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 static_prio
)
666 rq
->prio_bias
+= MAX_PRIO
- static_prio
;
669 static inline void dec_prio_bias(runqueue_t
*rq
, int static_prio
)
671 rq
->prio_bias
-= MAX_PRIO
- static_prio
;
674 static inline void inc_prio_bias(runqueue_t
*rq
, int static_prio
)
678 static inline void dec_prio_bias(runqueue_t
*rq
, int static_prio
)
683 static inline void inc_nr_running(task_t
*p
, runqueue_t
*rq
)
686 inc_prio_bias(rq
, p
->static_prio
);
689 static inline void dec_nr_running(task_t
*p
, runqueue_t
*rq
)
692 dec_prio_bias(rq
, p
->static_prio
);
696 * __activate_task - move a task to the runqueue.
698 static inline void __activate_task(task_t
*p
, runqueue_t
*rq
)
700 enqueue_task(p
, rq
->active
);
701 inc_nr_running(p
, rq
);
705 * __activate_idle_task - move idle task to the _front_ of runqueue.
707 static inline void __activate_idle_task(task_t
*p
, runqueue_t
*rq
)
709 enqueue_task_head(p
, rq
->active
);
710 inc_nr_running(p
, rq
);
713 static int recalc_task_prio(task_t
*p
, unsigned long long now
)
715 /* Caller must always ensure 'now >= p->timestamp' */
716 unsigned long long __sleep_time
= now
- p
->timestamp
;
717 unsigned long sleep_time
;
719 if (__sleep_time
> NS_MAX_SLEEP_AVG
)
720 sleep_time
= NS_MAX_SLEEP_AVG
;
722 sleep_time
= (unsigned long)__sleep_time
;
724 if (likely(sleep_time
> 0)) {
726 * User tasks that sleep a long time are categorised as
727 * idle and will get just interactive status to stay active &
728 * prevent them suddenly becoming cpu hogs and starving
731 if (p
->mm
&& p
->activated
!= -1 &&
732 sleep_time
> INTERACTIVE_SLEEP(p
)) {
733 p
->sleep_avg
= JIFFIES_TO_NS(MAX_SLEEP_AVG
-
737 * The lower the sleep avg a task has the more
738 * rapidly it will rise with sleep time.
740 sleep_time
*= (MAX_BONUS
- CURRENT_BONUS(p
)) ? : 1;
743 * Tasks waking from uninterruptible sleep are
744 * limited in their sleep_avg rise as they
745 * are likely to be waiting on I/O
747 if (p
->activated
== -1 && p
->mm
) {
748 if (p
->sleep_avg
>= INTERACTIVE_SLEEP(p
))
750 else if (p
->sleep_avg
+ sleep_time
>=
751 INTERACTIVE_SLEEP(p
)) {
752 p
->sleep_avg
= INTERACTIVE_SLEEP(p
);
758 * This code gives a bonus to interactive tasks.
760 * The boost works by updating the 'average sleep time'
761 * value here, based on ->timestamp. The more time a
762 * task spends sleeping, the higher the average gets -
763 * and the higher the priority boost gets as well.
765 p
->sleep_avg
+= sleep_time
;
767 if (p
->sleep_avg
> NS_MAX_SLEEP_AVG
)
768 p
->sleep_avg
= NS_MAX_SLEEP_AVG
;
772 return effective_prio(p
);
776 * activate_task - move a task to the runqueue and do priority recalculation
778 * Update all the scheduling statistics stuff. (sleep average
779 * calculation, priority modifiers, etc.)
781 static void activate_task(task_t
*p
, runqueue_t
*rq
, int local
)
783 unsigned long long now
;
788 /* Compensate for drifting sched_clock */
789 runqueue_t
*this_rq
= this_rq();
790 now
= (now
- this_rq
->timestamp_last_tick
)
791 + rq
->timestamp_last_tick
;
795 p
->prio
= recalc_task_prio(p
, now
);
798 * This checks to make sure it's not an uninterruptible task
799 * that is now waking up.
803 * Tasks which were woken up by interrupts (ie. hw events)
804 * are most likely of interactive nature. So we give them
805 * the credit of extending their sleep time to the period
806 * of time they spend on the runqueue, waiting for execution
807 * on a CPU, first time around:
813 * Normal first-time wakeups get a credit too for
814 * on-runqueue time, but it will be weighted down:
821 __activate_task(p
, rq
);
825 * deactivate_task - remove a task from the runqueue.
827 static void deactivate_task(struct task_struct
*p
, runqueue_t
*rq
)
829 dec_nr_running(p
, rq
);
830 dequeue_task(p
, p
->array
);
835 * resched_task - mark a task 'to be rescheduled now'.
837 * On UP this means the setting of the need_resched flag, on SMP it
838 * might also involve a cross-CPU call to trigger the scheduler on
842 static void resched_task(task_t
*p
)
844 int need_resched
, nrpolling
;
846 assert_spin_locked(&task_rq(p
)->lock
);
848 /* minimise the chance of sending an interrupt to poll_idle() */
849 nrpolling
= test_tsk_thread_flag(p
,TIF_POLLING_NRFLAG
);
850 need_resched
= test_and_set_tsk_thread_flag(p
,TIF_NEED_RESCHED
);
851 nrpolling
|= test_tsk_thread_flag(p
,TIF_POLLING_NRFLAG
);
853 if (!need_resched
&& !nrpolling
&& (task_cpu(p
) != smp_processor_id()))
854 smp_send_reschedule(task_cpu(p
));
857 static inline void resched_task(task_t
*p
)
859 set_tsk_need_resched(p
);
864 * task_curr - is this task currently executing on a CPU?
865 * @p: the task in question.
867 inline int task_curr(const task_t
*p
)
869 return cpu_curr(task_cpu(p
)) == p
;
874 struct list_head list
;
879 struct completion done
;
883 * The task's runqueue lock must be held.
884 * Returns true if you have to wait for migration thread.
886 static int migrate_task(task_t
*p
, int dest_cpu
, migration_req_t
*req
)
888 runqueue_t
*rq
= task_rq(p
);
891 * If the task is not on a runqueue (and not running), then
892 * it is sufficient to simply update the task's cpu field.
894 if (!p
->array
&& !task_running(rq
, p
)) {
895 set_task_cpu(p
, dest_cpu
);
899 init_completion(&req
->done
);
901 req
->dest_cpu
= dest_cpu
;
902 list_add(&req
->list
, &rq
->migration_queue
);
907 * wait_task_inactive - wait for a thread to unschedule.
909 * The caller must ensure that the task *will* unschedule sometime soon,
910 * else this function might spin for a *long* time. This function can't
911 * be called with interrupts off, or it may introduce deadlock with
912 * smp_call_function() if an IPI is sent by the same process we are
913 * waiting to become inactive.
915 void wait_task_inactive(task_t
*p
)
922 rq
= task_rq_lock(p
, &flags
);
923 /* Must be off runqueue entirely, not preempted. */
924 if (unlikely(p
->array
|| task_running(rq
, p
))) {
925 /* If it's preempted, we yield. It could be a while. */
926 preempted
= !task_running(rq
, p
);
927 task_rq_unlock(rq
, &flags
);
933 task_rq_unlock(rq
, &flags
);
937 * kick_process - kick a running thread to enter/exit the kernel
938 * @p: the to-be-kicked thread
940 * Cause a process which is running on another CPU to enter
941 * kernel-mode, without any delay. (to get signals handled.)
943 * NOTE: this function doesnt have to take the runqueue lock,
944 * because all it wants to ensure is that the remote task enters
945 * the kernel. If the IPI races and the task has been migrated
946 * to another CPU then no harm is done and the purpose has been
949 void kick_process(task_t
*p
)
955 if ((cpu
!= smp_processor_id()) && task_curr(p
))
956 smp_send_reschedule(cpu
);
961 * Return a low guess at the load of a migration-source cpu.
963 * We want to under-estimate the load of migration sources, to
964 * balance conservatively.
966 static inline unsigned long __source_load(int cpu
, int type
, enum idle_type idle
)
968 runqueue_t
*rq
= cpu_rq(cpu
);
969 unsigned long cpu_load
= rq
->cpu_load
[type
-1],
970 load_now
= rq
->nr_running
* SCHED_LOAD_SCALE
;
972 if (idle
== NOT_IDLE
) {
974 * If we are balancing busy runqueues the load is biased by
975 * priority to create 'nice' support across cpus.
977 cpu_load
*= rq
->prio_bias
;
978 load_now
*= rq
->prio_bias
;
984 return min(cpu_load
, load_now
);
987 static inline unsigned long source_load(int cpu
, int type
)
989 return __source_load(cpu
, type
, NOT_IDLE
);
993 * Return a high guess at the load of a migration-target cpu
995 static inline unsigned long __target_load(int cpu
, int type
, enum idle_type idle
)
997 runqueue_t
*rq
= cpu_rq(cpu
);
998 unsigned long cpu_load
= rq
->cpu_load
[type
-1],
999 load_now
= rq
->nr_running
* SCHED_LOAD_SCALE
;
1004 if (idle
== NOT_IDLE
) {
1005 cpu_load
*= rq
->prio_bias
;
1006 load_now
*= rq
->prio_bias
;
1008 return max(cpu_load
, load_now
);
1011 static inline unsigned long target_load(int cpu
, int type
)
1013 return __target_load(cpu
, type
, NOT_IDLE
);
1017 * find_idlest_group finds and returns the least busy CPU group within the
1020 static struct sched_group
*
1021 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1023 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1024 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1025 int load_idx
= sd
->forkexec_idx
;
1026 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1029 unsigned long load
, avg_load
;
1033 /* Skip over this group if it has no CPUs allowed */
1034 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1037 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1039 /* Tally up the load of all CPUs in the group */
1042 for_each_cpu_mask(i
, group
->cpumask
) {
1043 /* Bias balancing toward cpus of our domain */
1045 load
= source_load(i
, load_idx
);
1047 load
= target_load(i
, load_idx
);
1052 /* Adjust by relative CPU power of the group */
1053 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
1056 this_load
= avg_load
;
1058 } else if (avg_load
< min_load
) {
1059 min_load
= avg_load
;
1063 group
= group
->next
;
1064 } while (group
!= sd
->groups
);
1066 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1072 * find_idlest_queue - find the idlest runqueue among the cpus in group.
1075 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1078 unsigned long load
, min_load
= ULONG_MAX
;
1082 /* Traverse only the allowed CPUs */
1083 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1085 for_each_cpu_mask(i
, tmp
) {
1086 load
= source_load(i
, 0);
1088 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1098 * sched_balance_self: balance the current task (running on cpu) in domains
1099 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1102 * Balance, ie. select the least loaded group.
1104 * Returns the target CPU number, or the same CPU if no balancing is needed.
1106 * preempt must be disabled.
1108 static int sched_balance_self(int cpu
, int flag
)
1110 struct task_struct
*t
= current
;
1111 struct sched_domain
*tmp
, *sd
= NULL
;
1113 for_each_domain(cpu
, tmp
)
1114 if (tmp
->flags
& flag
)
1119 struct sched_group
*group
;
1124 group
= find_idlest_group(sd
, t
, cpu
);
1128 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1129 if (new_cpu
== -1 || new_cpu
== cpu
)
1132 /* Now try balancing at a lower domain level */
1136 weight
= cpus_weight(span
);
1137 for_each_domain(cpu
, tmp
) {
1138 if (weight
<= cpus_weight(tmp
->span
))
1140 if (tmp
->flags
& flag
)
1143 /* while loop will break here if sd == NULL */
1149 #endif /* CONFIG_SMP */
1152 * wake_idle() will wake a task on an idle cpu if task->cpu is
1153 * not idle and an idle cpu is available. The span of cpus to
1154 * search starts with cpus closest then further out as needed,
1155 * so we always favor a closer, idle cpu.
1157 * Returns the CPU we should wake onto.
1159 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1160 static int wake_idle(int cpu
, task_t
*p
)
1163 struct sched_domain
*sd
;
1169 for_each_domain(cpu
, sd
) {
1170 if (sd
->flags
& SD_WAKE_IDLE
) {
1171 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1172 for_each_cpu_mask(i
, tmp
) {
1183 static inline int wake_idle(int cpu
, task_t
*p
)
1190 * try_to_wake_up - wake up a thread
1191 * @p: the to-be-woken-up thread
1192 * @state: the mask of task states that can be woken
1193 * @sync: do a synchronous wakeup?
1195 * Put it on the run-queue if it's not already there. The "current"
1196 * thread is always on the run-queue (except when the actual
1197 * re-schedule is in progress), and as such you're allowed to do
1198 * the simpler "current->state = TASK_RUNNING" to mark yourself
1199 * runnable without the overhead of this.
1201 * returns failure only if the task is already active.
1203 static int try_to_wake_up(task_t
*p
, unsigned int state
, int sync
)
1205 int cpu
, this_cpu
, success
= 0;
1206 unsigned long flags
;
1210 unsigned long load
, this_load
;
1211 struct sched_domain
*sd
, *this_sd
= NULL
;
1215 rq
= task_rq_lock(p
, &flags
);
1216 old_state
= p
->state
;
1217 if (!(old_state
& state
))
1224 this_cpu
= smp_processor_id();
1227 if (unlikely(task_running(rq
, p
)))
1232 schedstat_inc(rq
, ttwu_cnt
);
1233 if (cpu
== this_cpu
) {
1234 schedstat_inc(rq
, ttwu_local
);
1238 for_each_domain(this_cpu
, sd
) {
1239 if (cpu_isset(cpu
, sd
->span
)) {
1240 schedstat_inc(sd
, ttwu_wake_remote
);
1246 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1250 * Check for affine wakeup and passive balancing possibilities.
1253 int idx
= this_sd
->wake_idx
;
1254 unsigned int imbalance
;
1256 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1258 load
= source_load(cpu
, idx
);
1259 this_load
= target_load(this_cpu
, idx
);
1261 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1263 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1264 unsigned long tl
= this_load
;
1266 * If sync wakeup then subtract the (maximum possible)
1267 * effect of the currently running task from the load
1268 * of the current CPU:
1271 tl
-= SCHED_LOAD_SCALE
;
1274 tl
+ target_load(cpu
, idx
) <= SCHED_LOAD_SCALE
) ||
1275 100*(tl
+ SCHED_LOAD_SCALE
) <= imbalance
*load
) {
1277 * This domain has SD_WAKE_AFFINE and
1278 * p is cache cold in this domain, and
1279 * there is no bad imbalance.
1281 schedstat_inc(this_sd
, ttwu_move_affine
);
1287 * Start passive balancing when half the imbalance_pct
1290 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1291 if (imbalance
*this_load
<= 100*load
) {
1292 schedstat_inc(this_sd
, ttwu_move_balance
);
1298 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1300 new_cpu
= wake_idle(new_cpu
, p
);
1301 if (new_cpu
!= cpu
) {
1302 set_task_cpu(p
, new_cpu
);
1303 task_rq_unlock(rq
, &flags
);
1304 /* might preempt at this point */
1305 rq
= task_rq_lock(p
, &flags
);
1306 old_state
= p
->state
;
1307 if (!(old_state
& state
))
1312 this_cpu
= smp_processor_id();
1317 #endif /* CONFIG_SMP */
1318 if (old_state
== TASK_UNINTERRUPTIBLE
) {
1319 rq
->nr_uninterruptible
--;
1321 * Tasks on involuntary sleep don't earn
1322 * sleep_avg beyond just interactive state.
1328 * Tasks that have marked their sleep as noninteractive get
1329 * woken up without updating their sleep average. (i.e. their
1330 * sleep is handled in a priority-neutral manner, no priority
1331 * boost and no penalty.)
1333 if (old_state
& TASK_NONINTERACTIVE
)
1334 __activate_task(p
, rq
);
1336 activate_task(p
, rq
, cpu
== this_cpu
);
1338 * Sync wakeups (i.e. those types of wakeups where the waker
1339 * has indicated that it will leave the CPU in short order)
1340 * don't trigger a preemption, if the woken up task will run on
1341 * this cpu. (in this case the 'I will reschedule' promise of
1342 * the waker guarantees that the freshly woken up task is going
1343 * to be considered on this CPU.)
1345 if (!sync
|| cpu
!= this_cpu
) {
1346 if (TASK_PREEMPTS_CURR(p
, rq
))
1347 resched_task(rq
->curr
);
1352 p
->state
= TASK_RUNNING
;
1354 task_rq_unlock(rq
, &flags
);
1359 int fastcall
wake_up_process(task_t
*p
)
1361 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1362 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1365 EXPORT_SYMBOL(wake_up_process
);
1367 int fastcall
wake_up_state(task_t
*p
, unsigned int state
)
1369 return try_to_wake_up(p
, state
, 0);
1373 * Perform scheduler related setup for a newly forked process p.
1374 * p is forked by current.
1376 void fastcall
sched_fork(task_t
*p
, int clone_flags
)
1378 int cpu
= get_cpu();
1381 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1383 set_task_cpu(p
, cpu
);
1386 * We mark the process as running here, but have not actually
1387 * inserted it onto the runqueue yet. This guarantees that
1388 * nobody will actually run it, and a signal or other external
1389 * event cannot wake it up and insert it on the runqueue either.
1391 p
->state
= TASK_RUNNING
;
1392 INIT_LIST_HEAD(&p
->run_list
);
1394 #ifdef CONFIG_SCHEDSTATS
1395 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1397 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1400 #ifdef CONFIG_PREEMPT
1401 /* Want to start with kernel preemption disabled. */
1402 p
->thread_info
->preempt_count
= 1;
1405 * Share the timeslice between parent and child, thus the
1406 * total amount of pending timeslices in the system doesn't change,
1407 * resulting in more scheduling fairness.
1409 local_irq_disable();
1410 p
->time_slice
= (current
->time_slice
+ 1) >> 1;
1412 * The remainder of the first timeslice might be recovered by
1413 * the parent if the child exits early enough.
1415 p
->first_time_slice
= 1;
1416 current
->time_slice
>>= 1;
1417 p
->timestamp
= sched_clock();
1418 if (unlikely(!current
->time_slice
)) {
1420 * This case is rare, it happens when the parent has only
1421 * a single jiffy left from its timeslice. Taking the
1422 * runqueue lock is not a problem.
1424 current
->time_slice
= 1;
1432 * wake_up_new_task - wake up a newly created task for the first time.
1434 * This function will do some initial scheduler statistics housekeeping
1435 * that must be done for every newly created context, then puts the task
1436 * on the runqueue and wakes it.
1438 void fastcall
wake_up_new_task(task_t
*p
, unsigned long clone_flags
)
1440 unsigned long flags
;
1442 runqueue_t
*rq
, *this_rq
;
1444 rq
= task_rq_lock(p
, &flags
);
1445 BUG_ON(p
->state
!= TASK_RUNNING
);
1446 this_cpu
= smp_processor_id();
1450 * We decrease the sleep average of forking parents
1451 * and children as well, to keep max-interactive tasks
1452 * from forking tasks that are max-interactive. The parent
1453 * (current) is done further down, under its lock.
1455 p
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(p
) *
1456 CHILD_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1458 p
->prio
= effective_prio(p
);
1460 if (likely(cpu
== this_cpu
)) {
1461 if (!(clone_flags
& CLONE_VM
)) {
1463 * The VM isn't cloned, so we're in a good position to
1464 * do child-runs-first in anticipation of an exec. This
1465 * usually avoids a lot of COW overhead.
1467 if (unlikely(!current
->array
))
1468 __activate_task(p
, rq
);
1470 p
->prio
= current
->prio
;
1471 list_add_tail(&p
->run_list
, ¤t
->run_list
);
1472 p
->array
= current
->array
;
1473 p
->array
->nr_active
++;
1474 inc_nr_running(p
, rq
);
1478 /* Run child last */
1479 __activate_task(p
, rq
);
1481 * We skip the following code due to cpu == this_cpu
1483 * task_rq_unlock(rq, &flags);
1484 * this_rq = task_rq_lock(current, &flags);
1488 this_rq
= cpu_rq(this_cpu
);
1491 * Not the local CPU - must adjust timestamp. This should
1492 * get optimised away in the !CONFIG_SMP case.
1494 p
->timestamp
= (p
->timestamp
- this_rq
->timestamp_last_tick
)
1495 + rq
->timestamp_last_tick
;
1496 __activate_task(p
, rq
);
1497 if (TASK_PREEMPTS_CURR(p
, rq
))
1498 resched_task(rq
->curr
);
1501 * Parent and child are on different CPUs, now get the
1502 * parent runqueue to update the parent's ->sleep_avg:
1504 task_rq_unlock(rq
, &flags
);
1505 this_rq
= task_rq_lock(current
, &flags
);
1507 current
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(current
) *
1508 PARENT_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1509 task_rq_unlock(this_rq
, &flags
);
1513 * Potentially available exiting-child timeslices are
1514 * retrieved here - this way the parent does not get
1515 * penalized for creating too many threads.
1517 * (this cannot be used to 'generate' timeslices
1518 * artificially, because any timeslice recovered here
1519 * was given away by the parent in the first place.)
1521 void fastcall
sched_exit(task_t
*p
)
1523 unsigned long flags
;
1527 * If the child was a (relative-) CPU hog then decrease
1528 * the sleep_avg of the parent as well.
1530 rq
= task_rq_lock(p
->parent
, &flags
);
1531 if (p
->first_time_slice
&& task_cpu(p
) == task_cpu(p
->parent
)) {
1532 p
->parent
->time_slice
+= p
->time_slice
;
1533 if (unlikely(p
->parent
->time_slice
> task_timeslice(p
)))
1534 p
->parent
->time_slice
= task_timeslice(p
);
1536 if (p
->sleep_avg
< p
->parent
->sleep_avg
)
1537 p
->parent
->sleep_avg
= p
->parent
->sleep_avg
/
1538 (EXIT_WEIGHT
+ 1) * EXIT_WEIGHT
+ p
->sleep_avg
/
1540 task_rq_unlock(rq
, &flags
);
1544 * prepare_task_switch - prepare to switch tasks
1545 * @rq: the runqueue preparing to switch
1546 * @next: the task we are going to switch to.
1548 * This is called with the rq lock held and interrupts off. It must
1549 * be paired with a subsequent finish_task_switch after the context
1552 * prepare_task_switch sets up locking and calls architecture specific
1555 static inline void prepare_task_switch(runqueue_t
*rq
, task_t
*next
)
1557 prepare_lock_switch(rq
, next
);
1558 prepare_arch_switch(next
);
1562 * finish_task_switch - clean up after a task-switch
1563 * @rq: runqueue associated with task-switch
1564 * @prev: the thread we just switched away from.
1566 * finish_task_switch must be called after the context switch, paired
1567 * with a prepare_task_switch call before the context switch.
1568 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1569 * and do any other architecture-specific cleanup actions.
1571 * Note that we may have delayed dropping an mm in context_switch(). If
1572 * so, we finish that here outside of the runqueue lock. (Doing it
1573 * with the lock held can cause deadlocks; see schedule() for
1576 static inline void finish_task_switch(runqueue_t
*rq
, task_t
*prev
)
1577 __releases(rq
->lock
)
1579 struct mm_struct
*mm
= rq
->prev_mm
;
1580 unsigned long prev_task_flags
;
1585 * A task struct has one reference for the use as "current".
1586 * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1587 * calls schedule one last time. The schedule call will never return,
1588 * and the scheduled task must drop that reference.
1589 * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1590 * still held, otherwise prev could be scheduled on another cpu, die
1591 * there before we look at prev->state, and then the reference would
1593 * Manfred Spraul <manfred@colorfullife.com>
1595 prev_task_flags
= prev
->flags
;
1596 finish_arch_switch(prev
);
1597 finish_lock_switch(rq
, prev
);
1600 if (unlikely(prev_task_flags
& PF_DEAD
))
1601 put_task_struct(prev
);
1605 * schedule_tail - first thing a freshly forked thread must call.
1606 * @prev: the thread we just switched away from.
1608 asmlinkage
void schedule_tail(task_t
*prev
)
1609 __releases(rq
->lock
)
1611 runqueue_t
*rq
= this_rq();
1612 finish_task_switch(rq
, prev
);
1613 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1614 /* In this case, finish_task_switch does not reenable preemption */
1617 if (current
->set_child_tid
)
1618 put_user(current
->pid
, current
->set_child_tid
);
1622 * context_switch - switch to the new MM and the new
1623 * thread's register state.
1626 task_t
* context_switch(runqueue_t
*rq
, task_t
*prev
, task_t
*next
)
1628 struct mm_struct
*mm
= next
->mm
;
1629 struct mm_struct
*oldmm
= prev
->active_mm
;
1631 if (unlikely(!mm
)) {
1632 next
->active_mm
= oldmm
;
1633 atomic_inc(&oldmm
->mm_count
);
1634 enter_lazy_tlb(oldmm
, next
);
1636 switch_mm(oldmm
, mm
, next
);
1638 if (unlikely(!prev
->mm
)) {
1639 prev
->active_mm
= NULL
;
1640 WARN_ON(rq
->prev_mm
);
1641 rq
->prev_mm
= oldmm
;
1644 /* Here we just switch the register state and the stack. */
1645 switch_to(prev
, next
, prev
);
1651 * nr_running, nr_uninterruptible and nr_context_switches:
1653 * externally visible scheduler statistics: current number of runnable
1654 * threads, current number of uninterruptible-sleeping threads, total
1655 * number of context switches performed since bootup.
1657 unsigned long nr_running(void)
1659 unsigned long i
, sum
= 0;
1661 for_each_online_cpu(i
)
1662 sum
+= cpu_rq(i
)->nr_running
;
1667 unsigned long nr_uninterruptible(void)
1669 unsigned long i
, sum
= 0;
1672 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1675 * Since we read the counters lockless, it might be slightly
1676 * inaccurate. Do not allow it to go below zero though:
1678 if (unlikely((long)sum
< 0))
1684 unsigned long long nr_context_switches(void)
1686 unsigned long long i
, sum
= 0;
1689 sum
+= cpu_rq(i
)->nr_switches
;
1694 unsigned long nr_iowait(void)
1696 unsigned long i
, sum
= 0;
1699 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1707 * double_rq_lock - safely lock two runqueues
1709 * Note this does not disable interrupts like task_rq_lock,
1710 * you need to do so manually before calling.
1712 static void double_rq_lock(runqueue_t
*rq1
, runqueue_t
*rq2
)
1713 __acquires(rq1
->lock
)
1714 __acquires(rq2
->lock
)
1717 spin_lock(&rq1
->lock
);
1718 __acquire(rq2
->lock
); /* Fake it out ;) */
1721 spin_lock(&rq1
->lock
);
1722 spin_lock(&rq2
->lock
);
1724 spin_lock(&rq2
->lock
);
1725 spin_lock(&rq1
->lock
);
1731 * double_rq_unlock - safely unlock two runqueues
1733 * Note this does not restore interrupts like task_rq_unlock,
1734 * you need to do so manually after calling.
1736 static void double_rq_unlock(runqueue_t
*rq1
, runqueue_t
*rq2
)
1737 __releases(rq1
->lock
)
1738 __releases(rq2
->lock
)
1740 spin_unlock(&rq1
->lock
);
1742 spin_unlock(&rq2
->lock
);
1744 __release(rq2
->lock
);
1748 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1750 static void double_lock_balance(runqueue_t
*this_rq
, runqueue_t
*busiest
)
1751 __releases(this_rq
->lock
)
1752 __acquires(busiest
->lock
)
1753 __acquires(this_rq
->lock
)
1755 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1756 if (busiest
< this_rq
) {
1757 spin_unlock(&this_rq
->lock
);
1758 spin_lock(&busiest
->lock
);
1759 spin_lock(&this_rq
->lock
);
1761 spin_lock(&busiest
->lock
);
1766 * If dest_cpu is allowed for this process, migrate the task to it.
1767 * This is accomplished by forcing the cpu_allowed mask to only
1768 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
1769 * the cpu_allowed mask is restored.
1771 static void sched_migrate_task(task_t
*p
, int dest_cpu
)
1773 migration_req_t req
;
1775 unsigned long flags
;
1777 rq
= task_rq_lock(p
, &flags
);
1778 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
1779 || unlikely(cpu_is_offline(dest_cpu
)))
1782 /* force the process onto the specified CPU */
1783 if (migrate_task(p
, dest_cpu
, &req
)) {
1784 /* Need to wait for migration thread (might exit: take ref). */
1785 struct task_struct
*mt
= rq
->migration_thread
;
1786 get_task_struct(mt
);
1787 task_rq_unlock(rq
, &flags
);
1788 wake_up_process(mt
);
1789 put_task_struct(mt
);
1790 wait_for_completion(&req
.done
);
1794 task_rq_unlock(rq
, &flags
);
1798 * sched_exec - execve() is a valuable balancing opportunity, because at
1799 * this point the task has the smallest effective memory and cache footprint.
1801 void sched_exec(void)
1803 int new_cpu
, this_cpu
= get_cpu();
1804 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
1806 if (new_cpu
!= this_cpu
)
1807 sched_migrate_task(current
, new_cpu
);
1811 * pull_task - move a task from a remote runqueue to the local runqueue.
1812 * Both runqueues must be locked.
1815 void pull_task(runqueue_t
*src_rq
, prio_array_t
*src_array
, task_t
*p
,
1816 runqueue_t
*this_rq
, prio_array_t
*this_array
, int this_cpu
)
1818 dequeue_task(p
, src_array
);
1819 dec_nr_running(p
, src_rq
);
1820 set_task_cpu(p
, this_cpu
);
1821 inc_nr_running(p
, this_rq
);
1822 enqueue_task(p
, this_array
);
1823 p
->timestamp
= (p
->timestamp
- src_rq
->timestamp_last_tick
)
1824 + this_rq
->timestamp_last_tick
;
1826 * Note that idle threads have a prio of MAX_PRIO, for this test
1827 * to be always true for them.
1829 if (TASK_PREEMPTS_CURR(p
, this_rq
))
1830 resched_task(this_rq
->curr
);
1834 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1837 int can_migrate_task(task_t
*p
, runqueue_t
*rq
, int this_cpu
,
1838 struct sched_domain
*sd
, enum idle_type idle
,
1842 * We do not migrate tasks that are:
1843 * 1) running (obviously), or
1844 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1845 * 3) are cache-hot on their current CPU.
1847 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
1851 if (task_running(rq
, p
))
1855 * Aggressive migration if:
1856 * 1) task is cache cold, or
1857 * 2) too many balance attempts have failed.
1860 if (sd
->nr_balance_failed
> sd
->cache_nice_tries
)
1863 if (task_hot(p
, rq
->timestamp_last_tick
, sd
))
1869 * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq,
1870 * as part of a balancing operation within "domain". Returns the number of
1873 * Called with both runqueues locked.
1875 static int move_tasks(runqueue_t
*this_rq
, int this_cpu
, runqueue_t
*busiest
,
1876 unsigned long max_nr_move
, struct sched_domain
*sd
,
1877 enum idle_type idle
, int *all_pinned
)
1879 prio_array_t
*array
, *dst_array
;
1880 struct list_head
*head
, *curr
;
1881 int idx
, pulled
= 0, pinned
= 0;
1884 if (max_nr_move
== 0)
1890 * We first consider expired tasks. Those will likely not be
1891 * executed in the near future, and they are most likely to
1892 * be cache-cold, thus switching CPUs has the least effect
1895 if (busiest
->expired
->nr_active
) {
1896 array
= busiest
->expired
;
1897 dst_array
= this_rq
->expired
;
1899 array
= busiest
->active
;
1900 dst_array
= this_rq
->active
;
1904 /* Start searching at priority 0: */
1908 idx
= sched_find_first_bit(array
->bitmap
);
1910 idx
= find_next_bit(array
->bitmap
, MAX_PRIO
, idx
);
1911 if (idx
>= MAX_PRIO
) {
1912 if (array
== busiest
->expired
&& busiest
->active
->nr_active
) {
1913 array
= busiest
->active
;
1914 dst_array
= this_rq
->active
;
1920 head
= array
->queue
+ idx
;
1923 tmp
= list_entry(curr
, task_t
, run_list
);
1927 if (!can_migrate_task(tmp
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
1934 #ifdef CONFIG_SCHEDSTATS
1935 if (task_hot(tmp
, busiest
->timestamp_last_tick
, sd
))
1936 schedstat_inc(sd
, lb_hot_gained
[idle
]);
1939 pull_task(busiest
, array
, tmp
, this_rq
, dst_array
, this_cpu
);
1942 /* We only want to steal up to the prescribed number of tasks. */
1943 if (pulled
< max_nr_move
) {
1951 * Right now, this is the only place pull_task() is called,
1952 * so we can safely collect pull_task() stats here rather than
1953 * inside pull_task().
1955 schedstat_add(sd
, lb_gained
[idle
], pulled
);
1958 *all_pinned
= pinned
;
1963 * find_busiest_group finds and returns the busiest CPU group within the
1964 * domain. It calculates and returns the number of tasks which should be
1965 * moved to restore balance via the imbalance parameter.
1967 static struct sched_group
*
1968 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
1969 unsigned long *imbalance
, enum idle_type idle
, int *sd_idle
)
1971 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1972 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
1973 unsigned long max_pull
;
1976 max_load
= this_load
= total_load
= total_pwr
= 0;
1977 if (idle
== NOT_IDLE
)
1978 load_idx
= sd
->busy_idx
;
1979 else if (idle
== NEWLY_IDLE
)
1980 load_idx
= sd
->newidle_idx
;
1982 load_idx
= sd
->idle_idx
;
1989 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1991 /* Tally up the load of all CPUs in the group */
1994 for_each_cpu_mask(i
, group
->cpumask
) {
1995 if (*sd_idle
&& !idle_cpu(i
))
1998 /* Bias balancing toward cpus of our domain */
2000 load
= __target_load(i
, load_idx
, idle
);
2002 load
= __source_load(i
, load_idx
, idle
);
2007 total_load
+= avg_load
;
2008 total_pwr
+= group
->cpu_power
;
2010 /* Adjust by relative CPU power of the group */
2011 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
2014 this_load
= avg_load
;
2016 } else if (avg_load
> max_load
) {
2017 max_load
= avg_load
;
2020 group
= group
->next
;
2021 } while (group
!= sd
->groups
);
2023 if (!busiest
|| this_load
>= max_load
|| max_load
<= SCHED_LOAD_SCALE
)
2026 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2028 if (this_load
>= avg_load
||
2029 100*max_load
<= sd
->imbalance_pct
*this_load
)
2033 * We're trying to get all the cpus to the average_load, so we don't
2034 * want to push ourselves above the average load, nor do we wish to
2035 * reduce the max loaded cpu below the average load, as either of these
2036 * actions would just result in more rebalancing later, and ping-pong
2037 * tasks around. Thus we look for the minimum possible imbalance.
2038 * Negative imbalances (*we* are more loaded than anyone else) will
2039 * be counted as no imbalance for these purposes -- we can't fix that
2040 * by pulling tasks to us. Be careful of negative numbers as they'll
2041 * appear as very large values with unsigned longs.
2044 /* Don't want to pull so many tasks that a group would go idle */
2045 max_pull
= min(max_load
- avg_load
, max_load
- SCHED_LOAD_SCALE
);
2047 /* How much load to actually move to equalise the imbalance */
2048 *imbalance
= min(max_pull
* busiest
->cpu_power
,
2049 (avg_load
- this_load
) * this->cpu_power
)
2052 if (*imbalance
< SCHED_LOAD_SCALE
) {
2053 unsigned long pwr_now
= 0, pwr_move
= 0;
2056 if (max_load
- this_load
>= SCHED_LOAD_SCALE
*2) {
2062 * OK, we don't have enough imbalance to justify moving tasks,
2063 * however we may be able to increase total CPU power used by
2067 pwr_now
+= busiest
->cpu_power
*min(SCHED_LOAD_SCALE
, max_load
);
2068 pwr_now
+= this->cpu_power
*min(SCHED_LOAD_SCALE
, this_load
);
2069 pwr_now
/= SCHED_LOAD_SCALE
;
2071 /* Amount of load we'd subtract */
2072 tmp
= SCHED_LOAD_SCALE
*SCHED_LOAD_SCALE
/busiest
->cpu_power
;
2074 pwr_move
+= busiest
->cpu_power
*min(SCHED_LOAD_SCALE
,
2077 /* Amount of load we'd add */
2078 if (max_load
*busiest
->cpu_power
<
2079 SCHED_LOAD_SCALE
*SCHED_LOAD_SCALE
)
2080 tmp
= max_load
*busiest
->cpu_power
/this->cpu_power
;
2082 tmp
= SCHED_LOAD_SCALE
*SCHED_LOAD_SCALE
/this->cpu_power
;
2083 pwr_move
+= this->cpu_power
*min(SCHED_LOAD_SCALE
, this_load
+ tmp
);
2084 pwr_move
/= SCHED_LOAD_SCALE
;
2086 /* Move if we gain throughput */
2087 if (pwr_move
<= pwr_now
)
2094 /* Get rid of the scaling factor, rounding down as we divide */
2095 *imbalance
= *imbalance
/ SCHED_LOAD_SCALE
;
2105 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2107 static runqueue_t
*find_busiest_queue(struct sched_group
*group
,
2108 enum idle_type idle
)
2110 unsigned long load
, max_load
= 0;
2111 runqueue_t
*busiest
= NULL
;
2114 for_each_cpu_mask(i
, group
->cpumask
) {
2115 load
= __source_load(i
, 0, idle
);
2117 if (load
> max_load
) {
2119 busiest
= cpu_rq(i
);
2127 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2128 * so long as it is large enough.
2130 #define MAX_PINNED_INTERVAL 512
2133 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2134 * tasks if there is an imbalance.
2136 * Called with this_rq unlocked.
2138 static int load_balance(int this_cpu
, runqueue_t
*this_rq
,
2139 struct sched_domain
*sd
, enum idle_type idle
)
2141 struct sched_group
*group
;
2142 runqueue_t
*busiest
;
2143 unsigned long imbalance
;
2144 int nr_moved
, all_pinned
= 0;
2145 int active_balance
= 0;
2148 if (idle
!= NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2151 schedstat_inc(sd
, lb_cnt
[idle
]);
2153 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
);
2155 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2159 busiest
= find_busiest_queue(group
, idle
);
2161 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2165 BUG_ON(busiest
== this_rq
);
2167 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2170 if (busiest
->nr_running
> 1) {
2172 * Attempt to move tasks. If find_busiest_group has found
2173 * an imbalance but busiest->nr_running <= 1, the group is
2174 * still unbalanced. nr_moved simply stays zero, so it is
2175 * correctly treated as an imbalance.
2177 double_rq_lock(this_rq
, busiest
);
2178 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2179 imbalance
, sd
, idle
, &all_pinned
);
2180 double_rq_unlock(this_rq
, busiest
);
2182 /* All tasks on this runqueue were pinned by CPU affinity */
2183 if (unlikely(all_pinned
))
2188 schedstat_inc(sd
, lb_failed
[idle
]);
2189 sd
->nr_balance_failed
++;
2191 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2193 spin_lock(&busiest
->lock
);
2195 /* don't kick the migration_thread, if the curr
2196 * task on busiest cpu can't be moved to this_cpu
2198 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2199 spin_unlock(&busiest
->lock
);
2201 goto out_one_pinned
;
2204 if (!busiest
->active_balance
) {
2205 busiest
->active_balance
= 1;
2206 busiest
->push_cpu
= this_cpu
;
2209 spin_unlock(&busiest
->lock
);
2211 wake_up_process(busiest
->migration_thread
);
2214 * We've kicked active balancing, reset the failure
2217 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2220 sd
->nr_balance_failed
= 0;
2222 if (likely(!active_balance
)) {
2223 /* We were unbalanced, so reset the balancing interval */
2224 sd
->balance_interval
= sd
->min_interval
;
2227 * If we've begun active balancing, start to back off. This
2228 * case may not be covered by the all_pinned logic if there
2229 * is only 1 task on the busy runqueue (because we don't call
2232 if (sd
->balance_interval
< sd
->max_interval
)
2233 sd
->balance_interval
*= 2;
2236 if (!nr_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2241 schedstat_inc(sd
, lb_balanced
[idle
]);
2243 sd
->nr_balance_failed
= 0;
2246 /* tune up the balancing interval */
2247 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2248 (sd
->balance_interval
< sd
->max_interval
))
2249 sd
->balance_interval
*= 2;
2251 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2257 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2258 * tasks if there is an imbalance.
2260 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2261 * this_rq is locked.
2263 static int load_balance_newidle(int this_cpu
, runqueue_t
*this_rq
,
2264 struct sched_domain
*sd
)
2266 struct sched_group
*group
;
2267 runqueue_t
*busiest
= NULL
;
2268 unsigned long imbalance
;
2272 if (sd
->flags
& SD_SHARE_CPUPOWER
)
2275 schedstat_inc(sd
, lb_cnt
[NEWLY_IDLE
]);
2276 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, NEWLY_IDLE
, &sd_idle
);
2278 schedstat_inc(sd
, lb_nobusyg
[NEWLY_IDLE
]);
2282 busiest
= find_busiest_queue(group
, NEWLY_IDLE
);
2284 schedstat_inc(sd
, lb_nobusyq
[NEWLY_IDLE
]);
2288 BUG_ON(busiest
== this_rq
);
2290 schedstat_add(sd
, lb_imbalance
[NEWLY_IDLE
], imbalance
);
2293 if (busiest
->nr_running
> 1) {
2294 /* Attempt to move tasks */
2295 double_lock_balance(this_rq
, busiest
);
2296 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2297 imbalance
, sd
, NEWLY_IDLE
, NULL
);
2298 spin_unlock(&busiest
->lock
);
2302 schedstat_inc(sd
, lb_failed
[NEWLY_IDLE
]);
2303 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2306 sd
->nr_balance_failed
= 0;
2311 schedstat_inc(sd
, lb_balanced
[NEWLY_IDLE
]);
2312 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2314 sd
->nr_balance_failed
= 0;
2319 * idle_balance is called by schedule() if this_cpu is about to become
2320 * idle. Attempts to pull tasks from other CPUs.
2322 static inline void idle_balance(int this_cpu
, runqueue_t
*this_rq
)
2324 struct sched_domain
*sd
;
2326 for_each_domain(this_cpu
, sd
) {
2327 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
2328 if (load_balance_newidle(this_cpu
, this_rq
, sd
)) {
2329 /* We've pulled tasks over so stop searching */
2337 * active_load_balance is run by migration threads. It pushes running tasks
2338 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2339 * running on each physical CPU where possible, and avoids physical /
2340 * logical imbalances.
2342 * Called with busiest_rq locked.
2344 static void active_load_balance(runqueue_t
*busiest_rq
, int busiest_cpu
)
2346 struct sched_domain
*sd
;
2347 runqueue_t
*target_rq
;
2348 int target_cpu
= busiest_rq
->push_cpu
;
2350 if (busiest_rq
->nr_running
<= 1)
2351 /* no task to move */
2354 target_rq
= cpu_rq(target_cpu
);
2357 * This condition is "impossible", if it occurs
2358 * we need to fix it. Originally reported by
2359 * Bjorn Helgaas on a 128-cpu setup.
2361 BUG_ON(busiest_rq
== target_rq
);
2363 /* move a task from busiest_rq to target_rq */
2364 double_lock_balance(busiest_rq
, target_rq
);
2366 /* Search for an sd spanning us and the target CPU. */
2367 for_each_domain(target_cpu
, sd
)
2368 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2369 cpu_isset(busiest_cpu
, sd
->span
))
2372 if (unlikely(sd
== NULL
))
2375 schedstat_inc(sd
, alb_cnt
);
2377 if (move_tasks(target_rq
, target_cpu
, busiest_rq
, 1, sd
, SCHED_IDLE
, NULL
))
2378 schedstat_inc(sd
, alb_pushed
);
2380 schedstat_inc(sd
, alb_failed
);
2382 spin_unlock(&target_rq
->lock
);
2386 * rebalance_tick will get called every timer tick, on every CPU.
2388 * It checks each scheduling domain to see if it is due to be balanced,
2389 * and initiates a balancing operation if so.
2391 * Balancing parameters are set up in arch_init_sched_domains.
2394 /* Don't have all balancing operations going off at once */
2395 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2397 static void rebalance_tick(int this_cpu
, runqueue_t
*this_rq
,
2398 enum idle_type idle
)
2400 unsigned long old_load
, this_load
;
2401 unsigned long j
= jiffies
+ CPU_OFFSET(this_cpu
);
2402 struct sched_domain
*sd
;
2405 this_load
= this_rq
->nr_running
* SCHED_LOAD_SCALE
;
2406 /* Update our load */
2407 for (i
= 0; i
< 3; i
++) {
2408 unsigned long new_load
= this_load
;
2410 old_load
= this_rq
->cpu_load
[i
];
2412 * Round up the averaging division if load is increasing. This
2413 * prevents us from getting stuck on 9 if the load is 10, for
2416 if (new_load
> old_load
)
2417 new_load
+= scale
-1;
2418 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) / scale
;
2421 for_each_domain(this_cpu
, sd
) {
2422 unsigned long interval
;
2424 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2427 interval
= sd
->balance_interval
;
2428 if (idle
!= SCHED_IDLE
)
2429 interval
*= sd
->busy_factor
;
2431 /* scale ms to jiffies */
2432 interval
= msecs_to_jiffies(interval
);
2433 if (unlikely(!interval
))
2436 if (j
- sd
->last_balance
>= interval
) {
2437 if (load_balance(this_cpu
, this_rq
, sd
, idle
)) {
2439 * We've pulled tasks over so either we're no
2440 * longer idle, or one of our SMT siblings is
2445 sd
->last_balance
+= interval
;
2451 * on UP we do not need to balance between CPUs:
2453 static inline void rebalance_tick(int cpu
, runqueue_t
*rq
, enum idle_type idle
)
2456 static inline void idle_balance(int cpu
, runqueue_t
*rq
)
2461 static inline int wake_priority_sleeper(runqueue_t
*rq
)
2464 #ifdef CONFIG_SCHED_SMT
2465 spin_lock(&rq
->lock
);
2467 * If an SMT sibling task has been put to sleep for priority
2468 * reasons reschedule the idle task to see if it can now run.
2470 if (rq
->nr_running
) {
2471 resched_task(rq
->idle
);
2474 spin_unlock(&rq
->lock
);
2479 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2481 EXPORT_PER_CPU_SYMBOL(kstat
);
2484 * This is called on clock ticks and on context switches.
2485 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2487 static inline void update_cpu_clock(task_t
*p
, runqueue_t
*rq
,
2488 unsigned long long now
)
2490 unsigned long long last
= max(p
->timestamp
, rq
->timestamp_last_tick
);
2491 p
->sched_time
+= now
- last
;
2495 * Return current->sched_time plus any more ns on the sched_clock
2496 * that have not yet been banked.
2498 unsigned long long current_sched_time(const task_t
*tsk
)
2500 unsigned long long ns
;
2501 unsigned long flags
;
2502 local_irq_save(flags
);
2503 ns
= max(tsk
->timestamp
, task_rq(tsk
)->timestamp_last_tick
);
2504 ns
= tsk
->sched_time
+ (sched_clock() - ns
);
2505 local_irq_restore(flags
);
2510 * We place interactive tasks back into the active array, if possible.
2512 * To guarantee that this does not starve expired tasks we ignore the
2513 * interactivity of a task if the first expired task had to wait more
2514 * than a 'reasonable' amount of time. This deadline timeout is
2515 * load-dependent, as the frequency of array switched decreases with
2516 * increasing number of running tasks. We also ignore the interactivity
2517 * if a better static_prio task has expired:
2519 #define EXPIRED_STARVING(rq) \
2520 ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2521 (jiffies - (rq)->expired_timestamp >= \
2522 STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2523 ((rq)->curr->static_prio > (rq)->best_expired_prio))
2526 * Account user cpu time to a process.
2527 * @p: the process that the cpu time gets accounted to
2528 * @hardirq_offset: the offset to subtract from hardirq_count()
2529 * @cputime: the cpu time spent in user space since the last update
2531 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
2533 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2536 p
->utime
= cputime_add(p
->utime
, cputime
);
2538 /* Add user time to cpustat. */
2539 tmp
= cputime_to_cputime64(cputime
);
2540 if (TASK_NICE(p
) > 0)
2541 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
2543 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
2547 * Account system cpu time to a process.
2548 * @p: the process that the cpu time gets accounted to
2549 * @hardirq_offset: the offset to subtract from hardirq_count()
2550 * @cputime: the cpu time spent in kernel space since the last update
2552 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
2555 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2556 runqueue_t
*rq
= this_rq();
2559 p
->stime
= cputime_add(p
->stime
, cputime
);
2561 /* Add system time to cpustat. */
2562 tmp
= cputime_to_cputime64(cputime
);
2563 if (hardirq_count() - hardirq_offset
)
2564 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
2565 else if (softirq_count())
2566 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
2567 else if (p
!= rq
->idle
)
2568 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
2569 else if (atomic_read(&rq
->nr_iowait
) > 0)
2570 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
2572 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
2573 /* Account for system time used */
2574 acct_update_integrals(p
);
2578 * Account for involuntary wait time.
2579 * @p: the process from which the cpu time has been stolen
2580 * @steal: the cpu time spent in involuntary wait
2582 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
2584 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2585 cputime64_t tmp
= cputime_to_cputime64(steal
);
2586 runqueue_t
*rq
= this_rq();
2588 if (p
== rq
->idle
) {
2589 p
->stime
= cputime_add(p
->stime
, steal
);
2590 if (atomic_read(&rq
->nr_iowait
) > 0)
2591 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
2593 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
2595 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
2599 * This function gets called by the timer code, with HZ frequency.
2600 * We call it with interrupts disabled.
2602 * It also gets called by the fork code, when changing the parent's
2605 void scheduler_tick(void)
2607 int cpu
= smp_processor_id();
2608 runqueue_t
*rq
= this_rq();
2609 task_t
*p
= current
;
2610 unsigned long long now
= sched_clock();
2612 update_cpu_clock(p
, rq
, now
);
2614 rq
->timestamp_last_tick
= now
;
2616 if (p
== rq
->idle
) {
2617 if (wake_priority_sleeper(rq
))
2619 rebalance_tick(cpu
, rq
, SCHED_IDLE
);
2623 /* Task might have expired already, but not scheduled off yet */
2624 if (p
->array
!= rq
->active
) {
2625 set_tsk_need_resched(p
);
2628 spin_lock(&rq
->lock
);
2630 * The task was running during this tick - update the
2631 * time slice counter. Note: we do not update a thread's
2632 * priority until it either goes to sleep or uses up its
2633 * timeslice. This makes it possible for interactive tasks
2634 * to use up their timeslices at their highest priority levels.
2638 * RR tasks need a special form of timeslice management.
2639 * FIFO tasks have no timeslices.
2641 if ((p
->policy
== SCHED_RR
) && !--p
->time_slice
) {
2642 p
->time_slice
= task_timeslice(p
);
2643 p
->first_time_slice
= 0;
2644 set_tsk_need_resched(p
);
2646 /* put it at the end of the queue: */
2647 requeue_task(p
, rq
->active
);
2651 if (!--p
->time_slice
) {
2652 dequeue_task(p
, rq
->active
);
2653 set_tsk_need_resched(p
);
2654 p
->prio
= effective_prio(p
);
2655 p
->time_slice
= task_timeslice(p
);
2656 p
->first_time_slice
= 0;
2658 if (!rq
->expired_timestamp
)
2659 rq
->expired_timestamp
= jiffies
;
2660 if (!TASK_INTERACTIVE(p
) || EXPIRED_STARVING(rq
)) {
2661 enqueue_task(p
, rq
->expired
);
2662 if (p
->static_prio
< rq
->best_expired_prio
)
2663 rq
->best_expired_prio
= p
->static_prio
;
2665 enqueue_task(p
, rq
->active
);
2668 * Prevent a too long timeslice allowing a task to monopolize
2669 * the CPU. We do this by splitting up the timeslice into
2672 * Note: this does not mean the task's timeslices expire or
2673 * get lost in any way, they just might be preempted by
2674 * another task of equal priority. (one with higher
2675 * priority would have preempted this task already.) We
2676 * requeue this task to the end of the list on this priority
2677 * level, which is in essence a round-robin of tasks with
2680 * This only applies to tasks in the interactive
2681 * delta range with at least TIMESLICE_GRANULARITY to requeue.
2683 if (TASK_INTERACTIVE(p
) && !((task_timeslice(p
) -
2684 p
->time_slice
) % TIMESLICE_GRANULARITY(p
)) &&
2685 (p
->time_slice
>= TIMESLICE_GRANULARITY(p
)) &&
2686 (p
->array
== rq
->active
)) {
2688 requeue_task(p
, rq
->active
);
2689 set_tsk_need_resched(p
);
2693 spin_unlock(&rq
->lock
);
2695 rebalance_tick(cpu
, rq
, NOT_IDLE
);
2698 #ifdef CONFIG_SCHED_SMT
2699 static inline void wakeup_busy_runqueue(runqueue_t
*rq
)
2701 /* If an SMT runqueue is sleeping due to priority reasons wake it up */
2702 if (rq
->curr
== rq
->idle
&& rq
->nr_running
)
2703 resched_task(rq
->idle
);
2706 static inline void wake_sleeping_dependent(int this_cpu
, runqueue_t
*this_rq
)
2708 struct sched_domain
*tmp
, *sd
= NULL
;
2709 cpumask_t sibling_map
;
2712 for_each_domain(this_cpu
, tmp
)
2713 if (tmp
->flags
& SD_SHARE_CPUPOWER
)
2720 * Unlock the current runqueue because we have to lock in
2721 * CPU order to avoid deadlocks. Caller knows that we might
2722 * unlock. We keep IRQs disabled.
2724 spin_unlock(&this_rq
->lock
);
2726 sibling_map
= sd
->span
;
2728 for_each_cpu_mask(i
, sibling_map
)
2729 spin_lock(&cpu_rq(i
)->lock
);
2731 * We clear this CPU from the mask. This both simplifies the
2732 * inner loop and keps this_rq locked when we exit:
2734 cpu_clear(this_cpu
, sibling_map
);
2736 for_each_cpu_mask(i
, sibling_map
) {
2737 runqueue_t
*smt_rq
= cpu_rq(i
);
2739 wakeup_busy_runqueue(smt_rq
);
2742 for_each_cpu_mask(i
, sibling_map
)
2743 spin_unlock(&cpu_rq(i
)->lock
);
2745 * We exit with this_cpu's rq still held and IRQs
2751 * number of 'lost' timeslices this task wont be able to fully
2752 * utilize, if another task runs on a sibling. This models the
2753 * slowdown effect of other tasks running on siblings:
2755 static inline unsigned long smt_slice(task_t
*p
, struct sched_domain
*sd
)
2757 return p
->time_slice
* (100 - sd
->per_cpu_gain
) / 100;
2760 static inline int dependent_sleeper(int this_cpu
, runqueue_t
*this_rq
)
2762 struct sched_domain
*tmp
, *sd
= NULL
;
2763 cpumask_t sibling_map
;
2764 prio_array_t
*array
;
2768 for_each_domain(this_cpu
, tmp
)
2769 if (tmp
->flags
& SD_SHARE_CPUPOWER
)
2776 * The same locking rules and details apply as for
2777 * wake_sleeping_dependent():
2779 spin_unlock(&this_rq
->lock
);
2780 sibling_map
= sd
->span
;
2781 for_each_cpu_mask(i
, sibling_map
)
2782 spin_lock(&cpu_rq(i
)->lock
);
2783 cpu_clear(this_cpu
, sibling_map
);
2786 * Establish next task to be run - it might have gone away because
2787 * we released the runqueue lock above:
2789 if (!this_rq
->nr_running
)
2791 array
= this_rq
->active
;
2792 if (!array
->nr_active
)
2793 array
= this_rq
->expired
;
2794 BUG_ON(!array
->nr_active
);
2796 p
= list_entry(array
->queue
[sched_find_first_bit(array
->bitmap
)].next
,
2799 for_each_cpu_mask(i
, sibling_map
) {
2800 runqueue_t
*smt_rq
= cpu_rq(i
);
2801 task_t
*smt_curr
= smt_rq
->curr
;
2803 /* Kernel threads do not participate in dependent sleeping */
2804 if (!p
->mm
|| !smt_curr
->mm
|| rt_task(p
))
2805 goto check_smt_task
;
2808 * If a user task with lower static priority than the
2809 * running task on the SMT sibling is trying to schedule,
2810 * delay it till there is proportionately less timeslice
2811 * left of the sibling task to prevent a lower priority
2812 * task from using an unfair proportion of the
2813 * physical cpu's resources. -ck
2815 if (rt_task(smt_curr
)) {
2817 * With real time tasks we run non-rt tasks only
2818 * per_cpu_gain% of the time.
2820 if ((jiffies
% DEF_TIMESLICE
) >
2821 (sd
->per_cpu_gain
* DEF_TIMESLICE
/ 100))
2824 if (smt_curr
->static_prio
< p
->static_prio
&&
2825 !TASK_PREEMPTS_CURR(p
, smt_rq
) &&
2826 smt_slice(smt_curr
, sd
) > task_timeslice(p
))
2830 if ((!smt_curr
->mm
&& smt_curr
!= smt_rq
->idle
) ||
2834 wakeup_busy_runqueue(smt_rq
);
2839 * Reschedule a lower priority task on the SMT sibling for
2840 * it to be put to sleep, or wake it up if it has been put to
2841 * sleep for priority reasons to see if it should run now.
2844 if ((jiffies
% DEF_TIMESLICE
) >
2845 (sd
->per_cpu_gain
* DEF_TIMESLICE
/ 100))
2846 resched_task(smt_curr
);
2848 if (TASK_PREEMPTS_CURR(p
, smt_rq
) &&
2849 smt_slice(p
, sd
) > task_timeslice(smt_curr
))
2850 resched_task(smt_curr
);
2852 wakeup_busy_runqueue(smt_rq
);
2856 for_each_cpu_mask(i
, sibling_map
)
2857 spin_unlock(&cpu_rq(i
)->lock
);
2861 static inline void wake_sleeping_dependent(int this_cpu
, runqueue_t
*this_rq
)
2865 static inline int dependent_sleeper(int this_cpu
, runqueue_t
*this_rq
)
2871 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
2873 void fastcall
add_preempt_count(int val
)
2878 BUG_ON((preempt_count() < 0));
2879 preempt_count() += val
;
2881 * Spinlock count overflowing soon?
2883 BUG_ON((preempt_count() & PREEMPT_MASK
) >= PREEMPT_MASK
-10);
2885 EXPORT_SYMBOL(add_preempt_count
);
2887 void fastcall
sub_preempt_count(int val
)
2892 BUG_ON(val
> preempt_count());
2894 * Is the spinlock portion underflowing?
2896 BUG_ON((val
< PREEMPT_MASK
) && !(preempt_count() & PREEMPT_MASK
));
2897 preempt_count() -= val
;
2899 EXPORT_SYMBOL(sub_preempt_count
);
2904 * schedule() is the main scheduler function.
2906 asmlinkage
void __sched
schedule(void)
2909 task_t
*prev
, *next
;
2911 prio_array_t
*array
;
2912 struct list_head
*queue
;
2913 unsigned long long now
;
2914 unsigned long run_time
;
2915 int cpu
, idx
, new_prio
;
2918 * Test if we are atomic. Since do_exit() needs to call into
2919 * schedule() atomically, we ignore that path for now.
2920 * Otherwise, whine if we are scheduling when we should not be.
2922 if (likely(!current
->exit_state
)) {
2923 if (unlikely(in_atomic())) {
2924 printk(KERN_ERR
"scheduling while atomic: "
2926 current
->comm
, preempt_count(), current
->pid
);
2930 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2935 release_kernel_lock(prev
);
2936 need_resched_nonpreemptible
:
2940 * The idle thread is not allowed to schedule!
2941 * Remove this check after it has been exercised a bit.
2943 if (unlikely(prev
== rq
->idle
) && prev
->state
!= TASK_RUNNING
) {
2944 printk(KERN_ERR
"bad: scheduling from the idle thread!\n");
2948 schedstat_inc(rq
, sched_cnt
);
2949 now
= sched_clock();
2950 if (likely((long long)(now
- prev
->timestamp
) < NS_MAX_SLEEP_AVG
)) {
2951 run_time
= now
- prev
->timestamp
;
2952 if (unlikely((long long)(now
- prev
->timestamp
) < 0))
2955 run_time
= NS_MAX_SLEEP_AVG
;
2958 * Tasks charged proportionately less run_time at high sleep_avg to
2959 * delay them losing their interactive status
2961 run_time
/= (CURRENT_BONUS(prev
) ? : 1);
2963 spin_lock_irq(&rq
->lock
);
2965 if (unlikely(prev
->flags
& PF_DEAD
))
2966 prev
->state
= EXIT_DEAD
;
2968 switch_count
= &prev
->nivcsw
;
2969 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
2970 switch_count
= &prev
->nvcsw
;
2971 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
2972 unlikely(signal_pending(prev
))))
2973 prev
->state
= TASK_RUNNING
;
2975 if (prev
->state
== TASK_UNINTERRUPTIBLE
)
2976 rq
->nr_uninterruptible
++;
2977 deactivate_task(prev
, rq
);
2981 cpu
= smp_processor_id();
2982 if (unlikely(!rq
->nr_running
)) {
2984 idle_balance(cpu
, rq
);
2985 if (!rq
->nr_running
) {
2987 rq
->expired_timestamp
= 0;
2988 wake_sleeping_dependent(cpu
, rq
);
2990 * wake_sleeping_dependent() might have released
2991 * the runqueue, so break out if we got new
2994 if (!rq
->nr_running
)
2998 if (dependent_sleeper(cpu
, rq
)) {
3003 * dependent_sleeper() releases and reacquires the runqueue
3004 * lock, hence go into the idle loop if the rq went
3007 if (unlikely(!rq
->nr_running
))
3012 if (unlikely(!array
->nr_active
)) {
3014 * Switch the active and expired arrays.
3016 schedstat_inc(rq
, sched_switch
);
3017 rq
->active
= rq
->expired
;
3018 rq
->expired
= array
;
3020 rq
->expired_timestamp
= 0;
3021 rq
->best_expired_prio
= MAX_PRIO
;
3024 idx
= sched_find_first_bit(array
->bitmap
);
3025 queue
= array
->queue
+ idx
;
3026 next
= list_entry(queue
->next
, task_t
, run_list
);
3028 if (!rt_task(next
) && next
->activated
> 0) {
3029 unsigned long long delta
= now
- next
->timestamp
;
3030 if (unlikely((long long)(now
- next
->timestamp
) < 0))
3033 if (next
->activated
== 1)
3034 delta
= delta
* (ON_RUNQUEUE_WEIGHT
* 128 / 100) / 128;
3036 array
= next
->array
;
3037 new_prio
= recalc_task_prio(next
, next
->timestamp
+ delta
);
3039 if (unlikely(next
->prio
!= new_prio
)) {
3040 dequeue_task(next
, array
);
3041 next
->prio
= new_prio
;
3042 enqueue_task(next
, array
);
3044 requeue_task(next
, array
);
3046 next
->activated
= 0;
3048 if (next
== rq
->idle
)
3049 schedstat_inc(rq
, sched_goidle
);
3051 prefetch_stack(next
);
3052 clear_tsk_need_resched(prev
);
3053 rcu_qsctr_inc(task_cpu(prev
));
3055 update_cpu_clock(prev
, rq
, now
);
3057 prev
->sleep_avg
-= run_time
;
3058 if ((long)prev
->sleep_avg
<= 0)
3059 prev
->sleep_avg
= 0;
3060 prev
->timestamp
= prev
->last_ran
= now
;
3062 sched_info_switch(prev
, next
);
3063 if (likely(prev
!= next
)) {
3064 next
->timestamp
= now
;
3069 prepare_task_switch(rq
, next
);
3070 prev
= context_switch(rq
, prev
, next
);
3073 * this_rq must be evaluated again because prev may have moved
3074 * CPUs since it called schedule(), thus the 'rq' on its stack
3075 * frame will be invalid.
3077 finish_task_switch(this_rq(), prev
);
3079 spin_unlock_irq(&rq
->lock
);
3082 if (unlikely(reacquire_kernel_lock(prev
) < 0))
3083 goto need_resched_nonpreemptible
;
3084 preempt_enable_no_resched();
3085 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3089 EXPORT_SYMBOL(schedule
);
3091 #ifdef CONFIG_PREEMPT
3093 * this is is the entry point to schedule() from in-kernel preemption
3094 * off of preempt_enable. Kernel preemptions off return from interrupt
3095 * occur there and call schedule directly.
3097 asmlinkage
void __sched
preempt_schedule(void)
3099 struct thread_info
*ti
= current_thread_info();
3100 #ifdef CONFIG_PREEMPT_BKL
3101 struct task_struct
*task
= current
;
3102 int saved_lock_depth
;
3105 * If there is a non-zero preempt_count or interrupts are disabled,
3106 * we do not want to preempt the current task. Just return..
3108 if (unlikely(ti
->preempt_count
|| irqs_disabled()))
3112 add_preempt_count(PREEMPT_ACTIVE
);
3114 * We keep the big kernel semaphore locked, but we
3115 * clear ->lock_depth so that schedule() doesnt
3116 * auto-release the semaphore:
3118 #ifdef CONFIG_PREEMPT_BKL
3119 saved_lock_depth
= task
->lock_depth
;
3120 task
->lock_depth
= -1;
3123 #ifdef CONFIG_PREEMPT_BKL
3124 task
->lock_depth
= saved_lock_depth
;
3126 sub_preempt_count(PREEMPT_ACTIVE
);
3128 /* we could miss a preemption opportunity between schedule and now */
3130 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3134 EXPORT_SYMBOL(preempt_schedule
);
3137 * this is is the entry point to schedule() from kernel preemption
3138 * off of irq context.
3139 * Note, that this is called and return with irqs disabled. This will
3140 * protect us against recursive calling from irq.
3142 asmlinkage
void __sched
preempt_schedule_irq(void)
3144 struct thread_info
*ti
= current_thread_info();
3145 #ifdef CONFIG_PREEMPT_BKL
3146 struct task_struct
*task
= current
;
3147 int saved_lock_depth
;
3149 /* Catch callers which need to be fixed*/
3150 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3153 add_preempt_count(PREEMPT_ACTIVE
);
3155 * We keep the big kernel semaphore locked, but we
3156 * clear ->lock_depth so that schedule() doesnt
3157 * auto-release the semaphore:
3159 #ifdef CONFIG_PREEMPT_BKL
3160 saved_lock_depth
= task
->lock_depth
;
3161 task
->lock_depth
= -1;
3165 local_irq_disable();
3166 #ifdef CONFIG_PREEMPT_BKL
3167 task
->lock_depth
= saved_lock_depth
;
3169 sub_preempt_count(PREEMPT_ACTIVE
);
3171 /* we could miss a preemption opportunity between schedule and now */
3173 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3177 #endif /* CONFIG_PREEMPT */
3179 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3182 task_t
*p
= curr
->private;
3183 return try_to_wake_up(p
, mode
, sync
);
3186 EXPORT_SYMBOL(default_wake_function
);
3189 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3190 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3191 * number) then we wake all the non-exclusive tasks and one exclusive task.
3193 * There are circumstances in which we can try to wake a task which has already
3194 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3195 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3197 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3198 int nr_exclusive
, int sync
, void *key
)
3200 struct list_head
*tmp
, *next
;
3202 list_for_each_safe(tmp
, next
, &q
->task_list
) {
3205 curr
= list_entry(tmp
, wait_queue_t
, task_list
);
3206 flags
= curr
->flags
;
3207 if (curr
->func(curr
, mode
, sync
, key
) &&
3208 (flags
& WQ_FLAG_EXCLUSIVE
) &&
3215 * __wake_up - wake up threads blocked on a waitqueue.
3217 * @mode: which threads
3218 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3219 * @key: is directly passed to the wakeup function
3221 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3222 int nr_exclusive
, void *key
)
3224 unsigned long flags
;
3226 spin_lock_irqsave(&q
->lock
, flags
);
3227 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3228 spin_unlock_irqrestore(&q
->lock
, flags
);
3231 EXPORT_SYMBOL(__wake_up
);
3234 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3236 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3238 __wake_up_common(q
, mode
, 1, 0, NULL
);
3242 * __wake_up_sync - wake up threads blocked on a waitqueue.
3244 * @mode: which threads
3245 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3247 * The sync wakeup differs that the waker knows that it will schedule
3248 * away soon, so while the target thread will be woken up, it will not
3249 * be migrated to another CPU - ie. the two threads are 'synchronized'
3250 * with each other. This can prevent needless bouncing between CPUs.
3252 * On UP it can prevent extra preemption.
3255 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3257 unsigned long flags
;
3263 if (unlikely(!nr_exclusive
))
3266 spin_lock_irqsave(&q
->lock
, flags
);
3267 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3268 spin_unlock_irqrestore(&q
->lock
, flags
);
3270 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3272 void fastcall
complete(struct completion
*x
)
3274 unsigned long flags
;
3276 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3278 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3280 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3282 EXPORT_SYMBOL(complete
);
3284 void fastcall
complete_all(struct completion
*x
)
3286 unsigned long flags
;
3288 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3289 x
->done
+= UINT_MAX
/2;
3290 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3292 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3294 EXPORT_SYMBOL(complete_all
);
3296 void fastcall __sched
wait_for_completion(struct completion
*x
)
3299 spin_lock_irq(&x
->wait
.lock
);
3301 DECLARE_WAITQUEUE(wait
, current
);
3303 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3304 __add_wait_queue_tail(&x
->wait
, &wait
);
3306 __set_current_state(TASK_UNINTERRUPTIBLE
);
3307 spin_unlock_irq(&x
->wait
.lock
);
3309 spin_lock_irq(&x
->wait
.lock
);
3311 __remove_wait_queue(&x
->wait
, &wait
);
3314 spin_unlock_irq(&x
->wait
.lock
);
3316 EXPORT_SYMBOL(wait_for_completion
);
3318 unsigned long fastcall __sched
3319 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3323 spin_lock_irq(&x
->wait
.lock
);
3325 DECLARE_WAITQUEUE(wait
, current
);
3327 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3328 __add_wait_queue_tail(&x
->wait
, &wait
);
3330 __set_current_state(TASK_UNINTERRUPTIBLE
);
3331 spin_unlock_irq(&x
->wait
.lock
);
3332 timeout
= schedule_timeout(timeout
);
3333 spin_lock_irq(&x
->wait
.lock
);
3335 __remove_wait_queue(&x
->wait
, &wait
);
3339 __remove_wait_queue(&x
->wait
, &wait
);
3343 spin_unlock_irq(&x
->wait
.lock
);
3346 EXPORT_SYMBOL(wait_for_completion_timeout
);
3348 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3354 spin_lock_irq(&x
->wait
.lock
);
3356 DECLARE_WAITQUEUE(wait
, current
);
3358 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3359 __add_wait_queue_tail(&x
->wait
, &wait
);
3361 if (signal_pending(current
)) {
3363 __remove_wait_queue(&x
->wait
, &wait
);
3366 __set_current_state(TASK_INTERRUPTIBLE
);
3367 spin_unlock_irq(&x
->wait
.lock
);
3369 spin_lock_irq(&x
->wait
.lock
);
3371 __remove_wait_queue(&x
->wait
, &wait
);
3375 spin_unlock_irq(&x
->wait
.lock
);
3379 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3381 unsigned long fastcall __sched
3382 wait_for_completion_interruptible_timeout(struct completion
*x
,
3383 unsigned long timeout
)
3387 spin_lock_irq(&x
->wait
.lock
);
3389 DECLARE_WAITQUEUE(wait
, current
);
3391 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3392 __add_wait_queue_tail(&x
->wait
, &wait
);
3394 if (signal_pending(current
)) {
3395 timeout
= -ERESTARTSYS
;
3396 __remove_wait_queue(&x
->wait
, &wait
);
3399 __set_current_state(TASK_INTERRUPTIBLE
);
3400 spin_unlock_irq(&x
->wait
.lock
);
3401 timeout
= schedule_timeout(timeout
);
3402 spin_lock_irq(&x
->wait
.lock
);
3404 __remove_wait_queue(&x
->wait
, &wait
);
3408 __remove_wait_queue(&x
->wait
, &wait
);
3412 spin_unlock_irq(&x
->wait
.lock
);
3415 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3418 #define SLEEP_ON_VAR \
3419 unsigned long flags; \
3420 wait_queue_t wait; \
3421 init_waitqueue_entry(&wait, current);
3423 #define SLEEP_ON_HEAD \
3424 spin_lock_irqsave(&q->lock,flags); \
3425 __add_wait_queue(q, &wait); \
3426 spin_unlock(&q->lock);
3428 #define SLEEP_ON_TAIL \
3429 spin_lock_irq(&q->lock); \
3430 __remove_wait_queue(q, &wait); \
3431 spin_unlock_irqrestore(&q->lock, flags);
3433 void fastcall __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3437 current
->state
= TASK_INTERRUPTIBLE
;
3444 EXPORT_SYMBOL(interruptible_sleep_on
);
3446 long fastcall __sched
3447 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3451 current
->state
= TASK_INTERRUPTIBLE
;
3454 timeout
= schedule_timeout(timeout
);
3460 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3462 void fastcall __sched
sleep_on(wait_queue_head_t
*q
)
3466 current
->state
= TASK_UNINTERRUPTIBLE
;
3473 EXPORT_SYMBOL(sleep_on
);
3475 long fastcall __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3479 current
->state
= TASK_UNINTERRUPTIBLE
;
3482 timeout
= schedule_timeout(timeout
);
3488 EXPORT_SYMBOL(sleep_on_timeout
);
3490 void set_user_nice(task_t
*p
, long nice
)
3492 unsigned long flags
;
3493 prio_array_t
*array
;
3495 int old_prio
, new_prio
, delta
;
3497 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3500 * We have to be careful, if called from sys_setpriority(),
3501 * the task might be in the middle of scheduling on another CPU.
3503 rq
= task_rq_lock(p
, &flags
);
3505 * The RT priorities are set via sched_setscheduler(), but we still
3506 * allow the 'normal' nice value to be set - but as expected
3507 * it wont have any effect on scheduling until the task is
3511 dec_prio_bias(rq
, p
->static_prio
);
3512 p
->static_prio
= NICE_TO_PRIO(nice
);
3513 inc_prio_bias(rq
, p
->static_prio
);
3518 dequeue_task(p
, array
);
3521 new_prio
= NICE_TO_PRIO(nice
);
3522 delta
= new_prio
- old_prio
;
3523 dec_prio_bias(rq
, p
->static_prio
);
3524 p
->static_prio
= NICE_TO_PRIO(nice
);
3525 inc_prio_bias(rq
, p
->static_prio
);
3529 enqueue_task(p
, array
);
3531 * If the task increased its priority or is running and
3532 * lowered its priority, then reschedule its CPU:
3534 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3535 resched_task(rq
->curr
);
3538 task_rq_unlock(rq
, &flags
);
3541 EXPORT_SYMBOL(set_user_nice
);
3544 * can_nice - check if a task can reduce its nice value
3548 int can_nice(const task_t
*p
, const int nice
)
3550 /* convert nice value [19,-20] to rlimit style value [1,40] */
3551 int nice_rlim
= 20 - nice
;
3552 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
3553 capable(CAP_SYS_NICE
));
3556 #ifdef __ARCH_WANT_SYS_NICE
3559 * sys_nice - change the priority of the current process.
3560 * @increment: priority increment
3562 * sys_setpriority is a more generic, but much slower function that
3563 * does similar things.
3565 asmlinkage
long sys_nice(int increment
)
3571 * Setpriority might change our priority at the same moment.
3572 * We don't have to worry. Conceptually one call occurs first
3573 * and we have a single winner.
3575 if (increment
< -40)
3580 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
3586 if (increment
< 0 && !can_nice(current
, nice
))
3589 retval
= security_task_setnice(current
, nice
);
3593 set_user_nice(current
, nice
);
3600 * task_prio - return the priority value of a given task.
3601 * @p: the task in question.
3603 * This is the priority value as seen by users in /proc.
3604 * RT tasks are offset by -200. Normal tasks are centered
3605 * around 0, value goes from -16 to +15.
3607 int task_prio(const task_t
*p
)
3609 return p
->prio
- MAX_RT_PRIO
;
3613 * task_nice - return the nice value of a given task.
3614 * @p: the task in question.
3616 int task_nice(const task_t
*p
)
3618 return TASK_NICE(p
);
3620 EXPORT_SYMBOL_GPL(task_nice
);
3623 * idle_cpu - is a given cpu idle currently?
3624 * @cpu: the processor in question.
3626 int idle_cpu(int cpu
)
3628 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
3632 * idle_task - return the idle task for a given cpu.
3633 * @cpu: the processor in question.
3635 task_t
*idle_task(int cpu
)
3637 return cpu_rq(cpu
)->idle
;
3641 * find_process_by_pid - find a process with a matching PID value.
3642 * @pid: the pid in question.
3644 static inline task_t
*find_process_by_pid(pid_t pid
)
3646 return pid
? find_task_by_pid(pid
) : current
;
3649 /* Actually do priority change: must hold rq lock. */
3650 static void __setscheduler(struct task_struct
*p
, int policy
, int prio
)
3654 p
->rt_priority
= prio
;
3655 if (policy
!= SCHED_NORMAL
)
3656 p
->prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
3658 p
->prio
= p
->static_prio
;
3662 * sched_setscheduler - change the scheduling policy and/or RT priority of
3664 * @p: the task in question.
3665 * @policy: new policy.
3666 * @param: structure containing the new RT priority.
3668 int sched_setscheduler(struct task_struct
*p
, int policy
,
3669 struct sched_param
*param
)
3672 int oldprio
, oldpolicy
= -1;
3673 prio_array_t
*array
;
3674 unsigned long flags
;
3678 /* double check policy once rq lock held */
3680 policy
= oldpolicy
= p
->policy
;
3681 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3682 policy
!= SCHED_NORMAL
)
3685 * Valid priorities for SCHED_FIFO and SCHED_RR are
3686 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL is 0.
3688 if (param
->sched_priority
< 0 ||
3689 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3690 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
3692 if ((policy
== SCHED_NORMAL
) != (param
->sched_priority
== 0))
3696 * Allow unprivileged RT tasks to decrease priority:
3698 if (!capable(CAP_SYS_NICE
)) {
3699 /* can't change policy */
3700 if (policy
!= p
->policy
&&
3701 !p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
)
3703 /* can't increase priority */
3704 if (policy
!= SCHED_NORMAL
&&
3705 param
->sched_priority
> p
->rt_priority
&&
3706 param
->sched_priority
>
3707 p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
)
3709 /* can't change other user's priorities */
3710 if ((current
->euid
!= p
->euid
) &&
3711 (current
->euid
!= p
->uid
))
3715 retval
= security_task_setscheduler(p
, policy
, param
);
3719 * To be able to change p->policy safely, the apropriate
3720 * runqueue lock must be held.
3722 rq
= task_rq_lock(p
, &flags
);
3723 /* recheck policy now with rq lock held */
3724 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3725 policy
= oldpolicy
= -1;
3726 task_rq_unlock(rq
, &flags
);
3731 deactivate_task(p
, rq
);
3733 __setscheduler(p
, policy
, param
->sched_priority
);
3735 __activate_task(p
, rq
);
3737 * Reschedule if we are currently running on this runqueue and
3738 * our priority decreased, or if we are not currently running on
3739 * this runqueue and our priority is higher than the current's
3741 if (task_running(rq
, p
)) {
3742 if (p
->prio
> oldprio
)
3743 resched_task(rq
->curr
);
3744 } else if (TASK_PREEMPTS_CURR(p
, rq
))
3745 resched_task(rq
->curr
);
3747 task_rq_unlock(rq
, &flags
);
3750 EXPORT_SYMBOL_GPL(sched_setscheduler
);
3753 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
3756 struct sched_param lparam
;
3757 struct task_struct
*p
;
3759 if (!param
|| pid
< 0)
3761 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
3763 read_lock_irq(&tasklist_lock
);
3764 p
= find_process_by_pid(pid
);
3766 read_unlock_irq(&tasklist_lock
);
3769 retval
= sched_setscheduler(p
, policy
, &lparam
);
3770 read_unlock_irq(&tasklist_lock
);
3775 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3776 * @pid: the pid in question.
3777 * @policy: new policy.
3778 * @param: structure containing the new RT priority.
3780 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
3781 struct sched_param __user
*param
)
3783 return do_sched_setscheduler(pid
, policy
, param
);
3787 * sys_sched_setparam - set/change the RT priority of a thread
3788 * @pid: the pid in question.
3789 * @param: structure containing the new RT priority.
3791 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
3793 return do_sched_setscheduler(pid
, -1, param
);
3797 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3798 * @pid: the pid in question.
3800 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
3802 int retval
= -EINVAL
;
3809 read_lock(&tasklist_lock
);
3810 p
= find_process_by_pid(pid
);
3812 retval
= security_task_getscheduler(p
);
3816 read_unlock(&tasklist_lock
);
3823 * sys_sched_getscheduler - get the RT priority of a thread
3824 * @pid: the pid in question.
3825 * @param: structure containing the RT priority.
3827 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
3829 struct sched_param lp
;
3830 int retval
= -EINVAL
;
3833 if (!param
|| pid
< 0)
3836 read_lock(&tasklist_lock
);
3837 p
= find_process_by_pid(pid
);
3842 retval
= security_task_getscheduler(p
);
3846 lp
.sched_priority
= p
->rt_priority
;
3847 read_unlock(&tasklist_lock
);
3850 * This one might sleep, we cannot do it with a spinlock held ...
3852 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
3858 read_unlock(&tasklist_lock
);
3862 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
3866 cpumask_t cpus_allowed
;
3869 read_lock(&tasklist_lock
);
3871 p
= find_process_by_pid(pid
);
3873 read_unlock(&tasklist_lock
);
3874 unlock_cpu_hotplug();
3879 * It is not safe to call set_cpus_allowed with the
3880 * tasklist_lock held. We will bump the task_struct's
3881 * usage count and then drop tasklist_lock.
3884 read_unlock(&tasklist_lock
);
3887 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
3888 !capable(CAP_SYS_NICE
))
3891 cpus_allowed
= cpuset_cpus_allowed(p
);
3892 cpus_and(new_mask
, new_mask
, cpus_allowed
);
3893 retval
= set_cpus_allowed(p
, new_mask
);
3897 unlock_cpu_hotplug();
3901 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
3902 cpumask_t
*new_mask
)
3904 if (len
< sizeof(cpumask_t
)) {
3905 memset(new_mask
, 0, sizeof(cpumask_t
));
3906 } else if (len
> sizeof(cpumask_t
)) {
3907 len
= sizeof(cpumask_t
);
3909 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
3913 * sys_sched_setaffinity - set the cpu affinity of a process
3914 * @pid: pid of the process
3915 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3916 * @user_mask_ptr: user-space pointer to the new cpu mask
3918 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
3919 unsigned long __user
*user_mask_ptr
)
3924 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
3928 return sched_setaffinity(pid
, new_mask
);
3932 * Represents all cpu's present in the system
3933 * In systems capable of hotplug, this map could dynamically grow
3934 * as new cpu's are detected in the system via any platform specific
3935 * method, such as ACPI for e.g.
3938 cpumask_t cpu_present_map
;
3939 EXPORT_SYMBOL(cpu_present_map
);
3942 cpumask_t cpu_online_map
= CPU_MASK_ALL
;
3943 cpumask_t cpu_possible_map
= CPU_MASK_ALL
;
3946 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
3952 read_lock(&tasklist_lock
);
3955 p
= find_process_by_pid(pid
);
3960 cpus_and(*mask
, p
->cpus_allowed
, cpu_possible_map
);
3963 read_unlock(&tasklist_lock
);
3964 unlock_cpu_hotplug();
3972 * sys_sched_getaffinity - get the cpu affinity of a process
3973 * @pid: pid of the process
3974 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3975 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3977 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
3978 unsigned long __user
*user_mask_ptr
)
3983 if (len
< sizeof(cpumask_t
))
3986 ret
= sched_getaffinity(pid
, &mask
);
3990 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
3993 return sizeof(cpumask_t
);
3997 * sys_sched_yield - yield the current processor to other threads.
3999 * this function yields the current CPU by moving the calling thread
4000 * to the expired array. If there are no other threads running on this
4001 * CPU then this function will return.
4003 asmlinkage
long sys_sched_yield(void)
4005 runqueue_t
*rq
= this_rq_lock();
4006 prio_array_t
*array
= current
->array
;
4007 prio_array_t
*target
= rq
->expired
;
4009 schedstat_inc(rq
, yld_cnt
);
4011 * We implement yielding by moving the task into the expired
4014 * (special rule: RT tasks will just roundrobin in the active
4017 if (rt_task(current
))
4018 target
= rq
->active
;
4020 if (array
->nr_active
== 1) {
4021 schedstat_inc(rq
, yld_act_empty
);
4022 if (!rq
->expired
->nr_active
)
4023 schedstat_inc(rq
, yld_both_empty
);
4024 } else if (!rq
->expired
->nr_active
)
4025 schedstat_inc(rq
, yld_exp_empty
);
4027 if (array
!= target
) {
4028 dequeue_task(current
, array
);
4029 enqueue_task(current
, target
);
4032 * requeue_task is cheaper so perform that if possible.
4034 requeue_task(current
, array
);
4037 * Since we are going to call schedule() anyway, there's
4038 * no need to preempt or enable interrupts:
4040 __release(rq
->lock
);
4041 _raw_spin_unlock(&rq
->lock
);
4042 preempt_enable_no_resched();
4049 static inline void __cond_resched(void)
4052 * The BKS might be reacquired before we have dropped
4053 * PREEMPT_ACTIVE, which could trigger a second
4054 * cond_resched() call.
4056 if (unlikely(preempt_count()))
4059 add_preempt_count(PREEMPT_ACTIVE
);
4061 sub_preempt_count(PREEMPT_ACTIVE
);
4062 } while (need_resched());
4065 int __sched
cond_resched(void)
4067 if (need_resched()) {
4074 EXPORT_SYMBOL(cond_resched
);
4077 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4078 * call schedule, and on return reacquire the lock.
4080 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4081 * operations here to prevent schedule() from being called twice (once via
4082 * spin_unlock(), once by hand).
4084 int cond_resched_lock(spinlock_t
*lock
)
4088 if (need_lockbreak(lock
)) {
4094 if (need_resched()) {
4095 _raw_spin_unlock(lock
);
4096 preempt_enable_no_resched();
4104 EXPORT_SYMBOL(cond_resched_lock
);
4106 int __sched
cond_resched_softirq(void)
4108 BUG_ON(!in_softirq());
4110 if (need_resched()) {
4111 __local_bh_enable();
4119 EXPORT_SYMBOL(cond_resched_softirq
);
4123 * yield - yield the current processor to other threads.
4125 * this is a shortcut for kernel-space yielding - it marks the
4126 * thread runnable and calls sys_sched_yield().
4128 void __sched
yield(void)
4130 set_current_state(TASK_RUNNING
);
4134 EXPORT_SYMBOL(yield
);
4137 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4138 * that process accounting knows that this is a task in IO wait state.
4140 * But don't do that if it is a deliberate, throttling IO wait (this task
4141 * has set its backing_dev_info: the queue against which it should throttle)
4143 void __sched
io_schedule(void)
4145 struct runqueue
*rq
= &per_cpu(runqueues
, raw_smp_processor_id());
4147 atomic_inc(&rq
->nr_iowait
);
4149 atomic_dec(&rq
->nr_iowait
);
4152 EXPORT_SYMBOL(io_schedule
);
4154 long __sched
io_schedule_timeout(long timeout
)
4156 struct runqueue
*rq
= &per_cpu(runqueues
, raw_smp_processor_id());
4159 atomic_inc(&rq
->nr_iowait
);
4160 ret
= schedule_timeout(timeout
);
4161 atomic_dec(&rq
->nr_iowait
);
4166 * sys_sched_get_priority_max - return maximum RT priority.
4167 * @policy: scheduling class.
4169 * this syscall returns the maximum rt_priority that can be used
4170 * by a given scheduling class.
4172 asmlinkage
long sys_sched_get_priority_max(int policy
)
4179 ret
= MAX_USER_RT_PRIO
-1;
4189 * sys_sched_get_priority_min - return minimum RT priority.
4190 * @policy: scheduling class.
4192 * this syscall returns the minimum rt_priority that can be used
4193 * by a given scheduling class.
4195 asmlinkage
long sys_sched_get_priority_min(int policy
)
4211 * sys_sched_rr_get_interval - return the default timeslice of a process.
4212 * @pid: pid of the process.
4213 * @interval: userspace pointer to the timeslice value.
4215 * this syscall writes the default timeslice value of a given process
4216 * into the user-space timespec buffer. A value of '0' means infinity.
4219 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4221 int retval
= -EINVAL
;
4229 read_lock(&tasklist_lock
);
4230 p
= find_process_by_pid(pid
);
4234 retval
= security_task_getscheduler(p
);
4238 jiffies_to_timespec(p
->policy
& SCHED_FIFO
?
4239 0 : task_timeslice(p
), &t
);
4240 read_unlock(&tasklist_lock
);
4241 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4245 read_unlock(&tasklist_lock
);
4249 static inline struct task_struct
*eldest_child(struct task_struct
*p
)
4251 if (list_empty(&p
->children
)) return NULL
;
4252 return list_entry(p
->children
.next
,struct task_struct
,sibling
);
4255 static inline struct task_struct
*older_sibling(struct task_struct
*p
)
4257 if (p
->sibling
.prev
==&p
->parent
->children
) return NULL
;
4258 return list_entry(p
->sibling
.prev
,struct task_struct
,sibling
);
4261 static inline struct task_struct
*younger_sibling(struct task_struct
*p
)
4263 if (p
->sibling
.next
==&p
->parent
->children
) return NULL
;
4264 return list_entry(p
->sibling
.next
,struct task_struct
,sibling
);
4267 static void show_task(task_t
*p
)
4271 unsigned long free
= 0;
4272 static const char *stat_nam
[] = { "R", "S", "D", "T", "t", "Z", "X" };
4274 printk("%-13.13s ", p
->comm
);
4275 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4276 if (state
< ARRAY_SIZE(stat_nam
))
4277 printk(stat_nam
[state
]);
4280 #if (BITS_PER_LONG == 32)
4281 if (state
== TASK_RUNNING
)
4282 printk(" running ");
4284 printk(" %08lX ", thread_saved_pc(p
));
4286 if (state
== TASK_RUNNING
)
4287 printk(" running task ");
4289 printk(" %016lx ", thread_saved_pc(p
));
4291 #ifdef CONFIG_DEBUG_STACK_USAGE
4293 unsigned long *n
= (unsigned long *) (p
->thread_info
+1);
4296 free
= (unsigned long) n
- (unsigned long)(p
->thread_info
+1);
4299 printk("%5lu %5d %6d ", free
, p
->pid
, p
->parent
->pid
);
4300 if ((relative
= eldest_child(p
)))
4301 printk("%5d ", relative
->pid
);
4304 if ((relative
= younger_sibling(p
)))
4305 printk("%7d", relative
->pid
);
4308 if ((relative
= older_sibling(p
)))
4309 printk(" %5d", relative
->pid
);
4313 printk(" (L-TLB)\n");
4315 printk(" (NOTLB)\n");
4317 if (state
!= TASK_RUNNING
)
4318 show_stack(p
, NULL
);
4321 void show_state(void)
4325 #if (BITS_PER_LONG == 32)
4328 printk(" task PC pid father child younger older\n");
4332 printk(" task PC pid father child younger older\n");
4334 read_lock(&tasklist_lock
);
4335 do_each_thread(g
, p
) {
4337 * reset the NMI-timeout, listing all files on a slow
4338 * console might take alot of time:
4340 touch_nmi_watchdog();
4342 } while_each_thread(g
, p
);
4344 read_unlock(&tasklist_lock
);
4348 * init_idle - set up an idle thread for a given CPU
4349 * @idle: task in question
4350 * @cpu: cpu the idle task belongs to
4352 * NOTE: this function does not set the idle thread's NEED_RESCHED
4353 * flag, to make booting more robust.
4355 void __devinit
init_idle(task_t
*idle
, int cpu
)
4357 runqueue_t
*rq
= cpu_rq(cpu
);
4358 unsigned long flags
;
4360 idle
->sleep_avg
= 0;
4362 idle
->prio
= MAX_PRIO
;
4363 idle
->state
= TASK_RUNNING
;
4364 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4365 set_task_cpu(idle
, cpu
);
4367 spin_lock_irqsave(&rq
->lock
, flags
);
4368 rq
->curr
= rq
->idle
= idle
;
4369 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4372 spin_unlock_irqrestore(&rq
->lock
, flags
);
4374 /* Set the preempt count _outside_ the spinlocks! */
4375 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4376 idle
->thread_info
->preempt_count
= (idle
->lock_depth
>= 0);
4378 idle
->thread_info
->preempt_count
= 0;
4383 * In a system that switches off the HZ timer nohz_cpu_mask
4384 * indicates which cpus entered this state. This is used
4385 * in the rcu update to wait only for active cpus. For system
4386 * which do not switch off the HZ timer nohz_cpu_mask should
4387 * always be CPU_MASK_NONE.
4389 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4393 * This is how migration works:
4395 * 1) we queue a migration_req_t structure in the source CPU's
4396 * runqueue and wake up that CPU's migration thread.
4397 * 2) we down() the locked semaphore => thread blocks.
4398 * 3) migration thread wakes up (implicitly it forces the migrated
4399 * thread off the CPU)
4400 * 4) it gets the migration request and checks whether the migrated
4401 * task is still in the wrong runqueue.
4402 * 5) if it's in the wrong runqueue then the migration thread removes
4403 * it and puts it into the right queue.
4404 * 6) migration thread up()s the semaphore.
4405 * 7) we wake up and the migration is done.
4409 * Change a given task's CPU affinity. Migrate the thread to a
4410 * proper CPU and schedule it away if the CPU it's executing on
4411 * is removed from the allowed bitmask.
4413 * NOTE: the caller must have a valid reference to the task, the
4414 * task must not exit() & deallocate itself prematurely. The
4415 * call is not atomic; no spinlocks may be held.
4417 int set_cpus_allowed(task_t
*p
, cpumask_t new_mask
)
4419 unsigned long flags
;
4421 migration_req_t req
;
4424 rq
= task_rq_lock(p
, &flags
);
4425 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4430 p
->cpus_allowed
= new_mask
;
4431 /* Can the task run on the task's current CPU? If so, we're done */
4432 if (cpu_isset(task_cpu(p
), new_mask
))
4435 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4436 /* Need help from migration thread: drop lock and wait. */
4437 task_rq_unlock(rq
, &flags
);
4438 wake_up_process(rq
->migration_thread
);
4439 wait_for_completion(&req
.done
);
4440 tlb_migrate_finish(p
->mm
);
4444 task_rq_unlock(rq
, &flags
);
4448 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4451 * Move (not current) task off this cpu, onto dest cpu. We're doing
4452 * this because either it can't run here any more (set_cpus_allowed()
4453 * away from this CPU, or CPU going down), or because we're
4454 * attempting to rebalance this task on exec (sched_exec).
4456 * So we race with normal scheduler movements, but that's OK, as long
4457 * as the task is no longer on this CPU.
4459 static void __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4461 runqueue_t
*rq_dest
, *rq_src
;
4463 if (unlikely(cpu_is_offline(dest_cpu
)))
4466 rq_src
= cpu_rq(src_cpu
);
4467 rq_dest
= cpu_rq(dest_cpu
);
4469 double_rq_lock(rq_src
, rq_dest
);
4470 /* Already moved. */
4471 if (task_cpu(p
) != src_cpu
)
4473 /* Affinity changed (again). */
4474 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
4477 set_task_cpu(p
, dest_cpu
);
4480 * Sync timestamp with rq_dest's before activating.
4481 * The same thing could be achieved by doing this step
4482 * afterwards, and pretending it was a local activate.
4483 * This way is cleaner and logically correct.
4485 p
->timestamp
= p
->timestamp
- rq_src
->timestamp_last_tick
4486 + rq_dest
->timestamp_last_tick
;
4487 deactivate_task(p
, rq_src
);
4488 activate_task(p
, rq_dest
, 0);
4489 if (TASK_PREEMPTS_CURR(p
, rq_dest
))
4490 resched_task(rq_dest
->curr
);
4494 double_rq_unlock(rq_src
, rq_dest
);
4498 * migration_thread - this is a highprio system thread that performs
4499 * thread migration by bumping thread off CPU then 'pushing' onto
4502 static int migration_thread(void *data
)
4505 int cpu
= (long)data
;
4508 BUG_ON(rq
->migration_thread
!= current
);
4510 set_current_state(TASK_INTERRUPTIBLE
);
4511 while (!kthread_should_stop()) {
4512 struct list_head
*head
;
4513 migration_req_t
*req
;
4517 spin_lock_irq(&rq
->lock
);
4519 if (cpu_is_offline(cpu
)) {
4520 spin_unlock_irq(&rq
->lock
);
4524 if (rq
->active_balance
) {
4525 active_load_balance(rq
, cpu
);
4526 rq
->active_balance
= 0;
4529 head
= &rq
->migration_queue
;
4531 if (list_empty(head
)) {
4532 spin_unlock_irq(&rq
->lock
);
4534 set_current_state(TASK_INTERRUPTIBLE
);
4537 req
= list_entry(head
->next
, migration_req_t
, list
);
4538 list_del_init(head
->next
);
4540 spin_unlock(&rq
->lock
);
4541 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
4544 complete(&req
->done
);
4546 __set_current_state(TASK_RUNNING
);
4550 /* Wait for kthread_stop */
4551 set_current_state(TASK_INTERRUPTIBLE
);
4552 while (!kthread_should_stop()) {
4554 set_current_state(TASK_INTERRUPTIBLE
);
4556 __set_current_state(TASK_RUNNING
);
4560 #ifdef CONFIG_HOTPLUG_CPU
4561 /* Figure out where task on dead CPU should go, use force if neccessary. */
4562 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*tsk
)
4568 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
4569 cpus_and(mask
, mask
, tsk
->cpus_allowed
);
4570 dest_cpu
= any_online_cpu(mask
);
4572 /* On any allowed CPU? */
4573 if (dest_cpu
== NR_CPUS
)
4574 dest_cpu
= any_online_cpu(tsk
->cpus_allowed
);
4576 /* No more Mr. Nice Guy. */
4577 if (dest_cpu
== NR_CPUS
) {
4578 cpus_setall(tsk
->cpus_allowed
);
4579 dest_cpu
= any_online_cpu(tsk
->cpus_allowed
);
4582 * Don't tell them about moving exiting tasks or
4583 * kernel threads (both mm NULL), since they never
4586 if (tsk
->mm
&& printk_ratelimit())
4587 printk(KERN_INFO
"process %d (%s) no "
4588 "longer affine to cpu%d\n",
4589 tsk
->pid
, tsk
->comm
, dead_cpu
);
4591 __migrate_task(tsk
, dead_cpu
, dest_cpu
);
4595 * While a dead CPU has no uninterruptible tasks queued at this point,
4596 * it might still have a nonzero ->nr_uninterruptible counter, because
4597 * for performance reasons the counter is not stricly tracking tasks to
4598 * their home CPUs. So we just add the counter to another CPU's counter,
4599 * to keep the global sum constant after CPU-down:
4601 static void migrate_nr_uninterruptible(runqueue_t
*rq_src
)
4603 runqueue_t
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
4604 unsigned long flags
;
4606 local_irq_save(flags
);
4607 double_rq_lock(rq_src
, rq_dest
);
4608 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
4609 rq_src
->nr_uninterruptible
= 0;
4610 double_rq_unlock(rq_src
, rq_dest
);
4611 local_irq_restore(flags
);
4614 /* Run through task list and migrate tasks from the dead cpu. */
4615 static void migrate_live_tasks(int src_cpu
)
4617 struct task_struct
*tsk
, *t
;
4619 write_lock_irq(&tasklist_lock
);
4621 do_each_thread(t
, tsk
) {
4625 if (task_cpu(tsk
) == src_cpu
)
4626 move_task_off_dead_cpu(src_cpu
, tsk
);
4627 } while_each_thread(t
, tsk
);
4629 write_unlock_irq(&tasklist_lock
);
4632 /* Schedules idle task to be the next runnable task on current CPU.
4633 * It does so by boosting its priority to highest possible and adding it to
4634 * the _front_ of runqueue. Used by CPU offline code.
4636 void sched_idle_next(void)
4638 int cpu
= smp_processor_id();
4639 runqueue_t
*rq
= this_rq();
4640 struct task_struct
*p
= rq
->idle
;
4641 unsigned long flags
;
4643 /* cpu has to be offline */
4644 BUG_ON(cpu_online(cpu
));
4646 /* Strictly not necessary since rest of the CPUs are stopped by now
4647 * and interrupts disabled on current cpu.
4649 spin_lock_irqsave(&rq
->lock
, flags
);
4651 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
4652 /* Add idle task to _front_ of it's priority queue */
4653 __activate_idle_task(p
, rq
);
4655 spin_unlock_irqrestore(&rq
->lock
, flags
);
4658 /* Ensures that the idle task is using init_mm right before its cpu goes
4661 void idle_task_exit(void)
4663 struct mm_struct
*mm
= current
->active_mm
;
4665 BUG_ON(cpu_online(smp_processor_id()));
4668 switch_mm(mm
, &init_mm
, current
);
4672 static void migrate_dead(unsigned int dead_cpu
, task_t
*tsk
)
4674 struct runqueue
*rq
= cpu_rq(dead_cpu
);
4676 /* Must be exiting, otherwise would be on tasklist. */
4677 BUG_ON(tsk
->exit_state
!= EXIT_ZOMBIE
&& tsk
->exit_state
!= EXIT_DEAD
);
4679 /* Cannot have done final schedule yet: would have vanished. */
4680 BUG_ON(tsk
->flags
& PF_DEAD
);
4682 get_task_struct(tsk
);
4685 * Drop lock around migration; if someone else moves it,
4686 * that's OK. No task can be added to this CPU, so iteration is
4689 spin_unlock_irq(&rq
->lock
);
4690 move_task_off_dead_cpu(dead_cpu
, tsk
);
4691 spin_lock_irq(&rq
->lock
);
4693 put_task_struct(tsk
);
4696 /* release_task() removes task from tasklist, so we won't find dead tasks. */
4697 static void migrate_dead_tasks(unsigned int dead_cpu
)
4700 struct runqueue
*rq
= cpu_rq(dead_cpu
);
4702 for (arr
= 0; arr
< 2; arr
++) {
4703 for (i
= 0; i
< MAX_PRIO
; i
++) {
4704 struct list_head
*list
= &rq
->arrays
[arr
].queue
[i
];
4705 while (!list_empty(list
))
4706 migrate_dead(dead_cpu
,
4707 list_entry(list
->next
, task_t
,
4712 #endif /* CONFIG_HOTPLUG_CPU */
4715 * migration_call - callback that gets triggered when a CPU is added.
4716 * Here we can start up the necessary migration thread for the new CPU.
4718 static int migration_call(struct notifier_block
*nfb
, unsigned long action
,
4721 int cpu
= (long)hcpu
;
4722 struct task_struct
*p
;
4723 struct runqueue
*rq
;
4724 unsigned long flags
;
4727 case CPU_UP_PREPARE
:
4728 p
= kthread_create(migration_thread
, hcpu
, "migration/%d",cpu
);
4731 p
->flags
|= PF_NOFREEZE
;
4732 kthread_bind(p
, cpu
);
4733 /* Must be high prio: stop_machine expects to yield to it. */
4734 rq
= task_rq_lock(p
, &flags
);
4735 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
4736 task_rq_unlock(rq
, &flags
);
4737 cpu_rq(cpu
)->migration_thread
= p
;
4740 /* Strictly unneccessary, as first user will wake it. */
4741 wake_up_process(cpu_rq(cpu
)->migration_thread
);
4743 #ifdef CONFIG_HOTPLUG_CPU
4744 case CPU_UP_CANCELED
:
4745 /* Unbind it from offline cpu so it can run. Fall thru. */
4746 kthread_bind(cpu_rq(cpu
)->migration_thread
,
4747 any_online_cpu(cpu_online_map
));
4748 kthread_stop(cpu_rq(cpu
)->migration_thread
);
4749 cpu_rq(cpu
)->migration_thread
= NULL
;
4752 migrate_live_tasks(cpu
);
4754 kthread_stop(rq
->migration_thread
);
4755 rq
->migration_thread
= NULL
;
4756 /* Idle task back to normal (off runqueue, low prio) */
4757 rq
= task_rq_lock(rq
->idle
, &flags
);
4758 deactivate_task(rq
->idle
, rq
);
4759 rq
->idle
->static_prio
= MAX_PRIO
;
4760 __setscheduler(rq
->idle
, SCHED_NORMAL
, 0);
4761 migrate_dead_tasks(cpu
);
4762 task_rq_unlock(rq
, &flags
);
4763 migrate_nr_uninterruptible(rq
);
4764 BUG_ON(rq
->nr_running
!= 0);
4766 /* No need to migrate the tasks: it was best-effort if
4767 * they didn't do lock_cpu_hotplug(). Just wake up
4768 * the requestors. */
4769 spin_lock_irq(&rq
->lock
);
4770 while (!list_empty(&rq
->migration_queue
)) {
4771 migration_req_t
*req
;
4772 req
= list_entry(rq
->migration_queue
.next
,
4773 migration_req_t
, list
);
4774 list_del_init(&req
->list
);
4775 complete(&req
->done
);
4777 spin_unlock_irq(&rq
->lock
);
4784 /* Register at highest priority so that task migration (migrate_all_tasks)
4785 * happens before everything else.
4787 static struct notifier_block __devinitdata migration_notifier
= {
4788 .notifier_call
= migration_call
,
4792 int __init
migration_init(void)
4794 void *cpu
= (void *)(long)smp_processor_id();
4795 /* Start one for boot CPU. */
4796 migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
4797 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
4798 register_cpu_notifier(&migration_notifier
);
4804 #undef SCHED_DOMAIN_DEBUG
4805 #ifdef SCHED_DOMAIN_DEBUG
4806 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
4811 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
4815 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
4820 struct sched_group
*group
= sd
->groups
;
4821 cpumask_t groupmask
;
4823 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
4824 cpus_clear(groupmask
);
4827 for (i
= 0; i
< level
+ 1; i
++)
4829 printk("domain %d: ", level
);
4831 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
4832 printk("does not load-balance\n");
4834 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain has parent");
4838 printk("span %s\n", str
);
4840 if (!cpu_isset(cpu
, sd
->span
))
4841 printk(KERN_ERR
"ERROR: domain->span does not contain CPU%d\n", cpu
);
4842 if (!cpu_isset(cpu
, group
->cpumask
))
4843 printk(KERN_ERR
"ERROR: domain->groups does not contain CPU%d\n", cpu
);
4846 for (i
= 0; i
< level
+ 2; i
++)
4852 printk(KERN_ERR
"ERROR: group is NULL\n");
4856 if (!group
->cpu_power
) {
4858 printk(KERN_ERR
"ERROR: domain->cpu_power not set\n");
4861 if (!cpus_weight(group
->cpumask
)) {
4863 printk(KERN_ERR
"ERROR: empty group\n");
4866 if (cpus_intersects(groupmask
, group
->cpumask
)) {
4868 printk(KERN_ERR
"ERROR: repeated CPUs\n");
4871 cpus_or(groupmask
, groupmask
, group
->cpumask
);
4873 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
4876 group
= group
->next
;
4877 } while (group
!= sd
->groups
);
4880 if (!cpus_equal(sd
->span
, groupmask
))
4881 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
4887 if (!cpus_subset(groupmask
, sd
->span
))
4888 printk(KERN_ERR
"ERROR: parent span is not a superset of domain->span\n");
4894 #define sched_domain_debug(sd, cpu) {}
4897 static int sd_degenerate(struct sched_domain
*sd
)
4899 if (cpus_weight(sd
->span
) == 1)
4902 /* Following flags need at least 2 groups */
4903 if (sd
->flags
& (SD_LOAD_BALANCE
|
4904 SD_BALANCE_NEWIDLE
|
4907 if (sd
->groups
!= sd
->groups
->next
)
4911 /* Following flags don't use groups */
4912 if (sd
->flags
& (SD_WAKE_IDLE
|
4920 static int sd_parent_degenerate(struct sched_domain
*sd
,
4921 struct sched_domain
*parent
)
4923 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
4925 if (sd_degenerate(parent
))
4928 if (!cpus_equal(sd
->span
, parent
->span
))
4931 /* Does parent contain flags not in child? */
4932 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
4933 if (cflags
& SD_WAKE_AFFINE
)
4934 pflags
&= ~SD_WAKE_BALANCE
;
4935 /* Flags needing groups don't count if only 1 group in parent */
4936 if (parent
->groups
== parent
->groups
->next
) {
4937 pflags
&= ~(SD_LOAD_BALANCE
|
4938 SD_BALANCE_NEWIDLE
|
4942 if (~cflags
& pflags
)
4949 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
4950 * hold the hotplug lock.
4952 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
4954 runqueue_t
*rq
= cpu_rq(cpu
);
4955 struct sched_domain
*tmp
;
4957 /* Remove the sched domains which do not contribute to scheduling. */
4958 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
4959 struct sched_domain
*parent
= tmp
->parent
;
4962 if (sd_parent_degenerate(tmp
, parent
))
4963 tmp
->parent
= parent
->parent
;
4966 if (sd
&& sd_degenerate(sd
))
4969 sched_domain_debug(sd
, cpu
);
4971 rcu_assign_pointer(rq
->sd
, sd
);
4974 /* cpus with isolated domains */
4975 static cpumask_t __devinitdata cpu_isolated_map
= CPU_MASK_NONE
;
4977 /* Setup the mask of cpus configured for isolated domains */
4978 static int __init
isolated_cpu_setup(char *str
)
4980 int ints
[NR_CPUS
], i
;
4982 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
4983 cpus_clear(cpu_isolated_map
);
4984 for (i
= 1; i
<= ints
[0]; i
++)
4985 if (ints
[i
] < NR_CPUS
)
4986 cpu_set(ints
[i
], cpu_isolated_map
);
4990 __setup ("isolcpus=", isolated_cpu_setup
);
4993 * init_sched_build_groups takes an array of groups, the cpumask we wish
4994 * to span, and a pointer to a function which identifies what group a CPU
4995 * belongs to. The return value of group_fn must be a valid index into the
4996 * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
4997 * keep track of groups covered with a cpumask_t).
4999 * init_sched_build_groups will build a circular linked list of the groups
5000 * covered by the given span, and will set each group's ->cpumask correctly,
5001 * and ->cpu_power to 0.
5003 static void init_sched_build_groups(struct sched_group groups
[], cpumask_t span
,
5004 int (*group_fn
)(int cpu
))
5006 struct sched_group
*first
= NULL
, *last
= NULL
;
5007 cpumask_t covered
= CPU_MASK_NONE
;
5010 for_each_cpu_mask(i
, span
) {
5011 int group
= group_fn(i
);
5012 struct sched_group
*sg
= &groups
[group
];
5015 if (cpu_isset(i
, covered
))
5018 sg
->cpumask
= CPU_MASK_NONE
;
5021 for_each_cpu_mask(j
, span
) {
5022 if (group_fn(j
) != group
)
5025 cpu_set(j
, covered
);
5026 cpu_set(j
, sg
->cpumask
);
5037 #define SD_NODES_PER_DOMAIN 16
5041 * find_next_best_node - find the next node to include in a sched_domain
5042 * @node: node whose sched_domain we're building
5043 * @used_nodes: nodes already in the sched_domain
5045 * Find the next node to include in a given scheduling domain. Simply
5046 * finds the closest node not already in the @used_nodes map.
5048 * Should use nodemask_t.
5050 static int find_next_best_node(int node
, unsigned long *used_nodes
)
5052 int i
, n
, val
, min_val
, best_node
= 0;
5056 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5057 /* Start at @node */
5058 n
= (node
+ i
) % MAX_NUMNODES
;
5060 if (!nr_cpus_node(n
))
5063 /* Skip already used nodes */
5064 if (test_bit(n
, used_nodes
))
5067 /* Simple min distance search */
5068 val
= node_distance(node
, n
);
5070 if (val
< min_val
) {
5076 set_bit(best_node
, used_nodes
);
5081 * sched_domain_node_span - get a cpumask for a node's sched_domain
5082 * @node: node whose cpumask we're constructing
5083 * @size: number of nodes to include in this span
5085 * Given a node, construct a good cpumask for its sched_domain to span. It
5086 * should be one that prevents unnecessary balancing, but also spreads tasks
5089 static cpumask_t
sched_domain_node_span(int node
)
5092 cpumask_t span
, nodemask
;
5093 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
5096 bitmap_zero(used_nodes
, MAX_NUMNODES
);
5098 nodemask
= node_to_cpumask(node
);
5099 cpus_or(span
, span
, nodemask
);
5100 set_bit(node
, used_nodes
);
5102 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5103 int next_node
= find_next_best_node(node
, used_nodes
);
5104 nodemask
= node_to_cpumask(next_node
);
5105 cpus_or(span
, span
, nodemask
);
5113 * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we
5114 * can switch it on easily if needed.
5116 #ifdef CONFIG_SCHED_SMT
5117 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
5118 static struct sched_group sched_group_cpus
[NR_CPUS
];
5119 static int cpu_to_cpu_group(int cpu
)
5125 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
5126 static struct sched_group sched_group_phys
[NR_CPUS
];
5127 static int cpu_to_phys_group(int cpu
)
5129 #ifdef CONFIG_SCHED_SMT
5130 return first_cpu(cpu_sibling_map
[cpu
]);
5138 * The init_sched_build_groups can't handle what we want to do with node
5139 * groups, so roll our own. Now each node has its own list of groups which
5140 * gets dynamically allocated.
5142 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
5143 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
5145 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
5146 static struct sched_group
*sched_group_allnodes_bycpu
[NR_CPUS
];
5148 static int cpu_to_allnodes_group(int cpu
)
5150 return cpu_to_node(cpu
);
5155 * Build sched domains for a given set of cpus and attach the sched domains
5156 * to the individual cpus
5158 void build_sched_domains(const cpumask_t
*cpu_map
)
5162 struct sched_group
**sched_group_nodes
= NULL
;
5163 struct sched_group
*sched_group_allnodes
= NULL
;
5166 * Allocate the per-node list of sched groups
5168 sched_group_nodes
= kmalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
5170 if (!sched_group_nodes
) {
5171 printk(KERN_WARNING
"Can not alloc sched group node list\n");
5174 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
5178 * Set up domains for cpus specified by the cpu_map.
5180 for_each_cpu_mask(i
, *cpu_map
) {
5182 struct sched_domain
*sd
= NULL
, *p
;
5183 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
5185 cpus_and(nodemask
, nodemask
, *cpu_map
);
5188 if (cpus_weight(*cpu_map
)
5189 > SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
5190 if (!sched_group_allnodes
) {
5191 sched_group_allnodes
5192 = kmalloc(sizeof(struct sched_group
)
5195 if (!sched_group_allnodes
) {
5197 "Can not alloc allnodes sched group\n");
5200 sched_group_allnodes_bycpu
[i
]
5201 = sched_group_allnodes
;
5203 sd
= &per_cpu(allnodes_domains
, i
);
5204 *sd
= SD_ALLNODES_INIT
;
5205 sd
->span
= *cpu_map
;
5206 group
= cpu_to_allnodes_group(i
);
5207 sd
->groups
= &sched_group_allnodes
[group
];
5212 sd
= &per_cpu(node_domains
, i
);
5214 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
5216 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
5220 sd
= &per_cpu(phys_domains
, i
);
5221 group
= cpu_to_phys_group(i
);
5223 sd
->span
= nodemask
;
5225 sd
->groups
= &sched_group_phys
[group
];
5227 #ifdef CONFIG_SCHED_SMT
5229 sd
= &per_cpu(cpu_domains
, i
);
5230 group
= cpu_to_cpu_group(i
);
5231 *sd
= SD_SIBLING_INIT
;
5232 sd
->span
= cpu_sibling_map
[i
];
5233 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
5235 sd
->groups
= &sched_group_cpus
[group
];
5239 #ifdef CONFIG_SCHED_SMT
5240 /* Set up CPU (sibling) groups */
5241 for_each_cpu_mask(i
, *cpu_map
) {
5242 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
5243 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
5244 if (i
!= first_cpu(this_sibling_map
))
5247 init_sched_build_groups(sched_group_cpus
, this_sibling_map
,
5252 /* Set up physical groups */
5253 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5254 cpumask_t nodemask
= node_to_cpumask(i
);
5256 cpus_and(nodemask
, nodemask
, *cpu_map
);
5257 if (cpus_empty(nodemask
))
5260 init_sched_build_groups(sched_group_phys
, nodemask
,
5261 &cpu_to_phys_group
);
5265 /* Set up node groups */
5266 if (sched_group_allnodes
)
5267 init_sched_build_groups(sched_group_allnodes
, *cpu_map
,
5268 &cpu_to_allnodes_group
);
5270 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5271 /* Set up node groups */
5272 struct sched_group
*sg
, *prev
;
5273 cpumask_t nodemask
= node_to_cpumask(i
);
5274 cpumask_t domainspan
;
5275 cpumask_t covered
= CPU_MASK_NONE
;
5278 cpus_and(nodemask
, nodemask
, *cpu_map
);
5279 if (cpus_empty(nodemask
)) {
5280 sched_group_nodes
[i
] = NULL
;
5284 domainspan
= sched_domain_node_span(i
);
5285 cpus_and(domainspan
, domainspan
, *cpu_map
);
5287 sg
= kmalloc(sizeof(struct sched_group
), GFP_KERNEL
);
5288 sched_group_nodes
[i
] = sg
;
5289 for_each_cpu_mask(j
, nodemask
) {
5290 struct sched_domain
*sd
;
5291 sd
= &per_cpu(node_domains
, j
);
5293 if (sd
->groups
== NULL
) {
5294 /* Turn off balancing if we have no groups */
5300 "Can not alloc domain group for node %d\n", i
);
5304 sg
->cpumask
= nodemask
;
5305 cpus_or(covered
, covered
, nodemask
);
5308 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
5309 cpumask_t tmp
, notcovered
;
5310 int n
= (i
+ j
) % MAX_NUMNODES
;
5312 cpus_complement(notcovered
, covered
);
5313 cpus_and(tmp
, notcovered
, *cpu_map
);
5314 cpus_and(tmp
, tmp
, domainspan
);
5315 if (cpus_empty(tmp
))
5318 nodemask
= node_to_cpumask(n
);
5319 cpus_and(tmp
, tmp
, nodemask
);
5320 if (cpus_empty(tmp
))
5323 sg
= kmalloc(sizeof(struct sched_group
), GFP_KERNEL
);
5326 "Can not alloc domain group for node %d\n", j
);
5331 cpus_or(covered
, covered
, tmp
);
5335 prev
->next
= sched_group_nodes
[i
];
5339 /* Calculate CPU power for physical packages and nodes */
5340 for_each_cpu_mask(i
, *cpu_map
) {
5342 struct sched_domain
*sd
;
5343 #ifdef CONFIG_SCHED_SMT
5344 sd
= &per_cpu(cpu_domains
, i
);
5345 power
= SCHED_LOAD_SCALE
;
5346 sd
->groups
->cpu_power
= power
;
5349 sd
= &per_cpu(phys_domains
, i
);
5350 power
= SCHED_LOAD_SCALE
+ SCHED_LOAD_SCALE
*
5351 (cpus_weight(sd
->groups
->cpumask
)-1) / 10;
5352 sd
->groups
->cpu_power
= power
;
5355 sd
= &per_cpu(allnodes_domains
, i
);
5357 power
= SCHED_LOAD_SCALE
+ SCHED_LOAD_SCALE
*
5358 (cpus_weight(sd
->groups
->cpumask
)-1) / 10;
5359 sd
->groups
->cpu_power
= power
;
5365 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5366 struct sched_group
*sg
= sched_group_nodes
[i
];
5372 for_each_cpu_mask(j
, sg
->cpumask
) {
5373 struct sched_domain
*sd
;
5376 sd
= &per_cpu(phys_domains
, j
);
5377 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
5379 * Only add "power" once for each
5384 power
= SCHED_LOAD_SCALE
+ SCHED_LOAD_SCALE
*
5385 (cpus_weight(sd
->groups
->cpumask
)-1) / 10;
5387 sg
->cpu_power
+= power
;
5390 if (sg
!= sched_group_nodes
[i
])
5395 /* Attach the domains */
5396 for_each_cpu_mask(i
, *cpu_map
) {
5397 struct sched_domain
*sd
;
5398 #ifdef CONFIG_SCHED_SMT
5399 sd
= &per_cpu(cpu_domains
, i
);
5401 sd
= &per_cpu(phys_domains
, i
);
5403 cpu_attach_domain(sd
, i
);
5407 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
5409 static void arch_init_sched_domains(const cpumask_t
*cpu_map
)
5411 cpumask_t cpu_default_map
;
5414 * Setup mask for cpus without special case scheduling requirements.
5415 * For now this just excludes isolated cpus, but could be used to
5416 * exclude other special cases in the future.
5418 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
5420 build_sched_domains(&cpu_default_map
);
5423 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
5429 for_each_cpu_mask(cpu
, *cpu_map
) {
5430 struct sched_group
*sched_group_allnodes
5431 = sched_group_allnodes_bycpu
[cpu
];
5432 struct sched_group
**sched_group_nodes
5433 = sched_group_nodes_bycpu
[cpu
];
5435 if (sched_group_allnodes
) {
5436 kfree(sched_group_allnodes
);
5437 sched_group_allnodes_bycpu
[cpu
] = NULL
;
5440 if (!sched_group_nodes
)
5443 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5444 cpumask_t nodemask
= node_to_cpumask(i
);
5445 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
5447 cpus_and(nodemask
, nodemask
, *cpu_map
);
5448 if (cpus_empty(nodemask
))
5458 if (oldsg
!= sched_group_nodes
[i
])
5461 kfree(sched_group_nodes
);
5462 sched_group_nodes_bycpu
[cpu
] = NULL
;
5468 * Detach sched domains from a group of cpus specified in cpu_map
5469 * These cpus will now be attached to the NULL domain
5471 static inline void detach_destroy_domains(const cpumask_t
*cpu_map
)
5475 for_each_cpu_mask(i
, *cpu_map
)
5476 cpu_attach_domain(NULL
, i
);
5477 synchronize_sched();
5478 arch_destroy_sched_domains(cpu_map
);
5482 * Partition sched domains as specified by the cpumasks below.
5483 * This attaches all cpus from the cpumasks to the NULL domain,
5484 * waits for a RCU quiescent period, recalculates sched
5485 * domain information and then attaches them back to the
5486 * correct sched domains
5487 * Call with hotplug lock held
5489 void partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
5491 cpumask_t change_map
;
5493 cpus_and(*partition1
, *partition1
, cpu_online_map
);
5494 cpus_and(*partition2
, *partition2
, cpu_online_map
);
5495 cpus_or(change_map
, *partition1
, *partition2
);
5497 /* Detach sched domains from all of the affected cpus */
5498 detach_destroy_domains(&change_map
);
5499 if (!cpus_empty(*partition1
))
5500 build_sched_domains(partition1
);
5501 if (!cpus_empty(*partition2
))
5502 build_sched_domains(partition2
);
5505 #ifdef CONFIG_HOTPLUG_CPU
5507 * Force a reinitialization of the sched domains hierarchy. The domains
5508 * and groups cannot be updated in place without racing with the balancing
5509 * code, so we temporarily attach all running cpus to the NULL domain
5510 * which will prevent rebalancing while the sched domains are recalculated.
5512 static int update_sched_domains(struct notifier_block
*nfb
,
5513 unsigned long action
, void *hcpu
)
5516 case CPU_UP_PREPARE
:
5517 case CPU_DOWN_PREPARE
:
5518 detach_destroy_domains(&cpu_online_map
);
5521 case CPU_UP_CANCELED
:
5522 case CPU_DOWN_FAILED
:
5526 * Fall through and re-initialise the domains.
5533 /* The hotplug lock is already held by cpu_up/cpu_down */
5534 arch_init_sched_domains(&cpu_online_map
);
5540 void __init
sched_init_smp(void)
5543 arch_init_sched_domains(&cpu_online_map
);
5544 unlock_cpu_hotplug();
5545 /* XXX: Theoretical race here - CPU may be hotplugged now */
5546 hotcpu_notifier(update_sched_domains
, 0);
5549 void __init
sched_init_smp(void)
5552 #endif /* CONFIG_SMP */
5554 int in_sched_functions(unsigned long addr
)
5556 /* Linker adds these: start and end of __sched functions */
5557 extern char __sched_text_start
[], __sched_text_end
[];
5558 return in_lock_functions(addr
) ||
5559 (addr
>= (unsigned long)__sched_text_start
5560 && addr
< (unsigned long)__sched_text_end
);
5563 void __init
sched_init(void)
5568 for (i
= 0; i
< NR_CPUS
; i
++) {
5569 prio_array_t
*array
;
5572 spin_lock_init(&rq
->lock
);
5574 rq
->active
= rq
->arrays
;
5575 rq
->expired
= rq
->arrays
+ 1;
5576 rq
->best_expired_prio
= MAX_PRIO
;
5580 for (j
= 1; j
< 3; j
++)
5581 rq
->cpu_load
[j
] = 0;
5582 rq
->active_balance
= 0;
5584 rq
->migration_thread
= NULL
;
5585 INIT_LIST_HEAD(&rq
->migration_queue
);
5587 atomic_set(&rq
->nr_iowait
, 0);
5589 for (j
= 0; j
< 2; j
++) {
5590 array
= rq
->arrays
+ j
;
5591 for (k
= 0; k
< MAX_PRIO
; k
++) {
5592 INIT_LIST_HEAD(array
->queue
+ k
);
5593 __clear_bit(k
, array
->bitmap
);
5595 // delimiter for bitsearch
5596 __set_bit(MAX_PRIO
, array
->bitmap
);
5601 * The boot idle thread does lazy MMU switching as well:
5603 atomic_inc(&init_mm
.mm_count
);
5604 enter_lazy_tlb(&init_mm
, current
);
5607 * Make us the idle thread. Technically, schedule() should not be
5608 * called from this thread, however somewhere below it might be,
5609 * but because we are the idle thread, we just pick up running again
5610 * when this runqueue becomes "idle".
5612 init_idle(current
, smp_processor_id());
5615 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5616 void __might_sleep(char *file
, int line
)
5618 #if defined(in_atomic)
5619 static unsigned long prev_jiffy
; /* ratelimiting */
5621 if ((in_atomic() || irqs_disabled()) &&
5622 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
5623 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
5625 prev_jiffy
= jiffies
;
5626 printk(KERN_ERR
"Debug: sleeping function called from invalid"
5627 " context at %s:%d\n", file
, line
);
5628 printk("in_atomic():%d, irqs_disabled():%d\n",
5629 in_atomic(), irqs_disabled());
5634 EXPORT_SYMBOL(__might_sleep
);
5637 #ifdef CONFIG_MAGIC_SYSRQ
5638 void normalize_rt_tasks(void)
5640 struct task_struct
*p
;
5641 prio_array_t
*array
;
5642 unsigned long flags
;
5645 read_lock_irq(&tasklist_lock
);
5646 for_each_process (p
) {
5650 rq
= task_rq_lock(p
, &flags
);
5654 deactivate_task(p
, task_rq(p
));
5655 __setscheduler(p
, SCHED_NORMAL
, 0);
5657 __activate_task(p
, task_rq(p
));
5658 resched_task(rq
->curr
);
5661 task_rq_unlock(rq
, &flags
);
5663 read_unlock_irq(&tasklist_lock
);
5666 #endif /* CONFIG_MAGIC_SYSRQ */
5670 * These functions are only useful for the IA64 MCA handling.
5672 * They can only be called when the whole system has been
5673 * stopped - every CPU needs to be quiescent, and no scheduling
5674 * activity can take place. Using them for anything else would
5675 * be a serious bug, and as a result, they aren't even visible
5676 * under any other configuration.
5680 * curr_task - return the current task for a given cpu.
5681 * @cpu: the processor in question.
5683 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
5685 task_t
*curr_task(int cpu
)
5687 return cpu_curr(cpu
);
5691 * set_curr_task - set the current task for a given cpu.
5692 * @cpu: the processor in question.
5693 * @p: the task pointer to set.
5695 * Description: This function must only be used when non-maskable interrupts
5696 * are serviced on a separate stack. It allows the architecture to switch the
5697 * notion of the current task on a cpu in a non-blocking manner. This function
5698 * must be called with all CPU's synchronized, and interrupts disabled, the
5699 * and caller must save the original value of the current task (see
5700 * curr_task() above) and restore that value before reenabling interrupts and
5701 * re-starting the system.
5703 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
5705 void set_curr_task(int cpu
, task_t
*p
)