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 cpu_load
[3];
211 unsigned long long nr_switches
;
214 * This is part of a global counter where only the total sum
215 * over all CPUs matters. A task can increase this counter on
216 * one CPU and if it got migrated afterwards it may decrease
217 * it on another CPU. Always updated under the runqueue lock:
219 unsigned long nr_uninterruptible
;
221 unsigned long expired_timestamp
;
222 unsigned long long timestamp_last_tick
;
224 struct mm_struct
*prev_mm
;
225 prio_array_t
*active
, *expired
, arrays
[2];
226 int best_expired_prio
;
230 struct sched_domain
*sd
;
232 /* For active balancing */
236 task_t
*migration_thread
;
237 struct list_head migration_queue
;
240 #ifdef CONFIG_SCHEDSTATS
242 struct sched_info rq_sched_info
;
244 /* sys_sched_yield() stats */
245 unsigned long yld_exp_empty
;
246 unsigned long yld_act_empty
;
247 unsigned long yld_both_empty
;
248 unsigned long yld_cnt
;
250 /* schedule() stats */
251 unsigned long sched_switch
;
252 unsigned long sched_cnt
;
253 unsigned long sched_goidle
;
255 /* try_to_wake_up() stats */
256 unsigned long ttwu_cnt
;
257 unsigned long ttwu_local
;
261 static DEFINE_PER_CPU(struct runqueue
, runqueues
);
264 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
265 * See detach_destroy_domains: synchronize_sched for details.
267 * The domain tree of any CPU may only be accessed from within
268 * preempt-disabled sections.
270 #define for_each_domain(cpu, domain) \
271 for (domain = rcu_dereference(cpu_rq(cpu)->sd); domain; domain = domain->parent)
273 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
274 #define this_rq() (&__get_cpu_var(runqueues))
275 #define task_rq(p) cpu_rq(task_cpu(p))
276 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
278 #ifndef prepare_arch_switch
279 # define prepare_arch_switch(next) do { } while (0)
281 #ifndef finish_arch_switch
282 # define finish_arch_switch(prev) do { } while (0)
285 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
286 static inline int task_running(runqueue_t
*rq
, task_t
*p
)
288 return rq
->curr
== p
;
291 static inline void prepare_lock_switch(runqueue_t
*rq
, task_t
*next
)
295 static inline void finish_lock_switch(runqueue_t
*rq
, task_t
*prev
)
297 spin_unlock_irq(&rq
->lock
);
300 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
301 static inline int task_running(runqueue_t
*rq
, task_t
*p
)
306 return rq
->curr
== p
;
310 static inline void prepare_lock_switch(runqueue_t
*rq
, task_t
*next
)
314 * We can optimise this out completely for !SMP, because the
315 * SMP rebalancing from interrupt is the only thing that cares
320 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
321 spin_unlock_irq(&rq
->lock
);
323 spin_unlock(&rq
->lock
);
327 static inline void finish_lock_switch(runqueue_t
*rq
, task_t
*prev
)
331 * After ->oncpu is cleared, the task can be moved to a different CPU.
332 * We must ensure this doesn't happen until the switch is completely
338 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
342 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
345 * task_rq_lock - lock the runqueue a given task resides on and disable
346 * interrupts. Note the ordering: we can safely lookup the task_rq without
347 * explicitly disabling preemption.
349 static inline runqueue_t
*task_rq_lock(task_t
*p
, unsigned long *flags
)
355 local_irq_save(*flags
);
357 spin_lock(&rq
->lock
);
358 if (unlikely(rq
!= task_rq(p
))) {
359 spin_unlock_irqrestore(&rq
->lock
, *flags
);
360 goto repeat_lock_task
;
365 static inline void task_rq_unlock(runqueue_t
*rq
, unsigned long *flags
)
368 spin_unlock_irqrestore(&rq
->lock
, *flags
);
371 #ifdef CONFIG_SCHEDSTATS
373 * bump this up when changing the output format or the meaning of an existing
374 * format, so that tools can adapt (or abort)
376 #define SCHEDSTAT_VERSION 12
378 static int show_schedstat(struct seq_file
*seq
, void *v
)
382 seq_printf(seq
, "version %d\n", SCHEDSTAT_VERSION
);
383 seq_printf(seq
, "timestamp %lu\n", jiffies
);
384 for_each_online_cpu(cpu
) {
385 runqueue_t
*rq
= cpu_rq(cpu
);
387 struct sched_domain
*sd
;
391 /* runqueue-specific stats */
393 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
394 cpu
, rq
->yld_both_empty
,
395 rq
->yld_act_empty
, rq
->yld_exp_empty
, rq
->yld_cnt
,
396 rq
->sched_switch
, rq
->sched_cnt
, rq
->sched_goidle
,
397 rq
->ttwu_cnt
, rq
->ttwu_local
,
398 rq
->rq_sched_info
.cpu_time
,
399 rq
->rq_sched_info
.run_delay
, rq
->rq_sched_info
.pcnt
);
401 seq_printf(seq
, "\n");
404 /* domain-specific stats */
406 for_each_domain(cpu
, sd
) {
407 enum idle_type itype
;
408 char mask_str
[NR_CPUS
];
410 cpumask_scnprintf(mask_str
, NR_CPUS
, sd
->span
);
411 seq_printf(seq
, "domain%d %s", dcnt
++, mask_str
);
412 for (itype
= SCHED_IDLE
; itype
< MAX_IDLE_TYPES
;
414 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu %lu",
416 sd
->lb_balanced
[itype
],
417 sd
->lb_failed
[itype
],
418 sd
->lb_imbalance
[itype
],
419 sd
->lb_gained
[itype
],
420 sd
->lb_hot_gained
[itype
],
421 sd
->lb_nobusyq
[itype
],
422 sd
->lb_nobusyg
[itype
]);
424 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
425 sd
->alb_cnt
, sd
->alb_failed
, sd
->alb_pushed
,
426 sd
->sbe_cnt
, sd
->sbe_balanced
, sd
->sbe_pushed
,
427 sd
->sbf_cnt
, sd
->sbf_balanced
, sd
->sbf_pushed
,
428 sd
->ttwu_wake_remote
, sd
->ttwu_move_affine
, sd
->ttwu_move_balance
);
436 static int schedstat_open(struct inode
*inode
, struct file
*file
)
438 unsigned int size
= PAGE_SIZE
* (1 + num_online_cpus() / 32);
439 char *buf
= kmalloc(size
, GFP_KERNEL
);
445 res
= single_open(file
, show_schedstat
, NULL
);
447 m
= file
->private_data
;
455 struct file_operations proc_schedstat_operations
= {
456 .open
= schedstat_open
,
459 .release
= single_release
,
462 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
463 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
464 #else /* !CONFIG_SCHEDSTATS */
465 # define schedstat_inc(rq, field) do { } while (0)
466 # define schedstat_add(rq, field, amt) do { } while (0)
470 * rq_lock - lock a given runqueue and disable interrupts.
472 static inline runqueue_t
*this_rq_lock(void)
479 spin_lock(&rq
->lock
);
484 #ifdef CONFIG_SCHEDSTATS
486 * Called when a process is dequeued from the active array and given
487 * the cpu. We should note that with the exception of interactive
488 * tasks, the expired queue will become the active queue after the active
489 * queue is empty, without explicitly dequeuing and requeuing tasks in the
490 * expired queue. (Interactive tasks may be requeued directly to the
491 * active queue, thus delaying tasks in the expired queue from running;
492 * see scheduler_tick()).
494 * This function is only called from sched_info_arrive(), rather than
495 * dequeue_task(). Even though a task may be queued and dequeued multiple
496 * times as it is shuffled about, we're really interested in knowing how
497 * long it was from the *first* time it was queued to the time that it
500 static inline void sched_info_dequeued(task_t
*t
)
502 t
->sched_info
.last_queued
= 0;
506 * Called when a task finally hits the cpu. We can now calculate how
507 * long it was waiting to run. We also note when it began so that we
508 * can keep stats on how long its timeslice is.
510 static inline void sched_info_arrive(task_t
*t
)
512 unsigned long now
= jiffies
, diff
= 0;
513 struct runqueue
*rq
= task_rq(t
);
515 if (t
->sched_info
.last_queued
)
516 diff
= now
- t
->sched_info
.last_queued
;
517 sched_info_dequeued(t
);
518 t
->sched_info
.run_delay
+= diff
;
519 t
->sched_info
.last_arrival
= now
;
520 t
->sched_info
.pcnt
++;
525 rq
->rq_sched_info
.run_delay
+= diff
;
526 rq
->rq_sched_info
.pcnt
++;
530 * Called when a process is queued into either the active or expired
531 * array. The time is noted and later used to determine how long we
532 * had to wait for us to reach the cpu. Since the expired queue will
533 * become the active queue after active queue is empty, without dequeuing
534 * and requeuing any tasks, we are interested in queuing to either. It
535 * is unusual but not impossible for tasks to be dequeued and immediately
536 * requeued in the same or another array: this can happen in sched_yield(),
537 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
540 * This function is only called from enqueue_task(), but also only updates
541 * the timestamp if it is already not set. It's assumed that
542 * sched_info_dequeued() will clear that stamp when appropriate.
544 static inline void sched_info_queued(task_t
*t
)
546 if (!t
->sched_info
.last_queued
)
547 t
->sched_info
.last_queued
= jiffies
;
551 * Called when a process ceases being the active-running process, either
552 * voluntarily or involuntarily. Now we can calculate how long we ran.
554 static inline void sched_info_depart(task_t
*t
)
556 struct runqueue
*rq
= task_rq(t
);
557 unsigned long diff
= jiffies
- t
->sched_info
.last_arrival
;
559 t
->sched_info
.cpu_time
+= diff
;
562 rq
->rq_sched_info
.cpu_time
+= diff
;
566 * Called when tasks are switched involuntarily due, typically, to expiring
567 * their time slice. (This may also be called when switching to or from
568 * the idle task.) We are only called when prev != next.
570 static inline void sched_info_switch(task_t
*prev
, task_t
*next
)
572 struct runqueue
*rq
= task_rq(prev
);
575 * prev now departs the cpu. It's not interesting to record
576 * stats about how efficient we were at scheduling the idle
579 if (prev
!= rq
->idle
)
580 sched_info_depart(prev
);
582 if (next
!= rq
->idle
)
583 sched_info_arrive(next
);
586 #define sched_info_queued(t) do { } while (0)
587 #define sched_info_switch(t, next) do { } while (0)
588 #endif /* CONFIG_SCHEDSTATS */
591 * Adding/removing a task to/from a priority array:
593 static void dequeue_task(struct task_struct
*p
, prio_array_t
*array
)
596 list_del(&p
->run_list
);
597 if (list_empty(array
->queue
+ p
->prio
))
598 __clear_bit(p
->prio
, array
->bitmap
);
601 static void enqueue_task(struct task_struct
*p
, prio_array_t
*array
)
603 sched_info_queued(p
);
604 list_add_tail(&p
->run_list
, array
->queue
+ p
->prio
);
605 __set_bit(p
->prio
, array
->bitmap
);
611 * Put task to the end of the run list without the overhead of dequeue
612 * followed by enqueue.
614 static void requeue_task(struct task_struct
*p
, prio_array_t
*array
)
616 list_move_tail(&p
->run_list
, array
->queue
+ p
->prio
);
619 static inline void enqueue_task_head(struct task_struct
*p
, prio_array_t
*array
)
621 list_add(&p
->run_list
, array
->queue
+ p
->prio
);
622 __set_bit(p
->prio
, array
->bitmap
);
628 * effective_prio - return the priority that is based on the static
629 * priority but is modified by bonuses/penalties.
631 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
632 * into the -5 ... 0 ... +5 bonus/penalty range.
634 * We use 25% of the full 0...39 priority range so that:
636 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
637 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
639 * Both properties are important to certain workloads.
641 static int effective_prio(task_t
*p
)
648 bonus
= CURRENT_BONUS(p
) - MAX_BONUS
/ 2;
650 prio
= p
->static_prio
- bonus
;
651 if (prio
< MAX_RT_PRIO
)
653 if (prio
> MAX_PRIO
-1)
659 * __activate_task - move a task to the runqueue.
661 static inline void __activate_task(task_t
*p
, runqueue_t
*rq
)
663 enqueue_task(p
, rq
->active
);
668 * __activate_idle_task - move idle task to the _front_ of runqueue.
670 static inline void __activate_idle_task(task_t
*p
, runqueue_t
*rq
)
672 enqueue_task_head(p
, rq
->active
);
676 static int recalc_task_prio(task_t
*p
, unsigned long long now
)
678 /* Caller must always ensure 'now >= p->timestamp' */
679 unsigned long long __sleep_time
= now
- p
->timestamp
;
680 unsigned long sleep_time
;
682 if (__sleep_time
> NS_MAX_SLEEP_AVG
)
683 sleep_time
= NS_MAX_SLEEP_AVG
;
685 sleep_time
= (unsigned long)__sleep_time
;
687 if (likely(sleep_time
> 0)) {
689 * User tasks that sleep a long time are categorised as
690 * idle and will get just interactive status to stay active &
691 * prevent them suddenly becoming cpu hogs and starving
694 if (p
->mm
&& p
->activated
!= -1 &&
695 sleep_time
> INTERACTIVE_SLEEP(p
)) {
696 p
->sleep_avg
= JIFFIES_TO_NS(MAX_SLEEP_AVG
-
700 * The lower the sleep avg a task has the more
701 * rapidly it will rise with sleep time.
703 sleep_time
*= (MAX_BONUS
- CURRENT_BONUS(p
)) ? : 1;
706 * Tasks waking from uninterruptible sleep are
707 * limited in their sleep_avg rise as they
708 * are likely to be waiting on I/O
710 if (p
->activated
== -1 && p
->mm
) {
711 if (p
->sleep_avg
>= INTERACTIVE_SLEEP(p
))
713 else if (p
->sleep_avg
+ sleep_time
>=
714 INTERACTIVE_SLEEP(p
)) {
715 p
->sleep_avg
= INTERACTIVE_SLEEP(p
);
721 * This code gives a bonus to interactive tasks.
723 * The boost works by updating the 'average sleep time'
724 * value here, based on ->timestamp. The more time a
725 * task spends sleeping, the higher the average gets -
726 * and the higher the priority boost gets as well.
728 p
->sleep_avg
+= sleep_time
;
730 if (p
->sleep_avg
> NS_MAX_SLEEP_AVG
)
731 p
->sleep_avg
= NS_MAX_SLEEP_AVG
;
735 return effective_prio(p
);
739 * activate_task - move a task to the runqueue and do priority recalculation
741 * Update all the scheduling statistics stuff. (sleep average
742 * calculation, priority modifiers, etc.)
744 static void activate_task(task_t
*p
, runqueue_t
*rq
, int local
)
746 unsigned long long now
;
751 /* Compensate for drifting sched_clock */
752 runqueue_t
*this_rq
= this_rq();
753 now
= (now
- this_rq
->timestamp_last_tick
)
754 + rq
->timestamp_last_tick
;
758 p
->prio
= recalc_task_prio(p
, now
);
761 * This checks to make sure it's not an uninterruptible task
762 * that is now waking up.
766 * Tasks which were woken up by interrupts (ie. hw events)
767 * are most likely of interactive nature. So we give them
768 * the credit of extending their sleep time to the period
769 * of time they spend on the runqueue, waiting for execution
770 * on a CPU, first time around:
776 * Normal first-time wakeups get a credit too for
777 * on-runqueue time, but it will be weighted down:
784 __activate_task(p
, rq
);
788 * deactivate_task - remove a task from the runqueue.
790 static void deactivate_task(struct task_struct
*p
, runqueue_t
*rq
)
793 dequeue_task(p
, p
->array
);
798 * resched_task - mark a task 'to be rescheduled now'.
800 * On UP this means the setting of the need_resched flag, on SMP it
801 * might also involve a cross-CPU call to trigger the scheduler on
805 static void resched_task(task_t
*p
)
807 int need_resched
, nrpolling
;
809 assert_spin_locked(&task_rq(p
)->lock
);
811 /* minimise the chance of sending an interrupt to poll_idle() */
812 nrpolling
= test_tsk_thread_flag(p
,TIF_POLLING_NRFLAG
);
813 need_resched
= test_and_set_tsk_thread_flag(p
,TIF_NEED_RESCHED
);
814 nrpolling
|= test_tsk_thread_flag(p
,TIF_POLLING_NRFLAG
);
816 if (!need_resched
&& !nrpolling
&& (task_cpu(p
) != smp_processor_id()))
817 smp_send_reschedule(task_cpu(p
));
820 static inline void resched_task(task_t
*p
)
822 set_tsk_need_resched(p
);
827 * task_curr - is this task currently executing on a CPU?
828 * @p: the task in question.
830 inline int task_curr(const task_t
*p
)
832 return cpu_curr(task_cpu(p
)) == p
;
837 struct list_head list
;
842 struct completion done
;
846 * The task's runqueue lock must be held.
847 * Returns true if you have to wait for migration thread.
849 static int migrate_task(task_t
*p
, int dest_cpu
, migration_req_t
*req
)
851 runqueue_t
*rq
= task_rq(p
);
854 * If the task is not on a runqueue (and not running), then
855 * it is sufficient to simply update the task's cpu field.
857 if (!p
->array
&& !task_running(rq
, p
)) {
858 set_task_cpu(p
, dest_cpu
);
862 init_completion(&req
->done
);
864 req
->dest_cpu
= dest_cpu
;
865 list_add(&req
->list
, &rq
->migration_queue
);
870 * wait_task_inactive - wait for a thread to unschedule.
872 * The caller must ensure that the task *will* unschedule sometime soon,
873 * else this function might spin for a *long* time. This function can't
874 * be called with interrupts off, or it may introduce deadlock with
875 * smp_call_function() if an IPI is sent by the same process we are
876 * waiting to become inactive.
878 void wait_task_inactive(task_t
*p
)
885 rq
= task_rq_lock(p
, &flags
);
886 /* Must be off runqueue entirely, not preempted. */
887 if (unlikely(p
->array
|| task_running(rq
, p
))) {
888 /* If it's preempted, we yield. It could be a while. */
889 preempted
= !task_running(rq
, p
);
890 task_rq_unlock(rq
, &flags
);
896 task_rq_unlock(rq
, &flags
);
900 * kick_process - kick a running thread to enter/exit the kernel
901 * @p: the to-be-kicked thread
903 * Cause a process which is running on another CPU to enter
904 * kernel-mode, without any delay. (to get signals handled.)
906 * NOTE: this function doesnt have to take the runqueue lock,
907 * because all it wants to ensure is that the remote task enters
908 * the kernel. If the IPI races and the task has been migrated
909 * to another CPU then no harm is done and the purpose has been
912 void kick_process(task_t
*p
)
918 if ((cpu
!= smp_processor_id()) && task_curr(p
))
919 smp_send_reschedule(cpu
);
924 * Return a low guess at the load of a migration-source cpu.
926 * We want to under-estimate the load of migration sources, to
927 * balance conservatively.
929 static inline unsigned long source_load(int cpu
, int type
)
931 runqueue_t
*rq
= cpu_rq(cpu
);
932 unsigned long load_now
= rq
->nr_running
* SCHED_LOAD_SCALE
;
936 return min(rq
->cpu_load
[type
-1], load_now
);
940 * Return a high guess at the load of a migration-target cpu
942 static inline unsigned long target_load(int cpu
, int type
)
944 runqueue_t
*rq
= cpu_rq(cpu
);
945 unsigned long load_now
= rq
->nr_running
* SCHED_LOAD_SCALE
;
949 return max(rq
->cpu_load
[type
-1], load_now
);
953 * find_idlest_group finds and returns the least busy CPU group within the
956 static struct sched_group
*
957 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
959 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
960 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
961 int load_idx
= sd
->forkexec_idx
;
962 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
965 unsigned long load
, avg_load
;
969 /* Skip over this group if it has no CPUs allowed */
970 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
973 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
975 /* Tally up the load of all CPUs in the group */
978 for_each_cpu_mask(i
, group
->cpumask
) {
979 /* Bias balancing toward cpus of our domain */
981 load
= source_load(i
, load_idx
);
983 load
= target_load(i
, load_idx
);
988 /* Adjust by relative CPU power of the group */
989 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
992 this_load
= avg_load
;
994 } else if (avg_load
< min_load
) {
1000 } while (group
!= sd
->groups
);
1002 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1008 * find_idlest_queue - find the idlest runqueue among the cpus in group.
1011 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1014 unsigned long load
, min_load
= ULONG_MAX
;
1018 /* Traverse only the allowed CPUs */
1019 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1021 for_each_cpu_mask(i
, tmp
) {
1022 load
= source_load(i
, 0);
1024 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1034 * sched_balance_self: balance the current task (running on cpu) in domains
1035 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1038 * Balance, ie. select the least loaded group.
1040 * Returns the target CPU number, or the same CPU if no balancing is needed.
1042 * preempt must be disabled.
1044 static int sched_balance_self(int cpu
, int flag
)
1046 struct task_struct
*t
= current
;
1047 struct sched_domain
*tmp
, *sd
= NULL
;
1049 for_each_domain(cpu
, tmp
)
1050 if (tmp
->flags
& flag
)
1055 struct sched_group
*group
;
1060 group
= find_idlest_group(sd
, t
, cpu
);
1064 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1065 if (new_cpu
== -1 || new_cpu
== cpu
)
1068 /* Now try balancing at a lower domain level */
1072 weight
= cpus_weight(span
);
1073 for_each_domain(cpu
, tmp
) {
1074 if (weight
<= cpus_weight(tmp
->span
))
1076 if (tmp
->flags
& flag
)
1079 /* while loop will break here if sd == NULL */
1085 #endif /* CONFIG_SMP */
1088 * wake_idle() will wake a task on an idle cpu if task->cpu is
1089 * not idle and an idle cpu is available. The span of cpus to
1090 * search starts with cpus closest then further out as needed,
1091 * so we always favor a closer, idle cpu.
1093 * Returns the CPU we should wake onto.
1095 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1096 static int wake_idle(int cpu
, task_t
*p
)
1099 struct sched_domain
*sd
;
1105 for_each_domain(cpu
, sd
) {
1106 if (sd
->flags
& SD_WAKE_IDLE
) {
1107 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1108 for_each_cpu_mask(i
, tmp
) {
1119 static inline int wake_idle(int cpu
, task_t
*p
)
1126 * try_to_wake_up - wake up a thread
1127 * @p: the to-be-woken-up thread
1128 * @state: the mask of task states that can be woken
1129 * @sync: do a synchronous wakeup?
1131 * Put it on the run-queue if it's not already there. The "current"
1132 * thread is always on the run-queue (except when the actual
1133 * re-schedule is in progress), and as such you're allowed to do
1134 * the simpler "current->state = TASK_RUNNING" to mark yourself
1135 * runnable without the overhead of this.
1137 * returns failure only if the task is already active.
1139 static int try_to_wake_up(task_t
*p
, unsigned int state
, int sync
)
1141 int cpu
, this_cpu
, success
= 0;
1142 unsigned long flags
;
1146 unsigned long load
, this_load
;
1147 struct sched_domain
*sd
, *this_sd
= NULL
;
1151 rq
= task_rq_lock(p
, &flags
);
1152 old_state
= p
->state
;
1153 if (!(old_state
& state
))
1160 this_cpu
= smp_processor_id();
1163 if (unlikely(task_running(rq
, p
)))
1168 schedstat_inc(rq
, ttwu_cnt
);
1169 if (cpu
== this_cpu
) {
1170 schedstat_inc(rq
, ttwu_local
);
1174 for_each_domain(this_cpu
, sd
) {
1175 if (cpu_isset(cpu
, sd
->span
)) {
1176 schedstat_inc(sd
, ttwu_wake_remote
);
1182 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1186 * Check for affine wakeup and passive balancing possibilities.
1189 int idx
= this_sd
->wake_idx
;
1190 unsigned int imbalance
;
1192 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1194 load
= source_load(cpu
, idx
);
1195 this_load
= target_load(this_cpu
, idx
);
1197 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1199 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1200 unsigned long tl
= this_load
;
1202 * If sync wakeup then subtract the (maximum possible)
1203 * effect of the currently running task from the load
1204 * of the current CPU:
1207 tl
-= SCHED_LOAD_SCALE
;
1210 tl
+ target_load(cpu
, idx
) <= SCHED_LOAD_SCALE
) ||
1211 100*(tl
+ SCHED_LOAD_SCALE
) <= imbalance
*load
) {
1213 * This domain has SD_WAKE_AFFINE and
1214 * p is cache cold in this domain, and
1215 * there is no bad imbalance.
1217 schedstat_inc(this_sd
, ttwu_move_affine
);
1223 * Start passive balancing when half the imbalance_pct
1226 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1227 if (imbalance
*this_load
<= 100*load
) {
1228 schedstat_inc(this_sd
, ttwu_move_balance
);
1234 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1236 new_cpu
= wake_idle(new_cpu
, p
);
1237 if (new_cpu
!= cpu
) {
1238 set_task_cpu(p
, new_cpu
);
1239 task_rq_unlock(rq
, &flags
);
1240 /* might preempt at this point */
1241 rq
= task_rq_lock(p
, &flags
);
1242 old_state
= p
->state
;
1243 if (!(old_state
& state
))
1248 this_cpu
= smp_processor_id();
1253 #endif /* CONFIG_SMP */
1254 if (old_state
== TASK_UNINTERRUPTIBLE
) {
1255 rq
->nr_uninterruptible
--;
1257 * Tasks on involuntary sleep don't earn
1258 * sleep_avg beyond just interactive state.
1264 * Tasks that have marked their sleep as noninteractive get
1265 * woken up without updating their sleep average. (i.e. their
1266 * sleep is handled in a priority-neutral manner, no priority
1267 * boost and no penalty.)
1269 if (old_state
& TASK_NONINTERACTIVE
)
1270 __activate_task(p
, rq
);
1272 activate_task(p
, rq
, cpu
== this_cpu
);
1274 * Sync wakeups (i.e. those types of wakeups where the waker
1275 * has indicated that it will leave the CPU in short order)
1276 * don't trigger a preemption, if the woken up task will run on
1277 * this cpu. (in this case the 'I will reschedule' promise of
1278 * the waker guarantees that the freshly woken up task is going
1279 * to be considered on this CPU.)
1281 if (!sync
|| cpu
!= this_cpu
) {
1282 if (TASK_PREEMPTS_CURR(p
, rq
))
1283 resched_task(rq
->curr
);
1288 p
->state
= TASK_RUNNING
;
1290 task_rq_unlock(rq
, &flags
);
1295 int fastcall
wake_up_process(task_t
*p
)
1297 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1298 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1301 EXPORT_SYMBOL(wake_up_process
);
1303 int fastcall
wake_up_state(task_t
*p
, unsigned int state
)
1305 return try_to_wake_up(p
, state
, 0);
1309 * Perform scheduler related setup for a newly forked process p.
1310 * p is forked by current.
1312 void fastcall
sched_fork(task_t
*p
, int clone_flags
)
1314 int cpu
= get_cpu();
1317 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1319 set_task_cpu(p
, cpu
);
1322 * We mark the process as running here, but have not actually
1323 * inserted it onto the runqueue yet. This guarantees that
1324 * nobody will actually run it, and a signal or other external
1325 * event cannot wake it up and insert it on the runqueue either.
1327 p
->state
= TASK_RUNNING
;
1328 INIT_LIST_HEAD(&p
->run_list
);
1330 #ifdef CONFIG_SCHEDSTATS
1331 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1333 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1336 #ifdef CONFIG_PREEMPT
1337 /* Want to start with kernel preemption disabled. */
1338 p
->thread_info
->preempt_count
= 1;
1341 * Share the timeslice between parent and child, thus the
1342 * total amount of pending timeslices in the system doesn't change,
1343 * resulting in more scheduling fairness.
1345 local_irq_disable();
1346 p
->time_slice
= (current
->time_slice
+ 1) >> 1;
1348 * The remainder of the first timeslice might be recovered by
1349 * the parent if the child exits early enough.
1351 p
->first_time_slice
= 1;
1352 current
->time_slice
>>= 1;
1353 p
->timestamp
= sched_clock();
1354 if (unlikely(!current
->time_slice
)) {
1356 * This case is rare, it happens when the parent has only
1357 * a single jiffy left from its timeslice. Taking the
1358 * runqueue lock is not a problem.
1360 current
->time_slice
= 1;
1368 * wake_up_new_task - wake up a newly created task for the first time.
1370 * This function will do some initial scheduler statistics housekeeping
1371 * that must be done for every newly created context, then puts the task
1372 * on the runqueue and wakes it.
1374 void fastcall
wake_up_new_task(task_t
*p
, unsigned long clone_flags
)
1376 unsigned long flags
;
1378 runqueue_t
*rq
, *this_rq
;
1380 rq
= task_rq_lock(p
, &flags
);
1381 BUG_ON(p
->state
!= TASK_RUNNING
);
1382 this_cpu
= smp_processor_id();
1386 * We decrease the sleep average of forking parents
1387 * and children as well, to keep max-interactive tasks
1388 * from forking tasks that are max-interactive. The parent
1389 * (current) is done further down, under its lock.
1391 p
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(p
) *
1392 CHILD_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1394 p
->prio
= effective_prio(p
);
1396 if (likely(cpu
== this_cpu
)) {
1397 if (!(clone_flags
& CLONE_VM
)) {
1399 * The VM isn't cloned, so we're in a good position to
1400 * do child-runs-first in anticipation of an exec. This
1401 * usually avoids a lot of COW overhead.
1403 if (unlikely(!current
->array
))
1404 __activate_task(p
, rq
);
1406 p
->prio
= current
->prio
;
1407 list_add_tail(&p
->run_list
, ¤t
->run_list
);
1408 p
->array
= current
->array
;
1409 p
->array
->nr_active
++;
1414 /* Run child last */
1415 __activate_task(p
, rq
);
1417 * We skip the following code due to cpu == this_cpu
1419 * task_rq_unlock(rq, &flags);
1420 * this_rq = task_rq_lock(current, &flags);
1424 this_rq
= cpu_rq(this_cpu
);
1427 * Not the local CPU - must adjust timestamp. This should
1428 * get optimised away in the !CONFIG_SMP case.
1430 p
->timestamp
= (p
->timestamp
- this_rq
->timestamp_last_tick
)
1431 + rq
->timestamp_last_tick
;
1432 __activate_task(p
, rq
);
1433 if (TASK_PREEMPTS_CURR(p
, rq
))
1434 resched_task(rq
->curr
);
1437 * Parent and child are on different CPUs, now get the
1438 * parent runqueue to update the parent's ->sleep_avg:
1440 task_rq_unlock(rq
, &flags
);
1441 this_rq
= task_rq_lock(current
, &flags
);
1443 current
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(current
) *
1444 PARENT_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1445 task_rq_unlock(this_rq
, &flags
);
1449 * Potentially available exiting-child timeslices are
1450 * retrieved here - this way the parent does not get
1451 * penalized for creating too many threads.
1453 * (this cannot be used to 'generate' timeslices
1454 * artificially, because any timeslice recovered here
1455 * was given away by the parent in the first place.)
1457 void fastcall
sched_exit(task_t
*p
)
1459 unsigned long flags
;
1463 * If the child was a (relative-) CPU hog then decrease
1464 * the sleep_avg of the parent as well.
1466 rq
= task_rq_lock(p
->parent
, &flags
);
1467 if (p
->first_time_slice
) {
1468 p
->parent
->time_slice
+= p
->time_slice
;
1469 if (unlikely(p
->parent
->time_slice
> task_timeslice(p
)))
1470 p
->parent
->time_slice
= task_timeslice(p
);
1472 if (p
->sleep_avg
< p
->parent
->sleep_avg
)
1473 p
->parent
->sleep_avg
= p
->parent
->sleep_avg
/
1474 (EXIT_WEIGHT
+ 1) * EXIT_WEIGHT
+ p
->sleep_avg
/
1476 task_rq_unlock(rq
, &flags
);
1480 * prepare_task_switch - prepare to switch tasks
1481 * @rq: the runqueue preparing to switch
1482 * @next: the task we are going to switch to.
1484 * This is called with the rq lock held and interrupts off. It must
1485 * be paired with a subsequent finish_task_switch after the context
1488 * prepare_task_switch sets up locking and calls architecture specific
1491 static inline void prepare_task_switch(runqueue_t
*rq
, task_t
*next
)
1493 prepare_lock_switch(rq
, next
);
1494 prepare_arch_switch(next
);
1498 * finish_task_switch - clean up after a task-switch
1499 * @rq: runqueue associated with task-switch
1500 * @prev: the thread we just switched away from.
1502 * finish_task_switch must be called after the context switch, paired
1503 * with a prepare_task_switch call before the context switch.
1504 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1505 * and do any other architecture-specific cleanup actions.
1507 * Note that we may have delayed dropping an mm in context_switch(). If
1508 * so, we finish that here outside of the runqueue lock. (Doing it
1509 * with the lock held can cause deadlocks; see schedule() for
1512 static inline void finish_task_switch(runqueue_t
*rq
, task_t
*prev
)
1513 __releases(rq
->lock
)
1515 struct mm_struct
*mm
= rq
->prev_mm
;
1516 unsigned long prev_task_flags
;
1521 * A task struct has one reference for the use as "current".
1522 * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1523 * calls schedule one last time. The schedule call will never return,
1524 * and the scheduled task must drop that reference.
1525 * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1526 * still held, otherwise prev could be scheduled on another cpu, die
1527 * there before we look at prev->state, and then the reference would
1529 * Manfred Spraul <manfred@colorfullife.com>
1531 prev_task_flags
= prev
->flags
;
1532 #ifdef CONFIG_DEBUG_SPINLOCK
1533 /* this is a valid case when another task releases the spinlock */
1534 rq
->lock
.owner
= current
;
1536 finish_arch_switch(prev
);
1537 finish_lock_switch(rq
, prev
);
1540 if (unlikely(prev_task_flags
& PF_DEAD
))
1541 put_task_struct(prev
);
1545 * schedule_tail - first thing a freshly forked thread must call.
1546 * @prev: the thread we just switched away from.
1548 asmlinkage
void schedule_tail(task_t
*prev
)
1549 __releases(rq
->lock
)
1551 runqueue_t
*rq
= this_rq();
1552 finish_task_switch(rq
, prev
);
1553 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1554 /* In this case, finish_task_switch does not reenable preemption */
1557 if (current
->set_child_tid
)
1558 put_user(current
->pid
, current
->set_child_tid
);
1562 * context_switch - switch to the new MM and the new
1563 * thread's register state.
1566 task_t
* context_switch(runqueue_t
*rq
, task_t
*prev
, task_t
*next
)
1568 struct mm_struct
*mm
= next
->mm
;
1569 struct mm_struct
*oldmm
= prev
->active_mm
;
1571 if (unlikely(!mm
)) {
1572 next
->active_mm
= oldmm
;
1573 atomic_inc(&oldmm
->mm_count
);
1574 enter_lazy_tlb(oldmm
, next
);
1576 switch_mm(oldmm
, mm
, next
);
1578 if (unlikely(!prev
->mm
)) {
1579 prev
->active_mm
= NULL
;
1580 WARN_ON(rq
->prev_mm
);
1581 rq
->prev_mm
= oldmm
;
1584 /* Here we just switch the register state and the stack. */
1585 switch_to(prev
, next
, prev
);
1591 * nr_running, nr_uninterruptible and nr_context_switches:
1593 * externally visible scheduler statistics: current number of runnable
1594 * threads, current number of uninterruptible-sleeping threads, total
1595 * number of context switches performed since bootup.
1597 unsigned long nr_running(void)
1599 unsigned long i
, sum
= 0;
1601 for_each_online_cpu(i
)
1602 sum
+= cpu_rq(i
)->nr_running
;
1607 unsigned long nr_uninterruptible(void)
1609 unsigned long i
, sum
= 0;
1612 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1615 * Since we read the counters lockless, it might be slightly
1616 * inaccurate. Do not allow it to go below zero though:
1618 if (unlikely((long)sum
< 0))
1624 unsigned long long nr_context_switches(void)
1626 unsigned long long i
, sum
= 0;
1629 sum
+= cpu_rq(i
)->nr_switches
;
1634 unsigned long nr_iowait(void)
1636 unsigned long i
, sum
= 0;
1639 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1647 * double_rq_lock - safely lock two runqueues
1649 * Note this does not disable interrupts like task_rq_lock,
1650 * you need to do so manually before calling.
1652 static void double_rq_lock(runqueue_t
*rq1
, runqueue_t
*rq2
)
1653 __acquires(rq1
->lock
)
1654 __acquires(rq2
->lock
)
1657 spin_lock(&rq1
->lock
);
1658 __acquire(rq2
->lock
); /* Fake it out ;) */
1661 spin_lock(&rq1
->lock
);
1662 spin_lock(&rq2
->lock
);
1664 spin_lock(&rq2
->lock
);
1665 spin_lock(&rq1
->lock
);
1671 * double_rq_unlock - safely unlock two runqueues
1673 * Note this does not restore interrupts like task_rq_unlock,
1674 * you need to do so manually after calling.
1676 static void double_rq_unlock(runqueue_t
*rq1
, runqueue_t
*rq2
)
1677 __releases(rq1
->lock
)
1678 __releases(rq2
->lock
)
1680 spin_unlock(&rq1
->lock
);
1682 spin_unlock(&rq2
->lock
);
1684 __release(rq2
->lock
);
1688 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1690 static void double_lock_balance(runqueue_t
*this_rq
, runqueue_t
*busiest
)
1691 __releases(this_rq
->lock
)
1692 __acquires(busiest
->lock
)
1693 __acquires(this_rq
->lock
)
1695 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1696 if (busiest
< this_rq
) {
1697 spin_unlock(&this_rq
->lock
);
1698 spin_lock(&busiest
->lock
);
1699 spin_lock(&this_rq
->lock
);
1701 spin_lock(&busiest
->lock
);
1706 * If dest_cpu is allowed for this process, migrate the task to it.
1707 * This is accomplished by forcing the cpu_allowed mask to only
1708 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
1709 * the cpu_allowed mask is restored.
1711 static void sched_migrate_task(task_t
*p
, int dest_cpu
)
1713 migration_req_t req
;
1715 unsigned long flags
;
1717 rq
= task_rq_lock(p
, &flags
);
1718 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
1719 || unlikely(cpu_is_offline(dest_cpu
)))
1722 /* force the process onto the specified CPU */
1723 if (migrate_task(p
, dest_cpu
, &req
)) {
1724 /* Need to wait for migration thread (might exit: take ref). */
1725 struct task_struct
*mt
= rq
->migration_thread
;
1726 get_task_struct(mt
);
1727 task_rq_unlock(rq
, &flags
);
1728 wake_up_process(mt
);
1729 put_task_struct(mt
);
1730 wait_for_completion(&req
.done
);
1734 task_rq_unlock(rq
, &flags
);
1738 * sched_exec - execve() is a valuable balancing opportunity, because at
1739 * this point the task has the smallest effective memory and cache footprint.
1741 void sched_exec(void)
1743 int new_cpu
, this_cpu
= get_cpu();
1744 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
1746 if (new_cpu
!= this_cpu
)
1747 sched_migrate_task(current
, new_cpu
);
1751 * pull_task - move a task from a remote runqueue to the local runqueue.
1752 * Both runqueues must be locked.
1755 void pull_task(runqueue_t
*src_rq
, prio_array_t
*src_array
, task_t
*p
,
1756 runqueue_t
*this_rq
, prio_array_t
*this_array
, int this_cpu
)
1758 dequeue_task(p
, src_array
);
1759 src_rq
->nr_running
--;
1760 set_task_cpu(p
, this_cpu
);
1761 this_rq
->nr_running
++;
1762 enqueue_task(p
, this_array
);
1763 p
->timestamp
= (p
->timestamp
- src_rq
->timestamp_last_tick
)
1764 + this_rq
->timestamp_last_tick
;
1766 * Note that idle threads have a prio of MAX_PRIO, for this test
1767 * to be always true for them.
1769 if (TASK_PREEMPTS_CURR(p
, this_rq
))
1770 resched_task(this_rq
->curr
);
1774 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1777 int can_migrate_task(task_t
*p
, runqueue_t
*rq
, int this_cpu
,
1778 struct sched_domain
*sd
, enum idle_type idle
,
1782 * We do not migrate tasks that are:
1783 * 1) running (obviously), or
1784 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1785 * 3) are cache-hot on their current CPU.
1787 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
1791 if (task_running(rq
, p
))
1795 * Aggressive migration if:
1796 * 1) task is cache cold, or
1797 * 2) too many balance attempts have failed.
1800 if (sd
->nr_balance_failed
> sd
->cache_nice_tries
)
1803 if (task_hot(p
, rq
->timestamp_last_tick
, sd
))
1809 * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq,
1810 * as part of a balancing operation within "domain". Returns the number of
1813 * Called with both runqueues locked.
1815 static int move_tasks(runqueue_t
*this_rq
, int this_cpu
, runqueue_t
*busiest
,
1816 unsigned long max_nr_move
, struct sched_domain
*sd
,
1817 enum idle_type idle
, int *all_pinned
)
1819 prio_array_t
*array
, *dst_array
;
1820 struct list_head
*head
, *curr
;
1821 int idx
, pulled
= 0, pinned
= 0;
1824 if (max_nr_move
== 0)
1830 * We first consider expired tasks. Those will likely not be
1831 * executed in the near future, and they are most likely to
1832 * be cache-cold, thus switching CPUs has the least effect
1835 if (busiest
->expired
->nr_active
) {
1836 array
= busiest
->expired
;
1837 dst_array
= this_rq
->expired
;
1839 array
= busiest
->active
;
1840 dst_array
= this_rq
->active
;
1844 /* Start searching at priority 0: */
1848 idx
= sched_find_first_bit(array
->bitmap
);
1850 idx
= find_next_bit(array
->bitmap
, MAX_PRIO
, idx
);
1851 if (idx
>= MAX_PRIO
) {
1852 if (array
== busiest
->expired
&& busiest
->active
->nr_active
) {
1853 array
= busiest
->active
;
1854 dst_array
= this_rq
->active
;
1860 head
= array
->queue
+ idx
;
1863 tmp
= list_entry(curr
, task_t
, run_list
);
1867 if (!can_migrate_task(tmp
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
1874 #ifdef CONFIG_SCHEDSTATS
1875 if (task_hot(tmp
, busiest
->timestamp_last_tick
, sd
))
1876 schedstat_inc(sd
, lb_hot_gained
[idle
]);
1879 pull_task(busiest
, array
, tmp
, this_rq
, dst_array
, this_cpu
);
1882 /* We only want to steal up to the prescribed number of tasks. */
1883 if (pulled
< max_nr_move
) {
1891 * Right now, this is the only place pull_task() is called,
1892 * so we can safely collect pull_task() stats here rather than
1893 * inside pull_task().
1895 schedstat_add(sd
, lb_gained
[idle
], pulled
);
1898 *all_pinned
= pinned
;
1903 * find_busiest_group finds and returns the busiest CPU group within the
1904 * domain. It calculates and returns the number of tasks which should be
1905 * moved to restore balance via the imbalance parameter.
1907 static struct sched_group
*
1908 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
1909 unsigned long *imbalance
, enum idle_type idle
)
1911 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1912 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
1915 max_load
= this_load
= total_load
= total_pwr
= 0;
1916 if (idle
== NOT_IDLE
)
1917 load_idx
= sd
->busy_idx
;
1918 else if (idle
== NEWLY_IDLE
)
1919 load_idx
= sd
->newidle_idx
;
1921 load_idx
= sd
->idle_idx
;
1928 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1930 /* Tally up the load of all CPUs in the group */
1933 for_each_cpu_mask(i
, group
->cpumask
) {
1934 /* Bias balancing toward cpus of our domain */
1936 load
= target_load(i
, load_idx
);
1938 load
= source_load(i
, load_idx
);
1943 total_load
+= avg_load
;
1944 total_pwr
+= group
->cpu_power
;
1946 /* Adjust by relative CPU power of the group */
1947 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
1950 this_load
= avg_load
;
1952 } else if (avg_load
> max_load
) {
1953 max_load
= avg_load
;
1956 group
= group
->next
;
1957 } while (group
!= sd
->groups
);
1959 if (!busiest
|| this_load
>= max_load
)
1962 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
1964 if (this_load
>= avg_load
||
1965 100*max_load
<= sd
->imbalance_pct
*this_load
)
1969 * We're trying to get all the cpus to the average_load, so we don't
1970 * want to push ourselves above the average load, nor do we wish to
1971 * reduce the max loaded cpu below the average load, as either of these
1972 * actions would just result in more rebalancing later, and ping-pong
1973 * tasks around. Thus we look for the minimum possible imbalance.
1974 * Negative imbalances (*we* are more loaded than anyone else) will
1975 * be counted as no imbalance for these purposes -- we can't fix that
1976 * by pulling tasks to us. Be careful of negative numbers as they'll
1977 * appear as very large values with unsigned longs.
1979 /* How much load to actually move to equalise the imbalance */
1980 *imbalance
= min((max_load
- avg_load
) * busiest
->cpu_power
,
1981 (avg_load
- this_load
) * this->cpu_power
)
1984 if (*imbalance
< SCHED_LOAD_SCALE
) {
1985 unsigned long pwr_now
= 0, pwr_move
= 0;
1988 if (max_load
- this_load
>= SCHED_LOAD_SCALE
*2) {
1994 * OK, we don't have enough imbalance to justify moving tasks,
1995 * however we may be able to increase total CPU power used by
1999 pwr_now
+= busiest
->cpu_power
*min(SCHED_LOAD_SCALE
, max_load
);
2000 pwr_now
+= this->cpu_power
*min(SCHED_LOAD_SCALE
, this_load
);
2001 pwr_now
/= SCHED_LOAD_SCALE
;
2003 /* Amount of load we'd subtract */
2004 tmp
= SCHED_LOAD_SCALE
*SCHED_LOAD_SCALE
/busiest
->cpu_power
;
2006 pwr_move
+= busiest
->cpu_power
*min(SCHED_LOAD_SCALE
,
2009 /* Amount of load we'd add */
2010 if (max_load
*busiest
->cpu_power
<
2011 SCHED_LOAD_SCALE
*SCHED_LOAD_SCALE
)
2012 tmp
= max_load
*busiest
->cpu_power
/this->cpu_power
;
2014 tmp
= SCHED_LOAD_SCALE
*SCHED_LOAD_SCALE
/this->cpu_power
;
2015 pwr_move
+= this->cpu_power
*min(SCHED_LOAD_SCALE
, this_load
+ tmp
);
2016 pwr_move
/= SCHED_LOAD_SCALE
;
2018 /* Move if we gain throughput */
2019 if (pwr_move
<= pwr_now
)
2026 /* Get rid of the scaling factor, rounding down as we divide */
2027 *imbalance
= *imbalance
/ SCHED_LOAD_SCALE
;
2037 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2039 static runqueue_t
*find_busiest_queue(struct sched_group
*group
)
2041 unsigned long load
, max_load
= 0;
2042 runqueue_t
*busiest
= NULL
;
2045 for_each_cpu_mask(i
, group
->cpumask
) {
2046 load
= source_load(i
, 0);
2048 if (load
> max_load
) {
2050 busiest
= cpu_rq(i
);
2058 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2059 * so long as it is large enough.
2061 #define MAX_PINNED_INTERVAL 512
2064 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2065 * tasks if there is an imbalance.
2067 * Called with this_rq unlocked.
2069 static int load_balance(int this_cpu
, runqueue_t
*this_rq
,
2070 struct sched_domain
*sd
, enum idle_type idle
)
2072 struct sched_group
*group
;
2073 runqueue_t
*busiest
;
2074 unsigned long imbalance
;
2075 int nr_moved
, all_pinned
= 0;
2076 int active_balance
= 0;
2078 spin_lock(&this_rq
->lock
);
2079 schedstat_inc(sd
, lb_cnt
[idle
]);
2081 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
);
2083 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2087 busiest
= find_busiest_queue(group
);
2089 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2093 BUG_ON(busiest
== this_rq
);
2095 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2098 if (busiest
->nr_running
> 1) {
2100 * Attempt to move tasks. If find_busiest_group has found
2101 * an imbalance but busiest->nr_running <= 1, the group is
2102 * still unbalanced. nr_moved simply stays zero, so it is
2103 * correctly treated as an imbalance.
2105 double_lock_balance(this_rq
, busiest
);
2106 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2107 imbalance
, sd
, idle
, &all_pinned
);
2108 spin_unlock(&busiest
->lock
);
2110 /* All tasks on this runqueue were pinned by CPU affinity */
2111 if (unlikely(all_pinned
))
2115 spin_unlock(&this_rq
->lock
);
2118 schedstat_inc(sd
, lb_failed
[idle
]);
2119 sd
->nr_balance_failed
++;
2121 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2123 spin_lock(&busiest
->lock
);
2124 if (!busiest
->active_balance
) {
2125 busiest
->active_balance
= 1;
2126 busiest
->push_cpu
= this_cpu
;
2129 spin_unlock(&busiest
->lock
);
2131 wake_up_process(busiest
->migration_thread
);
2134 * We've kicked active balancing, reset the failure
2137 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2140 sd
->nr_balance_failed
= 0;
2142 if (likely(!active_balance
)) {
2143 /* We were unbalanced, so reset the balancing interval */
2144 sd
->balance_interval
= sd
->min_interval
;
2147 * If we've begun active balancing, start to back off. This
2148 * case may not be covered by the all_pinned logic if there
2149 * is only 1 task on the busy runqueue (because we don't call
2152 if (sd
->balance_interval
< sd
->max_interval
)
2153 sd
->balance_interval
*= 2;
2159 spin_unlock(&this_rq
->lock
);
2161 schedstat_inc(sd
, lb_balanced
[idle
]);
2163 sd
->nr_balance_failed
= 0;
2164 /* tune up the balancing interval */
2165 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2166 (sd
->balance_interval
< sd
->max_interval
))
2167 sd
->balance_interval
*= 2;
2173 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2174 * tasks if there is an imbalance.
2176 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2177 * this_rq is locked.
2179 static int load_balance_newidle(int this_cpu
, runqueue_t
*this_rq
,
2180 struct sched_domain
*sd
)
2182 struct sched_group
*group
;
2183 runqueue_t
*busiest
= NULL
;
2184 unsigned long imbalance
;
2187 schedstat_inc(sd
, lb_cnt
[NEWLY_IDLE
]);
2188 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, NEWLY_IDLE
);
2190 schedstat_inc(sd
, lb_nobusyg
[NEWLY_IDLE
]);
2194 busiest
= find_busiest_queue(group
);
2196 schedstat_inc(sd
, lb_nobusyq
[NEWLY_IDLE
]);
2200 BUG_ON(busiest
== this_rq
);
2202 schedstat_add(sd
, lb_imbalance
[NEWLY_IDLE
], imbalance
);
2205 if (busiest
->nr_running
> 1) {
2206 /* Attempt to move tasks */
2207 double_lock_balance(this_rq
, busiest
);
2208 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2209 imbalance
, sd
, NEWLY_IDLE
, NULL
);
2210 spin_unlock(&busiest
->lock
);
2214 schedstat_inc(sd
, lb_failed
[NEWLY_IDLE
]);
2216 sd
->nr_balance_failed
= 0;
2221 schedstat_inc(sd
, lb_balanced
[NEWLY_IDLE
]);
2222 sd
->nr_balance_failed
= 0;
2227 * idle_balance is called by schedule() if this_cpu is about to become
2228 * idle. Attempts to pull tasks from other CPUs.
2230 static inline void idle_balance(int this_cpu
, runqueue_t
*this_rq
)
2232 struct sched_domain
*sd
;
2234 for_each_domain(this_cpu
, sd
) {
2235 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
2236 if (load_balance_newidle(this_cpu
, this_rq
, sd
)) {
2237 /* We've pulled tasks over so stop searching */
2245 * active_load_balance is run by migration threads. It pushes running tasks
2246 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2247 * running on each physical CPU where possible, and avoids physical /
2248 * logical imbalances.
2250 * Called with busiest_rq locked.
2252 static void active_load_balance(runqueue_t
*busiest_rq
, int busiest_cpu
)
2254 struct sched_domain
*sd
;
2255 runqueue_t
*target_rq
;
2256 int target_cpu
= busiest_rq
->push_cpu
;
2258 if (busiest_rq
->nr_running
<= 1)
2259 /* no task to move */
2262 target_rq
= cpu_rq(target_cpu
);
2265 * This condition is "impossible", if it occurs
2266 * we need to fix it. Originally reported by
2267 * Bjorn Helgaas on a 128-cpu setup.
2269 BUG_ON(busiest_rq
== target_rq
);
2271 /* move a task from busiest_rq to target_rq */
2272 double_lock_balance(busiest_rq
, target_rq
);
2274 /* Search for an sd spanning us and the target CPU. */
2275 for_each_domain(target_cpu
, sd
)
2276 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2277 cpu_isset(busiest_cpu
, sd
->span
))
2280 if (unlikely(sd
== NULL
))
2283 schedstat_inc(sd
, alb_cnt
);
2285 if (move_tasks(target_rq
, target_cpu
, busiest_rq
, 1, sd
, SCHED_IDLE
, NULL
))
2286 schedstat_inc(sd
, alb_pushed
);
2288 schedstat_inc(sd
, alb_failed
);
2290 spin_unlock(&target_rq
->lock
);
2294 * rebalance_tick will get called every timer tick, on every CPU.
2296 * It checks each scheduling domain to see if it is due to be balanced,
2297 * and initiates a balancing operation if so.
2299 * Balancing parameters are set up in arch_init_sched_domains.
2302 /* Don't have all balancing operations going off at once */
2303 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2305 static void rebalance_tick(int this_cpu
, runqueue_t
*this_rq
,
2306 enum idle_type idle
)
2308 unsigned long old_load
, this_load
;
2309 unsigned long j
= jiffies
+ CPU_OFFSET(this_cpu
);
2310 struct sched_domain
*sd
;
2313 this_load
= this_rq
->nr_running
* SCHED_LOAD_SCALE
;
2314 /* Update our load */
2315 for (i
= 0; i
< 3; i
++) {
2316 unsigned long new_load
= this_load
;
2318 old_load
= this_rq
->cpu_load
[i
];
2320 * Round up the averaging division if load is increasing. This
2321 * prevents us from getting stuck on 9 if the load is 10, for
2324 if (new_load
> old_load
)
2325 new_load
+= scale
-1;
2326 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) / scale
;
2329 for_each_domain(this_cpu
, sd
) {
2330 unsigned long interval
;
2332 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2335 interval
= sd
->balance_interval
;
2336 if (idle
!= SCHED_IDLE
)
2337 interval
*= sd
->busy_factor
;
2339 /* scale ms to jiffies */
2340 interval
= msecs_to_jiffies(interval
);
2341 if (unlikely(!interval
))
2344 if (j
- sd
->last_balance
>= interval
) {
2345 if (load_balance(this_cpu
, this_rq
, sd
, idle
)) {
2346 /* We've pulled tasks over so no longer idle */
2349 sd
->last_balance
+= interval
;
2355 * on UP we do not need to balance between CPUs:
2357 static inline void rebalance_tick(int cpu
, runqueue_t
*rq
, enum idle_type idle
)
2360 static inline void idle_balance(int cpu
, runqueue_t
*rq
)
2365 static inline int wake_priority_sleeper(runqueue_t
*rq
)
2368 #ifdef CONFIG_SCHED_SMT
2369 spin_lock(&rq
->lock
);
2371 * If an SMT sibling task has been put to sleep for priority
2372 * reasons reschedule the idle task to see if it can now run.
2374 if (rq
->nr_running
) {
2375 resched_task(rq
->idle
);
2378 spin_unlock(&rq
->lock
);
2383 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2385 EXPORT_PER_CPU_SYMBOL(kstat
);
2388 * This is called on clock ticks and on context switches.
2389 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2391 static inline void update_cpu_clock(task_t
*p
, runqueue_t
*rq
,
2392 unsigned long long now
)
2394 unsigned long long last
= max(p
->timestamp
, rq
->timestamp_last_tick
);
2395 p
->sched_time
+= now
- last
;
2399 * Return current->sched_time plus any more ns on the sched_clock
2400 * that have not yet been banked.
2402 unsigned long long current_sched_time(const task_t
*tsk
)
2404 unsigned long long ns
;
2405 unsigned long flags
;
2406 local_irq_save(flags
);
2407 ns
= max(tsk
->timestamp
, task_rq(tsk
)->timestamp_last_tick
);
2408 ns
= tsk
->sched_time
+ (sched_clock() - ns
);
2409 local_irq_restore(flags
);
2414 * We place interactive tasks back into the active array, if possible.
2416 * To guarantee that this does not starve expired tasks we ignore the
2417 * interactivity of a task if the first expired task had to wait more
2418 * than a 'reasonable' amount of time. This deadline timeout is
2419 * load-dependent, as the frequency of array switched decreases with
2420 * increasing number of running tasks. We also ignore the interactivity
2421 * if a better static_prio task has expired:
2423 #define EXPIRED_STARVING(rq) \
2424 ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2425 (jiffies - (rq)->expired_timestamp >= \
2426 STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2427 ((rq)->curr->static_prio > (rq)->best_expired_prio))
2430 * Account user cpu time to a process.
2431 * @p: the process that the cpu time gets accounted to
2432 * @hardirq_offset: the offset to subtract from hardirq_count()
2433 * @cputime: the cpu time spent in user space since the last update
2435 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
2437 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2440 p
->utime
= cputime_add(p
->utime
, cputime
);
2442 /* Add user time to cpustat. */
2443 tmp
= cputime_to_cputime64(cputime
);
2444 if (TASK_NICE(p
) > 0)
2445 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
2447 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
2451 * Account system cpu time to a process.
2452 * @p: the process that the cpu time gets accounted to
2453 * @hardirq_offset: the offset to subtract from hardirq_count()
2454 * @cputime: the cpu time spent in kernel space since the last update
2456 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
2459 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2460 runqueue_t
*rq
= this_rq();
2463 p
->stime
= cputime_add(p
->stime
, cputime
);
2465 /* Add system time to cpustat. */
2466 tmp
= cputime_to_cputime64(cputime
);
2467 if (hardirq_count() - hardirq_offset
)
2468 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
2469 else if (softirq_count())
2470 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
2471 else if (p
!= rq
->idle
)
2472 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
2473 else if (atomic_read(&rq
->nr_iowait
) > 0)
2474 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
2476 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
2477 /* Account for system time used */
2478 acct_update_integrals(p
);
2479 /* Update rss highwater mark */
2480 update_mem_hiwater(p
);
2484 * Account for involuntary wait time.
2485 * @p: the process from which the cpu time has been stolen
2486 * @steal: the cpu time spent in involuntary wait
2488 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
2490 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2491 cputime64_t tmp
= cputime_to_cputime64(steal
);
2492 runqueue_t
*rq
= this_rq();
2494 if (p
== rq
->idle
) {
2495 p
->stime
= cputime_add(p
->stime
, steal
);
2496 if (atomic_read(&rq
->nr_iowait
) > 0)
2497 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
2499 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
2501 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
2505 * This function gets called by the timer code, with HZ frequency.
2506 * We call it with interrupts disabled.
2508 * It also gets called by the fork code, when changing the parent's
2511 void scheduler_tick(void)
2513 int cpu
= smp_processor_id();
2514 runqueue_t
*rq
= this_rq();
2515 task_t
*p
= current
;
2516 unsigned long long now
= sched_clock();
2518 update_cpu_clock(p
, rq
, now
);
2520 rq
->timestamp_last_tick
= now
;
2522 if (p
== rq
->idle
) {
2523 if (wake_priority_sleeper(rq
))
2525 rebalance_tick(cpu
, rq
, SCHED_IDLE
);
2529 /* Task might have expired already, but not scheduled off yet */
2530 if (p
->array
!= rq
->active
) {
2531 set_tsk_need_resched(p
);
2534 spin_lock(&rq
->lock
);
2536 * The task was running during this tick - update the
2537 * time slice counter. Note: we do not update a thread's
2538 * priority until it either goes to sleep or uses up its
2539 * timeslice. This makes it possible for interactive tasks
2540 * to use up their timeslices at their highest priority levels.
2544 * RR tasks need a special form of timeslice management.
2545 * FIFO tasks have no timeslices.
2547 if ((p
->policy
== SCHED_RR
) && !--p
->time_slice
) {
2548 p
->time_slice
= task_timeslice(p
);
2549 p
->first_time_slice
= 0;
2550 set_tsk_need_resched(p
);
2552 /* put it at the end of the queue: */
2553 requeue_task(p
, rq
->active
);
2557 if (!--p
->time_slice
) {
2558 dequeue_task(p
, rq
->active
);
2559 set_tsk_need_resched(p
);
2560 p
->prio
= effective_prio(p
);
2561 p
->time_slice
= task_timeslice(p
);
2562 p
->first_time_slice
= 0;
2564 if (!rq
->expired_timestamp
)
2565 rq
->expired_timestamp
= jiffies
;
2566 if (!TASK_INTERACTIVE(p
) || EXPIRED_STARVING(rq
)) {
2567 enqueue_task(p
, rq
->expired
);
2568 if (p
->static_prio
< rq
->best_expired_prio
)
2569 rq
->best_expired_prio
= p
->static_prio
;
2571 enqueue_task(p
, rq
->active
);
2574 * Prevent a too long timeslice allowing a task to monopolize
2575 * the CPU. We do this by splitting up the timeslice into
2578 * Note: this does not mean the task's timeslices expire or
2579 * get lost in any way, they just might be preempted by
2580 * another task of equal priority. (one with higher
2581 * priority would have preempted this task already.) We
2582 * requeue this task to the end of the list on this priority
2583 * level, which is in essence a round-robin of tasks with
2586 * This only applies to tasks in the interactive
2587 * delta range with at least TIMESLICE_GRANULARITY to requeue.
2589 if (TASK_INTERACTIVE(p
) && !((task_timeslice(p
) -
2590 p
->time_slice
) % TIMESLICE_GRANULARITY(p
)) &&
2591 (p
->time_slice
>= TIMESLICE_GRANULARITY(p
)) &&
2592 (p
->array
== rq
->active
)) {
2594 requeue_task(p
, rq
->active
);
2595 set_tsk_need_resched(p
);
2599 spin_unlock(&rq
->lock
);
2601 rebalance_tick(cpu
, rq
, NOT_IDLE
);
2604 #ifdef CONFIG_SCHED_SMT
2605 static inline void wakeup_busy_runqueue(runqueue_t
*rq
)
2607 /* If an SMT runqueue is sleeping due to priority reasons wake it up */
2608 if (rq
->curr
== rq
->idle
&& rq
->nr_running
)
2609 resched_task(rq
->idle
);
2612 static inline void wake_sleeping_dependent(int this_cpu
, runqueue_t
*this_rq
)
2614 struct sched_domain
*tmp
, *sd
= NULL
;
2615 cpumask_t sibling_map
;
2618 for_each_domain(this_cpu
, tmp
)
2619 if (tmp
->flags
& SD_SHARE_CPUPOWER
)
2626 * Unlock the current runqueue because we have to lock in
2627 * CPU order to avoid deadlocks. Caller knows that we might
2628 * unlock. We keep IRQs disabled.
2630 spin_unlock(&this_rq
->lock
);
2632 sibling_map
= sd
->span
;
2634 for_each_cpu_mask(i
, sibling_map
)
2635 spin_lock(&cpu_rq(i
)->lock
);
2637 * We clear this CPU from the mask. This both simplifies the
2638 * inner loop and keps this_rq locked when we exit:
2640 cpu_clear(this_cpu
, sibling_map
);
2642 for_each_cpu_mask(i
, sibling_map
) {
2643 runqueue_t
*smt_rq
= cpu_rq(i
);
2645 wakeup_busy_runqueue(smt_rq
);
2648 for_each_cpu_mask(i
, sibling_map
)
2649 spin_unlock(&cpu_rq(i
)->lock
);
2651 * We exit with this_cpu's rq still held and IRQs
2657 * number of 'lost' timeslices this task wont be able to fully
2658 * utilize, if another task runs on a sibling. This models the
2659 * slowdown effect of other tasks running on siblings:
2661 static inline unsigned long smt_slice(task_t
*p
, struct sched_domain
*sd
)
2663 return p
->time_slice
* (100 - sd
->per_cpu_gain
) / 100;
2666 static inline int dependent_sleeper(int this_cpu
, runqueue_t
*this_rq
)
2668 struct sched_domain
*tmp
, *sd
= NULL
;
2669 cpumask_t sibling_map
;
2670 prio_array_t
*array
;
2674 for_each_domain(this_cpu
, tmp
)
2675 if (tmp
->flags
& SD_SHARE_CPUPOWER
)
2682 * The same locking rules and details apply as for
2683 * wake_sleeping_dependent():
2685 spin_unlock(&this_rq
->lock
);
2686 sibling_map
= sd
->span
;
2687 for_each_cpu_mask(i
, sibling_map
)
2688 spin_lock(&cpu_rq(i
)->lock
);
2689 cpu_clear(this_cpu
, sibling_map
);
2692 * Establish next task to be run - it might have gone away because
2693 * we released the runqueue lock above:
2695 if (!this_rq
->nr_running
)
2697 array
= this_rq
->active
;
2698 if (!array
->nr_active
)
2699 array
= this_rq
->expired
;
2700 BUG_ON(!array
->nr_active
);
2702 p
= list_entry(array
->queue
[sched_find_first_bit(array
->bitmap
)].next
,
2705 for_each_cpu_mask(i
, sibling_map
) {
2706 runqueue_t
*smt_rq
= cpu_rq(i
);
2707 task_t
*smt_curr
= smt_rq
->curr
;
2709 /* Kernel threads do not participate in dependent sleeping */
2710 if (!p
->mm
|| !smt_curr
->mm
|| rt_task(p
))
2711 goto check_smt_task
;
2714 * If a user task with lower static priority than the
2715 * running task on the SMT sibling is trying to schedule,
2716 * delay it till there is proportionately less timeslice
2717 * left of the sibling task to prevent a lower priority
2718 * task from using an unfair proportion of the
2719 * physical cpu's resources. -ck
2721 if (rt_task(smt_curr
)) {
2723 * With real time tasks we run non-rt tasks only
2724 * per_cpu_gain% of the time.
2726 if ((jiffies
% DEF_TIMESLICE
) >
2727 (sd
->per_cpu_gain
* DEF_TIMESLICE
/ 100))
2730 if (smt_curr
->static_prio
< p
->static_prio
&&
2731 !TASK_PREEMPTS_CURR(p
, smt_rq
) &&
2732 smt_slice(smt_curr
, sd
) > task_timeslice(p
))
2736 if ((!smt_curr
->mm
&& smt_curr
!= smt_rq
->idle
) ||
2740 wakeup_busy_runqueue(smt_rq
);
2745 * Reschedule a lower priority task on the SMT sibling for
2746 * it to be put to sleep, or wake it up if it has been put to
2747 * sleep for priority reasons to see if it should run now.
2750 if ((jiffies
% DEF_TIMESLICE
) >
2751 (sd
->per_cpu_gain
* DEF_TIMESLICE
/ 100))
2752 resched_task(smt_curr
);
2754 if (TASK_PREEMPTS_CURR(p
, smt_rq
) &&
2755 smt_slice(p
, sd
) > task_timeslice(smt_curr
))
2756 resched_task(smt_curr
);
2758 wakeup_busy_runqueue(smt_rq
);
2762 for_each_cpu_mask(i
, sibling_map
)
2763 spin_unlock(&cpu_rq(i
)->lock
);
2767 static inline void wake_sleeping_dependent(int this_cpu
, runqueue_t
*this_rq
)
2771 static inline int dependent_sleeper(int this_cpu
, runqueue_t
*this_rq
)
2777 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
2779 void fastcall
add_preempt_count(int val
)
2784 BUG_ON((preempt_count() < 0));
2785 preempt_count() += val
;
2787 * Spinlock count overflowing soon?
2789 BUG_ON((preempt_count() & PREEMPT_MASK
) >= PREEMPT_MASK
-10);
2791 EXPORT_SYMBOL(add_preempt_count
);
2793 void fastcall
sub_preempt_count(int val
)
2798 BUG_ON(val
> preempt_count());
2800 * Is the spinlock portion underflowing?
2802 BUG_ON((val
< PREEMPT_MASK
) && !(preempt_count() & PREEMPT_MASK
));
2803 preempt_count() -= val
;
2805 EXPORT_SYMBOL(sub_preempt_count
);
2810 * schedule() is the main scheduler function.
2812 asmlinkage
void __sched
schedule(void)
2815 task_t
*prev
, *next
;
2817 prio_array_t
*array
;
2818 struct list_head
*queue
;
2819 unsigned long long now
;
2820 unsigned long run_time
;
2821 int cpu
, idx
, new_prio
;
2824 * Test if we are atomic. Since do_exit() needs to call into
2825 * schedule() atomically, we ignore that path for now.
2826 * Otherwise, whine if we are scheduling when we should not be.
2828 if (likely(!current
->exit_state
)) {
2829 if (unlikely(in_atomic())) {
2830 printk(KERN_ERR
"scheduling while atomic: "
2832 current
->comm
, preempt_count(), current
->pid
);
2836 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2841 release_kernel_lock(prev
);
2842 need_resched_nonpreemptible
:
2846 * The idle thread is not allowed to schedule!
2847 * Remove this check after it has been exercised a bit.
2849 if (unlikely(prev
== rq
->idle
) && prev
->state
!= TASK_RUNNING
) {
2850 printk(KERN_ERR
"bad: scheduling from the idle thread!\n");
2854 schedstat_inc(rq
, sched_cnt
);
2855 now
= sched_clock();
2856 if (likely((long long)(now
- prev
->timestamp
) < NS_MAX_SLEEP_AVG
)) {
2857 run_time
= now
- prev
->timestamp
;
2858 if (unlikely((long long)(now
- prev
->timestamp
) < 0))
2861 run_time
= NS_MAX_SLEEP_AVG
;
2864 * Tasks charged proportionately less run_time at high sleep_avg to
2865 * delay them losing their interactive status
2867 run_time
/= (CURRENT_BONUS(prev
) ? : 1);
2869 spin_lock_irq(&rq
->lock
);
2871 if (unlikely(prev
->flags
& PF_DEAD
))
2872 prev
->state
= EXIT_DEAD
;
2874 switch_count
= &prev
->nivcsw
;
2875 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
2876 switch_count
= &prev
->nvcsw
;
2877 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
2878 unlikely(signal_pending(prev
))))
2879 prev
->state
= TASK_RUNNING
;
2881 if (prev
->state
== TASK_UNINTERRUPTIBLE
)
2882 rq
->nr_uninterruptible
++;
2883 deactivate_task(prev
, rq
);
2887 cpu
= smp_processor_id();
2888 if (unlikely(!rq
->nr_running
)) {
2890 idle_balance(cpu
, rq
);
2891 if (!rq
->nr_running
) {
2893 rq
->expired_timestamp
= 0;
2894 wake_sleeping_dependent(cpu
, rq
);
2896 * wake_sleeping_dependent() might have released
2897 * the runqueue, so break out if we got new
2900 if (!rq
->nr_running
)
2904 if (dependent_sleeper(cpu
, rq
)) {
2909 * dependent_sleeper() releases and reacquires the runqueue
2910 * lock, hence go into the idle loop if the rq went
2913 if (unlikely(!rq
->nr_running
))
2918 if (unlikely(!array
->nr_active
)) {
2920 * Switch the active and expired arrays.
2922 schedstat_inc(rq
, sched_switch
);
2923 rq
->active
= rq
->expired
;
2924 rq
->expired
= array
;
2926 rq
->expired_timestamp
= 0;
2927 rq
->best_expired_prio
= MAX_PRIO
;
2930 idx
= sched_find_first_bit(array
->bitmap
);
2931 queue
= array
->queue
+ idx
;
2932 next
= list_entry(queue
->next
, task_t
, run_list
);
2934 if (!rt_task(next
) && next
->activated
> 0) {
2935 unsigned long long delta
= now
- next
->timestamp
;
2936 if (unlikely((long long)(now
- next
->timestamp
) < 0))
2939 if (next
->activated
== 1)
2940 delta
= delta
* (ON_RUNQUEUE_WEIGHT
* 128 / 100) / 128;
2942 array
= next
->array
;
2943 new_prio
= recalc_task_prio(next
, next
->timestamp
+ delta
);
2945 if (unlikely(next
->prio
!= new_prio
)) {
2946 dequeue_task(next
, array
);
2947 next
->prio
= new_prio
;
2948 enqueue_task(next
, array
);
2950 requeue_task(next
, array
);
2952 next
->activated
= 0;
2954 if (next
== rq
->idle
)
2955 schedstat_inc(rq
, sched_goidle
);
2957 prefetch_stack(next
);
2958 clear_tsk_need_resched(prev
);
2959 rcu_qsctr_inc(task_cpu(prev
));
2961 update_cpu_clock(prev
, rq
, now
);
2963 prev
->sleep_avg
-= run_time
;
2964 if ((long)prev
->sleep_avg
<= 0)
2965 prev
->sleep_avg
= 0;
2966 prev
->timestamp
= prev
->last_ran
= now
;
2968 sched_info_switch(prev
, next
);
2969 if (likely(prev
!= next
)) {
2970 next
->timestamp
= now
;
2975 prepare_task_switch(rq
, next
);
2976 prev
= context_switch(rq
, prev
, next
);
2979 * this_rq must be evaluated again because prev may have moved
2980 * CPUs since it called schedule(), thus the 'rq' on its stack
2981 * frame will be invalid.
2983 finish_task_switch(this_rq(), prev
);
2985 spin_unlock_irq(&rq
->lock
);
2988 if (unlikely(reacquire_kernel_lock(prev
) < 0))
2989 goto need_resched_nonpreemptible
;
2990 preempt_enable_no_resched();
2991 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
2995 EXPORT_SYMBOL(schedule
);
2997 #ifdef CONFIG_PREEMPT
2999 * this is is the entry point to schedule() from in-kernel preemption
3000 * off of preempt_enable. Kernel preemptions off return from interrupt
3001 * occur there and call schedule directly.
3003 asmlinkage
void __sched
preempt_schedule(void)
3005 struct thread_info
*ti
= current_thread_info();
3006 #ifdef CONFIG_PREEMPT_BKL
3007 struct task_struct
*task
= current
;
3008 int saved_lock_depth
;
3011 * If there is a non-zero preempt_count or interrupts are disabled,
3012 * we do not want to preempt the current task. Just return..
3014 if (unlikely(ti
->preempt_count
|| irqs_disabled()))
3018 add_preempt_count(PREEMPT_ACTIVE
);
3020 * We keep the big kernel semaphore locked, but we
3021 * clear ->lock_depth so that schedule() doesnt
3022 * auto-release the semaphore:
3024 #ifdef CONFIG_PREEMPT_BKL
3025 saved_lock_depth
= task
->lock_depth
;
3026 task
->lock_depth
= -1;
3029 #ifdef CONFIG_PREEMPT_BKL
3030 task
->lock_depth
= saved_lock_depth
;
3032 sub_preempt_count(PREEMPT_ACTIVE
);
3034 /* we could miss a preemption opportunity between schedule and now */
3036 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3040 EXPORT_SYMBOL(preempt_schedule
);
3043 * this is is the entry point to schedule() from kernel preemption
3044 * off of irq context.
3045 * Note, that this is called and return with irqs disabled. This will
3046 * protect us against recursive calling from irq.
3048 asmlinkage
void __sched
preempt_schedule_irq(void)
3050 struct thread_info
*ti
= current_thread_info();
3051 #ifdef CONFIG_PREEMPT_BKL
3052 struct task_struct
*task
= current
;
3053 int saved_lock_depth
;
3055 /* Catch callers which need to be fixed*/
3056 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3059 add_preempt_count(PREEMPT_ACTIVE
);
3061 * We keep the big kernel semaphore locked, but we
3062 * clear ->lock_depth so that schedule() doesnt
3063 * auto-release the semaphore:
3065 #ifdef CONFIG_PREEMPT_BKL
3066 saved_lock_depth
= task
->lock_depth
;
3067 task
->lock_depth
= -1;
3071 local_irq_disable();
3072 #ifdef CONFIG_PREEMPT_BKL
3073 task
->lock_depth
= saved_lock_depth
;
3075 sub_preempt_count(PREEMPT_ACTIVE
);
3077 /* we could miss a preemption opportunity between schedule and now */
3079 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3083 #endif /* CONFIG_PREEMPT */
3085 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3088 task_t
*p
= curr
->private;
3089 return try_to_wake_up(p
, mode
, sync
);
3092 EXPORT_SYMBOL(default_wake_function
);
3095 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3096 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3097 * number) then we wake all the non-exclusive tasks and one exclusive task.
3099 * There are circumstances in which we can try to wake a task which has already
3100 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3101 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3103 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3104 int nr_exclusive
, int sync
, void *key
)
3106 struct list_head
*tmp
, *next
;
3108 list_for_each_safe(tmp
, next
, &q
->task_list
) {
3111 curr
= list_entry(tmp
, wait_queue_t
, task_list
);
3112 flags
= curr
->flags
;
3113 if (curr
->func(curr
, mode
, sync
, key
) &&
3114 (flags
& WQ_FLAG_EXCLUSIVE
) &&
3121 * __wake_up - wake up threads blocked on a waitqueue.
3123 * @mode: which threads
3124 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3125 * @key: is directly passed to the wakeup function
3127 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3128 int nr_exclusive
, void *key
)
3130 unsigned long flags
;
3132 spin_lock_irqsave(&q
->lock
, flags
);
3133 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3134 spin_unlock_irqrestore(&q
->lock
, flags
);
3137 EXPORT_SYMBOL(__wake_up
);
3140 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3142 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3144 __wake_up_common(q
, mode
, 1, 0, NULL
);
3148 * __wake_up_sync - wake up threads blocked on a waitqueue.
3150 * @mode: which threads
3151 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3153 * The sync wakeup differs that the waker knows that it will schedule
3154 * away soon, so while the target thread will be woken up, it will not
3155 * be migrated to another CPU - ie. the two threads are 'synchronized'
3156 * with each other. This can prevent needless bouncing between CPUs.
3158 * On UP it can prevent extra preemption.
3161 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3163 unsigned long flags
;
3169 if (unlikely(!nr_exclusive
))
3172 spin_lock_irqsave(&q
->lock
, flags
);
3173 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3174 spin_unlock_irqrestore(&q
->lock
, flags
);
3176 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3178 void fastcall
complete(struct completion
*x
)
3180 unsigned long flags
;
3182 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3184 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3186 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3188 EXPORT_SYMBOL(complete
);
3190 void fastcall
complete_all(struct completion
*x
)
3192 unsigned long flags
;
3194 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3195 x
->done
+= UINT_MAX
/2;
3196 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3198 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3200 EXPORT_SYMBOL(complete_all
);
3202 void fastcall __sched
wait_for_completion(struct completion
*x
)
3205 spin_lock_irq(&x
->wait
.lock
);
3207 DECLARE_WAITQUEUE(wait
, current
);
3209 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3210 __add_wait_queue_tail(&x
->wait
, &wait
);
3212 __set_current_state(TASK_UNINTERRUPTIBLE
);
3213 spin_unlock_irq(&x
->wait
.lock
);
3215 spin_lock_irq(&x
->wait
.lock
);
3217 __remove_wait_queue(&x
->wait
, &wait
);
3220 spin_unlock_irq(&x
->wait
.lock
);
3222 EXPORT_SYMBOL(wait_for_completion
);
3224 unsigned long fastcall __sched
3225 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3229 spin_lock_irq(&x
->wait
.lock
);
3231 DECLARE_WAITQUEUE(wait
, current
);
3233 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3234 __add_wait_queue_tail(&x
->wait
, &wait
);
3236 __set_current_state(TASK_UNINTERRUPTIBLE
);
3237 spin_unlock_irq(&x
->wait
.lock
);
3238 timeout
= schedule_timeout(timeout
);
3239 spin_lock_irq(&x
->wait
.lock
);
3241 __remove_wait_queue(&x
->wait
, &wait
);
3245 __remove_wait_queue(&x
->wait
, &wait
);
3249 spin_unlock_irq(&x
->wait
.lock
);
3252 EXPORT_SYMBOL(wait_for_completion_timeout
);
3254 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3260 spin_lock_irq(&x
->wait
.lock
);
3262 DECLARE_WAITQUEUE(wait
, current
);
3264 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3265 __add_wait_queue_tail(&x
->wait
, &wait
);
3267 if (signal_pending(current
)) {
3269 __remove_wait_queue(&x
->wait
, &wait
);
3272 __set_current_state(TASK_INTERRUPTIBLE
);
3273 spin_unlock_irq(&x
->wait
.lock
);
3275 spin_lock_irq(&x
->wait
.lock
);
3277 __remove_wait_queue(&x
->wait
, &wait
);
3281 spin_unlock_irq(&x
->wait
.lock
);
3285 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3287 unsigned long fastcall __sched
3288 wait_for_completion_interruptible_timeout(struct completion
*x
,
3289 unsigned long timeout
)
3293 spin_lock_irq(&x
->wait
.lock
);
3295 DECLARE_WAITQUEUE(wait
, current
);
3297 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3298 __add_wait_queue_tail(&x
->wait
, &wait
);
3300 if (signal_pending(current
)) {
3301 timeout
= -ERESTARTSYS
;
3302 __remove_wait_queue(&x
->wait
, &wait
);
3305 __set_current_state(TASK_INTERRUPTIBLE
);
3306 spin_unlock_irq(&x
->wait
.lock
);
3307 timeout
= schedule_timeout(timeout
);
3308 spin_lock_irq(&x
->wait
.lock
);
3310 __remove_wait_queue(&x
->wait
, &wait
);
3314 __remove_wait_queue(&x
->wait
, &wait
);
3318 spin_unlock_irq(&x
->wait
.lock
);
3321 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3324 #define SLEEP_ON_VAR \
3325 unsigned long flags; \
3326 wait_queue_t wait; \
3327 init_waitqueue_entry(&wait, current);
3329 #define SLEEP_ON_HEAD \
3330 spin_lock_irqsave(&q->lock,flags); \
3331 __add_wait_queue(q, &wait); \
3332 spin_unlock(&q->lock);
3334 #define SLEEP_ON_TAIL \
3335 spin_lock_irq(&q->lock); \
3336 __remove_wait_queue(q, &wait); \
3337 spin_unlock_irqrestore(&q->lock, flags);
3339 void fastcall __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3343 current
->state
= TASK_INTERRUPTIBLE
;
3350 EXPORT_SYMBOL(interruptible_sleep_on
);
3352 long fastcall __sched
3353 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3357 current
->state
= TASK_INTERRUPTIBLE
;
3360 timeout
= schedule_timeout(timeout
);
3366 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3368 void fastcall __sched
sleep_on(wait_queue_head_t
*q
)
3372 current
->state
= TASK_UNINTERRUPTIBLE
;
3379 EXPORT_SYMBOL(sleep_on
);
3381 long fastcall __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3385 current
->state
= TASK_UNINTERRUPTIBLE
;
3388 timeout
= schedule_timeout(timeout
);
3394 EXPORT_SYMBOL(sleep_on_timeout
);
3396 void set_user_nice(task_t
*p
, long nice
)
3398 unsigned long flags
;
3399 prio_array_t
*array
;
3401 int old_prio
, new_prio
, delta
;
3403 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3406 * We have to be careful, if called from sys_setpriority(),
3407 * the task might be in the middle of scheduling on another CPU.
3409 rq
= task_rq_lock(p
, &flags
);
3411 * The RT priorities are set via sched_setscheduler(), but we still
3412 * allow the 'normal' nice value to be set - but as expected
3413 * it wont have any effect on scheduling until the task is
3417 p
->static_prio
= NICE_TO_PRIO(nice
);
3422 dequeue_task(p
, array
);
3425 new_prio
= NICE_TO_PRIO(nice
);
3426 delta
= new_prio
- old_prio
;
3427 p
->static_prio
= NICE_TO_PRIO(nice
);
3431 enqueue_task(p
, array
);
3433 * If the task increased its priority or is running and
3434 * lowered its priority, then reschedule its CPU:
3436 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3437 resched_task(rq
->curr
);
3440 task_rq_unlock(rq
, &flags
);
3443 EXPORT_SYMBOL(set_user_nice
);
3446 * can_nice - check if a task can reduce its nice value
3450 int can_nice(const task_t
*p
, const int nice
)
3452 /* convert nice value [19,-20] to rlimit style value [1,40] */
3453 int nice_rlim
= 20 - nice
;
3454 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
3455 capable(CAP_SYS_NICE
));
3458 #ifdef __ARCH_WANT_SYS_NICE
3461 * sys_nice - change the priority of the current process.
3462 * @increment: priority increment
3464 * sys_setpriority is a more generic, but much slower function that
3465 * does similar things.
3467 asmlinkage
long sys_nice(int increment
)
3473 * Setpriority might change our priority at the same moment.
3474 * We don't have to worry. Conceptually one call occurs first
3475 * and we have a single winner.
3477 if (increment
< -40)
3482 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
3488 if (increment
< 0 && !can_nice(current
, nice
))
3491 retval
= security_task_setnice(current
, nice
);
3495 set_user_nice(current
, nice
);
3502 * task_prio - return the priority value of a given task.
3503 * @p: the task in question.
3505 * This is the priority value as seen by users in /proc.
3506 * RT tasks are offset by -200. Normal tasks are centered
3507 * around 0, value goes from -16 to +15.
3509 int task_prio(const task_t
*p
)
3511 return p
->prio
- MAX_RT_PRIO
;
3515 * task_nice - return the nice value of a given task.
3516 * @p: the task in question.
3518 int task_nice(const task_t
*p
)
3520 return TASK_NICE(p
);
3522 EXPORT_SYMBOL_GPL(task_nice
);
3525 * idle_cpu - is a given cpu idle currently?
3526 * @cpu: the processor in question.
3528 int idle_cpu(int cpu
)
3530 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
3533 EXPORT_SYMBOL_GPL(idle_cpu
);
3536 * idle_task - return the idle task for a given cpu.
3537 * @cpu: the processor in question.
3539 task_t
*idle_task(int cpu
)
3541 return cpu_rq(cpu
)->idle
;
3545 * find_process_by_pid - find a process with a matching PID value.
3546 * @pid: the pid in question.
3548 static inline task_t
*find_process_by_pid(pid_t pid
)
3550 return pid
? find_task_by_pid(pid
) : current
;
3553 /* Actually do priority change: must hold rq lock. */
3554 static void __setscheduler(struct task_struct
*p
, int policy
, int prio
)
3558 p
->rt_priority
= prio
;
3559 if (policy
!= SCHED_NORMAL
)
3560 p
->prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
3562 p
->prio
= p
->static_prio
;
3566 * sched_setscheduler - change the scheduling policy and/or RT priority of
3568 * @p: the task in question.
3569 * @policy: new policy.
3570 * @param: structure containing the new RT priority.
3572 int sched_setscheduler(struct task_struct
*p
, int policy
,
3573 struct sched_param
*param
)
3576 int oldprio
, oldpolicy
= -1;
3577 prio_array_t
*array
;
3578 unsigned long flags
;
3582 /* double check policy once rq lock held */
3584 policy
= oldpolicy
= p
->policy
;
3585 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3586 policy
!= SCHED_NORMAL
)
3589 * Valid priorities for SCHED_FIFO and SCHED_RR are
3590 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL is 0.
3592 if (param
->sched_priority
< 0 ||
3593 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3594 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
3596 if ((policy
== SCHED_NORMAL
) != (param
->sched_priority
== 0))
3600 * Allow unprivileged RT tasks to decrease priority:
3602 if (!capable(CAP_SYS_NICE
)) {
3603 /* can't change policy */
3604 if (policy
!= p
->policy
&&
3605 !p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
)
3607 /* can't increase priority */
3608 if (policy
!= SCHED_NORMAL
&&
3609 param
->sched_priority
> p
->rt_priority
&&
3610 param
->sched_priority
>
3611 p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
)
3613 /* can't change other user's priorities */
3614 if ((current
->euid
!= p
->euid
) &&
3615 (current
->euid
!= p
->uid
))
3619 retval
= security_task_setscheduler(p
, policy
, param
);
3623 * To be able to change p->policy safely, the apropriate
3624 * runqueue lock must be held.
3626 rq
= task_rq_lock(p
, &flags
);
3627 /* recheck policy now with rq lock held */
3628 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3629 policy
= oldpolicy
= -1;
3630 task_rq_unlock(rq
, &flags
);
3635 deactivate_task(p
, rq
);
3637 __setscheduler(p
, policy
, param
->sched_priority
);
3639 __activate_task(p
, rq
);
3641 * Reschedule if we are currently running on this runqueue and
3642 * our priority decreased, or if we are not currently running on
3643 * this runqueue and our priority is higher than the current's
3645 if (task_running(rq
, p
)) {
3646 if (p
->prio
> oldprio
)
3647 resched_task(rq
->curr
);
3648 } else if (TASK_PREEMPTS_CURR(p
, rq
))
3649 resched_task(rq
->curr
);
3651 task_rq_unlock(rq
, &flags
);
3654 EXPORT_SYMBOL_GPL(sched_setscheduler
);
3657 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
3660 struct sched_param lparam
;
3661 struct task_struct
*p
;
3663 if (!param
|| pid
< 0)
3665 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
3667 read_lock_irq(&tasklist_lock
);
3668 p
= find_process_by_pid(pid
);
3670 read_unlock_irq(&tasklist_lock
);
3673 retval
= sched_setscheduler(p
, policy
, &lparam
);
3674 read_unlock_irq(&tasklist_lock
);
3679 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3680 * @pid: the pid in question.
3681 * @policy: new policy.
3682 * @param: structure containing the new RT priority.
3684 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
3685 struct sched_param __user
*param
)
3687 return do_sched_setscheduler(pid
, policy
, param
);
3691 * sys_sched_setparam - set/change the RT priority of a thread
3692 * @pid: the pid in question.
3693 * @param: structure containing the new RT priority.
3695 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
3697 return do_sched_setscheduler(pid
, -1, param
);
3701 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3702 * @pid: the pid in question.
3704 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
3706 int retval
= -EINVAL
;
3713 read_lock(&tasklist_lock
);
3714 p
= find_process_by_pid(pid
);
3716 retval
= security_task_getscheduler(p
);
3720 read_unlock(&tasklist_lock
);
3727 * sys_sched_getscheduler - get the RT priority of a thread
3728 * @pid: the pid in question.
3729 * @param: structure containing the RT priority.
3731 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
3733 struct sched_param lp
;
3734 int retval
= -EINVAL
;
3737 if (!param
|| pid
< 0)
3740 read_lock(&tasklist_lock
);
3741 p
= find_process_by_pid(pid
);
3746 retval
= security_task_getscheduler(p
);
3750 lp
.sched_priority
= p
->rt_priority
;
3751 read_unlock(&tasklist_lock
);
3754 * This one might sleep, we cannot do it with a spinlock held ...
3756 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
3762 read_unlock(&tasklist_lock
);
3766 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
3770 cpumask_t cpus_allowed
;
3773 read_lock(&tasklist_lock
);
3775 p
= find_process_by_pid(pid
);
3777 read_unlock(&tasklist_lock
);
3778 unlock_cpu_hotplug();
3783 * It is not safe to call set_cpus_allowed with the
3784 * tasklist_lock held. We will bump the task_struct's
3785 * usage count and then drop tasklist_lock.
3788 read_unlock(&tasklist_lock
);
3791 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
3792 !capable(CAP_SYS_NICE
))
3795 cpus_allowed
= cpuset_cpus_allowed(p
);
3796 cpus_and(new_mask
, new_mask
, cpus_allowed
);
3797 retval
= set_cpus_allowed(p
, new_mask
);
3801 unlock_cpu_hotplug();
3805 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
3806 cpumask_t
*new_mask
)
3808 if (len
< sizeof(cpumask_t
)) {
3809 memset(new_mask
, 0, sizeof(cpumask_t
));
3810 } else if (len
> sizeof(cpumask_t
)) {
3811 len
= sizeof(cpumask_t
);
3813 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
3817 * sys_sched_setaffinity - set the cpu affinity of a process
3818 * @pid: pid of the process
3819 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3820 * @user_mask_ptr: user-space pointer to the new cpu mask
3822 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
3823 unsigned long __user
*user_mask_ptr
)
3828 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
3832 return sched_setaffinity(pid
, new_mask
);
3836 * Represents all cpu's present in the system
3837 * In systems capable of hotplug, this map could dynamically grow
3838 * as new cpu's are detected in the system via any platform specific
3839 * method, such as ACPI for e.g.
3842 cpumask_t cpu_present_map
;
3843 EXPORT_SYMBOL(cpu_present_map
);
3846 cpumask_t cpu_online_map
= CPU_MASK_ALL
;
3847 cpumask_t cpu_possible_map
= CPU_MASK_ALL
;
3850 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
3856 read_lock(&tasklist_lock
);
3859 p
= find_process_by_pid(pid
);
3864 cpus_and(*mask
, p
->cpus_allowed
, cpu_possible_map
);
3867 read_unlock(&tasklist_lock
);
3868 unlock_cpu_hotplug();
3876 * sys_sched_getaffinity - get the cpu affinity of a process
3877 * @pid: pid of the process
3878 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3879 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3881 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
3882 unsigned long __user
*user_mask_ptr
)
3887 if (len
< sizeof(cpumask_t
))
3890 ret
= sched_getaffinity(pid
, &mask
);
3894 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
3897 return sizeof(cpumask_t
);
3901 * sys_sched_yield - yield the current processor to other threads.
3903 * this function yields the current CPU by moving the calling thread
3904 * to the expired array. If there are no other threads running on this
3905 * CPU then this function will return.
3907 asmlinkage
long sys_sched_yield(void)
3909 runqueue_t
*rq
= this_rq_lock();
3910 prio_array_t
*array
= current
->array
;
3911 prio_array_t
*target
= rq
->expired
;
3913 schedstat_inc(rq
, yld_cnt
);
3915 * We implement yielding by moving the task into the expired
3918 * (special rule: RT tasks will just roundrobin in the active
3921 if (rt_task(current
))
3922 target
= rq
->active
;
3924 if (current
->array
->nr_active
== 1) {
3925 schedstat_inc(rq
, yld_act_empty
);
3926 if (!rq
->expired
->nr_active
)
3927 schedstat_inc(rq
, yld_both_empty
);
3928 } else if (!rq
->expired
->nr_active
)
3929 schedstat_inc(rq
, yld_exp_empty
);
3931 if (array
!= target
) {
3932 dequeue_task(current
, array
);
3933 enqueue_task(current
, target
);
3936 * requeue_task is cheaper so perform that if possible.
3938 requeue_task(current
, array
);
3941 * Since we are going to call schedule() anyway, there's
3942 * no need to preempt or enable interrupts:
3944 __release(rq
->lock
);
3945 _raw_spin_unlock(&rq
->lock
);
3946 preempt_enable_no_resched();
3953 static inline void __cond_resched(void)
3956 * The BKS might be reacquired before we have dropped
3957 * PREEMPT_ACTIVE, which could trigger a second
3958 * cond_resched() call.
3960 if (unlikely(preempt_count()))
3963 add_preempt_count(PREEMPT_ACTIVE
);
3965 sub_preempt_count(PREEMPT_ACTIVE
);
3966 } while (need_resched());
3969 int __sched
cond_resched(void)
3971 if (need_resched()) {
3978 EXPORT_SYMBOL(cond_resched
);
3981 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
3982 * call schedule, and on return reacquire the lock.
3984 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
3985 * operations here to prevent schedule() from being called twice (once via
3986 * spin_unlock(), once by hand).
3988 int cond_resched_lock(spinlock_t
*lock
)
3992 if (need_lockbreak(lock
)) {
3998 if (need_resched()) {
3999 _raw_spin_unlock(lock
);
4000 preempt_enable_no_resched();
4008 EXPORT_SYMBOL(cond_resched_lock
);
4010 int __sched
cond_resched_softirq(void)
4012 BUG_ON(!in_softirq());
4014 if (need_resched()) {
4015 __local_bh_enable();
4023 EXPORT_SYMBOL(cond_resched_softirq
);
4027 * yield - yield the current processor to other threads.
4029 * this is a shortcut for kernel-space yielding - it marks the
4030 * thread runnable and calls sys_sched_yield().
4032 void __sched
yield(void)
4034 set_current_state(TASK_RUNNING
);
4038 EXPORT_SYMBOL(yield
);
4041 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4042 * that process accounting knows that this is a task in IO wait state.
4044 * But don't do that if it is a deliberate, throttling IO wait (this task
4045 * has set its backing_dev_info: the queue against which it should throttle)
4047 void __sched
io_schedule(void)
4049 struct runqueue
*rq
= &per_cpu(runqueues
, raw_smp_processor_id());
4051 atomic_inc(&rq
->nr_iowait
);
4053 atomic_dec(&rq
->nr_iowait
);
4056 EXPORT_SYMBOL(io_schedule
);
4058 long __sched
io_schedule_timeout(long timeout
)
4060 struct runqueue
*rq
= &per_cpu(runqueues
, raw_smp_processor_id());
4063 atomic_inc(&rq
->nr_iowait
);
4064 ret
= schedule_timeout(timeout
);
4065 atomic_dec(&rq
->nr_iowait
);
4070 * sys_sched_get_priority_max - return maximum RT priority.
4071 * @policy: scheduling class.
4073 * this syscall returns the maximum rt_priority that can be used
4074 * by a given scheduling class.
4076 asmlinkage
long sys_sched_get_priority_max(int policy
)
4083 ret
= MAX_USER_RT_PRIO
-1;
4093 * sys_sched_get_priority_min - return minimum RT priority.
4094 * @policy: scheduling class.
4096 * this syscall returns the minimum rt_priority that can be used
4097 * by a given scheduling class.
4099 asmlinkage
long sys_sched_get_priority_min(int policy
)
4115 * sys_sched_rr_get_interval - return the default timeslice of a process.
4116 * @pid: pid of the process.
4117 * @interval: userspace pointer to the timeslice value.
4119 * this syscall writes the default timeslice value of a given process
4120 * into the user-space timespec buffer. A value of '0' means infinity.
4123 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4125 int retval
= -EINVAL
;
4133 read_lock(&tasklist_lock
);
4134 p
= find_process_by_pid(pid
);
4138 retval
= security_task_getscheduler(p
);
4142 jiffies_to_timespec(p
->policy
& SCHED_FIFO
?
4143 0 : task_timeslice(p
), &t
);
4144 read_unlock(&tasklist_lock
);
4145 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4149 read_unlock(&tasklist_lock
);
4153 static inline struct task_struct
*eldest_child(struct task_struct
*p
)
4155 if (list_empty(&p
->children
)) return NULL
;
4156 return list_entry(p
->children
.next
,struct task_struct
,sibling
);
4159 static inline struct task_struct
*older_sibling(struct task_struct
*p
)
4161 if (p
->sibling
.prev
==&p
->parent
->children
) return NULL
;
4162 return list_entry(p
->sibling
.prev
,struct task_struct
,sibling
);
4165 static inline struct task_struct
*younger_sibling(struct task_struct
*p
)
4167 if (p
->sibling
.next
==&p
->parent
->children
) return NULL
;
4168 return list_entry(p
->sibling
.next
,struct task_struct
,sibling
);
4171 static void show_task(task_t
*p
)
4175 unsigned long free
= 0;
4176 static const char *stat_nam
[] = { "R", "S", "D", "T", "t", "Z", "X" };
4178 printk("%-13.13s ", p
->comm
);
4179 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4180 if (state
< ARRAY_SIZE(stat_nam
))
4181 printk(stat_nam
[state
]);
4184 #if (BITS_PER_LONG == 32)
4185 if (state
== TASK_RUNNING
)
4186 printk(" running ");
4188 printk(" %08lX ", thread_saved_pc(p
));
4190 if (state
== TASK_RUNNING
)
4191 printk(" running task ");
4193 printk(" %016lx ", thread_saved_pc(p
));
4195 #ifdef CONFIG_DEBUG_STACK_USAGE
4197 unsigned long *n
= (unsigned long *) (p
->thread_info
+1);
4200 free
= (unsigned long) n
- (unsigned long)(p
->thread_info
+1);
4203 printk("%5lu %5d %6d ", free
, p
->pid
, p
->parent
->pid
);
4204 if ((relative
= eldest_child(p
)))
4205 printk("%5d ", relative
->pid
);
4208 if ((relative
= younger_sibling(p
)))
4209 printk("%7d", relative
->pid
);
4212 if ((relative
= older_sibling(p
)))
4213 printk(" %5d", relative
->pid
);
4217 printk(" (L-TLB)\n");
4219 printk(" (NOTLB)\n");
4221 if (state
!= TASK_RUNNING
)
4222 show_stack(p
, NULL
);
4225 void show_state(void)
4229 #if (BITS_PER_LONG == 32)
4232 printk(" task PC pid father child younger older\n");
4236 printk(" task PC pid father child younger older\n");
4238 read_lock(&tasklist_lock
);
4239 do_each_thread(g
, p
) {
4241 * reset the NMI-timeout, listing all files on a slow
4242 * console might take alot of time:
4244 touch_nmi_watchdog();
4246 } while_each_thread(g
, p
);
4248 read_unlock(&tasklist_lock
);
4252 * init_idle - set up an idle thread for a given CPU
4253 * @idle: task in question
4254 * @cpu: cpu the idle task belongs to
4256 * NOTE: this function does not set the idle thread's NEED_RESCHED
4257 * flag, to make booting more robust.
4259 void __devinit
init_idle(task_t
*idle
, int cpu
)
4261 runqueue_t
*rq
= cpu_rq(cpu
);
4262 unsigned long flags
;
4264 idle
->sleep_avg
= 0;
4266 idle
->prio
= MAX_PRIO
;
4267 idle
->state
= TASK_RUNNING
;
4268 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4269 set_task_cpu(idle
, cpu
);
4271 spin_lock_irqsave(&rq
->lock
, flags
);
4272 rq
->curr
= rq
->idle
= idle
;
4273 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4276 spin_unlock_irqrestore(&rq
->lock
, flags
);
4278 /* Set the preempt count _outside_ the spinlocks! */
4279 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4280 idle
->thread_info
->preempt_count
= (idle
->lock_depth
>= 0);
4282 idle
->thread_info
->preempt_count
= 0;
4287 * In a system that switches off the HZ timer nohz_cpu_mask
4288 * indicates which cpus entered this state. This is used
4289 * in the rcu update to wait only for active cpus. For system
4290 * which do not switch off the HZ timer nohz_cpu_mask should
4291 * always be CPU_MASK_NONE.
4293 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4297 * This is how migration works:
4299 * 1) we queue a migration_req_t structure in the source CPU's
4300 * runqueue and wake up that CPU's migration thread.
4301 * 2) we down() the locked semaphore => thread blocks.
4302 * 3) migration thread wakes up (implicitly it forces the migrated
4303 * thread off the CPU)
4304 * 4) it gets the migration request and checks whether the migrated
4305 * task is still in the wrong runqueue.
4306 * 5) if it's in the wrong runqueue then the migration thread removes
4307 * it and puts it into the right queue.
4308 * 6) migration thread up()s the semaphore.
4309 * 7) we wake up and the migration is done.
4313 * Change a given task's CPU affinity. Migrate the thread to a
4314 * proper CPU and schedule it away if the CPU it's executing on
4315 * is removed from the allowed bitmask.
4317 * NOTE: the caller must have a valid reference to the task, the
4318 * task must not exit() & deallocate itself prematurely. The
4319 * call is not atomic; no spinlocks may be held.
4321 int set_cpus_allowed(task_t
*p
, cpumask_t new_mask
)
4323 unsigned long flags
;
4325 migration_req_t req
;
4328 rq
= task_rq_lock(p
, &flags
);
4329 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4334 p
->cpus_allowed
= new_mask
;
4335 /* Can the task run on the task's current CPU? If so, we're done */
4336 if (cpu_isset(task_cpu(p
), new_mask
))
4339 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4340 /* Need help from migration thread: drop lock and wait. */
4341 task_rq_unlock(rq
, &flags
);
4342 wake_up_process(rq
->migration_thread
);
4343 wait_for_completion(&req
.done
);
4344 tlb_migrate_finish(p
->mm
);
4348 task_rq_unlock(rq
, &flags
);
4352 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4355 * Move (not current) task off this cpu, onto dest cpu. We're doing
4356 * this because either it can't run here any more (set_cpus_allowed()
4357 * away from this CPU, or CPU going down), or because we're
4358 * attempting to rebalance this task on exec (sched_exec).
4360 * So we race with normal scheduler movements, but that's OK, as long
4361 * as the task is no longer on this CPU.
4363 static void __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4365 runqueue_t
*rq_dest
, *rq_src
;
4367 if (unlikely(cpu_is_offline(dest_cpu
)))
4370 rq_src
= cpu_rq(src_cpu
);
4371 rq_dest
= cpu_rq(dest_cpu
);
4373 double_rq_lock(rq_src
, rq_dest
);
4374 /* Already moved. */
4375 if (task_cpu(p
) != src_cpu
)
4377 /* Affinity changed (again). */
4378 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
4381 set_task_cpu(p
, dest_cpu
);
4384 * Sync timestamp with rq_dest's before activating.
4385 * The same thing could be achieved by doing this step
4386 * afterwards, and pretending it was a local activate.
4387 * This way is cleaner and logically correct.
4389 p
->timestamp
= p
->timestamp
- rq_src
->timestamp_last_tick
4390 + rq_dest
->timestamp_last_tick
;
4391 deactivate_task(p
, rq_src
);
4392 activate_task(p
, rq_dest
, 0);
4393 if (TASK_PREEMPTS_CURR(p
, rq_dest
))
4394 resched_task(rq_dest
->curr
);
4398 double_rq_unlock(rq_src
, rq_dest
);
4402 * migration_thread - this is a highprio system thread that performs
4403 * thread migration by bumping thread off CPU then 'pushing' onto
4406 static int migration_thread(void *data
)
4409 int cpu
= (long)data
;
4412 BUG_ON(rq
->migration_thread
!= current
);
4414 set_current_state(TASK_INTERRUPTIBLE
);
4415 while (!kthread_should_stop()) {
4416 struct list_head
*head
;
4417 migration_req_t
*req
;
4421 spin_lock_irq(&rq
->lock
);
4423 if (cpu_is_offline(cpu
)) {
4424 spin_unlock_irq(&rq
->lock
);
4428 if (rq
->active_balance
) {
4429 active_load_balance(rq
, cpu
);
4430 rq
->active_balance
= 0;
4433 head
= &rq
->migration_queue
;
4435 if (list_empty(head
)) {
4436 spin_unlock_irq(&rq
->lock
);
4438 set_current_state(TASK_INTERRUPTIBLE
);
4441 req
= list_entry(head
->next
, migration_req_t
, list
);
4442 list_del_init(head
->next
);
4444 spin_unlock(&rq
->lock
);
4445 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
4448 complete(&req
->done
);
4450 __set_current_state(TASK_RUNNING
);
4454 /* Wait for kthread_stop */
4455 set_current_state(TASK_INTERRUPTIBLE
);
4456 while (!kthread_should_stop()) {
4458 set_current_state(TASK_INTERRUPTIBLE
);
4460 __set_current_state(TASK_RUNNING
);
4464 #ifdef CONFIG_HOTPLUG_CPU
4465 /* Figure out where task on dead CPU should go, use force if neccessary. */
4466 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*tsk
)
4472 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
4473 cpus_and(mask
, mask
, tsk
->cpus_allowed
);
4474 dest_cpu
= any_online_cpu(mask
);
4476 /* On any allowed CPU? */
4477 if (dest_cpu
== NR_CPUS
)
4478 dest_cpu
= any_online_cpu(tsk
->cpus_allowed
);
4480 /* No more Mr. Nice Guy. */
4481 if (dest_cpu
== NR_CPUS
) {
4482 cpus_setall(tsk
->cpus_allowed
);
4483 dest_cpu
= any_online_cpu(tsk
->cpus_allowed
);
4486 * Don't tell them about moving exiting tasks or
4487 * kernel threads (both mm NULL), since they never
4490 if (tsk
->mm
&& printk_ratelimit())
4491 printk(KERN_INFO
"process %d (%s) no "
4492 "longer affine to cpu%d\n",
4493 tsk
->pid
, tsk
->comm
, dead_cpu
);
4495 __migrate_task(tsk
, dead_cpu
, dest_cpu
);
4499 * While a dead CPU has no uninterruptible tasks queued at this point,
4500 * it might still have a nonzero ->nr_uninterruptible counter, because
4501 * for performance reasons the counter is not stricly tracking tasks to
4502 * their home CPUs. So we just add the counter to another CPU's counter,
4503 * to keep the global sum constant after CPU-down:
4505 static void migrate_nr_uninterruptible(runqueue_t
*rq_src
)
4507 runqueue_t
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
4508 unsigned long flags
;
4510 local_irq_save(flags
);
4511 double_rq_lock(rq_src
, rq_dest
);
4512 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
4513 rq_src
->nr_uninterruptible
= 0;
4514 double_rq_unlock(rq_src
, rq_dest
);
4515 local_irq_restore(flags
);
4518 /* Run through task list and migrate tasks from the dead cpu. */
4519 static void migrate_live_tasks(int src_cpu
)
4521 struct task_struct
*tsk
, *t
;
4523 write_lock_irq(&tasklist_lock
);
4525 do_each_thread(t
, tsk
) {
4529 if (task_cpu(tsk
) == src_cpu
)
4530 move_task_off_dead_cpu(src_cpu
, tsk
);
4531 } while_each_thread(t
, tsk
);
4533 write_unlock_irq(&tasklist_lock
);
4536 /* Schedules idle task to be the next runnable task on current CPU.
4537 * It does so by boosting its priority to highest possible and adding it to
4538 * the _front_ of runqueue. Used by CPU offline code.
4540 void sched_idle_next(void)
4542 int cpu
= smp_processor_id();
4543 runqueue_t
*rq
= this_rq();
4544 struct task_struct
*p
= rq
->idle
;
4545 unsigned long flags
;
4547 /* cpu has to be offline */
4548 BUG_ON(cpu_online(cpu
));
4550 /* Strictly not necessary since rest of the CPUs are stopped by now
4551 * and interrupts disabled on current cpu.
4553 spin_lock_irqsave(&rq
->lock
, flags
);
4555 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
4556 /* Add idle task to _front_ of it's priority queue */
4557 __activate_idle_task(p
, rq
);
4559 spin_unlock_irqrestore(&rq
->lock
, flags
);
4562 /* Ensures that the idle task is using init_mm right before its cpu goes
4565 void idle_task_exit(void)
4567 struct mm_struct
*mm
= current
->active_mm
;
4569 BUG_ON(cpu_online(smp_processor_id()));
4572 switch_mm(mm
, &init_mm
, current
);
4576 static void migrate_dead(unsigned int dead_cpu
, task_t
*tsk
)
4578 struct runqueue
*rq
= cpu_rq(dead_cpu
);
4580 /* Must be exiting, otherwise would be on tasklist. */
4581 BUG_ON(tsk
->exit_state
!= EXIT_ZOMBIE
&& tsk
->exit_state
!= EXIT_DEAD
);
4583 /* Cannot have done final schedule yet: would have vanished. */
4584 BUG_ON(tsk
->flags
& PF_DEAD
);
4586 get_task_struct(tsk
);
4589 * Drop lock around migration; if someone else moves it,
4590 * that's OK. No task can be added to this CPU, so iteration is
4593 spin_unlock_irq(&rq
->lock
);
4594 move_task_off_dead_cpu(dead_cpu
, tsk
);
4595 spin_lock_irq(&rq
->lock
);
4597 put_task_struct(tsk
);
4600 /* release_task() removes task from tasklist, so we won't find dead tasks. */
4601 static void migrate_dead_tasks(unsigned int dead_cpu
)
4604 struct runqueue
*rq
= cpu_rq(dead_cpu
);
4606 for (arr
= 0; arr
< 2; arr
++) {
4607 for (i
= 0; i
< MAX_PRIO
; i
++) {
4608 struct list_head
*list
= &rq
->arrays
[arr
].queue
[i
];
4609 while (!list_empty(list
))
4610 migrate_dead(dead_cpu
,
4611 list_entry(list
->next
, task_t
,
4616 #endif /* CONFIG_HOTPLUG_CPU */
4619 * migration_call - callback that gets triggered when a CPU is added.
4620 * Here we can start up the necessary migration thread for the new CPU.
4622 static int migration_call(struct notifier_block
*nfb
, unsigned long action
,
4625 int cpu
= (long)hcpu
;
4626 struct task_struct
*p
;
4627 struct runqueue
*rq
;
4628 unsigned long flags
;
4631 case CPU_UP_PREPARE
:
4632 p
= kthread_create(migration_thread
, hcpu
, "migration/%d",cpu
);
4635 p
->flags
|= PF_NOFREEZE
;
4636 kthread_bind(p
, cpu
);
4637 /* Must be high prio: stop_machine expects to yield to it. */
4638 rq
= task_rq_lock(p
, &flags
);
4639 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
4640 task_rq_unlock(rq
, &flags
);
4641 cpu_rq(cpu
)->migration_thread
= p
;
4644 /* Strictly unneccessary, as first user will wake it. */
4645 wake_up_process(cpu_rq(cpu
)->migration_thread
);
4647 #ifdef CONFIG_HOTPLUG_CPU
4648 case CPU_UP_CANCELED
:
4649 /* Unbind it from offline cpu so it can run. Fall thru. */
4650 kthread_bind(cpu_rq(cpu
)->migration_thread
,smp_processor_id());
4651 kthread_stop(cpu_rq(cpu
)->migration_thread
);
4652 cpu_rq(cpu
)->migration_thread
= NULL
;
4655 migrate_live_tasks(cpu
);
4657 kthread_stop(rq
->migration_thread
);
4658 rq
->migration_thread
= NULL
;
4659 /* Idle task back to normal (off runqueue, low prio) */
4660 rq
= task_rq_lock(rq
->idle
, &flags
);
4661 deactivate_task(rq
->idle
, rq
);
4662 rq
->idle
->static_prio
= MAX_PRIO
;
4663 __setscheduler(rq
->idle
, SCHED_NORMAL
, 0);
4664 migrate_dead_tasks(cpu
);
4665 task_rq_unlock(rq
, &flags
);
4666 migrate_nr_uninterruptible(rq
);
4667 BUG_ON(rq
->nr_running
!= 0);
4669 /* No need to migrate the tasks: it was best-effort if
4670 * they didn't do lock_cpu_hotplug(). Just wake up
4671 * the requestors. */
4672 spin_lock_irq(&rq
->lock
);
4673 while (!list_empty(&rq
->migration_queue
)) {
4674 migration_req_t
*req
;
4675 req
= list_entry(rq
->migration_queue
.next
,
4676 migration_req_t
, list
);
4677 list_del_init(&req
->list
);
4678 complete(&req
->done
);
4680 spin_unlock_irq(&rq
->lock
);
4687 /* Register at highest priority so that task migration (migrate_all_tasks)
4688 * happens before everything else.
4690 static struct notifier_block __devinitdata migration_notifier
= {
4691 .notifier_call
= migration_call
,
4695 int __init
migration_init(void)
4697 void *cpu
= (void *)(long)smp_processor_id();
4698 /* Start one for boot CPU. */
4699 migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
4700 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
4701 register_cpu_notifier(&migration_notifier
);
4707 #undef SCHED_DOMAIN_DEBUG
4708 #ifdef SCHED_DOMAIN_DEBUG
4709 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
4714 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
4718 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
4723 struct sched_group
*group
= sd
->groups
;
4724 cpumask_t groupmask
;
4726 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
4727 cpus_clear(groupmask
);
4730 for (i
= 0; i
< level
+ 1; i
++)
4732 printk("domain %d: ", level
);
4734 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
4735 printk("does not load-balance\n");
4737 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain has parent");
4741 printk("span %s\n", str
);
4743 if (!cpu_isset(cpu
, sd
->span
))
4744 printk(KERN_ERR
"ERROR: domain->span does not contain CPU%d\n", cpu
);
4745 if (!cpu_isset(cpu
, group
->cpumask
))
4746 printk(KERN_ERR
"ERROR: domain->groups does not contain CPU%d\n", cpu
);
4749 for (i
= 0; i
< level
+ 2; i
++)
4755 printk(KERN_ERR
"ERROR: group is NULL\n");
4759 if (!group
->cpu_power
) {
4761 printk(KERN_ERR
"ERROR: domain->cpu_power not set\n");
4764 if (!cpus_weight(group
->cpumask
)) {
4766 printk(KERN_ERR
"ERROR: empty group\n");
4769 if (cpus_intersects(groupmask
, group
->cpumask
)) {
4771 printk(KERN_ERR
"ERROR: repeated CPUs\n");
4774 cpus_or(groupmask
, groupmask
, group
->cpumask
);
4776 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
4779 group
= group
->next
;
4780 } while (group
!= sd
->groups
);
4783 if (!cpus_equal(sd
->span
, groupmask
))
4784 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
4790 if (!cpus_subset(groupmask
, sd
->span
))
4791 printk(KERN_ERR
"ERROR: parent span is not a superset of domain->span\n");
4797 #define sched_domain_debug(sd, cpu) {}
4800 static int sd_degenerate(struct sched_domain
*sd
)
4802 if (cpus_weight(sd
->span
) == 1)
4805 /* Following flags need at least 2 groups */
4806 if (sd
->flags
& (SD_LOAD_BALANCE
|
4807 SD_BALANCE_NEWIDLE
|
4810 if (sd
->groups
!= sd
->groups
->next
)
4814 /* Following flags don't use groups */
4815 if (sd
->flags
& (SD_WAKE_IDLE
|
4823 static int sd_parent_degenerate(struct sched_domain
*sd
,
4824 struct sched_domain
*parent
)
4826 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
4828 if (sd_degenerate(parent
))
4831 if (!cpus_equal(sd
->span
, parent
->span
))
4834 /* Does parent contain flags not in child? */
4835 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
4836 if (cflags
& SD_WAKE_AFFINE
)
4837 pflags
&= ~SD_WAKE_BALANCE
;
4838 /* Flags needing groups don't count if only 1 group in parent */
4839 if (parent
->groups
== parent
->groups
->next
) {
4840 pflags
&= ~(SD_LOAD_BALANCE
|
4841 SD_BALANCE_NEWIDLE
|
4845 if (~cflags
& pflags
)
4852 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
4853 * hold the hotplug lock.
4855 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
4857 runqueue_t
*rq
= cpu_rq(cpu
);
4858 struct sched_domain
*tmp
;
4860 /* Remove the sched domains which do not contribute to scheduling. */
4861 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
4862 struct sched_domain
*parent
= tmp
->parent
;
4865 if (sd_parent_degenerate(tmp
, parent
))
4866 tmp
->parent
= parent
->parent
;
4869 if (sd
&& sd_degenerate(sd
))
4872 sched_domain_debug(sd
, cpu
);
4874 rcu_assign_pointer(rq
->sd
, sd
);
4877 /* cpus with isolated domains */
4878 static cpumask_t __devinitdata cpu_isolated_map
= CPU_MASK_NONE
;
4880 /* Setup the mask of cpus configured for isolated domains */
4881 static int __init
isolated_cpu_setup(char *str
)
4883 int ints
[NR_CPUS
], i
;
4885 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
4886 cpus_clear(cpu_isolated_map
);
4887 for (i
= 1; i
<= ints
[0]; i
++)
4888 if (ints
[i
] < NR_CPUS
)
4889 cpu_set(ints
[i
], cpu_isolated_map
);
4893 __setup ("isolcpus=", isolated_cpu_setup
);
4896 * init_sched_build_groups takes an array of groups, the cpumask we wish
4897 * to span, and a pointer to a function which identifies what group a CPU
4898 * belongs to. The return value of group_fn must be a valid index into the
4899 * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
4900 * keep track of groups covered with a cpumask_t).
4902 * init_sched_build_groups will build a circular linked list of the groups
4903 * covered by the given span, and will set each group's ->cpumask correctly,
4904 * and ->cpu_power to 0.
4906 static void init_sched_build_groups(struct sched_group groups
[], cpumask_t span
,
4907 int (*group_fn
)(int cpu
))
4909 struct sched_group
*first
= NULL
, *last
= NULL
;
4910 cpumask_t covered
= CPU_MASK_NONE
;
4913 for_each_cpu_mask(i
, span
) {
4914 int group
= group_fn(i
);
4915 struct sched_group
*sg
= &groups
[group
];
4918 if (cpu_isset(i
, covered
))
4921 sg
->cpumask
= CPU_MASK_NONE
;
4924 for_each_cpu_mask(j
, span
) {
4925 if (group_fn(j
) != group
)
4928 cpu_set(j
, covered
);
4929 cpu_set(j
, sg
->cpumask
);
4940 #define SD_NODES_PER_DOMAIN 16
4944 * find_next_best_node - find the next node to include in a sched_domain
4945 * @node: node whose sched_domain we're building
4946 * @used_nodes: nodes already in the sched_domain
4948 * Find the next node to include in a given scheduling domain. Simply
4949 * finds the closest node not already in the @used_nodes map.
4951 * Should use nodemask_t.
4953 static int find_next_best_node(int node
, unsigned long *used_nodes
)
4955 int i
, n
, val
, min_val
, best_node
= 0;
4959 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
4960 /* Start at @node */
4961 n
= (node
+ i
) % MAX_NUMNODES
;
4963 if (!nr_cpus_node(n
))
4966 /* Skip already used nodes */
4967 if (test_bit(n
, used_nodes
))
4970 /* Simple min distance search */
4971 val
= node_distance(node
, n
);
4973 if (val
< min_val
) {
4979 set_bit(best_node
, used_nodes
);
4984 * sched_domain_node_span - get a cpumask for a node's sched_domain
4985 * @node: node whose cpumask we're constructing
4986 * @size: number of nodes to include in this span
4988 * Given a node, construct a good cpumask for its sched_domain to span. It
4989 * should be one that prevents unnecessary balancing, but also spreads tasks
4992 static cpumask_t
sched_domain_node_span(int node
)
4995 cpumask_t span
, nodemask
;
4996 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
4999 bitmap_zero(used_nodes
, MAX_NUMNODES
);
5001 nodemask
= node_to_cpumask(node
);
5002 cpus_or(span
, span
, nodemask
);
5003 set_bit(node
, used_nodes
);
5005 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5006 int next_node
= find_next_best_node(node
, used_nodes
);
5007 nodemask
= node_to_cpumask(next_node
);
5008 cpus_or(span
, span
, nodemask
);
5016 * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we
5017 * can switch it on easily if needed.
5019 #ifdef CONFIG_SCHED_SMT
5020 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
5021 static struct sched_group sched_group_cpus
[NR_CPUS
];
5022 static int cpu_to_cpu_group(int cpu
)
5028 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
5029 static struct sched_group sched_group_phys
[NR_CPUS
];
5030 static int cpu_to_phys_group(int cpu
)
5032 #ifdef CONFIG_SCHED_SMT
5033 return first_cpu(cpu_sibling_map
[cpu
]);
5041 * The init_sched_build_groups can't handle what we want to do with node
5042 * groups, so roll our own. Now each node has its own list of groups which
5043 * gets dynamically allocated.
5045 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
5046 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
5048 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
5049 static struct sched_group
*sched_group_allnodes_bycpu
[NR_CPUS
];
5051 static int cpu_to_allnodes_group(int cpu
)
5053 return cpu_to_node(cpu
);
5058 * Build sched domains for a given set of cpus and attach the sched domains
5059 * to the individual cpus
5061 void build_sched_domains(const cpumask_t
*cpu_map
)
5065 struct sched_group
**sched_group_nodes
= NULL
;
5066 struct sched_group
*sched_group_allnodes
= NULL
;
5069 * Allocate the per-node list of sched groups
5071 sched_group_nodes
= kmalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
5073 if (!sched_group_nodes
) {
5074 printk(KERN_WARNING
"Can not alloc sched group node list\n");
5077 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
5081 * Set up domains for cpus specified by the cpu_map.
5083 for_each_cpu_mask(i
, *cpu_map
) {
5085 struct sched_domain
*sd
= NULL
, *p
;
5086 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
5088 cpus_and(nodemask
, nodemask
, *cpu_map
);
5091 if (cpus_weight(*cpu_map
)
5092 > SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
5093 if (!sched_group_allnodes
) {
5094 sched_group_allnodes
5095 = kmalloc(sizeof(struct sched_group
)
5098 if (!sched_group_allnodes
) {
5100 "Can not alloc allnodes sched group\n");
5103 sched_group_allnodes_bycpu
[i
]
5104 = sched_group_allnodes
;
5106 sd
= &per_cpu(allnodes_domains
, i
);
5107 *sd
= SD_ALLNODES_INIT
;
5108 sd
->span
= *cpu_map
;
5109 group
= cpu_to_allnodes_group(i
);
5110 sd
->groups
= &sched_group_allnodes
[group
];
5115 sd
= &per_cpu(node_domains
, i
);
5117 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
5119 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
5123 sd
= &per_cpu(phys_domains
, i
);
5124 group
= cpu_to_phys_group(i
);
5126 sd
->span
= nodemask
;
5128 sd
->groups
= &sched_group_phys
[group
];
5130 #ifdef CONFIG_SCHED_SMT
5132 sd
= &per_cpu(cpu_domains
, i
);
5133 group
= cpu_to_cpu_group(i
);
5134 *sd
= SD_SIBLING_INIT
;
5135 sd
->span
= cpu_sibling_map
[i
];
5136 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
5138 sd
->groups
= &sched_group_cpus
[group
];
5142 #ifdef CONFIG_SCHED_SMT
5143 /* Set up CPU (sibling) groups */
5144 for_each_cpu_mask(i
, *cpu_map
) {
5145 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
5146 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
5147 if (i
!= first_cpu(this_sibling_map
))
5150 init_sched_build_groups(sched_group_cpus
, this_sibling_map
,
5155 /* Set up physical groups */
5156 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5157 cpumask_t nodemask
= node_to_cpumask(i
);
5159 cpus_and(nodemask
, nodemask
, *cpu_map
);
5160 if (cpus_empty(nodemask
))
5163 init_sched_build_groups(sched_group_phys
, nodemask
,
5164 &cpu_to_phys_group
);
5168 /* Set up node groups */
5169 if (sched_group_allnodes
)
5170 init_sched_build_groups(sched_group_allnodes
, *cpu_map
,
5171 &cpu_to_allnodes_group
);
5173 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5174 /* Set up node groups */
5175 struct sched_group
*sg
, *prev
;
5176 cpumask_t nodemask
= node_to_cpumask(i
);
5177 cpumask_t domainspan
;
5178 cpumask_t covered
= CPU_MASK_NONE
;
5181 cpus_and(nodemask
, nodemask
, *cpu_map
);
5182 if (cpus_empty(nodemask
)) {
5183 sched_group_nodes
[i
] = NULL
;
5187 domainspan
= sched_domain_node_span(i
);
5188 cpus_and(domainspan
, domainspan
, *cpu_map
);
5190 sg
= kmalloc(sizeof(struct sched_group
), GFP_KERNEL
);
5191 sched_group_nodes
[i
] = sg
;
5192 for_each_cpu_mask(j
, nodemask
) {
5193 struct sched_domain
*sd
;
5194 sd
= &per_cpu(node_domains
, j
);
5196 if (sd
->groups
== NULL
) {
5197 /* Turn off balancing if we have no groups */
5203 "Can not alloc domain group for node %d\n", i
);
5207 sg
->cpumask
= nodemask
;
5208 cpus_or(covered
, covered
, nodemask
);
5211 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
5212 cpumask_t tmp
, notcovered
;
5213 int n
= (i
+ j
) % MAX_NUMNODES
;
5215 cpus_complement(notcovered
, covered
);
5216 cpus_and(tmp
, notcovered
, *cpu_map
);
5217 cpus_and(tmp
, tmp
, domainspan
);
5218 if (cpus_empty(tmp
))
5221 nodemask
= node_to_cpumask(n
);
5222 cpus_and(tmp
, tmp
, nodemask
);
5223 if (cpus_empty(tmp
))
5226 sg
= kmalloc(sizeof(struct sched_group
), GFP_KERNEL
);
5229 "Can not alloc domain group for node %d\n", j
);
5234 cpus_or(covered
, covered
, tmp
);
5238 prev
->next
= sched_group_nodes
[i
];
5242 /* Calculate CPU power for physical packages and nodes */
5243 for_each_cpu_mask(i
, *cpu_map
) {
5245 struct sched_domain
*sd
;
5246 #ifdef CONFIG_SCHED_SMT
5247 sd
= &per_cpu(cpu_domains
, i
);
5248 power
= SCHED_LOAD_SCALE
;
5249 sd
->groups
->cpu_power
= power
;
5252 sd
= &per_cpu(phys_domains
, i
);
5253 power
= SCHED_LOAD_SCALE
+ SCHED_LOAD_SCALE
*
5254 (cpus_weight(sd
->groups
->cpumask
)-1) / 10;
5255 sd
->groups
->cpu_power
= power
;
5258 sd
= &per_cpu(allnodes_domains
, i
);
5260 power
= SCHED_LOAD_SCALE
+ SCHED_LOAD_SCALE
*
5261 (cpus_weight(sd
->groups
->cpumask
)-1) / 10;
5262 sd
->groups
->cpu_power
= power
;
5268 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5269 struct sched_group
*sg
= sched_group_nodes
[i
];
5275 for_each_cpu_mask(j
, sg
->cpumask
) {
5276 struct sched_domain
*sd
;
5279 sd
= &per_cpu(phys_domains
, j
);
5280 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
5282 * Only add "power" once for each
5287 power
= SCHED_LOAD_SCALE
+ SCHED_LOAD_SCALE
*
5288 (cpus_weight(sd
->groups
->cpumask
)-1) / 10;
5290 sg
->cpu_power
+= power
;
5293 if (sg
!= sched_group_nodes
[i
])
5298 /* Attach the domains */
5299 for_each_cpu_mask(i
, *cpu_map
) {
5300 struct sched_domain
*sd
;
5301 #ifdef CONFIG_SCHED_SMT
5302 sd
= &per_cpu(cpu_domains
, i
);
5304 sd
= &per_cpu(phys_domains
, i
);
5306 cpu_attach_domain(sd
, i
);
5310 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
5312 static void arch_init_sched_domains(const cpumask_t
*cpu_map
)
5314 cpumask_t cpu_default_map
;
5317 * Setup mask for cpus without special case scheduling requirements.
5318 * For now this just excludes isolated cpus, but could be used to
5319 * exclude other special cases in the future.
5321 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
5323 build_sched_domains(&cpu_default_map
);
5326 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
5332 for_each_cpu_mask(cpu
, *cpu_map
) {
5333 struct sched_group
*sched_group_allnodes
5334 = sched_group_allnodes_bycpu
[cpu
];
5335 struct sched_group
**sched_group_nodes
5336 = sched_group_nodes_bycpu
[cpu
];
5338 if (sched_group_allnodes
) {
5339 kfree(sched_group_allnodes
);
5340 sched_group_allnodes_bycpu
[cpu
] = NULL
;
5343 if (!sched_group_nodes
)
5346 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5347 cpumask_t nodemask
= node_to_cpumask(i
);
5348 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
5350 cpus_and(nodemask
, nodemask
, *cpu_map
);
5351 if (cpus_empty(nodemask
))
5361 if (oldsg
!= sched_group_nodes
[i
])
5364 kfree(sched_group_nodes
);
5365 sched_group_nodes_bycpu
[cpu
] = NULL
;
5371 * Detach sched domains from a group of cpus specified in cpu_map
5372 * These cpus will now be attached to the NULL domain
5374 static inline void detach_destroy_domains(const cpumask_t
*cpu_map
)
5378 for_each_cpu_mask(i
, *cpu_map
)
5379 cpu_attach_domain(NULL
, i
);
5380 synchronize_sched();
5381 arch_destroy_sched_domains(cpu_map
);
5385 * Partition sched domains as specified by the cpumasks below.
5386 * This attaches all cpus from the cpumasks to the NULL domain,
5387 * waits for a RCU quiescent period, recalculates sched
5388 * domain information and then attaches them back to the
5389 * correct sched domains
5390 * Call with hotplug lock held
5392 void partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
5394 cpumask_t change_map
;
5396 cpus_and(*partition1
, *partition1
, cpu_online_map
);
5397 cpus_and(*partition2
, *partition2
, cpu_online_map
);
5398 cpus_or(change_map
, *partition1
, *partition2
);
5400 /* Detach sched domains from all of the affected cpus */
5401 detach_destroy_domains(&change_map
);
5402 if (!cpus_empty(*partition1
))
5403 build_sched_domains(partition1
);
5404 if (!cpus_empty(*partition2
))
5405 build_sched_domains(partition2
);
5408 #ifdef CONFIG_HOTPLUG_CPU
5410 * Force a reinitialization of the sched domains hierarchy. The domains
5411 * and groups cannot be updated in place without racing with the balancing
5412 * code, so we temporarily attach all running cpus to the NULL domain
5413 * which will prevent rebalancing while the sched domains are recalculated.
5415 static int update_sched_domains(struct notifier_block
*nfb
,
5416 unsigned long action
, void *hcpu
)
5419 case CPU_UP_PREPARE
:
5420 case CPU_DOWN_PREPARE
:
5421 detach_destroy_domains(&cpu_online_map
);
5424 case CPU_UP_CANCELED
:
5425 case CPU_DOWN_FAILED
:
5429 * Fall through and re-initialise the domains.
5436 /* The hotplug lock is already held by cpu_up/cpu_down */
5437 arch_init_sched_domains(&cpu_online_map
);
5443 void __init
sched_init_smp(void)
5446 arch_init_sched_domains(&cpu_online_map
);
5447 unlock_cpu_hotplug();
5448 /* XXX: Theoretical race here - CPU may be hotplugged now */
5449 hotcpu_notifier(update_sched_domains
, 0);
5452 void __init
sched_init_smp(void)
5455 #endif /* CONFIG_SMP */
5457 int in_sched_functions(unsigned long addr
)
5459 /* Linker adds these: start and end of __sched functions */
5460 extern char __sched_text_start
[], __sched_text_end
[];
5461 return in_lock_functions(addr
) ||
5462 (addr
>= (unsigned long)__sched_text_start
5463 && addr
< (unsigned long)__sched_text_end
);
5466 void __init
sched_init(void)
5471 for (i
= 0; i
< NR_CPUS
; i
++) {
5472 prio_array_t
*array
;
5475 spin_lock_init(&rq
->lock
);
5477 rq
->active
= rq
->arrays
;
5478 rq
->expired
= rq
->arrays
+ 1;
5479 rq
->best_expired_prio
= MAX_PRIO
;
5483 for (j
= 1; j
< 3; j
++)
5484 rq
->cpu_load
[j
] = 0;
5485 rq
->active_balance
= 0;
5487 rq
->migration_thread
= NULL
;
5488 INIT_LIST_HEAD(&rq
->migration_queue
);
5490 atomic_set(&rq
->nr_iowait
, 0);
5492 for (j
= 0; j
< 2; j
++) {
5493 array
= rq
->arrays
+ j
;
5494 for (k
= 0; k
< MAX_PRIO
; k
++) {
5495 INIT_LIST_HEAD(array
->queue
+ k
);
5496 __clear_bit(k
, array
->bitmap
);
5498 // delimiter for bitsearch
5499 __set_bit(MAX_PRIO
, array
->bitmap
);
5504 * The boot idle thread does lazy MMU switching as well:
5506 atomic_inc(&init_mm
.mm_count
);
5507 enter_lazy_tlb(&init_mm
, current
);
5510 * Make us the idle thread. Technically, schedule() should not be
5511 * called from this thread, however somewhere below it might be,
5512 * but because we are the idle thread, we just pick up running again
5513 * when this runqueue becomes "idle".
5515 init_idle(current
, smp_processor_id());
5518 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5519 void __might_sleep(char *file
, int line
)
5521 #if defined(in_atomic)
5522 static unsigned long prev_jiffy
; /* ratelimiting */
5524 if ((in_atomic() || irqs_disabled()) &&
5525 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
5526 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
5528 prev_jiffy
= jiffies
;
5529 printk(KERN_ERR
"Debug: sleeping function called from invalid"
5530 " context at %s:%d\n", file
, line
);
5531 printk("in_atomic():%d, irqs_disabled():%d\n",
5532 in_atomic(), irqs_disabled());
5537 EXPORT_SYMBOL(__might_sleep
);
5540 #ifdef CONFIG_MAGIC_SYSRQ
5541 void normalize_rt_tasks(void)
5543 struct task_struct
*p
;
5544 prio_array_t
*array
;
5545 unsigned long flags
;
5548 read_lock_irq(&tasklist_lock
);
5549 for_each_process (p
) {
5553 rq
= task_rq_lock(p
, &flags
);
5557 deactivate_task(p
, task_rq(p
));
5558 __setscheduler(p
, SCHED_NORMAL
, 0);
5560 __activate_task(p
, task_rq(p
));
5561 resched_task(rq
->curr
);
5564 task_rq_unlock(rq
, &flags
);
5566 read_unlock_irq(&tasklist_lock
);
5569 #endif /* CONFIG_MAGIC_SYSRQ */