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
);
263 #define for_each_domain(cpu, domain) \
264 for (domain = cpu_rq(cpu)->sd; domain; domain = domain->parent)
266 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
267 #define this_rq() (&__get_cpu_var(runqueues))
268 #define task_rq(p) cpu_rq(task_cpu(p))
269 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
271 #ifndef prepare_arch_switch
272 # define prepare_arch_switch(next) do { } while (0)
274 #ifndef finish_arch_switch
275 # define finish_arch_switch(prev) do { } while (0)
278 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
279 static inline int task_running(runqueue_t
*rq
, task_t
*p
)
281 return rq
->curr
== p
;
284 static inline void prepare_lock_switch(runqueue_t
*rq
, task_t
*next
)
288 static inline void finish_lock_switch(runqueue_t
*rq
, task_t
*prev
)
290 spin_unlock_irq(&rq
->lock
);
293 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
294 static inline int task_running(runqueue_t
*rq
, task_t
*p
)
299 return rq
->curr
== p
;
303 static inline void prepare_lock_switch(runqueue_t
*rq
, task_t
*next
)
307 * We can optimise this out completely for !SMP, because the
308 * SMP rebalancing from interrupt is the only thing that cares
313 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
314 spin_unlock_irq(&rq
->lock
);
316 spin_unlock(&rq
->lock
);
320 static inline void finish_lock_switch(runqueue_t
*rq
, task_t
*prev
)
324 * After ->oncpu is cleared, the task can be moved to a different CPU.
325 * We must ensure this doesn't happen until the switch is completely
331 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
335 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
338 * task_rq_lock - lock the runqueue a given task resides on and disable
339 * interrupts. Note the ordering: we can safely lookup the task_rq without
340 * explicitly disabling preemption.
342 static inline runqueue_t
*task_rq_lock(task_t
*p
, unsigned long *flags
)
348 local_irq_save(*flags
);
350 spin_lock(&rq
->lock
);
351 if (unlikely(rq
!= task_rq(p
))) {
352 spin_unlock_irqrestore(&rq
->lock
, *flags
);
353 goto repeat_lock_task
;
358 static inline void task_rq_unlock(runqueue_t
*rq
, unsigned long *flags
)
361 spin_unlock_irqrestore(&rq
->lock
, *flags
);
364 #ifdef CONFIG_SCHEDSTATS
366 * bump this up when changing the output format or the meaning of an existing
367 * format, so that tools can adapt (or abort)
369 #define SCHEDSTAT_VERSION 12
371 static int show_schedstat(struct seq_file
*seq
, void *v
)
375 seq_printf(seq
, "version %d\n", SCHEDSTAT_VERSION
);
376 seq_printf(seq
, "timestamp %lu\n", jiffies
);
377 for_each_online_cpu(cpu
) {
378 runqueue_t
*rq
= cpu_rq(cpu
);
380 struct sched_domain
*sd
;
384 /* runqueue-specific stats */
386 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
387 cpu
, rq
->yld_both_empty
,
388 rq
->yld_act_empty
, rq
->yld_exp_empty
, rq
->yld_cnt
,
389 rq
->sched_switch
, rq
->sched_cnt
, rq
->sched_goidle
,
390 rq
->ttwu_cnt
, rq
->ttwu_local
,
391 rq
->rq_sched_info
.cpu_time
,
392 rq
->rq_sched_info
.run_delay
, rq
->rq_sched_info
.pcnt
);
394 seq_printf(seq
, "\n");
397 /* domain-specific stats */
398 for_each_domain(cpu
, sd
) {
399 enum idle_type itype
;
400 char mask_str
[NR_CPUS
];
402 cpumask_scnprintf(mask_str
, NR_CPUS
, sd
->span
);
403 seq_printf(seq
, "domain%d %s", dcnt
++, mask_str
);
404 for (itype
= SCHED_IDLE
; itype
< MAX_IDLE_TYPES
;
406 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu %lu",
408 sd
->lb_balanced
[itype
],
409 sd
->lb_failed
[itype
],
410 sd
->lb_imbalance
[itype
],
411 sd
->lb_gained
[itype
],
412 sd
->lb_hot_gained
[itype
],
413 sd
->lb_nobusyq
[itype
],
414 sd
->lb_nobusyg
[itype
]);
416 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
417 sd
->alb_cnt
, sd
->alb_failed
, sd
->alb_pushed
,
418 sd
->sbe_cnt
, sd
->sbe_balanced
, sd
->sbe_pushed
,
419 sd
->sbf_cnt
, sd
->sbf_balanced
, sd
->sbf_pushed
,
420 sd
->ttwu_wake_remote
, sd
->ttwu_move_affine
, sd
->ttwu_move_balance
);
427 static int schedstat_open(struct inode
*inode
, struct file
*file
)
429 unsigned int size
= PAGE_SIZE
* (1 + num_online_cpus() / 32);
430 char *buf
= kmalloc(size
, GFP_KERNEL
);
436 res
= single_open(file
, show_schedstat
, NULL
);
438 m
= file
->private_data
;
446 struct file_operations proc_schedstat_operations
= {
447 .open
= schedstat_open
,
450 .release
= single_release
,
453 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
454 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
455 #else /* !CONFIG_SCHEDSTATS */
456 # define schedstat_inc(rq, field) do { } while (0)
457 # define schedstat_add(rq, field, amt) do { } while (0)
461 * rq_lock - lock a given runqueue and disable interrupts.
463 static inline runqueue_t
*this_rq_lock(void)
470 spin_lock(&rq
->lock
);
475 #ifdef CONFIG_SCHEDSTATS
477 * Called when a process is dequeued from the active array and given
478 * the cpu. We should note that with the exception of interactive
479 * tasks, the expired queue will become the active queue after the active
480 * queue is empty, without explicitly dequeuing and requeuing tasks in the
481 * expired queue. (Interactive tasks may be requeued directly to the
482 * active queue, thus delaying tasks in the expired queue from running;
483 * see scheduler_tick()).
485 * This function is only called from sched_info_arrive(), rather than
486 * dequeue_task(). Even though a task may be queued and dequeued multiple
487 * times as it is shuffled about, we're really interested in knowing how
488 * long it was from the *first* time it was queued to the time that it
491 static inline void sched_info_dequeued(task_t
*t
)
493 t
->sched_info
.last_queued
= 0;
497 * Called when a task finally hits the cpu. We can now calculate how
498 * long it was waiting to run. We also note when it began so that we
499 * can keep stats on how long its timeslice is.
501 static inline void sched_info_arrive(task_t
*t
)
503 unsigned long now
= jiffies
, diff
= 0;
504 struct runqueue
*rq
= task_rq(t
);
506 if (t
->sched_info
.last_queued
)
507 diff
= now
- t
->sched_info
.last_queued
;
508 sched_info_dequeued(t
);
509 t
->sched_info
.run_delay
+= diff
;
510 t
->sched_info
.last_arrival
= now
;
511 t
->sched_info
.pcnt
++;
516 rq
->rq_sched_info
.run_delay
+= diff
;
517 rq
->rq_sched_info
.pcnt
++;
521 * Called when a process is queued into either the active or expired
522 * array. The time is noted and later used to determine how long we
523 * had to wait for us to reach the cpu. Since the expired queue will
524 * become the active queue after active queue is empty, without dequeuing
525 * and requeuing any tasks, we are interested in queuing to either. It
526 * is unusual but not impossible for tasks to be dequeued and immediately
527 * requeued in the same or another array: this can happen in sched_yield(),
528 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
531 * This function is only called from enqueue_task(), but also only updates
532 * the timestamp if it is already not set. It's assumed that
533 * sched_info_dequeued() will clear that stamp when appropriate.
535 static inline void sched_info_queued(task_t
*t
)
537 if (!t
->sched_info
.last_queued
)
538 t
->sched_info
.last_queued
= jiffies
;
542 * Called when a process ceases being the active-running process, either
543 * voluntarily or involuntarily. Now we can calculate how long we ran.
545 static inline void sched_info_depart(task_t
*t
)
547 struct runqueue
*rq
= task_rq(t
);
548 unsigned long diff
= jiffies
- t
->sched_info
.last_arrival
;
550 t
->sched_info
.cpu_time
+= diff
;
553 rq
->rq_sched_info
.cpu_time
+= diff
;
557 * Called when tasks are switched involuntarily due, typically, to expiring
558 * their time slice. (This may also be called when switching to or from
559 * the idle task.) We are only called when prev != next.
561 static inline void sched_info_switch(task_t
*prev
, task_t
*next
)
563 struct runqueue
*rq
= task_rq(prev
);
566 * prev now departs the cpu. It's not interesting to record
567 * stats about how efficient we were at scheduling the idle
570 if (prev
!= rq
->idle
)
571 sched_info_depart(prev
);
573 if (next
!= rq
->idle
)
574 sched_info_arrive(next
);
577 #define sched_info_queued(t) do { } while (0)
578 #define sched_info_switch(t, next) do { } while (0)
579 #endif /* CONFIG_SCHEDSTATS */
582 * Adding/removing a task to/from a priority array:
584 static void dequeue_task(struct task_struct
*p
, prio_array_t
*array
)
587 list_del(&p
->run_list
);
588 if (list_empty(array
->queue
+ p
->prio
))
589 __clear_bit(p
->prio
, array
->bitmap
);
592 static void enqueue_task(struct task_struct
*p
, prio_array_t
*array
)
594 sched_info_queued(p
);
595 list_add_tail(&p
->run_list
, array
->queue
+ p
->prio
);
596 __set_bit(p
->prio
, array
->bitmap
);
602 * Put task to the end of the run list without the overhead of dequeue
603 * followed by enqueue.
605 static void requeue_task(struct task_struct
*p
, prio_array_t
*array
)
607 list_move_tail(&p
->run_list
, array
->queue
+ p
->prio
);
610 static inline void enqueue_task_head(struct task_struct
*p
, prio_array_t
*array
)
612 list_add(&p
->run_list
, array
->queue
+ p
->prio
);
613 __set_bit(p
->prio
, array
->bitmap
);
619 * effective_prio - return the priority that is based on the static
620 * priority but is modified by bonuses/penalties.
622 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
623 * into the -5 ... 0 ... +5 bonus/penalty range.
625 * We use 25% of the full 0...39 priority range so that:
627 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
628 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
630 * Both properties are important to certain workloads.
632 static int effective_prio(task_t
*p
)
639 bonus
= CURRENT_BONUS(p
) - MAX_BONUS
/ 2;
641 prio
= p
->static_prio
- bonus
;
642 if (prio
< MAX_RT_PRIO
)
644 if (prio
> MAX_PRIO
-1)
650 * __activate_task - move a task to the runqueue.
652 static inline void __activate_task(task_t
*p
, runqueue_t
*rq
)
654 enqueue_task(p
, rq
->active
);
659 * __activate_idle_task - move idle task to the _front_ of runqueue.
661 static inline void __activate_idle_task(task_t
*p
, runqueue_t
*rq
)
663 enqueue_task_head(p
, rq
->active
);
667 static void recalc_task_prio(task_t
*p
, unsigned long long now
)
669 /* Caller must always ensure 'now >= p->timestamp' */
670 unsigned long long __sleep_time
= now
- p
->timestamp
;
671 unsigned long sleep_time
;
673 if (__sleep_time
> NS_MAX_SLEEP_AVG
)
674 sleep_time
= NS_MAX_SLEEP_AVG
;
676 sleep_time
= (unsigned long)__sleep_time
;
678 if (likely(sleep_time
> 0)) {
680 * User tasks that sleep a long time are categorised as
681 * idle and will get just interactive status to stay active &
682 * prevent them suddenly becoming cpu hogs and starving
685 if (p
->mm
&& p
->activated
!= -1 &&
686 sleep_time
> INTERACTIVE_SLEEP(p
)) {
687 p
->sleep_avg
= JIFFIES_TO_NS(MAX_SLEEP_AVG
-
691 * The lower the sleep avg a task has the more
692 * rapidly it will rise with sleep time.
694 sleep_time
*= (MAX_BONUS
- CURRENT_BONUS(p
)) ? : 1;
697 * Tasks waking from uninterruptible sleep are
698 * limited in their sleep_avg rise as they
699 * are likely to be waiting on I/O
701 if (p
->activated
== -1 && p
->mm
) {
702 if (p
->sleep_avg
>= INTERACTIVE_SLEEP(p
))
704 else if (p
->sleep_avg
+ sleep_time
>=
705 INTERACTIVE_SLEEP(p
)) {
706 p
->sleep_avg
= INTERACTIVE_SLEEP(p
);
712 * This code gives a bonus to interactive tasks.
714 * The boost works by updating the 'average sleep time'
715 * value here, based on ->timestamp. The more time a
716 * task spends sleeping, the higher the average gets -
717 * and the higher the priority boost gets as well.
719 p
->sleep_avg
+= sleep_time
;
721 if (p
->sleep_avg
> NS_MAX_SLEEP_AVG
)
722 p
->sleep_avg
= NS_MAX_SLEEP_AVG
;
726 p
->prio
= effective_prio(p
);
730 * activate_task - move a task to the runqueue and do priority recalculation
732 * Update all the scheduling statistics stuff. (sleep average
733 * calculation, priority modifiers, etc.)
735 static void activate_task(task_t
*p
, runqueue_t
*rq
, int local
)
737 unsigned long long now
;
742 /* Compensate for drifting sched_clock */
743 runqueue_t
*this_rq
= this_rq();
744 now
= (now
- this_rq
->timestamp_last_tick
)
745 + rq
->timestamp_last_tick
;
749 recalc_task_prio(p
, now
);
752 * This checks to make sure it's not an uninterruptible task
753 * that is now waking up.
757 * Tasks which were woken up by interrupts (ie. hw events)
758 * are most likely of interactive nature. So we give them
759 * the credit of extending their sleep time to the period
760 * of time they spend on the runqueue, waiting for execution
761 * on a CPU, first time around:
767 * Normal first-time wakeups get a credit too for
768 * on-runqueue time, but it will be weighted down:
775 __activate_task(p
, rq
);
779 * deactivate_task - remove a task from the runqueue.
781 static void deactivate_task(struct task_struct
*p
, runqueue_t
*rq
)
784 dequeue_task(p
, p
->array
);
789 * resched_task - mark a task 'to be rescheduled now'.
791 * On UP this means the setting of the need_resched flag, on SMP it
792 * might also involve a cross-CPU call to trigger the scheduler on
796 static void resched_task(task_t
*p
)
798 int need_resched
, nrpolling
;
800 assert_spin_locked(&task_rq(p
)->lock
);
802 /* minimise the chance of sending an interrupt to poll_idle() */
803 nrpolling
= test_tsk_thread_flag(p
,TIF_POLLING_NRFLAG
);
804 need_resched
= test_and_set_tsk_thread_flag(p
,TIF_NEED_RESCHED
);
805 nrpolling
|= test_tsk_thread_flag(p
,TIF_POLLING_NRFLAG
);
807 if (!need_resched
&& !nrpolling
&& (task_cpu(p
) != smp_processor_id()))
808 smp_send_reschedule(task_cpu(p
));
811 static inline void resched_task(task_t
*p
)
813 set_tsk_need_resched(p
);
818 * task_curr - is this task currently executing on a CPU?
819 * @p: the task in question.
821 inline int task_curr(const task_t
*p
)
823 return cpu_curr(task_cpu(p
)) == p
;
833 struct list_head list
;
834 enum request_type type
;
836 /* For REQ_MOVE_TASK */
840 /* For REQ_SET_DOMAIN */
841 struct sched_domain
*sd
;
843 struct completion done
;
847 * The task's runqueue lock must be held.
848 * Returns true if you have to wait for migration thread.
850 static int migrate_task(task_t
*p
, int dest_cpu
, migration_req_t
*req
)
852 runqueue_t
*rq
= task_rq(p
);
855 * If the task is not on a runqueue (and not running), then
856 * it is sufficient to simply update the task's cpu field.
858 if (!p
->array
&& !task_running(rq
, p
)) {
859 set_task_cpu(p
, dest_cpu
);
863 init_completion(&req
->done
);
864 req
->type
= REQ_MOVE_TASK
;
866 req
->dest_cpu
= dest_cpu
;
867 list_add(&req
->list
, &rq
->migration_queue
);
872 * wait_task_inactive - wait for a thread to unschedule.
874 * The caller must ensure that the task *will* unschedule sometime soon,
875 * else this function might spin for a *long* time. This function can't
876 * be called with interrupts off, or it may introduce deadlock with
877 * smp_call_function() if an IPI is sent by the same process we are
878 * waiting to become inactive.
880 void wait_task_inactive(task_t
* p
)
887 rq
= task_rq_lock(p
, &flags
);
888 /* Must be off runqueue entirely, not preempted. */
889 if (unlikely(p
->array
|| task_running(rq
, p
))) {
890 /* If it's preempted, we yield. It could be a while. */
891 preempted
= !task_running(rq
, p
);
892 task_rq_unlock(rq
, &flags
);
898 task_rq_unlock(rq
, &flags
);
902 * kick_process - kick a running thread to enter/exit the kernel
903 * @p: the to-be-kicked thread
905 * Cause a process which is running on another CPU to enter
906 * kernel-mode, without any delay. (to get signals handled.)
908 * NOTE: this function doesnt have to take the runqueue lock,
909 * because all it wants to ensure is that the remote task enters
910 * the kernel. If the IPI races and the task has been migrated
911 * to another CPU then no harm is done and the purpose has been
914 void kick_process(task_t
*p
)
920 if ((cpu
!= smp_processor_id()) && task_curr(p
))
921 smp_send_reschedule(cpu
);
926 * Return a low guess at the load of a migration-source cpu.
928 * We want to under-estimate the load of migration sources, to
929 * balance conservatively.
931 static inline unsigned long source_load(int cpu
, int type
)
933 runqueue_t
*rq
= cpu_rq(cpu
);
934 unsigned long load_now
= rq
->nr_running
* SCHED_LOAD_SCALE
;
938 return min(rq
->cpu_load
[type
-1], load_now
);
942 * Return a high guess at the load of a migration-target cpu
944 static inline unsigned long target_load(int cpu
, int type
)
946 runqueue_t
*rq
= cpu_rq(cpu
);
947 unsigned long load_now
= rq
->nr_running
* SCHED_LOAD_SCALE
;
951 return max(rq
->cpu_load
[type
-1], load_now
);
955 * find_idlest_group finds and returns the least busy CPU group within the
958 static struct sched_group
*
959 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
961 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
962 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
963 int load_idx
= sd
->forkexec_idx
;
964 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
967 unsigned long load
, avg_load
;
971 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
972 /* XXX: put a cpus allowed check */
974 /* Tally up the load of all CPUs in the group */
977 for_each_cpu_mask(i
, group
->cpumask
) {
978 /* Bias balancing toward cpus of our domain */
980 load
= source_load(i
, load_idx
);
982 load
= target_load(i
, load_idx
);
987 /* Adjust by relative CPU power of the group */
988 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
991 this_load
= avg_load
;
993 } else if (avg_load
< min_load
) {
998 } while (group
!= sd
->groups
);
1000 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1006 * find_idlest_queue - find the idlest runqueue among the cpus in group.
1008 static int find_idlest_cpu(struct sched_group
*group
, int this_cpu
)
1010 unsigned long load
, min_load
= ULONG_MAX
;
1014 for_each_cpu_mask(i
, group
->cpumask
) {
1015 load
= source_load(i
, 0);
1017 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1030 * wake_idle() will wake a task on an idle cpu if task->cpu is
1031 * not idle and an idle cpu is available. The span of cpus to
1032 * search starts with cpus closest then further out as needed,
1033 * so we always favor a closer, idle cpu.
1035 * Returns the CPU we should wake onto.
1037 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1038 static int wake_idle(int cpu
, task_t
*p
)
1041 struct sched_domain
*sd
;
1047 for_each_domain(cpu
, sd
) {
1048 if (sd
->flags
& SD_WAKE_IDLE
) {
1049 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1050 for_each_cpu_mask(i
, tmp
) {
1061 static inline int wake_idle(int cpu
, task_t
*p
)
1068 * try_to_wake_up - wake up a thread
1069 * @p: the to-be-woken-up thread
1070 * @state: the mask of task states that can be woken
1071 * @sync: do a synchronous wakeup?
1073 * Put it on the run-queue if it's not already there. The "current"
1074 * thread is always on the run-queue (except when the actual
1075 * re-schedule is in progress), and as such you're allowed to do
1076 * the simpler "current->state = TASK_RUNNING" to mark yourself
1077 * runnable without the overhead of this.
1079 * returns failure only if the task is already active.
1081 static int try_to_wake_up(task_t
* p
, unsigned int state
, int sync
)
1083 int cpu
, this_cpu
, success
= 0;
1084 unsigned long flags
;
1088 unsigned long load
, this_load
;
1089 struct sched_domain
*sd
, *this_sd
= NULL
;
1093 rq
= task_rq_lock(p
, &flags
);
1094 old_state
= p
->state
;
1095 if (!(old_state
& state
))
1102 this_cpu
= smp_processor_id();
1105 if (unlikely(task_running(rq
, p
)))
1110 schedstat_inc(rq
, ttwu_cnt
);
1111 if (cpu
== this_cpu
) {
1112 schedstat_inc(rq
, ttwu_local
);
1116 for_each_domain(this_cpu
, sd
) {
1117 if (cpu_isset(cpu
, sd
->span
)) {
1118 schedstat_inc(sd
, ttwu_wake_remote
);
1124 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1128 * Check for affine wakeup and passive balancing possibilities.
1131 int idx
= this_sd
->wake_idx
;
1132 unsigned int imbalance
;
1134 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1136 load
= source_load(cpu
, idx
);
1137 this_load
= target_load(this_cpu
, idx
);
1139 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1141 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1142 unsigned long tl
= this_load
;
1144 * If sync wakeup then subtract the (maximum possible)
1145 * effect of the currently running task from the load
1146 * of the current CPU:
1149 tl
-= SCHED_LOAD_SCALE
;
1152 tl
+ target_load(cpu
, idx
) <= SCHED_LOAD_SCALE
) ||
1153 100*(tl
+ SCHED_LOAD_SCALE
) <= imbalance
*load
) {
1155 * This domain has SD_WAKE_AFFINE and
1156 * p is cache cold in this domain, and
1157 * there is no bad imbalance.
1159 schedstat_inc(this_sd
, ttwu_move_affine
);
1165 * Start passive balancing when half the imbalance_pct
1168 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1169 if (imbalance
*this_load
<= 100*load
) {
1170 schedstat_inc(this_sd
, ttwu_move_balance
);
1176 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1178 new_cpu
= wake_idle(new_cpu
, p
);
1179 if (new_cpu
!= cpu
) {
1180 set_task_cpu(p
, new_cpu
);
1181 task_rq_unlock(rq
, &flags
);
1182 /* might preempt at this point */
1183 rq
= task_rq_lock(p
, &flags
);
1184 old_state
= p
->state
;
1185 if (!(old_state
& state
))
1190 this_cpu
= smp_processor_id();
1195 #endif /* CONFIG_SMP */
1196 if (old_state
== TASK_UNINTERRUPTIBLE
) {
1197 rq
->nr_uninterruptible
--;
1199 * Tasks on involuntary sleep don't earn
1200 * sleep_avg beyond just interactive state.
1206 * Sync wakeups (i.e. those types of wakeups where the waker
1207 * has indicated that it will leave the CPU in short order)
1208 * don't trigger a preemption, if the woken up task will run on
1209 * this cpu. (in this case the 'I will reschedule' promise of
1210 * the waker guarantees that the freshly woken up task is going
1211 * to be considered on this CPU.)
1213 activate_task(p
, rq
, cpu
== this_cpu
);
1214 if (!sync
|| cpu
!= this_cpu
) {
1215 if (TASK_PREEMPTS_CURR(p
, rq
))
1216 resched_task(rq
->curr
);
1221 p
->state
= TASK_RUNNING
;
1223 task_rq_unlock(rq
, &flags
);
1228 int fastcall
wake_up_process(task_t
* p
)
1230 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1231 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1234 EXPORT_SYMBOL(wake_up_process
);
1236 int fastcall
wake_up_state(task_t
*p
, unsigned int state
)
1238 return try_to_wake_up(p
, state
, 0);
1242 * Perform scheduler related setup for a newly forked process p.
1243 * p is forked by current.
1245 void fastcall
sched_fork(task_t
*p
)
1248 * We mark the process as running here, but have not actually
1249 * inserted it onto the runqueue yet. This guarantees that
1250 * nobody will actually run it, and a signal or other external
1251 * event cannot wake it up and insert it on the runqueue either.
1253 p
->state
= TASK_RUNNING
;
1254 INIT_LIST_HEAD(&p
->run_list
);
1256 #ifdef CONFIG_SCHEDSTATS
1257 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1259 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1262 #ifdef CONFIG_PREEMPT
1263 /* Want to start with kernel preemption disabled. */
1264 p
->thread_info
->preempt_count
= 1;
1267 * Share the timeslice between parent and child, thus the
1268 * total amount of pending timeslices in the system doesn't change,
1269 * resulting in more scheduling fairness.
1271 local_irq_disable();
1272 p
->time_slice
= (current
->time_slice
+ 1) >> 1;
1274 * The remainder of the first timeslice might be recovered by
1275 * the parent if the child exits early enough.
1277 p
->first_time_slice
= 1;
1278 current
->time_slice
>>= 1;
1279 p
->timestamp
= sched_clock();
1280 if (unlikely(!current
->time_slice
)) {
1282 * This case is rare, it happens when the parent has only
1283 * a single jiffy left from its timeslice. Taking the
1284 * runqueue lock is not a problem.
1286 current
->time_slice
= 1;
1296 * wake_up_new_task - wake up a newly created task for the first time.
1298 * This function will do some initial scheduler statistics housekeeping
1299 * that must be done for every newly created context, then puts the task
1300 * on the runqueue and wakes it.
1302 void fastcall
wake_up_new_task(task_t
* p
, unsigned long clone_flags
)
1304 unsigned long flags
;
1306 runqueue_t
*rq
, *this_rq
;
1308 struct sched_domain
*tmp
, *sd
= NULL
;
1311 rq
= task_rq_lock(p
, &flags
);
1312 BUG_ON(p
->state
!= TASK_RUNNING
);
1313 this_cpu
= smp_processor_id();
1317 for_each_domain(cpu
, tmp
)
1318 if (tmp
->flags
& SD_BALANCE_FORK
)
1324 struct sched_group
*group
;
1327 schedstat_inc(sd
, sbf_cnt
);
1330 group
= find_idlest_group(sd
, p
, cpu
);
1332 schedstat_inc(sd
, sbf_balanced
);
1336 new_cpu
= find_idlest_cpu(group
, cpu
);
1337 if (new_cpu
== -1 || new_cpu
== cpu
) {
1338 schedstat_inc(sd
, sbf_balanced
);
1342 if (cpu_isset(new_cpu
, p
->cpus_allowed
)) {
1343 schedstat_inc(sd
, sbf_pushed
);
1344 set_task_cpu(p
, new_cpu
);
1345 task_rq_unlock(rq
, &flags
);
1346 rq
= task_rq_lock(p
, &flags
);
1350 /* Now try balancing at a lower domain level */
1353 for_each_domain(cpu
, tmp
) {
1354 if (cpus_subset(span
, tmp
->span
))
1356 if (tmp
->flags
& SD_BALANCE_FORK
)
1366 * We decrease the sleep average of forking parents
1367 * and children as well, to keep max-interactive tasks
1368 * from forking tasks that are max-interactive. The parent
1369 * (current) is done further down, under its lock.
1371 p
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(p
) *
1372 CHILD_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1374 p
->prio
= effective_prio(p
);
1376 if (likely(cpu
== this_cpu
)) {
1377 if (!(clone_flags
& CLONE_VM
)) {
1379 * The VM isn't cloned, so we're in a good position to
1380 * do child-runs-first in anticipation of an exec. This
1381 * usually avoids a lot of COW overhead.
1383 if (unlikely(!current
->array
))
1384 __activate_task(p
, rq
);
1386 p
->prio
= current
->prio
;
1387 list_add_tail(&p
->run_list
, ¤t
->run_list
);
1388 p
->array
= current
->array
;
1389 p
->array
->nr_active
++;
1394 /* Run child last */
1395 __activate_task(p
, rq
);
1397 * We skip the following code due to cpu == this_cpu
1399 * task_rq_unlock(rq, &flags);
1400 * this_rq = task_rq_lock(current, &flags);
1404 this_rq
= cpu_rq(this_cpu
);
1407 * Not the local CPU - must adjust timestamp. This should
1408 * get optimised away in the !CONFIG_SMP case.
1410 p
->timestamp
= (p
->timestamp
- this_rq
->timestamp_last_tick
)
1411 + rq
->timestamp_last_tick
;
1412 __activate_task(p
, rq
);
1413 if (TASK_PREEMPTS_CURR(p
, rq
))
1414 resched_task(rq
->curr
);
1417 * Parent and child are on different CPUs, now get the
1418 * parent runqueue to update the parent's ->sleep_avg:
1420 task_rq_unlock(rq
, &flags
);
1421 this_rq
= task_rq_lock(current
, &flags
);
1423 current
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(current
) *
1424 PARENT_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1425 task_rq_unlock(this_rq
, &flags
);
1429 * Potentially available exiting-child timeslices are
1430 * retrieved here - this way the parent does not get
1431 * penalized for creating too many threads.
1433 * (this cannot be used to 'generate' timeslices
1434 * artificially, because any timeslice recovered here
1435 * was given away by the parent in the first place.)
1437 void fastcall
sched_exit(task_t
* p
)
1439 unsigned long flags
;
1443 * If the child was a (relative-) CPU hog then decrease
1444 * the sleep_avg of the parent as well.
1446 rq
= task_rq_lock(p
->parent
, &flags
);
1447 if (p
->first_time_slice
) {
1448 p
->parent
->time_slice
+= p
->time_slice
;
1449 if (unlikely(p
->parent
->time_slice
> task_timeslice(p
)))
1450 p
->parent
->time_slice
= task_timeslice(p
);
1452 if (p
->sleep_avg
< p
->parent
->sleep_avg
)
1453 p
->parent
->sleep_avg
= p
->parent
->sleep_avg
/
1454 (EXIT_WEIGHT
+ 1) * EXIT_WEIGHT
+ p
->sleep_avg
/
1456 task_rq_unlock(rq
, &flags
);
1460 * prepare_task_switch - prepare to switch tasks
1461 * @rq: the runqueue preparing to switch
1462 * @next: the task we are going to switch to.
1464 * This is called with the rq lock held and interrupts off. It must
1465 * be paired with a subsequent finish_task_switch after the context
1468 * prepare_task_switch sets up locking and calls architecture specific
1471 static inline void prepare_task_switch(runqueue_t
*rq
, task_t
*next
)
1473 prepare_lock_switch(rq
, next
);
1474 prepare_arch_switch(next
);
1478 * finish_task_switch - clean up after a task-switch
1479 * @prev: the thread we just switched away from.
1481 * finish_task_switch must be called after the context switch, paired
1482 * with a prepare_task_switch call before the context switch.
1483 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1484 * and do any other architecture-specific cleanup actions.
1486 * Note that we may have delayed dropping an mm in context_switch(). If
1487 * so, we finish that here outside of the runqueue lock. (Doing it
1488 * with the lock held can cause deadlocks; see schedule() for
1491 static inline void finish_task_switch(runqueue_t
*rq
, task_t
*prev
)
1492 __releases(rq
->lock
)
1494 struct mm_struct
*mm
= rq
->prev_mm
;
1495 unsigned long prev_task_flags
;
1500 * A task struct has one reference for the use as "current".
1501 * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1502 * calls schedule one last time. The schedule call will never return,
1503 * and the scheduled task must drop that reference.
1504 * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1505 * still held, otherwise prev could be scheduled on another cpu, die
1506 * there before we look at prev->state, and then the reference would
1508 * Manfred Spraul <manfred@colorfullife.com>
1510 prev_task_flags
= prev
->flags
;
1511 finish_arch_switch(prev
);
1512 finish_lock_switch(rq
, prev
);
1515 if (unlikely(prev_task_flags
& PF_DEAD
))
1516 put_task_struct(prev
);
1520 * schedule_tail - first thing a freshly forked thread must call.
1521 * @prev: the thread we just switched away from.
1523 asmlinkage
void schedule_tail(task_t
*prev
)
1524 __releases(rq
->lock
)
1526 runqueue_t
*rq
= this_rq();
1527 finish_task_switch(rq
, prev
);
1528 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1529 /* In this case, finish_task_switch does not reenable preemption */
1532 if (current
->set_child_tid
)
1533 put_user(current
->pid
, current
->set_child_tid
);
1537 * context_switch - switch to the new MM and the new
1538 * thread's register state.
1541 task_t
* context_switch(runqueue_t
*rq
, task_t
*prev
, task_t
*next
)
1543 struct mm_struct
*mm
= next
->mm
;
1544 struct mm_struct
*oldmm
= prev
->active_mm
;
1546 if (unlikely(!mm
)) {
1547 next
->active_mm
= oldmm
;
1548 atomic_inc(&oldmm
->mm_count
);
1549 enter_lazy_tlb(oldmm
, next
);
1551 switch_mm(oldmm
, mm
, next
);
1553 if (unlikely(!prev
->mm
)) {
1554 prev
->active_mm
= NULL
;
1555 WARN_ON(rq
->prev_mm
);
1556 rq
->prev_mm
= oldmm
;
1559 /* Here we just switch the register state and the stack. */
1560 switch_to(prev
, next
, prev
);
1566 * nr_running, nr_uninterruptible and nr_context_switches:
1568 * externally visible scheduler statistics: current number of runnable
1569 * threads, current number of uninterruptible-sleeping threads, total
1570 * number of context switches performed since bootup.
1572 unsigned long nr_running(void)
1574 unsigned long i
, sum
= 0;
1576 for_each_online_cpu(i
)
1577 sum
+= cpu_rq(i
)->nr_running
;
1582 unsigned long nr_uninterruptible(void)
1584 unsigned long i
, sum
= 0;
1587 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1590 * Since we read the counters lockless, it might be slightly
1591 * inaccurate. Do not allow it to go below zero though:
1593 if (unlikely((long)sum
< 0))
1599 unsigned long long nr_context_switches(void)
1601 unsigned long long i
, sum
= 0;
1604 sum
+= cpu_rq(i
)->nr_switches
;
1609 unsigned long nr_iowait(void)
1611 unsigned long i
, sum
= 0;
1614 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1622 * double_rq_lock - safely lock two runqueues
1624 * Note this does not disable interrupts like task_rq_lock,
1625 * you need to do so manually before calling.
1627 static void double_rq_lock(runqueue_t
*rq1
, runqueue_t
*rq2
)
1628 __acquires(rq1
->lock
)
1629 __acquires(rq2
->lock
)
1632 spin_lock(&rq1
->lock
);
1633 __acquire(rq2
->lock
); /* Fake it out ;) */
1636 spin_lock(&rq1
->lock
);
1637 spin_lock(&rq2
->lock
);
1639 spin_lock(&rq2
->lock
);
1640 spin_lock(&rq1
->lock
);
1646 * double_rq_unlock - safely unlock two runqueues
1648 * Note this does not restore interrupts like task_rq_unlock,
1649 * you need to do so manually after calling.
1651 static void double_rq_unlock(runqueue_t
*rq1
, runqueue_t
*rq2
)
1652 __releases(rq1
->lock
)
1653 __releases(rq2
->lock
)
1655 spin_unlock(&rq1
->lock
);
1657 spin_unlock(&rq2
->lock
);
1659 __release(rq2
->lock
);
1663 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1665 static void double_lock_balance(runqueue_t
*this_rq
, runqueue_t
*busiest
)
1666 __releases(this_rq
->lock
)
1667 __acquires(busiest
->lock
)
1668 __acquires(this_rq
->lock
)
1670 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1671 if (busiest
< this_rq
) {
1672 spin_unlock(&this_rq
->lock
);
1673 spin_lock(&busiest
->lock
);
1674 spin_lock(&this_rq
->lock
);
1676 spin_lock(&busiest
->lock
);
1681 * If dest_cpu is allowed for this process, migrate the task to it.
1682 * This is accomplished by forcing the cpu_allowed mask to only
1683 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
1684 * the cpu_allowed mask is restored.
1686 static void sched_migrate_task(task_t
*p
, int dest_cpu
)
1688 migration_req_t req
;
1690 unsigned long flags
;
1692 rq
= task_rq_lock(p
, &flags
);
1693 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
1694 || unlikely(cpu_is_offline(dest_cpu
)))
1697 /* force the process onto the specified CPU */
1698 if (migrate_task(p
, dest_cpu
, &req
)) {
1699 /* Need to wait for migration thread (might exit: take ref). */
1700 struct task_struct
*mt
= rq
->migration_thread
;
1701 get_task_struct(mt
);
1702 task_rq_unlock(rq
, &flags
);
1703 wake_up_process(mt
);
1704 put_task_struct(mt
);
1705 wait_for_completion(&req
.done
);
1709 task_rq_unlock(rq
, &flags
);
1713 * sched_exec(): find the highest-level, exec-balance-capable
1714 * domain and try to migrate the task to the least loaded CPU.
1716 * execve() is a valuable balancing opportunity, because at this point
1717 * the task has the smallest effective memory and cache footprint.
1719 void sched_exec(void)
1721 struct sched_domain
*tmp
, *sd
= NULL
;
1722 int new_cpu
, this_cpu
= get_cpu();
1724 for_each_domain(this_cpu
, tmp
)
1725 if (tmp
->flags
& SD_BALANCE_EXEC
)
1730 struct sched_group
*group
;
1732 schedstat_inc(sd
, sbe_cnt
);
1734 group
= find_idlest_group(sd
, current
, this_cpu
);
1736 schedstat_inc(sd
, sbe_balanced
);
1739 new_cpu
= find_idlest_cpu(group
, this_cpu
);
1740 if (new_cpu
== -1 || new_cpu
== this_cpu
) {
1741 schedstat_inc(sd
, sbe_balanced
);
1745 schedstat_inc(sd
, sbe_pushed
);
1747 sched_migrate_task(current
, new_cpu
);
1749 /* Now try balancing at a lower domain level */
1750 this_cpu
= get_cpu();
1753 for_each_domain(this_cpu
, tmp
) {
1754 if (cpus_subset(span
, tmp
->span
))
1756 if (tmp
->flags
& SD_BALANCE_EXEC
)
1768 * pull_task - move a task from a remote runqueue to the local runqueue.
1769 * Both runqueues must be locked.
1772 void pull_task(runqueue_t
*src_rq
, prio_array_t
*src_array
, task_t
*p
,
1773 runqueue_t
*this_rq
, prio_array_t
*this_array
, int this_cpu
)
1775 dequeue_task(p
, src_array
);
1776 src_rq
->nr_running
--;
1777 set_task_cpu(p
, this_cpu
);
1778 this_rq
->nr_running
++;
1779 enqueue_task(p
, this_array
);
1780 p
->timestamp
= (p
->timestamp
- src_rq
->timestamp_last_tick
)
1781 + this_rq
->timestamp_last_tick
;
1783 * Note that idle threads have a prio of MAX_PRIO, for this test
1784 * to be always true for them.
1786 if (TASK_PREEMPTS_CURR(p
, this_rq
))
1787 resched_task(this_rq
->curr
);
1791 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1794 int can_migrate_task(task_t
*p
, runqueue_t
*rq
, int this_cpu
,
1795 struct sched_domain
*sd
, enum idle_type idle
, int *all_pinned
)
1798 * We do not migrate tasks that are:
1799 * 1) running (obviously), or
1800 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1801 * 3) are cache-hot on their current CPU.
1803 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
1807 if (task_running(rq
, p
))
1811 * Aggressive migration if:
1812 * 1) task is cache cold, or
1813 * 2) too many balance attempts have failed.
1816 if (sd
->nr_balance_failed
> sd
->cache_nice_tries
)
1819 if (task_hot(p
, rq
->timestamp_last_tick
, sd
))
1825 * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq,
1826 * as part of a balancing operation within "domain". Returns the number of
1829 * Called with both runqueues locked.
1831 static int move_tasks(runqueue_t
*this_rq
, int this_cpu
, runqueue_t
*busiest
,
1832 unsigned long max_nr_move
, struct sched_domain
*sd
,
1833 enum idle_type idle
, int *all_pinned
)
1835 prio_array_t
*array
, *dst_array
;
1836 struct list_head
*head
, *curr
;
1837 int idx
, pulled
= 0, pinned
= 0;
1840 if (max_nr_move
== 0)
1846 * We first consider expired tasks. Those will likely not be
1847 * executed in the near future, and they are most likely to
1848 * be cache-cold, thus switching CPUs has the least effect
1851 if (busiest
->expired
->nr_active
) {
1852 array
= busiest
->expired
;
1853 dst_array
= this_rq
->expired
;
1855 array
= busiest
->active
;
1856 dst_array
= this_rq
->active
;
1860 /* Start searching at priority 0: */
1864 idx
= sched_find_first_bit(array
->bitmap
);
1866 idx
= find_next_bit(array
->bitmap
, MAX_PRIO
, idx
);
1867 if (idx
>= MAX_PRIO
) {
1868 if (array
== busiest
->expired
&& busiest
->active
->nr_active
) {
1869 array
= busiest
->active
;
1870 dst_array
= this_rq
->active
;
1876 head
= array
->queue
+ idx
;
1879 tmp
= list_entry(curr
, task_t
, run_list
);
1883 if (!can_migrate_task(tmp
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
1890 #ifdef CONFIG_SCHEDSTATS
1891 if (task_hot(tmp
, busiest
->timestamp_last_tick
, sd
))
1892 schedstat_inc(sd
, lb_hot_gained
[idle
]);
1895 pull_task(busiest
, array
, tmp
, this_rq
, dst_array
, this_cpu
);
1898 /* We only want to steal up to the prescribed number of tasks. */
1899 if (pulled
< max_nr_move
) {
1907 * Right now, this is the only place pull_task() is called,
1908 * so we can safely collect pull_task() stats here rather than
1909 * inside pull_task().
1911 schedstat_add(sd
, lb_gained
[idle
], pulled
);
1914 *all_pinned
= pinned
;
1919 * find_busiest_group finds and returns the busiest CPU group within the
1920 * domain. It calculates and returns the number of tasks which should be
1921 * moved to restore balance via the imbalance parameter.
1923 static struct sched_group
*
1924 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
1925 unsigned long *imbalance
, enum idle_type idle
)
1927 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1928 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
1931 max_load
= this_load
= total_load
= total_pwr
= 0;
1932 if (idle
== NOT_IDLE
)
1933 load_idx
= sd
->busy_idx
;
1934 else if (idle
== NEWLY_IDLE
)
1935 load_idx
= sd
->newidle_idx
;
1937 load_idx
= sd
->idle_idx
;
1944 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1946 /* Tally up the load of all CPUs in the group */
1949 for_each_cpu_mask(i
, group
->cpumask
) {
1950 /* Bias balancing toward cpus of our domain */
1952 load
= target_load(i
, load_idx
);
1954 load
= source_load(i
, load_idx
);
1959 total_load
+= avg_load
;
1960 total_pwr
+= group
->cpu_power
;
1962 /* Adjust by relative CPU power of the group */
1963 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
1966 this_load
= avg_load
;
1968 } else if (avg_load
> max_load
) {
1969 max_load
= avg_load
;
1972 group
= group
->next
;
1973 } while (group
!= sd
->groups
);
1975 if (!busiest
|| this_load
>= max_load
)
1978 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
1980 if (this_load
>= avg_load
||
1981 100*max_load
<= sd
->imbalance_pct
*this_load
)
1985 * We're trying to get all the cpus to the average_load, so we don't
1986 * want to push ourselves above the average load, nor do we wish to
1987 * reduce the max loaded cpu below the average load, as either of these
1988 * actions would just result in more rebalancing later, and ping-pong
1989 * tasks around. Thus we look for the minimum possible imbalance.
1990 * Negative imbalances (*we* are more loaded than anyone else) will
1991 * be counted as no imbalance for these purposes -- we can't fix that
1992 * by pulling tasks to us. Be careful of negative numbers as they'll
1993 * appear as very large values with unsigned longs.
1995 /* How much load to actually move to equalise the imbalance */
1996 *imbalance
= min((max_load
- avg_load
) * busiest
->cpu_power
,
1997 (avg_load
- this_load
) * this->cpu_power
)
2000 if (*imbalance
< SCHED_LOAD_SCALE
) {
2001 unsigned long pwr_now
= 0, pwr_move
= 0;
2004 if (max_load
- this_load
>= SCHED_LOAD_SCALE
*2) {
2010 * OK, we don't have enough imbalance to justify moving tasks,
2011 * however we may be able to increase total CPU power used by
2015 pwr_now
+= busiest
->cpu_power
*min(SCHED_LOAD_SCALE
, max_load
);
2016 pwr_now
+= this->cpu_power
*min(SCHED_LOAD_SCALE
, this_load
);
2017 pwr_now
/= SCHED_LOAD_SCALE
;
2019 /* Amount of load we'd subtract */
2020 tmp
= SCHED_LOAD_SCALE
*SCHED_LOAD_SCALE
/busiest
->cpu_power
;
2022 pwr_move
+= busiest
->cpu_power
*min(SCHED_LOAD_SCALE
,
2025 /* Amount of load we'd add */
2026 if (max_load
*busiest
->cpu_power
<
2027 SCHED_LOAD_SCALE
*SCHED_LOAD_SCALE
)
2028 tmp
= max_load
*busiest
->cpu_power
/this->cpu_power
;
2030 tmp
= SCHED_LOAD_SCALE
*SCHED_LOAD_SCALE
/this->cpu_power
;
2031 pwr_move
+= this->cpu_power
*min(SCHED_LOAD_SCALE
, this_load
+ tmp
);
2032 pwr_move
/= SCHED_LOAD_SCALE
;
2034 /* Move if we gain throughput */
2035 if (pwr_move
<= pwr_now
)
2042 /* Get rid of the scaling factor, rounding down as we divide */
2043 *imbalance
= *imbalance
/ SCHED_LOAD_SCALE
;
2053 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2055 static runqueue_t
*find_busiest_queue(struct sched_group
*group
)
2057 unsigned long load
, max_load
= 0;
2058 runqueue_t
*busiest
= NULL
;
2061 for_each_cpu_mask(i
, group
->cpumask
) {
2062 load
= source_load(i
, 0);
2064 if (load
> max_load
) {
2066 busiest
= cpu_rq(i
);
2074 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2075 * tasks if there is an imbalance.
2077 * Called with this_rq unlocked.
2079 static int load_balance(int this_cpu
, runqueue_t
*this_rq
,
2080 struct sched_domain
*sd
, enum idle_type idle
)
2082 struct sched_group
*group
;
2083 runqueue_t
*busiest
;
2084 unsigned long imbalance
;
2085 int nr_moved
, all_pinned
;
2086 int active_balance
= 0;
2088 spin_lock(&this_rq
->lock
);
2089 schedstat_inc(sd
, lb_cnt
[idle
]);
2091 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
);
2093 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2097 busiest
= find_busiest_queue(group
);
2099 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2103 BUG_ON(busiest
== this_rq
);
2105 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2108 if (busiest
->nr_running
> 1) {
2110 * Attempt to move tasks. If find_busiest_group has found
2111 * an imbalance but busiest->nr_running <= 1, the group is
2112 * still unbalanced. nr_moved simply stays zero, so it is
2113 * correctly treated as an imbalance.
2115 double_lock_balance(this_rq
, busiest
);
2116 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2117 imbalance
, sd
, idle
,
2119 spin_unlock(&busiest
->lock
);
2121 /* All tasks on this runqueue were pinned by CPU affinity */
2122 if (unlikely(all_pinned
))
2126 spin_unlock(&this_rq
->lock
);
2129 schedstat_inc(sd
, lb_failed
[idle
]);
2130 sd
->nr_balance_failed
++;
2132 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2134 spin_lock(&busiest
->lock
);
2135 if (!busiest
->active_balance
) {
2136 busiest
->active_balance
= 1;
2137 busiest
->push_cpu
= this_cpu
;
2140 spin_unlock(&busiest
->lock
);
2142 wake_up_process(busiest
->migration_thread
);
2145 * We've kicked active balancing, reset the failure
2148 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2151 sd
->nr_balance_failed
= 0;
2153 if (likely(!active_balance
)) {
2154 /* We were unbalanced, so reset the balancing interval */
2155 sd
->balance_interval
= sd
->min_interval
;
2158 * If we've begun active balancing, start to back off. This
2159 * case may not be covered by the all_pinned logic if there
2160 * is only 1 task on the busy runqueue (because we don't call
2163 if (sd
->balance_interval
< sd
->max_interval
)
2164 sd
->balance_interval
*= 2;
2170 spin_unlock(&this_rq
->lock
);
2172 schedstat_inc(sd
, lb_balanced
[idle
]);
2174 sd
->nr_balance_failed
= 0;
2175 /* tune up the balancing interval */
2176 if (sd
->balance_interval
< sd
->max_interval
)
2177 sd
->balance_interval
*= 2;
2183 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2184 * tasks if there is an imbalance.
2186 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2187 * this_rq is locked.
2189 static int load_balance_newidle(int this_cpu
, runqueue_t
*this_rq
,
2190 struct sched_domain
*sd
)
2192 struct sched_group
*group
;
2193 runqueue_t
*busiest
= NULL
;
2194 unsigned long imbalance
;
2197 schedstat_inc(sd
, lb_cnt
[NEWLY_IDLE
]);
2198 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, NEWLY_IDLE
);
2200 schedstat_inc(sd
, lb_nobusyg
[NEWLY_IDLE
]);
2204 busiest
= find_busiest_queue(group
);
2206 schedstat_inc(sd
, lb_nobusyq
[NEWLY_IDLE
]);
2210 BUG_ON(busiest
== this_rq
);
2212 /* Attempt to move tasks */
2213 double_lock_balance(this_rq
, busiest
);
2215 schedstat_add(sd
, lb_imbalance
[NEWLY_IDLE
], imbalance
);
2216 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2217 imbalance
, sd
, NEWLY_IDLE
, NULL
);
2219 schedstat_inc(sd
, lb_failed
[NEWLY_IDLE
]);
2221 sd
->nr_balance_failed
= 0;
2223 spin_unlock(&busiest
->lock
);
2227 schedstat_inc(sd
, lb_balanced
[NEWLY_IDLE
]);
2228 sd
->nr_balance_failed
= 0;
2233 * idle_balance is called by schedule() if this_cpu is about to become
2234 * idle. Attempts to pull tasks from other CPUs.
2236 static inline void idle_balance(int this_cpu
, runqueue_t
*this_rq
)
2238 struct sched_domain
*sd
;
2240 for_each_domain(this_cpu
, sd
) {
2241 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
2242 if (load_balance_newidle(this_cpu
, this_rq
, sd
)) {
2243 /* We've pulled tasks over so stop searching */
2251 * active_load_balance is run by migration threads. It pushes running tasks
2252 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2253 * running on each physical CPU where possible, and avoids physical /
2254 * logical imbalances.
2256 * Called with busiest_rq locked.
2258 static void active_load_balance(runqueue_t
*busiest_rq
, int busiest_cpu
)
2260 struct sched_domain
*sd
;
2261 runqueue_t
*target_rq
;
2262 int target_cpu
= busiest_rq
->push_cpu
;
2264 if (busiest_rq
->nr_running
<= 1)
2265 /* no task to move */
2268 target_rq
= cpu_rq(target_cpu
);
2271 * This condition is "impossible", if it occurs
2272 * we need to fix it. Originally reported by
2273 * Bjorn Helgaas on a 128-cpu setup.
2275 BUG_ON(busiest_rq
== target_rq
);
2277 /* move a task from busiest_rq to target_rq */
2278 double_lock_balance(busiest_rq
, target_rq
);
2280 /* Search for an sd spanning us and the target CPU. */
2281 for_each_domain(target_cpu
, sd
)
2282 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2283 cpu_isset(busiest_cpu
, sd
->span
))
2286 if (unlikely(sd
== NULL
))
2289 schedstat_inc(sd
, alb_cnt
);
2291 if (move_tasks(target_rq
, target_cpu
, busiest_rq
, 1, sd
, SCHED_IDLE
, NULL
))
2292 schedstat_inc(sd
, alb_pushed
);
2294 schedstat_inc(sd
, alb_failed
);
2296 spin_unlock(&target_rq
->lock
);
2300 * rebalance_tick will get called every timer tick, on every CPU.
2302 * It checks each scheduling domain to see if it is due to be balanced,
2303 * and initiates a balancing operation if so.
2305 * Balancing parameters are set up in arch_init_sched_domains.
2308 /* Don't have all balancing operations going off at once */
2309 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2311 static void rebalance_tick(int this_cpu
, runqueue_t
*this_rq
,
2312 enum idle_type idle
)
2314 unsigned long old_load
, this_load
;
2315 unsigned long j
= jiffies
+ CPU_OFFSET(this_cpu
);
2316 struct sched_domain
*sd
;
2319 this_load
= this_rq
->nr_running
* SCHED_LOAD_SCALE
;
2320 /* Update our load */
2321 for (i
= 0; i
< 3; i
++) {
2322 unsigned long new_load
= this_load
;
2324 old_load
= this_rq
->cpu_load
[i
];
2326 * Round up the averaging division if load is increasing. This
2327 * prevents us from getting stuck on 9 if the load is 10, for
2330 if (new_load
> old_load
)
2331 new_load
+= scale
-1;
2332 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) / scale
;
2335 for_each_domain(this_cpu
, sd
) {
2336 unsigned long interval
;
2338 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2341 interval
= sd
->balance_interval
;
2342 if (idle
!= SCHED_IDLE
)
2343 interval
*= sd
->busy_factor
;
2345 /* scale ms to jiffies */
2346 interval
= msecs_to_jiffies(interval
);
2347 if (unlikely(!interval
))
2350 if (j
- sd
->last_balance
>= interval
) {
2351 if (load_balance(this_cpu
, this_rq
, sd
, idle
)) {
2352 /* We've pulled tasks over so no longer idle */
2355 sd
->last_balance
+= interval
;
2361 * on UP we do not need to balance between CPUs:
2363 static inline void rebalance_tick(int cpu
, runqueue_t
*rq
, enum idle_type idle
)
2366 static inline void idle_balance(int cpu
, runqueue_t
*rq
)
2371 static inline int wake_priority_sleeper(runqueue_t
*rq
)
2374 #ifdef CONFIG_SCHED_SMT
2375 spin_lock(&rq
->lock
);
2377 * If an SMT sibling task has been put to sleep for priority
2378 * reasons reschedule the idle task to see if it can now run.
2380 if (rq
->nr_running
) {
2381 resched_task(rq
->idle
);
2384 spin_unlock(&rq
->lock
);
2389 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2391 EXPORT_PER_CPU_SYMBOL(kstat
);
2394 * This is called on clock ticks and on context switches.
2395 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2397 static inline void update_cpu_clock(task_t
*p
, runqueue_t
*rq
,
2398 unsigned long long now
)
2400 unsigned long long last
= max(p
->timestamp
, rq
->timestamp_last_tick
);
2401 p
->sched_time
+= now
- last
;
2405 * Return current->sched_time plus any more ns on the sched_clock
2406 * that have not yet been banked.
2408 unsigned long long current_sched_time(const task_t
*tsk
)
2410 unsigned long long ns
;
2411 unsigned long flags
;
2412 local_irq_save(flags
);
2413 ns
= max(tsk
->timestamp
, task_rq(tsk
)->timestamp_last_tick
);
2414 ns
= tsk
->sched_time
+ (sched_clock() - ns
);
2415 local_irq_restore(flags
);
2420 * We place interactive tasks back into the active array, if possible.
2422 * To guarantee that this does not starve expired tasks we ignore the
2423 * interactivity of a task if the first expired task had to wait more
2424 * than a 'reasonable' amount of time. This deadline timeout is
2425 * load-dependent, as the frequency of array switched decreases with
2426 * increasing number of running tasks. We also ignore the interactivity
2427 * if a better static_prio task has expired:
2429 #define EXPIRED_STARVING(rq) \
2430 ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2431 (jiffies - (rq)->expired_timestamp >= \
2432 STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2433 ((rq)->curr->static_prio > (rq)->best_expired_prio))
2436 * Account user cpu time to a process.
2437 * @p: the process that the cpu time gets accounted to
2438 * @hardirq_offset: the offset to subtract from hardirq_count()
2439 * @cputime: the cpu time spent in user space since the last update
2441 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
2443 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2446 p
->utime
= cputime_add(p
->utime
, cputime
);
2448 /* Add user time to cpustat. */
2449 tmp
= cputime_to_cputime64(cputime
);
2450 if (TASK_NICE(p
) > 0)
2451 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
2453 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
2457 * Account system cpu time to a process.
2458 * @p: the process that the cpu time gets accounted to
2459 * @hardirq_offset: the offset to subtract from hardirq_count()
2460 * @cputime: the cpu time spent in kernel space since the last update
2462 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
2465 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2466 runqueue_t
*rq
= this_rq();
2469 p
->stime
= cputime_add(p
->stime
, cputime
);
2471 /* Add system time to cpustat. */
2472 tmp
= cputime_to_cputime64(cputime
);
2473 if (hardirq_count() - hardirq_offset
)
2474 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
2475 else if (softirq_count())
2476 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
2477 else if (p
!= rq
->idle
)
2478 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
2479 else if (atomic_read(&rq
->nr_iowait
) > 0)
2480 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
2482 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
2483 /* Account for system time used */
2484 acct_update_integrals(p
);
2485 /* Update rss highwater mark */
2486 update_mem_hiwater(p
);
2490 * Account for involuntary wait time.
2491 * @p: the process from which the cpu time has been stolen
2492 * @steal: the cpu time spent in involuntary wait
2494 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
2496 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2497 cputime64_t tmp
= cputime_to_cputime64(steal
);
2498 runqueue_t
*rq
= this_rq();
2500 if (p
== rq
->idle
) {
2501 p
->stime
= cputime_add(p
->stime
, steal
);
2502 if (atomic_read(&rq
->nr_iowait
) > 0)
2503 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
2505 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
2507 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
2511 * This function gets called by the timer code, with HZ frequency.
2512 * We call it with interrupts disabled.
2514 * It also gets called by the fork code, when changing the parent's
2517 void scheduler_tick(void)
2519 int cpu
= smp_processor_id();
2520 runqueue_t
*rq
= this_rq();
2521 task_t
*p
= current
;
2522 unsigned long long now
= sched_clock();
2524 update_cpu_clock(p
, rq
, now
);
2526 rq
->timestamp_last_tick
= now
;
2528 if (p
== rq
->idle
) {
2529 if (wake_priority_sleeper(rq
))
2531 rebalance_tick(cpu
, rq
, SCHED_IDLE
);
2535 /* Task might have expired already, but not scheduled off yet */
2536 if (p
->array
!= rq
->active
) {
2537 set_tsk_need_resched(p
);
2540 spin_lock(&rq
->lock
);
2542 * The task was running during this tick - update the
2543 * time slice counter. Note: we do not update a thread's
2544 * priority until it either goes to sleep or uses up its
2545 * timeslice. This makes it possible for interactive tasks
2546 * to use up their timeslices at their highest priority levels.
2550 * RR tasks need a special form of timeslice management.
2551 * FIFO tasks have no timeslices.
2553 if ((p
->policy
== SCHED_RR
) && !--p
->time_slice
) {
2554 p
->time_slice
= task_timeslice(p
);
2555 p
->first_time_slice
= 0;
2556 set_tsk_need_resched(p
);
2558 /* put it at the end of the queue: */
2559 requeue_task(p
, rq
->active
);
2563 if (!--p
->time_slice
) {
2564 dequeue_task(p
, rq
->active
);
2565 set_tsk_need_resched(p
);
2566 p
->prio
= effective_prio(p
);
2567 p
->time_slice
= task_timeslice(p
);
2568 p
->first_time_slice
= 0;
2570 if (!rq
->expired_timestamp
)
2571 rq
->expired_timestamp
= jiffies
;
2572 if (!TASK_INTERACTIVE(p
) || EXPIRED_STARVING(rq
)) {
2573 enqueue_task(p
, rq
->expired
);
2574 if (p
->static_prio
< rq
->best_expired_prio
)
2575 rq
->best_expired_prio
= p
->static_prio
;
2577 enqueue_task(p
, rq
->active
);
2580 * Prevent a too long timeslice allowing a task to monopolize
2581 * the CPU. We do this by splitting up the timeslice into
2584 * Note: this does not mean the task's timeslices expire or
2585 * get lost in any way, they just might be preempted by
2586 * another task of equal priority. (one with higher
2587 * priority would have preempted this task already.) We
2588 * requeue this task to the end of the list on this priority
2589 * level, which is in essence a round-robin of tasks with
2592 * This only applies to tasks in the interactive
2593 * delta range with at least TIMESLICE_GRANULARITY to requeue.
2595 if (TASK_INTERACTIVE(p
) && !((task_timeslice(p
) -
2596 p
->time_slice
) % TIMESLICE_GRANULARITY(p
)) &&
2597 (p
->time_slice
>= TIMESLICE_GRANULARITY(p
)) &&
2598 (p
->array
== rq
->active
)) {
2600 requeue_task(p
, rq
->active
);
2601 set_tsk_need_resched(p
);
2605 spin_unlock(&rq
->lock
);
2607 rebalance_tick(cpu
, rq
, NOT_IDLE
);
2610 #ifdef CONFIG_SCHED_SMT
2611 static inline void wake_sleeping_dependent(int this_cpu
, runqueue_t
*this_rq
)
2613 struct sched_domain
*tmp
, *sd
= NULL
;
2614 cpumask_t sibling_map
;
2617 for_each_domain(this_cpu
, tmp
)
2618 if (tmp
->flags
& SD_SHARE_CPUPOWER
)
2625 * Unlock the current runqueue because we have to lock in
2626 * CPU order to avoid deadlocks. Caller knows that we might
2627 * unlock. We keep IRQs disabled.
2629 spin_unlock(&this_rq
->lock
);
2631 sibling_map
= sd
->span
;
2633 for_each_cpu_mask(i
, sibling_map
)
2634 spin_lock(&cpu_rq(i
)->lock
);
2636 * We clear this CPU from the mask. This both simplifies the
2637 * inner loop and keps this_rq locked when we exit:
2639 cpu_clear(this_cpu
, sibling_map
);
2641 for_each_cpu_mask(i
, sibling_map
) {
2642 runqueue_t
*smt_rq
= cpu_rq(i
);
2645 * If an SMT sibling task is sleeping due to priority
2646 * reasons wake it up now.
2648 if (smt_rq
->curr
== smt_rq
->idle
&& smt_rq
->nr_running
)
2649 resched_task(smt_rq
->idle
);
2652 for_each_cpu_mask(i
, sibling_map
)
2653 spin_unlock(&cpu_rq(i
)->lock
);
2655 * We exit with this_cpu's rq still held and IRQs
2660 static inline int dependent_sleeper(int this_cpu
, runqueue_t
*this_rq
)
2662 struct sched_domain
*tmp
, *sd
= NULL
;
2663 cpumask_t sibling_map
;
2664 prio_array_t
*array
;
2668 for_each_domain(this_cpu
, tmp
)
2669 if (tmp
->flags
& SD_SHARE_CPUPOWER
)
2676 * The same locking rules and details apply as for
2677 * wake_sleeping_dependent():
2679 spin_unlock(&this_rq
->lock
);
2680 sibling_map
= sd
->span
;
2681 for_each_cpu_mask(i
, sibling_map
)
2682 spin_lock(&cpu_rq(i
)->lock
);
2683 cpu_clear(this_cpu
, sibling_map
);
2686 * Establish next task to be run - it might have gone away because
2687 * we released the runqueue lock above:
2689 if (!this_rq
->nr_running
)
2691 array
= this_rq
->active
;
2692 if (!array
->nr_active
)
2693 array
= this_rq
->expired
;
2694 BUG_ON(!array
->nr_active
);
2696 p
= list_entry(array
->queue
[sched_find_first_bit(array
->bitmap
)].next
,
2699 for_each_cpu_mask(i
, sibling_map
) {
2700 runqueue_t
*smt_rq
= cpu_rq(i
);
2701 task_t
*smt_curr
= smt_rq
->curr
;
2704 * If a user task with lower static priority than the
2705 * running task on the SMT sibling is trying to schedule,
2706 * delay it till there is proportionately less timeslice
2707 * left of the sibling task to prevent a lower priority
2708 * task from using an unfair proportion of the
2709 * physical cpu's resources. -ck
2711 if (((smt_curr
->time_slice
* (100 - sd
->per_cpu_gain
) / 100) >
2712 task_timeslice(p
) || rt_task(smt_curr
)) &&
2713 p
->mm
&& smt_curr
->mm
&& !rt_task(p
))
2717 * Reschedule a lower priority task on the SMT sibling,
2718 * or wake it up if it has been put to sleep for priority
2721 if ((((p
->time_slice
* (100 - sd
->per_cpu_gain
) / 100) >
2722 task_timeslice(smt_curr
) || rt_task(p
)) &&
2723 smt_curr
->mm
&& p
->mm
&& !rt_task(smt_curr
)) ||
2724 (smt_curr
== smt_rq
->idle
&& smt_rq
->nr_running
))
2725 resched_task(smt_curr
);
2728 for_each_cpu_mask(i
, sibling_map
)
2729 spin_unlock(&cpu_rq(i
)->lock
);
2733 static inline void wake_sleeping_dependent(int this_cpu
, runqueue_t
*this_rq
)
2737 static inline int dependent_sleeper(int this_cpu
, runqueue_t
*this_rq
)
2743 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
2745 void fastcall
add_preempt_count(int val
)
2750 BUG_ON((preempt_count() < 0));
2751 preempt_count() += val
;
2753 * Spinlock count overflowing soon?
2755 BUG_ON((preempt_count() & PREEMPT_MASK
) >= PREEMPT_MASK
-10);
2757 EXPORT_SYMBOL(add_preempt_count
);
2759 void fastcall
sub_preempt_count(int val
)
2764 BUG_ON(val
> preempt_count());
2766 * Is the spinlock portion underflowing?
2768 BUG_ON((val
< PREEMPT_MASK
) && !(preempt_count() & PREEMPT_MASK
));
2769 preempt_count() -= val
;
2771 EXPORT_SYMBOL(sub_preempt_count
);
2776 * schedule() is the main scheduler function.
2778 asmlinkage
void __sched
schedule(void)
2781 task_t
*prev
, *next
;
2783 prio_array_t
*array
;
2784 struct list_head
*queue
;
2785 unsigned long long now
;
2786 unsigned long run_time
;
2790 * Test if we are atomic. Since do_exit() needs to call into
2791 * schedule() atomically, we ignore that path for now.
2792 * Otherwise, whine if we are scheduling when we should not be.
2794 if (likely(!current
->exit_state
)) {
2795 if (unlikely(in_atomic())) {
2796 printk(KERN_ERR
"scheduling while atomic: "
2798 current
->comm
, preempt_count(), current
->pid
);
2802 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2807 release_kernel_lock(prev
);
2808 need_resched_nonpreemptible
:
2812 * The idle thread is not allowed to schedule!
2813 * Remove this check after it has been exercised a bit.
2815 if (unlikely(prev
== rq
->idle
) && prev
->state
!= TASK_RUNNING
) {
2816 printk(KERN_ERR
"bad: scheduling from the idle thread!\n");
2820 schedstat_inc(rq
, sched_cnt
);
2821 now
= sched_clock();
2822 if (likely((long long)(now
- prev
->timestamp
) < NS_MAX_SLEEP_AVG
)) {
2823 run_time
= now
- prev
->timestamp
;
2824 if (unlikely((long long)(now
- prev
->timestamp
) < 0))
2827 run_time
= NS_MAX_SLEEP_AVG
;
2830 * Tasks charged proportionately less run_time at high sleep_avg to
2831 * delay them losing their interactive status
2833 run_time
/= (CURRENT_BONUS(prev
) ? : 1);
2835 spin_lock_irq(&rq
->lock
);
2837 if (unlikely(prev
->flags
& PF_DEAD
))
2838 prev
->state
= EXIT_DEAD
;
2840 switch_count
= &prev
->nivcsw
;
2841 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
2842 switch_count
= &prev
->nvcsw
;
2843 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
2844 unlikely(signal_pending(prev
))))
2845 prev
->state
= TASK_RUNNING
;
2847 if (prev
->state
== TASK_UNINTERRUPTIBLE
)
2848 rq
->nr_uninterruptible
++;
2849 deactivate_task(prev
, rq
);
2853 cpu
= smp_processor_id();
2854 if (unlikely(!rq
->nr_running
)) {
2856 idle_balance(cpu
, rq
);
2857 if (!rq
->nr_running
) {
2859 rq
->expired_timestamp
= 0;
2860 wake_sleeping_dependent(cpu
, rq
);
2862 * wake_sleeping_dependent() might have released
2863 * the runqueue, so break out if we got new
2866 if (!rq
->nr_running
)
2870 if (dependent_sleeper(cpu
, rq
)) {
2875 * dependent_sleeper() releases and reacquires the runqueue
2876 * lock, hence go into the idle loop if the rq went
2879 if (unlikely(!rq
->nr_running
))
2884 if (unlikely(!array
->nr_active
)) {
2886 * Switch the active and expired arrays.
2888 schedstat_inc(rq
, sched_switch
);
2889 rq
->active
= rq
->expired
;
2890 rq
->expired
= array
;
2892 rq
->expired_timestamp
= 0;
2893 rq
->best_expired_prio
= MAX_PRIO
;
2896 idx
= sched_find_first_bit(array
->bitmap
);
2897 queue
= array
->queue
+ idx
;
2898 next
= list_entry(queue
->next
, task_t
, run_list
);
2900 if (!rt_task(next
) && next
->activated
> 0) {
2901 unsigned long long delta
= now
- next
->timestamp
;
2902 if (unlikely((long long)(now
- next
->timestamp
) < 0))
2905 if (next
->activated
== 1)
2906 delta
= delta
* (ON_RUNQUEUE_WEIGHT
* 128 / 100) / 128;
2908 array
= next
->array
;
2909 dequeue_task(next
, array
);
2910 recalc_task_prio(next
, next
->timestamp
+ delta
);
2911 enqueue_task(next
, array
);
2913 next
->activated
= 0;
2915 if (next
== rq
->idle
)
2916 schedstat_inc(rq
, sched_goidle
);
2918 clear_tsk_need_resched(prev
);
2919 rcu_qsctr_inc(task_cpu(prev
));
2921 update_cpu_clock(prev
, rq
, now
);
2923 prev
->sleep_avg
-= run_time
;
2924 if ((long)prev
->sleep_avg
<= 0)
2925 prev
->sleep_avg
= 0;
2926 prev
->timestamp
= prev
->last_ran
= now
;
2928 sched_info_switch(prev
, next
);
2929 if (likely(prev
!= next
)) {
2930 next
->timestamp
= now
;
2935 prepare_task_switch(rq
, next
);
2936 prev
= context_switch(rq
, prev
, next
);
2939 * this_rq must be evaluated again because prev may have moved
2940 * CPUs since it called schedule(), thus the 'rq' on its stack
2941 * frame will be invalid.
2943 finish_task_switch(this_rq(), prev
);
2945 spin_unlock_irq(&rq
->lock
);
2948 if (unlikely(reacquire_kernel_lock(prev
) < 0))
2949 goto need_resched_nonpreemptible
;
2950 preempt_enable_no_resched();
2951 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
2955 EXPORT_SYMBOL(schedule
);
2957 #ifdef CONFIG_PREEMPT
2959 * this is is the entry point to schedule() from in-kernel preemption
2960 * off of preempt_enable. Kernel preemptions off return from interrupt
2961 * occur there and call schedule directly.
2963 asmlinkage
void __sched
preempt_schedule(void)
2965 struct thread_info
*ti
= current_thread_info();
2966 #ifdef CONFIG_PREEMPT_BKL
2967 struct task_struct
*task
= current
;
2968 int saved_lock_depth
;
2971 * If there is a non-zero preempt_count or interrupts are disabled,
2972 * we do not want to preempt the current task. Just return..
2974 if (unlikely(ti
->preempt_count
|| irqs_disabled()))
2978 add_preempt_count(PREEMPT_ACTIVE
);
2980 * We keep the big kernel semaphore locked, but we
2981 * clear ->lock_depth so that schedule() doesnt
2982 * auto-release the semaphore:
2984 #ifdef CONFIG_PREEMPT_BKL
2985 saved_lock_depth
= task
->lock_depth
;
2986 task
->lock_depth
= -1;
2989 #ifdef CONFIG_PREEMPT_BKL
2990 task
->lock_depth
= saved_lock_depth
;
2992 sub_preempt_count(PREEMPT_ACTIVE
);
2994 /* we could miss a preemption opportunity between schedule and now */
2996 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3000 EXPORT_SYMBOL(preempt_schedule
);
3003 * this is is the entry point to schedule() from kernel preemption
3004 * off of irq context.
3005 * Note, that this is called and return with irqs disabled. This will
3006 * protect us against recursive calling from irq.
3008 asmlinkage
void __sched
preempt_schedule_irq(void)
3010 struct thread_info
*ti
= current_thread_info();
3011 #ifdef CONFIG_PREEMPT_BKL
3012 struct task_struct
*task
= current
;
3013 int saved_lock_depth
;
3015 /* Catch callers which need to be fixed*/
3016 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3019 add_preempt_count(PREEMPT_ACTIVE
);
3021 * We keep the big kernel semaphore locked, but we
3022 * clear ->lock_depth so that schedule() doesnt
3023 * auto-release the semaphore:
3025 #ifdef CONFIG_PREEMPT_BKL
3026 saved_lock_depth
= task
->lock_depth
;
3027 task
->lock_depth
= -1;
3031 local_irq_disable();
3032 #ifdef CONFIG_PREEMPT_BKL
3033 task
->lock_depth
= saved_lock_depth
;
3035 sub_preempt_count(PREEMPT_ACTIVE
);
3037 /* we could miss a preemption opportunity between schedule and now */
3039 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3043 #endif /* CONFIG_PREEMPT */
3045 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
, void *key
)
3047 task_t
*p
= curr
->private;
3048 return try_to_wake_up(p
, mode
, sync
);
3051 EXPORT_SYMBOL(default_wake_function
);
3054 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3055 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3056 * number) then we wake all the non-exclusive tasks and one exclusive task.
3058 * There are circumstances in which we can try to wake a task which has already
3059 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3060 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3062 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3063 int nr_exclusive
, int sync
, void *key
)
3065 struct list_head
*tmp
, *next
;
3067 list_for_each_safe(tmp
, next
, &q
->task_list
) {
3070 curr
= list_entry(tmp
, wait_queue_t
, task_list
);
3071 flags
= curr
->flags
;
3072 if (curr
->func(curr
, mode
, sync
, key
) &&
3073 (flags
& WQ_FLAG_EXCLUSIVE
) &&
3080 * __wake_up - wake up threads blocked on a waitqueue.
3082 * @mode: which threads
3083 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3084 * @key: is directly passed to the wakeup function
3086 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3087 int nr_exclusive
, void *key
)
3089 unsigned long flags
;
3091 spin_lock_irqsave(&q
->lock
, flags
);
3092 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3093 spin_unlock_irqrestore(&q
->lock
, flags
);
3096 EXPORT_SYMBOL(__wake_up
);
3099 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3101 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3103 __wake_up_common(q
, mode
, 1, 0, NULL
);
3107 * __wake_up_sync - wake up threads blocked on a waitqueue.
3109 * @mode: which threads
3110 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3112 * The sync wakeup differs that the waker knows that it will schedule
3113 * away soon, so while the target thread will be woken up, it will not
3114 * be migrated to another CPU - ie. the two threads are 'synchronized'
3115 * with each other. This can prevent needless bouncing between CPUs.
3117 * On UP it can prevent extra preemption.
3119 void fastcall
__wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3121 unsigned long flags
;
3127 if (unlikely(!nr_exclusive
))
3130 spin_lock_irqsave(&q
->lock
, flags
);
3131 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3132 spin_unlock_irqrestore(&q
->lock
, flags
);
3134 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3136 void fastcall
complete(struct completion
*x
)
3138 unsigned long flags
;
3140 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3142 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3144 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3146 EXPORT_SYMBOL(complete
);
3148 void fastcall
complete_all(struct completion
*x
)
3150 unsigned long flags
;
3152 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3153 x
->done
+= UINT_MAX
/2;
3154 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3156 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3158 EXPORT_SYMBOL(complete_all
);
3160 void fastcall __sched
wait_for_completion(struct completion
*x
)
3163 spin_lock_irq(&x
->wait
.lock
);
3165 DECLARE_WAITQUEUE(wait
, current
);
3167 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3168 __add_wait_queue_tail(&x
->wait
, &wait
);
3170 __set_current_state(TASK_UNINTERRUPTIBLE
);
3171 spin_unlock_irq(&x
->wait
.lock
);
3173 spin_lock_irq(&x
->wait
.lock
);
3175 __remove_wait_queue(&x
->wait
, &wait
);
3178 spin_unlock_irq(&x
->wait
.lock
);
3180 EXPORT_SYMBOL(wait_for_completion
);
3182 unsigned long fastcall __sched
3183 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3187 spin_lock_irq(&x
->wait
.lock
);
3189 DECLARE_WAITQUEUE(wait
, current
);
3191 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3192 __add_wait_queue_tail(&x
->wait
, &wait
);
3194 __set_current_state(TASK_UNINTERRUPTIBLE
);
3195 spin_unlock_irq(&x
->wait
.lock
);
3196 timeout
= schedule_timeout(timeout
);
3197 spin_lock_irq(&x
->wait
.lock
);
3199 __remove_wait_queue(&x
->wait
, &wait
);
3203 __remove_wait_queue(&x
->wait
, &wait
);
3207 spin_unlock_irq(&x
->wait
.lock
);
3210 EXPORT_SYMBOL(wait_for_completion_timeout
);
3212 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3218 spin_lock_irq(&x
->wait
.lock
);
3220 DECLARE_WAITQUEUE(wait
, current
);
3222 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3223 __add_wait_queue_tail(&x
->wait
, &wait
);
3225 if (signal_pending(current
)) {
3227 __remove_wait_queue(&x
->wait
, &wait
);
3230 __set_current_state(TASK_INTERRUPTIBLE
);
3231 spin_unlock_irq(&x
->wait
.lock
);
3233 spin_lock_irq(&x
->wait
.lock
);
3235 __remove_wait_queue(&x
->wait
, &wait
);
3239 spin_unlock_irq(&x
->wait
.lock
);
3243 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3245 unsigned long fastcall __sched
3246 wait_for_completion_interruptible_timeout(struct completion
*x
,
3247 unsigned long timeout
)
3251 spin_lock_irq(&x
->wait
.lock
);
3253 DECLARE_WAITQUEUE(wait
, current
);
3255 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3256 __add_wait_queue_tail(&x
->wait
, &wait
);
3258 if (signal_pending(current
)) {
3259 timeout
= -ERESTARTSYS
;
3260 __remove_wait_queue(&x
->wait
, &wait
);
3263 __set_current_state(TASK_INTERRUPTIBLE
);
3264 spin_unlock_irq(&x
->wait
.lock
);
3265 timeout
= schedule_timeout(timeout
);
3266 spin_lock_irq(&x
->wait
.lock
);
3268 __remove_wait_queue(&x
->wait
, &wait
);
3272 __remove_wait_queue(&x
->wait
, &wait
);
3276 spin_unlock_irq(&x
->wait
.lock
);
3279 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3282 #define SLEEP_ON_VAR \
3283 unsigned long flags; \
3284 wait_queue_t wait; \
3285 init_waitqueue_entry(&wait, current);
3287 #define SLEEP_ON_HEAD \
3288 spin_lock_irqsave(&q->lock,flags); \
3289 __add_wait_queue(q, &wait); \
3290 spin_unlock(&q->lock);
3292 #define SLEEP_ON_TAIL \
3293 spin_lock_irq(&q->lock); \
3294 __remove_wait_queue(q, &wait); \
3295 spin_unlock_irqrestore(&q->lock, flags);
3297 void fastcall __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3301 current
->state
= TASK_INTERRUPTIBLE
;
3308 EXPORT_SYMBOL(interruptible_sleep_on
);
3310 long fastcall __sched
interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3314 current
->state
= TASK_INTERRUPTIBLE
;
3317 timeout
= schedule_timeout(timeout
);
3323 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3325 void fastcall __sched
sleep_on(wait_queue_head_t
*q
)
3329 current
->state
= TASK_UNINTERRUPTIBLE
;
3336 EXPORT_SYMBOL(sleep_on
);
3338 long fastcall __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3342 current
->state
= TASK_UNINTERRUPTIBLE
;
3345 timeout
= schedule_timeout(timeout
);
3351 EXPORT_SYMBOL(sleep_on_timeout
);
3353 void set_user_nice(task_t
*p
, long nice
)
3355 unsigned long flags
;
3356 prio_array_t
*array
;
3358 int old_prio
, new_prio
, delta
;
3360 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3363 * We have to be careful, if called from sys_setpriority(),
3364 * the task might be in the middle of scheduling on another CPU.
3366 rq
= task_rq_lock(p
, &flags
);
3368 * The RT priorities are set via sched_setscheduler(), but we still
3369 * allow the 'normal' nice value to be set - but as expected
3370 * it wont have any effect on scheduling until the task is
3374 p
->static_prio
= NICE_TO_PRIO(nice
);
3379 dequeue_task(p
, array
);
3382 new_prio
= NICE_TO_PRIO(nice
);
3383 delta
= new_prio
- old_prio
;
3384 p
->static_prio
= NICE_TO_PRIO(nice
);
3388 enqueue_task(p
, array
);
3390 * If the task increased its priority or is running and
3391 * lowered its priority, then reschedule its CPU:
3393 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3394 resched_task(rq
->curr
);
3397 task_rq_unlock(rq
, &flags
);
3400 EXPORT_SYMBOL(set_user_nice
);
3403 * can_nice - check if a task can reduce its nice value
3407 int can_nice(const task_t
*p
, const int nice
)
3409 /* convert nice value [19,-20] to rlimit style value [0,39] */
3410 int nice_rlim
= 19 - nice
;
3411 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
3412 capable(CAP_SYS_NICE
));
3415 #ifdef __ARCH_WANT_SYS_NICE
3418 * sys_nice - change the priority of the current process.
3419 * @increment: priority increment
3421 * sys_setpriority is a more generic, but much slower function that
3422 * does similar things.
3424 asmlinkage
long sys_nice(int increment
)
3430 * Setpriority might change our priority at the same moment.
3431 * We don't have to worry. Conceptually one call occurs first
3432 * and we have a single winner.
3434 if (increment
< -40)
3439 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
3445 if (increment
< 0 && !can_nice(current
, nice
))
3448 retval
= security_task_setnice(current
, nice
);
3452 set_user_nice(current
, nice
);
3459 * task_prio - return the priority value of a given task.
3460 * @p: the task in question.
3462 * This is the priority value as seen by users in /proc.
3463 * RT tasks are offset by -200. Normal tasks are centered
3464 * around 0, value goes from -16 to +15.
3466 int task_prio(const task_t
*p
)
3468 return p
->prio
- MAX_RT_PRIO
;
3472 * task_nice - return the nice value of a given task.
3473 * @p: the task in question.
3475 int task_nice(const task_t
*p
)
3477 return TASK_NICE(p
);
3481 * The only users of task_nice are binfmt_elf and binfmt_elf32.
3482 * binfmt_elf is no longer modular, but binfmt_elf32 still is.
3483 * Therefore, task_nice is needed if there is a compat_mode.
3485 #ifdef CONFIG_COMPAT
3486 EXPORT_SYMBOL_GPL(task_nice
);
3490 * idle_cpu - is a given cpu idle currently?
3491 * @cpu: the processor in question.
3493 int idle_cpu(int cpu
)
3495 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
3498 EXPORT_SYMBOL_GPL(idle_cpu
);
3501 * idle_task - return the idle task for a given cpu.
3502 * @cpu: the processor in question.
3504 task_t
*idle_task(int cpu
)
3506 return cpu_rq(cpu
)->idle
;
3510 * find_process_by_pid - find a process with a matching PID value.
3511 * @pid: the pid in question.
3513 static inline task_t
*find_process_by_pid(pid_t pid
)
3515 return pid
? find_task_by_pid(pid
) : current
;
3518 /* Actually do priority change: must hold rq lock. */
3519 static void __setscheduler(struct task_struct
*p
, int policy
, int prio
)
3523 p
->rt_priority
= prio
;
3524 if (policy
!= SCHED_NORMAL
)
3525 p
->prio
= MAX_USER_RT_PRIO
-1 - p
->rt_priority
;
3527 p
->prio
= p
->static_prio
;
3531 * sched_setscheduler - change the scheduling policy and/or RT priority of
3533 * @p: the task in question.
3534 * @policy: new policy.
3535 * @param: structure containing the new RT priority.
3537 int sched_setscheduler(struct task_struct
*p
, int policy
, struct sched_param
*param
)
3540 int oldprio
, oldpolicy
= -1;
3541 prio_array_t
*array
;
3542 unsigned long flags
;
3546 /* double check policy once rq lock held */
3548 policy
= oldpolicy
= p
->policy
;
3549 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3550 policy
!= SCHED_NORMAL
)
3553 * Valid priorities for SCHED_FIFO and SCHED_RR are
3554 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL is 0.
3556 if (param
->sched_priority
< 0 ||
3557 param
->sched_priority
> MAX_USER_RT_PRIO
-1)
3559 if ((policy
== SCHED_NORMAL
) != (param
->sched_priority
== 0))
3562 if ((policy
== SCHED_FIFO
|| policy
== SCHED_RR
) &&
3563 param
->sched_priority
> p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
&&
3564 !capable(CAP_SYS_NICE
))
3566 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
3567 !capable(CAP_SYS_NICE
))
3570 retval
= security_task_setscheduler(p
, policy
, param
);
3574 * To be able to change p->policy safely, the apropriate
3575 * runqueue lock must be held.
3577 rq
= task_rq_lock(p
, &flags
);
3578 /* recheck policy now with rq lock held */
3579 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3580 policy
= oldpolicy
= -1;
3581 task_rq_unlock(rq
, &flags
);
3586 deactivate_task(p
, rq
);
3588 __setscheduler(p
, policy
, param
->sched_priority
);
3590 __activate_task(p
, rq
);
3592 * Reschedule if we are currently running on this runqueue and
3593 * our priority decreased, or if we are not currently running on
3594 * this runqueue and our priority is higher than the current's
3596 if (task_running(rq
, p
)) {
3597 if (p
->prio
> oldprio
)
3598 resched_task(rq
->curr
);
3599 } else if (TASK_PREEMPTS_CURR(p
, rq
))
3600 resched_task(rq
->curr
);
3602 task_rq_unlock(rq
, &flags
);
3605 EXPORT_SYMBOL_GPL(sched_setscheduler
);
3607 static int do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
3610 struct sched_param lparam
;
3611 struct task_struct
*p
;
3613 if (!param
|| pid
< 0)
3615 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
3617 read_lock_irq(&tasklist_lock
);
3618 p
= find_process_by_pid(pid
);
3620 read_unlock_irq(&tasklist_lock
);
3623 retval
= sched_setscheduler(p
, policy
, &lparam
);
3624 read_unlock_irq(&tasklist_lock
);
3629 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3630 * @pid: the pid in question.
3631 * @policy: new policy.
3632 * @param: structure containing the new RT priority.
3634 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
3635 struct sched_param __user
*param
)
3637 return do_sched_setscheduler(pid
, policy
, param
);
3641 * sys_sched_setparam - set/change the RT priority of a thread
3642 * @pid: the pid in question.
3643 * @param: structure containing the new RT priority.
3645 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
3647 return do_sched_setscheduler(pid
, -1, param
);
3651 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3652 * @pid: the pid in question.
3654 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
3656 int retval
= -EINVAL
;
3663 read_lock(&tasklist_lock
);
3664 p
= find_process_by_pid(pid
);
3666 retval
= security_task_getscheduler(p
);
3670 read_unlock(&tasklist_lock
);
3677 * sys_sched_getscheduler - get the RT priority of a thread
3678 * @pid: the pid in question.
3679 * @param: structure containing the RT priority.
3681 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
3683 struct sched_param lp
;
3684 int retval
= -EINVAL
;
3687 if (!param
|| pid
< 0)
3690 read_lock(&tasklist_lock
);
3691 p
= find_process_by_pid(pid
);
3696 retval
= security_task_getscheduler(p
);
3700 lp
.sched_priority
= p
->rt_priority
;
3701 read_unlock(&tasklist_lock
);
3704 * This one might sleep, we cannot do it with a spinlock held ...
3706 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
3712 read_unlock(&tasklist_lock
);
3716 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
3720 cpumask_t cpus_allowed
;
3723 read_lock(&tasklist_lock
);
3725 p
= find_process_by_pid(pid
);
3727 read_unlock(&tasklist_lock
);
3728 unlock_cpu_hotplug();
3733 * It is not safe to call set_cpus_allowed with the
3734 * tasklist_lock held. We will bump the task_struct's
3735 * usage count and then drop tasklist_lock.
3738 read_unlock(&tasklist_lock
);
3741 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
3742 !capable(CAP_SYS_NICE
))
3745 cpus_allowed
= cpuset_cpus_allowed(p
);
3746 cpus_and(new_mask
, new_mask
, cpus_allowed
);
3747 retval
= set_cpus_allowed(p
, new_mask
);
3751 unlock_cpu_hotplug();
3755 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
3756 cpumask_t
*new_mask
)
3758 if (len
< sizeof(cpumask_t
)) {
3759 memset(new_mask
, 0, sizeof(cpumask_t
));
3760 } else if (len
> sizeof(cpumask_t
)) {
3761 len
= sizeof(cpumask_t
);
3763 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
3767 * sys_sched_setaffinity - set the cpu affinity of a process
3768 * @pid: pid of the process
3769 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3770 * @user_mask_ptr: user-space pointer to the new cpu mask
3772 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
3773 unsigned long __user
*user_mask_ptr
)
3778 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
3782 return sched_setaffinity(pid
, new_mask
);
3786 * Represents all cpu's present in the system
3787 * In systems capable of hotplug, this map could dynamically grow
3788 * as new cpu's are detected in the system via any platform specific
3789 * method, such as ACPI for e.g.
3792 cpumask_t cpu_present_map
;
3793 EXPORT_SYMBOL(cpu_present_map
);
3796 cpumask_t cpu_online_map
= CPU_MASK_ALL
;
3797 cpumask_t cpu_possible_map
= CPU_MASK_ALL
;
3800 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
3806 read_lock(&tasklist_lock
);
3809 p
= find_process_by_pid(pid
);
3814 cpus_and(*mask
, p
->cpus_allowed
, cpu_possible_map
);
3817 read_unlock(&tasklist_lock
);
3818 unlock_cpu_hotplug();
3826 * sys_sched_getaffinity - get the cpu affinity of a process
3827 * @pid: pid of the process
3828 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3829 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3831 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
3832 unsigned long __user
*user_mask_ptr
)
3837 if (len
< sizeof(cpumask_t
))
3840 ret
= sched_getaffinity(pid
, &mask
);
3844 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
3847 return sizeof(cpumask_t
);
3851 * sys_sched_yield - yield the current processor to other threads.
3853 * this function yields the current CPU by moving the calling thread
3854 * to the expired array. If there are no other threads running on this
3855 * CPU then this function will return.
3857 asmlinkage
long sys_sched_yield(void)
3859 runqueue_t
*rq
= this_rq_lock();
3860 prio_array_t
*array
= current
->array
;
3861 prio_array_t
*target
= rq
->expired
;
3863 schedstat_inc(rq
, yld_cnt
);
3865 * We implement yielding by moving the task into the expired
3868 * (special rule: RT tasks will just roundrobin in the active
3871 if (rt_task(current
))
3872 target
= rq
->active
;
3874 if (current
->array
->nr_active
== 1) {
3875 schedstat_inc(rq
, yld_act_empty
);
3876 if (!rq
->expired
->nr_active
)
3877 schedstat_inc(rq
, yld_both_empty
);
3878 } else if (!rq
->expired
->nr_active
)
3879 schedstat_inc(rq
, yld_exp_empty
);
3881 if (array
!= target
) {
3882 dequeue_task(current
, array
);
3883 enqueue_task(current
, target
);
3886 * requeue_task is cheaper so perform that if possible.
3888 requeue_task(current
, array
);
3891 * Since we are going to call schedule() anyway, there's
3892 * no need to preempt or enable interrupts:
3894 __release(rq
->lock
);
3895 _raw_spin_unlock(&rq
->lock
);
3896 preempt_enable_no_resched();
3903 static inline void __cond_resched(void)
3906 add_preempt_count(PREEMPT_ACTIVE
);
3908 sub_preempt_count(PREEMPT_ACTIVE
);
3909 } while (need_resched());
3912 int __sched
cond_resched(void)
3914 if (need_resched()) {
3921 EXPORT_SYMBOL(cond_resched
);
3924 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
3925 * call schedule, and on return reacquire the lock.
3927 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
3928 * operations here to prevent schedule() from being called twice (once via
3929 * spin_unlock(), once by hand).
3931 int cond_resched_lock(spinlock_t
* lock
)
3935 if (need_lockbreak(lock
)) {
3941 if (need_resched()) {
3942 _raw_spin_unlock(lock
);
3943 preempt_enable_no_resched();
3951 EXPORT_SYMBOL(cond_resched_lock
);
3953 int __sched
cond_resched_softirq(void)
3955 BUG_ON(!in_softirq());
3957 if (need_resched()) {
3958 __local_bh_enable();
3966 EXPORT_SYMBOL(cond_resched_softirq
);
3970 * yield - yield the current processor to other threads.
3972 * this is a shortcut for kernel-space yielding - it marks the
3973 * thread runnable and calls sys_sched_yield().
3975 void __sched
yield(void)
3977 set_current_state(TASK_RUNNING
);
3981 EXPORT_SYMBOL(yield
);
3984 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
3985 * that process accounting knows that this is a task in IO wait state.
3987 * But don't do that if it is a deliberate, throttling IO wait (this task
3988 * has set its backing_dev_info: the queue against which it should throttle)
3990 void __sched
io_schedule(void)
3992 struct runqueue
*rq
= &per_cpu(runqueues
, raw_smp_processor_id());
3994 atomic_inc(&rq
->nr_iowait
);
3996 atomic_dec(&rq
->nr_iowait
);
3999 EXPORT_SYMBOL(io_schedule
);
4001 long __sched
io_schedule_timeout(long timeout
)
4003 struct runqueue
*rq
= &per_cpu(runqueues
, raw_smp_processor_id());
4006 atomic_inc(&rq
->nr_iowait
);
4007 ret
= schedule_timeout(timeout
);
4008 atomic_dec(&rq
->nr_iowait
);
4013 * sys_sched_get_priority_max - return maximum RT priority.
4014 * @policy: scheduling class.
4016 * this syscall returns the maximum rt_priority that can be used
4017 * by a given scheduling class.
4019 asmlinkage
long sys_sched_get_priority_max(int policy
)
4026 ret
= MAX_USER_RT_PRIO
-1;
4036 * sys_sched_get_priority_min - return minimum RT priority.
4037 * @policy: scheduling class.
4039 * this syscall returns the minimum rt_priority that can be used
4040 * by a given scheduling class.
4042 asmlinkage
long sys_sched_get_priority_min(int policy
)
4058 * sys_sched_rr_get_interval - return the default timeslice of a process.
4059 * @pid: pid of the process.
4060 * @interval: userspace pointer to the timeslice value.
4062 * this syscall writes the default timeslice value of a given process
4063 * into the user-space timespec buffer. A value of '0' means infinity.
4066 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4068 int retval
= -EINVAL
;
4076 read_lock(&tasklist_lock
);
4077 p
= find_process_by_pid(pid
);
4081 retval
= security_task_getscheduler(p
);
4085 jiffies_to_timespec(p
->policy
& SCHED_FIFO
?
4086 0 : task_timeslice(p
), &t
);
4087 read_unlock(&tasklist_lock
);
4088 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4092 read_unlock(&tasklist_lock
);
4096 static inline struct task_struct
*eldest_child(struct task_struct
*p
)
4098 if (list_empty(&p
->children
)) return NULL
;
4099 return list_entry(p
->children
.next
,struct task_struct
,sibling
);
4102 static inline struct task_struct
*older_sibling(struct task_struct
*p
)
4104 if (p
->sibling
.prev
==&p
->parent
->children
) return NULL
;
4105 return list_entry(p
->sibling
.prev
,struct task_struct
,sibling
);
4108 static inline struct task_struct
*younger_sibling(struct task_struct
*p
)
4110 if (p
->sibling
.next
==&p
->parent
->children
) return NULL
;
4111 return list_entry(p
->sibling
.next
,struct task_struct
,sibling
);
4114 static void show_task(task_t
* p
)
4118 unsigned long free
= 0;
4119 static const char *stat_nam
[] = { "R", "S", "D", "T", "t", "Z", "X" };
4121 printk("%-13.13s ", p
->comm
);
4122 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4123 if (state
< ARRAY_SIZE(stat_nam
))
4124 printk(stat_nam
[state
]);
4127 #if (BITS_PER_LONG == 32)
4128 if (state
== TASK_RUNNING
)
4129 printk(" running ");
4131 printk(" %08lX ", thread_saved_pc(p
));
4133 if (state
== TASK_RUNNING
)
4134 printk(" running task ");
4136 printk(" %016lx ", thread_saved_pc(p
));
4138 #ifdef CONFIG_DEBUG_STACK_USAGE
4140 unsigned long * n
= (unsigned long *) (p
->thread_info
+1);
4143 free
= (unsigned long) n
- (unsigned long)(p
->thread_info
+1);
4146 printk("%5lu %5d %6d ", free
, p
->pid
, p
->parent
->pid
);
4147 if ((relative
= eldest_child(p
)))
4148 printk("%5d ", relative
->pid
);
4151 if ((relative
= younger_sibling(p
)))
4152 printk("%7d", relative
->pid
);
4155 if ((relative
= older_sibling(p
)))
4156 printk(" %5d", relative
->pid
);
4160 printk(" (L-TLB)\n");
4162 printk(" (NOTLB)\n");
4164 if (state
!= TASK_RUNNING
)
4165 show_stack(p
, NULL
);
4168 void show_state(void)
4172 #if (BITS_PER_LONG == 32)
4175 printk(" task PC pid father child younger older\n");
4179 printk(" task PC pid father child younger older\n");
4181 read_lock(&tasklist_lock
);
4182 do_each_thread(g
, p
) {
4184 * reset the NMI-timeout, listing all files on a slow
4185 * console might take alot of time:
4187 touch_nmi_watchdog();
4189 } while_each_thread(g
, p
);
4191 read_unlock(&tasklist_lock
);
4194 void __devinit
init_idle(task_t
*idle
, int cpu
)
4196 runqueue_t
*rq
= cpu_rq(cpu
);
4197 unsigned long flags
;
4199 idle
->sleep_avg
= 0;
4201 idle
->prio
= MAX_PRIO
;
4202 idle
->state
= TASK_RUNNING
;
4203 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4204 set_task_cpu(idle
, cpu
);
4206 spin_lock_irqsave(&rq
->lock
, flags
);
4207 rq
->curr
= rq
->idle
= idle
;
4208 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4211 set_tsk_need_resched(idle
);
4212 spin_unlock_irqrestore(&rq
->lock
, flags
);
4214 /* Set the preempt count _outside_ the spinlocks! */
4215 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4216 idle
->thread_info
->preempt_count
= (idle
->lock_depth
>= 0);
4218 idle
->thread_info
->preempt_count
= 0;
4223 * In a system that switches off the HZ timer nohz_cpu_mask
4224 * indicates which cpus entered this state. This is used
4225 * in the rcu update to wait only for active cpus. For system
4226 * which do not switch off the HZ timer nohz_cpu_mask should
4227 * always be CPU_MASK_NONE.
4229 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4233 * This is how migration works:
4235 * 1) we queue a migration_req_t structure in the source CPU's
4236 * runqueue and wake up that CPU's migration thread.
4237 * 2) we down() the locked semaphore => thread blocks.
4238 * 3) migration thread wakes up (implicitly it forces the migrated
4239 * thread off the CPU)
4240 * 4) it gets the migration request and checks whether the migrated
4241 * task is still in the wrong runqueue.
4242 * 5) if it's in the wrong runqueue then the migration thread removes
4243 * it and puts it into the right queue.
4244 * 6) migration thread up()s the semaphore.
4245 * 7) we wake up and the migration is done.
4249 * Change a given task's CPU affinity. Migrate the thread to a
4250 * proper CPU and schedule it away if the CPU it's executing on
4251 * is removed from the allowed bitmask.
4253 * NOTE: the caller must have a valid reference to the task, the
4254 * task must not exit() & deallocate itself prematurely. The
4255 * call is not atomic; no spinlocks may be held.
4257 int set_cpus_allowed(task_t
*p
, cpumask_t new_mask
)
4259 unsigned long flags
;
4261 migration_req_t req
;
4264 rq
= task_rq_lock(p
, &flags
);
4265 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4270 p
->cpus_allowed
= new_mask
;
4271 /* Can the task run on the task's current CPU? If so, we're done */
4272 if (cpu_isset(task_cpu(p
), new_mask
))
4275 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4276 /* Need help from migration thread: drop lock and wait. */
4277 task_rq_unlock(rq
, &flags
);
4278 wake_up_process(rq
->migration_thread
);
4279 wait_for_completion(&req
.done
);
4280 tlb_migrate_finish(p
->mm
);
4284 task_rq_unlock(rq
, &flags
);
4288 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4291 * Move (not current) task off this cpu, onto dest cpu. We're doing
4292 * this because either it can't run here any more (set_cpus_allowed()
4293 * away from this CPU, or CPU going down), or because we're
4294 * attempting to rebalance this task on exec (sched_exec).
4296 * So we race with normal scheduler movements, but that's OK, as long
4297 * as the task is no longer on this CPU.
4299 static void __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4301 runqueue_t
*rq_dest
, *rq_src
;
4303 if (unlikely(cpu_is_offline(dest_cpu
)))
4306 rq_src
= cpu_rq(src_cpu
);
4307 rq_dest
= cpu_rq(dest_cpu
);
4309 double_rq_lock(rq_src
, rq_dest
);
4310 /* Already moved. */
4311 if (task_cpu(p
) != src_cpu
)
4313 /* Affinity changed (again). */
4314 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
4317 set_task_cpu(p
, dest_cpu
);
4320 * Sync timestamp with rq_dest's before activating.
4321 * The same thing could be achieved by doing this step
4322 * afterwards, and pretending it was a local activate.
4323 * This way is cleaner and logically correct.
4325 p
->timestamp
= p
->timestamp
- rq_src
->timestamp_last_tick
4326 + rq_dest
->timestamp_last_tick
;
4327 deactivate_task(p
, rq_src
);
4328 activate_task(p
, rq_dest
, 0);
4329 if (TASK_PREEMPTS_CURR(p
, rq_dest
))
4330 resched_task(rq_dest
->curr
);
4334 double_rq_unlock(rq_src
, rq_dest
);
4338 * migration_thread - this is a highprio system thread that performs
4339 * thread migration by bumping thread off CPU then 'pushing' onto
4342 static int migration_thread(void * data
)
4345 int cpu
= (long)data
;
4348 BUG_ON(rq
->migration_thread
!= current
);
4350 set_current_state(TASK_INTERRUPTIBLE
);
4351 while (!kthread_should_stop()) {
4352 struct list_head
*head
;
4353 migration_req_t
*req
;
4355 if (current
->flags
& PF_FREEZE
)
4356 refrigerator(PF_FREEZE
);
4358 spin_lock_irq(&rq
->lock
);
4360 if (cpu_is_offline(cpu
)) {
4361 spin_unlock_irq(&rq
->lock
);
4365 if (rq
->active_balance
) {
4366 active_load_balance(rq
, cpu
);
4367 rq
->active_balance
= 0;
4370 head
= &rq
->migration_queue
;
4372 if (list_empty(head
)) {
4373 spin_unlock_irq(&rq
->lock
);
4375 set_current_state(TASK_INTERRUPTIBLE
);
4378 req
= list_entry(head
->next
, migration_req_t
, list
);
4379 list_del_init(head
->next
);
4381 if (req
->type
== REQ_MOVE_TASK
) {
4382 spin_unlock(&rq
->lock
);
4383 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
4385 } else if (req
->type
== REQ_SET_DOMAIN
) {
4387 spin_unlock_irq(&rq
->lock
);
4389 spin_unlock_irq(&rq
->lock
);
4393 complete(&req
->done
);
4395 __set_current_state(TASK_RUNNING
);
4399 /* Wait for kthread_stop */
4400 set_current_state(TASK_INTERRUPTIBLE
);
4401 while (!kthread_should_stop()) {
4403 set_current_state(TASK_INTERRUPTIBLE
);
4405 __set_current_state(TASK_RUNNING
);
4409 #ifdef CONFIG_HOTPLUG_CPU
4410 /* Figure out where task on dead CPU should go, use force if neccessary. */
4411 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*tsk
)
4417 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
4418 cpus_and(mask
, mask
, tsk
->cpus_allowed
);
4419 dest_cpu
= any_online_cpu(mask
);
4421 /* On any allowed CPU? */
4422 if (dest_cpu
== NR_CPUS
)
4423 dest_cpu
= any_online_cpu(tsk
->cpus_allowed
);
4425 /* No more Mr. Nice Guy. */
4426 if (dest_cpu
== NR_CPUS
) {
4427 cpus_setall(tsk
->cpus_allowed
);
4428 dest_cpu
= any_online_cpu(tsk
->cpus_allowed
);
4431 * Don't tell them about moving exiting tasks or
4432 * kernel threads (both mm NULL), since they never
4435 if (tsk
->mm
&& printk_ratelimit())
4436 printk(KERN_INFO
"process %d (%s) no "
4437 "longer affine to cpu%d\n",
4438 tsk
->pid
, tsk
->comm
, dead_cpu
);
4440 __migrate_task(tsk
, dead_cpu
, dest_cpu
);
4444 * While a dead CPU has no uninterruptible tasks queued at this point,
4445 * it might still have a nonzero ->nr_uninterruptible counter, because
4446 * for performance reasons the counter is not stricly tracking tasks to
4447 * their home CPUs. So we just add the counter to another CPU's counter,
4448 * to keep the global sum constant after CPU-down:
4450 static void migrate_nr_uninterruptible(runqueue_t
*rq_src
)
4452 runqueue_t
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
4453 unsigned long flags
;
4455 local_irq_save(flags
);
4456 double_rq_lock(rq_src
, rq_dest
);
4457 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
4458 rq_src
->nr_uninterruptible
= 0;
4459 double_rq_unlock(rq_src
, rq_dest
);
4460 local_irq_restore(flags
);
4463 /* Run through task list and migrate tasks from the dead cpu. */
4464 static void migrate_live_tasks(int src_cpu
)
4466 struct task_struct
*tsk
, *t
;
4468 write_lock_irq(&tasklist_lock
);
4470 do_each_thread(t
, tsk
) {
4474 if (task_cpu(tsk
) == src_cpu
)
4475 move_task_off_dead_cpu(src_cpu
, tsk
);
4476 } while_each_thread(t
, tsk
);
4478 write_unlock_irq(&tasklist_lock
);
4481 /* Schedules idle task to be the next runnable task on current CPU.
4482 * It does so by boosting its priority to highest possible and adding it to
4483 * the _front_ of runqueue. Used by CPU offline code.
4485 void sched_idle_next(void)
4487 int cpu
= smp_processor_id();
4488 runqueue_t
*rq
= this_rq();
4489 struct task_struct
*p
= rq
->idle
;
4490 unsigned long flags
;
4492 /* cpu has to be offline */
4493 BUG_ON(cpu_online(cpu
));
4495 /* Strictly not necessary since rest of the CPUs are stopped by now
4496 * and interrupts disabled on current cpu.
4498 spin_lock_irqsave(&rq
->lock
, flags
);
4500 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
4501 /* Add idle task to _front_ of it's priority queue */
4502 __activate_idle_task(p
, rq
);
4504 spin_unlock_irqrestore(&rq
->lock
, flags
);
4507 /* Ensures that the idle task is using init_mm right before its cpu goes
4510 void idle_task_exit(void)
4512 struct mm_struct
*mm
= current
->active_mm
;
4514 BUG_ON(cpu_online(smp_processor_id()));
4517 switch_mm(mm
, &init_mm
, current
);
4521 static void migrate_dead(unsigned int dead_cpu
, task_t
*tsk
)
4523 struct runqueue
*rq
= cpu_rq(dead_cpu
);
4525 /* Must be exiting, otherwise would be on tasklist. */
4526 BUG_ON(tsk
->exit_state
!= EXIT_ZOMBIE
&& tsk
->exit_state
!= EXIT_DEAD
);
4528 /* Cannot have done final schedule yet: would have vanished. */
4529 BUG_ON(tsk
->flags
& PF_DEAD
);
4531 get_task_struct(tsk
);
4534 * Drop lock around migration; if someone else moves it,
4535 * that's OK. No task can be added to this CPU, so iteration is
4538 spin_unlock_irq(&rq
->lock
);
4539 move_task_off_dead_cpu(dead_cpu
, tsk
);
4540 spin_lock_irq(&rq
->lock
);
4542 put_task_struct(tsk
);
4545 /* release_task() removes task from tasklist, so we won't find dead tasks. */
4546 static void migrate_dead_tasks(unsigned int dead_cpu
)
4549 struct runqueue
*rq
= cpu_rq(dead_cpu
);
4551 for (arr
= 0; arr
< 2; arr
++) {
4552 for (i
= 0; i
< MAX_PRIO
; i
++) {
4553 struct list_head
*list
= &rq
->arrays
[arr
].queue
[i
];
4554 while (!list_empty(list
))
4555 migrate_dead(dead_cpu
,
4556 list_entry(list
->next
, task_t
,
4561 #endif /* CONFIG_HOTPLUG_CPU */
4564 * migration_call - callback that gets triggered when a CPU is added.
4565 * Here we can start up the necessary migration thread for the new CPU.
4567 static int migration_call(struct notifier_block
*nfb
, unsigned long action
,
4570 int cpu
= (long)hcpu
;
4571 struct task_struct
*p
;
4572 struct runqueue
*rq
;
4573 unsigned long flags
;
4576 case CPU_UP_PREPARE
:
4577 p
= kthread_create(migration_thread
, hcpu
, "migration/%d",cpu
);
4580 p
->flags
|= PF_NOFREEZE
;
4581 kthread_bind(p
, cpu
);
4582 /* Must be high prio: stop_machine expects to yield to it. */
4583 rq
= task_rq_lock(p
, &flags
);
4584 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
4585 task_rq_unlock(rq
, &flags
);
4586 cpu_rq(cpu
)->migration_thread
= p
;
4589 /* Strictly unneccessary, as first user will wake it. */
4590 wake_up_process(cpu_rq(cpu
)->migration_thread
);
4592 #ifdef CONFIG_HOTPLUG_CPU
4593 case CPU_UP_CANCELED
:
4594 /* Unbind it from offline cpu so it can run. Fall thru. */
4595 kthread_bind(cpu_rq(cpu
)->migration_thread
,smp_processor_id());
4596 kthread_stop(cpu_rq(cpu
)->migration_thread
);
4597 cpu_rq(cpu
)->migration_thread
= NULL
;
4600 migrate_live_tasks(cpu
);
4602 kthread_stop(rq
->migration_thread
);
4603 rq
->migration_thread
= NULL
;
4604 /* Idle task back to normal (off runqueue, low prio) */
4605 rq
= task_rq_lock(rq
->idle
, &flags
);
4606 deactivate_task(rq
->idle
, rq
);
4607 rq
->idle
->static_prio
= MAX_PRIO
;
4608 __setscheduler(rq
->idle
, SCHED_NORMAL
, 0);
4609 migrate_dead_tasks(cpu
);
4610 task_rq_unlock(rq
, &flags
);
4611 migrate_nr_uninterruptible(rq
);
4612 BUG_ON(rq
->nr_running
!= 0);
4614 /* No need to migrate the tasks: it was best-effort if
4615 * they didn't do lock_cpu_hotplug(). Just wake up
4616 * the requestors. */
4617 spin_lock_irq(&rq
->lock
);
4618 while (!list_empty(&rq
->migration_queue
)) {
4619 migration_req_t
*req
;
4620 req
= list_entry(rq
->migration_queue
.next
,
4621 migration_req_t
, list
);
4622 BUG_ON(req
->type
!= REQ_MOVE_TASK
);
4623 list_del_init(&req
->list
);
4624 complete(&req
->done
);
4626 spin_unlock_irq(&rq
->lock
);
4633 /* Register at highest priority so that task migration (migrate_all_tasks)
4634 * happens before everything else.
4636 static struct notifier_block __devinitdata migration_notifier
= {
4637 .notifier_call
= migration_call
,
4641 int __init
migration_init(void)
4643 void *cpu
= (void *)(long)smp_processor_id();
4644 /* Start one for boot CPU. */
4645 migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
4646 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
4647 register_cpu_notifier(&migration_notifier
);
4653 #define SCHED_DOMAIN_DEBUG
4654 #ifdef SCHED_DOMAIN_DEBUG
4655 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
4660 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
4664 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
4669 struct sched_group
*group
= sd
->groups
;
4670 cpumask_t groupmask
;
4672 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
4673 cpus_clear(groupmask
);
4676 for (i
= 0; i
< level
+ 1; i
++)
4678 printk("domain %d: ", level
);
4680 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
4681 printk("does not load-balance\n");
4683 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain has parent");
4687 printk("span %s\n", str
);
4689 if (!cpu_isset(cpu
, sd
->span
))
4690 printk(KERN_ERR
"ERROR: domain->span does not contain CPU%d\n", cpu
);
4691 if (!cpu_isset(cpu
, group
->cpumask
))
4692 printk(KERN_ERR
"ERROR: domain->groups does not contain CPU%d\n", cpu
);
4695 for (i
= 0; i
< level
+ 2; i
++)
4701 printk(KERN_ERR
"ERROR: group is NULL\n");
4705 if (!group
->cpu_power
) {
4707 printk(KERN_ERR
"ERROR: domain->cpu_power not set\n");
4710 if (!cpus_weight(group
->cpumask
)) {
4712 printk(KERN_ERR
"ERROR: empty group\n");
4715 if (cpus_intersects(groupmask
, group
->cpumask
)) {
4717 printk(KERN_ERR
"ERROR: repeated CPUs\n");
4720 cpus_or(groupmask
, groupmask
, group
->cpumask
);
4722 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
4725 group
= group
->next
;
4726 } while (group
!= sd
->groups
);
4729 if (!cpus_equal(sd
->span
, groupmask
))
4730 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
4736 if (!cpus_subset(groupmask
, sd
->span
))
4737 printk(KERN_ERR
"ERROR: parent span is not a superset of domain->span\n");
4743 #define sched_domain_debug(sd, cpu) {}
4746 static int __devinit
sd_degenerate(struct sched_domain
*sd
)
4748 if (cpus_weight(sd
->span
) == 1)
4751 /* Following flags need at least 2 groups */
4752 if (sd
->flags
& (SD_LOAD_BALANCE
|
4753 SD_BALANCE_NEWIDLE
|
4756 if (sd
->groups
!= sd
->groups
->next
)
4760 /* Following flags don't use groups */
4761 if (sd
->flags
& (SD_WAKE_IDLE
|
4769 static int __devinit
sd_parent_degenerate(struct sched_domain
*sd
,
4770 struct sched_domain
*parent
)
4772 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
4774 if (sd_degenerate(parent
))
4777 if (!cpus_equal(sd
->span
, parent
->span
))
4780 /* Does parent contain flags not in child? */
4781 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
4782 if (cflags
& SD_WAKE_AFFINE
)
4783 pflags
&= ~SD_WAKE_BALANCE
;
4784 /* Flags needing groups don't count if only 1 group in parent */
4785 if (parent
->groups
== parent
->groups
->next
) {
4786 pflags
&= ~(SD_LOAD_BALANCE
|
4787 SD_BALANCE_NEWIDLE
|
4791 if (~cflags
& pflags
)
4798 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
4799 * hold the hotplug lock.
4801 void __devinit
cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
4803 migration_req_t req
;
4804 unsigned long flags
;
4805 runqueue_t
*rq
= cpu_rq(cpu
);
4807 struct sched_domain
*tmp
;
4809 /* Remove the sched domains which do not contribute to scheduling. */
4810 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
4811 struct sched_domain
*parent
= tmp
->parent
;
4814 if (sd_parent_degenerate(tmp
, parent
))
4815 tmp
->parent
= parent
->parent
;
4818 if (sd
&& sd_degenerate(sd
))
4821 sched_domain_debug(sd
, cpu
);
4823 spin_lock_irqsave(&rq
->lock
, flags
);
4825 if (cpu
== smp_processor_id() || !cpu_online(cpu
)) {
4828 init_completion(&req
.done
);
4829 req
.type
= REQ_SET_DOMAIN
;
4831 list_add(&req
.list
, &rq
->migration_queue
);
4835 spin_unlock_irqrestore(&rq
->lock
, flags
);
4838 wake_up_process(rq
->migration_thread
);
4839 wait_for_completion(&req
.done
);
4843 /* cpus with isolated domains */
4844 cpumask_t __devinitdata cpu_isolated_map
= CPU_MASK_NONE
;
4846 /* Setup the mask of cpus configured for isolated domains */
4847 static int __init
isolated_cpu_setup(char *str
)
4849 int ints
[NR_CPUS
], i
;
4851 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
4852 cpus_clear(cpu_isolated_map
);
4853 for (i
= 1; i
<= ints
[0]; i
++)
4854 if (ints
[i
] < NR_CPUS
)
4855 cpu_set(ints
[i
], cpu_isolated_map
);
4859 __setup ("isolcpus=", isolated_cpu_setup
);
4862 * init_sched_build_groups takes an array of groups, the cpumask we wish
4863 * to span, and a pointer to a function which identifies what group a CPU
4864 * belongs to. The return value of group_fn must be a valid index into the
4865 * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
4866 * keep track of groups covered with a cpumask_t).
4868 * init_sched_build_groups will build a circular linked list of the groups
4869 * covered by the given span, and will set each group's ->cpumask correctly,
4870 * and ->cpu_power to 0.
4872 void __devinit
init_sched_build_groups(struct sched_group groups
[],
4873 cpumask_t span
, int (*group_fn
)(int cpu
))
4875 struct sched_group
*first
= NULL
, *last
= NULL
;
4876 cpumask_t covered
= CPU_MASK_NONE
;
4879 for_each_cpu_mask(i
, span
) {
4880 int group
= group_fn(i
);
4881 struct sched_group
*sg
= &groups
[group
];
4884 if (cpu_isset(i
, covered
))
4887 sg
->cpumask
= CPU_MASK_NONE
;
4890 for_each_cpu_mask(j
, span
) {
4891 if (group_fn(j
) != group
)
4894 cpu_set(j
, covered
);
4895 cpu_set(j
, sg
->cpumask
);
4907 #ifdef ARCH_HAS_SCHED_DOMAIN
4908 extern void __devinit
arch_init_sched_domains(void);
4909 extern void __devinit
arch_destroy_sched_domains(void);
4911 #ifdef CONFIG_SCHED_SMT
4912 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
4913 static struct sched_group sched_group_cpus
[NR_CPUS
];
4914 static int __devinit
cpu_to_cpu_group(int cpu
)
4920 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
4921 static struct sched_group sched_group_phys
[NR_CPUS
];
4922 static int __devinit
cpu_to_phys_group(int cpu
)
4924 #ifdef CONFIG_SCHED_SMT
4925 return first_cpu(cpu_sibling_map
[cpu
]);
4933 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
4934 static struct sched_group sched_group_nodes
[MAX_NUMNODES
];
4935 static int __devinit
cpu_to_node_group(int cpu
)
4937 return cpu_to_node(cpu
);
4941 #if defined(CONFIG_SCHED_SMT) && defined(CONFIG_NUMA)
4943 * The domains setup code relies on siblings not spanning
4944 * multiple nodes. Make sure the architecture has a proper
4947 static void check_sibling_maps(void)
4951 for_each_online_cpu(i
) {
4952 for_each_cpu_mask(j
, cpu_sibling_map
[i
]) {
4953 if (cpu_to_node(i
) != cpu_to_node(j
)) {
4954 printk(KERN_INFO
"warning: CPU %d siblings map "
4955 "to different node - isolating "
4957 cpu_sibling_map
[i
] = cpumask_of_cpu(i
);
4966 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
4968 static void __devinit
arch_init_sched_domains(void)
4971 cpumask_t cpu_default_map
;
4973 #if defined(CONFIG_SCHED_SMT) && defined(CONFIG_NUMA)
4974 check_sibling_maps();
4977 * Setup mask for cpus without special case scheduling requirements.
4978 * For now this just excludes isolated cpus, but could be used to
4979 * exclude other special cases in the future.
4981 cpus_complement(cpu_default_map
, cpu_isolated_map
);
4982 cpus_and(cpu_default_map
, cpu_default_map
, cpu_online_map
);
4985 * Set up domains. Isolated domains just stay on the NULL domain.
4987 for_each_cpu_mask(i
, cpu_default_map
) {
4989 struct sched_domain
*sd
= NULL
, *p
;
4990 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
4992 cpus_and(nodemask
, nodemask
, cpu_default_map
);
4995 sd
= &per_cpu(node_domains
, i
);
4996 group
= cpu_to_node_group(i
);
4998 sd
->span
= cpu_default_map
;
4999 sd
->groups
= &sched_group_nodes
[group
];
5003 sd
= &per_cpu(phys_domains
, i
);
5004 group
= cpu_to_phys_group(i
);
5006 sd
->span
= nodemask
;
5008 sd
->groups
= &sched_group_phys
[group
];
5010 #ifdef CONFIG_SCHED_SMT
5012 sd
= &per_cpu(cpu_domains
, i
);
5013 group
= cpu_to_cpu_group(i
);
5014 *sd
= SD_SIBLING_INIT
;
5015 sd
->span
= cpu_sibling_map
[i
];
5016 cpus_and(sd
->span
, sd
->span
, cpu_default_map
);
5018 sd
->groups
= &sched_group_cpus
[group
];
5022 #ifdef CONFIG_SCHED_SMT
5023 /* Set up CPU (sibling) groups */
5024 for_each_online_cpu(i
) {
5025 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
5026 cpus_and(this_sibling_map
, this_sibling_map
, cpu_default_map
);
5027 if (i
!= first_cpu(this_sibling_map
))
5030 init_sched_build_groups(sched_group_cpus
, this_sibling_map
,
5035 /* Set up physical groups */
5036 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5037 cpumask_t nodemask
= node_to_cpumask(i
);
5039 cpus_and(nodemask
, nodemask
, cpu_default_map
);
5040 if (cpus_empty(nodemask
))
5043 init_sched_build_groups(sched_group_phys
, nodemask
,
5044 &cpu_to_phys_group
);
5048 /* Set up node groups */
5049 init_sched_build_groups(sched_group_nodes
, cpu_default_map
,
5050 &cpu_to_node_group
);
5053 /* Calculate CPU power for physical packages and nodes */
5054 for_each_cpu_mask(i
, cpu_default_map
) {
5056 struct sched_domain
*sd
;
5057 #ifdef CONFIG_SCHED_SMT
5058 sd
= &per_cpu(cpu_domains
, i
);
5059 power
= SCHED_LOAD_SCALE
;
5060 sd
->groups
->cpu_power
= power
;
5063 sd
= &per_cpu(phys_domains
, i
);
5064 power
= SCHED_LOAD_SCALE
+ SCHED_LOAD_SCALE
*
5065 (cpus_weight(sd
->groups
->cpumask
)-1) / 10;
5066 sd
->groups
->cpu_power
= power
;
5069 if (i
== first_cpu(sd
->groups
->cpumask
)) {
5070 /* Only add "power" once for each physical package. */
5071 sd
= &per_cpu(node_domains
, i
);
5072 sd
->groups
->cpu_power
+= power
;
5077 /* Attach the domains */
5078 for_each_online_cpu(i
) {
5079 struct sched_domain
*sd
;
5080 #ifdef CONFIG_SCHED_SMT
5081 sd
= &per_cpu(cpu_domains
, i
);
5083 sd
= &per_cpu(phys_domains
, i
);
5085 cpu_attach_domain(sd
, i
);
5089 #ifdef CONFIG_HOTPLUG_CPU
5090 static void __devinit
arch_destroy_sched_domains(void)
5092 /* Do nothing: everything is statically allocated. */
5096 #endif /* ARCH_HAS_SCHED_DOMAIN */
5098 #ifdef CONFIG_HOTPLUG_CPU
5100 * Force a reinitialization of the sched domains hierarchy. The domains
5101 * and groups cannot be updated in place without racing with the balancing
5102 * code, so we temporarily attach all running cpus to the NULL domain
5103 * which will prevent rebalancing while the sched domains are recalculated.
5105 static int update_sched_domains(struct notifier_block
*nfb
,
5106 unsigned long action
, void *hcpu
)
5111 case CPU_UP_PREPARE
:
5112 case CPU_DOWN_PREPARE
:
5113 for_each_online_cpu(i
)
5114 cpu_attach_domain(NULL
, i
);
5115 arch_destroy_sched_domains();
5118 case CPU_UP_CANCELED
:
5119 case CPU_DOWN_FAILED
:
5123 * Fall through and re-initialise the domains.
5130 /* The hotplug lock is already held by cpu_up/cpu_down */
5131 arch_init_sched_domains();
5137 void __init
sched_init_smp(void)
5140 arch_init_sched_domains();
5141 unlock_cpu_hotplug();
5142 /* XXX: Theoretical race here - CPU may be hotplugged now */
5143 hotcpu_notifier(update_sched_domains
, 0);
5146 void __init
sched_init_smp(void)
5149 #endif /* CONFIG_SMP */
5151 int in_sched_functions(unsigned long addr
)
5153 /* Linker adds these: start and end of __sched functions */
5154 extern char __sched_text_start
[], __sched_text_end
[];
5155 return in_lock_functions(addr
) ||
5156 (addr
>= (unsigned long)__sched_text_start
5157 && addr
< (unsigned long)__sched_text_end
);
5160 void __init
sched_init(void)
5165 for (i
= 0; i
< NR_CPUS
; i
++) {
5166 prio_array_t
*array
;
5169 spin_lock_init(&rq
->lock
);
5171 rq
->active
= rq
->arrays
;
5172 rq
->expired
= rq
->arrays
+ 1;
5173 rq
->best_expired_prio
= MAX_PRIO
;
5177 for (j
= 1; j
< 3; j
++)
5178 rq
->cpu_load
[j
] = 0;
5179 rq
->active_balance
= 0;
5181 rq
->migration_thread
= NULL
;
5182 INIT_LIST_HEAD(&rq
->migration_queue
);
5184 atomic_set(&rq
->nr_iowait
, 0);
5186 for (j
= 0; j
< 2; j
++) {
5187 array
= rq
->arrays
+ j
;
5188 for (k
= 0; k
< MAX_PRIO
; k
++) {
5189 INIT_LIST_HEAD(array
->queue
+ k
);
5190 __clear_bit(k
, array
->bitmap
);
5192 // delimiter for bitsearch
5193 __set_bit(MAX_PRIO
, array
->bitmap
);
5198 * The boot idle thread does lazy MMU switching as well:
5200 atomic_inc(&init_mm
.mm_count
);
5201 enter_lazy_tlb(&init_mm
, current
);
5204 * Make us the idle thread. Technically, schedule() should not be
5205 * called from this thread, however somewhere below it might be,
5206 * but because we are the idle thread, we just pick up running again
5207 * when this runqueue becomes "idle".
5209 init_idle(current
, smp_processor_id());
5212 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5213 void __might_sleep(char *file
, int line
)
5215 #if defined(in_atomic)
5216 static unsigned long prev_jiffy
; /* ratelimiting */
5218 if ((in_atomic() || irqs_disabled()) &&
5219 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
5220 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
5222 prev_jiffy
= jiffies
;
5223 printk(KERN_ERR
"Debug: sleeping function called from invalid"
5224 " context at %s:%d\n", file
, line
);
5225 printk("in_atomic():%d, irqs_disabled():%d\n",
5226 in_atomic(), irqs_disabled());
5231 EXPORT_SYMBOL(__might_sleep
);
5234 #ifdef CONFIG_MAGIC_SYSRQ
5235 void normalize_rt_tasks(void)
5237 struct task_struct
*p
;
5238 prio_array_t
*array
;
5239 unsigned long flags
;
5242 read_lock_irq(&tasklist_lock
);
5243 for_each_process (p
) {
5247 rq
= task_rq_lock(p
, &flags
);
5251 deactivate_task(p
, task_rq(p
));
5252 __setscheduler(p
, SCHED_NORMAL
, 0);
5254 __activate_task(p
, task_rq(p
));
5255 resched_task(rq
->curr
);
5258 task_rq_unlock(rq
, &flags
);
5260 read_unlock_irq(&tasklist_lock
);
5263 #endif /* CONFIG_MAGIC_SYSRQ */