4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
22 #include <linux/module.h>
23 #include <linux/nmi.h>
24 #include <linux/init.h>
25 #include <asm/uaccess.h>
26 #include <linux/highmem.h>
27 #include <linux/smp_lock.h>
28 #include <asm/mmu_context.h>
29 #include <linux/interrupt.h>
30 #include <linux/capability.h>
31 #include <linux/completion.h>
32 #include <linux/kernel_stat.h>
33 #include <linux/security.h>
34 #include <linux/notifier.h>
35 #include <linux/profile.h>
36 #include <linux/suspend.h>
37 #include <linux/vmalloc.h>
38 #include <linux/blkdev.h>
39 #include <linux/delay.h>
40 #include <linux/smp.h>
41 #include <linux/threads.h>
42 #include <linux/timer.h>
43 #include <linux/rcupdate.h>
44 #include <linux/cpu.h>
45 #include <linux/cpuset.h>
46 #include <linux/percpu.h>
47 #include <linux/kthread.h>
48 #include <linux/seq_file.h>
49 #include <linux/syscalls.h>
50 #include <linux/times.h>
51 #include <linux/acct.h>
52 #include <linux/kprobes.h>
55 #include <asm/unistd.h>
58 * Convert user-nice values [ -20 ... 0 ... 19 ]
59 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
62 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
63 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
64 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
67 * 'User priority' is the nice value converted to something we
68 * can work with better when scaling various scheduler parameters,
69 * it's a [ 0 ... 39 ] range.
71 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
72 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
73 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
76 * Some helpers for converting nanosecond timing to jiffy resolution
78 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
79 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
82 * These are the 'tuning knobs' of the scheduler:
84 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
85 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
86 * Timeslices get refilled after they expire.
88 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
89 #define DEF_TIMESLICE (100 * HZ / 1000)
90 #define ON_RUNQUEUE_WEIGHT 30
91 #define CHILD_PENALTY 95
92 #define PARENT_PENALTY 100
94 #define PRIO_BONUS_RATIO 25
95 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
96 #define INTERACTIVE_DELTA 2
97 #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
98 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
99 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
102 * If a task is 'interactive' then we reinsert it in the active
103 * array after it has expired its current timeslice. (it will not
104 * continue to run immediately, it will still roundrobin with
105 * other interactive tasks.)
107 * This part scales the interactivity limit depending on niceness.
109 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
110 * Here are a few examples of different nice levels:
112 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
113 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
114 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
115 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
116 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
118 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
119 * priority range a task can explore, a value of '1' means the
120 * task is rated interactive.)
122 * Ie. nice +19 tasks can never get 'interactive' enough to be
123 * reinserted into the active array. And only heavily CPU-hog nice -20
124 * tasks will be expired. Default nice 0 tasks are somewhere between,
125 * it takes some effort for them to get interactive, but it's not
129 #define CURRENT_BONUS(p) \
130 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
133 #define GRANULARITY (10 * HZ / 1000 ? : 1)
136 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
137 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
140 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
141 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
144 #define SCALE(v1,v1_max,v2_max) \
145 (v1) * (v2_max) / (v1_max)
148 (SCALE(TASK_NICE(p) + 20, 40, MAX_BONUS) - 20 * MAX_BONUS / 40 + \
151 #define TASK_INTERACTIVE(p) \
152 ((p)->prio <= (p)->static_prio - DELTA(p))
154 #define INTERACTIVE_SLEEP(p) \
155 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
156 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
158 #define TASK_PREEMPTS_CURR(p, rq) \
159 ((p)->prio < (rq)->curr->prio)
162 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
163 * to time slice values: [800ms ... 100ms ... 5ms]
165 * The higher a thread's priority, the bigger timeslices
166 * it gets during one round of execution. But even the lowest
167 * priority thread gets MIN_TIMESLICE worth of execution time.
170 #define SCALE_PRIO(x, prio) \
171 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO/2), MIN_TIMESLICE)
173 static unsigned int task_timeslice(task_t *p)
175 if (p->static_prio < NICE_TO_PRIO(0))
176 return SCALE_PRIO(DEF_TIMESLICE*4, p->static_prio);
178 return SCALE_PRIO(DEF_TIMESLICE, p->static_prio);
180 #define task_hot(p, now, sd) ((long long) ((now) - (p)->last_ran) \
181 < (long long) (sd)->cache_hot_time)
184 * These are the runqueue data structures:
187 #define BITMAP_SIZE ((((MAX_PRIO+1+7)/8)+sizeof(long)-1)/sizeof(long))
189 typedef struct runqueue runqueue_t;
192 unsigned int nr_active;
193 unsigned long bitmap[BITMAP_SIZE];
194 struct list_head queue[MAX_PRIO];
198 * This is the main, per-CPU runqueue data structure.
200 * Locking rule: those places that want to lock multiple runqueues
201 * (such as the load balancing or the thread migration code), lock
202 * acquire operations must be ordered by ascending &runqueue.
208 * nr_running and cpu_load should be in the same cacheline because
209 * remote CPUs use both these fields when doing load calculation.
211 unsigned long nr_running;
213 unsigned long cpu_load[3];
215 unsigned long long nr_switches;
218 * This is part of a global counter where only the total sum
219 * over all CPUs matters. A task can increase this counter on
220 * one CPU and if it got migrated afterwards it may decrease
221 * it on another CPU. Always updated under the runqueue lock:
223 unsigned long nr_uninterruptible;
225 unsigned long expired_timestamp;
226 unsigned long long timestamp_last_tick;
228 struct mm_struct *prev_mm;
229 prio_array_t *active, *expired, arrays[2];
230 int best_expired_prio;
234 struct sched_domain *sd;
236 /* For active balancing */
240 task_t *migration_thread;
241 struct list_head migration_queue;
245 #ifdef CONFIG_SCHEDSTATS
247 struct sched_info rq_sched_info;
249 /* sys_sched_yield() stats */
250 unsigned long yld_exp_empty;
251 unsigned long yld_act_empty;
252 unsigned long yld_both_empty;
253 unsigned long yld_cnt;
255 /* schedule() stats */
256 unsigned long sched_switch;
257 unsigned long sched_cnt;
258 unsigned long sched_goidle;
260 /* try_to_wake_up() stats */
261 unsigned long ttwu_cnt;
262 unsigned long ttwu_local;
266 static DEFINE_PER_CPU(struct runqueue, runqueues);
269 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
270 * See detach_destroy_domains: synchronize_sched for details.
272 * The domain tree of any CPU may only be accessed from within
273 * preempt-disabled sections.
275 #define for_each_domain(cpu, domain) \
276 for (domain = rcu_dereference(cpu_rq(cpu)->sd); domain; domain = domain->parent)
278 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
279 #define this_rq() (&__get_cpu_var(runqueues))
280 #define task_rq(p) cpu_rq(task_cpu(p))
281 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
283 #ifndef prepare_arch_switch
284 # define prepare_arch_switch(next) do { } while (0)
286 #ifndef finish_arch_switch
287 # define finish_arch_switch(prev) do { } while (0)
290 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
291 static inline int task_running(runqueue_t *rq, task_t *p)
293 return rq->curr == p;
296 static inline void prepare_lock_switch(runqueue_t *rq, task_t *next)
300 static inline void finish_lock_switch(runqueue_t *rq, task_t *prev)
302 #ifdef CONFIG_DEBUG_SPINLOCK
303 /* this is a valid case when another task releases the spinlock */
304 rq->lock.owner = current;
306 spin_unlock_irq(&rq->lock);
309 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
310 static inline int task_running(runqueue_t *rq, task_t *p)
315 return rq->curr == p;
319 static inline void prepare_lock_switch(runqueue_t *rq, task_t *next)
323 * We can optimise this out completely for !SMP, because the
324 * SMP rebalancing from interrupt is the only thing that cares
329 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
330 spin_unlock_irq(&rq->lock);
332 spin_unlock(&rq->lock);
336 static inline void finish_lock_switch(runqueue_t *rq, task_t *prev)
340 * After ->oncpu is cleared, the task can be moved to a different CPU.
341 * We must ensure this doesn't happen until the switch is completely
347 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
351 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
354 * task_rq_lock - lock the runqueue a given task resides on and disable
355 * interrupts. Note the ordering: we can safely lookup the task_rq without
356 * explicitly disabling preemption.
358 static inline runqueue_t *task_rq_lock(task_t *p, unsigned long *flags)
364 local_irq_save(*flags);
366 spin_lock(&rq->lock);
367 if (unlikely(rq != task_rq(p))) {
368 spin_unlock_irqrestore(&rq->lock, *flags);
369 goto repeat_lock_task;
374 static inline void task_rq_unlock(runqueue_t *rq, unsigned long *flags)
377 spin_unlock_irqrestore(&rq->lock, *flags);
380 #ifdef CONFIG_SCHEDSTATS
382 * bump this up when changing the output format or the meaning of an existing
383 * format, so that tools can adapt (or abort)
385 #define SCHEDSTAT_VERSION 12
387 static int show_schedstat(struct seq_file *seq, void *v)
391 seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION);
392 seq_printf(seq, "timestamp %lu\n", jiffies);
393 for_each_online_cpu(cpu) {
394 runqueue_t *rq = cpu_rq(cpu);
396 struct sched_domain *sd;
400 /* runqueue-specific stats */
402 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
403 cpu, rq->yld_both_empty,
404 rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt,
405 rq->sched_switch, rq->sched_cnt, rq->sched_goidle,
406 rq->ttwu_cnt, rq->ttwu_local,
407 rq->rq_sched_info.cpu_time,
408 rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt);
410 seq_printf(seq, "\n");
413 /* domain-specific stats */
415 for_each_domain(cpu, sd) {
416 enum idle_type itype;
417 char mask_str[NR_CPUS];
419 cpumask_scnprintf(mask_str, NR_CPUS, sd->span);
420 seq_printf(seq, "domain%d %s", dcnt++, mask_str);
421 for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES;
423 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu",
425 sd->lb_balanced[itype],
426 sd->lb_failed[itype],
427 sd->lb_imbalance[itype],
428 sd->lb_gained[itype],
429 sd->lb_hot_gained[itype],
430 sd->lb_nobusyq[itype],
431 sd->lb_nobusyg[itype]);
433 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
434 sd->alb_cnt, sd->alb_failed, sd->alb_pushed,
435 sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed,
436 sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed,
437 sd->ttwu_wake_remote, sd->ttwu_move_affine, sd->ttwu_move_balance);
445 static int schedstat_open(struct inode *inode, struct file *file)
447 unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32);
448 char *buf = kmalloc(size, GFP_KERNEL);
454 res = single_open(file, show_schedstat, NULL);
456 m = file->private_data;
464 struct file_operations proc_schedstat_operations = {
465 .open = schedstat_open,
468 .release = single_release,
471 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
472 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
473 #else /* !CONFIG_SCHEDSTATS */
474 # define schedstat_inc(rq, field) do { } while (0)
475 # define schedstat_add(rq, field, amt) do { } while (0)
479 * rq_lock - lock a given runqueue and disable interrupts.
481 static inline runqueue_t *this_rq_lock(void)
488 spin_lock(&rq->lock);
493 #ifdef CONFIG_SCHEDSTATS
495 * Called when a process is dequeued from the active array and given
496 * the cpu. We should note that with the exception of interactive
497 * tasks, the expired queue will become the active queue after the active
498 * queue is empty, without explicitly dequeuing and requeuing tasks in the
499 * expired queue. (Interactive tasks may be requeued directly to the
500 * active queue, thus delaying tasks in the expired queue from running;
501 * see scheduler_tick()).
503 * This function is only called from sched_info_arrive(), rather than
504 * dequeue_task(). Even though a task may be queued and dequeued multiple
505 * times as it is shuffled about, we're really interested in knowing how
506 * long it was from the *first* time it was queued to the time that it
509 static inline void sched_info_dequeued(task_t *t)
511 t->sched_info.last_queued = 0;
515 * Called when a task finally hits the cpu. We can now calculate how
516 * long it was waiting to run. We also note when it began so that we
517 * can keep stats on how long its timeslice is.
519 static void sched_info_arrive(task_t *t)
521 unsigned long now = jiffies, diff = 0;
522 struct runqueue *rq = task_rq(t);
524 if (t->sched_info.last_queued)
525 diff = now - t->sched_info.last_queued;
526 sched_info_dequeued(t);
527 t->sched_info.run_delay += diff;
528 t->sched_info.last_arrival = now;
529 t->sched_info.pcnt++;
534 rq->rq_sched_info.run_delay += diff;
535 rq->rq_sched_info.pcnt++;
539 * Called when a process is queued into either the active or expired
540 * array. The time is noted and later used to determine how long we
541 * had to wait for us to reach the cpu. Since the expired queue will
542 * become the active queue after active queue is empty, without dequeuing
543 * and requeuing any tasks, we are interested in queuing to either. It
544 * is unusual but not impossible for tasks to be dequeued and immediately
545 * requeued in the same or another array: this can happen in sched_yield(),
546 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
549 * This function is only called from enqueue_task(), but also only updates
550 * the timestamp if it is already not set. It's assumed that
551 * sched_info_dequeued() will clear that stamp when appropriate.
553 static inline void sched_info_queued(task_t *t)
555 if (!t->sched_info.last_queued)
556 t->sched_info.last_queued = jiffies;
560 * Called when a process ceases being the active-running process, either
561 * voluntarily or involuntarily. Now we can calculate how long we ran.
563 static inline void sched_info_depart(task_t *t)
565 struct runqueue *rq = task_rq(t);
566 unsigned long diff = jiffies - t->sched_info.last_arrival;
568 t->sched_info.cpu_time += diff;
571 rq->rq_sched_info.cpu_time += diff;
575 * Called when tasks are switched involuntarily due, typically, to expiring
576 * their time slice. (This may also be called when switching to or from
577 * the idle task.) We are only called when prev != next.
579 static inline void sched_info_switch(task_t *prev, task_t *next)
581 struct runqueue *rq = task_rq(prev);
584 * prev now departs the cpu. It's not interesting to record
585 * stats about how efficient we were at scheduling the idle
588 if (prev != rq->idle)
589 sched_info_depart(prev);
591 if (next != rq->idle)
592 sched_info_arrive(next);
595 #define sched_info_queued(t) do { } while (0)
596 #define sched_info_switch(t, next) do { } while (0)
597 #endif /* CONFIG_SCHEDSTATS */
600 * Adding/removing a task to/from a priority array:
602 static void dequeue_task(struct task_struct *p, prio_array_t *array)
605 list_del(&p->run_list);
606 if (list_empty(array->queue + p->prio))
607 __clear_bit(p->prio, array->bitmap);
610 static void enqueue_task(struct task_struct *p, prio_array_t *array)
612 sched_info_queued(p);
613 list_add_tail(&p->run_list, array->queue + p->prio);
614 __set_bit(p->prio, array->bitmap);
620 * Put task to the end of the run list without the overhead of dequeue
621 * followed by enqueue.
623 static void requeue_task(struct task_struct *p, prio_array_t *array)
625 list_move_tail(&p->run_list, array->queue + p->prio);
628 static inline void enqueue_task_head(struct task_struct *p, prio_array_t *array)
630 list_add(&p->run_list, array->queue + p->prio);
631 __set_bit(p->prio, array->bitmap);
637 * effective_prio - return the priority that is based on the static
638 * priority but is modified by bonuses/penalties.
640 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
641 * into the -5 ... 0 ... +5 bonus/penalty range.
643 * We use 25% of the full 0...39 priority range so that:
645 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
646 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
648 * Both properties are important to certain workloads.
650 static int effective_prio(task_t *p)
657 bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
659 prio = p->static_prio - bonus;
660 if (prio < MAX_RT_PRIO)
662 if (prio > MAX_PRIO-1)
668 * __activate_task - move a task to the runqueue.
670 static inline void __activate_task(task_t *p, runqueue_t *rq)
672 enqueue_task(p, rq->active);
677 * __activate_idle_task - move idle task to the _front_ of runqueue.
679 static inline void __activate_idle_task(task_t *p, runqueue_t *rq)
681 enqueue_task_head(p, rq->active);
685 static int recalc_task_prio(task_t *p, unsigned long long now)
687 /* Caller must always ensure 'now >= p->timestamp' */
688 unsigned long long __sleep_time = now - p->timestamp;
689 unsigned long sleep_time;
691 if (unlikely(p->policy == SCHED_BATCH))
694 if (__sleep_time > NS_MAX_SLEEP_AVG)
695 sleep_time = NS_MAX_SLEEP_AVG;
697 sleep_time = (unsigned long)__sleep_time;
700 if (likely(sleep_time > 0)) {
702 * User tasks that sleep a long time are categorised as
703 * idle and will get just interactive status to stay active &
704 * prevent them suddenly becoming cpu hogs and starving
707 if (p->mm && p->activated != -1 &&
708 sleep_time > INTERACTIVE_SLEEP(p)) {
709 p->sleep_avg = JIFFIES_TO_NS(MAX_SLEEP_AVG -
713 * Tasks waking from uninterruptible sleep are
714 * limited in their sleep_avg rise as they
715 * are likely to be waiting on I/O
717 if (p->activated == -1 && p->mm) {
718 if (p->sleep_avg >= INTERACTIVE_SLEEP(p))
720 else if (p->sleep_avg + sleep_time >=
721 INTERACTIVE_SLEEP(p)) {
722 p->sleep_avg = INTERACTIVE_SLEEP(p);
728 * This code gives a bonus to interactive tasks.
730 * The boost works by updating the 'average sleep time'
731 * value here, based on ->timestamp. The more time a
732 * task spends sleeping, the higher the average gets -
733 * and the higher the priority boost gets as well.
735 p->sleep_avg += sleep_time;
737 if (p->sleep_avg > NS_MAX_SLEEP_AVG)
738 p->sleep_avg = NS_MAX_SLEEP_AVG;
742 return effective_prio(p);
746 * activate_task - move a task to the runqueue and do priority recalculation
748 * Update all the scheduling statistics stuff. (sleep average
749 * calculation, priority modifiers, etc.)
751 static void activate_task(task_t *p, runqueue_t *rq, int local)
753 unsigned long long now;
758 /* Compensate for drifting sched_clock */
759 runqueue_t *this_rq = this_rq();
760 now = (now - this_rq->timestamp_last_tick)
761 + rq->timestamp_last_tick;
766 p->prio = recalc_task_prio(p, now);
769 * This checks to make sure it's not an uninterruptible task
770 * that is now waking up.
774 * Tasks which were woken up by interrupts (ie. hw events)
775 * are most likely of interactive nature. So we give them
776 * the credit of extending their sleep time to the period
777 * of time they spend on the runqueue, waiting for execution
778 * on a CPU, first time around:
784 * Normal first-time wakeups get a credit too for
785 * on-runqueue time, but it will be weighted down:
792 __activate_task(p, rq);
796 * deactivate_task - remove a task from the runqueue.
798 static void deactivate_task(struct task_struct *p, runqueue_t *rq)
801 dequeue_task(p, p->array);
806 * resched_task - mark a task 'to be rescheduled now'.
808 * On UP this means the setting of the need_resched flag, on SMP it
809 * might also involve a cross-CPU call to trigger the scheduler on
813 static void resched_task(task_t *p)
817 assert_spin_locked(&task_rq(p)->lock);
819 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
822 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
825 if (cpu == smp_processor_id())
828 /* NEED_RESCHED must be visible before we test POLLING_NRFLAG */
830 if (!test_tsk_thread_flag(p, TIF_POLLING_NRFLAG))
831 smp_send_reschedule(cpu);
834 static inline void resched_task(task_t *p)
836 assert_spin_locked(&task_rq(p)->lock);
837 set_tsk_need_resched(p);
842 * task_curr - is this task currently executing on a CPU?
843 * @p: the task in question.
845 inline int task_curr(const task_t *p)
847 return cpu_curr(task_cpu(p)) == p;
852 struct list_head list;
857 struct completion done;
861 * The task's runqueue lock must be held.
862 * Returns true if you have to wait for migration thread.
864 static int migrate_task(task_t *p, int dest_cpu, migration_req_t *req)
866 runqueue_t *rq = task_rq(p);
869 * If the task is not on a runqueue (and not running), then
870 * it is sufficient to simply update the task's cpu field.
872 if (!p->array && !task_running(rq, p)) {
873 set_task_cpu(p, dest_cpu);
877 init_completion(&req->done);
879 req->dest_cpu = dest_cpu;
880 list_add(&req->list, &rq->migration_queue);
885 * wait_task_inactive - wait for a thread to unschedule.
887 * The caller must ensure that the task *will* unschedule sometime soon,
888 * else this function might spin for a *long* time. This function can't
889 * be called with interrupts off, or it may introduce deadlock with
890 * smp_call_function() if an IPI is sent by the same process we are
891 * waiting to become inactive.
893 void wait_task_inactive(task_t *p)
900 rq = task_rq_lock(p, &flags);
901 /* Must be off runqueue entirely, not preempted. */
902 if (unlikely(p->array || task_running(rq, p))) {
903 /* If it's preempted, we yield. It could be a while. */
904 preempted = !task_running(rq, p);
905 task_rq_unlock(rq, &flags);
911 task_rq_unlock(rq, &flags);
915 * kick_process - kick a running thread to enter/exit the kernel
916 * @p: the to-be-kicked thread
918 * Cause a process which is running on another CPU to enter
919 * kernel-mode, without any delay. (to get signals handled.)
921 * NOTE: this function doesnt have to take the runqueue lock,
922 * because all it wants to ensure is that the remote task enters
923 * the kernel. If the IPI races and the task has been migrated
924 * to another CPU then no harm is done and the purpose has been
927 void kick_process(task_t *p)
933 if ((cpu != smp_processor_id()) && task_curr(p))
934 smp_send_reschedule(cpu);
939 * Return a low guess at the load of a migration-source cpu.
941 * We want to under-estimate the load of migration sources, to
942 * balance conservatively.
944 static inline unsigned long source_load(int cpu, int type)
946 runqueue_t *rq = cpu_rq(cpu);
947 unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE;
951 return min(rq->cpu_load[type-1], load_now);
955 * Return a high guess at the load of a migration-target cpu
957 static inline unsigned long target_load(int cpu, int type)
959 runqueue_t *rq = cpu_rq(cpu);
960 unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE;
964 return max(rq->cpu_load[type-1], load_now);
968 * find_idlest_group finds and returns the least busy CPU group within the
971 static struct sched_group *
972 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
974 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
975 unsigned long min_load = ULONG_MAX, this_load = 0;
976 int load_idx = sd->forkexec_idx;
977 int imbalance = 100 + (sd->imbalance_pct-100)/2;
980 unsigned long load, avg_load;
984 /* Skip over this group if it has no CPUs allowed */
985 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
988 local_group = cpu_isset(this_cpu, group->cpumask);
990 /* Tally up the load of all CPUs in the group */
993 for_each_cpu_mask(i, group->cpumask) {
994 /* Bias balancing toward cpus of our domain */
996 load = source_load(i, load_idx);
998 load = target_load(i, load_idx);
1003 /* Adjust by relative CPU power of the group */
1004 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1007 this_load = avg_load;
1009 } else if (avg_load < min_load) {
1010 min_load = avg_load;
1014 group = group->next;
1015 } while (group != sd->groups);
1017 if (!idlest || 100*this_load < imbalance*min_load)
1023 * find_idlest_queue - find the idlest runqueue among the cpus in group.
1026 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1029 unsigned long load, min_load = ULONG_MAX;
1033 /* Traverse only the allowed CPUs */
1034 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1036 for_each_cpu_mask(i, tmp) {
1037 load = source_load(i, 0);
1039 if (load < min_load || (load == min_load && i == this_cpu)) {
1049 * sched_balance_self: balance the current task (running on cpu) in domains
1050 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1053 * Balance, ie. select the least loaded group.
1055 * Returns the target CPU number, or the same CPU if no balancing is needed.
1057 * preempt must be disabled.
1059 static int sched_balance_self(int cpu, int flag)
1061 struct task_struct *t = current;
1062 struct sched_domain *tmp, *sd = NULL;
1064 for_each_domain(cpu, tmp)
1065 if (tmp->flags & flag)
1070 struct sched_group *group;
1075 group = find_idlest_group(sd, t, cpu);
1079 new_cpu = find_idlest_cpu(group, t, cpu);
1080 if (new_cpu == -1 || new_cpu == cpu)
1083 /* Now try balancing at a lower domain level */
1087 weight = cpus_weight(span);
1088 for_each_domain(cpu, tmp) {
1089 if (weight <= cpus_weight(tmp->span))
1091 if (tmp->flags & flag)
1094 /* while loop will break here if sd == NULL */
1100 #endif /* CONFIG_SMP */
1103 * wake_idle() will wake a task on an idle cpu if task->cpu is
1104 * not idle and an idle cpu is available. The span of cpus to
1105 * search starts with cpus closest then further out as needed,
1106 * so we always favor a closer, idle cpu.
1108 * Returns the CPU we should wake onto.
1110 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1111 static int wake_idle(int cpu, task_t *p)
1114 struct sched_domain *sd;
1120 for_each_domain(cpu, sd) {
1121 if (sd->flags & SD_WAKE_IDLE) {
1122 cpus_and(tmp, sd->span, p->cpus_allowed);
1123 for_each_cpu_mask(i, tmp) {
1134 static inline int wake_idle(int cpu, task_t *p)
1141 * try_to_wake_up - wake up a thread
1142 * @p: the to-be-woken-up thread
1143 * @state: the mask of task states that can be woken
1144 * @sync: do a synchronous wakeup?
1146 * Put it on the run-queue if it's not already there. The "current"
1147 * thread is always on the run-queue (except when the actual
1148 * re-schedule is in progress), and as such you're allowed to do
1149 * the simpler "current->state = TASK_RUNNING" to mark yourself
1150 * runnable without the overhead of this.
1152 * returns failure only if the task is already active.
1154 static int try_to_wake_up(task_t *p, unsigned int state, int sync)
1156 int cpu, this_cpu, success = 0;
1157 unsigned long flags;
1161 unsigned long load, this_load;
1162 struct sched_domain *sd, *this_sd = NULL;
1166 rq = task_rq_lock(p, &flags);
1167 old_state = p->state;
1168 if (!(old_state & state))
1175 this_cpu = smp_processor_id();
1178 if (unlikely(task_running(rq, p)))
1183 schedstat_inc(rq, ttwu_cnt);
1184 if (cpu == this_cpu) {
1185 schedstat_inc(rq, ttwu_local);
1189 for_each_domain(this_cpu, sd) {
1190 if (cpu_isset(cpu, sd->span)) {
1191 schedstat_inc(sd, ttwu_wake_remote);
1197 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1201 * Check for affine wakeup and passive balancing possibilities.
1204 int idx = this_sd->wake_idx;
1205 unsigned int imbalance;
1207 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1209 load = source_load(cpu, idx);
1210 this_load = target_load(this_cpu, idx);
1212 new_cpu = this_cpu; /* Wake to this CPU if we can */
1214 if (this_sd->flags & SD_WAKE_AFFINE) {
1215 unsigned long tl = this_load;
1217 * If sync wakeup then subtract the (maximum possible)
1218 * effect of the currently running task from the load
1219 * of the current CPU:
1222 tl -= SCHED_LOAD_SCALE;
1225 tl + target_load(cpu, idx) <= SCHED_LOAD_SCALE) ||
1226 100*(tl + SCHED_LOAD_SCALE) <= imbalance*load) {
1228 * This domain has SD_WAKE_AFFINE and
1229 * p is cache cold in this domain, and
1230 * there is no bad imbalance.
1232 schedstat_inc(this_sd, ttwu_move_affine);
1238 * Start passive balancing when half the imbalance_pct
1241 if (this_sd->flags & SD_WAKE_BALANCE) {
1242 if (imbalance*this_load <= 100*load) {
1243 schedstat_inc(this_sd, ttwu_move_balance);
1249 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1251 new_cpu = wake_idle(new_cpu, p);
1252 if (new_cpu != cpu) {
1253 set_task_cpu(p, new_cpu);
1254 task_rq_unlock(rq, &flags);
1255 /* might preempt at this point */
1256 rq = task_rq_lock(p, &flags);
1257 old_state = p->state;
1258 if (!(old_state & state))
1263 this_cpu = smp_processor_id();
1268 #endif /* CONFIG_SMP */
1269 if (old_state == TASK_UNINTERRUPTIBLE) {
1270 rq->nr_uninterruptible--;
1272 * Tasks on involuntary sleep don't earn
1273 * sleep_avg beyond just interactive state.
1279 * Tasks that have marked their sleep as noninteractive get
1280 * woken up without updating their sleep average. (i.e. their
1281 * sleep is handled in a priority-neutral manner, no priority
1282 * boost and no penalty.)
1284 if (old_state & TASK_NONINTERACTIVE)
1285 __activate_task(p, rq);
1287 activate_task(p, rq, cpu == this_cpu);
1289 * Sync wakeups (i.e. those types of wakeups where the waker
1290 * has indicated that it will leave the CPU in short order)
1291 * don't trigger a preemption, if the woken up task will run on
1292 * this cpu. (in this case the 'I will reschedule' promise of
1293 * the waker guarantees that the freshly woken up task is going
1294 * to be considered on this CPU.)
1296 if (!sync || cpu != this_cpu) {
1297 if (TASK_PREEMPTS_CURR(p, rq))
1298 resched_task(rq->curr);
1303 p->state = TASK_RUNNING;
1305 task_rq_unlock(rq, &flags);
1310 int fastcall wake_up_process(task_t *p)
1312 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1313 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1316 EXPORT_SYMBOL(wake_up_process);
1318 int fastcall wake_up_state(task_t *p, unsigned int state)
1320 return try_to_wake_up(p, state, 0);
1324 * Perform scheduler related setup for a newly forked process p.
1325 * p is forked by current.
1327 void fastcall sched_fork(task_t *p, int clone_flags)
1329 int cpu = get_cpu();
1332 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1334 set_task_cpu(p, cpu);
1337 * We mark the process as running here, but have not actually
1338 * inserted it onto the runqueue yet. This guarantees that
1339 * nobody will actually run it, and a signal or other external
1340 * event cannot wake it up and insert it on the runqueue either.
1342 p->state = TASK_RUNNING;
1343 INIT_LIST_HEAD(&p->run_list);
1345 #ifdef CONFIG_SCHEDSTATS
1346 memset(&p->sched_info, 0, sizeof(p->sched_info));
1348 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1351 #ifdef CONFIG_PREEMPT
1352 /* Want to start with kernel preemption disabled. */
1353 task_thread_info(p)->preempt_count = 1;
1356 * Share the timeslice between parent and child, thus the
1357 * total amount of pending timeslices in the system doesn't change,
1358 * resulting in more scheduling fairness.
1360 local_irq_disable();
1361 p->time_slice = (current->time_slice + 1) >> 1;
1363 * The remainder of the first timeslice might be recovered by
1364 * the parent if the child exits early enough.
1366 p->first_time_slice = 1;
1367 current->time_slice >>= 1;
1368 p->timestamp = sched_clock();
1369 if (unlikely(!current->time_slice)) {
1371 * This case is rare, it happens when the parent has only
1372 * a single jiffy left from its timeslice. Taking the
1373 * runqueue lock is not a problem.
1375 current->time_slice = 1;
1383 * wake_up_new_task - wake up a newly created task for the first time.
1385 * This function will do some initial scheduler statistics housekeeping
1386 * that must be done for every newly created context, then puts the task
1387 * on the runqueue and wakes it.
1389 void fastcall wake_up_new_task(task_t *p, unsigned long clone_flags)
1391 unsigned long flags;
1393 runqueue_t *rq, *this_rq;
1395 rq = task_rq_lock(p, &flags);
1396 BUG_ON(p->state != TASK_RUNNING);
1397 this_cpu = smp_processor_id();
1401 * We decrease the sleep average of forking parents
1402 * and children as well, to keep max-interactive tasks
1403 * from forking tasks that are max-interactive. The parent
1404 * (current) is done further down, under its lock.
1406 p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1407 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1409 p->prio = effective_prio(p);
1411 if (likely(cpu == this_cpu)) {
1412 if (!(clone_flags & CLONE_VM)) {
1414 * The VM isn't cloned, so we're in a good position to
1415 * do child-runs-first in anticipation of an exec. This
1416 * usually avoids a lot of COW overhead.
1418 if (unlikely(!current->array))
1419 __activate_task(p, rq);
1421 p->prio = current->prio;
1422 list_add_tail(&p->run_list, ¤t->run_list);
1423 p->array = current->array;
1424 p->array->nr_active++;
1429 /* Run child last */
1430 __activate_task(p, rq);
1432 * We skip the following code due to cpu == this_cpu
1434 * task_rq_unlock(rq, &flags);
1435 * this_rq = task_rq_lock(current, &flags);
1439 this_rq = cpu_rq(this_cpu);
1442 * Not the local CPU - must adjust timestamp. This should
1443 * get optimised away in the !CONFIG_SMP case.
1445 p->timestamp = (p->timestamp - this_rq->timestamp_last_tick)
1446 + rq->timestamp_last_tick;
1447 __activate_task(p, rq);
1448 if (TASK_PREEMPTS_CURR(p, rq))
1449 resched_task(rq->curr);
1452 * Parent and child are on different CPUs, now get the
1453 * parent runqueue to update the parent's ->sleep_avg:
1455 task_rq_unlock(rq, &flags);
1456 this_rq = task_rq_lock(current, &flags);
1458 current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1459 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1460 task_rq_unlock(this_rq, &flags);
1464 * Potentially available exiting-child timeslices are
1465 * retrieved here - this way the parent does not get
1466 * penalized for creating too many threads.
1468 * (this cannot be used to 'generate' timeslices
1469 * artificially, because any timeslice recovered here
1470 * was given away by the parent in the first place.)
1472 void fastcall sched_exit(task_t *p)
1474 unsigned long flags;
1478 * If the child was a (relative-) CPU hog then decrease
1479 * the sleep_avg of the parent as well.
1481 rq = task_rq_lock(p->parent, &flags);
1482 if (p->first_time_slice && task_cpu(p) == task_cpu(p->parent)) {
1483 p->parent->time_slice += p->time_slice;
1484 if (unlikely(p->parent->time_slice > task_timeslice(p)))
1485 p->parent->time_slice = task_timeslice(p);
1487 if (p->sleep_avg < p->parent->sleep_avg)
1488 p->parent->sleep_avg = p->parent->sleep_avg /
1489 (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg /
1491 task_rq_unlock(rq, &flags);
1495 * prepare_task_switch - prepare to switch tasks
1496 * @rq: the runqueue preparing to switch
1497 * @next: the task we are going to switch to.
1499 * This is called with the rq lock held and interrupts off. It must
1500 * be paired with a subsequent finish_task_switch after the context
1503 * prepare_task_switch sets up locking and calls architecture specific
1506 static inline void prepare_task_switch(runqueue_t *rq, task_t *next)
1508 prepare_lock_switch(rq, next);
1509 prepare_arch_switch(next);
1513 * finish_task_switch - clean up after a task-switch
1514 * @rq: runqueue associated with task-switch
1515 * @prev: the thread we just switched away from.
1517 * finish_task_switch must be called after the context switch, paired
1518 * with a prepare_task_switch call before the context switch.
1519 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1520 * and do any other architecture-specific cleanup actions.
1522 * Note that we may have delayed dropping an mm in context_switch(). If
1523 * so, we finish that here outside of the runqueue lock. (Doing it
1524 * with the lock held can cause deadlocks; see schedule() for
1527 static inline void finish_task_switch(runqueue_t *rq, task_t *prev)
1528 __releases(rq->lock)
1530 struct mm_struct *mm = rq->prev_mm;
1531 unsigned long prev_task_flags;
1536 * A task struct has one reference for the use as "current".
1537 * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1538 * calls schedule one last time. The schedule call will never return,
1539 * and the scheduled task must drop that reference.
1540 * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1541 * still held, otherwise prev could be scheduled on another cpu, die
1542 * there before we look at prev->state, and then the reference would
1544 * Manfred Spraul <manfred@colorfullife.com>
1546 prev_task_flags = prev->flags;
1547 finish_arch_switch(prev);
1548 finish_lock_switch(rq, prev);
1551 if (unlikely(prev_task_flags & PF_DEAD)) {
1553 * Remove function-return probe instances associated with this
1554 * task and put them back on the free list.
1556 kprobe_flush_task(prev);
1557 put_task_struct(prev);
1562 * schedule_tail - first thing a freshly forked thread must call.
1563 * @prev: the thread we just switched away from.
1565 asmlinkage void schedule_tail(task_t *prev)
1566 __releases(rq->lock)
1568 runqueue_t *rq = this_rq();
1569 finish_task_switch(rq, prev);
1570 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1571 /* In this case, finish_task_switch does not reenable preemption */
1574 if (current->set_child_tid)
1575 put_user(current->pid, current->set_child_tid);
1579 * context_switch - switch to the new MM and the new
1580 * thread's register state.
1583 task_t * context_switch(runqueue_t *rq, task_t *prev, task_t *next)
1585 struct mm_struct *mm = next->mm;
1586 struct mm_struct *oldmm = prev->active_mm;
1588 if (unlikely(!mm)) {
1589 next->active_mm = oldmm;
1590 atomic_inc(&oldmm->mm_count);
1591 enter_lazy_tlb(oldmm, next);
1593 switch_mm(oldmm, mm, next);
1595 if (unlikely(!prev->mm)) {
1596 prev->active_mm = NULL;
1597 WARN_ON(rq->prev_mm);
1598 rq->prev_mm = oldmm;
1601 /* Here we just switch the register state and the stack. */
1602 switch_to(prev, next, prev);
1608 * nr_running, nr_uninterruptible and nr_context_switches:
1610 * externally visible scheduler statistics: current number of runnable
1611 * threads, current number of uninterruptible-sleeping threads, total
1612 * number of context switches performed since bootup.
1614 unsigned long nr_running(void)
1616 unsigned long i, sum = 0;
1618 for_each_online_cpu(i)
1619 sum += cpu_rq(i)->nr_running;
1624 unsigned long nr_uninterruptible(void)
1626 unsigned long i, sum = 0;
1628 for_each_possible_cpu(i)
1629 sum += cpu_rq(i)->nr_uninterruptible;
1632 * Since we read the counters lockless, it might be slightly
1633 * inaccurate. Do not allow it to go below zero though:
1635 if (unlikely((long)sum < 0))
1641 unsigned long long nr_context_switches(void)
1643 unsigned long long i, sum = 0;
1645 for_each_possible_cpu(i)
1646 sum += cpu_rq(i)->nr_switches;
1651 unsigned long nr_iowait(void)
1653 unsigned long i, sum = 0;
1655 for_each_possible_cpu(i)
1656 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1661 unsigned long nr_active(void)
1663 unsigned long i, running = 0, uninterruptible = 0;
1665 for_each_online_cpu(i) {
1666 running += cpu_rq(i)->nr_running;
1667 uninterruptible += cpu_rq(i)->nr_uninterruptible;
1670 if (unlikely((long)uninterruptible < 0))
1671 uninterruptible = 0;
1673 return running + uninterruptible;
1679 * double_rq_lock - safely lock two runqueues
1681 * We must take them in cpu order to match code in
1682 * dependent_sleeper and wake_dependent_sleeper.
1684 * Note this does not disable interrupts like task_rq_lock,
1685 * you need to do so manually before calling.
1687 static void double_rq_lock(runqueue_t *rq1, runqueue_t *rq2)
1688 __acquires(rq1->lock)
1689 __acquires(rq2->lock)
1692 spin_lock(&rq1->lock);
1693 __acquire(rq2->lock); /* Fake it out ;) */
1695 if (rq1->cpu < rq2->cpu) {
1696 spin_lock(&rq1->lock);
1697 spin_lock(&rq2->lock);
1699 spin_lock(&rq2->lock);
1700 spin_lock(&rq1->lock);
1706 * double_rq_unlock - safely unlock two runqueues
1708 * Note this does not restore interrupts like task_rq_unlock,
1709 * you need to do so manually after calling.
1711 static void double_rq_unlock(runqueue_t *rq1, runqueue_t *rq2)
1712 __releases(rq1->lock)
1713 __releases(rq2->lock)
1715 spin_unlock(&rq1->lock);
1717 spin_unlock(&rq2->lock);
1719 __release(rq2->lock);
1723 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1725 static void double_lock_balance(runqueue_t *this_rq, runqueue_t *busiest)
1726 __releases(this_rq->lock)
1727 __acquires(busiest->lock)
1728 __acquires(this_rq->lock)
1730 if (unlikely(!spin_trylock(&busiest->lock))) {
1731 if (busiest->cpu < this_rq->cpu) {
1732 spin_unlock(&this_rq->lock);
1733 spin_lock(&busiest->lock);
1734 spin_lock(&this_rq->lock);
1736 spin_lock(&busiest->lock);
1741 * If dest_cpu is allowed for this process, migrate the task to it.
1742 * This is accomplished by forcing the cpu_allowed mask to only
1743 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
1744 * the cpu_allowed mask is restored.
1746 static void sched_migrate_task(task_t *p, int dest_cpu)
1748 migration_req_t req;
1750 unsigned long flags;
1752 rq = task_rq_lock(p, &flags);
1753 if (!cpu_isset(dest_cpu, p->cpus_allowed)
1754 || unlikely(cpu_is_offline(dest_cpu)))
1757 /* force the process onto the specified CPU */
1758 if (migrate_task(p, dest_cpu, &req)) {
1759 /* Need to wait for migration thread (might exit: take ref). */
1760 struct task_struct *mt = rq->migration_thread;
1761 get_task_struct(mt);
1762 task_rq_unlock(rq, &flags);
1763 wake_up_process(mt);
1764 put_task_struct(mt);
1765 wait_for_completion(&req.done);
1769 task_rq_unlock(rq, &flags);
1773 * sched_exec - execve() is a valuable balancing opportunity, because at
1774 * this point the task has the smallest effective memory and cache footprint.
1776 void sched_exec(void)
1778 int new_cpu, this_cpu = get_cpu();
1779 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
1781 if (new_cpu != this_cpu)
1782 sched_migrate_task(current, new_cpu);
1786 * pull_task - move a task from a remote runqueue to the local runqueue.
1787 * Both runqueues must be locked.
1790 void pull_task(runqueue_t *src_rq, prio_array_t *src_array, task_t *p,
1791 runqueue_t *this_rq, prio_array_t *this_array, int this_cpu)
1793 dequeue_task(p, src_array);
1794 src_rq->nr_running--;
1795 set_task_cpu(p, this_cpu);
1796 this_rq->nr_running++;
1797 enqueue_task(p, this_array);
1798 p->timestamp = (p->timestamp - src_rq->timestamp_last_tick)
1799 + this_rq->timestamp_last_tick;
1801 * Note that idle threads have a prio of MAX_PRIO, for this test
1802 * to be always true for them.
1804 if (TASK_PREEMPTS_CURR(p, this_rq))
1805 resched_task(this_rq->curr);
1809 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1812 int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu,
1813 struct sched_domain *sd, enum idle_type idle,
1817 * We do not migrate tasks that are:
1818 * 1) running (obviously), or
1819 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1820 * 3) are cache-hot on their current CPU.
1822 if (!cpu_isset(this_cpu, p->cpus_allowed))
1826 if (task_running(rq, p))
1830 * Aggressive migration if:
1831 * 1) task is cache cold, or
1832 * 2) too many balance attempts have failed.
1835 if (sd->nr_balance_failed > sd->cache_nice_tries)
1838 if (task_hot(p, rq->timestamp_last_tick, sd))
1844 * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq,
1845 * as part of a balancing operation within "domain". Returns the number of
1848 * Called with both runqueues locked.
1850 static int move_tasks(runqueue_t *this_rq, int this_cpu, runqueue_t *busiest,
1851 unsigned long max_nr_move, struct sched_domain *sd,
1852 enum idle_type idle, int *all_pinned)
1854 prio_array_t *array, *dst_array;
1855 struct list_head *head, *curr;
1856 int idx, pulled = 0, pinned = 0;
1859 if (max_nr_move == 0)
1865 * We first consider expired tasks. Those will likely not be
1866 * executed in the near future, and they are most likely to
1867 * be cache-cold, thus switching CPUs has the least effect
1870 if (busiest->expired->nr_active) {
1871 array = busiest->expired;
1872 dst_array = this_rq->expired;
1874 array = busiest->active;
1875 dst_array = this_rq->active;
1879 /* Start searching at priority 0: */
1883 idx = sched_find_first_bit(array->bitmap);
1885 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
1886 if (idx >= MAX_PRIO) {
1887 if (array == busiest->expired && busiest->active->nr_active) {
1888 array = busiest->active;
1889 dst_array = this_rq->active;
1895 head = array->queue + idx;
1898 tmp = list_entry(curr, task_t, run_list);
1902 if (!can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
1909 #ifdef CONFIG_SCHEDSTATS
1910 if (task_hot(tmp, busiest->timestamp_last_tick, sd))
1911 schedstat_inc(sd, lb_hot_gained[idle]);
1914 pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
1917 /* We only want to steal up to the prescribed number of tasks. */
1918 if (pulled < max_nr_move) {
1926 * Right now, this is the only place pull_task() is called,
1927 * so we can safely collect pull_task() stats here rather than
1928 * inside pull_task().
1930 schedstat_add(sd, lb_gained[idle], pulled);
1933 *all_pinned = pinned;
1938 * find_busiest_group finds and returns the busiest CPU group within the
1939 * domain. It calculates and returns the number of tasks which should be
1940 * moved to restore balance via the imbalance parameter.
1942 static struct sched_group *
1943 find_busiest_group(struct sched_domain *sd, int this_cpu,
1944 unsigned long *imbalance, enum idle_type idle, int *sd_idle)
1946 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
1947 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
1948 unsigned long max_pull;
1951 max_load = this_load = total_load = total_pwr = 0;
1952 if (idle == NOT_IDLE)
1953 load_idx = sd->busy_idx;
1954 else if (idle == NEWLY_IDLE)
1955 load_idx = sd->newidle_idx;
1957 load_idx = sd->idle_idx;
1964 local_group = cpu_isset(this_cpu, group->cpumask);
1966 /* Tally up the load of all CPUs in the group */
1969 for_each_cpu_mask(i, group->cpumask) {
1970 if (*sd_idle && !idle_cpu(i))
1973 /* Bias balancing toward cpus of our domain */
1975 load = target_load(i, load_idx);
1977 load = source_load(i, load_idx);
1982 total_load += avg_load;
1983 total_pwr += group->cpu_power;
1985 /* Adjust by relative CPU power of the group */
1986 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1989 this_load = avg_load;
1991 } else if (avg_load > max_load) {
1992 max_load = avg_load;
1995 group = group->next;
1996 } while (group != sd->groups);
1998 if (!busiest || this_load >= max_load || max_load <= SCHED_LOAD_SCALE)
2001 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2003 if (this_load >= avg_load ||
2004 100*max_load <= sd->imbalance_pct*this_load)
2008 * We're trying to get all the cpus to the average_load, so we don't
2009 * want to push ourselves above the average load, nor do we wish to
2010 * reduce the max loaded cpu below the average load, as either of these
2011 * actions would just result in more rebalancing later, and ping-pong
2012 * tasks around. Thus we look for the minimum possible imbalance.
2013 * Negative imbalances (*we* are more loaded than anyone else) will
2014 * be counted as no imbalance for these purposes -- we can't fix that
2015 * by pulling tasks to us. Be careful of negative numbers as they'll
2016 * appear as very large values with unsigned longs.
2019 /* Don't want to pull so many tasks that a group would go idle */
2020 max_pull = min(max_load - avg_load, max_load - SCHED_LOAD_SCALE);
2022 /* How much load to actually move to equalise the imbalance */
2023 *imbalance = min(max_pull * busiest->cpu_power,
2024 (avg_load - this_load) * this->cpu_power)
2027 if (*imbalance < SCHED_LOAD_SCALE) {
2028 unsigned long pwr_now = 0, pwr_move = 0;
2031 if (max_load - this_load >= SCHED_LOAD_SCALE*2) {
2037 * OK, we don't have enough imbalance to justify moving tasks,
2038 * however we may be able to increase total CPU power used by
2042 pwr_now += busiest->cpu_power*min(SCHED_LOAD_SCALE, max_load);
2043 pwr_now += this->cpu_power*min(SCHED_LOAD_SCALE, this_load);
2044 pwr_now /= SCHED_LOAD_SCALE;
2046 /* Amount of load we'd subtract */
2047 tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/busiest->cpu_power;
2049 pwr_move += busiest->cpu_power*min(SCHED_LOAD_SCALE,
2052 /* Amount of load we'd add */
2053 if (max_load*busiest->cpu_power <
2054 SCHED_LOAD_SCALE*SCHED_LOAD_SCALE)
2055 tmp = max_load*busiest->cpu_power/this->cpu_power;
2057 tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/this->cpu_power;
2058 pwr_move += this->cpu_power*min(SCHED_LOAD_SCALE, this_load + tmp);
2059 pwr_move /= SCHED_LOAD_SCALE;
2061 /* Move if we gain throughput */
2062 if (pwr_move <= pwr_now)
2069 /* Get rid of the scaling factor, rounding down as we divide */
2070 *imbalance = *imbalance / SCHED_LOAD_SCALE;
2080 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2082 static runqueue_t *find_busiest_queue(struct sched_group *group,
2083 enum idle_type idle)
2085 unsigned long load, max_load = 0;
2086 runqueue_t *busiest = NULL;
2089 for_each_cpu_mask(i, group->cpumask) {
2090 load = source_load(i, 0);
2092 if (load > max_load) {
2094 busiest = cpu_rq(i);
2102 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2103 * so long as it is large enough.
2105 #define MAX_PINNED_INTERVAL 512
2108 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2109 * tasks if there is an imbalance.
2111 * Called with this_rq unlocked.
2113 static int load_balance(int this_cpu, runqueue_t *this_rq,
2114 struct sched_domain *sd, enum idle_type idle)
2116 struct sched_group *group;
2117 runqueue_t *busiest;
2118 unsigned long imbalance;
2119 int nr_moved, all_pinned = 0;
2120 int active_balance = 0;
2123 if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER)
2126 schedstat_inc(sd, lb_cnt[idle]);
2128 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle);
2130 schedstat_inc(sd, lb_nobusyg[idle]);
2134 busiest = find_busiest_queue(group, idle);
2136 schedstat_inc(sd, lb_nobusyq[idle]);
2140 BUG_ON(busiest == this_rq);
2142 schedstat_add(sd, lb_imbalance[idle], imbalance);
2145 if (busiest->nr_running > 1) {
2147 * Attempt to move tasks. If find_busiest_group has found
2148 * an imbalance but busiest->nr_running <= 1, the group is
2149 * still unbalanced. nr_moved simply stays zero, so it is
2150 * correctly treated as an imbalance.
2152 double_rq_lock(this_rq, busiest);
2153 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2154 imbalance, sd, idle, &all_pinned);
2155 double_rq_unlock(this_rq, busiest);
2157 /* All tasks on this runqueue were pinned by CPU affinity */
2158 if (unlikely(all_pinned))
2163 schedstat_inc(sd, lb_failed[idle]);
2164 sd->nr_balance_failed++;
2166 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2168 spin_lock(&busiest->lock);
2170 /* don't kick the migration_thread, if the curr
2171 * task on busiest cpu can't be moved to this_cpu
2173 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2174 spin_unlock(&busiest->lock);
2176 goto out_one_pinned;
2179 if (!busiest->active_balance) {
2180 busiest->active_balance = 1;
2181 busiest->push_cpu = this_cpu;
2184 spin_unlock(&busiest->lock);
2186 wake_up_process(busiest->migration_thread);
2189 * We've kicked active balancing, reset the failure
2192 sd->nr_balance_failed = sd->cache_nice_tries+1;
2195 sd->nr_balance_failed = 0;
2197 if (likely(!active_balance)) {
2198 /* We were unbalanced, so reset the balancing interval */
2199 sd->balance_interval = sd->min_interval;
2202 * If we've begun active balancing, start to back off. This
2203 * case may not be covered by the all_pinned logic if there
2204 * is only 1 task on the busy runqueue (because we don't call
2207 if (sd->balance_interval < sd->max_interval)
2208 sd->balance_interval *= 2;
2211 if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2216 schedstat_inc(sd, lb_balanced[idle]);
2218 sd->nr_balance_failed = 0;
2221 /* tune up the balancing interval */
2222 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2223 (sd->balance_interval < sd->max_interval))
2224 sd->balance_interval *= 2;
2226 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2232 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2233 * tasks if there is an imbalance.
2235 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2236 * this_rq is locked.
2238 static int load_balance_newidle(int this_cpu, runqueue_t *this_rq,
2239 struct sched_domain *sd)
2241 struct sched_group *group;
2242 runqueue_t *busiest = NULL;
2243 unsigned long imbalance;
2247 if (sd->flags & SD_SHARE_CPUPOWER)
2250 schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
2251 group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE, &sd_idle);
2253 schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]);
2257 busiest = find_busiest_queue(group, NEWLY_IDLE);
2259 schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]);
2263 BUG_ON(busiest == this_rq);
2265 schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
2268 if (busiest->nr_running > 1) {
2269 /* Attempt to move tasks */
2270 double_lock_balance(this_rq, busiest);
2271 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2272 imbalance, sd, NEWLY_IDLE, NULL);
2273 spin_unlock(&busiest->lock);
2277 schedstat_inc(sd, lb_failed[NEWLY_IDLE]);
2278 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2281 sd->nr_balance_failed = 0;
2286 schedstat_inc(sd, lb_balanced[NEWLY_IDLE]);
2287 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2289 sd->nr_balance_failed = 0;
2294 * idle_balance is called by schedule() if this_cpu is about to become
2295 * idle. Attempts to pull tasks from other CPUs.
2297 static void idle_balance(int this_cpu, runqueue_t *this_rq)
2299 struct sched_domain *sd;
2301 for_each_domain(this_cpu, sd) {
2302 if (sd->flags & SD_BALANCE_NEWIDLE) {
2303 if (load_balance_newidle(this_cpu, this_rq, sd)) {
2304 /* We've pulled tasks over so stop searching */
2312 * active_load_balance is run by migration threads. It pushes running tasks
2313 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2314 * running on each physical CPU where possible, and avoids physical /
2315 * logical imbalances.
2317 * Called with busiest_rq locked.
2319 static void active_load_balance(runqueue_t *busiest_rq, int busiest_cpu)
2321 struct sched_domain *sd;
2322 runqueue_t *target_rq;
2323 int target_cpu = busiest_rq->push_cpu;
2325 if (busiest_rq->nr_running <= 1)
2326 /* no task to move */
2329 target_rq = cpu_rq(target_cpu);
2332 * This condition is "impossible", if it occurs
2333 * we need to fix it. Originally reported by
2334 * Bjorn Helgaas on a 128-cpu setup.
2336 BUG_ON(busiest_rq == target_rq);
2338 /* move a task from busiest_rq to target_rq */
2339 double_lock_balance(busiest_rq, target_rq);
2341 /* Search for an sd spanning us and the target CPU. */
2342 for_each_domain(target_cpu, sd)
2343 if ((sd->flags & SD_LOAD_BALANCE) &&
2344 cpu_isset(busiest_cpu, sd->span))
2347 if (unlikely(sd == NULL))
2350 schedstat_inc(sd, alb_cnt);
2352 if (move_tasks(target_rq, target_cpu, busiest_rq, 1, sd, SCHED_IDLE, NULL))
2353 schedstat_inc(sd, alb_pushed);
2355 schedstat_inc(sd, alb_failed);
2357 spin_unlock(&target_rq->lock);
2361 * rebalance_tick will get called every timer tick, on every CPU.
2363 * It checks each scheduling domain to see if it is due to be balanced,
2364 * and initiates a balancing operation if so.
2366 * Balancing parameters are set up in arch_init_sched_domains.
2369 /* Don't have all balancing operations going off at once */
2370 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2372 static void rebalance_tick(int this_cpu, runqueue_t *this_rq,
2373 enum idle_type idle)
2375 unsigned long old_load, this_load;
2376 unsigned long j = jiffies + CPU_OFFSET(this_cpu);
2377 struct sched_domain *sd;
2380 this_load = this_rq->nr_running * SCHED_LOAD_SCALE;
2381 /* Update our load */
2382 for (i = 0; i < 3; i++) {
2383 unsigned long new_load = this_load;
2385 old_load = this_rq->cpu_load[i];
2387 * Round up the averaging division if load is increasing. This
2388 * prevents us from getting stuck on 9 if the load is 10, for
2391 if (new_load > old_load)
2392 new_load += scale-1;
2393 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale;
2396 for_each_domain(this_cpu, sd) {
2397 unsigned long interval;
2399 if (!(sd->flags & SD_LOAD_BALANCE))
2402 interval = sd->balance_interval;
2403 if (idle != SCHED_IDLE)
2404 interval *= sd->busy_factor;
2406 /* scale ms to jiffies */
2407 interval = msecs_to_jiffies(interval);
2408 if (unlikely(!interval))
2411 if (j - sd->last_balance >= interval) {
2412 if (load_balance(this_cpu, this_rq, sd, idle)) {
2414 * We've pulled tasks over so either we're no
2415 * longer idle, or one of our SMT siblings is
2420 sd->last_balance += interval;
2426 * on UP we do not need to balance between CPUs:
2428 static inline void rebalance_tick(int cpu, runqueue_t *rq, enum idle_type idle)
2431 static inline void idle_balance(int cpu, runqueue_t *rq)
2436 static inline int wake_priority_sleeper(runqueue_t *rq)
2439 #ifdef CONFIG_SCHED_SMT
2440 spin_lock(&rq->lock);
2442 * If an SMT sibling task has been put to sleep for priority
2443 * reasons reschedule the idle task to see if it can now run.
2445 if (rq->nr_running) {
2446 resched_task(rq->idle);
2449 spin_unlock(&rq->lock);
2454 DEFINE_PER_CPU(struct kernel_stat, kstat);
2456 EXPORT_PER_CPU_SYMBOL(kstat);
2459 * This is called on clock ticks and on context switches.
2460 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2462 static inline void update_cpu_clock(task_t *p, runqueue_t *rq,
2463 unsigned long long now)
2465 unsigned long long last = max(p->timestamp, rq->timestamp_last_tick);
2466 p->sched_time += now - last;
2470 * Return current->sched_time plus any more ns on the sched_clock
2471 * that have not yet been banked.
2473 unsigned long long current_sched_time(const task_t *tsk)
2475 unsigned long long ns;
2476 unsigned long flags;
2477 local_irq_save(flags);
2478 ns = max(tsk->timestamp, task_rq(tsk)->timestamp_last_tick);
2479 ns = tsk->sched_time + (sched_clock() - ns);
2480 local_irq_restore(flags);
2485 * We place interactive tasks back into the active array, if possible.
2487 * To guarantee that this does not starve expired tasks we ignore the
2488 * interactivity of a task if the first expired task had to wait more
2489 * than a 'reasonable' amount of time. This deadline timeout is
2490 * load-dependent, as the frequency of array switched decreases with
2491 * increasing number of running tasks. We also ignore the interactivity
2492 * if a better static_prio task has expired:
2494 #define EXPIRED_STARVING(rq) \
2495 ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2496 (jiffies - (rq)->expired_timestamp >= \
2497 STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2498 ((rq)->curr->static_prio > (rq)->best_expired_prio))
2501 * Account user cpu time to a process.
2502 * @p: the process that the cpu time gets accounted to
2503 * @hardirq_offset: the offset to subtract from hardirq_count()
2504 * @cputime: the cpu time spent in user space since the last update
2506 void account_user_time(struct task_struct *p, cputime_t cputime)
2508 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2511 p->utime = cputime_add(p->utime, cputime);
2513 /* Add user time to cpustat. */
2514 tmp = cputime_to_cputime64(cputime);
2515 if (TASK_NICE(p) > 0)
2516 cpustat->nice = cputime64_add(cpustat->nice, tmp);
2518 cpustat->user = cputime64_add(cpustat->user, tmp);
2522 * Account system cpu time to a process.
2523 * @p: the process that the cpu time gets accounted to
2524 * @hardirq_offset: the offset to subtract from hardirq_count()
2525 * @cputime: the cpu time spent in kernel space since the last update
2527 void account_system_time(struct task_struct *p, int hardirq_offset,
2530 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2531 runqueue_t *rq = this_rq();
2534 p->stime = cputime_add(p->stime, cputime);
2536 /* Add system time to cpustat. */
2537 tmp = cputime_to_cputime64(cputime);
2538 if (hardirq_count() - hardirq_offset)
2539 cpustat->irq = cputime64_add(cpustat->irq, tmp);
2540 else if (softirq_count())
2541 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
2542 else if (p != rq->idle)
2543 cpustat->system = cputime64_add(cpustat->system, tmp);
2544 else if (atomic_read(&rq->nr_iowait) > 0)
2545 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2547 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2548 /* Account for system time used */
2549 acct_update_integrals(p);
2553 * Account for involuntary wait time.
2554 * @p: the process from which the cpu time has been stolen
2555 * @steal: the cpu time spent in involuntary wait
2557 void account_steal_time(struct task_struct *p, cputime_t steal)
2559 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2560 cputime64_t tmp = cputime_to_cputime64(steal);
2561 runqueue_t *rq = this_rq();
2563 if (p == rq->idle) {
2564 p->stime = cputime_add(p->stime, steal);
2565 if (atomic_read(&rq->nr_iowait) > 0)
2566 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2568 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2570 cpustat->steal = cputime64_add(cpustat->steal, tmp);
2574 * This function gets called by the timer code, with HZ frequency.
2575 * We call it with interrupts disabled.
2577 * It also gets called by the fork code, when changing the parent's
2580 void scheduler_tick(void)
2582 int cpu = smp_processor_id();
2583 runqueue_t *rq = this_rq();
2584 task_t *p = current;
2585 unsigned long long now = sched_clock();
2587 update_cpu_clock(p, rq, now);
2589 rq->timestamp_last_tick = now;
2591 if (p == rq->idle) {
2592 if (wake_priority_sleeper(rq))
2594 rebalance_tick(cpu, rq, SCHED_IDLE);
2598 /* Task might have expired already, but not scheduled off yet */
2599 if (p->array != rq->active) {
2600 set_tsk_need_resched(p);
2603 spin_lock(&rq->lock);
2605 * The task was running during this tick - update the
2606 * time slice counter. Note: we do not update a thread's
2607 * priority until it either goes to sleep or uses up its
2608 * timeslice. This makes it possible for interactive tasks
2609 * to use up their timeslices at their highest priority levels.
2613 * RR tasks need a special form of timeslice management.
2614 * FIFO tasks have no timeslices.
2616 if ((p->policy == SCHED_RR) && !--p->time_slice) {
2617 p->time_slice = task_timeslice(p);
2618 p->first_time_slice = 0;
2619 set_tsk_need_resched(p);
2621 /* put it at the end of the queue: */
2622 requeue_task(p, rq->active);
2626 if (!--p->time_slice) {
2627 dequeue_task(p, rq->active);
2628 set_tsk_need_resched(p);
2629 p->prio = effective_prio(p);
2630 p->time_slice = task_timeslice(p);
2631 p->first_time_slice = 0;
2633 if (!rq->expired_timestamp)
2634 rq->expired_timestamp = jiffies;
2635 if (!TASK_INTERACTIVE(p) || EXPIRED_STARVING(rq)) {
2636 enqueue_task(p, rq->expired);
2637 if (p->static_prio < rq->best_expired_prio)
2638 rq->best_expired_prio = p->static_prio;
2640 enqueue_task(p, rq->active);
2643 * Prevent a too long timeslice allowing a task to monopolize
2644 * the CPU. We do this by splitting up the timeslice into
2647 * Note: this does not mean the task's timeslices expire or
2648 * get lost in any way, they just might be preempted by
2649 * another task of equal priority. (one with higher
2650 * priority would have preempted this task already.) We
2651 * requeue this task to the end of the list on this priority
2652 * level, which is in essence a round-robin of tasks with
2655 * This only applies to tasks in the interactive
2656 * delta range with at least TIMESLICE_GRANULARITY to requeue.
2658 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
2659 p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
2660 (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
2661 (p->array == rq->active)) {
2663 requeue_task(p, rq->active);
2664 set_tsk_need_resched(p);
2668 spin_unlock(&rq->lock);
2670 rebalance_tick(cpu, rq, NOT_IDLE);
2673 #ifdef CONFIG_SCHED_SMT
2674 static inline void wakeup_busy_runqueue(runqueue_t *rq)
2676 /* If an SMT runqueue is sleeping due to priority reasons wake it up */
2677 if (rq->curr == rq->idle && rq->nr_running)
2678 resched_task(rq->idle);
2681 static void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2683 struct sched_domain *tmp, *sd = NULL;
2684 cpumask_t sibling_map;
2687 for_each_domain(this_cpu, tmp)
2688 if (tmp->flags & SD_SHARE_CPUPOWER)
2695 * Unlock the current runqueue because we have to lock in
2696 * CPU order to avoid deadlocks. Caller knows that we might
2697 * unlock. We keep IRQs disabled.
2699 spin_unlock(&this_rq->lock);
2701 sibling_map = sd->span;
2703 for_each_cpu_mask(i, sibling_map)
2704 spin_lock(&cpu_rq(i)->lock);
2706 * We clear this CPU from the mask. This both simplifies the
2707 * inner loop and keps this_rq locked when we exit:
2709 cpu_clear(this_cpu, sibling_map);
2711 for_each_cpu_mask(i, sibling_map) {
2712 runqueue_t *smt_rq = cpu_rq(i);
2714 wakeup_busy_runqueue(smt_rq);
2717 for_each_cpu_mask(i, sibling_map)
2718 spin_unlock(&cpu_rq(i)->lock);
2720 * We exit with this_cpu's rq still held and IRQs
2726 * number of 'lost' timeslices this task wont be able to fully
2727 * utilize, if another task runs on a sibling. This models the
2728 * slowdown effect of other tasks running on siblings:
2730 static inline unsigned long smt_slice(task_t *p, struct sched_domain *sd)
2732 return p->time_slice * (100 - sd->per_cpu_gain) / 100;
2735 static int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2737 struct sched_domain *tmp, *sd = NULL;
2738 cpumask_t sibling_map;
2739 prio_array_t *array;
2743 for_each_domain(this_cpu, tmp)
2744 if (tmp->flags & SD_SHARE_CPUPOWER)
2751 * The same locking rules and details apply as for
2752 * wake_sleeping_dependent():
2754 spin_unlock(&this_rq->lock);
2755 sibling_map = sd->span;
2756 for_each_cpu_mask(i, sibling_map)
2757 spin_lock(&cpu_rq(i)->lock);
2758 cpu_clear(this_cpu, sibling_map);
2761 * Establish next task to be run - it might have gone away because
2762 * we released the runqueue lock above:
2764 if (!this_rq->nr_running)
2766 array = this_rq->active;
2767 if (!array->nr_active)
2768 array = this_rq->expired;
2769 BUG_ON(!array->nr_active);
2771 p = list_entry(array->queue[sched_find_first_bit(array->bitmap)].next,
2774 for_each_cpu_mask(i, sibling_map) {
2775 runqueue_t *smt_rq = cpu_rq(i);
2776 task_t *smt_curr = smt_rq->curr;
2778 /* Kernel threads do not participate in dependent sleeping */
2779 if (!p->mm || !smt_curr->mm || rt_task(p))
2780 goto check_smt_task;
2783 * If a user task with lower static priority than the
2784 * running task on the SMT sibling is trying to schedule,
2785 * delay it till there is proportionately less timeslice
2786 * left of the sibling task to prevent a lower priority
2787 * task from using an unfair proportion of the
2788 * physical cpu's resources. -ck
2790 if (rt_task(smt_curr)) {
2792 * With real time tasks we run non-rt tasks only
2793 * per_cpu_gain% of the time.
2795 if ((jiffies % DEF_TIMESLICE) >
2796 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
2799 if (smt_curr->static_prio < p->static_prio &&
2800 !TASK_PREEMPTS_CURR(p, smt_rq) &&
2801 smt_slice(smt_curr, sd) > task_timeslice(p))
2805 if ((!smt_curr->mm && smt_curr != smt_rq->idle) ||
2809 wakeup_busy_runqueue(smt_rq);
2814 * Reschedule a lower priority task on the SMT sibling for
2815 * it to be put to sleep, or wake it up if it has been put to
2816 * sleep for priority reasons to see if it should run now.
2819 if ((jiffies % DEF_TIMESLICE) >
2820 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
2821 resched_task(smt_curr);
2823 if (TASK_PREEMPTS_CURR(p, smt_rq) &&
2824 smt_slice(p, sd) > task_timeslice(smt_curr))
2825 resched_task(smt_curr);
2827 wakeup_busy_runqueue(smt_rq);
2831 for_each_cpu_mask(i, sibling_map)
2832 spin_unlock(&cpu_rq(i)->lock);
2836 static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2840 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2846 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
2848 void fastcall add_preempt_count(int val)
2853 BUG_ON((preempt_count() < 0));
2854 preempt_count() += val;
2856 * Spinlock count overflowing soon?
2858 BUG_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10);
2860 EXPORT_SYMBOL(add_preempt_count);
2862 void fastcall sub_preempt_count(int val)
2867 BUG_ON(val > preempt_count());
2869 * Is the spinlock portion underflowing?
2871 BUG_ON((val < PREEMPT_MASK) && !(preempt_count() & PREEMPT_MASK));
2872 preempt_count() -= val;
2874 EXPORT_SYMBOL(sub_preempt_count);
2879 * schedule() is the main scheduler function.
2881 asmlinkage void __sched schedule(void)
2884 task_t *prev, *next;
2886 prio_array_t *array;
2887 struct list_head *queue;
2888 unsigned long long now;
2889 unsigned long run_time;
2890 int cpu, idx, new_prio;
2893 * Test if we are atomic. Since do_exit() needs to call into
2894 * schedule() atomically, we ignore that path for now.
2895 * Otherwise, whine if we are scheduling when we should not be.
2897 if (unlikely(in_atomic() && !current->exit_state)) {
2898 printk(KERN_ERR "BUG: scheduling while atomic: "
2900 current->comm, preempt_count(), current->pid);
2903 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2908 release_kernel_lock(prev);
2909 need_resched_nonpreemptible:
2913 * The idle thread is not allowed to schedule!
2914 * Remove this check after it has been exercised a bit.
2916 if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
2917 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
2921 schedstat_inc(rq, sched_cnt);
2922 now = sched_clock();
2923 if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
2924 run_time = now - prev->timestamp;
2925 if (unlikely((long long)(now - prev->timestamp) < 0))
2928 run_time = NS_MAX_SLEEP_AVG;
2931 * Tasks charged proportionately less run_time at high sleep_avg to
2932 * delay them losing their interactive status
2934 run_time /= (CURRENT_BONUS(prev) ? : 1);
2936 spin_lock_irq(&rq->lock);
2938 if (unlikely(prev->flags & PF_DEAD))
2939 prev->state = EXIT_DEAD;
2941 switch_count = &prev->nivcsw;
2942 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2943 switch_count = &prev->nvcsw;
2944 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
2945 unlikely(signal_pending(prev))))
2946 prev->state = TASK_RUNNING;
2948 if (prev->state == TASK_UNINTERRUPTIBLE)
2949 rq->nr_uninterruptible++;
2950 deactivate_task(prev, rq);
2954 cpu = smp_processor_id();
2955 if (unlikely(!rq->nr_running)) {
2957 idle_balance(cpu, rq);
2958 if (!rq->nr_running) {
2960 rq->expired_timestamp = 0;
2961 wake_sleeping_dependent(cpu, rq);
2963 * wake_sleeping_dependent() might have released
2964 * the runqueue, so break out if we got new
2967 if (!rq->nr_running)
2971 if (dependent_sleeper(cpu, rq)) {
2976 * dependent_sleeper() releases and reacquires the runqueue
2977 * lock, hence go into the idle loop if the rq went
2980 if (unlikely(!rq->nr_running))
2985 if (unlikely(!array->nr_active)) {
2987 * Switch the active and expired arrays.
2989 schedstat_inc(rq, sched_switch);
2990 rq->active = rq->expired;
2991 rq->expired = array;
2993 rq->expired_timestamp = 0;
2994 rq->best_expired_prio = MAX_PRIO;
2997 idx = sched_find_first_bit(array->bitmap);
2998 queue = array->queue + idx;
2999 next = list_entry(queue->next, task_t, run_list);
3001 if (!rt_task(next) && next->activated > 0) {
3002 unsigned long long delta = now - next->timestamp;
3003 if (unlikely((long long)(now - next->timestamp) < 0))
3006 if (next->activated == 1)
3007 delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
3009 array = next->array;
3010 new_prio = recalc_task_prio(next, next->timestamp + delta);
3012 if (unlikely(next->prio != new_prio)) {
3013 dequeue_task(next, array);
3014 next->prio = new_prio;
3015 enqueue_task(next, array);
3017 requeue_task(next, array);
3019 next->activated = 0;
3021 if (next == rq->idle)
3022 schedstat_inc(rq, sched_goidle);
3024 prefetch_stack(next);
3025 clear_tsk_need_resched(prev);
3026 rcu_qsctr_inc(task_cpu(prev));
3028 update_cpu_clock(prev, rq, now);
3030 prev->sleep_avg -= run_time;
3031 if ((long)prev->sleep_avg <= 0)
3032 prev->sleep_avg = 0;
3033 prev->timestamp = prev->last_ran = now;
3035 sched_info_switch(prev, next);
3036 if (likely(prev != next)) {
3037 next->timestamp = now;
3042 prepare_task_switch(rq, next);
3043 prev = context_switch(rq, prev, next);
3046 * this_rq must be evaluated again because prev may have moved
3047 * CPUs since it called schedule(), thus the 'rq' on its stack
3048 * frame will be invalid.
3050 finish_task_switch(this_rq(), prev);
3052 spin_unlock_irq(&rq->lock);
3055 if (unlikely(reacquire_kernel_lock(prev) < 0))
3056 goto need_resched_nonpreemptible;
3057 preempt_enable_no_resched();
3058 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3062 EXPORT_SYMBOL(schedule);
3064 #ifdef CONFIG_PREEMPT
3066 * this is is the entry point to schedule() from in-kernel preemption
3067 * off of preempt_enable. Kernel preemptions off return from interrupt
3068 * occur there and call schedule directly.
3070 asmlinkage void __sched preempt_schedule(void)
3072 struct thread_info *ti = current_thread_info();
3073 #ifdef CONFIG_PREEMPT_BKL
3074 struct task_struct *task = current;
3075 int saved_lock_depth;
3078 * If there is a non-zero preempt_count or interrupts are disabled,
3079 * we do not want to preempt the current task. Just return..
3081 if (unlikely(ti->preempt_count || irqs_disabled()))
3085 add_preempt_count(PREEMPT_ACTIVE);
3087 * We keep the big kernel semaphore locked, but we
3088 * clear ->lock_depth so that schedule() doesnt
3089 * auto-release the semaphore:
3091 #ifdef CONFIG_PREEMPT_BKL
3092 saved_lock_depth = task->lock_depth;
3093 task->lock_depth = -1;
3096 #ifdef CONFIG_PREEMPT_BKL
3097 task->lock_depth = saved_lock_depth;
3099 sub_preempt_count(PREEMPT_ACTIVE);
3101 /* we could miss a preemption opportunity between schedule and now */
3103 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3107 EXPORT_SYMBOL(preempt_schedule);
3110 * this is is the entry point to schedule() from kernel preemption
3111 * off of irq context.
3112 * Note, that this is called and return with irqs disabled. This will
3113 * protect us against recursive calling from irq.
3115 asmlinkage void __sched preempt_schedule_irq(void)
3117 struct thread_info *ti = current_thread_info();
3118 #ifdef CONFIG_PREEMPT_BKL
3119 struct task_struct *task = current;
3120 int saved_lock_depth;
3122 /* Catch callers which need to be fixed*/
3123 BUG_ON(ti->preempt_count || !irqs_disabled());
3126 add_preempt_count(PREEMPT_ACTIVE);
3128 * We keep the big kernel semaphore locked, but we
3129 * clear ->lock_depth so that schedule() doesnt
3130 * auto-release the semaphore:
3132 #ifdef CONFIG_PREEMPT_BKL
3133 saved_lock_depth = task->lock_depth;
3134 task->lock_depth = -1;
3138 local_irq_disable();
3139 #ifdef CONFIG_PREEMPT_BKL
3140 task->lock_depth = saved_lock_depth;
3142 sub_preempt_count(PREEMPT_ACTIVE);
3144 /* we could miss a preemption opportunity between schedule and now */
3146 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3150 #endif /* CONFIG_PREEMPT */
3152 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3155 task_t *p = curr->private;
3156 return try_to_wake_up(p, mode, sync);
3159 EXPORT_SYMBOL(default_wake_function);
3162 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3163 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3164 * number) then we wake all the non-exclusive tasks and one exclusive task.
3166 * There are circumstances in which we can try to wake a task which has already
3167 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3168 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3170 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3171 int nr_exclusive, int sync, void *key)
3173 struct list_head *tmp, *next;
3175 list_for_each_safe(tmp, next, &q->task_list) {
3178 curr = list_entry(tmp, wait_queue_t, task_list);
3179 flags = curr->flags;
3180 if (curr->func(curr, mode, sync, key) &&
3181 (flags & WQ_FLAG_EXCLUSIVE) &&
3188 * __wake_up - wake up threads blocked on a waitqueue.
3190 * @mode: which threads
3191 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3192 * @key: is directly passed to the wakeup function
3194 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3195 int nr_exclusive, void *key)
3197 unsigned long flags;
3199 spin_lock_irqsave(&q->lock, flags);
3200 __wake_up_common(q, mode, nr_exclusive, 0, key);
3201 spin_unlock_irqrestore(&q->lock, flags);
3204 EXPORT_SYMBOL(__wake_up);
3207 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3209 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3211 __wake_up_common(q, mode, 1, 0, NULL);
3215 * __wake_up_sync - wake up threads blocked on a waitqueue.
3217 * @mode: which threads
3218 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3220 * The sync wakeup differs that the waker knows that it will schedule
3221 * away soon, so while the target thread will be woken up, it will not
3222 * be migrated to another CPU - ie. the two threads are 'synchronized'
3223 * with each other. This can prevent needless bouncing between CPUs.
3225 * On UP it can prevent extra preemption.
3228 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3230 unsigned long flags;
3236 if (unlikely(!nr_exclusive))
3239 spin_lock_irqsave(&q->lock, flags);
3240 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3241 spin_unlock_irqrestore(&q->lock, flags);
3243 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3245 void fastcall complete(struct completion *x)
3247 unsigned long flags;
3249 spin_lock_irqsave(&x->wait.lock, flags);
3251 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3253 spin_unlock_irqrestore(&x->wait.lock, flags);
3255 EXPORT_SYMBOL(complete);
3257 void fastcall complete_all(struct completion *x)
3259 unsigned long flags;
3261 spin_lock_irqsave(&x->wait.lock, flags);
3262 x->done += UINT_MAX/2;
3263 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3265 spin_unlock_irqrestore(&x->wait.lock, flags);
3267 EXPORT_SYMBOL(complete_all);
3269 void fastcall __sched wait_for_completion(struct completion *x)
3272 spin_lock_irq(&x->wait.lock);
3274 DECLARE_WAITQUEUE(wait, current);
3276 wait.flags |= WQ_FLAG_EXCLUSIVE;
3277 __add_wait_queue_tail(&x->wait, &wait);
3279 __set_current_state(TASK_UNINTERRUPTIBLE);
3280 spin_unlock_irq(&x->wait.lock);
3282 spin_lock_irq(&x->wait.lock);
3284 __remove_wait_queue(&x->wait, &wait);
3287 spin_unlock_irq(&x->wait.lock);
3289 EXPORT_SYMBOL(wait_for_completion);
3291 unsigned long fastcall __sched
3292 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3296 spin_lock_irq(&x->wait.lock);
3298 DECLARE_WAITQUEUE(wait, current);
3300 wait.flags |= WQ_FLAG_EXCLUSIVE;
3301 __add_wait_queue_tail(&x->wait, &wait);
3303 __set_current_state(TASK_UNINTERRUPTIBLE);
3304 spin_unlock_irq(&x->wait.lock);
3305 timeout = schedule_timeout(timeout);
3306 spin_lock_irq(&x->wait.lock);
3308 __remove_wait_queue(&x->wait, &wait);
3312 __remove_wait_queue(&x->wait, &wait);
3316 spin_unlock_irq(&x->wait.lock);
3319 EXPORT_SYMBOL(wait_for_completion_timeout);
3321 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3327 spin_lock_irq(&x->wait.lock);
3329 DECLARE_WAITQUEUE(wait, current);
3331 wait.flags |= WQ_FLAG_EXCLUSIVE;
3332 __add_wait_queue_tail(&x->wait, &wait);
3334 if (signal_pending(current)) {
3336 __remove_wait_queue(&x->wait, &wait);
3339 __set_current_state(TASK_INTERRUPTIBLE);
3340 spin_unlock_irq(&x->wait.lock);
3342 spin_lock_irq(&x->wait.lock);
3344 __remove_wait_queue(&x->wait, &wait);
3348 spin_unlock_irq(&x->wait.lock);
3352 EXPORT_SYMBOL(wait_for_completion_interruptible);
3354 unsigned long fastcall __sched
3355 wait_for_completion_interruptible_timeout(struct completion *x,
3356 unsigned long timeout)
3360 spin_lock_irq(&x->wait.lock);
3362 DECLARE_WAITQUEUE(wait, current);
3364 wait.flags |= WQ_FLAG_EXCLUSIVE;
3365 __add_wait_queue_tail(&x->wait, &wait);
3367 if (signal_pending(current)) {
3368 timeout = -ERESTARTSYS;
3369 __remove_wait_queue(&x->wait, &wait);
3372 __set_current_state(TASK_INTERRUPTIBLE);
3373 spin_unlock_irq(&x->wait.lock);
3374 timeout = schedule_timeout(timeout);
3375 spin_lock_irq(&x->wait.lock);
3377 __remove_wait_queue(&x->wait, &wait);
3381 __remove_wait_queue(&x->wait, &wait);
3385 spin_unlock_irq(&x->wait.lock);
3388 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3391 #define SLEEP_ON_VAR \
3392 unsigned long flags; \
3393 wait_queue_t wait; \
3394 init_waitqueue_entry(&wait, current);
3396 #define SLEEP_ON_HEAD \
3397 spin_lock_irqsave(&q->lock,flags); \
3398 __add_wait_queue(q, &wait); \
3399 spin_unlock(&q->lock);
3401 #define SLEEP_ON_TAIL \
3402 spin_lock_irq(&q->lock); \
3403 __remove_wait_queue(q, &wait); \
3404 spin_unlock_irqrestore(&q->lock, flags);
3406 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
3410 current->state = TASK_INTERRUPTIBLE;
3417 EXPORT_SYMBOL(interruptible_sleep_on);
3419 long fastcall __sched
3420 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3424 current->state = TASK_INTERRUPTIBLE;
3427 timeout = schedule_timeout(timeout);
3433 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3435 void fastcall __sched sleep_on(wait_queue_head_t *q)
3439 current->state = TASK_UNINTERRUPTIBLE;
3446 EXPORT_SYMBOL(sleep_on);
3448 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3452 current->state = TASK_UNINTERRUPTIBLE;
3455 timeout = schedule_timeout(timeout);
3461 EXPORT_SYMBOL(sleep_on_timeout);
3463 void set_user_nice(task_t *p, long nice)
3465 unsigned long flags;
3466 prio_array_t *array;
3468 int old_prio, new_prio, delta;
3470 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3473 * We have to be careful, if called from sys_setpriority(),
3474 * the task might be in the middle of scheduling on another CPU.
3476 rq = task_rq_lock(p, &flags);
3478 * The RT priorities are set via sched_setscheduler(), but we still
3479 * allow the 'normal' nice value to be set - but as expected
3480 * it wont have any effect on scheduling until the task is
3481 * not SCHED_NORMAL/SCHED_BATCH:
3484 p->static_prio = NICE_TO_PRIO(nice);
3489 dequeue_task(p, array);
3492 new_prio = NICE_TO_PRIO(nice);
3493 delta = new_prio - old_prio;
3494 p->static_prio = NICE_TO_PRIO(nice);
3498 enqueue_task(p, array);
3500 * If the task increased its priority or is running and
3501 * lowered its priority, then reschedule its CPU:
3503 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3504 resched_task(rq->curr);
3507 task_rq_unlock(rq, &flags);
3510 EXPORT_SYMBOL(set_user_nice);
3513 * can_nice - check if a task can reduce its nice value
3517 int can_nice(const task_t *p, const int nice)
3519 /* convert nice value [19,-20] to rlimit style value [1,40] */
3520 int nice_rlim = 20 - nice;
3521 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
3522 capable(CAP_SYS_NICE));
3525 #ifdef __ARCH_WANT_SYS_NICE
3528 * sys_nice - change the priority of the current process.
3529 * @increment: priority increment
3531 * sys_setpriority is a more generic, but much slower function that
3532 * does similar things.
3534 asmlinkage long sys_nice(int increment)
3540 * Setpriority might change our priority at the same moment.
3541 * We don't have to worry. Conceptually one call occurs first
3542 * and we have a single winner.
3544 if (increment < -40)
3549 nice = PRIO_TO_NICE(current->static_prio) + increment;
3555 if (increment < 0 && !can_nice(current, nice))
3558 retval = security_task_setnice(current, nice);
3562 set_user_nice(current, nice);
3569 * task_prio - return the priority value of a given task.
3570 * @p: the task in question.
3572 * This is the priority value as seen by users in /proc.
3573 * RT tasks are offset by -200. Normal tasks are centered
3574 * around 0, value goes from -16 to +15.
3576 int task_prio(const task_t *p)
3578 return p->prio - MAX_RT_PRIO;
3582 * task_nice - return the nice value of a given task.
3583 * @p: the task in question.
3585 int task_nice(const task_t *p)
3587 return TASK_NICE(p);
3589 EXPORT_SYMBOL_GPL(task_nice);
3592 * idle_cpu - is a given cpu idle currently?
3593 * @cpu: the processor in question.
3595 int idle_cpu(int cpu)
3597 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
3601 * idle_task - return the idle task for a given cpu.
3602 * @cpu: the processor in question.
3604 task_t *idle_task(int cpu)
3606 return cpu_rq(cpu)->idle;
3610 * find_process_by_pid - find a process with a matching PID value.
3611 * @pid: the pid in question.
3613 static inline task_t *find_process_by_pid(pid_t pid)
3615 return pid ? find_task_by_pid(pid) : current;
3618 /* Actually do priority change: must hold rq lock. */
3619 static void __setscheduler(struct task_struct *p, int policy, int prio)
3623 p->rt_priority = prio;
3624 if (policy != SCHED_NORMAL && policy != SCHED_BATCH) {
3625 p->prio = MAX_RT_PRIO-1 - p->rt_priority;
3627 p->prio = p->static_prio;
3629 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
3631 if (policy == SCHED_BATCH)
3637 * sched_setscheduler - change the scheduling policy and/or RT priority of
3639 * @p: the task in question.
3640 * @policy: new policy.
3641 * @param: structure containing the new RT priority.
3643 int sched_setscheduler(struct task_struct *p, int policy,
3644 struct sched_param *param)
3647 int oldprio, oldpolicy = -1;
3648 prio_array_t *array;
3649 unsigned long flags;
3653 /* double check policy once rq lock held */
3655 policy = oldpolicy = p->policy;
3656 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
3657 policy != SCHED_NORMAL && policy != SCHED_BATCH)
3660 * Valid priorities for SCHED_FIFO and SCHED_RR are
3661 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
3664 if (param->sched_priority < 0 ||
3665 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3666 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3668 if ((policy == SCHED_NORMAL || policy == SCHED_BATCH)
3669 != (param->sched_priority == 0))
3673 * Allow unprivileged RT tasks to decrease priority:
3675 if (!capable(CAP_SYS_NICE)) {
3677 * can't change policy, except between SCHED_NORMAL
3680 if (((policy != SCHED_NORMAL && p->policy != SCHED_BATCH) &&
3681 (policy != SCHED_BATCH && p->policy != SCHED_NORMAL)) &&
3682 !p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
3684 /* can't increase priority */
3685 if ((policy != SCHED_NORMAL && policy != SCHED_BATCH) &&
3686 param->sched_priority > p->rt_priority &&
3687 param->sched_priority >
3688 p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
3690 /* can't change other user's priorities */
3691 if ((current->euid != p->euid) &&
3692 (current->euid != p->uid))
3696 retval = security_task_setscheduler(p, policy, param);
3700 * To be able to change p->policy safely, the apropriate
3701 * runqueue lock must be held.
3703 rq = task_rq_lock(p, &flags);
3704 /* recheck policy now with rq lock held */
3705 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3706 policy = oldpolicy = -1;
3707 task_rq_unlock(rq, &flags);
3712 deactivate_task(p, rq);
3714 __setscheduler(p, policy, param->sched_priority);
3716 __activate_task(p, rq);
3718 * Reschedule if we are currently running on this runqueue and
3719 * our priority decreased, or if we are not currently running on
3720 * this runqueue and our priority is higher than the current's
3722 if (task_running(rq, p)) {
3723 if (p->prio > oldprio)
3724 resched_task(rq->curr);
3725 } else if (TASK_PREEMPTS_CURR(p, rq))
3726 resched_task(rq->curr);
3728 task_rq_unlock(rq, &flags);
3731 EXPORT_SYMBOL_GPL(sched_setscheduler);
3734 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3737 struct sched_param lparam;
3738 struct task_struct *p;
3740 if (!param || pid < 0)
3742 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3744 read_lock_irq(&tasklist_lock);
3745 p = find_process_by_pid(pid);
3747 read_unlock_irq(&tasklist_lock);
3750 retval = sched_setscheduler(p, policy, &lparam);
3751 read_unlock_irq(&tasklist_lock);
3756 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3757 * @pid: the pid in question.
3758 * @policy: new policy.
3759 * @param: structure containing the new RT priority.
3761 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
3762 struct sched_param __user *param)
3764 /* negative values for policy are not valid */
3768 return do_sched_setscheduler(pid, policy, param);
3772 * sys_sched_setparam - set/change the RT priority of a thread
3773 * @pid: the pid in question.
3774 * @param: structure containing the new RT priority.
3776 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
3778 return do_sched_setscheduler(pid, -1, param);
3782 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3783 * @pid: the pid in question.
3785 asmlinkage long sys_sched_getscheduler(pid_t pid)
3787 int retval = -EINVAL;
3794 read_lock(&tasklist_lock);
3795 p = find_process_by_pid(pid);
3797 retval = security_task_getscheduler(p);
3801 read_unlock(&tasklist_lock);
3808 * sys_sched_getscheduler - get the RT priority of a thread
3809 * @pid: the pid in question.
3810 * @param: structure containing the RT priority.
3812 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
3814 struct sched_param lp;
3815 int retval = -EINVAL;
3818 if (!param || pid < 0)
3821 read_lock(&tasklist_lock);
3822 p = find_process_by_pid(pid);
3827 retval = security_task_getscheduler(p);
3831 lp.sched_priority = p->rt_priority;
3832 read_unlock(&tasklist_lock);
3835 * This one might sleep, we cannot do it with a spinlock held ...
3837 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3843 read_unlock(&tasklist_lock);
3847 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
3851 cpumask_t cpus_allowed;
3854 read_lock(&tasklist_lock);
3856 p = find_process_by_pid(pid);
3858 read_unlock(&tasklist_lock);
3859 unlock_cpu_hotplug();
3864 * It is not safe to call set_cpus_allowed with the
3865 * tasklist_lock held. We will bump the task_struct's
3866 * usage count and then drop tasklist_lock.
3869 read_unlock(&tasklist_lock);
3872 if ((current->euid != p->euid) && (current->euid != p->uid) &&
3873 !capable(CAP_SYS_NICE))
3876 cpus_allowed = cpuset_cpus_allowed(p);
3877 cpus_and(new_mask, new_mask, cpus_allowed);
3878 retval = set_cpus_allowed(p, new_mask);
3882 unlock_cpu_hotplug();
3886 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3887 cpumask_t *new_mask)
3889 if (len < sizeof(cpumask_t)) {
3890 memset(new_mask, 0, sizeof(cpumask_t));
3891 } else if (len > sizeof(cpumask_t)) {
3892 len = sizeof(cpumask_t);
3894 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3898 * sys_sched_setaffinity - set the cpu affinity of a process
3899 * @pid: pid of the process
3900 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3901 * @user_mask_ptr: user-space pointer to the new cpu mask
3903 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
3904 unsigned long __user *user_mask_ptr)
3909 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
3913 return sched_setaffinity(pid, new_mask);
3917 * Represents all cpu's present in the system
3918 * In systems capable of hotplug, this map could dynamically grow
3919 * as new cpu's are detected in the system via any platform specific
3920 * method, such as ACPI for e.g.
3923 cpumask_t cpu_present_map __read_mostly;
3924 EXPORT_SYMBOL(cpu_present_map);
3927 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
3928 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
3931 long sched_getaffinity(pid_t pid, cpumask_t *mask)
3937 read_lock(&tasklist_lock);
3940 p = find_process_by_pid(pid);
3945 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
3948 read_unlock(&tasklist_lock);
3949 unlock_cpu_hotplug();
3957 * sys_sched_getaffinity - get the cpu affinity of a process
3958 * @pid: pid of the process
3959 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3960 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3962 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
3963 unsigned long __user *user_mask_ptr)
3968 if (len < sizeof(cpumask_t))
3971 ret = sched_getaffinity(pid, &mask);
3975 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
3978 return sizeof(cpumask_t);
3982 * sys_sched_yield - yield the current processor to other threads.
3984 * this function yields the current CPU by moving the calling thread
3985 * to the expired array. If there are no other threads running on this
3986 * CPU then this function will return.
3988 asmlinkage long sys_sched_yield(void)
3990 runqueue_t *rq = this_rq_lock();
3991 prio_array_t *array = current->array;
3992 prio_array_t *target = rq->expired;
3994 schedstat_inc(rq, yld_cnt);
3996 * We implement yielding by moving the task into the expired
3999 * (special rule: RT tasks will just roundrobin in the active
4002 if (rt_task(current))
4003 target = rq->active;
4005 if (array->nr_active == 1) {
4006 schedstat_inc(rq, yld_act_empty);
4007 if (!rq->expired->nr_active)
4008 schedstat_inc(rq, yld_both_empty);
4009 } else if (!rq->expired->nr_active)
4010 schedstat_inc(rq, yld_exp_empty);
4012 if (array != target) {
4013 dequeue_task(current, array);
4014 enqueue_task(current, target);
4017 * requeue_task is cheaper so perform that if possible.
4019 requeue_task(current, array);
4022 * Since we are going to call schedule() anyway, there's
4023 * no need to preempt or enable interrupts:
4025 __release(rq->lock);
4026 _raw_spin_unlock(&rq->lock);
4027 preempt_enable_no_resched();
4034 static inline void __cond_resched(void)
4037 * The BKS might be reacquired before we have dropped
4038 * PREEMPT_ACTIVE, which could trigger a second
4039 * cond_resched() call.
4041 if (unlikely(preempt_count()))
4043 if (unlikely(system_state != SYSTEM_RUNNING))
4046 add_preempt_count(PREEMPT_ACTIVE);
4048 sub_preempt_count(PREEMPT_ACTIVE);
4049 } while (need_resched());
4052 int __sched cond_resched(void)
4054 if (need_resched()) {
4061 EXPORT_SYMBOL(cond_resched);
4064 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4065 * call schedule, and on return reacquire the lock.
4067 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4068 * operations here to prevent schedule() from being called twice (once via
4069 * spin_unlock(), once by hand).
4071 int cond_resched_lock(spinlock_t *lock)
4075 if (need_lockbreak(lock)) {
4081 if (need_resched()) {
4082 _raw_spin_unlock(lock);
4083 preempt_enable_no_resched();
4091 EXPORT_SYMBOL(cond_resched_lock);
4093 int __sched cond_resched_softirq(void)
4095 BUG_ON(!in_softirq());
4097 if (need_resched()) {
4098 __local_bh_enable();
4106 EXPORT_SYMBOL(cond_resched_softirq);
4110 * yield - yield the current processor to other threads.
4112 * this is a shortcut for kernel-space yielding - it marks the
4113 * thread runnable and calls sys_sched_yield().
4115 void __sched yield(void)
4117 set_current_state(TASK_RUNNING);
4121 EXPORT_SYMBOL(yield);
4124 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4125 * that process accounting knows that this is a task in IO wait state.
4127 * But don't do that if it is a deliberate, throttling IO wait (this task
4128 * has set its backing_dev_info: the queue against which it should throttle)
4130 void __sched io_schedule(void)
4132 struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id());
4134 atomic_inc(&rq->nr_iowait);
4136 atomic_dec(&rq->nr_iowait);
4139 EXPORT_SYMBOL(io_schedule);
4141 long __sched io_schedule_timeout(long timeout)
4143 struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id());
4146 atomic_inc(&rq->nr_iowait);
4147 ret = schedule_timeout(timeout);
4148 atomic_dec(&rq->nr_iowait);
4153 * sys_sched_get_priority_max - return maximum RT priority.
4154 * @policy: scheduling class.
4156 * this syscall returns the maximum rt_priority that can be used
4157 * by a given scheduling class.
4159 asmlinkage long sys_sched_get_priority_max(int policy)
4166 ret = MAX_USER_RT_PRIO-1;
4177 * sys_sched_get_priority_min - return minimum RT priority.
4178 * @policy: scheduling class.
4180 * this syscall returns the minimum rt_priority that can be used
4181 * by a given scheduling class.
4183 asmlinkage long sys_sched_get_priority_min(int policy)
4200 * sys_sched_rr_get_interval - return the default timeslice of a process.
4201 * @pid: pid of the process.
4202 * @interval: userspace pointer to the timeslice value.
4204 * this syscall writes the default timeslice value of a given process
4205 * into the user-space timespec buffer. A value of '0' means infinity.
4208 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4210 int retval = -EINVAL;
4218 read_lock(&tasklist_lock);
4219 p = find_process_by_pid(pid);
4223 retval = security_task_getscheduler(p);
4227 jiffies_to_timespec(p->policy & SCHED_FIFO ?
4228 0 : task_timeslice(p), &t);
4229 read_unlock(&tasklist_lock);
4230 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4234 read_unlock(&tasklist_lock);
4238 static inline struct task_struct *eldest_child(struct task_struct *p)
4240 if (list_empty(&p->children)) return NULL;
4241 return list_entry(p->children.next,struct task_struct,sibling);
4244 static inline struct task_struct *older_sibling(struct task_struct *p)
4246 if (p->sibling.prev==&p->parent->children) return NULL;
4247 return list_entry(p->sibling.prev,struct task_struct,sibling);
4250 static inline struct task_struct *younger_sibling(struct task_struct *p)
4252 if (p->sibling.next==&p->parent->children) return NULL;
4253 return list_entry(p->sibling.next,struct task_struct,sibling);
4256 static void show_task(task_t *p)
4260 unsigned long free = 0;
4261 static const char *stat_nam[] = { "R", "S", "D", "T", "t", "Z", "X" };
4263 printk("%-13.13s ", p->comm);
4264 state = p->state ? __ffs(p->state) + 1 : 0;
4265 if (state < ARRAY_SIZE(stat_nam))
4266 printk(stat_nam[state]);
4269 #if (BITS_PER_LONG == 32)
4270 if (state == TASK_RUNNING)
4271 printk(" running ");
4273 printk(" %08lX ", thread_saved_pc(p));
4275 if (state == TASK_RUNNING)
4276 printk(" running task ");
4278 printk(" %016lx ", thread_saved_pc(p));
4280 #ifdef CONFIG_DEBUG_STACK_USAGE
4282 unsigned long *n = end_of_stack(p);
4285 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4288 printk("%5lu %5d %6d ", free, p->pid, p->parent->pid);
4289 if ((relative = eldest_child(p)))
4290 printk("%5d ", relative->pid);
4293 if ((relative = younger_sibling(p)))
4294 printk("%7d", relative->pid);
4297 if ((relative = older_sibling(p)))
4298 printk(" %5d", relative->pid);
4302 printk(" (L-TLB)\n");
4304 printk(" (NOTLB)\n");
4306 if (state != TASK_RUNNING)
4307 show_stack(p, NULL);
4310 void show_state(void)
4314 #if (BITS_PER_LONG == 32)
4317 printk(" task PC pid father child younger older\n");
4321 printk(" task PC pid father child younger older\n");
4323 read_lock(&tasklist_lock);
4324 do_each_thread(g, p) {
4326 * reset the NMI-timeout, listing all files on a slow
4327 * console might take alot of time:
4329 touch_nmi_watchdog();
4331 } while_each_thread(g, p);
4333 read_unlock(&tasklist_lock);
4334 mutex_debug_show_all_locks();
4338 * init_idle - set up an idle thread for a given CPU
4339 * @idle: task in question
4340 * @cpu: cpu the idle task belongs to
4342 * NOTE: this function does not set the idle thread's NEED_RESCHED
4343 * flag, to make booting more robust.
4345 void __devinit init_idle(task_t *idle, int cpu)
4347 runqueue_t *rq = cpu_rq(cpu);
4348 unsigned long flags;
4350 idle->timestamp = sched_clock();
4351 idle->sleep_avg = 0;
4353 idle->prio = MAX_PRIO;
4354 idle->state = TASK_RUNNING;
4355 idle->cpus_allowed = cpumask_of_cpu(cpu);
4356 set_task_cpu(idle, cpu);
4358 spin_lock_irqsave(&rq->lock, flags);
4359 rq->curr = rq->idle = idle;
4360 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4363 spin_unlock_irqrestore(&rq->lock, flags);
4365 /* Set the preempt count _outside_ the spinlocks! */
4366 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4367 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4369 task_thread_info(idle)->preempt_count = 0;
4374 * In a system that switches off the HZ timer nohz_cpu_mask
4375 * indicates which cpus entered this state. This is used
4376 * in the rcu update to wait only for active cpus. For system
4377 * which do not switch off the HZ timer nohz_cpu_mask should
4378 * always be CPU_MASK_NONE.
4380 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4384 * This is how migration works:
4386 * 1) we queue a migration_req_t structure in the source CPU's
4387 * runqueue and wake up that CPU's migration thread.
4388 * 2) we down() the locked semaphore => thread blocks.
4389 * 3) migration thread wakes up (implicitly it forces the migrated
4390 * thread off the CPU)
4391 * 4) it gets the migration request and checks whether the migrated
4392 * task is still in the wrong runqueue.
4393 * 5) if it's in the wrong runqueue then the migration thread removes
4394 * it and puts it into the right queue.
4395 * 6) migration thread up()s the semaphore.
4396 * 7) we wake up and the migration is done.
4400 * Change a given task's CPU affinity. Migrate the thread to a
4401 * proper CPU and schedule it away if the CPU it's executing on
4402 * is removed from the allowed bitmask.
4404 * NOTE: the caller must have a valid reference to the task, the
4405 * task must not exit() & deallocate itself prematurely. The
4406 * call is not atomic; no spinlocks may be held.
4408 int set_cpus_allowed(task_t *p, cpumask_t new_mask)
4410 unsigned long flags;
4412 migration_req_t req;
4415 rq = task_rq_lock(p, &flags);
4416 if (!cpus_intersects(new_mask, cpu_online_map)) {
4421 p->cpus_allowed = new_mask;
4422 /* Can the task run on the task's current CPU? If so, we're done */
4423 if (cpu_isset(task_cpu(p), new_mask))
4426 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4427 /* Need help from migration thread: drop lock and wait. */
4428 task_rq_unlock(rq, &flags);
4429 wake_up_process(rq->migration_thread);
4430 wait_for_completion(&req.done);
4431 tlb_migrate_finish(p->mm);
4435 task_rq_unlock(rq, &flags);
4439 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4442 * Move (not current) task off this cpu, onto dest cpu. We're doing
4443 * this because either it can't run here any more (set_cpus_allowed()
4444 * away from this CPU, or CPU going down), or because we're
4445 * attempting to rebalance this task on exec (sched_exec).
4447 * So we race with normal scheduler movements, but that's OK, as long
4448 * as the task is no longer on this CPU.
4450 static void __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4452 runqueue_t *rq_dest, *rq_src;
4454 if (unlikely(cpu_is_offline(dest_cpu)))
4457 rq_src = cpu_rq(src_cpu);
4458 rq_dest = cpu_rq(dest_cpu);
4460 double_rq_lock(rq_src, rq_dest);
4461 /* Already moved. */
4462 if (task_cpu(p) != src_cpu)
4464 /* Affinity changed (again). */
4465 if (!cpu_isset(dest_cpu, p->cpus_allowed))
4468 set_task_cpu(p, dest_cpu);
4471 * Sync timestamp with rq_dest's before activating.
4472 * The same thing could be achieved by doing this step
4473 * afterwards, and pretending it was a local activate.
4474 * This way is cleaner and logically correct.
4476 p->timestamp = p->timestamp - rq_src->timestamp_last_tick
4477 + rq_dest->timestamp_last_tick;
4478 deactivate_task(p, rq_src);
4479 activate_task(p, rq_dest, 0);
4480 if (TASK_PREEMPTS_CURR(p, rq_dest))
4481 resched_task(rq_dest->curr);
4485 double_rq_unlock(rq_src, rq_dest);
4489 * migration_thread - this is a highprio system thread that performs
4490 * thread migration by bumping thread off CPU then 'pushing' onto
4493 static int migration_thread(void *data)
4496 int cpu = (long)data;
4499 BUG_ON(rq->migration_thread != current);
4501 set_current_state(TASK_INTERRUPTIBLE);
4502 while (!kthread_should_stop()) {
4503 struct list_head *head;
4504 migration_req_t *req;
4508 spin_lock_irq(&rq->lock);
4510 if (cpu_is_offline(cpu)) {
4511 spin_unlock_irq(&rq->lock);
4515 if (rq->active_balance) {
4516 active_load_balance(rq, cpu);
4517 rq->active_balance = 0;
4520 head = &rq->migration_queue;
4522 if (list_empty(head)) {
4523 spin_unlock_irq(&rq->lock);
4525 set_current_state(TASK_INTERRUPTIBLE);
4528 req = list_entry(head->next, migration_req_t, list);
4529 list_del_init(head->next);
4531 spin_unlock(&rq->lock);
4532 __migrate_task(req->task, cpu, req->dest_cpu);
4535 complete(&req->done);
4537 __set_current_state(TASK_RUNNING);
4541 /* Wait for kthread_stop */
4542 set_current_state(TASK_INTERRUPTIBLE);
4543 while (!kthread_should_stop()) {
4545 set_current_state(TASK_INTERRUPTIBLE);
4547 __set_current_state(TASK_RUNNING);
4551 #ifdef CONFIG_HOTPLUG_CPU
4552 /* Figure out where task on dead CPU should go, use force if neccessary. */
4553 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *tsk)
4559 mask = node_to_cpumask(cpu_to_node(dead_cpu));
4560 cpus_and(mask, mask, tsk->cpus_allowed);
4561 dest_cpu = any_online_cpu(mask);
4563 /* On any allowed CPU? */
4564 if (dest_cpu == NR_CPUS)
4565 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4567 /* No more Mr. Nice Guy. */
4568 if (dest_cpu == NR_CPUS) {
4569 cpus_setall(tsk->cpus_allowed);
4570 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4573 * Don't tell them about moving exiting tasks or
4574 * kernel threads (both mm NULL), since they never
4577 if (tsk->mm && printk_ratelimit())
4578 printk(KERN_INFO "process %d (%s) no "
4579 "longer affine to cpu%d\n",
4580 tsk->pid, tsk->comm, dead_cpu);
4582 __migrate_task(tsk, dead_cpu, dest_cpu);
4586 * While a dead CPU has no uninterruptible tasks queued at this point,
4587 * it might still have a nonzero ->nr_uninterruptible counter, because
4588 * for performance reasons the counter is not stricly tracking tasks to
4589 * their home CPUs. So we just add the counter to another CPU's counter,
4590 * to keep the global sum constant after CPU-down:
4592 static void migrate_nr_uninterruptible(runqueue_t *rq_src)
4594 runqueue_t *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
4595 unsigned long flags;
4597 local_irq_save(flags);
4598 double_rq_lock(rq_src, rq_dest);
4599 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
4600 rq_src->nr_uninterruptible = 0;
4601 double_rq_unlock(rq_src, rq_dest);
4602 local_irq_restore(flags);
4605 /* Run through task list and migrate tasks from the dead cpu. */
4606 static void migrate_live_tasks(int src_cpu)
4608 struct task_struct *tsk, *t;
4610 write_lock_irq(&tasklist_lock);
4612 do_each_thread(t, tsk) {
4616 if (task_cpu(tsk) == src_cpu)
4617 move_task_off_dead_cpu(src_cpu, tsk);
4618 } while_each_thread(t, tsk);
4620 write_unlock_irq(&tasklist_lock);
4623 /* Schedules idle task to be the next runnable task on current CPU.
4624 * It does so by boosting its priority to highest possible and adding it to
4625 * the _front_ of runqueue. Used by CPU offline code.
4627 void sched_idle_next(void)
4629 int cpu = smp_processor_id();
4630 runqueue_t *rq = this_rq();
4631 struct task_struct *p = rq->idle;
4632 unsigned long flags;
4634 /* cpu has to be offline */
4635 BUG_ON(cpu_online(cpu));
4637 /* Strictly not necessary since rest of the CPUs are stopped by now
4638 * and interrupts disabled on current cpu.
4640 spin_lock_irqsave(&rq->lock, flags);
4642 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4643 /* Add idle task to _front_ of it's priority queue */
4644 __activate_idle_task(p, rq);
4646 spin_unlock_irqrestore(&rq->lock, flags);
4649 /* Ensures that the idle task is using init_mm right before its cpu goes
4652 void idle_task_exit(void)
4654 struct mm_struct *mm = current->active_mm;
4656 BUG_ON(cpu_online(smp_processor_id()));
4659 switch_mm(mm, &init_mm, current);
4663 static void migrate_dead(unsigned int dead_cpu, task_t *tsk)
4665 struct runqueue *rq = cpu_rq(dead_cpu);
4667 /* Must be exiting, otherwise would be on tasklist. */
4668 BUG_ON(tsk->exit_state != EXIT_ZOMBIE && tsk->exit_state != EXIT_DEAD);
4670 /* Cannot have done final schedule yet: would have vanished. */
4671 BUG_ON(tsk->flags & PF_DEAD);
4673 get_task_struct(tsk);
4676 * Drop lock around migration; if someone else moves it,
4677 * that's OK. No task can be added to this CPU, so iteration is
4680 spin_unlock_irq(&rq->lock);
4681 move_task_off_dead_cpu(dead_cpu, tsk);
4682 spin_lock_irq(&rq->lock);
4684 put_task_struct(tsk);
4687 /* release_task() removes task from tasklist, so we won't find dead tasks. */
4688 static void migrate_dead_tasks(unsigned int dead_cpu)
4691 struct runqueue *rq = cpu_rq(dead_cpu);
4693 for (arr = 0; arr < 2; arr++) {
4694 for (i = 0; i < MAX_PRIO; i++) {
4695 struct list_head *list = &rq->arrays[arr].queue[i];
4696 while (!list_empty(list))
4697 migrate_dead(dead_cpu,
4698 list_entry(list->next, task_t,
4703 #endif /* CONFIG_HOTPLUG_CPU */
4706 * migration_call - callback that gets triggered when a CPU is added.
4707 * Here we can start up the necessary migration thread for the new CPU.
4709 static int migration_call(struct notifier_block *nfb, unsigned long action,
4712 int cpu = (long)hcpu;
4713 struct task_struct *p;
4714 struct runqueue *rq;
4715 unsigned long flags;
4718 case CPU_UP_PREPARE:
4719 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
4722 p->flags |= PF_NOFREEZE;
4723 kthread_bind(p, cpu);
4724 /* Must be high prio: stop_machine expects to yield to it. */
4725 rq = task_rq_lock(p, &flags);
4726 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4727 task_rq_unlock(rq, &flags);
4728 cpu_rq(cpu)->migration_thread = p;
4731 /* Strictly unneccessary, as first user will wake it. */
4732 wake_up_process(cpu_rq(cpu)->migration_thread);
4734 #ifdef CONFIG_HOTPLUG_CPU
4735 case CPU_UP_CANCELED:
4736 /* Unbind it from offline cpu so it can run. Fall thru. */
4737 kthread_bind(cpu_rq(cpu)->migration_thread,
4738 any_online_cpu(cpu_online_map));
4739 kthread_stop(cpu_rq(cpu)->migration_thread);
4740 cpu_rq(cpu)->migration_thread = NULL;
4743 migrate_live_tasks(cpu);
4745 kthread_stop(rq->migration_thread);
4746 rq->migration_thread = NULL;
4747 /* Idle task back to normal (off runqueue, low prio) */
4748 rq = task_rq_lock(rq->idle, &flags);
4749 deactivate_task(rq->idle, rq);
4750 rq->idle->static_prio = MAX_PRIO;
4751 __setscheduler(rq->idle, SCHED_NORMAL, 0);
4752 migrate_dead_tasks(cpu);
4753 task_rq_unlock(rq, &flags);
4754 migrate_nr_uninterruptible(rq);
4755 BUG_ON(rq->nr_running != 0);
4757 /* No need to migrate the tasks: it was best-effort if
4758 * they didn't do lock_cpu_hotplug(). Just wake up
4759 * the requestors. */
4760 spin_lock_irq(&rq->lock);
4761 while (!list_empty(&rq->migration_queue)) {
4762 migration_req_t *req;
4763 req = list_entry(rq->migration_queue.next,
4764 migration_req_t, list);
4765 list_del_init(&req->list);
4766 complete(&req->done);
4768 spin_unlock_irq(&rq->lock);
4775 /* Register at highest priority so that task migration (migrate_all_tasks)
4776 * happens before everything else.
4778 static struct notifier_block __devinitdata migration_notifier = {
4779 .notifier_call = migration_call,
4783 int __init migration_init(void)
4785 void *cpu = (void *)(long)smp_processor_id();
4786 /* Start one for boot CPU. */
4787 migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
4788 migration_call(&migration_notifier, CPU_ONLINE, cpu);
4789 register_cpu_notifier(&migration_notifier);
4795 #undef SCHED_DOMAIN_DEBUG
4796 #ifdef SCHED_DOMAIN_DEBUG
4797 static void sched_domain_debug(struct sched_domain *sd, int cpu)
4802 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
4806 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
4811 struct sched_group *group = sd->groups;
4812 cpumask_t groupmask;
4814 cpumask_scnprintf(str, NR_CPUS, sd->span);
4815 cpus_clear(groupmask);
4818 for (i = 0; i < level + 1; i++)
4820 printk("domain %d: ", level);
4822 if (!(sd->flags & SD_LOAD_BALANCE)) {
4823 printk("does not load-balance\n");
4825 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent");
4829 printk("span %s\n", str);
4831 if (!cpu_isset(cpu, sd->span))
4832 printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
4833 if (!cpu_isset(cpu, group->cpumask))
4834 printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
4837 for (i = 0; i < level + 2; i++)
4843 printk(KERN_ERR "ERROR: group is NULL\n");
4847 if (!group->cpu_power) {
4849 printk(KERN_ERR "ERROR: domain->cpu_power not set\n");
4852 if (!cpus_weight(group->cpumask)) {
4854 printk(KERN_ERR "ERROR: empty group\n");
4857 if (cpus_intersects(groupmask, group->cpumask)) {
4859 printk(KERN_ERR "ERROR: repeated CPUs\n");
4862 cpus_or(groupmask, groupmask, group->cpumask);
4864 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
4867 group = group->next;
4868 } while (group != sd->groups);
4871 if (!cpus_equal(sd->span, groupmask))
4872 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
4878 if (!cpus_subset(groupmask, sd->span))
4879 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
4885 #define sched_domain_debug(sd, cpu) {}
4888 static int sd_degenerate(struct sched_domain *sd)
4890 if (cpus_weight(sd->span) == 1)
4893 /* Following flags need at least 2 groups */
4894 if (sd->flags & (SD_LOAD_BALANCE |
4895 SD_BALANCE_NEWIDLE |
4898 if (sd->groups != sd->groups->next)
4902 /* Following flags don't use groups */
4903 if (sd->flags & (SD_WAKE_IDLE |
4911 static int sd_parent_degenerate(struct sched_domain *sd,
4912 struct sched_domain *parent)
4914 unsigned long cflags = sd->flags, pflags = parent->flags;
4916 if (sd_degenerate(parent))
4919 if (!cpus_equal(sd->span, parent->span))
4922 /* Does parent contain flags not in child? */
4923 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
4924 if (cflags & SD_WAKE_AFFINE)
4925 pflags &= ~SD_WAKE_BALANCE;
4926 /* Flags needing groups don't count if only 1 group in parent */
4927 if (parent->groups == parent->groups->next) {
4928 pflags &= ~(SD_LOAD_BALANCE |
4929 SD_BALANCE_NEWIDLE |
4933 if (~cflags & pflags)
4940 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
4941 * hold the hotplug lock.
4943 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
4945 runqueue_t *rq = cpu_rq(cpu);
4946 struct sched_domain *tmp;
4948 /* Remove the sched domains which do not contribute to scheduling. */
4949 for (tmp = sd; tmp; tmp = tmp->parent) {
4950 struct sched_domain *parent = tmp->parent;
4953 if (sd_parent_degenerate(tmp, parent))
4954 tmp->parent = parent->parent;
4957 if (sd && sd_degenerate(sd))
4960 sched_domain_debug(sd, cpu);
4962 rcu_assign_pointer(rq->sd, sd);
4965 /* cpus with isolated domains */
4966 static cpumask_t __devinitdata cpu_isolated_map = CPU_MASK_NONE;
4968 /* Setup the mask of cpus configured for isolated domains */
4969 static int __init isolated_cpu_setup(char *str)
4971 int ints[NR_CPUS], i;
4973 str = get_options(str, ARRAY_SIZE(ints), ints);
4974 cpus_clear(cpu_isolated_map);
4975 for (i = 1; i <= ints[0]; i++)
4976 if (ints[i] < NR_CPUS)
4977 cpu_set(ints[i], cpu_isolated_map);
4981 __setup ("isolcpus=", isolated_cpu_setup);
4984 * init_sched_build_groups takes an array of groups, the cpumask we wish
4985 * to span, and a pointer to a function which identifies what group a CPU
4986 * belongs to. The return value of group_fn must be a valid index into the
4987 * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
4988 * keep track of groups covered with a cpumask_t).
4990 * init_sched_build_groups will build a circular linked list of the groups
4991 * covered by the given span, and will set each group's ->cpumask correctly,
4992 * and ->cpu_power to 0.
4994 static void init_sched_build_groups(struct sched_group groups[], cpumask_t span,
4995 int (*group_fn)(int cpu))
4997 struct sched_group *first = NULL, *last = NULL;
4998 cpumask_t covered = CPU_MASK_NONE;
5001 for_each_cpu_mask(i, span) {
5002 int group = group_fn(i);
5003 struct sched_group *sg = &groups[group];
5006 if (cpu_isset(i, covered))
5009 sg->cpumask = CPU_MASK_NONE;
5012 for_each_cpu_mask(j, span) {
5013 if (group_fn(j) != group)
5016 cpu_set(j, covered);
5017 cpu_set(j, sg->cpumask);
5028 #define SD_NODES_PER_DOMAIN 16
5031 * Self-tuning task migration cost measurement between source and target CPUs.
5033 * This is done by measuring the cost of manipulating buffers of varying
5034 * sizes. For a given buffer-size here are the steps that are taken:
5036 * 1) the source CPU reads+dirties a shared buffer
5037 * 2) the target CPU reads+dirties the same shared buffer
5039 * We measure how long they take, in the following 4 scenarios:
5041 * - source: CPU1, target: CPU2 | cost1
5042 * - source: CPU2, target: CPU1 | cost2
5043 * - source: CPU1, target: CPU1 | cost3
5044 * - source: CPU2, target: CPU2 | cost4
5046 * We then calculate the cost3+cost4-cost1-cost2 difference - this is
5047 * the cost of migration.
5049 * We then start off from a small buffer-size and iterate up to larger
5050 * buffer sizes, in 5% steps - measuring each buffer-size separately, and
5051 * doing a maximum search for the cost. (The maximum cost for a migration
5052 * normally occurs when the working set size is around the effective cache
5055 #define SEARCH_SCOPE 2
5056 #define MIN_CACHE_SIZE (64*1024U)
5057 #define DEFAULT_CACHE_SIZE (5*1024*1024U)
5058 #define ITERATIONS 1
5059 #define SIZE_THRESH 130
5060 #define COST_THRESH 130
5063 * The migration cost is a function of 'domain distance'. Domain
5064 * distance is the number of steps a CPU has to iterate down its
5065 * domain tree to share a domain with the other CPU. The farther
5066 * two CPUs are from each other, the larger the distance gets.
5068 * Note that we use the distance only to cache measurement results,
5069 * the distance value is not used numerically otherwise. When two
5070 * CPUs have the same distance it is assumed that the migration
5071 * cost is the same. (this is a simplification but quite practical)
5073 #define MAX_DOMAIN_DISTANCE 32
5075 static unsigned long long migration_cost[MAX_DOMAIN_DISTANCE] =
5076 { [ 0 ... MAX_DOMAIN_DISTANCE-1 ] =
5078 * Architectures may override the migration cost and thus avoid
5079 * boot-time calibration. Unit is nanoseconds. Mostly useful for
5080 * virtualized hardware:
5082 #ifdef CONFIG_DEFAULT_MIGRATION_COST
5083 CONFIG_DEFAULT_MIGRATION_COST
5090 * Allow override of migration cost - in units of microseconds.
5091 * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost
5092 * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs:
5094 static int __init migration_cost_setup(char *str)
5096 int ints[MAX_DOMAIN_DISTANCE+1], i;
5098 str = get_options(str, ARRAY_SIZE(ints), ints);
5100 printk("#ints: %d\n", ints[0]);
5101 for (i = 1; i <= ints[0]; i++) {
5102 migration_cost[i-1] = (unsigned long long)ints[i]*1000;
5103 printk("migration_cost[%d]: %Ld\n", i-1, migration_cost[i-1]);
5108 __setup ("migration_cost=", migration_cost_setup);
5111 * Global multiplier (divisor) for migration-cutoff values,
5112 * in percentiles. E.g. use a value of 150 to get 1.5 times
5113 * longer cache-hot cutoff times.
5115 * (We scale it from 100 to 128 to long long handling easier.)
5118 #define MIGRATION_FACTOR_SCALE 128
5120 static unsigned int migration_factor = MIGRATION_FACTOR_SCALE;
5122 static int __init setup_migration_factor(char *str)
5124 get_option(&str, &migration_factor);
5125 migration_factor = migration_factor * MIGRATION_FACTOR_SCALE / 100;
5129 __setup("migration_factor=", setup_migration_factor);
5132 * Estimated distance of two CPUs, measured via the number of domains
5133 * we have to pass for the two CPUs to be in the same span:
5135 static unsigned long domain_distance(int cpu1, int cpu2)
5137 unsigned long distance = 0;
5138 struct sched_domain *sd;
5140 for_each_domain(cpu1, sd) {
5141 WARN_ON(!cpu_isset(cpu1, sd->span));
5142 if (cpu_isset(cpu2, sd->span))
5146 if (distance >= MAX_DOMAIN_DISTANCE) {
5148 distance = MAX_DOMAIN_DISTANCE-1;
5154 static unsigned int migration_debug;
5156 static int __init setup_migration_debug(char *str)
5158 get_option(&str, &migration_debug);
5162 __setup("migration_debug=", setup_migration_debug);
5165 * Maximum cache-size that the scheduler should try to measure.
5166 * Architectures with larger caches should tune this up during
5167 * bootup. Gets used in the domain-setup code (i.e. during SMP
5170 unsigned int max_cache_size;
5172 static int __init setup_max_cache_size(char *str)
5174 get_option(&str, &max_cache_size);
5178 __setup("max_cache_size=", setup_max_cache_size);
5181 * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This
5182 * is the operation that is timed, so we try to generate unpredictable
5183 * cachemisses that still end up filling the L2 cache:
5185 static void touch_cache(void *__cache, unsigned long __size)
5187 unsigned long size = __size/sizeof(long), chunk1 = size/3,
5189 unsigned long *cache = __cache;
5192 for (i = 0; i < size/6; i += 8) {
5195 case 1: cache[size-1-i]++;
5196 case 2: cache[chunk1-i]++;
5197 case 3: cache[chunk1+i]++;
5198 case 4: cache[chunk2-i]++;
5199 case 5: cache[chunk2+i]++;
5205 * Measure the cache-cost of one task migration. Returns in units of nsec.
5207 static unsigned long long measure_one(void *cache, unsigned long size,
5208 int source, int target)
5210 cpumask_t mask, saved_mask;
5211 unsigned long long t0, t1, t2, t3, cost;
5213 saved_mask = current->cpus_allowed;
5216 * Flush source caches to RAM and invalidate them:
5221 * Migrate to the source CPU:
5223 mask = cpumask_of_cpu(source);
5224 set_cpus_allowed(current, mask);
5225 WARN_ON(smp_processor_id() != source);
5228 * Dirty the working set:
5231 touch_cache(cache, size);
5235 * Migrate to the target CPU, dirty the L2 cache and access
5236 * the shared buffer. (which represents the working set
5237 * of a migrated task.)
5239 mask = cpumask_of_cpu(target);
5240 set_cpus_allowed(current, mask);
5241 WARN_ON(smp_processor_id() != target);
5244 touch_cache(cache, size);
5247 cost = t1-t0 + t3-t2;
5249 if (migration_debug >= 2)
5250 printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n",
5251 source, target, t1-t0, t1-t0, t3-t2, cost);
5253 * Flush target caches to RAM and invalidate them:
5257 set_cpus_allowed(current, saved_mask);
5263 * Measure a series of task migrations and return the average
5264 * result. Since this code runs early during bootup the system
5265 * is 'undisturbed' and the average latency makes sense.
5267 * The algorithm in essence auto-detects the relevant cache-size,
5268 * so it will properly detect different cachesizes for different
5269 * cache-hierarchies, depending on how the CPUs are connected.
5271 * Architectures can prime the upper limit of the search range via
5272 * max_cache_size, otherwise the search range defaults to 20MB...64K.
5274 static unsigned long long
5275 measure_cost(int cpu1, int cpu2, void *cache, unsigned int size)
5277 unsigned long long cost1, cost2;
5281 * Measure the migration cost of 'size' bytes, over an
5282 * average of 10 runs:
5284 * (We perturb the cache size by a small (0..4k)
5285 * value to compensate size/alignment related artifacts.
5286 * We also subtract the cost of the operation done on
5292 * dry run, to make sure we start off cache-cold on cpu1,
5293 * and to get any vmalloc pagefaults in advance:
5295 measure_one(cache, size, cpu1, cpu2);
5296 for (i = 0; i < ITERATIONS; i++)
5297 cost1 += measure_one(cache, size - i*1024, cpu1, cpu2);
5299 measure_one(cache, size, cpu2, cpu1);
5300 for (i = 0; i < ITERATIONS; i++)
5301 cost1 += measure_one(cache, size - i*1024, cpu2, cpu1);
5304 * (We measure the non-migrating [cached] cost on both
5305 * cpu1 and cpu2, to handle CPUs with different speeds)
5309 measure_one(cache, size, cpu1, cpu1);
5310 for (i = 0; i < ITERATIONS; i++)
5311 cost2 += measure_one(cache, size - i*1024, cpu1, cpu1);
5313 measure_one(cache, size, cpu2, cpu2);
5314 for (i = 0; i < ITERATIONS; i++)
5315 cost2 += measure_one(cache, size - i*1024, cpu2, cpu2);
5318 * Get the per-iteration migration cost:
5320 do_div(cost1, 2*ITERATIONS);
5321 do_div(cost2, 2*ITERATIONS);
5323 return cost1 - cost2;
5326 static unsigned long long measure_migration_cost(int cpu1, int cpu2)
5328 unsigned long long max_cost = 0, fluct = 0, avg_fluct = 0;
5329 unsigned int max_size, size, size_found = 0;
5330 long long cost = 0, prev_cost;
5334 * Search from max_cache_size*5 down to 64K - the real relevant
5335 * cachesize has to lie somewhere inbetween.
5337 if (max_cache_size) {
5338 max_size = max(max_cache_size * SEARCH_SCOPE, MIN_CACHE_SIZE);
5339 size = max(max_cache_size / SEARCH_SCOPE, MIN_CACHE_SIZE);
5342 * Since we have no estimation about the relevant
5345 max_size = DEFAULT_CACHE_SIZE * SEARCH_SCOPE;
5346 size = MIN_CACHE_SIZE;
5349 if (!cpu_online(cpu1) || !cpu_online(cpu2)) {
5350 printk("cpu %d and %d not both online!\n", cpu1, cpu2);
5355 * Allocate the working set:
5357 cache = vmalloc(max_size);
5359 printk("could not vmalloc %d bytes for cache!\n", 2*max_size);
5360 return 1000000; // return 1 msec on very small boxen
5363 while (size <= max_size) {
5365 cost = measure_cost(cpu1, cpu2, cache, size);
5371 if (max_cost < cost) {
5377 * Calculate average fluctuation, we use this to prevent
5378 * noise from triggering an early break out of the loop:
5380 fluct = abs(cost - prev_cost);
5381 avg_fluct = (avg_fluct + fluct)/2;
5383 if (migration_debug)
5384 printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): (%8Ld %8Ld)\n",
5386 (long)cost / 1000000,
5387 ((long)cost / 100000) % 10,
5388 (long)max_cost / 1000000,
5389 ((long)max_cost / 100000) % 10,
5390 domain_distance(cpu1, cpu2),
5394 * If we iterated at least 20% past the previous maximum,
5395 * and the cost has dropped by more than 20% already,
5396 * (taking fluctuations into account) then we assume to
5397 * have found the maximum and break out of the loop early:
5399 if (size_found && (size*100 > size_found*SIZE_THRESH))
5400 if (cost+avg_fluct <= 0 ||
5401 max_cost*100 > (cost+avg_fluct)*COST_THRESH) {
5403 if (migration_debug)
5404 printk("-> found max.\n");
5408 * Increase the cachesize in 10% steps:
5410 size = size * 10 / 9;
5413 if (migration_debug)
5414 printk("[%d][%d] working set size found: %d, cost: %Ld\n",
5415 cpu1, cpu2, size_found, max_cost);
5420 * A task is considered 'cache cold' if at least 2 times
5421 * the worst-case cost of migration has passed.
5423 * (this limit is only listened to if the load-balancing
5424 * situation is 'nice' - if there is a large imbalance we
5425 * ignore it for the sake of CPU utilization and
5426 * processing fairness.)
5428 return 2 * max_cost * migration_factor / MIGRATION_FACTOR_SCALE;
5431 static void calibrate_migration_costs(const cpumask_t *cpu_map)
5433 int cpu1 = -1, cpu2 = -1, cpu, orig_cpu = raw_smp_processor_id();
5434 unsigned long j0, j1, distance, max_distance = 0;
5435 struct sched_domain *sd;
5440 * First pass - calculate the cacheflush times:
5442 for_each_cpu_mask(cpu1, *cpu_map) {
5443 for_each_cpu_mask(cpu2, *cpu_map) {
5446 distance = domain_distance(cpu1, cpu2);
5447 max_distance = max(max_distance, distance);
5449 * No result cached yet?
5451 if (migration_cost[distance] == -1LL)
5452 migration_cost[distance] =
5453 measure_migration_cost(cpu1, cpu2);
5457 * Second pass - update the sched domain hierarchy with
5458 * the new cache-hot-time estimations:
5460 for_each_cpu_mask(cpu, *cpu_map) {
5462 for_each_domain(cpu, sd) {
5463 sd->cache_hot_time = migration_cost[distance];
5470 if (migration_debug)
5471 printk("migration: max_cache_size: %d, cpu: %d MHz:\n",
5479 if (system_state == SYSTEM_BOOTING) {
5480 printk("migration_cost=");
5481 for (distance = 0; distance <= max_distance; distance++) {
5484 printk("%ld", (long)migration_cost[distance] / 1000);
5489 if (migration_debug)
5490 printk("migration: %ld seconds\n", (j1-j0)/HZ);
5493 * Move back to the original CPU. NUMA-Q gets confused
5494 * if we migrate to another quad during bootup.
5496 if (raw_smp_processor_id() != orig_cpu) {
5497 cpumask_t mask = cpumask_of_cpu(orig_cpu),
5498 saved_mask = current->cpus_allowed;
5500 set_cpus_allowed(current, mask);
5501 set_cpus_allowed(current, saved_mask);
5508 * find_next_best_node - find the next node to include in a sched_domain
5509 * @node: node whose sched_domain we're building
5510 * @used_nodes: nodes already in the sched_domain
5512 * Find the next node to include in a given scheduling domain. Simply
5513 * finds the closest node not already in the @used_nodes map.
5515 * Should use nodemask_t.
5517 static int find_next_best_node(int node, unsigned long *used_nodes)
5519 int i, n, val, min_val, best_node = 0;
5523 for (i = 0; i < MAX_NUMNODES; i++) {
5524 /* Start at @node */
5525 n = (node + i) % MAX_NUMNODES;
5527 if (!nr_cpus_node(n))
5530 /* Skip already used nodes */
5531 if (test_bit(n, used_nodes))
5534 /* Simple min distance search */
5535 val = node_distance(node, n);
5537 if (val < min_val) {
5543 set_bit(best_node, used_nodes);
5548 * sched_domain_node_span - get a cpumask for a node's sched_domain
5549 * @node: node whose cpumask we're constructing
5550 * @size: number of nodes to include in this span
5552 * Given a node, construct a good cpumask for its sched_domain to span. It
5553 * should be one that prevents unnecessary balancing, but also spreads tasks
5556 static cpumask_t sched_domain_node_span(int node)
5559 cpumask_t span, nodemask;
5560 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5563 bitmap_zero(used_nodes, MAX_NUMNODES);
5565 nodemask = node_to_cpumask(node);
5566 cpus_or(span, span, nodemask);
5567 set_bit(node, used_nodes);
5569 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5570 int next_node = find_next_best_node(node, used_nodes);
5571 nodemask = node_to_cpumask(next_node);
5572 cpus_or(span, span, nodemask);
5580 * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we
5581 * can switch it on easily if needed.
5583 #ifdef CONFIG_SCHED_SMT
5584 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5585 static struct sched_group sched_group_cpus[NR_CPUS];
5586 static int cpu_to_cpu_group(int cpu)
5592 #ifdef CONFIG_SCHED_MC
5593 static DEFINE_PER_CPU(struct sched_domain, core_domains);
5594 static struct sched_group sched_group_core[NR_CPUS];
5597 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5598 static int cpu_to_core_group(int cpu)
5600 return first_cpu(cpu_sibling_map[cpu]);
5602 #elif defined(CONFIG_SCHED_MC)
5603 static int cpu_to_core_group(int cpu)
5609 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5610 static struct sched_group sched_group_phys[NR_CPUS];
5611 static int cpu_to_phys_group(int cpu)
5613 #if defined(CONFIG_SCHED_MC)
5614 cpumask_t mask = cpu_coregroup_map(cpu);
5615 return first_cpu(mask);
5616 #elif defined(CONFIG_SCHED_SMT)
5617 return first_cpu(cpu_sibling_map[cpu]);
5625 * The init_sched_build_groups can't handle what we want to do with node
5626 * groups, so roll our own. Now each node has its own list of groups which
5627 * gets dynamically allocated.
5629 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5630 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5632 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5633 static struct sched_group *sched_group_allnodes_bycpu[NR_CPUS];
5635 static int cpu_to_allnodes_group(int cpu)
5637 return cpu_to_node(cpu);
5639 static void init_numa_sched_groups_power(struct sched_group *group_head)
5641 struct sched_group *sg = group_head;
5647 for_each_cpu_mask(j, sg->cpumask) {
5648 struct sched_domain *sd;
5650 sd = &per_cpu(phys_domains, j);
5651 if (j != first_cpu(sd->groups->cpumask)) {
5653 * Only add "power" once for each
5659 sg->cpu_power += sd->groups->cpu_power;
5662 if (sg != group_head)
5668 * Build sched domains for a given set of cpus and attach the sched domains
5669 * to the individual cpus
5671 void build_sched_domains(const cpumask_t *cpu_map)
5675 struct sched_group **sched_group_nodes = NULL;
5676 struct sched_group *sched_group_allnodes = NULL;
5679 * Allocate the per-node list of sched groups
5681 sched_group_nodes = kmalloc(sizeof(struct sched_group*)*MAX_NUMNODES,
5683 if (!sched_group_nodes) {
5684 printk(KERN_WARNING "Can not alloc sched group node list\n");
5687 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
5691 * Set up domains for cpus specified by the cpu_map.
5693 for_each_cpu_mask(i, *cpu_map) {
5695 struct sched_domain *sd = NULL, *p;
5696 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
5698 cpus_and(nodemask, nodemask, *cpu_map);
5701 if (cpus_weight(*cpu_map)
5702 > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
5703 if (!sched_group_allnodes) {
5704 sched_group_allnodes
5705 = kmalloc(sizeof(struct sched_group)
5708 if (!sched_group_allnodes) {
5710 "Can not alloc allnodes sched group\n");
5713 sched_group_allnodes_bycpu[i]
5714 = sched_group_allnodes;
5716 sd = &per_cpu(allnodes_domains, i);
5717 *sd = SD_ALLNODES_INIT;
5718 sd->span = *cpu_map;
5719 group = cpu_to_allnodes_group(i);
5720 sd->groups = &sched_group_allnodes[group];
5725 sd = &per_cpu(node_domains, i);
5727 sd->span = sched_domain_node_span(cpu_to_node(i));
5729 cpus_and(sd->span, sd->span, *cpu_map);
5733 sd = &per_cpu(phys_domains, i);
5734 group = cpu_to_phys_group(i);
5736 sd->span = nodemask;
5738 sd->groups = &sched_group_phys[group];
5740 #ifdef CONFIG_SCHED_MC
5742 sd = &per_cpu(core_domains, i);
5743 group = cpu_to_core_group(i);
5745 sd->span = cpu_coregroup_map(i);
5746 cpus_and(sd->span, sd->span, *cpu_map);
5748 sd->groups = &sched_group_core[group];
5751 #ifdef CONFIG_SCHED_SMT
5753 sd = &per_cpu(cpu_domains, i);
5754 group = cpu_to_cpu_group(i);
5755 *sd = SD_SIBLING_INIT;
5756 sd->span = cpu_sibling_map[i];
5757 cpus_and(sd->span, sd->span, *cpu_map);
5759 sd->groups = &sched_group_cpus[group];
5763 #ifdef CONFIG_SCHED_SMT
5764 /* Set up CPU (sibling) groups */
5765 for_each_cpu_mask(i, *cpu_map) {
5766 cpumask_t this_sibling_map = cpu_sibling_map[i];
5767 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
5768 if (i != first_cpu(this_sibling_map))
5771 init_sched_build_groups(sched_group_cpus, this_sibling_map,
5776 #ifdef CONFIG_SCHED_MC
5777 /* Set up multi-core groups */
5778 for_each_cpu_mask(i, *cpu_map) {
5779 cpumask_t this_core_map = cpu_coregroup_map(i);
5780 cpus_and(this_core_map, this_core_map, *cpu_map);
5781 if (i != first_cpu(this_core_map))
5783 init_sched_build_groups(sched_group_core, this_core_map,
5784 &cpu_to_core_group);
5789 /* Set up physical groups */
5790 for (i = 0; i < MAX_NUMNODES; i++) {
5791 cpumask_t nodemask = node_to_cpumask(i);
5793 cpus_and(nodemask, nodemask, *cpu_map);
5794 if (cpus_empty(nodemask))
5797 init_sched_build_groups(sched_group_phys, nodemask,
5798 &cpu_to_phys_group);
5802 /* Set up node groups */
5803 if (sched_group_allnodes)
5804 init_sched_build_groups(sched_group_allnodes, *cpu_map,
5805 &cpu_to_allnodes_group);
5807 for (i = 0; i < MAX_NUMNODES; i++) {
5808 /* Set up node groups */
5809 struct sched_group *sg, *prev;
5810 cpumask_t nodemask = node_to_cpumask(i);
5811 cpumask_t domainspan;
5812 cpumask_t covered = CPU_MASK_NONE;
5815 cpus_and(nodemask, nodemask, *cpu_map);
5816 if (cpus_empty(nodemask)) {
5817 sched_group_nodes[i] = NULL;
5821 domainspan = sched_domain_node_span(i);
5822 cpus_and(domainspan, domainspan, *cpu_map);
5824 sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL);
5825 sched_group_nodes[i] = sg;
5826 for_each_cpu_mask(j, nodemask) {
5827 struct sched_domain *sd;
5828 sd = &per_cpu(node_domains, j);
5830 if (sd->groups == NULL) {
5831 /* Turn off balancing if we have no groups */
5837 "Can not alloc domain group for node %d\n", i);
5841 sg->cpumask = nodemask;
5842 cpus_or(covered, covered, nodemask);
5845 for (j = 0; j < MAX_NUMNODES; j++) {
5846 cpumask_t tmp, notcovered;
5847 int n = (i + j) % MAX_NUMNODES;
5849 cpus_complement(notcovered, covered);
5850 cpus_and(tmp, notcovered, *cpu_map);
5851 cpus_and(tmp, tmp, domainspan);
5852 if (cpus_empty(tmp))
5855 nodemask = node_to_cpumask(n);
5856 cpus_and(tmp, tmp, nodemask);
5857 if (cpus_empty(tmp))
5860 sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL);
5863 "Can not alloc domain group for node %d\n", j);
5868 cpus_or(covered, covered, tmp);
5872 prev->next = sched_group_nodes[i];
5876 /* Calculate CPU power for physical packages and nodes */
5877 for_each_cpu_mask(i, *cpu_map) {
5879 struct sched_domain *sd;
5880 #ifdef CONFIG_SCHED_SMT
5881 sd = &per_cpu(cpu_domains, i);
5882 power = SCHED_LOAD_SCALE;
5883 sd->groups->cpu_power = power;
5885 #ifdef CONFIG_SCHED_MC
5886 sd = &per_cpu(core_domains, i);
5887 power = SCHED_LOAD_SCALE + (cpus_weight(sd->groups->cpumask)-1)
5888 * SCHED_LOAD_SCALE / 10;
5889 sd->groups->cpu_power = power;
5891 sd = &per_cpu(phys_domains, i);
5894 * This has to be < 2 * SCHED_LOAD_SCALE
5895 * Lets keep it SCHED_LOAD_SCALE, so that
5896 * while calculating NUMA group's cpu_power
5898 * numa_group->cpu_power += phys_group->cpu_power;
5900 * See "only add power once for each physical pkg"
5903 sd->groups->cpu_power = SCHED_LOAD_SCALE;
5905 sd = &per_cpu(phys_domains, i);
5906 power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
5907 (cpus_weight(sd->groups->cpumask)-1) / 10;
5908 sd->groups->cpu_power = power;
5913 for (i = 0; i < MAX_NUMNODES; i++)
5914 init_numa_sched_groups_power(sched_group_nodes[i]);
5916 init_numa_sched_groups_power(sched_group_allnodes);
5919 /* Attach the domains */
5920 for_each_cpu_mask(i, *cpu_map) {
5921 struct sched_domain *sd;
5922 #ifdef CONFIG_SCHED_SMT
5923 sd = &per_cpu(cpu_domains, i);
5924 #elif defined(CONFIG_SCHED_MC)
5925 sd = &per_cpu(core_domains, i);
5927 sd = &per_cpu(phys_domains, i);
5929 cpu_attach_domain(sd, i);
5932 * Tune cache-hot values:
5934 calibrate_migration_costs(cpu_map);
5937 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
5939 static void arch_init_sched_domains(const cpumask_t *cpu_map)
5941 cpumask_t cpu_default_map;
5944 * Setup mask for cpus without special case scheduling requirements.
5945 * For now this just excludes isolated cpus, but could be used to
5946 * exclude other special cases in the future.
5948 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
5950 build_sched_domains(&cpu_default_map);
5953 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
5959 for_each_cpu_mask(cpu, *cpu_map) {
5960 struct sched_group *sched_group_allnodes
5961 = sched_group_allnodes_bycpu[cpu];
5962 struct sched_group **sched_group_nodes
5963 = sched_group_nodes_bycpu[cpu];
5965 if (sched_group_allnodes) {
5966 kfree(sched_group_allnodes);
5967 sched_group_allnodes_bycpu[cpu] = NULL;
5970 if (!sched_group_nodes)
5973 for (i = 0; i < MAX_NUMNODES; i++) {
5974 cpumask_t nodemask = node_to_cpumask(i);
5975 struct sched_group *oldsg, *sg = sched_group_nodes[i];
5977 cpus_and(nodemask, nodemask, *cpu_map);
5978 if (cpus_empty(nodemask))
5988 if (oldsg != sched_group_nodes[i])
5991 kfree(sched_group_nodes);
5992 sched_group_nodes_bycpu[cpu] = NULL;
5998 * Detach sched domains from a group of cpus specified in cpu_map
5999 * These cpus will now be attached to the NULL domain
6001 static void detach_destroy_domains(const cpumask_t *cpu_map)
6005 for_each_cpu_mask(i, *cpu_map)
6006 cpu_attach_domain(NULL, i);
6007 synchronize_sched();
6008 arch_destroy_sched_domains(cpu_map);
6012 * Partition sched domains as specified by the cpumasks below.
6013 * This attaches all cpus from the cpumasks to the NULL domain,
6014 * waits for a RCU quiescent period, recalculates sched
6015 * domain information and then attaches them back to the
6016 * correct sched domains
6017 * Call with hotplug lock held
6019 void partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6021 cpumask_t change_map;
6023 cpus_and(*partition1, *partition1, cpu_online_map);
6024 cpus_and(*partition2, *partition2, cpu_online_map);
6025 cpus_or(change_map, *partition1, *partition2);
6027 /* Detach sched domains from all of the affected cpus */
6028 detach_destroy_domains(&change_map);
6029 if (!cpus_empty(*partition1))
6030 build_sched_domains(partition1);
6031 if (!cpus_empty(*partition2))
6032 build_sched_domains(partition2);
6035 #ifdef CONFIG_HOTPLUG_CPU
6037 * Force a reinitialization of the sched domains hierarchy. The domains
6038 * and groups cannot be updated in place without racing with the balancing
6039 * code, so we temporarily attach all running cpus to the NULL domain
6040 * which will prevent rebalancing while the sched domains are recalculated.
6042 static int update_sched_domains(struct notifier_block *nfb,
6043 unsigned long action, void *hcpu)
6046 case CPU_UP_PREPARE:
6047 case CPU_DOWN_PREPARE:
6048 detach_destroy_domains(&cpu_online_map);
6051 case CPU_UP_CANCELED:
6052 case CPU_DOWN_FAILED:
6056 * Fall through and re-initialise the domains.
6063 /* The hotplug lock is already held by cpu_up/cpu_down */
6064 arch_init_sched_domains(&cpu_online_map);
6070 void __init sched_init_smp(void)
6073 arch_init_sched_domains(&cpu_online_map);
6074 unlock_cpu_hotplug();
6075 /* XXX: Theoretical race here - CPU may be hotplugged now */
6076 hotcpu_notifier(update_sched_domains, 0);
6079 void __init sched_init_smp(void)
6082 #endif /* CONFIG_SMP */
6084 int in_sched_functions(unsigned long addr)
6086 /* Linker adds these: start and end of __sched functions */
6087 extern char __sched_text_start[], __sched_text_end[];
6088 return in_lock_functions(addr) ||
6089 (addr >= (unsigned long)__sched_text_start
6090 && addr < (unsigned long)__sched_text_end);
6093 void __init sched_init(void)
6098 for_each_possible_cpu(i) {
6099 prio_array_t *array;
6102 spin_lock_init(&rq->lock);
6104 rq->active = rq->arrays;
6105 rq->expired = rq->arrays + 1;
6106 rq->best_expired_prio = MAX_PRIO;
6110 for (j = 1; j < 3; j++)
6111 rq->cpu_load[j] = 0;
6112 rq->active_balance = 0;
6114 rq->migration_thread = NULL;
6115 INIT_LIST_HEAD(&rq->migration_queue);
6118 atomic_set(&rq->nr_iowait, 0);
6120 for (j = 0; j < 2; j++) {
6121 array = rq->arrays + j;
6122 for (k = 0; k < MAX_PRIO; k++) {
6123 INIT_LIST_HEAD(array->queue + k);
6124 __clear_bit(k, array->bitmap);
6126 // delimiter for bitsearch
6127 __set_bit(MAX_PRIO, array->bitmap);
6132 * The boot idle thread does lazy MMU switching as well:
6134 atomic_inc(&init_mm.mm_count);
6135 enter_lazy_tlb(&init_mm, current);
6138 * Make us the idle thread. Technically, schedule() should not be
6139 * called from this thread, however somewhere below it might be,
6140 * but because we are the idle thread, we just pick up running again
6141 * when this runqueue becomes "idle".
6143 init_idle(current, smp_processor_id());
6146 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6147 void __might_sleep(char *file, int line)
6149 #if defined(in_atomic)
6150 static unsigned long prev_jiffy; /* ratelimiting */
6152 if ((in_atomic() || irqs_disabled()) &&
6153 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6154 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6156 prev_jiffy = jiffies;
6157 printk(KERN_ERR "BUG: sleeping function called from invalid"
6158 " context at %s:%d\n", file, line);
6159 printk("in_atomic():%d, irqs_disabled():%d\n",
6160 in_atomic(), irqs_disabled());
6165 EXPORT_SYMBOL(__might_sleep);
6168 #ifdef CONFIG_MAGIC_SYSRQ
6169 void normalize_rt_tasks(void)
6171 struct task_struct *p;
6172 prio_array_t *array;
6173 unsigned long flags;
6176 read_lock_irq(&tasklist_lock);
6177 for_each_process (p) {
6181 rq = task_rq_lock(p, &flags);
6185 deactivate_task(p, task_rq(p));
6186 __setscheduler(p, SCHED_NORMAL, 0);
6188 __activate_task(p, task_rq(p));
6189 resched_task(rq->curr);
6192 task_rq_unlock(rq, &flags);
6194 read_unlock_irq(&tasklist_lock);
6197 #endif /* CONFIG_MAGIC_SYSRQ */
6201 * These functions are only useful for the IA64 MCA handling.
6203 * They can only be called when the whole system has been
6204 * stopped - every CPU needs to be quiescent, and no scheduling
6205 * activity can take place. Using them for anything else would
6206 * be a serious bug, and as a result, they aren't even visible
6207 * under any other configuration.
6211 * curr_task - return the current task for a given cpu.
6212 * @cpu: the processor in question.
6214 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6216 task_t *curr_task(int cpu)
6218 return cpu_curr(cpu);
6222 * set_curr_task - set the current task for a given cpu.
6223 * @cpu: the processor in question.
6224 * @p: the task pointer to set.
6226 * Description: This function must only be used when non-maskable interrupts
6227 * are serviced on a separate stack. It allows the architecture to switch the
6228 * notion of the current task on a cpu in a non-blocking manner. This function
6229 * must be called with all CPU's synchronized, and interrupts disabled, the
6230 * and caller must save the original value of the current task (see
6231 * curr_task() above) and restore that value before reenabling interrupts and
6232 * re-starting the system.
6234 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6236 void set_curr_task(int cpu, task_t *p)