4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
22 #include <linux/module.h>
23 #include <linux/nmi.h>
24 #include <linux/init.h>
25 #include <asm/uaccess.h>
26 #include <linux/highmem.h>
27 #include <linux/smp_lock.h>
28 #include <asm/mmu_context.h>
29 #include <linux/interrupt.h>
30 #include <linux/capability.h>
31 #include <linux/completion.h>
32 #include <linux/kernel_stat.h>
33 #include <linux/debug_locks.h>
34 #include <linux/security.h>
35 #include <linux/notifier.h>
36 #include <linux/profile.h>
37 #include <linux/suspend.h>
38 #include <linux/vmalloc.h>
39 #include <linux/blkdev.h>
40 #include <linux/delay.h>
41 #include <linux/smp.h>
42 #include <linux/threads.h>
43 #include <linux/timer.h>
44 #include <linux/rcupdate.h>
45 #include <linux/cpu.h>
46 #include <linux/cpuset.h>
47 #include <linux/percpu.h>
48 #include <linux/kthread.h>
49 #include <linux/seq_file.h>
50 #include <linux/syscalls.h>
51 #include <linux/times.h>
52 #include <linux/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 unsigned long long timestamp_last_tick
;
229 struct task_struct
*curr
, *idle
;
230 struct mm_struct
*prev_mm
;
231 struct prio_array
*active
, *expired
, arrays
[2];
232 int best_expired_prio
;
236 struct sched_domain
*sd
;
238 /* For active balancing */
241 int cpu
; /* cpu of this runqueue */
243 struct task_struct
*migration_thread
;
244 struct list_head migration_queue
;
247 #ifdef CONFIG_SCHEDSTATS
249 struct sched_info rq_sched_info
;
251 /* sys_sched_yield() stats */
252 unsigned long yld_exp_empty
;
253 unsigned long yld_act_empty
;
254 unsigned long yld_both_empty
;
255 unsigned long yld_cnt
;
257 /* schedule() stats */
258 unsigned long sched_switch
;
259 unsigned long sched_cnt
;
260 unsigned long sched_goidle
;
262 /* try_to_wake_up() stats */
263 unsigned long ttwu_cnt
;
264 unsigned long ttwu_local
;
266 struct lock_class_key rq_lock_key
;
269 static DEFINE_PER_CPU(struct rq
, runqueues
);
271 static inline int cpu_of(struct rq
*rq
)
281 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
282 * See detach_destroy_domains: synchronize_sched for details.
284 * The domain tree of any CPU may only be accessed from within
285 * preempt-disabled sections.
287 #define for_each_domain(cpu, __sd) \
288 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
290 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
291 #define this_rq() (&__get_cpu_var(runqueues))
292 #define task_rq(p) cpu_rq(task_cpu(p))
293 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
295 #ifndef prepare_arch_switch
296 # define prepare_arch_switch(next) do { } while (0)
298 #ifndef finish_arch_switch
299 # define finish_arch_switch(prev) do { } while (0)
302 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
303 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
305 return rq
->curr
== p
;
308 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
312 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
314 #ifdef CONFIG_DEBUG_SPINLOCK
315 /* this is a valid case when another task releases the spinlock */
316 rq
->lock
.owner
= current
;
319 * If we are tracking spinlock dependencies then we have to
320 * fix up the runqueue lock - which gets 'carried over' from
323 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
325 spin_unlock_irq(&rq
->lock
);
328 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
329 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
334 return rq
->curr
== p
;
338 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
342 * We can optimise this out completely for !SMP, because the
343 * SMP rebalancing from interrupt is the only thing that cares
348 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
349 spin_unlock_irq(&rq
->lock
);
351 spin_unlock(&rq
->lock
);
355 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
359 * After ->oncpu is cleared, the task can be moved to a different CPU.
360 * We must ensure this doesn't happen until the switch is completely
366 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
370 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
373 * __task_rq_lock - lock the runqueue a given task resides on.
374 * Must be called interrupts disabled.
376 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
383 spin_lock(&rq
->lock
);
384 if (unlikely(rq
!= task_rq(p
))) {
385 spin_unlock(&rq
->lock
);
386 goto repeat_lock_task
;
392 * task_rq_lock - lock the runqueue a given task resides on and disable
393 * interrupts. Note the ordering: we can safely lookup the task_rq without
394 * explicitly disabling preemption.
396 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
402 local_irq_save(*flags
);
404 spin_lock(&rq
->lock
);
405 if (unlikely(rq
!= task_rq(p
))) {
406 spin_unlock_irqrestore(&rq
->lock
, *flags
);
407 goto repeat_lock_task
;
412 static inline void __task_rq_unlock(struct rq
*rq
)
415 spin_unlock(&rq
->lock
);
418 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
421 spin_unlock_irqrestore(&rq
->lock
, *flags
);
424 #ifdef CONFIG_SCHEDSTATS
426 * bump this up when changing the output format or the meaning of an existing
427 * format, so that tools can adapt (or abort)
429 #define SCHEDSTAT_VERSION 12
431 static int show_schedstat(struct seq_file
*seq
, void *v
)
435 seq_printf(seq
, "version %d\n", SCHEDSTAT_VERSION
);
436 seq_printf(seq
, "timestamp %lu\n", jiffies
);
437 for_each_online_cpu(cpu
) {
438 struct rq
*rq
= cpu_rq(cpu
);
440 struct sched_domain
*sd
;
444 /* runqueue-specific stats */
446 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
447 cpu
, rq
->yld_both_empty
,
448 rq
->yld_act_empty
, rq
->yld_exp_empty
, rq
->yld_cnt
,
449 rq
->sched_switch
, rq
->sched_cnt
, rq
->sched_goidle
,
450 rq
->ttwu_cnt
, rq
->ttwu_local
,
451 rq
->rq_sched_info
.cpu_time
,
452 rq
->rq_sched_info
.run_delay
, rq
->rq_sched_info
.pcnt
);
454 seq_printf(seq
, "\n");
457 /* domain-specific stats */
459 for_each_domain(cpu
, sd
) {
460 enum idle_type itype
;
461 char mask_str
[NR_CPUS
];
463 cpumask_scnprintf(mask_str
, NR_CPUS
, sd
->span
);
464 seq_printf(seq
, "domain%d %s", dcnt
++, mask_str
);
465 for (itype
= SCHED_IDLE
; itype
< MAX_IDLE_TYPES
;
467 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu %lu",
469 sd
->lb_balanced
[itype
],
470 sd
->lb_failed
[itype
],
471 sd
->lb_imbalance
[itype
],
472 sd
->lb_gained
[itype
],
473 sd
->lb_hot_gained
[itype
],
474 sd
->lb_nobusyq
[itype
],
475 sd
->lb_nobusyg
[itype
]);
477 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
478 sd
->alb_cnt
, sd
->alb_failed
, sd
->alb_pushed
,
479 sd
->sbe_cnt
, sd
->sbe_balanced
, sd
->sbe_pushed
,
480 sd
->sbf_cnt
, sd
->sbf_balanced
, sd
->sbf_pushed
,
481 sd
->ttwu_wake_remote
, sd
->ttwu_move_affine
, sd
->ttwu_move_balance
);
489 static int schedstat_open(struct inode
*inode
, struct file
*file
)
491 unsigned int size
= PAGE_SIZE
* (1 + num_online_cpus() / 32);
492 char *buf
= kmalloc(size
, GFP_KERNEL
);
498 res
= single_open(file
, show_schedstat
, NULL
);
500 m
= file
->private_data
;
508 struct file_operations proc_schedstat_operations
= {
509 .open
= schedstat_open
,
512 .release
= single_release
,
516 * Expects runqueue lock to be held for atomicity of update
519 rq_sched_info_arrive(struct rq
*rq
, unsigned long delta_jiffies
)
522 rq
->rq_sched_info
.run_delay
+= delta_jiffies
;
523 rq
->rq_sched_info
.pcnt
++;
528 * Expects runqueue lock to be held for atomicity of update
531 rq_sched_info_depart(struct rq
*rq
, unsigned long delta_jiffies
)
534 rq
->rq_sched_info
.cpu_time
+= delta_jiffies
;
536 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
537 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
538 #else /* !CONFIG_SCHEDSTATS */
540 rq_sched_info_arrive(struct rq
*rq
, unsigned long delta_jiffies
)
543 rq_sched_info_depart(struct rq
*rq
, unsigned long delta_jiffies
)
545 # define schedstat_inc(rq, field) do { } while (0)
546 # define schedstat_add(rq, field, amt) do { } while (0)
550 * rq_lock - lock a given runqueue and disable interrupts.
552 static inline struct rq
*this_rq_lock(void)
559 spin_lock(&rq
->lock
);
564 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
566 * Called when a process is dequeued from the active array and given
567 * the cpu. We should note that with the exception of interactive
568 * tasks, the expired queue will become the active queue after the active
569 * queue is empty, without explicitly dequeuing and requeuing tasks in the
570 * expired queue. (Interactive tasks may be requeued directly to the
571 * active queue, thus delaying tasks in the expired queue from running;
572 * see scheduler_tick()).
574 * This function is only called from sched_info_arrive(), rather than
575 * dequeue_task(). Even though a task may be queued and dequeued multiple
576 * times as it is shuffled about, we're really interested in knowing how
577 * long it was from the *first* time it was queued to the time that it
580 static inline void sched_info_dequeued(struct task_struct
*t
)
582 t
->sched_info
.last_queued
= 0;
586 * Called when a task finally hits the cpu. We can now calculate how
587 * long it was waiting to run. We also note when it began so that we
588 * can keep stats on how long its timeslice is.
590 static void sched_info_arrive(struct task_struct
*t
)
592 unsigned long now
= jiffies
, delta_jiffies
= 0;
594 if (t
->sched_info
.last_queued
)
595 delta_jiffies
= now
- t
->sched_info
.last_queued
;
596 sched_info_dequeued(t
);
597 t
->sched_info
.run_delay
+= delta_jiffies
;
598 t
->sched_info
.last_arrival
= now
;
599 t
->sched_info
.pcnt
++;
601 rq_sched_info_arrive(task_rq(t
), delta_jiffies
);
605 * Called when a process is queued into either the active or expired
606 * array. The time is noted and later used to determine how long we
607 * had to wait for us to reach the cpu. Since the expired queue will
608 * become the active queue after active queue is empty, without dequeuing
609 * and requeuing any tasks, we are interested in queuing to either. It
610 * is unusual but not impossible for tasks to be dequeued and immediately
611 * requeued in the same or another array: this can happen in sched_yield(),
612 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
615 * This function is only called from enqueue_task(), but also only updates
616 * the timestamp if it is already not set. It's assumed that
617 * sched_info_dequeued() will clear that stamp when appropriate.
619 static inline void sched_info_queued(struct task_struct
*t
)
621 if (unlikely(sched_info_on()))
622 if (!t
->sched_info
.last_queued
)
623 t
->sched_info
.last_queued
= jiffies
;
627 * Called when a process ceases being the active-running process, either
628 * voluntarily or involuntarily. Now we can calculate how long we ran.
630 static inline void sched_info_depart(struct task_struct
*t
)
632 unsigned long delta_jiffies
= jiffies
- t
->sched_info
.last_arrival
;
634 t
->sched_info
.cpu_time
+= delta_jiffies
;
635 rq_sched_info_depart(task_rq(t
), delta_jiffies
);
639 * Called when tasks are switched involuntarily due, typically, to expiring
640 * their time slice. (This may also be called when switching to or from
641 * the idle task.) We are only called when prev != next.
644 __sched_info_switch(struct task_struct
*prev
, struct task_struct
*next
)
646 struct rq
*rq
= task_rq(prev
);
649 * prev now departs the cpu. It's not interesting to record
650 * stats about how efficient we were at scheduling the idle
653 if (prev
!= rq
->idle
)
654 sched_info_depart(prev
);
656 if (next
!= rq
->idle
)
657 sched_info_arrive(next
);
660 sched_info_switch(struct task_struct
*prev
, struct task_struct
*next
)
662 if (unlikely(sched_info_on()))
663 __sched_info_switch(prev
, next
);
666 #define sched_info_queued(t) do { } while (0)
667 #define sched_info_switch(t, next) do { } while (0)
668 #endif /* CONFIG_SCHEDSTATS || CONFIG_TASK_DELAY_ACCT */
671 * Adding/removing a task to/from a priority array:
673 static void dequeue_task(struct task_struct
*p
, struct prio_array
*array
)
676 list_del(&p
->run_list
);
677 if (list_empty(array
->queue
+ p
->prio
))
678 __clear_bit(p
->prio
, array
->bitmap
);
681 static void enqueue_task(struct task_struct
*p
, struct prio_array
*array
)
683 sched_info_queued(p
);
684 list_add_tail(&p
->run_list
, array
->queue
+ p
->prio
);
685 __set_bit(p
->prio
, array
->bitmap
);
691 * Put task to the end of the run list without the overhead of dequeue
692 * followed by enqueue.
694 static void requeue_task(struct task_struct
*p
, struct prio_array
*array
)
696 list_move_tail(&p
->run_list
, array
->queue
+ p
->prio
);
700 enqueue_task_head(struct task_struct
*p
, struct prio_array
*array
)
702 list_add(&p
->run_list
, array
->queue
+ p
->prio
);
703 __set_bit(p
->prio
, array
->bitmap
);
709 * __normal_prio - return the priority that is based on the static
710 * priority but is modified by bonuses/penalties.
712 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
713 * into the -5 ... 0 ... +5 bonus/penalty range.
715 * We use 25% of the full 0...39 priority range so that:
717 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
718 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
720 * Both properties are important to certain workloads.
723 static inline int __normal_prio(struct task_struct
*p
)
727 bonus
= CURRENT_BONUS(p
) - MAX_BONUS
/ 2;
729 prio
= p
->static_prio
- bonus
;
730 if (prio
< MAX_RT_PRIO
)
732 if (prio
> MAX_PRIO
-1)
738 * To aid in avoiding the subversion of "niceness" due to uneven distribution
739 * of tasks with abnormal "nice" values across CPUs the contribution that
740 * each task makes to its run queue's load is weighted according to its
741 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
742 * scaled version of the new time slice allocation that they receive on time
747 * Assume: static_prio_timeslice(NICE_TO_PRIO(0)) == DEF_TIMESLICE
748 * If static_prio_timeslice() is ever changed to break this assumption then
749 * this code will need modification
751 #define TIME_SLICE_NICE_ZERO DEF_TIMESLICE
752 #define LOAD_WEIGHT(lp) \
753 (((lp) * SCHED_LOAD_SCALE) / TIME_SLICE_NICE_ZERO)
754 #define PRIO_TO_LOAD_WEIGHT(prio) \
755 LOAD_WEIGHT(static_prio_timeslice(prio))
756 #define RTPRIO_TO_LOAD_WEIGHT(rp) \
757 (PRIO_TO_LOAD_WEIGHT(MAX_RT_PRIO) + LOAD_WEIGHT(rp))
759 static void set_load_weight(struct task_struct
*p
)
761 if (has_rt_policy(p
)) {
763 if (p
== task_rq(p
)->migration_thread
)
765 * The migration thread does the actual balancing.
766 * Giving its load any weight will skew balancing
772 p
->load_weight
= RTPRIO_TO_LOAD_WEIGHT(p
->rt_priority
);
774 p
->load_weight
= PRIO_TO_LOAD_WEIGHT(p
->static_prio
);
778 inc_raw_weighted_load(struct rq
*rq
, const struct task_struct
*p
)
780 rq
->raw_weighted_load
+= p
->load_weight
;
784 dec_raw_weighted_load(struct rq
*rq
, const struct task_struct
*p
)
786 rq
->raw_weighted_load
-= p
->load_weight
;
789 static inline void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
792 inc_raw_weighted_load(rq
, p
);
795 static inline void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
798 dec_raw_weighted_load(rq
, p
);
802 * Calculate the expected normal priority: i.e. priority
803 * without taking RT-inheritance into account. Might be
804 * boosted by interactivity modifiers. Changes upon fork,
805 * setprio syscalls, and whenever the interactivity
806 * estimator recalculates.
808 static inline int normal_prio(struct task_struct
*p
)
812 if (has_rt_policy(p
))
813 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
815 prio
= __normal_prio(p
);
820 * Calculate the current priority, i.e. the priority
821 * taken into account by the scheduler. This value might
822 * be boosted by RT tasks, or might be boosted by
823 * interactivity modifiers. Will be RT if the task got
824 * RT-boosted. If not then it returns p->normal_prio.
826 static int effective_prio(struct task_struct
*p
)
828 p
->normal_prio
= normal_prio(p
);
830 * If we are RT tasks or we were boosted to RT priority,
831 * keep the priority unchanged. Otherwise, update priority
832 * to the normal priority:
834 if (!rt_prio(p
->prio
))
835 return p
->normal_prio
;
840 * __activate_task - move a task to the runqueue.
842 static void __activate_task(struct task_struct
*p
, struct rq
*rq
)
844 struct prio_array
*target
= rq
->active
;
847 target
= rq
->expired
;
848 enqueue_task(p
, target
);
849 inc_nr_running(p
, rq
);
853 * __activate_idle_task - move idle task to the _front_ of runqueue.
855 static inline void __activate_idle_task(struct task_struct
*p
, struct rq
*rq
)
857 enqueue_task_head(p
, rq
->active
);
858 inc_nr_running(p
, rq
);
862 * Recalculate p->normal_prio and p->prio after having slept,
863 * updating the sleep-average too:
865 static int recalc_task_prio(struct task_struct
*p
, unsigned long long now
)
867 /* Caller must always ensure 'now >= p->timestamp' */
868 unsigned long sleep_time
= now
- p
->timestamp
;
873 if (likely(sleep_time
> 0)) {
875 * This ceiling is set to the lowest priority that would allow
876 * a task to be reinserted into the active array on timeslice
879 unsigned long ceiling
= INTERACTIVE_SLEEP(p
);
881 if (p
->mm
&& sleep_time
> ceiling
&& p
->sleep_avg
< ceiling
) {
883 * Prevents user tasks from achieving best priority
884 * with one single large enough sleep.
886 p
->sleep_avg
= ceiling
;
888 * Using INTERACTIVE_SLEEP() as a ceiling places a
889 * nice(0) task 1ms sleep away from promotion, and
890 * gives it 700ms to round-robin with no chance of
891 * being demoted. This is more than generous, so
892 * mark this sleep as non-interactive to prevent the
893 * on-runqueue bonus logic from intervening should
894 * this task not receive cpu immediately.
896 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
899 * Tasks waking from uninterruptible sleep are
900 * limited in their sleep_avg rise as they
901 * are likely to be waiting on I/O
903 if (p
->sleep_type
== SLEEP_NONINTERACTIVE
&& p
->mm
) {
904 if (p
->sleep_avg
>= ceiling
)
906 else if (p
->sleep_avg
+ sleep_time
>=
908 p
->sleep_avg
= ceiling
;
914 * This code gives a bonus to interactive tasks.
916 * The boost works by updating the 'average sleep time'
917 * value here, based on ->timestamp. The more time a
918 * task spends sleeping, the higher the average gets -
919 * and the higher the priority boost gets as well.
921 p
->sleep_avg
+= sleep_time
;
924 if (p
->sleep_avg
> NS_MAX_SLEEP_AVG
)
925 p
->sleep_avg
= NS_MAX_SLEEP_AVG
;
928 return effective_prio(p
);
932 * activate_task - move a task to the runqueue and do priority recalculation
934 * Update all the scheduling statistics stuff. (sleep average
935 * calculation, priority modifiers, etc.)
937 static void activate_task(struct task_struct
*p
, struct rq
*rq
, int local
)
939 unsigned long long now
;
944 /* Compensate for drifting sched_clock */
945 struct rq
*this_rq
= this_rq();
946 now
= (now
- this_rq
->timestamp_last_tick
)
947 + rq
->timestamp_last_tick
;
952 p
->prio
= recalc_task_prio(p
, now
);
955 * This checks to make sure it's not an uninterruptible task
956 * that is now waking up.
958 if (p
->sleep_type
== SLEEP_NORMAL
) {
960 * Tasks which were woken up by interrupts (ie. hw events)
961 * are most likely of interactive nature. So we give them
962 * the credit of extending their sleep time to the period
963 * of time they spend on the runqueue, waiting for execution
964 * on a CPU, first time around:
967 p
->sleep_type
= SLEEP_INTERRUPTED
;
970 * Normal first-time wakeups get a credit too for
971 * on-runqueue time, but it will be weighted down:
973 p
->sleep_type
= SLEEP_INTERACTIVE
;
978 __activate_task(p
, rq
);
982 * deactivate_task - remove a task from the runqueue.
984 static void deactivate_task(struct task_struct
*p
, struct rq
*rq
)
986 dec_nr_running(p
, rq
);
987 dequeue_task(p
, p
->array
);
992 * resched_task - mark a task 'to be rescheduled now'.
994 * On UP this means the setting of the need_resched flag, on SMP it
995 * might also involve a cross-CPU call to trigger the scheduler on
1000 #ifndef tsk_is_polling
1001 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1004 static void resched_task(struct task_struct
*p
)
1008 assert_spin_locked(&task_rq(p
)->lock
);
1010 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
1013 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
1016 if (cpu
== smp_processor_id())
1019 /* NEED_RESCHED must be visible before we test polling */
1021 if (!tsk_is_polling(p
))
1022 smp_send_reschedule(cpu
);
1025 static inline void resched_task(struct task_struct
*p
)
1027 assert_spin_locked(&task_rq(p
)->lock
);
1028 set_tsk_need_resched(p
);
1033 * task_curr - is this task currently executing on a CPU?
1034 * @p: the task in question.
1036 inline int task_curr(const struct task_struct
*p
)
1038 return cpu_curr(task_cpu(p
)) == p
;
1041 /* Used instead of source_load when we know the type == 0 */
1042 unsigned long weighted_cpuload(const int cpu
)
1044 return cpu_rq(cpu
)->raw_weighted_load
;
1048 struct migration_req
{
1049 struct list_head list
;
1051 struct task_struct
*task
;
1054 struct completion done
;
1058 * The task's runqueue lock must be held.
1059 * Returns true if you have to wait for migration thread.
1062 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1064 struct rq
*rq
= task_rq(p
);
1067 * If the task is not on a runqueue (and not running), then
1068 * it is sufficient to simply update the task's cpu field.
1070 if (!p
->array
&& !task_running(rq
, p
)) {
1071 set_task_cpu(p
, dest_cpu
);
1075 init_completion(&req
->done
);
1077 req
->dest_cpu
= dest_cpu
;
1078 list_add(&req
->list
, &rq
->migration_queue
);
1084 * wait_task_inactive - wait for a thread to unschedule.
1086 * The caller must ensure that the task *will* unschedule sometime soon,
1087 * else this function might spin for a *long* time. This function can't
1088 * be called with interrupts off, or it may introduce deadlock with
1089 * smp_call_function() if an IPI is sent by the same process we are
1090 * waiting to become inactive.
1092 void wait_task_inactive(struct task_struct
*p
)
1094 unsigned long flags
;
1099 rq
= task_rq_lock(p
, &flags
);
1100 /* Must be off runqueue entirely, not preempted. */
1101 if (unlikely(p
->array
|| task_running(rq
, p
))) {
1102 /* If it's preempted, we yield. It could be a while. */
1103 preempted
= !task_running(rq
, p
);
1104 task_rq_unlock(rq
, &flags
);
1110 task_rq_unlock(rq
, &flags
);
1114 * kick_process - kick a running thread to enter/exit the kernel
1115 * @p: the to-be-kicked thread
1117 * Cause a process which is running on another CPU to enter
1118 * kernel-mode, without any delay. (to get signals handled.)
1120 * NOTE: this function doesnt have to take the runqueue lock,
1121 * because all it wants to ensure is that the remote task enters
1122 * the kernel. If the IPI races and the task has been migrated
1123 * to another CPU then no harm is done and the purpose has been
1126 void kick_process(struct task_struct
*p
)
1132 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1133 smp_send_reschedule(cpu
);
1138 * Return a low guess at the load of a migration-source cpu weighted
1139 * according to the scheduling class and "nice" value.
1141 * We want to under-estimate the load of migration sources, to
1142 * balance conservatively.
1144 static inline unsigned long source_load(int cpu
, int type
)
1146 struct rq
*rq
= cpu_rq(cpu
);
1149 return rq
->raw_weighted_load
;
1151 return min(rq
->cpu_load
[type
-1], rq
->raw_weighted_load
);
1155 * Return a high guess at the load of a migration-target cpu weighted
1156 * according to the scheduling class and "nice" value.
1158 static inline unsigned long target_load(int cpu
, int type
)
1160 struct rq
*rq
= cpu_rq(cpu
);
1163 return rq
->raw_weighted_load
;
1165 return max(rq
->cpu_load
[type
-1], rq
->raw_weighted_load
);
1169 * Return the average load per task on the cpu's run queue
1171 static inline unsigned long cpu_avg_load_per_task(int cpu
)
1173 struct rq
*rq
= cpu_rq(cpu
);
1174 unsigned long n
= rq
->nr_running
;
1176 return n
? rq
->raw_weighted_load
/ n
: SCHED_LOAD_SCALE
;
1180 * find_idlest_group finds and returns the least busy CPU group within the
1183 static struct sched_group
*
1184 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1186 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1187 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1188 int load_idx
= sd
->forkexec_idx
;
1189 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1192 unsigned long load
, avg_load
;
1196 /* Skip over this group if it has no CPUs allowed */
1197 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1200 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1202 /* Tally up the load of all CPUs in the group */
1205 for_each_cpu_mask(i
, group
->cpumask
) {
1206 /* Bias balancing toward cpus of our domain */
1208 load
= source_load(i
, load_idx
);
1210 load
= target_load(i
, load_idx
);
1215 /* Adjust by relative CPU power of the group */
1216 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
1219 this_load
= avg_load
;
1221 } else if (avg_load
< min_load
) {
1222 min_load
= avg_load
;
1226 group
= group
->next
;
1227 } while (group
!= sd
->groups
);
1229 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1235 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1238 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1241 unsigned long load
, min_load
= ULONG_MAX
;
1245 /* Traverse only the allowed CPUs */
1246 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1248 for_each_cpu_mask(i
, tmp
) {
1249 load
= weighted_cpuload(i
);
1251 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1261 * sched_balance_self: balance the current task (running on cpu) in domains
1262 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1265 * Balance, ie. select the least loaded group.
1267 * Returns the target CPU number, or the same CPU if no balancing is needed.
1269 * preempt must be disabled.
1271 static int sched_balance_self(int cpu
, int flag
)
1273 struct task_struct
*t
= current
;
1274 struct sched_domain
*tmp
, *sd
= NULL
;
1276 for_each_domain(cpu
, tmp
) {
1278 * If power savings logic is enabled for a domain, stop there.
1280 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1282 if (tmp
->flags
& flag
)
1288 struct sched_group
*group
;
1289 int new_cpu
, weight
;
1291 if (!(sd
->flags
& flag
)) {
1297 group
= find_idlest_group(sd
, t
, cpu
);
1303 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1304 if (new_cpu
== -1 || new_cpu
== cpu
) {
1305 /* Now try balancing at a lower domain level of cpu */
1310 /* Now try balancing at a lower domain level of new_cpu */
1313 weight
= cpus_weight(span
);
1314 for_each_domain(cpu
, tmp
) {
1315 if (weight
<= cpus_weight(tmp
->span
))
1317 if (tmp
->flags
& flag
)
1320 /* while loop will break here if sd == NULL */
1326 #endif /* CONFIG_SMP */
1329 * wake_idle() will wake a task on an idle cpu if task->cpu is
1330 * not idle and an idle cpu is available. The span of cpus to
1331 * search starts with cpus closest then further out as needed,
1332 * so we always favor a closer, idle cpu.
1334 * Returns the CPU we should wake onto.
1336 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1337 static int wake_idle(int cpu
, struct task_struct
*p
)
1340 struct sched_domain
*sd
;
1346 for_each_domain(cpu
, sd
) {
1347 if (sd
->flags
& SD_WAKE_IDLE
) {
1348 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1349 for_each_cpu_mask(i
, tmp
) {
1360 static inline int wake_idle(int cpu
, struct task_struct
*p
)
1367 * try_to_wake_up - wake up a thread
1368 * @p: the to-be-woken-up thread
1369 * @state: the mask of task states that can be woken
1370 * @sync: do a synchronous wakeup?
1372 * Put it on the run-queue if it's not already there. The "current"
1373 * thread is always on the run-queue (except when the actual
1374 * re-schedule is in progress), and as such you're allowed to do
1375 * the simpler "current->state = TASK_RUNNING" to mark yourself
1376 * runnable without the overhead of this.
1378 * returns failure only if the task is already active.
1380 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1382 int cpu
, this_cpu
, success
= 0;
1383 unsigned long flags
;
1387 struct sched_domain
*sd
, *this_sd
= NULL
;
1388 unsigned long load
, this_load
;
1392 rq
= task_rq_lock(p
, &flags
);
1393 old_state
= p
->state
;
1394 if (!(old_state
& state
))
1401 this_cpu
= smp_processor_id();
1404 if (unlikely(task_running(rq
, p
)))
1409 schedstat_inc(rq
, ttwu_cnt
);
1410 if (cpu
== this_cpu
) {
1411 schedstat_inc(rq
, ttwu_local
);
1415 for_each_domain(this_cpu
, sd
) {
1416 if (cpu_isset(cpu
, sd
->span
)) {
1417 schedstat_inc(sd
, ttwu_wake_remote
);
1423 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1427 * Check for affine wakeup and passive balancing possibilities.
1430 int idx
= this_sd
->wake_idx
;
1431 unsigned int imbalance
;
1433 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1435 load
= source_load(cpu
, idx
);
1436 this_load
= target_load(this_cpu
, idx
);
1438 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1440 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1441 unsigned long tl
= this_load
;
1442 unsigned long tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1445 * If sync wakeup then subtract the (maximum possible)
1446 * effect of the currently running task from the load
1447 * of the current CPU:
1450 tl
-= current
->load_weight
;
1453 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1454 100*(tl
+ p
->load_weight
) <= imbalance
*load
) {
1456 * This domain has SD_WAKE_AFFINE and
1457 * p is cache cold in this domain, and
1458 * there is no bad imbalance.
1460 schedstat_inc(this_sd
, ttwu_move_affine
);
1466 * Start passive balancing when half the imbalance_pct
1469 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1470 if (imbalance
*this_load
<= 100*load
) {
1471 schedstat_inc(this_sd
, ttwu_move_balance
);
1477 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1479 new_cpu
= wake_idle(new_cpu
, p
);
1480 if (new_cpu
!= cpu
) {
1481 set_task_cpu(p
, new_cpu
);
1482 task_rq_unlock(rq
, &flags
);
1483 /* might preempt at this point */
1484 rq
= task_rq_lock(p
, &flags
);
1485 old_state
= p
->state
;
1486 if (!(old_state
& state
))
1491 this_cpu
= smp_processor_id();
1496 #endif /* CONFIG_SMP */
1497 if (old_state
== TASK_UNINTERRUPTIBLE
) {
1498 rq
->nr_uninterruptible
--;
1500 * Tasks on involuntary sleep don't earn
1501 * sleep_avg beyond just interactive state.
1503 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
1507 * Tasks that have marked their sleep as noninteractive get
1508 * woken up with their sleep average not weighted in an
1511 if (old_state
& TASK_NONINTERACTIVE
)
1512 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
1515 activate_task(p
, rq
, cpu
== this_cpu
);
1517 * Sync wakeups (i.e. those types of wakeups where the waker
1518 * has indicated that it will leave the CPU in short order)
1519 * don't trigger a preemption, if the woken up task will run on
1520 * this cpu. (in this case the 'I will reschedule' promise of
1521 * the waker guarantees that the freshly woken up task is going
1522 * to be considered on this CPU.)
1524 if (!sync
|| cpu
!= this_cpu
) {
1525 if (TASK_PREEMPTS_CURR(p
, rq
))
1526 resched_task(rq
->curr
);
1531 p
->state
= TASK_RUNNING
;
1533 task_rq_unlock(rq
, &flags
);
1538 int fastcall
wake_up_process(struct task_struct
*p
)
1540 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1541 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1543 EXPORT_SYMBOL(wake_up_process
);
1545 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1547 return try_to_wake_up(p
, state
, 0);
1551 * Perform scheduler related setup for a newly forked process p.
1552 * p is forked by current.
1554 void fastcall
sched_fork(struct task_struct
*p
, int clone_flags
)
1556 int cpu
= get_cpu();
1559 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1561 set_task_cpu(p
, cpu
);
1564 * We mark the process as running here, but have not actually
1565 * inserted it onto the runqueue yet. This guarantees that
1566 * nobody will actually run it, and a signal or other external
1567 * event cannot wake it up and insert it on the runqueue either.
1569 p
->state
= TASK_RUNNING
;
1572 * Make sure we do not leak PI boosting priority to the child:
1574 p
->prio
= current
->normal_prio
;
1576 INIT_LIST_HEAD(&p
->run_list
);
1578 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1579 if (unlikely(sched_info_on()))
1580 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1582 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1585 #ifdef CONFIG_PREEMPT
1586 /* Want to start with kernel preemption disabled. */
1587 task_thread_info(p
)->preempt_count
= 1;
1590 * Share the timeslice between parent and child, thus the
1591 * total amount of pending timeslices in the system doesn't change,
1592 * resulting in more scheduling fairness.
1594 local_irq_disable();
1595 p
->time_slice
= (current
->time_slice
+ 1) >> 1;
1597 * The remainder of the first timeslice might be recovered by
1598 * the parent if the child exits early enough.
1600 p
->first_time_slice
= 1;
1601 current
->time_slice
>>= 1;
1602 p
->timestamp
= sched_clock();
1603 if (unlikely(!current
->time_slice
)) {
1605 * This case is rare, it happens when the parent has only
1606 * a single jiffy left from its timeslice. Taking the
1607 * runqueue lock is not a problem.
1609 current
->time_slice
= 1;
1617 * wake_up_new_task - wake up a newly created task for the first time.
1619 * This function will do some initial scheduler statistics housekeeping
1620 * that must be done for every newly created context, then puts the task
1621 * on the runqueue and wakes it.
1623 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1625 struct rq
*rq
, *this_rq
;
1626 unsigned long flags
;
1629 rq
= task_rq_lock(p
, &flags
);
1630 BUG_ON(p
->state
!= TASK_RUNNING
);
1631 this_cpu
= smp_processor_id();
1635 * We decrease the sleep average of forking parents
1636 * and children as well, to keep max-interactive tasks
1637 * from forking tasks that are max-interactive. The parent
1638 * (current) is done further down, under its lock.
1640 p
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(p
) *
1641 CHILD_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1643 p
->prio
= effective_prio(p
);
1645 if (likely(cpu
== this_cpu
)) {
1646 if (!(clone_flags
& CLONE_VM
)) {
1648 * The VM isn't cloned, so we're in a good position to
1649 * do child-runs-first in anticipation of an exec. This
1650 * usually avoids a lot of COW overhead.
1652 if (unlikely(!current
->array
))
1653 __activate_task(p
, rq
);
1655 p
->prio
= current
->prio
;
1656 p
->normal_prio
= current
->normal_prio
;
1657 list_add_tail(&p
->run_list
, ¤t
->run_list
);
1658 p
->array
= current
->array
;
1659 p
->array
->nr_active
++;
1660 inc_nr_running(p
, rq
);
1664 /* Run child last */
1665 __activate_task(p
, rq
);
1667 * We skip the following code due to cpu == this_cpu
1669 * task_rq_unlock(rq, &flags);
1670 * this_rq = task_rq_lock(current, &flags);
1674 this_rq
= cpu_rq(this_cpu
);
1677 * Not the local CPU - must adjust timestamp. This should
1678 * get optimised away in the !CONFIG_SMP case.
1680 p
->timestamp
= (p
->timestamp
- this_rq
->timestamp_last_tick
)
1681 + rq
->timestamp_last_tick
;
1682 __activate_task(p
, rq
);
1683 if (TASK_PREEMPTS_CURR(p
, rq
))
1684 resched_task(rq
->curr
);
1687 * Parent and child are on different CPUs, now get the
1688 * parent runqueue to update the parent's ->sleep_avg:
1690 task_rq_unlock(rq
, &flags
);
1691 this_rq
= task_rq_lock(current
, &flags
);
1693 current
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(current
) *
1694 PARENT_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1695 task_rq_unlock(this_rq
, &flags
);
1699 * Potentially available exiting-child timeslices are
1700 * retrieved here - this way the parent does not get
1701 * penalized for creating too many threads.
1703 * (this cannot be used to 'generate' timeslices
1704 * artificially, because any timeslice recovered here
1705 * was given away by the parent in the first place.)
1707 void fastcall
sched_exit(struct task_struct
*p
)
1709 unsigned long flags
;
1713 * If the child was a (relative-) CPU hog then decrease
1714 * the sleep_avg of the parent as well.
1716 rq
= task_rq_lock(p
->parent
, &flags
);
1717 if (p
->first_time_slice
&& task_cpu(p
) == task_cpu(p
->parent
)) {
1718 p
->parent
->time_slice
+= p
->time_slice
;
1719 if (unlikely(p
->parent
->time_slice
> task_timeslice(p
)))
1720 p
->parent
->time_slice
= task_timeslice(p
);
1722 if (p
->sleep_avg
< p
->parent
->sleep_avg
)
1723 p
->parent
->sleep_avg
= p
->parent
->sleep_avg
/
1724 (EXIT_WEIGHT
+ 1) * EXIT_WEIGHT
+ p
->sleep_avg
/
1726 task_rq_unlock(rq
, &flags
);
1730 * prepare_task_switch - prepare to switch tasks
1731 * @rq: the runqueue preparing to switch
1732 * @next: the task we are going to switch to.
1734 * This is called with the rq lock held and interrupts off. It must
1735 * be paired with a subsequent finish_task_switch after the context
1738 * prepare_task_switch sets up locking and calls architecture specific
1741 static inline void prepare_task_switch(struct rq
*rq
, struct task_struct
*next
)
1743 prepare_lock_switch(rq
, next
);
1744 prepare_arch_switch(next
);
1748 * finish_task_switch - clean up after a task-switch
1749 * @rq: runqueue associated with task-switch
1750 * @prev: the thread we just switched away from.
1752 * finish_task_switch must be called after the context switch, paired
1753 * with a prepare_task_switch call before the context switch.
1754 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1755 * and do any other architecture-specific cleanup actions.
1757 * Note that we may have delayed dropping an mm in context_switch(). If
1758 * so, we finish that here outside of the runqueue lock. (Doing it
1759 * with the lock held can cause deadlocks; see schedule() for
1762 static inline void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1763 __releases(rq
->lock
)
1765 struct mm_struct
*mm
= rq
->prev_mm
;
1771 * A task struct has one reference for the use as "current".
1772 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1773 * schedule one last time. The schedule call will never return, and
1774 * the scheduled task must drop that reference.
1775 * The test for TASK_DEAD must occur while the runqueue locks are
1776 * still held, otherwise prev could be scheduled on another cpu, die
1777 * there before we look at prev->state, and then the reference would
1779 * Manfred Spraul <manfred@colorfullife.com>
1781 prev_state
= prev
->state
;
1782 finish_arch_switch(prev
);
1783 finish_lock_switch(rq
, prev
);
1786 if (unlikely(prev_state
== TASK_DEAD
)) {
1788 * Remove function-return probe instances associated with this
1789 * task and put them back on the free list.
1791 kprobe_flush_task(prev
);
1792 put_task_struct(prev
);
1797 * schedule_tail - first thing a freshly forked thread must call.
1798 * @prev: the thread we just switched away from.
1800 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1801 __releases(rq
->lock
)
1803 struct rq
*rq
= this_rq();
1805 finish_task_switch(rq
, prev
);
1806 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1807 /* In this case, finish_task_switch does not reenable preemption */
1810 if (current
->set_child_tid
)
1811 put_user(current
->pid
, current
->set_child_tid
);
1815 * context_switch - switch to the new MM and the new
1816 * thread's register state.
1818 static inline struct task_struct
*
1819 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1820 struct task_struct
*next
)
1822 struct mm_struct
*mm
= next
->mm
;
1823 struct mm_struct
*oldmm
= prev
->active_mm
;
1826 next
->active_mm
= oldmm
;
1827 atomic_inc(&oldmm
->mm_count
);
1828 enter_lazy_tlb(oldmm
, next
);
1830 switch_mm(oldmm
, mm
, next
);
1833 prev
->active_mm
= NULL
;
1834 WARN_ON(rq
->prev_mm
);
1835 rq
->prev_mm
= oldmm
;
1838 * Since the runqueue lock will be released by the next
1839 * task (which is an invalid locking op but in the case
1840 * of the scheduler it's an obvious special-case), so we
1841 * do an early lockdep release here:
1843 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1844 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1847 /* Here we just switch the register state and the stack. */
1848 switch_to(prev
, next
, prev
);
1854 * nr_running, nr_uninterruptible and nr_context_switches:
1856 * externally visible scheduler statistics: current number of runnable
1857 * threads, current number of uninterruptible-sleeping threads, total
1858 * number of context switches performed since bootup.
1860 unsigned long nr_running(void)
1862 unsigned long i
, sum
= 0;
1864 for_each_online_cpu(i
)
1865 sum
+= cpu_rq(i
)->nr_running
;
1870 unsigned long nr_uninterruptible(void)
1872 unsigned long i
, sum
= 0;
1874 for_each_possible_cpu(i
)
1875 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1878 * Since we read the counters lockless, it might be slightly
1879 * inaccurate. Do not allow it to go below zero though:
1881 if (unlikely((long)sum
< 0))
1887 unsigned long long nr_context_switches(void)
1890 unsigned long long sum
= 0;
1892 for_each_possible_cpu(i
)
1893 sum
+= cpu_rq(i
)->nr_switches
;
1898 unsigned long nr_iowait(void)
1900 unsigned long i
, sum
= 0;
1902 for_each_possible_cpu(i
)
1903 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1908 unsigned long nr_active(void)
1910 unsigned long i
, running
= 0, uninterruptible
= 0;
1912 for_each_online_cpu(i
) {
1913 running
+= cpu_rq(i
)->nr_running
;
1914 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
1917 if (unlikely((long)uninterruptible
< 0))
1918 uninterruptible
= 0;
1920 return running
+ uninterruptible
;
1926 * Is this task likely cache-hot:
1929 task_hot(struct task_struct
*p
, unsigned long long now
, struct sched_domain
*sd
)
1931 return (long long)(now
- p
->last_ran
) < (long long)sd
->cache_hot_time
;
1935 * double_rq_lock - safely lock two runqueues
1937 * Note this does not disable interrupts like task_rq_lock,
1938 * you need to do so manually before calling.
1940 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1941 __acquires(rq1
->lock
)
1942 __acquires(rq2
->lock
)
1944 BUG_ON(!irqs_disabled());
1946 spin_lock(&rq1
->lock
);
1947 __acquire(rq2
->lock
); /* Fake it out ;) */
1950 spin_lock(&rq1
->lock
);
1951 spin_lock(&rq2
->lock
);
1953 spin_lock(&rq2
->lock
);
1954 spin_lock(&rq1
->lock
);
1960 * double_rq_unlock - safely unlock two runqueues
1962 * Note this does not restore interrupts like task_rq_unlock,
1963 * you need to do so manually after calling.
1965 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
1966 __releases(rq1
->lock
)
1967 __releases(rq2
->lock
)
1969 spin_unlock(&rq1
->lock
);
1971 spin_unlock(&rq2
->lock
);
1973 __release(rq2
->lock
);
1977 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1979 static void double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1980 __releases(this_rq
->lock
)
1981 __acquires(busiest
->lock
)
1982 __acquires(this_rq
->lock
)
1984 if (unlikely(!irqs_disabled())) {
1985 /* printk() doesn't work good under rq->lock */
1986 spin_unlock(&this_rq
->lock
);
1989 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1990 if (busiest
< this_rq
) {
1991 spin_unlock(&this_rq
->lock
);
1992 spin_lock(&busiest
->lock
);
1993 spin_lock(&this_rq
->lock
);
1995 spin_lock(&busiest
->lock
);
2000 * If dest_cpu is allowed for this process, migrate the task to it.
2001 * This is accomplished by forcing the cpu_allowed mask to only
2002 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2003 * the cpu_allowed mask is restored.
2005 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2007 struct migration_req req
;
2008 unsigned long flags
;
2011 rq
= task_rq_lock(p
, &flags
);
2012 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2013 || unlikely(cpu_is_offline(dest_cpu
)))
2016 /* force the process onto the specified CPU */
2017 if (migrate_task(p
, dest_cpu
, &req
)) {
2018 /* Need to wait for migration thread (might exit: take ref). */
2019 struct task_struct
*mt
= rq
->migration_thread
;
2021 get_task_struct(mt
);
2022 task_rq_unlock(rq
, &flags
);
2023 wake_up_process(mt
);
2024 put_task_struct(mt
);
2025 wait_for_completion(&req
.done
);
2030 task_rq_unlock(rq
, &flags
);
2034 * sched_exec - execve() is a valuable balancing opportunity, because at
2035 * this point the task has the smallest effective memory and cache footprint.
2037 void sched_exec(void)
2039 int new_cpu
, this_cpu
= get_cpu();
2040 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2042 if (new_cpu
!= this_cpu
)
2043 sched_migrate_task(current
, new_cpu
);
2047 * pull_task - move a task from a remote runqueue to the local runqueue.
2048 * Both runqueues must be locked.
2050 static void pull_task(struct rq
*src_rq
, struct prio_array
*src_array
,
2051 struct task_struct
*p
, struct rq
*this_rq
,
2052 struct prio_array
*this_array
, int this_cpu
)
2054 dequeue_task(p
, src_array
);
2055 dec_nr_running(p
, src_rq
);
2056 set_task_cpu(p
, this_cpu
);
2057 inc_nr_running(p
, this_rq
);
2058 enqueue_task(p
, this_array
);
2059 p
->timestamp
= (p
->timestamp
- src_rq
->timestamp_last_tick
)
2060 + this_rq
->timestamp_last_tick
;
2062 * Note that idle threads have a prio of MAX_PRIO, for this test
2063 * to be always true for them.
2065 if (TASK_PREEMPTS_CURR(p
, this_rq
))
2066 resched_task(this_rq
->curr
);
2070 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2073 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2074 struct sched_domain
*sd
, enum idle_type idle
,
2078 * We do not migrate tasks that are:
2079 * 1) running (obviously), or
2080 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2081 * 3) are cache-hot on their current CPU.
2083 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
2087 if (task_running(rq
, p
))
2091 * Aggressive migration if:
2092 * 1) task is cache cold, or
2093 * 2) too many balance attempts have failed.
2096 if (sd
->nr_balance_failed
> sd
->cache_nice_tries
)
2099 if (task_hot(p
, rq
->timestamp_last_tick
, sd
))
2104 #define rq_best_prio(rq) min((rq)->curr->prio, (rq)->best_expired_prio)
2107 * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted
2108 * load from busiest to this_rq, as part of a balancing operation within
2109 * "domain". Returns the number of tasks moved.
2111 * Called with both runqueues locked.
2113 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2114 unsigned long max_nr_move
, unsigned long max_load_move
,
2115 struct sched_domain
*sd
, enum idle_type idle
,
2118 int idx
, pulled
= 0, pinned
= 0, this_best_prio
, best_prio
,
2119 best_prio_seen
, skip_for_load
;
2120 struct prio_array
*array
, *dst_array
;
2121 struct list_head
*head
, *curr
;
2122 struct task_struct
*tmp
;
2125 if (max_nr_move
== 0 || max_load_move
== 0)
2128 rem_load_move
= max_load_move
;
2130 this_best_prio
= rq_best_prio(this_rq
);
2131 best_prio
= rq_best_prio(busiest
);
2133 * Enable handling of the case where there is more than one task
2134 * with the best priority. If the current running task is one
2135 * of those with prio==best_prio we know it won't be moved
2136 * and therefore it's safe to override the skip (based on load) of
2137 * any task we find with that prio.
2139 best_prio_seen
= best_prio
== busiest
->curr
->prio
;
2142 * We first consider expired tasks. Those will likely not be
2143 * executed in the near future, and they are most likely to
2144 * be cache-cold, thus switching CPUs has the least effect
2147 if (busiest
->expired
->nr_active
) {
2148 array
= busiest
->expired
;
2149 dst_array
= this_rq
->expired
;
2151 array
= busiest
->active
;
2152 dst_array
= this_rq
->active
;
2156 /* Start searching at priority 0: */
2160 idx
= sched_find_first_bit(array
->bitmap
);
2162 idx
= find_next_bit(array
->bitmap
, MAX_PRIO
, idx
);
2163 if (idx
>= MAX_PRIO
) {
2164 if (array
== busiest
->expired
&& busiest
->active
->nr_active
) {
2165 array
= busiest
->active
;
2166 dst_array
= this_rq
->active
;
2172 head
= array
->queue
+ idx
;
2175 tmp
= list_entry(curr
, struct task_struct
, run_list
);
2180 * To help distribute high priority tasks accross CPUs we don't
2181 * skip a task if it will be the highest priority task (i.e. smallest
2182 * prio value) on its new queue regardless of its load weight
2184 skip_for_load
= tmp
->load_weight
> rem_load_move
;
2185 if (skip_for_load
&& idx
< this_best_prio
)
2186 skip_for_load
= !best_prio_seen
&& idx
== best_prio
;
2187 if (skip_for_load
||
2188 !can_migrate_task(tmp
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2190 best_prio_seen
|= idx
== best_prio
;
2197 #ifdef CONFIG_SCHEDSTATS
2198 if (task_hot(tmp
, busiest
->timestamp_last_tick
, sd
))
2199 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2202 pull_task(busiest
, array
, tmp
, this_rq
, dst_array
, this_cpu
);
2204 rem_load_move
-= tmp
->load_weight
;
2207 * We only want to steal up to the prescribed number of tasks
2208 * and the prescribed amount of weighted load.
2210 if (pulled
< max_nr_move
&& rem_load_move
> 0) {
2211 if (idx
< this_best_prio
)
2212 this_best_prio
= idx
;
2220 * Right now, this is the only place pull_task() is called,
2221 * so we can safely collect pull_task() stats here rather than
2222 * inside pull_task().
2224 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2227 *all_pinned
= pinned
;
2232 * find_busiest_group finds and returns the busiest CPU group within the
2233 * domain. It calculates and returns the amount of weighted load which
2234 * should be moved to restore balance via the imbalance parameter.
2236 static struct sched_group
*
2237 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2238 unsigned long *imbalance
, enum idle_type idle
, int *sd_idle
,
2241 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2242 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2243 unsigned long max_pull
;
2244 unsigned long busiest_load_per_task
, busiest_nr_running
;
2245 unsigned long this_load_per_task
, this_nr_running
;
2247 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2248 int power_savings_balance
= 1;
2249 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2250 unsigned long min_nr_running
= ULONG_MAX
;
2251 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2254 max_load
= this_load
= total_load
= total_pwr
= 0;
2255 busiest_load_per_task
= busiest_nr_running
= 0;
2256 this_load_per_task
= this_nr_running
= 0;
2257 if (idle
== NOT_IDLE
)
2258 load_idx
= sd
->busy_idx
;
2259 else if (idle
== NEWLY_IDLE
)
2260 load_idx
= sd
->newidle_idx
;
2262 load_idx
= sd
->idle_idx
;
2265 unsigned long load
, group_capacity
;
2268 unsigned long sum_nr_running
, sum_weighted_load
;
2270 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2272 /* Tally up the load of all CPUs in the group */
2273 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2275 for_each_cpu_mask(i
, group
->cpumask
) {
2278 if (!cpu_isset(i
, *cpus
))
2283 if (*sd_idle
&& !idle_cpu(i
))
2286 /* Bias balancing toward cpus of our domain */
2288 load
= target_load(i
, load_idx
);
2290 load
= source_load(i
, load_idx
);
2293 sum_nr_running
+= rq
->nr_running
;
2294 sum_weighted_load
+= rq
->raw_weighted_load
;
2297 total_load
+= avg_load
;
2298 total_pwr
+= group
->cpu_power
;
2300 /* Adjust by relative CPU power of the group */
2301 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
2303 group_capacity
= group
->cpu_power
/ SCHED_LOAD_SCALE
;
2306 this_load
= avg_load
;
2308 this_nr_running
= sum_nr_running
;
2309 this_load_per_task
= sum_weighted_load
;
2310 } else if (avg_load
> max_load
&&
2311 sum_nr_running
> group_capacity
) {
2312 max_load
= avg_load
;
2314 busiest_nr_running
= sum_nr_running
;
2315 busiest_load_per_task
= sum_weighted_load
;
2318 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2320 * Busy processors will not participate in power savings
2323 if (idle
== NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2327 * If the local group is idle or completely loaded
2328 * no need to do power savings balance at this domain
2330 if (local_group
&& (this_nr_running
>= group_capacity
||
2332 power_savings_balance
= 0;
2335 * If a group is already running at full capacity or idle,
2336 * don't include that group in power savings calculations
2338 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2343 * Calculate the group which has the least non-idle load.
2344 * This is the group from where we need to pick up the load
2347 if ((sum_nr_running
< min_nr_running
) ||
2348 (sum_nr_running
== min_nr_running
&&
2349 first_cpu(group
->cpumask
) <
2350 first_cpu(group_min
->cpumask
))) {
2352 min_nr_running
= sum_nr_running
;
2353 min_load_per_task
= sum_weighted_load
/
2358 * Calculate the group which is almost near its
2359 * capacity but still has some space to pick up some load
2360 * from other group and save more power
2362 if (sum_nr_running
<= group_capacity
- 1) {
2363 if (sum_nr_running
> leader_nr_running
||
2364 (sum_nr_running
== leader_nr_running
&&
2365 first_cpu(group
->cpumask
) >
2366 first_cpu(group_leader
->cpumask
))) {
2367 group_leader
= group
;
2368 leader_nr_running
= sum_nr_running
;
2373 group
= group
->next
;
2374 } while (group
!= sd
->groups
);
2376 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2379 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2381 if (this_load
>= avg_load
||
2382 100*max_load
<= sd
->imbalance_pct
*this_load
)
2385 busiest_load_per_task
/= busiest_nr_running
;
2387 * We're trying to get all the cpus to the average_load, so we don't
2388 * want to push ourselves above the average load, nor do we wish to
2389 * reduce the max loaded cpu below the average load, as either of these
2390 * actions would just result in more rebalancing later, and ping-pong
2391 * tasks around. Thus we look for the minimum possible imbalance.
2392 * Negative imbalances (*we* are more loaded than anyone else) will
2393 * be counted as no imbalance for these purposes -- we can't fix that
2394 * by pulling tasks to us. Be careful of negative numbers as they'll
2395 * appear as very large values with unsigned longs.
2397 if (max_load
<= busiest_load_per_task
)
2401 * In the presence of smp nice balancing, certain scenarios can have
2402 * max load less than avg load(as we skip the groups at or below
2403 * its cpu_power, while calculating max_load..)
2405 if (max_load
< avg_load
) {
2407 goto small_imbalance
;
2410 /* Don't want to pull so many tasks that a group would go idle */
2411 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2413 /* How much load to actually move to equalise the imbalance */
2414 *imbalance
= min(max_pull
* busiest
->cpu_power
,
2415 (avg_load
- this_load
) * this->cpu_power
)
2419 * if *imbalance is less than the average load per runnable task
2420 * there is no gaurantee that any tasks will be moved so we'll have
2421 * a think about bumping its value to force at least one task to be
2424 if (*imbalance
< busiest_load_per_task
) {
2425 unsigned long tmp
, pwr_now
, pwr_move
;
2429 pwr_move
= pwr_now
= 0;
2431 if (this_nr_running
) {
2432 this_load_per_task
/= this_nr_running
;
2433 if (busiest_load_per_task
> this_load_per_task
)
2436 this_load_per_task
= SCHED_LOAD_SCALE
;
2438 if (max_load
- this_load
>= busiest_load_per_task
* imbn
) {
2439 *imbalance
= busiest_load_per_task
;
2444 * OK, we don't have enough imbalance to justify moving tasks,
2445 * however we may be able to increase total CPU power used by
2449 pwr_now
+= busiest
->cpu_power
*
2450 min(busiest_load_per_task
, max_load
);
2451 pwr_now
+= this->cpu_power
*
2452 min(this_load_per_task
, this_load
);
2453 pwr_now
/= SCHED_LOAD_SCALE
;
2455 /* Amount of load we'd subtract */
2456 tmp
= busiest_load_per_task
*SCHED_LOAD_SCALE
/busiest
->cpu_power
;
2458 pwr_move
+= busiest
->cpu_power
*
2459 min(busiest_load_per_task
, max_load
- tmp
);
2461 /* Amount of load we'd add */
2462 if (max_load
*busiest
->cpu_power
<
2463 busiest_load_per_task
*SCHED_LOAD_SCALE
)
2464 tmp
= max_load
*busiest
->cpu_power
/this->cpu_power
;
2466 tmp
= busiest_load_per_task
*SCHED_LOAD_SCALE
/this->cpu_power
;
2467 pwr_move
+= this->cpu_power
*min(this_load_per_task
, this_load
+ tmp
);
2468 pwr_move
/= SCHED_LOAD_SCALE
;
2470 /* Move if we gain throughput */
2471 if (pwr_move
<= pwr_now
)
2474 *imbalance
= busiest_load_per_task
;
2480 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2481 if (idle
== NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2484 if (this == group_leader
&& group_leader
!= group_min
) {
2485 *imbalance
= min_load_per_task
;
2495 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2498 find_busiest_queue(struct sched_group
*group
, enum idle_type idle
,
2499 unsigned long imbalance
, cpumask_t
*cpus
)
2501 struct rq
*busiest
= NULL
, *rq
;
2502 unsigned long max_load
= 0;
2505 for_each_cpu_mask(i
, group
->cpumask
) {
2507 if (!cpu_isset(i
, *cpus
))
2512 if (rq
->nr_running
== 1 && rq
->raw_weighted_load
> imbalance
)
2515 if (rq
->raw_weighted_load
> max_load
) {
2516 max_load
= rq
->raw_weighted_load
;
2525 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2526 * so long as it is large enough.
2528 #define MAX_PINNED_INTERVAL 512
2530 static inline unsigned long minus_1_or_zero(unsigned long n
)
2532 return n
> 0 ? n
- 1 : 0;
2536 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2537 * tasks if there is an imbalance.
2539 * Called with this_rq unlocked.
2541 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2542 struct sched_domain
*sd
, enum idle_type idle
)
2544 int nr_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2545 struct sched_group
*group
;
2546 unsigned long imbalance
;
2548 cpumask_t cpus
= CPU_MASK_ALL
;
2551 * When power savings policy is enabled for the parent domain, idle
2552 * sibling can pick up load irrespective of busy siblings. In this case,
2553 * let the state of idle sibling percolate up as IDLE, instead of
2554 * portraying it as NOT_IDLE.
2556 if (idle
!= NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2557 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2560 schedstat_inc(sd
, lb_cnt
[idle
]);
2563 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2566 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2570 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2572 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2576 BUG_ON(busiest
== this_rq
);
2578 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2581 if (busiest
->nr_running
> 1) {
2583 * Attempt to move tasks. If find_busiest_group has found
2584 * an imbalance but busiest->nr_running <= 1, the group is
2585 * still unbalanced. nr_moved simply stays zero, so it is
2586 * correctly treated as an imbalance.
2588 double_rq_lock(this_rq
, busiest
);
2589 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2590 minus_1_or_zero(busiest
->nr_running
),
2591 imbalance
, sd
, idle
, &all_pinned
);
2592 double_rq_unlock(this_rq
, busiest
);
2594 /* All tasks on this runqueue were pinned by CPU affinity */
2595 if (unlikely(all_pinned
)) {
2596 cpu_clear(cpu_of(busiest
), cpus
);
2597 if (!cpus_empty(cpus
))
2604 schedstat_inc(sd
, lb_failed
[idle
]);
2605 sd
->nr_balance_failed
++;
2607 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2609 spin_lock(&busiest
->lock
);
2611 /* don't kick the migration_thread, if the curr
2612 * task on busiest cpu can't be moved to this_cpu
2614 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2615 spin_unlock(&busiest
->lock
);
2617 goto out_one_pinned
;
2620 if (!busiest
->active_balance
) {
2621 busiest
->active_balance
= 1;
2622 busiest
->push_cpu
= this_cpu
;
2625 spin_unlock(&busiest
->lock
);
2627 wake_up_process(busiest
->migration_thread
);
2630 * We've kicked active balancing, reset the failure
2633 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2636 sd
->nr_balance_failed
= 0;
2638 if (likely(!active_balance
)) {
2639 /* We were unbalanced, so reset the balancing interval */
2640 sd
->balance_interval
= sd
->min_interval
;
2643 * If we've begun active balancing, start to back off. This
2644 * case may not be covered by the all_pinned logic if there
2645 * is only 1 task on the busy runqueue (because we don't call
2648 if (sd
->balance_interval
< sd
->max_interval
)
2649 sd
->balance_interval
*= 2;
2652 if (!nr_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2653 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2658 schedstat_inc(sd
, lb_balanced
[idle
]);
2660 sd
->nr_balance_failed
= 0;
2663 /* tune up the balancing interval */
2664 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2665 (sd
->balance_interval
< sd
->max_interval
))
2666 sd
->balance_interval
*= 2;
2668 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2669 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2675 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2676 * tasks if there is an imbalance.
2678 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2679 * this_rq is locked.
2682 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
2684 struct sched_group
*group
;
2685 struct rq
*busiest
= NULL
;
2686 unsigned long imbalance
;
2689 cpumask_t cpus
= CPU_MASK_ALL
;
2692 * When power savings policy is enabled for the parent domain, idle
2693 * sibling can pick up load irrespective of busy siblings. In this case,
2694 * let the state of idle sibling percolate up as IDLE, instead of
2695 * portraying it as NOT_IDLE.
2697 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
2698 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2701 schedstat_inc(sd
, lb_cnt
[NEWLY_IDLE
]);
2703 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, NEWLY_IDLE
,
2706 schedstat_inc(sd
, lb_nobusyg
[NEWLY_IDLE
]);
2710 busiest
= find_busiest_queue(group
, NEWLY_IDLE
, imbalance
,
2713 schedstat_inc(sd
, lb_nobusyq
[NEWLY_IDLE
]);
2717 BUG_ON(busiest
== this_rq
);
2719 schedstat_add(sd
, lb_imbalance
[NEWLY_IDLE
], imbalance
);
2722 if (busiest
->nr_running
> 1) {
2723 /* Attempt to move tasks */
2724 double_lock_balance(this_rq
, busiest
);
2725 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2726 minus_1_or_zero(busiest
->nr_running
),
2727 imbalance
, sd
, NEWLY_IDLE
, NULL
);
2728 spin_unlock(&busiest
->lock
);
2731 cpu_clear(cpu_of(busiest
), cpus
);
2732 if (!cpus_empty(cpus
))
2738 schedstat_inc(sd
, lb_failed
[NEWLY_IDLE
]);
2739 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2740 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2743 sd
->nr_balance_failed
= 0;
2748 schedstat_inc(sd
, lb_balanced
[NEWLY_IDLE
]);
2749 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2750 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2752 sd
->nr_balance_failed
= 0;
2758 * idle_balance is called by schedule() if this_cpu is about to become
2759 * idle. Attempts to pull tasks from other CPUs.
2761 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
2763 struct sched_domain
*sd
;
2765 for_each_domain(this_cpu
, sd
) {
2766 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
2767 /* If we've pulled tasks over stop searching: */
2768 if (load_balance_newidle(this_cpu
, this_rq
, sd
))
2775 * active_load_balance is run by migration threads. It pushes running tasks
2776 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2777 * running on each physical CPU where possible, and avoids physical /
2778 * logical imbalances.
2780 * Called with busiest_rq locked.
2782 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
2784 int target_cpu
= busiest_rq
->push_cpu
;
2785 struct sched_domain
*sd
;
2786 struct rq
*target_rq
;
2788 /* Is there any task to move? */
2789 if (busiest_rq
->nr_running
<= 1)
2792 target_rq
= cpu_rq(target_cpu
);
2795 * This condition is "impossible", if it occurs
2796 * we need to fix it. Originally reported by
2797 * Bjorn Helgaas on a 128-cpu setup.
2799 BUG_ON(busiest_rq
== target_rq
);
2801 /* move a task from busiest_rq to target_rq */
2802 double_lock_balance(busiest_rq
, target_rq
);
2804 /* Search for an sd spanning us and the target CPU. */
2805 for_each_domain(target_cpu
, sd
) {
2806 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2807 cpu_isset(busiest_cpu
, sd
->span
))
2812 schedstat_inc(sd
, alb_cnt
);
2814 if (move_tasks(target_rq
, target_cpu
, busiest_rq
, 1,
2815 RTPRIO_TO_LOAD_WEIGHT(100), sd
, SCHED_IDLE
,
2817 schedstat_inc(sd
, alb_pushed
);
2819 schedstat_inc(sd
, alb_failed
);
2821 spin_unlock(&target_rq
->lock
);
2825 * rebalance_tick will get called every timer tick, on every CPU.
2827 * It checks each scheduling domain to see if it is due to be balanced,
2828 * and initiates a balancing operation if so.
2830 * Balancing parameters are set up in arch_init_sched_domains.
2833 /* Don't have all balancing operations going off at once: */
2834 static inline unsigned long cpu_offset(int cpu
)
2836 return jiffies
+ cpu
* HZ
/ NR_CPUS
;
2840 rebalance_tick(int this_cpu
, struct rq
*this_rq
, enum idle_type idle
)
2842 unsigned long this_load
, interval
, j
= cpu_offset(this_cpu
);
2843 struct sched_domain
*sd
;
2846 this_load
= this_rq
->raw_weighted_load
;
2848 /* Update our load: */
2849 for (i
= 0, scale
= 1; i
< 3; i
++, scale
<<= 1) {
2850 unsigned long old_load
, new_load
;
2852 old_load
= this_rq
->cpu_load
[i
];
2853 new_load
= this_load
;
2855 * Round up the averaging division if load is increasing. This
2856 * prevents us from getting stuck on 9 if the load is 10, for
2859 if (new_load
> old_load
)
2860 new_load
+= scale
-1;
2861 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) / scale
;
2864 for_each_domain(this_cpu
, sd
) {
2865 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2868 interval
= sd
->balance_interval
;
2869 if (idle
!= SCHED_IDLE
)
2870 interval
*= sd
->busy_factor
;
2872 /* scale ms to jiffies */
2873 interval
= msecs_to_jiffies(interval
);
2874 if (unlikely(!interval
))
2877 if (j
- sd
->last_balance
>= interval
) {
2878 if (load_balance(this_cpu
, this_rq
, sd
, idle
)) {
2880 * We've pulled tasks over so either we're no
2881 * longer idle, or one of our SMT siblings is
2886 sd
->last_balance
+= interval
;
2892 * on UP we do not need to balance between CPUs:
2894 static inline void rebalance_tick(int cpu
, struct rq
*rq
, enum idle_type idle
)
2897 static inline void idle_balance(int cpu
, struct rq
*rq
)
2902 static inline int wake_priority_sleeper(struct rq
*rq
)
2906 #ifdef CONFIG_SCHED_SMT
2907 spin_lock(&rq
->lock
);
2909 * If an SMT sibling task has been put to sleep for priority
2910 * reasons reschedule the idle task to see if it can now run.
2912 if (rq
->nr_running
) {
2913 resched_task(rq
->idle
);
2916 spin_unlock(&rq
->lock
);
2921 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2923 EXPORT_PER_CPU_SYMBOL(kstat
);
2926 * This is called on clock ticks and on context switches.
2927 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2930 update_cpu_clock(struct task_struct
*p
, struct rq
*rq
, unsigned long long now
)
2932 p
->sched_time
+= now
- max(p
->timestamp
, rq
->timestamp_last_tick
);
2936 * Return current->sched_time plus any more ns on the sched_clock
2937 * that have not yet been banked.
2939 unsigned long long current_sched_time(const struct task_struct
*p
)
2941 unsigned long long ns
;
2942 unsigned long flags
;
2944 local_irq_save(flags
);
2945 ns
= max(p
->timestamp
, task_rq(p
)->timestamp_last_tick
);
2946 ns
= p
->sched_time
+ sched_clock() - ns
;
2947 local_irq_restore(flags
);
2953 * We place interactive tasks back into the active array, if possible.
2955 * To guarantee that this does not starve expired tasks we ignore the
2956 * interactivity of a task if the first expired task had to wait more
2957 * than a 'reasonable' amount of time. This deadline timeout is
2958 * load-dependent, as the frequency of array switched decreases with
2959 * increasing number of running tasks. We also ignore the interactivity
2960 * if a better static_prio task has expired:
2962 static inline int expired_starving(struct rq
*rq
)
2964 if (rq
->curr
->static_prio
> rq
->best_expired_prio
)
2966 if (!STARVATION_LIMIT
|| !rq
->expired_timestamp
)
2968 if (jiffies
- rq
->expired_timestamp
> STARVATION_LIMIT
* rq
->nr_running
)
2974 * Account user cpu time to a process.
2975 * @p: the process that the cpu time gets accounted to
2976 * @hardirq_offset: the offset to subtract from hardirq_count()
2977 * @cputime: the cpu time spent in user space since the last update
2979 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
2981 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2984 p
->utime
= cputime_add(p
->utime
, cputime
);
2986 /* Add user time to cpustat. */
2987 tmp
= cputime_to_cputime64(cputime
);
2988 if (TASK_NICE(p
) > 0)
2989 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
2991 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
2995 * Account system cpu time to a process.
2996 * @p: the process that the cpu time gets accounted to
2997 * @hardirq_offset: the offset to subtract from hardirq_count()
2998 * @cputime: the cpu time spent in kernel space since the last update
3000 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3003 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3004 struct rq
*rq
= this_rq();
3007 p
->stime
= cputime_add(p
->stime
, cputime
);
3009 /* Add system time to cpustat. */
3010 tmp
= cputime_to_cputime64(cputime
);
3011 if (hardirq_count() - hardirq_offset
)
3012 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3013 else if (softirq_count())
3014 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3015 else if (p
!= rq
->idle
)
3016 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3017 else if (atomic_read(&rq
->nr_iowait
) > 0)
3018 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3020 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3021 /* Account for system time used */
3022 acct_update_integrals(p
);
3026 * Account for involuntary wait time.
3027 * @p: the process from which the cpu time has been stolen
3028 * @steal: the cpu time spent in involuntary wait
3030 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3032 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3033 cputime64_t tmp
= cputime_to_cputime64(steal
);
3034 struct rq
*rq
= this_rq();
3036 if (p
== rq
->idle
) {
3037 p
->stime
= cputime_add(p
->stime
, steal
);
3038 if (atomic_read(&rq
->nr_iowait
) > 0)
3039 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3041 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3043 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3047 * This function gets called by the timer code, with HZ frequency.
3048 * We call it with interrupts disabled.
3050 * It also gets called by the fork code, when changing the parent's
3053 void scheduler_tick(void)
3055 unsigned long long now
= sched_clock();
3056 struct task_struct
*p
= current
;
3057 int cpu
= smp_processor_id();
3058 struct rq
*rq
= cpu_rq(cpu
);
3060 update_cpu_clock(p
, rq
, now
);
3062 rq
->timestamp_last_tick
= now
;
3064 if (p
== rq
->idle
) {
3065 if (wake_priority_sleeper(rq
))
3067 rebalance_tick(cpu
, rq
, SCHED_IDLE
);
3071 /* Task might have expired already, but not scheduled off yet */
3072 if (p
->array
!= rq
->active
) {
3073 set_tsk_need_resched(p
);
3076 spin_lock(&rq
->lock
);
3078 * The task was running during this tick - update the
3079 * time slice counter. Note: we do not update a thread's
3080 * priority until it either goes to sleep or uses up its
3081 * timeslice. This makes it possible for interactive tasks
3082 * to use up their timeslices at their highest priority levels.
3086 * RR tasks need a special form of timeslice management.
3087 * FIFO tasks have no timeslices.
3089 if ((p
->policy
== SCHED_RR
) && !--p
->time_slice
) {
3090 p
->time_slice
= task_timeslice(p
);
3091 p
->first_time_slice
= 0;
3092 set_tsk_need_resched(p
);
3094 /* put it at the end of the queue: */
3095 requeue_task(p
, rq
->active
);
3099 if (!--p
->time_slice
) {
3100 dequeue_task(p
, rq
->active
);
3101 set_tsk_need_resched(p
);
3102 p
->prio
= effective_prio(p
);
3103 p
->time_slice
= task_timeslice(p
);
3104 p
->first_time_slice
= 0;
3106 if (!rq
->expired_timestamp
)
3107 rq
->expired_timestamp
= jiffies
;
3108 if (!TASK_INTERACTIVE(p
) || expired_starving(rq
)) {
3109 enqueue_task(p
, rq
->expired
);
3110 if (p
->static_prio
< rq
->best_expired_prio
)
3111 rq
->best_expired_prio
= p
->static_prio
;
3113 enqueue_task(p
, rq
->active
);
3116 * Prevent a too long timeslice allowing a task to monopolize
3117 * the CPU. We do this by splitting up the timeslice into
3120 * Note: this does not mean the task's timeslices expire or
3121 * get lost in any way, they just might be preempted by
3122 * another task of equal priority. (one with higher
3123 * priority would have preempted this task already.) We
3124 * requeue this task to the end of the list on this priority
3125 * level, which is in essence a round-robin of tasks with
3128 * This only applies to tasks in the interactive
3129 * delta range with at least TIMESLICE_GRANULARITY to requeue.
3131 if (TASK_INTERACTIVE(p
) && !((task_timeslice(p
) -
3132 p
->time_slice
) % TIMESLICE_GRANULARITY(p
)) &&
3133 (p
->time_slice
>= TIMESLICE_GRANULARITY(p
)) &&
3134 (p
->array
== rq
->active
)) {
3136 requeue_task(p
, rq
->active
);
3137 set_tsk_need_resched(p
);
3141 spin_unlock(&rq
->lock
);
3143 rebalance_tick(cpu
, rq
, NOT_IDLE
);
3146 #ifdef CONFIG_SCHED_SMT
3147 static inline void wakeup_busy_runqueue(struct rq
*rq
)
3149 /* If an SMT runqueue is sleeping due to priority reasons wake it up */
3150 if (rq
->curr
== rq
->idle
&& rq
->nr_running
)
3151 resched_task(rq
->idle
);
3155 * Called with interrupt disabled and this_rq's runqueue locked.
3157 static void wake_sleeping_dependent(int this_cpu
)
3159 struct sched_domain
*tmp
, *sd
= NULL
;
3162 for_each_domain(this_cpu
, tmp
) {
3163 if (tmp
->flags
& SD_SHARE_CPUPOWER
) {
3172 for_each_cpu_mask(i
, sd
->span
) {
3173 struct rq
*smt_rq
= cpu_rq(i
);
3177 if (unlikely(!spin_trylock(&smt_rq
->lock
)))
3180 wakeup_busy_runqueue(smt_rq
);
3181 spin_unlock(&smt_rq
->lock
);
3186 * number of 'lost' timeslices this task wont be able to fully
3187 * utilize, if another task runs on a sibling. This models the
3188 * slowdown effect of other tasks running on siblings:
3190 static inline unsigned long
3191 smt_slice(struct task_struct
*p
, struct sched_domain
*sd
)
3193 return p
->time_slice
* (100 - sd
->per_cpu_gain
) / 100;
3197 * To minimise lock contention and not have to drop this_rq's runlock we only
3198 * trylock the sibling runqueues and bypass those runqueues if we fail to
3199 * acquire their lock. As we only trylock the normal locking order does not
3200 * need to be obeyed.
3203 dependent_sleeper(int this_cpu
, struct rq
*this_rq
, struct task_struct
*p
)
3205 struct sched_domain
*tmp
, *sd
= NULL
;
3208 /* kernel/rt threads do not participate in dependent sleeping */
3209 if (!p
->mm
|| rt_task(p
))
3212 for_each_domain(this_cpu
, tmp
) {
3213 if (tmp
->flags
& SD_SHARE_CPUPOWER
) {
3222 for_each_cpu_mask(i
, sd
->span
) {
3223 struct task_struct
*smt_curr
;
3230 if (unlikely(!spin_trylock(&smt_rq
->lock
)))
3233 smt_curr
= smt_rq
->curr
;
3239 * If a user task with lower static priority than the
3240 * running task on the SMT sibling is trying to schedule,
3241 * delay it till there is proportionately less timeslice
3242 * left of the sibling task to prevent a lower priority
3243 * task from using an unfair proportion of the
3244 * physical cpu's resources. -ck
3246 if (rt_task(smt_curr
)) {
3248 * With real time tasks we run non-rt tasks only
3249 * per_cpu_gain% of the time.
3251 if ((jiffies
% DEF_TIMESLICE
) >
3252 (sd
->per_cpu_gain
* DEF_TIMESLICE
/ 100))
3255 if (smt_curr
->static_prio
< p
->static_prio
&&
3256 !TASK_PREEMPTS_CURR(p
, smt_rq
) &&
3257 smt_slice(smt_curr
, sd
) > task_timeslice(p
))
3261 spin_unlock(&smt_rq
->lock
);
3266 static inline void wake_sleeping_dependent(int this_cpu
)
3270 dependent_sleeper(int this_cpu
, struct rq
*this_rq
, struct task_struct
*p
)
3276 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3278 void fastcall
add_preempt_count(int val
)
3283 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3285 preempt_count() += val
;
3287 * Spinlock count overflowing soon?
3289 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >= PREEMPT_MASK
-10);
3291 EXPORT_SYMBOL(add_preempt_count
);
3293 void fastcall
sub_preempt_count(int val
)
3298 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3301 * Is the spinlock portion underflowing?
3303 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3304 !(preempt_count() & PREEMPT_MASK
)))
3307 preempt_count() -= val
;
3309 EXPORT_SYMBOL(sub_preempt_count
);
3313 static inline int interactive_sleep(enum sleep_type sleep_type
)
3315 return (sleep_type
== SLEEP_INTERACTIVE
||
3316 sleep_type
== SLEEP_INTERRUPTED
);
3320 * schedule() is the main scheduler function.
3322 asmlinkage
void __sched
schedule(void)
3324 struct task_struct
*prev
, *next
;
3325 struct prio_array
*array
;
3326 struct list_head
*queue
;
3327 unsigned long long now
;
3328 unsigned long run_time
;
3329 int cpu
, idx
, new_prio
;
3334 * Test if we are atomic. Since do_exit() needs to call into
3335 * schedule() atomically, we ignore that path for now.
3336 * Otherwise, whine if we are scheduling when we should not be.
3338 if (unlikely(in_atomic() && !current
->exit_state
)) {
3339 printk(KERN_ERR
"BUG: scheduling while atomic: "
3341 current
->comm
, preempt_count(), current
->pid
);
3344 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3349 release_kernel_lock(prev
);
3350 need_resched_nonpreemptible
:
3354 * The idle thread is not allowed to schedule!
3355 * Remove this check after it has been exercised a bit.
3357 if (unlikely(prev
== rq
->idle
) && prev
->state
!= TASK_RUNNING
) {
3358 printk(KERN_ERR
"bad: scheduling from the idle thread!\n");
3362 schedstat_inc(rq
, sched_cnt
);
3363 now
= sched_clock();
3364 if (likely((long long)(now
- prev
->timestamp
) < NS_MAX_SLEEP_AVG
)) {
3365 run_time
= now
- prev
->timestamp
;
3366 if (unlikely((long long)(now
- prev
->timestamp
) < 0))
3369 run_time
= NS_MAX_SLEEP_AVG
;
3372 * Tasks charged proportionately less run_time at high sleep_avg to
3373 * delay them losing their interactive status
3375 run_time
/= (CURRENT_BONUS(prev
) ? : 1);
3377 spin_lock_irq(&rq
->lock
);
3379 switch_count
= &prev
->nivcsw
;
3380 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3381 switch_count
= &prev
->nvcsw
;
3382 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3383 unlikely(signal_pending(prev
))))
3384 prev
->state
= TASK_RUNNING
;
3386 if (prev
->state
== TASK_UNINTERRUPTIBLE
)
3387 rq
->nr_uninterruptible
++;
3388 deactivate_task(prev
, rq
);
3392 cpu
= smp_processor_id();
3393 if (unlikely(!rq
->nr_running
)) {
3394 idle_balance(cpu
, rq
);
3395 if (!rq
->nr_running
) {
3397 rq
->expired_timestamp
= 0;
3398 wake_sleeping_dependent(cpu
);
3404 if (unlikely(!array
->nr_active
)) {
3406 * Switch the active and expired arrays.
3408 schedstat_inc(rq
, sched_switch
);
3409 rq
->active
= rq
->expired
;
3410 rq
->expired
= array
;
3412 rq
->expired_timestamp
= 0;
3413 rq
->best_expired_prio
= MAX_PRIO
;
3416 idx
= sched_find_first_bit(array
->bitmap
);
3417 queue
= array
->queue
+ idx
;
3418 next
= list_entry(queue
->next
, struct task_struct
, run_list
);
3420 if (!rt_task(next
) && interactive_sleep(next
->sleep_type
)) {
3421 unsigned long long delta
= now
- next
->timestamp
;
3422 if (unlikely((long long)(now
- next
->timestamp
) < 0))
3425 if (next
->sleep_type
== SLEEP_INTERACTIVE
)
3426 delta
= delta
* (ON_RUNQUEUE_WEIGHT
* 128 / 100) / 128;
3428 array
= next
->array
;
3429 new_prio
= recalc_task_prio(next
, next
->timestamp
+ delta
);
3431 if (unlikely(next
->prio
!= new_prio
)) {
3432 dequeue_task(next
, array
);
3433 next
->prio
= new_prio
;
3434 enqueue_task(next
, array
);
3437 next
->sleep_type
= SLEEP_NORMAL
;
3438 if (dependent_sleeper(cpu
, rq
, next
))
3441 if (next
== rq
->idle
)
3442 schedstat_inc(rq
, sched_goidle
);
3444 prefetch_stack(next
);
3445 clear_tsk_need_resched(prev
);
3446 rcu_qsctr_inc(task_cpu(prev
));
3448 update_cpu_clock(prev
, rq
, now
);
3450 prev
->sleep_avg
-= run_time
;
3451 if ((long)prev
->sleep_avg
<= 0)
3452 prev
->sleep_avg
= 0;
3453 prev
->timestamp
= prev
->last_ran
= now
;
3455 sched_info_switch(prev
, next
);
3456 if (likely(prev
!= next
)) {
3457 next
->timestamp
= now
;
3462 prepare_task_switch(rq
, next
);
3463 prev
= context_switch(rq
, prev
, next
);
3466 * this_rq must be evaluated again because prev may have moved
3467 * CPUs since it called schedule(), thus the 'rq' on its stack
3468 * frame will be invalid.
3470 finish_task_switch(this_rq(), prev
);
3472 spin_unlock_irq(&rq
->lock
);
3475 if (unlikely(reacquire_kernel_lock(prev
) < 0))
3476 goto need_resched_nonpreemptible
;
3477 preempt_enable_no_resched();
3478 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3481 EXPORT_SYMBOL(schedule
);
3483 #ifdef CONFIG_PREEMPT
3485 * this is the entry point to schedule() from in-kernel preemption
3486 * off of preempt_enable. Kernel preemptions off return from interrupt
3487 * occur there and call schedule directly.
3489 asmlinkage
void __sched
preempt_schedule(void)
3491 struct thread_info
*ti
= current_thread_info();
3492 #ifdef CONFIG_PREEMPT_BKL
3493 struct task_struct
*task
= current
;
3494 int saved_lock_depth
;
3497 * If there is a non-zero preempt_count or interrupts are disabled,
3498 * we do not want to preempt the current task. Just return..
3500 if (likely(ti
->preempt_count
|| irqs_disabled()))
3504 add_preempt_count(PREEMPT_ACTIVE
);
3506 * We keep the big kernel semaphore locked, but we
3507 * clear ->lock_depth so that schedule() doesnt
3508 * auto-release the semaphore:
3510 #ifdef CONFIG_PREEMPT_BKL
3511 saved_lock_depth
= task
->lock_depth
;
3512 task
->lock_depth
= -1;
3515 #ifdef CONFIG_PREEMPT_BKL
3516 task
->lock_depth
= saved_lock_depth
;
3518 sub_preempt_count(PREEMPT_ACTIVE
);
3520 /* we could miss a preemption opportunity between schedule and now */
3522 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3525 EXPORT_SYMBOL(preempt_schedule
);
3528 * this is the entry point to schedule() from kernel preemption
3529 * off of irq context.
3530 * Note, that this is called and return with irqs disabled. This will
3531 * protect us against recursive calling from irq.
3533 asmlinkage
void __sched
preempt_schedule_irq(void)
3535 struct thread_info
*ti
= current_thread_info();
3536 #ifdef CONFIG_PREEMPT_BKL
3537 struct task_struct
*task
= current
;
3538 int saved_lock_depth
;
3540 /* Catch callers which need to be fixed */
3541 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3544 add_preempt_count(PREEMPT_ACTIVE
);
3546 * We keep the big kernel semaphore locked, but we
3547 * clear ->lock_depth so that schedule() doesnt
3548 * auto-release the semaphore:
3550 #ifdef CONFIG_PREEMPT_BKL
3551 saved_lock_depth
= task
->lock_depth
;
3552 task
->lock_depth
= -1;
3556 local_irq_disable();
3557 #ifdef CONFIG_PREEMPT_BKL
3558 task
->lock_depth
= saved_lock_depth
;
3560 sub_preempt_count(PREEMPT_ACTIVE
);
3562 /* we could miss a preemption opportunity between schedule and now */
3564 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3568 #endif /* CONFIG_PREEMPT */
3570 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3573 return try_to_wake_up(curr
->private, mode
, sync
);
3575 EXPORT_SYMBOL(default_wake_function
);
3578 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3579 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3580 * number) then we wake all the non-exclusive tasks and one exclusive task.
3582 * There are circumstances in which we can try to wake a task which has already
3583 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3584 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3586 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3587 int nr_exclusive
, int sync
, void *key
)
3589 struct list_head
*tmp
, *next
;
3591 list_for_each_safe(tmp
, next
, &q
->task_list
) {
3592 wait_queue_t
*curr
= list_entry(tmp
, wait_queue_t
, task_list
);
3593 unsigned flags
= curr
->flags
;
3595 if (curr
->func(curr
, mode
, sync
, key
) &&
3596 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3602 * __wake_up - wake up threads blocked on a waitqueue.
3604 * @mode: which threads
3605 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3606 * @key: is directly passed to the wakeup function
3608 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3609 int nr_exclusive
, void *key
)
3611 unsigned long flags
;
3613 spin_lock_irqsave(&q
->lock
, flags
);
3614 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3615 spin_unlock_irqrestore(&q
->lock
, flags
);
3617 EXPORT_SYMBOL(__wake_up
);
3620 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3622 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3624 __wake_up_common(q
, mode
, 1, 0, NULL
);
3628 * __wake_up_sync - wake up threads blocked on a waitqueue.
3630 * @mode: which threads
3631 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3633 * The sync wakeup differs that the waker knows that it will schedule
3634 * away soon, so while the target thread will be woken up, it will not
3635 * be migrated to another CPU - ie. the two threads are 'synchronized'
3636 * with each other. This can prevent needless bouncing between CPUs.
3638 * On UP it can prevent extra preemption.
3641 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3643 unsigned long flags
;
3649 if (unlikely(!nr_exclusive
))
3652 spin_lock_irqsave(&q
->lock
, flags
);
3653 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3654 spin_unlock_irqrestore(&q
->lock
, flags
);
3656 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3658 void fastcall
complete(struct completion
*x
)
3660 unsigned long flags
;
3662 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3664 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3666 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3668 EXPORT_SYMBOL(complete
);
3670 void fastcall
complete_all(struct completion
*x
)
3672 unsigned long flags
;
3674 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3675 x
->done
+= UINT_MAX
/2;
3676 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3678 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3680 EXPORT_SYMBOL(complete_all
);
3682 void fastcall __sched
wait_for_completion(struct completion
*x
)
3686 spin_lock_irq(&x
->wait
.lock
);
3688 DECLARE_WAITQUEUE(wait
, current
);
3690 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3691 __add_wait_queue_tail(&x
->wait
, &wait
);
3693 __set_current_state(TASK_UNINTERRUPTIBLE
);
3694 spin_unlock_irq(&x
->wait
.lock
);
3696 spin_lock_irq(&x
->wait
.lock
);
3698 __remove_wait_queue(&x
->wait
, &wait
);
3701 spin_unlock_irq(&x
->wait
.lock
);
3703 EXPORT_SYMBOL(wait_for_completion
);
3705 unsigned long fastcall __sched
3706 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3710 spin_lock_irq(&x
->wait
.lock
);
3712 DECLARE_WAITQUEUE(wait
, current
);
3714 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3715 __add_wait_queue_tail(&x
->wait
, &wait
);
3717 __set_current_state(TASK_UNINTERRUPTIBLE
);
3718 spin_unlock_irq(&x
->wait
.lock
);
3719 timeout
= schedule_timeout(timeout
);
3720 spin_lock_irq(&x
->wait
.lock
);
3722 __remove_wait_queue(&x
->wait
, &wait
);
3726 __remove_wait_queue(&x
->wait
, &wait
);
3730 spin_unlock_irq(&x
->wait
.lock
);
3733 EXPORT_SYMBOL(wait_for_completion_timeout
);
3735 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3741 spin_lock_irq(&x
->wait
.lock
);
3743 DECLARE_WAITQUEUE(wait
, current
);
3745 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3746 __add_wait_queue_tail(&x
->wait
, &wait
);
3748 if (signal_pending(current
)) {
3750 __remove_wait_queue(&x
->wait
, &wait
);
3753 __set_current_state(TASK_INTERRUPTIBLE
);
3754 spin_unlock_irq(&x
->wait
.lock
);
3756 spin_lock_irq(&x
->wait
.lock
);
3758 __remove_wait_queue(&x
->wait
, &wait
);
3762 spin_unlock_irq(&x
->wait
.lock
);
3766 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3768 unsigned long fastcall __sched
3769 wait_for_completion_interruptible_timeout(struct completion
*x
,
3770 unsigned long timeout
)
3774 spin_lock_irq(&x
->wait
.lock
);
3776 DECLARE_WAITQUEUE(wait
, current
);
3778 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3779 __add_wait_queue_tail(&x
->wait
, &wait
);
3781 if (signal_pending(current
)) {
3782 timeout
= -ERESTARTSYS
;
3783 __remove_wait_queue(&x
->wait
, &wait
);
3786 __set_current_state(TASK_INTERRUPTIBLE
);
3787 spin_unlock_irq(&x
->wait
.lock
);
3788 timeout
= schedule_timeout(timeout
);
3789 spin_lock_irq(&x
->wait
.lock
);
3791 __remove_wait_queue(&x
->wait
, &wait
);
3795 __remove_wait_queue(&x
->wait
, &wait
);
3799 spin_unlock_irq(&x
->wait
.lock
);
3802 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3805 #define SLEEP_ON_VAR \
3806 unsigned long flags; \
3807 wait_queue_t wait; \
3808 init_waitqueue_entry(&wait, current);
3810 #define SLEEP_ON_HEAD \
3811 spin_lock_irqsave(&q->lock,flags); \
3812 __add_wait_queue(q, &wait); \
3813 spin_unlock(&q->lock);
3815 #define SLEEP_ON_TAIL \
3816 spin_lock_irq(&q->lock); \
3817 __remove_wait_queue(q, &wait); \
3818 spin_unlock_irqrestore(&q->lock, flags);
3820 void fastcall __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3824 current
->state
= TASK_INTERRUPTIBLE
;
3830 EXPORT_SYMBOL(interruptible_sleep_on
);
3832 long fastcall __sched
3833 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3837 current
->state
= TASK_INTERRUPTIBLE
;
3840 timeout
= schedule_timeout(timeout
);
3845 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3847 void fastcall __sched
sleep_on(wait_queue_head_t
*q
)
3851 current
->state
= TASK_UNINTERRUPTIBLE
;
3857 EXPORT_SYMBOL(sleep_on
);
3859 long fastcall __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3863 current
->state
= TASK_UNINTERRUPTIBLE
;
3866 timeout
= schedule_timeout(timeout
);
3872 EXPORT_SYMBOL(sleep_on_timeout
);
3874 #ifdef CONFIG_RT_MUTEXES
3877 * rt_mutex_setprio - set the current priority of a task
3879 * @prio: prio value (kernel-internal form)
3881 * This function changes the 'effective' priority of a task. It does
3882 * not touch ->normal_prio like __setscheduler().
3884 * Used by the rt_mutex code to implement priority inheritance logic.
3886 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3888 struct prio_array
*array
;
3889 unsigned long flags
;
3893 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3895 rq
= task_rq_lock(p
, &flags
);
3900 dequeue_task(p
, array
);
3905 * If changing to an RT priority then queue it
3906 * in the active array!
3910 enqueue_task(p
, array
);
3912 * Reschedule if we are currently running on this runqueue and
3913 * our priority decreased, or if we are not currently running on
3914 * this runqueue and our priority is higher than the current's
3916 if (task_running(rq
, p
)) {
3917 if (p
->prio
> oldprio
)
3918 resched_task(rq
->curr
);
3919 } else if (TASK_PREEMPTS_CURR(p
, rq
))
3920 resched_task(rq
->curr
);
3922 task_rq_unlock(rq
, &flags
);
3927 void set_user_nice(struct task_struct
*p
, long nice
)
3929 struct prio_array
*array
;
3930 int old_prio
, delta
;
3931 unsigned long flags
;
3934 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3937 * We have to be careful, if called from sys_setpriority(),
3938 * the task might be in the middle of scheduling on another CPU.
3940 rq
= task_rq_lock(p
, &flags
);
3942 * The RT priorities are set via sched_setscheduler(), but we still
3943 * allow the 'normal' nice value to be set - but as expected
3944 * it wont have any effect on scheduling until the task is
3945 * not SCHED_NORMAL/SCHED_BATCH:
3947 if (has_rt_policy(p
)) {
3948 p
->static_prio
= NICE_TO_PRIO(nice
);
3953 dequeue_task(p
, array
);
3954 dec_raw_weighted_load(rq
, p
);
3957 p
->static_prio
= NICE_TO_PRIO(nice
);
3960 p
->prio
= effective_prio(p
);
3961 delta
= p
->prio
- old_prio
;
3964 enqueue_task(p
, array
);
3965 inc_raw_weighted_load(rq
, p
);
3967 * If the task increased its priority or is running and
3968 * lowered its priority, then reschedule its CPU:
3970 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3971 resched_task(rq
->curr
);
3974 task_rq_unlock(rq
, &flags
);
3976 EXPORT_SYMBOL(set_user_nice
);
3979 * can_nice - check if a task can reduce its nice value
3983 int can_nice(const struct task_struct
*p
, const int nice
)
3985 /* convert nice value [19,-20] to rlimit style value [1,40] */
3986 int nice_rlim
= 20 - nice
;
3988 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
3989 capable(CAP_SYS_NICE
));
3992 #ifdef __ARCH_WANT_SYS_NICE
3995 * sys_nice - change the priority of the current process.
3996 * @increment: priority increment
3998 * sys_setpriority is a more generic, but much slower function that
3999 * does similar things.
4001 asmlinkage
long sys_nice(int increment
)
4006 * Setpriority might change our priority at the same moment.
4007 * We don't have to worry. Conceptually one call occurs first
4008 * and we have a single winner.
4010 if (increment
< -40)
4015 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4021 if (increment
< 0 && !can_nice(current
, nice
))
4024 retval
= security_task_setnice(current
, nice
);
4028 set_user_nice(current
, nice
);
4035 * task_prio - return the priority value of a given task.
4036 * @p: the task in question.
4038 * This is the priority value as seen by users in /proc.
4039 * RT tasks are offset by -200. Normal tasks are centered
4040 * around 0, value goes from -16 to +15.
4042 int task_prio(const struct task_struct
*p
)
4044 return p
->prio
- MAX_RT_PRIO
;
4048 * task_nice - return the nice value of a given task.
4049 * @p: the task in question.
4051 int task_nice(const struct task_struct
*p
)
4053 return TASK_NICE(p
);
4055 EXPORT_SYMBOL_GPL(task_nice
);
4058 * idle_cpu - is a given cpu idle currently?
4059 * @cpu: the processor in question.
4061 int idle_cpu(int cpu
)
4063 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4067 * idle_task - return the idle task for a given cpu.
4068 * @cpu: the processor in question.
4070 struct task_struct
*idle_task(int cpu
)
4072 return cpu_rq(cpu
)->idle
;
4076 * find_process_by_pid - find a process with a matching PID value.
4077 * @pid: the pid in question.
4079 static inline struct task_struct
*find_process_by_pid(pid_t pid
)
4081 return pid
? find_task_by_pid(pid
) : current
;
4084 /* Actually do priority change: must hold rq lock. */
4085 static void __setscheduler(struct task_struct
*p
, int policy
, int prio
)
4090 p
->rt_priority
= prio
;
4091 p
->normal_prio
= normal_prio(p
);
4092 /* we are holding p->pi_lock already */
4093 p
->prio
= rt_mutex_getprio(p
);
4095 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
4097 if (policy
== SCHED_BATCH
)
4103 * sched_setscheduler - change the scheduling policy and/or RT priority of
4105 * @p: the task in question.
4106 * @policy: new policy.
4107 * @param: structure containing the new RT priority.
4109 * NOTE: the task may be already dead
4111 int sched_setscheduler(struct task_struct
*p
, int policy
,
4112 struct sched_param
*param
)
4114 int retval
, oldprio
, oldpolicy
= -1;
4115 struct prio_array
*array
;
4116 unsigned long flags
;
4119 /* may grab non-irq protected spin_locks */
4120 BUG_ON(in_interrupt());
4122 /* double check policy once rq lock held */
4124 policy
= oldpolicy
= p
->policy
;
4125 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4126 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
)
4129 * Valid priorities for SCHED_FIFO and SCHED_RR are
4130 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
4133 if (param
->sched_priority
< 0 ||
4134 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4135 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4137 if (is_rt_policy(policy
) != (param
->sched_priority
!= 0))
4141 * Allow unprivileged RT tasks to decrease priority:
4143 if (!capable(CAP_SYS_NICE
)) {
4144 if (is_rt_policy(policy
)) {
4145 unsigned long rlim_rtprio
;
4146 unsigned long flags
;
4148 if (!lock_task_sighand(p
, &flags
))
4150 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4151 unlock_task_sighand(p
, &flags
);
4153 /* can't set/change the rt policy */
4154 if (policy
!= p
->policy
&& !rlim_rtprio
)
4157 /* can't increase priority */
4158 if (param
->sched_priority
> p
->rt_priority
&&
4159 param
->sched_priority
> rlim_rtprio
)
4163 /* can't change other user's priorities */
4164 if ((current
->euid
!= p
->euid
) &&
4165 (current
->euid
!= p
->uid
))
4169 retval
= security_task_setscheduler(p
, policy
, param
);
4173 * make sure no PI-waiters arrive (or leave) while we are
4174 * changing the priority of the task:
4176 spin_lock_irqsave(&p
->pi_lock
, flags
);
4178 * To be able to change p->policy safely, the apropriate
4179 * runqueue lock must be held.
4181 rq
= __task_rq_lock(p
);
4182 /* recheck policy now with rq lock held */
4183 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4184 policy
= oldpolicy
= -1;
4185 __task_rq_unlock(rq
);
4186 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4191 deactivate_task(p
, rq
);
4193 __setscheduler(p
, policy
, param
->sched_priority
);
4195 __activate_task(p
, rq
);
4197 * Reschedule if we are currently running on this runqueue and
4198 * our priority decreased, or if we are not currently running on
4199 * this runqueue and our priority is higher than the current's
4201 if (task_running(rq
, p
)) {
4202 if (p
->prio
> oldprio
)
4203 resched_task(rq
->curr
);
4204 } else if (TASK_PREEMPTS_CURR(p
, rq
))
4205 resched_task(rq
->curr
);
4207 __task_rq_unlock(rq
);
4208 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4210 rt_mutex_adjust_pi(p
);
4214 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4217 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4219 struct sched_param lparam
;
4220 struct task_struct
*p
;
4223 if (!param
|| pid
< 0)
4225 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4230 p
= find_process_by_pid(pid
);
4232 retval
= sched_setscheduler(p
, policy
, &lparam
);
4239 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4240 * @pid: the pid in question.
4241 * @policy: new policy.
4242 * @param: structure containing the new RT priority.
4244 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
4245 struct sched_param __user
*param
)
4247 /* negative values for policy are not valid */
4251 return do_sched_setscheduler(pid
, policy
, param
);
4255 * sys_sched_setparam - set/change the RT priority of a thread
4256 * @pid: the pid in question.
4257 * @param: structure containing the new RT priority.
4259 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4261 return do_sched_setscheduler(pid
, -1, param
);
4265 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4266 * @pid: the pid in question.
4268 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4270 struct task_struct
*p
;
4271 int retval
= -EINVAL
;
4277 read_lock(&tasklist_lock
);
4278 p
= find_process_by_pid(pid
);
4280 retval
= security_task_getscheduler(p
);
4284 read_unlock(&tasklist_lock
);
4291 * sys_sched_getscheduler - get the RT priority of a thread
4292 * @pid: the pid in question.
4293 * @param: structure containing the RT priority.
4295 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4297 struct sched_param lp
;
4298 struct task_struct
*p
;
4299 int retval
= -EINVAL
;
4301 if (!param
|| pid
< 0)
4304 read_lock(&tasklist_lock
);
4305 p
= find_process_by_pid(pid
);
4310 retval
= security_task_getscheduler(p
);
4314 lp
.sched_priority
= p
->rt_priority
;
4315 read_unlock(&tasklist_lock
);
4318 * This one might sleep, we cannot do it with a spinlock held ...
4320 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4326 read_unlock(&tasklist_lock
);
4330 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4332 cpumask_t cpus_allowed
;
4333 struct task_struct
*p
;
4337 read_lock(&tasklist_lock
);
4339 p
= find_process_by_pid(pid
);
4341 read_unlock(&tasklist_lock
);
4342 unlock_cpu_hotplug();
4347 * It is not safe to call set_cpus_allowed with the
4348 * tasklist_lock held. We will bump the task_struct's
4349 * usage count and then drop tasklist_lock.
4352 read_unlock(&tasklist_lock
);
4355 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4356 !capable(CAP_SYS_NICE
))
4359 retval
= security_task_setscheduler(p
, 0, NULL
);
4363 cpus_allowed
= cpuset_cpus_allowed(p
);
4364 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4365 retval
= set_cpus_allowed(p
, new_mask
);
4369 unlock_cpu_hotplug();
4373 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4374 cpumask_t
*new_mask
)
4376 if (len
< sizeof(cpumask_t
)) {
4377 memset(new_mask
, 0, sizeof(cpumask_t
));
4378 } else if (len
> sizeof(cpumask_t
)) {
4379 len
= sizeof(cpumask_t
);
4381 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4385 * sys_sched_setaffinity - set the cpu affinity of a process
4386 * @pid: pid of the process
4387 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4388 * @user_mask_ptr: user-space pointer to the new cpu mask
4390 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4391 unsigned long __user
*user_mask_ptr
)
4396 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4400 return sched_setaffinity(pid
, new_mask
);
4404 * Represents all cpu's present in the system
4405 * In systems capable of hotplug, this map could dynamically grow
4406 * as new cpu's are detected in the system via any platform specific
4407 * method, such as ACPI for e.g.
4410 cpumask_t cpu_present_map __read_mostly
;
4411 EXPORT_SYMBOL(cpu_present_map
);
4414 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4415 EXPORT_SYMBOL(cpu_online_map
);
4417 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4418 EXPORT_SYMBOL(cpu_possible_map
);
4421 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4423 struct task_struct
*p
;
4427 read_lock(&tasklist_lock
);
4430 p
= find_process_by_pid(pid
);
4434 retval
= security_task_getscheduler(p
);
4438 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4441 read_unlock(&tasklist_lock
);
4442 unlock_cpu_hotplug();
4450 * sys_sched_getaffinity - get the cpu affinity of a process
4451 * @pid: pid of the process
4452 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4453 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4455 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4456 unsigned long __user
*user_mask_ptr
)
4461 if (len
< sizeof(cpumask_t
))
4464 ret
= sched_getaffinity(pid
, &mask
);
4468 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4471 return sizeof(cpumask_t
);
4475 * sys_sched_yield - yield the current processor to other threads.
4477 * this function yields the current CPU by moving the calling thread
4478 * to the expired array. If there are no other threads running on this
4479 * CPU then this function will return.
4481 asmlinkage
long sys_sched_yield(void)
4483 struct rq
*rq
= this_rq_lock();
4484 struct prio_array
*array
= current
->array
, *target
= rq
->expired
;
4486 schedstat_inc(rq
, yld_cnt
);
4488 * We implement yielding by moving the task into the expired
4491 * (special rule: RT tasks will just roundrobin in the active
4494 if (rt_task(current
))
4495 target
= rq
->active
;
4497 if (array
->nr_active
== 1) {
4498 schedstat_inc(rq
, yld_act_empty
);
4499 if (!rq
->expired
->nr_active
)
4500 schedstat_inc(rq
, yld_both_empty
);
4501 } else if (!rq
->expired
->nr_active
)
4502 schedstat_inc(rq
, yld_exp_empty
);
4504 if (array
!= target
) {
4505 dequeue_task(current
, array
);
4506 enqueue_task(current
, target
);
4509 * requeue_task is cheaper so perform that if possible.
4511 requeue_task(current
, array
);
4514 * Since we are going to call schedule() anyway, there's
4515 * no need to preempt or enable interrupts:
4517 __release(rq
->lock
);
4518 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4519 _raw_spin_unlock(&rq
->lock
);
4520 preempt_enable_no_resched();
4527 static void __cond_resched(void)
4529 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4530 __might_sleep(__FILE__
, __LINE__
);
4533 * The BKS might be reacquired before we have dropped
4534 * PREEMPT_ACTIVE, which could trigger a second
4535 * cond_resched() call.
4538 add_preempt_count(PREEMPT_ACTIVE
);
4540 sub_preempt_count(PREEMPT_ACTIVE
);
4541 } while (need_resched());
4544 int __sched
cond_resched(void)
4546 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4547 system_state
== SYSTEM_RUNNING
) {
4553 EXPORT_SYMBOL(cond_resched
);
4556 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4557 * call schedule, and on return reacquire the lock.
4559 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4560 * operations here to prevent schedule() from being called twice (once via
4561 * spin_unlock(), once by hand).
4563 int cond_resched_lock(spinlock_t
*lock
)
4567 if (need_lockbreak(lock
)) {
4573 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4574 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4575 _raw_spin_unlock(lock
);
4576 preempt_enable_no_resched();
4583 EXPORT_SYMBOL(cond_resched_lock
);
4585 int __sched
cond_resched_softirq(void)
4587 BUG_ON(!in_softirq());
4589 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4590 raw_local_irq_disable();
4592 raw_local_irq_enable();
4599 EXPORT_SYMBOL(cond_resched_softirq
);
4602 * yield - yield the current processor to other threads.
4604 * this is a shortcut for kernel-space yielding - it marks the
4605 * thread runnable and calls sys_sched_yield().
4607 void __sched
yield(void)
4609 set_current_state(TASK_RUNNING
);
4612 EXPORT_SYMBOL(yield
);
4615 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4616 * that process accounting knows that this is a task in IO wait state.
4618 * But don't do that if it is a deliberate, throttling IO wait (this task
4619 * has set its backing_dev_info: the queue against which it should throttle)
4621 void __sched
io_schedule(void)
4623 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4625 delayacct_blkio_start();
4626 atomic_inc(&rq
->nr_iowait
);
4628 atomic_dec(&rq
->nr_iowait
);
4629 delayacct_blkio_end();
4631 EXPORT_SYMBOL(io_schedule
);
4633 long __sched
io_schedule_timeout(long timeout
)
4635 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4638 delayacct_blkio_start();
4639 atomic_inc(&rq
->nr_iowait
);
4640 ret
= schedule_timeout(timeout
);
4641 atomic_dec(&rq
->nr_iowait
);
4642 delayacct_blkio_end();
4647 * sys_sched_get_priority_max - return maximum RT priority.
4648 * @policy: scheduling class.
4650 * this syscall returns the maximum rt_priority that can be used
4651 * by a given scheduling class.
4653 asmlinkage
long sys_sched_get_priority_max(int policy
)
4660 ret
= MAX_USER_RT_PRIO
-1;
4671 * sys_sched_get_priority_min - return minimum RT priority.
4672 * @policy: scheduling class.
4674 * this syscall returns the minimum rt_priority that can be used
4675 * by a given scheduling class.
4677 asmlinkage
long sys_sched_get_priority_min(int policy
)
4694 * sys_sched_rr_get_interval - return the default timeslice of a process.
4695 * @pid: pid of the process.
4696 * @interval: userspace pointer to the timeslice value.
4698 * this syscall writes the default timeslice value of a given process
4699 * into the user-space timespec buffer. A value of '0' means infinity.
4702 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4704 struct task_struct
*p
;
4705 int retval
= -EINVAL
;
4712 read_lock(&tasklist_lock
);
4713 p
= find_process_by_pid(pid
);
4717 retval
= security_task_getscheduler(p
);
4721 jiffies_to_timespec(p
->policy
== SCHED_FIFO
?
4722 0 : task_timeslice(p
), &t
);
4723 read_unlock(&tasklist_lock
);
4724 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4728 read_unlock(&tasklist_lock
);
4732 static inline struct task_struct
*eldest_child(struct task_struct
*p
)
4734 if (list_empty(&p
->children
))
4736 return list_entry(p
->children
.next
,struct task_struct
,sibling
);
4739 static inline struct task_struct
*older_sibling(struct task_struct
*p
)
4741 if (p
->sibling
.prev
==&p
->parent
->children
)
4743 return list_entry(p
->sibling
.prev
,struct task_struct
,sibling
);
4746 static inline struct task_struct
*younger_sibling(struct task_struct
*p
)
4748 if (p
->sibling
.next
==&p
->parent
->children
)
4750 return list_entry(p
->sibling
.next
,struct task_struct
,sibling
);
4753 static const char stat_nam
[] = "RSDTtZX";
4755 static void show_task(struct task_struct
*p
)
4757 struct task_struct
*relative
;
4758 unsigned long free
= 0;
4761 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4762 printk("%-13.13s %c", p
->comm
,
4763 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4764 #if (BITS_PER_LONG == 32)
4765 if (state
== TASK_RUNNING
)
4766 printk(" running ");
4768 printk(" %08lX ", thread_saved_pc(p
));
4770 if (state
== TASK_RUNNING
)
4771 printk(" running task ");
4773 printk(" %016lx ", thread_saved_pc(p
));
4775 #ifdef CONFIG_DEBUG_STACK_USAGE
4777 unsigned long *n
= end_of_stack(p
);
4780 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4783 printk("%5lu %5d %6d ", free
, p
->pid
, p
->parent
->pid
);
4784 if ((relative
= eldest_child(p
)))
4785 printk("%5d ", relative
->pid
);
4788 if ((relative
= younger_sibling(p
)))
4789 printk("%7d", relative
->pid
);
4792 if ((relative
= older_sibling(p
)))
4793 printk(" %5d", relative
->pid
);
4797 printk(" (L-TLB)\n");
4799 printk(" (NOTLB)\n");
4801 if (state
!= TASK_RUNNING
)
4802 show_stack(p
, NULL
);
4805 void show_state(void)
4807 struct task_struct
*g
, *p
;
4809 #if (BITS_PER_LONG == 32)
4812 printk(" task PC pid father child younger older\n");
4816 printk(" task PC pid father child younger older\n");
4818 read_lock(&tasklist_lock
);
4819 do_each_thread(g
, p
) {
4821 * reset the NMI-timeout, listing all files on a slow
4822 * console might take alot of time:
4824 touch_nmi_watchdog();
4826 } while_each_thread(g
, p
);
4828 read_unlock(&tasklist_lock
);
4829 debug_show_all_locks();
4833 * init_idle - set up an idle thread for a given CPU
4834 * @idle: task in question
4835 * @cpu: cpu the idle task belongs to
4837 * NOTE: this function does not set the idle thread's NEED_RESCHED
4838 * flag, to make booting more robust.
4840 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4842 struct rq
*rq
= cpu_rq(cpu
);
4843 unsigned long flags
;
4845 idle
->timestamp
= sched_clock();
4846 idle
->sleep_avg
= 0;
4848 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
4849 idle
->state
= TASK_RUNNING
;
4850 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4851 set_task_cpu(idle
, cpu
);
4853 spin_lock_irqsave(&rq
->lock
, flags
);
4854 rq
->curr
= rq
->idle
= idle
;
4855 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4858 spin_unlock_irqrestore(&rq
->lock
, flags
);
4860 /* Set the preempt count _outside_ the spinlocks! */
4861 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4862 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4864 task_thread_info(idle
)->preempt_count
= 0;
4869 * In a system that switches off the HZ timer nohz_cpu_mask
4870 * indicates which cpus entered this state. This is used
4871 * in the rcu update to wait only for active cpus. For system
4872 * which do not switch off the HZ timer nohz_cpu_mask should
4873 * always be CPU_MASK_NONE.
4875 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4879 * This is how migration works:
4881 * 1) we queue a struct migration_req structure in the source CPU's
4882 * runqueue and wake up that CPU's migration thread.
4883 * 2) we down() the locked semaphore => thread blocks.
4884 * 3) migration thread wakes up (implicitly it forces the migrated
4885 * thread off the CPU)
4886 * 4) it gets the migration request and checks whether the migrated
4887 * task is still in the wrong runqueue.
4888 * 5) if it's in the wrong runqueue then the migration thread removes
4889 * it and puts it into the right queue.
4890 * 6) migration thread up()s the semaphore.
4891 * 7) we wake up and the migration is done.
4895 * Change a given task's CPU affinity. Migrate the thread to a
4896 * proper CPU and schedule it away if the CPU it's executing on
4897 * is removed from the allowed bitmask.
4899 * NOTE: the caller must have a valid reference to the task, the
4900 * task must not exit() & deallocate itself prematurely. The
4901 * call is not atomic; no spinlocks may be held.
4903 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
4905 struct migration_req req
;
4906 unsigned long flags
;
4910 rq
= task_rq_lock(p
, &flags
);
4911 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4916 p
->cpus_allowed
= new_mask
;
4917 /* Can the task run on the task's current CPU? If so, we're done */
4918 if (cpu_isset(task_cpu(p
), new_mask
))
4921 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4922 /* Need help from migration thread: drop lock and wait. */
4923 task_rq_unlock(rq
, &flags
);
4924 wake_up_process(rq
->migration_thread
);
4925 wait_for_completion(&req
.done
);
4926 tlb_migrate_finish(p
->mm
);
4930 task_rq_unlock(rq
, &flags
);
4934 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4937 * Move (not current) task off this cpu, onto dest cpu. We're doing
4938 * this because either it can't run here any more (set_cpus_allowed()
4939 * away from this CPU, or CPU going down), or because we're
4940 * attempting to rebalance this task on exec (sched_exec).
4942 * So we race with normal scheduler movements, but that's OK, as long
4943 * as the task is no longer on this CPU.
4945 * Returns non-zero if task was successfully migrated.
4947 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4949 struct rq
*rq_dest
, *rq_src
;
4952 if (unlikely(cpu_is_offline(dest_cpu
)))
4955 rq_src
= cpu_rq(src_cpu
);
4956 rq_dest
= cpu_rq(dest_cpu
);
4958 double_rq_lock(rq_src
, rq_dest
);
4959 /* Already moved. */
4960 if (task_cpu(p
) != src_cpu
)
4962 /* Affinity changed (again). */
4963 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
4966 set_task_cpu(p
, dest_cpu
);
4969 * Sync timestamp with rq_dest's before activating.
4970 * The same thing could be achieved by doing this step
4971 * afterwards, and pretending it was a local activate.
4972 * This way is cleaner and logically correct.
4974 p
->timestamp
= p
->timestamp
- rq_src
->timestamp_last_tick
4975 + rq_dest
->timestamp_last_tick
;
4976 deactivate_task(p
, rq_src
);
4977 __activate_task(p
, rq_dest
);
4978 if (TASK_PREEMPTS_CURR(p
, rq_dest
))
4979 resched_task(rq_dest
->curr
);
4983 double_rq_unlock(rq_src
, rq_dest
);
4988 * migration_thread - this is a highprio system thread that performs
4989 * thread migration by bumping thread off CPU then 'pushing' onto
4992 static int migration_thread(void *data
)
4994 int cpu
= (long)data
;
4998 BUG_ON(rq
->migration_thread
!= current
);
5000 set_current_state(TASK_INTERRUPTIBLE
);
5001 while (!kthread_should_stop()) {
5002 struct migration_req
*req
;
5003 struct list_head
*head
;
5007 spin_lock_irq(&rq
->lock
);
5009 if (cpu_is_offline(cpu
)) {
5010 spin_unlock_irq(&rq
->lock
);
5014 if (rq
->active_balance
) {
5015 active_load_balance(rq
, cpu
);
5016 rq
->active_balance
= 0;
5019 head
= &rq
->migration_queue
;
5021 if (list_empty(head
)) {
5022 spin_unlock_irq(&rq
->lock
);
5024 set_current_state(TASK_INTERRUPTIBLE
);
5027 req
= list_entry(head
->next
, struct migration_req
, list
);
5028 list_del_init(head
->next
);
5030 spin_unlock(&rq
->lock
);
5031 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5034 complete(&req
->done
);
5036 __set_current_state(TASK_RUNNING
);
5040 /* Wait for kthread_stop */
5041 set_current_state(TASK_INTERRUPTIBLE
);
5042 while (!kthread_should_stop()) {
5044 set_current_state(TASK_INTERRUPTIBLE
);
5046 __set_current_state(TASK_RUNNING
);
5050 #ifdef CONFIG_HOTPLUG_CPU
5052 * Figure out where task on dead CPU should go, use force if neccessary.
5053 * NOTE: interrupts should be disabled by the caller
5055 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5057 unsigned long flags
;
5064 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5065 cpus_and(mask
, mask
, p
->cpus_allowed
);
5066 dest_cpu
= any_online_cpu(mask
);
5068 /* On any allowed CPU? */
5069 if (dest_cpu
== NR_CPUS
)
5070 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5072 /* No more Mr. Nice Guy. */
5073 if (dest_cpu
== NR_CPUS
) {
5074 rq
= task_rq_lock(p
, &flags
);
5075 cpus_setall(p
->cpus_allowed
);
5076 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5077 task_rq_unlock(rq
, &flags
);
5080 * Don't tell them about moving exiting tasks or
5081 * kernel threads (both mm NULL), since they never
5084 if (p
->mm
&& printk_ratelimit())
5085 printk(KERN_INFO
"process %d (%s) no "
5086 "longer affine to cpu%d\n",
5087 p
->pid
, p
->comm
, dead_cpu
);
5089 if (!__migrate_task(p
, dead_cpu
, dest_cpu
))
5094 * While a dead CPU has no uninterruptible tasks queued at this point,
5095 * it might still have a nonzero ->nr_uninterruptible counter, because
5096 * for performance reasons the counter is not stricly tracking tasks to
5097 * their home CPUs. So we just add the counter to another CPU's counter,
5098 * to keep the global sum constant after CPU-down:
5100 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5102 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5103 unsigned long flags
;
5105 local_irq_save(flags
);
5106 double_rq_lock(rq_src
, rq_dest
);
5107 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5108 rq_src
->nr_uninterruptible
= 0;
5109 double_rq_unlock(rq_src
, rq_dest
);
5110 local_irq_restore(flags
);
5113 /* Run through task list and migrate tasks from the dead cpu. */
5114 static void migrate_live_tasks(int src_cpu
)
5116 struct task_struct
*p
, *t
;
5118 write_lock_irq(&tasklist_lock
);
5120 do_each_thread(t
, p
) {
5124 if (task_cpu(p
) == src_cpu
)
5125 move_task_off_dead_cpu(src_cpu
, p
);
5126 } while_each_thread(t
, p
);
5128 write_unlock_irq(&tasklist_lock
);
5131 /* Schedules idle task to be the next runnable task on current CPU.
5132 * It does so by boosting its priority to highest possible and adding it to
5133 * the _front_ of the runqueue. Used by CPU offline code.
5135 void sched_idle_next(void)
5137 int this_cpu
= smp_processor_id();
5138 struct rq
*rq
= cpu_rq(this_cpu
);
5139 struct task_struct
*p
= rq
->idle
;
5140 unsigned long flags
;
5142 /* cpu has to be offline */
5143 BUG_ON(cpu_online(this_cpu
));
5146 * Strictly not necessary since rest of the CPUs are stopped by now
5147 * and interrupts disabled on the current cpu.
5149 spin_lock_irqsave(&rq
->lock
, flags
);
5151 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5153 /* Add idle task to the _front_ of its priority queue: */
5154 __activate_idle_task(p
, rq
);
5156 spin_unlock_irqrestore(&rq
->lock
, flags
);
5160 * Ensures that the idle task is using init_mm right before its cpu goes
5163 void idle_task_exit(void)
5165 struct mm_struct
*mm
= current
->active_mm
;
5167 BUG_ON(cpu_online(smp_processor_id()));
5170 switch_mm(mm
, &init_mm
, current
);
5174 /* called under rq->lock with disabled interrupts */
5175 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5177 struct rq
*rq
= cpu_rq(dead_cpu
);
5179 /* Must be exiting, otherwise would be on tasklist. */
5180 BUG_ON(p
->exit_state
!= EXIT_ZOMBIE
&& p
->exit_state
!= EXIT_DEAD
);
5182 /* Cannot have done final schedule yet: would have vanished. */
5183 BUG_ON(p
->state
== TASK_DEAD
);
5188 * Drop lock around migration; if someone else moves it,
5189 * that's OK. No task can be added to this CPU, so iteration is
5191 * NOTE: interrupts should be left disabled --dev@
5193 spin_unlock(&rq
->lock
);
5194 move_task_off_dead_cpu(dead_cpu
, p
);
5195 spin_lock(&rq
->lock
);
5200 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5201 static void migrate_dead_tasks(unsigned int dead_cpu
)
5203 struct rq
*rq
= cpu_rq(dead_cpu
);
5204 unsigned int arr
, i
;
5206 for (arr
= 0; arr
< 2; arr
++) {
5207 for (i
= 0; i
< MAX_PRIO
; i
++) {
5208 struct list_head
*list
= &rq
->arrays
[arr
].queue
[i
];
5210 while (!list_empty(list
))
5211 migrate_dead(dead_cpu
, list_entry(list
->next
,
5212 struct task_struct
, run_list
));
5216 #endif /* CONFIG_HOTPLUG_CPU */
5219 * migration_call - callback that gets triggered when a CPU is added.
5220 * Here we can start up the necessary migration thread for the new CPU.
5222 static int __cpuinit
5223 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5225 struct task_struct
*p
;
5226 int cpu
= (long)hcpu
;
5227 unsigned long flags
;
5231 case CPU_UP_PREPARE
:
5232 p
= kthread_create(migration_thread
, hcpu
, "migration/%d",cpu
);
5235 p
->flags
|= PF_NOFREEZE
;
5236 kthread_bind(p
, cpu
);
5237 /* Must be high prio: stop_machine expects to yield to it. */
5238 rq
= task_rq_lock(p
, &flags
);
5239 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5240 task_rq_unlock(rq
, &flags
);
5241 cpu_rq(cpu
)->migration_thread
= p
;
5245 /* Strictly unneccessary, as first user will wake it. */
5246 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5249 #ifdef CONFIG_HOTPLUG_CPU
5250 case CPU_UP_CANCELED
:
5251 if (!cpu_rq(cpu
)->migration_thread
)
5253 /* Unbind it from offline cpu so it can run. Fall thru. */
5254 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5255 any_online_cpu(cpu_online_map
));
5256 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5257 cpu_rq(cpu
)->migration_thread
= NULL
;
5261 migrate_live_tasks(cpu
);
5263 kthread_stop(rq
->migration_thread
);
5264 rq
->migration_thread
= NULL
;
5265 /* Idle task back to normal (off runqueue, low prio) */
5266 rq
= task_rq_lock(rq
->idle
, &flags
);
5267 deactivate_task(rq
->idle
, rq
);
5268 rq
->idle
->static_prio
= MAX_PRIO
;
5269 __setscheduler(rq
->idle
, SCHED_NORMAL
, 0);
5270 migrate_dead_tasks(cpu
);
5271 task_rq_unlock(rq
, &flags
);
5272 migrate_nr_uninterruptible(rq
);
5273 BUG_ON(rq
->nr_running
!= 0);
5275 /* No need to migrate the tasks: it was best-effort if
5276 * they didn't do lock_cpu_hotplug(). Just wake up
5277 * the requestors. */
5278 spin_lock_irq(&rq
->lock
);
5279 while (!list_empty(&rq
->migration_queue
)) {
5280 struct migration_req
*req
;
5282 req
= list_entry(rq
->migration_queue
.next
,
5283 struct migration_req
, list
);
5284 list_del_init(&req
->list
);
5285 complete(&req
->done
);
5287 spin_unlock_irq(&rq
->lock
);
5294 /* Register at highest priority so that task migration (migrate_all_tasks)
5295 * happens before everything else.
5297 static struct notifier_block __cpuinitdata migration_notifier
= {
5298 .notifier_call
= migration_call
,
5302 int __init
migration_init(void)
5304 void *cpu
= (void *)(long)smp_processor_id();
5307 /* Start one for the boot CPU: */
5308 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5309 BUG_ON(err
== NOTIFY_BAD
);
5310 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5311 register_cpu_notifier(&migration_notifier
);
5318 #undef SCHED_DOMAIN_DEBUG
5319 #ifdef SCHED_DOMAIN_DEBUG
5320 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5325 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5329 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5334 struct sched_group
*group
= sd
->groups
;
5335 cpumask_t groupmask
;
5337 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5338 cpus_clear(groupmask
);
5341 for (i
= 0; i
< level
+ 1; i
++)
5343 printk("domain %d: ", level
);
5345 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5346 printk("does not load-balance\n");
5348 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain has parent");
5352 printk("span %s\n", str
);
5354 if (!cpu_isset(cpu
, sd
->span
))
5355 printk(KERN_ERR
"ERROR: domain->span does not contain CPU%d\n", cpu
);
5356 if (!cpu_isset(cpu
, group
->cpumask
))
5357 printk(KERN_ERR
"ERROR: domain->groups does not contain CPU%d\n", cpu
);
5360 for (i
= 0; i
< level
+ 2; i
++)
5366 printk(KERN_ERR
"ERROR: group is NULL\n");
5370 if (!group
->cpu_power
) {
5372 printk(KERN_ERR
"ERROR: domain->cpu_power not set\n");
5375 if (!cpus_weight(group
->cpumask
)) {
5377 printk(KERN_ERR
"ERROR: empty group\n");
5380 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5382 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5385 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5387 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5390 group
= group
->next
;
5391 } while (group
!= sd
->groups
);
5394 if (!cpus_equal(sd
->span
, groupmask
))
5395 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5401 if (!cpus_subset(groupmask
, sd
->span
))
5402 printk(KERN_ERR
"ERROR: parent span is not a superset of domain->span\n");
5408 # define sched_domain_debug(sd, cpu) do { } while (0)
5411 static int sd_degenerate(struct sched_domain
*sd
)
5413 if (cpus_weight(sd
->span
) == 1)
5416 /* Following flags need at least 2 groups */
5417 if (sd
->flags
& (SD_LOAD_BALANCE
|
5418 SD_BALANCE_NEWIDLE
|
5422 SD_SHARE_PKG_RESOURCES
)) {
5423 if (sd
->groups
!= sd
->groups
->next
)
5427 /* Following flags don't use groups */
5428 if (sd
->flags
& (SD_WAKE_IDLE
|
5437 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5439 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5441 if (sd_degenerate(parent
))
5444 if (!cpus_equal(sd
->span
, parent
->span
))
5447 /* Does parent contain flags not in child? */
5448 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5449 if (cflags
& SD_WAKE_AFFINE
)
5450 pflags
&= ~SD_WAKE_BALANCE
;
5451 /* Flags needing groups don't count if only 1 group in parent */
5452 if (parent
->groups
== parent
->groups
->next
) {
5453 pflags
&= ~(SD_LOAD_BALANCE
|
5454 SD_BALANCE_NEWIDLE
|
5458 SD_SHARE_PKG_RESOURCES
);
5460 if (~cflags
& pflags
)
5467 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5468 * hold the hotplug lock.
5470 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5472 struct rq
*rq
= cpu_rq(cpu
);
5473 struct sched_domain
*tmp
;
5475 /* Remove the sched domains which do not contribute to scheduling. */
5476 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5477 struct sched_domain
*parent
= tmp
->parent
;
5480 if (sd_parent_degenerate(tmp
, parent
)) {
5481 tmp
->parent
= parent
->parent
;
5483 parent
->parent
->child
= tmp
;
5487 if (sd
&& sd_degenerate(sd
)) {
5493 sched_domain_debug(sd
, cpu
);
5495 rcu_assign_pointer(rq
->sd
, sd
);
5498 /* cpus with isolated domains */
5499 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
5501 /* Setup the mask of cpus configured for isolated domains */
5502 static int __init
isolated_cpu_setup(char *str
)
5504 int ints
[NR_CPUS
], i
;
5506 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5507 cpus_clear(cpu_isolated_map
);
5508 for (i
= 1; i
<= ints
[0]; i
++)
5509 if (ints
[i
] < NR_CPUS
)
5510 cpu_set(ints
[i
], cpu_isolated_map
);
5514 __setup ("isolcpus=", isolated_cpu_setup
);
5517 * init_sched_build_groups takes an array of groups, the cpumask we wish
5518 * to span, and a pointer to a function which identifies what group a CPU
5519 * belongs to. The return value of group_fn must be a valid index into the
5520 * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
5521 * keep track of groups covered with a cpumask_t).
5523 * init_sched_build_groups will build a circular linked list of the groups
5524 * covered by the given span, and will set each group's ->cpumask correctly,
5525 * and ->cpu_power to 0.
5528 init_sched_build_groups(struct sched_group groups
[], cpumask_t span
,
5529 const cpumask_t
*cpu_map
,
5530 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
))
5532 struct sched_group
*first
= NULL
, *last
= NULL
;
5533 cpumask_t covered
= CPU_MASK_NONE
;
5536 for_each_cpu_mask(i
, span
) {
5537 int group
= group_fn(i
, cpu_map
);
5538 struct sched_group
*sg
= &groups
[group
];
5541 if (cpu_isset(i
, covered
))
5544 sg
->cpumask
= CPU_MASK_NONE
;
5547 for_each_cpu_mask(j
, span
) {
5548 if (group_fn(j
, cpu_map
) != group
)
5551 cpu_set(j
, covered
);
5552 cpu_set(j
, sg
->cpumask
);
5563 #define SD_NODES_PER_DOMAIN 16
5566 * Self-tuning task migration cost measurement between source and target CPUs.
5568 * This is done by measuring the cost of manipulating buffers of varying
5569 * sizes. For a given buffer-size here are the steps that are taken:
5571 * 1) the source CPU reads+dirties a shared buffer
5572 * 2) the target CPU reads+dirties the same shared buffer
5574 * We measure how long they take, in the following 4 scenarios:
5576 * - source: CPU1, target: CPU2 | cost1
5577 * - source: CPU2, target: CPU1 | cost2
5578 * - source: CPU1, target: CPU1 | cost3
5579 * - source: CPU2, target: CPU2 | cost4
5581 * We then calculate the cost3+cost4-cost1-cost2 difference - this is
5582 * the cost of migration.
5584 * We then start off from a small buffer-size and iterate up to larger
5585 * buffer sizes, in 5% steps - measuring each buffer-size separately, and
5586 * doing a maximum search for the cost. (The maximum cost for a migration
5587 * normally occurs when the working set size is around the effective cache
5590 #define SEARCH_SCOPE 2
5591 #define MIN_CACHE_SIZE (64*1024U)
5592 #define DEFAULT_CACHE_SIZE (5*1024*1024U)
5593 #define ITERATIONS 1
5594 #define SIZE_THRESH 130
5595 #define COST_THRESH 130
5598 * The migration cost is a function of 'domain distance'. Domain
5599 * distance is the number of steps a CPU has to iterate down its
5600 * domain tree to share a domain with the other CPU. The farther
5601 * two CPUs are from each other, the larger the distance gets.
5603 * Note that we use the distance only to cache measurement results,
5604 * the distance value is not used numerically otherwise. When two
5605 * CPUs have the same distance it is assumed that the migration
5606 * cost is the same. (this is a simplification but quite practical)
5608 #define MAX_DOMAIN_DISTANCE 32
5610 static unsigned long long migration_cost
[MAX_DOMAIN_DISTANCE
] =
5611 { [ 0 ... MAX_DOMAIN_DISTANCE
-1 ] =
5613 * Architectures may override the migration cost and thus avoid
5614 * boot-time calibration. Unit is nanoseconds. Mostly useful for
5615 * virtualized hardware:
5617 #ifdef CONFIG_DEFAULT_MIGRATION_COST
5618 CONFIG_DEFAULT_MIGRATION_COST
5625 * Allow override of migration cost - in units of microseconds.
5626 * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost
5627 * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs:
5629 static int __init
migration_cost_setup(char *str
)
5631 int ints
[MAX_DOMAIN_DISTANCE
+1], i
;
5633 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5635 printk("#ints: %d\n", ints
[0]);
5636 for (i
= 1; i
<= ints
[0]; i
++) {
5637 migration_cost
[i
-1] = (unsigned long long)ints
[i
]*1000;
5638 printk("migration_cost[%d]: %Ld\n", i
-1, migration_cost
[i
-1]);
5643 __setup ("migration_cost=", migration_cost_setup
);
5646 * Global multiplier (divisor) for migration-cutoff values,
5647 * in percentiles. E.g. use a value of 150 to get 1.5 times
5648 * longer cache-hot cutoff times.
5650 * (We scale it from 100 to 128 to long long handling easier.)
5653 #define MIGRATION_FACTOR_SCALE 128
5655 static unsigned int migration_factor
= MIGRATION_FACTOR_SCALE
;
5657 static int __init
setup_migration_factor(char *str
)
5659 get_option(&str
, &migration_factor
);
5660 migration_factor
= migration_factor
* MIGRATION_FACTOR_SCALE
/ 100;
5664 __setup("migration_factor=", setup_migration_factor
);
5667 * Estimated distance of two CPUs, measured via the number of domains
5668 * we have to pass for the two CPUs to be in the same span:
5670 static unsigned long domain_distance(int cpu1
, int cpu2
)
5672 unsigned long distance
= 0;
5673 struct sched_domain
*sd
;
5675 for_each_domain(cpu1
, sd
) {
5676 WARN_ON(!cpu_isset(cpu1
, sd
->span
));
5677 if (cpu_isset(cpu2
, sd
->span
))
5681 if (distance
>= MAX_DOMAIN_DISTANCE
) {
5683 distance
= MAX_DOMAIN_DISTANCE
-1;
5689 static unsigned int migration_debug
;
5691 static int __init
setup_migration_debug(char *str
)
5693 get_option(&str
, &migration_debug
);
5697 __setup("migration_debug=", setup_migration_debug
);
5700 * Maximum cache-size that the scheduler should try to measure.
5701 * Architectures with larger caches should tune this up during
5702 * bootup. Gets used in the domain-setup code (i.e. during SMP
5705 unsigned int max_cache_size
;
5707 static int __init
setup_max_cache_size(char *str
)
5709 get_option(&str
, &max_cache_size
);
5713 __setup("max_cache_size=", setup_max_cache_size
);
5716 * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This
5717 * is the operation that is timed, so we try to generate unpredictable
5718 * cachemisses that still end up filling the L2 cache:
5720 static void touch_cache(void *__cache
, unsigned long __size
)
5722 unsigned long size
= __size
/sizeof(long), chunk1
= size
/3,
5724 unsigned long *cache
= __cache
;
5727 for (i
= 0; i
< size
/6; i
+= 8) {
5730 case 1: cache
[size
-1-i
]++;
5731 case 2: cache
[chunk1
-i
]++;
5732 case 3: cache
[chunk1
+i
]++;
5733 case 4: cache
[chunk2
-i
]++;
5734 case 5: cache
[chunk2
+i
]++;
5740 * Measure the cache-cost of one task migration. Returns in units of nsec.
5742 static unsigned long long
5743 measure_one(void *cache
, unsigned long size
, int source
, int target
)
5745 cpumask_t mask
, saved_mask
;
5746 unsigned long long t0
, t1
, t2
, t3
, cost
;
5748 saved_mask
= current
->cpus_allowed
;
5751 * Flush source caches to RAM and invalidate them:
5756 * Migrate to the source CPU:
5758 mask
= cpumask_of_cpu(source
);
5759 set_cpus_allowed(current
, mask
);
5760 WARN_ON(smp_processor_id() != source
);
5763 * Dirty the working set:
5766 touch_cache(cache
, size
);
5770 * Migrate to the target CPU, dirty the L2 cache and access
5771 * the shared buffer. (which represents the working set
5772 * of a migrated task.)
5774 mask
= cpumask_of_cpu(target
);
5775 set_cpus_allowed(current
, mask
);
5776 WARN_ON(smp_processor_id() != target
);
5779 touch_cache(cache
, size
);
5782 cost
= t1
-t0
+ t3
-t2
;
5784 if (migration_debug
>= 2)
5785 printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n",
5786 source
, target
, t1
-t0
, t1
-t0
, t3
-t2
, cost
);
5788 * Flush target caches to RAM and invalidate them:
5792 set_cpus_allowed(current
, saved_mask
);
5798 * Measure a series of task migrations and return the average
5799 * result. Since this code runs early during bootup the system
5800 * is 'undisturbed' and the average latency makes sense.
5802 * The algorithm in essence auto-detects the relevant cache-size,
5803 * so it will properly detect different cachesizes for different
5804 * cache-hierarchies, depending on how the CPUs are connected.
5806 * Architectures can prime the upper limit of the search range via
5807 * max_cache_size, otherwise the search range defaults to 20MB...64K.
5809 static unsigned long long
5810 measure_cost(int cpu1
, int cpu2
, void *cache
, unsigned int size
)
5812 unsigned long long cost1
, cost2
;
5816 * Measure the migration cost of 'size' bytes, over an
5817 * average of 10 runs:
5819 * (We perturb the cache size by a small (0..4k)
5820 * value to compensate size/alignment related artifacts.
5821 * We also subtract the cost of the operation done on
5827 * dry run, to make sure we start off cache-cold on cpu1,
5828 * and to get any vmalloc pagefaults in advance:
5830 measure_one(cache
, size
, cpu1
, cpu2
);
5831 for (i
= 0; i
< ITERATIONS
; i
++)
5832 cost1
+= measure_one(cache
, size
- i
*1024, cpu1
, cpu2
);
5834 measure_one(cache
, size
, cpu2
, cpu1
);
5835 for (i
= 0; i
< ITERATIONS
; i
++)
5836 cost1
+= measure_one(cache
, size
- i
*1024, cpu2
, cpu1
);
5839 * (We measure the non-migrating [cached] cost on both
5840 * cpu1 and cpu2, to handle CPUs with different speeds)
5844 measure_one(cache
, size
, cpu1
, cpu1
);
5845 for (i
= 0; i
< ITERATIONS
; i
++)
5846 cost2
+= measure_one(cache
, size
- i
*1024, cpu1
, cpu1
);
5848 measure_one(cache
, size
, cpu2
, cpu2
);
5849 for (i
= 0; i
< ITERATIONS
; i
++)
5850 cost2
+= measure_one(cache
, size
- i
*1024, cpu2
, cpu2
);
5853 * Get the per-iteration migration cost:
5855 do_div(cost1
, 2*ITERATIONS
);
5856 do_div(cost2
, 2*ITERATIONS
);
5858 return cost1
- cost2
;
5861 static unsigned long long measure_migration_cost(int cpu1
, int cpu2
)
5863 unsigned long long max_cost
= 0, fluct
= 0, avg_fluct
= 0;
5864 unsigned int max_size
, size
, size_found
= 0;
5865 long long cost
= 0, prev_cost
;
5869 * Search from max_cache_size*5 down to 64K - the real relevant
5870 * cachesize has to lie somewhere inbetween.
5872 if (max_cache_size
) {
5873 max_size
= max(max_cache_size
* SEARCH_SCOPE
, MIN_CACHE_SIZE
);
5874 size
= max(max_cache_size
/ SEARCH_SCOPE
, MIN_CACHE_SIZE
);
5877 * Since we have no estimation about the relevant
5880 max_size
= DEFAULT_CACHE_SIZE
* SEARCH_SCOPE
;
5881 size
= MIN_CACHE_SIZE
;
5884 if (!cpu_online(cpu1
) || !cpu_online(cpu2
)) {
5885 printk("cpu %d and %d not both online!\n", cpu1
, cpu2
);
5890 * Allocate the working set:
5892 cache
= vmalloc(max_size
);
5894 printk("could not vmalloc %d bytes for cache!\n", 2*max_size
);
5895 return 1000000; /* return 1 msec on very small boxen */
5898 while (size
<= max_size
) {
5900 cost
= measure_cost(cpu1
, cpu2
, cache
, size
);
5906 if (max_cost
< cost
) {
5912 * Calculate average fluctuation, we use this to prevent
5913 * noise from triggering an early break out of the loop:
5915 fluct
= abs(cost
- prev_cost
);
5916 avg_fluct
= (avg_fluct
+ fluct
)/2;
5918 if (migration_debug
)
5919 printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): (%8Ld %8Ld)\n",
5921 (long)cost
/ 1000000,
5922 ((long)cost
/ 100000) % 10,
5923 (long)max_cost
/ 1000000,
5924 ((long)max_cost
/ 100000) % 10,
5925 domain_distance(cpu1
, cpu2
),
5929 * If we iterated at least 20% past the previous maximum,
5930 * and the cost has dropped by more than 20% already,
5931 * (taking fluctuations into account) then we assume to
5932 * have found the maximum and break out of the loop early:
5934 if (size_found
&& (size
*100 > size_found
*SIZE_THRESH
))
5935 if (cost
+avg_fluct
<= 0 ||
5936 max_cost
*100 > (cost
+avg_fluct
)*COST_THRESH
) {
5938 if (migration_debug
)
5939 printk("-> found max.\n");
5943 * Increase the cachesize in 10% steps:
5945 size
= size
* 10 / 9;
5948 if (migration_debug
)
5949 printk("[%d][%d] working set size found: %d, cost: %Ld\n",
5950 cpu1
, cpu2
, size_found
, max_cost
);
5955 * A task is considered 'cache cold' if at least 2 times
5956 * the worst-case cost of migration has passed.
5958 * (this limit is only listened to if the load-balancing
5959 * situation is 'nice' - if there is a large imbalance we
5960 * ignore it for the sake of CPU utilization and
5961 * processing fairness.)
5963 return 2 * max_cost
* migration_factor
/ MIGRATION_FACTOR_SCALE
;
5966 static void calibrate_migration_costs(const cpumask_t
*cpu_map
)
5968 int cpu1
= -1, cpu2
= -1, cpu
, orig_cpu
= raw_smp_processor_id();
5969 unsigned long j0
, j1
, distance
, max_distance
= 0;
5970 struct sched_domain
*sd
;
5975 * First pass - calculate the cacheflush times:
5977 for_each_cpu_mask(cpu1
, *cpu_map
) {
5978 for_each_cpu_mask(cpu2
, *cpu_map
) {
5981 distance
= domain_distance(cpu1
, cpu2
);
5982 max_distance
= max(max_distance
, distance
);
5984 * No result cached yet?
5986 if (migration_cost
[distance
] == -1LL)
5987 migration_cost
[distance
] =
5988 measure_migration_cost(cpu1
, cpu2
);
5992 * Second pass - update the sched domain hierarchy with
5993 * the new cache-hot-time estimations:
5995 for_each_cpu_mask(cpu
, *cpu_map
) {
5997 for_each_domain(cpu
, sd
) {
5998 sd
->cache_hot_time
= migration_cost
[distance
];
6005 if (migration_debug
)
6006 printk("migration: max_cache_size: %d, cpu: %d MHz:\n",
6014 if (system_state
== SYSTEM_BOOTING
) {
6015 if (num_online_cpus() > 1) {
6016 printk("migration_cost=");
6017 for (distance
= 0; distance
<= max_distance
; distance
++) {
6020 printk("%ld", (long)migration_cost
[distance
] / 1000);
6026 if (migration_debug
)
6027 printk("migration: %ld seconds\n", (j1
-j0
)/HZ
);
6030 * Move back to the original CPU. NUMA-Q gets confused
6031 * if we migrate to another quad during bootup.
6033 if (raw_smp_processor_id() != orig_cpu
) {
6034 cpumask_t mask
= cpumask_of_cpu(orig_cpu
),
6035 saved_mask
= current
->cpus_allowed
;
6037 set_cpus_allowed(current
, mask
);
6038 set_cpus_allowed(current
, saved_mask
);
6045 * find_next_best_node - find the next node to include in a sched_domain
6046 * @node: node whose sched_domain we're building
6047 * @used_nodes: nodes already in the sched_domain
6049 * Find the next node to include in a given scheduling domain. Simply
6050 * finds the closest node not already in the @used_nodes map.
6052 * Should use nodemask_t.
6054 static int find_next_best_node(int node
, unsigned long *used_nodes
)
6056 int i
, n
, val
, min_val
, best_node
= 0;
6060 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6061 /* Start at @node */
6062 n
= (node
+ i
) % MAX_NUMNODES
;
6064 if (!nr_cpus_node(n
))
6067 /* Skip already used nodes */
6068 if (test_bit(n
, used_nodes
))
6071 /* Simple min distance search */
6072 val
= node_distance(node
, n
);
6074 if (val
< min_val
) {
6080 set_bit(best_node
, used_nodes
);
6085 * sched_domain_node_span - get a cpumask for a node's sched_domain
6086 * @node: node whose cpumask we're constructing
6087 * @size: number of nodes to include in this span
6089 * Given a node, construct a good cpumask for its sched_domain to span. It
6090 * should be one that prevents unnecessary balancing, but also spreads tasks
6093 static cpumask_t
sched_domain_node_span(int node
)
6095 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
6096 cpumask_t span
, nodemask
;
6100 bitmap_zero(used_nodes
, MAX_NUMNODES
);
6102 nodemask
= node_to_cpumask(node
);
6103 cpus_or(span
, span
, nodemask
);
6104 set_bit(node
, used_nodes
);
6106 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6107 int next_node
= find_next_best_node(node
, used_nodes
);
6109 nodemask
= node_to_cpumask(next_node
);
6110 cpus_or(span
, span
, nodemask
);
6117 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6120 * SMT sched-domains:
6122 #ifdef CONFIG_SCHED_SMT
6123 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
6124 static struct sched_group sched_group_cpus
[NR_CPUS
];
6126 static int cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
)
6133 * multi-core sched-domains:
6135 #ifdef CONFIG_SCHED_MC
6136 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
6137 static struct sched_group sched_group_core
[NR_CPUS
];
6140 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6141 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
)
6143 cpumask_t mask
= cpu_sibling_map
[cpu
];
6144 cpus_and(mask
, mask
, *cpu_map
);
6145 return first_cpu(mask
);
6147 #elif defined(CONFIG_SCHED_MC)
6148 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
)
6154 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
6155 static struct sched_group sched_group_phys
[NR_CPUS
];
6157 static int cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
)
6159 #ifdef CONFIG_SCHED_MC
6160 cpumask_t mask
= cpu_coregroup_map(cpu
);
6161 cpus_and(mask
, mask
, *cpu_map
);
6162 return first_cpu(mask
);
6163 #elif defined(CONFIG_SCHED_SMT)
6164 cpumask_t mask
= cpu_sibling_map
[cpu
];
6165 cpus_and(mask
, mask
, *cpu_map
);
6166 return first_cpu(mask
);
6174 * The init_sched_build_groups can't handle what we want to do with node
6175 * groups, so roll our own. Now each node has its own list of groups which
6176 * gets dynamically allocated.
6178 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
6179 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
6181 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
6182 static struct sched_group
*sched_group_allnodes_bycpu
[NR_CPUS
];
6184 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
)
6186 return cpu_to_node(cpu
);
6188 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6190 struct sched_group
*sg
= group_head
;
6196 for_each_cpu_mask(j
, sg
->cpumask
) {
6197 struct sched_domain
*sd
;
6199 sd
= &per_cpu(phys_domains
, j
);
6200 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
6202 * Only add "power" once for each
6208 sg
->cpu_power
+= sd
->groups
->cpu_power
;
6211 if (sg
!= group_head
)
6217 /* Free memory allocated for various sched_group structures */
6218 static void free_sched_groups(const cpumask_t
*cpu_map
)
6222 for_each_cpu_mask(cpu
, *cpu_map
) {
6223 struct sched_group
*sched_group_allnodes
6224 = sched_group_allnodes_bycpu
[cpu
];
6225 struct sched_group
**sched_group_nodes
6226 = sched_group_nodes_bycpu
[cpu
];
6228 if (sched_group_allnodes
) {
6229 kfree(sched_group_allnodes
);
6230 sched_group_allnodes_bycpu
[cpu
] = NULL
;
6233 if (!sched_group_nodes
)
6236 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6237 cpumask_t nodemask
= node_to_cpumask(i
);
6238 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6240 cpus_and(nodemask
, nodemask
, *cpu_map
);
6241 if (cpus_empty(nodemask
))
6251 if (oldsg
!= sched_group_nodes
[i
])
6254 kfree(sched_group_nodes
);
6255 sched_group_nodes_bycpu
[cpu
] = NULL
;
6259 static void free_sched_groups(const cpumask_t
*cpu_map
)
6265 * Initialize sched groups cpu_power.
6267 * cpu_power indicates the capacity of sched group, which is used while
6268 * distributing the load between different sched groups in a sched domain.
6269 * Typically cpu_power for all the groups in a sched domain will be same unless
6270 * there are asymmetries in the topology. If there are asymmetries, group
6271 * having more cpu_power will pickup more load compared to the group having
6274 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6275 * the maximum number of tasks a group can handle in the presence of other idle
6276 * or lightly loaded groups in the same sched domain.
6278 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6280 struct sched_domain
*child
;
6281 struct sched_group
*group
;
6283 WARN_ON(!sd
|| !sd
->groups
);
6285 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
6291 * For perf policy, if the groups in child domain share resources
6292 * (for example cores sharing some portions of the cache hierarchy
6293 * or SMT), then set this domain groups cpu_power such that each group
6294 * can handle only one task, when there are other idle groups in the
6295 * same sched domain.
6297 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
6299 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
6300 sd
->groups
->cpu_power
= SCHED_LOAD_SCALE
;
6304 sd
->groups
->cpu_power
= 0;
6307 * add cpu_power of each child group to this groups cpu_power
6309 group
= child
->groups
;
6311 sd
->groups
->cpu_power
+= group
->cpu_power
;
6312 group
= group
->next
;
6313 } while (group
!= child
->groups
);
6317 * Build sched domains for a given set of cpus and attach the sched domains
6318 * to the individual cpus
6320 static int build_sched_domains(const cpumask_t
*cpu_map
)
6323 struct sched_domain
*sd
;
6325 struct sched_group
**sched_group_nodes
= NULL
;
6326 struct sched_group
*sched_group_allnodes
= NULL
;
6329 * Allocate the per-node list of sched groups
6331 sched_group_nodes
= kzalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
6333 if (!sched_group_nodes
) {
6334 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6337 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6341 * Set up domains for cpus specified by the cpu_map.
6343 for_each_cpu_mask(i
, *cpu_map
) {
6345 struct sched_domain
*sd
= NULL
, *p
;
6346 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6348 cpus_and(nodemask
, nodemask
, *cpu_map
);
6351 if (cpus_weight(*cpu_map
)
6352 > SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6353 if (!sched_group_allnodes
) {
6354 sched_group_allnodes
6355 = kmalloc_node(sizeof(struct sched_group
)
6359 if (!sched_group_allnodes
) {
6361 "Can not alloc allnodes sched group\n");
6364 sched_group_allnodes_bycpu
[i
]
6365 = sched_group_allnodes
;
6367 sd
= &per_cpu(allnodes_domains
, i
);
6368 *sd
= SD_ALLNODES_INIT
;
6369 sd
->span
= *cpu_map
;
6370 group
= cpu_to_allnodes_group(i
, cpu_map
);
6371 sd
->groups
= &sched_group_allnodes
[group
];
6376 sd
= &per_cpu(node_domains
, i
);
6378 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6382 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6386 sd
= &per_cpu(phys_domains
, i
);
6387 group
= cpu_to_phys_group(i
, cpu_map
);
6389 sd
->span
= nodemask
;
6393 sd
->groups
= &sched_group_phys
[group
];
6395 #ifdef CONFIG_SCHED_MC
6397 sd
= &per_cpu(core_domains
, i
);
6398 group
= cpu_to_core_group(i
, cpu_map
);
6400 sd
->span
= cpu_coregroup_map(i
);
6401 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6404 sd
->groups
= &sched_group_core
[group
];
6407 #ifdef CONFIG_SCHED_SMT
6409 sd
= &per_cpu(cpu_domains
, i
);
6410 group
= cpu_to_cpu_group(i
, cpu_map
);
6411 *sd
= SD_SIBLING_INIT
;
6412 sd
->span
= cpu_sibling_map
[i
];
6413 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6416 sd
->groups
= &sched_group_cpus
[group
];
6420 #ifdef CONFIG_SCHED_SMT
6421 /* Set up CPU (sibling) groups */
6422 for_each_cpu_mask(i
, *cpu_map
) {
6423 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
6424 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6425 if (i
!= first_cpu(this_sibling_map
))
6428 init_sched_build_groups(sched_group_cpus
, this_sibling_map
,
6429 cpu_map
, &cpu_to_cpu_group
);
6433 #ifdef CONFIG_SCHED_MC
6434 /* Set up multi-core groups */
6435 for_each_cpu_mask(i
, *cpu_map
) {
6436 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6437 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6438 if (i
!= first_cpu(this_core_map
))
6440 init_sched_build_groups(sched_group_core
, this_core_map
,
6441 cpu_map
, &cpu_to_core_group
);
6446 /* Set up physical groups */
6447 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6448 cpumask_t nodemask
= node_to_cpumask(i
);
6450 cpus_and(nodemask
, nodemask
, *cpu_map
);
6451 if (cpus_empty(nodemask
))
6454 init_sched_build_groups(sched_group_phys
, nodemask
,
6455 cpu_map
, &cpu_to_phys_group
);
6459 /* Set up node groups */
6460 if (sched_group_allnodes
)
6461 init_sched_build_groups(sched_group_allnodes
, *cpu_map
,
6462 cpu_map
, &cpu_to_allnodes_group
);
6464 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6465 /* Set up node groups */
6466 struct sched_group
*sg
, *prev
;
6467 cpumask_t nodemask
= node_to_cpumask(i
);
6468 cpumask_t domainspan
;
6469 cpumask_t covered
= CPU_MASK_NONE
;
6472 cpus_and(nodemask
, nodemask
, *cpu_map
);
6473 if (cpus_empty(nodemask
)) {
6474 sched_group_nodes
[i
] = NULL
;
6478 domainspan
= sched_domain_node_span(i
);
6479 cpus_and(domainspan
, domainspan
, *cpu_map
);
6481 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6483 printk(KERN_WARNING
"Can not alloc domain group for "
6487 sched_group_nodes
[i
] = sg
;
6488 for_each_cpu_mask(j
, nodemask
) {
6489 struct sched_domain
*sd
;
6490 sd
= &per_cpu(node_domains
, j
);
6494 sg
->cpumask
= nodemask
;
6496 cpus_or(covered
, covered
, nodemask
);
6499 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6500 cpumask_t tmp
, notcovered
;
6501 int n
= (i
+ j
) % MAX_NUMNODES
;
6503 cpus_complement(notcovered
, covered
);
6504 cpus_and(tmp
, notcovered
, *cpu_map
);
6505 cpus_and(tmp
, tmp
, domainspan
);
6506 if (cpus_empty(tmp
))
6509 nodemask
= node_to_cpumask(n
);
6510 cpus_and(tmp
, tmp
, nodemask
);
6511 if (cpus_empty(tmp
))
6514 sg
= kmalloc_node(sizeof(struct sched_group
),
6518 "Can not alloc domain group for node %d\n", j
);
6523 sg
->next
= prev
->next
;
6524 cpus_or(covered
, covered
, tmp
);
6531 /* Calculate CPU power for physical packages and nodes */
6532 #ifdef CONFIG_SCHED_SMT
6533 for_each_cpu_mask(i
, *cpu_map
) {
6534 sd
= &per_cpu(cpu_domains
, i
);
6535 init_sched_groups_power(i
, sd
);
6538 #ifdef CONFIG_SCHED_MC
6539 for_each_cpu_mask(i
, *cpu_map
) {
6540 sd
= &per_cpu(core_domains
, i
);
6541 init_sched_groups_power(i
, sd
);
6545 for_each_cpu_mask(i
, *cpu_map
) {
6546 sd
= &per_cpu(phys_domains
, i
);
6547 init_sched_groups_power(i
, sd
);
6551 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6552 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6554 if (sched_group_allnodes
) {
6555 int group
= cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
);
6556 struct sched_group
*sg
= &sched_group_allnodes
[group
];
6558 init_numa_sched_groups_power(sg
);
6562 /* Attach the domains */
6563 for_each_cpu_mask(i
, *cpu_map
) {
6564 struct sched_domain
*sd
;
6565 #ifdef CONFIG_SCHED_SMT
6566 sd
= &per_cpu(cpu_domains
, i
);
6567 #elif defined(CONFIG_SCHED_MC)
6568 sd
= &per_cpu(core_domains
, i
);
6570 sd
= &per_cpu(phys_domains
, i
);
6572 cpu_attach_domain(sd
, i
);
6575 * Tune cache-hot values:
6577 calibrate_migration_costs(cpu_map
);
6583 free_sched_groups(cpu_map
);
6588 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6590 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6592 cpumask_t cpu_default_map
;
6596 * Setup mask for cpus without special case scheduling requirements.
6597 * For now this just excludes isolated cpus, but could be used to
6598 * exclude other special cases in the future.
6600 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
6602 err
= build_sched_domains(&cpu_default_map
);
6607 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6609 free_sched_groups(cpu_map
);
6613 * Detach sched domains from a group of cpus specified in cpu_map
6614 * These cpus will now be attached to the NULL domain
6616 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6620 for_each_cpu_mask(i
, *cpu_map
)
6621 cpu_attach_domain(NULL
, i
);
6622 synchronize_sched();
6623 arch_destroy_sched_domains(cpu_map
);
6627 * Partition sched domains as specified by the cpumasks below.
6628 * This attaches all cpus from the cpumasks to the NULL domain,
6629 * waits for a RCU quiescent period, recalculates sched
6630 * domain information and then attaches them back to the
6631 * correct sched domains
6632 * Call with hotplug lock held
6634 int partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
6636 cpumask_t change_map
;
6639 cpus_and(*partition1
, *partition1
, cpu_online_map
);
6640 cpus_and(*partition2
, *partition2
, cpu_online_map
);
6641 cpus_or(change_map
, *partition1
, *partition2
);
6643 /* Detach sched domains from all of the affected cpus */
6644 detach_destroy_domains(&change_map
);
6645 if (!cpus_empty(*partition1
))
6646 err
= build_sched_domains(partition1
);
6647 if (!err
&& !cpus_empty(*partition2
))
6648 err
= build_sched_domains(partition2
);
6653 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6654 int arch_reinit_sched_domains(void)
6659 detach_destroy_domains(&cpu_online_map
);
6660 err
= arch_init_sched_domains(&cpu_online_map
);
6661 unlock_cpu_hotplug();
6666 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6670 if (buf
[0] != '0' && buf
[0] != '1')
6674 sched_smt_power_savings
= (buf
[0] == '1');
6676 sched_mc_power_savings
= (buf
[0] == '1');
6678 ret
= arch_reinit_sched_domains();
6680 return ret
? ret
: count
;
6683 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6687 #ifdef CONFIG_SCHED_SMT
6689 err
= sysfs_create_file(&cls
->kset
.kobj
,
6690 &attr_sched_smt_power_savings
.attr
);
6692 #ifdef CONFIG_SCHED_MC
6693 if (!err
&& mc_capable())
6694 err
= sysfs_create_file(&cls
->kset
.kobj
,
6695 &attr_sched_mc_power_savings
.attr
);
6701 #ifdef CONFIG_SCHED_MC
6702 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6704 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6706 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6707 const char *buf
, size_t count
)
6709 return sched_power_savings_store(buf
, count
, 0);
6711 SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6712 sched_mc_power_savings_store
);
6715 #ifdef CONFIG_SCHED_SMT
6716 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6718 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6720 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6721 const char *buf
, size_t count
)
6723 return sched_power_savings_store(buf
, count
, 1);
6725 SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6726 sched_smt_power_savings_store
);
6730 #ifdef CONFIG_HOTPLUG_CPU
6732 * Force a reinitialization of the sched domains hierarchy. The domains
6733 * and groups cannot be updated in place without racing with the balancing
6734 * code, so we temporarily attach all running cpus to the NULL domain
6735 * which will prevent rebalancing while the sched domains are recalculated.
6737 static int update_sched_domains(struct notifier_block
*nfb
,
6738 unsigned long action
, void *hcpu
)
6741 case CPU_UP_PREPARE
:
6742 case CPU_DOWN_PREPARE
:
6743 detach_destroy_domains(&cpu_online_map
);
6746 case CPU_UP_CANCELED
:
6747 case CPU_DOWN_FAILED
:
6751 * Fall through and re-initialise the domains.
6758 /* The hotplug lock is already held by cpu_up/cpu_down */
6759 arch_init_sched_domains(&cpu_online_map
);
6765 void __init
sched_init_smp(void)
6767 cpumask_t non_isolated_cpus
;
6770 arch_init_sched_domains(&cpu_online_map
);
6771 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
6772 if (cpus_empty(non_isolated_cpus
))
6773 cpu_set(smp_processor_id(), non_isolated_cpus
);
6774 unlock_cpu_hotplug();
6775 /* XXX: Theoretical race here - CPU may be hotplugged now */
6776 hotcpu_notifier(update_sched_domains
, 0);
6778 /* Move init over to a non-isolated CPU */
6779 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
6783 void __init
sched_init_smp(void)
6786 #endif /* CONFIG_SMP */
6788 int in_sched_functions(unsigned long addr
)
6790 /* Linker adds these: start and end of __sched functions */
6791 extern char __sched_text_start
[], __sched_text_end
[];
6793 return in_lock_functions(addr
) ||
6794 (addr
>= (unsigned long)__sched_text_start
6795 && addr
< (unsigned long)__sched_text_end
);
6798 void __init
sched_init(void)
6802 for_each_possible_cpu(i
) {
6803 struct prio_array
*array
;
6807 spin_lock_init(&rq
->lock
);
6808 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
6810 rq
->active
= rq
->arrays
;
6811 rq
->expired
= rq
->arrays
+ 1;
6812 rq
->best_expired_prio
= MAX_PRIO
;
6816 for (j
= 1; j
< 3; j
++)
6817 rq
->cpu_load
[j
] = 0;
6818 rq
->active_balance
= 0;
6821 rq
->migration_thread
= NULL
;
6822 INIT_LIST_HEAD(&rq
->migration_queue
);
6824 atomic_set(&rq
->nr_iowait
, 0);
6826 for (j
= 0; j
< 2; j
++) {
6827 array
= rq
->arrays
+ j
;
6828 for (k
= 0; k
< MAX_PRIO
; k
++) {
6829 INIT_LIST_HEAD(array
->queue
+ k
);
6830 __clear_bit(k
, array
->bitmap
);
6832 // delimiter for bitsearch
6833 __set_bit(MAX_PRIO
, array
->bitmap
);
6837 set_load_weight(&init_task
);
6839 #ifdef CONFIG_RT_MUTEXES
6840 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
6844 * The boot idle thread does lazy MMU switching as well:
6846 atomic_inc(&init_mm
.mm_count
);
6847 enter_lazy_tlb(&init_mm
, current
);
6850 * Make us the idle thread. Technically, schedule() should not be
6851 * called from this thread, however somewhere below it might be,
6852 * but because we are the idle thread, we just pick up running again
6853 * when this runqueue becomes "idle".
6855 init_idle(current
, smp_processor_id());
6858 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6859 void __might_sleep(char *file
, int line
)
6862 static unsigned long prev_jiffy
; /* ratelimiting */
6864 if ((in_atomic() || irqs_disabled()) &&
6865 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6866 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6868 prev_jiffy
= jiffies
;
6869 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6870 " context at %s:%d\n", file
, line
);
6871 printk("in_atomic():%d, irqs_disabled():%d\n",
6872 in_atomic(), irqs_disabled());
6877 EXPORT_SYMBOL(__might_sleep
);
6880 #ifdef CONFIG_MAGIC_SYSRQ
6881 void normalize_rt_tasks(void)
6883 struct prio_array
*array
;
6884 struct task_struct
*p
;
6885 unsigned long flags
;
6888 read_lock_irq(&tasklist_lock
);
6889 for_each_process(p
) {
6893 spin_lock_irqsave(&p
->pi_lock
, flags
);
6894 rq
= __task_rq_lock(p
);
6898 deactivate_task(p
, task_rq(p
));
6899 __setscheduler(p
, SCHED_NORMAL
, 0);
6901 __activate_task(p
, task_rq(p
));
6902 resched_task(rq
->curr
);
6905 __task_rq_unlock(rq
);
6906 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6908 read_unlock_irq(&tasklist_lock
);
6911 #endif /* CONFIG_MAGIC_SYSRQ */
6915 * These functions are only useful for the IA64 MCA handling.
6917 * They can only be called when the whole system has been
6918 * stopped - every CPU needs to be quiescent, and no scheduling
6919 * activity can take place. Using them for anything else would
6920 * be a serious bug, and as a result, they aren't even visible
6921 * under any other configuration.
6925 * curr_task - return the current task for a given cpu.
6926 * @cpu: the processor in question.
6928 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6930 struct task_struct
*curr_task(int cpu
)
6932 return cpu_curr(cpu
);
6936 * set_curr_task - set the current task for a given cpu.
6937 * @cpu: the processor in question.
6938 * @p: the task pointer to set.
6940 * Description: This function must only be used when non-maskable interrupts
6941 * are serviced on a separate stack. It allows the architecture to switch the
6942 * notion of the current task on a cpu in a non-blocking manner. This function
6943 * must be called with all CPU's synchronized, and interrupts disabled, the
6944 * and caller must save the original value of the current task (see
6945 * curr_task() above) and restore that value before reenabling interrupts and
6946 * re-starting the system.
6948 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6950 void set_curr_task(int cpu
, struct task_struct
*p
)