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 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) {
1166 * If power savings logic is enabled for a domain, stop there.
1168 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1170 if (tmp->flags & flag)
1176 struct sched_group *group;
1181 group = find_idlest_group(sd, t, cpu);
1185 new_cpu = find_idlest_cpu(group, t, cpu);
1186 if (new_cpu == -1 || new_cpu == cpu)
1189 /* Now try balancing at a lower domain level */
1193 weight = cpus_weight(span);
1194 for_each_domain(cpu, tmp) {
1195 if (weight <= cpus_weight(tmp->span))
1197 if (tmp->flags & flag)
1200 /* while loop will break here if sd == NULL */
1206 #endif /* CONFIG_SMP */
1209 * wake_idle() will wake a task on an idle cpu if task->cpu is
1210 * not idle and an idle cpu is available. The span of cpus to
1211 * search starts with cpus closest then further out as needed,
1212 * so we always favor a closer, idle cpu.
1214 * Returns the CPU we should wake onto.
1216 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1217 static int wake_idle(int cpu, task_t *p)
1220 struct sched_domain *sd;
1226 for_each_domain(cpu, sd) {
1227 if (sd->flags & SD_WAKE_IDLE) {
1228 cpus_and(tmp, sd->span, p->cpus_allowed);
1229 for_each_cpu_mask(i, tmp) {
1240 static inline int wake_idle(int cpu, task_t *p)
1247 * try_to_wake_up - wake up a thread
1248 * @p: the to-be-woken-up thread
1249 * @state: the mask of task states that can be woken
1250 * @sync: do a synchronous wakeup?
1252 * Put it on the run-queue if it's not already there. The "current"
1253 * thread is always on the run-queue (except when the actual
1254 * re-schedule is in progress), and as such you're allowed to do
1255 * the simpler "current->state = TASK_RUNNING" to mark yourself
1256 * runnable without the overhead of this.
1258 * returns failure only if the task is already active.
1260 static int try_to_wake_up(task_t *p, unsigned int state, int sync)
1262 int cpu, this_cpu, success = 0;
1263 unsigned long flags;
1267 unsigned long load, this_load;
1268 struct sched_domain *sd, *this_sd = NULL;
1272 rq = task_rq_lock(p, &flags);
1273 old_state = p->state;
1274 if (!(old_state & state))
1281 this_cpu = smp_processor_id();
1284 if (unlikely(task_running(rq, p)))
1289 schedstat_inc(rq, ttwu_cnt);
1290 if (cpu == this_cpu) {
1291 schedstat_inc(rq, ttwu_local);
1295 for_each_domain(this_cpu, sd) {
1296 if (cpu_isset(cpu, sd->span)) {
1297 schedstat_inc(sd, ttwu_wake_remote);
1303 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1307 * Check for affine wakeup and passive balancing possibilities.
1310 int idx = this_sd->wake_idx;
1311 unsigned int imbalance;
1313 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1315 load = source_load(cpu, idx);
1316 this_load = target_load(this_cpu, idx);
1318 new_cpu = this_cpu; /* Wake to this CPU if we can */
1320 if (this_sd->flags & SD_WAKE_AFFINE) {
1321 unsigned long tl = this_load;
1322 unsigned long tl_per_task = cpu_avg_load_per_task(this_cpu);
1325 * If sync wakeup then subtract the (maximum possible)
1326 * effect of the currently running task from the load
1327 * of the current CPU:
1330 tl -= current->load_weight;
1333 tl + target_load(cpu, idx) <= tl_per_task) ||
1334 100*(tl + p->load_weight) <= imbalance*load) {
1336 * This domain has SD_WAKE_AFFINE and
1337 * p is cache cold in this domain, and
1338 * there is no bad imbalance.
1340 schedstat_inc(this_sd, ttwu_move_affine);
1346 * Start passive balancing when half the imbalance_pct
1349 if (this_sd->flags & SD_WAKE_BALANCE) {
1350 if (imbalance*this_load <= 100*load) {
1351 schedstat_inc(this_sd, ttwu_move_balance);
1357 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1359 new_cpu = wake_idle(new_cpu, p);
1360 if (new_cpu != cpu) {
1361 set_task_cpu(p, new_cpu);
1362 task_rq_unlock(rq, &flags);
1363 /* might preempt at this point */
1364 rq = task_rq_lock(p, &flags);
1365 old_state = p->state;
1366 if (!(old_state & state))
1371 this_cpu = smp_processor_id();
1376 #endif /* CONFIG_SMP */
1377 if (old_state == TASK_UNINTERRUPTIBLE) {
1378 rq->nr_uninterruptible--;
1380 * Tasks on involuntary sleep don't earn
1381 * sleep_avg beyond just interactive state.
1383 p->sleep_type = SLEEP_NONINTERACTIVE;
1387 * Tasks that have marked their sleep as noninteractive get
1388 * woken up with their sleep average not weighted in an
1391 if (old_state & TASK_NONINTERACTIVE)
1392 p->sleep_type = SLEEP_NONINTERACTIVE;
1395 activate_task(p, rq, cpu == this_cpu);
1397 * Sync wakeups (i.e. those types of wakeups where the waker
1398 * has indicated that it will leave the CPU in short order)
1399 * don't trigger a preemption, if the woken up task will run on
1400 * this cpu. (in this case the 'I will reschedule' promise of
1401 * the waker guarantees that the freshly woken up task is going
1402 * to be considered on this CPU.)
1404 if (!sync || cpu != this_cpu) {
1405 if (TASK_PREEMPTS_CURR(p, rq))
1406 resched_task(rq->curr);
1411 p->state = TASK_RUNNING;
1413 task_rq_unlock(rq, &flags);
1418 int fastcall wake_up_process(task_t *p)
1420 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1421 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1424 EXPORT_SYMBOL(wake_up_process);
1426 int fastcall wake_up_state(task_t *p, unsigned int state)
1428 return try_to_wake_up(p, state, 0);
1432 * Perform scheduler related setup for a newly forked process p.
1433 * p is forked by current.
1435 void fastcall sched_fork(task_t *p, int clone_flags)
1437 int cpu = get_cpu();
1440 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1442 set_task_cpu(p, cpu);
1445 * We mark the process as running here, but have not actually
1446 * inserted it onto the runqueue yet. This guarantees that
1447 * nobody will actually run it, and a signal or other external
1448 * event cannot wake it up and insert it on the runqueue either.
1450 p->state = TASK_RUNNING;
1451 INIT_LIST_HEAD(&p->run_list);
1453 #ifdef CONFIG_SCHEDSTATS
1454 memset(&p->sched_info, 0, sizeof(p->sched_info));
1456 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1459 #ifdef CONFIG_PREEMPT
1460 /* Want to start with kernel preemption disabled. */
1461 task_thread_info(p)->preempt_count = 1;
1464 * Share the timeslice between parent and child, thus the
1465 * total amount of pending timeslices in the system doesn't change,
1466 * resulting in more scheduling fairness.
1468 local_irq_disable();
1469 p->time_slice = (current->time_slice + 1) >> 1;
1471 * The remainder of the first timeslice might be recovered by
1472 * the parent if the child exits early enough.
1474 p->first_time_slice = 1;
1475 current->time_slice >>= 1;
1476 p->timestamp = sched_clock();
1477 if (unlikely(!current->time_slice)) {
1479 * This case is rare, it happens when the parent has only
1480 * a single jiffy left from its timeslice. Taking the
1481 * runqueue lock is not a problem.
1483 current->time_slice = 1;
1491 * wake_up_new_task - wake up a newly created task for the first time.
1493 * This function will do some initial scheduler statistics housekeeping
1494 * that must be done for every newly created context, then puts the task
1495 * on the runqueue and wakes it.
1497 void fastcall wake_up_new_task(task_t *p, unsigned long clone_flags)
1499 unsigned long flags;
1501 runqueue_t *rq, *this_rq;
1503 rq = task_rq_lock(p, &flags);
1504 BUG_ON(p->state != TASK_RUNNING);
1505 this_cpu = smp_processor_id();
1509 * We decrease the sleep average of forking parents
1510 * and children as well, to keep max-interactive tasks
1511 * from forking tasks that are max-interactive. The parent
1512 * (current) is done further down, under its lock.
1514 p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1515 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1517 p->prio = effective_prio(p);
1519 if (likely(cpu == this_cpu)) {
1520 if (!(clone_flags & CLONE_VM)) {
1522 * The VM isn't cloned, so we're in a good position to
1523 * do child-runs-first in anticipation of an exec. This
1524 * usually avoids a lot of COW overhead.
1526 if (unlikely(!current->array))
1527 __activate_task(p, rq);
1529 p->prio = current->prio;
1530 list_add_tail(&p->run_list, ¤t->run_list);
1531 p->array = current->array;
1532 p->array->nr_active++;
1533 inc_nr_running(p, rq);
1537 /* Run child last */
1538 __activate_task(p, rq);
1540 * We skip the following code due to cpu == this_cpu
1542 * task_rq_unlock(rq, &flags);
1543 * this_rq = task_rq_lock(current, &flags);
1547 this_rq = cpu_rq(this_cpu);
1550 * Not the local CPU - must adjust timestamp. This should
1551 * get optimised away in the !CONFIG_SMP case.
1553 p->timestamp = (p->timestamp - this_rq->timestamp_last_tick)
1554 + rq->timestamp_last_tick;
1555 __activate_task(p, rq);
1556 if (TASK_PREEMPTS_CURR(p, rq))
1557 resched_task(rq->curr);
1560 * Parent and child are on different CPUs, now get the
1561 * parent runqueue to update the parent's ->sleep_avg:
1563 task_rq_unlock(rq, &flags);
1564 this_rq = task_rq_lock(current, &flags);
1566 current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1567 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1568 task_rq_unlock(this_rq, &flags);
1572 * Potentially available exiting-child timeslices are
1573 * retrieved here - this way the parent does not get
1574 * penalized for creating too many threads.
1576 * (this cannot be used to 'generate' timeslices
1577 * artificially, because any timeslice recovered here
1578 * was given away by the parent in the first place.)
1580 void fastcall sched_exit(task_t *p)
1582 unsigned long flags;
1586 * If the child was a (relative-) CPU hog then decrease
1587 * the sleep_avg of the parent as well.
1589 rq = task_rq_lock(p->parent, &flags);
1590 if (p->first_time_slice && task_cpu(p) == task_cpu(p->parent)) {
1591 p->parent->time_slice += p->time_slice;
1592 if (unlikely(p->parent->time_slice > task_timeslice(p)))
1593 p->parent->time_slice = task_timeslice(p);
1595 if (p->sleep_avg < p->parent->sleep_avg)
1596 p->parent->sleep_avg = p->parent->sleep_avg /
1597 (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg /
1599 task_rq_unlock(rq, &flags);
1603 * prepare_task_switch - prepare to switch tasks
1604 * @rq: the runqueue preparing to switch
1605 * @next: the task we are going to switch to.
1607 * This is called with the rq lock held and interrupts off. It must
1608 * be paired with a subsequent finish_task_switch after the context
1611 * prepare_task_switch sets up locking and calls architecture specific
1614 static inline void prepare_task_switch(runqueue_t *rq, task_t *next)
1616 prepare_lock_switch(rq, next);
1617 prepare_arch_switch(next);
1621 * finish_task_switch - clean up after a task-switch
1622 * @rq: runqueue associated with task-switch
1623 * @prev: the thread we just switched away from.
1625 * finish_task_switch must be called after the context switch, paired
1626 * with a prepare_task_switch call before the context switch.
1627 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1628 * and do any other architecture-specific cleanup actions.
1630 * Note that we may have delayed dropping an mm in context_switch(). If
1631 * so, we finish that here outside of the runqueue lock. (Doing it
1632 * with the lock held can cause deadlocks; see schedule() for
1635 static inline void finish_task_switch(runqueue_t *rq, task_t *prev)
1636 __releases(rq->lock)
1638 struct mm_struct *mm = rq->prev_mm;
1639 unsigned long prev_task_flags;
1644 * A task struct has one reference for the use as "current".
1645 * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1646 * calls schedule one last time. The schedule call will never return,
1647 * and the scheduled task must drop that reference.
1648 * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1649 * still held, otherwise prev could be scheduled on another cpu, die
1650 * there before we look at prev->state, and then the reference would
1652 * Manfred Spraul <manfred@colorfullife.com>
1654 prev_task_flags = prev->flags;
1655 finish_arch_switch(prev);
1656 finish_lock_switch(rq, prev);
1659 if (unlikely(prev_task_flags & PF_DEAD)) {
1661 * Remove function-return probe instances associated with this
1662 * task and put them back on the free list.
1664 kprobe_flush_task(prev);
1665 put_task_struct(prev);
1670 * schedule_tail - first thing a freshly forked thread must call.
1671 * @prev: the thread we just switched away from.
1673 asmlinkage void schedule_tail(task_t *prev)
1674 __releases(rq->lock)
1676 runqueue_t *rq = this_rq();
1677 finish_task_switch(rq, prev);
1678 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1679 /* In this case, finish_task_switch does not reenable preemption */
1682 if (current->set_child_tid)
1683 put_user(current->pid, current->set_child_tid);
1687 * context_switch - switch to the new MM and the new
1688 * thread's register state.
1691 task_t * context_switch(runqueue_t *rq, task_t *prev, task_t *next)
1693 struct mm_struct *mm = next->mm;
1694 struct mm_struct *oldmm = prev->active_mm;
1696 if (unlikely(!mm)) {
1697 next->active_mm = oldmm;
1698 atomic_inc(&oldmm->mm_count);
1699 enter_lazy_tlb(oldmm, next);
1701 switch_mm(oldmm, mm, next);
1703 if (unlikely(!prev->mm)) {
1704 prev->active_mm = NULL;
1705 WARN_ON(rq->prev_mm);
1706 rq->prev_mm = oldmm;
1709 /* Here we just switch the register state and the stack. */
1710 switch_to(prev, next, prev);
1716 * nr_running, nr_uninterruptible and nr_context_switches:
1718 * externally visible scheduler statistics: current number of runnable
1719 * threads, current number of uninterruptible-sleeping threads, total
1720 * number of context switches performed since bootup.
1722 unsigned long nr_running(void)
1724 unsigned long i, sum = 0;
1726 for_each_online_cpu(i)
1727 sum += cpu_rq(i)->nr_running;
1732 unsigned long nr_uninterruptible(void)
1734 unsigned long i, sum = 0;
1736 for_each_possible_cpu(i)
1737 sum += cpu_rq(i)->nr_uninterruptible;
1740 * Since we read the counters lockless, it might be slightly
1741 * inaccurate. Do not allow it to go below zero though:
1743 if (unlikely((long)sum < 0))
1749 unsigned long long nr_context_switches(void)
1752 unsigned long long sum = 0;
1754 for_each_possible_cpu(i)
1755 sum += cpu_rq(i)->nr_switches;
1760 unsigned long nr_iowait(void)
1762 unsigned long i, sum = 0;
1764 for_each_possible_cpu(i)
1765 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1770 unsigned long nr_active(void)
1772 unsigned long i, running = 0, uninterruptible = 0;
1774 for_each_online_cpu(i) {
1775 running += cpu_rq(i)->nr_running;
1776 uninterruptible += cpu_rq(i)->nr_uninterruptible;
1779 if (unlikely((long)uninterruptible < 0))
1780 uninterruptible = 0;
1782 return running + uninterruptible;
1788 * double_rq_lock - safely lock two runqueues
1790 * Note this does not disable interrupts like task_rq_lock,
1791 * you need to do so manually before calling.
1793 static void double_rq_lock(runqueue_t *rq1, runqueue_t *rq2)
1794 __acquires(rq1->lock)
1795 __acquires(rq2->lock)
1798 spin_lock(&rq1->lock);
1799 __acquire(rq2->lock); /* Fake it out ;) */
1802 spin_lock(&rq1->lock);
1803 spin_lock(&rq2->lock);
1805 spin_lock(&rq2->lock);
1806 spin_lock(&rq1->lock);
1812 * double_rq_unlock - safely unlock two runqueues
1814 * Note this does not restore interrupts like task_rq_unlock,
1815 * you need to do so manually after calling.
1817 static void double_rq_unlock(runqueue_t *rq1, runqueue_t *rq2)
1818 __releases(rq1->lock)
1819 __releases(rq2->lock)
1821 spin_unlock(&rq1->lock);
1823 spin_unlock(&rq2->lock);
1825 __release(rq2->lock);
1829 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1831 static void double_lock_balance(runqueue_t *this_rq, runqueue_t *busiest)
1832 __releases(this_rq->lock)
1833 __acquires(busiest->lock)
1834 __acquires(this_rq->lock)
1836 if (unlikely(!spin_trylock(&busiest->lock))) {
1837 if (busiest < this_rq) {
1838 spin_unlock(&this_rq->lock);
1839 spin_lock(&busiest->lock);
1840 spin_lock(&this_rq->lock);
1842 spin_lock(&busiest->lock);
1847 * If dest_cpu is allowed for this process, migrate the task to it.
1848 * This is accomplished by forcing the cpu_allowed mask to only
1849 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
1850 * the cpu_allowed mask is restored.
1852 static void sched_migrate_task(task_t *p, int dest_cpu)
1854 migration_req_t req;
1856 unsigned long flags;
1858 rq = task_rq_lock(p, &flags);
1859 if (!cpu_isset(dest_cpu, p->cpus_allowed)
1860 || unlikely(cpu_is_offline(dest_cpu)))
1863 /* force the process onto the specified CPU */
1864 if (migrate_task(p, dest_cpu, &req)) {
1865 /* Need to wait for migration thread (might exit: take ref). */
1866 struct task_struct *mt = rq->migration_thread;
1867 get_task_struct(mt);
1868 task_rq_unlock(rq, &flags);
1869 wake_up_process(mt);
1870 put_task_struct(mt);
1871 wait_for_completion(&req.done);
1875 task_rq_unlock(rq, &flags);
1879 * sched_exec - execve() is a valuable balancing opportunity, because at
1880 * this point the task has the smallest effective memory and cache footprint.
1882 void sched_exec(void)
1884 int new_cpu, this_cpu = get_cpu();
1885 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
1887 if (new_cpu != this_cpu)
1888 sched_migrate_task(current, new_cpu);
1892 * pull_task - move a task from a remote runqueue to the local runqueue.
1893 * Both runqueues must be locked.
1896 void pull_task(runqueue_t *src_rq, prio_array_t *src_array, task_t *p,
1897 runqueue_t *this_rq, prio_array_t *this_array, int this_cpu)
1899 dequeue_task(p, src_array);
1900 dec_nr_running(p, src_rq);
1901 set_task_cpu(p, this_cpu);
1902 inc_nr_running(p, this_rq);
1903 enqueue_task(p, this_array);
1904 p->timestamp = (p->timestamp - src_rq->timestamp_last_tick)
1905 + this_rq->timestamp_last_tick;
1907 * Note that idle threads have a prio of MAX_PRIO, for this test
1908 * to be always true for them.
1910 if (TASK_PREEMPTS_CURR(p, this_rq))
1911 resched_task(this_rq->curr);
1915 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1918 int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu,
1919 struct sched_domain *sd, enum idle_type idle,
1923 * We do not migrate tasks that are:
1924 * 1) running (obviously), or
1925 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1926 * 3) are cache-hot on their current CPU.
1928 if (!cpu_isset(this_cpu, p->cpus_allowed))
1932 if (task_running(rq, p))
1936 * Aggressive migration if:
1937 * 1) task is cache cold, or
1938 * 2) too many balance attempts have failed.
1941 if (sd->nr_balance_failed > sd->cache_nice_tries)
1944 if (task_hot(p, rq->timestamp_last_tick, sd))
1949 #define rq_best_prio(rq) min((rq)->curr->prio, (rq)->best_expired_prio)
1951 * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted
1952 * load from busiest to this_rq, as part of a balancing operation within
1953 * "domain". Returns the number of tasks moved.
1955 * Called with both runqueues locked.
1957 static int move_tasks(runqueue_t *this_rq, int this_cpu, runqueue_t *busiest,
1958 unsigned long max_nr_move, unsigned long max_load_move,
1959 struct sched_domain *sd, enum idle_type idle,
1962 prio_array_t *array, *dst_array;
1963 struct list_head *head, *curr;
1964 int idx, pulled = 0, pinned = 0, this_best_prio, busiest_best_prio;
1965 int busiest_best_prio_seen;
1966 int skip_for_load; /* skip the task based on weighted load issues */
1970 if (max_nr_move == 0 || max_load_move == 0)
1973 rem_load_move = max_load_move;
1975 this_best_prio = rq_best_prio(this_rq);
1976 busiest_best_prio = rq_best_prio(busiest);
1978 * Enable handling of the case where there is more than one task
1979 * with the best priority. If the current running task is one
1980 * of those with prio==busiest_best_prio we know it won't be moved
1981 * and therefore it's safe to override the skip (based on load) of
1982 * any task we find with that prio.
1984 busiest_best_prio_seen = busiest_best_prio == busiest->curr->prio;
1987 * We first consider expired tasks. Those will likely not be
1988 * executed in the near future, and they are most likely to
1989 * be cache-cold, thus switching CPUs has the least effect
1992 if (busiest->expired->nr_active) {
1993 array = busiest->expired;
1994 dst_array = this_rq->expired;
1996 array = busiest->active;
1997 dst_array = this_rq->active;
2001 /* Start searching at priority 0: */
2005 idx = sched_find_first_bit(array->bitmap);
2007 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
2008 if (idx >= MAX_PRIO) {
2009 if (array == busiest->expired && busiest->active->nr_active) {
2010 array = busiest->active;
2011 dst_array = this_rq->active;
2017 head = array->queue + idx;
2020 tmp = list_entry(curr, task_t, run_list);
2025 * To help distribute high priority tasks accross CPUs we don't
2026 * skip a task if it will be the highest priority task (i.e. smallest
2027 * prio value) on its new queue regardless of its load weight
2029 skip_for_load = tmp->load_weight > rem_load_move;
2030 if (skip_for_load && idx < this_best_prio)
2031 skip_for_load = !busiest_best_prio_seen && idx == busiest_best_prio;
2032 if (skip_for_load ||
2033 !can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
2034 busiest_best_prio_seen |= idx == busiest_best_prio;
2041 #ifdef CONFIG_SCHEDSTATS
2042 if (task_hot(tmp, busiest->timestamp_last_tick, sd))
2043 schedstat_inc(sd, lb_hot_gained[idle]);
2046 pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
2048 rem_load_move -= tmp->load_weight;
2051 * We only want to steal up to the prescribed number of tasks
2052 * and the prescribed amount of weighted load.
2054 if (pulled < max_nr_move && rem_load_move > 0) {
2055 if (idx < this_best_prio)
2056 this_best_prio = idx;
2064 * Right now, this is the only place pull_task() is called,
2065 * so we can safely collect pull_task() stats here rather than
2066 * inside pull_task().
2068 schedstat_add(sd, lb_gained[idle], pulled);
2071 *all_pinned = pinned;
2076 * find_busiest_group finds and returns the busiest CPU group within the
2077 * domain. It calculates and returns the amount of weighted load which should be
2078 * moved to restore balance via the imbalance parameter.
2080 static struct sched_group *
2081 find_busiest_group(struct sched_domain *sd, int this_cpu,
2082 unsigned long *imbalance, enum idle_type idle, int *sd_idle)
2084 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2085 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2086 unsigned long max_pull;
2087 unsigned long busiest_load_per_task, busiest_nr_running;
2088 unsigned long this_load_per_task, this_nr_running;
2090 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2091 int power_savings_balance = 1;
2092 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2093 unsigned long min_nr_running = ULONG_MAX;
2094 struct sched_group *group_min = NULL, *group_leader = NULL;
2097 max_load = this_load = total_load = total_pwr = 0;
2098 busiest_load_per_task = busiest_nr_running = 0;
2099 this_load_per_task = this_nr_running = 0;
2100 if (idle == NOT_IDLE)
2101 load_idx = sd->busy_idx;
2102 else if (idle == NEWLY_IDLE)
2103 load_idx = sd->newidle_idx;
2105 load_idx = sd->idle_idx;
2108 unsigned long load, group_capacity;
2111 unsigned long sum_nr_running, sum_weighted_load;
2113 local_group = cpu_isset(this_cpu, group->cpumask);
2115 /* Tally up the load of all CPUs in the group */
2116 sum_weighted_load = sum_nr_running = avg_load = 0;
2118 for_each_cpu_mask(i, group->cpumask) {
2119 runqueue_t *rq = cpu_rq(i);
2121 if (*sd_idle && !idle_cpu(i))
2124 /* Bias balancing toward cpus of our domain */
2126 load = target_load(i, load_idx);
2128 load = source_load(i, load_idx);
2131 sum_nr_running += rq->nr_running;
2132 sum_weighted_load += rq->raw_weighted_load;
2135 total_load += avg_load;
2136 total_pwr += group->cpu_power;
2138 /* Adjust by relative CPU power of the group */
2139 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
2141 group_capacity = group->cpu_power / SCHED_LOAD_SCALE;
2144 this_load = avg_load;
2146 this_nr_running = sum_nr_running;
2147 this_load_per_task = sum_weighted_load;
2148 } else if (avg_load > max_load &&
2149 sum_nr_running > group_capacity) {
2150 max_load = avg_load;
2152 busiest_nr_running = sum_nr_running;
2153 busiest_load_per_task = sum_weighted_load;
2156 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2158 * Busy processors will not participate in power savings
2161 if (idle == NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2165 * If the local group is idle or completely loaded
2166 * no need to do power savings balance at this domain
2168 if (local_group && (this_nr_running >= group_capacity ||
2170 power_savings_balance = 0;
2173 * If a group is already running at full capacity or idle,
2174 * don't include that group in power savings calculations
2176 if (!power_savings_balance || sum_nr_running >= group_capacity
2181 * Calculate the group which has the least non-idle load.
2182 * This is the group from where we need to pick up the load
2185 if ((sum_nr_running < min_nr_running) ||
2186 (sum_nr_running == min_nr_running &&
2187 first_cpu(group->cpumask) <
2188 first_cpu(group_min->cpumask))) {
2190 min_nr_running = sum_nr_running;
2191 min_load_per_task = sum_weighted_load /
2196 * Calculate the group which is almost near its
2197 * capacity but still has some space to pick up some load
2198 * from other group and save more power
2200 if (sum_nr_running <= group_capacity - 1)
2201 if (sum_nr_running > leader_nr_running ||
2202 (sum_nr_running == leader_nr_running &&
2203 first_cpu(group->cpumask) >
2204 first_cpu(group_leader->cpumask))) {
2205 group_leader = group;
2206 leader_nr_running = sum_nr_running;
2211 group = group->next;
2212 } while (group != sd->groups);
2214 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2217 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2219 if (this_load >= avg_load ||
2220 100*max_load <= sd->imbalance_pct*this_load)
2223 busiest_load_per_task /= busiest_nr_running;
2225 * We're trying to get all the cpus to the average_load, so we don't
2226 * want to push ourselves above the average load, nor do we wish to
2227 * reduce the max loaded cpu below the average load, as either of these
2228 * actions would just result in more rebalancing later, and ping-pong
2229 * tasks around. Thus we look for the minimum possible imbalance.
2230 * Negative imbalances (*we* are more loaded than anyone else) will
2231 * be counted as no imbalance for these purposes -- we can't fix that
2232 * by pulling tasks to us. Be careful of negative numbers as they'll
2233 * appear as very large values with unsigned longs.
2235 if (max_load <= busiest_load_per_task)
2239 * In the presence of smp nice balancing, certain scenarios can have
2240 * max load less than avg load(as we skip the groups at or below
2241 * its cpu_power, while calculating max_load..)
2243 if (max_load < avg_load) {
2245 goto small_imbalance;
2248 /* Don't want to pull so many tasks that a group would go idle */
2249 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2251 /* How much load to actually move to equalise the imbalance */
2252 *imbalance = min(max_pull * busiest->cpu_power,
2253 (avg_load - this_load) * this->cpu_power)
2257 * if *imbalance is less than the average load per runnable task
2258 * there is no gaurantee that any tasks will be moved so we'll have
2259 * a think about bumping its value to force at least one task to be
2262 if (*imbalance < busiest_load_per_task) {
2263 unsigned long pwr_now, pwr_move;
2268 pwr_move = pwr_now = 0;
2270 if (this_nr_running) {
2271 this_load_per_task /= this_nr_running;
2272 if (busiest_load_per_task > this_load_per_task)
2275 this_load_per_task = SCHED_LOAD_SCALE;
2277 if (max_load - this_load >= busiest_load_per_task * imbn) {
2278 *imbalance = busiest_load_per_task;
2283 * OK, we don't have enough imbalance to justify moving tasks,
2284 * however we may be able to increase total CPU power used by
2288 pwr_now += busiest->cpu_power *
2289 min(busiest_load_per_task, max_load);
2290 pwr_now += this->cpu_power *
2291 min(this_load_per_task, this_load);
2292 pwr_now /= SCHED_LOAD_SCALE;
2294 /* Amount of load we'd subtract */
2295 tmp = busiest_load_per_task*SCHED_LOAD_SCALE/busiest->cpu_power;
2297 pwr_move += busiest->cpu_power *
2298 min(busiest_load_per_task, max_load - tmp);
2300 /* Amount of load we'd add */
2301 if (max_load*busiest->cpu_power <
2302 busiest_load_per_task*SCHED_LOAD_SCALE)
2303 tmp = max_load*busiest->cpu_power/this->cpu_power;
2305 tmp = busiest_load_per_task*SCHED_LOAD_SCALE/this->cpu_power;
2306 pwr_move += this->cpu_power*min(this_load_per_task, this_load + tmp);
2307 pwr_move /= SCHED_LOAD_SCALE;
2309 /* Move if we gain throughput */
2310 if (pwr_move <= pwr_now)
2313 *imbalance = busiest_load_per_task;
2319 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2320 if (idle == NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2323 if (this == group_leader && group_leader != group_min) {
2324 *imbalance = min_load_per_task;
2334 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2336 static runqueue_t *find_busiest_queue(struct sched_group *group,
2337 enum idle_type idle, unsigned long imbalance)
2339 unsigned long max_load = 0;
2340 runqueue_t *busiest = NULL, *rqi;
2343 for_each_cpu_mask(i, group->cpumask) {
2346 if (rqi->nr_running == 1 && rqi->raw_weighted_load > imbalance)
2349 if (rqi->raw_weighted_load > max_load) {
2350 max_load = rqi->raw_weighted_load;
2359 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2360 * so long as it is large enough.
2362 #define MAX_PINNED_INTERVAL 512
2364 #define minus_1_or_zero(n) ((n) > 0 ? (n) - 1 : 0)
2366 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2367 * tasks if there is an imbalance.
2369 * Called with this_rq unlocked.
2371 static int load_balance(int this_cpu, runqueue_t *this_rq,
2372 struct sched_domain *sd, enum idle_type idle)
2374 struct sched_group *group;
2375 runqueue_t *busiest;
2376 unsigned long imbalance;
2377 int nr_moved, all_pinned = 0;
2378 int active_balance = 0;
2381 if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2382 !sched_smt_power_savings)
2385 schedstat_inc(sd, lb_cnt[idle]);
2387 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle);
2389 schedstat_inc(sd, lb_nobusyg[idle]);
2393 busiest = find_busiest_queue(group, idle, imbalance);
2395 schedstat_inc(sd, lb_nobusyq[idle]);
2399 BUG_ON(busiest == this_rq);
2401 schedstat_add(sd, lb_imbalance[idle], imbalance);
2404 if (busiest->nr_running > 1) {
2406 * Attempt to move tasks. If find_busiest_group has found
2407 * an imbalance but busiest->nr_running <= 1, the group is
2408 * still unbalanced. nr_moved simply stays zero, so it is
2409 * correctly treated as an imbalance.
2411 double_rq_lock(this_rq, busiest);
2412 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2413 minus_1_or_zero(busiest->nr_running),
2414 imbalance, sd, idle, &all_pinned);
2415 double_rq_unlock(this_rq, busiest);
2417 /* All tasks on this runqueue were pinned by CPU affinity */
2418 if (unlikely(all_pinned))
2423 schedstat_inc(sd, lb_failed[idle]);
2424 sd->nr_balance_failed++;
2426 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2428 spin_lock(&busiest->lock);
2430 /* don't kick the migration_thread, if the curr
2431 * task on busiest cpu can't be moved to this_cpu
2433 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2434 spin_unlock(&busiest->lock);
2436 goto out_one_pinned;
2439 if (!busiest->active_balance) {
2440 busiest->active_balance = 1;
2441 busiest->push_cpu = this_cpu;
2444 spin_unlock(&busiest->lock);
2446 wake_up_process(busiest->migration_thread);
2449 * We've kicked active balancing, reset the failure
2452 sd->nr_balance_failed = sd->cache_nice_tries+1;
2455 sd->nr_balance_failed = 0;
2457 if (likely(!active_balance)) {
2458 /* We were unbalanced, so reset the balancing interval */
2459 sd->balance_interval = sd->min_interval;
2462 * If we've begun active balancing, start to back off. This
2463 * case may not be covered by the all_pinned logic if there
2464 * is only 1 task on the busy runqueue (because we don't call
2467 if (sd->balance_interval < sd->max_interval)
2468 sd->balance_interval *= 2;
2471 if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2472 !sched_smt_power_savings)
2477 schedstat_inc(sd, lb_balanced[idle]);
2479 sd->nr_balance_failed = 0;
2482 /* tune up the balancing interval */
2483 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2484 (sd->balance_interval < sd->max_interval))
2485 sd->balance_interval *= 2;
2487 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && !sched_smt_power_savings)
2493 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2494 * tasks if there is an imbalance.
2496 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2497 * this_rq is locked.
2499 static int load_balance_newidle(int this_cpu, runqueue_t *this_rq,
2500 struct sched_domain *sd)
2502 struct sched_group *group;
2503 runqueue_t *busiest = NULL;
2504 unsigned long imbalance;
2508 if (sd->flags & SD_SHARE_CPUPOWER && !sched_smt_power_savings)
2511 schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
2512 group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE, &sd_idle);
2514 schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]);
2518 busiest = find_busiest_queue(group, NEWLY_IDLE, imbalance);
2520 schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]);
2524 BUG_ON(busiest == this_rq);
2526 schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
2529 if (busiest->nr_running > 1) {
2530 /* Attempt to move tasks */
2531 double_lock_balance(this_rq, busiest);
2532 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2533 minus_1_or_zero(busiest->nr_running),
2534 imbalance, sd, NEWLY_IDLE, NULL);
2535 spin_unlock(&busiest->lock);
2539 schedstat_inc(sd, lb_failed[NEWLY_IDLE]);
2540 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2543 sd->nr_balance_failed = 0;
2548 schedstat_inc(sd, lb_balanced[NEWLY_IDLE]);
2549 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && !sched_smt_power_savings)
2551 sd->nr_balance_failed = 0;
2556 * idle_balance is called by schedule() if this_cpu is about to become
2557 * idle. Attempts to pull tasks from other CPUs.
2559 static void idle_balance(int this_cpu, runqueue_t *this_rq)
2561 struct sched_domain *sd;
2563 for_each_domain(this_cpu, sd) {
2564 if (sd->flags & SD_BALANCE_NEWIDLE) {
2565 if (load_balance_newidle(this_cpu, this_rq, sd)) {
2566 /* We've pulled tasks over so stop searching */
2574 * active_load_balance is run by migration threads. It pushes running tasks
2575 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2576 * running on each physical CPU where possible, and avoids physical /
2577 * logical imbalances.
2579 * Called with busiest_rq locked.
2581 static void active_load_balance(runqueue_t *busiest_rq, int busiest_cpu)
2583 struct sched_domain *sd;
2584 runqueue_t *target_rq;
2585 int target_cpu = busiest_rq->push_cpu;
2587 if (busiest_rq->nr_running <= 1)
2588 /* no task to move */
2591 target_rq = cpu_rq(target_cpu);
2594 * This condition is "impossible", if it occurs
2595 * we need to fix it. Originally reported by
2596 * Bjorn Helgaas on a 128-cpu setup.
2598 BUG_ON(busiest_rq == target_rq);
2600 /* move a task from busiest_rq to target_rq */
2601 double_lock_balance(busiest_rq, target_rq);
2603 /* Search for an sd spanning us and the target CPU. */
2604 for_each_domain(target_cpu, sd) {
2605 if ((sd->flags & SD_LOAD_BALANCE) &&
2606 cpu_isset(busiest_cpu, sd->span))
2610 if (unlikely(sd == NULL))
2613 schedstat_inc(sd, alb_cnt);
2615 if (move_tasks(target_rq, target_cpu, busiest_rq, 1,
2616 RTPRIO_TO_LOAD_WEIGHT(100), sd, SCHED_IDLE, NULL))
2617 schedstat_inc(sd, alb_pushed);
2619 schedstat_inc(sd, alb_failed);
2621 spin_unlock(&target_rq->lock);
2625 * rebalance_tick will get called every timer tick, on every CPU.
2627 * It checks each scheduling domain to see if it is due to be balanced,
2628 * and initiates a balancing operation if so.
2630 * Balancing parameters are set up in arch_init_sched_domains.
2633 /* Don't have all balancing operations going off at once */
2634 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2636 static void rebalance_tick(int this_cpu, runqueue_t *this_rq,
2637 enum idle_type idle)
2639 unsigned long old_load, this_load;
2640 unsigned long j = jiffies + CPU_OFFSET(this_cpu);
2641 struct sched_domain *sd;
2644 this_load = this_rq->raw_weighted_load;
2645 /* Update our load */
2646 for (i = 0; i < 3; i++) {
2647 unsigned long new_load = this_load;
2649 old_load = this_rq->cpu_load[i];
2651 * Round up the averaging division if load is increasing. This
2652 * prevents us from getting stuck on 9 if the load is 10, for
2655 if (new_load > old_load)
2656 new_load += scale-1;
2657 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale;
2660 for_each_domain(this_cpu, sd) {
2661 unsigned long interval;
2663 if (!(sd->flags & SD_LOAD_BALANCE))
2666 interval = sd->balance_interval;
2667 if (idle != SCHED_IDLE)
2668 interval *= sd->busy_factor;
2670 /* scale ms to jiffies */
2671 interval = msecs_to_jiffies(interval);
2672 if (unlikely(!interval))
2675 if (j - sd->last_balance >= interval) {
2676 if (load_balance(this_cpu, this_rq, sd, idle)) {
2678 * We've pulled tasks over so either we're no
2679 * longer idle, or one of our SMT siblings is
2684 sd->last_balance += interval;
2690 * on UP we do not need to balance between CPUs:
2692 static inline void rebalance_tick(int cpu, runqueue_t *rq, enum idle_type idle)
2695 static inline void idle_balance(int cpu, runqueue_t *rq)
2700 static inline int wake_priority_sleeper(runqueue_t *rq)
2703 #ifdef CONFIG_SCHED_SMT
2704 spin_lock(&rq->lock);
2706 * If an SMT sibling task has been put to sleep for priority
2707 * reasons reschedule the idle task to see if it can now run.
2709 if (rq->nr_running) {
2710 resched_task(rq->idle);
2713 spin_unlock(&rq->lock);
2718 DEFINE_PER_CPU(struct kernel_stat, kstat);
2720 EXPORT_PER_CPU_SYMBOL(kstat);
2723 * This is called on clock ticks and on context switches.
2724 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2726 static inline void update_cpu_clock(task_t *p, runqueue_t *rq,
2727 unsigned long long now)
2729 unsigned long long last = max(p->timestamp, rq->timestamp_last_tick);
2730 p->sched_time += now - last;
2734 * Return current->sched_time plus any more ns on the sched_clock
2735 * that have not yet been banked.
2737 unsigned long long current_sched_time(const task_t *tsk)
2739 unsigned long long ns;
2740 unsigned long flags;
2741 local_irq_save(flags);
2742 ns = max(tsk->timestamp, task_rq(tsk)->timestamp_last_tick);
2743 ns = tsk->sched_time + (sched_clock() - ns);
2744 local_irq_restore(flags);
2749 * We place interactive tasks back into the active array, if possible.
2751 * To guarantee that this does not starve expired tasks we ignore the
2752 * interactivity of a task if the first expired task had to wait more
2753 * than a 'reasonable' amount of time. This deadline timeout is
2754 * load-dependent, as the frequency of array switched decreases with
2755 * increasing number of running tasks. We also ignore the interactivity
2756 * if a better static_prio task has expired:
2758 #define EXPIRED_STARVING(rq) \
2759 ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2760 (jiffies - (rq)->expired_timestamp >= \
2761 STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2762 ((rq)->curr->static_prio > (rq)->best_expired_prio))
2765 * Account user cpu time to a process.
2766 * @p: the process that the cpu time gets accounted to
2767 * @hardirq_offset: the offset to subtract from hardirq_count()
2768 * @cputime: the cpu time spent in user space since the last update
2770 void account_user_time(struct task_struct *p, cputime_t cputime)
2772 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2775 p->utime = cputime_add(p->utime, cputime);
2777 /* Add user time to cpustat. */
2778 tmp = cputime_to_cputime64(cputime);
2779 if (TASK_NICE(p) > 0)
2780 cpustat->nice = cputime64_add(cpustat->nice, tmp);
2782 cpustat->user = cputime64_add(cpustat->user, tmp);
2786 * Account system cpu time to a process.
2787 * @p: the process that the cpu time gets accounted to
2788 * @hardirq_offset: the offset to subtract from hardirq_count()
2789 * @cputime: the cpu time spent in kernel space since the last update
2791 void account_system_time(struct task_struct *p, int hardirq_offset,
2794 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2795 runqueue_t *rq = this_rq();
2798 p->stime = cputime_add(p->stime, cputime);
2800 /* Add system time to cpustat. */
2801 tmp = cputime_to_cputime64(cputime);
2802 if (hardirq_count() - hardirq_offset)
2803 cpustat->irq = cputime64_add(cpustat->irq, tmp);
2804 else if (softirq_count())
2805 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
2806 else if (p != rq->idle)
2807 cpustat->system = cputime64_add(cpustat->system, tmp);
2808 else if (atomic_read(&rq->nr_iowait) > 0)
2809 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2811 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2812 /* Account for system time used */
2813 acct_update_integrals(p);
2817 * Account for involuntary wait time.
2818 * @p: the process from which the cpu time has been stolen
2819 * @steal: the cpu time spent in involuntary wait
2821 void account_steal_time(struct task_struct *p, cputime_t steal)
2823 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2824 cputime64_t tmp = cputime_to_cputime64(steal);
2825 runqueue_t *rq = this_rq();
2827 if (p == rq->idle) {
2828 p->stime = cputime_add(p->stime, steal);
2829 if (atomic_read(&rq->nr_iowait) > 0)
2830 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2832 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2834 cpustat->steal = cputime64_add(cpustat->steal, tmp);
2838 * This function gets called by the timer code, with HZ frequency.
2839 * We call it with interrupts disabled.
2841 * It also gets called by the fork code, when changing the parent's
2844 void scheduler_tick(void)
2846 int cpu = smp_processor_id();
2847 runqueue_t *rq = this_rq();
2848 task_t *p = current;
2849 unsigned long long now = sched_clock();
2851 update_cpu_clock(p, rq, now);
2853 rq->timestamp_last_tick = now;
2855 if (p == rq->idle) {
2856 if (wake_priority_sleeper(rq))
2858 rebalance_tick(cpu, rq, SCHED_IDLE);
2862 /* Task might have expired already, but not scheduled off yet */
2863 if (p->array != rq->active) {
2864 set_tsk_need_resched(p);
2867 spin_lock(&rq->lock);
2869 * The task was running during this tick - update the
2870 * time slice counter. Note: we do not update a thread's
2871 * priority until it either goes to sleep or uses up its
2872 * timeslice. This makes it possible for interactive tasks
2873 * to use up their timeslices at their highest priority levels.
2877 * RR tasks need a special form of timeslice management.
2878 * FIFO tasks have no timeslices.
2880 if ((p->policy == SCHED_RR) && !--p->time_slice) {
2881 p->time_slice = task_timeslice(p);
2882 p->first_time_slice = 0;
2883 set_tsk_need_resched(p);
2885 /* put it at the end of the queue: */
2886 requeue_task(p, rq->active);
2890 if (!--p->time_slice) {
2891 dequeue_task(p, rq->active);
2892 set_tsk_need_resched(p);
2893 p->prio = effective_prio(p);
2894 p->time_slice = task_timeslice(p);
2895 p->first_time_slice = 0;
2897 if (!rq->expired_timestamp)
2898 rq->expired_timestamp = jiffies;
2899 if (!TASK_INTERACTIVE(p) || EXPIRED_STARVING(rq)) {
2900 enqueue_task(p, rq->expired);
2901 if (p->static_prio < rq->best_expired_prio)
2902 rq->best_expired_prio = p->static_prio;
2904 enqueue_task(p, rq->active);
2907 * Prevent a too long timeslice allowing a task to monopolize
2908 * the CPU. We do this by splitting up the timeslice into
2911 * Note: this does not mean the task's timeslices expire or
2912 * get lost in any way, they just might be preempted by
2913 * another task of equal priority. (one with higher
2914 * priority would have preempted this task already.) We
2915 * requeue this task to the end of the list on this priority
2916 * level, which is in essence a round-robin of tasks with
2919 * This only applies to tasks in the interactive
2920 * delta range with at least TIMESLICE_GRANULARITY to requeue.
2922 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
2923 p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
2924 (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
2925 (p->array == rq->active)) {
2927 requeue_task(p, rq->active);
2928 set_tsk_need_resched(p);
2932 spin_unlock(&rq->lock);
2934 rebalance_tick(cpu, rq, NOT_IDLE);
2937 #ifdef CONFIG_SCHED_SMT
2938 static inline void wakeup_busy_runqueue(runqueue_t *rq)
2940 /* If an SMT runqueue is sleeping due to priority reasons wake it up */
2941 if (rq->curr == rq->idle && rq->nr_running)
2942 resched_task(rq->idle);
2946 * Called with interrupt disabled and this_rq's runqueue locked.
2948 static void wake_sleeping_dependent(int this_cpu)
2950 struct sched_domain *tmp, *sd = NULL;
2953 for_each_domain(this_cpu, tmp) {
2954 if (tmp->flags & SD_SHARE_CPUPOWER) {
2963 for_each_cpu_mask(i, sd->span) {
2964 runqueue_t *smt_rq = cpu_rq(i);
2968 if (unlikely(!spin_trylock(&smt_rq->lock)))
2971 wakeup_busy_runqueue(smt_rq);
2972 spin_unlock(&smt_rq->lock);
2977 * number of 'lost' timeslices this task wont be able to fully
2978 * utilize, if another task runs on a sibling. This models the
2979 * slowdown effect of other tasks running on siblings:
2981 static inline unsigned long smt_slice(task_t *p, struct sched_domain *sd)
2983 return p->time_slice * (100 - sd->per_cpu_gain) / 100;
2987 * To minimise lock contention and not have to drop this_rq's runlock we only
2988 * trylock the sibling runqueues and bypass those runqueues if we fail to
2989 * acquire their lock. As we only trylock the normal locking order does not
2990 * need to be obeyed.
2992 static int dependent_sleeper(int this_cpu, runqueue_t *this_rq, task_t *p)
2994 struct sched_domain *tmp, *sd = NULL;
2997 /* kernel/rt threads do not participate in dependent sleeping */
2998 if (!p->mm || rt_task(p))
3001 for_each_domain(this_cpu, tmp) {
3002 if (tmp->flags & SD_SHARE_CPUPOWER) {
3011 for_each_cpu_mask(i, sd->span) {
3019 if (unlikely(!spin_trylock(&smt_rq->lock)))
3022 smt_curr = smt_rq->curr;
3028 * If a user task with lower static priority than the
3029 * running task on the SMT sibling is trying to schedule,
3030 * delay it till there is proportionately less timeslice
3031 * left of the sibling task to prevent a lower priority
3032 * task from using an unfair proportion of the
3033 * physical cpu's resources. -ck
3035 if (rt_task(smt_curr)) {
3037 * With real time tasks we run non-rt tasks only
3038 * per_cpu_gain% of the time.
3040 if ((jiffies % DEF_TIMESLICE) >
3041 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
3044 if (smt_curr->static_prio < p->static_prio &&
3045 !TASK_PREEMPTS_CURR(p, smt_rq) &&
3046 smt_slice(smt_curr, sd) > task_timeslice(p))
3050 spin_unlock(&smt_rq->lock);
3055 static inline void wake_sleeping_dependent(int this_cpu)
3059 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq,
3066 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3068 void fastcall add_preempt_count(int val)
3073 BUG_ON((preempt_count() < 0));
3074 preempt_count() += val;
3076 * Spinlock count overflowing soon?
3078 BUG_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10);
3080 EXPORT_SYMBOL(add_preempt_count);
3082 void fastcall sub_preempt_count(int val)
3087 BUG_ON(val > preempt_count());
3089 * Is the spinlock portion underflowing?
3091 BUG_ON((val < PREEMPT_MASK) && !(preempt_count() & PREEMPT_MASK));
3092 preempt_count() -= val;
3094 EXPORT_SYMBOL(sub_preempt_count);
3098 static inline int interactive_sleep(enum sleep_type sleep_type)
3100 return (sleep_type == SLEEP_INTERACTIVE ||
3101 sleep_type == SLEEP_INTERRUPTED);
3105 * schedule() is the main scheduler function.
3107 asmlinkage void __sched schedule(void)
3110 task_t *prev, *next;
3112 prio_array_t *array;
3113 struct list_head *queue;
3114 unsigned long long now;
3115 unsigned long run_time;
3116 int cpu, idx, new_prio;
3119 * Test if we are atomic. Since do_exit() needs to call into
3120 * schedule() atomically, we ignore that path for now.
3121 * Otherwise, whine if we are scheduling when we should not be.
3123 if (unlikely(in_atomic() && !current->exit_state)) {
3124 printk(KERN_ERR "BUG: scheduling while atomic: "
3126 current->comm, preempt_count(), current->pid);
3129 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3134 release_kernel_lock(prev);
3135 need_resched_nonpreemptible:
3139 * The idle thread is not allowed to schedule!
3140 * Remove this check after it has been exercised a bit.
3142 if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
3143 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
3147 schedstat_inc(rq, sched_cnt);
3148 now = sched_clock();
3149 if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
3150 run_time = now - prev->timestamp;
3151 if (unlikely((long long)(now - prev->timestamp) < 0))
3154 run_time = NS_MAX_SLEEP_AVG;
3157 * Tasks charged proportionately less run_time at high sleep_avg to
3158 * delay them losing their interactive status
3160 run_time /= (CURRENT_BONUS(prev) ? : 1);
3162 spin_lock_irq(&rq->lock);
3164 if (unlikely(prev->flags & PF_DEAD))
3165 prev->state = EXIT_DEAD;
3167 switch_count = &prev->nivcsw;
3168 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3169 switch_count = &prev->nvcsw;
3170 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3171 unlikely(signal_pending(prev))))
3172 prev->state = TASK_RUNNING;
3174 if (prev->state == TASK_UNINTERRUPTIBLE)
3175 rq->nr_uninterruptible++;
3176 deactivate_task(prev, rq);
3180 cpu = smp_processor_id();
3181 if (unlikely(!rq->nr_running)) {
3182 idle_balance(cpu, rq);
3183 if (!rq->nr_running) {
3185 rq->expired_timestamp = 0;
3186 wake_sleeping_dependent(cpu);
3192 if (unlikely(!array->nr_active)) {
3194 * Switch the active and expired arrays.
3196 schedstat_inc(rq, sched_switch);
3197 rq->active = rq->expired;
3198 rq->expired = array;
3200 rq->expired_timestamp = 0;
3201 rq->best_expired_prio = MAX_PRIO;
3204 idx = sched_find_first_bit(array->bitmap);
3205 queue = array->queue + idx;
3206 next = list_entry(queue->next, task_t, run_list);
3208 if (!rt_task(next) && interactive_sleep(next->sleep_type)) {
3209 unsigned long long delta = now - next->timestamp;
3210 if (unlikely((long long)(now - next->timestamp) < 0))
3213 if (next->sleep_type == SLEEP_INTERACTIVE)
3214 delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
3216 array = next->array;
3217 new_prio = recalc_task_prio(next, next->timestamp + delta);
3219 if (unlikely(next->prio != new_prio)) {
3220 dequeue_task(next, array);
3221 next->prio = new_prio;
3222 enqueue_task(next, array);
3225 next->sleep_type = SLEEP_NORMAL;
3226 if (dependent_sleeper(cpu, rq, next))
3229 if (next == rq->idle)
3230 schedstat_inc(rq, sched_goidle);
3232 prefetch_stack(next);
3233 clear_tsk_need_resched(prev);
3234 rcu_qsctr_inc(task_cpu(prev));
3236 update_cpu_clock(prev, rq, now);
3238 prev->sleep_avg -= run_time;
3239 if ((long)prev->sleep_avg <= 0)
3240 prev->sleep_avg = 0;
3241 prev->timestamp = prev->last_ran = now;
3243 sched_info_switch(prev, next);
3244 if (likely(prev != next)) {
3245 next->timestamp = now;
3250 prepare_task_switch(rq, next);
3251 prev = context_switch(rq, prev, next);
3254 * this_rq must be evaluated again because prev may have moved
3255 * CPUs since it called schedule(), thus the 'rq' on its stack
3256 * frame will be invalid.
3258 finish_task_switch(this_rq(), prev);
3260 spin_unlock_irq(&rq->lock);
3263 if (unlikely(reacquire_kernel_lock(prev) < 0))
3264 goto need_resched_nonpreemptible;
3265 preempt_enable_no_resched();
3266 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3270 EXPORT_SYMBOL(schedule);
3272 #ifdef CONFIG_PREEMPT
3274 * this is is the entry point to schedule() from in-kernel preemption
3275 * off of preempt_enable. Kernel preemptions off return from interrupt
3276 * occur there and call schedule directly.
3278 asmlinkage void __sched preempt_schedule(void)
3280 struct thread_info *ti = current_thread_info();
3281 #ifdef CONFIG_PREEMPT_BKL
3282 struct task_struct *task = current;
3283 int saved_lock_depth;
3286 * If there is a non-zero preempt_count or interrupts are disabled,
3287 * we do not want to preempt the current task. Just return..
3289 if (unlikely(ti->preempt_count || irqs_disabled()))
3293 add_preempt_count(PREEMPT_ACTIVE);
3295 * We keep the big kernel semaphore locked, but we
3296 * clear ->lock_depth so that schedule() doesnt
3297 * auto-release the semaphore:
3299 #ifdef CONFIG_PREEMPT_BKL
3300 saved_lock_depth = task->lock_depth;
3301 task->lock_depth = -1;
3304 #ifdef CONFIG_PREEMPT_BKL
3305 task->lock_depth = saved_lock_depth;
3307 sub_preempt_count(PREEMPT_ACTIVE);
3309 /* we could miss a preemption opportunity between schedule and now */
3311 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3315 EXPORT_SYMBOL(preempt_schedule);
3318 * this is is the entry point to schedule() from kernel preemption
3319 * off of irq context.
3320 * Note, that this is called and return with irqs disabled. This will
3321 * protect us against recursive calling from irq.
3323 asmlinkage void __sched preempt_schedule_irq(void)
3325 struct thread_info *ti = current_thread_info();
3326 #ifdef CONFIG_PREEMPT_BKL
3327 struct task_struct *task = current;
3328 int saved_lock_depth;
3330 /* Catch callers which need to be fixed*/
3331 BUG_ON(ti->preempt_count || !irqs_disabled());
3334 add_preempt_count(PREEMPT_ACTIVE);
3336 * We keep the big kernel semaphore locked, but we
3337 * clear ->lock_depth so that schedule() doesnt
3338 * auto-release the semaphore:
3340 #ifdef CONFIG_PREEMPT_BKL
3341 saved_lock_depth = task->lock_depth;
3342 task->lock_depth = -1;
3346 local_irq_disable();
3347 #ifdef CONFIG_PREEMPT_BKL
3348 task->lock_depth = saved_lock_depth;
3350 sub_preempt_count(PREEMPT_ACTIVE);
3352 /* we could miss a preemption opportunity between schedule and now */
3354 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3358 #endif /* CONFIG_PREEMPT */
3360 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3363 task_t *p = curr->private;
3364 return try_to_wake_up(p, mode, sync);
3367 EXPORT_SYMBOL(default_wake_function);
3370 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3371 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3372 * number) then we wake all the non-exclusive tasks and one exclusive task.
3374 * There are circumstances in which we can try to wake a task which has already
3375 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3376 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3378 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3379 int nr_exclusive, int sync, void *key)
3381 struct list_head *tmp, *next;
3383 list_for_each_safe(tmp, next, &q->task_list) {
3386 curr = list_entry(tmp, wait_queue_t, task_list);
3387 flags = curr->flags;
3388 if (curr->func(curr, mode, sync, key) &&
3389 (flags & WQ_FLAG_EXCLUSIVE) &&
3396 * __wake_up - wake up threads blocked on a waitqueue.
3398 * @mode: which threads
3399 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3400 * @key: is directly passed to the wakeup function
3402 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3403 int nr_exclusive, void *key)
3405 unsigned long flags;
3407 spin_lock_irqsave(&q->lock, flags);
3408 __wake_up_common(q, mode, nr_exclusive, 0, key);
3409 spin_unlock_irqrestore(&q->lock, flags);
3412 EXPORT_SYMBOL(__wake_up);
3415 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3417 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3419 __wake_up_common(q, mode, 1, 0, NULL);
3423 * __wake_up_sync - wake up threads blocked on a waitqueue.
3425 * @mode: which threads
3426 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3428 * The sync wakeup differs that the waker knows that it will schedule
3429 * away soon, so while the target thread will be woken up, it will not
3430 * be migrated to another CPU - ie. the two threads are 'synchronized'
3431 * with each other. This can prevent needless bouncing between CPUs.
3433 * On UP it can prevent extra preemption.
3436 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3438 unsigned long flags;
3444 if (unlikely(!nr_exclusive))
3447 spin_lock_irqsave(&q->lock, flags);
3448 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3449 spin_unlock_irqrestore(&q->lock, flags);
3451 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3453 void fastcall complete(struct completion *x)
3455 unsigned long flags;
3457 spin_lock_irqsave(&x->wait.lock, flags);
3459 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3461 spin_unlock_irqrestore(&x->wait.lock, flags);
3463 EXPORT_SYMBOL(complete);
3465 void fastcall complete_all(struct completion *x)
3467 unsigned long flags;
3469 spin_lock_irqsave(&x->wait.lock, flags);
3470 x->done += UINT_MAX/2;
3471 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3473 spin_unlock_irqrestore(&x->wait.lock, flags);
3475 EXPORT_SYMBOL(complete_all);
3477 void fastcall __sched wait_for_completion(struct completion *x)
3480 spin_lock_irq(&x->wait.lock);
3482 DECLARE_WAITQUEUE(wait, current);
3484 wait.flags |= WQ_FLAG_EXCLUSIVE;
3485 __add_wait_queue_tail(&x->wait, &wait);
3487 __set_current_state(TASK_UNINTERRUPTIBLE);
3488 spin_unlock_irq(&x->wait.lock);
3490 spin_lock_irq(&x->wait.lock);
3492 __remove_wait_queue(&x->wait, &wait);
3495 spin_unlock_irq(&x->wait.lock);
3497 EXPORT_SYMBOL(wait_for_completion);
3499 unsigned long fastcall __sched
3500 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3504 spin_lock_irq(&x->wait.lock);
3506 DECLARE_WAITQUEUE(wait, current);
3508 wait.flags |= WQ_FLAG_EXCLUSIVE;
3509 __add_wait_queue_tail(&x->wait, &wait);
3511 __set_current_state(TASK_UNINTERRUPTIBLE);
3512 spin_unlock_irq(&x->wait.lock);
3513 timeout = schedule_timeout(timeout);
3514 spin_lock_irq(&x->wait.lock);
3516 __remove_wait_queue(&x->wait, &wait);
3520 __remove_wait_queue(&x->wait, &wait);
3524 spin_unlock_irq(&x->wait.lock);
3527 EXPORT_SYMBOL(wait_for_completion_timeout);
3529 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3535 spin_lock_irq(&x->wait.lock);
3537 DECLARE_WAITQUEUE(wait, current);
3539 wait.flags |= WQ_FLAG_EXCLUSIVE;
3540 __add_wait_queue_tail(&x->wait, &wait);
3542 if (signal_pending(current)) {
3544 __remove_wait_queue(&x->wait, &wait);
3547 __set_current_state(TASK_INTERRUPTIBLE);
3548 spin_unlock_irq(&x->wait.lock);
3550 spin_lock_irq(&x->wait.lock);
3552 __remove_wait_queue(&x->wait, &wait);
3556 spin_unlock_irq(&x->wait.lock);
3560 EXPORT_SYMBOL(wait_for_completion_interruptible);
3562 unsigned long fastcall __sched
3563 wait_for_completion_interruptible_timeout(struct completion *x,
3564 unsigned long timeout)
3568 spin_lock_irq(&x->wait.lock);
3570 DECLARE_WAITQUEUE(wait, current);
3572 wait.flags |= WQ_FLAG_EXCLUSIVE;
3573 __add_wait_queue_tail(&x->wait, &wait);
3575 if (signal_pending(current)) {
3576 timeout = -ERESTARTSYS;
3577 __remove_wait_queue(&x->wait, &wait);
3580 __set_current_state(TASK_INTERRUPTIBLE);
3581 spin_unlock_irq(&x->wait.lock);
3582 timeout = schedule_timeout(timeout);
3583 spin_lock_irq(&x->wait.lock);
3585 __remove_wait_queue(&x->wait, &wait);
3589 __remove_wait_queue(&x->wait, &wait);
3593 spin_unlock_irq(&x->wait.lock);
3596 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3599 #define SLEEP_ON_VAR \
3600 unsigned long flags; \
3601 wait_queue_t wait; \
3602 init_waitqueue_entry(&wait, current);
3604 #define SLEEP_ON_HEAD \
3605 spin_lock_irqsave(&q->lock,flags); \
3606 __add_wait_queue(q, &wait); \
3607 spin_unlock(&q->lock);
3609 #define SLEEP_ON_TAIL \
3610 spin_lock_irq(&q->lock); \
3611 __remove_wait_queue(q, &wait); \
3612 spin_unlock_irqrestore(&q->lock, flags);
3614 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
3618 current->state = TASK_INTERRUPTIBLE;
3625 EXPORT_SYMBOL(interruptible_sleep_on);
3627 long fastcall __sched
3628 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3632 current->state = TASK_INTERRUPTIBLE;
3635 timeout = schedule_timeout(timeout);
3641 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3643 void fastcall __sched sleep_on(wait_queue_head_t *q)
3647 current->state = TASK_UNINTERRUPTIBLE;
3654 EXPORT_SYMBOL(sleep_on);
3656 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3660 current->state = TASK_UNINTERRUPTIBLE;
3663 timeout = schedule_timeout(timeout);
3669 EXPORT_SYMBOL(sleep_on_timeout);
3671 void set_user_nice(task_t *p, long nice)
3673 unsigned long flags;
3674 prio_array_t *array;
3676 int old_prio, new_prio, delta;
3678 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3681 * We have to be careful, if called from sys_setpriority(),
3682 * the task might be in the middle of scheduling on another CPU.
3684 rq = task_rq_lock(p, &flags);
3686 * The RT priorities are set via sched_setscheduler(), but we still
3687 * allow the 'normal' nice value to be set - but as expected
3688 * it wont have any effect on scheduling until the task is
3689 * not SCHED_NORMAL/SCHED_BATCH:
3692 p->static_prio = NICE_TO_PRIO(nice);
3697 dequeue_task(p, array);
3698 dec_raw_weighted_load(rq, p);
3702 new_prio = NICE_TO_PRIO(nice);
3703 delta = new_prio - old_prio;
3704 p->static_prio = NICE_TO_PRIO(nice);
3709 enqueue_task(p, array);
3710 inc_raw_weighted_load(rq, p);
3712 * If the task increased its priority or is running and
3713 * lowered its priority, then reschedule its CPU:
3715 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3716 resched_task(rq->curr);
3719 task_rq_unlock(rq, &flags);
3722 EXPORT_SYMBOL(set_user_nice);
3725 * can_nice - check if a task can reduce its nice value
3729 int can_nice(const task_t *p, const int nice)
3731 /* convert nice value [19,-20] to rlimit style value [1,40] */
3732 int nice_rlim = 20 - nice;
3733 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
3734 capable(CAP_SYS_NICE));
3737 #ifdef __ARCH_WANT_SYS_NICE
3740 * sys_nice - change the priority of the current process.
3741 * @increment: priority increment
3743 * sys_setpriority is a more generic, but much slower function that
3744 * does similar things.
3746 asmlinkage long sys_nice(int increment)
3752 * Setpriority might change our priority at the same moment.
3753 * We don't have to worry. Conceptually one call occurs first
3754 * and we have a single winner.
3756 if (increment < -40)
3761 nice = PRIO_TO_NICE(current->static_prio) + increment;
3767 if (increment < 0 && !can_nice(current, nice))
3770 retval = security_task_setnice(current, nice);
3774 set_user_nice(current, nice);
3781 * task_prio - return the priority value of a given task.
3782 * @p: the task in question.
3784 * This is the priority value as seen by users in /proc.
3785 * RT tasks are offset by -200. Normal tasks are centered
3786 * around 0, value goes from -16 to +15.
3788 int task_prio(const task_t *p)
3790 return p->prio - MAX_RT_PRIO;
3794 * task_nice - return the nice value of a given task.
3795 * @p: the task in question.
3797 int task_nice(const task_t *p)
3799 return TASK_NICE(p);
3801 EXPORT_SYMBOL_GPL(task_nice);
3804 * idle_cpu - is a given cpu idle currently?
3805 * @cpu: the processor in question.
3807 int idle_cpu(int cpu)
3809 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
3813 * idle_task - return the idle task for a given cpu.
3814 * @cpu: the processor in question.
3816 task_t *idle_task(int cpu)
3818 return cpu_rq(cpu)->idle;
3822 * find_process_by_pid - find a process with a matching PID value.
3823 * @pid: the pid in question.
3825 static inline task_t *find_process_by_pid(pid_t pid)
3827 return pid ? find_task_by_pid(pid) : current;
3830 /* Actually do priority change: must hold rq lock. */
3831 static void __setscheduler(struct task_struct *p, int policy, int prio)
3835 p->rt_priority = prio;
3836 if (policy != SCHED_NORMAL && policy != SCHED_BATCH) {
3837 p->prio = MAX_RT_PRIO-1 - p->rt_priority;
3839 p->prio = p->static_prio;
3841 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
3843 if (policy == SCHED_BATCH)
3850 * sched_setscheduler - change the scheduling policy and/or RT priority of
3852 * @p: the task in question.
3853 * @policy: new policy.
3854 * @param: structure containing the new RT priority.
3856 int sched_setscheduler(struct task_struct *p, int policy,
3857 struct sched_param *param)
3860 int oldprio, oldpolicy = -1;
3861 prio_array_t *array;
3862 unsigned long flags;
3866 /* double check policy once rq lock held */
3868 policy = oldpolicy = p->policy;
3869 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
3870 policy != SCHED_NORMAL && policy != SCHED_BATCH)
3873 * Valid priorities for SCHED_FIFO and SCHED_RR are
3874 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
3877 if (param->sched_priority < 0 ||
3878 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3879 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3881 if ((policy == SCHED_NORMAL || policy == SCHED_BATCH)
3882 != (param->sched_priority == 0))
3886 * Allow unprivileged RT tasks to decrease priority:
3888 if (!capable(CAP_SYS_NICE)) {
3890 * can't change policy, except between SCHED_NORMAL
3893 if (((policy != SCHED_NORMAL && p->policy != SCHED_BATCH) &&
3894 (policy != SCHED_BATCH && p->policy != SCHED_NORMAL)) &&
3895 !p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
3897 /* can't increase priority */
3898 if ((policy != SCHED_NORMAL && policy != SCHED_BATCH) &&
3899 param->sched_priority > p->rt_priority &&
3900 param->sched_priority >
3901 p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
3903 /* can't change other user's priorities */
3904 if ((current->euid != p->euid) &&
3905 (current->euid != p->uid))
3909 retval = security_task_setscheduler(p, policy, param);
3913 * To be able to change p->policy safely, the apropriate
3914 * runqueue lock must be held.
3916 rq = task_rq_lock(p, &flags);
3917 /* recheck policy now with rq lock held */
3918 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3919 policy = oldpolicy = -1;
3920 task_rq_unlock(rq, &flags);
3925 deactivate_task(p, rq);
3927 __setscheduler(p, policy, param->sched_priority);
3929 __activate_task(p, rq);
3931 * Reschedule if we are currently running on this runqueue and
3932 * our priority decreased, or if we are not currently running on
3933 * this runqueue and our priority is higher than the current's
3935 if (task_running(rq, p)) {
3936 if (p->prio > oldprio)
3937 resched_task(rq->curr);
3938 } else if (TASK_PREEMPTS_CURR(p, rq))
3939 resched_task(rq->curr);
3941 task_rq_unlock(rq, &flags);
3944 EXPORT_SYMBOL_GPL(sched_setscheduler);
3947 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3950 struct sched_param lparam;
3951 struct task_struct *p;
3953 if (!param || pid < 0)
3955 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3957 read_lock_irq(&tasklist_lock);
3958 p = find_process_by_pid(pid);
3960 read_unlock_irq(&tasklist_lock);
3963 retval = sched_setscheduler(p, policy, &lparam);
3964 read_unlock_irq(&tasklist_lock);
3969 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3970 * @pid: the pid in question.
3971 * @policy: new policy.
3972 * @param: structure containing the new RT priority.
3974 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
3975 struct sched_param __user *param)
3977 /* negative values for policy are not valid */
3981 return do_sched_setscheduler(pid, policy, param);
3985 * sys_sched_setparam - set/change the RT priority of a thread
3986 * @pid: the pid in question.
3987 * @param: structure containing the new RT priority.
3989 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
3991 return do_sched_setscheduler(pid, -1, param);
3995 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3996 * @pid: the pid in question.
3998 asmlinkage long sys_sched_getscheduler(pid_t pid)
4000 int retval = -EINVAL;
4007 read_lock(&tasklist_lock);
4008 p = find_process_by_pid(pid);
4010 retval = security_task_getscheduler(p);
4014 read_unlock(&tasklist_lock);
4021 * sys_sched_getscheduler - get the RT priority of a thread
4022 * @pid: the pid in question.
4023 * @param: structure containing the RT priority.
4025 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4027 struct sched_param lp;
4028 int retval = -EINVAL;
4031 if (!param || pid < 0)
4034 read_lock(&tasklist_lock);
4035 p = find_process_by_pid(pid);
4040 retval = security_task_getscheduler(p);
4044 lp.sched_priority = p->rt_priority;
4045 read_unlock(&tasklist_lock);
4048 * This one might sleep, we cannot do it with a spinlock held ...
4050 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4056 read_unlock(&tasklist_lock);
4060 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4064 cpumask_t cpus_allowed;
4067 read_lock(&tasklist_lock);
4069 p = find_process_by_pid(pid);
4071 read_unlock(&tasklist_lock);
4072 unlock_cpu_hotplug();
4077 * It is not safe to call set_cpus_allowed with the
4078 * tasklist_lock held. We will bump the task_struct's
4079 * usage count and then drop tasklist_lock.
4082 read_unlock(&tasklist_lock);
4085 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4086 !capable(CAP_SYS_NICE))
4089 retval = security_task_setscheduler(p, 0, NULL);
4093 cpus_allowed = cpuset_cpus_allowed(p);
4094 cpus_and(new_mask, new_mask, cpus_allowed);
4095 retval = set_cpus_allowed(p, new_mask);
4099 unlock_cpu_hotplug();
4103 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4104 cpumask_t *new_mask)
4106 if (len < sizeof(cpumask_t)) {
4107 memset(new_mask, 0, sizeof(cpumask_t));
4108 } else if (len > sizeof(cpumask_t)) {
4109 len = sizeof(cpumask_t);
4111 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4115 * sys_sched_setaffinity - set the cpu affinity of a process
4116 * @pid: pid of the process
4117 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4118 * @user_mask_ptr: user-space pointer to the new cpu mask
4120 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4121 unsigned long __user *user_mask_ptr)
4126 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4130 return sched_setaffinity(pid, new_mask);
4134 * Represents all cpu's present in the system
4135 * In systems capable of hotplug, this map could dynamically grow
4136 * as new cpu's are detected in the system via any platform specific
4137 * method, such as ACPI for e.g.
4140 cpumask_t cpu_present_map __read_mostly;
4141 EXPORT_SYMBOL(cpu_present_map);
4144 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4145 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4148 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4154 read_lock(&tasklist_lock);
4157 p = find_process_by_pid(pid);
4161 retval = security_task_getscheduler(p);
4165 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4168 read_unlock(&tasklist_lock);
4169 unlock_cpu_hotplug();
4177 * sys_sched_getaffinity - get the cpu affinity of a process
4178 * @pid: pid of the process
4179 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4180 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4182 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4183 unsigned long __user *user_mask_ptr)
4188 if (len < sizeof(cpumask_t))
4191 ret = sched_getaffinity(pid, &mask);
4195 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4198 return sizeof(cpumask_t);
4202 * sys_sched_yield - yield the current processor to other threads.
4204 * this function yields the current CPU by moving the calling thread
4205 * to the expired array. If there are no other threads running on this
4206 * CPU then this function will return.
4208 asmlinkage long sys_sched_yield(void)
4210 runqueue_t *rq = this_rq_lock();
4211 prio_array_t *array = current->array;
4212 prio_array_t *target = rq->expired;
4214 schedstat_inc(rq, yld_cnt);
4216 * We implement yielding by moving the task into the expired
4219 * (special rule: RT tasks will just roundrobin in the active
4222 if (rt_task(current))
4223 target = rq->active;
4225 if (array->nr_active == 1) {
4226 schedstat_inc(rq, yld_act_empty);
4227 if (!rq->expired->nr_active)
4228 schedstat_inc(rq, yld_both_empty);
4229 } else if (!rq->expired->nr_active)
4230 schedstat_inc(rq, yld_exp_empty);
4232 if (array != target) {
4233 dequeue_task(current, array);
4234 enqueue_task(current, target);
4237 * requeue_task is cheaper so perform that if possible.
4239 requeue_task(current, array);
4242 * Since we are going to call schedule() anyway, there's
4243 * no need to preempt or enable interrupts:
4245 __release(rq->lock);
4246 _raw_spin_unlock(&rq->lock);
4247 preempt_enable_no_resched();
4254 static inline void __cond_resched(void)
4256 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4257 __might_sleep(__FILE__, __LINE__);
4260 * The BKS might be reacquired before we have dropped
4261 * PREEMPT_ACTIVE, which could trigger a second
4262 * cond_resched() call.
4264 if (unlikely(preempt_count()))
4266 if (unlikely(system_state != SYSTEM_RUNNING))
4269 add_preempt_count(PREEMPT_ACTIVE);
4271 sub_preempt_count(PREEMPT_ACTIVE);
4272 } while (need_resched());
4275 int __sched cond_resched(void)
4277 if (need_resched()) {
4284 EXPORT_SYMBOL(cond_resched);
4287 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4288 * call schedule, and on return reacquire the lock.
4290 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4291 * operations here to prevent schedule() from being called twice (once via
4292 * spin_unlock(), once by hand).
4294 int cond_resched_lock(spinlock_t *lock)
4298 if (need_lockbreak(lock)) {
4304 if (need_resched()) {
4305 _raw_spin_unlock(lock);
4306 preempt_enable_no_resched();
4314 EXPORT_SYMBOL(cond_resched_lock);
4316 int __sched cond_resched_softirq(void)
4318 BUG_ON(!in_softirq());
4320 if (need_resched()) {
4321 __local_bh_enable();
4329 EXPORT_SYMBOL(cond_resched_softirq);
4333 * yield - yield the current processor to other threads.
4335 * this is a shortcut for kernel-space yielding - it marks the
4336 * thread runnable and calls sys_sched_yield().
4338 void __sched yield(void)
4340 set_current_state(TASK_RUNNING);
4344 EXPORT_SYMBOL(yield);
4347 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4348 * that process accounting knows that this is a task in IO wait state.
4350 * But don't do that if it is a deliberate, throttling IO wait (this task
4351 * has set its backing_dev_info: the queue against which it should throttle)
4353 void __sched io_schedule(void)
4355 struct runqueue *rq = &__raw_get_cpu_var(runqueues);
4357 atomic_inc(&rq->nr_iowait);
4359 atomic_dec(&rq->nr_iowait);
4362 EXPORT_SYMBOL(io_schedule);
4364 long __sched io_schedule_timeout(long timeout)
4366 struct runqueue *rq = &__raw_get_cpu_var(runqueues);
4369 atomic_inc(&rq->nr_iowait);
4370 ret = schedule_timeout(timeout);
4371 atomic_dec(&rq->nr_iowait);
4376 * sys_sched_get_priority_max - return maximum RT priority.
4377 * @policy: scheduling class.
4379 * this syscall returns the maximum rt_priority that can be used
4380 * by a given scheduling class.
4382 asmlinkage long sys_sched_get_priority_max(int policy)
4389 ret = MAX_USER_RT_PRIO-1;
4400 * sys_sched_get_priority_min - return minimum RT priority.
4401 * @policy: scheduling class.
4403 * this syscall returns the minimum rt_priority that can be used
4404 * by a given scheduling class.
4406 asmlinkage long sys_sched_get_priority_min(int policy)
4423 * sys_sched_rr_get_interval - return the default timeslice of a process.
4424 * @pid: pid of the process.
4425 * @interval: userspace pointer to the timeslice value.
4427 * this syscall writes the default timeslice value of a given process
4428 * into the user-space timespec buffer. A value of '0' means infinity.
4431 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4433 int retval = -EINVAL;
4441 read_lock(&tasklist_lock);
4442 p = find_process_by_pid(pid);
4446 retval = security_task_getscheduler(p);
4450 jiffies_to_timespec(p->policy == SCHED_FIFO ?
4451 0 : task_timeslice(p), &t);
4452 read_unlock(&tasklist_lock);
4453 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4457 read_unlock(&tasklist_lock);
4461 static inline struct task_struct *eldest_child(struct task_struct *p)
4463 if (list_empty(&p->children)) return NULL;
4464 return list_entry(p->children.next,struct task_struct,sibling);
4467 static inline struct task_struct *older_sibling(struct task_struct *p)
4469 if (p->sibling.prev==&p->parent->children) return NULL;
4470 return list_entry(p->sibling.prev,struct task_struct,sibling);
4473 static inline struct task_struct *younger_sibling(struct task_struct *p)
4475 if (p->sibling.next==&p->parent->children) return NULL;
4476 return list_entry(p->sibling.next,struct task_struct,sibling);
4479 static void show_task(task_t *p)
4483 unsigned long free = 0;
4484 static const char *stat_nam[] = { "R", "S", "D", "T", "t", "Z", "X" };
4486 printk("%-13.13s ", p->comm);
4487 state = p->state ? __ffs(p->state) + 1 : 0;
4488 if (state < ARRAY_SIZE(stat_nam))
4489 printk(stat_nam[state]);
4492 #if (BITS_PER_LONG == 32)
4493 if (state == TASK_RUNNING)
4494 printk(" running ");
4496 printk(" %08lX ", thread_saved_pc(p));
4498 if (state == TASK_RUNNING)
4499 printk(" running task ");
4501 printk(" %016lx ", thread_saved_pc(p));
4503 #ifdef CONFIG_DEBUG_STACK_USAGE
4505 unsigned long *n = end_of_stack(p);
4508 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4511 printk("%5lu %5d %6d ", free, p->pid, p->parent->pid);
4512 if ((relative = eldest_child(p)))
4513 printk("%5d ", relative->pid);
4516 if ((relative = younger_sibling(p)))
4517 printk("%7d", relative->pid);
4520 if ((relative = older_sibling(p)))
4521 printk(" %5d", relative->pid);
4525 printk(" (L-TLB)\n");
4527 printk(" (NOTLB)\n");
4529 if (state != TASK_RUNNING)
4530 show_stack(p, NULL);
4533 void show_state(void)
4537 #if (BITS_PER_LONG == 32)
4540 printk(" task PC pid father child younger older\n");
4544 printk(" task PC pid father child younger older\n");
4546 read_lock(&tasklist_lock);
4547 do_each_thread(g, p) {
4549 * reset the NMI-timeout, listing all files on a slow
4550 * console might take alot of time:
4552 touch_nmi_watchdog();
4554 } while_each_thread(g, p);
4556 read_unlock(&tasklist_lock);
4557 mutex_debug_show_all_locks();
4561 * init_idle - set up an idle thread for a given CPU
4562 * @idle: task in question
4563 * @cpu: cpu the idle task belongs to
4565 * NOTE: this function does not set the idle thread's NEED_RESCHED
4566 * flag, to make booting more robust.
4568 void __devinit init_idle(task_t *idle, int cpu)
4570 runqueue_t *rq = cpu_rq(cpu);
4571 unsigned long flags;
4573 idle->timestamp = sched_clock();
4574 idle->sleep_avg = 0;
4576 idle->prio = MAX_PRIO;
4577 idle->state = TASK_RUNNING;
4578 idle->cpus_allowed = cpumask_of_cpu(cpu);
4579 set_task_cpu(idle, cpu);
4581 spin_lock_irqsave(&rq->lock, flags);
4582 rq->curr = rq->idle = idle;
4583 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4586 spin_unlock_irqrestore(&rq->lock, flags);
4588 /* Set the preempt count _outside_ the spinlocks! */
4589 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4590 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4592 task_thread_info(idle)->preempt_count = 0;
4597 * In a system that switches off the HZ timer nohz_cpu_mask
4598 * indicates which cpus entered this state. This is used
4599 * in the rcu update to wait only for active cpus. For system
4600 * which do not switch off the HZ timer nohz_cpu_mask should
4601 * always be CPU_MASK_NONE.
4603 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4607 * This is how migration works:
4609 * 1) we queue a migration_req_t structure in the source CPU's
4610 * runqueue and wake up that CPU's migration thread.
4611 * 2) we down() the locked semaphore => thread blocks.
4612 * 3) migration thread wakes up (implicitly it forces the migrated
4613 * thread off the CPU)
4614 * 4) it gets the migration request and checks whether the migrated
4615 * task is still in the wrong runqueue.
4616 * 5) if it's in the wrong runqueue then the migration thread removes
4617 * it and puts it into the right queue.
4618 * 6) migration thread up()s the semaphore.
4619 * 7) we wake up and the migration is done.
4623 * Change a given task's CPU affinity. Migrate the thread to a
4624 * proper CPU and schedule it away if the CPU it's executing on
4625 * is removed from the allowed bitmask.
4627 * NOTE: the caller must have a valid reference to the task, the
4628 * task must not exit() & deallocate itself prematurely. The
4629 * call is not atomic; no spinlocks may be held.
4631 int set_cpus_allowed(task_t *p, cpumask_t new_mask)
4633 unsigned long flags;
4635 migration_req_t req;
4638 rq = task_rq_lock(p, &flags);
4639 if (!cpus_intersects(new_mask, cpu_online_map)) {
4644 p->cpus_allowed = new_mask;
4645 /* Can the task run on the task's current CPU? If so, we're done */
4646 if (cpu_isset(task_cpu(p), new_mask))
4649 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4650 /* Need help from migration thread: drop lock and wait. */
4651 task_rq_unlock(rq, &flags);
4652 wake_up_process(rq->migration_thread);
4653 wait_for_completion(&req.done);
4654 tlb_migrate_finish(p->mm);
4658 task_rq_unlock(rq, &flags);
4662 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4665 * Move (not current) task off this cpu, onto dest cpu. We're doing
4666 * this because either it can't run here any more (set_cpus_allowed()
4667 * away from this CPU, or CPU going down), or because we're
4668 * attempting to rebalance this task on exec (sched_exec).
4670 * So we race with normal scheduler movements, but that's OK, as long
4671 * as the task is no longer on this CPU.
4673 * Returns non-zero if task was successfully migrated.
4675 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4677 runqueue_t *rq_dest, *rq_src;
4680 if (unlikely(cpu_is_offline(dest_cpu)))
4683 rq_src = cpu_rq(src_cpu);
4684 rq_dest = cpu_rq(dest_cpu);
4686 double_rq_lock(rq_src, rq_dest);
4687 /* Already moved. */
4688 if (task_cpu(p) != src_cpu)
4690 /* Affinity changed (again). */
4691 if (!cpu_isset(dest_cpu, p->cpus_allowed))
4694 set_task_cpu(p, dest_cpu);
4697 * Sync timestamp with rq_dest's before activating.
4698 * The same thing could be achieved by doing this step
4699 * afterwards, and pretending it was a local activate.
4700 * This way is cleaner and logically correct.
4702 p->timestamp = p->timestamp - rq_src->timestamp_last_tick
4703 + rq_dest->timestamp_last_tick;
4704 deactivate_task(p, rq_src);
4705 activate_task(p, rq_dest, 0);
4706 if (TASK_PREEMPTS_CURR(p, rq_dest))
4707 resched_task(rq_dest->curr);
4711 double_rq_unlock(rq_src, rq_dest);
4716 * migration_thread - this is a highprio system thread that performs
4717 * thread migration by bumping thread off CPU then 'pushing' onto
4720 static int migration_thread(void *data)
4723 int cpu = (long)data;
4726 BUG_ON(rq->migration_thread != current);
4728 set_current_state(TASK_INTERRUPTIBLE);
4729 while (!kthread_should_stop()) {
4730 struct list_head *head;
4731 migration_req_t *req;
4735 spin_lock_irq(&rq->lock);
4737 if (cpu_is_offline(cpu)) {
4738 spin_unlock_irq(&rq->lock);
4742 if (rq->active_balance) {
4743 active_load_balance(rq, cpu);
4744 rq->active_balance = 0;
4747 head = &rq->migration_queue;
4749 if (list_empty(head)) {
4750 spin_unlock_irq(&rq->lock);
4752 set_current_state(TASK_INTERRUPTIBLE);
4755 req = list_entry(head->next, migration_req_t, list);
4756 list_del_init(head->next);
4758 spin_unlock(&rq->lock);
4759 __migrate_task(req->task, cpu, req->dest_cpu);
4762 complete(&req->done);
4764 __set_current_state(TASK_RUNNING);
4768 /* Wait for kthread_stop */
4769 set_current_state(TASK_INTERRUPTIBLE);
4770 while (!kthread_should_stop()) {
4772 set_current_state(TASK_INTERRUPTIBLE);
4774 __set_current_state(TASK_RUNNING);
4778 #ifdef CONFIG_HOTPLUG_CPU
4779 /* Figure out where task on dead CPU should go, use force if neccessary. */
4780 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *tsk)
4783 unsigned long flags;
4789 mask = node_to_cpumask(cpu_to_node(dead_cpu));
4790 cpus_and(mask, mask, tsk->cpus_allowed);
4791 dest_cpu = any_online_cpu(mask);
4793 /* On any allowed CPU? */
4794 if (dest_cpu == NR_CPUS)
4795 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4797 /* No more Mr. Nice Guy. */
4798 if (dest_cpu == NR_CPUS) {
4799 rq = task_rq_lock(tsk, &flags);
4800 cpus_setall(tsk->cpus_allowed);
4801 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4802 task_rq_unlock(rq, &flags);
4805 * Don't tell them about moving exiting tasks or
4806 * kernel threads (both mm NULL), since they never
4809 if (tsk->mm && printk_ratelimit())
4810 printk(KERN_INFO "process %d (%s) no "
4811 "longer affine to cpu%d\n",
4812 tsk->pid, tsk->comm, dead_cpu);
4814 if (!__migrate_task(tsk, dead_cpu, dest_cpu))
4819 * While a dead CPU has no uninterruptible tasks queued at this point,
4820 * it might still have a nonzero ->nr_uninterruptible counter, because
4821 * for performance reasons the counter is not stricly tracking tasks to
4822 * their home CPUs. So we just add the counter to another CPU's counter,
4823 * to keep the global sum constant after CPU-down:
4825 static void migrate_nr_uninterruptible(runqueue_t *rq_src)
4827 runqueue_t *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
4828 unsigned long flags;
4830 local_irq_save(flags);
4831 double_rq_lock(rq_src, rq_dest);
4832 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
4833 rq_src->nr_uninterruptible = 0;
4834 double_rq_unlock(rq_src, rq_dest);
4835 local_irq_restore(flags);
4838 /* Run through task list and migrate tasks from the dead cpu. */
4839 static void migrate_live_tasks(int src_cpu)
4841 struct task_struct *tsk, *t;
4843 write_lock_irq(&tasklist_lock);
4845 do_each_thread(t, tsk) {
4849 if (task_cpu(tsk) == src_cpu)
4850 move_task_off_dead_cpu(src_cpu, tsk);
4851 } while_each_thread(t, tsk);
4853 write_unlock_irq(&tasklist_lock);
4856 /* Schedules idle task to be the next runnable task on current CPU.
4857 * It does so by boosting its priority to highest possible and adding it to
4858 * the _front_ of runqueue. Used by CPU offline code.
4860 void sched_idle_next(void)
4862 int cpu = smp_processor_id();
4863 runqueue_t *rq = this_rq();
4864 struct task_struct *p = rq->idle;
4865 unsigned long flags;
4867 /* cpu has to be offline */
4868 BUG_ON(cpu_online(cpu));
4870 /* Strictly not necessary since rest of the CPUs are stopped by now
4871 * and interrupts disabled on current cpu.
4873 spin_lock_irqsave(&rq->lock, flags);
4875 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4876 /* Add idle task to _front_ of it's priority queue */
4877 __activate_idle_task(p, rq);
4879 spin_unlock_irqrestore(&rq->lock, flags);
4882 /* Ensures that the idle task is using init_mm right before its cpu goes
4885 void idle_task_exit(void)
4887 struct mm_struct *mm = current->active_mm;
4889 BUG_ON(cpu_online(smp_processor_id()));
4892 switch_mm(mm, &init_mm, current);
4896 static void migrate_dead(unsigned int dead_cpu, task_t *tsk)
4898 struct runqueue *rq = cpu_rq(dead_cpu);
4900 /* Must be exiting, otherwise would be on tasklist. */
4901 BUG_ON(tsk->exit_state != EXIT_ZOMBIE && tsk->exit_state != EXIT_DEAD);
4903 /* Cannot have done final schedule yet: would have vanished. */
4904 BUG_ON(tsk->flags & PF_DEAD);
4906 get_task_struct(tsk);
4909 * Drop lock around migration; if someone else moves it,
4910 * that's OK. No task can be added to this CPU, so iteration is
4913 spin_unlock_irq(&rq->lock);
4914 move_task_off_dead_cpu(dead_cpu, tsk);
4915 spin_lock_irq(&rq->lock);
4917 put_task_struct(tsk);
4920 /* release_task() removes task from tasklist, so we won't find dead tasks. */
4921 static void migrate_dead_tasks(unsigned int dead_cpu)
4924 struct runqueue *rq = cpu_rq(dead_cpu);
4926 for (arr = 0; arr < 2; arr++) {
4927 for (i = 0; i < MAX_PRIO; i++) {
4928 struct list_head *list = &rq->arrays[arr].queue[i];
4929 while (!list_empty(list))
4930 migrate_dead(dead_cpu,
4931 list_entry(list->next, task_t,
4936 #endif /* CONFIG_HOTPLUG_CPU */
4939 * migration_call - callback that gets triggered when a CPU is added.
4940 * Here we can start up the necessary migration thread for the new CPU.
4942 static int __cpuinit migration_call(struct notifier_block *nfb,
4943 unsigned long action,
4946 int cpu = (long)hcpu;
4947 struct task_struct *p;
4948 struct runqueue *rq;
4949 unsigned long flags;
4952 case CPU_UP_PREPARE:
4953 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
4956 p->flags |= PF_NOFREEZE;
4957 kthread_bind(p, cpu);
4958 /* Must be high prio: stop_machine expects to yield to it. */
4959 rq = task_rq_lock(p, &flags);
4960 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4961 task_rq_unlock(rq, &flags);
4962 cpu_rq(cpu)->migration_thread = p;
4965 /* Strictly unneccessary, as first user will wake it. */
4966 wake_up_process(cpu_rq(cpu)->migration_thread);
4968 #ifdef CONFIG_HOTPLUG_CPU
4969 case CPU_UP_CANCELED:
4970 if (!cpu_rq(cpu)->migration_thread)
4972 /* Unbind it from offline cpu so it can run. Fall thru. */
4973 kthread_bind(cpu_rq(cpu)->migration_thread,
4974 any_online_cpu(cpu_online_map));
4975 kthread_stop(cpu_rq(cpu)->migration_thread);
4976 cpu_rq(cpu)->migration_thread = NULL;
4979 migrate_live_tasks(cpu);
4981 kthread_stop(rq->migration_thread);
4982 rq->migration_thread = NULL;
4983 /* Idle task back to normal (off runqueue, low prio) */
4984 rq = task_rq_lock(rq->idle, &flags);
4985 deactivate_task(rq->idle, rq);
4986 rq->idle->static_prio = MAX_PRIO;
4987 __setscheduler(rq->idle, SCHED_NORMAL, 0);
4988 migrate_dead_tasks(cpu);
4989 task_rq_unlock(rq, &flags);
4990 migrate_nr_uninterruptible(rq);
4991 BUG_ON(rq->nr_running != 0);
4993 /* No need to migrate the tasks: it was best-effort if
4994 * they didn't do lock_cpu_hotplug(). Just wake up
4995 * the requestors. */
4996 spin_lock_irq(&rq->lock);
4997 while (!list_empty(&rq->migration_queue)) {
4998 migration_req_t *req;
4999 req = list_entry(rq->migration_queue.next,
5000 migration_req_t, list);
5001 list_del_init(&req->list);
5002 complete(&req->done);
5004 spin_unlock_irq(&rq->lock);
5011 /* Register at highest priority so that task migration (migrate_all_tasks)
5012 * happens before everything else.
5014 static struct notifier_block __cpuinitdata migration_notifier = {
5015 .notifier_call = migration_call,
5019 int __init migration_init(void)
5021 void *cpu = (void *)(long)smp_processor_id();
5022 /* Start one for boot CPU. */
5023 migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5024 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5025 register_cpu_notifier(&migration_notifier);
5031 #undef SCHED_DOMAIN_DEBUG
5032 #ifdef SCHED_DOMAIN_DEBUG
5033 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5038 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5042 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5047 struct sched_group *group = sd->groups;
5048 cpumask_t groupmask;
5050 cpumask_scnprintf(str, NR_CPUS, sd->span);
5051 cpus_clear(groupmask);
5054 for (i = 0; i < level + 1; i++)
5056 printk("domain %d: ", level);
5058 if (!(sd->flags & SD_LOAD_BALANCE)) {
5059 printk("does not load-balance\n");
5061 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent");
5065 printk("span %s\n", str);
5067 if (!cpu_isset(cpu, sd->span))
5068 printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
5069 if (!cpu_isset(cpu, group->cpumask))
5070 printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
5073 for (i = 0; i < level + 2; i++)
5079 printk(KERN_ERR "ERROR: group is NULL\n");
5083 if (!group->cpu_power) {
5085 printk(KERN_ERR "ERROR: domain->cpu_power not set\n");
5088 if (!cpus_weight(group->cpumask)) {
5090 printk(KERN_ERR "ERROR: empty group\n");
5093 if (cpus_intersects(groupmask, group->cpumask)) {
5095 printk(KERN_ERR "ERROR: repeated CPUs\n");
5098 cpus_or(groupmask, groupmask, group->cpumask);
5100 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5103 group = group->next;
5104 } while (group != sd->groups);
5107 if (!cpus_equal(sd->span, groupmask))
5108 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5114 if (!cpus_subset(groupmask, sd->span))
5115 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
5121 #define sched_domain_debug(sd, cpu) {}
5124 static int sd_degenerate(struct sched_domain *sd)
5126 if (cpus_weight(sd->span) == 1)
5129 /* Following flags need at least 2 groups */
5130 if (sd->flags & (SD_LOAD_BALANCE |
5131 SD_BALANCE_NEWIDLE |
5134 if (sd->groups != sd->groups->next)
5138 /* Following flags don't use groups */
5139 if (sd->flags & (SD_WAKE_IDLE |
5147 static int sd_parent_degenerate(struct sched_domain *sd,
5148 struct sched_domain *parent)
5150 unsigned long cflags = sd->flags, pflags = parent->flags;
5152 if (sd_degenerate(parent))
5155 if (!cpus_equal(sd->span, parent->span))
5158 /* Does parent contain flags not in child? */
5159 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5160 if (cflags & SD_WAKE_AFFINE)
5161 pflags &= ~SD_WAKE_BALANCE;
5162 /* Flags needing groups don't count if only 1 group in parent */
5163 if (parent->groups == parent->groups->next) {
5164 pflags &= ~(SD_LOAD_BALANCE |
5165 SD_BALANCE_NEWIDLE |
5169 if (~cflags & pflags)
5176 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5177 * hold the hotplug lock.
5179 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5181 runqueue_t *rq = cpu_rq(cpu);
5182 struct sched_domain *tmp;
5184 /* Remove the sched domains which do not contribute to scheduling. */
5185 for (tmp = sd; tmp; tmp = tmp->parent) {
5186 struct sched_domain *parent = tmp->parent;
5189 if (sd_parent_degenerate(tmp, parent))
5190 tmp->parent = parent->parent;
5193 if (sd && sd_degenerate(sd))
5196 sched_domain_debug(sd, cpu);
5198 rcu_assign_pointer(rq->sd, sd);
5201 /* cpus with isolated domains */
5202 static cpumask_t __devinitdata cpu_isolated_map = CPU_MASK_NONE;
5204 /* Setup the mask of cpus configured for isolated domains */
5205 static int __init isolated_cpu_setup(char *str)
5207 int ints[NR_CPUS], i;
5209 str = get_options(str, ARRAY_SIZE(ints), ints);
5210 cpus_clear(cpu_isolated_map);
5211 for (i = 1; i <= ints[0]; i++)
5212 if (ints[i] < NR_CPUS)
5213 cpu_set(ints[i], cpu_isolated_map);
5217 __setup ("isolcpus=", isolated_cpu_setup);
5220 * init_sched_build_groups takes an array of groups, the cpumask we wish
5221 * to span, and a pointer to a function which identifies what group a CPU
5222 * belongs to. The return value of group_fn must be a valid index into the
5223 * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
5224 * keep track of groups covered with a cpumask_t).
5226 * init_sched_build_groups will build a circular linked list of the groups
5227 * covered by the given span, and will set each group's ->cpumask correctly,
5228 * and ->cpu_power to 0.
5230 static void init_sched_build_groups(struct sched_group groups[], cpumask_t span,
5231 int (*group_fn)(int cpu))
5233 struct sched_group *first = NULL, *last = NULL;
5234 cpumask_t covered = CPU_MASK_NONE;
5237 for_each_cpu_mask(i, span) {
5238 int group = group_fn(i);
5239 struct sched_group *sg = &groups[group];
5242 if (cpu_isset(i, covered))
5245 sg->cpumask = CPU_MASK_NONE;
5248 for_each_cpu_mask(j, span) {
5249 if (group_fn(j) != group)
5252 cpu_set(j, covered);
5253 cpu_set(j, sg->cpumask);
5264 #define SD_NODES_PER_DOMAIN 16
5267 * Self-tuning task migration cost measurement between source and target CPUs.
5269 * This is done by measuring the cost of manipulating buffers of varying
5270 * sizes. For a given buffer-size here are the steps that are taken:
5272 * 1) the source CPU reads+dirties a shared buffer
5273 * 2) the target CPU reads+dirties the same shared buffer
5275 * We measure how long they take, in the following 4 scenarios:
5277 * - source: CPU1, target: CPU2 | cost1
5278 * - source: CPU2, target: CPU1 | cost2
5279 * - source: CPU1, target: CPU1 | cost3
5280 * - source: CPU2, target: CPU2 | cost4
5282 * We then calculate the cost3+cost4-cost1-cost2 difference - this is
5283 * the cost of migration.
5285 * We then start off from a small buffer-size and iterate up to larger
5286 * buffer sizes, in 5% steps - measuring each buffer-size separately, and
5287 * doing a maximum search for the cost. (The maximum cost for a migration
5288 * normally occurs when the working set size is around the effective cache
5291 #define SEARCH_SCOPE 2
5292 #define MIN_CACHE_SIZE (64*1024U)
5293 #define DEFAULT_CACHE_SIZE (5*1024*1024U)
5294 #define ITERATIONS 1
5295 #define SIZE_THRESH 130
5296 #define COST_THRESH 130
5299 * The migration cost is a function of 'domain distance'. Domain
5300 * distance is the number of steps a CPU has to iterate down its
5301 * domain tree to share a domain with the other CPU. The farther
5302 * two CPUs are from each other, the larger the distance gets.
5304 * Note that we use the distance only to cache measurement results,
5305 * the distance value is not used numerically otherwise. When two
5306 * CPUs have the same distance it is assumed that the migration
5307 * cost is the same. (this is a simplification but quite practical)
5309 #define MAX_DOMAIN_DISTANCE 32
5311 static unsigned long long migration_cost[MAX_DOMAIN_DISTANCE] =
5312 { [ 0 ... MAX_DOMAIN_DISTANCE-1 ] =
5314 * Architectures may override the migration cost and thus avoid
5315 * boot-time calibration. Unit is nanoseconds. Mostly useful for
5316 * virtualized hardware:
5318 #ifdef CONFIG_DEFAULT_MIGRATION_COST
5319 CONFIG_DEFAULT_MIGRATION_COST
5326 * Allow override of migration cost - in units of microseconds.
5327 * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost
5328 * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs:
5330 static int __init migration_cost_setup(char *str)
5332 int ints[MAX_DOMAIN_DISTANCE+1], i;
5334 str = get_options(str, ARRAY_SIZE(ints), ints);
5336 printk("#ints: %d\n", ints[0]);
5337 for (i = 1; i <= ints[0]; i++) {
5338 migration_cost[i-1] = (unsigned long long)ints[i]*1000;
5339 printk("migration_cost[%d]: %Ld\n", i-1, migration_cost[i-1]);
5344 __setup ("migration_cost=", migration_cost_setup);
5347 * Global multiplier (divisor) for migration-cutoff values,
5348 * in percentiles. E.g. use a value of 150 to get 1.5 times
5349 * longer cache-hot cutoff times.
5351 * (We scale it from 100 to 128 to long long handling easier.)
5354 #define MIGRATION_FACTOR_SCALE 128
5356 static unsigned int migration_factor = MIGRATION_FACTOR_SCALE;
5358 static int __init setup_migration_factor(char *str)
5360 get_option(&str, &migration_factor);
5361 migration_factor = migration_factor * MIGRATION_FACTOR_SCALE / 100;
5365 __setup("migration_factor=", setup_migration_factor);
5368 * Estimated distance of two CPUs, measured via the number of domains
5369 * we have to pass for the two CPUs to be in the same span:
5371 static unsigned long domain_distance(int cpu1, int cpu2)
5373 unsigned long distance = 0;
5374 struct sched_domain *sd;
5376 for_each_domain(cpu1, sd) {
5377 WARN_ON(!cpu_isset(cpu1, sd->span));
5378 if (cpu_isset(cpu2, sd->span))
5382 if (distance >= MAX_DOMAIN_DISTANCE) {
5384 distance = MAX_DOMAIN_DISTANCE-1;
5390 static unsigned int migration_debug;
5392 static int __init setup_migration_debug(char *str)
5394 get_option(&str, &migration_debug);
5398 __setup("migration_debug=", setup_migration_debug);
5401 * Maximum cache-size that the scheduler should try to measure.
5402 * Architectures with larger caches should tune this up during
5403 * bootup. Gets used in the domain-setup code (i.e. during SMP
5406 unsigned int max_cache_size;
5408 static int __init setup_max_cache_size(char *str)
5410 get_option(&str, &max_cache_size);
5414 __setup("max_cache_size=", setup_max_cache_size);
5417 * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This
5418 * is the operation that is timed, so we try to generate unpredictable
5419 * cachemisses that still end up filling the L2 cache:
5421 static void touch_cache(void *__cache, unsigned long __size)
5423 unsigned long size = __size/sizeof(long), chunk1 = size/3,
5425 unsigned long *cache = __cache;
5428 for (i = 0; i < size/6; i += 8) {
5431 case 1: cache[size-1-i]++;
5432 case 2: cache[chunk1-i]++;
5433 case 3: cache[chunk1+i]++;
5434 case 4: cache[chunk2-i]++;
5435 case 5: cache[chunk2+i]++;
5441 * Measure the cache-cost of one task migration. Returns in units of nsec.
5443 static unsigned long long measure_one(void *cache, unsigned long size,
5444 int source, int target)
5446 cpumask_t mask, saved_mask;
5447 unsigned long long t0, t1, t2, t3, cost;
5449 saved_mask = current->cpus_allowed;
5452 * Flush source caches to RAM and invalidate them:
5457 * Migrate to the source CPU:
5459 mask = cpumask_of_cpu(source);
5460 set_cpus_allowed(current, mask);
5461 WARN_ON(smp_processor_id() != source);
5464 * Dirty the working set:
5467 touch_cache(cache, size);
5471 * Migrate to the target CPU, dirty the L2 cache and access
5472 * the shared buffer. (which represents the working set
5473 * of a migrated task.)
5475 mask = cpumask_of_cpu(target);
5476 set_cpus_allowed(current, mask);
5477 WARN_ON(smp_processor_id() != target);
5480 touch_cache(cache, size);
5483 cost = t1-t0 + t3-t2;
5485 if (migration_debug >= 2)
5486 printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n",
5487 source, target, t1-t0, t1-t0, t3-t2, cost);
5489 * Flush target caches to RAM and invalidate them:
5493 set_cpus_allowed(current, saved_mask);
5499 * Measure a series of task migrations and return the average
5500 * result. Since this code runs early during bootup the system
5501 * is 'undisturbed' and the average latency makes sense.
5503 * The algorithm in essence auto-detects the relevant cache-size,
5504 * so it will properly detect different cachesizes for different
5505 * cache-hierarchies, depending on how the CPUs are connected.
5507 * Architectures can prime the upper limit of the search range via
5508 * max_cache_size, otherwise the search range defaults to 20MB...64K.
5510 static unsigned long long
5511 measure_cost(int cpu1, int cpu2, void *cache, unsigned int size)
5513 unsigned long long cost1, cost2;
5517 * Measure the migration cost of 'size' bytes, over an
5518 * average of 10 runs:
5520 * (We perturb the cache size by a small (0..4k)
5521 * value to compensate size/alignment related artifacts.
5522 * We also subtract the cost of the operation done on
5528 * dry run, to make sure we start off cache-cold on cpu1,
5529 * and to get any vmalloc pagefaults in advance:
5531 measure_one(cache, size, cpu1, cpu2);
5532 for (i = 0; i < ITERATIONS; i++)
5533 cost1 += measure_one(cache, size - i*1024, cpu1, cpu2);
5535 measure_one(cache, size, cpu2, cpu1);
5536 for (i = 0; i < ITERATIONS; i++)
5537 cost1 += measure_one(cache, size - i*1024, cpu2, cpu1);
5540 * (We measure the non-migrating [cached] cost on both
5541 * cpu1 and cpu2, to handle CPUs with different speeds)
5545 measure_one(cache, size, cpu1, cpu1);
5546 for (i = 0; i < ITERATIONS; i++)
5547 cost2 += measure_one(cache, size - i*1024, cpu1, cpu1);
5549 measure_one(cache, size, cpu2, cpu2);
5550 for (i = 0; i < ITERATIONS; i++)
5551 cost2 += measure_one(cache, size - i*1024, cpu2, cpu2);
5554 * Get the per-iteration migration cost:
5556 do_div(cost1, 2*ITERATIONS);
5557 do_div(cost2, 2*ITERATIONS);
5559 return cost1 - cost2;
5562 static unsigned long long measure_migration_cost(int cpu1, int cpu2)
5564 unsigned long long max_cost = 0, fluct = 0, avg_fluct = 0;
5565 unsigned int max_size, size, size_found = 0;
5566 long long cost = 0, prev_cost;
5570 * Search from max_cache_size*5 down to 64K - the real relevant
5571 * cachesize has to lie somewhere inbetween.
5573 if (max_cache_size) {
5574 max_size = max(max_cache_size * SEARCH_SCOPE, MIN_CACHE_SIZE);
5575 size = max(max_cache_size / SEARCH_SCOPE, MIN_CACHE_SIZE);
5578 * Since we have no estimation about the relevant
5581 max_size = DEFAULT_CACHE_SIZE * SEARCH_SCOPE;
5582 size = MIN_CACHE_SIZE;
5585 if (!cpu_online(cpu1) || !cpu_online(cpu2)) {
5586 printk("cpu %d and %d not both online!\n", cpu1, cpu2);
5591 * Allocate the working set:
5593 cache = vmalloc(max_size);
5595 printk("could not vmalloc %d bytes for cache!\n", 2*max_size);
5596 return 1000000; // return 1 msec on very small boxen
5599 while (size <= max_size) {
5601 cost = measure_cost(cpu1, cpu2, cache, size);
5607 if (max_cost < cost) {
5613 * Calculate average fluctuation, we use this to prevent
5614 * noise from triggering an early break out of the loop:
5616 fluct = abs(cost - prev_cost);
5617 avg_fluct = (avg_fluct + fluct)/2;
5619 if (migration_debug)
5620 printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): (%8Ld %8Ld)\n",
5622 (long)cost / 1000000,
5623 ((long)cost / 100000) % 10,
5624 (long)max_cost / 1000000,
5625 ((long)max_cost / 100000) % 10,
5626 domain_distance(cpu1, cpu2),
5630 * If we iterated at least 20% past the previous maximum,
5631 * and the cost has dropped by more than 20% already,
5632 * (taking fluctuations into account) then we assume to
5633 * have found the maximum and break out of the loop early:
5635 if (size_found && (size*100 > size_found*SIZE_THRESH))
5636 if (cost+avg_fluct <= 0 ||
5637 max_cost*100 > (cost+avg_fluct)*COST_THRESH) {
5639 if (migration_debug)
5640 printk("-> found max.\n");
5644 * Increase the cachesize in 10% steps:
5646 size = size * 10 / 9;
5649 if (migration_debug)
5650 printk("[%d][%d] working set size found: %d, cost: %Ld\n",
5651 cpu1, cpu2, size_found, max_cost);
5656 * A task is considered 'cache cold' if at least 2 times
5657 * the worst-case cost of migration has passed.
5659 * (this limit is only listened to if the load-balancing
5660 * situation is 'nice' - if there is a large imbalance we
5661 * ignore it for the sake of CPU utilization and
5662 * processing fairness.)
5664 return 2 * max_cost * migration_factor / MIGRATION_FACTOR_SCALE;
5667 static void calibrate_migration_costs(const cpumask_t *cpu_map)
5669 int cpu1 = -1, cpu2 = -1, cpu, orig_cpu = raw_smp_processor_id();
5670 unsigned long j0, j1, distance, max_distance = 0;
5671 struct sched_domain *sd;
5676 * First pass - calculate the cacheflush times:
5678 for_each_cpu_mask(cpu1, *cpu_map) {
5679 for_each_cpu_mask(cpu2, *cpu_map) {
5682 distance = domain_distance(cpu1, cpu2);
5683 max_distance = max(max_distance, distance);
5685 * No result cached yet?
5687 if (migration_cost[distance] == -1LL)
5688 migration_cost[distance] =
5689 measure_migration_cost(cpu1, cpu2);
5693 * Second pass - update the sched domain hierarchy with
5694 * the new cache-hot-time estimations:
5696 for_each_cpu_mask(cpu, *cpu_map) {
5698 for_each_domain(cpu, sd) {
5699 sd->cache_hot_time = migration_cost[distance];
5706 if (migration_debug)
5707 printk("migration: max_cache_size: %d, cpu: %d MHz:\n",
5715 if (system_state == SYSTEM_BOOTING) {
5716 printk("migration_cost=");
5717 for (distance = 0; distance <= max_distance; distance++) {
5720 printk("%ld", (long)migration_cost[distance] / 1000);
5725 if (migration_debug)
5726 printk("migration: %ld seconds\n", (j1-j0)/HZ);
5729 * Move back to the original CPU. NUMA-Q gets confused
5730 * if we migrate to another quad during bootup.
5732 if (raw_smp_processor_id() != orig_cpu) {
5733 cpumask_t mask = cpumask_of_cpu(orig_cpu),
5734 saved_mask = current->cpus_allowed;
5736 set_cpus_allowed(current, mask);
5737 set_cpus_allowed(current, saved_mask);
5744 * find_next_best_node - find the next node to include in a sched_domain
5745 * @node: node whose sched_domain we're building
5746 * @used_nodes: nodes already in the sched_domain
5748 * Find the next node to include in a given scheduling domain. Simply
5749 * finds the closest node not already in the @used_nodes map.
5751 * Should use nodemask_t.
5753 static int find_next_best_node(int node, unsigned long *used_nodes)
5755 int i, n, val, min_val, best_node = 0;
5759 for (i = 0; i < MAX_NUMNODES; i++) {
5760 /* Start at @node */
5761 n = (node + i) % MAX_NUMNODES;
5763 if (!nr_cpus_node(n))
5766 /* Skip already used nodes */
5767 if (test_bit(n, used_nodes))
5770 /* Simple min distance search */
5771 val = node_distance(node, n);
5773 if (val < min_val) {
5779 set_bit(best_node, used_nodes);
5784 * sched_domain_node_span - get a cpumask for a node's sched_domain
5785 * @node: node whose cpumask we're constructing
5786 * @size: number of nodes to include in this span
5788 * Given a node, construct a good cpumask for its sched_domain to span. It
5789 * should be one that prevents unnecessary balancing, but also spreads tasks
5792 static cpumask_t sched_domain_node_span(int node)
5795 cpumask_t span, nodemask;
5796 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5799 bitmap_zero(used_nodes, MAX_NUMNODES);
5801 nodemask = node_to_cpumask(node);
5802 cpus_or(span, span, nodemask);
5803 set_bit(node, used_nodes);
5805 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5806 int next_node = find_next_best_node(node, used_nodes);
5807 nodemask = node_to_cpumask(next_node);
5808 cpus_or(span, span, nodemask);
5815 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5817 * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we
5818 * can switch it on easily if needed.
5820 #ifdef CONFIG_SCHED_SMT
5821 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5822 static struct sched_group sched_group_cpus[NR_CPUS];
5823 static int cpu_to_cpu_group(int cpu)
5829 #ifdef CONFIG_SCHED_MC
5830 static DEFINE_PER_CPU(struct sched_domain, core_domains);
5831 static struct sched_group *sched_group_core_bycpu[NR_CPUS];
5834 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5835 static int cpu_to_core_group(int cpu)
5837 return first_cpu(cpu_sibling_map[cpu]);
5839 #elif defined(CONFIG_SCHED_MC)
5840 static int cpu_to_core_group(int cpu)
5846 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5847 static struct sched_group *sched_group_phys_bycpu[NR_CPUS];
5848 static int cpu_to_phys_group(int cpu)
5850 #if defined(CONFIG_SCHED_MC)
5851 cpumask_t mask = cpu_coregroup_map(cpu);
5852 return first_cpu(mask);
5853 #elif defined(CONFIG_SCHED_SMT)
5854 return first_cpu(cpu_sibling_map[cpu]);
5862 * The init_sched_build_groups can't handle what we want to do with node
5863 * groups, so roll our own. Now each node has its own list of groups which
5864 * gets dynamically allocated.
5866 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5867 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5869 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5870 static struct sched_group *sched_group_allnodes_bycpu[NR_CPUS];
5872 static int cpu_to_allnodes_group(int cpu)
5874 return cpu_to_node(cpu);
5876 static void init_numa_sched_groups_power(struct sched_group *group_head)
5878 struct sched_group *sg = group_head;
5884 for_each_cpu_mask(j, sg->cpumask) {
5885 struct sched_domain *sd;
5887 sd = &per_cpu(phys_domains, j);
5888 if (j != first_cpu(sd->groups->cpumask)) {
5890 * Only add "power" once for each
5896 sg->cpu_power += sd->groups->cpu_power;
5899 if (sg != group_head)
5904 /* Free memory allocated for various sched_group structures */
5905 static void free_sched_groups(const cpumask_t *cpu_map)
5911 for_each_cpu_mask(cpu, *cpu_map) {
5912 struct sched_group *sched_group_allnodes
5913 = sched_group_allnodes_bycpu[cpu];
5914 struct sched_group **sched_group_nodes
5915 = sched_group_nodes_bycpu[cpu];
5917 if (sched_group_allnodes) {
5918 kfree(sched_group_allnodes);
5919 sched_group_allnodes_bycpu[cpu] = NULL;
5922 if (!sched_group_nodes)
5925 for (i = 0; i < MAX_NUMNODES; i++) {
5926 cpumask_t nodemask = node_to_cpumask(i);
5927 struct sched_group *oldsg, *sg = sched_group_nodes[i];
5929 cpus_and(nodemask, nodemask, *cpu_map);
5930 if (cpus_empty(nodemask))
5940 if (oldsg != sched_group_nodes[i])
5943 kfree(sched_group_nodes);
5944 sched_group_nodes_bycpu[cpu] = NULL;
5947 for_each_cpu_mask(cpu, *cpu_map) {
5948 if (sched_group_phys_bycpu[cpu]) {
5949 kfree(sched_group_phys_bycpu[cpu]);
5950 sched_group_phys_bycpu[cpu] = NULL;
5952 #ifdef CONFIG_SCHED_MC
5953 if (sched_group_core_bycpu[cpu]) {
5954 kfree(sched_group_core_bycpu[cpu]);
5955 sched_group_core_bycpu[cpu] = NULL;
5962 * Build sched domains for a given set of cpus and attach the sched domains
5963 * to the individual cpus
5965 static int build_sched_domains(const cpumask_t *cpu_map)
5968 struct sched_group *sched_group_phys = NULL;
5969 #ifdef CONFIG_SCHED_MC
5970 struct sched_group *sched_group_core = NULL;
5973 struct sched_group **sched_group_nodes = NULL;
5974 struct sched_group *sched_group_allnodes = NULL;
5977 * Allocate the per-node list of sched groups
5979 sched_group_nodes = kzalloc(sizeof(struct sched_group*)*MAX_NUMNODES,
5981 if (!sched_group_nodes) {
5982 printk(KERN_WARNING "Can not alloc sched group node list\n");
5985 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
5989 * Set up domains for cpus specified by the cpu_map.
5991 for_each_cpu_mask(i, *cpu_map) {
5993 struct sched_domain *sd = NULL, *p;
5994 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
5996 cpus_and(nodemask, nodemask, *cpu_map);
5999 if (cpus_weight(*cpu_map)
6000 > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6001 if (!sched_group_allnodes) {
6002 sched_group_allnodes
6003 = kmalloc(sizeof(struct sched_group)
6006 if (!sched_group_allnodes) {
6008 "Can not alloc allnodes sched group\n");
6011 sched_group_allnodes_bycpu[i]
6012 = sched_group_allnodes;
6014 sd = &per_cpu(allnodes_domains, i);
6015 *sd = SD_ALLNODES_INIT;
6016 sd->span = *cpu_map;
6017 group = cpu_to_allnodes_group(i);
6018 sd->groups = &sched_group_allnodes[group];
6023 sd = &per_cpu(node_domains, i);
6025 sd->span = sched_domain_node_span(cpu_to_node(i));
6027 cpus_and(sd->span, sd->span, *cpu_map);
6030 if (!sched_group_phys) {
6032 = kmalloc(sizeof(struct sched_group) * NR_CPUS,
6034 if (!sched_group_phys) {
6035 printk (KERN_WARNING "Can not alloc phys sched"
6039 sched_group_phys_bycpu[i] = sched_group_phys;
6043 sd = &per_cpu(phys_domains, i);
6044 group = cpu_to_phys_group(i);
6046 sd->span = nodemask;
6048 sd->groups = &sched_group_phys[group];
6050 #ifdef CONFIG_SCHED_MC
6051 if (!sched_group_core) {
6053 = kmalloc(sizeof(struct sched_group) * NR_CPUS,
6055 if (!sched_group_core) {
6056 printk (KERN_WARNING "Can not alloc core sched"
6060 sched_group_core_bycpu[i] = sched_group_core;
6064 sd = &per_cpu(core_domains, i);
6065 group = cpu_to_core_group(i);
6067 sd->span = cpu_coregroup_map(i);
6068 cpus_and(sd->span, sd->span, *cpu_map);
6070 sd->groups = &sched_group_core[group];
6073 #ifdef CONFIG_SCHED_SMT
6075 sd = &per_cpu(cpu_domains, i);
6076 group = cpu_to_cpu_group(i);
6077 *sd = SD_SIBLING_INIT;
6078 sd->span = cpu_sibling_map[i];
6079 cpus_and(sd->span, sd->span, *cpu_map);
6081 sd->groups = &sched_group_cpus[group];
6085 #ifdef CONFIG_SCHED_SMT
6086 /* Set up CPU (sibling) groups */
6087 for_each_cpu_mask(i, *cpu_map) {
6088 cpumask_t this_sibling_map = cpu_sibling_map[i];
6089 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6090 if (i != first_cpu(this_sibling_map))
6093 init_sched_build_groups(sched_group_cpus, this_sibling_map,
6098 #ifdef CONFIG_SCHED_MC
6099 /* Set up multi-core groups */
6100 for_each_cpu_mask(i, *cpu_map) {
6101 cpumask_t this_core_map = cpu_coregroup_map(i);
6102 cpus_and(this_core_map, this_core_map, *cpu_map);
6103 if (i != first_cpu(this_core_map))
6105 init_sched_build_groups(sched_group_core, this_core_map,
6106 &cpu_to_core_group);
6111 /* Set up physical groups */
6112 for (i = 0; i < MAX_NUMNODES; i++) {
6113 cpumask_t nodemask = node_to_cpumask(i);
6115 cpus_and(nodemask, nodemask, *cpu_map);
6116 if (cpus_empty(nodemask))
6119 init_sched_build_groups(sched_group_phys, nodemask,
6120 &cpu_to_phys_group);
6124 /* Set up node groups */
6125 if (sched_group_allnodes)
6126 init_sched_build_groups(sched_group_allnodes, *cpu_map,
6127 &cpu_to_allnodes_group);
6129 for (i = 0; i < MAX_NUMNODES; i++) {
6130 /* Set up node groups */
6131 struct sched_group *sg, *prev;
6132 cpumask_t nodemask = node_to_cpumask(i);
6133 cpumask_t domainspan;
6134 cpumask_t covered = CPU_MASK_NONE;
6137 cpus_and(nodemask, nodemask, *cpu_map);
6138 if (cpus_empty(nodemask)) {
6139 sched_group_nodes[i] = NULL;
6143 domainspan = sched_domain_node_span(i);
6144 cpus_and(domainspan, domainspan, *cpu_map);
6146 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6148 printk(KERN_WARNING "Can not alloc domain group for "
6152 sched_group_nodes[i] = sg;
6153 for_each_cpu_mask(j, nodemask) {
6154 struct sched_domain *sd;
6155 sd = &per_cpu(node_domains, j);
6159 sg->cpumask = nodemask;
6161 cpus_or(covered, covered, nodemask);
6164 for (j = 0; j < MAX_NUMNODES; j++) {
6165 cpumask_t tmp, notcovered;
6166 int n = (i + j) % MAX_NUMNODES;
6168 cpus_complement(notcovered, covered);
6169 cpus_and(tmp, notcovered, *cpu_map);
6170 cpus_and(tmp, tmp, domainspan);
6171 if (cpus_empty(tmp))
6174 nodemask = node_to_cpumask(n);
6175 cpus_and(tmp, tmp, nodemask);
6176 if (cpus_empty(tmp))
6179 sg = kmalloc_node(sizeof(struct sched_group),
6183 "Can not alloc domain group for node %d\n", j);
6188 sg->next = prev->next;
6189 cpus_or(covered, covered, tmp);
6196 /* Calculate CPU power for physical packages and nodes */
6197 #ifdef CONFIG_SCHED_SMT
6198 for_each_cpu_mask(i, *cpu_map) {
6199 struct sched_domain *sd;
6200 sd = &per_cpu(cpu_domains, i);
6201 sd->groups->cpu_power = SCHED_LOAD_SCALE;
6204 #ifdef CONFIG_SCHED_MC
6205 for_each_cpu_mask(i, *cpu_map) {
6207 struct sched_domain *sd;
6208 sd = &per_cpu(core_domains, i);
6209 if (sched_smt_power_savings)
6210 power = SCHED_LOAD_SCALE * cpus_weight(sd->groups->cpumask);
6212 power = SCHED_LOAD_SCALE + (cpus_weight(sd->groups->cpumask)-1)
6213 * SCHED_LOAD_SCALE / 10;
6214 sd->groups->cpu_power = power;
6218 for_each_cpu_mask(i, *cpu_map) {
6219 struct sched_domain *sd;
6220 #ifdef CONFIG_SCHED_MC
6221 sd = &per_cpu(phys_domains, i);
6222 if (i != first_cpu(sd->groups->cpumask))
6225 sd->groups->cpu_power = 0;
6226 if (sched_mc_power_savings || sched_smt_power_savings) {
6229 for_each_cpu_mask(j, sd->groups->cpumask) {
6230 struct sched_domain *sd1;
6231 sd1 = &per_cpu(core_domains, j);
6233 * for each core we will add once
6234 * to the group in physical domain
6236 if (j != first_cpu(sd1->groups->cpumask))
6239 if (sched_smt_power_savings)
6240 sd->groups->cpu_power += sd1->groups->cpu_power;
6242 sd->groups->cpu_power += SCHED_LOAD_SCALE;
6246 * This has to be < 2 * SCHED_LOAD_SCALE
6247 * Lets keep it SCHED_LOAD_SCALE, so that
6248 * while calculating NUMA group's cpu_power
6250 * numa_group->cpu_power += phys_group->cpu_power;
6252 * See "only add power once for each physical pkg"
6255 sd->groups->cpu_power = SCHED_LOAD_SCALE;
6258 sd = &per_cpu(phys_domains, i);
6259 if (sched_smt_power_savings)
6260 power = SCHED_LOAD_SCALE * cpus_weight(sd->groups->cpumask);
6262 power = SCHED_LOAD_SCALE;
6263 sd->groups->cpu_power = power;
6268 for (i = 0; i < MAX_NUMNODES; i++)
6269 init_numa_sched_groups_power(sched_group_nodes[i]);
6271 init_numa_sched_groups_power(sched_group_allnodes);
6274 /* Attach the domains */
6275 for_each_cpu_mask(i, *cpu_map) {
6276 struct sched_domain *sd;
6277 #ifdef CONFIG_SCHED_SMT
6278 sd = &per_cpu(cpu_domains, i);
6279 #elif defined(CONFIG_SCHED_MC)
6280 sd = &per_cpu(core_domains, i);
6282 sd = &per_cpu(phys_domains, i);
6284 cpu_attach_domain(sd, i);
6287 * Tune cache-hot values:
6289 calibrate_migration_costs(cpu_map);
6294 free_sched_groups(cpu_map);
6298 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6300 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6302 cpumask_t cpu_default_map;
6306 * Setup mask for cpus without special case scheduling requirements.
6307 * For now this just excludes isolated cpus, but could be used to
6308 * exclude other special cases in the future.
6310 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6312 err = build_sched_domains(&cpu_default_map);
6317 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6319 free_sched_groups(cpu_map);
6323 * Detach sched domains from a group of cpus specified in cpu_map
6324 * These cpus will now be attached to the NULL domain
6326 static void detach_destroy_domains(const cpumask_t *cpu_map)
6330 for_each_cpu_mask(i, *cpu_map)
6331 cpu_attach_domain(NULL, i);
6332 synchronize_sched();
6333 arch_destroy_sched_domains(cpu_map);
6337 * Partition sched domains as specified by the cpumasks below.
6338 * This attaches all cpus from the cpumasks to the NULL domain,
6339 * waits for a RCU quiescent period, recalculates sched
6340 * domain information and then attaches them back to the
6341 * correct sched domains
6342 * Call with hotplug lock held
6344 int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6346 cpumask_t change_map;
6349 cpus_and(*partition1, *partition1, cpu_online_map);
6350 cpus_and(*partition2, *partition2, cpu_online_map);
6351 cpus_or(change_map, *partition1, *partition2);
6353 /* Detach sched domains from all of the affected cpus */
6354 detach_destroy_domains(&change_map);
6355 if (!cpus_empty(*partition1))
6356 err = build_sched_domains(partition1);
6357 if (!err && !cpus_empty(*partition2))
6358 err = build_sched_domains(partition2);
6363 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6364 int arch_reinit_sched_domains(void)
6369 detach_destroy_domains(&cpu_online_map);
6370 err = arch_init_sched_domains(&cpu_online_map);
6371 unlock_cpu_hotplug();
6376 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6380 if (buf[0] != '0' && buf[0] != '1')
6384 sched_smt_power_savings = (buf[0] == '1');
6386 sched_mc_power_savings = (buf[0] == '1');
6388 ret = arch_reinit_sched_domains();
6390 return ret ? ret : count;
6393 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6396 #ifdef CONFIG_SCHED_SMT
6398 err = sysfs_create_file(&cls->kset.kobj,
6399 &attr_sched_smt_power_savings.attr);
6401 #ifdef CONFIG_SCHED_MC
6402 if (!err && mc_capable())
6403 err = sysfs_create_file(&cls->kset.kobj,
6404 &attr_sched_mc_power_savings.attr);
6410 #ifdef CONFIG_SCHED_MC
6411 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6413 return sprintf(page, "%u\n", sched_mc_power_savings);
6415 static ssize_t sched_mc_power_savings_store(struct sys_device *dev, const char *buf, size_t count)
6417 return sched_power_savings_store(buf, count, 0);
6419 SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6420 sched_mc_power_savings_store);
6423 #ifdef CONFIG_SCHED_SMT
6424 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6426 return sprintf(page, "%u\n", sched_smt_power_savings);
6428 static ssize_t sched_smt_power_savings_store(struct sys_device *dev, const char *buf, size_t count)
6430 return sched_power_savings_store(buf, count, 1);
6432 SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6433 sched_smt_power_savings_store);
6437 #ifdef CONFIG_HOTPLUG_CPU
6439 * Force a reinitialization of the sched domains hierarchy. The domains
6440 * and groups cannot be updated in place without racing with the balancing
6441 * code, so we temporarily attach all running cpus to the NULL domain
6442 * which will prevent rebalancing while the sched domains are recalculated.
6444 static int update_sched_domains(struct notifier_block *nfb,
6445 unsigned long action, void *hcpu)
6448 case CPU_UP_PREPARE:
6449 case CPU_DOWN_PREPARE:
6450 detach_destroy_domains(&cpu_online_map);
6453 case CPU_UP_CANCELED:
6454 case CPU_DOWN_FAILED:
6458 * Fall through and re-initialise the domains.
6465 /* The hotplug lock is already held by cpu_up/cpu_down */
6466 arch_init_sched_domains(&cpu_online_map);
6472 void __init sched_init_smp(void)
6475 arch_init_sched_domains(&cpu_online_map);
6476 unlock_cpu_hotplug();
6477 /* XXX: Theoretical race here - CPU may be hotplugged now */
6478 hotcpu_notifier(update_sched_domains, 0);
6481 void __init sched_init_smp(void)
6484 #endif /* CONFIG_SMP */
6486 int in_sched_functions(unsigned long addr)
6488 /* Linker adds these: start and end of __sched functions */
6489 extern char __sched_text_start[], __sched_text_end[];
6490 return in_lock_functions(addr) ||
6491 (addr >= (unsigned long)__sched_text_start
6492 && addr < (unsigned long)__sched_text_end);
6495 void __init sched_init(void)
6500 for_each_possible_cpu(i) {
6501 prio_array_t *array;
6504 spin_lock_init(&rq->lock);
6506 rq->active = rq->arrays;
6507 rq->expired = rq->arrays + 1;
6508 rq->best_expired_prio = MAX_PRIO;
6512 for (j = 1; j < 3; j++)
6513 rq->cpu_load[j] = 0;
6514 rq->active_balance = 0;
6516 rq->migration_thread = NULL;
6517 INIT_LIST_HEAD(&rq->migration_queue);
6519 atomic_set(&rq->nr_iowait, 0);
6521 for (j = 0; j < 2; j++) {
6522 array = rq->arrays + j;
6523 for (k = 0; k < MAX_PRIO; k++) {
6524 INIT_LIST_HEAD(array->queue + k);
6525 __clear_bit(k, array->bitmap);
6527 // delimiter for bitsearch
6528 __set_bit(MAX_PRIO, array->bitmap);
6532 set_load_weight(&init_task);
6534 * The boot idle thread does lazy MMU switching as well:
6536 atomic_inc(&init_mm.mm_count);
6537 enter_lazy_tlb(&init_mm, current);
6540 * Make us the idle thread. Technically, schedule() should not be
6541 * called from this thread, however somewhere below it might be,
6542 * but because we are the idle thread, we just pick up running again
6543 * when this runqueue becomes "idle".
6545 init_idle(current, smp_processor_id());
6548 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6549 void __might_sleep(char *file, int line)
6551 #if defined(in_atomic)
6552 static unsigned long prev_jiffy; /* ratelimiting */
6554 if ((in_atomic() || irqs_disabled()) &&
6555 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6556 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6558 prev_jiffy = jiffies;
6559 printk(KERN_ERR "BUG: sleeping function called from invalid"
6560 " context at %s:%d\n", file, line);
6561 printk("in_atomic():%d, irqs_disabled():%d\n",
6562 in_atomic(), irqs_disabled());
6567 EXPORT_SYMBOL(__might_sleep);
6570 #ifdef CONFIG_MAGIC_SYSRQ
6571 void normalize_rt_tasks(void)
6573 struct task_struct *p;
6574 prio_array_t *array;
6575 unsigned long flags;
6578 read_lock_irq(&tasklist_lock);
6579 for_each_process(p) {
6583 rq = task_rq_lock(p, &flags);
6587 deactivate_task(p, task_rq(p));
6588 __setscheduler(p, SCHED_NORMAL, 0);
6590 __activate_task(p, task_rq(p));
6591 resched_task(rq->curr);
6594 task_rq_unlock(rq, &flags);
6596 read_unlock_irq(&tasklist_lock);
6599 #endif /* CONFIG_MAGIC_SYSRQ */
6603 * These functions are only useful for the IA64 MCA handling.
6605 * They can only be called when the whole system has been
6606 * stopped - every CPU needs to be quiescent, and no scheduling
6607 * activity can take place. Using them for anything else would
6608 * be a serious bug, and as a result, they aren't even visible
6609 * under any other configuration.
6613 * curr_task - return the current task for a given cpu.
6614 * @cpu: the processor in question.
6616 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6618 task_t *curr_task(int cpu)
6620 return cpu_curr(cpu);
6624 * set_curr_task - set the current task for a given cpu.
6625 * @cpu: the processor in question.
6626 * @p: the task pointer to set.
6628 * Description: This function must only be used when non-maskable interrupts
6629 * are serviced on a separate stack. It allows the architecture to switch the
6630 * notion of the current task on a cpu in a non-blocking manner. This function
6631 * must be called with all CPU's synchronized, and interrupts disabled, the
6632 * and caller must save the original value of the current task (see
6633 * curr_task() above) and restore that value before reenabling interrupts and
6634 * re-starting the system.
6636 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6638 void set_curr_task(int cpu, task_t *p)