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 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[3];
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);
264 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
265 * See detach_destroy_domains: synchronize_sched for details.
267 * The domain tree of any CPU may only be accessed from within
268 * preempt-disabled sections.
270 #define for_each_domain(cpu, domain) \
271 for (domain = rcu_dereference(cpu_rq(cpu)->sd); domain; domain = domain->parent)
273 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
274 #define this_rq() (&__get_cpu_var(runqueues))
275 #define task_rq(p) cpu_rq(task_cpu(p))
276 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
278 #ifndef prepare_arch_switch
279 # define prepare_arch_switch(next) do { } while (0)
281 #ifndef finish_arch_switch
282 # define finish_arch_switch(prev) do { } while (0)
285 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
286 static inline int task_running(runqueue_t *rq, task_t *p)
288 return rq->curr == p;
291 static inline void prepare_lock_switch(runqueue_t *rq, task_t *next)
295 static inline void finish_lock_switch(runqueue_t *rq, task_t *prev)
297 spin_unlock_irq(&rq->lock);
300 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
301 static inline int task_running(runqueue_t *rq, task_t *p)
306 return rq->curr == p;
310 static inline void prepare_lock_switch(runqueue_t *rq, task_t *next)
314 * We can optimise this out completely for !SMP, because the
315 * SMP rebalancing from interrupt is the only thing that cares
320 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
321 spin_unlock_irq(&rq->lock);
323 spin_unlock(&rq->lock);
327 static inline void finish_lock_switch(runqueue_t *rq, task_t *prev)
331 * After ->oncpu is cleared, the task can be moved to a different CPU.
332 * We must ensure this doesn't happen until the switch is completely
338 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
342 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
345 * task_rq_lock - lock the runqueue a given task resides on and disable
346 * interrupts. Note the ordering: we can safely lookup the task_rq without
347 * explicitly disabling preemption.
349 static inline runqueue_t *task_rq_lock(task_t *p, unsigned long *flags)
355 local_irq_save(*flags);
357 spin_lock(&rq->lock);
358 if (unlikely(rq != task_rq(p))) {
359 spin_unlock_irqrestore(&rq->lock, *flags);
360 goto repeat_lock_task;
365 static inline void task_rq_unlock(runqueue_t *rq, unsigned long *flags)
368 spin_unlock_irqrestore(&rq->lock, *flags);
371 #ifdef CONFIG_SCHEDSTATS
373 * bump this up when changing the output format or the meaning of an existing
374 * format, so that tools can adapt (or abort)
376 #define SCHEDSTAT_VERSION 12
378 static int show_schedstat(struct seq_file *seq, void *v)
382 seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION);
383 seq_printf(seq, "timestamp %lu\n", jiffies);
384 for_each_online_cpu(cpu) {
385 runqueue_t *rq = cpu_rq(cpu);
387 struct sched_domain *sd;
391 /* runqueue-specific stats */
393 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
394 cpu, rq->yld_both_empty,
395 rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt,
396 rq->sched_switch, rq->sched_cnt, rq->sched_goidle,
397 rq->ttwu_cnt, rq->ttwu_local,
398 rq->rq_sched_info.cpu_time,
399 rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt);
401 seq_printf(seq, "\n");
404 /* domain-specific stats */
406 for_each_domain(cpu, sd) {
407 enum idle_type itype;
408 char mask_str[NR_CPUS];
410 cpumask_scnprintf(mask_str, NR_CPUS, sd->span);
411 seq_printf(seq, "domain%d %s", dcnt++, mask_str);
412 for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES;
414 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu",
416 sd->lb_balanced[itype],
417 sd->lb_failed[itype],
418 sd->lb_imbalance[itype],
419 sd->lb_gained[itype],
420 sd->lb_hot_gained[itype],
421 sd->lb_nobusyq[itype],
422 sd->lb_nobusyg[itype]);
424 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
425 sd->alb_cnt, sd->alb_failed, sd->alb_pushed,
426 sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed,
427 sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed,
428 sd->ttwu_wake_remote, sd->ttwu_move_affine, sd->ttwu_move_balance);
436 static int schedstat_open(struct inode *inode, struct file *file)
438 unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32);
439 char *buf = kmalloc(size, GFP_KERNEL);
445 res = single_open(file, show_schedstat, NULL);
447 m = file->private_data;
455 struct file_operations proc_schedstat_operations = {
456 .open = schedstat_open,
459 .release = single_release,
462 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
463 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
464 #else /* !CONFIG_SCHEDSTATS */
465 # define schedstat_inc(rq, field) do { } while (0)
466 # define schedstat_add(rq, field, amt) do { } while (0)
470 * rq_lock - lock a given runqueue and disable interrupts.
472 static inline runqueue_t *this_rq_lock(void)
479 spin_lock(&rq->lock);
484 #ifdef CONFIG_SCHEDSTATS
486 * Called when a process is dequeued from the active array and given
487 * the cpu. We should note that with the exception of interactive
488 * tasks, the expired queue will become the active queue after the active
489 * queue is empty, without explicitly dequeuing and requeuing tasks in the
490 * expired queue. (Interactive tasks may be requeued directly to the
491 * active queue, thus delaying tasks in the expired queue from running;
492 * see scheduler_tick()).
494 * This function is only called from sched_info_arrive(), rather than
495 * dequeue_task(). Even though a task may be queued and dequeued multiple
496 * times as it is shuffled about, we're really interested in knowing how
497 * long it was from the *first* time it was queued to the time that it
500 static inline void sched_info_dequeued(task_t *t)
502 t->sched_info.last_queued = 0;
506 * Called when a task finally hits the cpu. We can now calculate how
507 * long it was waiting to run. We also note when it began so that we
508 * can keep stats on how long its timeslice is.
510 static inline void sched_info_arrive(task_t *t)
512 unsigned long now = jiffies, diff = 0;
513 struct runqueue *rq = task_rq(t);
515 if (t->sched_info.last_queued)
516 diff = now - t->sched_info.last_queued;
517 sched_info_dequeued(t);
518 t->sched_info.run_delay += diff;
519 t->sched_info.last_arrival = now;
520 t->sched_info.pcnt++;
525 rq->rq_sched_info.run_delay += diff;
526 rq->rq_sched_info.pcnt++;
530 * Called when a process is queued into either the active or expired
531 * array. The time is noted and later used to determine how long we
532 * had to wait for us to reach the cpu. Since the expired queue will
533 * become the active queue after active queue is empty, without dequeuing
534 * and requeuing any tasks, we are interested in queuing to either. It
535 * is unusual but not impossible for tasks to be dequeued and immediately
536 * requeued in the same or another array: this can happen in sched_yield(),
537 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
540 * This function is only called from enqueue_task(), but also only updates
541 * the timestamp if it is already not set. It's assumed that
542 * sched_info_dequeued() will clear that stamp when appropriate.
544 static inline void sched_info_queued(task_t *t)
546 if (!t->sched_info.last_queued)
547 t->sched_info.last_queued = jiffies;
551 * Called when a process ceases being the active-running process, either
552 * voluntarily or involuntarily. Now we can calculate how long we ran.
554 static inline void sched_info_depart(task_t *t)
556 struct runqueue *rq = task_rq(t);
557 unsigned long diff = jiffies - t->sched_info.last_arrival;
559 t->sched_info.cpu_time += diff;
562 rq->rq_sched_info.cpu_time += diff;
566 * Called when tasks are switched involuntarily due, typically, to expiring
567 * their time slice. (This may also be called when switching to or from
568 * the idle task.) We are only called when prev != next.
570 static inline void sched_info_switch(task_t *prev, task_t *next)
572 struct runqueue *rq = task_rq(prev);
575 * prev now departs the cpu. It's not interesting to record
576 * stats about how efficient we were at scheduling the idle
579 if (prev != rq->idle)
580 sched_info_depart(prev);
582 if (next != rq->idle)
583 sched_info_arrive(next);
586 #define sched_info_queued(t) do { } while (0)
587 #define sched_info_switch(t, next) do { } while (0)
588 #endif /* CONFIG_SCHEDSTATS */
591 * Adding/removing a task to/from a priority array:
593 static void dequeue_task(struct task_struct *p, prio_array_t *array)
596 list_del(&p->run_list);
597 if (list_empty(array->queue + p->prio))
598 __clear_bit(p->prio, array->bitmap);
601 static void enqueue_task(struct task_struct *p, prio_array_t *array)
603 sched_info_queued(p);
604 list_add_tail(&p->run_list, array->queue + p->prio);
605 __set_bit(p->prio, array->bitmap);
611 * Put task to the end of the run list without the overhead of dequeue
612 * followed by enqueue.
614 static void requeue_task(struct task_struct *p, prio_array_t *array)
616 list_move_tail(&p->run_list, array->queue + p->prio);
619 static inline void enqueue_task_head(struct task_struct *p, prio_array_t *array)
621 list_add(&p->run_list, array->queue + p->prio);
622 __set_bit(p->prio, array->bitmap);
628 * effective_prio - return the priority that is based on the static
629 * priority but is modified by bonuses/penalties.
631 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
632 * into the -5 ... 0 ... +5 bonus/penalty range.
634 * We use 25% of the full 0...39 priority range so that:
636 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
637 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
639 * Both properties are important to certain workloads.
641 static int effective_prio(task_t *p)
648 bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
650 prio = p->static_prio - bonus;
651 if (prio < MAX_RT_PRIO)
653 if (prio > MAX_PRIO-1)
659 * __activate_task - move a task to the runqueue.
661 static inline void __activate_task(task_t *p, runqueue_t *rq)
663 enqueue_task(p, rq->active);
668 * __activate_idle_task - move idle task to the _front_ of runqueue.
670 static inline void __activate_idle_task(task_t *p, runqueue_t *rq)
672 enqueue_task_head(p, rq->active);
676 static int recalc_task_prio(task_t *p, unsigned long long now)
678 /* Caller must always ensure 'now >= p->timestamp' */
679 unsigned long long __sleep_time = now - p->timestamp;
680 unsigned long sleep_time;
682 if (__sleep_time > NS_MAX_SLEEP_AVG)
683 sleep_time = NS_MAX_SLEEP_AVG;
685 sleep_time = (unsigned long)__sleep_time;
687 if (likely(sleep_time > 0)) {
689 * User tasks that sleep a long time are categorised as
690 * idle and will get just interactive status to stay active &
691 * prevent them suddenly becoming cpu hogs and starving
694 if (p->mm && p->activated != -1 &&
695 sleep_time > INTERACTIVE_SLEEP(p)) {
696 p->sleep_avg = JIFFIES_TO_NS(MAX_SLEEP_AVG -
700 * The lower the sleep avg a task has the more
701 * rapidly it will rise with sleep time.
703 sleep_time *= (MAX_BONUS - CURRENT_BONUS(p)) ? : 1;
706 * Tasks waking from uninterruptible sleep are
707 * limited in their sleep_avg rise as they
708 * are likely to be waiting on I/O
710 if (p->activated == -1 && p->mm) {
711 if (p->sleep_avg >= INTERACTIVE_SLEEP(p))
713 else if (p->sleep_avg + sleep_time >=
714 INTERACTIVE_SLEEP(p)) {
715 p->sleep_avg = INTERACTIVE_SLEEP(p);
721 * This code gives a bonus to interactive tasks.
723 * The boost works by updating the 'average sleep time'
724 * value here, based on ->timestamp. The more time a
725 * task spends sleeping, the higher the average gets -
726 * and the higher the priority boost gets as well.
728 p->sleep_avg += sleep_time;
730 if (p->sleep_avg > NS_MAX_SLEEP_AVG)
731 p->sleep_avg = NS_MAX_SLEEP_AVG;
735 return effective_prio(p);
739 * activate_task - move a task to the runqueue and do priority recalculation
741 * Update all the scheduling statistics stuff. (sleep average
742 * calculation, priority modifiers, etc.)
744 static void activate_task(task_t *p, runqueue_t *rq, int local)
746 unsigned long long now;
751 /* Compensate for drifting sched_clock */
752 runqueue_t *this_rq = this_rq();
753 now = (now - this_rq->timestamp_last_tick)
754 + rq->timestamp_last_tick;
758 p->prio = recalc_task_prio(p, now);
761 * This checks to make sure it's not an uninterruptible task
762 * that is now waking up.
766 * Tasks which were woken up by interrupts (ie. hw events)
767 * are most likely of interactive nature. So we give them
768 * the credit of extending their sleep time to the period
769 * of time they spend on the runqueue, waiting for execution
770 * on a CPU, first time around:
776 * Normal first-time wakeups get a credit too for
777 * on-runqueue time, but it will be weighted down:
784 __activate_task(p, rq);
788 * deactivate_task - remove a task from the runqueue.
790 static void deactivate_task(struct task_struct *p, runqueue_t *rq)
793 dequeue_task(p, p->array);
798 * resched_task - mark a task 'to be rescheduled now'.
800 * On UP this means the setting of the need_resched flag, on SMP it
801 * might also involve a cross-CPU call to trigger the scheduler on
805 static void resched_task(task_t *p)
807 int need_resched, nrpolling;
809 assert_spin_locked(&task_rq(p)->lock);
811 /* minimise the chance of sending an interrupt to poll_idle() */
812 nrpolling = test_tsk_thread_flag(p,TIF_POLLING_NRFLAG);
813 need_resched = test_and_set_tsk_thread_flag(p,TIF_NEED_RESCHED);
814 nrpolling |= test_tsk_thread_flag(p,TIF_POLLING_NRFLAG);
816 if (!need_resched && !nrpolling && (task_cpu(p) != smp_processor_id()))
817 smp_send_reschedule(task_cpu(p));
820 static inline void resched_task(task_t *p)
822 set_tsk_need_resched(p);
827 * task_curr - is this task currently executing on a CPU?
828 * @p: the task in question.
830 inline int task_curr(const task_t *p)
832 return cpu_curr(task_cpu(p)) == p;
837 struct list_head list;
842 struct completion done;
846 * The task's runqueue lock must be held.
847 * Returns true if you have to wait for migration thread.
849 static int migrate_task(task_t *p, int dest_cpu, migration_req_t *req)
851 runqueue_t *rq = task_rq(p);
854 * If the task is not on a runqueue (and not running), then
855 * it is sufficient to simply update the task's cpu field.
857 if (!p->array && !task_running(rq, p)) {
858 set_task_cpu(p, dest_cpu);
862 init_completion(&req->done);
864 req->dest_cpu = dest_cpu;
865 list_add(&req->list, &rq->migration_queue);
870 * wait_task_inactive - wait for a thread to unschedule.
872 * The caller must ensure that the task *will* unschedule sometime soon,
873 * else this function might spin for a *long* time. This function can't
874 * be called with interrupts off, or it may introduce deadlock with
875 * smp_call_function() if an IPI is sent by the same process we are
876 * waiting to become inactive.
878 void wait_task_inactive(task_t *p)
885 rq = task_rq_lock(p, &flags);
886 /* Must be off runqueue entirely, not preempted. */
887 if (unlikely(p->array || task_running(rq, p))) {
888 /* If it's preempted, we yield. It could be a while. */
889 preempted = !task_running(rq, p);
890 task_rq_unlock(rq, &flags);
896 task_rq_unlock(rq, &flags);
900 * kick_process - kick a running thread to enter/exit the kernel
901 * @p: the to-be-kicked thread
903 * Cause a process which is running on another CPU to enter
904 * kernel-mode, without any delay. (to get signals handled.)
906 * NOTE: this function doesnt have to take the runqueue lock,
907 * because all it wants to ensure is that the remote task enters
908 * the kernel. If the IPI races and the task has been migrated
909 * to another CPU then no harm is done and the purpose has been
912 void kick_process(task_t *p)
918 if ((cpu != smp_processor_id()) && task_curr(p))
919 smp_send_reschedule(cpu);
924 * Return a low guess at the load of a migration-source cpu.
926 * We want to under-estimate the load of migration sources, to
927 * balance conservatively.
929 static inline unsigned long source_load(int cpu, int type)
931 runqueue_t *rq = cpu_rq(cpu);
932 unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE;
936 return min(rq->cpu_load[type-1], load_now);
940 * Return a high guess at the load of a migration-target cpu
942 static inline unsigned long target_load(int cpu, int type)
944 runqueue_t *rq = cpu_rq(cpu);
945 unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE;
949 return max(rq->cpu_load[type-1], load_now);
953 * find_idlest_group finds and returns the least busy CPU group within the
956 static struct sched_group *
957 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
959 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
960 unsigned long min_load = ULONG_MAX, this_load = 0;
961 int load_idx = sd->forkexec_idx;
962 int imbalance = 100 + (sd->imbalance_pct-100)/2;
965 unsigned long load, avg_load;
969 /* Skip over this group if it has no CPUs allowed */
970 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
973 local_group = cpu_isset(this_cpu, group->cpumask);
975 /* Tally up the load of all CPUs in the group */
978 for_each_cpu_mask(i, group->cpumask) {
979 /* Bias balancing toward cpus of our domain */
981 load = source_load(i, load_idx);
983 load = target_load(i, load_idx);
988 /* Adjust by relative CPU power of the group */
989 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
992 this_load = avg_load;
994 } else if (avg_load < min_load) {
1000 } while (group != sd->groups);
1002 if (!idlest || 100*this_load < imbalance*min_load)
1008 * find_idlest_queue - find the idlest runqueue among the cpus in group.
1011 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1014 unsigned long load, min_load = ULONG_MAX;
1018 /* Traverse only the allowed CPUs */
1019 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1021 for_each_cpu_mask(i, tmp) {
1022 load = source_load(i, 0);
1024 if (load < min_load || (load == min_load && i == this_cpu)) {
1034 * sched_balance_self: balance the current task (running on cpu) in domains
1035 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1038 * Balance, ie. select the least loaded group.
1040 * Returns the target CPU number, or the same CPU if no balancing is needed.
1042 * preempt must be disabled.
1044 static int sched_balance_self(int cpu, int flag)
1046 struct task_struct *t = current;
1047 struct sched_domain *tmp, *sd = NULL;
1049 for_each_domain(cpu, tmp)
1050 if (tmp->flags & flag)
1055 struct sched_group *group;
1060 group = find_idlest_group(sd, t, cpu);
1064 new_cpu = find_idlest_cpu(group, t, cpu);
1065 if (new_cpu == -1 || new_cpu == cpu)
1068 /* Now try balancing at a lower domain level */
1072 weight = cpus_weight(span);
1073 for_each_domain(cpu, tmp) {
1074 if (weight <= cpus_weight(tmp->span))
1076 if (tmp->flags & flag)
1079 /* while loop will break here if sd == NULL */
1085 #endif /* CONFIG_SMP */
1088 * wake_idle() will wake a task on an idle cpu if task->cpu is
1089 * not idle and an idle cpu is available. The span of cpus to
1090 * search starts with cpus closest then further out as needed,
1091 * so we always favor a closer, idle cpu.
1093 * Returns the CPU we should wake onto.
1095 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1096 static int wake_idle(int cpu, task_t *p)
1099 struct sched_domain *sd;
1105 for_each_domain(cpu, sd) {
1106 if (sd->flags & SD_WAKE_IDLE) {
1107 cpus_and(tmp, sd->span, p->cpus_allowed);
1108 for_each_cpu_mask(i, tmp) {
1119 static inline int wake_idle(int cpu, task_t *p)
1126 * try_to_wake_up - wake up a thread
1127 * @p: the to-be-woken-up thread
1128 * @state: the mask of task states that can be woken
1129 * @sync: do a synchronous wakeup?
1131 * Put it on the run-queue if it's not already there. The "current"
1132 * thread is always on the run-queue (except when the actual
1133 * re-schedule is in progress), and as such you're allowed to do
1134 * the simpler "current->state = TASK_RUNNING" to mark yourself
1135 * runnable without the overhead of this.
1137 * returns failure only if the task is already active.
1139 static int try_to_wake_up(task_t *p, unsigned int state, int sync)
1141 int cpu, this_cpu, success = 0;
1142 unsigned long flags;
1146 unsigned long load, this_load;
1147 struct sched_domain *sd, *this_sd = NULL;
1151 rq = task_rq_lock(p, &flags);
1152 old_state = p->state;
1153 if (!(old_state & state))
1160 this_cpu = smp_processor_id();
1163 if (unlikely(task_running(rq, p)))
1168 schedstat_inc(rq, ttwu_cnt);
1169 if (cpu == this_cpu) {
1170 schedstat_inc(rq, ttwu_local);
1174 for_each_domain(this_cpu, sd) {
1175 if (cpu_isset(cpu, sd->span)) {
1176 schedstat_inc(sd, ttwu_wake_remote);
1182 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1186 * Check for affine wakeup and passive balancing possibilities.
1189 int idx = this_sd->wake_idx;
1190 unsigned int imbalance;
1192 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1194 load = source_load(cpu, idx);
1195 this_load = target_load(this_cpu, idx);
1197 new_cpu = this_cpu; /* Wake to this CPU if we can */
1199 if (this_sd->flags & SD_WAKE_AFFINE) {
1200 unsigned long tl = this_load;
1202 * If sync wakeup then subtract the (maximum possible)
1203 * effect of the currently running task from the load
1204 * of the current CPU:
1207 tl -= SCHED_LOAD_SCALE;
1210 tl + target_load(cpu, idx) <= SCHED_LOAD_SCALE) ||
1211 100*(tl + SCHED_LOAD_SCALE) <= imbalance*load) {
1213 * This domain has SD_WAKE_AFFINE and
1214 * p is cache cold in this domain, and
1215 * there is no bad imbalance.
1217 schedstat_inc(this_sd, ttwu_move_affine);
1223 * Start passive balancing when half the imbalance_pct
1226 if (this_sd->flags & SD_WAKE_BALANCE) {
1227 if (imbalance*this_load <= 100*load) {
1228 schedstat_inc(this_sd, ttwu_move_balance);
1234 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1236 new_cpu = wake_idle(new_cpu, p);
1237 if (new_cpu != cpu) {
1238 set_task_cpu(p, new_cpu);
1239 task_rq_unlock(rq, &flags);
1240 /* might preempt at this point */
1241 rq = task_rq_lock(p, &flags);
1242 old_state = p->state;
1243 if (!(old_state & state))
1248 this_cpu = smp_processor_id();
1253 #endif /* CONFIG_SMP */
1254 if (old_state == TASK_UNINTERRUPTIBLE) {
1255 rq->nr_uninterruptible--;
1257 * Tasks on involuntary sleep don't earn
1258 * sleep_avg beyond just interactive state.
1264 * Sync wakeups (i.e. those types of wakeups where the waker
1265 * has indicated that it will leave the CPU in short order)
1266 * don't trigger a preemption, if the woken up task will run on
1267 * this cpu. (in this case the 'I will reschedule' promise of
1268 * the waker guarantees that the freshly woken up task is going
1269 * to be considered on this CPU.)
1271 activate_task(p, rq, cpu == this_cpu);
1272 if (!sync || cpu != this_cpu) {
1273 if (TASK_PREEMPTS_CURR(p, rq))
1274 resched_task(rq->curr);
1279 p->state = TASK_RUNNING;
1281 task_rq_unlock(rq, &flags);
1286 int fastcall wake_up_process(task_t *p)
1288 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1289 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1292 EXPORT_SYMBOL(wake_up_process);
1294 int fastcall wake_up_state(task_t *p, unsigned int state)
1296 return try_to_wake_up(p, state, 0);
1300 * Perform scheduler related setup for a newly forked process p.
1301 * p is forked by current.
1303 void fastcall sched_fork(task_t *p, int clone_flags)
1305 int cpu = get_cpu();
1308 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1310 set_task_cpu(p, cpu);
1313 * We mark the process as running here, but have not actually
1314 * inserted it onto the runqueue yet. This guarantees that
1315 * nobody will actually run it, and a signal or other external
1316 * event cannot wake it up and insert it on the runqueue either.
1318 p->state = TASK_RUNNING;
1319 INIT_LIST_HEAD(&p->run_list);
1321 #ifdef CONFIG_SCHEDSTATS
1322 memset(&p->sched_info, 0, sizeof(p->sched_info));
1324 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1327 #ifdef CONFIG_PREEMPT
1328 /* Want to start with kernel preemption disabled. */
1329 p->thread_info->preempt_count = 1;
1332 * Share the timeslice between parent and child, thus the
1333 * total amount of pending timeslices in the system doesn't change,
1334 * resulting in more scheduling fairness.
1336 local_irq_disable();
1337 p->time_slice = (current->time_slice + 1) >> 1;
1339 * The remainder of the first timeslice might be recovered by
1340 * the parent if the child exits early enough.
1342 p->first_time_slice = 1;
1343 current->time_slice >>= 1;
1344 p->timestamp = sched_clock();
1345 if (unlikely(!current->time_slice)) {
1347 * This case is rare, it happens when the parent has only
1348 * a single jiffy left from its timeslice. Taking the
1349 * runqueue lock is not a problem.
1351 current->time_slice = 1;
1359 * wake_up_new_task - wake up a newly created task for the first time.
1361 * This function will do some initial scheduler statistics housekeeping
1362 * that must be done for every newly created context, then puts the task
1363 * on the runqueue and wakes it.
1365 void fastcall wake_up_new_task(task_t *p, unsigned long clone_flags)
1367 unsigned long flags;
1369 runqueue_t *rq, *this_rq;
1371 rq = task_rq_lock(p, &flags);
1372 BUG_ON(p->state != TASK_RUNNING);
1373 this_cpu = smp_processor_id();
1377 * We decrease the sleep average of forking parents
1378 * and children as well, to keep max-interactive tasks
1379 * from forking tasks that are max-interactive. The parent
1380 * (current) is done further down, under its lock.
1382 p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1383 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1385 p->prio = effective_prio(p);
1387 if (likely(cpu == this_cpu)) {
1388 if (!(clone_flags & CLONE_VM)) {
1390 * The VM isn't cloned, so we're in a good position to
1391 * do child-runs-first in anticipation of an exec. This
1392 * usually avoids a lot of COW overhead.
1394 if (unlikely(!current->array))
1395 __activate_task(p, rq);
1397 p->prio = current->prio;
1398 list_add_tail(&p->run_list, ¤t->run_list);
1399 p->array = current->array;
1400 p->array->nr_active++;
1405 /* Run child last */
1406 __activate_task(p, rq);
1408 * We skip the following code due to cpu == this_cpu
1410 * task_rq_unlock(rq, &flags);
1411 * this_rq = task_rq_lock(current, &flags);
1415 this_rq = cpu_rq(this_cpu);
1418 * Not the local CPU - must adjust timestamp. This should
1419 * get optimised away in the !CONFIG_SMP case.
1421 p->timestamp = (p->timestamp - this_rq->timestamp_last_tick)
1422 + rq->timestamp_last_tick;
1423 __activate_task(p, rq);
1424 if (TASK_PREEMPTS_CURR(p, rq))
1425 resched_task(rq->curr);
1428 * Parent and child are on different CPUs, now get the
1429 * parent runqueue to update the parent's ->sleep_avg:
1431 task_rq_unlock(rq, &flags);
1432 this_rq = task_rq_lock(current, &flags);
1434 current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1435 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1436 task_rq_unlock(this_rq, &flags);
1440 * Potentially available exiting-child timeslices are
1441 * retrieved here - this way the parent does not get
1442 * penalized for creating too many threads.
1444 * (this cannot be used to 'generate' timeslices
1445 * artificially, because any timeslice recovered here
1446 * was given away by the parent in the first place.)
1448 void fastcall sched_exit(task_t *p)
1450 unsigned long flags;
1454 * If the child was a (relative-) CPU hog then decrease
1455 * the sleep_avg of the parent as well.
1457 rq = task_rq_lock(p->parent, &flags);
1458 if (p->first_time_slice) {
1459 p->parent->time_slice += p->time_slice;
1460 if (unlikely(p->parent->time_slice > task_timeslice(p)))
1461 p->parent->time_slice = task_timeslice(p);
1463 if (p->sleep_avg < p->parent->sleep_avg)
1464 p->parent->sleep_avg = p->parent->sleep_avg /
1465 (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg /
1467 task_rq_unlock(rq, &flags);
1471 * prepare_task_switch - prepare to switch tasks
1472 * @rq: the runqueue preparing to switch
1473 * @next: the task we are going to switch to.
1475 * This is called with the rq lock held and interrupts off. It must
1476 * be paired with a subsequent finish_task_switch after the context
1479 * prepare_task_switch sets up locking and calls architecture specific
1482 static inline void prepare_task_switch(runqueue_t *rq, task_t *next)
1484 prepare_lock_switch(rq, next);
1485 prepare_arch_switch(next);
1489 * finish_task_switch - clean up after a task-switch
1490 * @rq: runqueue associated with task-switch
1491 * @prev: the thread we just switched away from.
1493 * finish_task_switch must be called after the context switch, paired
1494 * with a prepare_task_switch call before the context switch.
1495 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1496 * and do any other architecture-specific cleanup actions.
1498 * Note that we may have delayed dropping an mm in context_switch(). If
1499 * so, we finish that here outside of the runqueue lock. (Doing it
1500 * with the lock held can cause deadlocks; see schedule() for
1503 static inline void finish_task_switch(runqueue_t *rq, task_t *prev)
1504 __releases(rq->lock)
1506 struct mm_struct *mm = rq->prev_mm;
1507 unsigned long prev_task_flags;
1512 * A task struct has one reference for the use as "current".
1513 * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1514 * calls schedule one last time. The schedule call will never return,
1515 * and the scheduled task must drop that reference.
1516 * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1517 * still held, otherwise prev could be scheduled on another cpu, die
1518 * there before we look at prev->state, and then the reference would
1520 * Manfred Spraul <manfred@colorfullife.com>
1522 prev_task_flags = prev->flags;
1523 #ifdef CONFIG_DEBUG_SPINLOCK
1524 /* this is a valid case when another task releases the spinlock */
1525 rq->lock.owner = current;
1527 finish_arch_switch(prev);
1528 finish_lock_switch(rq, prev);
1531 if (unlikely(prev_task_flags & PF_DEAD))
1532 put_task_struct(prev);
1536 * schedule_tail - first thing a freshly forked thread must call.
1537 * @prev: the thread we just switched away from.
1539 asmlinkage void schedule_tail(task_t *prev)
1540 __releases(rq->lock)
1542 runqueue_t *rq = this_rq();
1543 finish_task_switch(rq, prev);
1544 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1545 /* In this case, finish_task_switch does not reenable preemption */
1548 if (current->set_child_tid)
1549 put_user(current->pid, current->set_child_tid);
1553 * context_switch - switch to the new MM and the new
1554 * thread's register state.
1557 task_t * context_switch(runqueue_t *rq, task_t *prev, task_t *next)
1559 struct mm_struct *mm = next->mm;
1560 struct mm_struct *oldmm = prev->active_mm;
1562 if (unlikely(!mm)) {
1563 next->active_mm = oldmm;
1564 atomic_inc(&oldmm->mm_count);
1565 enter_lazy_tlb(oldmm, next);
1567 switch_mm(oldmm, mm, next);
1569 if (unlikely(!prev->mm)) {
1570 prev->active_mm = NULL;
1571 WARN_ON(rq->prev_mm);
1572 rq->prev_mm = oldmm;
1575 /* Here we just switch the register state and the stack. */
1576 switch_to(prev, next, prev);
1582 * nr_running, nr_uninterruptible and nr_context_switches:
1584 * externally visible scheduler statistics: current number of runnable
1585 * threads, current number of uninterruptible-sleeping threads, total
1586 * number of context switches performed since bootup.
1588 unsigned long nr_running(void)
1590 unsigned long i, sum = 0;
1592 for_each_online_cpu(i)
1593 sum += cpu_rq(i)->nr_running;
1598 unsigned long nr_uninterruptible(void)
1600 unsigned long i, sum = 0;
1603 sum += cpu_rq(i)->nr_uninterruptible;
1606 * Since we read the counters lockless, it might be slightly
1607 * inaccurate. Do not allow it to go below zero though:
1609 if (unlikely((long)sum < 0))
1615 unsigned long long nr_context_switches(void)
1617 unsigned long long i, sum = 0;
1620 sum += cpu_rq(i)->nr_switches;
1625 unsigned long nr_iowait(void)
1627 unsigned long i, sum = 0;
1630 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1638 * double_rq_lock - safely lock two runqueues
1640 * Note this does not disable interrupts like task_rq_lock,
1641 * you need to do so manually before calling.
1643 static void double_rq_lock(runqueue_t *rq1, runqueue_t *rq2)
1644 __acquires(rq1->lock)
1645 __acquires(rq2->lock)
1648 spin_lock(&rq1->lock);
1649 __acquire(rq2->lock); /* Fake it out ;) */
1652 spin_lock(&rq1->lock);
1653 spin_lock(&rq2->lock);
1655 spin_lock(&rq2->lock);
1656 spin_lock(&rq1->lock);
1662 * double_rq_unlock - safely unlock two runqueues
1664 * Note this does not restore interrupts like task_rq_unlock,
1665 * you need to do so manually after calling.
1667 static void double_rq_unlock(runqueue_t *rq1, runqueue_t *rq2)
1668 __releases(rq1->lock)
1669 __releases(rq2->lock)
1671 spin_unlock(&rq1->lock);
1673 spin_unlock(&rq2->lock);
1675 __release(rq2->lock);
1679 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1681 static void double_lock_balance(runqueue_t *this_rq, runqueue_t *busiest)
1682 __releases(this_rq->lock)
1683 __acquires(busiest->lock)
1684 __acquires(this_rq->lock)
1686 if (unlikely(!spin_trylock(&busiest->lock))) {
1687 if (busiest < this_rq) {
1688 spin_unlock(&this_rq->lock);
1689 spin_lock(&busiest->lock);
1690 spin_lock(&this_rq->lock);
1692 spin_lock(&busiest->lock);
1697 * If dest_cpu is allowed for this process, migrate the task to it.
1698 * This is accomplished by forcing the cpu_allowed mask to only
1699 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
1700 * the cpu_allowed mask is restored.
1702 static void sched_migrate_task(task_t *p, int dest_cpu)
1704 migration_req_t req;
1706 unsigned long flags;
1708 rq = task_rq_lock(p, &flags);
1709 if (!cpu_isset(dest_cpu, p->cpus_allowed)
1710 || unlikely(cpu_is_offline(dest_cpu)))
1713 /* force the process onto the specified CPU */
1714 if (migrate_task(p, dest_cpu, &req)) {
1715 /* Need to wait for migration thread (might exit: take ref). */
1716 struct task_struct *mt = rq->migration_thread;
1717 get_task_struct(mt);
1718 task_rq_unlock(rq, &flags);
1719 wake_up_process(mt);
1720 put_task_struct(mt);
1721 wait_for_completion(&req.done);
1725 task_rq_unlock(rq, &flags);
1729 * sched_exec - execve() is a valuable balancing opportunity, because at
1730 * this point the task has the smallest effective memory and cache footprint.
1732 void sched_exec(void)
1734 int new_cpu, this_cpu = get_cpu();
1735 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
1737 if (new_cpu != this_cpu)
1738 sched_migrate_task(current, new_cpu);
1742 * pull_task - move a task from a remote runqueue to the local runqueue.
1743 * Both runqueues must be locked.
1746 void pull_task(runqueue_t *src_rq, prio_array_t *src_array, task_t *p,
1747 runqueue_t *this_rq, prio_array_t *this_array, int this_cpu)
1749 dequeue_task(p, src_array);
1750 src_rq->nr_running--;
1751 set_task_cpu(p, this_cpu);
1752 this_rq->nr_running++;
1753 enqueue_task(p, this_array);
1754 p->timestamp = (p->timestamp - src_rq->timestamp_last_tick)
1755 + this_rq->timestamp_last_tick;
1757 * Note that idle threads have a prio of MAX_PRIO, for this test
1758 * to be always true for them.
1760 if (TASK_PREEMPTS_CURR(p, this_rq))
1761 resched_task(this_rq->curr);
1765 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1768 int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu,
1769 struct sched_domain *sd, enum idle_type idle,
1773 * We do not migrate tasks that are:
1774 * 1) running (obviously), or
1775 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1776 * 3) are cache-hot on their current CPU.
1778 if (!cpu_isset(this_cpu, p->cpus_allowed))
1782 if (task_running(rq, p))
1786 * Aggressive migration if:
1787 * 1) task is cache cold, or
1788 * 2) too many balance attempts have failed.
1791 if (sd->nr_balance_failed > sd->cache_nice_tries)
1794 if (task_hot(p, rq->timestamp_last_tick, sd))
1800 * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq,
1801 * as part of a balancing operation within "domain". Returns the number of
1804 * Called with both runqueues locked.
1806 static int move_tasks(runqueue_t *this_rq, int this_cpu, runqueue_t *busiest,
1807 unsigned long max_nr_move, struct sched_domain *sd,
1808 enum idle_type idle, int *all_pinned)
1810 prio_array_t *array, *dst_array;
1811 struct list_head *head, *curr;
1812 int idx, pulled = 0, pinned = 0;
1815 if (max_nr_move == 0)
1821 * We first consider expired tasks. Those will likely not be
1822 * executed in the near future, and they are most likely to
1823 * be cache-cold, thus switching CPUs has the least effect
1826 if (busiest->expired->nr_active) {
1827 array = busiest->expired;
1828 dst_array = this_rq->expired;
1830 array = busiest->active;
1831 dst_array = this_rq->active;
1835 /* Start searching at priority 0: */
1839 idx = sched_find_first_bit(array->bitmap);
1841 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
1842 if (idx >= MAX_PRIO) {
1843 if (array == busiest->expired && busiest->active->nr_active) {
1844 array = busiest->active;
1845 dst_array = this_rq->active;
1851 head = array->queue + idx;
1854 tmp = list_entry(curr, task_t, run_list);
1858 if (!can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
1865 #ifdef CONFIG_SCHEDSTATS
1866 if (task_hot(tmp, busiest->timestamp_last_tick, sd))
1867 schedstat_inc(sd, lb_hot_gained[idle]);
1870 pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
1873 /* We only want to steal up to the prescribed number of tasks. */
1874 if (pulled < max_nr_move) {
1882 * Right now, this is the only place pull_task() is called,
1883 * so we can safely collect pull_task() stats here rather than
1884 * inside pull_task().
1886 schedstat_add(sd, lb_gained[idle], pulled);
1889 *all_pinned = pinned;
1894 * find_busiest_group finds and returns the busiest CPU group within the
1895 * domain. It calculates and returns the number of tasks which should be
1896 * moved to restore balance via the imbalance parameter.
1898 static struct sched_group *
1899 find_busiest_group(struct sched_domain *sd, int this_cpu,
1900 unsigned long *imbalance, enum idle_type idle)
1902 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
1903 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
1906 max_load = this_load = total_load = total_pwr = 0;
1907 if (idle == NOT_IDLE)
1908 load_idx = sd->busy_idx;
1909 else if (idle == NEWLY_IDLE)
1910 load_idx = sd->newidle_idx;
1912 load_idx = sd->idle_idx;
1919 local_group = cpu_isset(this_cpu, group->cpumask);
1921 /* Tally up the load of all CPUs in the group */
1924 for_each_cpu_mask(i, group->cpumask) {
1925 /* Bias balancing toward cpus of our domain */
1927 load = target_load(i, load_idx);
1929 load = source_load(i, load_idx);
1934 total_load += avg_load;
1935 total_pwr += group->cpu_power;
1937 /* Adjust by relative CPU power of the group */
1938 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1941 this_load = avg_load;
1943 } else if (avg_load > max_load) {
1944 max_load = avg_load;
1947 group = group->next;
1948 } while (group != sd->groups);
1950 if (!busiest || this_load >= max_load)
1953 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
1955 if (this_load >= avg_load ||
1956 100*max_load <= sd->imbalance_pct*this_load)
1960 * We're trying to get all the cpus to the average_load, so we don't
1961 * want to push ourselves above the average load, nor do we wish to
1962 * reduce the max loaded cpu below the average load, as either of these
1963 * actions would just result in more rebalancing later, and ping-pong
1964 * tasks around. Thus we look for the minimum possible imbalance.
1965 * Negative imbalances (*we* are more loaded than anyone else) will
1966 * be counted as no imbalance for these purposes -- we can't fix that
1967 * by pulling tasks to us. Be careful of negative numbers as they'll
1968 * appear as very large values with unsigned longs.
1970 /* How much load to actually move to equalise the imbalance */
1971 *imbalance = min((max_load - avg_load) * busiest->cpu_power,
1972 (avg_load - this_load) * this->cpu_power)
1975 if (*imbalance < SCHED_LOAD_SCALE) {
1976 unsigned long pwr_now = 0, pwr_move = 0;
1979 if (max_load - this_load >= SCHED_LOAD_SCALE*2) {
1985 * OK, we don't have enough imbalance to justify moving tasks,
1986 * however we may be able to increase total CPU power used by
1990 pwr_now += busiest->cpu_power*min(SCHED_LOAD_SCALE, max_load);
1991 pwr_now += this->cpu_power*min(SCHED_LOAD_SCALE, this_load);
1992 pwr_now /= SCHED_LOAD_SCALE;
1994 /* Amount of load we'd subtract */
1995 tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/busiest->cpu_power;
1997 pwr_move += busiest->cpu_power*min(SCHED_LOAD_SCALE,
2000 /* Amount of load we'd add */
2001 if (max_load*busiest->cpu_power <
2002 SCHED_LOAD_SCALE*SCHED_LOAD_SCALE)
2003 tmp = max_load*busiest->cpu_power/this->cpu_power;
2005 tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/this->cpu_power;
2006 pwr_move += this->cpu_power*min(SCHED_LOAD_SCALE, this_load + tmp);
2007 pwr_move /= SCHED_LOAD_SCALE;
2009 /* Move if we gain throughput */
2010 if (pwr_move <= pwr_now)
2017 /* Get rid of the scaling factor, rounding down as we divide */
2018 *imbalance = *imbalance / SCHED_LOAD_SCALE;
2028 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2030 static runqueue_t *find_busiest_queue(struct sched_group *group)
2032 unsigned long load, max_load = 0;
2033 runqueue_t *busiest = NULL;
2036 for_each_cpu_mask(i, group->cpumask) {
2037 load = source_load(i, 0);
2039 if (load > max_load) {
2041 busiest = cpu_rq(i);
2049 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2050 * so long as it is large enough.
2052 #define MAX_PINNED_INTERVAL 512
2055 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2056 * tasks if there is an imbalance.
2058 * Called with this_rq unlocked.
2060 static int load_balance(int this_cpu, runqueue_t *this_rq,
2061 struct sched_domain *sd, enum idle_type idle)
2063 struct sched_group *group;
2064 runqueue_t *busiest;
2065 unsigned long imbalance;
2066 int nr_moved, all_pinned = 0;
2067 int active_balance = 0;
2069 spin_lock(&this_rq->lock);
2070 schedstat_inc(sd, lb_cnt[idle]);
2072 group = find_busiest_group(sd, this_cpu, &imbalance, idle);
2074 schedstat_inc(sd, lb_nobusyg[idle]);
2078 busiest = find_busiest_queue(group);
2080 schedstat_inc(sd, lb_nobusyq[idle]);
2084 BUG_ON(busiest == this_rq);
2086 schedstat_add(sd, lb_imbalance[idle], imbalance);
2089 if (busiest->nr_running > 1) {
2091 * Attempt to move tasks. If find_busiest_group has found
2092 * an imbalance but busiest->nr_running <= 1, the group is
2093 * still unbalanced. nr_moved simply stays zero, so it is
2094 * correctly treated as an imbalance.
2096 double_lock_balance(this_rq, busiest);
2097 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2098 imbalance, sd, idle,
2100 spin_unlock(&busiest->lock);
2102 /* All tasks on this runqueue were pinned by CPU affinity */
2103 if (unlikely(all_pinned))
2107 spin_unlock(&this_rq->lock);
2110 schedstat_inc(sd, lb_failed[idle]);
2111 sd->nr_balance_failed++;
2113 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2115 spin_lock(&busiest->lock);
2116 if (!busiest->active_balance) {
2117 busiest->active_balance = 1;
2118 busiest->push_cpu = this_cpu;
2121 spin_unlock(&busiest->lock);
2123 wake_up_process(busiest->migration_thread);
2126 * We've kicked active balancing, reset the failure
2129 sd->nr_balance_failed = sd->cache_nice_tries+1;
2132 sd->nr_balance_failed = 0;
2134 if (likely(!active_balance)) {
2135 /* We were unbalanced, so reset the balancing interval */
2136 sd->balance_interval = sd->min_interval;
2139 * If we've begun active balancing, start to back off. This
2140 * case may not be covered by the all_pinned logic if there
2141 * is only 1 task on the busy runqueue (because we don't call
2144 if (sd->balance_interval < sd->max_interval)
2145 sd->balance_interval *= 2;
2151 spin_unlock(&this_rq->lock);
2153 schedstat_inc(sd, lb_balanced[idle]);
2155 sd->nr_balance_failed = 0;
2156 /* tune up the balancing interval */
2157 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2158 (sd->balance_interval < sd->max_interval))
2159 sd->balance_interval *= 2;
2165 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2166 * tasks if there is an imbalance.
2168 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2169 * this_rq is locked.
2171 static int load_balance_newidle(int this_cpu, runqueue_t *this_rq,
2172 struct sched_domain *sd)
2174 struct sched_group *group;
2175 runqueue_t *busiest = NULL;
2176 unsigned long imbalance;
2179 schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
2180 group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE);
2182 schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]);
2186 busiest = find_busiest_queue(group);
2188 schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]);
2192 BUG_ON(busiest == this_rq);
2194 /* Attempt to move tasks */
2195 double_lock_balance(this_rq, busiest);
2197 schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
2198 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2199 imbalance, sd, NEWLY_IDLE, NULL);
2201 schedstat_inc(sd, lb_failed[NEWLY_IDLE]);
2203 sd->nr_balance_failed = 0;
2205 spin_unlock(&busiest->lock);
2209 schedstat_inc(sd, lb_balanced[NEWLY_IDLE]);
2210 sd->nr_balance_failed = 0;
2215 * idle_balance is called by schedule() if this_cpu is about to become
2216 * idle. Attempts to pull tasks from other CPUs.
2218 static inline void idle_balance(int this_cpu, runqueue_t *this_rq)
2220 struct sched_domain *sd;
2222 for_each_domain(this_cpu, sd) {
2223 if (sd->flags & SD_BALANCE_NEWIDLE) {
2224 if (load_balance_newidle(this_cpu, this_rq, sd)) {
2225 /* We've pulled tasks over so stop searching */
2233 * active_load_balance is run by migration threads. It pushes running tasks
2234 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2235 * running on each physical CPU where possible, and avoids physical /
2236 * logical imbalances.
2238 * Called with busiest_rq locked.
2240 static void active_load_balance(runqueue_t *busiest_rq, int busiest_cpu)
2242 struct sched_domain *sd;
2243 runqueue_t *target_rq;
2244 int target_cpu = busiest_rq->push_cpu;
2246 if (busiest_rq->nr_running <= 1)
2247 /* no task to move */
2250 target_rq = cpu_rq(target_cpu);
2253 * This condition is "impossible", if it occurs
2254 * we need to fix it. Originally reported by
2255 * Bjorn Helgaas on a 128-cpu setup.
2257 BUG_ON(busiest_rq == target_rq);
2259 /* move a task from busiest_rq to target_rq */
2260 double_lock_balance(busiest_rq, target_rq);
2262 /* Search for an sd spanning us and the target CPU. */
2263 for_each_domain(target_cpu, sd)
2264 if ((sd->flags & SD_LOAD_BALANCE) &&
2265 cpu_isset(busiest_cpu, sd->span))
2268 if (unlikely(sd == NULL))
2271 schedstat_inc(sd, alb_cnt);
2273 if (move_tasks(target_rq, target_cpu, busiest_rq, 1, sd, SCHED_IDLE, NULL))
2274 schedstat_inc(sd, alb_pushed);
2276 schedstat_inc(sd, alb_failed);
2278 spin_unlock(&target_rq->lock);
2282 * rebalance_tick will get called every timer tick, on every CPU.
2284 * It checks each scheduling domain to see if it is due to be balanced,
2285 * and initiates a balancing operation if so.
2287 * Balancing parameters are set up in arch_init_sched_domains.
2290 /* Don't have all balancing operations going off at once */
2291 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2293 static void rebalance_tick(int this_cpu, runqueue_t *this_rq,
2294 enum idle_type idle)
2296 unsigned long old_load, this_load;
2297 unsigned long j = jiffies + CPU_OFFSET(this_cpu);
2298 struct sched_domain *sd;
2301 this_load = this_rq->nr_running * SCHED_LOAD_SCALE;
2302 /* Update our load */
2303 for (i = 0; i < 3; i++) {
2304 unsigned long new_load = this_load;
2306 old_load = this_rq->cpu_load[i];
2308 * Round up the averaging division if load is increasing. This
2309 * prevents us from getting stuck on 9 if the load is 10, for
2312 if (new_load > old_load)
2313 new_load += scale-1;
2314 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale;
2317 for_each_domain(this_cpu, sd) {
2318 unsigned long interval;
2320 if (!(sd->flags & SD_LOAD_BALANCE))
2323 interval = sd->balance_interval;
2324 if (idle != SCHED_IDLE)
2325 interval *= sd->busy_factor;
2327 /* scale ms to jiffies */
2328 interval = msecs_to_jiffies(interval);
2329 if (unlikely(!interval))
2332 if (j - sd->last_balance >= interval) {
2333 if (load_balance(this_cpu, this_rq, sd, idle)) {
2334 /* We've pulled tasks over so no longer idle */
2337 sd->last_balance += interval;
2343 * on UP we do not need to balance between CPUs:
2345 static inline void rebalance_tick(int cpu, runqueue_t *rq, enum idle_type idle)
2348 static inline void idle_balance(int cpu, runqueue_t *rq)
2353 static inline int wake_priority_sleeper(runqueue_t *rq)
2356 #ifdef CONFIG_SCHED_SMT
2357 spin_lock(&rq->lock);
2359 * If an SMT sibling task has been put to sleep for priority
2360 * reasons reschedule the idle task to see if it can now run.
2362 if (rq->nr_running) {
2363 resched_task(rq->idle);
2366 spin_unlock(&rq->lock);
2371 DEFINE_PER_CPU(struct kernel_stat, kstat);
2373 EXPORT_PER_CPU_SYMBOL(kstat);
2376 * This is called on clock ticks and on context switches.
2377 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2379 static inline void update_cpu_clock(task_t *p, runqueue_t *rq,
2380 unsigned long long now)
2382 unsigned long long last = max(p->timestamp, rq->timestamp_last_tick);
2383 p->sched_time += now - last;
2387 * Return current->sched_time plus any more ns on the sched_clock
2388 * that have not yet been banked.
2390 unsigned long long current_sched_time(const task_t *tsk)
2392 unsigned long long ns;
2393 unsigned long flags;
2394 local_irq_save(flags);
2395 ns = max(tsk->timestamp, task_rq(tsk)->timestamp_last_tick);
2396 ns = tsk->sched_time + (sched_clock() - ns);
2397 local_irq_restore(flags);
2402 * We place interactive tasks back into the active array, if possible.
2404 * To guarantee that this does not starve expired tasks we ignore the
2405 * interactivity of a task if the first expired task had to wait more
2406 * than a 'reasonable' amount of time. This deadline timeout is
2407 * load-dependent, as the frequency of array switched decreases with
2408 * increasing number of running tasks. We also ignore the interactivity
2409 * if a better static_prio task has expired:
2411 #define EXPIRED_STARVING(rq) \
2412 ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2413 (jiffies - (rq)->expired_timestamp >= \
2414 STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2415 ((rq)->curr->static_prio > (rq)->best_expired_prio))
2418 * Account user cpu time to a process.
2419 * @p: the process that the cpu time gets accounted to
2420 * @hardirq_offset: the offset to subtract from hardirq_count()
2421 * @cputime: the cpu time spent in user space since the last update
2423 void account_user_time(struct task_struct *p, cputime_t cputime)
2425 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2428 p->utime = cputime_add(p->utime, cputime);
2430 /* Add user time to cpustat. */
2431 tmp = cputime_to_cputime64(cputime);
2432 if (TASK_NICE(p) > 0)
2433 cpustat->nice = cputime64_add(cpustat->nice, tmp);
2435 cpustat->user = cputime64_add(cpustat->user, tmp);
2439 * Account system cpu time to a process.
2440 * @p: the process that the cpu time gets accounted to
2441 * @hardirq_offset: the offset to subtract from hardirq_count()
2442 * @cputime: the cpu time spent in kernel space since the last update
2444 void account_system_time(struct task_struct *p, int hardirq_offset,
2447 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2448 runqueue_t *rq = this_rq();
2451 p->stime = cputime_add(p->stime, cputime);
2453 /* Add system time to cpustat. */
2454 tmp = cputime_to_cputime64(cputime);
2455 if (hardirq_count() - hardirq_offset)
2456 cpustat->irq = cputime64_add(cpustat->irq, tmp);
2457 else if (softirq_count())
2458 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
2459 else if (p != rq->idle)
2460 cpustat->system = cputime64_add(cpustat->system, tmp);
2461 else if (atomic_read(&rq->nr_iowait) > 0)
2462 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2464 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2465 /* Account for system time used */
2466 acct_update_integrals(p);
2467 /* Update rss highwater mark */
2468 update_mem_hiwater(p);
2472 * Account for involuntary wait time.
2473 * @p: the process from which the cpu time has been stolen
2474 * @steal: the cpu time spent in involuntary wait
2476 void account_steal_time(struct task_struct *p, cputime_t steal)
2478 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2479 cputime64_t tmp = cputime_to_cputime64(steal);
2480 runqueue_t *rq = this_rq();
2482 if (p == rq->idle) {
2483 p->stime = cputime_add(p->stime, steal);
2484 if (atomic_read(&rq->nr_iowait) > 0)
2485 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2487 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2489 cpustat->steal = cputime64_add(cpustat->steal, tmp);
2493 * This function gets called by the timer code, with HZ frequency.
2494 * We call it with interrupts disabled.
2496 * It also gets called by the fork code, when changing the parent's
2499 void scheduler_tick(void)
2501 int cpu = smp_processor_id();
2502 runqueue_t *rq = this_rq();
2503 task_t *p = current;
2504 unsigned long long now = sched_clock();
2506 update_cpu_clock(p, rq, now);
2508 rq->timestamp_last_tick = now;
2510 if (p == rq->idle) {
2511 if (wake_priority_sleeper(rq))
2513 rebalance_tick(cpu, rq, SCHED_IDLE);
2517 /* Task might have expired already, but not scheduled off yet */
2518 if (p->array != rq->active) {
2519 set_tsk_need_resched(p);
2522 spin_lock(&rq->lock);
2524 * The task was running during this tick - update the
2525 * time slice counter. Note: we do not update a thread's
2526 * priority until it either goes to sleep or uses up its
2527 * timeslice. This makes it possible for interactive tasks
2528 * to use up their timeslices at their highest priority levels.
2532 * RR tasks need a special form of timeslice management.
2533 * FIFO tasks have no timeslices.
2535 if ((p->policy == SCHED_RR) && !--p->time_slice) {
2536 p->time_slice = task_timeslice(p);
2537 p->first_time_slice = 0;
2538 set_tsk_need_resched(p);
2540 /* put it at the end of the queue: */
2541 requeue_task(p, rq->active);
2545 if (!--p->time_slice) {
2546 dequeue_task(p, rq->active);
2547 set_tsk_need_resched(p);
2548 p->prio = effective_prio(p);
2549 p->time_slice = task_timeslice(p);
2550 p->first_time_slice = 0;
2552 if (!rq->expired_timestamp)
2553 rq->expired_timestamp = jiffies;
2554 if (!TASK_INTERACTIVE(p) || EXPIRED_STARVING(rq)) {
2555 enqueue_task(p, rq->expired);
2556 if (p->static_prio < rq->best_expired_prio)
2557 rq->best_expired_prio = p->static_prio;
2559 enqueue_task(p, rq->active);
2562 * Prevent a too long timeslice allowing a task to monopolize
2563 * the CPU. We do this by splitting up the timeslice into
2566 * Note: this does not mean the task's timeslices expire or
2567 * get lost in any way, they just might be preempted by
2568 * another task of equal priority. (one with higher
2569 * priority would have preempted this task already.) We
2570 * requeue this task to the end of the list on this priority
2571 * level, which is in essence a round-robin of tasks with
2574 * This only applies to tasks in the interactive
2575 * delta range with at least TIMESLICE_GRANULARITY to requeue.
2577 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
2578 p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
2579 (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
2580 (p->array == rq->active)) {
2582 requeue_task(p, rq->active);
2583 set_tsk_need_resched(p);
2587 spin_unlock(&rq->lock);
2589 rebalance_tick(cpu, rq, NOT_IDLE);
2592 #ifdef CONFIG_SCHED_SMT
2593 static inline void wakeup_busy_runqueue(runqueue_t *rq)
2595 /* If an SMT runqueue is sleeping due to priority reasons wake it up */
2596 if (rq->curr == rq->idle && rq->nr_running)
2597 resched_task(rq->idle);
2600 static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2602 struct sched_domain *tmp, *sd = NULL;
2603 cpumask_t sibling_map;
2606 for_each_domain(this_cpu, tmp)
2607 if (tmp->flags & SD_SHARE_CPUPOWER)
2614 * Unlock the current runqueue because we have to lock in
2615 * CPU order to avoid deadlocks. Caller knows that we might
2616 * unlock. We keep IRQs disabled.
2618 spin_unlock(&this_rq->lock);
2620 sibling_map = sd->span;
2622 for_each_cpu_mask(i, sibling_map)
2623 spin_lock(&cpu_rq(i)->lock);
2625 * We clear this CPU from the mask. This both simplifies the
2626 * inner loop and keps this_rq locked when we exit:
2628 cpu_clear(this_cpu, sibling_map);
2630 for_each_cpu_mask(i, sibling_map) {
2631 runqueue_t *smt_rq = cpu_rq(i);
2633 wakeup_busy_runqueue(smt_rq);
2636 for_each_cpu_mask(i, sibling_map)
2637 spin_unlock(&cpu_rq(i)->lock);
2639 * We exit with this_cpu's rq still held and IRQs
2644 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2646 struct sched_domain *tmp, *sd = NULL;
2647 cpumask_t sibling_map;
2648 prio_array_t *array;
2652 for_each_domain(this_cpu, tmp)
2653 if (tmp->flags & SD_SHARE_CPUPOWER)
2660 * The same locking rules and details apply as for
2661 * wake_sleeping_dependent():
2663 spin_unlock(&this_rq->lock);
2664 sibling_map = sd->span;
2665 for_each_cpu_mask(i, sibling_map)
2666 spin_lock(&cpu_rq(i)->lock);
2667 cpu_clear(this_cpu, sibling_map);
2670 * Establish next task to be run - it might have gone away because
2671 * we released the runqueue lock above:
2673 if (!this_rq->nr_running)
2675 array = this_rq->active;
2676 if (!array->nr_active)
2677 array = this_rq->expired;
2678 BUG_ON(!array->nr_active);
2680 p = list_entry(array->queue[sched_find_first_bit(array->bitmap)].next,
2683 for_each_cpu_mask(i, sibling_map) {
2684 runqueue_t *smt_rq = cpu_rq(i);
2685 task_t *smt_curr = smt_rq->curr;
2687 /* Kernel threads do not participate in dependent sleeping */
2688 if (!p->mm || !smt_curr->mm || rt_task(p))
2689 goto check_smt_task;
2692 * If a user task with lower static priority than the
2693 * running task on the SMT sibling is trying to schedule,
2694 * delay it till there is proportionately less timeslice
2695 * left of the sibling task to prevent a lower priority
2696 * task from using an unfair proportion of the
2697 * physical cpu's resources. -ck
2699 if (rt_task(smt_curr)) {
2701 * With real time tasks we run non-rt tasks only
2702 * per_cpu_gain% of the time.
2704 if ((jiffies % DEF_TIMESLICE) >
2705 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
2708 if (((smt_curr->time_slice * (100 - sd->per_cpu_gain) /
2709 100) > task_timeslice(p)))
2713 if ((!smt_curr->mm && smt_curr != smt_rq->idle) ||
2717 wakeup_busy_runqueue(smt_rq);
2722 * Reschedule a lower priority task on the SMT sibling for
2723 * it to be put to sleep, or wake it up if it has been put to
2724 * sleep for priority reasons to see if it should run now.
2727 if ((jiffies % DEF_TIMESLICE) >
2728 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
2729 resched_task(smt_curr);
2731 if ((p->time_slice * (100 - sd->per_cpu_gain) / 100) >
2732 task_timeslice(smt_curr))
2733 resched_task(smt_curr);
2735 wakeup_busy_runqueue(smt_rq);
2739 for_each_cpu_mask(i, sibling_map)
2740 spin_unlock(&cpu_rq(i)->lock);
2744 static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2748 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2754 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
2756 void fastcall add_preempt_count(int val)
2761 BUG_ON((preempt_count() < 0));
2762 preempt_count() += val;
2764 * Spinlock count overflowing soon?
2766 BUG_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10);
2768 EXPORT_SYMBOL(add_preempt_count);
2770 void fastcall sub_preempt_count(int val)
2775 BUG_ON(val > preempt_count());
2777 * Is the spinlock portion underflowing?
2779 BUG_ON((val < PREEMPT_MASK) && !(preempt_count() & PREEMPT_MASK));
2780 preempt_count() -= val;
2782 EXPORT_SYMBOL(sub_preempt_count);
2787 * schedule() is the main scheduler function.
2789 asmlinkage void __sched schedule(void)
2792 task_t *prev, *next;
2794 prio_array_t *array;
2795 struct list_head *queue;
2796 unsigned long long now;
2797 unsigned long run_time;
2798 int cpu, idx, new_prio;
2801 * Test if we are atomic. Since do_exit() needs to call into
2802 * schedule() atomically, we ignore that path for now.
2803 * Otherwise, whine if we are scheduling when we should not be.
2805 if (likely(!current->exit_state)) {
2806 if (unlikely(in_atomic())) {
2807 printk(KERN_ERR "scheduling while atomic: "
2809 current->comm, preempt_count(), current->pid);
2813 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2818 release_kernel_lock(prev);
2819 need_resched_nonpreemptible:
2823 * The idle thread is not allowed to schedule!
2824 * Remove this check after it has been exercised a bit.
2826 if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
2827 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
2831 schedstat_inc(rq, sched_cnt);
2832 now = sched_clock();
2833 if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
2834 run_time = now - prev->timestamp;
2835 if (unlikely((long long)(now - prev->timestamp) < 0))
2838 run_time = NS_MAX_SLEEP_AVG;
2841 * Tasks charged proportionately less run_time at high sleep_avg to
2842 * delay them losing their interactive status
2844 run_time /= (CURRENT_BONUS(prev) ? : 1);
2846 spin_lock_irq(&rq->lock);
2848 if (unlikely(prev->flags & PF_DEAD))
2849 prev->state = EXIT_DEAD;
2851 switch_count = &prev->nivcsw;
2852 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2853 switch_count = &prev->nvcsw;
2854 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
2855 unlikely(signal_pending(prev))))
2856 prev->state = TASK_RUNNING;
2858 if (prev->state == TASK_UNINTERRUPTIBLE)
2859 rq->nr_uninterruptible++;
2860 deactivate_task(prev, rq);
2864 cpu = smp_processor_id();
2865 if (unlikely(!rq->nr_running)) {
2867 idle_balance(cpu, rq);
2868 if (!rq->nr_running) {
2870 rq->expired_timestamp = 0;
2871 wake_sleeping_dependent(cpu, rq);
2873 * wake_sleeping_dependent() might have released
2874 * the runqueue, so break out if we got new
2877 if (!rq->nr_running)
2881 if (dependent_sleeper(cpu, rq)) {
2886 * dependent_sleeper() releases and reacquires the runqueue
2887 * lock, hence go into the idle loop if the rq went
2890 if (unlikely(!rq->nr_running))
2895 if (unlikely(!array->nr_active)) {
2897 * Switch the active and expired arrays.
2899 schedstat_inc(rq, sched_switch);
2900 rq->active = rq->expired;
2901 rq->expired = array;
2903 rq->expired_timestamp = 0;
2904 rq->best_expired_prio = MAX_PRIO;
2907 idx = sched_find_first_bit(array->bitmap);
2908 queue = array->queue + idx;
2909 next = list_entry(queue->next, task_t, run_list);
2911 if (!rt_task(next) && next->activated > 0) {
2912 unsigned long long delta = now - next->timestamp;
2913 if (unlikely((long long)(now - next->timestamp) < 0))
2916 if (next->activated == 1)
2917 delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
2919 array = next->array;
2920 new_prio = recalc_task_prio(next, next->timestamp + delta);
2922 if (unlikely(next->prio != new_prio)) {
2923 dequeue_task(next, array);
2924 next->prio = new_prio;
2925 enqueue_task(next, array);
2927 requeue_task(next, array);
2929 next->activated = 0;
2931 if (next == rq->idle)
2932 schedstat_inc(rq, sched_goidle);
2934 prefetch_stack(next);
2935 clear_tsk_need_resched(prev);
2936 rcu_qsctr_inc(task_cpu(prev));
2938 update_cpu_clock(prev, rq, now);
2940 prev->sleep_avg -= run_time;
2941 if ((long)prev->sleep_avg <= 0)
2942 prev->sleep_avg = 0;
2943 prev->timestamp = prev->last_ran = now;
2945 sched_info_switch(prev, next);
2946 if (likely(prev != next)) {
2947 next->timestamp = now;
2952 prepare_task_switch(rq, next);
2953 prev = context_switch(rq, prev, next);
2956 * this_rq must be evaluated again because prev may have moved
2957 * CPUs since it called schedule(), thus the 'rq' on its stack
2958 * frame will be invalid.
2960 finish_task_switch(this_rq(), prev);
2962 spin_unlock_irq(&rq->lock);
2965 if (unlikely(reacquire_kernel_lock(prev) < 0))
2966 goto need_resched_nonpreemptible;
2967 preempt_enable_no_resched();
2968 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
2972 EXPORT_SYMBOL(schedule);
2974 #ifdef CONFIG_PREEMPT
2976 * this is is the entry point to schedule() from in-kernel preemption
2977 * off of preempt_enable. Kernel preemptions off return from interrupt
2978 * occur there and call schedule directly.
2980 asmlinkage void __sched preempt_schedule(void)
2982 struct thread_info *ti = current_thread_info();
2983 #ifdef CONFIG_PREEMPT_BKL
2984 struct task_struct *task = current;
2985 int saved_lock_depth;
2988 * If there is a non-zero preempt_count or interrupts are disabled,
2989 * we do not want to preempt the current task. Just return..
2991 if (unlikely(ti->preempt_count || irqs_disabled()))
2995 add_preempt_count(PREEMPT_ACTIVE);
2997 * We keep the big kernel semaphore locked, but we
2998 * clear ->lock_depth so that schedule() doesnt
2999 * auto-release the semaphore:
3001 #ifdef CONFIG_PREEMPT_BKL
3002 saved_lock_depth = task->lock_depth;
3003 task->lock_depth = -1;
3006 #ifdef CONFIG_PREEMPT_BKL
3007 task->lock_depth = saved_lock_depth;
3009 sub_preempt_count(PREEMPT_ACTIVE);
3011 /* we could miss a preemption opportunity between schedule and now */
3013 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3017 EXPORT_SYMBOL(preempt_schedule);
3020 * this is is the entry point to schedule() from kernel preemption
3021 * off of irq context.
3022 * Note, that this is called and return with irqs disabled. This will
3023 * protect us against recursive calling from irq.
3025 asmlinkage void __sched preempt_schedule_irq(void)
3027 struct thread_info *ti = current_thread_info();
3028 #ifdef CONFIG_PREEMPT_BKL
3029 struct task_struct *task = current;
3030 int saved_lock_depth;
3032 /* Catch callers which need to be fixed*/
3033 BUG_ON(ti->preempt_count || !irqs_disabled());
3036 add_preempt_count(PREEMPT_ACTIVE);
3038 * We keep the big kernel semaphore locked, but we
3039 * clear ->lock_depth so that schedule() doesnt
3040 * auto-release the semaphore:
3042 #ifdef CONFIG_PREEMPT_BKL
3043 saved_lock_depth = task->lock_depth;
3044 task->lock_depth = -1;
3048 local_irq_disable();
3049 #ifdef CONFIG_PREEMPT_BKL
3050 task->lock_depth = saved_lock_depth;
3052 sub_preempt_count(PREEMPT_ACTIVE);
3054 /* we could miss a preemption opportunity between schedule and now */
3056 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3060 #endif /* CONFIG_PREEMPT */
3062 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3065 task_t *p = curr->private;
3066 return try_to_wake_up(p, mode, sync);
3069 EXPORT_SYMBOL(default_wake_function);
3072 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3073 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3074 * number) then we wake all the non-exclusive tasks and one exclusive task.
3076 * There are circumstances in which we can try to wake a task which has already
3077 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3078 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3080 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3081 int nr_exclusive, int sync, void *key)
3083 struct list_head *tmp, *next;
3085 list_for_each_safe(tmp, next, &q->task_list) {
3088 curr = list_entry(tmp, wait_queue_t, task_list);
3089 flags = curr->flags;
3090 if (curr->func(curr, mode, sync, key) &&
3091 (flags & WQ_FLAG_EXCLUSIVE) &&
3098 * __wake_up - wake up threads blocked on a waitqueue.
3100 * @mode: which threads
3101 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3102 * @key: is directly passed to the wakeup function
3104 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3105 int nr_exclusive, void *key)
3107 unsigned long flags;
3109 spin_lock_irqsave(&q->lock, flags);
3110 __wake_up_common(q, mode, nr_exclusive, 0, key);
3111 spin_unlock_irqrestore(&q->lock, flags);
3114 EXPORT_SYMBOL(__wake_up);
3117 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3119 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3121 __wake_up_common(q, mode, 1, 0, NULL);
3125 * __wake_up_sync - wake up threads blocked on a waitqueue.
3127 * @mode: which threads
3128 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3130 * The sync wakeup differs that the waker knows that it will schedule
3131 * away soon, so while the target thread will be woken up, it will not
3132 * be migrated to another CPU - ie. the two threads are 'synchronized'
3133 * with each other. This can prevent needless bouncing between CPUs.
3135 * On UP it can prevent extra preemption.
3138 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3140 unsigned long flags;
3146 if (unlikely(!nr_exclusive))
3149 spin_lock_irqsave(&q->lock, flags);
3150 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3151 spin_unlock_irqrestore(&q->lock, flags);
3153 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3155 void fastcall complete(struct completion *x)
3157 unsigned long flags;
3159 spin_lock_irqsave(&x->wait.lock, flags);
3161 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3163 spin_unlock_irqrestore(&x->wait.lock, flags);
3165 EXPORT_SYMBOL(complete);
3167 void fastcall complete_all(struct completion *x)
3169 unsigned long flags;
3171 spin_lock_irqsave(&x->wait.lock, flags);
3172 x->done += UINT_MAX/2;
3173 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3175 spin_unlock_irqrestore(&x->wait.lock, flags);
3177 EXPORT_SYMBOL(complete_all);
3179 void fastcall __sched wait_for_completion(struct completion *x)
3182 spin_lock_irq(&x->wait.lock);
3184 DECLARE_WAITQUEUE(wait, current);
3186 wait.flags |= WQ_FLAG_EXCLUSIVE;
3187 __add_wait_queue_tail(&x->wait, &wait);
3189 __set_current_state(TASK_UNINTERRUPTIBLE);
3190 spin_unlock_irq(&x->wait.lock);
3192 spin_lock_irq(&x->wait.lock);
3194 __remove_wait_queue(&x->wait, &wait);
3197 spin_unlock_irq(&x->wait.lock);
3199 EXPORT_SYMBOL(wait_for_completion);
3201 unsigned long fastcall __sched
3202 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3206 spin_lock_irq(&x->wait.lock);
3208 DECLARE_WAITQUEUE(wait, current);
3210 wait.flags |= WQ_FLAG_EXCLUSIVE;
3211 __add_wait_queue_tail(&x->wait, &wait);
3213 __set_current_state(TASK_UNINTERRUPTIBLE);
3214 spin_unlock_irq(&x->wait.lock);
3215 timeout = schedule_timeout(timeout);
3216 spin_lock_irq(&x->wait.lock);
3218 __remove_wait_queue(&x->wait, &wait);
3222 __remove_wait_queue(&x->wait, &wait);
3226 spin_unlock_irq(&x->wait.lock);
3229 EXPORT_SYMBOL(wait_for_completion_timeout);
3231 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3237 spin_lock_irq(&x->wait.lock);
3239 DECLARE_WAITQUEUE(wait, current);
3241 wait.flags |= WQ_FLAG_EXCLUSIVE;
3242 __add_wait_queue_tail(&x->wait, &wait);
3244 if (signal_pending(current)) {
3246 __remove_wait_queue(&x->wait, &wait);
3249 __set_current_state(TASK_INTERRUPTIBLE);
3250 spin_unlock_irq(&x->wait.lock);
3252 spin_lock_irq(&x->wait.lock);
3254 __remove_wait_queue(&x->wait, &wait);
3258 spin_unlock_irq(&x->wait.lock);
3262 EXPORT_SYMBOL(wait_for_completion_interruptible);
3264 unsigned long fastcall __sched
3265 wait_for_completion_interruptible_timeout(struct completion *x,
3266 unsigned long timeout)
3270 spin_lock_irq(&x->wait.lock);
3272 DECLARE_WAITQUEUE(wait, current);
3274 wait.flags |= WQ_FLAG_EXCLUSIVE;
3275 __add_wait_queue_tail(&x->wait, &wait);
3277 if (signal_pending(current)) {
3278 timeout = -ERESTARTSYS;
3279 __remove_wait_queue(&x->wait, &wait);
3282 __set_current_state(TASK_INTERRUPTIBLE);
3283 spin_unlock_irq(&x->wait.lock);
3284 timeout = schedule_timeout(timeout);
3285 spin_lock_irq(&x->wait.lock);
3287 __remove_wait_queue(&x->wait, &wait);
3291 __remove_wait_queue(&x->wait, &wait);
3295 spin_unlock_irq(&x->wait.lock);
3298 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3301 #define SLEEP_ON_VAR \
3302 unsigned long flags; \
3303 wait_queue_t wait; \
3304 init_waitqueue_entry(&wait, current);
3306 #define SLEEP_ON_HEAD \
3307 spin_lock_irqsave(&q->lock,flags); \
3308 __add_wait_queue(q, &wait); \
3309 spin_unlock(&q->lock);
3311 #define SLEEP_ON_TAIL \
3312 spin_lock_irq(&q->lock); \
3313 __remove_wait_queue(q, &wait); \
3314 spin_unlock_irqrestore(&q->lock, flags);
3316 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
3320 current->state = TASK_INTERRUPTIBLE;
3327 EXPORT_SYMBOL(interruptible_sleep_on);
3329 long fastcall __sched
3330 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3334 current->state = TASK_INTERRUPTIBLE;
3337 timeout = schedule_timeout(timeout);
3343 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3345 void fastcall __sched sleep_on(wait_queue_head_t *q)
3349 current->state = TASK_UNINTERRUPTIBLE;
3356 EXPORT_SYMBOL(sleep_on);
3358 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3362 current->state = TASK_UNINTERRUPTIBLE;
3365 timeout = schedule_timeout(timeout);
3371 EXPORT_SYMBOL(sleep_on_timeout);
3373 void set_user_nice(task_t *p, long nice)
3375 unsigned long flags;
3376 prio_array_t *array;
3378 int old_prio, new_prio, delta;
3380 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3383 * We have to be careful, if called from sys_setpriority(),
3384 * the task might be in the middle of scheduling on another CPU.
3386 rq = task_rq_lock(p, &flags);
3388 * The RT priorities are set via sched_setscheduler(), but we still
3389 * allow the 'normal' nice value to be set - but as expected
3390 * it wont have any effect on scheduling until the task is
3394 p->static_prio = NICE_TO_PRIO(nice);
3399 dequeue_task(p, array);
3402 new_prio = NICE_TO_PRIO(nice);
3403 delta = new_prio - old_prio;
3404 p->static_prio = NICE_TO_PRIO(nice);
3408 enqueue_task(p, array);
3410 * If the task increased its priority or is running and
3411 * lowered its priority, then reschedule its CPU:
3413 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3414 resched_task(rq->curr);
3417 task_rq_unlock(rq, &flags);
3420 EXPORT_SYMBOL(set_user_nice);
3423 * can_nice - check if a task can reduce its nice value
3427 int can_nice(const task_t *p, const int nice)
3429 /* convert nice value [19,-20] to rlimit style value [1,40] */
3430 int nice_rlim = 20 - nice;
3431 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
3432 capable(CAP_SYS_NICE));
3435 #ifdef __ARCH_WANT_SYS_NICE
3438 * sys_nice - change the priority of the current process.
3439 * @increment: priority increment
3441 * sys_setpriority is a more generic, but much slower function that
3442 * does similar things.
3444 asmlinkage long sys_nice(int increment)
3450 * Setpriority might change our priority at the same moment.
3451 * We don't have to worry. Conceptually one call occurs first
3452 * and we have a single winner.
3454 if (increment < -40)
3459 nice = PRIO_TO_NICE(current->static_prio) + increment;
3465 if (increment < 0 && !can_nice(current, nice))
3468 retval = security_task_setnice(current, nice);
3472 set_user_nice(current, nice);
3479 * task_prio - return the priority value of a given task.
3480 * @p: the task in question.
3482 * This is the priority value as seen by users in /proc.
3483 * RT tasks are offset by -200. Normal tasks are centered
3484 * around 0, value goes from -16 to +15.
3486 int task_prio(const task_t *p)
3488 return p->prio - MAX_RT_PRIO;
3492 * task_nice - return the nice value of a given task.
3493 * @p: the task in question.
3495 int task_nice(const task_t *p)
3497 return TASK_NICE(p);
3499 EXPORT_SYMBOL_GPL(task_nice);
3502 * idle_cpu - is a given cpu idle currently?
3503 * @cpu: the processor in question.
3505 int idle_cpu(int cpu)
3507 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
3510 EXPORT_SYMBOL_GPL(idle_cpu);
3513 * idle_task - return the idle task for a given cpu.
3514 * @cpu: the processor in question.
3516 task_t *idle_task(int cpu)
3518 return cpu_rq(cpu)->idle;
3522 * find_process_by_pid - find a process with a matching PID value.
3523 * @pid: the pid in question.
3525 static inline task_t *find_process_by_pid(pid_t pid)
3527 return pid ? find_task_by_pid(pid) : current;
3530 /* Actually do priority change: must hold rq lock. */
3531 static void __setscheduler(struct task_struct *p, int policy, int prio)
3535 p->rt_priority = prio;
3536 if (policy != SCHED_NORMAL)
3537 p->prio = MAX_RT_PRIO-1 - p->rt_priority;
3539 p->prio = p->static_prio;
3543 * sched_setscheduler - change the scheduling policy and/or RT priority of
3545 * @p: the task in question.
3546 * @policy: new policy.
3547 * @param: structure containing the new RT priority.
3549 int sched_setscheduler(struct task_struct *p, int policy,
3550 struct sched_param *param)
3553 int oldprio, oldpolicy = -1;
3554 prio_array_t *array;
3555 unsigned long flags;
3559 /* double check policy once rq lock held */
3561 policy = oldpolicy = p->policy;
3562 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
3563 policy != SCHED_NORMAL)
3566 * Valid priorities for SCHED_FIFO and SCHED_RR are
3567 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL is 0.
3569 if (param->sched_priority < 0 ||
3570 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3571 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3573 if ((policy == SCHED_NORMAL) != (param->sched_priority == 0))
3577 * Allow unprivileged RT tasks to decrease priority:
3579 if (!capable(CAP_SYS_NICE)) {
3580 /* can't change policy */
3581 if (policy != p->policy &&
3582 !p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
3584 /* can't increase priority */
3585 if (policy != SCHED_NORMAL &&
3586 param->sched_priority > p->rt_priority &&
3587 param->sched_priority >
3588 p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
3590 /* can't change other user's priorities */
3591 if ((current->euid != p->euid) &&
3592 (current->euid != p->uid))
3596 retval = security_task_setscheduler(p, policy, param);
3600 * To be able to change p->policy safely, the apropriate
3601 * runqueue lock must be held.
3603 rq = task_rq_lock(p, &flags);
3604 /* recheck policy now with rq lock held */
3605 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3606 policy = oldpolicy = -1;
3607 task_rq_unlock(rq, &flags);
3612 deactivate_task(p, rq);
3614 __setscheduler(p, policy, param->sched_priority);
3616 __activate_task(p, rq);
3618 * Reschedule if we are currently running on this runqueue and
3619 * our priority decreased, or if we are not currently running on
3620 * this runqueue and our priority is higher than the current's
3622 if (task_running(rq, p)) {
3623 if (p->prio > oldprio)
3624 resched_task(rq->curr);
3625 } else if (TASK_PREEMPTS_CURR(p, rq))
3626 resched_task(rq->curr);
3628 task_rq_unlock(rq, &flags);
3631 EXPORT_SYMBOL_GPL(sched_setscheduler);
3634 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3637 struct sched_param lparam;
3638 struct task_struct *p;
3640 if (!param || pid < 0)
3642 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3644 read_lock_irq(&tasklist_lock);
3645 p = find_process_by_pid(pid);
3647 read_unlock_irq(&tasklist_lock);
3650 retval = sched_setscheduler(p, policy, &lparam);
3651 read_unlock_irq(&tasklist_lock);
3656 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3657 * @pid: the pid in question.
3658 * @policy: new policy.
3659 * @param: structure containing the new RT priority.
3661 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
3662 struct sched_param __user *param)
3664 return do_sched_setscheduler(pid, policy, param);
3668 * sys_sched_setparam - set/change the RT priority of a thread
3669 * @pid: the pid in question.
3670 * @param: structure containing the new RT priority.
3672 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
3674 return do_sched_setscheduler(pid, -1, param);
3678 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3679 * @pid: the pid in question.
3681 asmlinkage long sys_sched_getscheduler(pid_t pid)
3683 int retval = -EINVAL;
3690 read_lock(&tasklist_lock);
3691 p = find_process_by_pid(pid);
3693 retval = security_task_getscheduler(p);
3697 read_unlock(&tasklist_lock);
3704 * sys_sched_getscheduler - get the RT priority of a thread
3705 * @pid: the pid in question.
3706 * @param: structure containing the RT priority.
3708 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
3710 struct sched_param lp;
3711 int retval = -EINVAL;
3714 if (!param || pid < 0)
3717 read_lock(&tasklist_lock);
3718 p = find_process_by_pid(pid);
3723 retval = security_task_getscheduler(p);
3727 lp.sched_priority = p->rt_priority;
3728 read_unlock(&tasklist_lock);
3731 * This one might sleep, we cannot do it with a spinlock held ...
3733 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3739 read_unlock(&tasklist_lock);
3743 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
3747 cpumask_t cpus_allowed;
3750 read_lock(&tasklist_lock);
3752 p = find_process_by_pid(pid);
3754 read_unlock(&tasklist_lock);
3755 unlock_cpu_hotplug();
3760 * It is not safe to call set_cpus_allowed with the
3761 * tasklist_lock held. We will bump the task_struct's
3762 * usage count and then drop tasklist_lock.
3765 read_unlock(&tasklist_lock);
3768 if ((current->euid != p->euid) && (current->euid != p->uid) &&
3769 !capable(CAP_SYS_NICE))
3772 cpus_allowed = cpuset_cpus_allowed(p);
3773 cpus_and(new_mask, new_mask, cpus_allowed);
3774 retval = set_cpus_allowed(p, new_mask);
3778 unlock_cpu_hotplug();
3782 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3783 cpumask_t *new_mask)
3785 if (len < sizeof(cpumask_t)) {
3786 memset(new_mask, 0, sizeof(cpumask_t));
3787 } else if (len > sizeof(cpumask_t)) {
3788 len = sizeof(cpumask_t);
3790 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3794 * sys_sched_setaffinity - set the cpu affinity of a process
3795 * @pid: pid of the process
3796 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3797 * @user_mask_ptr: user-space pointer to the new cpu mask
3799 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
3800 unsigned long __user *user_mask_ptr)
3805 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
3809 return sched_setaffinity(pid, new_mask);
3813 * Represents all cpu's present in the system
3814 * In systems capable of hotplug, this map could dynamically grow
3815 * as new cpu's are detected in the system via any platform specific
3816 * method, such as ACPI for e.g.
3819 cpumask_t cpu_present_map;
3820 EXPORT_SYMBOL(cpu_present_map);
3823 cpumask_t cpu_online_map = CPU_MASK_ALL;
3824 cpumask_t cpu_possible_map = CPU_MASK_ALL;
3827 long sched_getaffinity(pid_t pid, cpumask_t *mask)
3833 read_lock(&tasklist_lock);
3836 p = find_process_by_pid(pid);
3841 cpus_and(*mask, p->cpus_allowed, cpu_possible_map);
3844 read_unlock(&tasklist_lock);
3845 unlock_cpu_hotplug();
3853 * sys_sched_getaffinity - get the cpu affinity of a process
3854 * @pid: pid of the process
3855 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3856 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3858 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
3859 unsigned long __user *user_mask_ptr)
3864 if (len < sizeof(cpumask_t))
3867 ret = sched_getaffinity(pid, &mask);
3871 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
3874 return sizeof(cpumask_t);
3878 * sys_sched_yield - yield the current processor to other threads.
3880 * this function yields the current CPU by moving the calling thread
3881 * to the expired array. If there are no other threads running on this
3882 * CPU then this function will return.
3884 asmlinkage long sys_sched_yield(void)
3886 runqueue_t *rq = this_rq_lock();
3887 prio_array_t *array = current->array;
3888 prio_array_t *target = rq->expired;
3890 schedstat_inc(rq, yld_cnt);
3892 * We implement yielding by moving the task into the expired
3895 * (special rule: RT tasks will just roundrobin in the active
3898 if (rt_task(current))
3899 target = rq->active;
3901 if (current->array->nr_active == 1) {
3902 schedstat_inc(rq, yld_act_empty);
3903 if (!rq->expired->nr_active)
3904 schedstat_inc(rq, yld_both_empty);
3905 } else if (!rq->expired->nr_active)
3906 schedstat_inc(rq, yld_exp_empty);
3908 if (array != target) {
3909 dequeue_task(current, array);
3910 enqueue_task(current, target);
3913 * requeue_task is cheaper so perform that if possible.
3915 requeue_task(current, array);
3918 * Since we are going to call schedule() anyway, there's
3919 * no need to preempt or enable interrupts:
3921 __release(rq->lock);
3922 _raw_spin_unlock(&rq->lock);
3923 preempt_enable_no_resched();
3930 static inline void __cond_resched(void)
3933 * The BKS might be reacquired before we have dropped
3934 * PREEMPT_ACTIVE, which could trigger a second
3935 * cond_resched() call.
3937 if (unlikely(preempt_count()))
3940 add_preempt_count(PREEMPT_ACTIVE);
3942 sub_preempt_count(PREEMPT_ACTIVE);
3943 } while (need_resched());
3946 int __sched cond_resched(void)
3948 if (need_resched()) {
3955 EXPORT_SYMBOL(cond_resched);
3958 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
3959 * call schedule, and on return reacquire the lock.
3961 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
3962 * operations here to prevent schedule() from being called twice (once via
3963 * spin_unlock(), once by hand).
3965 int cond_resched_lock(spinlock_t *lock)
3969 if (need_lockbreak(lock)) {
3975 if (need_resched()) {
3976 _raw_spin_unlock(lock);
3977 preempt_enable_no_resched();
3985 EXPORT_SYMBOL(cond_resched_lock);
3987 int __sched cond_resched_softirq(void)
3989 BUG_ON(!in_softirq());
3991 if (need_resched()) {
3992 __local_bh_enable();
4000 EXPORT_SYMBOL(cond_resched_softirq);
4004 * yield - yield the current processor to other threads.
4006 * this is a shortcut for kernel-space yielding - it marks the
4007 * thread runnable and calls sys_sched_yield().
4009 void __sched yield(void)
4011 set_current_state(TASK_RUNNING);
4015 EXPORT_SYMBOL(yield);
4018 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4019 * that process accounting knows that this is a task in IO wait state.
4021 * But don't do that if it is a deliberate, throttling IO wait (this task
4022 * has set its backing_dev_info: the queue against which it should throttle)
4024 void __sched io_schedule(void)
4026 struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id());
4028 atomic_inc(&rq->nr_iowait);
4030 atomic_dec(&rq->nr_iowait);
4033 EXPORT_SYMBOL(io_schedule);
4035 long __sched io_schedule_timeout(long timeout)
4037 struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id());
4040 atomic_inc(&rq->nr_iowait);
4041 ret = schedule_timeout(timeout);
4042 atomic_dec(&rq->nr_iowait);
4047 * sys_sched_get_priority_max - return maximum RT priority.
4048 * @policy: scheduling class.
4050 * this syscall returns the maximum rt_priority that can be used
4051 * by a given scheduling class.
4053 asmlinkage long sys_sched_get_priority_max(int policy)
4060 ret = MAX_USER_RT_PRIO-1;
4070 * sys_sched_get_priority_min - return minimum RT priority.
4071 * @policy: scheduling class.
4073 * this syscall returns the minimum rt_priority that can be used
4074 * by a given scheduling class.
4076 asmlinkage long sys_sched_get_priority_min(int policy)
4092 * sys_sched_rr_get_interval - return the default timeslice of a process.
4093 * @pid: pid of the process.
4094 * @interval: userspace pointer to the timeslice value.
4096 * this syscall writes the default timeslice value of a given process
4097 * into the user-space timespec buffer. A value of '0' means infinity.
4100 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4102 int retval = -EINVAL;
4110 read_lock(&tasklist_lock);
4111 p = find_process_by_pid(pid);
4115 retval = security_task_getscheduler(p);
4119 jiffies_to_timespec(p->policy & SCHED_FIFO ?
4120 0 : task_timeslice(p), &t);
4121 read_unlock(&tasklist_lock);
4122 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4126 read_unlock(&tasklist_lock);
4130 static inline struct task_struct *eldest_child(struct task_struct *p)
4132 if (list_empty(&p->children)) return NULL;
4133 return list_entry(p->children.next,struct task_struct,sibling);
4136 static inline struct task_struct *older_sibling(struct task_struct *p)
4138 if (p->sibling.prev==&p->parent->children) return NULL;
4139 return list_entry(p->sibling.prev,struct task_struct,sibling);
4142 static inline struct task_struct *younger_sibling(struct task_struct *p)
4144 if (p->sibling.next==&p->parent->children) return NULL;
4145 return list_entry(p->sibling.next,struct task_struct,sibling);
4148 static void show_task(task_t *p)
4152 unsigned long free = 0;
4153 static const char *stat_nam[] = { "R", "S", "D", "T", "t", "Z", "X" };
4155 printk("%-13.13s ", p->comm);
4156 state = p->state ? __ffs(p->state) + 1 : 0;
4157 if (state < ARRAY_SIZE(stat_nam))
4158 printk(stat_nam[state]);
4161 #if (BITS_PER_LONG == 32)
4162 if (state == TASK_RUNNING)
4163 printk(" running ");
4165 printk(" %08lX ", thread_saved_pc(p));
4167 if (state == TASK_RUNNING)
4168 printk(" running task ");
4170 printk(" %016lx ", thread_saved_pc(p));
4172 #ifdef CONFIG_DEBUG_STACK_USAGE
4174 unsigned long *n = (unsigned long *) (p->thread_info+1);
4177 free = (unsigned long) n - (unsigned long)(p->thread_info+1);
4180 printk("%5lu %5d %6d ", free, p->pid, p->parent->pid);
4181 if ((relative = eldest_child(p)))
4182 printk("%5d ", relative->pid);
4185 if ((relative = younger_sibling(p)))
4186 printk("%7d", relative->pid);
4189 if ((relative = older_sibling(p)))
4190 printk(" %5d", relative->pid);
4194 printk(" (L-TLB)\n");
4196 printk(" (NOTLB)\n");
4198 if (state != TASK_RUNNING)
4199 show_stack(p, NULL);
4202 void show_state(void)
4206 #if (BITS_PER_LONG == 32)
4209 printk(" task PC pid father child younger older\n");
4213 printk(" task PC pid father child younger older\n");
4215 read_lock(&tasklist_lock);
4216 do_each_thread(g, p) {
4218 * reset the NMI-timeout, listing all files on a slow
4219 * console might take alot of time:
4221 touch_nmi_watchdog();
4223 } while_each_thread(g, p);
4225 read_unlock(&tasklist_lock);
4229 * init_idle - set up an idle thread for a given CPU
4230 * @idle: task in question
4231 * @cpu: cpu the idle task belongs to
4233 * NOTE: this function does not set the idle thread's NEED_RESCHED
4234 * flag, to make booting more robust.
4236 void __devinit init_idle(task_t *idle, int cpu)
4238 runqueue_t *rq = cpu_rq(cpu);
4239 unsigned long flags;
4241 idle->sleep_avg = 0;
4243 idle->prio = MAX_PRIO;
4244 idle->state = TASK_RUNNING;
4245 idle->cpus_allowed = cpumask_of_cpu(cpu);
4246 set_task_cpu(idle, cpu);
4248 spin_lock_irqsave(&rq->lock, flags);
4249 rq->curr = rq->idle = idle;
4250 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4253 spin_unlock_irqrestore(&rq->lock, flags);
4255 /* Set the preempt count _outside_ the spinlocks! */
4256 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4257 idle->thread_info->preempt_count = (idle->lock_depth >= 0);
4259 idle->thread_info->preempt_count = 0;
4264 * In a system that switches off the HZ timer nohz_cpu_mask
4265 * indicates which cpus entered this state. This is used
4266 * in the rcu update to wait only for active cpus. For system
4267 * which do not switch off the HZ timer nohz_cpu_mask should
4268 * always be CPU_MASK_NONE.
4270 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4274 * This is how migration works:
4276 * 1) we queue a migration_req_t structure in the source CPU's
4277 * runqueue and wake up that CPU's migration thread.
4278 * 2) we down() the locked semaphore => thread blocks.
4279 * 3) migration thread wakes up (implicitly it forces the migrated
4280 * thread off the CPU)
4281 * 4) it gets the migration request and checks whether the migrated
4282 * task is still in the wrong runqueue.
4283 * 5) if it's in the wrong runqueue then the migration thread removes
4284 * it and puts it into the right queue.
4285 * 6) migration thread up()s the semaphore.
4286 * 7) we wake up and the migration is done.
4290 * Change a given task's CPU affinity. Migrate the thread to a
4291 * proper CPU and schedule it away if the CPU it's executing on
4292 * is removed from the allowed bitmask.
4294 * NOTE: the caller must have a valid reference to the task, the
4295 * task must not exit() & deallocate itself prematurely. The
4296 * call is not atomic; no spinlocks may be held.
4298 int set_cpus_allowed(task_t *p, cpumask_t new_mask)
4300 unsigned long flags;
4302 migration_req_t req;
4305 rq = task_rq_lock(p, &flags);
4306 if (!cpus_intersects(new_mask, cpu_online_map)) {
4311 p->cpus_allowed = new_mask;
4312 /* Can the task run on the task's current CPU? If so, we're done */
4313 if (cpu_isset(task_cpu(p), new_mask))
4316 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4317 /* Need help from migration thread: drop lock and wait. */
4318 task_rq_unlock(rq, &flags);
4319 wake_up_process(rq->migration_thread);
4320 wait_for_completion(&req.done);
4321 tlb_migrate_finish(p->mm);
4325 task_rq_unlock(rq, &flags);
4329 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4332 * Move (not current) task off this cpu, onto dest cpu. We're doing
4333 * this because either it can't run here any more (set_cpus_allowed()
4334 * away from this CPU, or CPU going down), or because we're
4335 * attempting to rebalance this task on exec (sched_exec).
4337 * So we race with normal scheduler movements, but that's OK, as long
4338 * as the task is no longer on this CPU.
4340 static void __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4342 runqueue_t *rq_dest, *rq_src;
4344 if (unlikely(cpu_is_offline(dest_cpu)))
4347 rq_src = cpu_rq(src_cpu);
4348 rq_dest = cpu_rq(dest_cpu);
4350 double_rq_lock(rq_src, rq_dest);
4351 /* Already moved. */
4352 if (task_cpu(p) != src_cpu)
4354 /* Affinity changed (again). */
4355 if (!cpu_isset(dest_cpu, p->cpus_allowed))
4358 set_task_cpu(p, dest_cpu);
4361 * Sync timestamp with rq_dest's before activating.
4362 * The same thing could be achieved by doing this step
4363 * afterwards, and pretending it was a local activate.
4364 * This way is cleaner and logically correct.
4366 p->timestamp = p->timestamp - rq_src->timestamp_last_tick
4367 + rq_dest->timestamp_last_tick;
4368 deactivate_task(p, rq_src);
4369 activate_task(p, rq_dest, 0);
4370 if (TASK_PREEMPTS_CURR(p, rq_dest))
4371 resched_task(rq_dest->curr);
4375 double_rq_unlock(rq_src, rq_dest);
4379 * migration_thread - this is a highprio system thread that performs
4380 * thread migration by bumping thread off CPU then 'pushing' onto
4383 static int migration_thread(void *data)
4386 int cpu = (long)data;
4389 BUG_ON(rq->migration_thread != current);
4391 set_current_state(TASK_INTERRUPTIBLE);
4392 while (!kthread_should_stop()) {
4393 struct list_head *head;
4394 migration_req_t *req;
4398 spin_lock_irq(&rq->lock);
4400 if (cpu_is_offline(cpu)) {
4401 spin_unlock_irq(&rq->lock);
4405 if (rq->active_balance) {
4406 active_load_balance(rq, cpu);
4407 rq->active_balance = 0;
4410 head = &rq->migration_queue;
4412 if (list_empty(head)) {
4413 spin_unlock_irq(&rq->lock);
4415 set_current_state(TASK_INTERRUPTIBLE);
4418 req = list_entry(head->next, migration_req_t, list);
4419 list_del_init(head->next);
4421 spin_unlock(&rq->lock);
4422 __migrate_task(req->task, cpu, req->dest_cpu);
4425 complete(&req->done);
4427 __set_current_state(TASK_RUNNING);
4431 /* Wait for kthread_stop */
4432 set_current_state(TASK_INTERRUPTIBLE);
4433 while (!kthread_should_stop()) {
4435 set_current_state(TASK_INTERRUPTIBLE);
4437 __set_current_state(TASK_RUNNING);
4441 #ifdef CONFIG_HOTPLUG_CPU
4442 /* Figure out where task on dead CPU should go, use force if neccessary. */
4443 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *tsk)
4449 mask = node_to_cpumask(cpu_to_node(dead_cpu));
4450 cpus_and(mask, mask, tsk->cpus_allowed);
4451 dest_cpu = any_online_cpu(mask);
4453 /* On any allowed CPU? */
4454 if (dest_cpu == NR_CPUS)
4455 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4457 /* No more Mr. Nice Guy. */
4458 if (dest_cpu == NR_CPUS) {
4459 cpus_setall(tsk->cpus_allowed);
4460 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4463 * Don't tell them about moving exiting tasks or
4464 * kernel threads (both mm NULL), since they never
4467 if (tsk->mm && printk_ratelimit())
4468 printk(KERN_INFO "process %d (%s) no "
4469 "longer affine to cpu%d\n",
4470 tsk->pid, tsk->comm, dead_cpu);
4472 __migrate_task(tsk, dead_cpu, dest_cpu);
4476 * While a dead CPU has no uninterruptible tasks queued at this point,
4477 * it might still have a nonzero ->nr_uninterruptible counter, because
4478 * for performance reasons the counter is not stricly tracking tasks to
4479 * their home CPUs. So we just add the counter to another CPU's counter,
4480 * to keep the global sum constant after CPU-down:
4482 static void migrate_nr_uninterruptible(runqueue_t *rq_src)
4484 runqueue_t *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
4485 unsigned long flags;
4487 local_irq_save(flags);
4488 double_rq_lock(rq_src, rq_dest);
4489 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
4490 rq_src->nr_uninterruptible = 0;
4491 double_rq_unlock(rq_src, rq_dest);
4492 local_irq_restore(flags);
4495 /* Run through task list and migrate tasks from the dead cpu. */
4496 static void migrate_live_tasks(int src_cpu)
4498 struct task_struct *tsk, *t;
4500 write_lock_irq(&tasklist_lock);
4502 do_each_thread(t, tsk) {
4506 if (task_cpu(tsk) == src_cpu)
4507 move_task_off_dead_cpu(src_cpu, tsk);
4508 } while_each_thread(t, tsk);
4510 write_unlock_irq(&tasklist_lock);
4513 /* Schedules idle task to be the next runnable task on current CPU.
4514 * It does so by boosting its priority to highest possible and adding it to
4515 * the _front_ of runqueue. Used by CPU offline code.
4517 void sched_idle_next(void)
4519 int cpu = smp_processor_id();
4520 runqueue_t *rq = this_rq();
4521 struct task_struct *p = rq->idle;
4522 unsigned long flags;
4524 /* cpu has to be offline */
4525 BUG_ON(cpu_online(cpu));
4527 /* Strictly not necessary since rest of the CPUs are stopped by now
4528 * and interrupts disabled on current cpu.
4530 spin_lock_irqsave(&rq->lock, flags);
4532 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4533 /* Add idle task to _front_ of it's priority queue */
4534 __activate_idle_task(p, rq);
4536 spin_unlock_irqrestore(&rq->lock, flags);
4539 /* Ensures that the idle task is using init_mm right before its cpu goes
4542 void idle_task_exit(void)
4544 struct mm_struct *mm = current->active_mm;
4546 BUG_ON(cpu_online(smp_processor_id()));
4549 switch_mm(mm, &init_mm, current);
4553 static void migrate_dead(unsigned int dead_cpu, task_t *tsk)
4555 struct runqueue *rq = cpu_rq(dead_cpu);
4557 /* Must be exiting, otherwise would be on tasklist. */
4558 BUG_ON(tsk->exit_state != EXIT_ZOMBIE && tsk->exit_state != EXIT_DEAD);
4560 /* Cannot have done final schedule yet: would have vanished. */
4561 BUG_ON(tsk->flags & PF_DEAD);
4563 get_task_struct(tsk);
4566 * Drop lock around migration; if someone else moves it,
4567 * that's OK. No task can be added to this CPU, so iteration is
4570 spin_unlock_irq(&rq->lock);
4571 move_task_off_dead_cpu(dead_cpu, tsk);
4572 spin_lock_irq(&rq->lock);
4574 put_task_struct(tsk);
4577 /* release_task() removes task from tasklist, so we won't find dead tasks. */
4578 static void migrate_dead_tasks(unsigned int dead_cpu)
4581 struct runqueue *rq = cpu_rq(dead_cpu);
4583 for (arr = 0; arr < 2; arr++) {
4584 for (i = 0; i < MAX_PRIO; i++) {
4585 struct list_head *list = &rq->arrays[arr].queue[i];
4586 while (!list_empty(list))
4587 migrate_dead(dead_cpu,
4588 list_entry(list->next, task_t,
4593 #endif /* CONFIG_HOTPLUG_CPU */
4596 * migration_call - callback that gets triggered when a CPU is added.
4597 * Here we can start up the necessary migration thread for the new CPU.
4599 static int migration_call(struct notifier_block *nfb, unsigned long action,
4602 int cpu = (long)hcpu;
4603 struct task_struct *p;
4604 struct runqueue *rq;
4605 unsigned long flags;
4608 case CPU_UP_PREPARE:
4609 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
4612 p->flags |= PF_NOFREEZE;
4613 kthread_bind(p, cpu);
4614 /* Must be high prio: stop_machine expects to yield to it. */
4615 rq = task_rq_lock(p, &flags);
4616 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4617 task_rq_unlock(rq, &flags);
4618 cpu_rq(cpu)->migration_thread = p;
4621 /* Strictly unneccessary, as first user will wake it. */
4622 wake_up_process(cpu_rq(cpu)->migration_thread);
4624 #ifdef CONFIG_HOTPLUG_CPU
4625 case CPU_UP_CANCELED:
4626 /* Unbind it from offline cpu so it can run. Fall thru. */
4627 kthread_bind(cpu_rq(cpu)->migration_thread,smp_processor_id());
4628 kthread_stop(cpu_rq(cpu)->migration_thread);
4629 cpu_rq(cpu)->migration_thread = NULL;
4632 migrate_live_tasks(cpu);
4634 kthread_stop(rq->migration_thread);
4635 rq->migration_thread = NULL;
4636 /* Idle task back to normal (off runqueue, low prio) */
4637 rq = task_rq_lock(rq->idle, &flags);
4638 deactivate_task(rq->idle, rq);
4639 rq->idle->static_prio = MAX_PRIO;
4640 __setscheduler(rq->idle, SCHED_NORMAL, 0);
4641 migrate_dead_tasks(cpu);
4642 task_rq_unlock(rq, &flags);
4643 migrate_nr_uninterruptible(rq);
4644 BUG_ON(rq->nr_running != 0);
4646 /* No need to migrate the tasks: it was best-effort if
4647 * they didn't do lock_cpu_hotplug(). Just wake up
4648 * the requestors. */
4649 spin_lock_irq(&rq->lock);
4650 while (!list_empty(&rq->migration_queue)) {
4651 migration_req_t *req;
4652 req = list_entry(rq->migration_queue.next,
4653 migration_req_t, list);
4654 list_del_init(&req->list);
4655 complete(&req->done);
4657 spin_unlock_irq(&rq->lock);
4664 /* Register at highest priority so that task migration (migrate_all_tasks)
4665 * happens before everything else.
4667 static struct notifier_block __devinitdata migration_notifier = {
4668 .notifier_call = migration_call,
4672 int __init migration_init(void)
4674 void *cpu = (void *)(long)smp_processor_id();
4675 /* Start one for boot CPU. */
4676 migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
4677 migration_call(&migration_notifier, CPU_ONLINE, cpu);
4678 register_cpu_notifier(&migration_notifier);
4684 #undef SCHED_DOMAIN_DEBUG
4685 #ifdef SCHED_DOMAIN_DEBUG
4686 static void sched_domain_debug(struct sched_domain *sd, int cpu)
4691 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
4695 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
4700 struct sched_group *group = sd->groups;
4701 cpumask_t groupmask;
4703 cpumask_scnprintf(str, NR_CPUS, sd->span);
4704 cpus_clear(groupmask);
4707 for (i = 0; i < level + 1; i++)
4709 printk("domain %d: ", level);
4711 if (!(sd->flags & SD_LOAD_BALANCE)) {
4712 printk("does not load-balance\n");
4714 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent");
4718 printk("span %s\n", str);
4720 if (!cpu_isset(cpu, sd->span))
4721 printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
4722 if (!cpu_isset(cpu, group->cpumask))
4723 printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
4726 for (i = 0; i < level + 2; i++)
4732 printk(KERN_ERR "ERROR: group is NULL\n");
4736 if (!group->cpu_power) {
4738 printk(KERN_ERR "ERROR: domain->cpu_power not set\n");
4741 if (!cpus_weight(group->cpumask)) {
4743 printk(KERN_ERR "ERROR: empty group\n");
4746 if (cpus_intersects(groupmask, group->cpumask)) {
4748 printk(KERN_ERR "ERROR: repeated CPUs\n");
4751 cpus_or(groupmask, groupmask, group->cpumask);
4753 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
4756 group = group->next;
4757 } while (group != sd->groups);
4760 if (!cpus_equal(sd->span, groupmask))
4761 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
4767 if (!cpus_subset(groupmask, sd->span))
4768 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
4774 #define sched_domain_debug(sd, cpu) {}
4777 static int sd_degenerate(struct sched_domain *sd)
4779 if (cpus_weight(sd->span) == 1)
4782 /* Following flags need at least 2 groups */
4783 if (sd->flags & (SD_LOAD_BALANCE |
4784 SD_BALANCE_NEWIDLE |
4787 if (sd->groups != sd->groups->next)
4791 /* Following flags don't use groups */
4792 if (sd->flags & (SD_WAKE_IDLE |
4800 static int sd_parent_degenerate(struct sched_domain *sd,
4801 struct sched_domain *parent)
4803 unsigned long cflags = sd->flags, pflags = parent->flags;
4805 if (sd_degenerate(parent))
4808 if (!cpus_equal(sd->span, parent->span))
4811 /* Does parent contain flags not in child? */
4812 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
4813 if (cflags & SD_WAKE_AFFINE)
4814 pflags &= ~SD_WAKE_BALANCE;
4815 /* Flags needing groups don't count if only 1 group in parent */
4816 if (parent->groups == parent->groups->next) {
4817 pflags &= ~(SD_LOAD_BALANCE |
4818 SD_BALANCE_NEWIDLE |
4822 if (~cflags & pflags)
4829 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
4830 * hold the hotplug lock.
4832 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
4834 runqueue_t *rq = cpu_rq(cpu);
4835 struct sched_domain *tmp;
4837 /* Remove the sched domains which do not contribute to scheduling. */
4838 for (tmp = sd; tmp; tmp = tmp->parent) {
4839 struct sched_domain *parent = tmp->parent;
4842 if (sd_parent_degenerate(tmp, parent))
4843 tmp->parent = parent->parent;
4846 if (sd && sd_degenerate(sd))
4849 sched_domain_debug(sd, cpu);
4851 rcu_assign_pointer(rq->sd, sd);
4854 /* cpus with isolated domains */
4855 static cpumask_t __devinitdata cpu_isolated_map = CPU_MASK_NONE;
4857 /* Setup the mask of cpus configured for isolated domains */
4858 static int __init isolated_cpu_setup(char *str)
4860 int ints[NR_CPUS], i;
4862 str = get_options(str, ARRAY_SIZE(ints), ints);
4863 cpus_clear(cpu_isolated_map);
4864 for (i = 1; i <= ints[0]; i++)
4865 if (ints[i] < NR_CPUS)
4866 cpu_set(ints[i], cpu_isolated_map);
4870 __setup ("isolcpus=", isolated_cpu_setup);
4873 * init_sched_build_groups takes an array of groups, the cpumask we wish
4874 * to span, and a pointer to a function which identifies what group a CPU
4875 * belongs to. The return value of group_fn must be a valid index into the
4876 * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
4877 * keep track of groups covered with a cpumask_t).
4879 * init_sched_build_groups will build a circular linked list of the groups
4880 * covered by the given span, and will set each group's ->cpumask correctly,
4881 * and ->cpu_power to 0.
4883 static void init_sched_build_groups(struct sched_group groups[], cpumask_t span,
4884 int (*group_fn)(int cpu))
4886 struct sched_group *first = NULL, *last = NULL;
4887 cpumask_t covered = CPU_MASK_NONE;
4890 for_each_cpu_mask(i, span) {
4891 int group = group_fn(i);
4892 struct sched_group *sg = &groups[group];
4895 if (cpu_isset(i, covered))
4898 sg->cpumask = CPU_MASK_NONE;
4901 for_each_cpu_mask(j, span) {
4902 if (group_fn(j) != group)
4905 cpu_set(j, covered);
4906 cpu_set(j, sg->cpumask);
4917 #define SD_NODES_PER_DOMAIN 16
4921 * find_next_best_node - find the next node to include in a sched_domain
4922 * @node: node whose sched_domain we're building
4923 * @used_nodes: nodes already in the sched_domain
4925 * Find the next node to include in a given scheduling domain. Simply
4926 * finds the closest node not already in the @used_nodes map.
4928 * Should use nodemask_t.
4930 static int find_next_best_node(int node, unsigned long *used_nodes)
4932 int i, n, val, min_val, best_node = 0;
4936 for (i = 0; i < MAX_NUMNODES; i++) {
4937 /* Start at @node */
4938 n = (node + i) % MAX_NUMNODES;
4940 if (!nr_cpus_node(n))
4943 /* Skip already used nodes */
4944 if (test_bit(n, used_nodes))
4947 /* Simple min distance search */
4948 val = node_distance(node, n);
4950 if (val < min_val) {
4956 set_bit(best_node, used_nodes);
4961 * sched_domain_node_span - get a cpumask for a node's sched_domain
4962 * @node: node whose cpumask we're constructing
4963 * @size: number of nodes to include in this span
4965 * Given a node, construct a good cpumask for its sched_domain to span. It
4966 * should be one that prevents unnecessary balancing, but also spreads tasks
4969 static cpumask_t sched_domain_node_span(int node)
4972 cpumask_t span, nodemask;
4973 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
4976 bitmap_zero(used_nodes, MAX_NUMNODES);
4978 nodemask = node_to_cpumask(node);
4979 cpus_or(span, span, nodemask);
4980 set_bit(node, used_nodes);
4982 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
4983 int next_node = find_next_best_node(node, used_nodes);
4984 nodemask = node_to_cpumask(next_node);
4985 cpus_or(span, span, nodemask);
4993 * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we
4994 * can switch it on easily if needed.
4996 #ifdef CONFIG_SCHED_SMT
4997 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
4998 static struct sched_group sched_group_cpus[NR_CPUS];
4999 static int cpu_to_cpu_group(int cpu)
5005 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5006 static struct sched_group sched_group_phys[NR_CPUS];
5007 static int cpu_to_phys_group(int cpu)
5009 #ifdef CONFIG_SCHED_SMT
5010 return first_cpu(cpu_sibling_map[cpu]);
5018 * The init_sched_build_groups can't handle what we want to do with node
5019 * groups, so roll our own. Now each node has its own list of groups which
5020 * gets dynamically allocated.
5022 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5023 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5025 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5026 static struct sched_group *sched_group_allnodes_bycpu[NR_CPUS];
5028 static int cpu_to_allnodes_group(int cpu)
5030 return cpu_to_node(cpu);
5035 * Build sched domains for a given set of cpus and attach the sched domains
5036 * to the individual cpus
5038 void build_sched_domains(const cpumask_t *cpu_map)
5042 struct sched_group **sched_group_nodes = NULL;
5043 struct sched_group *sched_group_allnodes = NULL;
5046 * Allocate the per-node list of sched groups
5048 sched_group_nodes = kmalloc(sizeof(struct sched_group*)*MAX_NUMNODES,
5050 if (!sched_group_nodes) {
5051 printk(KERN_WARNING "Can not alloc sched group node list\n");
5054 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
5058 * Set up domains for cpus specified by the cpu_map.
5060 for_each_cpu_mask(i, *cpu_map) {
5062 struct sched_domain *sd = NULL, *p;
5063 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
5065 cpus_and(nodemask, nodemask, *cpu_map);
5068 if (cpus_weight(*cpu_map)
5069 > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
5070 if (!sched_group_allnodes) {
5071 sched_group_allnodes
5072 = kmalloc(sizeof(struct sched_group)
5075 if (!sched_group_allnodes) {
5077 "Can not alloc allnodes sched group\n");
5080 sched_group_allnodes_bycpu[i]
5081 = sched_group_allnodes;
5083 sd = &per_cpu(allnodes_domains, i);
5084 *sd = SD_ALLNODES_INIT;
5085 sd->span = *cpu_map;
5086 group = cpu_to_allnodes_group(i);
5087 sd->groups = &sched_group_allnodes[group];
5092 sd = &per_cpu(node_domains, i);
5094 sd->span = sched_domain_node_span(cpu_to_node(i));
5096 cpus_and(sd->span, sd->span, *cpu_map);
5100 sd = &per_cpu(phys_domains, i);
5101 group = cpu_to_phys_group(i);
5103 sd->span = nodemask;
5105 sd->groups = &sched_group_phys[group];
5107 #ifdef CONFIG_SCHED_SMT
5109 sd = &per_cpu(cpu_domains, i);
5110 group = cpu_to_cpu_group(i);
5111 *sd = SD_SIBLING_INIT;
5112 sd->span = cpu_sibling_map[i];
5113 cpus_and(sd->span, sd->span, *cpu_map);
5115 sd->groups = &sched_group_cpus[group];
5119 #ifdef CONFIG_SCHED_SMT
5120 /* Set up CPU (sibling) groups */
5121 for_each_cpu_mask(i, *cpu_map) {
5122 cpumask_t this_sibling_map = cpu_sibling_map[i];
5123 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
5124 if (i != first_cpu(this_sibling_map))
5127 init_sched_build_groups(sched_group_cpus, this_sibling_map,
5132 /* Set up physical groups */
5133 for (i = 0; i < MAX_NUMNODES; i++) {
5134 cpumask_t nodemask = node_to_cpumask(i);
5136 cpus_and(nodemask, nodemask, *cpu_map);
5137 if (cpus_empty(nodemask))
5140 init_sched_build_groups(sched_group_phys, nodemask,
5141 &cpu_to_phys_group);
5145 /* Set up node groups */
5146 if (sched_group_allnodes)
5147 init_sched_build_groups(sched_group_allnodes, *cpu_map,
5148 &cpu_to_allnodes_group);
5150 for (i = 0; i < MAX_NUMNODES; i++) {
5151 /* Set up node groups */
5152 struct sched_group *sg, *prev;
5153 cpumask_t nodemask = node_to_cpumask(i);
5154 cpumask_t domainspan;
5155 cpumask_t covered = CPU_MASK_NONE;
5158 cpus_and(nodemask, nodemask, *cpu_map);
5159 if (cpus_empty(nodemask)) {
5160 sched_group_nodes[i] = NULL;
5164 domainspan = sched_domain_node_span(i);
5165 cpus_and(domainspan, domainspan, *cpu_map);
5167 sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL);
5168 sched_group_nodes[i] = sg;
5169 for_each_cpu_mask(j, nodemask) {
5170 struct sched_domain *sd;
5171 sd = &per_cpu(node_domains, j);
5173 if (sd->groups == NULL) {
5174 /* Turn off balancing if we have no groups */
5180 "Can not alloc domain group for node %d\n", i);
5184 sg->cpumask = nodemask;
5185 cpus_or(covered, covered, nodemask);
5188 for (j = 0; j < MAX_NUMNODES; j++) {
5189 cpumask_t tmp, notcovered;
5190 int n = (i + j) % MAX_NUMNODES;
5192 cpus_complement(notcovered, covered);
5193 cpus_and(tmp, notcovered, *cpu_map);
5194 cpus_and(tmp, tmp, domainspan);
5195 if (cpus_empty(tmp))
5198 nodemask = node_to_cpumask(n);
5199 cpus_and(tmp, tmp, nodemask);
5200 if (cpus_empty(tmp))
5203 sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL);
5206 "Can not alloc domain group for node %d\n", j);
5211 cpus_or(covered, covered, tmp);
5215 prev->next = sched_group_nodes[i];
5219 /* Calculate CPU power for physical packages and nodes */
5220 for_each_cpu_mask(i, *cpu_map) {
5222 struct sched_domain *sd;
5223 #ifdef CONFIG_SCHED_SMT
5224 sd = &per_cpu(cpu_domains, i);
5225 power = SCHED_LOAD_SCALE;
5226 sd->groups->cpu_power = power;
5229 sd = &per_cpu(phys_domains, i);
5230 power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
5231 (cpus_weight(sd->groups->cpumask)-1) / 10;
5232 sd->groups->cpu_power = power;
5235 sd = &per_cpu(allnodes_domains, i);
5237 power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
5238 (cpus_weight(sd->groups->cpumask)-1) / 10;
5239 sd->groups->cpu_power = power;
5245 for (i = 0; i < MAX_NUMNODES; i++) {
5246 struct sched_group *sg = sched_group_nodes[i];
5252 for_each_cpu_mask(j, sg->cpumask) {
5253 struct sched_domain *sd;
5256 sd = &per_cpu(phys_domains, j);
5257 if (j != first_cpu(sd->groups->cpumask)) {
5259 * Only add "power" once for each
5264 power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
5265 (cpus_weight(sd->groups->cpumask)-1) / 10;
5267 sg->cpu_power += power;
5270 if (sg != sched_group_nodes[i])
5275 /* Attach the domains */
5276 for_each_cpu_mask(i, *cpu_map) {
5277 struct sched_domain *sd;
5278 #ifdef CONFIG_SCHED_SMT
5279 sd = &per_cpu(cpu_domains, i);
5281 sd = &per_cpu(phys_domains, i);
5283 cpu_attach_domain(sd, i);
5287 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
5289 static void arch_init_sched_domains(const cpumask_t *cpu_map)
5291 cpumask_t cpu_default_map;
5294 * Setup mask for cpus without special case scheduling requirements.
5295 * For now this just excludes isolated cpus, but could be used to
5296 * exclude other special cases in the future.
5298 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
5300 build_sched_domains(&cpu_default_map);
5303 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
5309 for_each_cpu_mask(cpu, *cpu_map) {
5310 struct sched_group *sched_group_allnodes
5311 = sched_group_allnodes_bycpu[cpu];
5312 struct sched_group **sched_group_nodes
5313 = sched_group_nodes_bycpu[cpu];
5315 if (sched_group_allnodes) {
5316 kfree(sched_group_allnodes);
5317 sched_group_allnodes_bycpu[cpu] = NULL;
5320 if (!sched_group_nodes)
5323 for (i = 0; i < MAX_NUMNODES; i++) {
5324 cpumask_t nodemask = node_to_cpumask(i);
5325 struct sched_group *oldsg, *sg = sched_group_nodes[i];
5327 cpus_and(nodemask, nodemask, *cpu_map);
5328 if (cpus_empty(nodemask))
5338 if (oldsg != sched_group_nodes[i])
5341 kfree(sched_group_nodes);
5342 sched_group_nodes_bycpu[cpu] = NULL;
5348 * Detach sched domains from a group of cpus specified in cpu_map
5349 * These cpus will now be attached to the NULL domain
5351 static inline void detach_destroy_domains(const cpumask_t *cpu_map)
5355 for_each_cpu_mask(i, *cpu_map)
5356 cpu_attach_domain(NULL, i);
5357 synchronize_sched();
5358 arch_destroy_sched_domains(cpu_map);
5362 * Partition sched domains as specified by the cpumasks below.
5363 * This attaches all cpus from the cpumasks to the NULL domain,
5364 * waits for a RCU quiescent period, recalculates sched
5365 * domain information and then attaches them back to the
5366 * correct sched domains
5367 * Call with hotplug lock held
5369 void partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
5371 cpumask_t change_map;
5373 cpus_and(*partition1, *partition1, cpu_online_map);
5374 cpus_and(*partition2, *partition2, cpu_online_map);
5375 cpus_or(change_map, *partition1, *partition2);
5377 /* Detach sched domains from all of the affected cpus */
5378 detach_destroy_domains(&change_map);
5379 if (!cpus_empty(*partition1))
5380 build_sched_domains(partition1);
5381 if (!cpus_empty(*partition2))
5382 build_sched_domains(partition2);
5385 #ifdef CONFIG_HOTPLUG_CPU
5387 * Force a reinitialization of the sched domains hierarchy. The domains
5388 * and groups cannot be updated in place without racing with the balancing
5389 * code, so we temporarily attach all running cpus to the NULL domain
5390 * which will prevent rebalancing while the sched domains are recalculated.
5392 static int update_sched_domains(struct notifier_block *nfb,
5393 unsigned long action, void *hcpu)
5396 case CPU_UP_PREPARE:
5397 case CPU_DOWN_PREPARE:
5398 detach_destroy_domains(&cpu_online_map);
5401 case CPU_UP_CANCELED:
5402 case CPU_DOWN_FAILED:
5406 * Fall through and re-initialise the domains.
5413 /* The hotplug lock is already held by cpu_up/cpu_down */
5414 arch_init_sched_domains(&cpu_online_map);
5420 void __init sched_init_smp(void)
5423 arch_init_sched_domains(&cpu_online_map);
5424 unlock_cpu_hotplug();
5425 /* XXX: Theoretical race here - CPU may be hotplugged now */
5426 hotcpu_notifier(update_sched_domains, 0);
5429 void __init sched_init_smp(void)
5432 #endif /* CONFIG_SMP */
5434 int in_sched_functions(unsigned long addr)
5436 /* Linker adds these: start and end of __sched functions */
5437 extern char __sched_text_start[], __sched_text_end[];
5438 return in_lock_functions(addr) ||
5439 (addr >= (unsigned long)__sched_text_start
5440 && addr < (unsigned long)__sched_text_end);
5443 void __init sched_init(void)
5448 for (i = 0; i < NR_CPUS; i++) {
5449 prio_array_t *array;
5452 spin_lock_init(&rq->lock);
5454 rq->active = rq->arrays;
5455 rq->expired = rq->arrays + 1;
5456 rq->best_expired_prio = MAX_PRIO;
5460 for (j = 1; j < 3; j++)
5461 rq->cpu_load[j] = 0;
5462 rq->active_balance = 0;
5464 rq->migration_thread = NULL;
5465 INIT_LIST_HEAD(&rq->migration_queue);
5467 atomic_set(&rq->nr_iowait, 0);
5469 for (j = 0; j < 2; j++) {
5470 array = rq->arrays + j;
5471 for (k = 0; k < MAX_PRIO; k++) {
5472 INIT_LIST_HEAD(array->queue + k);
5473 __clear_bit(k, array->bitmap);
5475 // delimiter for bitsearch
5476 __set_bit(MAX_PRIO, array->bitmap);
5481 * The boot idle thread does lazy MMU switching as well:
5483 atomic_inc(&init_mm.mm_count);
5484 enter_lazy_tlb(&init_mm, current);
5487 * Make us the idle thread. Technically, schedule() should not be
5488 * called from this thread, however somewhere below it might be,
5489 * but because we are the idle thread, we just pick up running again
5490 * when this runqueue becomes "idle".
5492 init_idle(current, smp_processor_id());
5495 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5496 void __might_sleep(char *file, int line)
5498 #if defined(in_atomic)
5499 static unsigned long prev_jiffy; /* ratelimiting */
5501 if ((in_atomic() || irqs_disabled()) &&
5502 system_state == SYSTEM_RUNNING && !oops_in_progress) {
5503 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
5505 prev_jiffy = jiffies;
5506 printk(KERN_ERR "Debug: sleeping function called from invalid"
5507 " context at %s:%d\n", file, line);
5508 printk("in_atomic():%d, irqs_disabled():%d\n",
5509 in_atomic(), irqs_disabled());
5514 EXPORT_SYMBOL(__might_sleep);
5517 #ifdef CONFIG_MAGIC_SYSRQ
5518 void normalize_rt_tasks(void)
5520 struct task_struct *p;
5521 prio_array_t *array;
5522 unsigned long flags;
5525 read_lock_irq(&tasklist_lock);
5526 for_each_process (p) {
5530 rq = task_rq_lock(p, &flags);
5534 deactivate_task(p, task_rq(p));
5535 __setscheduler(p, SCHED_NORMAL, 0);
5537 __activate_task(p, task_rq(p));
5538 resched_task(rq->curr);
5541 task_rq_unlock(rq, &flags);
5543 read_unlock_irq(&tasklist_lock);
5546 #endif /* CONFIG_MAGIC_SYSRQ */