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/capability.h>
31 #include <linux/completion.h>
32 #include <linux/kernel_stat.h>
33 #include <linux/debug_locks.h>
34 #include <linux/security.h>
35 #include <linux/notifier.h>
36 #include <linux/profile.h>
37 #include <linux/suspend.h>
38 #include <linux/vmalloc.h>
39 #include <linux/blkdev.h>
40 #include <linux/delay.h>
41 #include <linux/smp.h>
42 #include <linux/threads.h>
43 #include <linux/timer.h>
44 #include <linux/rcupdate.h>
45 #include <linux/cpu.h>
46 #include <linux/cpuset.h>
47 #include <linux/percpu.h>
48 #include <linux/kthread.h>
49 #include <linux/seq_file.h>
50 #include <linux/syscalls.h>
51 #include <linux/times.h>
52 #include <linux/acct.h>
53 #include <linux/kprobes.h>
56 #include <asm/unistd.h>
59 * Convert user-nice values [ -20 ... 0 ... 19 ]
60 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
63 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
64 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
65 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
68 * 'User priority' is the nice value converted to something we
69 * can work with better when scaling various scheduler parameters,
70 * it's a [ 0 ... 39 ] range.
72 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
73 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
74 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
77 * Some helpers for converting nanosecond timing to jiffy resolution
79 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
80 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
83 * These are the 'tuning knobs' of the scheduler:
85 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
86 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
87 * Timeslices get refilled after they expire.
89 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
90 #define DEF_TIMESLICE (100 * HZ / 1000)
91 #define ON_RUNQUEUE_WEIGHT 30
92 #define CHILD_PENALTY 95
93 #define PARENT_PENALTY 100
95 #define PRIO_BONUS_RATIO 25
96 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
97 #define INTERACTIVE_DELTA 2
98 #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
99 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
100 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
103 * If a task is 'interactive' then we reinsert it in the active
104 * array after it has expired its current timeslice. (it will not
105 * continue to run immediately, it will still roundrobin with
106 * other interactive tasks.)
108 * This part scales the interactivity limit depending on niceness.
110 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
111 * Here are a few examples of different nice levels:
113 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
114 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
115 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
116 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
117 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
119 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
120 * priority range a task can explore, a value of '1' means the
121 * task is rated interactive.)
123 * Ie. nice +19 tasks can never get 'interactive' enough to be
124 * reinserted into the active array. And only heavily CPU-hog nice -20
125 * tasks will be expired. Default nice 0 tasks are somewhere between,
126 * it takes some effort for them to get interactive, but it's not
130 #define CURRENT_BONUS(p) \
131 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
134 #define GRANULARITY (10 * HZ / 1000 ? : 1)
137 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
138 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
141 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
142 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
145 #define SCALE(v1,v1_max,v2_max) \
146 (v1) * (v2_max) / (v1_max)
149 (SCALE(TASK_NICE(p) + 20, 40, MAX_BONUS) - 20 * MAX_BONUS / 40 + \
152 #define TASK_INTERACTIVE(p) \
153 ((p)->prio <= (p)->static_prio - DELTA(p))
155 #define INTERACTIVE_SLEEP(p) \
156 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
157 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
159 #define TASK_PREEMPTS_CURR(p, rq) \
160 ((p)->prio < (rq)->curr->prio)
163 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
164 * to time slice values: [800ms ... 100ms ... 5ms]
166 * The higher a thread's priority, the bigger timeslices
167 * it gets during one round of execution. But even the lowest
168 * priority thread gets MIN_TIMESLICE worth of execution time.
171 #define SCALE_PRIO(x, prio) \
172 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
174 static unsigned int static_prio_timeslice(int static_prio
)
176 if (static_prio
< NICE_TO_PRIO(0))
177 return SCALE_PRIO(DEF_TIMESLICE
* 4, static_prio
);
179 return SCALE_PRIO(DEF_TIMESLICE
, static_prio
);
182 static inline unsigned int task_timeslice(task_t
*p
)
184 return static_prio_timeslice(p
->static_prio
);
187 #define task_hot(p, now, sd) ((long long) ((now) - (p)->last_ran) \
188 < (long long) (sd)->cache_hot_time)
191 * These are the runqueue data structures:
194 typedef struct runqueue runqueue_t
;
197 unsigned int nr_active
;
198 DECLARE_BITMAP(bitmap
, MAX_PRIO
+1); /* include 1 bit for delimiter */
199 struct list_head queue
[MAX_PRIO
];
203 * This is the main, per-CPU runqueue data structure.
205 * Locking rule: those places that want to lock multiple runqueues
206 * (such as the load balancing or the thread migration code), lock
207 * acquire operations must be ordered by ascending &runqueue.
213 * nr_running and cpu_load should be in the same cacheline because
214 * remote CPUs use both these fields when doing load calculation.
216 unsigned long nr_running
;
217 unsigned long raw_weighted_load
;
219 unsigned long cpu_load
[3];
221 unsigned long long nr_switches
;
224 * This is part of a global counter where only the total sum
225 * over all CPUs matters. A task can increase this counter on
226 * one CPU and if it got migrated afterwards it may decrease
227 * it on another CPU. Always updated under the runqueue lock:
229 unsigned long nr_uninterruptible
;
231 unsigned long expired_timestamp
;
232 unsigned long long timestamp_last_tick
;
234 struct mm_struct
*prev_mm
;
235 prio_array_t
*active
, *expired
, arrays
[2];
236 int best_expired_prio
;
240 struct sched_domain
*sd
;
242 /* For active balancing */
246 task_t
*migration_thread
;
247 struct list_head migration_queue
;
250 #ifdef CONFIG_SCHEDSTATS
252 struct sched_info rq_sched_info
;
254 /* sys_sched_yield() stats */
255 unsigned long yld_exp_empty
;
256 unsigned long yld_act_empty
;
257 unsigned long yld_both_empty
;
258 unsigned long yld_cnt
;
260 /* schedule() stats */
261 unsigned long sched_switch
;
262 unsigned long sched_cnt
;
263 unsigned long sched_goidle
;
265 /* try_to_wake_up() stats */
266 unsigned long ttwu_cnt
;
267 unsigned long ttwu_local
;
269 struct lock_class_key rq_lock_key
;
272 static DEFINE_PER_CPU(struct runqueue
, runqueues
);
275 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
276 * See detach_destroy_domains: synchronize_sched for details.
278 * The domain tree of any CPU may only be accessed from within
279 * preempt-disabled sections.
281 #define for_each_domain(cpu, domain) \
282 for (domain = rcu_dereference(cpu_rq(cpu)->sd); domain; domain = domain->parent)
284 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
285 #define this_rq() (&__get_cpu_var(runqueues))
286 #define task_rq(p) cpu_rq(task_cpu(p))
287 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
289 #ifndef prepare_arch_switch
290 # define prepare_arch_switch(next) do { } while (0)
292 #ifndef finish_arch_switch
293 # define finish_arch_switch(prev) do { } while (0)
296 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
297 static inline int task_running(runqueue_t
*rq
, task_t
*p
)
299 return rq
->curr
== p
;
302 static inline void prepare_lock_switch(runqueue_t
*rq
, task_t
*next
)
306 static inline void finish_lock_switch(runqueue_t
*rq
, task_t
*prev
)
308 #ifdef CONFIG_DEBUG_SPINLOCK
309 /* this is a valid case when another task releases the spinlock */
310 rq
->lock
.owner
= current
;
313 * If we are tracking spinlock dependencies then we have to
314 * fix up the runqueue lock - which gets 'carried over' from
317 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
319 spin_unlock_irq(&rq
->lock
);
322 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
323 static inline int task_running(runqueue_t
*rq
, task_t
*p
)
328 return rq
->curr
== p
;
332 static inline void prepare_lock_switch(runqueue_t
*rq
, task_t
*next
)
336 * We can optimise this out completely for !SMP, because the
337 * SMP rebalancing from interrupt is the only thing that cares
342 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
343 spin_unlock_irq(&rq
->lock
);
345 spin_unlock(&rq
->lock
);
349 static inline void finish_lock_switch(runqueue_t
*rq
, task_t
*prev
)
353 * After ->oncpu is cleared, the task can be moved to a different CPU.
354 * We must ensure this doesn't happen until the switch is completely
360 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
364 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
367 * __task_rq_lock - lock the runqueue a given task resides on.
368 * Must be called interrupts disabled.
370 static inline runqueue_t
*__task_rq_lock(task_t
*p
)
377 spin_lock(&rq
->lock
);
378 if (unlikely(rq
!= task_rq(p
))) {
379 spin_unlock(&rq
->lock
);
380 goto repeat_lock_task
;
386 * task_rq_lock - lock the runqueue a given task resides on and disable
387 * interrupts. Note the ordering: we can safely lookup the task_rq without
388 * explicitly disabling preemption.
390 static runqueue_t
*task_rq_lock(task_t
*p
, unsigned long *flags
)
396 local_irq_save(*flags
);
398 spin_lock(&rq
->lock
);
399 if (unlikely(rq
!= task_rq(p
))) {
400 spin_unlock_irqrestore(&rq
->lock
, *flags
);
401 goto repeat_lock_task
;
406 static inline void __task_rq_unlock(runqueue_t
*rq
)
409 spin_unlock(&rq
->lock
);
412 static inline void task_rq_unlock(runqueue_t
*rq
, unsigned long *flags
)
415 spin_unlock_irqrestore(&rq
->lock
, *flags
);
418 #ifdef CONFIG_SCHEDSTATS
420 * bump this up when changing the output format or the meaning of an existing
421 * format, so that tools can adapt (or abort)
423 #define SCHEDSTAT_VERSION 12
425 static int show_schedstat(struct seq_file
*seq
, void *v
)
429 seq_printf(seq
, "version %d\n", SCHEDSTAT_VERSION
);
430 seq_printf(seq
, "timestamp %lu\n", jiffies
);
431 for_each_online_cpu(cpu
) {
432 runqueue_t
*rq
= cpu_rq(cpu
);
434 struct sched_domain
*sd
;
438 /* runqueue-specific stats */
440 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
441 cpu
, rq
->yld_both_empty
,
442 rq
->yld_act_empty
, rq
->yld_exp_empty
, rq
->yld_cnt
,
443 rq
->sched_switch
, rq
->sched_cnt
, rq
->sched_goidle
,
444 rq
->ttwu_cnt
, rq
->ttwu_local
,
445 rq
->rq_sched_info
.cpu_time
,
446 rq
->rq_sched_info
.run_delay
, rq
->rq_sched_info
.pcnt
);
448 seq_printf(seq
, "\n");
451 /* domain-specific stats */
453 for_each_domain(cpu
, sd
) {
454 enum idle_type itype
;
455 char mask_str
[NR_CPUS
];
457 cpumask_scnprintf(mask_str
, NR_CPUS
, sd
->span
);
458 seq_printf(seq
, "domain%d %s", dcnt
++, mask_str
);
459 for (itype
= SCHED_IDLE
; itype
< MAX_IDLE_TYPES
;
461 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu %lu",
463 sd
->lb_balanced
[itype
],
464 sd
->lb_failed
[itype
],
465 sd
->lb_imbalance
[itype
],
466 sd
->lb_gained
[itype
],
467 sd
->lb_hot_gained
[itype
],
468 sd
->lb_nobusyq
[itype
],
469 sd
->lb_nobusyg
[itype
]);
471 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
472 sd
->alb_cnt
, sd
->alb_failed
, sd
->alb_pushed
,
473 sd
->sbe_cnt
, sd
->sbe_balanced
, sd
->sbe_pushed
,
474 sd
->sbf_cnt
, sd
->sbf_balanced
, sd
->sbf_pushed
,
475 sd
->ttwu_wake_remote
, sd
->ttwu_move_affine
, sd
->ttwu_move_balance
);
483 static int schedstat_open(struct inode
*inode
, struct file
*file
)
485 unsigned int size
= PAGE_SIZE
* (1 + num_online_cpus() / 32);
486 char *buf
= kmalloc(size
, GFP_KERNEL
);
492 res
= single_open(file
, show_schedstat
, NULL
);
494 m
= file
->private_data
;
502 struct file_operations proc_schedstat_operations
= {
503 .open
= schedstat_open
,
506 .release
= single_release
,
509 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
510 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
511 #else /* !CONFIG_SCHEDSTATS */
512 # define schedstat_inc(rq, field) do { } while (0)
513 # define schedstat_add(rq, field, amt) do { } while (0)
517 * rq_lock - lock a given runqueue and disable interrupts.
519 static inline runqueue_t
*this_rq_lock(void)
526 spin_lock(&rq
->lock
);
531 #ifdef CONFIG_SCHEDSTATS
533 * Called when a process is dequeued from the active array and given
534 * the cpu. We should note that with the exception of interactive
535 * tasks, the expired queue will become the active queue after the active
536 * queue is empty, without explicitly dequeuing and requeuing tasks in the
537 * expired queue. (Interactive tasks may be requeued directly to the
538 * active queue, thus delaying tasks in the expired queue from running;
539 * see scheduler_tick()).
541 * This function is only called from sched_info_arrive(), rather than
542 * dequeue_task(). Even though a task may be queued and dequeued multiple
543 * times as it is shuffled about, we're really interested in knowing how
544 * long it was from the *first* time it was queued to the time that it
547 static inline void sched_info_dequeued(task_t
*t
)
549 t
->sched_info
.last_queued
= 0;
553 * Called when a task finally hits the cpu. We can now calculate how
554 * long it was waiting to run. We also note when it began so that we
555 * can keep stats on how long its timeslice is.
557 static void sched_info_arrive(task_t
*t
)
559 unsigned long now
= jiffies
, diff
= 0;
560 struct runqueue
*rq
= task_rq(t
);
562 if (t
->sched_info
.last_queued
)
563 diff
= now
- t
->sched_info
.last_queued
;
564 sched_info_dequeued(t
);
565 t
->sched_info
.run_delay
+= diff
;
566 t
->sched_info
.last_arrival
= now
;
567 t
->sched_info
.pcnt
++;
572 rq
->rq_sched_info
.run_delay
+= diff
;
573 rq
->rq_sched_info
.pcnt
++;
577 * Called when a process is queued into either the active or expired
578 * array. The time is noted and later used to determine how long we
579 * had to wait for us to reach the cpu. Since the expired queue will
580 * become the active queue after active queue is empty, without dequeuing
581 * and requeuing any tasks, we are interested in queuing to either. It
582 * is unusual but not impossible for tasks to be dequeued and immediately
583 * requeued in the same or another array: this can happen in sched_yield(),
584 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
587 * This function is only called from enqueue_task(), but also only updates
588 * the timestamp if it is already not set. It's assumed that
589 * sched_info_dequeued() will clear that stamp when appropriate.
591 static inline void sched_info_queued(task_t
*t
)
593 if (!t
->sched_info
.last_queued
)
594 t
->sched_info
.last_queued
= jiffies
;
598 * Called when a process ceases being the active-running process, either
599 * voluntarily or involuntarily. Now we can calculate how long we ran.
601 static inline void sched_info_depart(task_t
*t
)
603 struct runqueue
*rq
= task_rq(t
);
604 unsigned long diff
= jiffies
- t
->sched_info
.last_arrival
;
606 t
->sched_info
.cpu_time
+= diff
;
609 rq
->rq_sched_info
.cpu_time
+= diff
;
613 * Called when tasks are switched involuntarily due, typically, to expiring
614 * their time slice. (This may also be called when switching to or from
615 * the idle task.) We are only called when prev != next.
617 static inline void sched_info_switch(task_t
*prev
, task_t
*next
)
619 struct runqueue
*rq
= task_rq(prev
);
622 * prev now departs the cpu. It's not interesting to record
623 * stats about how efficient we were at scheduling the idle
626 if (prev
!= rq
->idle
)
627 sched_info_depart(prev
);
629 if (next
!= rq
->idle
)
630 sched_info_arrive(next
);
633 #define sched_info_queued(t) do { } while (0)
634 #define sched_info_switch(t, next) do { } while (0)
635 #endif /* CONFIG_SCHEDSTATS */
638 * Adding/removing a task to/from a priority array:
640 static void dequeue_task(struct task_struct
*p
, prio_array_t
*array
)
643 list_del(&p
->run_list
);
644 if (list_empty(array
->queue
+ p
->prio
))
645 __clear_bit(p
->prio
, array
->bitmap
);
648 static void enqueue_task(struct task_struct
*p
, prio_array_t
*array
)
650 sched_info_queued(p
);
651 list_add_tail(&p
->run_list
, array
->queue
+ p
->prio
);
652 __set_bit(p
->prio
, array
->bitmap
);
658 * Put task to the end of the run list without the overhead of dequeue
659 * followed by enqueue.
661 static void requeue_task(struct task_struct
*p
, prio_array_t
*array
)
663 list_move_tail(&p
->run_list
, array
->queue
+ p
->prio
);
666 static inline void enqueue_task_head(struct task_struct
*p
, prio_array_t
*array
)
668 list_add(&p
->run_list
, array
->queue
+ p
->prio
);
669 __set_bit(p
->prio
, array
->bitmap
);
675 * __normal_prio - return the priority that is based on the static
676 * priority but is modified by bonuses/penalties.
678 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
679 * into the -5 ... 0 ... +5 bonus/penalty range.
681 * We use 25% of the full 0...39 priority range so that:
683 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
684 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
686 * Both properties are important to certain workloads.
689 static inline int __normal_prio(task_t
*p
)
693 bonus
= CURRENT_BONUS(p
) - MAX_BONUS
/ 2;
695 prio
= p
->static_prio
- bonus
;
696 if (prio
< MAX_RT_PRIO
)
698 if (prio
> MAX_PRIO
-1)
704 * To aid in avoiding the subversion of "niceness" due to uneven distribution
705 * of tasks with abnormal "nice" values across CPUs the contribution that
706 * each task makes to its run queue's load is weighted according to its
707 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
708 * scaled version of the new time slice allocation that they receive on time
713 * Assume: static_prio_timeslice(NICE_TO_PRIO(0)) == DEF_TIMESLICE
714 * If static_prio_timeslice() is ever changed to break this assumption then
715 * this code will need modification
717 #define TIME_SLICE_NICE_ZERO DEF_TIMESLICE
718 #define LOAD_WEIGHT(lp) \
719 (((lp) * SCHED_LOAD_SCALE) / TIME_SLICE_NICE_ZERO)
720 #define PRIO_TO_LOAD_WEIGHT(prio) \
721 LOAD_WEIGHT(static_prio_timeslice(prio))
722 #define RTPRIO_TO_LOAD_WEIGHT(rp) \
723 (PRIO_TO_LOAD_WEIGHT(MAX_RT_PRIO) + LOAD_WEIGHT(rp))
725 static void set_load_weight(task_t
*p
)
727 if (has_rt_policy(p
)) {
729 if (p
== task_rq(p
)->migration_thread
)
731 * The migration thread does the actual balancing.
732 * Giving its load any weight will skew balancing
738 p
->load_weight
= RTPRIO_TO_LOAD_WEIGHT(p
->rt_priority
);
740 p
->load_weight
= PRIO_TO_LOAD_WEIGHT(p
->static_prio
);
743 static inline void inc_raw_weighted_load(runqueue_t
*rq
, const task_t
*p
)
745 rq
->raw_weighted_load
+= p
->load_weight
;
748 static inline void dec_raw_weighted_load(runqueue_t
*rq
, const task_t
*p
)
750 rq
->raw_weighted_load
-= p
->load_weight
;
753 static inline void inc_nr_running(task_t
*p
, runqueue_t
*rq
)
756 inc_raw_weighted_load(rq
, p
);
759 static inline void dec_nr_running(task_t
*p
, runqueue_t
*rq
)
762 dec_raw_weighted_load(rq
, p
);
766 * Calculate the expected normal priority: i.e. priority
767 * without taking RT-inheritance into account. Might be
768 * boosted by interactivity modifiers. Changes upon fork,
769 * setprio syscalls, and whenever the interactivity
770 * estimator recalculates.
772 static inline int normal_prio(task_t
*p
)
776 if (has_rt_policy(p
))
777 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
779 prio
= __normal_prio(p
);
784 * Calculate the current priority, i.e. the priority
785 * taken into account by the scheduler. This value might
786 * be boosted by RT tasks, or might be boosted by
787 * interactivity modifiers. Will be RT if the task got
788 * RT-boosted. If not then it returns p->normal_prio.
790 static int effective_prio(task_t
*p
)
792 p
->normal_prio
= normal_prio(p
);
794 * If we are RT tasks or we were boosted to RT priority,
795 * keep the priority unchanged. Otherwise, update priority
796 * to the normal priority:
798 if (!rt_prio(p
->prio
))
799 return p
->normal_prio
;
804 * __activate_task - move a task to the runqueue.
806 static void __activate_task(task_t
*p
, runqueue_t
*rq
)
808 prio_array_t
*target
= rq
->active
;
811 target
= rq
->expired
;
812 enqueue_task(p
, target
);
813 inc_nr_running(p
, rq
);
817 * __activate_idle_task - move idle task to the _front_ of runqueue.
819 static inline void __activate_idle_task(task_t
*p
, runqueue_t
*rq
)
821 enqueue_task_head(p
, rq
->active
);
822 inc_nr_running(p
, rq
);
826 * Recalculate p->normal_prio and p->prio after having slept,
827 * updating the sleep-average too:
829 static int recalc_task_prio(task_t
*p
, unsigned long long now
)
831 /* Caller must always ensure 'now >= p->timestamp' */
832 unsigned long sleep_time
= now
- p
->timestamp
;
837 if (likely(sleep_time
> 0)) {
839 * This ceiling is set to the lowest priority that would allow
840 * a task to be reinserted into the active array on timeslice
843 unsigned long ceiling
= INTERACTIVE_SLEEP(p
);
845 if (p
->mm
&& sleep_time
> ceiling
&& p
->sleep_avg
< ceiling
) {
847 * Prevents user tasks from achieving best priority
848 * with one single large enough sleep.
850 p
->sleep_avg
= ceiling
;
852 * Using INTERACTIVE_SLEEP() as a ceiling places a
853 * nice(0) task 1ms sleep away from promotion, and
854 * gives it 700ms to round-robin with no chance of
855 * being demoted. This is more than generous, so
856 * mark this sleep as non-interactive to prevent the
857 * on-runqueue bonus logic from intervening should
858 * this task not receive cpu immediately.
860 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
863 * Tasks waking from uninterruptible sleep are
864 * limited in their sleep_avg rise as they
865 * are likely to be waiting on I/O
867 if (p
->sleep_type
== SLEEP_NONINTERACTIVE
&& p
->mm
) {
868 if (p
->sleep_avg
>= ceiling
)
870 else if (p
->sleep_avg
+ sleep_time
>=
872 p
->sleep_avg
= ceiling
;
878 * This code gives a bonus to interactive tasks.
880 * The boost works by updating the 'average sleep time'
881 * value here, based on ->timestamp. The more time a
882 * task spends sleeping, the higher the average gets -
883 * and the higher the priority boost gets as well.
885 p
->sleep_avg
+= sleep_time
;
888 if (p
->sleep_avg
> NS_MAX_SLEEP_AVG
)
889 p
->sleep_avg
= NS_MAX_SLEEP_AVG
;
892 return effective_prio(p
);
896 * activate_task - move a task to the runqueue and do priority recalculation
898 * Update all the scheduling statistics stuff. (sleep average
899 * calculation, priority modifiers, etc.)
901 static void activate_task(task_t
*p
, runqueue_t
*rq
, int local
)
903 unsigned long long now
;
908 /* Compensate for drifting sched_clock */
909 runqueue_t
*this_rq
= this_rq();
910 now
= (now
- this_rq
->timestamp_last_tick
)
911 + rq
->timestamp_last_tick
;
916 p
->prio
= recalc_task_prio(p
, now
);
919 * This checks to make sure it's not an uninterruptible task
920 * that is now waking up.
922 if (p
->sleep_type
== SLEEP_NORMAL
) {
924 * Tasks which were woken up by interrupts (ie. hw events)
925 * are most likely of interactive nature. So we give them
926 * the credit of extending their sleep time to the period
927 * of time they spend on the runqueue, waiting for execution
928 * on a CPU, first time around:
931 p
->sleep_type
= SLEEP_INTERRUPTED
;
934 * Normal first-time wakeups get a credit too for
935 * on-runqueue time, but it will be weighted down:
937 p
->sleep_type
= SLEEP_INTERACTIVE
;
942 __activate_task(p
, rq
);
946 * deactivate_task - remove a task from the runqueue.
948 static void deactivate_task(struct task_struct
*p
, runqueue_t
*rq
)
950 dec_nr_running(p
, rq
);
951 dequeue_task(p
, p
->array
);
956 * resched_task - mark a task 'to be rescheduled now'.
958 * On UP this means the setting of the need_resched flag, on SMP it
959 * might also involve a cross-CPU call to trigger the scheduler on
964 #ifndef tsk_is_polling
965 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
968 static void resched_task(task_t
*p
)
972 assert_spin_locked(&task_rq(p
)->lock
);
974 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
977 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
980 if (cpu
== smp_processor_id())
983 /* NEED_RESCHED must be visible before we test polling */
985 if (!tsk_is_polling(p
))
986 smp_send_reschedule(cpu
);
989 static inline void resched_task(task_t
*p
)
991 assert_spin_locked(&task_rq(p
)->lock
);
992 set_tsk_need_resched(p
);
997 * task_curr - is this task currently executing on a CPU?
998 * @p: the task in question.
1000 inline int task_curr(const task_t
*p
)
1002 return cpu_curr(task_cpu(p
)) == p
;
1005 /* Used instead of source_load when we know the type == 0 */
1006 unsigned long weighted_cpuload(const int cpu
)
1008 return cpu_rq(cpu
)->raw_weighted_load
;
1013 struct list_head list
;
1018 struct completion done
;
1022 * The task's runqueue lock must be held.
1023 * Returns true if you have to wait for migration thread.
1025 static int migrate_task(task_t
*p
, int dest_cpu
, migration_req_t
*req
)
1027 runqueue_t
*rq
= task_rq(p
);
1030 * If the task is not on a runqueue (and not running), then
1031 * it is sufficient to simply update the task's cpu field.
1033 if (!p
->array
&& !task_running(rq
, p
)) {
1034 set_task_cpu(p
, dest_cpu
);
1038 init_completion(&req
->done
);
1040 req
->dest_cpu
= dest_cpu
;
1041 list_add(&req
->list
, &rq
->migration_queue
);
1046 * wait_task_inactive - wait for a thread to unschedule.
1048 * The caller must ensure that the task *will* unschedule sometime soon,
1049 * else this function might spin for a *long* time. This function can't
1050 * be called with interrupts off, or it may introduce deadlock with
1051 * smp_call_function() if an IPI is sent by the same process we are
1052 * waiting to become inactive.
1054 void wait_task_inactive(task_t
*p
)
1056 unsigned long flags
;
1061 rq
= task_rq_lock(p
, &flags
);
1062 /* Must be off runqueue entirely, not preempted. */
1063 if (unlikely(p
->array
|| task_running(rq
, p
))) {
1064 /* If it's preempted, we yield. It could be a while. */
1065 preempted
= !task_running(rq
, p
);
1066 task_rq_unlock(rq
, &flags
);
1072 task_rq_unlock(rq
, &flags
);
1076 * kick_process - kick a running thread to enter/exit the kernel
1077 * @p: the to-be-kicked thread
1079 * Cause a process which is running on another CPU to enter
1080 * kernel-mode, without any delay. (to get signals handled.)
1082 * NOTE: this function doesnt have to take the runqueue lock,
1083 * because all it wants to ensure is that the remote task enters
1084 * the kernel. If the IPI races and the task has been migrated
1085 * to another CPU then no harm is done and the purpose has been
1088 void kick_process(task_t
*p
)
1094 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1095 smp_send_reschedule(cpu
);
1100 * Return a low guess at the load of a migration-source cpu weighted
1101 * according to the scheduling class and "nice" value.
1103 * We want to under-estimate the load of migration sources, to
1104 * balance conservatively.
1106 static inline unsigned long source_load(int cpu
, int type
)
1108 runqueue_t
*rq
= cpu_rq(cpu
);
1111 return rq
->raw_weighted_load
;
1113 return min(rq
->cpu_load
[type
-1], rq
->raw_weighted_load
);
1117 * Return a high guess at the load of a migration-target cpu weighted
1118 * according to the scheduling class and "nice" value.
1120 static inline unsigned long target_load(int cpu
, int type
)
1122 runqueue_t
*rq
= cpu_rq(cpu
);
1125 return rq
->raw_weighted_load
;
1127 return max(rq
->cpu_load
[type
-1], rq
->raw_weighted_load
);
1131 * Return the average load per task on the cpu's run queue
1133 static inline unsigned long cpu_avg_load_per_task(int cpu
)
1135 runqueue_t
*rq
= cpu_rq(cpu
);
1136 unsigned long n
= rq
->nr_running
;
1138 return n
? rq
->raw_weighted_load
/ n
: SCHED_LOAD_SCALE
;
1142 * find_idlest_group finds and returns the least busy CPU group within the
1145 static struct sched_group
*
1146 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1148 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1149 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1150 int load_idx
= sd
->forkexec_idx
;
1151 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1154 unsigned long load
, avg_load
;
1158 /* Skip over this group if it has no CPUs allowed */
1159 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1162 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1164 /* Tally up the load of all CPUs in the group */
1167 for_each_cpu_mask(i
, group
->cpumask
) {
1168 /* Bias balancing toward cpus of our domain */
1170 load
= source_load(i
, load_idx
);
1172 load
= target_load(i
, load_idx
);
1177 /* Adjust by relative CPU power of the group */
1178 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
1181 this_load
= avg_load
;
1183 } else if (avg_load
< min_load
) {
1184 min_load
= avg_load
;
1188 group
= group
->next
;
1189 } while (group
!= sd
->groups
);
1191 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1197 * find_idlest_queue - find the idlest runqueue among the cpus in group.
1200 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1203 unsigned long load
, min_load
= ULONG_MAX
;
1207 /* Traverse only the allowed CPUs */
1208 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1210 for_each_cpu_mask(i
, tmp
) {
1211 load
= weighted_cpuload(i
);
1213 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1223 * sched_balance_self: balance the current task (running on cpu) in domains
1224 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1227 * Balance, ie. select the least loaded group.
1229 * Returns the target CPU number, or the same CPU if no balancing is needed.
1231 * preempt must be disabled.
1233 static int sched_balance_self(int cpu
, int flag
)
1235 struct task_struct
*t
= current
;
1236 struct sched_domain
*tmp
, *sd
= NULL
;
1238 for_each_domain(cpu
, tmp
) {
1240 * If power savings logic is enabled for a domain, stop there.
1242 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1244 if (tmp
->flags
& flag
)
1250 struct sched_group
*group
;
1255 group
= find_idlest_group(sd
, t
, cpu
);
1259 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1260 if (new_cpu
== -1 || new_cpu
== cpu
)
1263 /* Now try balancing at a lower domain level */
1267 weight
= cpus_weight(span
);
1268 for_each_domain(cpu
, tmp
) {
1269 if (weight
<= cpus_weight(tmp
->span
))
1271 if (tmp
->flags
& flag
)
1274 /* while loop will break here if sd == NULL */
1280 #endif /* CONFIG_SMP */
1283 * wake_idle() will wake a task on an idle cpu if task->cpu is
1284 * not idle and an idle cpu is available. The span of cpus to
1285 * search starts with cpus closest then further out as needed,
1286 * so we always favor a closer, idle cpu.
1288 * Returns the CPU we should wake onto.
1290 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1291 static int wake_idle(int cpu
, task_t
*p
)
1294 struct sched_domain
*sd
;
1300 for_each_domain(cpu
, sd
) {
1301 if (sd
->flags
& SD_WAKE_IDLE
) {
1302 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1303 for_each_cpu_mask(i
, tmp
) {
1314 static inline int wake_idle(int cpu
, task_t
*p
)
1321 * try_to_wake_up - wake up a thread
1322 * @p: the to-be-woken-up thread
1323 * @state: the mask of task states that can be woken
1324 * @sync: do a synchronous wakeup?
1326 * Put it on the run-queue if it's not already there. The "current"
1327 * thread is always on the run-queue (except when the actual
1328 * re-schedule is in progress), and as such you're allowed to do
1329 * the simpler "current->state = TASK_RUNNING" to mark yourself
1330 * runnable without the overhead of this.
1332 * returns failure only if the task is already active.
1334 static int try_to_wake_up(task_t
*p
, unsigned int state
, int sync
)
1336 int cpu
, this_cpu
, success
= 0;
1337 unsigned long flags
;
1341 unsigned long load
, this_load
;
1342 struct sched_domain
*sd
, *this_sd
= NULL
;
1346 rq
= task_rq_lock(p
, &flags
);
1347 old_state
= p
->state
;
1348 if (!(old_state
& state
))
1355 this_cpu
= smp_processor_id();
1358 if (unlikely(task_running(rq
, p
)))
1363 schedstat_inc(rq
, ttwu_cnt
);
1364 if (cpu
== this_cpu
) {
1365 schedstat_inc(rq
, ttwu_local
);
1369 for_each_domain(this_cpu
, sd
) {
1370 if (cpu_isset(cpu
, sd
->span
)) {
1371 schedstat_inc(sd
, ttwu_wake_remote
);
1377 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1381 * Check for affine wakeup and passive balancing possibilities.
1384 int idx
= this_sd
->wake_idx
;
1385 unsigned int imbalance
;
1387 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1389 load
= source_load(cpu
, idx
);
1390 this_load
= target_load(this_cpu
, idx
);
1392 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1394 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1395 unsigned long tl
= this_load
;
1396 unsigned long tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1399 * If sync wakeup then subtract the (maximum possible)
1400 * effect of the currently running task from the load
1401 * of the current CPU:
1404 tl
-= current
->load_weight
;
1407 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1408 100*(tl
+ p
->load_weight
) <= imbalance
*load
) {
1410 * This domain has SD_WAKE_AFFINE and
1411 * p is cache cold in this domain, and
1412 * there is no bad imbalance.
1414 schedstat_inc(this_sd
, ttwu_move_affine
);
1420 * Start passive balancing when half the imbalance_pct
1423 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1424 if (imbalance
*this_load
<= 100*load
) {
1425 schedstat_inc(this_sd
, ttwu_move_balance
);
1431 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1433 new_cpu
= wake_idle(new_cpu
, p
);
1434 if (new_cpu
!= cpu
) {
1435 set_task_cpu(p
, new_cpu
);
1436 task_rq_unlock(rq
, &flags
);
1437 /* might preempt at this point */
1438 rq
= task_rq_lock(p
, &flags
);
1439 old_state
= p
->state
;
1440 if (!(old_state
& state
))
1445 this_cpu
= smp_processor_id();
1450 #endif /* CONFIG_SMP */
1451 if (old_state
== TASK_UNINTERRUPTIBLE
) {
1452 rq
->nr_uninterruptible
--;
1454 * Tasks on involuntary sleep don't earn
1455 * sleep_avg beyond just interactive state.
1457 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
1461 * Tasks that have marked their sleep as noninteractive get
1462 * woken up with their sleep average not weighted in an
1465 if (old_state
& TASK_NONINTERACTIVE
)
1466 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
1469 activate_task(p
, rq
, cpu
== this_cpu
);
1471 * Sync wakeups (i.e. those types of wakeups where the waker
1472 * has indicated that it will leave the CPU in short order)
1473 * don't trigger a preemption, if the woken up task will run on
1474 * this cpu. (in this case the 'I will reschedule' promise of
1475 * the waker guarantees that the freshly woken up task is going
1476 * to be considered on this CPU.)
1478 if (!sync
|| cpu
!= this_cpu
) {
1479 if (TASK_PREEMPTS_CURR(p
, rq
))
1480 resched_task(rq
->curr
);
1485 p
->state
= TASK_RUNNING
;
1487 task_rq_unlock(rq
, &flags
);
1492 int fastcall
wake_up_process(task_t
*p
)
1494 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1495 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1498 EXPORT_SYMBOL(wake_up_process
);
1500 int fastcall
wake_up_state(task_t
*p
, unsigned int state
)
1502 return try_to_wake_up(p
, state
, 0);
1506 * Perform scheduler related setup for a newly forked process p.
1507 * p is forked by current.
1509 void fastcall
sched_fork(task_t
*p
, int clone_flags
)
1511 int cpu
= get_cpu();
1514 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1516 set_task_cpu(p
, cpu
);
1519 * We mark the process as running here, but have not actually
1520 * inserted it onto the runqueue yet. This guarantees that
1521 * nobody will actually run it, and a signal or other external
1522 * event cannot wake it up and insert it on the runqueue either.
1524 p
->state
= TASK_RUNNING
;
1527 * Make sure we do not leak PI boosting priority to the child:
1529 p
->prio
= current
->normal_prio
;
1531 INIT_LIST_HEAD(&p
->run_list
);
1533 #ifdef CONFIG_SCHEDSTATS
1534 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1536 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1539 #ifdef CONFIG_PREEMPT
1540 /* Want to start with kernel preemption disabled. */
1541 task_thread_info(p
)->preempt_count
= 1;
1544 * Share the timeslice between parent and child, thus the
1545 * total amount of pending timeslices in the system doesn't change,
1546 * resulting in more scheduling fairness.
1548 local_irq_disable();
1549 p
->time_slice
= (current
->time_slice
+ 1) >> 1;
1551 * The remainder of the first timeslice might be recovered by
1552 * the parent if the child exits early enough.
1554 p
->first_time_slice
= 1;
1555 current
->time_slice
>>= 1;
1556 p
->timestamp
= sched_clock();
1557 if (unlikely(!current
->time_slice
)) {
1559 * This case is rare, it happens when the parent has only
1560 * a single jiffy left from its timeslice. Taking the
1561 * runqueue lock is not a problem.
1563 current
->time_slice
= 1;
1571 * wake_up_new_task - wake up a newly created task for the first time.
1573 * This function will do some initial scheduler statistics housekeeping
1574 * that must be done for every newly created context, then puts the task
1575 * on the runqueue and wakes it.
1577 void fastcall
wake_up_new_task(task_t
*p
, unsigned long clone_flags
)
1579 unsigned long flags
;
1581 runqueue_t
*rq
, *this_rq
;
1583 rq
= task_rq_lock(p
, &flags
);
1584 BUG_ON(p
->state
!= TASK_RUNNING
);
1585 this_cpu
= smp_processor_id();
1589 * We decrease the sleep average of forking parents
1590 * and children as well, to keep max-interactive tasks
1591 * from forking tasks that are max-interactive. The parent
1592 * (current) is done further down, under its lock.
1594 p
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(p
) *
1595 CHILD_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1597 p
->prio
= effective_prio(p
);
1599 if (likely(cpu
== this_cpu
)) {
1600 if (!(clone_flags
& CLONE_VM
)) {
1602 * The VM isn't cloned, so we're in a good position to
1603 * do child-runs-first in anticipation of an exec. This
1604 * usually avoids a lot of COW overhead.
1606 if (unlikely(!current
->array
))
1607 __activate_task(p
, rq
);
1609 p
->prio
= current
->prio
;
1610 p
->normal_prio
= current
->normal_prio
;
1611 list_add_tail(&p
->run_list
, ¤t
->run_list
);
1612 p
->array
= current
->array
;
1613 p
->array
->nr_active
++;
1614 inc_nr_running(p
, rq
);
1618 /* Run child last */
1619 __activate_task(p
, rq
);
1621 * We skip the following code due to cpu == this_cpu
1623 * task_rq_unlock(rq, &flags);
1624 * this_rq = task_rq_lock(current, &flags);
1628 this_rq
= cpu_rq(this_cpu
);
1631 * Not the local CPU - must adjust timestamp. This should
1632 * get optimised away in the !CONFIG_SMP case.
1634 p
->timestamp
= (p
->timestamp
- this_rq
->timestamp_last_tick
)
1635 + rq
->timestamp_last_tick
;
1636 __activate_task(p
, rq
);
1637 if (TASK_PREEMPTS_CURR(p
, rq
))
1638 resched_task(rq
->curr
);
1641 * Parent and child are on different CPUs, now get the
1642 * parent runqueue to update the parent's ->sleep_avg:
1644 task_rq_unlock(rq
, &flags
);
1645 this_rq
= task_rq_lock(current
, &flags
);
1647 current
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(current
) *
1648 PARENT_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1649 task_rq_unlock(this_rq
, &flags
);
1653 * Potentially available exiting-child timeslices are
1654 * retrieved here - this way the parent does not get
1655 * penalized for creating too many threads.
1657 * (this cannot be used to 'generate' timeslices
1658 * artificially, because any timeslice recovered here
1659 * was given away by the parent in the first place.)
1661 void fastcall
sched_exit(task_t
*p
)
1663 unsigned long flags
;
1667 * If the child was a (relative-) CPU hog then decrease
1668 * the sleep_avg of the parent as well.
1670 rq
= task_rq_lock(p
->parent
, &flags
);
1671 if (p
->first_time_slice
&& task_cpu(p
) == task_cpu(p
->parent
)) {
1672 p
->parent
->time_slice
+= p
->time_slice
;
1673 if (unlikely(p
->parent
->time_slice
> task_timeslice(p
)))
1674 p
->parent
->time_slice
= task_timeslice(p
);
1676 if (p
->sleep_avg
< p
->parent
->sleep_avg
)
1677 p
->parent
->sleep_avg
= p
->parent
->sleep_avg
/
1678 (EXIT_WEIGHT
+ 1) * EXIT_WEIGHT
+ p
->sleep_avg
/
1680 task_rq_unlock(rq
, &flags
);
1684 * prepare_task_switch - prepare to switch tasks
1685 * @rq: the runqueue preparing to switch
1686 * @next: the task we are going to switch to.
1688 * This is called with the rq lock held and interrupts off. It must
1689 * be paired with a subsequent finish_task_switch after the context
1692 * prepare_task_switch sets up locking and calls architecture specific
1695 static inline void prepare_task_switch(runqueue_t
*rq
, task_t
*next
)
1697 prepare_lock_switch(rq
, next
);
1698 prepare_arch_switch(next
);
1702 * finish_task_switch - clean up after a task-switch
1703 * @rq: runqueue associated with task-switch
1704 * @prev: the thread we just switched away from.
1706 * finish_task_switch must be called after the context switch, paired
1707 * with a prepare_task_switch call before the context switch.
1708 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1709 * and do any other architecture-specific cleanup actions.
1711 * Note that we may have delayed dropping an mm in context_switch(). If
1712 * so, we finish that here outside of the runqueue lock. (Doing it
1713 * with the lock held can cause deadlocks; see schedule() for
1716 static inline void finish_task_switch(runqueue_t
*rq
, task_t
*prev
)
1717 __releases(rq
->lock
)
1719 struct mm_struct
*mm
= rq
->prev_mm
;
1720 unsigned long prev_task_flags
;
1725 * A task struct has one reference for the use as "current".
1726 * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1727 * calls schedule one last time. The schedule call will never return,
1728 * and the scheduled task must drop that reference.
1729 * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1730 * still held, otherwise prev could be scheduled on another cpu, die
1731 * there before we look at prev->state, and then the reference would
1733 * Manfred Spraul <manfred@colorfullife.com>
1735 prev_task_flags
= prev
->flags
;
1736 finish_arch_switch(prev
);
1737 finish_lock_switch(rq
, prev
);
1740 if (unlikely(prev_task_flags
& PF_DEAD
)) {
1742 * Remove function-return probe instances associated with this
1743 * task and put them back on the free list.
1745 kprobe_flush_task(prev
);
1746 put_task_struct(prev
);
1751 * schedule_tail - first thing a freshly forked thread must call.
1752 * @prev: the thread we just switched away from.
1754 asmlinkage
void schedule_tail(task_t
*prev
)
1755 __releases(rq
->lock
)
1757 runqueue_t
*rq
= this_rq();
1758 finish_task_switch(rq
, prev
);
1759 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1760 /* In this case, finish_task_switch does not reenable preemption */
1763 if (current
->set_child_tid
)
1764 put_user(current
->pid
, current
->set_child_tid
);
1768 * context_switch - switch to the new MM and the new
1769 * thread's register state.
1772 task_t
* context_switch(runqueue_t
*rq
, task_t
*prev
, task_t
*next
)
1774 struct mm_struct
*mm
= next
->mm
;
1775 struct mm_struct
*oldmm
= prev
->active_mm
;
1777 if (unlikely(!mm
)) {
1778 next
->active_mm
= oldmm
;
1779 atomic_inc(&oldmm
->mm_count
);
1780 enter_lazy_tlb(oldmm
, next
);
1782 switch_mm(oldmm
, mm
, next
);
1784 if (unlikely(!prev
->mm
)) {
1785 prev
->active_mm
= NULL
;
1786 WARN_ON(rq
->prev_mm
);
1787 rq
->prev_mm
= oldmm
;
1789 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1791 /* Here we just switch the register state and the stack. */
1792 switch_to(prev
, next
, prev
);
1798 * nr_running, nr_uninterruptible and nr_context_switches:
1800 * externally visible scheduler statistics: current number of runnable
1801 * threads, current number of uninterruptible-sleeping threads, total
1802 * number of context switches performed since bootup.
1804 unsigned long nr_running(void)
1806 unsigned long i
, sum
= 0;
1808 for_each_online_cpu(i
)
1809 sum
+= cpu_rq(i
)->nr_running
;
1814 unsigned long nr_uninterruptible(void)
1816 unsigned long i
, sum
= 0;
1818 for_each_possible_cpu(i
)
1819 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1822 * Since we read the counters lockless, it might be slightly
1823 * inaccurate. Do not allow it to go below zero though:
1825 if (unlikely((long)sum
< 0))
1831 unsigned long long nr_context_switches(void)
1834 unsigned long long sum
= 0;
1836 for_each_possible_cpu(i
)
1837 sum
+= cpu_rq(i
)->nr_switches
;
1842 unsigned long nr_iowait(void)
1844 unsigned long i
, sum
= 0;
1846 for_each_possible_cpu(i
)
1847 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1852 unsigned long nr_active(void)
1854 unsigned long i
, running
= 0, uninterruptible
= 0;
1856 for_each_online_cpu(i
) {
1857 running
+= cpu_rq(i
)->nr_running
;
1858 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
1861 if (unlikely((long)uninterruptible
< 0))
1862 uninterruptible
= 0;
1864 return running
+ uninterruptible
;
1870 * double_rq_lock - safely lock two runqueues
1872 * Note this does not disable interrupts like task_rq_lock,
1873 * you need to do so manually before calling.
1875 static void double_rq_lock(runqueue_t
*rq1
, runqueue_t
*rq2
)
1876 __acquires(rq1
->lock
)
1877 __acquires(rq2
->lock
)
1880 spin_lock(&rq1
->lock
);
1881 __acquire(rq2
->lock
); /* Fake it out ;) */
1884 spin_lock(&rq1
->lock
);
1885 spin_lock(&rq2
->lock
);
1887 spin_lock(&rq2
->lock
);
1888 spin_lock(&rq1
->lock
);
1894 * double_rq_unlock - safely unlock two runqueues
1896 * Note this does not restore interrupts like task_rq_unlock,
1897 * you need to do so manually after calling.
1899 static void double_rq_unlock(runqueue_t
*rq1
, runqueue_t
*rq2
)
1900 __releases(rq1
->lock
)
1901 __releases(rq2
->lock
)
1903 spin_unlock(&rq1
->lock
);
1905 spin_unlock(&rq2
->lock
);
1907 __release(rq2
->lock
);
1911 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1913 static void double_lock_balance(runqueue_t
*this_rq
, runqueue_t
*busiest
)
1914 __releases(this_rq
->lock
)
1915 __acquires(busiest
->lock
)
1916 __acquires(this_rq
->lock
)
1918 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1919 if (busiest
< this_rq
) {
1920 spin_unlock(&this_rq
->lock
);
1921 spin_lock(&busiest
->lock
);
1922 spin_lock(&this_rq
->lock
);
1924 spin_lock(&busiest
->lock
);
1929 * If dest_cpu is allowed for this process, migrate the task to it.
1930 * This is accomplished by forcing the cpu_allowed mask to only
1931 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
1932 * the cpu_allowed mask is restored.
1934 static void sched_migrate_task(task_t
*p
, int dest_cpu
)
1936 migration_req_t req
;
1938 unsigned long flags
;
1940 rq
= task_rq_lock(p
, &flags
);
1941 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
1942 || unlikely(cpu_is_offline(dest_cpu
)))
1945 /* force the process onto the specified CPU */
1946 if (migrate_task(p
, dest_cpu
, &req
)) {
1947 /* Need to wait for migration thread (might exit: take ref). */
1948 struct task_struct
*mt
= rq
->migration_thread
;
1949 get_task_struct(mt
);
1950 task_rq_unlock(rq
, &flags
);
1951 wake_up_process(mt
);
1952 put_task_struct(mt
);
1953 wait_for_completion(&req
.done
);
1957 task_rq_unlock(rq
, &flags
);
1961 * sched_exec - execve() is a valuable balancing opportunity, because at
1962 * this point the task has the smallest effective memory and cache footprint.
1964 void sched_exec(void)
1966 int new_cpu
, this_cpu
= get_cpu();
1967 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
1969 if (new_cpu
!= this_cpu
)
1970 sched_migrate_task(current
, new_cpu
);
1974 * pull_task - move a task from a remote runqueue to the local runqueue.
1975 * Both runqueues must be locked.
1978 void pull_task(runqueue_t
*src_rq
, prio_array_t
*src_array
, task_t
*p
,
1979 runqueue_t
*this_rq
, prio_array_t
*this_array
, int this_cpu
)
1981 dequeue_task(p
, src_array
);
1982 dec_nr_running(p
, src_rq
);
1983 set_task_cpu(p
, this_cpu
);
1984 inc_nr_running(p
, this_rq
);
1985 enqueue_task(p
, this_array
);
1986 p
->timestamp
= (p
->timestamp
- src_rq
->timestamp_last_tick
)
1987 + this_rq
->timestamp_last_tick
;
1989 * Note that idle threads have a prio of MAX_PRIO, for this test
1990 * to be always true for them.
1992 if (TASK_PREEMPTS_CURR(p
, this_rq
))
1993 resched_task(this_rq
->curr
);
1997 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2000 int can_migrate_task(task_t
*p
, runqueue_t
*rq
, int this_cpu
,
2001 struct sched_domain
*sd
, enum idle_type idle
,
2005 * We do not migrate tasks that are:
2006 * 1) running (obviously), or
2007 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2008 * 3) are cache-hot on their current CPU.
2010 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
2014 if (task_running(rq
, p
))
2018 * Aggressive migration if:
2019 * 1) task is cache cold, or
2020 * 2) too many balance attempts have failed.
2023 if (sd
->nr_balance_failed
> sd
->cache_nice_tries
)
2026 if (task_hot(p
, rq
->timestamp_last_tick
, sd
))
2031 #define rq_best_prio(rq) min((rq)->curr->prio, (rq)->best_expired_prio)
2033 * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted
2034 * load from busiest to this_rq, as part of a balancing operation within
2035 * "domain". Returns the number of tasks moved.
2037 * Called with both runqueues locked.
2039 static int move_tasks(runqueue_t
*this_rq
, int this_cpu
, runqueue_t
*busiest
,
2040 unsigned long max_nr_move
, unsigned long max_load_move
,
2041 struct sched_domain
*sd
, enum idle_type idle
,
2044 prio_array_t
*array
, *dst_array
;
2045 struct list_head
*head
, *curr
;
2046 int idx
, pulled
= 0, pinned
= 0, this_best_prio
, busiest_best_prio
;
2047 int busiest_best_prio_seen
;
2048 int skip_for_load
; /* skip the task based on weighted load issues */
2052 if (max_nr_move
== 0 || max_load_move
== 0)
2055 rem_load_move
= max_load_move
;
2057 this_best_prio
= rq_best_prio(this_rq
);
2058 busiest_best_prio
= rq_best_prio(busiest
);
2060 * Enable handling of the case where there is more than one task
2061 * with the best priority. If the current running task is one
2062 * of those with prio==busiest_best_prio we know it won't be moved
2063 * and therefore it's safe to override the skip (based on load) of
2064 * any task we find with that prio.
2066 busiest_best_prio_seen
= busiest_best_prio
== busiest
->curr
->prio
;
2069 * We first consider expired tasks. Those will likely not be
2070 * executed in the near future, and they are most likely to
2071 * be cache-cold, thus switching CPUs has the least effect
2074 if (busiest
->expired
->nr_active
) {
2075 array
= busiest
->expired
;
2076 dst_array
= this_rq
->expired
;
2078 array
= busiest
->active
;
2079 dst_array
= this_rq
->active
;
2083 /* Start searching at priority 0: */
2087 idx
= sched_find_first_bit(array
->bitmap
);
2089 idx
= find_next_bit(array
->bitmap
, MAX_PRIO
, idx
);
2090 if (idx
>= MAX_PRIO
) {
2091 if (array
== busiest
->expired
&& busiest
->active
->nr_active
) {
2092 array
= busiest
->active
;
2093 dst_array
= this_rq
->active
;
2099 head
= array
->queue
+ idx
;
2102 tmp
= list_entry(curr
, task_t
, run_list
);
2107 * To help distribute high priority tasks accross CPUs we don't
2108 * skip a task if it will be the highest priority task (i.e. smallest
2109 * prio value) on its new queue regardless of its load weight
2111 skip_for_load
= tmp
->load_weight
> rem_load_move
;
2112 if (skip_for_load
&& idx
< this_best_prio
)
2113 skip_for_load
= !busiest_best_prio_seen
&& idx
== busiest_best_prio
;
2114 if (skip_for_load
||
2115 !can_migrate_task(tmp
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2116 busiest_best_prio_seen
|= idx
== busiest_best_prio
;
2123 #ifdef CONFIG_SCHEDSTATS
2124 if (task_hot(tmp
, busiest
->timestamp_last_tick
, sd
))
2125 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2128 pull_task(busiest
, array
, tmp
, this_rq
, dst_array
, this_cpu
);
2130 rem_load_move
-= tmp
->load_weight
;
2133 * We only want to steal up to the prescribed number of tasks
2134 * and the prescribed amount of weighted load.
2136 if (pulled
< max_nr_move
&& rem_load_move
> 0) {
2137 if (idx
< this_best_prio
)
2138 this_best_prio
= idx
;
2146 * Right now, this is the only place pull_task() is called,
2147 * so we can safely collect pull_task() stats here rather than
2148 * inside pull_task().
2150 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2153 *all_pinned
= pinned
;
2158 * find_busiest_group finds and returns the busiest CPU group within the
2159 * domain. It calculates and returns the amount of weighted load which should be
2160 * moved to restore balance via the imbalance parameter.
2162 static struct sched_group
*
2163 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2164 unsigned long *imbalance
, enum idle_type idle
, int *sd_idle
)
2166 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2167 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2168 unsigned long max_pull
;
2169 unsigned long busiest_load_per_task
, busiest_nr_running
;
2170 unsigned long this_load_per_task
, this_nr_running
;
2172 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2173 int power_savings_balance
= 1;
2174 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2175 unsigned long min_nr_running
= ULONG_MAX
;
2176 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2179 max_load
= this_load
= total_load
= total_pwr
= 0;
2180 busiest_load_per_task
= busiest_nr_running
= 0;
2181 this_load_per_task
= this_nr_running
= 0;
2182 if (idle
== NOT_IDLE
)
2183 load_idx
= sd
->busy_idx
;
2184 else if (idle
== NEWLY_IDLE
)
2185 load_idx
= sd
->newidle_idx
;
2187 load_idx
= sd
->idle_idx
;
2190 unsigned long load
, group_capacity
;
2193 unsigned long sum_nr_running
, sum_weighted_load
;
2195 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2197 /* Tally up the load of all CPUs in the group */
2198 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2200 for_each_cpu_mask(i
, group
->cpumask
) {
2201 runqueue_t
*rq
= cpu_rq(i
);
2203 if (*sd_idle
&& !idle_cpu(i
))
2206 /* Bias balancing toward cpus of our domain */
2208 load
= target_load(i
, load_idx
);
2210 load
= source_load(i
, load_idx
);
2213 sum_nr_running
+= rq
->nr_running
;
2214 sum_weighted_load
+= rq
->raw_weighted_load
;
2217 total_load
+= avg_load
;
2218 total_pwr
+= group
->cpu_power
;
2220 /* Adjust by relative CPU power of the group */
2221 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
2223 group_capacity
= group
->cpu_power
/ SCHED_LOAD_SCALE
;
2226 this_load
= avg_load
;
2228 this_nr_running
= sum_nr_running
;
2229 this_load_per_task
= sum_weighted_load
;
2230 } else if (avg_load
> max_load
&&
2231 sum_nr_running
> group_capacity
) {
2232 max_load
= avg_load
;
2234 busiest_nr_running
= sum_nr_running
;
2235 busiest_load_per_task
= sum_weighted_load
;
2238 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2240 * Busy processors will not participate in power savings
2243 if (idle
== NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2247 * If the local group is idle or completely loaded
2248 * no need to do power savings balance at this domain
2250 if (local_group
&& (this_nr_running
>= group_capacity
||
2252 power_savings_balance
= 0;
2255 * If a group is already running at full capacity or idle,
2256 * don't include that group in power savings calculations
2258 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2263 * Calculate the group which has the least non-idle load.
2264 * This is the group from where we need to pick up the load
2267 if ((sum_nr_running
< min_nr_running
) ||
2268 (sum_nr_running
== min_nr_running
&&
2269 first_cpu(group
->cpumask
) <
2270 first_cpu(group_min
->cpumask
))) {
2272 min_nr_running
= sum_nr_running
;
2273 min_load_per_task
= sum_weighted_load
/
2278 * Calculate the group which is almost near its
2279 * capacity but still has some space to pick up some load
2280 * from other group and save more power
2282 if (sum_nr_running
<= group_capacity
- 1)
2283 if (sum_nr_running
> leader_nr_running
||
2284 (sum_nr_running
== leader_nr_running
&&
2285 first_cpu(group
->cpumask
) >
2286 first_cpu(group_leader
->cpumask
))) {
2287 group_leader
= group
;
2288 leader_nr_running
= sum_nr_running
;
2293 group
= group
->next
;
2294 } while (group
!= sd
->groups
);
2296 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2299 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2301 if (this_load
>= avg_load
||
2302 100*max_load
<= sd
->imbalance_pct
*this_load
)
2305 busiest_load_per_task
/= busiest_nr_running
;
2307 * We're trying to get all the cpus to the average_load, so we don't
2308 * want to push ourselves above the average load, nor do we wish to
2309 * reduce the max loaded cpu below the average load, as either of these
2310 * actions would just result in more rebalancing later, and ping-pong
2311 * tasks around. Thus we look for the minimum possible imbalance.
2312 * Negative imbalances (*we* are more loaded than anyone else) will
2313 * be counted as no imbalance for these purposes -- we can't fix that
2314 * by pulling tasks to us. Be careful of negative numbers as they'll
2315 * appear as very large values with unsigned longs.
2317 if (max_load
<= busiest_load_per_task
)
2321 * In the presence of smp nice balancing, certain scenarios can have
2322 * max load less than avg load(as we skip the groups at or below
2323 * its cpu_power, while calculating max_load..)
2325 if (max_load
< avg_load
) {
2327 goto small_imbalance
;
2330 /* Don't want to pull so many tasks that a group would go idle */
2331 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2333 /* How much load to actually move to equalise the imbalance */
2334 *imbalance
= min(max_pull
* busiest
->cpu_power
,
2335 (avg_load
- this_load
) * this->cpu_power
)
2339 * if *imbalance is less than the average load per runnable task
2340 * there is no gaurantee that any tasks will be moved so we'll have
2341 * a think about bumping its value to force at least one task to be
2344 if (*imbalance
< busiest_load_per_task
) {
2345 unsigned long pwr_now
, pwr_move
;
2350 pwr_move
= pwr_now
= 0;
2352 if (this_nr_running
) {
2353 this_load_per_task
/= this_nr_running
;
2354 if (busiest_load_per_task
> this_load_per_task
)
2357 this_load_per_task
= SCHED_LOAD_SCALE
;
2359 if (max_load
- this_load
>= busiest_load_per_task
* imbn
) {
2360 *imbalance
= busiest_load_per_task
;
2365 * OK, we don't have enough imbalance to justify moving tasks,
2366 * however we may be able to increase total CPU power used by
2370 pwr_now
+= busiest
->cpu_power
*
2371 min(busiest_load_per_task
, max_load
);
2372 pwr_now
+= this->cpu_power
*
2373 min(this_load_per_task
, this_load
);
2374 pwr_now
/= SCHED_LOAD_SCALE
;
2376 /* Amount of load we'd subtract */
2377 tmp
= busiest_load_per_task
*SCHED_LOAD_SCALE
/busiest
->cpu_power
;
2379 pwr_move
+= busiest
->cpu_power
*
2380 min(busiest_load_per_task
, max_load
- tmp
);
2382 /* Amount of load we'd add */
2383 if (max_load
*busiest
->cpu_power
<
2384 busiest_load_per_task
*SCHED_LOAD_SCALE
)
2385 tmp
= max_load
*busiest
->cpu_power
/this->cpu_power
;
2387 tmp
= busiest_load_per_task
*SCHED_LOAD_SCALE
/this->cpu_power
;
2388 pwr_move
+= this->cpu_power
*min(this_load_per_task
, this_load
+ tmp
);
2389 pwr_move
/= SCHED_LOAD_SCALE
;
2391 /* Move if we gain throughput */
2392 if (pwr_move
<= pwr_now
)
2395 *imbalance
= busiest_load_per_task
;
2401 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2402 if (idle
== NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2405 if (this == group_leader
&& group_leader
!= group_min
) {
2406 *imbalance
= min_load_per_task
;
2416 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2418 static runqueue_t
*find_busiest_queue(struct sched_group
*group
,
2419 enum idle_type idle
, unsigned long imbalance
)
2421 unsigned long max_load
= 0;
2422 runqueue_t
*busiest
= NULL
, *rqi
;
2425 for_each_cpu_mask(i
, group
->cpumask
) {
2428 if (rqi
->nr_running
== 1 && rqi
->raw_weighted_load
> imbalance
)
2431 if (rqi
->raw_weighted_load
> max_load
) {
2432 max_load
= rqi
->raw_weighted_load
;
2441 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2442 * so long as it is large enough.
2444 #define MAX_PINNED_INTERVAL 512
2446 #define minus_1_or_zero(n) ((n) > 0 ? (n) - 1 : 0)
2448 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2449 * tasks if there is an imbalance.
2451 * Called with this_rq unlocked.
2453 static int load_balance(int this_cpu
, runqueue_t
*this_rq
,
2454 struct sched_domain
*sd
, enum idle_type idle
)
2456 struct sched_group
*group
;
2457 runqueue_t
*busiest
;
2458 unsigned long imbalance
;
2459 int nr_moved
, all_pinned
= 0;
2460 int active_balance
= 0;
2463 if (idle
!= NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2464 !sched_smt_power_savings
)
2467 schedstat_inc(sd
, lb_cnt
[idle
]);
2469 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
);
2471 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2475 busiest
= find_busiest_queue(group
, idle
, imbalance
);
2477 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2481 BUG_ON(busiest
== this_rq
);
2483 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2486 if (busiest
->nr_running
> 1) {
2488 * Attempt to move tasks. If find_busiest_group has found
2489 * an imbalance but busiest->nr_running <= 1, the group is
2490 * still unbalanced. nr_moved simply stays zero, so it is
2491 * correctly treated as an imbalance.
2493 double_rq_lock(this_rq
, busiest
);
2494 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2495 minus_1_or_zero(busiest
->nr_running
),
2496 imbalance
, sd
, idle
, &all_pinned
);
2497 double_rq_unlock(this_rq
, busiest
);
2499 /* All tasks on this runqueue were pinned by CPU affinity */
2500 if (unlikely(all_pinned
))
2505 schedstat_inc(sd
, lb_failed
[idle
]);
2506 sd
->nr_balance_failed
++;
2508 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2510 spin_lock(&busiest
->lock
);
2512 /* don't kick the migration_thread, if the curr
2513 * task on busiest cpu can't be moved to this_cpu
2515 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2516 spin_unlock(&busiest
->lock
);
2518 goto out_one_pinned
;
2521 if (!busiest
->active_balance
) {
2522 busiest
->active_balance
= 1;
2523 busiest
->push_cpu
= this_cpu
;
2526 spin_unlock(&busiest
->lock
);
2528 wake_up_process(busiest
->migration_thread
);
2531 * We've kicked active balancing, reset the failure
2534 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2537 sd
->nr_balance_failed
= 0;
2539 if (likely(!active_balance
)) {
2540 /* We were unbalanced, so reset the balancing interval */
2541 sd
->balance_interval
= sd
->min_interval
;
2544 * If we've begun active balancing, start to back off. This
2545 * case may not be covered by the all_pinned logic if there
2546 * is only 1 task on the busy runqueue (because we don't call
2549 if (sd
->balance_interval
< sd
->max_interval
)
2550 sd
->balance_interval
*= 2;
2553 if (!nr_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2554 !sched_smt_power_savings
)
2559 schedstat_inc(sd
, lb_balanced
[idle
]);
2561 sd
->nr_balance_failed
= 0;
2564 /* tune up the balancing interval */
2565 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2566 (sd
->balance_interval
< sd
->max_interval
))
2567 sd
->balance_interval
*= 2;
2569 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&& !sched_smt_power_savings
)
2575 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2576 * tasks if there is an imbalance.
2578 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2579 * this_rq is locked.
2581 static int load_balance_newidle(int this_cpu
, runqueue_t
*this_rq
,
2582 struct sched_domain
*sd
)
2584 struct sched_group
*group
;
2585 runqueue_t
*busiest
= NULL
;
2586 unsigned long imbalance
;
2590 if (sd
->flags
& SD_SHARE_CPUPOWER
&& !sched_smt_power_savings
)
2593 schedstat_inc(sd
, lb_cnt
[NEWLY_IDLE
]);
2594 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, NEWLY_IDLE
, &sd_idle
);
2596 schedstat_inc(sd
, lb_nobusyg
[NEWLY_IDLE
]);
2600 busiest
= find_busiest_queue(group
, NEWLY_IDLE
, imbalance
);
2602 schedstat_inc(sd
, lb_nobusyq
[NEWLY_IDLE
]);
2606 BUG_ON(busiest
== this_rq
);
2608 schedstat_add(sd
, lb_imbalance
[NEWLY_IDLE
], imbalance
);
2611 if (busiest
->nr_running
> 1) {
2612 /* Attempt to move tasks */
2613 double_lock_balance(this_rq
, busiest
);
2614 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2615 minus_1_or_zero(busiest
->nr_running
),
2616 imbalance
, sd
, NEWLY_IDLE
, NULL
);
2617 spin_unlock(&busiest
->lock
);
2621 schedstat_inc(sd
, lb_failed
[NEWLY_IDLE
]);
2622 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2625 sd
->nr_balance_failed
= 0;
2630 schedstat_inc(sd
, lb_balanced
[NEWLY_IDLE
]);
2631 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&& !sched_smt_power_savings
)
2633 sd
->nr_balance_failed
= 0;
2638 * idle_balance is called by schedule() if this_cpu is about to become
2639 * idle. Attempts to pull tasks from other CPUs.
2641 static void idle_balance(int this_cpu
, runqueue_t
*this_rq
)
2643 struct sched_domain
*sd
;
2645 for_each_domain(this_cpu
, sd
) {
2646 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
2647 if (load_balance_newidle(this_cpu
, this_rq
, sd
)) {
2648 /* We've pulled tasks over so stop searching */
2656 * active_load_balance is run by migration threads. It pushes running tasks
2657 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2658 * running on each physical CPU where possible, and avoids physical /
2659 * logical imbalances.
2661 * Called with busiest_rq locked.
2663 static void active_load_balance(runqueue_t
*busiest_rq
, int busiest_cpu
)
2665 struct sched_domain
*sd
;
2666 runqueue_t
*target_rq
;
2667 int target_cpu
= busiest_rq
->push_cpu
;
2669 if (busiest_rq
->nr_running
<= 1)
2670 /* no task to move */
2673 target_rq
= cpu_rq(target_cpu
);
2676 * This condition is "impossible", if it occurs
2677 * we need to fix it. Originally reported by
2678 * Bjorn Helgaas on a 128-cpu setup.
2680 BUG_ON(busiest_rq
== target_rq
);
2682 /* move a task from busiest_rq to target_rq */
2683 double_lock_balance(busiest_rq
, target_rq
);
2685 /* Search for an sd spanning us and the target CPU. */
2686 for_each_domain(target_cpu
, sd
) {
2687 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2688 cpu_isset(busiest_cpu
, sd
->span
))
2692 if (unlikely(sd
== NULL
))
2695 schedstat_inc(sd
, alb_cnt
);
2697 if (move_tasks(target_rq
, target_cpu
, busiest_rq
, 1,
2698 RTPRIO_TO_LOAD_WEIGHT(100), sd
, SCHED_IDLE
, NULL
))
2699 schedstat_inc(sd
, alb_pushed
);
2701 schedstat_inc(sd
, alb_failed
);
2703 spin_unlock(&target_rq
->lock
);
2707 * rebalance_tick will get called every timer tick, on every CPU.
2709 * It checks each scheduling domain to see if it is due to be balanced,
2710 * and initiates a balancing operation if so.
2712 * Balancing parameters are set up in arch_init_sched_domains.
2715 /* Don't have all balancing operations going off at once */
2716 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2718 static void rebalance_tick(int this_cpu
, runqueue_t
*this_rq
,
2719 enum idle_type idle
)
2721 unsigned long old_load
, this_load
;
2722 unsigned long j
= jiffies
+ CPU_OFFSET(this_cpu
);
2723 struct sched_domain
*sd
;
2726 this_load
= this_rq
->raw_weighted_load
;
2727 /* Update our load */
2728 for (i
= 0; i
< 3; i
++) {
2729 unsigned long new_load
= this_load
;
2731 old_load
= this_rq
->cpu_load
[i
];
2733 * Round up the averaging division if load is increasing. This
2734 * prevents us from getting stuck on 9 if the load is 10, for
2737 if (new_load
> old_load
)
2738 new_load
+= scale
-1;
2739 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) / scale
;
2742 for_each_domain(this_cpu
, sd
) {
2743 unsigned long interval
;
2745 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2748 interval
= sd
->balance_interval
;
2749 if (idle
!= SCHED_IDLE
)
2750 interval
*= sd
->busy_factor
;
2752 /* scale ms to jiffies */
2753 interval
= msecs_to_jiffies(interval
);
2754 if (unlikely(!interval
))
2757 if (j
- sd
->last_balance
>= interval
) {
2758 if (load_balance(this_cpu
, this_rq
, sd
, idle
)) {
2760 * We've pulled tasks over so either we're no
2761 * longer idle, or one of our SMT siblings is
2766 sd
->last_balance
+= interval
;
2772 * on UP we do not need to balance between CPUs:
2774 static inline void rebalance_tick(int cpu
, runqueue_t
*rq
, enum idle_type idle
)
2777 static inline void idle_balance(int cpu
, runqueue_t
*rq
)
2782 static inline int wake_priority_sleeper(runqueue_t
*rq
)
2785 #ifdef CONFIG_SCHED_SMT
2786 spin_lock(&rq
->lock
);
2788 * If an SMT sibling task has been put to sleep for priority
2789 * reasons reschedule the idle task to see if it can now run.
2791 if (rq
->nr_running
) {
2792 resched_task(rq
->idle
);
2795 spin_unlock(&rq
->lock
);
2800 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2802 EXPORT_PER_CPU_SYMBOL(kstat
);
2805 * This is called on clock ticks and on context switches.
2806 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2808 static inline void update_cpu_clock(task_t
*p
, runqueue_t
*rq
,
2809 unsigned long long now
)
2811 unsigned long long last
= max(p
->timestamp
, rq
->timestamp_last_tick
);
2812 p
->sched_time
+= now
- last
;
2816 * Return current->sched_time plus any more ns on the sched_clock
2817 * that have not yet been banked.
2819 unsigned long long current_sched_time(const task_t
*tsk
)
2821 unsigned long long ns
;
2822 unsigned long flags
;
2823 local_irq_save(flags
);
2824 ns
= max(tsk
->timestamp
, task_rq(tsk
)->timestamp_last_tick
);
2825 ns
= tsk
->sched_time
+ (sched_clock() - ns
);
2826 local_irq_restore(flags
);
2831 * We place interactive tasks back into the active array, if possible.
2833 * To guarantee that this does not starve expired tasks we ignore the
2834 * interactivity of a task if the first expired task had to wait more
2835 * than a 'reasonable' amount of time. This deadline timeout is
2836 * load-dependent, as the frequency of array switched decreases with
2837 * increasing number of running tasks. We also ignore the interactivity
2838 * if a better static_prio task has expired:
2840 #define EXPIRED_STARVING(rq) \
2841 ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2842 (jiffies - (rq)->expired_timestamp >= \
2843 STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2844 ((rq)->curr->static_prio > (rq)->best_expired_prio))
2847 * Account user cpu time to a process.
2848 * @p: the process that the cpu time gets accounted to
2849 * @hardirq_offset: the offset to subtract from hardirq_count()
2850 * @cputime: the cpu time spent in user space since the last update
2852 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
2854 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2857 p
->utime
= cputime_add(p
->utime
, cputime
);
2859 /* Add user time to cpustat. */
2860 tmp
= cputime_to_cputime64(cputime
);
2861 if (TASK_NICE(p
) > 0)
2862 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
2864 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
2868 * Account system cpu time to a process.
2869 * @p: the process that the cpu time gets accounted to
2870 * @hardirq_offset: the offset to subtract from hardirq_count()
2871 * @cputime: the cpu time spent in kernel space since the last update
2873 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
2876 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2877 runqueue_t
*rq
= this_rq();
2880 p
->stime
= cputime_add(p
->stime
, cputime
);
2882 /* Add system time to cpustat. */
2883 tmp
= cputime_to_cputime64(cputime
);
2884 if (hardirq_count() - hardirq_offset
)
2885 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
2886 else if (softirq_count())
2887 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
2888 else if (p
!= rq
->idle
)
2889 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
2890 else if (atomic_read(&rq
->nr_iowait
) > 0)
2891 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
2893 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
2894 /* Account for system time used */
2895 acct_update_integrals(p
);
2899 * Account for involuntary wait time.
2900 * @p: the process from which the cpu time has been stolen
2901 * @steal: the cpu time spent in involuntary wait
2903 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
2905 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2906 cputime64_t tmp
= cputime_to_cputime64(steal
);
2907 runqueue_t
*rq
= this_rq();
2909 if (p
== rq
->idle
) {
2910 p
->stime
= cputime_add(p
->stime
, steal
);
2911 if (atomic_read(&rq
->nr_iowait
) > 0)
2912 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
2914 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
2916 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
2920 * This function gets called by the timer code, with HZ frequency.
2921 * We call it with interrupts disabled.
2923 * It also gets called by the fork code, when changing the parent's
2926 void scheduler_tick(void)
2928 int cpu
= smp_processor_id();
2929 runqueue_t
*rq
= this_rq();
2930 task_t
*p
= current
;
2931 unsigned long long now
= sched_clock();
2933 update_cpu_clock(p
, rq
, now
);
2935 rq
->timestamp_last_tick
= now
;
2937 if (p
== rq
->idle
) {
2938 if (wake_priority_sleeper(rq
))
2940 rebalance_tick(cpu
, rq
, SCHED_IDLE
);
2944 /* Task might have expired already, but not scheduled off yet */
2945 if (p
->array
!= rq
->active
) {
2946 set_tsk_need_resched(p
);
2949 spin_lock(&rq
->lock
);
2951 * The task was running during this tick - update the
2952 * time slice counter. Note: we do not update a thread's
2953 * priority until it either goes to sleep or uses up its
2954 * timeslice. This makes it possible for interactive tasks
2955 * to use up their timeslices at their highest priority levels.
2959 * RR tasks need a special form of timeslice management.
2960 * FIFO tasks have no timeslices.
2962 if ((p
->policy
== SCHED_RR
) && !--p
->time_slice
) {
2963 p
->time_slice
= task_timeslice(p
);
2964 p
->first_time_slice
= 0;
2965 set_tsk_need_resched(p
);
2967 /* put it at the end of the queue: */
2968 requeue_task(p
, rq
->active
);
2972 if (!--p
->time_slice
) {
2973 dequeue_task(p
, rq
->active
);
2974 set_tsk_need_resched(p
);
2975 p
->prio
= effective_prio(p
);
2976 p
->time_slice
= task_timeslice(p
);
2977 p
->first_time_slice
= 0;
2979 if (!rq
->expired_timestamp
)
2980 rq
->expired_timestamp
= jiffies
;
2981 if (!TASK_INTERACTIVE(p
) || EXPIRED_STARVING(rq
)) {
2982 enqueue_task(p
, rq
->expired
);
2983 if (p
->static_prio
< rq
->best_expired_prio
)
2984 rq
->best_expired_prio
= p
->static_prio
;
2986 enqueue_task(p
, rq
->active
);
2989 * Prevent a too long timeslice allowing a task to monopolize
2990 * the CPU. We do this by splitting up the timeslice into
2993 * Note: this does not mean the task's timeslices expire or
2994 * get lost in any way, they just might be preempted by
2995 * another task of equal priority. (one with higher
2996 * priority would have preempted this task already.) We
2997 * requeue this task to the end of the list on this priority
2998 * level, which is in essence a round-robin of tasks with
3001 * This only applies to tasks in the interactive
3002 * delta range with at least TIMESLICE_GRANULARITY to requeue.
3004 if (TASK_INTERACTIVE(p
) && !((task_timeslice(p
) -
3005 p
->time_slice
) % TIMESLICE_GRANULARITY(p
)) &&
3006 (p
->time_slice
>= TIMESLICE_GRANULARITY(p
)) &&
3007 (p
->array
== rq
->active
)) {
3009 requeue_task(p
, rq
->active
);
3010 set_tsk_need_resched(p
);
3014 spin_unlock(&rq
->lock
);
3016 rebalance_tick(cpu
, rq
, NOT_IDLE
);
3019 #ifdef CONFIG_SCHED_SMT
3020 static inline void wakeup_busy_runqueue(runqueue_t
*rq
)
3022 /* If an SMT runqueue is sleeping due to priority reasons wake it up */
3023 if (rq
->curr
== rq
->idle
&& rq
->nr_running
)
3024 resched_task(rq
->idle
);
3028 * Called with interrupt disabled and this_rq's runqueue locked.
3030 static void wake_sleeping_dependent(int this_cpu
)
3032 struct sched_domain
*tmp
, *sd
= NULL
;
3035 for_each_domain(this_cpu
, tmp
) {
3036 if (tmp
->flags
& SD_SHARE_CPUPOWER
) {
3045 for_each_cpu_mask(i
, sd
->span
) {
3046 runqueue_t
*smt_rq
= cpu_rq(i
);
3050 if (unlikely(!spin_trylock(&smt_rq
->lock
)))
3053 wakeup_busy_runqueue(smt_rq
);
3054 spin_unlock(&smt_rq
->lock
);
3059 * number of 'lost' timeslices this task wont be able to fully
3060 * utilize, if another task runs on a sibling. This models the
3061 * slowdown effect of other tasks running on siblings:
3063 static inline unsigned long smt_slice(task_t
*p
, struct sched_domain
*sd
)
3065 return p
->time_slice
* (100 - sd
->per_cpu_gain
) / 100;
3069 * To minimise lock contention and not have to drop this_rq's runlock we only
3070 * trylock the sibling runqueues and bypass those runqueues if we fail to
3071 * acquire their lock. As we only trylock the normal locking order does not
3072 * need to be obeyed.
3074 static int dependent_sleeper(int this_cpu
, runqueue_t
*this_rq
, task_t
*p
)
3076 struct sched_domain
*tmp
, *sd
= NULL
;
3079 /* kernel/rt threads do not participate in dependent sleeping */
3080 if (!p
->mm
|| rt_task(p
))
3083 for_each_domain(this_cpu
, tmp
) {
3084 if (tmp
->flags
& SD_SHARE_CPUPOWER
) {
3093 for_each_cpu_mask(i
, sd
->span
) {
3101 if (unlikely(!spin_trylock(&smt_rq
->lock
)))
3104 smt_curr
= smt_rq
->curr
;
3110 * If a user task with lower static priority than the
3111 * running task on the SMT sibling is trying to schedule,
3112 * delay it till there is proportionately less timeslice
3113 * left of the sibling task to prevent a lower priority
3114 * task from using an unfair proportion of the
3115 * physical cpu's resources. -ck
3117 if (rt_task(smt_curr
)) {
3119 * With real time tasks we run non-rt tasks only
3120 * per_cpu_gain% of the time.
3122 if ((jiffies
% DEF_TIMESLICE
) >
3123 (sd
->per_cpu_gain
* DEF_TIMESLICE
/ 100))
3126 if (smt_curr
->static_prio
< p
->static_prio
&&
3127 !TASK_PREEMPTS_CURR(p
, smt_rq
) &&
3128 smt_slice(smt_curr
, sd
) > task_timeslice(p
))
3132 spin_unlock(&smt_rq
->lock
);
3137 static inline void wake_sleeping_dependent(int this_cpu
)
3141 static inline int dependent_sleeper(int this_cpu
, runqueue_t
*this_rq
,
3148 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3150 void fastcall
add_preempt_count(int val
)
3155 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3157 preempt_count() += val
;
3159 * Spinlock count overflowing soon?
3161 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >= PREEMPT_MASK
-10);
3163 EXPORT_SYMBOL(add_preempt_count
);
3165 void fastcall
sub_preempt_count(int val
)
3170 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3173 * Is the spinlock portion underflowing?
3175 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3176 !(preempt_count() & PREEMPT_MASK
)))
3179 preempt_count() -= val
;
3181 EXPORT_SYMBOL(sub_preempt_count
);
3185 static inline int interactive_sleep(enum sleep_type sleep_type
)
3187 return (sleep_type
== SLEEP_INTERACTIVE
||
3188 sleep_type
== SLEEP_INTERRUPTED
);
3192 * schedule() is the main scheduler function.
3194 asmlinkage
void __sched
schedule(void)
3197 task_t
*prev
, *next
;
3199 prio_array_t
*array
;
3200 struct list_head
*queue
;
3201 unsigned long long now
;
3202 unsigned long run_time
;
3203 int cpu
, idx
, new_prio
;
3206 * Test if we are atomic. Since do_exit() needs to call into
3207 * schedule() atomically, we ignore that path for now.
3208 * Otherwise, whine if we are scheduling when we should not be.
3210 if (unlikely(in_atomic() && !current
->exit_state
)) {
3211 printk(KERN_ERR
"BUG: scheduling while atomic: "
3213 current
->comm
, preempt_count(), current
->pid
);
3216 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3221 release_kernel_lock(prev
);
3222 need_resched_nonpreemptible
:
3226 * The idle thread is not allowed to schedule!
3227 * Remove this check after it has been exercised a bit.
3229 if (unlikely(prev
== rq
->idle
) && prev
->state
!= TASK_RUNNING
) {
3230 printk(KERN_ERR
"bad: scheduling from the idle thread!\n");
3234 schedstat_inc(rq
, sched_cnt
);
3235 now
= sched_clock();
3236 if (likely((long long)(now
- prev
->timestamp
) < NS_MAX_SLEEP_AVG
)) {
3237 run_time
= now
- prev
->timestamp
;
3238 if (unlikely((long long)(now
- prev
->timestamp
) < 0))
3241 run_time
= NS_MAX_SLEEP_AVG
;
3244 * Tasks charged proportionately less run_time at high sleep_avg to
3245 * delay them losing their interactive status
3247 run_time
/= (CURRENT_BONUS(prev
) ? : 1);
3249 spin_lock_irq(&rq
->lock
);
3251 if (unlikely(prev
->flags
& PF_DEAD
))
3252 prev
->state
= EXIT_DEAD
;
3254 switch_count
= &prev
->nivcsw
;
3255 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3256 switch_count
= &prev
->nvcsw
;
3257 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3258 unlikely(signal_pending(prev
))))
3259 prev
->state
= TASK_RUNNING
;
3261 if (prev
->state
== TASK_UNINTERRUPTIBLE
)
3262 rq
->nr_uninterruptible
++;
3263 deactivate_task(prev
, rq
);
3267 cpu
= smp_processor_id();
3268 if (unlikely(!rq
->nr_running
)) {
3269 idle_balance(cpu
, rq
);
3270 if (!rq
->nr_running
) {
3272 rq
->expired_timestamp
= 0;
3273 wake_sleeping_dependent(cpu
);
3279 if (unlikely(!array
->nr_active
)) {
3281 * Switch the active and expired arrays.
3283 schedstat_inc(rq
, sched_switch
);
3284 rq
->active
= rq
->expired
;
3285 rq
->expired
= array
;
3287 rq
->expired_timestamp
= 0;
3288 rq
->best_expired_prio
= MAX_PRIO
;
3291 idx
= sched_find_first_bit(array
->bitmap
);
3292 queue
= array
->queue
+ idx
;
3293 next
= list_entry(queue
->next
, task_t
, run_list
);
3295 if (!rt_task(next
) && interactive_sleep(next
->sleep_type
)) {
3296 unsigned long long delta
= now
- next
->timestamp
;
3297 if (unlikely((long long)(now
- next
->timestamp
) < 0))
3300 if (next
->sleep_type
== SLEEP_INTERACTIVE
)
3301 delta
= delta
* (ON_RUNQUEUE_WEIGHT
* 128 / 100) / 128;
3303 array
= next
->array
;
3304 new_prio
= recalc_task_prio(next
, next
->timestamp
+ delta
);
3306 if (unlikely(next
->prio
!= new_prio
)) {
3307 dequeue_task(next
, array
);
3308 next
->prio
= new_prio
;
3309 enqueue_task(next
, array
);
3312 next
->sleep_type
= SLEEP_NORMAL
;
3313 if (dependent_sleeper(cpu
, rq
, next
))
3316 if (next
== rq
->idle
)
3317 schedstat_inc(rq
, sched_goidle
);
3319 prefetch_stack(next
);
3320 clear_tsk_need_resched(prev
);
3321 rcu_qsctr_inc(task_cpu(prev
));
3323 update_cpu_clock(prev
, rq
, now
);
3325 prev
->sleep_avg
-= run_time
;
3326 if ((long)prev
->sleep_avg
<= 0)
3327 prev
->sleep_avg
= 0;
3328 prev
->timestamp
= prev
->last_ran
= now
;
3330 sched_info_switch(prev
, next
);
3331 if (likely(prev
!= next
)) {
3332 next
->timestamp
= now
;
3337 prepare_task_switch(rq
, next
);
3338 prev
= context_switch(rq
, prev
, next
);
3341 * this_rq must be evaluated again because prev may have moved
3342 * CPUs since it called schedule(), thus the 'rq' on its stack
3343 * frame will be invalid.
3345 finish_task_switch(this_rq(), prev
);
3347 spin_unlock_irq(&rq
->lock
);
3350 if (unlikely(reacquire_kernel_lock(prev
) < 0))
3351 goto need_resched_nonpreemptible
;
3352 preempt_enable_no_resched();
3353 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3357 EXPORT_SYMBOL(schedule
);
3359 #ifdef CONFIG_PREEMPT
3361 * this is is the entry point to schedule() from in-kernel preemption
3362 * off of preempt_enable. Kernel preemptions off return from interrupt
3363 * occur there and call schedule directly.
3365 asmlinkage
void __sched
preempt_schedule(void)
3367 struct thread_info
*ti
= current_thread_info();
3368 #ifdef CONFIG_PREEMPT_BKL
3369 struct task_struct
*task
= current
;
3370 int saved_lock_depth
;
3373 * If there is a non-zero preempt_count or interrupts are disabled,
3374 * we do not want to preempt the current task. Just return..
3376 if (unlikely(ti
->preempt_count
|| irqs_disabled()))
3380 add_preempt_count(PREEMPT_ACTIVE
);
3382 * We keep the big kernel semaphore locked, but we
3383 * clear ->lock_depth so that schedule() doesnt
3384 * auto-release the semaphore:
3386 #ifdef CONFIG_PREEMPT_BKL
3387 saved_lock_depth
= task
->lock_depth
;
3388 task
->lock_depth
= -1;
3391 #ifdef CONFIG_PREEMPT_BKL
3392 task
->lock_depth
= saved_lock_depth
;
3394 sub_preempt_count(PREEMPT_ACTIVE
);
3396 /* we could miss a preemption opportunity between schedule and now */
3398 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3402 EXPORT_SYMBOL(preempt_schedule
);
3405 * this is is the entry point to schedule() from kernel preemption
3406 * off of irq context.
3407 * Note, that this is called and return with irqs disabled. This will
3408 * protect us against recursive calling from irq.
3410 asmlinkage
void __sched
preempt_schedule_irq(void)
3412 struct thread_info
*ti
= current_thread_info();
3413 #ifdef CONFIG_PREEMPT_BKL
3414 struct task_struct
*task
= current
;
3415 int saved_lock_depth
;
3417 /* Catch callers which need to be fixed*/
3418 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3421 add_preempt_count(PREEMPT_ACTIVE
);
3423 * We keep the big kernel semaphore locked, but we
3424 * clear ->lock_depth so that schedule() doesnt
3425 * auto-release the semaphore:
3427 #ifdef CONFIG_PREEMPT_BKL
3428 saved_lock_depth
= task
->lock_depth
;
3429 task
->lock_depth
= -1;
3433 local_irq_disable();
3434 #ifdef CONFIG_PREEMPT_BKL
3435 task
->lock_depth
= saved_lock_depth
;
3437 sub_preempt_count(PREEMPT_ACTIVE
);
3439 /* we could miss a preemption opportunity between schedule and now */
3441 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3445 #endif /* CONFIG_PREEMPT */
3447 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3450 task_t
*p
= curr
->private;
3451 return try_to_wake_up(p
, mode
, sync
);
3454 EXPORT_SYMBOL(default_wake_function
);
3457 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3458 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3459 * number) then we wake all the non-exclusive tasks and one exclusive task.
3461 * There are circumstances in which we can try to wake a task which has already
3462 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3463 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3465 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3466 int nr_exclusive
, int sync
, void *key
)
3468 struct list_head
*tmp
, *next
;
3470 list_for_each_safe(tmp
, next
, &q
->task_list
) {
3473 curr
= list_entry(tmp
, wait_queue_t
, task_list
);
3474 flags
= curr
->flags
;
3475 if (curr
->func(curr
, mode
, sync
, key
) &&
3476 (flags
& WQ_FLAG_EXCLUSIVE
) &&
3483 * __wake_up - wake up threads blocked on a waitqueue.
3485 * @mode: which threads
3486 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3487 * @key: is directly passed to the wakeup function
3489 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3490 int nr_exclusive
, void *key
)
3492 unsigned long flags
;
3494 spin_lock_irqsave(&q
->lock
, flags
);
3495 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3496 spin_unlock_irqrestore(&q
->lock
, flags
);
3499 EXPORT_SYMBOL(__wake_up
);
3502 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3504 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3506 __wake_up_common(q
, mode
, 1, 0, NULL
);
3510 * __wake_up_sync - wake up threads blocked on a waitqueue.
3512 * @mode: which threads
3513 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3515 * The sync wakeup differs that the waker knows that it will schedule
3516 * away soon, so while the target thread will be woken up, it will not
3517 * be migrated to another CPU - ie. the two threads are 'synchronized'
3518 * with each other. This can prevent needless bouncing between CPUs.
3520 * On UP it can prevent extra preemption.
3523 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3525 unsigned long flags
;
3531 if (unlikely(!nr_exclusive
))
3534 spin_lock_irqsave(&q
->lock
, flags
);
3535 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3536 spin_unlock_irqrestore(&q
->lock
, flags
);
3538 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3540 void fastcall
complete(struct completion
*x
)
3542 unsigned long flags
;
3544 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3546 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3548 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3550 EXPORT_SYMBOL(complete
);
3552 void fastcall
complete_all(struct completion
*x
)
3554 unsigned long flags
;
3556 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3557 x
->done
+= UINT_MAX
/2;
3558 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3560 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3562 EXPORT_SYMBOL(complete_all
);
3564 void fastcall __sched
wait_for_completion(struct completion
*x
)
3567 spin_lock_irq(&x
->wait
.lock
);
3569 DECLARE_WAITQUEUE(wait
, current
);
3571 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3572 __add_wait_queue_tail(&x
->wait
, &wait
);
3574 __set_current_state(TASK_UNINTERRUPTIBLE
);
3575 spin_unlock_irq(&x
->wait
.lock
);
3577 spin_lock_irq(&x
->wait
.lock
);
3579 __remove_wait_queue(&x
->wait
, &wait
);
3582 spin_unlock_irq(&x
->wait
.lock
);
3584 EXPORT_SYMBOL(wait_for_completion
);
3586 unsigned long fastcall __sched
3587 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3591 spin_lock_irq(&x
->wait
.lock
);
3593 DECLARE_WAITQUEUE(wait
, current
);
3595 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3596 __add_wait_queue_tail(&x
->wait
, &wait
);
3598 __set_current_state(TASK_UNINTERRUPTIBLE
);
3599 spin_unlock_irq(&x
->wait
.lock
);
3600 timeout
= schedule_timeout(timeout
);
3601 spin_lock_irq(&x
->wait
.lock
);
3603 __remove_wait_queue(&x
->wait
, &wait
);
3607 __remove_wait_queue(&x
->wait
, &wait
);
3611 spin_unlock_irq(&x
->wait
.lock
);
3614 EXPORT_SYMBOL(wait_for_completion_timeout
);
3616 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3622 spin_lock_irq(&x
->wait
.lock
);
3624 DECLARE_WAITQUEUE(wait
, current
);
3626 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3627 __add_wait_queue_tail(&x
->wait
, &wait
);
3629 if (signal_pending(current
)) {
3631 __remove_wait_queue(&x
->wait
, &wait
);
3634 __set_current_state(TASK_INTERRUPTIBLE
);
3635 spin_unlock_irq(&x
->wait
.lock
);
3637 spin_lock_irq(&x
->wait
.lock
);
3639 __remove_wait_queue(&x
->wait
, &wait
);
3643 spin_unlock_irq(&x
->wait
.lock
);
3647 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3649 unsigned long fastcall __sched
3650 wait_for_completion_interruptible_timeout(struct completion
*x
,
3651 unsigned long timeout
)
3655 spin_lock_irq(&x
->wait
.lock
);
3657 DECLARE_WAITQUEUE(wait
, current
);
3659 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3660 __add_wait_queue_tail(&x
->wait
, &wait
);
3662 if (signal_pending(current
)) {
3663 timeout
= -ERESTARTSYS
;
3664 __remove_wait_queue(&x
->wait
, &wait
);
3667 __set_current_state(TASK_INTERRUPTIBLE
);
3668 spin_unlock_irq(&x
->wait
.lock
);
3669 timeout
= schedule_timeout(timeout
);
3670 spin_lock_irq(&x
->wait
.lock
);
3672 __remove_wait_queue(&x
->wait
, &wait
);
3676 __remove_wait_queue(&x
->wait
, &wait
);
3680 spin_unlock_irq(&x
->wait
.lock
);
3683 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3686 #define SLEEP_ON_VAR \
3687 unsigned long flags; \
3688 wait_queue_t wait; \
3689 init_waitqueue_entry(&wait, current);
3691 #define SLEEP_ON_HEAD \
3692 spin_lock_irqsave(&q->lock,flags); \
3693 __add_wait_queue(q, &wait); \
3694 spin_unlock(&q->lock);
3696 #define SLEEP_ON_TAIL \
3697 spin_lock_irq(&q->lock); \
3698 __remove_wait_queue(q, &wait); \
3699 spin_unlock_irqrestore(&q->lock, flags);
3701 void fastcall __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3705 current
->state
= TASK_INTERRUPTIBLE
;
3712 EXPORT_SYMBOL(interruptible_sleep_on
);
3714 long fastcall __sched
3715 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3719 current
->state
= TASK_INTERRUPTIBLE
;
3722 timeout
= schedule_timeout(timeout
);
3728 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3730 void fastcall __sched
sleep_on(wait_queue_head_t
*q
)
3734 current
->state
= TASK_UNINTERRUPTIBLE
;
3741 EXPORT_SYMBOL(sleep_on
);
3743 long fastcall __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3747 current
->state
= TASK_UNINTERRUPTIBLE
;
3750 timeout
= schedule_timeout(timeout
);
3756 EXPORT_SYMBOL(sleep_on_timeout
);
3758 #ifdef CONFIG_RT_MUTEXES
3761 * rt_mutex_setprio - set the current priority of a task
3763 * @prio: prio value (kernel-internal form)
3765 * This function changes the 'effective' priority of a task. It does
3766 * not touch ->normal_prio like __setscheduler().
3768 * Used by the rt_mutex code to implement priority inheritance logic.
3770 void rt_mutex_setprio(task_t
*p
, int prio
)
3772 unsigned long flags
;
3773 prio_array_t
*array
;
3777 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3779 rq
= task_rq_lock(p
, &flags
);
3784 dequeue_task(p
, array
);
3789 * If changing to an RT priority then queue it
3790 * in the active array!
3794 enqueue_task(p
, array
);
3796 * Reschedule if we are currently running on this runqueue and
3797 * our priority decreased, or if we are not currently running on
3798 * this runqueue and our priority is higher than the current's
3800 if (task_running(rq
, p
)) {
3801 if (p
->prio
> oldprio
)
3802 resched_task(rq
->curr
);
3803 } else if (TASK_PREEMPTS_CURR(p
, rq
))
3804 resched_task(rq
->curr
);
3806 task_rq_unlock(rq
, &flags
);
3811 void set_user_nice(task_t
*p
, long nice
)
3813 unsigned long flags
;
3814 prio_array_t
*array
;
3816 int old_prio
, delta
;
3818 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3821 * We have to be careful, if called from sys_setpriority(),
3822 * the task might be in the middle of scheduling on another CPU.
3824 rq
= task_rq_lock(p
, &flags
);
3826 * The RT priorities are set via sched_setscheduler(), but we still
3827 * allow the 'normal' nice value to be set - but as expected
3828 * it wont have any effect on scheduling until the task is
3829 * not SCHED_NORMAL/SCHED_BATCH:
3831 if (has_rt_policy(p
)) {
3832 p
->static_prio
= NICE_TO_PRIO(nice
);
3837 dequeue_task(p
, array
);
3838 dec_raw_weighted_load(rq
, p
);
3841 p
->static_prio
= NICE_TO_PRIO(nice
);
3844 p
->prio
= effective_prio(p
);
3845 delta
= p
->prio
- old_prio
;
3848 enqueue_task(p
, array
);
3849 inc_raw_weighted_load(rq
, p
);
3851 * If the task increased its priority or is running and
3852 * lowered its priority, then reschedule its CPU:
3854 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3855 resched_task(rq
->curr
);
3858 task_rq_unlock(rq
, &flags
);
3860 EXPORT_SYMBOL(set_user_nice
);
3863 * can_nice - check if a task can reduce its nice value
3867 int can_nice(const task_t
*p
, const int nice
)
3869 /* convert nice value [19,-20] to rlimit style value [1,40] */
3870 int nice_rlim
= 20 - nice
;
3871 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
3872 capable(CAP_SYS_NICE
));
3875 #ifdef __ARCH_WANT_SYS_NICE
3878 * sys_nice - change the priority of the current process.
3879 * @increment: priority increment
3881 * sys_setpriority is a more generic, but much slower function that
3882 * does similar things.
3884 asmlinkage
long sys_nice(int increment
)
3890 * Setpriority might change our priority at the same moment.
3891 * We don't have to worry. Conceptually one call occurs first
3892 * and we have a single winner.
3894 if (increment
< -40)
3899 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
3905 if (increment
< 0 && !can_nice(current
, nice
))
3908 retval
= security_task_setnice(current
, nice
);
3912 set_user_nice(current
, nice
);
3919 * task_prio - return the priority value of a given task.
3920 * @p: the task in question.
3922 * This is the priority value as seen by users in /proc.
3923 * RT tasks are offset by -200. Normal tasks are centered
3924 * around 0, value goes from -16 to +15.
3926 int task_prio(const task_t
*p
)
3928 return p
->prio
- MAX_RT_PRIO
;
3932 * task_nice - return the nice value of a given task.
3933 * @p: the task in question.
3935 int task_nice(const task_t
*p
)
3937 return TASK_NICE(p
);
3939 EXPORT_SYMBOL_GPL(task_nice
);
3942 * idle_cpu - is a given cpu idle currently?
3943 * @cpu: the processor in question.
3945 int idle_cpu(int cpu
)
3947 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
3951 * idle_task - return the idle task for a given cpu.
3952 * @cpu: the processor in question.
3954 task_t
*idle_task(int cpu
)
3956 return cpu_rq(cpu
)->idle
;
3960 * find_process_by_pid - find a process with a matching PID value.
3961 * @pid: the pid in question.
3963 static inline task_t
*find_process_by_pid(pid_t pid
)
3965 return pid
? find_task_by_pid(pid
) : current
;
3968 /* Actually do priority change: must hold rq lock. */
3969 static void __setscheduler(struct task_struct
*p
, int policy
, int prio
)
3973 p
->rt_priority
= prio
;
3974 p
->normal_prio
= normal_prio(p
);
3975 /* we are holding p->pi_lock already */
3976 p
->prio
= rt_mutex_getprio(p
);
3978 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
3980 if (policy
== SCHED_BATCH
)
3986 * sched_setscheduler - change the scheduling policy and/or RT priority of
3988 * @p: the task in question.
3989 * @policy: new policy.
3990 * @param: structure containing the new RT priority.
3992 int sched_setscheduler(struct task_struct
*p
, int policy
,
3993 struct sched_param
*param
)
3996 int oldprio
, oldpolicy
= -1;
3997 prio_array_t
*array
;
3998 unsigned long flags
;
4001 /* may grab non-irq protected spin_locks */
4002 BUG_ON(in_interrupt());
4004 /* double check policy once rq lock held */
4006 policy
= oldpolicy
= p
->policy
;
4007 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4008 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
)
4011 * Valid priorities for SCHED_FIFO and SCHED_RR are
4012 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
4015 if (param
->sched_priority
< 0 ||
4016 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4017 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4019 if ((policy
== SCHED_NORMAL
|| policy
== SCHED_BATCH
)
4020 != (param
->sched_priority
== 0))
4024 * Allow unprivileged RT tasks to decrease priority:
4026 if (!capable(CAP_SYS_NICE
)) {
4028 * can't change policy, except between SCHED_NORMAL
4031 if (((policy
!= SCHED_NORMAL
&& p
->policy
!= SCHED_BATCH
) &&
4032 (policy
!= SCHED_BATCH
&& p
->policy
!= SCHED_NORMAL
)) &&
4033 !p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
)
4035 /* can't increase priority */
4036 if ((policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
) &&
4037 param
->sched_priority
> p
->rt_priority
&&
4038 param
->sched_priority
>
4039 p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
)
4041 /* can't change other user's priorities */
4042 if ((current
->euid
!= p
->euid
) &&
4043 (current
->euid
!= p
->uid
))
4047 retval
= security_task_setscheduler(p
, policy
, param
);
4051 * make sure no PI-waiters arrive (or leave) while we are
4052 * changing the priority of the task:
4054 spin_lock_irqsave(&p
->pi_lock
, flags
);
4056 * To be able to change p->policy safely, the apropriate
4057 * runqueue lock must be held.
4059 rq
= __task_rq_lock(p
);
4060 /* recheck policy now with rq lock held */
4061 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4062 policy
= oldpolicy
= -1;
4063 __task_rq_unlock(rq
);
4064 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4069 deactivate_task(p
, rq
);
4071 __setscheduler(p
, policy
, param
->sched_priority
);
4073 __activate_task(p
, rq
);
4075 * Reschedule if we are currently running on this runqueue and
4076 * our priority decreased, or if we are not currently running on
4077 * this runqueue and our priority is higher than the current's
4079 if (task_running(rq
, p
)) {
4080 if (p
->prio
> oldprio
)
4081 resched_task(rq
->curr
);
4082 } else if (TASK_PREEMPTS_CURR(p
, rq
))
4083 resched_task(rq
->curr
);
4085 __task_rq_unlock(rq
);
4086 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4088 rt_mutex_adjust_pi(p
);
4092 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4095 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4098 struct sched_param lparam
;
4099 struct task_struct
*p
;
4101 if (!param
|| pid
< 0)
4103 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4105 read_lock_irq(&tasklist_lock
);
4106 p
= find_process_by_pid(pid
);
4108 read_unlock_irq(&tasklist_lock
);
4112 read_unlock_irq(&tasklist_lock
);
4113 retval
= sched_setscheduler(p
, policy
, &lparam
);
4119 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4120 * @pid: the pid in question.
4121 * @policy: new policy.
4122 * @param: structure containing the new RT priority.
4124 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
4125 struct sched_param __user
*param
)
4127 /* negative values for policy are not valid */
4131 return do_sched_setscheduler(pid
, policy
, param
);
4135 * sys_sched_setparam - set/change the RT priority of a thread
4136 * @pid: the pid in question.
4137 * @param: structure containing the new RT priority.
4139 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4141 return do_sched_setscheduler(pid
, -1, param
);
4145 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4146 * @pid: the pid in question.
4148 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4150 int retval
= -EINVAL
;
4157 read_lock(&tasklist_lock
);
4158 p
= find_process_by_pid(pid
);
4160 retval
= security_task_getscheduler(p
);
4164 read_unlock(&tasklist_lock
);
4171 * sys_sched_getscheduler - get the RT priority of a thread
4172 * @pid: the pid in question.
4173 * @param: structure containing the RT priority.
4175 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4177 struct sched_param lp
;
4178 int retval
= -EINVAL
;
4181 if (!param
|| pid
< 0)
4184 read_lock(&tasklist_lock
);
4185 p
= find_process_by_pid(pid
);
4190 retval
= security_task_getscheduler(p
);
4194 lp
.sched_priority
= p
->rt_priority
;
4195 read_unlock(&tasklist_lock
);
4198 * This one might sleep, we cannot do it with a spinlock held ...
4200 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4206 read_unlock(&tasklist_lock
);
4210 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4214 cpumask_t cpus_allowed
;
4217 read_lock(&tasklist_lock
);
4219 p
= find_process_by_pid(pid
);
4221 read_unlock(&tasklist_lock
);
4222 unlock_cpu_hotplug();
4227 * It is not safe to call set_cpus_allowed with the
4228 * tasklist_lock held. We will bump the task_struct's
4229 * usage count and then drop tasklist_lock.
4232 read_unlock(&tasklist_lock
);
4235 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4236 !capable(CAP_SYS_NICE
))
4239 retval
= security_task_setscheduler(p
, 0, NULL
);
4243 cpus_allowed
= cpuset_cpus_allowed(p
);
4244 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4245 retval
= set_cpus_allowed(p
, new_mask
);
4249 unlock_cpu_hotplug();
4253 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4254 cpumask_t
*new_mask
)
4256 if (len
< sizeof(cpumask_t
)) {
4257 memset(new_mask
, 0, sizeof(cpumask_t
));
4258 } else if (len
> sizeof(cpumask_t
)) {
4259 len
= sizeof(cpumask_t
);
4261 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4265 * sys_sched_setaffinity - set the cpu affinity of a process
4266 * @pid: pid of the process
4267 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4268 * @user_mask_ptr: user-space pointer to the new cpu mask
4270 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4271 unsigned long __user
*user_mask_ptr
)
4276 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4280 return sched_setaffinity(pid
, new_mask
);
4284 * Represents all cpu's present in the system
4285 * In systems capable of hotplug, this map could dynamically grow
4286 * as new cpu's are detected in the system via any platform specific
4287 * method, such as ACPI for e.g.
4290 cpumask_t cpu_present_map __read_mostly
;
4291 EXPORT_SYMBOL(cpu_present_map
);
4294 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4295 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4298 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4304 read_lock(&tasklist_lock
);
4307 p
= find_process_by_pid(pid
);
4311 retval
= security_task_getscheduler(p
);
4315 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4318 read_unlock(&tasklist_lock
);
4319 unlock_cpu_hotplug();
4327 * sys_sched_getaffinity - get the cpu affinity of a process
4328 * @pid: pid of the process
4329 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4330 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4332 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4333 unsigned long __user
*user_mask_ptr
)
4338 if (len
< sizeof(cpumask_t
))
4341 ret
= sched_getaffinity(pid
, &mask
);
4345 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4348 return sizeof(cpumask_t
);
4352 * sys_sched_yield - yield the current processor to other threads.
4354 * this function yields the current CPU by moving the calling thread
4355 * to the expired array. If there are no other threads running on this
4356 * CPU then this function will return.
4358 asmlinkage
long sys_sched_yield(void)
4360 runqueue_t
*rq
= this_rq_lock();
4361 prio_array_t
*array
= current
->array
;
4362 prio_array_t
*target
= rq
->expired
;
4364 schedstat_inc(rq
, yld_cnt
);
4366 * We implement yielding by moving the task into the expired
4369 * (special rule: RT tasks will just roundrobin in the active
4372 if (rt_task(current
))
4373 target
= rq
->active
;
4375 if (array
->nr_active
== 1) {
4376 schedstat_inc(rq
, yld_act_empty
);
4377 if (!rq
->expired
->nr_active
)
4378 schedstat_inc(rq
, yld_both_empty
);
4379 } else if (!rq
->expired
->nr_active
)
4380 schedstat_inc(rq
, yld_exp_empty
);
4382 if (array
!= target
) {
4383 dequeue_task(current
, array
);
4384 enqueue_task(current
, target
);
4387 * requeue_task is cheaper so perform that if possible.
4389 requeue_task(current
, array
);
4392 * Since we are going to call schedule() anyway, there's
4393 * no need to preempt or enable interrupts:
4395 __release(rq
->lock
);
4396 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4397 _raw_spin_unlock(&rq
->lock
);
4398 preempt_enable_no_resched();
4405 static inline int __resched_legal(void)
4407 if (unlikely(preempt_count()))
4409 if (unlikely(system_state
!= SYSTEM_RUNNING
))
4414 static void __cond_resched(void)
4416 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4417 __might_sleep(__FILE__
, __LINE__
);
4420 * The BKS might be reacquired before we have dropped
4421 * PREEMPT_ACTIVE, which could trigger a second
4422 * cond_resched() call.
4425 add_preempt_count(PREEMPT_ACTIVE
);
4427 sub_preempt_count(PREEMPT_ACTIVE
);
4428 } while (need_resched());
4431 int __sched
cond_resched(void)
4433 if (need_resched() && __resched_legal()) {
4439 EXPORT_SYMBOL(cond_resched
);
4442 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4443 * call schedule, and on return reacquire the lock.
4445 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4446 * operations here to prevent schedule() from being called twice (once via
4447 * spin_unlock(), once by hand).
4449 int cond_resched_lock(spinlock_t
*lock
)
4453 if (need_lockbreak(lock
)) {
4459 if (need_resched() && __resched_legal()) {
4460 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4461 _raw_spin_unlock(lock
);
4462 preempt_enable_no_resched();
4469 EXPORT_SYMBOL(cond_resched_lock
);
4471 int __sched
cond_resched_softirq(void)
4473 BUG_ON(!in_softirq());
4475 if (need_resched() && __resched_legal()) {
4476 raw_local_irq_disable();
4478 raw_local_irq_enable();
4485 EXPORT_SYMBOL(cond_resched_softirq
);
4488 * yield - yield the current processor to other threads.
4490 * this is a shortcut for kernel-space yielding - it marks the
4491 * thread runnable and calls sys_sched_yield().
4493 void __sched
yield(void)
4495 set_current_state(TASK_RUNNING
);
4499 EXPORT_SYMBOL(yield
);
4502 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4503 * that process accounting knows that this is a task in IO wait state.
4505 * But don't do that if it is a deliberate, throttling IO wait (this task
4506 * has set its backing_dev_info: the queue against which it should throttle)
4508 void __sched
io_schedule(void)
4510 struct runqueue
*rq
= &__raw_get_cpu_var(runqueues
);
4512 atomic_inc(&rq
->nr_iowait
);
4514 atomic_dec(&rq
->nr_iowait
);
4517 EXPORT_SYMBOL(io_schedule
);
4519 long __sched
io_schedule_timeout(long timeout
)
4521 struct runqueue
*rq
= &__raw_get_cpu_var(runqueues
);
4524 atomic_inc(&rq
->nr_iowait
);
4525 ret
= schedule_timeout(timeout
);
4526 atomic_dec(&rq
->nr_iowait
);
4531 * sys_sched_get_priority_max - return maximum RT priority.
4532 * @policy: scheduling class.
4534 * this syscall returns the maximum rt_priority that can be used
4535 * by a given scheduling class.
4537 asmlinkage
long sys_sched_get_priority_max(int policy
)
4544 ret
= MAX_USER_RT_PRIO
-1;
4555 * sys_sched_get_priority_min - return minimum RT priority.
4556 * @policy: scheduling class.
4558 * this syscall returns the minimum rt_priority that can be used
4559 * by a given scheduling class.
4561 asmlinkage
long sys_sched_get_priority_min(int policy
)
4578 * sys_sched_rr_get_interval - return the default timeslice of a process.
4579 * @pid: pid of the process.
4580 * @interval: userspace pointer to the timeslice value.
4582 * this syscall writes the default timeslice value of a given process
4583 * into the user-space timespec buffer. A value of '0' means infinity.
4586 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4588 int retval
= -EINVAL
;
4596 read_lock(&tasklist_lock
);
4597 p
= find_process_by_pid(pid
);
4601 retval
= security_task_getscheduler(p
);
4605 jiffies_to_timespec(p
->policy
== SCHED_FIFO
?
4606 0 : task_timeslice(p
), &t
);
4607 read_unlock(&tasklist_lock
);
4608 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4612 read_unlock(&tasklist_lock
);
4616 static inline struct task_struct
*eldest_child(struct task_struct
*p
)
4618 if (list_empty(&p
->children
)) return NULL
;
4619 return list_entry(p
->children
.next
,struct task_struct
,sibling
);
4622 static inline struct task_struct
*older_sibling(struct task_struct
*p
)
4624 if (p
->sibling
.prev
==&p
->parent
->children
) return NULL
;
4625 return list_entry(p
->sibling
.prev
,struct task_struct
,sibling
);
4628 static inline struct task_struct
*younger_sibling(struct task_struct
*p
)
4630 if (p
->sibling
.next
==&p
->parent
->children
) return NULL
;
4631 return list_entry(p
->sibling
.next
,struct task_struct
,sibling
);
4634 static void show_task(task_t
*p
)
4638 unsigned long free
= 0;
4639 static const char *stat_nam
[] = { "R", "S", "D", "T", "t", "Z", "X" };
4641 printk("%-13.13s ", p
->comm
);
4642 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4643 if (state
< ARRAY_SIZE(stat_nam
))
4644 printk(stat_nam
[state
]);
4647 #if (BITS_PER_LONG == 32)
4648 if (state
== TASK_RUNNING
)
4649 printk(" running ");
4651 printk(" %08lX ", thread_saved_pc(p
));
4653 if (state
== TASK_RUNNING
)
4654 printk(" running task ");
4656 printk(" %016lx ", thread_saved_pc(p
));
4658 #ifdef CONFIG_DEBUG_STACK_USAGE
4660 unsigned long *n
= end_of_stack(p
);
4663 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4666 printk("%5lu %5d %6d ", free
, p
->pid
, p
->parent
->pid
);
4667 if ((relative
= eldest_child(p
)))
4668 printk("%5d ", relative
->pid
);
4671 if ((relative
= younger_sibling(p
)))
4672 printk("%7d", relative
->pid
);
4675 if ((relative
= older_sibling(p
)))
4676 printk(" %5d", relative
->pid
);
4680 printk(" (L-TLB)\n");
4682 printk(" (NOTLB)\n");
4684 if (state
!= TASK_RUNNING
)
4685 show_stack(p
, NULL
);
4688 void show_state(void)
4692 #if (BITS_PER_LONG == 32)
4695 printk(" task PC pid father child younger older\n");
4699 printk(" task PC pid father child younger older\n");
4701 read_lock(&tasklist_lock
);
4702 do_each_thread(g
, p
) {
4704 * reset the NMI-timeout, listing all files on a slow
4705 * console might take alot of time:
4707 touch_nmi_watchdog();
4709 } while_each_thread(g
, p
);
4711 read_unlock(&tasklist_lock
);
4712 debug_show_all_locks();
4716 * init_idle - set up an idle thread for a given CPU
4717 * @idle: task in question
4718 * @cpu: cpu the idle task belongs to
4720 * NOTE: this function does not set the idle thread's NEED_RESCHED
4721 * flag, to make booting more robust.
4723 void __devinit
init_idle(task_t
*idle
, int cpu
)
4725 runqueue_t
*rq
= cpu_rq(cpu
);
4726 unsigned long flags
;
4728 idle
->timestamp
= sched_clock();
4729 idle
->sleep_avg
= 0;
4731 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
4732 idle
->state
= TASK_RUNNING
;
4733 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4734 set_task_cpu(idle
, cpu
);
4736 spin_lock_irqsave(&rq
->lock
, flags
);
4737 rq
->curr
= rq
->idle
= idle
;
4738 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4741 spin_unlock_irqrestore(&rq
->lock
, flags
);
4743 /* Set the preempt count _outside_ the spinlocks! */
4744 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4745 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4747 task_thread_info(idle
)->preempt_count
= 0;
4752 * In a system that switches off the HZ timer nohz_cpu_mask
4753 * indicates which cpus entered this state. This is used
4754 * in the rcu update to wait only for active cpus. For system
4755 * which do not switch off the HZ timer nohz_cpu_mask should
4756 * always be CPU_MASK_NONE.
4758 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4762 * This is how migration works:
4764 * 1) we queue a migration_req_t structure in the source CPU's
4765 * runqueue and wake up that CPU's migration thread.
4766 * 2) we down() the locked semaphore => thread blocks.
4767 * 3) migration thread wakes up (implicitly it forces the migrated
4768 * thread off the CPU)
4769 * 4) it gets the migration request and checks whether the migrated
4770 * task is still in the wrong runqueue.
4771 * 5) if it's in the wrong runqueue then the migration thread removes
4772 * it and puts it into the right queue.
4773 * 6) migration thread up()s the semaphore.
4774 * 7) we wake up and the migration is done.
4778 * Change a given task's CPU affinity. Migrate the thread to a
4779 * proper CPU and schedule it away if the CPU it's executing on
4780 * is removed from the allowed bitmask.
4782 * NOTE: the caller must have a valid reference to the task, the
4783 * task must not exit() & deallocate itself prematurely. The
4784 * call is not atomic; no spinlocks may be held.
4786 int set_cpus_allowed(task_t
*p
, cpumask_t new_mask
)
4788 unsigned long flags
;
4790 migration_req_t req
;
4793 rq
= task_rq_lock(p
, &flags
);
4794 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4799 p
->cpus_allowed
= new_mask
;
4800 /* Can the task run on the task's current CPU? If so, we're done */
4801 if (cpu_isset(task_cpu(p
), new_mask
))
4804 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4805 /* Need help from migration thread: drop lock and wait. */
4806 task_rq_unlock(rq
, &flags
);
4807 wake_up_process(rq
->migration_thread
);
4808 wait_for_completion(&req
.done
);
4809 tlb_migrate_finish(p
->mm
);
4813 task_rq_unlock(rq
, &flags
);
4817 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4820 * Move (not current) task off this cpu, onto dest cpu. We're doing
4821 * this because either it can't run here any more (set_cpus_allowed()
4822 * away from this CPU, or CPU going down), or because we're
4823 * attempting to rebalance this task on exec (sched_exec).
4825 * So we race with normal scheduler movements, but that's OK, as long
4826 * as the task is no longer on this CPU.
4828 * Returns non-zero if task was successfully migrated.
4830 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4832 runqueue_t
*rq_dest
, *rq_src
;
4835 if (unlikely(cpu_is_offline(dest_cpu
)))
4838 rq_src
= cpu_rq(src_cpu
);
4839 rq_dest
= cpu_rq(dest_cpu
);
4841 double_rq_lock(rq_src
, rq_dest
);
4842 /* Already moved. */
4843 if (task_cpu(p
) != src_cpu
)
4845 /* Affinity changed (again). */
4846 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
4849 set_task_cpu(p
, dest_cpu
);
4852 * Sync timestamp with rq_dest's before activating.
4853 * The same thing could be achieved by doing this step
4854 * afterwards, and pretending it was a local activate.
4855 * This way is cleaner and logically correct.
4857 p
->timestamp
= p
->timestamp
- rq_src
->timestamp_last_tick
4858 + rq_dest
->timestamp_last_tick
;
4859 deactivate_task(p
, rq_src
);
4860 activate_task(p
, rq_dest
, 0);
4861 if (TASK_PREEMPTS_CURR(p
, rq_dest
))
4862 resched_task(rq_dest
->curr
);
4866 double_rq_unlock(rq_src
, rq_dest
);
4871 * migration_thread - this is a highprio system thread that performs
4872 * thread migration by bumping thread off CPU then 'pushing' onto
4875 static int migration_thread(void *data
)
4878 int cpu
= (long)data
;
4881 BUG_ON(rq
->migration_thread
!= current
);
4883 set_current_state(TASK_INTERRUPTIBLE
);
4884 while (!kthread_should_stop()) {
4885 struct list_head
*head
;
4886 migration_req_t
*req
;
4890 spin_lock_irq(&rq
->lock
);
4892 if (cpu_is_offline(cpu
)) {
4893 spin_unlock_irq(&rq
->lock
);
4897 if (rq
->active_balance
) {
4898 active_load_balance(rq
, cpu
);
4899 rq
->active_balance
= 0;
4902 head
= &rq
->migration_queue
;
4904 if (list_empty(head
)) {
4905 spin_unlock_irq(&rq
->lock
);
4907 set_current_state(TASK_INTERRUPTIBLE
);
4910 req
= list_entry(head
->next
, migration_req_t
, list
);
4911 list_del_init(head
->next
);
4913 spin_unlock(&rq
->lock
);
4914 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
4917 complete(&req
->done
);
4919 __set_current_state(TASK_RUNNING
);
4923 /* Wait for kthread_stop */
4924 set_current_state(TASK_INTERRUPTIBLE
);
4925 while (!kthread_should_stop()) {
4927 set_current_state(TASK_INTERRUPTIBLE
);
4929 __set_current_state(TASK_RUNNING
);
4933 #ifdef CONFIG_HOTPLUG_CPU
4934 /* Figure out where task on dead CPU should go, use force if neccessary. */
4935 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*tsk
)
4938 unsigned long flags
;
4944 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
4945 cpus_and(mask
, mask
, tsk
->cpus_allowed
);
4946 dest_cpu
= any_online_cpu(mask
);
4948 /* On any allowed CPU? */
4949 if (dest_cpu
== NR_CPUS
)
4950 dest_cpu
= any_online_cpu(tsk
->cpus_allowed
);
4952 /* No more Mr. Nice Guy. */
4953 if (dest_cpu
== NR_CPUS
) {
4954 rq
= task_rq_lock(tsk
, &flags
);
4955 cpus_setall(tsk
->cpus_allowed
);
4956 dest_cpu
= any_online_cpu(tsk
->cpus_allowed
);
4957 task_rq_unlock(rq
, &flags
);
4960 * Don't tell them about moving exiting tasks or
4961 * kernel threads (both mm NULL), since they never
4964 if (tsk
->mm
&& printk_ratelimit())
4965 printk(KERN_INFO
"process %d (%s) no "
4966 "longer affine to cpu%d\n",
4967 tsk
->pid
, tsk
->comm
, dead_cpu
);
4969 if (!__migrate_task(tsk
, dead_cpu
, dest_cpu
))
4974 * While a dead CPU has no uninterruptible tasks queued at this point,
4975 * it might still have a nonzero ->nr_uninterruptible counter, because
4976 * for performance reasons the counter is not stricly tracking tasks to
4977 * their home CPUs. So we just add the counter to another CPU's counter,
4978 * to keep the global sum constant after CPU-down:
4980 static void migrate_nr_uninterruptible(runqueue_t
*rq_src
)
4982 runqueue_t
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
4983 unsigned long flags
;
4985 local_irq_save(flags
);
4986 double_rq_lock(rq_src
, rq_dest
);
4987 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
4988 rq_src
->nr_uninterruptible
= 0;
4989 double_rq_unlock(rq_src
, rq_dest
);
4990 local_irq_restore(flags
);
4993 /* Run through task list and migrate tasks from the dead cpu. */
4994 static void migrate_live_tasks(int src_cpu
)
4996 struct task_struct
*tsk
, *t
;
4998 write_lock_irq(&tasklist_lock
);
5000 do_each_thread(t
, tsk
) {
5004 if (task_cpu(tsk
) == src_cpu
)
5005 move_task_off_dead_cpu(src_cpu
, tsk
);
5006 } while_each_thread(t
, tsk
);
5008 write_unlock_irq(&tasklist_lock
);
5011 /* Schedules idle task to be the next runnable task on current CPU.
5012 * It does so by boosting its priority to highest possible and adding it to
5013 * the _front_ of runqueue. Used by CPU offline code.
5015 void sched_idle_next(void)
5017 int cpu
= smp_processor_id();
5018 runqueue_t
*rq
= this_rq();
5019 struct task_struct
*p
= rq
->idle
;
5020 unsigned long flags
;
5022 /* cpu has to be offline */
5023 BUG_ON(cpu_online(cpu
));
5025 /* Strictly not necessary since rest of the CPUs are stopped by now
5026 * and interrupts disabled on current cpu.
5028 spin_lock_irqsave(&rq
->lock
, flags
);
5030 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5031 /* Add idle task to _front_ of it's priority queue */
5032 __activate_idle_task(p
, rq
);
5034 spin_unlock_irqrestore(&rq
->lock
, flags
);
5037 /* Ensures that the idle task is using init_mm right before its cpu goes
5040 void idle_task_exit(void)
5042 struct mm_struct
*mm
= current
->active_mm
;
5044 BUG_ON(cpu_online(smp_processor_id()));
5047 switch_mm(mm
, &init_mm
, current
);
5051 static void migrate_dead(unsigned int dead_cpu
, task_t
*tsk
)
5053 struct runqueue
*rq
= cpu_rq(dead_cpu
);
5055 /* Must be exiting, otherwise would be on tasklist. */
5056 BUG_ON(tsk
->exit_state
!= EXIT_ZOMBIE
&& tsk
->exit_state
!= EXIT_DEAD
);
5058 /* Cannot have done final schedule yet: would have vanished. */
5059 BUG_ON(tsk
->flags
& PF_DEAD
);
5061 get_task_struct(tsk
);
5064 * Drop lock around migration; if someone else moves it,
5065 * that's OK. No task can be added to this CPU, so iteration is
5068 spin_unlock_irq(&rq
->lock
);
5069 move_task_off_dead_cpu(dead_cpu
, tsk
);
5070 spin_lock_irq(&rq
->lock
);
5072 put_task_struct(tsk
);
5075 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5076 static void migrate_dead_tasks(unsigned int dead_cpu
)
5079 struct runqueue
*rq
= cpu_rq(dead_cpu
);
5081 for (arr
= 0; arr
< 2; arr
++) {
5082 for (i
= 0; i
< MAX_PRIO
; i
++) {
5083 struct list_head
*list
= &rq
->arrays
[arr
].queue
[i
];
5084 while (!list_empty(list
))
5085 migrate_dead(dead_cpu
,
5086 list_entry(list
->next
, task_t
,
5091 #endif /* CONFIG_HOTPLUG_CPU */
5094 * migration_call - callback that gets triggered when a CPU is added.
5095 * Here we can start up the necessary migration thread for the new CPU.
5097 static int __cpuinit
migration_call(struct notifier_block
*nfb
,
5098 unsigned long action
,
5101 int cpu
= (long)hcpu
;
5102 struct task_struct
*p
;
5103 struct runqueue
*rq
;
5104 unsigned long flags
;
5107 case CPU_UP_PREPARE
:
5108 p
= kthread_create(migration_thread
, hcpu
, "migration/%d",cpu
);
5111 p
->flags
|= PF_NOFREEZE
;
5112 kthread_bind(p
, cpu
);
5113 /* Must be high prio: stop_machine expects to yield to it. */
5114 rq
= task_rq_lock(p
, &flags
);
5115 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5116 task_rq_unlock(rq
, &flags
);
5117 cpu_rq(cpu
)->migration_thread
= p
;
5120 /* Strictly unneccessary, as first user will wake it. */
5121 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5123 #ifdef CONFIG_HOTPLUG_CPU
5124 case CPU_UP_CANCELED
:
5125 if (!cpu_rq(cpu
)->migration_thread
)
5127 /* Unbind it from offline cpu so it can run. Fall thru. */
5128 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5129 any_online_cpu(cpu_online_map
));
5130 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5131 cpu_rq(cpu
)->migration_thread
= NULL
;
5134 migrate_live_tasks(cpu
);
5136 kthread_stop(rq
->migration_thread
);
5137 rq
->migration_thread
= NULL
;
5138 /* Idle task back to normal (off runqueue, low prio) */
5139 rq
= task_rq_lock(rq
->idle
, &flags
);
5140 deactivate_task(rq
->idle
, rq
);
5141 rq
->idle
->static_prio
= MAX_PRIO
;
5142 __setscheduler(rq
->idle
, SCHED_NORMAL
, 0);
5143 migrate_dead_tasks(cpu
);
5144 task_rq_unlock(rq
, &flags
);
5145 migrate_nr_uninterruptible(rq
);
5146 BUG_ON(rq
->nr_running
!= 0);
5148 /* No need to migrate the tasks: it was best-effort if
5149 * they didn't do lock_cpu_hotplug(). Just wake up
5150 * the requestors. */
5151 spin_lock_irq(&rq
->lock
);
5152 while (!list_empty(&rq
->migration_queue
)) {
5153 migration_req_t
*req
;
5154 req
= list_entry(rq
->migration_queue
.next
,
5155 migration_req_t
, list
);
5156 list_del_init(&req
->list
);
5157 complete(&req
->done
);
5159 spin_unlock_irq(&rq
->lock
);
5166 /* Register at highest priority so that task migration (migrate_all_tasks)
5167 * happens before everything else.
5169 static struct notifier_block __cpuinitdata migration_notifier
= {
5170 .notifier_call
= migration_call
,
5174 int __init
migration_init(void)
5176 void *cpu
= (void *)(long)smp_processor_id();
5177 /* Start one for boot CPU. */
5178 migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5179 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5180 register_cpu_notifier(&migration_notifier
);
5186 #undef SCHED_DOMAIN_DEBUG
5187 #ifdef SCHED_DOMAIN_DEBUG
5188 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5193 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5197 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5202 struct sched_group
*group
= sd
->groups
;
5203 cpumask_t groupmask
;
5205 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5206 cpus_clear(groupmask
);
5209 for (i
= 0; i
< level
+ 1; i
++)
5211 printk("domain %d: ", level
);
5213 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5214 printk("does not load-balance\n");
5216 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain has parent");
5220 printk("span %s\n", str
);
5222 if (!cpu_isset(cpu
, sd
->span
))
5223 printk(KERN_ERR
"ERROR: domain->span does not contain CPU%d\n", cpu
);
5224 if (!cpu_isset(cpu
, group
->cpumask
))
5225 printk(KERN_ERR
"ERROR: domain->groups does not contain CPU%d\n", cpu
);
5228 for (i
= 0; i
< level
+ 2; i
++)
5234 printk(KERN_ERR
"ERROR: group is NULL\n");
5238 if (!group
->cpu_power
) {
5240 printk(KERN_ERR
"ERROR: domain->cpu_power not set\n");
5243 if (!cpus_weight(group
->cpumask
)) {
5245 printk(KERN_ERR
"ERROR: empty group\n");
5248 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5250 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5253 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5255 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5258 group
= group
->next
;
5259 } while (group
!= sd
->groups
);
5262 if (!cpus_equal(sd
->span
, groupmask
))
5263 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5269 if (!cpus_subset(groupmask
, sd
->span
))
5270 printk(KERN_ERR
"ERROR: parent span is not a superset of domain->span\n");
5276 #define sched_domain_debug(sd, cpu) {}
5279 static int sd_degenerate(struct sched_domain
*sd
)
5281 if (cpus_weight(sd
->span
) == 1)
5284 /* Following flags need at least 2 groups */
5285 if (sd
->flags
& (SD_LOAD_BALANCE
|
5286 SD_BALANCE_NEWIDLE
|
5289 if (sd
->groups
!= sd
->groups
->next
)
5293 /* Following flags don't use groups */
5294 if (sd
->flags
& (SD_WAKE_IDLE
|
5302 static int sd_parent_degenerate(struct sched_domain
*sd
,
5303 struct sched_domain
*parent
)
5305 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5307 if (sd_degenerate(parent
))
5310 if (!cpus_equal(sd
->span
, parent
->span
))
5313 /* Does parent contain flags not in child? */
5314 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5315 if (cflags
& SD_WAKE_AFFINE
)
5316 pflags
&= ~SD_WAKE_BALANCE
;
5317 /* Flags needing groups don't count if only 1 group in parent */
5318 if (parent
->groups
== parent
->groups
->next
) {
5319 pflags
&= ~(SD_LOAD_BALANCE
|
5320 SD_BALANCE_NEWIDLE
|
5324 if (~cflags
& pflags
)
5331 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5332 * hold the hotplug lock.
5334 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5336 runqueue_t
*rq
= cpu_rq(cpu
);
5337 struct sched_domain
*tmp
;
5339 /* Remove the sched domains which do not contribute to scheduling. */
5340 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5341 struct sched_domain
*parent
= tmp
->parent
;
5344 if (sd_parent_degenerate(tmp
, parent
))
5345 tmp
->parent
= parent
->parent
;
5348 if (sd
&& sd_degenerate(sd
))
5351 sched_domain_debug(sd
, cpu
);
5353 rcu_assign_pointer(rq
->sd
, sd
);
5356 /* cpus with isolated domains */
5357 static cpumask_t __devinitdata cpu_isolated_map
= CPU_MASK_NONE
;
5359 /* Setup the mask of cpus configured for isolated domains */
5360 static int __init
isolated_cpu_setup(char *str
)
5362 int ints
[NR_CPUS
], i
;
5364 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5365 cpus_clear(cpu_isolated_map
);
5366 for (i
= 1; i
<= ints
[0]; i
++)
5367 if (ints
[i
] < NR_CPUS
)
5368 cpu_set(ints
[i
], cpu_isolated_map
);
5372 __setup ("isolcpus=", isolated_cpu_setup
);
5375 * init_sched_build_groups takes an array of groups, the cpumask we wish
5376 * to span, and a pointer to a function which identifies what group a CPU
5377 * belongs to. The return value of group_fn must be a valid index into the
5378 * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
5379 * keep track of groups covered with a cpumask_t).
5381 * init_sched_build_groups will build a circular linked list of the groups
5382 * covered by the given span, and will set each group's ->cpumask correctly,
5383 * and ->cpu_power to 0.
5385 static void init_sched_build_groups(struct sched_group groups
[], cpumask_t span
,
5386 int (*group_fn
)(int cpu
))
5388 struct sched_group
*first
= NULL
, *last
= NULL
;
5389 cpumask_t covered
= CPU_MASK_NONE
;
5392 for_each_cpu_mask(i
, span
) {
5393 int group
= group_fn(i
);
5394 struct sched_group
*sg
= &groups
[group
];
5397 if (cpu_isset(i
, covered
))
5400 sg
->cpumask
= CPU_MASK_NONE
;
5403 for_each_cpu_mask(j
, span
) {
5404 if (group_fn(j
) != group
)
5407 cpu_set(j
, covered
);
5408 cpu_set(j
, sg
->cpumask
);
5419 #define SD_NODES_PER_DOMAIN 16
5422 * Self-tuning task migration cost measurement between source and target CPUs.
5424 * This is done by measuring the cost of manipulating buffers of varying
5425 * sizes. For a given buffer-size here are the steps that are taken:
5427 * 1) the source CPU reads+dirties a shared buffer
5428 * 2) the target CPU reads+dirties the same shared buffer
5430 * We measure how long they take, in the following 4 scenarios:
5432 * - source: CPU1, target: CPU2 | cost1
5433 * - source: CPU2, target: CPU1 | cost2
5434 * - source: CPU1, target: CPU1 | cost3
5435 * - source: CPU2, target: CPU2 | cost4
5437 * We then calculate the cost3+cost4-cost1-cost2 difference - this is
5438 * the cost of migration.
5440 * We then start off from a small buffer-size and iterate up to larger
5441 * buffer sizes, in 5% steps - measuring each buffer-size separately, and
5442 * doing a maximum search for the cost. (The maximum cost for a migration
5443 * normally occurs when the working set size is around the effective cache
5446 #define SEARCH_SCOPE 2
5447 #define MIN_CACHE_SIZE (64*1024U)
5448 #define DEFAULT_CACHE_SIZE (5*1024*1024U)
5449 #define ITERATIONS 1
5450 #define SIZE_THRESH 130
5451 #define COST_THRESH 130
5454 * The migration cost is a function of 'domain distance'. Domain
5455 * distance is the number of steps a CPU has to iterate down its
5456 * domain tree to share a domain with the other CPU. The farther
5457 * two CPUs are from each other, the larger the distance gets.
5459 * Note that we use the distance only to cache measurement results,
5460 * the distance value is not used numerically otherwise. When two
5461 * CPUs have the same distance it is assumed that the migration
5462 * cost is the same. (this is a simplification but quite practical)
5464 #define MAX_DOMAIN_DISTANCE 32
5466 static unsigned long long migration_cost
[MAX_DOMAIN_DISTANCE
] =
5467 { [ 0 ... MAX_DOMAIN_DISTANCE
-1 ] =
5469 * Architectures may override the migration cost and thus avoid
5470 * boot-time calibration. Unit is nanoseconds. Mostly useful for
5471 * virtualized hardware:
5473 #ifdef CONFIG_DEFAULT_MIGRATION_COST
5474 CONFIG_DEFAULT_MIGRATION_COST
5481 * Allow override of migration cost - in units of microseconds.
5482 * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost
5483 * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs:
5485 static int __init
migration_cost_setup(char *str
)
5487 int ints
[MAX_DOMAIN_DISTANCE
+1], i
;
5489 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5491 printk("#ints: %d\n", ints
[0]);
5492 for (i
= 1; i
<= ints
[0]; i
++) {
5493 migration_cost
[i
-1] = (unsigned long long)ints
[i
]*1000;
5494 printk("migration_cost[%d]: %Ld\n", i
-1, migration_cost
[i
-1]);
5499 __setup ("migration_cost=", migration_cost_setup
);
5502 * Global multiplier (divisor) for migration-cutoff values,
5503 * in percentiles. E.g. use a value of 150 to get 1.5 times
5504 * longer cache-hot cutoff times.
5506 * (We scale it from 100 to 128 to long long handling easier.)
5509 #define MIGRATION_FACTOR_SCALE 128
5511 static unsigned int migration_factor
= MIGRATION_FACTOR_SCALE
;
5513 static int __init
setup_migration_factor(char *str
)
5515 get_option(&str
, &migration_factor
);
5516 migration_factor
= migration_factor
* MIGRATION_FACTOR_SCALE
/ 100;
5520 __setup("migration_factor=", setup_migration_factor
);
5523 * Estimated distance of two CPUs, measured via the number of domains
5524 * we have to pass for the two CPUs to be in the same span:
5526 static unsigned long domain_distance(int cpu1
, int cpu2
)
5528 unsigned long distance
= 0;
5529 struct sched_domain
*sd
;
5531 for_each_domain(cpu1
, sd
) {
5532 WARN_ON(!cpu_isset(cpu1
, sd
->span
));
5533 if (cpu_isset(cpu2
, sd
->span
))
5537 if (distance
>= MAX_DOMAIN_DISTANCE
) {
5539 distance
= MAX_DOMAIN_DISTANCE
-1;
5545 static unsigned int migration_debug
;
5547 static int __init
setup_migration_debug(char *str
)
5549 get_option(&str
, &migration_debug
);
5553 __setup("migration_debug=", setup_migration_debug
);
5556 * Maximum cache-size that the scheduler should try to measure.
5557 * Architectures with larger caches should tune this up during
5558 * bootup. Gets used in the domain-setup code (i.e. during SMP
5561 unsigned int max_cache_size
;
5563 static int __init
setup_max_cache_size(char *str
)
5565 get_option(&str
, &max_cache_size
);
5569 __setup("max_cache_size=", setup_max_cache_size
);
5572 * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This
5573 * is the operation that is timed, so we try to generate unpredictable
5574 * cachemisses that still end up filling the L2 cache:
5576 static void touch_cache(void *__cache
, unsigned long __size
)
5578 unsigned long size
= __size
/sizeof(long), chunk1
= size
/3,
5580 unsigned long *cache
= __cache
;
5583 for (i
= 0; i
< size
/6; i
+= 8) {
5586 case 1: cache
[size
-1-i
]++;
5587 case 2: cache
[chunk1
-i
]++;
5588 case 3: cache
[chunk1
+i
]++;
5589 case 4: cache
[chunk2
-i
]++;
5590 case 5: cache
[chunk2
+i
]++;
5596 * Measure the cache-cost of one task migration. Returns in units of nsec.
5598 static unsigned long long measure_one(void *cache
, unsigned long size
,
5599 int source
, int target
)
5601 cpumask_t mask
, saved_mask
;
5602 unsigned long long t0
, t1
, t2
, t3
, cost
;
5604 saved_mask
= current
->cpus_allowed
;
5607 * Flush source caches to RAM and invalidate them:
5612 * Migrate to the source CPU:
5614 mask
= cpumask_of_cpu(source
);
5615 set_cpus_allowed(current
, mask
);
5616 WARN_ON(smp_processor_id() != source
);
5619 * Dirty the working set:
5622 touch_cache(cache
, size
);
5626 * Migrate to the target CPU, dirty the L2 cache and access
5627 * the shared buffer. (which represents the working set
5628 * of a migrated task.)
5630 mask
= cpumask_of_cpu(target
);
5631 set_cpus_allowed(current
, mask
);
5632 WARN_ON(smp_processor_id() != target
);
5635 touch_cache(cache
, size
);
5638 cost
= t1
-t0
+ t3
-t2
;
5640 if (migration_debug
>= 2)
5641 printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n",
5642 source
, target
, t1
-t0
, t1
-t0
, t3
-t2
, cost
);
5644 * Flush target caches to RAM and invalidate them:
5648 set_cpus_allowed(current
, saved_mask
);
5654 * Measure a series of task migrations and return the average
5655 * result. Since this code runs early during bootup the system
5656 * is 'undisturbed' and the average latency makes sense.
5658 * The algorithm in essence auto-detects the relevant cache-size,
5659 * so it will properly detect different cachesizes for different
5660 * cache-hierarchies, depending on how the CPUs are connected.
5662 * Architectures can prime the upper limit of the search range via
5663 * max_cache_size, otherwise the search range defaults to 20MB...64K.
5665 static unsigned long long
5666 measure_cost(int cpu1
, int cpu2
, void *cache
, unsigned int size
)
5668 unsigned long long cost1
, cost2
;
5672 * Measure the migration cost of 'size' bytes, over an
5673 * average of 10 runs:
5675 * (We perturb the cache size by a small (0..4k)
5676 * value to compensate size/alignment related artifacts.
5677 * We also subtract the cost of the operation done on
5683 * dry run, to make sure we start off cache-cold on cpu1,
5684 * and to get any vmalloc pagefaults in advance:
5686 measure_one(cache
, size
, cpu1
, cpu2
);
5687 for (i
= 0; i
< ITERATIONS
; i
++)
5688 cost1
+= measure_one(cache
, size
- i
*1024, cpu1
, cpu2
);
5690 measure_one(cache
, size
, cpu2
, cpu1
);
5691 for (i
= 0; i
< ITERATIONS
; i
++)
5692 cost1
+= measure_one(cache
, size
- i
*1024, cpu2
, cpu1
);
5695 * (We measure the non-migrating [cached] cost on both
5696 * cpu1 and cpu2, to handle CPUs with different speeds)
5700 measure_one(cache
, size
, cpu1
, cpu1
);
5701 for (i
= 0; i
< ITERATIONS
; i
++)
5702 cost2
+= measure_one(cache
, size
- i
*1024, cpu1
, cpu1
);
5704 measure_one(cache
, size
, cpu2
, cpu2
);
5705 for (i
= 0; i
< ITERATIONS
; i
++)
5706 cost2
+= measure_one(cache
, size
- i
*1024, cpu2
, cpu2
);
5709 * Get the per-iteration migration cost:
5711 do_div(cost1
, 2*ITERATIONS
);
5712 do_div(cost2
, 2*ITERATIONS
);
5714 return cost1
- cost2
;
5717 static unsigned long long measure_migration_cost(int cpu1
, int cpu2
)
5719 unsigned long long max_cost
= 0, fluct
= 0, avg_fluct
= 0;
5720 unsigned int max_size
, size
, size_found
= 0;
5721 long long cost
= 0, prev_cost
;
5725 * Search from max_cache_size*5 down to 64K - the real relevant
5726 * cachesize has to lie somewhere inbetween.
5728 if (max_cache_size
) {
5729 max_size
= max(max_cache_size
* SEARCH_SCOPE
, MIN_CACHE_SIZE
);
5730 size
= max(max_cache_size
/ SEARCH_SCOPE
, MIN_CACHE_SIZE
);
5733 * Since we have no estimation about the relevant
5736 max_size
= DEFAULT_CACHE_SIZE
* SEARCH_SCOPE
;
5737 size
= MIN_CACHE_SIZE
;
5740 if (!cpu_online(cpu1
) || !cpu_online(cpu2
)) {
5741 printk("cpu %d and %d not both online!\n", cpu1
, cpu2
);
5746 * Allocate the working set:
5748 cache
= vmalloc(max_size
);
5750 printk("could not vmalloc %d bytes for cache!\n", 2*max_size
);
5751 return 1000000; // return 1 msec on very small boxen
5754 while (size
<= max_size
) {
5756 cost
= measure_cost(cpu1
, cpu2
, cache
, size
);
5762 if (max_cost
< cost
) {
5768 * Calculate average fluctuation, we use this to prevent
5769 * noise from triggering an early break out of the loop:
5771 fluct
= abs(cost
- prev_cost
);
5772 avg_fluct
= (avg_fluct
+ fluct
)/2;
5774 if (migration_debug
)
5775 printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): (%8Ld %8Ld)\n",
5777 (long)cost
/ 1000000,
5778 ((long)cost
/ 100000) % 10,
5779 (long)max_cost
/ 1000000,
5780 ((long)max_cost
/ 100000) % 10,
5781 domain_distance(cpu1
, cpu2
),
5785 * If we iterated at least 20% past the previous maximum,
5786 * and the cost has dropped by more than 20% already,
5787 * (taking fluctuations into account) then we assume to
5788 * have found the maximum and break out of the loop early:
5790 if (size_found
&& (size
*100 > size_found
*SIZE_THRESH
))
5791 if (cost
+avg_fluct
<= 0 ||
5792 max_cost
*100 > (cost
+avg_fluct
)*COST_THRESH
) {
5794 if (migration_debug
)
5795 printk("-> found max.\n");
5799 * Increase the cachesize in 10% steps:
5801 size
= size
* 10 / 9;
5804 if (migration_debug
)
5805 printk("[%d][%d] working set size found: %d, cost: %Ld\n",
5806 cpu1
, cpu2
, size_found
, max_cost
);
5811 * A task is considered 'cache cold' if at least 2 times
5812 * the worst-case cost of migration has passed.
5814 * (this limit is only listened to if the load-balancing
5815 * situation is 'nice' - if there is a large imbalance we
5816 * ignore it for the sake of CPU utilization and
5817 * processing fairness.)
5819 return 2 * max_cost
* migration_factor
/ MIGRATION_FACTOR_SCALE
;
5822 static void calibrate_migration_costs(const cpumask_t
*cpu_map
)
5824 int cpu1
= -1, cpu2
= -1, cpu
, orig_cpu
= raw_smp_processor_id();
5825 unsigned long j0
, j1
, distance
, max_distance
= 0;
5826 struct sched_domain
*sd
;
5831 * First pass - calculate the cacheflush times:
5833 for_each_cpu_mask(cpu1
, *cpu_map
) {
5834 for_each_cpu_mask(cpu2
, *cpu_map
) {
5837 distance
= domain_distance(cpu1
, cpu2
);
5838 max_distance
= max(max_distance
, distance
);
5840 * No result cached yet?
5842 if (migration_cost
[distance
] == -1LL)
5843 migration_cost
[distance
] =
5844 measure_migration_cost(cpu1
, cpu2
);
5848 * Second pass - update the sched domain hierarchy with
5849 * the new cache-hot-time estimations:
5851 for_each_cpu_mask(cpu
, *cpu_map
) {
5853 for_each_domain(cpu
, sd
) {
5854 sd
->cache_hot_time
= migration_cost
[distance
];
5861 if (migration_debug
)
5862 printk("migration: max_cache_size: %d, cpu: %d MHz:\n",
5870 if (system_state
== SYSTEM_BOOTING
) {
5871 printk("migration_cost=");
5872 for (distance
= 0; distance
<= max_distance
; distance
++) {
5875 printk("%ld", (long)migration_cost
[distance
] / 1000);
5880 if (migration_debug
)
5881 printk("migration: %ld seconds\n", (j1
-j0
)/HZ
);
5884 * Move back to the original CPU. NUMA-Q gets confused
5885 * if we migrate to another quad during bootup.
5887 if (raw_smp_processor_id() != orig_cpu
) {
5888 cpumask_t mask
= cpumask_of_cpu(orig_cpu
),
5889 saved_mask
= current
->cpus_allowed
;
5891 set_cpus_allowed(current
, mask
);
5892 set_cpus_allowed(current
, saved_mask
);
5899 * find_next_best_node - find the next node to include in a sched_domain
5900 * @node: node whose sched_domain we're building
5901 * @used_nodes: nodes already in the sched_domain
5903 * Find the next node to include in a given scheduling domain. Simply
5904 * finds the closest node not already in the @used_nodes map.
5906 * Should use nodemask_t.
5908 static int find_next_best_node(int node
, unsigned long *used_nodes
)
5910 int i
, n
, val
, min_val
, best_node
= 0;
5914 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5915 /* Start at @node */
5916 n
= (node
+ i
) % MAX_NUMNODES
;
5918 if (!nr_cpus_node(n
))
5921 /* Skip already used nodes */
5922 if (test_bit(n
, used_nodes
))
5925 /* Simple min distance search */
5926 val
= node_distance(node
, n
);
5928 if (val
< min_val
) {
5934 set_bit(best_node
, used_nodes
);
5939 * sched_domain_node_span - get a cpumask for a node's sched_domain
5940 * @node: node whose cpumask we're constructing
5941 * @size: number of nodes to include in this span
5943 * Given a node, construct a good cpumask for its sched_domain to span. It
5944 * should be one that prevents unnecessary balancing, but also spreads tasks
5947 static cpumask_t
sched_domain_node_span(int node
)
5950 cpumask_t span
, nodemask
;
5951 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
5954 bitmap_zero(used_nodes
, MAX_NUMNODES
);
5956 nodemask
= node_to_cpumask(node
);
5957 cpus_or(span
, span
, nodemask
);
5958 set_bit(node
, used_nodes
);
5960 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5961 int next_node
= find_next_best_node(node
, used_nodes
);
5962 nodemask
= node_to_cpumask(next_node
);
5963 cpus_or(span
, span
, nodemask
);
5970 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
5972 * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we
5973 * can switch it on easily if needed.
5975 #ifdef CONFIG_SCHED_SMT
5976 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
5977 static struct sched_group sched_group_cpus
[NR_CPUS
];
5978 static int cpu_to_cpu_group(int cpu
)
5984 #ifdef CONFIG_SCHED_MC
5985 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
5986 static struct sched_group
*sched_group_core_bycpu
[NR_CPUS
];
5989 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5990 static int cpu_to_core_group(int cpu
)
5992 return first_cpu(cpu_sibling_map
[cpu
]);
5994 #elif defined(CONFIG_SCHED_MC)
5995 static int cpu_to_core_group(int cpu
)
6001 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
6002 static struct sched_group
*sched_group_phys_bycpu
[NR_CPUS
];
6003 static int cpu_to_phys_group(int cpu
)
6005 #if defined(CONFIG_SCHED_MC)
6006 cpumask_t mask
= cpu_coregroup_map(cpu
);
6007 return first_cpu(mask
);
6008 #elif defined(CONFIG_SCHED_SMT)
6009 return first_cpu(cpu_sibling_map
[cpu
]);
6017 * The init_sched_build_groups can't handle what we want to do with node
6018 * groups, so roll our own. Now each node has its own list of groups which
6019 * gets dynamically allocated.
6021 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
6022 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
6024 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
6025 static struct sched_group
*sched_group_allnodes_bycpu
[NR_CPUS
];
6027 static int cpu_to_allnodes_group(int cpu
)
6029 return cpu_to_node(cpu
);
6031 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6033 struct sched_group
*sg
= group_head
;
6039 for_each_cpu_mask(j
, sg
->cpumask
) {
6040 struct sched_domain
*sd
;
6042 sd
= &per_cpu(phys_domains
, j
);
6043 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
6045 * Only add "power" once for each
6051 sg
->cpu_power
+= sd
->groups
->cpu_power
;
6054 if (sg
!= group_head
)
6059 /* Free memory allocated for various sched_group structures */
6060 static void free_sched_groups(const cpumask_t
*cpu_map
)
6066 for_each_cpu_mask(cpu
, *cpu_map
) {
6067 struct sched_group
*sched_group_allnodes
6068 = sched_group_allnodes_bycpu
[cpu
];
6069 struct sched_group
**sched_group_nodes
6070 = sched_group_nodes_bycpu
[cpu
];
6072 if (sched_group_allnodes
) {
6073 kfree(sched_group_allnodes
);
6074 sched_group_allnodes_bycpu
[cpu
] = NULL
;
6077 if (!sched_group_nodes
)
6080 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6081 cpumask_t nodemask
= node_to_cpumask(i
);
6082 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6084 cpus_and(nodemask
, nodemask
, *cpu_map
);
6085 if (cpus_empty(nodemask
))
6095 if (oldsg
!= sched_group_nodes
[i
])
6098 kfree(sched_group_nodes
);
6099 sched_group_nodes_bycpu
[cpu
] = NULL
;
6102 for_each_cpu_mask(cpu
, *cpu_map
) {
6103 if (sched_group_phys_bycpu
[cpu
]) {
6104 kfree(sched_group_phys_bycpu
[cpu
]);
6105 sched_group_phys_bycpu
[cpu
] = NULL
;
6107 #ifdef CONFIG_SCHED_MC
6108 if (sched_group_core_bycpu
[cpu
]) {
6109 kfree(sched_group_core_bycpu
[cpu
]);
6110 sched_group_core_bycpu
[cpu
] = NULL
;
6117 * Build sched domains for a given set of cpus and attach the sched domains
6118 * to the individual cpus
6120 static int build_sched_domains(const cpumask_t
*cpu_map
)
6123 struct sched_group
*sched_group_phys
= NULL
;
6124 #ifdef CONFIG_SCHED_MC
6125 struct sched_group
*sched_group_core
= NULL
;
6128 struct sched_group
**sched_group_nodes
= NULL
;
6129 struct sched_group
*sched_group_allnodes
= NULL
;
6132 * Allocate the per-node list of sched groups
6134 sched_group_nodes
= kzalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
6136 if (!sched_group_nodes
) {
6137 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6140 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6144 * Set up domains for cpus specified by the cpu_map.
6146 for_each_cpu_mask(i
, *cpu_map
) {
6148 struct sched_domain
*sd
= NULL
, *p
;
6149 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6151 cpus_and(nodemask
, nodemask
, *cpu_map
);
6154 if (cpus_weight(*cpu_map
)
6155 > SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6156 if (!sched_group_allnodes
) {
6157 sched_group_allnodes
6158 = kmalloc(sizeof(struct sched_group
)
6161 if (!sched_group_allnodes
) {
6163 "Can not alloc allnodes sched group\n");
6166 sched_group_allnodes_bycpu
[i
]
6167 = sched_group_allnodes
;
6169 sd
= &per_cpu(allnodes_domains
, i
);
6170 *sd
= SD_ALLNODES_INIT
;
6171 sd
->span
= *cpu_map
;
6172 group
= cpu_to_allnodes_group(i
);
6173 sd
->groups
= &sched_group_allnodes
[group
];
6178 sd
= &per_cpu(node_domains
, i
);
6180 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6182 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6185 if (!sched_group_phys
) {
6187 = kmalloc(sizeof(struct sched_group
) * NR_CPUS
,
6189 if (!sched_group_phys
) {
6190 printk (KERN_WARNING
"Can not alloc phys sched"
6194 sched_group_phys_bycpu
[i
] = sched_group_phys
;
6198 sd
= &per_cpu(phys_domains
, i
);
6199 group
= cpu_to_phys_group(i
);
6201 sd
->span
= nodemask
;
6203 sd
->groups
= &sched_group_phys
[group
];
6205 #ifdef CONFIG_SCHED_MC
6206 if (!sched_group_core
) {
6208 = kmalloc(sizeof(struct sched_group
) * NR_CPUS
,
6210 if (!sched_group_core
) {
6211 printk (KERN_WARNING
"Can not alloc core sched"
6215 sched_group_core_bycpu
[i
] = sched_group_core
;
6219 sd
= &per_cpu(core_domains
, i
);
6220 group
= cpu_to_core_group(i
);
6222 sd
->span
= cpu_coregroup_map(i
);
6223 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6225 sd
->groups
= &sched_group_core
[group
];
6228 #ifdef CONFIG_SCHED_SMT
6230 sd
= &per_cpu(cpu_domains
, i
);
6231 group
= cpu_to_cpu_group(i
);
6232 *sd
= SD_SIBLING_INIT
;
6233 sd
->span
= cpu_sibling_map
[i
];
6234 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6236 sd
->groups
= &sched_group_cpus
[group
];
6240 #ifdef CONFIG_SCHED_SMT
6241 /* Set up CPU (sibling) groups */
6242 for_each_cpu_mask(i
, *cpu_map
) {
6243 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
6244 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6245 if (i
!= first_cpu(this_sibling_map
))
6248 init_sched_build_groups(sched_group_cpus
, this_sibling_map
,
6253 #ifdef CONFIG_SCHED_MC
6254 /* Set up multi-core groups */
6255 for_each_cpu_mask(i
, *cpu_map
) {
6256 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6257 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6258 if (i
!= first_cpu(this_core_map
))
6260 init_sched_build_groups(sched_group_core
, this_core_map
,
6261 &cpu_to_core_group
);
6266 /* Set up physical groups */
6267 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6268 cpumask_t nodemask
= node_to_cpumask(i
);
6270 cpus_and(nodemask
, nodemask
, *cpu_map
);
6271 if (cpus_empty(nodemask
))
6274 init_sched_build_groups(sched_group_phys
, nodemask
,
6275 &cpu_to_phys_group
);
6279 /* Set up node groups */
6280 if (sched_group_allnodes
)
6281 init_sched_build_groups(sched_group_allnodes
, *cpu_map
,
6282 &cpu_to_allnodes_group
);
6284 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6285 /* Set up node groups */
6286 struct sched_group
*sg
, *prev
;
6287 cpumask_t nodemask
= node_to_cpumask(i
);
6288 cpumask_t domainspan
;
6289 cpumask_t covered
= CPU_MASK_NONE
;
6292 cpus_and(nodemask
, nodemask
, *cpu_map
);
6293 if (cpus_empty(nodemask
)) {
6294 sched_group_nodes
[i
] = NULL
;
6298 domainspan
= sched_domain_node_span(i
);
6299 cpus_and(domainspan
, domainspan
, *cpu_map
);
6301 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6303 printk(KERN_WARNING
"Can not alloc domain group for "
6307 sched_group_nodes
[i
] = sg
;
6308 for_each_cpu_mask(j
, nodemask
) {
6309 struct sched_domain
*sd
;
6310 sd
= &per_cpu(node_domains
, j
);
6314 sg
->cpumask
= nodemask
;
6316 cpus_or(covered
, covered
, nodemask
);
6319 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6320 cpumask_t tmp
, notcovered
;
6321 int n
= (i
+ j
) % MAX_NUMNODES
;
6323 cpus_complement(notcovered
, covered
);
6324 cpus_and(tmp
, notcovered
, *cpu_map
);
6325 cpus_and(tmp
, tmp
, domainspan
);
6326 if (cpus_empty(tmp
))
6329 nodemask
= node_to_cpumask(n
);
6330 cpus_and(tmp
, tmp
, nodemask
);
6331 if (cpus_empty(tmp
))
6334 sg
= kmalloc_node(sizeof(struct sched_group
),
6338 "Can not alloc domain group for node %d\n", j
);
6343 sg
->next
= prev
->next
;
6344 cpus_or(covered
, covered
, tmp
);
6351 /* Calculate CPU power for physical packages and nodes */
6352 #ifdef CONFIG_SCHED_SMT
6353 for_each_cpu_mask(i
, *cpu_map
) {
6354 struct sched_domain
*sd
;
6355 sd
= &per_cpu(cpu_domains
, i
);
6356 sd
->groups
->cpu_power
= SCHED_LOAD_SCALE
;
6359 #ifdef CONFIG_SCHED_MC
6360 for_each_cpu_mask(i
, *cpu_map
) {
6362 struct sched_domain
*sd
;
6363 sd
= &per_cpu(core_domains
, i
);
6364 if (sched_smt_power_savings
)
6365 power
= SCHED_LOAD_SCALE
* cpus_weight(sd
->groups
->cpumask
);
6367 power
= SCHED_LOAD_SCALE
+ (cpus_weight(sd
->groups
->cpumask
)-1)
6368 * SCHED_LOAD_SCALE
/ 10;
6369 sd
->groups
->cpu_power
= power
;
6373 for_each_cpu_mask(i
, *cpu_map
) {
6374 struct sched_domain
*sd
;
6375 #ifdef CONFIG_SCHED_MC
6376 sd
= &per_cpu(phys_domains
, i
);
6377 if (i
!= first_cpu(sd
->groups
->cpumask
))
6380 sd
->groups
->cpu_power
= 0;
6381 if (sched_mc_power_savings
|| sched_smt_power_savings
) {
6384 for_each_cpu_mask(j
, sd
->groups
->cpumask
) {
6385 struct sched_domain
*sd1
;
6386 sd1
= &per_cpu(core_domains
, j
);
6388 * for each core we will add once
6389 * to the group in physical domain
6391 if (j
!= first_cpu(sd1
->groups
->cpumask
))
6394 if (sched_smt_power_savings
)
6395 sd
->groups
->cpu_power
+= sd1
->groups
->cpu_power
;
6397 sd
->groups
->cpu_power
+= SCHED_LOAD_SCALE
;
6401 * This has to be < 2 * SCHED_LOAD_SCALE
6402 * Lets keep it SCHED_LOAD_SCALE, so that
6403 * while calculating NUMA group's cpu_power
6405 * numa_group->cpu_power += phys_group->cpu_power;
6407 * See "only add power once for each physical pkg"
6410 sd
->groups
->cpu_power
= SCHED_LOAD_SCALE
;
6413 sd
= &per_cpu(phys_domains
, i
);
6414 if (sched_smt_power_savings
)
6415 power
= SCHED_LOAD_SCALE
* cpus_weight(sd
->groups
->cpumask
);
6417 power
= SCHED_LOAD_SCALE
;
6418 sd
->groups
->cpu_power
= power
;
6423 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6424 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6426 init_numa_sched_groups_power(sched_group_allnodes
);
6429 /* Attach the domains */
6430 for_each_cpu_mask(i
, *cpu_map
) {
6431 struct sched_domain
*sd
;
6432 #ifdef CONFIG_SCHED_SMT
6433 sd
= &per_cpu(cpu_domains
, i
);
6434 #elif defined(CONFIG_SCHED_MC)
6435 sd
= &per_cpu(core_domains
, i
);
6437 sd
= &per_cpu(phys_domains
, i
);
6439 cpu_attach_domain(sd
, i
);
6442 * Tune cache-hot values:
6444 calibrate_migration_costs(cpu_map
);
6449 free_sched_groups(cpu_map
);
6453 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6455 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6457 cpumask_t cpu_default_map
;
6461 * Setup mask for cpus without special case scheduling requirements.
6462 * For now this just excludes isolated cpus, but could be used to
6463 * exclude other special cases in the future.
6465 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
6467 err
= build_sched_domains(&cpu_default_map
);
6472 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6474 free_sched_groups(cpu_map
);
6478 * Detach sched domains from a group of cpus specified in cpu_map
6479 * These cpus will now be attached to the NULL domain
6481 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6485 for_each_cpu_mask(i
, *cpu_map
)
6486 cpu_attach_domain(NULL
, i
);
6487 synchronize_sched();
6488 arch_destroy_sched_domains(cpu_map
);
6492 * Partition sched domains as specified by the cpumasks below.
6493 * This attaches all cpus from the cpumasks to the NULL domain,
6494 * waits for a RCU quiescent period, recalculates sched
6495 * domain information and then attaches them back to the
6496 * correct sched domains
6497 * Call with hotplug lock held
6499 int partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
6501 cpumask_t change_map
;
6504 cpus_and(*partition1
, *partition1
, cpu_online_map
);
6505 cpus_and(*partition2
, *partition2
, cpu_online_map
);
6506 cpus_or(change_map
, *partition1
, *partition2
);
6508 /* Detach sched domains from all of the affected cpus */
6509 detach_destroy_domains(&change_map
);
6510 if (!cpus_empty(*partition1
))
6511 err
= build_sched_domains(partition1
);
6512 if (!err
&& !cpus_empty(*partition2
))
6513 err
= build_sched_domains(partition2
);
6518 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6519 int arch_reinit_sched_domains(void)
6524 detach_destroy_domains(&cpu_online_map
);
6525 err
= arch_init_sched_domains(&cpu_online_map
);
6526 unlock_cpu_hotplug();
6531 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6535 if (buf
[0] != '0' && buf
[0] != '1')
6539 sched_smt_power_savings
= (buf
[0] == '1');
6541 sched_mc_power_savings
= (buf
[0] == '1');
6543 ret
= arch_reinit_sched_domains();
6545 return ret
? ret
: count
;
6548 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6551 #ifdef CONFIG_SCHED_SMT
6553 err
= sysfs_create_file(&cls
->kset
.kobj
,
6554 &attr_sched_smt_power_savings
.attr
);
6556 #ifdef CONFIG_SCHED_MC
6557 if (!err
&& mc_capable())
6558 err
= sysfs_create_file(&cls
->kset
.kobj
,
6559 &attr_sched_mc_power_savings
.attr
);
6565 #ifdef CONFIG_SCHED_MC
6566 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6568 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6570 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
, const char *buf
, size_t count
)
6572 return sched_power_savings_store(buf
, count
, 0);
6574 SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6575 sched_mc_power_savings_store
);
6578 #ifdef CONFIG_SCHED_SMT
6579 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6581 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6583 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
, const char *buf
, size_t count
)
6585 return sched_power_savings_store(buf
, count
, 1);
6587 SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6588 sched_smt_power_savings_store
);
6592 #ifdef CONFIG_HOTPLUG_CPU
6594 * Force a reinitialization of the sched domains hierarchy. The domains
6595 * and groups cannot be updated in place without racing with the balancing
6596 * code, so we temporarily attach all running cpus to the NULL domain
6597 * which will prevent rebalancing while the sched domains are recalculated.
6599 static int update_sched_domains(struct notifier_block
*nfb
,
6600 unsigned long action
, void *hcpu
)
6603 case CPU_UP_PREPARE
:
6604 case CPU_DOWN_PREPARE
:
6605 detach_destroy_domains(&cpu_online_map
);
6608 case CPU_UP_CANCELED
:
6609 case CPU_DOWN_FAILED
:
6613 * Fall through and re-initialise the domains.
6620 /* The hotplug lock is already held by cpu_up/cpu_down */
6621 arch_init_sched_domains(&cpu_online_map
);
6627 void __init
sched_init_smp(void)
6630 arch_init_sched_domains(&cpu_online_map
);
6631 unlock_cpu_hotplug();
6632 /* XXX: Theoretical race here - CPU may be hotplugged now */
6633 hotcpu_notifier(update_sched_domains
, 0);
6636 void __init
sched_init_smp(void)
6639 #endif /* CONFIG_SMP */
6641 int in_sched_functions(unsigned long addr
)
6643 /* Linker adds these: start and end of __sched functions */
6644 extern char __sched_text_start
[], __sched_text_end
[];
6645 return in_lock_functions(addr
) ||
6646 (addr
>= (unsigned long)__sched_text_start
6647 && addr
< (unsigned long)__sched_text_end
);
6650 void __init
sched_init(void)
6655 for_each_possible_cpu(i
) {
6656 prio_array_t
*array
;
6659 spin_lock_init(&rq
->lock
);
6660 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
6662 rq
->active
= rq
->arrays
;
6663 rq
->expired
= rq
->arrays
+ 1;
6664 rq
->best_expired_prio
= MAX_PRIO
;
6668 for (j
= 1; j
< 3; j
++)
6669 rq
->cpu_load
[j
] = 0;
6670 rq
->active_balance
= 0;
6672 rq
->migration_thread
= NULL
;
6673 INIT_LIST_HEAD(&rq
->migration_queue
);
6675 atomic_set(&rq
->nr_iowait
, 0);
6677 for (j
= 0; j
< 2; j
++) {
6678 array
= rq
->arrays
+ j
;
6679 for (k
= 0; k
< MAX_PRIO
; k
++) {
6680 INIT_LIST_HEAD(array
->queue
+ k
);
6681 __clear_bit(k
, array
->bitmap
);
6683 // delimiter for bitsearch
6684 __set_bit(MAX_PRIO
, array
->bitmap
);
6688 set_load_weight(&init_task
);
6690 * The boot idle thread does lazy MMU switching as well:
6692 atomic_inc(&init_mm
.mm_count
);
6693 enter_lazy_tlb(&init_mm
, current
);
6696 * Make us the idle thread. Technically, schedule() should not be
6697 * called from this thread, however somewhere below it might be,
6698 * but because we are the idle thread, we just pick up running again
6699 * when this runqueue becomes "idle".
6701 init_idle(current
, smp_processor_id());
6704 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6705 void __might_sleep(char *file
, int line
)
6707 #if defined(in_atomic)
6708 static unsigned long prev_jiffy
; /* ratelimiting */
6710 if ((in_atomic() || irqs_disabled()) &&
6711 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6712 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6714 prev_jiffy
= jiffies
;
6715 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6716 " context at %s:%d\n", file
, line
);
6717 printk("in_atomic():%d, irqs_disabled():%d\n",
6718 in_atomic(), irqs_disabled());
6723 EXPORT_SYMBOL(__might_sleep
);
6726 #ifdef CONFIG_MAGIC_SYSRQ
6727 void normalize_rt_tasks(void)
6729 struct task_struct
*p
;
6730 prio_array_t
*array
;
6731 unsigned long flags
;
6734 read_lock_irq(&tasklist_lock
);
6735 for_each_process(p
) {
6739 spin_lock_irqsave(&p
->pi_lock
, flags
);
6740 rq
= __task_rq_lock(p
);
6744 deactivate_task(p
, task_rq(p
));
6745 __setscheduler(p
, SCHED_NORMAL
, 0);
6747 __activate_task(p
, task_rq(p
));
6748 resched_task(rq
->curr
);
6751 __task_rq_unlock(rq
);
6752 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6754 read_unlock_irq(&tasklist_lock
);
6757 #endif /* CONFIG_MAGIC_SYSRQ */
6761 * These functions are only useful for the IA64 MCA handling.
6763 * They can only be called when the whole system has been
6764 * stopped - every CPU needs to be quiescent, and no scheduling
6765 * activity can take place. Using them for anything else would
6766 * be a serious bug, and as a result, they aren't even visible
6767 * under any other configuration.
6771 * curr_task - return the current task for a given cpu.
6772 * @cpu: the processor in question.
6774 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6776 task_t
*curr_task(int cpu
)
6778 return cpu_curr(cpu
);
6782 * set_curr_task - set the current task for a given cpu.
6783 * @cpu: the processor in question.
6784 * @p: the task pointer to set.
6786 * Description: This function must only be used when non-maskable interrupts
6787 * are serviced on a separate stack. It allows the architecture to switch the
6788 * notion of the current task on a cpu in a non-blocking manner. This function
6789 * must be called with all CPU's synchronized, and interrupts disabled, the
6790 * and caller must save the original value of the current task (see
6791 * curr_task() above) and restore that value before reenabling interrupts and
6792 * re-starting the system.
6794 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6796 void set_curr_task(int cpu
, task_t
*p
)