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/security.h>
34 #include <linux/notifier.h>
35 #include <linux/profile.h>
36 #include <linux/suspend.h>
37 #include <linux/vmalloc.h>
38 #include <linux/blkdev.h>
39 #include <linux/delay.h>
40 #include <linux/smp.h>
41 #include <linux/threads.h>
42 #include <linux/timer.h>
43 #include <linux/rcupdate.h>
44 #include <linux/cpu.h>
45 #include <linux/cpuset.h>
46 #include <linux/percpu.h>
47 #include <linux/kthread.h>
48 #include <linux/seq_file.h>
49 #include <linux/syscalls.h>
50 #include <linux/times.h>
51 #include <linux/acct.h>
54 #include <asm/unistd.h>
57 * Convert user-nice values [ -20 ... 0 ... 19 ]
58 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
61 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
62 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
63 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
66 * 'User priority' is the nice value converted to something we
67 * can work with better when scaling various scheduler parameters,
68 * it's a [ 0 ... 39 ] range.
70 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
71 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
72 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
75 * Some helpers for converting nanosecond timing to jiffy resolution
77 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
78 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
81 * These are the 'tuning knobs' of the scheduler:
83 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
84 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
85 * Timeslices get refilled after they expire.
87 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
88 #define DEF_TIMESLICE (100 * HZ / 1000)
89 #define ON_RUNQUEUE_WEIGHT 30
90 #define CHILD_PENALTY 95
91 #define PARENT_PENALTY 100
93 #define PRIO_BONUS_RATIO 25
94 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
95 #define INTERACTIVE_DELTA 2
96 #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
97 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
98 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
101 * If a task is 'interactive' then we reinsert it in the active
102 * array after it has expired its current timeslice. (it will not
103 * continue to run immediately, it will still roundrobin with
104 * other interactive tasks.)
106 * This part scales the interactivity limit depending on niceness.
108 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
109 * Here are a few examples of different nice levels:
111 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
112 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
113 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
114 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
115 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
117 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
118 * priority range a task can explore, a value of '1' means the
119 * task is rated interactive.)
121 * Ie. nice +19 tasks can never get 'interactive' enough to be
122 * reinserted into the active array. And only heavily CPU-hog nice -20
123 * tasks will be expired. Default nice 0 tasks are somewhere between,
124 * it takes some effort for them to get interactive, but it's not
128 #define CURRENT_BONUS(p) \
129 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
132 #define GRANULARITY (10 * HZ / 1000 ? : 1)
135 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
136 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
139 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
140 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
143 #define SCALE(v1,v1_max,v2_max) \
144 (v1) * (v2_max) / (v1_max)
147 (SCALE(TASK_NICE(p), 40, MAX_BONUS) + INTERACTIVE_DELTA)
149 #define TASK_INTERACTIVE(p) \
150 ((p)->prio <= (p)->static_prio - DELTA(p))
152 #define INTERACTIVE_SLEEP(p) \
153 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
154 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
156 #define TASK_PREEMPTS_CURR(p, rq) \
157 ((p)->prio < (rq)->curr->prio)
160 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
161 * to time slice values: [800ms ... 100ms ... 5ms]
163 * The higher a thread's priority, the bigger timeslices
164 * it gets during one round of execution. But even the lowest
165 * priority thread gets MIN_TIMESLICE worth of execution time.
168 #define SCALE_PRIO(x, prio) \
169 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO/2), MIN_TIMESLICE)
171 static unsigned int task_timeslice(task_t
*p
)
173 if (p
->static_prio
< NICE_TO_PRIO(0))
174 return SCALE_PRIO(DEF_TIMESLICE
*4, p
->static_prio
);
176 return SCALE_PRIO(DEF_TIMESLICE
, p
->static_prio
);
178 #define task_hot(p, now, sd) ((long long) ((now) - (p)->last_ran) \
179 < (long long) (sd)->cache_hot_time)
182 * These are the runqueue data structures:
185 #define BITMAP_SIZE ((((MAX_PRIO+1+7)/8)+sizeof(long)-1)/sizeof(long))
187 typedef struct runqueue runqueue_t
;
190 unsigned int nr_active
;
191 unsigned long bitmap
[BITMAP_SIZE
];
192 struct list_head queue
[MAX_PRIO
];
196 * This is the main, per-CPU runqueue data structure.
198 * Locking rule: those places that want to lock multiple runqueues
199 * (such as the load balancing or the thread migration code), lock
200 * acquire operations must be ordered by ascending &runqueue.
206 * nr_running and cpu_load should be in the same cacheline because
207 * remote CPUs use both these fields when doing load calculation.
209 unsigned long nr_running
;
211 unsigned long cpu_load
[3];
213 unsigned long long nr_switches
;
216 * This is part of a global counter where only the total sum
217 * over all CPUs matters. A task can increase this counter on
218 * one CPU and if it got migrated afterwards it may decrease
219 * it on another CPU. Always updated under the runqueue lock:
221 unsigned long nr_uninterruptible
;
223 unsigned long expired_timestamp
;
224 unsigned long long timestamp_last_tick
;
226 struct mm_struct
*prev_mm
;
227 prio_array_t
*active
, *expired
, arrays
[2];
228 int best_expired_prio
;
232 struct sched_domain
*sd
;
234 /* For active balancing */
238 task_t
*migration_thread
;
239 struct list_head migration_queue
;
242 #ifdef CONFIG_SCHEDSTATS
244 struct sched_info rq_sched_info
;
246 /* sys_sched_yield() stats */
247 unsigned long yld_exp_empty
;
248 unsigned long yld_act_empty
;
249 unsigned long yld_both_empty
;
250 unsigned long yld_cnt
;
252 /* schedule() stats */
253 unsigned long sched_switch
;
254 unsigned long sched_cnt
;
255 unsigned long sched_goidle
;
257 /* try_to_wake_up() stats */
258 unsigned long ttwu_cnt
;
259 unsigned long ttwu_local
;
263 static DEFINE_PER_CPU(struct runqueue
, runqueues
);
266 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
267 * See detach_destroy_domains: synchronize_sched for details.
269 * The domain tree of any CPU may only be accessed from within
270 * preempt-disabled sections.
272 #define for_each_domain(cpu, domain) \
273 for (domain = rcu_dereference(cpu_rq(cpu)->sd); domain; domain = domain->parent)
275 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
276 #define this_rq() (&__get_cpu_var(runqueues))
277 #define task_rq(p) cpu_rq(task_cpu(p))
278 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
280 #ifndef prepare_arch_switch
281 # define prepare_arch_switch(next) do { } while (0)
283 #ifndef finish_arch_switch
284 # define finish_arch_switch(prev) do { } while (0)
287 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
288 static inline int task_running(runqueue_t
*rq
, task_t
*p
)
290 return rq
->curr
== p
;
293 static inline void prepare_lock_switch(runqueue_t
*rq
, task_t
*next
)
297 static inline void finish_lock_switch(runqueue_t
*rq
, task_t
*prev
)
299 #ifdef CONFIG_DEBUG_SPINLOCK
300 /* this is a valid case when another task releases the spinlock */
301 rq
->lock
.owner
= current
;
303 spin_unlock_irq(&rq
->lock
);
306 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
307 static inline int task_running(runqueue_t
*rq
, task_t
*p
)
312 return rq
->curr
== p
;
316 static inline void prepare_lock_switch(runqueue_t
*rq
, task_t
*next
)
320 * We can optimise this out completely for !SMP, because the
321 * SMP rebalancing from interrupt is the only thing that cares
326 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
327 spin_unlock_irq(&rq
->lock
);
329 spin_unlock(&rq
->lock
);
333 static inline void finish_lock_switch(runqueue_t
*rq
, task_t
*prev
)
337 * After ->oncpu is cleared, the task can be moved to a different CPU.
338 * We must ensure this doesn't happen until the switch is completely
344 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
348 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
351 * task_rq_lock - lock the runqueue a given task resides on and disable
352 * interrupts. Note the ordering: we can safely lookup the task_rq without
353 * explicitly disabling preemption.
355 static inline runqueue_t
*task_rq_lock(task_t
*p
, unsigned long *flags
)
361 local_irq_save(*flags
);
363 spin_lock(&rq
->lock
);
364 if (unlikely(rq
!= task_rq(p
))) {
365 spin_unlock_irqrestore(&rq
->lock
, *flags
);
366 goto repeat_lock_task
;
371 static inline void task_rq_unlock(runqueue_t
*rq
, unsigned long *flags
)
374 spin_unlock_irqrestore(&rq
->lock
, *flags
);
377 #ifdef CONFIG_SCHEDSTATS
379 * bump this up when changing the output format or the meaning of an existing
380 * format, so that tools can adapt (or abort)
382 #define SCHEDSTAT_VERSION 12
384 static int show_schedstat(struct seq_file
*seq
, void *v
)
388 seq_printf(seq
, "version %d\n", SCHEDSTAT_VERSION
);
389 seq_printf(seq
, "timestamp %lu\n", jiffies
);
390 for_each_online_cpu(cpu
) {
391 runqueue_t
*rq
= cpu_rq(cpu
);
393 struct sched_domain
*sd
;
397 /* runqueue-specific stats */
399 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
400 cpu
, rq
->yld_both_empty
,
401 rq
->yld_act_empty
, rq
->yld_exp_empty
, rq
->yld_cnt
,
402 rq
->sched_switch
, rq
->sched_cnt
, rq
->sched_goidle
,
403 rq
->ttwu_cnt
, rq
->ttwu_local
,
404 rq
->rq_sched_info
.cpu_time
,
405 rq
->rq_sched_info
.run_delay
, rq
->rq_sched_info
.pcnt
);
407 seq_printf(seq
, "\n");
410 /* domain-specific stats */
412 for_each_domain(cpu
, sd
) {
413 enum idle_type itype
;
414 char mask_str
[NR_CPUS
];
416 cpumask_scnprintf(mask_str
, NR_CPUS
, sd
->span
);
417 seq_printf(seq
, "domain%d %s", dcnt
++, mask_str
);
418 for (itype
= SCHED_IDLE
; itype
< MAX_IDLE_TYPES
;
420 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu %lu",
422 sd
->lb_balanced
[itype
],
423 sd
->lb_failed
[itype
],
424 sd
->lb_imbalance
[itype
],
425 sd
->lb_gained
[itype
],
426 sd
->lb_hot_gained
[itype
],
427 sd
->lb_nobusyq
[itype
],
428 sd
->lb_nobusyg
[itype
]);
430 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
431 sd
->alb_cnt
, sd
->alb_failed
, sd
->alb_pushed
,
432 sd
->sbe_cnt
, sd
->sbe_balanced
, sd
->sbe_pushed
,
433 sd
->sbf_cnt
, sd
->sbf_balanced
, sd
->sbf_pushed
,
434 sd
->ttwu_wake_remote
, sd
->ttwu_move_affine
, sd
->ttwu_move_balance
);
442 static int schedstat_open(struct inode
*inode
, struct file
*file
)
444 unsigned int size
= PAGE_SIZE
* (1 + num_online_cpus() / 32);
445 char *buf
= kmalloc(size
, GFP_KERNEL
);
451 res
= single_open(file
, show_schedstat
, NULL
);
453 m
= file
->private_data
;
461 struct file_operations proc_schedstat_operations
= {
462 .open
= schedstat_open
,
465 .release
= single_release
,
468 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
469 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
470 #else /* !CONFIG_SCHEDSTATS */
471 # define schedstat_inc(rq, field) do { } while (0)
472 # define schedstat_add(rq, field, amt) do { } while (0)
476 * rq_lock - lock a given runqueue and disable interrupts.
478 static inline runqueue_t
*this_rq_lock(void)
485 spin_lock(&rq
->lock
);
490 #ifdef CONFIG_SCHEDSTATS
492 * Called when a process is dequeued from the active array and given
493 * the cpu. We should note that with the exception of interactive
494 * tasks, the expired queue will become the active queue after the active
495 * queue is empty, without explicitly dequeuing and requeuing tasks in the
496 * expired queue. (Interactive tasks may be requeued directly to the
497 * active queue, thus delaying tasks in the expired queue from running;
498 * see scheduler_tick()).
500 * This function is only called from sched_info_arrive(), rather than
501 * dequeue_task(). Even though a task may be queued and dequeued multiple
502 * times as it is shuffled about, we're really interested in knowing how
503 * long it was from the *first* time it was queued to the time that it
506 static inline void sched_info_dequeued(task_t
*t
)
508 t
->sched_info
.last_queued
= 0;
512 * Called when a task finally hits the cpu. We can now calculate how
513 * long it was waiting to run. We also note when it began so that we
514 * can keep stats on how long its timeslice is.
516 static void sched_info_arrive(task_t
*t
)
518 unsigned long now
= jiffies
, diff
= 0;
519 struct runqueue
*rq
= task_rq(t
);
521 if (t
->sched_info
.last_queued
)
522 diff
= now
- t
->sched_info
.last_queued
;
523 sched_info_dequeued(t
);
524 t
->sched_info
.run_delay
+= diff
;
525 t
->sched_info
.last_arrival
= now
;
526 t
->sched_info
.pcnt
++;
531 rq
->rq_sched_info
.run_delay
+= diff
;
532 rq
->rq_sched_info
.pcnt
++;
536 * Called when a process is queued into either the active or expired
537 * array. The time is noted and later used to determine how long we
538 * had to wait for us to reach the cpu. Since the expired queue will
539 * become the active queue after active queue is empty, without dequeuing
540 * and requeuing any tasks, we are interested in queuing to either. It
541 * is unusual but not impossible for tasks to be dequeued and immediately
542 * requeued in the same or another array: this can happen in sched_yield(),
543 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
546 * This function is only called from enqueue_task(), but also only updates
547 * the timestamp if it is already not set. It's assumed that
548 * sched_info_dequeued() will clear that stamp when appropriate.
550 static inline void sched_info_queued(task_t
*t
)
552 if (!t
->sched_info
.last_queued
)
553 t
->sched_info
.last_queued
= jiffies
;
557 * Called when a process ceases being the active-running process, either
558 * voluntarily or involuntarily. Now we can calculate how long we ran.
560 static inline void sched_info_depart(task_t
*t
)
562 struct runqueue
*rq
= task_rq(t
);
563 unsigned long diff
= jiffies
- t
->sched_info
.last_arrival
;
565 t
->sched_info
.cpu_time
+= diff
;
568 rq
->rq_sched_info
.cpu_time
+= diff
;
572 * Called when tasks are switched involuntarily due, typically, to expiring
573 * their time slice. (This may also be called when switching to or from
574 * the idle task.) We are only called when prev != next.
576 static inline void sched_info_switch(task_t
*prev
, task_t
*next
)
578 struct runqueue
*rq
= task_rq(prev
);
581 * prev now departs the cpu. It's not interesting to record
582 * stats about how efficient we were at scheduling the idle
585 if (prev
!= rq
->idle
)
586 sched_info_depart(prev
);
588 if (next
!= rq
->idle
)
589 sched_info_arrive(next
);
592 #define sched_info_queued(t) do { } while (0)
593 #define sched_info_switch(t, next) do { } while (0)
594 #endif /* CONFIG_SCHEDSTATS */
597 * Adding/removing a task to/from a priority array:
599 static void dequeue_task(struct task_struct
*p
, prio_array_t
*array
)
602 list_del(&p
->run_list
);
603 if (list_empty(array
->queue
+ p
->prio
))
604 __clear_bit(p
->prio
, array
->bitmap
);
607 static void enqueue_task(struct task_struct
*p
, prio_array_t
*array
)
609 sched_info_queued(p
);
610 list_add_tail(&p
->run_list
, array
->queue
+ p
->prio
);
611 __set_bit(p
->prio
, array
->bitmap
);
617 * Put task to the end of the run list without the overhead of dequeue
618 * followed by enqueue.
620 static void requeue_task(struct task_struct
*p
, prio_array_t
*array
)
622 list_move_tail(&p
->run_list
, array
->queue
+ p
->prio
);
625 static inline void enqueue_task_head(struct task_struct
*p
, prio_array_t
*array
)
627 list_add(&p
->run_list
, array
->queue
+ p
->prio
);
628 __set_bit(p
->prio
, array
->bitmap
);
634 * effective_prio - return the priority that is based on the static
635 * priority but is modified by bonuses/penalties.
637 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
638 * into the -5 ... 0 ... +5 bonus/penalty range.
640 * We use 25% of the full 0...39 priority range so that:
642 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
643 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
645 * Both properties are important to certain workloads.
647 static int effective_prio(task_t
*p
)
654 bonus
= CURRENT_BONUS(p
) - MAX_BONUS
/ 2;
656 prio
= p
->static_prio
- bonus
;
657 if (prio
< MAX_RT_PRIO
)
659 if (prio
> MAX_PRIO
-1)
665 * __activate_task - move a task to the runqueue.
667 static inline void __activate_task(task_t
*p
, runqueue_t
*rq
)
669 enqueue_task(p
, rq
->active
);
674 * __activate_idle_task - move idle task to the _front_ of runqueue.
676 static inline void __activate_idle_task(task_t
*p
, runqueue_t
*rq
)
678 enqueue_task_head(p
, rq
->active
);
682 static int recalc_task_prio(task_t
*p
, unsigned long long now
)
684 /* Caller must always ensure 'now >= p->timestamp' */
685 unsigned long long __sleep_time
= now
- p
->timestamp
;
686 unsigned long sleep_time
;
688 if (unlikely(p
->policy
== SCHED_BATCH
))
691 if (__sleep_time
> NS_MAX_SLEEP_AVG
)
692 sleep_time
= NS_MAX_SLEEP_AVG
;
694 sleep_time
= (unsigned long)__sleep_time
;
697 if (likely(sleep_time
> 0)) {
699 * User tasks that sleep a long time are categorised as
700 * idle and will get just interactive status to stay active &
701 * prevent them suddenly becoming cpu hogs and starving
704 if (p
->mm
&& p
->activated
!= -1 &&
705 sleep_time
> INTERACTIVE_SLEEP(p
)) {
706 p
->sleep_avg
= JIFFIES_TO_NS(MAX_SLEEP_AVG
-
710 * Tasks waking from uninterruptible sleep are
711 * limited in their sleep_avg rise as they
712 * are likely to be waiting on I/O
714 if (p
->activated
== -1 && p
->mm
) {
715 if (p
->sleep_avg
>= INTERACTIVE_SLEEP(p
))
717 else if (p
->sleep_avg
+ sleep_time
>=
718 INTERACTIVE_SLEEP(p
)) {
719 p
->sleep_avg
= INTERACTIVE_SLEEP(p
);
725 * This code gives a bonus to interactive tasks.
727 * The boost works by updating the 'average sleep time'
728 * value here, based on ->timestamp. The more time a
729 * task spends sleeping, the higher the average gets -
730 * and the higher the priority boost gets as well.
732 p
->sleep_avg
+= sleep_time
;
734 if (p
->sleep_avg
> NS_MAX_SLEEP_AVG
)
735 p
->sleep_avg
= NS_MAX_SLEEP_AVG
;
739 return effective_prio(p
);
743 * activate_task - move a task to the runqueue and do priority recalculation
745 * Update all the scheduling statistics stuff. (sleep average
746 * calculation, priority modifiers, etc.)
748 static void activate_task(task_t
*p
, runqueue_t
*rq
, int local
)
750 unsigned long long now
;
755 /* Compensate for drifting sched_clock */
756 runqueue_t
*this_rq
= this_rq();
757 now
= (now
- this_rq
->timestamp_last_tick
)
758 + rq
->timestamp_last_tick
;
763 p
->prio
= recalc_task_prio(p
, now
);
766 * This checks to make sure it's not an uninterruptible task
767 * that is now waking up.
771 * Tasks which were woken up by interrupts (ie. hw events)
772 * are most likely of interactive nature. So we give them
773 * the credit of extending their sleep time to the period
774 * of time they spend on the runqueue, waiting for execution
775 * on a CPU, first time around:
781 * Normal first-time wakeups get a credit too for
782 * on-runqueue time, but it will be weighted down:
789 __activate_task(p
, rq
);
793 * deactivate_task - remove a task from the runqueue.
795 static void deactivate_task(struct task_struct
*p
, runqueue_t
*rq
)
798 dequeue_task(p
, p
->array
);
803 * resched_task - mark a task 'to be rescheduled now'.
805 * On UP this means the setting of the need_resched flag, on SMP it
806 * might also involve a cross-CPU call to trigger the scheduler on
810 static void resched_task(task_t
*p
)
814 assert_spin_locked(&task_rq(p
)->lock
);
816 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
819 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
822 if (cpu
== smp_processor_id())
825 /* NEED_RESCHED must be visible before we test POLLING_NRFLAG */
827 if (!test_tsk_thread_flag(p
, TIF_POLLING_NRFLAG
))
828 smp_send_reschedule(cpu
);
831 static inline void resched_task(task_t
*p
)
833 assert_spin_locked(&task_rq(p
)->lock
);
834 set_tsk_need_resched(p
);
839 * task_curr - is this task currently executing on a CPU?
840 * @p: the task in question.
842 inline int task_curr(const task_t
*p
)
844 return cpu_curr(task_cpu(p
)) == p
;
849 struct list_head list
;
854 struct completion done
;
858 * The task's runqueue lock must be held.
859 * Returns true if you have to wait for migration thread.
861 static int migrate_task(task_t
*p
, int dest_cpu
, migration_req_t
*req
)
863 runqueue_t
*rq
= task_rq(p
);
866 * If the task is not on a runqueue (and not running), then
867 * it is sufficient to simply update the task's cpu field.
869 if (!p
->array
&& !task_running(rq
, p
)) {
870 set_task_cpu(p
, dest_cpu
);
874 init_completion(&req
->done
);
876 req
->dest_cpu
= dest_cpu
;
877 list_add(&req
->list
, &rq
->migration_queue
);
882 * wait_task_inactive - wait for a thread to unschedule.
884 * The caller must ensure that the task *will* unschedule sometime soon,
885 * else this function might spin for a *long* time. This function can't
886 * be called with interrupts off, or it may introduce deadlock with
887 * smp_call_function() if an IPI is sent by the same process we are
888 * waiting to become inactive.
890 void wait_task_inactive(task_t
*p
)
897 rq
= task_rq_lock(p
, &flags
);
898 /* Must be off runqueue entirely, not preempted. */
899 if (unlikely(p
->array
|| task_running(rq
, p
))) {
900 /* If it's preempted, we yield. It could be a while. */
901 preempted
= !task_running(rq
, p
);
902 task_rq_unlock(rq
, &flags
);
908 task_rq_unlock(rq
, &flags
);
912 * kick_process - kick a running thread to enter/exit the kernel
913 * @p: the to-be-kicked thread
915 * Cause a process which is running on another CPU to enter
916 * kernel-mode, without any delay. (to get signals handled.)
918 * NOTE: this function doesnt have to take the runqueue lock,
919 * because all it wants to ensure is that the remote task enters
920 * the kernel. If the IPI races and the task has been migrated
921 * to another CPU then no harm is done and the purpose has been
924 void kick_process(task_t
*p
)
930 if ((cpu
!= smp_processor_id()) && task_curr(p
))
931 smp_send_reschedule(cpu
);
936 * Return a low guess at the load of a migration-source cpu.
938 * We want to under-estimate the load of migration sources, to
939 * balance conservatively.
941 static inline unsigned long source_load(int cpu
, int type
)
943 runqueue_t
*rq
= cpu_rq(cpu
);
944 unsigned long load_now
= rq
->nr_running
* SCHED_LOAD_SCALE
;
948 return min(rq
->cpu_load
[type
-1], load_now
);
952 * Return a high guess at the load of a migration-target cpu
954 static inline unsigned long target_load(int cpu
, int type
)
956 runqueue_t
*rq
= cpu_rq(cpu
);
957 unsigned long load_now
= rq
->nr_running
* SCHED_LOAD_SCALE
;
961 return max(rq
->cpu_load
[type
-1], load_now
);
965 * find_idlest_group finds and returns the least busy CPU group within the
968 static struct sched_group
*
969 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
971 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
972 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
973 int load_idx
= sd
->forkexec_idx
;
974 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
977 unsigned long load
, avg_load
;
981 /* Skip over this group if it has no CPUs allowed */
982 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
985 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
987 /* Tally up the load of all CPUs in the group */
990 for_each_cpu_mask(i
, group
->cpumask
) {
991 /* Bias balancing toward cpus of our domain */
993 load
= source_load(i
, load_idx
);
995 load
= target_load(i
, load_idx
);
1000 /* Adjust by relative CPU power of the group */
1001 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
1004 this_load
= avg_load
;
1006 } else if (avg_load
< min_load
) {
1007 min_load
= avg_load
;
1011 group
= group
->next
;
1012 } while (group
!= sd
->groups
);
1014 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1020 * find_idlest_queue - find the idlest runqueue among the cpus in group.
1023 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1026 unsigned long load
, min_load
= ULONG_MAX
;
1030 /* Traverse only the allowed CPUs */
1031 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1033 for_each_cpu_mask(i
, tmp
) {
1034 load
= source_load(i
, 0);
1036 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1046 * sched_balance_self: balance the current task (running on cpu) in domains
1047 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1050 * Balance, ie. select the least loaded group.
1052 * Returns the target CPU number, or the same CPU if no balancing is needed.
1054 * preempt must be disabled.
1056 static int sched_balance_self(int cpu
, int flag
)
1058 struct task_struct
*t
= current
;
1059 struct sched_domain
*tmp
, *sd
= NULL
;
1061 for_each_domain(cpu
, tmp
)
1062 if (tmp
->flags
& flag
)
1067 struct sched_group
*group
;
1072 group
= find_idlest_group(sd
, t
, cpu
);
1076 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1077 if (new_cpu
== -1 || new_cpu
== cpu
)
1080 /* Now try balancing at a lower domain level */
1084 weight
= cpus_weight(span
);
1085 for_each_domain(cpu
, tmp
) {
1086 if (weight
<= cpus_weight(tmp
->span
))
1088 if (tmp
->flags
& flag
)
1091 /* while loop will break here if sd == NULL */
1097 #endif /* CONFIG_SMP */
1100 * wake_idle() will wake a task on an idle cpu if task->cpu is
1101 * not idle and an idle cpu is available. The span of cpus to
1102 * search starts with cpus closest then further out as needed,
1103 * so we always favor a closer, idle cpu.
1105 * Returns the CPU we should wake onto.
1107 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1108 static int wake_idle(int cpu
, task_t
*p
)
1111 struct sched_domain
*sd
;
1117 for_each_domain(cpu
, sd
) {
1118 if (sd
->flags
& SD_WAKE_IDLE
) {
1119 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1120 for_each_cpu_mask(i
, tmp
) {
1131 static inline int wake_idle(int cpu
, task_t
*p
)
1138 * try_to_wake_up - wake up a thread
1139 * @p: the to-be-woken-up thread
1140 * @state: the mask of task states that can be woken
1141 * @sync: do a synchronous wakeup?
1143 * Put it on the run-queue if it's not already there. The "current"
1144 * thread is always on the run-queue (except when the actual
1145 * re-schedule is in progress), and as such you're allowed to do
1146 * the simpler "current->state = TASK_RUNNING" to mark yourself
1147 * runnable without the overhead of this.
1149 * returns failure only if the task is already active.
1151 static int try_to_wake_up(task_t
*p
, unsigned int state
, int sync
)
1153 int cpu
, this_cpu
, success
= 0;
1154 unsigned long flags
;
1158 unsigned long load
, this_load
;
1159 struct sched_domain
*sd
, *this_sd
= NULL
;
1163 rq
= task_rq_lock(p
, &flags
);
1164 old_state
= p
->state
;
1165 if (!(old_state
& state
))
1172 this_cpu
= smp_processor_id();
1175 if (unlikely(task_running(rq
, p
)))
1180 schedstat_inc(rq
, ttwu_cnt
);
1181 if (cpu
== this_cpu
) {
1182 schedstat_inc(rq
, ttwu_local
);
1186 for_each_domain(this_cpu
, sd
) {
1187 if (cpu_isset(cpu
, sd
->span
)) {
1188 schedstat_inc(sd
, ttwu_wake_remote
);
1194 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1198 * Check for affine wakeup and passive balancing possibilities.
1201 int idx
= this_sd
->wake_idx
;
1202 unsigned int imbalance
;
1204 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1206 load
= source_load(cpu
, idx
);
1207 this_load
= target_load(this_cpu
, idx
);
1209 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1211 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1212 unsigned long tl
= this_load
;
1214 * If sync wakeup then subtract the (maximum possible)
1215 * effect of the currently running task from the load
1216 * of the current CPU:
1219 tl
-= SCHED_LOAD_SCALE
;
1222 tl
+ target_load(cpu
, idx
) <= SCHED_LOAD_SCALE
) ||
1223 100*(tl
+ SCHED_LOAD_SCALE
) <= imbalance
*load
) {
1225 * This domain has SD_WAKE_AFFINE and
1226 * p is cache cold in this domain, and
1227 * there is no bad imbalance.
1229 schedstat_inc(this_sd
, ttwu_move_affine
);
1235 * Start passive balancing when half the imbalance_pct
1238 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1239 if (imbalance
*this_load
<= 100*load
) {
1240 schedstat_inc(this_sd
, ttwu_move_balance
);
1246 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1248 new_cpu
= wake_idle(new_cpu
, p
);
1249 if (new_cpu
!= cpu
) {
1250 set_task_cpu(p
, new_cpu
);
1251 task_rq_unlock(rq
, &flags
);
1252 /* might preempt at this point */
1253 rq
= task_rq_lock(p
, &flags
);
1254 old_state
= p
->state
;
1255 if (!(old_state
& state
))
1260 this_cpu
= smp_processor_id();
1265 #endif /* CONFIG_SMP */
1266 if (old_state
== TASK_UNINTERRUPTIBLE
) {
1267 rq
->nr_uninterruptible
--;
1269 * Tasks on involuntary sleep don't earn
1270 * sleep_avg beyond just interactive state.
1276 * Tasks that have marked their sleep as noninteractive get
1277 * woken up without updating their sleep average. (i.e. their
1278 * sleep is handled in a priority-neutral manner, no priority
1279 * boost and no penalty.)
1281 if (old_state
& TASK_NONINTERACTIVE
)
1282 __activate_task(p
, rq
);
1284 activate_task(p
, rq
, cpu
== this_cpu
);
1286 * Sync wakeups (i.e. those types of wakeups where the waker
1287 * has indicated that it will leave the CPU in short order)
1288 * don't trigger a preemption, if the woken up task will run on
1289 * this cpu. (in this case the 'I will reschedule' promise of
1290 * the waker guarantees that the freshly woken up task is going
1291 * to be considered on this CPU.)
1293 if (!sync
|| cpu
!= this_cpu
) {
1294 if (TASK_PREEMPTS_CURR(p
, rq
))
1295 resched_task(rq
->curr
);
1300 p
->state
= TASK_RUNNING
;
1302 task_rq_unlock(rq
, &flags
);
1307 int fastcall
wake_up_process(task_t
*p
)
1309 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1310 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1313 EXPORT_SYMBOL(wake_up_process
);
1315 int fastcall
wake_up_state(task_t
*p
, unsigned int state
)
1317 return try_to_wake_up(p
, state
, 0);
1321 * Perform scheduler related setup for a newly forked process p.
1322 * p is forked by current.
1324 void fastcall
sched_fork(task_t
*p
, int clone_flags
)
1326 int cpu
= get_cpu();
1329 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1331 set_task_cpu(p
, cpu
);
1334 * We mark the process as running here, but have not actually
1335 * inserted it onto the runqueue yet. This guarantees that
1336 * nobody will actually run it, and a signal or other external
1337 * event cannot wake it up and insert it on the runqueue either.
1339 p
->state
= TASK_RUNNING
;
1340 INIT_LIST_HEAD(&p
->run_list
);
1342 #ifdef CONFIG_SCHEDSTATS
1343 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1345 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1348 #ifdef CONFIG_PREEMPT
1349 /* Want to start with kernel preemption disabled. */
1350 task_thread_info(p
)->preempt_count
= 1;
1353 * Share the timeslice between parent and child, thus the
1354 * total amount of pending timeslices in the system doesn't change,
1355 * resulting in more scheduling fairness.
1357 local_irq_disable();
1358 p
->time_slice
= (current
->time_slice
+ 1) >> 1;
1360 * The remainder of the first timeslice might be recovered by
1361 * the parent if the child exits early enough.
1363 p
->first_time_slice
= 1;
1364 current
->time_slice
>>= 1;
1365 p
->timestamp
= sched_clock();
1366 if (unlikely(!current
->time_slice
)) {
1368 * This case is rare, it happens when the parent has only
1369 * a single jiffy left from its timeslice. Taking the
1370 * runqueue lock is not a problem.
1372 current
->time_slice
= 1;
1380 * wake_up_new_task - wake up a newly created task for the first time.
1382 * This function will do some initial scheduler statistics housekeeping
1383 * that must be done for every newly created context, then puts the task
1384 * on the runqueue and wakes it.
1386 void fastcall
wake_up_new_task(task_t
*p
, unsigned long clone_flags
)
1388 unsigned long flags
;
1390 runqueue_t
*rq
, *this_rq
;
1392 rq
= task_rq_lock(p
, &flags
);
1393 BUG_ON(p
->state
!= TASK_RUNNING
);
1394 this_cpu
= smp_processor_id();
1398 * We decrease the sleep average of forking parents
1399 * and children as well, to keep max-interactive tasks
1400 * from forking tasks that are max-interactive. The parent
1401 * (current) is done further down, under its lock.
1403 p
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(p
) *
1404 CHILD_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1406 p
->prio
= effective_prio(p
);
1408 if (likely(cpu
== this_cpu
)) {
1409 if (!(clone_flags
& CLONE_VM
)) {
1411 * The VM isn't cloned, so we're in a good position to
1412 * do child-runs-first in anticipation of an exec. This
1413 * usually avoids a lot of COW overhead.
1415 if (unlikely(!current
->array
))
1416 __activate_task(p
, rq
);
1418 p
->prio
= current
->prio
;
1419 list_add_tail(&p
->run_list
, ¤t
->run_list
);
1420 p
->array
= current
->array
;
1421 p
->array
->nr_active
++;
1426 /* Run child last */
1427 __activate_task(p
, rq
);
1429 * We skip the following code due to cpu == this_cpu
1431 * task_rq_unlock(rq, &flags);
1432 * this_rq = task_rq_lock(current, &flags);
1436 this_rq
= cpu_rq(this_cpu
);
1439 * Not the local CPU - must adjust timestamp. This should
1440 * get optimised away in the !CONFIG_SMP case.
1442 p
->timestamp
= (p
->timestamp
- this_rq
->timestamp_last_tick
)
1443 + rq
->timestamp_last_tick
;
1444 __activate_task(p
, rq
);
1445 if (TASK_PREEMPTS_CURR(p
, rq
))
1446 resched_task(rq
->curr
);
1449 * Parent and child are on different CPUs, now get the
1450 * parent runqueue to update the parent's ->sleep_avg:
1452 task_rq_unlock(rq
, &flags
);
1453 this_rq
= task_rq_lock(current
, &flags
);
1455 current
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(current
) *
1456 PARENT_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1457 task_rq_unlock(this_rq
, &flags
);
1461 * Potentially available exiting-child timeslices are
1462 * retrieved here - this way the parent does not get
1463 * penalized for creating too many threads.
1465 * (this cannot be used to 'generate' timeslices
1466 * artificially, because any timeslice recovered here
1467 * was given away by the parent in the first place.)
1469 void fastcall
sched_exit(task_t
*p
)
1471 unsigned long flags
;
1475 * If the child was a (relative-) CPU hog then decrease
1476 * the sleep_avg of the parent as well.
1478 rq
= task_rq_lock(p
->parent
, &flags
);
1479 if (p
->first_time_slice
&& task_cpu(p
) == task_cpu(p
->parent
)) {
1480 p
->parent
->time_slice
+= p
->time_slice
;
1481 if (unlikely(p
->parent
->time_slice
> task_timeslice(p
)))
1482 p
->parent
->time_slice
= task_timeslice(p
);
1484 if (p
->sleep_avg
< p
->parent
->sleep_avg
)
1485 p
->parent
->sleep_avg
= p
->parent
->sleep_avg
/
1486 (EXIT_WEIGHT
+ 1) * EXIT_WEIGHT
+ p
->sleep_avg
/
1488 task_rq_unlock(rq
, &flags
);
1492 * prepare_task_switch - prepare to switch tasks
1493 * @rq: the runqueue preparing to switch
1494 * @next: the task we are going to switch to.
1496 * This is called with the rq lock held and interrupts off. It must
1497 * be paired with a subsequent finish_task_switch after the context
1500 * prepare_task_switch sets up locking and calls architecture specific
1503 static inline void prepare_task_switch(runqueue_t
*rq
, task_t
*next
)
1505 prepare_lock_switch(rq
, next
);
1506 prepare_arch_switch(next
);
1510 * finish_task_switch - clean up after a task-switch
1511 * @rq: runqueue associated with task-switch
1512 * @prev: the thread we just switched away from.
1514 * finish_task_switch must be called after the context switch, paired
1515 * with a prepare_task_switch call before the context switch.
1516 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1517 * and do any other architecture-specific cleanup actions.
1519 * Note that we may have delayed dropping an mm in context_switch(). If
1520 * so, we finish that here outside of the runqueue lock. (Doing it
1521 * with the lock held can cause deadlocks; see schedule() for
1524 static inline void finish_task_switch(runqueue_t
*rq
, task_t
*prev
)
1525 __releases(rq
->lock
)
1527 struct mm_struct
*mm
= rq
->prev_mm
;
1528 unsigned long prev_task_flags
;
1533 * A task struct has one reference for the use as "current".
1534 * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1535 * calls schedule one last time. The schedule call will never return,
1536 * and the scheduled task must drop that reference.
1537 * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1538 * still held, otherwise prev could be scheduled on another cpu, die
1539 * there before we look at prev->state, and then the reference would
1541 * Manfred Spraul <manfred@colorfullife.com>
1543 prev_task_flags
= prev
->flags
;
1544 finish_arch_switch(prev
);
1545 finish_lock_switch(rq
, prev
);
1548 if (unlikely(prev_task_flags
& PF_DEAD
))
1549 put_task_struct(prev
);
1553 * schedule_tail - first thing a freshly forked thread must call.
1554 * @prev: the thread we just switched away from.
1556 asmlinkage
void schedule_tail(task_t
*prev
)
1557 __releases(rq
->lock
)
1559 runqueue_t
*rq
= this_rq();
1560 finish_task_switch(rq
, prev
);
1561 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1562 /* In this case, finish_task_switch does not reenable preemption */
1565 if (current
->set_child_tid
)
1566 put_user(current
->pid
, current
->set_child_tid
);
1570 * context_switch - switch to the new MM and the new
1571 * thread's register state.
1574 task_t
* context_switch(runqueue_t
*rq
, task_t
*prev
, task_t
*next
)
1576 struct mm_struct
*mm
= next
->mm
;
1577 struct mm_struct
*oldmm
= prev
->active_mm
;
1579 if (unlikely(!mm
)) {
1580 next
->active_mm
= oldmm
;
1581 atomic_inc(&oldmm
->mm_count
);
1582 enter_lazy_tlb(oldmm
, next
);
1584 switch_mm(oldmm
, mm
, next
);
1586 if (unlikely(!prev
->mm
)) {
1587 prev
->active_mm
= NULL
;
1588 WARN_ON(rq
->prev_mm
);
1589 rq
->prev_mm
= oldmm
;
1592 /* Here we just switch the register state and the stack. */
1593 switch_to(prev
, next
, prev
);
1599 * nr_running, nr_uninterruptible and nr_context_switches:
1601 * externally visible scheduler statistics: current number of runnable
1602 * threads, current number of uninterruptible-sleeping threads, total
1603 * number of context switches performed since bootup.
1605 unsigned long nr_running(void)
1607 unsigned long i
, sum
= 0;
1609 for_each_online_cpu(i
)
1610 sum
+= cpu_rq(i
)->nr_running
;
1615 unsigned long nr_uninterruptible(void)
1617 unsigned long i
, sum
= 0;
1620 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1623 * Since we read the counters lockless, it might be slightly
1624 * inaccurate. Do not allow it to go below zero though:
1626 if (unlikely((long)sum
< 0))
1632 unsigned long long nr_context_switches(void)
1634 unsigned long long i
, sum
= 0;
1637 sum
+= cpu_rq(i
)->nr_switches
;
1642 unsigned long nr_iowait(void)
1644 unsigned long i
, sum
= 0;
1647 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1655 * double_rq_lock - safely lock two runqueues
1657 * Note this does not disable interrupts like task_rq_lock,
1658 * you need to do so manually before calling.
1660 static void double_rq_lock(runqueue_t
*rq1
, runqueue_t
*rq2
)
1661 __acquires(rq1
->lock
)
1662 __acquires(rq2
->lock
)
1665 spin_lock(&rq1
->lock
);
1666 __acquire(rq2
->lock
); /* Fake it out ;) */
1669 spin_lock(&rq1
->lock
);
1670 spin_lock(&rq2
->lock
);
1672 spin_lock(&rq2
->lock
);
1673 spin_lock(&rq1
->lock
);
1679 * double_rq_unlock - safely unlock two runqueues
1681 * Note this does not restore interrupts like task_rq_unlock,
1682 * you need to do so manually after calling.
1684 static void double_rq_unlock(runqueue_t
*rq1
, runqueue_t
*rq2
)
1685 __releases(rq1
->lock
)
1686 __releases(rq2
->lock
)
1688 spin_unlock(&rq1
->lock
);
1690 spin_unlock(&rq2
->lock
);
1692 __release(rq2
->lock
);
1696 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1698 static void double_lock_balance(runqueue_t
*this_rq
, runqueue_t
*busiest
)
1699 __releases(this_rq
->lock
)
1700 __acquires(busiest
->lock
)
1701 __acquires(this_rq
->lock
)
1703 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1704 if (busiest
< this_rq
) {
1705 spin_unlock(&this_rq
->lock
);
1706 spin_lock(&busiest
->lock
);
1707 spin_lock(&this_rq
->lock
);
1709 spin_lock(&busiest
->lock
);
1714 * If dest_cpu is allowed for this process, migrate the task to it.
1715 * This is accomplished by forcing the cpu_allowed mask to only
1716 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
1717 * the cpu_allowed mask is restored.
1719 static void sched_migrate_task(task_t
*p
, int dest_cpu
)
1721 migration_req_t req
;
1723 unsigned long flags
;
1725 rq
= task_rq_lock(p
, &flags
);
1726 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
1727 || unlikely(cpu_is_offline(dest_cpu
)))
1730 /* force the process onto the specified CPU */
1731 if (migrate_task(p
, dest_cpu
, &req
)) {
1732 /* Need to wait for migration thread (might exit: take ref). */
1733 struct task_struct
*mt
= rq
->migration_thread
;
1734 get_task_struct(mt
);
1735 task_rq_unlock(rq
, &flags
);
1736 wake_up_process(mt
);
1737 put_task_struct(mt
);
1738 wait_for_completion(&req
.done
);
1742 task_rq_unlock(rq
, &flags
);
1746 * sched_exec - execve() is a valuable balancing opportunity, because at
1747 * this point the task has the smallest effective memory and cache footprint.
1749 void sched_exec(void)
1751 int new_cpu
, this_cpu
= get_cpu();
1752 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
1754 if (new_cpu
!= this_cpu
)
1755 sched_migrate_task(current
, new_cpu
);
1759 * pull_task - move a task from a remote runqueue to the local runqueue.
1760 * Both runqueues must be locked.
1763 void pull_task(runqueue_t
*src_rq
, prio_array_t
*src_array
, task_t
*p
,
1764 runqueue_t
*this_rq
, prio_array_t
*this_array
, int this_cpu
)
1766 dequeue_task(p
, src_array
);
1767 src_rq
->nr_running
--;
1768 set_task_cpu(p
, this_cpu
);
1769 this_rq
->nr_running
++;
1770 enqueue_task(p
, this_array
);
1771 p
->timestamp
= (p
->timestamp
- src_rq
->timestamp_last_tick
)
1772 + this_rq
->timestamp_last_tick
;
1774 * Note that idle threads have a prio of MAX_PRIO, for this test
1775 * to be always true for them.
1777 if (TASK_PREEMPTS_CURR(p
, this_rq
))
1778 resched_task(this_rq
->curr
);
1782 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1785 int can_migrate_task(task_t
*p
, runqueue_t
*rq
, int this_cpu
,
1786 struct sched_domain
*sd
, enum idle_type idle
,
1790 * We do not migrate tasks that are:
1791 * 1) running (obviously), or
1792 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1793 * 3) are cache-hot on their current CPU.
1795 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
1799 if (task_running(rq
, p
))
1803 * Aggressive migration if:
1804 * 1) task is cache cold, or
1805 * 2) too many balance attempts have failed.
1808 if (sd
->nr_balance_failed
> sd
->cache_nice_tries
)
1811 if (task_hot(p
, rq
->timestamp_last_tick
, sd
))
1817 * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq,
1818 * as part of a balancing operation within "domain". Returns the number of
1821 * Called with both runqueues locked.
1823 static int move_tasks(runqueue_t
*this_rq
, int this_cpu
, runqueue_t
*busiest
,
1824 unsigned long max_nr_move
, struct sched_domain
*sd
,
1825 enum idle_type idle
, int *all_pinned
)
1827 prio_array_t
*array
, *dst_array
;
1828 struct list_head
*head
, *curr
;
1829 int idx
, pulled
= 0, pinned
= 0;
1832 if (max_nr_move
== 0)
1838 * We first consider expired tasks. Those will likely not be
1839 * executed in the near future, and they are most likely to
1840 * be cache-cold, thus switching CPUs has the least effect
1843 if (busiest
->expired
->nr_active
) {
1844 array
= busiest
->expired
;
1845 dst_array
= this_rq
->expired
;
1847 array
= busiest
->active
;
1848 dst_array
= this_rq
->active
;
1852 /* Start searching at priority 0: */
1856 idx
= sched_find_first_bit(array
->bitmap
);
1858 idx
= find_next_bit(array
->bitmap
, MAX_PRIO
, idx
);
1859 if (idx
>= MAX_PRIO
) {
1860 if (array
== busiest
->expired
&& busiest
->active
->nr_active
) {
1861 array
= busiest
->active
;
1862 dst_array
= this_rq
->active
;
1868 head
= array
->queue
+ idx
;
1871 tmp
= list_entry(curr
, task_t
, run_list
);
1875 if (!can_migrate_task(tmp
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
1882 #ifdef CONFIG_SCHEDSTATS
1883 if (task_hot(tmp
, busiest
->timestamp_last_tick
, sd
))
1884 schedstat_inc(sd
, lb_hot_gained
[idle
]);
1887 pull_task(busiest
, array
, tmp
, this_rq
, dst_array
, this_cpu
);
1890 /* We only want to steal up to the prescribed number of tasks. */
1891 if (pulled
< max_nr_move
) {
1899 * Right now, this is the only place pull_task() is called,
1900 * so we can safely collect pull_task() stats here rather than
1901 * inside pull_task().
1903 schedstat_add(sd
, lb_gained
[idle
], pulled
);
1906 *all_pinned
= pinned
;
1911 * find_busiest_group finds and returns the busiest CPU group within the
1912 * domain. It calculates and returns the number of tasks which should be
1913 * moved to restore balance via the imbalance parameter.
1915 static struct sched_group
*
1916 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
1917 unsigned long *imbalance
, enum idle_type idle
, int *sd_idle
)
1919 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1920 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
1921 unsigned long max_pull
;
1924 max_load
= this_load
= total_load
= total_pwr
= 0;
1925 if (idle
== NOT_IDLE
)
1926 load_idx
= sd
->busy_idx
;
1927 else if (idle
== NEWLY_IDLE
)
1928 load_idx
= sd
->newidle_idx
;
1930 load_idx
= sd
->idle_idx
;
1937 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1939 /* Tally up the load of all CPUs in the group */
1942 for_each_cpu_mask(i
, group
->cpumask
) {
1943 if (*sd_idle
&& !idle_cpu(i
))
1946 /* Bias balancing toward cpus of our domain */
1948 load
= target_load(i
, load_idx
);
1950 load
= source_load(i
, load_idx
);
1955 total_load
+= avg_load
;
1956 total_pwr
+= group
->cpu_power
;
1958 /* Adjust by relative CPU power of the group */
1959 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
1962 this_load
= avg_load
;
1964 } else if (avg_load
> max_load
) {
1965 max_load
= avg_load
;
1968 group
= group
->next
;
1969 } while (group
!= sd
->groups
);
1971 if (!busiest
|| this_load
>= max_load
|| max_load
<= SCHED_LOAD_SCALE
)
1974 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
1976 if (this_load
>= avg_load
||
1977 100*max_load
<= sd
->imbalance_pct
*this_load
)
1981 * We're trying to get all the cpus to the average_load, so we don't
1982 * want to push ourselves above the average load, nor do we wish to
1983 * reduce the max loaded cpu below the average load, as either of these
1984 * actions would just result in more rebalancing later, and ping-pong
1985 * tasks around. Thus we look for the minimum possible imbalance.
1986 * Negative imbalances (*we* are more loaded than anyone else) will
1987 * be counted as no imbalance for these purposes -- we can't fix that
1988 * by pulling tasks to us. Be careful of negative numbers as they'll
1989 * appear as very large values with unsigned longs.
1992 /* Don't want to pull so many tasks that a group would go idle */
1993 max_pull
= min(max_load
- avg_load
, max_load
- SCHED_LOAD_SCALE
);
1995 /* How much load to actually move to equalise the imbalance */
1996 *imbalance
= min(max_pull
* busiest
->cpu_power
,
1997 (avg_load
- this_load
) * this->cpu_power
)
2000 if (*imbalance
< SCHED_LOAD_SCALE
) {
2001 unsigned long pwr_now
= 0, pwr_move
= 0;
2004 if (max_load
- this_load
>= SCHED_LOAD_SCALE
*2) {
2010 * OK, we don't have enough imbalance to justify moving tasks,
2011 * however we may be able to increase total CPU power used by
2015 pwr_now
+= busiest
->cpu_power
*min(SCHED_LOAD_SCALE
, max_load
);
2016 pwr_now
+= this->cpu_power
*min(SCHED_LOAD_SCALE
, this_load
);
2017 pwr_now
/= SCHED_LOAD_SCALE
;
2019 /* Amount of load we'd subtract */
2020 tmp
= SCHED_LOAD_SCALE
*SCHED_LOAD_SCALE
/busiest
->cpu_power
;
2022 pwr_move
+= busiest
->cpu_power
*min(SCHED_LOAD_SCALE
,
2025 /* Amount of load we'd add */
2026 if (max_load
*busiest
->cpu_power
<
2027 SCHED_LOAD_SCALE
*SCHED_LOAD_SCALE
)
2028 tmp
= max_load
*busiest
->cpu_power
/this->cpu_power
;
2030 tmp
= SCHED_LOAD_SCALE
*SCHED_LOAD_SCALE
/this->cpu_power
;
2031 pwr_move
+= this->cpu_power
*min(SCHED_LOAD_SCALE
, this_load
+ tmp
);
2032 pwr_move
/= SCHED_LOAD_SCALE
;
2034 /* Move if we gain throughput */
2035 if (pwr_move
<= pwr_now
)
2042 /* Get rid of the scaling factor, rounding down as we divide */
2043 *imbalance
= *imbalance
/ SCHED_LOAD_SCALE
;
2053 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2055 static runqueue_t
*find_busiest_queue(struct sched_group
*group
,
2056 enum idle_type idle
)
2058 unsigned long load
, max_load
= 0;
2059 runqueue_t
*busiest
= NULL
;
2062 for_each_cpu_mask(i
, group
->cpumask
) {
2063 load
= source_load(i
, 0);
2065 if (load
> max_load
) {
2067 busiest
= cpu_rq(i
);
2075 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2076 * so long as it is large enough.
2078 #define MAX_PINNED_INTERVAL 512
2081 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2082 * tasks if there is an imbalance.
2084 * Called with this_rq unlocked.
2086 static int load_balance(int this_cpu
, runqueue_t
*this_rq
,
2087 struct sched_domain
*sd
, enum idle_type idle
)
2089 struct sched_group
*group
;
2090 runqueue_t
*busiest
;
2091 unsigned long imbalance
;
2092 int nr_moved
, all_pinned
= 0;
2093 int active_balance
= 0;
2096 if (idle
!= NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2099 schedstat_inc(sd
, lb_cnt
[idle
]);
2101 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
);
2103 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2107 busiest
= find_busiest_queue(group
, idle
);
2109 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2113 BUG_ON(busiest
== this_rq
);
2115 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2118 if (busiest
->nr_running
> 1) {
2120 * Attempt to move tasks. If find_busiest_group has found
2121 * an imbalance but busiest->nr_running <= 1, the group is
2122 * still unbalanced. nr_moved simply stays zero, so it is
2123 * correctly treated as an imbalance.
2125 double_rq_lock(this_rq
, busiest
);
2126 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2127 imbalance
, sd
, idle
, &all_pinned
);
2128 double_rq_unlock(this_rq
, busiest
);
2130 /* All tasks on this runqueue were pinned by CPU affinity */
2131 if (unlikely(all_pinned
))
2136 schedstat_inc(sd
, lb_failed
[idle
]);
2137 sd
->nr_balance_failed
++;
2139 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2141 spin_lock(&busiest
->lock
);
2143 /* don't kick the migration_thread, if the curr
2144 * task on busiest cpu can't be moved to this_cpu
2146 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2147 spin_unlock(&busiest
->lock
);
2149 goto out_one_pinned
;
2152 if (!busiest
->active_balance
) {
2153 busiest
->active_balance
= 1;
2154 busiest
->push_cpu
= this_cpu
;
2157 spin_unlock(&busiest
->lock
);
2159 wake_up_process(busiest
->migration_thread
);
2162 * We've kicked active balancing, reset the failure
2165 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2168 sd
->nr_balance_failed
= 0;
2170 if (likely(!active_balance
)) {
2171 /* We were unbalanced, so reset the balancing interval */
2172 sd
->balance_interval
= sd
->min_interval
;
2175 * If we've begun active balancing, start to back off. This
2176 * case may not be covered by the all_pinned logic if there
2177 * is only 1 task on the busy runqueue (because we don't call
2180 if (sd
->balance_interval
< sd
->max_interval
)
2181 sd
->balance_interval
*= 2;
2184 if (!nr_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2189 schedstat_inc(sd
, lb_balanced
[idle
]);
2191 sd
->nr_balance_failed
= 0;
2194 /* tune up the balancing interval */
2195 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2196 (sd
->balance_interval
< sd
->max_interval
))
2197 sd
->balance_interval
*= 2;
2199 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2205 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2206 * tasks if there is an imbalance.
2208 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2209 * this_rq is locked.
2211 static int load_balance_newidle(int this_cpu
, runqueue_t
*this_rq
,
2212 struct sched_domain
*sd
)
2214 struct sched_group
*group
;
2215 runqueue_t
*busiest
= NULL
;
2216 unsigned long imbalance
;
2220 if (sd
->flags
& SD_SHARE_CPUPOWER
)
2223 schedstat_inc(sd
, lb_cnt
[NEWLY_IDLE
]);
2224 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, NEWLY_IDLE
, &sd_idle
);
2226 schedstat_inc(sd
, lb_nobusyg
[NEWLY_IDLE
]);
2230 busiest
= find_busiest_queue(group
, NEWLY_IDLE
);
2232 schedstat_inc(sd
, lb_nobusyq
[NEWLY_IDLE
]);
2236 BUG_ON(busiest
== this_rq
);
2238 schedstat_add(sd
, lb_imbalance
[NEWLY_IDLE
], imbalance
);
2241 if (busiest
->nr_running
> 1) {
2242 /* Attempt to move tasks */
2243 double_lock_balance(this_rq
, busiest
);
2244 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2245 imbalance
, sd
, NEWLY_IDLE
, NULL
);
2246 spin_unlock(&busiest
->lock
);
2250 schedstat_inc(sd
, lb_failed
[NEWLY_IDLE
]);
2251 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2254 sd
->nr_balance_failed
= 0;
2259 schedstat_inc(sd
, lb_balanced
[NEWLY_IDLE
]);
2260 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2262 sd
->nr_balance_failed
= 0;
2267 * idle_balance is called by schedule() if this_cpu is about to become
2268 * idle. Attempts to pull tasks from other CPUs.
2270 static void idle_balance(int this_cpu
, runqueue_t
*this_rq
)
2272 struct sched_domain
*sd
;
2274 for_each_domain(this_cpu
, sd
) {
2275 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
2276 if (load_balance_newidle(this_cpu
, this_rq
, sd
)) {
2277 /* We've pulled tasks over so stop searching */
2285 * active_load_balance is run by migration threads. It pushes running tasks
2286 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2287 * running on each physical CPU where possible, and avoids physical /
2288 * logical imbalances.
2290 * Called with busiest_rq locked.
2292 static void active_load_balance(runqueue_t
*busiest_rq
, int busiest_cpu
)
2294 struct sched_domain
*sd
;
2295 runqueue_t
*target_rq
;
2296 int target_cpu
= busiest_rq
->push_cpu
;
2298 if (busiest_rq
->nr_running
<= 1)
2299 /* no task to move */
2302 target_rq
= cpu_rq(target_cpu
);
2305 * This condition is "impossible", if it occurs
2306 * we need to fix it. Originally reported by
2307 * Bjorn Helgaas on a 128-cpu setup.
2309 BUG_ON(busiest_rq
== target_rq
);
2311 /* move a task from busiest_rq to target_rq */
2312 double_lock_balance(busiest_rq
, target_rq
);
2314 /* Search for an sd spanning us and the target CPU. */
2315 for_each_domain(target_cpu
, sd
)
2316 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2317 cpu_isset(busiest_cpu
, sd
->span
))
2320 if (unlikely(sd
== NULL
))
2323 schedstat_inc(sd
, alb_cnt
);
2325 if (move_tasks(target_rq
, target_cpu
, busiest_rq
, 1, sd
, SCHED_IDLE
, NULL
))
2326 schedstat_inc(sd
, alb_pushed
);
2328 schedstat_inc(sd
, alb_failed
);
2330 spin_unlock(&target_rq
->lock
);
2334 * rebalance_tick will get called every timer tick, on every CPU.
2336 * It checks each scheduling domain to see if it is due to be balanced,
2337 * and initiates a balancing operation if so.
2339 * Balancing parameters are set up in arch_init_sched_domains.
2342 /* Don't have all balancing operations going off at once */
2343 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2345 static void rebalance_tick(int this_cpu
, runqueue_t
*this_rq
,
2346 enum idle_type idle
)
2348 unsigned long old_load
, this_load
;
2349 unsigned long j
= jiffies
+ CPU_OFFSET(this_cpu
);
2350 struct sched_domain
*sd
;
2353 this_load
= this_rq
->nr_running
* SCHED_LOAD_SCALE
;
2354 /* Update our load */
2355 for (i
= 0; i
< 3; i
++) {
2356 unsigned long new_load
= this_load
;
2358 old_load
= this_rq
->cpu_load
[i
];
2360 * Round up the averaging division if load is increasing. This
2361 * prevents us from getting stuck on 9 if the load is 10, for
2364 if (new_load
> old_load
)
2365 new_load
+= scale
-1;
2366 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) / scale
;
2369 for_each_domain(this_cpu
, sd
) {
2370 unsigned long interval
;
2372 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2375 interval
= sd
->balance_interval
;
2376 if (idle
!= SCHED_IDLE
)
2377 interval
*= sd
->busy_factor
;
2379 /* scale ms to jiffies */
2380 interval
= msecs_to_jiffies(interval
);
2381 if (unlikely(!interval
))
2384 if (j
- sd
->last_balance
>= interval
) {
2385 if (load_balance(this_cpu
, this_rq
, sd
, idle
)) {
2387 * We've pulled tasks over so either we're no
2388 * longer idle, or one of our SMT siblings is
2393 sd
->last_balance
+= interval
;
2399 * on UP we do not need to balance between CPUs:
2401 static inline void rebalance_tick(int cpu
, runqueue_t
*rq
, enum idle_type idle
)
2404 static inline void idle_balance(int cpu
, runqueue_t
*rq
)
2409 static inline int wake_priority_sleeper(runqueue_t
*rq
)
2412 #ifdef CONFIG_SCHED_SMT
2413 spin_lock(&rq
->lock
);
2415 * If an SMT sibling task has been put to sleep for priority
2416 * reasons reschedule the idle task to see if it can now run.
2418 if (rq
->nr_running
) {
2419 resched_task(rq
->idle
);
2422 spin_unlock(&rq
->lock
);
2427 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2429 EXPORT_PER_CPU_SYMBOL(kstat
);
2432 * This is called on clock ticks and on context switches.
2433 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2435 static inline void update_cpu_clock(task_t
*p
, runqueue_t
*rq
,
2436 unsigned long long now
)
2438 unsigned long long last
= max(p
->timestamp
, rq
->timestamp_last_tick
);
2439 p
->sched_time
+= now
- last
;
2443 * Return current->sched_time plus any more ns on the sched_clock
2444 * that have not yet been banked.
2446 unsigned long long current_sched_time(const task_t
*tsk
)
2448 unsigned long long ns
;
2449 unsigned long flags
;
2450 local_irq_save(flags
);
2451 ns
= max(tsk
->timestamp
, task_rq(tsk
)->timestamp_last_tick
);
2452 ns
= tsk
->sched_time
+ (sched_clock() - ns
);
2453 local_irq_restore(flags
);
2458 * We place interactive tasks back into the active array, if possible.
2460 * To guarantee that this does not starve expired tasks we ignore the
2461 * interactivity of a task if the first expired task had to wait more
2462 * than a 'reasonable' amount of time. This deadline timeout is
2463 * load-dependent, as the frequency of array switched decreases with
2464 * increasing number of running tasks. We also ignore the interactivity
2465 * if a better static_prio task has expired:
2467 #define EXPIRED_STARVING(rq) \
2468 ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2469 (jiffies - (rq)->expired_timestamp >= \
2470 STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2471 ((rq)->curr->static_prio > (rq)->best_expired_prio))
2474 * Account user cpu time to a process.
2475 * @p: the process that the cpu time gets accounted to
2476 * @hardirq_offset: the offset to subtract from hardirq_count()
2477 * @cputime: the cpu time spent in user space since the last update
2479 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
2481 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2484 p
->utime
= cputime_add(p
->utime
, cputime
);
2486 /* Add user time to cpustat. */
2487 tmp
= cputime_to_cputime64(cputime
);
2488 if (TASK_NICE(p
) > 0)
2489 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
2491 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
2495 * Account system cpu time to a process.
2496 * @p: the process that the cpu time gets accounted to
2497 * @hardirq_offset: the offset to subtract from hardirq_count()
2498 * @cputime: the cpu time spent in kernel space since the last update
2500 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
2503 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2504 runqueue_t
*rq
= this_rq();
2507 p
->stime
= cputime_add(p
->stime
, cputime
);
2509 /* Add system time to cpustat. */
2510 tmp
= cputime_to_cputime64(cputime
);
2511 if (hardirq_count() - hardirq_offset
)
2512 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
2513 else if (softirq_count())
2514 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
2515 else if (p
!= rq
->idle
)
2516 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
2517 else if (atomic_read(&rq
->nr_iowait
) > 0)
2518 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
2520 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
2521 /* Account for system time used */
2522 acct_update_integrals(p
);
2526 * Account for involuntary wait time.
2527 * @p: the process from which the cpu time has been stolen
2528 * @steal: the cpu time spent in involuntary wait
2530 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
2532 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2533 cputime64_t tmp
= cputime_to_cputime64(steal
);
2534 runqueue_t
*rq
= this_rq();
2536 if (p
== rq
->idle
) {
2537 p
->stime
= cputime_add(p
->stime
, steal
);
2538 if (atomic_read(&rq
->nr_iowait
) > 0)
2539 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
2541 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
2543 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
2547 * This function gets called by the timer code, with HZ frequency.
2548 * We call it with interrupts disabled.
2550 * It also gets called by the fork code, when changing the parent's
2553 void scheduler_tick(void)
2555 int cpu
= smp_processor_id();
2556 runqueue_t
*rq
= this_rq();
2557 task_t
*p
= current
;
2558 unsigned long long now
= sched_clock();
2560 update_cpu_clock(p
, rq
, now
);
2562 rq
->timestamp_last_tick
= now
;
2564 if (p
== rq
->idle
) {
2565 if (wake_priority_sleeper(rq
))
2567 rebalance_tick(cpu
, rq
, SCHED_IDLE
);
2571 /* Task might have expired already, but not scheduled off yet */
2572 if (p
->array
!= rq
->active
) {
2573 set_tsk_need_resched(p
);
2576 spin_lock(&rq
->lock
);
2578 * The task was running during this tick - update the
2579 * time slice counter. Note: we do not update a thread's
2580 * priority until it either goes to sleep or uses up its
2581 * timeslice. This makes it possible for interactive tasks
2582 * to use up their timeslices at their highest priority levels.
2586 * RR tasks need a special form of timeslice management.
2587 * FIFO tasks have no timeslices.
2589 if ((p
->policy
== SCHED_RR
) && !--p
->time_slice
) {
2590 p
->time_slice
= task_timeslice(p
);
2591 p
->first_time_slice
= 0;
2592 set_tsk_need_resched(p
);
2594 /* put it at the end of the queue: */
2595 requeue_task(p
, rq
->active
);
2599 if (!--p
->time_slice
) {
2600 dequeue_task(p
, rq
->active
);
2601 set_tsk_need_resched(p
);
2602 p
->prio
= effective_prio(p
);
2603 p
->time_slice
= task_timeslice(p
);
2604 p
->first_time_slice
= 0;
2606 if (!rq
->expired_timestamp
)
2607 rq
->expired_timestamp
= jiffies
;
2608 if (!TASK_INTERACTIVE(p
) || EXPIRED_STARVING(rq
)) {
2609 enqueue_task(p
, rq
->expired
);
2610 if (p
->static_prio
< rq
->best_expired_prio
)
2611 rq
->best_expired_prio
= p
->static_prio
;
2613 enqueue_task(p
, rq
->active
);
2616 * Prevent a too long timeslice allowing a task to monopolize
2617 * the CPU. We do this by splitting up the timeslice into
2620 * Note: this does not mean the task's timeslices expire or
2621 * get lost in any way, they just might be preempted by
2622 * another task of equal priority. (one with higher
2623 * priority would have preempted this task already.) We
2624 * requeue this task to the end of the list on this priority
2625 * level, which is in essence a round-robin of tasks with
2628 * This only applies to tasks in the interactive
2629 * delta range with at least TIMESLICE_GRANULARITY to requeue.
2631 if (TASK_INTERACTIVE(p
) && !((task_timeslice(p
) -
2632 p
->time_slice
) % TIMESLICE_GRANULARITY(p
)) &&
2633 (p
->time_slice
>= TIMESLICE_GRANULARITY(p
)) &&
2634 (p
->array
== rq
->active
)) {
2636 requeue_task(p
, rq
->active
);
2637 set_tsk_need_resched(p
);
2641 spin_unlock(&rq
->lock
);
2643 rebalance_tick(cpu
, rq
, NOT_IDLE
);
2646 #ifdef CONFIG_SCHED_SMT
2647 static inline void wakeup_busy_runqueue(runqueue_t
*rq
)
2649 /* If an SMT runqueue is sleeping due to priority reasons wake it up */
2650 if (rq
->curr
== rq
->idle
&& rq
->nr_running
)
2651 resched_task(rq
->idle
);
2654 static void wake_sleeping_dependent(int this_cpu
, runqueue_t
*this_rq
)
2656 struct sched_domain
*tmp
, *sd
= NULL
;
2657 cpumask_t sibling_map
;
2660 for_each_domain(this_cpu
, tmp
)
2661 if (tmp
->flags
& SD_SHARE_CPUPOWER
)
2668 * Unlock the current runqueue because we have to lock in
2669 * CPU order to avoid deadlocks. Caller knows that we might
2670 * unlock. We keep IRQs disabled.
2672 spin_unlock(&this_rq
->lock
);
2674 sibling_map
= sd
->span
;
2676 for_each_cpu_mask(i
, sibling_map
)
2677 spin_lock(&cpu_rq(i
)->lock
);
2679 * We clear this CPU from the mask. This both simplifies the
2680 * inner loop and keps this_rq locked when we exit:
2682 cpu_clear(this_cpu
, sibling_map
);
2684 for_each_cpu_mask(i
, sibling_map
) {
2685 runqueue_t
*smt_rq
= cpu_rq(i
);
2687 wakeup_busy_runqueue(smt_rq
);
2690 for_each_cpu_mask(i
, sibling_map
)
2691 spin_unlock(&cpu_rq(i
)->lock
);
2693 * We exit with this_cpu's rq still held and IRQs
2699 * number of 'lost' timeslices this task wont be able to fully
2700 * utilize, if another task runs on a sibling. This models the
2701 * slowdown effect of other tasks running on siblings:
2703 static inline unsigned long smt_slice(task_t
*p
, struct sched_domain
*sd
)
2705 return p
->time_slice
* (100 - sd
->per_cpu_gain
) / 100;
2708 static int dependent_sleeper(int this_cpu
, runqueue_t
*this_rq
)
2710 struct sched_domain
*tmp
, *sd
= NULL
;
2711 cpumask_t sibling_map
;
2712 prio_array_t
*array
;
2716 for_each_domain(this_cpu
, tmp
)
2717 if (tmp
->flags
& SD_SHARE_CPUPOWER
)
2724 * The same locking rules and details apply as for
2725 * wake_sleeping_dependent():
2727 spin_unlock(&this_rq
->lock
);
2728 sibling_map
= sd
->span
;
2729 for_each_cpu_mask(i
, sibling_map
)
2730 spin_lock(&cpu_rq(i
)->lock
);
2731 cpu_clear(this_cpu
, sibling_map
);
2734 * Establish next task to be run - it might have gone away because
2735 * we released the runqueue lock above:
2737 if (!this_rq
->nr_running
)
2739 array
= this_rq
->active
;
2740 if (!array
->nr_active
)
2741 array
= this_rq
->expired
;
2742 BUG_ON(!array
->nr_active
);
2744 p
= list_entry(array
->queue
[sched_find_first_bit(array
->bitmap
)].next
,
2747 for_each_cpu_mask(i
, sibling_map
) {
2748 runqueue_t
*smt_rq
= cpu_rq(i
);
2749 task_t
*smt_curr
= smt_rq
->curr
;
2751 /* Kernel threads do not participate in dependent sleeping */
2752 if (!p
->mm
|| !smt_curr
->mm
|| rt_task(p
))
2753 goto check_smt_task
;
2756 * If a user task with lower static priority than the
2757 * running task on the SMT sibling is trying to schedule,
2758 * delay it till there is proportionately less timeslice
2759 * left of the sibling task to prevent a lower priority
2760 * task from using an unfair proportion of the
2761 * physical cpu's resources. -ck
2763 if (rt_task(smt_curr
)) {
2765 * With real time tasks we run non-rt tasks only
2766 * per_cpu_gain% of the time.
2768 if ((jiffies
% DEF_TIMESLICE
) >
2769 (sd
->per_cpu_gain
* DEF_TIMESLICE
/ 100))
2772 if (smt_curr
->static_prio
< p
->static_prio
&&
2773 !TASK_PREEMPTS_CURR(p
, smt_rq
) &&
2774 smt_slice(smt_curr
, sd
) > task_timeslice(p
))
2778 if ((!smt_curr
->mm
&& smt_curr
!= smt_rq
->idle
) ||
2782 wakeup_busy_runqueue(smt_rq
);
2787 * Reschedule a lower priority task on the SMT sibling for
2788 * it to be put to sleep, or wake it up if it has been put to
2789 * sleep for priority reasons to see if it should run now.
2792 if ((jiffies
% DEF_TIMESLICE
) >
2793 (sd
->per_cpu_gain
* DEF_TIMESLICE
/ 100))
2794 resched_task(smt_curr
);
2796 if (TASK_PREEMPTS_CURR(p
, smt_rq
) &&
2797 smt_slice(p
, sd
) > task_timeslice(smt_curr
))
2798 resched_task(smt_curr
);
2800 wakeup_busy_runqueue(smt_rq
);
2804 for_each_cpu_mask(i
, sibling_map
)
2805 spin_unlock(&cpu_rq(i
)->lock
);
2809 static inline void wake_sleeping_dependent(int this_cpu
, runqueue_t
*this_rq
)
2813 static inline int dependent_sleeper(int this_cpu
, runqueue_t
*this_rq
)
2819 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
2821 void fastcall
add_preempt_count(int val
)
2826 BUG_ON((preempt_count() < 0));
2827 preempt_count() += val
;
2829 * Spinlock count overflowing soon?
2831 BUG_ON((preempt_count() & PREEMPT_MASK
) >= PREEMPT_MASK
-10);
2833 EXPORT_SYMBOL(add_preempt_count
);
2835 void fastcall
sub_preempt_count(int val
)
2840 BUG_ON(val
> preempt_count());
2842 * Is the spinlock portion underflowing?
2844 BUG_ON((val
< PREEMPT_MASK
) && !(preempt_count() & PREEMPT_MASK
));
2845 preempt_count() -= val
;
2847 EXPORT_SYMBOL(sub_preempt_count
);
2852 * schedule() is the main scheduler function.
2854 asmlinkage
void __sched
schedule(void)
2857 task_t
*prev
, *next
;
2859 prio_array_t
*array
;
2860 struct list_head
*queue
;
2861 unsigned long long now
;
2862 unsigned long run_time
;
2863 int cpu
, idx
, new_prio
;
2866 * Test if we are atomic. Since do_exit() needs to call into
2867 * schedule() atomically, we ignore that path for now.
2868 * Otherwise, whine if we are scheduling when we should not be.
2870 if (likely(!current
->exit_state
)) {
2871 if (unlikely(in_atomic())) {
2872 printk(KERN_ERR
"scheduling while atomic: "
2874 current
->comm
, preempt_count(), current
->pid
);
2878 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2883 release_kernel_lock(prev
);
2884 need_resched_nonpreemptible
:
2888 * The idle thread is not allowed to schedule!
2889 * Remove this check after it has been exercised a bit.
2891 if (unlikely(prev
== rq
->idle
) && prev
->state
!= TASK_RUNNING
) {
2892 printk(KERN_ERR
"bad: scheduling from the idle thread!\n");
2896 schedstat_inc(rq
, sched_cnt
);
2897 now
= sched_clock();
2898 if (likely((long long)(now
- prev
->timestamp
) < NS_MAX_SLEEP_AVG
)) {
2899 run_time
= now
- prev
->timestamp
;
2900 if (unlikely((long long)(now
- prev
->timestamp
) < 0))
2903 run_time
= NS_MAX_SLEEP_AVG
;
2906 * Tasks charged proportionately less run_time at high sleep_avg to
2907 * delay them losing their interactive status
2909 run_time
/= (CURRENT_BONUS(prev
) ? : 1);
2911 spin_lock_irq(&rq
->lock
);
2913 if (unlikely(prev
->flags
& PF_DEAD
))
2914 prev
->state
= EXIT_DEAD
;
2916 switch_count
= &prev
->nivcsw
;
2917 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
2918 switch_count
= &prev
->nvcsw
;
2919 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
2920 unlikely(signal_pending(prev
))))
2921 prev
->state
= TASK_RUNNING
;
2923 if (prev
->state
== TASK_UNINTERRUPTIBLE
)
2924 rq
->nr_uninterruptible
++;
2925 deactivate_task(prev
, rq
);
2929 cpu
= smp_processor_id();
2930 if (unlikely(!rq
->nr_running
)) {
2932 idle_balance(cpu
, rq
);
2933 if (!rq
->nr_running
) {
2935 rq
->expired_timestamp
= 0;
2936 wake_sleeping_dependent(cpu
, rq
);
2938 * wake_sleeping_dependent() might have released
2939 * the runqueue, so break out if we got new
2942 if (!rq
->nr_running
)
2946 if (dependent_sleeper(cpu
, rq
)) {
2951 * dependent_sleeper() releases and reacquires the runqueue
2952 * lock, hence go into the idle loop if the rq went
2955 if (unlikely(!rq
->nr_running
))
2960 if (unlikely(!array
->nr_active
)) {
2962 * Switch the active and expired arrays.
2964 schedstat_inc(rq
, sched_switch
);
2965 rq
->active
= rq
->expired
;
2966 rq
->expired
= array
;
2968 rq
->expired_timestamp
= 0;
2969 rq
->best_expired_prio
= MAX_PRIO
;
2972 idx
= sched_find_first_bit(array
->bitmap
);
2973 queue
= array
->queue
+ idx
;
2974 next
= list_entry(queue
->next
, task_t
, run_list
);
2976 if (!rt_task(next
) && next
->activated
> 0) {
2977 unsigned long long delta
= now
- next
->timestamp
;
2978 if (unlikely((long long)(now
- next
->timestamp
) < 0))
2981 if (next
->activated
== 1)
2982 delta
= delta
* (ON_RUNQUEUE_WEIGHT
* 128 / 100) / 128;
2984 array
= next
->array
;
2985 new_prio
= recalc_task_prio(next
, next
->timestamp
+ delta
);
2987 if (unlikely(next
->prio
!= new_prio
)) {
2988 dequeue_task(next
, array
);
2989 next
->prio
= new_prio
;
2990 enqueue_task(next
, array
);
2992 requeue_task(next
, array
);
2994 next
->activated
= 0;
2996 if (next
== rq
->idle
)
2997 schedstat_inc(rq
, sched_goidle
);
2999 prefetch_stack(next
);
3000 clear_tsk_need_resched(prev
);
3001 rcu_qsctr_inc(task_cpu(prev
));
3003 update_cpu_clock(prev
, rq
, now
);
3005 prev
->sleep_avg
-= run_time
;
3006 if ((long)prev
->sleep_avg
<= 0)
3007 prev
->sleep_avg
= 0;
3008 prev
->timestamp
= prev
->last_ran
= now
;
3010 sched_info_switch(prev
, next
);
3011 if (likely(prev
!= next
)) {
3012 next
->timestamp
= now
;
3017 prepare_task_switch(rq
, next
);
3018 prev
= context_switch(rq
, prev
, next
);
3021 * this_rq must be evaluated again because prev may have moved
3022 * CPUs since it called schedule(), thus the 'rq' on its stack
3023 * frame will be invalid.
3025 finish_task_switch(this_rq(), prev
);
3027 spin_unlock_irq(&rq
->lock
);
3030 if (unlikely(reacquire_kernel_lock(prev
) < 0))
3031 goto need_resched_nonpreemptible
;
3032 preempt_enable_no_resched();
3033 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3037 EXPORT_SYMBOL(schedule
);
3039 #ifdef CONFIG_PREEMPT
3041 * this is is the entry point to schedule() from in-kernel preemption
3042 * off of preempt_enable. Kernel preemptions off return from interrupt
3043 * occur there and call schedule directly.
3045 asmlinkage
void __sched
preempt_schedule(void)
3047 struct thread_info
*ti
= current_thread_info();
3048 #ifdef CONFIG_PREEMPT_BKL
3049 struct task_struct
*task
= current
;
3050 int saved_lock_depth
;
3053 * If there is a non-zero preempt_count or interrupts are disabled,
3054 * we do not want to preempt the current task. Just return..
3056 if (unlikely(ti
->preempt_count
|| irqs_disabled()))
3060 add_preempt_count(PREEMPT_ACTIVE
);
3062 * We keep the big kernel semaphore locked, but we
3063 * clear ->lock_depth so that schedule() doesnt
3064 * auto-release the semaphore:
3066 #ifdef CONFIG_PREEMPT_BKL
3067 saved_lock_depth
= task
->lock_depth
;
3068 task
->lock_depth
= -1;
3071 #ifdef CONFIG_PREEMPT_BKL
3072 task
->lock_depth
= saved_lock_depth
;
3074 sub_preempt_count(PREEMPT_ACTIVE
);
3076 /* we could miss a preemption opportunity between schedule and now */
3078 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3082 EXPORT_SYMBOL(preempt_schedule
);
3085 * this is is the entry point to schedule() from kernel preemption
3086 * off of irq context.
3087 * Note, that this is called and return with irqs disabled. This will
3088 * protect us against recursive calling from irq.
3090 asmlinkage
void __sched
preempt_schedule_irq(void)
3092 struct thread_info
*ti
= current_thread_info();
3093 #ifdef CONFIG_PREEMPT_BKL
3094 struct task_struct
*task
= current
;
3095 int saved_lock_depth
;
3097 /* Catch callers which need to be fixed*/
3098 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3101 add_preempt_count(PREEMPT_ACTIVE
);
3103 * We keep the big kernel semaphore locked, but we
3104 * clear ->lock_depth so that schedule() doesnt
3105 * auto-release the semaphore:
3107 #ifdef CONFIG_PREEMPT_BKL
3108 saved_lock_depth
= task
->lock_depth
;
3109 task
->lock_depth
= -1;
3113 local_irq_disable();
3114 #ifdef CONFIG_PREEMPT_BKL
3115 task
->lock_depth
= saved_lock_depth
;
3117 sub_preempt_count(PREEMPT_ACTIVE
);
3119 /* we could miss a preemption opportunity between schedule and now */
3121 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3125 #endif /* CONFIG_PREEMPT */
3127 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3130 task_t
*p
= curr
->private;
3131 return try_to_wake_up(p
, mode
, sync
);
3134 EXPORT_SYMBOL(default_wake_function
);
3137 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3138 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3139 * number) then we wake all the non-exclusive tasks and one exclusive task.
3141 * There are circumstances in which we can try to wake a task which has already
3142 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3143 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3145 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3146 int nr_exclusive
, int sync
, void *key
)
3148 struct list_head
*tmp
, *next
;
3150 list_for_each_safe(tmp
, next
, &q
->task_list
) {
3153 curr
= list_entry(tmp
, wait_queue_t
, task_list
);
3154 flags
= curr
->flags
;
3155 if (curr
->func(curr
, mode
, sync
, key
) &&
3156 (flags
& WQ_FLAG_EXCLUSIVE
) &&
3163 * __wake_up - wake up threads blocked on a waitqueue.
3165 * @mode: which threads
3166 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3167 * @key: is directly passed to the wakeup function
3169 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3170 int nr_exclusive
, void *key
)
3172 unsigned long flags
;
3174 spin_lock_irqsave(&q
->lock
, flags
);
3175 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3176 spin_unlock_irqrestore(&q
->lock
, flags
);
3179 EXPORT_SYMBOL(__wake_up
);
3182 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3184 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3186 __wake_up_common(q
, mode
, 1, 0, NULL
);
3190 * __wake_up_sync - wake up threads blocked on a waitqueue.
3192 * @mode: which threads
3193 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3195 * The sync wakeup differs that the waker knows that it will schedule
3196 * away soon, so while the target thread will be woken up, it will not
3197 * be migrated to another CPU - ie. the two threads are 'synchronized'
3198 * with each other. This can prevent needless bouncing between CPUs.
3200 * On UP it can prevent extra preemption.
3203 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3205 unsigned long flags
;
3211 if (unlikely(!nr_exclusive
))
3214 spin_lock_irqsave(&q
->lock
, flags
);
3215 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3216 spin_unlock_irqrestore(&q
->lock
, flags
);
3218 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3220 void fastcall
complete(struct completion
*x
)
3222 unsigned long flags
;
3224 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3226 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3228 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3230 EXPORT_SYMBOL(complete
);
3232 void fastcall
complete_all(struct completion
*x
)
3234 unsigned long flags
;
3236 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3237 x
->done
+= UINT_MAX
/2;
3238 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3240 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3242 EXPORT_SYMBOL(complete_all
);
3244 void fastcall __sched
wait_for_completion(struct completion
*x
)
3247 spin_lock_irq(&x
->wait
.lock
);
3249 DECLARE_WAITQUEUE(wait
, current
);
3251 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3252 __add_wait_queue_tail(&x
->wait
, &wait
);
3254 __set_current_state(TASK_UNINTERRUPTIBLE
);
3255 spin_unlock_irq(&x
->wait
.lock
);
3257 spin_lock_irq(&x
->wait
.lock
);
3259 __remove_wait_queue(&x
->wait
, &wait
);
3262 spin_unlock_irq(&x
->wait
.lock
);
3264 EXPORT_SYMBOL(wait_for_completion
);
3266 unsigned long fastcall __sched
3267 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3271 spin_lock_irq(&x
->wait
.lock
);
3273 DECLARE_WAITQUEUE(wait
, current
);
3275 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3276 __add_wait_queue_tail(&x
->wait
, &wait
);
3278 __set_current_state(TASK_UNINTERRUPTIBLE
);
3279 spin_unlock_irq(&x
->wait
.lock
);
3280 timeout
= schedule_timeout(timeout
);
3281 spin_lock_irq(&x
->wait
.lock
);
3283 __remove_wait_queue(&x
->wait
, &wait
);
3287 __remove_wait_queue(&x
->wait
, &wait
);
3291 spin_unlock_irq(&x
->wait
.lock
);
3294 EXPORT_SYMBOL(wait_for_completion_timeout
);
3296 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3302 spin_lock_irq(&x
->wait
.lock
);
3304 DECLARE_WAITQUEUE(wait
, current
);
3306 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3307 __add_wait_queue_tail(&x
->wait
, &wait
);
3309 if (signal_pending(current
)) {
3311 __remove_wait_queue(&x
->wait
, &wait
);
3314 __set_current_state(TASK_INTERRUPTIBLE
);
3315 spin_unlock_irq(&x
->wait
.lock
);
3317 spin_lock_irq(&x
->wait
.lock
);
3319 __remove_wait_queue(&x
->wait
, &wait
);
3323 spin_unlock_irq(&x
->wait
.lock
);
3327 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3329 unsigned long fastcall __sched
3330 wait_for_completion_interruptible_timeout(struct completion
*x
,
3331 unsigned long timeout
)
3335 spin_lock_irq(&x
->wait
.lock
);
3337 DECLARE_WAITQUEUE(wait
, current
);
3339 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3340 __add_wait_queue_tail(&x
->wait
, &wait
);
3342 if (signal_pending(current
)) {
3343 timeout
= -ERESTARTSYS
;
3344 __remove_wait_queue(&x
->wait
, &wait
);
3347 __set_current_state(TASK_INTERRUPTIBLE
);
3348 spin_unlock_irq(&x
->wait
.lock
);
3349 timeout
= schedule_timeout(timeout
);
3350 spin_lock_irq(&x
->wait
.lock
);
3352 __remove_wait_queue(&x
->wait
, &wait
);
3356 __remove_wait_queue(&x
->wait
, &wait
);
3360 spin_unlock_irq(&x
->wait
.lock
);
3363 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3366 #define SLEEP_ON_VAR \
3367 unsigned long flags; \
3368 wait_queue_t wait; \
3369 init_waitqueue_entry(&wait, current);
3371 #define SLEEP_ON_HEAD \
3372 spin_lock_irqsave(&q->lock,flags); \
3373 __add_wait_queue(q, &wait); \
3374 spin_unlock(&q->lock);
3376 #define SLEEP_ON_TAIL \
3377 spin_lock_irq(&q->lock); \
3378 __remove_wait_queue(q, &wait); \
3379 spin_unlock_irqrestore(&q->lock, flags);
3381 void fastcall __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3385 current
->state
= TASK_INTERRUPTIBLE
;
3392 EXPORT_SYMBOL(interruptible_sleep_on
);
3394 long fastcall __sched
3395 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3399 current
->state
= TASK_INTERRUPTIBLE
;
3402 timeout
= schedule_timeout(timeout
);
3408 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3410 void fastcall __sched
sleep_on(wait_queue_head_t
*q
)
3414 current
->state
= TASK_UNINTERRUPTIBLE
;
3421 EXPORT_SYMBOL(sleep_on
);
3423 long fastcall __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3427 current
->state
= TASK_UNINTERRUPTIBLE
;
3430 timeout
= schedule_timeout(timeout
);
3436 EXPORT_SYMBOL(sleep_on_timeout
);
3438 void set_user_nice(task_t
*p
, long nice
)
3440 unsigned long flags
;
3441 prio_array_t
*array
;
3443 int old_prio
, new_prio
, delta
;
3445 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3448 * We have to be careful, if called from sys_setpriority(),
3449 * the task might be in the middle of scheduling on another CPU.
3451 rq
= task_rq_lock(p
, &flags
);
3453 * The RT priorities are set via sched_setscheduler(), but we still
3454 * allow the 'normal' nice value to be set - but as expected
3455 * it wont have any effect on scheduling until the task is
3456 * not SCHED_NORMAL/SCHED_BATCH:
3459 p
->static_prio
= NICE_TO_PRIO(nice
);
3464 dequeue_task(p
, array
);
3467 new_prio
= NICE_TO_PRIO(nice
);
3468 delta
= new_prio
- old_prio
;
3469 p
->static_prio
= NICE_TO_PRIO(nice
);
3473 enqueue_task(p
, array
);
3475 * If the task increased its priority or is running and
3476 * lowered its priority, then reschedule its CPU:
3478 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3479 resched_task(rq
->curr
);
3482 task_rq_unlock(rq
, &flags
);
3485 EXPORT_SYMBOL(set_user_nice
);
3488 * can_nice - check if a task can reduce its nice value
3492 int can_nice(const task_t
*p
, const int nice
)
3494 /* convert nice value [19,-20] to rlimit style value [1,40] */
3495 int nice_rlim
= 20 - nice
;
3496 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
3497 capable(CAP_SYS_NICE
));
3500 #ifdef __ARCH_WANT_SYS_NICE
3503 * sys_nice - change the priority of the current process.
3504 * @increment: priority increment
3506 * sys_setpriority is a more generic, but much slower function that
3507 * does similar things.
3509 asmlinkage
long sys_nice(int increment
)
3515 * Setpriority might change our priority at the same moment.
3516 * We don't have to worry. Conceptually one call occurs first
3517 * and we have a single winner.
3519 if (increment
< -40)
3524 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
3530 if (increment
< 0 && !can_nice(current
, nice
))
3533 retval
= security_task_setnice(current
, nice
);
3537 set_user_nice(current
, nice
);
3544 * task_prio - return the priority value of a given task.
3545 * @p: the task in question.
3547 * This is the priority value as seen by users in /proc.
3548 * RT tasks are offset by -200. Normal tasks are centered
3549 * around 0, value goes from -16 to +15.
3551 int task_prio(const task_t
*p
)
3553 return p
->prio
- MAX_RT_PRIO
;
3557 * task_nice - return the nice value of a given task.
3558 * @p: the task in question.
3560 int task_nice(const task_t
*p
)
3562 return TASK_NICE(p
);
3564 EXPORT_SYMBOL_GPL(task_nice
);
3567 * idle_cpu - is a given cpu idle currently?
3568 * @cpu: the processor in question.
3570 int idle_cpu(int cpu
)
3572 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
3576 * idle_task - return the idle task for a given cpu.
3577 * @cpu: the processor in question.
3579 task_t
*idle_task(int cpu
)
3581 return cpu_rq(cpu
)->idle
;
3585 * find_process_by_pid - find a process with a matching PID value.
3586 * @pid: the pid in question.
3588 static inline task_t
*find_process_by_pid(pid_t pid
)
3590 return pid
? find_task_by_pid(pid
) : current
;
3593 /* Actually do priority change: must hold rq lock. */
3594 static void __setscheduler(struct task_struct
*p
, int policy
, int prio
)
3598 p
->rt_priority
= prio
;
3599 if (policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
) {
3600 p
->prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
3602 p
->prio
= p
->static_prio
;
3604 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
3606 if (policy
== SCHED_BATCH
)
3612 * sched_setscheduler - change the scheduling policy and/or RT priority of
3614 * @p: the task in question.
3615 * @policy: new policy.
3616 * @param: structure containing the new RT priority.
3618 int sched_setscheduler(struct task_struct
*p
, int policy
,
3619 struct sched_param
*param
)
3622 int oldprio
, oldpolicy
= -1;
3623 prio_array_t
*array
;
3624 unsigned long flags
;
3628 /* double check policy once rq lock held */
3630 policy
= oldpolicy
= p
->policy
;
3631 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3632 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
)
3635 * Valid priorities for SCHED_FIFO and SCHED_RR are
3636 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
3639 if (param
->sched_priority
< 0 ||
3640 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3641 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
3643 if ((policy
== SCHED_NORMAL
|| policy
== SCHED_BATCH
)
3644 != (param
->sched_priority
== 0))
3648 * Allow unprivileged RT tasks to decrease priority:
3650 if (!capable(CAP_SYS_NICE
)) {
3652 * can't change policy, except between SCHED_NORMAL
3655 if (((policy
!= SCHED_NORMAL
&& p
->policy
!= SCHED_BATCH
) &&
3656 (policy
!= SCHED_BATCH
&& p
->policy
!= SCHED_NORMAL
)) &&
3657 !p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
)
3659 /* can't increase priority */
3660 if ((policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
) &&
3661 param
->sched_priority
> p
->rt_priority
&&
3662 param
->sched_priority
>
3663 p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
)
3665 /* can't change other user's priorities */
3666 if ((current
->euid
!= p
->euid
) &&
3667 (current
->euid
!= p
->uid
))
3671 retval
= security_task_setscheduler(p
, policy
, param
);
3675 * To be able to change p->policy safely, the apropriate
3676 * runqueue lock must be held.
3678 rq
= task_rq_lock(p
, &flags
);
3679 /* recheck policy now with rq lock held */
3680 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3681 policy
= oldpolicy
= -1;
3682 task_rq_unlock(rq
, &flags
);
3687 deactivate_task(p
, rq
);
3689 __setscheduler(p
, policy
, param
->sched_priority
);
3691 __activate_task(p
, rq
);
3693 * Reschedule if we are currently running on this runqueue and
3694 * our priority decreased, or if we are not currently running on
3695 * this runqueue and our priority is higher than the current's
3697 if (task_running(rq
, p
)) {
3698 if (p
->prio
> oldprio
)
3699 resched_task(rq
->curr
);
3700 } else if (TASK_PREEMPTS_CURR(p
, rq
))
3701 resched_task(rq
->curr
);
3703 task_rq_unlock(rq
, &flags
);
3706 EXPORT_SYMBOL_GPL(sched_setscheduler
);
3709 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
3712 struct sched_param lparam
;
3713 struct task_struct
*p
;
3715 if (!param
|| pid
< 0)
3717 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
3719 read_lock_irq(&tasklist_lock
);
3720 p
= find_process_by_pid(pid
);
3722 read_unlock_irq(&tasklist_lock
);
3725 retval
= sched_setscheduler(p
, policy
, &lparam
);
3726 read_unlock_irq(&tasklist_lock
);
3731 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3732 * @pid: the pid in question.
3733 * @policy: new policy.
3734 * @param: structure containing the new RT priority.
3736 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
3737 struct sched_param __user
*param
)
3739 /* negative values for policy are not valid */
3743 return do_sched_setscheduler(pid
, policy
, param
);
3747 * sys_sched_setparam - set/change the RT priority of a thread
3748 * @pid: the pid in question.
3749 * @param: structure containing the new RT priority.
3751 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
3753 return do_sched_setscheduler(pid
, -1, param
);
3757 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3758 * @pid: the pid in question.
3760 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
3762 int retval
= -EINVAL
;
3769 read_lock(&tasklist_lock
);
3770 p
= find_process_by_pid(pid
);
3772 retval
= security_task_getscheduler(p
);
3776 read_unlock(&tasklist_lock
);
3783 * sys_sched_getscheduler - get the RT priority of a thread
3784 * @pid: the pid in question.
3785 * @param: structure containing the RT priority.
3787 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
3789 struct sched_param lp
;
3790 int retval
= -EINVAL
;
3793 if (!param
|| pid
< 0)
3796 read_lock(&tasklist_lock
);
3797 p
= find_process_by_pid(pid
);
3802 retval
= security_task_getscheduler(p
);
3806 lp
.sched_priority
= p
->rt_priority
;
3807 read_unlock(&tasklist_lock
);
3810 * This one might sleep, we cannot do it with a spinlock held ...
3812 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
3818 read_unlock(&tasklist_lock
);
3822 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
3826 cpumask_t cpus_allowed
;
3829 read_lock(&tasklist_lock
);
3831 p
= find_process_by_pid(pid
);
3833 read_unlock(&tasklist_lock
);
3834 unlock_cpu_hotplug();
3839 * It is not safe to call set_cpus_allowed with the
3840 * tasklist_lock held. We will bump the task_struct's
3841 * usage count and then drop tasklist_lock.
3844 read_unlock(&tasklist_lock
);
3847 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
3848 !capable(CAP_SYS_NICE
))
3851 cpus_allowed
= cpuset_cpus_allowed(p
);
3852 cpus_and(new_mask
, new_mask
, cpus_allowed
);
3853 retval
= set_cpus_allowed(p
, new_mask
);
3857 unlock_cpu_hotplug();
3861 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
3862 cpumask_t
*new_mask
)
3864 if (len
< sizeof(cpumask_t
)) {
3865 memset(new_mask
, 0, sizeof(cpumask_t
));
3866 } else if (len
> sizeof(cpumask_t
)) {
3867 len
= sizeof(cpumask_t
);
3869 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
3873 * sys_sched_setaffinity - set the cpu affinity of a process
3874 * @pid: pid of the process
3875 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3876 * @user_mask_ptr: user-space pointer to the new cpu mask
3878 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
3879 unsigned long __user
*user_mask_ptr
)
3884 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
3888 return sched_setaffinity(pid
, new_mask
);
3892 * Represents all cpu's present in the system
3893 * In systems capable of hotplug, this map could dynamically grow
3894 * as new cpu's are detected in the system via any platform specific
3895 * method, such as ACPI for e.g.
3898 cpumask_t cpu_present_map __read_mostly
;
3899 EXPORT_SYMBOL(cpu_present_map
);
3902 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
3903 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
3906 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
3912 read_lock(&tasklist_lock
);
3915 p
= find_process_by_pid(pid
);
3920 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
3923 read_unlock(&tasklist_lock
);
3924 unlock_cpu_hotplug();
3932 * sys_sched_getaffinity - get the cpu affinity of a process
3933 * @pid: pid of the process
3934 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3935 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3937 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
3938 unsigned long __user
*user_mask_ptr
)
3943 if (len
< sizeof(cpumask_t
))
3946 ret
= sched_getaffinity(pid
, &mask
);
3950 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
3953 return sizeof(cpumask_t
);
3957 * sys_sched_yield - yield the current processor to other threads.
3959 * this function yields the current CPU by moving the calling thread
3960 * to the expired array. If there are no other threads running on this
3961 * CPU then this function will return.
3963 asmlinkage
long sys_sched_yield(void)
3965 runqueue_t
*rq
= this_rq_lock();
3966 prio_array_t
*array
= current
->array
;
3967 prio_array_t
*target
= rq
->expired
;
3969 schedstat_inc(rq
, yld_cnt
);
3971 * We implement yielding by moving the task into the expired
3974 * (special rule: RT tasks will just roundrobin in the active
3977 if (rt_task(current
))
3978 target
= rq
->active
;
3980 if (array
->nr_active
== 1) {
3981 schedstat_inc(rq
, yld_act_empty
);
3982 if (!rq
->expired
->nr_active
)
3983 schedstat_inc(rq
, yld_both_empty
);
3984 } else if (!rq
->expired
->nr_active
)
3985 schedstat_inc(rq
, yld_exp_empty
);
3987 if (array
!= target
) {
3988 dequeue_task(current
, array
);
3989 enqueue_task(current
, target
);
3992 * requeue_task is cheaper so perform that if possible.
3994 requeue_task(current
, array
);
3997 * Since we are going to call schedule() anyway, there's
3998 * no need to preempt or enable interrupts:
4000 __release(rq
->lock
);
4001 _raw_spin_unlock(&rq
->lock
);
4002 preempt_enable_no_resched();
4009 static inline void __cond_resched(void)
4012 * The BKS might be reacquired before we have dropped
4013 * PREEMPT_ACTIVE, which could trigger a second
4014 * cond_resched() call.
4016 if (unlikely(preempt_count()))
4018 if (unlikely(system_state
!= SYSTEM_RUNNING
))
4021 add_preempt_count(PREEMPT_ACTIVE
);
4023 sub_preempt_count(PREEMPT_ACTIVE
);
4024 } while (need_resched());
4027 int __sched
cond_resched(void)
4029 if (need_resched()) {
4036 EXPORT_SYMBOL(cond_resched
);
4039 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4040 * call schedule, and on return reacquire the lock.
4042 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4043 * operations here to prevent schedule() from being called twice (once via
4044 * spin_unlock(), once by hand).
4046 int cond_resched_lock(spinlock_t
*lock
)
4050 if (need_lockbreak(lock
)) {
4056 if (need_resched()) {
4057 _raw_spin_unlock(lock
);
4058 preempt_enable_no_resched();
4066 EXPORT_SYMBOL(cond_resched_lock
);
4068 int __sched
cond_resched_softirq(void)
4070 BUG_ON(!in_softirq());
4072 if (need_resched()) {
4073 __local_bh_enable();
4081 EXPORT_SYMBOL(cond_resched_softirq
);
4085 * yield - yield the current processor to other threads.
4087 * this is a shortcut for kernel-space yielding - it marks the
4088 * thread runnable and calls sys_sched_yield().
4090 void __sched
yield(void)
4092 set_current_state(TASK_RUNNING
);
4096 EXPORT_SYMBOL(yield
);
4099 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4100 * that process accounting knows that this is a task in IO wait state.
4102 * But don't do that if it is a deliberate, throttling IO wait (this task
4103 * has set its backing_dev_info: the queue against which it should throttle)
4105 void __sched
io_schedule(void)
4107 struct runqueue
*rq
= &per_cpu(runqueues
, raw_smp_processor_id());
4109 atomic_inc(&rq
->nr_iowait
);
4111 atomic_dec(&rq
->nr_iowait
);
4114 EXPORT_SYMBOL(io_schedule
);
4116 long __sched
io_schedule_timeout(long timeout
)
4118 struct runqueue
*rq
= &per_cpu(runqueues
, raw_smp_processor_id());
4121 atomic_inc(&rq
->nr_iowait
);
4122 ret
= schedule_timeout(timeout
);
4123 atomic_dec(&rq
->nr_iowait
);
4128 * sys_sched_get_priority_max - return maximum RT priority.
4129 * @policy: scheduling class.
4131 * this syscall returns the maximum rt_priority that can be used
4132 * by a given scheduling class.
4134 asmlinkage
long sys_sched_get_priority_max(int policy
)
4141 ret
= MAX_USER_RT_PRIO
-1;
4152 * sys_sched_get_priority_min - return minimum RT priority.
4153 * @policy: scheduling class.
4155 * this syscall returns the minimum rt_priority that can be used
4156 * by a given scheduling class.
4158 asmlinkage
long sys_sched_get_priority_min(int policy
)
4175 * sys_sched_rr_get_interval - return the default timeslice of a process.
4176 * @pid: pid of the process.
4177 * @interval: userspace pointer to the timeslice value.
4179 * this syscall writes the default timeslice value of a given process
4180 * into the user-space timespec buffer. A value of '0' means infinity.
4183 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4185 int retval
= -EINVAL
;
4193 read_lock(&tasklist_lock
);
4194 p
= find_process_by_pid(pid
);
4198 retval
= security_task_getscheduler(p
);
4202 jiffies_to_timespec(p
->policy
& SCHED_FIFO
?
4203 0 : task_timeslice(p
), &t
);
4204 read_unlock(&tasklist_lock
);
4205 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4209 read_unlock(&tasklist_lock
);
4213 static inline struct task_struct
*eldest_child(struct task_struct
*p
)
4215 if (list_empty(&p
->children
)) return NULL
;
4216 return list_entry(p
->children
.next
,struct task_struct
,sibling
);
4219 static inline struct task_struct
*older_sibling(struct task_struct
*p
)
4221 if (p
->sibling
.prev
==&p
->parent
->children
) return NULL
;
4222 return list_entry(p
->sibling
.prev
,struct task_struct
,sibling
);
4225 static inline struct task_struct
*younger_sibling(struct task_struct
*p
)
4227 if (p
->sibling
.next
==&p
->parent
->children
) return NULL
;
4228 return list_entry(p
->sibling
.next
,struct task_struct
,sibling
);
4231 static void show_task(task_t
*p
)
4235 unsigned long free
= 0;
4236 static const char *stat_nam
[] = { "R", "S", "D", "T", "t", "Z", "X" };
4238 printk("%-13.13s ", p
->comm
);
4239 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4240 if (state
< ARRAY_SIZE(stat_nam
))
4241 printk(stat_nam
[state
]);
4244 #if (BITS_PER_LONG == 32)
4245 if (state
== TASK_RUNNING
)
4246 printk(" running ");
4248 printk(" %08lX ", thread_saved_pc(p
));
4250 if (state
== TASK_RUNNING
)
4251 printk(" running task ");
4253 printk(" %016lx ", thread_saved_pc(p
));
4255 #ifdef CONFIG_DEBUG_STACK_USAGE
4257 unsigned long *n
= end_of_stack(p
);
4260 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4263 printk("%5lu %5d %6d ", free
, p
->pid
, p
->parent
->pid
);
4264 if ((relative
= eldest_child(p
)))
4265 printk("%5d ", relative
->pid
);
4268 if ((relative
= younger_sibling(p
)))
4269 printk("%7d", relative
->pid
);
4272 if ((relative
= older_sibling(p
)))
4273 printk(" %5d", relative
->pid
);
4277 printk(" (L-TLB)\n");
4279 printk(" (NOTLB)\n");
4281 if (state
!= TASK_RUNNING
)
4282 show_stack(p
, NULL
);
4285 void show_state(void)
4289 #if (BITS_PER_LONG == 32)
4292 printk(" task PC pid father child younger older\n");
4296 printk(" task PC pid father child younger older\n");
4298 read_lock(&tasklist_lock
);
4299 do_each_thread(g
, p
) {
4301 * reset the NMI-timeout, listing all files on a slow
4302 * console might take alot of time:
4304 touch_nmi_watchdog();
4306 } while_each_thread(g
, p
);
4308 read_unlock(&tasklist_lock
);
4309 mutex_debug_show_all_locks();
4313 * init_idle - set up an idle thread for a given CPU
4314 * @idle: task in question
4315 * @cpu: cpu the idle task belongs to
4317 * NOTE: this function does not set the idle thread's NEED_RESCHED
4318 * flag, to make booting more robust.
4320 void __devinit
init_idle(task_t
*idle
, int cpu
)
4322 runqueue_t
*rq
= cpu_rq(cpu
);
4323 unsigned long flags
;
4325 idle
->timestamp
= sched_clock();
4326 idle
->sleep_avg
= 0;
4328 idle
->prio
= MAX_PRIO
;
4329 idle
->state
= TASK_RUNNING
;
4330 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4331 set_task_cpu(idle
, cpu
);
4333 spin_lock_irqsave(&rq
->lock
, flags
);
4334 rq
->curr
= rq
->idle
= idle
;
4335 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4338 spin_unlock_irqrestore(&rq
->lock
, flags
);
4340 /* Set the preempt count _outside_ the spinlocks! */
4341 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4342 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4344 task_thread_info(idle
)->preempt_count
= 0;
4349 * In a system that switches off the HZ timer nohz_cpu_mask
4350 * indicates which cpus entered this state. This is used
4351 * in the rcu update to wait only for active cpus. For system
4352 * which do not switch off the HZ timer nohz_cpu_mask should
4353 * always be CPU_MASK_NONE.
4355 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4359 * This is how migration works:
4361 * 1) we queue a migration_req_t structure in the source CPU's
4362 * runqueue and wake up that CPU's migration thread.
4363 * 2) we down() the locked semaphore => thread blocks.
4364 * 3) migration thread wakes up (implicitly it forces the migrated
4365 * thread off the CPU)
4366 * 4) it gets the migration request and checks whether the migrated
4367 * task is still in the wrong runqueue.
4368 * 5) if it's in the wrong runqueue then the migration thread removes
4369 * it and puts it into the right queue.
4370 * 6) migration thread up()s the semaphore.
4371 * 7) we wake up and the migration is done.
4375 * Change a given task's CPU affinity. Migrate the thread to a
4376 * proper CPU and schedule it away if the CPU it's executing on
4377 * is removed from the allowed bitmask.
4379 * NOTE: the caller must have a valid reference to the task, the
4380 * task must not exit() & deallocate itself prematurely. The
4381 * call is not atomic; no spinlocks may be held.
4383 int set_cpus_allowed(task_t
*p
, cpumask_t new_mask
)
4385 unsigned long flags
;
4387 migration_req_t req
;
4390 rq
= task_rq_lock(p
, &flags
);
4391 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4396 p
->cpus_allowed
= new_mask
;
4397 /* Can the task run on the task's current CPU? If so, we're done */
4398 if (cpu_isset(task_cpu(p
), new_mask
))
4401 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4402 /* Need help from migration thread: drop lock and wait. */
4403 task_rq_unlock(rq
, &flags
);
4404 wake_up_process(rq
->migration_thread
);
4405 wait_for_completion(&req
.done
);
4406 tlb_migrate_finish(p
->mm
);
4410 task_rq_unlock(rq
, &flags
);
4414 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4417 * Move (not current) task off this cpu, onto dest cpu. We're doing
4418 * this because either it can't run here any more (set_cpus_allowed()
4419 * away from this CPU, or CPU going down), or because we're
4420 * attempting to rebalance this task on exec (sched_exec).
4422 * So we race with normal scheduler movements, but that's OK, as long
4423 * as the task is no longer on this CPU.
4425 static void __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4427 runqueue_t
*rq_dest
, *rq_src
;
4429 if (unlikely(cpu_is_offline(dest_cpu
)))
4432 rq_src
= cpu_rq(src_cpu
);
4433 rq_dest
= cpu_rq(dest_cpu
);
4435 double_rq_lock(rq_src
, rq_dest
);
4436 /* Already moved. */
4437 if (task_cpu(p
) != src_cpu
)
4439 /* Affinity changed (again). */
4440 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
4443 set_task_cpu(p
, dest_cpu
);
4446 * Sync timestamp with rq_dest's before activating.
4447 * The same thing could be achieved by doing this step
4448 * afterwards, and pretending it was a local activate.
4449 * This way is cleaner and logically correct.
4451 p
->timestamp
= p
->timestamp
- rq_src
->timestamp_last_tick
4452 + rq_dest
->timestamp_last_tick
;
4453 deactivate_task(p
, rq_src
);
4454 activate_task(p
, rq_dest
, 0);
4455 if (TASK_PREEMPTS_CURR(p
, rq_dest
))
4456 resched_task(rq_dest
->curr
);
4460 double_rq_unlock(rq_src
, rq_dest
);
4464 * migration_thread - this is a highprio system thread that performs
4465 * thread migration by bumping thread off CPU then 'pushing' onto
4468 static int migration_thread(void *data
)
4471 int cpu
= (long)data
;
4474 BUG_ON(rq
->migration_thread
!= current
);
4476 set_current_state(TASK_INTERRUPTIBLE
);
4477 while (!kthread_should_stop()) {
4478 struct list_head
*head
;
4479 migration_req_t
*req
;
4483 spin_lock_irq(&rq
->lock
);
4485 if (cpu_is_offline(cpu
)) {
4486 spin_unlock_irq(&rq
->lock
);
4490 if (rq
->active_balance
) {
4491 active_load_balance(rq
, cpu
);
4492 rq
->active_balance
= 0;
4495 head
= &rq
->migration_queue
;
4497 if (list_empty(head
)) {
4498 spin_unlock_irq(&rq
->lock
);
4500 set_current_state(TASK_INTERRUPTIBLE
);
4503 req
= list_entry(head
->next
, migration_req_t
, list
);
4504 list_del_init(head
->next
);
4506 spin_unlock(&rq
->lock
);
4507 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
4510 complete(&req
->done
);
4512 __set_current_state(TASK_RUNNING
);
4516 /* Wait for kthread_stop */
4517 set_current_state(TASK_INTERRUPTIBLE
);
4518 while (!kthread_should_stop()) {
4520 set_current_state(TASK_INTERRUPTIBLE
);
4522 __set_current_state(TASK_RUNNING
);
4526 #ifdef CONFIG_HOTPLUG_CPU
4527 /* Figure out where task on dead CPU should go, use force if neccessary. */
4528 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*tsk
)
4534 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
4535 cpus_and(mask
, mask
, tsk
->cpus_allowed
);
4536 dest_cpu
= any_online_cpu(mask
);
4538 /* On any allowed CPU? */
4539 if (dest_cpu
== NR_CPUS
)
4540 dest_cpu
= any_online_cpu(tsk
->cpus_allowed
);
4542 /* No more Mr. Nice Guy. */
4543 if (dest_cpu
== NR_CPUS
) {
4544 cpus_setall(tsk
->cpus_allowed
);
4545 dest_cpu
= any_online_cpu(tsk
->cpus_allowed
);
4548 * Don't tell them about moving exiting tasks or
4549 * kernel threads (both mm NULL), since they never
4552 if (tsk
->mm
&& printk_ratelimit())
4553 printk(KERN_INFO
"process %d (%s) no "
4554 "longer affine to cpu%d\n",
4555 tsk
->pid
, tsk
->comm
, dead_cpu
);
4557 __migrate_task(tsk
, dead_cpu
, dest_cpu
);
4561 * While a dead CPU has no uninterruptible tasks queued at this point,
4562 * it might still have a nonzero ->nr_uninterruptible counter, because
4563 * for performance reasons the counter is not stricly tracking tasks to
4564 * their home CPUs. So we just add the counter to another CPU's counter,
4565 * to keep the global sum constant after CPU-down:
4567 static void migrate_nr_uninterruptible(runqueue_t
*rq_src
)
4569 runqueue_t
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
4570 unsigned long flags
;
4572 local_irq_save(flags
);
4573 double_rq_lock(rq_src
, rq_dest
);
4574 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
4575 rq_src
->nr_uninterruptible
= 0;
4576 double_rq_unlock(rq_src
, rq_dest
);
4577 local_irq_restore(flags
);
4580 /* Run through task list and migrate tasks from the dead cpu. */
4581 static void migrate_live_tasks(int src_cpu
)
4583 struct task_struct
*tsk
, *t
;
4585 write_lock_irq(&tasklist_lock
);
4587 do_each_thread(t
, tsk
) {
4591 if (task_cpu(tsk
) == src_cpu
)
4592 move_task_off_dead_cpu(src_cpu
, tsk
);
4593 } while_each_thread(t
, tsk
);
4595 write_unlock_irq(&tasklist_lock
);
4598 /* Schedules idle task to be the next runnable task on current CPU.
4599 * It does so by boosting its priority to highest possible and adding it to
4600 * the _front_ of runqueue. Used by CPU offline code.
4602 void sched_idle_next(void)
4604 int cpu
= smp_processor_id();
4605 runqueue_t
*rq
= this_rq();
4606 struct task_struct
*p
= rq
->idle
;
4607 unsigned long flags
;
4609 /* cpu has to be offline */
4610 BUG_ON(cpu_online(cpu
));
4612 /* Strictly not necessary since rest of the CPUs are stopped by now
4613 * and interrupts disabled on current cpu.
4615 spin_lock_irqsave(&rq
->lock
, flags
);
4617 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
4618 /* Add idle task to _front_ of it's priority queue */
4619 __activate_idle_task(p
, rq
);
4621 spin_unlock_irqrestore(&rq
->lock
, flags
);
4624 /* Ensures that the idle task is using init_mm right before its cpu goes
4627 void idle_task_exit(void)
4629 struct mm_struct
*mm
= current
->active_mm
;
4631 BUG_ON(cpu_online(smp_processor_id()));
4634 switch_mm(mm
, &init_mm
, current
);
4638 static void migrate_dead(unsigned int dead_cpu
, task_t
*tsk
)
4640 struct runqueue
*rq
= cpu_rq(dead_cpu
);
4642 /* Must be exiting, otherwise would be on tasklist. */
4643 BUG_ON(tsk
->exit_state
!= EXIT_ZOMBIE
&& tsk
->exit_state
!= EXIT_DEAD
);
4645 /* Cannot have done final schedule yet: would have vanished. */
4646 BUG_ON(tsk
->flags
& PF_DEAD
);
4648 get_task_struct(tsk
);
4651 * Drop lock around migration; if someone else moves it,
4652 * that's OK. No task can be added to this CPU, so iteration is
4655 spin_unlock_irq(&rq
->lock
);
4656 move_task_off_dead_cpu(dead_cpu
, tsk
);
4657 spin_lock_irq(&rq
->lock
);
4659 put_task_struct(tsk
);
4662 /* release_task() removes task from tasklist, so we won't find dead tasks. */
4663 static void migrate_dead_tasks(unsigned int dead_cpu
)
4666 struct runqueue
*rq
= cpu_rq(dead_cpu
);
4668 for (arr
= 0; arr
< 2; arr
++) {
4669 for (i
= 0; i
< MAX_PRIO
; i
++) {
4670 struct list_head
*list
= &rq
->arrays
[arr
].queue
[i
];
4671 while (!list_empty(list
))
4672 migrate_dead(dead_cpu
,
4673 list_entry(list
->next
, task_t
,
4678 #endif /* CONFIG_HOTPLUG_CPU */
4681 * migration_call - callback that gets triggered when a CPU is added.
4682 * Here we can start up the necessary migration thread for the new CPU.
4684 static int migration_call(struct notifier_block
*nfb
, unsigned long action
,
4687 int cpu
= (long)hcpu
;
4688 struct task_struct
*p
;
4689 struct runqueue
*rq
;
4690 unsigned long flags
;
4693 case CPU_UP_PREPARE
:
4694 p
= kthread_create(migration_thread
, hcpu
, "migration/%d",cpu
);
4697 p
->flags
|= PF_NOFREEZE
;
4698 kthread_bind(p
, cpu
);
4699 /* Must be high prio: stop_machine expects to yield to it. */
4700 rq
= task_rq_lock(p
, &flags
);
4701 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
4702 task_rq_unlock(rq
, &flags
);
4703 cpu_rq(cpu
)->migration_thread
= p
;
4706 /* Strictly unneccessary, as first user will wake it. */
4707 wake_up_process(cpu_rq(cpu
)->migration_thread
);
4709 #ifdef CONFIG_HOTPLUG_CPU
4710 case CPU_UP_CANCELED
:
4711 /* Unbind it from offline cpu so it can run. Fall thru. */
4712 kthread_bind(cpu_rq(cpu
)->migration_thread
,
4713 any_online_cpu(cpu_online_map
));
4714 kthread_stop(cpu_rq(cpu
)->migration_thread
);
4715 cpu_rq(cpu
)->migration_thread
= NULL
;
4718 migrate_live_tasks(cpu
);
4720 kthread_stop(rq
->migration_thread
);
4721 rq
->migration_thread
= NULL
;
4722 /* Idle task back to normal (off runqueue, low prio) */
4723 rq
= task_rq_lock(rq
->idle
, &flags
);
4724 deactivate_task(rq
->idle
, rq
);
4725 rq
->idle
->static_prio
= MAX_PRIO
;
4726 __setscheduler(rq
->idle
, SCHED_NORMAL
, 0);
4727 migrate_dead_tasks(cpu
);
4728 task_rq_unlock(rq
, &flags
);
4729 migrate_nr_uninterruptible(rq
);
4730 BUG_ON(rq
->nr_running
!= 0);
4732 /* No need to migrate the tasks: it was best-effort if
4733 * they didn't do lock_cpu_hotplug(). Just wake up
4734 * the requestors. */
4735 spin_lock_irq(&rq
->lock
);
4736 while (!list_empty(&rq
->migration_queue
)) {
4737 migration_req_t
*req
;
4738 req
= list_entry(rq
->migration_queue
.next
,
4739 migration_req_t
, list
);
4740 list_del_init(&req
->list
);
4741 complete(&req
->done
);
4743 spin_unlock_irq(&rq
->lock
);
4750 /* Register at highest priority so that task migration (migrate_all_tasks)
4751 * happens before everything else.
4753 static struct notifier_block __devinitdata migration_notifier
= {
4754 .notifier_call
= migration_call
,
4758 int __init
migration_init(void)
4760 void *cpu
= (void *)(long)smp_processor_id();
4761 /* Start one for boot CPU. */
4762 migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
4763 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
4764 register_cpu_notifier(&migration_notifier
);
4770 #undef SCHED_DOMAIN_DEBUG
4771 #ifdef SCHED_DOMAIN_DEBUG
4772 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
4777 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
4781 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
4786 struct sched_group
*group
= sd
->groups
;
4787 cpumask_t groupmask
;
4789 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
4790 cpus_clear(groupmask
);
4793 for (i
= 0; i
< level
+ 1; i
++)
4795 printk("domain %d: ", level
);
4797 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
4798 printk("does not load-balance\n");
4800 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain has parent");
4804 printk("span %s\n", str
);
4806 if (!cpu_isset(cpu
, sd
->span
))
4807 printk(KERN_ERR
"ERROR: domain->span does not contain CPU%d\n", cpu
);
4808 if (!cpu_isset(cpu
, group
->cpumask
))
4809 printk(KERN_ERR
"ERROR: domain->groups does not contain CPU%d\n", cpu
);
4812 for (i
= 0; i
< level
+ 2; i
++)
4818 printk(KERN_ERR
"ERROR: group is NULL\n");
4822 if (!group
->cpu_power
) {
4824 printk(KERN_ERR
"ERROR: domain->cpu_power not set\n");
4827 if (!cpus_weight(group
->cpumask
)) {
4829 printk(KERN_ERR
"ERROR: empty group\n");
4832 if (cpus_intersects(groupmask
, group
->cpumask
)) {
4834 printk(KERN_ERR
"ERROR: repeated CPUs\n");
4837 cpus_or(groupmask
, groupmask
, group
->cpumask
);
4839 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
4842 group
= group
->next
;
4843 } while (group
!= sd
->groups
);
4846 if (!cpus_equal(sd
->span
, groupmask
))
4847 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
4853 if (!cpus_subset(groupmask
, sd
->span
))
4854 printk(KERN_ERR
"ERROR: parent span is not a superset of domain->span\n");
4860 #define sched_domain_debug(sd, cpu) {}
4863 static int sd_degenerate(struct sched_domain
*sd
)
4865 if (cpus_weight(sd
->span
) == 1)
4868 /* Following flags need at least 2 groups */
4869 if (sd
->flags
& (SD_LOAD_BALANCE
|
4870 SD_BALANCE_NEWIDLE
|
4873 if (sd
->groups
!= sd
->groups
->next
)
4877 /* Following flags don't use groups */
4878 if (sd
->flags
& (SD_WAKE_IDLE
|
4886 static int sd_parent_degenerate(struct sched_domain
*sd
,
4887 struct sched_domain
*parent
)
4889 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
4891 if (sd_degenerate(parent
))
4894 if (!cpus_equal(sd
->span
, parent
->span
))
4897 /* Does parent contain flags not in child? */
4898 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
4899 if (cflags
& SD_WAKE_AFFINE
)
4900 pflags
&= ~SD_WAKE_BALANCE
;
4901 /* Flags needing groups don't count if only 1 group in parent */
4902 if (parent
->groups
== parent
->groups
->next
) {
4903 pflags
&= ~(SD_LOAD_BALANCE
|
4904 SD_BALANCE_NEWIDLE
|
4908 if (~cflags
& pflags
)
4915 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
4916 * hold the hotplug lock.
4918 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
4920 runqueue_t
*rq
= cpu_rq(cpu
);
4921 struct sched_domain
*tmp
;
4923 /* Remove the sched domains which do not contribute to scheduling. */
4924 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
4925 struct sched_domain
*parent
= tmp
->parent
;
4928 if (sd_parent_degenerate(tmp
, parent
))
4929 tmp
->parent
= parent
->parent
;
4932 if (sd
&& sd_degenerate(sd
))
4935 sched_domain_debug(sd
, cpu
);
4937 rcu_assign_pointer(rq
->sd
, sd
);
4940 /* cpus with isolated domains */
4941 static cpumask_t __devinitdata cpu_isolated_map
= CPU_MASK_NONE
;
4943 /* Setup the mask of cpus configured for isolated domains */
4944 static int __init
isolated_cpu_setup(char *str
)
4946 int ints
[NR_CPUS
], i
;
4948 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
4949 cpus_clear(cpu_isolated_map
);
4950 for (i
= 1; i
<= ints
[0]; i
++)
4951 if (ints
[i
] < NR_CPUS
)
4952 cpu_set(ints
[i
], cpu_isolated_map
);
4956 __setup ("isolcpus=", isolated_cpu_setup
);
4959 * init_sched_build_groups takes an array of groups, the cpumask we wish
4960 * to span, and a pointer to a function which identifies what group a CPU
4961 * belongs to. The return value of group_fn must be a valid index into the
4962 * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
4963 * keep track of groups covered with a cpumask_t).
4965 * init_sched_build_groups will build a circular linked list of the groups
4966 * covered by the given span, and will set each group's ->cpumask correctly,
4967 * and ->cpu_power to 0.
4969 static void init_sched_build_groups(struct sched_group groups
[], cpumask_t span
,
4970 int (*group_fn
)(int cpu
))
4972 struct sched_group
*first
= NULL
, *last
= NULL
;
4973 cpumask_t covered
= CPU_MASK_NONE
;
4976 for_each_cpu_mask(i
, span
) {
4977 int group
= group_fn(i
);
4978 struct sched_group
*sg
= &groups
[group
];
4981 if (cpu_isset(i
, covered
))
4984 sg
->cpumask
= CPU_MASK_NONE
;
4987 for_each_cpu_mask(j
, span
) {
4988 if (group_fn(j
) != group
)
4991 cpu_set(j
, covered
);
4992 cpu_set(j
, sg
->cpumask
);
5003 #define SD_NODES_PER_DOMAIN 16
5006 * Self-tuning task migration cost measurement between source and target CPUs.
5008 * This is done by measuring the cost of manipulating buffers of varying
5009 * sizes. For a given buffer-size here are the steps that are taken:
5011 * 1) the source CPU reads+dirties a shared buffer
5012 * 2) the target CPU reads+dirties the same shared buffer
5014 * We measure how long they take, in the following 4 scenarios:
5016 * - source: CPU1, target: CPU2 | cost1
5017 * - source: CPU2, target: CPU1 | cost2
5018 * - source: CPU1, target: CPU1 | cost3
5019 * - source: CPU2, target: CPU2 | cost4
5021 * We then calculate the cost3+cost4-cost1-cost2 difference - this is
5022 * the cost of migration.
5024 * We then start off from a small buffer-size and iterate up to larger
5025 * buffer sizes, in 5% steps - measuring each buffer-size separately, and
5026 * doing a maximum search for the cost. (The maximum cost for a migration
5027 * normally occurs when the working set size is around the effective cache
5030 #define SEARCH_SCOPE 2
5031 #define MIN_CACHE_SIZE (64*1024U)
5032 #define DEFAULT_CACHE_SIZE (5*1024*1024U)
5033 #define ITERATIONS 1
5034 #define SIZE_THRESH 130
5035 #define COST_THRESH 130
5038 * The migration cost is a function of 'domain distance'. Domain
5039 * distance is the number of steps a CPU has to iterate down its
5040 * domain tree to share a domain with the other CPU. The farther
5041 * two CPUs are from each other, the larger the distance gets.
5043 * Note that we use the distance only to cache measurement results,
5044 * the distance value is not used numerically otherwise. When two
5045 * CPUs have the same distance it is assumed that the migration
5046 * cost is the same. (this is a simplification but quite practical)
5048 #define MAX_DOMAIN_DISTANCE 32
5050 static unsigned long long migration_cost
[MAX_DOMAIN_DISTANCE
] =
5051 { [ 0 ... MAX_DOMAIN_DISTANCE
-1 ] =
5053 * Architectures may override the migration cost and thus avoid
5054 * boot-time calibration. Unit is nanoseconds. Mostly useful for
5055 * virtualized hardware:
5057 #ifdef CONFIG_DEFAULT_MIGRATION_COST
5058 CONFIG_DEFAULT_MIGRATION_COST
5065 * Allow override of migration cost - in units of microseconds.
5066 * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost
5067 * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs:
5069 static int __init
migration_cost_setup(char *str
)
5071 int ints
[MAX_DOMAIN_DISTANCE
+1], i
;
5073 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5075 printk("#ints: %d\n", ints
[0]);
5076 for (i
= 1; i
<= ints
[0]; i
++) {
5077 migration_cost
[i
-1] = (unsigned long long)ints
[i
]*1000;
5078 printk("migration_cost[%d]: %Ld\n", i
-1, migration_cost
[i
-1]);
5083 __setup ("migration_cost=", migration_cost_setup
);
5086 * Global multiplier (divisor) for migration-cutoff values,
5087 * in percentiles. E.g. use a value of 150 to get 1.5 times
5088 * longer cache-hot cutoff times.
5090 * (We scale it from 100 to 128 to long long handling easier.)
5093 #define MIGRATION_FACTOR_SCALE 128
5095 static unsigned int migration_factor
= MIGRATION_FACTOR_SCALE
;
5097 static int __init
setup_migration_factor(char *str
)
5099 get_option(&str
, &migration_factor
);
5100 migration_factor
= migration_factor
* MIGRATION_FACTOR_SCALE
/ 100;
5104 __setup("migration_factor=", setup_migration_factor
);
5107 * Estimated distance of two CPUs, measured via the number of domains
5108 * we have to pass for the two CPUs to be in the same span:
5110 static unsigned long domain_distance(int cpu1
, int cpu2
)
5112 unsigned long distance
= 0;
5113 struct sched_domain
*sd
;
5115 for_each_domain(cpu1
, sd
) {
5116 WARN_ON(!cpu_isset(cpu1
, sd
->span
));
5117 if (cpu_isset(cpu2
, sd
->span
))
5121 if (distance
>= MAX_DOMAIN_DISTANCE
) {
5123 distance
= MAX_DOMAIN_DISTANCE
-1;
5129 static unsigned int migration_debug
;
5131 static int __init
setup_migration_debug(char *str
)
5133 get_option(&str
, &migration_debug
);
5137 __setup("migration_debug=", setup_migration_debug
);
5140 * Maximum cache-size that the scheduler should try to measure.
5141 * Architectures with larger caches should tune this up during
5142 * bootup. Gets used in the domain-setup code (i.e. during SMP
5145 unsigned int max_cache_size
;
5147 static int __init
setup_max_cache_size(char *str
)
5149 get_option(&str
, &max_cache_size
);
5153 __setup("max_cache_size=", setup_max_cache_size
);
5156 * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This
5157 * is the operation that is timed, so we try to generate unpredictable
5158 * cachemisses that still end up filling the L2 cache:
5160 static void touch_cache(void *__cache
, unsigned long __size
)
5162 unsigned long size
= __size
/sizeof(long), chunk1
= size
/3,
5164 unsigned long *cache
= __cache
;
5167 for (i
= 0; i
< size
/6; i
+= 8) {
5170 case 1: cache
[size
-1-i
]++;
5171 case 2: cache
[chunk1
-i
]++;
5172 case 3: cache
[chunk1
+i
]++;
5173 case 4: cache
[chunk2
-i
]++;
5174 case 5: cache
[chunk2
+i
]++;
5180 * Measure the cache-cost of one task migration. Returns in units of nsec.
5182 static unsigned long long measure_one(void *cache
, unsigned long size
,
5183 int source
, int target
)
5185 cpumask_t mask
, saved_mask
;
5186 unsigned long long t0
, t1
, t2
, t3
, cost
;
5188 saved_mask
= current
->cpus_allowed
;
5191 * Flush source caches to RAM and invalidate them:
5196 * Migrate to the source CPU:
5198 mask
= cpumask_of_cpu(source
);
5199 set_cpus_allowed(current
, mask
);
5200 WARN_ON(smp_processor_id() != source
);
5203 * Dirty the working set:
5206 touch_cache(cache
, size
);
5210 * Migrate to the target CPU, dirty the L2 cache and access
5211 * the shared buffer. (which represents the working set
5212 * of a migrated task.)
5214 mask
= cpumask_of_cpu(target
);
5215 set_cpus_allowed(current
, mask
);
5216 WARN_ON(smp_processor_id() != target
);
5219 touch_cache(cache
, size
);
5222 cost
= t1
-t0
+ t3
-t2
;
5224 if (migration_debug
>= 2)
5225 printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n",
5226 source
, target
, t1
-t0
, t1
-t0
, t3
-t2
, cost
);
5228 * Flush target caches to RAM and invalidate them:
5232 set_cpus_allowed(current
, saved_mask
);
5238 * Measure a series of task migrations and return the average
5239 * result. Since this code runs early during bootup the system
5240 * is 'undisturbed' and the average latency makes sense.
5242 * The algorithm in essence auto-detects the relevant cache-size,
5243 * so it will properly detect different cachesizes for different
5244 * cache-hierarchies, depending on how the CPUs are connected.
5246 * Architectures can prime the upper limit of the search range via
5247 * max_cache_size, otherwise the search range defaults to 20MB...64K.
5249 static unsigned long long
5250 measure_cost(int cpu1
, int cpu2
, void *cache
, unsigned int size
)
5252 unsigned long long cost1
, cost2
;
5256 * Measure the migration cost of 'size' bytes, over an
5257 * average of 10 runs:
5259 * (We perturb the cache size by a small (0..4k)
5260 * value to compensate size/alignment related artifacts.
5261 * We also subtract the cost of the operation done on
5267 * dry run, to make sure we start off cache-cold on cpu1,
5268 * and to get any vmalloc pagefaults in advance:
5270 measure_one(cache
, size
, cpu1
, cpu2
);
5271 for (i
= 0; i
< ITERATIONS
; i
++)
5272 cost1
+= measure_one(cache
, size
- i
*1024, cpu1
, cpu2
);
5274 measure_one(cache
, size
, cpu2
, cpu1
);
5275 for (i
= 0; i
< ITERATIONS
; i
++)
5276 cost1
+= measure_one(cache
, size
- i
*1024, cpu2
, cpu1
);
5279 * (We measure the non-migrating [cached] cost on both
5280 * cpu1 and cpu2, to handle CPUs with different speeds)
5284 measure_one(cache
, size
, cpu1
, cpu1
);
5285 for (i
= 0; i
< ITERATIONS
; i
++)
5286 cost2
+= measure_one(cache
, size
- i
*1024, cpu1
, cpu1
);
5288 measure_one(cache
, size
, cpu2
, cpu2
);
5289 for (i
= 0; i
< ITERATIONS
; i
++)
5290 cost2
+= measure_one(cache
, size
- i
*1024, cpu2
, cpu2
);
5293 * Get the per-iteration migration cost:
5295 do_div(cost1
, 2*ITERATIONS
);
5296 do_div(cost2
, 2*ITERATIONS
);
5298 return cost1
- cost2
;
5301 static unsigned long long measure_migration_cost(int cpu1
, int cpu2
)
5303 unsigned long long max_cost
= 0, fluct
= 0, avg_fluct
= 0;
5304 unsigned int max_size
, size
, size_found
= 0;
5305 long long cost
= 0, prev_cost
;
5309 * Search from max_cache_size*5 down to 64K - the real relevant
5310 * cachesize has to lie somewhere inbetween.
5312 if (max_cache_size
) {
5313 max_size
= max(max_cache_size
* SEARCH_SCOPE
, MIN_CACHE_SIZE
);
5314 size
= max(max_cache_size
/ SEARCH_SCOPE
, MIN_CACHE_SIZE
);
5317 * Since we have no estimation about the relevant
5320 max_size
= DEFAULT_CACHE_SIZE
* SEARCH_SCOPE
;
5321 size
= MIN_CACHE_SIZE
;
5324 if (!cpu_online(cpu1
) || !cpu_online(cpu2
)) {
5325 printk("cpu %d and %d not both online!\n", cpu1
, cpu2
);
5330 * Allocate the working set:
5332 cache
= vmalloc(max_size
);
5334 printk("could not vmalloc %d bytes for cache!\n", 2*max_size
);
5335 return 1000000; // return 1 msec on very small boxen
5338 while (size
<= max_size
) {
5340 cost
= measure_cost(cpu1
, cpu2
, cache
, size
);
5346 if (max_cost
< cost
) {
5352 * Calculate average fluctuation, we use this to prevent
5353 * noise from triggering an early break out of the loop:
5355 fluct
= abs(cost
- prev_cost
);
5356 avg_fluct
= (avg_fluct
+ fluct
)/2;
5358 if (migration_debug
)
5359 printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): (%8Ld %8Ld)\n",
5361 (long)cost
/ 1000000,
5362 ((long)cost
/ 100000) % 10,
5363 (long)max_cost
/ 1000000,
5364 ((long)max_cost
/ 100000) % 10,
5365 domain_distance(cpu1
, cpu2
),
5369 * If we iterated at least 20% past the previous maximum,
5370 * and the cost has dropped by more than 20% already,
5371 * (taking fluctuations into account) then we assume to
5372 * have found the maximum and break out of the loop early:
5374 if (size_found
&& (size
*100 > size_found
*SIZE_THRESH
))
5375 if (cost
+avg_fluct
<= 0 ||
5376 max_cost
*100 > (cost
+avg_fluct
)*COST_THRESH
) {
5378 if (migration_debug
)
5379 printk("-> found max.\n");
5383 * Increase the cachesize in 10% steps:
5385 size
= size
* 10 / 9;
5388 if (migration_debug
)
5389 printk("[%d][%d] working set size found: %d, cost: %Ld\n",
5390 cpu1
, cpu2
, size_found
, max_cost
);
5395 * A task is considered 'cache cold' if at least 2 times
5396 * the worst-case cost of migration has passed.
5398 * (this limit is only listened to if the load-balancing
5399 * situation is 'nice' - if there is a large imbalance we
5400 * ignore it for the sake of CPU utilization and
5401 * processing fairness.)
5403 return 2 * max_cost
* migration_factor
/ MIGRATION_FACTOR_SCALE
;
5406 static void calibrate_migration_costs(const cpumask_t
*cpu_map
)
5408 int cpu1
= -1, cpu2
= -1, cpu
, orig_cpu
= raw_smp_processor_id();
5409 unsigned long j0
, j1
, distance
, max_distance
= 0;
5410 struct sched_domain
*sd
;
5415 * First pass - calculate the cacheflush times:
5417 for_each_cpu_mask(cpu1
, *cpu_map
) {
5418 for_each_cpu_mask(cpu2
, *cpu_map
) {
5421 distance
= domain_distance(cpu1
, cpu2
);
5422 max_distance
= max(max_distance
, distance
);
5424 * No result cached yet?
5426 if (migration_cost
[distance
] == -1LL)
5427 migration_cost
[distance
] =
5428 measure_migration_cost(cpu1
, cpu2
);
5432 * Second pass - update the sched domain hierarchy with
5433 * the new cache-hot-time estimations:
5435 for_each_cpu_mask(cpu
, *cpu_map
) {
5437 for_each_domain(cpu
, sd
) {
5438 sd
->cache_hot_time
= migration_cost
[distance
];
5445 if (migration_debug
)
5446 printk("migration: max_cache_size: %d, cpu: %d MHz:\n",
5454 if (system_state
== SYSTEM_BOOTING
) {
5455 printk("migration_cost=");
5456 for (distance
= 0; distance
<= max_distance
; distance
++) {
5459 printk("%ld", (long)migration_cost
[distance
] / 1000);
5464 if (migration_debug
)
5465 printk("migration: %ld seconds\n", (j1
-j0
)/HZ
);
5468 * Move back to the original CPU. NUMA-Q gets confused
5469 * if we migrate to another quad during bootup.
5471 if (raw_smp_processor_id() != orig_cpu
) {
5472 cpumask_t mask
= cpumask_of_cpu(orig_cpu
),
5473 saved_mask
= current
->cpus_allowed
;
5475 set_cpus_allowed(current
, mask
);
5476 set_cpus_allowed(current
, saved_mask
);
5483 * find_next_best_node - find the next node to include in a sched_domain
5484 * @node: node whose sched_domain we're building
5485 * @used_nodes: nodes already in the sched_domain
5487 * Find the next node to include in a given scheduling domain. Simply
5488 * finds the closest node not already in the @used_nodes map.
5490 * Should use nodemask_t.
5492 static int find_next_best_node(int node
, unsigned long *used_nodes
)
5494 int i
, n
, val
, min_val
, best_node
= 0;
5498 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5499 /* Start at @node */
5500 n
= (node
+ i
) % MAX_NUMNODES
;
5502 if (!nr_cpus_node(n
))
5505 /* Skip already used nodes */
5506 if (test_bit(n
, used_nodes
))
5509 /* Simple min distance search */
5510 val
= node_distance(node
, n
);
5512 if (val
< min_val
) {
5518 set_bit(best_node
, used_nodes
);
5523 * sched_domain_node_span - get a cpumask for a node's sched_domain
5524 * @node: node whose cpumask we're constructing
5525 * @size: number of nodes to include in this span
5527 * Given a node, construct a good cpumask for its sched_domain to span. It
5528 * should be one that prevents unnecessary balancing, but also spreads tasks
5531 static cpumask_t
sched_domain_node_span(int node
)
5534 cpumask_t span
, nodemask
;
5535 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
5538 bitmap_zero(used_nodes
, MAX_NUMNODES
);
5540 nodemask
= node_to_cpumask(node
);
5541 cpus_or(span
, span
, nodemask
);
5542 set_bit(node
, used_nodes
);
5544 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5545 int next_node
= find_next_best_node(node
, used_nodes
);
5546 nodemask
= node_to_cpumask(next_node
);
5547 cpus_or(span
, span
, nodemask
);
5555 * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we
5556 * can switch it on easily if needed.
5558 #ifdef CONFIG_SCHED_SMT
5559 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
5560 static struct sched_group sched_group_cpus
[NR_CPUS
];
5561 static int cpu_to_cpu_group(int cpu
)
5567 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
5568 static struct sched_group sched_group_phys
[NR_CPUS
];
5569 static int cpu_to_phys_group(int cpu
)
5571 #ifdef CONFIG_SCHED_SMT
5572 return first_cpu(cpu_sibling_map
[cpu
]);
5580 * The init_sched_build_groups can't handle what we want to do with node
5581 * groups, so roll our own. Now each node has its own list of groups which
5582 * gets dynamically allocated.
5584 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
5585 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
5587 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
5588 static struct sched_group
*sched_group_allnodes_bycpu
[NR_CPUS
];
5590 static int cpu_to_allnodes_group(int cpu
)
5592 return cpu_to_node(cpu
);
5597 * Build sched domains for a given set of cpus and attach the sched domains
5598 * to the individual cpus
5600 void build_sched_domains(const cpumask_t
*cpu_map
)
5604 struct sched_group
**sched_group_nodes
= NULL
;
5605 struct sched_group
*sched_group_allnodes
= NULL
;
5608 * Allocate the per-node list of sched groups
5610 sched_group_nodes
= kmalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
5612 if (!sched_group_nodes
) {
5613 printk(KERN_WARNING
"Can not alloc sched group node list\n");
5616 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
5620 * Set up domains for cpus specified by the cpu_map.
5622 for_each_cpu_mask(i
, *cpu_map
) {
5624 struct sched_domain
*sd
= NULL
, *p
;
5625 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
5627 cpus_and(nodemask
, nodemask
, *cpu_map
);
5630 if (cpus_weight(*cpu_map
)
5631 > SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
5632 if (!sched_group_allnodes
) {
5633 sched_group_allnodes
5634 = kmalloc(sizeof(struct sched_group
)
5637 if (!sched_group_allnodes
) {
5639 "Can not alloc allnodes sched group\n");
5642 sched_group_allnodes_bycpu
[i
]
5643 = sched_group_allnodes
;
5645 sd
= &per_cpu(allnodes_domains
, i
);
5646 *sd
= SD_ALLNODES_INIT
;
5647 sd
->span
= *cpu_map
;
5648 group
= cpu_to_allnodes_group(i
);
5649 sd
->groups
= &sched_group_allnodes
[group
];
5654 sd
= &per_cpu(node_domains
, i
);
5656 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
5658 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
5662 sd
= &per_cpu(phys_domains
, i
);
5663 group
= cpu_to_phys_group(i
);
5665 sd
->span
= nodemask
;
5667 sd
->groups
= &sched_group_phys
[group
];
5669 #ifdef CONFIG_SCHED_SMT
5671 sd
= &per_cpu(cpu_domains
, i
);
5672 group
= cpu_to_cpu_group(i
);
5673 *sd
= SD_SIBLING_INIT
;
5674 sd
->span
= cpu_sibling_map
[i
];
5675 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
5677 sd
->groups
= &sched_group_cpus
[group
];
5681 #ifdef CONFIG_SCHED_SMT
5682 /* Set up CPU (sibling) groups */
5683 for_each_cpu_mask(i
, *cpu_map
) {
5684 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
5685 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
5686 if (i
!= first_cpu(this_sibling_map
))
5689 init_sched_build_groups(sched_group_cpus
, this_sibling_map
,
5694 /* Set up physical groups */
5695 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5696 cpumask_t nodemask
= node_to_cpumask(i
);
5698 cpus_and(nodemask
, nodemask
, *cpu_map
);
5699 if (cpus_empty(nodemask
))
5702 init_sched_build_groups(sched_group_phys
, nodemask
,
5703 &cpu_to_phys_group
);
5707 /* Set up node groups */
5708 if (sched_group_allnodes
)
5709 init_sched_build_groups(sched_group_allnodes
, *cpu_map
,
5710 &cpu_to_allnodes_group
);
5712 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5713 /* Set up node groups */
5714 struct sched_group
*sg
, *prev
;
5715 cpumask_t nodemask
= node_to_cpumask(i
);
5716 cpumask_t domainspan
;
5717 cpumask_t covered
= CPU_MASK_NONE
;
5720 cpus_and(nodemask
, nodemask
, *cpu_map
);
5721 if (cpus_empty(nodemask
)) {
5722 sched_group_nodes
[i
] = NULL
;
5726 domainspan
= sched_domain_node_span(i
);
5727 cpus_and(domainspan
, domainspan
, *cpu_map
);
5729 sg
= kmalloc(sizeof(struct sched_group
), GFP_KERNEL
);
5730 sched_group_nodes
[i
] = sg
;
5731 for_each_cpu_mask(j
, nodemask
) {
5732 struct sched_domain
*sd
;
5733 sd
= &per_cpu(node_domains
, j
);
5735 if (sd
->groups
== NULL
) {
5736 /* Turn off balancing if we have no groups */
5742 "Can not alloc domain group for node %d\n", i
);
5746 sg
->cpumask
= nodemask
;
5747 cpus_or(covered
, covered
, nodemask
);
5750 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
5751 cpumask_t tmp
, notcovered
;
5752 int n
= (i
+ j
) % MAX_NUMNODES
;
5754 cpus_complement(notcovered
, covered
);
5755 cpus_and(tmp
, notcovered
, *cpu_map
);
5756 cpus_and(tmp
, tmp
, domainspan
);
5757 if (cpus_empty(tmp
))
5760 nodemask
= node_to_cpumask(n
);
5761 cpus_and(tmp
, tmp
, nodemask
);
5762 if (cpus_empty(tmp
))
5765 sg
= kmalloc(sizeof(struct sched_group
), GFP_KERNEL
);
5768 "Can not alloc domain group for node %d\n", j
);
5773 cpus_or(covered
, covered
, tmp
);
5777 prev
->next
= sched_group_nodes
[i
];
5781 /* Calculate CPU power for physical packages and nodes */
5782 for_each_cpu_mask(i
, *cpu_map
) {
5784 struct sched_domain
*sd
;
5785 #ifdef CONFIG_SCHED_SMT
5786 sd
= &per_cpu(cpu_domains
, i
);
5787 power
= SCHED_LOAD_SCALE
;
5788 sd
->groups
->cpu_power
= power
;
5791 sd
= &per_cpu(phys_domains
, i
);
5792 power
= SCHED_LOAD_SCALE
+ SCHED_LOAD_SCALE
*
5793 (cpus_weight(sd
->groups
->cpumask
)-1) / 10;
5794 sd
->groups
->cpu_power
= power
;
5797 sd
= &per_cpu(allnodes_domains
, i
);
5799 power
= SCHED_LOAD_SCALE
+ SCHED_LOAD_SCALE
*
5800 (cpus_weight(sd
->groups
->cpumask
)-1) / 10;
5801 sd
->groups
->cpu_power
= power
;
5807 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5808 struct sched_group
*sg
= sched_group_nodes
[i
];
5814 for_each_cpu_mask(j
, sg
->cpumask
) {
5815 struct sched_domain
*sd
;
5818 sd
= &per_cpu(phys_domains
, j
);
5819 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
5821 * Only add "power" once for each
5826 power
= SCHED_LOAD_SCALE
+ SCHED_LOAD_SCALE
*
5827 (cpus_weight(sd
->groups
->cpumask
)-1) / 10;
5829 sg
->cpu_power
+= power
;
5832 if (sg
!= sched_group_nodes
[i
])
5837 /* Attach the domains */
5838 for_each_cpu_mask(i
, *cpu_map
) {
5839 struct sched_domain
*sd
;
5840 #ifdef CONFIG_SCHED_SMT
5841 sd
= &per_cpu(cpu_domains
, i
);
5843 sd
= &per_cpu(phys_domains
, i
);
5845 cpu_attach_domain(sd
, i
);
5848 * Tune cache-hot values:
5850 calibrate_migration_costs(cpu_map
);
5853 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
5855 static void arch_init_sched_domains(const cpumask_t
*cpu_map
)
5857 cpumask_t cpu_default_map
;
5860 * Setup mask for cpus without special case scheduling requirements.
5861 * For now this just excludes isolated cpus, but could be used to
5862 * exclude other special cases in the future.
5864 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
5866 build_sched_domains(&cpu_default_map
);
5869 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
5875 for_each_cpu_mask(cpu
, *cpu_map
) {
5876 struct sched_group
*sched_group_allnodes
5877 = sched_group_allnodes_bycpu
[cpu
];
5878 struct sched_group
**sched_group_nodes
5879 = sched_group_nodes_bycpu
[cpu
];
5881 if (sched_group_allnodes
) {
5882 kfree(sched_group_allnodes
);
5883 sched_group_allnodes_bycpu
[cpu
] = NULL
;
5886 if (!sched_group_nodes
)
5889 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5890 cpumask_t nodemask
= node_to_cpumask(i
);
5891 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
5893 cpus_and(nodemask
, nodemask
, *cpu_map
);
5894 if (cpus_empty(nodemask
))
5904 if (oldsg
!= sched_group_nodes
[i
])
5907 kfree(sched_group_nodes
);
5908 sched_group_nodes_bycpu
[cpu
] = NULL
;
5914 * Detach sched domains from a group of cpus specified in cpu_map
5915 * These cpus will now be attached to the NULL domain
5917 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
5921 for_each_cpu_mask(i
, *cpu_map
)
5922 cpu_attach_domain(NULL
, i
);
5923 synchronize_sched();
5924 arch_destroy_sched_domains(cpu_map
);
5928 * Partition sched domains as specified by the cpumasks below.
5929 * This attaches all cpus from the cpumasks to the NULL domain,
5930 * waits for a RCU quiescent period, recalculates sched
5931 * domain information and then attaches them back to the
5932 * correct sched domains
5933 * Call with hotplug lock held
5935 void partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
5937 cpumask_t change_map
;
5939 cpus_and(*partition1
, *partition1
, cpu_online_map
);
5940 cpus_and(*partition2
, *partition2
, cpu_online_map
);
5941 cpus_or(change_map
, *partition1
, *partition2
);
5943 /* Detach sched domains from all of the affected cpus */
5944 detach_destroy_domains(&change_map
);
5945 if (!cpus_empty(*partition1
))
5946 build_sched_domains(partition1
);
5947 if (!cpus_empty(*partition2
))
5948 build_sched_domains(partition2
);
5951 #ifdef CONFIG_HOTPLUG_CPU
5953 * Force a reinitialization of the sched domains hierarchy. The domains
5954 * and groups cannot be updated in place without racing with the balancing
5955 * code, so we temporarily attach all running cpus to the NULL domain
5956 * which will prevent rebalancing while the sched domains are recalculated.
5958 static int update_sched_domains(struct notifier_block
*nfb
,
5959 unsigned long action
, void *hcpu
)
5962 case CPU_UP_PREPARE
:
5963 case CPU_DOWN_PREPARE
:
5964 detach_destroy_domains(&cpu_online_map
);
5967 case CPU_UP_CANCELED
:
5968 case CPU_DOWN_FAILED
:
5972 * Fall through and re-initialise the domains.
5979 /* The hotplug lock is already held by cpu_up/cpu_down */
5980 arch_init_sched_domains(&cpu_online_map
);
5986 void __init
sched_init_smp(void)
5989 arch_init_sched_domains(&cpu_online_map
);
5990 unlock_cpu_hotplug();
5991 /* XXX: Theoretical race here - CPU may be hotplugged now */
5992 hotcpu_notifier(update_sched_domains
, 0);
5995 void __init
sched_init_smp(void)
5998 #endif /* CONFIG_SMP */
6000 int in_sched_functions(unsigned long addr
)
6002 /* Linker adds these: start and end of __sched functions */
6003 extern char __sched_text_start
[], __sched_text_end
[];
6004 return in_lock_functions(addr
) ||
6005 (addr
>= (unsigned long)__sched_text_start
6006 && addr
< (unsigned long)__sched_text_end
);
6009 void __init
sched_init(void)
6015 prio_array_t
*array
;
6018 spin_lock_init(&rq
->lock
);
6020 rq
->active
= rq
->arrays
;
6021 rq
->expired
= rq
->arrays
+ 1;
6022 rq
->best_expired_prio
= MAX_PRIO
;
6026 for (j
= 1; j
< 3; j
++)
6027 rq
->cpu_load
[j
] = 0;
6028 rq
->active_balance
= 0;
6030 rq
->migration_thread
= NULL
;
6031 INIT_LIST_HEAD(&rq
->migration_queue
);
6033 atomic_set(&rq
->nr_iowait
, 0);
6035 for (j
= 0; j
< 2; j
++) {
6036 array
= rq
->arrays
+ j
;
6037 for (k
= 0; k
< MAX_PRIO
; k
++) {
6038 INIT_LIST_HEAD(array
->queue
+ k
);
6039 __clear_bit(k
, array
->bitmap
);
6041 // delimiter for bitsearch
6042 __set_bit(MAX_PRIO
, array
->bitmap
);
6047 * The boot idle thread does lazy MMU switching as well:
6049 atomic_inc(&init_mm
.mm_count
);
6050 enter_lazy_tlb(&init_mm
, current
);
6053 * Make us the idle thread. Technically, schedule() should not be
6054 * called from this thread, however somewhere below it might be,
6055 * but because we are the idle thread, we just pick up running again
6056 * when this runqueue becomes "idle".
6058 init_idle(current
, smp_processor_id());
6061 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6062 void __might_sleep(char *file
, int line
)
6064 #if defined(in_atomic)
6065 static unsigned long prev_jiffy
; /* ratelimiting */
6067 if ((in_atomic() || irqs_disabled()) &&
6068 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6069 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6071 prev_jiffy
= jiffies
;
6072 printk(KERN_ERR
"Debug: sleeping function called from invalid"
6073 " context at %s:%d\n", file
, line
);
6074 printk("in_atomic():%d, irqs_disabled():%d\n",
6075 in_atomic(), irqs_disabled());
6080 EXPORT_SYMBOL(__might_sleep
);
6083 #ifdef CONFIG_MAGIC_SYSRQ
6084 void normalize_rt_tasks(void)
6086 struct task_struct
*p
;
6087 prio_array_t
*array
;
6088 unsigned long flags
;
6091 read_lock_irq(&tasklist_lock
);
6092 for_each_process (p
) {
6096 rq
= task_rq_lock(p
, &flags
);
6100 deactivate_task(p
, task_rq(p
));
6101 __setscheduler(p
, SCHED_NORMAL
, 0);
6103 __activate_task(p
, task_rq(p
));
6104 resched_task(rq
->curr
);
6107 task_rq_unlock(rq
, &flags
);
6109 read_unlock_irq(&tasklist_lock
);
6112 #endif /* CONFIG_MAGIC_SYSRQ */
6116 * These functions are only useful for the IA64 MCA handling.
6118 * They can only be called when the whole system has been
6119 * stopped - every CPU needs to be quiescent, and no scheduling
6120 * activity can take place. Using them for anything else would
6121 * be a serious bug, and as a result, they aren't even visible
6122 * under any other configuration.
6126 * curr_task - return the current task for a given cpu.
6127 * @cpu: the processor in question.
6129 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6131 task_t
*curr_task(int cpu
)
6133 return cpu_curr(cpu
);
6137 * set_curr_task - set the current task for a given cpu.
6138 * @cpu: the processor in question.
6139 * @p: the task pointer to set.
6141 * Description: This function must only be used when non-maskable interrupts
6142 * are serviced on a separate stack. It allows the architecture to switch the
6143 * notion of the current task on a cpu in a non-blocking manner. This function
6144 * must be called with all CPU's synchronized, and interrupts disabled, the
6145 * and caller must save the original value of the current task (see
6146 * curr_task() above) and restore that value before reenabling interrupts and
6147 * re-starting the system.
6149 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6151 void set_curr_task(int cpu
, task_t
*p
)