4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
22 #include <linux/module.h>
23 #include <linux/nmi.h>
24 #include <linux/init.h>
25 #include <asm/uaccess.h>
26 #include <linux/highmem.h>
27 #include <linux/smp_lock.h>
28 #include <asm/mmu_context.h>
29 #include <linux/interrupt.h>
30 #include <linux/capability.h>
31 #include <linux/completion.h>
32 #include <linux/kernel_stat.h>
33 #include <linux/security.h>
34 #include <linux/notifier.h>
35 #include <linux/profile.h>
36 #include <linux/suspend.h>
37 #include <linux/vmalloc.h>
38 #include <linux/blkdev.h>
39 #include <linux/delay.h>
40 #include <linux/smp.h>
41 #include <linux/threads.h>
42 #include <linux/timer.h>
43 #include <linux/rcupdate.h>
44 #include <linux/cpu.h>
45 #include <linux/cpuset.h>
46 #include <linux/percpu.h>
47 #include <linux/kthread.h>
48 #include <linux/seq_file.h>
49 #include <linux/syscalls.h>
50 #include <linux/times.h>
51 #include <linux/acct.h>
52 #include <linux/kprobes.h>
55 #include <asm/unistd.h>
58 * Convert user-nice values [ -20 ... 0 ... 19 ]
59 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
62 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
63 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
64 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
67 * 'User priority' is the nice value converted to something we
68 * can work with better when scaling various scheduler parameters,
69 * it's a [ 0 ... 39 ] range.
71 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
72 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
73 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
76 * Some helpers for converting nanosecond timing to jiffy resolution
78 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
79 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
82 * These are the 'tuning knobs' of the scheduler:
84 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
85 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
86 * Timeslices get refilled after they expire.
88 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
89 #define DEF_TIMESLICE (100 * HZ / 1000)
90 #define ON_RUNQUEUE_WEIGHT 30
91 #define CHILD_PENALTY 95
92 #define PARENT_PENALTY 100
94 #define PRIO_BONUS_RATIO 25
95 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
96 #define INTERACTIVE_DELTA 2
97 #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
98 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
99 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
102 * If a task is 'interactive' then we reinsert it in the active
103 * array after it has expired its current timeslice. (it will not
104 * continue to run immediately, it will still roundrobin with
105 * other interactive tasks.)
107 * This part scales the interactivity limit depending on niceness.
109 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
110 * Here are a few examples of different nice levels:
112 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
113 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
114 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
115 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
116 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
118 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
119 * priority range a task can explore, a value of '1' means the
120 * task is rated interactive.)
122 * Ie. nice +19 tasks can never get 'interactive' enough to be
123 * reinserted into the active array. And only heavily CPU-hog nice -20
124 * tasks will be expired. Default nice 0 tasks are somewhere between,
125 * it takes some effort for them to get interactive, but it's not
129 #define CURRENT_BONUS(p) \
130 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
133 #define GRANULARITY (10 * HZ / 1000 ? : 1)
136 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
137 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
140 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
141 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
144 #define SCALE(v1,v1_max,v2_max) \
145 (v1) * (v2_max) / (v1_max)
148 (SCALE(TASK_NICE(p) + 20, 40, MAX_BONUS) - 20 * MAX_BONUS / 40 + \
151 #define TASK_INTERACTIVE(p) \
152 ((p)->prio <= (p)->static_prio - DELTA(p))
154 #define INTERACTIVE_SLEEP(p) \
155 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
156 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
158 #define TASK_PREEMPTS_CURR(p, rq) \
159 ((p)->prio < (rq)->curr->prio)
162 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
163 * to time slice values: [800ms ... 100ms ... 5ms]
165 * The higher a thread's priority, the bigger timeslices
166 * it gets during one round of execution. But even the lowest
167 * priority thread gets MIN_TIMESLICE worth of execution time.
170 #define SCALE_PRIO(x, prio) \
171 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO/2), MIN_TIMESLICE)
173 static unsigned int task_timeslice(task_t *p)
175 if (p->static_prio < NICE_TO_PRIO(0))
176 return SCALE_PRIO(DEF_TIMESLICE*4, p->static_prio);
178 return SCALE_PRIO(DEF_TIMESLICE, p->static_prio);
180 #define task_hot(p, now, sd) ((long long) ((now) - (p)->last_ran) \
181 < (long long) (sd)->cache_hot_time)
184 * These are the runqueue data structures:
187 #define BITMAP_SIZE ((((MAX_PRIO+1+7)/8)+sizeof(long)-1)/sizeof(long))
189 typedef struct runqueue runqueue_t;
192 unsigned int nr_active;
193 unsigned long bitmap[BITMAP_SIZE];
194 struct list_head queue[MAX_PRIO];
198 * This is the main, per-CPU runqueue data structure.
200 * Locking rule: those places that want to lock multiple runqueues
201 * (such as the load balancing or the thread migration code), lock
202 * acquire operations must be ordered by ascending &runqueue.
208 * nr_running and cpu_load should be in the same cacheline because
209 * remote CPUs use both these fields when doing load calculation.
211 unsigned long nr_running;
213 unsigned long cpu_load[3];
215 unsigned long long nr_switches;
218 * This is part of a global counter where only the total sum
219 * over all CPUs matters. A task can increase this counter on
220 * one CPU and if it got migrated afterwards it may decrease
221 * it on another CPU. Always updated under the runqueue lock:
223 unsigned long nr_uninterruptible;
225 unsigned long expired_timestamp;
226 unsigned long long timestamp_last_tick;
228 struct mm_struct *prev_mm;
229 prio_array_t *active, *expired, arrays[2];
230 int best_expired_prio;
234 struct sched_domain *sd;
236 /* For active balancing */
240 task_t *migration_thread;
241 struct list_head migration_queue;
245 #ifdef CONFIG_SCHEDSTATS
247 struct sched_info rq_sched_info;
249 /* sys_sched_yield() stats */
250 unsigned long yld_exp_empty;
251 unsigned long yld_act_empty;
252 unsigned long yld_both_empty;
253 unsigned long yld_cnt;
255 /* schedule() stats */
256 unsigned long sched_switch;
257 unsigned long sched_cnt;
258 unsigned long sched_goidle;
260 /* try_to_wake_up() stats */
261 unsigned long ttwu_cnt;
262 unsigned long ttwu_local;
266 static DEFINE_PER_CPU(struct runqueue, runqueues);
269 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
270 * See detach_destroy_domains: synchronize_sched for details.
272 * The domain tree of any CPU may only be accessed from within
273 * preempt-disabled sections.
275 #define for_each_domain(cpu, domain) \
276 for (domain = rcu_dereference(cpu_rq(cpu)->sd); domain; domain = domain->parent)
278 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
279 #define this_rq() (&__get_cpu_var(runqueues))
280 #define task_rq(p) cpu_rq(task_cpu(p))
281 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
283 #ifndef prepare_arch_switch
284 # define prepare_arch_switch(next) do { } while (0)
286 #ifndef finish_arch_switch
287 # define finish_arch_switch(prev) do { } while (0)
290 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
291 static inline int task_running(runqueue_t *rq, task_t *p)
293 return rq->curr == p;
296 static inline void prepare_lock_switch(runqueue_t *rq, task_t *next)
300 static inline void finish_lock_switch(runqueue_t *rq, task_t *prev)
302 #ifdef CONFIG_DEBUG_SPINLOCK
303 /* this is a valid case when another task releases the spinlock */
304 rq->lock.owner = current;
306 spin_unlock_irq(&rq->lock);
309 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
310 static inline int task_running(runqueue_t *rq, task_t *p)
315 return rq->curr == p;
319 static inline void prepare_lock_switch(runqueue_t *rq, task_t *next)
323 * We can optimise this out completely for !SMP, because the
324 * SMP rebalancing from interrupt is the only thing that cares
329 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
330 spin_unlock_irq(&rq->lock);
332 spin_unlock(&rq->lock);
336 static inline void finish_lock_switch(runqueue_t *rq, task_t *prev)
340 * After ->oncpu is cleared, the task can be moved to a different CPU.
341 * We must ensure this doesn't happen until the switch is completely
347 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
351 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
354 * task_rq_lock - lock the runqueue a given task resides on and disable
355 * interrupts. Note the ordering: we can safely lookup the task_rq without
356 * explicitly disabling preemption.
358 static inline runqueue_t *task_rq_lock(task_t *p, unsigned long *flags)
364 local_irq_save(*flags);
366 spin_lock(&rq->lock);
367 if (unlikely(rq != task_rq(p))) {
368 spin_unlock_irqrestore(&rq->lock, *flags);
369 goto repeat_lock_task;
374 static inline void task_rq_unlock(runqueue_t *rq, unsigned long *flags)
377 spin_unlock_irqrestore(&rq->lock, *flags);
380 #ifdef CONFIG_SCHEDSTATS
382 * bump this up when changing the output format or the meaning of an existing
383 * format, so that tools can adapt (or abort)
385 #define SCHEDSTAT_VERSION 12
387 static int show_schedstat(struct seq_file *seq, void *v)
391 seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION);
392 seq_printf(seq, "timestamp %lu\n", jiffies);
393 for_each_online_cpu(cpu) {
394 runqueue_t *rq = cpu_rq(cpu);
396 struct sched_domain *sd;
400 /* runqueue-specific stats */
402 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
403 cpu, rq->yld_both_empty,
404 rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt,
405 rq->sched_switch, rq->sched_cnt, rq->sched_goidle,
406 rq->ttwu_cnt, rq->ttwu_local,
407 rq->rq_sched_info.cpu_time,
408 rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt);
410 seq_printf(seq, "\n");
413 /* domain-specific stats */
415 for_each_domain(cpu, sd) {
416 enum idle_type itype;
417 char mask_str[NR_CPUS];
419 cpumask_scnprintf(mask_str, NR_CPUS, sd->span);
420 seq_printf(seq, "domain%d %s", dcnt++, mask_str);
421 for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES;
423 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu",
425 sd->lb_balanced[itype],
426 sd->lb_failed[itype],
427 sd->lb_imbalance[itype],
428 sd->lb_gained[itype],
429 sd->lb_hot_gained[itype],
430 sd->lb_nobusyq[itype],
431 sd->lb_nobusyg[itype]);
433 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
434 sd->alb_cnt, sd->alb_failed, sd->alb_pushed,
435 sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed,
436 sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed,
437 sd->ttwu_wake_remote, sd->ttwu_move_affine, sd->ttwu_move_balance);
445 static int schedstat_open(struct inode *inode, struct file *file)
447 unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32);
448 char *buf = kmalloc(size, GFP_KERNEL);
454 res = single_open(file, show_schedstat, NULL);
456 m = file->private_data;
464 struct file_operations proc_schedstat_operations = {
465 .open = schedstat_open,
468 .release = single_release,
471 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
472 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
473 #else /* !CONFIG_SCHEDSTATS */
474 # define schedstat_inc(rq, field) do { } while (0)
475 # define schedstat_add(rq, field, amt) do { } while (0)
479 * rq_lock - lock a given runqueue and disable interrupts.
481 static inline runqueue_t *this_rq_lock(void)
488 spin_lock(&rq->lock);
493 #ifdef CONFIG_SCHEDSTATS
495 * Called when a process is dequeued from the active array and given
496 * the cpu. We should note that with the exception of interactive
497 * tasks, the expired queue will become the active queue after the active
498 * queue is empty, without explicitly dequeuing and requeuing tasks in the
499 * expired queue. (Interactive tasks may be requeued directly to the
500 * active queue, thus delaying tasks in the expired queue from running;
501 * see scheduler_tick()).
503 * This function is only called from sched_info_arrive(), rather than
504 * dequeue_task(). Even though a task may be queued and dequeued multiple
505 * times as it is shuffled about, we're really interested in knowing how
506 * long it was from the *first* time it was queued to the time that it
509 static inline void sched_info_dequeued(task_t *t)
511 t->sched_info.last_queued = 0;
515 * Called when a task finally hits the cpu. We can now calculate how
516 * long it was waiting to run. We also note when it began so that we
517 * can keep stats on how long its timeslice is.
519 static void sched_info_arrive(task_t *t)
521 unsigned long now = jiffies, diff = 0;
522 struct runqueue *rq = task_rq(t);
524 if (t->sched_info.last_queued)
525 diff = now - t->sched_info.last_queued;
526 sched_info_dequeued(t);
527 t->sched_info.run_delay += diff;
528 t->sched_info.last_arrival = now;
529 t->sched_info.pcnt++;
534 rq->rq_sched_info.run_delay += diff;
535 rq->rq_sched_info.pcnt++;
539 * Called when a process is queued into either the active or expired
540 * array. The time is noted and later used to determine how long we
541 * had to wait for us to reach the cpu. Since the expired queue will
542 * become the active queue after active queue is empty, without dequeuing
543 * and requeuing any tasks, we are interested in queuing to either. It
544 * is unusual but not impossible for tasks to be dequeued and immediately
545 * requeued in the same or another array: this can happen in sched_yield(),
546 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
549 * This function is only called from enqueue_task(), but also only updates
550 * the timestamp if it is already not set. It's assumed that
551 * sched_info_dequeued() will clear that stamp when appropriate.
553 static inline void sched_info_queued(task_t *t)
555 if (!t->sched_info.last_queued)
556 t->sched_info.last_queued = jiffies;
560 * Called when a process ceases being the active-running process, either
561 * voluntarily or involuntarily. Now we can calculate how long we ran.
563 static inline void sched_info_depart(task_t *t)
565 struct runqueue *rq = task_rq(t);
566 unsigned long diff = jiffies - t->sched_info.last_arrival;
568 t->sched_info.cpu_time += diff;
571 rq->rq_sched_info.cpu_time += diff;
575 * Called when tasks are switched involuntarily due, typically, to expiring
576 * their time slice. (This may also be called when switching to or from
577 * the idle task.) We are only called when prev != next.
579 static inline void sched_info_switch(task_t *prev, task_t *next)
581 struct runqueue *rq = task_rq(prev);
584 * prev now departs the cpu. It's not interesting to record
585 * stats about how efficient we were at scheduling the idle
588 if (prev != rq->idle)
589 sched_info_depart(prev);
591 if (next != rq->idle)
592 sched_info_arrive(next);
595 #define sched_info_queued(t) do { } while (0)
596 #define sched_info_switch(t, next) do { } while (0)
597 #endif /* CONFIG_SCHEDSTATS */
600 * Adding/removing a task to/from a priority array:
602 static void dequeue_task(struct task_struct *p, prio_array_t *array)
605 list_del(&p->run_list);
606 if (list_empty(array->queue + p->prio))
607 __clear_bit(p->prio, array->bitmap);
610 static void enqueue_task(struct task_struct *p, prio_array_t *array)
612 sched_info_queued(p);
613 list_add_tail(&p->run_list, array->queue + p->prio);
614 __set_bit(p->prio, array->bitmap);
620 * Put task to the end of the run list without the overhead of dequeue
621 * followed by enqueue.
623 static void requeue_task(struct task_struct *p, prio_array_t *array)
625 list_move_tail(&p->run_list, array->queue + p->prio);
628 static inline void enqueue_task_head(struct task_struct *p, prio_array_t *array)
630 list_add(&p->run_list, array->queue + p->prio);
631 __set_bit(p->prio, array->bitmap);
637 * effective_prio - return the priority that is based on the static
638 * priority but is modified by bonuses/penalties.
640 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
641 * into the -5 ... 0 ... +5 bonus/penalty range.
643 * We use 25% of the full 0...39 priority range so that:
645 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
646 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
648 * Both properties are important to certain workloads.
650 static int effective_prio(task_t *p)
657 bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
659 prio = p->static_prio - bonus;
660 if (prio < MAX_RT_PRIO)
662 if (prio > MAX_PRIO-1)
668 * __activate_task - move a task to the runqueue.
670 static inline void __activate_task(task_t *p, runqueue_t *rq)
672 enqueue_task(p, rq->active);
677 * __activate_idle_task - move idle task to the _front_ of runqueue.
679 static inline void __activate_idle_task(task_t *p, runqueue_t *rq)
681 enqueue_task_head(p, rq->active);
685 static int recalc_task_prio(task_t *p, unsigned long long now)
687 /* Caller must always ensure 'now >= p->timestamp' */
688 unsigned long long __sleep_time = now - p->timestamp;
689 unsigned long sleep_time;
691 if (unlikely(p->policy == SCHED_BATCH))
694 if (__sleep_time > NS_MAX_SLEEP_AVG)
695 sleep_time = NS_MAX_SLEEP_AVG;
697 sleep_time = (unsigned long)__sleep_time;
700 if (likely(sleep_time > 0)) {
702 * User tasks that sleep a long time are categorised as
703 * idle. They will only have their sleep_avg increased to a
704 * level that makes them just interactive priority to stay
705 * active yet prevent them suddenly becoming cpu hogs and
706 * starving other processes.
708 if (p->mm && sleep_time > INTERACTIVE_SLEEP(p)) {
709 unsigned long ceiling;
711 ceiling = JIFFIES_TO_NS(MAX_SLEEP_AVG -
713 if (p->sleep_avg < ceiling)
714 p->sleep_avg = ceiling;
717 * Tasks waking from uninterruptible sleep are
718 * limited in their sleep_avg rise as they
719 * are likely to be waiting on I/O
721 if (p->sleep_type == SLEEP_NONINTERACTIVE && p->mm) {
722 if (p->sleep_avg >= INTERACTIVE_SLEEP(p))
724 else if (p->sleep_avg + sleep_time >=
725 INTERACTIVE_SLEEP(p)) {
726 p->sleep_avg = INTERACTIVE_SLEEP(p);
732 * This code gives a bonus to interactive tasks.
734 * The boost works by updating the 'average sleep time'
735 * value here, based on ->timestamp. The more time a
736 * task spends sleeping, the higher the average gets -
737 * and the higher the priority boost gets as well.
739 p->sleep_avg += sleep_time;
741 if (p->sleep_avg > NS_MAX_SLEEP_AVG)
742 p->sleep_avg = NS_MAX_SLEEP_AVG;
746 return effective_prio(p);
750 * activate_task - move a task to the runqueue and do priority recalculation
752 * Update all the scheduling statistics stuff. (sleep average
753 * calculation, priority modifiers, etc.)
755 static void activate_task(task_t *p, runqueue_t *rq, int local)
757 unsigned long long now;
762 /* Compensate for drifting sched_clock */
763 runqueue_t *this_rq = this_rq();
764 now = (now - this_rq->timestamp_last_tick)
765 + rq->timestamp_last_tick;
770 p->prio = recalc_task_prio(p, now);
773 * This checks to make sure it's not an uninterruptible task
774 * that is now waking up.
776 if (p->sleep_type == SLEEP_NORMAL) {
778 * Tasks which were woken up by interrupts (ie. hw events)
779 * are most likely of interactive nature. So we give them
780 * the credit of extending their sleep time to the period
781 * of time they spend on the runqueue, waiting for execution
782 * on a CPU, first time around:
785 p->sleep_type = SLEEP_INTERRUPTED;
788 * Normal first-time wakeups get a credit too for
789 * on-runqueue time, but it will be weighted down:
791 p->sleep_type = SLEEP_INTERACTIVE;
796 __activate_task(p, rq);
800 * deactivate_task - remove a task from the runqueue.
802 static void deactivate_task(struct task_struct *p, runqueue_t *rq)
805 dequeue_task(p, p->array);
810 * resched_task - mark a task 'to be rescheduled now'.
812 * On UP this means the setting of the need_resched flag, on SMP it
813 * might also involve a cross-CPU call to trigger the scheduler on
817 static void resched_task(task_t *p)
821 assert_spin_locked(&task_rq(p)->lock);
823 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
826 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
829 if (cpu == smp_processor_id())
832 /* NEED_RESCHED must be visible before we test POLLING_NRFLAG */
834 if (!test_tsk_thread_flag(p, TIF_POLLING_NRFLAG))
835 smp_send_reschedule(cpu);
838 static inline void resched_task(task_t *p)
840 assert_spin_locked(&task_rq(p)->lock);
841 set_tsk_need_resched(p);
846 * task_curr - is this task currently executing on a CPU?
847 * @p: the task in question.
849 inline int task_curr(const task_t *p)
851 return cpu_curr(task_cpu(p)) == p;
856 struct list_head list;
861 struct completion done;
865 * The task's runqueue lock must be held.
866 * Returns true if you have to wait for migration thread.
868 static int migrate_task(task_t *p, int dest_cpu, migration_req_t *req)
870 runqueue_t *rq = task_rq(p);
873 * If the task is not on a runqueue (and not running), then
874 * it is sufficient to simply update the task's cpu field.
876 if (!p->array && !task_running(rq, p)) {
877 set_task_cpu(p, dest_cpu);
881 init_completion(&req->done);
883 req->dest_cpu = dest_cpu;
884 list_add(&req->list, &rq->migration_queue);
889 * wait_task_inactive - wait for a thread to unschedule.
891 * The caller must ensure that the task *will* unschedule sometime soon,
892 * else this function might spin for a *long* time. This function can't
893 * be called with interrupts off, or it may introduce deadlock with
894 * smp_call_function() if an IPI is sent by the same process we are
895 * waiting to become inactive.
897 void wait_task_inactive(task_t *p)
904 rq = task_rq_lock(p, &flags);
905 /* Must be off runqueue entirely, not preempted. */
906 if (unlikely(p->array || task_running(rq, p))) {
907 /* If it's preempted, we yield. It could be a while. */
908 preempted = !task_running(rq, p);
909 task_rq_unlock(rq, &flags);
915 task_rq_unlock(rq, &flags);
919 * kick_process - kick a running thread to enter/exit the kernel
920 * @p: the to-be-kicked thread
922 * Cause a process which is running on another CPU to enter
923 * kernel-mode, without any delay. (to get signals handled.)
925 * NOTE: this function doesnt have to take the runqueue lock,
926 * because all it wants to ensure is that the remote task enters
927 * the kernel. If the IPI races and the task has been migrated
928 * to another CPU then no harm is done and the purpose has been
931 void kick_process(task_t *p)
937 if ((cpu != smp_processor_id()) && task_curr(p))
938 smp_send_reschedule(cpu);
943 * Return a low guess at the load of a migration-source cpu.
945 * We want to under-estimate the load of migration sources, to
946 * balance conservatively.
948 static inline unsigned long source_load(int cpu, int type)
950 runqueue_t *rq = cpu_rq(cpu);
951 unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE;
955 return min(rq->cpu_load[type-1], load_now);
959 * Return a high guess at the load of a migration-target cpu
961 static inline unsigned long target_load(int cpu, int type)
963 runqueue_t *rq = cpu_rq(cpu);
964 unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE;
968 return max(rq->cpu_load[type-1], load_now);
972 * find_idlest_group finds and returns the least busy CPU group within the
975 static struct sched_group *
976 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
978 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
979 unsigned long min_load = ULONG_MAX, this_load = 0;
980 int load_idx = sd->forkexec_idx;
981 int imbalance = 100 + (sd->imbalance_pct-100)/2;
984 unsigned long load, avg_load;
988 /* Skip over this group if it has no CPUs allowed */
989 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
992 local_group = cpu_isset(this_cpu, group->cpumask);
994 /* Tally up the load of all CPUs in the group */
997 for_each_cpu_mask(i, group->cpumask) {
998 /* Bias balancing toward cpus of our domain */
1000 load = source_load(i, load_idx);
1002 load = target_load(i, load_idx);
1007 /* Adjust by relative CPU power of the group */
1008 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1011 this_load = avg_load;
1013 } else if (avg_load < min_load) {
1014 min_load = avg_load;
1018 group = group->next;
1019 } while (group != sd->groups);
1021 if (!idlest || 100*this_load < imbalance*min_load)
1027 * find_idlest_queue - find the idlest runqueue among the cpus in group.
1030 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1033 unsigned long load, min_load = ULONG_MAX;
1037 /* Traverse only the allowed CPUs */
1038 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1040 for_each_cpu_mask(i, tmp) {
1041 load = source_load(i, 0);
1043 if (load < min_load || (load == min_load && i == this_cpu)) {
1053 * sched_balance_self: balance the current task (running on cpu) in domains
1054 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1057 * Balance, ie. select the least loaded group.
1059 * Returns the target CPU number, or the same CPU if no balancing is needed.
1061 * preempt must be disabled.
1063 static int sched_balance_self(int cpu, int flag)
1065 struct task_struct *t = current;
1066 struct sched_domain *tmp, *sd = NULL;
1068 for_each_domain(cpu, tmp)
1069 if (tmp->flags & flag)
1074 struct sched_group *group;
1079 group = find_idlest_group(sd, t, cpu);
1083 new_cpu = find_idlest_cpu(group, t, cpu);
1084 if (new_cpu == -1 || new_cpu == cpu)
1087 /* Now try balancing at a lower domain level */
1091 weight = cpus_weight(span);
1092 for_each_domain(cpu, tmp) {
1093 if (weight <= cpus_weight(tmp->span))
1095 if (tmp->flags & flag)
1098 /* while loop will break here if sd == NULL */
1104 #endif /* CONFIG_SMP */
1107 * wake_idle() will wake a task on an idle cpu if task->cpu is
1108 * not idle and an idle cpu is available. The span of cpus to
1109 * search starts with cpus closest then further out as needed,
1110 * so we always favor a closer, idle cpu.
1112 * Returns the CPU we should wake onto.
1114 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1115 static int wake_idle(int cpu, task_t *p)
1118 struct sched_domain *sd;
1124 for_each_domain(cpu, sd) {
1125 if (sd->flags & SD_WAKE_IDLE) {
1126 cpus_and(tmp, sd->span, p->cpus_allowed);
1127 for_each_cpu_mask(i, tmp) {
1138 static inline int wake_idle(int cpu, task_t *p)
1145 * try_to_wake_up - wake up a thread
1146 * @p: the to-be-woken-up thread
1147 * @state: the mask of task states that can be woken
1148 * @sync: do a synchronous wakeup?
1150 * Put it on the run-queue if it's not already there. The "current"
1151 * thread is always on the run-queue (except when the actual
1152 * re-schedule is in progress), and as such you're allowed to do
1153 * the simpler "current->state = TASK_RUNNING" to mark yourself
1154 * runnable without the overhead of this.
1156 * returns failure only if the task is already active.
1158 static int try_to_wake_up(task_t *p, unsigned int state, int sync)
1160 int cpu, this_cpu, success = 0;
1161 unsigned long flags;
1165 unsigned long load, this_load;
1166 struct sched_domain *sd, *this_sd = NULL;
1170 rq = task_rq_lock(p, &flags);
1171 old_state = p->state;
1172 if (!(old_state & state))
1179 this_cpu = smp_processor_id();
1182 if (unlikely(task_running(rq, p)))
1187 schedstat_inc(rq, ttwu_cnt);
1188 if (cpu == this_cpu) {
1189 schedstat_inc(rq, ttwu_local);
1193 for_each_domain(this_cpu, sd) {
1194 if (cpu_isset(cpu, sd->span)) {
1195 schedstat_inc(sd, ttwu_wake_remote);
1201 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1205 * Check for affine wakeup and passive balancing possibilities.
1208 int idx = this_sd->wake_idx;
1209 unsigned int imbalance;
1211 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1213 load = source_load(cpu, idx);
1214 this_load = target_load(this_cpu, idx);
1216 new_cpu = this_cpu; /* Wake to this CPU if we can */
1218 if (this_sd->flags & SD_WAKE_AFFINE) {
1219 unsigned long tl = this_load;
1221 * If sync wakeup then subtract the (maximum possible)
1222 * effect of the currently running task from the load
1223 * of the current CPU:
1226 tl -= SCHED_LOAD_SCALE;
1229 tl + target_load(cpu, idx) <= SCHED_LOAD_SCALE) ||
1230 100*(tl + SCHED_LOAD_SCALE) <= imbalance*load) {
1232 * This domain has SD_WAKE_AFFINE and
1233 * p is cache cold in this domain, and
1234 * there is no bad imbalance.
1236 schedstat_inc(this_sd, ttwu_move_affine);
1242 * Start passive balancing when half the imbalance_pct
1245 if (this_sd->flags & SD_WAKE_BALANCE) {
1246 if (imbalance*this_load <= 100*load) {
1247 schedstat_inc(this_sd, ttwu_move_balance);
1253 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1255 new_cpu = wake_idle(new_cpu, p);
1256 if (new_cpu != cpu) {
1257 set_task_cpu(p, new_cpu);
1258 task_rq_unlock(rq, &flags);
1259 /* might preempt at this point */
1260 rq = task_rq_lock(p, &flags);
1261 old_state = p->state;
1262 if (!(old_state & state))
1267 this_cpu = smp_processor_id();
1272 #endif /* CONFIG_SMP */
1273 if (old_state == TASK_UNINTERRUPTIBLE) {
1274 rq->nr_uninterruptible--;
1276 * Tasks on involuntary sleep don't earn
1277 * sleep_avg beyond just interactive state.
1279 p->sleep_type = SLEEP_NONINTERACTIVE;
1283 * Tasks that have marked their sleep as noninteractive get
1284 * woken up with their sleep average not weighted in an
1287 if (old_state & TASK_NONINTERACTIVE)
1288 p->sleep_type = SLEEP_NONINTERACTIVE;
1291 activate_task(p, rq, cpu == this_cpu);
1293 * Sync wakeups (i.e. those types of wakeups where the waker
1294 * has indicated that it will leave the CPU in short order)
1295 * don't trigger a preemption, if the woken up task will run on
1296 * this cpu. (in this case the 'I will reschedule' promise of
1297 * the waker guarantees that the freshly woken up task is going
1298 * to be considered on this CPU.)
1300 if (!sync || cpu != this_cpu) {
1301 if (TASK_PREEMPTS_CURR(p, rq))
1302 resched_task(rq->curr);
1307 p->state = TASK_RUNNING;
1309 task_rq_unlock(rq, &flags);
1314 int fastcall wake_up_process(task_t *p)
1316 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1317 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1320 EXPORT_SYMBOL(wake_up_process);
1322 int fastcall wake_up_state(task_t *p, unsigned int state)
1324 return try_to_wake_up(p, state, 0);
1328 * Perform scheduler related setup for a newly forked process p.
1329 * p is forked by current.
1331 void fastcall sched_fork(task_t *p, int clone_flags)
1333 int cpu = get_cpu();
1336 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1338 set_task_cpu(p, cpu);
1341 * We mark the process as running here, but have not actually
1342 * inserted it onto the runqueue yet. This guarantees that
1343 * nobody will actually run it, and a signal or other external
1344 * event cannot wake it up and insert it on the runqueue either.
1346 p->state = TASK_RUNNING;
1347 INIT_LIST_HEAD(&p->run_list);
1349 #ifdef CONFIG_SCHEDSTATS
1350 memset(&p->sched_info, 0, sizeof(p->sched_info));
1352 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1355 #ifdef CONFIG_PREEMPT
1356 /* Want to start with kernel preemption disabled. */
1357 task_thread_info(p)->preempt_count = 1;
1360 * Share the timeslice between parent and child, thus the
1361 * total amount of pending timeslices in the system doesn't change,
1362 * resulting in more scheduling fairness.
1364 local_irq_disable();
1365 p->time_slice = (current->time_slice + 1) >> 1;
1367 * The remainder of the first timeslice might be recovered by
1368 * the parent if the child exits early enough.
1370 p->first_time_slice = 1;
1371 current->time_slice >>= 1;
1372 p->timestamp = sched_clock();
1373 if (unlikely(!current->time_slice)) {
1375 * This case is rare, it happens when the parent has only
1376 * a single jiffy left from its timeslice. Taking the
1377 * runqueue lock is not a problem.
1379 current->time_slice = 1;
1387 * wake_up_new_task - wake up a newly created task for the first time.
1389 * This function will do some initial scheduler statistics housekeeping
1390 * that must be done for every newly created context, then puts the task
1391 * on the runqueue and wakes it.
1393 void fastcall wake_up_new_task(task_t *p, unsigned long clone_flags)
1395 unsigned long flags;
1397 runqueue_t *rq, *this_rq;
1399 rq = task_rq_lock(p, &flags);
1400 BUG_ON(p->state != TASK_RUNNING);
1401 this_cpu = smp_processor_id();
1405 * We decrease the sleep average of forking parents
1406 * and children as well, to keep max-interactive tasks
1407 * from forking tasks that are max-interactive. The parent
1408 * (current) is done further down, under its lock.
1410 p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1411 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1413 p->prio = effective_prio(p);
1415 if (likely(cpu == this_cpu)) {
1416 if (!(clone_flags & CLONE_VM)) {
1418 * The VM isn't cloned, so we're in a good position to
1419 * do child-runs-first in anticipation of an exec. This
1420 * usually avoids a lot of COW overhead.
1422 if (unlikely(!current->array))
1423 __activate_task(p, rq);
1425 p->prio = current->prio;
1426 list_add_tail(&p->run_list, ¤t->run_list);
1427 p->array = current->array;
1428 p->array->nr_active++;
1433 /* Run child last */
1434 __activate_task(p, rq);
1436 * We skip the following code due to cpu == this_cpu
1438 * task_rq_unlock(rq, &flags);
1439 * this_rq = task_rq_lock(current, &flags);
1443 this_rq = cpu_rq(this_cpu);
1446 * Not the local CPU - must adjust timestamp. This should
1447 * get optimised away in the !CONFIG_SMP case.
1449 p->timestamp = (p->timestamp - this_rq->timestamp_last_tick)
1450 + rq->timestamp_last_tick;
1451 __activate_task(p, rq);
1452 if (TASK_PREEMPTS_CURR(p, rq))
1453 resched_task(rq->curr);
1456 * Parent and child are on different CPUs, now get the
1457 * parent runqueue to update the parent's ->sleep_avg:
1459 task_rq_unlock(rq, &flags);
1460 this_rq = task_rq_lock(current, &flags);
1462 current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1463 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1464 task_rq_unlock(this_rq, &flags);
1468 * Potentially available exiting-child timeslices are
1469 * retrieved here - this way the parent does not get
1470 * penalized for creating too many threads.
1472 * (this cannot be used to 'generate' timeslices
1473 * artificially, because any timeslice recovered here
1474 * was given away by the parent in the first place.)
1476 void fastcall sched_exit(task_t *p)
1478 unsigned long flags;
1482 * If the child was a (relative-) CPU hog then decrease
1483 * the sleep_avg of the parent as well.
1485 rq = task_rq_lock(p->parent, &flags);
1486 if (p->first_time_slice && task_cpu(p) == task_cpu(p->parent)) {
1487 p->parent->time_slice += p->time_slice;
1488 if (unlikely(p->parent->time_slice > task_timeslice(p)))
1489 p->parent->time_slice = task_timeslice(p);
1491 if (p->sleep_avg < p->parent->sleep_avg)
1492 p->parent->sleep_avg = p->parent->sleep_avg /
1493 (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg /
1495 task_rq_unlock(rq, &flags);
1499 * prepare_task_switch - prepare to switch tasks
1500 * @rq: the runqueue preparing to switch
1501 * @next: the task we are going to switch to.
1503 * This is called with the rq lock held and interrupts off. It must
1504 * be paired with a subsequent finish_task_switch after the context
1507 * prepare_task_switch sets up locking and calls architecture specific
1510 static inline void prepare_task_switch(runqueue_t *rq, task_t *next)
1512 prepare_lock_switch(rq, next);
1513 prepare_arch_switch(next);
1517 * finish_task_switch - clean up after a task-switch
1518 * @rq: runqueue associated with task-switch
1519 * @prev: the thread we just switched away from.
1521 * finish_task_switch must be called after the context switch, paired
1522 * with a prepare_task_switch call before the context switch.
1523 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1524 * and do any other architecture-specific cleanup actions.
1526 * Note that we may have delayed dropping an mm in context_switch(). If
1527 * so, we finish that here outside of the runqueue lock. (Doing it
1528 * with the lock held can cause deadlocks; see schedule() for
1531 static inline void finish_task_switch(runqueue_t *rq, task_t *prev)
1532 __releases(rq->lock)
1534 struct mm_struct *mm = rq->prev_mm;
1535 unsigned long prev_task_flags;
1540 * A task struct has one reference for the use as "current".
1541 * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1542 * calls schedule one last time. The schedule call will never return,
1543 * and the scheduled task must drop that reference.
1544 * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1545 * still held, otherwise prev could be scheduled on another cpu, die
1546 * there before we look at prev->state, and then the reference would
1548 * Manfred Spraul <manfred@colorfullife.com>
1550 prev_task_flags = prev->flags;
1551 finish_arch_switch(prev);
1552 finish_lock_switch(rq, prev);
1555 if (unlikely(prev_task_flags & PF_DEAD)) {
1557 * Remove function-return probe instances associated with this
1558 * task and put them back on the free list.
1560 kprobe_flush_task(prev);
1561 put_task_struct(prev);
1566 * schedule_tail - first thing a freshly forked thread must call.
1567 * @prev: the thread we just switched away from.
1569 asmlinkage void schedule_tail(task_t *prev)
1570 __releases(rq->lock)
1572 runqueue_t *rq = this_rq();
1573 finish_task_switch(rq, prev);
1574 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1575 /* In this case, finish_task_switch does not reenable preemption */
1578 if (current->set_child_tid)
1579 put_user(current->pid, current->set_child_tid);
1583 * context_switch - switch to the new MM and the new
1584 * thread's register state.
1587 task_t * context_switch(runqueue_t *rq, task_t *prev, task_t *next)
1589 struct mm_struct *mm = next->mm;
1590 struct mm_struct *oldmm = prev->active_mm;
1592 if (unlikely(!mm)) {
1593 next->active_mm = oldmm;
1594 atomic_inc(&oldmm->mm_count);
1595 enter_lazy_tlb(oldmm, next);
1597 switch_mm(oldmm, mm, next);
1599 if (unlikely(!prev->mm)) {
1600 prev->active_mm = NULL;
1601 WARN_ON(rq->prev_mm);
1602 rq->prev_mm = oldmm;
1605 /* Here we just switch the register state and the stack. */
1606 switch_to(prev, next, prev);
1612 * nr_running, nr_uninterruptible and nr_context_switches:
1614 * externally visible scheduler statistics: current number of runnable
1615 * threads, current number of uninterruptible-sleeping threads, total
1616 * number of context switches performed since bootup.
1618 unsigned long nr_running(void)
1620 unsigned long i, sum = 0;
1622 for_each_online_cpu(i)
1623 sum += cpu_rq(i)->nr_running;
1628 unsigned long nr_uninterruptible(void)
1630 unsigned long i, sum = 0;
1632 for_each_possible_cpu(i)
1633 sum += cpu_rq(i)->nr_uninterruptible;
1636 * Since we read the counters lockless, it might be slightly
1637 * inaccurate. Do not allow it to go below zero though:
1639 if (unlikely((long)sum < 0))
1645 unsigned long long nr_context_switches(void)
1647 unsigned long long i, sum = 0;
1649 for_each_possible_cpu(i)
1650 sum += cpu_rq(i)->nr_switches;
1655 unsigned long nr_iowait(void)
1657 unsigned long i, sum = 0;
1659 for_each_possible_cpu(i)
1660 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1665 unsigned long nr_active(void)
1667 unsigned long i, running = 0, uninterruptible = 0;
1669 for_each_online_cpu(i) {
1670 running += cpu_rq(i)->nr_running;
1671 uninterruptible += cpu_rq(i)->nr_uninterruptible;
1674 if (unlikely((long)uninterruptible < 0))
1675 uninterruptible = 0;
1677 return running + uninterruptible;
1683 * double_rq_lock - safely lock two runqueues
1685 * We must take them in cpu order to match code in
1686 * dependent_sleeper and wake_dependent_sleeper.
1688 * Note this does not disable interrupts like task_rq_lock,
1689 * you need to do so manually before calling.
1691 static void double_rq_lock(runqueue_t *rq1, runqueue_t *rq2)
1692 __acquires(rq1->lock)
1693 __acquires(rq2->lock)
1696 spin_lock(&rq1->lock);
1697 __acquire(rq2->lock); /* Fake it out ;) */
1699 if (rq1->cpu < rq2->cpu) {
1700 spin_lock(&rq1->lock);
1701 spin_lock(&rq2->lock);
1703 spin_lock(&rq2->lock);
1704 spin_lock(&rq1->lock);
1710 * double_rq_unlock - safely unlock two runqueues
1712 * Note this does not restore interrupts like task_rq_unlock,
1713 * you need to do so manually after calling.
1715 static void double_rq_unlock(runqueue_t *rq1, runqueue_t *rq2)
1716 __releases(rq1->lock)
1717 __releases(rq2->lock)
1719 spin_unlock(&rq1->lock);
1721 spin_unlock(&rq2->lock);
1723 __release(rq2->lock);
1727 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1729 static void double_lock_balance(runqueue_t *this_rq, runqueue_t *busiest)
1730 __releases(this_rq->lock)
1731 __acquires(busiest->lock)
1732 __acquires(this_rq->lock)
1734 if (unlikely(!spin_trylock(&busiest->lock))) {
1735 if (busiest->cpu < this_rq->cpu) {
1736 spin_unlock(&this_rq->lock);
1737 spin_lock(&busiest->lock);
1738 spin_lock(&this_rq->lock);
1740 spin_lock(&busiest->lock);
1745 * If dest_cpu is allowed for this process, migrate the task to it.
1746 * This is accomplished by forcing the cpu_allowed mask to only
1747 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
1748 * the cpu_allowed mask is restored.
1750 static void sched_migrate_task(task_t *p, int dest_cpu)
1752 migration_req_t req;
1754 unsigned long flags;
1756 rq = task_rq_lock(p, &flags);
1757 if (!cpu_isset(dest_cpu, p->cpus_allowed)
1758 || unlikely(cpu_is_offline(dest_cpu)))
1761 /* force the process onto the specified CPU */
1762 if (migrate_task(p, dest_cpu, &req)) {
1763 /* Need to wait for migration thread (might exit: take ref). */
1764 struct task_struct *mt = rq->migration_thread;
1765 get_task_struct(mt);
1766 task_rq_unlock(rq, &flags);
1767 wake_up_process(mt);
1768 put_task_struct(mt);
1769 wait_for_completion(&req.done);
1773 task_rq_unlock(rq, &flags);
1777 * sched_exec - execve() is a valuable balancing opportunity, because at
1778 * this point the task has the smallest effective memory and cache footprint.
1780 void sched_exec(void)
1782 int new_cpu, this_cpu = get_cpu();
1783 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
1785 if (new_cpu != this_cpu)
1786 sched_migrate_task(current, new_cpu);
1790 * pull_task - move a task from a remote runqueue to the local runqueue.
1791 * Both runqueues must be locked.
1794 void pull_task(runqueue_t *src_rq, prio_array_t *src_array, task_t *p,
1795 runqueue_t *this_rq, prio_array_t *this_array, int this_cpu)
1797 dequeue_task(p, src_array);
1798 src_rq->nr_running--;
1799 set_task_cpu(p, this_cpu);
1800 this_rq->nr_running++;
1801 enqueue_task(p, this_array);
1802 p->timestamp = (p->timestamp - src_rq->timestamp_last_tick)
1803 + this_rq->timestamp_last_tick;
1805 * Note that idle threads have a prio of MAX_PRIO, for this test
1806 * to be always true for them.
1808 if (TASK_PREEMPTS_CURR(p, this_rq))
1809 resched_task(this_rq->curr);
1813 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1816 int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu,
1817 struct sched_domain *sd, enum idle_type idle,
1821 * We do not migrate tasks that are:
1822 * 1) running (obviously), or
1823 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1824 * 3) are cache-hot on their current CPU.
1826 if (!cpu_isset(this_cpu, p->cpus_allowed))
1830 if (task_running(rq, p))
1834 * Aggressive migration if:
1835 * 1) task is cache cold, or
1836 * 2) too many balance attempts have failed.
1839 if (sd->nr_balance_failed > sd->cache_nice_tries)
1842 if (task_hot(p, rq->timestamp_last_tick, sd))
1848 * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq,
1849 * as part of a balancing operation within "domain". Returns the number of
1852 * Called with both runqueues locked.
1854 static int move_tasks(runqueue_t *this_rq, int this_cpu, runqueue_t *busiest,
1855 unsigned long max_nr_move, struct sched_domain *sd,
1856 enum idle_type idle, int *all_pinned)
1858 prio_array_t *array, *dst_array;
1859 struct list_head *head, *curr;
1860 int idx, pulled = 0, pinned = 0;
1863 if (max_nr_move == 0)
1869 * We first consider expired tasks. Those will likely not be
1870 * executed in the near future, and they are most likely to
1871 * be cache-cold, thus switching CPUs has the least effect
1874 if (busiest->expired->nr_active) {
1875 array = busiest->expired;
1876 dst_array = this_rq->expired;
1878 array = busiest->active;
1879 dst_array = this_rq->active;
1883 /* Start searching at priority 0: */
1887 idx = sched_find_first_bit(array->bitmap);
1889 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
1890 if (idx >= MAX_PRIO) {
1891 if (array == busiest->expired && busiest->active->nr_active) {
1892 array = busiest->active;
1893 dst_array = this_rq->active;
1899 head = array->queue + idx;
1902 tmp = list_entry(curr, task_t, run_list);
1906 if (!can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
1913 #ifdef CONFIG_SCHEDSTATS
1914 if (task_hot(tmp, busiest->timestamp_last_tick, sd))
1915 schedstat_inc(sd, lb_hot_gained[idle]);
1918 pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
1921 /* We only want to steal up to the prescribed number of tasks. */
1922 if (pulled < max_nr_move) {
1930 * Right now, this is the only place pull_task() is called,
1931 * so we can safely collect pull_task() stats here rather than
1932 * inside pull_task().
1934 schedstat_add(sd, lb_gained[idle], pulled);
1937 *all_pinned = pinned;
1942 * find_busiest_group finds and returns the busiest CPU group within the
1943 * domain. It calculates and returns the number of tasks which should be
1944 * moved to restore balance via the imbalance parameter.
1946 static struct sched_group *
1947 find_busiest_group(struct sched_domain *sd, int this_cpu,
1948 unsigned long *imbalance, enum idle_type idle, int *sd_idle)
1950 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
1951 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
1952 unsigned long max_pull;
1955 max_load = this_load = total_load = total_pwr = 0;
1956 if (idle == NOT_IDLE)
1957 load_idx = sd->busy_idx;
1958 else if (idle == NEWLY_IDLE)
1959 load_idx = sd->newidle_idx;
1961 load_idx = sd->idle_idx;
1968 local_group = cpu_isset(this_cpu, group->cpumask);
1970 /* Tally up the load of all CPUs in the group */
1973 for_each_cpu_mask(i, group->cpumask) {
1974 if (*sd_idle && !idle_cpu(i))
1977 /* Bias balancing toward cpus of our domain */
1979 load = target_load(i, load_idx);
1981 load = source_load(i, load_idx);
1986 total_load += avg_load;
1987 total_pwr += group->cpu_power;
1989 /* Adjust by relative CPU power of the group */
1990 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1993 this_load = avg_load;
1995 } else if (avg_load > max_load) {
1996 max_load = avg_load;
1999 group = group->next;
2000 } while (group != sd->groups);
2002 if (!busiest || this_load >= max_load || max_load <= SCHED_LOAD_SCALE)
2005 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2007 if (this_load >= avg_load ||
2008 100*max_load <= sd->imbalance_pct*this_load)
2012 * We're trying to get all the cpus to the average_load, so we don't
2013 * want to push ourselves above the average load, nor do we wish to
2014 * reduce the max loaded cpu below the average load, as either of these
2015 * actions would just result in more rebalancing later, and ping-pong
2016 * tasks around. Thus we look for the minimum possible imbalance.
2017 * Negative imbalances (*we* are more loaded than anyone else) will
2018 * be counted as no imbalance for these purposes -- we can't fix that
2019 * by pulling tasks to us. Be careful of negative numbers as they'll
2020 * appear as very large values with unsigned longs.
2023 /* Don't want to pull so many tasks that a group would go idle */
2024 max_pull = min(max_load - avg_load, max_load - SCHED_LOAD_SCALE);
2026 /* How much load to actually move to equalise the imbalance */
2027 *imbalance = min(max_pull * busiest->cpu_power,
2028 (avg_load - this_load) * this->cpu_power)
2031 if (*imbalance < SCHED_LOAD_SCALE) {
2032 unsigned long pwr_now = 0, pwr_move = 0;
2035 if (max_load - this_load >= SCHED_LOAD_SCALE*2) {
2041 * OK, we don't have enough imbalance to justify moving tasks,
2042 * however we may be able to increase total CPU power used by
2046 pwr_now += busiest->cpu_power*min(SCHED_LOAD_SCALE, max_load);
2047 pwr_now += this->cpu_power*min(SCHED_LOAD_SCALE, this_load);
2048 pwr_now /= SCHED_LOAD_SCALE;
2050 /* Amount of load we'd subtract */
2051 tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/busiest->cpu_power;
2053 pwr_move += busiest->cpu_power*min(SCHED_LOAD_SCALE,
2056 /* Amount of load we'd add */
2057 if (max_load*busiest->cpu_power <
2058 SCHED_LOAD_SCALE*SCHED_LOAD_SCALE)
2059 tmp = max_load*busiest->cpu_power/this->cpu_power;
2061 tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/this->cpu_power;
2062 pwr_move += this->cpu_power*min(SCHED_LOAD_SCALE, this_load + tmp);
2063 pwr_move /= SCHED_LOAD_SCALE;
2065 /* Move if we gain throughput */
2066 if (pwr_move <= pwr_now)
2073 /* Get rid of the scaling factor, rounding down as we divide */
2074 *imbalance = *imbalance / SCHED_LOAD_SCALE;
2084 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2086 static runqueue_t *find_busiest_queue(struct sched_group *group,
2087 enum idle_type idle)
2089 unsigned long load, max_load = 0;
2090 runqueue_t *busiest = NULL;
2093 for_each_cpu_mask(i, group->cpumask) {
2094 load = source_load(i, 0);
2096 if (load > max_load) {
2098 busiest = cpu_rq(i);
2106 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2107 * so long as it is large enough.
2109 #define MAX_PINNED_INTERVAL 512
2112 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2113 * tasks if there is an imbalance.
2115 * Called with this_rq unlocked.
2117 static int load_balance(int this_cpu, runqueue_t *this_rq,
2118 struct sched_domain *sd, enum idle_type idle)
2120 struct sched_group *group;
2121 runqueue_t *busiest;
2122 unsigned long imbalance;
2123 int nr_moved, all_pinned = 0;
2124 int active_balance = 0;
2127 if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER)
2130 schedstat_inc(sd, lb_cnt[idle]);
2132 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle);
2134 schedstat_inc(sd, lb_nobusyg[idle]);
2138 busiest = find_busiest_queue(group, idle);
2140 schedstat_inc(sd, lb_nobusyq[idle]);
2144 BUG_ON(busiest == this_rq);
2146 schedstat_add(sd, lb_imbalance[idle], imbalance);
2149 if (busiest->nr_running > 1) {
2151 * Attempt to move tasks. If find_busiest_group has found
2152 * an imbalance but busiest->nr_running <= 1, the group is
2153 * still unbalanced. nr_moved simply stays zero, so it is
2154 * correctly treated as an imbalance.
2156 double_rq_lock(this_rq, busiest);
2157 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2158 imbalance, sd, idle, &all_pinned);
2159 double_rq_unlock(this_rq, busiest);
2161 /* All tasks on this runqueue were pinned by CPU affinity */
2162 if (unlikely(all_pinned))
2167 schedstat_inc(sd, lb_failed[idle]);
2168 sd->nr_balance_failed++;
2170 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2172 spin_lock(&busiest->lock);
2174 /* don't kick the migration_thread, if the curr
2175 * task on busiest cpu can't be moved to this_cpu
2177 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2178 spin_unlock(&busiest->lock);
2180 goto out_one_pinned;
2183 if (!busiest->active_balance) {
2184 busiest->active_balance = 1;
2185 busiest->push_cpu = this_cpu;
2188 spin_unlock(&busiest->lock);
2190 wake_up_process(busiest->migration_thread);
2193 * We've kicked active balancing, reset the failure
2196 sd->nr_balance_failed = sd->cache_nice_tries+1;
2199 sd->nr_balance_failed = 0;
2201 if (likely(!active_balance)) {
2202 /* We were unbalanced, so reset the balancing interval */
2203 sd->balance_interval = sd->min_interval;
2206 * If we've begun active balancing, start to back off. This
2207 * case may not be covered by the all_pinned logic if there
2208 * is only 1 task on the busy runqueue (because we don't call
2211 if (sd->balance_interval < sd->max_interval)
2212 sd->balance_interval *= 2;
2215 if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2220 schedstat_inc(sd, lb_balanced[idle]);
2222 sd->nr_balance_failed = 0;
2225 /* tune up the balancing interval */
2226 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2227 (sd->balance_interval < sd->max_interval))
2228 sd->balance_interval *= 2;
2230 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2236 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2237 * tasks if there is an imbalance.
2239 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2240 * this_rq is locked.
2242 static int load_balance_newidle(int this_cpu, runqueue_t *this_rq,
2243 struct sched_domain *sd)
2245 struct sched_group *group;
2246 runqueue_t *busiest = NULL;
2247 unsigned long imbalance;
2251 if (sd->flags & SD_SHARE_CPUPOWER)
2254 schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
2255 group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE, &sd_idle);
2257 schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]);
2261 busiest = find_busiest_queue(group, NEWLY_IDLE);
2263 schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]);
2267 BUG_ON(busiest == this_rq);
2269 schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
2272 if (busiest->nr_running > 1) {
2273 /* Attempt to move tasks */
2274 double_lock_balance(this_rq, busiest);
2275 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2276 imbalance, sd, NEWLY_IDLE, NULL);
2277 spin_unlock(&busiest->lock);
2281 schedstat_inc(sd, lb_failed[NEWLY_IDLE]);
2282 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2285 sd->nr_balance_failed = 0;
2290 schedstat_inc(sd, lb_balanced[NEWLY_IDLE]);
2291 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2293 sd->nr_balance_failed = 0;
2298 * idle_balance is called by schedule() if this_cpu is about to become
2299 * idle. Attempts to pull tasks from other CPUs.
2301 static void idle_balance(int this_cpu, runqueue_t *this_rq)
2303 struct sched_domain *sd;
2305 for_each_domain(this_cpu, sd) {
2306 if (sd->flags & SD_BALANCE_NEWIDLE) {
2307 if (load_balance_newidle(this_cpu, this_rq, sd)) {
2308 /* We've pulled tasks over so stop searching */
2316 * active_load_balance is run by migration threads. It pushes running tasks
2317 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2318 * running on each physical CPU where possible, and avoids physical /
2319 * logical imbalances.
2321 * Called with busiest_rq locked.
2323 static void active_load_balance(runqueue_t *busiest_rq, int busiest_cpu)
2325 struct sched_domain *sd;
2326 runqueue_t *target_rq;
2327 int target_cpu = busiest_rq->push_cpu;
2329 if (busiest_rq->nr_running <= 1)
2330 /* no task to move */
2333 target_rq = cpu_rq(target_cpu);
2336 * This condition is "impossible", if it occurs
2337 * we need to fix it. Originally reported by
2338 * Bjorn Helgaas on a 128-cpu setup.
2340 BUG_ON(busiest_rq == target_rq);
2342 /* move a task from busiest_rq to target_rq */
2343 double_lock_balance(busiest_rq, target_rq);
2345 /* Search for an sd spanning us and the target CPU. */
2346 for_each_domain(target_cpu, sd)
2347 if ((sd->flags & SD_LOAD_BALANCE) &&
2348 cpu_isset(busiest_cpu, sd->span))
2351 if (unlikely(sd == NULL))
2354 schedstat_inc(sd, alb_cnt);
2356 if (move_tasks(target_rq, target_cpu, busiest_rq, 1, sd, SCHED_IDLE, NULL))
2357 schedstat_inc(sd, alb_pushed);
2359 schedstat_inc(sd, alb_failed);
2361 spin_unlock(&target_rq->lock);
2365 * rebalance_tick will get called every timer tick, on every CPU.
2367 * It checks each scheduling domain to see if it is due to be balanced,
2368 * and initiates a balancing operation if so.
2370 * Balancing parameters are set up in arch_init_sched_domains.
2373 /* Don't have all balancing operations going off at once */
2374 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2376 static void rebalance_tick(int this_cpu, runqueue_t *this_rq,
2377 enum idle_type idle)
2379 unsigned long old_load, this_load;
2380 unsigned long j = jiffies + CPU_OFFSET(this_cpu);
2381 struct sched_domain *sd;
2384 this_load = this_rq->nr_running * SCHED_LOAD_SCALE;
2385 /* Update our load */
2386 for (i = 0; i < 3; i++) {
2387 unsigned long new_load = this_load;
2389 old_load = this_rq->cpu_load[i];
2391 * Round up the averaging division if load is increasing. This
2392 * prevents us from getting stuck on 9 if the load is 10, for
2395 if (new_load > old_load)
2396 new_load += scale-1;
2397 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale;
2400 for_each_domain(this_cpu, sd) {
2401 unsigned long interval;
2403 if (!(sd->flags & SD_LOAD_BALANCE))
2406 interval = sd->balance_interval;
2407 if (idle != SCHED_IDLE)
2408 interval *= sd->busy_factor;
2410 /* scale ms to jiffies */
2411 interval = msecs_to_jiffies(interval);
2412 if (unlikely(!interval))
2415 if (j - sd->last_balance >= interval) {
2416 if (load_balance(this_cpu, this_rq, sd, idle)) {
2418 * We've pulled tasks over so either we're no
2419 * longer idle, or one of our SMT siblings is
2424 sd->last_balance += interval;
2430 * on UP we do not need to balance between CPUs:
2432 static inline void rebalance_tick(int cpu, runqueue_t *rq, enum idle_type idle)
2435 static inline void idle_balance(int cpu, runqueue_t *rq)
2440 static inline int wake_priority_sleeper(runqueue_t *rq)
2443 #ifdef CONFIG_SCHED_SMT
2444 spin_lock(&rq->lock);
2446 * If an SMT sibling task has been put to sleep for priority
2447 * reasons reschedule the idle task to see if it can now run.
2449 if (rq->nr_running) {
2450 resched_task(rq->idle);
2453 spin_unlock(&rq->lock);
2458 DEFINE_PER_CPU(struct kernel_stat, kstat);
2460 EXPORT_PER_CPU_SYMBOL(kstat);
2463 * This is called on clock ticks and on context switches.
2464 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2466 static inline void update_cpu_clock(task_t *p, runqueue_t *rq,
2467 unsigned long long now)
2469 unsigned long long last = max(p->timestamp, rq->timestamp_last_tick);
2470 p->sched_time += now - last;
2474 * Return current->sched_time plus any more ns on the sched_clock
2475 * that have not yet been banked.
2477 unsigned long long current_sched_time(const task_t *tsk)
2479 unsigned long long ns;
2480 unsigned long flags;
2481 local_irq_save(flags);
2482 ns = max(tsk->timestamp, task_rq(tsk)->timestamp_last_tick);
2483 ns = tsk->sched_time + (sched_clock() - ns);
2484 local_irq_restore(flags);
2489 * We place interactive tasks back into the active array, if possible.
2491 * To guarantee that this does not starve expired tasks we ignore the
2492 * interactivity of a task if the first expired task had to wait more
2493 * than a 'reasonable' amount of time. This deadline timeout is
2494 * load-dependent, as the frequency of array switched decreases with
2495 * increasing number of running tasks. We also ignore the interactivity
2496 * if a better static_prio task has expired:
2498 #define EXPIRED_STARVING(rq) \
2499 ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2500 (jiffies - (rq)->expired_timestamp >= \
2501 STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2502 ((rq)->curr->static_prio > (rq)->best_expired_prio))
2505 * Account user cpu time to a process.
2506 * @p: the process that the cpu time gets accounted to
2507 * @hardirq_offset: the offset to subtract from hardirq_count()
2508 * @cputime: the cpu time spent in user space since the last update
2510 void account_user_time(struct task_struct *p, cputime_t cputime)
2512 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2515 p->utime = cputime_add(p->utime, cputime);
2517 /* Add user time to cpustat. */
2518 tmp = cputime_to_cputime64(cputime);
2519 if (TASK_NICE(p) > 0)
2520 cpustat->nice = cputime64_add(cpustat->nice, tmp);
2522 cpustat->user = cputime64_add(cpustat->user, tmp);
2526 * Account system cpu time to a process.
2527 * @p: the process that the cpu time gets accounted to
2528 * @hardirq_offset: the offset to subtract from hardirq_count()
2529 * @cputime: the cpu time spent in kernel space since the last update
2531 void account_system_time(struct task_struct *p, int hardirq_offset,
2534 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2535 runqueue_t *rq = this_rq();
2538 p->stime = cputime_add(p->stime, cputime);
2540 /* Add system time to cpustat. */
2541 tmp = cputime_to_cputime64(cputime);
2542 if (hardirq_count() - hardirq_offset)
2543 cpustat->irq = cputime64_add(cpustat->irq, tmp);
2544 else if (softirq_count())
2545 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
2546 else if (p != rq->idle)
2547 cpustat->system = cputime64_add(cpustat->system, tmp);
2548 else if (atomic_read(&rq->nr_iowait) > 0)
2549 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2551 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2552 /* Account for system time used */
2553 acct_update_integrals(p);
2557 * Account for involuntary wait time.
2558 * @p: the process from which the cpu time has been stolen
2559 * @steal: the cpu time spent in involuntary wait
2561 void account_steal_time(struct task_struct *p, cputime_t steal)
2563 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2564 cputime64_t tmp = cputime_to_cputime64(steal);
2565 runqueue_t *rq = this_rq();
2567 if (p == rq->idle) {
2568 p->stime = cputime_add(p->stime, steal);
2569 if (atomic_read(&rq->nr_iowait) > 0)
2570 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2572 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2574 cpustat->steal = cputime64_add(cpustat->steal, tmp);
2578 * This function gets called by the timer code, with HZ frequency.
2579 * We call it with interrupts disabled.
2581 * It also gets called by the fork code, when changing the parent's
2584 void scheduler_tick(void)
2586 int cpu = smp_processor_id();
2587 runqueue_t *rq = this_rq();
2588 task_t *p = current;
2589 unsigned long long now = sched_clock();
2591 update_cpu_clock(p, rq, now);
2593 rq->timestamp_last_tick = now;
2595 if (p == rq->idle) {
2596 if (wake_priority_sleeper(rq))
2598 rebalance_tick(cpu, rq, SCHED_IDLE);
2602 /* Task might have expired already, but not scheduled off yet */
2603 if (p->array != rq->active) {
2604 set_tsk_need_resched(p);
2607 spin_lock(&rq->lock);
2609 * The task was running during this tick - update the
2610 * time slice counter. Note: we do not update a thread's
2611 * priority until it either goes to sleep or uses up its
2612 * timeslice. This makes it possible for interactive tasks
2613 * to use up their timeslices at their highest priority levels.
2617 * RR tasks need a special form of timeslice management.
2618 * FIFO tasks have no timeslices.
2620 if ((p->policy == SCHED_RR) && !--p->time_slice) {
2621 p->time_slice = task_timeslice(p);
2622 p->first_time_slice = 0;
2623 set_tsk_need_resched(p);
2625 /* put it at the end of the queue: */
2626 requeue_task(p, rq->active);
2630 if (!--p->time_slice) {
2631 dequeue_task(p, rq->active);
2632 set_tsk_need_resched(p);
2633 p->prio = effective_prio(p);
2634 p->time_slice = task_timeslice(p);
2635 p->first_time_slice = 0;
2637 if (!rq->expired_timestamp)
2638 rq->expired_timestamp = jiffies;
2639 if (!TASK_INTERACTIVE(p) || EXPIRED_STARVING(rq)) {
2640 enqueue_task(p, rq->expired);
2641 if (p->static_prio < rq->best_expired_prio)
2642 rq->best_expired_prio = p->static_prio;
2644 enqueue_task(p, rq->active);
2647 * Prevent a too long timeslice allowing a task to monopolize
2648 * the CPU. We do this by splitting up the timeslice into
2651 * Note: this does not mean the task's timeslices expire or
2652 * get lost in any way, they just might be preempted by
2653 * another task of equal priority. (one with higher
2654 * priority would have preempted this task already.) We
2655 * requeue this task to the end of the list on this priority
2656 * level, which is in essence a round-robin of tasks with
2659 * This only applies to tasks in the interactive
2660 * delta range with at least TIMESLICE_GRANULARITY to requeue.
2662 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
2663 p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
2664 (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
2665 (p->array == rq->active)) {
2667 requeue_task(p, rq->active);
2668 set_tsk_need_resched(p);
2672 spin_unlock(&rq->lock);
2674 rebalance_tick(cpu, rq, NOT_IDLE);
2677 #ifdef CONFIG_SCHED_SMT
2678 static inline void wakeup_busy_runqueue(runqueue_t *rq)
2680 /* If an SMT runqueue is sleeping due to priority reasons wake it up */
2681 if (rq->curr == rq->idle && rq->nr_running)
2682 resched_task(rq->idle);
2685 static void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2687 struct sched_domain *tmp, *sd = NULL;
2688 cpumask_t sibling_map;
2691 for_each_domain(this_cpu, tmp)
2692 if (tmp->flags & SD_SHARE_CPUPOWER)
2699 * Unlock the current runqueue because we have to lock in
2700 * CPU order to avoid deadlocks. Caller knows that we might
2701 * unlock. We keep IRQs disabled.
2703 spin_unlock(&this_rq->lock);
2705 sibling_map = sd->span;
2707 for_each_cpu_mask(i, sibling_map)
2708 spin_lock(&cpu_rq(i)->lock);
2710 * We clear this CPU from the mask. This both simplifies the
2711 * inner loop and keps this_rq locked when we exit:
2713 cpu_clear(this_cpu, sibling_map);
2715 for_each_cpu_mask(i, sibling_map) {
2716 runqueue_t *smt_rq = cpu_rq(i);
2718 wakeup_busy_runqueue(smt_rq);
2721 for_each_cpu_mask(i, sibling_map)
2722 spin_unlock(&cpu_rq(i)->lock);
2724 * We exit with this_cpu's rq still held and IRQs
2730 * number of 'lost' timeslices this task wont be able to fully
2731 * utilize, if another task runs on a sibling. This models the
2732 * slowdown effect of other tasks running on siblings:
2734 static inline unsigned long smt_slice(task_t *p, struct sched_domain *sd)
2736 return p->time_slice * (100 - sd->per_cpu_gain) / 100;
2739 static int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2741 struct sched_domain *tmp, *sd = NULL;
2742 cpumask_t sibling_map;
2743 prio_array_t *array;
2747 for_each_domain(this_cpu, tmp)
2748 if (tmp->flags & SD_SHARE_CPUPOWER)
2755 * The same locking rules and details apply as for
2756 * wake_sleeping_dependent():
2758 spin_unlock(&this_rq->lock);
2759 sibling_map = sd->span;
2760 for_each_cpu_mask(i, sibling_map)
2761 spin_lock(&cpu_rq(i)->lock);
2762 cpu_clear(this_cpu, sibling_map);
2765 * Establish next task to be run - it might have gone away because
2766 * we released the runqueue lock above:
2768 if (!this_rq->nr_running)
2770 array = this_rq->active;
2771 if (!array->nr_active)
2772 array = this_rq->expired;
2773 BUG_ON(!array->nr_active);
2775 p = list_entry(array->queue[sched_find_first_bit(array->bitmap)].next,
2778 for_each_cpu_mask(i, sibling_map) {
2779 runqueue_t *smt_rq = cpu_rq(i);
2780 task_t *smt_curr = smt_rq->curr;
2782 /* Kernel threads do not participate in dependent sleeping */
2783 if (!p->mm || !smt_curr->mm || rt_task(p))
2784 goto check_smt_task;
2787 * If a user task with lower static priority than the
2788 * running task on the SMT sibling is trying to schedule,
2789 * delay it till there is proportionately less timeslice
2790 * left of the sibling task to prevent a lower priority
2791 * task from using an unfair proportion of the
2792 * physical cpu's resources. -ck
2794 if (rt_task(smt_curr)) {
2796 * With real time tasks we run non-rt tasks only
2797 * per_cpu_gain% of the time.
2799 if ((jiffies % DEF_TIMESLICE) >
2800 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
2803 if (smt_curr->static_prio < p->static_prio &&
2804 !TASK_PREEMPTS_CURR(p, smt_rq) &&
2805 smt_slice(smt_curr, sd) > task_timeslice(p))
2809 if ((!smt_curr->mm && smt_curr != smt_rq->idle) ||
2813 wakeup_busy_runqueue(smt_rq);
2818 * Reschedule a lower priority task on the SMT sibling for
2819 * it to be put to sleep, or wake it up if it has been put to
2820 * sleep for priority reasons to see if it should run now.
2823 if ((jiffies % DEF_TIMESLICE) >
2824 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
2825 resched_task(smt_curr);
2827 if (TASK_PREEMPTS_CURR(p, smt_rq) &&
2828 smt_slice(p, sd) > task_timeslice(smt_curr))
2829 resched_task(smt_curr);
2831 wakeup_busy_runqueue(smt_rq);
2835 for_each_cpu_mask(i, sibling_map)
2836 spin_unlock(&cpu_rq(i)->lock);
2840 static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2844 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2850 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
2852 void fastcall add_preempt_count(int val)
2857 BUG_ON((preempt_count() < 0));
2858 preempt_count() += val;
2860 * Spinlock count overflowing soon?
2862 BUG_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10);
2864 EXPORT_SYMBOL(add_preempt_count);
2866 void fastcall sub_preempt_count(int val)
2871 BUG_ON(val > preempt_count());
2873 * Is the spinlock portion underflowing?
2875 BUG_ON((val < PREEMPT_MASK) && !(preempt_count() & PREEMPT_MASK));
2876 preempt_count() -= val;
2878 EXPORT_SYMBOL(sub_preempt_count);
2882 static inline int interactive_sleep(enum sleep_type sleep_type)
2884 return (sleep_type == SLEEP_INTERACTIVE ||
2885 sleep_type == SLEEP_INTERRUPTED);
2889 * schedule() is the main scheduler function.
2891 asmlinkage void __sched schedule(void)
2894 task_t *prev, *next;
2896 prio_array_t *array;
2897 struct list_head *queue;
2898 unsigned long long now;
2899 unsigned long run_time;
2900 int cpu, idx, new_prio;
2903 * Test if we are atomic. Since do_exit() needs to call into
2904 * schedule() atomically, we ignore that path for now.
2905 * Otherwise, whine if we are scheduling when we should not be.
2907 if (unlikely(in_atomic() && !current->exit_state)) {
2908 printk(KERN_ERR "BUG: scheduling while atomic: "
2910 current->comm, preempt_count(), current->pid);
2913 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2918 release_kernel_lock(prev);
2919 need_resched_nonpreemptible:
2923 * The idle thread is not allowed to schedule!
2924 * Remove this check after it has been exercised a bit.
2926 if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
2927 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
2931 schedstat_inc(rq, sched_cnt);
2932 now = sched_clock();
2933 if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
2934 run_time = now - prev->timestamp;
2935 if (unlikely((long long)(now - prev->timestamp) < 0))
2938 run_time = NS_MAX_SLEEP_AVG;
2941 * Tasks charged proportionately less run_time at high sleep_avg to
2942 * delay them losing their interactive status
2944 run_time /= (CURRENT_BONUS(prev) ? : 1);
2946 spin_lock_irq(&rq->lock);
2948 if (unlikely(prev->flags & PF_DEAD))
2949 prev->state = EXIT_DEAD;
2951 switch_count = &prev->nivcsw;
2952 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2953 switch_count = &prev->nvcsw;
2954 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
2955 unlikely(signal_pending(prev))))
2956 prev->state = TASK_RUNNING;
2958 if (prev->state == TASK_UNINTERRUPTIBLE)
2959 rq->nr_uninterruptible++;
2960 deactivate_task(prev, rq);
2964 cpu = smp_processor_id();
2965 if (unlikely(!rq->nr_running)) {
2967 idle_balance(cpu, rq);
2968 if (!rq->nr_running) {
2970 rq->expired_timestamp = 0;
2971 wake_sleeping_dependent(cpu, rq);
2973 * wake_sleeping_dependent() might have released
2974 * the runqueue, so break out if we got new
2977 if (!rq->nr_running)
2981 if (dependent_sleeper(cpu, rq)) {
2986 * dependent_sleeper() releases and reacquires the runqueue
2987 * lock, hence go into the idle loop if the rq went
2990 if (unlikely(!rq->nr_running))
2995 if (unlikely(!array->nr_active)) {
2997 * Switch the active and expired arrays.
2999 schedstat_inc(rq, sched_switch);
3000 rq->active = rq->expired;
3001 rq->expired = array;
3003 rq->expired_timestamp = 0;
3004 rq->best_expired_prio = MAX_PRIO;
3007 idx = sched_find_first_bit(array->bitmap);
3008 queue = array->queue + idx;
3009 next = list_entry(queue->next, task_t, run_list);
3011 if (!rt_task(next) && interactive_sleep(next->sleep_type)) {
3012 unsigned long long delta = now - next->timestamp;
3013 if (unlikely((long long)(now - next->timestamp) < 0))
3016 if (next->sleep_type == SLEEP_INTERACTIVE)
3017 delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
3019 array = next->array;
3020 new_prio = recalc_task_prio(next, next->timestamp + delta);
3022 if (unlikely(next->prio != new_prio)) {
3023 dequeue_task(next, array);
3024 next->prio = new_prio;
3025 enqueue_task(next, array);
3027 requeue_task(next, array);
3029 next->sleep_type = SLEEP_NORMAL;
3031 if (next == rq->idle)
3032 schedstat_inc(rq, sched_goidle);
3034 prefetch_stack(next);
3035 clear_tsk_need_resched(prev);
3036 rcu_qsctr_inc(task_cpu(prev));
3038 update_cpu_clock(prev, rq, now);
3040 prev->sleep_avg -= run_time;
3041 if ((long)prev->sleep_avg <= 0)
3042 prev->sleep_avg = 0;
3043 prev->timestamp = prev->last_ran = now;
3045 sched_info_switch(prev, next);
3046 if (likely(prev != next)) {
3047 next->timestamp = now;
3052 prepare_task_switch(rq, next);
3053 prev = context_switch(rq, prev, next);
3056 * this_rq must be evaluated again because prev may have moved
3057 * CPUs since it called schedule(), thus the 'rq' on its stack
3058 * frame will be invalid.
3060 finish_task_switch(this_rq(), prev);
3062 spin_unlock_irq(&rq->lock);
3065 if (unlikely(reacquire_kernel_lock(prev) < 0))
3066 goto need_resched_nonpreemptible;
3067 preempt_enable_no_resched();
3068 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3072 EXPORT_SYMBOL(schedule);
3074 #ifdef CONFIG_PREEMPT
3076 * this is is the entry point to schedule() from in-kernel preemption
3077 * off of preempt_enable. Kernel preemptions off return from interrupt
3078 * occur there and call schedule directly.
3080 asmlinkage void __sched preempt_schedule(void)
3082 struct thread_info *ti = current_thread_info();
3083 #ifdef CONFIG_PREEMPT_BKL
3084 struct task_struct *task = current;
3085 int saved_lock_depth;
3088 * If there is a non-zero preempt_count or interrupts are disabled,
3089 * we do not want to preempt the current task. Just return..
3091 if (unlikely(ti->preempt_count || irqs_disabled()))
3095 add_preempt_count(PREEMPT_ACTIVE);
3097 * We keep the big kernel semaphore locked, but we
3098 * clear ->lock_depth so that schedule() doesnt
3099 * auto-release the semaphore:
3101 #ifdef CONFIG_PREEMPT_BKL
3102 saved_lock_depth = task->lock_depth;
3103 task->lock_depth = -1;
3106 #ifdef CONFIG_PREEMPT_BKL
3107 task->lock_depth = saved_lock_depth;
3109 sub_preempt_count(PREEMPT_ACTIVE);
3111 /* we could miss a preemption opportunity between schedule and now */
3113 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3117 EXPORT_SYMBOL(preempt_schedule);
3120 * this is is the entry point to schedule() from kernel preemption
3121 * off of irq context.
3122 * Note, that this is called and return with irqs disabled. This will
3123 * protect us against recursive calling from irq.
3125 asmlinkage void __sched preempt_schedule_irq(void)
3127 struct thread_info *ti = current_thread_info();
3128 #ifdef CONFIG_PREEMPT_BKL
3129 struct task_struct *task = current;
3130 int saved_lock_depth;
3132 /* Catch callers which need to be fixed*/
3133 BUG_ON(ti->preempt_count || !irqs_disabled());
3136 add_preempt_count(PREEMPT_ACTIVE);
3138 * We keep the big kernel semaphore locked, but we
3139 * clear ->lock_depth so that schedule() doesnt
3140 * auto-release the semaphore:
3142 #ifdef CONFIG_PREEMPT_BKL
3143 saved_lock_depth = task->lock_depth;
3144 task->lock_depth = -1;
3148 local_irq_disable();
3149 #ifdef CONFIG_PREEMPT_BKL
3150 task->lock_depth = saved_lock_depth;
3152 sub_preempt_count(PREEMPT_ACTIVE);
3154 /* we could miss a preemption opportunity between schedule and now */
3156 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3160 #endif /* CONFIG_PREEMPT */
3162 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3165 task_t *p = curr->private;
3166 return try_to_wake_up(p, mode, sync);
3169 EXPORT_SYMBOL(default_wake_function);
3172 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3173 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3174 * number) then we wake all the non-exclusive tasks and one exclusive task.
3176 * There are circumstances in which we can try to wake a task which has already
3177 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3178 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3180 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3181 int nr_exclusive, int sync, void *key)
3183 struct list_head *tmp, *next;
3185 list_for_each_safe(tmp, next, &q->task_list) {
3188 curr = list_entry(tmp, wait_queue_t, task_list);
3189 flags = curr->flags;
3190 if (curr->func(curr, mode, sync, key) &&
3191 (flags & WQ_FLAG_EXCLUSIVE) &&
3198 * __wake_up - wake up threads blocked on a waitqueue.
3200 * @mode: which threads
3201 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3202 * @key: is directly passed to the wakeup function
3204 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3205 int nr_exclusive, void *key)
3207 unsigned long flags;
3209 spin_lock_irqsave(&q->lock, flags);
3210 __wake_up_common(q, mode, nr_exclusive, 0, key);
3211 spin_unlock_irqrestore(&q->lock, flags);
3214 EXPORT_SYMBOL(__wake_up);
3217 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3219 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3221 __wake_up_common(q, mode, 1, 0, NULL);
3225 * __wake_up_sync - wake up threads blocked on a waitqueue.
3227 * @mode: which threads
3228 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3230 * The sync wakeup differs that the waker knows that it will schedule
3231 * away soon, so while the target thread will be woken up, it will not
3232 * be migrated to another CPU - ie. the two threads are 'synchronized'
3233 * with each other. This can prevent needless bouncing between CPUs.
3235 * On UP it can prevent extra preemption.
3238 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3240 unsigned long flags;
3246 if (unlikely(!nr_exclusive))
3249 spin_lock_irqsave(&q->lock, flags);
3250 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3251 spin_unlock_irqrestore(&q->lock, flags);
3253 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3255 void fastcall complete(struct completion *x)
3257 unsigned long flags;
3259 spin_lock_irqsave(&x->wait.lock, flags);
3261 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3263 spin_unlock_irqrestore(&x->wait.lock, flags);
3265 EXPORT_SYMBOL(complete);
3267 void fastcall complete_all(struct completion *x)
3269 unsigned long flags;
3271 spin_lock_irqsave(&x->wait.lock, flags);
3272 x->done += UINT_MAX/2;
3273 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3275 spin_unlock_irqrestore(&x->wait.lock, flags);
3277 EXPORT_SYMBOL(complete_all);
3279 void fastcall __sched wait_for_completion(struct completion *x)
3282 spin_lock_irq(&x->wait.lock);
3284 DECLARE_WAITQUEUE(wait, current);
3286 wait.flags |= WQ_FLAG_EXCLUSIVE;
3287 __add_wait_queue_tail(&x->wait, &wait);
3289 __set_current_state(TASK_UNINTERRUPTIBLE);
3290 spin_unlock_irq(&x->wait.lock);
3292 spin_lock_irq(&x->wait.lock);
3294 __remove_wait_queue(&x->wait, &wait);
3297 spin_unlock_irq(&x->wait.lock);
3299 EXPORT_SYMBOL(wait_for_completion);
3301 unsigned long fastcall __sched
3302 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3306 spin_lock_irq(&x->wait.lock);
3308 DECLARE_WAITQUEUE(wait, current);
3310 wait.flags |= WQ_FLAG_EXCLUSIVE;
3311 __add_wait_queue_tail(&x->wait, &wait);
3313 __set_current_state(TASK_UNINTERRUPTIBLE);
3314 spin_unlock_irq(&x->wait.lock);
3315 timeout = schedule_timeout(timeout);
3316 spin_lock_irq(&x->wait.lock);
3318 __remove_wait_queue(&x->wait, &wait);
3322 __remove_wait_queue(&x->wait, &wait);
3326 spin_unlock_irq(&x->wait.lock);
3329 EXPORT_SYMBOL(wait_for_completion_timeout);
3331 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3337 spin_lock_irq(&x->wait.lock);
3339 DECLARE_WAITQUEUE(wait, current);
3341 wait.flags |= WQ_FLAG_EXCLUSIVE;
3342 __add_wait_queue_tail(&x->wait, &wait);
3344 if (signal_pending(current)) {
3346 __remove_wait_queue(&x->wait, &wait);
3349 __set_current_state(TASK_INTERRUPTIBLE);
3350 spin_unlock_irq(&x->wait.lock);
3352 spin_lock_irq(&x->wait.lock);
3354 __remove_wait_queue(&x->wait, &wait);
3358 spin_unlock_irq(&x->wait.lock);
3362 EXPORT_SYMBOL(wait_for_completion_interruptible);
3364 unsigned long fastcall __sched
3365 wait_for_completion_interruptible_timeout(struct completion *x,
3366 unsigned long timeout)
3370 spin_lock_irq(&x->wait.lock);
3372 DECLARE_WAITQUEUE(wait, current);
3374 wait.flags |= WQ_FLAG_EXCLUSIVE;
3375 __add_wait_queue_tail(&x->wait, &wait);
3377 if (signal_pending(current)) {
3378 timeout = -ERESTARTSYS;
3379 __remove_wait_queue(&x->wait, &wait);
3382 __set_current_state(TASK_INTERRUPTIBLE);
3383 spin_unlock_irq(&x->wait.lock);
3384 timeout = schedule_timeout(timeout);
3385 spin_lock_irq(&x->wait.lock);
3387 __remove_wait_queue(&x->wait, &wait);
3391 __remove_wait_queue(&x->wait, &wait);
3395 spin_unlock_irq(&x->wait.lock);
3398 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3401 #define SLEEP_ON_VAR \
3402 unsigned long flags; \
3403 wait_queue_t wait; \
3404 init_waitqueue_entry(&wait, current);
3406 #define SLEEP_ON_HEAD \
3407 spin_lock_irqsave(&q->lock,flags); \
3408 __add_wait_queue(q, &wait); \
3409 spin_unlock(&q->lock);
3411 #define SLEEP_ON_TAIL \
3412 spin_lock_irq(&q->lock); \
3413 __remove_wait_queue(q, &wait); \
3414 spin_unlock_irqrestore(&q->lock, flags);
3416 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
3420 current->state = TASK_INTERRUPTIBLE;
3427 EXPORT_SYMBOL(interruptible_sleep_on);
3429 long fastcall __sched
3430 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3434 current->state = TASK_INTERRUPTIBLE;
3437 timeout = schedule_timeout(timeout);
3443 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3445 void fastcall __sched sleep_on(wait_queue_head_t *q)
3449 current->state = TASK_UNINTERRUPTIBLE;
3456 EXPORT_SYMBOL(sleep_on);
3458 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3462 current->state = TASK_UNINTERRUPTIBLE;
3465 timeout = schedule_timeout(timeout);
3471 EXPORT_SYMBOL(sleep_on_timeout);
3473 void set_user_nice(task_t *p, long nice)
3475 unsigned long flags;
3476 prio_array_t *array;
3478 int old_prio, new_prio, delta;
3480 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3483 * We have to be careful, if called from sys_setpriority(),
3484 * the task might be in the middle of scheduling on another CPU.
3486 rq = task_rq_lock(p, &flags);
3488 * The RT priorities are set via sched_setscheduler(), but we still
3489 * allow the 'normal' nice value to be set - but as expected
3490 * it wont have any effect on scheduling until the task is
3491 * not SCHED_NORMAL/SCHED_BATCH:
3494 p->static_prio = NICE_TO_PRIO(nice);
3499 dequeue_task(p, array);
3502 new_prio = NICE_TO_PRIO(nice);
3503 delta = new_prio - old_prio;
3504 p->static_prio = NICE_TO_PRIO(nice);
3508 enqueue_task(p, array);
3510 * If the task increased its priority or is running and
3511 * lowered its priority, then reschedule its CPU:
3513 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3514 resched_task(rq->curr);
3517 task_rq_unlock(rq, &flags);
3520 EXPORT_SYMBOL(set_user_nice);
3523 * can_nice - check if a task can reduce its nice value
3527 int can_nice(const task_t *p, const int nice)
3529 /* convert nice value [19,-20] to rlimit style value [1,40] */
3530 int nice_rlim = 20 - nice;
3531 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
3532 capable(CAP_SYS_NICE));
3535 #ifdef __ARCH_WANT_SYS_NICE
3538 * sys_nice - change the priority of the current process.
3539 * @increment: priority increment
3541 * sys_setpriority is a more generic, but much slower function that
3542 * does similar things.
3544 asmlinkage long sys_nice(int increment)
3550 * Setpriority might change our priority at the same moment.
3551 * We don't have to worry. Conceptually one call occurs first
3552 * and we have a single winner.
3554 if (increment < -40)
3559 nice = PRIO_TO_NICE(current->static_prio) + increment;
3565 if (increment < 0 && !can_nice(current, nice))
3568 retval = security_task_setnice(current, nice);
3572 set_user_nice(current, nice);
3579 * task_prio - return the priority value of a given task.
3580 * @p: the task in question.
3582 * This is the priority value as seen by users in /proc.
3583 * RT tasks are offset by -200. Normal tasks are centered
3584 * around 0, value goes from -16 to +15.
3586 int task_prio(const task_t *p)
3588 return p->prio - MAX_RT_PRIO;
3592 * task_nice - return the nice value of a given task.
3593 * @p: the task in question.
3595 int task_nice(const task_t *p)
3597 return TASK_NICE(p);
3599 EXPORT_SYMBOL_GPL(task_nice);
3602 * idle_cpu - is a given cpu idle currently?
3603 * @cpu: the processor in question.
3605 int idle_cpu(int cpu)
3607 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
3611 * idle_task - return the idle task for a given cpu.
3612 * @cpu: the processor in question.
3614 task_t *idle_task(int cpu)
3616 return cpu_rq(cpu)->idle;
3620 * find_process_by_pid - find a process with a matching PID value.
3621 * @pid: the pid in question.
3623 static inline task_t *find_process_by_pid(pid_t pid)
3625 return pid ? find_task_by_pid(pid) : current;
3628 /* Actually do priority change: must hold rq lock. */
3629 static void __setscheduler(struct task_struct *p, int policy, int prio)
3633 p->rt_priority = prio;
3634 if (policy != SCHED_NORMAL && policy != SCHED_BATCH) {
3635 p->prio = MAX_RT_PRIO-1 - p->rt_priority;
3637 p->prio = p->static_prio;
3639 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
3641 if (policy == SCHED_BATCH)
3647 * sched_setscheduler - change the scheduling policy and/or RT priority of
3649 * @p: the task in question.
3650 * @policy: new policy.
3651 * @param: structure containing the new RT priority.
3653 int sched_setscheduler(struct task_struct *p, int policy,
3654 struct sched_param *param)
3657 int oldprio, oldpolicy = -1;
3658 prio_array_t *array;
3659 unsigned long flags;
3663 /* double check policy once rq lock held */
3665 policy = oldpolicy = p->policy;
3666 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
3667 policy != SCHED_NORMAL && policy != SCHED_BATCH)
3670 * Valid priorities for SCHED_FIFO and SCHED_RR are
3671 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
3674 if (param->sched_priority < 0 ||
3675 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3676 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3678 if ((policy == SCHED_NORMAL || policy == SCHED_BATCH)
3679 != (param->sched_priority == 0))
3683 * Allow unprivileged RT tasks to decrease priority:
3685 if (!capable(CAP_SYS_NICE)) {
3687 * can't change policy, except between SCHED_NORMAL
3690 if (((policy != SCHED_NORMAL && p->policy != SCHED_BATCH) &&
3691 (policy != SCHED_BATCH && p->policy != SCHED_NORMAL)) &&
3692 !p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
3694 /* can't increase priority */
3695 if ((policy != SCHED_NORMAL && policy != SCHED_BATCH) &&
3696 param->sched_priority > p->rt_priority &&
3697 param->sched_priority >
3698 p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
3700 /* can't change other user's priorities */
3701 if ((current->euid != p->euid) &&
3702 (current->euid != p->uid))
3706 retval = security_task_setscheduler(p, policy, param);
3710 * To be able to change p->policy safely, the apropriate
3711 * runqueue lock must be held.
3713 rq = task_rq_lock(p, &flags);
3714 /* recheck policy now with rq lock held */
3715 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3716 policy = oldpolicy = -1;
3717 task_rq_unlock(rq, &flags);
3722 deactivate_task(p, rq);
3724 __setscheduler(p, policy, param->sched_priority);
3726 __activate_task(p, rq);
3728 * Reschedule if we are currently running on this runqueue and
3729 * our priority decreased, or if we are not currently running on
3730 * this runqueue and our priority is higher than the current's
3732 if (task_running(rq, p)) {
3733 if (p->prio > oldprio)
3734 resched_task(rq->curr);
3735 } else if (TASK_PREEMPTS_CURR(p, rq))
3736 resched_task(rq->curr);
3738 task_rq_unlock(rq, &flags);
3741 EXPORT_SYMBOL_GPL(sched_setscheduler);
3744 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3747 struct sched_param lparam;
3748 struct task_struct *p;
3750 if (!param || pid < 0)
3752 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3754 read_lock_irq(&tasklist_lock);
3755 p = find_process_by_pid(pid);
3757 read_unlock_irq(&tasklist_lock);
3760 retval = sched_setscheduler(p, policy, &lparam);
3761 read_unlock_irq(&tasklist_lock);
3766 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3767 * @pid: the pid in question.
3768 * @policy: new policy.
3769 * @param: structure containing the new RT priority.
3771 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
3772 struct sched_param __user *param)
3774 /* negative values for policy are not valid */
3778 return do_sched_setscheduler(pid, policy, param);
3782 * sys_sched_setparam - set/change the RT priority of a thread
3783 * @pid: the pid in question.
3784 * @param: structure containing the new RT priority.
3786 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
3788 return do_sched_setscheduler(pid, -1, param);
3792 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3793 * @pid: the pid in question.
3795 asmlinkage long sys_sched_getscheduler(pid_t pid)
3797 int retval = -EINVAL;
3804 read_lock(&tasklist_lock);
3805 p = find_process_by_pid(pid);
3807 retval = security_task_getscheduler(p);
3811 read_unlock(&tasklist_lock);
3818 * sys_sched_getscheduler - get the RT priority of a thread
3819 * @pid: the pid in question.
3820 * @param: structure containing the RT priority.
3822 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
3824 struct sched_param lp;
3825 int retval = -EINVAL;
3828 if (!param || pid < 0)
3831 read_lock(&tasklist_lock);
3832 p = find_process_by_pid(pid);
3837 retval = security_task_getscheduler(p);
3841 lp.sched_priority = p->rt_priority;
3842 read_unlock(&tasklist_lock);
3845 * This one might sleep, we cannot do it with a spinlock held ...
3847 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3853 read_unlock(&tasklist_lock);
3857 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
3861 cpumask_t cpus_allowed;
3864 read_lock(&tasklist_lock);
3866 p = find_process_by_pid(pid);
3868 read_unlock(&tasklist_lock);
3869 unlock_cpu_hotplug();
3874 * It is not safe to call set_cpus_allowed with the
3875 * tasklist_lock held. We will bump the task_struct's
3876 * usage count and then drop tasklist_lock.
3879 read_unlock(&tasklist_lock);
3882 if ((current->euid != p->euid) && (current->euid != p->uid) &&
3883 !capable(CAP_SYS_NICE))
3886 cpus_allowed = cpuset_cpus_allowed(p);
3887 cpus_and(new_mask, new_mask, cpus_allowed);
3888 retval = set_cpus_allowed(p, new_mask);
3892 unlock_cpu_hotplug();
3896 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3897 cpumask_t *new_mask)
3899 if (len < sizeof(cpumask_t)) {
3900 memset(new_mask, 0, sizeof(cpumask_t));
3901 } else if (len > sizeof(cpumask_t)) {
3902 len = sizeof(cpumask_t);
3904 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3908 * sys_sched_setaffinity - set the cpu affinity of a process
3909 * @pid: pid of the process
3910 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3911 * @user_mask_ptr: user-space pointer to the new cpu mask
3913 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
3914 unsigned long __user *user_mask_ptr)
3919 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
3923 return sched_setaffinity(pid, new_mask);
3927 * Represents all cpu's present in the system
3928 * In systems capable of hotplug, this map could dynamically grow
3929 * as new cpu's are detected in the system via any platform specific
3930 * method, such as ACPI for e.g.
3933 cpumask_t cpu_present_map __read_mostly;
3934 EXPORT_SYMBOL(cpu_present_map);
3937 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
3938 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
3941 long sched_getaffinity(pid_t pid, cpumask_t *mask)
3947 read_lock(&tasklist_lock);
3950 p = find_process_by_pid(pid);
3955 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
3958 read_unlock(&tasklist_lock);
3959 unlock_cpu_hotplug();
3967 * sys_sched_getaffinity - get the cpu affinity of a process
3968 * @pid: pid of the process
3969 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3970 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3972 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
3973 unsigned long __user *user_mask_ptr)
3978 if (len < sizeof(cpumask_t))
3981 ret = sched_getaffinity(pid, &mask);
3985 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
3988 return sizeof(cpumask_t);
3992 * sys_sched_yield - yield the current processor to other threads.
3994 * this function yields the current CPU by moving the calling thread
3995 * to the expired array. If there are no other threads running on this
3996 * CPU then this function will return.
3998 asmlinkage long sys_sched_yield(void)
4000 runqueue_t *rq = this_rq_lock();
4001 prio_array_t *array = current->array;
4002 prio_array_t *target = rq->expired;
4004 schedstat_inc(rq, yld_cnt);
4006 * We implement yielding by moving the task into the expired
4009 * (special rule: RT tasks will just roundrobin in the active
4012 if (rt_task(current))
4013 target = rq->active;
4015 if (array->nr_active == 1) {
4016 schedstat_inc(rq, yld_act_empty);
4017 if (!rq->expired->nr_active)
4018 schedstat_inc(rq, yld_both_empty);
4019 } else if (!rq->expired->nr_active)
4020 schedstat_inc(rq, yld_exp_empty);
4022 if (array != target) {
4023 dequeue_task(current, array);
4024 enqueue_task(current, target);
4027 * requeue_task is cheaper so perform that if possible.
4029 requeue_task(current, array);
4032 * Since we are going to call schedule() anyway, there's
4033 * no need to preempt or enable interrupts:
4035 __release(rq->lock);
4036 _raw_spin_unlock(&rq->lock);
4037 preempt_enable_no_resched();
4044 static inline void __cond_resched(void)
4047 * The BKS might be reacquired before we have dropped
4048 * PREEMPT_ACTIVE, which could trigger a second
4049 * cond_resched() call.
4051 if (unlikely(preempt_count()))
4053 if (unlikely(system_state != SYSTEM_RUNNING))
4056 add_preempt_count(PREEMPT_ACTIVE);
4058 sub_preempt_count(PREEMPT_ACTIVE);
4059 } while (need_resched());
4062 int __sched cond_resched(void)
4064 if (need_resched()) {
4071 EXPORT_SYMBOL(cond_resched);
4074 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4075 * call schedule, and on return reacquire the lock.
4077 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4078 * operations here to prevent schedule() from being called twice (once via
4079 * spin_unlock(), once by hand).
4081 int cond_resched_lock(spinlock_t *lock)
4085 if (need_lockbreak(lock)) {
4091 if (need_resched()) {
4092 _raw_spin_unlock(lock);
4093 preempt_enable_no_resched();
4101 EXPORT_SYMBOL(cond_resched_lock);
4103 int __sched cond_resched_softirq(void)
4105 BUG_ON(!in_softirq());
4107 if (need_resched()) {
4108 __local_bh_enable();
4116 EXPORT_SYMBOL(cond_resched_softirq);
4120 * yield - yield the current processor to other threads.
4122 * this is a shortcut for kernel-space yielding - it marks the
4123 * thread runnable and calls sys_sched_yield().
4125 void __sched yield(void)
4127 set_current_state(TASK_RUNNING);
4131 EXPORT_SYMBOL(yield);
4134 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4135 * that process accounting knows that this is a task in IO wait state.
4137 * But don't do that if it is a deliberate, throttling IO wait (this task
4138 * has set its backing_dev_info: the queue against which it should throttle)
4140 void __sched io_schedule(void)
4142 struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id());
4144 atomic_inc(&rq->nr_iowait);
4146 atomic_dec(&rq->nr_iowait);
4149 EXPORT_SYMBOL(io_schedule);
4151 long __sched io_schedule_timeout(long timeout)
4153 struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id());
4156 atomic_inc(&rq->nr_iowait);
4157 ret = schedule_timeout(timeout);
4158 atomic_dec(&rq->nr_iowait);
4163 * sys_sched_get_priority_max - return maximum RT priority.
4164 * @policy: scheduling class.
4166 * this syscall returns the maximum rt_priority that can be used
4167 * by a given scheduling class.
4169 asmlinkage long sys_sched_get_priority_max(int policy)
4176 ret = MAX_USER_RT_PRIO-1;
4187 * sys_sched_get_priority_min - return minimum RT priority.
4188 * @policy: scheduling class.
4190 * this syscall returns the minimum rt_priority that can be used
4191 * by a given scheduling class.
4193 asmlinkage long sys_sched_get_priority_min(int policy)
4210 * sys_sched_rr_get_interval - return the default timeslice of a process.
4211 * @pid: pid of the process.
4212 * @interval: userspace pointer to the timeslice value.
4214 * this syscall writes the default timeslice value of a given process
4215 * into the user-space timespec buffer. A value of '0' means infinity.
4218 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4220 int retval = -EINVAL;
4228 read_lock(&tasklist_lock);
4229 p = find_process_by_pid(pid);
4233 retval = security_task_getscheduler(p);
4237 jiffies_to_timespec(p->policy & SCHED_FIFO ?
4238 0 : task_timeslice(p), &t);
4239 read_unlock(&tasklist_lock);
4240 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4244 read_unlock(&tasklist_lock);
4248 static inline struct task_struct *eldest_child(struct task_struct *p)
4250 if (list_empty(&p->children)) return NULL;
4251 return list_entry(p->children.next,struct task_struct,sibling);
4254 static inline struct task_struct *older_sibling(struct task_struct *p)
4256 if (p->sibling.prev==&p->parent->children) return NULL;
4257 return list_entry(p->sibling.prev,struct task_struct,sibling);
4260 static inline struct task_struct *younger_sibling(struct task_struct *p)
4262 if (p->sibling.next==&p->parent->children) return NULL;
4263 return list_entry(p->sibling.next,struct task_struct,sibling);
4266 static void show_task(task_t *p)
4270 unsigned long free = 0;
4271 static const char *stat_nam[] = { "R", "S", "D", "T", "t", "Z", "X" };
4273 printk("%-13.13s ", p->comm);
4274 state = p->state ? __ffs(p->state) + 1 : 0;
4275 if (state < ARRAY_SIZE(stat_nam))
4276 printk(stat_nam[state]);
4279 #if (BITS_PER_LONG == 32)
4280 if (state == TASK_RUNNING)
4281 printk(" running ");
4283 printk(" %08lX ", thread_saved_pc(p));
4285 if (state == TASK_RUNNING)
4286 printk(" running task ");
4288 printk(" %016lx ", thread_saved_pc(p));
4290 #ifdef CONFIG_DEBUG_STACK_USAGE
4292 unsigned long *n = end_of_stack(p);
4295 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4298 printk("%5lu %5d %6d ", free, p->pid, p->parent->pid);
4299 if ((relative = eldest_child(p)))
4300 printk("%5d ", relative->pid);
4303 if ((relative = younger_sibling(p)))
4304 printk("%7d", relative->pid);
4307 if ((relative = older_sibling(p)))
4308 printk(" %5d", relative->pid);
4312 printk(" (L-TLB)\n");
4314 printk(" (NOTLB)\n");
4316 if (state != TASK_RUNNING)
4317 show_stack(p, NULL);
4320 void show_state(void)
4324 #if (BITS_PER_LONG == 32)
4327 printk(" task PC pid father child younger older\n");
4331 printk(" task PC pid father child younger older\n");
4333 read_lock(&tasklist_lock);
4334 do_each_thread(g, p) {
4336 * reset the NMI-timeout, listing all files on a slow
4337 * console might take alot of time:
4339 touch_nmi_watchdog();
4341 } while_each_thread(g, p);
4343 read_unlock(&tasklist_lock);
4344 mutex_debug_show_all_locks();
4348 * init_idle - set up an idle thread for a given CPU
4349 * @idle: task in question
4350 * @cpu: cpu the idle task belongs to
4352 * NOTE: this function does not set the idle thread's NEED_RESCHED
4353 * flag, to make booting more robust.
4355 void __devinit init_idle(task_t *idle, int cpu)
4357 runqueue_t *rq = cpu_rq(cpu);
4358 unsigned long flags;
4360 idle->timestamp = sched_clock();
4361 idle->sleep_avg = 0;
4363 idle->prio = MAX_PRIO;
4364 idle->state = TASK_RUNNING;
4365 idle->cpus_allowed = cpumask_of_cpu(cpu);
4366 set_task_cpu(idle, cpu);
4368 spin_lock_irqsave(&rq->lock, flags);
4369 rq->curr = rq->idle = idle;
4370 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4373 spin_unlock_irqrestore(&rq->lock, flags);
4375 /* Set the preempt count _outside_ the spinlocks! */
4376 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4377 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4379 task_thread_info(idle)->preempt_count = 0;
4384 * In a system that switches off the HZ timer nohz_cpu_mask
4385 * indicates which cpus entered this state. This is used
4386 * in the rcu update to wait only for active cpus. For system
4387 * which do not switch off the HZ timer nohz_cpu_mask should
4388 * always be CPU_MASK_NONE.
4390 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4394 * This is how migration works:
4396 * 1) we queue a migration_req_t structure in the source CPU's
4397 * runqueue and wake up that CPU's migration thread.
4398 * 2) we down() the locked semaphore => thread blocks.
4399 * 3) migration thread wakes up (implicitly it forces the migrated
4400 * thread off the CPU)
4401 * 4) it gets the migration request and checks whether the migrated
4402 * task is still in the wrong runqueue.
4403 * 5) if it's in the wrong runqueue then the migration thread removes
4404 * it and puts it into the right queue.
4405 * 6) migration thread up()s the semaphore.
4406 * 7) we wake up and the migration is done.
4410 * Change a given task's CPU affinity. Migrate the thread to a
4411 * proper CPU and schedule it away if the CPU it's executing on
4412 * is removed from the allowed bitmask.
4414 * NOTE: the caller must have a valid reference to the task, the
4415 * task must not exit() & deallocate itself prematurely. The
4416 * call is not atomic; no spinlocks may be held.
4418 int set_cpus_allowed(task_t *p, cpumask_t new_mask)
4420 unsigned long flags;
4422 migration_req_t req;
4425 rq = task_rq_lock(p, &flags);
4426 if (!cpus_intersects(new_mask, cpu_online_map)) {
4431 p->cpus_allowed = new_mask;
4432 /* Can the task run on the task's current CPU? If so, we're done */
4433 if (cpu_isset(task_cpu(p), new_mask))
4436 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4437 /* Need help from migration thread: drop lock and wait. */
4438 task_rq_unlock(rq, &flags);
4439 wake_up_process(rq->migration_thread);
4440 wait_for_completion(&req.done);
4441 tlb_migrate_finish(p->mm);
4445 task_rq_unlock(rq, &flags);
4449 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4452 * Move (not current) task off this cpu, onto dest cpu. We're doing
4453 * this because either it can't run here any more (set_cpus_allowed()
4454 * away from this CPU, or CPU going down), or because we're
4455 * attempting to rebalance this task on exec (sched_exec).
4457 * So we race with normal scheduler movements, but that's OK, as long
4458 * as the task is no longer on this CPU.
4460 static void __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4462 runqueue_t *rq_dest, *rq_src;
4464 if (unlikely(cpu_is_offline(dest_cpu)))
4467 rq_src = cpu_rq(src_cpu);
4468 rq_dest = cpu_rq(dest_cpu);
4470 double_rq_lock(rq_src, rq_dest);
4471 /* Already moved. */
4472 if (task_cpu(p) != src_cpu)
4474 /* Affinity changed (again). */
4475 if (!cpu_isset(dest_cpu, p->cpus_allowed))
4478 set_task_cpu(p, dest_cpu);
4481 * Sync timestamp with rq_dest's before activating.
4482 * The same thing could be achieved by doing this step
4483 * afterwards, and pretending it was a local activate.
4484 * This way is cleaner and logically correct.
4486 p->timestamp = p->timestamp - rq_src->timestamp_last_tick
4487 + rq_dest->timestamp_last_tick;
4488 deactivate_task(p, rq_src);
4489 activate_task(p, rq_dest, 0);
4490 if (TASK_PREEMPTS_CURR(p, rq_dest))
4491 resched_task(rq_dest->curr);
4495 double_rq_unlock(rq_src, rq_dest);
4499 * migration_thread - this is a highprio system thread that performs
4500 * thread migration by bumping thread off CPU then 'pushing' onto
4503 static int migration_thread(void *data)
4506 int cpu = (long)data;
4509 BUG_ON(rq->migration_thread != current);
4511 set_current_state(TASK_INTERRUPTIBLE);
4512 while (!kthread_should_stop()) {
4513 struct list_head *head;
4514 migration_req_t *req;
4518 spin_lock_irq(&rq->lock);
4520 if (cpu_is_offline(cpu)) {
4521 spin_unlock_irq(&rq->lock);
4525 if (rq->active_balance) {
4526 active_load_balance(rq, cpu);
4527 rq->active_balance = 0;
4530 head = &rq->migration_queue;
4532 if (list_empty(head)) {
4533 spin_unlock_irq(&rq->lock);
4535 set_current_state(TASK_INTERRUPTIBLE);
4538 req = list_entry(head->next, migration_req_t, list);
4539 list_del_init(head->next);
4541 spin_unlock(&rq->lock);
4542 __migrate_task(req->task, cpu, req->dest_cpu);
4545 complete(&req->done);
4547 __set_current_state(TASK_RUNNING);
4551 /* Wait for kthread_stop */
4552 set_current_state(TASK_INTERRUPTIBLE);
4553 while (!kthread_should_stop()) {
4555 set_current_state(TASK_INTERRUPTIBLE);
4557 __set_current_state(TASK_RUNNING);
4561 #ifdef CONFIG_HOTPLUG_CPU
4562 /* Figure out where task on dead CPU should go, use force if neccessary. */
4563 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *tsk)
4569 mask = node_to_cpumask(cpu_to_node(dead_cpu));
4570 cpus_and(mask, mask, tsk->cpus_allowed);
4571 dest_cpu = any_online_cpu(mask);
4573 /* On any allowed CPU? */
4574 if (dest_cpu == NR_CPUS)
4575 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4577 /* No more Mr. Nice Guy. */
4578 if (dest_cpu == NR_CPUS) {
4579 cpus_setall(tsk->cpus_allowed);
4580 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4583 * Don't tell them about moving exiting tasks or
4584 * kernel threads (both mm NULL), since they never
4587 if (tsk->mm && printk_ratelimit())
4588 printk(KERN_INFO "process %d (%s) no "
4589 "longer affine to cpu%d\n",
4590 tsk->pid, tsk->comm, dead_cpu);
4592 __migrate_task(tsk, dead_cpu, dest_cpu);
4596 * While a dead CPU has no uninterruptible tasks queued at this point,
4597 * it might still have a nonzero ->nr_uninterruptible counter, because
4598 * for performance reasons the counter is not stricly tracking tasks to
4599 * their home CPUs. So we just add the counter to another CPU's counter,
4600 * to keep the global sum constant after CPU-down:
4602 static void migrate_nr_uninterruptible(runqueue_t *rq_src)
4604 runqueue_t *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
4605 unsigned long flags;
4607 local_irq_save(flags);
4608 double_rq_lock(rq_src, rq_dest);
4609 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
4610 rq_src->nr_uninterruptible = 0;
4611 double_rq_unlock(rq_src, rq_dest);
4612 local_irq_restore(flags);
4615 /* Run through task list and migrate tasks from the dead cpu. */
4616 static void migrate_live_tasks(int src_cpu)
4618 struct task_struct *tsk, *t;
4620 write_lock_irq(&tasklist_lock);
4622 do_each_thread(t, tsk) {
4626 if (task_cpu(tsk) == src_cpu)
4627 move_task_off_dead_cpu(src_cpu, tsk);
4628 } while_each_thread(t, tsk);
4630 write_unlock_irq(&tasklist_lock);
4633 /* Schedules idle task to be the next runnable task on current CPU.
4634 * It does so by boosting its priority to highest possible and adding it to
4635 * the _front_ of runqueue. Used by CPU offline code.
4637 void sched_idle_next(void)
4639 int cpu = smp_processor_id();
4640 runqueue_t *rq = this_rq();
4641 struct task_struct *p = rq->idle;
4642 unsigned long flags;
4644 /* cpu has to be offline */
4645 BUG_ON(cpu_online(cpu));
4647 /* Strictly not necessary since rest of the CPUs are stopped by now
4648 * and interrupts disabled on current cpu.
4650 spin_lock_irqsave(&rq->lock, flags);
4652 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4653 /* Add idle task to _front_ of it's priority queue */
4654 __activate_idle_task(p, rq);
4656 spin_unlock_irqrestore(&rq->lock, flags);
4659 /* Ensures that the idle task is using init_mm right before its cpu goes
4662 void idle_task_exit(void)
4664 struct mm_struct *mm = current->active_mm;
4666 BUG_ON(cpu_online(smp_processor_id()));
4669 switch_mm(mm, &init_mm, current);
4673 static void migrate_dead(unsigned int dead_cpu, task_t *tsk)
4675 struct runqueue *rq = cpu_rq(dead_cpu);
4677 /* Must be exiting, otherwise would be on tasklist. */
4678 BUG_ON(tsk->exit_state != EXIT_ZOMBIE && tsk->exit_state != EXIT_DEAD);
4680 /* Cannot have done final schedule yet: would have vanished. */
4681 BUG_ON(tsk->flags & PF_DEAD);
4683 get_task_struct(tsk);
4686 * Drop lock around migration; if someone else moves it,
4687 * that's OK. No task can be added to this CPU, so iteration is
4690 spin_unlock_irq(&rq->lock);
4691 move_task_off_dead_cpu(dead_cpu, tsk);
4692 spin_lock_irq(&rq->lock);
4694 put_task_struct(tsk);
4697 /* release_task() removes task from tasklist, so we won't find dead tasks. */
4698 static void migrate_dead_tasks(unsigned int dead_cpu)
4701 struct runqueue *rq = cpu_rq(dead_cpu);
4703 for (arr = 0; arr < 2; arr++) {
4704 for (i = 0; i < MAX_PRIO; i++) {
4705 struct list_head *list = &rq->arrays[arr].queue[i];
4706 while (!list_empty(list))
4707 migrate_dead(dead_cpu,
4708 list_entry(list->next, task_t,
4713 #endif /* CONFIG_HOTPLUG_CPU */
4716 * migration_call - callback that gets triggered when a CPU is added.
4717 * Here we can start up the necessary migration thread for the new CPU.
4719 static int migration_call(struct notifier_block *nfb, unsigned long action,
4722 int cpu = (long)hcpu;
4723 struct task_struct *p;
4724 struct runqueue *rq;
4725 unsigned long flags;
4728 case CPU_UP_PREPARE:
4729 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
4732 p->flags |= PF_NOFREEZE;
4733 kthread_bind(p, cpu);
4734 /* Must be high prio: stop_machine expects to yield to it. */
4735 rq = task_rq_lock(p, &flags);
4736 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4737 task_rq_unlock(rq, &flags);
4738 cpu_rq(cpu)->migration_thread = p;
4741 /* Strictly unneccessary, as first user will wake it. */
4742 wake_up_process(cpu_rq(cpu)->migration_thread);
4744 #ifdef CONFIG_HOTPLUG_CPU
4745 case CPU_UP_CANCELED:
4746 /* Unbind it from offline cpu so it can run. Fall thru. */
4747 kthread_bind(cpu_rq(cpu)->migration_thread,
4748 any_online_cpu(cpu_online_map));
4749 kthread_stop(cpu_rq(cpu)->migration_thread);
4750 cpu_rq(cpu)->migration_thread = NULL;
4753 migrate_live_tasks(cpu);
4755 kthread_stop(rq->migration_thread);
4756 rq->migration_thread = NULL;
4757 /* Idle task back to normal (off runqueue, low prio) */
4758 rq = task_rq_lock(rq->idle, &flags);
4759 deactivate_task(rq->idle, rq);
4760 rq->idle->static_prio = MAX_PRIO;
4761 __setscheduler(rq->idle, SCHED_NORMAL, 0);
4762 migrate_dead_tasks(cpu);
4763 task_rq_unlock(rq, &flags);
4764 migrate_nr_uninterruptible(rq);
4765 BUG_ON(rq->nr_running != 0);
4767 /* No need to migrate the tasks: it was best-effort if
4768 * they didn't do lock_cpu_hotplug(). Just wake up
4769 * the requestors. */
4770 spin_lock_irq(&rq->lock);
4771 while (!list_empty(&rq->migration_queue)) {
4772 migration_req_t *req;
4773 req = list_entry(rq->migration_queue.next,
4774 migration_req_t, list);
4775 list_del_init(&req->list);
4776 complete(&req->done);
4778 spin_unlock_irq(&rq->lock);
4785 /* Register at highest priority so that task migration (migrate_all_tasks)
4786 * happens before everything else.
4788 static struct notifier_block __devinitdata migration_notifier = {
4789 .notifier_call = migration_call,
4793 int __init migration_init(void)
4795 void *cpu = (void *)(long)smp_processor_id();
4796 /* Start one for boot CPU. */
4797 migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
4798 migration_call(&migration_notifier, CPU_ONLINE, cpu);
4799 register_cpu_notifier(&migration_notifier);
4805 #undef SCHED_DOMAIN_DEBUG
4806 #ifdef SCHED_DOMAIN_DEBUG
4807 static void sched_domain_debug(struct sched_domain *sd, int cpu)
4812 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
4816 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
4821 struct sched_group *group = sd->groups;
4822 cpumask_t groupmask;
4824 cpumask_scnprintf(str, NR_CPUS, sd->span);
4825 cpus_clear(groupmask);
4828 for (i = 0; i < level + 1; i++)
4830 printk("domain %d: ", level);
4832 if (!(sd->flags & SD_LOAD_BALANCE)) {
4833 printk("does not load-balance\n");
4835 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent");
4839 printk("span %s\n", str);
4841 if (!cpu_isset(cpu, sd->span))
4842 printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
4843 if (!cpu_isset(cpu, group->cpumask))
4844 printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
4847 for (i = 0; i < level + 2; i++)
4853 printk(KERN_ERR "ERROR: group is NULL\n");
4857 if (!group->cpu_power) {
4859 printk(KERN_ERR "ERROR: domain->cpu_power not set\n");
4862 if (!cpus_weight(group->cpumask)) {
4864 printk(KERN_ERR "ERROR: empty group\n");
4867 if (cpus_intersects(groupmask, group->cpumask)) {
4869 printk(KERN_ERR "ERROR: repeated CPUs\n");
4872 cpus_or(groupmask, groupmask, group->cpumask);
4874 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
4877 group = group->next;
4878 } while (group != sd->groups);
4881 if (!cpus_equal(sd->span, groupmask))
4882 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
4888 if (!cpus_subset(groupmask, sd->span))
4889 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
4895 #define sched_domain_debug(sd, cpu) {}
4898 static int sd_degenerate(struct sched_domain *sd)
4900 if (cpus_weight(sd->span) == 1)
4903 /* Following flags need at least 2 groups */
4904 if (sd->flags & (SD_LOAD_BALANCE |
4905 SD_BALANCE_NEWIDLE |
4908 if (sd->groups != sd->groups->next)
4912 /* Following flags don't use groups */
4913 if (sd->flags & (SD_WAKE_IDLE |
4921 static int sd_parent_degenerate(struct sched_domain *sd,
4922 struct sched_domain *parent)
4924 unsigned long cflags = sd->flags, pflags = parent->flags;
4926 if (sd_degenerate(parent))
4929 if (!cpus_equal(sd->span, parent->span))
4932 /* Does parent contain flags not in child? */
4933 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
4934 if (cflags & SD_WAKE_AFFINE)
4935 pflags &= ~SD_WAKE_BALANCE;
4936 /* Flags needing groups don't count if only 1 group in parent */
4937 if (parent->groups == parent->groups->next) {
4938 pflags &= ~(SD_LOAD_BALANCE |
4939 SD_BALANCE_NEWIDLE |
4943 if (~cflags & pflags)
4950 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
4951 * hold the hotplug lock.
4953 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
4955 runqueue_t *rq = cpu_rq(cpu);
4956 struct sched_domain *tmp;
4958 /* Remove the sched domains which do not contribute to scheduling. */
4959 for (tmp = sd; tmp; tmp = tmp->parent) {
4960 struct sched_domain *parent = tmp->parent;
4963 if (sd_parent_degenerate(tmp, parent))
4964 tmp->parent = parent->parent;
4967 if (sd && sd_degenerate(sd))
4970 sched_domain_debug(sd, cpu);
4972 rcu_assign_pointer(rq->sd, sd);
4975 /* cpus with isolated domains */
4976 static cpumask_t __devinitdata cpu_isolated_map = CPU_MASK_NONE;
4978 /* Setup the mask of cpus configured for isolated domains */
4979 static int __init isolated_cpu_setup(char *str)
4981 int ints[NR_CPUS], i;
4983 str = get_options(str, ARRAY_SIZE(ints), ints);
4984 cpus_clear(cpu_isolated_map);
4985 for (i = 1; i <= ints[0]; i++)
4986 if (ints[i] < NR_CPUS)
4987 cpu_set(ints[i], cpu_isolated_map);
4991 __setup ("isolcpus=", isolated_cpu_setup);
4994 * init_sched_build_groups takes an array of groups, the cpumask we wish
4995 * to span, and a pointer to a function which identifies what group a CPU
4996 * belongs to. The return value of group_fn must be a valid index into the
4997 * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
4998 * keep track of groups covered with a cpumask_t).
5000 * init_sched_build_groups will build a circular linked list of the groups
5001 * covered by the given span, and will set each group's ->cpumask correctly,
5002 * and ->cpu_power to 0.
5004 static void init_sched_build_groups(struct sched_group groups[], cpumask_t span,
5005 int (*group_fn)(int cpu))
5007 struct sched_group *first = NULL, *last = NULL;
5008 cpumask_t covered = CPU_MASK_NONE;
5011 for_each_cpu_mask(i, span) {
5012 int group = group_fn(i);
5013 struct sched_group *sg = &groups[group];
5016 if (cpu_isset(i, covered))
5019 sg->cpumask = CPU_MASK_NONE;
5022 for_each_cpu_mask(j, span) {
5023 if (group_fn(j) != group)
5026 cpu_set(j, covered);
5027 cpu_set(j, sg->cpumask);
5038 #define SD_NODES_PER_DOMAIN 16
5041 * Self-tuning task migration cost measurement between source and target CPUs.
5043 * This is done by measuring the cost of manipulating buffers of varying
5044 * sizes. For a given buffer-size here are the steps that are taken:
5046 * 1) the source CPU reads+dirties a shared buffer
5047 * 2) the target CPU reads+dirties the same shared buffer
5049 * We measure how long they take, in the following 4 scenarios:
5051 * - source: CPU1, target: CPU2 | cost1
5052 * - source: CPU2, target: CPU1 | cost2
5053 * - source: CPU1, target: CPU1 | cost3
5054 * - source: CPU2, target: CPU2 | cost4
5056 * We then calculate the cost3+cost4-cost1-cost2 difference - this is
5057 * the cost of migration.
5059 * We then start off from a small buffer-size and iterate up to larger
5060 * buffer sizes, in 5% steps - measuring each buffer-size separately, and
5061 * doing a maximum search for the cost. (The maximum cost for a migration
5062 * normally occurs when the working set size is around the effective cache
5065 #define SEARCH_SCOPE 2
5066 #define MIN_CACHE_SIZE (64*1024U)
5067 #define DEFAULT_CACHE_SIZE (5*1024*1024U)
5068 #define ITERATIONS 1
5069 #define SIZE_THRESH 130
5070 #define COST_THRESH 130
5073 * The migration cost is a function of 'domain distance'. Domain
5074 * distance is the number of steps a CPU has to iterate down its
5075 * domain tree to share a domain with the other CPU. The farther
5076 * two CPUs are from each other, the larger the distance gets.
5078 * Note that we use the distance only to cache measurement results,
5079 * the distance value is not used numerically otherwise. When two
5080 * CPUs have the same distance it is assumed that the migration
5081 * cost is the same. (this is a simplification but quite practical)
5083 #define MAX_DOMAIN_DISTANCE 32
5085 static unsigned long long migration_cost[MAX_DOMAIN_DISTANCE] =
5086 { [ 0 ... MAX_DOMAIN_DISTANCE-1 ] =
5088 * Architectures may override the migration cost and thus avoid
5089 * boot-time calibration. Unit is nanoseconds. Mostly useful for
5090 * virtualized hardware:
5092 #ifdef CONFIG_DEFAULT_MIGRATION_COST
5093 CONFIG_DEFAULT_MIGRATION_COST
5100 * Allow override of migration cost - in units of microseconds.
5101 * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost
5102 * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs:
5104 static int __init migration_cost_setup(char *str)
5106 int ints[MAX_DOMAIN_DISTANCE+1], i;
5108 str = get_options(str, ARRAY_SIZE(ints), ints);
5110 printk("#ints: %d\n", ints[0]);
5111 for (i = 1; i <= ints[0]; i++) {
5112 migration_cost[i-1] = (unsigned long long)ints[i]*1000;
5113 printk("migration_cost[%d]: %Ld\n", i-1, migration_cost[i-1]);
5118 __setup ("migration_cost=", migration_cost_setup);
5121 * Global multiplier (divisor) for migration-cutoff values,
5122 * in percentiles. E.g. use a value of 150 to get 1.5 times
5123 * longer cache-hot cutoff times.
5125 * (We scale it from 100 to 128 to long long handling easier.)
5128 #define MIGRATION_FACTOR_SCALE 128
5130 static unsigned int migration_factor = MIGRATION_FACTOR_SCALE;
5132 static int __init setup_migration_factor(char *str)
5134 get_option(&str, &migration_factor);
5135 migration_factor = migration_factor * MIGRATION_FACTOR_SCALE / 100;
5139 __setup("migration_factor=", setup_migration_factor);
5142 * Estimated distance of two CPUs, measured via the number of domains
5143 * we have to pass for the two CPUs to be in the same span:
5145 static unsigned long domain_distance(int cpu1, int cpu2)
5147 unsigned long distance = 0;
5148 struct sched_domain *sd;
5150 for_each_domain(cpu1, sd) {
5151 WARN_ON(!cpu_isset(cpu1, sd->span));
5152 if (cpu_isset(cpu2, sd->span))
5156 if (distance >= MAX_DOMAIN_DISTANCE) {
5158 distance = MAX_DOMAIN_DISTANCE-1;
5164 static unsigned int migration_debug;
5166 static int __init setup_migration_debug(char *str)
5168 get_option(&str, &migration_debug);
5172 __setup("migration_debug=", setup_migration_debug);
5175 * Maximum cache-size that the scheduler should try to measure.
5176 * Architectures with larger caches should tune this up during
5177 * bootup. Gets used in the domain-setup code (i.e. during SMP
5180 unsigned int max_cache_size;
5182 static int __init setup_max_cache_size(char *str)
5184 get_option(&str, &max_cache_size);
5188 __setup("max_cache_size=", setup_max_cache_size);
5191 * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This
5192 * is the operation that is timed, so we try to generate unpredictable
5193 * cachemisses that still end up filling the L2 cache:
5195 static void touch_cache(void *__cache, unsigned long __size)
5197 unsigned long size = __size/sizeof(long), chunk1 = size/3,
5199 unsigned long *cache = __cache;
5202 for (i = 0; i < size/6; i += 8) {
5205 case 1: cache[size-1-i]++;
5206 case 2: cache[chunk1-i]++;
5207 case 3: cache[chunk1+i]++;
5208 case 4: cache[chunk2-i]++;
5209 case 5: cache[chunk2+i]++;
5215 * Measure the cache-cost of one task migration. Returns in units of nsec.
5217 static unsigned long long measure_one(void *cache, unsigned long size,
5218 int source, int target)
5220 cpumask_t mask, saved_mask;
5221 unsigned long long t0, t1, t2, t3, cost;
5223 saved_mask = current->cpus_allowed;
5226 * Flush source caches to RAM and invalidate them:
5231 * Migrate to the source CPU:
5233 mask = cpumask_of_cpu(source);
5234 set_cpus_allowed(current, mask);
5235 WARN_ON(smp_processor_id() != source);
5238 * Dirty the working set:
5241 touch_cache(cache, size);
5245 * Migrate to the target CPU, dirty the L2 cache and access
5246 * the shared buffer. (which represents the working set
5247 * of a migrated task.)
5249 mask = cpumask_of_cpu(target);
5250 set_cpus_allowed(current, mask);
5251 WARN_ON(smp_processor_id() != target);
5254 touch_cache(cache, size);
5257 cost = t1-t0 + t3-t2;
5259 if (migration_debug >= 2)
5260 printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n",
5261 source, target, t1-t0, t1-t0, t3-t2, cost);
5263 * Flush target caches to RAM and invalidate them:
5267 set_cpus_allowed(current, saved_mask);
5273 * Measure a series of task migrations and return the average
5274 * result. Since this code runs early during bootup the system
5275 * is 'undisturbed' and the average latency makes sense.
5277 * The algorithm in essence auto-detects the relevant cache-size,
5278 * so it will properly detect different cachesizes for different
5279 * cache-hierarchies, depending on how the CPUs are connected.
5281 * Architectures can prime the upper limit of the search range via
5282 * max_cache_size, otherwise the search range defaults to 20MB...64K.
5284 static unsigned long long
5285 measure_cost(int cpu1, int cpu2, void *cache, unsigned int size)
5287 unsigned long long cost1, cost2;
5291 * Measure the migration cost of 'size' bytes, over an
5292 * average of 10 runs:
5294 * (We perturb the cache size by a small (0..4k)
5295 * value to compensate size/alignment related artifacts.
5296 * We also subtract the cost of the operation done on
5302 * dry run, to make sure we start off cache-cold on cpu1,
5303 * and to get any vmalloc pagefaults in advance:
5305 measure_one(cache, size, cpu1, cpu2);
5306 for (i = 0; i < ITERATIONS; i++)
5307 cost1 += measure_one(cache, size - i*1024, cpu1, cpu2);
5309 measure_one(cache, size, cpu2, cpu1);
5310 for (i = 0; i < ITERATIONS; i++)
5311 cost1 += measure_one(cache, size - i*1024, cpu2, cpu1);
5314 * (We measure the non-migrating [cached] cost on both
5315 * cpu1 and cpu2, to handle CPUs with different speeds)
5319 measure_one(cache, size, cpu1, cpu1);
5320 for (i = 0; i < ITERATIONS; i++)
5321 cost2 += measure_one(cache, size - i*1024, cpu1, cpu1);
5323 measure_one(cache, size, cpu2, cpu2);
5324 for (i = 0; i < ITERATIONS; i++)
5325 cost2 += measure_one(cache, size - i*1024, cpu2, cpu2);
5328 * Get the per-iteration migration cost:
5330 do_div(cost1, 2*ITERATIONS);
5331 do_div(cost2, 2*ITERATIONS);
5333 return cost1 - cost2;
5336 static unsigned long long measure_migration_cost(int cpu1, int cpu2)
5338 unsigned long long max_cost = 0, fluct = 0, avg_fluct = 0;
5339 unsigned int max_size, size, size_found = 0;
5340 long long cost = 0, prev_cost;
5344 * Search from max_cache_size*5 down to 64K - the real relevant
5345 * cachesize has to lie somewhere inbetween.
5347 if (max_cache_size) {
5348 max_size = max(max_cache_size * SEARCH_SCOPE, MIN_CACHE_SIZE);
5349 size = max(max_cache_size / SEARCH_SCOPE, MIN_CACHE_SIZE);
5352 * Since we have no estimation about the relevant
5355 max_size = DEFAULT_CACHE_SIZE * SEARCH_SCOPE;
5356 size = MIN_CACHE_SIZE;
5359 if (!cpu_online(cpu1) || !cpu_online(cpu2)) {
5360 printk("cpu %d and %d not both online!\n", cpu1, cpu2);
5365 * Allocate the working set:
5367 cache = vmalloc(max_size);
5369 printk("could not vmalloc %d bytes for cache!\n", 2*max_size);
5370 return 1000000; // return 1 msec on very small boxen
5373 while (size <= max_size) {
5375 cost = measure_cost(cpu1, cpu2, cache, size);
5381 if (max_cost < cost) {
5387 * Calculate average fluctuation, we use this to prevent
5388 * noise from triggering an early break out of the loop:
5390 fluct = abs(cost - prev_cost);
5391 avg_fluct = (avg_fluct + fluct)/2;
5393 if (migration_debug)
5394 printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): (%8Ld %8Ld)\n",
5396 (long)cost / 1000000,
5397 ((long)cost / 100000) % 10,
5398 (long)max_cost / 1000000,
5399 ((long)max_cost / 100000) % 10,
5400 domain_distance(cpu1, cpu2),
5404 * If we iterated at least 20% past the previous maximum,
5405 * and the cost has dropped by more than 20% already,
5406 * (taking fluctuations into account) then we assume to
5407 * have found the maximum and break out of the loop early:
5409 if (size_found && (size*100 > size_found*SIZE_THRESH))
5410 if (cost+avg_fluct <= 0 ||
5411 max_cost*100 > (cost+avg_fluct)*COST_THRESH) {
5413 if (migration_debug)
5414 printk("-> found max.\n");
5418 * Increase the cachesize in 10% steps:
5420 size = size * 10 / 9;
5423 if (migration_debug)
5424 printk("[%d][%d] working set size found: %d, cost: %Ld\n",
5425 cpu1, cpu2, size_found, max_cost);
5430 * A task is considered 'cache cold' if at least 2 times
5431 * the worst-case cost of migration has passed.
5433 * (this limit is only listened to if the load-balancing
5434 * situation is 'nice' - if there is a large imbalance we
5435 * ignore it for the sake of CPU utilization and
5436 * processing fairness.)
5438 return 2 * max_cost * migration_factor / MIGRATION_FACTOR_SCALE;
5441 static void calibrate_migration_costs(const cpumask_t *cpu_map)
5443 int cpu1 = -1, cpu2 = -1, cpu, orig_cpu = raw_smp_processor_id();
5444 unsigned long j0, j1, distance, max_distance = 0;
5445 struct sched_domain *sd;
5450 * First pass - calculate the cacheflush times:
5452 for_each_cpu_mask(cpu1, *cpu_map) {
5453 for_each_cpu_mask(cpu2, *cpu_map) {
5456 distance = domain_distance(cpu1, cpu2);
5457 max_distance = max(max_distance, distance);
5459 * No result cached yet?
5461 if (migration_cost[distance] == -1LL)
5462 migration_cost[distance] =
5463 measure_migration_cost(cpu1, cpu2);
5467 * Second pass - update the sched domain hierarchy with
5468 * the new cache-hot-time estimations:
5470 for_each_cpu_mask(cpu, *cpu_map) {
5472 for_each_domain(cpu, sd) {
5473 sd->cache_hot_time = migration_cost[distance];
5480 if (migration_debug)
5481 printk("migration: max_cache_size: %d, cpu: %d MHz:\n",
5489 if (system_state == SYSTEM_BOOTING) {
5490 printk("migration_cost=");
5491 for (distance = 0; distance <= max_distance; distance++) {
5494 printk("%ld", (long)migration_cost[distance] / 1000);
5499 if (migration_debug)
5500 printk("migration: %ld seconds\n", (j1-j0)/HZ);
5503 * Move back to the original CPU. NUMA-Q gets confused
5504 * if we migrate to another quad during bootup.
5506 if (raw_smp_processor_id() != orig_cpu) {
5507 cpumask_t mask = cpumask_of_cpu(orig_cpu),
5508 saved_mask = current->cpus_allowed;
5510 set_cpus_allowed(current, mask);
5511 set_cpus_allowed(current, saved_mask);
5518 * find_next_best_node - find the next node to include in a sched_domain
5519 * @node: node whose sched_domain we're building
5520 * @used_nodes: nodes already in the sched_domain
5522 * Find the next node to include in a given scheduling domain. Simply
5523 * finds the closest node not already in the @used_nodes map.
5525 * Should use nodemask_t.
5527 static int find_next_best_node(int node, unsigned long *used_nodes)
5529 int i, n, val, min_val, best_node = 0;
5533 for (i = 0; i < MAX_NUMNODES; i++) {
5534 /* Start at @node */
5535 n = (node + i) % MAX_NUMNODES;
5537 if (!nr_cpus_node(n))
5540 /* Skip already used nodes */
5541 if (test_bit(n, used_nodes))
5544 /* Simple min distance search */
5545 val = node_distance(node, n);
5547 if (val < min_val) {
5553 set_bit(best_node, used_nodes);
5558 * sched_domain_node_span - get a cpumask for a node's sched_domain
5559 * @node: node whose cpumask we're constructing
5560 * @size: number of nodes to include in this span
5562 * Given a node, construct a good cpumask for its sched_domain to span. It
5563 * should be one that prevents unnecessary balancing, but also spreads tasks
5566 static cpumask_t sched_domain_node_span(int node)
5569 cpumask_t span, nodemask;
5570 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5573 bitmap_zero(used_nodes, MAX_NUMNODES);
5575 nodemask = node_to_cpumask(node);
5576 cpus_or(span, span, nodemask);
5577 set_bit(node, used_nodes);
5579 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5580 int next_node = find_next_best_node(node, used_nodes);
5581 nodemask = node_to_cpumask(next_node);
5582 cpus_or(span, span, nodemask);
5590 * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we
5591 * can switch it on easily if needed.
5593 #ifdef CONFIG_SCHED_SMT
5594 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5595 static struct sched_group sched_group_cpus[NR_CPUS];
5596 static int cpu_to_cpu_group(int cpu)
5602 #ifdef CONFIG_SCHED_MC
5603 static DEFINE_PER_CPU(struct sched_domain, core_domains);
5604 static struct sched_group sched_group_core[NR_CPUS];
5607 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5608 static int cpu_to_core_group(int cpu)
5610 return first_cpu(cpu_sibling_map[cpu]);
5612 #elif defined(CONFIG_SCHED_MC)
5613 static int cpu_to_core_group(int cpu)
5619 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5620 static struct sched_group sched_group_phys[NR_CPUS];
5621 static int cpu_to_phys_group(int cpu)
5623 #if defined(CONFIG_SCHED_MC)
5624 cpumask_t mask = cpu_coregroup_map(cpu);
5625 return first_cpu(mask);
5626 #elif defined(CONFIG_SCHED_SMT)
5627 return first_cpu(cpu_sibling_map[cpu]);
5635 * The init_sched_build_groups can't handle what we want to do with node
5636 * groups, so roll our own. Now each node has its own list of groups which
5637 * gets dynamically allocated.
5639 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5640 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5642 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5643 static struct sched_group *sched_group_allnodes_bycpu[NR_CPUS];
5645 static int cpu_to_allnodes_group(int cpu)
5647 return cpu_to_node(cpu);
5649 static void init_numa_sched_groups_power(struct sched_group *group_head)
5651 struct sched_group *sg = group_head;
5657 for_each_cpu_mask(j, sg->cpumask) {
5658 struct sched_domain *sd;
5660 sd = &per_cpu(phys_domains, j);
5661 if (j != first_cpu(sd->groups->cpumask)) {
5663 * Only add "power" once for each
5669 sg->cpu_power += sd->groups->cpu_power;
5672 if (sg != group_head)
5678 * Build sched domains for a given set of cpus and attach the sched domains
5679 * to the individual cpus
5681 void build_sched_domains(const cpumask_t *cpu_map)
5685 struct sched_group **sched_group_nodes = NULL;
5686 struct sched_group *sched_group_allnodes = NULL;
5689 * Allocate the per-node list of sched groups
5691 sched_group_nodes = kmalloc(sizeof(struct sched_group*)*MAX_NUMNODES,
5693 if (!sched_group_nodes) {
5694 printk(KERN_WARNING "Can not alloc sched group node list\n");
5697 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
5701 * Set up domains for cpus specified by the cpu_map.
5703 for_each_cpu_mask(i, *cpu_map) {
5705 struct sched_domain *sd = NULL, *p;
5706 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
5708 cpus_and(nodemask, nodemask, *cpu_map);
5711 if (cpus_weight(*cpu_map)
5712 > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
5713 if (!sched_group_allnodes) {
5714 sched_group_allnodes
5715 = kmalloc(sizeof(struct sched_group)
5718 if (!sched_group_allnodes) {
5720 "Can not alloc allnodes sched group\n");
5723 sched_group_allnodes_bycpu[i]
5724 = sched_group_allnodes;
5726 sd = &per_cpu(allnodes_domains, i);
5727 *sd = SD_ALLNODES_INIT;
5728 sd->span = *cpu_map;
5729 group = cpu_to_allnodes_group(i);
5730 sd->groups = &sched_group_allnodes[group];
5735 sd = &per_cpu(node_domains, i);
5737 sd->span = sched_domain_node_span(cpu_to_node(i));
5739 cpus_and(sd->span, sd->span, *cpu_map);
5743 sd = &per_cpu(phys_domains, i);
5744 group = cpu_to_phys_group(i);
5746 sd->span = nodemask;
5748 sd->groups = &sched_group_phys[group];
5750 #ifdef CONFIG_SCHED_MC
5752 sd = &per_cpu(core_domains, i);
5753 group = cpu_to_core_group(i);
5755 sd->span = cpu_coregroup_map(i);
5756 cpus_and(sd->span, sd->span, *cpu_map);
5758 sd->groups = &sched_group_core[group];
5761 #ifdef CONFIG_SCHED_SMT
5763 sd = &per_cpu(cpu_domains, i);
5764 group = cpu_to_cpu_group(i);
5765 *sd = SD_SIBLING_INIT;
5766 sd->span = cpu_sibling_map[i];
5767 cpus_and(sd->span, sd->span, *cpu_map);
5769 sd->groups = &sched_group_cpus[group];
5773 #ifdef CONFIG_SCHED_SMT
5774 /* Set up CPU (sibling) groups */
5775 for_each_cpu_mask(i, *cpu_map) {
5776 cpumask_t this_sibling_map = cpu_sibling_map[i];
5777 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
5778 if (i != first_cpu(this_sibling_map))
5781 init_sched_build_groups(sched_group_cpus, this_sibling_map,
5786 #ifdef CONFIG_SCHED_MC
5787 /* Set up multi-core groups */
5788 for_each_cpu_mask(i, *cpu_map) {
5789 cpumask_t this_core_map = cpu_coregroup_map(i);
5790 cpus_and(this_core_map, this_core_map, *cpu_map);
5791 if (i != first_cpu(this_core_map))
5793 init_sched_build_groups(sched_group_core, this_core_map,
5794 &cpu_to_core_group);
5799 /* Set up physical groups */
5800 for (i = 0; i < MAX_NUMNODES; i++) {
5801 cpumask_t nodemask = node_to_cpumask(i);
5803 cpus_and(nodemask, nodemask, *cpu_map);
5804 if (cpus_empty(nodemask))
5807 init_sched_build_groups(sched_group_phys, nodemask,
5808 &cpu_to_phys_group);
5812 /* Set up node groups */
5813 if (sched_group_allnodes)
5814 init_sched_build_groups(sched_group_allnodes, *cpu_map,
5815 &cpu_to_allnodes_group);
5817 for (i = 0; i < MAX_NUMNODES; i++) {
5818 /* Set up node groups */
5819 struct sched_group *sg, *prev;
5820 cpumask_t nodemask = node_to_cpumask(i);
5821 cpumask_t domainspan;
5822 cpumask_t covered = CPU_MASK_NONE;
5825 cpus_and(nodemask, nodemask, *cpu_map);
5826 if (cpus_empty(nodemask)) {
5827 sched_group_nodes[i] = NULL;
5831 domainspan = sched_domain_node_span(i);
5832 cpus_and(domainspan, domainspan, *cpu_map);
5834 sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL);
5835 sched_group_nodes[i] = sg;
5836 for_each_cpu_mask(j, nodemask) {
5837 struct sched_domain *sd;
5838 sd = &per_cpu(node_domains, j);
5840 if (sd->groups == NULL) {
5841 /* Turn off balancing if we have no groups */
5847 "Can not alloc domain group for node %d\n", i);
5851 sg->cpumask = nodemask;
5852 cpus_or(covered, covered, nodemask);
5855 for (j = 0; j < MAX_NUMNODES; j++) {
5856 cpumask_t tmp, notcovered;
5857 int n = (i + j) % MAX_NUMNODES;
5859 cpus_complement(notcovered, covered);
5860 cpus_and(tmp, notcovered, *cpu_map);
5861 cpus_and(tmp, tmp, domainspan);
5862 if (cpus_empty(tmp))
5865 nodemask = node_to_cpumask(n);
5866 cpus_and(tmp, tmp, nodemask);
5867 if (cpus_empty(tmp))
5870 sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL);
5873 "Can not alloc domain group for node %d\n", j);
5878 cpus_or(covered, covered, tmp);
5882 prev->next = sched_group_nodes[i];
5886 /* Calculate CPU power for physical packages and nodes */
5887 for_each_cpu_mask(i, *cpu_map) {
5889 struct sched_domain *sd;
5890 #ifdef CONFIG_SCHED_SMT
5891 sd = &per_cpu(cpu_domains, i);
5892 power = SCHED_LOAD_SCALE;
5893 sd->groups->cpu_power = power;
5895 #ifdef CONFIG_SCHED_MC
5896 sd = &per_cpu(core_domains, i);
5897 power = SCHED_LOAD_SCALE + (cpus_weight(sd->groups->cpumask)-1)
5898 * SCHED_LOAD_SCALE / 10;
5899 sd->groups->cpu_power = power;
5901 sd = &per_cpu(phys_domains, i);
5904 * This has to be < 2 * SCHED_LOAD_SCALE
5905 * Lets keep it SCHED_LOAD_SCALE, so that
5906 * while calculating NUMA group's cpu_power
5908 * numa_group->cpu_power += phys_group->cpu_power;
5910 * See "only add power once for each physical pkg"
5913 sd->groups->cpu_power = SCHED_LOAD_SCALE;
5915 sd = &per_cpu(phys_domains, i);
5916 power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
5917 (cpus_weight(sd->groups->cpumask)-1) / 10;
5918 sd->groups->cpu_power = power;
5923 for (i = 0; i < MAX_NUMNODES; i++)
5924 init_numa_sched_groups_power(sched_group_nodes[i]);
5926 init_numa_sched_groups_power(sched_group_allnodes);
5929 /* Attach the domains */
5930 for_each_cpu_mask(i, *cpu_map) {
5931 struct sched_domain *sd;
5932 #ifdef CONFIG_SCHED_SMT
5933 sd = &per_cpu(cpu_domains, i);
5934 #elif defined(CONFIG_SCHED_MC)
5935 sd = &per_cpu(core_domains, i);
5937 sd = &per_cpu(phys_domains, i);
5939 cpu_attach_domain(sd, i);
5942 * Tune cache-hot values:
5944 calibrate_migration_costs(cpu_map);
5947 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
5949 static void arch_init_sched_domains(const cpumask_t *cpu_map)
5951 cpumask_t cpu_default_map;
5954 * Setup mask for cpus without special case scheduling requirements.
5955 * For now this just excludes isolated cpus, but could be used to
5956 * exclude other special cases in the future.
5958 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
5960 build_sched_domains(&cpu_default_map);
5963 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
5969 for_each_cpu_mask(cpu, *cpu_map) {
5970 struct sched_group *sched_group_allnodes
5971 = sched_group_allnodes_bycpu[cpu];
5972 struct sched_group **sched_group_nodes
5973 = sched_group_nodes_bycpu[cpu];
5975 if (sched_group_allnodes) {
5976 kfree(sched_group_allnodes);
5977 sched_group_allnodes_bycpu[cpu] = NULL;
5980 if (!sched_group_nodes)
5983 for (i = 0; i < MAX_NUMNODES; i++) {
5984 cpumask_t nodemask = node_to_cpumask(i);
5985 struct sched_group *oldsg, *sg = sched_group_nodes[i];
5987 cpus_and(nodemask, nodemask, *cpu_map);
5988 if (cpus_empty(nodemask))
5998 if (oldsg != sched_group_nodes[i])
6001 kfree(sched_group_nodes);
6002 sched_group_nodes_bycpu[cpu] = NULL;
6008 * Detach sched domains from a group of cpus specified in cpu_map
6009 * These cpus will now be attached to the NULL domain
6011 static void detach_destroy_domains(const cpumask_t *cpu_map)
6015 for_each_cpu_mask(i, *cpu_map)
6016 cpu_attach_domain(NULL, i);
6017 synchronize_sched();
6018 arch_destroy_sched_domains(cpu_map);
6022 * Partition sched domains as specified by the cpumasks below.
6023 * This attaches all cpus from the cpumasks to the NULL domain,
6024 * waits for a RCU quiescent period, recalculates sched
6025 * domain information and then attaches them back to the
6026 * correct sched domains
6027 * Call with hotplug lock held
6029 void partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6031 cpumask_t change_map;
6033 cpus_and(*partition1, *partition1, cpu_online_map);
6034 cpus_and(*partition2, *partition2, cpu_online_map);
6035 cpus_or(change_map, *partition1, *partition2);
6037 /* Detach sched domains from all of the affected cpus */
6038 detach_destroy_domains(&change_map);
6039 if (!cpus_empty(*partition1))
6040 build_sched_domains(partition1);
6041 if (!cpus_empty(*partition2))
6042 build_sched_domains(partition2);
6045 #ifdef CONFIG_HOTPLUG_CPU
6047 * Force a reinitialization of the sched domains hierarchy. The domains
6048 * and groups cannot be updated in place without racing with the balancing
6049 * code, so we temporarily attach all running cpus to the NULL domain
6050 * which will prevent rebalancing while the sched domains are recalculated.
6052 static int update_sched_domains(struct notifier_block *nfb,
6053 unsigned long action, void *hcpu)
6056 case CPU_UP_PREPARE:
6057 case CPU_DOWN_PREPARE:
6058 detach_destroy_domains(&cpu_online_map);
6061 case CPU_UP_CANCELED:
6062 case CPU_DOWN_FAILED:
6066 * Fall through and re-initialise the domains.
6073 /* The hotplug lock is already held by cpu_up/cpu_down */
6074 arch_init_sched_domains(&cpu_online_map);
6080 void __init sched_init_smp(void)
6083 arch_init_sched_domains(&cpu_online_map);
6084 unlock_cpu_hotplug();
6085 /* XXX: Theoretical race here - CPU may be hotplugged now */
6086 hotcpu_notifier(update_sched_domains, 0);
6089 void __init sched_init_smp(void)
6092 #endif /* CONFIG_SMP */
6094 int in_sched_functions(unsigned long addr)
6096 /* Linker adds these: start and end of __sched functions */
6097 extern char __sched_text_start[], __sched_text_end[];
6098 return in_lock_functions(addr) ||
6099 (addr >= (unsigned long)__sched_text_start
6100 && addr < (unsigned long)__sched_text_end);
6103 void __init sched_init(void)
6108 for_each_possible_cpu(i) {
6109 prio_array_t *array;
6112 spin_lock_init(&rq->lock);
6114 rq->active = rq->arrays;
6115 rq->expired = rq->arrays + 1;
6116 rq->best_expired_prio = MAX_PRIO;
6120 for (j = 1; j < 3; j++)
6121 rq->cpu_load[j] = 0;
6122 rq->active_balance = 0;
6124 rq->migration_thread = NULL;
6125 INIT_LIST_HEAD(&rq->migration_queue);
6128 atomic_set(&rq->nr_iowait, 0);
6130 for (j = 0; j < 2; j++) {
6131 array = rq->arrays + j;
6132 for (k = 0; k < MAX_PRIO; k++) {
6133 INIT_LIST_HEAD(array->queue + k);
6134 __clear_bit(k, array->bitmap);
6136 // delimiter for bitsearch
6137 __set_bit(MAX_PRIO, array->bitmap);
6142 * The boot idle thread does lazy MMU switching as well:
6144 atomic_inc(&init_mm.mm_count);
6145 enter_lazy_tlb(&init_mm, current);
6148 * Make us the idle thread. Technically, schedule() should not be
6149 * called from this thread, however somewhere below it might be,
6150 * but because we are the idle thread, we just pick up running again
6151 * when this runqueue becomes "idle".
6153 init_idle(current, smp_processor_id());
6156 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6157 void __might_sleep(char *file, int line)
6159 #if defined(in_atomic)
6160 static unsigned long prev_jiffy; /* ratelimiting */
6162 if ((in_atomic() || irqs_disabled()) &&
6163 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6164 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6166 prev_jiffy = jiffies;
6167 printk(KERN_ERR "BUG: sleeping function called from invalid"
6168 " context at %s:%d\n", file, line);
6169 printk("in_atomic():%d, irqs_disabled():%d\n",
6170 in_atomic(), irqs_disabled());
6175 EXPORT_SYMBOL(__might_sleep);
6178 #ifdef CONFIG_MAGIC_SYSRQ
6179 void normalize_rt_tasks(void)
6181 struct task_struct *p;
6182 prio_array_t *array;
6183 unsigned long flags;
6186 read_lock_irq(&tasklist_lock);
6187 for_each_process (p) {
6191 rq = task_rq_lock(p, &flags);
6195 deactivate_task(p, task_rq(p));
6196 __setscheduler(p, SCHED_NORMAL, 0);
6198 __activate_task(p, task_rq(p));
6199 resched_task(rq->curr);
6202 task_rq_unlock(rq, &flags);
6204 read_unlock_irq(&tasklist_lock);
6207 #endif /* CONFIG_MAGIC_SYSRQ */
6211 * These functions are only useful for the IA64 MCA handling.
6213 * They can only be called when the whole system has been
6214 * stopped - every CPU needs to be quiescent, and no scheduling
6215 * activity can take place. Using them for anything else would
6216 * be a serious bug, and as a result, they aren't even visible
6217 * under any other configuration.
6221 * curr_task - return the current task for a given cpu.
6222 * @cpu: the processor in question.
6224 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6226 task_t *curr_task(int cpu)
6228 return cpu_curr(cpu);
6232 * set_curr_task - set the current task for a given cpu.
6233 * @cpu: the processor in question.
6234 * @p: the task pointer to set.
6236 * Description: This function must only be used when non-maskable interrupts
6237 * are serviced on a separate stack. It allows the architecture to switch the
6238 * notion of the current task on a cpu in a non-blocking manner. This function
6239 * must be called with all CPU's synchronized, and interrupts disabled, the
6240 * and caller must save the original value of the current task (see
6241 * curr_task() above) and restore that value before reenabling interrupts and
6242 * re-starting the system.
6244 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6246 void set_curr_task(int cpu, task_t *p)