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
) ____cacheline_aligned_in_smp
;
308 static DEFINE_MUTEX(sched_hotcpu_mutex
);
310 static inline int cpu_of(struct rq
*rq
)
320 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
321 * See detach_destroy_domains: synchronize_sched for details.
323 * The domain tree of any CPU may only be accessed from within
324 * preempt-disabled sections.
326 #define for_each_domain(cpu, __sd) \
327 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
329 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
330 #define this_rq() (&__get_cpu_var(runqueues))
331 #define task_rq(p) cpu_rq(task_cpu(p))
332 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
334 #ifndef prepare_arch_switch
335 # define prepare_arch_switch(next) do { } while (0)
337 #ifndef finish_arch_switch
338 # define finish_arch_switch(prev) do { } while (0)
341 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
342 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
344 return rq
->curr
== p
;
347 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
351 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
353 #ifdef CONFIG_DEBUG_SPINLOCK
354 /* this is a valid case when another task releases the spinlock */
355 rq
->lock
.owner
= current
;
358 * If we are tracking spinlock dependencies then we have to
359 * fix up the runqueue lock - which gets 'carried over' from
362 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
364 spin_unlock_irq(&rq
->lock
);
367 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
368 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
373 return rq
->curr
== p
;
377 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
381 * We can optimise this out completely for !SMP, because the
382 * SMP rebalancing from interrupt is the only thing that cares
387 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
388 spin_unlock_irq(&rq
->lock
);
390 spin_unlock(&rq
->lock
);
394 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
398 * After ->oncpu is cleared, the task can be moved to a different CPU.
399 * We must ensure this doesn't happen until the switch is completely
405 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
409 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
412 * __task_rq_lock - lock the runqueue a given task resides on.
413 * Must be called interrupts disabled.
415 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
422 spin_lock(&rq
->lock
);
423 if (unlikely(rq
!= task_rq(p
))) {
424 spin_unlock(&rq
->lock
);
425 goto repeat_lock_task
;
431 * task_rq_lock - lock the runqueue a given task resides on and disable
432 * interrupts. Note the ordering: we can safely lookup the task_rq without
433 * explicitly disabling preemption.
435 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
441 local_irq_save(*flags
);
443 spin_lock(&rq
->lock
);
444 if (unlikely(rq
!= task_rq(p
))) {
445 spin_unlock_irqrestore(&rq
->lock
, *flags
);
446 goto repeat_lock_task
;
451 static inline void __task_rq_unlock(struct rq
*rq
)
454 spin_unlock(&rq
->lock
);
457 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
460 spin_unlock_irqrestore(&rq
->lock
, *flags
);
463 #ifdef CONFIG_SCHEDSTATS
465 * bump this up when changing the output format or the meaning of an existing
466 * format, so that tools can adapt (or abort)
468 #define SCHEDSTAT_VERSION 14
470 static int show_schedstat(struct seq_file
*seq
, void *v
)
474 seq_printf(seq
, "version %d\n", SCHEDSTAT_VERSION
);
475 seq_printf(seq
, "timestamp %lu\n", jiffies
);
476 for_each_online_cpu(cpu
) {
477 struct rq
*rq
= cpu_rq(cpu
);
479 struct sched_domain
*sd
;
483 /* runqueue-specific stats */
485 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
486 cpu
, rq
->yld_both_empty
,
487 rq
->yld_act_empty
, rq
->yld_exp_empty
, rq
->yld_cnt
,
488 rq
->sched_switch
, rq
->sched_cnt
, rq
->sched_goidle
,
489 rq
->ttwu_cnt
, rq
->ttwu_local
,
490 rq
->rq_sched_info
.cpu_time
,
491 rq
->rq_sched_info
.run_delay
, rq
->rq_sched_info
.pcnt
);
493 seq_printf(seq
, "\n");
496 /* domain-specific stats */
498 for_each_domain(cpu
, sd
) {
499 enum idle_type itype
;
500 char mask_str
[NR_CPUS
];
502 cpumask_scnprintf(mask_str
, NR_CPUS
, sd
->span
);
503 seq_printf(seq
, "domain%d %s", dcnt
++, mask_str
);
504 for (itype
= SCHED_IDLE
; itype
< MAX_IDLE_TYPES
;
506 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu "
509 sd
->lb_balanced
[itype
],
510 sd
->lb_failed
[itype
],
511 sd
->lb_imbalance
[itype
],
512 sd
->lb_gained
[itype
],
513 sd
->lb_hot_gained
[itype
],
514 sd
->lb_nobusyq
[itype
],
515 sd
->lb_nobusyg
[itype
]);
517 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu %lu %lu"
519 sd
->alb_cnt
, sd
->alb_failed
, sd
->alb_pushed
,
520 sd
->sbe_cnt
, sd
->sbe_balanced
, sd
->sbe_pushed
,
521 sd
->sbf_cnt
, sd
->sbf_balanced
, sd
->sbf_pushed
,
522 sd
->ttwu_wake_remote
, sd
->ttwu_move_affine
,
523 sd
->ttwu_move_balance
);
531 static int schedstat_open(struct inode
*inode
, struct file
*file
)
533 unsigned int size
= PAGE_SIZE
* (1 + num_online_cpus() / 32);
534 char *buf
= kmalloc(size
, GFP_KERNEL
);
540 res
= single_open(file
, show_schedstat
, NULL
);
542 m
= file
->private_data
;
550 const struct file_operations proc_schedstat_operations
= {
551 .open
= schedstat_open
,
554 .release
= single_release
,
558 * Expects runqueue lock to be held for atomicity of update
561 rq_sched_info_arrive(struct rq
*rq
, unsigned long delta_jiffies
)
564 rq
->rq_sched_info
.run_delay
+= delta_jiffies
;
565 rq
->rq_sched_info
.pcnt
++;
570 * Expects runqueue lock to be held for atomicity of update
573 rq_sched_info_depart(struct rq
*rq
, unsigned long delta_jiffies
)
576 rq
->rq_sched_info
.cpu_time
+= delta_jiffies
;
578 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
579 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
580 #else /* !CONFIG_SCHEDSTATS */
582 rq_sched_info_arrive(struct rq
*rq
, unsigned long delta_jiffies
)
585 rq_sched_info_depart(struct rq
*rq
, unsigned long delta_jiffies
)
587 # define schedstat_inc(rq, field) do { } while (0)
588 # define schedstat_add(rq, field, amt) do { } while (0)
592 * this_rq_lock - lock this runqueue and disable interrupts.
594 static inline struct rq
*this_rq_lock(void)
601 spin_lock(&rq
->lock
);
606 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
608 * Called when a process is dequeued from the active array and given
609 * the cpu. We should note that with the exception of interactive
610 * tasks, the expired queue will become the active queue after the active
611 * queue is empty, without explicitly dequeuing and requeuing tasks in the
612 * expired queue. (Interactive tasks may be requeued directly to the
613 * active queue, thus delaying tasks in the expired queue from running;
614 * see scheduler_tick()).
616 * This function is only called from sched_info_arrive(), rather than
617 * dequeue_task(). Even though a task may be queued and dequeued multiple
618 * times as it is shuffled about, we're really interested in knowing how
619 * long it was from the *first* time it was queued to the time that it
622 static inline void sched_info_dequeued(struct task_struct
*t
)
624 t
->sched_info
.last_queued
= 0;
628 * Called when a task finally hits the cpu. We can now calculate how
629 * long it was waiting to run. We also note when it began so that we
630 * can keep stats on how long its timeslice is.
632 static void sched_info_arrive(struct task_struct
*t
)
634 unsigned long now
= jiffies
, delta_jiffies
= 0;
636 if (t
->sched_info
.last_queued
)
637 delta_jiffies
= now
- t
->sched_info
.last_queued
;
638 sched_info_dequeued(t
);
639 t
->sched_info
.run_delay
+= delta_jiffies
;
640 t
->sched_info
.last_arrival
= now
;
641 t
->sched_info
.pcnt
++;
643 rq_sched_info_arrive(task_rq(t
), delta_jiffies
);
647 * Called when a process is queued into either the active or expired
648 * array. The time is noted and later used to determine how long we
649 * had to wait for us to reach the cpu. Since the expired queue will
650 * become the active queue after active queue is empty, without dequeuing
651 * and requeuing any tasks, we are interested in queuing to either. It
652 * is unusual but not impossible for tasks to be dequeued and immediately
653 * requeued in the same or another array: this can happen in sched_yield(),
654 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
657 * This function is only called from enqueue_task(), but also only updates
658 * the timestamp if it is already not set. It's assumed that
659 * sched_info_dequeued() will clear that stamp when appropriate.
661 static inline void sched_info_queued(struct task_struct
*t
)
663 if (unlikely(sched_info_on()))
664 if (!t
->sched_info
.last_queued
)
665 t
->sched_info
.last_queued
= jiffies
;
669 * Called when a process ceases being the active-running process, either
670 * voluntarily or involuntarily. Now we can calculate how long we ran.
672 static inline void sched_info_depart(struct task_struct
*t
)
674 unsigned long delta_jiffies
= jiffies
- t
->sched_info
.last_arrival
;
676 t
->sched_info
.cpu_time
+= delta_jiffies
;
677 rq_sched_info_depart(task_rq(t
), delta_jiffies
);
681 * Called when tasks are switched involuntarily due, typically, to expiring
682 * their time slice. (This may also be called when switching to or from
683 * the idle task.) We are only called when prev != next.
686 __sched_info_switch(struct task_struct
*prev
, struct task_struct
*next
)
688 struct rq
*rq
= task_rq(prev
);
691 * prev now departs the cpu. It's not interesting to record
692 * stats about how efficient we were at scheduling the idle
695 if (prev
!= rq
->idle
)
696 sched_info_depart(prev
);
698 if (next
!= rq
->idle
)
699 sched_info_arrive(next
);
702 sched_info_switch(struct task_struct
*prev
, struct task_struct
*next
)
704 if (unlikely(sched_info_on()))
705 __sched_info_switch(prev
, next
);
708 #define sched_info_queued(t) do { } while (0)
709 #define sched_info_switch(t, next) do { } while (0)
710 #endif /* CONFIG_SCHEDSTATS || CONFIG_TASK_DELAY_ACCT */
713 * Adding/removing a task to/from a priority array:
715 static void dequeue_task(struct task_struct
*p
, struct prio_array
*array
)
718 list_del(&p
->run_list
);
719 if (list_empty(array
->queue
+ p
->prio
))
720 __clear_bit(p
->prio
, array
->bitmap
);
723 static void enqueue_task(struct task_struct
*p
, struct prio_array
*array
)
725 sched_info_queued(p
);
726 list_add_tail(&p
->run_list
, array
->queue
+ p
->prio
);
727 __set_bit(p
->prio
, array
->bitmap
);
733 * Put task to the end of the run list without the overhead of dequeue
734 * followed by enqueue.
736 static void requeue_task(struct task_struct
*p
, struct prio_array
*array
)
738 list_move_tail(&p
->run_list
, array
->queue
+ p
->prio
);
742 enqueue_task_head(struct task_struct
*p
, struct prio_array
*array
)
744 list_add(&p
->run_list
, array
->queue
+ p
->prio
);
745 __set_bit(p
->prio
, array
->bitmap
);
751 * __normal_prio - return the priority that is based on the static
752 * priority but is modified by bonuses/penalties.
754 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
755 * into the -5 ... 0 ... +5 bonus/penalty range.
757 * We use 25% of the full 0...39 priority range so that:
759 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
760 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
762 * Both properties are important to certain workloads.
765 static inline int __normal_prio(struct task_struct
*p
)
769 bonus
= CURRENT_BONUS(p
) - MAX_BONUS
/ 2;
771 prio
= p
->static_prio
- bonus
;
772 if (prio
< MAX_RT_PRIO
)
774 if (prio
> MAX_PRIO
-1)
780 * To aid in avoiding the subversion of "niceness" due to uneven distribution
781 * of tasks with abnormal "nice" values across CPUs the contribution that
782 * each task makes to its run queue's load is weighted according to its
783 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
784 * scaled version of the new time slice allocation that they receive on time
789 * Assume: static_prio_timeslice(NICE_TO_PRIO(0)) == DEF_TIMESLICE
790 * If static_prio_timeslice() is ever changed to break this assumption then
791 * this code will need modification
793 #define TIME_SLICE_NICE_ZERO DEF_TIMESLICE
794 #define LOAD_WEIGHT(lp) \
795 (((lp) * SCHED_LOAD_SCALE) / TIME_SLICE_NICE_ZERO)
796 #define PRIO_TO_LOAD_WEIGHT(prio) \
797 LOAD_WEIGHT(static_prio_timeslice(prio))
798 #define RTPRIO_TO_LOAD_WEIGHT(rp) \
799 (PRIO_TO_LOAD_WEIGHT(MAX_RT_PRIO) + LOAD_WEIGHT(rp))
801 static void set_load_weight(struct task_struct
*p
)
803 if (has_rt_policy(p
)) {
805 if (p
== task_rq(p
)->migration_thread
)
807 * The migration thread does the actual balancing.
808 * Giving its load any weight will skew balancing
814 p
->load_weight
= RTPRIO_TO_LOAD_WEIGHT(p
->rt_priority
);
816 p
->load_weight
= PRIO_TO_LOAD_WEIGHT(p
->static_prio
);
820 inc_raw_weighted_load(struct rq
*rq
, const struct task_struct
*p
)
822 rq
->raw_weighted_load
+= p
->load_weight
;
826 dec_raw_weighted_load(struct rq
*rq
, const struct task_struct
*p
)
828 rq
->raw_weighted_load
-= p
->load_weight
;
831 static inline void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
834 inc_raw_weighted_load(rq
, p
);
837 static inline void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
840 dec_raw_weighted_load(rq
, p
);
844 * Calculate the expected normal priority: i.e. priority
845 * without taking RT-inheritance into account. Might be
846 * boosted by interactivity modifiers. Changes upon fork,
847 * setprio syscalls, and whenever the interactivity
848 * estimator recalculates.
850 static inline int normal_prio(struct task_struct
*p
)
854 if (has_rt_policy(p
))
855 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
857 prio
= __normal_prio(p
);
862 * Calculate the current priority, i.e. the priority
863 * taken into account by the scheduler. This value might
864 * be boosted by RT tasks, or might be boosted by
865 * interactivity modifiers. Will be RT if the task got
866 * RT-boosted. If not then it returns p->normal_prio.
868 static int effective_prio(struct task_struct
*p
)
870 p
->normal_prio
= normal_prio(p
);
872 * If we are RT tasks or we were boosted to RT priority,
873 * keep the priority unchanged. Otherwise, update priority
874 * to the normal priority:
876 if (!rt_prio(p
->prio
))
877 return p
->normal_prio
;
882 * __activate_task - move a task to the runqueue.
884 static void __activate_task(struct task_struct
*p
, struct rq
*rq
)
886 struct prio_array
*target
= rq
->active
;
889 target
= rq
->expired
;
890 enqueue_task(p
, target
);
891 inc_nr_running(p
, rq
);
895 * __activate_idle_task - move idle task to the _front_ of runqueue.
897 static inline void __activate_idle_task(struct task_struct
*p
, struct rq
*rq
)
899 enqueue_task_head(p
, rq
->active
);
900 inc_nr_running(p
, rq
);
904 * Recalculate p->normal_prio and p->prio after having slept,
905 * updating the sleep-average too:
907 static int recalc_task_prio(struct task_struct
*p
, unsigned long long now
)
909 /* Caller must always ensure 'now >= p->timestamp' */
910 unsigned long sleep_time
= now
- p
->timestamp
;
915 if (likely(sleep_time
> 0)) {
917 * This ceiling is set to the lowest priority that would allow
918 * a task to be reinserted into the active array on timeslice
921 unsigned long ceiling
= INTERACTIVE_SLEEP(p
);
923 if (p
->mm
&& sleep_time
> ceiling
&& p
->sleep_avg
< ceiling
) {
925 * Prevents user tasks from achieving best priority
926 * with one single large enough sleep.
928 p
->sleep_avg
= ceiling
;
930 * Using INTERACTIVE_SLEEP() as a ceiling places a
931 * nice(0) task 1ms sleep away from promotion, and
932 * gives it 700ms to round-robin with no chance of
933 * being demoted. This is more than generous, so
934 * mark this sleep as non-interactive to prevent the
935 * on-runqueue bonus logic from intervening should
936 * this task not receive cpu immediately.
938 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
941 * Tasks waking from uninterruptible sleep are
942 * limited in their sleep_avg rise as they
943 * are likely to be waiting on I/O
945 if (p
->sleep_type
== SLEEP_NONINTERACTIVE
&& p
->mm
) {
946 if (p
->sleep_avg
>= ceiling
)
948 else if (p
->sleep_avg
+ sleep_time
>=
950 p
->sleep_avg
= ceiling
;
956 * This code gives a bonus to interactive tasks.
958 * The boost works by updating the 'average sleep time'
959 * value here, based on ->timestamp. The more time a
960 * task spends sleeping, the higher the average gets -
961 * and the higher the priority boost gets as well.
963 p
->sleep_avg
+= sleep_time
;
966 if (p
->sleep_avg
> NS_MAX_SLEEP_AVG
)
967 p
->sleep_avg
= NS_MAX_SLEEP_AVG
;
970 return effective_prio(p
);
974 * activate_task - move a task to the runqueue and do priority recalculation
976 * Update all the scheduling statistics stuff. (sleep average
977 * calculation, priority modifiers, etc.)
979 static void activate_task(struct task_struct
*p
, struct rq
*rq
, int local
)
981 unsigned long long now
;
989 /* Compensate for drifting sched_clock */
990 struct rq
*this_rq
= this_rq();
991 now
= (now
- this_rq
->most_recent_timestamp
)
992 + rq
->most_recent_timestamp
;
997 * Sleep time is in units of nanosecs, so shift by 20 to get a
998 * milliseconds-range estimation of the amount of time that the task
1001 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
1002 if (p
->state
== TASK_UNINTERRUPTIBLE
)
1003 profile_hits(SLEEP_PROFILING
, (void *)get_wchan(p
),
1004 (now
- p
->timestamp
) >> 20);
1007 p
->prio
= recalc_task_prio(p
, now
);
1010 * This checks to make sure it's not an uninterruptible task
1011 * that is now waking up.
1013 if (p
->sleep_type
== SLEEP_NORMAL
) {
1015 * Tasks which were woken up by interrupts (ie. hw events)
1016 * are most likely of interactive nature. So we give them
1017 * the credit of extending their sleep time to the period
1018 * of time they spend on the runqueue, waiting for execution
1019 * on a CPU, first time around:
1022 p
->sleep_type
= SLEEP_INTERRUPTED
;
1025 * Normal first-time wakeups get a credit too for
1026 * on-runqueue time, but it will be weighted down:
1028 p
->sleep_type
= SLEEP_INTERACTIVE
;
1033 __activate_task(p
, rq
);
1037 * deactivate_task - remove a task from the runqueue.
1039 static void deactivate_task(struct task_struct
*p
, struct rq
*rq
)
1041 dec_nr_running(p
, rq
);
1042 dequeue_task(p
, p
->array
);
1047 * resched_task - mark a task 'to be rescheduled now'.
1049 * On UP this means the setting of the need_resched flag, on SMP it
1050 * might also involve a cross-CPU call to trigger the scheduler on
1055 #ifndef tsk_is_polling
1056 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1059 static void resched_task(struct task_struct
*p
)
1063 assert_spin_locked(&task_rq(p
)->lock
);
1065 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
1068 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
1071 if (cpu
== smp_processor_id())
1074 /* NEED_RESCHED must be visible before we test polling */
1076 if (!tsk_is_polling(p
))
1077 smp_send_reschedule(cpu
);
1080 static void resched_cpu(int cpu
)
1082 struct rq
*rq
= cpu_rq(cpu
);
1083 unsigned long flags
;
1085 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1087 resched_task(cpu_curr(cpu
));
1088 spin_unlock_irqrestore(&rq
->lock
, flags
);
1091 static inline void resched_task(struct task_struct
*p
)
1093 assert_spin_locked(&task_rq(p
)->lock
);
1094 set_tsk_need_resched(p
);
1099 * task_curr - is this task currently executing on a CPU?
1100 * @p: the task in question.
1102 inline int task_curr(const struct task_struct
*p
)
1104 return cpu_curr(task_cpu(p
)) == p
;
1107 /* Used instead of source_load when we know the type == 0 */
1108 unsigned long weighted_cpuload(const int cpu
)
1110 return cpu_rq(cpu
)->raw_weighted_load
;
1114 struct migration_req
{
1115 struct list_head list
;
1117 struct task_struct
*task
;
1120 struct completion done
;
1124 * The task's runqueue lock must be held.
1125 * Returns true if you have to wait for migration thread.
1128 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1130 struct rq
*rq
= task_rq(p
);
1133 * If the task is not on a runqueue (and not running), then
1134 * it is sufficient to simply update the task's cpu field.
1136 if (!p
->array
&& !task_running(rq
, p
)) {
1137 set_task_cpu(p
, dest_cpu
);
1141 init_completion(&req
->done
);
1143 req
->dest_cpu
= dest_cpu
;
1144 list_add(&req
->list
, &rq
->migration_queue
);
1150 * wait_task_inactive - wait for a thread to unschedule.
1152 * The caller must ensure that the task *will* unschedule sometime soon,
1153 * else this function might spin for a *long* time. This function can't
1154 * be called with interrupts off, or it may introduce deadlock with
1155 * smp_call_function() if an IPI is sent by the same process we are
1156 * waiting to become inactive.
1158 void wait_task_inactive(struct task_struct
*p
)
1160 unsigned long flags
;
1162 struct prio_array
*array
;
1167 * We do the initial early heuristics without holding
1168 * any task-queue locks at all. We'll only try to get
1169 * the runqueue lock when things look like they will
1175 * If the task is actively running on another CPU
1176 * still, just relax and busy-wait without holding
1179 * NOTE! Since we don't hold any locks, it's not
1180 * even sure that "rq" stays as the right runqueue!
1181 * But we don't care, since "task_running()" will
1182 * return false if the runqueue has changed and p
1183 * is actually now running somewhere else!
1185 while (task_running(rq
, p
))
1189 * Ok, time to look more closely! We need the rq
1190 * lock now, to be *sure*. If we're wrong, we'll
1191 * just go back and repeat.
1193 rq
= task_rq_lock(p
, &flags
);
1194 running
= task_running(rq
, p
);
1196 task_rq_unlock(rq
, &flags
);
1199 * Was it really running after all now that we
1200 * checked with the proper locks actually held?
1202 * Oops. Go back and try again..
1204 if (unlikely(running
)) {
1210 * It's not enough that it's not actively running,
1211 * it must be off the runqueue _entirely_, and not
1214 * So if it wa still runnable (but just not actively
1215 * running right now), it's preempted, and we should
1216 * yield - it could be a while.
1218 if (unlikely(array
)) {
1224 * Ahh, all good. It wasn't running, and it wasn't
1225 * runnable, which means that it will never become
1226 * running in the future either. We're all done!
1231 * kick_process - kick a running thread to enter/exit the kernel
1232 * @p: the to-be-kicked thread
1234 * Cause a process which is running on another CPU to enter
1235 * kernel-mode, without any delay. (to get signals handled.)
1237 * NOTE: this function doesnt have to take the runqueue lock,
1238 * because all it wants to ensure is that the remote task enters
1239 * the kernel. If the IPI races and the task has been migrated
1240 * to another CPU then no harm is done and the purpose has been
1243 void kick_process(struct task_struct
*p
)
1249 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1250 smp_send_reschedule(cpu
);
1255 * Return a low guess at the load of a migration-source cpu weighted
1256 * according to the scheduling class and "nice" value.
1258 * We want to under-estimate the load of migration sources, to
1259 * balance conservatively.
1261 static inline unsigned long source_load(int cpu
, int type
)
1263 struct rq
*rq
= cpu_rq(cpu
);
1266 return rq
->raw_weighted_load
;
1268 return min(rq
->cpu_load
[type
-1], rq
->raw_weighted_load
);
1272 * Return a high guess at the load of a migration-target cpu weighted
1273 * according to the scheduling class and "nice" value.
1275 static inline unsigned long target_load(int cpu
, int type
)
1277 struct rq
*rq
= cpu_rq(cpu
);
1280 return rq
->raw_weighted_load
;
1282 return max(rq
->cpu_load
[type
-1], rq
->raw_weighted_load
);
1286 * Return the average load per task on the cpu's run queue
1288 static inline unsigned long cpu_avg_load_per_task(int cpu
)
1290 struct rq
*rq
= cpu_rq(cpu
);
1291 unsigned long n
= rq
->nr_running
;
1293 return n
? rq
->raw_weighted_load
/ n
: SCHED_LOAD_SCALE
;
1297 * find_idlest_group finds and returns the least busy CPU group within the
1300 static struct sched_group
*
1301 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1303 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1304 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1305 int load_idx
= sd
->forkexec_idx
;
1306 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1309 unsigned long load
, avg_load
;
1313 /* Skip over this group if it has no CPUs allowed */
1314 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1317 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1319 /* Tally up the load of all CPUs in the group */
1322 for_each_cpu_mask(i
, group
->cpumask
) {
1323 /* Bias balancing toward cpus of our domain */
1325 load
= source_load(i
, load_idx
);
1327 load
= target_load(i
, load_idx
);
1332 /* Adjust by relative CPU power of the group */
1333 avg_load
= sg_div_cpu_power(group
,
1334 avg_load
* SCHED_LOAD_SCALE
);
1337 this_load
= avg_load
;
1339 } else if (avg_load
< min_load
) {
1340 min_load
= avg_load
;
1344 group
= group
->next
;
1345 } while (group
!= sd
->groups
);
1347 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1353 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1356 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1359 unsigned long load
, min_load
= ULONG_MAX
;
1363 /* Traverse only the allowed CPUs */
1364 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1366 for_each_cpu_mask(i
, tmp
) {
1367 load
= weighted_cpuload(i
);
1369 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1379 * sched_balance_self: balance the current task (running on cpu) in domains
1380 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1383 * Balance, ie. select the least loaded group.
1385 * Returns the target CPU number, or the same CPU if no balancing is needed.
1387 * preempt must be disabled.
1389 static int sched_balance_self(int cpu
, int flag
)
1391 struct task_struct
*t
= current
;
1392 struct sched_domain
*tmp
, *sd
= NULL
;
1394 for_each_domain(cpu
, tmp
) {
1396 * If power savings logic is enabled for a domain, stop there.
1398 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1400 if (tmp
->flags
& flag
)
1406 struct sched_group
*group
;
1407 int new_cpu
, weight
;
1409 if (!(sd
->flags
& flag
)) {
1415 group
= find_idlest_group(sd
, t
, cpu
);
1421 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1422 if (new_cpu
== -1 || new_cpu
== cpu
) {
1423 /* Now try balancing at a lower domain level of cpu */
1428 /* Now try balancing at a lower domain level of new_cpu */
1431 weight
= cpus_weight(span
);
1432 for_each_domain(cpu
, tmp
) {
1433 if (weight
<= cpus_weight(tmp
->span
))
1435 if (tmp
->flags
& flag
)
1438 /* while loop will break here if sd == NULL */
1444 #endif /* CONFIG_SMP */
1447 * wake_idle() will wake a task on an idle cpu if task->cpu is
1448 * not idle and an idle cpu is available. The span of cpus to
1449 * search starts with cpus closest then further out as needed,
1450 * so we always favor a closer, idle cpu.
1452 * Returns the CPU we should wake onto.
1454 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1455 static int wake_idle(int cpu
, struct task_struct
*p
)
1458 struct sched_domain
*sd
;
1462 * If it is idle, then it is the best cpu to run this task.
1464 * This cpu is also the best, if it has more than one task already.
1465 * Siblings must be also busy(in most cases) as they didn't already
1466 * pickup the extra load from this cpu and hence we need not check
1467 * sibling runqueue info. This will avoid the checks and cache miss
1468 * penalities associated with that.
1470 if (idle_cpu(cpu
) || cpu_rq(cpu
)->nr_running
> 1)
1473 for_each_domain(cpu
, sd
) {
1474 if (sd
->flags
& SD_WAKE_IDLE
) {
1475 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1476 for_each_cpu_mask(i
, tmp
) {
1487 static inline int wake_idle(int cpu
, struct task_struct
*p
)
1494 * try_to_wake_up - wake up a thread
1495 * @p: the to-be-woken-up thread
1496 * @state: the mask of task states that can be woken
1497 * @sync: do a synchronous wakeup?
1499 * Put it on the run-queue if it's not already there. The "current"
1500 * thread is always on the run-queue (except when the actual
1501 * re-schedule is in progress), and as such you're allowed to do
1502 * the simpler "current->state = TASK_RUNNING" to mark yourself
1503 * runnable without the overhead of this.
1505 * returns failure only if the task is already active.
1507 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1509 int cpu
, this_cpu
, success
= 0;
1510 unsigned long flags
;
1514 struct sched_domain
*sd
, *this_sd
= NULL
;
1515 unsigned long load
, this_load
;
1519 rq
= task_rq_lock(p
, &flags
);
1520 old_state
= p
->state
;
1521 if (!(old_state
& state
))
1528 this_cpu
= smp_processor_id();
1531 if (unlikely(task_running(rq
, p
)))
1536 schedstat_inc(rq
, ttwu_cnt
);
1537 if (cpu
== this_cpu
) {
1538 schedstat_inc(rq
, ttwu_local
);
1542 for_each_domain(this_cpu
, sd
) {
1543 if (cpu_isset(cpu
, sd
->span
)) {
1544 schedstat_inc(sd
, ttwu_wake_remote
);
1550 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1554 * Check for affine wakeup and passive balancing possibilities.
1557 int idx
= this_sd
->wake_idx
;
1558 unsigned int imbalance
;
1560 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1562 load
= source_load(cpu
, idx
);
1563 this_load
= target_load(this_cpu
, idx
);
1565 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1567 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1568 unsigned long tl
= this_load
;
1569 unsigned long tl_per_task
;
1571 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1574 * If sync wakeup then subtract the (maximum possible)
1575 * effect of the currently running task from the load
1576 * of the current CPU:
1579 tl
-= current
->load_weight
;
1582 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1583 100*(tl
+ p
->load_weight
) <= imbalance
*load
) {
1585 * This domain has SD_WAKE_AFFINE and
1586 * p is cache cold in this domain, and
1587 * there is no bad imbalance.
1589 schedstat_inc(this_sd
, ttwu_move_affine
);
1595 * Start passive balancing when half the imbalance_pct
1598 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1599 if (imbalance
*this_load
<= 100*load
) {
1600 schedstat_inc(this_sd
, ttwu_move_balance
);
1606 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1608 new_cpu
= wake_idle(new_cpu
, p
);
1609 if (new_cpu
!= cpu
) {
1610 set_task_cpu(p
, new_cpu
);
1611 task_rq_unlock(rq
, &flags
);
1612 /* might preempt at this point */
1613 rq
= task_rq_lock(p
, &flags
);
1614 old_state
= p
->state
;
1615 if (!(old_state
& state
))
1620 this_cpu
= smp_processor_id();
1625 #endif /* CONFIG_SMP */
1626 if (old_state
== TASK_UNINTERRUPTIBLE
) {
1627 rq
->nr_uninterruptible
--;
1629 * Tasks on involuntary sleep don't earn
1630 * sleep_avg beyond just interactive state.
1632 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
1636 * Tasks that have marked their sleep as noninteractive get
1637 * woken up with their sleep average not weighted in an
1640 if (old_state
& TASK_NONINTERACTIVE
)
1641 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
1644 activate_task(p
, rq
, cpu
== this_cpu
);
1646 * Sync wakeups (i.e. those types of wakeups where the waker
1647 * has indicated that it will leave the CPU in short order)
1648 * don't trigger a preemption, if the woken up task will run on
1649 * this cpu. (in this case the 'I will reschedule' promise of
1650 * the waker guarantees that the freshly woken up task is going
1651 * to be considered on this CPU.)
1653 if (!sync
|| cpu
!= this_cpu
) {
1654 if (TASK_PREEMPTS_CURR(p
, rq
))
1655 resched_task(rq
->curr
);
1660 p
->state
= TASK_RUNNING
;
1662 task_rq_unlock(rq
, &flags
);
1667 int fastcall
wake_up_process(struct task_struct
*p
)
1669 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1670 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1672 EXPORT_SYMBOL(wake_up_process
);
1674 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1676 return try_to_wake_up(p
, state
, 0);
1679 static void task_running_tick(struct rq
*rq
, struct task_struct
*p
);
1681 * Perform scheduler related setup for a newly forked process p.
1682 * p is forked by current.
1684 void fastcall
sched_fork(struct task_struct
*p
, int clone_flags
)
1686 int cpu
= get_cpu();
1689 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1691 set_task_cpu(p
, cpu
);
1694 * We mark the process as running here, but have not actually
1695 * inserted it onto the runqueue yet. This guarantees that
1696 * nobody will actually run it, and a signal or other external
1697 * event cannot wake it up and insert it on the runqueue either.
1699 p
->state
= TASK_RUNNING
;
1702 * Make sure we do not leak PI boosting priority to the child:
1704 p
->prio
= current
->normal_prio
;
1706 INIT_LIST_HEAD(&p
->run_list
);
1708 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1709 if (unlikely(sched_info_on()))
1710 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1712 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1715 #ifdef CONFIG_PREEMPT
1716 /* Want to start with kernel preemption disabled. */
1717 task_thread_info(p
)->preempt_count
= 1;
1720 * Share the timeslice between parent and child, thus the
1721 * total amount of pending timeslices in the system doesn't change,
1722 * resulting in more scheduling fairness.
1724 local_irq_disable();
1725 p
->time_slice
= (current
->time_slice
+ 1) >> 1;
1727 * The remainder of the first timeslice might be recovered by
1728 * the parent if the child exits early enough.
1730 p
->first_time_slice
= 1;
1731 current
->time_slice
>>= 1;
1732 p
->timestamp
= sched_clock();
1733 if (unlikely(!current
->time_slice
)) {
1735 * This case is rare, it happens when the parent has only
1736 * a single jiffy left from its timeslice. Taking the
1737 * runqueue lock is not a problem.
1739 current
->time_slice
= 1;
1740 task_running_tick(cpu_rq(cpu
), current
);
1747 * wake_up_new_task - wake up a newly created task for the first time.
1749 * This function will do some initial scheduler statistics housekeeping
1750 * that must be done for every newly created context, then puts the task
1751 * on the runqueue and wakes it.
1753 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1755 struct rq
*rq
, *this_rq
;
1756 unsigned long flags
;
1759 rq
= task_rq_lock(p
, &flags
);
1760 BUG_ON(p
->state
!= TASK_RUNNING
);
1761 this_cpu
= smp_processor_id();
1765 * We decrease the sleep average of forking parents
1766 * and children as well, to keep max-interactive tasks
1767 * from forking tasks that are max-interactive. The parent
1768 * (current) is done further down, under its lock.
1770 p
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(p
) *
1771 CHILD_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1773 p
->prio
= effective_prio(p
);
1775 if (likely(cpu
== this_cpu
)) {
1776 if (!(clone_flags
& CLONE_VM
)) {
1778 * The VM isn't cloned, so we're in a good position to
1779 * do child-runs-first in anticipation of an exec. This
1780 * usually avoids a lot of COW overhead.
1782 if (unlikely(!current
->array
))
1783 __activate_task(p
, rq
);
1785 p
->prio
= current
->prio
;
1786 p
->normal_prio
= current
->normal_prio
;
1787 list_add_tail(&p
->run_list
, ¤t
->run_list
);
1788 p
->array
= current
->array
;
1789 p
->array
->nr_active
++;
1790 inc_nr_running(p
, rq
);
1794 /* Run child last */
1795 __activate_task(p
, rq
);
1797 * We skip the following code due to cpu == this_cpu
1799 * task_rq_unlock(rq, &flags);
1800 * this_rq = task_rq_lock(current, &flags);
1804 this_rq
= cpu_rq(this_cpu
);
1807 * Not the local CPU - must adjust timestamp. This should
1808 * get optimised away in the !CONFIG_SMP case.
1810 p
->timestamp
= (p
->timestamp
- this_rq
->most_recent_timestamp
)
1811 + rq
->most_recent_timestamp
;
1812 __activate_task(p
, rq
);
1813 if (TASK_PREEMPTS_CURR(p
, rq
))
1814 resched_task(rq
->curr
);
1817 * Parent and child are on different CPUs, now get the
1818 * parent runqueue to update the parent's ->sleep_avg:
1820 task_rq_unlock(rq
, &flags
);
1821 this_rq
= task_rq_lock(current
, &flags
);
1823 current
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(current
) *
1824 PARENT_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1825 task_rq_unlock(this_rq
, &flags
);
1829 * Potentially available exiting-child timeslices are
1830 * retrieved here - this way the parent does not get
1831 * penalized for creating too many threads.
1833 * (this cannot be used to 'generate' timeslices
1834 * artificially, because any timeslice recovered here
1835 * was given away by the parent in the first place.)
1837 void fastcall
sched_exit(struct task_struct
*p
)
1839 unsigned long flags
;
1843 * If the child was a (relative-) CPU hog then decrease
1844 * the sleep_avg of the parent as well.
1846 rq
= task_rq_lock(p
->parent
, &flags
);
1847 if (p
->first_time_slice
&& task_cpu(p
) == task_cpu(p
->parent
)) {
1848 p
->parent
->time_slice
+= p
->time_slice
;
1849 if (unlikely(p
->parent
->time_slice
> task_timeslice(p
)))
1850 p
->parent
->time_slice
= task_timeslice(p
);
1852 if (p
->sleep_avg
< p
->parent
->sleep_avg
)
1853 p
->parent
->sleep_avg
= p
->parent
->sleep_avg
/
1854 (EXIT_WEIGHT
+ 1) * EXIT_WEIGHT
+ p
->sleep_avg
/
1856 task_rq_unlock(rq
, &flags
);
1860 * prepare_task_switch - prepare to switch tasks
1861 * @rq: the runqueue preparing to switch
1862 * @next: the task we are going to switch to.
1864 * This is called with the rq lock held and interrupts off. It must
1865 * be paired with a subsequent finish_task_switch after the context
1868 * prepare_task_switch sets up locking and calls architecture specific
1871 static inline void prepare_task_switch(struct rq
*rq
, struct task_struct
*next
)
1873 prepare_lock_switch(rq
, next
);
1874 prepare_arch_switch(next
);
1878 * finish_task_switch - clean up after a task-switch
1879 * @rq: runqueue associated with task-switch
1880 * @prev: the thread we just switched away from.
1882 * finish_task_switch must be called after the context switch, paired
1883 * with a prepare_task_switch call before the context switch.
1884 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1885 * and do any other architecture-specific cleanup actions.
1887 * Note that we may have delayed dropping an mm in context_switch(). If
1888 * so, we finish that here outside of the runqueue lock. (Doing it
1889 * with the lock held can cause deadlocks; see schedule() for
1892 static inline void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1893 __releases(rq
->lock
)
1895 struct mm_struct
*mm
= rq
->prev_mm
;
1901 * A task struct has one reference for the use as "current".
1902 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1903 * schedule one last time. The schedule call will never return, and
1904 * the scheduled task must drop that reference.
1905 * The test for TASK_DEAD must occur while the runqueue locks are
1906 * still held, otherwise prev could be scheduled on another cpu, die
1907 * there before we look at prev->state, and then the reference would
1909 * Manfred Spraul <manfred@colorfullife.com>
1911 prev_state
= prev
->state
;
1912 finish_arch_switch(prev
);
1913 finish_lock_switch(rq
, prev
);
1916 if (unlikely(prev_state
== TASK_DEAD
)) {
1918 * Remove function-return probe instances associated with this
1919 * task and put them back on the free list.
1921 kprobe_flush_task(prev
);
1922 put_task_struct(prev
);
1927 * schedule_tail - first thing a freshly forked thread must call.
1928 * @prev: the thread we just switched away from.
1930 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1931 __releases(rq
->lock
)
1933 struct rq
*rq
= this_rq();
1935 finish_task_switch(rq
, prev
);
1936 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1937 /* In this case, finish_task_switch does not reenable preemption */
1940 if (current
->set_child_tid
)
1941 put_user(current
->pid
, current
->set_child_tid
);
1945 * context_switch - switch to the new MM and the new
1946 * thread's register state.
1948 static inline struct task_struct
*
1949 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1950 struct task_struct
*next
)
1952 struct mm_struct
*mm
= next
->mm
;
1953 struct mm_struct
*oldmm
= prev
->active_mm
;
1956 * For paravirt, this is coupled with an exit in switch_to to
1957 * combine the page table reload and the switch backend into
1960 arch_enter_lazy_cpu_mode();
1963 next
->active_mm
= oldmm
;
1964 atomic_inc(&oldmm
->mm_count
);
1965 enter_lazy_tlb(oldmm
, next
);
1967 switch_mm(oldmm
, mm
, next
);
1970 prev
->active_mm
= NULL
;
1971 WARN_ON(rq
->prev_mm
);
1972 rq
->prev_mm
= oldmm
;
1975 * Since the runqueue lock will be released by the next
1976 * task (which is an invalid locking op but in the case
1977 * of the scheduler it's an obvious special-case), so we
1978 * do an early lockdep release here:
1980 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1981 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1984 /* Here we just switch the register state and the stack. */
1985 switch_to(prev
, next
, prev
);
1991 * nr_running, nr_uninterruptible and nr_context_switches:
1993 * externally visible scheduler statistics: current number of runnable
1994 * threads, current number of uninterruptible-sleeping threads, total
1995 * number of context switches performed since bootup.
1997 unsigned long nr_running(void)
1999 unsigned long i
, sum
= 0;
2001 for_each_online_cpu(i
)
2002 sum
+= cpu_rq(i
)->nr_running
;
2007 unsigned long nr_uninterruptible(void)
2009 unsigned long i
, sum
= 0;
2011 for_each_possible_cpu(i
)
2012 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2015 * Since we read the counters lockless, it might be slightly
2016 * inaccurate. Do not allow it to go below zero though:
2018 if (unlikely((long)sum
< 0))
2024 unsigned long long nr_context_switches(void)
2027 unsigned long long sum
= 0;
2029 for_each_possible_cpu(i
)
2030 sum
+= cpu_rq(i
)->nr_switches
;
2035 unsigned long nr_iowait(void)
2037 unsigned long i
, sum
= 0;
2039 for_each_possible_cpu(i
)
2040 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2045 unsigned long nr_active(void)
2047 unsigned long i
, running
= 0, uninterruptible
= 0;
2049 for_each_online_cpu(i
) {
2050 running
+= cpu_rq(i
)->nr_running
;
2051 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2054 if (unlikely((long)uninterruptible
< 0))
2055 uninterruptible
= 0;
2057 return running
+ uninterruptible
;
2063 * Is this task likely cache-hot:
2066 task_hot(struct task_struct
*p
, unsigned long long now
, struct sched_domain
*sd
)
2068 return (long long)(now
- p
->last_ran
) < (long long)sd
->cache_hot_time
;
2072 * double_rq_lock - safely lock two runqueues
2074 * Note this does not disable interrupts like task_rq_lock,
2075 * you need to do so manually before calling.
2077 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2078 __acquires(rq1
->lock
)
2079 __acquires(rq2
->lock
)
2081 BUG_ON(!irqs_disabled());
2083 spin_lock(&rq1
->lock
);
2084 __acquire(rq2
->lock
); /* Fake it out ;) */
2087 spin_lock(&rq1
->lock
);
2088 spin_lock(&rq2
->lock
);
2090 spin_lock(&rq2
->lock
);
2091 spin_lock(&rq1
->lock
);
2097 * double_rq_unlock - safely unlock two runqueues
2099 * Note this does not restore interrupts like task_rq_unlock,
2100 * you need to do so manually after calling.
2102 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2103 __releases(rq1
->lock
)
2104 __releases(rq2
->lock
)
2106 spin_unlock(&rq1
->lock
);
2108 spin_unlock(&rq2
->lock
);
2110 __release(rq2
->lock
);
2114 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2116 static void double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2117 __releases(this_rq
->lock
)
2118 __acquires(busiest
->lock
)
2119 __acquires(this_rq
->lock
)
2121 if (unlikely(!irqs_disabled())) {
2122 /* printk() doesn't work good under rq->lock */
2123 spin_unlock(&this_rq
->lock
);
2126 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2127 if (busiest
< this_rq
) {
2128 spin_unlock(&this_rq
->lock
);
2129 spin_lock(&busiest
->lock
);
2130 spin_lock(&this_rq
->lock
);
2132 spin_lock(&busiest
->lock
);
2137 * If dest_cpu is allowed for this process, migrate the task to it.
2138 * This is accomplished by forcing the cpu_allowed mask to only
2139 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2140 * the cpu_allowed mask is restored.
2142 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2144 struct migration_req req
;
2145 unsigned long flags
;
2148 rq
= task_rq_lock(p
, &flags
);
2149 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2150 || unlikely(cpu_is_offline(dest_cpu
)))
2153 /* force the process onto the specified CPU */
2154 if (migrate_task(p
, dest_cpu
, &req
)) {
2155 /* Need to wait for migration thread (might exit: take ref). */
2156 struct task_struct
*mt
= rq
->migration_thread
;
2158 get_task_struct(mt
);
2159 task_rq_unlock(rq
, &flags
);
2160 wake_up_process(mt
);
2161 put_task_struct(mt
);
2162 wait_for_completion(&req
.done
);
2167 task_rq_unlock(rq
, &flags
);
2171 * sched_exec - execve() is a valuable balancing opportunity, because at
2172 * this point the task has the smallest effective memory and cache footprint.
2174 void sched_exec(void)
2176 int new_cpu
, this_cpu
= get_cpu();
2177 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2179 if (new_cpu
!= this_cpu
)
2180 sched_migrate_task(current
, new_cpu
);
2184 * pull_task - move a task from a remote runqueue to the local runqueue.
2185 * Both runqueues must be locked.
2187 static void pull_task(struct rq
*src_rq
, struct prio_array
*src_array
,
2188 struct task_struct
*p
, struct rq
*this_rq
,
2189 struct prio_array
*this_array
, int this_cpu
)
2191 dequeue_task(p
, src_array
);
2192 dec_nr_running(p
, src_rq
);
2193 set_task_cpu(p
, this_cpu
);
2194 inc_nr_running(p
, this_rq
);
2195 enqueue_task(p
, this_array
);
2196 p
->timestamp
= (p
->timestamp
- src_rq
->most_recent_timestamp
)
2197 + this_rq
->most_recent_timestamp
;
2199 * Note that idle threads have a prio of MAX_PRIO, for this test
2200 * to be always true for them.
2202 if (TASK_PREEMPTS_CURR(p
, this_rq
))
2203 resched_task(this_rq
->curr
);
2207 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2210 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2211 struct sched_domain
*sd
, enum idle_type idle
,
2215 * We do not migrate tasks that are:
2216 * 1) running (obviously), or
2217 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2218 * 3) are cache-hot on their current CPU.
2220 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
2224 if (task_running(rq
, p
))
2228 * Aggressive migration if:
2229 * 1) task is cache cold, or
2230 * 2) too many balance attempts have failed.
2233 if (sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2234 #ifdef CONFIG_SCHEDSTATS
2235 if (task_hot(p
, rq
->most_recent_timestamp
, sd
))
2236 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2241 if (task_hot(p
, rq
->most_recent_timestamp
, sd
))
2246 #define rq_best_prio(rq) min((rq)->curr->prio, (rq)->best_expired_prio)
2249 * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted
2250 * load from busiest to this_rq, as part of a balancing operation within
2251 * "domain". Returns the number of tasks moved.
2253 * Called with both runqueues locked.
2255 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2256 unsigned long max_nr_move
, unsigned long max_load_move
,
2257 struct sched_domain
*sd
, enum idle_type idle
,
2260 int idx
, pulled
= 0, pinned
= 0, this_best_prio
, best_prio
,
2261 best_prio_seen
, skip_for_load
;
2262 struct prio_array
*array
, *dst_array
;
2263 struct list_head
*head
, *curr
;
2264 struct task_struct
*tmp
;
2267 if (max_nr_move
== 0 || max_load_move
== 0)
2270 rem_load_move
= max_load_move
;
2272 this_best_prio
= rq_best_prio(this_rq
);
2273 best_prio
= rq_best_prio(busiest
);
2275 * Enable handling of the case where there is more than one task
2276 * with the best priority. If the current running task is one
2277 * of those with prio==best_prio we know it won't be moved
2278 * and therefore it's safe to override the skip (based on load) of
2279 * any task we find with that prio.
2281 best_prio_seen
= best_prio
== busiest
->curr
->prio
;
2284 * We first consider expired tasks. Those will likely not be
2285 * executed in the near future, and they are most likely to
2286 * be cache-cold, thus switching CPUs has the least effect
2289 if (busiest
->expired
->nr_active
) {
2290 array
= busiest
->expired
;
2291 dst_array
= this_rq
->expired
;
2293 array
= busiest
->active
;
2294 dst_array
= this_rq
->active
;
2298 /* Start searching at priority 0: */
2302 idx
= sched_find_first_bit(array
->bitmap
);
2304 idx
= find_next_bit(array
->bitmap
, MAX_PRIO
, idx
);
2305 if (idx
>= MAX_PRIO
) {
2306 if (array
== busiest
->expired
&& busiest
->active
->nr_active
) {
2307 array
= busiest
->active
;
2308 dst_array
= this_rq
->active
;
2314 head
= array
->queue
+ idx
;
2317 tmp
= list_entry(curr
, struct task_struct
, run_list
);
2322 * To help distribute high priority tasks accross CPUs we don't
2323 * skip a task if it will be the highest priority task (i.e. smallest
2324 * prio value) on its new queue regardless of its load weight
2326 skip_for_load
= tmp
->load_weight
> rem_load_move
;
2327 if (skip_for_load
&& idx
< this_best_prio
)
2328 skip_for_load
= !best_prio_seen
&& idx
== best_prio
;
2329 if (skip_for_load
||
2330 !can_migrate_task(tmp
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2332 best_prio_seen
|= idx
== best_prio
;
2339 pull_task(busiest
, array
, tmp
, this_rq
, dst_array
, this_cpu
);
2341 rem_load_move
-= tmp
->load_weight
;
2344 * We only want to steal up to the prescribed number of tasks
2345 * and the prescribed amount of weighted load.
2347 if (pulled
< max_nr_move
&& rem_load_move
> 0) {
2348 if (idx
< this_best_prio
)
2349 this_best_prio
= idx
;
2357 * Right now, this is the only place pull_task() is called,
2358 * so we can safely collect pull_task() stats here rather than
2359 * inside pull_task().
2361 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2364 *all_pinned
= pinned
;
2369 * find_busiest_group finds and returns the busiest CPU group within the
2370 * domain. It calculates and returns the amount of weighted load which
2371 * should be moved to restore balance via the imbalance parameter.
2373 static struct sched_group
*
2374 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2375 unsigned long *imbalance
, enum idle_type idle
, int *sd_idle
,
2376 cpumask_t
*cpus
, int *balance
)
2378 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2379 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2380 unsigned long max_pull
;
2381 unsigned long busiest_load_per_task
, busiest_nr_running
;
2382 unsigned long this_load_per_task
, this_nr_running
;
2384 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2385 int power_savings_balance
= 1;
2386 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2387 unsigned long min_nr_running
= ULONG_MAX
;
2388 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2391 max_load
= this_load
= total_load
= total_pwr
= 0;
2392 busiest_load_per_task
= busiest_nr_running
= 0;
2393 this_load_per_task
= this_nr_running
= 0;
2394 if (idle
== NOT_IDLE
)
2395 load_idx
= sd
->busy_idx
;
2396 else if (idle
== NEWLY_IDLE
)
2397 load_idx
= sd
->newidle_idx
;
2399 load_idx
= sd
->idle_idx
;
2402 unsigned long load
, group_capacity
;
2405 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2406 unsigned long sum_nr_running
, sum_weighted_load
;
2408 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2411 balance_cpu
= first_cpu(group
->cpumask
);
2413 /* Tally up the load of all CPUs in the group */
2414 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2416 for_each_cpu_mask(i
, group
->cpumask
) {
2419 if (!cpu_isset(i
, *cpus
))
2424 if (*sd_idle
&& !idle_cpu(i
))
2427 /* Bias balancing toward cpus of our domain */
2429 if (idle_cpu(i
) && !first_idle_cpu
) {
2434 load
= target_load(i
, load_idx
);
2436 load
= source_load(i
, load_idx
);
2439 sum_nr_running
+= rq
->nr_running
;
2440 sum_weighted_load
+= rq
->raw_weighted_load
;
2444 * First idle cpu or the first cpu(busiest) in this sched group
2445 * is eligible for doing load balancing at this and above
2448 if (local_group
&& balance_cpu
!= this_cpu
&& balance
) {
2453 total_load
+= avg_load
;
2454 total_pwr
+= group
->__cpu_power
;
2456 /* Adjust by relative CPU power of the group */
2457 avg_load
= sg_div_cpu_power(group
,
2458 avg_load
* SCHED_LOAD_SCALE
);
2460 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
2463 this_load
= avg_load
;
2465 this_nr_running
= sum_nr_running
;
2466 this_load_per_task
= sum_weighted_load
;
2467 } else if (avg_load
> max_load
&&
2468 sum_nr_running
> group_capacity
) {
2469 max_load
= avg_load
;
2471 busiest_nr_running
= sum_nr_running
;
2472 busiest_load_per_task
= sum_weighted_load
;
2475 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2477 * Busy processors will not participate in power savings
2480 if (idle
== NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2484 * If the local group is idle or completely loaded
2485 * no need to do power savings balance at this domain
2487 if (local_group
&& (this_nr_running
>= group_capacity
||
2489 power_savings_balance
= 0;
2492 * If a group is already running at full capacity or idle,
2493 * don't include that group in power savings calculations
2495 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2500 * Calculate the group which has the least non-idle load.
2501 * This is the group from where we need to pick up the load
2504 if ((sum_nr_running
< min_nr_running
) ||
2505 (sum_nr_running
== min_nr_running
&&
2506 first_cpu(group
->cpumask
) <
2507 first_cpu(group_min
->cpumask
))) {
2509 min_nr_running
= sum_nr_running
;
2510 min_load_per_task
= sum_weighted_load
/
2515 * Calculate the group which is almost near its
2516 * capacity but still has some space to pick up some load
2517 * from other group and save more power
2519 if (sum_nr_running
<= group_capacity
- 1) {
2520 if (sum_nr_running
> leader_nr_running
||
2521 (sum_nr_running
== leader_nr_running
&&
2522 first_cpu(group
->cpumask
) >
2523 first_cpu(group_leader
->cpumask
))) {
2524 group_leader
= group
;
2525 leader_nr_running
= sum_nr_running
;
2530 group
= group
->next
;
2531 } while (group
!= sd
->groups
);
2533 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2536 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2538 if (this_load
>= avg_load
||
2539 100*max_load
<= sd
->imbalance_pct
*this_load
)
2542 busiest_load_per_task
/= busiest_nr_running
;
2544 * We're trying to get all the cpus to the average_load, so we don't
2545 * want to push ourselves above the average load, nor do we wish to
2546 * reduce the max loaded cpu below the average load, as either of these
2547 * actions would just result in more rebalancing later, and ping-pong
2548 * tasks around. Thus we look for the minimum possible imbalance.
2549 * Negative imbalances (*we* are more loaded than anyone else) will
2550 * be counted as no imbalance for these purposes -- we can't fix that
2551 * by pulling tasks to us. Be careful of negative numbers as they'll
2552 * appear as very large values with unsigned longs.
2554 if (max_load
<= busiest_load_per_task
)
2558 * In the presence of smp nice balancing, certain scenarios can have
2559 * max load less than avg load(as we skip the groups at or below
2560 * its cpu_power, while calculating max_load..)
2562 if (max_load
< avg_load
) {
2564 goto small_imbalance
;
2567 /* Don't want to pull so many tasks that a group would go idle */
2568 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2570 /* How much load to actually move to equalise the imbalance */
2571 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
2572 (avg_load
- this_load
) * this->__cpu_power
)
2576 * if *imbalance is less than the average load per runnable task
2577 * there is no gaurantee that any tasks will be moved so we'll have
2578 * a think about bumping its value to force at least one task to be
2581 if (*imbalance
< busiest_load_per_task
) {
2582 unsigned long tmp
, pwr_now
, pwr_move
;
2586 pwr_move
= pwr_now
= 0;
2588 if (this_nr_running
) {
2589 this_load_per_task
/= this_nr_running
;
2590 if (busiest_load_per_task
> this_load_per_task
)
2593 this_load_per_task
= SCHED_LOAD_SCALE
;
2595 if (max_load
- this_load
>= busiest_load_per_task
* imbn
) {
2596 *imbalance
= busiest_load_per_task
;
2601 * OK, we don't have enough imbalance to justify moving tasks,
2602 * however we may be able to increase total CPU power used by
2606 pwr_now
+= busiest
->__cpu_power
*
2607 min(busiest_load_per_task
, max_load
);
2608 pwr_now
+= this->__cpu_power
*
2609 min(this_load_per_task
, this_load
);
2610 pwr_now
/= SCHED_LOAD_SCALE
;
2612 /* Amount of load we'd subtract */
2613 tmp
= sg_div_cpu_power(busiest
,
2614 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2616 pwr_move
+= busiest
->__cpu_power
*
2617 min(busiest_load_per_task
, max_load
- tmp
);
2619 /* Amount of load we'd add */
2620 if (max_load
* busiest
->__cpu_power
<
2621 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2622 tmp
= sg_div_cpu_power(this,
2623 max_load
* busiest
->__cpu_power
);
2625 tmp
= sg_div_cpu_power(this,
2626 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2627 pwr_move
+= this->__cpu_power
*
2628 min(this_load_per_task
, this_load
+ tmp
);
2629 pwr_move
/= SCHED_LOAD_SCALE
;
2631 /* Move if we gain throughput */
2632 if (pwr_move
<= pwr_now
)
2635 *imbalance
= busiest_load_per_task
;
2641 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2642 if (idle
== NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2645 if (this == group_leader
&& group_leader
!= group_min
) {
2646 *imbalance
= min_load_per_task
;
2656 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2659 find_busiest_queue(struct sched_group
*group
, enum idle_type idle
,
2660 unsigned long imbalance
, cpumask_t
*cpus
)
2662 struct rq
*busiest
= NULL
, *rq
;
2663 unsigned long max_load
= 0;
2666 for_each_cpu_mask(i
, group
->cpumask
) {
2668 if (!cpu_isset(i
, *cpus
))
2673 if (rq
->nr_running
== 1 && rq
->raw_weighted_load
> imbalance
)
2676 if (rq
->raw_weighted_load
> max_load
) {
2677 max_load
= rq
->raw_weighted_load
;
2686 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2687 * so long as it is large enough.
2689 #define MAX_PINNED_INTERVAL 512
2691 static inline unsigned long minus_1_or_zero(unsigned long n
)
2693 return n
> 0 ? n
- 1 : 0;
2697 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2698 * tasks if there is an imbalance.
2700 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2701 struct sched_domain
*sd
, enum idle_type idle
,
2704 int nr_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2705 struct sched_group
*group
;
2706 unsigned long imbalance
;
2708 cpumask_t cpus
= CPU_MASK_ALL
;
2709 unsigned long flags
;
2712 * When power savings policy is enabled for the parent domain, idle
2713 * sibling can pick up load irrespective of busy siblings. In this case,
2714 * let the state of idle sibling percolate up as IDLE, instead of
2715 * portraying it as NOT_IDLE.
2717 if (idle
!= NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2718 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2721 schedstat_inc(sd
, lb_cnt
[idle
]);
2724 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2731 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2735 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2737 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2741 BUG_ON(busiest
== this_rq
);
2743 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2746 if (busiest
->nr_running
> 1) {
2748 * Attempt to move tasks. If find_busiest_group has found
2749 * an imbalance but busiest->nr_running <= 1, the group is
2750 * still unbalanced. nr_moved simply stays zero, so it is
2751 * correctly treated as an imbalance.
2753 local_irq_save(flags
);
2754 double_rq_lock(this_rq
, busiest
);
2755 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2756 minus_1_or_zero(busiest
->nr_running
),
2757 imbalance
, sd
, idle
, &all_pinned
);
2758 double_rq_unlock(this_rq
, busiest
);
2759 local_irq_restore(flags
);
2762 * some other cpu did the load balance for us.
2764 if (nr_moved
&& this_cpu
!= smp_processor_id())
2765 resched_cpu(this_cpu
);
2767 /* All tasks on this runqueue were pinned by CPU affinity */
2768 if (unlikely(all_pinned
)) {
2769 cpu_clear(cpu_of(busiest
), cpus
);
2770 if (!cpus_empty(cpus
))
2777 schedstat_inc(sd
, lb_failed
[idle
]);
2778 sd
->nr_balance_failed
++;
2780 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2782 spin_lock_irqsave(&busiest
->lock
, flags
);
2784 /* don't kick the migration_thread, if the curr
2785 * task on busiest cpu can't be moved to this_cpu
2787 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2788 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2790 goto out_one_pinned
;
2793 if (!busiest
->active_balance
) {
2794 busiest
->active_balance
= 1;
2795 busiest
->push_cpu
= this_cpu
;
2798 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2800 wake_up_process(busiest
->migration_thread
);
2803 * We've kicked active balancing, reset the failure
2806 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2809 sd
->nr_balance_failed
= 0;
2811 if (likely(!active_balance
)) {
2812 /* We were unbalanced, so reset the balancing interval */
2813 sd
->balance_interval
= sd
->min_interval
;
2816 * If we've begun active balancing, start to back off. This
2817 * case may not be covered by the all_pinned logic if there
2818 * is only 1 task on the busy runqueue (because we don't call
2821 if (sd
->balance_interval
< sd
->max_interval
)
2822 sd
->balance_interval
*= 2;
2825 if (!nr_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2826 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2831 schedstat_inc(sd
, lb_balanced
[idle
]);
2833 sd
->nr_balance_failed
= 0;
2836 /* tune up the balancing interval */
2837 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2838 (sd
->balance_interval
< sd
->max_interval
))
2839 sd
->balance_interval
*= 2;
2841 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2842 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2848 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2849 * tasks if there is an imbalance.
2851 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2852 * this_rq is locked.
2855 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
2857 struct sched_group
*group
;
2858 struct rq
*busiest
= NULL
;
2859 unsigned long imbalance
;
2862 cpumask_t cpus
= CPU_MASK_ALL
;
2865 * When power savings policy is enabled for the parent domain, idle
2866 * sibling can pick up load irrespective of busy siblings. In this case,
2867 * let the state of idle sibling percolate up as IDLE, instead of
2868 * portraying it as NOT_IDLE.
2870 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
2871 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2874 schedstat_inc(sd
, lb_cnt
[NEWLY_IDLE
]);
2876 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, NEWLY_IDLE
,
2877 &sd_idle
, &cpus
, NULL
);
2879 schedstat_inc(sd
, lb_nobusyg
[NEWLY_IDLE
]);
2883 busiest
= find_busiest_queue(group
, NEWLY_IDLE
, imbalance
,
2886 schedstat_inc(sd
, lb_nobusyq
[NEWLY_IDLE
]);
2890 BUG_ON(busiest
== this_rq
);
2892 schedstat_add(sd
, lb_imbalance
[NEWLY_IDLE
], imbalance
);
2895 if (busiest
->nr_running
> 1) {
2896 /* Attempt to move tasks */
2897 double_lock_balance(this_rq
, busiest
);
2898 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2899 minus_1_or_zero(busiest
->nr_running
),
2900 imbalance
, sd
, NEWLY_IDLE
, NULL
);
2901 spin_unlock(&busiest
->lock
);
2904 cpu_clear(cpu_of(busiest
), cpus
);
2905 if (!cpus_empty(cpus
))
2911 schedstat_inc(sd
, lb_failed
[NEWLY_IDLE
]);
2912 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2913 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2916 sd
->nr_balance_failed
= 0;
2921 schedstat_inc(sd
, lb_balanced
[NEWLY_IDLE
]);
2922 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2923 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2925 sd
->nr_balance_failed
= 0;
2931 * idle_balance is called by schedule() if this_cpu is about to become
2932 * idle. Attempts to pull tasks from other CPUs.
2934 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
2936 struct sched_domain
*sd
;
2937 int pulled_task
= 0;
2938 unsigned long next_balance
= jiffies
+ 60 * HZ
;
2940 for_each_domain(this_cpu
, sd
) {
2941 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
2942 /* If we've pulled tasks over stop searching: */
2943 pulled_task
= load_balance_newidle(this_cpu
,
2945 if (time_after(next_balance
,
2946 sd
->last_balance
+ sd
->balance_interval
))
2947 next_balance
= sd
->last_balance
2948 + sd
->balance_interval
;
2955 * We are going idle. next_balance may be set based on
2956 * a busy processor. So reset next_balance.
2958 this_rq
->next_balance
= next_balance
;
2962 * active_load_balance is run by migration threads. It pushes running tasks
2963 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2964 * running on each physical CPU where possible, and avoids physical /
2965 * logical imbalances.
2967 * Called with busiest_rq locked.
2969 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
2971 int target_cpu
= busiest_rq
->push_cpu
;
2972 struct sched_domain
*sd
;
2973 struct rq
*target_rq
;
2975 /* Is there any task to move? */
2976 if (busiest_rq
->nr_running
<= 1)
2979 target_rq
= cpu_rq(target_cpu
);
2982 * This condition is "impossible", if it occurs
2983 * we need to fix it. Originally reported by
2984 * Bjorn Helgaas on a 128-cpu setup.
2986 BUG_ON(busiest_rq
== target_rq
);
2988 /* move a task from busiest_rq to target_rq */
2989 double_lock_balance(busiest_rq
, target_rq
);
2991 /* Search for an sd spanning us and the target CPU. */
2992 for_each_domain(target_cpu
, sd
) {
2993 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2994 cpu_isset(busiest_cpu
, sd
->span
))
2999 schedstat_inc(sd
, alb_cnt
);
3001 if (move_tasks(target_rq
, target_cpu
, busiest_rq
, 1,
3002 RTPRIO_TO_LOAD_WEIGHT(100), sd
, SCHED_IDLE
,
3004 schedstat_inc(sd
, alb_pushed
);
3006 schedstat_inc(sd
, alb_failed
);
3008 spin_unlock(&target_rq
->lock
);
3011 static void update_load(struct rq
*this_rq
)
3013 unsigned long this_load
;
3014 unsigned int i
, scale
;
3016 this_load
= this_rq
->raw_weighted_load
;
3018 /* Update our load: */
3019 for (i
= 0, scale
= 1; i
< 3; i
++, scale
+= scale
) {
3020 unsigned long old_load
, new_load
;
3022 /* scale is effectively 1 << i now, and >> i divides by scale */
3024 old_load
= this_rq
->cpu_load
[i
];
3025 new_load
= this_load
;
3027 * Round up the averaging division if load is increasing. This
3028 * prevents us from getting stuck on 9 if the load is 10, for
3031 if (new_load
> old_load
)
3032 new_load
+= scale
-1;
3033 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
3039 atomic_t load_balancer
;
3041 } nohz ____cacheline_aligned
= {
3042 .load_balancer
= ATOMIC_INIT(-1),
3043 .cpu_mask
= CPU_MASK_NONE
,
3047 * This routine will try to nominate the ilb (idle load balancing)
3048 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3049 * load balancing on behalf of all those cpus. If all the cpus in the system
3050 * go into this tickless mode, then there will be no ilb owner (as there is
3051 * no need for one) and all the cpus will sleep till the next wakeup event
3054 * For the ilb owner, tick is not stopped. And this tick will be used
3055 * for idle load balancing. ilb owner will still be part of
3058 * While stopping the tick, this cpu will become the ilb owner if there
3059 * is no other owner. And will be the owner till that cpu becomes busy
3060 * or if all cpus in the system stop their ticks at which point
3061 * there is no need for ilb owner.
3063 * When the ilb owner becomes busy, it nominates another owner, during the
3064 * next busy scheduler_tick()
3066 int select_nohz_load_balancer(int stop_tick
)
3068 int cpu
= smp_processor_id();
3071 cpu_set(cpu
, nohz
.cpu_mask
);
3072 cpu_rq(cpu
)->in_nohz_recently
= 1;
3075 * If we are going offline and still the leader, give up!
3077 if (cpu_is_offline(cpu
) &&
3078 atomic_read(&nohz
.load_balancer
) == cpu
) {
3079 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3084 /* time for ilb owner also to sleep */
3085 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3086 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3087 atomic_set(&nohz
.load_balancer
, -1);
3091 if (atomic_read(&nohz
.load_balancer
) == -1) {
3092 /* make me the ilb owner */
3093 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3095 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3098 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3101 cpu_clear(cpu
, nohz
.cpu_mask
);
3103 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3104 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3111 static DEFINE_SPINLOCK(balancing
);
3114 * It checks each scheduling domain to see if it is due to be balanced,
3115 * and initiates a balancing operation if so.
3117 * Balancing parameters are set up in arch_init_sched_domains.
3119 static inline void rebalance_domains(int cpu
, enum idle_type idle
)
3122 struct rq
*rq
= cpu_rq(cpu
);
3123 unsigned long interval
;
3124 struct sched_domain
*sd
;
3125 /* Earliest time when we have to do rebalance again */
3126 unsigned long next_balance
= jiffies
+ 60*HZ
;
3128 for_each_domain(cpu
, sd
) {
3129 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3132 interval
= sd
->balance_interval
;
3133 if (idle
!= SCHED_IDLE
)
3134 interval
*= sd
->busy_factor
;
3136 /* scale ms to jiffies */
3137 interval
= msecs_to_jiffies(interval
);
3138 if (unlikely(!interval
))
3141 if (sd
->flags
& SD_SERIALIZE
) {
3142 if (!spin_trylock(&balancing
))
3146 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3147 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3149 * We've pulled tasks over so either we're no
3150 * longer idle, or one of our SMT siblings is
3155 sd
->last_balance
= jiffies
;
3157 if (sd
->flags
& SD_SERIALIZE
)
3158 spin_unlock(&balancing
);
3160 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3161 next_balance
= sd
->last_balance
+ interval
;
3164 * Stop the load balance at this level. There is another
3165 * CPU in our sched group which is doing load balancing more
3171 rq
->next_balance
= next_balance
;
3175 * run_rebalance_domains is triggered when needed from the scheduler tick.
3176 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3177 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3179 static void run_rebalance_domains(struct softirq_action
*h
)
3181 int local_cpu
= smp_processor_id();
3182 struct rq
*local_rq
= cpu_rq(local_cpu
);
3183 enum idle_type idle
= local_rq
->idle_at_tick
? SCHED_IDLE
: NOT_IDLE
;
3185 rebalance_domains(local_cpu
, idle
);
3189 * If this cpu is the owner for idle load balancing, then do the
3190 * balancing on behalf of the other idle cpus whose ticks are
3193 if (local_rq
->idle_at_tick
&&
3194 atomic_read(&nohz
.load_balancer
) == local_cpu
) {
3195 cpumask_t cpus
= nohz
.cpu_mask
;
3199 cpu_clear(local_cpu
, cpus
);
3200 for_each_cpu_mask(balance_cpu
, cpus
) {
3202 * If this cpu gets work to do, stop the load balancing
3203 * work being done for other cpus. Next load
3204 * balancing owner will pick it up.
3209 rebalance_domains(balance_cpu
, SCHED_IDLE
);
3211 rq
= cpu_rq(balance_cpu
);
3212 if (time_after(local_rq
->next_balance
, rq
->next_balance
))
3213 local_rq
->next_balance
= rq
->next_balance
;
3220 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3222 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3223 * idle load balancing owner or decide to stop the periodic load balancing,
3224 * if the whole system is idle.
3226 static inline void trigger_load_balance(int cpu
)
3228 struct rq
*rq
= cpu_rq(cpu
);
3231 * If we were in the nohz mode recently and busy at the current
3232 * scheduler tick, then check if we need to nominate new idle
3235 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3236 rq
->in_nohz_recently
= 0;
3238 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3239 cpu_clear(cpu
, nohz
.cpu_mask
);
3240 atomic_set(&nohz
.load_balancer
, -1);
3243 if (atomic_read(&nohz
.load_balancer
) == -1) {
3245 * simple selection for now: Nominate the
3246 * first cpu in the nohz list to be the next
3249 * TBD: Traverse the sched domains and nominate
3250 * the nearest cpu in the nohz.cpu_mask.
3252 int ilb
= first_cpu(nohz
.cpu_mask
);
3260 * If this cpu is idle and doing idle load balancing for all the
3261 * cpus with ticks stopped, is it time for that to stop?
3263 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3264 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3270 * If this cpu is idle and the idle load balancing is done by
3271 * someone else, then no need raise the SCHED_SOFTIRQ
3273 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3274 cpu_isset(cpu
, nohz
.cpu_mask
))
3277 if (time_after_eq(jiffies
, rq
->next_balance
))
3278 raise_softirq(SCHED_SOFTIRQ
);
3282 * on UP we do not need to balance between CPUs:
3284 static inline void idle_balance(int cpu
, struct rq
*rq
)
3289 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3291 EXPORT_PER_CPU_SYMBOL(kstat
);
3294 * This is called on clock ticks and on context switches.
3295 * Bank in p->sched_time the ns elapsed since the last tick or switch.
3298 update_cpu_clock(struct task_struct
*p
, struct rq
*rq
, unsigned long long now
)
3300 p
->sched_time
+= now
- p
->last_ran
;
3301 p
->last_ran
= rq
->most_recent_timestamp
= now
;
3305 * Return current->sched_time plus any more ns on the sched_clock
3306 * that have not yet been banked.
3308 unsigned long long current_sched_time(const struct task_struct
*p
)
3310 unsigned long long ns
;
3311 unsigned long flags
;
3313 local_irq_save(flags
);
3314 ns
= p
->sched_time
+ sched_clock() - p
->last_ran
;
3315 local_irq_restore(flags
);
3321 * We place interactive tasks back into the active array, if possible.
3323 * To guarantee that this does not starve expired tasks we ignore the
3324 * interactivity of a task if the first expired task had to wait more
3325 * than a 'reasonable' amount of time. This deadline timeout is
3326 * load-dependent, as the frequency of array switched decreases with
3327 * increasing number of running tasks. We also ignore the interactivity
3328 * if a better static_prio task has expired:
3330 static inline int expired_starving(struct rq
*rq
)
3332 if (rq
->curr
->static_prio
> rq
->best_expired_prio
)
3334 if (!STARVATION_LIMIT
|| !rq
->expired_timestamp
)
3336 if (jiffies
- rq
->expired_timestamp
> STARVATION_LIMIT
* rq
->nr_running
)
3342 * Account user cpu time to a process.
3343 * @p: the process that the cpu time gets accounted to
3344 * @hardirq_offset: the offset to subtract from hardirq_count()
3345 * @cputime: the cpu time spent in user space since the last update
3347 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3349 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3352 p
->utime
= cputime_add(p
->utime
, cputime
);
3354 /* Add user time to cpustat. */
3355 tmp
= cputime_to_cputime64(cputime
);
3356 if (TASK_NICE(p
) > 0)
3357 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3359 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3363 * Account system cpu time to a process.
3364 * @p: the process that the cpu time gets accounted to
3365 * @hardirq_offset: the offset to subtract from hardirq_count()
3366 * @cputime: the cpu time spent in kernel space since the last update
3368 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3371 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3372 struct rq
*rq
= this_rq();
3375 p
->stime
= cputime_add(p
->stime
, cputime
);
3377 /* Add system time to cpustat. */
3378 tmp
= cputime_to_cputime64(cputime
);
3379 if (hardirq_count() - hardirq_offset
)
3380 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3381 else if (softirq_count())
3382 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3383 else if (p
!= rq
->idle
)
3384 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3385 else if (atomic_read(&rq
->nr_iowait
) > 0)
3386 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3388 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3389 /* Account for system time used */
3390 acct_update_integrals(p
);
3394 * Account for involuntary wait time.
3395 * @p: the process from which the cpu time has been stolen
3396 * @steal: the cpu time spent in involuntary wait
3398 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3400 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3401 cputime64_t tmp
= cputime_to_cputime64(steal
);
3402 struct rq
*rq
= this_rq();
3404 if (p
== rq
->idle
) {
3405 p
->stime
= cputime_add(p
->stime
, steal
);
3406 if (atomic_read(&rq
->nr_iowait
) > 0)
3407 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3409 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3411 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3414 static void task_running_tick(struct rq
*rq
, struct task_struct
*p
)
3416 if (p
->array
!= rq
->active
) {
3417 /* Task has expired but was not scheduled yet */
3418 set_tsk_need_resched(p
);
3421 spin_lock(&rq
->lock
);
3423 * The task was running during this tick - update the
3424 * time slice counter. Note: we do not update a thread's
3425 * priority until it either goes to sleep or uses up its
3426 * timeslice. This makes it possible for interactive tasks
3427 * to use up their timeslices at their highest priority levels.
3431 * RR tasks need a special form of timeslice management.
3432 * FIFO tasks have no timeslices.
3434 if ((p
->policy
== SCHED_RR
) && !--p
->time_slice
) {
3435 p
->time_slice
= task_timeslice(p
);
3436 p
->first_time_slice
= 0;
3437 set_tsk_need_resched(p
);
3439 /* put it at the end of the queue: */
3440 requeue_task(p
, rq
->active
);
3444 if (!--p
->time_slice
) {
3445 dequeue_task(p
, rq
->active
);
3446 set_tsk_need_resched(p
);
3447 p
->prio
= effective_prio(p
);
3448 p
->time_slice
= task_timeslice(p
);
3449 p
->first_time_slice
= 0;
3451 if (!rq
->expired_timestamp
)
3452 rq
->expired_timestamp
= jiffies
;
3453 if (!TASK_INTERACTIVE(p
) || expired_starving(rq
)) {
3454 enqueue_task(p
, rq
->expired
);
3455 if (p
->static_prio
< rq
->best_expired_prio
)
3456 rq
->best_expired_prio
= p
->static_prio
;
3458 enqueue_task(p
, rq
->active
);
3461 * Prevent a too long timeslice allowing a task to monopolize
3462 * the CPU. We do this by splitting up the timeslice into
3465 * Note: this does not mean the task's timeslices expire or
3466 * get lost in any way, they just might be preempted by
3467 * another task of equal priority. (one with higher
3468 * priority would have preempted this task already.) We
3469 * requeue this task to the end of the list on this priority
3470 * level, which is in essence a round-robin of tasks with
3473 * This only applies to tasks in the interactive
3474 * delta range with at least TIMESLICE_GRANULARITY to requeue.
3476 if (TASK_INTERACTIVE(p
) && !((task_timeslice(p
) -
3477 p
->time_slice
) % TIMESLICE_GRANULARITY(p
)) &&
3478 (p
->time_slice
>= TIMESLICE_GRANULARITY(p
)) &&
3479 (p
->array
== rq
->active
)) {
3481 requeue_task(p
, rq
->active
);
3482 set_tsk_need_resched(p
);
3486 spin_unlock(&rq
->lock
);
3490 * This function gets called by the timer code, with HZ frequency.
3491 * We call it with interrupts disabled.
3493 * It also gets called by the fork code, when changing the parent's
3496 void scheduler_tick(void)
3498 unsigned long long now
= sched_clock();
3499 struct task_struct
*p
= current
;
3500 int cpu
= smp_processor_id();
3501 int idle_at_tick
= idle_cpu(cpu
);
3502 struct rq
*rq
= cpu_rq(cpu
);
3504 update_cpu_clock(p
, rq
, now
);
3507 task_running_tick(rq
, p
);
3510 rq
->idle_at_tick
= idle_at_tick
;
3511 trigger_load_balance(cpu
);
3515 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3517 void fastcall
add_preempt_count(int val
)
3522 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3524 preempt_count() += val
;
3526 * Spinlock count overflowing soon?
3528 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3531 EXPORT_SYMBOL(add_preempt_count
);
3533 void fastcall
sub_preempt_count(int val
)
3538 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3541 * Is the spinlock portion underflowing?
3543 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3544 !(preempt_count() & PREEMPT_MASK
)))
3547 preempt_count() -= val
;
3549 EXPORT_SYMBOL(sub_preempt_count
);
3553 static inline int interactive_sleep(enum sleep_type sleep_type
)
3555 return (sleep_type
== SLEEP_INTERACTIVE
||
3556 sleep_type
== SLEEP_INTERRUPTED
);
3560 * schedule() is the main scheduler function.
3562 asmlinkage
void __sched
schedule(void)
3564 struct task_struct
*prev
, *next
;
3565 struct prio_array
*array
;
3566 struct list_head
*queue
;
3567 unsigned long long now
;
3568 unsigned long run_time
;
3569 int cpu
, idx
, new_prio
;
3574 * Test if we are atomic. Since do_exit() needs to call into
3575 * schedule() atomically, we ignore that path for now.
3576 * Otherwise, whine if we are scheduling when we should not be.
3578 if (unlikely(in_atomic() && !current
->exit_state
)) {
3579 printk(KERN_ERR
"BUG: scheduling while atomic: "
3581 current
->comm
, preempt_count(), current
->pid
);
3582 debug_show_held_locks(current
);
3583 if (irqs_disabled())
3584 print_irqtrace_events(current
);
3587 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3592 release_kernel_lock(prev
);
3593 need_resched_nonpreemptible
:
3597 * The idle thread is not allowed to schedule!
3598 * Remove this check after it has been exercised a bit.
3600 if (unlikely(prev
== rq
->idle
) && prev
->state
!= TASK_RUNNING
) {
3601 printk(KERN_ERR
"bad: scheduling from the idle thread!\n");
3605 schedstat_inc(rq
, sched_cnt
);
3606 now
= sched_clock();
3607 if (likely((long long)(now
- prev
->timestamp
) < NS_MAX_SLEEP_AVG
)) {
3608 run_time
= now
- prev
->timestamp
;
3609 if (unlikely((long long)(now
- prev
->timestamp
) < 0))
3612 run_time
= NS_MAX_SLEEP_AVG
;
3615 * Tasks charged proportionately less run_time at high sleep_avg to
3616 * delay them losing their interactive status
3618 run_time
/= (CURRENT_BONUS(prev
) ? : 1);
3620 spin_lock_irq(&rq
->lock
);
3622 switch_count
= &prev
->nivcsw
;
3623 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3624 switch_count
= &prev
->nvcsw
;
3625 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3626 unlikely(signal_pending(prev
))))
3627 prev
->state
= TASK_RUNNING
;
3629 if (prev
->state
== TASK_UNINTERRUPTIBLE
)
3630 rq
->nr_uninterruptible
++;
3631 deactivate_task(prev
, rq
);
3635 cpu
= smp_processor_id();
3636 if (unlikely(!rq
->nr_running
)) {
3637 idle_balance(cpu
, rq
);
3638 if (!rq
->nr_running
) {
3640 rq
->expired_timestamp
= 0;
3646 if (unlikely(!array
->nr_active
)) {
3648 * Switch the active and expired arrays.
3650 schedstat_inc(rq
, sched_switch
);
3651 rq
->active
= rq
->expired
;
3652 rq
->expired
= array
;
3654 rq
->expired_timestamp
= 0;
3655 rq
->best_expired_prio
= MAX_PRIO
;
3658 idx
= sched_find_first_bit(array
->bitmap
);
3659 queue
= array
->queue
+ idx
;
3660 next
= list_entry(queue
->next
, struct task_struct
, run_list
);
3662 if (!rt_task(next
) && interactive_sleep(next
->sleep_type
)) {
3663 unsigned long long delta
= now
- next
->timestamp
;
3664 if (unlikely((long long)(now
- next
->timestamp
) < 0))
3667 if (next
->sleep_type
== SLEEP_INTERACTIVE
)
3668 delta
= delta
* (ON_RUNQUEUE_WEIGHT
* 128 / 100) / 128;
3670 array
= next
->array
;
3671 new_prio
= recalc_task_prio(next
, next
->timestamp
+ delta
);
3673 if (unlikely(next
->prio
!= new_prio
)) {
3674 dequeue_task(next
, array
);
3675 next
->prio
= new_prio
;
3676 enqueue_task(next
, array
);
3679 next
->sleep_type
= SLEEP_NORMAL
;
3681 if (next
== rq
->idle
)
3682 schedstat_inc(rq
, sched_goidle
);
3684 prefetch_stack(next
);
3685 clear_tsk_need_resched(prev
);
3686 rcu_qsctr_inc(task_cpu(prev
));
3688 update_cpu_clock(prev
, rq
, now
);
3690 prev
->sleep_avg
-= run_time
;
3691 if ((long)prev
->sleep_avg
<= 0)
3692 prev
->sleep_avg
= 0;
3693 prev
->timestamp
= prev
->last_ran
= now
;
3695 sched_info_switch(prev
, next
);
3696 if (likely(prev
!= next
)) {
3697 next
->timestamp
= next
->last_ran
= now
;
3702 prepare_task_switch(rq
, next
);
3703 prev
= context_switch(rq
, prev
, next
);
3706 * this_rq must be evaluated again because prev may have moved
3707 * CPUs since it called schedule(), thus the 'rq' on its stack
3708 * frame will be invalid.
3710 finish_task_switch(this_rq(), prev
);
3712 spin_unlock_irq(&rq
->lock
);
3715 if (unlikely(reacquire_kernel_lock(prev
) < 0))
3716 goto need_resched_nonpreemptible
;
3717 preempt_enable_no_resched();
3718 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3721 EXPORT_SYMBOL(schedule
);
3723 #ifdef CONFIG_PREEMPT
3725 * this is the entry point to schedule() from in-kernel preemption
3726 * off of preempt_enable. Kernel preemptions off return from interrupt
3727 * occur there and call schedule directly.
3729 asmlinkage
void __sched
preempt_schedule(void)
3731 struct thread_info
*ti
= current_thread_info();
3732 #ifdef CONFIG_PREEMPT_BKL
3733 struct task_struct
*task
= current
;
3734 int saved_lock_depth
;
3737 * If there is a non-zero preempt_count or interrupts are disabled,
3738 * we do not want to preempt the current task. Just return..
3740 if (likely(ti
->preempt_count
|| irqs_disabled()))
3744 add_preempt_count(PREEMPT_ACTIVE
);
3746 * We keep the big kernel semaphore locked, but we
3747 * clear ->lock_depth so that schedule() doesnt
3748 * auto-release the semaphore:
3750 #ifdef CONFIG_PREEMPT_BKL
3751 saved_lock_depth
= task
->lock_depth
;
3752 task
->lock_depth
= -1;
3755 #ifdef CONFIG_PREEMPT_BKL
3756 task
->lock_depth
= saved_lock_depth
;
3758 sub_preempt_count(PREEMPT_ACTIVE
);
3760 /* we could miss a preemption opportunity between schedule and now */
3762 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3765 EXPORT_SYMBOL(preempt_schedule
);
3768 * this is the entry point to schedule() from kernel preemption
3769 * off of irq context.
3770 * Note, that this is called and return with irqs disabled. This will
3771 * protect us against recursive calling from irq.
3773 asmlinkage
void __sched
preempt_schedule_irq(void)
3775 struct thread_info
*ti
= current_thread_info();
3776 #ifdef CONFIG_PREEMPT_BKL
3777 struct task_struct
*task
= current
;
3778 int saved_lock_depth
;
3780 /* Catch callers which need to be fixed */
3781 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3784 add_preempt_count(PREEMPT_ACTIVE
);
3786 * We keep the big kernel semaphore locked, but we
3787 * clear ->lock_depth so that schedule() doesnt
3788 * auto-release the semaphore:
3790 #ifdef CONFIG_PREEMPT_BKL
3791 saved_lock_depth
= task
->lock_depth
;
3792 task
->lock_depth
= -1;
3796 local_irq_disable();
3797 #ifdef CONFIG_PREEMPT_BKL
3798 task
->lock_depth
= saved_lock_depth
;
3800 sub_preempt_count(PREEMPT_ACTIVE
);
3802 /* we could miss a preemption opportunity between schedule and now */
3804 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3808 #endif /* CONFIG_PREEMPT */
3810 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3813 return try_to_wake_up(curr
->private, mode
, sync
);
3815 EXPORT_SYMBOL(default_wake_function
);
3818 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3819 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3820 * number) then we wake all the non-exclusive tasks and one exclusive task.
3822 * There are circumstances in which we can try to wake a task which has already
3823 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3824 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3826 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3827 int nr_exclusive
, int sync
, void *key
)
3829 struct list_head
*tmp
, *next
;
3831 list_for_each_safe(tmp
, next
, &q
->task_list
) {
3832 wait_queue_t
*curr
= list_entry(tmp
, wait_queue_t
, task_list
);
3833 unsigned flags
= curr
->flags
;
3835 if (curr
->func(curr
, mode
, sync
, key
) &&
3836 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3842 * __wake_up - wake up threads blocked on a waitqueue.
3844 * @mode: which threads
3845 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3846 * @key: is directly passed to the wakeup function
3848 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3849 int nr_exclusive
, void *key
)
3851 unsigned long flags
;
3853 spin_lock_irqsave(&q
->lock
, flags
);
3854 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3855 spin_unlock_irqrestore(&q
->lock
, flags
);
3857 EXPORT_SYMBOL(__wake_up
);
3860 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3862 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3864 __wake_up_common(q
, mode
, 1, 0, NULL
);
3868 * __wake_up_sync - wake up threads blocked on a waitqueue.
3870 * @mode: which threads
3871 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3873 * The sync wakeup differs that the waker knows that it will schedule
3874 * away soon, so while the target thread will be woken up, it will not
3875 * be migrated to another CPU - ie. the two threads are 'synchronized'
3876 * with each other. This can prevent needless bouncing between CPUs.
3878 * On UP it can prevent extra preemption.
3881 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3883 unsigned long flags
;
3889 if (unlikely(!nr_exclusive
))
3892 spin_lock_irqsave(&q
->lock
, flags
);
3893 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3894 spin_unlock_irqrestore(&q
->lock
, flags
);
3896 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3898 void fastcall
complete(struct completion
*x
)
3900 unsigned long flags
;
3902 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3904 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3906 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3908 EXPORT_SYMBOL(complete
);
3910 void fastcall
complete_all(struct completion
*x
)
3912 unsigned long flags
;
3914 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3915 x
->done
+= UINT_MAX
/2;
3916 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3918 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3920 EXPORT_SYMBOL(complete_all
);
3922 void fastcall __sched
wait_for_completion(struct completion
*x
)
3926 spin_lock_irq(&x
->wait
.lock
);
3928 DECLARE_WAITQUEUE(wait
, current
);
3930 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3931 __add_wait_queue_tail(&x
->wait
, &wait
);
3933 __set_current_state(TASK_UNINTERRUPTIBLE
);
3934 spin_unlock_irq(&x
->wait
.lock
);
3936 spin_lock_irq(&x
->wait
.lock
);
3938 __remove_wait_queue(&x
->wait
, &wait
);
3941 spin_unlock_irq(&x
->wait
.lock
);
3943 EXPORT_SYMBOL(wait_for_completion
);
3945 unsigned long fastcall __sched
3946 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3950 spin_lock_irq(&x
->wait
.lock
);
3952 DECLARE_WAITQUEUE(wait
, current
);
3954 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3955 __add_wait_queue_tail(&x
->wait
, &wait
);
3957 __set_current_state(TASK_UNINTERRUPTIBLE
);
3958 spin_unlock_irq(&x
->wait
.lock
);
3959 timeout
= schedule_timeout(timeout
);
3960 spin_lock_irq(&x
->wait
.lock
);
3962 __remove_wait_queue(&x
->wait
, &wait
);
3966 __remove_wait_queue(&x
->wait
, &wait
);
3970 spin_unlock_irq(&x
->wait
.lock
);
3973 EXPORT_SYMBOL(wait_for_completion_timeout
);
3975 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3981 spin_lock_irq(&x
->wait
.lock
);
3983 DECLARE_WAITQUEUE(wait
, current
);
3985 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3986 __add_wait_queue_tail(&x
->wait
, &wait
);
3988 if (signal_pending(current
)) {
3990 __remove_wait_queue(&x
->wait
, &wait
);
3993 __set_current_state(TASK_INTERRUPTIBLE
);
3994 spin_unlock_irq(&x
->wait
.lock
);
3996 spin_lock_irq(&x
->wait
.lock
);
3998 __remove_wait_queue(&x
->wait
, &wait
);
4002 spin_unlock_irq(&x
->wait
.lock
);
4006 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4008 unsigned long fastcall __sched
4009 wait_for_completion_interruptible_timeout(struct completion
*x
,
4010 unsigned long timeout
)
4014 spin_lock_irq(&x
->wait
.lock
);
4016 DECLARE_WAITQUEUE(wait
, current
);
4018 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4019 __add_wait_queue_tail(&x
->wait
, &wait
);
4021 if (signal_pending(current
)) {
4022 timeout
= -ERESTARTSYS
;
4023 __remove_wait_queue(&x
->wait
, &wait
);
4026 __set_current_state(TASK_INTERRUPTIBLE
);
4027 spin_unlock_irq(&x
->wait
.lock
);
4028 timeout
= schedule_timeout(timeout
);
4029 spin_lock_irq(&x
->wait
.lock
);
4031 __remove_wait_queue(&x
->wait
, &wait
);
4035 __remove_wait_queue(&x
->wait
, &wait
);
4039 spin_unlock_irq(&x
->wait
.lock
);
4042 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4045 #define SLEEP_ON_VAR \
4046 unsigned long flags; \
4047 wait_queue_t wait; \
4048 init_waitqueue_entry(&wait, current);
4050 #define SLEEP_ON_HEAD \
4051 spin_lock_irqsave(&q->lock,flags); \
4052 __add_wait_queue(q, &wait); \
4053 spin_unlock(&q->lock);
4055 #define SLEEP_ON_TAIL \
4056 spin_lock_irq(&q->lock); \
4057 __remove_wait_queue(q, &wait); \
4058 spin_unlock_irqrestore(&q->lock, flags);
4060 void fastcall __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4064 current
->state
= TASK_INTERRUPTIBLE
;
4070 EXPORT_SYMBOL(interruptible_sleep_on
);
4072 long fastcall __sched
4073 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4077 current
->state
= TASK_INTERRUPTIBLE
;
4080 timeout
= schedule_timeout(timeout
);
4085 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4087 void fastcall __sched
sleep_on(wait_queue_head_t
*q
)
4091 current
->state
= TASK_UNINTERRUPTIBLE
;
4097 EXPORT_SYMBOL(sleep_on
);
4099 long fastcall __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4103 current
->state
= TASK_UNINTERRUPTIBLE
;
4106 timeout
= schedule_timeout(timeout
);
4112 EXPORT_SYMBOL(sleep_on_timeout
);
4114 #ifdef CONFIG_RT_MUTEXES
4117 * rt_mutex_setprio - set the current priority of a task
4119 * @prio: prio value (kernel-internal form)
4121 * This function changes the 'effective' priority of a task. It does
4122 * not touch ->normal_prio like __setscheduler().
4124 * Used by the rt_mutex code to implement priority inheritance logic.
4126 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4128 struct prio_array
*array
;
4129 unsigned long flags
;
4133 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4135 rq
= task_rq_lock(p
, &flags
);
4140 dequeue_task(p
, array
);
4145 * If changing to an RT priority then queue it
4146 * in the active array!
4150 enqueue_task(p
, array
);
4152 * Reschedule if we are currently running on this runqueue and
4153 * our priority decreased, or if we are not currently running on
4154 * this runqueue and our priority is higher than the current's
4156 if (task_running(rq
, p
)) {
4157 if (p
->prio
> oldprio
)
4158 resched_task(rq
->curr
);
4159 } else if (TASK_PREEMPTS_CURR(p
, rq
))
4160 resched_task(rq
->curr
);
4162 task_rq_unlock(rq
, &flags
);
4167 void set_user_nice(struct task_struct
*p
, long nice
)
4169 struct prio_array
*array
;
4170 int old_prio
, delta
;
4171 unsigned long flags
;
4174 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4177 * We have to be careful, if called from sys_setpriority(),
4178 * the task might be in the middle of scheduling on another CPU.
4180 rq
= task_rq_lock(p
, &flags
);
4182 * The RT priorities are set via sched_setscheduler(), but we still
4183 * allow the 'normal' nice value to be set - but as expected
4184 * it wont have any effect on scheduling until the task is
4185 * not SCHED_NORMAL/SCHED_BATCH:
4187 if (has_rt_policy(p
)) {
4188 p
->static_prio
= NICE_TO_PRIO(nice
);
4193 dequeue_task(p
, array
);
4194 dec_raw_weighted_load(rq
, p
);
4197 p
->static_prio
= NICE_TO_PRIO(nice
);
4200 p
->prio
= effective_prio(p
);
4201 delta
= p
->prio
- old_prio
;
4204 enqueue_task(p
, array
);
4205 inc_raw_weighted_load(rq
, p
);
4207 * If the task increased its priority or is running and
4208 * lowered its priority, then reschedule its CPU:
4210 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4211 resched_task(rq
->curr
);
4214 task_rq_unlock(rq
, &flags
);
4216 EXPORT_SYMBOL(set_user_nice
);
4219 * can_nice - check if a task can reduce its nice value
4223 int can_nice(const struct task_struct
*p
, const int nice
)
4225 /* convert nice value [19,-20] to rlimit style value [1,40] */
4226 int nice_rlim
= 20 - nice
;
4228 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4229 capable(CAP_SYS_NICE
));
4232 #ifdef __ARCH_WANT_SYS_NICE
4235 * sys_nice - change the priority of the current process.
4236 * @increment: priority increment
4238 * sys_setpriority is a more generic, but much slower function that
4239 * does similar things.
4241 asmlinkage
long sys_nice(int increment
)
4246 * Setpriority might change our priority at the same moment.
4247 * We don't have to worry. Conceptually one call occurs first
4248 * and we have a single winner.
4250 if (increment
< -40)
4255 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4261 if (increment
< 0 && !can_nice(current
, nice
))
4264 retval
= security_task_setnice(current
, nice
);
4268 set_user_nice(current
, nice
);
4275 * task_prio - return the priority value of a given task.
4276 * @p: the task in question.
4278 * This is the priority value as seen by users in /proc.
4279 * RT tasks are offset by -200. Normal tasks are centered
4280 * around 0, value goes from -16 to +15.
4282 int task_prio(const struct task_struct
*p
)
4284 return p
->prio
- MAX_RT_PRIO
;
4288 * task_nice - return the nice value of a given task.
4289 * @p: the task in question.
4291 int task_nice(const struct task_struct
*p
)
4293 return TASK_NICE(p
);
4295 EXPORT_SYMBOL_GPL(task_nice
);
4298 * idle_cpu - is a given cpu idle currently?
4299 * @cpu: the processor in question.
4301 int idle_cpu(int cpu
)
4303 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4307 * idle_task - return the idle task for a given cpu.
4308 * @cpu: the processor in question.
4310 struct task_struct
*idle_task(int cpu
)
4312 return cpu_rq(cpu
)->idle
;
4316 * find_process_by_pid - find a process with a matching PID value.
4317 * @pid: the pid in question.
4319 static inline struct task_struct
*find_process_by_pid(pid_t pid
)
4321 return pid
? find_task_by_pid(pid
) : current
;
4324 /* Actually do priority change: must hold rq lock. */
4325 static void __setscheduler(struct task_struct
*p
, int policy
, int prio
)
4330 p
->rt_priority
= prio
;
4331 p
->normal_prio
= normal_prio(p
);
4332 /* we are holding p->pi_lock already */
4333 p
->prio
= rt_mutex_getprio(p
);
4335 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
4337 if (policy
== SCHED_BATCH
)
4343 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4344 * @p: the task in question.
4345 * @policy: new policy.
4346 * @param: structure containing the new RT priority.
4348 * NOTE that the task may be already dead.
4350 int sched_setscheduler(struct task_struct
*p
, int policy
,
4351 struct sched_param
*param
)
4353 int retval
, oldprio
, oldpolicy
= -1;
4354 struct prio_array
*array
;
4355 unsigned long flags
;
4358 /* may grab non-irq protected spin_locks */
4359 BUG_ON(in_interrupt());
4361 /* double check policy once rq lock held */
4363 policy
= oldpolicy
= p
->policy
;
4364 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4365 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
)
4368 * Valid priorities for SCHED_FIFO and SCHED_RR are
4369 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
4372 if (param
->sched_priority
< 0 ||
4373 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4374 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4376 if (is_rt_policy(policy
) != (param
->sched_priority
!= 0))
4380 * Allow unprivileged RT tasks to decrease priority:
4382 if (!capable(CAP_SYS_NICE
)) {
4383 if (is_rt_policy(policy
)) {
4384 unsigned long rlim_rtprio
;
4385 unsigned long flags
;
4387 if (!lock_task_sighand(p
, &flags
))
4389 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4390 unlock_task_sighand(p
, &flags
);
4392 /* can't set/change the rt policy */
4393 if (policy
!= p
->policy
&& !rlim_rtprio
)
4396 /* can't increase priority */
4397 if (param
->sched_priority
> p
->rt_priority
&&
4398 param
->sched_priority
> rlim_rtprio
)
4402 /* can't change other user's priorities */
4403 if ((current
->euid
!= p
->euid
) &&
4404 (current
->euid
!= p
->uid
))
4408 retval
= security_task_setscheduler(p
, policy
, param
);
4412 * make sure no PI-waiters arrive (or leave) while we are
4413 * changing the priority of the task:
4415 spin_lock_irqsave(&p
->pi_lock
, flags
);
4417 * To be able to change p->policy safely, the apropriate
4418 * runqueue lock must be held.
4420 rq
= __task_rq_lock(p
);
4421 /* recheck policy now with rq lock held */
4422 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4423 policy
= oldpolicy
= -1;
4424 __task_rq_unlock(rq
);
4425 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4430 deactivate_task(p
, rq
);
4432 __setscheduler(p
, policy
, param
->sched_priority
);
4434 __activate_task(p
, rq
);
4436 * Reschedule if we are currently running on this runqueue and
4437 * our priority decreased, or if we are not currently running on
4438 * this runqueue and our priority is higher than the current's
4440 if (task_running(rq
, p
)) {
4441 if (p
->prio
> oldprio
)
4442 resched_task(rq
->curr
);
4443 } else if (TASK_PREEMPTS_CURR(p
, rq
))
4444 resched_task(rq
->curr
);
4446 __task_rq_unlock(rq
);
4447 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4449 rt_mutex_adjust_pi(p
);
4453 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4456 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4458 struct sched_param lparam
;
4459 struct task_struct
*p
;
4462 if (!param
|| pid
< 0)
4464 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4469 p
= find_process_by_pid(pid
);
4471 retval
= sched_setscheduler(p
, policy
, &lparam
);
4478 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4479 * @pid: the pid in question.
4480 * @policy: new policy.
4481 * @param: structure containing the new RT priority.
4483 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
4484 struct sched_param __user
*param
)
4486 /* negative values for policy are not valid */
4490 return do_sched_setscheduler(pid
, policy
, param
);
4494 * sys_sched_setparam - set/change the RT priority of a thread
4495 * @pid: the pid in question.
4496 * @param: structure containing the new RT priority.
4498 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4500 return do_sched_setscheduler(pid
, -1, param
);
4504 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4505 * @pid: the pid in question.
4507 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4509 struct task_struct
*p
;
4510 int retval
= -EINVAL
;
4516 read_lock(&tasklist_lock
);
4517 p
= find_process_by_pid(pid
);
4519 retval
= security_task_getscheduler(p
);
4523 read_unlock(&tasklist_lock
);
4530 * sys_sched_getscheduler - get the RT priority of a thread
4531 * @pid: the pid in question.
4532 * @param: structure containing the RT priority.
4534 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4536 struct sched_param lp
;
4537 struct task_struct
*p
;
4538 int retval
= -EINVAL
;
4540 if (!param
|| pid
< 0)
4543 read_lock(&tasklist_lock
);
4544 p
= find_process_by_pid(pid
);
4549 retval
= security_task_getscheduler(p
);
4553 lp
.sched_priority
= p
->rt_priority
;
4554 read_unlock(&tasklist_lock
);
4557 * This one might sleep, we cannot do it with a spinlock held ...
4559 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4565 read_unlock(&tasklist_lock
);
4569 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4571 cpumask_t cpus_allowed
;
4572 struct task_struct
*p
;
4575 mutex_lock(&sched_hotcpu_mutex
);
4576 read_lock(&tasklist_lock
);
4578 p
= find_process_by_pid(pid
);
4580 read_unlock(&tasklist_lock
);
4581 mutex_unlock(&sched_hotcpu_mutex
);
4586 * It is not safe to call set_cpus_allowed with the
4587 * tasklist_lock held. We will bump the task_struct's
4588 * usage count and then drop tasklist_lock.
4591 read_unlock(&tasklist_lock
);
4594 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4595 !capable(CAP_SYS_NICE
))
4598 retval
= security_task_setscheduler(p
, 0, NULL
);
4602 cpus_allowed
= cpuset_cpus_allowed(p
);
4603 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4604 retval
= set_cpus_allowed(p
, new_mask
);
4608 mutex_unlock(&sched_hotcpu_mutex
);
4612 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4613 cpumask_t
*new_mask
)
4615 if (len
< sizeof(cpumask_t
)) {
4616 memset(new_mask
, 0, sizeof(cpumask_t
));
4617 } else if (len
> sizeof(cpumask_t
)) {
4618 len
= sizeof(cpumask_t
);
4620 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4624 * sys_sched_setaffinity - set the cpu affinity of a process
4625 * @pid: pid of the process
4626 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4627 * @user_mask_ptr: user-space pointer to the new cpu mask
4629 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4630 unsigned long __user
*user_mask_ptr
)
4635 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4639 return sched_setaffinity(pid
, new_mask
);
4643 * Represents all cpu's present in the system
4644 * In systems capable of hotplug, this map could dynamically grow
4645 * as new cpu's are detected in the system via any platform specific
4646 * method, such as ACPI for e.g.
4649 cpumask_t cpu_present_map __read_mostly
;
4650 EXPORT_SYMBOL(cpu_present_map
);
4653 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4654 EXPORT_SYMBOL(cpu_online_map
);
4656 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4657 EXPORT_SYMBOL(cpu_possible_map
);
4660 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4662 struct task_struct
*p
;
4665 mutex_lock(&sched_hotcpu_mutex
);
4666 read_lock(&tasklist_lock
);
4669 p
= find_process_by_pid(pid
);
4673 retval
= security_task_getscheduler(p
);
4677 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4680 read_unlock(&tasklist_lock
);
4681 mutex_unlock(&sched_hotcpu_mutex
);
4689 * sys_sched_getaffinity - get the cpu affinity of a process
4690 * @pid: pid of the process
4691 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4692 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4694 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4695 unsigned long __user
*user_mask_ptr
)
4700 if (len
< sizeof(cpumask_t
))
4703 ret
= sched_getaffinity(pid
, &mask
);
4707 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4710 return sizeof(cpumask_t
);
4714 * sys_sched_yield - yield the current processor to other threads.
4716 * This function yields the current CPU by moving the calling thread
4717 * to the expired array. If there are no other threads running on this
4718 * CPU then this function will return.
4720 asmlinkage
long sys_sched_yield(void)
4722 struct rq
*rq
= this_rq_lock();
4723 struct prio_array
*array
= current
->array
, *target
= rq
->expired
;
4725 schedstat_inc(rq
, yld_cnt
);
4727 * We implement yielding by moving the task into the expired
4730 * (special rule: RT tasks will just roundrobin in the active
4733 if (rt_task(current
))
4734 target
= rq
->active
;
4736 if (array
->nr_active
== 1) {
4737 schedstat_inc(rq
, yld_act_empty
);
4738 if (!rq
->expired
->nr_active
)
4739 schedstat_inc(rq
, yld_both_empty
);
4740 } else if (!rq
->expired
->nr_active
)
4741 schedstat_inc(rq
, yld_exp_empty
);
4743 if (array
!= target
) {
4744 dequeue_task(current
, array
);
4745 enqueue_task(current
, target
);
4748 * requeue_task is cheaper so perform that if possible.
4750 requeue_task(current
, array
);
4753 * Since we are going to call schedule() anyway, there's
4754 * no need to preempt or enable interrupts:
4756 __release(rq
->lock
);
4757 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4758 _raw_spin_unlock(&rq
->lock
);
4759 preempt_enable_no_resched();
4766 static void __cond_resched(void)
4768 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4769 __might_sleep(__FILE__
, __LINE__
);
4772 * The BKS might be reacquired before we have dropped
4773 * PREEMPT_ACTIVE, which could trigger a second
4774 * cond_resched() call.
4777 add_preempt_count(PREEMPT_ACTIVE
);
4779 sub_preempt_count(PREEMPT_ACTIVE
);
4780 } while (need_resched());
4783 int __sched
cond_resched(void)
4785 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4786 system_state
== SYSTEM_RUNNING
) {
4792 EXPORT_SYMBOL(cond_resched
);
4795 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4796 * call schedule, and on return reacquire the lock.
4798 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4799 * operations here to prevent schedule() from being called twice (once via
4800 * spin_unlock(), once by hand).
4802 int cond_resched_lock(spinlock_t
*lock
)
4806 if (need_lockbreak(lock
)) {
4812 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4813 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4814 _raw_spin_unlock(lock
);
4815 preempt_enable_no_resched();
4822 EXPORT_SYMBOL(cond_resched_lock
);
4824 int __sched
cond_resched_softirq(void)
4826 BUG_ON(!in_softirq());
4828 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4836 EXPORT_SYMBOL(cond_resched_softirq
);
4839 * yield - yield the current processor to other threads.
4841 * This is a shortcut for kernel-space yielding - it marks the
4842 * thread runnable and calls sys_sched_yield().
4844 void __sched
yield(void)
4846 set_current_state(TASK_RUNNING
);
4849 EXPORT_SYMBOL(yield
);
4852 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4853 * that process accounting knows that this is a task in IO wait state.
4855 * But don't do that if it is a deliberate, throttling IO wait (this task
4856 * has set its backing_dev_info: the queue against which it should throttle)
4858 void __sched
io_schedule(void)
4860 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4862 delayacct_blkio_start();
4863 atomic_inc(&rq
->nr_iowait
);
4865 atomic_dec(&rq
->nr_iowait
);
4866 delayacct_blkio_end();
4868 EXPORT_SYMBOL(io_schedule
);
4870 long __sched
io_schedule_timeout(long timeout
)
4872 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4875 delayacct_blkio_start();
4876 atomic_inc(&rq
->nr_iowait
);
4877 ret
= schedule_timeout(timeout
);
4878 atomic_dec(&rq
->nr_iowait
);
4879 delayacct_blkio_end();
4884 * sys_sched_get_priority_max - return maximum RT priority.
4885 * @policy: scheduling class.
4887 * this syscall returns the maximum rt_priority that can be used
4888 * by a given scheduling class.
4890 asmlinkage
long sys_sched_get_priority_max(int policy
)
4897 ret
= MAX_USER_RT_PRIO
-1;
4908 * sys_sched_get_priority_min - return minimum RT priority.
4909 * @policy: scheduling class.
4911 * this syscall returns the minimum rt_priority that can be used
4912 * by a given scheduling class.
4914 asmlinkage
long sys_sched_get_priority_min(int policy
)
4931 * sys_sched_rr_get_interval - return the default timeslice of a process.
4932 * @pid: pid of the process.
4933 * @interval: userspace pointer to the timeslice value.
4935 * this syscall writes the default timeslice value of a given process
4936 * into the user-space timespec buffer. A value of '0' means infinity.
4939 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4941 struct task_struct
*p
;
4942 int retval
= -EINVAL
;
4949 read_lock(&tasklist_lock
);
4950 p
= find_process_by_pid(pid
);
4954 retval
= security_task_getscheduler(p
);
4958 jiffies_to_timespec(p
->policy
== SCHED_FIFO
?
4959 0 : task_timeslice(p
), &t
);
4960 read_unlock(&tasklist_lock
);
4961 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4965 read_unlock(&tasklist_lock
);
4969 static const char stat_nam
[] = "RSDTtZX";
4971 static void show_task(struct task_struct
*p
)
4973 unsigned long free
= 0;
4976 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4977 printk("%-13.13s %c", p
->comm
,
4978 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4979 #if (BITS_PER_LONG == 32)
4980 if (state
== TASK_RUNNING
)
4981 printk(" running ");
4983 printk(" %08lX ", thread_saved_pc(p
));
4985 if (state
== TASK_RUNNING
)
4986 printk(" running task ");
4988 printk(" %016lx ", thread_saved_pc(p
));
4990 #ifdef CONFIG_DEBUG_STACK_USAGE
4992 unsigned long *n
= end_of_stack(p
);
4995 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4998 printk("%5lu %5d %6d", free
, p
->pid
, p
->parent
->pid
);
5000 printk(" (L-TLB)\n");
5002 printk(" (NOTLB)\n");
5004 if (state
!= TASK_RUNNING
)
5005 show_stack(p
, NULL
);
5008 void show_state_filter(unsigned long state_filter
)
5010 struct task_struct
*g
, *p
;
5012 #if (BITS_PER_LONG == 32)
5015 printk(" task PC stack pid father child younger older\n");
5019 printk(" task PC stack pid father child younger older\n");
5021 read_lock(&tasklist_lock
);
5022 do_each_thread(g
, p
) {
5024 * reset the NMI-timeout, listing all files on a slow
5025 * console might take alot of time:
5027 touch_nmi_watchdog();
5028 if (!state_filter
|| (p
->state
& state_filter
))
5030 } while_each_thread(g
, p
);
5032 touch_all_softlockup_watchdogs();
5034 read_unlock(&tasklist_lock
);
5036 * Only show locks if all tasks are dumped:
5038 if (state_filter
== -1)
5039 debug_show_all_locks();
5043 * init_idle - set up an idle thread for a given CPU
5044 * @idle: task in question
5045 * @cpu: cpu the idle task belongs to
5047 * NOTE: this function does not set the idle thread's NEED_RESCHED
5048 * flag, to make booting more robust.
5050 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5052 struct rq
*rq
= cpu_rq(cpu
);
5053 unsigned long flags
;
5055 idle
->timestamp
= sched_clock();
5056 idle
->sleep_avg
= 0;
5058 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
5059 idle
->state
= TASK_RUNNING
;
5060 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
5061 set_task_cpu(idle
, cpu
);
5063 spin_lock_irqsave(&rq
->lock
, flags
);
5064 rq
->curr
= rq
->idle
= idle
;
5065 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5068 spin_unlock_irqrestore(&rq
->lock
, flags
);
5070 /* Set the preempt count _outside_ the spinlocks! */
5071 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
5072 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5074 task_thread_info(idle
)->preempt_count
= 0;
5079 * In a system that switches off the HZ timer nohz_cpu_mask
5080 * indicates which cpus entered this state. This is used
5081 * in the rcu update to wait only for active cpus. For system
5082 * which do not switch off the HZ timer nohz_cpu_mask should
5083 * always be CPU_MASK_NONE.
5085 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
5089 * This is how migration works:
5091 * 1) we queue a struct migration_req structure in the source CPU's
5092 * runqueue and wake up that CPU's migration thread.
5093 * 2) we down() the locked semaphore => thread blocks.
5094 * 3) migration thread wakes up (implicitly it forces the migrated
5095 * thread off the CPU)
5096 * 4) it gets the migration request and checks whether the migrated
5097 * task is still in the wrong runqueue.
5098 * 5) if it's in the wrong runqueue then the migration thread removes
5099 * it and puts it into the right queue.
5100 * 6) migration thread up()s the semaphore.
5101 * 7) we wake up and the migration is done.
5105 * Change a given task's CPU affinity. Migrate the thread to a
5106 * proper CPU and schedule it away if the CPU it's executing on
5107 * is removed from the allowed bitmask.
5109 * NOTE: the caller must have a valid reference to the task, the
5110 * task must not exit() & deallocate itself prematurely. The
5111 * call is not atomic; no spinlocks may be held.
5113 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
5115 struct migration_req req
;
5116 unsigned long flags
;
5120 rq
= task_rq_lock(p
, &flags
);
5121 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
5126 p
->cpus_allowed
= new_mask
;
5127 /* Can the task run on the task's current CPU? If so, we're done */
5128 if (cpu_isset(task_cpu(p
), new_mask
))
5131 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
5132 /* Need help from migration thread: drop lock and wait. */
5133 task_rq_unlock(rq
, &flags
);
5134 wake_up_process(rq
->migration_thread
);
5135 wait_for_completion(&req
.done
);
5136 tlb_migrate_finish(p
->mm
);
5140 task_rq_unlock(rq
, &flags
);
5144 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
5147 * Move (not current) task off this cpu, onto dest cpu. We're doing
5148 * this because either it can't run here any more (set_cpus_allowed()
5149 * away from this CPU, or CPU going down), or because we're
5150 * attempting to rebalance this task on exec (sched_exec).
5152 * So we race with normal scheduler movements, but that's OK, as long
5153 * as the task is no longer on this CPU.
5155 * Returns non-zero if task was successfully migrated.
5157 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5159 struct rq
*rq_dest
, *rq_src
;
5162 if (unlikely(cpu_is_offline(dest_cpu
)))
5165 rq_src
= cpu_rq(src_cpu
);
5166 rq_dest
= cpu_rq(dest_cpu
);
5168 double_rq_lock(rq_src
, rq_dest
);
5169 /* Already moved. */
5170 if (task_cpu(p
) != src_cpu
)
5172 /* Affinity changed (again). */
5173 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
5176 set_task_cpu(p
, dest_cpu
);
5179 * Sync timestamp with rq_dest's before activating.
5180 * The same thing could be achieved by doing this step
5181 * afterwards, and pretending it was a local activate.
5182 * This way is cleaner and logically correct.
5184 p
->timestamp
= p
->timestamp
- rq_src
->most_recent_timestamp
5185 + rq_dest
->most_recent_timestamp
;
5186 deactivate_task(p
, rq_src
);
5187 __activate_task(p
, rq_dest
);
5188 if (TASK_PREEMPTS_CURR(p
, rq_dest
))
5189 resched_task(rq_dest
->curr
);
5193 double_rq_unlock(rq_src
, rq_dest
);
5198 * migration_thread - this is a highprio system thread that performs
5199 * thread migration by bumping thread off CPU then 'pushing' onto
5202 static int migration_thread(void *data
)
5204 int cpu
= (long)data
;
5208 BUG_ON(rq
->migration_thread
!= current
);
5210 set_current_state(TASK_INTERRUPTIBLE
);
5211 while (!kthread_should_stop()) {
5212 struct migration_req
*req
;
5213 struct list_head
*head
;
5217 spin_lock_irq(&rq
->lock
);
5219 if (cpu_is_offline(cpu
)) {
5220 spin_unlock_irq(&rq
->lock
);
5224 if (rq
->active_balance
) {
5225 active_load_balance(rq
, cpu
);
5226 rq
->active_balance
= 0;
5229 head
= &rq
->migration_queue
;
5231 if (list_empty(head
)) {
5232 spin_unlock_irq(&rq
->lock
);
5234 set_current_state(TASK_INTERRUPTIBLE
);
5237 req
= list_entry(head
->next
, struct migration_req
, list
);
5238 list_del_init(head
->next
);
5240 spin_unlock(&rq
->lock
);
5241 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5244 complete(&req
->done
);
5246 __set_current_state(TASK_RUNNING
);
5250 /* Wait for kthread_stop */
5251 set_current_state(TASK_INTERRUPTIBLE
);
5252 while (!kthread_should_stop()) {
5254 set_current_state(TASK_INTERRUPTIBLE
);
5256 __set_current_state(TASK_RUNNING
);
5260 #ifdef CONFIG_HOTPLUG_CPU
5262 * Figure out where task on dead CPU should go, use force if neccessary.
5263 * NOTE: interrupts should be disabled by the caller
5265 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5267 unsigned long flags
;
5274 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5275 cpus_and(mask
, mask
, p
->cpus_allowed
);
5276 dest_cpu
= any_online_cpu(mask
);
5278 /* On any allowed CPU? */
5279 if (dest_cpu
== NR_CPUS
)
5280 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5282 /* No more Mr. Nice Guy. */
5283 if (dest_cpu
== NR_CPUS
) {
5284 rq
= task_rq_lock(p
, &flags
);
5285 cpus_setall(p
->cpus_allowed
);
5286 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5287 task_rq_unlock(rq
, &flags
);
5290 * Don't tell them about moving exiting tasks or
5291 * kernel threads (both mm NULL), since they never
5294 if (p
->mm
&& printk_ratelimit())
5295 printk(KERN_INFO
"process %d (%s) no "
5296 "longer affine to cpu%d\n",
5297 p
->pid
, p
->comm
, dead_cpu
);
5299 if (!__migrate_task(p
, dead_cpu
, dest_cpu
))
5304 * While a dead CPU has no uninterruptible tasks queued at this point,
5305 * it might still have a nonzero ->nr_uninterruptible counter, because
5306 * for performance reasons the counter is not stricly tracking tasks to
5307 * their home CPUs. So we just add the counter to another CPU's counter,
5308 * to keep the global sum constant after CPU-down:
5310 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5312 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5313 unsigned long flags
;
5315 local_irq_save(flags
);
5316 double_rq_lock(rq_src
, rq_dest
);
5317 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5318 rq_src
->nr_uninterruptible
= 0;
5319 double_rq_unlock(rq_src
, rq_dest
);
5320 local_irq_restore(flags
);
5323 /* Run through task list and migrate tasks from the dead cpu. */
5324 static void migrate_live_tasks(int src_cpu
)
5326 struct task_struct
*p
, *t
;
5328 write_lock_irq(&tasklist_lock
);
5330 do_each_thread(t
, p
) {
5334 if (task_cpu(p
) == src_cpu
)
5335 move_task_off_dead_cpu(src_cpu
, p
);
5336 } while_each_thread(t
, p
);
5338 write_unlock_irq(&tasklist_lock
);
5341 /* Schedules idle task to be the next runnable task on current CPU.
5342 * It does so by boosting its priority to highest possible and adding it to
5343 * the _front_ of the runqueue. Used by CPU offline code.
5345 void sched_idle_next(void)
5347 int this_cpu
= smp_processor_id();
5348 struct rq
*rq
= cpu_rq(this_cpu
);
5349 struct task_struct
*p
= rq
->idle
;
5350 unsigned long flags
;
5352 /* cpu has to be offline */
5353 BUG_ON(cpu_online(this_cpu
));
5356 * Strictly not necessary since rest of the CPUs are stopped by now
5357 * and interrupts disabled on the current cpu.
5359 spin_lock_irqsave(&rq
->lock
, flags
);
5361 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5363 /* Add idle task to the _front_ of its priority queue: */
5364 __activate_idle_task(p
, rq
);
5366 spin_unlock_irqrestore(&rq
->lock
, flags
);
5370 * Ensures that the idle task is using init_mm right before its cpu goes
5373 void idle_task_exit(void)
5375 struct mm_struct
*mm
= current
->active_mm
;
5377 BUG_ON(cpu_online(smp_processor_id()));
5380 switch_mm(mm
, &init_mm
, current
);
5384 /* called under rq->lock with disabled interrupts */
5385 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5387 struct rq
*rq
= cpu_rq(dead_cpu
);
5389 /* Must be exiting, otherwise would be on tasklist. */
5390 BUG_ON(p
->exit_state
!= EXIT_ZOMBIE
&& p
->exit_state
!= EXIT_DEAD
);
5392 /* Cannot have done final schedule yet: would have vanished. */
5393 BUG_ON(p
->state
== TASK_DEAD
);
5398 * Drop lock around migration; if someone else moves it,
5399 * that's OK. No task can be added to this CPU, so iteration is
5401 * NOTE: interrupts should be left disabled --dev@
5403 spin_unlock(&rq
->lock
);
5404 move_task_off_dead_cpu(dead_cpu
, p
);
5405 spin_lock(&rq
->lock
);
5410 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5411 static void migrate_dead_tasks(unsigned int dead_cpu
)
5413 struct rq
*rq
= cpu_rq(dead_cpu
);
5414 unsigned int arr
, i
;
5416 for (arr
= 0; arr
< 2; arr
++) {
5417 for (i
= 0; i
< MAX_PRIO
; i
++) {
5418 struct list_head
*list
= &rq
->arrays
[arr
].queue
[i
];
5420 while (!list_empty(list
))
5421 migrate_dead(dead_cpu
, list_entry(list
->next
,
5422 struct task_struct
, run_list
));
5426 #endif /* CONFIG_HOTPLUG_CPU */
5429 * migration_call - callback that gets triggered when a CPU is added.
5430 * Here we can start up the necessary migration thread for the new CPU.
5432 static int __cpuinit
5433 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5435 struct task_struct
*p
;
5436 int cpu
= (long)hcpu
;
5437 unsigned long flags
;
5441 case CPU_LOCK_ACQUIRE
:
5442 mutex_lock(&sched_hotcpu_mutex
);
5445 case CPU_UP_PREPARE
:
5446 case CPU_UP_PREPARE_FROZEN
:
5447 p
= kthread_create(migration_thread
, hcpu
, "migration/%d",cpu
);
5450 p
->flags
|= PF_NOFREEZE
;
5451 kthread_bind(p
, cpu
);
5452 /* Must be high prio: stop_machine expects to yield to it. */
5453 rq
= task_rq_lock(p
, &flags
);
5454 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5455 task_rq_unlock(rq
, &flags
);
5456 cpu_rq(cpu
)->migration_thread
= p
;
5460 case CPU_ONLINE_FROZEN
:
5461 /* Strictly unneccessary, as first user will wake it. */
5462 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5465 #ifdef CONFIG_HOTPLUG_CPU
5466 case CPU_UP_CANCELED
:
5467 case CPU_UP_CANCELED_FROZEN
:
5468 if (!cpu_rq(cpu
)->migration_thread
)
5470 /* Unbind it from offline cpu so it can run. Fall thru. */
5471 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5472 any_online_cpu(cpu_online_map
));
5473 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5474 cpu_rq(cpu
)->migration_thread
= NULL
;
5478 case CPU_DEAD_FROZEN
:
5479 migrate_live_tasks(cpu
);
5481 kthread_stop(rq
->migration_thread
);
5482 rq
->migration_thread
= NULL
;
5483 /* Idle task back to normal (off runqueue, low prio) */
5484 rq
= task_rq_lock(rq
->idle
, &flags
);
5485 deactivate_task(rq
->idle
, rq
);
5486 rq
->idle
->static_prio
= MAX_PRIO
;
5487 __setscheduler(rq
->idle
, SCHED_NORMAL
, 0);
5488 migrate_dead_tasks(cpu
);
5489 task_rq_unlock(rq
, &flags
);
5490 migrate_nr_uninterruptible(rq
);
5491 BUG_ON(rq
->nr_running
!= 0);
5493 /* No need to migrate the tasks: it was best-effort if
5494 * they didn't take sched_hotcpu_mutex. Just wake up
5495 * the requestors. */
5496 spin_lock_irq(&rq
->lock
);
5497 while (!list_empty(&rq
->migration_queue
)) {
5498 struct migration_req
*req
;
5500 req
= list_entry(rq
->migration_queue
.next
,
5501 struct migration_req
, list
);
5502 list_del_init(&req
->list
);
5503 complete(&req
->done
);
5505 spin_unlock_irq(&rq
->lock
);
5508 case CPU_LOCK_RELEASE
:
5509 mutex_unlock(&sched_hotcpu_mutex
);
5515 /* Register at highest priority so that task migration (migrate_all_tasks)
5516 * happens before everything else.
5518 static struct notifier_block __cpuinitdata migration_notifier
= {
5519 .notifier_call
= migration_call
,
5523 int __init
migration_init(void)
5525 void *cpu
= (void *)(long)smp_processor_id();
5528 /* Start one for the boot CPU: */
5529 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5530 BUG_ON(err
== NOTIFY_BAD
);
5531 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5532 register_cpu_notifier(&migration_notifier
);
5540 /* Number of possible processor ids */
5541 int nr_cpu_ids __read_mostly
= NR_CPUS
;
5542 EXPORT_SYMBOL(nr_cpu_ids
);
5544 #undef SCHED_DOMAIN_DEBUG
5545 #ifdef SCHED_DOMAIN_DEBUG
5546 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5551 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5555 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5560 struct sched_group
*group
= sd
->groups
;
5561 cpumask_t groupmask
;
5563 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5564 cpus_clear(groupmask
);
5567 for (i
= 0; i
< level
+ 1; i
++)
5569 printk("domain %d: ", level
);
5571 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5572 printk("does not load-balance\n");
5574 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5579 printk("span %s\n", str
);
5581 if (!cpu_isset(cpu
, sd
->span
))
5582 printk(KERN_ERR
"ERROR: domain->span does not contain "
5584 if (!cpu_isset(cpu
, group
->cpumask
))
5585 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5589 for (i
= 0; i
< level
+ 2; i
++)
5595 printk(KERN_ERR
"ERROR: group is NULL\n");
5599 if (!group
->__cpu_power
) {
5601 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5605 if (!cpus_weight(group
->cpumask
)) {
5607 printk(KERN_ERR
"ERROR: empty group\n");
5610 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5612 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5615 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5617 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5620 group
= group
->next
;
5621 } while (group
!= sd
->groups
);
5624 if (!cpus_equal(sd
->span
, groupmask
))
5625 printk(KERN_ERR
"ERROR: groups don't span "
5633 if (!cpus_subset(groupmask
, sd
->span
))
5634 printk(KERN_ERR
"ERROR: parent span is not a superset "
5635 "of domain->span\n");
5640 # define sched_domain_debug(sd, cpu) do { } while (0)
5643 static int sd_degenerate(struct sched_domain
*sd
)
5645 if (cpus_weight(sd
->span
) == 1)
5648 /* Following flags need at least 2 groups */
5649 if (sd
->flags
& (SD_LOAD_BALANCE
|
5650 SD_BALANCE_NEWIDLE
|
5654 SD_SHARE_PKG_RESOURCES
)) {
5655 if (sd
->groups
!= sd
->groups
->next
)
5659 /* Following flags don't use groups */
5660 if (sd
->flags
& (SD_WAKE_IDLE
|
5669 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5671 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5673 if (sd_degenerate(parent
))
5676 if (!cpus_equal(sd
->span
, parent
->span
))
5679 /* Does parent contain flags not in child? */
5680 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5681 if (cflags
& SD_WAKE_AFFINE
)
5682 pflags
&= ~SD_WAKE_BALANCE
;
5683 /* Flags needing groups don't count if only 1 group in parent */
5684 if (parent
->groups
== parent
->groups
->next
) {
5685 pflags
&= ~(SD_LOAD_BALANCE
|
5686 SD_BALANCE_NEWIDLE
|
5690 SD_SHARE_PKG_RESOURCES
);
5692 if (~cflags
& pflags
)
5699 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5700 * hold the hotplug lock.
5702 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5704 struct rq
*rq
= cpu_rq(cpu
);
5705 struct sched_domain
*tmp
;
5707 /* Remove the sched domains which do not contribute to scheduling. */
5708 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5709 struct sched_domain
*parent
= tmp
->parent
;
5712 if (sd_parent_degenerate(tmp
, parent
)) {
5713 tmp
->parent
= parent
->parent
;
5715 parent
->parent
->child
= tmp
;
5719 if (sd
&& sd_degenerate(sd
)) {
5725 sched_domain_debug(sd
, cpu
);
5727 rcu_assign_pointer(rq
->sd
, sd
);
5730 /* cpus with isolated domains */
5731 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
5733 /* Setup the mask of cpus configured for isolated domains */
5734 static int __init
isolated_cpu_setup(char *str
)
5736 int ints
[NR_CPUS
], i
;
5738 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5739 cpus_clear(cpu_isolated_map
);
5740 for (i
= 1; i
<= ints
[0]; i
++)
5741 if (ints
[i
] < NR_CPUS
)
5742 cpu_set(ints
[i
], cpu_isolated_map
);
5746 __setup ("isolcpus=", isolated_cpu_setup
);
5749 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5750 * to a function which identifies what group(along with sched group) a CPU
5751 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5752 * (due to the fact that we keep track of groups covered with a cpumask_t).
5754 * init_sched_build_groups will build a circular linked list of the groups
5755 * covered by the given span, and will set each group's ->cpumask correctly,
5756 * and ->cpu_power to 0.
5759 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
5760 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
5761 struct sched_group
**sg
))
5763 struct sched_group
*first
= NULL
, *last
= NULL
;
5764 cpumask_t covered
= CPU_MASK_NONE
;
5767 for_each_cpu_mask(i
, span
) {
5768 struct sched_group
*sg
;
5769 int group
= group_fn(i
, cpu_map
, &sg
);
5772 if (cpu_isset(i
, covered
))
5775 sg
->cpumask
= CPU_MASK_NONE
;
5776 sg
->__cpu_power
= 0;
5778 for_each_cpu_mask(j
, span
) {
5779 if (group_fn(j
, cpu_map
, NULL
) != group
)
5782 cpu_set(j
, covered
);
5783 cpu_set(j
, sg
->cpumask
);
5794 #define SD_NODES_PER_DOMAIN 16
5797 * Self-tuning task migration cost measurement between source and target CPUs.
5799 * This is done by measuring the cost of manipulating buffers of varying
5800 * sizes. For a given buffer-size here are the steps that are taken:
5802 * 1) the source CPU reads+dirties a shared buffer
5803 * 2) the target CPU reads+dirties the same shared buffer
5805 * We measure how long they take, in the following 4 scenarios:
5807 * - source: CPU1, target: CPU2 | cost1
5808 * - source: CPU2, target: CPU1 | cost2
5809 * - source: CPU1, target: CPU1 | cost3
5810 * - source: CPU2, target: CPU2 | cost4
5812 * We then calculate the cost3+cost4-cost1-cost2 difference - this is
5813 * the cost of migration.
5815 * We then start off from a small buffer-size and iterate up to larger
5816 * buffer sizes, in 5% steps - measuring each buffer-size separately, and
5817 * doing a maximum search for the cost. (The maximum cost for a migration
5818 * normally occurs when the working set size is around the effective cache
5821 #define SEARCH_SCOPE 2
5822 #define MIN_CACHE_SIZE (64*1024U)
5823 #define DEFAULT_CACHE_SIZE (5*1024*1024U)
5824 #define ITERATIONS 1
5825 #define SIZE_THRESH 130
5826 #define COST_THRESH 130
5829 * The migration cost is a function of 'domain distance'. Domain
5830 * distance is the number of steps a CPU has to iterate down its
5831 * domain tree to share a domain with the other CPU. The farther
5832 * two CPUs are from each other, the larger the distance gets.
5834 * Note that we use the distance only to cache measurement results,
5835 * the distance value is not used numerically otherwise. When two
5836 * CPUs have the same distance it is assumed that the migration
5837 * cost is the same. (this is a simplification but quite practical)
5839 #define MAX_DOMAIN_DISTANCE 32
5841 static unsigned long long migration_cost
[MAX_DOMAIN_DISTANCE
] =
5842 { [ 0 ... MAX_DOMAIN_DISTANCE
-1 ] =
5844 * Architectures may override the migration cost and thus avoid
5845 * boot-time calibration. Unit is nanoseconds. Mostly useful for
5846 * virtualized hardware:
5848 #ifdef CONFIG_DEFAULT_MIGRATION_COST
5849 CONFIG_DEFAULT_MIGRATION_COST
5856 * Allow override of migration cost - in units of microseconds.
5857 * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost
5858 * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs:
5860 static int __init
migration_cost_setup(char *str
)
5862 int ints
[MAX_DOMAIN_DISTANCE
+1], i
;
5864 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5866 printk("#ints: %d\n", ints
[0]);
5867 for (i
= 1; i
<= ints
[0]; i
++) {
5868 migration_cost
[i
-1] = (unsigned long long)ints
[i
]*1000;
5869 printk("migration_cost[%d]: %Ld\n", i
-1, migration_cost
[i
-1]);
5874 __setup ("migration_cost=", migration_cost_setup
);
5877 * Global multiplier (divisor) for migration-cutoff values,
5878 * in percentiles. E.g. use a value of 150 to get 1.5 times
5879 * longer cache-hot cutoff times.
5881 * (We scale it from 100 to 128 to long long handling easier.)
5884 #define MIGRATION_FACTOR_SCALE 128
5886 static unsigned int migration_factor
= MIGRATION_FACTOR_SCALE
;
5888 static int __init
setup_migration_factor(char *str
)
5890 get_option(&str
, &migration_factor
);
5891 migration_factor
= migration_factor
* MIGRATION_FACTOR_SCALE
/ 100;
5895 __setup("migration_factor=", setup_migration_factor
);
5898 * Estimated distance of two CPUs, measured via the number of domains
5899 * we have to pass for the two CPUs to be in the same span:
5901 static unsigned long domain_distance(int cpu1
, int cpu2
)
5903 unsigned long distance
= 0;
5904 struct sched_domain
*sd
;
5906 for_each_domain(cpu1
, sd
) {
5907 WARN_ON(!cpu_isset(cpu1
, sd
->span
));
5908 if (cpu_isset(cpu2
, sd
->span
))
5912 if (distance
>= MAX_DOMAIN_DISTANCE
) {
5914 distance
= MAX_DOMAIN_DISTANCE
-1;
5920 static unsigned int migration_debug
;
5922 static int __init
setup_migration_debug(char *str
)
5924 get_option(&str
, &migration_debug
);
5928 __setup("migration_debug=", setup_migration_debug
);
5931 * Maximum cache-size that the scheduler should try to measure.
5932 * Architectures with larger caches should tune this up during
5933 * bootup. Gets used in the domain-setup code (i.e. during SMP
5936 unsigned int max_cache_size
;
5938 static int __init
setup_max_cache_size(char *str
)
5940 get_option(&str
, &max_cache_size
);
5944 __setup("max_cache_size=", setup_max_cache_size
);
5947 * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This
5948 * is the operation that is timed, so we try to generate unpredictable
5949 * cachemisses that still end up filling the L2 cache:
5951 static void touch_cache(void *__cache
, unsigned long __size
)
5953 unsigned long size
= __size
/ sizeof(long);
5954 unsigned long chunk1
= size
/ 3;
5955 unsigned long chunk2
= 2 * size
/ 3;
5956 unsigned long *cache
= __cache
;
5959 for (i
= 0; i
< size
/6; i
+= 8) {
5962 case 1: cache
[size
-1-i
]++;
5963 case 2: cache
[chunk1
-i
]++;
5964 case 3: cache
[chunk1
+i
]++;
5965 case 4: cache
[chunk2
-i
]++;
5966 case 5: cache
[chunk2
+i
]++;
5972 * Measure the cache-cost of one task migration. Returns in units of nsec.
5974 static unsigned long long
5975 measure_one(void *cache
, unsigned long size
, int source
, int target
)
5977 cpumask_t mask
, saved_mask
;
5978 unsigned long long t0
, t1
, t2
, t3
, cost
;
5980 saved_mask
= current
->cpus_allowed
;
5983 * Flush source caches to RAM and invalidate them:
5988 * Migrate to the source CPU:
5990 mask
= cpumask_of_cpu(source
);
5991 set_cpus_allowed(current
, mask
);
5992 WARN_ON(smp_processor_id() != source
);
5995 * Dirty the working set:
5998 touch_cache(cache
, size
);
6002 * Migrate to the target CPU, dirty the L2 cache and access
6003 * the shared buffer. (which represents the working set
6004 * of a migrated task.)
6006 mask
= cpumask_of_cpu(target
);
6007 set_cpus_allowed(current
, mask
);
6008 WARN_ON(smp_processor_id() != target
);
6011 touch_cache(cache
, size
);
6014 cost
= t1
-t0
+ t3
-t2
;
6016 if (migration_debug
>= 2)
6017 printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n",
6018 source
, target
, t1
-t0
, t1
-t0
, t3
-t2
, cost
);
6020 * Flush target caches to RAM and invalidate them:
6024 set_cpus_allowed(current
, saved_mask
);
6030 * Measure a series of task migrations and return the average
6031 * result. Since this code runs early during bootup the system
6032 * is 'undisturbed' and the average latency makes sense.
6034 * The algorithm in essence auto-detects the relevant cache-size,
6035 * so it will properly detect different cachesizes for different
6036 * cache-hierarchies, depending on how the CPUs are connected.
6038 * Architectures can prime the upper limit of the search range via
6039 * max_cache_size, otherwise the search range defaults to 20MB...64K.
6041 static unsigned long long
6042 measure_cost(int cpu1
, int cpu2
, void *cache
, unsigned int size
)
6044 unsigned long long cost1
, cost2
;
6048 * Measure the migration cost of 'size' bytes, over an
6049 * average of 10 runs:
6051 * (We perturb the cache size by a small (0..4k)
6052 * value to compensate size/alignment related artifacts.
6053 * We also subtract the cost of the operation done on
6059 * dry run, to make sure we start off cache-cold on cpu1,
6060 * and to get any vmalloc pagefaults in advance:
6062 measure_one(cache
, size
, cpu1
, cpu2
);
6063 for (i
= 0; i
< ITERATIONS
; i
++)
6064 cost1
+= measure_one(cache
, size
- i
* 1024, cpu1
, cpu2
);
6066 measure_one(cache
, size
, cpu2
, cpu1
);
6067 for (i
= 0; i
< ITERATIONS
; i
++)
6068 cost1
+= measure_one(cache
, size
- i
* 1024, cpu2
, cpu1
);
6071 * (We measure the non-migrating [cached] cost on both
6072 * cpu1 and cpu2, to handle CPUs with different speeds)
6076 measure_one(cache
, size
, cpu1
, cpu1
);
6077 for (i
= 0; i
< ITERATIONS
; i
++)
6078 cost2
+= measure_one(cache
, size
- i
* 1024, cpu1
, cpu1
);
6080 measure_one(cache
, size
, cpu2
, cpu2
);
6081 for (i
= 0; i
< ITERATIONS
; i
++)
6082 cost2
+= measure_one(cache
, size
- i
* 1024, cpu2
, cpu2
);
6085 * Get the per-iteration migration cost:
6087 do_div(cost1
, 2 * ITERATIONS
);
6088 do_div(cost2
, 2 * ITERATIONS
);
6090 return cost1
- cost2
;
6093 static unsigned long long measure_migration_cost(int cpu1
, int cpu2
)
6095 unsigned long long max_cost
= 0, fluct
= 0, avg_fluct
= 0;
6096 unsigned int max_size
, size
, size_found
= 0;
6097 long long cost
= 0, prev_cost
;
6101 * Search from max_cache_size*5 down to 64K - the real relevant
6102 * cachesize has to lie somewhere inbetween.
6104 if (max_cache_size
) {
6105 max_size
= max(max_cache_size
* SEARCH_SCOPE
, MIN_CACHE_SIZE
);
6106 size
= max(max_cache_size
/ SEARCH_SCOPE
, MIN_CACHE_SIZE
);
6109 * Since we have no estimation about the relevant
6112 max_size
= DEFAULT_CACHE_SIZE
* SEARCH_SCOPE
;
6113 size
= MIN_CACHE_SIZE
;
6116 if (!cpu_online(cpu1
) || !cpu_online(cpu2
)) {
6117 printk("cpu %d and %d not both online!\n", cpu1
, cpu2
);
6122 * Allocate the working set:
6124 cache
= vmalloc(max_size
);
6126 printk("could not vmalloc %d bytes for cache!\n", 2 * max_size
);
6127 return 1000000; /* return 1 msec on very small boxen */
6130 while (size
<= max_size
) {
6132 cost
= measure_cost(cpu1
, cpu2
, cache
, size
);
6138 if (max_cost
< cost
) {
6144 * Calculate average fluctuation, we use this to prevent
6145 * noise from triggering an early break out of the loop:
6147 fluct
= abs(cost
- prev_cost
);
6148 avg_fluct
= (avg_fluct
+ fluct
)/2;
6150 if (migration_debug
)
6151 printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): "
6154 (long)cost
/ 1000000,
6155 ((long)cost
/ 100000) % 10,
6156 (long)max_cost
/ 1000000,
6157 ((long)max_cost
/ 100000) % 10,
6158 domain_distance(cpu1
, cpu2
),
6162 * If we iterated at least 20% past the previous maximum,
6163 * and the cost has dropped by more than 20% already,
6164 * (taking fluctuations into account) then we assume to
6165 * have found the maximum and break out of the loop early:
6167 if (size_found
&& (size
*100 > size_found
*SIZE_THRESH
))
6168 if (cost
+avg_fluct
<= 0 ||
6169 max_cost
*100 > (cost
+avg_fluct
)*COST_THRESH
) {
6171 if (migration_debug
)
6172 printk("-> found max.\n");
6176 * Increase the cachesize in 10% steps:
6178 size
= size
* 10 / 9;
6181 if (migration_debug
)
6182 printk("[%d][%d] working set size found: %d, cost: %Ld\n",
6183 cpu1
, cpu2
, size_found
, max_cost
);
6188 * A task is considered 'cache cold' if at least 2 times
6189 * the worst-case cost of migration has passed.
6191 * (this limit is only listened to if the load-balancing
6192 * situation is 'nice' - if there is a large imbalance we
6193 * ignore it for the sake of CPU utilization and
6194 * processing fairness.)
6196 return 2 * max_cost
* migration_factor
/ MIGRATION_FACTOR_SCALE
;
6199 static void calibrate_migration_costs(const cpumask_t
*cpu_map
)
6201 int cpu1
= -1, cpu2
= -1, cpu
, orig_cpu
= raw_smp_processor_id();
6202 unsigned long j0
, j1
, distance
, max_distance
= 0;
6203 struct sched_domain
*sd
;
6208 * First pass - calculate the cacheflush times:
6210 for_each_cpu_mask(cpu1
, *cpu_map
) {
6211 for_each_cpu_mask(cpu2
, *cpu_map
) {
6214 distance
= domain_distance(cpu1
, cpu2
);
6215 max_distance
= max(max_distance
, distance
);
6217 * No result cached yet?
6219 if (migration_cost
[distance
] == -1LL)
6220 migration_cost
[distance
] =
6221 measure_migration_cost(cpu1
, cpu2
);
6225 * Second pass - update the sched domain hierarchy with
6226 * the new cache-hot-time estimations:
6228 for_each_cpu_mask(cpu
, *cpu_map
) {
6230 for_each_domain(cpu
, sd
) {
6231 sd
->cache_hot_time
= migration_cost
[distance
];
6238 if (migration_debug
)
6239 printk("migration: max_cache_size: %d, cpu: %d MHz:\n",
6247 if (system_state
== SYSTEM_BOOTING
&& num_online_cpus() > 1) {
6248 printk("migration_cost=");
6249 for (distance
= 0; distance
<= max_distance
; distance
++) {
6252 printk("%ld", (long)migration_cost
[distance
] / 1000);
6257 if (migration_debug
)
6258 printk("migration: %ld seconds\n", (j1
-j0
) / HZ
);
6261 * Move back to the original CPU. NUMA-Q gets confused
6262 * if we migrate to another quad during bootup.
6264 if (raw_smp_processor_id() != orig_cpu
) {
6265 cpumask_t mask
= cpumask_of_cpu(orig_cpu
),
6266 saved_mask
= current
->cpus_allowed
;
6268 set_cpus_allowed(current
, mask
);
6269 set_cpus_allowed(current
, saved_mask
);
6276 * find_next_best_node - find the next node to include in a sched_domain
6277 * @node: node whose sched_domain we're building
6278 * @used_nodes: nodes already in the sched_domain
6280 * Find the next node to include in a given scheduling domain. Simply
6281 * finds the closest node not already in the @used_nodes map.
6283 * Should use nodemask_t.
6285 static int find_next_best_node(int node
, unsigned long *used_nodes
)
6287 int i
, n
, val
, min_val
, best_node
= 0;
6291 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6292 /* Start at @node */
6293 n
= (node
+ i
) % MAX_NUMNODES
;
6295 if (!nr_cpus_node(n
))
6298 /* Skip already used nodes */
6299 if (test_bit(n
, used_nodes
))
6302 /* Simple min distance search */
6303 val
= node_distance(node
, n
);
6305 if (val
< min_val
) {
6311 set_bit(best_node
, used_nodes
);
6316 * sched_domain_node_span - get a cpumask for a node's sched_domain
6317 * @node: node whose cpumask we're constructing
6318 * @size: number of nodes to include in this span
6320 * Given a node, construct a good cpumask for its sched_domain to span. It
6321 * should be one that prevents unnecessary balancing, but also spreads tasks
6324 static cpumask_t
sched_domain_node_span(int node
)
6326 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
6327 cpumask_t span
, nodemask
;
6331 bitmap_zero(used_nodes
, MAX_NUMNODES
);
6333 nodemask
= node_to_cpumask(node
);
6334 cpus_or(span
, span
, nodemask
);
6335 set_bit(node
, used_nodes
);
6337 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6338 int next_node
= find_next_best_node(node
, used_nodes
);
6340 nodemask
= node_to_cpumask(next_node
);
6341 cpus_or(span
, span
, nodemask
);
6348 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6351 * SMT sched-domains:
6353 #ifdef CONFIG_SCHED_SMT
6354 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
6355 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
6357 static int cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
,
6358 struct sched_group
**sg
)
6361 *sg
= &per_cpu(sched_group_cpus
, cpu
);
6367 * multi-core sched-domains:
6369 #ifdef CONFIG_SCHED_MC
6370 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
6371 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
6374 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6375 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
6376 struct sched_group
**sg
)
6379 cpumask_t mask
= cpu_sibling_map
[cpu
];
6380 cpus_and(mask
, mask
, *cpu_map
);
6381 group
= first_cpu(mask
);
6383 *sg
= &per_cpu(sched_group_core
, group
);
6386 #elif defined(CONFIG_SCHED_MC)
6387 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
6388 struct sched_group
**sg
)
6391 *sg
= &per_cpu(sched_group_core
, cpu
);
6396 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
6397 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
6399 static int cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
,
6400 struct sched_group
**sg
)
6403 #ifdef CONFIG_SCHED_MC
6404 cpumask_t mask
= cpu_coregroup_map(cpu
);
6405 cpus_and(mask
, mask
, *cpu_map
);
6406 group
= first_cpu(mask
);
6407 #elif defined(CONFIG_SCHED_SMT)
6408 cpumask_t mask
= cpu_sibling_map
[cpu
];
6409 cpus_and(mask
, mask
, *cpu_map
);
6410 group
= first_cpu(mask
);
6415 *sg
= &per_cpu(sched_group_phys
, group
);
6421 * The init_sched_build_groups can't handle what we want to do with node
6422 * groups, so roll our own. Now each node has its own list of groups which
6423 * gets dynamically allocated.
6425 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
6426 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
6428 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
6429 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
6431 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
6432 struct sched_group
**sg
)
6434 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
6437 cpus_and(nodemask
, nodemask
, *cpu_map
);
6438 group
= first_cpu(nodemask
);
6441 *sg
= &per_cpu(sched_group_allnodes
, group
);
6445 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6447 struct sched_group
*sg
= group_head
;
6453 for_each_cpu_mask(j
, sg
->cpumask
) {
6454 struct sched_domain
*sd
;
6456 sd
= &per_cpu(phys_domains
, j
);
6457 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
6459 * Only add "power" once for each
6465 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
6468 if (sg
!= group_head
)
6474 /* Free memory allocated for various sched_group structures */
6475 static void free_sched_groups(const cpumask_t
*cpu_map
)
6479 for_each_cpu_mask(cpu
, *cpu_map
) {
6480 struct sched_group
**sched_group_nodes
6481 = sched_group_nodes_bycpu
[cpu
];
6483 if (!sched_group_nodes
)
6486 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6487 cpumask_t nodemask
= node_to_cpumask(i
);
6488 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6490 cpus_and(nodemask
, nodemask
, *cpu_map
);
6491 if (cpus_empty(nodemask
))
6501 if (oldsg
!= sched_group_nodes
[i
])
6504 kfree(sched_group_nodes
);
6505 sched_group_nodes_bycpu
[cpu
] = NULL
;
6509 static void free_sched_groups(const cpumask_t
*cpu_map
)
6515 * Initialize sched groups cpu_power.
6517 * cpu_power indicates the capacity of sched group, which is used while
6518 * distributing the load between different sched groups in a sched domain.
6519 * Typically cpu_power for all the groups in a sched domain will be same unless
6520 * there are asymmetries in the topology. If there are asymmetries, group
6521 * having more cpu_power will pickup more load compared to the group having
6524 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6525 * the maximum number of tasks a group can handle in the presence of other idle
6526 * or lightly loaded groups in the same sched domain.
6528 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6530 struct sched_domain
*child
;
6531 struct sched_group
*group
;
6533 WARN_ON(!sd
|| !sd
->groups
);
6535 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
6540 sd
->groups
->__cpu_power
= 0;
6543 * For perf policy, if the groups in child domain share resources
6544 * (for example cores sharing some portions of the cache hierarchy
6545 * or SMT), then set this domain groups cpu_power such that each group
6546 * can handle only one task, when there are other idle groups in the
6547 * same sched domain.
6549 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
6551 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
6552 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
6557 * add cpu_power of each child group to this groups cpu_power
6559 group
= child
->groups
;
6561 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
6562 group
= group
->next
;
6563 } while (group
!= child
->groups
);
6567 * Build sched domains for a given set of cpus and attach the sched domains
6568 * to the individual cpus
6570 static int build_sched_domains(const cpumask_t
*cpu_map
)
6573 struct sched_domain
*sd
;
6575 struct sched_group
**sched_group_nodes
= NULL
;
6576 int sd_allnodes
= 0;
6579 * Allocate the per-node list of sched groups
6581 sched_group_nodes
= kzalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
6583 if (!sched_group_nodes
) {
6584 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6587 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6591 * Set up domains for cpus specified by the cpu_map.
6593 for_each_cpu_mask(i
, *cpu_map
) {
6594 struct sched_domain
*sd
= NULL
, *p
;
6595 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6597 cpus_and(nodemask
, nodemask
, *cpu_map
);
6600 if (cpus_weight(*cpu_map
)
6601 > SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6602 sd
= &per_cpu(allnodes_domains
, i
);
6603 *sd
= SD_ALLNODES_INIT
;
6604 sd
->span
= *cpu_map
;
6605 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6611 sd
= &per_cpu(node_domains
, i
);
6613 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6617 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6621 sd
= &per_cpu(phys_domains
, i
);
6623 sd
->span
= nodemask
;
6627 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6629 #ifdef CONFIG_SCHED_MC
6631 sd
= &per_cpu(core_domains
, i
);
6633 sd
->span
= cpu_coregroup_map(i
);
6634 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6637 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6640 #ifdef CONFIG_SCHED_SMT
6642 sd
= &per_cpu(cpu_domains
, i
);
6643 *sd
= SD_SIBLING_INIT
;
6644 sd
->span
= cpu_sibling_map
[i
];
6645 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6648 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6652 #ifdef CONFIG_SCHED_SMT
6653 /* Set up CPU (sibling) groups */
6654 for_each_cpu_mask(i
, *cpu_map
) {
6655 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
6656 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6657 if (i
!= first_cpu(this_sibling_map
))
6660 init_sched_build_groups(this_sibling_map
, cpu_map
, &cpu_to_cpu_group
);
6664 #ifdef CONFIG_SCHED_MC
6665 /* Set up multi-core groups */
6666 for_each_cpu_mask(i
, *cpu_map
) {
6667 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6668 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6669 if (i
!= first_cpu(this_core_map
))
6671 init_sched_build_groups(this_core_map
, cpu_map
, &cpu_to_core_group
);
6676 /* Set up physical groups */
6677 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6678 cpumask_t nodemask
= node_to_cpumask(i
);
6680 cpus_and(nodemask
, nodemask
, *cpu_map
);
6681 if (cpus_empty(nodemask
))
6684 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6688 /* Set up node groups */
6690 init_sched_build_groups(*cpu_map
, cpu_map
, &cpu_to_allnodes_group
);
6692 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6693 /* Set up node groups */
6694 struct sched_group
*sg
, *prev
;
6695 cpumask_t nodemask
= node_to_cpumask(i
);
6696 cpumask_t domainspan
;
6697 cpumask_t covered
= CPU_MASK_NONE
;
6700 cpus_and(nodemask
, nodemask
, *cpu_map
);
6701 if (cpus_empty(nodemask
)) {
6702 sched_group_nodes
[i
] = NULL
;
6706 domainspan
= sched_domain_node_span(i
);
6707 cpus_and(domainspan
, domainspan
, *cpu_map
);
6709 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6711 printk(KERN_WARNING
"Can not alloc domain group for "
6715 sched_group_nodes
[i
] = sg
;
6716 for_each_cpu_mask(j
, nodemask
) {
6717 struct sched_domain
*sd
;
6718 sd
= &per_cpu(node_domains
, j
);
6721 sg
->__cpu_power
= 0;
6722 sg
->cpumask
= nodemask
;
6724 cpus_or(covered
, covered
, nodemask
);
6727 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6728 cpumask_t tmp
, notcovered
;
6729 int n
= (i
+ j
) % MAX_NUMNODES
;
6731 cpus_complement(notcovered
, covered
);
6732 cpus_and(tmp
, notcovered
, *cpu_map
);
6733 cpus_and(tmp
, tmp
, domainspan
);
6734 if (cpus_empty(tmp
))
6737 nodemask
= node_to_cpumask(n
);
6738 cpus_and(tmp
, tmp
, nodemask
);
6739 if (cpus_empty(tmp
))
6742 sg
= kmalloc_node(sizeof(struct sched_group
),
6746 "Can not alloc domain group for node %d\n", j
);
6749 sg
->__cpu_power
= 0;
6751 sg
->next
= prev
->next
;
6752 cpus_or(covered
, covered
, tmp
);
6759 /* Calculate CPU power for physical packages and nodes */
6760 #ifdef CONFIG_SCHED_SMT
6761 for_each_cpu_mask(i
, *cpu_map
) {
6762 sd
= &per_cpu(cpu_domains
, i
);
6763 init_sched_groups_power(i
, sd
);
6766 #ifdef CONFIG_SCHED_MC
6767 for_each_cpu_mask(i
, *cpu_map
) {
6768 sd
= &per_cpu(core_domains
, i
);
6769 init_sched_groups_power(i
, sd
);
6773 for_each_cpu_mask(i
, *cpu_map
) {
6774 sd
= &per_cpu(phys_domains
, i
);
6775 init_sched_groups_power(i
, sd
);
6779 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6780 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6783 struct sched_group
*sg
;
6785 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6786 init_numa_sched_groups_power(sg
);
6790 /* Attach the domains */
6791 for_each_cpu_mask(i
, *cpu_map
) {
6792 struct sched_domain
*sd
;
6793 #ifdef CONFIG_SCHED_SMT
6794 sd
= &per_cpu(cpu_domains
, i
);
6795 #elif defined(CONFIG_SCHED_MC)
6796 sd
= &per_cpu(core_domains
, i
);
6798 sd
= &per_cpu(phys_domains
, i
);
6800 cpu_attach_domain(sd
, i
);
6803 * Tune cache-hot values:
6805 calibrate_migration_costs(cpu_map
);
6811 free_sched_groups(cpu_map
);
6816 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6818 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6820 cpumask_t cpu_default_map
;
6824 * Setup mask for cpus without special case scheduling requirements.
6825 * For now this just excludes isolated cpus, but could be used to
6826 * exclude other special cases in the future.
6828 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
6830 err
= build_sched_domains(&cpu_default_map
);
6835 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6837 free_sched_groups(cpu_map
);
6841 * Detach sched domains from a group of cpus specified in cpu_map
6842 * These cpus will now be attached to the NULL domain
6844 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6848 for_each_cpu_mask(i
, *cpu_map
)
6849 cpu_attach_domain(NULL
, i
);
6850 synchronize_sched();
6851 arch_destroy_sched_domains(cpu_map
);
6855 * Partition sched domains as specified by the cpumasks below.
6856 * This attaches all cpus from the cpumasks to the NULL domain,
6857 * waits for a RCU quiescent period, recalculates sched
6858 * domain information and then attaches them back to the
6859 * correct sched domains
6860 * Call with hotplug lock held
6862 int partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
6864 cpumask_t change_map
;
6867 cpus_and(*partition1
, *partition1
, cpu_online_map
);
6868 cpus_and(*partition2
, *partition2
, cpu_online_map
);
6869 cpus_or(change_map
, *partition1
, *partition2
);
6871 /* Detach sched domains from all of the affected cpus */
6872 detach_destroy_domains(&change_map
);
6873 if (!cpus_empty(*partition1
))
6874 err
= build_sched_domains(partition1
);
6875 if (!err
&& !cpus_empty(*partition2
))
6876 err
= build_sched_domains(partition2
);
6881 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6882 int arch_reinit_sched_domains(void)
6886 mutex_lock(&sched_hotcpu_mutex
);
6887 detach_destroy_domains(&cpu_online_map
);
6888 err
= arch_init_sched_domains(&cpu_online_map
);
6889 mutex_unlock(&sched_hotcpu_mutex
);
6894 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6898 if (buf
[0] != '0' && buf
[0] != '1')
6902 sched_smt_power_savings
= (buf
[0] == '1');
6904 sched_mc_power_savings
= (buf
[0] == '1');
6906 ret
= arch_reinit_sched_domains();
6908 return ret
? ret
: count
;
6911 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6915 #ifdef CONFIG_SCHED_SMT
6917 err
= sysfs_create_file(&cls
->kset
.kobj
,
6918 &attr_sched_smt_power_savings
.attr
);
6920 #ifdef CONFIG_SCHED_MC
6921 if (!err
&& mc_capable())
6922 err
= sysfs_create_file(&cls
->kset
.kobj
,
6923 &attr_sched_mc_power_savings
.attr
);
6929 #ifdef CONFIG_SCHED_MC
6930 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6932 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6934 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6935 const char *buf
, size_t count
)
6937 return sched_power_savings_store(buf
, count
, 0);
6939 SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6940 sched_mc_power_savings_store
);
6943 #ifdef CONFIG_SCHED_SMT
6944 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6946 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6948 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6949 const char *buf
, size_t count
)
6951 return sched_power_savings_store(buf
, count
, 1);
6953 SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6954 sched_smt_power_savings_store
);
6958 * Force a reinitialization of the sched domains hierarchy. The domains
6959 * and groups cannot be updated in place without racing with the balancing
6960 * code, so we temporarily attach all running cpus to the NULL domain
6961 * which will prevent rebalancing while the sched domains are recalculated.
6963 static int update_sched_domains(struct notifier_block
*nfb
,
6964 unsigned long action
, void *hcpu
)
6967 case CPU_UP_PREPARE
:
6968 case CPU_UP_PREPARE_FROZEN
:
6969 case CPU_DOWN_PREPARE
:
6970 case CPU_DOWN_PREPARE_FROZEN
:
6971 detach_destroy_domains(&cpu_online_map
);
6974 case CPU_UP_CANCELED
:
6975 case CPU_UP_CANCELED_FROZEN
:
6976 case CPU_DOWN_FAILED
:
6977 case CPU_DOWN_FAILED_FROZEN
:
6979 case CPU_ONLINE_FROZEN
:
6981 case CPU_DEAD_FROZEN
:
6983 * Fall through and re-initialise the domains.
6990 /* The hotplug lock is already held by cpu_up/cpu_down */
6991 arch_init_sched_domains(&cpu_online_map
);
6996 void __init
sched_init_smp(void)
6998 cpumask_t non_isolated_cpus
;
7000 mutex_lock(&sched_hotcpu_mutex
);
7001 arch_init_sched_domains(&cpu_online_map
);
7002 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
7003 if (cpus_empty(non_isolated_cpus
))
7004 cpu_set(smp_processor_id(), non_isolated_cpus
);
7005 mutex_unlock(&sched_hotcpu_mutex
);
7006 /* XXX: Theoretical race here - CPU may be hotplugged now */
7007 hotcpu_notifier(update_sched_domains
, 0);
7009 /* Move init over to a non-isolated CPU */
7010 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
7014 void __init
sched_init_smp(void)
7017 #endif /* CONFIG_SMP */
7019 int in_sched_functions(unsigned long addr
)
7021 /* Linker adds these: start and end of __sched functions */
7022 extern char __sched_text_start
[], __sched_text_end
[];
7024 return in_lock_functions(addr
) ||
7025 (addr
>= (unsigned long)__sched_text_start
7026 && addr
< (unsigned long)__sched_text_end
);
7029 void __init
sched_init(void)
7032 int highest_cpu
= 0;
7034 for_each_possible_cpu(i
) {
7035 struct prio_array
*array
;
7039 spin_lock_init(&rq
->lock
);
7040 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
7042 rq
->active
= rq
->arrays
;
7043 rq
->expired
= rq
->arrays
+ 1;
7044 rq
->best_expired_prio
= MAX_PRIO
;
7048 for (j
= 1; j
< 3; j
++)
7049 rq
->cpu_load
[j
] = 0;
7050 rq
->active_balance
= 0;
7053 rq
->migration_thread
= NULL
;
7054 INIT_LIST_HEAD(&rq
->migration_queue
);
7056 atomic_set(&rq
->nr_iowait
, 0);
7058 for (j
= 0; j
< 2; j
++) {
7059 array
= rq
->arrays
+ j
;
7060 for (k
= 0; k
< MAX_PRIO
; k
++) {
7061 INIT_LIST_HEAD(array
->queue
+ k
);
7062 __clear_bit(k
, array
->bitmap
);
7064 // delimiter for bitsearch
7065 __set_bit(MAX_PRIO
, array
->bitmap
);
7070 set_load_weight(&init_task
);
7073 nr_cpu_ids
= highest_cpu
+ 1;
7074 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
7077 #ifdef CONFIG_RT_MUTEXES
7078 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
7082 * The boot idle thread does lazy MMU switching as well:
7084 atomic_inc(&init_mm
.mm_count
);
7085 enter_lazy_tlb(&init_mm
, current
);
7088 * Make us the idle thread. Technically, schedule() should not be
7089 * called from this thread, however somewhere below it might be,
7090 * but because we are the idle thread, we just pick up running again
7091 * when this runqueue becomes "idle".
7093 init_idle(current
, smp_processor_id());
7096 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7097 void __might_sleep(char *file
, int line
)
7100 static unsigned long prev_jiffy
; /* ratelimiting */
7102 if ((in_atomic() || irqs_disabled()) &&
7103 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
7104 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7106 prev_jiffy
= jiffies
;
7107 printk(KERN_ERR
"BUG: sleeping function called from invalid"
7108 " context at %s:%d\n", file
, line
);
7109 printk("in_atomic():%d, irqs_disabled():%d\n",
7110 in_atomic(), irqs_disabled());
7111 debug_show_held_locks(current
);
7112 if (irqs_disabled())
7113 print_irqtrace_events(current
);
7118 EXPORT_SYMBOL(__might_sleep
);
7121 #ifdef CONFIG_MAGIC_SYSRQ
7122 void normalize_rt_tasks(void)
7124 struct prio_array
*array
;
7125 struct task_struct
*g
, *p
;
7126 unsigned long flags
;
7129 read_lock_irq(&tasklist_lock
);
7131 do_each_thread(g
, p
) {
7135 spin_lock_irqsave(&p
->pi_lock
, flags
);
7136 rq
= __task_rq_lock(p
);
7140 deactivate_task(p
, task_rq(p
));
7141 __setscheduler(p
, SCHED_NORMAL
, 0);
7143 __activate_task(p
, task_rq(p
));
7144 resched_task(rq
->curr
);
7147 __task_rq_unlock(rq
);
7148 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
7149 } while_each_thread(g
, p
);
7151 read_unlock_irq(&tasklist_lock
);
7154 #endif /* CONFIG_MAGIC_SYSRQ */
7158 * These functions are only useful for the IA64 MCA handling.
7160 * They can only be called when the whole system has been
7161 * stopped - every CPU needs to be quiescent, and no scheduling
7162 * activity can take place. Using them for anything else would
7163 * be a serious bug, and as a result, they aren't even visible
7164 * under any other configuration.
7168 * curr_task - return the current task for a given cpu.
7169 * @cpu: the processor in question.
7171 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7173 struct task_struct
*curr_task(int cpu
)
7175 return cpu_curr(cpu
);
7179 * set_curr_task - set the current task for a given cpu.
7180 * @cpu: the processor in question.
7181 * @p: the task pointer to set.
7183 * Description: This function must only be used when non-maskable interrupts
7184 * are serviced on a separate stack. It allows the architecture to switch the
7185 * notion of the current task on a cpu in a non-blocking manner. This function
7186 * must be called with all CPU's synchronized, and interrupts disabled, the
7187 * and caller must save the original value of the current task (see
7188 * curr_task() above) and restore that value before reenabling interrupts and
7189 * re-starting the system.
7191 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7193 void set_curr_task(int cpu
, struct task_struct
*p
)