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/freezer.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/tsacct_kern.h>
53 #include <linux/kprobes.h>
54 #include <linux/delayacct.h>
55 #include <linux/reciprocal_div.h>
58 #include <asm/unistd.h>
61 * Scheduler clock - returns current time in nanosec units.
62 * This is default implementation.
63 * Architectures and sub-architectures can override this.
65 unsigned long long __attribute__((weak
)) sched_clock(void)
67 return (unsigned long long)jiffies
* (1000000000 / HZ
);
71 * Convert user-nice values [ -20 ... 0 ... 19 ]
72 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
75 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
76 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
77 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
80 * 'User priority' is the nice value converted to something we
81 * can work with better when scaling various scheduler parameters,
82 * it's a [ 0 ... 39 ] range.
84 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
85 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
86 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
89 * Some helpers for converting nanosecond timing to jiffy resolution
91 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
92 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
95 * These are the 'tuning knobs' of the scheduler:
97 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
98 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
99 * Timeslices get refilled after they expire.
101 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
102 #define DEF_TIMESLICE (100 * HZ / 1000)
103 #define ON_RUNQUEUE_WEIGHT 30
104 #define CHILD_PENALTY 95
105 #define PARENT_PENALTY 100
106 #define EXIT_WEIGHT 3
107 #define PRIO_BONUS_RATIO 25
108 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
109 #define INTERACTIVE_DELTA 2
110 #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
111 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
112 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
115 * If a task is 'interactive' then we reinsert it in the active
116 * array after it has expired its current timeslice. (it will not
117 * continue to run immediately, it will still roundrobin with
118 * other interactive tasks.)
120 * This part scales the interactivity limit depending on niceness.
122 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
123 * Here are a few examples of different nice levels:
125 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
126 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
127 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
128 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
129 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
131 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
132 * priority range a task can explore, a value of '1' means the
133 * task is rated interactive.)
135 * Ie. nice +19 tasks can never get 'interactive' enough to be
136 * reinserted into the active array. And only heavily CPU-hog nice -20
137 * tasks will be expired. Default nice 0 tasks are somewhere between,
138 * it takes some effort for them to get interactive, but it's not
142 #define CURRENT_BONUS(p) \
143 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
146 #define GRANULARITY (10 * HZ / 1000 ? : 1)
149 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
150 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
153 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
154 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
157 #define SCALE(v1,v1_max,v2_max) \
158 (v1) * (v2_max) / (v1_max)
161 (SCALE(TASK_NICE(p) + 20, 40, MAX_BONUS) - 20 * MAX_BONUS / 40 + \
164 #define TASK_INTERACTIVE(p) \
165 ((p)->prio <= (p)->static_prio - DELTA(p))
167 #define INTERACTIVE_SLEEP(p) \
168 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
169 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
171 #define TASK_PREEMPTS_CURR(p, rq) \
172 ((p)->prio < (rq)->curr->prio)
174 #define SCALE_PRIO(x, prio) \
175 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
177 static unsigned int static_prio_timeslice(int static_prio
)
179 if (static_prio
< NICE_TO_PRIO(0))
180 return SCALE_PRIO(DEF_TIMESLICE
* 4, static_prio
);
182 return SCALE_PRIO(DEF_TIMESLICE
, static_prio
);
187 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
188 * Since cpu_power is a 'constant', we can use a reciprocal divide.
190 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
192 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
196 * Each time a sched group cpu_power is changed,
197 * we must compute its reciprocal value
199 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
201 sg
->__cpu_power
+= val
;
202 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
207 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
208 * to time slice values: [800ms ... 100ms ... 5ms]
210 * The higher a thread's priority, the bigger timeslices
211 * it gets during one round of execution. But even the lowest
212 * priority thread gets MIN_TIMESLICE worth of execution time.
215 static inline unsigned int task_timeslice(struct task_struct
*p
)
217 return static_prio_timeslice(p
->static_prio
);
221 * These are the runqueue data structures:
225 unsigned int nr_active
;
226 DECLARE_BITMAP(bitmap
, MAX_PRIO
+1); /* include 1 bit for delimiter */
227 struct list_head queue
[MAX_PRIO
];
231 * This is the main, per-CPU runqueue data structure.
233 * Locking rule: those places that want to lock multiple runqueues
234 * (such as the load balancing or the thread migration code), lock
235 * acquire operations must be ordered by ascending &runqueue.
241 * nr_running and cpu_load should be in the same cacheline because
242 * remote CPUs use both these fields when doing load calculation.
244 unsigned long nr_running
;
245 unsigned long raw_weighted_load
;
247 unsigned long cpu_load
[3];
248 unsigned char idle_at_tick
;
250 unsigned char in_nohz_recently
;
253 unsigned long long nr_switches
;
256 * This is part of a global counter where only the total sum
257 * over all CPUs matters. A task can increase this counter on
258 * one CPU and if it got migrated afterwards it may decrease
259 * it on another CPU. Always updated under the runqueue lock:
261 unsigned long nr_uninterruptible
;
263 unsigned long expired_timestamp
;
264 /* Cached timestamp set by update_cpu_clock() */
265 unsigned long long most_recent_timestamp
;
266 struct task_struct
*curr
, *idle
;
267 unsigned long next_balance
;
268 struct mm_struct
*prev_mm
;
269 struct prio_array
*active
, *expired
, arrays
[2];
270 int best_expired_prio
;
274 struct sched_domain
*sd
;
276 /* For active balancing */
279 int cpu
; /* cpu of this runqueue */
281 struct task_struct
*migration_thread
;
282 struct list_head migration_queue
;
285 #ifdef CONFIG_SCHEDSTATS
287 struct sched_info rq_sched_info
;
289 /* sys_sched_yield() stats */
290 unsigned long yld_exp_empty
;
291 unsigned long yld_act_empty
;
292 unsigned long yld_both_empty
;
293 unsigned long yld_cnt
;
295 /* schedule() stats */
296 unsigned long sched_switch
;
297 unsigned long sched_cnt
;
298 unsigned long sched_goidle
;
300 /* try_to_wake_up() stats */
301 unsigned long ttwu_cnt
;
302 unsigned long ttwu_local
;
304 struct lock_class_key rq_lock_key
;
307 static DEFINE_PER_CPU(struct rq
, runqueues
);
309 static inline int cpu_of(struct rq
*rq
)
319 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
320 * See detach_destroy_domains: synchronize_sched for details.
322 * The domain tree of any CPU may only be accessed from within
323 * preempt-disabled sections.
325 #define for_each_domain(cpu, __sd) \
326 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
328 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
329 #define this_rq() (&__get_cpu_var(runqueues))
330 #define task_rq(p) cpu_rq(task_cpu(p))
331 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
333 #ifndef prepare_arch_switch
334 # define prepare_arch_switch(next) do { } while (0)
336 #ifndef finish_arch_switch
337 # define finish_arch_switch(prev) do { } while (0)
340 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
341 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
343 return rq
->curr
== p
;
346 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
350 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
352 #ifdef CONFIG_DEBUG_SPINLOCK
353 /* this is a valid case when another task releases the spinlock */
354 rq
->lock
.owner
= current
;
357 * If we are tracking spinlock dependencies then we have to
358 * fix up the runqueue lock - which gets 'carried over' from
361 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
363 spin_unlock_irq(&rq
->lock
);
366 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
367 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
372 return rq
->curr
== p
;
376 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
380 * We can optimise this out completely for !SMP, because the
381 * SMP rebalancing from interrupt is the only thing that cares
386 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
387 spin_unlock_irq(&rq
->lock
);
389 spin_unlock(&rq
->lock
);
393 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
397 * After ->oncpu is cleared, the task can be moved to a different CPU.
398 * We must ensure this doesn't happen until the switch is completely
404 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
408 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
411 * __task_rq_lock - lock the runqueue a given task resides on.
412 * Must be called interrupts disabled.
414 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
421 spin_lock(&rq
->lock
);
422 if (unlikely(rq
!= task_rq(p
))) {
423 spin_unlock(&rq
->lock
);
424 goto repeat_lock_task
;
430 * task_rq_lock - lock the runqueue a given task resides on and disable
431 * interrupts. Note the ordering: we can safely lookup the task_rq without
432 * explicitly disabling preemption.
434 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
440 local_irq_save(*flags
);
442 spin_lock(&rq
->lock
);
443 if (unlikely(rq
!= task_rq(p
))) {
444 spin_unlock_irqrestore(&rq
->lock
, *flags
);
445 goto repeat_lock_task
;
450 static inline void __task_rq_unlock(struct rq
*rq
)
453 spin_unlock(&rq
->lock
);
456 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
459 spin_unlock_irqrestore(&rq
->lock
, *flags
);
462 #ifdef CONFIG_SCHEDSTATS
464 * bump this up when changing the output format or the meaning of an existing
465 * format, so that tools can adapt (or abort)
467 #define SCHEDSTAT_VERSION 14
469 static int show_schedstat(struct seq_file
*seq
, void *v
)
473 seq_printf(seq
, "version %d\n", SCHEDSTAT_VERSION
);
474 seq_printf(seq
, "timestamp %lu\n", jiffies
);
475 for_each_online_cpu(cpu
) {
476 struct rq
*rq
= cpu_rq(cpu
);
478 struct sched_domain
*sd
;
482 /* runqueue-specific stats */
484 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
485 cpu
, rq
->yld_both_empty
,
486 rq
->yld_act_empty
, rq
->yld_exp_empty
, rq
->yld_cnt
,
487 rq
->sched_switch
, rq
->sched_cnt
, rq
->sched_goidle
,
488 rq
->ttwu_cnt
, rq
->ttwu_local
,
489 rq
->rq_sched_info
.cpu_time
,
490 rq
->rq_sched_info
.run_delay
, rq
->rq_sched_info
.pcnt
);
492 seq_printf(seq
, "\n");
495 /* domain-specific stats */
497 for_each_domain(cpu
, sd
) {
498 enum idle_type itype
;
499 char mask_str
[NR_CPUS
];
501 cpumask_scnprintf(mask_str
, NR_CPUS
, sd
->span
);
502 seq_printf(seq
, "domain%d %s", dcnt
++, mask_str
);
503 for (itype
= SCHED_IDLE
; itype
< MAX_IDLE_TYPES
;
505 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu "
508 sd
->lb_balanced
[itype
],
509 sd
->lb_failed
[itype
],
510 sd
->lb_imbalance
[itype
],
511 sd
->lb_gained
[itype
],
512 sd
->lb_hot_gained
[itype
],
513 sd
->lb_nobusyq
[itype
],
514 sd
->lb_nobusyg
[itype
]);
516 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu %lu %lu"
518 sd
->alb_cnt
, sd
->alb_failed
, sd
->alb_pushed
,
519 sd
->sbe_cnt
, sd
->sbe_balanced
, sd
->sbe_pushed
,
520 sd
->sbf_cnt
, sd
->sbf_balanced
, sd
->sbf_pushed
,
521 sd
->ttwu_wake_remote
, sd
->ttwu_move_affine
,
522 sd
->ttwu_move_balance
);
530 static int schedstat_open(struct inode
*inode
, struct file
*file
)
532 unsigned int size
= PAGE_SIZE
* (1 + num_online_cpus() / 32);
533 char *buf
= kmalloc(size
, GFP_KERNEL
);
539 res
= single_open(file
, show_schedstat
, NULL
);
541 m
= file
->private_data
;
549 const struct file_operations proc_schedstat_operations
= {
550 .open
= schedstat_open
,
553 .release
= single_release
,
557 * Expects runqueue lock to be held for atomicity of update
560 rq_sched_info_arrive(struct rq
*rq
, unsigned long delta_jiffies
)
563 rq
->rq_sched_info
.run_delay
+= delta_jiffies
;
564 rq
->rq_sched_info
.pcnt
++;
569 * Expects runqueue lock to be held for atomicity of update
572 rq_sched_info_depart(struct rq
*rq
, unsigned long delta_jiffies
)
575 rq
->rq_sched_info
.cpu_time
+= delta_jiffies
;
577 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
578 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
579 #else /* !CONFIG_SCHEDSTATS */
581 rq_sched_info_arrive(struct rq
*rq
, unsigned long delta_jiffies
)
584 rq_sched_info_depart(struct rq
*rq
, unsigned long delta_jiffies
)
586 # define schedstat_inc(rq, field) do { } while (0)
587 # define schedstat_add(rq, field, amt) do { } while (0)
591 * this_rq_lock - lock this runqueue and disable interrupts.
593 static inline struct rq
*this_rq_lock(void)
600 spin_lock(&rq
->lock
);
605 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
607 * Called when a process is dequeued from the active array and given
608 * the cpu. We should note that with the exception of interactive
609 * tasks, the expired queue will become the active queue after the active
610 * queue is empty, without explicitly dequeuing and requeuing tasks in the
611 * expired queue. (Interactive tasks may be requeued directly to the
612 * active queue, thus delaying tasks in the expired queue from running;
613 * see scheduler_tick()).
615 * This function is only called from sched_info_arrive(), rather than
616 * dequeue_task(). Even though a task may be queued and dequeued multiple
617 * times as it is shuffled about, we're really interested in knowing how
618 * long it was from the *first* time it was queued to the time that it
621 static inline void sched_info_dequeued(struct task_struct
*t
)
623 t
->sched_info
.last_queued
= 0;
627 * Called when a task finally hits the cpu. We can now calculate how
628 * long it was waiting to run. We also note when it began so that we
629 * can keep stats on how long its timeslice is.
631 static void sched_info_arrive(struct task_struct
*t
)
633 unsigned long now
= jiffies
, delta_jiffies
= 0;
635 if (t
->sched_info
.last_queued
)
636 delta_jiffies
= now
- t
->sched_info
.last_queued
;
637 sched_info_dequeued(t
);
638 t
->sched_info
.run_delay
+= delta_jiffies
;
639 t
->sched_info
.last_arrival
= now
;
640 t
->sched_info
.pcnt
++;
642 rq_sched_info_arrive(task_rq(t
), delta_jiffies
);
646 * Called when a process is queued into either the active or expired
647 * array. The time is noted and later used to determine how long we
648 * had to wait for us to reach the cpu. Since the expired queue will
649 * become the active queue after active queue is empty, without dequeuing
650 * and requeuing any tasks, we are interested in queuing to either. It
651 * is unusual but not impossible for tasks to be dequeued and immediately
652 * requeued in the same or another array: this can happen in sched_yield(),
653 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
656 * This function is only called from enqueue_task(), but also only updates
657 * the timestamp if it is already not set. It's assumed that
658 * sched_info_dequeued() will clear that stamp when appropriate.
660 static inline void sched_info_queued(struct task_struct
*t
)
662 if (unlikely(sched_info_on()))
663 if (!t
->sched_info
.last_queued
)
664 t
->sched_info
.last_queued
= jiffies
;
668 * Called when a process ceases being the active-running process, either
669 * voluntarily or involuntarily. Now we can calculate how long we ran.
671 static inline void sched_info_depart(struct task_struct
*t
)
673 unsigned long delta_jiffies
= jiffies
- t
->sched_info
.last_arrival
;
675 t
->sched_info
.cpu_time
+= delta_jiffies
;
676 rq_sched_info_depart(task_rq(t
), delta_jiffies
);
680 * Called when tasks are switched involuntarily due, typically, to expiring
681 * their time slice. (This may also be called when switching to or from
682 * the idle task.) We are only called when prev != next.
685 __sched_info_switch(struct task_struct
*prev
, struct task_struct
*next
)
687 struct rq
*rq
= task_rq(prev
);
690 * prev now departs the cpu. It's not interesting to record
691 * stats about how efficient we were at scheduling the idle
694 if (prev
!= rq
->idle
)
695 sched_info_depart(prev
);
697 if (next
!= rq
->idle
)
698 sched_info_arrive(next
);
701 sched_info_switch(struct task_struct
*prev
, struct task_struct
*next
)
703 if (unlikely(sched_info_on()))
704 __sched_info_switch(prev
, next
);
707 #define sched_info_queued(t) do { } while (0)
708 #define sched_info_switch(t, next) do { } while (0)
709 #endif /* CONFIG_SCHEDSTATS || CONFIG_TASK_DELAY_ACCT */
712 * Adding/removing a task to/from a priority array:
714 static void dequeue_task(struct task_struct
*p
, struct prio_array
*array
)
717 list_del(&p
->run_list
);
718 if (list_empty(array
->queue
+ p
->prio
))
719 __clear_bit(p
->prio
, array
->bitmap
);
722 static void enqueue_task(struct task_struct
*p
, struct prio_array
*array
)
724 sched_info_queued(p
);
725 list_add_tail(&p
->run_list
, array
->queue
+ p
->prio
);
726 __set_bit(p
->prio
, array
->bitmap
);
732 * Put task to the end of the run list without the overhead of dequeue
733 * followed by enqueue.
735 static void requeue_task(struct task_struct
*p
, struct prio_array
*array
)
737 list_move_tail(&p
->run_list
, array
->queue
+ p
->prio
);
741 enqueue_task_head(struct task_struct
*p
, struct prio_array
*array
)
743 list_add(&p
->run_list
, array
->queue
+ p
->prio
);
744 __set_bit(p
->prio
, array
->bitmap
);
750 * __normal_prio - return the priority that is based on the static
751 * priority but is modified by bonuses/penalties.
753 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
754 * into the -5 ... 0 ... +5 bonus/penalty range.
756 * We use 25% of the full 0...39 priority range so that:
758 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
759 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
761 * Both properties are important to certain workloads.
764 static inline int __normal_prio(struct task_struct
*p
)
768 bonus
= CURRENT_BONUS(p
) - MAX_BONUS
/ 2;
770 prio
= p
->static_prio
- bonus
;
771 if (prio
< MAX_RT_PRIO
)
773 if (prio
> MAX_PRIO
-1)
779 * To aid in avoiding the subversion of "niceness" due to uneven distribution
780 * of tasks with abnormal "nice" values across CPUs the contribution that
781 * each task makes to its run queue's load is weighted according to its
782 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
783 * scaled version of the new time slice allocation that they receive on time
788 * Assume: static_prio_timeslice(NICE_TO_PRIO(0)) == DEF_TIMESLICE
789 * If static_prio_timeslice() is ever changed to break this assumption then
790 * this code will need modification
792 #define TIME_SLICE_NICE_ZERO DEF_TIMESLICE
793 #define LOAD_WEIGHT(lp) \
794 (((lp) * SCHED_LOAD_SCALE) / TIME_SLICE_NICE_ZERO)
795 #define PRIO_TO_LOAD_WEIGHT(prio) \
796 LOAD_WEIGHT(static_prio_timeslice(prio))
797 #define RTPRIO_TO_LOAD_WEIGHT(rp) \
798 (PRIO_TO_LOAD_WEIGHT(MAX_RT_PRIO) + LOAD_WEIGHT(rp))
800 static void set_load_weight(struct task_struct
*p
)
802 if (has_rt_policy(p
)) {
804 if (p
== task_rq(p
)->migration_thread
)
806 * The migration thread does the actual balancing.
807 * Giving its load any weight will skew balancing
813 p
->load_weight
= RTPRIO_TO_LOAD_WEIGHT(p
->rt_priority
);
815 p
->load_weight
= PRIO_TO_LOAD_WEIGHT(p
->static_prio
);
819 inc_raw_weighted_load(struct rq
*rq
, const struct task_struct
*p
)
821 rq
->raw_weighted_load
+= p
->load_weight
;
825 dec_raw_weighted_load(struct rq
*rq
, const struct task_struct
*p
)
827 rq
->raw_weighted_load
-= p
->load_weight
;
830 static inline void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
833 inc_raw_weighted_load(rq
, p
);
836 static inline void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
839 dec_raw_weighted_load(rq
, p
);
843 * Calculate the expected normal priority: i.e. priority
844 * without taking RT-inheritance into account. Might be
845 * boosted by interactivity modifiers. Changes upon fork,
846 * setprio syscalls, and whenever the interactivity
847 * estimator recalculates.
849 static inline int normal_prio(struct task_struct
*p
)
853 if (has_rt_policy(p
))
854 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
856 prio
= __normal_prio(p
);
861 * Calculate the current priority, i.e. the priority
862 * taken into account by the scheduler. This value might
863 * be boosted by RT tasks, or might be boosted by
864 * interactivity modifiers. Will be RT if the task got
865 * RT-boosted. If not then it returns p->normal_prio.
867 static int effective_prio(struct task_struct
*p
)
869 p
->normal_prio
= normal_prio(p
);
871 * If we are RT tasks or we were boosted to RT priority,
872 * keep the priority unchanged. Otherwise, update priority
873 * to the normal priority:
875 if (!rt_prio(p
->prio
))
876 return p
->normal_prio
;
881 * __activate_task - move a task to the runqueue.
883 static void __activate_task(struct task_struct
*p
, struct rq
*rq
)
885 struct prio_array
*target
= rq
->active
;
888 target
= rq
->expired
;
889 enqueue_task(p
, target
);
890 inc_nr_running(p
, rq
);
894 * __activate_idle_task - move idle task to the _front_ of runqueue.
896 static inline void __activate_idle_task(struct task_struct
*p
, struct rq
*rq
)
898 enqueue_task_head(p
, rq
->active
);
899 inc_nr_running(p
, rq
);
903 * Recalculate p->normal_prio and p->prio after having slept,
904 * updating the sleep-average too:
906 static int recalc_task_prio(struct task_struct
*p
, unsigned long long now
)
908 /* Caller must always ensure 'now >= p->timestamp' */
909 unsigned long sleep_time
= now
- p
->timestamp
;
914 if (likely(sleep_time
> 0)) {
916 * This ceiling is set to the lowest priority that would allow
917 * a task to be reinserted into the active array on timeslice
920 unsigned long ceiling
= INTERACTIVE_SLEEP(p
);
922 if (p
->mm
&& sleep_time
> ceiling
&& p
->sleep_avg
< ceiling
) {
924 * Prevents user tasks from achieving best priority
925 * with one single large enough sleep.
927 p
->sleep_avg
= ceiling
;
929 * Using INTERACTIVE_SLEEP() as a ceiling places a
930 * nice(0) task 1ms sleep away from promotion, and
931 * gives it 700ms to round-robin with no chance of
932 * being demoted. This is more than generous, so
933 * mark this sleep as non-interactive to prevent the
934 * on-runqueue bonus logic from intervening should
935 * this task not receive cpu immediately.
937 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
940 * Tasks waking from uninterruptible sleep are
941 * limited in their sleep_avg rise as they
942 * are likely to be waiting on I/O
944 if (p
->sleep_type
== SLEEP_NONINTERACTIVE
&& p
->mm
) {
945 if (p
->sleep_avg
>= ceiling
)
947 else if (p
->sleep_avg
+ sleep_time
>=
949 p
->sleep_avg
= ceiling
;
955 * This code gives a bonus to interactive tasks.
957 * The boost works by updating the 'average sleep time'
958 * value here, based on ->timestamp. The more time a
959 * task spends sleeping, the higher the average gets -
960 * and the higher the priority boost gets as well.
962 p
->sleep_avg
+= sleep_time
;
965 if (p
->sleep_avg
> NS_MAX_SLEEP_AVG
)
966 p
->sleep_avg
= NS_MAX_SLEEP_AVG
;
969 return effective_prio(p
);
973 * activate_task - move a task to the runqueue and do priority recalculation
975 * Update all the scheduling statistics stuff. (sleep average
976 * calculation, priority modifiers, etc.)
978 static void activate_task(struct task_struct
*p
, struct rq
*rq
, int local
)
980 unsigned long long now
;
988 /* Compensate for drifting sched_clock */
989 struct rq
*this_rq
= this_rq();
990 now
= (now
- this_rq
->most_recent_timestamp
)
991 + rq
->most_recent_timestamp
;
996 * Sleep time is in units of nanosecs, so shift by 20 to get a
997 * milliseconds-range estimation of the amount of time that the task
1000 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
1001 if (p
->state
== TASK_UNINTERRUPTIBLE
)
1002 profile_hits(SLEEP_PROFILING
, (void *)get_wchan(p
),
1003 (now
- p
->timestamp
) >> 20);
1006 p
->prio
= recalc_task_prio(p
, now
);
1009 * This checks to make sure it's not an uninterruptible task
1010 * that is now waking up.
1012 if (p
->sleep_type
== SLEEP_NORMAL
) {
1014 * Tasks which were woken up by interrupts (ie. hw events)
1015 * are most likely of interactive nature. So we give them
1016 * the credit of extending their sleep time to the period
1017 * of time they spend on the runqueue, waiting for execution
1018 * on a CPU, first time around:
1021 p
->sleep_type
= SLEEP_INTERRUPTED
;
1024 * Normal first-time wakeups get a credit too for
1025 * on-runqueue time, but it will be weighted down:
1027 p
->sleep_type
= SLEEP_INTERACTIVE
;
1032 __activate_task(p
, rq
);
1036 * deactivate_task - remove a task from the runqueue.
1038 static void deactivate_task(struct task_struct
*p
, struct rq
*rq
)
1040 dec_nr_running(p
, rq
);
1041 dequeue_task(p
, p
->array
);
1046 * resched_task - mark a task 'to be rescheduled now'.
1048 * On UP this means the setting of the need_resched flag, on SMP it
1049 * might also involve a cross-CPU call to trigger the scheduler on
1054 #ifndef tsk_is_polling
1055 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1058 static void resched_task(struct task_struct
*p
)
1062 assert_spin_locked(&task_rq(p
)->lock
);
1064 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
1067 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
1070 if (cpu
== smp_processor_id())
1073 /* NEED_RESCHED must be visible before we test polling */
1075 if (!tsk_is_polling(p
))
1076 smp_send_reschedule(cpu
);
1079 static void resched_cpu(int cpu
)
1081 struct rq
*rq
= cpu_rq(cpu
);
1082 unsigned long flags
;
1084 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1086 resched_task(cpu_curr(cpu
));
1087 spin_unlock_irqrestore(&rq
->lock
, flags
);
1090 static inline void resched_task(struct task_struct
*p
)
1092 assert_spin_locked(&task_rq(p
)->lock
);
1093 set_tsk_need_resched(p
);
1098 * task_curr - is this task currently executing on a CPU?
1099 * @p: the task in question.
1101 inline int task_curr(const struct task_struct
*p
)
1103 return cpu_curr(task_cpu(p
)) == p
;
1106 /* Used instead of source_load when we know the type == 0 */
1107 unsigned long weighted_cpuload(const int cpu
)
1109 return cpu_rq(cpu
)->raw_weighted_load
;
1113 struct migration_req
{
1114 struct list_head list
;
1116 struct task_struct
*task
;
1119 struct completion done
;
1123 * The task's runqueue lock must be held.
1124 * Returns true if you have to wait for migration thread.
1127 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1129 struct rq
*rq
= task_rq(p
);
1132 * If the task is not on a runqueue (and not running), then
1133 * it is sufficient to simply update the task's cpu field.
1135 if (!p
->array
&& !task_running(rq
, p
)) {
1136 set_task_cpu(p
, dest_cpu
);
1140 init_completion(&req
->done
);
1142 req
->dest_cpu
= dest_cpu
;
1143 list_add(&req
->list
, &rq
->migration_queue
);
1149 * wait_task_inactive - wait for a thread to unschedule.
1151 * The caller must ensure that the task *will* unschedule sometime soon,
1152 * else this function might spin for a *long* time. This function can't
1153 * be called with interrupts off, or it may introduce deadlock with
1154 * smp_call_function() if an IPI is sent by the same process we are
1155 * waiting to become inactive.
1157 void wait_task_inactive(struct task_struct
*p
)
1159 unsigned long flags
;
1164 rq
= task_rq_lock(p
, &flags
);
1165 /* Must be off runqueue entirely, not preempted. */
1166 if (unlikely(p
->array
|| task_running(rq
, p
))) {
1167 /* If it's preempted, we yield. It could be a while. */
1168 preempted
= !task_running(rq
, p
);
1169 task_rq_unlock(rq
, &flags
);
1175 task_rq_unlock(rq
, &flags
);
1179 * kick_process - kick a running thread to enter/exit the kernel
1180 * @p: the to-be-kicked thread
1182 * Cause a process which is running on another CPU to enter
1183 * kernel-mode, without any delay. (to get signals handled.)
1185 * NOTE: this function doesnt have to take the runqueue lock,
1186 * because all it wants to ensure is that the remote task enters
1187 * the kernel. If the IPI races and the task has been migrated
1188 * to another CPU then no harm is done and the purpose has been
1191 void kick_process(struct task_struct
*p
)
1197 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1198 smp_send_reschedule(cpu
);
1203 * Return a low guess at the load of a migration-source cpu weighted
1204 * according to the scheduling class and "nice" value.
1206 * We want to under-estimate the load of migration sources, to
1207 * balance conservatively.
1209 static inline unsigned long source_load(int cpu
, int type
)
1211 struct rq
*rq
= cpu_rq(cpu
);
1214 return rq
->raw_weighted_load
;
1216 return min(rq
->cpu_load
[type
-1], rq
->raw_weighted_load
);
1220 * Return a high guess at the load of a migration-target cpu weighted
1221 * according to the scheduling class and "nice" value.
1223 static inline unsigned long target_load(int cpu
, int type
)
1225 struct rq
*rq
= cpu_rq(cpu
);
1228 return rq
->raw_weighted_load
;
1230 return max(rq
->cpu_load
[type
-1], rq
->raw_weighted_load
);
1234 * Return the average load per task on the cpu's run queue
1236 static inline unsigned long cpu_avg_load_per_task(int cpu
)
1238 struct rq
*rq
= cpu_rq(cpu
);
1239 unsigned long n
= rq
->nr_running
;
1241 return n
? rq
->raw_weighted_load
/ n
: SCHED_LOAD_SCALE
;
1245 * find_idlest_group finds and returns the least busy CPU group within the
1248 static struct sched_group
*
1249 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1251 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1252 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1253 int load_idx
= sd
->forkexec_idx
;
1254 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1257 unsigned long load
, avg_load
;
1261 /* Skip over this group if it has no CPUs allowed */
1262 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1265 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1267 /* Tally up the load of all CPUs in the group */
1270 for_each_cpu_mask(i
, group
->cpumask
) {
1271 /* Bias balancing toward cpus of our domain */
1273 load
= source_load(i
, load_idx
);
1275 load
= target_load(i
, load_idx
);
1280 /* Adjust by relative CPU power of the group */
1281 avg_load
= sg_div_cpu_power(group
,
1282 avg_load
* SCHED_LOAD_SCALE
);
1285 this_load
= avg_load
;
1287 } else if (avg_load
< min_load
) {
1288 min_load
= avg_load
;
1292 group
= group
->next
;
1293 } while (group
!= sd
->groups
);
1295 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1301 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1304 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1307 unsigned long load
, min_load
= ULONG_MAX
;
1311 /* Traverse only the allowed CPUs */
1312 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1314 for_each_cpu_mask(i
, tmp
) {
1315 load
= weighted_cpuload(i
);
1317 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1327 * sched_balance_self: balance the current task (running on cpu) in domains
1328 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1331 * Balance, ie. select the least loaded group.
1333 * Returns the target CPU number, or the same CPU if no balancing is needed.
1335 * preempt must be disabled.
1337 static int sched_balance_self(int cpu
, int flag
)
1339 struct task_struct
*t
= current
;
1340 struct sched_domain
*tmp
, *sd
= NULL
;
1342 for_each_domain(cpu
, tmp
) {
1344 * If power savings logic is enabled for a domain, stop there.
1346 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1348 if (tmp
->flags
& flag
)
1354 struct sched_group
*group
;
1355 int new_cpu
, weight
;
1357 if (!(sd
->flags
& flag
)) {
1363 group
= find_idlest_group(sd
, t
, cpu
);
1369 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1370 if (new_cpu
== -1 || new_cpu
== cpu
) {
1371 /* Now try balancing at a lower domain level of cpu */
1376 /* Now try balancing at a lower domain level of new_cpu */
1379 weight
= cpus_weight(span
);
1380 for_each_domain(cpu
, tmp
) {
1381 if (weight
<= cpus_weight(tmp
->span
))
1383 if (tmp
->flags
& flag
)
1386 /* while loop will break here if sd == NULL */
1392 #endif /* CONFIG_SMP */
1395 * wake_idle() will wake a task on an idle cpu if task->cpu is
1396 * not idle and an idle cpu is available. The span of cpus to
1397 * search starts with cpus closest then further out as needed,
1398 * so we always favor a closer, idle cpu.
1400 * Returns the CPU we should wake onto.
1402 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1403 static int wake_idle(int cpu
, struct task_struct
*p
)
1406 struct sched_domain
*sd
;
1410 * If it is idle, then it is the best cpu to run this task.
1412 * This cpu is also the best, if it has more than one task already.
1413 * Siblings must be also busy(in most cases) as they didn't already
1414 * pickup the extra load from this cpu and hence we need not check
1415 * sibling runqueue info. This will avoid the checks and cache miss
1416 * penalities associated with that.
1418 if (idle_cpu(cpu
) || cpu_rq(cpu
)->nr_running
> 1)
1421 for_each_domain(cpu
, sd
) {
1422 if (sd
->flags
& SD_WAKE_IDLE
) {
1423 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1424 for_each_cpu_mask(i
, tmp
) {
1435 static inline int wake_idle(int cpu
, struct task_struct
*p
)
1442 * try_to_wake_up - wake up a thread
1443 * @p: the to-be-woken-up thread
1444 * @state: the mask of task states that can be woken
1445 * @sync: do a synchronous wakeup?
1447 * Put it on the run-queue if it's not already there. The "current"
1448 * thread is always on the run-queue (except when the actual
1449 * re-schedule is in progress), and as such you're allowed to do
1450 * the simpler "current->state = TASK_RUNNING" to mark yourself
1451 * runnable without the overhead of this.
1453 * returns failure only if the task is already active.
1455 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1457 int cpu
, this_cpu
, success
= 0;
1458 unsigned long flags
;
1462 struct sched_domain
*sd
, *this_sd
= NULL
;
1463 unsigned long load
, this_load
;
1467 rq
= task_rq_lock(p
, &flags
);
1468 old_state
= p
->state
;
1469 if (!(old_state
& state
))
1476 this_cpu
= smp_processor_id();
1479 if (unlikely(task_running(rq
, p
)))
1484 schedstat_inc(rq
, ttwu_cnt
);
1485 if (cpu
== this_cpu
) {
1486 schedstat_inc(rq
, ttwu_local
);
1490 for_each_domain(this_cpu
, sd
) {
1491 if (cpu_isset(cpu
, sd
->span
)) {
1492 schedstat_inc(sd
, ttwu_wake_remote
);
1498 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1502 * Check for affine wakeup and passive balancing possibilities.
1505 int idx
= this_sd
->wake_idx
;
1506 unsigned int imbalance
;
1508 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1510 load
= source_load(cpu
, idx
);
1511 this_load
= target_load(this_cpu
, idx
);
1513 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1515 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1516 unsigned long tl
= this_load
;
1517 unsigned long tl_per_task
;
1519 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1522 * If sync wakeup then subtract the (maximum possible)
1523 * effect of the currently running task from the load
1524 * of the current CPU:
1527 tl
-= current
->load_weight
;
1530 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1531 100*(tl
+ p
->load_weight
) <= imbalance
*load
) {
1533 * This domain has SD_WAKE_AFFINE and
1534 * p is cache cold in this domain, and
1535 * there is no bad imbalance.
1537 schedstat_inc(this_sd
, ttwu_move_affine
);
1543 * Start passive balancing when half the imbalance_pct
1546 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1547 if (imbalance
*this_load
<= 100*load
) {
1548 schedstat_inc(this_sd
, ttwu_move_balance
);
1554 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1556 new_cpu
= wake_idle(new_cpu
, p
);
1557 if (new_cpu
!= cpu
) {
1558 set_task_cpu(p
, new_cpu
);
1559 task_rq_unlock(rq
, &flags
);
1560 /* might preempt at this point */
1561 rq
= task_rq_lock(p
, &flags
);
1562 old_state
= p
->state
;
1563 if (!(old_state
& state
))
1568 this_cpu
= smp_processor_id();
1573 #endif /* CONFIG_SMP */
1574 if (old_state
== TASK_UNINTERRUPTIBLE
) {
1575 rq
->nr_uninterruptible
--;
1577 * Tasks on involuntary sleep don't earn
1578 * sleep_avg beyond just interactive state.
1580 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
1584 * Tasks that have marked their sleep as noninteractive get
1585 * woken up with their sleep average not weighted in an
1588 if (old_state
& TASK_NONINTERACTIVE
)
1589 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
1592 activate_task(p
, rq
, cpu
== this_cpu
);
1594 * Sync wakeups (i.e. those types of wakeups where the waker
1595 * has indicated that it will leave the CPU in short order)
1596 * don't trigger a preemption, if the woken up task will run on
1597 * this cpu. (in this case the 'I will reschedule' promise of
1598 * the waker guarantees that the freshly woken up task is going
1599 * to be considered on this CPU.)
1601 if (!sync
|| cpu
!= this_cpu
) {
1602 if (TASK_PREEMPTS_CURR(p
, rq
))
1603 resched_task(rq
->curr
);
1608 p
->state
= TASK_RUNNING
;
1610 task_rq_unlock(rq
, &flags
);
1615 int fastcall
wake_up_process(struct task_struct
*p
)
1617 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1618 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1620 EXPORT_SYMBOL(wake_up_process
);
1622 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1624 return try_to_wake_up(p
, state
, 0);
1627 static void task_running_tick(struct rq
*rq
, struct task_struct
*p
);
1629 * Perform scheduler related setup for a newly forked process p.
1630 * p is forked by current.
1632 void fastcall
sched_fork(struct task_struct
*p
, int clone_flags
)
1634 int cpu
= get_cpu();
1637 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1639 set_task_cpu(p
, cpu
);
1642 * We mark the process as running here, but have not actually
1643 * inserted it onto the runqueue yet. This guarantees that
1644 * nobody will actually run it, and a signal or other external
1645 * event cannot wake it up and insert it on the runqueue either.
1647 p
->state
= TASK_RUNNING
;
1650 * Make sure we do not leak PI boosting priority to the child:
1652 p
->prio
= current
->normal_prio
;
1654 INIT_LIST_HEAD(&p
->run_list
);
1656 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1657 if (unlikely(sched_info_on()))
1658 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1660 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1663 #ifdef CONFIG_PREEMPT
1664 /* Want to start with kernel preemption disabled. */
1665 task_thread_info(p
)->preempt_count
= 1;
1668 * Share the timeslice between parent and child, thus the
1669 * total amount of pending timeslices in the system doesn't change,
1670 * resulting in more scheduling fairness.
1672 local_irq_disable();
1673 p
->time_slice
= (current
->time_slice
+ 1) >> 1;
1675 * The remainder of the first timeslice might be recovered by
1676 * the parent if the child exits early enough.
1678 p
->first_time_slice
= 1;
1679 current
->time_slice
>>= 1;
1680 p
->timestamp
= sched_clock();
1681 if (unlikely(!current
->time_slice
)) {
1683 * This case is rare, it happens when the parent has only
1684 * a single jiffy left from its timeslice. Taking the
1685 * runqueue lock is not a problem.
1687 current
->time_slice
= 1;
1688 task_running_tick(cpu_rq(cpu
), current
);
1695 * wake_up_new_task - wake up a newly created task for the first time.
1697 * This function will do some initial scheduler statistics housekeeping
1698 * that must be done for every newly created context, then puts the task
1699 * on the runqueue and wakes it.
1701 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1703 struct rq
*rq
, *this_rq
;
1704 unsigned long flags
;
1707 rq
= task_rq_lock(p
, &flags
);
1708 BUG_ON(p
->state
!= TASK_RUNNING
);
1709 this_cpu
= smp_processor_id();
1713 * We decrease the sleep average of forking parents
1714 * and children as well, to keep max-interactive tasks
1715 * from forking tasks that are max-interactive. The parent
1716 * (current) is done further down, under its lock.
1718 p
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(p
) *
1719 CHILD_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1721 p
->prio
= effective_prio(p
);
1723 if (likely(cpu
== this_cpu
)) {
1724 if (!(clone_flags
& CLONE_VM
)) {
1726 * The VM isn't cloned, so we're in a good position to
1727 * do child-runs-first in anticipation of an exec. This
1728 * usually avoids a lot of COW overhead.
1730 if (unlikely(!current
->array
))
1731 __activate_task(p
, rq
);
1733 p
->prio
= current
->prio
;
1734 p
->normal_prio
= current
->normal_prio
;
1735 list_add_tail(&p
->run_list
, ¤t
->run_list
);
1736 p
->array
= current
->array
;
1737 p
->array
->nr_active
++;
1738 inc_nr_running(p
, rq
);
1742 /* Run child last */
1743 __activate_task(p
, rq
);
1745 * We skip the following code due to cpu == this_cpu
1747 * task_rq_unlock(rq, &flags);
1748 * this_rq = task_rq_lock(current, &flags);
1752 this_rq
= cpu_rq(this_cpu
);
1755 * Not the local CPU - must adjust timestamp. This should
1756 * get optimised away in the !CONFIG_SMP case.
1758 p
->timestamp
= (p
->timestamp
- this_rq
->most_recent_timestamp
)
1759 + rq
->most_recent_timestamp
;
1760 __activate_task(p
, rq
);
1761 if (TASK_PREEMPTS_CURR(p
, rq
))
1762 resched_task(rq
->curr
);
1765 * Parent and child are on different CPUs, now get the
1766 * parent runqueue to update the parent's ->sleep_avg:
1768 task_rq_unlock(rq
, &flags
);
1769 this_rq
= task_rq_lock(current
, &flags
);
1771 current
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(current
) *
1772 PARENT_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1773 task_rq_unlock(this_rq
, &flags
);
1777 * Potentially available exiting-child timeslices are
1778 * retrieved here - this way the parent does not get
1779 * penalized for creating too many threads.
1781 * (this cannot be used to 'generate' timeslices
1782 * artificially, because any timeslice recovered here
1783 * was given away by the parent in the first place.)
1785 void fastcall
sched_exit(struct task_struct
*p
)
1787 unsigned long flags
;
1791 * If the child was a (relative-) CPU hog then decrease
1792 * the sleep_avg of the parent as well.
1794 rq
= task_rq_lock(p
->parent
, &flags
);
1795 if (p
->first_time_slice
&& task_cpu(p
) == task_cpu(p
->parent
)) {
1796 p
->parent
->time_slice
+= p
->time_slice
;
1797 if (unlikely(p
->parent
->time_slice
> task_timeslice(p
)))
1798 p
->parent
->time_slice
= task_timeslice(p
);
1800 if (p
->sleep_avg
< p
->parent
->sleep_avg
)
1801 p
->parent
->sleep_avg
= p
->parent
->sleep_avg
/
1802 (EXIT_WEIGHT
+ 1) * EXIT_WEIGHT
+ p
->sleep_avg
/
1804 task_rq_unlock(rq
, &flags
);
1808 * prepare_task_switch - prepare to switch tasks
1809 * @rq: the runqueue preparing to switch
1810 * @next: the task we are going to switch to.
1812 * This is called with the rq lock held and interrupts off. It must
1813 * be paired with a subsequent finish_task_switch after the context
1816 * prepare_task_switch sets up locking and calls architecture specific
1819 static inline void prepare_task_switch(struct rq
*rq
, struct task_struct
*next
)
1821 prepare_lock_switch(rq
, next
);
1822 prepare_arch_switch(next
);
1826 * finish_task_switch - clean up after a task-switch
1827 * @rq: runqueue associated with task-switch
1828 * @prev: the thread we just switched away from.
1830 * finish_task_switch must be called after the context switch, paired
1831 * with a prepare_task_switch call before the context switch.
1832 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1833 * and do any other architecture-specific cleanup actions.
1835 * Note that we may have delayed dropping an mm in context_switch(). If
1836 * so, we finish that here outside of the runqueue lock. (Doing it
1837 * with the lock held can cause deadlocks; see schedule() for
1840 static inline void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1841 __releases(rq
->lock
)
1843 struct mm_struct
*mm
= rq
->prev_mm
;
1849 * A task struct has one reference for the use as "current".
1850 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1851 * schedule one last time. The schedule call will never return, and
1852 * the scheduled task must drop that reference.
1853 * The test for TASK_DEAD must occur while the runqueue locks are
1854 * still held, otherwise prev could be scheduled on another cpu, die
1855 * there before we look at prev->state, and then the reference would
1857 * Manfred Spraul <manfred@colorfullife.com>
1859 prev_state
= prev
->state
;
1860 finish_arch_switch(prev
);
1861 finish_lock_switch(rq
, prev
);
1864 if (unlikely(prev_state
== TASK_DEAD
)) {
1866 * Remove function-return probe instances associated with this
1867 * task and put them back on the free list.
1869 kprobe_flush_task(prev
);
1870 put_task_struct(prev
);
1875 * schedule_tail - first thing a freshly forked thread must call.
1876 * @prev: the thread we just switched away from.
1878 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1879 __releases(rq
->lock
)
1881 struct rq
*rq
= this_rq();
1883 finish_task_switch(rq
, prev
);
1884 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1885 /* In this case, finish_task_switch does not reenable preemption */
1888 if (current
->set_child_tid
)
1889 put_user(current
->pid
, current
->set_child_tid
);
1893 * context_switch - switch to the new MM and the new
1894 * thread's register state.
1896 static inline struct task_struct
*
1897 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1898 struct task_struct
*next
)
1900 struct mm_struct
*mm
= next
->mm
;
1901 struct mm_struct
*oldmm
= prev
->active_mm
;
1904 * For paravirt, this is coupled with an exit in switch_to to
1905 * combine the page table reload and the switch backend into
1908 arch_enter_lazy_cpu_mode();
1911 next
->active_mm
= oldmm
;
1912 atomic_inc(&oldmm
->mm_count
);
1913 enter_lazy_tlb(oldmm
, next
);
1915 switch_mm(oldmm
, mm
, next
);
1918 prev
->active_mm
= NULL
;
1919 WARN_ON(rq
->prev_mm
);
1920 rq
->prev_mm
= oldmm
;
1923 * Since the runqueue lock will be released by the next
1924 * task (which is an invalid locking op but in the case
1925 * of the scheduler it's an obvious special-case), so we
1926 * do an early lockdep release here:
1928 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1929 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1932 /* Here we just switch the register state and the stack. */
1933 switch_to(prev
, next
, prev
);
1939 * nr_running, nr_uninterruptible and nr_context_switches:
1941 * externally visible scheduler statistics: current number of runnable
1942 * threads, current number of uninterruptible-sleeping threads, total
1943 * number of context switches performed since bootup.
1945 unsigned long nr_running(void)
1947 unsigned long i
, sum
= 0;
1949 for_each_online_cpu(i
)
1950 sum
+= cpu_rq(i
)->nr_running
;
1955 unsigned long nr_uninterruptible(void)
1957 unsigned long i
, sum
= 0;
1959 for_each_possible_cpu(i
)
1960 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1963 * Since we read the counters lockless, it might be slightly
1964 * inaccurate. Do not allow it to go below zero though:
1966 if (unlikely((long)sum
< 0))
1972 unsigned long long nr_context_switches(void)
1975 unsigned long long sum
= 0;
1977 for_each_possible_cpu(i
)
1978 sum
+= cpu_rq(i
)->nr_switches
;
1983 unsigned long nr_iowait(void)
1985 unsigned long i
, sum
= 0;
1987 for_each_possible_cpu(i
)
1988 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1993 unsigned long nr_active(void)
1995 unsigned long i
, running
= 0, uninterruptible
= 0;
1997 for_each_online_cpu(i
) {
1998 running
+= cpu_rq(i
)->nr_running
;
1999 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2002 if (unlikely((long)uninterruptible
< 0))
2003 uninterruptible
= 0;
2005 return running
+ uninterruptible
;
2011 * Is this task likely cache-hot:
2014 task_hot(struct task_struct
*p
, unsigned long long now
, struct sched_domain
*sd
)
2016 return (long long)(now
- p
->last_ran
) < (long long)sd
->cache_hot_time
;
2020 * double_rq_lock - safely lock two runqueues
2022 * Note this does not disable interrupts like task_rq_lock,
2023 * you need to do so manually before calling.
2025 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2026 __acquires(rq1
->lock
)
2027 __acquires(rq2
->lock
)
2029 BUG_ON(!irqs_disabled());
2031 spin_lock(&rq1
->lock
);
2032 __acquire(rq2
->lock
); /* Fake it out ;) */
2035 spin_lock(&rq1
->lock
);
2036 spin_lock(&rq2
->lock
);
2038 spin_lock(&rq2
->lock
);
2039 spin_lock(&rq1
->lock
);
2045 * double_rq_unlock - safely unlock two runqueues
2047 * Note this does not restore interrupts like task_rq_unlock,
2048 * you need to do so manually after calling.
2050 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2051 __releases(rq1
->lock
)
2052 __releases(rq2
->lock
)
2054 spin_unlock(&rq1
->lock
);
2056 spin_unlock(&rq2
->lock
);
2058 __release(rq2
->lock
);
2062 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2064 static void double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2065 __releases(this_rq
->lock
)
2066 __acquires(busiest
->lock
)
2067 __acquires(this_rq
->lock
)
2069 if (unlikely(!irqs_disabled())) {
2070 /* printk() doesn't work good under rq->lock */
2071 spin_unlock(&this_rq
->lock
);
2074 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2075 if (busiest
< this_rq
) {
2076 spin_unlock(&this_rq
->lock
);
2077 spin_lock(&busiest
->lock
);
2078 spin_lock(&this_rq
->lock
);
2080 spin_lock(&busiest
->lock
);
2085 * If dest_cpu is allowed for this process, migrate the task to it.
2086 * This is accomplished by forcing the cpu_allowed mask to only
2087 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2088 * the cpu_allowed mask is restored.
2090 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2092 struct migration_req req
;
2093 unsigned long flags
;
2096 rq
= task_rq_lock(p
, &flags
);
2097 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2098 || unlikely(cpu_is_offline(dest_cpu
)))
2101 /* force the process onto the specified CPU */
2102 if (migrate_task(p
, dest_cpu
, &req
)) {
2103 /* Need to wait for migration thread (might exit: take ref). */
2104 struct task_struct
*mt
= rq
->migration_thread
;
2106 get_task_struct(mt
);
2107 task_rq_unlock(rq
, &flags
);
2108 wake_up_process(mt
);
2109 put_task_struct(mt
);
2110 wait_for_completion(&req
.done
);
2115 task_rq_unlock(rq
, &flags
);
2119 * sched_exec - execve() is a valuable balancing opportunity, because at
2120 * this point the task has the smallest effective memory and cache footprint.
2122 void sched_exec(void)
2124 int new_cpu
, this_cpu
= get_cpu();
2125 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2127 if (new_cpu
!= this_cpu
)
2128 sched_migrate_task(current
, new_cpu
);
2132 * pull_task - move a task from a remote runqueue to the local runqueue.
2133 * Both runqueues must be locked.
2135 static void pull_task(struct rq
*src_rq
, struct prio_array
*src_array
,
2136 struct task_struct
*p
, struct rq
*this_rq
,
2137 struct prio_array
*this_array
, int this_cpu
)
2139 dequeue_task(p
, src_array
);
2140 dec_nr_running(p
, src_rq
);
2141 set_task_cpu(p
, this_cpu
);
2142 inc_nr_running(p
, this_rq
);
2143 enqueue_task(p
, this_array
);
2144 p
->timestamp
= (p
->timestamp
- src_rq
->most_recent_timestamp
)
2145 + this_rq
->most_recent_timestamp
;
2147 * Note that idle threads have a prio of MAX_PRIO, for this test
2148 * to be always true for them.
2150 if (TASK_PREEMPTS_CURR(p
, this_rq
))
2151 resched_task(this_rq
->curr
);
2155 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2158 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2159 struct sched_domain
*sd
, enum idle_type idle
,
2163 * We do not migrate tasks that are:
2164 * 1) running (obviously), or
2165 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2166 * 3) are cache-hot on their current CPU.
2168 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
2172 if (task_running(rq
, p
))
2176 * Aggressive migration if:
2177 * 1) task is cache cold, or
2178 * 2) too many balance attempts have failed.
2181 if (sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2182 #ifdef CONFIG_SCHEDSTATS
2183 if (task_hot(p
, rq
->most_recent_timestamp
, sd
))
2184 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2189 if (task_hot(p
, rq
->most_recent_timestamp
, sd
))
2194 #define rq_best_prio(rq) min((rq)->curr->prio, (rq)->best_expired_prio)
2197 * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted
2198 * load from busiest to this_rq, as part of a balancing operation within
2199 * "domain". Returns the number of tasks moved.
2201 * Called with both runqueues locked.
2203 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2204 unsigned long max_nr_move
, unsigned long max_load_move
,
2205 struct sched_domain
*sd
, enum idle_type idle
,
2208 int idx
, pulled
= 0, pinned
= 0, this_best_prio
, best_prio
,
2209 best_prio_seen
, skip_for_load
;
2210 struct prio_array
*array
, *dst_array
;
2211 struct list_head
*head
, *curr
;
2212 struct task_struct
*tmp
;
2215 if (max_nr_move
== 0 || max_load_move
== 0)
2218 rem_load_move
= max_load_move
;
2220 this_best_prio
= rq_best_prio(this_rq
);
2221 best_prio
= rq_best_prio(busiest
);
2223 * Enable handling of the case where there is more than one task
2224 * with the best priority. If the current running task is one
2225 * of those with prio==best_prio we know it won't be moved
2226 * and therefore it's safe to override the skip (based on load) of
2227 * any task we find with that prio.
2229 best_prio_seen
= best_prio
== busiest
->curr
->prio
;
2232 * We first consider expired tasks. Those will likely not be
2233 * executed in the near future, and they are most likely to
2234 * be cache-cold, thus switching CPUs has the least effect
2237 if (busiest
->expired
->nr_active
) {
2238 array
= busiest
->expired
;
2239 dst_array
= this_rq
->expired
;
2241 array
= busiest
->active
;
2242 dst_array
= this_rq
->active
;
2246 /* Start searching at priority 0: */
2250 idx
= sched_find_first_bit(array
->bitmap
);
2252 idx
= find_next_bit(array
->bitmap
, MAX_PRIO
, idx
);
2253 if (idx
>= MAX_PRIO
) {
2254 if (array
== busiest
->expired
&& busiest
->active
->nr_active
) {
2255 array
= busiest
->active
;
2256 dst_array
= this_rq
->active
;
2262 head
= array
->queue
+ idx
;
2265 tmp
= list_entry(curr
, struct task_struct
, run_list
);
2270 * To help distribute high priority tasks accross CPUs we don't
2271 * skip a task if it will be the highest priority task (i.e. smallest
2272 * prio value) on its new queue regardless of its load weight
2274 skip_for_load
= tmp
->load_weight
> rem_load_move
;
2275 if (skip_for_load
&& idx
< this_best_prio
)
2276 skip_for_load
= !best_prio_seen
&& idx
== best_prio
;
2277 if (skip_for_load
||
2278 !can_migrate_task(tmp
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2280 best_prio_seen
|= idx
== best_prio
;
2287 pull_task(busiest
, array
, tmp
, this_rq
, dst_array
, this_cpu
);
2289 rem_load_move
-= tmp
->load_weight
;
2292 * We only want to steal up to the prescribed number of tasks
2293 * and the prescribed amount of weighted load.
2295 if (pulled
< max_nr_move
&& rem_load_move
> 0) {
2296 if (idx
< this_best_prio
)
2297 this_best_prio
= idx
;
2305 * Right now, this is the only place pull_task() is called,
2306 * so we can safely collect pull_task() stats here rather than
2307 * inside pull_task().
2309 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2312 *all_pinned
= pinned
;
2317 * find_busiest_group finds and returns the busiest CPU group within the
2318 * domain. It calculates and returns the amount of weighted load which
2319 * should be moved to restore balance via the imbalance parameter.
2321 static struct sched_group
*
2322 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2323 unsigned long *imbalance
, enum idle_type idle
, int *sd_idle
,
2324 cpumask_t
*cpus
, int *balance
)
2326 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2327 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2328 unsigned long max_pull
;
2329 unsigned long busiest_load_per_task
, busiest_nr_running
;
2330 unsigned long this_load_per_task
, this_nr_running
;
2332 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2333 int power_savings_balance
= 1;
2334 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2335 unsigned long min_nr_running
= ULONG_MAX
;
2336 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2339 max_load
= this_load
= total_load
= total_pwr
= 0;
2340 busiest_load_per_task
= busiest_nr_running
= 0;
2341 this_load_per_task
= this_nr_running
= 0;
2342 if (idle
== NOT_IDLE
)
2343 load_idx
= sd
->busy_idx
;
2344 else if (idle
== NEWLY_IDLE
)
2345 load_idx
= sd
->newidle_idx
;
2347 load_idx
= sd
->idle_idx
;
2350 unsigned long load
, group_capacity
;
2353 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2354 unsigned long sum_nr_running
, sum_weighted_load
;
2356 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2359 balance_cpu
= first_cpu(group
->cpumask
);
2361 /* Tally up the load of all CPUs in the group */
2362 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2364 for_each_cpu_mask(i
, group
->cpumask
) {
2367 if (!cpu_isset(i
, *cpus
))
2372 if (*sd_idle
&& !idle_cpu(i
))
2375 /* Bias balancing toward cpus of our domain */
2377 if (idle_cpu(i
) && !first_idle_cpu
) {
2382 load
= target_load(i
, load_idx
);
2384 load
= source_load(i
, load_idx
);
2387 sum_nr_running
+= rq
->nr_running
;
2388 sum_weighted_load
+= rq
->raw_weighted_load
;
2392 * First idle cpu or the first cpu(busiest) in this sched group
2393 * is eligible for doing load balancing at this and above
2396 if (local_group
&& balance_cpu
!= this_cpu
&& balance
) {
2401 total_load
+= avg_load
;
2402 total_pwr
+= group
->__cpu_power
;
2404 /* Adjust by relative CPU power of the group */
2405 avg_load
= sg_div_cpu_power(group
,
2406 avg_load
* SCHED_LOAD_SCALE
);
2408 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
2411 this_load
= avg_load
;
2413 this_nr_running
= sum_nr_running
;
2414 this_load_per_task
= sum_weighted_load
;
2415 } else if (avg_load
> max_load
&&
2416 sum_nr_running
> group_capacity
) {
2417 max_load
= avg_load
;
2419 busiest_nr_running
= sum_nr_running
;
2420 busiest_load_per_task
= sum_weighted_load
;
2423 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2425 * Busy processors will not participate in power savings
2428 if (idle
== NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2432 * If the local group is idle or completely loaded
2433 * no need to do power savings balance at this domain
2435 if (local_group
&& (this_nr_running
>= group_capacity
||
2437 power_savings_balance
= 0;
2440 * If a group is already running at full capacity or idle,
2441 * don't include that group in power savings calculations
2443 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2448 * Calculate the group which has the least non-idle load.
2449 * This is the group from where we need to pick up the load
2452 if ((sum_nr_running
< min_nr_running
) ||
2453 (sum_nr_running
== min_nr_running
&&
2454 first_cpu(group
->cpumask
) <
2455 first_cpu(group_min
->cpumask
))) {
2457 min_nr_running
= sum_nr_running
;
2458 min_load_per_task
= sum_weighted_load
/
2463 * Calculate the group which is almost near its
2464 * capacity but still has some space to pick up some load
2465 * from other group and save more power
2467 if (sum_nr_running
<= group_capacity
- 1) {
2468 if (sum_nr_running
> leader_nr_running
||
2469 (sum_nr_running
== leader_nr_running
&&
2470 first_cpu(group
->cpumask
) >
2471 first_cpu(group_leader
->cpumask
))) {
2472 group_leader
= group
;
2473 leader_nr_running
= sum_nr_running
;
2478 group
= group
->next
;
2479 } while (group
!= sd
->groups
);
2481 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2484 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2486 if (this_load
>= avg_load
||
2487 100*max_load
<= sd
->imbalance_pct
*this_load
)
2490 busiest_load_per_task
/= busiest_nr_running
;
2492 * We're trying to get all the cpus to the average_load, so we don't
2493 * want to push ourselves above the average load, nor do we wish to
2494 * reduce the max loaded cpu below the average load, as either of these
2495 * actions would just result in more rebalancing later, and ping-pong
2496 * tasks around. Thus we look for the minimum possible imbalance.
2497 * Negative imbalances (*we* are more loaded than anyone else) will
2498 * be counted as no imbalance for these purposes -- we can't fix that
2499 * by pulling tasks to us. Be careful of negative numbers as they'll
2500 * appear as very large values with unsigned longs.
2502 if (max_load
<= busiest_load_per_task
)
2506 * In the presence of smp nice balancing, certain scenarios can have
2507 * max load less than avg load(as we skip the groups at or below
2508 * its cpu_power, while calculating max_load..)
2510 if (max_load
< avg_load
) {
2512 goto small_imbalance
;
2515 /* Don't want to pull so many tasks that a group would go idle */
2516 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2518 /* How much load to actually move to equalise the imbalance */
2519 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
2520 (avg_load
- this_load
) * this->__cpu_power
)
2524 * if *imbalance is less than the average load per runnable task
2525 * there is no gaurantee that any tasks will be moved so we'll have
2526 * a think about bumping its value to force at least one task to be
2529 if (*imbalance
< busiest_load_per_task
) {
2530 unsigned long tmp
, pwr_now
, pwr_move
;
2534 pwr_move
= pwr_now
= 0;
2536 if (this_nr_running
) {
2537 this_load_per_task
/= this_nr_running
;
2538 if (busiest_load_per_task
> this_load_per_task
)
2541 this_load_per_task
= SCHED_LOAD_SCALE
;
2543 if (max_load
- this_load
>= busiest_load_per_task
* imbn
) {
2544 *imbalance
= busiest_load_per_task
;
2549 * OK, we don't have enough imbalance to justify moving tasks,
2550 * however we may be able to increase total CPU power used by
2554 pwr_now
+= busiest
->__cpu_power
*
2555 min(busiest_load_per_task
, max_load
);
2556 pwr_now
+= this->__cpu_power
*
2557 min(this_load_per_task
, this_load
);
2558 pwr_now
/= SCHED_LOAD_SCALE
;
2560 /* Amount of load we'd subtract */
2561 tmp
= sg_div_cpu_power(busiest
,
2562 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2564 pwr_move
+= busiest
->__cpu_power
*
2565 min(busiest_load_per_task
, max_load
- tmp
);
2567 /* Amount of load we'd add */
2568 if (max_load
* busiest
->__cpu_power
<
2569 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2570 tmp
= sg_div_cpu_power(this,
2571 max_load
* busiest
->__cpu_power
);
2573 tmp
= sg_div_cpu_power(this,
2574 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2575 pwr_move
+= this->__cpu_power
*
2576 min(this_load_per_task
, this_load
+ tmp
);
2577 pwr_move
/= SCHED_LOAD_SCALE
;
2579 /* Move if we gain throughput */
2580 if (pwr_move
<= pwr_now
)
2583 *imbalance
= busiest_load_per_task
;
2589 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2590 if (idle
== NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2593 if (this == group_leader
&& group_leader
!= group_min
) {
2594 *imbalance
= min_load_per_task
;
2604 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2607 find_busiest_queue(struct sched_group
*group
, enum idle_type idle
,
2608 unsigned long imbalance
, cpumask_t
*cpus
)
2610 struct rq
*busiest
= NULL
, *rq
;
2611 unsigned long max_load
= 0;
2614 for_each_cpu_mask(i
, group
->cpumask
) {
2616 if (!cpu_isset(i
, *cpus
))
2621 if (rq
->nr_running
== 1 && rq
->raw_weighted_load
> imbalance
)
2624 if (rq
->raw_weighted_load
> max_load
) {
2625 max_load
= rq
->raw_weighted_load
;
2634 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2635 * so long as it is large enough.
2637 #define MAX_PINNED_INTERVAL 512
2639 static inline unsigned long minus_1_or_zero(unsigned long n
)
2641 return n
> 0 ? n
- 1 : 0;
2645 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2646 * tasks if there is an imbalance.
2648 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2649 struct sched_domain
*sd
, enum idle_type idle
,
2652 int nr_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2653 struct sched_group
*group
;
2654 unsigned long imbalance
;
2656 cpumask_t cpus
= CPU_MASK_ALL
;
2657 unsigned long flags
;
2660 * When power savings policy is enabled for the parent domain, idle
2661 * sibling can pick up load irrespective of busy siblings. In this case,
2662 * let the state of idle sibling percolate up as IDLE, instead of
2663 * portraying it as NOT_IDLE.
2665 if (idle
!= NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2666 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2669 schedstat_inc(sd
, lb_cnt
[idle
]);
2672 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2679 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2683 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2685 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2689 BUG_ON(busiest
== this_rq
);
2691 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2694 if (busiest
->nr_running
> 1) {
2696 * Attempt to move tasks. If find_busiest_group has found
2697 * an imbalance but busiest->nr_running <= 1, the group is
2698 * still unbalanced. nr_moved simply stays zero, so it is
2699 * correctly treated as an imbalance.
2701 local_irq_save(flags
);
2702 double_rq_lock(this_rq
, busiest
);
2703 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2704 minus_1_or_zero(busiest
->nr_running
),
2705 imbalance
, sd
, idle
, &all_pinned
);
2706 double_rq_unlock(this_rq
, busiest
);
2707 local_irq_restore(flags
);
2710 * some other cpu did the load balance for us.
2712 if (nr_moved
&& this_cpu
!= smp_processor_id())
2713 resched_cpu(this_cpu
);
2715 /* All tasks on this runqueue were pinned by CPU affinity */
2716 if (unlikely(all_pinned
)) {
2717 cpu_clear(cpu_of(busiest
), cpus
);
2718 if (!cpus_empty(cpus
))
2725 schedstat_inc(sd
, lb_failed
[idle
]);
2726 sd
->nr_balance_failed
++;
2728 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2730 spin_lock_irqsave(&busiest
->lock
, flags
);
2732 /* don't kick the migration_thread, if the curr
2733 * task on busiest cpu can't be moved to this_cpu
2735 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2736 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2738 goto out_one_pinned
;
2741 if (!busiest
->active_balance
) {
2742 busiest
->active_balance
= 1;
2743 busiest
->push_cpu
= this_cpu
;
2746 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2748 wake_up_process(busiest
->migration_thread
);
2751 * We've kicked active balancing, reset the failure
2754 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2757 sd
->nr_balance_failed
= 0;
2759 if (likely(!active_balance
)) {
2760 /* We were unbalanced, so reset the balancing interval */
2761 sd
->balance_interval
= sd
->min_interval
;
2764 * If we've begun active balancing, start to back off. This
2765 * case may not be covered by the all_pinned logic if there
2766 * is only 1 task on the busy runqueue (because we don't call
2769 if (sd
->balance_interval
< sd
->max_interval
)
2770 sd
->balance_interval
*= 2;
2773 if (!nr_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2774 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2779 schedstat_inc(sd
, lb_balanced
[idle
]);
2781 sd
->nr_balance_failed
= 0;
2784 /* tune up the balancing interval */
2785 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2786 (sd
->balance_interval
< sd
->max_interval
))
2787 sd
->balance_interval
*= 2;
2789 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2790 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2796 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2797 * tasks if there is an imbalance.
2799 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2800 * this_rq is locked.
2803 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
2805 struct sched_group
*group
;
2806 struct rq
*busiest
= NULL
;
2807 unsigned long imbalance
;
2810 cpumask_t cpus
= CPU_MASK_ALL
;
2813 * When power savings policy is enabled for the parent domain, idle
2814 * sibling can pick up load irrespective of busy siblings. In this case,
2815 * let the state of idle sibling percolate up as IDLE, instead of
2816 * portraying it as NOT_IDLE.
2818 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
2819 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2822 schedstat_inc(sd
, lb_cnt
[NEWLY_IDLE
]);
2824 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, NEWLY_IDLE
,
2825 &sd_idle
, &cpus
, NULL
);
2827 schedstat_inc(sd
, lb_nobusyg
[NEWLY_IDLE
]);
2831 busiest
= find_busiest_queue(group
, NEWLY_IDLE
, imbalance
,
2834 schedstat_inc(sd
, lb_nobusyq
[NEWLY_IDLE
]);
2838 BUG_ON(busiest
== this_rq
);
2840 schedstat_add(sd
, lb_imbalance
[NEWLY_IDLE
], imbalance
);
2843 if (busiest
->nr_running
> 1) {
2844 /* Attempt to move tasks */
2845 double_lock_balance(this_rq
, busiest
);
2846 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2847 minus_1_or_zero(busiest
->nr_running
),
2848 imbalance
, sd
, NEWLY_IDLE
, NULL
);
2849 spin_unlock(&busiest
->lock
);
2852 cpu_clear(cpu_of(busiest
), cpus
);
2853 if (!cpus_empty(cpus
))
2859 schedstat_inc(sd
, lb_failed
[NEWLY_IDLE
]);
2860 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2861 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2864 sd
->nr_balance_failed
= 0;
2869 schedstat_inc(sd
, lb_balanced
[NEWLY_IDLE
]);
2870 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2871 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2873 sd
->nr_balance_failed
= 0;
2879 * idle_balance is called by schedule() if this_cpu is about to become
2880 * idle. Attempts to pull tasks from other CPUs.
2882 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
2884 struct sched_domain
*sd
;
2885 int pulled_task
= 0;
2886 unsigned long next_balance
= jiffies
+ 60 * HZ
;
2888 for_each_domain(this_cpu
, sd
) {
2889 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
2890 /* If we've pulled tasks over stop searching: */
2891 pulled_task
= load_balance_newidle(this_cpu
,
2893 if (time_after(next_balance
,
2894 sd
->last_balance
+ sd
->balance_interval
))
2895 next_balance
= sd
->last_balance
2896 + sd
->balance_interval
;
2903 * We are going idle. next_balance may be set based on
2904 * a busy processor. So reset next_balance.
2906 this_rq
->next_balance
= next_balance
;
2910 * active_load_balance is run by migration threads. It pushes running tasks
2911 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2912 * running on each physical CPU where possible, and avoids physical /
2913 * logical imbalances.
2915 * Called with busiest_rq locked.
2917 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
2919 int target_cpu
= busiest_rq
->push_cpu
;
2920 struct sched_domain
*sd
;
2921 struct rq
*target_rq
;
2923 /* Is there any task to move? */
2924 if (busiest_rq
->nr_running
<= 1)
2927 target_rq
= cpu_rq(target_cpu
);
2930 * This condition is "impossible", if it occurs
2931 * we need to fix it. Originally reported by
2932 * Bjorn Helgaas on a 128-cpu setup.
2934 BUG_ON(busiest_rq
== target_rq
);
2936 /* move a task from busiest_rq to target_rq */
2937 double_lock_balance(busiest_rq
, target_rq
);
2939 /* Search for an sd spanning us and the target CPU. */
2940 for_each_domain(target_cpu
, sd
) {
2941 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2942 cpu_isset(busiest_cpu
, sd
->span
))
2947 schedstat_inc(sd
, alb_cnt
);
2949 if (move_tasks(target_rq
, target_cpu
, busiest_rq
, 1,
2950 RTPRIO_TO_LOAD_WEIGHT(100), sd
, SCHED_IDLE
,
2952 schedstat_inc(sd
, alb_pushed
);
2954 schedstat_inc(sd
, alb_failed
);
2956 spin_unlock(&target_rq
->lock
);
2959 static void update_load(struct rq
*this_rq
)
2961 unsigned long this_load
;
2962 unsigned int i
, scale
;
2964 this_load
= this_rq
->raw_weighted_load
;
2966 /* Update our load: */
2967 for (i
= 0, scale
= 1; i
< 3; i
++, scale
+= scale
) {
2968 unsigned long old_load
, new_load
;
2970 /* scale is effectively 1 << i now, and >> i divides by scale */
2972 old_load
= this_rq
->cpu_load
[i
];
2973 new_load
= this_load
;
2975 * Round up the averaging division if load is increasing. This
2976 * prevents us from getting stuck on 9 if the load is 10, for
2979 if (new_load
> old_load
)
2980 new_load
+= scale
-1;
2981 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2987 atomic_t load_balancer
;
2989 } nohz ____cacheline_aligned
= {
2990 .load_balancer
= ATOMIC_INIT(-1),
2991 .cpu_mask
= CPU_MASK_NONE
,
2995 * This routine will try to nominate the ilb (idle load balancing)
2996 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2997 * load balancing on behalf of all those cpus. If all the cpus in the system
2998 * go into this tickless mode, then there will be no ilb owner (as there is
2999 * no need for one) and all the cpus will sleep till the next wakeup event
3002 * For the ilb owner, tick is not stopped. And this tick will be used
3003 * for idle load balancing. ilb owner will still be part of
3006 * While stopping the tick, this cpu will become the ilb owner if there
3007 * is no other owner. And will be the owner till that cpu becomes busy
3008 * or if all cpus in the system stop their ticks at which point
3009 * there is no need for ilb owner.
3011 * When the ilb owner becomes busy, it nominates another owner, during the
3012 * next busy scheduler_tick()
3014 int select_nohz_load_balancer(int stop_tick
)
3016 int cpu
= smp_processor_id();
3019 cpu_set(cpu
, nohz
.cpu_mask
);
3020 cpu_rq(cpu
)->in_nohz_recently
= 1;
3023 * If we are going offline and still the leader, give up!
3025 if (cpu_is_offline(cpu
) &&
3026 atomic_read(&nohz
.load_balancer
) == cpu
) {
3027 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3032 /* time for ilb owner also to sleep */
3033 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3034 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3035 atomic_set(&nohz
.load_balancer
, -1);
3039 if (atomic_read(&nohz
.load_balancer
) == -1) {
3040 /* make me the ilb owner */
3041 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3043 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3046 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3049 cpu_clear(cpu
, nohz
.cpu_mask
);
3051 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3052 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3059 static DEFINE_SPINLOCK(balancing
);
3062 * It checks each scheduling domain to see if it is due to be balanced,
3063 * and initiates a balancing operation if so.
3065 * Balancing parameters are set up in arch_init_sched_domains.
3067 static inline void rebalance_domains(int cpu
, enum idle_type idle
)
3070 struct rq
*rq
= cpu_rq(cpu
);
3071 unsigned long interval
;
3072 struct sched_domain
*sd
;
3073 /* Earliest time when we have to do rebalance again */
3074 unsigned long next_balance
= jiffies
+ 60*HZ
;
3076 for_each_domain(cpu
, sd
) {
3077 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3080 interval
= sd
->balance_interval
;
3081 if (idle
!= SCHED_IDLE
)
3082 interval
*= sd
->busy_factor
;
3084 /* scale ms to jiffies */
3085 interval
= msecs_to_jiffies(interval
);
3086 if (unlikely(!interval
))
3089 if (sd
->flags
& SD_SERIALIZE
) {
3090 if (!spin_trylock(&balancing
))
3094 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3095 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3097 * We've pulled tasks over so either we're no
3098 * longer idle, or one of our SMT siblings is
3103 sd
->last_balance
= jiffies
;
3105 if (sd
->flags
& SD_SERIALIZE
)
3106 spin_unlock(&balancing
);
3108 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3109 next_balance
= sd
->last_balance
+ interval
;
3112 * Stop the load balance at this level. There is another
3113 * CPU in our sched group which is doing load balancing more
3119 rq
->next_balance
= next_balance
;
3123 * run_rebalance_domains is triggered when needed from the scheduler tick.
3124 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3125 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3127 static void run_rebalance_domains(struct softirq_action
*h
)
3129 int local_cpu
= smp_processor_id();
3130 struct rq
*local_rq
= cpu_rq(local_cpu
);
3131 enum idle_type idle
= local_rq
->idle_at_tick
? SCHED_IDLE
: NOT_IDLE
;
3133 rebalance_domains(local_cpu
, idle
);
3137 * If this cpu is the owner for idle load balancing, then do the
3138 * balancing on behalf of the other idle cpus whose ticks are
3141 if (local_rq
->idle_at_tick
&&
3142 atomic_read(&nohz
.load_balancer
) == local_cpu
) {
3143 cpumask_t cpus
= nohz
.cpu_mask
;
3147 cpu_clear(local_cpu
, cpus
);
3148 for_each_cpu_mask(balance_cpu
, cpus
) {
3150 * If this cpu gets work to do, stop the load balancing
3151 * work being done for other cpus. Next load
3152 * balancing owner will pick it up.
3157 rebalance_domains(balance_cpu
, SCHED_IDLE
);
3159 rq
= cpu_rq(balance_cpu
);
3160 if (time_after(local_rq
->next_balance
, rq
->next_balance
))
3161 local_rq
->next_balance
= rq
->next_balance
;
3168 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3170 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3171 * idle load balancing owner or decide to stop the periodic load balancing,
3172 * if the whole system is idle.
3174 static inline void trigger_load_balance(int cpu
)
3176 struct rq
*rq
= cpu_rq(cpu
);
3179 * If we were in the nohz mode recently and busy at the current
3180 * scheduler tick, then check if we need to nominate new idle
3183 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3184 rq
->in_nohz_recently
= 0;
3186 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3187 cpu_clear(cpu
, nohz
.cpu_mask
);
3188 atomic_set(&nohz
.load_balancer
, -1);
3191 if (atomic_read(&nohz
.load_balancer
) == -1) {
3193 * simple selection for now: Nominate the
3194 * first cpu in the nohz list to be the next
3197 * TBD: Traverse the sched domains and nominate
3198 * the nearest cpu in the nohz.cpu_mask.
3200 int ilb
= first_cpu(nohz
.cpu_mask
);
3208 * If this cpu is idle and doing idle load balancing for all the
3209 * cpus with ticks stopped, is it time for that to stop?
3211 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3212 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3218 * If this cpu is idle and the idle load balancing is done by
3219 * someone else, then no need raise the SCHED_SOFTIRQ
3221 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3222 cpu_isset(cpu
, nohz
.cpu_mask
))
3225 if (time_after_eq(jiffies
, rq
->next_balance
))
3226 raise_softirq(SCHED_SOFTIRQ
);
3230 * on UP we do not need to balance between CPUs:
3232 static inline void idle_balance(int cpu
, struct rq
*rq
)
3237 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3239 EXPORT_PER_CPU_SYMBOL(kstat
);
3242 * This is called on clock ticks and on context switches.
3243 * Bank in p->sched_time the ns elapsed since the last tick or switch.
3246 update_cpu_clock(struct task_struct
*p
, struct rq
*rq
, unsigned long long now
)
3248 p
->sched_time
+= now
- p
->last_ran
;
3249 p
->last_ran
= rq
->most_recent_timestamp
= now
;
3253 * Return current->sched_time plus any more ns on the sched_clock
3254 * that have not yet been banked.
3256 unsigned long long current_sched_time(const struct task_struct
*p
)
3258 unsigned long long ns
;
3259 unsigned long flags
;
3261 local_irq_save(flags
);
3262 ns
= p
->sched_time
+ sched_clock() - p
->last_ran
;
3263 local_irq_restore(flags
);
3269 * We place interactive tasks back into the active array, if possible.
3271 * To guarantee that this does not starve expired tasks we ignore the
3272 * interactivity of a task if the first expired task had to wait more
3273 * than a 'reasonable' amount of time. This deadline timeout is
3274 * load-dependent, as the frequency of array switched decreases with
3275 * increasing number of running tasks. We also ignore the interactivity
3276 * if a better static_prio task has expired:
3278 static inline int expired_starving(struct rq
*rq
)
3280 if (rq
->curr
->static_prio
> rq
->best_expired_prio
)
3282 if (!STARVATION_LIMIT
|| !rq
->expired_timestamp
)
3284 if (jiffies
- rq
->expired_timestamp
> STARVATION_LIMIT
* rq
->nr_running
)
3290 * Account user cpu time to a process.
3291 * @p: the process that the cpu time gets accounted to
3292 * @hardirq_offset: the offset to subtract from hardirq_count()
3293 * @cputime: the cpu time spent in user space since the last update
3295 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3297 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3300 p
->utime
= cputime_add(p
->utime
, cputime
);
3302 /* Add user time to cpustat. */
3303 tmp
= cputime_to_cputime64(cputime
);
3304 if (TASK_NICE(p
) > 0)
3305 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3307 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3311 * Account system cpu time to a process.
3312 * @p: the process that the cpu time gets accounted to
3313 * @hardirq_offset: the offset to subtract from hardirq_count()
3314 * @cputime: the cpu time spent in kernel space since the last update
3316 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3319 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3320 struct rq
*rq
= this_rq();
3323 p
->stime
= cputime_add(p
->stime
, cputime
);
3325 /* Add system time to cpustat. */
3326 tmp
= cputime_to_cputime64(cputime
);
3327 if (hardirq_count() - hardirq_offset
)
3328 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3329 else if (softirq_count())
3330 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3331 else if (p
!= rq
->idle
)
3332 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3333 else if (atomic_read(&rq
->nr_iowait
) > 0)
3334 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3336 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3337 /* Account for system time used */
3338 acct_update_integrals(p
);
3342 * Account for involuntary wait time.
3343 * @p: the process from which the cpu time has been stolen
3344 * @steal: the cpu time spent in involuntary wait
3346 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3348 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3349 cputime64_t tmp
= cputime_to_cputime64(steal
);
3350 struct rq
*rq
= this_rq();
3352 if (p
== rq
->idle
) {
3353 p
->stime
= cputime_add(p
->stime
, steal
);
3354 if (atomic_read(&rq
->nr_iowait
) > 0)
3355 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3357 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3359 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3362 static void task_running_tick(struct rq
*rq
, struct task_struct
*p
)
3364 if (p
->array
!= rq
->active
) {
3365 /* Task has expired but was not scheduled yet */
3366 set_tsk_need_resched(p
);
3369 spin_lock(&rq
->lock
);
3371 * The task was running during this tick - update the
3372 * time slice counter. Note: we do not update a thread's
3373 * priority until it either goes to sleep or uses up its
3374 * timeslice. This makes it possible for interactive tasks
3375 * to use up their timeslices at their highest priority levels.
3379 * RR tasks need a special form of timeslice management.
3380 * FIFO tasks have no timeslices.
3382 if ((p
->policy
== SCHED_RR
) && !--p
->time_slice
) {
3383 p
->time_slice
= task_timeslice(p
);
3384 p
->first_time_slice
= 0;
3385 set_tsk_need_resched(p
);
3387 /* put it at the end of the queue: */
3388 requeue_task(p
, rq
->active
);
3392 if (!--p
->time_slice
) {
3393 dequeue_task(p
, rq
->active
);
3394 set_tsk_need_resched(p
);
3395 p
->prio
= effective_prio(p
);
3396 p
->time_slice
= task_timeslice(p
);
3397 p
->first_time_slice
= 0;
3399 if (!rq
->expired_timestamp
)
3400 rq
->expired_timestamp
= jiffies
;
3401 if (!TASK_INTERACTIVE(p
) || expired_starving(rq
)) {
3402 enqueue_task(p
, rq
->expired
);
3403 if (p
->static_prio
< rq
->best_expired_prio
)
3404 rq
->best_expired_prio
= p
->static_prio
;
3406 enqueue_task(p
, rq
->active
);
3409 * Prevent a too long timeslice allowing a task to monopolize
3410 * the CPU. We do this by splitting up the timeslice into
3413 * Note: this does not mean the task's timeslices expire or
3414 * get lost in any way, they just might be preempted by
3415 * another task of equal priority. (one with higher
3416 * priority would have preempted this task already.) We
3417 * requeue this task to the end of the list on this priority
3418 * level, which is in essence a round-robin of tasks with
3421 * This only applies to tasks in the interactive
3422 * delta range with at least TIMESLICE_GRANULARITY to requeue.
3424 if (TASK_INTERACTIVE(p
) && !((task_timeslice(p
) -
3425 p
->time_slice
) % TIMESLICE_GRANULARITY(p
)) &&
3426 (p
->time_slice
>= TIMESLICE_GRANULARITY(p
)) &&
3427 (p
->array
== rq
->active
)) {
3429 requeue_task(p
, rq
->active
);
3430 set_tsk_need_resched(p
);
3434 spin_unlock(&rq
->lock
);
3438 * This function gets called by the timer code, with HZ frequency.
3439 * We call it with interrupts disabled.
3441 * It also gets called by the fork code, when changing the parent's
3444 void scheduler_tick(void)
3446 unsigned long long now
= sched_clock();
3447 struct task_struct
*p
= current
;
3448 int cpu
= smp_processor_id();
3449 int idle_at_tick
= idle_cpu(cpu
);
3450 struct rq
*rq
= cpu_rq(cpu
);
3452 update_cpu_clock(p
, rq
, now
);
3455 task_running_tick(rq
, p
);
3458 rq
->idle_at_tick
= idle_at_tick
;
3459 trigger_load_balance(cpu
);
3463 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3465 void fastcall
add_preempt_count(int val
)
3470 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3472 preempt_count() += val
;
3474 * Spinlock count overflowing soon?
3476 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3479 EXPORT_SYMBOL(add_preempt_count
);
3481 void fastcall
sub_preempt_count(int val
)
3486 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3489 * Is the spinlock portion underflowing?
3491 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3492 !(preempt_count() & PREEMPT_MASK
)))
3495 preempt_count() -= val
;
3497 EXPORT_SYMBOL(sub_preempt_count
);
3501 static inline int interactive_sleep(enum sleep_type sleep_type
)
3503 return (sleep_type
== SLEEP_INTERACTIVE
||
3504 sleep_type
== SLEEP_INTERRUPTED
);
3508 * schedule() is the main scheduler function.
3510 asmlinkage
void __sched
schedule(void)
3512 struct task_struct
*prev
, *next
;
3513 struct prio_array
*array
;
3514 struct list_head
*queue
;
3515 unsigned long long now
;
3516 unsigned long run_time
;
3517 int cpu
, idx
, new_prio
;
3522 * Test if we are atomic. Since do_exit() needs to call into
3523 * schedule() atomically, we ignore that path for now.
3524 * Otherwise, whine if we are scheduling when we should not be.
3526 if (unlikely(in_atomic() && !current
->exit_state
)) {
3527 printk(KERN_ERR
"BUG: scheduling while atomic: "
3529 current
->comm
, preempt_count(), current
->pid
);
3530 debug_show_held_locks(current
);
3531 if (irqs_disabled())
3532 print_irqtrace_events(current
);
3535 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3540 release_kernel_lock(prev
);
3541 need_resched_nonpreemptible
:
3545 * The idle thread is not allowed to schedule!
3546 * Remove this check after it has been exercised a bit.
3548 if (unlikely(prev
== rq
->idle
) && prev
->state
!= TASK_RUNNING
) {
3549 printk(KERN_ERR
"bad: scheduling from the idle thread!\n");
3553 schedstat_inc(rq
, sched_cnt
);
3554 now
= sched_clock();
3555 if (likely((long long)(now
- prev
->timestamp
) < NS_MAX_SLEEP_AVG
)) {
3556 run_time
= now
- prev
->timestamp
;
3557 if (unlikely((long long)(now
- prev
->timestamp
) < 0))
3560 run_time
= NS_MAX_SLEEP_AVG
;
3563 * Tasks charged proportionately less run_time at high sleep_avg to
3564 * delay them losing their interactive status
3566 run_time
/= (CURRENT_BONUS(prev
) ? : 1);
3568 spin_lock_irq(&rq
->lock
);
3570 switch_count
= &prev
->nivcsw
;
3571 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3572 switch_count
= &prev
->nvcsw
;
3573 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3574 unlikely(signal_pending(prev
))))
3575 prev
->state
= TASK_RUNNING
;
3577 if (prev
->state
== TASK_UNINTERRUPTIBLE
)
3578 rq
->nr_uninterruptible
++;
3579 deactivate_task(prev
, rq
);
3583 cpu
= smp_processor_id();
3584 if (unlikely(!rq
->nr_running
)) {
3585 idle_balance(cpu
, rq
);
3586 if (!rq
->nr_running
) {
3588 rq
->expired_timestamp
= 0;
3594 if (unlikely(!array
->nr_active
)) {
3596 * Switch the active and expired arrays.
3598 schedstat_inc(rq
, sched_switch
);
3599 rq
->active
= rq
->expired
;
3600 rq
->expired
= array
;
3602 rq
->expired_timestamp
= 0;
3603 rq
->best_expired_prio
= MAX_PRIO
;
3606 idx
= sched_find_first_bit(array
->bitmap
);
3607 queue
= array
->queue
+ idx
;
3608 next
= list_entry(queue
->next
, struct task_struct
, run_list
);
3610 if (!rt_task(next
) && interactive_sleep(next
->sleep_type
)) {
3611 unsigned long long delta
= now
- next
->timestamp
;
3612 if (unlikely((long long)(now
- next
->timestamp
) < 0))
3615 if (next
->sleep_type
== SLEEP_INTERACTIVE
)
3616 delta
= delta
* (ON_RUNQUEUE_WEIGHT
* 128 / 100) / 128;
3618 array
= next
->array
;
3619 new_prio
= recalc_task_prio(next
, next
->timestamp
+ delta
);
3621 if (unlikely(next
->prio
!= new_prio
)) {
3622 dequeue_task(next
, array
);
3623 next
->prio
= new_prio
;
3624 enqueue_task(next
, array
);
3627 next
->sleep_type
= SLEEP_NORMAL
;
3629 if (next
== rq
->idle
)
3630 schedstat_inc(rq
, sched_goidle
);
3632 prefetch_stack(next
);
3633 clear_tsk_need_resched(prev
);
3634 rcu_qsctr_inc(task_cpu(prev
));
3636 update_cpu_clock(prev
, rq
, now
);
3638 prev
->sleep_avg
-= run_time
;
3639 if ((long)prev
->sleep_avg
<= 0)
3640 prev
->sleep_avg
= 0;
3641 prev
->timestamp
= prev
->last_ran
= now
;
3643 sched_info_switch(prev
, next
);
3644 if (likely(prev
!= next
)) {
3645 next
->timestamp
= next
->last_ran
= now
;
3650 prepare_task_switch(rq
, next
);
3651 prev
= context_switch(rq
, prev
, next
);
3654 * this_rq must be evaluated again because prev may have moved
3655 * CPUs since it called schedule(), thus the 'rq' on its stack
3656 * frame will be invalid.
3658 finish_task_switch(this_rq(), prev
);
3660 spin_unlock_irq(&rq
->lock
);
3663 if (unlikely(reacquire_kernel_lock(prev
) < 0))
3664 goto need_resched_nonpreemptible
;
3665 preempt_enable_no_resched();
3666 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3669 EXPORT_SYMBOL(schedule
);
3671 #ifdef CONFIG_PREEMPT
3673 * this is the entry point to schedule() from in-kernel preemption
3674 * off of preempt_enable. Kernel preemptions off return from interrupt
3675 * occur there and call schedule directly.
3677 asmlinkage
void __sched
preempt_schedule(void)
3679 struct thread_info
*ti
= current_thread_info();
3680 #ifdef CONFIG_PREEMPT_BKL
3681 struct task_struct
*task
= current
;
3682 int saved_lock_depth
;
3685 * If there is a non-zero preempt_count or interrupts are disabled,
3686 * we do not want to preempt the current task. Just return..
3688 if (likely(ti
->preempt_count
|| irqs_disabled()))
3692 add_preempt_count(PREEMPT_ACTIVE
);
3694 * We keep the big kernel semaphore locked, but we
3695 * clear ->lock_depth so that schedule() doesnt
3696 * auto-release the semaphore:
3698 #ifdef CONFIG_PREEMPT_BKL
3699 saved_lock_depth
= task
->lock_depth
;
3700 task
->lock_depth
= -1;
3703 #ifdef CONFIG_PREEMPT_BKL
3704 task
->lock_depth
= saved_lock_depth
;
3706 sub_preempt_count(PREEMPT_ACTIVE
);
3708 /* we could miss a preemption opportunity between schedule and now */
3710 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3713 EXPORT_SYMBOL(preempt_schedule
);
3716 * this is the entry point to schedule() from kernel preemption
3717 * off of irq context.
3718 * Note, that this is called and return with irqs disabled. This will
3719 * protect us against recursive calling from irq.
3721 asmlinkage
void __sched
preempt_schedule_irq(void)
3723 struct thread_info
*ti
= current_thread_info();
3724 #ifdef CONFIG_PREEMPT_BKL
3725 struct task_struct
*task
= current
;
3726 int saved_lock_depth
;
3728 /* Catch callers which need to be fixed */
3729 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3732 add_preempt_count(PREEMPT_ACTIVE
);
3734 * We keep the big kernel semaphore locked, but we
3735 * clear ->lock_depth so that schedule() doesnt
3736 * auto-release the semaphore:
3738 #ifdef CONFIG_PREEMPT_BKL
3739 saved_lock_depth
= task
->lock_depth
;
3740 task
->lock_depth
= -1;
3744 local_irq_disable();
3745 #ifdef CONFIG_PREEMPT_BKL
3746 task
->lock_depth
= saved_lock_depth
;
3748 sub_preempt_count(PREEMPT_ACTIVE
);
3750 /* we could miss a preemption opportunity between schedule and now */
3752 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3756 #endif /* CONFIG_PREEMPT */
3758 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3761 return try_to_wake_up(curr
->private, mode
, sync
);
3763 EXPORT_SYMBOL(default_wake_function
);
3766 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3767 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3768 * number) then we wake all the non-exclusive tasks and one exclusive task.
3770 * There are circumstances in which we can try to wake a task which has already
3771 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3772 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3774 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3775 int nr_exclusive
, int sync
, void *key
)
3777 struct list_head
*tmp
, *next
;
3779 list_for_each_safe(tmp
, next
, &q
->task_list
) {
3780 wait_queue_t
*curr
= list_entry(tmp
, wait_queue_t
, task_list
);
3781 unsigned flags
= curr
->flags
;
3783 if (curr
->func(curr
, mode
, sync
, key
) &&
3784 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3790 * __wake_up - wake up threads blocked on a waitqueue.
3792 * @mode: which threads
3793 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3794 * @key: is directly passed to the wakeup function
3796 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3797 int nr_exclusive
, void *key
)
3799 unsigned long flags
;
3801 spin_lock_irqsave(&q
->lock
, flags
);
3802 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3803 spin_unlock_irqrestore(&q
->lock
, flags
);
3805 EXPORT_SYMBOL(__wake_up
);
3808 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3810 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3812 __wake_up_common(q
, mode
, 1, 0, NULL
);
3816 * __wake_up_sync - wake up threads blocked on a waitqueue.
3818 * @mode: which threads
3819 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3821 * The sync wakeup differs that the waker knows that it will schedule
3822 * away soon, so while the target thread will be woken up, it will not
3823 * be migrated to another CPU - ie. the two threads are 'synchronized'
3824 * with each other. This can prevent needless bouncing between CPUs.
3826 * On UP it can prevent extra preemption.
3829 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3831 unsigned long flags
;
3837 if (unlikely(!nr_exclusive
))
3840 spin_lock_irqsave(&q
->lock
, flags
);
3841 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3842 spin_unlock_irqrestore(&q
->lock
, flags
);
3844 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3846 void fastcall
complete(struct completion
*x
)
3848 unsigned long flags
;
3850 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3852 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3854 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3856 EXPORT_SYMBOL(complete
);
3858 void fastcall
complete_all(struct completion
*x
)
3860 unsigned long flags
;
3862 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3863 x
->done
+= UINT_MAX
/2;
3864 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3866 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3868 EXPORT_SYMBOL(complete_all
);
3870 void fastcall __sched
wait_for_completion(struct completion
*x
)
3874 spin_lock_irq(&x
->wait
.lock
);
3876 DECLARE_WAITQUEUE(wait
, current
);
3878 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3879 __add_wait_queue_tail(&x
->wait
, &wait
);
3881 __set_current_state(TASK_UNINTERRUPTIBLE
);
3882 spin_unlock_irq(&x
->wait
.lock
);
3884 spin_lock_irq(&x
->wait
.lock
);
3886 __remove_wait_queue(&x
->wait
, &wait
);
3889 spin_unlock_irq(&x
->wait
.lock
);
3891 EXPORT_SYMBOL(wait_for_completion
);
3893 unsigned long fastcall __sched
3894 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3898 spin_lock_irq(&x
->wait
.lock
);
3900 DECLARE_WAITQUEUE(wait
, current
);
3902 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3903 __add_wait_queue_tail(&x
->wait
, &wait
);
3905 __set_current_state(TASK_UNINTERRUPTIBLE
);
3906 spin_unlock_irq(&x
->wait
.lock
);
3907 timeout
= schedule_timeout(timeout
);
3908 spin_lock_irq(&x
->wait
.lock
);
3910 __remove_wait_queue(&x
->wait
, &wait
);
3914 __remove_wait_queue(&x
->wait
, &wait
);
3918 spin_unlock_irq(&x
->wait
.lock
);
3921 EXPORT_SYMBOL(wait_for_completion_timeout
);
3923 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3929 spin_lock_irq(&x
->wait
.lock
);
3931 DECLARE_WAITQUEUE(wait
, current
);
3933 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3934 __add_wait_queue_tail(&x
->wait
, &wait
);
3936 if (signal_pending(current
)) {
3938 __remove_wait_queue(&x
->wait
, &wait
);
3941 __set_current_state(TASK_INTERRUPTIBLE
);
3942 spin_unlock_irq(&x
->wait
.lock
);
3944 spin_lock_irq(&x
->wait
.lock
);
3946 __remove_wait_queue(&x
->wait
, &wait
);
3950 spin_unlock_irq(&x
->wait
.lock
);
3954 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3956 unsigned long fastcall __sched
3957 wait_for_completion_interruptible_timeout(struct completion
*x
,
3958 unsigned long timeout
)
3962 spin_lock_irq(&x
->wait
.lock
);
3964 DECLARE_WAITQUEUE(wait
, current
);
3966 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3967 __add_wait_queue_tail(&x
->wait
, &wait
);
3969 if (signal_pending(current
)) {
3970 timeout
= -ERESTARTSYS
;
3971 __remove_wait_queue(&x
->wait
, &wait
);
3974 __set_current_state(TASK_INTERRUPTIBLE
);
3975 spin_unlock_irq(&x
->wait
.lock
);
3976 timeout
= schedule_timeout(timeout
);
3977 spin_lock_irq(&x
->wait
.lock
);
3979 __remove_wait_queue(&x
->wait
, &wait
);
3983 __remove_wait_queue(&x
->wait
, &wait
);
3987 spin_unlock_irq(&x
->wait
.lock
);
3990 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3993 #define SLEEP_ON_VAR \
3994 unsigned long flags; \
3995 wait_queue_t wait; \
3996 init_waitqueue_entry(&wait, current);
3998 #define SLEEP_ON_HEAD \
3999 spin_lock_irqsave(&q->lock,flags); \
4000 __add_wait_queue(q, &wait); \
4001 spin_unlock(&q->lock);
4003 #define SLEEP_ON_TAIL \
4004 spin_lock_irq(&q->lock); \
4005 __remove_wait_queue(q, &wait); \
4006 spin_unlock_irqrestore(&q->lock, flags);
4008 void fastcall __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4012 current
->state
= TASK_INTERRUPTIBLE
;
4018 EXPORT_SYMBOL(interruptible_sleep_on
);
4020 long fastcall __sched
4021 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4025 current
->state
= TASK_INTERRUPTIBLE
;
4028 timeout
= schedule_timeout(timeout
);
4033 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4035 void fastcall __sched
sleep_on(wait_queue_head_t
*q
)
4039 current
->state
= TASK_UNINTERRUPTIBLE
;
4045 EXPORT_SYMBOL(sleep_on
);
4047 long fastcall __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4051 current
->state
= TASK_UNINTERRUPTIBLE
;
4054 timeout
= schedule_timeout(timeout
);
4060 EXPORT_SYMBOL(sleep_on_timeout
);
4062 #ifdef CONFIG_RT_MUTEXES
4065 * rt_mutex_setprio - set the current priority of a task
4067 * @prio: prio value (kernel-internal form)
4069 * This function changes the 'effective' priority of a task. It does
4070 * not touch ->normal_prio like __setscheduler().
4072 * Used by the rt_mutex code to implement priority inheritance logic.
4074 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4076 struct prio_array
*array
;
4077 unsigned long flags
;
4081 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4083 rq
= task_rq_lock(p
, &flags
);
4088 dequeue_task(p
, array
);
4093 * If changing to an RT priority then queue it
4094 * in the active array!
4098 enqueue_task(p
, array
);
4100 * Reschedule if we are currently running on this runqueue and
4101 * our priority decreased, or if we are not currently running on
4102 * this runqueue and our priority is higher than the current's
4104 if (task_running(rq
, p
)) {
4105 if (p
->prio
> oldprio
)
4106 resched_task(rq
->curr
);
4107 } else if (TASK_PREEMPTS_CURR(p
, rq
))
4108 resched_task(rq
->curr
);
4110 task_rq_unlock(rq
, &flags
);
4115 void set_user_nice(struct task_struct
*p
, long nice
)
4117 struct prio_array
*array
;
4118 int old_prio
, delta
;
4119 unsigned long flags
;
4122 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4125 * We have to be careful, if called from sys_setpriority(),
4126 * the task might be in the middle of scheduling on another CPU.
4128 rq
= task_rq_lock(p
, &flags
);
4130 * The RT priorities are set via sched_setscheduler(), but we still
4131 * allow the 'normal' nice value to be set - but as expected
4132 * it wont have any effect on scheduling until the task is
4133 * not SCHED_NORMAL/SCHED_BATCH:
4135 if (has_rt_policy(p
)) {
4136 p
->static_prio
= NICE_TO_PRIO(nice
);
4141 dequeue_task(p
, array
);
4142 dec_raw_weighted_load(rq
, p
);
4145 p
->static_prio
= NICE_TO_PRIO(nice
);
4148 p
->prio
= effective_prio(p
);
4149 delta
= p
->prio
- old_prio
;
4152 enqueue_task(p
, array
);
4153 inc_raw_weighted_load(rq
, p
);
4155 * If the task increased its priority or is running and
4156 * lowered its priority, then reschedule its CPU:
4158 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4159 resched_task(rq
->curr
);
4162 task_rq_unlock(rq
, &flags
);
4164 EXPORT_SYMBOL(set_user_nice
);
4167 * can_nice - check if a task can reduce its nice value
4171 int can_nice(const struct task_struct
*p
, const int nice
)
4173 /* convert nice value [19,-20] to rlimit style value [1,40] */
4174 int nice_rlim
= 20 - nice
;
4176 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4177 capable(CAP_SYS_NICE
));
4180 #ifdef __ARCH_WANT_SYS_NICE
4183 * sys_nice - change the priority of the current process.
4184 * @increment: priority increment
4186 * sys_setpriority is a more generic, but much slower function that
4187 * does similar things.
4189 asmlinkage
long sys_nice(int increment
)
4194 * Setpriority might change our priority at the same moment.
4195 * We don't have to worry. Conceptually one call occurs first
4196 * and we have a single winner.
4198 if (increment
< -40)
4203 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4209 if (increment
< 0 && !can_nice(current
, nice
))
4212 retval
= security_task_setnice(current
, nice
);
4216 set_user_nice(current
, nice
);
4223 * task_prio - return the priority value of a given task.
4224 * @p: the task in question.
4226 * This is the priority value as seen by users in /proc.
4227 * RT tasks are offset by -200. Normal tasks are centered
4228 * around 0, value goes from -16 to +15.
4230 int task_prio(const struct task_struct
*p
)
4232 return p
->prio
- MAX_RT_PRIO
;
4236 * task_nice - return the nice value of a given task.
4237 * @p: the task in question.
4239 int task_nice(const struct task_struct
*p
)
4241 return TASK_NICE(p
);
4243 EXPORT_SYMBOL_GPL(task_nice
);
4246 * idle_cpu - is a given cpu idle currently?
4247 * @cpu: the processor in question.
4249 int idle_cpu(int cpu
)
4251 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4255 * idle_task - return the idle task for a given cpu.
4256 * @cpu: the processor in question.
4258 struct task_struct
*idle_task(int cpu
)
4260 return cpu_rq(cpu
)->idle
;
4264 * find_process_by_pid - find a process with a matching PID value.
4265 * @pid: the pid in question.
4267 static inline struct task_struct
*find_process_by_pid(pid_t pid
)
4269 return pid
? find_task_by_pid(pid
) : current
;
4272 /* Actually do priority change: must hold rq lock. */
4273 static void __setscheduler(struct task_struct
*p
, int policy
, int prio
)
4278 p
->rt_priority
= prio
;
4279 p
->normal_prio
= normal_prio(p
);
4280 /* we are holding p->pi_lock already */
4281 p
->prio
= rt_mutex_getprio(p
);
4283 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
4285 if (policy
== SCHED_BATCH
)
4291 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4292 * @p: the task in question.
4293 * @policy: new policy.
4294 * @param: structure containing the new RT priority.
4296 * NOTE that the task may be already dead.
4298 int sched_setscheduler(struct task_struct
*p
, int policy
,
4299 struct sched_param
*param
)
4301 int retval
, oldprio
, oldpolicy
= -1;
4302 struct prio_array
*array
;
4303 unsigned long flags
;
4306 /* may grab non-irq protected spin_locks */
4307 BUG_ON(in_interrupt());
4309 /* double check policy once rq lock held */
4311 policy
= oldpolicy
= p
->policy
;
4312 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4313 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
)
4316 * Valid priorities for SCHED_FIFO and SCHED_RR are
4317 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
4320 if (param
->sched_priority
< 0 ||
4321 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4322 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4324 if (is_rt_policy(policy
) != (param
->sched_priority
!= 0))
4328 * Allow unprivileged RT tasks to decrease priority:
4330 if (!capable(CAP_SYS_NICE
)) {
4331 if (is_rt_policy(policy
)) {
4332 unsigned long rlim_rtprio
;
4333 unsigned long flags
;
4335 if (!lock_task_sighand(p
, &flags
))
4337 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4338 unlock_task_sighand(p
, &flags
);
4340 /* can't set/change the rt policy */
4341 if (policy
!= p
->policy
&& !rlim_rtprio
)
4344 /* can't increase priority */
4345 if (param
->sched_priority
> p
->rt_priority
&&
4346 param
->sched_priority
> rlim_rtprio
)
4350 /* can't change other user's priorities */
4351 if ((current
->euid
!= p
->euid
) &&
4352 (current
->euid
!= p
->uid
))
4356 retval
= security_task_setscheduler(p
, policy
, param
);
4360 * make sure no PI-waiters arrive (or leave) while we are
4361 * changing the priority of the task:
4363 spin_lock_irqsave(&p
->pi_lock
, flags
);
4365 * To be able to change p->policy safely, the apropriate
4366 * runqueue lock must be held.
4368 rq
= __task_rq_lock(p
);
4369 /* recheck policy now with rq lock held */
4370 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4371 policy
= oldpolicy
= -1;
4372 __task_rq_unlock(rq
);
4373 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4378 deactivate_task(p
, rq
);
4380 __setscheduler(p
, policy
, param
->sched_priority
);
4382 __activate_task(p
, rq
);
4384 * Reschedule if we are currently running on this runqueue and
4385 * our priority decreased, or if we are not currently running on
4386 * this runqueue and our priority is higher than the current's
4388 if (task_running(rq
, p
)) {
4389 if (p
->prio
> oldprio
)
4390 resched_task(rq
->curr
);
4391 } else if (TASK_PREEMPTS_CURR(p
, rq
))
4392 resched_task(rq
->curr
);
4394 __task_rq_unlock(rq
);
4395 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4397 rt_mutex_adjust_pi(p
);
4401 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4404 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4406 struct sched_param lparam
;
4407 struct task_struct
*p
;
4410 if (!param
|| pid
< 0)
4412 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4417 p
= find_process_by_pid(pid
);
4419 retval
= sched_setscheduler(p
, policy
, &lparam
);
4426 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4427 * @pid: the pid in question.
4428 * @policy: new policy.
4429 * @param: structure containing the new RT priority.
4431 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
4432 struct sched_param __user
*param
)
4434 /* negative values for policy are not valid */
4438 return do_sched_setscheduler(pid
, policy
, param
);
4442 * sys_sched_setparam - set/change the RT priority of a thread
4443 * @pid: the pid in question.
4444 * @param: structure containing the new RT priority.
4446 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4448 return do_sched_setscheduler(pid
, -1, param
);
4452 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4453 * @pid: the pid in question.
4455 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4457 struct task_struct
*p
;
4458 int retval
= -EINVAL
;
4464 read_lock(&tasklist_lock
);
4465 p
= find_process_by_pid(pid
);
4467 retval
= security_task_getscheduler(p
);
4471 read_unlock(&tasklist_lock
);
4478 * sys_sched_getscheduler - get the RT priority of a thread
4479 * @pid: the pid in question.
4480 * @param: structure containing the RT priority.
4482 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4484 struct sched_param lp
;
4485 struct task_struct
*p
;
4486 int retval
= -EINVAL
;
4488 if (!param
|| pid
< 0)
4491 read_lock(&tasklist_lock
);
4492 p
= find_process_by_pid(pid
);
4497 retval
= security_task_getscheduler(p
);
4501 lp
.sched_priority
= p
->rt_priority
;
4502 read_unlock(&tasklist_lock
);
4505 * This one might sleep, we cannot do it with a spinlock held ...
4507 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4513 read_unlock(&tasklist_lock
);
4517 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4519 cpumask_t cpus_allowed
;
4520 struct task_struct
*p
;
4524 read_lock(&tasklist_lock
);
4526 p
= find_process_by_pid(pid
);
4528 read_unlock(&tasklist_lock
);
4529 unlock_cpu_hotplug();
4534 * It is not safe to call set_cpus_allowed with the
4535 * tasklist_lock held. We will bump the task_struct's
4536 * usage count and then drop tasklist_lock.
4539 read_unlock(&tasklist_lock
);
4542 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4543 !capable(CAP_SYS_NICE
))
4546 retval
= security_task_setscheduler(p
, 0, NULL
);
4550 cpus_allowed
= cpuset_cpus_allowed(p
);
4551 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4552 retval
= set_cpus_allowed(p
, new_mask
);
4556 unlock_cpu_hotplug();
4560 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4561 cpumask_t
*new_mask
)
4563 if (len
< sizeof(cpumask_t
)) {
4564 memset(new_mask
, 0, sizeof(cpumask_t
));
4565 } else if (len
> sizeof(cpumask_t
)) {
4566 len
= sizeof(cpumask_t
);
4568 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4572 * sys_sched_setaffinity - set the cpu affinity of a process
4573 * @pid: pid of the process
4574 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4575 * @user_mask_ptr: user-space pointer to the new cpu mask
4577 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4578 unsigned long __user
*user_mask_ptr
)
4583 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4587 return sched_setaffinity(pid
, new_mask
);
4591 * Represents all cpu's present in the system
4592 * In systems capable of hotplug, this map could dynamically grow
4593 * as new cpu's are detected in the system via any platform specific
4594 * method, such as ACPI for e.g.
4597 cpumask_t cpu_present_map __read_mostly
;
4598 EXPORT_SYMBOL(cpu_present_map
);
4601 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4602 EXPORT_SYMBOL(cpu_online_map
);
4604 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4605 EXPORT_SYMBOL(cpu_possible_map
);
4608 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4610 struct task_struct
*p
;
4614 read_lock(&tasklist_lock
);
4617 p
= find_process_by_pid(pid
);
4621 retval
= security_task_getscheduler(p
);
4625 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4628 read_unlock(&tasklist_lock
);
4629 unlock_cpu_hotplug();
4637 * sys_sched_getaffinity - get the cpu affinity of a process
4638 * @pid: pid of the process
4639 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4640 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4642 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4643 unsigned long __user
*user_mask_ptr
)
4648 if (len
< sizeof(cpumask_t
))
4651 ret
= sched_getaffinity(pid
, &mask
);
4655 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4658 return sizeof(cpumask_t
);
4662 * sys_sched_yield - yield the current processor to other threads.
4664 * This function yields the current CPU by moving the calling thread
4665 * to the expired array. If there are no other threads running on this
4666 * CPU then this function will return.
4668 asmlinkage
long sys_sched_yield(void)
4670 struct rq
*rq
= this_rq_lock();
4671 struct prio_array
*array
= current
->array
, *target
= rq
->expired
;
4673 schedstat_inc(rq
, yld_cnt
);
4675 * We implement yielding by moving the task into the expired
4678 * (special rule: RT tasks will just roundrobin in the active
4681 if (rt_task(current
))
4682 target
= rq
->active
;
4684 if (array
->nr_active
== 1) {
4685 schedstat_inc(rq
, yld_act_empty
);
4686 if (!rq
->expired
->nr_active
)
4687 schedstat_inc(rq
, yld_both_empty
);
4688 } else if (!rq
->expired
->nr_active
)
4689 schedstat_inc(rq
, yld_exp_empty
);
4691 if (array
!= target
) {
4692 dequeue_task(current
, array
);
4693 enqueue_task(current
, target
);
4696 * requeue_task is cheaper so perform that if possible.
4698 requeue_task(current
, array
);
4701 * Since we are going to call schedule() anyway, there's
4702 * no need to preempt or enable interrupts:
4704 __release(rq
->lock
);
4705 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4706 _raw_spin_unlock(&rq
->lock
);
4707 preempt_enable_no_resched();
4714 static void __cond_resched(void)
4716 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4717 __might_sleep(__FILE__
, __LINE__
);
4720 * The BKS might be reacquired before we have dropped
4721 * PREEMPT_ACTIVE, which could trigger a second
4722 * cond_resched() call.
4725 add_preempt_count(PREEMPT_ACTIVE
);
4727 sub_preempt_count(PREEMPT_ACTIVE
);
4728 } while (need_resched());
4731 int __sched
cond_resched(void)
4733 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4734 system_state
== SYSTEM_RUNNING
) {
4740 EXPORT_SYMBOL(cond_resched
);
4743 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4744 * call schedule, and on return reacquire the lock.
4746 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4747 * operations here to prevent schedule() from being called twice (once via
4748 * spin_unlock(), once by hand).
4750 int cond_resched_lock(spinlock_t
*lock
)
4754 if (need_lockbreak(lock
)) {
4760 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4761 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4762 _raw_spin_unlock(lock
);
4763 preempt_enable_no_resched();
4770 EXPORT_SYMBOL(cond_resched_lock
);
4772 int __sched
cond_resched_softirq(void)
4774 BUG_ON(!in_softirq());
4776 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4777 raw_local_irq_disable();
4779 raw_local_irq_enable();
4786 EXPORT_SYMBOL(cond_resched_softirq
);
4789 * yield - yield the current processor to other threads.
4791 * This is a shortcut for kernel-space yielding - it marks the
4792 * thread runnable and calls sys_sched_yield().
4794 void __sched
yield(void)
4796 set_current_state(TASK_RUNNING
);
4799 EXPORT_SYMBOL(yield
);
4802 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4803 * that process accounting knows that this is a task in IO wait state.
4805 * But don't do that if it is a deliberate, throttling IO wait (this task
4806 * has set its backing_dev_info: the queue against which it should throttle)
4808 void __sched
io_schedule(void)
4810 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4812 delayacct_blkio_start();
4813 atomic_inc(&rq
->nr_iowait
);
4815 atomic_dec(&rq
->nr_iowait
);
4816 delayacct_blkio_end();
4818 EXPORT_SYMBOL(io_schedule
);
4820 long __sched
io_schedule_timeout(long timeout
)
4822 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4825 delayacct_blkio_start();
4826 atomic_inc(&rq
->nr_iowait
);
4827 ret
= schedule_timeout(timeout
);
4828 atomic_dec(&rq
->nr_iowait
);
4829 delayacct_blkio_end();
4834 * sys_sched_get_priority_max - return maximum RT priority.
4835 * @policy: scheduling class.
4837 * this syscall returns the maximum rt_priority that can be used
4838 * by a given scheduling class.
4840 asmlinkage
long sys_sched_get_priority_max(int policy
)
4847 ret
= MAX_USER_RT_PRIO
-1;
4858 * sys_sched_get_priority_min - return minimum RT priority.
4859 * @policy: scheduling class.
4861 * this syscall returns the minimum rt_priority that can be used
4862 * by a given scheduling class.
4864 asmlinkage
long sys_sched_get_priority_min(int policy
)
4881 * sys_sched_rr_get_interval - return the default timeslice of a process.
4882 * @pid: pid of the process.
4883 * @interval: userspace pointer to the timeslice value.
4885 * this syscall writes the default timeslice value of a given process
4886 * into the user-space timespec buffer. A value of '0' means infinity.
4889 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4891 struct task_struct
*p
;
4892 int retval
= -EINVAL
;
4899 read_lock(&tasklist_lock
);
4900 p
= find_process_by_pid(pid
);
4904 retval
= security_task_getscheduler(p
);
4908 jiffies_to_timespec(p
->policy
== SCHED_FIFO
?
4909 0 : task_timeslice(p
), &t
);
4910 read_unlock(&tasklist_lock
);
4911 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4915 read_unlock(&tasklist_lock
);
4919 static const char stat_nam
[] = "RSDTtZX";
4921 static void show_task(struct task_struct
*p
)
4923 unsigned long free
= 0;
4926 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4927 printk("%-13.13s %c", p
->comm
,
4928 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4929 #if (BITS_PER_LONG == 32)
4930 if (state
== TASK_RUNNING
)
4931 printk(" running ");
4933 printk(" %08lX ", thread_saved_pc(p
));
4935 if (state
== TASK_RUNNING
)
4936 printk(" running task ");
4938 printk(" %016lx ", thread_saved_pc(p
));
4940 #ifdef CONFIG_DEBUG_STACK_USAGE
4942 unsigned long *n
= end_of_stack(p
);
4945 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4948 printk("%5lu %5d %6d", free
, p
->pid
, p
->parent
->pid
);
4950 printk(" (L-TLB)\n");
4952 printk(" (NOTLB)\n");
4954 if (state
!= TASK_RUNNING
)
4955 show_stack(p
, NULL
);
4958 void show_state_filter(unsigned long state_filter
)
4960 struct task_struct
*g
, *p
;
4962 #if (BITS_PER_LONG == 32)
4965 printk(" task PC stack pid father child younger older\n");
4969 printk(" task PC stack pid father child younger older\n");
4971 read_lock(&tasklist_lock
);
4972 do_each_thread(g
, p
) {
4974 * reset the NMI-timeout, listing all files on a slow
4975 * console might take alot of time:
4977 touch_nmi_watchdog();
4978 if (!state_filter
|| (p
->state
& state_filter
))
4980 } while_each_thread(g
, p
);
4982 touch_all_softlockup_watchdogs();
4984 read_unlock(&tasklist_lock
);
4986 * Only show locks if all tasks are dumped:
4988 if (state_filter
== -1)
4989 debug_show_all_locks();
4993 * init_idle - set up an idle thread for a given CPU
4994 * @idle: task in question
4995 * @cpu: cpu the idle task belongs to
4997 * NOTE: this function does not set the idle thread's NEED_RESCHED
4998 * flag, to make booting more robust.
5000 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5002 struct rq
*rq
= cpu_rq(cpu
);
5003 unsigned long flags
;
5005 idle
->timestamp
= sched_clock();
5006 idle
->sleep_avg
= 0;
5008 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
5009 idle
->state
= TASK_RUNNING
;
5010 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
5011 set_task_cpu(idle
, cpu
);
5013 spin_lock_irqsave(&rq
->lock
, flags
);
5014 rq
->curr
= rq
->idle
= idle
;
5015 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5018 spin_unlock_irqrestore(&rq
->lock
, flags
);
5020 /* Set the preempt count _outside_ the spinlocks! */
5021 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
5022 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5024 task_thread_info(idle
)->preempt_count
= 0;
5029 * In a system that switches off the HZ timer nohz_cpu_mask
5030 * indicates which cpus entered this state. This is used
5031 * in the rcu update to wait only for active cpus. For system
5032 * which do not switch off the HZ timer nohz_cpu_mask should
5033 * always be CPU_MASK_NONE.
5035 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
5039 * This is how migration works:
5041 * 1) we queue a struct migration_req structure in the source CPU's
5042 * runqueue and wake up that CPU's migration thread.
5043 * 2) we down() the locked semaphore => thread blocks.
5044 * 3) migration thread wakes up (implicitly it forces the migrated
5045 * thread off the CPU)
5046 * 4) it gets the migration request and checks whether the migrated
5047 * task is still in the wrong runqueue.
5048 * 5) if it's in the wrong runqueue then the migration thread removes
5049 * it and puts it into the right queue.
5050 * 6) migration thread up()s the semaphore.
5051 * 7) we wake up and the migration is done.
5055 * Change a given task's CPU affinity. Migrate the thread to a
5056 * proper CPU and schedule it away if the CPU it's executing on
5057 * is removed from the allowed bitmask.
5059 * NOTE: the caller must have a valid reference to the task, the
5060 * task must not exit() & deallocate itself prematurely. The
5061 * call is not atomic; no spinlocks may be held.
5063 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
5065 struct migration_req req
;
5066 unsigned long flags
;
5070 rq
= task_rq_lock(p
, &flags
);
5071 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
5076 p
->cpus_allowed
= new_mask
;
5077 /* Can the task run on the task's current CPU? If so, we're done */
5078 if (cpu_isset(task_cpu(p
), new_mask
))
5081 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
5082 /* Need help from migration thread: drop lock and wait. */
5083 task_rq_unlock(rq
, &flags
);
5084 wake_up_process(rq
->migration_thread
);
5085 wait_for_completion(&req
.done
);
5086 tlb_migrate_finish(p
->mm
);
5090 task_rq_unlock(rq
, &flags
);
5094 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
5097 * Move (not current) task off this cpu, onto dest cpu. We're doing
5098 * this because either it can't run here any more (set_cpus_allowed()
5099 * away from this CPU, or CPU going down), or because we're
5100 * attempting to rebalance this task on exec (sched_exec).
5102 * So we race with normal scheduler movements, but that's OK, as long
5103 * as the task is no longer on this CPU.
5105 * Returns non-zero if task was successfully migrated.
5107 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5109 struct rq
*rq_dest
, *rq_src
;
5112 if (unlikely(cpu_is_offline(dest_cpu
)))
5115 rq_src
= cpu_rq(src_cpu
);
5116 rq_dest
= cpu_rq(dest_cpu
);
5118 double_rq_lock(rq_src
, rq_dest
);
5119 /* Already moved. */
5120 if (task_cpu(p
) != src_cpu
)
5122 /* Affinity changed (again). */
5123 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
5126 set_task_cpu(p
, dest_cpu
);
5129 * Sync timestamp with rq_dest's before activating.
5130 * The same thing could be achieved by doing this step
5131 * afterwards, and pretending it was a local activate.
5132 * This way is cleaner and logically correct.
5134 p
->timestamp
= p
->timestamp
- rq_src
->most_recent_timestamp
5135 + rq_dest
->most_recent_timestamp
;
5136 deactivate_task(p
, rq_src
);
5137 __activate_task(p
, rq_dest
);
5138 if (TASK_PREEMPTS_CURR(p
, rq_dest
))
5139 resched_task(rq_dest
->curr
);
5143 double_rq_unlock(rq_src
, rq_dest
);
5148 * migration_thread - this is a highprio system thread that performs
5149 * thread migration by bumping thread off CPU then 'pushing' onto
5152 static int migration_thread(void *data
)
5154 int cpu
= (long)data
;
5158 BUG_ON(rq
->migration_thread
!= current
);
5160 set_current_state(TASK_INTERRUPTIBLE
);
5161 while (!kthread_should_stop()) {
5162 struct migration_req
*req
;
5163 struct list_head
*head
;
5167 spin_lock_irq(&rq
->lock
);
5169 if (cpu_is_offline(cpu
)) {
5170 spin_unlock_irq(&rq
->lock
);
5174 if (rq
->active_balance
) {
5175 active_load_balance(rq
, cpu
);
5176 rq
->active_balance
= 0;
5179 head
= &rq
->migration_queue
;
5181 if (list_empty(head
)) {
5182 spin_unlock_irq(&rq
->lock
);
5184 set_current_state(TASK_INTERRUPTIBLE
);
5187 req
= list_entry(head
->next
, struct migration_req
, list
);
5188 list_del_init(head
->next
);
5190 spin_unlock(&rq
->lock
);
5191 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5194 complete(&req
->done
);
5196 __set_current_state(TASK_RUNNING
);
5200 /* Wait for kthread_stop */
5201 set_current_state(TASK_INTERRUPTIBLE
);
5202 while (!kthread_should_stop()) {
5204 set_current_state(TASK_INTERRUPTIBLE
);
5206 __set_current_state(TASK_RUNNING
);
5210 #ifdef CONFIG_HOTPLUG_CPU
5212 * Figure out where task on dead CPU should go, use force if neccessary.
5213 * NOTE: interrupts should be disabled by the caller
5215 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5217 unsigned long flags
;
5224 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5225 cpus_and(mask
, mask
, p
->cpus_allowed
);
5226 dest_cpu
= any_online_cpu(mask
);
5228 /* On any allowed CPU? */
5229 if (dest_cpu
== NR_CPUS
)
5230 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5232 /* No more Mr. Nice Guy. */
5233 if (dest_cpu
== NR_CPUS
) {
5234 rq
= task_rq_lock(p
, &flags
);
5235 cpus_setall(p
->cpus_allowed
);
5236 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5237 task_rq_unlock(rq
, &flags
);
5240 * Don't tell them about moving exiting tasks or
5241 * kernel threads (both mm NULL), since they never
5244 if (p
->mm
&& printk_ratelimit())
5245 printk(KERN_INFO
"process %d (%s) no "
5246 "longer affine to cpu%d\n",
5247 p
->pid
, p
->comm
, dead_cpu
);
5249 if (!__migrate_task(p
, dead_cpu
, dest_cpu
))
5254 * While a dead CPU has no uninterruptible tasks queued at this point,
5255 * it might still have a nonzero ->nr_uninterruptible counter, because
5256 * for performance reasons the counter is not stricly tracking tasks to
5257 * their home CPUs. So we just add the counter to another CPU's counter,
5258 * to keep the global sum constant after CPU-down:
5260 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5262 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5263 unsigned long flags
;
5265 local_irq_save(flags
);
5266 double_rq_lock(rq_src
, rq_dest
);
5267 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5268 rq_src
->nr_uninterruptible
= 0;
5269 double_rq_unlock(rq_src
, rq_dest
);
5270 local_irq_restore(flags
);
5273 /* Run through task list and migrate tasks from the dead cpu. */
5274 static void migrate_live_tasks(int src_cpu
)
5276 struct task_struct
*p
, *t
;
5278 write_lock_irq(&tasklist_lock
);
5280 do_each_thread(t
, p
) {
5284 if (task_cpu(p
) == src_cpu
)
5285 move_task_off_dead_cpu(src_cpu
, p
);
5286 } while_each_thread(t
, p
);
5288 write_unlock_irq(&tasklist_lock
);
5291 /* Schedules idle task to be the next runnable task on current CPU.
5292 * It does so by boosting its priority to highest possible and adding it to
5293 * the _front_ of the runqueue. Used by CPU offline code.
5295 void sched_idle_next(void)
5297 int this_cpu
= smp_processor_id();
5298 struct rq
*rq
= cpu_rq(this_cpu
);
5299 struct task_struct
*p
= rq
->idle
;
5300 unsigned long flags
;
5302 /* cpu has to be offline */
5303 BUG_ON(cpu_online(this_cpu
));
5306 * Strictly not necessary since rest of the CPUs are stopped by now
5307 * and interrupts disabled on the current cpu.
5309 spin_lock_irqsave(&rq
->lock
, flags
);
5311 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5313 /* Add idle task to the _front_ of its priority queue: */
5314 __activate_idle_task(p
, rq
);
5316 spin_unlock_irqrestore(&rq
->lock
, flags
);
5320 * Ensures that the idle task is using init_mm right before its cpu goes
5323 void idle_task_exit(void)
5325 struct mm_struct
*mm
= current
->active_mm
;
5327 BUG_ON(cpu_online(smp_processor_id()));
5330 switch_mm(mm
, &init_mm
, current
);
5334 /* called under rq->lock with disabled interrupts */
5335 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5337 struct rq
*rq
= cpu_rq(dead_cpu
);
5339 /* Must be exiting, otherwise would be on tasklist. */
5340 BUG_ON(p
->exit_state
!= EXIT_ZOMBIE
&& p
->exit_state
!= EXIT_DEAD
);
5342 /* Cannot have done final schedule yet: would have vanished. */
5343 BUG_ON(p
->state
== TASK_DEAD
);
5348 * Drop lock around migration; if someone else moves it,
5349 * that's OK. No task can be added to this CPU, so iteration is
5351 * NOTE: interrupts should be left disabled --dev@
5353 spin_unlock(&rq
->lock
);
5354 move_task_off_dead_cpu(dead_cpu
, p
);
5355 spin_lock(&rq
->lock
);
5360 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5361 static void migrate_dead_tasks(unsigned int dead_cpu
)
5363 struct rq
*rq
= cpu_rq(dead_cpu
);
5364 unsigned int arr
, i
;
5366 for (arr
= 0; arr
< 2; arr
++) {
5367 for (i
= 0; i
< MAX_PRIO
; i
++) {
5368 struct list_head
*list
= &rq
->arrays
[arr
].queue
[i
];
5370 while (!list_empty(list
))
5371 migrate_dead(dead_cpu
, list_entry(list
->next
,
5372 struct task_struct
, run_list
));
5376 #endif /* CONFIG_HOTPLUG_CPU */
5379 * migration_call - callback that gets triggered when a CPU is added.
5380 * Here we can start up the necessary migration thread for the new CPU.
5382 static int __cpuinit
5383 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5385 struct task_struct
*p
;
5386 int cpu
= (long)hcpu
;
5387 unsigned long flags
;
5391 case CPU_UP_PREPARE
:
5392 p
= kthread_create(migration_thread
, hcpu
, "migration/%d",cpu
);
5395 p
->flags
|= PF_NOFREEZE
;
5396 kthread_bind(p
, cpu
);
5397 /* Must be high prio: stop_machine expects to yield to it. */
5398 rq
= task_rq_lock(p
, &flags
);
5399 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5400 task_rq_unlock(rq
, &flags
);
5401 cpu_rq(cpu
)->migration_thread
= p
;
5405 /* Strictly unneccessary, as first user will wake it. */
5406 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5409 #ifdef CONFIG_HOTPLUG_CPU
5410 case CPU_UP_CANCELED
:
5411 if (!cpu_rq(cpu
)->migration_thread
)
5413 /* Unbind it from offline cpu so it can run. Fall thru. */
5414 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5415 any_online_cpu(cpu_online_map
));
5416 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5417 cpu_rq(cpu
)->migration_thread
= NULL
;
5421 migrate_live_tasks(cpu
);
5423 kthread_stop(rq
->migration_thread
);
5424 rq
->migration_thread
= NULL
;
5425 /* Idle task back to normal (off runqueue, low prio) */
5426 rq
= task_rq_lock(rq
->idle
, &flags
);
5427 deactivate_task(rq
->idle
, rq
);
5428 rq
->idle
->static_prio
= MAX_PRIO
;
5429 __setscheduler(rq
->idle
, SCHED_NORMAL
, 0);
5430 migrate_dead_tasks(cpu
);
5431 task_rq_unlock(rq
, &flags
);
5432 migrate_nr_uninterruptible(rq
);
5433 BUG_ON(rq
->nr_running
!= 0);
5435 /* No need to migrate the tasks: it was best-effort if
5436 * they didn't do lock_cpu_hotplug(). Just wake up
5437 * the requestors. */
5438 spin_lock_irq(&rq
->lock
);
5439 while (!list_empty(&rq
->migration_queue
)) {
5440 struct migration_req
*req
;
5442 req
= list_entry(rq
->migration_queue
.next
,
5443 struct migration_req
, list
);
5444 list_del_init(&req
->list
);
5445 complete(&req
->done
);
5447 spin_unlock_irq(&rq
->lock
);
5454 /* Register at highest priority so that task migration (migrate_all_tasks)
5455 * happens before everything else.
5457 static struct notifier_block __cpuinitdata migration_notifier
= {
5458 .notifier_call
= migration_call
,
5462 int __init
migration_init(void)
5464 void *cpu
= (void *)(long)smp_processor_id();
5467 /* Start one for the boot CPU: */
5468 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5469 BUG_ON(err
== NOTIFY_BAD
);
5470 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5471 register_cpu_notifier(&migration_notifier
);
5479 /* Number of possible processor ids */
5480 int nr_cpu_ids __read_mostly
= NR_CPUS
;
5481 EXPORT_SYMBOL(nr_cpu_ids
);
5483 #undef SCHED_DOMAIN_DEBUG
5484 #ifdef SCHED_DOMAIN_DEBUG
5485 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5490 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5494 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5499 struct sched_group
*group
= sd
->groups
;
5500 cpumask_t groupmask
;
5502 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5503 cpus_clear(groupmask
);
5506 for (i
= 0; i
< level
+ 1; i
++)
5508 printk("domain %d: ", level
);
5510 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5511 printk("does not load-balance\n");
5513 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5518 printk("span %s\n", str
);
5520 if (!cpu_isset(cpu
, sd
->span
))
5521 printk(KERN_ERR
"ERROR: domain->span does not contain "
5523 if (!cpu_isset(cpu
, group
->cpumask
))
5524 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5528 for (i
= 0; i
< level
+ 2; i
++)
5534 printk(KERN_ERR
"ERROR: group is NULL\n");
5538 if (!group
->__cpu_power
) {
5540 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5544 if (!cpus_weight(group
->cpumask
)) {
5546 printk(KERN_ERR
"ERROR: empty group\n");
5549 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5551 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5554 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5556 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5559 group
= group
->next
;
5560 } while (group
!= sd
->groups
);
5563 if (!cpus_equal(sd
->span
, groupmask
))
5564 printk(KERN_ERR
"ERROR: groups don't span "
5572 if (!cpus_subset(groupmask
, sd
->span
))
5573 printk(KERN_ERR
"ERROR: parent span is not a superset "
5574 "of domain->span\n");
5579 # define sched_domain_debug(sd, cpu) do { } while (0)
5582 static int sd_degenerate(struct sched_domain
*sd
)
5584 if (cpus_weight(sd
->span
) == 1)
5587 /* Following flags need at least 2 groups */
5588 if (sd
->flags
& (SD_LOAD_BALANCE
|
5589 SD_BALANCE_NEWIDLE
|
5593 SD_SHARE_PKG_RESOURCES
)) {
5594 if (sd
->groups
!= sd
->groups
->next
)
5598 /* Following flags don't use groups */
5599 if (sd
->flags
& (SD_WAKE_IDLE
|
5608 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5610 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5612 if (sd_degenerate(parent
))
5615 if (!cpus_equal(sd
->span
, parent
->span
))
5618 /* Does parent contain flags not in child? */
5619 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5620 if (cflags
& SD_WAKE_AFFINE
)
5621 pflags
&= ~SD_WAKE_BALANCE
;
5622 /* Flags needing groups don't count if only 1 group in parent */
5623 if (parent
->groups
== parent
->groups
->next
) {
5624 pflags
&= ~(SD_LOAD_BALANCE
|
5625 SD_BALANCE_NEWIDLE
|
5629 SD_SHARE_PKG_RESOURCES
);
5631 if (~cflags
& pflags
)
5638 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5639 * hold the hotplug lock.
5641 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5643 struct rq
*rq
= cpu_rq(cpu
);
5644 struct sched_domain
*tmp
;
5646 /* Remove the sched domains which do not contribute to scheduling. */
5647 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5648 struct sched_domain
*parent
= tmp
->parent
;
5651 if (sd_parent_degenerate(tmp
, parent
)) {
5652 tmp
->parent
= parent
->parent
;
5654 parent
->parent
->child
= tmp
;
5658 if (sd
&& sd_degenerate(sd
)) {
5664 sched_domain_debug(sd
, cpu
);
5666 rcu_assign_pointer(rq
->sd
, sd
);
5669 /* cpus with isolated domains */
5670 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
5672 /* Setup the mask of cpus configured for isolated domains */
5673 static int __init
isolated_cpu_setup(char *str
)
5675 int ints
[NR_CPUS
], i
;
5677 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5678 cpus_clear(cpu_isolated_map
);
5679 for (i
= 1; i
<= ints
[0]; i
++)
5680 if (ints
[i
] < NR_CPUS
)
5681 cpu_set(ints
[i
], cpu_isolated_map
);
5685 __setup ("isolcpus=", isolated_cpu_setup
);
5688 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5689 * to a function which identifies what group(along with sched group) a CPU
5690 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5691 * (due to the fact that we keep track of groups covered with a cpumask_t).
5693 * init_sched_build_groups will build a circular linked list of the groups
5694 * covered by the given span, and will set each group's ->cpumask correctly,
5695 * and ->cpu_power to 0.
5698 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
5699 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
5700 struct sched_group
**sg
))
5702 struct sched_group
*first
= NULL
, *last
= NULL
;
5703 cpumask_t covered
= CPU_MASK_NONE
;
5706 for_each_cpu_mask(i
, span
) {
5707 struct sched_group
*sg
;
5708 int group
= group_fn(i
, cpu_map
, &sg
);
5711 if (cpu_isset(i
, covered
))
5714 sg
->cpumask
= CPU_MASK_NONE
;
5715 sg
->__cpu_power
= 0;
5717 for_each_cpu_mask(j
, span
) {
5718 if (group_fn(j
, cpu_map
, NULL
) != group
)
5721 cpu_set(j
, covered
);
5722 cpu_set(j
, sg
->cpumask
);
5733 #define SD_NODES_PER_DOMAIN 16
5736 * Self-tuning task migration cost measurement between source and target CPUs.
5738 * This is done by measuring the cost of manipulating buffers of varying
5739 * sizes. For a given buffer-size here are the steps that are taken:
5741 * 1) the source CPU reads+dirties a shared buffer
5742 * 2) the target CPU reads+dirties the same shared buffer
5744 * We measure how long they take, in the following 4 scenarios:
5746 * - source: CPU1, target: CPU2 | cost1
5747 * - source: CPU2, target: CPU1 | cost2
5748 * - source: CPU1, target: CPU1 | cost3
5749 * - source: CPU2, target: CPU2 | cost4
5751 * We then calculate the cost3+cost4-cost1-cost2 difference - this is
5752 * the cost of migration.
5754 * We then start off from a small buffer-size and iterate up to larger
5755 * buffer sizes, in 5% steps - measuring each buffer-size separately, and
5756 * doing a maximum search for the cost. (The maximum cost for a migration
5757 * normally occurs when the working set size is around the effective cache
5760 #define SEARCH_SCOPE 2
5761 #define MIN_CACHE_SIZE (64*1024U)
5762 #define DEFAULT_CACHE_SIZE (5*1024*1024U)
5763 #define ITERATIONS 1
5764 #define SIZE_THRESH 130
5765 #define COST_THRESH 130
5768 * The migration cost is a function of 'domain distance'. Domain
5769 * distance is the number of steps a CPU has to iterate down its
5770 * domain tree to share a domain with the other CPU. The farther
5771 * two CPUs are from each other, the larger the distance gets.
5773 * Note that we use the distance only to cache measurement results,
5774 * the distance value is not used numerically otherwise. When two
5775 * CPUs have the same distance it is assumed that the migration
5776 * cost is the same. (this is a simplification but quite practical)
5778 #define MAX_DOMAIN_DISTANCE 32
5780 static unsigned long long migration_cost
[MAX_DOMAIN_DISTANCE
] =
5781 { [ 0 ... MAX_DOMAIN_DISTANCE
-1 ] =
5783 * Architectures may override the migration cost and thus avoid
5784 * boot-time calibration. Unit is nanoseconds. Mostly useful for
5785 * virtualized hardware:
5787 #ifdef CONFIG_DEFAULT_MIGRATION_COST
5788 CONFIG_DEFAULT_MIGRATION_COST
5795 * Allow override of migration cost - in units of microseconds.
5796 * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost
5797 * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs:
5799 static int __init
migration_cost_setup(char *str
)
5801 int ints
[MAX_DOMAIN_DISTANCE
+1], i
;
5803 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5805 printk("#ints: %d\n", ints
[0]);
5806 for (i
= 1; i
<= ints
[0]; i
++) {
5807 migration_cost
[i
-1] = (unsigned long long)ints
[i
]*1000;
5808 printk("migration_cost[%d]: %Ld\n", i
-1, migration_cost
[i
-1]);
5813 __setup ("migration_cost=", migration_cost_setup
);
5816 * Global multiplier (divisor) for migration-cutoff values,
5817 * in percentiles. E.g. use a value of 150 to get 1.5 times
5818 * longer cache-hot cutoff times.
5820 * (We scale it from 100 to 128 to long long handling easier.)
5823 #define MIGRATION_FACTOR_SCALE 128
5825 static unsigned int migration_factor
= MIGRATION_FACTOR_SCALE
;
5827 static int __init
setup_migration_factor(char *str
)
5829 get_option(&str
, &migration_factor
);
5830 migration_factor
= migration_factor
* MIGRATION_FACTOR_SCALE
/ 100;
5834 __setup("migration_factor=", setup_migration_factor
);
5837 * Estimated distance of two CPUs, measured via the number of domains
5838 * we have to pass for the two CPUs to be in the same span:
5840 static unsigned long domain_distance(int cpu1
, int cpu2
)
5842 unsigned long distance
= 0;
5843 struct sched_domain
*sd
;
5845 for_each_domain(cpu1
, sd
) {
5846 WARN_ON(!cpu_isset(cpu1
, sd
->span
));
5847 if (cpu_isset(cpu2
, sd
->span
))
5851 if (distance
>= MAX_DOMAIN_DISTANCE
) {
5853 distance
= MAX_DOMAIN_DISTANCE
-1;
5859 static unsigned int migration_debug
;
5861 static int __init
setup_migration_debug(char *str
)
5863 get_option(&str
, &migration_debug
);
5867 __setup("migration_debug=", setup_migration_debug
);
5870 * Maximum cache-size that the scheduler should try to measure.
5871 * Architectures with larger caches should tune this up during
5872 * bootup. Gets used in the domain-setup code (i.e. during SMP
5875 unsigned int max_cache_size
;
5877 static int __init
setup_max_cache_size(char *str
)
5879 get_option(&str
, &max_cache_size
);
5883 __setup("max_cache_size=", setup_max_cache_size
);
5886 * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This
5887 * is the operation that is timed, so we try to generate unpredictable
5888 * cachemisses that still end up filling the L2 cache:
5890 static void touch_cache(void *__cache
, unsigned long __size
)
5892 unsigned long size
= __size
/ sizeof(long);
5893 unsigned long chunk1
= size
/ 3;
5894 unsigned long chunk2
= 2 * size
/ 3;
5895 unsigned long *cache
= __cache
;
5898 for (i
= 0; i
< size
/6; i
+= 8) {
5901 case 1: cache
[size
-1-i
]++;
5902 case 2: cache
[chunk1
-i
]++;
5903 case 3: cache
[chunk1
+i
]++;
5904 case 4: cache
[chunk2
-i
]++;
5905 case 5: cache
[chunk2
+i
]++;
5911 * Measure the cache-cost of one task migration. Returns in units of nsec.
5913 static unsigned long long
5914 measure_one(void *cache
, unsigned long size
, int source
, int target
)
5916 cpumask_t mask
, saved_mask
;
5917 unsigned long long t0
, t1
, t2
, t3
, cost
;
5919 saved_mask
= current
->cpus_allowed
;
5922 * Flush source caches to RAM and invalidate them:
5927 * Migrate to the source CPU:
5929 mask
= cpumask_of_cpu(source
);
5930 set_cpus_allowed(current
, mask
);
5931 WARN_ON(smp_processor_id() != source
);
5934 * Dirty the working set:
5937 touch_cache(cache
, size
);
5941 * Migrate to the target CPU, dirty the L2 cache and access
5942 * the shared buffer. (which represents the working set
5943 * of a migrated task.)
5945 mask
= cpumask_of_cpu(target
);
5946 set_cpus_allowed(current
, mask
);
5947 WARN_ON(smp_processor_id() != target
);
5950 touch_cache(cache
, size
);
5953 cost
= t1
-t0
+ t3
-t2
;
5955 if (migration_debug
>= 2)
5956 printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n",
5957 source
, target
, t1
-t0
, t1
-t0
, t3
-t2
, cost
);
5959 * Flush target caches to RAM and invalidate them:
5963 set_cpus_allowed(current
, saved_mask
);
5969 * Measure a series of task migrations and return the average
5970 * result. Since this code runs early during bootup the system
5971 * is 'undisturbed' and the average latency makes sense.
5973 * The algorithm in essence auto-detects the relevant cache-size,
5974 * so it will properly detect different cachesizes for different
5975 * cache-hierarchies, depending on how the CPUs are connected.
5977 * Architectures can prime the upper limit of the search range via
5978 * max_cache_size, otherwise the search range defaults to 20MB...64K.
5980 static unsigned long long
5981 measure_cost(int cpu1
, int cpu2
, void *cache
, unsigned int size
)
5983 unsigned long long cost1
, cost2
;
5987 * Measure the migration cost of 'size' bytes, over an
5988 * average of 10 runs:
5990 * (We perturb the cache size by a small (0..4k)
5991 * value to compensate size/alignment related artifacts.
5992 * We also subtract the cost of the operation done on
5998 * dry run, to make sure we start off cache-cold on cpu1,
5999 * and to get any vmalloc pagefaults in advance:
6001 measure_one(cache
, size
, cpu1
, cpu2
);
6002 for (i
= 0; i
< ITERATIONS
; i
++)
6003 cost1
+= measure_one(cache
, size
- i
* 1024, cpu1
, cpu2
);
6005 measure_one(cache
, size
, cpu2
, cpu1
);
6006 for (i
= 0; i
< ITERATIONS
; i
++)
6007 cost1
+= measure_one(cache
, size
- i
* 1024, cpu2
, cpu1
);
6010 * (We measure the non-migrating [cached] cost on both
6011 * cpu1 and cpu2, to handle CPUs with different speeds)
6015 measure_one(cache
, size
, cpu1
, cpu1
);
6016 for (i
= 0; i
< ITERATIONS
; i
++)
6017 cost2
+= measure_one(cache
, size
- i
* 1024, cpu1
, cpu1
);
6019 measure_one(cache
, size
, cpu2
, cpu2
);
6020 for (i
= 0; i
< ITERATIONS
; i
++)
6021 cost2
+= measure_one(cache
, size
- i
* 1024, cpu2
, cpu2
);
6024 * Get the per-iteration migration cost:
6026 do_div(cost1
, 2 * ITERATIONS
);
6027 do_div(cost2
, 2 * ITERATIONS
);
6029 return cost1
- cost2
;
6032 static unsigned long long measure_migration_cost(int cpu1
, int cpu2
)
6034 unsigned long long max_cost
= 0, fluct
= 0, avg_fluct
= 0;
6035 unsigned int max_size
, size
, size_found
= 0;
6036 long long cost
= 0, prev_cost
;
6040 * Search from max_cache_size*5 down to 64K - the real relevant
6041 * cachesize has to lie somewhere inbetween.
6043 if (max_cache_size
) {
6044 max_size
= max(max_cache_size
* SEARCH_SCOPE
, MIN_CACHE_SIZE
);
6045 size
= max(max_cache_size
/ SEARCH_SCOPE
, MIN_CACHE_SIZE
);
6048 * Since we have no estimation about the relevant
6051 max_size
= DEFAULT_CACHE_SIZE
* SEARCH_SCOPE
;
6052 size
= MIN_CACHE_SIZE
;
6055 if (!cpu_online(cpu1
) || !cpu_online(cpu2
)) {
6056 printk("cpu %d and %d not both online!\n", cpu1
, cpu2
);
6061 * Allocate the working set:
6063 cache
= vmalloc(max_size
);
6065 printk("could not vmalloc %d bytes for cache!\n", 2 * max_size
);
6066 return 1000000; /* return 1 msec on very small boxen */
6069 while (size
<= max_size
) {
6071 cost
= measure_cost(cpu1
, cpu2
, cache
, size
);
6077 if (max_cost
< cost
) {
6083 * Calculate average fluctuation, we use this to prevent
6084 * noise from triggering an early break out of the loop:
6086 fluct
= abs(cost
- prev_cost
);
6087 avg_fluct
= (avg_fluct
+ fluct
)/2;
6089 if (migration_debug
)
6090 printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): "
6093 (long)cost
/ 1000000,
6094 ((long)cost
/ 100000) % 10,
6095 (long)max_cost
/ 1000000,
6096 ((long)max_cost
/ 100000) % 10,
6097 domain_distance(cpu1
, cpu2
),
6101 * If we iterated at least 20% past the previous maximum,
6102 * and the cost has dropped by more than 20% already,
6103 * (taking fluctuations into account) then we assume to
6104 * have found the maximum and break out of the loop early:
6106 if (size_found
&& (size
*100 > size_found
*SIZE_THRESH
))
6107 if (cost
+avg_fluct
<= 0 ||
6108 max_cost
*100 > (cost
+avg_fluct
)*COST_THRESH
) {
6110 if (migration_debug
)
6111 printk("-> found max.\n");
6115 * Increase the cachesize in 10% steps:
6117 size
= size
* 10 / 9;
6120 if (migration_debug
)
6121 printk("[%d][%d] working set size found: %d, cost: %Ld\n",
6122 cpu1
, cpu2
, size_found
, max_cost
);
6127 * A task is considered 'cache cold' if at least 2 times
6128 * the worst-case cost of migration has passed.
6130 * (this limit is only listened to if the load-balancing
6131 * situation is 'nice' - if there is a large imbalance we
6132 * ignore it for the sake of CPU utilization and
6133 * processing fairness.)
6135 return 2 * max_cost
* migration_factor
/ MIGRATION_FACTOR_SCALE
;
6138 static void calibrate_migration_costs(const cpumask_t
*cpu_map
)
6140 int cpu1
= -1, cpu2
= -1, cpu
, orig_cpu
= raw_smp_processor_id();
6141 unsigned long j0
, j1
, distance
, max_distance
= 0;
6142 struct sched_domain
*sd
;
6147 * First pass - calculate the cacheflush times:
6149 for_each_cpu_mask(cpu1
, *cpu_map
) {
6150 for_each_cpu_mask(cpu2
, *cpu_map
) {
6153 distance
= domain_distance(cpu1
, cpu2
);
6154 max_distance
= max(max_distance
, distance
);
6156 * No result cached yet?
6158 if (migration_cost
[distance
] == -1LL)
6159 migration_cost
[distance
] =
6160 measure_migration_cost(cpu1
, cpu2
);
6164 * Second pass - update the sched domain hierarchy with
6165 * the new cache-hot-time estimations:
6167 for_each_cpu_mask(cpu
, *cpu_map
) {
6169 for_each_domain(cpu
, sd
) {
6170 sd
->cache_hot_time
= migration_cost
[distance
];
6177 if (migration_debug
)
6178 printk("migration: max_cache_size: %d, cpu: %d MHz:\n",
6186 if (system_state
== SYSTEM_BOOTING
&& num_online_cpus() > 1) {
6187 printk("migration_cost=");
6188 for (distance
= 0; distance
<= max_distance
; distance
++) {
6191 printk("%ld", (long)migration_cost
[distance
] / 1000);
6196 if (migration_debug
)
6197 printk("migration: %ld seconds\n", (j1
-j0
) / HZ
);
6200 * Move back to the original CPU. NUMA-Q gets confused
6201 * if we migrate to another quad during bootup.
6203 if (raw_smp_processor_id() != orig_cpu
) {
6204 cpumask_t mask
= cpumask_of_cpu(orig_cpu
),
6205 saved_mask
= current
->cpus_allowed
;
6207 set_cpus_allowed(current
, mask
);
6208 set_cpus_allowed(current
, saved_mask
);
6215 * find_next_best_node - find the next node to include in a sched_domain
6216 * @node: node whose sched_domain we're building
6217 * @used_nodes: nodes already in the sched_domain
6219 * Find the next node to include in a given scheduling domain. Simply
6220 * finds the closest node not already in the @used_nodes map.
6222 * Should use nodemask_t.
6224 static int find_next_best_node(int node
, unsigned long *used_nodes
)
6226 int i
, n
, val
, min_val
, best_node
= 0;
6230 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6231 /* Start at @node */
6232 n
= (node
+ i
) % MAX_NUMNODES
;
6234 if (!nr_cpus_node(n
))
6237 /* Skip already used nodes */
6238 if (test_bit(n
, used_nodes
))
6241 /* Simple min distance search */
6242 val
= node_distance(node
, n
);
6244 if (val
< min_val
) {
6250 set_bit(best_node
, used_nodes
);
6255 * sched_domain_node_span - get a cpumask for a node's sched_domain
6256 * @node: node whose cpumask we're constructing
6257 * @size: number of nodes to include in this span
6259 * Given a node, construct a good cpumask for its sched_domain to span. It
6260 * should be one that prevents unnecessary balancing, but also spreads tasks
6263 static cpumask_t
sched_domain_node_span(int node
)
6265 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
6266 cpumask_t span
, nodemask
;
6270 bitmap_zero(used_nodes
, MAX_NUMNODES
);
6272 nodemask
= node_to_cpumask(node
);
6273 cpus_or(span
, span
, nodemask
);
6274 set_bit(node
, used_nodes
);
6276 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6277 int next_node
= find_next_best_node(node
, used_nodes
);
6279 nodemask
= node_to_cpumask(next_node
);
6280 cpus_or(span
, span
, nodemask
);
6287 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6290 * SMT sched-domains:
6292 #ifdef CONFIG_SCHED_SMT
6293 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
6294 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
6296 static int cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
,
6297 struct sched_group
**sg
)
6300 *sg
= &per_cpu(sched_group_cpus
, cpu
);
6306 * multi-core sched-domains:
6308 #ifdef CONFIG_SCHED_MC
6309 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
6310 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
6313 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6314 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
6315 struct sched_group
**sg
)
6318 cpumask_t mask
= cpu_sibling_map
[cpu
];
6319 cpus_and(mask
, mask
, *cpu_map
);
6320 group
= first_cpu(mask
);
6322 *sg
= &per_cpu(sched_group_core
, group
);
6325 #elif defined(CONFIG_SCHED_MC)
6326 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
6327 struct sched_group
**sg
)
6330 *sg
= &per_cpu(sched_group_core
, cpu
);
6335 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
6336 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
6338 static int cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
,
6339 struct sched_group
**sg
)
6342 #ifdef CONFIG_SCHED_MC
6343 cpumask_t mask
= cpu_coregroup_map(cpu
);
6344 cpus_and(mask
, mask
, *cpu_map
);
6345 group
= first_cpu(mask
);
6346 #elif defined(CONFIG_SCHED_SMT)
6347 cpumask_t mask
= cpu_sibling_map
[cpu
];
6348 cpus_and(mask
, mask
, *cpu_map
);
6349 group
= first_cpu(mask
);
6354 *sg
= &per_cpu(sched_group_phys
, group
);
6360 * The init_sched_build_groups can't handle what we want to do with node
6361 * groups, so roll our own. Now each node has its own list of groups which
6362 * gets dynamically allocated.
6364 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
6365 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
6367 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
6368 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
6370 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
6371 struct sched_group
**sg
)
6373 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
6376 cpus_and(nodemask
, nodemask
, *cpu_map
);
6377 group
= first_cpu(nodemask
);
6380 *sg
= &per_cpu(sched_group_allnodes
, group
);
6384 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6386 struct sched_group
*sg
= group_head
;
6392 for_each_cpu_mask(j
, sg
->cpumask
) {
6393 struct sched_domain
*sd
;
6395 sd
= &per_cpu(phys_domains
, j
);
6396 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
6398 * Only add "power" once for each
6404 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
6407 if (sg
!= group_head
)
6413 /* Free memory allocated for various sched_group structures */
6414 static void free_sched_groups(const cpumask_t
*cpu_map
)
6418 for_each_cpu_mask(cpu
, *cpu_map
) {
6419 struct sched_group
**sched_group_nodes
6420 = sched_group_nodes_bycpu
[cpu
];
6422 if (!sched_group_nodes
)
6425 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6426 cpumask_t nodemask
= node_to_cpumask(i
);
6427 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6429 cpus_and(nodemask
, nodemask
, *cpu_map
);
6430 if (cpus_empty(nodemask
))
6440 if (oldsg
!= sched_group_nodes
[i
])
6443 kfree(sched_group_nodes
);
6444 sched_group_nodes_bycpu
[cpu
] = NULL
;
6448 static void free_sched_groups(const cpumask_t
*cpu_map
)
6454 * Initialize sched groups cpu_power.
6456 * cpu_power indicates the capacity of sched group, which is used while
6457 * distributing the load between different sched groups in a sched domain.
6458 * Typically cpu_power for all the groups in a sched domain will be same unless
6459 * there are asymmetries in the topology. If there are asymmetries, group
6460 * having more cpu_power will pickup more load compared to the group having
6463 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6464 * the maximum number of tasks a group can handle in the presence of other idle
6465 * or lightly loaded groups in the same sched domain.
6467 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6469 struct sched_domain
*child
;
6470 struct sched_group
*group
;
6472 WARN_ON(!sd
|| !sd
->groups
);
6474 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
6479 sd
->groups
->__cpu_power
= 0;
6482 * For perf policy, if the groups in child domain share resources
6483 * (for example cores sharing some portions of the cache hierarchy
6484 * or SMT), then set this domain groups cpu_power such that each group
6485 * can handle only one task, when there are other idle groups in the
6486 * same sched domain.
6488 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
6490 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
6491 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
6496 * add cpu_power of each child group to this groups cpu_power
6498 group
= child
->groups
;
6500 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
6501 group
= group
->next
;
6502 } while (group
!= child
->groups
);
6506 * Build sched domains for a given set of cpus and attach the sched domains
6507 * to the individual cpus
6509 static int build_sched_domains(const cpumask_t
*cpu_map
)
6512 struct sched_domain
*sd
;
6514 struct sched_group
**sched_group_nodes
= NULL
;
6515 int sd_allnodes
= 0;
6518 * Allocate the per-node list of sched groups
6520 sched_group_nodes
= kzalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
6522 if (!sched_group_nodes
) {
6523 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6526 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6530 * Set up domains for cpus specified by the cpu_map.
6532 for_each_cpu_mask(i
, *cpu_map
) {
6533 struct sched_domain
*sd
= NULL
, *p
;
6534 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6536 cpus_and(nodemask
, nodemask
, *cpu_map
);
6539 if (cpus_weight(*cpu_map
)
6540 > SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6541 sd
= &per_cpu(allnodes_domains
, i
);
6542 *sd
= SD_ALLNODES_INIT
;
6543 sd
->span
= *cpu_map
;
6544 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6550 sd
= &per_cpu(node_domains
, i
);
6552 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6556 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6560 sd
= &per_cpu(phys_domains
, i
);
6562 sd
->span
= nodemask
;
6566 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6568 #ifdef CONFIG_SCHED_MC
6570 sd
= &per_cpu(core_domains
, i
);
6572 sd
->span
= cpu_coregroup_map(i
);
6573 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6576 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6579 #ifdef CONFIG_SCHED_SMT
6581 sd
= &per_cpu(cpu_domains
, i
);
6582 *sd
= SD_SIBLING_INIT
;
6583 sd
->span
= cpu_sibling_map
[i
];
6584 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6587 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6591 #ifdef CONFIG_SCHED_SMT
6592 /* Set up CPU (sibling) groups */
6593 for_each_cpu_mask(i
, *cpu_map
) {
6594 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
6595 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6596 if (i
!= first_cpu(this_sibling_map
))
6599 init_sched_build_groups(this_sibling_map
, cpu_map
, &cpu_to_cpu_group
);
6603 #ifdef CONFIG_SCHED_MC
6604 /* Set up multi-core groups */
6605 for_each_cpu_mask(i
, *cpu_map
) {
6606 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6607 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6608 if (i
!= first_cpu(this_core_map
))
6610 init_sched_build_groups(this_core_map
, cpu_map
, &cpu_to_core_group
);
6615 /* Set up physical groups */
6616 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6617 cpumask_t nodemask
= node_to_cpumask(i
);
6619 cpus_and(nodemask
, nodemask
, *cpu_map
);
6620 if (cpus_empty(nodemask
))
6623 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6627 /* Set up node groups */
6629 init_sched_build_groups(*cpu_map
, cpu_map
, &cpu_to_allnodes_group
);
6631 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6632 /* Set up node groups */
6633 struct sched_group
*sg
, *prev
;
6634 cpumask_t nodemask
= node_to_cpumask(i
);
6635 cpumask_t domainspan
;
6636 cpumask_t covered
= CPU_MASK_NONE
;
6639 cpus_and(nodemask
, nodemask
, *cpu_map
);
6640 if (cpus_empty(nodemask
)) {
6641 sched_group_nodes
[i
] = NULL
;
6645 domainspan
= sched_domain_node_span(i
);
6646 cpus_and(domainspan
, domainspan
, *cpu_map
);
6648 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6650 printk(KERN_WARNING
"Can not alloc domain group for "
6654 sched_group_nodes
[i
] = sg
;
6655 for_each_cpu_mask(j
, nodemask
) {
6656 struct sched_domain
*sd
;
6657 sd
= &per_cpu(node_domains
, j
);
6660 sg
->__cpu_power
= 0;
6661 sg
->cpumask
= nodemask
;
6663 cpus_or(covered
, covered
, nodemask
);
6666 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6667 cpumask_t tmp
, notcovered
;
6668 int n
= (i
+ j
) % MAX_NUMNODES
;
6670 cpus_complement(notcovered
, covered
);
6671 cpus_and(tmp
, notcovered
, *cpu_map
);
6672 cpus_and(tmp
, tmp
, domainspan
);
6673 if (cpus_empty(tmp
))
6676 nodemask
= node_to_cpumask(n
);
6677 cpus_and(tmp
, tmp
, nodemask
);
6678 if (cpus_empty(tmp
))
6681 sg
= kmalloc_node(sizeof(struct sched_group
),
6685 "Can not alloc domain group for node %d\n", j
);
6688 sg
->__cpu_power
= 0;
6690 sg
->next
= prev
->next
;
6691 cpus_or(covered
, covered
, tmp
);
6698 /* Calculate CPU power for physical packages and nodes */
6699 #ifdef CONFIG_SCHED_SMT
6700 for_each_cpu_mask(i
, *cpu_map
) {
6701 sd
= &per_cpu(cpu_domains
, i
);
6702 init_sched_groups_power(i
, sd
);
6705 #ifdef CONFIG_SCHED_MC
6706 for_each_cpu_mask(i
, *cpu_map
) {
6707 sd
= &per_cpu(core_domains
, i
);
6708 init_sched_groups_power(i
, sd
);
6712 for_each_cpu_mask(i
, *cpu_map
) {
6713 sd
= &per_cpu(phys_domains
, i
);
6714 init_sched_groups_power(i
, sd
);
6718 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6719 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6722 struct sched_group
*sg
;
6724 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6725 init_numa_sched_groups_power(sg
);
6729 /* Attach the domains */
6730 for_each_cpu_mask(i
, *cpu_map
) {
6731 struct sched_domain
*sd
;
6732 #ifdef CONFIG_SCHED_SMT
6733 sd
= &per_cpu(cpu_domains
, i
);
6734 #elif defined(CONFIG_SCHED_MC)
6735 sd
= &per_cpu(core_domains
, i
);
6737 sd
= &per_cpu(phys_domains
, i
);
6739 cpu_attach_domain(sd
, i
);
6742 * Tune cache-hot values:
6744 calibrate_migration_costs(cpu_map
);
6750 free_sched_groups(cpu_map
);
6755 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6757 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6759 cpumask_t cpu_default_map
;
6763 * Setup mask for cpus without special case scheduling requirements.
6764 * For now this just excludes isolated cpus, but could be used to
6765 * exclude other special cases in the future.
6767 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
6769 err
= build_sched_domains(&cpu_default_map
);
6774 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6776 free_sched_groups(cpu_map
);
6780 * Detach sched domains from a group of cpus specified in cpu_map
6781 * These cpus will now be attached to the NULL domain
6783 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6787 for_each_cpu_mask(i
, *cpu_map
)
6788 cpu_attach_domain(NULL
, i
);
6789 synchronize_sched();
6790 arch_destroy_sched_domains(cpu_map
);
6794 * Partition sched domains as specified by the cpumasks below.
6795 * This attaches all cpus from the cpumasks to the NULL domain,
6796 * waits for a RCU quiescent period, recalculates sched
6797 * domain information and then attaches them back to the
6798 * correct sched domains
6799 * Call with hotplug lock held
6801 int partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
6803 cpumask_t change_map
;
6806 cpus_and(*partition1
, *partition1
, cpu_online_map
);
6807 cpus_and(*partition2
, *partition2
, cpu_online_map
);
6808 cpus_or(change_map
, *partition1
, *partition2
);
6810 /* Detach sched domains from all of the affected cpus */
6811 detach_destroy_domains(&change_map
);
6812 if (!cpus_empty(*partition1
))
6813 err
= build_sched_domains(partition1
);
6814 if (!err
&& !cpus_empty(*partition2
))
6815 err
= build_sched_domains(partition2
);
6820 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6821 int arch_reinit_sched_domains(void)
6826 detach_destroy_domains(&cpu_online_map
);
6827 err
= arch_init_sched_domains(&cpu_online_map
);
6828 unlock_cpu_hotplug();
6833 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6837 if (buf
[0] != '0' && buf
[0] != '1')
6841 sched_smt_power_savings
= (buf
[0] == '1');
6843 sched_mc_power_savings
= (buf
[0] == '1');
6845 ret
= arch_reinit_sched_domains();
6847 return ret
? ret
: count
;
6850 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6854 #ifdef CONFIG_SCHED_SMT
6856 err
= sysfs_create_file(&cls
->kset
.kobj
,
6857 &attr_sched_smt_power_savings
.attr
);
6859 #ifdef CONFIG_SCHED_MC
6860 if (!err
&& mc_capable())
6861 err
= sysfs_create_file(&cls
->kset
.kobj
,
6862 &attr_sched_mc_power_savings
.attr
);
6868 #ifdef CONFIG_SCHED_MC
6869 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6871 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6873 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6874 const char *buf
, size_t count
)
6876 return sched_power_savings_store(buf
, count
, 0);
6878 SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6879 sched_mc_power_savings_store
);
6882 #ifdef CONFIG_SCHED_SMT
6883 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6885 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6887 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6888 const char *buf
, size_t count
)
6890 return sched_power_savings_store(buf
, count
, 1);
6892 SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6893 sched_smt_power_savings_store
);
6897 * Force a reinitialization of the sched domains hierarchy. The domains
6898 * and groups cannot be updated in place without racing with the balancing
6899 * code, so we temporarily attach all running cpus to the NULL domain
6900 * which will prevent rebalancing while the sched domains are recalculated.
6902 static int update_sched_domains(struct notifier_block
*nfb
,
6903 unsigned long action
, void *hcpu
)
6906 case CPU_UP_PREPARE
:
6907 case CPU_DOWN_PREPARE
:
6908 detach_destroy_domains(&cpu_online_map
);
6911 case CPU_UP_CANCELED
:
6912 case CPU_DOWN_FAILED
:
6916 * Fall through and re-initialise the domains.
6923 /* The hotplug lock is already held by cpu_up/cpu_down */
6924 arch_init_sched_domains(&cpu_online_map
);
6929 void __init
sched_init_smp(void)
6931 cpumask_t non_isolated_cpus
;
6934 arch_init_sched_domains(&cpu_online_map
);
6935 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
6936 if (cpus_empty(non_isolated_cpus
))
6937 cpu_set(smp_processor_id(), non_isolated_cpus
);
6938 unlock_cpu_hotplug();
6939 /* XXX: Theoretical race here - CPU may be hotplugged now */
6940 hotcpu_notifier(update_sched_domains
, 0);
6942 /* Move init over to a non-isolated CPU */
6943 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
6947 void __init
sched_init_smp(void)
6950 #endif /* CONFIG_SMP */
6952 int in_sched_functions(unsigned long addr
)
6954 /* Linker adds these: start and end of __sched functions */
6955 extern char __sched_text_start
[], __sched_text_end
[];
6957 return in_lock_functions(addr
) ||
6958 (addr
>= (unsigned long)__sched_text_start
6959 && addr
< (unsigned long)__sched_text_end
);
6962 void __init
sched_init(void)
6965 int highest_cpu
= 0;
6967 for_each_possible_cpu(i
) {
6968 struct prio_array
*array
;
6972 spin_lock_init(&rq
->lock
);
6973 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
6975 rq
->active
= rq
->arrays
;
6976 rq
->expired
= rq
->arrays
+ 1;
6977 rq
->best_expired_prio
= MAX_PRIO
;
6981 for (j
= 1; j
< 3; j
++)
6982 rq
->cpu_load
[j
] = 0;
6983 rq
->active_balance
= 0;
6986 rq
->migration_thread
= NULL
;
6987 INIT_LIST_HEAD(&rq
->migration_queue
);
6989 atomic_set(&rq
->nr_iowait
, 0);
6991 for (j
= 0; j
< 2; j
++) {
6992 array
= rq
->arrays
+ j
;
6993 for (k
= 0; k
< MAX_PRIO
; k
++) {
6994 INIT_LIST_HEAD(array
->queue
+ k
);
6995 __clear_bit(k
, array
->bitmap
);
6997 // delimiter for bitsearch
6998 __set_bit(MAX_PRIO
, array
->bitmap
);
7003 set_load_weight(&init_task
);
7006 nr_cpu_ids
= highest_cpu
+ 1;
7007 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
7010 #ifdef CONFIG_RT_MUTEXES
7011 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
7015 * The boot idle thread does lazy MMU switching as well:
7017 atomic_inc(&init_mm
.mm_count
);
7018 enter_lazy_tlb(&init_mm
, current
);
7021 * Make us the idle thread. Technically, schedule() should not be
7022 * called from this thread, however somewhere below it might be,
7023 * but because we are the idle thread, we just pick up running again
7024 * when this runqueue becomes "idle".
7026 init_idle(current
, smp_processor_id());
7029 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7030 void __might_sleep(char *file
, int line
)
7033 static unsigned long prev_jiffy
; /* ratelimiting */
7035 if ((in_atomic() || irqs_disabled()) &&
7036 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
7037 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7039 prev_jiffy
= jiffies
;
7040 printk(KERN_ERR
"BUG: sleeping function called from invalid"
7041 " context at %s:%d\n", file
, line
);
7042 printk("in_atomic():%d, irqs_disabled():%d\n",
7043 in_atomic(), irqs_disabled());
7044 debug_show_held_locks(current
);
7045 if (irqs_disabled())
7046 print_irqtrace_events(current
);
7051 EXPORT_SYMBOL(__might_sleep
);
7054 #ifdef CONFIG_MAGIC_SYSRQ
7055 void normalize_rt_tasks(void)
7057 struct prio_array
*array
;
7058 struct task_struct
*p
;
7059 unsigned long flags
;
7062 read_lock_irq(&tasklist_lock
);
7063 for_each_process(p
) {
7067 spin_lock_irqsave(&p
->pi_lock
, flags
);
7068 rq
= __task_rq_lock(p
);
7072 deactivate_task(p
, task_rq(p
));
7073 __setscheduler(p
, SCHED_NORMAL
, 0);
7075 __activate_task(p
, task_rq(p
));
7076 resched_task(rq
->curr
);
7079 __task_rq_unlock(rq
);
7080 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
7082 read_unlock_irq(&tasklist_lock
);
7085 #endif /* CONFIG_MAGIC_SYSRQ */
7089 * These functions are only useful for the IA64 MCA handling.
7091 * They can only be called when the whole system has been
7092 * stopped - every CPU needs to be quiescent, and no scheduling
7093 * activity can take place. Using them for anything else would
7094 * be a serious bug, and as a result, they aren't even visible
7095 * under any other configuration.
7099 * curr_task - return the current task for a given cpu.
7100 * @cpu: the processor in question.
7102 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7104 struct task_struct
*curr_task(int cpu
)
7106 return cpu_curr(cpu
);
7110 * set_curr_task - set the current task for a given cpu.
7111 * @cpu: the processor in question.
7112 * @p: the task pointer to set.
7114 * Description: This function must only be used when non-maskable interrupts
7115 * are serviced on a separate stack. It allows the architecture to switch the
7116 * notion of the current task on a cpu in a non-blocking manner. This function
7117 * must be called with all CPU's synchronized, and interrupts disabled, the
7118 * and caller must save the original value of the current task (see
7119 * curr_task() above) and restore that value before reenabling interrupts and
7120 * re-starting the system.
7122 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7124 void set_curr_task(int cpu
, struct task_struct
*p
)