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>
57 #include <asm/unistd.h>
60 * Convert user-nice values [ -20 ... 0 ... 19 ]
61 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
64 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
65 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
66 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
69 * 'User priority' is the nice value converted to something we
70 * can work with better when scaling various scheduler parameters,
71 * it's a [ 0 ... 39 ] range.
73 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
74 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
75 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
78 * Some helpers for converting nanosecond timing to jiffy resolution
80 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
81 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
84 * These are the 'tuning knobs' of the scheduler:
86 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
87 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
88 * Timeslices get refilled after they expire.
90 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
91 #define DEF_TIMESLICE (100 * HZ / 1000)
92 #define ON_RUNQUEUE_WEIGHT 30
93 #define CHILD_PENALTY 95
94 #define PARENT_PENALTY 100
96 #define PRIO_BONUS_RATIO 25
97 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
98 #define INTERACTIVE_DELTA 2
99 #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
100 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
101 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
104 * If a task is 'interactive' then we reinsert it in the active
105 * array after it has expired its current timeslice. (it will not
106 * continue to run immediately, it will still roundrobin with
107 * other interactive tasks.)
109 * This part scales the interactivity limit depending on niceness.
111 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
112 * Here are a few examples of different nice levels:
114 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
115 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
116 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
117 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
118 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
120 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
121 * priority range a task can explore, a value of '1' means the
122 * task is rated interactive.)
124 * Ie. nice +19 tasks can never get 'interactive' enough to be
125 * reinserted into the active array. And only heavily CPU-hog nice -20
126 * tasks will be expired. Default nice 0 tasks are somewhere between,
127 * it takes some effort for them to get interactive, but it's not
131 #define CURRENT_BONUS(p) \
132 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
135 #define GRANULARITY (10 * HZ / 1000 ? : 1)
138 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
139 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
142 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
143 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
146 #define SCALE(v1,v1_max,v2_max) \
147 (v1) * (v2_max) / (v1_max)
150 (SCALE(TASK_NICE(p) + 20, 40, MAX_BONUS) - 20 * MAX_BONUS / 40 + \
153 #define TASK_INTERACTIVE(p) \
154 ((p)->prio <= (p)->static_prio - DELTA(p))
156 #define INTERACTIVE_SLEEP(p) \
157 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
158 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
160 #define TASK_PREEMPTS_CURR(p, rq) \
161 ((p)->prio < (rq)->curr->prio)
163 #define SCALE_PRIO(x, prio) \
164 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
166 static unsigned int static_prio_timeslice(int static_prio
)
168 if (static_prio
< NICE_TO_PRIO(0))
169 return SCALE_PRIO(DEF_TIMESLICE
* 4, static_prio
);
171 return SCALE_PRIO(DEF_TIMESLICE
, static_prio
);
175 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
176 * to time slice values: [800ms ... 100ms ... 5ms]
178 * The higher a thread's priority, the bigger timeslices
179 * it gets during one round of execution. But even the lowest
180 * priority thread gets MIN_TIMESLICE worth of execution time.
183 static inline unsigned int task_timeslice(struct task_struct
*p
)
185 return static_prio_timeslice(p
->static_prio
);
189 * These are the runqueue data structures:
193 unsigned int nr_active
;
194 DECLARE_BITMAP(bitmap
, MAX_PRIO
+1); /* include 1 bit for delimiter */
195 struct list_head queue
[MAX_PRIO
];
199 * This is the main, per-CPU runqueue data structure.
201 * Locking rule: those places that want to lock multiple runqueues
202 * (such as the load balancing or the thread migration code), lock
203 * acquire operations must be ordered by ascending &runqueue.
209 * nr_running and cpu_load should be in the same cacheline because
210 * remote CPUs use both these fields when doing load calculation.
212 unsigned long nr_running
;
213 unsigned long raw_weighted_load
;
215 unsigned long cpu_load
[3];
217 unsigned long long nr_switches
;
220 * This is part of a global counter where only the total sum
221 * over all CPUs matters. A task can increase this counter on
222 * one CPU and if it got migrated afterwards it may decrease
223 * it on another CPU. Always updated under the runqueue lock:
225 unsigned long nr_uninterruptible
;
227 unsigned long expired_timestamp
;
228 /* Cached timestamp set by update_cpu_clock() */
229 unsigned long long most_recent_timestamp
;
230 struct task_struct
*curr
, *idle
;
231 unsigned long next_balance
;
232 struct mm_struct
*prev_mm
;
233 struct prio_array
*active
, *expired
, arrays
[2];
234 int best_expired_prio
;
238 struct sched_domain
*sd
;
240 /* For active balancing */
243 int cpu
; /* cpu of this runqueue */
245 struct task_struct
*migration_thread
;
246 struct list_head migration_queue
;
249 #ifdef CONFIG_SCHEDSTATS
251 struct sched_info rq_sched_info
;
253 /* sys_sched_yield() stats */
254 unsigned long yld_exp_empty
;
255 unsigned long yld_act_empty
;
256 unsigned long yld_both_empty
;
257 unsigned long yld_cnt
;
259 /* schedule() stats */
260 unsigned long sched_switch
;
261 unsigned long sched_cnt
;
262 unsigned long sched_goidle
;
264 /* try_to_wake_up() stats */
265 unsigned long ttwu_cnt
;
266 unsigned long ttwu_local
;
268 struct lock_class_key rq_lock_key
;
271 static DEFINE_PER_CPU(struct rq
, runqueues
);
273 static inline int cpu_of(struct rq
*rq
)
283 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
284 * See detach_destroy_domains: synchronize_sched for details.
286 * The domain tree of any CPU may only be accessed from within
287 * preempt-disabled sections.
289 #define for_each_domain(cpu, __sd) \
290 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
292 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
293 #define this_rq() (&__get_cpu_var(runqueues))
294 #define task_rq(p) cpu_rq(task_cpu(p))
295 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
297 #ifndef prepare_arch_switch
298 # define prepare_arch_switch(next) do { } while (0)
300 #ifndef finish_arch_switch
301 # define finish_arch_switch(prev) do { } while (0)
304 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
305 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
307 return rq
->curr
== p
;
310 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
314 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
316 #ifdef CONFIG_DEBUG_SPINLOCK
317 /* this is a valid case when another task releases the spinlock */
318 rq
->lock
.owner
= current
;
321 * If we are tracking spinlock dependencies then we have to
322 * fix up the runqueue lock - which gets 'carried over' from
325 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
327 spin_unlock_irq(&rq
->lock
);
330 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
331 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
336 return rq
->curr
== p
;
340 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
344 * We can optimise this out completely for !SMP, because the
345 * SMP rebalancing from interrupt is the only thing that cares
350 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
351 spin_unlock_irq(&rq
->lock
);
353 spin_unlock(&rq
->lock
);
357 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
361 * After ->oncpu is cleared, the task can be moved to a different CPU.
362 * We must ensure this doesn't happen until the switch is completely
368 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
372 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
375 * __task_rq_lock - lock the runqueue a given task resides on.
376 * Must be called interrupts disabled.
378 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
385 spin_lock(&rq
->lock
);
386 if (unlikely(rq
!= task_rq(p
))) {
387 spin_unlock(&rq
->lock
);
388 goto repeat_lock_task
;
394 * task_rq_lock - lock the runqueue a given task resides on and disable
395 * interrupts. Note the ordering: we can safely lookup the task_rq without
396 * explicitly disabling preemption.
398 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
404 local_irq_save(*flags
);
406 spin_lock(&rq
->lock
);
407 if (unlikely(rq
!= task_rq(p
))) {
408 spin_unlock_irqrestore(&rq
->lock
, *flags
);
409 goto repeat_lock_task
;
414 static inline void __task_rq_unlock(struct rq
*rq
)
417 spin_unlock(&rq
->lock
);
420 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
423 spin_unlock_irqrestore(&rq
->lock
, *flags
);
426 #ifdef CONFIG_SCHEDSTATS
428 * bump this up when changing the output format or the meaning of an existing
429 * format, so that tools can adapt (or abort)
431 #define SCHEDSTAT_VERSION 14
433 static int show_schedstat(struct seq_file
*seq
, void *v
)
437 seq_printf(seq
, "version %d\n", SCHEDSTAT_VERSION
);
438 seq_printf(seq
, "timestamp %lu\n", jiffies
);
439 for_each_online_cpu(cpu
) {
440 struct rq
*rq
= cpu_rq(cpu
);
442 struct sched_domain
*sd
;
446 /* runqueue-specific stats */
448 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
449 cpu
, rq
->yld_both_empty
,
450 rq
->yld_act_empty
, rq
->yld_exp_empty
, rq
->yld_cnt
,
451 rq
->sched_switch
, rq
->sched_cnt
, rq
->sched_goidle
,
452 rq
->ttwu_cnt
, rq
->ttwu_local
,
453 rq
->rq_sched_info
.cpu_time
,
454 rq
->rq_sched_info
.run_delay
, rq
->rq_sched_info
.pcnt
);
456 seq_printf(seq
, "\n");
459 /* domain-specific stats */
461 for_each_domain(cpu
, sd
) {
462 enum idle_type itype
;
463 char mask_str
[NR_CPUS
];
465 cpumask_scnprintf(mask_str
, NR_CPUS
, sd
->span
);
466 seq_printf(seq
, "domain%d %s", dcnt
++, mask_str
);
467 for (itype
= SCHED_IDLE
; itype
< MAX_IDLE_TYPES
;
469 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu "
472 sd
->lb_balanced
[itype
],
473 sd
->lb_failed
[itype
],
474 sd
->lb_imbalance
[itype
],
475 sd
->lb_gained
[itype
],
476 sd
->lb_hot_gained
[itype
],
477 sd
->lb_nobusyq
[itype
],
478 sd
->lb_nobusyg
[itype
]);
480 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu %lu %lu"
482 sd
->alb_cnt
, sd
->alb_failed
, sd
->alb_pushed
,
483 sd
->sbe_cnt
, sd
->sbe_balanced
, sd
->sbe_pushed
,
484 sd
->sbf_cnt
, sd
->sbf_balanced
, sd
->sbf_pushed
,
485 sd
->ttwu_wake_remote
, sd
->ttwu_move_affine
,
486 sd
->ttwu_move_balance
);
494 static int schedstat_open(struct inode
*inode
, struct file
*file
)
496 unsigned int size
= PAGE_SIZE
* (1 + num_online_cpus() / 32);
497 char *buf
= kmalloc(size
, GFP_KERNEL
);
503 res
= single_open(file
, show_schedstat
, NULL
);
505 m
= file
->private_data
;
513 const struct file_operations proc_schedstat_operations
= {
514 .open
= schedstat_open
,
517 .release
= single_release
,
521 * Expects runqueue lock to be held for atomicity of update
524 rq_sched_info_arrive(struct rq
*rq
, unsigned long delta_jiffies
)
527 rq
->rq_sched_info
.run_delay
+= delta_jiffies
;
528 rq
->rq_sched_info
.pcnt
++;
533 * Expects runqueue lock to be held for atomicity of update
536 rq_sched_info_depart(struct rq
*rq
, unsigned long delta_jiffies
)
539 rq
->rq_sched_info
.cpu_time
+= delta_jiffies
;
541 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
542 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
543 #else /* !CONFIG_SCHEDSTATS */
545 rq_sched_info_arrive(struct rq
*rq
, unsigned long delta_jiffies
)
548 rq_sched_info_depart(struct rq
*rq
, unsigned long delta_jiffies
)
550 # define schedstat_inc(rq, field) do { } while (0)
551 # define schedstat_add(rq, field, amt) do { } while (0)
555 * this_rq_lock - lock this runqueue and disable interrupts.
557 static inline struct rq
*this_rq_lock(void)
564 spin_lock(&rq
->lock
);
569 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
571 * Called when a process is dequeued from the active array and given
572 * the cpu. We should note that with the exception of interactive
573 * tasks, the expired queue will become the active queue after the active
574 * queue is empty, without explicitly dequeuing and requeuing tasks in the
575 * expired queue. (Interactive tasks may be requeued directly to the
576 * active queue, thus delaying tasks in the expired queue from running;
577 * see scheduler_tick()).
579 * This function is only called from sched_info_arrive(), rather than
580 * dequeue_task(). Even though a task may be queued and dequeued multiple
581 * times as it is shuffled about, we're really interested in knowing how
582 * long it was from the *first* time it was queued to the time that it
585 static inline void sched_info_dequeued(struct task_struct
*t
)
587 t
->sched_info
.last_queued
= 0;
591 * Called when a task finally hits the cpu. We can now calculate how
592 * long it was waiting to run. We also note when it began so that we
593 * can keep stats on how long its timeslice is.
595 static void sched_info_arrive(struct task_struct
*t
)
597 unsigned long now
= jiffies
, delta_jiffies
= 0;
599 if (t
->sched_info
.last_queued
)
600 delta_jiffies
= now
- t
->sched_info
.last_queued
;
601 sched_info_dequeued(t
);
602 t
->sched_info
.run_delay
+= delta_jiffies
;
603 t
->sched_info
.last_arrival
= now
;
604 t
->sched_info
.pcnt
++;
606 rq_sched_info_arrive(task_rq(t
), delta_jiffies
);
610 * Called when a process is queued into either the active or expired
611 * array. The time is noted and later used to determine how long we
612 * had to wait for us to reach the cpu. Since the expired queue will
613 * become the active queue after active queue is empty, without dequeuing
614 * and requeuing any tasks, we are interested in queuing to either. It
615 * is unusual but not impossible for tasks to be dequeued and immediately
616 * requeued in the same or another array: this can happen in sched_yield(),
617 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
620 * This function is only called from enqueue_task(), but also only updates
621 * the timestamp if it is already not set. It's assumed that
622 * sched_info_dequeued() will clear that stamp when appropriate.
624 static inline void sched_info_queued(struct task_struct
*t
)
626 if (unlikely(sched_info_on()))
627 if (!t
->sched_info
.last_queued
)
628 t
->sched_info
.last_queued
= jiffies
;
632 * Called when a process ceases being the active-running process, either
633 * voluntarily or involuntarily. Now we can calculate how long we ran.
635 static inline void sched_info_depart(struct task_struct
*t
)
637 unsigned long delta_jiffies
= jiffies
- t
->sched_info
.last_arrival
;
639 t
->sched_info
.cpu_time
+= delta_jiffies
;
640 rq_sched_info_depart(task_rq(t
), delta_jiffies
);
644 * Called when tasks are switched involuntarily due, typically, to expiring
645 * their time slice. (This may also be called when switching to or from
646 * the idle task.) We are only called when prev != next.
649 __sched_info_switch(struct task_struct
*prev
, struct task_struct
*next
)
651 struct rq
*rq
= task_rq(prev
);
654 * prev now departs the cpu. It's not interesting to record
655 * stats about how efficient we were at scheduling the idle
658 if (prev
!= rq
->idle
)
659 sched_info_depart(prev
);
661 if (next
!= rq
->idle
)
662 sched_info_arrive(next
);
665 sched_info_switch(struct task_struct
*prev
, struct task_struct
*next
)
667 if (unlikely(sched_info_on()))
668 __sched_info_switch(prev
, next
);
671 #define sched_info_queued(t) do { } while (0)
672 #define sched_info_switch(t, next) do { } while (0)
673 #endif /* CONFIG_SCHEDSTATS || CONFIG_TASK_DELAY_ACCT */
676 * Adding/removing a task to/from a priority array:
678 static void dequeue_task(struct task_struct
*p
, struct prio_array
*array
)
681 list_del(&p
->run_list
);
682 if (list_empty(array
->queue
+ p
->prio
))
683 __clear_bit(p
->prio
, array
->bitmap
);
686 static void enqueue_task(struct task_struct
*p
, struct prio_array
*array
)
688 sched_info_queued(p
);
689 list_add_tail(&p
->run_list
, array
->queue
+ p
->prio
);
690 __set_bit(p
->prio
, array
->bitmap
);
696 * Put task to the end of the run list without the overhead of dequeue
697 * followed by enqueue.
699 static void requeue_task(struct task_struct
*p
, struct prio_array
*array
)
701 list_move_tail(&p
->run_list
, array
->queue
+ p
->prio
);
705 enqueue_task_head(struct task_struct
*p
, struct prio_array
*array
)
707 list_add(&p
->run_list
, array
->queue
+ p
->prio
);
708 __set_bit(p
->prio
, array
->bitmap
);
714 * __normal_prio - return the priority that is based on the static
715 * priority but is modified by bonuses/penalties.
717 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
718 * into the -5 ... 0 ... +5 bonus/penalty range.
720 * We use 25% of the full 0...39 priority range so that:
722 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
723 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
725 * Both properties are important to certain workloads.
728 static inline int __normal_prio(struct task_struct
*p
)
732 bonus
= CURRENT_BONUS(p
) - MAX_BONUS
/ 2;
734 prio
= p
->static_prio
- bonus
;
735 if (prio
< MAX_RT_PRIO
)
737 if (prio
> MAX_PRIO
-1)
743 * To aid in avoiding the subversion of "niceness" due to uneven distribution
744 * of tasks with abnormal "nice" values across CPUs the contribution that
745 * each task makes to its run queue's load is weighted according to its
746 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
747 * scaled version of the new time slice allocation that they receive on time
752 * Assume: static_prio_timeslice(NICE_TO_PRIO(0)) == DEF_TIMESLICE
753 * If static_prio_timeslice() is ever changed to break this assumption then
754 * this code will need modification
756 #define TIME_SLICE_NICE_ZERO DEF_TIMESLICE
757 #define LOAD_WEIGHT(lp) \
758 (((lp) * SCHED_LOAD_SCALE) / TIME_SLICE_NICE_ZERO)
759 #define PRIO_TO_LOAD_WEIGHT(prio) \
760 LOAD_WEIGHT(static_prio_timeslice(prio))
761 #define RTPRIO_TO_LOAD_WEIGHT(rp) \
762 (PRIO_TO_LOAD_WEIGHT(MAX_RT_PRIO) + LOAD_WEIGHT(rp))
764 static void set_load_weight(struct task_struct
*p
)
766 if (has_rt_policy(p
)) {
768 if (p
== task_rq(p
)->migration_thread
)
770 * The migration thread does the actual balancing.
771 * Giving its load any weight will skew balancing
777 p
->load_weight
= RTPRIO_TO_LOAD_WEIGHT(p
->rt_priority
);
779 p
->load_weight
= PRIO_TO_LOAD_WEIGHT(p
->static_prio
);
783 inc_raw_weighted_load(struct rq
*rq
, const struct task_struct
*p
)
785 rq
->raw_weighted_load
+= p
->load_weight
;
789 dec_raw_weighted_load(struct rq
*rq
, const struct task_struct
*p
)
791 rq
->raw_weighted_load
-= p
->load_weight
;
794 static inline void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
797 inc_raw_weighted_load(rq
, p
);
800 static inline void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
803 dec_raw_weighted_load(rq
, p
);
807 * Calculate the expected normal priority: i.e. priority
808 * without taking RT-inheritance into account. Might be
809 * boosted by interactivity modifiers. Changes upon fork,
810 * setprio syscalls, and whenever the interactivity
811 * estimator recalculates.
813 static inline int normal_prio(struct task_struct
*p
)
817 if (has_rt_policy(p
))
818 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
820 prio
= __normal_prio(p
);
825 * Calculate the current priority, i.e. the priority
826 * taken into account by the scheduler. This value might
827 * be boosted by RT tasks, or might be boosted by
828 * interactivity modifiers. Will be RT if the task got
829 * RT-boosted. If not then it returns p->normal_prio.
831 static int effective_prio(struct task_struct
*p
)
833 p
->normal_prio
= normal_prio(p
);
835 * If we are RT tasks or we were boosted to RT priority,
836 * keep the priority unchanged. Otherwise, update priority
837 * to the normal priority:
839 if (!rt_prio(p
->prio
))
840 return p
->normal_prio
;
845 * __activate_task - move a task to the runqueue.
847 static void __activate_task(struct task_struct
*p
, struct rq
*rq
)
849 struct prio_array
*target
= rq
->active
;
852 target
= rq
->expired
;
853 enqueue_task(p
, target
);
854 inc_nr_running(p
, rq
);
858 * __activate_idle_task - move idle task to the _front_ of runqueue.
860 static inline void __activate_idle_task(struct task_struct
*p
, struct rq
*rq
)
862 enqueue_task_head(p
, rq
->active
);
863 inc_nr_running(p
, rq
);
867 * Recalculate p->normal_prio and p->prio after having slept,
868 * updating the sleep-average too:
870 static int recalc_task_prio(struct task_struct
*p
, unsigned long long now
)
872 /* Caller must always ensure 'now >= p->timestamp' */
873 unsigned long sleep_time
= now
- p
->timestamp
;
878 if (likely(sleep_time
> 0)) {
880 * This ceiling is set to the lowest priority that would allow
881 * a task to be reinserted into the active array on timeslice
884 unsigned long ceiling
= INTERACTIVE_SLEEP(p
);
886 if (p
->mm
&& sleep_time
> ceiling
&& p
->sleep_avg
< ceiling
) {
888 * Prevents user tasks from achieving best priority
889 * with one single large enough sleep.
891 p
->sleep_avg
= ceiling
;
893 * Using INTERACTIVE_SLEEP() as a ceiling places a
894 * nice(0) task 1ms sleep away from promotion, and
895 * gives it 700ms to round-robin with no chance of
896 * being demoted. This is more than generous, so
897 * mark this sleep as non-interactive to prevent the
898 * on-runqueue bonus logic from intervening should
899 * this task not receive cpu immediately.
901 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
904 * Tasks waking from uninterruptible sleep are
905 * limited in their sleep_avg rise as they
906 * are likely to be waiting on I/O
908 if (p
->sleep_type
== SLEEP_NONINTERACTIVE
&& p
->mm
) {
909 if (p
->sleep_avg
>= ceiling
)
911 else if (p
->sleep_avg
+ sleep_time
>=
913 p
->sleep_avg
= ceiling
;
919 * This code gives a bonus to interactive tasks.
921 * The boost works by updating the 'average sleep time'
922 * value here, based on ->timestamp. The more time a
923 * task spends sleeping, the higher the average gets -
924 * and the higher the priority boost gets as well.
926 p
->sleep_avg
+= sleep_time
;
929 if (p
->sleep_avg
> NS_MAX_SLEEP_AVG
)
930 p
->sleep_avg
= NS_MAX_SLEEP_AVG
;
933 return effective_prio(p
);
937 * activate_task - move a task to the runqueue and do priority recalculation
939 * Update all the scheduling statistics stuff. (sleep average
940 * calculation, priority modifiers, etc.)
942 static void activate_task(struct task_struct
*p
, struct rq
*rq
, int local
)
944 unsigned long long now
;
952 /* Compensate for drifting sched_clock */
953 struct rq
*this_rq
= this_rq();
954 now
= (now
- this_rq
->most_recent_timestamp
)
955 + rq
->most_recent_timestamp
;
960 * Sleep time is in units of nanosecs, so shift by 20 to get a
961 * milliseconds-range estimation of the amount of time that the task
964 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
965 if (p
->state
== TASK_UNINTERRUPTIBLE
)
966 profile_hits(SLEEP_PROFILING
, (void *)get_wchan(p
),
967 (now
- p
->timestamp
) >> 20);
970 p
->prio
= recalc_task_prio(p
, now
);
973 * This checks to make sure it's not an uninterruptible task
974 * that is now waking up.
976 if (p
->sleep_type
== SLEEP_NORMAL
) {
978 * Tasks which were woken up by interrupts (ie. hw events)
979 * are most likely of interactive nature. So we give them
980 * the credit of extending their sleep time to the period
981 * of time they spend on the runqueue, waiting for execution
982 * on a CPU, first time around:
985 p
->sleep_type
= SLEEP_INTERRUPTED
;
988 * Normal first-time wakeups get a credit too for
989 * on-runqueue time, but it will be weighted down:
991 p
->sleep_type
= SLEEP_INTERACTIVE
;
996 __activate_task(p
, rq
);
1000 * deactivate_task - remove a task from the runqueue.
1002 static void deactivate_task(struct task_struct
*p
, struct rq
*rq
)
1004 dec_nr_running(p
, rq
);
1005 dequeue_task(p
, p
->array
);
1010 * resched_task - mark a task 'to be rescheduled now'.
1012 * On UP this means the setting of the need_resched flag, on SMP it
1013 * might also involve a cross-CPU call to trigger the scheduler on
1018 #ifndef tsk_is_polling
1019 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1022 static void resched_task(struct task_struct
*p
)
1026 assert_spin_locked(&task_rq(p
)->lock
);
1028 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
1031 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
1034 if (cpu
== smp_processor_id())
1037 /* NEED_RESCHED must be visible before we test polling */
1039 if (!tsk_is_polling(p
))
1040 smp_send_reschedule(cpu
);
1043 static inline void resched_task(struct task_struct
*p
)
1045 assert_spin_locked(&task_rq(p
)->lock
);
1046 set_tsk_need_resched(p
);
1051 * task_curr - is this task currently executing on a CPU?
1052 * @p: the task in question.
1054 inline int task_curr(const struct task_struct
*p
)
1056 return cpu_curr(task_cpu(p
)) == p
;
1059 /* Used instead of source_load when we know the type == 0 */
1060 unsigned long weighted_cpuload(const int cpu
)
1062 return cpu_rq(cpu
)->raw_weighted_load
;
1066 struct migration_req
{
1067 struct list_head list
;
1069 struct task_struct
*task
;
1072 struct completion done
;
1076 * The task's runqueue lock must be held.
1077 * Returns true if you have to wait for migration thread.
1080 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1082 struct rq
*rq
= task_rq(p
);
1085 * If the task is not on a runqueue (and not running), then
1086 * it is sufficient to simply update the task's cpu field.
1088 if (!p
->array
&& !task_running(rq
, p
)) {
1089 set_task_cpu(p
, dest_cpu
);
1093 init_completion(&req
->done
);
1095 req
->dest_cpu
= dest_cpu
;
1096 list_add(&req
->list
, &rq
->migration_queue
);
1102 * wait_task_inactive - wait for a thread to unschedule.
1104 * The caller must ensure that the task *will* unschedule sometime soon,
1105 * else this function might spin for a *long* time. This function can't
1106 * be called with interrupts off, or it may introduce deadlock with
1107 * smp_call_function() if an IPI is sent by the same process we are
1108 * waiting to become inactive.
1110 void wait_task_inactive(struct task_struct
*p
)
1112 unsigned long flags
;
1117 rq
= task_rq_lock(p
, &flags
);
1118 /* Must be off runqueue entirely, not preempted. */
1119 if (unlikely(p
->array
|| task_running(rq
, p
))) {
1120 /* If it's preempted, we yield. It could be a while. */
1121 preempted
= !task_running(rq
, p
);
1122 task_rq_unlock(rq
, &flags
);
1128 task_rq_unlock(rq
, &flags
);
1132 * kick_process - kick a running thread to enter/exit the kernel
1133 * @p: the to-be-kicked thread
1135 * Cause a process which is running on another CPU to enter
1136 * kernel-mode, without any delay. (to get signals handled.)
1138 * NOTE: this function doesnt have to take the runqueue lock,
1139 * because all it wants to ensure is that the remote task enters
1140 * the kernel. If the IPI races and the task has been migrated
1141 * to another CPU then no harm is done and the purpose has been
1144 void kick_process(struct task_struct
*p
)
1150 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1151 smp_send_reschedule(cpu
);
1156 * Return a low guess at the load of a migration-source cpu weighted
1157 * according to the scheduling class and "nice" value.
1159 * We want to under-estimate the load of migration sources, to
1160 * balance conservatively.
1162 static inline unsigned long source_load(int cpu
, int type
)
1164 struct rq
*rq
= cpu_rq(cpu
);
1167 return rq
->raw_weighted_load
;
1169 return min(rq
->cpu_load
[type
-1], rq
->raw_weighted_load
);
1173 * Return a high guess at the load of a migration-target cpu weighted
1174 * according to the scheduling class and "nice" value.
1176 static inline unsigned long target_load(int cpu
, int type
)
1178 struct rq
*rq
= cpu_rq(cpu
);
1181 return rq
->raw_weighted_load
;
1183 return max(rq
->cpu_load
[type
-1], rq
->raw_weighted_load
);
1187 * Return the average load per task on the cpu's run queue
1189 static inline unsigned long cpu_avg_load_per_task(int cpu
)
1191 struct rq
*rq
= cpu_rq(cpu
);
1192 unsigned long n
= rq
->nr_running
;
1194 return n
? rq
->raw_weighted_load
/ n
: SCHED_LOAD_SCALE
;
1198 * find_idlest_group finds and returns the least busy CPU group within the
1201 static struct sched_group
*
1202 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1204 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1205 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1206 int load_idx
= sd
->forkexec_idx
;
1207 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1210 unsigned long load
, avg_load
;
1214 /* Skip over this group if it has no CPUs allowed */
1215 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1218 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1220 /* Tally up the load of all CPUs in the group */
1223 for_each_cpu_mask(i
, group
->cpumask
) {
1224 /* Bias balancing toward cpus of our domain */
1226 load
= source_load(i
, load_idx
);
1228 load
= target_load(i
, load_idx
);
1233 /* Adjust by relative CPU power of the group */
1234 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
1237 this_load
= avg_load
;
1239 } else if (avg_load
< min_load
) {
1240 min_load
= avg_load
;
1244 group
= group
->next
;
1245 } while (group
!= sd
->groups
);
1247 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1253 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1256 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1259 unsigned long load
, min_load
= ULONG_MAX
;
1263 /* Traverse only the allowed CPUs */
1264 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1266 for_each_cpu_mask(i
, tmp
) {
1267 load
= weighted_cpuload(i
);
1269 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1279 * sched_balance_self: balance the current task (running on cpu) in domains
1280 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1283 * Balance, ie. select the least loaded group.
1285 * Returns the target CPU number, or the same CPU if no balancing is needed.
1287 * preempt must be disabled.
1289 static int sched_balance_self(int cpu
, int flag
)
1291 struct task_struct
*t
= current
;
1292 struct sched_domain
*tmp
, *sd
= NULL
;
1294 for_each_domain(cpu
, tmp
) {
1296 * If power savings logic is enabled for a domain, stop there.
1298 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1300 if (tmp
->flags
& flag
)
1306 struct sched_group
*group
;
1307 int new_cpu
, weight
;
1309 if (!(sd
->flags
& flag
)) {
1315 group
= find_idlest_group(sd
, t
, cpu
);
1321 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1322 if (new_cpu
== -1 || new_cpu
== cpu
) {
1323 /* Now try balancing at a lower domain level of cpu */
1328 /* Now try balancing at a lower domain level of new_cpu */
1331 weight
= cpus_weight(span
);
1332 for_each_domain(cpu
, tmp
) {
1333 if (weight
<= cpus_weight(tmp
->span
))
1335 if (tmp
->flags
& flag
)
1338 /* while loop will break here if sd == NULL */
1344 #endif /* CONFIG_SMP */
1347 * wake_idle() will wake a task on an idle cpu if task->cpu is
1348 * not idle and an idle cpu is available. The span of cpus to
1349 * search starts with cpus closest then further out as needed,
1350 * so we always favor a closer, idle cpu.
1352 * Returns the CPU we should wake onto.
1354 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1355 static int wake_idle(int cpu
, struct task_struct
*p
)
1358 struct sched_domain
*sd
;
1364 for_each_domain(cpu
, sd
) {
1365 if (sd
->flags
& SD_WAKE_IDLE
) {
1366 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1367 for_each_cpu_mask(i
, tmp
) {
1378 static inline int wake_idle(int cpu
, struct task_struct
*p
)
1385 * try_to_wake_up - wake up a thread
1386 * @p: the to-be-woken-up thread
1387 * @state: the mask of task states that can be woken
1388 * @sync: do a synchronous wakeup?
1390 * Put it on the run-queue if it's not already there. The "current"
1391 * thread is always on the run-queue (except when the actual
1392 * re-schedule is in progress), and as such you're allowed to do
1393 * the simpler "current->state = TASK_RUNNING" to mark yourself
1394 * runnable without the overhead of this.
1396 * returns failure only if the task is already active.
1398 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1400 int cpu
, this_cpu
, success
= 0;
1401 unsigned long flags
;
1405 struct sched_domain
*sd
, *this_sd
= NULL
;
1406 unsigned long load
, this_load
;
1410 rq
= task_rq_lock(p
, &flags
);
1411 old_state
= p
->state
;
1412 if (!(old_state
& state
))
1419 this_cpu
= smp_processor_id();
1422 if (unlikely(task_running(rq
, p
)))
1427 schedstat_inc(rq
, ttwu_cnt
);
1428 if (cpu
== this_cpu
) {
1429 schedstat_inc(rq
, ttwu_local
);
1433 for_each_domain(this_cpu
, sd
) {
1434 if (cpu_isset(cpu
, sd
->span
)) {
1435 schedstat_inc(sd
, ttwu_wake_remote
);
1441 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1445 * Check for affine wakeup and passive balancing possibilities.
1448 int idx
= this_sd
->wake_idx
;
1449 unsigned int imbalance
;
1451 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1453 load
= source_load(cpu
, idx
);
1454 this_load
= target_load(this_cpu
, idx
);
1456 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1458 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1459 unsigned long tl
= this_load
;
1460 unsigned long tl_per_task
;
1462 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1465 * If sync wakeup then subtract the (maximum possible)
1466 * effect of the currently running task from the load
1467 * of the current CPU:
1470 tl
-= current
->load_weight
;
1473 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1474 100*(tl
+ p
->load_weight
) <= imbalance
*load
) {
1476 * This domain has SD_WAKE_AFFINE and
1477 * p is cache cold in this domain, and
1478 * there is no bad imbalance.
1480 schedstat_inc(this_sd
, ttwu_move_affine
);
1486 * Start passive balancing when half the imbalance_pct
1489 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1490 if (imbalance
*this_load
<= 100*load
) {
1491 schedstat_inc(this_sd
, ttwu_move_balance
);
1497 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1499 new_cpu
= wake_idle(new_cpu
, p
);
1500 if (new_cpu
!= cpu
) {
1501 set_task_cpu(p
, new_cpu
);
1502 task_rq_unlock(rq
, &flags
);
1503 /* might preempt at this point */
1504 rq
= task_rq_lock(p
, &flags
);
1505 old_state
= p
->state
;
1506 if (!(old_state
& state
))
1511 this_cpu
= smp_processor_id();
1516 #endif /* CONFIG_SMP */
1517 if (old_state
== TASK_UNINTERRUPTIBLE
) {
1518 rq
->nr_uninterruptible
--;
1520 * Tasks on involuntary sleep don't earn
1521 * sleep_avg beyond just interactive state.
1523 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
1527 * Tasks that have marked their sleep as noninteractive get
1528 * woken up with their sleep average not weighted in an
1531 if (old_state
& TASK_NONINTERACTIVE
)
1532 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
1535 activate_task(p
, rq
, cpu
== this_cpu
);
1537 * Sync wakeups (i.e. those types of wakeups where the waker
1538 * has indicated that it will leave the CPU in short order)
1539 * don't trigger a preemption, if the woken up task will run on
1540 * this cpu. (in this case the 'I will reschedule' promise of
1541 * the waker guarantees that the freshly woken up task is going
1542 * to be considered on this CPU.)
1544 if (!sync
|| cpu
!= this_cpu
) {
1545 if (TASK_PREEMPTS_CURR(p
, rq
))
1546 resched_task(rq
->curr
);
1551 p
->state
= TASK_RUNNING
;
1553 task_rq_unlock(rq
, &flags
);
1558 int fastcall
wake_up_process(struct task_struct
*p
)
1560 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1561 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1563 EXPORT_SYMBOL(wake_up_process
);
1565 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1567 return try_to_wake_up(p
, state
, 0);
1570 static void task_running_tick(struct rq
*rq
, struct task_struct
*p
);
1572 * Perform scheduler related setup for a newly forked process p.
1573 * p is forked by current.
1575 void fastcall
sched_fork(struct task_struct
*p
, int clone_flags
)
1577 int cpu
= get_cpu();
1580 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1582 set_task_cpu(p
, cpu
);
1585 * We mark the process as running here, but have not actually
1586 * inserted it onto the runqueue yet. This guarantees that
1587 * nobody will actually run it, and a signal or other external
1588 * event cannot wake it up and insert it on the runqueue either.
1590 p
->state
= TASK_RUNNING
;
1593 * Make sure we do not leak PI boosting priority to the child:
1595 p
->prio
= current
->normal_prio
;
1597 INIT_LIST_HEAD(&p
->run_list
);
1599 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1600 if (unlikely(sched_info_on()))
1601 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1603 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1606 #ifdef CONFIG_PREEMPT
1607 /* Want to start with kernel preemption disabled. */
1608 task_thread_info(p
)->preempt_count
= 1;
1611 * Share the timeslice between parent and child, thus the
1612 * total amount of pending timeslices in the system doesn't change,
1613 * resulting in more scheduling fairness.
1615 local_irq_disable();
1616 p
->time_slice
= (current
->time_slice
+ 1) >> 1;
1618 * The remainder of the first timeslice might be recovered by
1619 * the parent if the child exits early enough.
1621 p
->first_time_slice
= 1;
1622 current
->time_slice
>>= 1;
1623 p
->timestamp
= sched_clock();
1624 if (unlikely(!current
->time_slice
)) {
1626 * This case is rare, it happens when the parent has only
1627 * a single jiffy left from its timeslice. Taking the
1628 * runqueue lock is not a problem.
1630 current
->time_slice
= 1;
1631 task_running_tick(cpu_rq(cpu
), current
);
1638 * wake_up_new_task - wake up a newly created task for the first time.
1640 * This function will do some initial scheduler statistics housekeeping
1641 * that must be done for every newly created context, then puts the task
1642 * on the runqueue and wakes it.
1644 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1646 struct rq
*rq
, *this_rq
;
1647 unsigned long flags
;
1650 rq
= task_rq_lock(p
, &flags
);
1651 BUG_ON(p
->state
!= TASK_RUNNING
);
1652 this_cpu
= smp_processor_id();
1656 * We decrease the sleep average of forking parents
1657 * and children as well, to keep max-interactive tasks
1658 * from forking tasks that are max-interactive. The parent
1659 * (current) is done further down, under its lock.
1661 p
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(p
) *
1662 CHILD_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1664 p
->prio
= effective_prio(p
);
1666 if (likely(cpu
== this_cpu
)) {
1667 if (!(clone_flags
& CLONE_VM
)) {
1669 * The VM isn't cloned, so we're in a good position to
1670 * do child-runs-first in anticipation of an exec. This
1671 * usually avoids a lot of COW overhead.
1673 if (unlikely(!current
->array
))
1674 __activate_task(p
, rq
);
1676 p
->prio
= current
->prio
;
1677 p
->normal_prio
= current
->normal_prio
;
1678 list_add_tail(&p
->run_list
, ¤t
->run_list
);
1679 p
->array
= current
->array
;
1680 p
->array
->nr_active
++;
1681 inc_nr_running(p
, rq
);
1685 /* Run child last */
1686 __activate_task(p
, rq
);
1688 * We skip the following code due to cpu == this_cpu
1690 * task_rq_unlock(rq, &flags);
1691 * this_rq = task_rq_lock(current, &flags);
1695 this_rq
= cpu_rq(this_cpu
);
1698 * Not the local CPU - must adjust timestamp. This should
1699 * get optimised away in the !CONFIG_SMP case.
1701 p
->timestamp
= (p
->timestamp
- this_rq
->most_recent_timestamp
)
1702 + rq
->most_recent_timestamp
;
1703 __activate_task(p
, rq
);
1704 if (TASK_PREEMPTS_CURR(p
, rq
))
1705 resched_task(rq
->curr
);
1708 * Parent and child are on different CPUs, now get the
1709 * parent runqueue to update the parent's ->sleep_avg:
1711 task_rq_unlock(rq
, &flags
);
1712 this_rq
= task_rq_lock(current
, &flags
);
1714 current
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(current
) *
1715 PARENT_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1716 task_rq_unlock(this_rq
, &flags
);
1720 * Potentially available exiting-child timeslices are
1721 * retrieved here - this way the parent does not get
1722 * penalized for creating too many threads.
1724 * (this cannot be used to 'generate' timeslices
1725 * artificially, because any timeslice recovered here
1726 * was given away by the parent in the first place.)
1728 void fastcall
sched_exit(struct task_struct
*p
)
1730 unsigned long flags
;
1734 * If the child was a (relative-) CPU hog then decrease
1735 * the sleep_avg of the parent as well.
1737 rq
= task_rq_lock(p
->parent
, &flags
);
1738 if (p
->first_time_slice
&& task_cpu(p
) == task_cpu(p
->parent
)) {
1739 p
->parent
->time_slice
+= p
->time_slice
;
1740 if (unlikely(p
->parent
->time_slice
> task_timeslice(p
)))
1741 p
->parent
->time_slice
= task_timeslice(p
);
1743 if (p
->sleep_avg
< p
->parent
->sleep_avg
)
1744 p
->parent
->sleep_avg
= p
->parent
->sleep_avg
/
1745 (EXIT_WEIGHT
+ 1) * EXIT_WEIGHT
+ p
->sleep_avg
/
1747 task_rq_unlock(rq
, &flags
);
1751 * prepare_task_switch - prepare to switch tasks
1752 * @rq: the runqueue preparing to switch
1753 * @next: the task we are going to switch to.
1755 * This is called with the rq lock held and interrupts off. It must
1756 * be paired with a subsequent finish_task_switch after the context
1759 * prepare_task_switch sets up locking and calls architecture specific
1762 static inline void prepare_task_switch(struct rq
*rq
, struct task_struct
*next
)
1764 prepare_lock_switch(rq
, next
);
1765 prepare_arch_switch(next
);
1769 * finish_task_switch - clean up after a task-switch
1770 * @rq: runqueue associated with task-switch
1771 * @prev: the thread we just switched away from.
1773 * finish_task_switch must be called after the context switch, paired
1774 * with a prepare_task_switch call before the context switch.
1775 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1776 * and do any other architecture-specific cleanup actions.
1778 * Note that we may have delayed dropping an mm in context_switch(). If
1779 * so, we finish that here outside of the runqueue lock. (Doing it
1780 * with the lock held can cause deadlocks; see schedule() for
1783 static inline void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1784 __releases(rq
->lock
)
1786 struct mm_struct
*mm
= rq
->prev_mm
;
1792 * A task struct has one reference for the use as "current".
1793 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1794 * schedule one last time. The schedule call will never return, and
1795 * the scheduled task must drop that reference.
1796 * The test for TASK_DEAD must occur while the runqueue locks are
1797 * still held, otherwise prev could be scheduled on another cpu, die
1798 * there before we look at prev->state, and then the reference would
1800 * Manfred Spraul <manfred@colorfullife.com>
1802 prev_state
= prev
->state
;
1803 finish_arch_switch(prev
);
1804 finish_lock_switch(rq
, prev
);
1807 if (unlikely(prev_state
== TASK_DEAD
)) {
1809 * Remove function-return probe instances associated with this
1810 * task and put them back on the free list.
1812 kprobe_flush_task(prev
);
1813 put_task_struct(prev
);
1818 * schedule_tail - first thing a freshly forked thread must call.
1819 * @prev: the thread we just switched away from.
1821 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1822 __releases(rq
->lock
)
1824 struct rq
*rq
= this_rq();
1826 finish_task_switch(rq
, prev
);
1827 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1828 /* In this case, finish_task_switch does not reenable preemption */
1831 if (current
->set_child_tid
)
1832 put_user(current
->pid
, current
->set_child_tid
);
1836 * context_switch - switch to the new MM and the new
1837 * thread's register state.
1839 static inline struct task_struct
*
1840 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1841 struct task_struct
*next
)
1843 struct mm_struct
*mm
= next
->mm
;
1844 struct mm_struct
*oldmm
= prev
->active_mm
;
1847 next
->active_mm
= oldmm
;
1848 atomic_inc(&oldmm
->mm_count
);
1849 enter_lazy_tlb(oldmm
, next
);
1851 switch_mm(oldmm
, mm
, next
);
1854 prev
->active_mm
= NULL
;
1855 WARN_ON(rq
->prev_mm
);
1856 rq
->prev_mm
= oldmm
;
1859 * Since the runqueue lock will be released by the next
1860 * task (which is an invalid locking op but in the case
1861 * of the scheduler it's an obvious special-case), so we
1862 * do an early lockdep release here:
1864 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1865 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1868 /* Here we just switch the register state and the stack. */
1869 switch_to(prev
, next
, prev
);
1875 * nr_running, nr_uninterruptible and nr_context_switches:
1877 * externally visible scheduler statistics: current number of runnable
1878 * threads, current number of uninterruptible-sleeping threads, total
1879 * number of context switches performed since bootup.
1881 unsigned long nr_running(void)
1883 unsigned long i
, sum
= 0;
1885 for_each_online_cpu(i
)
1886 sum
+= cpu_rq(i
)->nr_running
;
1891 unsigned long nr_uninterruptible(void)
1893 unsigned long i
, sum
= 0;
1895 for_each_possible_cpu(i
)
1896 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1899 * Since we read the counters lockless, it might be slightly
1900 * inaccurate. Do not allow it to go below zero though:
1902 if (unlikely((long)sum
< 0))
1908 unsigned long long nr_context_switches(void)
1911 unsigned long long sum
= 0;
1913 for_each_possible_cpu(i
)
1914 sum
+= cpu_rq(i
)->nr_switches
;
1919 unsigned long nr_iowait(void)
1921 unsigned long i
, sum
= 0;
1923 for_each_possible_cpu(i
)
1924 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1929 unsigned long nr_active(void)
1931 unsigned long i
, running
= 0, uninterruptible
= 0;
1933 for_each_online_cpu(i
) {
1934 running
+= cpu_rq(i
)->nr_running
;
1935 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
1938 if (unlikely((long)uninterruptible
< 0))
1939 uninterruptible
= 0;
1941 return running
+ uninterruptible
;
1947 * Is this task likely cache-hot:
1950 task_hot(struct task_struct
*p
, unsigned long long now
, struct sched_domain
*sd
)
1952 return (long long)(now
- p
->last_ran
) < (long long)sd
->cache_hot_time
;
1956 * double_rq_lock - safely lock two runqueues
1958 * Note this does not disable interrupts like task_rq_lock,
1959 * you need to do so manually before calling.
1961 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1962 __acquires(rq1
->lock
)
1963 __acquires(rq2
->lock
)
1965 BUG_ON(!irqs_disabled());
1967 spin_lock(&rq1
->lock
);
1968 __acquire(rq2
->lock
); /* Fake it out ;) */
1971 spin_lock(&rq1
->lock
);
1972 spin_lock(&rq2
->lock
);
1974 spin_lock(&rq2
->lock
);
1975 spin_lock(&rq1
->lock
);
1981 * double_rq_unlock - safely unlock two runqueues
1983 * Note this does not restore interrupts like task_rq_unlock,
1984 * you need to do so manually after calling.
1986 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
1987 __releases(rq1
->lock
)
1988 __releases(rq2
->lock
)
1990 spin_unlock(&rq1
->lock
);
1992 spin_unlock(&rq2
->lock
);
1994 __release(rq2
->lock
);
1998 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2000 static void double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2001 __releases(this_rq
->lock
)
2002 __acquires(busiest
->lock
)
2003 __acquires(this_rq
->lock
)
2005 if (unlikely(!irqs_disabled())) {
2006 /* printk() doesn't work good under rq->lock */
2007 spin_unlock(&this_rq
->lock
);
2010 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2011 if (busiest
< this_rq
) {
2012 spin_unlock(&this_rq
->lock
);
2013 spin_lock(&busiest
->lock
);
2014 spin_lock(&this_rq
->lock
);
2016 spin_lock(&busiest
->lock
);
2021 * If dest_cpu is allowed for this process, migrate the task to it.
2022 * This is accomplished by forcing the cpu_allowed mask to only
2023 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2024 * the cpu_allowed mask is restored.
2026 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2028 struct migration_req req
;
2029 unsigned long flags
;
2032 rq
= task_rq_lock(p
, &flags
);
2033 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2034 || unlikely(cpu_is_offline(dest_cpu
)))
2037 /* force the process onto the specified CPU */
2038 if (migrate_task(p
, dest_cpu
, &req
)) {
2039 /* Need to wait for migration thread (might exit: take ref). */
2040 struct task_struct
*mt
= rq
->migration_thread
;
2042 get_task_struct(mt
);
2043 task_rq_unlock(rq
, &flags
);
2044 wake_up_process(mt
);
2045 put_task_struct(mt
);
2046 wait_for_completion(&req
.done
);
2051 task_rq_unlock(rq
, &flags
);
2055 * sched_exec - execve() is a valuable balancing opportunity, because at
2056 * this point the task has the smallest effective memory and cache footprint.
2058 void sched_exec(void)
2060 int new_cpu
, this_cpu
= get_cpu();
2061 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2063 if (new_cpu
!= this_cpu
)
2064 sched_migrate_task(current
, new_cpu
);
2068 * pull_task - move a task from a remote runqueue to the local runqueue.
2069 * Both runqueues must be locked.
2071 static void pull_task(struct rq
*src_rq
, struct prio_array
*src_array
,
2072 struct task_struct
*p
, struct rq
*this_rq
,
2073 struct prio_array
*this_array
, int this_cpu
)
2075 dequeue_task(p
, src_array
);
2076 dec_nr_running(p
, src_rq
);
2077 set_task_cpu(p
, this_cpu
);
2078 inc_nr_running(p
, this_rq
);
2079 enqueue_task(p
, this_array
);
2080 p
->timestamp
= (p
->timestamp
- src_rq
->most_recent_timestamp
)
2081 + this_rq
->most_recent_timestamp
;
2083 * Note that idle threads have a prio of MAX_PRIO, for this test
2084 * to be always true for them.
2086 if (TASK_PREEMPTS_CURR(p
, this_rq
))
2087 resched_task(this_rq
->curr
);
2091 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2094 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2095 struct sched_domain
*sd
, enum idle_type idle
,
2099 * We do not migrate tasks that are:
2100 * 1) running (obviously), or
2101 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2102 * 3) are cache-hot on their current CPU.
2104 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
2108 if (task_running(rq
, p
))
2112 * Aggressive migration if:
2113 * 1) task is cache cold, or
2114 * 2) too many balance attempts have failed.
2117 if (sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2118 #ifdef CONFIG_SCHEDSTATS
2119 if (task_hot(p
, rq
->most_recent_timestamp
, sd
))
2120 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2125 if (task_hot(p
, rq
->most_recent_timestamp
, sd
))
2130 #define rq_best_prio(rq) min((rq)->curr->prio, (rq)->best_expired_prio)
2133 * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted
2134 * load from busiest to this_rq, as part of a balancing operation within
2135 * "domain". Returns the number of tasks moved.
2137 * Called with both runqueues locked.
2139 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2140 unsigned long max_nr_move
, unsigned long max_load_move
,
2141 struct sched_domain
*sd
, enum idle_type idle
,
2144 int idx
, pulled
= 0, pinned
= 0, this_best_prio
, best_prio
,
2145 best_prio_seen
, skip_for_load
;
2146 struct prio_array
*array
, *dst_array
;
2147 struct list_head
*head
, *curr
;
2148 struct task_struct
*tmp
;
2151 if (max_nr_move
== 0 || max_load_move
== 0)
2154 rem_load_move
= max_load_move
;
2156 this_best_prio
= rq_best_prio(this_rq
);
2157 best_prio
= rq_best_prio(busiest
);
2159 * Enable handling of the case where there is more than one task
2160 * with the best priority. If the current running task is one
2161 * of those with prio==best_prio we know it won't be moved
2162 * and therefore it's safe to override the skip (based on load) of
2163 * any task we find with that prio.
2165 best_prio_seen
= best_prio
== busiest
->curr
->prio
;
2168 * We first consider expired tasks. Those will likely not be
2169 * executed in the near future, and they are most likely to
2170 * be cache-cold, thus switching CPUs has the least effect
2173 if (busiest
->expired
->nr_active
) {
2174 array
= busiest
->expired
;
2175 dst_array
= this_rq
->expired
;
2177 array
= busiest
->active
;
2178 dst_array
= this_rq
->active
;
2182 /* Start searching at priority 0: */
2186 idx
= sched_find_first_bit(array
->bitmap
);
2188 idx
= find_next_bit(array
->bitmap
, MAX_PRIO
, idx
);
2189 if (idx
>= MAX_PRIO
) {
2190 if (array
== busiest
->expired
&& busiest
->active
->nr_active
) {
2191 array
= busiest
->active
;
2192 dst_array
= this_rq
->active
;
2198 head
= array
->queue
+ idx
;
2201 tmp
= list_entry(curr
, struct task_struct
, run_list
);
2206 * To help distribute high priority tasks accross CPUs we don't
2207 * skip a task if it will be the highest priority task (i.e. smallest
2208 * prio value) on its new queue regardless of its load weight
2210 skip_for_load
= tmp
->load_weight
> rem_load_move
;
2211 if (skip_for_load
&& idx
< this_best_prio
)
2212 skip_for_load
= !best_prio_seen
&& idx
== best_prio
;
2213 if (skip_for_load
||
2214 !can_migrate_task(tmp
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2216 best_prio_seen
|= idx
== best_prio
;
2223 pull_task(busiest
, array
, tmp
, this_rq
, dst_array
, this_cpu
);
2225 rem_load_move
-= tmp
->load_weight
;
2228 * We only want to steal up to the prescribed number of tasks
2229 * and the prescribed amount of weighted load.
2231 if (pulled
< max_nr_move
&& rem_load_move
> 0) {
2232 if (idx
< this_best_prio
)
2233 this_best_prio
= idx
;
2241 * Right now, this is the only place pull_task() is called,
2242 * so we can safely collect pull_task() stats here rather than
2243 * inside pull_task().
2245 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2248 *all_pinned
= pinned
;
2253 * find_busiest_group finds and returns the busiest CPU group within the
2254 * domain. It calculates and returns the amount of weighted load which
2255 * should be moved to restore balance via the imbalance parameter.
2257 static struct sched_group
*
2258 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2259 unsigned long *imbalance
, enum idle_type idle
, int *sd_idle
,
2260 cpumask_t
*cpus
, int *balance
)
2262 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2263 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2264 unsigned long max_pull
;
2265 unsigned long busiest_load_per_task
, busiest_nr_running
;
2266 unsigned long this_load_per_task
, this_nr_running
;
2268 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2269 int power_savings_balance
= 1;
2270 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2271 unsigned long min_nr_running
= ULONG_MAX
;
2272 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2275 max_load
= this_load
= total_load
= total_pwr
= 0;
2276 busiest_load_per_task
= busiest_nr_running
= 0;
2277 this_load_per_task
= this_nr_running
= 0;
2278 if (idle
== NOT_IDLE
)
2279 load_idx
= sd
->busy_idx
;
2280 else if (idle
== NEWLY_IDLE
)
2281 load_idx
= sd
->newidle_idx
;
2283 load_idx
= sd
->idle_idx
;
2286 unsigned long load
, group_capacity
;
2289 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2290 unsigned long sum_nr_running
, sum_weighted_load
;
2292 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2295 balance_cpu
= first_cpu(group
->cpumask
);
2297 /* Tally up the load of all CPUs in the group */
2298 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2300 for_each_cpu_mask(i
, group
->cpumask
) {
2303 if (!cpu_isset(i
, *cpus
))
2308 if (*sd_idle
&& !idle_cpu(i
))
2311 /* Bias balancing toward cpus of our domain */
2313 if (idle_cpu(i
) && !first_idle_cpu
) {
2318 load
= target_load(i
, load_idx
);
2320 load
= source_load(i
, load_idx
);
2323 sum_nr_running
+= rq
->nr_running
;
2324 sum_weighted_load
+= rq
->raw_weighted_load
;
2328 * First idle cpu or the first cpu(busiest) in this sched group
2329 * is eligible for doing load balancing at this and above
2332 if (local_group
&& balance_cpu
!= this_cpu
&& balance
) {
2337 total_load
+= avg_load
;
2338 total_pwr
+= group
->cpu_power
;
2340 /* Adjust by relative CPU power of the group */
2341 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
2343 group_capacity
= group
->cpu_power
/ SCHED_LOAD_SCALE
;
2346 this_load
= avg_load
;
2348 this_nr_running
= sum_nr_running
;
2349 this_load_per_task
= sum_weighted_load
;
2350 } else if (avg_load
> max_load
&&
2351 sum_nr_running
> group_capacity
) {
2352 max_load
= avg_load
;
2354 busiest_nr_running
= sum_nr_running
;
2355 busiest_load_per_task
= sum_weighted_load
;
2358 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2360 * Busy processors will not participate in power savings
2363 if (idle
== NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2367 * If the local group is idle or completely loaded
2368 * no need to do power savings balance at this domain
2370 if (local_group
&& (this_nr_running
>= group_capacity
||
2372 power_savings_balance
= 0;
2375 * If a group is already running at full capacity or idle,
2376 * don't include that group in power savings calculations
2378 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2383 * Calculate the group which has the least non-idle load.
2384 * This is the group from where we need to pick up the load
2387 if ((sum_nr_running
< min_nr_running
) ||
2388 (sum_nr_running
== min_nr_running
&&
2389 first_cpu(group
->cpumask
) <
2390 first_cpu(group_min
->cpumask
))) {
2392 min_nr_running
= sum_nr_running
;
2393 min_load_per_task
= sum_weighted_load
/
2398 * Calculate the group which is almost near its
2399 * capacity but still has some space to pick up some load
2400 * from other group and save more power
2402 if (sum_nr_running
<= group_capacity
- 1) {
2403 if (sum_nr_running
> leader_nr_running
||
2404 (sum_nr_running
== leader_nr_running
&&
2405 first_cpu(group
->cpumask
) >
2406 first_cpu(group_leader
->cpumask
))) {
2407 group_leader
= group
;
2408 leader_nr_running
= sum_nr_running
;
2413 group
= group
->next
;
2414 } while (group
!= sd
->groups
);
2416 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2419 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2421 if (this_load
>= avg_load
||
2422 100*max_load
<= sd
->imbalance_pct
*this_load
)
2425 busiest_load_per_task
/= busiest_nr_running
;
2427 * We're trying to get all the cpus to the average_load, so we don't
2428 * want to push ourselves above the average load, nor do we wish to
2429 * reduce the max loaded cpu below the average load, as either of these
2430 * actions would just result in more rebalancing later, and ping-pong
2431 * tasks around. Thus we look for the minimum possible imbalance.
2432 * Negative imbalances (*we* are more loaded than anyone else) will
2433 * be counted as no imbalance for these purposes -- we can't fix that
2434 * by pulling tasks to us. Be careful of negative numbers as they'll
2435 * appear as very large values with unsigned longs.
2437 if (max_load
<= busiest_load_per_task
)
2441 * In the presence of smp nice balancing, certain scenarios can have
2442 * max load less than avg load(as we skip the groups at or below
2443 * its cpu_power, while calculating max_load..)
2445 if (max_load
< avg_load
) {
2447 goto small_imbalance
;
2450 /* Don't want to pull so many tasks that a group would go idle */
2451 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2453 /* How much load to actually move to equalise the imbalance */
2454 *imbalance
= min(max_pull
* busiest
->cpu_power
,
2455 (avg_load
- this_load
) * this->cpu_power
)
2459 * if *imbalance is less than the average load per runnable task
2460 * there is no gaurantee that any tasks will be moved so we'll have
2461 * a think about bumping its value to force at least one task to be
2464 if (*imbalance
< busiest_load_per_task
) {
2465 unsigned long tmp
, pwr_now
, pwr_move
;
2469 pwr_move
= pwr_now
= 0;
2471 if (this_nr_running
) {
2472 this_load_per_task
/= this_nr_running
;
2473 if (busiest_load_per_task
> this_load_per_task
)
2476 this_load_per_task
= SCHED_LOAD_SCALE
;
2478 if (max_load
- this_load
>= busiest_load_per_task
* imbn
) {
2479 *imbalance
= busiest_load_per_task
;
2484 * OK, we don't have enough imbalance to justify moving tasks,
2485 * however we may be able to increase total CPU power used by
2489 pwr_now
+= busiest
->cpu_power
*
2490 min(busiest_load_per_task
, max_load
);
2491 pwr_now
+= this->cpu_power
*
2492 min(this_load_per_task
, this_load
);
2493 pwr_now
/= SCHED_LOAD_SCALE
;
2495 /* Amount of load we'd subtract */
2496 tmp
= busiest_load_per_task
* SCHED_LOAD_SCALE
/
2499 pwr_move
+= busiest
->cpu_power
*
2500 min(busiest_load_per_task
, max_load
- tmp
);
2502 /* Amount of load we'd add */
2503 if (max_load
* busiest
->cpu_power
<
2504 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2505 tmp
= max_load
* busiest
->cpu_power
/ this->cpu_power
;
2507 tmp
= busiest_load_per_task
* SCHED_LOAD_SCALE
/
2509 pwr_move
+= this->cpu_power
*
2510 min(this_load_per_task
, this_load
+ tmp
);
2511 pwr_move
/= SCHED_LOAD_SCALE
;
2513 /* Move if we gain throughput */
2514 if (pwr_move
<= pwr_now
)
2517 *imbalance
= busiest_load_per_task
;
2523 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2524 if (idle
== NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2527 if (this == group_leader
&& group_leader
!= group_min
) {
2528 *imbalance
= min_load_per_task
;
2538 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2541 find_busiest_queue(struct sched_group
*group
, enum idle_type idle
,
2542 unsigned long imbalance
, cpumask_t
*cpus
)
2544 struct rq
*busiest
= NULL
, *rq
;
2545 unsigned long max_load
= 0;
2548 for_each_cpu_mask(i
, group
->cpumask
) {
2550 if (!cpu_isset(i
, *cpus
))
2555 if (rq
->nr_running
== 1 && rq
->raw_weighted_load
> imbalance
)
2558 if (rq
->raw_weighted_load
> max_load
) {
2559 max_load
= rq
->raw_weighted_load
;
2568 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2569 * so long as it is large enough.
2571 #define MAX_PINNED_INTERVAL 512
2573 static inline unsigned long minus_1_or_zero(unsigned long n
)
2575 return n
> 0 ? n
- 1 : 0;
2579 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2580 * tasks if there is an imbalance.
2582 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2583 struct sched_domain
*sd
, enum idle_type idle
,
2586 int nr_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2587 struct sched_group
*group
;
2588 unsigned long imbalance
;
2590 cpumask_t cpus
= CPU_MASK_ALL
;
2591 unsigned long flags
;
2594 * When power savings policy is enabled for the parent domain, idle
2595 * sibling can pick up load irrespective of busy siblings. In this case,
2596 * let the state of idle sibling percolate up as IDLE, instead of
2597 * portraying it as NOT_IDLE.
2599 if (idle
!= NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2600 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2603 schedstat_inc(sd
, lb_cnt
[idle
]);
2606 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2613 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2617 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2619 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2623 BUG_ON(busiest
== this_rq
);
2625 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2628 if (busiest
->nr_running
> 1) {
2630 * Attempt to move tasks. If find_busiest_group has found
2631 * an imbalance but busiest->nr_running <= 1, the group is
2632 * still unbalanced. nr_moved simply stays zero, so it is
2633 * correctly treated as an imbalance.
2635 local_irq_save(flags
);
2636 double_rq_lock(this_rq
, busiest
);
2637 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2638 minus_1_or_zero(busiest
->nr_running
),
2639 imbalance
, sd
, idle
, &all_pinned
);
2640 double_rq_unlock(this_rq
, busiest
);
2641 local_irq_restore(flags
);
2643 /* All tasks on this runqueue were pinned by CPU affinity */
2644 if (unlikely(all_pinned
)) {
2645 cpu_clear(cpu_of(busiest
), cpus
);
2646 if (!cpus_empty(cpus
))
2653 schedstat_inc(sd
, lb_failed
[idle
]);
2654 sd
->nr_balance_failed
++;
2656 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2658 spin_lock_irqsave(&busiest
->lock
, flags
);
2660 /* don't kick the migration_thread, if the curr
2661 * task on busiest cpu can't be moved to this_cpu
2663 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2664 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2666 goto out_one_pinned
;
2669 if (!busiest
->active_balance
) {
2670 busiest
->active_balance
= 1;
2671 busiest
->push_cpu
= this_cpu
;
2674 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2676 wake_up_process(busiest
->migration_thread
);
2679 * We've kicked active balancing, reset the failure
2682 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2685 sd
->nr_balance_failed
= 0;
2687 if (likely(!active_balance
)) {
2688 /* We were unbalanced, so reset the balancing interval */
2689 sd
->balance_interval
= sd
->min_interval
;
2692 * If we've begun active balancing, start to back off. This
2693 * case may not be covered by the all_pinned logic if there
2694 * is only 1 task on the busy runqueue (because we don't call
2697 if (sd
->balance_interval
< sd
->max_interval
)
2698 sd
->balance_interval
*= 2;
2701 if (!nr_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2702 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2707 schedstat_inc(sd
, lb_balanced
[idle
]);
2709 sd
->nr_balance_failed
= 0;
2712 /* tune up the balancing interval */
2713 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2714 (sd
->balance_interval
< sd
->max_interval
))
2715 sd
->balance_interval
*= 2;
2717 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2718 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2724 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2725 * tasks if there is an imbalance.
2727 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2728 * this_rq is locked.
2731 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
2733 struct sched_group
*group
;
2734 struct rq
*busiest
= NULL
;
2735 unsigned long imbalance
;
2738 cpumask_t cpus
= CPU_MASK_ALL
;
2741 * When power savings policy is enabled for the parent domain, idle
2742 * sibling can pick up load irrespective of busy siblings. In this case,
2743 * let the state of idle sibling percolate up as IDLE, instead of
2744 * portraying it as NOT_IDLE.
2746 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
2747 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2750 schedstat_inc(sd
, lb_cnt
[NEWLY_IDLE
]);
2752 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, NEWLY_IDLE
,
2753 &sd_idle
, &cpus
, NULL
);
2755 schedstat_inc(sd
, lb_nobusyg
[NEWLY_IDLE
]);
2759 busiest
= find_busiest_queue(group
, NEWLY_IDLE
, imbalance
,
2762 schedstat_inc(sd
, lb_nobusyq
[NEWLY_IDLE
]);
2766 BUG_ON(busiest
== this_rq
);
2768 schedstat_add(sd
, lb_imbalance
[NEWLY_IDLE
], imbalance
);
2771 if (busiest
->nr_running
> 1) {
2772 /* Attempt to move tasks */
2773 double_lock_balance(this_rq
, busiest
);
2774 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2775 minus_1_or_zero(busiest
->nr_running
),
2776 imbalance
, sd
, NEWLY_IDLE
, NULL
);
2777 spin_unlock(&busiest
->lock
);
2780 cpu_clear(cpu_of(busiest
), cpus
);
2781 if (!cpus_empty(cpus
))
2787 schedstat_inc(sd
, lb_failed
[NEWLY_IDLE
]);
2788 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2789 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2792 sd
->nr_balance_failed
= 0;
2797 schedstat_inc(sd
, lb_balanced
[NEWLY_IDLE
]);
2798 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2799 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2801 sd
->nr_balance_failed
= 0;
2807 * idle_balance is called by schedule() if this_cpu is about to become
2808 * idle. Attempts to pull tasks from other CPUs.
2810 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
2812 struct sched_domain
*sd
;
2813 int pulled_task
= 0;
2814 unsigned long next_balance
= jiffies
+ 60 * HZ
;
2816 for_each_domain(this_cpu
, sd
) {
2817 unsigned long interval
;
2819 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2822 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
2823 /* If we've pulled tasks over stop searching: */
2824 pulled_task
= load_balance_newidle(this_cpu
,
2827 interval
= msecs_to_jiffies(sd
->balance_interval
);
2828 if (time_after(next_balance
, sd
->last_balance
+ interval
))
2829 next_balance
= sd
->last_balance
+ interval
;
2835 * We are going idle. next_balance may be set based on
2836 * a busy processor. So reset next_balance.
2838 this_rq
->next_balance
= next_balance
;
2842 * active_load_balance is run by migration threads. It pushes running tasks
2843 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2844 * running on each physical CPU where possible, and avoids physical /
2845 * logical imbalances.
2847 * Called with busiest_rq locked.
2849 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
2851 int target_cpu
= busiest_rq
->push_cpu
;
2852 struct sched_domain
*sd
;
2853 struct rq
*target_rq
;
2855 /* Is there any task to move? */
2856 if (busiest_rq
->nr_running
<= 1)
2859 target_rq
= cpu_rq(target_cpu
);
2862 * This condition is "impossible", if it occurs
2863 * we need to fix it. Originally reported by
2864 * Bjorn Helgaas on a 128-cpu setup.
2866 BUG_ON(busiest_rq
== target_rq
);
2868 /* move a task from busiest_rq to target_rq */
2869 double_lock_balance(busiest_rq
, target_rq
);
2871 /* Search for an sd spanning us and the target CPU. */
2872 for_each_domain(target_cpu
, sd
) {
2873 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2874 cpu_isset(busiest_cpu
, sd
->span
))
2879 schedstat_inc(sd
, alb_cnt
);
2881 if (move_tasks(target_rq
, target_cpu
, busiest_rq
, 1,
2882 RTPRIO_TO_LOAD_WEIGHT(100), sd
, SCHED_IDLE
,
2884 schedstat_inc(sd
, alb_pushed
);
2886 schedstat_inc(sd
, alb_failed
);
2888 spin_unlock(&target_rq
->lock
);
2891 static void update_load(struct rq
*this_rq
)
2893 unsigned long this_load
;
2896 this_load
= this_rq
->raw_weighted_load
;
2898 /* Update our load: */
2899 for (i
= 0, scale
= 1; i
< 3; i
++, scale
<<= 1) {
2900 unsigned long old_load
, new_load
;
2902 old_load
= this_rq
->cpu_load
[i
];
2903 new_load
= this_load
;
2905 * Round up the averaging division if load is increasing. This
2906 * prevents us from getting stuck on 9 if the load is 10, for
2909 if (new_load
> old_load
)
2910 new_load
+= scale
-1;
2911 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) / scale
;
2916 * run_rebalance_domains is triggered when needed from the scheduler tick.
2918 * It checks each scheduling domain to see if it is due to be balanced,
2919 * and initiates a balancing operation if so.
2921 * Balancing parameters are set up in arch_init_sched_domains.
2923 static DEFINE_SPINLOCK(balancing
);
2925 static void run_rebalance_domains(struct softirq_action
*h
)
2927 int this_cpu
= smp_processor_id(), balance
= 1;
2928 struct rq
*this_rq
= cpu_rq(this_cpu
);
2929 unsigned long interval
;
2930 struct sched_domain
*sd
;
2932 * We are idle if there are no processes running. This
2933 * is valid even if we are the idle process (SMT).
2935 enum idle_type idle
= !this_rq
->nr_running
?
2936 SCHED_IDLE
: NOT_IDLE
;
2937 /* Earliest time when we have to call run_rebalance_domains again */
2938 unsigned long next_balance
= jiffies
+ 60*HZ
;
2940 for_each_domain(this_cpu
, sd
) {
2941 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2944 interval
= sd
->balance_interval
;
2945 if (idle
!= SCHED_IDLE
)
2946 interval
*= sd
->busy_factor
;
2948 /* scale ms to jiffies */
2949 interval
= msecs_to_jiffies(interval
);
2950 if (unlikely(!interval
))
2953 if (sd
->flags
& SD_SERIALIZE
) {
2954 if (!spin_trylock(&balancing
))
2958 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
2959 if (load_balance(this_cpu
, this_rq
, sd
, idle
, &balance
)) {
2961 * We've pulled tasks over so either we're no
2962 * longer idle, or one of our SMT siblings is
2967 sd
->last_balance
= jiffies
;
2969 if (sd
->flags
& SD_SERIALIZE
)
2970 spin_unlock(&balancing
);
2972 if (time_after(next_balance
, sd
->last_balance
+ interval
))
2973 next_balance
= sd
->last_balance
+ interval
;
2976 * Stop the load balance at this level. There is another
2977 * CPU in our sched group which is doing load balancing more
2983 this_rq
->next_balance
= next_balance
;
2987 * on UP we do not need to balance between CPUs:
2989 static inline void idle_balance(int cpu
, struct rq
*rq
)
2994 static inline void wake_priority_sleeper(struct rq
*rq
)
2996 #ifdef CONFIG_SCHED_SMT
2997 if (!rq
->nr_running
)
3000 spin_lock(&rq
->lock
);
3002 * If an SMT sibling task has been put to sleep for priority
3003 * reasons reschedule the idle task to see if it can now run.
3006 resched_task(rq
->idle
);
3007 spin_unlock(&rq
->lock
);
3011 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3013 EXPORT_PER_CPU_SYMBOL(kstat
);
3016 * This is called on clock ticks and on context switches.
3017 * Bank in p->sched_time the ns elapsed since the last tick or switch.
3020 update_cpu_clock(struct task_struct
*p
, struct rq
*rq
, unsigned long long now
)
3022 p
->sched_time
+= now
- p
->last_ran
;
3023 p
->last_ran
= rq
->most_recent_timestamp
= now
;
3027 * Return current->sched_time plus any more ns on the sched_clock
3028 * that have not yet been banked.
3030 unsigned long long current_sched_time(const struct task_struct
*p
)
3032 unsigned long long ns
;
3033 unsigned long flags
;
3035 local_irq_save(flags
);
3036 ns
= p
->sched_time
+ sched_clock() - p
->last_ran
;
3037 local_irq_restore(flags
);
3043 * We place interactive tasks back into the active array, if possible.
3045 * To guarantee that this does not starve expired tasks we ignore the
3046 * interactivity of a task if the first expired task had to wait more
3047 * than a 'reasonable' amount of time. This deadline timeout is
3048 * load-dependent, as the frequency of array switched decreases with
3049 * increasing number of running tasks. We also ignore the interactivity
3050 * if a better static_prio task has expired:
3052 static inline int expired_starving(struct rq
*rq
)
3054 if (rq
->curr
->static_prio
> rq
->best_expired_prio
)
3056 if (!STARVATION_LIMIT
|| !rq
->expired_timestamp
)
3058 if (jiffies
- rq
->expired_timestamp
> STARVATION_LIMIT
* rq
->nr_running
)
3064 * Account user cpu time to a process.
3065 * @p: the process that the cpu time gets accounted to
3066 * @hardirq_offset: the offset to subtract from hardirq_count()
3067 * @cputime: the cpu time spent in user space since the last update
3069 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3071 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3074 p
->utime
= cputime_add(p
->utime
, cputime
);
3076 /* Add user time to cpustat. */
3077 tmp
= cputime_to_cputime64(cputime
);
3078 if (TASK_NICE(p
) > 0)
3079 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3081 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3085 * Account system cpu time to a process.
3086 * @p: the process that the cpu time gets accounted to
3087 * @hardirq_offset: the offset to subtract from hardirq_count()
3088 * @cputime: the cpu time spent in kernel space since the last update
3090 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3093 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3094 struct rq
*rq
= this_rq();
3097 p
->stime
= cputime_add(p
->stime
, cputime
);
3099 /* Add system time to cpustat. */
3100 tmp
= cputime_to_cputime64(cputime
);
3101 if (hardirq_count() - hardirq_offset
)
3102 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3103 else if (softirq_count())
3104 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3105 else if (p
!= rq
->idle
)
3106 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3107 else if (atomic_read(&rq
->nr_iowait
) > 0)
3108 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3110 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3111 /* Account for system time used */
3112 acct_update_integrals(p
);
3116 * Account for involuntary wait time.
3117 * @p: the process from which the cpu time has been stolen
3118 * @steal: the cpu time spent in involuntary wait
3120 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3122 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3123 cputime64_t tmp
= cputime_to_cputime64(steal
);
3124 struct rq
*rq
= this_rq();
3126 if (p
== rq
->idle
) {
3127 p
->stime
= cputime_add(p
->stime
, steal
);
3128 if (atomic_read(&rq
->nr_iowait
) > 0)
3129 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3131 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3133 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3136 static void task_running_tick(struct rq
*rq
, struct task_struct
*p
)
3138 if (p
->array
!= rq
->active
) {
3139 /* Task has expired but was not scheduled yet */
3140 set_tsk_need_resched(p
);
3143 spin_lock(&rq
->lock
);
3145 * The task was running during this tick - update the
3146 * time slice counter. Note: we do not update a thread's
3147 * priority until it either goes to sleep or uses up its
3148 * timeslice. This makes it possible for interactive tasks
3149 * to use up their timeslices at their highest priority levels.
3153 * RR tasks need a special form of timeslice management.
3154 * FIFO tasks have no timeslices.
3156 if ((p
->policy
== SCHED_RR
) && !--p
->time_slice
) {
3157 p
->time_slice
= task_timeslice(p
);
3158 p
->first_time_slice
= 0;
3159 set_tsk_need_resched(p
);
3161 /* put it at the end of the queue: */
3162 requeue_task(p
, rq
->active
);
3166 if (!--p
->time_slice
) {
3167 dequeue_task(p
, rq
->active
);
3168 set_tsk_need_resched(p
);
3169 p
->prio
= effective_prio(p
);
3170 p
->time_slice
= task_timeslice(p
);
3171 p
->first_time_slice
= 0;
3173 if (!rq
->expired_timestamp
)
3174 rq
->expired_timestamp
= jiffies
;
3175 if (!TASK_INTERACTIVE(p
) || expired_starving(rq
)) {
3176 enqueue_task(p
, rq
->expired
);
3177 if (p
->static_prio
< rq
->best_expired_prio
)
3178 rq
->best_expired_prio
= p
->static_prio
;
3180 enqueue_task(p
, rq
->active
);
3183 * Prevent a too long timeslice allowing a task to monopolize
3184 * the CPU. We do this by splitting up the timeslice into
3187 * Note: this does not mean the task's timeslices expire or
3188 * get lost in any way, they just might be preempted by
3189 * another task of equal priority. (one with higher
3190 * priority would have preempted this task already.) We
3191 * requeue this task to the end of the list on this priority
3192 * level, which is in essence a round-robin of tasks with
3195 * This only applies to tasks in the interactive
3196 * delta range with at least TIMESLICE_GRANULARITY to requeue.
3198 if (TASK_INTERACTIVE(p
) && !((task_timeslice(p
) -
3199 p
->time_slice
) % TIMESLICE_GRANULARITY(p
)) &&
3200 (p
->time_slice
>= TIMESLICE_GRANULARITY(p
)) &&
3201 (p
->array
== rq
->active
)) {
3203 requeue_task(p
, rq
->active
);
3204 set_tsk_need_resched(p
);
3208 spin_unlock(&rq
->lock
);
3212 * This function gets called by the timer code, with HZ frequency.
3213 * We call it with interrupts disabled.
3215 * It also gets called by the fork code, when changing the parent's
3218 void scheduler_tick(void)
3220 unsigned long long now
= sched_clock();
3221 struct task_struct
*p
= current
;
3222 int cpu
= smp_processor_id();
3223 struct rq
*rq
= cpu_rq(cpu
);
3225 update_cpu_clock(p
, rq
, now
);
3228 /* Task on the idle queue */
3229 wake_priority_sleeper(rq
);
3231 task_running_tick(rq
, p
);
3234 if (time_after_eq(jiffies
, rq
->next_balance
))
3235 raise_softirq(SCHED_SOFTIRQ
);
3239 #ifdef CONFIG_SCHED_SMT
3240 static inline void wakeup_busy_runqueue(struct rq
*rq
)
3242 /* If an SMT runqueue is sleeping due to priority reasons wake it up */
3243 if (rq
->curr
== rq
->idle
&& rq
->nr_running
)
3244 resched_task(rq
->idle
);
3248 * Called with interrupt disabled and this_rq's runqueue locked.
3250 static void wake_sleeping_dependent(int this_cpu
)
3252 struct sched_domain
*tmp
, *sd
= NULL
;
3255 for_each_domain(this_cpu
, tmp
) {
3256 if (tmp
->flags
& SD_SHARE_CPUPOWER
) {
3265 for_each_cpu_mask(i
, sd
->span
) {
3266 struct rq
*smt_rq
= cpu_rq(i
);
3270 if (unlikely(!spin_trylock(&smt_rq
->lock
)))
3273 wakeup_busy_runqueue(smt_rq
);
3274 spin_unlock(&smt_rq
->lock
);
3279 * number of 'lost' timeslices this task wont be able to fully
3280 * utilize, if another task runs on a sibling. This models the
3281 * slowdown effect of other tasks running on siblings:
3283 static inline unsigned long
3284 smt_slice(struct task_struct
*p
, struct sched_domain
*sd
)
3286 return p
->time_slice
* (100 - sd
->per_cpu_gain
) / 100;
3290 * To minimise lock contention and not have to drop this_rq's runlock we only
3291 * trylock the sibling runqueues and bypass those runqueues if we fail to
3292 * acquire their lock. As we only trylock the normal locking order does not
3293 * need to be obeyed.
3296 dependent_sleeper(int this_cpu
, struct rq
*this_rq
, struct task_struct
*p
)
3298 struct sched_domain
*tmp
, *sd
= NULL
;
3301 /* kernel/rt threads do not participate in dependent sleeping */
3302 if (!p
->mm
|| rt_task(p
))
3305 for_each_domain(this_cpu
, tmp
) {
3306 if (tmp
->flags
& SD_SHARE_CPUPOWER
) {
3315 for_each_cpu_mask(i
, sd
->span
) {
3316 struct task_struct
*smt_curr
;
3323 if (unlikely(!spin_trylock(&smt_rq
->lock
)))
3326 smt_curr
= smt_rq
->curr
;
3332 * If a user task with lower static priority than the
3333 * running task on the SMT sibling is trying to schedule,
3334 * delay it till there is proportionately less timeslice
3335 * left of the sibling task to prevent a lower priority
3336 * task from using an unfair proportion of the
3337 * physical cpu's resources. -ck
3339 if (rt_task(smt_curr
)) {
3341 * With real time tasks we run non-rt tasks only
3342 * per_cpu_gain% of the time.
3344 if ((jiffies
% DEF_TIMESLICE
) >
3345 (sd
->per_cpu_gain
* DEF_TIMESLICE
/ 100))
3348 if (smt_curr
->static_prio
< p
->static_prio
&&
3349 !TASK_PREEMPTS_CURR(p
, smt_rq
) &&
3350 smt_slice(smt_curr
, sd
) > task_timeslice(p
))
3354 spin_unlock(&smt_rq
->lock
);
3359 static inline void wake_sleeping_dependent(int this_cpu
)
3363 dependent_sleeper(int this_cpu
, struct rq
*this_rq
, struct task_struct
*p
)
3369 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3371 void fastcall
add_preempt_count(int val
)
3376 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3378 preempt_count() += val
;
3380 * Spinlock count overflowing soon?
3382 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3385 EXPORT_SYMBOL(add_preempt_count
);
3387 void fastcall
sub_preempt_count(int val
)
3392 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3395 * Is the spinlock portion underflowing?
3397 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3398 !(preempt_count() & PREEMPT_MASK
)))
3401 preempt_count() -= val
;
3403 EXPORT_SYMBOL(sub_preempt_count
);
3407 static inline int interactive_sleep(enum sleep_type sleep_type
)
3409 return (sleep_type
== SLEEP_INTERACTIVE
||
3410 sleep_type
== SLEEP_INTERRUPTED
);
3414 * schedule() is the main scheduler function.
3416 asmlinkage
void __sched
schedule(void)
3418 struct task_struct
*prev
, *next
;
3419 struct prio_array
*array
;
3420 struct list_head
*queue
;
3421 unsigned long long now
;
3422 unsigned long run_time
;
3423 int cpu
, idx
, new_prio
;
3428 * Test if we are atomic. Since do_exit() needs to call into
3429 * schedule() atomically, we ignore that path for now.
3430 * Otherwise, whine if we are scheduling when we should not be.
3432 if (unlikely(in_atomic() && !current
->exit_state
)) {
3433 printk(KERN_ERR
"BUG: scheduling while atomic: "
3435 current
->comm
, preempt_count(), current
->pid
);
3436 debug_show_held_locks(current
);
3437 if (irqs_disabled())
3438 print_irqtrace_events(current
);
3441 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3446 release_kernel_lock(prev
);
3447 need_resched_nonpreemptible
:
3451 * The idle thread is not allowed to schedule!
3452 * Remove this check after it has been exercised a bit.
3454 if (unlikely(prev
== rq
->idle
) && prev
->state
!= TASK_RUNNING
) {
3455 printk(KERN_ERR
"bad: scheduling from the idle thread!\n");
3459 schedstat_inc(rq
, sched_cnt
);
3460 now
= sched_clock();
3461 if (likely((long long)(now
- prev
->timestamp
) < NS_MAX_SLEEP_AVG
)) {
3462 run_time
= now
- prev
->timestamp
;
3463 if (unlikely((long long)(now
- prev
->timestamp
) < 0))
3466 run_time
= NS_MAX_SLEEP_AVG
;
3469 * Tasks charged proportionately less run_time at high sleep_avg to
3470 * delay them losing their interactive status
3472 run_time
/= (CURRENT_BONUS(prev
) ? : 1);
3474 spin_lock_irq(&rq
->lock
);
3476 switch_count
= &prev
->nivcsw
;
3477 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3478 switch_count
= &prev
->nvcsw
;
3479 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3480 unlikely(signal_pending(prev
))))
3481 prev
->state
= TASK_RUNNING
;
3483 if (prev
->state
== TASK_UNINTERRUPTIBLE
)
3484 rq
->nr_uninterruptible
++;
3485 deactivate_task(prev
, rq
);
3489 cpu
= smp_processor_id();
3490 if (unlikely(!rq
->nr_running
)) {
3491 idle_balance(cpu
, rq
);
3492 if (!rq
->nr_running
) {
3494 rq
->expired_timestamp
= 0;
3495 wake_sleeping_dependent(cpu
);
3501 if (unlikely(!array
->nr_active
)) {
3503 * Switch the active and expired arrays.
3505 schedstat_inc(rq
, sched_switch
);
3506 rq
->active
= rq
->expired
;
3507 rq
->expired
= array
;
3509 rq
->expired_timestamp
= 0;
3510 rq
->best_expired_prio
= MAX_PRIO
;
3513 idx
= sched_find_first_bit(array
->bitmap
);
3514 queue
= array
->queue
+ idx
;
3515 next
= list_entry(queue
->next
, struct task_struct
, run_list
);
3517 if (!rt_task(next
) && interactive_sleep(next
->sleep_type
)) {
3518 unsigned long long delta
= now
- next
->timestamp
;
3519 if (unlikely((long long)(now
- next
->timestamp
) < 0))
3522 if (next
->sleep_type
== SLEEP_INTERACTIVE
)
3523 delta
= delta
* (ON_RUNQUEUE_WEIGHT
* 128 / 100) / 128;
3525 array
= next
->array
;
3526 new_prio
= recalc_task_prio(next
, next
->timestamp
+ delta
);
3528 if (unlikely(next
->prio
!= new_prio
)) {
3529 dequeue_task(next
, array
);
3530 next
->prio
= new_prio
;
3531 enqueue_task(next
, array
);
3534 next
->sleep_type
= SLEEP_NORMAL
;
3535 if (rq
->nr_running
== 1 && dependent_sleeper(cpu
, rq
, next
))
3538 if (next
== rq
->idle
)
3539 schedstat_inc(rq
, sched_goidle
);
3541 prefetch_stack(next
);
3542 clear_tsk_need_resched(prev
);
3543 rcu_qsctr_inc(task_cpu(prev
));
3545 update_cpu_clock(prev
, rq
, now
);
3547 prev
->sleep_avg
-= run_time
;
3548 if ((long)prev
->sleep_avg
<= 0)
3549 prev
->sleep_avg
= 0;
3550 prev
->timestamp
= prev
->last_ran
= now
;
3552 sched_info_switch(prev
, next
);
3553 if (likely(prev
!= next
)) {
3554 next
->timestamp
= next
->last_ran
= now
;
3559 prepare_task_switch(rq
, next
);
3560 prev
= context_switch(rq
, prev
, next
);
3563 * this_rq must be evaluated again because prev may have moved
3564 * CPUs since it called schedule(), thus the 'rq' on its stack
3565 * frame will be invalid.
3567 finish_task_switch(this_rq(), prev
);
3569 spin_unlock_irq(&rq
->lock
);
3572 if (unlikely(reacquire_kernel_lock(prev
) < 0))
3573 goto need_resched_nonpreemptible
;
3574 preempt_enable_no_resched();
3575 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3578 EXPORT_SYMBOL(schedule
);
3580 #ifdef CONFIG_PREEMPT
3582 * this is the entry point to schedule() from in-kernel preemption
3583 * off of preempt_enable. Kernel preemptions off return from interrupt
3584 * occur there and call schedule directly.
3586 asmlinkage
void __sched
preempt_schedule(void)
3588 struct thread_info
*ti
= current_thread_info();
3589 #ifdef CONFIG_PREEMPT_BKL
3590 struct task_struct
*task
= current
;
3591 int saved_lock_depth
;
3594 * If there is a non-zero preempt_count or interrupts are disabled,
3595 * we do not want to preempt the current task. Just return..
3597 if (likely(ti
->preempt_count
|| irqs_disabled()))
3601 add_preempt_count(PREEMPT_ACTIVE
);
3603 * We keep the big kernel semaphore locked, but we
3604 * clear ->lock_depth so that schedule() doesnt
3605 * auto-release the semaphore:
3607 #ifdef CONFIG_PREEMPT_BKL
3608 saved_lock_depth
= task
->lock_depth
;
3609 task
->lock_depth
= -1;
3612 #ifdef CONFIG_PREEMPT_BKL
3613 task
->lock_depth
= saved_lock_depth
;
3615 sub_preempt_count(PREEMPT_ACTIVE
);
3617 /* we could miss a preemption opportunity between schedule and now */
3619 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3622 EXPORT_SYMBOL(preempt_schedule
);
3625 * this is the entry point to schedule() from kernel preemption
3626 * off of irq context.
3627 * Note, that this is called and return with irqs disabled. This will
3628 * protect us against recursive calling from irq.
3630 asmlinkage
void __sched
preempt_schedule_irq(void)
3632 struct thread_info
*ti
= current_thread_info();
3633 #ifdef CONFIG_PREEMPT_BKL
3634 struct task_struct
*task
= current
;
3635 int saved_lock_depth
;
3637 /* Catch callers which need to be fixed */
3638 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3641 add_preempt_count(PREEMPT_ACTIVE
);
3643 * We keep the big kernel semaphore locked, but we
3644 * clear ->lock_depth so that schedule() doesnt
3645 * auto-release the semaphore:
3647 #ifdef CONFIG_PREEMPT_BKL
3648 saved_lock_depth
= task
->lock_depth
;
3649 task
->lock_depth
= -1;
3653 local_irq_disable();
3654 #ifdef CONFIG_PREEMPT_BKL
3655 task
->lock_depth
= saved_lock_depth
;
3657 sub_preempt_count(PREEMPT_ACTIVE
);
3659 /* we could miss a preemption opportunity between schedule and now */
3661 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3665 #endif /* CONFIG_PREEMPT */
3667 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3670 return try_to_wake_up(curr
->private, mode
, sync
);
3672 EXPORT_SYMBOL(default_wake_function
);
3675 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3676 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3677 * number) then we wake all the non-exclusive tasks and one exclusive task.
3679 * There are circumstances in which we can try to wake a task which has already
3680 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3681 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3683 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3684 int nr_exclusive
, int sync
, void *key
)
3686 struct list_head
*tmp
, *next
;
3688 list_for_each_safe(tmp
, next
, &q
->task_list
) {
3689 wait_queue_t
*curr
= list_entry(tmp
, wait_queue_t
, task_list
);
3690 unsigned flags
= curr
->flags
;
3692 if (curr
->func(curr
, mode
, sync
, key
) &&
3693 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3699 * __wake_up - wake up threads blocked on a waitqueue.
3701 * @mode: which threads
3702 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3703 * @key: is directly passed to the wakeup function
3705 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3706 int nr_exclusive
, void *key
)
3708 unsigned long flags
;
3710 spin_lock_irqsave(&q
->lock
, flags
);
3711 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3712 spin_unlock_irqrestore(&q
->lock
, flags
);
3714 EXPORT_SYMBOL(__wake_up
);
3717 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3719 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3721 __wake_up_common(q
, mode
, 1, 0, NULL
);
3725 * __wake_up_sync - wake up threads blocked on a waitqueue.
3727 * @mode: which threads
3728 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3730 * The sync wakeup differs that the waker knows that it will schedule
3731 * away soon, so while the target thread will be woken up, it will not
3732 * be migrated to another CPU - ie. the two threads are 'synchronized'
3733 * with each other. This can prevent needless bouncing between CPUs.
3735 * On UP it can prevent extra preemption.
3738 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3740 unsigned long flags
;
3746 if (unlikely(!nr_exclusive
))
3749 spin_lock_irqsave(&q
->lock
, flags
);
3750 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3751 spin_unlock_irqrestore(&q
->lock
, flags
);
3753 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3755 void fastcall
complete(struct completion
*x
)
3757 unsigned long flags
;
3759 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3761 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3763 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3765 EXPORT_SYMBOL(complete
);
3767 void fastcall
complete_all(struct completion
*x
)
3769 unsigned long flags
;
3771 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3772 x
->done
+= UINT_MAX
/2;
3773 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3775 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3777 EXPORT_SYMBOL(complete_all
);
3779 void fastcall __sched
wait_for_completion(struct completion
*x
)
3783 spin_lock_irq(&x
->wait
.lock
);
3785 DECLARE_WAITQUEUE(wait
, current
);
3787 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3788 __add_wait_queue_tail(&x
->wait
, &wait
);
3790 __set_current_state(TASK_UNINTERRUPTIBLE
);
3791 spin_unlock_irq(&x
->wait
.lock
);
3793 spin_lock_irq(&x
->wait
.lock
);
3795 __remove_wait_queue(&x
->wait
, &wait
);
3798 spin_unlock_irq(&x
->wait
.lock
);
3800 EXPORT_SYMBOL(wait_for_completion
);
3802 unsigned long fastcall __sched
3803 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3807 spin_lock_irq(&x
->wait
.lock
);
3809 DECLARE_WAITQUEUE(wait
, current
);
3811 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3812 __add_wait_queue_tail(&x
->wait
, &wait
);
3814 __set_current_state(TASK_UNINTERRUPTIBLE
);
3815 spin_unlock_irq(&x
->wait
.lock
);
3816 timeout
= schedule_timeout(timeout
);
3817 spin_lock_irq(&x
->wait
.lock
);
3819 __remove_wait_queue(&x
->wait
, &wait
);
3823 __remove_wait_queue(&x
->wait
, &wait
);
3827 spin_unlock_irq(&x
->wait
.lock
);
3830 EXPORT_SYMBOL(wait_for_completion_timeout
);
3832 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3838 spin_lock_irq(&x
->wait
.lock
);
3840 DECLARE_WAITQUEUE(wait
, current
);
3842 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3843 __add_wait_queue_tail(&x
->wait
, &wait
);
3845 if (signal_pending(current
)) {
3847 __remove_wait_queue(&x
->wait
, &wait
);
3850 __set_current_state(TASK_INTERRUPTIBLE
);
3851 spin_unlock_irq(&x
->wait
.lock
);
3853 spin_lock_irq(&x
->wait
.lock
);
3855 __remove_wait_queue(&x
->wait
, &wait
);
3859 spin_unlock_irq(&x
->wait
.lock
);
3863 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3865 unsigned long fastcall __sched
3866 wait_for_completion_interruptible_timeout(struct completion
*x
,
3867 unsigned long timeout
)
3871 spin_lock_irq(&x
->wait
.lock
);
3873 DECLARE_WAITQUEUE(wait
, current
);
3875 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3876 __add_wait_queue_tail(&x
->wait
, &wait
);
3878 if (signal_pending(current
)) {
3879 timeout
= -ERESTARTSYS
;
3880 __remove_wait_queue(&x
->wait
, &wait
);
3883 __set_current_state(TASK_INTERRUPTIBLE
);
3884 spin_unlock_irq(&x
->wait
.lock
);
3885 timeout
= schedule_timeout(timeout
);
3886 spin_lock_irq(&x
->wait
.lock
);
3888 __remove_wait_queue(&x
->wait
, &wait
);
3892 __remove_wait_queue(&x
->wait
, &wait
);
3896 spin_unlock_irq(&x
->wait
.lock
);
3899 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3902 #define SLEEP_ON_VAR \
3903 unsigned long flags; \
3904 wait_queue_t wait; \
3905 init_waitqueue_entry(&wait, current);
3907 #define SLEEP_ON_HEAD \
3908 spin_lock_irqsave(&q->lock,flags); \
3909 __add_wait_queue(q, &wait); \
3910 spin_unlock(&q->lock);
3912 #define SLEEP_ON_TAIL \
3913 spin_lock_irq(&q->lock); \
3914 __remove_wait_queue(q, &wait); \
3915 spin_unlock_irqrestore(&q->lock, flags);
3917 void fastcall __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3921 current
->state
= TASK_INTERRUPTIBLE
;
3927 EXPORT_SYMBOL(interruptible_sleep_on
);
3929 long fastcall __sched
3930 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3934 current
->state
= TASK_INTERRUPTIBLE
;
3937 timeout
= schedule_timeout(timeout
);
3942 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3944 void fastcall __sched
sleep_on(wait_queue_head_t
*q
)
3948 current
->state
= TASK_UNINTERRUPTIBLE
;
3954 EXPORT_SYMBOL(sleep_on
);
3956 long fastcall __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3960 current
->state
= TASK_UNINTERRUPTIBLE
;
3963 timeout
= schedule_timeout(timeout
);
3969 EXPORT_SYMBOL(sleep_on_timeout
);
3971 #ifdef CONFIG_RT_MUTEXES
3974 * rt_mutex_setprio - set the current priority of a task
3976 * @prio: prio value (kernel-internal form)
3978 * This function changes the 'effective' priority of a task. It does
3979 * not touch ->normal_prio like __setscheduler().
3981 * Used by the rt_mutex code to implement priority inheritance logic.
3983 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3985 struct prio_array
*array
;
3986 unsigned long flags
;
3990 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3992 rq
= task_rq_lock(p
, &flags
);
3997 dequeue_task(p
, array
);
4002 * If changing to an RT priority then queue it
4003 * in the active array!
4007 enqueue_task(p
, array
);
4009 * Reschedule if we are currently running on this runqueue and
4010 * our priority decreased, or if we are not currently running on
4011 * this runqueue and our priority is higher than the current's
4013 if (task_running(rq
, p
)) {
4014 if (p
->prio
> oldprio
)
4015 resched_task(rq
->curr
);
4016 } else if (TASK_PREEMPTS_CURR(p
, rq
))
4017 resched_task(rq
->curr
);
4019 task_rq_unlock(rq
, &flags
);
4024 void set_user_nice(struct task_struct
*p
, long nice
)
4026 struct prio_array
*array
;
4027 int old_prio
, delta
;
4028 unsigned long flags
;
4031 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4034 * We have to be careful, if called from sys_setpriority(),
4035 * the task might be in the middle of scheduling on another CPU.
4037 rq
= task_rq_lock(p
, &flags
);
4039 * The RT priorities are set via sched_setscheduler(), but we still
4040 * allow the 'normal' nice value to be set - but as expected
4041 * it wont have any effect on scheduling until the task is
4042 * not SCHED_NORMAL/SCHED_BATCH:
4044 if (has_rt_policy(p
)) {
4045 p
->static_prio
= NICE_TO_PRIO(nice
);
4050 dequeue_task(p
, array
);
4051 dec_raw_weighted_load(rq
, p
);
4054 p
->static_prio
= NICE_TO_PRIO(nice
);
4057 p
->prio
= effective_prio(p
);
4058 delta
= p
->prio
- old_prio
;
4061 enqueue_task(p
, array
);
4062 inc_raw_weighted_load(rq
, p
);
4064 * If the task increased its priority or is running and
4065 * lowered its priority, then reschedule its CPU:
4067 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4068 resched_task(rq
->curr
);
4071 task_rq_unlock(rq
, &flags
);
4073 EXPORT_SYMBOL(set_user_nice
);
4076 * can_nice - check if a task can reduce its nice value
4080 int can_nice(const struct task_struct
*p
, const int nice
)
4082 /* convert nice value [19,-20] to rlimit style value [1,40] */
4083 int nice_rlim
= 20 - nice
;
4085 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4086 capable(CAP_SYS_NICE
));
4089 #ifdef __ARCH_WANT_SYS_NICE
4092 * sys_nice - change the priority of the current process.
4093 * @increment: priority increment
4095 * sys_setpriority is a more generic, but much slower function that
4096 * does similar things.
4098 asmlinkage
long sys_nice(int increment
)
4103 * Setpriority might change our priority at the same moment.
4104 * We don't have to worry. Conceptually one call occurs first
4105 * and we have a single winner.
4107 if (increment
< -40)
4112 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4118 if (increment
< 0 && !can_nice(current
, nice
))
4121 retval
= security_task_setnice(current
, nice
);
4125 set_user_nice(current
, nice
);
4132 * task_prio - return the priority value of a given task.
4133 * @p: the task in question.
4135 * This is the priority value as seen by users in /proc.
4136 * RT tasks are offset by -200. Normal tasks are centered
4137 * around 0, value goes from -16 to +15.
4139 int task_prio(const struct task_struct
*p
)
4141 return p
->prio
- MAX_RT_PRIO
;
4145 * task_nice - return the nice value of a given task.
4146 * @p: the task in question.
4148 int task_nice(const struct task_struct
*p
)
4150 return TASK_NICE(p
);
4152 EXPORT_SYMBOL_GPL(task_nice
);
4155 * idle_cpu - is a given cpu idle currently?
4156 * @cpu: the processor in question.
4158 int idle_cpu(int cpu
)
4160 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4164 * idle_task - return the idle task for a given cpu.
4165 * @cpu: the processor in question.
4167 struct task_struct
*idle_task(int cpu
)
4169 return cpu_rq(cpu
)->idle
;
4173 * find_process_by_pid - find a process with a matching PID value.
4174 * @pid: the pid in question.
4176 static inline struct task_struct
*find_process_by_pid(pid_t pid
)
4178 return pid
? find_task_by_pid(pid
) : current
;
4181 /* Actually do priority change: must hold rq lock. */
4182 static void __setscheduler(struct task_struct
*p
, int policy
, int prio
)
4187 p
->rt_priority
= prio
;
4188 p
->normal_prio
= normal_prio(p
);
4189 /* we are holding p->pi_lock already */
4190 p
->prio
= rt_mutex_getprio(p
);
4192 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
4194 if (policy
== SCHED_BATCH
)
4200 * sched_setscheduler - change the scheduling policy and/or RT priority of
4202 * @p: the task in question.
4203 * @policy: new policy.
4204 * @param: structure containing the new RT priority.
4206 * NOTE: the task may be already dead
4208 int sched_setscheduler(struct task_struct
*p
, int policy
,
4209 struct sched_param
*param
)
4211 int retval
, oldprio
, oldpolicy
= -1;
4212 struct prio_array
*array
;
4213 unsigned long flags
;
4216 /* may grab non-irq protected spin_locks */
4217 BUG_ON(in_interrupt());
4219 /* double check policy once rq lock held */
4221 policy
= oldpolicy
= p
->policy
;
4222 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4223 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
)
4226 * Valid priorities for SCHED_FIFO and SCHED_RR are
4227 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
4230 if (param
->sched_priority
< 0 ||
4231 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4232 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4234 if (is_rt_policy(policy
) != (param
->sched_priority
!= 0))
4238 * Allow unprivileged RT tasks to decrease priority:
4240 if (!capable(CAP_SYS_NICE
)) {
4241 if (is_rt_policy(policy
)) {
4242 unsigned long rlim_rtprio
;
4243 unsigned long flags
;
4245 if (!lock_task_sighand(p
, &flags
))
4247 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4248 unlock_task_sighand(p
, &flags
);
4250 /* can't set/change the rt policy */
4251 if (policy
!= p
->policy
&& !rlim_rtprio
)
4254 /* can't increase priority */
4255 if (param
->sched_priority
> p
->rt_priority
&&
4256 param
->sched_priority
> rlim_rtprio
)
4260 /* can't change other user's priorities */
4261 if ((current
->euid
!= p
->euid
) &&
4262 (current
->euid
!= p
->uid
))
4266 retval
= security_task_setscheduler(p
, policy
, param
);
4270 * make sure no PI-waiters arrive (or leave) while we are
4271 * changing the priority of the task:
4273 spin_lock_irqsave(&p
->pi_lock
, flags
);
4275 * To be able to change p->policy safely, the apropriate
4276 * runqueue lock must be held.
4278 rq
= __task_rq_lock(p
);
4279 /* recheck policy now with rq lock held */
4280 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4281 policy
= oldpolicy
= -1;
4282 __task_rq_unlock(rq
);
4283 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4288 deactivate_task(p
, rq
);
4290 __setscheduler(p
, policy
, param
->sched_priority
);
4292 __activate_task(p
, rq
);
4294 * Reschedule if we are currently running on this runqueue and
4295 * our priority decreased, or if we are not currently running on
4296 * this runqueue and our priority is higher than the current's
4298 if (task_running(rq
, p
)) {
4299 if (p
->prio
> oldprio
)
4300 resched_task(rq
->curr
);
4301 } else if (TASK_PREEMPTS_CURR(p
, rq
))
4302 resched_task(rq
->curr
);
4304 __task_rq_unlock(rq
);
4305 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4307 rt_mutex_adjust_pi(p
);
4311 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4314 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4316 struct sched_param lparam
;
4317 struct task_struct
*p
;
4320 if (!param
|| pid
< 0)
4322 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4327 p
= find_process_by_pid(pid
);
4329 retval
= sched_setscheduler(p
, policy
, &lparam
);
4336 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4337 * @pid: the pid in question.
4338 * @policy: new policy.
4339 * @param: structure containing the new RT priority.
4341 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
4342 struct sched_param __user
*param
)
4344 /* negative values for policy are not valid */
4348 return do_sched_setscheduler(pid
, policy
, param
);
4352 * sys_sched_setparam - set/change the RT priority of a thread
4353 * @pid: the pid in question.
4354 * @param: structure containing the new RT priority.
4356 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4358 return do_sched_setscheduler(pid
, -1, param
);
4362 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4363 * @pid: the pid in question.
4365 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4367 struct task_struct
*p
;
4368 int retval
= -EINVAL
;
4374 read_lock(&tasklist_lock
);
4375 p
= find_process_by_pid(pid
);
4377 retval
= security_task_getscheduler(p
);
4381 read_unlock(&tasklist_lock
);
4388 * sys_sched_getscheduler - get the RT priority of a thread
4389 * @pid: the pid in question.
4390 * @param: structure containing the RT priority.
4392 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4394 struct sched_param lp
;
4395 struct task_struct
*p
;
4396 int retval
= -EINVAL
;
4398 if (!param
|| pid
< 0)
4401 read_lock(&tasklist_lock
);
4402 p
= find_process_by_pid(pid
);
4407 retval
= security_task_getscheduler(p
);
4411 lp
.sched_priority
= p
->rt_priority
;
4412 read_unlock(&tasklist_lock
);
4415 * This one might sleep, we cannot do it with a spinlock held ...
4417 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4423 read_unlock(&tasklist_lock
);
4427 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4429 cpumask_t cpus_allowed
;
4430 struct task_struct
*p
;
4434 read_lock(&tasklist_lock
);
4436 p
= find_process_by_pid(pid
);
4438 read_unlock(&tasklist_lock
);
4439 unlock_cpu_hotplug();
4444 * It is not safe to call set_cpus_allowed with the
4445 * tasklist_lock held. We will bump the task_struct's
4446 * usage count and then drop tasklist_lock.
4449 read_unlock(&tasklist_lock
);
4452 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4453 !capable(CAP_SYS_NICE
))
4456 retval
= security_task_setscheduler(p
, 0, NULL
);
4460 cpus_allowed
= cpuset_cpus_allowed(p
);
4461 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4462 retval
= set_cpus_allowed(p
, new_mask
);
4466 unlock_cpu_hotplug();
4470 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4471 cpumask_t
*new_mask
)
4473 if (len
< sizeof(cpumask_t
)) {
4474 memset(new_mask
, 0, sizeof(cpumask_t
));
4475 } else if (len
> sizeof(cpumask_t
)) {
4476 len
= sizeof(cpumask_t
);
4478 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4482 * sys_sched_setaffinity - set the cpu affinity of a process
4483 * @pid: pid of the process
4484 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4485 * @user_mask_ptr: user-space pointer to the new cpu mask
4487 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4488 unsigned long __user
*user_mask_ptr
)
4493 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4497 return sched_setaffinity(pid
, new_mask
);
4501 * Represents all cpu's present in the system
4502 * In systems capable of hotplug, this map could dynamically grow
4503 * as new cpu's are detected in the system via any platform specific
4504 * method, such as ACPI for e.g.
4507 cpumask_t cpu_present_map __read_mostly
;
4508 EXPORT_SYMBOL(cpu_present_map
);
4511 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4512 EXPORT_SYMBOL(cpu_online_map
);
4514 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4515 EXPORT_SYMBOL(cpu_possible_map
);
4518 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4520 struct task_struct
*p
;
4524 read_lock(&tasklist_lock
);
4527 p
= find_process_by_pid(pid
);
4531 retval
= security_task_getscheduler(p
);
4535 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4538 read_unlock(&tasklist_lock
);
4539 unlock_cpu_hotplug();
4547 * sys_sched_getaffinity - get the cpu affinity of a process
4548 * @pid: pid of the process
4549 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4550 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4552 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4553 unsigned long __user
*user_mask_ptr
)
4558 if (len
< sizeof(cpumask_t
))
4561 ret
= sched_getaffinity(pid
, &mask
);
4565 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4568 return sizeof(cpumask_t
);
4572 * sys_sched_yield - yield the current processor to other threads.
4574 * this function yields the current CPU by moving the calling thread
4575 * to the expired array. If there are no other threads running on this
4576 * CPU then this function will return.
4578 asmlinkage
long sys_sched_yield(void)
4580 struct rq
*rq
= this_rq_lock();
4581 struct prio_array
*array
= current
->array
, *target
= rq
->expired
;
4583 schedstat_inc(rq
, yld_cnt
);
4585 * We implement yielding by moving the task into the expired
4588 * (special rule: RT tasks will just roundrobin in the active
4591 if (rt_task(current
))
4592 target
= rq
->active
;
4594 if (array
->nr_active
== 1) {
4595 schedstat_inc(rq
, yld_act_empty
);
4596 if (!rq
->expired
->nr_active
)
4597 schedstat_inc(rq
, yld_both_empty
);
4598 } else if (!rq
->expired
->nr_active
)
4599 schedstat_inc(rq
, yld_exp_empty
);
4601 if (array
!= target
) {
4602 dequeue_task(current
, array
);
4603 enqueue_task(current
, target
);
4606 * requeue_task is cheaper so perform that if possible.
4608 requeue_task(current
, array
);
4611 * Since we are going to call schedule() anyway, there's
4612 * no need to preempt or enable interrupts:
4614 __release(rq
->lock
);
4615 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4616 _raw_spin_unlock(&rq
->lock
);
4617 preempt_enable_no_resched();
4624 static void __cond_resched(void)
4626 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4627 __might_sleep(__FILE__
, __LINE__
);
4630 * The BKS might be reacquired before we have dropped
4631 * PREEMPT_ACTIVE, which could trigger a second
4632 * cond_resched() call.
4635 add_preempt_count(PREEMPT_ACTIVE
);
4637 sub_preempt_count(PREEMPT_ACTIVE
);
4638 } while (need_resched());
4641 int __sched
cond_resched(void)
4643 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4644 system_state
== SYSTEM_RUNNING
) {
4650 EXPORT_SYMBOL(cond_resched
);
4653 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4654 * call schedule, and on return reacquire the lock.
4656 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4657 * operations here to prevent schedule() from being called twice (once via
4658 * spin_unlock(), once by hand).
4660 int cond_resched_lock(spinlock_t
*lock
)
4664 if (need_lockbreak(lock
)) {
4670 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4671 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4672 _raw_spin_unlock(lock
);
4673 preempt_enable_no_resched();
4680 EXPORT_SYMBOL(cond_resched_lock
);
4682 int __sched
cond_resched_softirq(void)
4684 BUG_ON(!in_softirq());
4686 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4687 raw_local_irq_disable();
4689 raw_local_irq_enable();
4696 EXPORT_SYMBOL(cond_resched_softirq
);
4699 * yield - yield the current processor to other threads.
4701 * this is a shortcut for kernel-space yielding - it marks the
4702 * thread runnable and calls sys_sched_yield().
4704 void __sched
yield(void)
4706 set_current_state(TASK_RUNNING
);
4709 EXPORT_SYMBOL(yield
);
4712 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4713 * that process accounting knows that this is a task in IO wait state.
4715 * But don't do that if it is a deliberate, throttling IO wait (this task
4716 * has set its backing_dev_info: the queue against which it should throttle)
4718 void __sched
io_schedule(void)
4720 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4722 delayacct_blkio_start();
4723 atomic_inc(&rq
->nr_iowait
);
4725 atomic_dec(&rq
->nr_iowait
);
4726 delayacct_blkio_end();
4728 EXPORT_SYMBOL(io_schedule
);
4730 long __sched
io_schedule_timeout(long timeout
)
4732 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4735 delayacct_blkio_start();
4736 atomic_inc(&rq
->nr_iowait
);
4737 ret
= schedule_timeout(timeout
);
4738 atomic_dec(&rq
->nr_iowait
);
4739 delayacct_blkio_end();
4744 * sys_sched_get_priority_max - return maximum RT priority.
4745 * @policy: scheduling class.
4747 * this syscall returns the maximum rt_priority that can be used
4748 * by a given scheduling class.
4750 asmlinkage
long sys_sched_get_priority_max(int policy
)
4757 ret
= MAX_USER_RT_PRIO
-1;
4768 * sys_sched_get_priority_min - return minimum RT priority.
4769 * @policy: scheduling class.
4771 * this syscall returns the minimum rt_priority that can be used
4772 * by a given scheduling class.
4774 asmlinkage
long sys_sched_get_priority_min(int policy
)
4791 * sys_sched_rr_get_interval - return the default timeslice of a process.
4792 * @pid: pid of the process.
4793 * @interval: userspace pointer to the timeslice value.
4795 * this syscall writes the default timeslice value of a given process
4796 * into the user-space timespec buffer. A value of '0' means infinity.
4799 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4801 struct task_struct
*p
;
4802 int retval
= -EINVAL
;
4809 read_lock(&tasklist_lock
);
4810 p
= find_process_by_pid(pid
);
4814 retval
= security_task_getscheduler(p
);
4818 jiffies_to_timespec(p
->policy
== SCHED_FIFO
?
4819 0 : task_timeslice(p
), &t
);
4820 read_unlock(&tasklist_lock
);
4821 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4825 read_unlock(&tasklist_lock
);
4829 static inline struct task_struct
*eldest_child(struct task_struct
*p
)
4831 if (list_empty(&p
->children
))
4833 return list_entry(p
->children
.next
,struct task_struct
,sibling
);
4836 static inline struct task_struct
*older_sibling(struct task_struct
*p
)
4838 if (p
->sibling
.prev
==&p
->parent
->children
)
4840 return list_entry(p
->sibling
.prev
,struct task_struct
,sibling
);
4843 static inline struct task_struct
*younger_sibling(struct task_struct
*p
)
4845 if (p
->sibling
.next
==&p
->parent
->children
)
4847 return list_entry(p
->sibling
.next
,struct task_struct
,sibling
);
4850 static const char stat_nam
[] = "RSDTtZX";
4852 static void show_task(struct task_struct
*p
)
4854 struct task_struct
*relative
;
4855 unsigned long free
= 0;
4858 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4859 printk("%-13.13s %c", p
->comm
,
4860 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4861 #if (BITS_PER_LONG == 32)
4862 if (state
== TASK_RUNNING
)
4863 printk(" running ");
4865 printk(" %08lX ", thread_saved_pc(p
));
4867 if (state
== TASK_RUNNING
)
4868 printk(" running task ");
4870 printk(" %016lx ", thread_saved_pc(p
));
4872 #ifdef CONFIG_DEBUG_STACK_USAGE
4874 unsigned long *n
= end_of_stack(p
);
4877 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4880 printk("%5lu %5d %6d ", free
, p
->pid
, p
->parent
->pid
);
4881 if ((relative
= eldest_child(p
)))
4882 printk("%5d ", relative
->pid
);
4885 if ((relative
= younger_sibling(p
)))
4886 printk("%7d", relative
->pid
);
4889 if ((relative
= older_sibling(p
)))
4890 printk(" %5d", relative
->pid
);
4894 printk(" (L-TLB)\n");
4896 printk(" (NOTLB)\n");
4898 if (state
!= TASK_RUNNING
)
4899 show_stack(p
, NULL
);
4902 void show_state_filter(unsigned long state_filter
)
4904 struct task_struct
*g
, *p
;
4906 #if (BITS_PER_LONG == 32)
4909 printk(" task PC stack pid father child younger older\n");
4913 printk(" task PC stack pid father child younger older\n");
4915 read_lock(&tasklist_lock
);
4916 do_each_thread(g
, p
) {
4918 * reset the NMI-timeout, listing all files on a slow
4919 * console might take alot of time:
4921 touch_nmi_watchdog();
4922 if (p
->state
& state_filter
)
4924 } while_each_thread(g
, p
);
4926 read_unlock(&tasklist_lock
);
4928 * Only show locks if all tasks are dumped:
4930 if (state_filter
== -1)
4931 debug_show_all_locks();
4935 * init_idle - set up an idle thread for a given CPU
4936 * @idle: task in question
4937 * @cpu: cpu the idle task belongs to
4939 * NOTE: this function does not set the idle thread's NEED_RESCHED
4940 * flag, to make booting more robust.
4942 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4944 struct rq
*rq
= cpu_rq(cpu
);
4945 unsigned long flags
;
4947 idle
->timestamp
= sched_clock();
4948 idle
->sleep_avg
= 0;
4950 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
4951 idle
->state
= TASK_RUNNING
;
4952 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4953 set_task_cpu(idle
, cpu
);
4955 spin_lock_irqsave(&rq
->lock
, flags
);
4956 rq
->curr
= rq
->idle
= idle
;
4957 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4960 spin_unlock_irqrestore(&rq
->lock
, flags
);
4962 /* Set the preempt count _outside_ the spinlocks! */
4963 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4964 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4966 task_thread_info(idle
)->preempt_count
= 0;
4971 * In a system that switches off the HZ timer nohz_cpu_mask
4972 * indicates which cpus entered this state. This is used
4973 * in the rcu update to wait only for active cpus. For system
4974 * which do not switch off the HZ timer nohz_cpu_mask should
4975 * always be CPU_MASK_NONE.
4977 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4981 * This is how migration works:
4983 * 1) we queue a struct migration_req structure in the source CPU's
4984 * runqueue and wake up that CPU's migration thread.
4985 * 2) we down() the locked semaphore => thread blocks.
4986 * 3) migration thread wakes up (implicitly it forces the migrated
4987 * thread off the CPU)
4988 * 4) it gets the migration request and checks whether the migrated
4989 * task is still in the wrong runqueue.
4990 * 5) if it's in the wrong runqueue then the migration thread removes
4991 * it and puts it into the right queue.
4992 * 6) migration thread up()s the semaphore.
4993 * 7) we wake up and the migration is done.
4997 * Change a given task's CPU affinity. Migrate the thread to a
4998 * proper CPU and schedule it away if the CPU it's executing on
4999 * is removed from the allowed bitmask.
5001 * NOTE: the caller must have a valid reference to the task, the
5002 * task must not exit() & deallocate itself prematurely. The
5003 * call is not atomic; no spinlocks may be held.
5005 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
5007 struct migration_req req
;
5008 unsigned long flags
;
5012 rq
= task_rq_lock(p
, &flags
);
5013 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
5018 p
->cpus_allowed
= new_mask
;
5019 /* Can the task run on the task's current CPU? If so, we're done */
5020 if (cpu_isset(task_cpu(p
), new_mask
))
5023 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
5024 /* Need help from migration thread: drop lock and wait. */
5025 task_rq_unlock(rq
, &flags
);
5026 wake_up_process(rq
->migration_thread
);
5027 wait_for_completion(&req
.done
);
5028 tlb_migrate_finish(p
->mm
);
5032 task_rq_unlock(rq
, &flags
);
5036 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
5039 * Move (not current) task off this cpu, onto dest cpu. We're doing
5040 * this because either it can't run here any more (set_cpus_allowed()
5041 * away from this CPU, or CPU going down), or because we're
5042 * attempting to rebalance this task on exec (sched_exec).
5044 * So we race with normal scheduler movements, but that's OK, as long
5045 * as the task is no longer on this CPU.
5047 * Returns non-zero if task was successfully migrated.
5049 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5051 struct rq
*rq_dest
, *rq_src
;
5054 if (unlikely(cpu_is_offline(dest_cpu
)))
5057 rq_src
= cpu_rq(src_cpu
);
5058 rq_dest
= cpu_rq(dest_cpu
);
5060 double_rq_lock(rq_src
, rq_dest
);
5061 /* Already moved. */
5062 if (task_cpu(p
) != src_cpu
)
5064 /* Affinity changed (again). */
5065 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
5068 set_task_cpu(p
, dest_cpu
);
5071 * Sync timestamp with rq_dest's before activating.
5072 * The same thing could be achieved by doing this step
5073 * afterwards, and pretending it was a local activate.
5074 * This way is cleaner and logically correct.
5076 p
->timestamp
= p
->timestamp
- rq_src
->most_recent_timestamp
5077 + rq_dest
->most_recent_timestamp
;
5078 deactivate_task(p
, rq_src
);
5079 __activate_task(p
, rq_dest
);
5080 if (TASK_PREEMPTS_CURR(p
, rq_dest
))
5081 resched_task(rq_dest
->curr
);
5085 double_rq_unlock(rq_src
, rq_dest
);
5090 * migration_thread - this is a highprio system thread that performs
5091 * thread migration by bumping thread off CPU then 'pushing' onto
5094 static int migration_thread(void *data
)
5096 int cpu
= (long)data
;
5100 BUG_ON(rq
->migration_thread
!= current
);
5102 set_current_state(TASK_INTERRUPTIBLE
);
5103 while (!kthread_should_stop()) {
5104 struct migration_req
*req
;
5105 struct list_head
*head
;
5109 spin_lock_irq(&rq
->lock
);
5111 if (cpu_is_offline(cpu
)) {
5112 spin_unlock_irq(&rq
->lock
);
5116 if (rq
->active_balance
) {
5117 active_load_balance(rq
, cpu
);
5118 rq
->active_balance
= 0;
5121 head
= &rq
->migration_queue
;
5123 if (list_empty(head
)) {
5124 spin_unlock_irq(&rq
->lock
);
5126 set_current_state(TASK_INTERRUPTIBLE
);
5129 req
= list_entry(head
->next
, struct migration_req
, list
);
5130 list_del_init(head
->next
);
5132 spin_unlock(&rq
->lock
);
5133 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5136 complete(&req
->done
);
5138 __set_current_state(TASK_RUNNING
);
5142 /* Wait for kthread_stop */
5143 set_current_state(TASK_INTERRUPTIBLE
);
5144 while (!kthread_should_stop()) {
5146 set_current_state(TASK_INTERRUPTIBLE
);
5148 __set_current_state(TASK_RUNNING
);
5152 #ifdef CONFIG_HOTPLUG_CPU
5154 * Figure out where task on dead CPU should go, use force if neccessary.
5155 * NOTE: interrupts should be disabled by the caller
5157 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5159 unsigned long flags
;
5166 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5167 cpus_and(mask
, mask
, p
->cpus_allowed
);
5168 dest_cpu
= any_online_cpu(mask
);
5170 /* On any allowed CPU? */
5171 if (dest_cpu
== NR_CPUS
)
5172 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5174 /* No more Mr. Nice Guy. */
5175 if (dest_cpu
== NR_CPUS
) {
5176 rq
= task_rq_lock(p
, &flags
);
5177 cpus_setall(p
->cpus_allowed
);
5178 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5179 task_rq_unlock(rq
, &flags
);
5182 * Don't tell them about moving exiting tasks or
5183 * kernel threads (both mm NULL), since they never
5186 if (p
->mm
&& printk_ratelimit())
5187 printk(KERN_INFO
"process %d (%s) no "
5188 "longer affine to cpu%d\n",
5189 p
->pid
, p
->comm
, dead_cpu
);
5191 if (!__migrate_task(p
, dead_cpu
, dest_cpu
))
5196 * While a dead CPU has no uninterruptible tasks queued at this point,
5197 * it might still have a nonzero ->nr_uninterruptible counter, because
5198 * for performance reasons the counter is not stricly tracking tasks to
5199 * their home CPUs. So we just add the counter to another CPU's counter,
5200 * to keep the global sum constant after CPU-down:
5202 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5204 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5205 unsigned long flags
;
5207 local_irq_save(flags
);
5208 double_rq_lock(rq_src
, rq_dest
);
5209 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5210 rq_src
->nr_uninterruptible
= 0;
5211 double_rq_unlock(rq_src
, rq_dest
);
5212 local_irq_restore(flags
);
5215 /* Run through task list and migrate tasks from the dead cpu. */
5216 static void migrate_live_tasks(int src_cpu
)
5218 struct task_struct
*p
, *t
;
5220 write_lock_irq(&tasklist_lock
);
5222 do_each_thread(t
, p
) {
5226 if (task_cpu(p
) == src_cpu
)
5227 move_task_off_dead_cpu(src_cpu
, p
);
5228 } while_each_thread(t
, p
);
5230 write_unlock_irq(&tasklist_lock
);
5233 /* Schedules idle task to be the next runnable task on current CPU.
5234 * It does so by boosting its priority to highest possible and adding it to
5235 * the _front_ of the runqueue. Used by CPU offline code.
5237 void sched_idle_next(void)
5239 int this_cpu
= smp_processor_id();
5240 struct rq
*rq
= cpu_rq(this_cpu
);
5241 struct task_struct
*p
= rq
->idle
;
5242 unsigned long flags
;
5244 /* cpu has to be offline */
5245 BUG_ON(cpu_online(this_cpu
));
5248 * Strictly not necessary since rest of the CPUs are stopped by now
5249 * and interrupts disabled on the current cpu.
5251 spin_lock_irqsave(&rq
->lock
, flags
);
5253 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5255 /* Add idle task to the _front_ of its priority queue: */
5256 __activate_idle_task(p
, rq
);
5258 spin_unlock_irqrestore(&rq
->lock
, flags
);
5262 * Ensures that the idle task is using init_mm right before its cpu goes
5265 void idle_task_exit(void)
5267 struct mm_struct
*mm
= current
->active_mm
;
5269 BUG_ON(cpu_online(smp_processor_id()));
5272 switch_mm(mm
, &init_mm
, current
);
5276 /* called under rq->lock with disabled interrupts */
5277 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5279 struct rq
*rq
= cpu_rq(dead_cpu
);
5281 /* Must be exiting, otherwise would be on tasklist. */
5282 BUG_ON(p
->exit_state
!= EXIT_ZOMBIE
&& p
->exit_state
!= EXIT_DEAD
);
5284 /* Cannot have done final schedule yet: would have vanished. */
5285 BUG_ON(p
->state
== TASK_DEAD
);
5290 * Drop lock around migration; if someone else moves it,
5291 * that's OK. No task can be added to this CPU, so iteration is
5293 * NOTE: interrupts should be left disabled --dev@
5295 spin_unlock(&rq
->lock
);
5296 move_task_off_dead_cpu(dead_cpu
, p
);
5297 spin_lock(&rq
->lock
);
5302 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5303 static void migrate_dead_tasks(unsigned int dead_cpu
)
5305 struct rq
*rq
= cpu_rq(dead_cpu
);
5306 unsigned int arr
, i
;
5308 for (arr
= 0; arr
< 2; arr
++) {
5309 for (i
= 0; i
< MAX_PRIO
; i
++) {
5310 struct list_head
*list
= &rq
->arrays
[arr
].queue
[i
];
5312 while (!list_empty(list
))
5313 migrate_dead(dead_cpu
, list_entry(list
->next
,
5314 struct task_struct
, run_list
));
5318 #endif /* CONFIG_HOTPLUG_CPU */
5321 * migration_call - callback that gets triggered when a CPU is added.
5322 * Here we can start up the necessary migration thread for the new CPU.
5324 static int __cpuinit
5325 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5327 struct task_struct
*p
;
5328 int cpu
= (long)hcpu
;
5329 unsigned long flags
;
5333 case CPU_UP_PREPARE
:
5334 p
= kthread_create(migration_thread
, hcpu
, "migration/%d",cpu
);
5337 p
->flags
|= PF_NOFREEZE
;
5338 kthread_bind(p
, cpu
);
5339 /* Must be high prio: stop_machine expects to yield to it. */
5340 rq
= task_rq_lock(p
, &flags
);
5341 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5342 task_rq_unlock(rq
, &flags
);
5343 cpu_rq(cpu
)->migration_thread
= p
;
5347 /* Strictly unneccessary, as first user will wake it. */
5348 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5351 #ifdef CONFIG_HOTPLUG_CPU
5352 case CPU_UP_CANCELED
:
5353 if (!cpu_rq(cpu
)->migration_thread
)
5355 /* Unbind it from offline cpu so it can run. Fall thru. */
5356 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5357 any_online_cpu(cpu_online_map
));
5358 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5359 cpu_rq(cpu
)->migration_thread
= NULL
;
5363 migrate_live_tasks(cpu
);
5365 kthread_stop(rq
->migration_thread
);
5366 rq
->migration_thread
= NULL
;
5367 /* Idle task back to normal (off runqueue, low prio) */
5368 rq
= task_rq_lock(rq
->idle
, &flags
);
5369 deactivate_task(rq
->idle
, rq
);
5370 rq
->idle
->static_prio
= MAX_PRIO
;
5371 __setscheduler(rq
->idle
, SCHED_NORMAL
, 0);
5372 migrate_dead_tasks(cpu
);
5373 task_rq_unlock(rq
, &flags
);
5374 migrate_nr_uninterruptible(rq
);
5375 BUG_ON(rq
->nr_running
!= 0);
5377 /* No need to migrate the tasks: it was best-effort if
5378 * they didn't do lock_cpu_hotplug(). Just wake up
5379 * the requestors. */
5380 spin_lock_irq(&rq
->lock
);
5381 while (!list_empty(&rq
->migration_queue
)) {
5382 struct migration_req
*req
;
5384 req
= list_entry(rq
->migration_queue
.next
,
5385 struct migration_req
, list
);
5386 list_del_init(&req
->list
);
5387 complete(&req
->done
);
5389 spin_unlock_irq(&rq
->lock
);
5396 /* Register at highest priority so that task migration (migrate_all_tasks)
5397 * happens before everything else.
5399 static struct notifier_block __cpuinitdata migration_notifier
= {
5400 .notifier_call
= migration_call
,
5404 int __init
migration_init(void)
5406 void *cpu
= (void *)(long)smp_processor_id();
5409 /* Start one for the boot CPU: */
5410 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5411 BUG_ON(err
== NOTIFY_BAD
);
5412 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5413 register_cpu_notifier(&migration_notifier
);
5420 #undef SCHED_DOMAIN_DEBUG
5421 #ifdef SCHED_DOMAIN_DEBUG
5422 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5427 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5431 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5436 struct sched_group
*group
= sd
->groups
;
5437 cpumask_t groupmask
;
5439 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5440 cpus_clear(groupmask
);
5443 for (i
= 0; i
< level
+ 1; i
++)
5445 printk("domain %d: ", level
);
5447 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5448 printk("does not load-balance\n");
5450 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5455 printk("span %s\n", str
);
5457 if (!cpu_isset(cpu
, sd
->span
))
5458 printk(KERN_ERR
"ERROR: domain->span does not contain "
5460 if (!cpu_isset(cpu
, group
->cpumask
))
5461 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5465 for (i
= 0; i
< level
+ 2; i
++)
5471 printk(KERN_ERR
"ERROR: group is NULL\n");
5475 if (!group
->cpu_power
) {
5477 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5481 if (!cpus_weight(group
->cpumask
)) {
5483 printk(KERN_ERR
"ERROR: empty group\n");
5486 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5488 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5491 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5493 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5496 group
= group
->next
;
5497 } while (group
!= sd
->groups
);
5500 if (!cpus_equal(sd
->span
, groupmask
))
5501 printk(KERN_ERR
"ERROR: groups don't span "
5509 if (!cpus_subset(groupmask
, sd
->span
))
5510 printk(KERN_ERR
"ERROR: parent span is not a superset "
5511 "of domain->span\n");
5516 # define sched_domain_debug(sd, cpu) do { } while (0)
5519 static int sd_degenerate(struct sched_domain
*sd
)
5521 if (cpus_weight(sd
->span
) == 1)
5524 /* Following flags need at least 2 groups */
5525 if (sd
->flags
& (SD_LOAD_BALANCE
|
5526 SD_BALANCE_NEWIDLE
|
5530 SD_SHARE_PKG_RESOURCES
)) {
5531 if (sd
->groups
!= sd
->groups
->next
)
5535 /* Following flags don't use groups */
5536 if (sd
->flags
& (SD_WAKE_IDLE
|
5545 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5547 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5549 if (sd_degenerate(parent
))
5552 if (!cpus_equal(sd
->span
, parent
->span
))
5555 /* Does parent contain flags not in child? */
5556 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5557 if (cflags
& SD_WAKE_AFFINE
)
5558 pflags
&= ~SD_WAKE_BALANCE
;
5559 /* Flags needing groups don't count if only 1 group in parent */
5560 if (parent
->groups
== parent
->groups
->next
) {
5561 pflags
&= ~(SD_LOAD_BALANCE
|
5562 SD_BALANCE_NEWIDLE
|
5566 SD_SHARE_PKG_RESOURCES
);
5568 if (~cflags
& pflags
)
5575 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5576 * hold the hotplug lock.
5578 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5580 struct rq
*rq
= cpu_rq(cpu
);
5581 struct sched_domain
*tmp
;
5583 /* Remove the sched domains which do not contribute to scheduling. */
5584 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5585 struct sched_domain
*parent
= tmp
->parent
;
5588 if (sd_parent_degenerate(tmp
, parent
)) {
5589 tmp
->parent
= parent
->parent
;
5591 parent
->parent
->child
= tmp
;
5595 if (sd
&& sd_degenerate(sd
)) {
5601 sched_domain_debug(sd
, cpu
);
5603 rcu_assign_pointer(rq
->sd
, sd
);
5606 /* cpus with isolated domains */
5607 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
5609 /* Setup the mask of cpus configured for isolated domains */
5610 static int __init
isolated_cpu_setup(char *str
)
5612 int ints
[NR_CPUS
], i
;
5614 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5615 cpus_clear(cpu_isolated_map
);
5616 for (i
= 1; i
<= ints
[0]; i
++)
5617 if (ints
[i
] < NR_CPUS
)
5618 cpu_set(ints
[i
], cpu_isolated_map
);
5622 __setup ("isolcpus=", isolated_cpu_setup
);
5625 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5626 * to a function which identifies what group(along with sched group) a CPU
5627 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5628 * (due to the fact that we keep track of groups covered with a cpumask_t).
5630 * init_sched_build_groups will build a circular linked list of the groups
5631 * covered by the given span, and will set each group's ->cpumask correctly,
5632 * and ->cpu_power to 0.
5635 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
5636 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
5637 struct sched_group
**sg
))
5639 struct sched_group
*first
= NULL
, *last
= NULL
;
5640 cpumask_t covered
= CPU_MASK_NONE
;
5643 for_each_cpu_mask(i
, span
) {
5644 struct sched_group
*sg
;
5645 int group
= group_fn(i
, cpu_map
, &sg
);
5648 if (cpu_isset(i
, covered
))
5651 sg
->cpumask
= CPU_MASK_NONE
;
5654 for_each_cpu_mask(j
, span
) {
5655 if (group_fn(j
, cpu_map
, NULL
) != group
)
5658 cpu_set(j
, covered
);
5659 cpu_set(j
, sg
->cpumask
);
5670 #define SD_NODES_PER_DOMAIN 16
5673 * Self-tuning task migration cost measurement between source and target CPUs.
5675 * This is done by measuring the cost of manipulating buffers of varying
5676 * sizes. For a given buffer-size here are the steps that are taken:
5678 * 1) the source CPU reads+dirties a shared buffer
5679 * 2) the target CPU reads+dirties the same shared buffer
5681 * We measure how long they take, in the following 4 scenarios:
5683 * - source: CPU1, target: CPU2 | cost1
5684 * - source: CPU2, target: CPU1 | cost2
5685 * - source: CPU1, target: CPU1 | cost3
5686 * - source: CPU2, target: CPU2 | cost4
5688 * We then calculate the cost3+cost4-cost1-cost2 difference - this is
5689 * the cost of migration.
5691 * We then start off from a small buffer-size and iterate up to larger
5692 * buffer sizes, in 5% steps - measuring each buffer-size separately, and
5693 * doing a maximum search for the cost. (The maximum cost for a migration
5694 * normally occurs when the working set size is around the effective cache
5697 #define SEARCH_SCOPE 2
5698 #define MIN_CACHE_SIZE (64*1024U)
5699 #define DEFAULT_CACHE_SIZE (5*1024*1024U)
5700 #define ITERATIONS 1
5701 #define SIZE_THRESH 130
5702 #define COST_THRESH 130
5705 * The migration cost is a function of 'domain distance'. Domain
5706 * distance is the number of steps a CPU has to iterate down its
5707 * domain tree to share a domain with the other CPU. The farther
5708 * two CPUs are from each other, the larger the distance gets.
5710 * Note that we use the distance only to cache measurement results,
5711 * the distance value is not used numerically otherwise. When two
5712 * CPUs have the same distance it is assumed that the migration
5713 * cost is the same. (this is a simplification but quite practical)
5715 #define MAX_DOMAIN_DISTANCE 32
5717 static unsigned long long migration_cost
[MAX_DOMAIN_DISTANCE
] =
5718 { [ 0 ... MAX_DOMAIN_DISTANCE
-1 ] =
5720 * Architectures may override the migration cost and thus avoid
5721 * boot-time calibration. Unit is nanoseconds. Mostly useful for
5722 * virtualized hardware:
5724 #ifdef CONFIG_DEFAULT_MIGRATION_COST
5725 CONFIG_DEFAULT_MIGRATION_COST
5732 * Allow override of migration cost - in units of microseconds.
5733 * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost
5734 * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs:
5736 static int __init
migration_cost_setup(char *str
)
5738 int ints
[MAX_DOMAIN_DISTANCE
+1], i
;
5740 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5742 printk("#ints: %d\n", ints
[0]);
5743 for (i
= 1; i
<= ints
[0]; i
++) {
5744 migration_cost
[i
-1] = (unsigned long long)ints
[i
]*1000;
5745 printk("migration_cost[%d]: %Ld\n", i
-1, migration_cost
[i
-1]);
5750 __setup ("migration_cost=", migration_cost_setup
);
5753 * Global multiplier (divisor) for migration-cutoff values,
5754 * in percentiles. E.g. use a value of 150 to get 1.5 times
5755 * longer cache-hot cutoff times.
5757 * (We scale it from 100 to 128 to long long handling easier.)
5760 #define MIGRATION_FACTOR_SCALE 128
5762 static unsigned int migration_factor
= MIGRATION_FACTOR_SCALE
;
5764 static int __init
setup_migration_factor(char *str
)
5766 get_option(&str
, &migration_factor
);
5767 migration_factor
= migration_factor
* MIGRATION_FACTOR_SCALE
/ 100;
5771 __setup("migration_factor=", setup_migration_factor
);
5774 * Estimated distance of two CPUs, measured via the number of domains
5775 * we have to pass for the two CPUs to be in the same span:
5777 static unsigned long domain_distance(int cpu1
, int cpu2
)
5779 unsigned long distance
= 0;
5780 struct sched_domain
*sd
;
5782 for_each_domain(cpu1
, sd
) {
5783 WARN_ON(!cpu_isset(cpu1
, sd
->span
));
5784 if (cpu_isset(cpu2
, sd
->span
))
5788 if (distance
>= MAX_DOMAIN_DISTANCE
) {
5790 distance
= MAX_DOMAIN_DISTANCE
-1;
5796 static unsigned int migration_debug
;
5798 static int __init
setup_migration_debug(char *str
)
5800 get_option(&str
, &migration_debug
);
5804 __setup("migration_debug=", setup_migration_debug
);
5807 * Maximum cache-size that the scheduler should try to measure.
5808 * Architectures with larger caches should tune this up during
5809 * bootup. Gets used in the domain-setup code (i.e. during SMP
5812 unsigned int max_cache_size
;
5814 static int __init
setup_max_cache_size(char *str
)
5816 get_option(&str
, &max_cache_size
);
5820 __setup("max_cache_size=", setup_max_cache_size
);
5823 * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This
5824 * is the operation that is timed, so we try to generate unpredictable
5825 * cachemisses that still end up filling the L2 cache:
5827 static void touch_cache(void *__cache
, unsigned long __size
)
5829 unsigned long size
= __size
/ sizeof(long);
5830 unsigned long chunk1
= size
/ 3;
5831 unsigned long chunk2
= 2 * size
/ 3;
5832 unsigned long *cache
= __cache
;
5835 for (i
= 0; i
< size
/6; i
+= 8) {
5838 case 1: cache
[size
-1-i
]++;
5839 case 2: cache
[chunk1
-i
]++;
5840 case 3: cache
[chunk1
+i
]++;
5841 case 4: cache
[chunk2
-i
]++;
5842 case 5: cache
[chunk2
+i
]++;
5848 * Measure the cache-cost of one task migration. Returns in units of nsec.
5850 static unsigned long long
5851 measure_one(void *cache
, unsigned long size
, int source
, int target
)
5853 cpumask_t mask
, saved_mask
;
5854 unsigned long long t0
, t1
, t2
, t3
, cost
;
5856 saved_mask
= current
->cpus_allowed
;
5859 * Flush source caches to RAM and invalidate them:
5864 * Migrate to the source CPU:
5866 mask
= cpumask_of_cpu(source
);
5867 set_cpus_allowed(current
, mask
);
5868 WARN_ON(smp_processor_id() != source
);
5871 * Dirty the working set:
5874 touch_cache(cache
, size
);
5878 * Migrate to the target CPU, dirty the L2 cache and access
5879 * the shared buffer. (which represents the working set
5880 * of a migrated task.)
5882 mask
= cpumask_of_cpu(target
);
5883 set_cpus_allowed(current
, mask
);
5884 WARN_ON(smp_processor_id() != target
);
5887 touch_cache(cache
, size
);
5890 cost
= t1
-t0
+ t3
-t2
;
5892 if (migration_debug
>= 2)
5893 printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n",
5894 source
, target
, t1
-t0
, t1
-t0
, t3
-t2
, cost
);
5896 * Flush target caches to RAM and invalidate them:
5900 set_cpus_allowed(current
, saved_mask
);
5906 * Measure a series of task migrations and return the average
5907 * result. Since this code runs early during bootup the system
5908 * is 'undisturbed' and the average latency makes sense.
5910 * The algorithm in essence auto-detects the relevant cache-size,
5911 * so it will properly detect different cachesizes for different
5912 * cache-hierarchies, depending on how the CPUs are connected.
5914 * Architectures can prime the upper limit of the search range via
5915 * max_cache_size, otherwise the search range defaults to 20MB...64K.
5917 static unsigned long long
5918 measure_cost(int cpu1
, int cpu2
, void *cache
, unsigned int size
)
5920 unsigned long long cost1
, cost2
;
5924 * Measure the migration cost of 'size' bytes, over an
5925 * average of 10 runs:
5927 * (We perturb the cache size by a small (0..4k)
5928 * value to compensate size/alignment related artifacts.
5929 * We also subtract the cost of the operation done on
5935 * dry run, to make sure we start off cache-cold on cpu1,
5936 * and to get any vmalloc pagefaults in advance:
5938 measure_one(cache
, size
, cpu1
, cpu2
);
5939 for (i
= 0; i
< ITERATIONS
; i
++)
5940 cost1
+= measure_one(cache
, size
- i
* 1024, cpu1
, cpu2
);
5942 measure_one(cache
, size
, cpu2
, cpu1
);
5943 for (i
= 0; i
< ITERATIONS
; i
++)
5944 cost1
+= measure_one(cache
, size
- i
* 1024, cpu2
, cpu1
);
5947 * (We measure the non-migrating [cached] cost on both
5948 * cpu1 and cpu2, to handle CPUs with different speeds)
5952 measure_one(cache
, size
, cpu1
, cpu1
);
5953 for (i
= 0; i
< ITERATIONS
; i
++)
5954 cost2
+= measure_one(cache
, size
- i
* 1024, cpu1
, cpu1
);
5956 measure_one(cache
, size
, cpu2
, cpu2
);
5957 for (i
= 0; i
< ITERATIONS
; i
++)
5958 cost2
+= measure_one(cache
, size
- i
* 1024, cpu2
, cpu2
);
5961 * Get the per-iteration migration cost:
5963 do_div(cost1
, 2 * ITERATIONS
);
5964 do_div(cost2
, 2 * ITERATIONS
);
5966 return cost1
- cost2
;
5969 static unsigned long long measure_migration_cost(int cpu1
, int cpu2
)
5971 unsigned long long max_cost
= 0, fluct
= 0, avg_fluct
= 0;
5972 unsigned int max_size
, size
, size_found
= 0;
5973 long long cost
= 0, prev_cost
;
5977 * Search from max_cache_size*5 down to 64K - the real relevant
5978 * cachesize has to lie somewhere inbetween.
5980 if (max_cache_size
) {
5981 max_size
= max(max_cache_size
* SEARCH_SCOPE
, MIN_CACHE_SIZE
);
5982 size
= max(max_cache_size
/ SEARCH_SCOPE
, MIN_CACHE_SIZE
);
5985 * Since we have no estimation about the relevant
5988 max_size
= DEFAULT_CACHE_SIZE
* SEARCH_SCOPE
;
5989 size
= MIN_CACHE_SIZE
;
5992 if (!cpu_online(cpu1
) || !cpu_online(cpu2
)) {
5993 printk("cpu %d and %d not both online!\n", cpu1
, cpu2
);
5998 * Allocate the working set:
6000 cache
= vmalloc(max_size
);
6002 printk("could not vmalloc %d bytes for cache!\n", 2 * max_size
);
6003 return 1000000; /* return 1 msec on very small boxen */
6006 while (size
<= max_size
) {
6008 cost
= measure_cost(cpu1
, cpu2
, cache
, size
);
6014 if (max_cost
< cost
) {
6020 * Calculate average fluctuation, we use this to prevent
6021 * noise from triggering an early break out of the loop:
6023 fluct
= abs(cost
- prev_cost
);
6024 avg_fluct
= (avg_fluct
+ fluct
)/2;
6026 if (migration_debug
)
6027 printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): "
6030 (long)cost
/ 1000000,
6031 ((long)cost
/ 100000) % 10,
6032 (long)max_cost
/ 1000000,
6033 ((long)max_cost
/ 100000) % 10,
6034 domain_distance(cpu1
, cpu2
),
6038 * If we iterated at least 20% past the previous maximum,
6039 * and the cost has dropped by more than 20% already,
6040 * (taking fluctuations into account) then we assume to
6041 * have found the maximum and break out of the loop early:
6043 if (size_found
&& (size
*100 > size_found
*SIZE_THRESH
))
6044 if (cost
+avg_fluct
<= 0 ||
6045 max_cost
*100 > (cost
+avg_fluct
)*COST_THRESH
) {
6047 if (migration_debug
)
6048 printk("-> found max.\n");
6052 * Increase the cachesize in 10% steps:
6054 size
= size
* 10 / 9;
6057 if (migration_debug
)
6058 printk("[%d][%d] working set size found: %d, cost: %Ld\n",
6059 cpu1
, cpu2
, size_found
, max_cost
);
6064 * A task is considered 'cache cold' if at least 2 times
6065 * the worst-case cost of migration has passed.
6067 * (this limit is only listened to if the load-balancing
6068 * situation is 'nice' - if there is a large imbalance we
6069 * ignore it for the sake of CPU utilization and
6070 * processing fairness.)
6072 return 2 * max_cost
* migration_factor
/ MIGRATION_FACTOR_SCALE
;
6075 static void calibrate_migration_costs(const cpumask_t
*cpu_map
)
6077 int cpu1
= -1, cpu2
= -1, cpu
, orig_cpu
= raw_smp_processor_id();
6078 unsigned long j0
, j1
, distance
, max_distance
= 0;
6079 struct sched_domain
*sd
;
6084 * First pass - calculate the cacheflush times:
6086 for_each_cpu_mask(cpu1
, *cpu_map
) {
6087 for_each_cpu_mask(cpu2
, *cpu_map
) {
6090 distance
= domain_distance(cpu1
, cpu2
);
6091 max_distance
= max(max_distance
, distance
);
6093 * No result cached yet?
6095 if (migration_cost
[distance
] == -1LL)
6096 migration_cost
[distance
] =
6097 measure_migration_cost(cpu1
, cpu2
);
6101 * Second pass - update the sched domain hierarchy with
6102 * the new cache-hot-time estimations:
6104 for_each_cpu_mask(cpu
, *cpu_map
) {
6106 for_each_domain(cpu
, sd
) {
6107 sd
->cache_hot_time
= migration_cost
[distance
];
6114 if (migration_debug
)
6115 printk("migration: max_cache_size: %d, cpu: %d MHz:\n",
6123 if (system_state
== SYSTEM_BOOTING
&& num_online_cpus() > 1) {
6124 printk("migration_cost=");
6125 for (distance
= 0; distance
<= max_distance
; distance
++) {
6128 printk("%ld", (long)migration_cost
[distance
] / 1000);
6133 if (migration_debug
)
6134 printk("migration: %ld seconds\n", (j1
-j0
) / HZ
);
6137 * Move back to the original CPU. NUMA-Q gets confused
6138 * if we migrate to another quad during bootup.
6140 if (raw_smp_processor_id() != orig_cpu
) {
6141 cpumask_t mask
= cpumask_of_cpu(orig_cpu
),
6142 saved_mask
= current
->cpus_allowed
;
6144 set_cpus_allowed(current
, mask
);
6145 set_cpus_allowed(current
, saved_mask
);
6152 * find_next_best_node - find the next node to include in a sched_domain
6153 * @node: node whose sched_domain we're building
6154 * @used_nodes: nodes already in the sched_domain
6156 * Find the next node to include in a given scheduling domain. Simply
6157 * finds the closest node not already in the @used_nodes map.
6159 * Should use nodemask_t.
6161 static int find_next_best_node(int node
, unsigned long *used_nodes
)
6163 int i
, n
, val
, min_val
, best_node
= 0;
6167 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6168 /* Start at @node */
6169 n
= (node
+ i
) % MAX_NUMNODES
;
6171 if (!nr_cpus_node(n
))
6174 /* Skip already used nodes */
6175 if (test_bit(n
, used_nodes
))
6178 /* Simple min distance search */
6179 val
= node_distance(node
, n
);
6181 if (val
< min_val
) {
6187 set_bit(best_node
, used_nodes
);
6192 * sched_domain_node_span - get a cpumask for a node's sched_domain
6193 * @node: node whose cpumask we're constructing
6194 * @size: number of nodes to include in this span
6196 * Given a node, construct a good cpumask for its sched_domain to span. It
6197 * should be one that prevents unnecessary balancing, but also spreads tasks
6200 static cpumask_t
sched_domain_node_span(int node
)
6202 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
6203 cpumask_t span
, nodemask
;
6207 bitmap_zero(used_nodes
, MAX_NUMNODES
);
6209 nodemask
= node_to_cpumask(node
);
6210 cpus_or(span
, span
, nodemask
);
6211 set_bit(node
, used_nodes
);
6213 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6214 int next_node
= find_next_best_node(node
, used_nodes
);
6216 nodemask
= node_to_cpumask(next_node
);
6217 cpus_or(span
, span
, nodemask
);
6224 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6227 * SMT sched-domains:
6229 #ifdef CONFIG_SCHED_SMT
6230 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
6231 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
6233 static int cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
,
6234 struct sched_group
**sg
)
6237 *sg
= &per_cpu(sched_group_cpus
, cpu
);
6243 * multi-core sched-domains:
6245 #ifdef CONFIG_SCHED_MC
6246 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
6247 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
6250 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6251 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
6252 struct sched_group
**sg
)
6255 cpumask_t mask
= cpu_sibling_map
[cpu
];
6256 cpus_and(mask
, mask
, *cpu_map
);
6257 group
= first_cpu(mask
);
6259 *sg
= &per_cpu(sched_group_core
, group
);
6262 #elif defined(CONFIG_SCHED_MC)
6263 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
6264 struct sched_group
**sg
)
6267 *sg
= &per_cpu(sched_group_core
, cpu
);
6272 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
6273 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
6275 static int cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
,
6276 struct sched_group
**sg
)
6279 #ifdef CONFIG_SCHED_MC
6280 cpumask_t mask
= cpu_coregroup_map(cpu
);
6281 cpus_and(mask
, mask
, *cpu_map
);
6282 group
= first_cpu(mask
);
6283 #elif defined(CONFIG_SCHED_SMT)
6284 cpumask_t mask
= cpu_sibling_map
[cpu
];
6285 cpus_and(mask
, mask
, *cpu_map
);
6286 group
= first_cpu(mask
);
6291 *sg
= &per_cpu(sched_group_phys
, group
);
6297 * The init_sched_build_groups can't handle what we want to do with node
6298 * groups, so roll our own. Now each node has its own list of groups which
6299 * gets dynamically allocated.
6301 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
6302 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
6304 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
6305 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
6307 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
6308 struct sched_group
**sg
)
6310 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
6313 cpus_and(nodemask
, nodemask
, *cpu_map
);
6314 group
= first_cpu(nodemask
);
6317 *sg
= &per_cpu(sched_group_allnodes
, group
);
6321 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6323 struct sched_group
*sg
= group_head
;
6329 for_each_cpu_mask(j
, sg
->cpumask
) {
6330 struct sched_domain
*sd
;
6332 sd
= &per_cpu(phys_domains
, j
);
6333 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
6335 * Only add "power" once for each
6341 sg
->cpu_power
+= sd
->groups
->cpu_power
;
6344 if (sg
!= group_head
)
6350 /* Free memory allocated for various sched_group structures */
6351 static void free_sched_groups(const cpumask_t
*cpu_map
)
6355 for_each_cpu_mask(cpu
, *cpu_map
) {
6356 struct sched_group
**sched_group_nodes
6357 = sched_group_nodes_bycpu
[cpu
];
6359 if (!sched_group_nodes
)
6362 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6363 cpumask_t nodemask
= node_to_cpumask(i
);
6364 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6366 cpus_and(nodemask
, nodemask
, *cpu_map
);
6367 if (cpus_empty(nodemask
))
6377 if (oldsg
!= sched_group_nodes
[i
])
6380 kfree(sched_group_nodes
);
6381 sched_group_nodes_bycpu
[cpu
] = NULL
;
6385 static void free_sched_groups(const cpumask_t
*cpu_map
)
6391 * Initialize sched groups cpu_power.
6393 * cpu_power indicates the capacity of sched group, which is used while
6394 * distributing the load between different sched groups in a sched domain.
6395 * Typically cpu_power for all the groups in a sched domain will be same unless
6396 * there are asymmetries in the topology. If there are asymmetries, group
6397 * having more cpu_power will pickup more load compared to the group having
6400 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6401 * the maximum number of tasks a group can handle in the presence of other idle
6402 * or lightly loaded groups in the same sched domain.
6404 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6406 struct sched_domain
*child
;
6407 struct sched_group
*group
;
6409 WARN_ON(!sd
|| !sd
->groups
);
6411 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
6417 * For perf policy, if the groups in child domain share resources
6418 * (for example cores sharing some portions of the cache hierarchy
6419 * or SMT), then set this domain groups cpu_power such that each group
6420 * can handle only one task, when there are other idle groups in the
6421 * same sched domain.
6423 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
6425 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
6426 sd
->groups
->cpu_power
= SCHED_LOAD_SCALE
;
6430 sd
->groups
->cpu_power
= 0;
6433 * add cpu_power of each child group to this groups cpu_power
6435 group
= child
->groups
;
6437 sd
->groups
->cpu_power
+= group
->cpu_power
;
6438 group
= group
->next
;
6439 } while (group
!= child
->groups
);
6443 * Build sched domains for a given set of cpus and attach the sched domains
6444 * to the individual cpus
6446 static int build_sched_domains(const cpumask_t
*cpu_map
)
6449 struct sched_domain
*sd
;
6451 struct sched_group
**sched_group_nodes
= NULL
;
6452 int sd_allnodes
= 0;
6455 * Allocate the per-node list of sched groups
6457 sched_group_nodes
= kzalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
6459 if (!sched_group_nodes
) {
6460 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6463 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6467 * Set up domains for cpus specified by the cpu_map.
6469 for_each_cpu_mask(i
, *cpu_map
) {
6470 struct sched_domain
*sd
= NULL
, *p
;
6471 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6473 cpus_and(nodemask
, nodemask
, *cpu_map
);
6476 if (cpus_weight(*cpu_map
)
6477 > SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6478 sd
= &per_cpu(allnodes_domains
, i
);
6479 *sd
= SD_ALLNODES_INIT
;
6480 sd
->span
= *cpu_map
;
6481 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6487 sd
= &per_cpu(node_domains
, i
);
6489 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6493 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6497 sd
= &per_cpu(phys_domains
, i
);
6499 sd
->span
= nodemask
;
6503 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6505 #ifdef CONFIG_SCHED_MC
6507 sd
= &per_cpu(core_domains
, i
);
6509 sd
->span
= cpu_coregroup_map(i
);
6510 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6513 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6516 #ifdef CONFIG_SCHED_SMT
6518 sd
= &per_cpu(cpu_domains
, i
);
6519 *sd
= SD_SIBLING_INIT
;
6520 sd
->span
= cpu_sibling_map
[i
];
6521 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6524 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6528 #ifdef CONFIG_SCHED_SMT
6529 /* Set up CPU (sibling) groups */
6530 for_each_cpu_mask(i
, *cpu_map
) {
6531 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
6532 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6533 if (i
!= first_cpu(this_sibling_map
))
6536 init_sched_build_groups(this_sibling_map
, cpu_map
, &cpu_to_cpu_group
);
6540 #ifdef CONFIG_SCHED_MC
6541 /* Set up multi-core groups */
6542 for_each_cpu_mask(i
, *cpu_map
) {
6543 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6544 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6545 if (i
!= first_cpu(this_core_map
))
6547 init_sched_build_groups(this_core_map
, cpu_map
, &cpu_to_core_group
);
6552 /* Set up physical groups */
6553 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6554 cpumask_t nodemask
= node_to_cpumask(i
);
6556 cpus_and(nodemask
, nodemask
, *cpu_map
);
6557 if (cpus_empty(nodemask
))
6560 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6564 /* Set up node groups */
6566 init_sched_build_groups(*cpu_map
, cpu_map
, &cpu_to_allnodes_group
);
6568 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6569 /* Set up node groups */
6570 struct sched_group
*sg
, *prev
;
6571 cpumask_t nodemask
= node_to_cpumask(i
);
6572 cpumask_t domainspan
;
6573 cpumask_t covered
= CPU_MASK_NONE
;
6576 cpus_and(nodemask
, nodemask
, *cpu_map
);
6577 if (cpus_empty(nodemask
)) {
6578 sched_group_nodes
[i
] = NULL
;
6582 domainspan
= sched_domain_node_span(i
);
6583 cpus_and(domainspan
, domainspan
, *cpu_map
);
6585 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6587 printk(KERN_WARNING
"Can not alloc domain group for "
6591 sched_group_nodes
[i
] = sg
;
6592 for_each_cpu_mask(j
, nodemask
) {
6593 struct sched_domain
*sd
;
6594 sd
= &per_cpu(node_domains
, j
);
6598 sg
->cpumask
= nodemask
;
6600 cpus_or(covered
, covered
, nodemask
);
6603 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6604 cpumask_t tmp
, notcovered
;
6605 int n
= (i
+ j
) % MAX_NUMNODES
;
6607 cpus_complement(notcovered
, covered
);
6608 cpus_and(tmp
, notcovered
, *cpu_map
);
6609 cpus_and(tmp
, tmp
, domainspan
);
6610 if (cpus_empty(tmp
))
6613 nodemask
= node_to_cpumask(n
);
6614 cpus_and(tmp
, tmp
, nodemask
);
6615 if (cpus_empty(tmp
))
6618 sg
= kmalloc_node(sizeof(struct sched_group
),
6622 "Can not alloc domain group for node %d\n", j
);
6627 sg
->next
= prev
->next
;
6628 cpus_or(covered
, covered
, tmp
);
6635 /* Calculate CPU power for physical packages and nodes */
6636 #ifdef CONFIG_SCHED_SMT
6637 for_each_cpu_mask(i
, *cpu_map
) {
6638 sd
= &per_cpu(cpu_domains
, i
);
6639 init_sched_groups_power(i
, sd
);
6642 #ifdef CONFIG_SCHED_MC
6643 for_each_cpu_mask(i
, *cpu_map
) {
6644 sd
= &per_cpu(core_domains
, i
);
6645 init_sched_groups_power(i
, sd
);
6649 for_each_cpu_mask(i
, *cpu_map
) {
6650 sd
= &per_cpu(phys_domains
, i
);
6651 init_sched_groups_power(i
, sd
);
6655 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6656 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6659 struct sched_group
*sg
;
6661 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6662 init_numa_sched_groups_power(sg
);
6666 /* Attach the domains */
6667 for_each_cpu_mask(i
, *cpu_map
) {
6668 struct sched_domain
*sd
;
6669 #ifdef CONFIG_SCHED_SMT
6670 sd
= &per_cpu(cpu_domains
, i
);
6671 #elif defined(CONFIG_SCHED_MC)
6672 sd
= &per_cpu(core_domains
, i
);
6674 sd
= &per_cpu(phys_domains
, i
);
6676 cpu_attach_domain(sd
, i
);
6679 * Tune cache-hot values:
6681 calibrate_migration_costs(cpu_map
);
6687 free_sched_groups(cpu_map
);
6692 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6694 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6696 cpumask_t cpu_default_map
;
6700 * Setup mask for cpus without special case scheduling requirements.
6701 * For now this just excludes isolated cpus, but could be used to
6702 * exclude other special cases in the future.
6704 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
6706 err
= build_sched_domains(&cpu_default_map
);
6711 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6713 free_sched_groups(cpu_map
);
6717 * Detach sched domains from a group of cpus specified in cpu_map
6718 * These cpus will now be attached to the NULL domain
6720 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6724 for_each_cpu_mask(i
, *cpu_map
)
6725 cpu_attach_domain(NULL
, i
);
6726 synchronize_sched();
6727 arch_destroy_sched_domains(cpu_map
);
6731 * Partition sched domains as specified by the cpumasks below.
6732 * This attaches all cpus from the cpumasks to the NULL domain,
6733 * waits for a RCU quiescent period, recalculates sched
6734 * domain information and then attaches them back to the
6735 * correct sched domains
6736 * Call with hotplug lock held
6738 int partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
6740 cpumask_t change_map
;
6743 cpus_and(*partition1
, *partition1
, cpu_online_map
);
6744 cpus_and(*partition2
, *partition2
, cpu_online_map
);
6745 cpus_or(change_map
, *partition1
, *partition2
);
6747 /* Detach sched domains from all of the affected cpus */
6748 detach_destroy_domains(&change_map
);
6749 if (!cpus_empty(*partition1
))
6750 err
= build_sched_domains(partition1
);
6751 if (!err
&& !cpus_empty(*partition2
))
6752 err
= build_sched_domains(partition2
);
6757 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6758 int arch_reinit_sched_domains(void)
6763 detach_destroy_domains(&cpu_online_map
);
6764 err
= arch_init_sched_domains(&cpu_online_map
);
6765 unlock_cpu_hotplug();
6770 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6774 if (buf
[0] != '0' && buf
[0] != '1')
6778 sched_smt_power_savings
= (buf
[0] == '1');
6780 sched_mc_power_savings
= (buf
[0] == '1');
6782 ret
= arch_reinit_sched_domains();
6784 return ret
? ret
: count
;
6787 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6791 #ifdef CONFIG_SCHED_SMT
6793 err
= sysfs_create_file(&cls
->kset
.kobj
,
6794 &attr_sched_smt_power_savings
.attr
);
6796 #ifdef CONFIG_SCHED_MC
6797 if (!err
&& mc_capable())
6798 err
= sysfs_create_file(&cls
->kset
.kobj
,
6799 &attr_sched_mc_power_savings
.attr
);
6805 #ifdef CONFIG_SCHED_MC
6806 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6808 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6810 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6811 const char *buf
, size_t count
)
6813 return sched_power_savings_store(buf
, count
, 0);
6815 SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6816 sched_mc_power_savings_store
);
6819 #ifdef CONFIG_SCHED_SMT
6820 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6822 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6824 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6825 const char *buf
, size_t count
)
6827 return sched_power_savings_store(buf
, count
, 1);
6829 SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6830 sched_smt_power_savings_store
);
6834 * Force a reinitialization of the sched domains hierarchy. The domains
6835 * and groups cannot be updated in place without racing with the balancing
6836 * code, so we temporarily attach all running cpus to the NULL domain
6837 * which will prevent rebalancing while the sched domains are recalculated.
6839 static int update_sched_domains(struct notifier_block
*nfb
,
6840 unsigned long action
, void *hcpu
)
6843 case CPU_UP_PREPARE
:
6844 case CPU_DOWN_PREPARE
:
6845 detach_destroy_domains(&cpu_online_map
);
6848 case CPU_UP_CANCELED
:
6849 case CPU_DOWN_FAILED
:
6853 * Fall through and re-initialise the domains.
6860 /* The hotplug lock is already held by cpu_up/cpu_down */
6861 arch_init_sched_domains(&cpu_online_map
);
6866 void __init
sched_init_smp(void)
6868 cpumask_t non_isolated_cpus
;
6871 arch_init_sched_domains(&cpu_online_map
);
6872 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
6873 if (cpus_empty(non_isolated_cpus
))
6874 cpu_set(smp_processor_id(), non_isolated_cpus
);
6875 unlock_cpu_hotplug();
6876 /* XXX: Theoretical race here - CPU may be hotplugged now */
6877 hotcpu_notifier(update_sched_domains
, 0);
6879 /* Move init over to a non-isolated CPU */
6880 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
6884 void __init
sched_init_smp(void)
6887 #endif /* CONFIG_SMP */
6889 int in_sched_functions(unsigned long addr
)
6891 /* Linker adds these: start and end of __sched functions */
6892 extern char __sched_text_start
[], __sched_text_end
[];
6894 return in_lock_functions(addr
) ||
6895 (addr
>= (unsigned long)__sched_text_start
6896 && addr
< (unsigned long)__sched_text_end
);
6899 void __init
sched_init(void)
6903 for_each_possible_cpu(i
) {
6904 struct prio_array
*array
;
6908 spin_lock_init(&rq
->lock
);
6909 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
6911 rq
->active
= rq
->arrays
;
6912 rq
->expired
= rq
->arrays
+ 1;
6913 rq
->best_expired_prio
= MAX_PRIO
;
6917 for (j
= 1; j
< 3; j
++)
6918 rq
->cpu_load
[j
] = 0;
6919 rq
->active_balance
= 0;
6922 rq
->migration_thread
= NULL
;
6923 INIT_LIST_HEAD(&rq
->migration_queue
);
6925 atomic_set(&rq
->nr_iowait
, 0);
6927 for (j
= 0; j
< 2; j
++) {
6928 array
= rq
->arrays
+ j
;
6929 for (k
= 0; k
< MAX_PRIO
; k
++) {
6930 INIT_LIST_HEAD(array
->queue
+ k
);
6931 __clear_bit(k
, array
->bitmap
);
6933 // delimiter for bitsearch
6934 __set_bit(MAX_PRIO
, array
->bitmap
);
6938 set_load_weight(&init_task
);
6941 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
6944 #ifdef CONFIG_RT_MUTEXES
6945 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
6949 * The boot idle thread does lazy MMU switching as well:
6951 atomic_inc(&init_mm
.mm_count
);
6952 enter_lazy_tlb(&init_mm
, current
);
6955 * Make us the idle thread. Technically, schedule() should not be
6956 * called from this thread, however somewhere below it might be,
6957 * but because we are the idle thread, we just pick up running again
6958 * when this runqueue becomes "idle".
6960 init_idle(current
, smp_processor_id());
6963 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6964 void __might_sleep(char *file
, int line
)
6967 static unsigned long prev_jiffy
; /* ratelimiting */
6969 if ((in_atomic() || irqs_disabled()) &&
6970 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6971 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6973 prev_jiffy
= jiffies
;
6974 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6975 " context at %s:%d\n", file
, line
);
6976 printk("in_atomic():%d, irqs_disabled():%d\n",
6977 in_atomic(), irqs_disabled());
6978 debug_show_held_locks(current
);
6979 if (irqs_disabled())
6980 print_irqtrace_events(current
);
6985 EXPORT_SYMBOL(__might_sleep
);
6988 #ifdef CONFIG_MAGIC_SYSRQ
6989 void normalize_rt_tasks(void)
6991 struct prio_array
*array
;
6992 struct task_struct
*p
;
6993 unsigned long flags
;
6996 read_lock_irq(&tasklist_lock
);
6997 for_each_process(p
) {
7001 spin_lock_irqsave(&p
->pi_lock
, flags
);
7002 rq
= __task_rq_lock(p
);
7006 deactivate_task(p
, task_rq(p
));
7007 __setscheduler(p
, SCHED_NORMAL
, 0);
7009 __activate_task(p
, task_rq(p
));
7010 resched_task(rq
->curr
);
7013 __task_rq_unlock(rq
);
7014 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
7016 read_unlock_irq(&tasklist_lock
);
7019 #endif /* CONFIG_MAGIC_SYSRQ */
7023 * These functions are only useful for the IA64 MCA handling.
7025 * They can only be called when the whole system has been
7026 * stopped - every CPU needs to be quiescent, and no scheduling
7027 * activity can take place. Using them for anything else would
7028 * be a serious bug, and as a result, they aren't even visible
7029 * under any other configuration.
7033 * curr_task - return the current task for a given cpu.
7034 * @cpu: the processor in question.
7036 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7038 struct task_struct
*curr_task(int cpu
)
7040 return cpu_curr(cpu
);
7044 * set_curr_task - set the current task for a given cpu.
7045 * @cpu: the processor in question.
7046 * @p: the task pointer to set.
7048 * Description: This function must only be used when non-maskable interrupts
7049 * are serviced on a separate stack. It allows the architecture to switch the
7050 * notion of the current task on a cpu in a non-blocking manner. This function
7051 * must be called with all CPU's synchronized, and interrupts disabled, the
7052 * and caller must save the original value of the current task (see
7053 * curr_task() above) and restore that value before reenabling interrupts and
7054 * re-starting the system.
7056 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7058 void set_curr_task(int cpu
, struct task_struct
*p
)