4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
22 #include <linux/module.h>
23 #include <linux/nmi.h>
24 #include <linux/init.h>
25 #include <asm/uaccess.h>
26 #include <linux/highmem.h>
27 #include <linux/smp_lock.h>
28 #include <asm/mmu_context.h>
29 #include <linux/interrupt.h>
30 #include <linux/capability.h>
31 #include <linux/completion.h>
32 #include <linux/kernel_stat.h>
33 #include <linux/security.h>
34 #include <linux/notifier.h>
35 #include <linux/profile.h>
36 #include <linux/suspend.h>
37 #include <linux/vmalloc.h>
38 #include <linux/blkdev.h>
39 #include <linux/delay.h>
40 #include <linux/smp.h>
41 #include <linux/threads.h>
42 #include <linux/timer.h>
43 #include <linux/rcupdate.h>
44 #include <linux/cpu.h>
45 #include <linux/cpuset.h>
46 #include <linux/percpu.h>
47 #include <linux/kthread.h>
48 #include <linux/seq_file.h>
49 #include <linux/syscalls.h>
50 #include <linux/times.h>
51 #include <linux/acct.h>
52 #include <linux/kprobes.h>
55 #include <asm/unistd.h>
58 * Convert user-nice values [ -20 ... 0 ... 19 ]
59 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
62 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
63 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
64 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
67 * 'User priority' is the nice value converted to something we
68 * can work with better when scaling various scheduler parameters,
69 * it's a [ 0 ... 39 ] range.
71 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
72 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
73 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
76 * Some helpers for converting nanosecond timing to jiffy resolution
78 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
79 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
82 * These are the 'tuning knobs' of the scheduler:
84 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
85 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
86 * Timeslices get refilled after they expire.
88 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
89 #define DEF_TIMESLICE (100 * HZ / 1000)
90 #define ON_RUNQUEUE_WEIGHT 30
91 #define CHILD_PENALTY 95
92 #define PARENT_PENALTY 100
94 #define PRIO_BONUS_RATIO 25
95 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
96 #define INTERACTIVE_DELTA 2
97 #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
98 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
99 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
102 * If a task is 'interactive' then we reinsert it in the active
103 * array after it has expired its current timeslice. (it will not
104 * continue to run immediately, it will still roundrobin with
105 * other interactive tasks.)
107 * This part scales the interactivity limit depending on niceness.
109 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
110 * Here are a few examples of different nice levels:
112 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
113 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
114 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
115 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
116 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
118 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
119 * priority range a task can explore, a value of '1' means the
120 * task is rated interactive.)
122 * Ie. nice +19 tasks can never get 'interactive' enough to be
123 * reinserted into the active array. And only heavily CPU-hog nice -20
124 * tasks will be expired. Default nice 0 tasks are somewhere between,
125 * it takes some effort for them to get interactive, but it's not
129 #define CURRENT_BONUS(p) \
130 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
133 #define GRANULARITY (10 * HZ / 1000 ? : 1)
136 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
137 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
140 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
141 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
144 #define SCALE(v1,v1_max,v2_max) \
145 (v1) * (v2_max) / (v1_max)
148 (SCALE(TASK_NICE(p) + 20, 40, MAX_BONUS) - 20 * MAX_BONUS / 40 + \
151 #define TASK_INTERACTIVE(p) \
152 ((p)->prio <= (p)->static_prio - DELTA(p))
154 #define INTERACTIVE_SLEEP(p) \
155 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
156 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
158 #define TASK_PREEMPTS_CURR(p, rq) \
159 ((p)->prio < (rq)->curr->prio)
162 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
163 * to time slice values: [800ms ... 100ms ... 5ms]
165 * The higher a thread's priority, the bigger timeslices
166 * it gets during one round of execution. But even the lowest
167 * priority thread gets MIN_TIMESLICE worth of execution time.
170 #define SCALE_PRIO(x, prio) \
171 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
173 static unsigned int static_prio_timeslice(int static_prio)
175 if (static_prio < NICE_TO_PRIO(0))
176 return SCALE_PRIO(DEF_TIMESLICE * 4, static_prio);
178 return SCALE_PRIO(DEF_TIMESLICE, static_prio);
181 static inline unsigned int task_timeslice(task_t *p)
183 return static_prio_timeslice(p->static_prio);
186 #define task_hot(p, now, sd) ((long long) ((now) - (p)->last_ran) \
187 < (long long) (sd)->cache_hot_time)
190 * These are the runqueue data structures:
193 typedef struct runqueue runqueue_t;
196 unsigned int nr_active;
197 DECLARE_BITMAP(bitmap, MAX_PRIO+1); /* include 1 bit for delimiter */
198 struct list_head queue[MAX_PRIO];
202 * This is the main, per-CPU runqueue data structure.
204 * Locking rule: those places that want to lock multiple runqueues
205 * (such as the load balancing or the thread migration code), lock
206 * acquire operations must be ordered by ascending &runqueue.
212 * nr_running and cpu_load should be in the same cacheline because
213 * remote CPUs use both these fields when doing load calculation.
215 unsigned long nr_running;
216 unsigned long raw_weighted_load;
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 * To aid in avoiding the subversion of "niceness" due to uneven distribution
673 * of tasks with abnormal "nice" values across CPUs the contribution that
674 * each task makes to its run queue's load is weighted according to its
675 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
676 * scaled version of the new time slice allocation that they receive on time
681 * Assume: static_prio_timeslice(NICE_TO_PRIO(0)) == DEF_TIMESLICE
682 * If static_prio_timeslice() is ever changed to break this assumption then
683 * this code will need modification
685 #define TIME_SLICE_NICE_ZERO DEF_TIMESLICE
686 #define LOAD_WEIGHT(lp) \
687 (((lp) * SCHED_LOAD_SCALE) / TIME_SLICE_NICE_ZERO)
688 #define PRIO_TO_LOAD_WEIGHT(prio) \
689 LOAD_WEIGHT(static_prio_timeslice(prio))
690 #define RTPRIO_TO_LOAD_WEIGHT(rp) \
691 (PRIO_TO_LOAD_WEIGHT(MAX_RT_PRIO) + LOAD_WEIGHT(rp))
693 static void set_load_weight(task_t *p)
697 if (p == task_rq(p)->migration_thread)
699 * The migration thread does the actual balancing.
700 * Giving its load any weight will skew balancing
706 p->load_weight = RTPRIO_TO_LOAD_WEIGHT(p->rt_priority);
708 p->load_weight = PRIO_TO_LOAD_WEIGHT(p->static_prio);
711 static inline void inc_raw_weighted_load(runqueue_t *rq, const task_t *p)
713 rq->raw_weighted_load += p->load_weight;
716 static inline void dec_raw_weighted_load(runqueue_t *rq, const task_t *p)
718 rq->raw_weighted_load -= p->load_weight;
721 static inline void inc_nr_running(task_t *p, runqueue_t *rq)
724 inc_raw_weighted_load(rq, p);
727 static inline void dec_nr_running(task_t *p, runqueue_t *rq)
730 dec_raw_weighted_load(rq, p);
734 * __activate_task - move a task to the runqueue.
736 static void __activate_task(task_t *p, runqueue_t *rq)
738 prio_array_t *target = rq->active;
741 target = rq->expired;
742 enqueue_task(p, target);
743 inc_nr_running(p, rq);
747 * __activate_idle_task - move idle task to the _front_ of runqueue.
749 static inline void __activate_idle_task(task_t *p, runqueue_t *rq)
751 enqueue_task_head(p, rq->active);
752 inc_nr_running(p, rq);
755 static int recalc_task_prio(task_t *p, unsigned long long now)
757 /* Caller must always ensure 'now >= p->timestamp' */
758 unsigned long sleep_time = now - p->timestamp;
763 if (likely(sleep_time > 0)) {
765 * This ceiling is set to the lowest priority that would allow
766 * a task to be reinserted into the active array on timeslice
769 unsigned long ceiling = INTERACTIVE_SLEEP(p);
771 if (p->mm && sleep_time > ceiling && p->sleep_avg < ceiling) {
773 * Prevents user tasks from achieving best priority
774 * with one single large enough sleep.
776 p->sleep_avg = ceiling;
778 * Using INTERACTIVE_SLEEP() as a ceiling places a
779 * nice(0) task 1ms sleep away from promotion, and
780 * gives it 700ms to round-robin with no chance of
781 * being demoted. This is more than generous, so
782 * mark this sleep as non-interactive to prevent the
783 * on-runqueue bonus logic from intervening should
784 * this task not receive cpu immediately.
786 p->sleep_type = SLEEP_NONINTERACTIVE;
789 * Tasks waking from uninterruptible sleep are
790 * limited in their sleep_avg rise as they
791 * are likely to be waiting on I/O
793 if (p->sleep_type == SLEEP_NONINTERACTIVE && p->mm) {
794 if (p->sleep_avg >= ceiling)
796 else if (p->sleep_avg + sleep_time >=
798 p->sleep_avg = ceiling;
804 * This code gives a bonus to interactive tasks.
806 * The boost works by updating the 'average sleep time'
807 * value here, based on ->timestamp. The more time a
808 * task spends sleeping, the higher the average gets -
809 * and the higher the priority boost gets as well.
811 p->sleep_avg += sleep_time;
814 if (p->sleep_avg > NS_MAX_SLEEP_AVG)
815 p->sleep_avg = NS_MAX_SLEEP_AVG;
818 return effective_prio(p);
822 * activate_task - move a task to the runqueue and do priority recalculation
824 * Update all the scheduling statistics stuff. (sleep average
825 * calculation, priority modifiers, etc.)
827 static void activate_task(task_t *p, runqueue_t *rq, int local)
829 unsigned long long now;
834 /* Compensate for drifting sched_clock */
835 runqueue_t *this_rq = this_rq();
836 now = (now - this_rq->timestamp_last_tick)
837 + rq->timestamp_last_tick;
842 p->prio = recalc_task_prio(p, now);
845 * This checks to make sure it's not an uninterruptible task
846 * that is now waking up.
848 if (p->sleep_type == SLEEP_NORMAL) {
850 * Tasks which were woken up by interrupts (ie. hw events)
851 * are most likely of interactive nature. So we give them
852 * the credit of extending their sleep time to the period
853 * of time they spend on the runqueue, waiting for execution
854 * on a CPU, first time around:
857 p->sleep_type = SLEEP_INTERRUPTED;
860 * Normal first-time wakeups get a credit too for
861 * on-runqueue time, but it will be weighted down:
863 p->sleep_type = SLEEP_INTERACTIVE;
868 __activate_task(p, rq);
872 * deactivate_task - remove a task from the runqueue.
874 static void deactivate_task(struct task_struct *p, runqueue_t *rq)
876 dec_nr_running(p, rq);
877 dequeue_task(p, p->array);
882 * resched_task - mark a task 'to be rescheduled now'.
884 * On UP this means the setting of the need_resched flag, on SMP it
885 * might also involve a cross-CPU call to trigger the scheduler on
890 #ifndef tsk_is_polling
891 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
894 static void resched_task(task_t *p)
898 assert_spin_locked(&task_rq(p)->lock);
900 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
903 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
906 if (cpu == smp_processor_id())
909 /* NEED_RESCHED must be visible before we test polling */
911 if (!tsk_is_polling(p))
912 smp_send_reschedule(cpu);
915 static inline void resched_task(task_t *p)
917 assert_spin_locked(&task_rq(p)->lock);
918 set_tsk_need_resched(p);
923 * task_curr - is this task currently executing on a CPU?
924 * @p: the task in question.
926 inline int task_curr(const task_t *p)
928 return cpu_curr(task_cpu(p)) == p;
931 /* Used instead of source_load when we know the type == 0 */
932 unsigned long weighted_cpuload(const int cpu)
934 return cpu_rq(cpu)->raw_weighted_load;
939 struct list_head list;
944 struct completion done;
948 * The task's runqueue lock must be held.
949 * Returns true if you have to wait for migration thread.
951 static int migrate_task(task_t *p, int dest_cpu, migration_req_t *req)
953 runqueue_t *rq = task_rq(p);
956 * If the task is not on a runqueue (and not running), then
957 * it is sufficient to simply update the task's cpu field.
959 if (!p->array && !task_running(rq, p)) {
960 set_task_cpu(p, dest_cpu);
964 init_completion(&req->done);
966 req->dest_cpu = dest_cpu;
967 list_add(&req->list, &rq->migration_queue);
972 * wait_task_inactive - wait for a thread to unschedule.
974 * The caller must ensure that the task *will* unschedule sometime soon,
975 * else this function might spin for a *long* time. This function can't
976 * be called with interrupts off, or it may introduce deadlock with
977 * smp_call_function() if an IPI is sent by the same process we are
978 * waiting to become inactive.
980 void wait_task_inactive(task_t *p)
987 rq = task_rq_lock(p, &flags);
988 /* Must be off runqueue entirely, not preempted. */
989 if (unlikely(p->array || task_running(rq, p))) {
990 /* If it's preempted, we yield. It could be a while. */
991 preempted = !task_running(rq, p);
992 task_rq_unlock(rq, &flags);
998 task_rq_unlock(rq, &flags);
1002 * kick_process - kick a running thread to enter/exit the kernel
1003 * @p: the to-be-kicked thread
1005 * Cause a process which is running on another CPU to enter
1006 * kernel-mode, without any delay. (to get signals handled.)
1008 * NOTE: this function doesnt have to take the runqueue lock,
1009 * because all it wants to ensure is that the remote task enters
1010 * the kernel. If the IPI races and the task has been migrated
1011 * to another CPU then no harm is done and the purpose has been
1014 void kick_process(task_t *p)
1020 if ((cpu != smp_processor_id()) && task_curr(p))
1021 smp_send_reschedule(cpu);
1026 * Return a low guess at the load of a migration-source cpu weighted
1027 * according to the scheduling class and "nice" value.
1029 * We want to under-estimate the load of migration sources, to
1030 * balance conservatively.
1032 static inline unsigned long source_load(int cpu, int type)
1034 runqueue_t *rq = cpu_rq(cpu);
1037 return rq->raw_weighted_load;
1039 return min(rq->cpu_load[type-1], rq->raw_weighted_load);
1043 * Return a high guess at the load of a migration-target cpu weighted
1044 * according to the scheduling class and "nice" value.
1046 static inline unsigned long target_load(int cpu, int type)
1048 runqueue_t *rq = cpu_rq(cpu);
1051 return rq->raw_weighted_load;
1053 return max(rq->cpu_load[type-1], rq->raw_weighted_load);
1057 * Return the average load per task on the cpu's run queue
1059 static inline unsigned long cpu_avg_load_per_task(int cpu)
1061 runqueue_t *rq = cpu_rq(cpu);
1062 unsigned long n = rq->nr_running;
1064 return n ? rq->raw_weighted_load / n : SCHED_LOAD_SCALE;
1068 * find_idlest_group finds and returns the least busy CPU group within the
1071 static struct sched_group *
1072 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1074 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1075 unsigned long min_load = ULONG_MAX, this_load = 0;
1076 int load_idx = sd->forkexec_idx;
1077 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1080 unsigned long load, avg_load;
1084 /* Skip over this group if it has no CPUs allowed */
1085 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1088 local_group = cpu_isset(this_cpu, group->cpumask);
1090 /* Tally up the load of all CPUs in the group */
1093 for_each_cpu_mask(i, group->cpumask) {
1094 /* Bias balancing toward cpus of our domain */
1096 load = source_load(i, load_idx);
1098 load = target_load(i, load_idx);
1103 /* Adjust by relative CPU power of the group */
1104 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1107 this_load = avg_load;
1109 } else if (avg_load < min_load) {
1110 min_load = avg_load;
1114 group = group->next;
1115 } while (group != sd->groups);
1117 if (!idlest || 100*this_load < imbalance*min_load)
1123 * find_idlest_queue - find the idlest runqueue among the cpus in group.
1126 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1129 unsigned long load, min_load = ULONG_MAX;
1133 /* Traverse only the allowed CPUs */
1134 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1136 for_each_cpu_mask(i, tmp) {
1137 load = weighted_cpuload(i);
1139 if (load < min_load || (load == min_load && i == this_cpu)) {
1149 * sched_balance_self: balance the current task (running on cpu) in domains
1150 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1153 * Balance, ie. select the least loaded group.
1155 * Returns the target CPU number, or the same CPU if no balancing is needed.
1157 * preempt must be disabled.
1159 static int sched_balance_self(int cpu, int flag)
1161 struct task_struct *t = current;
1162 struct sched_domain *tmp, *sd = NULL;
1164 for_each_domain(cpu, tmp) {
1165 if (tmp->flags & flag)
1171 struct sched_group *group;
1176 group = find_idlest_group(sd, t, cpu);
1180 new_cpu = find_idlest_cpu(group, t, cpu);
1181 if (new_cpu == -1 || new_cpu == cpu)
1184 /* Now try balancing at a lower domain level */
1188 weight = cpus_weight(span);
1189 for_each_domain(cpu, tmp) {
1190 if (weight <= cpus_weight(tmp->span))
1192 if (tmp->flags & flag)
1195 /* while loop will break here if sd == NULL */
1201 #endif /* CONFIG_SMP */
1204 * wake_idle() will wake a task on an idle cpu if task->cpu is
1205 * not idle and an idle cpu is available. The span of cpus to
1206 * search starts with cpus closest then further out as needed,
1207 * so we always favor a closer, idle cpu.
1209 * Returns the CPU we should wake onto.
1211 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1212 static int wake_idle(int cpu, task_t *p)
1215 struct sched_domain *sd;
1221 for_each_domain(cpu, sd) {
1222 if (sd->flags & SD_WAKE_IDLE) {
1223 cpus_and(tmp, sd->span, p->cpus_allowed);
1224 for_each_cpu_mask(i, tmp) {
1235 static inline int wake_idle(int cpu, task_t *p)
1242 * try_to_wake_up - wake up a thread
1243 * @p: the to-be-woken-up thread
1244 * @state: the mask of task states that can be woken
1245 * @sync: do a synchronous wakeup?
1247 * Put it on the run-queue if it's not already there. The "current"
1248 * thread is always on the run-queue (except when the actual
1249 * re-schedule is in progress), and as such you're allowed to do
1250 * the simpler "current->state = TASK_RUNNING" to mark yourself
1251 * runnable without the overhead of this.
1253 * returns failure only if the task is already active.
1255 static int try_to_wake_up(task_t *p, unsigned int state, int sync)
1257 int cpu, this_cpu, success = 0;
1258 unsigned long flags;
1262 unsigned long load, this_load;
1263 struct sched_domain *sd, *this_sd = NULL;
1267 rq = task_rq_lock(p, &flags);
1268 old_state = p->state;
1269 if (!(old_state & state))
1276 this_cpu = smp_processor_id();
1279 if (unlikely(task_running(rq, p)))
1284 schedstat_inc(rq, ttwu_cnt);
1285 if (cpu == this_cpu) {
1286 schedstat_inc(rq, ttwu_local);
1290 for_each_domain(this_cpu, sd) {
1291 if (cpu_isset(cpu, sd->span)) {
1292 schedstat_inc(sd, ttwu_wake_remote);
1298 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1302 * Check for affine wakeup and passive balancing possibilities.
1305 int idx = this_sd->wake_idx;
1306 unsigned int imbalance;
1308 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1310 load = source_load(cpu, idx);
1311 this_load = target_load(this_cpu, idx);
1313 new_cpu = this_cpu; /* Wake to this CPU if we can */
1315 if (this_sd->flags & SD_WAKE_AFFINE) {
1316 unsigned long tl = this_load;
1317 unsigned long tl_per_task = cpu_avg_load_per_task(this_cpu);
1320 * If sync wakeup then subtract the (maximum possible)
1321 * effect of the currently running task from the load
1322 * of the current CPU:
1325 tl -= current->load_weight;
1328 tl + target_load(cpu, idx) <= tl_per_task) ||
1329 100*(tl + p->load_weight) <= imbalance*load) {
1331 * This domain has SD_WAKE_AFFINE and
1332 * p is cache cold in this domain, and
1333 * there is no bad imbalance.
1335 schedstat_inc(this_sd, ttwu_move_affine);
1341 * Start passive balancing when half the imbalance_pct
1344 if (this_sd->flags & SD_WAKE_BALANCE) {
1345 if (imbalance*this_load <= 100*load) {
1346 schedstat_inc(this_sd, ttwu_move_balance);
1352 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1354 new_cpu = wake_idle(new_cpu, p);
1355 if (new_cpu != cpu) {
1356 set_task_cpu(p, new_cpu);
1357 task_rq_unlock(rq, &flags);
1358 /* might preempt at this point */
1359 rq = task_rq_lock(p, &flags);
1360 old_state = p->state;
1361 if (!(old_state & state))
1366 this_cpu = smp_processor_id();
1371 #endif /* CONFIG_SMP */
1372 if (old_state == TASK_UNINTERRUPTIBLE) {
1373 rq->nr_uninterruptible--;
1375 * Tasks on involuntary sleep don't earn
1376 * sleep_avg beyond just interactive state.
1378 p->sleep_type = SLEEP_NONINTERACTIVE;
1382 * Tasks that have marked their sleep as noninteractive get
1383 * woken up with their sleep average not weighted in an
1386 if (old_state & TASK_NONINTERACTIVE)
1387 p->sleep_type = SLEEP_NONINTERACTIVE;
1390 activate_task(p, rq, cpu == this_cpu);
1392 * Sync wakeups (i.e. those types of wakeups where the waker
1393 * has indicated that it will leave the CPU in short order)
1394 * don't trigger a preemption, if the woken up task will run on
1395 * this cpu. (in this case the 'I will reschedule' promise of
1396 * the waker guarantees that the freshly woken up task is going
1397 * to be considered on this CPU.)
1399 if (!sync || cpu != this_cpu) {
1400 if (TASK_PREEMPTS_CURR(p, rq))
1401 resched_task(rq->curr);
1406 p->state = TASK_RUNNING;
1408 task_rq_unlock(rq, &flags);
1413 int fastcall wake_up_process(task_t *p)
1415 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1416 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1419 EXPORT_SYMBOL(wake_up_process);
1421 int fastcall wake_up_state(task_t *p, unsigned int state)
1423 return try_to_wake_up(p, state, 0);
1427 * Perform scheduler related setup for a newly forked process p.
1428 * p is forked by current.
1430 void fastcall sched_fork(task_t *p, int clone_flags)
1432 int cpu = get_cpu();
1435 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1437 set_task_cpu(p, cpu);
1440 * We mark the process as running here, but have not actually
1441 * inserted it onto the runqueue yet. This guarantees that
1442 * nobody will actually run it, and a signal or other external
1443 * event cannot wake it up and insert it on the runqueue either.
1445 p->state = TASK_RUNNING;
1446 INIT_LIST_HEAD(&p->run_list);
1448 #ifdef CONFIG_SCHEDSTATS
1449 memset(&p->sched_info, 0, sizeof(p->sched_info));
1451 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1454 #ifdef CONFIG_PREEMPT
1455 /* Want to start with kernel preemption disabled. */
1456 task_thread_info(p)->preempt_count = 1;
1459 * Share the timeslice between parent and child, thus the
1460 * total amount of pending timeslices in the system doesn't change,
1461 * resulting in more scheduling fairness.
1463 local_irq_disable();
1464 p->time_slice = (current->time_slice + 1) >> 1;
1466 * The remainder of the first timeslice might be recovered by
1467 * the parent if the child exits early enough.
1469 p->first_time_slice = 1;
1470 current->time_slice >>= 1;
1471 p->timestamp = sched_clock();
1472 if (unlikely(!current->time_slice)) {
1474 * This case is rare, it happens when the parent has only
1475 * a single jiffy left from its timeslice. Taking the
1476 * runqueue lock is not a problem.
1478 current->time_slice = 1;
1486 * wake_up_new_task - wake up a newly created task for the first time.
1488 * This function will do some initial scheduler statistics housekeeping
1489 * that must be done for every newly created context, then puts the task
1490 * on the runqueue and wakes it.
1492 void fastcall wake_up_new_task(task_t *p, unsigned long clone_flags)
1494 unsigned long flags;
1496 runqueue_t *rq, *this_rq;
1498 rq = task_rq_lock(p, &flags);
1499 BUG_ON(p->state != TASK_RUNNING);
1500 this_cpu = smp_processor_id();
1504 * We decrease the sleep average of forking parents
1505 * and children as well, to keep max-interactive tasks
1506 * from forking tasks that are max-interactive. The parent
1507 * (current) is done further down, under its lock.
1509 p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1510 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1512 p->prio = effective_prio(p);
1514 if (likely(cpu == this_cpu)) {
1515 if (!(clone_flags & CLONE_VM)) {
1517 * The VM isn't cloned, so we're in a good position to
1518 * do child-runs-first in anticipation of an exec. This
1519 * usually avoids a lot of COW overhead.
1521 if (unlikely(!current->array))
1522 __activate_task(p, rq);
1524 p->prio = current->prio;
1525 list_add_tail(&p->run_list, ¤t->run_list);
1526 p->array = current->array;
1527 p->array->nr_active++;
1528 inc_nr_running(p, rq);
1532 /* Run child last */
1533 __activate_task(p, rq);
1535 * We skip the following code due to cpu == this_cpu
1537 * task_rq_unlock(rq, &flags);
1538 * this_rq = task_rq_lock(current, &flags);
1542 this_rq = cpu_rq(this_cpu);
1545 * Not the local CPU - must adjust timestamp. This should
1546 * get optimised away in the !CONFIG_SMP case.
1548 p->timestamp = (p->timestamp - this_rq->timestamp_last_tick)
1549 + rq->timestamp_last_tick;
1550 __activate_task(p, rq);
1551 if (TASK_PREEMPTS_CURR(p, rq))
1552 resched_task(rq->curr);
1555 * Parent and child are on different CPUs, now get the
1556 * parent runqueue to update the parent's ->sleep_avg:
1558 task_rq_unlock(rq, &flags);
1559 this_rq = task_rq_lock(current, &flags);
1561 current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1562 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1563 task_rq_unlock(this_rq, &flags);
1567 * Potentially available exiting-child timeslices are
1568 * retrieved here - this way the parent does not get
1569 * penalized for creating too many threads.
1571 * (this cannot be used to 'generate' timeslices
1572 * artificially, because any timeslice recovered here
1573 * was given away by the parent in the first place.)
1575 void fastcall sched_exit(task_t *p)
1577 unsigned long flags;
1581 * If the child was a (relative-) CPU hog then decrease
1582 * the sleep_avg of the parent as well.
1584 rq = task_rq_lock(p->parent, &flags);
1585 if (p->first_time_slice && task_cpu(p) == task_cpu(p->parent)) {
1586 p->parent->time_slice += p->time_slice;
1587 if (unlikely(p->parent->time_slice > task_timeslice(p)))
1588 p->parent->time_slice = task_timeslice(p);
1590 if (p->sleep_avg < p->parent->sleep_avg)
1591 p->parent->sleep_avg = p->parent->sleep_avg /
1592 (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg /
1594 task_rq_unlock(rq, &flags);
1598 * prepare_task_switch - prepare to switch tasks
1599 * @rq: the runqueue preparing to switch
1600 * @next: the task we are going to switch to.
1602 * This is called with the rq lock held and interrupts off. It must
1603 * be paired with a subsequent finish_task_switch after the context
1606 * prepare_task_switch sets up locking and calls architecture specific
1609 static inline void prepare_task_switch(runqueue_t *rq, task_t *next)
1611 prepare_lock_switch(rq, next);
1612 prepare_arch_switch(next);
1616 * finish_task_switch - clean up after a task-switch
1617 * @rq: runqueue associated with task-switch
1618 * @prev: the thread we just switched away from.
1620 * finish_task_switch must be called after the context switch, paired
1621 * with a prepare_task_switch call before the context switch.
1622 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1623 * and do any other architecture-specific cleanup actions.
1625 * Note that we may have delayed dropping an mm in context_switch(). If
1626 * so, we finish that here outside of the runqueue lock. (Doing it
1627 * with the lock held can cause deadlocks; see schedule() for
1630 static inline void finish_task_switch(runqueue_t *rq, task_t *prev)
1631 __releases(rq->lock)
1633 struct mm_struct *mm = rq->prev_mm;
1634 unsigned long prev_task_flags;
1639 * A task struct has one reference for the use as "current".
1640 * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1641 * calls schedule one last time. The schedule call will never return,
1642 * and the scheduled task must drop that reference.
1643 * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1644 * still held, otherwise prev could be scheduled on another cpu, die
1645 * there before we look at prev->state, and then the reference would
1647 * Manfred Spraul <manfred@colorfullife.com>
1649 prev_task_flags = prev->flags;
1650 finish_arch_switch(prev);
1651 finish_lock_switch(rq, prev);
1654 if (unlikely(prev_task_flags & PF_DEAD)) {
1656 * Remove function-return probe instances associated with this
1657 * task and put them back on the free list.
1659 kprobe_flush_task(prev);
1660 put_task_struct(prev);
1665 * schedule_tail - first thing a freshly forked thread must call.
1666 * @prev: the thread we just switched away from.
1668 asmlinkage void schedule_tail(task_t *prev)
1669 __releases(rq->lock)
1671 runqueue_t *rq = this_rq();
1672 finish_task_switch(rq, prev);
1673 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1674 /* In this case, finish_task_switch does not reenable preemption */
1677 if (current->set_child_tid)
1678 put_user(current->pid, current->set_child_tid);
1682 * context_switch - switch to the new MM and the new
1683 * thread's register state.
1686 task_t * context_switch(runqueue_t *rq, task_t *prev, task_t *next)
1688 struct mm_struct *mm = next->mm;
1689 struct mm_struct *oldmm = prev->active_mm;
1691 if (unlikely(!mm)) {
1692 next->active_mm = oldmm;
1693 atomic_inc(&oldmm->mm_count);
1694 enter_lazy_tlb(oldmm, next);
1696 switch_mm(oldmm, mm, next);
1698 if (unlikely(!prev->mm)) {
1699 prev->active_mm = NULL;
1700 WARN_ON(rq->prev_mm);
1701 rq->prev_mm = oldmm;
1704 /* Here we just switch the register state and the stack. */
1705 switch_to(prev, next, prev);
1711 * nr_running, nr_uninterruptible and nr_context_switches:
1713 * externally visible scheduler statistics: current number of runnable
1714 * threads, current number of uninterruptible-sleeping threads, total
1715 * number of context switches performed since bootup.
1717 unsigned long nr_running(void)
1719 unsigned long i, sum = 0;
1721 for_each_online_cpu(i)
1722 sum += cpu_rq(i)->nr_running;
1727 unsigned long nr_uninterruptible(void)
1729 unsigned long i, sum = 0;
1731 for_each_possible_cpu(i)
1732 sum += cpu_rq(i)->nr_uninterruptible;
1735 * Since we read the counters lockless, it might be slightly
1736 * inaccurate. Do not allow it to go below zero though:
1738 if (unlikely((long)sum < 0))
1744 unsigned long long nr_context_switches(void)
1747 unsigned long long sum = 0;
1749 for_each_possible_cpu(i)
1750 sum += cpu_rq(i)->nr_switches;
1755 unsigned long nr_iowait(void)
1757 unsigned long i, sum = 0;
1759 for_each_possible_cpu(i)
1760 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1765 unsigned long nr_active(void)
1767 unsigned long i, running = 0, uninterruptible = 0;
1769 for_each_online_cpu(i) {
1770 running += cpu_rq(i)->nr_running;
1771 uninterruptible += cpu_rq(i)->nr_uninterruptible;
1774 if (unlikely((long)uninterruptible < 0))
1775 uninterruptible = 0;
1777 return running + uninterruptible;
1783 * double_rq_lock - safely lock two runqueues
1785 * Note this does not disable interrupts like task_rq_lock,
1786 * you need to do so manually before calling.
1788 static void double_rq_lock(runqueue_t *rq1, runqueue_t *rq2)
1789 __acquires(rq1->lock)
1790 __acquires(rq2->lock)
1793 spin_lock(&rq1->lock);
1794 __acquire(rq2->lock); /* Fake it out ;) */
1797 spin_lock(&rq1->lock);
1798 spin_lock(&rq2->lock);
1800 spin_lock(&rq2->lock);
1801 spin_lock(&rq1->lock);
1807 * double_rq_unlock - safely unlock two runqueues
1809 * Note this does not restore interrupts like task_rq_unlock,
1810 * you need to do so manually after calling.
1812 static void double_rq_unlock(runqueue_t *rq1, runqueue_t *rq2)
1813 __releases(rq1->lock)
1814 __releases(rq2->lock)
1816 spin_unlock(&rq1->lock);
1818 spin_unlock(&rq2->lock);
1820 __release(rq2->lock);
1824 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1826 static void double_lock_balance(runqueue_t *this_rq, runqueue_t *busiest)
1827 __releases(this_rq->lock)
1828 __acquires(busiest->lock)
1829 __acquires(this_rq->lock)
1831 if (unlikely(!spin_trylock(&busiest->lock))) {
1832 if (busiest < this_rq) {
1833 spin_unlock(&this_rq->lock);
1834 spin_lock(&busiest->lock);
1835 spin_lock(&this_rq->lock);
1837 spin_lock(&busiest->lock);
1842 * If dest_cpu is allowed for this process, migrate the task to it.
1843 * This is accomplished by forcing the cpu_allowed mask to only
1844 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
1845 * the cpu_allowed mask is restored.
1847 static void sched_migrate_task(task_t *p, int dest_cpu)
1849 migration_req_t req;
1851 unsigned long flags;
1853 rq = task_rq_lock(p, &flags);
1854 if (!cpu_isset(dest_cpu, p->cpus_allowed)
1855 || unlikely(cpu_is_offline(dest_cpu)))
1858 /* force the process onto the specified CPU */
1859 if (migrate_task(p, dest_cpu, &req)) {
1860 /* Need to wait for migration thread (might exit: take ref). */
1861 struct task_struct *mt = rq->migration_thread;
1862 get_task_struct(mt);
1863 task_rq_unlock(rq, &flags);
1864 wake_up_process(mt);
1865 put_task_struct(mt);
1866 wait_for_completion(&req.done);
1870 task_rq_unlock(rq, &flags);
1874 * sched_exec - execve() is a valuable balancing opportunity, because at
1875 * this point the task has the smallest effective memory and cache footprint.
1877 void sched_exec(void)
1879 int new_cpu, this_cpu = get_cpu();
1880 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
1882 if (new_cpu != this_cpu)
1883 sched_migrate_task(current, new_cpu);
1887 * pull_task - move a task from a remote runqueue to the local runqueue.
1888 * Both runqueues must be locked.
1891 void pull_task(runqueue_t *src_rq, prio_array_t *src_array, task_t *p,
1892 runqueue_t *this_rq, prio_array_t *this_array, int this_cpu)
1894 dequeue_task(p, src_array);
1895 dec_nr_running(p, src_rq);
1896 set_task_cpu(p, this_cpu);
1897 inc_nr_running(p, this_rq);
1898 enqueue_task(p, this_array);
1899 p->timestamp = (p->timestamp - src_rq->timestamp_last_tick)
1900 + this_rq->timestamp_last_tick;
1902 * Note that idle threads have a prio of MAX_PRIO, for this test
1903 * to be always true for them.
1905 if (TASK_PREEMPTS_CURR(p, this_rq))
1906 resched_task(this_rq->curr);
1910 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1913 int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu,
1914 struct sched_domain *sd, enum idle_type idle,
1918 * We do not migrate tasks that are:
1919 * 1) running (obviously), or
1920 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1921 * 3) are cache-hot on their current CPU.
1923 if (!cpu_isset(this_cpu, p->cpus_allowed))
1927 if (task_running(rq, p))
1931 * Aggressive migration if:
1932 * 1) task is cache cold, or
1933 * 2) too many balance attempts have failed.
1936 if (sd->nr_balance_failed > sd->cache_nice_tries)
1939 if (task_hot(p, rq->timestamp_last_tick, sd))
1945 * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted
1946 * load from busiest to this_rq, as part of a balancing operation within
1947 * "domain". Returns the number of tasks moved.
1949 * Called with both runqueues locked.
1951 static int move_tasks(runqueue_t *this_rq, int this_cpu, runqueue_t *busiest,
1952 unsigned long max_nr_move, unsigned long max_load_move,
1953 struct sched_domain *sd, enum idle_type idle,
1956 prio_array_t *array, *dst_array;
1957 struct list_head *head, *curr;
1958 int idx, pulled = 0, pinned = 0, this_min_prio;
1962 if (max_nr_move == 0 || max_load_move == 0)
1965 rem_load_move = max_load_move;
1967 this_min_prio = this_rq->curr->prio;
1970 * We first consider expired tasks. Those will likely not be
1971 * executed in the near future, and they are most likely to
1972 * be cache-cold, thus switching CPUs has the least effect
1975 if (busiest->expired->nr_active) {
1976 array = busiest->expired;
1977 dst_array = this_rq->expired;
1979 array = busiest->active;
1980 dst_array = this_rq->active;
1984 /* Start searching at priority 0: */
1988 idx = sched_find_first_bit(array->bitmap);
1990 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
1991 if (idx >= MAX_PRIO) {
1992 if (array == busiest->expired && busiest->active->nr_active) {
1993 array = busiest->active;
1994 dst_array = this_rq->active;
2000 head = array->queue + idx;
2003 tmp = list_entry(curr, task_t, run_list);
2008 * To help distribute high priority tasks accross CPUs we don't
2009 * skip a task if it will be the highest priority task (i.e. smallest
2010 * prio value) on its new queue regardless of its load weight
2012 if ((idx >= this_min_prio && tmp->load_weight > rem_load_move) ||
2013 !can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
2020 #ifdef CONFIG_SCHEDSTATS
2021 if (task_hot(tmp, busiest->timestamp_last_tick, sd))
2022 schedstat_inc(sd, lb_hot_gained[idle]);
2025 pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
2027 rem_load_move -= tmp->load_weight;
2030 * We only want to steal up to the prescribed number of tasks
2031 * and the prescribed amount of weighted load.
2033 if (pulled < max_nr_move && rem_load_move > 0) {
2034 if (idx < this_min_prio)
2035 this_min_prio = idx;
2043 * Right now, this is the only place pull_task() is called,
2044 * so we can safely collect pull_task() stats here rather than
2045 * inside pull_task().
2047 schedstat_add(sd, lb_gained[idle], pulled);
2050 *all_pinned = pinned;
2055 * find_busiest_group finds and returns the busiest CPU group within the
2056 * domain. It calculates and returns the amount of weighted load which should be
2057 * moved to restore balance via the imbalance parameter.
2059 static struct sched_group *
2060 find_busiest_group(struct sched_domain *sd, int this_cpu,
2061 unsigned long *imbalance, enum idle_type idle, int *sd_idle)
2063 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2064 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2065 unsigned long max_pull;
2066 unsigned long busiest_load_per_task, busiest_nr_running;
2067 unsigned long this_load_per_task, this_nr_running;
2070 max_load = this_load = total_load = total_pwr = 0;
2071 busiest_load_per_task = busiest_nr_running = 0;
2072 this_load_per_task = this_nr_running = 0;
2073 if (idle == NOT_IDLE)
2074 load_idx = sd->busy_idx;
2075 else if (idle == NEWLY_IDLE)
2076 load_idx = sd->newidle_idx;
2078 load_idx = sd->idle_idx;
2084 unsigned long sum_nr_running, sum_weighted_load;
2086 local_group = cpu_isset(this_cpu, group->cpumask);
2088 /* Tally up the load of all CPUs in the group */
2089 sum_weighted_load = sum_nr_running = avg_load = 0;
2091 for_each_cpu_mask(i, group->cpumask) {
2092 runqueue_t *rq = cpu_rq(i);
2094 if (*sd_idle && !idle_cpu(i))
2097 /* Bias balancing toward cpus of our domain */
2099 load = target_load(i, load_idx);
2101 load = source_load(i, load_idx);
2104 sum_nr_running += rq->nr_running;
2105 sum_weighted_load += rq->raw_weighted_load;
2108 total_load += avg_load;
2109 total_pwr += group->cpu_power;
2111 /* Adjust by relative CPU power of the group */
2112 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
2115 this_load = avg_load;
2117 this_nr_running = sum_nr_running;
2118 this_load_per_task = sum_weighted_load;
2119 } else if (avg_load > max_load &&
2120 sum_nr_running > group->cpu_power / SCHED_LOAD_SCALE) {
2121 max_load = avg_load;
2123 busiest_nr_running = sum_nr_running;
2124 busiest_load_per_task = sum_weighted_load;
2126 group = group->next;
2127 } while (group != sd->groups);
2129 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2132 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2134 if (this_load >= avg_load ||
2135 100*max_load <= sd->imbalance_pct*this_load)
2138 busiest_load_per_task /= busiest_nr_running;
2140 * We're trying to get all the cpus to the average_load, so we don't
2141 * want to push ourselves above the average load, nor do we wish to
2142 * reduce the max loaded cpu below the average load, as either of these
2143 * actions would just result in more rebalancing later, and ping-pong
2144 * tasks around. Thus we look for the minimum possible imbalance.
2145 * Negative imbalances (*we* are more loaded than anyone else) will
2146 * be counted as no imbalance for these purposes -- we can't fix that
2147 * by pulling tasks to us. Be careful of negative numbers as they'll
2148 * appear as very large values with unsigned longs.
2150 if (max_load <= busiest_load_per_task)
2154 * In the presence of smp nice balancing, certain scenarios can have
2155 * max load less than avg load(as we skip the groups at or below
2156 * its cpu_power, while calculating max_load..)
2158 if (max_load < avg_load) {
2160 goto small_imbalance;
2163 /* Don't want to pull so many tasks that a group would go idle */
2164 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2166 /* How much load to actually move to equalise the imbalance */
2167 *imbalance = min(max_pull * busiest->cpu_power,
2168 (avg_load - this_load) * this->cpu_power)
2172 * if *imbalance is less than the average load per runnable task
2173 * there is no gaurantee that any tasks will be moved so we'll have
2174 * a think about bumping its value to force at least one task to be
2177 if (*imbalance < busiest_load_per_task) {
2178 unsigned long pwr_now, pwr_move;
2183 pwr_move = pwr_now = 0;
2185 if (this_nr_running) {
2186 this_load_per_task /= this_nr_running;
2187 if (busiest_load_per_task > this_load_per_task)
2190 this_load_per_task = SCHED_LOAD_SCALE;
2192 if (max_load - this_load >= busiest_load_per_task * imbn) {
2193 *imbalance = busiest_load_per_task;
2198 * OK, we don't have enough imbalance to justify moving tasks,
2199 * however we may be able to increase total CPU power used by
2203 pwr_now += busiest->cpu_power *
2204 min(busiest_load_per_task, max_load);
2205 pwr_now += this->cpu_power *
2206 min(this_load_per_task, this_load);
2207 pwr_now /= SCHED_LOAD_SCALE;
2209 /* Amount of load we'd subtract */
2210 tmp = busiest_load_per_task*SCHED_LOAD_SCALE/busiest->cpu_power;
2212 pwr_move += busiest->cpu_power *
2213 min(busiest_load_per_task, max_load - tmp);
2215 /* Amount of load we'd add */
2216 if (max_load*busiest->cpu_power <
2217 busiest_load_per_task*SCHED_LOAD_SCALE)
2218 tmp = max_load*busiest->cpu_power/this->cpu_power;
2220 tmp = busiest_load_per_task*SCHED_LOAD_SCALE/this->cpu_power;
2221 pwr_move += this->cpu_power*min(this_load_per_task, this_load + tmp);
2222 pwr_move /= SCHED_LOAD_SCALE;
2224 /* Move if we gain throughput */
2225 if (pwr_move <= pwr_now)
2228 *imbalance = busiest_load_per_task;
2240 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2242 static runqueue_t *find_busiest_queue(struct sched_group *group,
2243 enum idle_type idle, unsigned long imbalance)
2245 unsigned long max_load = 0;
2246 runqueue_t *busiest = NULL, *rqi;
2249 for_each_cpu_mask(i, group->cpumask) {
2252 if (rqi->nr_running == 1 && rqi->raw_weighted_load > imbalance)
2255 if (rqi->raw_weighted_load > max_load) {
2256 max_load = rqi->raw_weighted_load;
2265 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2266 * so long as it is large enough.
2268 #define MAX_PINNED_INTERVAL 512
2270 #define minus_1_or_zero(n) ((n) > 0 ? (n) - 1 : 0)
2272 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2273 * tasks if there is an imbalance.
2275 * Called with this_rq unlocked.
2277 static int load_balance(int this_cpu, runqueue_t *this_rq,
2278 struct sched_domain *sd, enum idle_type idle)
2280 struct sched_group *group;
2281 runqueue_t *busiest;
2282 unsigned long imbalance;
2283 int nr_moved, all_pinned = 0;
2284 int active_balance = 0;
2287 if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER)
2290 schedstat_inc(sd, lb_cnt[idle]);
2292 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle);
2294 schedstat_inc(sd, lb_nobusyg[idle]);
2298 busiest = find_busiest_queue(group, idle, imbalance);
2300 schedstat_inc(sd, lb_nobusyq[idle]);
2304 BUG_ON(busiest == this_rq);
2306 schedstat_add(sd, lb_imbalance[idle], imbalance);
2309 if (busiest->nr_running > 1) {
2311 * Attempt to move tasks. If find_busiest_group has found
2312 * an imbalance but busiest->nr_running <= 1, the group is
2313 * still unbalanced. nr_moved simply stays zero, so it is
2314 * correctly treated as an imbalance.
2316 double_rq_lock(this_rq, busiest);
2317 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2318 minus_1_or_zero(busiest->nr_running),
2319 imbalance, sd, idle, &all_pinned);
2320 double_rq_unlock(this_rq, busiest);
2322 /* All tasks on this runqueue were pinned by CPU affinity */
2323 if (unlikely(all_pinned))
2328 schedstat_inc(sd, lb_failed[idle]);
2329 sd->nr_balance_failed++;
2331 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2333 spin_lock(&busiest->lock);
2335 /* don't kick the migration_thread, if the curr
2336 * task on busiest cpu can't be moved to this_cpu
2338 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2339 spin_unlock(&busiest->lock);
2341 goto out_one_pinned;
2344 if (!busiest->active_balance) {
2345 busiest->active_balance = 1;
2346 busiest->push_cpu = this_cpu;
2349 spin_unlock(&busiest->lock);
2351 wake_up_process(busiest->migration_thread);
2354 * We've kicked active balancing, reset the failure
2357 sd->nr_balance_failed = sd->cache_nice_tries+1;
2360 sd->nr_balance_failed = 0;
2362 if (likely(!active_balance)) {
2363 /* We were unbalanced, so reset the balancing interval */
2364 sd->balance_interval = sd->min_interval;
2367 * If we've begun active balancing, start to back off. This
2368 * case may not be covered by the all_pinned logic if there
2369 * is only 1 task on the busy runqueue (because we don't call
2372 if (sd->balance_interval < sd->max_interval)
2373 sd->balance_interval *= 2;
2376 if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2381 schedstat_inc(sd, lb_balanced[idle]);
2383 sd->nr_balance_failed = 0;
2386 /* tune up the balancing interval */
2387 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2388 (sd->balance_interval < sd->max_interval))
2389 sd->balance_interval *= 2;
2391 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2397 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2398 * tasks if there is an imbalance.
2400 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2401 * this_rq is locked.
2403 static int load_balance_newidle(int this_cpu, runqueue_t *this_rq,
2404 struct sched_domain *sd)
2406 struct sched_group *group;
2407 runqueue_t *busiest = NULL;
2408 unsigned long imbalance;
2412 if (sd->flags & SD_SHARE_CPUPOWER)
2415 schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
2416 group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE, &sd_idle);
2418 schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]);
2422 busiest = find_busiest_queue(group, NEWLY_IDLE, imbalance);
2424 schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]);
2428 BUG_ON(busiest == this_rq);
2430 schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
2433 if (busiest->nr_running > 1) {
2434 /* Attempt to move tasks */
2435 double_lock_balance(this_rq, busiest);
2436 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2437 minus_1_or_zero(busiest->nr_running),
2438 imbalance, sd, NEWLY_IDLE, NULL);
2439 spin_unlock(&busiest->lock);
2443 schedstat_inc(sd, lb_failed[NEWLY_IDLE]);
2444 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2447 sd->nr_balance_failed = 0;
2452 schedstat_inc(sd, lb_balanced[NEWLY_IDLE]);
2453 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2455 sd->nr_balance_failed = 0;
2460 * idle_balance is called by schedule() if this_cpu is about to become
2461 * idle. Attempts to pull tasks from other CPUs.
2463 static void idle_balance(int this_cpu, runqueue_t *this_rq)
2465 struct sched_domain *sd;
2467 for_each_domain(this_cpu, sd) {
2468 if (sd->flags & SD_BALANCE_NEWIDLE) {
2469 if (load_balance_newidle(this_cpu, this_rq, sd)) {
2470 /* We've pulled tasks over so stop searching */
2478 * active_load_balance is run by migration threads. It pushes running tasks
2479 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2480 * running on each physical CPU where possible, and avoids physical /
2481 * logical imbalances.
2483 * Called with busiest_rq locked.
2485 static void active_load_balance(runqueue_t *busiest_rq, int busiest_cpu)
2487 struct sched_domain *sd;
2488 runqueue_t *target_rq;
2489 int target_cpu = busiest_rq->push_cpu;
2491 if (busiest_rq->nr_running <= 1)
2492 /* no task to move */
2495 target_rq = cpu_rq(target_cpu);
2498 * This condition is "impossible", if it occurs
2499 * we need to fix it. Originally reported by
2500 * Bjorn Helgaas on a 128-cpu setup.
2502 BUG_ON(busiest_rq == target_rq);
2504 /* move a task from busiest_rq to target_rq */
2505 double_lock_balance(busiest_rq, target_rq);
2507 /* Search for an sd spanning us and the target CPU. */
2508 for_each_domain(target_cpu, sd) {
2509 if ((sd->flags & SD_LOAD_BALANCE) &&
2510 cpu_isset(busiest_cpu, sd->span))
2514 if (unlikely(sd == NULL))
2517 schedstat_inc(sd, alb_cnt);
2519 if (move_tasks(target_rq, target_cpu, busiest_rq, 1,
2520 RTPRIO_TO_LOAD_WEIGHT(100), sd, SCHED_IDLE, NULL))
2521 schedstat_inc(sd, alb_pushed);
2523 schedstat_inc(sd, alb_failed);
2525 spin_unlock(&target_rq->lock);
2529 * rebalance_tick will get called every timer tick, on every CPU.
2531 * It checks each scheduling domain to see if it is due to be balanced,
2532 * and initiates a balancing operation if so.
2534 * Balancing parameters are set up in arch_init_sched_domains.
2537 /* Don't have all balancing operations going off at once */
2538 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2540 static void rebalance_tick(int this_cpu, runqueue_t *this_rq,
2541 enum idle_type idle)
2543 unsigned long old_load, this_load;
2544 unsigned long j = jiffies + CPU_OFFSET(this_cpu);
2545 struct sched_domain *sd;
2548 this_load = this_rq->raw_weighted_load;
2549 /* Update our load */
2550 for (i = 0; i < 3; i++) {
2551 unsigned long new_load = this_load;
2553 old_load = this_rq->cpu_load[i];
2555 * Round up the averaging division if load is increasing. This
2556 * prevents us from getting stuck on 9 if the load is 10, for
2559 if (new_load > old_load)
2560 new_load += scale-1;
2561 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale;
2564 for_each_domain(this_cpu, sd) {
2565 unsigned long interval;
2567 if (!(sd->flags & SD_LOAD_BALANCE))
2570 interval = sd->balance_interval;
2571 if (idle != SCHED_IDLE)
2572 interval *= sd->busy_factor;
2574 /* scale ms to jiffies */
2575 interval = msecs_to_jiffies(interval);
2576 if (unlikely(!interval))
2579 if (j - sd->last_balance >= interval) {
2580 if (load_balance(this_cpu, this_rq, sd, idle)) {
2582 * We've pulled tasks over so either we're no
2583 * longer idle, or one of our SMT siblings is
2588 sd->last_balance += interval;
2594 * on UP we do not need to balance between CPUs:
2596 static inline void rebalance_tick(int cpu, runqueue_t *rq, enum idle_type idle)
2599 static inline void idle_balance(int cpu, runqueue_t *rq)
2604 static inline int wake_priority_sleeper(runqueue_t *rq)
2607 #ifdef CONFIG_SCHED_SMT
2608 spin_lock(&rq->lock);
2610 * If an SMT sibling task has been put to sleep for priority
2611 * reasons reschedule the idle task to see if it can now run.
2613 if (rq->nr_running) {
2614 resched_task(rq->idle);
2617 spin_unlock(&rq->lock);
2622 DEFINE_PER_CPU(struct kernel_stat, kstat);
2624 EXPORT_PER_CPU_SYMBOL(kstat);
2627 * This is called on clock ticks and on context switches.
2628 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2630 static inline void update_cpu_clock(task_t *p, runqueue_t *rq,
2631 unsigned long long now)
2633 unsigned long long last = max(p->timestamp, rq->timestamp_last_tick);
2634 p->sched_time += now - last;
2638 * Return current->sched_time plus any more ns on the sched_clock
2639 * that have not yet been banked.
2641 unsigned long long current_sched_time(const task_t *tsk)
2643 unsigned long long ns;
2644 unsigned long flags;
2645 local_irq_save(flags);
2646 ns = max(tsk->timestamp, task_rq(tsk)->timestamp_last_tick);
2647 ns = tsk->sched_time + (sched_clock() - ns);
2648 local_irq_restore(flags);
2653 * We place interactive tasks back into the active array, if possible.
2655 * To guarantee that this does not starve expired tasks we ignore the
2656 * interactivity of a task if the first expired task had to wait more
2657 * than a 'reasonable' amount of time. This deadline timeout is
2658 * load-dependent, as the frequency of array switched decreases with
2659 * increasing number of running tasks. We also ignore the interactivity
2660 * if a better static_prio task has expired:
2662 #define EXPIRED_STARVING(rq) \
2663 ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2664 (jiffies - (rq)->expired_timestamp >= \
2665 STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2666 ((rq)->curr->static_prio > (rq)->best_expired_prio))
2669 * Account user cpu time to a process.
2670 * @p: the process that the cpu time gets accounted to
2671 * @hardirq_offset: the offset to subtract from hardirq_count()
2672 * @cputime: the cpu time spent in user space since the last update
2674 void account_user_time(struct task_struct *p, cputime_t cputime)
2676 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2679 p->utime = cputime_add(p->utime, cputime);
2681 /* Add user time to cpustat. */
2682 tmp = cputime_to_cputime64(cputime);
2683 if (TASK_NICE(p) > 0)
2684 cpustat->nice = cputime64_add(cpustat->nice, tmp);
2686 cpustat->user = cputime64_add(cpustat->user, tmp);
2690 * Account system cpu time to a process.
2691 * @p: the process that the cpu time gets accounted to
2692 * @hardirq_offset: the offset to subtract from hardirq_count()
2693 * @cputime: the cpu time spent in kernel space since the last update
2695 void account_system_time(struct task_struct *p, int hardirq_offset,
2698 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2699 runqueue_t *rq = this_rq();
2702 p->stime = cputime_add(p->stime, cputime);
2704 /* Add system time to cpustat. */
2705 tmp = cputime_to_cputime64(cputime);
2706 if (hardirq_count() - hardirq_offset)
2707 cpustat->irq = cputime64_add(cpustat->irq, tmp);
2708 else if (softirq_count())
2709 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
2710 else if (p != rq->idle)
2711 cpustat->system = cputime64_add(cpustat->system, tmp);
2712 else if (atomic_read(&rq->nr_iowait) > 0)
2713 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2715 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2716 /* Account for system time used */
2717 acct_update_integrals(p);
2721 * Account for involuntary wait time.
2722 * @p: the process from which the cpu time has been stolen
2723 * @steal: the cpu time spent in involuntary wait
2725 void account_steal_time(struct task_struct *p, cputime_t steal)
2727 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2728 cputime64_t tmp = cputime_to_cputime64(steal);
2729 runqueue_t *rq = this_rq();
2731 if (p == rq->idle) {
2732 p->stime = cputime_add(p->stime, steal);
2733 if (atomic_read(&rq->nr_iowait) > 0)
2734 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2736 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2738 cpustat->steal = cputime64_add(cpustat->steal, tmp);
2742 * This function gets called by the timer code, with HZ frequency.
2743 * We call it with interrupts disabled.
2745 * It also gets called by the fork code, when changing the parent's
2748 void scheduler_tick(void)
2750 int cpu = smp_processor_id();
2751 runqueue_t *rq = this_rq();
2752 task_t *p = current;
2753 unsigned long long now = sched_clock();
2755 update_cpu_clock(p, rq, now);
2757 rq->timestamp_last_tick = now;
2759 if (p == rq->idle) {
2760 if (wake_priority_sleeper(rq))
2762 rebalance_tick(cpu, rq, SCHED_IDLE);
2766 /* Task might have expired already, but not scheduled off yet */
2767 if (p->array != rq->active) {
2768 set_tsk_need_resched(p);
2771 spin_lock(&rq->lock);
2773 * The task was running during this tick - update the
2774 * time slice counter. Note: we do not update a thread's
2775 * priority until it either goes to sleep or uses up its
2776 * timeslice. This makes it possible for interactive tasks
2777 * to use up their timeslices at their highest priority levels.
2781 * RR tasks need a special form of timeslice management.
2782 * FIFO tasks have no timeslices.
2784 if ((p->policy == SCHED_RR) && !--p->time_slice) {
2785 p->time_slice = task_timeslice(p);
2786 p->first_time_slice = 0;
2787 set_tsk_need_resched(p);
2789 /* put it at the end of the queue: */
2790 requeue_task(p, rq->active);
2794 if (!--p->time_slice) {
2795 dequeue_task(p, rq->active);
2796 set_tsk_need_resched(p);
2797 p->prio = effective_prio(p);
2798 p->time_slice = task_timeslice(p);
2799 p->first_time_slice = 0;
2801 if (!rq->expired_timestamp)
2802 rq->expired_timestamp = jiffies;
2803 if (!TASK_INTERACTIVE(p) || EXPIRED_STARVING(rq)) {
2804 enqueue_task(p, rq->expired);
2805 if (p->static_prio < rq->best_expired_prio)
2806 rq->best_expired_prio = p->static_prio;
2808 enqueue_task(p, rq->active);
2811 * Prevent a too long timeslice allowing a task to monopolize
2812 * the CPU. We do this by splitting up the timeslice into
2815 * Note: this does not mean the task's timeslices expire or
2816 * get lost in any way, they just might be preempted by
2817 * another task of equal priority. (one with higher
2818 * priority would have preempted this task already.) We
2819 * requeue this task to the end of the list on this priority
2820 * level, which is in essence a round-robin of tasks with
2823 * This only applies to tasks in the interactive
2824 * delta range with at least TIMESLICE_GRANULARITY to requeue.
2826 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
2827 p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
2828 (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
2829 (p->array == rq->active)) {
2831 requeue_task(p, rq->active);
2832 set_tsk_need_resched(p);
2836 spin_unlock(&rq->lock);
2838 rebalance_tick(cpu, rq, NOT_IDLE);
2841 #ifdef CONFIG_SCHED_SMT
2842 static inline void wakeup_busy_runqueue(runqueue_t *rq)
2844 /* If an SMT runqueue is sleeping due to priority reasons wake it up */
2845 if (rq->curr == rq->idle && rq->nr_running)
2846 resched_task(rq->idle);
2850 * Called with interrupt disabled and this_rq's runqueue locked.
2852 static void wake_sleeping_dependent(int this_cpu)
2854 struct sched_domain *tmp, *sd = NULL;
2857 for_each_domain(this_cpu, tmp) {
2858 if (tmp->flags & SD_SHARE_CPUPOWER) {
2867 for_each_cpu_mask(i, sd->span) {
2868 runqueue_t *smt_rq = cpu_rq(i);
2872 if (unlikely(!spin_trylock(&smt_rq->lock)))
2875 wakeup_busy_runqueue(smt_rq);
2876 spin_unlock(&smt_rq->lock);
2881 * number of 'lost' timeslices this task wont be able to fully
2882 * utilize, if another task runs on a sibling. This models the
2883 * slowdown effect of other tasks running on siblings:
2885 static inline unsigned long smt_slice(task_t *p, struct sched_domain *sd)
2887 return p->time_slice * (100 - sd->per_cpu_gain) / 100;
2891 * To minimise lock contention and not have to drop this_rq's runlock we only
2892 * trylock the sibling runqueues and bypass those runqueues if we fail to
2893 * acquire their lock. As we only trylock the normal locking order does not
2894 * need to be obeyed.
2896 static int dependent_sleeper(int this_cpu, runqueue_t *this_rq, task_t *p)
2898 struct sched_domain *tmp, *sd = NULL;
2901 /* kernel/rt threads do not participate in dependent sleeping */
2902 if (!p->mm || rt_task(p))
2905 for_each_domain(this_cpu, tmp) {
2906 if (tmp->flags & SD_SHARE_CPUPOWER) {
2915 for_each_cpu_mask(i, sd->span) {
2923 if (unlikely(!spin_trylock(&smt_rq->lock)))
2926 smt_curr = smt_rq->curr;
2932 * If a user task with lower static priority than the
2933 * running task on the SMT sibling is trying to schedule,
2934 * delay it till there is proportionately less timeslice
2935 * left of the sibling task to prevent a lower priority
2936 * task from using an unfair proportion of the
2937 * physical cpu's resources. -ck
2939 if (rt_task(smt_curr)) {
2941 * With real time tasks we run non-rt tasks only
2942 * per_cpu_gain% of the time.
2944 if ((jiffies % DEF_TIMESLICE) >
2945 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
2948 if (smt_curr->static_prio < p->static_prio &&
2949 !TASK_PREEMPTS_CURR(p, smt_rq) &&
2950 smt_slice(smt_curr, sd) > task_timeslice(p))
2954 spin_unlock(&smt_rq->lock);
2959 static inline void wake_sleeping_dependent(int this_cpu)
2963 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq,
2970 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
2972 void fastcall add_preempt_count(int val)
2977 BUG_ON((preempt_count() < 0));
2978 preempt_count() += val;
2980 * Spinlock count overflowing soon?
2982 BUG_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10);
2984 EXPORT_SYMBOL(add_preempt_count);
2986 void fastcall sub_preempt_count(int val)
2991 BUG_ON(val > preempt_count());
2993 * Is the spinlock portion underflowing?
2995 BUG_ON((val < PREEMPT_MASK) && !(preempt_count() & PREEMPT_MASK));
2996 preempt_count() -= val;
2998 EXPORT_SYMBOL(sub_preempt_count);
3002 static inline int interactive_sleep(enum sleep_type sleep_type)
3004 return (sleep_type == SLEEP_INTERACTIVE ||
3005 sleep_type == SLEEP_INTERRUPTED);
3009 * schedule() is the main scheduler function.
3011 asmlinkage void __sched schedule(void)
3014 task_t *prev, *next;
3016 prio_array_t *array;
3017 struct list_head *queue;
3018 unsigned long long now;
3019 unsigned long run_time;
3020 int cpu, idx, new_prio;
3023 * Test if we are atomic. Since do_exit() needs to call into
3024 * schedule() atomically, we ignore that path for now.
3025 * Otherwise, whine if we are scheduling when we should not be.
3027 if (unlikely(in_atomic() && !current->exit_state)) {
3028 printk(KERN_ERR "BUG: scheduling while atomic: "
3030 current->comm, preempt_count(), current->pid);
3033 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3038 release_kernel_lock(prev);
3039 need_resched_nonpreemptible:
3043 * The idle thread is not allowed to schedule!
3044 * Remove this check after it has been exercised a bit.
3046 if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
3047 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
3051 schedstat_inc(rq, sched_cnt);
3052 now = sched_clock();
3053 if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
3054 run_time = now - prev->timestamp;
3055 if (unlikely((long long)(now - prev->timestamp) < 0))
3058 run_time = NS_MAX_SLEEP_AVG;
3061 * Tasks charged proportionately less run_time at high sleep_avg to
3062 * delay them losing their interactive status
3064 run_time /= (CURRENT_BONUS(prev) ? : 1);
3066 spin_lock_irq(&rq->lock);
3068 if (unlikely(prev->flags & PF_DEAD))
3069 prev->state = EXIT_DEAD;
3071 switch_count = &prev->nivcsw;
3072 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3073 switch_count = &prev->nvcsw;
3074 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3075 unlikely(signal_pending(prev))))
3076 prev->state = TASK_RUNNING;
3078 if (prev->state == TASK_UNINTERRUPTIBLE)
3079 rq->nr_uninterruptible++;
3080 deactivate_task(prev, rq);
3084 cpu = smp_processor_id();
3085 if (unlikely(!rq->nr_running)) {
3086 idle_balance(cpu, rq);
3087 if (!rq->nr_running) {
3089 rq->expired_timestamp = 0;
3090 wake_sleeping_dependent(cpu);
3096 if (unlikely(!array->nr_active)) {
3098 * Switch the active and expired arrays.
3100 schedstat_inc(rq, sched_switch);
3101 rq->active = rq->expired;
3102 rq->expired = array;
3104 rq->expired_timestamp = 0;
3105 rq->best_expired_prio = MAX_PRIO;
3108 idx = sched_find_first_bit(array->bitmap);
3109 queue = array->queue + idx;
3110 next = list_entry(queue->next, task_t, run_list);
3112 if (!rt_task(next) && interactive_sleep(next->sleep_type)) {
3113 unsigned long long delta = now - next->timestamp;
3114 if (unlikely((long long)(now - next->timestamp) < 0))
3117 if (next->sleep_type == SLEEP_INTERACTIVE)
3118 delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
3120 array = next->array;
3121 new_prio = recalc_task_prio(next, next->timestamp + delta);
3123 if (unlikely(next->prio != new_prio)) {
3124 dequeue_task(next, array);
3125 next->prio = new_prio;
3126 enqueue_task(next, array);
3129 next->sleep_type = SLEEP_NORMAL;
3130 if (dependent_sleeper(cpu, rq, next))
3133 if (next == rq->idle)
3134 schedstat_inc(rq, sched_goidle);
3136 prefetch_stack(next);
3137 clear_tsk_need_resched(prev);
3138 rcu_qsctr_inc(task_cpu(prev));
3140 update_cpu_clock(prev, rq, now);
3142 prev->sleep_avg -= run_time;
3143 if ((long)prev->sleep_avg <= 0)
3144 prev->sleep_avg = 0;
3145 prev->timestamp = prev->last_ran = now;
3147 sched_info_switch(prev, next);
3148 if (likely(prev != next)) {
3149 next->timestamp = now;
3154 prepare_task_switch(rq, next);
3155 prev = context_switch(rq, prev, next);
3158 * this_rq must be evaluated again because prev may have moved
3159 * CPUs since it called schedule(), thus the 'rq' on its stack
3160 * frame will be invalid.
3162 finish_task_switch(this_rq(), prev);
3164 spin_unlock_irq(&rq->lock);
3167 if (unlikely(reacquire_kernel_lock(prev) < 0))
3168 goto need_resched_nonpreemptible;
3169 preempt_enable_no_resched();
3170 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3174 EXPORT_SYMBOL(schedule);
3176 #ifdef CONFIG_PREEMPT
3178 * this is is the entry point to schedule() from in-kernel preemption
3179 * off of preempt_enable. Kernel preemptions off return from interrupt
3180 * occur there and call schedule directly.
3182 asmlinkage void __sched preempt_schedule(void)
3184 struct thread_info *ti = current_thread_info();
3185 #ifdef CONFIG_PREEMPT_BKL
3186 struct task_struct *task = current;
3187 int saved_lock_depth;
3190 * If there is a non-zero preempt_count or interrupts are disabled,
3191 * we do not want to preempt the current task. Just return..
3193 if (unlikely(ti->preempt_count || irqs_disabled()))
3197 add_preempt_count(PREEMPT_ACTIVE);
3199 * We keep the big kernel semaphore locked, but we
3200 * clear ->lock_depth so that schedule() doesnt
3201 * auto-release the semaphore:
3203 #ifdef CONFIG_PREEMPT_BKL
3204 saved_lock_depth = task->lock_depth;
3205 task->lock_depth = -1;
3208 #ifdef CONFIG_PREEMPT_BKL
3209 task->lock_depth = saved_lock_depth;
3211 sub_preempt_count(PREEMPT_ACTIVE);
3213 /* we could miss a preemption opportunity between schedule and now */
3215 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3219 EXPORT_SYMBOL(preempt_schedule);
3222 * this is is the entry point to schedule() from kernel preemption
3223 * off of irq context.
3224 * Note, that this is called and return with irqs disabled. This will
3225 * protect us against recursive calling from irq.
3227 asmlinkage void __sched preempt_schedule_irq(void)
3229 struct thread_info *ti = current_thread_info();
3230 #ifdef CONFIG_PREEMPT_BKL
3231 struct task_struct *task = current;
3232 int saved_lock_depth;
3234 /* Catch callers which need to be fixed*/
3235 BUG_ON(ti->preempt_count || !irqs_disabled());
3238 add_preempt_count(PREEMPT_ACTIVE);
3240 * We keep the big kernel semaphore locked, but we
3241 * clear ->lock_depth so that schedule() doesnt
3242 * auto-release the semaphore:
3244 #ifdef CONFIG_PREEMPT_BKL
3245 saved_lock_depth = task->lock_depth;
3246 task->lock_depth = -1;
3250 local_irq_disable();
3251 #ifdef CONFIG_PREEMPT_BKL
3252 task->lock_depth = saved_lock_depth;
3254 sub_preempt_count(PREEMPT_ACTIVE);
3256 /* we could miss a preemption opportunity between schedule and now */
3258 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3262 #endif /* CONFIG_PREEMPT */
3264 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3267 task_t *p = curr->private;
3268 return try_to_wake_up(p, mode, sync);
3271 EXPORT_SYMBOL(default_wake_function);
3274 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3275 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3276 * number) then we wake all the non-exclusive tasks and one exclusive task.
3278 * There are circumstances in which we can try to wake a task which has already
3279 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3280 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3282 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3283 int nr_exclusive, int sync, void *key)
3285 struct list_head *tmp, *next;
3287 list_for_each_safe(tmp, next, &q->task_list) {
3290 curr = list_entry(tmp, wait_queue_t, task_list);
3291 flags = curr->flags;
3292 if (curr->func(curr, mode, sync, key) &&
3293 (flags & WQ_FLAG_EXCLUSIVE) &&
3300 * __wake_up - wake up threads blocked on a waitqueue.
3302 * @mode: which threads
3303 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3304 * @key: is directly passed to the wakeup function
3306 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3307 int nr_exclusive, void *key)
3309 unsigned long flags;
3311 spin_lock_irqsave(&q->lock, flags);
3312 __wake_up_common(q, mode, nr_exclusive, 0, key);
3313 spin_unlock_irqrestore(&q->lock, flags);
3316 EXPORT_SYMBOL(__wake_up);
3319 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3321 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3323 __wake_up_common(q, mode, 1, 0, NULL);
3327 * __wake_up_sync - wake up threads blocked on a waitqueue.
3329 * @mode: which threads
3330 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3332 * The sync wakeup differs that the waker knows that it will schedule
3333 * away soon, so while the target thread will be woken up, it will not
3334 * be migrated to another CPU - ie. the two threads are 'synchronized'
3335 * with each other. This can prevent needless bouncing between CPUs.
3337 * On UP it can prevent extra preemption.
3340 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3342 unsigned long flags;
3348 if (unlikely(!nr_exclusive))
3351 spin_lock_irqsave(&q->lock, flags);
3352 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3353 spin_unlock_irqrestore(&q->lock, flags);
3355 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3357 void fastcall complete(struct completion *x)
3359 unsigned long flags;
3361 spin_lock_irqsave(&x->wait.lock, flags);
3363 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3365 spin_unlock_irqrestore(&x->wait.lock, flags);
3367 EXPORT_SYMBOL(complete);
3369 void fastcall complete_all(struct completion *x)
3371 unsigned long flags;
3373 spin_lock_irqsave(&x->wait.lock, flags);
3374 x->done += UINT_MAX/2;
3375 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3377 spin_unlock_irqrestore(&x->wait.lock, flags);
3379 EXPORT_SYMBOL(complete_all);
3381 void fastcall __sched wait_for_completion(struct completion *x)
3384 spin_lock_irq(&x->wait.lock);
3386 DECLARE_WAITQUEUE(wait, current);
3388 wait.flags |= WQ_FLAG_EXCLUSIVE;
3389 __add_wait_queue_tail(&x->wait, &wait);
3391 __set_current_state(TASK_UNINTERRUPTIBLE);
3392 spin_unlock_irq(&x->wait.lock);
3394 spin_lock_irq(&x->wait.lock);
3396 __remove_wait_queue(&x->wait, &wait);
3399 spin_unlock_irq(&x->wait.lock);
3401 EXPORT_SYMBOL(wait_for_completion);
3403 unsigned long fastcall __sched
3404 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3408 spin_lock_irq(&x->wait.lock);
3410 DECLARE_WAITQUEUE(wait, current);
3412 wait.flags |= WQ_FLAG_EXCLUSIVE;
3413 __add_wait_queue_tail(&x->wait, &wait);
3415 __set_current_state(TASK_UNINTERRUPTIBLE);
3416 spin_unlock_irq(&x->wait.lock);
3417 timeout = schedule_timeout(timeout);
3418 spin_lock_irq(&x->wait.lock);
3420 __remove_wait_queue(&x->wait, &wait);
3424 __remove_wait_queue(&x->wait, &wait);
3428 spin_unlock_irq(&x->wait.lock);
3431 EXPORT_SYMBOL(wait_for_completion_timeout);
3433 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3439 spin_lock_irq(&x->wait.lock);
3441 DECLARE_WAITQUEUE(wait, current);
3443 wait.flags |= WQ_FLAG_EXCLUSIVE;
3444 __add_wait_queue_tail(&x->wait, &wait);
3446 if (signal_pending(current)) {
3448 __remove_wait_queue(&x->wait, &wait);
3451 __set_current_state(TASK_INTERRUPTIBLE);
3452 spin_unlock_irq(&x->wait.lock);
3454 spin_lock_irq(&x->wait.lock);
3456 __remove_wait_queue(&x->wait, &wait);
3460 spin_unlock_irq(&x->wait.lock);
3464 EXPORT_SYMBOL(wait_for_completion_interruptible);
3466 unsigned long fastcall __sched
3467 wait_for_completion_interruptible_timeout(struct completion *x,
3468 unsigned long timeout)
3472 spin_lock_irq(&x->wait.lock);
3474 DECLARE_WAITQUEUE(wait, current);
3476 wait.flags |= WQ_FLAG_EXCLUSIVE;
3477 __add_wait_queue_tail(&x->wait, &wait);
3479 if (signal_pending(current)) {
3480 timeout = -ERESTARTSYS;
3481 __remove_wait_queue(&x->wait, &wait);
3484 __set_current_state(TASK_INTERRUPTIBLE);
3485 spin_unlock_irq(&x->wait.lock);
3486 timeout = schedule_timeout(timeout);
3487 spin_lock_irq(&x->wait.lock);
3489 __remove_wait_queue(&x->wait, &wait);
3493 __remove_wait_queue(&x->wait, &wait);
3497 spin_unlock_irq(&x->wait.lock);
3500 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3503 #define SLEEP_ON_VAR \
3504 unsigned long flags; \
3505 wait_queue_t wait; \
3506 init_waitqueue_entry(&wait, current);
3508 #define SLEEP_ON_HEAD \
3509 spin_lock_irqsave(&q->lock,flags); \
3510 __add_wait_queue(q, &wait); \
3511 spin_unlock(&q->lock);
3513 #define SLEEP_ON_TAIL \
3514 spin_lock_irq(&q->lock); \
3515 __remove_wait_queue(q, &wait); \
3516 spin_unlock_irqrestore(&q->lock, flags);
3518 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
3522 current->state = TASK_INTERRUPTIBLE;
3529 EXPORT_SYMBOL(interruptible_sleep_on);
3531 long fastcall __sched
3532 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3536 current->state = TASK_INTERRUPTIBLE;
3539 timeout = schedule_timeout(timeout);
3545 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3547 void fastcall __sched sleep_on(wait_queue_head_t *q)
3551 current->state = TASK_UNINTERRUPTIBLE;
3558 EXPORT_SYMBOL(sleep_on);
3560 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3564 current->state = TASK_UNINTERRUPTIBLE;
3567 timeout = schedule_timeout(timeout);
3573 EXPORT_SYMBOL(sleep_on_timeout);
3575 void set_user_nice(task_t *p, long nice)
3577 unsigned long flags;
3578 prio_array_t *array;
3580 int old_prio, new_prio, delta;
3582 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3585 * We have to be careful, if called from sys_setpriority(),
3586 * the task might be in the middle of scheduling on another CPU.
3588 rq = task_rq_lock(p, &flags);
3590 * The RT priorities are set via sched_setscheduler(), but we still
3591 * allow the 'normal' nice value to be set - but as expected
3592 * it wont have any effect on scheduling until the task is
3593 * not SCHED_NORMAL/SCHED_BATCH:
3596 p->static_prio = NICE_TO_PRIO(nice);
3601 dequeue_task(p, array);
3602 dec_raw_weighted_load(rq, p);
3606 new_prio = NICE_TO_PRIO(nice);
3607 delta = new_prio - old_prio;
3608 p->static_prio = NICE_TO_PRIO(nice);
3613 enqueue_task(p, array);
3614 inc_raw_weighted_load(rq, p);
3616 * If the task increased its priority or is running and
3617 * lowered its priority, then reschedule its CPU:
3619 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3620 resched_task(rq->curr);
3623 task_rq_unlock(rq, &flags);
3626 EXPORT_SYMBOL(set_user_nice);
3629 * can_nice - check if a task can reduce its nice value
3633 int can_nice(const task_t *p, const int nice)
3635 /* convert nice value [19,-20] to rlimit style value [1,40] */
3636 int nice_rlim = 20 - nice;
3637 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
3638 capable(CAP_SYS_NICE));
3641 #ifdef __ARCH_WANT_SYS_NICE
3644 * sys_nice - change the priority of the current process.
3645 * @increment: priority increment
3647 * sys_setpriority is a more generic, but much slower function that
3648 * does similar things.
3650 asmlinkage long sys_nice(int increment)
3656 * Setpriority might change our priority at the same moment.
3657 * We don't have to worry. Conceptually one call occurs first
3658 * and we have a single winner.
3660 if (increment < -40)
3665 nice = PRIO_TO_NICE(current->static_prio) + increment;
3671 if (increment < 0 && !can_nice(current, nice))
3674 retval = security_task_setnice(current, nice);
3678 set_user_nice(current, nice);
3685 * task_prio - return the priority value of a given task.
3686 * @p: the task in question.
3688 * This is the priority value as seen by users in /proc.
3689 * RT tasks are offset by -200. Normal tasks are centered
3690 * around 0, value goes from -16 to +15.
3692 int task_prio(const task_t *p)
3694 return p->prio - MAX_RT_PRIO;
3698 * task_nice - return the nice value of a given task.
3699 * @p: the task in question.
3701 int task_nice(const task_t *p)
3703 return TASK_NICE(p);
3705 EXPORT_SYMBOL_GPL(task_nice);
3708 * idle_cpu - is a given cpu idle currently?
3709 * @cpu: the processor in question.
3711 int idle_cpu(int cpu)
3713 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
3717 * idle_task - return the idle task for a given cpu.
3718 * @cpu: the processor in question.
3720 task_t *idle_task(int cpu)
3722 return cpu_rq(cpu)->idle;
3726 * find_process_by_pid - find a process with a matching PID value.
3727 * @pid: the pid in question.
3729 static inline task_t *find_process_by_pid(pid_t pid)
3731 return pid ? find_task_by_pid(pid) : current;
3734 /* Actually do priority change: must hold rq lock. */
3735 static void __setscheduler(struct task_struct *p, int policy, int prio)
3739 p->rt_priority = prio;
3740 if (policy != SCHED_NORMAL && policy != SCHED_BATCH) {
3741 p->prio = MAX_RT_PRIO-1 - p->rt_priority;
3743 p->prio = p->static_prio;
3745 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
3747 if (policy == SCHED_BATCH)
3754 * sched_setscheduler - change the scheduling policy and/or RT priority of
3756 * @p: the task in question.
3757 * @policy: new policy.
3758 * @param: structure containing the new RT priority.
3760 int sched_setscheduler(struct task_struct *p, int policy,
3761 struct sched_param *param)
3764 int oldprio, oldpolicy = -1;
3765 prio_array_t *array;
3766 unsigned long flags;
3770 /* double check policy once rq lock held */
3772 policy = oldpolicy = p->policy;
3773 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
3774 policy != SCHED_NORMAL && policy != SCHED_BATCH)
3777 * Valid priorities for SCHED_FIFO and SCHED_RR are
3778 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
3781 if (param->sched_priority < 0 ||
3782 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3783 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3785 if ((policy == SCHED_NORMAL || policy == SCHED_BATCH)
3786 != (param->sched_priority == 0))
3790 * Allow unprivileged RT tasks to decrease priority:
3792 if (!capable(CAP_SYS_NICE)) {
3794 * can't change policy, except between SCHED_NORMAL
3797 if (((policy != SCHED_NORMAL && p->policy != SCHED_BATCH) &&
3798 (policy != SCHED_BATCH && p->policy != SCHED_NORMAL)) &&
3799 !p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
3801 /* can't increase priority */
3802 if ((policy != SCHED_NORMAL && policy != SCHED_BATCH) &&
3803 param->sched_priority > p->rt_priority &&
3804 param->sched_priority >
3805 p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
3807 /* can't change other user's priorities */
3808 if ((current->euid != p->euid) &&
3809 (current->euid != p->uid))
3813 retval = security_task_setscheduler(p, policy, param);
3817 * To be able to change p->policy safely, the apropriate
3818 * runqueue lock must be held.
3820 rq = task_rq_lock(p, &flags);
3821 /* recheck policy now with rq lock held */
3822 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3823 policy = oldpolicy = -1;
3824 task_rq_unlock(rq, &flags);
3829 deactivate_task(p, rq);
3831 __setscheduler(p, policy, param->sched_priority);
3833 __activate_task(p, rq);
3835 * Reschedule if we are currently running on this runqueue and
3836 * our priority decreased, or if we are not currently running on
3837 * this runqueue and our priority is higher than the current's
3839 if (task_running(rq, p)) {
3840 if (p->prio > oldprio)
3841 resched_task(rq->curr);
3842 } else if (TASK_PREEMPTS_CURR(p, rq))
3843 resched_task(rq->curr);
3845 task_rq_unlock(rq, &flags);
3848 EXPORT_SYMBOL_GPL(sched_setscheduler);
3851 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3854 struct sched_param lparam;
3855 struct task_struct *p;
3857 if (!param || pid < 0)
3859 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3861 read_lock_irq(&tasklist_lock);
3862 p = find_process_by_pid(pid);
3864 read_unlock_irq(&tasklist_lock);
3867 retval = sched_setscheduler(p, policy, &lparam);
3868 read_unlock_irq(&tasklist_lock);
3873 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3874 * @pid: the pid in question.
3875 * @policy: new policy.
3876 * @param: structure containing the new RT priority.
3878 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
3879 struct sched_param __user *param)
3881 /* negative values for policy are not valid */
3885 return do_sched_setscheduler(pid, policy, param);
3889 * sys_sched_setparam - set/change the RT priority of a thread
3890 * @pid: the pid in question.
3891 * @param: structure containing the new RT priority.
3893 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
3895 return do_sched_setscheduler(pid, -1, param);
3899 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3900 * @pid: the pid in question.
3902 asmlinkage long sys_sched_getscheduler(pid_t pid)
3904 int retval = -EINVAL;
3911 read_lock(&tasklist_lock);
3912 p = find_process_by_pid(pid);
3914 retval = security_task_getscheduler(p);
3918 read_unlock(&tasklist_lock);
3925 * sys_sched_getscheduler - get the RT priority of a thread
3926 * @pid: the pid in question.
3927 * @param: structure containing the RT priority.
3929 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
3931 struct sched_param lp;
3932 int retval = -EINVAL;
3935 if (!param || pid < 0)
3938 read_lock(&tasklist_lock);
3939 p = find_process_by_pid(pid);
3944 retval = security_task_getscheduler(p);
3948 lp.sched_priority = p->rt_priority;
3949 read_unlock(&tasklist_lock);
3952 * This one might sleep, we cannot do it with a spinlock held ...
3954 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3960 read_unlock(&tasklist_lock);
3964 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
3968 cpumask_t cpus_allowed;
3971 read_lock(&tasklist_lock);
3973 p = find_process_by_pid(pid);
3975 read_unlock(&tasklist_lock);
3976 unlock_cpu_hotplug();
3981 * It is not safe to call set_cpus_allowed with the
3982 * tasklist_lock held. We will bump the task_struct's
3983 * usage count and then drop tasklist_lock.
3986 read_unlock(&tasklist_lock);
3989 if ((current->euid != p->euid) && (current->euid != p->uid) &&
3990 !capable(CAP_SYS_NICE))
3993 retval = security_task_setscheduler(p, 0, NULL);
3997 cpus_allowed = cpuset_cpus_allowed(p);
3998 cpus_and(new_mask, new_mask, cpus_allowed);
3999 retval = set_cpus_allowed(p, new_mask);
4003 unlock_cpu_hotplug();
4007 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4008 cpumask_t *new_mask)
4010 if (len < sizeof(cpumask_t)) {
4011 memset(new_mask, 0, sizeof(cpumask_t));
4012 } else if (len > sizeof(cpumask_t)) {
4013 len = sizeof(cpumask_t);
4015 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4019 * sys_sched_setaffinity - set the cpu affinity of a process
4020 * @pid: pid of the process
4021 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4022 * @user_mask_ptr: user-space pointer to the new cpu mask
4024 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4025 unsigned long __user *user_mask_ptr)
4030 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4034 return sched_setaffinity(pid, new_mask);
4038 * Represents all cpu's present in the system
4039 * In systems capable of hotplug, this map could dynamically grow
4040 * as new cpu's are detected in the system via any platform specific
4041 * method, such as ACPI for e.g.
4044 cpumask_t cpu_present_map __read_mostly;
4045 EXPORT_SYMBOL(cpu_present_map);
4048 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4049 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4052 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4058 read_lock(&tasklist_lock);
4061 p = find_process_by_pid(pid);
4065 retval = security_task_getscheduler(p);
4069 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4072 read_unlock(&tasklist_lock);
4073 unlock_cpu_hotplug();
4081 * sys_sched_getaffinity - get the cpu affinity of a process
4082 * @pid: pid of the process
4083 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4084 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4086 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4087 unsigned long __user *user_mask_ptr)
4092 if (len < sizeof(cpumask_t))
4095 ret = sched_getaffinity(pid, &mask);
4099 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4102 return sizeof(cpumask_t);
4106 * sys_sched_yield - yield the current processor to other threads.
4108 * this function yields the current CPU by moving the calling thread
4109 * to the expired array. If there are no other threads running on this
4110 * CPU then this function will return.
4112 asmlinkage long sys_sched_yield(void)
4114 runqueue_t *rq = this_rq_lock();
4115 prio_array_t *array = current->array;
4116 prio_array_t *target = rq->expired;
4118 schedstat_inc(rq, yld_cnt);
4120 * We implement yielding by moving the task into the expired
4123 * (special rule: RT tasks will just roundrobin in the active
4126 if (rt_task(current))
4127 target = rq->active;
4129 if (array->nr_active == 1) {
4130 schedstat_inc(rq, yld_act_empty);
4131 if (!rq->expired->nr_active)
4132 schedstat_inc(rq, yld_both_empty);
4133 } else if (!rq->expired->nr_active)
4134 schedstat_inc(rq, yld_exp_empty);
4136 if (array != target) {
4137 dequeue_task(current, array);
4138 enqueue_task(current, target);
4141 * requeue_task is cheaper so perform that if possible.
4143 requeue_task(current, array);
4146 * Since we are going to call schedule() anyway, there's
4147 * no need to preempt or enable interrupts:
4149 __release(rq->lock);
4150 _raw_spin_unlock(&rq->lock);
4151 preempt_enable_no_resched();
4158 static inline void __cond_resched(void)
4160 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4161 __might_sleep(__FILE__, __LINE__);
4164 * The BKS might be reacquired before we have dropped
4165 * PREEMPT_ACTIVE, which could trigger a second
4166 * cond_resched() call.
4168 if (unlikely(preempt_count()))
4170 if (unlikely(system_state != SYSTEM_RUNNING))
4173 add_preempt_count(PREEMPT_ACTIVE);
4175 sub_preempt_count(PREEMPT_ACTIVE);
4176 } while (need_resched());
4179 int __sched cond_resched(void)
4181 if (need_resched()) {
4188 EXPORT_SYMBOL(cond_resched);
4191 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4192 * call schedule, and on return reacquire the lock.
4194 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4195 * operations here to prevent schedule() from being called twice (once via
4196 * spin_unlock(), once by hand).
4198 int cond_resched_lock(spinlock_t *lock)
4202 if (need_lockbreak(lock)) {
4208 if (need_resched()) {
4209 _raw_spin_unlock(lock);
4210 preempt_enable_no_resched();
4218 EXPORT_SYMBOL(cond_resched_lock);
4220 int __sched cond_resched_softirq(void)
4222 BUG_ON(!in_softirq());
4224 if (need_resched()) {
4225 __local_bh_enable();
4233 EXPORT_SYMBOL(cond_resched_softirq);
4237 * yield - yield the current processor to other threads.
4239 * this is a shortcut for kernel-space yielding - it marks the
4240 * thread runnable and calls sys_sched_yield().
4242 void __sched yield(void)
4244 set_current_state(TASK_RUNNING);
4248 EXPORT_SYMBOL(yield);
4251 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4252 * that process accounting knows that this is a task in IO wait state.
4254 * But don't do that if it is a deliberate, throttling IO wait (this task
4255 * has set its backing_dev_info: the queue against which it should throttle)
4257 void __sched io_schedule(void)
4259 struct runqueue *rq = &__raw_get_cpu_var(runqueues);
4261 atomic_inc(&rq->nr_iowait);
4263 atomic_dec(&rq->nr_iowait);
4266 EXPORT_SYMBOL(io_schedule);
4268 long __sched io_schedule_timeout(long timeout)
4270 struct runqueue *rq = &__raw_get_cpu_var(runqueues);
4273 atomic_inc(&rq->nr_iowait);
4274 ret = schedule_timeout(timeout);
4275 atomic_dec(&rq->nr_iowait);
4280 * sys_sched_get_priority_max - return maximum RT priority.
4281 * @policy: scheduling class.
4283 * this syscall returns the maximum rt_priority that can be used
4284 * by a given scheduling class.
4286 asmlinkage long sys_sched_get_priority_max(int policy)
4293 ret = MAX_USER_RT_PRIO-1;
4304 * sys_sched_get_priority_min - return minimum RT priority.
4305 * @policy: scheduling class.
4307 * this syscall returns the minimum rt_priority that can be used
4308 * by a given scheduling class.
4310 asmlinkage long sys_sched_get_priority_min(int policy)
4327 * sys_sched_rr_get_interval - return the default timeslice of a process.
4328 * @pid: pid of the process.
4329 * @interval: userspace pointer to the timeslice value.
4331 * this syscall writes the default timeslice value of a given process
4332 * into the user-space timespec buffer. A value of '0' means infinity.
4335 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4337 int retval = -EINVAL;
4345 read_lock(&tasklist_lock);
4346 p = find_process_by_pid(pid);
4350 retval = security_task_getscheduler(p);
4354 jiffies_to_timespec(p->policy == SCHED_FIFO ?
4355 0 : task_timeslice(p), &t);
4356 read_unlock(&tasklist_lock);
4357 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4361 read_unlock(&tasklist_lock);
4365 static inline struct task_struct *eldest_child(struct task_struct *p)
4367 if (list_empty(&p->children)) return NULL;
4368 return list_entry(p->children.next,struct task_struct,sibling);
4371 static inline struct task_struct *older_sibling(struct task_struct *p)
4373 if (p->sibling.prev==&p->parent->children) return NULL;
4374 return list_entry(p->sibling.prev,struct task_struct,sibling);
4377 static inline struct task_struct *younger_sibling(struct task_struct *p)
4379 if (p->sibling.next==&p->parent->children) return NULL;
4380 return list_entry(p->sibling.next,struct task_struct,sibling);
4383 static void show_task(task_t *p)
4387 unsigned long free = 0;
4388 static const char *stat_nam[] = { "R", "S", "D", "T", "t", "Z", "X" };
4390 printk("%-13.13s ", p->comm);
4391 state = p->state ? __ffs(p->state) + 1 : 0;
4392 if (state < ARRAY_SIZE(stat_nam))
4393 printk(stat_nam[state]);
4396 #if (BITS_PER_LONG == 32)
4397 if (state == TASK_RUNNING)
4398 printk(" running ");
4400 printk(" %08lX ", thread_saved_pc(p));
4402 if (state == TASK_RUNNING)
4403 printk(" running task ");
4405 printk(" %016lx ", thread_saved_pc(p));
4407 #ifdef CONFIG_DEBUG_STACK_USAGE
4409 unsigned long *n = end_of_stack(p);
4412 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4415 printk("%5lu %5d %6d ", free, p->pid, p->parent->pid);
4416 if ((relative = eldest_child(p)))
4417 printk("%5d ", relative->pid);
4420 if ((relative = younger_sibling(p)))
4421 printk("%7d", relative->pid);
4424 if ((relative = older_sibling(p)))
4425 printk(" %5d", relative->pid);
4429 printk(" (L-TLB)\n");
4431 printk(" (NOTLB)\n");
4433 if (state != TASK_RUNNING)
4434 show_stack(p, NULL);
4437 void show_state(void)
4441 #if (BITS_PER_LONG == 32)
4444 printk(" task PC pid father child younger older\n");
4448 printk(" task PC pid father child younger older\n");
4450 read_lock(&tasklist_lock);
4451 do_each_thread(g, p) {
4453 * reset the NMI-timeout, listing all files on a slow
4454 * console might take alot of time:
4456 touch_nmi_watchdog();
4458 } while_each_thread(g, p);
4460 read_unlock(&tasklist_lock);
4461 mutex_debug_show_all_locks();
4465 * init_idle - set up an idle thread for a given CPU
4466 * @idle: task in question
4467 * @cpu: cpu the idle task belongs to
4469 * NOTE: this function does not set the idle thread's NEED_RESCHED
4470 * flag, to make booting more robust.
4472 void __devinit init_idle(task_t *idle, int cpu)
4474 runqueue_t *rq = cpu_rq(cpu);
4475 unsigned long flags;
4477 idle->timestamp = sched_clock();
4478 idle->sleep_avg = 0;
4480 idle->prio = MAX_PRIO;
4481 idle->state = TASK_RUNNING;
4482 idle->cpus_allowed = cpumask_of_cpu(cpu);
4483 set_task_cpu(idle, cpu);
4485 spin_lock_irqsave(&rq->lock, flags);
4486 rq->curr = rq->idle = idle;
4487 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4490 spin_unlock_irqrestore(&rq->lock, flags);
4492 /* Set the preempt count _outside_ the spinlocks! */
4493 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4494 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4496 task_thread_info(idle)->preempt_count = 0;
4501 * In a system that switches off the HZ timer nohz_cpu_mask
4502 * indicates which cpus entered this state. This is used
4503 * in the rcu update to wait only for active cpus. For system
4504 * which do not switch off the HZ timer nohz_cpu_mask should
4505 * always be CPU_MASK_NONE.
4507 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4511 * This is how migration works:
4513 * 1) we queue a migration_req_t structure in the source CPU's
4514 * runqueue and wake up that CPU's migration thread.
4515 * 2) we down() the locked semaphore => thread blocks.
4516 * 3) migration thread wakes up (implicitly it forces the migrated
4517 * thread off the CPU)
4518 * 4) it gets the migration request and checks whether the migrated
4519 * task is still in the wrong runqueue.
4520 * 5) if it's in the wrong runqueue then the migration thread removes
4521 * it and puts it into the right queue.
4522 * 6) migration thread up()s the semaphore.
4523 * 7) we wake up and the migration is done.
4527 * Change a given task's CPU affinity. Migrate the thread to a
4528 * proper CPU and schedule it away if the CPU it's executing on
4529 * is removed from the allowed bitmask.
4531 * NOTE: the caller must have a valid reference to the task, the
4532 * task must not exit() & deallocate itself prematurely. The
4533 * call is not atomic; no spinlocks may be held.
4535 int set_cpus_allowed(task_t *p, cpumask_t new_mask)
4537 unsigned long flags;
4539 migration_req_t req;
4542 rq = task_rq_lock(p, &flags);
4543 if (!cpus_intersects(new_mask, cpu_online_map)) {
4548 p->cpus_allowed = new_mask;
4549 /* Can the task run on the task's current CPU? If so, we're done */
4550 if (cpu_isset(task_cpu(p), new_mask))
4553 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4554 /* Need help from migration thread: drop lock and wait. */
4555 task_rq_unlock(rq, &flags);
4556 wake_up_process(rq->migration_thread);
4557 wait_for_completion(&req.done);
4558 tlb_migrate_finish(p->mm);
4562 task_rq_unlock(rq, &flags);
4566 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4569 * Move (not current) task off this cpu, onto dest cpu. We're doing
4570 * this because either it can't run here any more (set_cpus_allowed()
4571 * away from this CPU, or CPU going down), or because we're
4572 * attempting to rebalance this task on exec (sched_exec).
4574 * So we race with normal scheduler movements, but that's OK, as long
4575 * as the task is no longer on this CPU.
4577 * Returns non-zero if task was successfully migrated.
4579 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4581 runqueue_t *rq_dest, *rq_src;
4584 if (unlikely(cpu_is_offline(dest_cpu)))
4587 rq_src = cpu_rq(src_cpu);
4588 rq_dest = cpu_rq(dest_cpu);
4590 double_rq_lock(rq_src, rq_dest);
4591 /* Already moved. */
4592 if (task_cpu(p) != src_cpu)
4594 /* Affinity changed (again). */
4595 if (!cpu_isset(dest_cpu, p->cpus_allowed))
4598 set_task_cpu(p, dest_cpu);
4601 * Sync timestamp with rq_dest's before activating.
4602 * The same thing could be achieved by doing this step
4603 * afterwards, and pretending it was a local activate.
4604 * This way is cleaner and logically correct.
4606 p->timestamp = p->timestamp - rq_src->timestamp_last_tick
4607 + rq_dest->timestamp_last_tick;
4608 deactivate_task(p, rq_src);
4609 activate_task(p, rq_dest, 0);
4610 if (TASK_PREEMPTS_CURR(p, rq_dest))
4611 resched_task(rq_dest->curr);
4615 double_rq_unlock(rq_src, rq_dest);
4620 * migration_thread - this is a highprio system thread that performs
4621 * thread migration by bumping thread off CPU then 'pushing' onto
4624 static int migration_thread(void *data)
4627 int cpu = (long)data;
4630 BUG_ON(rq->migration_thread != current);
4632 set_current_state(TASK_INTERRUPTIBLE);
4633 while (!kthread_should_stop()) {
4634 struct list_head *head;
4635 migration_req_t *req;
4639 spin_lock_irq(&rq->lock);
4641 if (cpu_is_offline(cpu)) {
4642 spin_unlock_irq(&rq->lock);
4646 if (rq->active_balance) {
4647 active_load_balance(rq, cpu);
4648 rq->active_balance = 0;
4651 head = &rq->migration_queue;
4653 if (list_empty(head)) {
4654 spin_unlock_irq(&rq->lock);
4656 set_current_state(TASK_INTERRUPTIBLE);
4659 req = list_entry(head->next, migration_req_t, list);
4660 list_del_init(head->next);
4662 spin_unlock(&rq->lock);
4663 __migrate_task(req->task, cpu, req->dest_cpu);
4666 complete(&req->done);
4668 __set_current_state(TASK_RUNNING);
4672 /* Wait for kthread_stop */
4673 set_current_state(TASK_INTERRUPTIBLE);
4674 while (!kthread_should_stop()) {
4676 set_current_state(TASK_INTERRUPTIBLE);
4678 __set_current_state(TASK_RUNNING);
4682 #ifdef CONFIG_HOTPLUG_CPU
4683 /* Figure out where task on dead CPU should go, use force if neccessary. */
4684 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *tsk)
4687 unsigned long flags;
4693 mask = node_to_cpumask(cpu_to_node(dead_cpu));
4694 cpus_and(mask, mask, tsk->cpus_allowed);
4695 dest_cpu = any_online_cpu(mask);
4697 /* On any allowed CPU? */
4698 if (dest_cpu == NR_CPUS)
4699 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4701 /* No more Mr. Nice Guy. */
4702 if (dest_cpu == NR_CPUS) {
4703 rq = task_rq_lock(tsk, &flags);
4704 cpus_setall(tsk->cpus_allowed);
4705 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4706 task_rq_unlock(rq, &flags);
4709 * Don't tell them about moving exiting tasks or
4710 * kernel threads (both mm NULL), since they never
4713 if (tsk->mm && printk_ratelimit())
4714 printk(KERN_INFO "process %d (%s) no "
4715 "longer affine to cpu%d\n",
4716 tsk->pid, tsk->comm, dead_cpu);
4718 if (!__migrate_task(tsk, dead_cpu, dest_cpu))
4723 * While a dead CPU has no uninterruptible tasks queued at this point,
4724 * it might still have a nonzero ->nr_uninterruptible counter, because
4725 * for performance reasons the counter is not stricly tracking tasks to
4726 * their home CPUs. So we just add the counter to another CPU's counter,
4727 * to keep the global sum constant after CPU-down:
4729 static void migrate_nr_uninterruptible(runqueue_t *rq_src)
4731 runqueue_t *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
4732 unsigned long flags;
4734 local_irq_save(flags);
4735 double_rq_lock(rq_src, rq_dest);
4736 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
4737 rq_src->nr_uninterruptible = 0;
4738 double_rq_unlock(rq_src, rq_dest);
4739 local_irq_restore(flags);
4742 /* Run through task list and migrate tasks from the dead cpu. */
4743 static void migrate_live_tasks(int src_cpu)
4745 struct task_struct *tsk, *t;
4747 write_lock_irq(&tasklist_lock);
4749 do_each_thread(t, tsk) {
4753 if (task_cpu(tsk) == src_cpu)
4754 move_task_off_dead_cpu(src_cpu, tsk);
4755 } while_each_thread(t, tsk);
4757 write_unlock_irq(&tasklist_lock);
4760 /* Schedules idle task to be the next runnable task on current CPU.
4761 * It does so by boosting its priority to highest possible and adding it to
4762 * the _front_ of runqueue. Used by CPU offline code.
4764 void sched_idle_next(void)
4766 int cpu = smp_processor_id();
4767 runqueue_t *rq = this_rq();
4768 struct task_struct *p = rq->idle;
4769 unsigned long flags;
4771 /* cpu has to be offline */
4772 BUG_ON(cpu_online(cpu));
4774 /* Strictly not necessary since rest of the CPUs are stopped by now
4775 * and interrupts disabled on current cpu.
4777 spin_lock_irqsave(&rq->lock, flags);
4779 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4780 /* Add idle task to _front_ of it's priority queue */
4781 __activate_idle_task(p, rq);
4783 spin_unlock_irqrestore(&rq->lock, flags);
4786 /* Ensures that the idle task is using init_mm right before its cpu goes
4789 void idle_task_exit(void)
4791 struct mm_struct *mm = current->active_mm;
4793 BUG_ON(cpu_online(smp_processor_id()));
4796 switch_mm(mm, &init_mm, current);
4800 static void migrate_dead(unsigned int dead_cpu, task_t *tsk)
4802 struct runqueue *rq = cpu_rq(dead_cpu);
4804 /* Must be exiting, otherwise would be on tasklist. */
4805 BUG_ON(tsk->exit_state != EXIT_ZOMBIE && tsk->exit_state != EXIT_DEAD);
4807 /* Cannot have done final schedule yet: would have vanished. */
4808 BUG_ON(tsk->flags & PF_DEAD);
4810 get_task_struct(tsk);
4813 * Drop lock around migration; if someone else moves it,
4814 * that's OK. No task can be added to this CPU, so iteration is
4817 spin_unlock_irq(&rq->lock);
4818 move_task_off_dead_cpu(dead_cpu, tsk);
4819 spin_lock_irq(&rq->lock);
4821 put_task_struct(tsk);
4824 /* release_task() removes task from tasklist, so we won't find dead tasks. */
4825 static void migrate_dead_tasks(unsigned int dead_cpu)
4828 struct runqueue *rq = cpu_rq(dead_cpu);
4830 for (arr = 0; arr < 2; arr++) {
4831 for (i = 0; i < MAX_PRIO; i++) {
4832 struct list_head *list = &rq->arrays[arr].queue[i];
4833 while (!list_empty(list))
4834 migrate_dead(dead_cpu,
4835 list_entry(list->next, task_t,
4840 #endif /* CONFIG_HOTPLUG_CPU */
4843 * migration_call - callback that gets triggered when a CPU is added.
4844 * Here we can start up the necessary migration thread for the new CPU.
4846 static int __cpuinit migration_call(struct notifier_block *nfb,
4847 unsigned long action,
4850 int cpu = (long)hcpu;
4851 struct task_struct *p;
4852 struct runqueue *rq;
4853 unsigned long flags;
4856 case CPU_UP_PREPARE:
4857 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
4860 p->flags |= PF_NOFREEZE;
4861 kthread_bind(p, cpu);
4862 /* Must be high prio: stop_machine expects to yield to it. */
4863 rq = task_rq_lock(p, &flags);
4864 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4865 task_rq_unlock(rq, &flags);
4866 cpu_rq(cpu)->migration_thread = p;
4869 /* Strictly unneccessary, as first user will wake it. */
4870 wake_up_process(cpu_rq(cpu)->migration_thread);
4872 #ifdef CONFIG_HOTPLUG_CPU
4873 case CPU_UP_CANCELED:
4874 if (!cpu_rq(cpu)->migration_thread)
4876 /* Unbind it from offline cpu so it can run. Fall thru. */
4877 kthread_bind(cpu_rq(cpu)->migration_thread,
4878 any_online_cpu(cpu_online_map));
4879 kthread_stop(cpu_rq(cpu)->migration_thread);
4880 cpu_rq(cpu)->migration_thread = NULL;
4883 migrate_live_tasks(cpu);
4885 kthread_stop(rq->migration_thread);
4886 rq->migration_thread = NULL;
4887 /* Idle task back to normal (off runqueue, low prio) */
4888 rq = task_rq_lock(rq->idle, &flags);
4889 deactivate_task(rq->idle, rq);
4890 rq->idle->static_prio = MAX_PRIO;
4891 __setscheduler(rq->idle, SCHED_NORMAL, 0);
4892 migrate_dead_tasks(cpu);
4893 task_rq_unlock(rq, &flags);
4894 migrate_nr_uninterruptible(rq);
4895 BUG_ON(rq->nr_running != 0);
4897 /* No need to migrate the tasks: it was best-effort if
4898 * they didn't do lock_cpu_hotplug(). Just wake up
4899 * the requestors. */
4900 spin_lock_irq(&rq->lock);
4901 while (!list_empty(&rq->migration_queue)) {
4902 migration_req_t *req;
4903 req = list_entry(rq->migration_queue.next,
4904 migration_req_t, list);
4905 list_del_init(&req->list);
4906 complete(&req->done);
4908 spin_unlock_irq(&rq->lock);
4915 /* Register at highest priority so that task migration (migrate_all_tasks)
4916 * happens before everything else.
4918 static struct notifier_block __cpuinitdata migration_notifier = {
4919 .notifier_call = migration_call,
4923 int __init migration_init(void)
4925 void *cpu = (void *)(long)smp_processor_id();
4926 /* Start one for boot CPU. */
4927 migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
4928 migration_call(&migration_notifier, CPU_ONLINE, cpu);
4929 register_cpu_notifier(&migration_notifier);
4935 #undef SCHED_DOMAIN_DEBUG
4936 #ifdef SCHED_DOMAIN_DEBUG
4937 static void sched_domain_debug(struct sched_domain *sd, int cpu)
4942 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
4946 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
4951 struct sched_group *group = sd->groups;
4952 cpumask_t groupmask;
4954 cpumask_scnprintf(str, NR_CPUS, sd->span);
4955 cpus_clear(groupmask);
4958 for (i = 0; i < level + 1; i++)
4960 printk("domain %d: ", level);
4962 if (!(sd->flags & SD_LOAD_BALANCE)) {
4963 printk("does not load-balance\n");
4965 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent");
4969 printk("span %s\n", str);
4971 if (!cpu_isset(cpu, sd->span))
4972 printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
4973 if (!cpu_isset(cpu, group->cpumask))
4974 printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
4977 for (i = 0; i < level + 2; i++)
4983 printk(KERN_ERR "ERROR: group is NULL\n");
4987 if (!group->cpu_power) {
4989 printk(KERN_ERR "ERROR: domain->cpu_power not set\n");
4992 if (!cpus_weight(group->cpumask)) {
4994 printk(KERN_ERR "ERROR: empty group\n");
4997 if (cpus_intersects(groupmask, group->cpumask)) {
4999 printk(KERN_ERR "ERROR: repeated CPUs\n");
5002 cpus_or(groupmask, groupmask, group->cpumask);
5004 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5007 group = group->next;
5008 } while (group != sd->groups);
5011 if (!cpus_equal(sd->span, groupmask))
5012 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5018 if (!cpus_subset(groupmask, sd->span))
5019 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
5025 #define sched_domain_debug(sd, cpu) {}
5028 static int sd_degenerate(struct sched_domain *sd)
5030 if (cpus_weight(sd->span) == 1)
5033 /* Following flags need at least 2 groups */
5034 if (sd->flags & (SD_LOAD_BALANCE |
5035 SD_BALANCE_NEWIDLE |
5038 if (sd->groups != sd->groups->next)
5042 /* Following flags don't use groups */
5043 if (sd->flags & (SD_WAKE_IDLE |
5051 static int sd_parent_degenerate(struct sched_domain *sd,
5052 struct sched_domain *parent)
5054 unsigned long cflags = sd->flags, pflags = parent->flags;
5056 if (sd_degenerate(parent))
5059 if (!cpus_equal(sd->span, parent->span))
5062 /* Does parent contain flags not in child? */
5063 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5064 if (cflags & SD_WAKE_AFFINE)
5065 pflags &= ~SD_WAKE_BALANCE;
5066 /* Flags needing groups don't count if only 1 group in parent */
5067 if (parent->groups == parent->groups->next) {
5068 pflags &= ~(SD_LOAD_BALANCE |
5069 SD_BALANCE_NEWIDLE |
5073 if (~cflags & pflags)
5080 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5081 * hold the hotplug lock.
5083 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5085 runqueue_t *rq = cpu_rq(cpu);
5086 struct sched_domain *tmp;
5088 /* Remove the sched domains which do not contribute to scheduling. */
5089 for (tmp = sd; tmp; tmp = tmp->parent) {
5090 struct sched_domain *parent = tmp->parent;
5093 if (sd_parent_degenerate(tmp, parent))
5094 tmp->parent = parent->parent;
5097 if (sd && sd_degenerate(sd))
5100 sched_domain_debug(sd, cpu);
5102 rcu_assign_pointer(rq->sd, sd);
5105 /* cpus with isolated domains */
5106 static cpumask_t __devinitdata cpu_isolated_map = CPU_MASK_NONE;
5108 /* Setup the mask of cpus configured for isolated domains */
5109 static int __init isolated_cpu_setup(char *str)
5111 int ints[NR_CPUS], i;
5113 str = get_options(str, ARRAY_SIZE(ints), ints);
5114 cpus_clear(cpu_isolated_map);
5115 for (i = 1; i <= ints[0]; i++)
5116 if (ints[i] < NR_CPUS)
5117 cpu_set(ints[i], cpu_isolated_map);
5121 __setup ("isolcpus=", isolated_cpu_setup);
5124 * init_sched_build_groups takes an array of groups, the cpumask we wish
5125 * to span, and a pointer to a function which identifies what group a CPU
5126 * belongs to. The return value of group_fn must be a valid index into the
5127 * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
5128 * keep track of groups covered with a cpumask_t).
5130 * init_sched_build_groups will build a circular linked list of the groups
5131 * covered by the given span, and will set each group's ->cpumask correctly,
5132 * and ->cpu_power to 0.
5134 static void init_sched_build_groups(struct sched_group groups[], cpumask_t span,
5135 int (*group_fn)(int cpu))
5137 struct sched_group *first = NULL, *last = NULL;
5138 cpumask_t covered = CPU_MASK_NONE;
5141 for_each_cpu_mask(i, span) {
5142 int group = group_fn(i);
5143 struct sched_group *sg = &groups[group];
5146 if (cpu_isset(i, covered))
5149 sg->cpumask = CPU_MASK_NONE;
5152 for_each_cpu_mask(j, span) {
5153 if (group_fn(j) != group)
5156 cpu_set(j, covered);
5157 cpu_set(j, sg->cpumask);
5168 #define SD_NODES_PER_DOMAIN 16
5171 * Self-tuning task migration cost measurement between source and target CPUs.
5173 * This is done by measuring the cost of manipulating buffers of varying
5174 * sizes. For a given buffer-size here are the steps that are taken:
5176 * 1) the source CPU reads+dirties a shared buffer
5177 * 2) the target CPU reads+dirties the same shared buffer
5179 * We measure how long they take, in the following 4 scenarios:
5181 * - source: CPU1, target: CPU2 | cost1
5182 * - source: CPU2, target: CPU1 | cost2
5183 * - source: CPU1, target: CPU1 | cost3
5184 * - source: CPU2, target: CPU2 | cost4
5186 * We then calculate the cost3+cost4-cost1-cost2 difference - this is
5187 * the cost of migration.
5189 * We then start off from a small buffer-size and iterate up to larger
5190 * buffer sizes, in 5% steps - measuring each buffer-size separately, and
5191 * doing a maximum search for the cost. (The maximum cost for a migration
5192 * normally occurs when the working set size is around the effective cache
5195 #define SEARCH_SCOPE 2
5196 #define MIN_CACHE_SIZE (64*1024U)
5197 #define DEFAULT_CACHE_SIZE (5*1024*1024U)
5198 #define ITERATIONS 1
5199 #define SIZE_THRESH 130
5200 #define COST_THRESH 130
5203 * The migration cost is a function of 'domain distance'. Domain
5204 * distance is the number of steps a CPU has to iterate down its
5205 * domain tree to share a domain with the other CPU. The farther
5206 * two CPUs are from each other, the larger the distance gets.
5208 * Note that we use the distance only to cache measurement results,
5209 * the distance value is not used numerically otherwise. When two
5210 * CPUs have the same distance it is assumed that the migration
5211 * cost is the same. (this is a simplification but quite practical)
5213 #define MAX_DOMAIN_DISTANCE 32
5215 static unsigned long long migration_cost[MAX_DOMAIN_DISTANCE] =
5216 { [ 0 ... MAX_DOMAIN_DISTANCE-1 ] =
5218 * Architectures may override the migration cost and thus avoid
5219 * boot-time calibration. Unit is nanoseconds. Mostly useful for
5220 * virtualized hardware:
5222 #ifdef CONFIG_DEFAULT_MIGRATION_COST
5223 CONFIG_DEFAULT_MIGRATION_COST
5230 * Allow override of migration cost - in units of microseconds.
5231 * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost
5232 * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs:
5234 static int __init migration_cost_setup(char *str)
5236 int ints[MAX_DOMAIN_DISTANCE+1], i;
5238 str = get_options(str, ARRAY_SIZE(ints), ints);
5240 printk("#ints: %d\n", ints[0]);
5241 for (i = 1; i <= ints[0]; i++) {
5242 migration_cost[i-1] = (unsigned long long)ints[i]*1000;
5243 printk("migration_cost[%d]: %Ld\n", i-1, migration_cost[i-1]);
5248 __setup ("migration_cost=", migration_cost_setup);
5251 * Global multiplier (divisor) for migration-cutoff values,
5252 * in percentiles. E.g. use a value of 150 to get 1.5 times
5253 * longer cache-hot cutoff times.
5255 * (We scale it from 100 to 128 to long long handling easier.)
5258 #define MIGRATION_FACTOR_SCALE 128
5260 static unsigned int migration_factor = MIGRATION_FACTOR_SCALE;
5262 static int __init setup_migration_factor(char *str)
5264 get_option(&str, &migration_factor);
5265 migration_factor = migration_factor * MIGRATION_FACTOR_SCALE / 100;
5269 __setup("migration_factor=", setup_migration_factor);
5272 * Estimated distance of two CPUs, measured via the number of domains
5273 * we have to pass for the two CPUs to be in the same span:
5275 static unsigned long domain_distance(int cpu1, int cpu2)
5277 unsigned long distance = 0;
5278 struct sched_domain *sd;
5280 for_each_domain(cpu1, sd) {
5281 WARN_ON(!cpu_isset(cpu1, sd->span));
5282 if (cpu_isset(cpu2, sd->span))
5286 if (distance >= MAX_DOMAIN_DISTANCE) {
5288 distance = MAX_DOMAIN_DISTANCE-1;
5294 static unsigned int migration_debug;
5296 static int __init setup_migration_debug(char *str)
5298 get_option(&str, &migration_debug);
5302 __setup("migration_debug=", setup_migration_debug);
5305 * Maximum cache-size that the scheduler should try to measure.
5306 * Architectures with larger caches should tune this up during
5307 * bootup. Gets used in the domain-setup code (i.e. during SMP
5310 unsigned int max_cache_size;
5312 static int __init setup_max_cache_size(char *str)
5314 get_option(&str, &max_cache_size);
5318 __setup("max_cache_size=", setup_max_cache_size);
5321 * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This
5322 * is the operation that is timed, so we try to generate unpredictable
5323 * cachemisses that still end up filling the L2 cache:
5325 static void touch_cache(void *__cache, unsigned long __size)
5327 unsigned long size = __size/sizeof(long), chunk1 = size/3,
5329 unsigned long *cache = __cache;
5332 for (i = 0; i < size/6; i += 8) {
5335 case 1: cache[size-1-i]++;
5336 case 2: cache[chunk1-i]++;
5337 case 3: cache[chunk1+i]++;
5338 case 4: cache[chunk2-i]++;
5339 case 5: cache[chunk2+i]++;
5345 * Measure the cache-cost of one task migration. Returns in units of nsec.
5347 static unsigned long long measure_one(void *cache, unsigned long size,
5348 int source, int target)
5350 cpumask_t mask, saved_mask;
5351 unsigned long long t0, t1, t2, t3, cost;
5353 saved_mask = current->cpus_allowed;
5356 * Flush source caches to RAM and invalidate them:
5361 * Migrate to the source CPU:
5363 mask = cpumask_of_cpu(source);
5364 set_cpus_allowed(current, mask);
5365 WARN_ON(smp_processor_id() != source);
5368 * Dirty the working set:
5371 touch_cache(cache, size);
5375 * Migrate to the target CPU, dirty the L2 cache and access
5376 * the shared buffer. (which represents the working set
5377 * of a migrated task.)
5379 mask = cpumask_of_cpu(target);
5380 set_cpus_allowed(current, mask);
5381 WARN_ON(smp_processor_id() != target);
5384 touch_cache(cache, size);
5387 cost = t1-t0 + t3-t2;
5389 if (migration_debug >= 2)
5390 printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n",
5391 source, target, t1-t0, t1-t0, t3-t2, cost);
5393 * Flush target caches to RAM and invalidate them:
5397 set_cpus_allowed(current, saved_mask);
5403 * Measure a series of task migrations and return the average
5404 * result. Since this code runs early during bootup the system
5405 * is 'undisturbed' and the average latency makes sense.
5407 * The algorithm in essence auto-detects the relevant cache-size,
5408 * so it will properly detect different cachesizes for different
5409 * cache-hierarchies, depending on how the CPUs are connected.
5411 * Architectures can prime the upper limit of the search range via
5412 * max_cache_size, otherwise the search range defaults to 20MB...64K.
5414 static unsigned long long
5415 measure_cost(int cpu1, int cpu2, void *cache, unsigned int size)
5417 unsigned long long cost1, cost2;
5421 * Measure the migration cost of 'size' bytes, over an
5422 * average of 10 runs:
5424 * (We perturb the cache size by a small (0..4k)
5425 * value to compensate size/alignment related artifacts.
5426 * We also subtract the cost of the operation done on
5432 * dry run, to make sure we start off cache-cold on cpu1,
5433 * and to get any vmalloc pagefaults in advance:
5435 measure_one(cache, size, cpu1, cpu2);
5436 for (i = 0; i < ITERATIONS; i++)
5437 cost1 += measure_one(cache, size - i*1024, cpu1, cpu2);
5439 measure_one(cache, size, cpu2, cpu1);
5440 for (i = 0; i < ITERATIONS; i++)
5441 cost1 += measure_one(cache, size - i*1024, cpu2, cpu1);
5444 * (We measure the non-migrating [cached] cost on both
5445 * cpu1 and cpu2, to handle CPUs with different speeds)
5449 measure_one(cache, size, cpu1, cpu1);
5450 for (i = 0; i < ITERATIONS; i++)
5451 cost2 += measure_one(cache, size - i*1024, cpu1, cpu1);
5453 measure_one(cache, size, cpu2, cpu2);
5454 for (i = 0; i < ITERATIONS; i++)
5455 cost2 += measure_one(cache, size - i*1024, cpu2, cpu2);
5458 * Get the per-iteration migration cost:
5460 do_div(cost1, 2*ITERATIONS);
5461 do_div(cost2, 2*ITERATIONS);
5463 return cost1 - cost2;
5466 static unsigned long long measure_migration_cost(int cpu1, int cpu2)
5468 unsigned long long max_cost = 0, fluct = 0, avg_fluct = 0;
5469 unsigned int max_size, size, size_found = 0;
5470 long long cost = 0, prev_cost;
5474 * Search from max_cache_size*5 down to 64K - the real relevant
5475 * cachesize has to lie somewhere inbetween.
5477 if (max_cache_size) {
5478 max_size = max(max_cache_size * SEARCH_SCOPE, MIN_CACHE_SIZE);
5479 size = max(max_cache_size / SEARCH_SCOPE, MIN_CACHE_SIZE);
5482 * Since we have no estimation about the relevant
5485 max_size = DEFAULT_CACHE_SIZE * SEARCH_SCOPE;
5486 size = MIN_CACHE_SIZE;
5489 if (!cpu_online(cpu1) || !cpu_online(cpu2)) {
5490 printk("cpu %d and %d not both online!\n", cpu1, cpu2);
5495 * Allocate the working set:
5497 cache = vmalloc(max_size);
5499 printk("could not vmalloc %d bytes for cache!\n", 2*max_size);
5500 return 1000000; // return 1 msec on very small boxen
5503 while (size <= max_size) {
5505 cost = measure_cost(cpu1, cpu2, cache, size);
5511 if (max_cost < cost) {
5517 * Calculate average fluctuation, we use this to prevent
5518 * noise from triggering an early break out of the loop:
5520 fluct = abs(cost - prev_cost);
5521 avg_fluct = (avg_fluct + fluct)/2;
5523 if (migration_debug)
5524 printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): (%8Ld %8Ld)\n",
5526 (long)cost / 1000000,
5527 ((long)cost / 100000) % 10,
5528 (long)max_cost / 1000000,
5529 ((long)max_cost / 100000) % 10,
5530 domain_distance(cpu1, cpu2),
5534 * If we iterated at least 20% past the previous maximum,
5535 * and the cost has dropped by more than 20% already,
5536 * (taking fluctuations into account) then we assume to
5537 * have found the maximum and break out of the loop early:
5539 if (size_found && (size*100 > size_found*SIZE_THRESH))
5540 if (cost+avg_fluct <= 0 ||
5541 max_cost*100 > (cost+avg_fluct)*COST_THRESH) {
5543 if (migration_debug)
5544 printk("-> found max.\n");
5548 * Increase the cachesize in 10% steps:
5550 size = size * 10 / 9;
5553 if (migration_debug)
5554 printk("[%d][%d] working set size found: %d, cost: %Ld\n",
5555 cpu1, cpu2, size_found, max_cost);
5560 * A task is considered 'cache cold' if at least 2 times
5561 * the worst-case cost of migration has passed.
5563 * (this limit is only listened to if the load-balancing
5564 * situation is 'nice' - if there is a large imbalance we
5565 * ignore it for the sake of CPU utilization and
5566 * processing fairness.)
5568 return 2 * max_cost * migration_factor / MIGRATION_FACTOR_SCALE;
5571 static void calibrate_migration_costs(const cpumask_t *cpu_map)
5573 int cpu1 = -1, cpu2 = -1, cpu, orig_cpu = raw_smp_processor_id();
5574 unsigned long j0, j1, distance, max_distance = 0;
5575 struct sched_domain *sd;
5580 * First pass - calculate the cacheflush times:
5582 for_each_cpu_mask(cpu1, *cpu_map) {
5583 for_each_cpu_mask(cpu2, *cpu_map) {
5586 distance = domain_distance(cpu1, cpu2);
5587 max_distance = max(max_distance, distance);
5589 * No result cached yet?
5591 if (migration_cost[distance] == -1LL)
5592 migration_cost[distance] =
5593 measure_migration_cost(cpu1, cpu2);
5597 * Second pass - update the sched domain hierarchy with
5598 * the new cache-hot-time estimations:
5600 for_each_cpu_mask(cpu, *cpu_map) {
5602 for_each_domain(cpu, sd) {
5603 sd->cache_hot_time = migration_cost[distance];
5610 if (migration_debug)
5611 printk("migration: max_cache_size: %d, cpu: %d MHz:\n",
5619 if (system_state == SYSTEM_BOOTING) {
5620 printk("migration_cost=");
5621 for (distance = 0; distance <= max_distance; distance++) {
5624 printk("%ld", (long)migration_cost[distance] / 1000);
5629 if (migration_debug)
5630 printk("migration: %ld seconds\n", (j1-j0)/HZ);
5633 * Move back to the original CPU. NUMA-Q gets confused
5634 * if we migrate to another quad during bootup.
5636 if (raw_smp_processor_id() != orig_cpu) {
5637 cpumask_t mask = cpumask_of_cpu(orig_cpu),
5638 saved_mask = current->cpus_allowed;
5640 set_cpus_allowed(current, mask);
5641 set_cpus_allowed(current, saved_mask);
5648 * find_next_best_node - find the next node to include in a sched_domain
5649 * @node: node whose sched_domain we're building
5650 * @used_nodes: nodes already in the sched_domain
5652 * Find the next node to include in a given scheduling domain. Simply
5653 * finds the closest node not already in the @used_nodes map.
5655 * Should use nodemask_t.
5657 static int find_next_best_node(int node, unsigned long *used_nodes)
5659 int i, n, val, min_val, best_node = 0;
5663 for (i = 0; i < MAX_NUMNODES; i++) {
5664 /* Start at @node */
5665 n = (node + i) % MAX_NUMNODES;
5667 if (!nr_cpus_node(n))
5670 /* Skip already used nodes */
5671 if (test_bit(n, used_nodes))
5674 /* Simple min distance search */
5675 val = node_distance(node, n);
5677 if (val < min_val) {
5683 set_bit(best_node, used_nodes);
5688 * sched_domain_node_span - get a cpumask for a node's sched_domain
5689 * @node: node whose cpumask we're constructing
5690 * @size: number of nodes to include in this span
5692 * Given a node, construct a good cpumask for its sched_domain to span. It
5693 * should be one that prevents unnecessary balancing, but also spreads tasks
5696 static cpumask_t sched_domain_node_span(int node)
5699 cpumask_t span, nodemask;
5700 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5703 bitmap_zero(used_nodes, MAX_NUMNODES);
5705 nodemask = node_to_cpumask(node);
5706 cpus_or(span, span, nodemask);
5707 set_bit(node, used_nodes);
5709 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5710 int next_node = find_next_best_node(node, used_nodes);
5711 nodemask = node_to_cpumask(next_node);
5712 cpus_or(span, span, nodemask);
5720 * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we
5721 * can switch it on easily if needed.
5723 #ifdef CONFIG_SCHED_SMT
5724 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5725 static struct sched_group sched_group_cpus[NR_CPUS];
5726 static int cpu_to_cpu_group(int cpu)
5732 #ifdef CONFIG_SCHED_MC
5733 static DEFINE_PER_CPU(struct sched_domain, core_domains);
5734 static struct sched_group sched_group_core[NR_CPUS];
5737 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5738 static int cpu_to_core_group(int cpu)
5740 return first_cpu(cpu_sibling_map[cpu]);
5742 #elif defined(CONFIG_SCHED_MC)
5743 static int cpu_to_core_group(int cpu)
5749 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5750 static struct sched_group sched_group_phys[NR_CPUS];
5751 static int cpu_to_phys_group(int cpu)
5753 #if defined(CONFIG_SCHED_MC)
5754 cpumask_t mask = cpu_coregroup_map(cpu);
5755 return first_cpu(mask);
5756 #elif defined(CONFIG_SCHED_SMT)
5757 return first_cpu(cpu_sibling_map[cpu]);
5765 * The init_sched_build_groups can't handle what we want to do with node
5766 * groups, so roll our own. Now each node has its own list of groups which
5767 * gets dynamically allocated.
5769 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5770 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5772 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5773 static struct sched_group *sched_group_allnodes_bycpu[NR_CPUS];
5775 static int cpu_to_allnodes_group(int cpu)
5777 return cpu_to_node(cpu);
5779 static void init_numa_sched_groups_power(struct sched_group *group_head)
5781 struct sched_group *sg = group_head;
5787 for_each_cpu_mask(j, sg->cpumask) {
5788 struct sched_domain *sd;
5790 sd = &per_cpu(phys_domains, j);
5791 if (j != first_cpu(sd->groups->cpumask)) {
5793 * Only add "power" once for each
5799 sg->cpu_power += sd->groups->cpu_power;
5802 if (sg != group_head)
5808 * Build sched domains for a given set of cpus and attach the sched domains
5809 * to the individual cpus
5811 void build_sched_domains(const cpumask_t *cpu_map)
5815 struct sched_group **sched_group_nodes = NULL;
5816 struct sched_group *sched_group_allnodes = NULL;
5819 * Allocate the per-node list of sched groups
5821 sched_group_nodes = kmalloc(sizeof(struct sched_group*)*MAX_NUMNODES,
5823 if (!sched_group_nodes) {
5824 printk(KERN_WARNING "Can not alloc sched group node list\n");
5827 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
5831 * Set up domains for cpus specified by the cpu_map.
5833 for_each_cpu_mask(i, *cpu_map) {
5835 struct sched_domain *sd = NULL, *p;
5836 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
5838 cpus_and(nodemask, nodemask, *cpu_map);
5841 if (cpus_weight(*cpu_map)
5842 > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
5843 if (!sched_group_allnodes) {
5844 sched_group_allnodes
5845 = kmalloc(sizeof(struct sched_group)
5848 if (!sched_group_allnodes) {
5850 "Can not alloc allnodes sched group\n");
5853 sched_group_allnodes_bycpu[i]
5854 = sched_group_allnodes;
5856 sd = &per_cpu(allnodes_domains, i);
5857 *sd = SD_ALLNODES_INIT;
5858 sd->span = *cpu_map;
5859 group = cpu_to_allnodes_group(i);
5860 sd->groups = &sched_group_allnodes[group];
5865 sd = &per_cpu(node_domains, i);
5867 sd->span = sched_domain_node_span(cpu_to_node(i));
5869 cpus_and(sd->span, sd->span, *cpu_map);
5873 sd = &per_cpu(phys_domains, i);
5874 group = cpu_to_phys_group(i);
5876 sd->span = nodemask;
5878 sd->groups = &sched_group_phys[group];
5880 #ifdef CONFIG_SCHED_MC
5882 sd = &per_cpu(core_domains, i);
5883 group = cpu_to_core_group(i);
5885 sd->span = cpu_coregroup_map(i);
5886 cpus_and(sd->span, sd->span, *cpu_map);
5888 sd->groups = &sched_group_core[group];
5891 #ifdef CONFIG_SCHED_SMT
5893 sd = &per_cpu(cpu_domains, i);
5894 group = cpu_to_cpu_group(i);
5895 *sd = SD_SIBLING_INIT;
5896 sd->span = cpu_sibling_map[i];
5897 cpus_and(sd->span, sd->span, *cpu_map);
5899 sd->groups = &sched_group_cpus[group];
5903 #ifdef CONFIG_SCHED_SMT
5904 /* Set up CPU (sibling) groups */
5905 for_each_cpu_mask(i, *cpu_map) {
5906 cpumask_t this_sibling_map = cpu_sibling_map[i];
5907 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
5908 if (i != first_cpu(this_sibling_map))
5911 init_sched_build_groups(sched_group_cpus, this_sibling_map,
5916 #ifdef CONFIG_SCHED_MC
5917 /* Set up multi-core groups */
5918 for_each_cpu_mask(i, *cpu_map) {
5919 cpumask_t this_core_map = cpu_coregroup_map(i);
5920 cpus_and(this_core_map, this_core_map, *cpu_map);
5921 if (i != first_cpu(this_core_map))
5923 init_sched_build_groups(sched_group_core, this_core_map,
5924 &cpu_to_core_group);
5929 /* Set up physical groups */
5930 for (i = 0; i < MAX_NUMNODES; i++) {
5931 cpumask_t nodemask = node_to_cpumask(i);
5933 cpus_and(nodemask, nodemask, *cpu_map);
5934 if (cpus_empty(nodemask))
5937 init_sched_build_groups(sched_group_phys, nodemask,
5938 &cpu_to_phys_group);
5942 /* Set up node groups */
5943 if (sched_group_allnodes)
5944 init_sched_build_groups(sched_group_allnodes, *cpu_map,
5945 &cpu_to_allnodes_group);
5947 for (i = 0; i < MAX_NUMNODES; i++) {
5948 /* Set up node groups */
5949 struct sched_group *sg, *prev;
5950 cpumask_t nodemask = node_to_cpumask(i);
5951 cpumask_t domainspan;
5952 cpumask_t covered = CPU_MASK_NONE;
5955 cpus_and(nodemask, nodemask, *cpu_map);
5956 if (cpus_empty(nodemask)) {
5957 sched_group_nodes[i] = NULL;
5961 domainspan = sched_domain_node_span(i);
5962 cpus_and(domainspan, domainspan, *cpu_map);
5964 sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL);
5965 sched_group_nodes[i] = sg;
5966 for_each_cpu_mask(j, nodemask) {
5967 struct sched_domain *sd;
5968 sd = &per_cpu(node_domains, j);
5970 if (sd->groups == NULL) {
5971 /* Turn off balancing if we have no groups */
5977 "Can not alloc domain group for node %d\n", i);
5981 sg->cpumask = nodemask;
5982 cpus_or(covered, covered, nodemask);
5985 for (j = 0; j < MAX_NUMNODES; j++) {
5986 cpumask_t tmp, notcovered;
5987 int n = (i + j) % MAX_NUMNODES;
5989 cpus_complement(notcovered, covered);
5990 cpus_and(tmp, notcovered, *cpu_map);
5991 cpus_and(tmp, tmp, domainspan);
5992 if (cpus_empty(tmp))
5995 nodemask = node_to_cpumask(n);
5996 cpus_and(tmp, tmp, nodemask);
5997 if (cpus_empty(tmp))
6000 sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL);
6003 "Can not alloc domain group for node %d\n", j);
6008 cpus_or(covered, covered, tmp);
6012 prev->next = sched_group_nodes[i];
6016 /* Calculate CPU power for physical packages and nodes */
6017 for_each_cpu_mask(i, *cpu_map) {
6019 struct sched_domain *sd;
6020 #ifdef CONFIG_SCHED_SMT
6021 sd = &per_cpu(cpu_domains, i);
6022 power = SCHED_LOAD_SCALE;
6023 sd->groups->cpu_power = power;
6025 #ifdef CONFIG_SCHED_MC
6026 sd = &per_cpu(core_domains, i);
6027 power = SCHED_LOAD_SCALE + (cpus_weight(sd->groups->cpumask)-1)
6028 * SCHED_LOAD_SCALE / 10;
6029 sd->groups->cpu_power = power;
6031 sd = &per_cpu(phys_domains, i);
6034 * This has to be < 2 * SCHED_LOAD_SCALE
6035 * Lets keep it SCHED_LOAD_SCALE, so that
6036 * while calculating NUMA group's cpu_power
6038 * numa_group->cpu_power += phys_group->cpu_power;
6040 * See "only add power once for each physical pkg"
6043 sd->groups->cpu_power = SCHED_LOAD_SCALE;
6045 sd = &per_cpu(phys_domains, i);
6046 power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
6047 (cpus_weight(sd->groups->cpumask)-1) / 10;
6048 sd->groups->cpu_power = power;
6053 for (i = 0; i < MAX_NUMNODES; i++)
6054 init_numa_sched_groups_power(sched_group_nodes[i]);
6056 init_numa_sched_groups_power(sched_group_allnodes);
6059 /* Attach the domains */
6060 for_each_cpu_mask(i, *cpu_map) {
6061 struct sched_domain *sd;
6062 #ifdef CONFIG_SCHED_SMT
6063 sd = &per_cpu(cpu_domains, i);
6064 #elif defined(CONFIG_SCHED_MC)
6065 sd = &per_cpu(core_domains, i);
6067 sd = &per_cpu(phys_domains, i);
6069 cpu_attach_domain(sd, i);
6072 * Tune cache-hot values:
6074 calibrate_migration_costs(cpu_map);
6077 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6079 static void arch_init_sched_domains(const cpumask_t *cpu_map)
6081 cpumask_t cpu_default_map;
6084 * Setup mask for cpus without special case scheduling requirements.
6085 * For now this just excludes isolated cpus, but could be used to
6086 * exclude other special cases in the future.
6088 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6090 build_sched_domains(&cpu_default_map);
6093 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6099 for_each_cpu_mask(cpu, *cpu_map) {
6100 struct sched_group *sched_group_allnodes
6101 = sched_group_allnodes_bycpu[cpu];
6102 struct sched_group **sched_group_nodes
6103 = sched_group_nodes_bycpu[cpu];
6105 if (sched_group_allnodes) {
6106 kfree(sched_group_allnodes);
6107 sched_group_allnodes_bycpu[cpu] = NULL;
6110 if (!sched_group_nodes)
6113 for (i = 0; i < MAX_NUMNODES; i++) {
6114 cpumask_t nodemask = node_to_cpumask(i);
6115 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6117 cpus_and(nodemask, nodemask, *cpu_map);
6118 if (cpus_empty(nodemask))
6128 if (oldsg != sched_group_nodes[i])
6131 kfree(sched_group_nodes);
6132 sched_group_nodes_bycpu[cpu] = NULL;
6138 * Detach sched domains from a group of cpus specified in cpu_map
6139 * These cpus will now be attached to the NULL domain
6141 static void detach_destroy_domains(const cpumask_t *cpu_map)
6145 for_each_cpu_mask(i, *cpu_map)
6146 cpu_attach_domain(NULL, i);
6147 synchronize_sched();
6148 arch_destroy_sched_domains(cpu_map);
6152 * Partition sched domains as specified by the cpumasks below.
6153 * This attaches all cpus from the cpumasks to the NULL domain,
6154 * waits for a RCU quiescent period, recalculates sched
6155 * domain information and then attaches them back to the
6156 * correct sched domains
6157 * Call with hotplug lock held
6159 void partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6161 cpumask_t change_map;
6163 cpus_and(*partition1, *partition1, cpu_online_map);
6164 cpus_and(*partition2, *partition2, cpu_online_map);
6165 cpus_or(change_map, *partition1, *partition2);
6167 /* Detach sched domains from all of the affected cpus */
6168 detach_destroy_domains(&change_map);
6169 if (!cpus_empty(*partition1))
6170 build_sched_domains(partition1);
6171 if (!cpus_empty(*partition2))
6172 build_sched_domains(partition2);
6175 #ifdef CONFIG_HOTPLUG_CPU
6177 * Force a reinitialization of the sched domains hierarchy. The domains
6178 * and groups cannot be updated in place without racing with the balancing
6179 * code, so we temporarily attach all running cpus to the NULL domain
6180 * which will prevent rebalancing while the sched domains are recalculated.
6182 static int update_sched_domains(struct notifier_block *nfb,
6183 unsigned long action, void *hcpu)
6186 case CPU_UP_PREPARE:
6187 case CPU_DOWN_PREPARE:
6188 detach_destroy_domains(&cpu_online_map);
6191 case CPU_UP_CANCELED:
6192 case CPU_DOWN_FAILED:
6196 * Fall through and re-initialise the domains.
6203 /* The hotplug lock is already held by cpu_up/cpu_down */
6204 arch_init_sched_domains(&cpu_online_map);
6210 void __init sched_init_smp(void)
6213 arch_init_sched_domains(&cpu_online_map);
6214 unlock_cpu_hotplug();
6215 /* XXX: Theoretical race here - CPU may be hotplugged now */
6216 hotcpu_notifier(update_sched_domains, 0);
6219 void __init sched_init_smp(void)
6222 #endif /* CONFIG_SMP */
6224 int in_sched_functions(unsigned long addr)
6226 /* Linker adds these: start and end of __sched functions */
6227 extern char __sched_text_start[], __sched_text_end[];
6228 return in_lock_functions(addr) ||
6229 (addr >= (unsigned long)__sched_text_start
6230 && addr < (unsigned long)__sched_text_end);
6233 void __init sched_init(void)
6238 for_each_possible_cpu(i) {
6239 prio_array_t *array;
6242 spin_lock_init(&rq->lock);
6244 rq->active = rq->arrays;
6245 rq->expired = rq->arrays + 1;
6246 rq->best_expired_prio = MAX_PRIO;
6250 for (j = 1; j < 3; j++)
6251 rq->cpu_load[j] = 0;
6252 rq->active_balance = 0;
6254 rq->migration_thread = NULL;
6255 INIT_LIST_HEAD(&rq->migration_queue);
6257 atomic_set(&rq->nr_iowait, 0);
6259 for (j = 0; j < 2; j++) {
6260 array = rq->arrays + j;
6261 for (k = 0; k < MAX_PRIO; k++) {
6262 INIT_LIST_HEAD(array->queue + k);
6263 __clear_bit(k, array->bitmap);
6265 // delimiter for bitsearch
6266 __set_bit(MAX_PRIO, array->bitmap);
6270 set_load_weight(&init_task);
6272 * The boot idle thread does lazy MMU switching as well:
6274 atomic_inc(&init_mm.mm_count);
6275 enter_lazy_tlb(&init_mm, current);
6278 * Make us the idle thread. Technically, schedule() should not be
6279 * called from this thread, however somewhere below it might be,
6280 * but because we are the idle thread, we just pick up running again
6281 * when this runqueue becomes "idle".
6283 init_idle(current, smp_processor_id());
6286 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6287 void __might_sleep(char *file, int line)
6289 #if defined(in_atomic)
6290 static unsigned long prev_jiffy; /* ratelimiting */
6292 if ((in_atomic() || irqs_disabled()) &&
6293 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6294 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6296 prev_jiffy = jiffies;
6297 printk(KERN_ERR "BUG: sleeping function called from invalid"
6298 " context at %s:%d\n", file, line);
6299 printk("in_atomic():%d, irqs_disabled():%d\n",
6300 in_atomic(), irqs_disabled());
6305 EXPORT_SYMBOL(__might_sleep);
6308 #ifdef CONFIG_MAGIC_SYSRQ
6309 void normalize_rt_tasks(void)
6311 struct task_struct *p;
6312 prio_array_t *array;
6313 unsigned long flags;
6316 read_lock_irq(&tasklist_lock);
6317 for_each_process(p) {
6321 rq = task_rq_lock(p, &flags);
6325 deactivate_task(p, task_rq(p));
6326 __setscheduler(p, SCHED_NORMAL, 0);
6328 __activate_task(p, task_rq(p));
6329 resched_task(rq->curr);
6332 task_rq_unlock(rq, &flags);
6334 read_unlock_irq(&tasklist_lock);
6337 #endif /* CONFIG_MAGIC_SYSRQ */
6341 * These functions are only useful for the IA64 MCA handling.
6343 * They can only be called when the whole system has been
6344 * stopped - every CPU needs to be quiescent, and no scheduling
6345 * activity can take place. Using them for anything else would
6346 * be a serious bug, and as a result, they aren't even visible
6347 * under any other configuration.
6351 * curr_task - return the current task for a given cpu.
6352 * @cpu: the processor in question.
6354 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6356 task_t *curr_task(int cpu)
6358 return cpu_curr(cpu);
6362 * set_curr_task - set the current task for a given cpu.
6363 * @cpu: the processor in question.
6364 * @p: the task pointer to set.
6366 * Description: This function must only be used when non-maskable interrupts
6367 * are serviced on a separate stack. It allows the architecture to switch the
6368 * notion of the current task on a cpu in a non-blocking manner. This function
6369 * must be called with all CPU's synchronized, and interrupts disabled, the
6370 * and caller must save the original value of the current task (see
6371 * curr_task() above) and restore that value before reenabling interrupts and
6372 * re-starting the system.
6374 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6376 void set_curr_task(int cpu, task_t *p)