4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
22 #include <linux/module.h>
23 #include <linux/nmi.h>
24 #include <linux/init.h>
25 #include <asm/uaccess.h>
26 #include <linux/highmem.h>
27 #include <linux/smp_lock.h>
28 #include <asm/mmu_context.h>
29 #include <linux/interrupt.h>
30 #include <linux/capability.h>
31 #include <linux/completion.h>
32 #include <linux/kernel_stat.h>
33 #include <linux/debug_locks.h>
34 #include <linux/security.h>
35 #include <linux/notifier.h>
36 #include <linux/profile.h>
37 #include <linux/suspend.h>
38 #include <linux/vmalloc.h>
39 #include <linux/blkdev.h>
40 #include <linux/delay.h>
41 #include <linux/smp.h>
42 #include <linux/threads.h>
43 #include <linux/timer.h>
44 #include <linux/rcupdate.h>
45 #include <linux/cpu.h>
46 #include <linux/cpuset.h>
47 #include <linux/percpu.h>
48 #include <linux/kthread.h>
49 #include <linux/seq_file.h>
50 #include <linux/syscalls.h>
51 #include <linux/times.h>
52 #include <linux/acct.h>
53 #include <linux/kprobes.h>
56 #include <asm/unistd.h>
59 * Convert user-nice values [ -20 ... 0 ... 19 ]
60 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
63 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
64 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
65 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
68 * 'User priority' is the nice value converted to something we
69 * can work with better when scaling various scheduler parameters,
70 * it's a [ 0 ... 39 ] range.
72 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
73 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
74 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
77 * Some helpers for converting nanosecond timing to jiffy resolution
79 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
80 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
83 * These are the 'tuning knobs' of the scheduler:
85 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
86 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
87 * Timeslices get refilled after they expire.
89 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
90 #define DEF_TIMESLICE (100 * HZ / 1000)
91 #define ON_RUNQUEUE_WEIGHT 30
92 #define CHILD_PENALTY 95
93 #define PARENT_PENALTY 100
95 #define PRIO_BONUS_RATIO 25
96 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
97 #define INTERACTIVE_DELTA 2
98 #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
99 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
100 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
103 * If a task is 'interactive' then we reinsert it in the active
104 * array after it has expired its current timeslice. (it will not
105 * continue to run immediately, it will still roundrobin with
106 * other interactive tasks.)
108 * This part scales the interactivity limit depending on niceness.
110 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
111 * Here are a few examples of different nice levels:
113 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
114 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
115 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
116 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
117 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
119 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
120 * priority range a task can explore, a value of '1' means the
121 * task is rated interactive.)
123 * Ie. nice +19 tasks can never get 'interactive' enough to be
124 * reinserted into the active array. And only heavily CPU-hog nice -20
125 * tasks will be expired. Default nice 0 tasks are somewhere between,
126 * it takes some effort for them to get interactive, but it's not
130 #define CURRENT_BONUS(p) \
131 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
134 #define GRANULARITY (10 * HZ / 1000 ? : 1)
137 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
138 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
141 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
142 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
145 #define SCALE(v1,v1_max,v2_max) \
146 (v1) * (v2_max) / (v1_max)
149 (SCALE(TASK_NICE(p) + 20, 40, MAX_BONUS) - 20 * MAX_BONUS / 40 + \
152 #define TASK_INTERACTIVE(p) \
153 ((p)->prio <= (p)->static_prio - DELTA(p))
155 #define INTERACTIVE_SLEEP(p) \
156 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
157 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
159 #define TASK_PREEMPTS_CURR(p, rq) \
160 ((p)->prio < (rq)->curr->prio)
163 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
164 * to time slice values: [800ms ... 100ms ... 5ms]
166 * The higher a thread's priority, the bigger timeslices
167 * it gets during one round of execution. But even the lowest
168 * priority thread gets MIN_TIMESLICE worth of execution time.
171 #define SCALE_PRIO(x, prio) \
172 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
174 static unsigned int static_prio_timeslice(int static_prio
)
176 if (static_prio
< NICE_TO_PRIO(0))
177 return SCALE_PRIO(DEF_TIMESLICE
* 4, static_prio
);
179 return SCALE_PRIO(DEF_TIMESLICE
, static_prio
);
182 static inline unsigned int task_timeslice(task_t
*p
)
184 return static_prio_timeslice(p
->static_prio
);
187 #define task_hot(p, now, sd) ((long long) ((now) - (p)->last_ran) \
188 < (long long) (sd)->cache_hot_time)
191 * These are the runqueue data structures:
194 typedef struct runqueue runqueue_t
;
197 unsigned int nr_active
;
198 DECLARE_BITMAP(bitmap
, MAX_PRIO
+1); /* include 1 bit for delimiter */
199 struct list_head queue
[MAX_PRIO
];
203 * This is the main, per-CPU runqueue data structure.
205 * Locking rule: those places that want to lock multiple runqueues
206 * (such as the load balancing or the thread migration code), lock
207 * acquire operations must be ordered by ascending &runqueue.
213 * nr_running and cpu_load should be in the same cacheline because
214 * remote CPUs use both these fields when doing load calculation.
216 unsigned long nr_running
;
217 unsigned long raw_weighted_load
;
219 unsigned long cpu_load
[3];
221 unsigned long long nr_switches
;
224 * This is part of a global counter where only the total sum
225 * over all CPUs matters. A task can increase this counter on
226 * one CPU and if it got migrated afterwards it may decrease
227 * it on another CPU. Always updated under the runqueue lock:
229 unsigned long nr_uninterruptible
;
231 unsigned long expired_timestamp
;
232 unsigned long long timestamp_last_tick
;
234 struct mm_struct
*prev_mm
;
235 prio_array_t
*active
, *expired
, arrays
[2];
236 int best_expired_prio
;
240 struct sched_domain
*sd
;
242 /* For active balancing */
246 task_t
*migration_thread
;
247 struct list_head migration_queue
;
250 #ifdef CONFIG_SCHEDSTATS
252 struct sched_info rq_sched_info
;
254 /* sys_sched_yield() stats */
255 unsigned long yld_exp_empty
;
256 unsigned long yld_act_empty
;
257 unsigned long yld_both_empty
;
258 unsigned long yld_cnt
;
260 /* schedule() stats */
261 unsigned long sched_switch
;
262 unsigned long sched_cnt
;
263 unsigned long sched_goidle
;
265 /* try_to_wake_up() stats */
266 unsigned long ttwu_cnt
;
267 unsigned long ttwu_local
;
271 static DEFINE_PER_CPU(struct runqueue
, runqueues
);
274 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
275 * See detach_destroy_domains: synchronize_sched for details.
277 * The domain tree of any CPU may only be accessed from within
278 * preempt-disabled sections.
280 #define for_each_domain(cpu, domain) \
281 for (domain = rcu_dereference(cpu_rq(cpu)->sd); domain; domain = domain->parent)
283 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
284 #define this_rq() (&__get_cpu_var(runqueues))
285 #define task_rq(p) cpu_rq(task_cpu(p))
286 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
288 #ifndef prepare_arch_switch
289 # define prepare_arch_switch(next) do { } while (0)
291 #ifndef finish_arch_switch
292 # define finish_arch_switch(prev) do { } while (0)
295 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
296 static inline int task_running(runqueue_t
*rq
, task_t
*p
)
298 return rq
->curr
== p
;
301 static inline void prepare_lock_switch(runqueue_t
*rq
, task_t
*next
)
305 static inline void finish_lock_switch(runqueue_t
*rq
, task_t
*prev
)
307 #ifdef CONFIG_DEBUG_SPINLOCK
308 /* this is a valid case when another task releases the spinlock */
309 rq
->lock
.owner
= current
;
312 * If we are tracking spinlock dependencies then we have to
313 * fix up the runqueue lock - which gets 'carried over' from
316 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
318 spin_unlock_irq(&rq
->lock
);
321 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
322 static inline int task_running(runqueue_t
*rq
, task_t
*p
)
327 return rq
->curr
== p
;
331 static inline void prepare_lock_switch(runqueue_t
*rq
, task_t
*next
)
335 * We can optimise this out completely for !SMP, because the
336 * SMP rebalancing from interrupt is the only thing that cares
341 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
342 spin_unlock_irq(&rq
->lock
);
344 spin_unlock(&rq
->lock
);
348 static inline void finish_lock_switch(runqueue_t
*rq
, task_t
*prev
)
352 * After ->oncpu is cleared, the task can be moved to a different CPU.
353 * We must ensure this doesn't happen until the switch is completely
359 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
363 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
366 * __task_rq_lock - lock the runqueue a given task resides on.
367 * Must be called interrupts disabled.
369 static inline runqueue_t
*__task_rq_lock(task_t
*p
)
376 spin_lock(&rq
->lock
);
377 if (unlikely(rq
!= task_rq(p
))) {
378 spin_unlock(&rq
->lock
);
379 goto repeat_lock_task
;
385 * task_rq_lock - lock the runqueue a given task resides on and disable
386 * interrupts. Note the ordering: we can safely lookup the task_rq without
387 * explicitly disabling preemption.
389 static runqueue_t
*task_rq_lock(task_t
*p
, unsigned long *flags
)
395 local_irq_save(*flags
);
397 spin_lock(&rq
->lock
);
398 if (unlikely(rq
!= task_rq(p
))) {
399 spin_unlock_irqrestore(&rq
->lock
, *flags
);
400 goto repeat_lock_task
;
405 static inline void __task_rq_unlock(runqueue_t
*rq
)
408 spin_unlock(&rq
->lock
);
411 static inline void task_rq_unlock(runqueue_t
*rq
, unsigned long *flags
)
414 spin_unlock_irqrestore(&rq
->lock
, *flags
);
417 #ifdef CONFIG_SCHEDSTATS
419 * bump this up when changing the output format or the meaning of an existing
420 * format, so that tools can adapt (or abort)
422 #define SCHEDSTAT_VERSION 12
424 static int show_schedstat(struct seq_file
*seq
, void *v
)
428 seq_printf(seq
, "version %d\n", SCHEDSTAT_VERSION
);
429 seq_printf(seq
, "timestamp %lu\n", jiffies
);
430 for_each_online_cpu(cpu
) {
431 runqueue_t
*rq
= cpu_rq(cpu
);
433 struct sched_domain
*sd
;
437 /* runqueue-specific stats */
439 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
440 cpu
, rq
->yld_both_empty
,
441 rq
->yld_act_empty
, rq
->yld_exp_empty
, rq
->yld_cnt
,
442 rq
->sched_switch
, rq
->sched_cnt
, rq
->sched_goidle
,
443 rq
->ttwu_cnt
, rq
->ttwu_local
,
444 rq
->rq_sched_info
.cpu_time
,
445 rq
->rq_sched_info
.run_delay
, rq
->rq_sched_info
.pcnt
);
447 seq_printf(seq
, "\n");
450 /* domain-specific stats */
452 for_each_domain(cpu
, sd
) {
453 enum idle_type itype
;
454 char mask_str
[NR_CPUS
];
456 cpumask_scnprintf(mask_str
, NR_CPUS
, sd
->span
);
457 seq_printf(seq
, "domain%d %s", dcnt
++, mask_str
);
458 for (itype
= SCHED_IDLE
; itype
< MAX_IDLE_TYPES
;
460 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu %lu",
462 sd
->lb_balanced
[itype
],
463 sd
->lb_failed
[itype
],
464 sd
->lb_imbalance
[itype
],
465 sd
->lb_gained
[itype
],
466 sd
->lb_hot_gained
[itype
],
467 sd
->lb_nobusyq
[itype
],
468 sd
->lb_nobusyg
[itype
]);
470 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
471 sd
->alb_cnt
, sd
->alb_failed
, sd
->alb_pushed
,
472 sd
->sbe_cnt
, sd
->sbe_balanced
, sd
->sbe_pushed
,
473 sd
->sbf_cnt
, sd
->sbf_balanced
, sd
->sbf_pushed
,
474 sd
->ttwu_wake_remote
, sd
->ttwu_move_affine
, sd
->ttwu_move_balance
);
482 static int schedstat_open(struct inode
*inode
, struct file
*file
)
484 unsigned int size
= PAGE_SIZE
* (1 + num_online_cpus() / 32);
485 char *buf
= kmalloc(size
, GFP_KERNEL
);
491 res
= single_open(file
, show_schedstat
, NULL
);
493 m
= file
->private_data
;
501 struct file_operations proc_schedstat_operations
= {
502 .open
= schedstat_open
,
505 .release
= single_release
,
508 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
509 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
510 #else /* !CONFIG_SCHEDSTATS */
511 # define schedstat_inc(rq, field) do { } while (0)
512 # define schedstat_add(rq, field, amt) do { } while (0)
516 * rq_lock - lock a given runqueue and disable interrupts.
518 static inline runqueue_t
*this_rq_lock(void)
525 spin_lock(&rq
->lock
);
530 #ifdef CONFIG_SCHEDSTATS
532 * Called when a process is dequeued from the active array and given
533 * the cpu. We should note that with the exception of interactive
534 * tasks, the expired queue will become the active queue after the active
535 * queue is empty, without explicitly dequeuing and requeuing tasks in the
536 * expired queue. (Interactive tasks may be requeued directly to the
537 * active queue, thus delaying tasks in the expired queue from running;
538 * see scheduler_tick()).
540 * This function is only called from sched_info_arrive(), rather than
541 * dequeue_task(). Even though a task may be queued and dequeued multiple
542 * times as it is shuffled about, we're really interested in knowing how
543 * long it was from the *first* time it was queued to the time that it
546 static inline void sched_info_dequeued(task_t
*t
)
548 t
->sched_info
.last_queued
= 0;
552 * Called when a task finally hits the cpu. We can now calculate how
553 * long it was waiting to run. We also note when it began so that we
554 * can keep stats on how long its timeslice is.
556 static void sched_info_arrive(task_t
*t
)
558 unsigned long now
= jiffies
, diff
= 0;
559 struct runqueue
*rq
= task_rq(t
);
561 if (t
->sched_info
.last_queued
)
562 diff
= now
- t
->sched_info
.last_queued
;
563 sched_info_dequeued(t
);
564 t
->sched_info
.run_delay
+= diff
;
565 t
->sched_info
.last_arrival
= now
;
566 t
->sched_info
.pcnt
++;
571 rq
->rq_sched_info
.run_delay
+= diff
;
572 rq
->rq_sched_info
.pcnt
++;
576 * Called when a process is queued into either the active or expired
577 * array. The time is noted and later used to determine how long we
578 * had to wait for us to reach the cpu. Since the expired queue will
579 * become the active queue after active queue is empty, without dequeuing
580 * and requeuing any tasks, we are interested in queuing to either. It
581 * is unusual but not impossible for tasks to be dequeued and immediately
582 * requeued in the same or another array: this can happen in sched_yield(),
583 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
586 * This function is only called from enqueue_task(), but also only updates
587 * the timestamp if it is already not set. It's assumed that
588 * sched_info_dequeued() will clear that stamp when appropriate.
590 static inline void sched_info_queued(task_t
*t
)
592 if (!t
->sched_info
.last_queued
)
593 t
->sched_info
.last_queued
= jiffies
;
597 * Called when a process ceases being the active-running process, either
598 * voluntarily or involuntarily. Now we can calculate how long we ran.
600 static inline void sched_info_depart(task_t
*t
)
602 struct runqueue
*rq
= task_rq(t
);
603 unsigned long diff
= jiffies
- t
->sched_info
.last_arrival
;
605 t
->sched_info
.cpu_time
+= diff
;
608 rq
->rq_sched_info
.cpu_time
+= diff
;
612 * Called when tasks are switched involuntarily due, typically, to expiring
613 * their time slice. (This may also be called when switching to or from
614 * the idle task.) We are only called when prev != next.
616 static inline void sched_info_switch(task_t
*prev
, task_t
*next
)
618 struct runqueue
*rq
= task_rq(prev
);
621 * prev now departs the cpu. It's not interesting to record
622 * stats about how efficient we were at scheduling the idle
625 if (prev
!= rq
->idle
)
626 sched_info_depart(prev
);
628 if (next
!= rq
->idle
)
629 sched_info_arrive(next
);
632 #define sched_info_queued(t) do { } while (0)
633 #define sched_info_switch(t, next) do { } while (0)
634 #endif /* CONFIG_SCHEDSTATS */
637 * Adding/removing a task to/from a priority array:
639 static void dequeue_task(struct task_struct
*p
, prio_array_t
*array
)
642 list_del(&p
->run_list
);
643 if (list_empty(array
->queue
+ p
->prio
))
644 __clear_bit(p
->prio
, array
->bitmap
);
647 static void enqueue_task(struct task_struct
*p
, prio_array_t
*array
)
649 sched_info_queued(p
);
650 list_add_tail(&p
->run_list
, array
->queue
+ p
->prio
);
651 __set_bit(p
->prio
, array
->bitmap
);
657 * Put task to the end of the run list without the overhead of dequeue
658 * followed by enqueue.
660 static void requeue_task(struct task_struct
*p
, prio_array_t
*array
)
662 list_move_tail(&p
->run_list
, array
->queue
+ p
->prio
);
665 static inline void enqueue_task_head(struct task_struct
*p
, prio_array_t
*array
)
667 list_add(&p
->run_list
, array
->queue
+ p
->prio
);
668 __set_bit(p
->prio
, array
->bitmap
);
674 * __normal_prio - return the priority that is based on the static
675 * priority but is modified by bonuses/penalties.
677 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
678 * into the -5 ... 0 ... +5 bonus/penalty range.
680 * We use 25% of the full 0...39 priority range so that:
682 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
683 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
685 * Both properties are important to certain workloads.
688 static inline int __normal_prio(task_t
*p
)
692 bonus
= CURRENT_BONUS(p
) - MAX_BONUS
/ 2;
694 prio
= p
->static_prio
- bonus
;
695 if (prio
< MAX_RT_PRIO
)
697 if (prio
> MAX_PRIO
-1)
703 * To aid in avoiding the subversion of "niceness" due to uneven distribution
704 * of tasks with abnormal "nice" values across CPUs the contribution that
705 * each task makes to its run queue's load is weighted according to its
706 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
707 * scaled version of the new time slice allocation that they receive on time
712 * Assume: static_prio_timeslice(NICE_TO_PRIO(0)) == DEF_TIMESLICE
713 * If static_prio_timeslice() is ever changed to break this assumption then
714 * this code will need modification
716 #define TIME_SLICE_NICE_ZERO DEF_TIMESLICE
717 #define LOAD_WEIGHT(lp) \
718 (((lp) * SCHED_LOAD_SCALE) / TIME_SLICE_NICE_ZERO)
719 #define PRIO_TO_LOAD_WEIGHT(prio) \
720 LOAD_WEIGHT(static_prio_timeslice(prio))
721 #define RTPRIO_TO_LOAD_WEIGHT(rp) \
722 (PRIO_TO_LOAD_WEIGHT(MAX_RT_PRIO) + LOAD_WEIGHT(rp))
724 static void set_load_weight(task_t
*p
)
726 if (has_rt_policy(p
)) {
728 if (p
== task_rq(p
)->migration_thread
)
730 * The migration thread does the actual balancing.
731 * Giving its load any weight will skew balancing
737 p
->load_weight
= RTPRIO_TO_LOAD_WEIGHT(p
->rt_priority
);
739 p
->load_weight
= PRIO_TO_LOAD_WEIGHT(p
->static_prio
);
742 static inline void inc_raw_weighted_load(runqueue_t
*rq
, const task_t
*p
)
744 rq
->raw_weighted_load
+= p
->load_weight
;
747 static inline void dec_raw_weighted_load(runqueue_t
*rq
, const task_t
*p
)
749 rq
->raw_weighted_load
-= p
->load_weight
;
752 static inline void inc_nr_running(task_t
*p
, runqueue_t
*rq
)
755 inc_raw_weighted_load(rq
, p
);
758 static inline void dec_nr_running(task_t
*p
, runqueue_t
*rq
)
761 dec_raw_weighted_load(rq
, p
);
765 * Calculate the expected normal priority: i.e. priority
766 * without taking RT-inheritance into account. Might be
767 * boosted by interactivity modifiers. Changes upon fork,
768 * setprio syscalls, and whenever the interactivity
769 * estimator recalculates.
771 static inline int normal_prio(task_t
*p
)
775 if (has_rt_policy(p
))
776 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
778 prio
= __normal_prio(p
);
783 * Calculate the current priority, i.e. the priority
784 * taken into account by the scheduler. This value might
785 * be boosted by RT tasks, or might be boosted by
786 * interactivity modifiers. Will be RT if the task got
787 * RT-boosted. If not then it returns p->normal_prio.
789 static int effective_prio(task_t
*p
)
791 p
->normal_prio
= normal_prio(p
);
793 * If we are RT tasks or we were boosted to RT priority,
794 * keep the priority unchanged. Otherwise, update priority
795 * to the normal priority:
797 if (!rt_prio(p
->prio
))
798 return p
->normal_prio
;
803 * __activate_task - move a task to the runqueue.
805 static void __activate_task(task_t
*p
, runqueue_t
*rq
)
807 prio_array_t
*target
= rq
->active
;
810 target
= rq
->expired
;
811 enqueue_task(p
, target
);
812 inc_nr_running(p
, rq
);
816 * __activate_idle_task - move idle task to the _front_ of runqueue.
818 static inline void __activate_idle_task(task_t
*p
, runqueue_t
*rq
)
820 enqueue_task_head(p
, rq
->active
);
821 inc_nr_running(p
, rq
);
825 * Recalculate p->normal_prio and p->prio after having slept,
826 * updating the sleep-average too:
828 static int recalc_task_prio(task_t
*p
, unsigned long long now
)
830 /* Caller must always ensure 'now >= p->timestamp' */
831 unsigned long sleep_time
= now
- p
->timestamp
;
836 if (likely(sleep_time
> 0)) {
838 * This ceiling is set to the lowest priority that would allow
839 * a task to be reinserted into the active array on timeslice
842 unsigned long ceiling
= INTERACTIVE_SLEEP(p
);
844 if (p
->mm
&& sleep_time
> ceiling
&& p
->sleep_avg
< ceiling
) {
846 * Prevents user tasks from achieving best priority
847 * with one single large enough sleep.
849 p
->sleep_avg
= ceiling
;
851 * Using INTERACTIVE_SLEEP() as a ceiling places a
852 * nice(0) task 1ms sleep away from promotion, and
853 * gives it 700ms to round-robin with no chance of
854 * being demoted. This is more than generous, so
855 * mark this sleep as non-interactive to prevent the
856 * on-runqueue bonus logic from intervening should
857 * this task not receive cpu immediately.
859 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
862 * Tasks waking from uninterruptible sleep are
863 * limited in their sleep_avg rise as they
864 * are likely to be waiting on I/O
866 if (p
->sleep_type
== SLEEP_NONINTERACTIVE
&& p
->mm
) {
867 if (p
->sleep_avg
>= ceiling
)
869 else if (p
->sleep_avg
+ sleep_time
>=
871 p
->sleep_avg
= ceiling
;
877 * This code gives a bonus to interactive tasks.
879 * The boost works by updating the 'average sleep time'
880 * value here, based on ->timestamp. The more time a
881 * task spends sleeping, the higher the average gets -
882 * and the higher the priority boost gets as well.
884 p
->sleep_avg
+= sleep_time
;
887 if (p
->sleep_avg
> NS_MAX_SLEEP_AVG
)
888 p
->sleep_avg
= NS_MAX_SLEEP_AVG
;
891 return effective_prio(p
);
895 * activate_task - move a task to the runqueue and do priority recalculation
897 * Update all the scheduling statistics stuff. (sleep average
898 * calculation, priority modifiers, etc.)
900 static void activate_task(task_t
*p
, runqueue_t
*rq
, int local
)
902 unsigned long long now
;
907 /* Compensate for drifting sched_clock */
908 runqueue_t
*this_rq
= this_rq();
909 now
= (now
- this_rq
->timestamp_last_tick
)
910 + rq
->timestamp_last_tick
;
915 p
->prio
= recalc_task_prio(p
, now
);
918 * This checks to make sure it's not an uninterruptible task
919 * that is now waking up.
921 if (p
->sleep_type
== SLEEP_NORMAL
) {
923 * Tasks which were woken up by interrupts (ie. hw events)
924 * are most likely of interactive nature. So we give them
925 * the credit of extending their sleep time to the period
926 * of time they spend on the runqueue, waiting for execution
927 * on a CPU, first time around:
930 p
->sleep_type
= SLEEP_INTERRUPTED
;
933 * Normal first-time wakeups get a credit too for
934 * on-runqueue time, but it will be weighted down:
936 p
->sleep_type
= SLEEP_INTERACTIVE
;
941 __activate_task(p
, rq
);
945 * deactivate_task - remove a task from the runqueue.
947 static void deactivate_task(struct task_struct
*p
, runqueue_t
*rq
)
949 dec_nr_running(p
, rq
);
950 dequeue_task(p
, p
->array
);
955 * resched_task - mark a task 'to be rescheduled now'.
957 * On UP this means the setting of the need_resched flag, on SMP it
958 * might also involve a cross-CPU call to trigger the scheduler on
963 #ifndef tsk_is_polling
964 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
967 static void resched_task(task_t
*p
)
971 assert_spin_locked(&task_rq(p
)->lock
);
973 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
976 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
979 if (cpu
== smp_processor_id())
982 /* NEED_RESCHED must be visible before we test polling */
984 if (!tsk_is_polling(p
))
985 smp_send_reschedule(cpu
);
988 static inline void resched_task(task_t
*p
)
990 assert_spin_locked(&task_rq(p
)->lock
);
991 set_tsk_need_resched(p
);
996 * task_curr - is this task currently executing on a CPU?
997 * @p: the task in question.
999 inline int task_curr(const task_t
*p
)
1001 return cpu_curr(task_cpu(p
)) == p
;
1004 /* Used instead of source_load when we know the type == 0 */
1005 unsigned long weighted_cpuload(const int cpu
)
1007 return cpu_rq(cpu
)->raw_weighted_load
;
1012 struct list_head list
;
1017 struct completion done
;
1021 * The task's runqueue lock must be held.
1022 * Returns true if you have to wait for migration thread.
1024 static int migrate_task(task_t
*p
, int dest_cpu
, migration_req_t
*req
)
1026 runqueue_t
*rq
= task_rq(p
);
1029 * If the task is not on a runqueue (and not running), then
1030 * it is sufficient to simply update the task's cpu field.
1032 if (!p
->array
&& !task_running(rq
, p
)) {
1033 set_task_cpu(p
, dest_cpu
);
1037 init_completion(&req
->done
);
1039 req
->dest_cpu
= dest_cpu
;
1040 list_add(&req
->list
, &rq
->migration_queue
);
1045 * wait_task_inactive - wait for a thread to unschedule.
1047 * The caller must ensure that the task *will* unschedule sometime soon,
1048 * else this function might spin for a *long* time. This function can't
1049 * be called with interrupts off, or it may introduce deadlock with
1050 * smp_call_function() if an IPI is sent by the same process we are
1051 * waiting to become inactive.
1053 void wait_task_inactive(task_t
*p
)
1055 unsigned long flags
;
1060 rq
= task_rq_lock(p
, &flags
);
1061 /* Must be off runqueue entirely, not preempted. */
1062 if (unlikely(p
->array
|| task_running(rq
, p
))) {
1063 /* If it's preempted, we yield. It could be a while. */
1064 preempted
= !task_running(rq
, p
);
1065 task_rq_unlock(rq
, &flags
);
1071 task_rq_unlock(rq
, &flags
);
1075 * kick_process - kick a running thread to enter/exit the kernel
1076 * @p: the to-be-kicked thread
1078 * Cause a process which is running on another CPU to enter
1079 * kernel-mode, without any delay. (to get signals handled.)
1081 * NOTE: this function doesnt have to take the runqueue lock,
1082 * because all it wants to ensure is that the remote task enters
1083 * the kernel. If the IPI races and the task has been migrated
1084 * to another CPU then no harm is done and the purpose has been
1087 void kick_process(task_t
*p
)
1093 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1094 smp_send_reschedule(cpu
);
1099 * Return a low guess at the load of a migration-source cpu weighted
1100 * according to the scheduling class and "nice" value.
1102 * We want to under-estimate the load of migration sources, to
1103 * balance conservatively.
1105 static inline unsigned long source_load(int cpu
, int type
)
1107 runqueue_t
*rq
= cpu_rq(cpu
);
1110 return rq
->raw_weighted_load
;
1112 return min(rq
->cpu_load
[type
-1], rq
->raw_weighted_load
);
1116 * Return a high guess at the load of a migration-target cpu weighted
1117 * according to the scheduling class and "nice" value.
1119 static inline unsigned long target_load(int cpu
, int type
)
1121 runqueue_t
*rq
= cpu_rq(cpu
);
1124 return rq
->raw_weighted_load
;
1126 return max(rq
->cpu_load
[type
-1], rq
->raw_weighted_load
);
1130 * Return the average load per task on the cpu's run queue
1132 static inline unsigned long cpu_avg_load_per_task(int cpu
)
1134 runqueue_t
*rq
= cpu_rq(cpu
);
1135 unsigned long n
= rq
->nr_running
;
1137 return n
? rq
->raw_weighted_load
/ n
: SCHED_LOAD_SCALE
;
1141 * find_idlest_group finds and returns the least busy CPU group within the
1144 static struct sched_group
*
1145 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1147 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1148 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1149 int load_idx
= sd
->forkexec_idx
;
1150 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1153 unsigned long load
, avg_load
;
1157 /* Skip over this group if it has no CPUs allowed */
1158 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1161 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1163 /* Tally up the load of all CPUs in the group */
1166 for_each_cpu_mask(i
, group
->cpumask
) {
1167 /* Bias balancing toward cpus of our domain */
1169 load
= source_load(i
, load_idx
);
1171 load
= target_load(i
, load_idx
);
1176 /* Adjust by relative CPU power of the group */
1177 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
1180 this_load
= avg_load
;
1182 } else if (avg_load
< min_load
) {
1183 min_load
= avg_load
;
1187 group
= group
->next
;
1188 } while (group
!= sd
->groups
);
1190 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1196 * find_idlest_queue - find the idlest runqueue among the cpus in group.
1199 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1202 unsigned long load
, min_load
= ULONG_MAX
;
1206 /* Traverse only the allowed CPUs */
1207 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1209 for_each_cpu_mask(i
, tmp
) {
1210 load
= weighted_cpuload(i
);
1212 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1222 * sched_balance_self: balance the current task (running on cpu) in domains
1223 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1226 * Balance, ie. select the least loaded group.
1228 * Returns the target CPU number, or the same CPU if no balancing is needed.
1230 * preempt must be disabled.
1232 static int sched_balance_self(int cpu
, int flag
)
1234 struct task_struct
*t
= current
;
1235 struct sched_domain
*tmp
, *sd
= NULL
;
1237 for_each_domain(cpu
, tmp
) {
1239 * If power savings logic is enabled for a domain, stop there.
1241 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1243 if (tmp
->flags
& flag
)
1249 struct sched_group
*group
;
1254 group
= find_idlest_group(sd
, t
, cpu
);
1258 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1259 if (new_cpu
== -1 || new_cpu
== cpu
)
1262 /* Now try balancing at a lower domain level */
1266 weight
= cpus_weight(span
);
1267 for_each_domain(cpu
, tmp
) {
1268 if (weight
<= cpus_weight(tmp
->span
))
1270 if (tmp
->flags
& flag
)
1273 /* while loop will break here if sd == NULL */
1279 #endif /* CONFIG_SMP */
1282 * wake_idle() will wake a task on an idle cpu if task->cpu is
1283 * not idle and an idle cpu is available. The span of cpus to
1284 * search starts with cpus closest then further out as needed,
1285 * so we always favor a closer, idle cpu.
1287 * Returns the CPU we should wake onto.
1289 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1290 static int wake_idle(int cpu
, task_t
*p
)
1293 struct sched_domain
*sd
;
1299 for_each_domain(cpu
, sd
) {
1300 if (sd
->flags
& SD_WAKE_IDLE
) {
1301 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1302 for_each_cpu_mask(i
, tmp
) {
1313 static inline int wake_idle(int cpu
, task_t
*p
)
1320 * try_to_wake_up - wake up a thread
1321 * @p: the to-be-woken-up thread
1322 * @state: the mask of task states that can be woken
1323 * @sync: do a synchronous wakeup?
1325 * Put it on the run-queue if it's not already there. The "current"
1326 * thread is always on the run-queue (except when the actual
1327 * re-schedule is in progress), and as such you're allowed to do
1328 * the simpler "current->state = TASK_RUNNING" to mark yourself
1329 * runnable without the overhead of this.
1331 * returns failure only if the task is already active.
1333 static int try_to_wake_up(task_t
*p
, unsigned int state
, int sync
)
1335 int cpu
, this_cpu
, success
= 0;
1336 unsigned long flags
;
1340 unsigned long load
, this_load
;
1341 struct sched_domain
*sd
, *this_sd
= NULL
;
1345 rq
= task_rq_lock(p
, &flags
);
1346 old_state
= p
->state
;
1347 if (!(old_state
& state
))
1354 this_cpu
= smp_processor_id();
1357 if (unlikely(task_running(rq
, p
)))
1362 schedstat_inc(rq
, ttwu_cnt
);
1363 if (cpu
== this_cpu
) {
1364 schedstat_inc(rq
, ttwu_local
);
1368 for_each_domain(this_cpu
, sd
) {
1369 if (cpu_isset(cpu
, sd
->span
)) {
1370 schedstat_inc(sd
, ttwu_wake_remote
);
1376 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1380 * Check for affine wakeup and passive balancing possibilities.
1383 int idx
= this_sd
->wake_idx
;
1384 unsigned int imbalance
;
1386 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1388 load
= source_load(cpu
, idx
);
1389 this_load
= target_load(this_cpu
, idx
);
1391 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1393 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1394 unsigned long tl
= this_load
;
1395 unsigned long tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1398 * If sync wakeup then subtract the (maximum possible)
1399 * effect of the currently running task from the load
1400 * of the current CPU:
1403 tl
-= current
->load_weight
;
1406 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1407 100*(tl
+ p
->load_weight
) <= imbalance
*load
) {
1409 * This domain has SD_WAKE_AFFINE and
1410 * p is cache cold in this domain, and
1411 * there is no bad imbalance.
1413 schedstat_inc(this_sd
, ttwu_move_affine
);
1419 * Start passive balancing when half the imbalance_pct
1422 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1423 if (imbalance
*this_load
<= 100*load
) {
1424 schedstat_inc(this_sd
, ttwu_move_balance
);
1430 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1432 new_cpu
= wake_idle(new_cpu
, p
);
1433 if (new_cpu
!= cpu
) {
1434 set_task_cpu(p
, new_cpu
);
1435 task_rq_unlock(rq
, &flags
);
1436 /* might preempt at this point */
1437 rq
= task_rq_lock(p
, &flags
);
1438 old_state
= p
->state
;
1439 if (!(old_state
& state
))
1444 this_cpu
= smp_processor_id();
1449 #endif /* CONFIG_SMP */
1450 if (old_state
== TASK_UNINTERRUPTIBLE
) {
1451 rq
->nr_uninterruptible
--;
1453 * Tasks on involuntary sleep don't earn
1454 * sleep_avg beyond just interactive state.
1456 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
1460 * Tasks that have marked their sleep as noninteractive get
1461 * woken up with their sleep average not weighted in an
1464 if (old_state
& TASK_NONINTERACTIVE
)
1465 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
1468 activate_task(p
, rq
, cpu
== this_cpu
);
1470 * Sync wakeups (i.e. those types of wakeups where the waker
1471 * has indicated that it will leave the CPU in short order)
1472 * don't trigger a preemption, if the woken up task will run on
1473 * this cpu. (in this case the 'I will reschedule' promise of
1474 * the waker guarantees that the freshly woken up task is going
1475 * to be considered on this CPU.)
1477 if (!sync
|| cpu
!= this_cpu
) {
1478 if (TASK_PREEMPTS_CURR(p
, rq
))
1479 resched_task(rq
->curr
);
1484 p
->state
= TASK_RUNNING
;
1486 task_rq_unlock(rq
, &flags
);
1491 int fastcall
wake_up_process(task_t
*p
)
1493 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1494 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1497 EXPORT_SYMBOL(wake_up_process
);
1499 int fastcall
wake_up_state(task_t
*p
, unsigned int state
)
1501 return try_to_wake_up(p
, state
, 0);
1505 * Perform scheduler related setup for a newly forked process p.
1506 * p is forked by current.
1508 void fastcall
sched_fork(task_t
*p
, int clone_flags
)
1510 int cpu
= get_cpu();
1513 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1515 set_task_cpu(p
, cpu
);
1518 * We mark the process as running here, but have not actually
1519 * inserted it onto the runqueue yet. This guarantees that
1520 * nobody will actually run it, and a signal or other external
1521 * event cannot wake it up and insert it on the runqueue either.
1523 p
->state
= TASK_RUNNING
;
1526 * Make sure we do not leak PI boosting priority to the child:
1528 p
->prio
= current
->normal_prio
;
1530 INIT_LIST_HEAD(&p
->run_list
);
1532 #ifdef CONFIG_SCHEDSTATS
1533 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1535 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1538 #ifdef CONFIG_PREEMPT
1539 /* Want to start with kernel preemption disabled. */
1540 task_thread_info(p
)->preempt_count
= 1;
1543 * Share the timeslice between parent and child, thus the
1544 * total amount of pending timeslices in the system doesn't change,
1545 * resulting in more scheduling fairness.
1547 local_irq_disable();
1548 p
->time_slice
= (current
->time_slice
+ 1) >> 1;
1550 * The remainder of the first timeslice might be recovered by
1551 * the parent if the child exits early enough.
1553 p
->first_time_slice
= 1;
1554 current
->time_slice
>>= 1;
1555 p
->timestamp
= sched_clock();
1556 if (unlikely(!current
->time_slice
)) {
1558 * This case is rare, it happens when the parent has only
1559 * a single jiffy left from its timeslice. Taking the
1560 * runqueue lock is not a problem.
1562 current
->time_slice
= 1;
1570 * wake_up_new_task - wake up a newly created task for the first time.
1572 * This function will do some initial scheduler statistics housekeeping
1573 * that must be done for every newly created context, then puts the task
1574 * on the runqueue and wakes it.
1576 void fastcall
wake_up_new_task(task_t
*p
, unsigned long clone_flags
)
1578 unsigned long flags
;
1580 runqueue_t
*rq
, *this_rq
;
1582 rq
= task_rq_lock(p
, &flags
);
1583 BUG_ON(p
->state
!= TASK_RUNNING
);
1584 this_cpu
= smp_processor_id();
1588 * We decrease the sleep average of forking parents
1589 * and children as well, to keep max-interactive tasks
1590 * from forking tasks that are max-interactive. The parent
1591 * (current) is done further down, under its lock.
1593 p
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(p
) *
1594 CHILD_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1596 p
->prio
= effective_prio(p
);
1598 if (likely(cpu
== this_cpu
)) {
1599 if (!(clone_flags
& CLONE_VM
)) {
1601 * The VM isn't cloned, so we're in a good position to
1602 * do child-runs-first in anticipation of an exec. This
1603 * usually avoids a lot of COW overhead.
1605 if (unlikely(!current
->array
))
1606 __activate_task(p
, rq
);
1608 p
->prio
= current
->prio
;
1609 p
->normal_prio
= current
->normal_prio
;
1610 list_add_tail(&p
->run_list
, ¤t
->run_list
);
1611 p
->array
= current
->array
;
1612 p
->array
->nr_active
++;
1613 inc_nr_running(p
, rq
);
1617 /* Run child last */
1618 __activate_task(p
, rq
);
1620 * We skip the following code due to cpu == this_cpu
1622 * task_rq_unlock(rq, &flags);
1623 * this_rq = task_rq_lock(current, &flags);
1627 this_rq
= cpu_rq(this_cpu
);
1630 * Not the local CPU - must adjust timestamp. This should
1631 * get optimised away in the !CONFIG_SMP case.
1633 p
->timestamp
= (p
->timestamp
- this_rq
->timestamp_last_tick
)
1634 + rq
->timestamp_last_tick
;
1635 __activate_task(p
, rq
);
1636 if (TASK_PREEMPTS_CURR(p
, rq
))
1637 resched_task(rq
->curr
);
1640 * Parent and child are on different CPUs, now get the
1641 * parent runqueue to update the parent's ->sleep_avg:
1643 task_rq_unlock(rq
, &flags
);
1644 this_rq
= task_rq_lock(current
, &flags
);
1646 current
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(current
) *
1647 PARENT_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1648 task_rq_unlock(this_rq
, &flags
);
1652 * Potentially available exiting-child timeslices are
1653 * retrieved here - this way the parent does not get
1654 * penalized for creating too many threads.
1656 * (this cannot be used to 'generate' timeslices
1657 * artificially, because any timeslice recovered here
1658 * was given away by the parent in the first place.)
1660 void fastcall
sched_exit(task_t
*p
)
1662 unsigned long flags
;
1666 * If the child was a (relative-) CPU hog then decrease
1667 * the sleep_avg of the parent as well.
1669 rq
= task_rq_lock(p
->parent
, &flags
);
1670 if (p
->first_time_slice
&& task_cpu(p
) == task_cpu(p
->parent
)) {
1671 p
->parent
->time_slice
+= p
->time_slice
;
1672 if (unlikely(p
->parent
->time_slice
> task_timeslice(p
)))
1673 p
->parent
->time_slice
= task_timeslice(p
);
1675 if (p
->sleep_avg
< p
->parent
->sleep_avg
)
1676 p
->parent
->sleep_avg
= p
->parent
->sleep_avg
/
1677 (EXIT_WEIGHT
+ 1) * EXIT_WEIGHT
+ p
->sleep_avg
/
1679 task_rq_unlock(rq
, &flags
);
1683 * prepare_task_switch - prepare to switch tasks
1684 * @rq: the runqueue preparing to switch
1685 * @next: the task we are going to switch to.
1687 * This is called with the rq lock held and interrupts off. It must
1688 * be paired with a subsequent finish_task_switch after the context
1691 * prepare_task_switch sets up locking and calls architecture specific
1694 static inline void prepare_task_switch(runqueue_t
*rq
, task_t
*next
)
1696 prepare_lock_switch(rq
, next
);
1697 prepare_arch_switch(next
);
1701 * finish_task_switch - clean up after a task-switch
1702 * @rq: runqueue associated with task-switch
1703 * @prev: the thread we just switched away from.
1705 * finish_task_switch must be called after the context switch, paired
1706 * with a prepare_task_switch call before the context switch.
1707 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1708 * and do any other architecture-specific cleanup actions.
1710 * Note that we may have delayed dropping an mm in context_switch(). If
1711 * so, we finish that here outside of the runqueue lock. (Doing it
1712 * with the lock held can cause deadlocks; see schedule() for
1715 static inline void finish_task_switch(runqueue_t
*rq
, task_t
*prev
)
1716 __releases(rq
->lock
)
1718 struct mm_struct
*mm
= rq
->prev_mm
;
1719 unsigned long prev_task_flags
;
1724 * A task struct has one reference for the use as "current".
1725 * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1726 * calls schedule one last time. The schedule call will never return,
1727 * and the scheduled task must drop that reference.
1728 * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1729 * still held, otherwise prev could be scheduled on another cpu, die
1730 * there before we look at prev->state, and then the reference would
1732 * Manfred Spraul <manfred@colorfullife.com>
1734 prev_task_flags
= prev
->flags
;
1735 finish_arch_switch(prev
);
1736 finish_lock_switch(rq
, prev
);
1739 if (unlikely(prev_task_flags
& PF_DEAD
)) {
1741 * Remove function-return probe instances associated with this
1742 * task and put them back on the free list.
1744 kprobe_flush_task(prev
);
1745 put_task_struct(prev
);
1750 * schedule_tail - first thing a freshly forked thread must call.
1751 * @prev: the thread we just switched away from.
1753 asmlinkage
void schedule_tail(task_t
*prev
)
1754 __releases(rq
->lock
)
1756 runqueue_t
*rq
= this_rq();
1757 finish_task_switch(rq
, prev
);
1758 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1759 /* In this case, finish_task_switch does not reenable preemption */
1762 if (current
->set_child_tid
)
1763 put_user(current
->pid
, current
->set_child_tid
);
1767 * context_switch - switch to the new MM and the new
1768 * thread's register state.
1771 task_t
* context_switch(runqueue_t
*rq
, task_t
*prev
, task_t
*next
)
1773 struct mm_struct
*mm
= next
->mm
;
1774 struct mm_struct
*oldmm
= prev
->active_mm
;
1776 if (unlikely(!mm
)) {
1777 next
->active_mm
= oldmm
;
1778 atomic_inc(&oldmm
->mm_count
);
1779 enter_lazy_tlb(oldmm
, next
);
1781 switch_mm(oldmm
, mm
, next
);
1783 if (unlikely(!prev
->mm
)) {
1784 prev
->active_mm
= NULL
;
1785 WARN_ON(rq
->prev_mm
);
1786 rq
->prev_mm
= oldmm
;
1788 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1790 /* Here we just switch the register state and the stack. */
1791 switch_to(prev
, next
, prev
);
1797 * nr_running, nr_uninterruptible and nr_context_switches:
1799 * externally visible scheduler statistics: current number of runnable
1800 * threads, current number of uninterruptible-sleeping threads, total
1801 * number of context switches performed since bootup.
1803 unsigned long nr_running(void)
1805 unsigned long i
, sum
= 0;
1807 for_each_online_cpu(i
)
1808 sum
+= cpu_rq(i
)->nr_running
;
1813 unsigned long nr_uninterruptible(void)
1815 unsigned long i
, sum
= 0;
1817 for_each_possible_cpu(i
)
1818 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1821 * Since we read the counters lockless, it might be slightly
1822 * inaccurate. Do not allow it to go below zero though:
1824 if (unlikely((long)sum
< 0))
1830 unsigned long long nr_context_switches(void)
1833 unsigned long long sum
= 0;
1835 for_each_possible_cpu(i
)
1836 sum
+= cpu_rq(i
)->nr_switches
;
1841 unsigned long nr_iowait(void)
1843 unsigned long i
, sum
= 0;
1845 for_each_possible_cpu(i
)
1846 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1851 unsigned long nr_active(void)
1853 unsigned long i
, running
= 0, uninterruptible
= 0;
1855 for_each_online_cpu(i
) {
1856 running
+= cpu_rq(i
)->nr_running
;
1857 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
1860 if (unlikely((long)uninterruptible
< 0))
1861 uninterruptible
= 0;
1863 return running
+ uninterruptible
;
1869 * double_rq_lock - safely lock two runqueues
1871 * Note this does not disable interrupts like task_rq_lock,
1872 * you need to do so manually before calling.
1874 static void double_rq_lock(runqueue_t
*rq1
, runqueue_t
*rq2
)
1875 __acquires(rq1
->lock
)
1876 __acquires(rq2
->lock
)
1879 spin_lock(&rq1
->lock
);
1880 __acquire(rq2
->lock
); /* Fake it out ;) */
1883 spin_lock(&rq1
->lock
);
1884 spin_lock(&rq2
->lock
);
1886 spin_lock(&rq2
->lock
);
1887 spin_lock(&rq1
->lock
);
1893 * double_rq_unlock - safely unlock two runqueues
1895 * Note this does not restore interrupts like task_rq_unlock,
1896 * you need to do so manually after calling.
1898 static void double_rq_unlock(runqueue_t
*rq1
, runqueue_t
*rq2
)
1899 __releases(rq1
->lock
)
1900 __releases(rq2
->lock
)
1902 spin_unlock(&rq1
->lock
);
1904 spin_unlock(&rq2
->lock
);
1906 __release(rq2
->lock
);
1910 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1912 static void double_lock_balance(runqueue_t
*this_rq
, runqueue_t
*busiest
)
1913 __releases(this_rq
->lock
)
1914 __acquires(busiest
->lock
)
1915 __acquires(this_rq
->lock
)
1917 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1918 if (busiest
< this_rq
) {
1919 spin_unlock(&this_rq
->lock
);
1920 spin_lock(&busiest
->lock
);
1921 spin_lock(&this_rq
->lock
);
1923 spin_lock(&busiest
->lock
);
1928 * If dest_cpu is allowed for this process, migrate the task to it.
1929 * This is accomplished by forcing the cpu_allowed mask to only
1930 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
1931 * the cpu_allowed mask is restored.
1933 static void sched_migrate_task(task_t
*p
, int dest_cpu
)
1935 migration_req_t req
;
1937 unsigned long flags
;
1939 rq
= task_rq_lock(p
, &flags
);
1940 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
1941 || unlikely(cpu_is_offline(dest_cpu
)))
1944 /* force the process onto the specified CPU */
1945 if (migrate_task(p
, dest_cpu
, &req
)) {
1946 /* Need to wait for migration thread (might exit: take ref). */
1947 struct task_struct
*mt
= rq
->migration_thread
;
1948 get_task_struct(mt
);
1949 task_rq_unlock(rq
, &flags
);
1950 wake_up_process(mt
);
1951 put_task_struct(mt
);
1952 wait_for_completion(&req
.done
);
1956 task_rq_unlock(rq
, &flags
);
1960 * sched_exec - execve() is a valuable balancing opportunity, because at
1961 * this point the task has the smallest effective memory and cache footprint.
1963 void sched_exec(void)
1965 int new_cpu
, this_cpu
= get_cpu();
1966 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
1968 if (new_cpu
!= this_cpu
)
1969 sched_migrate_task(current
, new_cpu
);
1973 * pull_task - move a task from a remote runqueue to the local runqueue.
1974 * Both runqueues must be locked.
1977 void pull_task(runqueue_t
*src_rq
, prio_array_t
*src_array
, task_t
*p
,
1978 runqueue_t
*this_rq
, prio_array_t
*this_array
, int this_cpu
)
1980 dequeue_task(p
, src_array
);
1981 dec_nr_running(p
, src_rq
);
1982 set_task_cpu(p
, this_cpu
);
1983 inc_nr_running(p
, this_rq
);
1984 enqueue_task(p
, this_array
);
1985 p
->timestamp
= (p
->timestamp
- src_rq
->timestamp_last_tick
)
1986 + this_rq
->timestamp_last_tick
;
1988 * Note that idle threads have a prio of MAX_PRIO, for this test
1989 * to be always true for them.
1991 if (TASK_PREEMPTS_CURR(p
, this_rq
))
1992 resched_task(this_rq
->curr
);
1996 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1999 int can_migrate_task(task_t
*p
, runqueue_t
*rq
, int this_cpu
,
2000 struct sched_domain
*sd
, enum idle_type idle
,
2004 * We do not migrate tasks that are:
2005 * 1) running (obviously), or
2006 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2007 * 3) are cache-hot on their current CPU.
2009 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
2013 if (task_running(rq
, p
))
2017 * Aggressive migration if:
2018 * 1) task is cache cold, or
2019 * 2) too many balance attempts have failed.
2022 if (sd
->nr_balance_failed
> sd
->cache_nice_tries
)
2025 if (task_hot(p
, rq
->timestamp_last_tick
, sd
))
2030 #define rq_best_prio(rq) min((rq)->curr->prio, (rq)->best_expired_prio)
2032 * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted
2033 * load from busiest to this_rq, as part of a balancing operation within
2034 * "domain". Returns the number of tasks moved.
2036 * Called with both runqueues locked.
2038 static int move_tasks(runqueue_t
*this_rq
, int this_cpu
, runqueue_t
*busiest
,
2039 unsigned long max_nr_move
, unsigned long max_load_move
,
2040 struct sched_domain
*sd
, enum idle_type idle
,
2043 prio_array_t
*array
, *dst_array
;
2044 struct list_head
*head
, *curr
;
2045 int idx
, pulled
= 0, pinned
= 0, this_best_prio
, busiest_best_prio
;
2046 int busiest_best_prio_seen
;
2047 int skip_for_load
; /* skip the task based on weighted load issues */
2051 if (max_nr_move
== 0 || max_load_move
== 0)
2054 rem_load_move
= max_load_move
;
2056 this_best_prio
= rq_best_prio(this_rq
);
2057 busiest_best_prio
= rq_best_prio(busiest
);
2059 * Enable handling of the case where there is more than one task
2060 * with the best priority. If the current running task is one
2061 * of those with prio==busiest_best_prio we know it won't be moved
2062 * and therefore it's safe to override the skip (based on load) of
2063 * any task we find with that prio.
2065 busiest_best_prio_seen
= busiest_best_prio
== busiest
->curr
->prio
;
2068 * We first consider expired tasks. Those will likely not be
2069 * executed in the near future, and they are most likely to
2070 * be cache-cold, thus switching CPUs has the least effect
2073 if (busiest
->expired
->nr_active
) {
2074 array
= busiest
->expired
;
2075 dst_array
= this_rq
->expired
;
2077 array
= busiest
->active
;
2078 dst_array
= this_rq
->active
;
2082 /* Start searching at priority 0: */
2086 idx
= sched_find_first_bit(array
->bitmap
);
2088 idx
= find_next_bit(array
->bitmap
, MAX_PRIO
, idx
);
2089 if (idx
>= MAX_PRIO
) {
2090 if (array
== busiest
->expired
&& busiest
->active
->nr_active
) {
2091 array
= busiest
->active
;
2092 dst_array
= this_rq
->active
;
2098 head
= array
->queue
+ idx
;
2101 tmp
= list_entry(curr
, task_t
, run_list
);
2106 * To help distribute high priority tasks accross CPUs we don't
2107 * skip a task if it will be the highest priority task (i.e. smallest
2108 * prio value) on its new queue regardless of its load weight
2110 skip_for_load
= tmp
->load_weight
> rem_load_move
;
2111 if (skip_for_load
&& idx
< this_best_prio
)
2112 skip_for_load
= !busiest_best_prio_seen
&& idx
== busiest_best_prio
;
2113 if (skip_for_load
||
2114 !can_migrate_task(tmp
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2115 busiest_best_prio_seen
|= idx
== busiest_best_prio
;
2122 #ifdef CONFIG_SCHEDSTATS
2123 if (task_hot(tmp
, busiest
->timestamp_last_tick
, sd
))
2124 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2127 pull_task(busiest
, array
, tmp
, this_rq
, dst_array
, this_cpu
);
2129 rem_load_move
-= tmp
->load_weight
;
2132 * We only want to steal up to the prescribed number of tasks
2133 * and the prescribed amount of weighted load.
2135 if (pulled
< max_nr_move
&& rem_load_move
> 0) {
2136 if (idx
< this_best_prio
)
2137 this_best_prio
= idx
;
2145 * Right now, this is the only place pull_task() is called,
2146 * so we can safely collect pull_task() stats here rather than
2147 * inside pull_task().
2149 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2152 *all_pinned
= pinned
;
2157 * find_busiest_group finds and returns the busiest CPU group within the
2158 * domain. It calculates and returns the amount of weighted load which should be
2159 * moved to restore balance via the imbalance parameter.
2161 static struct sched_group
*
2162 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2163 unsigned long *imbalance
, enum idle_type idle
, int *sd_idle
)
2165 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2166 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2167 unsigned long max_pull
;
2168 unsigned long busiest_load_per_task
, busiest_nr_running
;
2169 unsigned long this_load_per_task
, this_nr_running
;
2171 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2172 int power_savings_balance
= 1;
2173 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2174 unsigned long min_nr_running
= ULONG_MAX
;
2175 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2178 max_load
= this_load
= total_load
= total_pwr
= 0;
2179 busiest_load_per_task
= busiest_nr_running
= 0;
2180 this_load_per_task
= this_nr_running
= 0;
2181 if (idle
== NOT_IDLE
)
2182 load_idx
= sd
->busy_idx
;
2183 else if (idle
== NEWLY_IDLE
)
2184 load_idx
= sd
->newidle_idx
;
2186 load_idx
= sd
->idle_idx
;
2189 unsigned long load
, group_capacity
;
2192 unsigned long sum_nr_running
, sum_weighted_load
;
2194 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2196 /* Tally up the load of all CPUs in the group */
2197 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2199 for_each_cpu_mask(i
, group
->cpumask
) {
2200 runqueue_t
*rq
= cpu_rq(i
);
2202 if (*sd_idle
&& !idle_cpu(i
))
2205 /* Bias balancing toward cpus of our domain */
2207 load
= target_load(i
, load_idx
);
2209 load
= source_load(i
, load_idx
);
2212 sum_nr_running
+= rq
->nr_running
;
2213 sum_weighted_load
+= rq
->raw_weighted_load
;
2216 total_load
+= avg_load
;
2217 total_pwr
+= group
->cpu_power
;
2219 /* Adjust by relative CPU power of the group */
2220 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
2222 group_capacity
= group
->cpu_power
/ SCHED_LOAD_SCALE
;
2225 this_load
= avg_load
;
2227 this_nr_running
= sum_nr_running
;
2228 this_load_per_task
= sum_weighted_load
;
2229 } else if (avg_load
> max_load
&&
2230 sum_nr_running
> group_capacity
) {
2231 max_load
= avg_load
;
2233 busiest_nr_running
= sum_nr_running
;
2234 busiest_load_per_task
= sum_weighted_load
;
2237 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2239 * Busy processors will not participate in power savings
2242 if (idle
== NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2246 * If the local group is idle or completely loaded
2247 * no need to do power savings balance at this domain
2249 if (local_group
&& (this_nr_running
>= group_capacity
||
2251 power_savings_balance
= 0;
2254 * If a group is already running at full capacity or idle,
2255 * don't include that group in power savings calculations
2257 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2262 * Calculate the group which has the least non-idle load.
2263 * This is the group from where we need to pick up the load
2266 if ((sum_nr_running
< min_nr_running
) ||
2267 (sum_nr_running
== min_nr_running
&&
2268 first_cpu(group
->cpumask
) <
2269 first_cpu(group_min
->cpumask
))) {
2271 min_nr_running
= sum_nr_running
;
2272 min_load_per_task
= sum_weighted_load
/
2277 * Calculate the group which is almost near its
2278 * capacity but still has some space to pick up some load
2279 * from other group and save more power
2281 if (sum_nr_running
<= group_capacity
- 1)
2282 if (sum_nr_running
> leader_nr_running
||
2283 (sum_nr_running
== leader_nr_running
&&
2284 first_cpu(group
->cpumask
) >
2285 first_cpu(group_leader
->cpumask
))) {
2286 group_leader
= group
;
2287 leader_nr_running
= sum_nr_running
;
2292 group
= group
->next
;
2293 } while (group
!= sd
->groups
);
2295 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2298 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2300 if (this_load
>= avg_load
||
2301 100*max_load
<= sd
->imbalance_pct
*this_load
)
2304 busiest_load_per_task
/= busiest_nr_running
;
2306 * We're trying to get all the cpus to the average_load, so we don't
2307 * want to push ourselves above the average load, nor do we wish to
2308 * reduce the max loaded cpu below the average load, as either of these
2309 * actions would just result in more rebalancing later, and ping-pong
2310 * tasks around. Thus we look for the minimum possible imbalance.
2311 * Negative imbalances (*we* are more loaded than anyone else) will
2312 * be counted as no imbalance for these purposes -- we can't fix that
2313 * by pulling tasks to us. Be careful of negative numbers as they'll
2314 * appear as very large values with unsigned longs.
2316 if (max_load
<= busiest_load_per_task
)
2320 * In the presence of smp nice balancing, certain scenarios can have
2321 * max load less than avg load(as we skip the groups at or below
2322 * its cpu_power, while calculating max_load..)
2324 if (max_load
< avg_load
) {
2326 goto small_imbalance
;
2329 /* Don't want to pull so many tasks that a group would go idle */
2330 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2332 /* How much load to actually move to equalise the imbalance */
2333 *imbalance
= min(max_pull
* busiest
->cpu_power
,
2334 (avg_load
- this_load
) * this->cpu_power
)
2338 * if *imbalance is less than the average load per runnable task
2339 * there is no gaurantee that any tasks will be moved so we'll have
2340 * a think about bumping its value to force at least one task to be
2343 if (*imbalance
< busiest_load_per_task
) {
2344 unsigned long pwr_now
, pwr_move
;
2349 pwr_move
= pwr_now
= 0;
2351 if (this_nr_running
) {
2352 this_load_per_task
/= this_nr_running
;
2353 if (busiest_load_per_task
> this_load_per_task
)
2356 this_load_per_task
= SCHED_LOAD_SCALE
;
2358 if (max_load
- this_load
>= busiest_load_per_task
* imbn
) {
2359 *imbalance
= busiest_load_per_task
;
2364 * OK, we don't have enough imbalance to justify moving tasks,
2365 * however we may be able to increase total CPU power used by
2369 pwr_now
+= busiest
->cpu_power
*
2370 min(busiest_load_per_task
, max_load
);
2371 pwr_now
+= this->cpu_power
*
2372 min(this_load_per_task
, this_load
);
2373 pwr_now
/= SCHED_LOAD_SCALE
;
2375 /* Amount of load we'd subtract */
2376 tmp
= busiest_load_per_task
*SCHED_LOAD_SCALE
/busiest
->cpu_power
;
2378 pwr_move
+= busiest
->cpu_power
*
2379 min(busiest_load_per_task
, max_load
- tmp
);
2381 /* Amount of load we'd add */
2382 if (max_load
*busiest
->cpu_power
<
2383 busiest_load_per_task
*SCHED_LOAD_SCALE
)
2384 tmp
= max_load
*busiest
->cpu_power
/this->cpu_power
;
2386 tmp
= busiest_load_per_task
*SCHED_LOAD_SCALE
/this->cpu_power
;
2387 pwr_move
+= this->cpu_power
*min(this_load_per_task
, this_load
+ tmp
);
2388 pwr_move
/= SCHED_LOAD_SCALE
;
2390 /* Move if we gain throughput */
2391 if (pwr_move
<= pwr_now
)
2394 *imbalance
= busiest_load_per_task
;
2400 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2401 if (idle
== NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2404 if (this == group_leader
&& group_leader
!= group_min
) {
2405 *imbalance
= min_load_per_task
;
2415 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2417 static runqueue_t
*find_busiest_queue(struct sched_group
*group
,
2418 enum idle_type idle
, unsigned long imbalance
)
2420 unsigned long max_load
= 0;
2421 runqueue_t
*busiest
= NULL
, *rqi
;
2424 for_each_cpu_mask(i
, group
->cpumask
) {
2427 if (rqi
->nr_running
== 1 && rqi
->raw_weighted_load
> imbalance
)
2430 if (rqi
->raw_weighted_load
> max_load
) {
2431 max_load
= rqi
->raw_weighted_load
;
2440 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2441 * so long as it is large enough.
2443 #define MAX_PINNED_INTERVAL 512
2445 #define minus_1_or_zero(n) ((n) > 0 ? (n) - 1 : 0)
2447 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2448 * tasks if there is an imbalance.
2450 * Called with this_rq unlocked.
2452 static int load_balance(int this_cpu
, runqueue_t
*this_rq
,
2453 struct sched_domain
*sd
, enum idle_type idle
)
2455 struct sched_group
*group
;
2456 runqueue_t
*busiest
;
2457 unsigned long imbalance
;
2458 int nr_moved
, all_pinned
= 0;
2459 int active_balance
= 0;
2462 if (idle
!= NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2463 !sched_smt_power_savings
)
2466 schedstat_inc(sd
, lb_cnt
[idle
]);
2468 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
);
2470 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2474 busiest
= find_busiest_queue(group
, idle
, imbalance
);
2476 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2480 BUG_ON(busiest
== this_rq
);
2482 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2485 if (busiest
->nr_running
> 1) {
2487 * Attempt to move tasks. If find_busiest_group has found
2488 * an imbalance but busiest->nr_running <= 1, the group is
2489 * still unbalanced. nr_moved simply stays zero, so it is
2490 * correctly treated as an imbalance.
2492 double_rq_lock(this_rq
, busiest
);
2493 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2494 minus_1_or_zero(busiest
->nr_running
),
2495 imbalance
, sd
, idle
, &all_pinned
);
2496 double_rq_unlock(this_rq
, busiest
);
2498 /* All tasks on this runqueue were pinned by CPU affinity */
2499 if (unlikely(all_pinned
))
2504 schedstat_inc(sd
, lb_failed
[idle
]);
2505 sd
->nr_balance_failed
++;
2507 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2509 spin_lock(&busiest
->lock
);
2511 /* don't kick the migration_thread, if the curr
2512 * task on busiest cpu can't be moved to this_cpu
2514 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2515 spin_unlock(&busiest
->lock
);
2517 goto out_one_pinned
;
2520 if (!busiest
->active_balance
) {
2521 busiest
->active_balance
= 1;
2522 busiest
->push_cpu
= this_cpu
;
2525 spin_unlock(&busiest
->lock
);
2527 wake_up_process(busiest
->migration_thread
);
2530 * We've kicked active balancing, reset the failure
2533 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2536 sd
->nr_balance_failed
= 0;
2538 if (likely(!active_balance
)) {
2539 /* We were unbalanced, so reset the balancing interval */
2540 sd
->balance_interval
= sd
->min_interval
;
2543 * If we've begun active balancing, start to back off. This
2544 * case may not be covered by the all_pinned logic if there
2545 * is only 1 task on the busy runqueue (because we don't call
2548 if (sd
->balance_interval
< sd
->max_interval
)
2549 sd
->balance_interval
*= 2;
2552 if (!nr_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2553 !sched_smt_power_savings
)
2558 schedstat_inc(sd
, lb_balanced
[idle
]);
2560 sd
->nr_balance_failed
= 0;
2563 /* tune up the balancing interval */
2564 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2565 (sd
->balance_interval
< sd
->max_interval
))
2566 sd
->balance_interval
*= 2;
2568 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&& !sched_smt_power_savings
)
2574 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2575 * tasks if there is an imbalance.
2577 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2578 * this_rq is locked.
2580 static int load_balance_newidle(int this_cpu
, runqueue_t
*this_rq
,
2581 struct sched_domain
*sd
)
2583 struct sched_group
*group
;
2584 runqueue_t
*busiest
= NULL
;
2585 unsigned long imbalance
;
2589 if (sd
->flags
& SD_SHARE_CPUPOWER
&& !sched_smt_power_savings
)
2592 schedstat_inc(sd
, lb_cnt
[NEWLY_IDLE
]);
2593 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, NEWLY_IDLE
, &sd_idle
);
2595 schedstat_inc(sd
, lb_nobusyg
[NEWLY_IDLE
]);
2599 busiest
= find_busiest_queue(group
, NEWLY_IDLE
, imbalance
);
2601 schedstat_inc(sd
, lb_nobusyq
[NEWLY_IDLE
]);
2605 BUG_ON(busiest
== this_rq
);
2607 schedstat_add(sd
, lb_imbalance
[NEWLY_IDLE
], imbalance
);
2610 if (busiest
->nr_running
> 1) {
2611 /* Attempt to move tasks */
2612 double_lock_balance(this_rq
, busiest
);
2613 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2614 minus_1_or_zero(busiest
->nr_running
),
2615 imbalance
, sd
, NEWLY_IDLE
, NULL
);
2616 spin_unlock(&busiest
->lock
);
2620 schedstat_inc(sd
, lb_failed
[NEWLY_IDLE
]);
2621 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2624 sd
->nr_balance_failed
= 0;
2629 schedstat_inc(sd
, lb_balanced
[NEWLY_IDLE
]);
2630 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&& !sched_smt_power_savings
)
2632 sd
->nr_balance_failed
= 0;
2637 * idle_balance is called by schedule() if this_cpu is about to become
2638 * idle. Attempts to pull tasks from other CPUs.
2640 static void idle_balance(int this_cpu
, runqueue_t
*this_rq
)
2642 struct sched_domain
*sd
;
2644 for_each_domain(this_cpu
, sd
) {
2645 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
2646 if (load_balance_newidle(this_cpu
, this_rq
, sd
)) {
2647 /* We've pulled tasks over so stop searching */
2655 * active_load_balance is run by migration threads. It pushes running tasks
2656 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2657 * running on each physical CPU where possible, and avoids physical /
2658 * logical imbalances.
2660 * Called with busiest_rq locked.
2662 static void active_load_balance(runqueue_t
*busiest_rq
, int busiest_cpu
)
2664 struct sched_domain
*sd
;
2665 runqueue_t
*target_rq
;
2666 int target_cpu
= busiest_rq
->push_cpu
;
2668 if (busiest_rq
->nr_running
<= 1)
2669 /* no task to move */
2672 target_rq
= cpu_rq(target_cpu
);
2675 * This condition is "impossible", if it occurs
2676 * we need to fix it. Originally reported by
2677 * Bjorn Helgaas on a 128-cpu setup.
2679 BUG_ON(busiest_rq
== target_rq
);
2681 /* move a task from busiest_rq to target_rq */
2682 double_lock_balance(busiest_rq
, target_rq
);
2684 /* Search for an sd spanning us and the target CPU. */
2685 for_each_domain(target_cpu
, sd
) {
2686 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2687 cpu_isset(busiest_cpu
, sd
->span
))
2691 if (unlikely(sd
== NULL
))
2694 schedstat_inc(sd
, alb_cnt
);
2696 if (move_tasks(target_rq
, target_cpu
, busiest_rq
, 1,
2697 RTPRIO_TO_LOAD_WEIGHT(100), sd
, SCHED_IDLE
, NULL
))
2698 schedstat_inc(sd
, alb_pushed
);
2700 schedstat_inc(sd
, alb_failed
);
2702 spin_unlock(&target_rq
->lock
);
2706 * rebalance_tick will get called every timer tick, on every CPU.
2708 * It checks each scheduling domain to see if it is due to be balanced,
2709 * and initiates a balancing operation if so.
2711 * Balancing parameters are set up in arch_init_sched_domains.
2714 /* Don't have all balancing operations going off at once */
2715 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2717 static void rebalance_tick(int this_cpu
, runqueue_t
*this_rq
,
2718 enum idle_type idle
)
2720 unsigned long old_load
, this_load
;
2721 unsigned long j
= jiffies
+ CPU_OFFSET(this_cpu
);
2722 struct sched_domain
*sd
;
2725 this_load
= this_rq
->raw_weighted_load
;
2726 /* Update our load */
2727 for (i
= 0; i
< 3; i
++) {
2728 unsigned long new_load
= this_load
;
2730 old_load
= this_rq
->cpu_load
[i
];
2732 * Round up the averaging division if load is increasing. This
2733 * prevents us from getting stuck on 9 if the load is 10, for
2736 if (new_load
> old_load
)
2737 new_load
+= scale
-1;
2738 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) / scale
;
2741 for_each_domain(this_cpu
, sd
) {
2742 unsigned long interval
;
2744 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2747 interval
= sd
->balance_interval
;
2748 if (idle
!= SCHED_IDLE
)
2749 interval
*= sd
->busy_factor
;
2751 /* scale ms to jiffies */
2752 interval
= msecs_to_jiffies(interval
);
2753 if (unlikely(!interval
))
2756 if (j
- sd
->last_balance
>= interval
) {
2757 if (load_balance(this_cpu
, this_rq
, sd
, idle
)) {
2759 * We've pulled tasks over so either we're no
2760 * longer idle, or one of our SMT siblings is
2765 sd
->last_balance
+= interval
;
2771 * on UP we do not need to balance between CPUs:
2773 static inline void rebalance_tick(int cpu
, runqueue_t
*rq
, enum idle_type idle
)
2776 static inline void idle_balance(int cpu
, runqueue_t
*rq
)
2781 static inline int wake_priority_sleeper(runqueue_t
*rq
)
2784 #ifdef CONFIG_SCHED_SMT
2785 spin_lock(&rq
->lock
);
2787 * If an SMT sibling task has been put to sleep for priority
2788 * reasons reschedule the idle task to see if it can now run.
2790 if (rq
->nr_running
) {
2791 resched_task(rq
->idle
);
2794 spin_unlock(&rq
->lock
);
2799 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2801 EXPORT_PER_CPU_SYMBOL(kstat
);
2804 * This is called on clock ticks and on context switches.
2805 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2807 static inline void update_cpu_clock(task_t
*p
, runqueue_t
*rq
,
2808 unsigned long long now
)
2810 unsigned long long last
= max(p
->timestamp
, rq
->timestamp_last_tick
);
2811 p
->sched_time
+= now
- last
;
2815 * Return current->sched_time plus any more ns on the sched_clock
2816 * that have not yet been banked.
2818 unsigned long long current_sched_time(const task_t
*tsk
)
2820 unsigned long long ns
;
2821 unsigned long flags
;
2822 local_irq_save(flags
);
2823 ns
= max(tsk
->timestamp
, task_rq(tsk
)->timestamp_last_tick
);
2824 ns
= tsk
->sched_time
+ (sched_clock() - ns
);
2825 local_irq_restore(flags
);
2830 * We place interactive tasks back into the active array, if possible.
2832 * To guarantee that this does not starve expired tasks we ignore the
2833 * interactivity of a task if the first expired task had to wait more
2834 * than a 'reasonable' amount of time. This deadline timeout is
2835 * load-dependent, as the frequency of array switched decreases with
2836 * increasing number of running tasks. We also ignore the interactivity
2837 * if a better static_prio task has expired:
2839 #define EXPIRED_STARVING(rq) \
2840 ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2841 (jiffies - (rq)->expired_timestamp >= \
2842 STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2843 ((rq)->curr->static_prio > (rq)->best_expired_prio))
2846 * Account user cpu time to a process.
2847 * @p: the process that the cpu time gets accounted to
2848 * @hardirq_offset: the offset to subtract from hardirq_count()
2849 * @cputime: the cpu time spent in user space since the last update
2851 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
2853 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2856 p
->utime
= cputime_add(p
->utime
, cputime
);
2858 /* Add user time to cpustat. */
2859 tmp
= cputime_to_cputime64(cputime
);
2860 if (TASK_NICE(p
) > 0)
2861 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
2863 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
2867 * Account system cpu time to a process.
2868 * @p: the process that the cpu time gets accounted to
2869 * @hardirq_offset: the offset to subtract from hardirq_count()
2870 * @cputime: the cpu time spent in kernel space since the last update
2872 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
2875 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2876 runqueue_t
*rq
= this_rq();
2879 p
->stime
= cputime_add(p
->stime
, cputime
);
2881 /* Add system time to cpustat. */
2882 tmp
= cputime_to_cputime64(cputime
);
2883 if (hardirq_count() - hardirq_offset
)
2884 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
2885 else if (softirq_count())
2886 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
2887 else if (p
!= rq
->idle
)
2888 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
2889 else if (atomic_read(&rq
->nr_iowait
) > 0)
2890 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
2892 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
2893 /* Account for system time used */
2894 acct_update_integrals(p
);
2898 * Account for involuntary wait time.
2899 * @p: the process from which the cpu time has been stolen
2900 * @steal: the cpu time spent in involuntary wait
2902 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
2904 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2905 cputime64_t tmp
= cputime_to_cputime64(steal
);
2906 runqueue_t
*rq
= this_rq();
2908 if (p
== rq
->idle
) {
2909 p
->stime
= cputime_add(p
->stime
, steal
);
2910 if (atomic_read(&rq
->nr_iowait
) > 0)
2911 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
2913 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
2915 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
2919 * This function gets called by the timer code, with HZ frequency.
2920 * We call it with interrupts disabled.
2922 * It also gets called by the fork code, when changing the parent's
2925 void scheduler_tick(void)
2927 int cpu
= smp_processor_id();
2928 runqueue_t
*rq
= this_rq();
2929 task_t
*p
= current
;
2930 unsigned long long now
= sched_clock();
2932 update_cpu_clock(p
, rq
, now
);
2934 rq
->timestamp_last_tick
= now
;
2936 if (p
== rq
->idle
) {
2937 if (wake_priority_sleeper(rq
))
2939 rebalance_tick(cpu
, rq
, SCHED_IDLE
);
2943 /* Task might have expired already, but not scheduled off yet */
2944 if (p
->array
!= rq
->active
) {
2945 set_tsk_need_resched(p
);
2948 spin_lock(&rq
->lock
);
2950 * The task was running during this tick - update the
2951 * time slice counter. Note: we do not update a thread's
2952 * priority until it either goes to sleep or uses up its
2953 * timeslice. This makes it possible for interactive tasks
2954 * to use up their timeslices at their highest priority levels.
2958 * RR tasks need a special form of timeslice management.
2959 * FIFO tasks have no timeslices.
2961 if ((p
->policy
== SCHED_RR
) && !--p
->time_slice
) {
2962 p
->time_slice
= task_timeslice(p
);
2963 p
->first_time_slice
= 0;
2964 set_tsk_need_resched(p
);
2966 /* put it at the end of the queue: */
2967 requeue_task(p
, rq
->active
);
2971 if (!--p
->time_slice
) {
2972 dequeue_task(p
, rq
->active
);
2973 set_tsk_need_resched(p
);
2974 p
->prio
= effective_prio(p
);
2975 p
->time_slice
= task_timeslice(p
);
2976 p
->first_time_slice
= 0;
2978 if (!rq
->expired_timestamp
)
2979 rq
->expired_timestamp
= jiffies
;
2980 if (!TASK_INTERACTIVE(p
) || EXPIRED_STARVING(rq
)) {
2981 enqueue_task(p
, rq
->expired
);
2982 if (p
->static_prio
< rq
->best_expired_prio
)
2983 rq
->best_expired_prio
= p
->static_prio
;
2985 enqueue_task(p
, rq
->active
);
2988 * Prevent a too long timeslice allowing a task to monopolize
2989 * the CPU. We do this by splitting up the timeslice into
2992 * Note: this does not mean the task's timeslices expire or
2993 * get lost in any way, they just might be preempted by
2994 * another task of equal priority. (one with higher
2995 * priority would have preempted this task already.) We
2996 * requeue this task to the end of the list on this priority
2997 * level, which is in essence a round-robin of tasks with
3000 * This only applies to tasks in the interactive
3001 * delta range with at least TIMESLICE_GRANULARITY to requeue.
3003 if (TASK_INTERACTIVE(p
) && !((task_timeslice(p
) -
3004 p
->time_slice
) % TIMESLICE_GRANULARITY(p
)) &&
3005 (p
->time_slice
>= TIMESLICE_GRANULARITY(p
)) &&
3006 (p
->array
== rq
->active
)) {
3008 requeue_task(p
, rq
->active
);
3009 set_tsk_need_resched(p
);
3013 spin_unlock(&rq
->lock
);
3015 rebalance_tick(cpu
, rq
, NOT_IDLE
);
3018 #ifdef CONFIG_SCHED_SMT
3019 static inline void wakeup_busy_runqueue(runqueue_t
*rq
)
3021 /* If an SMT runqueue is sleeping due to priority reasons wake it up */
3022 if (rq
->curr
== rq
->idle
&& rq
->nr_running
)
3023 resched_task(rq
->idle
);
3027 * Called with interrupt disabled and this_rq's runqueue locked.
3029 static void wake_sleeping_dependent(int this_cpu
)
3031 struct sched_domain
*tmp
, *sd
= NULL
;
3034 for_each_domain(this_cpu
, tmp
) {
3035 if (tmp
->flags
& SD_SHARE_CPUPOWER
) {
3044 for_each_cpu_mask(i
, sd
->span
) {
3045 runqueue_t
*smt_rq
= cpu_rq(i
);
3049 if (unlikely(!spin_trylock(&smt_rq
->lock
)))
3052 wakeup_busy_runqueue(smt_rq
);
3053 spin_unlock(&smt_rq
->lock
);
3058 * number of 'lost' timeslices this task wont be able to fully
3059 * utilize, if another task runs on a sibling. This models the
3060 * slowdown effect of other tasks running on siblings:
3062 static inline unsigned long smt_slice(task_t
*p
, struct sched_domain
*sd
)
3064 return p
->time_slice
* (100 - sd
->per_cpu_gain
) / 100;
3068 * To minimise lock contention and not have to drop this_rq's runlock we only
3069 * trylock the sibling runqueues and bypass those runqueues if we fail to
3070 * acquire their lock. As we only trylock the normal locking order does not
3071 * need to be obeyed.
3073 static int dependent_sleeper(int this_cpu
, runqueue_t
*this_rq
, task_t
*p
)
3075 struct sched_domain
*tmp
, *sd
= NULL
;
3078 /* kernel/rt threads do not participate in dependent sleeping */
3079 if (!p
->mm
|| rt_task(p
))
3082 for_each_domain(this_cpu
, tmp
) {
3083 if (tmp
->flags
& SD_SHARE_CPUPOWER
) {
3092 for_each_cpu_mask(i
, sd
->span
) {
3100 if (unlikely(!spin_trylock(&smt_rq
->lock
)))
3103 smt_curr
= smt_rq
->curr
;
3109 * If a user task with lower static priority than the
3110 * running task on the SMT sibling is trying to schedule,
3111 * delay it till there is proportionately less timeslice
3112 * left of the sibling task to prevent a lower priority
3113 * task from using an unfair proportion of the
3114 * physical cpu's resources. -ck
3116 if (rt_task(smt_curr
)) {
3118 * With real time tasks we run non-rt tasks only
3119 * per_cpu_gain% of the time.
3121 if ((jiffies
% DEF_TIMESLICE
) >
3122 (sd
->per_cpu_gain
* DEF_TIMESLICE
/ 100))
3125 if (smt_curr
->static_prio
< p
->static_prio
&&
3126 !TASK_PREEMPTS_CURR(p
, smt_rq
) &&
3127 smt_slice(smt_curr
, sd
) > task_timeslice(p
))
3131 spin_unlock(&smt_rq
->lock
);
3136 static inline void wake_sleeping_dependent(int this_cpu
)
3140 static inline int dependent_sleeper(int this_cpu
, runqueue_t
*this_rq
,
3147 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3149 void fastcall
add_preempt_count(int val
)
3154 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3156 preempt_count() += val
;
3158 * Spinlock count overflowing soon?
3160 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >= PREEMPT_MASK
-10);
3162 EXPORT_SYMBOL(add_preempt_count
);
3164 void fastcall
sub_preempt_count(int val
)
3169 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3172 * Is the spinlock portion underflowing?
3174 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3175 !(preempt_count() & PREEMPT_MASK
)))
3178 preempt_count() -= val
;
3180 EXPORT_SYMBOL(sub_preempt_count
);
3184 static inline int interactive_sleep(enum sleep_type sleep_type
)
3186 return (sleep_type
== SLEEP_INTERACTIVE
||
3187 sleep_type
== SLEEP_INTERRUPTED
);
3191 * schedule() is the main scheduler function.
3193 asmlinkage
void __sched
schedule(void)
3196 task_t
*prev
, *next
;
3198 prio_array_t
*array
;
3199 struct list_head
*queue
;
3200 unsigned long long now
;
3201 unsigned long run_time
;
3202 int cpu
, idx
, new_prio
;
3205 * Test if we are atomic. Since do_exit() needs to call into
3206 * schedule() atomically, we ignore that path for now.
3207 * Otherwise, whine if we are scheduling when we should not be.
3209 if (unlikely(in_atomic() && !current
->exit_state
)) {
3210 printk(KERN_ERR
"BUG: scheduling while atomic: "
3212 current
->comm
, preempt_count(), current
->pid
);
3215 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3220 release_kernel_lock(prev
);
3221 need_resched_nonpreemptible
:
3225 * The idle thread is not allowed to schedule!
3226 * Remove this check after it has been exercised a bit.
3228 if (unlikely(prev
== rq
->idle
) && prev
->state
!= TASK_RUNNING
) {
3229 printk(KERN_ERR
"bad: scheduling from the idle thread!\n");
3233 schedstat_inc(rq
, sched_cnt
);
3234 now
= sched_clock();
3235 if (likely((long long)(now
- prev
->timestamp
) < NS_MAX_SLEEP_AVG
)) {
3236 run_time
= now
- prev
->timestamp
;
3237 if (unlikely((long long)(now
- prev
->timestamp
) < 0))
3240 run_time
= NS_MAX_SLEEP_AVG
;
3243 * Tasks charged proportionately less run_time at high sleep_avg to
3244 * delay them losing their interactive status
3246 run_time
/= (CURRENT_BONUS(prev
) ? : 1);
3248 spin_lock_irq(&rq
->lock
);
3250 if (unlikely(prev
->flags
& PF_DEAD
))
3251 prev
->state
= EXIT_DEAD
;
3253 switch_count
= &prev
->nivcsw
;
3254 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3255 switch_count
= &prev
->nvcsw
;
3256 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3257 unlikely(signal_pending(prev
))))
3258 prev
->state
= TASK_RUNNING
;
3260 if (prev
->state
== TASK_UNINTERRUPTIBLE
)
3261 rq
->nr_uninterruptible
++;
3262 deactivate_task(prev
, rq
);
3266 cpu
= smp_processor_id();
3267 if (unlikely(!rq
->nr_running
)) {
3268 idle_balance(cpu
, rq
);
3269 if (!rq
->nr_running
) {
3271 rq
->expired_timestamp
= 0;
3272 wake_sleeping_dependent(cpu
);
3278 if (unlikely(!array
->nr_active
)) {
3280 * Switch the active and expired arrays.
3282 schedstat_inc(rq
, sched_switch
);
3283 rq
->active
= rq
->expired
;
3284 rq
->expired
= array
;
3286 rq
->expired_timestamp
= 0;
3287 rq
->best_expired_prio
= MAX_PRIO
;
3290 idx
= sched_find_first_bit(array
->bitmap
);
3291 queue
= array
->queue
+ idx
;
3292 next
= list_entry(queue
->next
, task_t
, run_list
);
3294 if (!rt_task(next
) && interactive_sleep(next
->sleep_type
)) {
3295 unsigned long long delta
= now
- next
->timestamp
;
3296 if (unlikely((long long)(now
- next
->timestamp
) < 0))
3299 if (next
->sleep_type
== SLEEP_INTERACTIVE
)
3300 delta
= delta
* (ON_RUNQUEUE_WEIGHT
* 128 / 100) / 128;
3302 array
= next
->array
;
3303 new_prio
= recalc_task_prio(next
, next
->timestamp
+ delta
);
3305 if (unlikely(next
->prio
!= new_prio
)) {
3306 dequeue_task(next
, array
);
3307 next
->prio
= new_prio
;
3308 enqueue_task(next
, array
);
3311 next
->sleep_type
= SLEEP_NORMAL
;
3312 if (dependent_sleeper(cpu
, rq
, next
))
3315 if (next
== rq
->idle
)
3316 schedstat_inc(rq
, sched_goidle
);
3318 prefetch_stack(next
);
3319 clear_tsk_need_resched(prev
);
3320 rcu_qsctr_inc(task_cpu(prev
));
3322 update_cpu_clock(prev
, rq
, now
);
3324 prev
->sleep_avg
-= run_time
;
3325 if ((long)prev
->sleep_avg
<= 0)
3326 prev
->sleep_avg
= 0;
3327 prev
->timestamp
= prev
->last_ran
= now
;
3329 sched_info_switch(prev
, next
);
3330 if (likely(prev
!= next
)) {
3331 next
->timestamp
= now
;
3336 prepare_task_switch(rq
, next
);
3337 prev
= context_switch(rq
, prev
, next
);
3340 * this_rq must be evaluated again because prev may have moved
3341 * CPUs since it called schedule(), thus the 'rq' on its stack
3342 * frame will be invalid.
3344 finish_task_switch(this_rq(), prev
);
3346 spin_unlock_irq(&rq
->lock
);
3349 if (unlikely(reacquire_kernel_lock(prev
) < 0))
3350 goto need_resched_nonpreemptible
;
3351 preempt_enable_no_resched();
3352 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3356 EXPORT_SYMBOL(schedule
);
3358 #ifdef CONFIG_PREEMPT
3360 * this is is the entry point to schedule() from in-kernel preemption
3361 * off of preempt_enable. Kernel preemptions off return from interrupt
3362 * occur there and call schedule directly.
3364 asmlinkage
void __sched
preempt_schedule(void)
3366 struct thread_info
*ti
= current_thread_info();
3367 #ifdef CONFIG_PREEMPT_BKL
3368 struct task_struct
*task
= current
;
3369 int saved_lock_depth
;
3372 * If there is a non-zero preempt_count or interrupts are disabled,
3373 * we do not want to preempt the current task. Just return..
3375 if (unlikely(ti
->preempt_count
|| irqs_disabled()))
3379 add_preempt_count(PREEMPT_ACTIVE
);
3381 * We keep the big kernel semaphore locked, but we
3382 * clear ->lock_depth so that schedule() doesnt
3383 * auto-release the semaphore:
3385 #ifdef CONFIG_PREEMPT_BKL
3386 saved_lock_depth
= task
->lock_depth
;
3387 task
->lock_depth
= -1;
3390 #ifdef CONFIG_PREEMPT_BKL
3391 task
->lock_depth
= saved_lock_depth
;
3393 sub_preempt_count(PREEMPT_ACTIVE
);
3395 /* we could miss a preemption opportunity between schedule and now */
3397 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3401 EXPORT_SYMBOL(preempt_schedule
);
3404 * this is is the entry point to schedule() from kernel preemption
3405 * off of irq context.
3406 * Note, that this is called and return with irqs disabled. This will
3407 * protect us against recursive calling from irq.
3409 asmlinkage
void __sched
preempt_schedule_irq(void)
3411 struct thread_info
*ti
= current_thread_info();
3412 #ifdef CONFIG_PREEMPT_BKL
3413 struct task_struct
*task
= current
;
3414 int saved_lock_depth
;
3416 /* Catch callers which need to be fixed*/
3417 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3420 add_preempt_count(PREEMPT_ACTIVE
);
3422 * We keep the big kernel semaphore locked, but we
3423 * clear ->lock_depth so that schedule() doesnt
3424 * auto-release the semaphore:
3426 #ifdef CONFIG_PREEMPT_BKL
3427 saved_lock_depth
= task
->lock_depth
;
3428 task
->lock_depth
= -1;
3432 local_irq_disable();
3433 #ifdef CONFIG_PREEMPT_BKL
3434 task
->lock_depth
= saved_lock_depth
;
3436 sub_preempt_count(PREEMPT_ACTIVE
);
3438 /* we could miss a preemption opportunity between schedule and now */
3440 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3444 #endif /* CONFIG_PREEMPT */
3446 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3449 task_t
*p
= curr
->private;
3450 return try_to_wake_up(p
, mode
, sync
);
3453 EXPORT_SYMBOL(default_wake_function
);
3456 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3457 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3458 * number) then we wake all the non-exclusive tasks and one exclusive task.
3460 * There are circumstances in which we can try to wake a task which has already
3461 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3462 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3464 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3465 int nr_exclusive
, int sync
, void *key
)
3467 struct list_head
*tmp
, *next
;
3469 list_for_each_safe(tmp
, next
, &q
->task_list
) {
3472 curr
= list_entry(tmp
, wait_queue_t
, task_list
);
3473 flags
= curr
->flags
;
3474 if (curr
->func(curr
, mode
, sync
, key
) &&
3475 (flags
& WQ_FLAG_EXCLUSIVE
) &&
3482 * __wake_up - wake up threads blocked on a waitqueue.
3484 * @mode: which threads
3485 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3486 * @key: is directly passed to the wakeup function
3488 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3489 int nr_exclusive
, void *key
)
3491 unsigned long flags
;
3493 spin_lock_irqsave(&q
->lock
, flags
);
3494 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3495 spin_unlock_irqrestore(&q
->lock
, flags
);
3498 EXPORT_SYMBOL(__wake_up
);
3501 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3503 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3505 __wake_up_common(q
, mode
, 1, 0, NULL
);
3509 * __wake_up_sync - wake up threads blocked on a waitqueue.
3511 * @mode: which threads
3512 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3514 * The sync wakeup differs that the waker knows that it will schedule
3515 * away soon, so while the target thread will be woken up, it will not
3516 * be migrated to another CPU - ie. the two threads are 'synchronized'
3517 * with each other. This can prevent needless bouncing between CPUs.
3519 * On UP it can prevent extra preemption.
3522 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3524 unsigned long flags
;
3530 if (unlikely(!nr_exclusive
))
3533 spin_lock_irqsave(&q
->lock
, flags
);
3534 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3535 spin_unlock_irqrestore(&q
->lock
, flags
);
3537 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3539 void fastcall
complete(struct completion
*x
)
3541 unsigned long flags
;
3543 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3545 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3547 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3549 EXPORT_SYMBOL(complete
);
3551 void fastcall
complete_all(struct completion
*x
)
3553 unsigned long flags
;
3555 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3556 x
->done
+= UINT_MAX
/2;
3557 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3559 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3561 EXPORT_SYMBOL(complete_all
);
3563 void fastcall __sched
wait_for_completion(struct completion
*x
)
3566 spin_lock_irq(&x
->wait
.lock
);
3568 DECLARE_WAITQUEUE(wait
, current
);
3570 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3571 __add_wait_queue_tail(&x
->wait
, &wait
);
3573 __set_current_state(TASK_UNINTERRUPTIBLE
);
3574 spin_unlock_irq(&x
->wait
.lock
);
3576 spin_lock_irq(&x
->wait
.lock
);
3578 __remove_wait_queue(&x
->wait
, &wait
);
3581 spin_unlock_irq(&x
->wait
.lock
);
3583 EXPORT_SYMBOL(wait_for_completion
);
3585 unsigned long fastcall __sched
3586 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3590 spin_lock_irq(&x
->wait
.lock
);
3592 DECLARE_WAITQUEUE(wait
, current
);
3594 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3595 __add_wait_queue_tail(&x
->wait
, &wait
);
3597 __set_current_state(TASK_UNINTERRUPTIBLE
);
3598 spin_unlock_irq(&x
->wait
.lock
);
3599 timeout
= schedule_timeout(timeout
);
3600 spin_lock_irq(&x
->wait
.lock
);
3602 __remove_wait_queue(&x
->wait
, &wait
);
3606 __remove_wait_queue(&x
->wait
, &wait
);
3610 spin_unlock_irq(&x
->wait
.lock
);
3613 EXPORT_SYMBOL(wait_for_completion_timeout
);
3615 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3621 spin_lock_irq(&x
->wait
.lock
);
3623 DECLARE_WAITQUEUE(wait
, current
);
3625 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3626 __add_wait_queue_tail(&x
->wait
, &wait
);
3628 if (signal_pending(current
)) {
3630 __remove_wait_queue(&x
->wait
, &wait
);
3633 __set_current_state(TASK_INTERRUPTIBLE
);
3634 spin_unlock_irq(&x
->wait
.lock
);
3636 spin_lock_irq(&x
->wait
.lock
);
3638 __remove_wait_queue(&x
->wait
, &wait
);
3642 spin_unlock_irq(&x
->wait
.lock
);
3646 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3648 unsigned long fastcall __sched
3649 wait_for_completion_interruptible_timeout(struct completion
*x
,
3650 unsigned long timeout
)
3654 spin_lock_irq(&x
->wait
.lock
);
3656 DECLARE_WAITQUEUE(wait
, current
);
3658 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3659 __add_wait_queue_tail(&x
->wait
, &wait
);
3661 if (signal_pending(current
)) {
3662 timeout
= -ERESTARTSYS
;
3663 __remove_wait_queue(&x
->wait
, &wait
);
3666 __set_current_state(TASK_INTERRUPTIBLE
);
3667 spin_unlock_irq(&x
->wait
.lock
);
3668 timeout
= schedule_timeout(timeout
);
3669 spin_lock_irq(&x
->wait
.lock
);
3671 __remove_wait_queue(&x
->wait
, &wait
);
3675 __remove_wait_queue(&x
->wait
, &wait
);
3679 spin_unlock_irq(&x
->wait
.lock
);
3682 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3685 #define SLEEP_ON_VAR \
3686 unsigned long flags; \
3687 wait_queue_t wait; \
3688 init_waitqueue_entry(&wait, current);
3690 #define SLEEP_ON_HEAD \
3691 spin_lock_irqsave(&q->lock,flags); \
3692 __add_wait_queue(q, &wait); \
3693 spin_unlock(&q->lock);
3695 #define SLEEP_ON_TAIL \
3696 spin_lock_irq(&q->lock); \
3697 __remove_wait_queue(q, &wait); \
3698 spin_unlock_irqrestore(&q->lock, flags);
3700 void fastcall __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3704 current
->state
= TASK_INTERRUPTIBLE
;
3711 EXPORT_SYMBOL(interruptible_sleep_on
);
3713 long fastcall __sched
3714 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3718 current
->state
= TASK_INTERRUPTIBLE
;
3721 timeout
= schedule_timeout(timeout
);
3727 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3729 void fastcall __sched
sleep_on(wait_queue_head_t
*q
)
3733 current
->state
= TASK_UNINTERRUPTIBLE
;
3740 EXPORT_SYMBOL(sleep_on
);
3742 long fastcall __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3746 current
->state
= TASK_UNINTERRUPTIBLE
;
3749 timeout
= schedule_timeout(timeout
);
3755 EXPORT_SYMBOL(sleep_on_timeout
);
3757 #ifdef CONFIG_RT_MUTEXES
3760 * rt_mutex_setprio - set the current priority of a task
3762 * @prio: prio value (kernel-internal form)
3764 * This function changes the 'effective' priority of a task. It does
3765 * not touch ->normal_prio like __setscheduler().
3767 * Used by the rt_mutex code to implement priority inheritance logic.
3769 void rt_mutex_setprio(task_t
*p
, int prio
)
3771 unsigned long flags
;
3772 prio_array_t
*array
;
3776 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3778 rq
= task_rq_lock(p
, &flags
);
3783 dequeue_task(p
, array
);
3788 * If changing to an RT priority then queue it
3789 * in the active array!
3793 enqueue_task(p
, array
);
3795 * Reschedule if we are currently running on this runqueue and
3796 * our priority decreased, or if we are not currently running on
3797 * this runqueue and our priority is higher than the current's
3799 if (task_running(rq
, p
)) {
3800 if (p
->prio
> oldprio
)
3801 resched_task(rq
->curr
);
3802 } else if (TASK_PREEMPTS_CURR(p
, rq
))
3803 resched_task(rq
->curr
);
3805 task_rq_unlock(rq
, &flags
);
3810 void set_user_nice(task_t
*p
, long nice
)
3812 unsigned long flags
;
3813 prio_array_t
*array
;
3815 int old_prio
, delta
;
3817 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3820 * We have to be careful, if called from sys_setpriority(),
3821 * the task might be in the middle of scheduling on another CPU.
3823 rq
= task_rq_lock(p
, &flags
);
3825 * The RT priorities are set via sched_setscheduler(), but we still
3826 * allow the 'normal' nice value to be set - but as expected
3827 * it wont have any effect on scheduling until the task is
3828 * not SCHED_NORMAL/SCHED_BATCH:
3830 if (has_rt_policy(p
)) {
3831 p
->static_prio
= NICE_TO_PRIO(nice
);
3836 dequeue_task(p
, array
);
3837 dec_raw_weighted_load(rq
, p
);
3840 p
->static_prio
= NICE_TO_PRIO(nice
);
3843 p
->prio
= effective_prio(p
);
3844 delta
= p
->prio
- old_prio
;
3847 enqueue_task(p
, array
);
3848 inc_raw_weighted_load(rq
, p
);
3850 * If the task increased its priority or is running and
3851 * lowered its priority, then reschedule its CPU:
3853 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3854 resched_task(rq
->curr
);
3857 task_rq_unlock(rq
, &flags
);
3859 EXPORT_SYMBOL(set_user_nice
);
3862 * can_nice - check if a task can reduce its nice value
3866 int can_nice(const task_t
*p
, const int nice
)
3868 /* convert nice value [19,-20] to rlimit style value [1,40] */
3869 int nice_rlim
= 20 - nice
;
3870 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
3871 capable(CAP_SYS_NICE
));
3874 #ifdef __ARCH_WANT_SYS_NICE
3877 * sys_nice - change the priority of the current process.
3878 * @increment: priority increment
3880 * sys_setpriority is a more generic, but much slower function that
3881 * does similar things.
3883 asmlinkage
long sys_nice(int increment
)
3889 * Setpriority might change our priority at the same moment.
3890 * We don't have to worry. Conceptually one call occurs first
3891 * and we have a single winner.
3893 if (increment
< -40)
3898 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
3904 if (increment
< 0 && !can_nice(current
, nice
))
3907 retval
= security_task_setnice(current
, nice
);
3911 set_user_nice(current
, nice
);
3918 * task_prio - return the priority value of a given task.
3919 * @p: the task in question.
3921 * This is the priority value as seen by users in /proc.
3922 * RT tasks are offset by -200. Normal tasks are centered
3923 * around 0, value goes from -16 to +15.
3925 int task_prio(const task_t
*p
)
3927 return p
->prio
- MAX_RT_PRIO
;
3931 * task_nice - return the nice value of a given task.
3932 * @p: the task in question.
3934 int task_nice(const task_t
*p
)
3936 return TASK_NICE(p
);
3938 EXPORT_SYMBOL_GPL(task_nice
);
3941 * idle_cpu - is a given cpu idle currently?
3942 * @cpu: the processor in question.
3944 int idle_cpu(int cpu
)
3946 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
3950 * idle_task - return the idle task for a given cpu.
3951 * @cpu: the processor in question.
3953 task_t
*idle_task(int cpu
)
3955 return cpu_rq(cpu
)->idle
;
3959 * find_process_by_pid - find a process with a matching PID value.
3960 * @pid: the pid in question.
3962 static inline task_t
*find_process_by_pid(pid_t pid
)
3964 return pid
? find_task_by_pid(pid
) : current
;
3967 /* Actually do priority change: must hold rq lock. */
3968 static void __setscheduler(struct task_struct
*p
, int policy
, int prio
)
3972 p
->rt_priority
= prio
;
3973 p
->normal_prio
= normal_prio(p
);
3974 /* we are holding p->pi_lock already */
3975 p
->prio
= rt_mutex_getprio(p
);
3977 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
3979 if (policy
== SCHED_BATCH
)
3985 * sched_setscheduler - change the scheduling policy and/or RT priority of
3987 * @p: the task in question.
3988 * @policy: new policy.
3989 * @param: structure containing the new RT priority.
3991 int sched_setscheduler(struct task_struct
*p
, int policy
,
3992 struct sched_param
*param
)
3995 int oldprio
, oldpolicy
= -1;
3996 prio_array_t
*array
;
3997 unsigned long flags
;
4000 /* may grab non-irq protected spin_locks */
4001 BUG_ON(in_interrupt());
4003 /* double check policy once rq lock held */
4005 policy
= oldpolicy
= p
->policy
;
4006 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4007 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
)
4010 * Valid priorities for SCHED_FIFO and SCHED_RR are
4011 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
4014 if (param
->sched_priority
< 0 ||
4015 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4016 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4018 if ((policy
== SCHED_NORMAL
|| policy
== SCHED_BATCH
)
4019 != (param
->sched_priority
== 0))
4023 * Allow unprivileged RT tasks to decrease priority:
4025 if (!capable(CAP_SYS_NICE
)) {
4027 * can't change policy, except between SCHED_NORMAL
4030 if (((policy
!= SCHED_NORMAL
&& p
->policy
!= SCHED_BATCH
) &&
4031 (policy
!= SCHED_BATCH
&& p
->policy
!= SCHED_NORMAL
)) &&
4032 !p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
)
4034 /* can't increase priority */
4035 if ((policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
) &&
4036 param
->sched_priority
> p
->rt_priority
&&
4037 param
->sched_priority
>
4038 p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
)
4040 /* can't change other user's priorities */
4041 if ((current
->euid
!= p
->euid
) &&
4042 (current
->euid
!= p
->uid
))
4046 retval
= security_task_setscheduler(p
, policy
, param
);
4050 * make sure no PI-waiters arrive (or leave) while we are
4051 * changing the priority of the task:
4053 spin_lock_irqsave(&p
->pi_lock
, flags
);
4055 * To be able to change p->policy safely, the apropriate
4056 * runqueue lock must be held.
4058 rq
= __task_rq_lock(p
);
4059 /* recheck policy now with rq lock held */
4060 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4061 policy
= oldpolicy
= -1;
4062 __task_rq_unlock(rq
);
4063 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4068 deactivate_task(p
, rq
);
4070 __setscheduler(p
, policy
, param
->sched_priority
);
4072 __activate_task(p
, rq
);
4074 * Reschedule if we are currently running on this runqueue and
4075 * our priority decreased, or if we are not currently running on
4076 * this runqueue and our priority is higher than the current's
4078 if (task_running(rq
, p
)) {
4079 if (p
->prio
> oldprio
)
4080 resched_task(rq
->curr
);
4081 } else if (TASK_PREEMPTS_CURR(p
, rq
))
4082 resched_task(rq
->curr
);
4084 __task_rq_unlock(rq
);
4085 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4087 rt_mutex_adjust_pi(p
);
4091 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4094 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4097 struct sched_param lparam
;
4098 struct task_struct
*p
;
4100 if (!param
|| pid
< 0)
4102 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4104 read_lock_irq(&tasklist_lock
);
4105 p
= find_process_by_pid(pid
);
4107 read_unlock_irq(&tasklist_lock
);
4111 read_unlock_irq(&tasklist_lock
);
4112 retval
= sched_setscheduler(p
, policy
, &lparam
);
4118 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4119 * @pid: the pid in question.
4120 * @policy: new policy.
4121 * @param: structure containing the new RT priority.
4123 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
4124 struct sched_param __user
*param
)
4126 /* negative values for policy are not valid */
4130 return do_sched_setscheduler(pid
, policy
, param
);
4134 * sys_sched_setparam - set/change the RT priority of a thread
4135 * @pid: the pid in question.
4136 * @param: structure containing the new RT priority.
4138 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4140 return do_sched_setscheduler(pid
, -1, param
);
4144 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4145 * @pid: the pid in question.
4147 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4149 int retval
= -EINVAL
;
4156 read_lock(&tasklist_lock
);
4157 p
= find_process_by_pid(pid
);
4159 retval
= security_task_getscheduler(p
);
4163 read_unlock(&tasklist_lock
);
4170 * sys_sched_getscheduler - get the RT priority of a thread
4171 * @pid: the pid in question.
4172 * @param: structure containing the RT priority.
4174 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4176 struct sched_param lp
;
4177 int retval
= -EINVAL
;
4180 if (!param
|| pid
< 0)
4183 read_lock(&tasklist_lock
);
4184 p
= find_process_by_pid(pid
);
4189 retval
= security_task_getscheduler(p
);
4193 lp
.sched_priority
= p
->rt_priority
;
4194 read_unlock(&tasklist_lock
);
4197 * This one might sleep, we cannot do it with a spinlock held ...
4199 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4205 read_unlock(&tasklist_lock
);
4209 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4213 cpumask_t cpus_allowed
;
4216 read_lock(&tasklist_lock
);
4218 p
= find_process_by_pid(pid
);
4220 read_unlock(&tasklist_lock
);
4221 unlock_cpu_hotplug();
4226 * It is not safe to call set_cpus_allowed with the
4227 * tasklist_lock held. We will bump the task_struct's
4228 * usage count and then drop tasklist_lock.
4231 read_unlock(&tasklist_lock
);
4234 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4235 !capable(CAP_SYS_NICE
))
4238 retval
= security_task_setscheduler(p
, 0, NULL
);
4242 cpus_allowed
= cpuset_cpus_allowed(p
);
4243 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4244 retval
= set_cpus_allowed(p
, new_mask
);
4248 unlock_cpu_hotplug();
4252 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4253 cpumask_t
*new_mask
)
4255 if (len
< sizeof(cpumask_t
)) {
4256 memset(new_mask
, 0, sizeof(cpumask_t
));
4257 } else if (len
> sizeof(cpumask_t
)) {
4258 len
= sizeof(cpumask_t
);
4260 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4264 * sys_sched_setaffinity - set the cpu affinity of a process
4265 * @pid: pid of the process
4266 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4267 * @user_mask_ptr: user-space pointer to the new cpu mask
4269 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4270 unsigned long __user
*user_mask_ptr
)
4275 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4279 return sched_setaffinity(pid
, new_mask
);
4283 * Represents all cpu's present in the system
4284 * In systems capable of hotplug, this map could dynamically grow
4285 * as new cpu's are detected in the system via any platform specific
4286 * method, such as ACPI for e.g.
4289 cpumask_t cpu_present_map __read_mostly
;
4290 EXPORT_SYMBOL(cpu_present_map
);
4293 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4294 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4297 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4303 read_lock(&tasklist_lock
);
4306 p
= find_process_by_pid(pid
);
4310 retval
= security_task_getscheduler(p
);
4314 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4317 read_unlock(&tasklist_lock
);
4318 unlock_cpu_hotplug();
4326 * sys_sched_getaffinity - get the cpu affinity of a process
4327 * @pid: pid of the process
4328 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4329 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4331 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4332 unsigned long __user
*user_mask_ptr
)
4337 if (len
< sizeof(cpumask_t
))
4340 ret
= sched_getaffinity(pid
, &mask
);
4344 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4347 return sizeof(cpumask_t
);
4351 * sys_sched_yield - yield the current processor to other threads.
4353 * this function yields the current CPU by moving the calling thread
4354 * to the expired array. If there are no other threads running on this
4355 * CPU then this function will return.
4357 asmlinkage
long sys_sched_yield(void)
4359 runqueue_t
*rq
= this_rq_lock();
4360 prio_array_t
*array
= current
->array
;
4361 prio_array_t
*target
= rq
->expired
;
4363 schedstat_inc(rq
, yld_cnt
);
4365 * We implement yielding by moving the task into the expired
4368 * (special rule: RT tasks will just roundrobin in the active
4371 if (rt_task(current
))
4372 target
= rq
->active
;
4374 if (array
->nr_active
== 1) {
4375 schedstat_inc(rq
, yld_act_empty
);
4376 if (!rq
->expired
->nr_active
)
4377 schedstat_inc(rq
, yld_both_empty
);
4378 } else if (!rq
->expired
->nr_active
)
4379 schedstat_inc(rq
, yld_exp_empty
);
4381 if (array
!= target
) {
4382 dequeue_task(current
, array
);
4383 enqueue_task(current
, target
);
4386 * requeue_task is cheaper so perform that if possible.
4388 requeue_task(current
, array
);
4391 * Since we are going to call schedule() anyway, there's
4392 * no need to preempt or enable interrupts:
4394 __release(rq
->lock
);
4395 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4396 _raw_spin_unlock(&rq
->lock
);
4397 preempt_enable_no_resched();
4404 static inline int __resched_legal(void)
4406 if (unlikely(preempt_count()))
4408 if (unlikely(system_state
!= SYSTEM_RUNNING
))
4413 static void __cond_resched(void)
4415 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4416 __might_sleep(__FILE__
, __LINE__
);
4419 * The BKS might be reacquired before we have dropped
4420 * PREEMPT_ACTIVE, which could trigger a second
4421 * cond_resched() call.
4424 add_preempt_count(PREEMPT_ACTIVE
);
4426 sub_preempt_count(PREEMPT_ACTIVE
);
4427 } while (need_resched());
4430 int __sched
cond_resched(void)
4432 if (need_resched() && __resched_legal()) {
4438 EXPORT_SYMBOL(cond_resched
);
4441 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4442 * call schedule, and on return reacquire the lock.
4444 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4445 * operations here to prevent schedule() from being called twice (once via
4446 * spin_unlock(), once by hand).
4448 int cond_resched_lock(spinlock_t
*lock
)
4452 if (need_lockbreak(lock
)) {
4458 if (need_resched() && __resched_legal()) {
4459 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4460 _raw_spin_unlock(lock
);
4461 preempt_enable_no_resched();
4468 EXPORT_SYMBOL(cond_resched_lock
);
4470 int __sched
cond_resched_softirq(void)
4472 BUG_ON(!in_softirq());
4474 if (need_resched() && __resched_legal()) {
4475 raw_local_irq_disable();
4477 raw_local_irq_enable();
4484 EXPORT_SYMBOL(cond_resched_softirq
);
4487 * yield - yield the current processor to other threads.
4489 * this is a shortcut for kernel-space yielding - it marks the
4490 * thread runnable and calls sys_sched_yield().
4492 void __sched
yield(void)
4494 set_current_state(TASK_RUNNING
);
4498 EXPORT_SYMBOL(yield
);
4501 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4502 * that process accounting knows that this is a task in IO wait state.
4504 * But don't do that if it is a deliberate, throttling IO wait (this task
4505 * has set its backing_dev_info: the queue against which it should throttle)
4507 void __sched
io_schedule(void)
4509 struct runqueue
*rq
= &__raw_get_cpu_var(runqueues
);
4511 atomic_inc(&rq
->nr_iowait
);
4513 atomic_dec(&rq
->nr_iowait
);
4516 EXPORT_SYMBOL(io_schedule
);
4518 long __sched
io_schedule_timeout(long timeout
)
4520 struct runqueue
*rq
= &__raw_get_cpu_var(runqueues
);
4523 atomic_inc(&rq
->nr_iowait
);
4524 ret
= schedule_timeout(timeout
);
4525 atomic_dec(&rq
->nr_iowait
);
4530 * sys_sched_get_priority_max - return maximum RT priority.
4531 * @policy: scheduling class.
4533 * this syscall returns the maximum rt_priority that can be used
4534 * by a given scheduling class.
4536 asmlinkage
long sys_sched_get_priority_max(int policy
)
4543 ret
= MAX_USER_RT_PRIO
-1;
4554 * sys_sched_get_priority_min - return minimum RT priority.
4555 * @policy: scheduling class.
4557 * this syscall returns the minimum rt_priority that can be used
4558 * by a given scheduling class.
4560 asmlinkage
long sys_sched_get_priority_min(int policy
)
4577 * sys_sched_rr_get_interval - return the default timeslice of a process.
4578 * @pid: pid of the process.
4579 * @interval: userspace pointer to the timeslice value.
4581 * this syscall writes the default timeslice value of a given process
4582 * into the user-space timespec buffer. A value of '0' means infinity.
4585 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4587 int retval
= -EINVAL
;
4595 read_lock(&tasklist_lock
);
4596 p
= find_process_by_pid(pid
);
4600 retval
= security_task_getscheduler(p
);
4604 jiffies_to_timespec(p
->policy
== SCHED_FIFO
?
4605 0 : task_timeslice(p
), &t
);
4606 read_unlock(&tasklist_lock
);
4607 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4611 read_unlock(&tasklist_lock
);
4615 static inline struct task_struct
*eldest_child(struct task_struct
*p
)
4617 if (list_empty(&p
->children
)) return NULL
;
4618 return list_entry(p
->children
.next
,struct task_struct
,sibling
);
4621 static inline struct task_struct
*older_sibling(struct task_struct
*p
)
4623 if (p
->sibling
.prev
==&p
->parent
->children
) return NULL
;
4624 return list_entry(p
->sibling
.prev
,struct task_struct
,sibling
);
4627 static inline struct task_struct
*younger_sibling(struct task_struct
*p
)
4629 if (p
->sibling
.next
==&p
->parent
->children
) return NULL
;
4630 return list_entry(p
->sibling
.next
,struct task_struct
,sibling
);
4633 static void show_task(task_t
*p
)
4637 unsigned long free
= 0;
4638 static const char *stat_nam
[] = { "R", "S", "D", "T", "t", "Z", "X" };
4640 printk("%-13.13s ", p
->comm
);
4641 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4642 if (state
< ARRAY_SIZE(stat_nam
))
4643 printk(stat_nam
[state
]);
4646 #if (BITS_PER_LONG == 32)
4647 if (state
== TASK_RUNNING
)
4648 printk(" running ");
4650 printk(" %08lX ", thread_saved_pc(p
));
4652 if (state
== TASK_RUNNING
)
4653 printk(" running task ");
4655 printk(" %016lx ", thread_saved_pc(p
));
4657 #ifdef CONFIG_DEBUG_STACK_USAGE
4659 unsigned long *n
= end_of_stack(p
);
4662 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4665 printk("%5lu %5d %6d ", free
, p
->pid
, p
->parent
->pid
);
4666 if ((relative
= eldest_child(p
)))
4667 printk("%5d ", relative
->pid
);
4670 if ((relative
= younger_sibling(p
)))
4671 printk("%7d", relative
->pid
);
4674 if ((relative
= older_sibling(p
)))
4675 printk(" %5d", relative
->pid
);
4679 printk(" (L-TLB)\n");
4681 printk(" (NOTLB)\n");
4683 if (state
!= TASK_RUNNING
)
4684 show_stack(p
, NULL
);
4687 void show_state(void)
4691 #if (BITS_PER_LONG == 32)
4694 printk(" task PC pid father child younger older\n");
4698 printk(" task PC pid father child younger older\n");
4700 read_lock(&tasklist_lock
);
4701 do_each_thread(g
, p
) {
4703 * reset the NMI-timeout, listing all files on a slow
4704 * console might take alot of time:
4706 touch_nmi_watchdog();
4708 } while_each_thread(g
, p
);
4710 read_unlock(&tasklist_lock
);
4711 debug_show_all_locks();
4715 * init_idle - set up an idle thread for a given CPU
4716 * @idle: task in question
4717 * @cpu: cpu the idle task belongs to
4719 * NOTE: this function does not set the idle thread's NEED_RESCHED
4720 * flag, to make booting more robust.
4722 void __devinit
init_idle(task_t
*idle
, int cpu
)
4724 runqueue_t
*rq
= cpu_rq(cpu
);
4725 unsigned long flags
;
4727 idle
->timestamp
= sched_clock();
4728 idle
->sleep_avg
= 0;
4730 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
4731 idle
->state
= TASK_RUNNING
;
4732 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4733 set_task_cpu(idle
, cpu
);
4735 spin_lock_irqsave(&rq
->lock
, flags
);
4736 rq
->curr
= rq
->idle
= idle
;
4737 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4740 spin_unlock_irqrestore(&rq
->lock
, flags
);
4742 /* Set the preempt count _outside_ the spinlocks! */
4743 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4744 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4746 task_thread_info(idle
)->preempt_count
= 0;
4751 * In a system that switches off the HZ timer nohz_cpu_mask
4752 * indicates which cpus entered this state. This is used
4753 * in the rcu update to wait only for active cpus. For system
4754 * which do not switch off the HZ timer nohz_cpu_mask should
4755 * always be CPU_MASK_NONE.
4757 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4761 * This is how migration works:
4763 * 1) we queue a migration_req_t structure in the source CPU's
4764 * runqueue and wake up that CPU's migration thread.
4765 * 2) we down() the locked semaphore => thread blocks.
4766 * 3) migration thread wakes up (implicitly it forces the migrated
4767 * thread off the CPU)
4768 * 4) it gets the migration request and checks whether the migrated
4769 * task is still in the wrong runqueue.
4770 * 5) if it's in the wrong runqueue then the migration thread removes
4771 * it and puts it into the right queue.
4772 * 6) migration thread up()s the semaphore.
4773 * 7) we wake up and the migration is done.
4777 * Change a given task's CPU affinity. Migrate the thread to a
4778 * proper CPU and schedule it away if the CPU it's executing on
4779 * is removed from the allowed bitmask.
4781 * NOTE: the caller must have a valid reference to the task, the
4782 * task must not exit() & deallocate itself prematurely. The
4783 * call is not atomic; no spinlocks may be held.
4785 int set_cpus_allowed(task_t
*p
, cpumask_t new_mask
)
4787 unsigned long flags
;
4789 migration_req_t req
;
4792 rq
= task_rq_lock(p
, &flags
);
4793 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4798 p
->cpus_allowed
= new_mask
;
4799 /* Can the task run on the task's current CPU? If so, we're done */
4800 if (cpu_isset(task_cpu(p
), new_mask
))
4803 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4804 /* Need help from migration thread: drop lock and wait. */
4805 task_rq_unlock(rq
, &flags
);
4806 wake_up_process(rq
->migration_thread
);
4807 wait_for_completion(&req
.done
);
4808 tlb_migrate_finish(p
->mm
);
4812 task_rq_unlock(rq
, &flags
);
4816 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4819 * Move (not current) task off this cpu, onto dest cpu. We're doing
4820 * this because either it can't run here any more (set_cpus_allowed()
4821 * away from this CPU, or CPU going down), or because we're
4822 * attempting to rebalance this task on exec (sched_exec).
4824 * So we race with normal scheduler movements, but that's OK, as long
4825 * as the task is no longer on this CPU.
4827 * Returns non-zero if task was successfully migrated.
4829 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4831 runqueue_t
*rq_dest
, *rq_src
;
4834 if (unlikely(cpu_is_offline(dest_cpu
)))
4837 rq_src
= cpu_rq(src_cpu
);
4838 rq_dest
= cpu_rq(dest_cpu
);
4840 double_rq_lock(rq_src
, rq_dest
);
4841 /* Already moved. */
4842 if (task_cpu(p
) != src_cpu
)
4844 /* Affinity changed (again). */
4845 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
4848 set_task_cpu(p
, dest_cpu
);
4851 * Sync timestamp with rq_dest's before activating.
4852 * The same thing could be achieved by doing this step
4853 * afterwards, and pretending it was a local activate.
4854 * This way is cleaner and logically correct.
4856 p
->timestamp
= p
->timestamp
- rq_src
->timestamp_last_tick
4857 + rq_dest
->timestamp_last_tick
;
4858 deactivate_task(p
, rq_src
);
4859 activate_task(p
, rq_dest
, 0);
4860 if (TASK_PREEMPTS_CURR(p
, rq_dest
))
4861 resched_task(rq_dest
->curr
);
4865 double_rq_unlock(rq_src
, rq_dest
);
4870 * migration_thread - this is a highprio system thread that performs
4871 * thread migration by bumping thread off CPU then 'pushing' onto
4874 static int migration_thread(void *data
)
4877 int cpu
= (long)data
;
4880 BUG_ON(rq
->migration_thread
!= current
);
4882 set_current_state(TASK_INTERRUPTIBLE
);
4883 while (!kthread_should_stop()) {
4884 struct list_head
*head
;
4885 migration_req_t
*req
;
4889 spin_lock_irq(&rq
->lock
);
4891 if (cpu_is_offline(cpu
)) {
4892 spin_unlock_irq(&rq
->lock
);
4896 if (rq
->active_balance
) {
4897 active_load_balance(rq
, cpu
);
4898 rq
->active_balance
= 0;
4901 head
= &rq
->migration_queue
;
4903 if (list_empty(head
)) {
4904 spin_unlock_irq(&rq
->lock
);
4906 set_current_state(TASK_INTERRUPTIBLE
);
4909 req
= list_entry(head
->next
, migration_req_t
, list
);
4910 list_del_init(head
->next
);
4912 spin_unlock(&rq
->lock
);
4913 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
4916 complete(&req
->done
);
4918 __set_current_state(TASK_RUNNING
);
4922 /* Wait for kthread_stop */
4923 set_current_state(TASK_INTERRUPTIBLE
);
4924 while (!kthread_should_stop()) {
4926 set_current_state(TASK_INTERRUPTIBLE
);
4928 __set_current_state(TASK_RUNNING
);
4932 #ifdef CONFIG_HOTPLUG_CPU
4933 /* Figure out where task on dead CPU should go, use force if neccessary. */
4934 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*tsk
)
4937 unsigned long flags
;
4943 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
4944 cpus_and(mask
, mask
, tsk
->cpus_allowed
);
4945 dest_cpu
= any_online_cpu(mask
);
4947 /* On any allowed CPU? */
4948 if (dest_cpu
== NR_CPUS
)
4949 dest_cpu
= any_online_cpu(tsk
->cpus_allowed
);
4951 /* No more Mr. Nice Guy. */
4952 if (dest_cpu
== NR_CPUS
) {
4953 rq
= task_rq_lock(tsk
, &flags
);
4954 cpus_setall(tsk
->cpus_allowed
);
4955 dest_cpu
= any_online_cpu(tsk
->cpus_allowed
);
4956 task_rq_unlock(rq
, &flags
);
4959 * Don't tell them about moving exiting tasks or
4960 * kernel threads (both mm NULL), since they never
4963 if (tsk
->mm
&& printk_ratelimit())
4964 printk(KERN_INFO
"process %d (%s) no "
4965 "longer affine to cpu%d\n",
4966 tsk
->pid
, tsk
->comm
, dead_cpu
);
4968 if (!__migrate_task(tsk
, dead_cpu
, dest_cpu
))
4973 * While a dead CPU has no uninterruptible tasks queued at this point,
4974 * it might still have a nonzero ->nr_uninterruptible counter, because
4975 * for performance reasons the counter is not stricly tracking tasks to
4976 * their home CPUs. So we just add the counter to another CPU's counter,
4977 * to keep the global sum constant after CPU-down:
4979 static void migrate_nr_uninterruptible(runqueue_t
*rq_src
)
4981 runqueue_t
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
4982 unsigned long flags
;
4984 local_irq_save(flags
);
4985 double_rq_lock(rq_src
, rq_dest
);
4986 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
4987 rq_src
->nr_uninterruptible
= 0;
4988 double_rq_unlock(rq_src
, rq_dest
);
4989 local_irq_restore(flags
);
4992 /* Run through task list and migrate tasks from the dead cpu. */
4993 static void migrate_live_tasks(int src_cpu
)
4995 struct task_struct
*tsk
, *t
;
4997 write_lock_irq(&tasklist_lock
);
4999 do_each_thread(t
, tsk
) {
5003 if (task_cpu(tsk
) == src_cpu
)
5004 move_task_off_dead_cpu(src_cpu
, tsk
);
5005 } while_each_thread(t
, tsk
);
5007 write_unlock_irq(&tasklist_lock
);
5010 /* Schedules idle task to be the next runnable task on current CPU.
5011 * It does so by boosting its priority to highest possible and adding it to
5012 * the _front_ of runqueue. Used by CPU offline code.
5014 void sched_idle_next(void)
5016 int cpu
= smp_processor_id();
5017 runqueue_t
*rq
= this_rq();
5018 struct task_struct
*p
= rq
->idle
;
5019 unsigned long flags
;
5021 /* cpu has to be offline */
5022 BUG_ON(cpu_online(cpu
));
5024 /* Strictly not necessary since rest of the CPUs are stopped by now
5025 * and interrupts disabled on current cpu.
5027 spin_lock_irqsave(&rq
->lock
, flags
);
5029 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5030 /* Add idle task to _front_ of it's priority queue */
5031 __activate_idle_task(p
, rq
);
5033 spin_unlock_irqrestore(&rq
->lock
, flags
);
5036 /* Ensures that the idle task is using init_mm right before its cpu goes
5039 void idle_task_exit(void)
5041 struct mm_struct
*mm
= current
->active_mm
;
5043 BUG_ON(cpu_online(smp_processor_id()));
5046 switch_mm(mm
, &init_mm
, current
);
5050 static void migrate_dead(unsigned int dead_cpu
, task_t
*tsk
)
5052 struct runqueue
*rq
= cpu_rq(dead_cpu
);
5054 /* Must be exiting, otherwise would be on tasklist. */
5055 BUG_ON(tsk
->exit_state
!= EXIT_ZOMBIE
&& tsk
->exit_state
!= EXIT_DEAD
);
5057 /* Cannot have done final schedule yet: would have vanished. */
5058 BUG_ON(tsk
->flags
& PF_DEAD
);
5060 get_task_struct(tsk
);
5063 * Drop lock around migration; if someone else moves it,
5064 * that's OK. No task can be added to this CPU, so iteration is
5067 spin_unlock_irq(&rq
->lock
);
5068 move_task_off_dead_cpu(dead_cpu
, tsk
);
5069 spin_lock_irq(&rq
->lock
);
5071 put_task_struct(tsk
);
5074 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5075 static void migrate_dead_tasks(unsigned int dead_cpu
)
5078 struct runqueue
*rq
= cpu_rq(dead_cpu
);
5080 for (arr
= 0; arr
< 2; arr
++) {
5081 for (i
= 0; i
< MAX_PRIO
; i
++) {
5082 struct list_head
*list
= &rq
->arrays
[arr
].queue
[i
];
5083 while (!list_empty(list
))
5084 migrate_dead(dead_cpu
,
5085 list_entry(list
->next
, task_t
,
5090 #endif /* CONFIG_HOTPLUG_CPU */
5093 * migration_call - callback that gets triggered when a CPU is added.
5094 * Here we can start up the necessary migration thread for the new CPU.
5096 static int __cpuinit
migration_call(struct notifier_block
*nfb
,
5097 unsigned long action
,
5100 int cpu
= (long)hcpu
;
5101 struct task_struct
*p
;
5102 struct runqueue
*rq
;
5103 unsigned long flags
;
5106 case CPU_UP_PREPARE
:
5107 p
= kthread_create(migration_thread
, hcpu
, "migration/%d",cpu
);
5110 p
->flags
|= PF_NOFREEZE
;
5111 kthread_bind(p
, cpu
);
5112 /* Must be high prio: stop_machine expects to yield to it. */
5113 rq
= task_rq_lock(p
, &flags
);
5114 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5115 task_rq_unlock(rq
, &flags
);
5116 cpu_rq(cpu
)->migration_thread
= p
;
5119 /* Strictly unneccessary, as first user will wake it. */
5120 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5122 #ifdef CONFIG_HOTPLUG_CPU
5123 case CPU_UP_CANCELED
:
5124 if (!cpu_rq(cpu
)->migration_thread
)
5126 /* Unbind it from offline cpu so it can run. Fall thru. */
5127 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5128 any_online_cpu(cpu_online_map
));
5129 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5130 cpu_rq(cpu
)->migration_thread
= NULL
;
5133 migrate_live_tasks(cpu
);
5135 kthread_stop(rq
->migration_thread
);
5136 rq
->migration_thread
= NULL
;
5137 /* Idle task back to normal (off runqueue, low prio) */
5138 rq
= task_rq_lock(rq
->idle
, &flags
);
5139 deactivate_task(rq
->idle
, rq
);
5140 rq
->idle
->static_prio
= MAX_PRIO
;
5141 __setscheduler(rq
->idle
, SCHED_NORMAL
, 0);
5142 migrate_dead_tasks(cpu
);
5143 task_rq_unlock(rq
, &flags
);
5144 migrate_nr_uninterruptible(rq
);
5145 BUG_ON(rq
->nr_running
!= 0);
5147 /* No need to migrate the tasks: it was best-effort if
5148 * they didn't do lock_cpu_hotplug(). Just wake up
5149 * the requestors. */
5150 spin_lock_irq(&rq
->lock
);
5151 while (!list_empty(&rq
->migration_queue
)) {
5152 migration_req_t
*req
;
5153 req
= list_entry(rq
->migration_queue
.next
,
5154 migration_req_t
, list
);
5155 list_del_init(&req
->list
);
5156 complete(&req
->done
);
5158 spin_unlock_irq(&rq
->lock
);
5165 /* Register at highest priority so that task migration (migrate_all_tasks)
5166 * happens before everything else.
5168 static struct notifier_block __cpuinitdata migration_notifier
= {
5169 .notifier_call
= migration_call
,
5173 int __init
migration_init(void)
5175 void *cpu
= (void *)(long)smp_processor_id();
5176 /* Start one for boot CPU. */
5177 migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5178 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5179 register_cpu_notifier(&migration_notifier
);
5185 #undef SCHED_DOMAIN_DEBUG
5186 #ifdef SCHED_DOMAIN_DEBUG
5187 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5192 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5196 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5201 struct sched_group
*group
= sd
->groups
;
5202 cpumask_t groupmask
;
5204 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5205 cpus_clear(groupmask
);
5208 for (i
= 0; i
< level
+ 1; i
++)
5210 printk("domain %d: ", level
);
5212 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5213 printk("does not load-balance\n");
5215 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain has parent");
5219 printk("span %s\n", str
);
5221 if (!cpu_isset(cpu
, sd
->span
))
5222 printk(KERN_ERR
"ERROR: domain->span does not contain CPU%d\n", cpu
);
5223 if (!cpu_isset(cpu
, group
->cpumask
))
5224 printk(KERN_ERR
"ERROR: domain->groups does not contain CPU%d\n", cpu
);
5227 for (i
= 0; i
< level
+ 2; i
++)
5233 printk(KERN_ERR
"ERROR: group is NULL\n");
5237 if (!group
->cpu_power
) {
5239 printk(KERN_ERR
"ERROR: domain->cpu_power not set\n");
5242 if (!cpus_weight(group
->cpumask
)) {
5244 printk(KERN_ERR
"ERROR: empty group\n");
5247 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5249 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5252 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5254 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5257 group
= group
->next
;
5258 } while (group
!= sd
->groups
);
5261 if (!cpus_equal(sd
->span
, groupmask
))
5262 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5268 if (!cpus_subset(groupmask
, sd
->span
))
5269 printk(KERN_ERR
"ERROR: parent span is not a superset of domain->span\n");
5275 #define sched_domain_debug(sd, cpu) {}
5278 static int sd_degenerate(struct sched_domain
*sd
)
5280 if (cpus_weight(sd
->span
) == 1)
5283 /* Following flags need at least 2 groups */
5284 if (sd
->flags
& (SD_LOAD_BALANCE
|
5285 SD_BALANCE_NEWIDLE
|
5288 if (sd
->groups
!= sd
->groups
->next
)
5292 /* Following flags don't use groups */
5293 if (sd
->flags
& (SD_WAKE_IDLE
|
5301 static int sd_parent_degenerate(struct sched_domain
*sd
,
5302 struct sched_domain
*parent
)
5304 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5306 if (sd_degenerate(parent
))
5309 if (!cpus_equal(sd
->span
, parent
->span
))
5312 /* Does parent contain flags not in child? */
5313 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5314 if (cflags
& SD_WAKE_AFFINE
)
5315 pflags
&= ~SD_WAKE_BALANCE
;
5316 /* Flags needing groups don't count if only 1 group in parent */
5317 if (parent
->groups
== parent
->groups
->next
) {
5318 pflags
&= ~(SD_LOAD_BALANCE
|
5319 SD_BALANCE_NEWIDLE
|
5323 if (~cflags
& pflags
)
5330 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5331 * hold the hotplug lock.
5333 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5335 runqueue_t
*rq
= cpu_rq(cpu
);
5336 struct sched_domain
*tmp
;
5338 /* Remove the sched domains which do not contribute to scheduling. */
5339 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5340 struct sched_domain
*parent
= tmp
->parent
;
5343 if (sd_parent_degenerate(tmp
, parent
))
5344 tmp
->parent
= parent
->parent
;
5347 if (sd
&& sd_degenerate(sd
))
5350 sched_domain_debug(sd
, cpu
);
5352 rcu_assign_pointer(rq
->sd
, sd
);
5355 /* cpus with isolated domains */
5356 static cpumask_t __devinitdata cpu_isolated_map
= CPU_MASK_NONE
;
5358 /* Setup the mask of cpus configured for isolated domains */
5359 static int __init
isolated_cpu_setup(char *str
)
5361 int ints
[NR_CPUS
], i
;
5363 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5364 cpus_clear(cpu_isolated_map
);
5365 for (i
= 1; i
<= ints
[0]; i
++)
5366 if (ints
[i
] < NR_CPUS
)
5367 cpu_set(ints
[i
], cpu_isolated_map
);
5371 __setup ("isolcpus=", isolated_cpu_setup
);
5374 * init_sched_build_groups takes an array of groups, the cpumask we wish
5375 * to span, and a pointer to a function which identifies what group a CPU
5376 * belongs to. The return value of group_fn must be a valid index into the
5377 * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
5378 * keep track of groups covered with a cpumask_t).
5380 * init_sched_build_groups will build a circular linked list of the groups
5381 * covered by the given span, and will set each group's ->cpumask correctly,
5382 * and ->cpu_power to 0.
5384 static void init_sched_build_groups(struct sched_group groups
[], cpumask_t span
,
5385 int (*group_fn
)(int cpu
))
5387 struct sched_group
*first
= NULL
, *last
= NULL
;
5388 cpumask_t covered
= CPU_MASK_NONE
;
5391 for_each_cpu_mask(i
, span
) {
5392 int group
= group_fn(i
);
5393 struct sched_group
*sg
= &groups
[group
];
5396 if (cpu_isset(i
, covered
))
5399 sg
->cpumask
= CPU_MASK_NONE
;
5402 for_each_cpu_mask(j
, span
) {
5403 if (group_fn(j
) != group
)
5406 cpu_set(j
, covered
);
5407 cpu_set(j
, sg
->cpumask
);
5418 #define SD_NODES_PER_DOMAIN 16
5421 * Self-tuning task migration cost measurement between source and target CPUs.
5423 * This is done by measuring the cost of manipulating buffers of varying
5424 * sizes. For a given buffer-size here are the steps that are taken:
5426 * 1) the source CPU reads+dirties a shared buffer
5427 * 2) the target CPU reads+dirties the same shared buffer
5429 * We measure how long they take, in the following 4 scenarios:
5431 * - source: CPU1, target: CPU2 | cost1
5432 * - source: CPU2, target: CPU1 | cost2
5433 * - source: CPU1, target: CPU1 | cost3
5434 * - source: CPU2, target: CPU2 | cost4
5436 * We then calculate the cost3+cost4-cost1-cost2 difference - this is
5437 * the cost of migration.
5439 * We then start off from a small buffer-size and iterate up to larger
5440 * buffer sizes, in 5% steps - measuring each buffer-size separately, and
5441 * doing a maximum search for the cost. (The maximum cost for a migration
5442 * normally occurs when the working set size is around the effective cache
5445 #define SEARCH_SCOPE 2
5446 #define MIN_CACHE_SIZE (64*1024U)
5447 #define DEFAULT_CACHE_SIZE (5*1024*1024U)
5448 #define ITERATIONS 1
5449 #define SIZE_THRESH 130
5450 #define COST_THRESH 130
5453 * The migration cost is a function of 'domain distance'. Domain
5454 * distance is the number of steps a CPU has to iterate down its
5455 * domain tree to share a domain with the other CPU. The farther
5456 * two CPUs are from each other, the larger the distance gets.
5458 * Note that we use the distance only to cache measurement results,
5459 * the distance value is not used numerically otherwise. When two
5460 * CPUs have the same distance it is assumed that the migration
5461 * cost is the same. (this is a simplification but quite practical)
5463 #define MAX_DOMAIN_DISTANCE 32
5465 static unsigned long long migration_cost
[MAX_DOMAIN_DISTANCE
] =
5466 { [ 0 ... MAX_DOMAIN_DISTANCE
-1 ] =
5468 * Architectures may override the migration cost and thus avoid
5469 * boot-time calibration. Unit is nanoseconds. Mostly useful for
5470 * virtualized hardware:
5472 #ifdef CONFIG_DEFAULT_MIGRATION_COST
5473 CONFIG_DEFAULT_MIGRATION_COST
5480 * Allow override of migration cost - in units of microseconds.
5481 * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost
5482 * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs:
5484 static int __init
migration_cost_setup(char *str
)
5486 int ints
[MAX_DOMAIN_DISTANCE
+1], i
;
5488 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5490 printk("#ints: %d\n", ints
[0]);
5491 for (i
= 1; i
<= ints
[0]; i
++) {
5492 migration_cost
[i
-1] = (unsigned long long)ints
[i
]*1000;
5493 printk("migration_cost[%d]: %Ld\n", i
-1, migration_cost
[i
-1]);
5498 __setup ("migration_cost=", migration_cost_setup
);
5501 * Global multiplier (divisor) for migration-cutoff values,
5502 * in percentiles. E.g. use a value of 150 to get 1.5 times
5503 * longer cache-hot cutoff times.
5505 * (We scale it from 100 to 128 to long long handling easier.)
5508 #define MIGRATION_FACTOR_SCALE 128
5510 static unsigned int migration_factor
= MIGRATION_FACTOR_SCALE
;
5512 static int __init
setup_migration_factor(char *str
)
5514 get_option(&str
, &migration_factor
);
5515 migration_factor
= migration_factor
* MIGRATION_FACTOR_SCALE
/ 100;
5519 __setup("migration_factor=", setup_migration_factor
);
5522 * Estimated distance of two CPUs, measured via the number of domains
5523 * we have to pass for the two CPUs to be in the same span:
5525 static unsigned long domain_distance(int cpu1
, int cpu2
)
5527 unsigned long distance
= 0;
5528 struct sched_domain
*sd
;
5530 for_each_domain(cpu1
, sd
) {
5531 WARN_ON(!cpu_isset(cpu1
, sd
->span
));
5532 if (cpu_isset(cpu2
, sd
->span
))
5536 if (distance
>= MAX_DOMAIN_DISTANCE
) {
5538 distance
= MAX_DOMAIN_DISTANCE
-1;
5544 static unsigned int migration_debug
;
5546 static int __init
setup_migration_debug(char *str
)
5548 get_option(&str
, &migration_debug
);
5552 __setup("migration_debug=", setup_migration_debug
);
5555 * Maximum cache-size that the scheduler should try to measure.
5556 * Architectures with larger caches should tune this up during
5557 * bootup. Gets used in the domain-setup code (i.e. during SMP
5560 unsigned int max_cache_size
;
5562 static int __init
setup_max_cache_size(char *str
)
5564 get_option(&str
, &max_cache_size
);
5568 __setup("max_cache_size=", setup_max_cache_size
);
5571 * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This
5572 * is the operation that is timed, so we try to generate unpredictable
5573 * cachemisses that still end up filling the L2 cache:
5575 static void touch_cache(void *__cache
, unsigned long __size
)
5577 unsigned long size
= __size
/sizeof(long), chunk1
= size
/3,
5579 unsigned long *cache
= __cache
;
5582 for (i
= 0; i
< size
/6; i
+= 8) {
5585 case 1: cache
[size
-1-i
]++;
5586 case 2: cache
[chunk1
-i
]++;
5587 case 3: cache
[chunk1
+i
]++;
5588 case 4: cache
[chunk2
-i
]++;
5589 case 5: cache
[chunk2
+i
]++;
5595 * Measure the cache-cost of one task migration. Returns in units of nsec.
5597 static unsigned long long measure_one(void *cache
, unsigned long size
,
5598 int source
, int target
)
5600 cpumask_t mask
, saved_mask
;
5601 unsigned long long t0
, t1
, t2
, t3
, cost
;
5603 saved_mask
= current
->cpus_allowed
;
5606 * Flush source caches to RAM and invalidate them:
5611 * Migrate to the source CPU:
5613 mask
= cpumask_of_cpu(source
);
5614 set_cpus_allowed(current
, mask
);
5615 WARN_ON(smp_processor_id() != source
);
5618 * Dirty the working set:
5621 touch_cache(cache
, size
);
5625 * Migrate to the target CPU, dirty the L2 cache and access
5626 * the shared buffer. (which represents the working set
5627 * of a migrated task.)
5629 mask
= cpumask_of_cpu(target
);
5630 set_cpus_allowed(current
, mask
);
5631 WARN_ON(smp_processor_id() != target
);
5634 touch_cache(cache
, size
);
5637 cost
= t1
-t0
+ t3
-t2
;
5639 if (migration_debug
>= 2)
5640 printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n",
5641 source
, target
, t1
-t0
, t1
-t0
, t3
-t2
, cost
);
5643 * Flush target caches to RAM and invalidate them:
5647 set_cpus_allowed(current
, saved_mask
);
5653 * Measure a series of task migrations and return the average
5654 * result. Since this code runs early during bootup the system
5655 * is 'undisturbed' and the average latency makes sense.
5657 * The algorithm in essence auto-detects the relevant cache-size,
5658 * so it will properly detect different cachesizes for different
5659 * cache-hierarchies, depending on how the CPUs are connected.
5661 * Architectures can prime the upper limit of the search range via
5662 * max_cache_size, otherwise the search range defaults to 20MB...64K.
5664 static unsigned long long
5665 measure_cost(int cpu1
, int cpu2
, void *cache
, unsigned int size
)
5667 unsigned long long cost1
, cost2
;
5671 * Measure the migration cost of 'size' bytes, over an
5672 * average of 10 runs:
5674 * (We perturb the cache size by a small (0..4k)
5675 * value to compensate size/alignment related artifacts.
5676 * We also subtract the cost of the operation done on
5682 * dry run, to make sure we start off cache-cold on cpu1,
5683 * and to get any vmalloc pagefaults in advance:
5685 measure_one(cache
, size
, cpu1
, cpu2
);
5686 for (i
= 0; i
< ITERATIONS
; i
++)
5687 cost1
+= measure_one(cache
, size
- i
*1024, cpu1
, cpu2
);
5689 measure_one(cache
, size
, cpu2
, cpu1
);
5690 for (i
= 0; i
< ITERATIONS
; i
++)
5691 cost1
+= measure_one(cache
, size
- i
*1024, cpu2
, cpu1
);
5694 * (We measure the non-migrating [cached] cost on both
5695 * cpu1 and cpu2, to handle CPUs with different speeds)
5699 measure_one(cache
, size
, cpu1
, cpu1
);
5700 for (i
= 0; i
< ITERATIONS
; i
++)
5701 cost2
+= measure_one(cache
, size
- i
*1024, cpu1
, cpu1
);
5703 measure_one(cache
, size
, cpu2
, cpu2
);
5704 for (i
= 0; i
< ITERATIONS
; i
++)
5705 cost2
+= measure_one(cache
, size
- i
*1024, cpu2
, cpu2
);
5708 * Get the per-iteration migration cost:
5710 do_div(cost1
, 2*ITERATIONS
);
5711 do_div(cost2
, 2*ITERATIONS
);
5713 return cost1
- cost2
;
5716 static unsigned long long measure_migration_cost(int cpu1
, int cpu2
)
5718 unsigned long long max_cost
= 0, fluct
= 0, avg_fluct
= 0;
5719 unsigned int max_size
, size
, size_found
= 0;
5720 long long cost
= 0, prev_cost
;
5724 * Search from max_cache_size*5 down to 64K - the real relevant
5725 * cachesize has to lie somewhere inbetween.
5727 if (max_cache_size
) {
5728 max_size
= max(max_cache_size
* SEARCH_SCOPE
, MIN_CACHE_SIZE
);
5729 size
= max(max_cache_size
/ SEARCH_SCOPE
, MIN_CACHE_SIZE
);
5732 * Since we have no estimation about the relevant
5735 max_size
= DEFAULT_CACHE_SIZE
* SEARCH_SCOPE
;
5736 size
= MIN_CACHE_SIZE
;
5739 if (!cpu_online(cpu1
) || !cpu_online(cpu2
)) {
5740 printk("cpu %d and %d not both online!\n", cpu1
, cpu2
);
5745 * Allocate the working set:
5747 cache
= vmalloc(max_size
);
5749 printk("could not vmalloc %d bytes for cache!\n", 2*max_size
);
5750 return 1000000; // return 1 msec on very small boxen
5753 while (size
<= max_size
) {
5755 cost
= measure_cost(cpu1
, cpu2
, cache
, size
);
5761 if (max_cost
< cost
) {
5767 * Calculate average fluctuation, we use this to prevent
5768 * noise from triggering an early break out of the loop:
5770 fluct
= abs(cost
- prev_cost
);
5771 avg_fluct
= (avg_fluct
+ fluct
)/2;
5773 if (migration_debug
)
5774 printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): (%8Ld %8Ld)\n",
5776 (long)cost
/ 1000000,
5777 ((long)cost
/ 100000) % 10,
5778 (long)max_cost
/ 1000000,
5779 ((long)max_cost
/ 100000) % 10,
5780 domain_distance(cpu1
, cpu2
),
5784 * If we iterated at least 20% past the previous maximum,
5785 * and the cost has dropped by more than 20% already,
5786 * (taking fluctuations into account) then we assume to
5787 * have found the maximum and break out of the loop early:
5789 if (size_found
&& (size
*100 > size_found
*SIZE_THRESH
))
5790 if (cost
+avg_fluct
<= 0 ||
5791 max_cost
*100 > (cost
+avg_fluct
)*COST_THRESH
) {
5793 if (migration_debug
)
5794 printk("-> found max.\n");
5798 * Increase the cachesize in 10% steps:
5800 size
= size
* 10 / 9;
5803 if (migration_debug
)
5804 printk("[%d][%d] working set size found: %d, cost: %Ld\n",
5805 cpu1
, cpu2
, size_found
, max_cost
);
5810 * A task is considered 'cache cold' if at least 2 times
5811 * the worst-case cost of migration has passed.
5813 * (this limit is only listened to if the load-balancing
5814 * situation is 'nice' - if there is a large imbalance we
5815 * ignore it for the sake of CPU utilization and
5816 * processing fairness.)
5818 return 2 * max_cost
* migration_factor
/ MIGRATION_FACTOR_SCALE
;
5821 static void calibrate_migration_costs(const cpumask_t
*cpu_map
)
5823 int cpu1
= -1, cpu2
= -1, cpu
, orig_cpu
= raw_smp_processor_id();
5824 unsigned long j0
, j1
, distance
, max_distance
= 0;
5825 struct sched_domain
*sd
;
5830 * First pass - calculate the cacheflush times:
5832 for_each_cpu_mask(cpu1
, *cpu_map
) {
5833 for_each_cpu_mask(cpu2
, *cpu_map
) {
5836 distance
= domain_distance(cpu1
, cpu2
);
5837 max_distance
= max(max_distance
, distance
);
5839 * No result cached yet?
5841 if (migration_cost
[distance
] == -1LL)
5842 migration_cost
[distance
] =
5843 measure_migration_cost(cpu1
, cpu2
);
5847 * Second pass - update the sched domain hierarchy with
5848 * the new cache-hot-time estimations:
5850 for_each_cpu_mask(cpu
, *cpu_map
) {
5852 for_each_domain(cpu
, sd
) {
5853 sd
->cache_hot_time
= migration_cost
[distance
];
5860 if (migration_debug
)
5861 printk("migration: max_cache_size: %d, cpu: %d MHz:\n",
5869 if (system_state
== SYSTEM_BOOTING
) {
5870 printk("migration_cost=");
5871 for (distance
= 0; distance
<= max_distance
; distance
++) {
5874 printk("%ld", (long)migration_cost
[distance
] / 1000);
5879 if (migration_debug
)
5880 printk("migration: %ld seconds\n", (j1
-j0
)/HZ
);
5883 * Move back to the original CPU. NUMA-Q gets confused
5884 * if we migrate to another quad during bootup.
5886 if (raw_smp_processor_id() != orig_cpu
) {
5887 cpumask_t mask
= cpumask_of_cpu(orig_cpu
),
5888 saved_mask
= current
->cpus_allowed
;
5890 set_cpus_allowed(current
, mask
);
5891 set_cpus_allowed(current
, saved_mask
);
5898 * find_next_best_node - find the next node to include in a sched_domain
5899 * @node: node whose sched_domain we're building
5900 * @used_nodes: nodes already in the sched_domain
5902 * Find the next node to include in a given scheduling domain. Simply
5903 * finds the closest node not already in the @used_nodes map.
5905 * Should use nodemask_t.
5907 static int find_next_best_node(int node
, unsigned long *used_nodes
)
5909 int i
, n
, val
, min_val
, best_node
= 0;
5913 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5914 /* Start at @node */
5915 n
= (node
+ i
) % MAX_NUMNODES
;
5917 if (!nr_cpus_node(n
))
5920 /* Skip already used nodes */
5921 if (test_bit(n
, used_nodes
))
5924 /* Simple min distance search */
5925 val
= node_distance(node
, n
);
5927 if (val
< min_val
) {
5933 set_bit(best_node
, used_nodes
);
5938 * sched_domain_node_span - get a cpumask for a node's sched_domain
5939 * @node: node whose cpumask we're constructing
5940 * @size: number of nodes to include in this span
5942 * Given a node, construct a good cpumask for its sched_domain to span. It
5943 * should be one that prevents unnecessary balancing, but also spreads tasks
5946 static cpumask_t
sched_domain_node_span(int node
)
5949 cpumask_t span
, nodemask
;
5950 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
5953 bitmap_zero(used_nodes
, MAX_NUMNODES
);
5955 nodemask
= node_to_cpumask(node
);
5956 cpus_or(span
, span
, nodemask
);
5957 set_bit(node
, used_nodes
);
5959 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5960 int next_node
= find_next_best_node(node
, used_nodes
);
5961 nodemask
= node_to_cpumask(next_node
);
5962 cpus_or(span
, span
, nodemask
);
5969 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
5971 * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we
5972 * can switch it on easily if needed.
5974 #ifdef CONFIG_SCHED_SMT
5975 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
5976 static struct sched_group sched_group_cpus
[NR_CPUS
];
5977 static int cpu_to_cpu_group(int cpu
)
5983 #ifdef CONFIG_SCHED_MC
5984 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
5985 static struct sched_group
*sched_group_core_bycpu
[NR_CPUS
];
5988 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5989 static int cpu_to_core_group(int cpu
)
5991 return first_cpu(cpu_sibling_map
[cpu
]);
5993 #elif defined(CONFIG_SCHED_MC)
5994 static int cpu_to_core_group(int cpu
)
6000 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
6001 static struct sched_group
*sched_group_phys_bycpu
[NR_CPUS
];
6002 static int cpu_to_phys_group(int cpu
)
6004 #if defined(CONFIG_SCHED_MC)
6005 cpumask_t mask
= cpu_coregroup_map(cpu
);
6006 return first_cpu(mask
);
6007 #elif defined(CONFIG_SCHED_SMT)
6008 return first_cpu(cpu_sibling_map
[cpu
]);
6016 * The init_sched_build_groups can't handle what we want to do with node
6017 * groups, so roll our own. Now each node has its own list of groups which
6018 * gets dynamically allocated.
6020 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
6021 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
6023 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
6024 static struct sched_group
*sched_group_allnodes_bycpu
[NR_CPUS
];
6026 static int cpu_to_allnodes_group(int cpu
)
6028 return cpu_to_node(cpu
);
6030 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6032 struct sched_group
*sg
= group_head
;
6038 for_each_cpu_mask(j
, sg
->cpumask
) {
6039 struct sched_domain
*sd
;
6041 sd
= &per_cpu(phys_domains
, j
);
6042 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
6044 * Only add "power" once for each
6050 sg
->cpu_power
+= sd
->groups
->cpu_power
;
6053 if (sg
!= group_head
)
6058 /* Free memory allocated for various sched_group structures */
6059 static void free_sched_groups(const cpumask_t
*cpu_map
)
6065 for_each_cpu_mask(cpu
, *cpu_map
) {
6066 struct sched_group
*sched_group_allnodes
6067 = sched_group_allnodes_bycpu
[cpu
];
6068 struct sched_group
**sched_group_nodes
6069 = sched_group_nodes_bycpu
[cpu
];
6071 if (sched_group_allnodes
) {
6072 kfree(sched_group_allnodes
);
6073 sched_group_allnodes_bycpu
[cpu
] = NULL
;
6076 if (!sched_group_nodes
)
6079 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6080 cpumask_t nodemask
= node_to_cpumask(i
);
6081 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6083 cpus_and(nodemask
, nodemask
, *cpu_map
);
6084 if (cpus_empty(nodemask
))
6094 if (oldsg
!= sched_group_nodes
[i
])
6097 kfree(sched_group_nodes
);
6098 sched_group_nodes_bycpu
[cpu
] = NULL
;
6101 for_each_cpu_mask(cpu
, *cpu_map
) {
6102 if (sched_group_phys_bycpu
[cpu
]) {
6103 kfree(sched_group_phys_bycpu
[cpu
]);
6104 sched_group_phys_bycpu
[cpu
] = NULL
;
6106 #ifdef CONFIG_SCHED_MC
6107 if (sched_group_core_bycpu
[cpu
]) {
6108 kfree(sched_group_core_bycpu
[cpu
]);
6109 sched_group_core_bycpu
[cpu
] = NULL
;
6116 * Build sched domains for a given set of cpus and attach the sched domains
6117 * to the individual cpus
6119 static int build_sched_domains(const cpumask_t
*cpu_map
)
6122 struct sched_group
*sched_group_phys
= NULL
;
6123 #ifdef CONFIG_SCHED_MC
6124 struct sched_group
*sched_group_core
= NULL
;
6127 struct sched_group
**sched_group_nodes
= NULL
;
6128 struct sched_group
*sched_group_allnodes
= NULL
;
6131 * Allocate the per-node list of sched groups
6133 sched_group_nodes
= kzalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
6135 if (!sched_group_nodes
) {
6136 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6139 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6143 * Set up domains for cpus specified by the cpu_map.
6145 for_each_cpu_mask(i
, *cpu_map
) {
6147 struct sched_domain
*sd
= NULL
, *p
;
6148 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6150 cpus_and(nodemask
, nodemask
, *cpu_map
);
6153 if (cpus_weight(*cpu_map
)
6154 > SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6155 if (!sched_group_allnodes
) {
6156 sched_group_allnodes
6157 = kmalloc(sizeof(struct sched_group
)
6160 if (!sched_group_allnodes
) {
6162 "Can not alloc allnodes sched group\n");
6165 sched_group_allnodes_bycpu
[i
]
6166 = sched_group_allnodes
;
6168 sd
= &per_cpu(allnodes_domains
, i
);
6169 *sd
= SD_ALLNODES_INIT
;
6170 sd
->span
= *cpu_map
;
6171 group
= cpu_to_allnodes_group(i
);
6172 sd
->groups
= &sched_group_allnodes
[group
];
6177 sd
= &per_cpu(node_domains
, i
);
6179 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6181 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6184 if (!sched_group_phys
) {
6186 = kmalloc(sizeof(struct sched_group
) * NR_CPUS
,
6188 if (!sched_group_phys
) {
6189 printk (KERN_WARNING
"Can not alloc phys sched"
6193 sched_group_phys_bycpu
[i
] = sched_group_phys
;
6197 sd
= &per_cpu(phys_domains
, i
);
6198 group
= cpu_to_phys_group(i
);
6200 sd
->span
= nodemask
;
6202 sd
->groups
= &sched_group_phys
[group
];
6204 #ifdef CONFIG_SCHED_MC
6205 if (!sched_group_core
) {
6207 = kmalloc(sizeof(struct sched_group
) * NR_CPUS
,
6209 if (!sched_group_core
) {
6210 printk (KERN_WARNING
"Can not alloc core sched"
6214 sched_group_core_bycpu
[i
] = sched_group_core
;
6218 sd
= &per_cpu(core_domains
, i
);
6219 group
= cpu_to_core_group(i
);
6221 sd
->span
= cpu_coregroup_map(i
);
6222 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6224 sd
->groups
= &sched_group_core
[group
];
6227 #ifdef CONFIG_SCHED_SMT
6229 sd
= &per_cpu(cpu_domains
, i
);
6230 group
= cpu_to_cpu_group(i
);
6231 *sd
= SD_SIBLING_INIT
;
6232 sd
->span
= cpu_sibling_map
[i
];
6233 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6235 sd
->groups
= &sched_group_cpus
[group
];
6239 #ifdef CONFIG_SCHED_SMT
6240 /* Set up CPU (sibling) groups */
6241 for_each_cpu_mask(i
, *cpu_map
) {
6242 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
6243 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6244 if (i
!= first_cpu(this_sibling_map
))
6247 init_sched_build_groups(sched_group_cpus
, this_sibling_map
,
6252 #ifdef CONFIG_SCHED_MC
6253 /* Set up multi-core groups */
6254 for_each_cpu_mask(i
, *cpu_map
) {
6255 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6256 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6257 if (i
!= first_cpu(this_core_map
))
6259 init_sched_build_groups(sched_group_core
, this_core_map
,
6260 &cpu_to_core_group
);
6265 /* Set up physical groups */
6266 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6267 cpumask_t nodemask
= node_to_cpumask(i
);
6269 cpus_and(nodemask
, nodemask
, *cpu_map
);
6270 if (cpus_empty(nodemask
))
6273 init_sched_build_groups(sched_group_phys
, nodemask
,
6274 &cpu_to_phys_group
);
6278 /* Set up node groups */
6279 if (sched_group_allnodes
)
6280 init_sched_build_groups(sched_group_allnodes
, *cpu_map
,
6281 &cpu_to_allnodes_group
);
6283 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6284 /* Set up node groups */
6285 struct sched_group
*sg
, *prev
;
6286 cpumask_t nodemask
= node_to_cpumask(i
);
6287 cpumask_t domainspan
;
6288 cpumask_t covered
= CPU_MASK_NONE
;
6291 cpus_and(nodemask
, nodemask
, *cpu_map
);
6292 if (cpus_empty(nodemask
)) {
6293 sched_group_nodes
[i
] = NULL
;
6297 domainspan
= sched_domain_node_span(i
);
6298 cpus_and(domainspan
, domainspan
, *cpu_map
);
6300 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6302 printk(KERN_WARNING
"Can not alloc domain group for "
6306 sched_group_nodes
[i
] = sg
;
6307 for_each_cpu_mask(j
, nodemask
) {
6308 struct sched_domain
*sd
;
6309 sd
= &per_cpu(node_domains
, j
);
6313 sg
->cpumask
= nodemask
;
6315 cpus_or(covered
, covered
, nodemask
);
6318 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6319 cpumask_t tmp
, notcovered
;
6320 int n
= (i
+ j
) % MAX_NUMNODES
;
6322 cpus_complement(notcovered
, covered
);
6323 cpus_and(tmp
, notcovered
, *cpu_map
);
6324 cpus_and(tmp
, tmp
, domainspan
);
6325 if (cpus_empty(tmp
))
6328 nodemask
= node_to_cpumask(n
);
6329 cpus_and(tmp
, tmp
, nodemask
);
6330 if (cpus_empty(tmp
))
6333 sg
= kmalloc_node(sizeof(struct sched_group
),
6337 "Can not alloc domain group for node %d\n", j
);
6342 sg
->next
= prev
->next
;
6343 cpus_or(covered
, covered
, tmp
);
6350 /* Calculate CPU power for physical packages and nodes */
6351 #ifdef CONFIG_SCHED_SMT
6352 for_each_cpu_mask(i
, *cpu_map
) {
6353 struct sched_domain
*sd
;
6354 sd
= &per_cpu(cpu_domains
, i
);
6355 sd
->groups
->cpu_power
= SCHED_LOAD_SCALE
;
6358 #ifdef CONFIG_SCHED_MC
6359 for_each_cpu_mask(i
, *cpu_map
) {
6361 struct sched_domain
*sd
;
6362 sd
= &per_cpu(core_domains
, i
);
6363 if (sched_smt_power_savings
)
6364 power
= SCHED_LOAD_SCALE
* cpus_weight(sd
->groups
->cpumask
);
6366 power
= SCHED_LOAD_SCALE
+ (cpus_weight(sd
->groups
->cpumask
)-1)
6367 * SCHED_LOAD_SCALE
/ 10;
6368 sd
->groups
->cpu_power
= power
;
6372 for_each_cpu_mask(i
, *cpu_map
) {
6373 struct sched_domain
*sd
;
6374 #ifdef CONFIG_SCHED_MC
6375 sd
= &per_cpu(phys_domains
, i
);
6376 if (i
!= first_cpu(sd
->groups
->cpumask
))
6379 sd
->groups
->cpu_power
= 0;
6380 if (sched_mc_power_savings
|| sched_smt_power_savings
) {
6383 for_each_cpu_mask(j
, sd
->groups
->cpumask
) {
6384 struct sched_domain
*sd1
;
6385 sd1
= &per_cpu(core_domains
, j
);
6387 * for each core we will add once
6388 * to the group in physical domain
6390 if (j
!= first_cpu(sd1
->groups
->cpumask
))
6393 if (sched_smt_power_savings
)
6394 sd
->groups
->cpu_power
+= sd1
->groups
->cpu_power
;
6396 sd
->groups
->cpu_power
+= SCHED_LOAD_SCALE
;
6400 * This has to be < 2 * SCHED_LOAD_SCALE
6401 * Lets keep it SCHED_LOAD_SCALE, so that
6402 * while calculating NUMA group's cpu_power
6404 * numa_group->cpu_power += phys_group->cpu_power;
6406 * See "only add power once for each physical pkg"
6409 sd
->groups
->cpu_power
= SCHED_LOAD_SCALE
;
6412 sd
= &per_cpu(phys_domains
, i
);
6413 if (sched_smt_power_savings
)
6414 power
= SCHED_LOAD_SCALE
* cpus_weight(sd
->groups
->cpumask
);
6416 power
= SCHED_LOAD_SCALE
;
6417 sd
->groups
->cpu_power
= power
;
6422 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6423 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6425 init_numa_sched_groups_power(sched_group_allnodes
);
6428 /* Attach the domains */
6429 for_each_cpu_mask(i
, *cpu_map
) {
6430 struct sched_domain
*sd
;
6431 #ifdef CONFIG_SCHED_SMT
6432 sd
= &per_cpu(cpu_domains
, i
);
6433 #elif defined(CONFIG_SCHED_MC)
6434 sd
= &per_cpu(core_domains
, i
);
6436 sd
= &per_cpu(phys_domains
, i
);
6438 cpu_attach_domain(sd
, i
);
6441 * Tune cache-hot values:
6443 calibrate_migration_costs(cpu_map
);
6448 free_sched_groups(cpu_map
);
6452 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6454 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6456 cpumask_t cpu_default_map
;
6460 * Setup mask for cpus without special case scheduling requirements.
6461 * For now this just excludes isolated cpus, but could be used to
6462 * exclude other special cases in the future.
6464 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
6466 err
= build_sched_domains(&cpu_default_map
);
6471 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6473 free_sched_groups(cpu_map
);
6477 * Detach sched domains from a group of cpus specified in cpu_map
6478 * These cpus will now be attached to the NULL domain
6480 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6484 for_each_cpu_mask(i
, *cpu_map
)
6485 cpu_attach_domain(NULL
, i
);
6486 synchronize_sched();
6487 arch_destroy_sched_domains(cpu_map
);
6491 * Partition sched domains as specified by the cpumasks below.
6492 * This attaches all cpus from the cpumasks to the NULL domain,
6493 * waits for a RCU quiescent period, recalculates sched
6494 * domain information and then attaches them back to the
6495 * correct sched domains
6496 * Call with hotplug lock held
6498 int partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
6500 cpumask_t change_map
;
6503 cpus_and(*partition1
, *partition1
, cpu_online_map
);
6504 cpus_and(*partition2
, *partition2
, cpu_online_map
);
6505 cpus_or(change_map
, *partition1
, *partition2
);
6507 /* Detach sched domains from all of the affected cpus */
6508 detach_destroy_domains(&change_map
);
6509 if (!cpus_empty(*partition1
))
6510 err
= build_sched_domains(partition1
);
6511 if (!err
&& !cpus_empty(*partition2
))
6512 err
= build_sched_domains(partition2
);
6517 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6518 int arch_reinit_sched_domains(void)
6523 detach_destroy_domains(&cpu_online_map
);
6524 err
= arch_init_sched_domains(&cpu_online_map
);
6525 unlock_cpu_hotplug();
6530 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6534 if (buf
[0] != '0' && buf
[0] != '1')
6538 sched_smt_power_savings
= (buf
[0] == '1');
6540 sched_mc_power_savings
= (buf
[0] == '1');
6542 ret
= arch_reinit_sched_domains();
6544 return ret
? ret
: count
;
6547 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6550 #ifdef CONFIG_SCHED_SMT
6552 err
= sysfs_create_file(&cls
->kset
.kobj
,
6553 &attr_sched_smt_power_savings
.attr
);
6555 #ifdef CONFIG_SCHED_MC
6556 if (!err
&& mc_capable())
6557 err
= sysfs_create_file(&cls
->kset
.kobj
,
6558 &attr_sched_mc_power_savings
.attr
);
6564 #ifdef CONFIG_SCHED_MC
6565 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6567 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6569 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
, const char *buf
, size_t count
)
6571 return sched_power_savings_store(buf
, count
, 0);
6573 SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6574 sched_mc_power_savings_store
);
6577 #ifdef CONFIG_SCHED_SMT
6578 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6580 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6582 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
, const char *buf
, size_t count
)
6584 return sched_power_savings_store(buf
, count
, 1);
6586 SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6587 sched_smt_power_savings_store
);
6591 #ifdef CONFIG_HOTPLUG_CPU
6593 * Force a reinitialization of the sched domains hierarchy. The domains
6594 * and groups cannot be updated in place without racing with the balancing
6595 * code, so we temporarily attach all running cpus to the NULL domain
6596 * which will prevent rebalancing while the sched domains are recalculated.
6598 static int update_sched_domains(struct notifier_block
*nfb
,
6599 unsigned long action
, void *hcpu
)
6602 case CPU_UP_PREPARE
:
6603 case CPU_DOWN_PREPARE
:
6604 detach_destroy_domains(&cpu_online_map
);
6607 case CPU_UP_CANCELED
:
6608 case CPU_DOWN_FAILED
:
6612 * Fall through and re-initialise the domains.
6619 /* The hotplug lock is already held by cpu_up/cpu_down */
6620 arch_init_sched_domains(&cpu_online_map
);
6626 void __init
sched_init_smp(void)
6629 arch_init_sched_domains(&cpu_online_map
);
6630 unlock_cpu_hotplug();
6631 /* XXX: Theoretical race here - CPU may be hotplugged now */
6632 hotcpu_notifier(update_sched_domains
, 0);
6635 void __init
sched_init_smp(void)
6638 #endif /* CONFIG_SMP */
6640 int in_sched_functions(unsigned long addr
)
6642 /* Linker adds these: start and end of __sched functions */
6643 extern char __sched_text_start
[], __sched_text_end
[];
6644 return in_lock_functions(addr
) ||
6645 (addr
>= (unsigned long)__sched_text_start
6646 && addr
< (unsigned long)__sched_text_end
);
6649 void __init
sched_init(void)
6654 for_each_possible_cpu(i
) {
6655 prio_array_t
*array
;
6658 spin_lock_init(&rq
->lock
);
6660 rq
->active
= rq
->arrays
;
6661 rq
->expired
= rq
->arrays
+ 1;
6662 rq
->best_expired_prio
= MAX_PRIO
;
6666 for (j
= 1; j
< 3; j
++)
6667 rq
->cpu_load
[j
] = 0;
6668 rq
->active_balance
= 0;
6670 rq
->migration_thread
= NULL
;
6671 INIT_LIST_HEAD(&rq
->migration_queue
);
6673 atomic_set(&rq
->nr_iowait
, 0);
6675 for (j
= 0; j
< 2; j
++) {
6676 array
= rq
->arrays
+ j
;
6677 for (k
= 0; k
< MAX_PRIO
; k
++) {
6678 INIT_LIST_HEAD(array
->queue
+ k
);
6679 __clear_bit(k
, array
->bitmap
);
6681 // delimiter for bitsearch
6682 __set_bit(MAX_PRIO
, array
->bitmap
);
6686 set_load_weight(&init_task
);
6688 * The boot idle thread does lazy MMU switching as well:
6690 atomic_inc(&init_mm
.mm_count
);
6691 enter_lazy_tlb(&init_mm
, current
);
6694 * Make us the idle thread. Technically, schedule() should not be
6695 * called from this thread, however somewhere below it might be,
6696 * but because we are the idle thread, we just pick up running again
6697 * when this runqueue becomes "idle".
6699 init_idle(current
, smp_processor_id());
6702 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6703 void __might_sleep(char *file
, int line
)
6705 #if defined(in_atomic)
6706 static unsigned long prev_jiffy
; /* ratelimiting */
6708 if ((in_atomic() || irqs_disabled()) &&
6709 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6710 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6712 prev_jiffy
= jiffies
;
6713 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6714 " context at %s:%d\n", file
, line
);
6715 printk("in_atomic():%d, irqs_disabled():%d\n",
6716 in_atomic(), irqs_disabled());
6721 EXPORT_SYMBOL(__might_sleep
);
6724 #ifdef CONFIG_MAGIC_SYSRQ
6725 void normalize_rt_tasks(void)
6727 struct task_struct
*p
;
6728 prio_array_t
*array
;
6729 unsigned long flags
;
6732 read_lock_irq(&tasklist_lock
);
6733 for_each_process(p
) {
6737 spin_lock_irqsave(&p
->pi_lock
, flags
);
6738 rq
= __task_rq_lock(p
);
6742 deactivate_task(p
, task_rq(p
));
6743 __setscheduler(p
, SCHED_NORMAL
, 0);
6745 __activate_task(p
, task_rq(p
));
6746 resched_task(rq
->curr
);
6749 __task_rq_unlock(rq
);
6750 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6752 read_unlock_irq(&tasklist_lock
);
6755 #endif /* CONFIG_MAGIC_SYSRQ */
6759 * These functions are only useful for the IA64 MCA handling.
6761 * They can only be called when the whole system has been
6762 * stopped - every CPU needs to be quiescent, and no scheduling
6763 * activity can take place. Using them for anything else would
6764 * be a serious bug, and as a result, they aren't even visible
6765 * under any other configuration.
6769 * curr_task - return the current task for a given cpu.
6770 * @cpu: the processor in question.
6772 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6774 task_t
*curr_task(int cpu
)
6776 return cpu_curr(cpu
);
6780 * set_curr_task - set the current task for a given cpu.
6781 * @cpu: the processor in question.
6782 * @p: the task pointer to set.
6784 * Description: This function must only be used when non-maskable interrupts
6785 * are serviced on a separate stack. It allows the architecture to switch the
6786 * notion of the current task on a cpu in a non-blocking manner. This function
6787 * must be called with all CPU's synchronized, and interrupts disabled, the
6788 * and caller must save the original value of the current task (see
6789 * curr_task() above) and restore that value before reenabling interrupts and
6790 * re-starting the system.
6792 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6794 void set_curr_task(int cpu
, task_t
*p
)