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_prio_bias(runqueue_t
*rq
, int prio
)
678 static inline void dec_prio_bias(runqueue_t
*rq
, int prio
)
683 static inline void inc_nr_running(task_t
*p
, runqueue_t
*rq
)
687 inc_prio_bias(rq
, p
->prio
);
689 inc_prio_bias(rq
, p
->static_prio
);
692 static inline void dec_nr_running(task_t
*p
, runqueue_t
*rq
)
696 dec_prio_bias(rq
, p
->prio
);
698 dec_prio_bias(rq
, p
->static_prio
);
702 * __activate_task - move a task to the runqueue.
704 static inline void __activate_task(task_t
*p
, runqueue_t
*rq
)
706 enqueue_task(p
, rq
->active
);
707 inc_nr_running(p
, rq
);
711 * __activate_idle_task - move idle task to the _front_ of runqueue.
713 static inline void __activate_idle_task(task_t
*p
, runqueue_t
*rq
)
715 enqueue_task_head(p
, rq
->active
);
716 inc_nr_running(p
, rq
);
719 static int recalc_task_prio(task_t
*p
, unsigned long long now
)
721 /* Caller must always ensure 'now >= p->timestamp' */
722 unsigned long long __sleep_time
= now
- p
->timestamp
;
723 unsigned long sleep_time
;
725 if (__sleep_time
> NS_MAX_SLEEP_AVG
)
726 sleep_time
= NS_MAX_SLEEP_AVG
;
728 sleep_time
= (unsigned long)__sleep_time
;
730 if (likely(sleep_time
> 0)) {
732 * User tasks that sleep a long time are categorised as
733 * idle and will get just interactive status to stay active &
734 * prevent them suddenly becoming cpu hogs and starving
737 if (p
->mm
&& p
->activated
!= -1 &&
738 sleep_time
> INTERACTIVE_SLEEP(p
)) {
739 p
->sleep_avg
= JIFFIES_TO_NS(MAX_SLEEP_AVG
-
743 * The lower the sleep avg a task has the more
744 * rapidly it will rise with sleep time.
746 sleep_time
*= (MAX_BONUS
- CURRENT_BONUS(p
)) ? : 1;
749 * Tasks waking from uninterruptible sleep are
750 * limited in their sleep_avg rise as they
751 * are likely to be waiting on I/O
753 if (p
->activated
== -1 && p
->mm
) {
754 if (p
->sleep_avg
>= INTERACTIVE_SLEEP(p
))
756 else if (p
->sleep_avg
+ sleep_time
>=
757 INTERACTIVE_SLEEP(p
)) {
758 p
->sleep_avg
= INTERACTIVE_SLEEP(p
);
764 * This code gives a bonus to interactive tasks.
766 * The boost works by updating the 'average sleep time'
767 * value here, based on ->timestamp. The more time a
768 * task spends sleeping, the higher the average gets -
769 * and the higher the priority boost gets as well.
771 p
->sleep_avg
+= sleep_time
;
773 if (p
->sleep_avg
> NS_MAX_SLEEP_AVG
)
774 p
->sleep_avg
= NS_MAX_SLEEP_AVG
;
778 return effective_prio(p
);
782 * activate_task - move a task to the runqueue and do priority recalculation
784 * Update all the scheduling statistics stuff. (sleep average
785 * calculation, priority modifiers, etc.)
787 static void activate_task(task_t
*p
, runqueue_t
*rq
, int local
)
789 unsigned long long now
;
794 /* Compensate for drifting sched_clock */
795 runqueue_t
*this_rq
= this_rq();
796 now
= (now
- this_rq
->timestamp_last_tick
)
797 + rq
->timestamp_last_tick
;
801 p
->prio
= recalc_task_prio(p
, now
);
804 * This checks to make sure it's not an uninterruptible task
805 * that is now waking up.
809 * Tasks which were woken up by interrupts (ie. hw events)
810 * are most likely of interactive nature. So we give them
811 * the credit of extending their sleep time to the period
812 * of time they spend on the runqueue, waiting for execution
813 * on a CPU, first time around:
819 * Normal first-time wakeups get a credit too for
820 * on-runqueue time, but it will be weighted down:
827 __activate_task(p
, rq
);
831 * deactivate_task - remove a task from the runqueue.
833 static void deactivate_task(struct task_struct
*p
, runqueue_t
*rq
)
835 dec_nr_running(p
, rq
);
836 dequeue_task(p
, p
->array
);
841 * resched_task - mark a task 'to be rescheduled now'.
843 * On UP this means the setting of the need_resched flag, on SMP it
844 * might also involve a cross-CPU call to trigger the scheduler on
848 static void resched_task(task_t
*p
)
850 int need_resched
, nrpolling
;
852 assert_spin_locked(&task_rq(p
)->lock
);
854 /* minimise the chance of sending an interrupt to poll_idle() */
855 nrpolling
= test_tsk_thread_flag(p
,TIF_POLLING_NRFLAG
);
856 need_resched
= test_and_set_tsk_thread_flag(p
,TIF_NEED_RESCHED
);
857 nrpolling
|= test_tsk_thread_flag(p
,TIF_POLLING_NRFLAG
);
859 if (!need_resched
&& !nrpolling
&& (task_cpu(p
) != smp_processor_id()))
860 smp_send_reschedule(task_cpu(p
));
863 static inline void resched_task(task_t
*p
)
865 set_tsk_need_resched(p
);
870 * task_curr - is this task currently executing on a CPU?
871 * @p: the task in question.
873 inline int task_curr(const task_t
*p
)
875 return cpu_curr(task_cpu(p
)) == p
;
880 struct list_head list
;
885 struct completion done
;
889 * The task's runqueue lock must be held.
890 * Returns true if you have to wait for migration thread.
892 static int migrate_task(task_t
*p
, int dest_cpu
, migration_req_t
*req
)
894 runqueue_t
*rq
= task_rq(p
);
897 * If the task is not on a runqueue (and not running), then
898 * it is sufficient to simply update the task's cpu field.
900 if (!p
->array
&& !task_running(rq
, p
)) {
901 set_task_cpu(p
, dest_cpu
);
905 init_completion(&req
->done
);
907 req
->dest_cpu
= dest_cpu
;
908 list_add(&req
->list
, &rq
->migration_queue
);
913 * wait_task_inactive - wait for a thread to unschedule.
915 * The caller must ensure that the task *will* unschedule sometime soon,
916 * else this function might spin for a *long* time. This function can't
917 * be called with interrupts off, or it may introduce deadlock with
918 * smp_call_function() if an IPI is sent by the same process we are
919 * waiting to become inactive.
921 void wait_task_inactive(task_t
*p
)
928 rq
= task_rq_lock(p
, &flags
);
929 /* Must be off runqueue entirely, not preempted. */
930 if (unlikely(p
->array
|| task_running(rq
, p
))) {
931 /* If it's preempted, we yield. It could be a while. */
932 preempted
= !task_running(rq
, p
);
933 task_rq_unlock(rq
, &flags
);
939 task_rq_unlock(rq
, &flags
);
943 * kick_process - kick a running thread to enter/exit the kernel
944 * @p: the to-be-kicked thread
946 * Cause a process which is running on another CPU to enter
947 * kernel-mode, without any delay. (to get signals handled.)
949 * NOTE: this function doesnt have to take the runqueue lock,
950 * because all it wants to ensure is that the remote task enters
951 * the kernel. If the IPI races and the task has been migrated
952 * to another CPU then no harm is done and the purpose has been
955 void kick_process(task_t
*p
)
961 if ((cpu
!= smp_processor_id()) && task_curr(p
))
962 smp_send_reschedule(cpu
);
967 * Return a low guess at the load of a migration-source cpu.
969 * We want to under-estimate the load of migration sources, to
970 * balance conservatively.
972 static inline unsigned long __source_load(int cpu
, int type
, enum idle_type idle
)
974 runqueue_t
*rq
= cpu_rq(cpu
);
975 unsigned long source_load
, cpu_load
= rq
->cpu_load
[type
-1],
976 load_now
= rq
->nr_running
* SCHED_LOAD_SCALE
;
979 source_load
= load_now
;
981 source_load
= min(cpu_load
, load_now
);
983 if (idle
== NOT_IDLE
|| rq
->nr_running
> 1)
985 * If we are busy rebalancing the load is biased by
986 * priority to create 'nice' support across cpus. When
987 * idle rebalancing we should only bias the source_load if
988 * there is more than one task running on that queue to
989 * prevent idle rebalance from trying to pull tasks from a
990 * queue with only one running task.
992 source_load
*= rq
->prio_bias
;
997 static inline unsigned long source_load(int cpu
, int type
)
999 return __source_load(cpu
, type
, NOT_IDLE
);
1003 * Return a high guess at the load of a migration-target cpu
1005 static inline unsigned long __target_load(int cpu
, int type
, enum idle_type idle
)
1007 runqueue_t
*rq
= cpu_rq(cpu
);
1008 unsigned long target_load
, cpu_load
= rq
->cpu_load
[type
-1],
1009 load_now
= rq
->nr_running
* SCHED_LOAD_SCALE
;
1012 target_load
= load_now
;
1014 target_load
= max(cpu_load
, load_now
);
1016 if (idle
== NOT_IDLE
|| rq
->nr_running
> 1)
1017 target_load
*= rq
->prio_bias
;
1022 static inline unsigned long target_load(int cpu
, int type
)
1024 return __target_load(cpu
, type
, NOT_IDLE
);
1028 * find_idlest_group finds and returns the least busy CPU group within the
1031 static struct sched_group
*
1032 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1034 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1035 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1036 int load_idx
= sd
->forkexec_idx
;
1037 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1040 unsigned long load
, avg_load
;
1044 /* Skip over this group if it has no CPUs allowed */
1045 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1048 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1050 /* Tally up the load of all CPUs in the group */
1053 for_each_cpu_mask(i
, group
->cpumask
) {
1054 /* Bias balancing toward cpus of our domain */
1056 load
= source_load(i
, load_idx
);
1058 load
= target_load(i
, load_idx
);
1063 /* Adjust by relative CPU power of the group */
1064 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
1067 this_load
= avg_load
;
1069 } else if (avg_load
< min_load
) {
1070 min_load
= avg_load
;
1074 group
= group
->next
;
1075 } while (group
!= sd
->groups
);
1077 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1083 * find_idlest_queue - find the idlest runqueue among the cpus in group.
1086 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1089 unsigned long load
, min_load
= ULONG_MAX
;
1093 /* Traverse only the allowed CPUs */
1094 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1096 for_each_cpu_mask(i
, tmp
) {
1097 load
= source_load(i
, 0);
1099 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1109 * sched_balance_self: balance the current task (running on cpu) in domains
1110 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1113 * Balance, ie. select the least loaded group.
1115 * Returns the target CPU number, or the same CPU if no balancing is needed.
1117 * preempt must be disabled.
1119 static int sched_balance_self(int cpu
, int flag
)
1121 struct task_struct
*t
= current
;
1122 struct sched_domain
*tmp
, *sd
= NULL
;
1124 for_each_domain(cpu
, tmp
)
1125 if (tmp
->flags
& flag
)
1130 struct sched_group
*group
;
1135 group
= find_idlest_group(sd
, t
, cpu
);
1139 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1140 if (new_cpu
== -1 || new_cpu
== cpu
)
1143 /* Now try balancing at a lower domain level */
1147 weight
= cpus_weight(span
);
1148 for_each_domain(cpu
, tmp
) {
1149 if (weight
<= cpus_weight(tmp
->span
))
1151 if (tmp
->flags
& flag
)
1154 /* while loop will break here if sd == NULL */
1160 #endif /* CONFIG_SMP */
1163 * wake_idle() will wake a task on an idle cpu if task->cpu is
1164 * not idle and an idle cpu is available. The span of cpus to
1165 * search starts with cpus closest then further out as needed,
1166 * so we always favor a closer, idle cpu.
1168 * Returns the CPU we should wake onto.
1170 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1171 static int wake_idle(int cpu
, task_t
*p
)
1174 struct sched_domain
*sd
;
1180 for_each_domain(cpu
, sd
) {
1181 if (sd
->flags
& SD_WAKE_IDLE
) {
1182 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1183 for_each_cpu_mask(i
, tmp
) {
1194 static inline int wake_idle(int cpu
, task_t
*p
)
1201 * try_to_wake_up - wake up a thread
1202 * @p: the to-be-woken-up thread
1203 * @state: the mask of task states that can be woken
1204 * @sync: do a synchronous wakeup?
1206 * Put it on the run-queue if it's not already there. The "current"
1207 * thread is always on the run-queue (except when the actual
1208 * re-schedule is in progress), and as such you're allowed to do
1209 * the simpler "current->state = TASK_RUNNING" to mark yourself
1210 * runnable without the overhead of this.
1212 * returns failure only if the task is already active.
1214 static int try_to_wake_up(task_t
*p
, unsigned int state
, int sync
)
1216 int cpu
, this_cpu
, success
= 0;
1217 unsigned long flags
;
1221 unsigned long load
, this_load
;
1222 struct sched_domain
*sd
, *this_sd
= NULL
;
1226 rq
= task_rq_lock(p
, &flags
);
1227 old_state
= p
->state
;
1228 if (!(old_state
& state
))
1235 this_cpu
= smp_processor_id();
1238 if (unlikely(task_running(rq
, p
)))
1243 schedstat_inc(rq
, ttwu_cnt
);
1244 if (cpu
== this_cpu
) {
1245 schedstat_inc(rq
, ttwu_local
);
1249 for_each_domain(this_cpu
, sd
) {
1250 if (cpu_isset(cpu
, sd
->span
)) {
1251 schedstat_inc(sd
, ttwu_wake_remote
);
1257 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1261 * Check for affine wakeup and passive balancing possibilities.
1264 int idx
= this_sd
->wake_idx
;
1265 unsigned int imbalance
;
1267 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1269 load
= source_load(cpu
, idx
);
1270 this_load
= target_load(this_cpu
, idx
);
1272 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1274 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1275 unsigned long tl
= this_load
;
1277 * If sync wakeup then subtract the (maximum possible)
1278 * effect of the currently running task from the load
1279 * of the current CPU:
1282 tl
-= SCHED_LOAD_SCALE
;
1285 tl
+ target_load(cpu
, idx
) <= SCHED_LOAD_SCALE
) ||
1286 100*(tl
+ SCHED_LOAD_SCALE
) <= imbalance
*load
) {
1288 * This domain has SD_WAKE_AFFINE and
1289 * p is cache cold in this domain, and
1290 * there is no bad imbalance.
1292 schedstat_inc(this_sd
, ttwu_move_affine
);
1298 * Start passive balancing when half the imbalance_pct
1301 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1302 if (imbalance
*this_load
<= 100*load
) {
1303 schedstat_inc(this_sd
, ttwu_move_balance
);
1309 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1311 new_cpu
= wake_idle(new_cpu
, p
);
1312 if (new_cpu
!= cpu
) {
1313 set_task_cpu(p
, new_cpu
);
1314 task_rq_unlock(rq
, &flags
);
1315 /* might preempt at this point */
1316 rq
= task_rq_lock(p
, &flags
);
1317 old_state
= p
->state
;
1318 if (!(old_state
& state
))
1323 this_cpu
= smp_processor_id();
1328 #endif /* CONFIG_SMP */
1329 if (old_state
== TASK_UNINTERRUPTIBLE
) {
1330 rq
->nr_uninterruptible
--;
1332 * Tasks on involuntary sleep don't earn
1333 * sleep_avg beyond just interactive state.
1339 * Tasks that have marked their sleep as noninteractive get
1340 * woken up without updating their sleep average. (i.e. their
1341 * sleep is handled in a priority-neutral manner, no priority
1342 * boost and no penalty.)
1344 if (old_state
& TASK_NONINTERACTIVE
)
1345 __activate_task(p
, rq
);
1347 activate_task(p
, rq
, cpu
== this_cpu
);
1349 * Sync wakeups (i.e. those types of wakeups where the waker
1350 * has indicated that it will leave the CPU in short order)
1351 * don't trigger a preemption, if the woken up task will run on
1352 * this cpu. (in this case the 'I will reschedule' promise of
1353 * the waker guarantees that the freshly woken up task is going
1354 * to be considered on this CPU.)
1356 if (!sync
|| cpu
!= this_cpu
) {
1357 if (TASK_PREEMPTS_CURR(p
, rq
))
1358 resched_task(rq
->curr
);
1363 p
->state
= TASK_RUNNING
;
1365 task_rq_unlock(rq
, &flags
);
1370 int fastcall
wake_up_process(task_t
*p
)
1372 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1373 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1376 EXPORT_SYMBOL(wake_up_process
);
1378 int fastcall
wake_up_state(task_t
*p
, unsigned int state
)
1380 return try_to_wake_up(p
, state
, 0);
1384 * Perform scheduler related setup for a newly forked process p.
1385 * p is forked by current.
1387 void fastcall
sched_fork(task_t
*p
, int clone_flags
)
1389 int cpu
= get_cpu();
1392 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1394 set_task_cpu(p
, cpu
);
1397 * We mark the process as running here, but have not actually
1398 * inserted it onto the runqueue yet. This guarantees that
1399 * nobody will actually run it, and a signal or other external
1400 * event cannot wake it up and insert it on the runqueue either.
1402 p
->state
= TASK_RUNNING
;
1403 INIT_LIST_HEAD(&p
->run_list
);
1405 #ifdef CONFIG_SCHEDSTATS
1406 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1408 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1411 #ifdef CONFIG_PREEMPT
1412 /* Want to start with kernel preemption disabled. */
1413 p
->thread_info
->preempt_count
= 1;
1416 * Share the timeslice between parent and child, thus the
1417 * total amount of pending timeslices in the system doesn't change,
1418 * resulting in more scheduling fairness.
1420 local_irq_disable();
1421 p
->time_slice
= (current
->time_slice
+ 1) >> 1;
1423 * The remainder of the first timeslice might be recovered by
1424 * the parent if the child exits early enough.
1426 p
->first_time_slice
= 1;
1427 current
->time_slice
>>= 1;
1428 p
->timestamp
= sched_clock();
1429 if (unlikely(!current
->time_slice
)) {
1431 * This case is rare, it happens when the parent has only
1432 * a single jiffy left from its timeslice. Taking the
1433 * runqueue lock is not a problem.
1435 current
->time_slice
= 1;
1443 * wake_up_new_task - wake up a newly created task for the first time.
1445 * This function will do some initial scheduler statistics housekeeping
1446 * that must be done for every newly created context, then puts the task
1447 * on the runqueue and wakes it.
1449 void fastcall
wake_up_new_task(task_t
*p
, unsigned long clone_flags
)
1451 unsigned long flags
;
1453 runqueue_t
*rq
, *this_rq
;
1455 rq
= task_rq_lock(p
, &flags
);
1456 BUG_ON(p
->state
!= TASK_RUNNING
);
1457 this_cpu
= smp_processor_id();
1461 * We decrease the sleep average of forking parents
1462 * and children as well, to keep max-interactive tasks
1463 * from forking tasks that are max-interactive. The parent
1464 * (current) is done further down, under its lock.
1466 p
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(p
) *
1467 CHILD_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1469 p
->prio
= effective_prio(p
);
1471 if (likely(cpu
== this_cpu
)) {
1472 if (!(clone_flags
& CLONE_VM
)) {
1474 * The VM isn't cloned, so we're in a good position to
1475 * do child-runs-first in anticipation of an exec. This
1476 * usually avoids a lot of COW overhead.
1478 if (unlikely(!current
->array
))
1479 __activate_task(p
, rq
);
1481 p
->prio
= current
->prio
;
1482 list_add_tail(&p
->run_list
, ¤t
->run_list
);
1483 p
->array
= current
->array
;
1484 p
->array
->nr_active
++;
1485 inc_nr_running(p
, rq
);
1489 /* Run child last */
1490 __activate_task(p
, rq
);
1492 * We skip the following code due to cpu == this_cpu
1494 * task_rq_unlock(rq, &flags);
1495 * this_rq = task_rq_lock(current, &flags);
1499 this_rq
= cpu_rq(this_cpu
);
1502 * Not the local CPU - must adjust timestamp. This should
1503 * get optimised away in the !CONFIG_SMP case.
1505 p
->timestamp
= (p
->timestamp
- this_rq
->timestamp_last_tick
)
1506 + rq
->timestamp_last_tick
;
1507 __activate_task(p
, rq
);
1508 if (TASK_PREEMPTS_CURR(p
, rq
))
1509 resched_task(rq
->curr
);
1512 * Parent and child are on different CPUs, now get the
1513 * parent runqueue to update the parent's ->sleep_avg:
1515 task_rq_unlock(rq
, &flags
);
1516 this_rq
= task_rq_lock(current
, &flags
);
1518 current
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(current
) *
1519 PARENT_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1520 task_rq_unlock(this_rq
, &flags
);
1524 * Potentially available exiting-child timeslices are
1525 * retrieved here - this way the parent does not get
1526 * penalized for creating too many threads.
1528 * (this cannot be used to 'generate' timeslices
1529 * artificially, because any timeslice recovered here
1530 * was given away by the parent in the first place.)
1532 void fastcall
sched_exit(task_t
*p
)
1534 unsigned long flags
;
1538 * If the child was a (relative-) CPU hog then decrease
1539 * the sleep_avg of the parent as well.
1541 rq
= task_rq_lock(p
->parent
, &flags
);
1542 if (p
->first_time_slice
&& task_cpu(p
) == task_cpu(p
->parent
)) {
1543 p
->parent
->time_slice
+= p
->time_slice
;
1544 if (unlikely(p
->parent
->time_slice
> task_timeslice(p
)))
1545 p
->parent
->time_slice
= task_timeslice(p
);
1547 if (p
->sleep_avg
< p
->parent
->sleep_avg
)
1548 p
->parent
->sleep_avg
= p
->parent
->sleep_avg
/
1549 (EXIT_WEIGHT
+ 1) * EXIT_WEIGHT
+ p
->sleep_avg
/
1551 task_rq_unlock(rq
, &flags
);
1555 * prepare_task_switch - prepare to switch tasks
1556 * @rq: the runqueue preparing to switch
1557 * @next: the task we are going to switch to.
1559 * This is called with the rq lock held and interrupts off. It must
1560 * be paired with a subsequent finish_task_switch after the context
1563 * prepare_task_switch sets up locking and calls architecture specific
1566 static inline void prepare_task_switch(runqueue_t
*rq
, task_t
*next
)
1568 prepare_lock_switch(rq
, next
);
1569 prepare_arch_switch(next
);
1573 * finish_task_switch - clean up after a task-switch
1574 * @rq: runqueue associated with task-switch
1575 * @prev: the thread we just switched away from.
1577 * finish_task_switch must be called after the context switch, paired
1578 * with a prepare_task_switch call before the context switch.
1579 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1580 * and do any other architecture-specific cleanup actions.
1582 * Note that we may have delayed dropping an mm in context_switch(). If
1583 * so, we finish that here outside of the runqueue lock. (Doing it
1584 * with the lock held can cause deadlocks; see schedule() for
1587 static inline void finish_task_switch(runqueue_t
*rq
, task_t
*prev
)
1588 __releases(rq
->lock
)
1590 struct mm_struct
*mm
= rq
->prev_mm
;
1591 unsigned long prev_task_flags
;
1596 * A task struct has one reference for the use as "current".
1597 * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1598 * calls schedule one last time. The schedule call will never return,
1599 * and the scheduled task must drop that reference.
1600 * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1601 * still held, otherwise prev could be scheduled on another cpu, die
1602 * there before we look at prev->state, and then the reference would
1604 * Manfred Spraul <manfred@colorfullife.com>
1606 prev_task_flags
= prev
->flags
;
1607 finish_arch_switch(prev
);
1608 finish_lock_switch(rq
, prev
);
1611 if (unlikely(prev_task_flags
& PF_DEAD
))
1612 put_task_struct(prev
);
1616 * schedule_tail - first thing a freshly forked thread must call.
1617 * @prev: the thread we just switched away from.
1619 asmlinkage
void schedule_tail(task_t
*prev
)
1620 __releases(rq
->lock
)
1622 runqueue_t
*rq
= this_rq();
1623 finish_task_switch(rq
, prev
);
1624 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1625 /* In this case, finish_task_switch does not reenable preemption */
1628 if (current
->set_child_tid
)
1629 put_user(current
->pid
, current
->set_child_tid
);
1633 * context_switch - switch to the new MM and the new
1634 * thread's register state.
1637 task_t
* context_switch(runqueue_t
*rq
, task_t
*prev
, task_t
*next
)
1639 struct mm_struct
*mm
= next
->mm
;
1640 struct mm_struct
*oldmm
= prev
->active_mm
;
1642 if (unlikely(!mm
)) {
1643 next
->active_mm
= oldmm
;
1644 atomic_inc(&oldmm
->mm_count
);
1645 enter_lazy_tlb(oldmm
, next
);
1647 switch_mm(oldmm
, mm
, next
);
1649 if (unlikely(!prev
->mm
)) {
1650 prev
->active_mm
= NULL
;
1651 WARN_ON(rq
->prev_mm
);
1652 rq
->prev_mm
= oldmm
;
1655 /* Here we just switch the register state and the stack. */
1656 switch_to(prev
, next
, prev
);
1662 * nr_running, nr_uninterruptible and nr_context_switches:
1664 * externally visible scheduler statistics: current number of runnable
1665 * threads, current number of uninterruptible-sleeping threads, total
1666 * number of context switches performed since bootup.
1668 unsigned long nr_running(void)
1670 unsigned long i
, sum
= 0;
1672 for_each_online_cpu(i
)
1673 sum
+= cpu_rq(i
)->nr_running
;
1678 unsigned long nr_uninterruptible(void)
1680 unsigned long i
, sum
= 0;
1683 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1686 * Since we read the counters lockless, it might be slightly
1687 * inaccurate. Do not allow it to go below zero though:
1689 if (unlikely((long)sum
< 0))
1695 unsigned long long nr_context_switches(void)
1697 unsigned long long i
, sum
= 0;
1700 sum
+= cpu_rq(i
)->nr_switches
;
1705 unsigned long nr_iowait(void)
1707 unsigned long i
, sum
= 0;
1710 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1718 * double_rq_lock - safely lock two runqueues
1720 * Note this does not disable interrupts like task_rq_lock,
1721 * you need to do so manually before calling.
1723 static void double_rq_lock(runqueue_t
*rq1
, runqueue_t
*rq2
)
1724 __acquires(rq1
->lock
)
1725 __acquires(rq2
->lock
)
1728 spin_lock(&rq1
->lock
);
1729 __acquire(rq2
->lock
); /* Fake it out ;) */
1732 spin_lock(&rq1
->lock
);
1733 spin_lock(&rq2
->lock
);
1735 spin_lock(&rq2
->lock
);
1736 spin_lock(&rq1
->lock
);
1742 * double_rq_unlock - safely unlock two runqueues
1744 * Note this does not restore interrupts like task_rq_unlock,
1745 * you need to do so manually after calling.
1747 static void double_rq_unlock(runqueue_t
*rq1
, runqueue_t
*rq2
)
1748 __releases(rq1
->lock
)
1749 __releases(rq2
->lock
)
1751 spin_unlock(&rq1
->lock
);
1753 spin_unlock(&rq2
->lock
);
1755 __release(rq2
->lock
);
1759 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1761 static void double_lock_balance(runqueue_t
*this_rq
, runqueue_t
*busiest
)
1762 __releases(this_rq
->lock
)
1763 __acquires(busiest
->lock
)
1764 __acquires(this_rq
->lock
)
1766 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1767 if (busiest
< this_rq
) {
1768 spin_unlock(&this_rq
->lock
);
1769 spin_lock(&busiest
->lock
);
1770 spin_lock(&this_rq
->lock
);
1772 spin_lock(&busiest
->lock
);
1777 * If dest_cpu is allowed for this process, migrate the task to it.
1778 * This is accomplished by forcing the cpu_allowed mask to only
1779 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
1780 * the cpu_allowed mask is restored.
1782 static void sched_migrate_task(task_t
*p
, int dest_cpu
)
1784 migration_req_t req
;
1786 unsigned long flags
;
1788 rq
= task_rq_lock(p
, &flags
);
1789 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
1790 || unlikely(cpu_is_offline(dest_cpu
)))
1793 /* force the process onto the specified CPU */
1794 if (migrate_task(p
, dest_cpu
, &req
)) {
1795 /* Need to wait for migration thread (might exit: take ref). */
1796 struct task_struct
*mt
= rq
->migration_thread
;
1797 get_task_struct(mt
);
1798 task_rq_unlock(rq
, &flags
);
1799 wake_up_process(mt
);
1800 put_task_struct(mt
);
1801 wait_for_completion(&req
.done
);
1805 task_rq_unlock(rq
, &flags
);
1809 * sched_exec - execve() is a valuable balancing opportunity, because at
1810 * this point the task has the smallest effective memory and cache footprint.
1812 void sched_exec(void)
1814 int new_cpu
, this_cpu
= get_cpu();
1815 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
1817 if (new_cpu
!= this_cpu
)
1818 sched_migrate_task(current
, new_cpu
);
1822 * pull_task - move a task from a remote runqueue to the local runqueue.
1823 * Both runqueues must be locked.
1826 void pull_task(runqueue_t
*src_rq
, prio_array_t
*src_array
, task_t
*p
,
1827 runqueue_t
*this_rq
, prio_array_t
*this_array
, int this_cpu
)
1829 dequeue_task(p
, src_array
);
1830 dec_nr_running(p
, src_rq
);
1831 set_task_cpu(p
, this_cpu
);
1832 inc_nr_running(p
, this_rq
);
1833 enqueue_task(p
, this_array
);
1834 p
->timestamp
= (p
->timestamp
- src_rq
->timestamp_last_tick
)
1835 + this_rq
->timestamp_last_tick
;
1837 * Note that idle threads have a prio of MAX_PRIO, for this test
1838 * to be always true for them.
1840 if (TASK_PREEMPTS_CURR(p
, this_rq
))
1841 resched_task(this_rq
->curr
);
1845 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1848 int can_migrate_task(task_t
*p
, runqueue_t
*rq
, int this_cpu
,
1849 struct sched_domain
*sd
, enum idle_type idle
,
1853 * We do not migrate tasks that are:
1854 * 1) running (obviously), or
1855 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1856 * 3) are cache-hot on their current CPU.
1858 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
1862 if (task_running(rq
, p
))
1866 * Aggressive migration if:
1867 * 1) task is cache cold, or
1868 * 2) too many balance attempts have failed.
1871 if (sd
->nr_balance_failed
> sd
->cache_nice_tries
)
1874 if (task_hot(p
, rq
->timestamp_last_tick
, sd
))
1880 * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq,
1881 * as part of a balancing operation within "domain". Returns the number of
1884 * Called with both runqueues locked.
1886 static int move_tasks(runqueue_t
*this_rq
, int this_cpu
, runqueue_t
*busiest
,
1887 unsigned long max_nr_move
, struct sched_domain
*sd
,
1888 enum idle_type idle
, int *all_pinned
)
1890 prio_array_t
*array
, *dst_array
;
1891 struct list_head
*head
, *curr
;
1892 int idx
, pulled
= 0, pinned
= 0;
1895 if (max_nr_move
== 0)
1901 * We first consider expired tasks. Those will likely not be
1902 * executed in the near future, and they are most likely to
1903 * be cache-cold, thus switching CPUs has the least effect
1906 if (busiest
->expired
->nr_active
) {
1907 array
= busiest
->expired
;
1908 dst_array
= this_rq
->expired
;
1910 array
= busiest
->active
;
1911 dst_array
= this_rq
->active
;
1915 /* Start searching at priority 0: */
1919 idx
= sched_find_first_bit(array
->bitmap
);
1921 idx
= find_next_bit(array
->bitmap
, MAX_PRIO
, idx
);
1922 if (idx
>= MAX_PRIO
) {
1923 if (array
== busiest
->expired
&& busiest
->active
->nr_active
) {
1924 array
= busiest
->active
;
1925 dst_array
= this_rq
->active
;
1931 head
= array
->queue
+ idx
;
1934 tmp
= list_entry(curr
, task_t
, run_list
);
1938 if (!can_migrate_task(tmp
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
1945 #ifdef CONFIG_SCHEDSTATS
1946 if (task_hot(tmp
, busiest
->timestamp_last_tick
, sd
))
1947 schedstat_inc(sd
, lb_hot_gained
[idle
]);
1950 pull_task(busiest
, array
, tmp
, this_rq
, dst_array
, this_cpu
);
1953 /* We only want to steal up to the prescribed number of tasks. */
1954 if (pulled
< max_nr_move
) {
1962 * Right now, this is the only place pull_task() is called,
1963 * so we can safely collect pull_task() stats here rather than
1964 * inside pull_task().
1966 schedstat_add(sd
, lb_gained
[idle
], pulled
);
1969 *all_pinned
= pinned
;
1974 * find_busiest_group finds and returns the busiest CPU group within the
1975 * domain. It calculates and returns the number of tasks which should be
1976 * moved to restore balance via the imbalance parameter.
1978 static struct sched_group
*
1979 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
1980 unsigned long *imbalance
, enum idle_type idle
, int *sd_idle
)
1982 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1983 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
1984 unsigned long max_pull
;
1987 max_load
= this_load
= total_load
= total_pwr
= 0;
1988 if (idle
== NOT_IDLE
)
1989 load_idx
= sd
->busy_idx
;
1990 else if (idle
== NEWLY_IDLE
)
1991 load_idx
= sd
->newidle_idx
;
1993 load_idx
= sd
->idle_idx
;
2000 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2002 /* Tally up the load of all CPUs in the group */
2005 for_each_cpu_mask(i
, group
->cpumask
) {
2006 if (*sd_idle
&& !idle_cpu(i
))
2009 /* Bias balancing toward cpus of our domain */
2011 load
= __target_load(i
, load_idx
, idle
);
2013 load
= __source_load(i
, load_idx
, idle
);
2018 total_load
+= avg_load
;
2019 total_pwr
+= group
->cpu_power
;
2021 /* Adjust by relative CPU power of the group */
2022 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
2025 this_load
= avg_load
;
2027 } else if (avg_load
> max_load
) {
2028 max_load
= avg_load
;
2031 group
= group
->next
;
2032 } while (group
!= sd
->groups
);
2034 if (!busiest
|| this_load
>= max_load
|| max_load
<= SCHED_LOAD_SCALE
)
2037 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2039 if (this_load
>= avg_load
||
2040 100*max_load
<= sd
->imbalance_pct
*this_load
)
2044 * We're trying to get all the cpus to the average_load, so we don't
2045 * want to push ourselves above the average load, nor do we wish to
2046 * reduce the max loaded cpu below the average load, as either of these
2047 * actions would just result in more rebalancing later, and ping-pong
2048 * tasks around. Thus we look for the minimum possible imbalance.
2049 * Negative imbalances (*we* are more loaded than anyone else) will
2050 * be counted as no imbalance for these purposes -- we can't fix that
2051 * by pulling tasks to us. Be careful of negative numbers as they'll
2052 * appear as very large values with unsigned longs.
2055 /* Don't want to pull so many tasks that a group would go idle */
2056 max_pull
= min(max_load
- avg_load
, max_load
- SCHED_LOAD_SCALE
);
2058 /* How much load to actually move to equalise the imbalance */
2059 *imbalance
= min(max_pull
* busiest
->cpu_power
,
2060 (avg_load
- this_load
) * this->cpu_power
)
2063 if (*imbalance
< SCHED_LOAD_SCALE
) {
2064 unsigned long pwr_now
= 0, pwr_move
= 0;
2067 if (max_load
- this_load
>= SCHED_LOAD_SCALE
*2) {
2073 * OK, we don't have enough imbalance to justify moving tasks,
2074 * however we may be able to increase total CPU power used by
2078 pwr_now
+= busiest
->cpu_power
*min(SCHED_LOAD_SCALE
, max_load
);
2079 pwr_now
+= this->cpu_power
*min(SCHED_LOAD_SCALE
, this_load
);
2080 pwr_now
/= SCHED_LOAD_SCALE
;
2082 /* Amount of load we'd subtract */
2083 tmp
= SCHED_LOAD_SCALE
*SCHED_LOAD_SCALE
/busiest
->cpu_power
;
2085 pwr_move
+= busiest
->cpu_power
*min(SCHED_LOAD_SCALE
,
2088 /* Amount of load we'd add */
2089 if (max_load
*busiest
->cpu_power
<
2090 SCHED_LOAD_SCALE
*SCHED_LOAD_SCALE
)
2091 tmp
= max_load
*busiest
->cpu_power
/this->cpu_power
;
2093 tmp
= SCHED_LOAD_SCALE
*SCHED_LOAD_SCALE
/this->cpu_power
;
2094 pwr_move
+= this->cpu_power
*min(SCHED_LOAD_SCALE
, this_load
+ tmp
);
2095 pwr_move
/= SCHED_LOAD_SCALE
;
2097 /* Move if we gain throughput */
2098 if (pwr_move
<= pwr_now
)
2105 /* Get rid of the scaling factor, rounding down as we divide */
2106 *imbalance
= *imbalance
/ SCHED_LOAD_SCALE
;
2116 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2118 static runqueue_t
*find_busiest_queue(struct sched_group
*group
,
2119 enum idle_type idle
)
2121 unsigned long load
, max_load
= 0;
2122 runqueue_t
*busiest
= NULL
;
2125 for_each_cpu_mask(i
, group
->cpumask
) {
2126 load
= __source_load(i
, 0, idle
);
2128 if (load
> max_load
) {
2130 busiest
= cpu_rq(i
);
2138 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2139 * so long as it is large enough.
2141 #define MAX_PINNED_INTERVAL 512
2144 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2145 * tasks if there is an imbalance.
2147 * Called with this_rq unlocked.
2149 static int load_balance(int this_cpu
, runqueue_t
*this_rq
,
2150 struct sched_domain
*sd
, enum idle_type idle
)
2152 struct sched_group
*group
;
2153 runqueue_t
*busiest
;
2154 unsigned long imbalance
;
2155 int nr_moved
, all_pinned
= 0;
2156 int active_balance
= 0;
2159 if (idle
!= NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2162 schedstat_inc(sd
, lb_cnt
[idle
]);
2164 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
);
2166 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2170 busiest
= find_busiest_queue(group
, idle
);
2172 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2176 BUG_ON(busiest
== this_rq
);
2178 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2181 if (busiest
->nr_running
> 1) {
2183 * Attempt to move tasks. If find_busiest_group has found
2184 * an imbalance but busiest->nr_running <= 1, the group is
2185 * still unbalanced. nr_moved simply stays zero, so it is
2186 * correctly treated as an imbalance.
2188 double_rq_lock(this_rq
, busiest
);
2189 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2190 imbalance
, sd
, idle
, &all_pinned
);
2191 double_rq_unlock(this_rq
, busiest
);
2193 /* All tasks on this runqueue were pinned by CPU affinity */
2194 if (unlikely(all_pinned
))
2199 schedstat_inc(sd
, lb_failed
[idle
]);
2200 sd
->nr_balance_failed
++;
2202 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2204 spin_lock(&busiest
->lock
);
2206 /* don't kick the migration_thread, if the curr
2207 * task on busiest cpu can't be moved to this_cpu
2209 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2210 spin_unlock(&busiest
->lock
);
2212 goto out_one_pinned
;
2215 if (!busiest
->active_balance
) {
2216 busiest
->active_balance
= 1;
2217 busiest
->push_cpu
= this_cpu
;
2220 spin_unlock(&busiest
->lock
);
2222 wake_up_process(busiest
->migration_thread
);
2225 * We've kicked active balancing, reset the failure
2228 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2231 sd
->nr_balance_failed
= 0;
2233 if (likely(!active_balance
)) {
2234 /* We were unbalanced, so reset the balancing interval */
2235 sd
->balance_interval
= sd
->min_interval
;
2238 * If we've begun active balancing, start to back off. This
2239 * case may not be covered by the all_pinned logic if there
2240 * is only 1 task on the busy runqueue (because we don't call
2243 if (sd
->balance_interval
< sd
->max_interval
)
2244 sd
->balance_interval
*= 2;
2247 if (!nr_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2252 schedstat_inc(sd
, lb_balanced
[idle
]);
2254 sd
->nr_balance_failed
= 0;
2257 /* tune up the balancing interval */
2258 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2259 (sd
->balance_interval
< sd
->max_interval
))
2260 sd
->balance_interval
*= 2;
2262 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2268 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2269 * tasks if there is an imbalance.
2271 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2272 * this_rq is locked.
2274 static int load_balance_newidle(int this_cpu
, runqueue_t
*this_rq
,
2275 struct sched_domain
*sd
)
2277 struct sched_group
*group
;
2278 runqueue_t
*busiest
= NULL
;
2279 unsigned long imbalance
;
2283 if (sd
->flags
& SD_SHARE_CPUPOWER
)
2286 schedstat_inc(sd
, lb_cnt
[NEWLY_IDLE
]);
2287 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, NEWLY_IDLE
, &sd_idle
);
2289 schedstat_inc(sd
, lb_nobusyg
[NEWLY_IDLE
]);
2293 busiest
= find_busiest_queue(group
, NEWLY_IDLE
);
2295 schedstat_inc(sd
, lb_nobusyq
[NEWLY_IDLE
]);
2299 BUG_ON(busiest
== this_rq
);
2301 schedstat_add(sd
, lb_imbalance
[NEWLY_IDLE
], imbalance
);
2304 if (busiest
->nr_running
> 1) {
2305 /* Attempt to move tasks */
2306 double_lock_balance(this_rq
, busiest
);
2307 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2308 imbalance
, sd
, NEWLY_IDLE
, NULL
);
2309 spin_unlock(&busiest
->lock
);
2313 schedstat_inc(sd
, lb_failed
[NEWLY_IDLE
]);
2314 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2317 sd
->nr_balance_failed
= 0;
2322 schedstat_inc(sd
, lb_balanced
[NEWLY_IDLE
]);
2323 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2325 sd
->nr_balance_failed
= 0;
2330 * idle_balance is called by schedule() if this_cpu is about to become
2331 * idle. Attempts to pull tasks from other CPUs.
2333 static inline void idle_balance(int this_cpu
, runqueue_t
*this_rq
)
2335 struct sched_domain
*sd
;
2337 for_each_domain(this_cpu
, sd
) {
2338 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
2339 if (load_balance_newidle(this_cpu
, this_rq
, sd
)) {
2340 /* We've pulled tasks over so stop searching */
2348 * active_load_balance is run by migration threads. It pushes running tasks
2349 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2350 * running on each physical CPU where possible, and avoids physical /
2351 * logical imbalances.
2353 * Called with busiest_rq locked.
2355 static void active_load_balance(runqueue_t
*busiest_rq
, int busiest_cpu
)
2357 struct sched_domain
*sd
;
2358 runqueue_t
*target_rq
;
2359 int target_cpu
= busiest_rq
->push_cpu
;
2361 if (busiest_rq
->nr_running
<= 1)
2362 /* no task to move */
2365 target_rq
= cpu_rq(target_cpu
);
2368 * This condition is "impossible", if it occurs
2369 * we need to fix it. Originally reported by
2370 * Bjorn Helgaas on a 128-cpu setup.
2372 BUG_ON(busiest_rq
== target_rq
);
2374 /* move a task from busiest_rq to target_rq */
2375 double_lock_balance(busiest_rq
, target_rq
);
2377 /* Search for an sd spanning us and the target CPU. */
2378 for_each_domain(target_cpu
, sd
)
2379 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2380 cpu_isset(busiest_cpu
, sd
->span
))
2383 if (unlikely(sd
== NULL
))
2386 schedstat_inc(sd
, alb_cnt
);
2388 if (move_tasks(target_rq
, target_cpu
, busiest_rq
, 1, sd
, SCHED_IDLE
, NULL
))
2389 schedstat_inc(sd
, alb_pushed
);
2391 schedstat_inc(sd
, alb_failed
);
2393 spin_unlock(&target_rq
->lock
);
2397 * rebalance_tick will get called every timer tick, on every CPU.
2399 * It checks each scheduling domain to see if it is due to be balanced,
2400 * and initiates a balancing operation if so.
2402 * Balancing parameters are set up in arch_init_sched_domains.
2405 /* Don't have all balancing operations going off at once */
2406 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2408 static void rebalance_tick(int this_cpu
, runqueue_t
*this_rq
,
2409 enum idle_type idle
)
2411 unsigned long old_load
, this_load
;
2412 unsigned long j
= jiffies
+ CPU_OFFSET(this_cpu
);
2413 struct sched_domain
*sd
;
2416 this_load
= this_rq
->nr_running
* SCHED_LOAD_SCALE
;
2417 /* Update our load */
2418 for (i
= 0; i
< 3; i
++) {
2419 unsigned long new_load
= this_load
;
2421 old_load
= this_rq
->cpu_load
[i
];
2423 * Round up the averaging division if load is increasing. This
2424 * prevents us from getting stuck on 9 if the load is 10, for
2427 if (new_load
> old_load
)
2428 new_load
+= scale
-1;
2429 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) / scale
;
2432 for_each_domain(this_cpu
, sd
) {
2433 unsigned long interval
;
2435 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2438 interval
= sd
->balance_interval
;
2439 if (idle
!= SCHED_IDLE
)
2440 interval
*= sd
->busy_factor
;
2442 /* scale ms to jiffies */
2443 interval
= msecs_to_jiffies(interval
);
2444 if (unlikely(!interval
))
2447 if (j
- sd
->last_balance
>= interval
) {
2448 if (load_balance(this_cpu
, this_rq
, sd
, idle
)) {
2450 * We've pulled tasks over so either we're no
2451 * longer idle, or one of our SMT siblings is
2456 sd
->last_balance
+= interval
;
2462 * on UP we do not need to balance between CPUs:
2464 static inline void rebalance_tick(int cpu
, runqueue_t
*rq
, enum idle_type idle
)
2467 static inline void idle_balance(int cpu
, runqueue_t
*rq
)
2472 static inline int wake_priority_sleeper(runqueue_t
*rq
)
2475 #ifdef CONFIG_SCHED_SMT
2476 spin_lock(&rq
->lock
);
2478 * If an SMT sibling task has been put to sleep for priority
2479 * reasons reschedule the idle task to see if it can now run.
2481 if (rq
->nr_running
) {
2482 resched_task(rq
->idle
);
2485 spin_unlock(&rq
->lock
);
2490 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2492 EXPORT_PER_CPU_SYMBOL(kstat
);
2495 * This is called on clock ticks and on context switches.
2496 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2498 static inline void update_cpu_clock(task_t
*p
, runqueue_t
*rq
,
2499 unsigned long long now
)
2501 unsigned long long last
= max(p
->timestamp
, rq
->timestamp_last_tick
);
2502 p
->sched_time
+= now
- last
;
2506 * Return current->sched_time plus any more ns on the sched_clock
2507 * that have not yet been banked.
2509 unsigned long long current_sched_time(const task_t
*tsk
)
2511 unsigned long long ns
;
2512 unsigned long flags
;
2513 local_irq_save(flags
);
2514 ns
= max(tsk
->timestamp
, task_rq(tsk
)->timestamp_last_tick
);
2515 ns
= tsk
->sched_time
+ (sched_clock() - ns
);
2516 local_irq_restore(flags
);
2521 * We place interactive tasks back into the active array, if possible.
2523 * To guarantee that this does not starve expired tasks we ignore the
2524 * interactivity of a task if the first expired task had to wait more
2525 * than a 'reasonable' amount of time. This deadline timeout is
2526 * load-dependent, as the frequency of array switched decreases with
2527 * increasing number of running tasks. We also ignore the interactivity
2528 * if a better static_prio task has expired:
2530 #define EXPIRED_STARVING(rq) \
2531 ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2532 (jiffies - (rq)->expired_timestamp >= \
2533 STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2534 ((rq)->curr->static_prio > (rq)->best_expired_prio))
2537 * Account user cpu time to a process.
2538 * @p: the process that the cpu time gets accounted to
2539 * @hardirq_offset: the offset to subtract from hardirq_count()
2540 * @cputime: the cpu time spent in user space since the last update
2542 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
2544 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2547 p
->utime
= cputime_add(p
->utime
, cputime
);
2549 /* Add user time to cpustat. */
2550 tmp
= cputime_to_cputime64(cputime
);
2551 if (TASK_NICE(p
) > 0)
2552 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
2554 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
2558 * Account system cpu time to a process.
2559 * @p: the process that the cpu time gets accounted to
2560 * @hardirq_offset: the offset to subtract from hardirq_count()
2561 * @cputime: the cpu time spent in kernel space since the last update
2563 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
2566 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2567 runqueue_t
*rq
= this_rq();
2570 p
->stime
= cputime_add(p
->stime
, cputime
);
2572 /* Add system time to cpustat. */
2573 tmp
= cputime_to_cputime64(cputime
);
2574 if (hardirq_count() - hardirq_offset
)
2575 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
2576 else if (softirq_count())
2577 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
2578 else if (p
!= rq
->idle
)
2579 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
2580 else if (atomic_read(&rq
->nr_iowait
) > 0)
2581 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
2583 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
2584 /* Account for system time used */
2585 acct_update_integrals(p
);
2589 * Account for involuntary wait time.
2590 * @p: the process from which the cpu time has been stolen
2591 * @steal: the cpu time spent in involuntary wait
2593 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
2595 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2596 cputime64_t tmp
= cputime_to_cputime64(steal
);
2597 runqueue_t
*rq
= this_rq();
2599 if (p
== rq
->idle
) {
2600 p
->stime
= cputime_add(p
->stime
, steal
);
2601 if (atomic_read(&rq
->nr_iowait
) > 0)
2602 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
2604 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
2606 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
2610 * This function gets called by the timer code, with HZ frequency.
2611 * We call it with interrupts disabled.
2613 * It also gets called by the fork code, when changing the parent's
2616 void scheduler_tick(void)
2618 int cpu
= smp_processor_id();
2619 runqueue_t
*rq
= this_rq();
2620 task_t
*p
= current
;
2621 unsigned long long now
= sched_clock();
2623 update_cpu_clock(p
, rq
, now
);
2625 rq
->timestamp_last_tick
= now
;
2627 if (p
== rq
->idle
) {
2628 if (wake_priority_sleeper(rq
))
2630 rebalance_tick(cpu
, rq
, SCHED_IDLE
);
2634 /* Task might have expired already, but not scheduled off yet */
2635 if (p
->array
!= rq
->active
) {
2636 set_tsk_need_resched(p
);
2639 spin_lock(&rq
->lock
);
2641 * The task was running during this tick - update the
2642 * time slice counter. Note: we do not update a thread's
2643 * priority until it either goes to sleep or uses up its
2644 * timeslice. This makes it possible for interactive tasks
2645 * to use up their timeslices at their highest priority levels.
2649 * RR tasks need a special form of timeslice management.
2650 * FIFO tasks have no timeslices.
2652 if ((p
->policy
== SCHED_RR
) && !--p
->time_slice
) {
2653 p
->time_slice
= task_timeslice(p
);
2654 p
->first_time_slice
= 0;
2655 set_tsk_need_resched(p
);
2657 /* put it at the end of the queue: */
2658 requeue_task(p
, rq
->active
);
2662 if (!--p
->time_slice
) {
2663 dequeue_task(p
, rq
->active
);
2664 set_tsk_need_resched(p
);
2665 p
->prio
= effective_prio(p
);
2666 p
->time_slice
= task_timeslice(p
);
2667 p
->first_time_slice
= 0;
2669 if (!rq
->expired_timestamp
)
2670 rq
->expired_timestamp
= jiffies
;
2671 if (!TASK_INTERACTIVE(p
) || EXPIRED_STARVING(rq
)) {
2672 enqueue_task(p
, rq
->expired
);
2673 if (p
->static_prio
< rq
->best_expired_prio
)
2674 rq
->best_expired_prio
= p
->static_prio
;
2676 enqueue_task(p
, rq
->active
);
2679 * Prevent a too long timeslice allowing a task to monopolize
2680 * the CPU. We do this by splitting up the timeslice into
2683 * Note: this does not mean the task's timeslices expire or
2684 * get lost in any way, they just might be preempted by
2685 * another task of equal priority. (one with higher
2686 * priority would have preempted this task already.) We
2687 * requeue this task to the end of the list on this priority
2688 * level, which is in essence a round-robin of tasks with
2691 * This only applies to tasks in the interactive
2692 * delta range with at least TIMESLICE_GRANULARITY to requeue.
2694 if (TASK_INTERACTIVE(p
) && !((task_timeslice(p
) -
2695 p
->time_slice
) % TIMESLICE_GRANULARITY(p
)) &&
2696 (p
->time_slice
>= TIMESLICE_GRANULARITY(p
)) &&
2697 (p
->array
== rq
->active
)) {
2699 requeue_task(p
, rq
->active
);
2700 set_tsk_need_resched(p
);
2704 spin_unlock(&rq
->lock
);
2706 rebalance_tick(cpu
, rq
, NOT_IDLE
);
2709 #ifdef CONFIG_SCHED_SMT
2710 static inline void wakeup_busy_runqueue(runqueue_t
*rq
)
2712 /* If an SMT runqueue is sleeping due to priority reasons wake it up */
2713 if (rq
->curr
== rq
->idle
&& rq
->nr_running
)
2714 resched_task(rq
->idle
);
2717 static inline void wake_sleeping_dependent(int this_cpu
, runqueue_t
*this_rq
)
2719 struct sched_domain
*tmp
, *sd
= NULL
;
2720 cpumask_t sibling_map
;
2723 for_each_domain(this_cpu
, tmp
)
2724 if (tmp
->flags
& SD_SHARE_CPUPOWER
)
2731 * Unlock the current runqueue because we have to lock in
2732 * CPU order to avoid deadlocks. Caller knows that we might
2733 * unlock. We keep IRQs disabled.
2735 spin_unlock(&this_rq
->lock
);
2737 sibling_map
= sd
->span
;
2739 for_each_cpu_mask(i
, sibling_map
)
2740 spin_lock(&cpu_rq(i
)->lock
);
2742 * We clear this CPU from the mask. This both simplifies the
2743 * inner loop and keps this_rq locked when we exit:
2745 cpu_clear(this_cpu
, sibling_map
);
2747 for_each_cpu_mask(i
, sibling_map
) {
2748 runqueue_t
*smt_rq
= cpu_rq(i
);
2750 wakeup_busy_runqueue(smt_rq
);
2753 for_each_cpu_mask(i
, sibling_map
)
2754 spin_unlock(&cpu_rq(i
)->lock
);
2756 * We exit with this_cpu's rq still held and IRQs
2762 * number of 'lost' timeslices this task wont be able to fully
2763 * utilize, if another task runs on a sibling. This models the
2764 * slowdown effect of other tasks running on siblings:
2766 static inline unsigned long smt_slice(task_t
*p
, struct sched_domain
*sd
)
2768 return p
->time_slice
* (100 - sd
->per_cpu_gain
) / 100;
2771 static inline int dependent_sleeper(int this_cpu
, runqueue_t
*this_rq
)
2773 struct sched_domain
*tmp
, *sd
= NULL
;
2774 cpumask_t sibling_map
;
2775 prio_array_t
*array
;
2779 for_each_domain(this_cpu
, tmp
)
2780 if (tmp
->flags
& SD_SHARE_CPUPOWER
)
2787 * The same locking rules and details apply as for
2788 * wake_sleeping_dependent():
2790 spin_unlock(&this_rq
->lock
);
2791 sibling_map
= sd
->span
;
2792 for_each_cpu_mask(i
, sibling_map
)
2793 spin_lock(&cpu_rq(i
)->lock
);
2794 cpu_clear(this_cpu
, sibling_map
);
2797 * Establish next task to be run - it might have gone away because
2798 * we released the runqueue lock above:
2800 if (!this_rq
->nr_running
)
2802 array
= this_rq
->active
;
2803 if (!array
->nr_active
)
2804 array
= this_rq
->expired
;
2805 BUG_ON(!array
->nr_active
);
2807 p
= list_entry(array
->queue
[sched_find_first_bit(array
->bitmap
)].next
,
2810 for_each_cpu_mask(i
, sibling_map
) {
2811 runqueue_t
*smt_rq
= cpu_rq(i
);
2812 task_t
*smt_curr
= smt_rq
->curr
;
2814 /* Kernel threads do not participate in dependent sleeping */
2815 if (!p
->mm
|| !smt_curr
->mm
|| rt_task(p
))
2816 goto check_smt_task
;
2819 * If a user task with lower static priority than the
2820 * running task on the SMT sibling is trying to schedule,
2821 * delay it till there is proportionately less timeslice
2822 * left of the sibling task to prevent a lower priority
2823 * task from using an unfair proportion of the
2824 * physical cpu's resources. -ck
2826 if (rt_task(smt_curr
)) {
2828 * With real time tasks we run non-rt tasks only
2829 * per_cpu_gain% of the time.
2831 if ((jiffies
% DEF_TIMESLICE
) >
2832 (sd
->per_cpu_gain
* DEF_TIMESLICE
/ 100))
2835 if (smt_curr
->static_prio
< p
->static_prio
&&
2836 !TASK_PREEMPTS_CURR(p
, smt_rq
) &&
2837 smt_slice(smt_curr
, sd
) > task_timeslice(p
))
2841 if ((!smt_curr
->mm
&& smt_curr
!= smt_rq
->idle
) ||
2845 wakeup_busy_runqueue(smt_rq
);
2850 * Reschedule a lower priority task on the SMT sibling for
2851 * it to be put to sleep, or wake it up if it has been put to
2852 * sleep for priority reasons to see if it should run now.
2855 if ((jiffies
% DEF_TIMESLICE
) >
2856 (sd
->per_cpu_gain
* DEF_TIMESLICE
/ 100))
2857 resched_task(smt_curr
);
2859 if (TASK_PREEMPTS_CURR(p
, smt_rq
) &&
2860 smt_slice(p
, sd
) > task_timeslice(smt_curr
))
2861 resched_task(smt_curr
);
2863 wakeup_busy_runqueue(smt_rq
);
2867 for_each_cpu_mask(i
, sibling_map
)
2868 spin_unlock(&cpu_rq(i
)->lock
);
2872 static inline void wake_sleeping_dependent(int this_cpu
, runqueue_t
*this_rq
)
2876 static inline int dependent_sleeper(int this_cpu
, runqueue_t
*this_rq
)
2882 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
2884 void fastcall
add_preempt_count(int val
)
2889 BUG_ON((preempt_count() < 0));
2890 preempt_count() += val
;
2892 * Spinlock count overflowing soon?
2894 BUG_ON((preempt_count() & PREEMPT_MASK
) >= PREEMPT_MASK
-10);
2896 EXPORT_SYMBOL(add_preempt_count
);
2898 void fastcall
sub_preempt_count(int val
)
2903 BUG_ON(val
> preempt_count());
2905 * Is the spinlock portion underflowing?
2907 BUG_ON((val
< PREEMPT_MASK
) && !(preempt_count() & PREEMPT_MASK
));
2908 preempt_count() -= val
;
2910 EXPORT_SYMBOL(sub_preempt_count
);
2915 * schedule() is the main scheduler function.
2917 asmlinkage
void __sched
schedule(void)
2920 task_t
*prev
, *next
;
2922 prio_array_t
*array
;
2923 struct list_head
*queue
;
2924 unsigned long long now
;
2925 unsigned long run_time
;
2926 int cpu
, idx
, new_prio
;
2929 * Test if we are atomic. Since do_exit() needs to call into
2930 * schedule() atomically, we ignore that path for now.
2931 * Otherwise, whine if we are scheduling when we should not be.
2933 if (likely(!current
->exit_state
)) {
2934 if (unlikely(in_atomic())) {
2935 printk(KERN_ERR
"scheduling while atomic: "
2937 current
->comm
, preempt_count(), current
->pid
);
2941 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2946 release_kernel_lock(prev
);
2947 need_resched_nonpreemptible
:
2951 * The idle thread is not allowed to schedule!
2952 * Remove this check after it has been exercised a bit.
2954 if (unlikely(prev
== rq
->idle
) && prev
->state
!= TASK_RUNNING
) {
2955 printk(KERN_ERR
"bad: scheduling from the idle thread!\n");
2959 schedstat_inc(rq
, sched_cnt
);
2960 now
= sched_clock();
2961 if (likely((long long)(now
- prev
->timestamp
) < NS_MAX_SLEEP_AVG
)) {
2962 run_time
= now
- prev
->timestamp
;
2963 if (unlikely((long long)(now
- prev
->timestamp
) < 0))
2966 run_time
= NS_MAX_SLEEP_AVG
;
2969 * Tasks charged proportionately less run_time at high sleep_avg to
2970 * delay them losing their interactive status
2972 run_time
/= (CURRENT_BONUS(prev
) ? : 1);
2974 spin_lock_irq(&rq
->lock
);
2976 if (unlikely(prev
->flags
& PF_DEAD
))
2977 prev
->state
= EXIT_DEAD
;
2979 switch_count
= &prev
->nivcsw
;
2980 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
2981 switch_count
= &prev
->nvcsw
;
2982 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
2983 unlikely(signal_pending(prev
))))
2984 prev
->state
= TASK_RUNNING
;
2986 if (prev
->state
== TASK_UNINTERRUPTIBLE
)
2987 rq
->nr_uninterruptible
++;
2988 deactivate_task(prev
, rq
);
2992 cpu
= smp_processor_id();
2993 if (unlikely(!rq
->nr_running
)) {
2995 idle_balance(cpu
, rq
);
2996 if (!rq
->nr_running
) {
2998 rq
->expired_timestamp
= 0;
2999 wake_sleeping_dependent(cpu
, rq
);
3001 * wake_sleeping_dependent() might have released
3002 * the runqueue, so break out if we got new
3005 if (!rq
->nr_running
)
3009 if (dependent_sleeper(cpu
, rq
)) {
3014 * dependent_sleeper() releases and reacquires the runqueue
3015 * lock, hence go into the idle loop if the rq went
3018 if (unlikely(!rq
->nr_running
))
3023 if (unlikely(!array
->nr_active
)) {
3025 * Switch the active and expired arrays.
3027 schedstat_inc(rq
, sched_switch
);
3028 rq
->active
= rq
->expired
;
3029 rq
->expired
= array
;
3031 rq
->expired_timestamp
= 0;
3032 rq
->best_expired_prio
= MAX_PRIO
;
3035 idx
= sched_find_first_bit(array
->bitmap
);
3036 queue
= array
->queue
+ idx
;
3037 next
= list_entry(queue
->next
, task_t
, run_list
);
3039 if (!rt_task(next
) && next
->activated
> 0) {
3040 unsigned long long delta
= now
- next
->timestamp
;
3041 if (unlikely((long long)(now
- next
->timestamp
) < 0))
3044 if (next
->activated
== 1)
3045 delta
= delta
* (ON_RUNQUEUE_WEIGHT
* 128 / 100) / 128;
3047 array
= next
->array
;
3048 new_prio
= recalc_task_prio(next
, next
->timestamp
+ delta
);
3050 if (unlikely(next
->prio
!= new_prio
)) {
3051 dequeue_task(next
, array
);
3052 next
->prio
= new_prio
;
3053 enqueue_task(next
, array
);
3055 requeue_task(next
, array
);
3057 next
->activated
= 0;
3059 if (next
== rq
->idle
)
3060 schedstat_inc(rq
, sched_goidle
);
3062 prefetch_stack(next
);
3063 clear_tsk_need_resched(prev
);
3064 rcu_qsctr_inc(task_cpu(prev
));
3066 update_cpu_clock(prev
, rq
, now
);
3068 prev
->sleep_avg
-= run_time
;
3069 if ((long)prev
->sleep_avg
<= 0)
3070 prev
->sleep_avg
= 0;
3071 prev
->timestamp
= prev
->last_ran
= now
;
3073 sched_info_switch(prev
, next
);
3074 if (likely(prev
!= next
)) {
3075 next
->timestamp
= now
;
3080 prepare_task_switch(rq
, next
);
3081 prev
= context_switch(rq
, prev
, next
);
3084 * this_rq must be evaluated again because prev may have moved
3085 * CPUs since it called schedule(), thus the 'rq' on its stack
3086 * frame will be invalid.
3088 finish_task_switch(this_rq(), prev
);
3090 spin_unlock_irq(&rq
->lock
);
3093 if (unlikely(reacquire_kernel_lock(prev
) < 0))
3094 goto need_resched_nonpreemptible
;
3095 preempt_enable_no_resched();
3096 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3100 EXPORT_SYMBOL(schedule
);
3102 #ifdef CONFIG_PREEMPT
3104 * this is is the entry point to schedule() from in-kernel preemption
3105 * off of preempt_enable. Kernel preemptions off return from interrupt
3106 * occur there and call schedule directly.
3108 asmlinkage
void __sched
preempt_schedule(void)
3110 struct thread_info
*ti
= current_thread_info();
3111 #ifdef CONFIG_PREEMPT_BKL
3112 struct task_struct
*task
= current
;
3113 int saved_lock_depth
;
3116 * If there is a non-zero preempt_count or interrupts are disabled,
3117 * we do not want to preempt the current task. Just return..
3119 if (unlikely(ti
->preempt_count
|| irqs_disabled()))
3123 add_preempt_count(PREEMPT_ACTIVE
);
3125 * We keep the big kernel semaphore locked, but we
3126 * clear ->lock_depth so that schedule() doesnt
3127 * auto-release the semaphore:
3129 #ifdef CONFIG_PREEMPT_BKL
3130 saved_lock_depth
= task
->lock_depth
;
3131 task
->lock_depth
= -1;
3134 #ifdef CONFIG_PREEMPT_BKL
3135 task
->lock_depth
= saved_lock_depth
;
3137 sub_preempt_count(PREEMPT_ACTIVE
);
3139 /* we could miss a preemption opportunity between schedule and now */
3141 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3145 EXPORT_SYMBOL(preempt_schedule
);
3148 * this is is the entry point to schedule() from kernel preemption
3149 * off of irq context.
3150 * Note, that this is called and return with irqs disabled. This will
3151 * protect us against recursive calling from irq.
3153 asmlinkage
void __sched
preempt_schedule_irq(void)
3155 struct thread_info
*ti
= current_thread_info();
3156 #ifdef CONFIG_PREEMPT_BKL
3157 struct task_struct
*task
= current
;
3158 int saved_lock_depth
;
3160 /* Catch callers which need to be fixed*/
3161 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3164 add_preempt_count(PREEMPT_ACTIVE
);
3166 * We keep the big kernel semaphore locked, but we
3167 * clear ->lock_depth so that schedule() doesnt
3168 * auto-release the semaphore:
3170 #ifdef CONFIG_PREEMPT_BKL
3171 saved_lock_depth
= task
->lock_depth
;
3172 task
->lock_depth
= -1;
3176 local_irq_disable();
3177 #ifdef CONFIG_PREEMPT_BKL
3178 task
->lock_depth
= saved_lock_depth
;
3180 sub_preempt_count(PREEMPT_ACTIVE
);
3182 /* we could miss a preemption opportunity between schedule and now */
3184 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3188 #endif /* CONFIG_PREEMPT */
3190 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3193 task_t
*p
= curr
->private;
3194 return try_to_wake_up(p
, mode
, sync
);
3197 EXPORT_SYMBOL(default_wake_function
);
3200 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3201 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3202 * number) then we wake all the non-exclusive tasks and one exclusive task.
3204 * There are circumstances in which we can try to wake a task which has already
3205 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3206 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3208 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3209 int nr_exclusive
, int sync
, void *key
)
3211 struct list_head
*tmp
, *next
;
3213 list_for_each_safe(tmp
, next
, &q
->task_list
) {
3216 curr
= list_entry(tmp
, wait_queue_t
, task_list
);
3217 flags
= curr
->flags
;
3218 if (curr
->func(curr
, mode
, sync
, key
) &&
3219 (flags
& WQ_FLAG_EXCLUSIVE
) &&
3226 * __wake_up - wake up threads blocked on a waitqueue.
3228 * @mode: which threads
3229 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3230 * @key: is directly passed to the wakeup function
3232 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3233 int nr_exclusive
, void *key
)
3235 unsigned long flags
;
3237 spin_lock_irqsave(&q
->lock
, flags
);
3238 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3239 spin_unlock_irqrestore(&q
->lock
, flags
);
3242 EXPORT_SYMBOL(__wake_up
);
3245 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3247 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3249 __wake_up_common(q
, mode
, 1, 0, NULL
);
3253 * __wake_up_sync - wake up threads blocked on a waitqueue.
3255 * @mode: which threads
3256 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3258 * The sync wakeup differs that the waker knows that it will schedule
3259 * away soon, so while the target thread will be woken up, it will not
3260 * be migrated to another CPU - ie. the two threads are 'synchronized'
3261 * with each other. This can prevent needless bouncing between CPUs.
3263 * On UP it can prevent extra preemption.
3266 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3268 unsigned long flags
;
3274 if (unlikely(!nr_exclusive
))
3277 spin_lock_irqsave(&q
->lock
, flags
);
3278 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3279 spin_unlock_irqrestore(&q
->lock
, flags
);
3281 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3283 void fastcall
complete(struct completion
*x
)
3285 unsigned long flags
;
3287 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3289 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3291 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3293 EXPORT_SYMBOL(complete
);
3295 void fastcall
complete_all(struct completion
*x
)
3297 unsigned long flags
;
3299 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3300 x
->done
+= UINT_MAX
/2;
3301 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3303 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3305 EXPORT_SYMBOL(complete_all
);
3307 void fastcall __sched
wait_for_completion(struct completion
*x
)
3310 spin_lock_irq(&x
->wait
.lock
);
3312 DECLARE_WAITQUEUE(wait
, current
);
3314 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3315 __add_wait_queue_tail(&x
->wait
, &wait
);
3317 __set_current_state(TASK_UNINTERRUPTIBLE
);
3318 spin_unlock_irq(&x
->wait
.lock
);
3320 spin_lock_irq(&x
->wait
.lock
);
3322 __remove_wait_queue(&x
->wait
, &wait
);
3325 spin_unlock_irq(&x
->wait
.lock
);
3327 EXPORT_SYMBOL(wait_for_completion
);
3329 unsigned long fastcall __sched
3330 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3334 spin_lock_irq(&x
->wait
.lock
);
3336 DECLARE_WAITQUEUE(wait
, current
);
3338 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3339 __add_wait_queue_tail(&x
->wait
, &wait
);
3341 __set_current_state(TASK_UNINTERRUPTIBLE
);
3342 spin_unlock_irq(&x
->wait
.lock
);
3343 timeout
= schedule_timeout(timeout
);
3344 spin_lock_irq(&x
->wait
.lock
);
3346 __remove_wait_queue(&x
->wait
, &wait
);
3350 __remove_wait_queue(&x
->wait
, &wait
);
3354 spin_unlock_irq(&x
->wait
.lock
);
3357 EXPORT_SYMBOL(wait_for_completion_timeout
);
3359 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3365 spin_lock_irq(&x
->wait
.lock
);
3367 DECLARE_WAITQUEUE(wait
, current
);
3369 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3370 __add_wait_queue_tail(&x
->wait
, &wait
);
3372 if (signal_pending(current
)) {
3374 __remove_wait_queue(&x
->wait
, &wait
);
3377 __set_current_state(TASK_INTERRUPTIBLE
);
3378 spin_unlock_irq(&x
->wait
.lock
);
3380 spin_lock_irq(&x
->wait
.lock
);
3382 __remove_wait_queue(&x
->wait
, &wait
);
3386 spin_unlock_irq(&x
->wait
.lock
);
3390 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3392 unsigned long fastcall __sched
3393 wait_for_completion_interruptible_timeout(struct completion
*x
,
3394 unsigned long timeout
)
3398 spin_lock_irq(&x
->wait
.lock
);
3400 DECLARE_WAITQUEUE(wait
, current
);
3402 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3403 __add_wait_queue_tail(&x
->wait
, &wait
);
3405 if (signal_pending(current
)) {
3406 timeout
= -ERESTARTSYS
;
3407 __remove_wait_queue(&x
->wait
, &wait
);
3410 __set_current_state(TASK_INTERRUPTIBLE
);
3411 spin_unlock_irq(&x
->wait
.lock
);
3412 timeout
= schedule_timeout(timeout
);
3413 spin_lock_irq(&x
->wait
.lock
);
3415 __remove_wait_queue(&x
->wait
, &wait
);
3419 __remove_wait_queue(&x
->wait
, &wait
);
3423 spin_unlock_irq(&x
->wait
.lock
);
3426 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3429 #define SLEEP_ON_VAR \
3430 unsigned long flags; \
3431 wait_queue_t wait; \
3432 init_waitqueue_entry(&wait, current);
3434 #define SLEEP_ON_HEAD \
3435 spin_lock_irqsave(&q->lock,flags); \
3436 __add_wait_queue(q, &wait); \
3437 spin_unlock(&q->lock);
3439 #define SLEEP_ON_TAIL \
3440 spin_lock_irq(&q->lock); \
3441 __remove_wait_queue(q, &wait); \
3442 spin_unlock_irqrestore(&q->lock, flags);
3444 void fastcall __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3448 current
->state
= TASK_INTERRUPTIBLE
;
3455 EXPORT_SYMBOL(interruptible_sleep_on
);
3457 long fastcall __sched
3458 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3462 current
->state
= TASK_INTERRUPTIBLE
;
3465 timeout
= schedule_timeout(timeout
);
3471 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3473 void fastcall __sched
sleep_on(wait_queue_head_t
*q
)
3477 current
->state
= TASK_UNINTERRUPTIBLE
;
3484 EXPORT_SYMBOL(sleep_on
);
3486 long fastcall __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3490 current
->state
= TASK_UNINTERRUPTIBLE
;
3493 timeout
= schedule_timeout(timeout
);
3499 EXPORT_SYMBOL(sleep_on_timeout
);
3501 void set_user_nice(task_t
*p
, long nice
)
3503 unsigned long flags
;
3504 prio_array_t
*array
;
3506 int old_prio
, new_prio
, delta
;
3508 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3511 * We have to be careful, if called from sys_setpriority(),
3512 * the task might be in the middle of scheduling on another CPU.
3514 rq
= task_rq_lock(p
, &flags
);
3516 * The RT priorities are set via sched_setscheduler(), but we still
3517 * allow the 'normal' nice value to be set - but as expected
3518 * it wont have any effect on scheduling until the task is
3522 p
->static_prio
= NICE_TO_PRIO(nice
);
3527 dequeue_task(p
, array
);
3528 dec_prio_bias(rq
, p
->static_prio
);
3532 new_prio
= NICE_TO_PRIO(nice
);
3533 delta
= new_prio
- old_prio
;
3534 p
->static_prio
= NICE_TO_PRIO(nice
);
3538 enqueue_task(p
, array
);
3539 inc_prio_bias(rq
, p
->static_prio
);
3541 * If the task increased its priority or is running and
3542 * lowered its priority, then reschedule its CPU:
3544 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3545 resched_task(rq
->curr
);
3548 task_rq_unlock(rq
, &flags
);
3551 EXPORT_SYMBOL(set_user_nice
);
3554 * can_nice - check if a task can reduce its nice value
3558 int can_nice(const task_t
*p
, const int nice
)
3560 /* convert nice value [19,-20] to rlimit style value [1,40] */
3561 int nice_rlim
= 20 - nice
;
3562 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
3563 capable(CAP_SYS_NICE
));
3566 #ifdef __ARCH_WANT_SYS_NICE
3569 * sys_nice - change the priority of the current process.
3570 * @increment: priority increment
3572 * sys_setpriority is a more generic, but much slower function that
3573 * does similar things.
3575 asmlinkage
long sys_nice(int increment
)
3581 * Setpriority might change our priority at the same moment.
3582 * We don't have to worry. Conceptually one call occurs first
3583 * and we have a single winner.
3585 if (increment
< -40)
3590 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
3596 if (increment
< 0 && !can_nice(current
, nice
))
3599 retval
= security_task_setnice(current
, nice
);
3603 set_user_nice(current
, nice
);
3610 * task_prio - return the priority value of a given task.
3611 * @p: the task in question.
3613 * This is the priority value as seen by users in /proc.
3614 * RT tasks are offset by -200. Normal tasks are centered
3615 * around 0, value goes from -16 to +15.
3617 int task_prio(const task_t
*p
)
3619 return p
->prio
- MAX_RT_PRIO
;
3623 * task_nice - return the nice value of a given task.
3624 * @p: the task in question.
3626 int task_nice(const task_t
*p
)
3628 return TASK_NICE(p
);
3630 EXPORT_SYMBOL_GPL(task_nice
);
3633 * idle_cpu - is a given cpu idle currently?
3634 * @cpu: the processor in question.
3636 int idle_cpu(int cpu
)
3638 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
3642 * idle_task - return the idle task for a given cpu.
3643 * @cpu: the processor in question.
3645 task_t
*idle_task(int cpu
)
3647 return cpu_rq(cpu
)->idle
;
3651 * find_process_by_pid - find a process with a matching PID value.
3652 * @pid: the pid in question.
3654 static inline task_t
*find_process_by_pid(pid_t pid
)
3656 return pid
? find_task_by_pid(pid
) : current
;
3659 /* Actually do priority change: must hold rq lock. */
3660 static void __setscheduler(struct task_struct
*p
, int policy
, int prio
)
3664 p
->rt_priority
= prio
;
3665 if (policy
!= SCHED_NORMAL
)
3666 p
->prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
3668 p
->prio
= p
->static_prio
;
3672 * sched_setscheduler - change the scheduling policy and/or RT priority of
3674 * @p: the task in question.
3675 * @policy: new policy.
3676 * @param: structure containing the new RT priority.
3678 int sched_setscheduler(struct task_struct
*p
, int policy
,
3679 struct sched_param
*param
)
3682 int oldprio
, oldpolicy
= -1;
3683 prio_array_t
*array
;
3684 unsigned long flags
;
3688 /* double check policy once rq lock held */
3690 policy
= oldpolicy
= p
->policy
;
3691 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3692 policy
!= SCHED_NORMAL
)
3695 * Valid priorities for SCHED_FIFO and SCHED_RR are
3696 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL is 0.
3698 if (param
->sched_priority
< 0 ||
3699 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3700 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
3702 if ((policy
== SCHED_NORMAL
) != (param
->sched_priority
== 0))
3706 * Allow unprivileged RT tasks to decrease priority:
3708 if (!capable(CAP_SYS_NICE
)) {
3709 /* can't change policy */
3710 if (policy
!= p
->policy
&&
3711 !p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
)
3713 /* can't increase priority */
3714 if (policy
!= SCHED_NORMAL
&&
3715 param
->sched_priority
> p
->rt_priority
&&
3716 param
->sched_priority
>
3717 p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
)
3719 /* can't change other user's priorities */
3720 if ((current
->euid
!= p
->euid
) &&
3721 (current
->euid
!= p
->uid
))
3725 retval
= security_task_setscheduler(p
, policy
, param
);
3729 * To be able to change p->policy safely, the apropriate
3730 * runqueue lock must be held.
3732 rq
= task_rq_lock(p
, &flags
);
3733 /* recheck policy now with rq lock held */
3734 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3735 policy
= oldpolicy
= -1;
3736 task_rq_unlock(rq
, &flags
);
3741 deactivate_task(p
, rq
);
3743 __setscheduler(p
, policy
, param
->sched_priority
);
3745 __activate_task(p
, rq
);
3747 * Reschedule if we are currently running on this runqueue and
3748 * our priority decreased, or if we are not currently running on
3749 * this runqueue and our priority is higher than the current's
3751 if (task_running(rq
, p
)) {
3752 if (p
->prio
> oldprio
)
3753 resched_task(rq
->curr
);
3754 } else if (TASK_PREEMPTS_CURR(p
, rq
))
3755 resched_task(rq
->curr
);
3757 task_rq_unlock(rq
, &flags
);
3760 EXPORT_SYMBOL_GPL(sched_setscheduler
);
3763 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
3766 struct sched_param lparam
;
3767 struct task_struct
*p
;
3769 if (!param
|| pid
< 0)
3771 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
3773 read_lock_irq(&tasklist_lock
);
3774 p
= find_process_by_pid(pid
);
3776 read_unlock_irq(&tasklist_lock
);
3779 retval
= sched_setscheduler(p
, policy
, &lparam
);
3780 read_unlock_irq(&tasklist_lock
);
3785 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3786 * @pid: the pid in question.
3787 * @policy: new policy.
3788 * @param: structure containing the new RT priority.
3790 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
3791 struct sched_param __user
*param
)
3793 return do_sched_setscheduler(pid
, policy
, param
);
3797 * sys_sched_setparam - set/change the RT priority of a thread
3798 * @pid: the pid in question.
3799 * @param: structure containing the new RT priority.
3801 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
3803 return do_sched_setscheduler(pid
, -1, param
);
3807 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3808 * @pid: the pid in question.
3810 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
3812 int retval
= -EINVAL
;
3819 read_lock(&tasklist_lock
);
3820 p
= find_process_by_pid(pid
);
3822 retval
= security_task_getscheduler(p
);
3826 read_unlock(&tasklist_lock
);
3833 * sys_sched_getscheduler - get the RT priority of a thread
3834 * @pid: the pid in question.
3835 * @param: structure containing the RT priority.
3837 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
3839 struct sched_param lp
;
3840 int retval
= -EINVAL
;
3843 if (!param
|| pid
< 0)
3846 read_lock(&tasklist_lock
);
3847 p
= find_process_by_pid(pid
);
3852 retval
= security_task_getscheduler(p
);
3856 lp
.sched_priority
= p
->rt_priority
;
3857 read_unlock(&tasklist_lock
);
3860 * This one might sleep, we cannot do it with a spinlock held ...
3862 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
3868 read_unlock(&tasklist_lock
);
3872 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
3876 cpumask_t cpus_allowed
;
3879 read_lock(&tasklist_lock
);
3881 p
= find_process_by_pid(pid
);
3883 read_unlock(&tasklist_lock
);
3884 unlock_cpu_hotplug();
3889 * It is not safe to call set_cpus_allowed with the
3890 * tasklist_lock held. We will bump the task_struct's
3891 * usage count and then drop tasklist_lock.
3894 read_unlock(&tasklist_lock
);
3897 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
3898 !capable(CAP_SYS_NICE
))
3901 cpus_allowed
= cpuset_cpus_allowed(p
);
3902 cpus_and(new_mask
, new_mask
, cpus_allowed
);
3903 retval
= set_cpus_allowed(p
, new_mask
);
3907 unlock_cpu_hotplug();
3911 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
3912 cpumask_t
*new_mask
)
3914 if (len
< sizeof(cpumask_t
)) {
3915 memset(new_mask
, 0, sizeof(cpumask_t
));
3916 } else if (len
> sizeof(cpumask_t
)) {
3917 len
= sizeof(cpumask_t
);
3919 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
3923 * sys_sched_setaffinity - set the cpu affinity of a process
3924 * @pid: pid of the process
3925 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3926 * @user_mask_ptr: user-space pointer to the new cpu mask
3928 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
3929 unsigned long __user
*user_mask_ptr
)
3934 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
3938 return sched_setaffinity(pid
, new_mask
);
3942 * Represents all cpu's present in the system
3943 * In systems capable of hotplug, this map could dynamically grow
3944 * as new cpu's are detected in the system via any platform specific
3945 * method, such as ACPI for e.g.
3948 cpumask_t cpu_present_map
;
3949 EXPORT_SYMBOL(cpu_present_map
);
3952 cpumask_t cpu_online_map
= CPU_MASK_ALL
;
3953 cpumask_t cpu_possible_map
= CPU_MASK_ALL
;
3956 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
3962 read_lock(&tasklist_lock
);
3965 p
= find_process_by_pid(pid
);
3970 cpus_and(*mask
, p
->cpus_allowed
, cpu_possible_map
);
3973 read_unlock(&tasklist_lock
);
3974 unlock_cpu_hotplug();
3982 * sys_sched_getaffinity - get the cpu affinity of a process
3983 * @pid: pid of the process
3984 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3985 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3987 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
3988 unsigned long __user
*user_mask_ptr
)
3993 if (len
< sizeof(cpumask_t
))
3996 ret
= sched_getaffinity(pid
, &mask
);
4000 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4003 return sizeof(cpumask_t
);
4007 * sys_sched_yield - yield the current processor to other threads.
4009 * this function yields the current CPU by moving the calling thread
4010 * to the expired array. If there are no other threads running on this
4011 * CPU then this function will return.
4013 asmlinkage
long sys_sched_yield(void)
4015 runqueue_t
*rq
= this_rq_lock();
4016 prio_array_t
*array
= current
->array
;
4017 prio_array_t
*target
= rq
->expired
;
4019 schedstat_inc(rq
, yld_cnt
);
4021 * We implement yielding by moving the task into the expired
4024 * (special rule: RT tasks will just roundrobin in the active
4027 if (rt_task(current
))
4028 target
= rq
->active
;
4030 if (array
->nr_active
== 1) {
4031 schedstat_inc(rq
, yld_act_empty
);
4032 if (!rq
->expired
->nr_active
)
4033 schedstat_inc(rq
, yld_both_empty
);
4034 } else if (!rq
->expired
->nr_active
)
4035 schedstat_inc(rq
, yld_exp_empty
);
4037 if (array
!= target
) {
4038 dequeue_task(current
, array
);
4039 enqueue_task(current
, target
);
4042 * requeue_task is cheaper so perform that if possible.
4044 requeue_task(current
, array
);
4047 * Since we are going to call schedule() anyway, there's
4048 * no need to preempt or enable interrupts:
4050 __release(rq
->lock
);
4051 _raw_spin_unlock(&rq
->lock
);
4052 preempt_enable_no_resched();
4059 static inline void __cond_resched(void)
4062 * The BKS might be reacquired before we have dropped
4063 * PREEMPT_ACTIVE, which could trigger a second
4064 * cond_resched() call.
4066 if (unlikely(preempt_count()))
4069 add_preempt_count(PREEMPT_ACTIVE
);
4071 sub_preempt_count(PREEMPT_ACTIVE
);
4072 } while (need_resched());
4075 int __sched
cond_resched(void)
4077 if (need_resched()) {
4084 EXPORT_SYMBOL(cond_resched
);
4087 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4088 * call schedule, and on return reacquire the lock.
4090 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4091 * operations here to prevent schedule() from being called twice (once via
4092 * spin_unlock(), once by hand).
4094 int cond_resched_lock(spinlock_t
*lock
)
4098 if (need_lockbreak(lock
)) {
4104 if (need_resched()) {
4105 _raw_spin_unlock(lock
);
4106 preempt_enable_no_resched();
4114 EXPORT_SYMBOL(cond_resched_lock
);
4116 int __sched
cond_resched_softirq(void)
4118 BUG_ON(!in_softirq());
4120 if (need_resched()) {
4121 __local_bh_enable();
4129 EXPORT_SYMBOL(cond_resched_softirq
);
4133 * yield - yield the current processor to other threads.
4135 * this is a shortcut for kernel-space yielding - it marks the
4136 * thread runnable and calls sys_sched_yield().
4138 void __sched
yield(void)
4140 set_current_state(TASK_RUNNING
);
4144 EXPORT_SYMBOL(yield
);
4147 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4148 * that process accounting knows that this is a task in IO wait state.
4150 * But don't do that if it is a deliberate, throttling IO wait (this task
4151 * has set its backing_dev_info: the queue against which it should throttle)
4153 void __sched
io_schedule(void)
4155 struct runqueue
*rq
= &per_cpu(runqueues
, raw_smp_processor_id());
4157 atomic_inc(&rq
->nr_iowait
);
4159 atomic_dec(&rq
->nr_iowait
);
4162 EXPORT_SYMBOL(io_schedule
);
4164 long __sched
io_schedule_timeout(long timeout
)
4166 struct runqueue
*rq
= &per_cpu(runqueues
, raw_smp_processor_id());
4169 atomic_inc(&rq
->nr_iowait
);
4170 ret
= schedule_timeout(timeout
);
4171 atomic_dec(&rq
->nr_iowait
);
4176 * sys_sched_get_priority_max - return maximum RT priority.
4177 * @policy: scheduling class.
4179 * this syscall returns the maximum rt_priority that can be used
4180 * by a given scheduling class.
4182 asmlinkage
long sys_sched_get_priority_max(int policy
)
4189 ret
= MAX_USER_RT_PRIO
-1;
4199 * sys_sched_get_priority_min - return minimum RT priority.
4200 * @policy: scheduling class.
4202 * this syscall returns the minimum rt_priority that can be used
4203 * by a given scheduling class.
4205 asmlinkage
long sys_sched_get_priority_min(int policy
)
4221 * sys_sched_rr_get_interval - return the default timeslice of a process.
4222 * @pid: pid of the process.
4223 * @interval: userspace pointer to the timeslice value.
4225 * this syscall writes the default timeslice value of a given process
4226 * into the user-space timespec buffer. A value of '0' means infinity.
4229 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4231 int retval
= -EINVAL
;
4239 read_lock(&tasklist_lock
);
4240 p
= find_process_by_pid(pid
);
4244 retval
= security_task_getscheduler(p
);
4248 jiffies_to_timespec(p
->policy
& SCHED_FIFO
?
4249 0 : task_timeslice(p
), &t
);
4250 read_unlock(&tasklist_lock
);
4251 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4255 read_unlock(&tasklist_lock
);
4259 static inline struct task_struct
*eldest_child(struct task_struct
*p
)
4261 if (list_empty(&p
->children
)) return NULL
;
4262 return list_entry(p
->children
.next
,struct task_struct
,sibling
);
4265 static inline struct task_struct
*older_sibling(struct task_struct
*p
)
4267 if (p
->sibling
.prev
==&p
->parent
->children
) return NULL
;
4268 return list_entry(p
->sibling
.prev
,struct task_struct
,sibling
);
4271 static inline struct task_struct
*younger_sibling(struct task_struct
*p
)
4273 if (p
->sibling
.next
==&p
->parent
->children
) return NULL
;
4274 return list_entry(p
->sibling
.next
,struct task_struct
,sibling
);
4277 static void show_task(task_t
*p
)
4281 unsigned long free
= 0;
4282 static const char *stat_nam
[] = { "R", "S", "D", "T", "t", "Z", "X" };
4284 printk("%-13.13s ", p
->comm
);
4285 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4286 if (state
< ARRAY_SIZE(stat_nam
))
4287 printk(stat_nam
[state
]);
4290 #if (BITS_PER_LONG == 32)
4291 if (state
== TASK_RUNNING
)
4292 printk(" running ");
4294 printk(" %08lX ", thread_saved_pc(p
));
4296 if (state
== TASK_RUNNING
)
4297 printk(" running task ");
4299 printk(" %016lx ", thread_saved_pc(p
));
4301 #ifdef CONFIG_DEBUG_STACK_USAGE
4303 unsigned long *n
= (unsigned long *) (p
->thread_info
+1);
4306 free
= (unsigned long) n
- (unsigned long)(p
->thread_info
+1);
4309 printk("%5lu %5d %6d ", free
, p
->pid
, p
->parent
->pid
);
4310 if ((relative
= eldest_child(p
)))
4311 printk("%5d ", relative
->pid
);
4314 if ((relative
= younger_sibling(p
)))
4315 printk("%7d", relative
->pid
);
4318 if ((relative
= older_sibling(p
)))
4319 printk(" %5d", relative
->pid
);
4323 printk(" (L-TLB)\n");
4325 printk(" (NOTLB)\n");
4327 if (state
!= TASK_RUNNING
)
4328 show_stack(p
, NULL
);
4331 void show_state(void)
4335 #if (BITS_PER_LONG == 32)
4338 printk(" task PC pid father child younger older\n");
4342 printk(" task PC pid father child younger older\n");
4344 read_lock(&tasklist_lock
);
4345 do_each_thread(g
, p
) {
4347 * reset the NMI-timeout, listing all files on a slow
4348 * console might take alot of time:
4350 touch_nmi_watchdog();
4352 } while_each_thread(g
, p
);
4354 read_unlock(&tasklist_lock
);
4358 * init_idle - set up an idle thread for a given CPU
4359 * @idle: task in question
4360 * @cpu: cpu the idle task belongs to
4362 * NOTE: this function does not set the idle thread's NEED_RESCHED
4363 * flag, to make booting more robust.
4365 void __devinit
init_idle(task_t
*idle
, int cpu
)
4367 runqueue_t
*rq
= cpu_rq(cpu
);
4368 unsigned long flags
;
4370 idle
->sleep_avg
= 0;
4372 idle
->prio
= MAX_PRIO
;
4373 idle
->state
= TASK_RUNNING
;
4374 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4375 set_task_cpu(idle
, cpu
);
4377 spin_lock_irqsave(&rq
->lock
, flags
);
4378 rq
->curr
= rq
->idle
= idle
;
4379 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4382 spin_unlock_irqrestore(&rq
->lock
, flags
);
4384 /* Set the preempt count _outside_ the spinlocks! */
4385 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4386 idle
->thread_info
->preempt_count
= (idle
->lock_depth
>= 0);
4388 idle
->thread_info
->preempt_count
= 0;
4393 * In a system that switches off the HZ timer nohz_cpu_mask
4394 * indicates which cpus entered this state. This is used
4395 * in the rcu update to wait only for active cpus. For system
4396 * which do not switch off the HZ timer nohz_cpu_mask should
4397 * always be CPU_MASK_NONE.
4399 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4403 * This is how migration works:
4405 * 1) we queue a migration_req_t structure in the source CPU's
4406 * runqueue and wake up that CPU's migration thread.
4407 * 2) we down() the locked semaphore => thread blocks.
4408 * 3) migration thread wakes up (implicitly it forces the migrated
4409 * thread off the CPU)
4410 * 4) it gets the migration request and checks whether the migrated
4411 * task is still in the wrong runqueue.
4412 * 5) if it's in the wrong runqueue then the migration thread removes
4413 * it and puts it into the right queue.
4414 * 6) migration thread up()s the semaphore.
4415 * 7) we wake up and the migration is done.
4419 * Change a given task's CPU affinity. Migrate the thread to a
4420 * proper CPU and schedule it away if the CPU it's executing on
4421 * is removed from the allowed bitmask.
4423 * NOTE: the caller must have a valid reference to the task, the
4424 * task must not exit() & deallocate itself prematurely. The
4425 * call is not atomic; no spinlocks may be held.
4427 int set_cpus_allowed(task_t
*p
, cpumask_t new_mask
)
4429 unsigned long flags
;
4431 migration_req_t req
;
4434 rq
= task_rq_lock(p
, &flags
);
4435 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4440 p
->cpus_allowed
= new_mask
;
4441 /* Can the task run on the task's current CPU? If so, we're done */
4442 if (cpu_isset(task_cpu(p
), new_mask
))
4445 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4446 /* Need help from migration thread: drop lock and wait. */
4447 task_rq_unlock(rq
, &flags
);
4448 wake_up_process(rq
->migration_thread
);
4449 wait_for_completion(&req
.done
);
4450 tlb_migrate_finish(p
->mm
);
4454 task_rq_unlock(rq
, &flags
);
4458 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4461 * Move (not current) task off this cpu, onto dest cpu. We're doing
4462 * this because either it can't run here any more (set_cpus_allowed()
4463 * away from this CPU, or CPU going down), or because we're
4464 * attempting to rebalance this task on exec (sched_exec).
4466 * So we race with normal scheduler movements, but that's OK, as long
4467 * as the task is no longer on this CPU.
4469 static void __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4471 runqueue_t
*rq_dest
, *rq_src
;
4473 if (unlikely(cpu_is_offline(dest_cpu
)))
4476 rq_src
= cpu_rq(src_cpu
);
4477 rq_dest
= cpu_rq(dest_cpu
);
4479 double_rq_lock(rq_src
, rq_dest
);
4480 /* Already moved. */
4481 if (task_cpu(p
) != src_cpu
)
4483 /* Affinity changed (again). */
4484 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
4487 set_task_cpu(p
, dest_cpu
);
4490 * Sync timestamp with rq_dest's before activating.
4491 * The same thing could be achieved by doing this step
4492 * afterwards, and pretending it was a local activate.
4493 * This way is cleaner and logically correct.
4495 p
->timestamp
= p
->timestamp
- rq_src
->timestamp_last_tick
4496 + rq_dest
->timestamp_last_tick
;
4497 deactivate_task(p
, rq_src
);
4498 activate_task(p
, rq_dest
, 0);
4499 if (TASK_PREEMPTS_CURR(p
, rq_dest
))
4500 resched_task(rq_dest
->curr
);
4504 double_rq_unlock(rq_src
, rq_dest
);
4508 * migration_thread - this is a highprio system thread that performs
4509 * thread migration by bumping thread off CPU then 'pushing' onto
4512 static int migration_thread(void *data
)
4515 int cpu
= (long)data
;
4518 BUG_ON(rq
->migration_thread
!= current
);
4520 set_current_state(TASK_INTERRUPTIBLE
);
4521 while (!kthread_should_stop()) {
4522 struct list_head
*head
;
4523 migration_req_t
*req
;
4527 spin_lock_irq(&rq
->lock
);
4529 if (cpu_is_offline(cpu
)) {
4530 spin_unlock_irq(&rq
->lock
);
4534 if (rq
->active_balance
) {
4535 active_load_balance(rq
, cpu
);
4536 rq
->active_balance
= 0;
4539 head
= &rq
->migration_queue
;
4541 if (list_empty(head
)) {
4542 spin_unlock_irq(&rq
->lock
);
4544 set_current_state(TASK_INTERRUPTIBLE
);
4547 req
= list_entry(head
->next
, migration_req_t
, list
);
4548 list_del_init(head
->next
);
4550 spin_unlock(&rq
->lock
);
4551 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
4554 complete(&req
->done
);
4556 __set_current_state(TASK_RUNNING
);
4560 /* Wait for kthread_stop */
4561 set_current_state(TASK_INTERRUPTIBLE
);
4562 while (!kthread_should_stop()) {
4564 set_current_state(TASK_INTERRUPTIBLE
);
4566 __set_current_state(TASK_RUNNING
);
4570 #ifdef CONFIG_HOTPLUG_CPU
4571 /* Figure out where task on dead CPU should go, use force if neccessary. */
4572 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*tsk
)
4578 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
4579 cpus_and(mask
, mask
, tsk
->cpus_allowed
);
4580 dest_cpu
= any_online_cpu(mask
);
4582 /* On any allowed CPU? */
4583 if (dest_cpu
== NR_CPUS
)
4584 dest_cpu
= any_online_cpu(tsk
->cpus_allowed
);
4586 /* No more Mr. Nice Guy. */
4587 if (dest_cpu
== NR_CPUS
) {
4588 cpus_setall(tsk
->cpus_allowed
);
4589 dest_cpu
= any_online_cpu(tsk
->cpus_allowed
);
4592 * Don't tell them about moving exiting tasks or
4593 * kernel threads (both mm NULL), since they never
4596 if (tsk
->mm
&& printk_ratelimit())
4597 printk(KERN_INFO
"process %d (%s) no "
4598 "longer affine to cpu%d\n",
4599 tsk
->pid
, tsk
->comm
, dead_cpu
);
4601 __migrate_task(tsk
, dead_cpu
, dest_cpu
);
4605 * While a dead CPU has no uninterruptible tasks queued at this point,
4606 * it might still have a nonzero ->nr_uninterruptible counter, because
4607 * for performance reasons the counter is not stricly tracking tasks to
4608 * their home CPUs. So we just add the counter to another CPU's counter,
4609 * to keep the global sum constant after CPU-down:
4611 static void migrate_nr_uninterruptible(runqueue_t
*rq_src
)
4613 runqueue_t
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
4614 unsigned long flags
;
4616 local_irq_save(flags
);
4617 double_rq_lock(rq_src
, rq_dest
);
4618 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
4619 rq_src
->nr_uninterruptible
= 0;
4620 double_rq_unlock(rq_src
, rq_dest
);
4621 local_irq_restore(flags
);
4624 /* Run through task list and migrate tasks from the dead cpu. */
4625 static void migrate_live_tasks(int src_cpu
)
4627 struct task_struct
*tsk
, *t
;
4629 write_lock_irq(&tasklist_lock
);
4631 do_each_thread(t
, tsk
) {
4635 if (task_cpu(tsk
) == src_cpu
)
4636 move_task_off_dead_cpu(src_cpu
, tsk
);
4637 } while_each_thread(t
, tsk
);
4639 write_unlock_irq(&tasklist_lock
);
4642 /* Schedules idle task to be the next runnable task on current CPU.
4643 * It does so by boosting its priority to highest possible and adding it to
4644 * the _front_ of runqueue. Used by CPU offline code.
4646 void sched_idle_next(void)
4648 int cpu
= smp_processor_id();
4649 runqueue_t
*rq
= this_rq();
4650 struct task_struct
*p
= rq
->idle
;
4651 unsigned long flags
;
4653 /* cpu has to be offline */
4654 BUG_ON(cpu_online(cpu
));
4656 /* Strictly not necessary since rest of the CPUs are stopped by now
4657 * and interrupts disabled on current cpu.
4659 spin_lock_irqsave(&rq
->lock
, flags
);
4661 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
4662 /* Add idle task to _front_ of it's priority queue */
4663 __activate_idle_task(p
, rq
);
4665 spin_unlock_irqrestore(&rq
->lock
, flags
);
4668 /* Ensures that the idle task is using init_mm right before its cpu goes
4671 void idle_task_exit(void)
4673 struct mm_struct
*mm
= current
->active_mm
;
4675 BUG_ON(cpu_online(smp_processor_id()));
4678 switch_mm(mm
, &init_mm
, current
);
4682 static void migrate_dead(unsigned int dead_cpu
, task_t
*tsk
)
4684 struct runqueue
*rq
= cpu_rq(dead_cpu
);
4686 /* Must be exiting, otherwise would be on tasklist. */
4687 BUG_ON(tsk
->exit_state
!= EXIT_ZOMBIE
&& tsk
->exit_state
!= EXIT_DEAD
);
4689 /* Cannot have done final schedule yet: would have vanished. */
4690 BUG_ON(tsk
->flags
& PF_DEAD
);
4692 get_task_struct(tsk
);
4695 * Drop lock around migration; if someone else moves it,
4696 * that's OK. No task can be added to this CPU, so iteration is
4699 spin_unlock_irq(&rq
->lock
);
4700 move_task_off_dead_cpu(dead_cpu
, tsk
);
4701 spin_lock_irq(&rq
->lock
);
4703 put_task_struct(tsk
);
4706 /* release_task() removes task from tasklist, so we won't find dead tasks. */
4707 static void migrate_dead_tasks(unsigned int dead_cpu
)
4710 struct runqueue
*rq
= cpu_rq(dead_cpu
);
4712 for (arr
= 0; arr
< 2; arr
++) {
4713 for (i
= 0; i
< MAX_PRIO
; i
++) {
4714 struct list_head
*list
= &rq
->arrays
[arr
].queue
[i
];
4715 while (!list_empty(list
))
4716 migrate_dead(dead_cpu
,
4717 list_entry(list
->next
, task_t
,
4722 #endif /* CONFIG_HOTPLUG_CPU */
4725 * migration_call - callback that gets triggered when a CPU is added.
4726 * Here we can start up the necessary migration thread for the new CPU.
4728 static int migration_call(struct notifier_block
*nfb
, unsigned long action
,
4731 int cpu
= (long)hcpu
;
4732 struct task_struct
*p
;
4733 struct runqueue
*rq
;
4734 unsigned long flags
;
4737 case CPU_UP_PREPARE
:
4738 p
= kthread_create(migration_thread
, hcpu
, "migration/%d",cpu
);
4741 p
->flags
|= PF_NOFREEZE
;
4742 kthread_bind(p
, cpu
);
4743 /* Must be high prio: stop_machine expects to yield to it. */
4744 rq
= task_rq_lock(p
, &flags
);
4745 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
4746 task_rq_unlock(rq
, &flags
);
4747 cpu_rq(cpu
)->migration_thread
= p
;
4750 /* Strictly unneccessary, as first user will wake it. */
4751 wake_up_process(cpu_rq(cpu
)->migration_thread
);
4753 #ifdef CONFIG_HOTPLUG_CPU
4754 case CPU_UP_CANCELED
:
4755 /* Unbind it from offline cpu so it can run. Fall thru. */
4756 kthread_bind(cpu_rq(cpu
)->migration_thread
,
4757 any_online_cpu(cpu_online_map
));
4758 kthread_stop(cpu_rq(cpu
)->migration_thread
);
4759 cpu_rq(cpu
)->migration_thread
= NULL
;
4762 migrate_live_tasks(cpu
);
4764 kthread_stop(rq
->migration_thread
);
4765 rq
->migration_thread
= NULL
;
4766 /* Idle task back to normal (off runqueue, low prio) */
4767 rq
= task_rq_lock(rq
->idle
, &flags
);
4768 deactivate_task(rq
->idle
, rq
);
4769 rq
->idle
->static_prio
= MAX_PRIO
;
4770 __setscheduler(rq
->idle
, SCHED_NORMAL
, 0);
4771 migrate_dead_tasks(cpu
);
4772 task_rq_unlock(rq
, &flags
);
4773 migrate_nr_uninterruptible(rq
);
4774 BUG_ON(rq
->nr_running
!= 0);
4776 /* No need to migrate the tasks: it was best-effort if
4777 * they didn't do lock_cpu_hotplug(). Just wake up
4778 * the requestors. */
4779 spin_lock_irq(&rq
->lock
);
4780 while (!list_empty(&rq
->migration_queue
)) {
4781 migration_req_t
*req
;
4782 req
= list_entry(rq
->migration_queue
.next
,
4783 migration_req_t
, list
);
4784 list_del_init(&req
->list
);
4785 complete(&req
->done
);
4787 spin_unlock_irq(&rq
->lock
);
4794 /* Register at highest priority so that task migration (migrate_all_tasks)
4795 * happens before everything else.
4797 static struct notifier_block __devinitdata migration_notifier
= {
4798 .notifier_call
= migration_call
,
4802 int __init
migration_init(void)
4804 void *cpu
= (void *)(long)smp_processor_id();
4805 /* Start one for boot CPU. */
4806 migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
4807 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
4808 register_cpu_notifier(&migration_notifier
);
4814 #undef SCHED_DOMAIN_DEBUG
4815 #ifdef SCHED_DOMAIN_DEBUG
4816 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
4821 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
4825 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
4830 struct sched_group
*group
= sd
->groups
;
4831 cpumask_t groupmask
;
4833 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
4834 cpus_clear(groupmask
);
4837 for (i
= 0; i
< level
+ 1; i
++)
4839 printk("domain %d: ", level
);
4841 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
4842 printk("does not load-balance\n");
4844 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain has parent");
4848 printk("span %s\n", str
);
4850 if (!cpu_isset(cpu
, sd
->span
))
4851 printk(KERN_ERR
"ERROR: domain->span does not contain CPU%d\n", cpu
);
4852 if (!cpu_isset(cpu
, group
->cpumask
))
4853 printk(KERN_ERR
"ERROR: domain->groups does not contain CPU%d\n", cpu
);
4856 for (i
= 0; i
< level
+ 2; i
++)
4862 printk(KERN_ERR
"ERROR: group is NULL\n");
4866 if (!group
->cpu_power
) {
4868 printk(KERN_ERR
"ERROR: domain->cpu_power not set\n");
4871 if (!cpus_weight(group
->cpumask
)) {
4873 printk(KERN_ERR
"ERROR: empty group\n");
4876 if (cpus_intersects(groupmask
, group
->cpumask
)) {
4878 printk(KERN_ERR
"ERROR: repeated CPUs\n");
4881 cpus_or(groupmask
, groupmask
, group
->cpumask
);
4883 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
4886 group
= group
->next
;
4887 } while (group
!= sd
->groups
);
4890 if (!cpus_equal(sd
->span
, groupmask
))
4891 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
4897 if (!cpus_subset(groupmask
, sd
->span
))
4898 printk(KERN_ERR
"ERROR: parent span is not a superset of domain->span\n");
4904 #define sched_domain_debug(sd, cpu) {}
4907 static int sd_degenerate(struct sched_domain
*sd
)
4909 if (cpus_weight(sd
->span
) == 1)
4912 /* Following flags need at least 2 groups */
4913 if (sd
->flags
& (SD_LOAD_BALANCE
|
4914 SD_BALANCE_NEWIDLE
|
4917 if (sd
->groups
!= sd
->groups
->next
)
4921 /* Following flags don't use groups */
4922 if (sd
->flags
& (SD_WAKE_IDLE
|
4930 static int sd_parent_degenerate(struct sched_domain
*sd
,
4931 struct sched_domain
*parent
)
4933 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
4935 if (sd_degenerate(parent
))
4938 if (!cpus_equal(sd
->span
, parent
->span
))
4941 /* Does parent contain flags not in child? */
4942 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
4943 if (cflags
& SD_WAKE_AFFINE
)
4944 pflags
&= ~SD_WAKE_BALANCE
;
4945 /* Flags needing groups don't count if only 1 group in parent */
4946 if (parent
->groups
== parent
->groups
->next
) {
4947 pflags
&= ~(SD_LOAD_BALANCE
|
4948 SD_BALANCE_NEWIDLE
|
4952 if (~cflags
& pflags
)
4959 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
4960 * hold the hotplug lock.
4962 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
4964 runqueue_t
*rq
= cpu_rq(cpu
);
4965 struct sched_domain
*tmp
;
4967 /* Remove the sched domains which do not contribute to scheduling. */
4968 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
4969 struct sched_domain
*parent
= tmp
->parent
;
4972 if (sd_parent_degenerate(tmp
, parent
))
4973 tmp
->parent
= parent
->parent
;
4976 if (sd
&& sd_degenerate(sd
))
4979 sched_domain_debug(sd
, cpu
);
4981 rcu_assign_pointer(rq
->sd
, sd
);
4984 /* cpus with isolated domains */
4985 static cpumask_t __devinitdata cpu_isolated_map
= CPU_MASK_NONE
;
4987 /* Setup the mask of cpus configured for isolated domains */
4988 static int __init
isolated_cpu_setup(char *str
)
4990 int ints
[NR_CPUS
], i
;
4992 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
4993 cpus_clear(cpu_isolated_map
);
4994 for (i
= 1; i
<= ints
[0]; i
++)
4995 if (ints
[i
] < NR_CPUS
)
4996 cpu_set(ints
[i
], cpu_isolated_map
);
5000 __setup ("isolcpus=", isolated_cpu_setup
);
5003 * init_sched_build_groups takes an array of groups, the cpumask we wish
5004 * to span, and a pointer to a function which identifies what group a CPU
5005 * belongs to. The return value of group_fn must be a valid index into the
5006 * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
5007 * keep track of groups covered with a cpumask_t).
5009 * init_sched_build_groups will build a circular linked list of the groups
5010 * covered by the given span, and will set each group's ->cpumask correctly,
5011 * and ->cpu_power to 0.
5013 static void init_sched_build_groups(struct sched_group groups
[], cpumask_t span
,
5014 int (*group_fn
)(int cpu
))
5016 struct sched_group
*first
= NULL
, *last
= NULL
;
5017 cpumask_t covered
= CPU_MASK_NONE
;
5020 for_each_cpu_mask(i
, span
) {
5021 int group
= group_fn(i
);
5022 struct sched_group
*sg
= &groups
[group
];
5025 if (cpu_isset(i
, covered
))
5028 sg
->cpumask
= CPU_MASK_NONE
;
5031 for_each_cpu_mask(j
, span
) {
5032 if (group_fn(j
) != group
)
5035 cpu_set(j
, covered
);
5036 cpu_set(j
, sg
->cpumask
);
5047 #define SD_NODES_PER_DOMAIN 16
5051 * find_next_best_node - find the next node to include in a sched_domain
5052 * @node: node whose sched_domain we're building
5053 * @used_nodes: nodes already in the sched_domain
5055 * Find the next node to include in a given scheduling domain. Simply
5056 * finds the closest node not already in the @used_nodes map.
5058 * Should use nodemask_t.
5060 static int find_next_best_node(int node
, unsigned long *used_nodes
)
5062 int i
, n
, val
, min_val
, best_node
= 0;
5066 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5067 /* Start at @node */
5068 n
= (node
+ i
) % MAX_NUMNODES
;
5070 if (!nr_cpus_node(n
))
5073 /* Skip already used nodes */
5074 if (test_bit(n
, used_nodes
))
5077 /* Simple min distance search */
5078 val
= node_distance(node
, n
);
5080 if (val
< min_val
) {
5086 set_bit(best_node
, used_nodes
);
5091 * sched_domain_node_span - get a cpumask for a node's sched_domain
5092 * @node: node whose cpumask we're constructing
5093 * @size: number of nodes to include in this span
5095 * Given a node, construct a good cpumask for its sched_domain to span. It
5096 * should be one that prevents unnecessary balancing, but also spreads tasks
5099 static cpumask_t
sched_domain_node_span(int node
)
5102 cpumask_t span
, nodemask
;
5103 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
5106 bitmap_zero(used_nodes
, MAX_NUMNODES
);
5108 nodemask
= node_to_cpumask(node
);
5109 cpus_or(span
, span
, nodemask
);
5110 set_bit(node
, used_nodes
);
5112 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5113 int next_node
= find_next_best_node(node
, used_nodes
);
5114 nodemask
= node_to_cpumask(next_node
);
5115 cpus_or(span
, span
, nodemask
);
5123 * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we
5124 * can switch it on easily if needed.
5126 #ifdef CONFIG_SCHED_SMT
5127 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
5128 static struct sched_group sched_group_cpus
[NR_CPUS
];
5129 static int cpu_to_cpu_group(int cpu
)
5135 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
5136 static struct sched_group sched_group_phys
[NR_CPUS
];
5137 static int cpu_to_phys_group(int cpu
)
5139 #ifdef CONFIG_SCHED_SMT
5140 return first_cpu(cpu_sibling_map
[cpu
]);
5148 * The init_sched_build_groups can't handle what we want to do with node
5149 * groups, so roll our own. Now each node has its own list of groups which
5150 * gets dynamically allocated.
5152 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
5153 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
5155 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
5156 static struct sched_group
*sched_group_allnodes_bycpu
[NR_CPUS
];
5158 static int cpu_to_allnodes_group(int cpu
)
5160 return cpu_to_node(cpu
);
5165 * Build sched domains for a given set of cpus and attach the sched domains
5166 * to the individual cpus
5168 void build_sched_domains(const cpumask_t
*cpu_map
)
5172 struct sched_group
**sched_group_nodes
= NULL
;
5173 struct sched_group
*sched_group_allnodes
= NULL
;
5176 * Allocate the per-node list of sched groups
5178 sched_group_nodes
= kmalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
5180 if (!sched_group_nodes
) {
5181 printk(KERN_WARNING
"Can not alloc sched group node list\n");
5184 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
5188 * Set up domains for cpus specified by the cpu_map.
5190 for_each_cpu_mask(i
, *cpu_map
) {
5192 struct sched_domain
*sd
= NULL
, *p
;
5193 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
5195 cpus_and(nodemask
, nodemask
, *cpu_map
);
5198 if (cpus_weight(*cpu_map
)
5199 > SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
5200 if (!sched_group_allnodes
) {
5201 sched_group_allnodes
5202 = kmalloc(sizeof(struct sched_group
)
5205 if (!sched_group_allnodes
) {
5207 "Can not alloc allnodes sched group\n");
5210 sched_group_allnodes_bycpu
[i
]
5211 = sched_group_allnodes
;
5213 sd
= &per_cpu(allnodes_domains
, i
);
5214 *sd
= SD_ALLNODES_INIT
;
5215 sd
->span
= *cpu_map
;
5216 group
= cpu_to_allnodes_group(i
);
5217 sd
->groups
= &sched_group_allnodes
[group
];
5222 sd
= &per_cpu(node_domains
, i
);
5224 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
5226 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
5230 sd
= &per_cpu(phys_domains
, i
);
5231 group
= cpu_to_phys_group(i
);
5233 sd
->span
= nodemask
;
5235 sd
->groups
= &sched_group_phys
[group
];
5237 #ifdef CONFIG_SCHED_SMT
5239 sd
= &per_cpu(cpu_domains
, i
);
5240 group
= cpu_to_cpu_group(i
);
5241 *sd
= SD_SIBLING_INIT
;
5242 sd
->span
= cpu_sibling_map
[i
];
5243 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
5245 sd
->groups
= &sched_group_cpus
[group
];
5249 #ifdef CONFIG_SCHED_SMT
5250 /* Set up CPU (sibling) groups */
5251 for_each_cpu_mask(i
, *cpu_map
) {
5252 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
5253 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
5254 if (i
!= first_cpu(this_sibling_map
))
5257 init_sched_build_groups(sched_group_cpus
, this_sibling_map
,
5262 /* Set up physical groups */
5263 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5264 cpumask_t nodemask
= node_to_cpumask(i
);
5266 cpus_and(nodemask
, nodemask
, *cpu_map
);
5267 if (cpus_empty(nodemask
))
5270 init_sched_build_groups(sched_group_phys
, nodemask
,
5271 &cpu_to_phys_group
);
5275 /* Set up node groups */
5276 if (sched_group_allnodes
)
5277 init_sched_build_groups(sched_group_allnodes
, *cpu_map
,
5278 &cpu_to_allnodes_group
);
5280 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5281 /* Set up node groups */
5282 struct sched_group
*sg
, *prev
;
5283 cpumask_t nodemask
= node_to_cpumask(i
);
5284 cpumask_t domainspan
;
5285 cpumask_t covered
= CPU_MASK_NONE
;
5288 cpus_and(nodemask
, nodemask
, *cpu_map
);
5289 if (cpus_empty(nodemask
)) {
5290 sched_group_nodes
[i
] = NULL
;
5294 domainspan
= sched_domain_node_span(i
);
5295 cpus_and(domainspan
, domainspan
, *cpu_map
);
5297 sg
= kmalloc(sizeof(struct sched_group
), GFP_KERNEL
);
5298 sched_group_nodes
[i
] = sg
;
5299 for_each_cpu_mask(j
, nodemask
) {
5300 struct sched_domain
*sd
;
5301 sd
= &per_cpu(node_domains
, j
);
5303 if (sd
->groups
== NULL
) {
5304 /* Turn off balancing if we have no groups */
5310 "Can not alloc domain group for node %d\n", i
);
5314 sg
->cpumask
= nodemask
;
5315 cpus_or(covered
, covered
, nodemask
);
5318 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
5319 cpumask_t tmp
, notcovered
;
5320 int n
= (i
+ j
) % MAX_NUMNODES
;
5322 cpus_complement(notcovered
, covered
);
5323 cpus_and(tmp
, notcovered
, *cpu_map
);
5324 cpus_and(tmp
, tmp
, domainspan
);
5325 if (cpus_empty(tmp
))
5328 nodemask
= node_to_cpumask(n
);
5329 cpus_and(tmp
, tmp
, nodemask
);
5330 if (cpus_empty(tmp
))
5333 sg
= kmalloc(sizeof(struct sched_group
), GFP_KERNEL
);
5336 "Can not alloc domain group for node %d\n", j
);
5341 cpus_or(covered
, covered
, tmp
);
5345 prev
->next
= sched_group_nodes
[i
];
5349 /* Calculate CPU power for physical packages and nodes */
5350 for_each_cpu_mask(i
, *cpu_map
) {
5352 struct sched_domain
*sd
;
5353 #ifdef CONFIG_SCHED_SMT
5354 sd
= &per_cpu(cpu_domains
, i
);
5355 power
= SCHED_LOAD_SCALE
;
5356 sd
->groups
->cpu_power
= power
;
5359 sd
= &per_cpu(phys_domains
, i
);
5360 power
= SCHED_LOAD_SCALE
+ SCHED_LOAD_SCALE
*
5361 (cpus_weight(sd
->groups
->cpumask
)-1) / 10;
5362 sd
->groups
->cpu_power
= power
;
5365 sd
= &per_cpu(allnodes_domains
, i
);
5367 power
= SCHED_LOAD_SCALE
+ SCHED_LOAD_SCALE
*
5368 (cpus_weight(sd
->groups
->cpumask
)-1) / 10;
5369 sd
->groups
->cpu_power
= power
;
5375 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5376 struct sched_group
*sg
= sched_group_nodes
[i
];
5382 for_each_cpu_mask(j
, sg
->cpumask
) {
5383 struct sched_domain
*sd
;
5386 sd
= &per_cpu(phys_domains
, j
);
5387 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
5389 * Only add "power" once for each
5394 power
= SCHED_LOAD_SCALE
+ SCHED_LOAD_SCALE
*
5395 (cpus_weight(sd
->groups
->cpumask
)-1) / 10;
5397 sg
->cpu_power
+= power
;
5400 if (sg
!= sched_group_nodes
[i
])
5405 /* Attach the domains */
5406 for_each_cpu_mask(i
, *cpu_map
) {
5407 struct sched_domain
*sd
;
5408 #ifdef CONFIG_SCHED_SMT
5409 sd
= &per_cpu(cpu_domains
, i
);
5411 sd
= &per_cpu(phys_domains
, i
);
5413 cpu_attach_domain(sd
, i
);
5417 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
5419 static void arch_init_sched_domains(const cpumask_t
*cpu_map
)
5421 cpumask_t cpu_default_map
;
5424 * Setup mask for cpus without special case scheduling requirements.
5425 * For now this just excludes isolated cpus, but could be used to
5426 * exclude other special cases in the future.
5428 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
5430 build_sched_domains(&cpu_default_map
);
5433 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
5439 for_each_cpu_mask(cpu
, *cpu_map
) {
5440 struct sched_group
*sched_group_allnodes
5441 = sched_group_allnodes_bycpu
[cpu
];
5442 struct sched_group
**sched_group_nodes
5443 = sched_group_nodes_bycpu
[cpu
];
5445 if (sched_group_allnodes
) {
5446 kfree(sched_group_allnodes
);
5447 sched_group_allnodes_bycpu
[cpu
] = NULL
;
5450 if (!sched_group_nodes
)
5453 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5454 cpumask_t nodemask
= node_to_cpumask(i
);
5455 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
5457 cpus_and(nodemask
, nodemask
, *cpu_map
);
5458 if (cpus_empty(nodemask
))
5468 if (oldsg
!= sched_group_nodes
[i
])
5471 kfree(sched_group_nodes
);
5472 sched_group_nodes_bycpu
[cpu
] = NULL
;
5478 * Detach sched domains from a group of cpus specified in cpu_map
5479 * These cpus will now be attached to the NULL domain
5481 static inline void detach_destroy_domains(const cpumask_t
*cpu_map
)
5485 for_each_cpu_mask(i
, *cpu_map
)
5486 cpu_attach_domain(NULL
, i
);
5487 synchronize_sched();
5488 arch_destroy_sched_domains(cpu_map
);
5492 * Partition sched domains as specified by the cpumasks below.
5493 * This attaches all cpus from the cpumasks to the NULL domain,
5494 * waits for a RCU quiescent period, recalculates sched
5495 * domain information and then attaches them back to the
5496 * correct sched domains
5497 * Call with hotplug lock held
5499 void partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
5501 cpumask_t change_map
;
5503 cpus_and(*partition1
, *partition1
, cpu_online_map
);
5504 cpus_and(*partition2
, *partition2
, cpu_online_map
);
5505 cpus_or(change_map
, *partition1
, *partition2
);
5507 /* Detach sched domains from all of the affected cpus */
5508 detach_destroy_domains(&change_map
);
5509 if (!cpus_empty(*partition1
))
5510 build_sched_domains(partition1
);
5511 if (!cpus_empty(*partition2
))
5512 build_sched_domains(partition2
);
5515 #ifdef CONFIG_HOTPLUG_CPU
5517 * Force a reinitialization of the sched domains hierarchy. The domains
5518 * and groups cannot be updated in place without racing with the balancing
5519 * code, so we temporarily attach all running cpus to the NULL domain
5520 * which will prevent rebalancing while the sched domains are recalculated.
5522 static int update_sched_domains(struct notifier_block
*nfb
,
5523 unsigned long action
, void *hcpu
)
5526 case CPU_UP_PREPARE
:
5527 case CPU_DOWN_PREPARE
:
5528 detach_destroy_domains(&cpu_online_map
);
5531 case CPU_UP_CANCELED
:
5532 case CPU_DOWN_FAILED
:
5536 * Fall through and re-initialise the domains.
5543 /* The hotplug lock is already held by cpu_up/cpu_down */
5544 arch_init_sched_domains(&cpu_online_map
);
5550 void __init
sched_init_smp(void)
5553 arch_init_sched_domains(&cpu_online_map
);
5554 unlock_cpu_hotplug();
5555 /* XXX: Theoretical race here - CPU may be hotplugged now */
5556 hotcpu_notifier(update_sched_domains
, 0);
5559 void __init
sched_init_smp(void)
5562 #endif /* CONFIG_SMP */
5564 int in_sched_functions(unsigned long addr
)
5566 /* Linker adds these: start and end of __sched functions */
5567 extern char __sched_text_start
[], __sched_text_end
[];
5568 return in_lock_functions(addr
) ||
5569 (addr
>= (unsigned long)__sched_text_start
5570 && addr
< (unsigned long)__sched_text_end
);
5573 void __init
sched_init(void)
5578 for (i
= 0; i
< NR_CPUS
; i
++) {
5579 prio_array_t
*array
;
5582 spin_lock_init(&rq
->lock
);
5584 rq
->active
= rq
->arrays
;
5585 rq
->expired
= rq
->arrays
+ 1;
5586 rq
->best_expired_prio
= MAX_PRIO
;
5590 for (j
= 1; j
< 3; j
++)
5591 rq
->cpu_load
[j
] = 0;
5592 rq
->active_balance
= 0;
5594 rq
->migration_thread
= NULL
;
5595 INIT_LIST_HEAD(&rq
->migration_queue
);
5597 atomic_set(&rq
->nr_iowait
, 0);
5599 for (j
= 0; j
< 2; j
++) {
5600 array
= rq
->arrays
+ j
;
5601 for (k
= 0; k
< MAX_PRIO
; k
++) {
5602 INIT_LIST_HEAD(array
->queue
+ k
);
5603 __clear_bit(k
, array
->bitmap
);
5605 // delimiter for bitsearch
5606 __set_bit(MAX_PRIO
, array
->bitmap
);
5611 * The boot idle thread does lazy MMU switching as well:
5613 atomic_inc(&init_mm
.mm_count
);
5614 enter_lazy_tlb(&init_mm
, current
);
5617 * Make us the idle thread. Technically, schedule() should not be
5618 * called from this thread, however somewhere below it might be,
5619 * but because we are the idle thread, we just pick up running again
5620 * when this runqueue becomes "idle".
5622 init_idle(current
, smp_processor_id());
5625 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5626 void __might_sleep(char *file
, int line
)
5628 #if defined(in_atomic)
5629 static unsigned long prev_jiffy
; /* ratelimiting */
5631 if ((in_atomic() || irqs_disabled()) &&
5632 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
5633 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
5635 prev_jiffy
= jiffies
;
5636 printk(KERN_ERR
"Debug: sleeping function called from invalid"
5637 " context at %s:%d\n", file
, line
);
5638 printk("in_atomic():%d, irqs_disabled():%d\n",
5639 in_atomic(), irqs_disabled());
5644 EXPORT_SYMBOL(__might_sleep
);
5647 #ifdef CONFIG_MAGIC_SYSRQ
5648 void normalize_rt_tasks(void)
5650 struct task_struct
*p
;
5651 prio_array_t
*array
;
5652 unsigned long flags
;
5655 read_lock_irq(&tasklist_lock
);
5656 for_each_process (p
) {
5660 rq
= task_rq_lock(p
, &flags
);
5664 deactivate_task(p
, task_rq(p
));
5665 __setscheduler(p
, SCHED_NORMAL
, 0);
5667 __activate_task(p
, task_rq(p
));
5668 resched_task(rq
->curr
);
5671 task_rq_unlock(rq
, &flags
);
5673 read_unlock_irq(&tasklist_lock
);
5676 #endif /* CONFIG_MAGIC_SYSRQ */
5680 * These functions are only useful for the IA64 MCA handling.
5682 * They can only be called when the whole system has been
5683 * stopped - every CPU needs to be quiescent, and no scheduling
5684 * activity can take place. Using them for anything else would
5685 * be a serious bug, and as a result, they aren't even visible
5686 * under any other configuration.
5690 * curr_task - return the current task for a given cpu.
5691 * @cpu: the processor in question.
5693 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
5695 task_t
*curr_task(int cpu
)
5697 return cpu_curr(cpu
);
5701 * set_curr_task - set the current task for a given cpu.
5702 * @cpu: the processor in question.
5703 * @p: the task pointer to set.
5705 * Description: This function must only be used when non-maskable interrupts
5706 * are serviced on a separate stack. It allows the architecture to switch the
5707 * notion of the current task on a cpu in a non-blocking manner. This function
5708 * must be called with all CPU's synchronized, and interrupts disabled, the
5709 * and caller must save the original value of the current task (see
5710 * curr_task() above) and restore that value before reenabling interrupts and
5711 * re-starting the system.
5713 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
5715 void set_curr_task(int cpu
, task_t
*p
)