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/completion.h>
31 #include <linux/kernel_stat.h>
32 #include <linux/security.h>
33 #include <linux/notifier.h>
34 #include <linux/profile.h>
35 #include <linux/suspend.h>
36 #include <linux/blkdev.h>
37 #include <linux/delay.h>
38 #include <linux/smp.h>
39 #include <linux/threads.h>
40 #include <linux/timer.h>
41 #include <linux/rcupdate.h>
42 #include <linux/cpu.h>
43 #include <linux/cpuset.h>
44 #include <linux/percpu.h>
45 #include <linux/kthread.h>
46 #include <linux/seq_file.h>
47 #include <linux/syscalls.h>
48 #include <linux/times.h>
49 #include <linux/acct.h>
52 #include <asm/unistd.h>
55 * Convert user-nice values [ -20 ... 0 ... 19 ]
56 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
59 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
60 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
61 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
64 * 'User priority' is the nice value converted to something we
65 * can work with better when scaling various scheduler parameters,
66 * it's a [ 0 ... 39 ] range.
68 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
69 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
70 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
73 * Some helpers for converting nanosecond timing to jiffy resolution
75 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
76 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
79 * These are the 'tuning knobs' of the scheduler:
81 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
82 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
83 * Timeslices get refilled after they expire.
85 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
86 #define DEF_TIMESLICE (100 * HZ / 1000)
87 #define ON_RUNQUEUE_WEIGHT 30
88 #define CHILD_PENALTY 95
89 #define PARENT_PENALTY 100
91 #define PRIO_BONUS_RATIO 25
92 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
93 #define INTERACTIVE_DELTA 2
94 #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
95 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
96 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
99 * If a task is 'interactive' then we reinsert it in the active
100 * array after it has expired its current timeslice. (it will not
101 * continue to run immediately, it will still roundrobin with
102 * other interactive tasks.)
104 * This part scales the interactivity limit depending on niceness.
106 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
107 * Here are a few examples of different nice levels:
109 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
110 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
111 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
112 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
113 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
115 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
116 * priority range a task can explore, a value of '1' means the
117 * task is rated interactive.)
119 * Ie. nice +19 tasks can never get 'interactive' enough to be
120 * reinserted into the active array. And only heavily CPU-hog nice -20
121 * tasks will be expired. Default nice 0 tasks are somewhere between,
122 * it takes some effort for them to get interactive, but it's not
126 #define CURRENT_BONUS(p) \
127 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
130 #define GRANULARITY (10 * HZ / 1000 ? : 1)
133 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
134 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
137 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
138 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
141 #define SCALE(v1,v1_max,v2_max) \
142 (v1) * (v2_max) / (v1_max)
145 (SCALE(TASK_NICE(p), 40, MAX_BONUS) + INTERACTIVE_DELTA)
147 #define TASK_INTERACTIVE(p) \
148 ((p)->prio <= (p)->static_prio - DELTA(p))
150 #define INTERACTIVE_SLEEP(p) \
151 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
152 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
154 #define TASK_PREEMPTS_CURR(p, rq) \
155 ((p)->prio < (rq)->curr->prio)
158 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
159 * to time slice values: [800ms ... 100ms ... 5ms]
161 * The higher a thread's priority, the bigger timeslices
162 * it gets during one round of execution. But even the lowest
163 * priority thread gets MIN_TIMESLICE worth of execution time.
166 #define SCALE_PRIO(x, prio) \
167 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO/2), MIN_TIMESLICE)
169 static inline unsigned int task_timeslice(task_t
*p
)
171 if (p
->static_prio
< NICE_TO_PRIO(0))
172 return SCALE_PRIO(DEF_TIMESLICE
*4, p
->static_prio
);
174 return SCALE_PRIO(DEF_TIMESLICE
, p
->static_prio
);
176 #define task_hot(p, now, sd) ((long long) ((now) - (p)->last_ran) \
177 < (long long) (sd)->cache_hot_time)
180 * These are the runqueue data structures:
183 #define BITMAP_SIZE ((((MAX_PRIO+1+7)/8)+sizeof(long)-1)/sizeof(long))
185 typedef struct runqueue runqueue_t
;
188 unsigned int nr_active
;
189 unsigned long bitmap
[BITMAP_SIZE
];
190 struct list_head queue
[MAX_PRIO
];
194 * This is the main, per-CPU runqueue data structure.
196 * Locking rule: those places that want to lock multiple runqueues
197 * (such as the load balancing or the thread migration code), lock
198 * acquire operations must be ordered by ascending &runqueue.
204 * nr_running and cpu_load should be in the same cacheline because
205 * remote CPUs use both these fields when doing load calculation.
207 unsigned long nr_running
;
209 unsigned long cpu_load
;
211 unsigned long long nr_switches
;
214 * This is part of a global counter where only the total sum
215 * over all CPUs matters. A task can increase this counter on
216 * one CPU and if it got migrated afterwards it may decrease
217 * it on another CPU. Always updated under the runqueue lock:
219 unsigned long nr_uninterruptible
;
221 unsigned long expired_timestamp
;
222 unsigned long long timestamp_last_tick
;
224 struct mm_struct
*prev_mm
;
225 prio_array_t
*active
, *expired
, arrays
[2];
226 int best_expired_prio
;
230 struct sched_domain
*sd
;
232 /* For active balancing */
236 task_t
*migration_thread
;
237 struct list_head migration_queue
;
240 #ifdef CONFIG_SCHEDSTATS
242 struct sched_info rq_sched_info
;
244 /* sys_sched_yield() stats */
245 unsigned long yld_exp_empty
;
246 unsigned long yld_act_empty
;
247 unsigned long yld_both_empty
;
248 unsigned long yld_cnt
;
250 /* schedule() stats */
251 unsigned long sched_switch
;
252 unsigned long sched_cnt
;
253 unsigned long sched_goidle
;
255 /* try_to_wake_up() stats */
256 unsigned long ttwu_cnt
;
257 unsigned long ttwu_local
;
261 static DEFINE_PER_CPU(struct runqueue
, runqueues
);
263 #define for_each_domain(cpu, domain) \
264 for (domain = cpu_rq(cpu)->sd; domain; domain = domain->parent)
266 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
267 #define this_rq() (&__get_cpu_var(runqueues))
268 #define task_rq(p) cpu_rq(task_cpu(p))
269 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
272 * Default context-switch locking:
274 #ifndef prepare_arch_switch
275 # define prepare_arch_switch(rq, next) do { } while (0)
276 # define finish_arch_switch(rq, next) spin_unlock_irq(&(rq)->lock)
277 # define task_running(rq, p) ((rq)->curr == (p))
281 * task_rq_lock - lock the runqueue a given task resides on and disable
282 * interrupts. Note the ordering: we can safely lookup the task_rq without
283 * explicitly disabling preemption.
285 static inline runqueue_t
*task_rq_lock(task_t
*p
, unsigned long *flags
)
291 local_irq_save(*flags
);
293 spin_lock(&rq
->lock
);
294 if (unlikely(rq
!= task_rq(p
))) {
295 spin_unlock_irqrestore(&rq
->lock
, *flags
);
296 goto repeat_lock_task
;
301 static inline void task_rq_unlock(runqueue_t
*rq
, unsigned long *flags
)
304 spin_unlock_irqrestore(&rq
->lock
, *flags
);
307 #ifdef CONFIG_SCHEDSTATS
309 * bump this up when changing the output format or the meaning of an existing
310 * format, so that tools can adapt (or abort)
312 #define SCHEDSTAT_VERSION 11
314 static int show_schedstat(struct seq_file
*seq
, void *v
)
318 seq_printf(seq
, "version %d\n", SCHEDSTAT_VERSION
);
319 seq_printf(seq
, "timestamp %lu\n", jiffies
);
320 for_each_online_cpu(cpu
) {
321 runqueue_t
*rq
= cpu_rq(cpu
);
323 struct sched_domain
*sd
;
327 /* runqueue-specific stats */
329 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
330 cpu
, rq
->yld_both_empty
,
331 rq
->yld_act_empty
, rq
->yld_exp_empty
, rq
->yld_cnt
,
332 rq
->sched_switch
, rq
->sched_cnt
, rq
->sched_goidle
,
333 rq
->ttwu_cnt
, rq
->ttwu_local
,
334 rq
->rq_sched_info
.cpu_time
,
335 rq
->rq_sched_info
.run_delay
, rq
->rq_sched_info
.pcnt
);
337 seq_printf(seq
, "\n");
340 /* domain-specific stats */
341 for_each_domain(cpu
, sd
) {
342 enum idle_type itype
;
343 char mask_str
[NR_CPUS
];
345 cpumask_scnprintf(mask_str
, NR_CPUS
, sd
->span
);
346 seq_printf(seq
, "domain%d %s", dcnt
++, mask_str
);
347 for (itype
= SCHED_IDLE
; itype
< MAX_IDLE_TYPES
;
349 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu %lu",
351 sd
->lb_balanced
[itype
],
352 sd
->lb_failed
[itype
],
353 sd
->lb_imbalance
[itype
],
354 sd
->lb_gained
[itype
],
355 sd
->lb_hot_gained
[itype
],
356 sd
->lb_nobusyq
[itype
],
357 sd
->lb_nobusyg
[itype
]);
359 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu %lu\n",
360 sd
->alb_cnt
, sd
->alb_failed
, sd
->alb_pushed
,
361 sd
->sbe_pushed
, sd
->sbe_attempts
,
362 sd
->ttwu_wake_remote
, sd
->ttwu_move_affine
, sd
->ttwu_move_balance
);
369 static int schedstat_open(struct inode
*inode
, struct file
*file
)
371 unsigned int size
= PAGE_SIZE
* (1 + num_online_cpus() / 32);
372 char *buf
= kmalloc(size
, GFP_KERNEL
);
378 res
= single_open(file
, show_schedstat
, NULL
);
380 m
= file
->private_data
;
388 struct file_operations proc_schedstat_operations
= {
389 .open
= schedstat_open
,
392 .release
= single_release
,
395 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
396 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
397 #else /* !CONFIG_SCHEDSTATS */
398 # define schedstat_inc(rq, field) do { } while (0)
399 # define schedstat_add(rq, field, amt) do { } while (0)
403 * rq_lock - lock a given runqueue and disable interrupts.
405 static inline runqueue_t
*this_rq_lock(void)
412 spin_lock(&rq
->lock
);
417 #ifdef CONFIG_SCHED_SMT
418 static int cpu_and_siblings_are_idle(int cpu
)
421 for_each_cpu_mask(sib
, cpu_sibling_map
[cpu
]) {
430 #define cpu_and_siblings_are_idle(A) idle_cpu(A)
433 #ifdef CONFIG_SCHEDSTATS
435 * Called when a process is dequeued from the active array and given
436 * the cpu. We should note that with the exception of interactive
437 * tasks, the expired queue will become the active queue after the active
438 * queue is empty, without explicitly dequeuing and requeuing tasks in the
439 * expired queue. (Interactive tasks may be requeued directly to the
440 * active queue, thus delaying tasks in the expired queue from running;
441 * see scheduler_tick()).
443 * This function is only called from sched_info_arrive(), rather than
444 * dequeue_task(). Even though a task may be queued and dequeued multiple
445 * times as it is shuffled about, we're really interested in knowing how
446 * long it was from the *first* time it was queued to the time that it
449 static inline void sched_info_dequeued(task_t
*t
)
451 t
->sched_info
.last_queued
= 0;
455 * Called when a task finally hits the cpu. We can now calculate how
456 * long it was waiting to run. We also note when it began so that we
457 * can keep stats on how long its timeslice is.
459 static inline void sched_info_arrive(task_t
*t
)
461 unsigned long now
= jiffies
, diff
= 0;
462 struct runqueue
*rq
= task_rq(t
);
464 if (t
->sched_info
.last_queued
)
465 diff
= now
- t
->sched_info
.last_queued
;
466 sched_info_dequeued(t
);
467 t
->sched_info
.run_delay
+= diff
;
468 t
->sched_info
.last_arrival
= now
;
469 t
->sched_info
.pcnt
++;
474 rq
->rq_sched_info
.run_delay
+= diff
;
475 rq
->rq_sched_info
.pcnt
++;
479 * Called when a process is queued into either the active or expired
480 * array. The time is noted and later used to determine how long we
481 * had to wait for us to reach the cpu. Since the expired queue will
482 * become the active queue after active queue is empty, without dequeuing
483 * and requeuing any tasks, we are interested in queuing to either. It
484 * is unusual but not impossible for tasks to be dequeued and immediately
485 * requeued in the same or another array: this can happen in sched_yield(),
486 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
489 * This function is only called from enqueue_task(), but also only updates
490 * the timestamp if it is already not set. It's assumed that
491 * sched_info_dequeued() will clear that stamp when appropriate.
493 static inline void sched_info_queued(task_t
*t
)
495 if (!t
->sched_info
.last_queued
)
496 t
->sched_info
.last_queued
= jiffies
;
500 * Called when a process ceases being the active-running process, either
501 * voluntarily or involuntarily. Now we can calculate how long we ran.
503 static inline void sched_info_depart(task_t
*t
)
505 struct runqueue
*rq
= task_rq(t
);
506 unsigned long diff
= jiffies
- t
->sched_info
.last_arrival
;
508 t
->sched_info
.cpu_time
+= diff
;
511 rq
->rq_sched_info
.cpu_time
+= diff
;
515 * Called when tasks are switched involuntarily due, typically, to expiring
516 * their time slice. (This may also be called when switching to or from
517 * the idle task.) We are only called when prev != next.
519 static inline void sched_info_switch(task_t
*prev
, task_t
*next
)
521 struct runqueue
*rq
= task_rq(prev
);
524 * prev now departs the cpu. It's not interesting to record
525 * stats about how efficient we were at scheduling the idle
528 if (prev
!= rq
->idle
)
529 sched_info_depart(prev
);
531 if (next
!= rq
->idle
)
532 sched_info_arrive(next
);
535 #define sched_info_queued(t) do { } while (0)
536 #define sched_info_switch(t, next) do { } while (0)
537 #endif /* CONFIG_SCHEDSTATS */
540 * Adding/removing a task to/from a priority array:
542 static void dequeue_task(struct task_struct
*p
, prio_array_t
*array
)
545 list_del(&p
->run_list
);
546 if (list_empty(array
->queue
+ p
->prio
))
547 __clear_bit(p
->prio
, array
->bitmap
);
550 static void enqueue_task(struct task_struct
*p
, prio_array_t
*array
)
552 sched_info_queued(p
);
553 list_add_tail(&p
->run_list
, array
->queue
+ p
->prio
);
554 __set_bit(p
->prio
, array
->bitmap
);
560 * Put task to the end of the run list without the overhead of dequeue
561 * followed by enqueue.
563 static void requeue_task(struct task_struct
*p
, prio_array_t
*array
)
565 list_move_tail(&p
->run_list
, array
->queue
+ p
->prio
);
568 static inline void enqueue_task_head(struct task_struct
*p
, prio_array_t
*array
)
570 list_add(&p
->run_list
, array
->queue
+ p
->prio
);
571 __set_bit(p
->prio
, array
->bitmap
);
577 * effective_prio - return the priority that is based on the static
578 * priority but is modified by bonuses/penalties.
580 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
581 * into the -5 ... 0 ... +5 bonus/penalty range.
583 * We use 25% of the full 0...39 priority range so that:
585 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
586 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
588 * Both properties are important to certain workloads.
590 static int effective_prio(task_t
*p
)
597 bonus
= CURRENT_BONUS(p
) - MAX_BONUS
/ 2;
599 prio
= p
->static_prio
- bonus
;
600 if (prio
< MAX_RT_PRIO
)
602 if (prio
> MAX_PRIO
-1)
608 * __activate_task - move a task to the runqueue.
610 static inline void __activate_task(task_t
*p
, runqueue_t
*rq
)
612 enqueue_task(p
, rq
->active
);
617 * __activate_idle_task - move idle task to the _front_ of runqueue.
619 static inline void __activate_idle_task(task_t
*p
, runqueue_t
*rq
)
621 enqueue_task_head(p
, rq
->active
);
625 static void recalc_task_prio(task_t
*p
, unsigned long long now
)
627 /* Caller must always ensure 'now >= p->timestamp' */
628 unsigned long long __sleep_time
= now
- p
->timestamp
;
629 unsigned long sleep_time
;
631 if (__sleep_time
> NS_MAX_SLEEP_AVG
)
632 sleep_time
= NS_MAX_SLEEP_AVG
;
634 sleep_time
= (unsigned long)__sleep_time
;
636 if (likely(sleep_time
> 0)) {
638 * User tasks that sleep a long time are categorised as
639 * idle and will get just interactive status to stay active &
640 * prevent them suddenly becoming cpu hogs and starving
643 if (p
->mm
&& p
->activated
!= -1 &&
644 sleep_time
> INTERACTIVE_SLEEP(p
)) {
645 p
->sleep_avg
= JIFFIES_TO_NS(MAX_SLEEP_AVG
-
649 * The lower the sleep avg a task has the more
650 * rapidly it will rise with sleep time.
652 sleep_time
*= (MAX_BONUS
- CURRENT_BONUS(p
)) ? : 1;
655 * Tasks waking from uninterruptible sleep are
656 * limited in their sleep_avg rise as they
657 * are likely to be waiting on I/O
659 if (p
->activated
== -1 && p
->mm
) {
660 if (p
->sleep_avg
>= INTERACTIVE_SLEEP(p
))
662 else if (p
->sleep_avg
+ sleep_time
>=
663 INTERACTIVE_SLEEP(p
)) {
664 p
->sleep_avg
= INTERACTIVE_SLEEP(p
);
670 * This code gives a bonus to interactive tasks.
672 * The boost works by updating the 'average sleep time'
673 * value here, based on ->timestamp. The more time a
674 * task spends sleeping, the higher the average gets -
675 * and the higher the priority boost gets as well.
677 p
->sleep_avg
+= sleep_time
;
679 if (p
->sleep_avg
> NS_MAX_SLEEP_AVG
)
680 p
->sleep_avg
= NS_MAX_SLEEP_AVG
;
684 p
->prio
= effective_prio(p
);
688 * activate_task - move a task to the runqueue and do priority recalculation
690 * Update all the scheduling statistics stuff. (sleep average
691 * calculation, priority modifiers, etc.)
693 static void activate_task(task_t
*p
, runqueue_t
*rq
, int local
)
695 unsigned long long now
;
700 /* Compensate for drifting sched_clock */
701 runqueue_t
*this_rq
= this_rq();
702 now
= (now
- this_rq
->timestamp_last_tick
)
703 + rq
->timestamp_last_tick
;
707 recalc_task_prio(p
, now
);
710 * This checks to make sure it's not an uninterruptible task
711 * that is now waking up.
715 * Tasks which were woken up by interrupts (ie. hw events)
716 * are most likely of interactive nature. So we give them
717 * the credit of extending their sleep time to the period
718 * of time they spend on the runqueue, waiting for execution
719 * on a CPU, first time around:
725 * Normal first-time wakeups get a credit too for
726 * on-runqueue time, but it will be weighted down:
733 __activate_task(p
, rq
);
737 * deactivate_task - remove a task from the runqueue.
739 static void deactivate_task(struct task_struct
*p
, runqueue_t
*rq
)
742 dequeue_task(p
, p
->array
);
747 * resched_task - mark a task 'to be rescheduled now'.
749 * On UP this means the setting of the need_resched flag, on SMP it
750 * might also involve a cross-CPU call to trigger the scheduler on
754 static void resched_task(task_t
*p
)
756 int need_resched
, nrpolling
;
758 assert_spin_locked(&task_rq(p
)->lock
);
760 /* minimise the chance of sending an interrupt to poll_idle() */
761 nrpolling
= test_tsk_thread_flag(p
,TIF_POLLING_NRFLAG
);
762 need_resched
= test_and_set_tsk_thread_flag(p
,TIF_NEED_RESCHED
);
763 nrpolling
|= test_tsk_thread_flag(p
,TIF_POLLING_NRFLAG
);
765 if (!need_resched
&& !nrpolling
&& (task_cpu(p
) != smp_processor_id()))
766 smp_send_reschedule(task_cpu(p
));
769 static inline void resched_task(task_t
*p
)
771 set_tsk_need_resched(p
);
776 * task_curr - is this task currently executing on a CPU?
777 * @p: the task in question.
779 inline int task_curr(const task_t
*p
)
781 return cpu_curr(task_cpu(p
)) == p
;
791 struct list_head list
;
792 enum request_type type
;
794 /* For REQ_MOVE_TASK */
798 /* For REQ_SET_DOMAIN */
799 struct sched_domain
*sd
;
801 struct completion done
;
805 * The task's runqueue lock must be held.
806 * Returns true if you have to wait for migration thread.
808 static int migrate_task(task_t
*p
, int dest_cpu
, migration_req_t
*req
)
810 runqueue_t
*rq
= task_rq(p
);
813 * If the task is not on a runqueue (and not running), then
814 * it is sufficient to simply update the task's cpu field.
816 if (!p
->array
&& !task_running(rq
, p
)) {
817 set_task_cpu(p
, dest_cpu
);
821 init_completion(&req
->done
);
822 req
->type
= REQ_MOVE_TASK
;
824 req
->dest_cpu
= dest_cpu
;
825 list_add(&req
->list
, &rq
->migration_queue
);
830 * wait_task_inactive - wait for a thread to unschedule.
832 * The caller must ensure that the task *will* unschedule sometime soon,
833 * else this function might spin for a *long* time. This function can't
834 * be called with interrupts off, or it may introduce deadlock with
835 * smp_call_function() if an IPI is sent by the same process we are
836 * waiting to become inactive.
838 void wait_task_inactive(task_t
* p
)
845 rq
= task_rq_lock(p
, &flags
);
846 /* Must be off runqueue entirely, not preempted. */
847 if (unlikely(p
->array
|| task_running(rq
, p
))) {
848 /* If it's preempted, we yield. It could be a while. */
849 preempted
= !task_running(rq
, p
);
850 task_rq_unlock(rq
, &flags
);
856 task_rq_unlock(rq
, &flags
);
860 * kick_process - kick a running thread to enter/exit the kernel
861 * @p: the to-be-kicked thread
863 * Cause a process which is running on another CPU to enter
864 * kernel-mode, without any delay. (to get signals handled.)
866 * NOTE: this function doesnt have to take the runqueue lock,
867 * because all it wants to ensure is that the remote task enters
868 * the kernel. If the IPI races and the task has been migrated
869 * to another CPU then no harm is done and the purpose has been
872 void kick_process(task_t
*p
)
878 if ((cpu
!= smp_processor_id()) && task_curr(p
))
879 smp_send_reschedule(cpu
);
884 * Return a low guess at the load of a migration-source cpu.
886 * We want to under-estimate the load of migration sources, to
887 * balance conservatively.
889 static inline unsigned long source_load(int cpu
)
891 runqueue_t
*rq
= cpu_rq(cpu
);
892 unsigned long load_now
= rq
->nr_running
* SCHED_LOAD_SCALE
;
894 return min(rq
->cpu_load
, load_now
);
898 * Return a high guess at the load of a migration-target cpu
900 static inline unsigned long target_load(int cpu
)
902 runqueue_t
*rq
= cpu_rq(cpu
);
903 unsigned long load_now
= rq
->nr_running
* SCHED_LOAD_SCALE
;
905 return max(rq
->cpu_load
, load_now
);
911 * wake_idle() will wake a task on an idle cpu if task->cpu is
912 * not idle and an idle cpu is available. The span of cpus to
913 * search starts with cpus closest then further out as needed,
914 * so we always favor a closer, idle cpu.
916 * Returns the CPU we should wake onto.
918 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
919 static int wake_idle(int cpu
, task_t
*p
)
922 struct sched_domain
*sd
;
928 for_each_domain(cpu
, sd
) {
929 if (sd
->flags
& SD_WAKE_IDLE
) {
930 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
931 for_each_cpu_mask(i
, tmp
) {
942 static inline int wake_idle(int cpu
, task_t
*p
)
949 * try_to_wake_up - wake up a thread
950 * @p: the to-be-woken-up thread
951 * @state: the mask of task states that can be woken
952 * @sync: do a synchronous wakeup?
954 * Put it on the run-queue if it's not already there. The "current"
955 * thread is always on the run-queue (except when the actual
956 * re-schedule is in progress), and as such you're allowed to do
957 * the simpler "current->state = TASK_RUNNING" to mark yourself
958 * runnable without the overhead of this.
960 * returns failure only if the task is already active.
962 static int try_to_wake_up(task_t
* p
, unsigned int state
, int sync
)
964 int cpu
, this_cpu
, success
= 0;
969 unsigned long load
, this_load
;
970 struct sched_domain
*sd
;
974 rq
= task_rq_lock(p
, &flags
);
975 old_state
= p
->state
;
976 if (!(old_state
& state
))
983 this_cpu
= smp_processor_id();
986 if (unlikely(task_running(rq
, p
)))
989 #ifdef CONFIG_SCHEDSTATS
990 schedstat_inc(rq
, ttwu_cnt
);
991 if (cpu
== this_cpu
) {
992 schedstat_inc(rq
, ttwu_local
);
994 for_each_domain(this_cpu
, sd
) {
995 if (cpu_isset(cpu
, sd
->span
)) {
996 schedstat_inc(sd
, ttwu_wake_remote
);
1004 if (cpu
== this_cpu
|| unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1007 load
= source_load(cpu
);
1008 this_load
= target_load(this_cpu
);
1011 * If sync wakeup then subtract the (maximum possible) effect of
1012 * the currently running task from the load of the current CPU:
1015 this_load
-= SCHED_LOAD_SCALE
;
1017 /* Don't pull the task off an idle CPU to a busy one */
1018 if (load
< SCHED_LOAD_SCALE
/2 && this_load
> SCHED_LOAD_SCALE
/2)
1021 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1024 * Scan domains for affine wakeup and passive balancing
1027 for_each_domain(this_cpu
, sd
) {
1028 unsigned int imbalance
;
1030 * Start passive balancing when half the imbalance_pct
1033 imbalance
= sd
->imbalance_pct
+ (sd
->imbalance_pct
- 100) / 2;
1035 if ((sd
->flags
& SD_WAKE_AFFINE
) &&
1036 !task_hot(p
, rq
->timestamp_last_tick
, sd
)) {
1038 * This domain has SD_WAKE_AFFINE and p is cache cold
1041 if (cpu_isset(cpu
, sd
->span
)) {
1042 schedstat_inc(sd
, ttwu_move_affine
);
1045 } else if ((sd
->flags
& SD_WAKE_BALANCE
) &&
1046 imbalance
*this_load
<= 100*load
) {
1048 * This domain has SD_WAKE_BALANCE and there is
1051 if (cpu_isset(cpu
, sd
->span
)) {
1052 schedstat_inc(sd
, ttwu_move_balance
);
1058 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1060 new_cpu
= wake_idle(new_cpu
, p
);
1061 if (new_cpu
!= cpu
) {
1062 set_task_cpu(p
, new_cpu
);
1063 task_rq_unlock(rq
, &flags
);
1064 /* might preempt at this point */
1065 rq
= task_rq_lock(p
, &flags
);
1066 old_state
= p
->state
;
1067 if (!(old_state
& state
))
1072 this_cpu
= smp_processor_id();
1077 #endif /* CONFIG_SMP */
1078 if (old_state
== TASK_UNINTERRUPTIBLE
) {
1079 rq
->nr_uninterruptible
--;
1081 * Tasks on involuntary sleep don't earn
1082 * sleep_avg beyond just interactive state.
1088 * Sync wakeups (i.e. those types of wakeups where the waker
1089 * has indicated that it will leave the CPU in short order)
1090 * don't trigger a preemption, if the woken up task will run on
1091 * this cpu. (in this case the 'I will reschedule' promise of
1092 * the waker guarantees that the freshly woken up task is going
1093 * to be considered on this CPU.)
1095 activate_task(p
, rq
, cpu
== this_cpu
);
1096 if (!sync
|| cpu
!= this_cpu
) {
1097 if (TASK_PREEMPTS_CURR(p
, rq
))
1098 resched_task(rq
->curr
);
1103 p
->state
= TASK_RUNNING
;
1105 task_rq_unlock(rq
, &flags
);
1110 int fastcall
wake_up_process(task_t
* p
)
1112 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1113 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1116 EXPORT_SYMBOL(wake_up_process
);
1118 int fastcall
wake_up_state(task_t
*p
, unsigned int state
)
1120 return try_to_wake_up(p
, state
, 0);
1124 static int find_idlest_cpu(struct task_struct
*p
, int this_cpu
,
1125 struct sched_domain
*sd
);
1129 * Perform scheduler related setup for a newly forked process p.
1130 * p is forked by current.
1132 void fastcall
sched_fork(task_t
*p
)
1135 * We mark the process as running here, but have not actually
1136 * inserted it onto the runqueue yet. This guarantees that
1137 * nobody will actually run it, and a signal or other external
1138 * event cannot wake it up and insert it on the runqueue either.
1140 p
->state
= TASK_RUNNING
;
1141 INIT_LIST_HEAD(&p
->run_list
);
1143 spin_lock_init(&p
->switch_lock
);
1144 #ifdef CONFIG_SCHEDSTATS
1145 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1147 #ifdef CONFIG_PREEMPT
1149 * During context-switch we hold precisely one spinlock, which
1150 * schedule_tail drops. (in the common case it's this_rq()->lock,
1151 * but it also can be p->switch_lock.) So we compensate with a count
1152 * of 1. Also, we want to start with kernel preemption disabled.
1154 p
->thread_info
->preempt_count
= 1;
1157 * Share the timeslice between parent and child, thus the
1158 * total amount of pending timeslices in the system doesn't change,
1159 * resulting in more scheduling fairness.
1161 local_irq_disable();
1162 p
->time_slice
= (current
->time_slice
+ 1) >> 1;
1164 * The remainder of the first timeslice might be recovered by
1165 * the parent if the child exits early enough.
1167 p
->first_time_slice
= 1;
1168 current
->time_slice
>>= 1;
1169 p
->timestamp
= sched_clock();
1170 if (unlikely(!current
->time_slice
)) {
1172 * This case is rare, it happens when the parent has only
1173 * a single jiffy left from its timeslice. Taking the
1174 * runqueue lock is not a problem.
1176 current
->time_slice
= 1;
1186 * wake_up_new_task - wake up a newly created task for the first time.
1188 * This function will do some initial scheduler statistics housekeeping
1189 * that must be done for every newly created context, then puts the task
1190 * on the runqueue and wakes it.
1192 void fastcall
wake_up_new_task(task_t
* p
, unsigned long clone_flags
)
1194 unsigned long flags
;
1196 runqueue_t
*rq
, *this_rq
;
1198 rq
= task_rq_lock(p
, &flags
);
1200 this_cpu
= smp_processor_id();
1202 BUG_ON(p
->state
!= TASK_RUNNING
);
1205 * We decrease the sleep average of forking parents
1206 * and children as well, to keep max-interactive tasks
1207 * from forking tasks that are max-interactive. The parent
1208 * (current) is done further down, under its lock.
1210 p
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(p
) *
1211 CHILD_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1213 p
->prio
= effective_prio(p
);
1215 if (likely(cpu
== this_cpu
)) {
1216 if (!(clone_flags
& CLONE_VM
)) {
1218 * The VM isn't cloned, so we're in a good position to
1219 * do child-runs-first in anticipation of an exec. This
1220 * usually avoids a lot of COW overhead.
1222 if (unlikely(!current
->array
))
1223 __activate_task(p
, rq
);
1225 p
->prio
= current
->prio
;
1226 list_add_tail(&p
->run_list
, ¤t
->run_list
);
1227 p
->array
= current
->array
;
1228 p
->array
->nr_active
++;
1233 /* Run child last */
1234 __activate_task(p
, rq
);
1236 * We skip the following code due to cpu == this_cpu
1238 * task_rq_unlock(rq, &flags);
1239 * this_rq = task_rq_lock(current, &flags);
1243 this_rq
= cpu_rq(this_cpu
);
1246 * Not the local CPU - must adjust timestamp. This should
1247 * get optimised away in the !CONFIG_SMP case.
1249 p
->timestamp
= (p
->timestamp
- this_rq
->timestamp_last_tick
)
1250 + rq
->timestamp_last_tick
;
1251 __activate_task(p
, rq
);
1252 if (TASK_PREEMPTS_CURR(p
, rq
))
1253 resched_task(rq
->curr
);
1256 * Parent and child are on different CPUs, now get the
1257 * parent runqueue to update the parent's ->sleep_avg:
1259 task_rq_unlock(rq
, &flags
);
1260 this_rq
= task_rq_lock(current
, &flags
);
1262 current
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(current
) *
1263 PARENT_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1264 task_rq_unlock(this_rq
, &flags
);
1268 * Potentially available exiting-child timeslices are
1269 * retrieved here - this way the parent does not get
1270 * penalized for creating too many threads.
1272 * (this cannot be used to 'generate' timeslices
1273 * artificially, because any timeslice recovered here
1274 * was given away by the parent in the first place.)
1276 void fastcall
sched_exit(task_t
* p
)
1278 unsigned long flags
;
1282 * If the child was a (relative-) CPU hog then decrease
1283 * the sleep_avg of the parent as well.
1285 rq
= task_rq_lock(p
->parent
, &flags
);
1286 if (p
->first_time_slice
) {
1287 p
->parent
->time_slice
+= p
->time_slice
;
1288 if (unlikely(p
->parent
->time_slice
> task_timeslice(p
)))
1289 p
->parent
->time_slice
= task_timeslice(p
);
1291 if (p
->sleep_avg
< p
->parent
->sleep_avg
)
1292 p
->parent
->sleep_avg
= p
->parent
->sleep_avg
/
1293 (EXIT_WEIGHT
+ 1) * EXIT_WEIGHT
+ p
->sleep_avg
/
1295 task_rq_unlock(rq
, &flags
);
1299 * finish_task_switch - clean up after a task-switch
1300 * @prev: the thread we just switched away from.
1302 * We enter this with the runqueue still locked, and finish_arch_switch()
1303 * will unlock it along with doing any other architecture-specific cleanup
1306 * Note that we may have delayed dropping an mm in context_switch(). If
1307 * so, we finish that here outside of the runqueue lock. (Doing it
1308 * with the lock held can cause deadlocks; see schedule() for
1311 static inline void finish_task_switch(task_t
*prev
)
1312 __releases(rq
->lock
)
1314 runqueue_t
*rq
= this_rq();
1315 struct mm_struct
*mm
= rq
->prev_mm
;
1316 unsigned long prev_task_flags
;
1321 * A task struct has one reference for the use as "current".
1322 * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1323 * calls schedule one last time. The schedule call will never return,
1324 * and the scheduled task must drop that reference.
1325 * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1326 * still held, otherwise prev could be scheduled on another cpu, die
1327 * there before we look at prev->state, and then the reference would
1329 * Manfred Spraul <manfred@colorfullife.com>
1331 prev_task_flags
= prev
->flags
;
1332 finish_arch_switch(rq
, prev
);
1335 if (unlikely(prev_task_flags
& PF_DEAD
))
1336 put_task_struct(prev
);
1340 * schedule_tail - first thing a freshly forked thread must call.
1341 * @prev: the thread we just switched away from.
1343 asmlinkage
void schedule_tail(task_t
*prev
)
1344 __releases(rq
->lock
)
1346 finish_task_switch(prev
);
1348 if (current
->set_child_tid
)
1349 put_user(current
->pid
, current
->set_child_tid
);
1353 * context_switch - switch to the new MM and the new
1354 * thread's register state.
1357 task_t
* context_switch(runqueue_t
*rq
, task_t
*prev
, task_t
*next
)
1359 struct mm_struct
*mm
= next
->mm
;
1360 struct mm_struct
*oldmm
= prev
->active_mm
;
1362 if (unlikely(!mm
)) {
1363 next
->active_mm
= oldmm
;
1364 atomic_inc(&oldmm
->mm_count
);
1365 enter_lazy_tlb(oldmm
, next
);
1367 switch_mm(oldmm
, mm
, next
);
1369 if (unlikely(!prev
->mm
)) {
1370 prev
->active_mm
= NULL
;
1371 WARN_ON(rq
->prev_mm
);
1372 rq
->prev_mm
= oldmm
;
1375 /* Here we just switch the register state and the stack. */
1376 switch_to(prev
, next
, prev
);
1382 * nr_running, nr_uninterruptible and nr_context_switches:
1384 * externally visible scheduler statistics: current number of runnable
1385 * threads, current number of uninterruptible-sleeping threads, total
1386 * number of context switches performed since bootup.
1388 unsigned long nr_running(void)
1390 unsigned long i
, sum
= 0;
1392 for_each_online_cpu(i
)
1393 sum
+= cpu_rq(i
)->nr_running
;
1398 unsigned long nr_uninterruptible(void)
1400 unsigned long i
, sum
= 0;
1403 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1406 * Since we read the counters lockless, it might be slightly
1407 * inaccurate. Do not allow it to go below zero though:
1409 if (unlikely((long)sum
< 0))
1415 unsigned long long nr_context_switches(void)
1417 unsigned long long i
, sum
= 0;
1420 sum
+= cpu_rq(i
)->nr_switches
;
1425 unsigned long nr_iowait(void)
1427 unsigned long i
, sum
= 0;
1430 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1438 * double_rq_lock - safely lock two runqueues
1440 * Note this does not disable interrupts like task_rq_lock,
1441 * you need to do so manually before calling.
1443 static void double_rq_lock(runqueue_t
*rq1
, runqueue_t
*rq2
)
1444 __acquires(rq1
->lock
)
1445 __acquires(rq2
->lock
)
1448 spin_lock(&rq1
->lock
);
1449 __acquire(rq2
->lock
); /* Fake it out ;) */
1452 spin_lock(&rq1
->lock
);
1453 spin_lock(&rq2
->lock
);
1455 spin_lock(&rq2
->lock
);
1456 spin_lock(&rq1
->lock
);
1462 * double_rq_unlock - safely unlock two runqueues
1464 * Note this does not restore interrupts like task_rq_unlock,
1465 * you need to do so manually after calling.
1467 static void double_rq_unlock(runqueue_t
*rq1
, runqueue_t
*rq2
)
1468 __releases(rq1
->lock
)
1469 __releases(rq2
->lock
)
1471 spin_unlock(&rq1
->lock
);
1473 spin_unlock(&rq2
->lock
);
1475 __release(rq2
->lock
);
1479 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1481 static void double_lock_balance(runqueue_t
*this_rq
, runqueue_t
*busiest
)
1482 __releases(this_rq
->lock
)
1483 __acquires(busiest
->lock
)
1484 __acquires(this_rq
->lock
)
1486 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1487 if (busiest
< this_rq
) {
1488 spin_unlock(&this_rq
->lock
);
1489 spin_lock(&busiest
->lock
);
1490 spin_lock(&this_rq
->lock
);
1492 spin_lock(&busiest
->lock
);
1497 * find_idlest_cpu - find the least busy runqueue.
1499 static int find_idlest_cpu(struct task_struct
*p
, int this_cpu
,
1500 struct sched_domain
*sd
)
1502 unsigned long load
, min_load
, this_load
;
1507 min_load
= ULONG_MAX
;
1509 cpus_and(mask
, sd
->span
, p
->cpus_allowed
);
1511 for_each_cpu_mask(i
, mask
) {
1512 load
= target_load(i
);
1514 if (load
< min_load
) {
1518 /* break out early on an idle CPU: */
1524 /* add +1 to account for the new task */
1525 this_load
= source_load(this_cpu
) + SCHED_LOAD_SCALE
;
1528 * Would with the addition of the new task to the
1529 * current CPU there be an imbalance between this
1530 * CPU and the idlest CPU?
1532 * Use half of the balancing threshold - new-context is
1533 * a good opportunity to balance.
1535 if (min_load
*(100 + (sd
->imbalance_pct
-100)/2) < this_load
*100)
1542 * If dest_cpu is allowed for this process, migrate the task to it.
1543 * This is accomplished by forcing the cpu_allowed mask to only
1544 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
1545 * the cpu_allowed mask is restored.
1547 static void sched_migrate_task(task_t
*p
, int dest_cpu
)
1549 migration_req_t req
;
1551 unsigned long flags
;
1553 rq
= task_rq_lock(p
, &flags
);
1554 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
1555 || unlikely(cpu_is_offline(dest_cpu
)))
1558 /* force the process onto the specified CPU */
1559 if (migrate_task(p
, dest_cpu
, &req
)) {
1560 /* Need to wait for migration thread (might exit: take ref). */
1561 struct task_struct
*mt
= rq
->migration_thread
;
1562 get_task_struct(mt
);
1563 task_rq_unlock(rq
, &flags
);
1564 wake_up_process(mt
);
1565 put_task_struct(mt
);
1566 wait_for_completion(&req
.done
);
1570 task_rq_unlock(rq
, &flags
);
1574 * sched_exec(): find the highest-level, exec-balance-capable
1575 * domain and try to migrate the task to the least loaded CPU.
1577 * execve() is a valuable balancing opportunity, because at this point
1578 * the task has the smallest effective memory and cache footprint.
1580 void sched_exec(void)
1582 struct sched_domain
*tmp
, *sd
= NULL
;
1583 int new_cpu
, this_cpu
= get_cpu();
1585 /* Prefer the current CPU if there's only this task running */
1586 if (this_rq()->nr_running
<= 1)
1589 for_each_domain(this_cpu
, tmp
)
1590 if (tmp
->flags
& SD_BALANCE_EXEC
)
1594 schedstat_inc(sd
, sbe_attempts
);
1595 new_cpu
= find_idlest_cpu(current
, this_cpu
, sd
);
1596 if (new_cpu
!= this_cpu
) {
1597 schedstat_inc(sd
, sbe_pushed
);
1599 sched_migrate_task(current
, new_cpu
);
1608 * pull_task - move a task from a remote runqueue to the local runqueue.
1609 * Both runqueues must be locked.
1612 void pull_task(runqueue_t
*src_rq
, prio_array_t
*src_array
, task_t
*p
,
1613 runqueue_t
*this_rq
, prio_array_t
*this_array
, int this_cpu
)
1615 dequeue_task(p
, src_array
);
1616 src_rq
->nr_running
--;
1617 set_task_cpu(p
, this_cpu
);
1618 this_rq
->nr_running
++;
1619 enqueue_task(p
, this_array
);
1620 p
->timestamp
= (p
->timestamp
- src_rq
->timestamp_last_tick
)
1621 + this_rq
->timestamp_last_tick
;
1623 * Note that idle threads have a prio of MAX_PRIO, for this test
1624 * to be always true for them.
1626 if (TASK_PREEMPTS_CURR(p
, this_rq
))
1627 resched_task(this_rq
->curr
);
1631 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1634 int can_migrate_task(task_t
*p
, runqueue_t
*rq
, int this_cpu
,
1635 struct sched_domain
*sd
, enum idle_type idle
, int *all_pinned
)
1638 * We do not migrate tasks that are:
1639 * 1) running (obviously), or
1640 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1641 * 3) are cache-hot on their current CPU.
1643 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
1647 if (task_running(rq
, p
))
1651 * Aggressive migration if:
1652 * 1) the [whole] cpu is idle, or
1653 * 2) too many balance attempts have failed.
1656 if (cpu_and_siblings_are_idle(this_cpu
) || \
1657 sd
->nr_balance_failed
> sd
->cache_nice_tries
)
1660 if (task_hot(p
, rq
->timestamp_last_tick
, sd
))
1666 * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq,
1667 * as part of a balancing operation within "domain". Returns the number of
1670 * Called with both runqueues locked.
1672 static int move_tasks(runqueue_t
*this_rq
, int this_cpu
, runqueue_t
*busiest
,
1673 unsigned long max_nr_move
, struct sched_domain
*sd
,
1674 enum idle_type idle
, int *all_pinned
)
1676 prio_array_t
*array
, *dst_array
;
1677 struct list_head
*head
, *curr
;
1678 int idx
, pulled
= 0, pinned
= 0;
1681 if (max_nr_move
== 0)
1687 * We first consider expired tasks. Those will likely not be
1688 * executed in the near future, and they are most likely to
1689 * be cache-cold, thus switching CPUs has the least effect
1692 if (busiest
->expired
->nr_active
) {
1693 array
= busiest
->expired
;
1694 dst_array
= this_rq
->expired
;
1696 array
= busiest
->active
;
1697 dst_array
= this_rq
->active
;
1701 /* Start searching at priority 0: */
1705 idx
= sched_find_first_bit(array
->bitmap
);
1707 idx
= find_next_bit(array
->bitmap
, MAX_PRIO
, idx
);
1708 if (idx
>= MAX_PRIO
) {
1709 if (array
== busiest
->expired
&& busiest
->active
->nr_active
) {
1710 array
= busiest
->active
;
1711 dst_array
= this_rq
->active
;
1717 head
= array
->queue
+ idx
;
1720 tmp
= list_entry(curr
, task_t
, run_list
);
1724 if (!can_migrate_task(tmp
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
1731 #ifdef CONFIG_SCHEDSTATS
1732 if (task_hot(tmp
, busiest
->timestamp_last_tick
, sd
))
1733 schedstat_inc(sd
, lb_hot_gained
[idle
]);
1736 pull_task(busiest
, array
, tmp
, this_rq
, dst_array
, this_cpu
);
1739 /* We only want to steal up to the prescribed number of tasks. */
1740 if (pulled
< max_nr_move
) {
1748 * Right now, this is the only place pull_task() is called,
1749 * so we can safely collect pull_task() stats here rather than
1750 * inside pull_task().
1752 schedstat_add(sd
, lb_gained
[idle
], pulled
);
1755 *all_pinned
= pinned
;
1760 * find_busiest_group finds and returns the busiest CPU group within the
1761 * domain. It calculates and returns the number of tasks which should be
1762 * moved to restore balance via the imbalance parameter.
1764 static struct sched_group
*
1765 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
1766 unsigned long *imbalance
, enum idle_type idle
)
1768 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1769 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
1771 max_load
= this_load
= total_load
= total_pwr
= 0;
1778 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1780 /* Tally up the load of all CPUs in the group */
1783 for_each_cpu_mask(i
, group
->cpumask
) {
1784 /* Bias balancing toward cpus of our domain */
1786 load
= target_load(i
);
1788 load
= source_load(i
);
1793 total_load
+= avg_load
;
1794 total_pwr
+= group
->cpu_power
;
1796 /* Adjust by relative CPU power of the group */
1797 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
1800 this_load
= avg_load
;
1803 } else if (avg_load
> max_load
) {
1804 max_load
= avg_load
;
1808 group
= group
->next
;
1809 } while (group
!= sd
->groups
);
1811 if (!busiest
|| this_load
>= max_load
)
1814 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
1816 if (this_load
>= avg_load
||
1817 100*max_load
<= sd
->imbalance_pct
*this_load
)
1821 * We're trying to get all the cpus to the average_load, so we don't
1822 * want to push ourselves above the average load, nor do we wish to
1823 * reduce the max loaded cpu below the average load, as either of these
1824 * actions would just result in more rebalancing later, and ping-pong
1825 * tasks around. Thus we look for the minimum possible imbalance.
1826 * Negative imbalances (*we* are more loaded than anyone else) will
1827 * be counted as no imbalance for these purposes -- we can't fix that
1828 * by pulling tasks to us. Be careful of negative numbers as they'll
1829 * appear as very large values with unsigned longs.
1831 /* How much load to actually move to equalise the imbalance */
1832 *imbalance
= min((max_load
- avg_load
) * busiest
->cpu_power
,
1833 (avg_load
- this_load
) * this->cpu_power
)
1836 if (*imbalance
< SCHED_LOAD_SCALE
) {
1837 unsigned long pwr_now
= 0, pwr_move
= 0;
1840 if (max_load
- this_load
>= SCHED_LOAD_SCALE
*2) {
1846 * OK, we don't have enough imbalance to justify moving tasks,
1847 * however we may be able to increase total CPU power used by
1851 pwr_now
+= busiest
->cpu_power
*min(SCHED_LOAD_SCALE
, max_load
);
1852 pwr_now
+= this->cpu_power
*min(SCHED_LOAD_SCALE
, this_load
);
1853 pwr_now
/= SCHED_LOAD_SCALE
;
1855 /* Amount of load we'd subtract */
1856 tmp
= SCHED_LOAD_SCALE
*SCHED_LOAD_SCALE
/busiest
->cpu_power
;
1858 pwr_move
+= busiest
->cpu_power
*min(SCHED_LOAD_SCALE
,
1861 /* Amount of load we'd add */
1862 if (max_load
*busiest
->cpu_power
<
1863 SCHED_LOAD_SCALE
*SCHED_LOAD_SCALE
)
1864 tmp
= max_load
*busiest
->cpu_power
/this->cpu_power
;
1866 tmp
= SCHED_LOAD_SCALE
*SCHED_LOAD_SCALE
/this->cpu_power
;
1867 pwr_move
+= this->cpu_power
*min(SCHED_LOAD_SCALE
, this_load
+ tmp
);
1868 pwr_move
/= SCHED_LOAD_SCALE
;
1870 /* Move if we gain throughput */
1871 if (pwr_move
<= pwr_now
)
1878 /* Get rid of the scaling factor, rounding down as we divide */
1879 *imbalance
= *imbalance
/ SCHED_LOAD_SCALE
;
1884 if (busiest
&& (idle
== NEWLY_IDLE
||
1885 (idle
== SCHED_IDLE
&& max_load
> SCHED_LOAD_SCALE
)) ) {
1895 * find_busiest_queue - find the busiest runqueue among the cpus in group.
1897 static runqueue_t
*find_busiest_queue(struct sched_group
*group
)
1899 unsigned long load
, max_load
= 0;
1900 runqueue_t
*busiest
= NULL
;
1903 for_each_cpu_mask(i
, group
->cpumask
) {
1904 load
= source_load(i
);
1906 if (load
> max_load
) {
1908 busiest
= cpu_rq(i
);
1916 * Check this_cpu to ensure it is balanced within domain. Attempt to move
1917 * tasks if there is an imbalance.
1919 * Called with this_rq unlocked.
1921 static int load_balance(int this_cpu
, runqueue_t
*this_rq
,
1922 struct sched_domain
*sd
, enum idle_type idle
)
1924 struct sched_group
*group
;
1925 runqueue_t
*busiest
;
1926 unsigned long imbalance
;
1927 int nr_moved
, all_pinned
;
1928 int active_balance
= 0;
1930 spin_lock(&this_rq
->lock
);
1931 schedstat_inc(sd
, lb_cnt
[idle
]);
1933 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
);
1935 schedstat_inc(sd
, lb_nobusyg
[idle
]);
1939 busiest
= find_busiest_queue(group
);
1941 schedstat_inc(sd
, lb_nobusyq
[idle
]);
1946 * This should be "impossible", but since load
1947 * balancing is inherently racy and statistical,
1948 * it could happen in theory.
1950 if (unlikely(busiest
== this_rq
)) {
1955 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
1958 if (busiest
->nr_running
> 1) {
1960 * Attempt to move tasks. If find_busiest_group has found
1961 * an imbalance but busiest->nr_running <= 1, the group is
1962 * still unbalanced. nr_moved simply stays zero, so it is
1963 * correctly treated as an imbalance.
1965 double_lock_balance(this_rq
, busiest
);
1966 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
1967 imbalance
, sd
, idle
,
1969 spin_unlock(&busiest
->lock
);
1971 /* All tasks on this runqueue were pinned by CPU affinity */
1972 if (unlikely(all_pinned
))
1976 spin_unlock(&this_rq
->lock
);
1979 schedstat_inc(sd
, lb_failed
[idle
]);
1980 sd
->nr_balance_failed
++;
1982 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
1984 spin_lock(&busiest
->lock
);
1985 if (!busiest
->active_balance
) {
1986 busiest
->active_balance
= 1;
1987 busiest
->push_cpu
= this_cpu
;
1990 spin_unlock(&busiest
->lock
);
1992 wake_up_process(busiest
->migration_thread
);
1995 * We've kicked active balancing, reset the failure
1998 sd
->nr_balance_failed
= sd
->cache_nice_tries
;
2001 sd
->nr_balance_failed
= 0;
2003 if (likely(!active_balance
)) {
2004 /* We were unbalanced, so reset the balancing interval */
2005 sd
->balance_interval
= sd
->min_interval
;
2008 * If we've begun active balancing, start to back off. This
2009 * case may not be covered by the all_pinned logic if there
2010 * is only 1 task on the busy runqueue (because we don't call
2013 if (sd
->balance_interval
< sd
->max_interval
)
2014 sd
->balance_interval
*= 2;
2020 spin_unlock(&this_rq
->lock
);
2022 schedstat_inc(sd
, lb_balanced
[idle
]);
2024 /* tune up the balancing interval */
2025 if (sd
->balance_interval
< sd
->max_interval
)
2026 sd
->balance_interval
*= 2;
2032 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2033 * tasks if there is an imbalance.
2035 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2036 * this_rq is locked.
2038 static int load_balance_newidle(int this_cpu
, runqueue_t
*this_rq
,
2039 struct sched_domain
*sd
)
2041 struct sched_group
*group
;
2042 runqueue_t
*busiest
= NULL
;
2043 unsigned long imbalance
;
2046 schedstat_inc(sd
, lb_cnt
[NEWLY_IDLE
]);
2047 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, NEWLY_IDLE
);
2049 schedstat_inc(sd
, lb_balanced
[NEWLY_IDLE
]);
2050 schedstat_inc(sd
, lb_nobusyg
[NEWLY_IDLE
]);
2054 busiest
= find_busiest_queue(group
);
2055 if (!busiest
|| busiest
== this_rq
) {
2056 schedstat_inc(sd
, lb_balanced
[NEWLY_IDLE
]);
2057 schedstat_inc(sd
, lb_nobusyq
[NEWLY_IDLE
]);
2061 /* Attempt to move tasks */
2062 double_lock_balance(this_rq
, busiest
);
2064 schedstat_add(sd
, lb_imbalance
[NEWLY_IDLE
], imbalance
);
2065 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2066 imbalance
, sd
, NEWLY_IDLE
, NULL
);
2068 schedstat_inc(sd
, lb_failed
[NEWLY_IDLE
]);
2070 spin_unlock(&busiest
->lock
);
2077 * idle_balance is called by schedule() if this_cpu is about to become
2078 * idle. Attempts to pull tasks from other CPUs.
2080 static inline void idle_balance(int this_cpu
, runqueue_t
*this_rq
)
2082 struct sched_domain
*sd
;
2084 for_each_domain(this_cpu
, sd
) {
2085 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
2086 if (load_balance_newidle(this_cpu
, this_rq
, sd
)) {
2087 /* We've pulled tasks over so stop searching */
2095 * active_load_balance is run by migration threads. It pushes running tasks
2096 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2097 * running on each physical CPU where possible, and avoids physical /
2098 * logical imbalances.
2100 * Called with busiest_rq locked.
2102 static void active_load_balance(runqueue_t
*busiest_rq
, int busiest_cpu
)
2104 struct sched_domain
*sd
;
2105 struct sched_group
*cpu_group
;
2106 runqueue_t
*target_rq
;
2107 cpumask_t visited_cpus
;
2111 * Search for suitable CPUs to push tasks to in successively higher
2112 * domains with SD_LOAD_BALANCE set.
2114 visited_cpus
= CPU_MASK_NONE
;
2115 for_each_domain(busiest_cpu
, sd
) {
2116 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2117 /* no more domains to search */
2120 schedstat_inc(sd
, alb_cnt
);
2122 cpu_group
= sd
->groups
;
2124 for_each_cpu_mask(cpu
, cpu_group
->cpumask
) {
2125 if (busiest_rq
->nr_running
<= 1)
2126 /* no more tasks left to move */
2128 if (cpu_isset(cpu
, visited_cpus
))
2130 cpu_set(cpu
, visited_cpus
);
2131 if (!cpu_and_siblings_are_idle(cpu
) || cpu
== busiest_cpu
)
2134 target_rq
= cpu_rq(cpu
);
2136 * This condition is "impossible", if it occurs
2137 * we need to fix it. Originally reported by
2138 * Bjorn Helgaas on a 128-cpu setup.
2140 BUG_ON(busiest_rq
== target_rq
);
2142 /* move a task from busiest_rq to target_rq */
2143 double_lock_balance(busiest_rq
, target_rq
);
2144 if (move_tasks(target_rq
, cpu
, busiest_rq
,
2145 1, sd
, SCHED_IDLE
, NULL
)) {
2146 schedstat_inc(sd
, alb_pushed
);
2148 schedstat_inc(sd
, alb_failed
);
2150 spin_unlock(&target_rq
->lock
);
2152 cpu_group
= cpu_group
->next
;
2153 } while (cpu_group
!= sd
->groups
);
2158 * rebalance_tick will get called every timer tick, on every CPU.
2160 * It checks each scheduling domain to see if it is due to be balanced,
2161 * and initiates a balancing operation if so.
2163 * Balancing parameters are set up in arch_init_sched_domains.
2166 /* Don't have all balancing operations going off at once */
2167 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2169 static void rebalance_tick(int this_cpu
, runqueue_t
*this_rq
,
2170 enum idle_type idle
)
2172 unsigned long old_load
, this_load
;
2173 unsigned long j
= jiffies
+ CPU_OFFSET(this_cpu
);
2174 struct sched_domain
*sd
;
2176 /* Update our load */
2177 old_load
= this_rq
->cpu_load
;
2178 this_load
= this_rq
->nr_running
* SCHED_LOAD_SCALE
;
2180 * Round up the averaging division if load is increasing. This
2181 * prevents us from getting stuck on 9 if the load is 10, for
2184 if (this_load
> old_load
)
2186 this_rq
->cpu_load
= (old_load
+ this_load
) / 2;
2188 for_each_domain(this_cpu
, sd
) {
2189 unsigned long interval
;
2191 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2194 interval
= sd
->balance_interval
;
2195 if (idle
!= SCHED_IDLE
)
2196 interval
*= sd
->busy_factor
;
2198 /* scale ms to jiffies */
2199 interval
= msecs_to_jiffies(interval
);
2200 if (unlikely(!interval
))
2203 if (j
- sd
->last_balance
>= interval
) {
2204 if (load_balance(this_cpu
, this_rq
, sd
, idle
)) {
2205 /* We've pulled tasks over so no longer idle */
2208 sd
->last_balance
+= interval
;
2214 * on UP we do not need to balance between CPUs:
2216 static inline void rebalance_tick(int cpu
, runqueue_t
*rq
, enum idle_type idle
)
2219 static inline void idle_balance(int cpu
, runqueue_t
*rq
)
2224 static inline int wake_priority_sleeper(runqueue_t
*rq
)
2227 #ifdef CONFIG_SCHED_SMT
2228 spin_lock(&rq
->lock
);
2230 * If an SMT sibling task has been put to sleep for priority
2231 * reasons reschedule the idle task to see if it can now run.
2233 if (rq
->nr_running
) {
2234 resched_task(rq
->idle
);
2237 spin_unlock(&rq
->lock
);
2242 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2244 EXPORT_PER_CPU_SYMBOL(kstat
);
2247 * This is called on clock ticks and on context switches.
2248 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2250 static inline void update_cpu_clock(task_t
*p
, runqueue_t
*rq
,
2251 unsigned long long now
)
2253 unsigned long long last
= max(p
->timestamp
, rq
->timestamp_last_tick
);
2254 p
->sched_time
+= now
- last
;
2258 * Return current->sched_time plus any more ns on the sched_clock
2259 * that have not yet been banked.
2261 unsigned long long current_sched_time(const task_t
*tsk
)
2263 unsigned long long ns
;
2264 unsigned long flags
;
2265 local_irq_save(flags
);
2266 ns
= max(tsk
->timestamp
, task_rq(tsk
)->timestamp_last_tick
);
2267 ns
= tsk
->sched_time
+ (sched_clock() - ns
);
2268 local_irq_restore(flags
);
2273 * We place interactive tasks back into the active array, if possible.
2275 * To guarantee that this does not starve expired tasks we ignore the
2276 * interactivity of a task if the first expired task had to wait more
2277 * than a 'reasonable' amount of time. This deadline timeout is
2278 * load-dependent, as the frequency of array switched decreases with
2279 * increasing number of running tasks. We also ignore the interactivity
2280 * if a better static_prio task has expired:
2282 #define EXPIRED_STARVING(rq) \
2283 ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2284 (jiffies - (rq)->expired_timestamp >= \
2285 STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2286 ((rq)->curr->static_prio > (rq)->best_expired_prio))
2289 * Account user cpu time to a process.
2290 * @p: the process that the cpu time gets accounted to
2291 * @hardirq_offset: the offset to subtract from hardirq_count()
2292 * @cputime: the cpu time spent in user space since the last update
2294 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
2296 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2299 p
->utime
= cputime_add(p
->utime
, cputime
);
2301 /* Add user time to cpustat. */
2302 tmp
= cputime_to_cputime64(cputime
);
2303 if (TASK_NICE(p
) > 0)
2304 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
2306 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
2310 * Account system cpu time to a process.
2311 * @p: the process that the cpu time gets accounted to
2312 * @hardirq_offset: the offset to subtract from hardirq_count()
2313 * @cputime: the cpu time spent in kernel space since the last update
2315 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
2318 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2319 runqueue_t
*rq
= this_rq();
2322 p
->stime
= cputime_add(p
->stime
, cputime
);
2324 /* Add system time to cpustat. */
2325 tmp
= cputime_to_cputime64(cputime
);
2326 if (hardirq_count() - hardirq_offset
)
2327 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
2328 else if (softirq_count())
2329 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
2330 else if (p
!= rq
->idle
)
2331 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
2332 else if (atomic_read(&rq
->nr_iowait
) > 0)
2333 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
2335 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
2336 /* Account for system time used */
2337 acct_update_integrals(p
);
2338 /* Update rss highwater mark */
2339 update_mem_hiwater(p
);
2343 * Account for involuntary wait time.
2344 * @p: the process from which the cpu time has been stolen
2345 * @steal: the cpu time spent in involuntary wait
2347 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
2349 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2350 cputime64_t tmp
= cputime_to_cputime64(steal
);
2351 runqueue_t
*rq
= this_rq();
2353 if (p
== rq
->idle
) {
2354 p
->stime
= cputime_add(p
->stime
, steal
);
2355 if (atomic_read(&rq
->nr_iowait
) > 0)
2356 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
2358 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
2360 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
2364 * This function gets called by the timer code, with HZ frequency.
2365 * We call it with interrupts disabled.
2367 * It also gets called by the fork code, when changing the parent's
2370 void scheduler_tick(void)
2372 int cpu
= smp_processor_id();
2373 runqueue_t
*rq
= this_rq();
2374 task_t
*p
= current
;
2375 unsigned long long now
= sched_clock();
2377 update_cpu_clock(p
, rq
, now
);
2379 rq
->timestamp_last_tick
= now
;
2381 if (p
== rq
->idle
) {
2382 if (wake_priority_sleeper(rq
))
2384 rebalance_tick(cpu
, rq
, SCHED_IDLE
);
2388 /* Task might have expired already, but not scheduled off yet */
2389 if (p
->array
!= rq
->active
) {
2390 set_tsk_need_resched(p
);
2393 spin_lock(&rq
->lock
);
2395 * The task was running during this tick - update the
2396 * time slice counter. Note: we do not update a thread's
2397 * priority until it either goes to sleep or uses up its
2398 * timeslice. This makes it possible for interactive tasks
2399 * to use up their timeslices at their highest priority levels.
2403 * RR tasks need a special form of timeslice management.
2404 * FIFO tasks have no timeslices.
2406 if ((p
->policy
== SCHED_RR
) && !--p
->time_slice
) {
2407 p
->time_slice
= task_timeslice(p
);
2408 p
->first_time_slice
= 0;
2409 set_tsk_need_resched(p
);
2411 /* put it at the end of the queue: */
2412 requeue_task(p
, rq
->active
);
2416 if (!--p
->time_slice
) {
2417 dequeue_task(p
, rq
->active
);
2418 set_tsk_need_resched(p
);
2419 p
->prio
= effective_prio(p
);
2420 p
->time_slice
= task_timeslice(p
);
2421 p
->first_time_slice
= 0;
2423 if (!rq
->expired_timestamp
)
2424 rq
->expired_timestamp
= jiffies
;
2425 if (!TASK_INTERACTIVE(p
) || EXPIRED_STARVING(rq
)) {
2426 enqueue_task(p
, rq
->expired
);
2427 if (p
->static_prio
< rq
->best_expired_prio
)
2428 rq
->best_expired_prio
= p
->static_prio
;
2430 enqueue_task(p
, rq
->active
);
2433 * Prevent a too long timeslice allowing a task to monopolize
2434 * the CPU. We do this by splitting up the timeslice into
2437 * Note: this does not mean the task's timeslices expire or
2438 * get lost in any way, they just might be preempted by
2439 * another task of equal priority. (one with higher
2440 * priority would have preempted this task already.) We
2441 * requeue this task to the end of the list on this priority
2442 * level, which is in essence a round-robin of tasks with
2445 * This only applies to tasks in the interactive
2446 * delta range with at least TIMESLICE_GRANULARITY to requeue.
2448 if (TASK_INTERACTIVE(p
) && !((task_timeslice(p
) -
2449 p
->time_slice
) % TIMESLICE_GRANULARITY(p
)) &&
2450 (p
->time_slice
>= TIMESLICE_GRANULARITY(p
)) &&
2451 (p
->array
== rq
->active
)) {
2453 requeue_task(p
, rq
->active
);
2454 set_tsk_need_resched(p
);
2458 spin_unlock(&rq
->lock
);
2460 rebalance_tick(cpu
, rq
, NOT_IDLE
);
2463 #ifdef CONFIG_SCHED_SMT
2464 static inline void wake_sleeping_dependent(int this_cpu
, runqueue_t
*this_rq
)
2466 struct sched_domain
*sd
= this_rq
->sd
;
2467 cpumask_t sibling_map
;
2470 if (!(sd
->flags
& SD_SHARE_CPUPOWER
))
2474 * Unlock the current runqueue because we have to lock in
2475 * CPU order to avoid deadlocks. Caller knows that we might
2476 * unlock. We keep IRQs disabled.
2478 spin_unlock(&this_rq
->lock
);
2480 sibling_map
= sd
->span
;
2482 for_each_cpu_mask(i
, sibling_map
)
2483 spin_lock(&cpu_rq(i
)->lock
);
2485 * We clear this CPU from the mask. This both simplifies the
2486 * inner loop and keps this_rq locked when we exit:
2488 cpu_clear(this_cpu
, sibling_map
);
2490 for_each_cpu_mask(i
, sibling_map
) {
2491 runqueue_t
*smt_rq
= cpu_rq(i
);
2494 * If an SMT sibling task is sleeping due to priority
2495 * reasons wake it up now.
2497 if (smt_rq
->curr
== smt_rq
->idle
&& smt_rq
->nr_running
)
2498 resched_task(smt_rq
->idle
);
2501 for_each_cpu_mask(i
, sibling_map
)
2502 spin_unlock(&cpu_rq(i
)->lock
);
2504 * We exit with this_cpu's rq still held and IRQs
2509 static inline int dependent_sleeper(int this_cpu
, runqueue_t
*this_rq
)
2511 struct sched_domain
*sd
= this_rq
->sd
;
2512 cpumask_t sibling_map
;
2513 prio_array_t
*array
;
2517 if (!(sd
->flags
& SD_SHARE_CPUPOWER
))
2521 * The same locking rules and details apply as for
2522 * wake_sleeping_dependent():
2524 spin_unlock(&this_rq
->lock
);
2525 sibling_map
= sd
->span
;
2526 for_each_cpu_mask(i
, sibling_map
)
2527 spin_lock(&cpu_rq(i
)->lock
);
2528 cpu_clear(this_cpu
, sibling_map
);
2531 * Establish next task to be run - it might have gone away because
2532 * we released the runqueue lock above:
2534 if (!this_rq
->nr_running
)
2536 array
= this_rq
->active
;
2537 if (!array
->nr_active
)
2538 array
= this_rq
->expired
;
2539 BUG_ON(!array
->nr_active
);
2541 p
= list_entry(array
->queue
[sched_find_first_bit(array
->bitmap
)].next
,
2544 for_each_cpu_mask(i
, sibling_map
) {
2545 runqueue_t
*smt_rq
= cpu_rq(i
);
2546 task_t
*smt_curr
= smt_rq
->curr
;
2549 * If a user task with lower static priority than the
2550 * running task on the SMT sibling is trying to schedule,
2551 * delay it till there is proportionately less timeslice
2552 * left of the sibling task to prevent a lower priority
2553 * task from using an unfair proportion of the
2554 * physical cpu's resources. -ck
2556 if (((smt_curr
->time_slice
* (100 - sd
->per_cpu_gain
) / 100) >
2557 task_timeslice(p
) || rt_task(smt_curr
)) &&
2558 p
->mm
&& smt_curr
->mm
&& !rt_task(p
))
2562 * Reschedule a lower priority task on the SMT sibling,
2563 * or wake it up if it has been put to sleep for priority
2566 if ((((p
->time_slice
* (100 - sd
->per_cpu_gain
) / 100) >
2567 task_timeslice(smt_curr
) || rt_task(p
)) &&
2568 smt_curr
->mm
&& p
->mm
&& !rt_task(smt_curr
)) ||
2569 (smt_curr
== smt_rq
->idle
&& smt_rq
->nr_running
))
2570 resched_task(smt_curr
);
2573 for_each_cpu_mask(i
, sibling_map
)
2574 spin_unlock(&cpu_rq(i
)->lock
);
2578 static inline void wake_sleeping_dependent(int this_cpu
, runqueue_t
*this_rq
)
2582 static inline int dependent_sleeper(int this_cpu
, runqueue_t
*this_rq
)
2588 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
2590 void fastcall
add_preempt_count(int val
)
2595 BUG_ON((preempt_count() < 0));
2596 preempt_count() += val
;
2598 * Spinlock count overflowing soon?
2600 BUG_ON((preempt_count() & PREEMPT_MASK
) >= PREEMPT_MASK
-10);
2602 EXPORT_SYMBOL(add_preempt_count
);
2604 void fastcall
sub_preempt_count(int val
)
2609 BUG_ON(val
> preempt_count());
2611 * Is the spinlock portion underflowing?
2613 BUG_ON((val
< PREEMPT_MASK
) && !(preempt_count() & PREEMPT_MASK
));
2614 preempt_count() -= val
;
2616 EXPORT_SYMBOL(sub_preempt_count
);
2621 * schedule() is the main scheduler function.
2623 asmlinkage
void __sched
schedule(void)
2626 task_t
*prev
, *next
;
2628 prio_array_t
*array
;
2629 struct list_head
*queue
;
2630 unsigned long long now
;
2631 unsigned long run_time
;
2635 * Test if we are atomic. Since do_exit() needs to call into
2636 * schedule() atomically, we ignore that path for now.
2637 * Otherwise, whine if we are scheduling when we should not be.
2639 if (likely(!current
->exit_state
)) {
2640 if (unlikely(in_atomic())) {
2641 printk(KERN_ERR
"scheduling while atomic: "
2643 current
->comm
, preempt_count(), current
->pid
);
2647 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2652 release_kernel_lock(prev
);
2653 need_resched_nonpreemptible
:
2657 * The idle thread is not allowed to schedule!
2658 * Remove this check after it has been exercised a bit.
2660 if (unlikely(prev
== rq
->idle
) && prev
->state
!= TASK_RUNNING
) {
2661 printk(KERN_ERR
"bad: scheduling from the idle thread!\n");
2665 schedstat_inc(rq
, sched_cnt
);
2666 now
= sched_clock();
2667 if (likely((long long)(now
- prev
->timestamp
) < NS_MAX_SLEEP_AVG
)) {
2668 run_time
= now
- prev
->timestamp
;
2669 if (unlikely((long long)(now
- prev
->timestamp
) < 0))
2672 run_time
= NS_MAX_SLEEP_AVG
;
2675 * Tasks charged proportionately less run_time at high sleep_avg to
2676 * delay them losing their interactive status
2678 run_time
/= (CURRENT_BONUS(prev
) ? : 1);
2680 spin_lock_irq(&rq
->lock
);
2682 if (unlikely(prev
->flags
& PF_DEAD
))
2683 prev
->state
= EXIT_DEAD
;
2685 switch_count
= &prev
->nivcsw
;
2686 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
2687 switch_count
= &prev
->nvcsw
;
2688 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
2689 unlikely(signal_pending(prev
))))
2690 prev
->state
= TASK_RUNNING
;
2692 if (prev
->state
== TASK_UNINTERRUPTIBLE
)
2693 rq
->nr_uninterruptible
++;
2694 deactivate_task(prev
, rq
);
2698 cpu
= smp_processor_id();
2699 if (unlikely(!rq
->nr_running
)) {
2701 idle_balance(cpu
, rq
);
2702 if (!rq
->nr_running
) {
2704 rq
->expired_timestamp
= 0;
2705 wake_sleeping_dependent(cpu
, rq
);
2707 * wake_sleeping_dependent() might have released
2708 * the runqueue, so break out if we got new
2711 if (!rq
->nr_running
)
2715 if (dependent_sleeper(cpu
, rq
)) {
2720 * dependent_sleeper() releases and reacquires the runqueue
2721 * lock, hence go into the idle loop if the rq went
2724 if (unlikely(!rq
->nr_running
))
2729 if (unlikely(!array
->nr_active
)) {
2731 * Switch the active and expired arrays.
2733 schedstat_inc(rq
, sched_switch
);
2734 rq
->active
= rq
->expired
;
2735 rq
->expired
= array
;
2737 rq
->expired_timestamp
= 0;
2738 rq
->best_expired_prio
= MAX_PRIO
;
2741 idx
= sched_find_first_bit(array
->bitmap
);
2742 queue
= array
->queue
+ idx
;
2743 next
= list_entry(queue
->next
, task_t
, run_list
);
2745 if (!rt_task(next
) && next
->activated
> 0) {
2746 unsigned long long delta
= now
- next
->timestamp
;
2747 if (unlikely((long long)(now
- next
->timestamp
) < 0))
2750 if (next
->activated
== 1)
2751 delta
= delta
* (ON_RUNQUEUE_WEIGHT
* 128 / 100) / 128;
2753 array
= next
->array
;
2754 dequeue_task(next
, array
);
2755 recalc_task_prio(next
, next
->timestamp
+ delta
);
2756 enqueue_task(next
, array
);
2758 next
->activated
= 0;
2760 if (next
== rq
->idle
)
2761 schedstat_inc(rq
, sched_goidle
);
2763 clear_tsk_need_resched(prev
);
2764 rcu_qsctr_inc(task_cpu(prev
));
2766 update_cpu_clock(prev
, rq
, now
);
2768 prev
->sleep_avg
-= run_time
;
2769 if ((long)prev
->sleep_avg
<= 0)
2770 prev
->sleep_avg
= 0;
2771 prev
->timestamp
= prev
->last_ran
= now
;
2773 sched_info_switch(prev
, next
);
2774 if (likely(prev
!= next
)) {
2775 next
->timestamp
= now
;
2780 prepare_arch_switch(rq
, next
);
2781 prev
= context_switch(rq
, prev
, next
);
2784 finish_task_switch(prev
);
2786 spin_unlock_irq(&rq
->lock
);
2789 if (unlikely(reacquire_kernel_lock(prev
) < 0))
2790 goto need_resched_nonpreemptible
;
2791 preempt_enable_no_resched();
2792 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
2796 EXPORT_SYMBOL(schedule
);
2798 #ifdef CONFIG_PREEMPT
2800 * this is is the entry point to schedule() from in-kernel preemption
2801 * off of preempt_enable. Kernel preemptions off return from interrupt
2802 * occur there and call schedule directly.
2804 asmlinkage
void __sched
preempt_schedule(void)
2806 struct thread_info
*ti
= current_thread_info();
2807 #ifdef CONFIG_PREEMPT_BKL
2808 struct task_struct
*task
= current
;
2809 int saved_lock_depth
;
2812 * If there is a non-zero preempt_count or interrupts are disabled,
2813 * we do not want to preempt the current task. Just return..
2815 if (unlikely(ti
->preempt_count
|| irqs_disabled()))
2819 add_preempt_count(PREEMPT_ACTIVE
);
2821 * We keep the big kernel semaphore locked, but we
2822 * clear ->lock_depth so that schedule() doesnt
2823 * auto-release the semaphore:
2825 #ifdef CONFIG_PREEMPT_BKL
2826 saved_lock_depth
= task
->lock_depth
;
2827 task
->lock_depth
= -1;
2830 #ifdef CONFIG_PREEMPT_BKL
2831 task
->lock_depth
= saved_lock_depth
;
2833 sub_preempt_count(PREEMPT_ACTIVE
);
2835 /* we could miss a preemption opportunity between schedule and now */
2837 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
2841 EXPORT_SYMBOL(preempt_schedule
);
2844 * this is is the entry point to schedule() from kernel preemption
2845 * off of irq context.
2846 * Note, that this is called and return with irqs disabled. This will
2847 * protect us against recursive calling from irq.
2849 asmlinkage
void __sched
preempt_schedule_irq(void)
2851 struct thread_info
*ti
= current_thread_info();
2852 #ifdef CONFIG_PREEMPT_BKL
2853 struct task_struct
*task
= current
;
2854 int saved_lock_depth
;
2856 /* Catch callers which need to be fixed*/
2857 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
2860 add_preempt_count(PREEMPT_ACTIVE
);
2862 * We keep the big kernel semaphore locked, but we
2863 * clear ->lock_depth so that schedule() doesnt
2864 * auto-release the semaphore:
2866 #ifdef CONFIG_PREEMPT_BKL
2867 saved_lock_depth
= task
->lock_depth
;
2868 task
->lock_depth
= -1;
2872 local_irq_disable();
2873 #ifdef CONFIG_PREEMPT_BKL
2874 task
->lock_depth
= saved_lock_depth
;
2876 sub_preempt_count(PREEMPT_ACTIVE
);
2878 /* we could miss a preemption opportunity between schedule and now */
2880 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
2884 #endif /* CONFIG_PREEMPT */
2886 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
, void *key
)
2888 task_t
*p
= curr
->private;
2889 return try_to_wake_up(p
, mode
, sync
);
2892 EXPORT_SYMBOL(default_wake_function
);
2895 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
2896 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
2897 * number) then we wake all the non-exclusive tasks and one exclusive task.
2899 * There are circumstances in which we can try to wake a task which has already
2900 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
2901 * zero in this (rare) case, and we handle it by continuing to scan the queue.
2903 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
2904 int nr_exclusive
, int sync
, void *key
)
2906 struct list_head
*tmp
, *next
;
2908 list_for_each_safe(tmp
, next
, &q
->task_list
) {
2911 curr
= list_entry(tmp
, wait_queue_t
, task_list
);
2912 flags
= curr
->flags
;
2913 if (curr
->func(curr
, mode
, sync
, key
) &&
2914 (flags
& WQ_FLAG_EXCLUSIVE
) &&
2921 * __wake_up - wake up threads blocked on a waitqueue.
2923 * @mode: which threads
2924 * @nr_exclusive: how many wake-one or wake-many threads to wake up
2925 * @key: is directly passed to the wakeup function
2927 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
2928 int nr_exclusive
, void *key
)
2930 unsigned long flags
;
2932 spin_lock_irqsave(&q
->lock
, flags
);
2933 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
2934 spin_unlock_irqrestore(&q
->lock
, flags
);
2937 EXPORT_SYMBOL(__wake_up
);
2940 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
2942 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
2944 __wake_up_common(q
, mode
, 1, 0, NULL
);
2948 * __wake_up_sync - wake up threads blocked on a waitqueue.
2950 * @mode: which threads
2951 * @nr_exclusive: how many wake-one or wake-many threads to wake up
2953 * The sync wakeup differs that the waker knows that it will schedule
2954 * away soon, so while the target thread will be woken up, it will not
2955 * be migrated to another CPU - ie. the two threads are 'synchronized'
2956 * with each other. This can prevent needless bouncing between CPUs.
2958 * On UP it can prevent extra preemption.
2960 void fastcall
__wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
2962 unsigned long flags
;
2968 if (unlikely(!nr_exclusive
))
2971 spin_lock_irqsave(&q
->lock
, flags
);
2972 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
2973 spin_unlock_irqrestore(&q
->lock
, flags
);
2975 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
2977 void fastcall
complete(struct completion
*x
)
2979 unsigned long flags
;
2981 spin_lock_irqsave(&x
->wait
.lock
, flags
);
2983 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
2985 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
2987 EXPORT_SYMBOL(complete
);
2989 void fastcall
complete_all(struct completion
*x
)
2991 unsigned long flags
;
2993 spin_lock_irqsave(&x
->wait
.lock
, flags
);
2994 x
->done
+= UINT_MAX
/2;
2995 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
2997 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
2999 EXPORT_SYMBOL(complete_all
);
3001 void fastcall __sched
wait_for_completion(struct completion
*x
)
3004 spin_lock_irq(&x
->wait
.lock
);
3006 DECLARE_WAITQUEUE(wait
, current
);
3008 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3009 __add_wait_queue_tail(&x
->wait
, &wait
);
3011 __set_current_state(TASK_UNINTERRUPTIBLE
);
3012 spin_unlock_irq(&x
->wait
.lock
);
3014 spin_lock_irq(&x
->wait
.lock
);
3016 __remove_wait_queue(&x
->wait
, &wait
);
3019 spin_unlock_irq(&x
->wait
.lock
);
3021 EXPORT_SYMBOL(wait_for_completion
);
3023 unsigned long fastcall __sched
3024 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3028 spin_lock_irq(&x
->wait
.lock
);
3030 DECLARE_WAITQUEUE(wait
, current
);
3032 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3033 __add_wait_queue_tail(&x
->wait
, &wait
);
3035 __set_current_state(TASK_UNINTERRUPTIBLE
);
3036 spin_unlock_irq(&x
->wait
.lock
);
3037 timeout
= schedule_timeout(timeout
);
3038 spin_lock_irq(&x
->wait
.lock
);
3040 __remove_wait_queue(&x
->wait
, &wait
);
3044 __remove_wait_queue(&x
->wait
, &wait
);
3048 spin_unlock_irq(&x
->wait
.lock
);
3051 EXPORT_SYMBOL(wait_for_completion_timeout
);
3053 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3059 spin_lock_irq(&x
->wait
.lock
);
3061 DECLARE_WAITQUEUE(wait
, current
);
3063 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3064 __add_wait_queue_tail(&x
->wait
, &wait
);
3066 if (signal_pending(current
)) {
3068 __remove_wait_queue(&x
->wait
, &wait
);
3071 __set_current_state(TASK_INTERRUPTIBLE
);
3072 spin_unlock_irq(&x
->wait
.lock
);
3074 spin_lock_irq(&x
->wait
.lock
);
3076 __remove_wait_queue(&x
->wait
, &wait
);
3080 spin_unlock_irq(&x
->wait
.lock
);
3084 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3086 unsigned long fastcall __sched
3087 wait_for_completion_interruptible_timeout(struct completion
*x
,
3088 unsigned long timeout
)
3092 spin_lock_irq(&x
->wait
.lock
);
3094 DECLARE_WAITQUEUE(wait
, current
);
3096 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3097 __add_wait_queue_tail(&x
->wait
, &wait
);
3099 if (signal_pending(current
)) {
3100 timeout
= -ERESTARTSYS
;
3101 __remove_wait_queue(&x
->wait
, &wait
);
3104 __set_current_state(TASK_INTERRUPTIBLE
);
3105 spin_unlock_irq(&x
->wait
.lock
);
3106 timeout
= schedule_timeout(timeout
);
3107 spin_lock_irq(&x
->wait
.lock
);
3109 __remove_wait_queue(&x
->wait
, &wait
);
3113 __remove_wait_queue(&x
->wait
, &wait
);
3117 spin_unlock_irq(&x
->wait
.lock
);
3120 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3123 #define SLEEP_ON_VAR \
3124 unsigned long flags; \
3125 wait_queue_t wait; \
3126 init_waitqueue_entry(&wait, current);
3128 #define SLEEP_ON_HEAD \
3129 spin_lock_irqsave(&q->lock,flags); \
3130 __add_wait_queue(q, &wait); \
3131 spin_unlock(&q->lock);
3133 #define SLEEP_ON_TAIL \
3134 spin_lock_irq(&q->lock); \
3135 __remove_wait_queue(q, &wait); \
3136 spin_unlock_irqrestore(&q->lock, flags);
3138 void fastcall __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3142 current
->state
= TASK_INTERRUPTIBLE
;
3149 EXPORT_SYMBOL(interruptible_sleep_on
);
3151 long fastcall __sched
interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3155 current
->state
= TASK_INTERRUPTIBLE
;
3158 timeout
= schedule_timeout(timeout
);
3164 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3166 void fastcall __sched
sleep_on(wait_queue_head_t
*q
)
3170 current
->state
= TASK_UNINTERRUPTIBLE
;
3177 EXPORT_SYMBOL(sleep_on
);
3179 long fastcall __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3183 current
->state
= TASK_UNINTERRUPTIBLE
;
3186 timeout
= schedule_timeout(timeout
);
3192 EXPORT_SYMBOL(sleep_on_timeout
);
3194 void set_user_nice(task_t
*p
, long nice
)
3196 unsigned long flags
;
3197 prio_array_t
*array
;
3199 int old_prio
, new_prio
, delta
;
3201 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3204 * We have to be careful, if called from sys_setpriority(),
3205 * the task might be in the middle of scheduling on another CPU.
3207 rq
= task_rq_lock(p
, &flags
);
3209 * The RT priorities are set via sched_setscheduler(), but we still
3210 * allow the 'normal' nice value to be set - but as expected
3211 * it wont have any effect on scheduling until the task is
3215 p
->static_prio
= NICE_TO_PRIO(nice
);
3220 dequeue_task(p
, array
);
3223 new_prio
= NICE_TO_PRIO(nice
);
3224 delta
= new_prio
- old_prio
;
3225 p
->static_prio
= NICE_TO_PRIO(nice
);
3229 enqueue_task(p
, array
);
3231 * If the task increased its priority or is running and
3232 * lowered its priority, then reschedule its CPU:
3234 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3235 resched_task(rq
->curr
);
3238 task_rq_unlock(rq
, &flags
);
3241 EXPORT_SYMBOL(set_user_nice
);
3244 * can_nice - check if a task can reduce its nice value
3248 int can_nice(const task_t
*p
, const int nice
)
3250 /* convert nice value [19,-20] to rlimit style value [0,39] */
3251 int nice_rlim
= 19 - nice
;
3252 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
3253 capable(CAP_SYS_NICE
));
3256 #ifdef __ARCH_WANT_SYS_NICE
3259 * sys_nice - change the priority of the current process.
3260 * @increment: priority increment
3262 * sys_setpriority is a more generic, but much slower function that
3263 * does similar things.
3265 asmlinkage
long sys_nice(int increment
)
3271 * Setpriority might change our priority at the same moment.
3272 * We don't have to worry. Conceptually one call occurs first
3273 * and we have a single winner.
3275 if (increment
< -40)
3280 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
3286 if (increment
< 0 && !can_nice(current
, nice
))
3289 retval
= security_task_setnice(current
, nice
);
3293 set_user_nice(current
, nice
);
3300 * task_prio - return the priority value of a given task.
3301 * @p: the task in question.
3303 * This is the priority value as seen by users in /proc.
3304 * RT tasks are offset by -200. Normal tasks are centered
3305 * around 0, value goes from -16 to +15.
3307 int task_prio(const task_t
*p
)
3309 return p
->prio
- MAX_RT_PRIO
;
3313 * task_nice - return the nice value of a given task.
3314 * @p: the task in question.
3316 int task_nice(const task_t
*p
)
3318 return TASK_NICE(p
);
3322 * The only users of task_nice are binfmt_elf and binfmt_elf32.
3323 * binfmt_elf is no longer modular, but binfmt_elf32 still is.
3324 * Therefore, task_nice is needed if there is a compat_mode.
3326 #ifdef CONFIG_COMPAT
3327 EXPORT_SYMBOL_GPL(task_nice
);
3331 * idle_cpu - is a given cpu idle currently?
3332 * @cpu: the processor in question.
3334 int idle_cpu(int cpu
)
3336 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
3339 EXPORT_SYMBOL_GPL(idle_cpu
);
3342 * idle_task - return the idle task for a given cpu.
3343 * @cpu: the processor in question.
3345 task_t
*idle_task(int cpu
)
3347 return cpu_rq(cpu
)->idle
;
3351 * find_process_by_pid - find a process with a matching PID value.
3352 * @pid: the pid in question.
3354 static inline task_t
*find_process_by_pid(pid_t pid
)
3356 return pid
? find_task_by_pid(pid
) : current
;
3359 /* Actually do priority change: must hold rq lock. */
3360 static void __setscheduler(struct task_struct
*p
, int policy
, int prio
)
3364 p
->rt_priority
= prio
;
3365 if (policy
!= SCHED_NORMAL
)
3366 p
->prio
= MAX_USER_RT_PRIO
-1 - p
->rt_priority
;
3368 p
->prio
= p
->static_prio
;
3372 * sched_setscheduler - change the scheduling policy and/or RT priority of
3374 * @p: the task in question.
3375 * @policy: new policy.
3376 * @param: structure containing the new RT priority.
3378 int sched_setscheduler(struct task_struct
*p
, int policy
, struct sched_param
*param
)
3381 int oldprio
, oldpolicy
= -1;
3382 prio_array_t
*array
;
3383 unsigned long flags
;
3387 /* double check policy once rq lock held */
3389 policy
= oldpolicy
= p
->policy
;
3390 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3391 policy
!= SCHED_NORMAL
)
3394 * Valid priorities for SCHED_FIFO and SCHED_RR are
3395 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL is 0.
3397 if (param
->sched_priority
< 0 ||
3398 param
->sched_priority
> MAX_USER_RT_PRIO
-1)
3400 if ((policy
== SCHED_NORMAL
) != (param
->sched_priority
== 0))
3403 if ((policy
== SCHED_FIFO
|| policy
== SCHED_RR
) &&
3404 param
->sched_priority
> p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
&&
3405 !capable(CAP_SYS_NICE
))
3407 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
3408 !capable(CAP_SYS_NICE
))
3411 retval
= security_task_setscheduler(p
, policy
, param
);
3415 * To be able to change p->policy safely, the apropriate
3416 * runqueue lock must be held.
3418 rq
= task_rq_lock(p
, &flags
);
3419 /* recheck policy now with rq lock held */
3420 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3421 policy
= oldpolicy
= -1;
3422 task_rq_unlock(rq
, &flags
);
3427 deactivate_task(p
, rq
);
3429 __setscheduler(p
, policy
, param
->sched_priority
);
3431 __activate_task(p
, rq
);
3433 * Reschedule if we are currently running on this runqueue and
3434 * our priority decreased, or if we are not currently running on
3435 * this runqueue and our priority is higher than the current's
3437 if (task_running(rq
, p
)) {
3438 if (p
->prio
> oldprio
)
3439 resched_task(rq
->curr
);
3440 } else if (TASK_PREEMPTS_CURR(p
, rq
))
3441 resched_task(rq
->curr
);
3443 task_rq_unlock(rq
, &flags
);
3446 EXPORT_SYMBOL_GPL(sched_setscheduler
);
3448 static int do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
3451 struct sched_param lparam
;
3452 struct task_struct
*p
;
3454 if (!param
|| pid
< 0)
3456 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
3458 read_lock_irq(&tasklist_lock
);
3459 p
= find_process_by_pid(pid
);
3461 read_unlock_irq(&tasklist_lock
);
3464 retval
= sched_setscheduler(p
, policy
, &lparam
);
3465 read_unlock_irq(&tasklist_lock
);
3470 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3471 * @pid: the pid in question.
3472 * @policy: new policy.
3473 * @param: structure containing the new RT priority.
3475 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
3476 struct sched_param __user
*param
)
3478 return do_sched_setscheduler(pid
, policy
, param
);
3482 * sys_sched_setparam - set/change the RT priority of a thread
3483 * @pid: the pid in question.
3484 * @param: structure containing the new RT priority.
3486 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
3488 return do_sched_setscheduler(pid
, -1, param
);
3492 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3493 * @pid: the pid in question.
3495 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
3497 int retval
= -EINVAL
;
3504 read_lock(&tasklist_lock
);
3505 p
= find_process_by_pid(pid
);
3507 retval
= security_task_getscheduler(p
);
3511 read_unlock(&tasklist_lock
);
3518 * sys_sched_getscheduler - get the RT priority of a thread
3519 * @pid: the pid in question.
3520 * @param: structure containing the RT priority.
3522 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
3524 struct sched_param lp
;
3525 int retval
= -EINVAL
;
3528 if (!param
|| pid
< 0)
3531 read_lock(&tasklist_lock
);
3532 p
= find_process_by_pid(pid
);
3537 retval
= security_task_getscheduler(p
);
3541 lp
.sched_priority
= p
->rt_priority
;
3542 read_unlock(&tasklist_lock
);
3545 * This one might sleep, we cannot do it with a spinlock held ...
3547 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
3553 read_unlock(&tasklist_lock
);
3557 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
3561 cpumask_t cpus_allowed
;
3564 read_lock(&tasklist_lock
);
3566 p
= find_process_by_pid(pid
);
3568 read_unlock(&tasklist_lock
);
3569 unlock_cpu_hotplug();
3574 * It is not safe to call set_cpus_allowed with the
3575 * tasklist_lock held. We will bump the task_struct's
3576 * usage count and then drop tasklist_lock.
3579 read_unlock(&tasklist_lock
);
3582 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
3583 !capable(CAP_SYS_NICE
))
3586 cpus_allowed
= cpuset_cpus_allowed(p
);
3587 cpus_and(new_mask
, new_mask
, cpus_allowed
);
3588 retval
= set_cpus_allowed(p
, new_mask
);
3592 unlock_cpu_hotplug();
3596 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
3597 cpumask_t
*new_mask
)
3599 if (len
< sizeof(cpumask_t
)) {
3600 memset(new_mask
, 0, sizeof(cpumask_t
));
3601 } else if (len
> sizeof(cpumask_t
)) {
3602 len
= sizeof(cpumask_t
);
3604 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
3608 * sys_sched_setaffinity - set the cpu affinity of a process
3609 * @pid: pid of the process
3610 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3611 * @user_mask_ptr: user-space pointer to the new cpu mask
3613 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
3614 unsigned long __user
*user_mask_ptr
)
3619 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
3623 return sched_setaffinity(pid
, new_mask
);
3627 * Represents all cpu's present in the system
3628 * In systems capable of hotplug, this map could dynamically grow
3629 * as new cpu's are detected in the system via any platform specific
3630 * method, such as ACPI for e.g.
3633 cpumask_t cpu_present_map
;
3634 EXPORT_SYMBOL(cpu_present_map
);
3637 cpumask_t cpu_online_map
= CPU_MASK_ALL
;
3638 cpumask_t cpu_possible_map
= CPU_MASK_ALL
;
3641 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
3647 read_lock(&tasklist_lock
);
3650 p
= find_process_by_pid(pid
);
3655 cpus_and(*mask
, p
->cpus_allowed
, cpu_possible_map
);
3658 read_unlock(&tasklist_lock
);
3659 unlock_cpu_hotplug();
3667 * sys_sched_getaffinity - get the cpu affinity of a process
3668 * @pid: pid of the process
3669 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3670 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3672 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
3673 unsigned long __user
*user_mask_ptr
)
3678 if (len
< sizeof(cpumask_t
))
3681 ret
= sched_getaffinity(pid
, &mask
);
3685 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
3688 return sizeof(cpumask_t
);
3692 * sys_sched_yield - yield the current processor to other threads.
3694 * this function yields the current CPU by moving the calling thread
3695 * to the expired array. If there are no other threads running on this
3696 * CPU then this function will return.
3698 asmlinkage
long sys_sched_yield(void)
3700 runqueue_t
*rq
= this_rq_lock();
3701 prio_array_t
*array
= current
->array
;
3702 prio_array_t
*target
= rq
->expired
;
3704 schedstat_inc(rq
, yld_cnt
);
3706 * We implement yielding by moving the task into the expired
3709 * (special rule: RT tasks will just roundrobin in the active
3712 if (rt_task(current
))
3713 target
= rq
->active
;
3715 if (current
->array
->nr_active
== 1) {
3716 schedstat_inc(rq
, yld_act_empty
);
3717 if (!rq
->expired
->nr_active
)
3718 schedstat_inc(rq
, yld_both_empty
);
3719 } else if (!rq
->expired
->nr_active
)
3720 schedstat_inc(rq
, yld_exp_empty
);
3722 if (array
!= target
) {
3723 dequeue_task(current
, array
);
3724 enqueue_task(current
, target
);
3727 * requeue_task is cheaper so perform that if possible.
3729 requeue_task(current
, array
);
3732 * Since we are going to call schedule() anyway, there's
3733 * no need to preempt or enable interrupts:
3735 __release(rq
->lock
);
3736 _raw_spin_unlock(&rq
->lock
);
3737 preempt_enable_no_resched();
3744 static inline void __cond_resched(void)
3747 add_preempt_count(PREEMPT_ACTIVE
);
3749 sub_preempt_count(PREEMPT_ACTIVE
);
3750 } while (need_resched());
3753 int __sched
cond_resched(void)
3755 if (need_resched()) {
3762 EXPORT_SYMBOL(cond_resched
);
3765 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
3766 * call schedule, and on return reacquire the lock.
3768 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
3769 * operations here to prevent schedule() from being called twice (once via
3770 * spin_unlock(), once by hand).
3772 int cond_resched_lock(spinlock_t
* lock
)
3776 if (need_lockbreak(lock
)) {
3782 if (need_resched()) {
3783 _raw_spin_unlock(lock
);
3784 preempt_enable_no_resched();
3792 EXPORT_SYMBOL(cond_resched_lock
);
3794 int __sched
cond_resched_softirq(void)
3796 BUG_ON(!in_softirq());
3798 if (need_resched()) {
3799 __local_bh_enable();
3807 EXPORT_SYMBOL(cond_resched_softirq
);
3811 * yield - yield the current processor to other threads.
3813 * this is a shortcut for kernel-space yielding - it marks the
3814 * thread runnable and calls sys_sched_yield().
3816 void __sched
yield(void)
3818 set_current_state(TASK_RUNNING
);
3822 EXPORT_SYMBOL(yield
);
3825 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
3826 * that process accounting knows that this is a task in IO wait state.
3828 * But don't do that if it is a deliberate, throttling IO wait (this task
3829 * has set its backing_dev_info: the queue against which it should throttle)
3831 void __sched
io_schedule(void)
3833 struct runqueue
*rq
= &per_cpu(runqueues
, raw_smp_processor_id());
3835 atomic_inc(&rq
->nr_iowait
);
3837 atomic_dec(&rq
->nr_iowait
);
3840 EXPORT_SYMBOL(io_schedule
);
3842 long __sched
io_schedule_timeout(long timeout
)
3844 struct runqueue
*rq
= &per_cpu(runqueues
, raw_smp_processor_id());
3847 atomic_inc(&rq
->nr_iowait
);
3848 ret
= schedule_timeout(timeout
);
3849 atomic_dec(&rq
->nr_iowait
);
3854 * sys_sched_get_priority_max - return maximum RT priority.
3855 * @policy: scheduling class.
3857 * this syscall returns the maximum rt_priority that can be used
3858 * by a given scheduling class.
3860 asmlinkage
long sys_sched_get_priority_max(int policy
)
3867 ret
= MAX_USER_RT_PRIO
-1;
3877 * sys_sched_get_priority_min - return minimum RT priority.
3878 * @policy: scheduling class.
3880 * this syscall returns the minimum rt_priority that can be used
3881 * by a given scheduling class.
3883 asmlinkage
long sys_sched_get_priority_min(int policy
)
3899 * sys_sched_rr_get_interval - return the default timeslice of a process.
3900 * @pid: pid of the process.
3901 * @interval: userspace pointer to the timeslice value.
3903 * this syscall writes the default timeslice value of a given process
3904 * into the user-space timespec buffer. A value of '0' means infinity.
3907 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
3909 int retval
= -EINVAL
;
3917 read_lock(&tasklist_lock
);
3918 p
= find_process_by_pid(pid
);
3922 retval
= security_task_getscheduler(p
);
3926 jiffies_to_timespec(p
->policy
& SCHED_FIFO
?
3927 0 : task_timeslice(p
), &t
);
3928 read_unlock(&tasklist_lock
);
3929 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
3933 read_unlock(&tasklist_lock
);
3937 static inline struct task_struct
*eldest_child(struct task_struct
*p
)
3939 if (list_empty(&p
->children
)) return NULL
;
3940 return list_entry(p
->children
.next
,struct task_struct
,sibling
);
3943 static inline struct task_struct
*older_sibling(struct task_struct
*p
)
3945 if (p
->sibling
.prev
==&p
->parent
->children
) return NULL
;
3946 return list_entry(p
->sibling
.prev
,struct task_struct
,sibling
);
3949 static inline struct task_struct
*younger_sibling(struct task_struct
*p
)
3951 if (p
->sibling
.next
==&p
->parent
->children
) return NULL
;
3952 return list_entry(p
->sibling
.next
,struct task_struct
,sibling
);
3955 static void show_task(task_t
* p
)
3959 unsigned long free
= 0;
3960 static const char *stat_nam
[] = { "R", "S", "D", "T", "t", "Z", "X" };
3962 printk("%-13.13s ", p
->comm
);
3963 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
3964 if (state
< ARRAY_SIZE(stat_nam
))
3965 printk(stat_nam
[state
]);
3968 #if (BITS_PER_LONG == 32)
3969 if (state
== TASK_RUNNING
)
3970 printk(" running ");
3972 printk(" %08lX ", thread_saved_pc(p
));
3974 if (state
== TASK_RUNNING
)
3975 printk(" running task ");
3977 printk(" %016lx ", thread_saved_pc(p
));
3979 #ifdef CONFIG_DEBUG_STACK_USAGE
3981 unsigned long * n
= (unsigned long *) (p
->thread_info
+1);
3984 free
= (unsigned long) n
- (unsigned long)(p
->thread_info
+1);
3987 printk("%5lu %5d %6d ", free
, p
->pid
, p
->parent
->pid
);
3988 if ((relative
= eldest_child(p
)))
3989 printk("%5d ", relative
->pid
);
3992 if ((relative
= younger_sibling(p
)))
3993 printk("%7d", relative
->pid
);
3996 if ((relative
= older_sibling(p
)))
3997 printk(" %5d", relative
->pid
);
4001 printk(" (L-TLB)\n");
4003 printk(" (NOTLB)\n");
4005 if (state
!= TASK_RUNNING
)
4006 show_stack(p
, NULL
);
4009 void show_state(void)
4013 #if (BITS_PER_LONG == 32)
4016 printk(" task PC pid father child younger older\n");
4020 printk(" task PC pid father child younger older\n");
4022 read_lock(&tasklist_lock
);
4023 do_each_thread(g
, p
) {
4025 * reset the NMI-timeout, listing all files on a slow
4026 * console might take alot of time:
4028 touch_nmi_watchdog();
4030 } while_each_thread(g
, p
);
4032 read_unlock(&tasklist_lock
);
4035 void __devinit
init_idle(task_t
*idle
, int cpu
)
4037 runqueue_t
*rq
= cpu_rq(cpu
);
4038 unsigned long flags
;
4040 idle
->sleep_avg
= 0;
4042 idle
->prio
= MAX_PRIO
;
4043 idle
->state
= TASK_RUNNING
;
4044 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4045 set_task_cpu(idle
, cpu
);
4047 spin_lock_irqsave(&rq
->lock
, flags
);
4048 rq
->curr
= rq
->idle
= idle
;
4049 set_tsk_need_resched(idle
);
4050 spin_unlock_irqrestore(&rq
->lock
, flags
);
4052 /* Set the preempt count _outside_ the spinlocks! */
4053 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4054 idle
->thread_info
->preempt_count
= (idle
->lock_depth
>= 0);
4056 idle
->thread_info
->preempt_count
= 0;
4061 * In a system that switches off the HZ timer nohz_cpu_mask
4062 * indicates which cpus entered this state. This is used
4063 * in the rcu update to wait only for active cpus. For system
4064 * which do not switch off the HZ timer nohz_cpu_mask should
4065 * always be CPU_MASK_NONE.
4067 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4071 * This is how migration works:
4073 * 1) we queue a migration_req_t structure in the source CPU's
4074 * runqueue and wake up that CPU's migration thread.
4075 * 2) we down() the locked semaphore => thread blocks.
4076 * 3) migration thread wakes up (implicitly it forces the migrated
4077 * thread off the CPU)
4078 * 4) it gets the migration request and checks whether the migrated
4079 * task is still in the wrong runqueue.
4080 * 5) if it's in the wrong runqueue then the migration thread removes
4081 * it and puts it into the right queue.
4082 * 6) migration thread up()s the semaphore.
4083 * 7) we wake up and the migration is done.
4087 * Change a given task's CPU affinity. Migrate the thread to a
4088 * proper CPU and schedule it away if the CPU it's executing on
4089 * is removed from the allowed bitmask.
4091 * NOTE: the caller must have a valid reference to the task, the
4092 * task must not exit() & deallocate itself prematurely. The
4093 * call is not atomic; no spinlocks may be held.
4095 int set_cpus_allowed(task_t
*p
, cpumask_t new_mask
)
4097 unsigned long flags
;
4099 migration_req_t req
;
4102 rq
= task_rq_lock(p
, &flags
);
4103 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4108 p
->cpus_allowed
= new_mask
;
4109 /* Can the task run on the task's current CPU? If so, we're done */
4110 if (cpu_isset(task_cpu(p
), new_mask
))
4113 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4114 /* Need help from migration thread: drop lock and wait. */
4115 task_rq_unlock(rq
, &flags
);
4116 wake_up_process(rq
->migration_thread
);
4117 wait_for_completion(&req
.done
);
4118 tlb_migrate_finish(p
->mm
);
4122 task_rq_unlock(rq
, &flags
);
4126 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4129 * Move (not current) task off this cpu, onto dest cpu. We're doing
4130 * this because either it can't run here any more (set_cpus_allowed()
4131 * away from this CPU, or CPU going down), or because we're
4132 * attempting to rebalance this task on exec (sched_exec).
4134 * So we race with normal scheduler movements, but that's OK, as long
4135 * as the task is no longer on this CPU.
4137 static void __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4139 runqueue_t
*rq_dest
, *rq_src
;
4141 if (unlikely(cpu_is_offline(dest_cpu
)))
4144 rq_src
= cpu_rq(src_cpu
);
4145 rq_dest
= cpu_rq(dest_cpu
);
4147 double_rq_lock(rq_src
, rq_dest
);
4148 /* Already moved. */
4149 if (task_cpu(p
) != src_cpu
)
4151 /* Affinity changed (again). */
4152 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
4155 set_task_cpu(p
, dest_cpu
);
4158 * Sync timestamp with rq_dest's before activating.
4159 * The same thing could be achieved by doing this step
4160 * afterwards, and pretending it was a local activate.
4161 * This way is cleaner and logically correct.
4163 p
->timestamp
= p
->timestamp
- rq_src
->timestamp_last_tick
4164 + rq_dest
->timestamp_last_tick
;
4165 deactivate_task(p
, rq_src
);
4166 activate_task(p
, rq_dest
, 0);
4167 if (TASK_PREEMPTS_CURR(p
, rq_dest
))
4168 resched_task(rq_dest
->curr
);
4172 double_rq_unlock(rq_src
, rq_dest
);
4176 * migration_thread - this is a highprio system thread that performs
4177 * thread migration by bumping thread off CPU then 'pushing' onto
4180 static int migration_thread(void * data
)
4183 int cpu
= (long)data
;
4186 BUG_ON(rq
->migration_thread
!= current
);
4188 set_current_state(TASK_INTERRUPTIBLE
);
4189 while (!kthread_should_stop()) {
4190 struct list_head
*head
;
4191 migration_req_t
*req
;
4193 if (current
->flags
& PF_FREEZE
)
4194 refrigerator(PF_FREEZE
);
4196 spin_lock_irq(&rq
->lock
);
4198 if (cpu_is_offline(cpu
)) {
4199 spin_unlock_irq(&rq
->lock
);
4203 if (rq
->active_balance
) {
4204 active_load_balance(rq
, cpu
);
4205 rq
->active_balance
= 0;
4208 head
= &rq
->migration_queue
;
4210 if (list_empty(head
)) {
4211 spin_unlock_irq(&rq
->lock
);
4213 set_current_state(TASK_INTERRUPTIBLE
);
4216 req
= list_entry(head
->next
, migration_req_t
, list
);
4217 list_del_init(head
->next
);
4219 if (req
->type
== REQ_MOVE_TASK
) {
4220 spin_unlock(&rq
->lock
);
4221 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
4223 } else if (req
->type
== REQ_SET_DOMAIN
) {
4225 spin_unlock_irq(&rq
->lock
);
4227 spin_unlock_irq(&rq
->lock
);
4231 complete(&req
->done
);
4233 __set_current_state(TASK_RUNNING
);
4237 /* Wait for kthread_stop */
4238 set_current_state(TASK_INTERRUPTIBLE
);
4239 while (!kthread_should_stop()) {
4241 set_current_state(TASK_INTERRUPTIBLE
);
4243 __set_current_state(TASK_RUNNING
);
4247 #ifdef CONFIG_HOTPLUG_CPU
4248 /* Figure out where task on dead CPU should go, use force if neccessary. */
4249 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*tsk
)
4255 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
4256 cpus_and(mask
, mask
, tsk
->cpus_allowed
);
4257 dest_cpu
= any_online_cpu(mask
);
4259 /* On any allowed CPU? */
4260 if (dest_cpu
== NR_CPUS
)
4261 dest_cpu
= any_online_cpu(tsk
->cpus_allowed
);
4263 /* No more Mr. Nice Guy. */
4264 if (dest_cpu
== NR_CPUS
) {
4265 cpus_setall(tsk
->cpus_allowed
);
4266 dest_cpu
= any_online_cpu(tsk
->cpus_allowed
);
4269 * Don't tell them about moving exiting tasks or
4270 * kernel threads (both mm NULL), since they never
4273 if (tsk
->mm
&& printk_ratelimit())
4274 printk(KERN_INFO
"process %d (%s) no "
4275 "longer affine to cpu%d\n",
4276 tsk
->pid
, tsk
->comm
, dead_cpu
);
4278 __migrate_task(tsk
, dead_cpu
, dest_cpu
);
4282 * While a dead CPU has no uninterruptible tasks queued at this point,
4283 * it might still have a nonzero ->nr_uninterruptible counter, because
4284 * for performance reasons the counter is not stricly tracking tasks to
4285 * their home CPUs. So we just add the counter to another CPU's counter,
4286 * to keep the global sum constant after CPU-down:
4288 static void migrate_nr_uninterruptible(runqueue_t
*rq_src
)
4290 runqueue_t
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
4291 unsigned long flags
;
4293 local_irq_save(flags
);
4294 double_rq_lock(rq_src
, rq_dest
);
4295 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
4296 rq_src
->nr_uninterruptible
= 0;
4297 double_rq_unlock(rq_src
, rq_dest
);
4298 local_irq_restore(flags
);
4301 /* Run through task list and migrate tasks from the dead cpu. */
4302 static void migrate_live_tasks(int src_cpu
)
4304 struct task_struct
*tsk
, *t
;
4306 write_lock_irq(&tasklist_lock
);
4308 do_each_thread(t
, tsk
) {
4312 if (task_cpu(tsk
) == src_cpu
)
4313 move_task_off_dead_cpu(src_cpu
, tsk
);
4314 } while_each_thread(t
, tsk
);
4316 write_unlock_irq(&tasklist_lock
);
4319 /* Schedules idle task to be the next runnable task on current CPU.
4320 * It does so by boosting its priority to highest possible and adding it to
4321 * the _front_ of runqueue. Used by CPU offline code.
4323 void sched_idle_next(void)
4325 int cpu
= smp_processor_id();
4326 runqueue_t
*rq
= this_rq();
4327 struct task_struct
*p
= rq
->idle
;
4328 unsigned long flags
;
4330 /* cpu has to be offline */
4331 BUG_ON(cpu_online(cpu
));
4333 /* Strictly not necessary since rest of the CPUs are stopped by now
4334 * and interrupts disabled on current cpu.
4336 spin_lock_irqsave(&rq
->lock
, flags
);
4338 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
4339 /* Add idle task to _front_ of it's priority queue */
4340 __activate_idle_task(p
, rq
);
4342 spin_unlock_irqrestore(&rq
->lock
, flags
);
4345 /* Ensures that the idle task is using init_mm right before its cpu goes
4348 void idle_task_exit(void)
4350 struct mm_struct
*mm
= current
->active_mm
;
4352 BUG_ON(cpu_online(smp_processor_id()));
4355 switch_mm(mm
, &init_mm
, current
);
4359 static void migrate_dead(unsigned int dead_cpu
, task_t
*tsk
)
4361 struct runqueue
*rq
= cpu_rq(dead_cpu
);
4363 /* Must be exiting, otherwise would be on tasklist. */
4364 BUG_ON(tsk
->exit_state
!= EXIT_ZOMBIE
&& tsk
->exit_state
!= EXIT_DEAD
);
4366 /* Cannot have done final schedule yet: would have vanished. */
4367 BUG_ON(tsk
->flags
& PF_DEAD
);
4369 get_task_struct(tsk
);
4372 * Drop lock around migration; if someone else moves it,
4373 * that's OK. No task can be added to this CPU, so iteration is
4376 spin_unlock_irq(&rq
->lock
);
4377 move_task_off_dead_cpu(dead_cpu
, tsk
);
4378 spin_lock_irq(&rq
->lock
);
4380 put_task_struct(tsk
);
4383 /* release_task() removes task from tasklist, so we won't find dead tasks. */
4384 static void migrate_dead_tasks(unsigned int dead_cpu
)
4387 struct runqueue
*rq
= cpu_rq(dead_cpu
);
4389 for (arr
= 0; arr
< 2; arr
++) {
4390 for (i
= 0; i
< MAX_PRIO
; i
++) {
4391 struct list_head
*list
= &rq
->arrays
[arr
].queue
[i
];
4392 while (!list_empty(list
))
4393 migrate_dead(dead_cpu
,
4394 list_entry(list
->next
, task_t
,
4399 #endif /* CONFIG_HOTPLUG_CPU */
4402 * migration_call - callback that gets triggered when a CPU is added.
4403 * Here we can start up the necessary migration thread for the new CPU.
4405 static int migration_call(struct notifier_block
*nfb
, unsigned long action
,
4408 int cpu
= (long)hcpu
;
4409 struct task_struct
*p
;
4410 struct runqueue
*rq
;
4411 unsigned long flags
;
4414 case CPU_UP_PREPARE
:
4415 p
= kthread_create(migration_thread
, hcpu
, "migration/%d",cpu
);
4418 p
->flags
|= PF_NOFREEZE
;
4419 kthread_bind(p
, cpu
);
4420 /* Must be high prio: stop_machine expects to yield to it. */
4421 rq
= task_rq_lock(p
, &flags
);
4422 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
4423 task_rq_unlock(rq
, &flags
);
4424 cpu_rq(cpu
)->migration_thread
= p
;
4427 /* Strictly unneccessary, as first user will wake it. */
4428 wake_up_process(cpu_rq(cpu
)->migration_thread
);
4430 #ifdef CONFIG_HOTPLUG_CPU
4431 case CPU_UP_CANCELED
:
4432 /* Unbind it from offline cpu so it can run. Fall thru. */
4433 kthread_bind(cpu_rq(cpu
)->migration_thread
,smp_processor_id());
4434 kthread_stop(cpu_rq(cpu
)->migration_thread
);
4435 cpu_rq(cpu
)->migration_thread
= NULL
;
4438 migrate_live_tasks(cpu
);
4440 kthread_stop(rq
->migration_thread
);
4441 rq
->migration_thread
= NULL
;
4442 /* Idle task back to normal (off runqueue, low prio) */
4443 rq
= task_rq_lock(rq
->idle
, &flags
);
4444 deactivate_task(rq
->idle
, rq
);
4445 rq
->idle
->static_prio
= MAX_PRIO
;
4446 __setscheduler(rq
->idle
, SCHED_NORMAL
, 0);
4447 migrate_dead_tasks(cpu
);
4448 task_rq_unlock(rq
, &flags
);
4449 migrate_nr_uninterruptible(rq
);
4450 BUG_ON(rq
->nr_running
!= 0);
4452 /* No need to migrate the tasks: it was best-effort if
4453 * they didn't do lock_cpu_hotplug(). Just wake up
4454 * the requestors. */
4455 spin_lock_irq(&rq
->lock
);
4456 while (!list_empty(&rq
->migration_queue
)) {
4457 migration_req_t
*req
;
4458 req
= list_entry(rq
->migration_queue
.next
,
4459 migration_req_t
, list
);
4460 BUG_ON(req
->type
!= REQ_MOVE_TASK
);
4461 list_del_init(&req
->list
);
4462 complete(&req
->done
);
4464 spin_unlock_irq(&rq
->lock
);
4471 /* Register at highest priority so that task migration (migrate_all_tasks)
4472 * happens before everything else.
4474 static struct notifier_block __devinitdata migration_notifier
= {
4475 .notifier_call
= migration_call
,
4479 int __init
migration_init(void)
4481 void *cpu
= (void *)(long)smp_processor_id();
4482 /* Start one for boot CPU. */
4483 migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
4484 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
4485 register_cpu_notifier(&migration_notifier
);
4491 #define SCHED_DOMAIN_DEBUG
4492 #ifdef SCHED_DOMAIN_DEBUG
4493 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
4497 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
4502 struct sched_group
*group
= sd
->groups
;
4503 cpumask_t groupmask
;
4505 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
4506 cpus_clear(groupmask
);
4509 for (i
= 0; i
< level
+ 1; i
++)
4511 printk("domain %d: ", level
);
4513 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
4514 printk("does not load-balance\n");
4516 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain has parent");
4520 printk("span %s\n", str
);
4522 if (!cpu_isset(cpu
, sd
->span
))
4523 printk(KERN_ERR
"ERROR: domain->span does not contain CPU%d\n", cpu
);
4524 if (!cpu_isset(cpu
, group
->cpumask
))
4525 printk(KERN_ERR
"ERROR: domain->groups does not contain CPU%d\n", cpu
);
4528 for (i
= 0; i
< level
+ 2; i
++)
4534 printk(KERN_ERR
"ERROR: group is NULL\n");
4538 if (!group
->cpu_power
) {
4540 printk(KERN_ERR
"ERROR: domain->cpu_power not set\n");
4543 if (!cpus_weight(group
->cpumask
)) {
4545 printk(KERN_ERR
"ERROR: empty group\n");
4548 if (cpus_intersects(groupmask
, group
->cpumask
)) {
4550 printk(KERN_ERR
"ERROR: repeated CPUs\n");
4553 cpus_or(groupmask
, groupmask
, group
->cpumask
);
4555 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
4558 group
= group
->next
;
4559 } while (group
!= sd
->groups
);
4562 if (!cpus_equal(sd
->span
, groupmask
))
4563 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
4569 if (!cpus_subset(groupmask
, sd
->span
))
4570 printk(KERN_ERR
"ERROR: parent span is not a superset of domain->span\n");
4576 #define sched_domain_debug(sd, cpu) {}
4580 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
4581 * hold the hotplug lock.
4583 void __devinit
cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
4585 migration_req_t req
;
4586 unsigned long flags
;
4587 runqueue_t
*rq
= cpu_rq(cpu
);
4590 sched_domain_debug(sd
, cpu
);
4592 spin_lock_irqsave(&rq
->lock
, flags
);
4594 if (cpu
== smp_processor_id() || !cpu_online(cpu
)) {
4597 init_completion(&req
.done
);
4598 req
.type
= REQ_SET_DOMAIN
;
4600 list_add(&req
.list
, &rq
->migration_queue
);
4604 spin_unlock_irqrestore(&rq
->lock
, flags
);
4607 wake_up_process(rq
->migration_thread
);
4608 wait_for_completion(&req
.done
);
4612 /* cpus with isolated domains */
4613 cpumask_t __devinitdata cpu_isolated_map
= CPU_MASK_NONE
;
4615 /* Setup the mask of cpus configured for isolated domains */
4616 static int __init
isolated_cpu_setup(char *str
)
4618 int ints
[NR_CPUS
], i
;
4620 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
4621 cpus_clear(cpu_isolated_map
);
4622 for (i
= 1; i
<= ints
[0]; i
++)
4623 if (ints
[i
] < NR_CPUS
)
4624 cpu_set(ints
[i
], cpu_isolated_map
);
4628 __setup ("isolcpus=", isolated_cpu_setup
);
4631 * init_sched_build_groups takes an array of groups, the cpumask we wish
4632 * to span, and a pointer to a function which identifies what group a CPU
4633 * belongs to. The return value of group_fn must be a valid index into the
4634 * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
4635 * keep track of groups covered with a cpumask_t).
4637 * init_sched_build_groups will build a circular linked list of the groups
4638 * covered by the given span, and will set each group's ->cpumask correctly,
4639 * and ->cpu_power to 0.
4641 void __devinit
init_sched_build_groups(struct sched_group groups
[],
4642 cpumask_t span
, int (*group_fn
)(int cpu
))
4644 struct sched_group
*first
= NULL
, *last
= NULL
;
4645 cpumask_t covered
= CPU_MASK_NONE
;
4648 for_each_cpu_mask(i
, span
) {
4649 int group
= group_fn(i
);
4650 struct sched_group
*sg
= &groups
[group
];
4653 if (cpu_isset(i
, covered
))
4656 sg
->cpumask
= CPU_MASK_NONE
;
4659 for_each_cpu_mask(j
, span
) {
4660 if (group_fn(j
) != group
)
4663 cpu_set(j
, covered
);
4664 cpu_set(j
, sg
->cpumask
);
4676 #ifdef ARCH_HAS_SCHED_DOMAIN
4677 extern void __devinit
arch_init_sched_domains(void);
4678 extern void __devinit
arch_destroy_sched_domains(void);
4680 #ifdef CONFIG_SCHED_SMT
4681 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
4682 static struct sched_group sched_group_cpus
[NR_CPUS
];
4683 static int __devinit
cpu_to_cpu_group(int cpu
)
4689 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
4690 static struct sched_group sched_group_phys
[NR_CPUS
];
4691 static int __devinit
cpu_to_phys_group(int cpu
)
4693 #ifdef CONFIG_SCHED_SMT
4694 return first_cpu(cpu_sibling_map
[cpu
]);
4702 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
4703 static struct sched_group sched_group_nodes
[MAX_NUMNODES
];
4704 static int __devinit
cpu_to_node_group(int cpu
)
4706 return cpu_to_node(cpu
);
4710 #if defined(CONFIG_SCHED_SMT) && defined(CONFIG_NUMA)
4712 * The domains setup code relies on siblings not spanning
4713 * multiple nodes. Make sure the architecture has a proper
4716 static void check_sibling_maps(void)
4720 for_each_online_cpu(i
) {
4721 for_each_cpu_mask(j
, cpu_sibling_map
[i
]) {
4722 if (cpu_to_node(i
) != cpu_to_node(j
)) {
4723 printk(KERN_INFO
"warning: CPU %d siblings map "
4724 "to different node - isolating "
4726 cpu_sibling_map
[i
] = cpumask_of_cpu(i
);
4735 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
4737 static void __devinit
arch_init_sched_domains(void)
4740 cpumask_t cpu_default_map
;
4742 #if defined(CONFIG_SCHED_SMT) && defined(CONFIG_NUMA)
4743 check_sibling_maps();
4746 * Setup mask for cpus without special case scheduling requirements.
4747 * For now this just excludes isolated cpus, but could be used to
4748 * exclude other special cases in the future.
4750 cpus_complement(cpu_default_map
, cpu_isolated_map
);
4751 cpus_and(cpu_default_map
, cpu_default_map
, cpu_online_map
);
4754 * Set up domains. Isolated domains just stay on the dummy domain.
4756 for_each_cpu_mask(i
, cpu_default_map
) {
4758 struct sched_domain
*sd
= NULL
, *p
;
4759 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
4761 cpus_and(nodemask
, nodemask
, cpu_default_map
);
4764 sd
= &per_cpu(node_domains
, i
);
4765 group
= cpu_to_node_group(i
);
4767 sd
->span
= cpu_default_map
;
4768 sd
->groups
= &sched_group_nodes
[group
];
4772 sd
= &per_cpu(phys_domains
, i
);
4773 group
= cpu_to_phys_group(i
);
4775 sd
->span
= nodemask
;
4777 sd
->groups
= &sched_group_phys
[group
];
4779 #ifdef CONFIG_SCHED_SMT
4781 sd
= &per_cpu(cpu_domains
, i
);
4782 group
= cpu_to_cpu_group(i
);
4783 *sd
= SD_SIBLING_INIT
;
4784 sd
->span
= cpu_sibling_map
[i
];
4785 cpus_and(sd
->span
, sd
->span
, cpu_default_map
);
4787 sd
->groups
= &sched_group_cpus
[group
];
4791 #ifdef CONFIG_SCHED_SMT
4792 /* Set up CPU (sibling) groups */
4793 for_each_online_cpu(i
) {
4794 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
4795 cpus_and(this_sibling_map
, this_sibling_map
, cpu_default_map
);
4796 if (i
!= first_cpu(this_sibling_map
))
4799 init_sched_build_groups(sched_group_cpus
, this_sibling_map
,
4804 /* Set up physical groups */
4805 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
4806 cpumask_t nodemask
= node_to_cpumask(i
);
4808 cpus_and(nodemask
, nodemask
, cpu_default_map
);
4809 if (cpus_empty(nodemask
))
4812 init_sched_build_groups(sched_group_phys
, nodemask
,
4813 &cpu_to_phys_group
);
4817 /* Set up node groups */
4818 init_sched_build_groups(sched_group_nodes
, cpu_default_map
,
4819 &cpu_to_node_group
);
4822 /* Calculate CPU power for physical packages and nodes */
4823 for_each_cpu_mask(i
, cpu_default_map
) {
4825 struct sched_domain
*sd
;
4826 #ifdef CONFIG_SCHED_SMT
4827 sd
= &per_cpu(cpu_domains
, i
);
4828 power
= SCHED_LOAD_SCALE
;
4829 sd
->groups
->cpu_power
= power
;
4832 sd
= &per_cpu(phys_domains
, i
);
4833 power
= SCHED_LOAD_SCALE
+ SCHED_LOAD_SCALE
*
4834 (cpus_weight(sd
->groups
->cpumask
)-1) / 10;
4835 sd
->groups
->cpu_power
= power
;
4838 if (i
== first_cpu(sd
->groups
->cpumask
)) {
4839 /* Only add "power" once for each physical package. */
4840 sd
= &per_cpu(node_domains
, i
);
4841 sd
->groups
->cpu_power
+= power
;
4846 /* Attach the domains */
4847 for_each_online_cpu(i
) {
4848 struct sched_domain
*sd
;
4849 #ifdef CONFIG_SCHED_SMT
4850 sd
= &per_cpu(cpu_domains
, i
);
4852 sd
= &per_cpu(phys_domains
, i
);
4854 cpu_attach_domain(sd
, i
);
4858 #ifdef CONFIG_HOTPLUG_CPU
4859 static void __devinit
arch_destroy_sched_domains(void)
4861 /* Do nothing: everything is statically allocated. */
4865 #endif /* ARCH_HAS_SCHED_DOMAIN */
4868 * Initial dummy domain for early boot and for hotplug cpu. Being static,
4869 * it is initialized to zero, so all balancing flags are cleared which is
4872 static struct sched_domain sched_domain_dummy
;
4874 #ifdef CONFIG_HOTPLUG_CPU
4876 * Force a reinitialization of the sched domains hierarchy. The domains
4877 * and groups cannot be updated in place without racing with the balancing
4878 * code, so we temporarily attach all running cpus to a "dummy" domain
4879 * which will prevent rebalancing while the sched domains are recalculated.
4881 static int update_sched_domains(struct notifier_block
*nfb
,
4882 unsigned long action
, void *hcpu
)
4887 case CPU_UP_PREPARE
:
4888 case CPU_DOWN_PREPARE
:
4889 for_each_online_cpu(i
)
4890 cpu_attach_domain(&sched_domain_dummy
, i
);
4891 arch_destroy_sched_domains();
4894 case CPU_UP_CANCELED
:
4895 case CPU_DOWN_FAILED
:
4899 * Fall through and re-initialise the domains.
4906 /* The hotplug lock is already held by cpu_up/cpu_down */
4907 arch_init_sched_domains();
4913 void __init
sched_init_smp(void)
4916 arch_init_sched_domains();
4917 unlock_cpu_hotplug();
4918 /* XXX: Theoretical race here - CPU may be hotplugged now */
4919 hotcpu_notifier(update_sched_domains
, 0);
4922 void __init
sched_init_smp(void)
4925 #endif /* CONFIG_SMP */
4927 int in_sched_functions(unsigned long addr
)
4929 /* Linker adds these: start and end of __sched functions */
4930 extern char __sched_text_start
[], __sched_text_end
[];
4931 return in_lock_functions(addr
) ||
4932 (addr
>= (unsigned long)__sched_text_start
4933 && addr
< (unsigned long)__sched_text_end
);
4936 void __init
sched_init(void)
4941 for (i
= 0; i
< NR_CPUS
; i
++) {
4942 prio_array_t
*array
;
4945 spin_lock_init(&rq
->lock
);
4946 rq
->active
= rq
->arrays
;
4947 rq
->expired
= rq
->arrays
+ 1;
4948 rq
->best_expired_prio
= MAX_PRIO
;
4951 rq
->sd
= &sched_domain_dummy
;
4953 rq
->active_balance
= 0;
4955 rq
->migration_thread
= NULL
;
4956 INIT_LIST_HEAD(&rq
->migration_queue
);
4958 atomic_set(&rq
->nr_iowait
, 0);
4960 for (j
= 0; j
< 2; j
++) {
4961 array
= rq
->arrays
+ j
;
4962 for (k
= 0; k
< MAX_PRIO
; k
++) {
4963 INIT_LIST_HEAD(array
->queue
+ k
);
4964 __clear_bit(k
, array
->bitmap
);
4966 // delimiter for bitsearch
4967 __set_bit(MAX_PRIO
, array
->bitmap
);
4972 * The boot idle thread does lazy MMU switching as well:
4974 atomic_inc(&init_mm
.mm_count
);
4975 enter_lazy_tlb(&init_mm
, current
);
4978 * Make us the idle thread. Technically, schedule() should not be
4979 * called from this thread, however somewhere below it might be,
4980 * but because we are the idle thread, we just pick up running again
4981 * when this runqueue becomes "idle".
4983 init_idle(current
, smp_processor_id());
4986 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4987 void __might_sleep(char *file
, int line
)
4989 #if defined(in_atomic)
4990 static unsigned long prev_jiffy
; /* ratelimiting */
4992 if ((in_atomic() || irqs_disabled()) &&
4993 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
4994 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
4996 prev_jiffy
= jiffies
;
4997 printk(KERN_ERR
"Debug: sleeping function called from invalid"
4998 " context at %s:%d\n", file
, line
);
4999 printk("in_atomic():%d, irqs_disabled():%d\n",
5000 in_atomic(), irqs_disabled());
5005 EXPORT_SYMBOL(__might_sleep
);
5008 #ifdef CONFIG_MAGIC_SYSRQ
5009 void normalize_rt_tasks(void)
5011 struct task_struct
*p
;
5012 prio_array_t
*array
;
5013 unsigned long flags
;
5016 read_lock_irq(&tasklist_lock
);
5017 for_each_process (p
) {
5021 rq
= task_rq_lock(p
, &flags
);
5025 deactivate_task(p
, task_rq(p
));
5026 __setscheduler(p
, SCHED_NORMAL
, 0);
5028 __activate_task(p
, task_rq(p
));
5029 resched_task(rq
->curr
);
5032 task_rq_unlock(rq
, &flags
);
5034 read_unlock_irq(&tasklist_lock
);
5037 #endif /* CONFIG_MAGIC_SYSRQ */