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>
54 #include <asm/unistd.h>
57 * Convert user-nice values [ -20 ... 0 ... 19 ]
58 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
61 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
62 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
63 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
66 * 'User priority' is the nice value converted to something we
67 * can work with better when scaling various scheduler parameters,
68 * it's a [ 0 ... 39 ] range.
70 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
71 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
72 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
75 * Some helpers for converting nanosecond timing to jiffy resolution
77 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
78 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
81 * These are the 'tuning knobs' of the scheduler:
83 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
84 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
85 * Timeslices get refilled after they expire.
87 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
88 #define DEF_TIMESLICE (100 * HZ / 1000)
89 #define ON_RUNQUEUE_WEIGHT 30
90 #define CHILD_PENALTY 95
91 #define PARENT_PENALTY 100
93 #define PRIO_BONUS_RATIO 25
94 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
95 #define INTERACTIVE_DELTA 2
96 #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
97 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
98 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
101 * If a task is 'interactive' then we reinsert it in the active
102 * array after it has expired its current timeslice. (it will not
103 * continue to run immediately, it will still roundrobin with
104 * other interactive tasks.)
106 * This part scales the interactivity limit depending on niceness.
108 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
109 * Here are a few examples of different nice levels:
111 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
112 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
113 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
114 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
115 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
117 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
118 * priority range a task can explore, a value of '1' means the
119 * task is rated interactive.)
121 * Ie. nice +19 tasks can never get 'interactive' enough to be
122 * reinserted into the active array. And only heavily CPU-hog nice -20
123 * tasks will be expired. Default nice 0 tasks are somewhere between,
124 * it takes some effort for them to get interactive, but it's not
128 #define CURRENT_BONUS(p) \
129 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
132 #define GRANULARITY (10 * HZ / 1000 ? : 1)
135 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
136 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
139 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
140 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
143 #define SCALE(v1,v1_max,v2_max) \
144 (v1) * (v2_max) / (v1_max)
147 (SCALE(TASK_NICE(p), 40, MAX_BONUS) + INTERACTIVE_DELTA)
149 #define TASK_INTERACTIVE(p) \
150 ((p)->prio <= (p)->static_prio - DELTA(p))
152 #define INTERACTIVE_SLEEP(p) \
153 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
154 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
156 #define TASK_PREEMPTS_CURR(p, rq) \
157 ((p)->prio < (rq)->curr->prio)
160 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
161 * to time slice values: [800ms ... 100ms ... 5ms]
163 * The higher a thread's priority, the bigger timeslices
164 * it gets during one round of execution. But even the lowest
165 * priority thread gets MIN_TIMESLICE worth of execution time.
168 #define SCALE_PRIO(x, prio) \
169 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO/2), MIN_TIMESLICE)
171 static unsigned int task_timeslice(task_t *p)
173 if (p->static_prio < NICE_TO_PRIO(0))
174 return SCALE_PRIO(DEF_TIMESLICE*4, p->static_prio);
176 return SCALE_PRIO(DEF_TIMESLICE, p->static_prio);
178 #define task_hot(p, now, sd) ((long long) ((now) - (p)->last_ran) \
179 < (long long) (sd)->cache_hot_time)
181 void __put_task_struct_cb(struct rcu_head *rhp)
183 __put_task_struct(container_of(rhp, struct task_struct, rcu));
186 EXPORT_SYMBOL_GPL(__put_task_struct_cb);
189 * These are the runqueue data structures:
192 #define BITMAP_SIZE ((((MAX_PRIO+1+7)/8)+sizeof(long)-1)/sizeof(long))
194 typedef struct runqueue runqueue_t;
197 unsigned int nr_active;
198 unsigned long bitmap[BITMAP_SIZE];
199 struct list_head queue[MAX_PRIO];
203 * This is the main, per-CPU runqueue data structure.
205 * Locking rule: those places that want to lock multiple runqueues
206 * (such as the load balancing or the thread migration code), lock
207 * acquire operations must be ordered by ascending &runqueue.
213 * nr_running and cpu_load should be in the same cacheline because
214 * remote CPUs use both these fields when doing load calculation.
216 unsigned long nr_running;
218 unsigned long cpu_load[3];
220 unsigned long long nr_switches;
223 * This is part of a global counter where only the total sum
224 * over all CPUs matters. A task can increase this counter on
225 * one CPU and if it got migrated afterwards it may decrease
226 * it on another CPU. Always updated under the runqueue lock:
228 unsigned long nr_uninterruptible;
230 unsigned long expired_timestamp;
231 unsigned long long timestamp_last_tick;
233 struct mm_struct *prev_mm;
234 prio_array_t *active, *expired, arrays[2];
235 int best_expired_prio;
239 struct sched_domain *sd;
241 /* For active balancing */
245 task_t *migration_thread;
246 struct list_head migration_queue;
249 #ifdef CONFIG_SCHEDSTATS
251 struct sched_info rq_sched_info;
253 /* sys_sched_yield() stats */
254 unsigned long yld_exp_empty;
255 unsigned long yld_act_empty;
256 unsigned long yld_both_empty;
257 unsigned long yld_cnt;
259 /* schedule() stats */
260 unsigned long sched_switch;
261 unsigned long sched_cnt;
262 unsigned long sched_goidle;
264 /* try_to_wake_up() stats */
265 unsigned long ttwu_cnt;
266 unsigned long ttwu_local;
270 static DEFINE_PER_CPU(struct runqueue, runqueues);
273 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
274 * See detach_destroy_domains: synchronize_sched for details.
276 * The domain tree of any CPU may only be accessed from within
277 * preempt-disabled sections.
279 #define for_each_domain(cpu, domain) \
280 for (domain = rcu_dereference(cpu_rq(cpu)->sd); domain; domain = domain->parent)
282 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
283 #define this_rq() (&__get_cpu_var(runqueues))
284 #define task_rq(p) cpu_rq(task_cpu(p))
285 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
287 #ifndef prepare_arch_switch
288 # define prepare_arch_switch(next) do { } while (0)
290 #ifndef finish_arch_switch
291 # define finish_arch_switch(prev) do { } while (0)
294 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
295 static inline int task_running(runqueue_t *rq, task_t *p)
297 return rq->curr == p;
300 static inline void prepare_lock_switch(runqueue_t *rq, task_t *next)
304 static inline void finish_lock_switch(runqueue_t *rq, task_t *prev)
306 #ifdef CONFIG_DEBUG_SPINLOCK
307 /* this is a valid case when another task releases the spinlock */
308 rq->lock.owner = current;
310 spin_unlock_irq(&rq->lock);
313 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
314 static inline int task_running(runqueue_t *rq, task_t *p)
319 return rq->curr == p;
323 static inline void prepare_lock_switch(runqueue_t *rq, task_t *next)
327 * We can optimise this out completely for !SMP, because the
328 * SMP rebalancing from interrupt is the only thing that cares
333 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
334 spin_unlock_irq(&rq->lock);
336 spin_unlock(&rq->lock);
340 static inline void finish_lock_switch(runqueue_t *rq, task_t *prev)
344 * After ->oncpu is cleared, the task can be moved to a different CPU.
345 * We must ensure this doesn't happen until the switch is completely
351 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
355 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
358 * task_rq_lock - lock the runqueue a given task resides on and disable
359 * interrupts. Note the ordering: we can safely lookup the task_rq without
360 * explicitly disabling preemption.
362 static inline runqueue_t *task_rq_lock(task_t *p, unsigned long *flags)
368 local_irq_save(*flags);
370 spin_lock(&rq->lock);
371 if (unlikely(rq != task_rq(p))) {
372 spin_unlock_irqrestore(&rq->lock, *flags);
373 goto repeat_lock_task;
378 static inline void task_rq_unlock(runqueue_t *rq, unsigned long *flags)
381 spin_unlock_irqrestore(&rq->lock, *flags);
384 #ifdef CONFIG_SCHEDSTATS
386 * bump this up when changing the output format or the meaning of an existing
387 * format, so that tools can adapt (or abort)
389 #define SCHEDSTAT_VERSION 12
391 static int show_schedstat(struct seq_file *seq, void *v)
395 seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION);
396 seq_printf(seq, "timestamp %lu\n", jiffies);
397 for_each_online_cpu(cpu) {
398 runqueue_t *rq = cpu_rq(cpu);
400 struct sched_domain *sd;
404 /* runqueue-specific stats */
406 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
407 cpu, rq->yld_both_empty,
408 rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt,
409 rq->sched_switch, rq->sched_cnt, rq->sched_goidle,
410 rq->ttwu_cnt, rq->ttwu_local,
411 rq->rq_sched_info.cpu_time,
412 rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt);
414 seq_printf(seq, "\n");
417 /* domain-specific stats */
419 for_each_domain(cpu, sd) {
420 enum idle_type itype;
421 char mask_str[NR_CPUS];
423 cpumask_scnprintf(mask_str, NR_CPUS, sd->span);
424 seq_printf(seq, "domain%d %s", dcnt++, mask_str);
425 for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES;
427 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu",
429 sd->lb_balanced[itype],
430 sd->lb_failed[itype],
431 sd->lb_imbalance[itype],
432 sd->lb_gained[itype],
433 sd->lb_hot_gained[itype],
434 sd->lb_nobusyq[itype],
435 sd->lb_nobusyg[itype]);
437 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
438 sd->alb_cnt, sd->alb_failed, sd->alb_pushed,
439 sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed,
440 sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed,
441 sd->ttwu_wake_remote, sd->ttwu_move_affine, sd->ttwu_move_balance);
449 static int schedstat_open(struct inode *inode, struct file *file)
451 unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32);
452 char *buf = kmalloc(size, GFP_KERNEL);
458 res = single_open(file, show_schedstat, NULL);
460 m = file->private_data;
468 struct file_operations proc_schedstat_operations = {
469 .open = schedstat_open,
472 .release = single_release,
475 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
476 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
477 #else /* !CONFIG_SCHEDSTATS */
478 # define schedstat_inc(rq, field) do { } while (0)
479 # define schedstat_add(rq, field, amt) do { } while (0)
483 * rq_lock - lock a given runqueue and disable interrupts.
485 static inline runqueue_t *this_rq_lock(void)
492 spin_lock(&rq->lock);
497 #ifdef CONFIG_SCHEDSTATS
499 * Called when a process is dequeued from the active array and given
500 * the cpu. We should note that with the exception of interactive
501 * tasks, the expired queue will become the active queue after the active
502 * queue is empty, without explicitly dequeuing and requeuing tasks in the
503 * expired queue. (Interactive tasks may be requeued directly to the
504 * active queue, thus delaying tasks in the expired queue from running;
505 * see scheduler_tick()).
507 * This function is only called from sched_info_arrive(), rather than
508 * dequeue_task(). Even though a task may be queued and dequeued multiple
509 * times as it is shuffled about, we're really interested in knowing how
510 * long it was from the *first* time it was queued to the time that it
513 static inline void sched_info_dequeued(task_t *t)
515 t->sched_info.last_queued = 0;
519 * Called when a task finally hits the cpu. We can now calculate how
520 * long it was waiting to run. We also note when it began so that we
521 * can keep stats on how long its timeslice is.
523 static void sched_info_arrive(task_t *t)
525 unsigned long now = jiffies, diff = 0;
526 struct runqueue *rq = task_rq(t);
528 if (t->sched_info.last_queued)
529 diff = now - t->sched_info.last_queued;
530 sched_info_dequeued(t);
531 t->sched_info.run_delay += diff;
532 t->sched_info.last_arrival = now;
533 t->sched_info.pcnt++;
538 rq->rq_sched_info.run_delay += diff;
539 rq->rq_sched_info.pcnt++;
543 * Called when a process is queued into either the active or expired
544 * array. The time is noted and later used to determine how long we
545 * had to wait for us to reach the cpu. Since the expired queue will
546 * become the active queue after active queue is empty, without dequeuing
547 * and requeuing any tasks, we are interested in queuing to either. It
548 * is unusual but not impossible for tasks to be dequeued and immediately
549 * requeued in the same or another array: this can happen in sched_yield(),
550 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
553 * This function is only called from enqueue_task(), but also only updates
554 * the timestamp if it is already not set. It's assumed that
555 * sched_info_dequeued() will clear that stamp when appropriate.
557 static inline void sched_info_queued(task_t *t)
559 if (!t->sched_info.last_queued)
560 t->sched_info.last_queued = jiffies;
564 * Called when a process ceases being the active-running process, either
565 * voluntarily or involuntarily. Now we can calculate how long we ran.
567 static inline void sched_info_depart(task_t *t)
569 struct runqueue *rq = task_rq(t);
570 unsigned long diff = jiffies - t->sched_info.last_arrival;
572 t->sched_info.cpu_time += diff;
575 rq->rq_sched_info.cpu_time += diff;
579 * Called when tasks are switched involuntarily due, typically, to expiring
580 * their time slice. (This may also be called when switching to or from
581 * the idle task.) We are only called when prev != next.
583 static inline void sched_info_switch(task_t *prev, task_t *next)
585 struct runqueue *rq = task_rq(prev);
588 * prev now departs the cpu. It's not interesting to record
589 * stats about how efficient we were at scheduling the idle
592 if (prev != rq->idle)
593 sched_info_depart(prev);
595 if (next != rq->idle)
596 sched_info_arrive(next);
599 #define sched_info_queued(t) do { } while (0)
600 #define sched_info_switch(t, next) do { } while (0)
601 #endif /* CONFIG_SCHEDSTATS */
604 * Adding/removing a task to/from a priority array:
606 static void dequeue_task(struct task_struct *p, prio_array_t *array)
609 list_del(&p->run_list);
610 if (list_empty(array->queue + p->prio))
611 __clear_bit(p->prio, array->bitmap);
614 static void enqueue_task(struct task_struct *p, prio_array_t *array)
616 sched_info_queued(p);
617 list_add_tail(&p->run_list, array->queue + p->prio);
618 __set_bit(p->prio, array->bitmap);
624 * Put task to the end of the run list without the overhead of dequeue
625 * followed by enqueue.
627 static void requeue_task(struct task_struct *p, prio_array_t *array)
629 list_move_tail(&p->run_list, array->queue + p->prio);
632 static inline void enqueue_task_head(struct task_struct *p, prio_array_t *array)
634 list_add(&p->run_list, array->queue + p->prio);
635 __set_bit(p->prio, array->bitmap);
641 * effective_prio - return the priority that is based on the static
642 * priority but is modified by bonuses/penalties.
644 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
645 * into the -5 ... 0 ... +5 bonus/penalty range.
647 * We use 25% of the full 0...39 priority range so that:
649 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
650 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
652 * Both properties are important to certain workloads.
654 static int effective_prio(task_t *p)
661 bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
663 prio = p->static_prio - bonus;
664 if (prio < MAX_RT_PRIO)
666 if (prio > MAX_PRIO-1)
672 * __activate_task - move a task to the runqueue.
674 static inline void __activate_task(task_t *p, runqueue_t *rq)
676 enqueue_task(p, rq->active);
681 * __activate_idle_task - move idle task to the _front_ of runqueue.
683 static inline void __activate_idle_task(task_t *p, runqueue_t *rq)
685 enqueue_task_head(p, rq->active);
689 static int recalc_task_prio(task_t *p, unsigned long long now)
691 /* Caller must always ensure 'now >= p->timestamp' */
692 unsigned long long __sleep_time = now - p->timestamp;
693 unsigned long sleep_time;
695 if (unlikely(p->policy == SCHED_BATCH))
698 if (__sleep_time > NS_MAX_SLEEP_AVG)
699 sleep_time = NS_MAX_SLEEP_AVG;
701 sleep_time = (unsigned long)__sleep_time;
704 if (likely(sleep_time > 0)) {
706 * User tasks that sleep a long time are categorised as
707 * idle and will get just interactive status to stay active &
708 * prevent them suddenly becoming cpu hogs and starving
711 if (p->mm && p->activated != -1 &&
712 sleep_time > INTERACTIVE_SLEEP(p)) {
713 p->sleep_avg = JIFFIES_TO_NS(MAX_SLEEP_AVG -
717 * The lower the sleep avg a task has the more
718 * rapidly it will rise with sleep time.
720 sleep_time *= (MAX_BONUS - CURRENT_BONUS(p)) ? : 1;
723 * Tasks waking from uninterruptible sleep are
724 * limited in their sleep_avg rise as they
725 * are likely to be waiting on I/O
727 if (p->activated == -1 && p->mm) {
728 if (p->sleep_avg >= INTERACTIVE_SLEEP(p))
730 else if (p->sleep_avg + sleep_time >=
731 INTERACTIVE_SLEEP(p)) {
732 p->sleep_avg = INTERACTIVE_SLEEP(p);
738 * This code gives a bonus to interactive tasks.
740 * The boost works by updating the 'average sleep time'
741 * value here, based on ->timestamp. The more time a
742 * task spends sleeping, the higher the average gets -
743 * and the higher the priority boost gets as well.
745 p->sleep_avg += sleep_time;
747 if (p->sleep_avg > NS_MAX_SLEEP_AVG)
748 p->sleep_avg = NS_MAX_SLEEP_AVG;
752 return effective_prio(p);
756 * activate_task - move a task to the runqueue and do priority recalculation
758 * Update all the scheduling statistics stuff. (sleep average
759 * calculation, priority modifiers, etc.)
761 static void activate_task(task_t *p, runqueue_t *rq, int local)
763 unsigned long long now;
768 /* Compensate for drifting sched_clock */
769 runqueue_t *this_rq = this_rq();
770 now = (now - this_rq->timestamp_last_tick)
771 + rq->timestamp_last_tick;
776 p->prio = recalc_task_prio(p, now);
779 * This checks to make sure it's not an uninterruptible task
780 * that is now waking up.
784 * Tasks which were woken up by interrupts (ie. hw events)
785 * are most likely of interactive nature. So we give them
786 * the credit of extending their sleep time to the period
787 * of time they spend on the runqueue, waiting for execution
788 * on a CPU, first time around:
794 * Normal first-time wakeups get a credit too for
795 * on-runqueue time, but it will be weighted down:
802 __activate_task(p, rq);
806 * deactivate_task - remove a task from the runqueue.
808 static void deactivate_task(struct task_struct *p, runqueue_t *rq)
811 dequeue_task(p, p->array);
816 * resched_task - mark a task 'to be rescheduled now'.
818 * On UP this means the setting of the need_resched flag, on SMP it
819 * might also involve a cross-CPU call to trigger the scheduler on
823 static void resched_task(task_t *p)
827 assert_spin_locked(&task_rq(p)->lock);
829 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
832 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
835 if (cpu == smp_processor_id())
838 /* NEED_RESCHED must be visible before we test POLLING_NRFLAG */
840 if (!test_tsk_thread_flag(p, TIF_POLLING_NRFLAG))
841 smp_send_reschedule(cpu);
844 static inline void resched_task(task_t *p)
846 assert_spin_locked(&task_rq(p)->lock);
847 set_tsk_need_resched(p);
852 * task_curr - is this task currently executing on a CPU?
853 * @p: the task in question.
855 inline int task_curr(const task_t *p)
857 return cpu_curr(task_cpu(p)) == p;
862 struct list_head list;
867 struct completion done;
871 * The task's runqueue lock must be held.
872 * Returns true if you have to wait for migration thread.
874 static int migrate_task(task_t *p, int dest_cpu, migration_req_t *req)
876 runqueue_t *rq = task_rq(p);
879 * If the task is not on a runqueue (and not running), then
880 * it is sufficient to simply update the task's cpu field.
882 if (!p->array && !task_running(rq, p)) {
883 set_task_cpu(p, dest_cpu);
887 init_completion(&req->done);
889 req->dest_cpu = dest_cpu;
890 list_add(&req->list, &rq->migration_queue);
895 * wait_task_inactive - wait for a thread to unschedule.
897 * The caller must ensure that the task *will* unschedule sometime soon,
898 * else this function might spin for a *long* time. This function can't
899 * be called with interrupts off, or it may introduce deadlock with
900 * smp_call_function() if an IPI is sent by the same process we are
901 * waiting to become inactive.
903 void wait_task_inactive(task_t *p)
910 rq = task_rq_lock(p, &flags);
911 /* Must be off runqueue entirely, not preempted. */
912 if (unlikely(p->array || task_running(rq, p))) {
913 /* If it's preempted, we yield. It could be a while. */
914 preempted = !task_running(rq, p);
915 task_rq_unlock(rq, &flags);
921 task_rq_unlock(rq, &flags);
925 * kick_process - kick a running thread to enter/exit the kernel
926 * @p: the to-be-kicked thread
928 * Cause a process which is running on another CPU to enter
929 * kernel-mode, without any delay. (to get signals handled.)
931 * NOTE: this function doesnt have to take the runqueue lock,
932 * because all it wants to ensure is that the remote task enters
933 * the kernel. If the IPI races and the task has been migrated
934 * to another CPU then no harm is done and the purpose has been
937 void kick_process(task_t *p)
943 if ((cpu != smp_processor_id()) && task_curr(p))
944 smp_send_reschedule(cpu);
949 * Return a low guess at the load of a migration-source cpu.
951 * We want to under-estimate the load of migration sources, to
952 * balance conservatively.
954 static inline unsigned long source_load(int cpu, int type)
956 runqueue_t *rq = cpu_rq(cpu);
957 unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE;
961 return min(rq->cpu_load[type-1], load_now);
965 * Return a high guess at the load of a migration-target cpu
967 static inline unsigned long target_load(int cpu, int type)
969 runqueue_t *rq = cpu_rq(cpu);
970 unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE;
974 return max(rq->cpu_load[type-1], load_now);
978 * find_idlest_group finds and returns the least busy CPU group within the
981 static struct sched_group *
982 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
984 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
985 unsigned long min_load = ULONG_MAX, this_load = 0;
986 int load_idx = sd->forkexec_idx;
987 int imbalance = 100 + (sd->imbalance_pct-100)/2;
990 unsigned long load, avg_load;
994 /* Skip over this group if it has no CPUs allowed */
995 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
998 local_group = cpu_isset(this_cpu, group->cpumask);
1000 /* Tally up the load of all CPUs in the group */
1003 for_each_cpu_mask(i, group->cpumask) {
1004 /* Bias balancing toward cpus of our domain */
1006 load = source_load(i, load_idx);
1008 load = target_load(i, load_idx);
1013 /* Adjust by relative CPU power of the group */
1014 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1017 this_load = avg_load;
1019 } else if (avg_load < min_load) {
1020 min_load = avg_load;
1024 group = group->next;
1025 } while (group != sd->groups);
1027 if (!idlest || 100*this_load < imbalance*min_load)
1033 * find_idlest_queue - find the idlest runqueue among the cpus in group.
1036 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1039 unsigned long load, min_load = ULONG_MAX;
1043 /* Traverse only the allowed CPUs */
1044 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1046 for_each_cpu_mask(i, tmp) {
1047 load = source_load(i, 0);
1049 if (load < min_load || (load == min_load && i == this_cpu)) {
1059 * sched_balance_self: balance the current task (running on cpu) in domains
1060 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1063 * Balance, ie. select the least loaded group.
1065 * Returns the target CPU number, or the same CPU if no balancing is needed.
1067 * preempt must be disabled.
1069 static int sched_balance_self(int cpu, int flag)
1071 struct task_struct *t = current;
1072 struct sched_domain *tmp, *sd = NULL;
1074 for_each_domain(cpu, tmp)
1075 if (tmp->flags & flag)
1080 struct sched_group *group;
1085 group = find_idlest_group(sd, t, cpu);
1089 new_cpu = find_idlest_cpu(group, t, cpu);
1090 if (new_cpu == -1 || new_cpu == cpu)
1093 /* Now try balancing at a lower domain level */
1097 weight = cpus_weight(span);
1098 for_each_domain(cpu, tmp) {
1099 if (weight <= cpus_weight(tmp->span))
1101 if (tmp->flags & flag)
1104 /* while loop will break here if sd == NULL */
1110 #endif /* CONFIG_SMP */
1113 * wake_idle() will wake a task on an idle cpu if task->cpu is
1114 * not idle and an idle cpu is available. The span of cpus to
1115 * search starts with cpus closest then further out as needed,
1116 * so we always favor a closer, idle cpu.
1118 * Returns the CPU we should wake onto.
1120 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1121 static int wake_idle(int cpu, task_t *p)
1124 struct sched_domain *sd;
1130 for_each_domain(cpu, sd) {
1131 if (sd->flags & SD_WAKE_IDLE) {
1132 cpus_and(tmp, sd->span, p->cpus_allowed);
1133 for_each_cpu_mask(i, tmp) {
1144 static inline int wake_idle(int cpu, task_t *p)
1151 * try_to_wake_up - wake up a thread
1152 * @p: the to-be-woken-up thread
1153 * @state: the mask of task states that can be woken
1154 * @sync: do a synchronous wakeup?
1156 * Put it on the run-queue if it's not already there. The "current"
1157 * thread is always on the run-queue (except when the actual
1158 * re-schedule is in progress), and as such you're allowed to do
1159 * the simpler "current->state = TASK_RUNNING" to mark yourself
1160 * runnable without the overhead of this.
1162 * returns failure only if the task is already active.
1164 static int try_to_wake_up(task_t *p, unsigned int state, int sync)
1166 int cpu, this_cpu, success = 0;
1167 unsigned long flags;
1171 unsigned long load, this_load;
1172 struct sched_domain *sd, *this_sd = NULL;
1176 rq = task_rq_lock(p, &flags);
1177 old_state = p->state;
1178 if (!(old_state & state))
1185 this_cpu = smp_processor_id();
1188 if (unlikely(task_running(rq, p)))
1193 schedstat_inc(rq, ttwu_cnt);
1194 if (cpu == this_cpu) {
1195 schedstat_inc(rq, ttwu_local);
1199 for_each_domain(this_cpu, sd) {
1200 if (cpu_isset(cpu, sd->span)) {
1201 schedstat_inc(sd, ttwu_wake_remote);
1207 if (p->last_waker_cpu != this_cpu)
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();
1280 p->last_waker_cpu = this_cpu;
1283 #endif /* CONFIG_SMP */
1284 if (old_state == TASK_UNINTERRUPTIBLE) {
1285 rq->nr_uninterruptible--;
1287 * Tasks on involuntary sleep don't earn
1288 * sleep_avg beyond just interactive state.
1294 * Tasks that have marked their sleep as noninteractive get
1295 * woken up without updating their sleep average. (i.e. their
1296 * sleep is handled in a priority-neutral manner, no priority
1297 * boost and no penalty.)
1299 if (old_state & TASK_NONINTERACTIVE)
1300 __activate_task(p, rq);
1302 activate_task(p, rq, cpu == this_cpu);
1304 * Sync wakeups (i.e. those types of wakeups where the waker
1305 * has indicated that it will leave the CPU in short order)
1306 * don't trigger a preemption, if the woken up task will run on
1307 * this cpu. (in this case the 'I will reschedule' promise of
1308 * the waker guarantees that the freshly woken up task is going
1309 * to be considered on this CPU.)
1311 if (!sync || cpu != this_cpu) {
1312 if (TASK_PREEMPTS_CURR(p, rq))
1313 resched_task(rq->curr);
1318 p->state = TASK_RUNNING;
1320 task_rq_unlock(rq, &flags);
1325 int fastcall wake_up_process(task_t *p)
1327 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1328 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1331 EXPORT_SYMBOL(wake_up_process);
1333 int fastcall wake_up_state(task_t *p, unsigned int state)
1335 return try_to_wake_up(p, state, 0);
1339 * Perform scheduler related setup for a newly forked process p.
1340 * p is forked by current.
1342 void fastcall sched_fork(task_t *p, int clone_flags)
1344 int cpu = get_cpu();
1347 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1349 set_task_cpu(p, cpu);
1352 * We mark the process as running here, but have not actually
1353 * inserted it onto the runqueue yet. This guarantees that
1354 * nobody will actually run it, and a signal or other external
1355 * event cannot wake it up and insert it on the runqueue either.
1357 p->state = TASK_RUNNING;
1358 INIT_LIST_HEAD(&p->run_list);
1360 #ifdef CONFIG_SCHEDSTATS
1361 memset(&p->sched_info, 0, sizeof(p->sched_info));
1363 #if defined(CONFIG_SMP)
1364 p->last_waker_cpu = cpu;
1365 #if defined(__ARCH_WANT_UNLOCKED_CTXSW)
1369 #ifdef CONFIG_PREEMPT
1370 /* Want to start with kernel preemption disabled. */
1371 task_thread_info(p)->preempt_count = 1;
1374 * Share the timeslice between parent and child, thus the
1375 * total amount of pending timeslices in the system doesn't change,
1376 * resulting in more scheduling fairness.
1378 local_irq_disable();
1379 p->time_slice = (current->time_slice + 1) >> 1;
1381 * The remainder of the first timeslice might be recovered by
1382 * the parent if the child exits early enough.
1384 p->first_time_slice = 1;
1385 current->time_slice >>= 1;
1386 p->timestamp = sched_clock();
1387 if (unlikely(!current->time_slice)) {
1389 * This case is rare, it happens when the parent has only
1390 * a single jiffy left from its timeslice. Taking the
1391 * runqueue lock is not a problem.
1393 current->time_slice = 1;
1401 * wake_up_new_task - wake up a newly created task for the first time.
1403 * This function will do some initial scheduler statistics housekeeping
1404 * that must be done for every newly created context, then puts the task
1405 * on the runqueue and wakes it.
1407 void fastcall wake_up_new_task(task_t *p, unsigned long clone_flags)
1409 unsigned long flags;
1411 runqueue_t *rq, *this_rq;
1413 rq = task_rq_lock(p, &flags);
1414 BUG_ON(p->state != TASK_RUNNING);
1415 this_cpu = smp_processor_id();
1419 * We decrease the sleep average of forking parents
1420 * and children as well, to keep max-interactive tasks
1421 * from forking tasks that are max-interactive. The parent
1422 * (current) is done further down, under its lock.
1424 p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1425 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1427 p->prio = effective_prio(p);
1429 if (likely(cpu == this_cpu)) {
1430 if (!(clone_flags & CLONE_VM)) {
1432 * The VM isn't cloned, so we're in a good position to
1433 * do child-runs-first in anticipation of an exec. This
1434 * usually avoids a lot of COW overhead.
1436 if (unlikely(!current->array))
1437 __activate_task(p, rq);
1439 p->prio = current->prio;
1440 list_add_tail(&p->run_list, ¤t->run_list);
1441 p->array = current->array;
1442 p->array->nr_active++;
1447 /* Run child last */
1448 __activate_task(p, rq);
1450 * We skip the following code due to cpu == this_cpu
1452 * task_rq_unlock(rq, &flags);
1453 * this_rq = task_rq_lock(current, &flags);
1457 this_rq = cpu_rq(this_cpu);
1460 * Not the local CPU - must adjust timestamp. This should
1461 * get optimised away in the !CONFIG_SMP case.
1463 p->timestamp = (p->timestamp - this_rq->timestamp_last_tick)
1464 + rq->timestamp_last_tick;
1465 __activate_task(p, rq);
1466 if (TASK_PREEMPTS_CURR(p, rq))
1467 resched_task(rq->curr);
1470 * Parent and child are on different CPUs, now get the
1471 * parent runqueue to update the parent's ->sleep_avg:
1473 task_rq_unlock(rq, &flags);
1474 this_rq = task_rq_lock(current, &flags);
1476 current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1477 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1478 task_rq_unlock(this_rq, &flags);
1482 * Potentially available exiting-child timeslices are
1483 * retrieved here - this way the parent does not get
1484 * penalized for creating too many threads.
1486 * (this cannot be used to 'generate' timeslices
1487 * artificially, because any timeslice recovered here
1488 * was given away by the parent in the first place.)
1490 void fastcall sched_exit(task_t *p)
1492 unsigned long flags;
1496 * If the child was a (relative-) CPU hog then decrease
1497 * the sleep_avg of the parent as well.
1499 rq = task_rq_lock(p->parent, &flags);
1500 if (p->first_time_slice && task_cpu(p) == task_cpu(p->parent)) {
1501 p->parent->time_slice += p->time_slice;
1502 if (unlikely(p->parent->time_slice > task_timeslice(p)))
1503 p->parent->time_slice = task_timeslice(p);
1505 if (p->sleep_avg < p->parent->sleep_avg)
1506 p->parent->sleep_avg = p->parent->sleep_avg /
1507 (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg /
1509 task_rq_unlock(rq, &flags);
1513 * prepare_task_switch - prepare to switch tasks
1514 * @rq: the runqueue preparing to switch
1515 * @next: the task we are going to switch to.
1517 * This is called with the rq lock held and interrupts off. It must
1518 * be paired with a subsequent finish_task_switch after the context
1521 * prepare_task_switch sets up locking and calls architecture specific
1524 static inline void prepare_task_switch(runqueue_t *rq, task_t *next)
1526 prepare_lock_switch(rq, next);
1527 prepare_arch_switch(next);
1531 * finish_task_switch - clean up after a task-switch
1532 * @rq: runqueue associated with task-switch
1533 * @prev: the thread we just switched away from.
1535 * finish_task_switch must be called after the context switch, paired
1536 * with a prepare_task_switch call before the context switch.
1537 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1538 * and do any other architecture-specific cleanup actions.
1540 * Note that we may have delayed dropping an mm in context_switch(). If
1541 * so, we finish that here outside of the runqueue lock. (Doing it
1542 * with the lock held can cause deadlocks; see schedule() for
1545 static inline void finish_task_switch(runqueue_t *rq, task_t *prev)
1546 __releases(rq->lock)
1548 struct mm_struct *mm = rq->prev_mm;
1549 unsigned long prev_task_flags;
1554 * A task struct has one reference for the use as "current".
1555 * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1556 * calls schedule one last time. The schedule call will never return,
1557 * and the scheduled task must drop that reference.
1558 * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1559 * still held, otherwise prev could be scheduled on another cpu, die
1560 * there before we look at prev->state, and then the reference would
1562 * Manfred Spraul <manfred@colorfullife.com>
1564 prev_task_flags = prev->flags;
1565 finish_arch_switch(prev);
1566 finish_lock_switch(rq, prev);
1569 if (unlikely(prev_task_flags & PF_DEAD))
1570 put_task_struct(prev);
1574 * schedule_tail - first thing a freshly forked thread must call.
1575 * @prev: the thread we just switched away from.
1577 asmlinkage void schedule_tail(task_t *prev)
1578 __releases(rq->lock)
1580 runqueue_t *rq = this_rq();
1581 finish_task_switch(rq, prev);
1582 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1583 /* In this case, finish_task_switch does not reenable preemption */
1586 if (current->set_child_tid)
1587 put_user(current->pid, current->set_child_tid);
1591 * context_switch - switch to the new MM and the new
1592 * thread's register state.
1595 task_t * context_switch(runqueue_t *rq, task_t *prev, task_t *next)
1597 struct mm_struct *mm = next->mm;
1598 struct mm_struct *oldmm = prev->active_mm;
1600 if (unlikely(!mm)) {
1601 next->active_mm = oldmm;
1602 atomic_inc(&oldmm->mm_count);
1603 enter_lazy_tlb(oldmm, next);
1605 switch_mm(oldmm, mm, next);
1607 if (unlikely(!prev->mm)) {
1608 prev->active_mm = NULL;
1609 WARN_ON(rq->prev_mm);
1610 rq->prev_mm = oldmm;
1613 /* Here we just switch the register state and the stack. */
1614 switch_to(prev, next, prev);
1620 * nr_running, nr_uninterruptible and nr_context_switches:
1622 * externally visible scheduler statistics: current number of runnable
1623 * threads, current number of uninterruptible-sleeping threads, total
1624 * number of context switches performed since bootup.
1626 unsigned long nr_running(void)
1628 unsigned long i, sum = 0;
1630 for_each_online_cpu(i)
1631 sum += cpu_rq(i)->nr_running;
1636 unsigned long nr_uninterruptible(void)
1638 unsigned long i, sum = 0;
1641 sum += cpu_rq(i)->nr_uninterruptible;
1644 * Since we read the counters lockless, it might be slightly
1645 * inaccurate. Do not allow it to go below zero though:
1647 if (unlikely((long)sum < 0))
1653 unsigned long long nr_context_switches(void)
1655 unsigned long long i, sum = 0;
1658 sum += cpu_rq(i)->nr_switches;
1663 unsigned long nr_iowait(void)
1665 unsigned long i, sum = 0;
1668 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1676 * double_rq_lock - safely lock two runqueues
1678 * Note this does not disable interrupts like task_rq_lock,
1679 * you need to do so manually before calling.
1681 static void double_rq_lock(runqueue_t *rq1, runqueue_t *rq2)
1682 __acquires(rq1->lock)
1683 __acquires(rq2->lock)
1686 spin_lock(&rq1->lock);
1687 __acquire(rq2->lock); /* Fake it out ;) */
1690 spin_lock(&rq1->lock);
1691 spin_lock(&rq2->lock);
1693 spin_lock(&rq2->lock);
1694 spin_lock(&rq1->lock);
1700 * double_rq_unlock - safely unlock two runqueues
1702 * Note this does not restore interrupts like task_rq_unlock,
1703 * you need to do so manually after calling.
1705 static void double_rq_unlock(runqueue_t *rq1, runqueue_t *rq2)
1706 __releases(rq1->lock)
1707 __releases(rq2->lock)
1709 spin_unlock(&rq1->lock);
1711 spin_unlock(&rq2->lock);
1713 __release(rq2->lock);
1717 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1719 static void double_lock_balance(runqueue_t *this_rq, runqueue_t *busiest)
1720 __releases(this_rq->lock)
1721 __acquires(busiest->lock)
1722 __acquires(this_rq->lock)
1724 if (unlikely(!spin_trylock(&busiest->lock))) {
1725 if (busiest < this_rq) {
1726 spin_unlock(&this_rq->lock);
1727 spin_lock(&busiest->lock);
1728 spin_lock(&this_rq->lock);
1730 spin_lock(&busiest->lock);
1735 * If dest_cpu is allowed for this process, migrate the task to it.
1736 * This is accomplished by forcing the cpu_allowed mask to only
1737 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
1738 * the cpu_allowed mask is restored.
1740 static void sched_migrate_task(task_t *p, int dest_cpu)
1742 migration_req_t req;
1744 unsigned long flags;
1746 rq = task_rq_lock(p, &flags);
1747 if (!cpu_isset(dest_cpu, p->cpus_allowed)
1748 || unlikely(cpu_is_offline(dest_cpu)))
1751 /* force the process onto the specified CPU */
1752 if (migrate_task(p, dest_cpu, &req)) {
1753 /* Need to wait for migration thread (might exit: take ref). */
1754 struct task_struct *mt = rq->migration_thread;
1755 get_task_struct(mt);
1756 task_rq_unlock(rq, &flags);
1757 wake_up_process(mt);
1758 put_task_struct(mt);
1759 wait_for_completion(&req.done);
1763 task_rq_unlock(rq, &flags);
1767 * sched_exec - execve() is a valuable balancing opportunity, because at
1768 * this point the task has the smallest effective memory and cache footprint.
1770 void sched_exec(void)
1772 int new_cpu, this_cpu = get_cpu();
1773 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
1775 if (new_cpu != this_cpu)
1776 sched_migrate_task(current, new_cpu);
1780 * pull_task - move a task from a remote runqueue to the local runqueue.
1781 * Both runqueues must be locked.
1784 void pull_task(runqueue_t *src_rq, prio_array_t *src_array, task_t *p,
1785 runqueue_t *this_rq, prio_array_t *this_array, int this_cpu)
1787 dequeue_task(p, src_array);
1788 src_rq->nr_running--;
1789 set_task_cpu(p, this_cpu);
1790 this_rq->nr_running++;
1791 enqueue_task(p, this_array);
1792 p->timestamp = (p->timestamp - src_rq->timestamp_last_tick)
1793 + this_rq->timestamp_last_tick;
1795 * Note that idle threads have a prio of MAX_PRIO, for this test
1796 * to be always true for them.
1798 if (TASK_PREEMPTS_CURR(p, this_rq))
1799 resched_task(this_rq->curr);
1803 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1806 int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu,
1807 struct sched_domain *sd, enum idle_type idle,
1811 * We do not migrate tasks that are:
1812 * 1) running (obviously), or
1813 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1814 * 3) are cache-hot on their current CPU.
1816 if (!cpu_isset(this_cpu, p->cpus_allowed))
1820 if (task_running(rq, p))
1824 * Aggressive migration if:
1825 * 1) task is cache cold, or
1826 * 2) too many balance attempts have failed.
1829 if (sd->nr_balance_failed > sd->cache_nice_tries)
1832 if (task_hot(p, rq->timestamp_last_tick, sd))
1838 * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq,
1839 * as part of a balancing operation within "domain". Returns the number of
1842 * Called with both runqueues locked.
1844 static int move_tasks(runqueue_t *this_rq, int this_cpu, runqueue_t *busiest,
1845 unsigned long max_nr_move, struct sched_domain *sd,
1846 enum idle_type idle, int *all_pinned)
1848 prio_array_t *array, *dst_array;
1849 struct list_head *head, *curr;
1850 int idx, pulled = 0, pinned = 0;
1853 if (max_nr_move == 0)
1859 * We first consider expired tasks. Those will likely not be
1860 * executed in the near future, and they are most likely to
1861 * be cache-cold, thus switching CPUs has the least effect
1864 if (busiest->expired->nr_active) {
1865 array = busiest->expired;
1866 dst_array = this_rq->expired;
1868 array = busiest->active;
1869 dst_array = this_rq->active;
1873 /* Start searching at priority 0: */
1877 idx = sched_find_first_bit(array->bitmap);
1879 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
1880 if (idx >= MAX_PRIO) {
1881 if (array == busiest->expired && busiest->active->nr_active) {
1882 array = busiest->active;
1883 dst_array = this_rq->active;
1889 head = array->queue + idx;
1892 tmp = list_entry(curr, task_t, run_list);
1896 if (!can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
1903 #ifdef CONFIG_SCHEDSTATS
1904 if (task_hot(tmp, busiest->timestamp_last_tick, sd))
1905 schedstat_inc(sd, lb_hot_gained[idle]);
1908 pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
1911 /* We only want to steal up to the prescribed number of tasks. */
1912 if (pulled < max_nr_move) {
1920 * Right now, this is the only place pull_task() is called,
1921 * so we can safely collect pull_task() stats here rather than
1922 * inside pull_task().
1924 schedstat_add(sd, lb_gained[idle], pulled);
1927 *all_pinned = pinned;
1932 * find_busiest_group finds and returns the busiest CPU group within the
1933 * domain. It calculates and returns the number of tasks which should be
1934 * moved to restore balance via the imbalance parameter.
1936 static struct sched_group *
1937 find_busiest_group(struct sched_domain *sd, int this_cpu,
1938 unsigned long *imbalance, enum idle_type idle, int *sd_idle)
1940 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
1941 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
1942 unsigned long max_pull;
1945 max_load = this_load = total_load = total_pwr = 0;
1946 if (idle == NOT_IDLE)
1947 load_idx = sd->busy_idx;
1948 else if (idle == NEWLY_IDLE)
1949 load_idx = sd->newidle_idx;
1951 load_idx = sd->idle_idx;
1958 local_group = cpu_isset(this_cpu, group->cpumask);
1960 /* Tally up the load of all CPUs in the group */
1963 for_each_cpu_mask(i, group->cpumask) {
1964 if (*sd_idle && !idle_cpu(i))
1967 /* Bias balancing toward cpus of our domain */
1969 load = target_load(i, load_idx);
1971 load = source_load(i, load_idx);
1976 total_load += avg_load;
1977 total_pwr += group->cpu_power;
1979 /* Adjust by relative CPU power of the group */
1980 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1983 this_load = avg_load;
1985 } else if (avg_load > max_load) {
1986 max_load = avg_load;
1989 group = group->next;
1990 } while (group != sd->groups);
1992 if (!busiest || this_load >= max_load || max_load <= SCHED_LOAD_SCALE)
1995 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
1997 if (this_load >= avg_load ||
1998 100*max_load <= sd->imbalance_pct*this_load)
2002 * We're trying to get all the cpus to the average_load, so we don't
2003 * want to push ourselves above the average load, nor do we wish to
2004 * reduce the max loaded cpu below the average load, as either of these
2005 * actions would just result in more rebalancing later, and ping-pong
2006 * tasks around. Thus we look for the minimum possible imbalance.
2007 * Negative imbalances (*we* are more loaded than anyone else) will
2008 * be counted as no imbalance for these purposes -- we can't fix that
2009 * by pulling tasks to us. Be careful of negative numbers as they'll
2010 * appear as very large values with unsigned longs.
2013 /* Don't want to pull so many tasks that a group would go idle */
2014 max_pull = min(max_load - avg_load, max_load - SCHED_LOAD_SCALE);
2016 /* How much load to actually move to equalise the imbalance */
2017 *imbalance = min(max_pull * busiest->cpu_power,
2018 (avg_load - this_load) * this->cpu_power)
2021 if (*imbalance < SCHED_LOAD_SCALE) {
2022 unsigned long pwr_now = 0, pwr_move = 0;
2025 if (max_load - this_load >= SCHED_LOAD_SCALE*2) {
2031 * OK, we don't have enough imbalance to justify moving tasks,
2032 * however we may be able to increase total CPU power used by
2036 pwr_now += busiest->cpu_power*min(SCHED_LOAD_SCALE, max_load);
2037 pwr_now += this->cpu_power*min(SCHED_LOAD_SCALE, this_load);
2038 pwr_now /= SCHED_LOAD_SCALE;
2040 /* Amount of load we'd subtract */
2041 tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/busiest->cpu_power;
2043 pwr_move += busiest->cpu_power*min(SCHED_LOAD_SCALE,
2046 /* Amount of load we'd add */
2047 if (max_load*busiest->cpu_power <
2048 SCHED_LOAD_SCALE*SCHED_LOAD_SCALE)
2049 tmp = max_load*busiest->cpu_power/this->cpu_power;
2051 tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/this->cpu_power;
2052 pwr_move += this->cpu_power*min(SCHED_LOAD_SCALE, this_load + tmp);
2053 pwr_move /= SCHED_LOAD_SCALE;
2055 /* Move if we gain throughput */
2056 if (pwr_move <= pwr_now)
2063 /* Get rid of the scaling factor, rounding down as we divide */
2064 *imbalance = *imbalance / SCHED_LOAD_SCALE;
2074 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2076 static runqueue_t *find_busiest_queue(struct sched_group *group,
2077 enum idle_type idle)
2079 unsigned long load, max_load = 0;
2080 runqueue_t *busiest = NULL;
2083 for_each_cpu_mask(i, group->cpumask) {
2084 load = source_load(i, 0);
2086 if (load > max_load) {
2088 busiest = cpu_rq(i);
2096 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2097 * so long as it is large enough.
2099 #define MAX_PINNED_INTERVAL 512
2102 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2103 * tasks if there is an imbalance.
2105 * Called with this_rq unlocked.
2107 static int load_balance(int this_cpu, runqueue_t *this_rq,
2108 struct sched_domain *sd, enum idle_type idle)
2110 struct sched_group *group;
2111 runqueue_t *busiest;
2112 unsigned long imbalance;
2113 int nr_moved, all_pinned = 0;
2114 int active_balance = 0;
2117 if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER)
2120 schedstat_inc(sd, lb_cnt[idle]);
2122 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle);
2124 schedstat_inc(sd, lb_nobusyg[idle]);
2128 busiest = find_busiest_queue(group, idle);
2130 schedstat_inc(sd, lb_nobusyq[idle]);
2134 BUG_ON(busiest == this_rq);
2136 schedstat_add(sd, lb_imbalance[idle], imbalance);
2139 if (busiest->nr_running > 1) {
2141 * Attempt to move tasks. If find_busiest_group has found
2142 * an imbalance but busiest->nr_running <= 1, the group is
2143 * still unbalanced. nr_moved simply stays zero, so it is
2144 * correctly treated as an imbalance.
2146 double_rq_lock(this_rq, busiest);
2147 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2148 imbalance, sd, idle, &all_pinned);
2149 double_rq_unlock(this_rq, busiest);
2151 /* All tasks on this runqueue were pinned by CPU affinity */
2152 if (unlikely(all_pinned))
2157 schedstat_inc(sd, lb_failed[idle]);
2158 sd->nr_balance_failed++;
2160 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2162 spin_lock(&busiest->lock);
2164 /* don't kick the migration_thread, if the curr
2165 * task on busiest cpu can't be moved to this_cpu
2167 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2168 spin_unlock(&busiest->lock);
2170 goto out_one_pinned;
2173 if (!busiest->active_balance) {
2174 busiest->active_balance = 1;
2175 busiest->push_cpu = this_cpu;
2178 spin_unlock(&busiest->lock);
2180 wake_up_process(busiest->migration_thread);
2183 * We've kicked active balancing, reset the failure
2186 sd->nr_balance_failed = sd->cache_nice_tries+1;
2189 sd->nr_balance_failed = 0;
2191 if (likely(!active_balance)) {
2192 /* We were unbalanced, so reset the balancing interval */
2193 sd->balance_interval = sd->min_interval;
2196 * If we've begun active balancing, start to back off. This
2197 * case may not be covered by the all_pinned logic if there
2198 * is only 1 task on the busy runqueue (because we don't call
2201 if (sd->balance_interval < sd->max_interval)
2202 sd->balance_interval *= 2;
2205 if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2210 schedstat_inc(sd, lb_balanced[idle]);
2212 sd->nr_balance_failed = 0;
2215 /* tune up the balancing interval */
2216 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2217 (sd->balance_interval < sd->max_interval))
2218 sd->balance_interval *= 2;
2220 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2226 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2227 * tasks if there is an imbalance.
2229 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2230 * this_rq is locked.
2232 static int load_balance_newidle(int this_cpu, runqueue_t *this_rq,
2233 struct sched_domain *sd)
2235 struct sched_group *group;
2236 runqueue_t *busiest = NULL;
2237 unsigned long imbalance;
2241 if (sd->flags & SD_SHARE_CPUPOWER)
2244 schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
2245 group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE, &sd_idle);
2247 schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]);
2251 busiest = find_busiest_queue(group, NEWLY_IDLE);
2253 schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]);
2257 BUG_ON(busiest == this_rq);
2259 schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
2262 if (busiest->nr_running > 1) {
2263 /* Attempt to move tasks */
2264 double_lock_balance(this_rq, busiest);
2265 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2266 imbalance, sd, NEWLY_IDLE, NULL);
2267 spin_unlock(&busiest->lock);
2271 schedstat_inc(sd, lb_failed[NEWLY_IDLE]);
2272 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2275 sd->nr_balance_failed = 0;
2280 schedstat_inc(sd, lb_balanced[NEWLY_IDLE]);
2281 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2283 sd->nr_balance_failed = 0;
2288 * idle_balance is called by schedule() if this_cpu is about to become
2289 * idle. Attempts to pull tasks from other CPUs.
2291 static void idle_balance(int this_cpu, runqueue_t *this_rq)
2293 struct sched_domain *sd;
2295 for_each_domain(this_cpu, sd) {
2296 if (sd->flags & SD_BALANCE_NEWIDLE) {
2297 if (load_balance_newidle(this_cpu, this_rq, sd)) {
2298 /* We've pulled tasks over so stop searching */
2306 * active_load_balance is run by migration threads. It pushes running tasks
2307 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2308 * running on each physical CPU where possible, and avoids physical /
2309 * logical imbalances.
2311 * Called with busiest_rq locked.
2313 static void active_load_balance(runqueue_t *busiest_rq, int busiest_cpu)
2315 struct sched_domain *sd;
2316 runqueue_t *target_rq;
2317 int target_cpu = busiest_rq->push_cpu;
2319 if (busiest_rq->nr_running <= 1)
2320 /* no task to move */
2323 target_rq = cpu_rq(target_cpu);
2326 * This condition is "impossible", if it occurs
2327 * we need to fix it. Originally reported by
2328 * Bjorn Helgaas on a 128-cpu setup.
2330 BUG_ON(busiest_rq == target_rq);
2332 /* move a task from busiest_rq to target_rq */
2333 double_lock_balance(busiest_rq, target_rq);
2335 /* Search for an sd spanning us and the target CPU. */
2336 for_each_domain(target_cpu, sd)
2337 if ((sd->flags & SD_LOAD_BALANCE) &&
2338 cpu_isset(busiest_cpu, sd->span))
2341 if (unlikely(sd == NULL))
2344 schedstat_inc(sd, alb_cnt);
2346 if (move_tasks(target_rq, target_cpu, busiest_rq, 1, sd, SCHED_IDLE, NULL))
2347 schedstat_inc(sd, alb_pushed);
2349 schedstat_inc(sd, alb_failed);
2351 spin_unlock(&target_rq->lock);
2355 * rebalance_tick will get called every timer tick, on every CPU.
2357 * It checks each scheduling domain to see if it is due to be balanced,
2358 * and initiates a balancing operation if so.
2360 * Balancing parameters are set up in arch_init_sched_domains.
2363 /* Don't have all balancing operations going off at once */
2364 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2366 static void rebalance_tick(int this_cpu, runqueue_t *this_rq,
2367 enum idle_type idle)
2369 unsigned long old_load, this_load;
2370 unsigned long j = jiffies + CPU_OFFSET(this_cpu);
2371 struct sched_domain *sd;
2374 this_load = this_rq->nr_running * SCHED_LOAD_SCALE;
2375 /* Update our load */
2376 for (i = 0; i < 3; i++) {
2377 unsigned long new_load = this_load;
2379 old_load = this_rq->cpu_load[i];
2381 * Round up the averaging division if load is increasing. This
2382 * prevents us from getting stuck on 9 if the load is 10, for
2385 if (new_load > old_load)
2386 new_load += scale-1;
2387 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale;
2390 for_each_domain(this_cpu, sd) {
2391 unsigned long interval;
2393 if (!(sd->flags & SD_LOAD_BALANCE))
2396 interval = sd->balance_interval;
2397 if (idle != SCHED_IDLE)
2398 interval *= sd->busy_factor;
2400 /* scale ms to jiffies */
2401 interval = msecs_to_jiffies(interval);
2402 if (unlikely(!interval))
2405 if (j - sd->last_balance >= interval) {
2406 if (load_balance(this_cpu, this_rq, sd, idle)) {
2408 * We've pulled tasks over so either we're no
2409 * longer idle, or one of our SMT siblings is
2414 sd->last_balance += interval;
2420 * on UP we do not need to balance between CPUs:
2422 static inline void rebalance_tick(int cpu, runqueue_t *rq, enum idle_type idle)
2425 static inline void idle_balance(int cpu, runqueue_t *rq)
2430 static inline int wake_priority_sleeper(runqueue_t *rq)
2433 #ifdef CONFIG_SCHED_SMT
2434 spin_lock(&rq->lock);
2436 * If an SMT sibling task has been put to sleep for priority
2437 * reasons reschedule the idle task to see if it can now run.
2439 if (rq->nr_running) {
2440 resched_task(rq->idle);
2443 spin_unlock(&rq->lock);
2448 DEFINE_PER_CPU(struct kernel_stat, kstat);
2450 EXPORT_PER_CPU_SYMBOL(kstat);
2453 * This is called on clock ticks and on context switches.
2454 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2456 static inline void update_cpu_clock(task_t *p, runqueue_t *rq,
2457 unsigned long long now)
2459 unsigned long long last = max(p->timestamp, rq->timestamp_last_tick);
2460 p->sched_time += now - last;
2464 * Return current->sched_time plus any more ns on the sched_clock
2465 * that have not yet been banked.
2467 unsigned long long current_sched_time(const task_t *tsk)
2469 unsigned long long ns;
2470 unsigned long flags;
2471 local_irq_save(flags);
2472 ns = max(tsk->timestamp, task_rq(tsk)->timestamp_last_tick);
2473 ns = tsk->sched_time + (sched_clock() - ns);
2474 local_irq_restore(flags);
2479 * We place interactive tasks back into the active array, if possible.
2481 * To guarantee that this does not starve expired tasks we ignore the
2482 * interactivity of a task if the first expired task had to wait more
2483 * than a 'reasonable' amount of time. This deadline timeout is
2484 * load-dependent, as the frequency of array switched decreases with
2485 * increasing number of running tasks. We also ignore the interactivity
2486 * if a better static_prio task has expired:
2488 #define EXPIRED_STARVING(rq) \
2489 ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2490 (jiffies - (rq)->expired_timestamp >= \
2491 STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2492 ((rq)->curr->static_prio > (rq)->best_expired_prio))
2495 * Account user cpu time to a process.
2496 * @p: the process that the cpu time gets accounted to
2497 * @hardirq_offset: the offset to subtract from hardirq_count()
2498 * @cputime: the cpu time spent in user space since the last update
2500 void account_user_time(struct task_struct *p, cputime_t cputime)
2502 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2505 p->utime = cputime_add(p->utime, cputime);
2507 /* Add user time to cpustat. */
2508 tmp = cputime_to_cputime64(cputime);
2509 if (TASK_NICE(p) > 0)
2510 cpustat->nice = cputime64_add(cpustat->nice, tmp);
2512 cpustat->user = cputime64_add(cpustat->user, tmp);
2516 * Account system cpu time to a process.
2517 * @p: the process that the cpu time gets accounted to
2518 * @hardirq_offset: the offset to subtract from hardirq_count()
2519 * @cputime: the cpu time spent in kernel space since the last update
2521 void account_system_time(struct task_struct *p, int hardirq_offset,
2524 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2525 runqueue_t *rq = this_rq();
2528 p->stime = cputime_add(p->stime, cputime);
2530 /* Add system time to cpustat. */
2531 tmp = cputime_to_cputime64(cputime);
2532 if (hardirq_count() - hardirq_offset)
2533 cpustat->irq = cputime64_add(cpustat->irq, tmp);
2534 else if (softirq_count())
2535 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
2536 else if (p != rq->idle)
2537 cpustat->system = cputime64_add(cpustat->system, tmp);
2538 else if (atomic_read(&rq->nr_iowait) > 0)
2539 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2541 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2542 /* Account for system time used */
2543 acct_update_integrals(p);
2547 * Account for involuntary wait time.
2548 * @p: the process from which the cpu time has been stolen
2549 * @steal: the cpu time spent in involuntary wait
2551 void account_steal_time(struct task_struct *p, cputime_t steal)
2553 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2554 cputime64_t tmp = cputime_to_cputime64(steal);
2555 runqueue_t *rq = this_rq();
2557 if (p == rq->idle) {
2558 p->stime = cputime_add(p->stime, steal);
2559 if (atomic_read(&rq->nr_iowait) > 0)
2560 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2562 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2564 cpustat->steal = cputime64_add(cpustat->steal, tmp);
2568 * This function gets called by the timer code, with HZ frequency.
2569 * We call it with interrupts disabled.
2571 * It also gets called by the fork code, when changing the parent's
2574 void scheduler_tick(void)
2576 int cpu = smp_processor_id();
2577 runqueue_t *rq = this_rq();
2578 task_t *p = current;
2579 unsigned long long now = sched_clock();
2581 update_cpu_clock(p, rq, now);
2583 rq->timestamp_last_tick = now;
2585 if (p == rq->idle) {
2586 if (wake_priority_sleeper(rq))
2588 rebalance_tick(cpu, rq, SCHED_IDLE);
2592 /* Task might have expired already, but not scheduled off yet */
2593 if (p->array != rq->active) {
2594 set_tsk_need_resched(p);
2597 spin_lock(&rq->lock);
2599 * The task was running during this tick - update the
2600 * time slice counter. Note: we do not update a thread's
2601 * priority until it either goes to sleep or uses up its
2602 * timeslice. This makes it possible for interactive tasks
2603 * to use up their timeslices at their highest priority levels.
2607 * RR tasks need a special form of timeslice management.
2608 * FIFO tasks have no timeslices.
2610 if ((p->policy == SCHED_RR) && !--p->time_slice) {
2611 p->time_slice = task_timeslice(p);
2612 p->first_time_slice = 0;
2613 set_tsk_need_resched(p);
2615 /* put it at the end of the queue: */
2616 requeue_task(p, rq->active);
2620 if (!--p->time_slice) {
2621 dequeue_task(p, rq->active);
2622 set_tsk_need_resched(p);
2623 p->prio = effective_prio(p);
2624 p->time_slice = task_timeslice(p);
2625 p->first_time_slice = 0;
2627 if (!rq->expired_timestamp)
2628 rq->expired_timestamp = jiffies;
2629 if (!TASK_INTERACTIVE(p) || EXPIRED_STARVING(rq)) {
2630 enqueue_task(p, rq->expired);
2631 if (p->static_prio < rq->best_expired_prio)
2632 rq->best_expired_prio = p->static_prio;
2634 enqueue_task(p, rq->active);
2637 * Prevent a too long timeslice allowing a task to monopolize
2638 * the CPU. We do this by splitting up the timeslice into
2641 * Note: this does not mean the task's timeslices expire or
2642 * get lost in any way, they just might be preempted by
2643 * another task of equal priority. (one with higher
2644 * priority would have preempted this task already.) We
2645 * requeue this task to the end of the list on this priority
2646 * level, which is in essence a round-robin of tasks with
2649 * This only applies to tasks in the interactive
2650 * delta range with at least TIMESLICE_GRANULARITY to requeue.
2652 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
2653 p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
2654 (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
2655 (p->array == rq->active)) {
2657 requeue_task(p, rq->active);
2658 set_tsk_need_resched(p);
2662 spin_unlock(&rq->lock);
2664 rebalance_tick(cpu, rq, NOT_IDLE);
2667 #ifdef CONFIG_SCHED_SMT
2668 static inline void wakeup_busy_runqueue(runqueue_t *rq)
2670 /* If an SMT runqueue is sleeping due to priority reasons wake it up */
2671 if (rq->curr == rq->idle && rq->nr_running)
2672 resched_task(rq->idle);
2675 static void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2677 struct sched_domain *tmp, *sd = NULL;
2678 cpumask_t sibling_map;
2681 for_each_domain(this_cpu, tmp)
2682 if (tmp->flags & SD_SHARE_CPUPOWER)
2689 * Unlock the current runqueue because we have to lock in
2690 * CPU order to avoid deadlocks. Caller knows that we might
2691 * unlock. We keep IRQs disabled.
2693 spin_unlock(&this_rq->lock);
2695 sibling_map = sd->span;
2697 for_each_cpu_mask(i, sibling_map)
2698 spin_lock(&cpu_rq(i)->lock);
2700 * We clear this CPU from the mask. This both simplifies the
2701 * inner loop and keps this_rq locked when we exit:
2703 cpu_clear(this_cpu, sibling_map);
2705 for_each_cpu_mask(i, sibling_map) {
2706 runqueue_t *smt_rq = cpu_rq(i);
2708 wakeup_busy_runqueue(smt_rq);
2711 for_each_cpu_mask(i, sibling_map)
2712 spin_unlock(&cpu_rq(i)->lock);
2714 * We exit with this_cpu's rq still held and IRQs
2720 * number of 'lost' timeslices this task wont be able to fully
2721 * utilize, if another task runs on a sibling. This models the
2722 * slowdown effect of other tasks running on siblings:
2724 static inline unsigned long smt_slice(task_t *p, struct sched_domain *sd)
2726 return p->time_slice * (100 - sd->per_cpu_gain) / 100;
2729 static int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2731 struct sched_domain *tmp, *sd = NULL;
2732 cpumask_t sibling_map;
2733 prio_array_t *array;
2737 for_each_domain(this_cpu, tmp)
2738 if (tmp->flags & SD_SHARE_CPUPOWER)
2745 * The same locking rules and details apply as for
2746 * wake_sleeping_dependent():
2748 spin_unlock(&this_rq->lock);
2749 sibling_map = sd->span;
2750 for_each_cpu_mask(i, sibling_map)
2751 spin_lock(&cpu_rq(i)->lock);
2752 cpu_clear(this_cpu, sibling_map);
2755 * Establish next task to be run - it might have gone away because
2756 * we released the runqueue lock above:
2758 if (!this_rq->nr_running)
2760 array = this_rq->active;
2761 if (!array->nr_active)
2762 array = this_rq->expired;
2763 BUG_ON(!array->nr_active);
2765 p = list_entry(array->queue[sched_find_first_bit(array->bitmap)].next,
2768 for_each_cpu_mask(i, sibling_map) {
2769 runqueue_t *smt_rq = cpu_rq(i);
2770 task_t *smt_curr = smt_rq->curr;
2772 /* Kernel threads do not participate in dependent sleeping */
2773 if (!p->mm || !smt_curr->mm || rt_task(p))
2774 goto check_smt_task;
2777 * If a user task with lower static priority than the
2778 * running task on the SMT sibling is trying to schedule,
2779 * delay it till there is proportionately less timeslice
2780 * left of the sibling task to prevent a lower priority
2781 * task from using an unfair proportion of the
2782 * physical cpu's resources. -ck
2784 if (rt_task(smt_curr)) {
2786 * With real time tasks we run non-rt tasks only
2787 * per_cpu_gain% of the time.
2789 if ((jiffies % DEF_TIMESLICE) >
2790 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
2793 if (smt_curr->static_prio < p->static_prio &&
2794 !TASK_PREEMPTS_CURR(p, smt_rq) &&
2795 smt_slice(smt_curr, sd) > task_timeslice(p))
2799 if ((!smt_curr->mm && smt_curr != smt_rq->idle) ||
2803 wakeup_busy_runqueue(smt_rq);
2808 * Reschedule a lower priority task on the SMT sibling for
2809 * it to be put to sleep, or wake it up if it has been put to
2810 * sleep for priority reasons to see if it should run now.
2813 if ((jiffies % DEF_TIMESLICE) >
2814 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
2815 resched_task(smt_curr);
2817 if (TASK_PREEMPTS_CURR(p, smt_rq) &&
2818 smt_slice(p, sd) > task_timeslice(smt_curr))
2819 resched_task(smt_curr);
2821 wakeup_busy_runqueue(smt_rq);
2825 for_each_cpu_mask(i, sibling_map)
2826 spin_unlock(&cpu_rq(i)->lock);
2830 static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2834 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2840 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
2842 void fastcall add_preempt_count(int val)
2847 BUG_ON((preempt_count() < 0));
2848 preempt_count() += val;
2850 * Spinlock count overflowing soon?
2852 BUG_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10);
2854 EXPORT_SYMBOL(add_preempt_count);
2856 void fastcall sub_preempt_count(int val)
2861 BUG_ON(val > preempt_count());
2863 * Is the spinlock portion underflowing?
2865 BUG_ON((val < PREEMPT_MASK) && !(preempt_count() & PREEMPT_MASK));
2866 preempt_count() -= val;
2868 EXPORT_SYMBOL(sub_preempt_count);
2873 * schedule() is the main scheduler function.
2875 asmlinkage void __sched schedule(void)
2878 task_t *prev, *next;
2880 prio_array_t *array;
2881 struct list_head *queue;
2882 unsigned long long now;
2883 unsigned long run_time;
2884 int cpu, idx, new_prio;
2887 * Test if we are atomic. Since do_exit() needs to call into
2888 * schedule() atomically, we ignore that path for now.
2889 * Otherwise, whine if we are scheduling when we should not be.
2891 if (likely(!current->exit_state)) {
2892 if (unlikely(in_atomic())) {
2893 printk(KERN_ERR "scheduling while atomic: "
2895 current->comm, preempt_count(), current->pid);
2899 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2904 release_kernel_lock(prev);
2905 need_resched_nonpreemptible:
2909 * The idle thread is not allowed to schedule!
2910 * Remove this check after it has been exercised a bit.
2912 if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
2913 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
2917 schedstat_inc(rq, sched_cnt);
2918 now = sched_clock();
2919 if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
2920 run_time = now - prev->timestamp;
2921 if (unlikely((long long)(now - prev->timestamp) < 0))
2924 run_time = NS_MAX_SLEEP_AVG;
2927 * Tasks charged proportionately less run_time at high sleep_avg to
2928 * delay them losing their interactive status
2930 run_time /= (CURRENT_BONUS(prev) ? : 1);
2932 spin_lock_irq(&rq->lock);
2934 if (unlikely(prev->flags & PF_DEAD))
2935 prev->state = EXIT_DEAD;
2937 switch_count = &prev->nivcsw;
2938 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2939 switch_count = &prev->nvcsw;
2940 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
2941 unlikely(signal_pending(prev))))
2942 prev->state = TASK_RUNNING;
2944 if (prev->state == TASK_UNINTERRUPTIBLE)
2945 rq->nr_uninterruptible++;
2946 deactivate_task(prev, rq);
2950 cpu = smp_processor_id();
2951 if (unlikely(!rq->nr_running)) {
2953 idle_balance(cpu, rq);
2954 if (!rq->nr_running) {
2956 rq->expired_timestamp = 0;
2957 wake_sleeping_dependent(cpu, rq);
2959 * wake_sleeping_dependent() might have released
2960 * the runqueue, so break out if we got new
2963 if (!rq->nr_running)
2967 if (dependent_sleeper(cpu, rq)) {
2972 * dependent_sleeper() releases and reacquires the runqueue
2973 * lock, hence go into the idle loop if the rq went
2976 if (unlikely(!rq->nr_running))
2981 if (unlikely(!array->nr_active)) {
2983 * Switch the active and expired arrays.
2985 schedstat_inc(rq, sched_switch);
2986 rq->active = rq->expired;
2987 rq->expired = array;
2989 rq->expired_timestamp = 0;
2990 rq->best_expired_prio = MAX_PRIO;
2993 idx = sched_find_first_bit(array->bitmap);
2994 queue = array->queue + idx;
2995 next = list_entry(queue->next, task_t, run_list);
2997 if (!rt_task(next) && next->activated > 0) {
2998 unsigned long long delta = now - next->timestamp;
2999 if (unlikely((long long)(now - next->timestamp) < 0))
3002 if (next->activated == 1)
3003 delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
3005 array = next->array;
3006 new_prio = recalc_task_prio(next, next->timestamp + delta);
3008 if (unlikely(next->prio != new_prio)) {
3009 dequeue_task(next, array);
3010 next->prio = new_prio;
3011 enqueue_task(next, array);
3013 requeue_task(next, array);
3015 next->activated = 0;
3017 if (next == rq->idle)
3018 schedstat_inc(rq, sched_goidle);
3020 prefetch_stack(next);
3021 clear_tsk_need_resched(prev);
3022 rcu_qsctr_inc(task_cpu(prev));
3024 update_cpu_clock(prev, rq, now);
3026 prev->sleep_avg -= run_time;
3027 if ((long)prev->sleep_avg <= 0)
3028 prev->sleep_avg = 0;
3029 prev->timestamp = prev->last_ran = now;
3031 sched_info_switch(prev, next);
3032 if (likely(prev != next)) {
3033 next->timestamp = now;
3038 prepare_task_switch(rq, next);
3039 prev = context_switch(rq, prev, next);
3042 * this_rq must be evaluated again because prev may have moved
3043 * CPUs since it called schedule(), thus the 'rq' on its stack
3044 * frame will be invalid.
3046 finish_task_switch(this_rq(), prev);
3048 spin_unlock_irq(&rq->lock);
3051 if (unlikely(reacquire_kernel_lock(prev) < 0))
3052 goto need_resched_nonpreemptible;
3053 preempt_enable_no_resched();
3054 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3058 EXPORT_SYMBOL(schedule);
3060 #ifdef CONFIG_PREEMPT
3062 * this is is the entry point to schedule() from in-kernel preemption
3063 * off of preempt_enable. Kernel preemptions off return from interrupt
3064 * occur there and call schedule directly.
3066 asmlinkage void __sched preempt_schedule(void)
3068 struct thread_info *ti = current_thread_info();
3069 #ifdef CONFIG_PREEMPT_BKL
3070 struct task_struct *task = current;
3071 int saved_lock_depth;
3074 * If there is a non-zero preempt_count or interrupts are disabled,
3075 * we do not want to preempt the current task. Just return..
3077 if (unlikely(ti->preempt_count || irqs_disabled()))
3081 add_preempt_count(PREEMPT_ACTIVE);
3083 * We keep the big kernel semaphore locked, but we
3084 * clear ->lock_depth so that schedule() doesnt
3085 * auto-release the semaphore:
3087 #ifdef CONFIG_PREEMPT_BKL
3088 saved_lock_depth = task->lock_depth;
3089 task->lock_depth = -1;
3092 #ifdef CONFIG_PREEMPT_BKL
3093 task->lock_depth = saved_lock_depth;
3095 sub_preempt_count(PREEMPT_ACTIVE);
3097 /* we could miss a preemption opportunity between schedule and now */
3099 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3103 EXPORT_SYMBOL(preempt_schedule);
3106 * this is is the entry point to schedule() from kernel preemption
3107 * off of irq context.
3108 * Note, that this is called and return with irqs disabled. This will
3109 * protect us against recursive calling from irq.
3111 asmlinkage void __sched preempt_schedule_irq(void)
3113 struct thread_info *ti = current_thread_info();
3114 #ifdef CONFIG_PREEMPT_BKL
3115 struct task_struct *task = current;
3116 int saved_lock_depth;
3118 /* Catch callers which need to be fixed*/
3119 BUG_ON(ti->preempt_count || !irqs_disabled());
3122 add_preempt_count(PREEMPT_ACTIVE);
3124 * We keep the big kernel semaphore locked, but we
3125 * clear ->lock_depth so that schedule() doesnt
3126 * auto-release the semaphore:
3128 #ifdef CONFIG_PREEMPT_BKL
3129 saved_lock_depth = task->lock_depth;
3130 task->lock_depth = -1;
3134 local_irq_disable();
3135 #ifdef CONFIG_PREEMPT_BKL
3136 task->lock_depth = saved_lock_depth;
3138 sub_preempt_count(PREEMPT_ACTIVE);
3140 /* we could miss a preemption opportunity between schedule and now */
3142 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3146 #endif /* CONFIG_PREEMPT */
3148 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3151 task_t *p = curr->private;
3152 return try_to_wake_up(p, mode, sync);
3155 EXPORT_SYMBOL(default_wake_function);
3158 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3159 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3160 * number) then we wake all the non-exclusive tasks and one exclusive task.
3162 * There are circumstances in which we can try to wake a task which has already
3163 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3164 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3166 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3167 int nr_exclusive, int sync, void *key)
3169 struct list_head *tmp, *next;
3171 list_for_each_safe(tmp, next, &q->task_list) {
3174 curr = list_entry(tmp, wait_queue_t, task_list);
3175 flags = curr->flags;
3176 if (curr->func(curr, mode, sync, key) &&
3177 (flags & WQ_FLAG_EXCLUSIVE) &&
3184 * __wake_up - wake up threads blocked on a waitqueue.
3186 * @mode: which threads
3187 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3188 * @key: is directly passed to the wakeup function
3190 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3191 int nr_exclusive, void *key)
3193 unsigned long flags;
3195 spin_lock_irqsave(&q->lock, flags);
3196 __wake_up_common(q, mode, nr_exclusive, 0, key);
3197 spin_unlock_irqrestore(&q->lock, flags);
3200 EXPORT_SYMBOL(__wake_up);
3203 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3205 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3207 __wake_up_common(q, mode, 1, 0, NULL);
3211 * __wake_up_sync - wake up threads blocked on a waitqueue.
3213 * @mode: which threads
3214 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3216 * The sync wakeup differs that the waker knows that it will schedule
3217 * away soon, so while the target thread will be woken up, it will not
3218 * be migrated to another CPU - ie. the two threads are 'synchronized'
3219 * with each other. This can prevent needless bouncing between CPUs.
3221 * On UP it can prevent extra preemption.
3224 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3226 unsigned long flags;
3232 if (unlikely(!nr_exclusive))
3235 spin_lock_irqsave(&q->lock, flags);
3236 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3237 spin_unlock_irqrestore(&q->lock, flags);
3239 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3241 void fastcall complete(struct completion *x)
3243 unsigned long flags;
3245 spin_lock_irqsave(&x->wait.lock, flags);
3247 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3249 spin_unlock_irqrestore(&x->wait.lock, flags);
3251 EXPORT_SYMBOL(complete);
3253 void fastcall complete_all(struct completion *x)
3255 unsigned long flags;
3257 spin_lock_irqsave(&x->wait.lock, flags);
3258 x->done += UINT_MAX/2;
3259 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3261 spin_unlock_irqrestore(&x->wait.lock, flags);
3263 EXPORT_SYMBOL(complete_all);
3265 void fastcall __sched wait_for_completion(struct completion *x)
3268 spin_lock_irq(&x->wait.lock);
3270 DECLARE_WAITQUEUE(wait, current);
3272 wait.flags |= WQ_FLAG_EXCLUSIVE;
3273 __add_wait_queue_tail(&x->wait, &wait);
3275 __set_current_state(TASK_UNINTERRUPTIBLE);
3276 spin_unlock_irq(&x->wait.lock);
3278 spin_lock_irq(&x->wait.lock);
3280 __remove_wait_queue(&x->wait, &wait);
3283 spin_unlock_irq(&x->wait.lock);
3285 EXPORT_SYMBOL(wait_for_completion);
3287 unsigned long fastcall __sched
3288 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3292 spin_lock_irq(&x->wait.lock);
3294 DECLARE_WAITQUEUE(wait, current);
3296 wait.flags |= WQ_FLAG_EXCLUSIVE;
3297 __add_wait_queue_tail(&x->wait, &wait);
3299 __set_current_state(TASK_UNINTERRUPTIBLE);
3300 spin_unlock_irq(&x->wait.lock);
3301 timeout = schedule_timeout(timeout);
3302 spin_lock_irq(&x->wait.lock);
3304 __remove_wait_queue(&x->wait, &wait);
3308 __remove_wait_queue(&x->wait, &wait);
3312 spin_unlock_irq(&x->wait.lock);
3315 EXPORT_SYMBOL(wait_for_completion_timeout);
3317 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3323 spin_lock_irq(&x->wait.lock);
3325 DECLARE_WAITQUEUE(wait, current);
3327 wait.flags |= WQ_FLAG_EXCLUSIVE;
3328 __add_wait_queue_tail(&x->wait, &wait);
3330 if (signal_pending(current)) {
3332 __remove_wait_queue(&x->wait, &wait);
3335 __set_current_state(TASK_INTERRUPTIBLE);
3336 spin_unlock_irq(&x->wait.lock);
3338 spin_lock_irq(&x->wait.lock);
3340 __remove_wait_queue(&x->wait, &wait);
3344 spin_unlock_irq(&x->wait.lock);
3348 EXPORT_SYMBOL(wait_for_completion_interruptible);
3350 unsigned long fastcall __sched
3351 wait_for_completion_interruptible_timeout(struct completion *x,
3352 unsigned long timeout)
3356 spin_lock_irq(&x->wait.lock);
3358 DECLARE_WAITQUEUE(wait, current);
3360 wait.flags |= WQ_FLAG_EXCLUSIVE;
3361 __add_wait_queue_tail(&x->wait, &wait);
3363 if (signal_pending(current)) {
3364 timeout = -ERESTARTSYS;
3365 __remove_wait_queue(&x->wait, &wait);
3368 __set_current_state(TASK_INTERRUPTIBLE);
3369 spin_unlock_irq(&x->wait.lock);
3370 timeout = schedule_timeout(timeout);
3371 spin_lock_irq(&x->wait.lock);
3373 __remove_wait_queue(&x->wait, &wait);
3377 __remove_wait_queue(&x->wait, &wait);
3381 spin_unlock_irq(&x->wait.lock);
3384 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3387 #define SLEEP_ON_VAR \
3388 unsigned long flags; \
3389 wait_queue_t wait; \
3390 init_waitqueue_entry(&wait, current);
3392 #define SLEEP_ON_HEAD \
3393 spin_lock_irqsave(&q->lock,flags); \
3394 __add_wait_queue(q, &wait); \
3395 spin_unlock(&q->lock);
3397 #define SLEEP_ON_TAIL \
3398 spin_lock_irq(&q->lock); \
3399 __remove_wait_queue(q, &wait); \
3400 spin_unlock_irqrestore(&q->lock, flags);
3402 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
3406 current->state = TASK_INTERRUPTIBLE;
3413 EXPORT_SYMBOL(interruptible_sleep_on);
3415 long fastcall __sched
3416 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3420 current->state = TASK_INTERRUPTIBLE;
3423 timeout = schedule_timeout(timeout);
3429 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3431 void fastcall __sched sleep_on(wait_queue_head_t *q)
3435 current->state = TASK_UNINTERRUPTIBLE;
3442 EXPORT_SYMBOL(sleep_on);
3444 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3448 current->state = TASK_UNINTERRUPTIBLE;
3451 timeout = schedule_timeout(timeout);
3457 EXPORT_SYMBOL(sleep_on_timeout);
3459 void set_user_nice(task_t *p, long nice)
3461 unsigned long flags;
3462 prio_array_t *array;
3464 int old_prio, new_prio, delta;
3466 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3469 * We have to be careful, if called from sys_setpriority(),
3470 * the task might be in the middle of scheduling on another CPU.
3472 rq = task_rq_lock(p, &flags);
3474 * The RT priorities are set via sched_setscheduler(), but we still
3475 * allow the 'normal' nice value to be set - but as expected
3476 * it wont have any effect on scheduling until the task is
3477 * not SCHED_NORMAL/SCHED_BATCH:
3480 p->static_prio = NICE_TO_PRIO(nice);
3485 dequeue_task(p, array);
3488 new_prio = NICE_TO_PRIO(nice);
3489 delta = new_prio - old_prio;
3490 p->static_prio = NICE_TO_PRIO(nice);
3494 enqueue_task(p, array);
3496 * If the task increased its priority or is running and
3497 * lowered its priority, then reschedule its CPU:
3499 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3500 resched_task(rq->curr);
3503 task_rq_unlock(rq, &flags);
3506 EXPORT_SYMBOL(set_user_nice);
3509 * can_nice - check if a task can reduce its nice value
3513 int can_nice(const task_t *p, const int nice)
3515 /* convert nice value [19,-20] to rlimit style value [1,40] */
3516 int nice_rlim = 20 - nice;
3517 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
3518 capable(CAP_SYS_NICE));
3521 #ifdef __ARCH_WANT_SYS_NICE
3524 * sys_nice - change the priority of the current process.
3525 * @increment: priority increment
3527 * sys_setpriority is a more generic, but much slower function that
3528 * does similar things.
3530 asmlinkage long sys_nice(int increment)
3536 * Setpriority might change our priority at the same moment.
3537 * We don't have to worry. Conceptually one call occurs first
3538 * and we have a single winner.
3540 if (increment < -40)
3545 nice = PRIO_TO_NICE(current->static_prio) + increment;
3551 if (increment < 0 && !can_nice(current, nice))
3554 retval = security_task_setnice(current, nice);
3558 set_user_nice(current, nice);
3565 * task_prio - return the priority value of a given task.
3566 * @p: the task in question.
3568 * This is the priority value as seen by users in /proc.
3569 * RT tasks are offset by -200. Normal tasks are centered
3570 * around 0, value goes from -16 to +15.
3572 int task_prio(const task_t *p)
3574 return p->prio - MAX_RT_PRIO;
3578 * task_nice - return the nice value of a given task.
3579 * @p: the task in question.
3581 int task_nice(const task_t *p)
3583 return TASK_NICE(p);
3585 EXPORT_SYMBOL_GPL(task_nice);
3588 * idle_cpu - is a given cpu idle currently?
3589 * @cpu: the processor in question.
3591 int idle_cpu(int cpu)
3593 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
3597 * idle_task - return the idle task for a given cpu.
3598 * @cpu: the processor in question.
3600 task_t *idle_task(int cpu)
3602 return cpu_rq(cpu)->idle;
3606 * find_process_by_pid - find a process with a matching PID value.
3607 * @pid: the pid in question.
3609 static inline task_t *find_process_by_pid(pid_t pid)
3611 return pid ? find_task_by_pid(pid) : current;
3614 /* Actually do priority change: must hold rq lock. */
3615 static void __setscheduler(struct task_struct *p, int policy, int prio)
3619 p->rt_priority = prio;
3620 if (policy != SCHED_NORMAL && policy != SCHED_BATCH) {
3621 p->prio = MAX_RT_PRIO-1 - p->rt_priority;
3623 p->prio = p->static_prio;
3625 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
3627 if (policy == SCHED_BATCH)
3633 * sched_setscheduler - change the scheduling policy and/or RT priority of
3635 * @p: the task in question.
3636 * @policy: new policy.
3637 * @param: structure containing the new RT priority.
3639 int sched_setscheduler(struct task_struct *p, int policy,
3640 struct sched_param *param)
3643 int oldprio, oldpolicy = -1;
3644 prio_array_t *array;
3645 unsigned long flags;
3649 /* double check policy once rq lock held */
3651 policy = oldpolicy = p->policy;
3652 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
3653 policy != SCHED_NORMAL && policy != SCHED_BATCH)
3656 * Valid priorities for SCHED_FIFO and SCHED_RR are
3657 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
3660 if (param->sched_priority < 0 ||
3661 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3662 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3664 if ((policy == SCHED_NORMAL || policy == SCHED_BATCH)
3665 != (param->sched_priority == 0))
3669 * Allow unprivileged RT tasks to decrease priority:
3671 if (!capable(CAP_SYS_NICE)) {
3673 * can't change policy, except between SCHED_NORMAL
3676 if (((policy != SCHED_NORMAL && p->policy != SCHED_BATCH) &&
3677 (policy != SCHED_BATCH && p->policy != SCHED_NORMAL)) &&
3678 !p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
3680 /* can't increase priority */
3681 if ((policy != SCHED_NORMAL && policy != SCHED_BATCH) &&
3682 param->sched_priority > p->rt_priority &&
3683 param->sched_priority >
3684 p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
3686 /* can't change other user's priorities */
3687 if ((current->euid != p->euid) &&
3688 (current->euid != p->uid))
3692 retval = security_task_setscheduler(p, policy, param);
3696 * To be able to change p->policy safely, the apropriate
3697 * runqueue lock must be held.
3699 rq = task_rq_lock(p, &flags);
3700 /* recheck policy now with rq lock held */
3701 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3702 policy = oldpolicy = -1;
3703 task_rq_unlock(rq, &flags);
3708 deactivate_task(p, rq);
3710 __setscheduler(p, policy, param->sched_priority);
3712 __activate_task(p, rq);
3714 * Reschedule if we are currently running on this runqueue and
3715 * our priority decreased, or if we are not currently running on
3716 * this runqueue and our priority is higher than the current's
3718 if (task_running(rq, p)) {
3719 if (p->prio > oldprio)
3720 resched_task(rq->curr);
3721 } else if (TASK_PREEMPTS_CURR(p, rq))
3722 resched_task(rq->curr);
3724 task_rq_unlock(rq, &flags);
3727 EXPORT_SYMBOL_GPL(sched_setscheduler);
3730 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3733 struct sched_param lparam;
3734 struct task_struct *p;
3736 if (!param || pid < 0)
3738 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3740 read_lock_irq(&tasklist_lock);
3741 p = find_process_by_pid(pid);
3743 read_unlock_irq(&tasklist_lock);
3746 retval = sched_setscheduler(p, policy, &lparam);
3747 read_unlock_irq(&tasklist_lock);
3752 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3753 * @pid: the pid in question.
3754 * @policy: new policy.
3755 * @param: structure containing the new RT priority.
3757 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
3758 struct sched_param __user *param)
3760 /* negative values for policy are not valid */
3764 return do_sched_setscheduler(pid, policy, param);
3768 * sys_sched_setparam - set/change the RT priority of a thread
3769 * @pid: the pid in question.
3770 * @param: structure containing the new RT priority.
3772 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
3774 return do_sched_setscheduler(pid, -1, param);
3778 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3779 * @pid: the pid in question.
3781 asmlinkage long sys_sched_getscheduler(pid_t pid)
3783 int retval = -EINVAL;
3790 read_lock(&tasklist_lock);
3791 p = find_process_by_pid(pid);
3793 retval = security_task_getscheduler(p);
3797 read_unlock(&tasklist_lock);
3804 * sys_sched_getscheduler - get the RT priority of a thread
3805 * @pid: the pid in question.
3806 * @param: structure containing the RT priority.
3808 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
3810 struct sched_param lp;
3811 int retval = -EINVAL;
3814 if (!param || pid < 0)
3817 read_lock(&tasklist_lock);
3818 p = find_process_by_pid(pid);
3823 retval = security_task_getscheduler(p);
3827 lp.sched_priority = p->rt_priority;
3828 read_unlock(&tasklist_lock);
3831 * This one might sleep, we cannot do it with a spinlock held ...
3833 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3839 read_unlock(&tasklist_lock);
3843 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
3847 cpumask_t cpus_allowed;
3850 read_lock(&tasklist_lock);
3852 p = find_process_by_pid(pid);
3854 read_unlock(&tasklist_lock);
3855 unlock_cpu_hotplug();
3860 * It is not safe to call set_cpus_allowed with the
3861 * tasklist_lock held. We will bump the task_struct's
3862 * usage count and then drop tasklist_lock.
3865 read_unlock(&tasklist_lock);
3868 if ((current->euid != p->euid) && (current->euid != p->uid) &&
3869 !capable(CAP_SYS_NICE))
3872 cpus_allowed = cpuset_cpus_allowed(p);
3873 cpus_and(new_mask, new_mask, cpus_allowed);
3874 retval = set_cpus_allowed(p, new_mask);
3878 unlock_cpu_hotplug();
3882 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3883 cpumask_t *new_mask)
3885 if (len < sizeof(cpumask_t)) {
3886 memset(new_mask, 0, sizeof(cpumask_t));
3887 } else if (len > sizeof(cpumask_t)) {
3888 len = sizeof(cpumask_t);
3890 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3894 * sys_sched_setaffinity - set the cpu affinity of a process
3895 * @pid: pid of the process
3896 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3897 * @user_mask_ptr: user-space pointer to the new cpu mask
3899 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
3900 unsigned long __user *user_mask_ptr)
3905 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
3909 return sched_setaffinity(pid, new_mask);
3913 * Represents all cpu's present in the system
3914 * In systems capable of hotplug, this map could dynamically grow
3915 * as new cpu's are detected in the system via any platform specific
3916 * method, such as ACPI for e.g.
3919 cpumask_t cpu_present_map __read_mostly;
3920 EXPORT_SYMBOL(cpu_present_map);
3923 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
3924 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
3927 long sched_getaffinity(pid_t pid, cpumask_t *mask)
3933 read_lock(&tasklist_lock);
3936 p = find_process_by_pid(pid);
3941 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
3944 read_unlock(&tasklist_lock);
3945 unlock_cpu_hotplug();
3953 * sys_sched_getaffinity - get the cpu affinity of a process
3954 * @pid: pid of the process
3955 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3956 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3958 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
3959 unsigned long __user *user_mask_ptr)
3964 if (len < sizeof(cpumask_t))
3967 ret = sched_getaffinity(pid, &mask);
3971 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
3974 return sizeof(cpumask_t);
3978 * sys_sched_yield - yield the current processor to other threads.
3980 * this function yields the current CPU by moving the calling thread
3981 * to the expired array. If there are no other threads running on this
3982 * CPU then this function will return.
3984 asmlinkage long sys_sched_yield(void)
3986 runqueue_t *rq = this_rq_lock();
3987 prio_array_t *array = current->array;
3988 prio_array_t *target = rq->expired;
3990 schedstat_inc(rq, yld_cnt);
3992 * We implement yielding by moving the task into the expired
3995 * (special rule: RT tasks will just roundrobin in the active
3998 if (rt_task(current))
3999 target = rq->active;
4001 if (array->nr_active == 1) {
4002 schedstat_inc(rq, yld_act_empty);
4003 if (!rq->expired->nr_active)
4004 schedstat_inc(rq, yld_both_empty);
4005 } else if (!rq->expired->nr_active)
4006 schedstat_inc(rq, yld_exp_empty);
4008 if (array != target) {
4009 dequeue_task(current, array);
4010 enqueue_task(current, target);
4013 * requeue_task is cheaper so perform that if possible.
4015 requeue_task(current, array);
4018 * Since we are going to call schedule() anyway, there's
4019 * no need to preempt or enable interrupts:
4021 __release(rq->lock);
4022 _raw_spin_unlock(&rq->lock);
4023 preempt_enable_no_resched();
4030 static inline void __cond_resched(void)
4033 * The BKS might be reacquired before we have dropped
4034 * PREEMPT_ACTIVE, which could trigger a second
4035 * cond_resched() call.
4037 if (unlikely(preempt_count()))
4040 add_preempt_count(PREEMPT_ACTIVE);
4042 sub_preempt_count(PREEMPT_ACTIVE);
4043 } while (need_resched());
4046 int __sched cond_resched(void)
4048 if (need_resched()) {
4055 EXPORT_SYMBOL(cond_resched);
4058 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4059 * call schedule, and on return reacquire the lock.
4061 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4062 * operations here to prevent schedule() from being called twice (once via
4063 * spin_unlock(), once by hand).
4065 int cond_resched_lock(spinlock_t *lock)
4069 if (need_lockbreak(lock)) {
4075 if (need_resched()) {
4076 _raw_spin_unlock(lock);
4077 preempt_enable_no_resched();
4085 EXPORT_SYMBOL(cond_resched_lock);
4087 int __sched cond_resched_softirq(void)
4089 BUG_ON(!in_softirq());
4091 if (need_resched()) {
4092 __local_bh_enable();
4100 EXPORT_SYMBOL(cond_resched_softirq);
4104 * yield - yield the current processor to other threads.
4106 * this is a shortcut for kernel-space yielding - it marks the
4107 * thread runnable and calls sys_sched_yield().
4109 void __sched yield(void)
4111 set_current_state(TASK_RUNNING);
4115 EXPORT_SYMBOL(yield);
4118 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4119 * that process accounting knows that this is a task in IO wait state.
4121 * But don't do that if it is a deliberate, throttling IO wait (this task
4122 * has set its backing_dev_info: the queue against which it should throttle)
4124 void __sched io_schedule(void)
4126 struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id());
4128 atomic_inc(&rq->nr_iowait);
4130 atomic_dec(&rq->nr_iowait);
4133 EXPORT_SYMBOL(io_schedule);
4135 long __sched io_schedule_timeout(long timeout)
4137 struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id());
4140 atomic_inc(&rq->nr_iowait);
4141 ret = schedule_timeout(timeout);
4142 atomic_dec(&rq->nr_iowait);
4147 * sys_sched_get_priority_max - return maximum RT priority.
4148 * @policy: scheduling class.
4150 * this syscall returns the maximum rt_priority that can be used
4151 * by a given scheduling class.
4153 asmlinkage long sys_sched_get_priority_max(int policy)
4160 ret = MAX_USER_RT_PRIO-1;
4171 * sys_sched_get_priority_min - return minimum RT priority.
4172 * @policy: scheduling class.
4174 * this syscall returns the minimum rt_priority that can be used
4175 * by a given scheduling class.
4177 asmlinkage long sys_sched_get_priority_min(int policy)
4194 * sys_sched_rr_get_interval - return the default timeslice of a process.
4195 * @pid: pid of the process.
4196 * @interval: userspace pointer to the timeslice value.
4198 * this syscall writes the default timeslice value of a given process
4199 * into the user-space timespec buffer. A value of '0' means infinity.
4202 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4204 int retval = -EINVAL;
4212 read_lock(&tasklist_lock);
4213 p = find_process_by_pid(pid);
4217 retval = security_task_getscheduler(p);
4221 jiffies_to_timespec(p->policy & SCHED_FIFO ?
4222 0 : task_timeslice(p), &t);
4223 read_unlock(&tasklist_lock);
4224 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4228 read_unlock(&tasklist_lock);
4232 static inline struct task_struct *eldest_child(struct task_struct *p)
4234 if (list_empty(&p->children)) return NULL;
4235 return list_entry(p->children.next,struct task_struct,sibling);
4238 static inline struct task_struct *older_sibling(struct task_struct *p)
4240 if (p->sibling.prev==&p->parent->children) return NULL;
4241 return list_entry(p->sibling.prev,struct task_struct,sibling);
4244 static inline struct task_struct *younger_sibling(struct task_struct *p)
4246 if (p->sibling.next==&p->parent->children) return NULL;
4247 return list_entry(p->sibling.next,struct task_struct,sibling);
4250 static void show_task(task_t *p)
4254 unsigned long free = 0;
4255 static const char *stat_nam[] = { "R", "S", "D", "T", "t", "Z", "X" };
4257 printk("%-13.13s ", p->comm);
4258 state = p->state ? __ffs(p->state) + 1 : 0;
4259 if (state < ARRAY_SIZE(stat_nam))
4260 printk(stat_nam[state]);
4263 #if (BITS_PER_LONG == 32)
4264 if (state == TASK_RUNNING)
4265 printk(" running ");
4267 printk(" %08lX ", thread_saved_pc(p));
4269 if (state == TASK_RUNNING)
4270 printk(" running task ");
4272 printk(" %016lx ", thread_saved_pc(p));
4274 #ifdef CONFIG_DEBUG_STACK_USAGE
4276 unsigned long *n = end_of_stack(p);
4279 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4282 printk("%5lu %5d %6d ", free, p->pid, p->parent->pid);
4283 if ((relative = eldest_child(p)))
4284 printk("%5d ", relative->pid);
4287 if ((relative = younger_sibling(p)))
4288 printk("%7d", relative->pid);
4291 if ((relative = older_sibling(p)))
4292 printk(" %5d", relative->pid);
4296 printk(" (L-TLB)\n");
4298 printk(" (NOTLB)\n");
4300 if (state != TASK_RUNNING)
4301 show_stack(p, NULL);
4304 void show_state(void)
4308 #if (BITS_PER_LONG == 32)
4311 printk(" task PC pid father child younger older\n");
4315 printk(" task PC pid father child younger older\n");
4317 read_lock(&tasklist_lock);
4318 do_each_thread(g, p) {
4320 * reset the NMI-timeout, listing all files on a slow
4321 * console might take alot of time:
4323 touch_nmi_watchdog();
4325 } while_each_thread(g, p);
4327 read_unlock(&tasklist_lock);
4328 mutex_debug_show_all_locks();
4332 * init_idle - set up an idle thread for a given CPU
4333 * @idle: task in question
4334 * @cpu: cpu the idle task belongs to
4336 * NOTE: this function does not set the idle thread's NEED_RESCHED
4337 * flag, to make booting more robust.
4339 void __devinit init_idle(task_t *idle, int cpu)
4341 runqueue_t *rq = cpu_rq(cpu);
4342 unsigned long flags;
4344 idle->sleep_avg = 0;
4346 idle->prio = MAX_PRIO;
4347 idle->state = TASK_RUNNING;
4348 idle->cpus_allowed = cpumask_of_cpu(cpu);
4349 set_task_cpu(idle, cpu);
4351 spin_lock_irqsave(&rq->lock, flags);
4352 rq->curr = rq->idle = idle;
4353 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4356 spin_unlock_irqrestore(&rq->lock, flags);
4358 /* Set the preempt count _outside_ the spinlocks! */
4359 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4360 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4362 task_thread_info(idle)->preempt_count = 0;
4367 * In a system that switches off the HZ timer nohz_cpu_mask
4368 * indicates which cpus entered this state. This is used
4369 * in the rcu update to wait only for active cpus. For system
4370 * which do not switch off the HZ timer nohz_cpu_mask should
4371 * always be CPU_MASK_NONE.
4373 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4377 * This is how migration works:
4379 * 1) we queue a migration_req_t structure in the source CPU's
4380 * runqueue and wake up that CPU's migration thread.
4381 * 2) we down() the locked semaphore => thread blocks.
4382 * 3) migration thread wakes up (implicitly it forces the migrated
4383 * thread off the CPU)
4384 * 4) it gets the migration request and checks whether the migrated
4385 * task is still in the wrong runqueue.
4386 * 5) if it's in the wrong runqueue then the migration thread removes
4387 * it and puts it into the right queue.
4388 * 6) migration thread up()s the semaphore.
4389 * 7) we wake up and the migration is done.
4393 * Change a given task's CPU affinity. Migrate the thread to a
4394 * proper CPU and schedule it away if the CPU it's executing on
4395 * is removed from the allowed bitmask.
4397 * NOTE: the caller must have a valid reference to the task, the
4398 * task must not exit() & deallocate itself prematurely. The
4399 * call is not atomic; no spinlocks may be held.
4401 int set_cpus_allowed(task_t *p, cpumask_t new_mask)
4403 unsigned long flags;
4405 migration_req_t req;
4408 rq = task_rq_lock(p, &flags);
4409 if (!cpus_intersects(new_mask, cpu_online_map)) {
4414 p->cpus_allowed = new_mask;
4415 /* Can the task run on the task's current CPU? If so, we're done */
4416 if (cpu_isset(task_cpu(p), new_mask))
4419 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4420 /* Need help from migration thread: drop lock and wait. */
4421 task_rq_unlock(rq, &flags);
4422 wake_up_process(rq->migration_thread);
4423 wait_for_completion(&req.done);
4424 tlb_migrate_finish(p->mm);
4428 task_rq_unlock(rq, &flags);
4432 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4435 * Move (not current) task off this cpu, onto dest cpu. We're doing
4436 * this because either it can't run here any more (set_cpus_allowed()
4437 * away from this CPU, or CPU going down), or because we're
4438 * attempting to rebalance this task on exec (sched_exec).
4440 * So we race with normal scheduler movements, but that's OK, as long
4441 * as the task is no longer on this CPU.
4443 static void __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4445 runqueue_t *rq_dest, *rq_src;
4447 if (unlikely(cpu_is_offline(dest_cpu)))
4450 rq_src = cpu_rq(src_cpu);
4451 rq_dest = cpu_rq(dest_cpu);
4453 double_rq_lock(rq_src, rq_dest);
4454 /* Already moved. */
4455 if (task_cpu(p) != src_cpu)
4457 /* Affinity changed (again). */
4458 if (!cpu_isset(dest_cpu, p->cpus_allowed))
4461 set_task_cpu(p, dest_cpu);
4464 * Sync timestamp with rq_dest's before activating.
4465 * The same thing could be achieved by doing this step
4466 * afterwards, and pretending it was a local activate.
4467 * This way is cleaner and logically correct.
4469 p->timestamp = p->timestamp - rq_src->timestamp_last_tick
4470 + rq_dest->timestamp_last_tick;
4471 deactivate_task(p, rq_src);
4472 activate_task(p, rq_dest, 0);
4473 if (TASK_PREEMPTS_CURR(p, rq_dest))
4474 resched_task(rq_dest->curr);
4478 double_rq_unlock(rq_src, rq_dest);
4482 * migration_thread - this is a highprio system thread that performs
4483 * thread migration by bumping thread off CPU then 'pushing' onto
4486 static int migration_thread(void *data)
4489 int cpu = (long)data;
4492 BUG_ON(rq->migration_thread != current);
4494 set_current_state(TASK_INTERRUPTIBLE);
4495 while (!kthread_should_stop()) {
4496 struct list_head *head;
4497 migration_req_t *req;
4501 spin_lock_irq(&rq->lock);
4503 if (cpu_is_offline(cpu)) {
4504 spin_unlock_irq(&rq->lock);
4508 if (rq->active_balance) {
4509 active_load_balance(rq, cpu);
4510 rq->active_balance = 0;
4513 head = &rq->migration_queue;
4515 if (list_empty(head)) {
4516 spin_unlock_irq(&rq->lock);
4518 set_current_state(TASK_INTERRUPTIBLE);
4521 req = list_entry(head->next, migration_req_t, list);
4522 list_del_init(head->next);
4524 spin_unlock(&rq->lock);
4525 __migrate_task(req->task, cpu, req->dest_cpu);
4528 complete(&req->done);
4530 __set_current_state(TASK_RUNNING);
4534 /* Wait for kthread_stop */
4535 set_current_state(TASK_INTERRUPTIBLE);
4536 while (!kthread_should_stop()) {
4538 set_current_state(TASK_INTERRUPTIBLE);
4540 __set_current_state(TASK_RUNNING);
4544 #ifdef CONFIG_HOTPLUG_CPU
4545 /* Figure out where task on dead CPU should go, use force if neccessary. */
4546 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *tsk)
4552 mask = node_to_cpumask(cpu_to_node(dead_cpu));
4553 cpus_and(mask, mask, tsk->cpus_allowed);
4554 dest_cpu = any_online_cpu(mask);
4556 /* On any allowed CPU? */
4557 if (dest_cpu == NR_CPUS)
4558 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4560 /* No more Mr. Nice Guy. */
4561 if (dest_cpu == NR_CPUS) {
4562 cpus_setall(tsk->cpus_allowed);
4563 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4566 * Don't tell them about moving exiting tasks or
4567 * kernel threads (both mm NULL), since they never
4570 if (tsk->mm && printk_ratelimit())
4571 printk(KERN_INFO "process %d (%s) no "
4572 "longer affine to cpu%d\n",
4573 tsk->pid, tsk->comm, dead_cpu);
4575 __migrate_task(tsk, dead_cpu, dest_cpu);
4579 * While a dead CPU has no uninterruptible tasks queued at this point,
4580 * it might still have a nonzero ->nr_uninterruptible counter, because
4581 * for performance reasons the counter is not stricly tracking tasks to
4582 * their home CPUs. So we just add the counter to another CPU's counter,
4583 * to keep the global sum constant after CPU-down:
4585 static void migrate_nr_uninterruptible(runqueue_t *rq_src)
4587 runqueue_t *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
4588 unsigned long flags;
4590 local_irq_save(flags);
4591 double_rq_lock(rq_src, rq_dest);
4592 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
4593 rq_src->nr_uninterruptible = 0;
4594 double_rq_unlock(rq_src, rq_dest);
4595 local_irq_restore(flags);
4598 /* Run through task list and migrate tasks from the dead cpu. */
4599 static void migrate_live_tasks(int src_cpu)
4601 struct task_struct *tsk, *t;
4603 write_lock_irq(&tasklist_lock);
4605 do_each_thread(t, tsk) {
4609 if (task_cpu(tsk) == src_cpu)
4610 move_task_off_dead_cpu(src_cpu, tsk);
4611 } while_each_thread(t, tsk);
4613 write_unlock_irq(&tasklist_lock);
4616 /* Schedules idle task to be the next runnable task on current CPU.
4617 * It does so by boosting its priority to highest possible and adding it to
4618 * the _front_ of runqueue. Used by CPU offline code.
4620 void sched_idle_next(void)
4622 int cpu = smp_processor_id();
4623 runqueue_t *rq = this_rq();
4624 struct task_struct *p = rq->idle;
4625 unsigned long flags;
4627 /* cpu has to be offline */
4628 BUG_ON(cpu_online(cpu));
4630 /* Strictly not necessary since rest of the CPUs are stopped by now
4631 * and interrupts disabled on current cpu.
4633 spin_lock_irqsave(&rq->lock, flags);
4635 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4636 /* Add idle task to _front_ of it's priority queue */
4637 __activate_idle_task(p, rq);
4639 spin_unlock_irqrestore(&rq->lock, flags);
4642 /* Ensures that the idle task is using init_mm right before its cpu goes
4645 void idle_task_exit(void)
4647 struct mm_struct *mm = current->active_mm;
4649 BUG_ON(cpu_online(smp_processor_id()));
4652 switch_mm(mm, &init_mm, current);
4656 static void migrate_dead(unsigned int dead_cpu, task_t *tsk)
4658 struct runqueue *rq = cpu_rq(dead_cpu);
4660 /* Must be exiting, otherwise would be on tasklist. */
4661 BUG_ON(tsk->exit_state != EXIT_ZOMBIE && tsk->exit_state != EXIT_DEAD);
4663 /* Cannot have done final schedule yet: would have vanished. */
4664 BUG_ON(tsk->flags & PF_DEAD);
4666 get_task_struct(tsk);
4669 * Drop lock around migration; if someone else moves it,
4670 * that's OK. No task can be added to this CPU, so iteration is
4673 spin_unlock_irq(&rq->lock);
4674 move_task_off_dead_cpu(dead_cpu, tsk);
4675 spin_lock_irq(&rq->lock);
4677 put_task_struct(tsk);
4680 /* release_task() removes task from tasklist, so we won't find dead tasks. */
4681 static void migrate_dead_tasks(unsigned int dead_cpu)
4684 struct runqueue *rq = cpu_rq(dead_cpu);
4686 for (arr = 0; arr < 2; arr++) {
4687 for (i = 0; i < MAX_PRIO; i++) {
4688 struct list_head *list = &rq->arrays[arr].queue[i];
4689 while (!list_empty(list))
4690 migrate_dead(dead_cpu,
4691 list_entry(list->next, task_t,
4696 #endif /* CONFIG_HOTPLUG_CPU */
4699 * migration_call - callback that gets triggered when a CPU is added.
4700 * Here we can start up the necessary migration thread for the new CPU.
4702 static int migration_call(struct notifier_block *nfb, unsigned long action,
4705 int cpu = (long)hcpu;
4706 struct task_struct *p;
4707 struct runqueue *rq;
4708 unsigned long flags;
4711 case CPU_UP_PREPARE:
4712 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
4715 p->flags |= PF_NOFREEZE;
4716 kthread_bind(p, cpu);
4717 /* Must be high prio: stop_machine expects to yield to it. */
4718 rq = task_rq_lock(p, &flags);
4719 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4720 task_rq_unlock(rq, &flags);
4721 cpu_rq(cpu)->migration_thread = p;
4724 /* Strictly unneccessary, as first user will wake it. */
4725 wake_up_process(cpu_rq(cpu)->migration_thread);
4727 #ifdef CONFIG_HOTPLUG_CPU
4728 case CPU_UP_CANCELED:
4729 /* Unbind it from offline cpu so it can run. Fall thru. */
4730 kthread_bind(cpu_rq(cpu)->migration_thread,
4731 any_online_cpu(cpu_online_map));
4732 kthread_stop(cpu_rq(cpu)->migration_thread);
4733 cpu_rq(cpu)->migration_thread = NULL;
4736 migrate_live_tasks(cpu);
4738 kthread_stop(rq->migration_thread);
4739 rq->migration_thread = NULL;
4740 /* Idle task back to normal (off runqueue, low prio) */
4741 rq = task_rq_lock(rq->idle, &flags);
4742 deactivate_task(rq->idle, rq);
4743 rq->idle->static_prio = MAX_PRIO;
4744 __setscheduler(rq->idle, SCHED_NORMAL, 0);
4745 migrate_dead_tasks(cpu);
4746 task_rq_unlock(rq, &flags);
4747 migrate_nr_uninterruptible(rq);
4748 BUG_ON(rq->nr_running != 0);
4750 /* No need to migrate the tasks: it was best-effort if
4751 * they didn't do lock_cpu_hotplug(). Just wake up
4752 * the requestors. */
4753 spin_lock_irq(&rq->lock);
4754 while (!list_empty(&rq->migration_queue)) {
4755 migration_req_t *req;
4756 req = list_entry(rq->migration_queue.next,
4757 migration_req_t, list);
4758 list_del_init(&req->list);
4759 complete(&req->done);
4761 spin_unlock_irq(&rq->lock);
4768 /* Register at highest priority so that task migration (migrate_all_tasks)
4769 * happens before everything else.
4771 static struct notifier_block __devinitdata migration_notifier = {
4772 .notifier_call = migration_call,
4776 int __init migration_init(void)
4778 void *cpu = (void *)(long)smp_processor_id();
4779 /* Start one for boot CPU. */
4780 migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
4781 migration_call(&migration_notifier, CPU_ONLINE, cpu);
4782 register_cpu_notifier(&migration_notifier);
4788 #undef SCHED_DOMAIN_DEBUG
4789 #ifdef SCHED_DOMAIN_DEBUG
4790 static void sched_domain_debug(struct sched_domain *sd, int cpu)
4795 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
4799 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
4804 struct sched_group *group = sd->groups;
4805 cpumask_t groupmask;
4807 cpumask_scnprintf(str, NR_CPUS, sd->span);
4808 cpus_clear(groupmask);
4811 for (i = 0; i < level + 1; i++)
4813 printk("domain %d: ", level);
4815 if (!(sd->flags & SD_LOAD_BALANCE)) {
4816 printk("does not load-balance\n");
4818 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent");
4822 printk("span %s\n", str);
4824 if (!cpu_isset(cpu, sd->span))
4825 printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
4826 if (!cpu_isset(cpu, group->cpumask))
4827 printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
4830 for (i = 0; i < level + 2; i++)
4836 printk(KERN_ERR "ERROR: group is NULL\n");
4840 if (!group->cpu_power) {
4842 printk(KERN_ERR "ERROR: domain->cpu_power not set\n");
4845 if (!cpus_weight(group->cpumask)) {
4847 printk(KERN_ERR "ERROR: empty group\n");
4850 if (cpus_intersects(groupmask, group->cpumask)) {
4852 printk(KERN_ERR "ERROR: repeated CPUs\n");
4855 cpus_or(groupmask, groupmask, group->cpumask);
4857 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
4860 group = group->next;
4861 } while (group != sd->groups);
4864 if (!cpus_equal(sd->span, groupmask))
4865 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
4871 if (!cpus_subset(groupmask, sd->span))
4872 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
4878 #define sched_domain_debug(sd, cpu) {}
4881 static int sd_degenerate(struct sched_domain *sd)
4883 if (cpus_weight(sd->span) == 1)
4886 /* Following flags need at least 2 groups */
4887 if (sd->flags & (SD_LOAD_BALANCE |
4888 SD_BALANCE_NEWIDLE |
4891 if (sd->groups != sd->groups->next)
4895 /* Following flags don't use groups */
4896 if (sd->flags & (SD_WAKE_IDLE |
4904 static int sd_parent_degenerate(struct sched_domain *sd,
4905 struct sched_domain *parent)
4907 unsigned long cflags = sd->flags, pflags = parent->flags;
4909 if (sd_degenerate(parent))
4912 if (!cpus_equal(sd->span, parent->span))
4915 /* Does parent contain flags not in child? */
4916 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
4917 if (cflags & SD_WAKE_AFFINE)
4918 pflags &= ~SD_WAKE_BALANCE;
4919 /* Flags needing groups don't count if only 1 group in parent */
4920 if (parent->groups == parent->groups->next) {
4921 pflags &= ~(SD_LOAD_BALANCE |
4922 SD_BALANCE_NEWIDLE |
4926 if (~cflags & pflags)
4933 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
4934 * hold the hotplug lock.
4936 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
4938 runqueue_t *rq = cpu_rq(cpu);
4939 struct sched_domain *tmp;
4941 /* Remove the sched domains which do not contribute to scheduling. */
4942 for (tmp = sd; tmp; tmp = tmp->parent) {
4943 struct sched_domain *parent = tmp->parent;
4946 if (sd_parent_degenerate(tmp, parent))
4947 tmp->parent = parent->parent;
4950 if (sd && sd_degenerate(sd))
4953 sched_domain_debug(sd, cpu);
4955 rcu_assign_pointer(rq->sd, sd);
4958 /* cpus with isolated domains */
4959 static cpumask_t __devinitdata cpu_isolated_map = CPU_MASK_NONE;
4961 /* Setup the mask of cpus configured for isolated domains */
4962 static int __init isolated_cpu_setup(char *str)
4964 int ints[NR_CPUS], i;
4966 str = get_options(str, ARRAY_SIZE(ints), ints);
4967 cpus_clear(cpu_isolated_map);
4968 for (i = 1; i <= ints[0]; i++)
4969 if (ints[i] < NR_CPUS)
4970 cpu_set(ints[i], cpu_isolated_map);
4974 __setup ("isolcpus=", isolated_cpu_setup);
4977 * init_sched_build_groups takes an array of groups, the cpumask we wish
4978 * to span, and a pointer to a function which identifies what group a CPU
4979 * belongs to. The return value of group_fn must be a valid index into the
4980 * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
4981 * keep track of groups covered with a cpumask_t).
4983 * init_sched_build_groups will build a circular linked list of the groups
4984 * covered by the given span, and will set each group's ->cpumask correctly,
4985 * and ->cpu_power to 0.
4987 static void init_sched_build_groups(struct sched_group groups[], cpumask_t span,
4988 int (*group_fn)(int cpu))
4990 struct sched_group *first = NULL, *last = NULL;
4991 cpumask_t covered = CPU_MASK_NONE;
4994 for_each_cpu_mask(i, span) {
4995 int group = group_fn(i);
4996 struct sched_group *sg = &groups[group];
4999 if (cpu_isset(i, covered))
5002 sg->cpumask = CPU_MASK_NONE;
5005 for_each_cpu_mask(j, span) {
5006 if (group_fn(j) != group)
5009 cpu_set(j, covered);
5010 cpu_set(j, sg->cpumask);
5021 #define SD_NODES_PER_DOMAIN 16
5024 * Self-tuning task migration cost measurement between source and target CPUs.
5026 * This is done by measuring the cost of manipulating buffers of varying
5027 * sizes. For a given buffer-size here are the steps that are taken:
5029 * 1) the source CPU reads+dirties a shared buffer
5030 * 2) the target CPU reads+dirties the same shared buffer
5032 * We measure how long they take, in the following 4 scenarios:
5034 * - source: CPU1, target: CPU2 | cost1
5035 * - source: CPU2, target: CPU1 | cost2
5036 * - source: CPU1, target: CPU1 | cost3
5037 * - source: CPU2, target: CPU2 | cost4
5039 * We then calculate the cost3+cost4-cost1-cost2 difference - this is
5040 * the cost of migration.
5042 * We then start off from a small buffer-size and iterate up to larger
5043 * buffer sizes, in 5% steps - measuring each buffer-size separately, and
5044 * doing a maximum search for the cost. (The maximum cost for a migration
5045 * normally occurs when the working set size is around the effective cache
5048 #define SEARCH_SCOPE 2
5049 #define MIN_CACHE_SIZE (64*1024U)
5050 #define DEFAULT_CACHE_SIZE (5*1024*1024U)
5051 #define ITERATIONS 1
5052 #define SIZE_THRESH 130
5053 #define COST_THRESH 130
5056 * The migration cost is a function of 'domain distance'. Domain
5057 * distance is the number of steps a CPU has to iterate down its
5058 * domain tree to share a domain with the other CPU. The farther
5059 * two CPUs are from each other, the larger the distance gets.
5061 * Note that we use the distance only to cache measurement results,
5062 * the distance value is not used numerically otherwise. When two
5063 * CPUs have the same distance it is assumed that the migration
5064 * cost is the same. (this is a simplification but quite practical)
5066 #define MAX_DOMAIN_DISTANCE 32
5068 static unsigned long long migration_cost[MAX_DOMAIN_DISTANCE] =
5069 { [ 0 ... MAX_DOMAIN_DISTANCE-1 ] = -1LL };
5072 * Allow override of migration cost - in units of microseconds.
5073 * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost
5074 * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs:
5076 static int __init migration_cost_setup(char *str)
5078 int ints[MAX_DOMAIN_DISTANCE+1], i;
5080 str = get_options(str, ARRAY_SIZE(ints), ints);
5082 printk("#ints: %d\n", ints[0]);
5083 for (i = 1; i <= ints[0]; i++) {
5084 migration_cost[i-1] = (unsigned long long)ints[i]*1000;
5085 printk("migration_cost[%d]: %Ld\n", i-1, migration_cost[i-1]);
5090 __setup ("migration_cost=", migration_cost_setup);
5093 * Global multiplier (divisor) for migration-cutoff values,
5094 * in percentiles. E.g. use a value of 150 to get 1.5 times
5095 * longer cache-hot cutoff times.
5097 * (We scale it from 100 to 128 to long long handling easier.)
5100 #define MIGRATION_FACTOR_SCALE 128
5102 static unsigned int migration_factor = MIGRATION_FACTOR_SCALE;
5104 static int __init setup_migration_factor(char *str)
5106 get_option(&str, &migration_factor);
5107 migration_factor = migration_factor * MIGRATION_FACTOR_SCALE / 100;
5111 __setup("migration_factor=", setup_migration_factor);
5114 * Estimated distance of two CPUs, measured via the number of domains
5115 * we have to pass for the two CPUs to be in the same span:
5117 static unsigned long domain_distance(int cpu1, int cpu2)
5119 unsigned long distance = 0;
5120 struct sched_domain *sd;
5122 for_each_domain(cpu1, sd) {
5123 WARN_ON(!cpu_isset(cpu1, sd->span));
5124 if (cpu_isset(cpu2, sd->span))
5128 if (distance >= MAX_DOMAIN_DISTANCE) {
5130 distance = MAX_DOMAIN_DISTANCE-1;
5136 static unsigned int migration_debug;
5138 static int __init setup_migration_debug(char *str)
5140 get_option(&str, &migration_debug);
5144 __setup("migration_debug=", setup_migration_debug);
5147 * Maximum cache-size that the scheduler should try to measure.
5148 * Architectures with larger caches should tune this up during
5149 * bootup. Gets used in the domain-setup code (i.e. during SMP
5152 unsigned int max_cache_size;
5154 static int __init setup_max_cache_size(char *str)
5156 get_option(&str, &max_cache_size);
5160 __setup("max_cache_size=", setup_max_cache_size);
5163 * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This
5164 * is the operation that is timed, so we try to generate unpredictable
5165 * cachemisses that still end up filling the L2 cache:
5167 static void touch_cache(void *__cache, unsigned long __size)
5169 unsigned long size = __size/sizeof(long), chunk1 = size/3,
5171 unsigned long *cache = __cache;
5174 for (i = 0; i < size/6; i += 8) {
5177 case 1: cache[size-1-i]++;
5178 case 2: cache[chunk1-i]++;
5179 case 3: cache[chunk1+i]++;
5180 case 4: cache[chunk2-i]++;
5181 case 5: cache[chunk2+i]++;
5187 * Measure the cache-cost of one task migration. Returns in units of nsec.
5189 static unsigned long long measure_one(void *cache, unsigned long size,
5190 int source, int target)
5192 cpumask_t mask, saved_mask;
5193 unsigned long long t0, t1, t2, t3, cost;
5195 saved_mask = current->cpus_allowed;
5198 * Flush source caches to RAM and invalidate them:
5203 * Migrate to the source CPU:
5205 mask = cpumask_of_cpu(source);
5206 set_cpus_allowed(current, mask);
5207 WARN_ON(smp_processor_id() != source);
5210 * Dirty the working set:
5213 touch_cache(cache, size);
5217 * Migrate to the target CPU, dirty the L2 cache and access
5218 * the shared buffer. (which represents the working set
5219 * of a migrated task.)
5221 mask = cpumask_of_cpu(target);
5222 set_cpus_allowed(current, mask);
5223 WARN_ON(smp_processor_id() != target);
5226 touch_cache(cache, size);
5229 cost = t1-t0 + t3-t2;
5231 if (migration_debug >= 2)
5232 printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n",
5233 source, target, t1-t0, t1-t0, t3-t2, cost);
5235 * Flush target caches to RAM and invalidate them:
5239 set_cpus_allowed(current, saved_mask);
5245 * Measure a series of task migrations and return the average
5246 * result. Since this code runs early during bootup the system
5247 * is 'undisturbed' and the average latency makes sense.
5249 * The algorithm in essence auto-detects the relevant cache-size,
5250 * so it will properly detect different cachesizes for different
5251 * cache-hierarchies, depending on how the CPUs are connected.
5253 * Architectures can prime the upper limit of the search range via
5254 * max_cache_size, otherwise the search range defaults to 20MB...64K.
5256 static unsigned long long
5257 measure_cost(int cpu1, int cpu2, void *cache, unsigned int size)
5259 unsigned long long cost1, cost2;
5263 * Measure the migration cost of 'size' bytes, over an
5264 * average of 10 runs:
5266 * (We perturb the cache size by a small (0..4k)
5267 * value to compensate size/alignment related artifacts.
5268 * We also subtract the cost of the operation done on
5274 * dry run, to make sure we start off cache-cold on cpu1,
5275 * and to get any vmalloc pagefaults in advance:
5277 measure_one(cache, size, cpu1, cpu2);
5278 for (i = 0; i < ITERATIONS; i++)
5279 cost1 += measure_one(cache, size - i*1024, cpu1, cpu2);
5281 measure_one(cache, size, cpu2, cpu1);
5282 for (i = 0; i < ITERATIONS; i++)
5283 cost1 += measure_one(cache, size - i*1024, cpu2, cpu1);
5286 * (We measure the non-migrating [cached] cost on both
5287 * cpu1 and cpu2, to handle CPUs with different speeds)
5291 measure_one(cache, size, cpu1, cpu1);
5292 for (i = 0; i < ITERATIONS; i++)
5293 cost2 += measure_one(cache, size - i*1024, cpu1, cpu1);
5295 measure_one(cache, size, cpu2, cpu2);
5296 for (i = 0; i < ITERATIONS; i++)
5297 cost2 += measure_one(cache, size - i*1024, cpu2, cpu2);
5300 * Get the per-iteration migration cost:
5302 do_div(cost1, 2*ITERATIONS);
5303 do_div(cost2, 2*ITERATIONS);
5305 return cost1 - cost2;
5308 static unsigned long long measure_migration_cost(int cpu1, int cpu2)
5310 unsigned long long max_cost = 0, fluct = 0, avg_fluct = 0;
5311 unsigned int max_size, size, size_found = 0;
5312 long long cost = 0, prev_cost;
5316 * Search from max_cache_size*5 down to 64K - the real relevant
5317 * cachesize has to lie somewhere inbetween.
5319 if (max_cache_size) {
5320 max_size = max(max_cache_size * SEARCH_SCOPE, MIN_CACHE_SIZE);
5321 size = max(max_cache_size / SEARCH_SCOPE, MIN_CACHE_SIZE);
5324 * Since we have no estimation about the relevant
5327 max_size = DEFAULT_CACHE_SIZE * SEARCH_SCOPE;
5328 size = MIN_CACHE_SIZE;
5331 if (!cpu_online(cpu1) || !cpu_online(cpu2)) {
5332 printk("cpu %d and %d not both online!\n", cpu1, cpu2);
5337 * Allocate the working set:
5339 cache = vmalloc(max_size);
5341 printk("could not vmalloc %d bytes for cache!\n", 2*max_size);
5342 return 1000000; // return 1 msec on very small boxen
5345 while (size <= max_size) {
5347 cost = measure_cost(cpu1, cpu2, cache, size);
5353 if (max_cost < cost) {
5359 * Calculate average fluctuation, we use this to prevent
5360 * noise from triggering an early break out of the loop:
5362 fluct = abs(cost - prev_cost);
5363 avg_fluct = (avg_fluct + fluct)/2;
5365 if (migration_debug)
5366 printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): (%8Ld %8Ld)\n",
5368 (long)cost / 1000000,
5369 ((long)cost / 100000) % 10,
5370 (long)max_cost / 1000000,
5371 ((long)max_cost / 100000) % 10,
5372 domain_distance(cpu1, cpu2),
5376 * If we iterated at least 20% past the previous maximum,
5377 * and the cost has dropped by more than 20% already,
5378 * (taking fluctuations into account) then we assume to
5379 * have found the maximum and break out of the loop early:
5381 if (size_found && (size*100 > size_found*SIZE_THRESH))
5382 if (cost+avg_fluct <= 0 ||
5383 max_cost*100 > (cost+avg_fluct)*COST_THRESH) {
5385 if (migration_debug)
5386 printk("-> found max.\n");
5390 * Increase the cachesize in 10% steps:
5392 size = size * 10 / 9;
5395 if (migration_debug)
5396 printk("[%d][%d] working set size found: %d, cost: %Ld\n",
5397 cpu1, cpu2, size_found, max_cost);
5402 * A task is considered 'cache cold' if at least 2 times
5403 * the worst-case cost of migration has passed.
5405 * (this limit is only listened to if the load-balancing
5406 * situation is 'nice' - if there is a large imbalance we
5407 * ignore it for the sake of CPU utilization and
5408 * processing fairness.)
5410 return 2 * max_cost * migration_factor / MIGRATION_FACTOR_SCALE;
5413 static void calibrate_migration_costs(const cpumask_t *cpu_map)
5415 int cpu1 = -1, cpu2 = -1, cpu, orig_cpu = raw_smp_processor_id();
5416 unsigned long j0, j1, distance, max_distance = 0;
5417 struct sched_domain *sd;
5422 * First pass - calculate the cacheflush times:
5424 for_each_cpu_mask(cpu1, *cpu_map) {
5425 for_each_cpu_mask(cpu2, *cpu_map) {
5428 distance = domain_distance(cpu1, cpu2);
5429 max_distance = max(max_distance, distance);
5431 * No result cached yet?
5433 if (migration_cost[distance] == -1LL)
5434 migration_cost[distance] =
5435 measure_migration_cost(cpu1, cpu2);
5439 * Second pass - update the sched domain hierarchy with
5440 * the new cache-hot-time estimations:
5442 for_each_cpu_mask(cpu, *cpu_map) {
5444 for_each_domain(cpu, sd) {
5445 sd->cache_hot_time = migration_cost[distance];
5452 if (migration_debug)
5453 printk("migration: max_cache_size: %d, cpu: %d MHz:\n",
5461 if (system_state == SYSTEM_BOOTING) {
5462 printk("migration_cost=");
5463 for (distance = 0; distance <= max_distance; distance++) {
5466 printk("%ld", (long)migration_cost[distance] / 1000);
5471 if (migration_debug)
5472 printk("migration: %ld seconds\n", (j1-j0)/HZ);
5475 * Move back to the original CPU. NUMA-Q gets confused
5476 * if we migrate to another quad during bootup.
5478 if (raw_smp_processor_id() != orig_cpu) {
5479 cpumask_t mask = cpumask_of_cpu(orig_cpu),
5480 saved_mask = current->cpus_allowed;
5482 set_cpus_allowed(current, mask);
5483 set_cpus_allowed(current, saved_mask);
5490 * find_next_best_node - find the next node to include in a sched_domain
5491 * @node: node whose sched_domain we're building
5492 * @used_nodes: nodes already in the sched_domain
5494 * Find the next node to include in a given scheduling domain. Simply
5495 * finds the closest node not already in the @used_nodes map.
5497 * Should use nodemask_t.
5499 static int find_next_best_node(int node, unsigned long *used_nodes)
5501 int i, n, val, min_val, best_node = 0;
5505 for (i = 0; i < MAX_NUMNODES; i++) {
5506 /* Start at @node */
5507 n = (node + i) % MAX_NUMNODES;
5509 if (!nr_cpus_node(n))
5512 /* Skip already used nodes */
5513 if (test_bit(n, used_nodes))
5516 /* Simple min distance search */
5517 val = node_distance(node, n);
5519 if (val < min_val) {
5525 set_bit(best_node, used_nodes);
5530 * sched_domain_node_span - get a cpumask for a node's sched_domain
5531 * @node: node whose cpumask we're constructing
5532 * @size: number of nodes to include in this span
5534 * Given a node, construct a good cpumask for its sched_domain to span. It
5535 * should be one that prevents unnecessary balancing, but also spreads tasks
5538 static cpumask_t sched_domain_node_span(int node)
5541 cpumask_t span, nodemask;
5542 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5545 bitmap_zero(used_nodes, MAX_NUMNODES);
5547 nodemask = node_to_cpumask(node);
5548 cpus_or(span, span, nodemask);
5549 set_bit(node, used_nodes);
5551 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5552 int next_node = find_next_best_node(node, used_nodes);
5553 nodemask = node_to_cpumask(next_node);
5554 cpus_or(span, span, nodemask);
5562 * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we
5563 * can switch it on easily if needed.
5565 #ifdef CONFIG_SCHED_SMT
5566 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5567 static struct sched_group sched_group_cpus[NR_CPUS];
5568 static int cpu_to_cpu_group(int cpu)
5574 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5575 static struct sched_group sched_group_phys[NR_CPUS];
5576 static int cpu_to_phys_group(int cpu)
5578 #ifdef CONFIG_SCHED_SMT
5579 return first_cpu(cpu_sibling_map[cpu]);
5587 * The init_sched_build_groups can't handle what we want to do with node
5588 * groups, so roll our own. Now each node has its own list of groups which
5589 * gets dynamically allocated.
5591 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5592 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5594 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5595 static struct sched_group *sched_group_allnodes_bycpu[NR_CPUS];
5597 static int cpu_to_allnodes_group(int cpu)
5599 return cpu_to_node(cpu);
5604 * Build sched domains for a given set of cpus and attach the sched domains
5605 * to the individual cpus
5607 void build_sched_domains(const cpumask_t *cpu_map)
5611 struct sched_group **sched_group_nodes = NULL;
5612 struct sched_group *sched_group_allnodes = NULL;
5615 * Allocate the per-node list of sched groups
5617 sched_group_nodes = kmalloc(sizeof(struct sched_group*)*MAX_NUMNODES,
5619 if (!sched_group_nodes) {
5620 printk(KERN_WARNING "Can not alloc sched group node list\n");
5623 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
5627 * Set up domains for cpus specified by the cpu_map.
5629 for_each_cpu_mask(i, *cpu_map) {
5631 struct sched_domain *sd = NULL, *p;
5632 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
5634 cpus_and(nodemask, nodemask, *cpu_map);
5637 if (cpus_weight(*cpu_map)
5638 > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
5639 if (!sched_group_allnodes) {
5640 sched_group_allnodes
5641 = kmalloc(sizeof(struct sched_group)
5644 if (!sched_group_allnodes) {
5646 "Can not alloc allnodes sched group\n");
5649 sched_group_allnodes_bycpu[i]
5650 = sched_group_allnodes;
5652 sd = &per_cpu(allnodes_domains, i);
5653 *sd = SD_ALLNODES_INIT;
5654 sd->span = *cpu_map;
5655 group = cpu_to_allnodes_group(i);
5656 sd->groups = &sched_group_allnodes[group];
5661 sd = &per_cpu(node_domains, i);
5663 sd->span = sched_domain_node_span(cpu_to_node(i));
5665 cpus_and(sd->span, sd->span, *cpu_map);
5669 sd = &per_cpu(phys_domains, i);
5670 group = cpu_to_phys_group(i);
5672 sd->span = nodemask;
5674 sd->groups = &sched_group_phys[group];
5676 #ifdef CONFIG_SCHED_SMT
5678 sd = &per_cpu(cpu_domains, i);
5679 group = cpu_to_cpu_group(i);
5680 *sd = SD_SIBLING_INIT;
5681 sd->span = cpu_sibling_map[i];
5682 cpus_and(sd->span, sd->span, *cpu_map);
5684 sd->groups = &sched_group_cpus[group];
5688 #ifdef CONFIG_SCHED_SMT
5689 /* Set up CPU (sibling) groups */
5690 for_each_cpu_mask(i, *cpu_map) {
5691 cpumask_t this_sibling_map = cpu_sibling_map[i];
5692 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
5693 if (i != first_cpu(this_sibling_map))
5696 init_sched_build_groups(sched_group_cpus, this_sibling_map,
5701 /* Set up physical groups */
5702 for (i = 0; i < MAX_NUMNODES; i++) {
5703 cpumask_t nodemask = node_to_cpumask(i);
5705 cpus_and(nodemask, nodemask, *cpu_map);
5706 if (cpus_empty(nodemask))
5709 init_sched_build_groups(sched_group_phys, nodemask,
5710 &cpu_to_phys_group);
5714 /* Set up node groups */
5715 if (sched_group_allnodes)
5716 init_sched_build_groups(sched_group_allnodes, *cpu_map,
5717 &cpu_to_allnodes_group);
5719 for (i = 0; i < MAX_NUMNODES; i++) {
5720 /* Set up node groups */
5721 struct sched_group *sg, *prev;
5722 cpumask_t nodemask = node_to_cpumask(i);
5723 cpumask_t domainspan;
5724 cpumask_t covered = CPU_MASK_NONE;
5727 cpus_and(nodemask, nodemask, *cpu_map);
5728 if (cpus_empty(nodemask)) {
5729 sched_group_nodes[i] = NULL;
5733 domainspan = sched_domain_node_span(i);
5734 cpus_and(domainspan, domainspan, *cpu_map);
5736 sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL);
5737 sched_group_nodes[i] = sg;
5738 for_each_cpu_mask(j, nodemask) {
5739 struct sched_domain *sd;
5740 sd = &per_cpu(node_domains, j);
5742 if (sd->groups == NULL) {
5743 /* Turn off balancing if we have no groups */
5749 "Can not alloc domain group for node %d\n", i);
5753 sg->cpumask = nodemask;
5754 cpus_or(covered, covered, nodemask);
5757 for (j = 0; j < MAX_NUMNODES; j++) {
5758 cpumask_t tmp, notcovered;
5759 int n = (i + j) % MAX_NUMNODES;
5761 cpus_complement(notcovered, covered);
5762 cpus_and(tmp, notcovered, *cpu_map);
5763 cpus_and(tmp, tmp, domainspan);
5764 if (cpus_empty(tmp))
5767 nodemask = node_to_cpumask(n);
5768 cpus_and(tmp, tmp, nodemask);
5769 if (cpus_empty(tmp))
5772 sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL);
5775 "Can not alloc domain group for node %d\n", j);
5780 cpus_or(covered, covered, tmp);
5784 prev->next = sched_group_nodes[i];
5788 /* Calculate CPU power for physical packages and nodes */
5789 for_each_cpu_mask(i, *cpu_map) {
5791 struct sched_domain *sd;
5792 #ifdef CONFIG_SCHED_SMT
5793 sd = &per_cpu(cpu_domains, i);
5794 power = SCHED_LOAD_SCALE;
5795 sd->groups->cpu_power = power;
5798 sd = &per_cpu(phys_domains, i);
5799 power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
5800 (cpus_weight(sd->groups->cpumask)-1) / 10;
5801 sd->groups->cpu_power = power;
5804 sd = &per_cpu(allnodes_domains, i);
5806 power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
5807 (cpus_weight(sd->groups->cpumask)-1) / 10;
5808 sd->groups->cpu_power = power;
5814 for (i = 0; i < MAX_NUMNODES; i++) {
5815 struct sched_group *sg = sched_group_nodes[i];
5821 for_each_cpu_mask(j, sg->cpumask) {
5822 struct sched_domain *sd;
5825 sd = &per_cpu(phys_domains, j);
5826 if (j != first_cpu(sd->groups->cpumask)) {
5828 * Only add "power" once for each
5833 power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
5834 (cpus_weight(sd->groups->cpumask)-1) / 10;
5836 sg->cpu_power += power;
5839 if (sg != sched_group_nodes[i])
5844 /* Attach the domains */
5845 for_each_cpu_mask(i, *cpu_map) {
5846 struct sched_domain *sd;
5847 #ifdef CONFIG_SCHED_SMT
5848 sd = &per_cpu(cpu_domains, i);
5850 sd = &per_cpu(phys_domains, i);
5852 cpu_attach_domain(sd, i);
5855 * Tune cache-hot values:
5857 calibrate_migration_costs(cpu_map);
5860 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
5862 static void arch_init_sched_domains(const cpumask_t *cpu_map)
5864 cpumask_t cpu_default_map;
5867 * Setup mask for cpus without special case scheduling requirements.
5868 * For now this just excludes isolated cpus, but could be used to
5869 * exclude other special cases in the future.
5871 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
5873 build_sched_domains(&cpu_default_map);
5876 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
5882 for_each_cpu_mask(cpu, *cpu_map) {
5883 struct sched_group *sched_group_allnodes
5884 = sched_group_allnodes_bycpu[cpu];
5885 struct sched_group **sched_group_nodes
5886 = sched_group_nodes_bycpu[cpu];
5888 if (sched_group_allnodes) {
5889 kfree(sched_group_allnodes);
5890 sched_group_allnodes_bycpu[cpu] = NULL;
5893 if (!sched_group_nodes)
5896 for (i = 0; i < MAX_NUMNODES; i++) {
5897 cpumask_t nodemask = node_to_cpumask(i);
5898 struct sched_group *oldsg, *sg = sched_group_nodes[i];
5900 cpus_and(nodemask, nodemask, *cpu_map);
5901 if (cpus_empty(nodemask))
5911 if (oldsg != sched_group_nodes[i])
5914 kfree(sched_group_nodes);
5915 sched_group_nodes_bycpu[cpu] = NULL;
5921 * Detach sched domains from a group of cpus specified in cpu_map
5922 * These cpus will now be attached to the NULL domain
5924 static void detach_destroy_domains(const cpumask_t *cpu_map)
5928 for_each_cpu_mask(i, *cpu_map)
5929 cpu_attach_domain(NULL, i);
5930 synchronize_sched();
5931 arch_destroy_sched_domains(cpu_map);
5935 * Partition sched domains as specified by the cpumasks below.
5936 * This attaches all cpus from the cpumasks to the NULL domain,
5937 * waits for a RCU quiescent period, recalculates sched
5938 * domain information and then attaches them back to the
5939 * correct sched domains
5940 * Call with hotplug lock held
5942 void partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
5944 cpumask_t change_map;
5946 cpus_and(*partition1, *partition1, cpu_online_map);
5947 cpus_and(*partition2, *partition2, cpu_online_map);
5948 cpus_or(change_map, *partition1, *partition2);
5950 /* Detach sched domains from all of the affected cpus */
5951 detach_destroy_domains(&change_map);
5952 if (!cpus_empty(*partition1))
5953 build_sched_domains(partition1);
5954 if (!cpus_empty(*partition2))
5955 build_sched_domains(partition2);
5958 #ifdef CONFIG_HOTPLUG_CPU
5960 * Force a reinitialization of the sched domains hierarchy. The domains
5961 * and groups cannot be updated in place without racing with the balancing
5962 * code, so we temporarily attach all running cpus to the NULL domain
5963 * which will prevent rebalancing while the sched domains are recalculated.
5965 static int update_sched_domains(struct notifier_block *nfb,
5966 unsigned long action, void *hcpu)
5969 case CPU_UP_PREPARE:
5970 case CPU_DOWN_PREPARE:
5971 detach_destroy_domains(&cpu_online_map);
5974 case CPU_UP_CANCELED:
5975 case CPU_DOWN_FAILED:
5979 * Fall through and re-initialise the domains.
5986 /* The hotplug lock is already held by cpu_up/cpu_down */
5987 arch_init_sched_domains(&cpu_online_map);
5993 void __init sched_init_smp(void)
5996 arch_init_sched_domains(&cpu_online_map);
5997 unlock_cpu_hotplug();
5998 /* XXX: Theoretical race here - CPU may be hotplugged now */
5999 hotcpu_notifier(update_sched_domains, 0);
6002 void __init sched_init_smp(void)
6005 #endif /* CONFIG_SMP */
6007 int in_sched_functions(unsigned long addr)
6009 /* Linker adds these: start and end of __sched functions */
6010 extern char __sched_text_start[], __sched_text_end[];
6011 return in_lock_functions(addr) ||
6012 (addr >= (unsigned long)__sched_text_start
6013 && addr < (unsigned long)__sched_text_end);
6016 void __init sched_init(void)
6022 prio_array_t *array;
6025 spin_lock_init(&rq->lock);
6027 rq->active = rq->arrays;
6028 rq->expired = rq->arrays + 1;
6029 rq->best_expired_prio = MAX_PRIO;
6033 for (j = 1; j < 3; j++)
6034 rq->cpu_load[j] = 0;
6035 rq->active_balance = 0;
6037 rq->migration_thread = NULL;
6038 INIT_LIST_HEAD(&rq->migration_queue);
6040 atomic_set(&rq->nr_iowait, 0);
6042 for (j = 0; j < 2; j++) {
6043 array = rq->arrays + j;
6044 for (k = 0; k < MAX_PRIO; k++) {
6045 INIT_LIST_HEAD(array->queue + k);
6046 __clear_bit(k, array->bitmap);
6048 // delimiter for bitsearch
6049 __set_bit(MAX_PRIO, array->bitmap);
6054 * The boot idle thread does lazy MMU switching as well:
6056 atomic_inc(&init_mm.mm_count);
6057 enter_lazy_tlb(&init_mm, current);
6060 * Make us the idle thread. Technically, schedule() should not be
6061 * called from this thread, however somewhere below it might be,
6062 * but because we are the idle thread, we just pick up running again
6063 * when this runqueue becomes "idle".
6065 init_idle(current, smp_processor_id());
6068 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6069 void __might_sleep(char *file, int line)
6071 #if defined(in_atomic)
6072 static unsigned long prev_jiffy; /* ratelimiting */
6074 if ((in_atomic() || irqs_disabled()) &&
6075 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6076 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6078 prev_jiffy = jiffies;
6079 printk(KERN_ERR "Debug: sleeping function called from invalid"
6080 " context at %s:%d\n", file, line);
6081 printk("in_atomic():%d, irqs_disabled():%d\n",
6082 in_atomic(), irqs_disabled());
6087 EXPORT_SYMBOL(__might_sleep);
6090 #ifdef CONFIG_MAGIC_SYSRQ
6091 void normalize_rt_tasks(void)
6093 struct task_struct *p;
6094 prio_array_t *array;
6095 unsigned long flags;
6098 read_lock_irq(&tasklist_lock);
6099 for_each_process (p) {
6103 rq = task_rq_lock(p, &flags);
6107 deactivate_task(p, task_rq(p));
6108 __setscheduler(p, SCHED_NORMAL, 0);
6110 __activate_task(p, task_rq(p));
6111 resched_task(rq->curr);
6114 task_rq_unlock(rq, &flags);
6116 read_unlock_irq(&tasklist_lock);
6119 #endif /* CONFIG_MAGIC_SYSRQ */
6123 * These functions are only useful for the IA64 MCA handling.
6125 * They can only be called when the whole system has been
6126 * stopped - every CPU needs to be quiescent, and no scheduling
6127 * activity can take place. Using them for anything else would
6128 * be a serious bug, and as a result, they aren't even visible
6129 * under any other configuration.
6133 * curr_task - return the current task for a given cpu.
6134 * @cpu: the processor in question.
6136 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6138 task_t *curr_task(int cpu)
6140 return cpu_curr(cpu);
6144 * set_curr_task - set the current task for a given cpu.
6145 * @cpu: the processor in question.
6146 * @p: the task pointer to set.
6148 * Description: This function must only be used when non-maskable interrupts
6149 * are serviced on a separate stack. It allows the architecture to switch the
6150 * notion of the current task on a cpu in a non-blocking manner. This function
6151 * must be called with all CPU's synchronized, and interrupts disabled, the
6152 * and caller must save the original value of the current task (see
6153 * curr_task() above) and restore that value before reenabling interrupts and
6154 * re-starting the system.
6156 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6158 void set_curr_task(int cpu, task_t *p)