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>
52 #include <linux/kprobes.h>
55 #include <asm/unistd.h>
58 * Convert user-nice values [ -20 ... 0 ... 19 ]
59 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
62 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
63 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
64 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
67 * 'User priority' is the nice value converted to something we
68 * can work with better when scaling various scheduler parameters,
69 * it's a [ 0 ... 39 ] range.
71 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
72 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
73 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
76 * Some helpers for converting nanosecond timing to jiffy resolution
78 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
79 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
82 * These are the 'tuning knobs' of the scheduler:
84 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
85 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
86 * Timeslices get refilled after they expire.
88 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
89 #define DEF_TIMESLICE (100 * HZ / 1000)
90 #define ON_RUNQUEUE_WEIGHT 30
91 #define CHILD_PENALTY 95
92 #define PARENT_PENALTY 100
94 #define PRIO_BONUS_RATIO 25
95 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
96 #define INTERACTIVE_DELTA 2
97 #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
98 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
99 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
102 * If a task is 'interactive' then we reinsert it in the active
103 * array after it has expired its current timeslice. (it will not
104 * continue to run immediately, it will still roundrobin with
105 * other interactive tasks.)
107 * This part scales the interactivity limit depending on niceness.
109 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
110 * Here are a few examples of different nice levels:
112 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
113 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
114 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
115 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
116 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
118 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
119 * priority range a task can explore, a value of '1' means the
120 * task is rated interactive.)
122 * Ie. nice +19 tasks can never get 'interactive' enough to be
123 * reinserted into the active array. And only heavily CPU-hog nice -20
124 * tasks will be expired. Default nice 0 tasks are somewhere between,
125 * it takes some effort for them to get interactive, but it's not
129 #define CURRENT_BONUS(p) \
130 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
133 #define GRANULARITY (10 * HZ / 1000 ? : 1)
136 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
137 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
140 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
141 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
144 #define SCALE(v1,v1_max,v2_max) \
145 (v1) * (v2_max) / (v1_max)
148 (SCALE(TASK_NICE(p) + 20, 40, MAX_BONUS) - 20 * MAX_BONUS / 40 + \
151 #define TASK_INTERACTIVE(p) \
152 ((p)->prio <= (p)->static_prio - DELTA(p))
154 #define INTERACTIVE_SLEEP(p) \
155 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
156 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
158 #define TASK_PREEMPTS_CURR(p, rq) \
159 ((p)->prio < (rq)->curr->prio)
162 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
163 * to time slice values: [800ms ... 100ms ... 5ms]
165 * The higher a thread's priority, the bigger timeslices
166 * it gets during one round of execution. But even the lowest
167 * priority thread gets MIN_TIMESLICE worth of execution time.
170 #define SCALE_PRIO(x, prio) \
171 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO/2), MIN_TIMESLICE)
173 static unsigned int task_timeslice(task_t
*p
)
175 if (p
->static_prio
< NICE_TO_PRIO(0))
176 return SCALE_PRIO(DEF_TIMESLICE
*4, p
->static_prio
);
178 return SCALE_PRIO(DEF_TIMESLICE
, p
->static_prio
);
180 #define task_hot(p, now, sd) ((long long) ((now) - (p)->last_ran) \
181 < (long long) (sd)->cache_hot_time)
184 * These are the runqueue data structures:
187 #define BITMAP_SIZE ((((MAX_PRIO+1+7)/8)+sizeof(long)-1)/sizeof(long))
189 typedef struct runqueue runqueue_t
;
192 unsigned int nr_active
;
193 unsigned long bitmap
[BITMAP_SIZE
];
194 struct list_head queue
[MAX_PRIO
];
198 * This is the main, per-CPU runqueue data structure.
200 * Locking rule: those places that want to lock multiple runqueues
201 * (such as the load balancing or the thread migration code), lock
202 * acquire operations must be ordered by ascending &runqueue.
208 * nr_running and cpu_load should be in the same cacheline because
209 * remote CPUs use both these fields when doing load calculation.
211 unsigned long nr_running
;
213 unsigned long cpu_load
[3];
215 unsigned long long nr_switches
;
218 * This is part of a global counter where only the total sum
219 * over all CPUs matters. A task can increase this counter on
220 * one CPU and if it got migrated afterwards it may decrease
221 * it on another CPU. Always updated under the runqueue lock:
223 unsigned long nr_uninterruptible
;
225 unsigned long expired_timestamp
;
226 unsigned long long timestamp_last_tick
;
228 struct mm_struct
*prev_mm
;
229 prio_array_t
*active
, *expired
, arrays
[2];
230 int best_expired_prio
;
234 struct sched_domain
*sd
;
236 /* For active balancing */
240 task_t
*migration_thread
;
241 struct list_head migration_queue
;
245 #ifdef CONFIG_SCHEDSTATS
247 struct sched_info rq_sched_info
;
249 /* sys_sched_yield() stats */
250 unsigned long yld_exp_empty
;
251 unsigned long yld_act_empty
;
252 unsigned long yld_both_empty
;
253 unsigned long yld_cnt
;
255 /* schedule() stats */
256 unsigned long sched_switch
;
257 unsigned long sched_cnt
;
258 unsigned long sched_goidle
;
260 /* try_to_wake_up() stats */
261 unsigned long ttwu_cnt
;
262 unsigned long ttwu_local
;
266 static DEFINE_PER_CPU(struct runqueue
, runqueues
);
269 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
270 * See detach_destroy_domains: synchronize_sched for details.
272 * The domain tree of any CPU may only be accessed from within
273 * preempt-disabled sections.
275 #define for_each_domain(cpu, domain) \
276 for (domain = rcu_dereference(cpu_rq(cpu)->sd); domain; domain = domain->parent)
278 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
279 #define this_rq() (&__get_cpu_var(runqueues))
280 #define task_rq(p) cpu_rq(task_cpu(p))
281 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
283 #ifndef prepare_arch_switch
284 # define prepare_arch_switch(next) do { } while (0)
286 #ifndef finish_arch_switch
287 # define finish_arch_switch(prev) do { } while (0)
290 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
291 static inline int task_running(runqueue_t
*rq
, task_t
*p
)
293 return rq
->curr
== p
;
296 static inline void prepare_lock_switch(runqueue_t
*rq
, task_t
*next
)
300 static inline void finish_lock_switch(runqueue_t
*rq
, task_t
*prev
)
302 #ifdef CONFIG_DEBUG_SPINLOCK
303 /* this is a valid case when another task releases the spinlock */
304 rq
->lock
.owner
= current
;
306 spin_unlock_irq(&rq
->lock
);
309 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
310 static inline int task_running(runqueue_t
*rq
, task_t
*p
)
315 return rq
->curr
== p
;
319 static inline void prepare_lock_switch(runqueue_t
*rq
, task_t
*next
)
323 * We can optimise this out completely for !SMP, because the
324 * SMP rebalancing from interrupt is the only thing that cares
329 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
330 spin_unlock_irq(&rq
->lock
);
332 spin_unlock(&rq
->lock
);
336 static inline void finish_lock_switch(runqueue_t
*rq
, task_t
*prev
)
340 * After ->oncpu is cleared, the task can be moved to a different CPU.
341 * We must ensure this doesn't happen until the switch is completely
347 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
351 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
354 * task_rq_lock - lock the runqueue a given task resides on and disable
355 * interrupts. Note the ordering: we can safely lookup the task_rq without
356 * explicitly disabling preemption.
358 static inline runqueue_t
*task_rq_lock(task_t
*p
, unsigned long *flags
)
364 local_irq_save(*flags
);
366 spin_lock(&rq
->lock
);
367 if (unlikely(rq
!= task_rq(p
))) {
368 spin_unlock_irqrestore(&rq
->lock
, *flags
);
369 goto repeat_lock_task
;
374 static inline void task_rq_unlock(runqueue_t
*rq
, unsigned long *flags
)
377 spin_unlock_irqrestore(&rq
->lock
, *flags
);
380 #ifdef CONFIG_SCHEDSTATS
382 * bump this up when changing the output format or the meaning of an existing
383 * format, so that tools can adapt (or abort)
385 #define SCHEDSTAT_VERSION 12
387 static int show_schedstat(struct seq_file
*seq
, void *v
)
391 seq_printf(seq
, "version %d\n", SCHEDSTAT_VERSION
);
392 seq_printf(seq
, "timestamp %lu\n", jiffies
);
393 for_each_online_cpu(cpu
) {
394 runqueue_t
*rq
= cpu_rq(cpu
);
396 struct sched_domain
*sd
;
400 /* runqueue-specific stats */
402 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
403 cpu
, rq
->yld_both_empty
,
404 rq
->yld_act_empty
, rq
->yld_exp_empty
, rq
->yld_cnt
,
405 rq
->sched_switch
, rq
->sched_cnt
, rq
->sched_goidle
,
406 rq
->ttwu_cnt
, rq
->ttwu_local
,
407 rq
->rq_sched_info
.cpu_time
,
408 rq
->rq_sched_info
.run_delay
, rq
->rq_sched_info
.pcnt
);
410 seq_printf(seq
, "\n");
413 /* domain-specific stats */
415 for_each_domain(cpu
, sd
) {
416 enum idle_type itype
;
417 char mask_str
[NR_CPUS
];
419 cpumask_scnprintf(mask_str
, NR_CPUS
, sd
->span
);
420 seq_printf(seq
, "domain%d %s", dcnt
++, mask_str
);
421 for (itype
= SCHED_IDLE
; itype
< MAX_IDLE_TYPES
;
423 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu %lu",
425 sd
->lb_balanced
[itype
],
426 sd
->lb_failed
[itype
],
427 sd
->lb_imbalance
[itype
],
428 sd
->lb_gained
[itype
],
429 sd
->lb_hot_gained
[itype
],
430 sd
->lb_nobusyq
[itype
],
431 sd
->lb_nobusyg
[itype
]);
433 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
434 sd
->alb_cnt
, sd
->alb_failed
, sd
->alb_pushed
,
435 sd
->sbe_cnt
, sd
->sbe_balanced
, sd
->sbe_pushed
,
436 sd
->sbf_cnt
, sd
->sbf_balanced
, sd
->sbf_pushed
,
437 sd
->ttwu_wake_remote
, sd
->ttwu_move_affine
, sd
->ttwu_move_balance
);
445 static int schedstat_open(struct inode
*inode
, struct file
*file
)
447 unsigned int size
= PAGE_SIZE
* (1 + num_online_cpus() / 32);
448 char *buf
= kmalloc(size
, GFP_KERNEL
);
454 res
= single_open(file
, show_schedstat
, NULL
);
456 m
= file
->private_data
;
464 struct file_operations proc_schedstat_operations
= {
465 .open
= schedstat_open
,
468 .release
= single_release
,
471 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
472 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
473 #else /* !CONFIG_SCHEDSTATS */
474 # define schedstat_inc(rq, field) do { } while (0)
475 # define schedstat_add(rq, field, amt) do { } while (0)
479 * rq_lock - lock a given runqueue and disable interrupts.
481 static inline runqueue_t
*this_rq_lock(void)
488 spin_lock(&rq
->lock
);
493 #ifdef CONFIG_SCHEDSTATS
495 * Called when a process is dequeued from the active array and given
496 * the cpu. We should note that with the exception of interactive
497 * tasks, the expired queue will become the active queue after the active
498 * queue is empty, without explicitly dequeuing and requeuing tasks in the
499 * expired queue. (Interactive tasks may be requeued directly to the
500 * active queue, thus delaying tasks in the expired queue from running;
501 * see scheduler_tick()).
503 * This function is only called from sched_info_arrive(), rather than
504 * dequeue_task(). Even though a task may be queued and dequeued multiple
505 * times as it is shuffled about, we're really interested in knowing how
506 * long it was from the *first* time it was queued to the time that it
509 static inline void sched_info_dequeued(task_t
*t
)
511 t
->sched_info
.last_queued
= 0;
515 * Called when a task finally hits the cpu. We can now calculate how
516 * long it was waiting to run. We also note when it began so that we
517 * can keep stats on how long its timeslice is.
519 static void sched_info_arrive(task_t
*t
)
521 unsigned long now
= jiffies
, diff
= 0;
522 struct runqueue
*rq
= task_rq(t
);
524 if (t
->sched_info
.last_queued
)
525 diff
= now
- t
->sched_info
.last_queued
;
526 sched_info_dequeued(t
);
527 t
->sched_info
.run_delay
+= diff
;
528 t
->sched_info
.last_arrival
= now
;
529 t
->sched_info
.pcnt
++;
534 rq
->rq_sched_info
.run_delay
+= diff
;
535 rq
->rq_sched_info
.pcnt
++;
539 * Called when a process is queued into either the active or expired
540 * array. The time is noted and later used to determine how long we
541 * had to wait for us to reach the cpu. Since the expired queue will
542 * become the active queue after active queue is empty, without dequeuing
543 * and requeuing any tasks, we are interested in queuing to either. It
544 * is unusual but not impossible for tasks to be dequeued and immediately
545 * requeued in the same or another array: this can happen in sched_yield(),
546 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
549 * This function is only called from enqueue_task(), but also only updates
550 * the timestamp if it is already not set. It's assumed that
551 * sched_info_dequeued() will clear that stamp when appropriate.
553 static inline void sched_info_queued(task_t
*t
)
555 if (!t
->sched_info
.last_queued
)
556 t
->sched_info
.last_queued
= jiffies
;
560 * Called when a process ceases being the active-running process, either
561 * voluntarily or involuntarily. Now we can calculate how long we ran.
563 static inline void sched_info_depart(task_t
*t
)
565 struct runqueue
*rq
= task_rq(t
);
566 unsigned long diff
= jiffies
- t
->sched_info
.last_arrival
;
568 t
->sched_info
.cpu_time
+= diff
;
571 rq
->rq_sched_info
.cpu_time
+= diff
;
575 * Called when tasks are switched involuntarily due, typically, to expiring
576 * their time slice. (This may also be called when switching to or from
577 * the idle task.) We are only called when prev != next.
579 static inline void sched_info_switch(task_t
*prev
, task_t
*next
)
581 struct runqueue
*rq
= task_rq(prev
);
584 * prev now departs the cpu. It's not interesting to record
585 * stats about how efficient we were at scheduling the idle
588 if (prev
!= rq
->idle
)
589 sched_info_depart(prev
);
591 if (next
!= rq
->idle
)
592 sched_info_arrive(next
);
595 #define sched_info_queued(t) do { } while (0)
596 #define sched_info_switch(t, next) do { } while (0)
597 #endif /* CONFIG_SCHEDSTATS */
600 * Adding/removing a task to/from a priority array:
602 static void dequeue_task(struct task_struct
*p
, prio_array_t
*array
)
605 list_del(&p
->run_list
);
606 if (list_empty(array
->queue
+ p
->prio
))
607 __clear_bit(p
->prio
, array
->bitmap
);
610 static void enqueue_task(struct task_struct
*p
, prio_array_t
*array
)
612 sched_info_queued(p
);
613 list_add_tail(&p
->run_list
, array
->queue
+ p
->prio
);
614 __set_bit(p
->prio
, array
->bitmap
);
620 * Put task to the end of the run list without the overhead of dequeue
621 * followed by enqueue.
623 static void requeue_task(struct task_struct
*p
, prio_array_t
*array
)
625 list_move_tail(&p
->run_list
, array
->queue
+ p
->prio
);
628 static inline void enqueue_task_head(struct task_struct
*p
, prio_array_t
*array
)
630 list_add(&p
->run_list
, array
->queue
+ p
->prio
);
631 __set_bit(p
->prio
, array
->bitmap
);
637 * effective_prio - return the priority that is based on the static
638 * priority but is modified by bonuses/penalties.
640 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
641 * into the -5 ... 0 ... +5 bonus/penalty range.
643 * We use 25% of the full 0...39 priority range so that:
645 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
646 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
648 * Both properties are important to certain workloads.
650 static int effective_prio(task_t
*p
)
657 bonus
= CURRENT_BONUS(p
) - MAX_BONUS
/ 2;
659 prio
= p
->static_prio
- bonus
;
660 if (prio
< MAX_RT_PRIO
)
662 if (prio
> MAX_PRIO
-1)
668 * __activate_task - move a task to the runqueue.
670 static inline void __activate_task(task_t
*p
, runqueue_t
*rq
)
672 enqueue_task(p
, rq
->active
);
677 * __activate_idle_task - move idle task to the _front_ of runqueue.
679 static inline void __activate_idle_task(task_t
*p
, runqueue_t
*rq
)
681 enqueue_task_head(p
, rq
->active
);
685 static int recalc_task_prio(task_t
*p
, unsigned long long now
)
687 /* Caller must always ensure 'now >= p->timestamp' */
688 unsigned long long __sleep_time
= now
- p
->timestamp
;
689 unsigned long sleep_time
;
691 if (unlikely(p
->policy
== SCHED_BATCH
))
694 if (__sleep_time
> NS_MAX_SLEEP_AVG
)
695 sleep_time
= NS_MAX_SLEEP_AVG
;
697 sleep_time
= (unsigned long)__sleep_time
;
700 if (likely(sleep_time
> 0)) {
702 * User tasks that sleep a long time are categorised as
703 * idle. They will only have their sleep_avg increased to a
704 * level that makes them just interactive priority to stay
705 * active yet prevent them suddenly becoming cpu hogs and
706 * starving other processes.
708 if (p
->mm
&& p
->sleep_type
!= SLEEP_NONINTERACTIVE
&&
709 sleep_time
> INTERACTIVE_SLEEP(p
)) {
710 unsigned long ceiling
;
712 ceiling
= JIFFIES_TO_NS(MAX_SLEEP_AVG
-
714 if (p
->sleep_avg
< ceiling
)
715 p
->sleep_avg
= ceiling
;
718 * Tasks waking from uninterruptible sleep are
719 * limited in their sleep_avg rise as they
720 * are likely to be waiting on I/O
722 if (p
->sleep_type
== SLEEP_NONINTERACTIVE
&& p
->mm
) {
723 if (p
->sleep_avg
>= INTERACTIVE_SLEEP(p
))
725 else if (p
->sleep_avg
+ sleep_time
>=
726 INTERACTIVE_SLEEP(p
)) {
727 p
->sleep_avg
= INTERACTIVE_SLEEP(p
);
733 * This code gives a bonus to interactive tasks.
735 * The boost works by updating the 'average sleep time'
736 * value here, based on ->timestamp. The more time a
737 * task spends sleeping, the higher the average gets -
738 * and the higher the priority boost gets as well.
740 p
->sleep_avg
+= sleep_time
;
742 if (p
->sleep_avg
> NS_MAX_SLEEP_AVG
)
743 p
->sleep_avg
= NS_MAX_SLEEP_AVG
;
747 return effective_prio(p
);
751 * activate_task - move a task to the runqueue and do priority recalculation
753 * Update all the scheduling statistics stuff. (sleep average
754 * calculation, priority modifiers, etc.)
756 static void activate_task(task_t
*p
, runqueue_t
*rq
, int local
)
758 unsigned long long now
;
763 /* Compensate for drifting sched_clock */
764 runqueue_t
*this_rq
= this_rq();
765 now
= (now
- this_rq
->timestamp_last_tick
)
766 + rq
->timestamp_last_tick
;
771 p
->prio
= recalc_task_prio(p
, now
);
774 * This checks to make sure it's not an uninterruptible task
775 * that is now waking up.
777 if (p
->sleep_type
== SLEEP_NORMAL
) {
779 * Tasks which were woken up by interrupts (ie. hw events)
780 * are most likely of interactive nature. So we give them
781 * the credit of extending their sleep time to the period
782 * of time they spend on the runqueue, waiting for execution
783 * on a CPU, first time around:
786 p
->sleep_type
= SLEEP_INTERRUPTED
;
789 * Normal first-time wakeups get a credit too for
790 * on-runqueue time, but it will be weighted down:
792 p
->sleep_type
= SLEEP_INTERACTIVE
;
797 __activate_task(p
, rq
);
801 * deactivate_task - remove a task from the runqueue.
803 static void deactivate_task(struct task_struct
*p
, runqueue_t
*rq
)
806 dequeue_task(p
, p
->array
);
811 * resched_task - mark a task 'to be rescheduled now'.
813 * On UP this means the setting of the need_resched flag, on SMP it
814 * might also involve a cross-CPU call to trigger the scheduler on
818 static void resched_task(task_t
*p
)
822 assert_spin_locked(&task_rq(p
)->lock
);
824 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
827 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
830 if (cpu
== smp_processor_id())
833 /* NEED_RESCHED must be visible before we test POLLING_NRFLAG */
835 if (!test_tsk_thread_flag(p
, TIF_POLLING_NRFLAG
))
836 smp_send_reschedule(cpu
);
839 static inline void resched_task(task_t
*p
)
841 assert_spin_locked(&task_rq(p
)->lock
);
842 set_tsk_need_resched(p
);
847 * task_curr - is this task currently executing on a CPU?
848 * @p: the task in question.
850 inline int task_curr(const task_t
*p
)
852 return cpu_curr(task_cpu(p
)) == p
;
857 struct list_head list
;
862 struct completion done
;
866 * The task's runqueue lock must be held.
867 * Returns true if you have to wait for migration thread.
869 static int migrate_task(task_t
*p
, int dest_cpu
, migration_req_t
*req
)
871 runqueue_t
*rq
= task_rq(p
);
874 * If the task is not on a runqueue (and not running), then
875 * it is sufficient to simply update the task's cpu field.
877 if (!p
->array
&& !task_running(rq
, p
)) {
878 set_task_cpu(p
, dest_cpu
);
882 init_completion(&req
->done
);
884 req
->dest_cpu
= dest_cpu
;
885 list_add(&req
->list
, &rq
->migration_queue
);
890 * wait_task_inactive - wait for a thread to unschedule.
892 * The caller must ensure that the task *will* unschedule sometime soon,
893 * else this function might spin for a *long* time. This function can't
894 * be called with interrupts off, or it may introduce deadlock with
895 * smp_call_function() if an IPI is sent by the same process we are
896 * waiting to become inactive.
898 void wait_task_inactive(task_t
*p
)
905 rq
= task_rq_lock(p
, &flags
);
906 /* Must be off runqueue entirely, not preempted. */
907 if (unlikely(p
->array
|| task_running(rq
, p
))) {
908 /* If it's preempted, we yield. It could be a while. */
909 preempted
= !task_running(rq
, p
);
910 task_rq_unlock(rq
, &flags
);
916 task_rq_unlock(rq
, &flags
);
920 * kick_process - kick a running thread to enter/exit the kernel
921 * @p: the to-be-kicked thread
923 * Cause a process which is running on another CPU to enter
924 * kernel-mode, without any delay. (to get signals handled.)
926 * NOTE: this function doesnt have to take the runqueue lock,
927 * because all it wants to ensure is that the remote task enters
928 * the kernel. If the IPI races and the task has been migrated
929 * to another CPU then no harm is done and the purpose has been
932 void kick_process(task_t
*p
)
938 if ((cpu
!= smp_processor_id()) && task_curr(p
))
939 smp_send_reschedule(cpu
);
944 * Return a low guess at the load of a migration-source cpu.
946 * We want to under-estimate the load of migration sources, to
947 * balance conservatively.
949 static inline unsigned long source_load(int cpu
, int type
)
951 runqueue_t
*rq
= cpu_rq(cpu
);
952 unsigned long load_now
= rq
->nr_running
* SCHED_LOAD_SCALE
;
956 return min(rq
->cpu_load
[type
-1], load_now
);
960 * Return a high guess at the load of a migration-target cpu
962 static inline unsigned long target_load(int cpu
, int type
)
964 runqueue_t
*rq
= cpu_rq(cpu
);
965 unsigned long load_now
= rq
->nr_running
* SCHED_LOAD_SCALE
;
969 return max(rq
->cpu_load
[type
-1], load_now
);
973 * find_idlest_group finds and returns the least busy CPU group within the
976 static struct sched_group
*
977 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
979 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
980 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
981 int load_idx
= sd
->forkexec_idx
;
982 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
985 unsigned long load
, avg_load
;
989 /* Skip over this group if it has no CPUs allowed */
990 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
993 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
995 /* Tally up the load of all CPUs in the group */
998 for_each_cpu_mask(i
, group
->cpumask
) {
999 /* Bias balancing toward cpus of our domain */
1001 load
= source_load(i
, load_idx
);
1003 load
= target_load(i
, load_idx
);
1008 /* Adjust by relative CPU power of the group */
1009 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
1012 this_load
= avg_load
;
1014 } else if (avg_load
< min_load
) {
1015 min_load
= avg_load
;
1019 group
= group
->next
;
1020 } while (group
!= sd
->groups
);
1022 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1028 * find_idlest_queue - find the idlest runqueue among the cpus in group.
1031 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1034 unsigned long load
, min_load
= ULONG_MAX
;
1038 /* Traverse only the allowed CPUs */
1039 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1041 for_each_cpu_mask(i
, tmp
) {
1042 load
= source_load(i
, 0);
1044 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1054 * sched_balance_self: balance the current task (running on cpu) in domains
1055 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1058 * Balance, ie. select the least loaded group.
1060 * Returns the target CPU number, or the same CPU if no balancing is needed.
1062 * preempt must be disabled.
1064 static int sched_balance_self(int cpu
, int flag
)
1066 struct task_struct
*t
= current
;
1067 struct sched_domain
*tmp
, *sd
= NULL
;
1069 for_each_domain(cpu
, tmp
)
1070 if (tmp
->flags
& flag
)
1075 struct sched_group
*group
;
1080 group
= find_idlest_group(sd
, t
, cpu
);
1084 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1085 if (new_cpu
== -1 || new_cpu
== cpu
)
1088 /* Now try balancing at a lower domain level */
1092 weight
= cpus_weight(span
);
1093 for_each_domain(cpu
, tmp
) {
1094 if (weight
<= cpus_weight(tmp
->span
))
1096 if (tmp
->flags
& flag
)
1099 /* while loop will break here if sd == NULL */
1105 #endif /* CONFIG_SMP */
1108 * wake_idle() will wake a task on an idle cpu if task->cpu is
1109 * not idle and an idle cpu is available. The span of cpus to
1110 * search starts with cpus closest then further out as needed,
1111 * so we always favor a closer, idle cpu.
1113 * Returns the CPU we should wake onto.
1115 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1116 static int wake_idle(int cpu
, task_t
*p
)
1119 struct sched_domain
*sd
;
1125 for_each_domain(cpu
, sd
) {
1126 if (sd
->flags
& SD_WAKE_IDLE
) {
1127 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1128 for_each_cpu_mask(i
, tmp
) {
1139 static inline int wake_idle(int cpu
, task_t
*p
)
1146 * try_to_wake_up - wake up a thread
1147 * @p: the to-be-woken-up thread
1148 * @state: the mask of task states that can be woken
1149 * @sync: do a synchronous wakeup?
1151 * Put it on the run-queue if it's not already there. The "current"
1152 * thread is always on the run-queue (except when the actual
1153 * re-schedule is in progress), and as such you're allowed to do
1154 * the simpler "current->state = TASK_RUNNING" to mark yourself
1155 * runnable without the overhead of this.
1157 * returns failure only if the task is already active.
1159 static int try_to_wake_up(task_t
*p
, unsigned int state
, int sync
)
1161 int cpu
, this_cpu
, success
= 0;
1162 unsigned long flags
;
1166 unsigned long load
, this_load
;
1167 struct sched_domain
*sd
, *this_sd
= NULL
;
1171 rq
= task_rq_lock(p
, &flags
);
1172 old_state
= p
->state
;
1173 if (!(old_state
& state
))
1180 this_cpu
= smp_processor_id();
1183 if (unlikely(task_running(rq
, p
)))
1188 schedstat_inc(rq
, ttwu_cnt
);
1189 if (cpu
== this_cpu
) {
1190 schedstat_inc(rq
, ttwu_local
);
1194 for_each_domain(this_cpu
, sd
) {
1195 if (cpu_isset(cpu
, sd
->span
)) {
1196 schedstat_inc(sd
, ttwu_wake_remote
);
1202 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1206 * Check for affine wakeup and passive balancing possibilities.
1209 int idx
= this_sd
->wake_idx
;
1210 unsigned int imbalance
;
1212 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1214 load
= source_load(cpu
, idx
);
1215 this_load
= target_load(this_cpu
, idx
);
1217 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1219 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1220 unsigned long tl
= this_load
;
1222 * If sync wakeup then subtract the (maximum possible)
1223 * effect of the currently running task from the load
1224 * of the current CPU:
1227 tl
-= SCHED_LOAD_SCALE
;
1230 tl
+ target_load(cpu
, idx
) <= SCHED_LOAD_SCALE
) ||
1231 100*(tl
+ SCHED_LOAD_SCALE
) <= imbalance
*load
) {
1233 * This domain has SD_WAKE_AFFINE and
1234 * p is cache cold in this domain, and
1235 * there is no bad imbalance.
1237 schedstat_inc(this_sd
, ttwu_move_affine
);
1243 * Start passive balancing when half the imbalance_pct
1246 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1247 if (imbalance
*this_load
<= 100*load
) {
1248 schedstat_inc(this_sd
, ttwu_move_balance
);
1254 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1256 new_cpu
= wake_idle(new_cpu
, p
);
1257 if (new_cpu
!= cpu
) {
1258 set_task_cpu(p
, new_cpu
);
1259 task_rq_unlock(rq
, &flags
);
1260 /* might preempt at this point */
1261 rq
= task_rq_lock(p
, &flags
);
1262 old_state
= p
->state
;
1263 if (!(old_state
& state
))
1268 this_cpu
= smp_processor_id();
1273 #endif /* CONFIG_SMP */
1274 if (old_state
== TASK_UNINTERRUPTIBLE
) {
1275 rq
->nr_uninterruptible
--;
1277 * Tasks on involuntary sleep don't earn
1278 * sleep_avg beyond just interactive state.
1280 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
1284 * Tasks that have marked their sleep as noninteractive get
1285 * woken up with their sleep average not weighted in an
1288 if (old_state
& TASK_NONINTERACTIVE
)
1289 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
1292 activate_task(p
, rq
, cpu
== this_cpu
);
1294 * Sync wakeups (i.e. those types of wakeups where the waker
1295 * has indicated that it will leave the CPU in short order)
1296 * don't trigger a preemption, if the woken up task will run on
1297 * this cpu. (in this case the 'I will reschedule' promise of
1298 * the waker guarantees that the freshly woken up task is going
1299 * to be considered on this CPU.)
1301 if (!sync
|| cpu
!= this_cpu
) {
1302 if (TASK_PREEMPTS_CURR(p
, rq
))
1303 resched_task(rq
->curr
);
1308 p
->state
= TASK_RUNNING
;
1310 task_rq_unlock(rq
, &flags
);
1315 int fastcall
wake_up_process(task_t
*p
)
1317 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1318 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1321 EXPORT_SYMBOL(wake_up_process
);
1323 int fastcall
wake_up_state(task_t
*p
, unsigned int state
)
1325 return try_to_wake_up(p
, state
, 0);
1329 * Perform scheduler related setup for a newly forked process p.
1330 * p is forked by current.
1332 void fastcall
sched_fork(task_t
*p
, int clone_flags
)
1334 int cpu
= get_cpu();
1337 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1339 set_task_cpu(p
, cpu
);
1342 * We mark the process as running here, but have not actually
1343 * inserted it onto the runqueue yet. This guarantees that
1344 * nobody will actually run it, and a signal or other external
1345 * event cannot wake it up and insert it on the runqueue either.
1347 p
->state
= TASK_RUNNING
;
1348 INIT_LIST_HEAD(&p
->run_list
);
1350 #ifdef CONFIG_SCHEDSTATS
1351 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1353 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1356 #ifdef CONFIG_PREEMPT
1357 /* Want to start with kernel preemption disabled. */
1358 task_thread_info(p
)->preempt_count
= 1;
1361 * Share the timeslice between parent and child, thus the
1362 * total amount of pending timeslices in the system doesn't change,
1363 * resulting in more scheduling fairness.
1365 local_irq_disable();
1366 p
->time_slice
= (current
->time_slice
+ 1) >> 1;
1368 * The remainder of the first timeslice might be recovered by
1369 * the parent if the child exits early enough.
1371 p
->first_time_slice
= 1;
1372 current
->time_slice
>>= 1;
1373 p
->timestamp
= sched_clock();
1374 if (unlikely(!current
->time_slice
)) {
1376 * This case is rare, it happens when the parent has only
1377 * a single jiffy left from its timeslice. Taking the
1378 * runqueue lock is not a problem.
1380 current
->time_slice
= 1;
1388 * wake_up_new_task - wake up a newly created task for the first time.
1390 * This function will do some initial scheduler statistics housekeeping
1391 * that must be done for every newly created context, then puts the task
1392 * on the runqueue and wakes it.
1394 void fastcall
wake_up_new_task(task_t
*p
, unsigned long clone_flags
)
1396 unsigned long flags
;
1398 runqueue_t
*rq
, *this_rq
;
1400 rq
= task_rq_lock(p
, &flags
);
1401 BUG_ON(p
->state
!= TASK_RUNNING
);
1402 this_cpu
= smp_processor_id();
1406 * We decrease the sleep average of forking parents
1407 * and children as well, to keep max-interactive tasks
1408 * from forking tasks that are max-interactive. The parent
1409 * (current) is done further down, under its lock.
1411 p
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(p
) *
1412 CHILD_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1414 p
->prio
= effective_prio(p
);
1416 if (likely(cpu
== this_cpu
)) {
1417 if (!(clone_flags
& CLONE_VM
)) {
1419 * The VM isn't cloned, so we're in a good position to
1420 * do child-runs-first in anticipation of an exec. This
1421 * usually avoids a lot of COW overhead.
1423 if (unlikely(!current
->array
))
1424 __activate_task(p
, rq
);
1426 p
->prio
= current
->prio
;
1427 list_add_tail(&p
->run_list
, ¤t
->run_list
);
1428 p
->array
= current
->array
;
1429 p
->array
->nr_active
++;
1434 /* Run child last */
1435 __activate_task(p
, rq
);
1437 * We skip the following code due to cpu == this_cpu
1439 * task_rq_unlock(rq, &flags);
1440 * this_rq = task_rq_lock(current, &flags);
1444 this_rq
= cpu_rq(this_cpu
);
1447 * Not the local CPU - must adjust timestamp. This should
1448 * get optimised away in the !CONFIG_SMP case.
1450 p
->timestamp
= (p
->timestamp
- this_rq
->timestamp_last_tick
)
1451 + rq
->timestamp_last_tick
;
1452 __activate_task(p
, rq
);
1453 if (TASK_PREEMPTS_CURR(p
, rq
))
1454 resched_task(rq
->curr
);
1457 * Parent and child are on different CPUs, now get the
1458 * parent runqueue to update the parent's ->sleep_avg:
1460 task_rq_unlock(rq
, &flags
);
1461 this_rq
= task_rq_lock(current
, &flags
);
1463 current
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(current
) *
1464 PARENT_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1465 task_rq_unlock(this_rq
, &flags
);
1469 * Potentially available exiting-child timeslices are
1470 * retrieved here - this way the parent does not get
1471 * penalized for creating too many threads.
1473 * (this cannot be used to 'generate' timeslices
1474 * artificially, because any timeslice recovered here
1475 * was given away by the parent in the first place.)
1477 void fastcall
sched_exit(task_t
*p
)
1479 unsigned long flags
;
1483 * If the child was a (relative-) CPU hog then decrease
1484 * the sleep_avg of the parent as well.
1486 rq
= task_rq_lock(p
->parent
, &flags
);
1487 if (p
->first_time_slice
&& task_cpu(p
) == task_cpu(p
->parent
)) {
1488 p
->parent
->time_slice
+= p
->time_slice
;
1489 if (unlikely(p
->parent
->time_slice
> task_timeslice(p
)))
1490 p
->parent
->time_slice
= task_timeslice(p
);
1492 if (p
->sleep_avg
< p
->parent
->sleep_avg
)
1493 p
->parent
->sleep_avg
= p
->parent
->sleep_avg
/
1494 (EXIT_WEIGHT
+ 1) * EXIT_WEIGHT
+ p
->sleep_avg
/
1496 task_rq_unlock(rq
, &flags
);
1500 * prepare_task_switch - prepare to switch tasks
1501 * @rq: the runqueue preparing to switch
1502 * @next: the task we are going to switch to.
1504 * This is called with the rq lock held and interrupts off. It must
1505 * be paired with a subsequent finish_task_switch after the context
1508 * prepare_task_switch sets up locking and calls architecture specific
1511 static inline void prepare_task_switch(runqueue_t
*rq
, task_t
*next
)
1513 prepare_lock_switch(rq
, next
);
1514 prepare_arch_switch(next
);
1518 * finish_task_switch - clean up after a task-switch
1519 * @rq: runqueue associated with task-switch
1520 * @prev: the thread we just switched away from.
1522 * finish_task_switch must be called after the context switch, paired
1523 * with a prepare_task_switch call before the context switch.
1524 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1525 * and do any other architecture-specific cleanup actions.
1527 * Note that we may have delayed dropping an mm in context_switch(). If
1528 * so, we finish that here outside of the runqueue lock. (Doing it
1529 * with the lock held can cause deadlocks; see schedule() for
1532 static inline void finish_task_switch(runqueue_t
*rq
, task_t
*prev
)
1533 __releases(rq
->lock
)
1535 struct mm_struct
*mm
= rq
->prev_mm
;
1536 unsigned long prev_task_flags
;
1541 * A task struct has one reference for the use as "current".
1542 * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1543 * calls schedule one last time. The schedule call will never return,
1544 * and the scheduled task must drop that reference.
1545 * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1546 * still held, otherwise prev could be scheduled on another cpu, die
1547 * there before we look at prev->state, and then the reference would
1549 * Manfred Spraul <manfred@colorfullife.com>
1551 prev_task_flags
= prev
->flags
;
1552 finish_arch_switch(prev
);
1553 finish_lock_switch(rq
, prev
);
1556 if (unlikely(prev_task_flags
& PF_DEAD
)) {
1558 * Remove function-return probe instances associated with this
1559 * task and put them back on the free list.
1561 kprobe_flush_task(prev
);
1562 put_task_struct(prev
);
1567 * schedule_tail - first thing a freshly forked thread must call.
1568 * @prev: the thread we just switched away from.
1570 asmlinkage
void schedule_tail(task_t
*prev
)
1571 __releases(rq
->lock
)
1573 runqueue_t
*rq
= this_rq();
1574 finish_task_switch(rq
, prev
);
1575 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1576 /* In this case, finish_task_switch does not reenable preemption */
1579 if (current
->set_child_tid
)
1580 put_user(current
->pid
, current
->set_child_tid
);
1584 * context_switch - switch to the new MM and the new
1585 * thread's register state.
1588 task_t
* context_switch(runqueue_t
*rq
, task_t
*prev
, task_t
*next
)
1590 struct mm_struct
*mm
= next
->mm
;
1591 struct mm_struct
*oldmm
= prev
->active_mm
;
1593 if (unlikely(!mm
)) {
1594 next
->active_mm
= oldmm
;
1595 atomic_inc(&oldmm
->mm_count
);
1596 enter_lazy_tlb(oldmm
, next
);
1598 switch_mm(oldmm
, mm
, next
);
1600 if (unlikely(!prev
->mm
)) {
1601 prev
->active_mm
= NULL
;
1602 WARN_ON(rq
->prev_mm
);
1603 rq
->prev_mm
= oldmm
;
1606 /* Here we just switch the register state and the stack. */
1607 switch_to(prev
, next
, prev
);
1613 * nr_running, nr_uninterruptible and nr_context_switches:
1615 * externally visible scheduler statistics: current number of runnable
1616 * threads, current number of uninterruptible-sleeping threads, total
1617 * number of context switches performed since bootup.
1619 unsigned long nr_running(void)
1621 unsigned long i
, sum
= 0;
1623 for_each_online_cpu(i
)
1624 sum
+= cpu_rq(i
)->nr_running
;
1629 unsigned long nr_uninterruptible(void)
1631 unsigned long i
, sum
= 0;
1633 for_each_possible_cpu(i
)
1634 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1637 * Since we read the counters lockless, it might be slightly
1638 * inaccurate. Do not allow it to go below zero though:
1640 if (unlikely((long)sum
< 0))
1646 unsigned long long nr_context_switches(void)
1648 unsigned long long i
, sum
= 0;
1650 for_each_possible_cpu(i
)
1651 sum
+= cpu_rq(i
)->nr_switches
;
1656 unsigned long nr_iowait(void)
1658 unsigned long i
, sum
= 0;
1660 for_each_possible_cpu(i
)
1661 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1666 unsigned long nr_active(void)
1668 unsigned long i
, running
= 0, uninterruptible
= 0;
1670 for_each_online_cpu(i
) {
1671 running
+= cpu_rq(i
)->nr_running
;
1672 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
1675 if (unlikely((long)uninterruptible
< 0))
1676 uninterruptible
= 0;
1678 return running
+ uninterruptible
;
1684 * double_rq_lock - safely lock two runqueues
1686 * We must take them in cpu order to match code in
1687 * dependent_sleeper and wake_dependent_sleeper.
1689 * Note this does not disable interrupts like task_rq_lock,
1690 * you need to do so manually before calling.
1692 static void double_rq_lock(runqueue_t
*rq1
, runqueue_t
*rq2
)
1693 __acquires(rq1
->lock
)
1694 __acquires(rq2
->lock
)
1697 spin_lock(&rq1
->lock
);
1698 __acquire(rq2
->lock
); /* Fake it out ;) */
1700 if (rq1
->cpu
< rq2
->cpu
) {
1701 spin_lock(&rq1
->lock
);
1702 spin_lock(&rq2
->lock
);
1704 spin_lock(&rq2
->lock
);
1705 spin_lock(&rq1
->lock
);
1711 * double_rq_unlock - safely unlock two runqueues
1713 * Note this does not restore interrupts like task_rq_unlock,
1714 * you need to do so manually after calling.
1716 static void double_rq_unlock(runqueue_t
*rq1
, runqueue_t
*rq2
)
1717 __releases(rq1
->lock
)
1718 __releases(rq2
->lock
)
1720 spin_unlock(&rq1
->lock
);
1722 spin_unlock(&rq2
->lock
);
1724 __release(rq2
->lock
);
1728 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1730 static void double_lock_balance(runqueue_t
*this_rq
, runqueue_t
*busiest
)
1731 __releases(this_rq
->lock
)
1732 __acquires(busiest
->lock
)
1733 __acquires(this_rq
->lock
)
1735 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1736 if (busiest
->cpu
< this_rq
->cpu
) {
1737 spin_unlock(&this_rq
->lock
);
1738 spin_lock(&busiest
->lock
);
1739 spin_lock(&this_rq
->lock
);
1741 spin_lock(&busiest
->lock
);
1746 * If dest_cpu is allowed for this process, migrate the task to it.
1747 * This is accomplished by forcing the cpu_allowed mask to only
1748 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
1749 * the cpu_allowed mask is restored.
1751 static void sched_migrate_task(task_t
*p
, int dest_cpu
)
1753 migration_req_t req
;
1755 unsigned long flags
;
1757 rq
= task_rq_lock(p
, &flags
);
1758 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
1759 || unlikely(cpu_is_offline(dest_cpu
)))
1762 /* force the process onto the specified CPU */
1763 if (migrate_task(p
, dest_cpu
, &req
)) {
1764 /* Need to wait for migration thread (might exit: take ref). */
1765 struct task_struct
*mt
= rq
->migration_thread
;
1766 get_task_struct(mt
);
1767 task_rq_unlock(rq
, &flags
);
1768 wake_up_process(mt
);
1769 put_task_struct(mt
);
1770 wait_for_completion(&req
.done
);
1774 task_rq_unlock(rq
, &flags
);
1778 * sched_exec - execve() is a valuable balancing opportunity, because at
1779 * this point the task has the smallest effective memory and cache footprint.
1781 void sched_exec(void)
1783 int new_cpu
, this_cpu
= get_cpu();
1784 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
1786 if (new_cpu
!= this_cpu
)
1787 sched_migrate_task(current
, new_cpu
);
1791 * pull_task - move a task from a remote runqueue to the local runqueue.
1792 * Both runqueues must be locked.
1795 void pull_task(runqueue_t
*src_rq
, prio_array_t
*src_array
, task_t
*p
,
1796 runqueue_t
*this_rq
, prio_array_t
*this_array
, int this_cpu
)
1798 dequeue_task(p
, src_array
);
1799 src_rq
->nr_running
--;
1800 set_task_cpu(p
, this_cpu
);
1801 this_rq
->nr_running
++;
1802 enqueue_task(p
, this_array
);
1803 p
->timestamp
= (p
->timestamp
- src_rq
->timestamp_last_tick
)
1804 + this_rq
->timestamp_last_tick
;
1806 * Note that idle threads have a prio of MAX_PRIO, for this test
1807 * to be always true for them.
1809 if (TASK_PREEMPTS_CURR(p
, this_rq
))
1810 resched_task(this_rq
->curr
);
1814 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1817 int can_migrate_task(task_t
*p
, runqueue_t
*rq
, int this_cpu
,
1818 struct sched_domain
*sd
, enum idle_type idle
,
1822 * We do not migrate tasks that are:
1823 * 1) running (obviously), or
1824 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1825 * 3) are cache-hot on their current CPU.
1827 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
1831 if (task_running(rq
, p
))
1835 * Aggressive migration if:
1836 * 1) task is cache cold, or
1837 * 2) too many balance attempts have failed.
1840 if (sd
->nr_balance_failed
> sd
->cache_nice_tries
)
1843 if (task_hot(p
, rq
->timestamp_last_tick
, sd
))
1849 * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq,
1850 * as part of a balancing operation within "domain". Returns the number of
1853 * Called with both runqueues locked.
1855 static int move_tasks(runqueue_t
*this_rq
, int this_cpu
, runqueue_t
*busiest
,
1856 unsigned long max_nr_move
, struct sched_domain
*sd
,
1857 enum idle_type idle
, int *all_pinned
)
1859 prio_array_t
*array
, *dst_array
;
1860 struct list_head
*head
, *curr
;
1861 int idx
, pulled
= 0, pinned
= 0;
1864 if (max_nr_move
== 0)
1870 * We first consider expired tasks. Those will likely not be
1871 * executed in the near future, and they are most likely to
1872 * be cache-cold, thus switching CPUs has the least effect
1875 if (busiest
->expired
->nr_active
) {
1876 array
= busiest
->expired
;
1877 dst_array
= this_rq
->expired
;
1879 array
= busiest
->active
;
1880 dst_array
= this_rq
->active
;
1884 /* Start searching at priority 0: */
1888 idx
= sched_find_first_bit(array
->bitmap
);
1890 idx
= find_next_bit(array
->bitmap
, MAX_PRIO
, idx
);
1891 if (idx
>= MAX_PRIO
) {
1892 if (array
== busiest
->expired
&& busiest
->active
->nr_active
) {
1893 array
= busiest
->active
;
1894 dst_array
= this_rq
->active
;
1900 head
= array
->queue
+ idx
;
1903 tmp
= list_entry(curr
, task_t
, run_list
);
1907 if (!can_migrate_task(tmp
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
1914 #ifdef CONFIG_SCHEDSTATS
1915 if (task_hot(tmp
, busiest
->timestamp_last_tick
, sd
))
1916 schedstat_inc(sd
, lb_hot_gained
[idle
]);
1919 pull_task(busiest
, array
, tmp
, this_rq
, dst_array
, this_cpu
);
1922 /* We only want to steal up to the prescribed number of tasks. */
1923 if (pulled
< max_nr_move
) {
1931 * Right now, this is the only place pull_task() is called,
1932 * so we can safely collect pull_task() stats here rather than
1933 * inside pull_task().
1935 schedstat_add(sd
, lb_gained
[idle
], pulled
);
1938 *all_pinned
= pinned
;
1943 * find_busiest_group finds and returns the busiest CPU group within the
1944 * domain. It calculates and returns the number of tasks which should be
1945 * moved to restore balance via the imbalance parameter.
1947 static struct sched_group
*
1948 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
1949 unsigned long *imbalance
, enum idle_type idle
, int *sd_idle
)
1951 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1952 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
1953 unsigned long max_pull
;
1956 max_load
= this_load
= total_load
= total_pwr
= 0;
1957 if (idle
== NOT_IDLE
)
1958 load_idx
= sd
->busy_idx
;
1959 else if (idle
== NEWLY_IDLE
)
1960 load_idx
= sd
->newidle_idx
;
1962 load_idx
= sd
->idle_idx
;
1969 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1971 /* Tally up the load of all CPUs in the group */
1974 for_each_cpu_mask(i
, group
->cpumask
) {
1975 if (*sd_idle
&& !idle_cpu(i
))
1978 /* Bias balancing toward cpus of our domain */
1980 load
= target_load(i
, load_idx
);
1982 load
= source_load(i
, load_idx
);
1987 total_load
+= avg_load
;
1988 total_pwr
+= group
->cpu_power
;
1990 /* Adjust by relative CPU power of the group */
1991 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
1994 this_load
= avg_load
;
1996 } else if (avg_load
> max_load
) {
1997 max_load
= avg_load
;
2000 group
= group
->next
;
2001 } while (group
!= sd
->groups
);
2003 if (!busiest
|| this_load
>= max_load
|| max_load
<= SCHED_LOAD_SCALE
)
2006 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2008 if (this_load
>= avg_load
||
2009 100*max_load
<= sd
->imbalance_pct
*this_load
)
2013 * We're trying to get all the cpus to the average_load, so we don't
2014 * want to push ourselves above the average load, nor do we wish to
2015 * reduce the max loaded cpu below the average load, as either of these
2016 * actions would just result in more rebalancing later, and ping-pong
2017 * tasks around. Thus we look for the minimum possible imbalance.
2018 * Negative imbalances (*we* are more loaded than anyone else) will
2019 * be counted as no imbalance for these purposes -- we can't fix that
2020 * by pulling tasks to us. Be careful of negative numbers as they'll
2021 * appear as very large values with unsigned longs.
2024 /* Don't want to pull so many tasks that a group would go idle */
2025 max_pull
= min(max_load
- avg_load
, max_load
- SCHED_LOAD_SCALE
);
2027 /* How much load to actually move to equalise the imbalance */
2028 *imbalance
= min(max_pull
* busiest
->cpu_power
,
2029 (avg_load
- this_load
) * this->cpu_power
)
2032 if (*imbalance
< SCHED_LOAD_SCALE
) {
2033 unsigned long pwr_now
= 0, pwr_move
= 0;
2036 if (max_load
- this_load
>= SCHED_LOAD_SCALE
*2) {
2042 * OK, we don't have enough imbalance to justify moving tasks,
2043 * however we may be able to increase total CPU power used by
2047 pwr_now
+= busiest
->cpu_power
*min(SCHED_LOAD_SCALE
, max_load
);
2048 pwr_now
+= this->cpu_power
*min(SCHED_LOAD_SCALE
, this_load
);
2049 pwr_now
/= SCHED_LOAD_SCALE
;
2051 /* Amount of load we'd subtract */
2052 tmp
= SCHED_LOAD_SCALE
*SCHED_LOAD_SCALE
/busiest
->cpu_power
;
2054 pwr_move
+= busiest
->cpu_power
*min(SCHED_LOAD_SCALE
,
2057 /* Amount of load we'd add */
2058 if (max_load
*busiest
->cpu_power
<
2059 SCHED_LOAD_SCALE
*SCHED_LOAD_SCALE
)
2060 tmp
= max_load
*busiest
->cpu_power
/this->cpu_power
;
2062 tmp
= SCHED_LOAD_SCALE
*SCHED_LOAD_SCALE
/this->cpu_power
;
2063 pwr_move
+= this->cpu_power
*min(SCHED_LOAD_SCALE
, this_load
+ tmp
);
2064 pwr_move
/= SCHED_LOAD_SCALE
;
2066 /* Move if we gain throughput */
2067 if (pwr_move
<= pwr_now
)
2074 /* Get rid of the scaling factor, rounding down as we divide */
2075 *imbalance
= *imbalance
/ SCHED_LOAD_SCALE
;
2085 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2087 static runqueue_t
*find_busiest_queue(struct sched_group
*group
,
2088 enum idle_type idle
)
2090 unsigned long load
, max_load
= 0;
2091 runqueue_t
*busiest
= NULL
;
2094 for_each_cpu_mask(i
, group
->cpumask
) {
2095 load
= source_load(i
, 0);
2097 if (load
> max_load
) {
2099 busiest
= cpu_rq(i
);
2107 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2108 * so long as it is large enough.
2110 #define MAX_PINNED_INTERVAL 512
2113 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2114 * tasks if there is an imbalance.
2116 * Called with this_rq unlocked.
2118 static int load_balance(int this_cpu
, runqueue_t
*this_rq
,
2119 struct sched_domain
*sd
, enum idle_type idle
)
2121 struct sched_group
*group
;
2122 runqueue_t
*busiest
;
2123 unsigned long imbalance
;
2124 int nr_moved
, all_pinned
= 0;
2125 int active_balance
= 0;
2128 if (idle
!= NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2131 schedstat_inc(sd
, lb_cnt
[idle
]);
2133 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
);
2135 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2139 busiest
= find_busiest_queue(group
, idle
);
2141 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2145 BUG_ON(busiest
== this_rq
);
2147 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2150 if (busiest
->nr_running
> 1) {
2152 * Attempt to move tasks. If find_busiest_group has found
2153 * an imbalance but busiest->nr_running <= 1, the group is
2154 * still unbalanced. nr_moved simply stays zero, so it is
2155 * correctly treated as an imbalance.
2157 double_rq_lock(this_rq
, busiest
);
2158 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2159 imbalance
, sd
, idle
, &all_pinned
);
2160 double_rq_unlock(this_rq
, busiest
);
2162 /* All tasks on this runqueue were pinned by CPU affinity */
2163 if (unlikely(all_pinned
))
2168 schedstat_inc(sd
, lb_failed
[idle
]);
2169 sd
->nr_balance_failed
++;
2171 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2173 spin_lock(&busiest
->lock
);
2175 /* don't kick the migration_thread, if the curr
2176 * task on busiest cpu can't be moved to this_cpu
2178 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2179 spin_unlock(&busiest
->lock
);
2181 goto out_one_pinned
;
2184 if (!busiest
->active_balance
) {
2185 busiest
->active_balance
= 1;
2186 busiest
->push_cpu
= this_cpu
;
2189 spin_unlock(&busiest
->lock
);
2191 wake_up_process(busiest
->migration_thread
);
2194 * We've kicked active balancing, reset the failure
2197 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2200 sd
->nr_balance_failed
= 0;
2202 if (likely(!active_balance
)) {
2203 /* We were unbalanced, so reset the balancing interval */
2204 sd
->balance_interval
= sd
->min_interval
;
2207 * If we've begun active balancing, start to back off. This
2208 * case may not be covered by the all_pinned logic if there
2209 * is only 1 task on the busy runqueue (because we don't call
2212 if (sd
->balance_interval
< sd
->max_interval
)
2213 sd
->balance_interval
*= 2;
2216 if (!nr_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2221 schedstat_inc(sd
, lb_balanced
[idle
]);
2223 sd
->nr_balance_failed
= 0;
2226 /* tune up the balancing interval */
2227 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2228 (sd
->balance_interval
< sd
->max_interval
))
2229 sd
->balance_interval
*= 2;
2231 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2237 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2238 * tasks if there is an imbalance.
2240 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2241 * this_rq is locked.
2243 static int load_balance_newidle(int this_cpu
, runqueue_t
*this_rq
,
2244 struct sched_domain
*sd
)
2246 struct sched_group
*group
;
2247 runqueue_t
*busiest
= NULL
;
2248 unsigned long imbalance
;
2252 if (sd
->flags
& SD_SHARE_CPUPOWER
)
2255 schedstat_inc(sd
, lb_cnt
[NEWLY_IDLE
]);
2256 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, NEWLY_IDLE
, &sd_idle
);
2258 schedstat_inc(sd
, lb_nobusyg
[NEWLY_IDLE
]);
2262 busiest
= find_busiest_queue(group
, NEWLY_IDLE
);
2264 schedstat_inc(sd
, lb_nobusyq
[NEWLY_IDLE
]);
2268 BUG_ON(busiest
== this_rq
);
2270 schedstat_add(sd
, lb_imbalance
[NEWLY_IDLE
], imbalance
);
2273 if (busiest
->nr_running
> 1) {
2274 /* Attempt to move tasks */
2275 double_lock_balance(this_rq
, busiest
);
2276 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2277 imbalance
, sd
, NEWLY_IDLE
, NULL
);
2278 spin_unlock(&busiest
->lock
);
2282 schedstat_inc(sd
, lb_failed
[NEWLY_IDLE
]);
2283 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2286 sd
->nr_balance_failed
= 0;
2291 schedstat_inc(sd
, lb_balanced
[NEWLY_IDLE
]);
2292 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2294 sd
->nr_balance_failed
= 0;
2299 * idle_balance is called by schedule() if this_cpu is about to become
2300 * idle. Attempts to pull tasks from other CPUs.
2302 static void idle_balance(int this_cpu
, runqueue_t
*this_rq
)
2304 struct sched_domain
*sd
;
2306 for_each_domain(this_cpu
, sd
) {
2307 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
2308 if (load_balance_newidle(this_cpu
, this_rq
, sd
)) {
2309 /* We've pulled tasks over so stop searching */
2317 * active_load_balance is run by migration threads. It pushes running tasks
2318 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2319 * running on each physical CPU where possible, and avoids physical /
2320 * logical imbalances.
2322 * Called with busiest_rq locked.
2324 static void active_load_balance(runqueue_t
*busiest_rq
, int busiest_cpu
)
2326 struct sched_domain
*sd
;
2327 runqueue_t
*target_rq
;
2328 int target_cpu
= busiest_rq
->push_cpu
;
2330 if (busiest_rq
->nr_running
<= 1)
2331 /* no task to move */
2334 target_rq
= cpu_rq(target_cpu
);
2337 * This condition is "impossible", if it occurs
2338 * we need to fix it. Originally reported by
2339 * Bjorn Helgaas on a 128-cpu setup.
2341 BUG_ON(busiest_rq
== target_rq
);
2343 /* move a task from busiest_rq to target_rq */
2344 double_lock_balance(busiest_rq
, target_rq
);
2346 /* Search for an sd spanning us and the target CPU. */
2347 for_each_domain(target_cpu
, sd
)
2348 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2349 cpu_isset(busiest_cpu
, sd
->span
))
2352 if (unlikely(sd
== NULL
))
2355 schedstat_inc(sd
, alb_cnt
);
2357 if (move_tasks(target_rq
, target_cpu
, busiest_rq
, 1, sd
, SCHED_IDLE
, NULL
))
2358 schedstat_inc(sd
, alb_pushed
);
2360 schedstat_inc(sd
, alb_failed
);
2362 spin_unlock(&target_rq
->lock
);
2366 * rebalance_tick will get called every timer tick, on every CPU.
2368 * It checks each scheduling domain to see if it is due to be balanced,
2369 * and initiates a balancing operation if so.
2371 * Balancing parameters are set up in arch_init_sched_domains.
2374 /* Don't have all balancing operations going off at once */
2375 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2377 static void rebalance_tick(int this_cpu
, runqueue_t
*this_rq
,
2378 enum idle_type idle
)
2380 unsigned long old_load
, this_load
;
2381 unsigned long j
= jiffies
+ CPU_OFFSET(this_cpu
);
2382 struct sched_domain
*sd
;
2385 this_load
= this_rq
->nr_running
* SCHED_LOAD_SCALE
;
2386 /* Update our load */
2387 for (i
= 0; i
< 3; i
++) {
2388 unsigned long new_load
= this_load
;
2390 old_load
= this_rq
->cpu_load
[i
];
2392 * Round up the averaging division if load is increasing. This
2393 * prevents us from getting stuck on 9 if the load is 10, for
2396 if (new_load
> old_load
)
2397 new_load
+= scale
-1;
2398 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) / scale
;
2401 for_each_domain(this_cpu
, sd
) {
2402 unsigned long interval
;
2404 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2407 interval
= sd
->balance_interval
;
2408 if (idle
!= SCHED_IDLE
)
2409 interval
*= sd
->busy_factor
;
2411 /* scale ms to jiffies */
2412 interval
= msecs_to_jiffies(interval
);
2413 if (unlikely(!interval
))
2416 if (j
- sd
->last_balance
>= interval
) {
2417 if (load_balance(this_cpu
, this_rq
, sd
, idle
)) {
2419 * We've pulled tasks over so either we're no
2420 * longer idle, or one of our SMT siblings is
2425 sd
->last_balance
+= interval
;
2431 * on UP we do not need to balance between CPUs:
2433 static inline void rebalance_tick(int cpu
, runqueue_t
*rq
, enum idle_type idle
)
2436 static inline void idle_balance(int cpu
, runqueue_t
*rq
)
2441 static inline int wake_priority_sleeper(runqueue_t
*rq
)
2444 #ifdef CONFIG_SCHED_SMT
2445 spin_lock(&rq
->lock
);
2447 * If an SMT sibling task has been put to sleep for priority
2448 * reasons reschedule the idle task to see if it can now run.
2450 if (rq
->nr_running
) {
2451 resched_task(rq
->idle
);
2454 spin_unlock(&rq
->lock
);
2459 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2461 EXPORT_PER_CPU_SYMBOL(kstat
);
2464 * This is called on clock ticks and on context switches.
2465 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2467 static inline void update_cpu_clock(task_t
*p
, runqueue_t
*rq
,
2468 unsigned long long now
)
2470 unsigned long long last
= max(p
->timestamp
, rq
->timestamp_last_tick
);
2471 p
->sched_time
+= now
- last
;
2475 * Return current->sched_time plus any more ns on the sched_clock
2476 * that have not yet been banked.
2478 unsigned long long current_sched_time(const task_t
*tsk
)
2480 unsigned long long ns
;
2481 unsigned long flags
;
2482 local_irq_save(flags
);
2483 ns
= max(tsk
->timestamp
, task_rq(tsk
)->timestamp_last_tick
);
2484 ns
= tsk
->sched_time
+ (sched_clock() - ns
);
2485 local_irq_restore(flags
);
2490 * We place interactive tasks back into the active array, if possible.
2492 * To guarantee that this does not starve expired tasks we ignore the
2493 * interactivity of a task if the first expired task had to wait more
2494 * than a 'reasonable' amount of time. This deadline timeout is
2495 * load-dependent, as the frequency of array switched decreases with
2496 * increasing number of running tasks. We also ignore the interactivity
2497 * if a better static_prio task has expired:
2499 #define EXPIRED_STARVING(rq) \
2500 ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2501 (jiffies - (rq)->expired_timestamp >= \
2502 STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2503 ((rq)->curr->static_prio > (rq)->best_expired_prio))
2506 * Account user cpu time to a process.
2507 * @p: the process that the cpu time gets accounted to
2508 * @hardirq_offset: the offset to subtract from hardirq_count()
2509 * @cputime: the cpu time spent in user space since the last update
2511 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
2513 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2516 p
->utime
= cputime_add(p
->utime
, cputime
);
2518 /* Add user time to cpustat. */
2519 tmp
= cputime_to_cputime64(cputime
);
2520 if (TASK_NICE(p
) > 0)
2521 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
2523 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
2527 * Account system cpu time to a process.
2528 * @p: the process that the cpu time gets accounted to
2529 * @hardirq_offset: the offset to subtract from hardirq_count()
2530 * @cputime: the cpu time spent in kernel space since the last update
2532 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
2535 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2536 runqueue_t
*rq
= this_rq();
2539 p
->stime
= cputime_add(p
->stime
, cputime
);
2541 /* Add system time to cpustat. */
2542 tmp
= cputime_to_cputime64(cputime
);
2543 if (hardirq_count() - hardirq_offset
)
2544 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
2545 else if (softirq_count())
2546 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
2547 else if (p
!= rq
->idle
)
2548 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
2549 else if (atomic_read(&rq
->nr_iowait
) > 0)
2550 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
2552 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
2553 /* Account for system time used */
2554 acct_update_integrals(p
);
2558 * Account for involuntary wait time.
2559 * @p: the process from which the cpu time has been stolen
2560 * @steal: the cpu time spent in involuntary wait
2562 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
2564 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2565 cputime64_t tmp
= cputime_to_cputime64(steal
);
2566 runqueue_t
*rq
= this_rq();
2568 if (p
== rq
->idle
) {
2569 p
->stime
= cputime_add(p
->stime
, steal
);
2570 if (atomic_read(&rq
->nr_iowait
) > 0)
2571 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
2573 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
2575 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
2579 * This function gets called by the timer code, with HZ frequency.
2580 * We call it with interrupts disabled.
2582 * It also gets called by the fork code, when changing the parent's
2585 void scheduler_tick(void)
2587 int cpu
= smp_processor_id();
2588 runqueue_t
*rq
= this_rq();
2589 task_t
*p
= current
;
2590 unsigned long long now
= sched_clock();
2592 update_cpu_clock(p
, rq
, now
);
2594 rq
->timestamp_last_tick
= now
;
2596 if (p
== rq
->idle
) {
2597 if (wake_priority_sleeper(rq
))
2599 rebalance_tick(cpu
, rq
, SCHED_IDLE
);
2603 /* Task might have expired already, but not scheduled off yet */
2604 if (p
->array
!= rq
->active
) {
2605 set_tsk_need_resched(p
);
2608 spin_lock(&rq
->lock
);
2610 * The task was running during this tick - update the
2611 * time slice counter. Note: we do not update a thread's
2612 * priority until it either goes to sleep or uses up its
2613 * timeslice. This makes it possible for interactive tasks
2614 * to use up their timeslices at their highest priority levels.
2618 * RR tasks need a special form of timeslice management.
2619 * FIFO tasks have no timeslices.
2621 if ((p
->policy
== SCHED_RR
) && !--p
->time_slice
) {
2622 p
->time_slice
= task_timeslice(p
);
2623 p
->first_time_slice
= 0;
2624 set_tsk_need_resched(p
);
2626 /* put it at the end of the queue: */
2627 requeue_task(p
, rq
->active
);
2631 if (!--p
->time_slice
) {
2632 dequeue_task(p
, rq
->active
);
2633 set_tsk_need_resched(p
);
2634 p
->prio
= effective_prio(p
);
2635 p
->time_slice
= task_timeslice(p
);
2636 p
->first_time_slice
= 0;
2638 if (!rq
->expired_timestamp
)
2639 rq
->expired_timestamp
= jiffies
;
2640 if (!TASK_INTERACTIVE(p
) || EXPIRED_STARVING(rq
)) {
2641 enqueue_task(p
, rq
->expired
);
2642 if (p
->static_prio
< rq
->best_expired_prio
)
2643 rq
->best_expired_prio
= p
->static_prio
;
2645 enqueue_task(p
, rq
->active
);
2648 * Prevent a too long timeslice allowing a task to monopolize
2649 * the CPU. We do this by splitting up the timeslice into
2652 * Note: this does not mean the task's timeslices expire or
2653 * get lost in any way, they just might be preempted by
2654 * another task of equal priority. (one with higher
2655 * priority would have preempted this task already.) We
2656 * requeue this task to the end of the list on this priority
2657 * level, which is in essence a round-robin of tasks with
2660 * This only applies to tasks in the interactive
2661 * delta range with at least TIMESLICE_GRANULARITY to requeue.
2663 if (TASK_INTERACTIVE(p
) && !((task_timeslice(p
) -
2664 p
->time_slice
) % TIMESLICE_GRANULARITY(p
)) &&
2665 (p
->time_slice
>= TIMESLICE_GRANULARITY(p
)) &&
2666 (p
->array
== rq
->active
)) {
2668 requeue_task(p
, rq
->active
);
2669 set_tsk_need_resched(p
);
2673 spin_unlock(&rq
->lock
);
2675 rebalance_tick(cpu
, rq
, NOT_IDLE
);
2678 #ifdef CONFIG_SCHED_SMT
2679 static inline void wakeup_busy_runqueue(runqueue_t
*rq
)
2681 /* If an SMT runqueue is sleeping due to priority reasons wake it up */
2682 if (rq
->curr
== rq
->idle
&& rq
->nr_running
)
2683 resched_task(rq
->idle
);
2686 static void wake_sleeping_dependent(int this_cpu
, runqueue_t
*this_rq
)
2688 struct sched_domain
*tmp
, *sd
= NULL
;
2689 cpumask_t sibling_map
;
2692 for_each_domain(this_cpu
, tmp
)
2693 if (tmp
->flags
& SD_SHARE_CPUPOWER
)
2700 * Unlock the current runqueue because we have to lock in
2701 * CPU order to avoid deadlocks. Caller knows that we might
2702 * unlock. We keep IRQs disabled.
2704 spin_unlock(&this_rq
->lock
);
2706 sibling_map
= sd
->span
;
2708 for_each_cpu_mask(i
, sibling_map
)
2709 spin_lock(&cpu_rq(i
)->lock
);
2711 * We clear this CPU from the mask. This both simplifies the
2712 * inner loop and keps this_rq locked when we exit:
2714 cpu_clear(this_cpu
, sibling_map
);
2716 for_each_cpu_mask(i
, sibling_map
) {
2717 runqueue_t
*smt_rq
= cpu_rq(i
);
2719 wakeup_busy_runqueue(smt_rq
);
2722 for_each_cpu_mask(i
, sibling_map
)
2723 spin_unlock(&cpu_rq(i
)->lock
);
2725 * We exit with this_cpu's rq still held and IRQs
2731 * number of 'lost' timeslices this task wont be able to fully
2732 * utilize, if another task runs on a sibling. This models the
2733 * slowdown effect of other tasks running on siblings:
2735 static inline unsigned long smt_slice(task_t
*p
, struct sched_domain
*sd
)
2737 return p
->time_slice
* (100 - sd
->per_cpu_gain
) / 100;
2740 static int dependent_sleeper(int this_cpu
, runqueue_t
*this_rq
)
2742 struct sched_domain
*tmp
, *sd
= NULL
;
2743 cpumask_t sibling_map
;
2744 prio_array_t
*array
;
2748 for_each_domain(this_cpu
, tmp
)
2749 if (tmp
->flags
& SD_SHARE_CPUPOWER
)
2756 * The same locking rules and details apply as for
2757 * wake_sleeping_dependent():
2759 spin_unlock(&this_rq
->lock
);
2760 sibling_map
= sd
->span
;
2761 for_each_cpu_mask(i
, sibling_map
)
2762 spin_lock(&cpu_rq(i
)->lock
);
2763 cpu_clear(this_cpu
, sibling_map
);
2766 * Establish next task to be run - it might have gone away because
2767 * we released the runqueue lock above:
2769 if (!this_rq
->nr_running
)
2771 array
= this_rq
->active
;
2772 if (!array
->nr_active
)
2773 array
= this_rq
->expired
;
2774 BUG_ON(!array
->nr_active
);
2776 p
= list_entry(array
->queue
[sched_find_first_bit(array
->bitmap
)].next
,
2779 for_each_cpu_mask(i
, sibling_map
) {
2780 runqueue_t
*smt_rq
= cpu_rq(i
);
2781 task_t
*smt_curr
= smt_rq
->curr
;
2783 /* Kernel threads do not participate in dependent sleeping */
2784 if (!p
->mm
|| !smt_curr
->mm
|| rt_task(p
))
2785 goto check_smt_task
;
2788 * If a user task with lower static priority than the
2789 * running task on the SMT sibling is trying to schedule,
2790 * delay it till there is proportionately less timeslice
2791 * left of the sibling task to prevent a lower priority
2792 * task from using an unfair proportion of the
2793 * physical cpu's resources. -ck
2795 if (rt_task(smt_curr
)) {
2797 * With real time tasks we run non-rt tasks only
2798 * per_cpu_gain% of the time.
2800 if ((jiffies
% DEF_TIMESLICE
) >
2801 (sd
->per_cpu_gain
* DEF_TIMESLICE
/ 100))
2804 if (smt_curr
->static_prio
< p
->static_prio
&&
2805 !TASK_PREEMPTS_CURR(p
, smt_rq
) &&
2806 smt_slice(smt_curr
, sd
) > task_timeslice(p
))
2810 if ((!smt_curr
->mm
&& smt_curr
!= smt_rq
->idle
) ||
2814 wakeup_busy_runqueue(smt_rq
);
2819 * Reschedule a lower priority task on the SMT sibling for
2820 * it to be put to sleep, or wake it up if it has been put to
2821 * sleep for priority reasons to see if it should run now.
2824 if ((jiffies
% DEF_TIMESLICE
) >
2825 (sd
->per_cpu_gain
* DEF_TIMESLICE
/ 100))
2826 resched_task(smt_curr
);
2828 if (TASK_PREEMPTS_CURR(p
, smt_rq
) &&
2829 smt_slice(p
, sd
) > task_timeslice(smt_curr
))
2830 resched_task(smt_curr
);
2832 wakeup_busy_runqueue(smt_rq
);
2836 for_each_cpu_mask(i
, sibling_map
)
2837 spin_unlock(&cpu_rq(i
)->lock
);
2841 static inline void wake_sleeping_dependent(int this_cpu
, runqueue_t
*this_rq
)
2845 static inline int dependent_sleeper(int this_cpu
, runqueue_t
*this_rq
)
2851 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
2853 void fastcall
add_preempt_count(int val
)
2858 BUG_ON((preempt_count() < 0));
2859 preempt_count() += val
;
2861 * Spinlock count overflowing soon?
2863 BUG_ON((preempt_count() & PREEMPT_MASK
) >= PREEMPT_MASK
-10);
2865 EXPORT_SYMBOL(add_preempt_count
);
2867 void fastcall
sub_preempt_count(int val
)
2872 BUG_ON(val
> preempt_count());
2874 * Is the spinlock portion underflowing?
2876 BUG_ON((val
< PREEMPT_MASK
) && !(preempt_count() & PREEMPT_MASK
));
2877 preempt_count() -= val
;
2879 EXPORT_SYMBOL(sub_preempt_count
);
2883 static inline int interactive_sleep(enum sleep_type sleep_type
)
2885 return (sleep_type
== SLEEP_INTERACTIVE
||
2886 sleep_type
== SLEEP_INTERRUPTED
);
2890 * schedule() is the main scheduler function.
2892 asmlinkage
void __sched
schedule(void)
2895 task_t
*prev
, *next
;
2897 prio_array_t
*array
;
2898 struct list_head
*queue
;
2899 unsigned long long now
;
2900 unsigned long run_time
;
2901 int cpu
, idx
, new_prio
;
2904 * Test if we are atomic. Since do_exit() needs to call into
2905 * schedule() atomically, we ignore that path for now.
2906 * Otherwise, whine if we are scheduling when we should not be.
2908 if (unlikely(in_atomic() && !current
->exit_state
)) {
2909 printk(KERN_ERR
"BUG: scheduling while atomic: "
2911 current
->comm
, preempt_count(), current
->pid
);
2914 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2919 release_kernel_lock(prev
);
2920 need_resched_nonpreemptible
:
2924 * The idle thread is not allowed to schedule!
2925 * Remove this check after it has been exercised a bit.
2927 if (unlikely(prev
== rq
->idle
) && prev
->state
!= TASK_RUNNING
) {
2928 printk(KERN_ERR
"bad: scheduling from the idle thread!\n");
2932 schedstat_inc(rq
, sched_cnt
);
2933 now
= sched_clock();
2934 if (likely((long long)(now
- prev
->timestamp
) < NS_MAX_SLEEP_AVG
)) {
2935 run_time
= now
- prev
->timestamp
;
2936 if (unlikely((long long)(now
- prev
->timestamp
) < 0))
2939 run_time
= NS_MAX_SLEEP_AVG
;
2942 * Tasks charged proportionately less run_time at high sleep_avg to
2943 * delay them losing their interactive status
2945 run_time
/= (CURRENT_BONUS(prev
) ? : 1);
2947 spin_lock_irq(&rq
->lock
);
2949 if (unlikely(prev
->flags
& PF_DEAD
))
2950 prev
->state
= EXIT_DEAD
;
2952 switch_count
= &prev
->nivcsw
;
2953 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
2954 switch_count
= &prev
->nvcsw
;
2955 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
2956 unlikely(signal_pending(prev
))))
2957 prev
->state
= TASK_RUNNING
;
2959 if (prev
->state
== TASK_UNINTERRUPTIBLE
)
2960 rq
->nr_uninterruptible
++;
2961 deactivate_task(prev
, rq
);
2965 cpu
= smp_processor_id();
2966 if (unlikely(!rq
->nr_running
)) {
2968 idle_balance(cpu
, rq
);
2969 if (!rq
->nr_running
) {
2971 rq
->expired_timestamp
= 0;
2972 wake_sleeping_dependent(cpu
, rq
);
2974 * wake_sleeping_dependent() might have released
2975 * the runqueue, so break out if we got new
2978 if (!rq
->nr_running
)
2982 if (dependent_sleeper(cpu
, rq
)) {
2987 * dependent_sleeper() releases and reacquires the runqueue
2988 * lock, hence go into the idle loop if the rq went
2991 if (unlikely(!rq
->nr_running
))
2996 if (unlikely(!array
->nr_active
)) {
2998 * Switch the active and expired arrays.
3000 schedstat_inc(rq
, sched_switch
);
3001 rq
->active
= rq
->expired
;
3002 rq
->expired
= array
;
3004 rq
->expired_timestamp
= 0;
3005 rq
->best_expired_prio
= MAX_PRIO
;
3008 idx
= sched_find_first_bit(array
->bitmap
);
3009 queue
= array
->queue
+ idx
;
3010 next
= list_entry(queue
->next
, task_t
, run_list
);
3012 if (!rt_task(next
) && interactive_sleep(next
->sleep_type
)) {
3013 unsigned long long delta
= now
- next
->timestamp
;
3014 if (unlikely((long long)(now
- next
->timestamp
) < 0))
3017 if (next
->sleep_type
== SLEEP_INTERACTIVE
)
3018 delta
= delta
* (ON_RUNQUEUE_WEIGHT
* 128 / 100) / 128;
3020 array
= next
->array
;
3021 new_prio
= recalc_task_prio(next
, next
->timestamp
+ delta
);
3023 if (unlikely(next
->prio
!= new_prio
)) {
3024 dequeue_task(next
, array
);
3025 next
->prio
= new_prio
;
3026 enqueue_task(next
, array
);
3028 requeue_task(next
, array
);
3030 next
->sleep_type
= SLEEP_NORMAL
;
3032 if (next
== rq
->idle
)
3033 schedstat_inc(rq
, sched_goidle
);
3035 prefetch_stack(next
);
3036 clear_tsk_need_resched(prev
);
3037 rcu_qsctr_inc(task_cpu(prev
));
3039 update_cpu_clock(prev
, rq
, now
);
3041 prev
->sleep_avg
-= run_time
;
3042 if ((long)prev
->sleep_avg
<= 0)
3043 prev
->sleep_avg
= 0;
3044 prev
->timestamp
= prev
->last_ran
= now
;
3046 sched_info_switch(prev
, next
);
3047 if (likely(prev
!= next
)) {
3048 next
->timestamp
= now
;
3053 prepare_task_switch(rq
, next
);
3054 prev
= context_switch(rq
, prev
, next
);
3057 * this_rq must be evaluated again because prev may have moved
3058 * CPUs since it called schedule(), thus the 'rq' on its stack
3059 * frame will be invalid.
3061 finish_task_switch(this_rq(), prev
);
3063 spin_unlock_irq(&rq
->lock
);
3066 if (unlikely(reacquire_kernel_lock(prev
) < 0))
3067 goto need_resched_nonpreemptible
;
3068 preempt_enable_no_resched();
3069 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3073 EXPORT_SYMBOL(schedule
);
3075 #ifdef CONFIG_PREEMPT
3077 * this is is the entry point to schedule() from in-kernel preemption
3078 * off of preempt_enable. Kernel preemptions off return from interrupt
3079 * occur there and call schedule directly.
3081 asmlinkage
void __sched
preempt_schedule(void)
3083 struct thread_info
*ti
= current_thread_info();
3084 #ifdef CONFIG_PREEMPT_BKL
3085 struct task_struct
*task
= current
;
3086 int saved_lock_depth
;
3089 * If there is a non-zero preempt_count or interrupts are disabled,
3090 * we do not want to preempt the current task. Just return..
3092 if (unlikely(ti
->preempt_count
|| irqs_disabled()))
3096 add_preempt_count(PREEMPT_ACTIVE
);
3098 * We keep the big kernel semaphore locked, but we
3099 * clear ->lock_depth so that schedule() doesnt
3100 * auto-release the semaphore:
3102 #ifdef CONFIG_PREEMPT_BKL
3103 saved_lock_depth
= task
->lock_depth
;
3104 task
->lock_depth
= -1;
3107 #ifdef CONFIG_PREEMPT_BKL
3108 task
->lock_depth
= saved_lock_depth
;
3110 sub_preempt_count(PREEMPT_ACTIVE
);
3112 /* we could miss a preemption opportunity between schedule and now */
3114 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3118 EXPORT_SYMBOL(preempt_schedule
);
3121 * this is is the entry point to schedule() from kernel preemption
3122 * off of irq context.
3123 * Note, that this is called and return with irqs disabled. This will
3124 * protect us against recursive calling from irq.
3126 asmlinkage
void __sched
preempt_schedule_irq(void)
3128 struct thread_info
*ti
= current_thread_info();
3129 #ifdef CONFIG_PREEMPT_BKL
3130 struct task_struct
*task
= current
;
3131 int saved_lock_depth
;
3133 /* Catch callers which need to be fixed*/
3134 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3137 add_preempt_count(PREEMPT_ACTIVE
);
3139 * We keep the big kernel semaphore locked, but we
3140 * clear ->lock_depth so that schedule() doesnt
3141 * auto-release the semaphore:
3143 #ifdef CONFIG_PREEMPT_BKL
3144 saved_lock_depth
= task
->lock_depth
;
3145 task
->lock_depth
= -1;
3149 local_irq_disable();
3150 #ifdef CONFIG_PREEMPT_BKL
3151 task
->lock_depth
= saved_lock_depth
;
3153 sub_preempt_count(PREEMPT_ACTIVE
);
3155 /* we could miss a preemption opportunity between schedule and now */
3157 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3161 #endif /* CONFIG_PREEMPT */
3163 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3166 task_t
*p
= curr
->private;
3167 return try_to_wake_up(p
, mode
, sync
);
3170 EXPORT_SYMBOL(default_wake_function
);
3173 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3174 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3175 * number) then we wake all the non-exclusive tasks and one exclusive task.
3177 * There are circumstances in which we can try to wake a task which has already
3178 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3179 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3181 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3182 int nr_exclusive
, int sync
, void *key
)
3184 struct list_head
*tmp
, *next
;
3186 list_for_each_safe(tmp
, next
, &q
->task_list
) {
3189 curr
= list_entry(tmp
, wait_queue_t
, task_list
);
3190 flags
= curr
->flags
;
3191 if (curr
->func(curr
, mode
, sync
, key
) &&
3192 (flags
& WQ_FLAG_EXCLUSIVE
) &&
3199 * __wake_up - wake up threads blocked on a waitqueue.
3201 * @mode: which threads
3202 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3203 * @key: is directly passed to the wakeup function
3205 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3206 int nr_exclusive
, void *key
)
3208 unsigned long flags
;
3210 spin_lock_irqsave(&q
->lock
, flags
);
3211 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3212 spin_unlock_irqrestore(&q
->lock
, flags
);
3215 EXPORT_SYMBOL(__wake_up
);
3218 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3220 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3222 __wake_up_common(q
, mode
, 1, 0, NULL
);
3226 * __wake_up_sync - wake up threads blocked on a waitqueue.
3228 * @mode: which threads
3229 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3231 * The sync wakeup differs that the waker knows that it will schedule
3232 * away soon, so while the target thread will be woken up, it will not
3233 * be migrated to another CPU - ie. the two threads are 'synchronized'
3234 * with each other. This can prevent needless bouncing between CPUs.
3236 * On UP it can prevent extra preemption.
3239 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3241 unsigned long flags
;
3247 if (unlikely(!nr_exclusive
))
3250 spin_lock_irqsave(&q
->lock
, flags
);
3251 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3252 spin_unlock_irqrestore(&q
->lock
, flags
);
3254 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3256 void fastcall
complete(struct completion
*x
)
3258 unsigned long flags
;
3260 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3262 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3264 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3266 EXPORT_SYMBOL(complete
);
3268 void fastcall
complete_all(struct completion
*x
)
3270 unsigned long flags
;
3272 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3273 x
->done
+= UINT_MAX
/2;
3274 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3276 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3278 EXPORT_SYMBOL(complete_all
);
3280 void fastcall __sched
wait_for_completion(struct completion
*x
)
3283 spin_lock_irq(&x
->wait
.lock
);
3285 DECLARE_WAITQUEUE(wait
, current
);
3287 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3288 __add_wait_queue_tail(&x
->wait
, &wait
);
3290 __set_current_state(TASK_UNINTERRUPTIBLE
);
3291 spin_unlock_irq(&x
->wait
.lock
);
3293 spin_lock_irq(&x
->wait
.lock
);
3295 __remove_wait_queue(&x
->wait
, &wait
);
3298 spin_unlock_irq(&x
->wait
.lock
);
3300 EXPORT_SYMBOL(wait_for_completion
);
3302 unsigned long fastcall __sched
3303 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3307 spin_lock_irq(&x
->wait
.lock
);
3309 DECLARE_WAITQUEUE(wait
, current
);
3311 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3312 __add_wait_queue_tail(&x
->wait
, &wait
);
3314 __set_current_state(TASK_UNINTERRUPTIBLE
);
3315 spin_unlock_irq(&x
->wait
.lock
);
3316 timeout
= schedule_timeout(timeout
);
3317 spin_lock_irq(&x
->wait
.lock
);
3319 __remove_wait_queue(&x
->wait
, &wait
);
3323 __remove_wait_queue(&x
->wait
, &wait
);
3327 spin_unlock_irq(&x
->wait
.lock
);
3330 EXPORT_SYMBOL(wait_for_completion_timeout
);
3332 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3338 spin_lock_irq(&x
->wait
.lock
);
3340 DECLARE_WAITQUEUE(wait
, current
);
3342 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3343 __add_wait_queue_tail(&x
->wait
, &wait
);
3345 if (signal_pending(current
)) {
3347 __remove_wait_queue(&x
->wait
, &wait
);
3350 __set_current_state(TASK_INTERRUPTIBLE
);
3351 spin_unlock_irq(&x
->wait
.lock
);
3353 spin_lock_irq(&x
->wait
.lock
);
3355 __remove_wait_queue(&x
->wait
, &wait
);
3359 spin_unlock_irq(&x
->wait
.lock
);
3363 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3365 unsigned long fastcall __sched
3366 wait_for_completion_interruptible_timeout(struct completion
*x
,
3367 unsigned long timeout
)
3371 spin_lock_irq(&x
->wait
.lock
);
3373 DECLARE_WAITQUEUE(wait
, current
);
3375 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3376 __add_wait_queue_tail(&x
->wait
, &wait
);
3378 if (signal_pending(current
)) {
3379 timeout
= -ERESTARTSYS
;
3380 __remove_wait_queue(&x
->wait
, &wait
);
3383 __set_current_state(TASK_INTERRUPTIBLE
);
3384 spin_unlock_irq(&x
->wait
.lock
);
3385 timeout
= schedule_timeout(timeout
);
3386 spin_lock_irq(&x
->wait
.lock
);
3388 __remove_wait_queue(&x
->wait
, &wait
);
3392 __remove_wait_queue(&x
->wait
, &wait
);
3396 spin_unlock_irq(&x
->wait
.lock
);
3399 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3402 #define SLEEP_ON_VAR \
3403 unsigned long flags; \
3404 wait_queue_t wait; \
3405 init_waitqueue_entry(&wait, current);
3407 #define SLEEP_ON_HEAD \
3408 spin_lock_irqsave(&q->lock,flags); \
3409 __add_wait_queue(q, &wait); \
3410 spin_unlock(&q->lock);
3412 #define SLEEP_ON_TAIL \
3413 spin_lock_irq(&q->lock); \
3414 __remove_wait_queue(q, &wait); \
3415 spin_unlock_irqrestore(&q->lock, flags);
3417 void fastcall __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3421 current
->state
= TASK_INTERRUPTIBLE
;
3428 EXPORT_SYMBOL(interruptible_sleep_on
);
3430 long fastcall __sched
3431 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3435 current
->state
= TASK_INTERRUPTIBLE
;
3438 timeout
= schedule_timeout(timeout
);
3444 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3446 void fastcall __sched
sleep_on(wait_queue_head_t
*q
)
3450 current
->state
= TASK_UNINTERRUPTIBLE
;
3457 EXPORT_SYMBOL(sleep_on
);
3459 long fastcall __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3463 current
->state
= TASK_UNINTERRUPTIBLE
;
3466 timeout
= schedule_timeout(timeout
);
3472 EXPORT_SYMBOL(sleep_on_timeout
);
3474 void set_user_nice(task_t
*p
, long nice
)
3476 unsigned long flags
;
3477 prio_array_t
*array
;
3479 int old_prio
, new_prio
, delta
;
3481 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3484 * We have to be careful, if called from sys_setpriority(),
3485 * the task might be in the middle of scheduling on another CPU.
3487 rq
= task_rq_lock(p
, &flags
);
3489 * The RT priorities are set via sched_setscheduler(), but we still
3490 * allow the 'normal' nice value to be set - but as expected
3491 * it wont have any effect on scheduling until the task is
3492 * not SCHED_NORMAL/SCHED_BATCH:
3495 p
->static_prio
= NICE_TO_PRIO(nice
);
3500 dequeue_task(p
, array
);
3503 new_prio
= NICE_TO_PRIO(nice
);
3504 delta
= new_prio
- old_prio
;
3505 p
->static_prio
= NICE_TO_PRIO(nice
);
3509 enqueue_task(p
, array
);
3511 * If the task increased its priority or is running and
3512 * lowered its priority, then reschedule its CPU:
3514 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3515 resched_task(rq
->curr
);
3518 task_rq_unlock(rq
, &flags
);
3521 EXPORT_SYMBOL(set_user_nice
);
3524 * can_nice - check if a task can reduce its nice value
3528 int can_nice(const task_t
*p
, const int nice
)
3530 /* convert nice value [19,-20] to rlimit style value [1,40] */
3531 int nice_rlim
= 20 - nice
;
3532 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
3533 capable(CAP_SYS_NICE
));
3536 #ifdef __ARCH_WANT_SYS_NICE
3539 * sys_nice - change the priority of the current process.
3540 * @increment: priority increment
3542 * sys_setpriority is a more generic, but much slower function that
3543 * does similar things.
3545 asmlinkage
long sys_nice(int increment
)
3551 * Setpriority might change our priority at the same moment.
3552 * We don't have to worry. Conceptually one call occurs first
3553 * and we have a single winner.
3555 if (increment
< -40)
3560 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
3566 if (increment
< 0 && !can_nice(current
, nice
))
3569 retval
= security_task_setnice(current
, nice
);
3573 set_user_nice(current
, nice
);
3580 * task_prio - return the priority value of a given task.
3581 * @p: the task in question.
3583 * This is the priority value as seen by users in /proc.
3584 * RT tasks are offset by -200. Normal tasks are centered
3585 * around 0, value goes from -16 to +15.
3587 int task_prio(const task_t
*p
)
3589 return p
->prio
- MAX_RT_PRIO
;
3593 * task_nice - return the nice value of a given task.
3594 * @p: the task in question.
3596 int task_nice(const task_t
*p
)
3598 return TASK_NICE(p
);
3600 EXPORT_SYMBOL_GPL(task_nice
);
3603 * idle_cpu - is a given cpu idle currently?
3604 * @cpu: the processor in question.
3606 int idle_cpu(int cpu
)
3608 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
3612 * idle_task - return the idle task for a given cpu.
3613 * @cpu: the processor in question.
3615 task_t
*idle_task(int cpu
)
3617 return cpu_rq(cpu
)->idle
;
3621 * find_process_by_pid - find a process with a matching PID value.
3622 * @pid: the pid in question.
3624 static inline task_t
*find_process_by_pid(pid_t pid
)
3626 return pid
? find_task_by_pid(pid
) : current
;
3629 /* Actually do priority change: must hold rq lock. */
3630 static void __setscheduler(struct task_struct
*p
, int policy
, int prio
)
3634 p
->rt_priority
= prio
;
3635 if (policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
) {
3636 p
->prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
3638 p
->prio
= p
->static_prio
;
3640 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
3642 if (policy
== SCHED_BATCH
)
3648 * sched_setscheduler - change the scheduling policy and/or RT priority of
3650 * @p: the task in question.
3651 * @policy: new policy.
3652 * @param: structure containing the new RT priority.
3654 int sched_setscheduler(struct task_struct
*p
, int policy
,
3655 struct sched_param
*param
)
3658 int oldprio
, oldpolicy
= -1;
3659 prio_array_t
*array
;
3660 unsigned long flags
;
3664 /* double check policy once rq lock held */
3666 policy
= oldpolicy
= p
->policy
;
3667 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3668 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
)
3671 * Valid priorities for SCHED_FIFO and SCHED_RR are
3672 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
3675 if (param
->sched_priority
< 0 ||
3676 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3677 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
3679 if ((policy
== SCHED_NORMAL
|| policy
== SCHED_BATCH
)
3680 != (param
->sched_priority
== 0))
3684 * Allow unprivileged RT tasks to decrease priority:
3686 if (!capable(CAP_SYS_NICE
)) {
3688 * can't change policy, except between SCHED_NORMAL
3691 if (((policy
!= SCHED_NORMAL
&& p
->policy
!= SCHED_BATCH
) &&
3692 (policy
!= SCHED_BATCH
&& p
->policy
!= SCHED_NORMAL
)) &&
3693 !p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
)
3695 /* can't increase priority */
3696 if ((policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
) &&
3697 param
->sched_priority
> p
->rt_priority
&&
3698 param
->sched_priority
>
3699 p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
)
3701 /* can't change other user's priorities */
3702 if ((current
->euid
!= p
->euid
) &&
3703 (current
->euid
!= p
->uid
))
3707 retval
= security_task_setscheduler(p
, policy
, param
);
3711 * To be able to change p->policy safely, the apropriate
3712 * runqueue lock must be held.
3714 rq
= task_rq_lock(p
, &flags
);
3715 /* recheck policy now with rq lock held */
3716 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3717 policy
= oldpolicy
= -1;
3718 task_rq_unlock(rq
, &flags
);
3723 deactivate_task(p
, rq
);
3725 __setscheduler(p
, policy
, param
->sched_priority
);
3727 __activate_task(p
, rq
);
3729 * Reschedule if we are currently running on this runqueue and
3730 * our priority decreased, or if we are not currently running on
3731 * this runqueue and our priority is higher than the current's
3733 if (task_running(rq
, p
)) {
3734 if (p
->prio
> oldprio
)
3735 resched_task(rq
->curr
);
3736 } else if (TASK_PREEMPTS_CURR(p
, rq
))
3737 resched_task(rq
->curr
);
3739 task_rq_unlock(rq
, &flags
);
3742 EXPORT_SYMBOL_GPL(sched_setscheduler
);
3745 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
3748 struct sched_param lparam
;
3749 struct task_struct
*p
;
3751 if (!param
|| pid
< 0)
3753 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
3755 read_lock_irq(&tasklist_lock
);
3756 p
= find_process_by_pid(pid
);
3758 read_unlock_irq(&tasklist_lock
);
3761 retval
= sched_setscheduler(p
, policy
, &lparam
);
3762 read_unlock_irq(&tasklist_lock
);
3767 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3768 * @pid: the pid in question.
3769 * @policy: new policy.
3770 * @param: structure containing the new RT priority.
3772 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
3773 struct sched_param __user
*param
)
3775 /* negative values for policy are not valid */
3779 return do_sched_setscheduler(pid
, policy
, param
);
3783 * sys_sched_setparam - set/change the RT priority of a thread
3784 * @pid: the pid in question.
3785 * @param: structure containing the new RT priority.
3787 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
3789 return do_sched_setscheduler(pid
, -1, param
);
3793 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3794 * @pid: the pid in question.
3796 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
3798 int retval
= -EINVAL
;
3805 read_lock(&tasklist_lock
);
3806 p
= find_process_by_pid(pid
);
3808 retval
= security_task_getscheduler(p
);
3812 read_unlock(&tasklist_lock
);
3819 * sys_sched_getscheduler - get the RT priority of a thread
3820 * @pid: the pid in question.
3821 * @param: structure containing the RT priority.
3823 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
3825 struct sched_param lp
;
3826 int retval
= -EINVAL
;
3829 if (!param
|| pid
< 0)
3832 read_lock(&tasklist_lock
);
3833 p
= find_process_by_pid(pid
);
3838 retval
= security_task_getscheduler(p
);
3842 lp
.sched_priority
= p
->rt_priority
;
3843 read_unlock(&tasklist_lock
);
3846 * This one might sleep, we cannot do it with a spinlock held ...
3848 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
3854 read_unlock(&tasklist_lock
);
3858 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
3862 cpumask_t cpus_allowed
;
3865 read_lock(&tasklist_lock
);
3867 p
= find_process_by_pid(pid
);
3869 read_unlock(&tasklist_lock
);
3870 unlock_cpu_hotplug();
3875 * It is not safe to call set_cpus_allowed with the
3876 * tasklist_lock held. We will bump the task_struct's
3877 * usage count and then drop tasklist_lock.
3880 read_unlock(&tasklist_lock
);
3883 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
3884 !capable(CAP_SYS_NICE
))
3887 cpus_allowed
= cpuset_cpus_allowed(p
);
3888 cpus_and(new_mask
, new_mask
, cpus_allowed
);
3889 retval
= set_cpus_allowed(p
, new_mask
);
3893 unlock_cpu_hotplug();
3897 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
3898 cpumask_t
*new_mask
)
3900 if (len
< sizeof(cpumask_t
)) {
3901 memset(new_mask
, 0, sizeof(cpumask_t
));
3902 } else if (len
> sizeof(cpumask_t
)) {
3903 len
= sizeof(cpumask_t
);
3905 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
3909 * sys_sched_setaffinity - set the cpu affinity of a process
3910 * @pid: pid of the process
3911 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3912 * @user_mask_ptr: user-space pointer to the new cpu mask
3914 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
3915 unsigned long __user
*user_mask_ptr
)
3920 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
3924 return sched_setaffinity(pid
, new_mask
);
3928 * Represents all cpu's present in the system
3929 * In systems capable of hotplug, this map could dynamically grow
3930 * as new cpu's are detected in the system via any platform specific
3931 * method, such as ACPI for e.g.
3934 cpumask_t cpu_present_map __read_mostly
;
3935 EXPORT_SYMBOL(cpu_present_map
);
3938 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
3939 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
3942 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
3948 read_lock(&tasklist_lock
);
3951 p
= find_process_by_pid(pid
);
3956 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
3959 read_unlock(&tasklist_lock
);
3960 unlock_cpu_hotplug();
3968 * sys_sched_getaffinity - get the cpu affinity of a process
3969 * @pid: pid of the process
3970 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3971 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3973 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
3974 unsigned long __user
*user_mask_ptr
)
3979 if (len
< sizeof(cpumask_t
))
3982 ret
= sched_getaffinity(pid
, &mask
);
3986 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
3989 return sizeof(cpumask_t
);
3993 * sys_sched_yield - yield the current processor to other threads.
3995 * this function yields the current CPU by moving the calling thread
3996 * to the expired array. If there are no other threads running on this
3997 * CPU then this function will return.
3999 asmlinkage
long sys_sched_yield(void)
4001 runqueue_t
*rq
= this_rq_lock();
4002 prio_array_t
*array
= current
->array
;
4003 prio_array_t
*target
= rq
->expired
;
4005 schedstat_inc(rq
, yld_cnt
);
4007 * We implement yielding by moving the task into the expired
4010 * (special rule: RT tasks will just roundrobin in the active
4013 if (rt_task(current
))
4014 target
= rq
->active
;
4016 if (array
->nr_active
== 1) {
4017 schedstat_inc(rq
, yld_act_empty
);
4018 if (!rq
->expired
->nr_active
)
4019 schedstat_inc(rq
, yld_both_empty
);
4020 } else if (!rq
->expired
->nr_active
)
4021 schedstat_inc(rq
, yld_exp_empty
);
4023 if (array
!= target
) {
4024 dequeue_task(current
, array
);
4025 enqueue_task(current
, target
);
4028 * requeue_task is cheaper so perform that if possible.
4030 requeue_task(current
, array
);
4033 * Since we are going to call schedule() anyway, there's
4034 * no need to preempt or enable interrupts:
4036 __release(rq
->lock
);
4037 _raw_spin_unlock(&rq
->lock
);
4038 preempt_enable_no_resched();
4045 static inline void __cond_resched(void)
4048 * The BKS might be reacquired before we have dropped
4049 * PREEMPT_ACTIVE, which could trigger a second
4050 * cond_resched() call.
4052 if (unlikely(preempt_count()))
4054 if (unlikely(system_state
!= SYSTEM_RUNNING
))
4057 add_preempt_count(PREEMPT_ACTIVE
);
4059 sub_preempt_count(PREEMPT_ACTIVE
);
4060 } while (need_resched());
4063 int __sched
cond_resched(void)
4065 if (need_resched()) {
4072 EXPORT_SYMBOL(cond_resched
);
4075 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4076 * call schedule, and on return reacquire the lock.
4078 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4079 * operations here to prevent schedule() from being called twice (once via
4080 * spin_unlock(), once by hand).
4082 int cond_resched_lock(spinlock_t
*lock
)
4086 if (need_lockbreak(lock
)) {
4092 if (need_resched()) {
4093 _raw_spin_unlock(lock
);
4094 preempt_enable_no_resched();
4102 EXPORT_SYMBOL(cond_resched_lock
);
4104 int __sched
cond_resched_softirq(void)
4106 BUG_ON(!in_softirq());
4108 if (need_resched()) {
4109 __local_bh_enable();
4117 EXPORT_SYMBOL(cond_resched_softirq
);
4121 * yield - yield the current processor to other threads.
4123 * this is a shortcut for kernel-space yielding - it marks the
4124 * thread runnable and calls sys_sched_yield().
4126 void __sched
yield(void)
4128 set_current_state(TASK_RUNNING
);
4132 EXPORT_SYMBOL(yield
);
4135 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4136 * that process accounting knows that this is a task in IO wait state.
4138 * But don't do that if it is a deliberate, throttling IO wait (this task
4139 * has set its backing_dev_info: the queue against which it should throttle)
4141 void __sched
io_schedule(void)
4143 struct runqueue
*rq
= &per_cpu(runqueues
, raw_smp_processor_id());
4145 atomic_inc(&rq
->nr_iowait
);
4147 atomic_dec(&rq
->nr_iowait
);
4150 EXPORT_SYMBOL(io_schedule
);
4152 long __sched
io_schedule_timeout(long timeout
)
4154 struct runqueue
*rq
= &per_cpu(runqueues
, raw_smp_processor_id());
4157 atomic_inc(&rq
->nr_iowait
);
4158 ret
= schedule_timeout(timeout
);
4159 atomic_dec(&rq
->nr_iowait
);
4164 * sys_sched_get_priority_max - return maximum RT priority.
4165 * @policy: scheduling class.
4167 * this syscall returns the maximum rt_priority that can be used
4168 * by a given scheduling class.
4170 asmlinkage
long sys_sched_get_priority_max(int policy
)
4177 ret
= MAX_USER_RT_PRIO
-1;
4188 * sys_sched_get_priority_min - return minimum RT priority.
4189 * @policy: scheduling class.
4191 * this syscall returns the minimum rt_priority that can be used
4192 * by a given scheduling class.
4194 asmlinkage
long sys_sched_get_priority_min(int policy
)
4211 * sys_sched_rr_get_interval - return the default timeslice of a process.
4212 * @pid: pid of the process.
4213 * @interval: userspace pointer to the timeslice value.
4215 * this syscall writes the default timeslice value of a given process
4216 * into the user-space timespec buffer. A value of '0' means infinity.
4219 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4221 int retval
= -EINVAL
;
4229 read_lock(&tasklist_lock
);
4230 p
= find_process_by_pid(pid
);
4234 retval
= security_task_getscheduler(p
);
4238 jiffies_to_timespec(p
->policy
& SCHED_FIFO
?
4239 0 : task_timeslice(p
), &t
);
4240 read_unlock(&tasklist_lock
);
4241 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4245 read_unlock(&tasklist_lock
);
4249 static inline struct task_struct
*eldest_child(struct task_struct
*p
)
4251 if (list_empty(&p
->children
)) return NULL
;
4252 return list_entry(p
->children
.next
,struct task_struct
,sibling
);
4255 static inline struct task_struct
*older_sibling(struct task_struct
*p
)
4257 if (p
->sibling
.prev
==&p
->parent
->children
) return NULL
;
4258 return list_entry(p
->sibling
.prev
,struct task_struct
,sibling
);
4261 static inline struct task_struct
*younger_sibling(struct task_struct
*p
)
4263 if (p
->sibling
.next
==&p
->parent
->children
) return NULL
;
4264 return list_entry(p
->sibling
.next
,struct task_struct
,sibling
);
4267 static void show_task(task_t
*p
)
4271 unsigned long free
= 0;
4272 static const char *stat_nam
[] = { "R", "S", "D", "T", "t", "Z", "X" };
4274 printk("%-13.13s ", p
->comm
);
4275 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4276 if (state
< ARRAY_SIZE(stat_nam
))
4277 printk(stat_nam
[state
]);
4280 #if (BITS_PER_LONG == 32)
4281 if (state
== TASK_RUNNING
)
4282 printk(" running ");
4284 printk(" %08lX ", thread_saved_pc(p
));
4286 if (state
== TASK_RUNNING
)
4287 printk(" running task ");
4289 printk(" %016lx ", thread_saved_pc(p
));
4291 #ifdef CONFIG_DEBUG_STACK_USAGE
4293 unsigned long *n
= end_of_stack(p
);
4296 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4299 printk("%5lu %5d %6d ", free
, p
->pid
, p
->parent
->pid
);
4300 if ((relative
= eldest_child(p
)))
4301 printk("%5d ", relative
->pid
);
4304 if ((relative
= younger_sibling(p
)))
4305 printk("%7d", relative
->pid
);
4308 if ((relative
= older_sibling(p
)))
4309 printk(" %5d", relative
->pid
);
4313 printk(" (L-TLB)\n");
4315 printk(" (NOTLB)\n");
4317 if (state
!= TASK_RUNNING
)
4318 show_stack(p
, NULL
);
4321 void show_state(void)
4325 #if (BITS_PER_LONG == 32)
4328 printk(" task PC pid father child younger older\n");
4332 printk(" task PC pid father child younger older\n");
4334 read_lock(&tasklist_lock
);
4335 do_each_thread(g
, p
) {
4337 * reset the NMI-timeout, listing all files on a slow
4338 * console might take alot of time:
4340 touch_nmi_watchdog();
4342 } while_each_thread(g
, p
);
4344 read_unlock(&tasklist_lock
);
4345 mutex_debug_show_all_locks();
4349 * init_idle - set up an idle thread for a given CPU
4350 * @idle: task in question
4351 * @cpu: cpu the idle task belongs to
4353 * NOTE: this function does not set the idle thread's NEED_RESCHED
4354 * flag, to make booting more robust.
4356 void __devinit
init_idle(task_t
*idle
, int cpu
)
4358 runqueue_t
*rq
= cpu_rq(cpu
);
4359 unsigned long flags
;
4361 idle
->timestamp
= sched_clock();
4362 idle
->sleep_avg
= 0;
4364 idle
->prio
= MAX_PRIO
;
4365 idle
->state
= TASK_RUNNING
;
4366 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4367 set_task_cpu(idle
, cpu
);
4369 spin_lock_irqsave(&rq
->lock
, flags
);
4370 rq
->curr
= rq
->idle
= idle
;
4371 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4374 spin_unlock_irqrestore(&rq
->lock
, flags
);
4376 /* Set the preempt count _outside_ the spinlocks! */
4377 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4378 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4380 task_thread_info(idle
)->preempt_count
= 0;
4385 * In a system that switches off the HZ timer nohz_cpu_mask
4386 * indicates which cpus entered this state. This is used
4387 * in the rcu update to wait only for active cpus. For system
4388 * which do not switch off the HZ timer nohz_cpu_mask should
4389 * always be CPU_MASK_NONE.
4391 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4395 * This is how migration works:
4397 * 1) we queue a migration_req_t structure in the source CPU's
4398 * runqueue and wake up that CPU's migration thread.
4399 * 2) we down() the locked semaphore => thread blocks.
4400 * 3) migration thread wakes up (implicitly it forces the migrated
4401 * thread off the CPU)
4402 * 4) it gets the migration request and checks whether the migrated
4403 * task is still in the wrong runqueue.
4404 * 5) if it's in the wrong runqueue then the migration thread removes
4405 * it and puts it into the right queue.
4406 * 6) migration thread up()s the semaphore.
4407 * 7) we wake up and the migration is done.
4411 * Change a given task's CPU affinity. Migrate the thread to a
4412 * proper CPU and schedule it away if the CPU it's executing on
4413 * is removed from the allowed bitmask.
4415 * NOTE: the caller must have a valid reference to the task, the
4416 * task must not exit() & deallocate itself prematurely. The
4417 * call is not atomic; no spinlocks may be held.
4419 int set_cpus_allowed(task_t
*p
, cpumask_t new_mask
)
4421 unsigned long flags
;
4423 migration_req_t req
;
4426 rq
= task_rq_lock(p
, &flags
);
4427 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4432 p
->cpus_allowed
= new_mask
;
4433 /* Can the task run on the task's current CPU? If so, we're done */
4434 if (cpu_isset(task_cpu(p
), new_mask
))
4437 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4438 /* Need help from migration thread: drop lock and wait. */
4439 task_rq_unlock(rq
, &flags
);
4440 wake_up_process(rq
->migration_thread
);
4441 wait_for_completion(&req
.done
);
4442 tlb_migrate_finish(p
->mm
);
4446 task_rq_unlock(rq
, &flags
);
4450 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4453 * Move (not current) task off this cpu, onto dest cpu. We're doing
4454 * this because either it can't run here any more (set_cpus_allowed()
4455 * away from this CPU, or CPU going down), or because we're
4456 * attempting to rebalance this task on exec (sched_exec).
4458 * So we race with normal scheduler movements, but that's OK, as long
4459 * as the task is no longer on this CPU.
4461 static void __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4463 runqueue_t
*rq_dest
, *rq_src
;
4465 if (unlikely(cpu_is_offline(dest_cpu
)))
4468 rq_src
= cpu_rq(src_cpu
);
4469 rq_dest
= cpu_rq(dest_cpu
);
4471 double_rq_lock(rq_src
, rq_dest
);
4472 /* Already moved. */
4473 if (task_cpu(p
) != src_cpu
)
4475 /* Affinity changed (again). */
4476 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
4479 set_task_cpu(p
, dest_cpu
);
4482 * Sync timestamp with rq_dest's before activating.
4483 * The same thing could be achieved by doing this step
4484 * afterwards, and pretending it was a local activate.
4485 * This way is cleaner and logically correct.
4487 p
->timestamp
= p
->timestamp
- rq_src
->timestamp_last_tick
4488 + rq_dest
->timestamp_last_tick
;
4489 deactivate_task(p
, rq_src
);
4490 activate_task(p
, rq_dest
, 0);
4491 if (TASK_PREEMPTS_CURR(p
, rq_dest
))
4492 resched_task(rq_dest
->curr
);
4496 double_rq_unlock(rq_src
, rq_dest
);
4500 * migration_thread - this is a highprio system thread that performs
4501 * thread migration by bumping thread off CPU then 'pushing' onto
4504 static int migration_thread(void *data
)
4507 int cpu
= (long)data
;
4510 BUG_ON(rq
->migration_thread
!= current
);
4512 set_current_state(TASK_INTERRUPTIBLE
);
4513 while (!kthread_should_stop()) {
4514 struct list_head
*head
;
4515 migration_req_t
*req
;
4519 spin_lock_irq(&rq
->lock
);
4521 if (cpu_is_offline(cpu
)) {
4522 spin_unlock_irq(&rq
->lock
);
4526 if (rq
->active_balance
) {
4527 active_load_balance(rq
, cpu
);
4528 rq
->active_balance
= 0;
4531 head
= &rq
->migration_queue
;
4533 if (list_empty(head
)) {
4534 spin_unlock_irq(&rq
->lock
);
4536 set_current_state(TASK_INTERRUPTIBLE
);
4539 req
= list_entry(head
->next
, migration_req_t
, list
);
4540 list_del_init(head
->next
);
4542 spin_unlock(&rq
->lock
);
4543 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
4546 complete(&req
->done
);
4548 __set_current_state(TASK_RUNNING
);
4552 /* Wait for kthread_stop */
4553 set_current_state(TASK_INTERRUPTIBLE
);
4554 while (!kthread_should_stop()) {
4556 set_current_state(TASK_INTERRUPTIBLE
);
4558 __set_current_state(TASK_RUNNING
);
4562 #ifdef CONFIG_HOTPLUG_CPU
4563 /* Figure out where task on dead CPU should go, use force if neccessary. */
4564 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*tsk
)
4570 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
4571 cpus_and(mask
, mask
, tsk
->cpus_allowed
);
4572 dest_cpu
= any_online_cpu(mask
);
4574 /* On any allowed CPU? */
4575 if (dest_cpu
== NR_CPUS
)
4576 dest_cpu
= any_online_cpu(tsk
->cpus_allowed
);
4578 /* No more Mr. Nice Guy. */
4579 if (dest_cpu
== NR_CPUS
) {
4580 cpus_setall(tsk
->cpus_allowed
);
4581 dest_cpu
= any_online_cpu(tsk
->cpus_allowed
);
4584 * Don't tell them about moving exiting tasks or
4585 * kernel threads (both mm NULL), since they never
4588 if (tsk
->mm
&& printk_ratelimit())
4589 printk(KERN_INFO
"process %d (%s) no "
4590 "longer affine to cpu%d\n",
4591 tsk
->pid
, tsk
->comm
, dead_cpu
);
4593 __migrate_task(tsk
, dead_cpu
, dest_cpu
);
4597 * While a dead CPU has no uninterruptible tasks queued at this point,
4598 * it might still have a nonzero ->nr_uninterruptible counter, because
4599 * for performance reasons the counter is not stricly tracking tasks to
4600 * their home CPUs. So we just add the counter to another CPU's counter,
4601 * to keep the global sum constant after CPU-down:
4603 static void migrate_nr_uninterruptible(runqueue_t
*rq_src
)
4605 runqueue_t
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
4606 unsigned long flags
;
4608 local_irq_save(flags
);
4609 double_rq_lock(rq_src
, rq_dest
);
4610 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
4611 rq_src
->nr_uninterruptible
= 0;
4612 double_rq_unlock(rq_src
, rq_dest
);
4613 local_irq_restore(flags
);
4616 /* Run through task list and migrate tasks from the dead cpu. */
4617 static void migrate_live_tasks(int src_cpu
)
4619 struct task_struct
*tsk
, *t
;
4621 write_lock_irq(&tasklist_lock
);
4623 do_each_thread(t
, tsk
) {
4627 if (task_cpu(tsk
) == src_cpu
)
4628 move_task_off_dead_cpu(src_cpu
, tsk
);
4629 } while_each_thread(t
, tsk
);
4631 write_unlock_irq(&tasklist_lock
);
4634 /* Schedules idle task to be the next runnable task on current CPU.
4635 * It does so by boosting its priority to highest possible and adding it to
4636 * the _front_ of runqueue. Used by CPU offline code.
4638 void sched_idle_next(void)
4640 int cpu
= smp_processor_id();
4641 runqueue_t
*rq
= this_rq();
4642 struct task_struct
*p
= rq
->idle
;
4643 unsigned long flags
;
4645 /* cpu has to be offline */
4646 BUG_ON(cpu_online(cpu
));
4648 /* Strictly not necessary since rest of the CPUs are stopped by now
4649 * and interrupts disabled on current cpu.
4651 spin_lock_irqsave(&rq
->lock
, flags
);
4653 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
4654 /* Add idle task to _front_ of it's priority queue */
4655 __activate_idle_task(p
, rq
);
4657 spin_unlock_irqrestore(&rq
->lock
, flags
);
4660 /* Ensures that the idle task is using init_mm right before its cpu goes
4663 void idle_task_exit(void)
4665 struct mm_struct
*mm
= current
->active_mm
;
4667 BUG_ON(cpu_online(smp_processor_id()));
4670 switch_mm(mm
, &init_mm
, current
);
4674 static void migrate_dead(unsigned int dead_cpu
, task_t
*tsk
)
4676 struct runqueue
*rq
= cpu_rq(dead_cpu
);
4678 /* Must be exiting, otherwise would be on tasklist. */
4679 BUG_ON(tsk
->exit_state
!= EXIT_ZOMBIE
&& tsk
->exit_state
!= EXIT_DEAD
);
4681 /* Cannot have done final schedule yet: would have vanished. */
4682 BUG_ON(tsk
->flags
& PF_DEAD
);
4684 get_task_struct(tsk
);
4687 * Drop lock around migration; if someone else moves it,
4688 * that's OK. No task can be added to this CPU, so iteration is
4691 spin_unlock_irq(&rq
->lock
);
4692 move_task_off_dead_cpu(dead_cpu
, tsk
);
4693 spin_lock_irq(&rq
->lock
);
4695 put_task_struct(tsk
);
4698 /* release_task() removes task from tasklist, so we won't find dead tasks. */
4699 static void migrate_dead_tasks(unsigned int dead_cpu
)
4702 struct runqueue
*rq
= cpu_rq(dead_cpu
);
4704 for (arr
= 0; arr
< 2; arr
++) {
4705 for (i
= 0; i
< MAX_PRIO
; i
++) {
4706 struct list_head
*list
= &rq
->arrays
[arr
].queue
[i
];
4707 while (!list_empty(list
))
4708 migrate_dead(dead_cpu
,
4709 list_entry(list
->next
, task_t
,
4714 #endif /* CONFIG_HOTPLUG_CPU */
4717 * migration_call - callback that gets triggered when a CPU is added.
4718 * Here we can start up the necessary migration thread for the new CPU.
4720 static int migration_call(struct notifier_block
*nfb
, unsigned long action
,
4723 int cpu
= (long)hcpu
;
4724 struct task_struct
*p
;
4725 struct runqueue
*rq
;
4726 unsigned long flags
;
4729 case CPU_UP_PREPARE
:
4730 p
= kthread_create(migration_thread
, hcpu
, "migration/%d",cpu
);
4733 p
->flags
|= PF_NOFREEZE
;
4734 kthread_bind(p
, cpu
);
4735 /* Must be high prio: stop_machine expects to yield to it. */
4736 rq
= task_rq_lock(p
, &flags
);
4737 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
4738 task_rq_unlock(rq
, &flags
);
4739 cpu_rq(cpu
)->migration_thread
= p
;
4742 /* Strictly unneccessary, as first user will wake it. */
4743 wake_up_process(cpu_rq(cpu
)->migration_thread
);
4745 #ifdef CONFIG_HOTPLUG_CPU
4746 case CPU_UP_CANCELED
:
4747 /* Unbind it from offline cpu so it can run. Fall thru. */
4748 kthread_bind(cpu_rq(cpu
)->migration_thread
,
4749 any_online_cpu(cpu_online_map
));
4750 kthread_stop(cpu_rq(cpu
)->migration_thread
);
4751 cpu_rq(cpu
)->migration_thread
= NULL
;
4754 migrate_live_tasks(cpu
);
4756 kthread_stop(rq
->migration_thread
);
4757 rq
->migration_thread
= NULL
;
4758 /* Idle task back to normal (off runqueue, low prio) */
4759 rq
= task_rq_lock(rq
->idle
, &flags
);
4760 deactivate_task(rq
->idle
, rq
);
4761 rq
->idle
->static_prio
= MAX_PRIO
;
4762 __setscheduler(rq
->idle
, SCHED_NORMAL
, 0);
4763 migrate_dead_tasks(cpu
);
4764 task_rq_unlock(rq
, &flags
);
4765 migrate_nr_uninterruptible(rq
);
4766 BUG_ON(rq
->nr_running
!= 0);
4768 /* No need to migrate the tasks: it was best-effort if
4769 * they didn't do lock_cpu_hotplug(). Just wake up
4770 * the requestors. */
4771 spin_lock_irq(&rq
->lock
);
4772 while (!list_empty(&rq
->migration_queue
)) {
4773 migration_req_t
*req
;
4774 req
= list_entry(rq
->migration_queue
.next
,
4775 migration_req_t
, list
);
4776 list_del_init(&req
->list
);
4777 complete(&req
->done
);
4779 spin_unlock_irq(&rq
->lock
);
4786 /* Register at highest priority so that task migration (migrate_all_tasks)
4787 * happens before everything else.
4789 static struct notifier_block __devinitdata migration_notifier
= {
4790 .notifier_call
= migration_call
,
4794 int __init
migration_init(void)
4796 void *cpu
= (void *)(long)smp_processor_id();
4797 /* Start one for boot CPU. */
4798 migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
4799 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
4800 register_cpu_notifier(&migration_notifier
);
4806 #undef SCHED_DOMAIN_DEBUG
4807 #ifdef SCHED_DOMAIN_DEBUG
4808 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
4813 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
4817 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
4822 struct sched_group
*group
= sd
->groups
;
4823 cpumask_t groupmask
;
4825 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
4826 cpus_clear(groupmask
);
4829 for (i
= 0; i
< level
+ 1; i
++)
4831 printk("domain %d: ", level
);
4833 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
4834 printk("does not load-balance\n");
4836 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain has parent");
4840 printk("span %s\n", str
);
4842 if (!cpu_isset(cpu
, sd
->span
))
4843 printk(KERN_ERR
"ERROR: domain->span does not contain CPU%d\n", cpu
);
4844 if (!cpu_isset(cpu
, group
->cpumask
))
4845 printk(KERN_ERR
"ERROR: domain->groups does not contain CPU%d\n", cpu
);
4848 for (i
= 0; i
< level
+ 2; i
++)
4854 printk(KERN_ERR
"ERROR: group is NULL\n");
4858 if (!group
->cpu_power
) {
4860 printk(KERN_ERR
"ERROR: domain->cpu_power not set\n");
4863 if (!cpus_weight(group
->cpumask
)) {
4865 printk(KERN_ERR
"ERROR: empty group\n");
4868 if (cpus_intersects(groupmask
, group
->cpumask
)) {
4870 printk(KERN_ERR
"ERROR: repeated CPUs\n");
4873 cpus_or(groupmask
, groupmask
, group
->cpumask
);
4875 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
4878 group
= group
->next
;
4879 } while (group
!= sd
->groups
);
4882 if (!cpus_equal(sd
->span
, groupmask
))
4883 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
4889 if (!cpus_subset(groupmask
, sd
->span
))
4890 printk(KERN_ERR
"ERROR: parent span is not a superset of domain->span\n");
4896 #define sched_domain_debug(sd, cpu) {}
4899 static int sd_degenerate(struct sched_domain
*sd
)
4901 if (cpus_weight(sd
->span
) == 1)
4904 /* Following flags need at least 2 groups */
4905 if (sd
->flags
& (SD_LOAD_BALANCE
|
4906 SD_BALANCE_NEWIDLE
|
4909 if (sd
->groups
!= sd
->groups
->next
)
4913 /* Following flags don't use groups */
4914 if (sd
->flags
& (SD_WAKE_IDLE
|
4922 static int sd_parent_degenerate(struct sched_domain
*sd
,
4923 struct sched_domain
*parent
)
4925 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
4927 if (sd_degenerate(parent
))
4930 if (!cpus_equal(sd
->span
, parent
->span
))
4933 /* Does parent contain flags not in child? */
4934 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
4935 if (cflags
& SD_WAKE_AFFINE
)
4936 pflags
&= ~SD_WAKE_BALANCE
;
4937 /* Flags needing groups don't count if only 1 group in parent */
4938 if (parent
->groups
== parent
->groups
->next
) {
4939 pflags
&= ~(SD_LOAD_BALANCE
|
4940 SD_BALANCE_NEWIDLE
|
4944 if (~cflags
& pflags
)
4951 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
4952 * hold the hotplug lock.
4954 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
4956 runqueue_t
*rq
= cpu_rq(cpu
);
4957 struct sched_domain
*tmp
;
4959 /* Remove the sched domains which do not contribute to scheduling. */
4960 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
4961 struct sched_domain
*parent
= tmp
->parent
;
4964 if (sd_parent_degenerate(tmp
, parent
))
4965 tmp
->parent
= parent
->parent
;
4968 if (sd
&& sd_degenerate(sd
))
4971 sched_domain_debug(sd
, cpu
);
4973 rcu_assign_pointer(rq
->sd
, sd
);
4976 /* cpus with isolated domains */
4977 static cpumask_t __devinitdata cpu_isolated_map
= CPU_MASK_NONE
;
4979 /* Setup the mask of cpus configured for isolated domains */
4980 static int __init
isolated_cpu_setup(char *str
)
4982 int ints
[NR_CPUS
], i
;
4984 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
4985 cpus_clear(cpu_isolated_map
);
4986 for (i
= 1; i
<= ints
[0]; i
++)
4987 if (ints
[i
] < NR_CPUS
)
4988 cpu_set(ints
[i
], cpu_isolated_map
);
4992 __setup ("isolcpus=", isolated_cpu_setup
);
4995 * init_sched_build_groups takes an array of groups, the cpumask we wish
4996 * to span, and a pointer to a function which identifies what group a CPU
4997 * belongs to. The return value of group_fn must be a valid index into the
4998 * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
4999 * keep track of groups covered with a cpumask_t).
5001 * init_sched_build_groups will build a circular linked list of the groups
5002 * covered by the given span, and will set each group's ->cpumask correctly,
5003 * and ->cpu_power to 0.
5005 static void init_sched_build_groups(struct sched_group groups
[], cpumask_t span
,
5006 int (*group_fn
)(int cpu
))
5008 struct sched_group
*first
= NULL
, *last
= NULL
;
5009 cpumask_t covered
= CPU_MASK_NONE
;
5012 for_each_cpu_mask(i
, span
) {
5013 int group
= group_fn(i
);
5014 struct sched_group
*sg
= &groups
[group
];
5017 if (cpu_isset(i
, covered
))
5020 sg
->cpumask
= CPU_MASK_NONE
;
5023 for_each_cpu_mask(j
, span
) {
5024 if (group_fn(j
) != group
)
5027 cpu_set(j
, covered
);
5028 cpu_set(j
, sg
->cpumask
);
5039 #define SD_NODES_PER_DOMAIN 16
5042 * Self-tuning task migration cost measurement between source and target CPUs.
5044 * This is done by measuring the cost of manipulating buffers of varying
5045 * sizes. For a given buffer-size here are the steps that are taken:
5047 * 1) the source CPU reads+dirties a shared buffer
5048 * 2) the target CPU reads+dirties the same shared buffer
5050 * We measure how long they take, in the following 4 scenarios:
5052 * - source: CPU1, target: CPU2 | cost1
5053 * - source: CPU2, target: CPU1 | cost2
5054 * - source: CPU1, target: CPU1 | cost3
5055 * - source: CPU2, target: CPU2 | cost4
5057 * We then calculate the cost3+cost4-cost1-cost2 difference - this is
5058 * the cost of migration.
5060 * We then start off from a small buffer-size and iterate up to larger
5061 * buffer sizes, in 5% steps - measuring each buffer-size separately, and
5062 * doing a maximum search for the cost. (The maximum cost for a migration
5063 * normally occurs when the working set size is around the effective cache
5066 #define SEARCH_SCOPE 2
5067 #define MIN_CACHE_SIZE (64*1024U)
5068 #define DEFAULT_CACHE_SIZE (5*1024*1024U)
5069 #define ITERATIONS 1
5070 #define SIZE_THRESH 130
5071 #define COST_THRESH 130
5074 * The migration cost is a function of 'domain distance'. Domain
5075 * distance is the number of steps a CPU has to iterate down its
5076 * domain tree to share a domain with the other CPU. The farther
5077 * two CPUs are from each other, the larger the distance gets.
5079 * Note that we use the distance only to cache measurement results,
5080 * the distance value is not used numerically otherwise. When two
5081 * CPUs have the same distance it is assumed that the migration
5082 * cost is the same. (this is a simplification but quite practical)
5084 #define MAX_DOMAIN_DISTANCE 32
5086 static unsigned long long migration_cost
[MAX_DOMAIN_DISTANCE
] =
5087 { [ 0 ... MAX_DOMAIN_DISTANCE
-1 ] =
5089 * Architectures may override the migration cost and thus avoid
5090 * boot-time calibration. Unit is nanoseconds. Mostly useful for
5091 * virtualized hardware:
5093 #ifdef CONFIG_DEFAULT_MIGRATION_COST
5094 CONFIG_DEFAULT_MIGRATION_COST
5101 * Allow override of migration cost - in units of microseconds.
5102 * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost
5103 * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs:
5105 static int __init
migration_cost_setup(char *str
)
5107 int ints
[MAX_DOMAIN_DISTANCE
+1], i
;
5109 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5111 printk("#ints: %d\n", ints
[0]);
5112 for (i
= 1; i
<= ints
[0]; i
++) {
5113 migration_cost
[i
-1] = (unsigned long long)ints
[i
]*1000;
5114 printk("migration_cost[%d]: %Ld\n", i
-1, migration_cost
[i
-1]);
5119 __setup ("migration_cost=", migration_cost_setup
);
5122 * Global multiplier (divisor) for migration-cutoff values,
5123 * in percentiles. E.g. use a value of 150 to get 1.5 times
5124 * longer cache-hot cutoff times.
5126 * (We scale it from 100 to 128 to long long handling easier.)
5129 #define MIGRATION_FACTOR_SCALE 128
5131 static unsigned int migration_factor
= MIGRATION_FACTOR_SCALE
;
5133 static int __init
setup_migration_factor(char *str
)
5135 get_option(&str
, &migration_factor
);
5136 migration_factor
= migration_factor
* MIGRATION_FACTOR_SCALE
/ 100;
5140 __setup("migration_factor=", setup_migration_factor
);
5143 * Estimated distance of two CPUs, measured via the number of domains
5144 * we have to pass for the two CPUs to be in the same span:
5146 static unsigned long domain_distance(int cpu1
, int cpu2
)
5148 unsigned long distance
= 0;
5149 struct sched_domain
*sd
;
5151 for_each_domain(cpu1
, sd
) {
5152 WARN_ON(!cpu_isset(cpu1
, sd
->span
));
5153 if (cpu_isset(cpu2
, sd
->span
))
5157 if (distance
>= MAX_DOMAIN_DISTANCE
) {
5159 distance
= MAX_DOMAIN_DISTANCE
-1;
5165 static unsigned int migration_debug
;
5167 static int __init
setup_migration_debug(char *str
)
5169 get_option(&str
, &migration_debug
);
5173 __setup("migration_debug=", setup_migration_debug
);
5176 * Maximum cache-size that the scheduler should try to measure.
5177 * Architectures with larger caches should tune this up during
5178 * bootup. Gets used in the domain-setup code (i.e. during SMP
5181 unsigned int max_cache_size
;
5183 static int __init
setup_max_cache_size(char *str
)
5185 get_option(&str
, &max_cache_size
);
5189 __setup("max_cache_size=", setup_max_cache_size
);
5192 * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This
5193 * is the operation that is timed, so we try to generate unpredictable
5194 * cachemisses that still end up filling the L2 cache:
5196 static void touch_cache(void *__cache
, unsigned long __size
)
5198 unsigned long size
= __size
/sizeof(long), chunk1
= size
/3,
5200 unsigned long *cache
= __cache
;
5203 for (i
= 0; i
< size
/6; i
+= 8) {
5206 case 1: cache
[size
-1-i
]++;
5207 case 2: cache
[chunk1
-i
]++;
5208 case 3: cache
[chunk1
+i
]++;
5209 case 4: cache
[chunk2
-i
]++;
5210 case 5: cache
[chunk2
+i
]++;
5216 * Measure the cache-cost of one task migration. Returns in units of nsec.
5218 static unsigned long long measure_one(void *cache
, unsigned long size
,
5219 int source
, int target
)
5221 cpumask_t mask
, saved_mask
;
5222 unsigned long long t0
, t1
, t2
, t3
, cost
;
5224 saved_mask
= current
->cpus_allowed
;
5227 * Flush source caches to RAM and invalidate them:
5232 * Migrate to the source CPU:
5234 mask
= cpumask_of_cpu(source
);
5235 set_cpus_allowed(current
, mask
);
5236 WARN_ON(smp_processor_id() != source
);
5239 * Dirty the working set:
5242 touch_cache(cache
, size
);
5246 * Migrate to the target CPU, dirty the L2 cache and access
5247 * the shared buffer. (which represents the working set
5248 * of a migrated task.)
5250 mask
= cpumask_of_cpu(target
);
5251 set_cpus_allowed(current
, mask
);
5252 WARN_ON(smp_processor_id() != target
);
5255 touch_cache(cache
, size
);
5258 cost
= t1
-t0
+ t3
-t2
;
5260 if (migration_debug
>= 2)
5261 printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n",
5262 source
, target
, t1
-t0
, t1
-t0
, t3
-t2
, cost
);
5264 * Flush target caches to RAM and invalidate them:
5268 set_cpus_allowed(current
, saved_mask
);
5274 * Measure a series of task migrations and return the average
5275 * result. Since this code runs early during bootup the system
5276 * is 'undisturbed' and the average latency makes sense.
5278 * The algorithm in essence auto-detects the relevant cache-size,
5279 * so it will properly detect different cachesizes for different
5280 * cache-hierarchies, depending on how the CPUs are connected.
5282 * Architectures can prime the upper limit of the search range via
5283 * max_cache_size, otherwise the search range defaults to 20MB...64K.
5285 static unsigned long long
5286 measure_cost(int cpu1
, int cpu2
, void *cache
, unsigned int size
)
5288 unsigned long long cost1
, cost2
;
5292 * Measure the migration cost of 'size' bytes, over an
5293 * average of 10 runs:
5295 * (We perturb the cache size by a small (0..4k)
5296 * value to compensate size/alignment related artifacts.
5297 * We also subtract the cost of the operation done on
5303 * dry run, to make sure we start off cache-cold on cpu1,
5304 * and to get any vmalloc pagefaults in advance:
5306 measure_one(cache
, size
, cpu1
, cpu2
);
5307 for (i
= 0; i
< ITERATIONS
; i
++)
5308 cost1
+= measure_one(cache
, size
- i
*1024, cpu1
, cpu2
);
5310 measure_one(cache
, size
, cpu2
, cpu1
);
5311 for (i
= 0; i
< ITERATIONS
; i
++)
5312 cost1
+= measure_one(cache
, size
- i
*1024, cpu2
, cpu1
);
5315 * (We measure the non-migrating [cached] cost on both
5316 * cpu1 and cpu2, to handle CPUs with different speeds)
5320 measure_one(cache
, size
, cpu1
, cpu1
);
5321 for (i
= 0; i
< ITERATIONS
; i
++)
5322 cost2
+= measure_one(cache
, size
- i
*1024, cpu1
, cpu1
);
5324 measure_one(cache
, size
, cpu2
, cpu2
);
5325 for (i
= 0; i
< ITERATIONS
; i
++)
5326 cost2
+= measure_one(cache
, size
- i
*1024, cpu2
, cpu2
);
5329 * Get the per-iteration migration cost:
5331 do_div(cost1
, 2*ITERATIONS
);
5332 do_div(cost2
, 2*ITERATIONS
);
5334 return cost1
- cost2
;
5337 static unsigned long long measure_migration_cost(int cpu1
, int cpu2
)
5339 unsigned long long max_cost
= 0, fluct
= 0, avg_fluct
= 0;
5340 unsigned int max_size
, size
, size_found
= 0;
5341 long long cost
= 0, prev_cost
;
5345 * Search from max_cache_size*5 down to 64K - the real relevant
5346 * cachesize has to lie somewhere inbetween.
5348 if (max_cache_size
) {
5349 max_size
= max(max_cache_size
* SEARCH_SCOPE
, MIN_CACHE_SIZE
);
5350 size
= max(max_cache_size
/ SEARCH_SCOPE
, MIN_CACHE_SIZE
);
5353 * Since we have no estimation about the relevant
5356 max_size
= DEFAULT_CACHE_SIZE
* SEARCH_SCOPE
;
5357 size
= MIN_CACHE_SIZE
;
5360 if (!cpu_online(cpu1
) || !cpu_online(cpu2
)) {
5361 printk("cpu %d and %d not both online!\n", cpu1
, cpu2
);
5366 * Allocate the working set:
5368 cache
= vmalloc(max_size
);
5370 printk("could not vmalloc %d bytes for cache!\n", 2*max_size
);
5371 return 1000000; // return 1 msec on very small boxen
5374 while (size
<= max_size
) {
5376 cost
= measure_cost(cpu1
, cpu2
, cache
, size
);
5382 if (max_cost
< cost
) {
5388 * Calculate average fluctuation, we use this to prevent
5389 * noise from triggering an early break out of the loop:
5391 fluct
= abs(cost
- prev_cost
);
5392 avg_fluct
= (avg_fluct
+ fluct
)/2;
5394 if (migration_debug
)
5395 printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): (%8Ld %8Ld)\n",
5397 (long)cost
/ 1000000,
5398 ((long)cost
/ 100000) % 10,
5399 (long)max_cost
/ 1000000,
5400 ((long)max_cost
/ 100000) % 10,
5401 domain_distance(cpu1
, cpu2
),
5405 * If we iterated at least 20% past the previous maximum,
5406 * and the cost has dropped by more than 20% already,
5407 * (taking fluctuations into account) then we assume to
5408 * have found the maximum and break out of the loop early:
5410 if (size_found
&& (size
*100 > size_found
*SIZE_THRESH
))
5411 if (cost
+avg_fluct
<= 0 ||
5412 max_cost
*100 > (cost
+avg_fluct
)*COST_THRESH
) {
5414 if (migration_debug
)
5415 printk("-> found max.\n");
5419 * Increase the cachesize in 10% steps:
5421 size
= size
* 10 / 9;
5424 if (migration_debug
)
5425 printk("[%d][%d] working set size found: %d, cost: %Ld\n",
5426 cpu1
, cpu2
, size_found
, max_cost
);
5431 * A task is considered 'cache cold' if at least 2 times
5432 * the worst-case cost of migration has passed.
5434 * (this limit is only listened to if the load-balancing
5435 * situation is 'nice' - if there is a large imbalance we
5436 * ignore it for the sake of CPU utilization and
5437 * processing fairness.)
5439 return 2 * max_cost
* migration_factor
/ MIGRATION_FACTOR_SCALE
;
5442 static void calibrate_migration_costs(const cpumask_t
*cpu_map
)
5444 int cpu1
= -1, cpu2
= -1, cpu
, orig_cpu
= raw_smp_processor_id();
5445 unsigned long j0
, j1
, distance
, max_distance
= 0;
5446 struct sched_domain
*sd
;
5451 * First pass - calculate the cacheflush times:
5453 for_each_cpu_mask(cpu1
, *cpu_map
) {
5454 for_each_cpu_mask(cpu2
, *cpu_map
) {
5457 distance
= domain_distance(cpu1
, cpu2
);
5458 max_distance
= max(max_distance
, distance
);
5460 * No result cached yet?
5462 if (migration_cost
[distance
] == -1LL)
5463 migration_cost
[distance
] =
5464 measure_migration_cost(cpu1
, cpu2
);
5468 * Second pass - update the sched domain hierarchy with
5469 * the new cache-hot-time estimations:
5471 for_each_cpu_mask(cpu
, *cpu_map
) {
5473 for_each_domain(cpu
, sd
) {
5474 sd
->cache_hot_time
= migration_cost
[distance
];
5481 if (migration_debug
)
5482 printk("migration: max_cache_size: %d, cpu: %d MHz:\n",
5490 if (system_state
== SYSTEM_BOOTING
) {
5491 printk("migration_cost=");
5492 for (distance
= 0; distance
<= max_distance
; distance
++) {
5495 printk("%ld", (long)migration_cost
[distance
] / 1000);
5500 if (migration_debug
)
5501 printk("migration: %ld seconds\n", (j1
-j0
)/HZ
);
5504 * Move back to the original CPU. NUMA-Q gets confused
5505 * if we migrate to another quad during bootup.
5507 if (raw_smp_processor_id() != orig_cpu
) {
5508 cpumask_t mask
= cpumask_of_cpu(orig_cpu
),
5509 saved_mask
= current
->cpus_allowed
;
5511 set_cpus_allowed(current
, mask
);
5512 set_cpus_allowed(current
, saved_mask
);
5519 * find_next_best_node - find the next node to include in a sched_domain
5520 * @node: node whose sched_domain we're building
5521 * @used_nodes: nodes already in the sched_domain
5523 * Find the next node to include in a given scheduling domain. Simply
5524 * finds the closest node not already in the @used_nodes map.
5526 * Should use nodemask_t.
5528 static int find_next_best_node(int node
, unsigned long *used_nodes
)
5530 int i
, n
, val
, min_val
, best_node
= 0;
5534 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5535 /* Start at @node */
5536 n
= (node
+ i
) % MAX_NUMNODES
;
5538 if (!nr_cpus_node(n
))
5541 /* Skip already used nodes */
5542 if (test_bit(n
, used_nodes
))
5545 /* Simple min distance search */
5546 val
= node_distance(node
, n
);
5548 if (val
< min_val
) {
5554 set_bit(best_node
, used_nodes
);
5559 * sched_domain_node_span - get a cpumask for a node's sched_domain
5560 * @node: node whose cpumask we're constructing
5561 * @size: number of nodes to include in this span
5563 * Given a node, construct a good cpumask for its sched_domain to span. It
5564 * should be one that prevents unnecessary balancing, but also spreads tasks
5567 static cpumask_t
sched_domain_node_span(int node
)
5570 cpumask_t span
, nodemask
;
5571 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
5574 bitmap_zero(used_nodes
, MAX_NUMNODES
);
5576 nodemask
= node_to_cpumask(node
);
5577 cpus_or(span
, span
, nodemask
);
5578 set_bit(node
, used_nodes
);
5580 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5581 int next_node
= find_next_best_node(node
, used_nodes
);
5582 nodemask
= node_to_cpumask(next_node
);
5583 cpus_or(span
, span
, nodemask
);
5591 * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we
5592 * can switch it on easily if needed.
5594 #ifdef CONFIG_SCHED_SMT
5595 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
5596 static struct sched_group sched_group_cpus
[NR_CPUS
];
5597 static int cpu_to_cpu_group(int cpu
)
5603 #ifdef CONFIG_SCHED_MC
5604 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
5605 static struct sched_group sched_group_core
[NR_CPUS
];
5608 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5609 static int cpu_to_core_group(int cpu
)
5611 return first_cpu(cpu_sibling_map
[cpu
]);
5613 #elif defined(CONFIG_SCHED_MC)
5614 static int cpu_to_core_group(int cpu
)
5620 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
5621 static struct sched_group sched_group_phys
[NR_CPUS
];
5622 static int cpu_to_phys_group(int cpu
)
5624 #if defined(CONFIG_SCHED_MC)
5625 cpumask_t mask
= cpu_coregroup_map(cpu
);
5626 return first_cpu(mask
);
5627 #elif defined(CONFIG_SCHED_SMT)
5628 return first_cpu(cpu_sibling_map
[cpu
]);
5636 * The init_sched_build_groups can't handle what we want to do with node
5637 * groups, so roll our own. Now each node has its own list of groups which
5638 * gets dynamically allocated.
5640 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
5641 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
5643 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
5644 static struct sched_group
*sched_group_allnodes_bycpu
[NR_CPUS
];
5646 static int cpu_to_allnodes_group(int cpu
)
5648 return cpu_to_node(cpu
);
5650 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
5652 struct sched_group
*sg
= group_head
;
5658 for_each_cpu_mask(j
, sg
->cpumask
) {
5659 struct sched_domain
*sd
;
5661 sd
= &per_cpu(phys_domains
, j
);
5662 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
5664 * Only add "power" once for each
5670 sg
->cpu_power
+= sd
->groups
->cpu_power
;
5673 if (sg
!= group_head
)
5679 * Build sched domains for a given set of cpus and attach the sched domains
5680 * to the individual cpus
5682 void build_sched_domains(const cpumask_t
*cpu_map
)
5686 struct sched_group
**sched_group_nodes
= NULL
;
5687 struct sched_group
*sched_group_allnodes
= NULL
;
5690 * Allocate the per-node list of sched groups
5692 sched_group_nodes
= kmalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
5694 if (!sched_group_nodes
) {
5695 printk(KERN_WARNING
"Can not alloc sched group node list\n");
5698 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
5702 * Set up domains for cpus specified by the cpu_map.
5704 for_each_cpu_mask(i
, *cpu_map
) {
5706 struct sched_domain
*sd
= NULL
, *p
;
5707 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
5709 cpus_and(nodemask
, nodemask
, *cpu_map
);
5712 if (cpus_weight(*cpu_map
)
5713 > SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
5714 if (!sched_group_allnodes
) {
5715 sched_group_allnodes
5716 = kmalloc(sizeof(struct sched_group
)
5719 if (!sched_group_allnodes
) {
5721 "Can not alloc allnodes sched group\n");
5724 sched_group_allnodes_bycpu
[i
]
5725 = sched_group_allnodes
;
5727 sd
= &per_cpu(allnodes_domains
, i
);
5728 *sd
= SD_ALLNODES_INIT
;
5729 sd
->span
= *cpu_map
;
5730 group
= cpu_to_allnodes_group(i
);
5731 sd
->groups
= &sched_group_allnodes
[group
];
5736 sd
= &per_cpu(node_domains
, i
);
5738 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
5740 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
5744 sd
= &per_cpu(phys_domains
, i
);
5745 group
= cpu_to_phys_group(i
);
5747 sd
->span
= nodemask
;
5749 sd
->groups
= &sched_group_phys
[group
];
5751 #ifdef CONFIG_SCHED_MC
5753 sd
= &per_cpu(core_domains
, i
);
5754 group
= cpu_to_core_group(i
);
5756 sd
->span
= cpu_coregroup_map(i
);
5757 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
5759 sd
->groups
= &sched_group_core
[group
];
5762 #ifdef CONFIG_SCHED_SMT
5764 sd
= &per_cpu(cpu_domains
, i
);
5765 group
= cpu_to_cpu_group(i
);
5766 *sd
= SD_SIBLING_INIT
;
5767 sd
->span
= cpu_sibling_map
[i
];
5768 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
5770 sd
->groups
= &sched_group_cpus
[group
];
5774 #ifdef CONFIG_SCHED_SMT
5775 /* Set up CPU (sibling) groups */
5776 for_each_cpu_mask(i
, *cpu_map
) {
5777 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
5778 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
5779 if (i
!= first_cpu(this_sibling_map
))
5782 init_sched_build_groups(sched_group_cpus
, this_sibling_map
,
5787 #ifdef CONFIG_SCHED_MC
5788 /* Set up multi-core groups */
5789 for_each_cpu_mask(i
, *cpu_map
) {
5790 cpumask_t this_core_map
= cpu_coregroup_map(i
);
5791 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
5792 if (i
!= first_cpu(this_core_map
))
5794 init_sched_build_groups(sched_group_core
, this_core_map
,
5795 &cpu_to_core_group
);
5800 /* Set up physical groups */
5801 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5802 cpumask_t nodemask
= node_to_cpumask(i
);
5804 cpus_and(nodemask
, nodemask
, *cpu_map
);
5805 if (cpus_empty(nodemask
))
5808 init_sched_build_groups(sched_group_phys
, nodemask
,
5809 &cpu_to_phys_group
);
5813 /* Set up node groups */
5814 if (sched_group_allnodes
)
5815 init_sched_build_groups(sched_group_allnodes
, *cpu_map
,
5816 &cpu_to_allnodes_group
);
5818 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5819 /* Set up node groups */
5820 struct sched_group
*sg
, *prev
;
5821 cpumask_t nodemask
= node_to_cpumask(i
);
5822 cpumask_t domainspan
;
5823 cpumask_t covered
= CPU_MASK_NONE
;
5826 cpus_and(nodemask
, nodemask
, *cpu_map
);
5827 if (cpus_empty(nodemask
)) {
5828 sched_group_nodes
[i
] = NULL
;
5832 domainspan
= sched_domain_node_span(i
);
5833 cpus_and(domainspan
, domainspan
, *cpu_map
);
5835 sg
= kmalloc(sizeof(struct sched_group
), GFP_KERNEL
);
5836 sched_group_nodes
[i
] = sg
;
5837 for_each_cpu_mask(j
, nodemask
) {
5838 struct sched_domain
*sd
;
5839 sd
= &per_cpu(node_domains
, j
);
5841 if (sd
->groups
== NULL
) {
5842 /* Turn off balancing if we have no groups */
5848 "Can not alloc domain group for node %d\n", i
);
5852 sg
->cpumask
= nodemask
;
5853 cpus_or(covered
, covered
, nodemask
);
5856 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
5857 cpumask_t tmp
, notcovered
;
5858 int n
= (i
+ j
) % MAX_NUMNODES
;
5860 cpus_complement(notcovered
, covered
);
5861 cpus_and(tmp
, notcovered
, *cpu_map
);
5862 cpus_and(tmp
, tmp
, domainspan
);
5863 if (cpus_empty(tmp
))
5866 nodemask
= node_to_cpumask(n
);
5867 cpus_and(tmp
, tmp
, nodemask
);
5868 if (cpus_empty(tmp
))
5871 sg
= kmalloc(sizeof(struct sched_group
), GFP_KERNEL
);
5874 "Can not alloc domain group for node %d\n", j
);
5879 cpus_or(covered
, covered
, tmp
);
5883 prev
->next
= sched_group_nodes
[i
];
5887 /* Calculate CPU power for physical packages and nodes */
5888 for_each_cpu_mask(i
, *cpu_map
) {
5890 struct sched_domain
*sd
;
5891 #ifdef CONFIG_SCHED_SMT
5892 sd
= &per_cpu(cpu_domains
, i
);
5893 power
= SCHED_LOAD_SCALE
;
5894 sd
->groups
->cpu_power
= power
;
5896 #ifdef CONFIG_SCHED_MC
5897 sd
= &per_cpu(core_domains
, i
);
5898 power
= SCHED_LOAD_SCALE
+ (cpus_weight(sd
->groups
->cpumask
)-1)
5899 * SCHED_LOAD_SCALE
/ 10;
5900 sd
->groups
->cpu_power
= power
;
5902 sd
= &per_cpu(phys_domains
, i
);
5905 * This has to be < 2 * SCHED_LOAD_SCALE
5906 * Lets keep it SCHED_LOAD_SCALE, so that
5907 * while calculating NUMA group's cpu_power
5909 * numa_group->cpu_power += phys_group->cpu_power;
5911 * See "only add power once for each physical pkg"
5914 sd
->groups
->cpu_power
= SCHED_LOAD_SCALE
;
5916 sd
= &per_cpu(phys_domains
, i
);
5917 power
= SCHED_LOAD_SCALE
+ SCHED_LOAD_SCALE
*
5918 (cpus_weight(sd
->groups
->cpumask
)-1) / 10;
5919 sd
->groups
->cpu_power
= power
;
5924 for (i
= 0; i
< MAX_NUMNODES
; i
++)
5925 init_numa_sched_groups_power(sched_group_nodes
[i
]);
5927 init_numa_sched_groups_power(sched_group_allnodes
);
5930 /* Attach the domains */
5931 for_each_cpu_mask(i
, *cpu_map
) {
5932 struct sched_domain
*sd
;
5933 #ifdef CONFIG_SCHED_SMT
5934 sd
= &per_cpu(cpu_domains
, i
);
5935 #elif defined(CONFIG_SCHED_MC)
5936 sd
= &per_cpu(core_domains
, i
);
5938 sd
= &per_cpu(phys_domains
, i
);
5940 cpu_attach_domain(sd
, i
);
5943 * Tune cache-hot values:
5945 calibrate_migration_costs(cpu_map
);
5948 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
5950 static void arch_init_sched_domains(const cpumask_t
*cpu_map
)
5952 cpumask_t cpu_default_map
;
5955 * Setup mask for cpus without special case scheduling requirements.
5956 * For now this just excludes isolated cpus, but could be used to
5957 * exclude other special cases in the future.
5959 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
5961 build_sched_domains(&cpu_default_map
);
5964 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
5970 for_each_cpu_mask(cpu
, *cpu_map
) {
5971 struct sched_group
*sched_group_allnodes
5972 = sched_group_allnodes_bycpu
[cpu
];
5973 struct sched_group
**sched_group_nodes
5974 = sched_group_nodes_bycpu
[cpu
];
5976 if (sched_group_allnodes
) {
5977 kfree(sched_group_allnodes
);
5978 sched_group_allnodes_bycpu
[cpu
] = NULL
;
5981 if (!sched_group_nodes
)
5984 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5985 cpumask_t nodemask
= node_to_cpumask(i
);
5986 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
5988 cpus_and(nodemask
, nodemask
, *cpu_map
);
5989 if (cpus_empty(nodemask
))
5999 if (oldsg
!= sched_group_nodes
[i
])
6002 kfree(sched_group_nodes
);
6003 sched_group_nodes_bycpu
[cpu
] = NULL
;
6009 * Detach sched domains from a group of cpus specified in cpu_map
6010 * These cpus will now be attached to the NULL domain
6012 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6016 for_each_cpu_mask(i
, *cpu_map
)
6017 cpu_attach_domain(NULL
, i
);
6018 synchronize_sched();
6019 arch_destroy_sched_domains(cpu_map
);
6023 * Partition sched domains as specified by the cpumasks below.
6024 * This attaches all cpus from the cpumasks to the NULL domain,
6025 * waits for a RCU quiescent period, recalculates sched
6026 * domain information and then attaches them back to the
6027 * correct sched domains
6028 * Call with hotplug lock held
6030 void partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
6032 cpumask_t change_map
;
6034 cpus_and(*partition1
, *partition1
, cpu_online_map
);
6035 cpus_and(*partition2
, *partition2
, cpu_online_map
);
6036 cpus_or(change_map
, *partition1
, *partition2
);
6038 /* Detach sched domains from all of the affected cpus */
6039 detach_destroy_domains(&change_map
);
6040 if (!cpus_empty(*partition1
))
6041 build_sched_domains(partition1
);
6042 if (!cpus_empty(*partition2
))
6043 build_sched_domains(partition2
);
6046 #ifdef CONFIG_HOTPLUG_CPU
6048 * Force a reinitialization of the sched domains hierarchy. The domains
6049 * and groups cannot be updated in place without racing with the balancing
6050 * code, so we temporarily attach all running cpus to the NULL domain
6051 * which will prevent rebalancing while the sched domains are recalculated.
6053 static int update_sched_domains(struct notifier_block
*nfb
,
6054 unsigned long action
, void *hcpu
)
6057 case CPU_UP_PREPARE
:
6058 case CPU_DOWN_PREPARE
:
6059 detach_destroy_domains(&cpu_online_map
);
6062 case CPU_UP_CANCELED
:
6063 case CPU_DOWN_FAILED
:
6067 * Fall through and re-initialise the domains.
6074 /* The hotplug lock is already held by cpu_up/cpu_down */
6075 arch_init_sched_domains(&cpu_online_map
);
6081 void __init
sched_init_smp(void)
6084 arch_init_sched_domains(&cpu_online_map
);
6085 unlock_cpu_hotplug();
6086 /* XXX: Theoretical race here - CPU may be hotplugged now */
6087 hotcpu_notifier(update_sched_domains
, 0);
6090 void __init
sched_init_smp(void)
6093 #endif /* CONFIG_SMP */
6095 int in_sched_functions(unsigned long addr
)
6097 /* Linker adds these: start and end of __sched functions */
6098 extern char __sched_text_start
[], __sched_text_end
[];
6099 return in_lock_functions(addr
) ||
6100 (addr
>= (unsigned long)__sched_text_start
6101 && addr
< (unsigned long)__sched_text_end
);
6104 void __init
sched_init(void)
6109 for_each_possible_cpu(i
) {
6110 prio_array_t
*array
;
6113 spin_lock_init(&rq
->lock
);
6115 rq
->active
= rq
->arrays
;
6116 rq
->expired
= rq
->arrays
+ 1;
6117 rq
->best_expired_prio
= MAX_PRIO
;
6121 for (j
= 1; j
< 3; j
++)
6122 rq
->cpu_load
[j
] = 0;
6123 rq
->active_balance
= 0;
6125 rq
->migration_thread
= NULL
;
6126 INIT_LIST_HEAD(&rq
->migration_queue
);
6129 atomic_set(&rq
->nr_iowait
, 0);
6131 for (j
= 0; j
< 2; j
++) {
6132 array
= rq
->arrays
+ j
;
6133 for (k
= 0; k
< MAX_PRIO
; k
++) {
6134 INIT_LIST_HEAD(array
->queue
+ k
);
6135 __clear_bit(k
, array
->bitmap
);
6137 // delimiter for bitsearch
6138 __set_bit(MAX_PRIO
, array
->bitmap
);
6143 * The boot idle thread does lazy MMU switching as well:
6145 atomic_inc(&init_mm
.mm_count
);
6146 enter_lazy_tlb(&init_mm
, current
);
6149 * Make us the idle thread. Technically, schedule() should not be
6150 * called from this thread, however somewhere below it might be,
6151 * but because we are the idle thread, we just pick up running again
6152 * when this runqueue becomes "idle".
6154 init_idle(current
, smp_processor_id());
6157 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6158 void __might_sleep(char *file
, int line
)
6160 #if defined(in_atomic)
6161 static unsigned long prev_jiffy
; /* ratelimiting */
6163 if ((in_atomic() || irqs_disabled()) &&
6164 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6165 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6167 prev_jiffy
= jiffies
;
6168 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6169 " context at %s:%d\n", file
, line
);
6170 printk("in_atomic():%d, irqs_disabled():%d\n",
6171 in_atomic(), irqs_disabled());
6176 EXPORT_SYMBOL(__might_sleep
);
6179 #ifdef CONFIG_MAGIC_SYSRQ
6180 void normalize_rt_tasks(void)
6182 struct task_struct
*p
;
6183 prio_array_t
*array
;
6184 unsigned long flags
;
6187 read_lock_irq(&tasklist_lock
);
6188 for_each_process (p
) {
6192 rq
= task_rq_lock(p
, &flags
);
6196 deactivate_task(p
, task_rq(p
));
6197 __setscheduler(p
, SCHED_NORMAL
, 0);
6199 __activate_task(p
, task_rq(p
));
6200 resched_task(rq
->curr
);
6203 task_rq_unlock(rq
, &flags
);
6205 read_unlock_irq(&tasklist_lock
);
6208 #endif /* CONFIG_MAGIC_SYSRQ */
6212 * These functions are only useful for the IA64 MCA handling.
6214 * They can only be called when the whole system has been
6215 * stopped - every CPU needs to be quiescent, and no scheduling
6216 * activity can take place. Using them for anything else would
6217 * be a serious bug, and as a result, they aren't even visible
6218 * under any other configuration.
6222 * curr_task - return the current task for a given cpu.
6223 * @cpu: the processor in question.
6225 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6227 task_t
*curr_task(int cpu
)
6229 return cpu_curr(cpu
);
6233 * set_curr_task - set the current task for a given cpu.
6234 * @cpu: the processor in question.
6235 * @p: the task pointer to set.
6237 * Description: This function must only be used when non-maskable interrupts
6238 * are serviced on a separate stack. It allows the architecture to switch the
6239 * notion of the current task on a cpu in a non-blocking manner. This function
6240 * must be called with all CPU's synchronized, and interrupts disabled, the
6241 * and caller must save the original value of the current task (see
6242 * curr_task() above) and restore that value before reenabling interrupts and
6243 * re-starting the system.
6245 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6247 void set_curr_task(int cpu
, task_t
*p
)