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.
1010 static int find_idlest_cpu(struct sched_group
*group
,
1011 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 * Sync wakeups (i.e. those types of wakeups where the waker
1265 * has indicated that it will leave the CPU in short order)
1266 * don't trigger a preemption, if the woken up task will run on
1267 * this cpu. (in this case the 'I will reschedule' promise of
1268 * the waker guarantees that the freshly woken up task is going
1269 * to be considered on this CPU.)
1271 activate_task(p
, rq
, cpu
== this_cpu
);
1272 if (!sync
|| cpu
!= this_cpu
) {
1273 if (TASK_PREEMPTS_CURR(p
, rq
))
1274 resched_task(rq
->curr
);
1279 p
->state
= TASK_RUNNING
;
1281 task_rq_unlock(rq
, &flags
);
1286 int fastcall
wake_up_process(task_t
* p
)
1288 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1289 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1292 EXPORT_SYMBOL(wake_up_process
);
1294 int fastcall
wake_up_state(task_t
*p
, unsigned int state
)
1296 return try_to_wake_up(p
, state
, 0);
1300 * Perform scheduler related setup for a newly forked process p.
1301 * p is forked by current.
1303 void fastcall
sched_fork(task_t
*p
, int clone_flags
)
1305 int cpu
= get_cpu();
1308 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1310 set_task_cpu(p
, cpu
);
1313 * We mark the process as running here, but have not actually
1314 * inserted it onto the runqueue yet. This guarantees that
1315 * nobody will actually run it, and a signal or other external
1316 * event cannot wake it up and insert it on the runqueue either.
1318 p
->state
= TASK_RUNNING
;
1319 INIT_LIST_HEAD(&p
->run_list
);
1321 #ifdef CONFIG_SCHEDSTATS
1322 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1324 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1327 #ifdef CONFIG_PREEMPT
1328 /* Want to start with kernel preemption disabled. */
1329 p
->thread_info
->preempt_count
= 1;
1332 * Share the timeslice between parent and child, thus the
1333 * total amount of pending timeslices in the system doesn't change,
1334 * resulting in more scheduling fairness.
1336 local_irq_disable();
1337 p
->time_slice
= (current
->time_slice
+ 1) >> 1;
1339 * The remainder of the first timeslice might be recovered by
1340 * the parent if the child exits early enough.
1342 p
->first_time_slice
= 1;
1343 current
->time_slice
>>= 1;
1344 p
->timestamp
= sched_clock();
1345 if (unlikely(!current
->time_slice
)) {
1347 * This case is rare, it happens when the parent has only
1348 * a single jiffy left from its timeslice. Taking the
1349 * runqueue lock is not a problem.
1351 current
->time_slice
= 1;
1359 * wake_up_new_task - wake up a newly created task for the first time.
1361 * This function will do some initial scheduler statistics housekeeping
1362 * that must be done for every newly created context, then puts the task
1363 * on the runqueue and wakes it.
1365 void fastcall
wake_up_new_task(task_t
* p
, unsigned long clone_flags
)
1367 unsigned long flags
;
1369 runqueue_t
*rq
, *this_rq
;
1371 rq
= task_rq_lock(p
, &flags
);
1372 BUG_ON(p
->state
!= TASK_RUNNING
);
1373 this_cpu
= smp_processor_id();
1377 * We decrease the sleep average of forking parents
1378 * and children as well, to keep max-interactive tasks
1379 * from forking tasks that are max-interactive. The parent
1380 * (current) is done further down, under its lock.
1382 p
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(p
) *
1383 CHILD_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1385 p
->prio
= effective_prio(p
);
1387 if (likely(cpu
== this_cpu
)) {
1388 if (!(clone_flags
& CLONE_VM
)) {
1390 * The VM isn't cloned, so we're in a good position to
1391 * do child-runs-first in anticipation of an exec. This
1392 * usually avoids a lot of COW overhead.
1394 if (unlikely(!current
->array
))
1395 __activate_task(p
, rq
);
1397 p
->prio
= current
->prio
;
1398 list_add_tail(&p
->run_list
, ¤t
->run_list
);
1399 p
->array
= current
->array
;
1400 p
->array
->nr_active
++;
1405 /* Run child last */
1406 __activate_task(p
, rq
);
1408 * We skip the following code due to cpu == this_cpu
1410 * task_rq_unlock(rq, &flags);
1411 * this_rq = task_rq_lock(current, &flags);
1415 this_rq
= cpu_rq(this_cpu
);
1418 * Not the local CPU - must adjust timestamp. This should
1419 * get optimised away in the !CONFIG_SMP case.
1421 p
->timestamp
= (p
->timestamp
- this_rq
->timestamp_last_tick
)
1422 + rq
->timestamp_last_tick
;
1423 __activate_task(p
, rq
);
1424 if (TASK_PREEMPTS_CURR(p
, rq
))
1425 resched_task(rq
->curr
);
1428 * Parent and child are on different CPUs, now get the
1429 * parent runqueue to update the parent's ->sleep_avg:
1431 task_rq_unlock(rq
, &flags
);
1432 this_rq
= task_rq_lock(current
, &flags
);
1434 current
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(current
) *
1435 PARENT_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1436 task_rq_unlock(this_rq
, &flags
);
1440 * Potentially available exiting-child timeslices are
1441 * retrieved here - this way the parent does not get
1442 * penalized for creating too many threads.
1444 * (this cannot be used to 'generate' timeslices
1445 * artificially, because any timeslice recovered here
1446 * was given away by the parent in the first place.)
1448 void fastcall
sched_exit(task_t
* p
)
1450 unsigned long flags
;
1454 * If the child was a (relative-) CPU hog then decrease
1455 * the sleep_avg of the parent as well.
1457 rq
= task_rq_lock(p
->parent
, &flags
);
1458 if (p
->first_time_slice
) {
1459 p
->parent
->time_slice
+= p
->time_slice
;
1460 if (unlikely(p
->parent
->time_slice
> task_timeslice(p
)))
1461 p
->parent
->time_slice
= task_timeslice(p
);
1463 if (p
->sleep_avg
< p
->parent
->sleep_avg
)
1464 p
->parent
->sleep_avg
= p
->parent
->sleep_avg
/
1465 (EXIT_WEIGHT
+ 1) * EXIT_WEIGHT
+ p
->sleep_avg
/
1467 task_rq_unlock(rq
, &flags
);
1471 * prepare_task_switch - prepare to switch tasks
1472 * @rq: the runqueue preparing to switch
1473 * @next: the task we are going to switch to.
1475 * This is called with the rq lock held and interrupts off. It must
1476 * be paired with a subsequent finish_task_switch after the context
1479 * prepare_task_switch sets up locking and calls architecture specific
1482 static inline void prepare_task_switch(runqueue_t
*rq
, task_t
*next
)
1484 prepare_lock_switch(rq
, next
);
1485 prepare_arch_switch(next
);
1489 * finish_task_switch - clean up after a task-switch
1490 * @rq: runqueue associated with task-switch
1491 * @prev: the thread we just switched away from.
1493 * finish_task_switch must be called after the context switch, paired
1494 * with a prepare_task_switch call before the context switch.
1495 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1496 * and do any other architecture-specific cleanup actions.
1498 * Note that we may have delayed dropping an mm in context_switch(). If
1499 * so, we finish that here outside of the runqueue lock. (Doing it
1500 * with the lock held can cause deadlocks; see schedule() for
1503 static inline void finish_task_switch(runqueue_t
*rq
, task_t
*prev
)
1504 __releases(rq
->lock
)
1506 struct mm_struct
*mm
= rq
->prev_mm
;
1507 unsigned long prev_task_flags
;
1512 * A task struct has one reference for the use as "current".
1513 * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1514 * calls schedule one last time. The schedule call will never return,
1515 * and the scheduled task must drop that reference.
1516 * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1517 * still held, otherwise prev could be scheduled on another cpu, die
1518 * there before we look at prev->state, and then the reference would
1520 * Manfred Spraul <manfred@colorfullife.com>
1522 prev_task_flags
= prev
->flags
;
1523 #ifdef CONFIG_DEBUG_SPINLOCK
1524 /* this is a valid case when another task releases the spinlock */
1525 rq
->lock
.owner
= current
;
1527 finish_arch_switch(prev
);
1528 finish_lock_switch(rq
, prev
);
1531 if (unlikely(prev_task_flags
& PF_DEAD
))
1532 put_task_struct(prev
);
1536 * schedule_tail - first thing a freshly forked thread must call.
1537 * @prev: the thread we just switched away from.
1539 asmlinkage
void schedule_tail(task_t
*prev
)
1540 __releases(rq
->lock
)
1542 runqueue_t
*rq
= this_rq();
1543 finish_task_switch(rq
, prev
);
1544 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1545 /* In this case, finish_task_switch does not reenable preemption */
1548 if (current
->set_child_tid
)
1549 put_user(current
->pid
, current
->set_child_tid
);
1553 * context_switch - switch to the new MM and the new
1554 * thread's register state.
1557 task_t
* context_switch(runqueue_t
*rq
, task_t
*prev
, task_t
*next
)
1559 struct mm_struct
*mm
= next
->mm
;
1560 struct mm_struct
*oldmm
= prev
->active_mm
;
1562 if (unlikely(!mm
)) {
1563 next
->active_mm
= oldmm
;
1564 atomic_inc(&oldmm
->mm_count
);
1565 enter_lazy_tlb(oldmm
, next
);
1567 switch_mm(oldmm
, mm
, next
);
1569 if (unlikely(!prev
->mm
)) {
1570 prev
->active_mm
= NULL
;
1571 WARN_ON(rq
->prev_mm
);
1572 rq
->prev_mm
= oldmm
;
1575 /* Here we just switch the register state and the stack. */
1576 switch_to(prev
, next
, prev
);
1582 * nr_running, nr_uninterruptible and nr_context_switches:
1584 * externally visible scheduler statistics: current number of runnable
1585 * threads, current number of uninterruptible-sleeping threads, total
1586 * number of context switches performed since bootup.
1588 unsigned long nr_running(void)
1590 unsigned long i
, sum
= 0;
1592 for_each_online_cpu(i
)
1593 sum
+= cpu_rq(i
)->nr_running
;
1598 unsigned long nr_uninterruptible(void)
1600 unsigned long i
, sum
= 0;
1603 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1606 * Since we read the counters lockless, it might be slightly
1607 * inaccurate. Do not allow it to go below zero though:
1609 if (unlikely((long)sum
< 0))
1615 unsigned long long nr_context_switches(void)
1617 unsigned long long i
, sum
= 0;
1620 sum
+= cpu_rq(i
)->nr_switches
;
1625 unsigned long nr_iowait(void)
1627 unsigned long i
, sum
= 0;
1630 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1638 * double_rq_lock - safely lock two runqueues
1640 * Note this does not disable interrupts like task_rq_lock,
1641 * you need to do so manually before calling.
1643 static void double_rq_lock(runqueue_t
*rq1
, runqueue_t
*rq2
)
1644 __acquires(rq1
->lock
)
1645 __acquires(rq2
->lock
)
1648 spin_lock(&rq1
->lock
);
1649 __acquire(rq2
->lock
); /* Fake it out ;) */
1652 spin_lock(&rq1
->lock
);
1653 spin_lock(&rq2
->lock
);
1655 spin_lock(&rq2
->lock
);
1656 spin_lock(&rq1
->lock
);
1662 * double_rq_unlock - safely unlock two runqueues
1664 * Note this does not restore interrupts like task_rq_unlock,
1665 * you need to do so manually after calling.
1667 static void double_rq_unlock(runqueue_t
*rq1
, runqueue_t
*rq2
)
1668 __releases(rq1
->lock
)
1669 __releases(rq2
->lock
)
1671 spin_unlock(&rq1
->lock
);
1673 spin_unlock(&rq2
->lock
);
1675 __release(rq2
->lock
);
1679 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1681 static void double_lock_balance(runqueue_t
*this_rq
, runqueue_t
*busiest
)
1682 __releases(this_rq
->lock
)
1683 __acquires(busiest
->lock
)
1684 __acquires(this_rq
->lock
)
1686 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1687 if (busiest
< this_rq
) {
1688 spin_unlock(&this_rq
->lock
);
1689 spin_lock(&busiest
->lock
);
1690 spin_lock(&this_rq
->lock
);
1692 spin_lock(&busiest
->lock
);
1697 * If dest_cpu is allowed for this process, migrate the task to it.
1698 * This is accomplished by forcing the cpu_allowed mask to only
1699 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
1700 * the cpu_allowed mask is restored.
1702 static void sched_migrate_task(task_t
*p
, int dest_cpu
)
1704 migration_req_t req
;
1706 unsigned long flags
;
1708 rq
= task_rq_lock(p
, &flags
);
1709 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
1710 || unlikely(cpu_is_offline(dest_cpu
)))
1713 /* force the process onto the specified CPU */
1714 if (migrate_task(p
, dest_cpu
, &req
)) {
1715 /* Need to wait for migration thread (might exit: take ref). */
1716 struct task_struct
*mt
= rq
->migration_thread
;
1717 get_task_struct(mt
);
1718 task_rq_unlock(rq
, &flags
);
1719 wake_up_process(mt
);
1720 put_task_struct(mt
);
1721 wait_for_completion(&req
.done
);
1725 task_rq_unlock(rq
, &flags
);
1729 * sched_exec - execve() is a valuable balancing opportunity, because at
1730 * this point the task has the smallest effective memory and cache footprint.
1732 void sched_exec(void)
1734 int new_cpu
, this_cpu
= get_cpu();
1735 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
1737 if (new_cpu
!= this_cpu
)
1738 sched_migrate_task(current
, new_cpu
);
1742 * pull_task - move a task from a remote runqueue to the local runqueue.
1743 * Both runqueues must be locked.
1746 void pull_task(runqueue_t
*src_rq
, prio_array_t
*src_array
, task_t
*p
,
1747 runqueue_t
*this_rq
, prio_array_t
*this_array
, int this_cpu
)
1749 dequeue_task(p
, src_array
);
1750 src_rq
->nr_running
--;
1751 set_task_cpu(p
, this_cpu
);
1752 this_rq
->nr_running
++;
1753 enqueue_task(p
, this_array
);
1754 p
->timestamp
= (p
->timestamp
- src_rq
->timestamp_last_tick
)
1755 + this_rq
->timestamp_last_tick
;
1757 * Note that idle threads have a prio of MAX_PRIO, for this test
1758 * to be always true for them.
1760 if (TASK_PREEMPTS_CURR(p
, this_rq
))
1761 resched_task(this_rq
->curr
);
1765 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1768 int can_migrate_task(task_t
*p
, runqueue_t
*rq
, int this_cpu
,
1769 struct sched_domain
*sd
, enum idle_type idle
, int *all_pinned
)
1772 * We do not migrate tasks that are:
1773 * 1) running (obviously), or
1774 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1775 * 3) are cache-hot on their current CPU.
1777 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
1781 if (task_running(rq
, p
))
1785 * Aggressive migration if:
1786 * 1) task is cache cold, or
1787 * 2) too many balance attempts have failed.
1790 if (sd
->nr_balance_failed
> sd
->cache_nice_tries
)
1793 if (task_hot(p
, rq
->timestamp_last_tick
, sd
))
1799 * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq,
1800 * as part of a balancing operation within "domain". Returns the number of
1803 * Called with both runqueues locked.
1805 static int move_tasks(runqueue_t
*this_rq
, int this_cpu
, runqueue_t
*busiest
,
1806 unsigned long max_nr_move
, struct sched_domain
*sd
,
1807 enum idle_type idle
, int *all_pinned
)
1809 prio_array_t
*array
, *dst_array
;
1810 struct list_head
*head
, *curr
;
1811 int idx
, pulled
= 0, pinned
= 0;
1814 if (max_nr_move
== 0)
1820 * We first consider expired tasks. Those will likely not be
1821 * executed in the near future, and they are most likely to
1822 * be cache-cold, thus switching CPUs has the least effect
1825 if (busiest
->expired
->nr_active
) {
1826 array
= busiest
->expired
;
1827 dst_array
= this_rq
->expired
;
1829 array
= busiest
->active
;
1830 dst_array
= this_rq
->active
;
1834 /* Start searching at priority 0: */
1838 idx
= sched_find_first_bit(array
->bitmap
);
1840 idx
= find_next_bit(array
->bitmap
, MAX_PRIO
, idx
);
1841 if (idx
>= MAX_PRIO
) {
1842 if (array
== busiest
->expired
&& busiest
->active
->nr_active
) {
1843 array
= busiest
->active
;
1844 dst_array
= this_rq
->active
;
1850 head
= array
->queue
+ idx
;
1853 tmp
= list_entry(curr
, task_t
, run_list
);
1857 if (!can_migrate_task(tmp
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
1864 #ifdef CONFIG_SCHEDSTATS
1865 if (task_hot(tmp
, busiest
->timestamp_last_tick
, sd
))
1866 schedstat_inc(sd
, lb_hot_gained
[idle
]);
1869 pull_task(busiest
, array
, tmp
, this_rq
, dst_array
, this_cpu
);
1872 /* We only want to steal up to the prescribed number of tasks. */
1873 if (pulled
< max_nr_move
) {
1881 * Right now, this is the only place pull_task() is called,
1882 * so we can safely collect pull_task() stats here rather than
1883 * inside pull_task().
1885 schedstat_add(sd
, lb_gained
[idle
], pulled
);
1888 *all_pinned
= pinned
;
1893 * find_busiest_group finds and returns the busiest CPU group within the
1894 * domain. It calculates and returns the number of tasks which should be
1895 * moved to restore balance via the imbalance parameter.
1897 static struct sched_group
*
1898 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
1899 unsigned long *imbalance
, enum idle_type idle
)
1901 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1902 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
1905 max_load
= this_load
= total_load
= total_pwr
= 0;
1906 if (idle
== NOT_IDLE
)
1907 load_idx
= sd
->busy_idx
;
1908 else if (idle
== NEWLY_IDLE
)
1909 load_idx
= sd
->newidle_idx
;
1911 load_idx
= sd
->idle_idx
;
1918 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1920 /* Tally up the load of all CPUs in the group */
1923 for_each_cpu_mask(i
, group
->cpumask
) {
1924 /* Bias balancing toward cpus of our domain */
1926 load
= target_load(i
, load_idx
);
1928 load
= source_load(i
, load_idx
);
1933 total_load
+= avg_load
;
1934 total_pwr
+= group
->cpu_power
;
1936 /* Adjust by relative CPU power of the group */
1937 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
1940 this_load
= avg_load
;
1942 } else if (avg_load
> max_load
) {
1943 max_load
= avg_load
;
1946 group
= group
->next
;
1947 } while (group
!= sd
->groups
);
1949 if (!busiest
|| this_load
>= max_load
)
1952 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
1954 if (this_load
>= avg_load
||
1955 100*max_load
<= sd
->imbalance_pct
*this_load
)
1959 * We're trying to get all the cpus to the average_load, so we don't
1960 * want to push ourselves above the average load, nor do we wish to
1961 * reduce the max loaded cpu below the average load, as either of these
1962 * actions would just result in more rebalancing later, and ping-pong
1963 * tasks around. Thus we look for the minimum possible imbalance.
1964 * Negative imbalances (*we* are more loaded than anyone else) will
1965 * be counted as no imbalance for these purposes -- we can't fix that
1966 * by pulling tasks to us. Be careful of negative numbers as they'll
1967 * appear as very large values with unsigned longs.
1969 /* How much load to actually move to equalise the imbalance */
1970 *imbalance
= min((max_load
- avg_load
) * busiest
->cpu_power
,
1971 (avg_load
- this_load
) * this->cpu_power
)
1974 if (*imbalance
< SCHED_LOAD_SCALE
) {
1975 unsigned long pwr_now
= 0, pwr_move
= 0;
1978 if (max_load
- this_load
>= SCHED_LOAD_SCALE
*2) {
1984 * OK, we don't have enough imbalance to justify moving tasks,
1985 * however we may be able to increase total CPU power used by
1989 pwr_now
+= busiest
->cpu_power
*min(SCHED_LOAD_SCALE
, max_load
);
1990 pwr_now
+= this->cpu_power
*min(SCHED_LOAD_SCALE
, this_load
);
1991 pwr_now
/= SCHED_LOAD_SCALE
;
1993 /* Amount of load we'd subtract */
1994 tmp
= SCHED_LOAD_SCALE
*SCHED_LOAD_SCALE
/busiest
->cpu_power
;
1996 pwr_move
+= busiest
->cpu_power
*min(SCHED_LOAD_SCALE
,
1999 /* Amount of load we'd add */
2000 if (max_load
*busiest
->cpu_power
<
2001 SCHED_LOAD_SCALE
*SCHED_LOAD_SCALE
)
2002 tmp
= max_load
*busiest
->cpu_power
/this->cpu_power
;
2004 tmp
= SCHED_LOAD_SCALE
*SCHED_LOAD_SCALE
/this->cpu_power
;
2005 pwr_move
+= this->cpu_power
*min(SCHED_LOAD_SCALE
, this_load
+ tmp
);
2006 pwr_move
/= SCHED_LOAD_SCALE
;
2008 /* Move if we gain throughput */
2009 if (pwr_move
<= pwr_now
)
2016 /* Get rid of the scaling factor, rounding down as we divide */
2017 *imbalance
= *imbalance
/ SCHED_LOAD_SCALE
;
2027 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2029 static runqueue_t
*find_busiest_queue(struct sched_group
*group
)
2031 unsigned long load
, max_load
= 0;
2032 runqueue_t
*busiest
= NULL
;
2035 for_each_cpu_mask(i
, group
->cpumask
) {
2036 load
= source_load(i
, 0);
2038 if (load
> max_load
) {
2040 busiest
= cpu_rq(i
);
2048 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2049 * so long as it is large enough.
2051 #define MAX_PINNED_INTERVAL 512
2054 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2055 * tasks if there is an imbalance.
2057 * Called with this_rq unlocked.
2059 static int load_balance(int this_cpu
, runqueue_t
*this_rq
,
2060 struct sched_domain
*sd
, enum idle_type idle
)
2062 struct sched_group
*group
;
2063 runqueue_t
*busiest
;
2064 unsigned long imbalance
;
2065 int nr_moved
, all_pinned
= 0;
2066 int active_balance
= 0;
2068 spin_lock(&this_rq
->lock
);
2069 schedstat_inc(sd
, lb_cnt
[idle
]);
2071 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
);
2073 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2077 busiest
= find_busiest_queue(group
);
2079 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2083 BUG_ON(busiest
== this_rq
);
2085 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2088 if (busiest
->nr_running
> 1) {
2090 * Attempt to move tasks. If find_busiest_group has found
2091 * an imbalance but busiest->nr_running <= 1, the group is
2092 * still unbalanced. nr_moved simply stays zero, so it is
2093 * correctly treated as an imbalance.
2095 double_lock_balance(this_rq
, busiest
);
2096 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2097 imbalance
, sd
, idle
,
2099 spin_unlock(&busiest
->lock
);
2101 /* All tasks on this runqueue were pinned by CPU affinity */
2102 if (unlikely(all_pinned
))
2106 spin_unlock(&this_rq
->lock
);
2109 schedstat_inc(sd
, lb_failed
[idle
]);
2110 sd
->nr_balance_failed
++;
2112 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2114 spin_lock(&busiest
->lock
);
2115 if (!busiest
->active_balance
) {
2116 busiest
->active_balance
= 1;
2117 busiest
->push_cpu
= this_cpu
;
2120 spin_unlock(&busiest
->lock
);
2122 wake_up_process(busiest
->migration_thread
);
2125 * We've kicked active balancing, reset the failure
2128 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2131 sd
->nr_balance_failed
= 0;
2133 if (likely(!active_balance
)) {
2134 /* We were unbalanced, so reset the balancing interval */
2135 sd
->balance_interval
= sd
->min_interval
;
2138 * If we've begun active balancing, start to back off. This
2139 * case may not be covered by the all_pinned logic if there
2140 * is only 1 task on the busy runqueue (because we don't call
2143 if (sd
->balance_interval
< sd
->max_interval
)
2144 sd
->balance_interval
*= 2;
2150 spin_unlock(&this_rq
->lock
);
2152 schedstat_inc(sd
, lb_balanced
[idle
]);
2154 sd
->nr_balance_failed
= 0;
2155 /* tune up the balancing interval */
2156 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2157 (sd
->balance_interval
< sd
->max_interval
))
2158 sd
->balance_interval
*= 2;
2164 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2165 * tasks if there is an imbalance.
2167 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2168 * this_rq is locked.
2170 static int load_balance_newidle(int this_cpu
, runqueue_t
*this_rq
,
2171 struct sched_domain
*sd
)
2173 struct sched_group
*group
;
2174 runqueue_t
*busiest
= NULL
;
2175 unsigned long imbalance
;
2178 schedstat_inc(sd
, lb_cnt
[NEWLY_IDLE
]);
2179 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, NEWLY_IDLE
);
2181 schedstat_inc(sd
, lb_nobusyg
[NEWLY_IDLE
]);
2185 busiest
= find_busiest_queue(group
);
2187 schedstat_inc(sd
, lb_nobusyq
[NEWLY_IDLE
]);
2191 BUG_ON(busiest
== this_rq
);
2193 /* Attempt to move tasks */
2194 double_lock_balance(this_rq
, busiest
);
2196 schedstat_add(sd
, lb_imbalance
[NEWLY_IDLE
], imbalance
);
2197 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2198 imbalance
, sd
, NEWLY_IDLE
, NULL
);
2200 schedstat_inc(sd
, lb_failed
[NEWLY_IDLE
]);
2202 sd
->nr_balance_failed
= 0;
2204 spin_unlock(&busiest
->lock
);
2208 schedstat_inc(sd
, lb_balanced
[NEWLY_IDLE
]);
2209 sd
->nr_balance_failed
= 0;
2214 * idle_balance is called by schedule() if this_cpu is about to become
2215 * idle. Attempts to pull tasks from other CPUs.
2217 static inline void idle_balance(int this_cpu
, runqueue_t
*this_rq
)
2219 struct sched_domain
*sd
;
2221 for_each_domain(this_cpu
, sd
) {
2222 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
2223 if (load_balance_newidle(this_cpu
, this_rq
, sd
)) {
2224 /* We've pulled tasks over so stop searching */
2232 * active_load_balance is run by migration threads. It pushes running tasks
2233 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2234 * running on each physical CPU where possible, and avoids physical /
2235 * logical imbalances.
2237 * Called with busiest_rq locked.
2239 static void active_load_balance(runqueue_t
*busiest_rq
, int busiest_cpu
)
2241 struct sched_domain
*sd
;
2242 runqueue_t
*target_rq
;
2243 int target_cpu
= busiest_rq
->push_cpu
;
2245 if (busiest_rq
->nr_running
<= 1)
2246 /* no task to move */
2249 target_rq
= cpu_rq(target_cpu
);
2252 * This condition is "impossible", if it occurs
2253 * we need to fix it. Originally reported by
2254 * Bjorn Helgaas on a 128-cpu setup.
2256 BUG_ON(busiest_rq
== target_rq
);
2258 /* move a task from busiest_rq to target_rq */
2259 double_lock_balance(busiest_rq
, target_rq
);
2261 /* Search for an sd spanning us and the target CPU. */
2262 for_each_domain(target_cpu
, sd
)
2263 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2264 cpu_isset(busiest_cpu
, sd
->span
))
2267 if (unlikely(sd
== NULL
))
2270 schedstat_inc(sd
, alb_cnt
);
2272 if (move_tasks(target_rq
, target_cpu
, busiest_rq
, 1, sd
, SCHED_IDLE
, NULL
))
2273 schedstat_inc(sd
, alb_pushed
);
2275 schedstat_inc(sd
, alb_failed
);
2277 spin_unlock(&target_rq
->lock
);
2281 * rebalance_tick will get called every timer tick, on every CPU.
2283 * It checks each scheduling domain to see if it is due to be balanced,
2284 * and initiates a balancing operation if so.
2286 * Balancing parameters are set up in arch_init_sched_domains.
2289 /* Don't have all balancing operations going off at once */
2290 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2292 static void rebalance_tick(int this_cpu
, runqueue_t
*this_rq
,
2293 enum idle_type idle
)
2295 unsigned long old_load
, this_load
;
2296 unsigned long j
= jiffies
+ CPU_OFFSET(this_cpu
);
2297 struct sched_domain
*sd
;
2300 this_load
= this_rq
->nr_running
* SCHED_LOAD_SCALE
;
2301 /* Update our load */
2302 for (i
= 0; i
< 3; i
++) {
2303 unsigned long new_load
= this_load
;
2305 old_load
= this_rq
->cpu_load
[i
];
2307 * Round up the averaging division if load is increasing. This
2308 * prevents us from getting stuck on 9 if the load is 10, for
2311 if (new_load
> old_load
)
2312 new_load
+= scale
-1;
2313 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) / scale
;
2316 for_each_domain(this_cpu
, sd
) {
2317 unsigned long interval
;
2319 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2322 interval
= sd
->balance_interval
;
2323 if (idle
!= SCHED_IDLE
)
2324 interval
*= sd
->busy_factor
;
2326 /* scale ms to jiffies */
2327 interval
= msecs_to_jiffies(interval
);
2328 if (unlikely(!interval
))
2331 if (j
- sd
->last_balance
>= interval
) {
2332 if (load_balance(this_cpu
, this_rq
, sd
, idle
)) {
2333 /* We've pulled tasks over so no longer idle */
2336 sd
->last_balance
+= interval
;
2342 * on UP we do not need to balance between CPUs:
2344 static inline void rebalance_tick(int cpu
, runqueue_t
*rq
, enum idle_type idle
)
2347 static inline void idle_balance(int cpu
, runqueue_t
*rq
)
2352 static inline int wake_priority_sleeper(runqueue_t
*rq
)
2355 #ifdef CONFIG_SCHED_SMT
2356 spin_lock(&rq
->lock
);
2358 * If an SMT sibling task has been put to sleep for priority
2359 * reasons reschedule the idle task to see if it can now run.
2361 if (rq
->nr_running
) {
2362 resched_task(rq
->idle
);
2365 spin_unlock(&rq
->lock
);
2370 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2372 EXPORT_PER_CPU_SYMBOL(kstat
);
2375 * This is called on clock ticks and on context switches.
2376 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2378 static inline void update_cpu_clock(task_t
*p
, runqueue_t
*rq
,
2379 unsigned long long now
)
2381 unsigned long long last
= max(p
->timestamp
, rq
->timestamp_last_tick
);
2382 p
->sched_time
+= now
- last
;
2386 * Return current->sched_time plus any more ns on the sched_clock
2387 * that have not yet been banked.
2389 unsigned long long current_sched_time(const task_t
*tsk
)
2391 unsigned long long ns
;
2392 unsigned long flags
;
2393 local_irq_save(flags
);
2394 ns
= max(tsk
->timestamp
, task_rq(tsk
)->timestamp_last_tick
);
2395 ns
= tsk
->sched_time
+ (sched_clock() - ns
);
2396 local_irq_restore(flags
);
2401 * We place interactive tasks back into the active array, if possible.
2403 * To guarantee that this does not starve expired tasks we ignore the
2404 * interactivity of a task if the first expired task had to wait more
2405 * than a 'reasonable' amount of time. This deadline timeout is
2406 * load-dependent, as the frequency of array switched decreases with
2407 * increasing number of running tasks. We also ignore the interactivity
2408 * if a better static_prio task has expired:
2410 #define EXPIRED_STARVING(rq) \
2411 ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2412 (jiffies - (rq)->expired_timestamp >= \
2413 STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2414 ((rq)->curr->static_prio > (rq)->best_expired_prio))
2417 * Account user cpu time to a process.
2418 * @p: the process that the cpu time gets accounted to
2419 * @hardirq_offset: the offset to subtract from hardirq_count()
2420 * @cputime: the cpu time spent in user space since the last update
2422 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
2424 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2427 p
->utime
= cputime_add(p
->utime
, cputime
);
2429 /* Add user time to cpustat. */
2430 tmp
= cputime_to_cputime64(cputime
);
2431 if (TASK_NICE(p
) > 0)
2432 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
2434 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
2438 * Account system cpu time to a process.
2439 * @p: the process that the cpu time gets accounted to
2440 * @hardirq_offset: the offset to subtract from hardirq_count()
2441 * @cputime: the cpu time spent in kernel space since the last update
2443 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
2446 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2447 runqueue_t
*rq
= this_rq();
2450 p
->stime
= cputime_add(p
->stime
, cputime
);
2452 /* Add system time to cpustat. */
2453 tmp
= cputime_to_cputime64(cputime
);
2454 if (hardirq_count() - hardirq_offset
)
2455 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
2456 else if (softirq_count())
2457 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
2458 else if (p
!= rq
->idle
)
2459 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
2460 else if (atomic_read(&rq
->nr_iowait
) > 0)
2461 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
2463 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
2464 /* Account for system time used */
2465 acct_update_integrals(p
);
2466 /* Update rss highwater mark */
2467 update_mem_hiwater(p
);
2471 * Account for involuntary wait time.
2472 * @p: the process from which the cpu time has been stolen
2473 * @steal: the cpu time spent in involuntary wait
2475 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
2477 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2478 cputime64_t tmp
= cputime_to_cputime64(steal
);
2479 runqueue_t
*rq
= this_rq();
2481 if (p
== rq
->idle
) {
2482 p
->stime
= cputime_add(p
->stime
, steal
);
2483 if (atomic_read(&rq
->nr_iowait
) > 0)
2484 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
2486 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
2488 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
2492 * This function gets called by the timer code, with HZ frequency.
2493 * We call it with interrupts disabled.
2495 * It also gets called by the fork code, when changing the parent's
2498 void scheduler_tick(void)
2500 int cpu
= smp_processor_id();
2501 runqueue_t
*rq
= this_rq();
2502 task_t
*p
= current
;
2503 unsigned long long now
= sched_clock();
2505 update_cpu_clock(p
, rq
, now
);
2507 rq
->timestamp_last_tick
= now
;
2509 if (p
== rq
->idle
) {
2510 if (wake_priority_sleeper(rq
))
2512 rebalance_tick(cpu
, rq
, SCHED_IDLE
);
2516 /* Task might have expired already, but not scheduled off yet */
2517 if (p
->array
!= rq
->active
) {
2518 set_tsk_need_resched(p
);
2521 spin_lock(&rq
->lock
);
2523 * The task was running during this tick - update the
2524 * time slice counter. Note: we do not update a thread's
2525 * priority until it either goes to sleep or uses up its
2526 * timeslice. This makes it possible for interactive tasks
2527 * to use up their timeslices at their highest priority levels.
2531 * RR tasks need a special form of timeslice management.
2532 * FIFO tasks have no timeslices.
2534 if ((p
->policy
== SCHED_RR
) && !--p
->time_slice
) {
2535 p
->time_slice
= task_timeslice(p
);
2536 p
->first_time_slice
= 0;
2537 set_tsk_need_resched(p
);
2539 /* put it at the end of the queue: */
2540 requeue_task(p
, rq
->active
);
2544 if (!--p
->time_slice
) {
2545 dequeue_task(p
, rq
->active
);
2546 set_tsk_need_resched(p
);
2547 p
->prio
= effective_prio(p
);
2548 p
->time_slice
= task_timeslice(p
);
2549 p
->first_time_slice
= 0;
2551 if (!rq
->expired_timestamp
)
2552 rq
->expired_timestamp
= jiffies
;
2553 if (!TASK_INTERACTIVE(p
) || EXPIRED_STARVING(rq
)) {
2554 enqueue_task(p
, rq
->expired
);
2555 if (p
->static_prio
< rq
->best_expired_prio
)
2556 rq
->best_expired_prio
= p
->static_prio
;
2558 enqueue_task(p
, rq
->active
);
2561 * Prevent a too long timeslice allowing a task to monopolize
2562 * the CPU. We do this by splitting up the timeslice into
2565 * Note: this does not mean the task's timeslices expire or
2566 * get lost in any way, they just might be preempted by
2567 * another task of equal priority. (one with higher
2568 * priority would have preempted this task already.) We
2569 * requeue this task to the end of the list on this priority
2570 * level, which is in essence a round-robin of tasks with
2573 * This only applies to tasks in the interactive
2574 * delta range with at least TIMESLICE_GRANULARITY to requeue.
2576 if (TASK_INTERACTIVE(p
) && !((task_timeslice(p
) -
2577 p
->time_slice
) % TIMESLICE_GRANULARITY(p
)) &&
2578 (p
->time_slice
>= TIMESLICE_GRANULARITY(p
)) &&
2579 (p
->array
== rq
->active
)) {
2581 requeue_task(p
, rq
->active
);
2582 set_tsk_need_resched(p
);
2586 spin_unlock(&rq
->lock
);
2588 rebalance_tick(cpu
, rq
, NOT_IDLE
);
2591 #ifdef CONFIG_SCHED_SMT
2592 static inline void wakeup_busy_runqueue(runqueue_t
*rq
)
2594 /* If an SMT runqueue is sleeping due to priority reasons wake it up */
2595 if (rq
->curr
== rq
->idle
&& rq
->nr_running
)
2596 resched_task(rq
->idle
);
2599 static inline void wake_sleeping_dependent(int this_cpu
, runqueue_t
*this_rq
)
2601 struct sched_domain
*tmp
, *sd
= NULL
;
2602 cpumask_t sibling_map
;
2605 for_each_domain(this_cpu
, tmp
)
2606 if (tmp
->flags
& SD_SHARE_CPUPOWER
)
2613 * Unlock the current runqueue because we have to lock in
2614 * CPU order to avoid deadlocks. Caller knows that we might
2615 * unlock. We keep IRQs disabled.
2617 spin_unlock(&this_rq
->lock
);
2619 sibling_map
= sd
->span
;
2621 for_each_cpu_mask(i
, sibling_map
)
2622 spin_lock(&cpu_rq(i
)->lock
);
2624 * We clear this CPU from the mask. This both simplifies the
2625 * inner loop and keps this_rq locked when we exit:
2627 cpu_clear(this_cpu
, sibling_map
);
2629 for_each_cpu_mask(i
, sibling_map
) {
2630 runqueue_t
*smt_rq
= cpu_rq(i
);
2632 wakeup_busy_runqueue(smt_rq
);
2635 for_each_cpu_mask(i
, sibling_map
)
2636 spin_unlock(&cpu_rq(i
)->lock
);
2638 * We exit with this_cpu's rq still held and IRQs
2643 static inline int dependent_sleeper(int this_cpu
, runqueue_t
*this_rq
)
2645 struct sched_domain
*tmp
, *sd
= NULL
;
2646 cpumask_t sibling_map
;
2647 prio_array_t
*array
;
2651 for_each_domain(this_cpu
, tmp
)
2652 if (tmp
->flags
& SD_SHARE_CPUPOWER
)
2659 * The same locking rules and details apply as for
2660 * wake_sleeping_dependent():
2662 spin_unlock(&this_rq
->lock
);
2663 sibling_map
= sd
->span
;
2664 for_each_cpu_mask(i
, sibling_map
)
2665 spin_lock(&cpu_rq(i
)->lock
);
2666 cpu_clear(this_cpu
, sibling_map
);
2669 * Establish next task to be run - it might have gone away because
2670 * we released the runqueue lock above:
2672 if (!this_rq
->nr_running
)
2674 array
= this_rq
->active
;
2675 if (!array
->nr_active
)
2676 array
= this_rq
->expired
;
2677 BUG_ON(!array
->nr_active
);
2679 p
= list_entry(array
->queue
[sched_find_first_bit(array
->bitmap
)].next
,
2682 for_each_cpu_mask(i
, sibling_map
) {
2683 runqueue_t
*smt_rq
= cpu_rq(i
);
2684 task_t
*smt_curr
= smt_rq
->curr
;
2686 /* Kernel threads do not participate in dependent sleeping */
2687 if (!p
->mm
|| !smt_curr
->mm
|| rt_task(p
))
2688 goto check_smt_task
;
2691 * If a user task with lower static priority than the
2692 * running task on the SMT sibling is trying to schedule,
2693 * delay it till there is proportionately less timeslice
2694 * left of the sibling task to prevent a lower priority
2695 * task from using an unfair proportion of the
2696 * physical cpu's resources. -ck
2698 if (rt_task(smt_curr
)) {
2700 * With real time tasks we run non-rt tasks only
2701 * per_cpu_gain% of the time.
2703 if ((jiffies
% DEF_TIMESLICE
) >
2704 (sd
->per_cpu_gain
* DEF_TIMESLICE
/ 100))
2707 if (((smt_curr
->time_slice
* (100 - sd
->per_cpu_gain
) /
2708 100) > task_timeslice(p
)))
2712 if ((!smt_curr
->mm
&& smt_curr
!= smt_rq
->idle
) ||
2716 wakeup_busy_runqueue(smt_rq
);
2721 * Reschedule a lower priority task on the SMT sibling for
2722 * it to be put to sleep, or wake it up if it has been put to
2723 * sleep for priority reasons to see if it should run now.
2726 if ((jiffies
% DEF_TIMESLICE
) >
2727 (sd
->per_cpu_gain
* DEF_TIMESLICE
/ 100))
2728 resched_task(smt_curr
);
2730 if ((p
->time_slice
* (100 - sd
->per_cpu_gain
) / 100) >
2731 task_timeslice(smt_curr
))
2732 resched_task(smt_curr
);
2734 wakeup_busy_runqueue(smt_rq
);
2738 for_each_cpu_mask(i
, sibling_map
)
2739 spin_unlock(&cpu_rq(i
)->lock
);
2743 static inline void wake_sleeping_dependent(int this_cpu
, runqueue_t
*this_rq
)
2747 static inline int dependent_sleeper(int this_cpu
, runqueue_t
*this_rq
)
2753 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
2755 void fastcall
add_preempt_count(int val
)
2760 BUG_ON((preempt_count() < 0));
2761 preempt_count() += val
;
2763 * Spinlock count overflowing soon?
2765 BUG_ON((preempt_count() & PREEMPT_MASK
) >= PREEMPT_MASK
-10);
2767 EXPORT_SYMBOL(add_preempt_count
);
2769 void fastcall
sub_preempt_count(int val
)
2774 BUG_ON(val
> preempt_count());
2776 * Is the spinlock portion underflowing?
2778 BUG_ON((val
< PREEMPT_MASK
) && !(preempt_count() & PREEMPT_MASK
));
2779 preempt_count() -= val
;
2781 EXPORT_SYMBOL(sub_preempt_count
);
2786 * schedule() is the main scheduler function.
2788 asmlinkage
void __sched
schedule(void)
2791 task_t
*prev
, *next
;
2793 prio_array_t
*array
;
2794 struct list_head
*queue
;
2795 unsigned long long now
;
2796 unsigned long run_time
;
2797 int cpu
, idx
, new_prio
;
2800 * Test if we are atomic. Since do_exit() needs to call into
2801 * schedule() atomically, we ignore that path for now.
2802 * Otherwise, whine if we are scheduling when we should not be.
2804 if (likely(!current
->exit_state
)) {
2805 if (unlikely(in_atomic())) {
2806 printk(KERN_ERR
"scheduling while atomic: "
2808 current
->comm
, preempt_count(), current
->pid
);
2812 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2817 release_kernel_lock(prev
);
2818 need_resched_nonpreemptible
:
2822 * The idle thread is not allowed to schedule!
2823 * Remove this check after it has been exercised a bit.
2825 if (unlikely(prev
== rq
->idle
) && prev
->state
!= TASK_RUNNING
) {
2826 printk(KERN_ERR
"bad: scheduling from the idle thread!\n");
2830 schedstat_inc(rq
, sched_cnt
);
2831 now
= sched_clock();
2832 if (likely((long long)(now
- prev
->timestamp
) < NS_MAX_SLEEP_AVG
)) {
2833 run_time
= now
- prev
->timestamp
;
2834 if (unlikely((long long)(now
- prev
->timestamp
) < 0))
2837 run_time
= NS_MAX_SLEEP_AVG
;
2840 * Tasks charged proportionately less run_time at high sleep_avg to
2841 * delay them losing their interactive status
2843 run_time
/= (CURRENT_BONUS(prev
) ? : 1);
2845 spin_lock_irq(&rq
->lock
);
2847 if (unlikely(prev
->flags
& PF_DEAD
))
2848 prev
->state
= EXIT_DEAD
;
2850 switch_count
= &prev
->nivcsw
;
2851 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
2852 switch_count
= &prev
->nvcsw
;
2853 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
2854 unlikely(signal_pending(prev
))))
2855 prev
->state
= TASK_RUNNING
;
2857 if (prev
->state
== TASK_UNINTERRUPTIBLE
)
2858 rq
->nr_uninterruptible
++;
2859 deactivate_task(prev
, rq
);
2863 cpu
= smp_processor_id();
2864 if (unlikely(!rq
->nr_running
)) {
2866 idle_balance(cpu
, rq
);
2867 if (!rq
->nr_running
) {
2869 rq
->expired_timestamp
= 0;
2870 wake_sleeping_dependent(cpu
, rq
);
2872 * wake_sleeping_dependent() might have released
2873 * the runqueue, so break out if we got new
2876 if (!rq
->nr_running
)
2880 if (dependent_sleeper(cpu
, rq
)) {
2885 * dependent_sleeper() releases and reacquires the runqueue
2886 * lock, hence go into the idle loop if the rq went
2889 if (unlikely(!rq
->nr_running
))
2894 if (unlikely(!array
->nr_active
)) {
2896 * Switch the active and expired arrays.
2898 schedstat_inc(rq
, sched_switch
);
2899 rq
->active
= rq
->expired
;
2900 rq
->expired
= array
;
2902 rq
->expired_timestamp
= 0;
2903 rq
->best_expired_prio
= MAX_PRIO
;
2906 idx
= sched_find_first_bit(array
->bitmap
);
2907 queue
= array
->queue
+ idx
;
2908 next
= list_entry(queue
->next
, task_t
, run_list
);
2910 if (!rt_task(next
) && next
->activated
> 0) {
2911 unsigned long long delta
= now
- next
->timestamp
;
2912 if (unlikely((long long)(now
- next
->timestamp
) < 0))
2915 if (next
->activated
== 1)
2916 delta
= delta
* (ON_RUNQUEUE_WEIGHT
* 128 / 100) / 128;
2918 array
= next
->array
;
2919 new_prio
= recalc_task_prio(next
, next
->timestamp
+ delta
);
2921 if (unlikely(next
->prio
!= new_prio
)) {
2922 dequeue_task(next
, array
);
2923 next
->prio
= new_prio
;
2924 enqueue_task(next
, array
);
2926 requeue_task(next
, array
);
2928 next
->activated
= 0;
2930 if (next
== rq
->idle
)
2931 schedstat_inc(rq
, sched_goidle
);
2933 prefetch_stack(next
);
2934 clear_tsk_need_resched(prev
);
2935 rcu_qsctr_inc(task_cpu(prev
));
2937 update_cpu_clock(prev
, rq
, now
);
2939 prev
->sleep_avg
-= run_time
;
2940 if ((long)prev
->sleep_avg
<= 0)
2941 prev
->sleep_avg
= 0;
2942 prev
->timestamp
= prev
->last_ran
= now
;
2944 sched_info_switch(prev
, next
);
2945 if (likely(prev
!= next
)) {
2946 next
->timestamp
= now
;
2951 prepare_task_switch(rq
, next
);
2952 prev
= context_switch(rq
, prev
, next
);
2955 * this_rq must be evaluated again because prev may have moved
2956 * CPUs since it called schedule(), thus the 'rq' on its stack
2957 * frame will be invalid.
2959 finish_task_switch(this_rq(), prev
);
2961 spin_unlock_irq(&rq
->lock
);
2964 if (unlikely(reacquire_kernel_lock(prev
) < 0))
2965 goto need_resched_nonpreemptible
;
2966 preempt_enable_no_resched();
2967 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
2971 EXPORT_SYMBOL(schedule
);
2973 #ifdef CONFIG_PREEMPT
2975 * this is is the entry point to schedule() from in-kernel preemption
2976 * off of preempt_enable. Kernel preemptions off return from interrupt
2977 * occur there and call schedule directly.
2979 asmlinkage
void __sched
preempt_schedule(void)
2981 struct thread_info
*ti
= current_thread_info();
2982 #ifdef CONFIG_PREEMPT_BKL
2983 struct task_struct
*task
= current
;
2984 int saved_lock_depth
;
2987 * If there is a non-zero preempt_count or interrupts are disabled,
2988 * we do not want to preempt the current task. Just return..
2990 if (unlikely(ti
->preempt_count
|| irqs_disabled()))
2994 add_preempt_count(PREEMPT_ACTIVE
);
2996 * We keep the big kernel semaphore locked, but we
2997 * clear ->lock_depth so that schedule() doesnt
2998 * auto-release the semaphore:
3000 #ifdef CONFIG_PREEMPT_BKL
3001 saved_lock_depth
= task
->lock_depth
;
3002 task
->lock_depth
= -1;
3005 #ifdef CONFIG_PREEMPT_BKL
3006 task
->lock_depth
= saved_lock_depth
;
3008 sub_preempt_count(PREEMPT_ACTIVE
);
3010 /* we could miss a preemption opportunity between schedule and now */
3012 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3016 EXPORT_SYMBOL(preempt_schedule
);
3019 * this is is the entry point to schedule() from kernel preemption
3020 * off of irq context.
3021 * Note, that this is called and return with irqs disabled. This will
3022 * protect us against recursive calling from irq.
3024 asmlinkage
void __sched
preempt_schedule_irq(void)
3026 struct thread_info
*ti
= current_thread_info();
3027 #ifdef CONFIG_PREEMPT_BKL
3028 struct task_struct
*task
= current
;
3029 int saved_lock_depth
;
3031 /* Catch callers which need to be fixed*/
3032 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3035 add_preempt_count(PREEMPT_ACTIVE
);
3037 * We keep the big kernel semaphore locked, but we
3038 * clear ->lock_depth so that schedule() doesnt
3039 * auto-release the semaphore:
3041 #ifdef CONFIG_PREEMPT_BKL
3042 saved_lock_depth
= task
->lock_depth
;
3043 task
->lock_depth
= -1;
3047 local_irq_disable();
3048 #ifdef CONFIG_PREEMPT_BKL
3049 task
->lock_depth
= saved_lock_depth
;
3051 sub_preempt_count(PREEMPT_ACTIVE
);
3053 /* we could miss a preemption opportunity between schedule and now */
3055 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3059 #endif /* CONFIG_PREEMPT */
3061 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
, void *key
)
3063 task_t
*p
= curr
->private;
3064 return try_to_wake_up(p
, mode
, sync
);
3067 EXPORT_SYMBOL(default_wake_function
);
3070 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3071 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3072 * number) then we wake all the non-exclusive tasks and one exclusive task.
3074 * There are circumstances in which we can try to wake a task which has already
3075 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3076 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3078 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3079 int nr_exclusive
, int sync
, void *key
)
3081 struct list_head
*tmp
, *next
;
3083 list_for_each_safe(tmp
, next
, &q
->task_list
) {
3086 curr
= list_entry(tmp
, wait_queue_t
, task_list
);
3087 flags
= curr
->flags
;
3088 if (curr
->func(curr
, mode
, sync
, key
) &&
3089 (flags
& WQ_FLAG_EXCLUSIVE
) &&
3096 * __wake_up - wake up threads blocked on a waitqueue.
3098 * @mode: which threads
3099 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3100 * @key: is directly passed to the wakeup function
3102 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3103 int nr_exclusive
, void *key
)
3105 unsigned long flags
;
3107 spin_lock_irqsave(&q
->lock
, flags
);
3108 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3109 spin_unlock_irqrestore(&q
->lock
, flags
);
3112 EXPORT_SYMBOL(__wake_up
);
3115 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3117 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3119 __wake_up_common(q
, mode
, 1, 0, NULL
);
3123 * __wake_up_sync - wake up threads blocked on a waitqueue.
3125 * @mode: which threads
3126 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3128 * The sync wakeup differs that the waker knows that it will schedule
3129 * away soon, so while the target thread will be woken up, it will not
3130 * be migrated to another CPU - ie. the two threads are 'synchronized'
3131 * with each other. This can prevent needless bouncing between CPUs.
3133 * On UP it can prevent extra preemption.
3135 void fastcall
__wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3137 unsigned long flags
;
3143 if (unlikely(!nr_exclusive
))
3146 spin_lock_irqsave(&q
->lock
, flags
);
3147 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3148 spin_unlock_irqrestore(&q
->lock
, flags
);
3150 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3152 void fastcall
complete(struct completion
*x
)
3154 unsigned long flags
;
3156 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3158 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3160 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3162 EXPORT_SYMBOL(complete
);
3164 void fastcall
complete_all(struct completion
*x
)
3166 unsigned long flags
;
3168 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3169 x
->done
+= UINT_MAX
/2;
3170 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3172 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3174 EXPORT_SYMBOL(complete_all
);
3176 void fastcall __sched
wait_for_completion(struct completion
*x
)
3179 spin_lock_irq(&x
->wait
.lock
);
3181 DECLARE_WAITQUEUE(wait
, current
);
3183 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3184 __add_wait_queue_tail(&x
->wait
, &wait
);
3186 __set_current_state(TASK_UNINTERRUPTIBLE
);
3187 spin_unlock_irq(&x
->wait
.lock
);
3189 spin_lock_irq(&x
->wait
.lock
);
3191 __remove_wait_queue(&x
->wait
, &wait
);
3194 spin_unlock_irq(&x
->wait
.lock
);
3196 EXPORT_SYMBOL(wait_for_completion
);
3198 unsigned long fastcall __sched
3199 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3203 spin_lock_irq(&x
->wait
.lock
);
3205 DECLARE_WAITQUEUE(wait
, current
);
3207 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3208 __add_wait_queue_tail(&x
->wait
, &wait
);
3210 __set_current_state(TASK_UNINTERRUPTIBLE
);
3211 spin_unlock_irq(&x
->wait
.lock
);
3212 timeout
= schedule_timeout(timeout
);
3213 spin_lock_irq(&x
->wait
.lock
);
3215 __remove_wait_queue(&x
->wait
, &wait
);
3219 __remove_wait_queue(&x
->wait
, &wait
);
3223 spin_unlock_irq(&x
->wait
.lock
);
3226 EXPORT_SYMBOL(wait_for_completion_timeout
);
3228 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3234 spin_lock_irq(&x
->wait
.lock
);
3236 DECLARE_WAITQUEUE(wait
, current
);
3238 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3239 __add_wait_queue_tail(&x
->wait
, &wait
);
3241 if (signal_pending(current
)) {
3243 __remove_wait_queue(&x
->wait
, &wait
);
3246 __set_current_state(TASK_INTERRUPTIBLE
);
3247 spin_unlock_irq(&x
->wait
.lock
);
3249 spin_lock_irq(&x
->wait
.lock
);
3251 __remove_wait_queue(&x
->wait
, &wait
);
3255 spin_unlock_irq(&x
->wait
.lock
);
3259 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3261 unsigned long fastcall __sched
3262 wait_for_completion_interruptible_timeout(struct completion
*x
,
3263 unsigned long timeout
)
3267 spin_lock_irq(&x
->wait
.lock
);
3269 DECLARE_WAITQUEUE(wait
, current
);
3271 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3272 __add_wait_queue_tail(&x
->wait
, &wait
);
3274 if (signal_pending(current
)) {
3275 timeout
= -ERESTARTSYS
;
3276 __remove_wait_queue(&x
->wait
, &wait
);
3279 __set_current_state(TASK_INTERRUPTIBLE
);
3280 spin_unlock_irq(&x
->wait
.lock
);
3281 timeout
= schedule_timeout(timeout
);
3282 spin_lock_irq(&x
->wait
.lock
);
3284 __remove_wait_queue(&x
->wait
, &wait
);
3288 __remove_wait_queue(&x
->wait
, &wait
);
3292 spin_unlock_irq(&x
->wait
.lock
);
3295 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3298 #define SLEEP_ON_VAR \
3299 unsigned long flags; \
3300 wait_queue_t wait; \
3301 init_waitqueue_entry(&wait, current);
3303 #define SLEEP_ON_HEAD \
3304 spin_lock_irqsave(&q->lock,flags); \
3305 __add_wait_queue(q, &wait); \
3306 spin_unlock(&q->lock);
3308 #define SLEEP_ON_TAIL \
3309 spin_lock_irq(&q->lock); \
3310 __remove_wait_queue(q, &wait); \
3311 spin_unlock_irqrestore(&q->lock, flags);
3313 void fastcall __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3317 current
->state
= TASK_INTERRUPTIBLE
;
3324 EXPORT_SYMBOL(interruptible_sleep_on
);
3326 long fastcall __sched
interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3330 current
->state
= TASK_INTERRUPTIBLE
;
3333 timeout
= schedule_timeout(timeout
);
3339 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3341 void fastcall __sched
sleep_on(wait_queue_head_t
*q
)
3345 current
->state
= TASK_UNINTERRUPTIBLE
;
3352 EXPORT_SYMBOL(sleep_on
);
3354 long fastcall __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3358 current
->state
= TASK_UNINTERRUPTIBLE
;
3361 timeout
= schedule_timeout(timeout
);
3367 EXPORT_SYMBOL(sleep_on_timeout
);
3369 void set_user_nice(task_t
*p
, long nice
)
3371 unsigned long flags
;
3372 prio_array_t
*array
;
3374 int old_prio
, new_prio
, delta
;
3376 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3379 * We have to be careful, if called from sys_setpriority(),
3380 * the task might be in the middle of scheduling on another CPU.
3382 rq
= task_rq_lock(p
, &flags
);
3384 * The RT priorities are set via sched_setscheduler(), but we still
3385 * allow the 'normal' nice value to be set - but as expected
3386 * it wont have any effect on scheduling until the task is
3390 p
->static_prio
= NICE_TO_PRIO(nice
);
3395 dequeue_task(p
, array
);
3398 new_prio
= NICE_TO_PRIO(nice
);
3399 delta
= new_prio
- old_prio
;
3400 p
->static_prio
= NICE_TO_PRIO(nice
);
3404 enqueue_task(p
, array
);
3406 * If the task increased its priority or is running and
3407 * lowered its priority, then reschedule its CPU:
3409 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3410 resched_task(rq
->curr
);
3413 task_rq_unlock(rq
, &flags
);
3416 EXPORT_SYMBOL(set_user_nice
);
3419 * can_nice - check if a task can reduce its nice value
3423 int can_nice(const task_t
*p
, const int nice
)
3425 /* convert nice value [19,-20] to rlimit style value [1,40] */
3426 int nice_rlim
= 20 - nice
;
3427 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
3428 capable(CAP_SYS_NICE
));
3431 #ifdef __ARCH_WANT_SYS_NICE
3434 * sys_nice - change the priority of the current process.
3435 * @increment: priority increment
3437 * sys_setpriority is a more generic, but much slower function that
3438 * does similar things.
3440 asmlinkage
long sys_nice(int increment
)
3446 * Setpriority might change our priority at the same moment.
3447 * We don't have to worry. Conceptually one call occurs first
3448 * and we have a single winner.
3450 if (increment
< -40)
3455 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
3461 if (increment
< 0 && !can_nice(current
, nice
))
3464 retval
= security_task_setnice(current
, nice
);
3468 set_user_nice(current
, nice
);
3475 * task_prio - return the priority value of a given task.
3476 * @p: the task in question.
3478 * This is the priority value as seen by users in /proc.
3479 * RT tasks are offset by -200. Normal tasks are centered
3480 * around 0, value goes from -16 to +15.
3482 int task_prio(const task_t
*p
)
3484 return p
->prio
- MAX_RT_PRIO
;
3488 * task_nice - return the nice value of a given task.
3489 * @p: the task in question.
3491 int task_nice(const task_t
*p
)
3493 return TASK_NICE(p
);
3495 EXPORT_SYMBOL_GPL(task_nice
);
3498 * idle_cpu - is a given cpu idle currently?
3499 * @cpu: the processor in question.
3501 int idle_cpu(int cpu
)
3503 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
3506 EXPORT_SYMBOL_GPL(idle_cpu
);
3509 * idle_task - return the idle task for a given cpu.
3510 * @cpu: the processor in question.
3512 task_t
*idle_task(int cpu
)
3514 return cpu_rq(cpu
)->idle
;
3518 * find_process_by_pid - find a process with a matching PID value.
3519 * @pid: the pid in question.
3521 static inline task_t
*find_process_by_pid(pid_t pid
)
3523 return pid
? find_task_by_pid(pid
) : current
;
3526 /* Actually do priority change: must hold rq lock. */
3527 static void __setscheduler(struct task_struct
*p
, int policy
, int prio
)
3531 p
->rt_priority
= prio
;
3532 if (policy
!= SCHED_NORMAL
)
3533 p
->prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
3535 p
->prio
= p
->static_prio
;
3539 * sched_setscheduler - change the scheduling policy and/or RT priority of
3541 * @p: the task in question.
3542 * @policy: new policy.
3543 * @param: structure containing the new RT priority.
3545 int sched_setscheduler(struct task_struct
*p
, int policy
, struct sched_param
*param
)
3548 int oldprio
, oldpolicy
= -1;
3549 prio_array_t
*array
;
3550 unsigned long flags
;
3554 /* double check policy once rq lock held */
3556 policy
= oldpolicy
= p
->policy
;
3557 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3558 policy
!= SCHED_NORMAL
)
3561 * Valid priorities for SCHED_FIFO and SCHED_RR are
3562 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL is 0.
3564 if (param
->sched_priority
< 0 ||
3565 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3566 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
3568 if ((policy
== SCHED_NORMAL
) != (param
->sched_priority
== 0))
3572 * Allow unprivileged RT tasks to decrease priority:
3574 if (!capable(CAP_SYS_NICE
)) {
3575 /* can't change policy */
3576 if (policy
!= p
->policy
&&
3577 !p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
)
3579 /* can't increase priority */
3580 if (policy
!= SCHED_NORMAL
&&
3581 param
->sched_priority
> p
->rt_priority
&&
3582 param
->sched_priority
>
3583 p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
)
3585 /* can't change other user's priorities */
3586 if ((current
->euid
!= p
->euid
) &&
3587 (current
->euid
!= p
->uid
))
3591 retval
= security_task_setscheduler(p
, policy
, param
);
3595 * To be able to change p->policy safely, the apropriate
3596 * runqueue lock must be held.
3598 rq
= task_rq_lock(p
, &flags
);
3599 /* recheck policy now with rq lock held */
3600 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3601 policy
= oldpolicy
= -1;
3602 task_rq_unlock(rq
, &flags
);
3607 deactivate_task(p
, rq
);
3609 __setscheduler(p
, policy
, param
->sched_priority
);
3611 __activate_task(p
, rq
);
3613 * Reschedule if we are currently running on this runqueue and
3614 * our priority decreased, or if we are not currently running on
3615 * this runqueue and our priority is higher than the current's
3617 if (task_running(rq
, p
)) {
3618 if (p
->prio
> oldprio
)
3619 resched_task(rq
->curr
);
3620 } else if (TASK_PREEMPTS_CURR(p
, rq
))
3621 resched_task(rq
->curr
);
3623 task_rq_unlock(rq
, &flags
);
3626 EXPORT_SYMBOL_GPL(sched_setscheduler
);
3628 static int do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
3631 struct sched_param lparam
;
3632 struct task_struct
*p
;
3634 if (!param
|| pid
< 0)
3636 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
3638 read_lock_irq(&tasklist_lock
);
3639 p
= find_process_by_pid(pid
);
3641 read_unlock_irq(&tasklist_lock
);
3644 retval
= sched_setscheduler(p
, policy
, &lparam
);
3645 read_unlock_irq(&tasklist_lock
);
3650 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3651 * @pid: the pid in question.
3652 * @policy: new policy.
3653 * @param: structure containing the new RT priority.
3655 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
3656 struct sched_param __user
*param
)
3658 return do_sched_setscheduler(pid
, policy
, param
);
3662 * sys_sched_setparam - set/change the RT priority of a thread
3663 * @pid: the pid in question.
3664 * @param: structure containing the new RT priority.
3666 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
3668 return do_sched_setscheduler(pid
, -1, param
);
3672 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3673 * @pid: the pid in question.
3675 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
3677 int retval
= -EINVAL
;
3684 read_lock(&tasklist_lock
);
3685 p
= find_process_by_pid(pid
);
3687 retval
= security_task_getscheduler(p
);
3691 read_unlock(&tasklist_lock
);
3698 * sys_sched_getscheduler - get the RT priority of a thread
3699 * @pid: the pid in question.
3700 * @param: structure containing the RT priority.
3702 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
3704 struct sched_param lp
;
3705 int retval
= -EINVAL
;
3708 if (!param
|| pid
< 0)
3711 read_lock(&tasklist_lock
);
3712 p
= find_process_by_pid(pid
);
3717 retval
= security_task_getscheduler(p
);
3721 lp
.sched_priority
= p
->rt_priority
;
3722 read_unlock(&tasklist_lock
);
3725 * This one might sleep, we cannot do it with a spinlock held ...
3727 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
3733 read_unlock(&tasklist_lock
);
3737 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
3741 cpumask_t cpus_allowed
;
3744 read_lock(&tasklist_lock
);
3746 p
= find_process_by_pid(pid
);
3748 read_unlock(&tasklist_lock
);
3749 unlock_cpu_hotplug();
3754 * It is not safe to call set_cpus_allowed with the
3755 * tasklist_lock held. We will bump the task_struct's
3756 * usage count and then drop tasklist_lock.
3759 read_unlock(&tasklist_lock
);
3762 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
3763 !capable(CAP_SYS_NICE
))
3766 cpus_allowed
= cpuset_cpus_allowed(p
);
3767 cpus_and(new_mask
, new_mask
, cpus_allowed
);
3768 retval
= set_cpus_allowed(p
, new_mask
);
3772 unlock_cpu_hotplug();
3776 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
3777 cpumask_t
*new_mask
)
3779 if (len
< sizeof(cpumask_t
)) {
3780 memset(new_mask
, 0, sizeof(cpumask_t
));
3781 } else if (len
> sizeof(cpumask_t
)) {
3782 len
= sizeof(cpumask_t
);
3784 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
3788 * sys_sched_setaffinity - set the cpu affinity of a process
3789 * @pid: pid of the process
3790 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3791 * @user_mask_ptr: user-space pointer to the new cpu mask
3793 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
3794 unsigned long __user
*user_mask_ptr
)
3799 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
3803 return sched_setaffinity(pid
, new_mask
);
3807 * Represents all cpu's present in the system
3808 * In systems capable of hotplug, this map could dynamically grow
3809 * as new cpu's are detected in the system via any platform specific
3810 * method, such as ACPI for e.g.
3813 cpumask_t cpu_present_map
;
3814 EXPORT_SYMBOL(cpu_present_map
);
3817 cpumask_t cpu_online_map
= CPU_MASK_ALL
;
3818 cpumask_t cpu_possible_map
= CPU_MASK_ALL
;
3821 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
3827 read_lock(&tasklist_lock
);
3830 p
= find_process_by_pid(pid
);
3835 cpus_and(*mask
, p
->cpus_allowed
, cpu_possible_map
);
3838 read_unlock(&tasklist_lock
);
3839 unlock_cpu_hotplug();
3847 * sys_sched_getaffinity - get the cpu affinity of a process
3848 * @pid: pid of the process
3849 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3850 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3852 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
3853 unsigned long __user
*user_mask_ptr
)
3858 if (len
< sizeof(cpumask_t
))
3861 ret
= sched_getaffinity(pid
, &mask
);
3865 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
3868 return sizeof(cpumask_t
);
3872 * sys_sched_yield - yield the current processor to other threads.
3874 * this function yields the current CPU by moving the calling thread
3875 * to the expired array. If there are no other threads running on this
3876 * CPU then this function will return.
3878 asmlinkage
long sys_sched_yield(void)
3880 runqueue_t
*rq
= this_rq_lock();
3881 prio_array_t
*array
= current
->array
;
3882 prio_array_t
*target
= rq
->expired
;
3884 schedstat_inc(rq
, yld_cnt
);
3886 * We implement yielding by moving the task into the expired
3889 * (special rule: RT tasks will just roundrobin in the active
3892 if (rt_task(current
))
3893 target
= rq
->active
;
3895 if (current
->array
->nr_active
== 1) {
3896 schedstat_inc(rq
, yld_act_empty
);
3897 if (!rq
->expired
->nr_active
)
3898 schedstat_inc(rq
, yld_both_empty
);
3899 } else if (!rq
->expired
->nr_active
)
3900 schedstat_inc(rq
, yld_exp_empty
);
3902 if (array
!= target
) {
3903 dequeue_task(current
, array
);
3904 enqueue_task(current
, target
);
3907 * requeue_task is cheaper so perform that if possible.
3909 requeue_task(current
, array
);
3912 * Since we are going to call schedule() anyway, there's
3913 * no need to preempt or enable interrupts:
3915 __release(rq
->lock
);
3916 _raw_spin_unlock(&rq
->lock
);
3917 preempt_enable_no_resched();
3924 static inline void __cond_resched(void)
3927 * The BKS might be reacquired before we have dropped
3928 * PREEMPT_ACTIVE, which could trigger a second
3929 * cond_resched() call.
3931 if (unlikely(preempt_count()))
3934 add_preempt_count(PREEMPT_ACTIVE
);
3936 sub_preempt_count(PREEMPT_ACTIVE
);
3937 } while (need_resched());
3940 int __sched
cond_resched(void)
3942 if (need_resched()) {
3949 EXPORT_SYMBOL(cond_resched
);
3952 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
3953 * call schedule, and on return reacquire the lock.
3955 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
3956 * operations here to prevent schedule() from being called twice (once via
3957 * spin_unlock(), once by hand).
3959 int cond_resched_lock(spinlock_t
* lock
)
3963 if (need_lockbreak(lock
)) {
3969 if (need_resched()) {
3970 _raw_spin_unlock(lock
);
3971 preempt_enable_no_resched();
3979 EXPORT_SYMBOL(cond_resched_lock
);
3981 int __sched
cond_resched_softirq(void)
3983 BUG_ON(!in_softirq());
3985 if (need_resched()) {
3986 __local_bh_enable();
3994 EXPORT_SYMBOL(cond_resched_softirq
);
3998 * yield - yield the current processor to other threads.
4000 * this is a shortcut for kernel-space yielding - it marks the
4001 * thread runnable and calls sys_sched_yield().
4003 void __sched
yield(void)
4005 set_current_state(TASK_RUNNING
);
4009 EXPORT_SYMBOL(yield
);
4012 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4013 * that process accounting knows that this is a task in IO wait state.
4015 * But don't do that if it is a deliberate, throttling IO wait (this task
4016 * has set its backing_dev_info: the queue against which it should throttle)
4018 void __sched
io_schedule(void)
4020 struct runqueue
*rq
= &per_cpu(runqueues
, raw_smp_processor_id());
4022 atomic_inc(&rq
->nr_iowait
);
4024 atomic_dec(&rq
->nr_iowait
);
4027 EXPORT_SYMBOL(io_schedule
);
4029 long __sched
io_schedule_timeout(long timeout
)
4031 struct runqueue
*rq
= &per_cpu(runqueues
, raw_smp_processor_id());
4034 atomic_inc(&rq
->nr_iowait
);
4035 ret
= schedule_timeout(timeout
);
4036 atomic_dec(&rq
->nr_iowait
);
4041 * sys_sched_get_priority_max - return maximum RT priority.
4042 * @policy: scheduling class.
4044 * this syscall returns the maximum rt_priority that can be used
4045 * by a given scheduling class.
4047 asmlinkage
long sys_sched_get_priority_max(int policy
)
4054 ret
= MAX_USER_RT_PRIO
-1;
4064 * sys_sched_get_priority_min - return minimum RT priority.
4065 * @policy: scheduling class.
4067 * this syscall returns the minimum rt_priority that can be used
4068 * by a given scheduling class.
4070 asmlinkage
long sys_sched_get_priority_min(int policy
)
4086 * sys_sched_rr_get_interval - return the default timeslice of a process.
4087 * @pid: pid of the process.
4088 * @interval: userspace pointer to the timeslice value.
4090 * this syscall writes the default timeslice value of a given process
4091 * into the user-space timespec buffer. A value of '0' means infinity.
4094 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4096 int retval
= -EINVAL
;
4104 read_lock(&tasklist_lock
);
4105 p
= find_process_by_pid(pid
);
4109 retval
= security_task_getscheduler(p
);
4113 jiffies_to_timespec(p
->policy
& SCHED_FIFO
?
4114 0 : task_timeslice(p
), &t
);
4115 read_unlock(&tasklist_lock
);
4116 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4120 read_unlock(&tasklist_lock
);
4124 static inline struct task_struct
*eldest_child(struct task_struct
*p
)
4126 if (list_empty(&p
->children
)) return NULL
;
4127 return list_entry(p
->children
.next
,struct task_struct
,sibling
);
4130 static inline struct task_struct
*older_sibling(struct task_struct
*p
)
4132 if (p
->sibling
.prev
==&p
->parent
->children
) return NULL
;
4133 return list_entry(p
->sibling
.prev
,struct task_struct
,sibling
);
4136 static inline struct task_struct
*younger_sibling(struct task_struct
*p
)
4138 if (p
->sibling
.next
==&p
->parent
->children
) return NULL
;
4139 return list_entry(p
->sibling
.next
,struct task_struct
,sibling
);
4142 static void show_task(task_t
* p
)
4146 unsigned long free
= 0;
4147 static const char *stat_nam
[] = { "R", "S", "D", "T", "t", "Z", "X" };
4149 printk("%-13.13s ", p
->comm
);
4150 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4151 if (state
< ARRAY_SIZE(stat_nam
))
4152 printk(stat_nam
[state
]);
4155 #if (BITS_PER_LONG == 32)
4156 if (state
== TASK_RUNNING
)
4157 printk(" running ");
4159 printk(" %08lX ", thread_saved_pc(p
));
4161 if (state
== TASK_RUNNING
)
4162 printk(" running task ");
4164 printk(" %016lx ", thread_saved_pc(p
));
4166 #ifdef CONFIG_DEBUG_STACK_USAGE
4168 unsigned long * n
= (unsigned long *) (p
->thread_info
+1);
4171 free
= (unsigned long) n
- (unsigned long)(p
->thread_info
+1);
4174 printk("%5lu %5d %6d ", free
, p
->pid
, p
->parent
->pid
);
4175 if ((relative
= eldest_child(p
)))
4176 printk("%5d ", relative
->pid
);
4179 if ((relative
= younger_sibling(p
)))
4180 printk("%7d", relative
->pid
);
4183 if ((relative
= older_sibling(p
)))
4184 printk(" %5d", relative
->pid
);
4188 printk(" (L-TLB)\n");
4190 printk(" (NOTLB)\n");
4192 if (state
!= TASK_RUNNING
)
4193 show_stack(p
, NULL
);
4196 void show_state(void)
4200 #if (BITS_PER_LONG == 32)
4203 printk(" task PC pid father child younger older\n");
4207 printk(" task PC pid father child younger older\n");
4209 read_lock(&tasklist_lock
);
4210 do_each_thread(g
, p
) {
4212 * reset the NMI-timeout, listing all files on a slow
4213 * console might take alot of time:
4215 touch_nmi_watchdog();
4217 } while_each_thread(g
, p
);
4219 read_unlock(&tasklist_lock
);
4223 * init_idle - set up an idle thread for a given CPU
4224 * @idle: task in question
4225 * @cpu: cpu the idle task belongs to
4227 * NOTE: this function does not set the idle thread's NEED_RESCHED
4228 * flag, to make booting more robust.
4230 void __devinit
init_idle(task_t
*idle
, int cpu
)
4232 runqueue_t
*rq
= cpu_rq(cpu
);
4233 unsigned long flags
;
4235 idle
->sleep_avg
= 0;
4237 idle
->prio
= MAX_PRIO
;
4238 idle
->state
= TASK_RUNNING
;
4239 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4240 set_task_cpu(idle
, cpu
);
4242 spin_lock_irqsave(&rq
->lock
, flags
);
4243 rq
->curr
= rq
->idle
= idle
;
4244 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4247 spin_unlock_irqrestore(&rq
->lock
, flags
);
4249 /* Set the preempt count _outside_ the spinlocks! */
4250 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4251 idle
->thread_info
->preempt_count
= (idle
->lock_depth
>= 0);
4253 idle
->thread_info
->preempt_count
= 0;
4258 * In a system that switches off the HZ timer nohz_cpu_mask
4259 * indicates which cpus entered this state. This is used
4260 * in the rcu update to wait only for active cpus. For system
4261 * which do not switch off the HZ timer nohz_cpu_mask should
4262 * always be CPU_MASK_NONE.
4264 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4268 * This is how migration works:
4270 * 1) we queue a migration_req_t structure in the source CPU's
4271 * runqueue and wake up that CPU's migration thread.
4272 * 2) we down() the locked semaphore => thread blocks.
4273 * 3) migration thread wakes up (implicitly it forces the migrated
4274 * thread off the CPU)
4275 * 4) it gets the migration request and checks whether the migrated
4276 * task is still in the wrong runqueue.
4277 * 5) if it's in the wrong runqueue then the migration thread removes
4278 * it and puts it into the right queue.
4279 * 6) migration thread up()s the semaphore.
4280 * 7) we wake up and the migration is done.
4284 * Change a given task's CPU affinity. Migrate the thread to a
4285 * proper CPU and schedule it away if the CPU it's executing on
4286 * is removed from the allowed bitmask.
4288 * NOTE: the caller must have a valid reference to the task, the
4289 * task must not exit() & deallocate itself prematurely. The
4290 * call is not atomic; no spinlocks may be held.
4292 int set_cpus_allowed(task_t
*p
, cpumask_t new_mask
)
4294 unsigned long flags
;
4296 migration_req_t req
;
4299 rq
= task_rq_lock(p
, &flags
);
4300 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4305 p
->cpus_allowed
= new_mask
;
4306 /* Can the task run on the task's current CPU? If so, we're done */
4307 if (cpu_isset(task_cpu(p
), new_mask
))
4310 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4311 /* Need help from migration thread: drop lock and wait. */
4312 task_rq_unlock(rq
, &flags
);
4313 wake_up_process(rq
->migration_thread
);
4314 wait_for_completion(&req
.done
);
4315 tlb_migrate_finish(p
->mm
);
4319 task_rq_unlock(rq
, &flags
);
4323 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4326 * Move (not current) task off this cpu, onto dest cpu. We're doing
4327 * this because either it can't run here any more (set_cpus_allowed()
4328 * away from this CPU, or CPU going down), or because we're
4329 * attempting to rebalance this task on exec (sched_exec).
4331 * So we race with normal scheduler movements, but that's OK, as long
4332 * as the task is no longer on this CPU.
4334 static void __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4336 runqueue_t
*rq_dest
, *rq_src
;
4338 if (unlikely(cpu_is_offline(dest_cpu
)))
4341 rq_src
= cpu_rq(src_cpu
);
4342 rq_dest
= cpu_rq(dest_cpu
);
4344 double_rq_lock(rq_src
, rq_dest
);
4345 /* Already moved. */
4346 if (task_cpu(p
) != src_cpu
)
4348 /* Affinity changed (again). */
4349 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
4352 set_task_cpu(p
, dest_cpu
);
4355 * Sync timestamp with rq_dest's before activating.
4356 * The same thing could be achieved by doing this step
4357 * afterwards, and pretending it was a local activate.
4358 * This way is cleaner and logically correct.
4360 p
->timestamp
= p
->timestamp
- rq_src
->timestamp_last_tick
4361 + rq_dest
->timestamp_last_tick
;
4362 deactivate_task(p
, rq_src
);
4363 activate_task(p
, rq_dest
, 0);
4364 if (TASK_PREEMPTS_CURR(p
, rq_dest
))
4365 resched_task(rq_dest
->curr
);
4369 double_rq_unlock(rq_src
, rq_dest
);
4373 * migration_thread - this is a highprio system thread that performs
4374 * thread migration by bumping thread off CPU then 'pushing' onto
4377 static int migration_thread(void * data
)
4380 int cpu
= (long)data
;
4383 BUG_ON(rq
->migration_thread
!= current
);
4385 set_current_state(TASK_INTERRUPTIBLE
);
4386 while (!kthread_should_stop()) {
4387 struct list_head
*head
;
4388 migration_req_t
*req
;
4392 spin_lock_irq(&rq
->lock
);
4394 if (cpu_is_offline(cpu
)) {
4395 spin_unlock_irq(&rq
->lock
);
4399 if (rq
->active_balance
) {
4400 active_load_balance(rq
, cpu
);
4401 rq
->active_balance
= 0;
4404 head
= &rq
->migration_queue
;
4406 if (list_empty(head
)) {
4407 spin_unlock_irq(&rq
->lock
);
4409 set_current_state(TASK_INTERRUPTIBLE
);
4412 req
= list_entry(head
->next
, migration_req_t
, list
);
4413 list_del_init(head
->next
);
4415 spin_unlock(&rq
->lock
);
4416 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
4419 complete(&req
->done
);
4421 __set_current_state(TASK_RUNNING
);
4425 /* Wait for kthread_stop */
4426 set_current_state(TASK_INTERRUPTIBLE
);
4427 while (!kthread_should_stop()) {
4429 set_current_state(TASK_INTERRUPTIBLE
);
4431 __set_current_state(TASK_RUNNING
);
4435 #ifdef CONFIG_HOTPLUG_CPU
4436 /* Figure out where task on dead CPU should go, use force if neccessary. */
4437 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*tsk
)
4443 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
4444 cpus_and(mask
, mask
, tsk
->cpus_allowed
);
4445 dest_cpu
= any_online_cpu(mask
);
4447 /* On any allowed CPU? */
4448 if (dest_cpu
== NR_CPUS
)
4449 dest_cpu
= any_online_cpu(tsk
->cpus_allowed
);
4451 /* No more Mr. Nice Guy. */
4452 if (dest_cpu
== NR_CPUS
) {
4453 cpus_setall(tsk
->cpus_allowed
);
4454 dest_cpu
= any_online_cpu(tsk
->cpus_allowed
);
4457 * Don't tell them about moving exiting tasks or
4458 * kernel threads (both mm NULL), since they never
4461 if (tsk
->mm
&& printk_ratelimit())
4462 printk(KERN_INFO
"process %d (%s) no "
4463 "longer affine to cpu%d\n",
4464 tsk
->pid
, tsk
->comm
, dead_cpu
);
4466 __migrate_task(tsk
, dead_cpu
, dest_cpu
);
4470 * While a dead CPU has no uninterruptible tasks queued at this point,
4471 * it might still have a nonzero ->nr_uninterruptible counter, because
4472 * for performance reasons the counter is not stricly tracking tasks to
4473 * their home CPUs. So we just add the counter to another CPU's counter,
4474 * to keep the global sum constant after CPU-down:
4476 static void migrate_nr_uninterruptible(runqueue_t
*rq_src
)
4478 runqueue_t
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
4479 unsigned long flags
;
4481 local_irq_save(flags
);
4482 double_rq_lock(rq_src
, rq_dest
);
4483 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
4484 rq_src
->nr_uninterruptible
= 0;
4485 double_rq_unlock(rq_src
, rq_dest
);
4486 local_irq_restore(flags
);
4489 /* Run through task list and migrate tasks from the dead cpu. */
4490 static void migrate_live_tasks(int src_cpu
)
4492 struct task_struct
*tsk
, *t
;
4494 write_lock_irq(&tasklist_lock
);
4496 do_each_thread(t
, tsk
) {
4500 if (task_cpu(tsk
) == src_cpu
)
4501 move_task_off_dead_cpu(src_cpu
, tsk
);
4502 } while_each_thread(t
, tsk
);
4504 write_unlock_irq(&tasklist_lock
);
4507 /* Schedules idle task to be the next runnable task on current CPU.
4508 * It does so by boosting its priority to highest possible and adding it to
4509 * the _front_ of runqueue. Used by CPU offline code.
4511 void sched_idle_next(void)
4513 int cpu
= smp_processor_id();
4514 runqueue_t
*rq
= this_rq();
4515 struct task_struct
*p
= rq
->idle
;
4516 unsigned long flags
;
4518 /* cpu has to be offline */
4519 BUG_ON(cpu_online(cpu
));
4521 /* Strictly not necessary since rest of the CPUs are stopped by now
4522 * and interrupts disabled on current cpu.
4524 spin_lock_irqsave(&rq
->lock
, flags
);
4526 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
4527 /* Add idle task to _front_ of it's priority queue */
4528 __activate_idle_task(p
, rq
);
4530 spin_unlock_irqrestore(&rq
->lock
, flags
);
4533 /* Ensures that the idle task is using init_mm right before its cpu goes
4536 void idle_task_exit(void)
4538 struct mm_struct
*mm
= current
->active_mm
;
4540 BUG_ON(cpu_online(smp_processor_id()));
4543 switch_mm(mm
, &init_mm
, current
);
4547 static void migrate_dead(unsigned int dead_cpu
, task_t
*tsk
)
4549 struct runqueue
*rq
= cpu_rq(dead_cpu
);
4551 /* Must be exiting, otherwise would be on tasklist. */
4552 BUG_ON(tsk
->exit_state
!= EXIT_ZOMBIE
&& tsk
->exit_state
!= EXIT_DEAD
);
4554 /* Cannot have done final schedule yet: would have vanished. */
4555 BUG_ON(tsk
->flags
& PF_DEAD
);
4557 get_task_struct(tsk
);
4560 * Drop lock around migration; if someone else moves it,
4561 * that's OK. No task can be added to this CPU, so iteration is
4564 spin_unlock_irq(&rq
->lock
);
4565 move_task_off_dead_cpu(dead_cpu
, tsk
);
4566 spin_lock_irq(&rq
->lock
);
4568 put_task_struct(tsk
);
4571 /* release_task() removes task from tasklist, so we won't find dead tasks. */
4572 static void migrate_dead_tasks(unsigned int dead_cpu
)
4575 struct runqueue
*rq
= cpu_rq(dead_cpu
);
4577 for (arr
= 0; arr
< 2; arr
++) {
4578 for (i
= 0; i
< MAX_PRIO
; i
++) {
4579 struct list_head
*list
= &rq
->arrays
[arr
].queue
[i
];
4580 while (!list_empty(list
))
4581 migrate_dead(dead_cpu
,
4582 list_entry(list
->next
, task_t
,
4587 #endif /* CONFIG_HOTPLUG_CPU */
4590 * migration_call - callback that gets triggered when a CPU is added.
4591 * Here we can start up the necessary migration thread for the new CPU.
4593 static int migration_call(struct notifier_block
*nfb
, unsigned long action
,
4596 int cpu
= (long)hcpu
;
4597 struct task_struct
*p
;
4598 struct runqueue
*rq
;
4599 unsigned long flags
;
4602 case CPU_UP_PREPARE
:
4603 p
= kthread_create(migration_thread
, hcpu
, "migration/%d",cpu
);
4606 p
->flags
|= PF_NOFREEZE
;
4607 kthread_bind(p
, cpu
);
4608 /* Must be high prio: stop_machine expects to yield to it. */
4609 rq
= task_rq_lock(p
, &flags
);
4610 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
4611 task_rq_unlock(rq
, &flags
);
4612 cpu_rq(cpu
)->migration_thread
= p
;
4615 /* Strictly unneccessary, as first user will wake it. */
4616 wake_up_process(cpu_rq(cpu
)->migration_thread
);
4618 #ifdef CONFIG_HOTPLUG_CPU
4619 case CPU_UP_CANCELED
:
4620 /* Unbind it from offline cpu so it can run. Fall thru. */
4621 kthread_bind(cpu_rq(cpu
)->migration_thread
,smp_processor_id());
4622 kthread_stop(cpu_rq(cpu
)->migration_thread
);
4623 cpu_rq(cpu
)->migration_thread
= NULL
;
4626 migrate_live_tasks(cpu
);
4628 kthread_stop(rq
->migration_thread
);
4629 rq
->migration_thread
= NULL
;
4630 /* Idle task back to normal (off runqueue, low prio) */
4631 rq
= task_rq_lock(rq
->idle
, &flags
);
4632 deactivate_task(rq
->idle
, rq
);
4633 rq
->idle
->static_prio
= MAX_PRIO
;
4634 __setscheduler(rq
->idle
, SCHED_NORMAL
, 0);
4635 migrate_dead_tasks(cpu
);
4636 task_rq_unlock(rq
, &flags
);
4637 migrate_nr_uninterruptible(rq
);
4638 BUG_ON(rq
->nr_running
!= 0);
4640 /* No need to migrate the tasks: it was best-effort if
4641 * they didn't do lock_cpu_hotplug(). Just wake up
4642 * the requestors. */
4643 spin_lock_irq(&rq
->lock
);
4644 while (!list_empty(&rq
->migration_queue
)) {
4645 migration_req_t
*req
;
4646 req
= list_entry(rq
->migration_queue
.next
,
4647 migration_req_t
, list
);
4648 list_del_init(&req
->list
);
4649 complete(&req
->done
);
4651 spin_unlock_irq(&rq
->lock
);
4658 /* Register at highest priority so that task migration (migrate_all_tasks)
4659 * happens before everything else.
4661 static struct notifier_block __devinitdata migration_notifier
= {
4662 .notifier_call
= migration_call
,
4666 int __init
migration_init(void)
4668 void *cpu
= (void *)(long)smp_processor_id();
4669 /* Start one for boot CPU. */
4670 migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
4671 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
4672 register_cpu_notifier(&migration_notifier
);
4678 #undef SCHED_DOMAIN_DEBUG
4679 #ifdef SCHED_DOMAIN_DEBUG
4680 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
4685 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
4689 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
4694 struct sched_group
*group
= sd
->groups
;
4695 cpumask_t groupmask
;
4697 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
4698 cpus_clear(groupmask
);
4701 for (i
= 0; i
< level
+ 1; i
++)
4703 printk("domain %d: ", level
);
4705 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
4706 printk("does not load-balance\n");
4708 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain has parent");
4712 printk("span %s\n", str
);
4714 if (!cpu_isset(cpu
, sd
->span
))
4715 printk(KERN_ERR
"ERROR: domain->span does not contain CPU%d\n", cpu
);
4716 if (!cpu_isset(cpu
, group
->cpumask
))
4717 printk(KERN_ERR
"ERROR: domain->groups does not contain CPU%d\n", cpu
);
4720 for (i
= 0; i
< level
+ 2; i
++)
4726 printk(KERN_ERR
"ERROR: group is NULL\n");
4730 if (!group
->cpu_power
) {
4732 printk(KERN_ERR
"ERROR: domain->cpu_power not set\n");
4735 if (!cpus_weight(group
->cpumask
)) {
4737 printk(KERN_ERR
"ERROR: empty group\n");
4740 if (cpus_intersects(groupmask
, group
->cpumask
)) {
4742 printk(KERN_ERR
"ERROR: repeated CPUs\n");
4745 cpus_or(groupmask
, groupmask
, group
->cpumask
);
4747 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
4750 group
= group
->next
;
4751 } while (group
!= sd
->groups
);
4754 if (!cpus_equal(sd
->span
, groupmask
))
4755 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
4761 if (!cpus_subset(groupmask
, sd
->span
))
4762 printk(KERN_ERR
"ERROR: parent span is not a superset of domain->span\n");
4768 #define sched_domain_debug(sd, cpu) {}
4771 static int sd_degenerate(struct sched_domain
*sd
)
4773 if (cpus_weight(sd
->span
) == 1)
4776 /* Following flags need at least 2 groups */
4777 if (sd
->flags
& (SD_LOAD_BALANCE
|
4778 SD_BALANCE_NEWIDLE
|
4781 if (sd
->groups
!= sd
->groups
->next
)
4785 /* Following flags don't use groups */
4786 if (sd
->flags
& (SD_WAKE_IDLE
|
4794 static int sd_parent_degenerate(struct sched_domain
*sd
,
4795 struct sched_domain
*parent
)
4797 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
4799 if (sd_degenerate(parent
))
4802 if (!cpus_equal(sd
->span
, parent
->span
))
4805 /* Does parent contain flags not in child? */
4806 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
4807 if (cflags
& SD_WAKE_AFFINE
)
4808 pflags
&= ~SD_WAKE_BALANCE
;
4809 /* Flags needing groups don't count if only 1 group in parent */
4810 if (parent
->groups
== parent
->groups
->next
) {
4811 pflags
&= ~(SD_LOAD_BALANCE
|
4812 SD_BALANCE_NEWIDLE
|
4816 if (~cflags
& pflags
)
4823 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
4824 * hold the hotplug lock.
4826 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
4828 runqueue_t
*rq
= cpu_rq(cpu
);
4829 struct sched_domain
*tmp
;
4831 /* Remove the sched domains which do not contribute to scheduling. */
4832 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
4833 struct sched_domain
*parent
= tmp
->parent
;
4836 if (sd_parent_degenerate(tmp
, parent
))
4837 tmp
->parent
= parent
->parent
;
4840 if (sd
&& sd_degenerate(sd
))
4843 sched_domain_debug(sd
, cpu
);
4845 rcu_assign_pointer(rq
->sd
, sd
);
4848 /* cpus with isolated domains */
4849 static cpumask_t __devinitdata cpu_isolated_map
= CPU_MASK_NONE
;
4851 /* Setup the mask of cpus configured for isolated domains */
4852 static int __init
isolated_cpu_setup(char *str
)
4854 int ints
[NR_CPUS
], i
;
4856 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
4857 cpus_clear(cpu_isolated_map
);
4858 for (i
= 1; i
<= ints
[0]; i
++)
4859 if (ints
[i
] < NR_CPUS
)
4860 cpu_set(ints
[i
], cpu_isolated_map
);
4864 __setup ("isolcpus=", isolated_cpu_setup
);
4867 * init_sched_build_groups takes an array of groups, the cpumask we wish
4868 * to span, and a pointer to a function which identifies what group a CPU
4869 * belongs to. The return value of group_fn must be a valid index into the
4870 * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
4871 * keep track of groups covered with a cpumask_t).
4873 * init_sched_build_groups will build a circular linked list of the groups
4874 * covered by the given span, and will set each group's ->cpumask correctly,
4875 * and ->cpu_power to 0.
4877 static void init_sched_build_groups(struct sched_group groups
[], cpumask_t span
,
4878 int (*group_fn
)(int cpu
))
4880 struct sched_group
*first
= NULL
, *last
= NULL
;
4881 cpumask_t covered
= CPU_MASK_NONE
;
4884 for_each_cpu_mask(i
, span
) {
4885 int group
= group_fn(i
);
4886 struct sched_group
*sg
= &groups
[group
];
4889 if (cpu_isset(i
, covered
))
4892 sg
->cpumask
= CPU_MASK_NONE
;
4895 for_each_cpu_mask(j
, span
) {
4896 if (group_fn(j
) != group
)
4899 cpu_set(j
, covered
);
4900 cpu_set(j
, sg
->cpumask
);
4911 #define SD_NODES_PER_DOMAIN 16
4915 * find_next_best_node - find the next node to include in a sched_domain
4916 * @node: node whose sched_domain we're building
4917 * @used_nodes: nodes already in the sched_domain
4919 * Find the next node to include in a given scheduling domain. Simply
4920 * finds the closest node not already in the @used_nodes map.
4922 * Should use nodemask_t.
4924 static int find_next_best_node(int node
, unsigned long *used_nodes
)
4926 int i
, n
, val
, min_val
, best_node
= 0;
4930 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
4931 /* Start at @node */
4932 n
= (node
+ i
) % MAX_NUMNODES
;
4934 if (!nr_cpus_node(n
))
4937 /* Skip already used nodes */
4938 if (test_bit(n
, used_nodes
))
4941 /* Simple min distance search */
4942 val
= node_distance(node
, n
);
4944 if (val
< min_val
) {
4950 set_bit(best_node
, used_nodes
);
4955 * sched_domain_node_span - get a cpumask for a node's sched_domain
4956 * @node: node whose cpumask we're constructing
4957 * @size: number of nodes to include in this span
4959 * Given a node, construct a good cpumask for its sched_domain to span. It
4960 * should be one that prevents unnecessary balancing, but also spreads tasks
4963 static cpumask_t
sched_domain_node_span(int node
)
4966 cpumask_t span
, nodemask
;
4967 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
4970 bitmap_zero(used_nodes
, MAX_NUMNODES
);
4972 nodemask
= node_to_cpumask(node
);
4973 cpus_or(span
, span
, nodemask
);
4974 set_bit(node
, used_nodes
);
4976 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
4977 int next_node
= find_next_best_node(node
, used_nodes
);
4978 nodemask
= node_to_cpumask(next_node
);
4979 cpus_or(span
, span
, nodemask
);
4987 * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we
4988 * can switch it on easily if needed.
4990 #ifdef CONFIG_SCHED_SMT
4991 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
4992 static struct sched_group sched_group_cpus
[NR_CPUS
];
4993 static int cpu_to_cpu_group(int cpu
)
4999 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
5000 static struct sched_group sched_group_phys
[NR_CPUS
];
5001 static int cpu_to_phys_group(int cpu
)
5003 #ifdef CONFIG_SCHED_SMT
5004 return first_cpu(cpu_sibling_map
[cpu
]);
5012 * The init_sched_build_groups can't handle what we want to do with node
5013 * groups, so roll our own. Now each node has its own list of groups which
5014 * gets dynamically allocated.
5016 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
5017 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
5019 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
5020 static struct sched_group
*sched_group_allnodes_bycpu
[NR_CPUS
];
5022 static int cpu_to_allnodes_group(int cpu
)
5024 return cpu_to_node(cpu
);
5029 * Build sched domains for a given set of cpus and attach the sched domains
5030 * to the individual cpus
5032 void build_sched_domains(const cpumask_t
*cpu_map
)
5036 struct sched_group
**sched_group_nodes
= NULL
;
5037 struct sched_group
*sched_group_allnodes
= NULL
;
5040 * Allocate the per-node list of sched groups
5042 sched_group_nodes
= kmalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
5044 if (!sched_group_nodes
) {
5045 printk(KERN_WARNING
"Can not alloc sched group node list\n");
5048 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
5052 * Set up domains for cpus specified by the cpu_map.
5054 for_each_cpu_mask(i
, *cpu_map
) {
5056 struct sched_domain
*sd
= NULL
, *p
;
5057 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
5059 cpus_and(nodemask
, nodemask
, *cpu_map
);
5062 if (cpus_weight(*cpu_map
)
5063 > SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
5064 if (!sched_group_allnodes
) {
5065 sched_group_allnodes
5066 = kmalloc(sizeof(struct sched_group
)
5069 if (!sched_group_allnodes
) {
5071 "Can not alloc allnodes sched group\n");
5074 sched_group_allnodes_bycpu
[i
]
5075 = sched_group_allnodes
;
5077 sd
= &per_cpu(allnodes_domains
, i
);
5078 *sd
= SD_ALLNODES_INIT
;
5079 sd
->span
= *cpu_map
;
5080 group
= cpu_to_allnodes_group(i
);
5081 sd
->groups
= &sched_group_allnodes
[group
];
5086 sd
= &per_cpu(node_domains
, i
);
5088 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
5090 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
5094 sd
= &per_cpu(phys_domains
, i
);
5095 group
= cpu_to_phys_group(i
);
5097 sd
->span
= nodemask
;
5099 sd
->groups
= &sched_group_phys
[group
];
5101 #ifdef CONFIG_SCHED_SMT
5103 sd
= &per_cpu(cpu_domains
, i
);
5104 group
= cpu_to_cpu_group(i
);
5105 *sd
= SD_SIBLING_INIT
;
5106 sd
->span
= cpu_sibling_map
[i
];
5107 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
5109 sd
->groups
= &sched_group_cpus
[group
];
5113 #ifdef CONFIG_SCHED_SMT
5114 /* Set up CPU (sibling) groups */
5115 for_each_cpu_mask(i
, *cpu_map
) {
5116 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
5117 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
5118 if (i
!= first_cpu(this_sibling_map
))
5121 init_sched_build_groups(sched_group_cpus
, this_sibling_map
,
5126 /* Set up physical groups */
5127 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5128 cpumask_t nodemask
= node_to_cpumask(i
);
5130 cpus_and(nodemask
, nodemask
, *cpu_map
);
5131 if (cpus_empty(nodemask
))
5134 init_sched_build_groups(sched_group_phys
, nodemask
,
5135 &cpu_to_phys_group
);
5139 /* Set up node groups */
5140 if (sched_group_allnodes
)
5141 init_sched_build_groups(sched_group_allnodes
, *cpu_map
,
5142 &cpu_to_allnodes_group
);
5144 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5145 /* Set up node groups */
5146 struct sched_group
*sg
, *prev
;
5147 cpumask_t nodemask
= node_to_cpumask(i
);
5148 cpumask_t domainspan
;
5149 cpumask_t covered
= CPU_MASK_NONE
;
5152 cpus_and(nodemask
, nodemask
, *cpu_map
);
5153 if (cpus_empty(nodemask
)) {
5154 sched_group_nodes
[i
] = NULL
;
5158 domainspan
= sched_domain_node_span(i
);
5159 cpus_and(domainspan
, domainspan
, *cpu_map
);
5161 sg
= kmalloc(sizeof(struct sched_group
), GFP_KERNEL
);
5162 sched_group_nodes
[i
] = sg
;
5163 for_each_cpu_mask(j
, nodemask
) {
5164 struct sched_domain
*sd
;
5165 sd
= &per_cpu(node_domains
, j
);
5167 if (sd
->groups
== NULL
) {
5168 /* Turn off balancing if we have no groups */
5174 "Can not alloc domain group for node %d\n", i
);
5178 sg
->cpumask
= nodemask
;
5179 cpus_or(covered
, covered
, nodemask
);
5182 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
5183 cpumask_t tmp
, notcovered
;
5184 int n
= (i
+ j
) % MAX_NUMNODES
;
5186 cpus_complement(notcovered
, covered
);
5187 cpus_and(tmp
, notcovered
, *cpu_map
);
5188 cpus_and(tmp
, tmp
, domainspan
);
5189 if (cpus_empty(tmp
))
5192 nodemask
= node_to_cpumask(n
);
5193 cpus_and(tmp
, tmp
, nodemask
);
5194 if (cpus_empty(tmp
))
5197 sg
= kmalloc(sizeof(struct sched_group
), GFP_KERNEL
);
5200 "Can not alloc domain group for node %d\n", j
);
5205 cpus_or(covered
, covered
, tmp
);
5209 prev
->next
= sched_group_nodes
[i
];
5213 /* Calculate CPU power for physical packages and nodes */
5214 for_each_cpu_mask(i
, *cpu_map
) {
5216 struct sched_domain
*sd
;
5217 #ifdef CONFIG_SCHED_SMT
5218 sd
= &per_cpu(cpu_domains
, i
);
5219 power
= SCHED_LOAD_SCALE
;
5220 sd
->groups
->cpu_power
= power
;
5223 sd
= &per_cpu(phys_domains
, i
);
5224 power
= SCHED_LOAD_SCALE
+ SCHED_LOAD_SCALE
*
5225 (cpus_weight(sd
->groups
->cpumask
)-1) / 10;
5226 sd
->groups
->cpu_power
= power
;
5229 sd
= &per_cpu(allnodes_domains
, i
);
5231 power
= SCHED_LOAD_SCALE
+ SCHED_LOAD_SCALE
*
5232 (cpus_weight(sd
->groups
->cpumask
)-1) / 10;
5233 sd
->groups
->cpu_power
= power
;
5239 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5240 struct sched_group
*sg
= sched_group_nodes
[i
];
5246 for_each_cpu_mask(j
, sg
->cpumask
) {
5247 struct sched_domain
*sd
;
5250 sd
= &per_cpu(phys_domains
, j
);
5251 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
5253 * Only add "power" once for each
5258 power
= SCHED_LOAD_SCALE
+ SCHED_LOAD_SCALE
*
5259 (cpus_weight(sd
->groups
->cpumask
)-1) / 10;
5261 sg
->cpu_power
+= power
;
5264 if (sg
!= sched_group_nodes
[i
])
5269 /* Attach the domains */
5270 for_each_cpu_mask(i
, *cpu_map
) {
5271 struct sched_domain
*sd
;
5272 #ifdef CONFIG_SCHED_SMT
5273 sd
= &per_cpu(cpu_domains
, i
);
5275 sd
= &per_cpu(phys_domains
, i
);
5277 cpu_attach_domain(sd
, i
);
5281 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
5283 static void arch_init_sched_domains(const cpumask_t
*cpu_map
)
5285 cpumask_t cpu_default_map
;
5288 * Setup mask for cpus without special case scheduling requirements.
5289 * For now this just excludes isolated cpus, but could be used to
5290 * exclude other special cases in the future.
5292 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
5294 build_sched_domains(&cpu_default_map
);
5297 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
5303 for_each_cpu_mask(cpu
, *cpu_map
) {
5304 struct sched_group
*sched_group_allnodes
5305 = sched_group_allnodes_bycpu
[cpu
];
5306 struct sched_group
**sched_group_nodes
5307 = sched_group_nodes_bycpu
[cpu
];
5309 if (sched_group_allnodes
) {
5310 kfree(sched_group_allnodes
);
5311 sched_group_allnodes_bycpu
[cpu
] = NULL
;
5314 if (!sched_group_nodes
)
5317 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5318 cpumask_t nodemask
= node_to_cpumask(i
);
5319 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
5321 cpus_and(nodemask
, nodemask
, *cpu_map
);
5322 if (cpus_empty(nodemask
))
5332 if (oldsg
!= sched_group_nodes
[i
])
5335 kfree(sched_group_nodes
);
5336 sched_group_nodes_bycpu
[cpu
] = NULL
;
5342 * Detach sched domains from a group of cpus specified in cpu_map
5343 * These cpus will now be attached to the NULL domain
5345 static inline void detach_destroy_domains(const cpumask_t
*cpu_map
)
5349 for_each_cpu_mask(i
, *cpu_map
)
5350 cpu_attach_domain(NULL
, i
);
5351 synchronize_sched();
5352 arch_destroy_sched_domains(cpu_map
);
5356 * Partition sched domains as specified by the cpumasks below.
5357 * This attaches all cpus from the cpumasks to the NULL domain,
5358 * waits for a RCU quiescent period, recalculates sched
5359 * domain information and then attaches them back to the
5360 * correct sched domains
5361 * Call with hotplug lock held
5363 void partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
5365 cpumask_t change_map
;
5367 cpus_and(*partition1
, *partition1
, cpu_online_map
);
5368 cpus_and(*partition2
, *partition2
, cpu_online_map
);
5369 cpus_or(change_map
, *partition1
, *partition2
);
5371 /* Detach sched domains from all of the affected cpus */
5372 detach_destroy_domains(&change_map
);
5373 if (!cpus_empty(*partition1
))
5374 build_sched_domains(partition1
);
5375 if (!cpus_empty(*partition2
))
5376 build_sched_domains(partition2
);
5379 #ifdef CONFIG_HOTPLUG_CPU
5381 * Force a reinitialization of the sched domains hierarchy. The domains
5382 * and groups cannot be updated in place without racing with the balancing
5383 * code, so we temporarily attach all running cpus to the NULL domain
5384 * which will prevent rebalancing while the sched domains are recalculated.
5386 static int update_sched_domains(struct notifier_block
*nfb
,
5387 unsigned long action
, void *hcpu
)
5390 case CPU_UP_PREPARE
:
5391 case CPU_DOWN_PREPARE
:
5392 detach_destroy_domains(&cpu_online_map
);
5395 case CPU_UP_CANCELED
:
5396 case CPU_DOWN_FAILED
:
5400 * Fall through and re-initialise the domains.
5407 /* The hotplug lock is already held by cpu_up/cpu_down */
5408 arch_init_sched_domains(&cpu_online_map
);
5414 void __init
sched_init_smp(void)
5417 arch_init_sched_domains(&cpu_online_map
);
5418 unlock_cpu_hotplug();
5419 /* XXX: Theoretical race here - CPU may be hotplugged now */
5420 hotcpu_notifier(update_sched_domains
, 0);
5423 void __init
sched_init_smp(void)
5426 #endif /* CONFIG_SMP */
5428 int in_sched_functions(unsigned long addr
)
5430 /* Linker adds these: start and end of __sched functions */
5431 extern char __sched_text_start
[], __sched_text_end
[];
5432 return in_lock_functions(addr
) ||
5433 (addr
>= (unsigned long)__sched_text_start
5434 && addr
< (unsigned long)__sched_text_end
);
5437 void __init
sched_init(void)
5442 for (i
= 0; i
< NR_CPUS
; i
++) {
5443 prio_array_t
*array
;
5446 spin_lock_init(&rq
->lock
);
5448 rq
->active
= rq
->arrays
;
5449 rq
->expired
= rq
->arrays
+ 1;
5450 rq
->best_expired_prio
= MAX_PRIO
;
5454 for (j
= 1; j
< 3; j
++)
5455 rq
->cpu_load
[j
] = 0;
5456 rq
->active_balance
= 0;
5458 rq
->migration_thread
= NULL
;
5459 INIT_LIST_HEAD(&rq
->migration_queue
);
5461 atomic_set(&rq
->nr_iowait
, 0);
5463 for (j
= 0; j
< 2; j
++) {
5464 array
= rq
->arrays
+ j
;
5465 for (k
= 0; k
< MAX_PRIO
; k
++) {
5466 INIT_LIST_HEAD(array
->queue
+ k
);
5467 __clear_bit(k
, array
->bitmap
);
5469 // delimiter for bitsearch
5470 __set_bit(MAX_PRIO
, array
->bitmap
);
5475 * The boot idle thread does lazy MMU switching as well:
5477 atomic_inc(&init_mm
.mm_count
);
5478 enter_lazy_tlb(&init_mm
, current
);
5481 * Make us the idle thread. Technically, schedule() should not be
5482 * called from this thread, however somewhere below it might be,
5483 * but because we are the idle thread, we just pick up running again
5484 * when this runqueue becomes "idle".
5486 init_idle(current
, smp_processor_id());
5489 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5490 void __might_sleep(char *file
, int line
)
5492 #if defined(in_atomic)
5493 static unsigned long prev_jiffy
; /* ratelimiting */
5495 if ((in_atomic() || irqs_disabled()) &&
5496 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
5497 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
5499 prev_jiffy
= jiffies
;
5500 printk(KERN_ERR
"Debug: sleeping function called from invalid"
5501 " context at %s:%d\n", file
, line
);
5502 printk("in_atomic():%d, irqs_disabled():%d\n",
5503 in_atomic(), irqs_disabled());
5508 EXPORT_SYMBOL(__might_sleep
);
5511 #ifdef CONFIG_MAGIC_SYSRQ
5512 void normalize_rt_tasks(void)
5514 struct task_struct
*p
;
5515 prio_array_t
*array
;
5516 unsigned long flags
;
5519 read_lock_irq(&tasklist_lock
);
5520 for_each_process (p
) {
5524 rq
= task_rq_lock(p
, &flags
);
5528 deactivate_task(p
, task_rq(p
));
5529 __setscheduler(p
, SCHED_NORMAL
, 0);
5531 __activate_task(p
, task_rq(p
));
5532 resched_task(rq
->curr
);
5535 task_rq_unlock(rq
, &flags
);
5537 read_unlock_irq(&tasklist_lock
);
5540 #endif /* CONFIG_MAGIC_SYSRQ */