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/completion.h>
31 #include <linux/kernel_stat.h>
32 #include <linux/security.h>
33 #include <linux/notifier.h>
34 #include <linux/profile.h>
35 #include <linux/suspend.h>
36 #include <linux/blkdev.h>
37 #include <linux/delay.h>
38 #include <linux/smp.h>
39 #include <linux/threads.h>
40 #include <linux/timer.h>
41 #include <linux/rcupdate.h>
42 #include <linux/cpu.h>
43 #include <linux/cpuset.h>
44 #include <linux/percpu.h>
45 #include <linux/kthread.h>
46 #include <linux/seq_file.h>
47 #include <linux/syscalls.h>
48 #include <linux/times.h>
49 #include <linux/acct.h>
52 #include <asm/unistd.h>
55 * Convert user-nice values [ -20 ... 0 ... 19 ]
56 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
59 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
60 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
61 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
64 * 'User priority' is the nice value converted to something we
65 * can work with better when scaling various scheduler parameters,
66 * it's a [ 0 ... 39 ] range.
68 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
69 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
70 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
73 * Some helpers for converting nanosecond timing to jiffy resolution
75 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
76 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
79 * These are the 'tuning knobs' of the scheduler:
81 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
82 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
83 * Timeslices get refilled after they expire.
85 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
86 #define DEF_TIMESLICE (100 * HZ / 1000)
87 #define ON_RUNQUEUE_WEIGHT 30
88 #define CHILD_PENALTY 95
89 #define PARENT_PENALTY 100
91 #define PRIO_BONUS_RATIO 25
92 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
93 #define INTERACTIVE_DELTA 2
94 #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
95 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
96 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
99 * If a task is 'interactive' then we reinsert it in the active
100 * array after it has expired its current timeslice. (it will not
101 * continue to run immediately, it will still roundrobin with
102 * other interactive tasks.)
104 * This part scales the interactivity limit depending on niceness.
106 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
107 * Here are a few examples of different nice levels:
109 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
110 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
111 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
112 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
113 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
115 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
116 * priority range a task can explore, a value of '1' means the
117 * task is rated interactive.)
119 * Ie. nice +19 tasks can never get 'interactive' enough to be
120 * reinserted into the active array. And only heavily CPU-hog nice -20
121 * tasks will be expired. Default nice 0 tasks are somewhere between,
122 * it takes some effort for them to get interactive, but it's not
126 #define CURRENT_BONUS(p) \
127 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
130 #define GRANULARITY (10 * HZ / 1000 ? : 1)
133 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
134 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
137 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
138 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
141 #define SCALE(v1,v1_max,v2_max) \
142 (v1) * (v2_max) / (v1_max)
145 (SCALE(TASK_NICE(p), 40, MAX_BONUS) + INTERACTIVE_DELTA)
147 #define TASK_INTERACTIVE(p) \
148 ((p)->prio <= (p)->static_prio - DELTA(p))
150 #define INTERACTIVE_SLEEP(p) \
151 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
152 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
154 #define TASK_PREEMPTS_CURR(p, rq) \
155 ((p)->prio < (rq)->curr->prio)
158 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
159 * to time slice values: [800ms ... 100ms ... 5ms]
161 * The higher a thread's priority, the bigger timeslices
162 * it gets during one round of execution. But even the lowest
163 * priority thread gets MIN_TIMESLICE worth of execution time.
166 #define SCALE_PRIO(x, prio) \
167 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO/2), MIN_TIMESLICE)
169 static unsigned int task_timeslice(task_t
*p
)
171 if (p
->static_prio
< NICE_TO_PRIO(0))
172 return SCALE_PRIO(DEF_TIMESLICE
*4, p
->static_prio
);
174 return SCALE_PRIO(DEF_TIMESLICE
, p
->static_prio
);
176 #define task_hot(p, now, sd) ((long long) ((now) - (p)->last_ran) \
177 < (long long) (sd)->cache_hot_time)
180 * These are the runqueue data structures:
183 #define BITMAP_SIZE ((((MAX_PRIO+1+7)/8)+sizeof(long)-1)/sizeof(long))
185 typedef struct runqueue runqueue_t
;
188 unsigned int nr_active
;
189 unsigned long bitmap
[BITMAP_SIZE
];
190 struct list_head queue
[MAX_PRIO
];
194 * This is the main, per-CPU runqueue data structure.
196 * Locking rule: those places that want to lock multiple runqueues
197 * (such as the load balancing or the thread migration code), lock
198 * acquire operations must be ordered by ascending &runqueue.
204 * nr_running and cpu_load should be in the same cacheline because
205 * remote CPUs use both these fields when doing load calculation.
207 unsigned long nr_running
;
209 unsigned long prio_bias
;
210 unsigned long cpu_load
[3];
212 unsigned long long nr_switches
;
215 * This is part of a global counter where only the total sum
216 * over all CPUs matters. A task can increase this counter on
217 * one CPU and if it got migrated afterwards it may decrease
218 * it on another CPU. Always updated under the runqueue lock:
220 unsigned long nr_uninterruptible
;
222 unsigned long expired_timestamp
;
223 unsigned long long timestamp_last_tick
;
225 struct mm_struct
*prev_mm
;
226 prio_array_t
*active
, *expired
, arrays
[2];
227 int best_expired_prio
;
231 struct sched_domain
*sd
;
233 /* For active balancing */
237 task_t
*migration_thread
;
238 struct list_head migration_queue
;
241 #ifdef CONFIG_SCHEDSTATS
243 struct sched_info rq_sched_info
;
245 /* sys_sched_yield() stats */
246 unsigned long yld_exp_empty
;
247 unsigned long yld_act_empty
;
248 unsigned long yld_both_empty
;
249 unsigned long yld_cnt
;
251 /* schedule() stats */
252 unsigned long sched_switch
;
253 unsigned long sched_cnt
;
254 unsigned long sched_goidle
;
256 /* try_to_wake_up() stats */
257 unsigned long ttwu_cnt
;
258 unsigned long ttwu_local
;
262 static DEFINE_PER_CPU(struct runqueue
, runqueues
);
265 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
266 * See detach_destroy_domains: synchronize_sched for details.
268 * The domain tree of any CPU may only be accessed from within
269 * preempt-disabled sections.
271 #define for_each_domain(cpu, domain) \
272 for (domain = rcu_dereference(cpu_rq(cpu)->sd); domain; domain = domain->parent)
274 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
275 #define this_rq() (&__get_cpu_var(runqueues))
276 #define task_rq(p) cpu_rq(task_cpu(p))
277 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
279 #ifndef prepare_arch_switch
280 # define prepare_arch_switch(next) do { } while (0)
282 #ifndef finish_arch_switch
283 # define finish_arch_switch(prev) do { } while (0)
286 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
287 static inline int task_running(runqueue_t
*rq
, task_t
*p
)
289 return rq
->curr
== p
;
292 static inline void prepare_lock_switch(runqueue_t
*rq
, task_t
*next
)
296 static inline void finish_lock_switch(runqueue_t
*rq
, task_t
*prev
)
298 #ifdef CONFIG_DEBUG_SPINLOCK
299 /* this is a valid case when another task releases the spinlock */
300 rq
->lock
.owner
= current
;
302 spin_unlock_irq(&rq
->lock
);
305 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
306 static inline int task_running(runqueue_t
*rq
, task_t
*p
)
311 return rq
->curr
== p
;
315 static inline void prepare_lock_switch(runqueue_t
*rq
, task_t
*next
)
319 * We can optimise this out completely for !SMP, because the
320 * SMP rebalancing from interrupt is the only thing that cares
325 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
326 spin_unlock_irq(&rq
->lock
);
328 spin_unlock(&rq
->lock
);
332 static inline void finish_lock_switch(runqueue_t
*rq
, task_t
*prev
)
336 * After ->oncpu is cleared, the task can be moved to a different CPU.
337 * We must ensure this doesn't happen until the switch is completely
343 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
347 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
350 * task_rq_lock - lock the runqueue a given task resides on and disable
351 * interrupts. Note the ordering: we can safely lookup the task_rq without
352 * explicitly disabling preemption.
354 static inline runqueue_t
*task_rq_lock(task_t
*p
, unsigned long *flags
)
360 local_irq_save(*flags
);
362 spin_lock(&rq
->lock
);
363 if (unlikely(rq
!= task_rq(p
))) {
364 spin_unlock_irqrestore(&rq
->lock
, *flags
);
365 goto repeat_lock_task
;
370 static inline void task_rq_unlock(runqueue_t
*rq
, unsigned long *flags
)
373 spin_unlock_irqrestore(&rq
->lock
, *flags
);
376 #ifdef CONFIG_SCHEDSTATS
378 * bump this up when changing the output format or the meaning of an existing
379 * format, so that tools can adapt (or abort)
381 #define SCHEDSTAT_VERSION 12
383 static int show_schedstat(struct seq_file
*seq
, void *v
)
387 seq_printf(seq
, "version %d\n", SCHEDSTAT_VERSION
);
388 seq_printf(seq
, "timestamp %lu\n", jiffies
);
389 for_each_online_cpu(cpu
) {
390 runqueue_t
*rq
= cpu_rq(cpu
);
392 struct sched_domain
*sd
;
396 /* runqueue-specific stats */
398 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
399 cpu
, rq
->yld_both_empty
,
400 rq
->yld_act_empty
, rq
->yld_exp_empty
, rq
->yld_cnt
,
401 rq
->sched_switch
, rq
->sched_cnt
, rq
->sched_goidle
,
402 rq
->ttwu_cnt
, rq
->ttwu_local
,
403 rq
->rq_sched_info
.cpu_time
,
404 rq
->rq_sched_info
.run_delay
, rq
->rq_sched_info
.pcnt
);
406 seq_printf(seq
, "\n");
409 /* domain-specific stats */
411 for_each_domain(cpu
, sd
) {
412 enum idle_type itype
;
413 char mask_str
[NR_CPUS
];
415 cpumask_scnprintf(mask_str
, NR_CPUS
, sd
->span
);
416 seq_printf(seq
, "domain%d %s", dcnt
++, mask_str
);
417 for (itype
= SCHED_IDLE
; itype
< MAX_IDLE_TYPES
;
419 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu %lu",
421 sd
->lb_balanced
[itype
],
422 sd
->lb_failed
[itype
],
423 sd
->lb_imbalance
[itype
],
424 sd
->lb_gained
[itype
],
425 sd
->lb_hot_gained
[itype
],
426 sd
->lb_nobusyq
[itype
],
427 sd
->lb_nobusyg
[itype
]);
429 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
430 sd
->alb_cnt
, sd
->alb_failed
, sd
->alb_pushed
,
431 sd
->sbe_cnt
, sd
->sbe_balanced
, sd
->sbe_pushed
,
432 sd
->sbf_cnt
, sd
->sbf_balanced
, sd
->sbf_pushed
,
433 sd
->ttwu_wake_remote
, sd
->ttwu_move_affine
, sd
->ttwu_move_balance
);
441 static int schedstat_open(struct inode
*inode
, struct file
*file
)
443 unsigned int size
= PAGE_SIZE
* (1 + num_online_cpus() / 32);
444 char *buf
= kmalloc(size
, GFP_KERNEL
);
450 res
= single_open(file
, show_schedstat
, NULL
);
452 m
= file
->private_data
;
460 struct file_operations proc_schedstat_operations
= {
461 .open
= schedstat_open
,
464 .release
= single_release
,
467 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
468 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
469 #else /* !CONFIG_SCHEDSTATS */
470 # define schedstat_inc(rq, field) do { } while (0)
471 # define schedstat_add(rq, field, amt) do { } while (0)
475 * rq_lock - lock a given runqueue and disable interrupts.
477 static inline runqueue_t
*this_rq_lock(void)
484 spin_lock(&rq
->lock
);
489 #ifdef CONFIG_SCHEDSTATS
491 * Called when a process is dequeued from the active array and given
492 * the cpu. We should note that with the exception of interactive
493 * tasks, the expired queue will become the active queue after the active
494 * queue is empty, without explicitly dequeuing and requeuing tasks in the
495 * expired queue. (Interactive tasks may be requeued directly to the
496 * active queue, thus delaying tasks in the expired queue from running;
497 * see scheduler_tick()).
499 * This function is only called from sched_info_arrive(), rather than
500 * dequeue_task(). Even though a task may be queued and dequeued multiple
501 * times as it is shuffled about, we're really interested in knowing how
502 * long it was from the *first* time it was queued to the time that it
505 static inline void sched_info_dequeued(task_t
*t
)
507 t
->sched_info
.last_queued
= 0;
511 * Called when a task finally hits the cpu. We can now calculate how
512 * long it was waiting to run. We also note when it began so that we
513 * can keep stats on how long its timeslice is.
515 static inline void sched_info_arrive(task_t
*t
)
517 unsigned long now
= jiffies
, diff
= 0;
518 struct runqueue
*rq
= task_rq(t
);
520 if (t
->sched_info
.last_queued
)
521 diff
= now
- t
->sched_info
.last_queued
;
522 sched_info_dequeued(t
);
523 t
->sched_info
.run_delay
+= diff
;
524 t
->sched_info
.last_arrival
= now
;
525 t
->sched_info
.pcnt
++;
530 rq
->rq_sched_info
.run_delay
+= diff
;
531 rq
->rq_sched_info
.pcnt
++;
535 * Called when a process is queued into either the active or expired
536 * array. The time is noted and later used to determine how long we
537 * had to wait for us to reach the cpu. Since the expired queue will
538 * become the active queue after active queue is empty, without dequeuing
539 * and requeuing any tasks, we are interested in queuing to either. It
540 * is unusual but not impossible for tasks to be dequeued and immediately
541 * requeued in the same or another array: this can happen in sched_yield(),
542 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
545 * This function is only called from enqueue_task(), but also only updates
546 * the timestamp if it is already not set. It's assumed that
547 * sched_info_dequeued() will clear that stamp when appropriate.
549 static inline void sched_info_queued(task_t
*t
)
551 if (!t
->sched_info
.last_queued
)
552 t
->sched_info
.last_queued
= jiffies
;
556 * Called when a process ceases being the active-running process, either
557 * voluntarily or involuntarily. Now we can calculate how long we ran.
559 static inline void sched_info_depart(task_t
*t
)
561 struct runqueue
*rq
= task_rq(t
);
562 unsigned long diff
= jiffies
- t
->sched_info
.last_arrival
;
564 t
->sched_info
.cpu_time
+= diff
;
567 rq
->rq_sched_info
.cpu_time
+= diff
;
571 * Called when tasks are switched involuntarily due, typically, to expiring
572 * their time slice. (This may also be called when switching to or from
573 * the idle task.) We are only called when prev != next.
575 static inline void sched_info_switch(task_t
*prev
, task_t
*next
)
577 struct runqueue
*rq
= task_rq(prev
);
580 * prev now departs the cpu. It's not interesting to record
581 * stats about how efficient we were at scheduling the idle
584 if (prev
!= rq
->idle
)
585 sched_info_depart(prev
);
587 if (next
!= rq
->idle
)
588 sched_info_arrive(next
);
591 #define sched_info_queued(t) do { } while (0)
592 #define sched_info_switch(t, next) do { } while (0)
593 #endif /* CONFIG_SCHEDSTATS */
596 * Adding/removing a task to/from a priority array:
598 static void dequeue_task(struct task_struct
*p
, prio_array_t
*array
)
601 list_del(&p
->run_list
);
602 if (list_empty(array
->queue
+ p
->prio
))
603 __clear_bit(p
->prio
, array
->bitmap
);
606 static void enqueue_task(struct task_struct
*p
, prio_array_t
*array
)
608 sched_info_queued(p
);
609 list_add_tail(&p
->run_list
, array
->queue
+ p
->prio
);
610 __set_bit(p
->prio
, array
->bitmap
);
616 * Put task to the end of the run list without the overhead of dequeue
617 * followed by enqueue.
619 static void requeue_task(struct task_struct
*p
, prio_array_t
*array
)
621 list_move_tail(&p
->run_list
, array
->queue
+ p
->prio
);
624 static inline void enqueue_task_head(struct task_struct
*p
, prio_array_t
*array
)
626 list_add(&p
->run_list
, array
->queue
+ p
->prio
);
627 __set_bit(p
->prio
, array
->bitmap
);
633 * effective_prio - return the priority that is based on the static
634 * priority but is modified by bonuses/penalties.
636 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
637 * into the -5 ... 0 ... +5 bonus/penalty range.
639 * We use 25% of the full 0...39 priority range so that:
641 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
642 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
644 * Both properties are important to certain workloads.
646 static int effective_prio(task_t
*p
)
653 bonus
= CURRENT_BONUS(p
) - MAX_BONUS
/ 2;
655 prio
= p
->static_prio
- bonus
;
656 if (prio
< MAX_RT_PRIO
)
658 if (prio
> MAX_PRIO
-1)
664 static inline void inc_prio_bias(runqueue_t
*rq
, int prio
)
666 rq
->prio_bias
+= MAX_PRIO
- prio
;
669 static inline void dec_prio_bias(runqueue_t
*rq
, int prio
)
671 rq
->prio_bias
-= MAX_PRIO
- prio
;
674 static inline void inc_nr_running(task_t
*p
, runqueue_t
*rq
)
678 if (p
!= rq
->migration_thread
)
680 * The migration thread does the actual balancing. Do
681 * not bias by its priority as the ultra high priority
682 * will skew balancing adversely.
684 inc_prio_bias(rq
, p
->prio
);
686 inc_prio_bias(rq
, p
->static_prio
);
689 static inline void dec_nr_running(task_t
*p
, runqueue_t
*rq
)
693 if (p
!= rq
->migration_thread
)
694 dec_prio_bias(rq
, p
->prio
);
696 dec_prio_bias(rq
, p
->static_prio
);
699 static inline void inc_prio_bias(runqueue_t
*rq
, int prio
)
703 static inline void dec_prio_bias(runqueue_t
*rq
, int prio
)
707 static inline void inc_nr_running(task_t
*p
, runqueue_t
*rq
)
712 static inline void dec_nr_running(task_t
*p
, runqueue_t
*rq
)
719 * __activate_task - move a task to the runqueue.
721 static inline void __activate_task(task_t
*p
, runqueue_t
*rq
)
723 enqueue_task(p
, rq
->active
);
724 inc_nr_running(p
, rq
);
728 * __activate_idle_task - move idle task to the _front_ of runqueue.
730 static inline void __activate_idle_task(task_t
*p
, runqueue_t
*rq
)
732 enqueue_task_head(p
, rq
->active
);
733 inc_nr_running(p
, rq
);
736 static int recalc_task_prio(task_t
*p
, unsigned long long now
)
738 /* Caller must always ensure 'now >= p->timestamp' */
739 unsigned long long __sleep_time
= now
- p
->timestamp
;
740 unsigned long sleep_time
;
742 if (__sleep_time
> NS_MAX_SLEEP_AVG
)
743 sleep_time
= NS_MAX_SLEEP_AVG
;
745 sleep_time
= (unsigned long)__sleep_time
;
747 if (likely(sleep_time
> 0)) {
749 * User tasks that sleep a long time are categorised as
750 * idle and will get just interactive status to stay active &
751 * prevent them suddenly becoming cpu hogs and starving
754 if (p
->mm
&& p
->activated
!= -1 &&
755 sleep_time
> INTERACTIVE_SLEEP(p
)) {
756 p
->sleep_avg
= JIFFIES_TO_NS(MAX_SLEEP_AVG
-
760 * The lower the sleep avg a task has the more
761 * rapidly it will rise with sleep time.
763 sleep_time
*= (MAX_BONUS
- CURRENT_BONUS(p
)) ? : 1;
766 * Tasks waking from uninterruptible sleep are
767 * limited in their sleep_avg rise as they
768 * are likely to be waiting on I/O
770 if (p
->activated
== -1 && p
->mm
) {
771 if (p
->sleep_avg
>= INTERACTIVE_SLEEP(p
))
773 else if (p
->sleep_avg
+ sleep_time
>=
774 INTERACTIVE_SLEEP(p
)) {
775 p
->sleep_avg
= INTERACTIVE_SLEEP(p
);
781 * This code gives a bonus to interactive tasks.
783 * The boost works by updating the 'average sleep time'
784 * value here, based on ->timestamp. The more time a
785 * task spends sleeping, the higher the average gets -
786 * and the higher the priority boost gets as well.
788 p
->sleep_avg
+= sleep_time
;
790 if (p
->sleep_avg
> NS_MAX_SLEEP_AVG
)
791 p
->sleep_avg
= NS_MAX_SLEEP_AVG
;
795 return effective_prio(p
);
799 * activate_task - move a task to the runqueue and do priority recalculation
801 * Update all the scheduling statistics stuff. (sleep average
802 * calculation, priority modifiers, etc.)
804 static void activate_task(task_t
*p
, runqueue_t
*rq
, int local
)
806 unsigned long long now
;
811 /* Compensate for drifting sched_clock */
812 runqueue_t
*this_rq
= this_rq();
813 now
= (now
- this_rq
->timestamp_last_tick
)
814 + rq
->timestamp_last_tick
;
818 p
->prio
= recalc_task_prio(p
, now
);
821 * This checks to make sure it's not an uninterruptible task
822 * that is now waking up.
826 * Tasks which were woken up by interrupts (ie. hw events)
827 * are most likely of interactive nature. So we give them
828 * the credit of extending their sleep time to the period
829 * of time they spend on the runqueue, waiting for execution
830 * on a CPU, first time around:
836 * Normal first-time wakeups get a credit too for
837 * on-runqueue time, but it will be weighted down:
844 __activate_task(p
, rq
);
848 * deactivate_task - remove a task from the runqueue.
850 static void deactivate_task(struct task_struct
*p
, runqueue_t
*rq
)
852 dec_nr_running(p
, rq
);
853 dequeue_task(p
, p
->array
);
858 * resched_task - mark a task 'to be rescheduled now'.
860 * On UP this means the setting of the need_resched flag, on SMP it
861 * might also involve a cross-CPU call to trigger the scheduler on
865 static void resched_task(task_t
*p
)
867 int need_resched
, nrpolling
;
869 assert_spin_locked(&task_rq(p
)->lock
);
871 /* minimise the chance of sending an interrupt to poll_idle() */
872 nrpolling
= test_tsk_thread_flag(p
,TIF_POLLING_NRFLAG
);
873 need_resched
= test_and_set_tsk_thread_flag(p
,TIF_NEED_RESCHED
);
874 nrpolling
|= test_tsk_thread_flag(p
,TIF_POLLING_NRFLAG
);
876 if (!need_resched
&& !nrpolling
&& (task_cpu(p
) != smp_processor_id()))
877 smp_send_reschedule(task_cpu(p
));
880 static inline void resched_task(task_t
*p
)
882 set_tsk_need_resched(p
);
887 * task_curr - is this task currently executing on a CPU?
888 * @p: the task in question.
890 inline int task_curr(const task_t
*p
)
892 return cpu_curr(task_cpu(p
)) == p
;
897 struct list_head list
;
902 struct completion done
;
906 * The task's runqueue lock must be held.
907 * Returns true if you have to wait for migration thread.
909 static int migrate_task(task_t
*p
, int dest_cpu
, migration_req_t
*req
)
911 runqueue_t
*rq
= task_rq(p
);
914 * If the task is not on a runqueue (and not running), then
915 * it is sufficient to simply update the task's cpu field.
917 if (!p
->array
&& !task_running(rq
, p
)) {
918 set_task_cpu(p
, dest_cpu
);
922 init_completion(&req
->done
);
924 req
->dest_cpu
= dest_cpu
;
925 list_add(&req
->list
, &rq
->migration_queue
);
930 * wait_task_inactive - wait for a thread to unschedule.
932 * The caller must ensure that the task *will* unschedule sometime soon,
933 * else this function might spin for a *long* time. This function can't
934 * be called with interrupts off, or it may introduce deadlock with
935 * smp_call_function() if an IPI is sent by the same process we are
936 * waiting to become inactive.
938 void wait_task_inactive(task_t
*p
)
945 rq
= task_rq_lock(p
, &flags
);
946 /* Must be off runqueue entirely, not preempted. */
947 if (unlikely(p
->array
|| task_running(rq
, p
))) {
948 /* If it's preempted, we yield. It could be a while. */
949 preempted
= !task_running(rq
, p
);
950 task_rq_unlock(rq
, &flags
);
956 task_rq_unlock(rq
, &flags
);
960 * kick_process - kick a running thread to enter/exit the kernel
961 * @p: the to-be-kicked thread
963 * Cause a process which is running on another CPU to enter
964 * kernel-mode, without any delay. (to get signals handled.)
966 * NOTE: this function doesnt have to take the runqueue lock,
967 * because all it wants to ensure is that the remote task enters
968 * the kernel. If the IPI races and the task has been migrated
969 * to another CPU then no harm is done and the purpose has been
972 void kick_process(task_t
*p
)
978 if ((cpu
!= smp_processor_id()) && task_curr(p
))
979 smp_send_reschedule(cpu
);
984 * Return a low guess at the load of a migration-source cpu.
986 * We want to under-estimate the load of migration sources, to
987 * balance conservatively.
989 static inline unsigned long __source_load(int cpu
, int type
, enum idle_type idle
)
991 runqueue_t
*rq
= cpu_rq(cpu
);
992 unsigned long running
= rq
->nr_running
;
993 unsigned long source_load
, cpu_load
= rq
->cpu_load
[type
-1],
994 load_now
= running
* SCHED_LOAD_SCALE
;
997 source_load
= load_now
;
999 source_load
= min(cpu_load
, load_now
);
1001 if (running
> 1 || (idle
== NOT_IDLE
&& running
))
1003 * If we are busy rebalancing the load is biased by
1004 * priority to create 'nice' support across cpus. When
1005 * idle rebalancing we should only bias the source_load if
1006 * there is more than one task running on that queue to
1007 * prevent idle rebalance from trying to pull tasks from a
1008 * queue with only one running task.
1010 source_load
= source_load
* rq
->prio_bias
/ running
;
1015 static inline unsigned long source_load(int cpu
, int type
)
1017 return __source_load(cpu
, type
, NOT_IDLE
);
1021 * Return a high guess at the load of a migration-target cpu
1023 static inline unsigned long __target_load(int cpu
, int type
, enum idle_type idle
)
1025 runqueue_t
*rq
= cpu_rq(cpu
);
1026 unsigned long running
= rq
->nr_running
;
1027 unsigned long target_load
, cpu_load
= rq
->cpu_load
[type
-1],
1028 load_now
= running
* SCHED_LOAD_SCALE
;
1031 target_load
= load_now
;
1033 target_load
= max(cpu_load
, load_now
);
1035 if (running
> 1 || (idle
== NOT_IDLE
&& running
))
1036 target_load
= target_load
* rq
->prio_bias
/ running
;
1041 static inline unsigned long target_load(int cpu
, int type
)
1043 return __target_load(cpu
, type
, NOT_IDLE
);
1047 * find_idlest_group finds and returns the least busy CPU group within the
1050 static struct sched_group
*
1051 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1053 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1054 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1055 int load_idx
= sd
->forkexec_idx
;
1056 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1059 unsigned long load
, avg_load
;
1063 /* Skip over this group if it has no CPUs allowed */
1064 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1067 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1069 /* Tally up the load of all CPUs in the group */
1072 for_each_cpu_mask(i
, group
->cpumask
) {
1073 /* Bias balancing toward cpus of our domain */
1075 load
= source_load(i
, load_idx
);
1077 load
= target_load(i
, load_idx
);
1082 /* Adjust by relative CPU power of the group */
1083 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
1086 this_load
= avg_load
;
1088 } else if (avg_load
< min_load
) {
1089 min_load
= avg_load
;
1093 group
= group
->next
;
1094 } while (group
!= sd
->groups
);
1096 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1102 * find_idlest_queue - find the idlest runqueue among the cpus in group.
1105 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1108 unsigned long load
, min_load
= ULONG_MAX
;
1112 /* Traverse only the allowed CPUs */
1113 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1115 for_each_cpu_mask(i
, tmp
) {
1116 load
= source_load(i
, 0);
1118 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1128 * sched_balance_self: balance the current task (running on cpu) in domains
1129 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1132 * Balance, ie. select the least loaded group.
1134 * Returns the target CPU number, or the same CPU if no balancing is needed.
1136 * preempt must be disabled.
1138 static int sched_balance_self(int cpu
, int flag
)
1140 struct task_struct
*t
= current
;
1141 struct sched_domain
*tmp
, *sd
= NULL
;
1143 for_each_domain(cpu
, tmp
)
1144 if (tmp
->flags
& flag
)
1149 struct sched_group
*group
;
1154 group
= find_idlest_group(sd
, t
, cpu
);
1158 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1159 if (new_cpu
== -1 || new_cpu
== cpu
)
1162 /* Now try balancing at a lower domain level */
1166 weight
= cpus_weight(span
);
1167 for_each_domain(cpu
, tmp
) {
1168 if (weight
<= cpus_weight(tmp
->span
))
1170 if (tmp
->flags
& flag
)
1173 /* while loop will break here if sd == NULL */
1179 #endif /* CONFIG_SMP */
1182 * wake_idle() will wake a task on an idle cpu if task->cpu is
1183 * not idle and an idle cpu is available. The span of cpus to
1184 * search starts with cpus closest then further out as needed,
1185 * so we always favor a closer, idle cpu.
1187 * Returns the CPU we should wake onto.
1189 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1190 static int wake_idle(int cpu
, task_t
*p
)
1193 struct sched_domain
*sd
;
1199 for_each_domain(cpu
, sd
) {
1200 if (sd
->flags
& SD_WAKE_IDLE
) {
1201 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1202 for_each_cpu_mask(i
, tmp
) {
1213 static inline int wake_idle(int cpu
, task_t
*p
)
1220 * try_to_wake_up - wake up a thread
1221 * @p: the to-be-woken-up thread
1222 * @state: the mask of task states that can be woken
1223 * @sync: do a synchronous wakeup?
1225 * Put it on the run-queue if it's not already there. The "current"
1226 * thread is always on the run-queue (except when the actual
1227 * re-schedule is in progress), and as such you're allowed to do
1228 * the simpler "current->state = TASK_RUNNING" to mark yourself
1229 * runnable without the overhead of this.
1231 * returns failure only if the task is already active.
1233 static int try_to_wake_up(task_t
*p
, unsigned int state
, int sync
)
1235 int cpu
, this_cpu
, success
= 0;
1236 unsigned long flags
;
1240 unsigned long load
, this_load
;
1241 struct sched_domain
*sd
, *this_sd
= NULL
;
1245 rq
= task_rq_lock(p
, &flags
);
1246 old_state
= p
->state
;
1247 if (!(old_state
& state
))
1254 this_cpu
= smp_processor_id();
1257 if (unlikely(task_running(rq
, p
)))
1262 schedstat_inc(rq
, ttwu_cnt
);
1263 if (cpu
== this_cpu
) {
1264 schedstat_inc(rq
, ttwu_local
);
1268 for_each_domain(this_cpu
, sd
) {
1269 if (cpu_isset(cpu
, sd
->span
)) {
1270 schedstat_inc(sd
, ttwu_wake_remote
);
1276 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1280 * Check for affine wakeup and passive balancing possibilities.
1283 int idx
= this_sd
->wake_idx
;
1284 unsigned int imbalance
;
1286 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1288 load
= source_load(cpu
, idx
);
1289 this_load
= target_load(this_cpu
, idx
);
1291 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1293 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1294 unsigned long tl
= this_load
;
1296 * If sync wakeup then subtract the (maximum possible)
1297 * effect of the currently running task from the load
1298 * of the current CPU:
1301 tl
-= SCHED_LOAD_SCALE
;
1304 tl
+ target_load(cpu
, idx
) <= SCHED_LOAD_SCALE
) ||
1305 100*(tl
+ SCHED_LOAD_SCALE
) <= imbalance
*load
) {
1307 * This domain has SD_WAKE_AFFINE and
1308 * p is cache cold in this domain, and
1309 * there is no bad imbalance.
1311 schedstat_inc(this_sd
, ttwu_move_affine
);
1317 * Start passive balancing when half the imbalance_pct
1320 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1321 if (imbalance
*this_load
<= 100*load
) {
1322 schedstat_inc(this_sd
, ttwu_move_balance
);
1328 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1330 new_cpu
= wake_idle(new_cpu
, p
);
1331 if (new_cpu
!= cpu
) {
1332 set_task_cpu(p
, new_cpu
);
1333 task_rq_unlock(rq
, &flags
);
1334 /* might preempt at this point */
1335 rq
= task_rq_lock(p
, &flags
);
1336 old_state
= p
->state
;
1337 if (!(old_state
& state
))
1342 this_cpu
= smp_processor_id();
1347 #endif /* CONFIG_SMP */
1348 if (old_state
== TASK_UNINTERRUPTIBLE
) {
1349 rq
->nr_uninterruptible
--;
1351 * Tasks on involuntary sleep don't earn
1352 * sleep_avg beyond just interactive state.
1358 * Tasks that have marked their sleep as noninteractive get
1359 * woken up without updating their sleep average. (i.e. their
1360 * sleep is handled in a priority-neutral manner, no priority
1361 * boost and no penalty.)
1363 if (old_state
& TASK_NONINTERACTIVE
)
1364 __activate_task(p
, rq
);
1366 activate_task(p
, rq
, cpu
== this_cpu
);
1368 * Sync wakeups (i.e. those types of wakeups where the waker
1369 * has indicated that it will leave the CPU in short order)
1370 * don't trigger a preemption, if the woken up task will run on
1371 * this cpu. (in this case the 'I will reschedule' promise of
1372 * the waker guarantees that the freshly woken up task is going
1373 * to be considered on this CPU.)
1375 if (!sync
|| cpu
!= this_cpu
) {
1376 if (TASK_PREEMPTS_CURR(p
, rq
))
1377 resched_task(rq
->curr
);
1382 p
->state
= TASK_RUNNING
;
1384 task_rq_unlock(rq
, &flags
);
1389 int fastcall
wake_up_process(task_t
*p
)
1391 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1392 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1395 EXPORT_SYMBOL(wake_up_process
);
1397 int fastcall
wake_up_state(task_t
*p
, unsigned int state
)
1399 return try_to_wake_up(p
, state
, 0);
1403 * Perform scheduler related setup for a newly forked process p.
1404 * p is forked by current.
1406 void fastcall
sched_fork(task_t
*p
, int clone_flags
)
1408 int cpu
= get_cpu();
1411 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1413 set_task_cpu(p
, cpu
);
1416 * We mark the process as running here, but have not actually
1417 * inserted it onto the runqueue yet. This guarantees that
1418 * nobody will actually run it, and a signal or other external
1419 * event cannot wake it up and insert it on the runqueue either.
1421 p
->state
= TASK_RUNNING
;
1422 INIT_LIST_HEAD(&p
->run_list
);
1424 #ifdef CONFIG_SCHEDSTATS
1425 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1427 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1430 #ifdef CONFIG_PREEMPT
1431 /* Want to start with kernel preemption disabled. */
1432 p
->thread_info
->preempt_count
= 1;
1435 * Share the timeslice between parent and child, thus the
1436 * total amount of pending timeslices in the system doesn't change,
1437 * resulting in more scheduling fairness.
1439 local_irq_disable();
1440 p
->time_slice
= (current
->time_slice
+ 1) >> 1;
1442 * The remainder of the first timeslice might be recovered by
1443 * the parent if the child exits early enough.
1445 p
->first_time_slice
= 1;
1446 current
->time_slice
>>= 1;
1447 p
->timestamp
= sched_clock();
1448 if (unlikely(!current
->time_slice
)) {
1450 * This case is rare, it happens when the parent has only
1451 * a single jiffy left from its timeslice. Taking the
1452 * runqueue lock is not a problem.
1454 current
->time_slice
= 1;
1462 * wake_up_new_task - wake up a newly created task for the first time.
1464 * This function will do some initial scheduler statistics housekeeping
1465 * that must be done for every newly created context, then puts the task
1466 * on the runqueue and wakes it.
1468 void fastcall
wake_up_new_task(task_t
*p
, unsigned long clone_flags
)
1470 unsigned long flags
;
1472 runqueue_t
*rq
, *this_rq
;
1474 rq
= task_rq_lock(p
, &flags
);
1475 BUG_ON(p
->state
!= TASK_RUNNING
);
1476 this_cpu
= smp_processor_id();
1480 * We decrease the sleep average of forking parents
1481 * and children as well, to keep max-interactive tasks
1482 * from forking tasks that are max-interactive. The parent
1483 * (current) is done further down, under its lock.
1485 p
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(p
) *
1486 CHILD_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1488 p
->prio
= effective_prio(p
);
1490 if (likely(cpu
== this_cpu
)) {
1491 if (!(clone_flags
& CLONE_VM
)) {
1493 * The VM isn't cloned, so we're in a good position to
1494 * do child-runs-first in anticipation of an exec. This
1495 * usually avoids a lot of COW overhead.
1497 if (unlikely(!current
->array
))
1498 __activate_task(p
, rq
);
1500 p
->prio
= current
->prio
;
1501 list_add_tail(&p
->run_list
, ¤t
->run_list
);
1502 p
->array
= current
->array
;
1503 p
->array
->nr_active
++;
1504 inc_nr_running(p
, rq
);
1508 /* Run child last */
1509 __activate_task(p
, rq
);
1511 * We skip the following code due to cpu == this_cpu
1513 * task_rq_unlock(rq, &flags);
1514 * this_rq = task_rq_lock(current, &flags);
1518 this_rq
= cpu_rq(this_cpu
);
1521 * Not the local CPU - must adjust timestamp. This should
1522 * get optimised away in the !CONFIG_SMP case.
1524 p
->timestamp
= (p
->timestamp
- this_rq
->timestamp_last_tick
)
1525 + rq
->timestamp_last_tick
;
1526 __activate_task(p
, rq
);
1527 if (TASK_PREEMPTS_CURR(p
, rq
))
1528 resched_task(rq
->curr
);
1531 * Parent and child are on different CPUs, now get the
1532 * parent runqueue to update the parent's ->sleep_avg:
1534 task_rq_unlock(rq
, &flags
);
1535 this_rq
= task_rq_lock(current
, &flags
);
1537 current
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(current
) *
1538 PARENT_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1539 task_rq_unlock(this_rq
, &flags
);
1543 * Potentially available exiting-child timeslices are
1544 * retrieved here - this way the parent does not get
1545 * penalized for creating too many threads.
1547 * (this cannot be used to 'generate' timeslices
1548 * artificially, because any timeslice recovered here
1549 * was given away by the parent in the first place.)
1551 void fastcall
sched_exit(task_t
*p
)
1553 unsigned long flags
;
1557 * If the child was a (relative-) CPU hog then decrease
1558 * the sleep_avg of the parent as well.
1560 rq
= task_rq_lock(p
->parent
, &flags
);
1561 if (p
->first_time_slice
&& task_cpu(p
) == task_cpu(p
->parent
)) {
1562 p
->parent
->time_slice
+= p
->time_slice
;
1563 if (unlikely(p
->parent
->time_slice
> task_timeslice(p
)))
1564 p
->parent
->time_slice
= task_timeslice(p
);
1566 if (p
->sleep_avg
< p
->parent
->sleep_avg
)
1567 p
->parent
->sleep_avg
= p
->parent
->sleep_avg
/
1568 (EXIT_WEIGHT
+ 1) * EXIT_WEIGHT
+ p
->sleep_avg
/
1570 task_rq_unlock(rq
, &flags
);
1574 * prepare_task_switch - prepare to switch tasks
1575 * @rq: the runqueue preparing to switch
1576 * @next: the task we are going to switch to.
1578 * This is called with the rq lock held and interrupts off. It must
1579 * be paired with a subsequent finish_task_switch after the context
1582 * prepare_task_switch sets up locking and calls architecture specific
1585 static inline void prepare_task_switch(runqueue_t
*rq
, task_t
*next
)
1587 prepare_lock_switch(rq
, next
);
1588 prepare_arch_switch(next
);
1592 * finish_task_switch - clean up after a task-switch
1593 * @rq: runqueue associated with task-switch
1594 * @prev: the thread we just switched away from.
1596 * finish_task_switch must be called after the context switch, paired
1597 * with a prepare_task_switch call before the context switch.
1598 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1599 * and do any other architecture-specific cleanup actions.
1601 * Note that we may have delayed dropping an mm in context_switch(). If
1602 * so, we finish that here outside of the runqueue lock. (Doing it
1603 * with the lock held can cause deadlocks; see schedule() for
1606 static inline void finish_task_switch(runqueue_t
*rq
, task_t
*prev
)
1607 __releases(rq
->lock
)
1609 struct mm_struct
*mm
= rq
->prev_mm
;
1610 unsigned long prev_task_flags
;
1615 * A task struct has one reference for the use as "current".
1616 * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1617 * calls schedule one last time. The schedule call will never return,
1618 * and the scheduled task must drop that reference.
1619 * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1620 * still held, otherwise prev could be scheduled on another cpu, die
1621 * there before we look at prev->state, and then the reference would
1623 * Manfred Spraul <manfred@colorfullife.com>
1625 prev_task_flags
= prev
->flags
;
1626 finish_arch_switch(prev
);
1627 finish_lock_switch(rq
, prev
);
1630 if (unlikely(prev_task_flags
& PF_DEAD
))
1631 put_task_struct(prev
);
1635 * schedule_tail - first thing a freshly forked thread must call.
1636 * @prev: the thread we just switched away from.
1638 asmlinkage
void schedule_tail(task_t
*prev
)
1639 __releases(rq
->lock
)
1641 runqueue_t
*rq
= this_rq();
1642 finish_task_switch(rq
, prev
);
1643 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1644 /* In this case, finish_task_switch does not reenable preemption */
1647 if (current
->set_child_tid
)
1648 put_user(current
->pid
, current
->set_child_tid
);
1652 * context_switch - switch to the new MM and the new
1653 * thread's register state.
1656 task_t
* context_switch(runqueue_t
*rq
, task_t
*prev
, task_t
*next
)
1658 struct mm_struct
*mm
= next
->mm
;
1659 struct mm_struct
*oldmm
= prev
->active_mm
;
1661 if (unlikely(!mm
)) {
1662 next
->active_mm
= oldmm
;
1663 atomic_inc(&oldmm
->mm_count
);
1664 enter_lazy_tlb(oldmm
, next
);
1666 switch_mm(oldmm
, mm
, next
);
1668 if (unlikely(!prev
->mm
)) {
1669 prev
->active_mm
= NULL
;
1670 WARN_ON(rq
->prev_mm
);
1671 rq
->prev_mm
= oldmm
;
1674 /* Here we just switch the register state and the stack. */
1675 switch_to(prev
, next
, prev
);
1681 * nr_running, nr_uninterruptible and nr_context_switches:
1683 * externally visible scheduler statistics: current number of runnable
1684 * threads, current number of uninterruptible-sleeping threads, total
1685 * number of context switches performed since bootup.
1687 unsigned long nr_running(void)
1689 unsigned long i
, sum
= 0;
1691 for_each_online_cpu(i
)
1692 sum
+= cpu_rq(i
)->nr_running
;
1697 unsigned long nr_uninterruptible(void)
1699 unsigned long i
, sum
= 0;
1702 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1705 * Since we read the counters lockless, it might be slightly
1706 * inaccurate. Do not allow it to go below zero though:
1708 if (unlikely((long)sum
< 0))
1714 unsigned long long nr_context_switches(void)
1716 unsigned long long i
, sum
= 0;
1719 sum
+= cpu_rq(i
)->nr_switches
;
1724 unsigned long nr_iowait(void)
1726 unsigned long i
, sum
= 0;
1729 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1737 * double_rq_lock - safely lock two runqueues
1739 * Note this does not disable interrupts like task_rq_lock,
1740 * you need to do so manually before calling.
1742 static void double_rq_lock(runqueue_t
*rq1
, runqueue_t
*rq2
)
1743 __acquires(rq1
->lock
)
1744 __acquires(rq2
->lock
)
1747 spin_lock(&rq1
->lock
);
1748 __acquire(rq2
->lock
); /* Fake it out ;) */
1751 spin_lock(&rq1
->lock
);
1752 spin_lock(&rq2
->lock
);
1754 spin_lock(&rq2
->lock
);
1755 spin_lock(&rq1
->lock
);
1761 * double_rq_unlock - safely unlock two runqueues
1763 * Note this does not restore interrupts like task_rq_unlock,
1764 * you need to do so manually after calling.
1766 static void double_rq_unlock(runqueue_t
*rq1
, runqueue_t
*rq2
)
1767 __releases(rq1
->lock
)
1768 __releases(rq2
->lock
)
1770 spin_unlock(&rq1
->lock
);
1772 spin_unlock(&rq2
->lock
);
1774 __release(rq2
->lock
);
1778 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1780 static void double_lock_balance(runqueue_t
*this_rq
, runqueue_t
*busiest
)
1781 __releases(this_rq
->lock
)
1782 __acquires(busiest
->lock
)
1783 __acquires(this_rq
->lock
)
1785 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1786 if (busiest
< this_rq
) {
1787 spin_unlock(&this_rq
->lock
);
1788 spin_lock(&busiest
->lock
);
1789 spin_lock(&this_rq
->lock
);
1791 spin_lock(&busiest
->lock
);
1796 * If dest_cpu is allowed for this process, migrate the task to it.
1797 * This is accomplished by forcing the cpu_allowed mask to only
1798 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
1799 * the cpu_allowed mask is restored.
1801 static void sched_migrate_task(task_t
*p
, int dest_cpu
)
1803 migration_req_t req
;
1805 unsigned long flags
;
1807 rq
= task_rq_lock(p
, &flags
);
1808 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
1809 || unlikely(cpu_is_offline(dest_cpu
)))
1812 /* force the process onto the specified CPU */
1813 if (migrate_task(p
, dest_cpu
, &req
)) {
1814 /* Need to wait for migration thread (might exit: take ref). */
1815 struct task_struct
*mt
= rq
->migration_thread
;
1816 get_task_struct(mt
);
1817 task_rq_unlock(rq
, &flags
);
1818 wake_up_process(mt
);
1819 put_task_struct(mt
);
1820 wait_for_completion(&req
.done
);
1824 task_rq_unlock(rq
, &flags
);
1828 * sched_exec - execve() is a valuable balancing opportunity, because at
1829 * this point the task has the smallest effective memory and cache footprint.
1831 void sched_exec(void)
1833 int new_cpu
, this_cpu
= get_cpu();
1834 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
1836 if (new_cpu
!= this_cpu
)
1837 sched_migrate_task(current
, new_cpu
);
1841 * pull_task - move a task from a remote runqueue to the local runqueue.
1842 * Both runqueues must be locked.
1845 void pull_task(runqueue_t
*src_rq
, prio_array_t
*src_array
, task_t
*p
,
1846 runqueue_t
*this_rq
, prio_array_t
*this_array
, int this_cpu
)
1848 dequeue_task(p
, src_array
);
1849 dec_nr_running(p
, src_rq
);
1850 set_task_cpu(p
, this_cpu
);
1851 inc_nr_running(p
, this_rq
);
1852 enqueue_task(p
, this_array
);
1853 p
->timestamp
= (p
->timestamp
- src_rq
->timestamp_last_tick
)
1854 + this_rq
->timestamp_last_tick
;
1856 * Note that idle threads have a prio of MAX_PRIO, for this test
1857 * to be always true for them.
1859 if (TASK_PREEMPTS_CURR(p
, this_rq
))
1860 resched_task(this_rq
->curr
);
1864 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1867 int can_migrate_task(task_t
*p
, runqueue_t
*rq
, int this_cpu
,
1868 struct sched_domain
*sd
, enum idle_type idle
,
1872 * We do not migrate tasks that are:
1873 * 1) running (obviously), or
1874 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1875 * 3) are cache-hot on their current CPU.
1877 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
1881 if (task_running(rq
, p
))
1885 * Aggressive migration if:
1886 * 1) task is cache cold, or
1887 * 2) too many balance attempts have failed.
1890 if (sd
->nr_balance_failed
> sd
->cache_nice_tries
)
1893 if (task_hot(p
, rq
->timestamp_last_tick
, sd
))
1899 * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq,
1900 * as part of a balancing operation within "domain". Returns the number of
1903 * Called with both runqueues locked.
1905 static int move_tasks(runqueue_t
*this_rq
, int this_cpu
, runqueue_t
*busiest
,
1906 unsigned long max_nr_move
, struct sched_domain
*sd
,
1907 enum idle_type idle
, int *all_pinned
)
1909 prio_array_t
*array
, *dst_array
;
1910 struct list_head
*head
, *curr
;
1911 int idx
, pulled
= 0, pinned
= 0;
1914 if (max_nr_move
== 0)
1920 * We first consider expired tasks. Those will likely not be
1921 * executed in the near future, and they are most likely to
1922 * be cache-cold, thus switching CPUs has the least effect
1925 if (busiest
->expired
->nr_active
) {
1926 array
= busiest
->expired
;
1927 dst_array
= this_rq
->expired
;
1929 array
= busiest
->active
;
1930 dst_array
= this_rq
->active
;
1934 /* Start searching at priority 0: */
1938 idx
= sched_find_first_bit(array
->bitmap
);
1940 idx
= find_next_bit(array
->bitmap
, MAX_PRIO
, idx
);
1941 if (idx
>= MAX_PRIO
) {
1942 if (array
== busiest
->expired
&& busiest
->active
->nr_active
) {
1943 array
= busiest
->active
;
1944 dst_array
= this_rq
->active
;
1950 head
= array
->queue
+ idx
;
1953 tmp
= list_entry(curr
, task_t
, run_list
);
1957 if (!can_migrate_task(tmp
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
1964 #ifdef CONFIG_SCHEDSTATS
1965 if (task_hot(tmp
, busiest
->timestamp_last_tick
, sd
))
1966 schedstat_inc(sd
, lb_hot_gained
[idle
]);
1969 pull_task(busiest
, array
, tmp
, this_rq
, dst_array
, this_cpu
);
1972 /* We only want to steal up to the prescribed number of tasks. */
1973 if (pulled
< max_nr_move
) {
1981 * Right now, this is the only place pull_task() is called,
1982 * so we can safely collect pull_task() stats here rather than
1983 * inside pull_task().
1985 schedstat_add(sd
, lb_gained
[idle
], pulled
);
1988 *all_pinned
= pinned
;
1993 * find_busiest_group finds and returns the busiest CPU group within the
1994 * domain. It calculates and returns the number of tasks which should be
1995 * moved to restore balance via the imbalance parameter.
1997 static struct sched_group
*
1998 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
1999 unsigned long *imbalance
, enum idle_type idle
, int *sd_idle
)
2001 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2002 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2003 unsigned long max_pull
;
2006 max_load
= this_load
= total_load
= total_pwr
= 0;
2007 if (idle
== NOT_IDLE
)
2008 load_idx
= sd
->busy_idx
;
2009 else if (idle
== NEWLY_IDLE
)
2010 load_idx
= sd
->newidle_idx
;
2012 load_idx
= sd
->idle_idx
;
2019 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2021 /* Tally up the load of all CPUs in the group */
2024 for_each_cpu_mask(i
, group
->cpumask
) {
2025 if (*sd_idle
&& !idle_cpu(i
))
2028 /* Bias balancing toward cpus of our domain */
2030 load
= __target_load(i
, load_idx
, idle
);
2032 load
= __source_load(i
, load_idx
, idle
);
2037 total_load
+= avg_load
;
2038 total_pwr
+= group
->cpu_power
;
2040 /* Adjust by relative CPU power of the group */
2041 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
2044 this_load
= avg_load
;
2046 } else if (avg_load
> max_load
) {
2047 max_load
= avg_load
;
2050 group
= group
->next
;
2051 } while (group
!= sd
->groups
);
2053 if (!busiest
|| this_load
>= max_load
|| max_load
<= SCHED_LOAD_SCALE
)
2056 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2058 if (this_load
>= avg_load
||
2059 100*max_load
<= sd
->imbalance_pct
*this_load
)
2063 * We're trying to get all the cpus to the average_load, so we don't
2064 * want to push ourselves above the average load, nor do we wish to
2065 * reduce the max loaded cpu below the average load, as either of these
2066 * actions would just result in more rebalancing later, and ping-pong
2067 * tasks around. Thus we look for the minimum possible imbalance.
2068 * Negative imbalances (*we* are more loaded than anyone else) will
2069 * be counted as no imbalance for these purposes -- we can't fix that
2070 * by pulling tasks to us. Be careful of negative numbers as they'll
2071 * appear as very large values with unsigned longs.
2074 /* Don't want to pull so many tasks that a group would go idle */
2075 max_pull
= min(max_load
- avg_load
, max_load
- SCHED_LOAD_SCALE
);
2077 /* How much load to actually move to equalise the imbalance */
2078 *imbalance
= min(max_pull
* busiest
->cpu_power
,
2079 (avg_load
- this_load
) * this->cpu_power
)
2082 if (*imbalance
< SCHED_LOAD_SCALE
) {
2083 unsigned long pwr_now
= 0, pwr_move
= 0;
2086 if (max_load
- this_load
>= SCHED_LOAD_SCALE
*2) {
2092 * OK, we don't have enough imbalance to justify moving tasks,
2093 * however we may be able to increase total CPU power used by
2097 pwr_now
+= busiest
->cpu_power
*min(SCHED_LOAD_SCALE
, max_load
);
2098 pwr_now
+= this->cpu_power
*min(SCHED_LOAD_SCALE
, this_load
);
2099 pwr_now
/= SCHED_LOAD_SCALE
;
2101 /* Amount of load we'd subtract */
2102 tmp
= SCHED_LOAD_SCALE
*SCHED_LOAD_SCALE
/busiest
->cpu_power
;
2104 pwr_move
+= busiest
->cpu_power
*min(SCHED_LOAD_SCALE
,
2107 /* Amount of load we'd add */
2108 if (max_load
*busiest
->cpu_power
<
2109 SCHED_LOAD_SCALE
*SCHED_LOAD_SCALE
)
2110 tmp
= max_load
*busiest
->cpu_power
/this->cpu_power
;
2112 tmp
= SCHED_LOAD_SCALE
*SCHED_LOAD_SCALE
/this->cpu_power
;
2113 pwr_move
+= this->cpu_power
*min(SCHED_LOAD_SCALE
, this_load
+ tmp
);
2114 pwr_move
/= SCHED_LOAD_SCALE
;
2116 /* Move if we gain throughput */
2117 if (pwr_move
<= pwr_now
)
2124 /* Get rid of the scaling factor, rounding down as we divide */
2125 *imbalance
= *imbalance
/ SCHED_LOAD_SCALE
;
2135 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2137 static runqueue_t
*find_busiest_queue(struct sched_group
*group
,
2138 enum idle_type idle
)
2140 unsigned long load
, max_load
= 0;
2141 runqueue_t
*busiest
= NULL
;
2144 for_each_cpu_mask(i
, group
->cpumask
) {
2145 load
= __source_load(i
, 0, idle
);
2147 if (load
> max_load
) {
2149 busiest
= cpu_rq(i
);
2157 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2158 * so long as it is large enough.
2160 #define MAX_PINNED_INTERVAL 512
2163 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2164 * tasks if there is an imbalance.
2166 * Called with this_rq unlocked.
2168 static int load_balance(int this_cpu
, runqueue_t
*this_rq
,
2169 struct sched_domain
*sd
, enum idle_type idle
)
2171 struct sched_group
*group
;
2172 runqueue_t
*busiest
;
2173 unsigned long imbalance
;
2174 int nr_moved
, all_pinned
= 0;
2175 int active_balance
= 0;
2178 if (idle
!= NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2181 schedstat_inc(sd
, lb_cnt
[idle
]);
2183 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
);
2185 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2189 busiest
= find_busiest_queue(group
, idle
);
2191 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2195 BUG_ON(busiest
== this_rq
);
2197 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2200 if (busiest
->nr_running
> 1) {
2202 * Attempt to move tasks. If find_busiest_group has found
2203 * an imbalance but busiest->nr_running <= 1, the group is
2204 * still unbalanced. nr_moved simply stays zero, so it is
2205 * correctly treated as an imbalance.
2207 double_rq_lock(this_rq
, busiest
);
2208 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2209 imbalance
, sd
, idle
, &all_pinned
);
2210 double_rq_unlock(this_rq
, busiest
);
2212 /* All tasks on this runqueue were pinned by CPU affinity */
2213 if (unlikely(all_pinned
))
2218 schedstat_inc(sd
, lb_failed
[idle
]);
2219 sd
->nr_balance_failed
++;
2221 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2223 spin_lock(&busiest
->lock
);
2225 /* don't kick the migration_thread, if the curr
2226 * task on busiest cpu can't be moved to this_cpu
2228 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2229 spin_unlock(&busiest
->lock
);
2231 goto out_one_pinned
;
2234 if (!busiest
->active_balance
) {
2235 busiest
->active_balance
= 1;
2236 busiest
->push_cpu
= this_cpu
;
2239 spin_unlock(&busiest
->lock
);
2241 wake_up_process(busiest
->migration_thread
);
2244 * We've kicked active balancing, reset the failure
2247 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2250 sd
->nr_balance_failed
= 0;
2252 if (likely(!active_balance
)) {
2253 /* We were unbalanced, so reset the balancing interval */
2254 sd
->balance_interval
= sd
->min_interval
;
2257 * If we've begun active balancing, start to back off. This
2258 * case may not be covered by the all_pinned logic if there
2259 * is only 1 task on the busy runqueue (because we don't call
2262 if (sd
->balance_interval
< sd
->max_interval
)
2263 sd
->balance_interval
*= 2;
2266 if (!nr_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2271 schedstat_inc(sd
, lb_balanced
[idle
]);
2273 sd
->nr_balance_failed
= 0;
2276 /* tune up the balancing interval */
2277 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2278 (sd
->balance_interval
< sd
->max_interval
))
2279 sd
->balance_interval
*= 2;
2281 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2287 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2288 * tasks if there is an imbalance.
2290 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2291 * this_rq is locked.
2293 static int load_balance_newidle(int this_cpu
, runqueue_t
*this_rq
,
2294 struct sched_domain
*sd
)
2296 struct sched_group
*group
;
2297 runqueue_t
*busiest
= NULL
;
2298 unsigned long imbalance
;
2302 if (sd
->flags
& SD_SHARE_CPUPOWER
)
2305 schedstat_inc(sd
, lb_cnt
[NEWLY_IDLE
]);
2306 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, NEWLY_IDLE
, &sd_idle
);
2308 schedstat_inc(sd
, lb_nobusyg
[NEWLY_IDLE
]);
2312 busiest
= find_busiest_queue(group
, NEWLY_IDLE
);
2314 schedstat_inc(sd
, lb_nobusyq
[NEWLY_IDLE
]);
2318 BUG_ON(busiest
== this_rq
);
2320 schedstat_add(sd
, lb_imbalance
[NEWLY_IDLE
], imbalance
);
2323 if (busiest
->nr_running
> 1) {
2324 /* Attempt to move tasks */
2325 double_lock_balance(this_rq
, busiest
);
2326 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2327 imbalance
, sd
, NEWLY_IDLE
, NULL
);
2328 spin_unlock(&busiest
->lock
);
2332 schedstat_inc(sd
, lb_failed
[NEWLY_IDLE
]);
2333 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2336 sd
->nr_balance_failed
= 0;
2341 schedstat_inc(sd
, lb_balanced
[NEWLY_IDLE
]);
2342 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2344 sd
->nr_balance_failed
= 0;
2349 * idle_balance is called by schedule() if this_cpu is about to become
2350 * idle. Attempts to pull tasks from other CPUs.
2352 static inline void idle_balance(int this_cpu
, runqueue_t
*this_rq
)
2354 struct sched_domain
*sd
;
2356 for_each_domain(this_cpu
, sd
) {
2357 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
2358 if (load_balance_newidle(this_cpu
, this_rq
, sd
)) {
2359 /* We've pulled tasks over so stop searching */
2367 * active_load_balance is run by migration threads. It pushes running tasks
2368 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2369 * running on each physical CPU where possible, and avoids physical /
2370 * logical imbalances.
2372 * Called with busiest_rq locked.
2374 static void active_load_balance(runqueue_t
*busiest_rq
, int busiest_cpu
)
2376 struct sched_domain
*sd
;
2377 runqueue_t
*target_rq
;
2378 int target_cpu
= busiest_rq
->push_cpu
;
2380 if (busiest_rq
->nr_running
<= 1)
2381 /* no task to move */
2384 target_rq
= cpu_rq(target_cpu
);
2387 * This condition is "impossible", if it occurs
2388 * we need to fix it. Originally reported by
2389 * Bjorn Helgaas on a 128-cpu setup.
2391 BUG_ON(busiest_rq
== target_rq
);
2393 /* move a task from busiest_rq to target_rq */
2394 double_lock_balance(busiest_rq
, target_rq
);
2396 /* Search for an sd spanning us and the target CPU. */
2397 for_each_domain(target_cpu
, sd
)
2398 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2399 cpu_isset(busiest_cpu
, sd
->span
))
2402 if (unlikely(sd
== NULL
))
2405 schedstat_inc(sd
, alb_cnt
);
2407 if (move_tasks(target_rq
, target_cpu
, busiest_rq
, 1, sd
, SCHED_IDLE
, NULL
))
2408 schedstat_inc(sd
, alb_pushed
);
2410 schedstat_inc(sd
, alb_failed
);
2412 spin_unlock(&target_rq
->lock
);
2416 * rebalance_tick will get called every timer tick, on every CPU.
2418 * It checks each scheduling domain to see if it is due to be balanced,
2419 * and initiates a balancing operation if so.
2421 * Balancing parameters are set up in arch_init_sched_domains.
2424 /* Don't have all balancing operations going off at once */
2425 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2427 static void rebalance_tick(int this_cpu
, runqueue_t
*this_rq
,
2428 enum idle_type idle
)
2430 unsigned long old_load
, this_load
;
2431 unsigned long j
= jiffies
+ CPU_OFFSET(this_cpu
);
2432 struct sched_domain
*sd
;
2435 this_load
= this_rq
->nr_running
* SCHED_LOAD_SCALE
;
2436 /* Update our load */
2437 for (i
= 0; i
< 3; i
++) {
2438 unsigned long new_load
= this_load
;
2440 old_load
= this_rq
->cpu_load
[i
];
2442 * Round up the averaging division if load is increasing. This
2443 * prevents us from getting stuck on 9 if the load is 10, for
2446 if (new_load
> old_load
)
2447 new_load
+= scale
-1;
2448 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) / scale
;
2451 for_each_domain(this_cpu
, sd
) {
2452 unsigned long interval
;
2454 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2457 interval
= sd
->balance_interval
;
2458 if (idle
!= SCHED_IDLE
)
2459 interval
*= sd
->busy_factor
;
2461 /* scale ms to jiffies */
2462 interval
= msecs_to_jiffies(interval
);
2463 if (unlikely(!interval
))
2466 if (j
- sd
->last_balance
>= interval
) {
2467 if (load_balance(this_cpu
, this_rq
, sd
, idle
)) {
2469 * We've pulled tasks over so either we're no
2470 * longer idle, or one of our SMT siblings is
2475 sd
->last_balance
+= interval
;
2481 * on UP we do not need to balance between CPUs:
2483 static inline void rebalance_tick(int cpu
, runqueue_t
*rq
, enum idle_type idle
)
2486 static inline void idle_balance(int cpu
, runqueue_t
*rq
)
2491 static inline int wake_priority_sleeper(runqueue_t
*rq
)
2494 #ifdef CONFIG_SCHED_SMT
2495 spin_lock(&rq
->lock
);
2497 * If an SMT sibling task has been put to sleep for priority
2498 * reasons reschedule the idle task to see if it can now run.
2500 if (rq
->nr_running
) {
2501 resched_task(rq
->idle
);
2504 spin_unlock(&rq
->lock
);
2509 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2511 EXPORT_PER_CPU_SYMBOL(kstat
);
2514 * This is called on clock ticks and on context switches.
2515 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2517 static inline void update_cpu_clock(task_t
*p
, runqueue_t
*rq
,
2518 unsigned long long now
)
2520 unsigned long long last
= max(p
->timestamp
, rq
->timestamp_last_tick
);
2521 p
->sched_time
+= now
- last
;
2525 * Return current->sched_time plus any more ns on the sched_clock
2526 * that have not yet been banked.
2528 unsigned long long current_sched_time(const task_t
*tsk
)
2530 unsigned long long ns
;
2531 unsigned long flags
;
2532 local_irq_save(flags
);
2533 ns
= max(tsk
->timestamp
, task_rq(tsk
)->timestamp_last_tick
);
2534 ns
= tsk
->sched_time
+ (sched_clock() - ns
);
2535 local_irq_restore(flags
);
2540 * We place interactive tasks back into the active array, if possible.
2542 * To guarantee that this does not starve expired tasks we ignore the
2543 * interactivity of a task if the first expired task had to wait more
2544 * than a 'reasonable' amount of time. This deadline timeout is
2545 * load-dependent, as the frequency of array switched decreases with
2546 * increasing number of running tasks. We also ignore the interactivity
2547 * if a better static_prio task has expired:
2549 #define EXPIRED_STARVING(rq) \
2550 ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2551 (jiffies - (rq)->expired_timestamp >= \
2552 STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2553 ((rq)->curr->static_prio > (rq)->best_expired_prio))
2556 * Account user cpu time to a process.
2557 * @p: the process that the cpu time gets accounted to
2558 * @hardirq_offset: the offset to subtract from hardirq_count()
2559 * @cputime: the cpu time spent in user space since the last update
2561 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
2563 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2566 p
->utime
= cputime_add(p
->utime
, cputime
);
2568 /* Add user time to cpustat. */
2569 tmp
= cputime_to_cputime64(cputime
);
2570 if (TASK_NICE(p
) > 0)
2571 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
2573 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
2577 * Account system cpu time to a process.
2578 * @p: the process that the cpu time gets accounted to
2579 * @hardirq_offset: the offset to subtract from hardirq_count()
2580 * @cputime: the cpu time spent in kernel space since the last update
2582 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
2585 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2586 runqueue_t
*rq
= this_rq();
2589 p
->stime
= cputime_add(p
->stime
, cputime
);
2591 /* Add system time to cpustat. */
2592 tmp
= cputime_to_cputime64(cputime
);
2593 if (hardirq_count() - hardirq_offset
)
2594 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
2595 else if (softirq_count())
2596 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
2597 else if (p
!= rq
->idle
)
2598 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
2599 else if (atomic_read(&rq
->nr_iowait
) > 0)
2600 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
2602 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
2603 /* Account for system time used */
2604 acct_update_integrals(p
);
2608 * Account for involuntary wait time.
2609 * @p: the process from which the cpu time has been stolen
2610 * @steal: the cpu time spent in involuntary wait
2612 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
2614 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2615 cputime64_t tmp
= cputime_to_cputime64(steal
);
2616 runqueue_t
*rq
= this_rq();
2618 if (p
== rq
->idle
) {
2619 p
->stime
= cputime_add(p
->stime
, steal
);
2620 if (atomic_read(&rq
->nr_iowait
) > 0)
2621 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
2623 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
2625 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
2629 * This function gets called by the timer code, with HZ frequency.
2630 * We call it with interrupts disabled.
2632 * It also gets called by the fork code, when changing the parent's
2635 void scheduler_tick(void)
2637 int cpu
= smp_processor_id();
2638 runqueue_t
*rq
= this_rq();
2639 task_t
*p
= current
;
2640 unsigned long long now
= sched_clock();
2642 update_cpu_clock(p
, rq
, now
);
2644 rq
->timestamp_last_tick
= now
;
2646 if (p
== rq
->idle
) {
2647 if (wake_priority_sleeper(rq
))
2649 rebalance_tick(cpu
, rq
, SCHED_IDLE
);
2653 /* Task might have expired already, but not scheduled off yet */
2654 if (p
->array
!= rq
->active
) {
2655 set_tsk_need_resched(p
);
2658 spin_lock(&rq
->lock
);
2660 * The task was running during this tick - update the
2661 * time slice counter. Note: we do not update a thread's
2662 * priority until it either goes to sleep or uses up its
2663 * timeslice. This makes it possible for interactive tasks
2664 * to use up their timeslices at their highest priority levels.
2668 * RR tasks need a special form of timeslice management.
2669 * FIFO tasks have no timeslices.
2671 if ((p
->policy
== SCHED_RR
) && !--p
->time_slice
) {
2672 p
->time_slice
= task_timeslice(p
);
2673 p
->first_time_slice
= 0;
2674 set_tsk_need_resched(p
);
2676 /* put it at the end of the queue: */
2677 requeue_task(p
, rq
->active
);
2681 if (!--p
->time_slice
) {
2682 dequeue_task(p
, rq
->active
);
2683 set_tsk_need_resched(p
);
2684 p
->prio
= effective_prio(p
);
2685 p
->time_slice
= task_timeslice(p
);
2686 p
->first_time_slice
= 0;
2688 if (!rq
->expired_timestamp
)
2689 rq
->expired_timestamp
= jiffies
;
2690 if (!TASK_INTERACTIVE(p
) || EXPIRED_STARVING(rq
)) {
2691 enqueue_task(p
, rq
->expired
);
2692 if (p
->static_prio
< rq
->best_expired_prio
)
2693 rq
->best_expired_prio
= p
->static_prio
;
2695 enqueue_task(p
, rq
->active
);
2698 * Prevent a too long timeslice allowing a task to monopolize
2699 * the CPU. We do this by splitting up the timeslice into
2702 * Note: this does not mean the task's timeslices expire or
2703 * get lost in any way, they just might be preempted by
2704 * another task of equal priority. (one with higher
2705 * priority would have preempted this task already.) We
2706 * requeue this task to the end of the list on this priority
2707 * level, which is in essence a round-robin of tasks with
2710 * This only applies to tasks in the interactive
2711 * delta range with at least TIMESLICE_GRANULARITY to requeue.
2713 if (TASK_INTERACTIVE(p
) && !((task_timeslice(p
) -
2714 p
->time_slice
) % TIMESLICE_GRANULARITY(p
)) &&
2715 (p
->time_slice
>= TIMESLICE_GRANULARITY(p
)) &&
2716 (p
->array
== rq
->active
)) {
2718 requeue_task(p
, rq
->active
);
2719 set_tsk_need_resched(p
);
2723 spin_unlock(&rq
->lock
);
2725 rebalance_tick(cpu
, rq
, NOT_IDLE
);
2728 #ifdef CONFIG_SCHED_SMT
2729 static inline void wakeup_busy_runqueue(runqueue_t
*rq
)
2731 /* If an SMT runqueue is sleeping due to priority reasons wake it up */
2732 if (rq
->curr
== rq
->idle
&& rq
->nr_running
)
2733 resched_task(rq
->idle
);
2736 static inline void wake_sleeping_dependent(int this_cpu
, runqueue_t
*this_rq
)
2738 struct sched_domain
*tmp
, *sd
= NULL
;
2739 cpumask_t sibling_map
;
2742 for_each_domain(this_cpu
, tmp
)
2743 if (tmp
->flags
& SD_SHARE_CPUPOWER
)
2750 * Unlock the current runqueue because we have to lock in
2751 * CPU order to avoid deadlocks. Caller knows that we might
2752 * unlock. We keep IRQs disabled.
2754 spin_unlock(&this_rq
->lock
);
2756 sibling_map
= sd
->span
;
2758 for_each_cpu_mask(i
, sibling_map
)
2759 spin_lock(&cpu_rq(i
)->lock
);
2761 * We clear this CPU from the mask. This both simplifies the
2762 * inner loop and keps this_rq locked when we exit:
2764 cpu_clear(this_cpu
, sibling_map
);
2766 for_each_cpu_mask(i
, sibling_map
) {
2767 runqueue_t
*smt_rq
= cpu_rq(i
);
2769 wakeup_busy_runqueue(smt_rq
);
2772 for_each_cpu_mask(i
, sibling_map
)
2773 spin_unlock(&cpu_rq(i
)->lock
);
2775 * We exit with this_cpu's rq still held and IRQs
2781 * number of 'lost' timeslices this task wont be able to fully
2782 * utilize, if another task runs on a sibling. This models the
2783 * slowdown effect of other tasks running on siblings:
2785 static inline unsigned long smt_slice(task_t
*p
, struct sched_domain
*sd
)
2787 return p
->time_slice
* (100 - sd
->per_cpu_gain
) / 100;
2790 static inline int dependent_sleeper(int this_cpu
, runqueue_t
*this_rq
)
2792 struct sched_domain
*tmp
, *sd
= NULL
;
2793 cpumask_t sibling_map
;
2794 prio_array_t
*array
;
2798 for_each_domain(this_cpu
, tmp
)
2799 if (tmp
->flags
& SD_SHARE_CPUPOWER
)
2806 * The same locking rules and details apply as for
2807 * wake_sleeping_dependent():
2809 spin_unlock(&this_rq
->lock
);
2810 sibling_map
= sd
->span
;
2811 for_each_cpu_mask(i
, sibling_map
)
2812 spin_lock(&cpu_rq(i
)->lock
);
2813 cpu_clear(this_cpu
, sibling_map
);
2816 * Establish next task to be run - it might have gone away because
2817 * we released the runqueue lock above:
2819 if (!this_rq
->nr_running
)
2821 array
= this_rq
->active
;
2822 if (!array
->nr_active
)
2823 array
= this_rq
->expired
;
2824 BUG_ON(!array
->nr_active
);
2826 p
= list_entry(array
->queue
[sched_find_first_bit(array
->bitmap
)].next
,
2829 for_each_cpu_mask(i
, sibling_map
) {
2830 runqueue_t
*smt_rq
= cpu_rq(i
);
2831 task_t
*smt_curr
= smt_rq
->curr
;
2833 /* Kernel threads do not participate in dependent sleeping */
2834 if (!p
->mm
|| !smt_curr
->mm
|| rt_task(p
))
2835 goto check_smt_task
;
2838 * If a user task with lower static priority than the
2839 * running task on the SMT sibling is trying to schedule,
2840 * delay it till there is proportionately less timeslice
2841 * left of the sibling task to prevent a lower priority
2842 * task from using an unfair proportion of the
2843 * physical cpu's resources. -ck
2845 if (rt_task(smt_curr
)) {
2847 * With real time tasks we run non-rt tasks only
2848 * per_cpu_gain% of the time.
2850 if ((jiffies
% DEF_TIMESLICE
) >
2851 (sd
->per_cpu_gain
* DEF_TIMESLICE
/ 100))
2854 if (smt_curr
->static_prio
< p
->static_prio
&&
2855 !TASK_PREEMPTS_CURR(p
, smt_rq
) &&
2856 smt_slice(smt_curr
, sd
) > task_timeslice(p
))
2860 if ((!smt_curr
->mm
&& smt_curr
!= smt_rq
->idle
) ||
2864 wakeup_busy_runqueue(smt_rq
);
2869 * Reschedule a lower priority task on the SMT sibling for
2870 * it to be put to sleep, or wake it up if it has been put to
2871 * sleep for priority reasons to see if it should run now.
2874 if ((jiffies
% DEF_TIMESLICE
) >
2875 (sd
->per_cpu_gain
* DEF_TIMESLICE
/ 100))
2876 resched_task(smt_curr
);
2878 if (TASK_PREEMPTS_CURR(p
, smt_rq
) &&
2879 smt_slice(p
, sd
) > task_timeslice(smt_curr
))
2880 resched_task(smt_curr
);
2882 wakeup_busy_runqueue(smt_rq
);
2886 for_each_cpu_mask(i
, sibling_map
)
2887 spin_unlock(&cpu_rq(i
)->lock
);
2891 static inline void wake_sleeping_dependent(int this_cpu
, runqueue_t
*this_rq
)
2895 static inline int dependent_sleeper(int this_cpu
, runqueue_t
*this_rq
)
2901 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
2903 void fastcall
add_preempt_count(int val
)
2908 BUG_ON((preempt_count() < 0));
2909 preempt_count() += val
;
2911 * Spinlock count overflowing soon?
2913 BUG_ON((preempt_count() & PREEMPT_MASK
) >= PREEMPT_MASK
-10);
2915 EXPORT_SYMBOL(add_preempt_count
);
2917 void fastcall
sub_preempt_count(int val
)
2922 BUG_ON(val
> preempt_count());
2924 * Is the spinlock portion underflowing?
2926 BUG_ON((val
< PREEMPT_MASK
) && !(preempt_count() & PREEMPT_MASK
));
2927 preempt_count() -= val
;
2929 EXPORT_SYMBOL(sub_preempt_count
);
2934 * schedule() is the main scheduler function.
2936 asmlinkage
void __sched
schedule(void)
2939 task_t
*prev
, *next
;
2941 prio_array_t
*array
;
2942 struct list_head
*queue
;
2943 unsigned long long now
;
2944 unsigned long run_time
;
2945 int cpu
, idx
, new_prio
;
2948 * Test if we are atomic. Since do_exit() needs to call into
2949 * schedule() atomically, we ignore that path for now.
2950 * Otherwise, whine if we are scheduling when we should not be.
2952 if (likely(!current
->exit_state
)) {
2953 if (unlikely(in_atomic())) {
2954 printk(KERN_ERR
"scheduling while atomic: "
2956 current
->comm
, preempt_count(), current
->pid
);
2960 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2965 release_kernel_lock(prev
);
2966 need_resched_nonpreemptible
:
2970 * The idle thread is not allowed to schedule!
2971 * Remove this check after it has been exercised a bit.
2973 if (unlikely(prev
== rq
->idle
) && prev
->state
!= TASK_RUNNING
) {
2974 printk(KERN_ERR
"bad: scheduling from the idle thread!\n");
2978 schedstat_inc(rq
, sched_cnt
);
2979 now
= sched_clock();
2980 if (likely((long long)(now
- prev
->timestamp
) < NS_MAX_SLEEP_AVG
)) {
2981 run_time
= now
- prev
->timestamp
;
2982 if (unlikely((long long)(now
- prev
->timestamp
) < 0))
2985 run_time
= NS_MAX_SLEEP_AVG
;
2988 * Tasks charged proportionately less run_time at high sleep_avg to
2989 * delay them losing their interactive status
2991 run_time
/= (CURRENT_BONUS(prev
) ? : 1);
2993 spin_lock_irq(&rq
->lock
);
2995 if (unlikely(prev
->flags
& PF_DEAD
))
2996 prev
->state
= EXIT_DEAD
;
2998 switch_count
= &prev
->nivcsw
;
2999 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3000 switch_count
= &prev
->nvcsw
;
3001 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3002 unlikely(signal_pending(prev
))))
3003 prev
->state
= TASK_RUNNING
;
3005 if (prev
->state
== TASK_UNINTERRUPTIBLE
)
3006 rq
->nr_uninterruptible
++;
3007 deactivate_task(prev
, rq
);
3011 cpu
= smp_processor_id();
3012 if (unlikely(!rq
->nr_running
)) {
3014 idle_balance(cpu
, rq
);
3015 if (!rq
->nr_running
) {
3017 rq
->expired_timestamp
= 0;
3018 wake_sleeping_dependent(cpu
, rq
);
3020 * wake_sleeping_dependent() might have released
3021 * the runqueue, so break out if we got new
3024 if (!rq
->nr_running
)
3028 if (dependent_sleeper(cpu
, rq
)) {
3033 * dependent_sleeper() releases and reacquires the runqueue
3034 * lock, hence go into the idle loop if the rq went
3037 if (unlikely(!rq
->nr_running
))
3042 if (unlikely(!array
->nr_active
)) {
3044 * Switch the active and expired arrays.
3046 schedstat_inc(rq
, sched_switch
);
3047 rq
->active
= rq
->expired
;
3048 rq
->expired
= array
;
3050 rq
->expired_timestamp
= 0;
3051 rq
->best_expired_prio
= MAX_PRIO
;
3054 idx
= sched_find_first_bit(array
->bitmap
);
3055 queue
= array
->queue
+ idx
;
3056 next
= list_entry(queue
->next
, task_t
, run_list
);
3058 if (!rt_task(next
) && next
->activated
> 0) {
3059 unsigned long long delta
= now
- next
->timestamp
;
3060 if (unlikely((long long)(now
- next
->timestamp
) < 0))
3063 if (next
->activated
== 1)
3064 delta
= delta
* (ON_RUNQUEUE_WEIGHT
* 128 / 100) / 128;
3066 array
= next
->array
;
3067 new_prio
= recalc_task_prio(next
, next
->timestamp
+ delta
);
3069 if (unlikely(next
->prio
!= new_prio
)) {
3070 dequeue_task(next
, array
);
3071 next
->prio
= new_prio
;
3072 enqueue_task(next
, array
);
3074 requeue_task(next
, array
);
3076 next
->activated
= 0;
3078 if (next
== rq
->idle
)
3079 schedstat_inc(rq
, sched_goidle
);
3081 prefetch_stack(next
);
3082 clear_tsk_need_resched(prev
);
3083 rcu_qsctr_inc(task_cpu(prev
));
3085 update_cpu_clock(prev
, rq
, now
);
3087 prev
->sleep_avg
-= run_time
;
3088 if ((long)prev
->sleep_avg
<= 0)
3089 prev
->sleep_avg
= 0;
3090 prev
->timestamp
= prev
->last_ran
= now
;
3092 sched_info_switch(prev
, next
);
3093 if (likely(prev
!= next
)) {
3094 next
->timestamp
= now
;
3099 prepare_task_switch(rq
, next
);
3100 prev
= context_switch(rq
, prev
, next
);
3103 * this_rq must be evaluated again because prev may have moved
3104 * CPUs since it called schedule(), thus the 'rq' on its stack
3105 * frame will be invalid.
3107 finish_task_switch(this_rq(), prev
);
3109 spin_unlock_irq(&rq
->lock
);
3112 if (unlikely(reacquire_kernel_lock(prev
) < 0))
3113 goto need_resched_nonpreemptible
;
3114 preempt_enable_no_resched();
3115 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3119 EXPORT_SYMBOL(schedule
);
3121 #ifdef CONFIG_PREEMPT
3123 * this is is the entry point to schedule() from in-kernel preemption
3124 * off of preempt_enable. Kernel preemptions off return from interrupt
3125 * occur there and call schedule directly.
3127 asmlinkage
void __sched
preempt_schedule(void)
3129 struct thread_info
*ti
= current_thread_info();
3130 #ifdef CONFIG_PREEMPT_BKL
3131 struct task_struct
*task
= current
;
3132 int saved_lock_depth
;
3135 * If there is a non-zero preempt_count or interrupts are disabled,
3136 * we do not want to preempt the current task. Just return..
3138 if (unlikely(ti
->preempt_count
|| irqs_disabled()))
3142 add_preempt_count(PREEMPT_ACTIVE
);
3144 * We keep the big kernel semaphore locked, but we
3145 * clear ->lock_depth so that schedule() doesnt
3146 * auto-release the semaphore:
3148 #ifdef CONFIG_PREEMPT_BKL
3149 saved_lock_depth
= task
->lock_depth
;
3150 task
->lock_depth
= -1;
3153 #ifdef CONFIG_PREEMPT_BKL
3154 task
->lock_depth
= saved_lock_depth
;
3156 sub_preempt_count(PREEMPT_ACTIVE
);
3158 /* we could miss a preemption opportunity between schedule and now */
3160 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3164 EXPORT_SYMBOL(preempt_schedule
);
3167 * this is is the entry point to schedule() from kernel preemption
3168 * off of irq context.
3169 * Note, that this is called and return with irqs disabled. This will
3170 * protect us against recursive calling from irq.
3172 asmlinkage
void __sched
preempt_schedule_irq(void)
3174 struct thread_info
*ti
= current_thread_info();
3175 #ifdef CONFIG_PREEMPT_BKL
3176 struct task_struct
*task
= current
;
3177 int saved_lock_depth
;
3179 /* Catch callers which need to be fixed*/
3180 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3183 add_preempt_count(PREEMPT_ACTIVE
);
3185 * We keep the big kernel semaphore locked, but we
3186 * clear ->lock_depth so that schedule() doesnt
3187 * auto-release the semaphore:
3189 #ifdef CONFIG_PREEMPT_BKL
3190 saved_lock_depth
= task
->lock_depth
;
3191 task
->lock_depth
= -1;
3195 local_irq_disable();
3196 #ifdef CONFIG_PREEMPT_BKL
3197 task
->lock_depth
= saved_lock_depth
;
3199 sub_preempt_count(PREEMPT_ACTIVE
);
3201 /* we could miss a preemption opportunity between schedule and now */
3203 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3207 #endif /* CONFIG_PREEMPT */
3209 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3212 task_t
*p
= curr
->private;
3213 return try_to_wake_up(p
, mode
, sync
);
3216 EXPORT_SYMBOL(default_wake_function
);
3219 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3220 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3221 * number) then we wake all the non-exclusive tasks and one exclusive task.
3223 * There are circumstances in which we can try to wake a task which has already
3224 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3225 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3227 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3228 int nr_exclusive
, int sync
, void *key
)
3230 struct list_head
*tmp
, *next
;
3232 list_for_each_safe(tmp
, next
, &q
->task_list
) {
3235 curr
= list_entry(tmp
, wait_queue_t
, task_list
);
3236 flags
= curr
->flags
;
3237 if (curr
->func(curr
, mode
, sync
, key
) &&
3238 (flags
& WQ_FLAG_EXCLUSIVE
) &&
3245 * __wake_up - wake up threads blocked on a waitqueue.
3247 * @mode: which threads
3248 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3249 * @key: is directly passed to the wakeup function
3251 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3252 int nr_exclusive
, void *key
)
3254 unsigned long flags
;
3256 spin_lock_irqsave(&q
->lock
, flags
);
3257 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3258 spin_unlock_irqrestore(&q
->lock
, flags
);
3261 EXPORT_SYMBOL(__wake_up
);
3264 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3266 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3268 __wake_up_common(q
, mode
, 1, 0, NULL
);
3272 * __wake_up_sync - wake up threads blocked on a waitqueue.
3274 * @mode: which threads
3275 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3277 * The sync wakeup differs that the waker knows that it will schedule
3278 * away soon, so while the target thread will be woken up, it will not
3279 * be migrated to another CPU - ie. the two threads are 'synchronized'
3280 * with each other. This can prevent needless bouncing between CPUs.
3282 * On UP it can prevent extra preemption.
3285 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3287 unsigned long flags
;
3293 if (unlikely(!nr_exclusive
))
3296 spin_lock_irqsave(&q
->lock
, flags
);
3297 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3298 spin_unlock_irqrestore(&q
->lock
, flags
);
3300 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3302 void fastcall
complete(struct completion
*x
)
3304 unsigned long flags
;
3306 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3308 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3310 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3312 EXPORT_SYMBOL(complete
);
3314 void fastcall
complete_all(struct completion
*x
)
3316 unsigned long flags
;
3318 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3319 x
->done
+= UINT_MAX
/2;
3320 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3322 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3324 EXPORT_SYMBOL(complete_all
);
3326 void fastcall __sched
wait_for_completion(struct completion
*x
)
3329 spin_lock_irq(&x
->wait
.lock
);
3331 DECLARE_WAITQUEUE(wait
, current
);
3333 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3334 __add_wait_queue_tail(&x
->wait
, &wait
);
3336 __set_current_state(TASK_UNINTERRUPTIBLE
);
3337 spin_unlock_irq(&x
->wait
.lock
);
3339 spin_lock_irq(&x
->wait
.lock
);
3341 __remove_wait_queue(&x
->wait
, &wait
);
3344 spin_unlock_irq(&x
->wait
.lock
);
3346 EXPORT_SYMBOL(wait_for_completion
);
3348 unsigned long fastcall __sched
3349 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3353 spin_lock_irq(&x
->wait
.lock
);
3355 DECLARE_WAITQUEUE(wait
, current
);
3357 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3358 __add_wait_queue_tail(&x
->wait
, &wait
);
3360 __set_current_state(TASK_UNINTERRUPTIBLE
);
3361 spin_unlock_irq(&x
->wait
.lock
);
3362 timeout
= schedule_timeout(timeout
);
3363 spin_lock_irq(&x
->wait
.lock
);
3365 __remove_wait_queue(&x
->wait
, &wait
);
3369 __remove_wait_queue(&x
->wait
, &wait
);
3373 spin_unlock_irq(&x
->wait
.lock
);
3376 EXPORT_SYMBOL(wait_for_completion_timeout
);
3378 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3384 spin_lock_irq(&x
->wait
.lock
);
3386 DECLARE_WAITQUEUE(wait
, current
);
3388 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3389 __add_wait_queue_tail(&x
->wait
, &wait
);
3391 if (signal_pending(current
)) {
3393 __remove_wait_queue(&x
->wait
, &wait
);
3396 __set_current_state(TASK_INTERRUPTIBLE
);
3397 spin_unlock_irq(&x
->wait
.lock
);
3399 spin_lock_irq(&x
->wait
.lock
);
3401 __remove_wait_queue(&x
->wait
, &wait
);
3405 spin_unlock_irq(&x
->wait
.lock
);
3409 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3411 unsigned long fastcall __sched
3412 wait_for_completion_interruptible_timeout(struct completion
*x
,
3413 unsigned long timeout
)
3417 spin_lock_irq(&x
->wait
.lock
);
3419 DECLARE_WAITQUEUE(wait
, current
);
3421 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3422 __add_wait_queue_tail(&x
->wait
, &wait
);
3424 if (signal_pending(current
)) {
3425 timeout
= -ERESTARTSYS
;
3426 __remove_wait_queue(&x
->wait
, &wait
);
3429 __set_current_state(TASK_INTERRUPTIBLE
);
3430 spin_unlock_irq(&x
->wait
.lock
);
3431 timeout
= schedule_timeout(timeout
);
3432 spin_lock_irq(&x
->wait
.lock
);
3434 __remove_wait_queue(&x
->wait
, &wait
);
3438 __remove_wait_queue(&x
->wait
, &wait
);
3442 spin_unlock_irq(&x
->wait
.lock
);
3445 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3448 #define SLEEP_ON_VAR \
3449 unsigned long flags; \
3450 wait_queue_t wait; \
3451 init_waitqueue_entry(&wait, current);
3453 #define SLEEP_ON_HEAD \
3454 spin_lock_irqsave(&q->lock,flags); \
3455 __add_wait_queue(q, &wait); \
3456 spin_unlock(&q->lock);
3458 #define SLEEP_ON_TAIL \
3459 spin_lock_irq(&q->lock); \
3460 __remove_wait_queue(q, &wait); \
3461 spin_unlock_irqrestore(&q->lock, flags);
3463 void fastcall __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3467 current
->state
= TASK_INTERRUPTIBLE
;
3474 EXPORT_SYMBOL(interruptible_sleep_on
);
3476 long fastcall __sched
3477 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3481 current
->state
= TASK_INTERRUPTIBLE
;
3484 timeout
= schedule_timeout(timeout
);
3490 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3492 void fastcall __sched
sleep_on(wait_queue_head_t
*q
)
3496 current
->state
= TASK_UNINTERRUPTIBLE
;
3503 EXPORT_SYMBOL(sleep_on
);
3505 long fastcall __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3509 current
->state
= TASK_UNINTERRUPTIBLE
;
3512 timeout
= schedule_timeout(timeout
);
3518 EXPORT_SYMBOL(sleep_on_timeout
);
3520 void set_user_nice(task_t
*p
, long nice
)
3522 unsigned long flags
;
3523 prio_array_t
*array
;
3525 int old_prio
, new_prio
, delta
;
3527 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3530 * We have to be careful, if called from sys_setpriority(),
3531 * the task might be in the middle of scheduling on another CPU.
3533 rq
= task_rq_lock(p
, &flags
);
3535 * The RT priorities are set via sched_setscheduler(), but we still
3536 * allow the 'normal' nice value to be set - but as expected
3537 * it wont have any effect on scheduling until the task is
3541 p
->static_prio
= NICE_TO_PRIO(nice
);
3546 dequeue_task(p
, array
);
3547 dec_prio_bias(rq
, p
->static_prio
);
3551 new_prio
= NICE_TO_PRIO(nice
);
3552 delta
= new_prio
- old_prio
;
3553 p
->static_prio
= NICE_TO_PRIO(nice
);
3557 enqueue_task(p
, array
);
3558 inc_prio_bias(rq
, p
->static_prio
);
3560 * If the task increased its priority or is running and
3561 * lowered its priority, then reschedule its CPU:
3563 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3564 resched_task(rq
->curr
);
3567 task_rq_unlock(rq
, &flags
);
3570 EXPORT_SYMBOL(set_user_nice
);
3573 * can_nice - check if a task can reduce its nice value
3577 int can_nice(const task_t
*p
, const int nice
)
3579 /* convert nice value [19,-20] to rlimit style value [1,40] */
3580 int nice_rlim
= 20 - nice
;
3581 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
3582 capable(CAP_SYS_NICE
));
3585 #ifdef __ARCH_WANT_SYS_NICE
3588 * sys_nice - change the priority of the current process.
3589 * @increment: priority increment
3591 * sys_setpriority is a more generic, but much slower function that
3592 * does similar things.
3594 asmlinkage
long sys_nice(int increment
)
3600 * Setpriority might change our priority at the same moment.
3601 * We don't have to worry. Conceptually one call occurs first
3602 * and we have a single winner.
3604 if (increment
< -40)
3609 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
3615 if (increment
< 0 && !can_nice(current
, nice
))
3618 retval
= security_task_setnice(current
, nice
);
3622 set_user_nice(current
, nice
);
3629 * task_prio - return the priority value of a given task.
3630 * @p: the task in question.
3632 * This is the priority value as seen by users in /proc.
3633 * RT tasks are offset by -200. Normal tasks are centered
3634 * around 0, value goes from -16 to +15.
3636 int task_prio(const task_t
*p
)
3638 return p
->prio
- MAX_RT_PRIO
;
3642 * task_nice - return the nice value of a given task.
3643 * @p: the task in question.
3645 int task_nice(const task_t
*p
)
3647 return TASK_NICE(p
);
3649 EXPORT_SYMBOL_GPL(task_nice
);
3652 * idle_cpu - is a given cpu idle currently?
3653 * @cpu: the processor in question.
3655 int idle_cpu(int cpu
)
3657 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
3661 * idle_task - return the idle task for a given cpu.
3662 * @cpu: the processor in question.
3664 task_t
*idle_task(int cpu
)
3666 return cpu_rq(cpu
)->idle
;
3670 * find_process_by_pid - find a process with a matching PID value.
3671 * @pid: the pid in question.
3673 static inline task_t
*find_process_by_pid(pid_t pid
)
3675 return pid
? find_task_by_pid(pid
) : current
;
3678 /* Actually do priority change: must hold rq lock. */
3679 static void __setscheduler(struct task_struct
*p
, int policy
, int prio
)
3683 p
->rt_priority
= prio
;
3684 if (policy
!= SCHED_NORMAL
)
3685 p
->prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
3687 p
->prio
= p
->static_prio
;
3691 * sched_setscheduler - change the scheduling policy and/or RT priority of
3693 * @p: the task in question.
3694 * @policy: new policy.
3695 * @param: structure containing the new RT priority.
3697 int sched_setscheduler(struct task_struct
*p
, int policy
,
3698 struct sched_param
*param
)
3701 int oldprio
, oldpolicy
= -1;
3702 prio_array_t
*array
;
3703 unsigned long flags
;
3707 /* double check policy once rq lock held */
3709 policy
= oldpolicy
= p
->policy
;
3710 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3711 policy
!= SCHED_NORMAL
)
3714 * Valid priorities for SCHED_FIFO and SCHED_RR are
3715 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL is 0.
3717 if (param
->sched_priority
< 0 ||
3718 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3719 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
3721 if ((policy
== SCHED_NORMAL
) != (param
->sched_priority
== 0))
3725 * Allow unprivileged RT tasks to decrease priority:
3727 if (!capable(CAP_SYS_NICE
)) {
3728 /* can't change policy */
3729 if (policy
!= p
->policy
&&
3730 !p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
)
3732 /* can't increase priority */
3733 if (policy
!= SCHED_NORMAL
&&
3734 param
->sched_priority
> p
->rt_priority
&&
3735 param
->sched_priority
>
3736 p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
)
3738 /* can't change other user's priorities */
3739 if ((current
->euid
!= p
->euid
) &&
3740 (current
->euid
!= p
->uid
))
3744 retval
= security_task_setscheduler(p
, policy
, param
);
3748 * To be able to change p->policy safely, the apropriate
3749 * runqueue lock must be held.
3751 rq
= task_rq_lock(p
, &flags
);
3752 /* recheck policy now with rq lock held */
3753 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3754 policy
= oldpolicy
= -1;
3755 task_rq_unlock(rq
, &flags
);
3760 deactivate_task(p
, rq
);
3762 __setscheduler(p
, policy
, param
->sched_priority
);
3764 __activate_task(p
, rq
);
3766 * Reschedule if we are currently running on this runqueue and
3767 * our priority decreased, or if we are not currently running on
3768 * this runqueue and our priority is higher than the current's
3770 if (task_running(rq
, p
)) {
3771 if (p
->prio
> oldprio
)
3772 resched_task(rq
->curr
);
3773 } else if (TASK_PREEMPTS_CURR(p
, rq
))
3774 resched_task(rq
->curr
);
3776 task_rq_unlock(rq
, &flags
);
3779 EXPORT_SYMBOL_GPL(sched_setscheduler
);
3782 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
3785 struct sched_param lparam
;
3786 struct task_struct
*p
;
3788 if (!param
|| pid
< 0)
3790 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
3792 read_lock_irq(&tasklist_lock
);
3793 p
= find_process_by_pid(pid
);
3795 read_unlock_irq(&tasklist_lock
);
3798 retval
= sched_setscheduler(p
, policy
, &lparam
);
3799 read_unlock_irq(&tasklist_lock
);
3804 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3805 * @pid: the pid in question.
3806 * @policy: new policy.
3807 * @param: structure containing the new RT priority.
3809 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
3810 struct sched_param __user
*param
)
3812 return do_sched_setscheduler(pid
, policy
, param
);
3816 * sys_sched_setparam - set/change the RT priority of a thread
3817 * @pid: the pid in question.
3818 * @param: structure containing the new RT priority.
3820 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
3822 return do_sched_setscheduler(pid
, -1, param
);
3826 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3827 * @pid: the pid in question.
3829 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
3831 int retval
= -EINVAL
;
3838 read_lock(&tasklist_lock
);
3839 p
= find_process_by_pid(pid
);
3841 retval
= security_task_getscheduler(p
);
3845 read_unlock(&tasklist_lock
);
3852 * sys_sched_getscheduler - get the RT priority of a thread
3853 * @pid: the pid in question.
3854 * @param: structure containing the RT priority.
3856 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
3858 struct sched_param lp
;
3859 int retval
= -EINVAL
;
3862 if (!param
|| pid
< 0)
3865 read_lock(&tasklist_lock
);
3866 p
= find_process_by_pid(pid
);
3871 retval
= security_task_getscheduler(p
);
3875 lp
.sched_priority
= p
->rt_priority
;
3876 read_unlock(&tasklist_lock
);
3879 * This one might sleep, we cannot do it with a spinlock held ...
3881 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
3887 read_unlock(&tasklist_lock
);
3891 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
3895 cpumask_t cpus_allowed
;
3898 read_lock(&tasklist_lock
);
3900 p
= find_process_by_pid(pid
);
3902 read_unlock(&tasklist_lock
);
3903 unlock_cpu_hotplug();
3908 * It is not safe to call set_cpus_allowed with the
3909 * tasklist_lock held. We will bump the task_struct's
3910 * usage count and then drop tasklist_lock.
3913 read_unlock(&tasklist_lock
);
3916 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
3917 !capable(CAP_SYS_NICE
))
3920 cpus_allowed
= cpuset_cpus_allowed(p
);
3921 cpus_and(new_mask
, new_mask
, cpus_allowed
);
3922 retval
= set_cpus_allowed(p
, new_mask
);
3926 unlock_cpu_hotplug();
3930 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
3931 cpumask_t
*new_mask
)
3933 if (len
< sizeof(cpumask_t
)) {
3934 memset(new_mask
, 0, sizeof(cpumask_t
));
3935 } else if (len
> sizeof(cpumask_t
)) {
3936 len
= sizeof(cpumask_t
);
3938 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
3942 * sys_sched_setaffinity - set the cpu affinity of a process
3943 * @pid: pid of the process
3944 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3945 * @user_mask_ptr: user-space pointer to the new cpu mask
3947 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
3948 unsigned long __user
*user_mask_ptr
)
3953 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
3957 return sched_setaffinity(pid
, new_mask
);
3961 * Represents all cpu's present in the system
3962 * In systems capable of hotplug, this map could dynamically grow
3963 * as new cpu's are detected in the system via any platform specific
3964 * method, such as ACPI for e.g.
3967 cpumask_t cpu_present_map
;
3968 EXPORT_SYMBOL(cpu_present_map
);
3971 cpumask_t cpu_online_map
= CPU_MASK_ALL
;
3972 cpumask_t cpu_possible_map
= CPU_MASK_ALL
;
3975 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
3981 read_lock(&tasklist_lock
);
3984 p
= find_process_by_pid(pid
);
3989 cpus_and(*mask
, p
->cpus_allowed
, cpu_possible_map
);
3992 read_unlock(&tasklist_lock
);
3993 unlock_cpu_hotplug();
4001 * sys_sched_getaffinity - get the cpu affinity of a process
4002 * @pid: pid of the process
4003 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4004 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4006 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4007 unsigned long __user
*user_mask_ptr
)
4012 if (len
< sizeof(cpumask_t
))
4015 ret
= sched_getaffinity(pid
, &mask
);
4019 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4022 return sizeof(cpumask_t
);
4026 * sys_sched_yield - yield the current processor to other threads.
4028 * this function yields the current CPU by moving the calling thread
4029 * to the expired array. If there are no other threads running on this
4030 * CPU then this function will return.
4032 asmlinkage
long sys_sched_yield(void)
4034 runqueue_t
*rq
= this_rq_lock();
4035 prio_array_t
*array
= current
->array
;
4036 prio_array_t
*target
= rq
->expired
;
4038 schedstat_inc(rq
, yld_cnt
);
4040 * We implement yielding by moving the task into the expired
4043 * (special rule: RT tasks will just roundrobin in the active
4046 if (rt_task(current
))
4047 target
= rq
->active
;
4049 if (array
->nr_active
== 1) {
4050 schedstat_inc(rq
, yld_act_empty
);
4051 if (!rq
->expired
->nr_active
)
4052 schedstat_inc(rq
, yld_both_empty
);
4053 } else if (!rq
->expired
->nr_active
)
4054 schedstat_inc(rq
, yld_exp_empty
);
4056 if (array
!= target
) {
4057 dequeue_task(current
, array
);
4058 enqueue_task(current
, target
);
4061 * requeue_task is cheaper so perform that if possible.
4063 requeue_task(current
, array
);
4066 * Since we are going to call schedule() anyway, there's
4067 * no need to preempt or enable interrupts:
4069 __release(rq
->lock
);
4070 _raw_spin_unlock(&rq
->lock
);
4071 preempt_enable_no_resched();
4078 static inline void __cond_resched(void)
4081 * The BKS might be reacquired before we have dropped
4082 * PREEMPT_ACTIVE, which could trigger a second
4083 * cond_resched() call.
4085 if (unlikely(preempt_count()))
4088 add_preempt_count(PREEMPT_ACTIVE
);
4090 sub_preempt_count(PREEMPT_ACTIVE
);
4091 } while (need_resched());
4094 int __sched
cond_resched(void)
4096 if (need_resched()) {
4103 EXPORT_SYMBOL(cond_resched
);
4106 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4107 * call schedule, and on return reacquire the lock.
4109 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4110 * operations here to prevent schedule() from being called twice (once via
4111 * spin_unlock(), once by hand).
4113 int cond_resched_lock(spinlock_t
*lock
)
4117 if (need_lockbreak(lock
)) {
4123 if (need_resched()) {
4124 _raw_spin_unlock(lock
);
4125 preempt_enable_no_resched();
4133 EXPORT_SYMBOL(cond_resched_lock
);
4135 int __sched
cond_resched_softirq(void)
4137 BUG_ON(!in_softirq());
4139 if (need_resched()) {
4140 __local_bh_enable();
4148 EXPORT_SYMBOL(cond_resched_softirq
);
4152 * yield - yield the current processor to other threads.
4154 * this is a shortcut for kernel-space yielding - it marks the
4155 * thread runnable and calls sys_sched_yield().
4157 void __sched
yield(void)
4159 set_current_state(TASK_RUNNING
);
4163 EXPORT_SYMBOL(yield
);
4166 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4167 * that process accounting knows that this is a task in IO wait state.
4169 * But don't do that if it is a deliberate, throttling IO wait (this task
4170 * has set its backing_dev_info: the queue against which it should throttle)
4172 void __sched
io_schedule(void)
4174 struct runqueue
*rq
= &per_cpu(runqueues
, raw_smp_processor_id());
4176 atomic_inc(&rq
->nr_iowait
);
4178 atomic_dec(&rq
->nr_iowait
);
4181 EXPORT_SYMBOL(io_schedule
);
4183 long __sched
io_schedule_timeout(long timeout
)
4185 struct runqueue
*rq
= &per_cpu(runqueues
, raw_smp_processor_id());
4188 atomic_inc(&rq
->nr_iowait
);
4189 ret
= schedule_timeout(timeout
);
4190 atomic_dec(&rq
->nr_iowait
);
4195 * sys_sched_get_priority_max - return maximum RT priority.
4196 * @policy: scheduling class.
4198 * this syscall returns the maximum rt_priority that can be used
4199 * by a given scheduling class.
4201 asmlinkage
long sys_sched_get_priority_max(int policy
)
4208 ret
= MAX_USER_RT_PRIO
-1;
4218 * sys_sched_get_priority_min - return minimum RT priority.
4219 * @policy: scheduling class.
4221 * this syscall returns the minimum rt_priority that can be used
4222 * by a given scheduling class.
4224 asmlinkage
long sys_sched_get_priority_min(int policy
)
4240 * sys_sched_rr_get_interval - return the default timeslice of a process.
4241 * @pid: pid of the process.
4242 * @interval: userspace pointer to the timeslice value.
4244 * this syscall writes the default timeslice value of a given process
4245 * into the user-space timespec buffer. A value of '0' means infinity.
4248 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4250 int retval
= -EINVAL
;
4258 read_lock(&tasklist_lock
);
4259 p
= find_process_by_pid(pid
);
4263 retval
= security_task_getscheduler(p
);
4267 jiffies_to_timespec(p
->policy
& SCHED_FIFO
?
4268 0 : task_timeslice(p
), &t
);
4269 read_unlock(&tasklist_lock
);
4270 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4274 read_unlock(&tasklist_lock
);
4278 static inline struct task_struct
*eldest_child(struct task_struct
*p
)
4280 if (list_empty(&p
->children
)) return NULL
;
4281 return list_entry(p
->children
.next
,struct task_struct
,sibling
);
4284 static inline struct task_struct
*older_sibling(struct task_struct
*p
)
4286 if (p
->sibling
.prev
==&p
->parent
->children
) return NULL
;
4287 return list_entry(p
->sibling
.prev
,struct task_struct
,sibling
);
4290 static inline struct task_struct
*younger_sibling(struct task_struct
*p
)
4292 if (p
->sibling
.next
==&p
->parent
->children
) return NULL
;
4293 return list_entry(p
->sibling
.next
,struct task_struct
,sibling
);
4296 static void show_task(task_t
*p
)
4300 unsigned long free
= 0;
4301 static const char *stat_nam
[] = { "R", "S", "D", "T", "t", "Z", "X" };
4303 printk("%-13.13s ", p
->comm
);
4304 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4305 if (state
< ARRAY_SIZE(stat_nam
))
4306 printk(stat_nam
[state
]);
4309 #if (BITS_PER_LONG == 32)
4310 if (state
== TASK_RUNNING
)
4311 printk(" running ");
4313 printk(" %08lX ", thread_saved_pc(p
));
4315 if (state
== TASK_RUNNING
)
4316 printk(" running task ");
4318 printk(" %016lx ", thread_saved_pc(p
));
4320 #ifdef CONFIG_DEBUG_STACK_USAGE
4322 unsigned long *n
= (unsigned long *) (p
->thread_info
+1);
4325 free
= (unsigned long) n
- (unsigned long)(p
->thread_info
+1);
4328 printk("%5lu %5d %6d ", free
, p
->pid
, p
->parent
->pid
);
4329 if ((relative
= eldest_child(p
)))
4330 printk("%5d ", relative
->pid
);
4333 if ((relative
= younger_sibling(p
)))
4334 printk("%7d", relative
->pid
);
4337 if ((relative
= older_sibling(p
)))
4338 printk(" %5d", relative
->pid
);
4342 printk(" (L-TLB)\n");
4344 printk(" (NOTLB)\n");
4346 if (state
!= TASK_RUNNING
)
4347 show_stack(p
, NULL
);
4350 void show_state(void)
4354 #if (BITS_PER_LONG == 32)
4357 printk(" task PC pid father child younger older\n");
4361 printk(" task PC pid father child younger older\n");
4363 read_lock(&tasklist_lock
);
4364 do_each_thread(g
, p
) {
4366 * reset the NMI-timeout, listing all files on a slow
4367 * console might take alot of time:
4369 touch_nmi_watchdog();
4371 } while_each_thread(g
, p
);
4373 read_unlock(&tasklist_lock
);
4377 * init_idle - set up an idle thread for a given CPU
4378 * @idle: task in question
4379 * @cpu: cpu the idle task belongs to
4381 * NOTE: this function does not set the idle thread's NEED_RESCHED
4382 * flag, to make booting more robust.
4384 void __devinit
init_idle(task_t
*idle
, int cpu
)
4386 runqueue_t
*rq
= cpu_rq(cpu
);
4387 unsigned long flags
;
4389 idle
->sleep_avg
= 0;
4391 idle
->prio
= MAX_PRIO
;
4392 idle
->state
= TASK_RUNNING
;
4393 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4394 set_task_cpu(idle
, cpu
);
4396 spin_lock_irqsave(&rq
->lock
, flags
);
4397 rq
->curr
= rq
->idle
= idle
;
4398 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4401 spin_unlock_irqrestore(&rq
->lock
, flags
);
4403 /* Set the preempt count _outside_ the spinlocks! */
4404 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4405 idle
->thread_info
->preempt_count
= (idle
->lock_depth
>= 0);
4407 idle
->thread_info
->preempt_count
= 0;
4412 * In a system that switches off the HZ timer nohz_cpu_mask
4413 * indicates which cpus entered this state. This is used
4414 * in the rcu update to wait only for active cpus. For system
4415 * which do not switch off the HZ timer nohz_cpu_mask should
4416 * always be CPU_MASK_NONE.
4418 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4422 * This is how migration works:
4424 * 1) we queue a migration_req_t structure in the source CPU's
4425 * runqueue and wake up that CPU's migration thread.
4426 * 2) we down() the locked semaphore => thread blocks.
4427 * 3) migration thread wakes up (implicitly it forces the migrated
4428 * thread off the CPU)
4429 * 4) it gets the migration request and checks whether the migrated
4430 * task is still in the wrong runqueue.
4431 * 5) if it's in the wrong runqueue then the migration thread removes
4432 * it and puts it into the right queue.
4433 * 6) migration thread up()s the semaphore.
4434 * 7) we wake up and the migration is done.
4438 * Change a given task's CPU affinity. Migrate the thread to a
4439 * proper CPU and schedule it away if the CPU it's executing on
4440 * is removed from the allowed bitmask.
4442 * NOTE: the caller must have a valid reference to the task, the
4443 * task must not exit() & deallocate itself prematurely. The
4444 * call is not atomic; no spinlocks may be held.
4446 int set_cpus_allowed(task_t
*p
, cpumask_t new_mask
)
4448 unsigned long flags
;
4450 migration_req_t req
;
4453 rq
= task_rq_lock(p
, &flags
);
4454 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4459 p
->cpus_allowed
= new_mask
;
4460 /* Can the task run on the task's current CPU? If so, we're done */
4461 if (cpu_isset(task_cpu(p
), new_mask
))
4464 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4465 /* Need help from migration thread: drop lock and wait. */
4466 task_rq_unlock(rq
, &flags
);
4467 wake_up_process(rq
->migration_thread
);
4468 wait_for_completion(&req
.done
);
4469 tlb_migrate_finish(p
->mm
);
4473 task_rq_unlock(rq
, &flags
);
4477 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4480 * Move (not current) task off this cpu, onto dest cpu. We're doing
4481 * this because either it can't run here any more (set_cpus_allowed()
4482 * away from this CPU, or CPU going down), or because we're
4483 * attempting to rebalance this task on exec (sched_exec).
4485 * So we race with normal scheduler movements, but that's OK, as long
4486 * as the task is no longer on this CPU.
4488 static void __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4490 runqueue_t
*rq_dest
, *rq_src
;
4492 if (unlikely(cpu_is_offline(dest_cpu
)))
4495 rq_src
= cpu_rq(src_cpu
);
4496 rq_dest
= cpu_rq(dest_cpu
);
4498 double_rq_lock(rq_src
, rq_dest
);
4499 /* Already moved. */
4500 if (task_cpu(p
) != src_cpu
)
4502 /* Affinity changed (again). */
4503 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
4506 set_task_cpu(p
, dest_cpu
);
4509 * Sync timestamp with rq_dest's before activating.
4510 * The same thing could be achieved by doing this step
4511 * afterwards, and pretending it was a local activate.
4512 * This way is cleaner and logically correct.
4514 p
->timestamp
= p
->timestamp
- rq_src
->timestamp_last_tick
4515 + rq_dest
->timestamp_last_tick
;
4516 deactivate_task(p
, rq_src
);
4517 activate_task(p
, rq_dest
, 0);
4518 if (TASK_PREEMPTS_CURR(p
, rq_dest
))
4519 resched_task(rq_dest
->curr
);
4523 double_rq_unlock(rq_src
, rq_dest
);
4527 * migration_thread - this is a highprio system thread that performs
4528 * thread migration by bumping thread off CPU then 'pushing' onto
4531 static int migration_thread(void *data
)
4534 int cpu
= (long)data
;
4537 BUG_ON(rq
->migration_thread
!= current
);
4539 set_current_state(TASK_INTERRUPTIBLE
);
4540 while (!kthread_should_stop()) {
4541 struct list_head
*head
;
4542 migration_req_t
*req
;
4546 spin_lock_irq(&rq
->lock
);
4548 if (cpu_is_offline(cpu
)) {
4549 spin_unlock_irq(&rq
->lock
);
4553 if (rq
->active_balance
) {
4554 active_load_balance(rq
, cpu
);
4555 rq
->active_balance
= 0;
4558 head
= &rq
->migration_queue
;
4560 if (list_empty(head
)) {
4561 spin_unlock_irq(&rq
->lock
);
4563 set_current_state(TASK_INTERRUPTIBLE
);
4566 req
= list_entry(head
->next
, migration_req_t
, list
);
4567 list_del_init(head
->next
);
4569 spin_unlock(&rq
->lock
);
4570 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
4573 complete(&req
->done
);
4575 __set_current_state(TASK_RUNNING
);
4579 /* Wait for kthread_stop */
4580 set_current_state(TASK_INTERRUPTIBLE
);
4581 while (!kthread_should_stop()) {
4583 set_current_state(TASK_INTERRUPTIBLE
);
4585 __set_current_state(TASK_RUNNING
);
4589 #ifdef CONFIG_HOTPLUG_CPU
4590 /* Figure out where task on dead CPU should go, use force if neccessary. */
4591 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*tsk
)
4597 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
4598 cpus_and(mask
, mask
, tsk
->cpus_allowed
);
4599 dest_cpu
= any_online_cpu(mask
);
4601 /* On any allowed CPU? */
4602 if (dest_cpu
== NR_CPUS
)
4603 dest_cpu
= any_online_cpu(tsk
->cpus_allowed
);
4605 /* No more Mr. Nice Guy. */
4606 if (dest_cpu
== NR_CPUS
) {
4607 cpus_setall(tsk
->cpus_allowed
);
4608 dest_cpu
= any_online_cpu(tsk
->cpus_allowed
);
4611 * Don't tell them about moving exiting tasks or
4612 * kernel threads (both mm NULL), since they never
4615 if (tsk
->mm
&& printk_ratelimit())
4616 printk(KERN_INFO
"process %d (%s) no "
4617 "longer affine to cpu%d\n",
4618 tsk
->pid
, tsk
->comm
, dead_cpu
);
4620 __migrate_task(tsk
, dead_cpu
, dest_cpu
);
4624 * While a dead CPU has no uninterruptible tasks queued at this point,
4625 * it might still have a nonzero ->nr_uninterruptible counter, because
4626 * for performance reasons the counter is not stricly tracking tasks to
4627 * their home CPUs. So we just add the counter to another CPU's counter,
4628 * to keep the global sum constant after CPU-down:
4630 static void migrate_nr_uninterruptible(runqueue_t
*rq_src
)
4632 runqueue_t
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
4633 unsigned long flags
;
4635 local_irq_save(flags
);
4636 double_rq_lock(rq_src
, rq_dest
);
4637 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
4638 rq_src
->nr_uninterruptible
= 0;
4639 double_rq_unlock(rq_src
, rq_dest
);
4640 local_irq_restore(flags
);
4643 /* Run through task list and migrate tasks from the dead cpu. */
4644 static void migrate_live_tasks(int src_cpu
)
4646 struct task_struct
*tsk
, *t
;
4648 write_lock_irq(&tasklist_lock
);
4650 do_each_thread(t
, tsk
) {
4654 if (task_cpu(tsk
) == src_cpu
)
4655 move_task_off_dead_cpu(src_cpu
, tsk
);
4656 } while_each_thread(t
, tsk
);
4658 write_unlock_irq(&tasklist_lock
);
4661 /* Schedules idle task to be the next runnable task on current CPU.
4662 * It does so by boosting its priority to highest possible and adding it to
4663 * the _front_ of runqueue. Used by CPU offline code.
4665 void sched_idle_next(void)
4667 int cpu
= smp_processor_id();
4668 runqueue_t
*rq
= this_rq();
4669 struct task_struct
*p
= rq
->idle
;
4670 unsigned long flags
;
4672 /* cpu has to be offline */
4673 BUG_ON(cpu_online(cpu
));
4675 /* Strictly not necessary since rest of the CPUs are stopped by now
4676 * and interrupts disabled on current cpu.
4678 spin_lock_irqsave(&rq
->lock
, flags
);
4680 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
4681 /* Add idle task to _front_ of it's priority queue */
4682 __activate_idle_task(p
, rq
);
4684 spin_unlock_irqrestore(&rq
->lock
, flags
);
4687 /* Ensures that the idle task is using init_mm right before its cpu goes
4690 void idle_task_exit(void)
4692 struct mm_struct
*mm
= current
->active_mm
;
4694 BUG_ON(cpu_online(smp_processor_id()));
4697 switch_mm(mm
, &init_mm
, current
);
4701 static void migrate_dead(unsigned int dead_cpu
, task_t
*tsk
)
4703 struct runqueue
*rq
= cpu_rq(dead_cpu
);
4705 /* Must be exiting, otherwise would be on tasklist. */
4706 BUG_ON(tsk
->exit_state
!= EXIT_ZOMBIE
&& tsk
->exit_state
!= EXIT_DEAD
);
4708 /* Cannot have done final schedule yet: would have vanished. */
4709 BUG_ON(tsk
->flags
& PF_DEAD
);
4711 get_task_struct(tsk
);
4714 * Drop lock around migration; if someone else moves it,
4715 * that's OK. No task can be added to this CPU, so iteration is
4718 spin_unlock_irq(&rq
->lock
);
4719 move_task_off_dead_cpu(dead_cpu
, tsk
);
4720 spin_lock_irq(&rq
->lock
);
4722 put_task_struct(tsk
);
4725 /* release_task() removes task from tasklist, so we won't find dead tasks. */
4726 static void migrate_dead_tasks(unsigned int dead_cpu
)
4729 struct runqueue
*rq
= cpu_rq(dead_cpu
);
4731 for (arr
= 0; arr
< 2; arr
++) {
4732 for (i
= 0; i
< MAX_PRIO
; i
++) {
4733 struct list_head
*list
= &rq
->arrays
[arr
].queue
[i
];
4734 while (!list_empty(list
))
4735 migrate_dead(dead_cpu
,
4736 list_entry(list
->next
, task_t
,
4741 #endif /* CONFIG_HOTPLUG_CPU */
4744 * migration_call - callback that gets triggered when a CPU is added.
4745 * Here we can start up the necessary migration thread for the new CPU.
4747 static int migration_call(struct notifier_block
*nfb
, unsigned long action
,
4750 int cpu
= (long)hcpu
;
4751 struct task_struct
*p
;
4752 struct runqueue
*rq
;
4753 unsigned long flags
;
4756 case CPU_UP_PREPARE
:
4757 p
= kthread_create(migration_thread
, hcpu
, "migration/%d",cpu
);
4760 p
->flags
|= PF_NOFREEZE
;
4761 kthread_bind(p
, cpu
);
4762 /* Must be high prio: stop_machine expects to yield to it. */
4763 rq
= task_rq_lock(p
, &flags
);
4764 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
4765 task_rq_unlock(rq
, &flags
);
4766 cpu_rq(cpu
)->migration_thread
= p
;
4769 /* Strictly unneccessary, as first user will wake it. */
4770 wake_up_process(cpu_rq(cpu
)->migration_thread
);
4772 #ifdef CONFIG_HOTPLUG_CPU
4773 case CPU_UP_CANCELED
:
4774 /* Unbind it from offline cpu so it can run. Fall thru. */
4775 kthread_bind(cpu_rq(cpu
)->migration_thread
,
4776 any_online_cpu(cpu_online_map
));
4777 kthread_stop(cpu_rq(cpu
)->migration_thread
);
4778 cpu_rq(cpu
)->migration_thread
= NULL
;
4781 migrate_live_tasks(cpu
);
4783 kthread_stop(rq
->migration_thread
);
4784 rq
->migration_thread
= NULL
;
4785 /* Idle task back to normal (off runqueue, low prio) */
4786 rq
= task_rq_lock(rq
->idle
, &flags
);
4787 deactivate_task(rq
->idle
, rq
);
4788 rq
->idle
->static_prio
= MAX_PRIO
;
4789 __setscheduler(rq
->idle
, SCHED_NORMAL
, 0);
4790 migrate_dead_tasks(cpu
);
4791 task_rq_unlock(rq
, &flags
);
4792 migrate_nr_uninterruptible(rq
);
4793 BUG_ON(rq
->nr_running
!= 0);
4795 /* No need to migrate the tasks: it was best-effort if
4796 * they didn't do lock_cpu_hotplug(). Just wake up
4797 * the requestors. */
4798 spin_lock_irq(&rq
->lock
);
4799 while (!list_empty(&rq
->migration_queue
)) {
4800 migration_req_t
*req
;
4801 req
= list_entry(rq
->migration_queue
.next
,
4802 migration_req_t
, list
);
4803 list_del_init(&req
->list
);
4804 complete(&req
->done
);
4806 spin_unlock_irq(&rq
->lock
);
4813 /* Register at highest priority so that task migration (migrate_all_tasks)
4814 * happens before everything else.
4816 static struct notifier_block __devinitdata migration_notifier
= {
4817 .notifier_call
= migration_call
,
4821 int __init
migration_init(void)
4823 void *cpu
= (void *)(long)smp_processor_id();
4824 /* Start one for boot CPU. */
4825 migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
4826 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
4827 register_cpu_notifier(&migration_notifier
);
4833 #undef SCHED_DOMAIN_DEBUG
4834 #ifdef SCHED_DOMAIN_DEBUG
4835 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
4840 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
4844 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
4849 struct sched_group
*group
= sd
->groups
;
4850 cpumask_t groupmask
;
4852 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
4853 cpus_clear(groupmask
);
4856 for (i
= 0; i
< level
+ 1; i
++)
4858 printk("domain %d: ", level
);
4860 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
4861 printk("does not load-balance\n");
4863 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain has parent");
4867 printk("span %s\n", str
);
4869 if (!cpu_isset(cpu
, sd
->span
))
4870 printk(KERN_ERR
"ERROR: domain->span does not contain CPU%d\n", cpu
);
4871 if (!cpu_isset(cpu
, group
->cpumask
))
4872 printk(KERN_ERR
"ERROR: domain->groups does not contain CPU%d\n", cpu
);
4875 for (i
= 0; i
< level
+ 2; i
++)
4881 printk(KERN_ERR
"ERROR: group is NULL\n");
4885 if (!group
->cpu_power
) {
4887 printk(KERN_ERR
"ERROR: domain->cpu_power not set\n");
4890 if (!cpus_weight(group
->cpumask
)) {
4892 printk(KERN_ERR
"ERROR: empty group\n");
4895 if (cpus_intersects(groupmask
, group
->cpumask
)) {
4897 printk(KERN_ERR
"ERROR: repeated CPUs\n");
4900 cpus_or(groupmask
, groupmask
, group
->cpumask
);
4902 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
4905 group
= group
->next
;
4906 } while (group
!= sd
->groups
);
4909 if (!cpus_equal(sd
->span
, groupmask
))
4910 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
4916 if (!cpus_subset(groupmask
, sd
->span
))
4917 printk(KERN_ERR
"ERROR: parent span is not a superset of domain->span\n");
4923 #define sched_domain_debug(sd, cpu) {}
4926 static int sd_degenerate(struct sched_domain
*sd
)
4928 if (cpus_weight(sd
->span
) == 1)
4931 /* Following flags need at least 2 groups */
4932 if (sd
->flags
& (SD_LOAD_BALANCE
|
4933 SD_BALANCE_NEWIDLE
|
4936 if (sd
->groups
!= sd
->groups
->next
)
4940 /* Following flags don't use groups */
4941 if (sd
->flags
& (SD_WAKE_IDLE
|
4949 static int sd_parent_degenerate(struct sched_domain
*sd
,
4950 struct sched_domain
*parent
)
4952 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
4954 if (sd_degenerate(parent
))
4957 if (!cpus_equal(sd
->span
, parent
->span
))
4960 /* Does parent contain flags not in child? */
4961 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
4962 if (cflags
& SD_WAKE_AFFINE
)
4963 pflags
&= ~SD_WAKE_BALANCE
;
4964 /* Flags needing groups don't count if only 1 group in parent */
4965 if (parent
->groups
== parent
->groups
->next
) {
4966 pflags
&= ~(SD_LOAD_BALANCE
|
4967 SD_BALANCE_NEWIDLE
|
4971 if (~cflags
& pflags
)
4978 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
4979 * hold the hotplug lock.
4981 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
4983 runqueue_t
*rq
= cpu_rq(cpu
);
4984 struct sched_domain
*tmp
;
4986 /* Remove the sched domains which do not contribute to scheduling. */
4987 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
4988 struct sched_domain
*parent
= tmp
->parent
;
4991 if (sd_parent_degenerate(tmp
, parent
))
4992 tmp
->parent
= parent
->parent
;
4995 if (sd
&& sd_degenerate(sd
))
4998 sched_domain_debug(sd
, cpu
);
5000 rcu_assign_pointer(rq
->sd
, sd
);
5003 /* cpus with isolated domains */
5004 static cpumask_t __devinitdata cpu_isolated_map
= CPU_MASK_NONE
;
5006 /* Setup the mask of cpus configured for isolated domains */
5007 static int __init
isolated_cpu_setup(char *str
)
5009 int ints
[NR_CPUS
], i
;
5011 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5012 cpus_clear(cpu_isolated_map
);
5013 for (i
= 1; i
<= ints
[0]; i
++)
5014 if (ints
[i
] < NR_CPUS
)
5015 cpu_set(ints
[i
], cpu_isolated_map
);
5019 __setup ("isolcpus=", isolated_cpu_setup
);
5022 * init_sched_build_groups takes an array of groups, the cpumask we wish
5023 * to span, and a pointer to a function which identifies what group a CPU
5024 * belongs to. The return value of group_fn must be a valid index into the
5025 * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
5026 * keep track of groups covered with a cpumask_t).
5028 * init_sched_build_groups will build a circular linked list of the groups
5029 * covered by the given span, and will set each group's ->cpumask correctly,
5030 * and ->cpu_power to 0.
5032 static void init_sched_build_groups(struct sched_group groups
[], cpumask_t span
,
5033 int (*group_fn
)(int cpu
))
5035 struct sched_group
*first
= NULL
, *last
= NULL
;
5036 cpumask_t covered
= CPU_MASK_NONE
;
5039 for_each_cpu_mask(i
, span
) {
5040 int group
= group_fn(i
);
5041 struct sched_group
*sg
= &groups
[group
];
5044 if (cpu_isset(i
, covered
))
5047 sg
->cpumask
= CPU_MASK_NONE
;
5050 for_each_cpu_mask(j
, span
) {
5051 if (group_fn(j
) != group
)
5054 cpu_set(j
, covered
);
5055 cpu_set(j
, sg
->cpumask
);
5066 #define SD_NODES_PER_DOMAIN 16
5070 * find_next_best_node - find the next node to include in a sched_domain
5071 * @node: node whose sched_domain we're building
5072 * @used_nodes: nodes already in the sched_domain
5074 * Find the next node to include in a given scheduling domain. Simply
5075 * finds the closest node not already in the @used_nodes map.
5077 * Should use nodemask_t.
5079 static int find_next_best_node(int node
, unsigned long *used_nodes
)
5081 int i
, n
, val
, min_val
, best_node
= 0;
5085 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5086 /* Start at @node */
5087 n
= (node
+ i
) % MAX_NUMNODES
;
5089 if (!nr_cpus_node(n
))
5092 /* Skip already used nodes */
5093 if (test_bit(n
, used_nodes
))
5096 /* Simple min distance search */
5097 val
= node_distance(node
, n
);
5099 if (val
< min_val
) {
5105 set_bit(best_node
, used_nodes
);
5110 * sched_domain_node_span - get a cpumask for a node's sched_domain
5111 * @node: node whose cpumask we're constructing
5112 * @size: number of nodes to include in this span
5114 * Given a node, construct a good cpumask for its sched_domain to span. It
5115 * should be one that prevents unnecessary balancing, but also spreads tasks
5118 static cpumask_t
sched_domain_node_span(int node
)
5121 cpumask_t span
, nodemask
;
5122 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
5125 bitmap_zero(used_nodes
, MAX_NUMNODES
);
5127 nodemask
= node_to_cpumask(node
);
5128 cpus_or(span
, span
, nodemask
);
5129 set_bit(node
, used_nodes
);
5131 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5132 int next_node
= find_next_best_node(node
, used_nodes
);
5133 nodemask
= node_to_cpumask(next_node
);
5134 cpus_or(span
, span
, nodemask
);
5142 * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we
5143 * can switch it on easily if needed.
5145 #ifdef CONFIG_SCHED_SMT
5146 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
5147 static struct sched_group sched_group_cpus
[NR_CPUS
];
5148 static int cpu_to_cpu_group(int cpu
)
5154 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
5155 static struct sched_group sched_group_phys
[NR_CPUS
];
5156 static int cpu_to_phys_group(int cpu
)
5158 #ifdef CONFIG_SCHED_SMT
5159 return first_cpu(cpu_sibling_map
[cpu
]);
5167 * The init_sched_build_groups can't handle what we want to do with node
5168 * groups, so roll our own. Now each node has its own list of groups which
5169 * gets dynamically allocated.
5171 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
5172 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
5174 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
5175 static struct sched_group
*sched_group_allnodes_bycpu
[NR_CPUS
];
5177 static int cpu_to_allnodes_group(int cpu
)
5179 return cpu_to_node(cpu
);
5184 * Build sched domains for a given set of cpus and attach the sched domains
5185 * to the individual cpus
5187 void build_sched_domains(const cpumask_t
*cpu_map
)
5191 struct sched_group
**sched_group_nodes
= NULL
;
5192 struct sched_group
*sched_group_allnodes
= NULL
;
5195 * Allocate the per-node list of sched groups
5197 sched_group_nodes
= kmalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
5199 if (!sched_group_nodes
) {
5200 printk(KERN_WARNING
"Can not alloc sched group node list\n");
5203 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
5207 * Set up domains for cpus specified by the cpu_map.
5209 for_each_cpu_mask(i
, *cpu_map
) {
5211 struct sched_domain
*sd
= NULL
, *p
;
5212 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
5214 cpus_and(nodemask
, nodemask
, *cpu_map
);
5217 if (cpus_weight(*cpu_map
)
5218 > SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
5219 if (!sched_group_allnodes
) {
5220 sched_group_allnodes
5221 = kmalloc(sizeof(struct sched_group
)
5224 if (!sched_group_allnodes
) {
5226 "Can not alloc allnodes sched group\n");
5229 sched_group_allnodes_bycpu
[i
]
5230 = sched_group_allnodes
;
5232 sd
= &per_cpu(allnodes_domains
, i
);
5233 *sd
= SD_ALLNODES_INIT
;
5234 sd
->span
= *cpu_map
;
5235 group
= cpu_to_allnodes_group(i
);
5236 sd
->groups
= &sched_group_allnodes
[group
];
5241 sd
= &per_cpu(node_domains
, i
);
5243 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
5245 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
5249 sd
= &per_cpu(phys_domains
, i
);
5250 group
= cpu_to_phys_group(i
);
5252 sd
->span
= nodemask
;
5254 sd
->groups
= &sched_group_phys
[group
];
5256 #ifdef CONFIG_SCHED_SMT
5258 sd
= &per_cpu(cpu_domains
, i
);
5259 group
= cpu_to_cpu_group(i
);
5260 *sd
= SD_SIBLING_INIT
;
5261 sd
->span
= cpu_sibling_map
[i
];
5262 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
5264 sd
->groups
= &sched_group_cpus
[group
];
5268 #ifdef CONFIG_SCHED_SMT
5269 /* Set up CPU (sibling) groups */
5270 for_each_cpu_mask(i
, *cpu_map
) {
5271 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
5272 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
5273 if (i
!= first_cpu(this_sibling_map
))
5276 init_sched_build_groups(sched_group_cpus
, this_sibling_map
,
5281 /* Set up physical groups */
5282 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5283 cpumask_t nodemask
= node_to_cpumask(i
);
5285 cpus_and(nodemask
, nodemask
, *cpu_map
);
5286 if (cpus_empty(nodemask
))
5289 init_sched_build_groups(sched_group_phys
, nodemask
,
5290 &cpu_to_phys_group
);
5294 /* Set up node groups */
5295 if (sched_group_allnodes
)
5296 init_sched_build_groups(sched_group_allnodes
, *cpu_map
,
5297 &cpu_to_allnodes_group
);
5299 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5300 /* Set up node groups */
5301 struct sched_group
*sg
, *prev
;
5302 cpumask_t nodemask
= node_to_cpumask(i
);
5303 cpumask_t domainspan
;
5304 cpumask_t covered
= CPU_MASK_NONE
;
5307 cpus_and(nodemask
, nodemask
, *cpu_map
);
5308 if (cpus_empty(nodemask
)) {
5309 sched_group_nodes
[i
] = NULL
;
5313 domainspan
= sched_domain_node_span(i
);
5314 cpus_and(domainspan
, domainspan
, *cpu_map
);
5316 sg
= kmalloc(sizeof(struct sched_group
), GFP_KERNEL
);
5317 sched_group_nodes
[i
] = sg
;
5318 for_each_cpu_mask(j
, nodemask
) {
5319 struct sched_domain
*sd
;
5320 sd
= &per_cpu(node_domains
, j
);
5322 if (sd
->groups
== NULL
) {
5323 /* Turn off balancing if we have no groups */
5329 "Can not alloc domain group for node %d\n", i
);
5333 sg
->cpumask
= nodemask
;
5334 cpus_or(covered
, covered
, nodemask
);
5337 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
5338 cpumask_t tmp
, notcovered
;
5339 int n
= (i
+ j
) % MAX_NUMNODES
;
5341 cpus_complement(notcovered
, covered
);
5342 cpus_and(tmp
, notcovered
, *cpu_map
);
5343 cpus_and(tmp
, tmp
, domainspan
);
5344 if (cpus_empty(tmp
))
5347 nodemask
= node_to_cpumask(n
);
5348 cpus_and(tmp
, tmp
, nodemask
);
5349 if (cpus_empty(tmp
))
5352 sg
= kmalloc(sizeof(struct sched_group
), GFP_KERNEL
);
5355 "Can not alloc domain group for node %d\n", j
);
5360 cpus_or(covered
, covered
, tmp
);
5364 prev
->next
= sched_group_nodes
[i
];
5368 /* Calculate CPU power for physical packages and nodes */
5369 for_each_cpu_mask(i
, *cpu_map
) {
5371 struct sched_domain
*sd
;
5372 #ifdef CONFIG_SCHED_SMT
5373 sd
= &per_cpu(cpu_domains
, i
);
5374 power
= SCHED_LOAD_SCALE
;
5375 sd
->groups
->cpu_power
= power
;
5378 sd
= &per_cpu(phys_domains
, i
);
5379 power
= SCHED_LOAD_SCALE
+ SCHED_LOAD_SCALE
*
5380 (cpus_weight(sd
->groups
->cpumask
)-1) / 10;
5381 sd
->groups
->cpu_power
= power
;
5384 sd
= &per_cpu(allnodes_domains
, i
);
5386 power
= SCHED_LOAD_SCALE
+ SCHED_LOAD_SCALE
*
5387 (cpus_weight(sd
->groups
->cpumask
)-1) / 10;
5388 sd
->groups
->cpu_power
= power
;
5394 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5395 struct sched_group
*sg
= sched_group_nodes
[i
];
5401 for_each_cpu_mask(j
, sg
->cpumask
) {
5402 struct sched_domain
*sd
;
5405 sd
= &per_cpu(phys_domains
, j
);
5406 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
5408 * Only add "power" once for each
5413 power
= SCHED_LOAD_SCALE
+ SCHED_LOAD_SCALE
*
5414 (cpus_weight(sd
->groups
->cpumask
)-1) / 10;
5416 sg
->cpu_power
+= power
;
5419 if (sg
!= sched_group_nodes
[i
])
5424 /* Attach the domains */
5425 for_each_cpu_mask(i
, *cpu_map
) {
5426 struct sched_domain
*sd
;
5427 #ifdef CONFIG_SCHED_SMT
5428 sd
= &per_cpu(cpu_domains
, i
);
5430 sd
= &per_cpu(phys_domains
, i
);
5432 cpu_attach_domain(sd
, i
);
5436 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
5438 static void arch_init_sched_domains(const cpumask_t
*cpu_map
)
5440 cpumask_t cpu_default_map
;
5443 * Setup mask for cpus without special case scheduling requirements.
5444 * For now this just excludes isolated cpus, but could be used to
5445 * exclude other special cases in the future.
5447 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
5449 build_sched_domains(&cpu_default_map
);
5452 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
5458 for_each_cpu_mask(cpu
, *cpu_map
) {
5459 struct sched_group
*sched_group_allnodes
5460 = sched_group_allnodes_bycpu
[cpu
];
5461 struct sched_group
**sched_group_nodes
5462 = sched_group_nodes_bycpu
[cpu
];
5464 if (sched_group_allnodes
) {
5465 kfree(sched_group_allnodes
);
5466 sched_group_allnodes_bycpu
[cpu
] = NULL
;
5469 if (!sched_group_nodes
)
5472 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5473 cpumask_t nodemask
= node_to_cpumask(i
);
5474 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
5476 cpus_and(nodemask
, nodemask
, *cpu_map
);
5477 if (cpus_empty(nodemask
))
5487 if (oldsg
!= sched_group_nodes
[i
])
5490 kfree(sched_group_nodes
);
5491 sched_group_nodes_bycpu
[cpu
] = NULL
;
5497 * Detach sched domains from a group of cpus specified in cpu_map
5498 * These cpus will now be attached to the NULL domain
5500 static inline void detach_destroy_domains(const cpumask_t
*cpu_map
)
5504 for_each_cpu_mask(i
, *cpu_map
)
5505 cpu_attach_domain(NULL
, i
);
5506 synchronize_sched();
5507 arch_destroy_sched_domains(cpu_map
);
5511 * Partition sched domains as specified by the cpumasks below.
5512 * This attaches all cpus from the cpumasks to the NULL domain,
5513 * waits for a RCU quiescent period, recalculates sched
5514 * domain information and then attaches them back to the
5515 * correct sched domains
5516 * Call with hotplug lock held
5518 void partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
5520 cpumask_t change_map
;
5522 cpus_and(*partition1
, *partition1
, cpu_online_map
);
5523 cpus_and(*partition2
, *partition2
, cpu_online_map
);
5524 cpus_or(change_map
, *partition1
, *partition2
);
5526 /* Detach sched domains from all of the affected cpus */
5527 detach_destroy_domains(&change_map
);
5528 if (!cpus_empty(*partition1
))
5529 build_sched_domains(partition1
);
5530 if (!cpus_empty(*partition2
))
5531 build_sched_domains(partition2
);
5534 #ifdef CONFIG_HOTPLUG_CPU
5536 * Force a reinitialization of the sched domains hierarchy. The domains
5537 * and groups cannot be updated in place without racing with the balancing
5538 * code, so we temporarily attach all running cpus to the NULL domain
5539 * which will prevent rebalancing while the sched domains are recalculated.
5541 static int update_sched_domains(struct notifier_block
*nfb
,
5542 unsigned long action
, void *hcpu
)
5545 case CPU_UP_PREPARE
:
5546 case CPU_DOWN_PREPARE
:
5547 detach_destroy_domains(&cpu_online_map
);
5550 case CPU_UP_CANCELED
:
5551 case CPU_DOWN_FAILED
:
5555 * Fall through and re-initialise the domains.
5562 /* The hotplug lock is already held by cpu_up/cpu_down */
5563 arch_init_sched_domains(&cpu_online_map
);
5569 void __init
sched_init_smp(void)
5572 arch_init_sched_domains(&cpu_online_map
);
5573 unlock_cpu_hotplug();
5574 /* XXX: Theoretical race here - CPU may be hotplugged now */
5575 hotcpu_notifier(update_sched_domains
, 0);
5578 void __init
sched_init_smp(void)
5581 #endif /* CONFIG_SMP */
5583 int in_sched_functions(unsigned long addr
)
5585 /* Linker adds these: start and end of __sched functions */
5586 extern char __sched_text_start
[], __sched_text_end
[];
5587 return in_lock_functions(addr
) ||
5588 (addr
>= (unsigned long)__sched_text_start
5589 && addr
< (unsigned long)__sched_text_end
);
5592 void __init
sched_init(void)
5597 for (i
= 0; i
< NR_CPUS
; i
++) {
5598 prio_array_t
*array
;
5601 spin_lock_init(&rq
->lock
);
5603 rq
->active
= rq
->arrays
;
5604 rq
->expired
= rq
->arrays
+ 1;
5605 rq
->best_expired_prio
= MAX_PRIO
;
5609 for (j
= 1; j
< 3; j
++)
5610 rq
->cpu_load
[j
] = 0;
5611 rq
->active_balance
= 0;
5613 rq
->migration_thread
= NULL
;
5614 INIT_LIST_HEAD(&rq
->migration_queue
);
5616 atomic_set(&rq
->nr_iowait
, 0);
5618 for (j
= 0; j
< 2; j
++) {
5619 array
= rq
->arrays
+ j
;
5620 for (k
= 0; k
< MAX_PRIO
; k
++) {
5621 INIT_LIST_HEAD(array
->queue
+ k
);
5622 __clear_bit(k
, array
->bitmap
);
5624 // delimiter for bitsearch
5625 __set_bit(MAX_PRIO
, array
->bitmap
);
5630 * The boot idle thread does lazy MMU switching as well:
5632 atomic_inc(&init_mm
.mm_count
);
5633 enter_lazy_tlb(&init_mm
, current
);
5636 * Make us the idle thread. Technically, schedule() should not be
5637 * called from this thread, however somewhere below it might be,
5638 * but because we are the idle thread, we just pick up running again
5639 * when this runqueue becomes "idle".
5641 init_idle(current
, smp_processor_id());
5644 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5645 void __might_sleep(char *file
, int line
)
5647 #if defined(in_atomic)
5648 static unsigned long prev_jiffy
; /* ratelimiting */
5650 if ((in_atomic() || irqs_disabled()) &&
5651 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
5652 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
5654 prev_jiffy
= jiffies
;
5655 printk(KERN_ERR
"Debug: sleeping function called from invalid"
5656 " context at %s:%d\n", file
, line
);
5657 printk("in_atomic():%d, irqs_disabled():%d\n",
5658 in_atomic(), irqs_disabled());
5663 EXPORT_SYMBOL(__might_sleep
);
5666 #ifdef CONFIG_MAGIC_SYSRQ
5667 void normalize_rt_tasks(void)
5669 struct task_struct
*p
;
5670 prio_array_t
*array
;
5671 unsigned long flags
;
5674 read_lock_irq(&tasklist_lock
);
5675 for_each_process (p
) {
5679 rq
= task_rq_lock(p
, &flags
);
5683 deactivate_task(p
, task_rq(p
));
5684 __setscheduler(p
, SCHED_NORMAL
, 0);
5686 __activate_task(p
, task_rq(p
));
5687 resched_task(rq
->curr
);
5690 task_rq_unlock(rq
, &flags
);
5692 read_unlock_irq(&tasklist_lock
);
5695 #endif /* CONFIG_MAGIC_SYSRQ */
5699 * These functions are only useful for the IA64 MCA handling.
5701 * They can only be called when the whole system has been
5702 * stopped - every CPU needs to be quiescent, and no scheduling
5703 * activity can take place. Using them for anything else would
5704 * be a serious bug, and as a result, they aren't even visible
5705 * under any other configuration.
5709 * curr_task - return the current task for a given cpu.
5710 * @cpu: the processor in question.
5712 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
5714 task_t
*curr_task(int cpu
)
5716 return cpu_curr(cpu
);
5720 * set_curr_task - set the current task for a given cpu.
5721 * @cpu: the processor in question.
5722 * @p: the task pointer to set.
5724 * Description: This function must only be used when non-maskable interrupts
5725 * are serviced on a separate stack. It allows the architecture to switch the
5726 * notion of the current task on a cpu in a non-blocking manner. This function
5727 * must be called with all CPU's synchronized, and interrupts disabled, the
5728 * and caller must save the original value of the current task (see
5729 * curr_task() above) and restore that value before reenabling interrupts and
5730 * re-starting the system.
5732 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
5734 void set_curr_task(int cpu
, task_t
*p
)