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 typedef struct runqueue runqueue_t;
190 unsigned int nr_active;
191 DECLARE_BITMAP(bitmap, MAX_PRIO+1); /* include 1 bit for delimiter */
192 struct list_head queue[MAX_PRIO];
196 * This is the main, per-CPU runqueue data structure.
198 * Locking rule: those places that want to lock multiple runqueues
199 * (such as the load balancing or the thread migration code), lock
200 * acquire operations must be ordered by ascending &runqueue.
206 * nr_running and cpu_load should be in the same cacheline because
207 * remote CPUs use both these fields when doing load calculation.
209 unsigned long nr_running;
211 unsigned long cpu_load[3];
213 unsigned long long nr_switches;
216 * This is part of a global counter where only the total sum
217 * over all CPUs matters. A task can increase this counter on
218 * one CPU and if it got migrated afterwards it may decrease
219 * it on another CPU. Always updated under the runqueue lock:
221 unsigned long nr_uninterruptible;
223 unsigned long expired_timestamp;
224 unsigned long long timestamp_last_tick;
226 struct mm_struct *prev_mm;
227 prio_array_t *active, *expired, arrays[2];
228 int best_expired_prio;
232 struct sched_domain *sd;
234 /* For active balancing */
238 task_t *migration_thread;
239 struct list_head migration_queue;
242 #ifdef CONFIG_SCHEDSTATS
244 struct sched_info rq_sched_info;
246 /* sys_sched_yield() stats */
247 unsigned long yld_exp_empty;
248 unsigned long yld_act_empty;
249 unsigned long yld_both_empty;
250 unsigned long yld_cnt;
252 /* schedule() stats */
253 unsigned long sched_switch;
254 unsigned long sched_cnt;
255 unsigned long sched_goidle;
257 /* try_to_wake_up() stats */
258 unsigned long ttwu_cnt;
259 unsigned long ttwu_local;
263 static DEFINE_PER_CPU(struct runqueue, runqueues);
266 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
267 * See detach_destroy_domains: synchronize_sched for details.
269 * The domain tree of any CPU may only be accessed from within
270 * preempt-disabled sections.
272 #define for_each_domain(cpu, domain) \
273 for (domain = rcu_dereference(cpu_rq(cpu)->sd); domain; domain = domain->parent)
275 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
276 #define this_rq() (&__get_cpu_var(runqueues))
277 #define task_rq(p) cpu_rq(task_cpu(p))
278 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
280 #ifndef prepare_arch_switch
281 # define prepare_arch_switch(next) do { } while (0)
283 #ifndef finish_arch_switch
284 # define finish_arch_switch(prev) do { } while (0)
287 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
288 static inline int task_running(runqueue_t *rq, task_t *p)
290 return rq->curr == p;
293 static inline void prepare_lock_switch(runqueue_t *rq, task_t *next)
297 static inline void finish_lock_switch(runqueue_t *rq, task_t *prev)
299 #ifdef CONFIG_DEBUG_SPINLOCK
300 /* this is a valid case when another task releases the spinlock */
301 rq->lock.owner = current;
303 spin_unlock_irq(&rq->lock);
306 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
307 static inline int task_running(runqueue_t *rq, task_t *p)
312 return rq->curr == p;
316 static inline void prepare_lock_switch(runqueue_t *rq, task_t *next)
320 * We can optimise this out completely for !SMP, because the
321 * SMP rebalancing from interrupt is the only thing that cares
326 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
327 spin_unlock_irq(&rq->lock);
329 spin_unlock(&rq->lock);
333 static inline void finish_lock_switch(runqueue_t *rq, task_t *prev)
337 * After ->oncpu is cleared, the task can be moved to a different CPU.
338 * We must ensure this doesn't happen until the switch is completely
344 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
348 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
351 * task_rq_lock - lock the runqueue a given task resides on and disable
352 * interrupts. Note the ordering: we can safely lookup the task_rq without
353 * explicitly disabling preemption.
355 static inline runqueue_t *task_rq_lock(task_t *p, unsigned long *flags)
361 local_irq_save(*flags);
363 spin_lock(&rq->lock);
364 if (unlikely(rq != task_rq(p))) {
365 spin_unlock_irqrestore(&rq->lock, *flags);
366 goto repeat_lock_task;
371 static inline void task_rq_unlock(runqueue_t *rq, unsigned long *flags)
374 spin_unlock_irqrestore(&rq->lock, *flags);
377 #ifdef CONFIG_SCHEDSTATS
379 * bump this up when changing the output format or the meaning of an existing
380 * format, so that tools can adapt (or abort)
382 #define SCHEDSTAT_VERSION 12
384 static int show_schedstat(struct seq_file *seq, void *v)
388 seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION);
389 seq_printf(seq, "timestamp %lu\n", jiffies);
390 for_each_online_cpu(cpu) {
391 runqueue_t *rq = cpu_rq(cpu);
393 struct sched_domain *sd;
397 /* runqueue-specific stats */
399 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
400 cpu, rq->yld_both_empty,
401 rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt,
402 rq->sched_switch, rq->sched_cnt, rq->sched_goidle,
403 rq->ttwu_cnt, rq->ttwu_local,
404 rq->rq_sched_info.cpu_time,
405 rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt);
407 seq_printf(seq, "\n");
410 /* domain-specific stats */
412 for_each_domain(cpu, sd) {
413 enum idle_type itype;
414 char mask_str[NR_CPUS];
416 cpumask_scnprintf(mask_str, NR_CPUS, sd->span);
417 seq_printf(seq, "domain%d %s", dcnt++, mask_str);
418 for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES;
420 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu",
422 sd->lb_balanced[itype],
423 sd->lb_failed[itype],
424 sd->lb_imbalance[itype],
425 sd->lb_gained[itype],
426 sd->lb_hot_gained[itype],
427 sd->lb_nobusyq[itype],
428 sd->lb_nobusyg[itype]);
430 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
431 sd->alb_cnt, sd->alb_failed, sd->alb_pushed,
432 sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed,
433 sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed,
434 sd->ttwu_wake_remote, sd->ttwu_move_affine, sd->ttwu_move_balance);
442 static int schedstat_open(struct inode *inode, struct file *file)
444 unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32);
445 char *buf = kmalloc(size, GFP_KERNEL);
451 res = single_open(file, show_schedstat, NULL);
453 m = file->private_data;
461 struct file_operations proc_schedstat_operations = {
462 .open = schedstat_open,
465 .release = single_release,
468 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
469 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
470 #else /* !CONFIG_SCHEDSTATS */
471 # define schedstat_inc(rq, field) do { } while (0)
472 # define schedstat_add(rq, field, amt) do { } while (0)
476 * rq_lock - lock a given runqueue and disable interrupts.
478 static inline runqueue_t *this_rq_lock(void)
485 spin_lock(&rq->lock);
490 #ifdef CONFIG_SCHEDSTATS
492 * Called when a process is dequeued from the active array and given
493 * the cpu. We should note that with the exception of interactive
494 * tasks, the expired queue will become the active queue after the active
495 * queue is empty, without explicitly dequeuing and requeuing tasks in the
496 * expired queue. (Interactive tasks may be requeued directly to the
497 * active queue, thus delaying tasks in the expired queue from running;
498 * see scheduler_tick()).
500 * This function is only called from sched_info_arrive(), rather than
501 * dequeue_task(). Even though a task may be queued and dequeued multiple
502 * times as it is shuffled about, we're really interested in knowing how
503 * long it was from the *first* time it was queued to the time that it
506 static inline void sched_info_dequeued(task_t *t)
508 t->sched_info.last_queued = 0;
512 * Called when a task finally hits the cpu. We can now calculate how
513 * long it was waiting to run. We also note when it began so that we
514 * can keep stats on how long its timeslice is.
516 static void sched_info_arrive(task_t *t)
518 unsigned long now = jiffies, diff = 0;
519 struct runqueue *rq = task_rq(t);
521 if (t->sched_info.last_queued)
522 diff = now - t->sched_info.last_queued;
523 sched_info_dequeued(t);
524 t->sched_info.run_delay += diff;
525 t->sched_info.last_arrival = now;
526 t->sched_info.pcnt++;
531 rq->rq_sched_info.run_delay += diff;
532 rq->rq_sched_info.pcnt++;
536 * Called when a process is queued into either the active or expired
537 * array. The time is noted and later used to determine how long we
538 * had to wait for us to reach the cpu. Since the expired queue will
539 * become the active queue after active queue is empty, without dequeuing
540 * and requeuing any tasks, we are interested in queuing to either. It
541 * is unusual but not impossible for tasks to be dequeued and immediately
542 * requeued in the same or another array: this can happen in sched_yield(),
543 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
546 * This function is only called from enqueue_task(), but also only updates
547 * the timestamp if it is already not set. It's assumed that
548 * sched_info_dequeued() will clear that stamp when appropriate.
550 static inline void sched_info_queued(task_t *t)
552 if (!t->sched_info.last_queued)
553 t->sched_info.last_queued = jiffies;
557 * Called when a process ceases being the active-running process, either
558 * voluntarily or involuntarily. Now we can calculate how long we ran.
560 static inline void sched_info_depart(task_t *t)
562 struct runqueue *rq = task_rq(t);
563 unsigned long diff = jiffies - t->sched_info.last_arrival;
565 t->sched_info.cpu_time += diff;
568 rq->rq_sched_info.cpu_time += diff;
572 * Called when tasks are switched involuntarily due, typically, to expiring
573 * their time slice. (This may also be called when switching to or from
574 * the idle task.) We are only called when prev != next.
576 static inline void sched_info_switch(task_t *prev, task_t *next)
578 struct runqueue *rq = task_rq(prev);
581 * prev now departs the cpu. It's not interesting to record
582 * stats about how efficient we were at scheduling the idle
585 if (prev != rq->idle)
586 sched_info_depart(prev);
588 if (next != rq->idle)
589 sched_info_arrive(next);
592 #define sched_info_queued(t) do { } while (0)
593 #define sched_info_switch(t, next) do { } while (0)
594 #endif /* CONFIG_SCHEDSTATS */
597 * Adding/removing a task to/from a priority array:
599 static void dequeue_task(struct task_struct *p, prio_array_t *array)
602 list_del(&p->run_list);
603 if (list_empty(array->queue + p->prio))
604 __clear_bit(p->prio, array->bitmap);
607 static void enqueue_task(struct task_struct *p, prio_array_t *array)
609 sched_info_queued(p);
610 list_add_tail(&p->run_list, array->queue + p->prio);
611 __set_bit(p->prio, array->bitmap);
617 * Put task to the end of the run list without the overhead of dequeue
618 * followed by enqueue.
620 static void requeue_task(struct task_struct *p, prio_array_t *array)
622 list_move_tail(&p->run_list, array->queue + p->prio);
625 static inline void enqueue_task_head(struct task_struct *p, prio_array_t *array)
627 list_add(&p->run_list, array->queue + p->prio);
628 __set_bit(p->prio, array->bitmap);
634 * effective_prio - return the priority that is based on the static
635 * priority but is modified by bonuses/penalties.
637 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
638 * into the -5 ... 0 ... +5 bonus/penalty range.
640 * We use 25% of the full 0...39 priority range so that:
642 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
643 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
645 * Both properties are important to certain workloads.
647 static int effective_prio(task_t *p)
654 bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
656 prio = p->static_prio - bonus;
657 if (prio < MAX_RT_PRIO)
659 if (prio > MAX_PRIO-1)
665 * __activate_task - move a task to the runqueue.
667 static void __activate_task(task_t *p, runqueue_t *rq)
669 prio_array_t *target = rq->active;
672 target = rq->expired;
673 enqueue_task(p, target);
678 * __activate_idle_task - move idle task to the _front_ of runqueue.
680 static inline void __activate_idle_task(task_t *p, runqueue_t *rq)
682 enqueue_task_head(p, rq->active);
686 static int recalc_task_prio(task_t *p, unsigned long long now)
688 /* Caller must always ensure 'now >= p->timestamp' */
689 unsigned long sleep_time = now - p->timestamp;
694 if (likely(sleep_time > 0)) {
696 * This ceiling is set to the lowest priority that would allow
697 * a task to be reinserted into the active array on timeslice
700 unsigned long ceiling = INTERACTIVE_SLEEP(p);
702 if (p->mm && sleep_time > ceiling && p->sleep_avg < ceiling) {
704 * Prevents user tasks from achieving best priority
705 * with one single large enough sleep.
707 p->sleep_avg = ceiling;
709 * Using INTERACTIVE_SLEEP() as a ceiling places a
710 * nice(0) task 1ms sleep away from promotion, and
711 * gives it 700ms to round-robin with no chance of
712 * being demoted. This is more than generous, so
713 * mark this sleep as non-interactive to prevent the
714 * on-runqueue bonus logic from intervening should
715 * this task not receive cpu immediately.
717 p->sleep_type = SLEEP_NONINTERACTIVE;
720 * Tasks waking from uninterruptible sleep are
721 * limited in their sleep_avg rise as they
722 * are likely to be waiting on I/O
724 if (p->sleep_type == SLEEP_NONINTERACTIVE && p->mm) {
725 if (p->sleep_avg >= ceiling)
727 else if (p->sleep_avg + sleep_time >=
729 p->sleep_avg = ceiling;
735 * This code gives a bonus to interactive tasks.
737 * The boost works by updating the 'average sleep time'
738 * value here, based on ->timestamp. The more time a
739 * task spends sleeping, the higher the average gets -
740 * and the higher the priority boost gets as well.
742 p->sleep_avg += sleep_time;
745 if (p->sleep_avg > NS_MAX_SLEEP_AVG)
746 p->sleep_avg = NS_MAX_SLEEP_AVG;
749 return effective_prio(p);
753 * activate_task - move a task to the runqueue and do priority recalculation
755 * Update all the scheduling statistics stuff. (sleep average
756 * calculation, priority modifiers, etc.)
758 static void activate_task(task_t *p, runqueue_t *rq, int local)
760 unsigned long long now;
765 /* Compensate for drifting sched_clock */
766 runqueue_t *this_rq = this_rq();
767 now = (now - this_rq->timestamp_last_tick)
768 + rq->timestamp_last_tick;
773 p->prio = recalc_task_prio(p, now);
776 * This checks to make sure it's not an uninterruptible task
777 * that is now waking up.
779 if (p->sleep_type == SLEEP_NORMAL) {
781 * Tasks which were woken up by interrupts (ie. hw events)
782 * are most likely of interactive nature. So we give them
783 * the credit of extending their sleep time to the period
784 * of time they spend on the runqueue, waiting for execution
785 * on a CPU, first time around:
788 p->sleep_type = SLEEP_INTERRUPTED;
791 * Normal first-time wakeups get a credit too for
792 * on-runqueue time, but it will be weighted down:
794 p->sleep_type = SLEEP_INTERACTIVE;
799 __activate_task(p, rq);
803 * deactivate_task - remove a task from the runqueue.
805 static void deactivate_task(struct task_struct *p, runqueue_t *rq)
808 dequeue_task(p, p->array);
813 * resched_task - mark a task 'to be rescheduled now'.
815 * On UP this means the setting of the need_resched flag, on SMP it
816 * might also involve a cross-CPU call to trigger the scheduler on
821 #ifndef tsk_is_polling
822 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
825 static void resched_task(task_t *p)
829 assert_spin_locked(&task_rq(p)->lock);
831 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
834 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
837 if (cpu == smp_processor_id())
840 /* NEED_RESCHED must be visible before we test polling */
842 if (!tsk_is_polling(p))
843 smp_send_reschedule(cpu);
846 static inline void resched_task(task_t *p)
848 assert_spin_locked(&task_rq(p)->lock);
849 set_tsk_need_resched(p);
854 * task_curr - is this task currently executing on a CPU?
855 * @p: the task in question.
857 inline int task_curr(const task_t *p)
859 return cpu_curr(task_cpu(p)) == p;
864 struct list_head list;
869 struct completion done;
873 * The task's runqueue lock must be held.
874 * Returns true if you have to wait for migration thread.
876 static int migrate_task(task_t *p, int dest_cpu, migration_req_t *req)
878 runqueue_t *rq = task_rq(p);
881 * If the task is not on a runqueue (and not running), then
882 * it is sufficient to simply update the task's cpu field.
884 if (!p->array && !task_running(rq, p)) {
885 set_task_cpu(p, dest_cpu);
889 init_completion(&req->done);
891 req->dest_cpu = dest_cpu;
892 list_add(&req->list, &rq->migration_queue);
897 * wait_task_inactive - wait for a thread to unschedule.
899 * The caller must ensure that the task *will* unschedule sometime soon,
900 * else this function might spin for a *long* time. This function can't
901 * be called with interrupts off, or it may introduce deadlock with
902 * smp_call_function() if an IPI is sent by the same process we are
903 * waiting to become inactive.
905 void wait_task_inactive(task_t *p)
912 rq = task_rq_lock(p, &flags);
913 /* Must be off runqueue entirely, not preempted. */
914 if (unlikely(p->array || task_running(rq, p))) {
915 /* If it's preempted, we yield. It could be a while. */
916 preempted = !task_running(rq, p);
917 task_rq_unlock(rq, &flags);
923 task_rq_unlock(rq, &flags);
927 * kick_process - kick a running thread to enter/exit the kernel
928 * @p: the to-be-kicked thread
930 * Cause a process which is running on another CPU to enter
931 * kernel-mode, without any delay. (to get signals handled.)
933 * NOTE: this function doesnt have to take the runqueue lock,
934 * because all it wants to ensure is that the remote task enters
935 * the kernel. If the IPI races and the task has been migrated
936 * to another CPU then no harm is done and the purpose has been
939 void kick_process(task_t *p)
945 if ((cpu != smp_processor_id()) && task_curr(p))
946 smp_send_reschedule(cpu);
951 * Return a low guess at the load of a migration-source cpu.
953 * We want to under-estimate the load of migration sources, to
954 * balance conservatively.
956 static inline unsigned long source_load(int cpu, int type)
958 runqueue_t *rq = cpu_rq(cpu);
959 unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE;
963 return min(rq->cpu_load[type-1], load_now);
967 * Return a high guess at the load of a migration-target cpu
969 static inline unsigned long target_load(int cpu, int type)
971 runqueue_t *rq = cpu_rq(cpu);
972 unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE;
976 return max(rq->cpu_load[type-1], load_now);
980 * find_idlest_group finds and returns the least busy CPU group within the
983 static struct sched_group *
984 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
986 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
987 unsigned long min_load = ULONG_MAX, this_load = 0;
988 int load_idx = sd->forkexec_idx;
989 int imbalance = 100 + (sd->imbalance_pct-100)/2;
992 unsigned long load, avg_load;
996 /* Skip over this group if it has no CPUs allowed */
997 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1000 local_group = cpu_isset(this_cpu, group->cpumask);
1002 /* Tally up the load of all CPUs in the group */
1005 for_each_cpu_mask(i, group->cpumask) {
1006 /* Bias balancing toward cpus of our domain */
1008 load = source_load(i, load_idx);
1010 load = target_load(i, load_idx);
1015 /* Adjust by relative CPU power of the group */
1016 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1019 this_load = avg_load;
1021 } else if (avg_load < min_load) {
1022 min_load = avg_load;
1026 group = group->next;
1027 } while (group != sd->groups);
1029 if (!idlest || 100*this_load < imbalance*min_load)
1035 * find_idlest_queue - find the idlest runqueue among the cpus in group.
1038 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1041 unsigned long load, min_load = ULONG_MAX;
1045 /* Traverse only the allowed CPUs */
1046 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1048 for_each_cpu_mask(i, tmp) {
1049 load = source_load(i, 0);
1051 if (load < min_load || (load == min_load && i == this_cpu)) {
1061 * sched_balance_self: balance the current task (running on cpu) in domains
1062 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1065 * Balance, ie. select the least loaded group.
1067 * Returns the target CPU number, or the same CPU if no balancing is needed.
1069 * preempt must be disabled.
1071 static int sched_balance_self(int cpu, int flag)
1073 struct task_struct *t = current;
1074 struct sched_domain *tmp, *sd = NULL;
1076 for_each_domain(cpu, tmp) {
1077 if (tmp->flags & flag)
1083 struct sched_group *group;
1088 group = find_idlest_group(sd, t, cpu);
1092 new_cpu = find_idlest_cpu(group, t, cpu);
1093 if (new_cpu == -1 || new_cpu == cpu)
1096 /* Now try balancing at a lower domain level */
1100 weight = cpus_weight(span);
1101 for_each_domain(cpu, tmp) {
1102 if (weight <= cpus_weight(tmp->span))
1104 if (tmp->flags & flag)
1107 /* while loop will break here if sd == NULL */
1113 #endif /* CONFIG_SMP */
1116 * wake_idle() will wake a task on an idle cpu if task->cpu is
1117 * not idle and an idle cpu is available. The span of cpus to
1118 * search starts with cpus closest then further out as needed,
1119 * so we always favor a closer, idle cpu.
1121 * Returns the CPU we should wake onto.
1123 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1124 static int wake_idle(int cpu, task_t *p)
1127 struct sched_domain *sd;
1133 for_each_domain(cpu, sd) {
1134 if (sd->flags & SD_WAKE_IDLE) {
1135 cpus_and(tmp, sd->span, p->cpus_allowed);
1136 for_each_cpu_mask(i, tmp) {
1147 static inline int wake_idle(int cpu, task_t *p)
1154 * try_to_wake_up - wake up a thread
1155 * @p: the to-be-woken-up thread
1156 * @state: the mask of task states that can be woken
1157 * @sync: do a synchronous wakeup?
1159 * Put it on the run-queue if it's not already there. The "current"
1160 * thread is always on the run-queue (except when the actual
1161 * re-schedule is in progress), and as such you're allowed to do
1162 * the simpler "current->state = TASK_RUNNING" to mark yourself
1163 * runnable without the overhead of this.
1165 * returns failure only if the task is already active.
1167 static int try_to_wake_up(task_t *p, unsigned int state, int sync)
1169 int cpu, this_cpu, success = 0;
1170 unsigned long flags;
1174 unsigned long load, this_load;
1175 struct sched_domain *sd, *this_sd = NULL;
1179 rq = task_rq_lock(p, &flags);
1180 old_state = p->state;
1181 if (!(old_state & state))
1188 this_cpu = smp_processor_id();
1191 if (unlikely(task_running(rq, p)))
1196 schedstat_inc(rq, ttwu_cnt);
1197 if (cpu == this_cpu) {
1198 schedstat_inc(rq, ttwu_local);
1202 for_each_domain(this_cpu, sd) {
1203 if (cpu_isset(cpu, sd->span)) {
1204 schedstat_inc(sd, ttwu_wake_remote);
1210 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1214 * Check for affine wakeup and passive balancing possibilities.
1217 int idx = this_sd->wake_idx;
1218 unsigned int imbalance;
1220 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1222 load = source_load(cpu, idx);
1223 this_load = target_load(this_cpu, idx);
1225 new_cpu = this_cpu; /* Wake to this CPU if we can */
1227 if (this_sd->flags & SD_WAKE_AFFINE) {
1228 unsigned long tl = this_load;
1230 * If sync wakeup then subtract the (maximum possible)
1231 * effect of the currently running task from the load
1232 * of the current CPU:
1235 tl -= SCHED_LOAD_SCALE;
1238 tl + target_load(cpu, idx) <= SCHED_LOAD_SCALE) ||
1239 100*(tl + SCHED_LOAD_SCALE) <= imbalance*load) {
1241 * This domain has SD_WAKE_AFFINE and
1242 * p is cache cold in this domain, and
1243 * there is no bad imbalance.
1245 schedstat_inc(this_sd, ttwu_move_affine);
1251 * Start passive balancing when half the imbalance_pct
1254 if (this_sd->flags & SD_WAKE_BALANCE) {
1255 if (imbalance*this_load <= 100*load) {
1256 schedstat_inc(this_sd, ttwu_move_balance);
1262 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1264 new_cpu = wake_idle(new_cpu, p);
1265 if (new_cpu != cpu) {
1266 set_task_cpu(p, new_cpu);
1267 task_rq_unlock(rq, &flags);
1268 /* might preempt at this point */
1269 rq = task_rq_lock(p, &flags);
1270 old_state = p->state;
1271 if (!(old_state & state))
1276 this_cpu = smp_processor_id();
1281 #endif /* CONFIG_SMP */
1282 if (old_state == TASK_UNINTERRUPTIBLE) {
1283 rq->nr_uninterruptible--;
1285 * Tasks on involuntary sleep don't earn
1286 * sleep_avg beyond just interactive state.
1288 p->sleep_type = SLEEP_NONINTERACTIVE;
1292 * Tasks that have marked their sleep as noninteractive get
1293 * woken up with their sleep average not weighted in an
1296 if (old_state & TASK_NONINTERACTIVE)
1297 p->sleep_type = SLEEP_NONINTERACTIVE;
1300 activate_task(p, rq, cpu == this_cpu);
1302 * Sync wakeups (i.e. those types of wakeups where the waker
1303 * has indicated that it will leave the CPU in short order)
1304 * don't trigger a preemption, if the woken up task will run on
1305 * this cpu. (in this case the 'I will reschedule' promise of
1306 * the waker guarantees that the freshly woken up task is going
1307 * to be considered on this CPU.)
1309 if (!sync || cpu != this_cpu) {
1310 if (TASK_PREEMPTS_CURR(p, rq))
1311 resched_task(rq->curr);
1316 p->state = TASK_RUNNING;
1318 task_rq_unlock(rq, &flags);
1323 int fastcall wake_up_process(task_t *p)
1325 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1326 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1329 EXPORT_SYMBOL(wake_up_process);
1331 int fastcall wake_up_state(task_t *p, unsigned int state)
1333 return try_to_wake_up(p, state, 0);
1337 * Perform scheduler related setup for a newly forked process p.
1338 * p is forked by current.
1340 void fastcall sched_fork(task_t *p, int clone_flags)
1342 int cpu = get_cpu();
1345 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1347 set_task_cpu(p, cpu);
1350 * We mark the process as running here, but have not actually
1351 * inserted it onto the runqueue yet. This guarantees that
1352 * nobody will actually run it, and a signal or other external
1353 * event cannot wake it up and insert it on the runqueue either.
1355 p->state = TASK_RUNNING;
1356 INIT_LIST_HEAD(&p->run_list);
1358 #ifdef CONFIG_SCHEDSTATS
1359 memset(&p->sched_info, 0, sizeof(p->sched_info));
1361 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1364 #ifdef CONFIG_PREEMPT
1365 /* Want to start with kernel preemption disabled. */
1366 task_thread_info(p)->preempt_count = 1;
1369 * Share the timeslice between parent and child, thus the
1370 * total amount of pending timeslices in the system doesn't change,
1371 * resulting in more scheduling fairness.
1373 local_irq_disable();
1374 p->time_slice = (current->time_slice + 1) >> 1;
1376 * The remainder of the first timeslice might be recovered by
1377 * the parent if the child exits early enough.
1379 p->first_time_slice = 1;
1380 current->time_slice >>= 1;
1381 p->timestamp = sched_clock();
1382 if (unlikely(!current->time_slice)) {
1384 * This case is rare, it happens when the parent has only
1385 * a single jiffy left from its timeslice. Taking the
1386 * runqueue lock is not a problem.
1388 current->time_slice = 1;
1396 * wake_up_new_task - wake up a newly created task for the first time.
1398 * This function will do some initial scheduler statistics housekeeping
1399 * that must be done for every newly created context, then puts the task
1400 * on the runqueue and wakes it.
1402 void fastcall wake_up_new_task(task_t *p, unsigned long clone_flags)
1404 unsigned long flags;
1406 runqueue_t *rq, *this_rq;
1408 rq = task_rq_lock(p, &flags);
1409 BUG_ON(p->state != TASK_RUNNING);
1410 this_cpu = smp_processor_id();
1414 * We decrease the sleep average of forking parents
1415 * and children as well, to keep max-interactive tasks
1416 * from forking tasks that are max-interactive. The parent
1417 * (current) is done further down, under its lock.
1419 p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1420 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1422 p->prio = effective_prio(p);
1424 if (likely(cpu == this_cpu)) {
1425 if (!(clone_flags & CLONE_VM)) {
1427 * The VM isn't cloned, so we're in a good position to
1428 * do child-runs-first in anticipation of an exec. This
1429 * usually avoids a lot of COW overhead.
1431 if (unlikely(!current->array))
1432 __activate_task(p, rq);
1434 p->prio = current->prio;
1435 list_add_tail(&p->run_list, ¤t->run_list);
1436 p->array = current->array;
1437 p->array->nr_active++;
1442 /* Run child last */
1443 __activate_task(p, rq);
1445 * We skip the following code due to cpu == this_cpu
1447 * task_rq_unlock(rq, &flags);
1448 * this_rq = task_rq_lock(current, &flags);
1452 this_rq = cpu_rq(this_cpu);
1455 * Not the local CPU - must adjust timestamp. This should
1456 * get optimised away in the !CONFIG_SMP case.
1458 p->timestamp = (p->timestamp - this_rq->timestamp_last_tick)
1459 + rq->timestamp_last_tick;
1460 __activate_task(p, rq);
1461 if (TASK_PREEMPTS_CURR(p, rq))
1462 resched_task(rq->curr);
1465 * Parent and child are on different CPUs, now get the
1466 * parent runqueue to update the parent's ->sleep_avg:
1468 task_rq_unlock(rq, &flags);
1469 this_rq = task_rq_lock(current, &flags);
1471 current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1472 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1473 task_rq_unlock(this_rq, &flags);
1477 * Potentially available exiting-child timeslices are
1478 * retrieved here - this way the parent does not get
1479 * penalized for creating too many threads.
1481 * (this cannot be used to 'generate' timeslices
1482 * artificially, because any timeslice recovered here
1483 * was given away by the parent in the first place.)
1485 void fastcall sched_exit(task_t *p)
1487 unsigned long flags;
1491 * If the child was a (relative-) CPU hog then decrease
1492 * the sleep_avg of the parent as well.
1494 rq = task_rq_lock(p->parent, &flags);
1495 if (p->first_time_slice && task_cpu(p) == task_cpu(p->parent)) {
1496 p->parent->time_slice += p->time_slice;
1497 if (unlikely(p->parent->time_slice > task_timeslice(p)))
1498 p->parent->time_slice = task_timeslice(p);
1500 if (p->sleep_avg < p->parent->sleep_avg)
1501 p->parent->sleep_avg = p->parent->sleep_avg /
1502 (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg /
1504 task_rq_unlock(rq, &flags);
1508 * prepare_task_switch - prepare to switch tasks
1509 * @rq: the runqueue preparing to switch
1510 * @next: the task we are going to switch to.
1512 * This is called with the rq lock held and interrupts off. It must
1513 * be paired with a subsequent finish_task_switch after the context
1516 * prepare_task_switch sets up locking and calls architecture specific
1519 static inline void prepare_task_switch(runqueue_t *rq, task_t *next)
1521 prepare_lock_switch(rq, next);
1522 prepare_arch_switch(next);
1526 * finish_task_switch - clean up after a task-switch
1527 * @rq: runqueue associated with task-switch
1528 * @prev: the thread we just switched away from.
1530 * finish_task_switch must be called after the context switch, paired
1531 * with a prepare_task_switch call before the context switch.
1532 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1533 * and do any other architecture-specific cleanup actions.
1535 * Note that we may have delayed dropping an mm in context_switch(). If
1536 * so, we finish that here outside of the runqueue lock. (Doing it
1537 * with the lock held can cause deadlocks; see schedule() for
1540 static inline void finish_task_switch(runqueue_t *rq, task_t *prev)
1541 __releases(rq->lock)
1543 struct mm_struct *mm = rq->prev_mm;
1544 unsigned long prev_task_flags;
1549 * A task struct has one reference for the use as "current".
1550 * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1551 * calls schedule one last time. The schedule call will never return,
1552 * and the scheduled task must drop that reference.
1553 * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1554 * still held, otherwise prev could be scheduled on another cpu, die
1555 * there before we look at prev->state, and then the reference would
1557 * Manfred Spraul <manfred@colorfullife.com>
1559 prev_task_flags = prev->flags;
1560 finish_arch_switch(prev);
1561 finish_lock_switch(rq, prev);
1564 if (unlikely(prev_task_flags & PF_DEAD)) {
1566 * Remove function-return probe instances associated with this
1567 * task and put them back on the free list.
1569 kprobe_flush_task(prev);
1570 put_task_struct(prev);
1575 * schedule_tail - first thing a freshly forked thread must call.
1576 * @prev: the thread we just switched away from.
1578 asmlinkage void schedule_tail(task_t *prev)
1579 __releases(rq->lock)
1581 runqueue_t *rq = this_rq();
1582 finish_task_switch(rq, prev);
1583 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1584 /* In this case, finish_task_switch does not reenable preemption */
1587 if (current->set_child_tid)
1588 put_user(current->pid, current->set_child_tid);
1592 * context_switch - switch to the new MM and the new
1593 * thread's register state.
1596 task_t * context_switch(runqueue_t *rq, task_t *prev, task_t *next)
1598 struct mm_struct *mm = next->mm;
1599 struct mm_struct *oldmm = prev->active_mm;
1601 if (unlikely(!mm)) {
1602 next->active_mm = oldmm;
1603 atomic_inc(&oldmm->mm_count);
1604 enter_lazy_tlb(oldmm, next);
1606 switch_mm(oldmm, mm, next);
1608 if (unlikely(!prev->mm)) {
1609 prev->active_mm = NULL;
1610 WARN_ON(rq->prev_mm);
1611 rq->prev_mm = oldmm;
1614 /* Here we just switch the register state and the stack. */
1615 switch_to(prev, next, prev);
1621 * nr_running, nr_uninterruptible and nr_context_switches:
1623 * externally visible scheduler statistics: current number of runnable
1624 * threads, current number of uninterruptible-sleeping threads, total
1625 * number of context switches performed since bootup.
1627 unsigned long nr_running(void)
1629 unsigned long i, sum = 0;
1631 for_each_online_cpu(i)
1632 sum += cpu_rq(i)->nr_running;
1637 unsigned long nr_uninterruptible(void)
1639 unsigned long i, sum = 0;
1641 for_each_possible_cpu(i)
1642 sum += cpu_rq(i)->nr_uninterruptible;
1645 * Since we read the counters lockless, it might be slightly
1646 * inaccurate. Do not allow it to go below zero though:
1648 if (unlikely((long)sum < 0))
1654 unsigned long long nr_context_switches(void)
1657 unsigned long long sum = 0;
1659 for_each_possible_cpu(i)
1660 sum += cpu_rq(i)->nr_switches;
1665 unsigned long nr_iowait(void)
1667 unsigned long i, sum = 0;
1669 for_each_possible_cpu(i)
1670 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1675 unsigned long nr_active(void)
1677 unsigned long i, running = 0, uninterruptible = 0;
1679 for_each_online_cpu(i) {
1680 running += cpu_rq(i)->nr_running;
1681 uninterruptible += cpu_rq(i)->nr_uninterruptible;
1684 if (unlikely((long)uninterruptible < 0))
1685 uninterruptible = 0;
1687 return running + uninterruptible;
1693 * double_rq_lock - safely lock two runqueues
1695 * Note this does not disable interrupts like task_rq_lock,
1696 * you need to do so manually before calling.
1698 static void double_rq_lock(runqueue_t *rq1, runqueue_t *rq2)
1699 __acquires(rq1->lock)
1700 __acquires(rq2->lock)
1703 spin_lock(&rq1->lock);
1704 __acquire(rq2->lock); /* Fake it out ;) */
1707 spin_lock(&rq1->lock);
1708 spin_lock(&rq2->lock);
1710 spin_lock(&rq2->lock);
1711 spin_lock(&rq1->lock);
1717 * double_rq_unlock - safely unlock two runqueues
1719 * Note this does not restore interrupts like task_rq_unlock,
1720 * you need to do so manually after calling.
1722 static void double_rq_unlock(runqueue_t *rq1, runqueue_t *rq2)
1723 __releases(rq1->lock)
1724 __releases(rq2->lock)
1726 spin_unlock(&rq1->lock);
1728 spin_unlock(&rq2->lock);
1730 __release(rq2->lock);
1734 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1736 static void double_lock_balance(runqueue_t *this_rq, runqueue_t *busiest)
1737 __releases(this_rq->lock)
1738 __acquires(busiest->lock)
1739 __acquires(this_rq->lock)
1741 if (unlikely(!spin_trylock(&busiest->lock))) {
1742 if (busiest < this_rq) {
1743 spin_unlock(&this_rq->lock);
1744 spin_lock(&busiest->lock);
1745 spin_lock(&this_rq->lock);
1747 spin_lock(&busiest->lock);
1752 * If dest_cpu is allowed for this process, migrate the task to it.
1753 * This is accomplished by forcing the cpu_allowed mask to only
1754 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
1755 * the cpu_allowed mask is restored.
1757 static void sched_migrate_task(task_t *p, int dest_cpu)
1759 migration_req_t req;
1761 unsigned long flags;
1763 rq = task_rq_lock(p, &flags);
1764 if (!cpu_isset(dest_cpu, p->cpus_allowed)
1765 || unlikely(cpu_is_offline(dest_cpu)))
1768 /* force the process onto the specified CPU */
1769 if (migrate_task(p, dest_cpu, &req)) {
1770 /* Need to wait for migration thread (might exit: take ref). */
1771 struct task_struct *mt = rq->migration_thread;
1772 get_task_struct(mt);
1773 task_rq_unlock(rq, &flags);
1774 wake_up_process(mt);
1775 put_task_struct(mt);
1776 wait_for_completion(&req.done);
1780 task_rq_unlock(rq, &flags);
1784 * sched_exec - execve() is a valuable balancing opportunity, because at
1785 * this point the task has the smallest effective memory and cache footprint.
1787 void sched_exec(void)
1789 int new_cpu, this_cpu = get_cpu();
1790 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
1792 if (new_cpu != this_cpu)
1793 sched_migrate_task(current, new_cpu);
1797 * pull_task - move a task from a remote runqueue to the local runqueue.
1798 * Both runqueues must be locked.
1801 void pull_task(runqueue_t *src_rq, prio_array_t *src_array, task_t *p,
1802 runqueue_t *this_rq, prio_array_t *this_array, int this_cpu)
1804 dequeue_task(p, src_array);
1805 src_rq->nr_running--;
1806 set_task_cpu(p, this_cpu);
1807 this_rq->nr_running++;
1808 enqueue_task(p, this_array);
1809 p->timestamp = (p->timestamp - src_rq->timestamp_last_tick)
1810 + this_rq->timestamp_last_tick;
1812 * Note that idle threads have a prio of MAX_PRIO, for this test
1813 * to be always true for them.
1815 if (TASK_PREEMPTS_CURR(p, this_rq))
1816 resched_task(this_rq->curr);
1820 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1823 int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu,
1824 struct sched_domain *sd, enum idle_type idle,
1828 * We do not migrate tasks that are:
1829 * 1) running (obviously), or
1830 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1831 * 3) are cache-hot on their current CPU.
1833 if (!cpu_isset(this_cpu, p->cpus_allowed))
1837 if (task_running(rq, p))
1841 * Aggressive migration if:
1842 * 1) task is cache cold, or
1843 * 2) too many balance attempts have failed.
1846 if (sd->nr_balance_failed > sd->cache_nice_tries)
1849 if (task_hot(p, rq->timestamp_last_tick, sd))
1855 * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq,
1856 * as part of a balancing operation within "domain". Returns the number of
1859 * Called with both runqueues locked.
1861 static int move_tasks(runqueue_t *this_rq, int this_cpu, runqueue_t *busiest,
1862 unsigned long max_nr_move, struct sched_domain *sd,
1863 enum idle_type idle, int *all_pinned)
1865 prio_array_t *array, *dst_array;
1866 struct list_head *head, *curr;
1867 int idx, pulled = 0, pinned = 0;
1870 if (max_nr_move == 0)
1876 * We first consider expired tasks. Those will likely not be
1877 * executed in the near future, and they are most likely to
1878 * be cache-cold, thus switching CPUs has the least effect
1881 if (busiest->expired->nr_active) {
1882 array = busiest->expired;
1883 dst_array = this_rq->expired;
1885 array = busiest->active;
1886 dst_array = this_rq->active;
1890 /* Start searching at priority 0: */
1894 idx = sched_find_first_bit(array->bitmap);
1896 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
1897 if (idx >= MAX_PRIO) {
1898 if (array == busiest->expired && busiest->active->nr_active) {
1899 array = busiest->active;
1900 dst_array = this_rq->active;
1906 head = array->queue + idx;
1909 tmp = list_entry(curr, task_t, run_list);
1913 if (!can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
1920 #ifdef CONFIG_SCHEDSTATS
1921 if (task_hot(tmp, busiest->timestamp_last_tick, sd))
1922 schedstat_inc(sd, lb_hot_gained[idle]);
1925 pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
1928 /* We only want to steal up to the prescribed number of tasks. */
1929 if (pulled < max_nr_move) {
1937 * Right now, this is the only place pull_task() is called,
1938 * so we can safely collect pull_task() stats here rather than
1939 * inside pull_task().
1941 schedstat_add(sd, lb_gained[idle], pulled);
1944 *all_pinned = pinned;
1949 * find_busiest_group finds and returns the busiest CPU group within the
1950 * domain. It calculates and returns the number of tasks which should be
1951 * moved to restore balance via the imbalance parameter.
1953 static struct sched_group *
1954 find_busiest_group(struct sched_domain *sd, int this_cpu,
1955 unsigned long *imbalance, enum idle_type idle, int *sd_idle)
1957 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
1958 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
1959 unsigned long max_pull;
1962 max_load = this_load = total_load = total_pwr = 0;
1963 if (idle == NOT_IDLE)
1964 load_idx = sd->busy_idx;
1965 else if (idle == NEWLY_IDLE)
1966 load_idx = sd->newidle_idx;
1968 load_idx = sd->idle_idx;
1975 local_group = cpu_isset(this_cpu, group->cpumask);
1977 /* Tally up the load of all CPUs in the group */
1980 for_each_cpu_mask(i, group->cpumask) {
1981 if (*sd_idle && !idle_cpu(i))
1984 /* Bias balancing toward cpus of our domain */
1986 load = target_load(i, load_idx);
1988 load = source_load(i, load_idx);
1993 total_load += avg_load;
1994 total_pwr += group->cpu_power;
1996 /* Adjust by relative CPU power of the group */
1997 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
2000 this_load = avg_load;
2002 } else if (avg_load > max_load) {
2003 max_load = avg_load;
2006 group = group->next;
2007 } while (group != sd->groups);
2009 if (!busiest || this_load >= max_load || max_load <= SCHED_LOAD_SCALE)
2012 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2014 if (this_load >= avg_load ||
2015 100*max_load <= sd->imbalance_pct*this_load)
2019 * We're trying to get all the cpus to the average_load, so we don't
2020 * want to push ourselves above the average load, nor do we wish to
2021 * reduce the max loaded cpu below the average load, as either of these
2022 * actions would just result in more rebalancing later, and ping-pong
2023 * tasks around. Thus we look for the minimum possible imbalance.
2024 * Negative imbalances (*we* are more loaded than anyone else) will
2025 * be counted as no imbalance for these purposes -- we can't fix that
2026 * by pulling tasks to us. Be careful of negative numbers as they'll
2027 * appear as very large values with unsigned longs.
2030 /* Don't want to pull so many tasks that a group would go idle */
2031 max_pull = min(max_load - avg_load, max_load - SCHED_LOAD_SCALE);
2033 /* How much load to actually move to equalise the imbalance */
2034 *imbalance = min(max_pull * busiest->cpu_power,
2035 (avg_load - this_load) * this->cpu_power)
2038 if (*imbalance < SCHED_LOAD_SCALE) {
2039 unsigned long pwr_now = 0, pwr_move = 0;
2042 if (max_load - this_load >= SCHED_LOAD_SCALE*2) {
2048 * OK, we don't have enough imbalance to justify moving tasks,
2049 * however we may be able to increase total CPU power used by
2053 pwr_now += busiest->cpu_power*min(SCHED_LOAD_SCALE, max_load);
2054 pwr_now += this->cpu_power*min(SCHED_LOAD_SCALE, this_load);
2055 pwr_now /= SCHED_LOAD_SCALE;
2057 /* Amount of load we'd subtract */
2058 tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/busiest->cpu_power;
2060 pwr_move += busiest->cpu_power*min(SCHED_LOAD_SCALE,
2063 /* Amount of load we'd add */
2064 if (max_load*busiest->cpu_power <
2065 SCHED_LOAD_SCALE*SCHED_LOAD_SCALE)
2066 tmp = max_load*busiest->cpu_power/this->cpu_power;
2068 tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/this->cpu_power;
2069 pwr_move += this->cpu_power*min(SCHED_LOAD_SCALE, this_load + tmp);
2070 pwr_move /= SCHED_LOAD_SCALE;
2072 /* Move if we gain throughput */
2073 if (pwr_move <= pwr_now)
2080 /* Get rid of the scaling factor, rounding down as we divide */
2081 *imbalance = *imbalance / SCHED_LOAD_SCALE;
2091 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2093 static runqueue_t *find_busiest_queue(struct sched_group *group,
2094 enum idle_type idle)
2096 unsigned long load, max_load = 0;
2097 runqueue_t *busiest = NULL;
2100 for_each_cpu_mask(i, group->cpumask) {
2101 load = source_load(i, 0);
2103 if (load > max_load) {
2105 busiest = cpu_rq(i);
2113 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2114 * so long as it is large enough.
2116 #define MAX_PINNED_INTERVAL 512
2119 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2120 * tasks if there is an imbalance.
2122 * Called with this_rq unlocked.
2124 static int load_balance(int this_cpu, runqueue_t *this_rq,
2125 struct sched_domain *sd, enum idle_type idle)
2127 struct sched_group *group;
2128 runqueue_t *busiest;
2129 unsigned long imbalance;
2130 int nr_moved, all_pinned = 0;
2131 int active_balance = 0;
2134 if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER)
2137 schedstat_inc(sd, lb_cnt[idle]);
2139 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle);
2141 schedstat_inc(sd, lb_nobusyg[idle]);
2145 busiest = find_busiest_queue(group, idle);
2147 schedstat_inc(sd, lb_nobusyq[idle]);
2151 BUG_ON(busiest == this_rq);
2153 schedstat_add(sd, lb_imbalance[idle], imbalance);
2156 if (busiest->nr_running > 1) {
2158 * Attempt to move tasks. If find_busiest_group has found
2159 * an imbalance but busiest->nr_running <= 1, the group is
2160 * still unbalanced. nr_moved simply stays zero, so it is
2161 * correctly treated as an imbalance.
2163 double_rq_lock(this_rq, busiest);
2164 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2165 imbalance, sd, idle, &all_pinned);
2166 double_rq_unlock(this_rq, busiest);
2168 /* All tasks on this runqueue were pinned by CPU affinity */
2169 if (unlikely(all_pinned))
2174 schedstat_inc(sd, lb_failed[idle]);
2175 sd->nr_balance_failed++;
2177 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2179 spin_lock(&busiest->lock);
2181 /* don't kick the migration_thread, if the curr
2182 * task on busiest cpu can't be moved to this_cpu
2184 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2185 spin_unlock(&busiest->lock);
2187 goto out_one_pinned;
2190 if (!busiest->active_balance) {
2191 busiest->active_balance = 1;
2192 busiest->push_cpu = this_cpu;
2195 spin_unlock(&busiest->lock);
2197 wake_up_process(busiest->migration_thread);
2200 * We've kicked active balancing, reset the failure
2203 sd->nr_balance_failed = sd->cache_nice_tries+1;
2206 sd->nr_balance_failed = 0;
2208 if (likely(!active_balance)) {
2209 /* We were unbalanced, so reset the balancing interval */
2210 sd->balance_interval = sd->min_interval;
2213 * If we've begun active balancing, start to back off. This
2214 * case may not be covered by the all_pinned logic if there
2215 * is only 1 task on the busy runqueue (because we don't call
2218 if (sd->balance_interval < sd->max_interval)
2219 sd->balance_interval *= 2;
2222 if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2227 schedstat_inc(sd, lb_balanced[idle]);
2229 sd->nr_balance_failed = 0;
2232 /* tune up the balancing interval */
2233 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2234 (sd->balance_interval < sd->max_interval))
2235 sd->balance_interval *= 2;
2237 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2243 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2244 * tasks if there is an imbalance.
2246 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2247 * this_rq is locked.
2249 static int load_balance_newidle(int this_cpu, runqueue_t *this_rq,
2250 struct sched_domain *sd)
2252 struct sched_group *group;
2253 runqueue_t *busiest = NULL;
2254 unsigned long imbalance;
2258 if (sd->flags & SD_SHARE_CPUPOWER)
2261 schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
2262 group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE, &sd_idle);
2264 schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]);
2268 busiest = find_busiest_queue(group, NEWLY_IDLE);
2270 schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]);
2274 BUG_ON(busiest == this_rq);
2276 schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
2279 if (busiest->nr_running > 1) {
2280 /* Attempt to move tasks */
2281 double_lock_balance(this_rq, busiest);
2282 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2283 imbalance, sd, NEWLY_IDLE, NULL);
2284 spin_unlock(&busiest->lock);
2288 schedstat_inc(sd, lb_failed[NEWLY_IDLE]);
2289 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2292 sd->nr_balance_failed = 0;
2297 schedstat_inc(sd, lb_balanced[NEWLY_IDLE]);
2298 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2300 sd->nr_balance_failed = 0;
2305 * idle_balance is called by schedule() if this_cpu is about to become
2306 * idle. Attempts to pull tasks from other CPUs.
2308 static void idle_balance(int this_cpu, runqueue_t *this_rq)
2310 struct sched_domain *sd;
2312 for_each_domain(this_cpu, sd) {
2313 if (sd->flags & SD_BALANCE_NEWIDLE) {
2314 if (load_balance_newidle(this_cpu, this_rq, sd)) {
2315 /* We've pulled tasks over so stop searching */
2323 * active_load_balance is run by migration threads. It pushes running tasks
2324 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2325 * running on each physical CPU where possible, and avoids physical /
2326 * logical imbalances.
2328 * Called with busiest_rq locked.
2330 static void active_load_balance(runqueue_t *busiest_rq, int busiest_cpu)
2332 struct sched_domain *sd;
2333 runqueue_t *target_rq;
2334 int target_cpu = busiest_rq->push_cpu;
2336 if (busiest_rq->nr_running <= 1)
2337 /* no task to move */
2340 target_rq = cpu_rq(target_cpu);
2343 * This condition is "impossible", if it occurs
2344 * we need to fix it. Originally reported by
2345 * Bjorn Helgaas on a 128-cpu setup.
2347 BUG_ON(busiest_rq == target_rq);
2349 /* move a task from busiest_rq to target_rq */
2350 double_lock_balance(busiest_rq, target_rq);
2352 /* Search for an sd spanning us and the target CPU. */
2353 for_each_domain(target_cpu, sd) {
2354 if ((sd->flags & SD_LOAD_BALANCE) &&
2355 cpu_isset(busiest_cpu, sd->span))
2359 if (unlikely(sd == NULL))
2362 schedstat_inc(sd, alb_cnt);
2364 if (move_tasks(target_rq, target_cpu, busiest_rq, 1, sd, SCHED_IDLE, NULL))
2365 schedstat_inc(sd, alb_pushed);
2367 schedstat_inc(sd, alb_failed);
2369 spin_unlock(&target_rq->lock);
2373 * rebalance_tick will get called every timer tick, on every CPU.
2375 * It checks each scheduling domain to see if it is due to be balanced,
2376 * and initiates a balancing operation if so.
2378 * Balancing parameters are set up in arch_init_sched_domains.
2381 /* Don't have all balancing operations going off at once */
2382 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2384 static void rebalance_tick(int this_cpu, runqueue_t *this_rq,
2385 enum idle_type idle)
2387 unsigned long old_load, this_load;
2388 unsigned long j = jiffies + CPU_OFFSET(this_cpu);
2389 struct sched_domain *sd;
2392 this_load = this_rq->nr_running * SCHED_LOAD_SCALE;
2393 /* Update our load */
2394 for (i = 0; i < 3; i++) {
2395 unsigned long new_load = this_load;
2397 old_load = this_rq->cpu_load[i];
2399 * Round up the averaging division if load is increasing. This
2400 * prevents us from getting stuck on 9 if the load is 10, for
2403 if (new_load > old_load)
2404 new_load += scale-1;
2405 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale;
2408 for_each_domain(this_cpu, sd) {
2409 unsigned long interval;
2411 if (!(sd->flags & SD_LOAD_BALANCE))
2414 interval = sd->balance_interval;
2415 if (idle != SCHED_IDLE)
2416 interval *= sd->busy_factor;
2418 /* scale ms to jiffies */
2419 interval = msecs_to_jiffies(interval);
2420 if (unlikely(!interval))
2423 if (j - sd->last_balance >= interval) {
2424 if (load_balance(this_cpu, this_rq, sd, idle)) {
2426 * We've pulled tasks over so either we're no
2427 * longer idle, or one of our SMT siblings is
2432 sd->last_balance += interval;
2438 * on UP we do not need to balance between CPUs:
2440 static inline void rebalance_tick(int cpu, runqueue_t *rq, enum idle_type idle)
2443 static inline void idle_balance(int cpu, runqueue_t *rq)
2448 static inline int wake_priority_sleeper(runqueue_t *rq)
2451 #ifdef CONFIG_SCHED_SMT
2452 spin_lock(&rq->lock);
2454 * If an SMT sibling task has been put to sleep for priority
2455 * reasons reschedule the idle task to see if it can now run.
2457 if (rq->nr_running) {
2458 resched_task(rq->idle);
2461 spin_unlock(&rq->lock);
2466 DEFINE_PER_CPU(struct kernel_stat, kstat);
2468 EXPORT_PER_CPU_SYMBOL(kstat);
2471 * This is called on clock ticks and on context switches.
2472 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2474 static inline void update_cpu_clock(task_t *p, runqueue_t *rq,
2475 unsigned long long now)
2477 unsigned long long last = max(p->timestamp, rq->timestamp_last_tick);
2478 p->sched_time += now - last;
2482 * Return current->sched_time plus any more ns on the sched_clock
2483 * that have not yet been banked.
2485 unsigned long long current_sched_time(const task_t *tsk)
2487 unsigned long long ns;
2488 unsigned long flags;
2489 local_irq_save(flags);
2490 ns = max(tsk->timestamp, task_rq(tsk)->timestamp_last_tick);
2491 ns = tsk->sched_time + (sched_clock() - ns);
2492 local_irq_restore(flags);
2497 * We place interactive tasks back into the active array, if possible.
2499 * To guarantee that this does not starve expired tasks we ignore the
2500 * interactivity of a task if the first expired task had to wait more
2501 * than a 'reasonable' amount of time. This deadline timeout is
2502 * load-dependent, as the frequency of array switched decreases with
2503 * increasing number of running tasks. We also ignore the interactivity
2504 * if a better static_prio task has expired:
2506 #define EXPIRED_STARVING(rq) \
2507 ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2508 (jiffies - (rq)->expired_timestamp >= \
2509 STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2510 ((rq)->curr->static_prio > (rq)->best_expired_prio))
2513 * Account user cpu time to a process.
2514 * @p: the process that the cpu time gets accounted to
2515 * @hardirq_offset: the offset to subtract from hardirq_count()
2516 * @cputime: the cpu time spent in user space since the last update
2518 void account_user_time(struct task_struct *p, cputime_t cputime)
2520 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2523 p->utime = cputime_add(p->utime, cputime);
2525 /* Add user time to cpustat. */
2526 tmp = cputime_to_cputime64(cputime);
2527 if (TASK_NICE(p) > 0)
2528 cpustat->nice = cputime64_add(cpustat->nice, tmp);
2530 cpustat->user = cputime64_add(cpustat->user, tmp);
2534 * Account system cpu time to a process.
2535 * @p: the process that the cpu time gets accounted to
2536 * @hardirq_offset: the offset to subtract from hardirq_count()
2537 * @cputime: the cpu time spent in kernel space since the last update
2539 void account_system_time(struct task_struct *p, int hardirq_offset,
2542 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2543 runqueue_t *rq = this_rq();
2546 p->stime = cputime_add(p->stime, cputime);
2548 /* Add system time to cpustat. */
2549 tmp = cputime_to_cputime64(cputime);
2550 if (hardirq_count() - hardirq_offset)
2551 cpustat->irq = cputime64_add(cpustat->irq, tmp);
2552 else if (softirq_count())
2553 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
2554 else if (p != rq->idle)
2555 cpustat->system = cputime64_add(cpustat->system, tmp);
2556 else if (atomic_read(&rq->nr_iowait) > 0)
2557 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2559 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2560 /* Account for system time used */
2561 acct_update_integrals(p);
2565 * Account for involuntary wait time.
2566 * @p: the process from which the cpu time has been stolen
2567 * @steal: the cpu time spent in involuntary wait
2569 void account_steal_time(struct task_struct *p, cputime_t steal)
2571 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2572 cputime64_t tmp = cputime_to_cputime64(steal);
2573 runqueue_t *rq = this_rq();
2575 if (p == rq->idle) {
2576 p->stime = cputime_add(p->stime, steal);
2577 if (atomic_read(&rq->nr_iowait) > 0)
2578 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2580 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2582 cpustat->steal = cputime64_add(cpustat->steal, tmp);
2586 * This function gets called by the timer code, with HZ frequency.
2587 * We call it with interrupts disabled.
2589 * It also gets called by the fork code, when changing the parent's
2592 void scheduler_tick(void)
2594 int cpu = smp_processor_id();
2595 runqueue_t *rq = this_rq();
2596 task_t *p = current;
2597 unsigned long long now = sched_clock();
2599 update_cpu_clock(p, rq, now);
2601 rq->timestamp_last_tick = now;
2603 if (p == rq->idle) {
2604 if (wake_priority_sleeper(rq))
2606 rebalance_tick(cpu, rq, SCHED_IDLE);
2610 /* Task might have expired already, but not scheduled off yet */
2611 if (p->array != rq->active) {
2612 set_tsk_need_resched(p);
2615 spin_lock(&rq->lock);
2617 * The task was running during this tick - update the
2618 * time slice counter. Note: we do not update a thread's
2619 * priority until it either goes to sleep or uses up its
2620 * timeslice. This makes it possible for interactive tasks
2621 * to use up their timeslices at their highest priority levels.
2625 * RR tasks need a special form of timeslice management.
2626 * FIFO tasks have no timeslices.
2628 if ((p->policy == SCHED_RR) && !--p->time_slice) {
2629 p->time_slice = task_timeslice(p);
2630 p->first_time_slice = 0;
2631 set_tsk_need_resched(p);
2633 /* put it at the end of the queue: */
2634 requeue_task(p, rq->active);
2638 if (!--p->time_slice) {
2639 dequeue_task(p, rq->active);
2640 set_tsk_need_resched(p);
2641 p->prio = effective_prio(p);
2642 p->time_slice = task_timeslice(p);
2643 p->first_time_slice = 0;
2645 if (!rq->expired_timestamp)
2646 rq->expired_timestamp = jiffies;
2647 if (!TASK_INTERACTIVE(p) || EXPIRED_STARVING(rq)) {
2648 enqueue_task(p, rq->expired);
2649 if (p->static_prio < rq->best_expired_prio)
2650 rq->best_expired_prio = p->static_prio;
2652 enqueue_task(p, rq->active);
2655 * Prevent a too long timeslice allowing a task to monopolize
2656 * the CPU. We do this by splitting up the timeslice into
2659 * Note: this does not mean the task's timeslices expire or
2660 * get lost in any way, they just might be preempted by
2661 * another task of equal priority. (one with higher
2662 * priority would have preempted this task already.) We
2663 * requeue this task to the end of the list on this priority
2664 * level, which is in essence a round-robin of tasks with
2667 * This only applies to tasks in the interactive
2668 * delta range with at least TIMESLICE_GRANULARITY to requeue.
2670 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
2671 p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
2672 (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
2673 (p->array == rq->active)) {
2675 requeue_task(p, rq->active);
2676 set_tsk_need_resched(p);
2680 spin_unlock(&rq->lock);
2682 rebalance_tick(cpu, rq, NOT_IDLE);
2685 #ifdef CONFIG_SCHED_SMT
2686 static inline void wakeup_busy_runqueue(runqueue_t *rq)
2688 /* If an SMT runqueue is sleeping due to priority reasons wake it up */
2689 if (rq->curr == rq->idle && rq->nr_running)
2690 resched_task(rq->idle);
2694 * Called with interrupt disabled and this_rq's runqueue locked.
2696 static void wake_sleeping_dependent(int this_cpu)
2698 struct sched_domain *tmp, *sd = NULL;
2701 for_each_domain(this_cpu, tmp) {
2702 if (tmp->flags & SD_SHARE_CPUPOWER) {
2711 for_each_cpu_mask(i, sd->span) {
2712 runqueue_t *smt_rq = cpu_rq(i);
2716 if (unlikely(!spin_trylock(&smt_rq->lock)))
2719 wakeup_busy_runqueue(smt_rq);
2720 spin_unlock(&smt_rq->lock);
2725 * number of 'lost' timeslices this task wont be able to fully
2726 * utilize, if another task runs on a sibling. This models the
2727 * slowdown effect of other tasks running on siblings:
2729 static inline unsigned long smt_slice(task_t *p, struct sched_domain *sd)
2731 return p->time_slice * (100 - sd->per_cpu_gain) / 100;
2735 * To minimise lock contention and not have to drop this_rq's runlock we only
2736 * trylock the sibling runqueues and bypass those runqueues if we fail to
2737 * acquire their lock. As we only trylock the normal locking order does not
2738 * need to be obeyed.
2740 static int dependent_sleeper(int this_cpu, runqueue_t *this_rq, task_t *p)
2742 struct sched_domain *tmp, *sd = NULL;
2745 /* kernel/rt threads do not participate in dependent sleeping */
2746 if (!p->mm || rt_task(p))
2749 for_each_domain(this_cpu, tmp) {
2750 if (tmp->flags & SD_SHARE_CPUPOWER) {
2759 for_each_cpu_mask(i, sd->span) {
2767 if (unlikely(!spin_trylock(&smt_rq->lock)))
2770 smt_curr = smt_rq->curr;
2776 * If a user task with lower static priority than the
2777 * running task on the SMT sibling is trying to schedule,
2778 * delay it till there is proportionately less timeslice
2779 * left of the sibling task to prevent a lower priority
2780 * task from using an unfair proportion of the
2781 * physical cpu's resources. -ck
2783 if (rt_task(smt_curr)) {
2785 * With real time tasks we run non-rt tasks only
2786 * per_cpu_gain% of the time.
2788 if ((jiffies % DEF_TIMESLICE) >
2789 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
2792 if (smt_curr->static_prio < p->static_prio &&
2793 !TASK_PREEMPTS_CURR(p, smt_rq) &&
2794 smt_slice(smt_curr, sd) > task_timeslice(p))
2798 spin_unlock(&smt_rq->lock);
2803 static inline void wake_sleeping_dependent(int this_cpu)
2807 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq,
2814 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
2816 void fastcall add_preempt_count(int val)
2821 BUG_ON((preempt_count() < 0));
2822 preempt_count() += val;
2824 * Spinlock count overflowing soon?
2826 BUG_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10);
2828 EXPORT_SYMBOL(add_preempt_count);
2830 void fastcall sub_preempt_count(int val)
2835 BUG_ON(val > preempt_count());
2837 * Is the spinlock portion underflowing?
2839 BUG_ON((val < PREEMPT_MASK) && !(preempt_count() & PREEMPT_MASK));
2840 preempt_count() -= val;
2842 EXPORT_SYMBOL(sub_preempt_count);
2846 static inline int interactive_sleep(enum sleep_type sleep_type)
2848 return (sleep_type == SLEEP_INTERACTIVE ||
2849 sleep_type == SLEEP_INTERRUPTED);
2853 * schedule() is the main scheduler function.
2855 asmlinkage void __sched schedule(void)
2858 task_t *prev, *next;
2860 prio_array_t *array;
2861 struct list_head *queue;
2862 unsigned long long now;
2863 unsigned long run_time;
2864 int cpu, idx, new_prio;
2867 * Test if we are atomic. Since do_exit() needs to call into
2868 * schedule() atomically, we ignore that path for now.
2869 * Otherwise, whine if we are scheduling when we should not be.
2871 if (unlikely(in_atomic() && !current->exit_state)) {
2872 printk(KERN_ERR "BUG: scheduling while atomic: "
2874 current->comm, preempt_count(), current->pid);
2877 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2882 release_kernel_lock(prev);
2883 need_resched_nonpreemptible:
2887 * The idle thread is not allowed to schedule!
2888 * Remove this check after it has been exercised a bit.
2890 if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
2891 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
2895 schedstat_inc(rq, sched_cnt);
2896 now = sched_clock();
2897 if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
2898 run_time = now - prev->timestamp;
2899 if (unlikely((long long)(now - prev->timestamp) < 0))
2902 run_time = NS_MAX_SLEEP_AVG;
2905 * Tasks charged proportionately less run_time at high sleep_avg to
2906 * delay them losing their interactive status
2908 run_time /= (CURRENT_BONUS(prev) ? : 1);
2910 spin_lock_irq(&rq->lock);
2912 if (unlikely(prev->flags & PF_DEAD))
2913 prev->state = EXIT_DEAD;
2915 switch_count = &prev->nivcsw;
2916 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2917 switch_count = &prev->nvcsw;
2918 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
2919 unlikely(signal_pending(prev))))
2920 prev->state = TASK_RUNNING;
2922 if (prev->state == TASK_UNINTERRUPTIBLE)
2923 rq->nr_uninterruptible++;
2924 deactivate_task(prev, rq);
2928 cpu = smp_processor_id();
2929 if (unlikely(!rq->nr_running)) {
2930 idle_balance(cpu, rq);
2931 if (!rq->nr_running) {
2933 rq->expired_timestamp = 0;
2934 wake_sleeping_dependent(cpu);
2940 if (unlikely(!array->nr_active)) {
2942 * Switch the active and expired arrays.
2944 schedstat_inc(rq, sched_switch);
2945 rq->active = rq->expired;
2946 rq->expired = array;
2948 rq->expired_timestamp = 0;
2949 rq->best_expired_prio = MAX_PRIO;
2952 idx = sched_find_first_bit(array->bitmap);
2953 queue = array->queue + idx;
2954 next = list_entry(queue->next, task_t, run_list);
2956 if (!rt_task(next) && interactive_sleep(next->sleep_type)) {
2957 unsigned long long delta = now - next->timestamp;
2958 if (unlikely((long long)(now - next->timestamp) < 0))
2961 if (next->sleep_type == SLEEP_INTERACTIVE)
2962 delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
2964 array = next->array;
2965 new_prio = recalc_task_prio(next, next->timestamp + delta);
2967 if (unlikely(next->prio != new_prio)) {
2968 dequeue_task(next, array);
2969 next->prio = new_prio;
2970 enqueue_task(next, array);
2973 next->sleep_type = SLEEP_NORMAL;
2974 if (dependent_sleeper(cpu, rq, next))
2977 if (next == rq->idle)
2978 schedstat_inc(rq, sched_goidle);
2980 prefetch_stack(next);
2981 clear_tsk_need_resched(prev);
2982 rcu_qsctr_inc(task_cpu(prev));
2984 update_cpu_clock(prev, rq, now);
2986 prev->sleep_avg -= run_time;
2987 if ((long)prev->sleep_avg <= 0)
2988 prev->sleep_avg = 0;
2989 prev->timestamp = prev->last_ran = now;
2991 sched_info_switch(prev, next);
2992 if (likely(prev != next)) {
2993 next->timestamp = now;
2998 prepare_task_switch(rq, next);
2999 prev = context_switch(rq, prev, next);
3002 * this_rq must be evaluated again because prev may have moved
3003 * CPUs since it called schedule(), thus the 'rq' on its stack
3004 * frame will be invalid.
3006 finish_task_switch(this_rq(), prev);
3008 spin_unlock_irq(&rq->lock);
3011 if (unlikely(reacquire_kernel_lock(prev) < 0))
3012 goto need_resched_nonpreemptible;
3013 preempt_enable_no_resched();
3014 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3018 EXPORT_SYMBOL(schedule);
3020 #ifdef CONFIG_PREEMPT
3022 * this is is the entry point to schedule() from in-kernel preemption
3023 * off of preempt_enable. Kernel preemptions off return from interrupt
3024 * occur there and call schedule directly.
3026 asmlinkage void __sched preempt_schedule(void)
3028 struct thread_info *ti = current_thread_info();
3029 #ifdef CONFIG_PREEMPT_BKL
3030 struct task_struct *task = current;
3031 int saved_lock_depth;
3034 * If there is a non-zero preempt_count or interrupts are disabled,
3035 * we do not want to preempt the current task. Just return..
3037 if (unlikely(ti->preempt_count || irqs_disabled()))
3041 add_preempt_count(PREEMPT_ACTIVE);
3043 * We keep the big kernel semaphore locked, but we
3044 * clear ->lock_depth so that schedule() doesnt
3045 * auto-release the semaphore:
3047 #ifdef CONFIG_PREEMPT_BKL
3048 saved_lock_depth = task->lock_depth;
3049 task->lock_depth = -1;
3052 #ifdef CONFIG_PREEMPT_BKL
3053 task->lock_depth = saved_lock_depth;
3055 sub_preempt_count(PREEMPT_ACTIVE);
3057 /* we could miss a preemption opportunity between schedule and now */
3059 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3063 EXPORT_SYMBOL(preempt_schedule);
3066 * this is is the entry point to schedule() from kernel preemption
3067 * off of irq context.
3068 * Note, that this is called and return with irqs disabled. This will
3069 * protect us against recursive calling from irq.
3071 asmlinkage void __sched preempt_schedule_irq(void)
3073 struct thread_info *ti = current_thread_info();
3074 #ifdef CONFIG_PREEMPT_BKL
3075 struct task_struct *task = current;
3076 int saved_lock_depth;
3078 /* Catch callers which need to be fixed*/
3079 BUG_ON(ti->preempt_count || !irqs_disabled());
3082 add_preempt_count(PREEMPT_ACTIVE);
3084 * We keep the big kernel semaphore locked, but we
3085 * clear ->lock_depth so that schedule() doesnt
3086 * auto-release the semaphore:
3088 #ifdef CONFIG_PREEMPT_BKL
3089 saved_lock_depth = task->lock_depth;
3090 task->lock_depth = -1;
3094 local_irq_disable();
3095 #ifdef CONFIG_PREEMPT_BKL
3096 task->lock_depth = saved_lock_depth;
3098 sub_preempt_count(PREEMPT_ACTIVE);
3100 /* we could miss a preemption opportunity between schedule and now */
3102 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3106 #endif /* CONFIG_PREEMPT */
3108 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3111 task_t *p = curr->private;
3112 return try_to_wake_up(p, mode, sync);
3115 EXPORT_SYMBOL(default_wake_function);
3118 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3119 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3120 * number) then we wake all the non-exclusive tasks and one exclusive task.
3122 * There are circumstances in which we can try to wake a task which has already
3123 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3124 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3126 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3127 int nr_exclusive, int sync, void *key)
3129 struct list_head *tmp, *next;
3131 list_for_each_safe(tmp, next, &q->task_list) {
3134 curr = list_entry(tmp, wait_queue_t, task_list);
3135 flags = curr->flags;
3136 if (curr->func(curr, mode, sync, key) &&
3137 (flags & WQ_FLAG_EXCLUSIVE) &&
3144 * __wake_up - wake up threads blocked on a waitqueue.
3146 * @mode: which threads
3147 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3148 * @key: is directly passed to the wakeup function
3150 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3151 int nr_exclusive, void *key)
3153 unsigned long flags;
3155 spin_lock_irqsave(&q->lock, flags);
3156 __wake_up_common(q, mode, nr_exclusive, 0, key);
3157 spin_unlock_irqrestore(&q->lock, flags);
3160 EXPORT_SYMBOL(__wake_up);
3163 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3165 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3167 __wake_up_common(q, mode, 1, 0, NULL);
3171 * __wake_up_sync - wake up threads blocked on a waitqueue.
3173 * @mode: which threads
3174 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3176 * The sync wakeup differs that the waker knows that it will schedule
3177 * away soon, so while the target thread will be woken up, it will not
3178 * be migrated to another CPU - ie. the two threads are 'synchronized'
3179 * with each other. This can prevent needless bouncing between CPUs.
3181 * On UP it can prevent extra preemption.
3184 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3186 unsigned long flags;
3192 if (unlikely(!nr_exclusive))
3195 spin_lock_irqsave(&q->lock, flags);
3196 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3197 spin_unlock_irqrestore(&q->lock, flags);
3199 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3201 void fastcall complete(struct completion *x)
3203 unsigned long flags;
3205 spin_lock_irqsave(&x->wait.lock, flags);
3207 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3209 spin_unlock_irqrestore(&x->wait.lock, flags);
3211 EXPORT_SYMBOL(complete);
3213 void fastcall complete_all(struct completion *x)
3215 unsigned long flags;
3217 spin_lock_irqsave(&x->wait.lock, flags);
3218 x->done += UINT_MAX/2;
3219 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3221 spin_unlock_irqrestore(&x->wait.lock, flags);
3223 EXPORT_SYMBOL(complete_all);
3225 void fastcall __sched wait_for_completion(struct completion *x)
3228 spin_lock_irq(&x->wait.lock);
3230 DECLARE_WAITQUEUE(wait, current);
3232 wait.flags |= WQ_FLAG_EXCLUSIVE;
3233 __add_wait_queue_tail(&x->wait, &wait);
3235 __set_current_state(TASK_UNINTERRUPTIBLE);
3236 spin_unlock_irq(&x->wait.lock);
3238 spin_lock_irq(&x->wait.lock);
3240 __remove_wait_queue(&x->wait, &wait);
3243 spin_unlock_irq(&x->wait.lock);
3245 EXPORT_SYMBOL(wait_for_completion);
3247 unsigned long fastcall __sched
3248 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3252 spin_lock_irq(&x->wait.lock);
3254 DECLARE_WAITQUEUE(wait, current);
3256 wait.flags |= WQ_FLAG_EXCLUSIVE;
3257 __add_wait_queue_tail(&x->wait, &wait);
3259 __set_current_state(TASK_UNINTERRUPTIBLE);
3260 spin_unlock_irq(&x->wait.lock);
3261 timeout = schedule_timeout(timeout);
3262 spin_lock_irq(&x->wait.lock);
3264 __remove_wait_queue(&x->wait, &wait);
3268 __remove_wait_queue(&x->wait, &wait);
3272 spin_unlock_irq(&x->wait.lock);
3275 EXPORT_SYMBOL(wait_for_completion_timeout);
3277 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3283 spin_lock_irq(&x->wait.lock);
3285 DECLARE_WAITQUEUE(wait, current);
3287 wait.flags |= WQ_FLAG_EXCLUSIVE;
3288 __add_wait_queue_tail(&x->wait, &wait);
3290 if (signal_pending(current)) {
3292 __remove_wait_queue(&x->wait, &wait);
3295 __set_current_state(TASK_INTERRUPTIBLE);
3296 spin_unlock_irq(&x->wait.lock);
3298 spin_lock_irq(&x->wait.lock);
3300 __remove_wait_queue(&x->wait, &wait);
3304 spin_unlock_irq(&x->wait.lock);
3308 EXPORT_SYMBOL(wait_for_completion_interruptible);
3310 unsigned long fastcall __sched
3311 wait_for_completion_interruptible_timeout(struct completion *x,
3312 unsigned long timeout)
3316 spin_lock_irq(&x->wait.lock);
3318 DECLARE_WAITQUEUE(wait, current);
3320 wait.flags |= WQ_FLAG_EXCLUSIVE;
3321 __add_wait_queue_tail(&x->wait, &wait);
3323 if (signal_pending(current)) {
3324 timeout = -ERESTARTSYS;
3325 __remove_wait_queue(&x->wait, &wait);
3328 __set_current_state(TASK_INTERRUPTIBLE);
3329 spin_unlock_irq(&x->wait.lock);
3330 timeout = schedule_timeout(timeout);
3331 spin_lock_irq(&x->wait.lock);
3333 __remove_wait_queue(&x->wait, &wait);
3337 __remove_wait_queue(&x->wait, &wait);
3341 spin_unlock_irq(&x->wait.lock);
3344 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3347 #define SLEEP_ON_VAR \
3348 unsigned long flags; \
3349 wait_queue_t wait; \
3350 init_waitqueue_entry(&wait, current);
3352 #define SLEEP_ON_HEAD \
3353 spin_lock_irqsave(&q->lock,flags); \
3354 __add_wait_queue(q, &wait); \
3355 spin_unlock(&q->lock);
3357 #define SLEEP_ON_TAIL \
3358 spin_lock_irq(&q->lock); \
3359 __remove_wait_queue(q, &wait); \
3360 spin_unlock_irqrestore(&q->lock, flags);
3362 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
3366 current->state = TASK_INTERRUPTIBLE;
3373 EXPORT_SYMBOL(interruptible_sleep_on);
3375 long fastcall __sched
3376 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3380 current->state = TASK_INTERRUPTIBLE;
3383 timeout = schedule_timeout(timeout);
3389 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3391 void fastcall __sched sleep_on(wait_queue_head_t *q)
3395 current->state = TASK_UNINTERRUPTIBLE;
3402 EXPORT_SYMBOL(sleep_on);
3404 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3408 current->state = TASK_UNINTERRUPTIBLE;
3411 timeout = schedule_timeout(timeout);
3417 EXPORT_SYMBOL(sleep_on_timeout);
3419 void set_user_nice(task_t *p, long nice)
3421 unsigned long flags;
3422 prio_array_t *array;
3424 int old_prio, new_prio, delta;
3426 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3429 * We have to be careful, if called from sys_setpriority(),
3430 * the task might be in the middle of scheduling on another CPU.
3432 rq = task_rq_lock(p, &flags);
3434 * The RT priorities are set via sched_setscheduler(), but we still
3435 * allow the 'normal' nice value to be set - but as expected
3436 * it wont have any effect on scheduling until the task is
3437 * not SCHED_NORMAL/SCHED_BATCH:
3440 p->static_prio = NICE_TO_PRIO(nice);
3445 dequeue_task(p, array);
3448 new_prio = NICE_TO_PRIO(nice);
3449 delta = new_prio - old_prio;
3450 p->static_prio = NICE_TO_PRIO(nice);
3454 enqueue_task(p, array);
3456 * If the task increased its priority or is running and
3457 * lowered its priority, then reschedule its CPU:
3459 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3460 resched_task(rq->curr);
3463 task_rq_unlock(rq, &flags);
3466 EXPORT_SYMBOL(set_user_nice);
3469 * can_nice - check if a task can reduce its nice value
3473 int can_nice(const task_t *p, const int nice)
3475 /* convert nice value [19,-20] to rlimit style value [1,40] */
3476 int nice_rlim = 20 - nice;
3477 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
3478 capable(CAP_SYS_NICE));
3481 #ifdef __ARCH_WANT_SYS_NICE
3484 * sys_nice - change the priority of the current process.
3485 * @increment: priority increment
3487 * sys_setpriority is a more generic, but much slower function that
3488 * does similar things.
3490 asmlinkage long sys_nice(int increment)
3496 * Setpriority might change our priority at the same moment.
3497 * We don't have to worry. Conceptually one call occurs first
3498 * and we have a single winner.
3500 if (increment < -40)
3505 nice = PRIO_TO_NICE(current->static_prio) + increment;
3511 if (increment < 0 && !can_nice(current, nice))
3514 retval = security_task_setnice(current, nice);
3518 set_user_nice(current, nice);
3525 * task_prio - return the priority value of a given task.
3526 * @p: the task in question.
3528 * This is the priority value as seen by users in /proc.
3529 * RT tasks are offset by -200. Normal tasks are centered
3530 * around 0, value goes from -16 to +15.
3532 int task_prio(const task_t *p)
3534 return p->prio - MAX_RT_PRIO;
3538 * task_nice - return the nice value of a given task.
3539 * @p: the task in question.
3541 int task_nice(const task_t *p)
3543 return TASK_NICE(p);
3545 EXPORT_SYMBOL_GPL(task_nice);
3548 * idle_cpu - is a given cpu idle currently?
3549 * @cpu: the processor in question.
3551 int idle_cpu(int cpu)
3553 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
3557 * idle_task - return the idle task for a given cpu.
3558 * @cpu: the processor in question.
3560 task_t *idle_task(int cpu)
3562 return cpu_rq(cpu)->idle;
3566 * find_process_by_pid - find a process with a matching PID value.
3567 * @pid: the pid in question.
3569 static inline task_t *find_process_by_pid(pid_t pid)
3571 return pid ? find_task_by_pid(pid) : current;
3574 /* Actually do priority change: must hold rq lock. */
3575 static void __setscheduler(struct task_struct *p, int policy, int prio)
3579 p->rt_priority = prio;
3580 if (policy != SCHED_NORMAL && policy != SCHED_BATCH) {
3581 p->prio = MAX_RT_PRIO-1 - p->rt_priority;
3583 p->prio = p->static_prio;
3585 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
3587 if (policy == SCHED_BATCH)
3593 * sched_setscheduler - change the scheduling policy and/or RT priority of
3595 * @p: the task in question.
3596 * @policy: new policy.
3597 * @param: structure containing the new RT priority.
3599 int sched_setscheduler(struct task_struct *p, int policy,
3600 struct sched_param *param)
3603 int oldprio, oldpolicy = -1;
3604 prio_array_t *array;
3605 unsigned long flags;
3609 /* double check policy once rq lock held */
3611 policy = oldpolicy = p->policy;
3612 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
3613 policy != SCHED_NORMAL && policy != SCHED_BATCH)
3616 * Valid priorities for SCHED_FIFO and SCHED_RR are
3617 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
3620 if (param->sched_priority < 0 ||
3621 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3622 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3624 if ((policy == SCHED_NORMAL || policy == SCHED_BATCH)
3625 != (param->sched_priority == 0))
3629 * Allow unprivileged RT tasks to decrease priority:
3631 if (!capable(CAP_SYS_NICE)) {
3633 * can't change policy, except between SCHED_NORMAL
3636 if (((policy != SCHED_NORMAL && p->policy != SCHED_BATCH) &&
3637 (policy != SCHED_BATCH && p->policy != SCHED_NORMAL)) &&
3638 !p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
3640 /* can't increase priority */
3641 if ((policy != SCHED_NORMAL && policy != SCHED_BATCH) &&
3642 param->sched_priority > p->rt_priority &&
3643 param->sched_priority >
3644 p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
3646 /* can't change other user's priorities */
3647 if ((current->euid != p->euid) &&
3648 (current->euid != p->uid))
3652 retval = security_task_setscheduler(p, policy, param);
3656 * To be able to change p->policy safely, the apropriate
3657 * runqueue lock must be held.
3659 rq = task_rq_lock(p, &flags);
3660 /* recheck policy now with rq lock held */
3661 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3662 policy = oldpolicy = -1;
3663 task_rq_unlock(rq, &flags);
3668 deactivate_task(p, rq);
3670 __setscheduler(p, policy, param->sched_priority);
3672 __activate_task(p, rq);
3674 * Reschedule if we are currently running on this runqueue and
3675 * our priority decreased, or if we are not currently running on
3676 * this runqueue and our priority is higher than the current's
3678 if (task_running(rq, p)) {
3679 if (p->prio > oldprio)
3680 resched_task(rq->curr);
3681 } else if (TASK_PREEMPTS_CURR(p, rq))
3682 resched_task(rq->curr);
3684 task_rq_unlock(rq, &flags);
3687 EXPORT_SYMBOL_GPL(sched_setscheduler);
3690 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3693 struct sched_param lparam;
3694 struct task_struct *p;
3696 if (!param || pid < 0)
3698 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3700 read_lock_irq(&tasklist_lock);
3701 p = find_process_by_pid(pid);
3703 read_unlock_irq(&tasklist_lock);
3706 retval = sched_setscheduler(p, policy, &lparam);
3707 read_unlock_irq(&tasklist_lock);
3712 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3713 * @pid: the pid in question.
3714 * @policy: new policy.
3715 * @param: structure containing the new RT priority.
3717 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
3718 struct sched_param __user *param)
3720 /* negative values for policy are not valid */
3724 return do_sched_setscheduler(pid, policy, param);
3728 * sys_sched_setparam - set/change the RT priority of a thread
3729 * @pid: the pid in question.
3730 * @param: structure containing the new RT priority.
3732 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
3734 return do_sched_setscheduler(pid, -1, param);
3738 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3739 * @pid: the pid in question.
3741 asmlinkage long sys_sched_getscheduler(pid_t pid)
3743 int retval = -EINVAL;
3750 read_lock(&tasklist_lock);
3751 p = find_process_by_pid(pid);
3753 retval = security_task_getscheduler(p);
3757 read_unlock(&tasklist_lock);
3764 * sys_sched_getscheduler - get the RT priority of a thread
3765 * @pid: the pid in question.
3766 * @param: structure containing the RT priority.
3768 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
3770 struct sched_param lp;
3771 int retval = -EINVAL;
3774 if (!param || pid < 0)
3777 read_lock(&tasklist_lock);
3778 p = find_process_by_pid(pid);
3783 retval = security_task_getscheduler(p);
3787 lp.sched_priority = p->rt_priority;
3788 read_unlock(&tasklist_lock);
3791 * This one might sleep, we cannot do it with a spinlock held ...
3793 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3799 read_unlock(&tasklist_lock);
3803 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
3807 cpumask_t cpus_allowed;
3810 read_lock(&tasklist_lock);
3812 p = find_process_by_pid(pid);
3814 read_unlock(&tasklist_lock);
3815 unlock_cpu_hotplug();
3820 * It is not safe to call set_cpus_allowed with the
3821 * tasklist_lock held. We will bump the task_struct's
3822 * usage count and then drop tasklist_lock.
3825 read_unlock(&tasklist_lock);
3828 if ((current->euid != p->euid) && (current->euid != p->uid) &&
3829 !capable(CAP_SYS_NICE))
3832 retval = security_task_setscheduler(p, 0, NULL);
3836 cpus_allowed = cpuset_cpus_allowed(p);
3837 cpus_and(new_mask, new_mask, cpus_allowed);
3838 retval = set_cpus_allowed(p, new_mask);
3842 unlock_cpu_hotplug();
3846 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3847 cpumask_t *new_mask)
3849 if (len < sizeof(cpumask_t)) {
3850 memset(new_mask, 0, sizeof(cpumask_t));
3851 } else if (len > sizeof(cpumask_t)) {
3852 len = sizeof(cpumask_t);
3854 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3858 * sys_sched_setaffinity - set the cpu affinity of a process
3859 * @pid: pid of the process
3860 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3861 * @user_mask_ptr: user-space pointer to the new cpu mask
3863 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
3864 unsigned long __user *user_mask_ptr)
3869 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
3873 return sched_setaffinity(pid, new_mask);
3877 * Represents all cpu's present in the system
3878 * In systems capable of hotplug, this map could dynamically grow
3879 * as new cpu's are detected in the system via any platform specific
3880 * method, such as ACPI for e.g.
3883 cpumask_t cpu_present_map __read_mostly;
3884 EXPORT_SYMBOL(cpu_present_map);
3887 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
3888 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
3891 long sched_getaffinity(pid_t pid, cpumask_t *mask)
3897 read_lock(&tasklist_lock);
3900 p = find_process_by_pid(pid);
3904 retval = security_task_getscheduler(p);
3908 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
3911 read_unlock(&tasklist_lock);
3912 unlock_cpu_hotplug();
3920 * sys_sched_getaffinity - get the cpu affinity of a process
3921 * @pid: pid of the process
3922 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3923 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3925 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
3926 unsigned long __user *user_mask_ptr)
3931 if (len < sizeof(cpumask_t))
3934 ret = sched_getaffinity(pid, &mask);
3938 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
3941 return sizeof(cpumask_t);
3945 * sys_sched_yield - yield the current processor to other threads.
3947 * this function yields the current CPU by moving the calling thread
3948 * to the expired array. If there are no other threads running on this
3949 * CPU then this function will return.
3951 asmlinkage long sys_sched_yield(void)
3953 runqueue_t *rq = this_rq_lock();
3954 prio_array_t *array = current->array;
3955 prio_array_t *target = rq->expired;
3957 schedstat_inc(rq, yld_cnt);
3959 * We implement yielding by moving the task into the expired
3962 * (special rule: RT tasks will just roundrobin in the active
3965 if (rt_task(current))
3966 target = rq->active;
3968 if (array->nr_active == 1) {
3969 schedstat_inc(rq, yld_act_empty);
3970 if (!rq->expired->nr_active)
3971 schedstat_inc(rq, yld_both_empty);
3972 } else if (!rq->expired->nr_active)
3973 schedstat_inc(rq, yld_exp_empty);
3975 if (array != target) {
3976 dequeue_task(current, array);
3977 enqueue_task(current, target);
3980 * requeue_task is cheaper so perform that if possible.
3982 requeue_task(current, array);
3985 * Since we are going to call schedule() anyway, there's
3986 * no need to preempt or enable interrupts:
3988 __release(rq->lock);
3989 _raw_spin_unlock(&rq->lock);
3990 preempt_enable_no_resched();
3997 static inline void __cond_resched(void)
3999 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4000 __might_sleep(__FILE__, __LINE__);
4003 * The BKS might be reacquired before we have dropped
4004 * PREEMPT_ACTIVE, which could trigger a second
4005 * cond_resched() call.
4007 if (unlikely(preempt_count()))
4009 if (unlikely(system_state != SYSTEM_RUNNING))
4012 add_preempt_count(PREEMPT_ACTIVE);
4014 sub_preempt_count(PREEMPT_ACTIVE);
4015 } while (need_resched());
4018 int __sched cond_resched(void)
4020 if (need_resched()) {
4027 EXPORT_SYMBOL(cond_resched);
4030 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4031 * call schedule, and on return reacquire the lock.
4033 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4034 * operations here to prevent schedule() from being called twice (once via
4035 * spin_unlock(), once by hand).
4037 int cond_resched_lock(spinlock_t *lock)
4041 if (need_lockbreak(lock)) {
4047 if (need_resched()) {
4048 _raw_spin_unlock(lock);
4049 preempt_enable_no_resched();
4057 EXPORT_SYMBOL(cond_resched_lock);
4059 int __sched cond_resched_softirq(void)
4061 BUG_ON(!in_softirq());
4063 if (need_resched()) {
4064 __local_bh_enable();
4072 EXPORT_SYMBOL(cond_resched_softirq);
4076 * yield - yield the current processor to other threads.
4078 * this is a shortcut for kernel-space yielding - it marks the
4079 * thread runnable and calls sys_sched_yield().
4081 void __sched yield(void)
4083 set_current_state(TASK_RUNNING);
4087 EXPORT_SYMBOL(yield);
4090 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4091 * that process accounting knows that this is a task in IO wait state.
4093 * But don't do that if it is a deliberate, throttling IO wait (this task
4094 * has set its backing_dev_info: the queue against which it should throttle)
4096 void __sched io_schedule(void)
4098 struct runqueue *rq = &__raw_get_cpu_var(runqueues);
4100 atomic_inc(&rq->nr_iowait);
4102 atomic_dec(&rq->nr_iowait);
4105 EXPORT_SYMBOL(io_schedule);
4107 long __sched io_schedule_timeout(long timeout)
4109 struct runqueue *rq = &__raw_get_cpu_var(runqueues);
4112 atomic_inc(&rq->nr_iowait);
4113 ret = schedule_timeout(timeout);
4114 atomic_dec(&rq->nr_iowait);
4119 * sys_sched_get_priority_max - return maximum RT priority.
4120 * @policy: scheduling class.
4122 * this syscall returns the maximum rt_priority that can be used
4123 * by a given scheduling class.
4125 asmlinkage long sys_sched_get_priority_max(int policy)
4132 ret = MAX_USER_RT_PRIO-1;
4143 * sys_sched_get_priority_min - return minimum RT priority.
4144 * @policy: scheduling class.
4146 * this syscall returns the minimum rt_priority that can be used
4147 * by a given scheduling class.
4149 asmlinkage long sys_sched_get_priority_min(int policy)
4166 * sys_sched_rr_get_interval - return the default timeslice of a process.
4167 * @pid: pid of the process.
4168 * @interval: userspace pointer to the timeslice value.
4170 * this syscall writes the default timeslice value of a given process
4171 * into the user-space timespec buffer. A value of '0' means infinity.
4174 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4176 int retval = -EINVAL;
4184 read_lock(&tasklist_lock);
4185 p = find_process_by_pid(pid);
4189 retval = security_task_getscheduler(p);
4193 jiffies_to_timespec(p->policy == SCHED_FIFO ?
4194 0 : task_timeslice(p), &t);
4195 read_unlock(&tasklist_lock);
4196 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4200 read_unlock(&tasklist_lock);
4204 static inline struct task_struct *eldest_child(struct task_struct *p)
4206 if (list_empty(&p->children)) return NULL;
4207 return list_entry(p->children.next,struct task_struct,sibling);
4210 static inline struct task_struct *older_sibling(struct task_struct *p)
4212 if (p->sibling.prev==&p->parent->children) return NULL;
4213 return list_entry(p->sibling.prev,struct task_struct,sibling);
4216 static inline struct task_struct *younger_sibling(struct task_struct *p)
4218 if (p->sibling.next==&p->parent->children) return NULL;
4219 return list_entry(p->sibling.next,struct task_struct,sibling);
4222 static void show_task(task_t *p)
4226 unsigned long free = 0;
4227 static const char *stat_nam[] = { "R", "S", "D", "T", "t", "Z", "X" };
4229 printk("%-13.13s ", p->comm);
4230 state = p->state ? __ffs(p->state) + 1 : 0;
4231 if (state < ARRAY_SIZE(stat_nam))
4232 printk(stat_nam[state]);
4235 #if (BITS_PER_LONG == 32)
4236 if (state == TASK_RUNNING)
4237 printk(" running ");
4239 printk(" %08lX ", thread_saved_pc(p));
4241 if (state == TASK_RUNNING)
4242 printk(" running task ");
4244 printk(" %016lx ", thread_saved_pc(p));
4246 #ifdef CONFIG_DEBUG_STACK_USAGE
4248 unsigned long *n = end_of_stack(p);
4251 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4254 printk("%5lu %5d %6d ", free, p->pid, p->parent->pid);
4255 if ((relative = eldest_child(p)))
4256 printk("%5d ", relative->pid);
4259 if ((relative = younger_sibling(p)))
4260 printk("%7d", relative->pid);
4263 if ((relative = older_sibling(p)))
4264 printk(" %5d", relative->pid);
4268 printk(" (L-TLB)\n");
4270 printk(" (NOTLB)\n");
4272 if (state != TASK_RUNNING)
4273 show_stack(p, NULL);
4276 void show_state(void)
4280 #if (BITS_PER_LONG == 32)
4283 printk(" task PC pid father child younger older\n");
4287 printk(" task PC pid father child younger older\n");
4289 read_lock(&tasklist_lock);
4290 do_each_thread(g, p) {
4292 * reset the NMI-timeout, listing all files on a slow
4293 * console might take alot of time:
4295 touch_nmi_watchdog();
4297 } while_each_thread(g, p);
4299 read_unlock(&tasklist_lock);
4300 mutex_debug_show_all_locks();
4304 * init_idle - set up an idle thread for a given CPU
4305 * @idle: task in question
4306 * @cpu: cpu the idle task belongs to
4308 * NOTE: this function does not set the idle thread's NEED_RESCHED
4309 * flag, to make booting more robust.
4311 void __devinit init_idle(task_t *idle, int cpu)
4313 runqueue_t *rq = cpu_rq(cpu);
4314 unsigned long flags;
4316 idle->timestamp = sched_clock();
4317 idle->sleep_avg = 0;
4319 idle->prio = MAX_PRIO;
4320 idle->state = TASK_RUNNING;
4321 idle->cpus_allowed = cpumask_of_cpu(cpu);
4322 set_task_cpu(idle, cpu);
4324 spin_lock_irqsave(&rq->lock, flags);
4325 rq->curr = rq->idle = idle;
4326 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4329 spin_unlock_irqrestore(&rq->lock, flags);
4331 /* Set the preempt count _outside_ the spinlocks! */
4332 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4333 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4335 task_thread_info(idle)->preempt_count = 0;
4340 * In a system that switches off the HZ timer nohz_cpu_mask
4341 * indicates which cpus entered this state. This is used
4342 * in the rcu update to wait only for active cpus. For system
4343 * which do not switch off the HZ timer nohz_cpu_mask should
4344 * always be CPU_MASK_NONE.
4346 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4350 * This is how migration works:
4352 * 1) we queue a migration_req_t structure in the source CPU's
4353 * runqueue and wake up that CPU's migration thread.
4354 * 2) we down() the locked semaphore => thread blocks.
4355 * 3) migration thread wakes up (implicitly it forces the migrated
4356 * thread off the CPU)
4357 * 4) it gets the migration request and checks whether the migrated
4358 * task is still in the wrong runqueue.
4359 * 5) if it's in the wrong runqueue then the migration thread removes
4360 * it and puts it into the right queue.
4361 * 6) migration thread up()s the semaphore.
4362 * 7) we wake up and the migration is done.
4366 * Change a given task's CPU affinity. Migrate the thread to a
4367 * proper CPU and schedule it away if the CPU it's executing on
4368 * is removed from the allowed bitmask.
4370 * NOTE: the caller must have a valid reference to the task, the
4371 * task must not exit() & deallocate itself prematurely. The
4372 * call is not atomic; no spinlocks may be held.
4374 int set_cpus_allowed(task_t *p, cpumask_t new_mask)
4376 unsigned long flags;
4378 migration_req_t req;
4381 rq = task_rq_lock(p, &flags);
4382 if (!cpus_intersects(new_mask, cpu_online_map)) {
4387 p->cpus_allowed = new_mask;
4388 /* Can the task run on the task's current CPU? If so, we're done */
4389 if (cpu_isset(task_cpu(p), new_mask))
4392 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4393 /* Need help from migration thread: drop lock and wait. */
4394 task_rq_unlock(rq, &flags);
4395 wake_up_process(rq->migration_thread);
4396 wait_for_completion(&req.done);
4397 tlb_migrate_finish(p->mm);
4401 task_rq_unlock(rq, &flags);
4405 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4408 * Move (not current) task off this cpu, onto dest cpu. We're doing
4409 * this because either it can't run here any more (set_cpus_allowed()
4410 * away from this CPU, or CPU going down), or because we're
4411 * attempting to rebalance this task on exec (sched_exec).
4413 * So we race with normal scheduler movements, but that's OK, as long
4414 * as the task is no longer on this CPU.
4416 * Returns non-zero if task was successfully migrated.
4418 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4420 runqueue_t *rq_dest, *rq_src;
4423 if (unlikely(cpu_is_offline(dest_cpu)))
4426 rq_src = cpu_rq(src_cpu);
4427 rq_dest = cpu_rq(dest_cpu);
4429 double_rq_lock(rq_src, rq_dest);
4430 /* Already moved. */
4431 if (task_cpu(p) != src_cpu)
4433 /* Affinity changed (again). */
4434 if (!cpu_isset(dest_cpu, p->cpus_allowed))
4437 set_task_cpu(p, dest_cpu);
4440 * Sync timestamp with rq_dest's before activating.
4441 * The same thing could be achieved by doing this step
4442 * afterwards, and pretending it was a local activate.
4443 * This way is cleaner and logically correct.
4445 p->timestamp = p->timestamp - rq_src->timestamp_last_tick
4446 + rq_dest->timestamp_last_tick;
4447 deactivate_task(p, rq_src);
4448 activate_task(p, rq_dest, 0);
4449 if (TASK_PREEMPTS_CURR(p, rq_dest))
4450 resched_task(rq_dest->curr);
4454 double_rq_unlock(rq_src, rq_dest);
4459 * migration_thread - this is a highprio system thread that performs
4460 * thread migration by bumping thread off CPU then 'pushing' onto
4463 static int migration_thread(void *data)
4466 int cpu = (long)data;
4469 BUG_ON(rq->migration_thread != current);
4471 set_current_state(TASK_INTERRUPTIBLE);
4472 while (!kthread_should_stop()) {
4473 struct list_head *head;
4474 migration_req_t *req;
4478 spin_lock_irq(&rq->lock);
4480 if (cpu_is_offline(cpu)) {
4481 spin_unlock_irq(&rq->lock);
4485 if (rq->active_balance) {
4486 active_load_balance(rq, cpu);
4487 rq->active_balance = 0;
4490 head = &rq->migration_queue;
4492 if (list_empty(head)) {
4493 spin_unlock_irq(&rq->lock);
4495 set_current_state(TASK_INTERRUPTIBLE);
4498 req = list_entry(head->next, migration_req_t, list);
4499 list_del_init(head->next);
4501 spin_unlock(&rq->lock);
4502 __migrate_task(req->task, cpu, req->dest_cpu);
4505 complete(&req->done);
4507 __set_current_state(TASK_RUNNING);
4511 /* Wait for kthread_stop */
4512 set_current_state(TASK_INTERRUPTIBLE);
4513 while (!kthread_should_stop()) {
4515 set_current_state(TASK_INTERRUPTIBLE);
4517 __set_current_state(TASK_RUNNING);
4521 #ifdef CONFIG_HOTPLUG_CPU
4522 /* Figure out where task on dead CPU should go, use force if neccessary. */
4523 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *tsk)
4526 unsigned long flags;
4532 mask = node_to_cpumask(cpu_to_node(dead_cpu));
4533 cpus_and(mask, mask, tsk->cpus_allowed);
4534 dest_cpu = any_online_cpu(mask);
4536 /* On any allowed CPU? */
4537 if (dest_cpu == NR_CPUS)
4538 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4540 /* No more Mr. Nice Guy. */
4541 if (dest_cpu == NR_CPUS) {
4542 rq = task_rq_lock(tsk, &flags);
4543 cpus_setall(tsk->cpus_allowed);
4544 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4545 task_rq_unlock(rq, &flags);
4548 * Don't tell them about moving exiting tasks or
4549 * kernel threads (both mm NULL), since they never
4552 if (tsk->mm && printk_ratelimit())
4553 printk(KERN_INFO "process %d (%s) no "
4554 "longer affine to cpu%d\n",
4555 tsk->pid, tsk->comm, dead_cpu);
4557 if (!__migrate_task(tsk, dead_cpu, dest_cpu))
4562 * While a dead CPU has no uninterruptible tasks queued at this point,
4563 * it might still have a nonzero ->nr_uninterruptible counter, because
4564 * for performance reasons the counter is not stricly tracking tasks to
4565 * their home CPUs. So we just add the counter to another CPU's counter,
4566 * to keep the global sum constant after CPU-down:
4568 static void migrate_nr_uninterruptible(runqueue_t *rq_src)
4570 runqueue_t *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
4571 unsigned long flags;
4573 local_irq_save(flags);
4574 double_rq_lock(rq_src, rq_dest);
4575 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
4576 rq_src->nr_uninterruptible = 0;
4577 double_rq_unlock(rq_src, rq_dest);
4578 local_irq_restore(flags);
4581 /* Run through task list and migrate tasks from the dead cpu. */
4582 static void migrate_live_tasks(int src_cpu)
4584 struct task_struct *tsk, *t;
4586 write_lock_irq(&tasklist_lock);
4588 do_each_thread(t, tsk) {
4592 if (task_cpu(tsk) == src_cpu)
4593 move_task_off_dead_cpu(src_cpu, tsk);
4594 } while_each_thread(t, tsk);
4596 write_unlock_irq(&tasklist_lock);
4599 /* Schedules idle task to be the next runnable task on current CPU.
4600 * It does so by boosting its priority to highest possible and adding it to
4601 * the _front_ of runqueue. Used by CPU offline code.
4603 void sched_idle_next(void)
4605 int cpu = smp_processor_id();
4606 runqueue_t *rq = this_rq();
4607 struct task_struct *p = rq->idle;
4608 unsigned long flags;
4610 /* cpu has to be offline */
4611 BUG_ON(cpu_online(cpu));
4613 /* Strictly not necessary since rest of the CPUs are stopped by now
4614 * and interrupts disabled on current cpu.
4616 spin_lock_irqsave(&rq->lock, flags);
4618 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4619 /* Add idle task to _front_ of it's priority queue */
4620 __activate_idle_task(p, rq);
4622 spin_unlock_irqrestore(&rq->lock, flags);
4625 /* Ensures that the idle task is using init_mm right before its cpu goes
4628 void idle_task_exit(void)
4630 struct mm_struct *mm = current->active_mm;
4632 BUG_ON(cpu_online(smp_processor_id()));
4635 switch_mm(mm, &init_mm, current);
4639 static void migrate_dead(unsigned int dead_cpu, task_t *tsk)
4641 struct runqueue *rq = cpu_rq(dead_cpu);
4643 /* Must be exiting, otherwise would be on tasklist. */
4644 BUG_ON(tsk->exit_state != EXIT_ZOMBIE && tsk->exit_state != EXIT_DEAD);
4646 /* Cannot have done final schedule yet: would have vanished. */
4647 BUG_ON(tsk->flags & PF_DEAD);
4649 get_task_struct(tsk);
4652 * Drop lock around migration; if someone else moves it,
4653 * that's OK. No task can be added to this CPU, so iteration is
4656 spin_unlock_irq(&rq->lock);
4657 move_task_off_dead_cpu(dead_cpu, tsk);
4658 spin_lock_irq(&rq->lock);
4660 put_task_struct(tsk);
4663 /* release_task() removes task from tasklist, so we won't find dead tasks. */
4664 static void migrate_dead_tasks(unsigned int dead_cpu)
4667 struct runqueue *rq = cpu_rq(dead_cpu);
4669 for (arr = 0; arr < 2; arr++) {
4670 for (i = 0; i < MAX_PRIO; i++) {
4671 struct list_head *list = &rq->arrays[arr].queue[i];
4672 while (!list_empty(list))
4673 migrate_dead(dead_cpu,
4674 list_entry(list->next, task_t,
4679 #endif /* CONFIG_HOTPLUG_CPU */
4682 * migration_call - callback that gets triggered when a CPU is added.
4683 * Here we can start up the necessary migration thread for the new CPU.
4685 static int __cpuinit migration_call(struct notifier_block *nfb,
4686 unsigned long action,
4689 int cpu = (long)hcpu;
4690 struct task_struct *p;
4691 struct runqueue *rq;
4692 unsigned long flags;
4695 case CPU_UP_PREPARE:
4696 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
4699 p->flags |= PF_NOFREEZE;
4700 kthread_bind(p, cpu);
4701 /* Must be high prio: stop_machine expects to yield to it. */
4702 rq = task_rq_lock(p, &flags);
4703 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4704 task_rq_unlock(rq, &flags);
4705 cpu_rq(cpu)->migration_thread = p;
4708 /* Strictly unneccessary, as first user will wake it. */
4709 wake_up_process(cpu_rq(cpu)->migration_thread);
4711 #ifdef CONFIG_HOTPLUG_CPU
4712 case CPU_UP_CANCELED:
4713 if (!cpu_rq(cpu)->migration_thread)
4715 /* Unbind it from offline cpu so it can run. Fall thru. */
4716 kthread_bind(cpu_rq(cpu)->migration_thread,
4717 any_online_cpu(cpu_online_map));
4718 kthread_stop(cpu_rq(cpu)->migration_thread);
4719 cpu_rq(cpu)->migration_thread = NULL;
4722 migrate_live_tasks(cpu);
4724 kthread_stop(rq->migration_thread);
4725 rq->migration_thread = NULL;
4726 /* Idle task back to normal (off runqueue, low prio) */
4727 rq = task_rq_lock(rq->idle, &flags);
4728 deactivate_task(rq->idle, rq);
4729 rq->idle->static_prio = MAX_PRIO;
4730 __setscheduler(rq->idle, SCHED_NORMAL, 0);
4731 migrate_dead_tasks(cpu);
4732 task_rq_unlock(rq, &flags);
4733 migrate_nr_uninterruptible(rq);
4734 BUG_ON(rq->nr_running != 0);
4736 /* No need to migrate the tasks: it was best-effort if
4737 * they didn't do lock_cpu_hotplug(). Just wake up
4738 * the requestors. */
4739 spin_lock_irq(&rq->lock);
4740 while (!list_empty(&rq->migration_queue)) {
4741 migration_req_t *req;
4742 req = list_entry(rq->migration_queue.next,
4743 migration_req_t, list);
4744 list_del_init(&req->list);
4745 complete(&req->done);
4747 spin_unlock_irq(&rq->lock);
4754 /* Register at highest priority so that task migration (migrate_all_tasks)
4755 * happens before everything else.
4757 static struct notifier_block __cpuinitdata migration_notifier = {
4758 .notifier_call = migration_call,
4762 int __init migration_init(void)
4764 void *cpu = (void *)(long)smp_processor_id();
4765 /* Start one for boot CPU. */
4766 migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
4767 migration_call(&migration_notifier, CPU_ONLINE, cpu);
4768 register_cpu_notifier(&migration_notifier);
4774 #undef SCHED_DOMAIN_DEBUG
4775 #ifdef SCHED_DOMAIN_DEBUG
4776 static void sched_domain_debug(struct sched_domain *sd, int cpu)
4781 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
4785 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
4790 struct sched_group *group = sd->groups;
4791 cpumask_t groupmask;
4793 cpumask_scnprintf(str, NR_CPUS, sd->span);
4794 cpus_clear(groupmask);
4797 for (i = 0; i < level + 1; i++)
4799 printk("domain %d: ", level);
4801 if (!(sd->flags & SD_LOAD_BALANCE)) {
4802 printk("does not load-balance\n");
4804 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent");
4808 printk("span %s\n", str);
4810 if (!cpu_isset(cpu, sd->span))
4811 printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
4812 if (!cpu_isset(cpu, group->cpumask))
4813 printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
4816 for (i = 0; i < level + 2; i++)
4822 printk(KERN_ERR "ERROR: group is NULL\n");
4826 if (!group->cpu_power) {
4828 printk(KERN_ERR "ERROR: domain->cpu_power not set\n");
4831 if (!cpus_weight(group->cpumask)) {
4833 printk(KERN_ERR "ERROR: empty group\n");
4836 if (cpus_intersects(groupmask, group->cpumask)) {
4838 printk(KERN_ERR "ERROR: repeated CPUs\n");
4841 cpus_or(groupmask, groupmask, group->cpumask);
4843 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
4846 group = group->next;
4847 } while (group != sd->groups);
4850 if (!cpus_equal(sd->span, groupmask))
4851 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
4857 if (!cpus_subset(groupmask, sd->span))
4858 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
4864 #define sched_domain_debug(sd, cpu) {}
4867 static int sd_degenerate(struct sched_domain *sd)
4869 if (cpus_weight(sd->span) == 1)
4872 /* Following flags need at least 2 groups */
4873 if (sd->flags & (SD_LOAD_BALANCE |
4874 SD_BALANCE_NEWIDLE |
4877 if (sd->groups != sd->groups->next)
4881 /* Following flags don't use groups */
4882 if (sd->flags & (SD_WAKE_IDLE |
4890 static int sd_parent_degenerate(struct sched_domain *sd,
4891 struct sched_domain *parent)
4893 unsigned long cflags = sd->flags, pflags = parent->flags;
4895 if (sd_degenerate(parent))
4898 if (!cpus_equal(sd->span, parent->span))
4901 /* Does parent contain flags not in child? */
4902 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
4903 if (cflags & SD_WAKE_AFFINE)
4904 pflags &= ~SD_WAKE_BALANCE;
4905 /* Flags needing groups don't count if only 1 group in parent */
4906 if (parent->groups == parent->groups->next) {
4907 pflags &= ~(SD_LOAD_BALANCE |
4908 SD_BALANCE_NEWIDLE |
4912 if (~cflags & pflags)
4919 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
4920 * hold the hotplug lock.
4922 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
4924 runqueue_t *rq = cpu_rq(cpu);
4925 struct sched_domain *tmp;
4927 /* Remove the sched domains which do not contribute to scheduling. */
4928 for (tmp = sd; tmp; tmp = tmp->parent) {
4929 struct sched_domain *parent = tmp->parent;
4932 if (sd_parent_degenerate(tmp, parent))
4933 tmp->parent = parent->parent;
4936 if (sd && sd_degenerate(sd))
4939 sched_domain_debug(sd, cpu);
4941 rcu_assign_pointer(rq->sd, sd);
4944 /* cpus with isolated domains */
4945 static cpumask_t __devinitdata cpu_isolated_map = CPU_MASK_NONE;
4947 /* Setup the mask of cpus configured for isolated domains */
4948 static int __init isolated_cpu_setup(char *str)
4950 int ints[NR_CPUS], i;
4952 str = get_options(str, ARRAY_SIZE(ints), ints);
4953 cpus_clear(cpu_isolated_map);
4954 for (i = 1; i <= ints[0]; i++)
4955 if (ints[i] < NR_CPUS)
4956 cpu_set(ints[i], cpu_isolated_map);
4960 __setup ("isolcpus=", isolated_cpu_setup);
4963 * init_sched_build_groups takes an array of groups, the cpumask we wish
4964 * to span, and a pointer to a function which identifies what group a CPU
4965 * belongs to. The return value of group_fn must be a valid index into the
4966 * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
4967 * keep track of groups covered with a cpumask_t).
4969 * init_sched_build_groups will build a circular linked list of the groups
4970 * covered by the given span, and will set each group's ->cpumask correctly,
4971 * and ->cpu_power to 0.
4973 static void init_sched_build_groups(struct sched_group groups[], cpumask_t span,
4974 int (*group_fn)(int cpu))
4976 struct sched_group *first = NULL, *last = NULL;
4977 cpumask_t covered = CPU_MASK_NONE;
4980 for_each_cpu_mask(i, span) {
4981 int group = group_fn(i);
4982 struct sched_group *sg = &groups[group];
4985 if (cpu_isset(i, covered))
4988 sg->cpumask = CPU_MASK_NONE;
4991 for_each_cpu_mask(j, span) {
4992 if (group_fn(j) != group)
4995 cpu_set(j, covered);
4996 cpu_set(j, sg->cpumask);
5007 #define SD_NODES_PER_DOMAIN 16
5010 * Self-tuning task migration cost measurement between source and target CPUs.
5012 * This is done by measuring the cost of manipulating buffers of varying
5013 * sizes. For a given buffer-size here are the steps that are taken:
5015 * 1) the source CPU reads+dirties a shared buffer
5016 * 2) the target CPU reads+dirties the same shared buffer
5018 * We measure how long they take, in the following 4 scenarios:
5020 * - source: CPU1, target: CPU2 | cost1
5021 * - source: CPU2, target: CPU1 | cost2
5022 * - source: CPU1, target: CPU1 | cost3
5023 * - source: CPU2, target: CPU2 | cost4
5025 * We then calculate the cost3+cost4-cost1-cost2 difference - this is
5026 * the cost of migration.
5028 * We then start off from a small buffer-size and iterate up to larger
5029 * buffer sizes, in 5% steps - measuring each buffer-size separately, and
5030 * doing a maximum search for the cost. (The maximum cost for a migration
5031 * normally occurs when the working set size is around the effective cache
5034 #define SEARCH_SCOPE 2
5035 #define MIN_CACHE_SIZE (64*1024U)
5036 #define DEFAULT_CACHE_SIZE (5*1024*1024U)
5037 #define ITERATIONS 1
5038 #define SIZE_THRESH 130
5039 #define COST_THRESH 130
5042 * The migration cost is a function of 'domain distance'. Domain
5043 * distance is the number of steps a CPU has to iterate down its
5044 * domain tree to share a domain with the other CPU. The farther
5045 * two CPUs are from each other, the larger the distance gets.
5047 * Note that we use the distance only to cache measurement results,
5048 * the distance value is not used numerically otherwise. When two
5049 * CPUs have the same distance it is assumed that the migration
5050 * cost is the same. (this is a simplification but quite practical)
5052 #define MAX_DOMAIN_DISTANCE 32
5054 static unsigned long long migration_cost[MAX_DOMAIN_DISTANCE] =
5055 { [ 0 ... MAX_DOMAIN_DISTANCE-1 ] =
5057 * Architectures may override the migration cost and thus avoid
5058 * boot-time calibration. Unit is nanoseconds. Mostly useful for
5059 * virtualized hardware:
5061 #ifdef CONFIG_DEFAULT_MIGRATION_COST
5062 CONFIG_DEFAULT_MIGRATION_COST
5069 * Allow override of migration cost - in units of microseconds.
5070 * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost
5071 * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs:
5073 static int __init migration_cost_setup(char *str)
5075 int ints[MAX_DOMAIN_DISTANCE+1], i;
5077 str = get_options(str, ARRAY_SIZE(ints), ints);
5079 printk("#ints: %d\n", ints[0]);
5080 for (i = 1; i <= ints[0]; i++) {
5081 migration_cost[i-1] = (unsigned long long)ints[i]*1000;
5082 printk("migration_cost[%d]: %Ld\n", i-1, migration_cost[i-1]);
5087 __setup ("migration_cost=", migration_cost_setup);
5090 * Global multiplier (divisor) for migration-cutoff values,
5091 * in percentiles. E.g. use a value of 150 to get 1.5 times
5092 * longer cache-hot cutoff times.
5094 * (We scale it from 100 to 128 to long long handling easier.)
5097 #define MIGRATION_FACTOR_SCALE 128
5099 static unsigned int migration_factor = MIGRATION_FACTOR_SCALE;
5101 static int __init setup_migration_factor(char *str)
5103 get_option(&str, &migration_factor);
5104 migration_factor = migration_factor * MIGRATION_FACTOR_SCALE / 100;
5108 __setup("migration_factor=", setup_migration_factor);
5111 * Estimated distance of two CPUs, measured via the number of domains
5112 * we have to pass for the two CPUs to be in the same span:
5114 static unsigned long domain_distance(int cpu1, int cpu2)
5116 unsigned long distance = 0;
5117 struct sched_domain *sd;
5119 for_each_domain(cpu1, sd) {
5120 WARN_ON(!cpu_isset(cpu1, sd->span));
5121 if (cpu_isset(cpu2, sd->span))
5125 if (distance >= MAX_DOMAIN_DISTANCE) {
5127 distance = MAX_DOMAIN_DISTANCE-1;
5133 static unsigned int migration_debug;
5135 static int __init setup_migration_debug(char *str)
5137 get_option(&str, &migration_debug);
5141 __setup("migration_debug=", setup_migration_debug);
5144 * Maximum cache-size that the scheduler should try to measure.
5145 * Architectures with larger caches should tune this up during
5146 * bootup. Gets used in the domain-setup code (i.e. during SMP
5149 unsigned int max_cache_size;
5151 static int __init setup_max_cache_size(char *str)
5153 get_option(&str, &max_cache_size);
5157 __setup("max_cache_size=", setup_max_cache_size);
5160 * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This
5161 * is the operation that is timed, so we try to generate unpredictable
5162 * cachemisses that still end up filling the L2 cache:
5164 static void touch_cache(void *__cache, unsigned long __size)
5166 unsigned long size = __size/sizeof(long), chunk1 = size/3,
5168 unsigned long *cache = __cache;
5171 for (i = 0; i < size/6; i += 8) {
5174 case 1: cache[size-1-i]++;
5175 case 2: cache[chunk1-i]++;
5176 case 3: cache[chunk1+i]++;
5177 case 4: cache[chunk2-i]++;
5178 case 5: cache[chunk2+i]++;
5184 * Measure the cache-cost of one task migration. Returns in units of nsec.
5186 static unsigned long long measure_one(void *cache, unsigned long size,
5187 int source, int target)
5189 cpumask_t mask, saved_mask;
5190 unsigned long long t0, t1, t2, t3, cost;
5192 saved_mask = current->cpus_allowed;
5195 * Flush source caches to RAM and invalidate them:
5200 * Migrate to the source CPU:
5202 mask = cpumask_of_cpu(source);
5203 set_cpus_allowed(current, mask);
5204 WARN_ON(smp_processor_id() != source);
5207 * Dirty the working set:
5210 touch_cache(cache, size);
5214 * Migrate to the target CPU, dirty the L2 cache and access
5215 * the shared buffer. (which represents the working set
5216 * of a migrated task.)
5218 mask = cpumask_of_cpu(target);
5219 set_cpus_allowed(current, mask);
5220 WARN_ON(smp_processor_id() != target);
5223 touch_cache(cache, size);
5226 cost = t1-t0 + t3-t2;
5228 if (migration_debug >= 2)
5229 printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n",
5230 source, target, t1-t0, t1-t0, t3-t2, cost);
5232 * Flush target caches to RAM and invalidate them:
5236 set_cpus_allowed(current, saved_mask);
5242 * Measure a series of task migrations and return the average
5243 * result. Since this code runs early during bootup the system
5244 * is 'undisturbed' and the average latency makes sense.
5246 * The algorithm in essence auto-detects the relevant cache-size,
5247 * so it will properly detect different cachesizes for different
5248 * cache-hierarchies, depending on how the CPUs are connected.
5250 * Architectures can prime the upper limit of the search range via
5251 * max_cache_size, otherwise the search range defaults to 20MB...64K.
5253 static unsigned long long
5254 measure_cost(int cpu1, int cpu2, void *cache, unsigned int size)
5256 unsigned long long cost1, cost2;
5260 * Measure the migration cost of 'size' bytes, over an
5261 * average of 10 runs:
5263 * (We perturb the cache size by a small (0..4k)
5264 * value to compensate size/alignment related artifacts.
5265 * We also subtract the cost of the operation done on
5271 * dry run, to make sure we start off cache-cold on cpu1,
5272 * and to get any vmalloc pagefaults in advance:
5274 measure_one(cache, size, cpu1, cpu2);
5275 for (i = 0; i < ITERATIONS; i++)
5276 cost1 += measure_one(cache, size - i*1024, cpu1, cpu2);
5278 measure_one(cache, size, cpu2, cpu1);
5279 for (i = 0; i < ITERATIONS; i++)
5280 cost1 += measure_one(cache, size - i*1024, cpu2, cpu1);
5283 * (We measure the non-migrating [cached] cost on both
5284 * cpu1 and cpu2, to handle CPUs with different speeds)
5288 measure_one(cache, size, cpu1, cpu1);
5289 for (i = 0; i < ITERATIONS; i++)
5290 cost2 += measure_one(cache, size - i*1024, cpu1, cpu1);
5292 measure_one(cache, size, cpu2, cpu2);
5293 for (i = 0; i < ITERATIONS; i++)
5294 cost2 += measure_one(cache, size - i*1024, cpu2, cpu2);
5297 * Get the per-iteration migration cost:
5299 do_div(cost1, 2*ITERATIONS);
5300 do_div(cost2, 2*ITERATIONS);
5302 return cost1 - cost2;
5305 static unsigned long long measure_migration_cost(int cpu1, int cpu2)
5307 unsigned long long max_cost = 0, fluct = 0, avg_fluct = 0;
5308 unsigned int max_size, size, size_found = 0;
5309 long long cost = 0, prev_cost;
5313 * Search from max_cache_size*5 down to 64K - the real relevant
5314 * cachesize has to lie somewhere inbetween.
5316 if (max_cache_size) {
5317 max_size = max(max_cache_size * SEARCH_SCOPE, MIN_CACHE_SIZE);
5318 size = max(max_cache_size / SEARCH_SCOPE, MIN_CACHE_SIZE);
5321 * Since we have no estimation about the relevant
5324 max_size = DEFAULT_CACHE_SIZE * SEARCH_SCOPE;
5325 size = MIN_CACHE_SIZE;
5328 if (!cpu_online(cpu1) || !cpu_online(cpu2)) {
5329 printk("cpu %d and %d not both online!\n", cpu1, cpu2);
5334 * Allocate the working set:
5336 cache = vmalloc(max_size);
5338 printk("could not vmalloc %d bytes for cache!\n", 2*max_size);
5339 return 1000000; // return 1 msec on very small boxen
5342 while (size <= max_size) {
5344 cost = measure_cost(cpu1, cpu2, cache, size);
5350 if (max_cost < cost) {
5356 * Calculate average fluctuation, we use this to prevent
5357 * noise from triggering an early break out of the loop:
5359 fluct = abs(cost - prev_cost);
5360 avg_fluct = (avg_fluct + fluct)/2;
5362 if (migration_debug)
5363 printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): (%8Ld %8Ld)\n",
5365 (long)cost / 1000000,
5366 ((long)cost / 100000) % 10,
5367 (long)max_cost / 1000000,
5368 ((long)max_cost / 100000) % 10,
5369 domain_distance(cpu1, cpu2),
5373 * If we iterated at least 20% past the previous maximum,
5374 * and the cost has dropped by more than 20% already,
5375 * (taking fluctuations into account) then we assume to
5376 * have found the maximum and break out of the loop early:
5378 if (size_found && (size*100 > size_found*SIZE_THRESH))
5379 if (cost+avg_fluct <= 0 ||
5380 max_cost*100 > (cost+avg_fluct)*COST_THRESH) {
5382 if (migration_debug)
5383 printk("-> found max.\n");
5387 * Increase the cachesize in 10% steps:
5389 size = size * 10 / 9;
5392 if (migration_debug)
5393 printk("[%d][%d] working set size found: %d, cost: %Ld\n",
5394 cpu1, cpu2, size_found, max_cost);
5399 * A task is considered 'cache cold' if at least 2 times
5400 * the worst-case cost of migration has passed.
5402 * (this limit is only listened to if the load-balancing
5403 * situation is 'nice' - if there is a large imbalance we
5404 * ignore it for the sake of CPU utilization and
5405 * processing fairness.)
5407 return 2 * max_cost * migration_factor / MIGRATION_FACTOR_SCALE;
5410 static void calibrate_migration_costs(const cpumask_t *cpu_map)
5412 int cpu1 = -1, cpu2 = -1, cpu, orig_cpu = raw_smp_processor_id();
5413 unsigned long j0, j1, distance, max_distance = 0;
5414 struct sched_domain *sd;
5419 * First pass - calculate the cacheflush times:
5421 for_each_cpu_mask(cpu1, *cpu_map) {
5422 for_each_cpu_mask(cpu2, *cpu_map) {
5425 distance = domain_distance(cpu1, cpu2);
5426 max_distance = max(max_distance, distance);
5428 * No result cached yet?
5430 if (migration_cost[distance] == -1LL)
5431 migration_cost[distance] =
5432 measure_migration_cost(cpu1, cpu2);
5436 * Second pass - update the sched domain hierarchy with
5437 * the new cache-hot-time estimations:
5439 for_each_cpu_mask(cpu, *cpu_map) {
5441 for_each_domain(cpu, sd) {
5442 sd->cache_hot_time = migration_cost[distance];
5449 if (migration_debug)
5450 printk("migration: max_cache_size: %d, cpu: %d MHz:\n",
5458 if (system_state == SYSTEM_BOOTING) {
5459 printk("migration_cost=");
5460 for (distance = 0; distance <= max_distance; distance++) {
5463 printk("%ld", (long)migration_cost[distance] / 1000);
5468 if (migration_debug)
5469 printk("migration: %ld seconds\n", (j1-j0)/HZ);
5472 * Move back to the original CPU. NUMA-Q gets confused
5473 * if we migrate to another quad during bootup.
5475 if (raw_smp_processor_id() != orig_cpu) {
5476 cpumask_t mask = cpumask_of_cpu(orig_cpu),
5477 saved_mask = current->cpus_allowed;
5479 set_cpus_allowed(current, mask);
5480 set_cpus_allowed(current, saved_mask);
5487 * find_next_best_node - find the next node to include in a sched_domain
5488 * @node: node whose sched_domain we're building
5489 * @used_nodes: nodes already in the sched_domain
5491 * Find the next node to include in a given scheduling domain. Simply
5492 * finds the closest node not already in the @used_nodes map.
5494 * Should use nodemask_t.
5496 static int find_next_best_node(int node, unsigned long *used_nodes)
5498 int i, n, val, min_val, best_node = 0;
5502 for (i = 0; i < MAX_NUMNODES; i++) {
5503 /* Start at @node */
5504 n = (node + i) % MAX_NUMNODES;
5506 if (!nr_cpus_node(n))
5509 /* Skip already used nodes */
5510 if (test_bit(n, used_nodes))
5513 /* Simple min distance search */
5514 val = node_distance(node, n);
5516 if (val < min_val) {
5522 set_bit(best_node, used_nodes);
5527 * sched_domain_node_span - get a cpumask for a node's sched_domain
5528 * @node: node whose cpumask we're constructing
5529 * @size: number of nodes to include in this span
5531 * Given a node, construct a good cpumask for its sched_domain to span. It
5532 * should be one that prevents unnecessary balancing, but also spreads tasks
5535 static cpumask_t sched_domain_node_span(int node)
5538 cpumask_t span, nodemask;
5539 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5542 bitmap_zero(used_nodes, MAX_NUMNODES);
5544 nodemask = node_to_cpumask(node);
5545 cpus_or(span, span, nodemask);
5546 set_bit(node, used_nodes);
5548 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5549 int next_node = find_next_best_node(node, used_nodes);
5550 nodemask = node_to_cpumask(next_node);
5551 cpus_or(span, span, nodemask);
5559 * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we
5560 * can switch it on easily if needed.
5562 #ifdef CONFIG_SCHED_SMT
5563 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5564 static struct sched_group sched_group_cpus[NR_CPUS];
5565 static int cpu_to_cpu_group(int cpu)
5571 #ifdef CONFIG_SCHED_MC
5572 static DEFINE_PER_CPU(struct sched_domain, core_domains);
5573 static struct sched_group sched_group_core[NR_CPUS];
5576 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5577 static int cpu_to_core_group(int cpu)
5579 return first_cpu(cpu_sibling_map[cpu]);
5581 #elif defined(CONFIG_SCHED_MC)
5582 static int cpu_to_core_group(int cpu)
5588 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5589 static struct sched_group sched_group_phys[NR_CPUS];
5590 static int cpu_to_phys_group(int cpu)
5592 #if defined(CONFIG_SCHED_MC)
5593 cpumask_t mask = cpu_coregroup_map(cpu);
5594 return first_cpu(mask);
5595 #elif defined(CONFIG_SCHED_SMT)
5596 return first_cpu(cpu_sibling_map[cpu]);
5604 * The init_sched_build_groups can't handle what we want to do with node
5605 * groups, so roll our own. Now each node has its own list of groups which
5606 * gets dynamically allocated.
5608 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5609 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5611 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5612 static struct sched_group *sched_group_allnodes_bycpu[NR_CPUS];
5614 static int cpu_to_allnodes_group(int cpu)
5616 return cpu_to_node(cpu);
5618 static void init_numa_sched_groups_power(struct sched_group *group_head)
5620 struct sched_group *sg = group_head;
5626 for_each_cpu_mask(j, sg->cpumask) {
5627 struct sched_domain *sd;
5629 sd = &per_cpu(phys_domains, j);
5630 if (j != first_cpu(sd->groups->cpumask)) {
5632 * Only add "power" once for each
5638 sg->cpu_power += sd->groups->cpu_power;
5641 if (sg != group_head)
5647 * Build sched domains for a given set of cpus and attach the sched domains
5648 * to the individual cpus
5650 void build_sched_domains(const cpumask_t *cpu_map)
5654 struct sched_group **sched_group_nodes = NULL;
5655 struct sched_group *sched_group_allnodes = NULL;
5658 * Allocate the per-node list of sched groups
5660 sched_group_nodes = kmalloc(sizeof(struct sched_group*)*MAX_NUMNODES,
5662 if (!sched_group_nodes) {
5663 printk(KERN_WARNING "Can not alloc sched group node list\n");
5666 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
5670 * Set up domains for cpus specified by the cpu_map.
5672 for_each_cpu_mask(i, *cpu_map) {
5674 struct sched_domain *sd = NULL, *p;
5675 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
5677 cpus_and(nodemask, nodemask, *cpu_map);
5680 if (cpus_weight(*cpu_map)
5681 > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
5682 if (!sched_group_allnodes) {
5683 sched_group_allnodes
5684 = kmalloc(sizeof(struct sched_group)
5687 if (!sched_group_allnodes) {
5689 "Can not alloc allnodes sched group\n");
5692 sched_group_allnodes_bycpu[i]
5693 = sched_group_allnodes;
5695 sd = &per_cpu(allnodes_domains, i);
5696 *sd = SD_ALLNODES_INIT;
5697 sd->span = *cpu_map;
5698 group = cpu_to_allnodes_group(i);
5699 sd->groups = &sched_group_allnodes[group];
5704 sd = &per_cpu(node_domains, i);
5706 sd->span = sched_domain_node_span(cpu_to_node(i));
5708 cpus_and(sd->span, sd->span, *cpu_map);
5712 sd = &per_cpu(phys_domains, i);
5713 group = cpu_to_phys_group(i);
5715 sd->span = nodemask;
5717 sd->groups = &sched_group_phys[group];
5719 #ifdef CONFIG_SCHED_MC
5721 sd = &per_cpu(core_domains, i);
5722 group = cpu_to_core_group(i);
5724 sd->span = cpu_coregroup_map(i);
5725 cpus_and(sd->span, sd->span, *cpu_map);
5727 sd->groups = &sched_group_core[group];
5730 #ifdef CONFIG_SCHED_SMT
5732 sd = &per_cpu(cpu_domains, i);
5733 group = cpu_to_cpu_group(i);
5734 *sd = SD_SIBLING_INIT;
5735 sd->span = cpu_sibling_map[i];
5736 cpus_and(sd->span, sd->span, *cpu_map);
5738 sd->groups = &sched_group_cpus[group];
5742 #ifdef CONFIG_SCHED_SMT
5743 /* Set up CPU (sibling) groups */
5744 for_each_cpu_mask(i, *cpu_map) {
5745 cpumask_t this_sibling_map = cpu_sibling_map[i];
5746 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
5747 if (i != first_cpu(this_sibling_map))
5750 init_sched_build_groups(sched_group_cpus, this_sibling_map,
5755 #ifdef CONFIG_SCHED_MC
5756 /* Set up multi-core groups */
5757 for_each_cpu_mask(i, *cpu_map) {
5758 cpumask_t this_core_map = cpu_coregroup_map(i);
5759 cpus_and(this_core_map, this_core_map, *cpu_map);
5760 if (i != first_cpu(this_core_map))
5762 init_sched_build_groups(sched_group_core, this_core_map,
5763 &cpu_to_core_group);
5768 /* Set up physical groups */
5769 for (i = 0; i < MAX_NUMNODES; i++) {
5770 cpumask_t nodemask = node_to_cpumask(i);
5772 cpus_and(nodemask, nodemask, *cpu_map);
5773 if (cpus_empty(nodemask))
5776 init_sched_build_groups(sched_group_phys, nodemask,
5777 &cpu_to_phys_group);
5781 /* Set up node groups */
5782 if (sched_group_allnodes)
5783 init_sched_build_groups(sched_group_allnodes, *cpu_map,
5784 &cpu_to_allnodes_group);
5786 for (i = 0; i < MAX_NUMNODES; i++) {
5787 /* Set up node groups */
5788 struct sched_group *sg, *prev;
5789 cpumask_t nodemask = node_to_cpumask(i);
5790 cpumask_t domainspan;
5791 cpumask_t covered = CPU_MASK_NONE;
5794 cpus_and(nodemask, nodemask, *cpu_map);
5795 if (cpus_empty(nodemask)) {
5796 sched_group_nodes[i] = NULL;
5800 domainspan = sched_domain_node_span(i);
5801 cpus_and(domainspan, domainspan, *cpu_map);
5803 sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL);
5804 sched_group_nodes[i] = sg;
5805 for_each_cpu_mask(j, nodemask) {
5806 struct sched_domain *sd;
5807 sd = &per_cpu(node_domains, j);
5809 if (sd->groups == NULL) {
5810 /* Turn off balancing if we have no groups */
5816 "Can not alloc domain group for node %d\n", i);
5820 sg->cpumask = nodemask;
5821 cpus_or(covered, covered, nodemask);
5824 for (j = 0; j < MAX_NUMNODES; j++) {
5825 cpumask_t tmp, notcovered;
5826 int n = (i + j) % MAX_NUMNODES;
5828 cpus_complement(notcovered, covered);
5829 cpus_and(tmp, notcovered, *cpu_map);
5830 cpus_and(tmp, tmp, domainspan);
5831 if (cpus_empty(tmp))
5834 nodemask = node_to_cpumask(n);
5835 cpus_and(tmp, tmp, nodemask);
5836 if (cpus_empty(tmp))
5839 sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL);
5842 "Can not alloc domain group for node %d\n", j);
5847 cpus_or(covered, covered, tmp);
5851 prev->next = sched_group_nodes[i];
5855 /* Calculate CPU power for physical packages and nodes */
5856 for_each_cpu_mask(i, *cpu_map) {
5858 struct sched_domain *sd;
5859 #ifdef CONFIG_SCHED_SMT
5860 sd = &per_cpu(cpu_domains, i);
5861 power = SCHED_LOAD_SCALE;
5862 sd->groups->cpu_power = power;
5864 #ifdef CONFIG_SCHED_MC
5865 sd = &per_cpu(core_domains, i);
5866 power = SCHED_LOAD_SCALE + (cpus_weight(sd->groups->cpumask)-1)
5867 * SCHED_LOAD_SCALE / 10;
5868 sd->groups->cpu_power = power;
5870 sd = &per_cpu(phys_domains, i);
5873 * This has to be < 2 * SCHED_LOAD_SCALE
5874 * Lets keep it SCHED_LOAD_SCALE, so that
5875 * while calculating NUMA group's cpu_power
5877 * numa_group->cpu_power += phys_group->cpu_power;
5879 * See "only add power once for each physical pkg"
5882 sd->groups->cpu_power = SCHED_LOAD_SCALE;
5884 sd = &per_cpu(phys_domains, i);
5885 power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
5886 (cpus_weight(sd->groups->cpumask)-1) / 10;
5887 sd->groups->cpu_power = power;
5892 for (i = 0; i < MAX_NUMNODES; i++)
5893 init_numa_sched_groups_power(sched_group_nodes[i]);
5895 init_numa_sched_groups_power(sched_group_allnodes);
5898 /* Attach the domains */
5899 for_each_cpu_mask(i, *cpu_map) {
5900 struct sched_domain *sd;
5901 #ifdef CONFIG_SCHED_SMT
5902 sd = &per_cpu(cpu_domains, i);
5903 #elif defined(CONFIG_SCHED_MC)
5904 sd = &per_cpu(core_domains, i);
5906 sd = &per_cpu(phys_domains, i);
5908 cpu_attach_domain(sd, i);
5911 * Tune cache-hot values:
5913 calibrate_migration_costs(cpu_map);
5916 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
5918 static void arch_init_sched_domains(const cpumask_t *cpu_map)
5920 cpumask_t cpu_default_map;
5923 * Setup mask for cpus without special case scheduling requirements.
5924 * For now this just excludes isolated cpus, but could be used to
5925 * exclude other special cases in the future.
5927 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
5929 build_sched_domains(&cpu_default_map);
5932 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
5938 for_each_cpu_mask(cpu, *cpu_map) {
5939 struct sched_group *sched_group_allnodes
5940 = sched_group_allnodes_bycpu[cpu];
5941 struct sched_group **sched_group_nodes
5942 = sched_group_nodes_bycpu[cpu];
5944 if (sched_group_allnodes) {
5945 kfree(sched_group_allnodes);
5946 sched_group_allnodes_bycpu[cpu] = NULL;
5949 if (!sched_group_nodes)
5952 for (i = 0; i < MAX_NUMNODES; i++) {
5953 cpumask_t nodemask = node_to_cpumask(i);
5954 struct sched_group *oldsg, *sg = sched_group_nodes[i];
5956 cpus_and(nodemask, nodemask, *cpu_map);
5957 if (cpus_empty(nodemask))
5967 if (oldsg != sched_group_nodes[i])
5970 kfree(sched_group_nodes);
5971 sched_group_nodes_bycpu[cpu] = NULL;
5977 * Detach sched domains from a group of cpus specified in cpu_map
5978 * These cpus will now be attached to the NULL domain
5980 static void detach_destroy_domains(const cpumask_t *cpu_map)
5984 for_each_cpu_mask(i, *cpu_map)
5985 cpu_attach_domain(NULL, i);
5986 synchronize_sched();
5987 arch_destroy_sched_domains(cpu_map);
5991 * Partition sched domains as specified by the cpumasks below.
5992 * This attaches all cpus from the cpumasks to the NULL domain,
5993 * waits for a RCU quiescent period, recalculates sched
5994 * domain information and then attaches them back to the
5995 * correct sched domains
5996 * Call with hotplug lock held
5998 void partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6000 cpumask_t change_map;
6002 cpus_and(*partition1, *partition1, cpu_online_map);
6003 cpus_and(*partition2, *partition2, cpu_online_map);
6004 cpus_or(change_map, *partition1, *partition2);
6006 /* Detach sched domains from all of the affected cpus */
6007 detach_destroy_domains(&change_map);
6008 if (!cpus_empty(*partition1))
6009 build_sched_domains(partition1);
6010 if (!cpus_empty(*partition2))
6011 build_sched_domains(partition2);
6014 #ifdef CONFIG_HOTPLUG_CPU
6016 * Force a reinitialization of the sched domains hierarchy. The domains
6017 * and groups cannot be updated in place without racing with the balancing
6018 * code, so we temporarily attach all running cpus to the NULL domain
6019 * which will prevent rebalancing while the sched domains are recalculated.
6021 static int update_sched_domains(struct notifier_block *nfb,
6022 unsigned long action, void *hcpu)
6025 case CPU_UP_PREPARE:
6026 case CPU_DOWN_PREPARE:
6027 detach_destroy_domains(&cpu_online_map);
6030 case CPU_UP_CANCELED:
6031 case CPU_DOWN_FAILED:
6035 * Fall through and re-initialise the domains.
6042 /* The hotplug lock is already held by cpu_up/cpu_down */
6043 arch_init_sched_domains(&cpu_online_map);
6049 void __init sched_init_smp(void)
6052 arch_init_sched_domains(&cpu_online_map);
6053 unlock_cpu_hotplug();
6054 /* XXX: Theoretical race here - CPU may be hotplugged now */
6055 hotcpu_notifier(update_sched_domains, 0);
6058 void __init sched_init_smp(void)
6061 #endif /* CONFIG_SMP */
6063 int in_sched_functions(unsigned long addr)
6065 /* Linker adds these: start and end of __sched functions */
6066 extern char __sched_text_start[], __sched_text_end[];
6067 return in_lock_functions(addr) ||
6068 (addr >= (unsigned long)__sched_text_start
6069 && addr < (unsigned long)__sched_text_end);
6072 void __init sched_init(void)
6077 for_each_possible_cpu(i) {
6078 prio_array_t *array;
6081 spin_lock_init(&rq->lock);
6083 rq->active = rq->arrays;
6084 rq->expired = rq->arrays + 1;
6085 rq->best_expired_prio = MAX_PRIO;
6089 for (j = 1; j < 3; j++)
6090 rq->cpu_load[j] = 0;
6091 rq->active_balance = 0;
6093 rq->migration_thread = NULL;
6094 INIT_LIST_HEAD(&rq->migration_queue);
6096 atomic_set(&rq->nr_iowait, 0);
6098 for (j = 0; j < 2; j++) {
6099 array = rq->arrays + j;
6100 for (k = 0; k < MAX_PRIO; k++) {
6101 INIT_LIST_HEAD(array->queue + k);
6102 __clear_bit(k, array->bitmap);
6104 // delimiter for bitsearch
6105 __set_bit(MAX_PRIO, array->bitmap);
6110 * The boot idle thread does lazy MMU switching as well:
6112 atomic_inc(&init_mm.mm_count);
6113 enter_lazy_tlb(&init_mm, current);
6116 * Make us the idle thread. Technically, schedule() should not be
6117 * called from this thread, however somewhere below it might be,
6118 * but because we are the idle thread, we just pick up running again
6119 * when this runqueue becomes "idle".
6121 init_idle(current, smp_processor_id());
6124 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6125 void __might_sleep(char *file, int line)
6127 #if defined(in_atomic)
6128 static unsigned long prev_jiffy; /* ratelimiting */
6130 if ((in_atomic() || irqs_disabled()) &&
6131 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6132 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6134 prev_jiffy = jiffies;
6135 printk(KERN_ERR "BUG: sleeping function called from invalid"
6136 " context at %s:%d\n", file, line);
6137 printk("in_atomic():%d, irqs_disabled():%d\n",
6138 in_atomic(), irqs_disabled());
6143 EXPORT_SYMBOL(__might_sleep);
6146 #ifdef CONFIG_MAGIC_SYSRQ
6147 void normalize_rt_tasks(void)
6149 struct task_struct *p;
6150 prio_array_t *array;
6151 unsigned long flags;
6154 read_lock_irq(&tasklist_lock);
6155 for_each_process(p) {
6159 rq = task_rq_lock(p, &flags);
6163 deactivate_task(p, task_rq(p));
6164 __setscheduler(p, SCHED_NORMAL, 0);
6166 __activate_task(p, task_rq(p));
6167 resched_task(rq->curr);
6170 task_rq_unlock(rq, &flags);
6172 read_unlock_irq(&tasklist_lock);
6175 #endif /* CONFIG_MAGIC_SYSRQ */
6179 * These functions are only useful for the IA64 MCA handling.
6181 * They can only be called when the whole system has been
6182 * stopped - every CPU needs to be quiescent, and no scheduling
6183 * activity can take place. Using them for anything else would
6184 * be a serious bug, and as a result, they aren't even visible
6185 * under any other configuration.
6189 * curr_task - return the current task for a given cpu.
6190 * @cpu: the processor in question.
6192 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6194 task_t *curr_task(int cpu)
6196 return cpu_curr(cpu);
6200 * set_curr_task - set the current task for a given cpu.
6201 * @cpu: the processor in question.
6202 * @p: the task pointer to set.
6204 * Description: This function must only be used when non-maskable interrupts
6205 * are serviced on a separate stack. It allows the architecture to switch the
6206 * notion of the current task on a cpu in a non-blocking manner. This function
6207 * must be called with all CPU's synchronized, and interrupts disabled, the
6208 * and caller must save the original value of the current task (see
6209 * curr_task() above) and restore that value before reenabling interrupts and
6210 * re-starting the system.
6212 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6214 void set_curr_task(int cpu, task_t *p)