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>
56 #include <linux/vs_context.h>
57 #include <linux/vs_cvirt.h>
58 #include <linux/vs_sched.h>
61 * Convert user-nice values [ -20 ... 0 ... 19 ]
62 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
65 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
66 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
67 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
70 * 'User priority' is the nice value converted to something we
71 * can work with better when scaling various scheduler parameters,
72 * it's a [ 0 ... 39 ] range.
74 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
75 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
76 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
79 * Some helpers for converting nanosecond timing to jiffy resolution
81 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
82 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
85 * These are the 'tuning knobs' of the scheduler:
87 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
88 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
89 * Timeslices get refilled after they expire.
91 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
92 #define DEF_TIMESLICE (100 * HZ / 1000)
93 #define ON_RUNQUEUE_WEIGHT 30
94 #define CHILD_PENALTY 95
95 #define PARENT_PENALTY 100
97 #define PRIO_BONUS_RATIO 25
98 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
99 #define INTERACTIVE_DELTA 2
100 #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
101 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
102 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
105 * If a task is 'interactive' then we reinsert it in the active
106 * array after it has expired its current timeslice. (it will not
107 * continue to run immediately, it will still roundrobin with
108 * other interactive tasks.)
110 * This part scales the interactivity limit depending on niceness.
112 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
113 * Here are a few examples of different nice levels:
115 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
116 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
117 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
118 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
119 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
121 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
122 * priority range a task can explore, a value of '1' means the
123 * task is rated interactive.)
125 * Ie. nice +19 tasks can never get 'interactive' enough to be
126 * reinserted into the active array. And only heavily CPU-hog nice -20
127 * tasks will be expired. Default nice 0 tasks are somewhere between,
128 * it takes some effort for them to get interactive, but it's not
132 #define CURRENT_BONUS(p) \
133 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
136 #define GRANULARITY (10 * HZ / 1000 ? : 1)
139 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
140 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
143 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
144 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
147 #define SCALE(v1,v1_max,v2_max) \
148 (v1) * (v2_max) / (v1_max)
151 (SCALE(TASK_NICE(p) + 20, 40, MAX_BONUS) - 20 * MAX_BONUS / 40 + \
154 #define TASK_INTERACTIVE(p) \
155 ((p)->prio <= (p)->static_prio - DELTA(p))
157 #define INTERACTIVE_SLEEP(p) \
158 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
159 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
161 #define TASK_PREEMPTS_CURR(p, rq) \
162 ((p)->prio < (rq)->curr->prio)
165 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
166 * to time slice values: [800ms ... 100ms ... 5ms]
168 * The higher a thread's priority, the bigger timeslices
169 * it gets during one round of execution. But even the lowest
170 * priority thread gets MIN_TIMESLICE worth of execution time.
173 #define SCALE_PRIO(x, prio) \
174 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO/2), MIN_TIMESLICE)
176 static unsigned int task_timeslice(task_t *p)
178 if (p->static_prio < NICE_TO_PRIO(0))
179 return SCALE_PRIO(DEF_TIMESLICE*4, p->static_prio);
181 return SCALE_PRIO(DEF_TIMESLICE, p->static_prio);
183 #define task_hot(p, now, sd) ((long long) ((now) - (p)->last_ran) \
184 < (long long) (sd)->cache_hot_time)
187 * These are the runqueue data structures:
190 #define BITMAP_SIZE ((((MAX_PRIO+1+7)/8)+sizeof(long)-1)/sizeof(long))
192 typedef struct runqueue runqueue_t;
195 unsigned int nr_active;
196 unsigned long bitmap[BITMAP_SIZE];
197 struct list_head queue[MAX_PRIO];
201 * This is the main, per-CPU runqueue data structure.
203 * Locking rule: those places that want to lock multiple runqueues
204 * (such as the load balancing or the thread migration code), lock
205 * acquire operations must be ordered by ascending &runqueue.
211 * nr_running and cpu_load should be in the same cacheline because
212 * remote CPUs use both these fields when doing load calculation.
214 unsigned long nr_running;
216 unsigned long cpu_load[3];
218 unsigned long long nr_switches;
221 * This is part of a global counter where only the total sum
222 * over all CPUs matters. A task can increase this counter on
223 * one CPU and if it got migrated afterwards it may decrease
224 * it on another CPU. Always updated under the runqueue lock:
226 unsigned long nr_uninterruptible;
228 unsigned long expired_timestamp;
229 unsigned long long timestamp_last_tick;
231 struct mm_struct *prev_mm;
232 prio_array_t *active, *expired, arrays[2];
233 int best_expired_prio;
237 struct sched_domain *sd;
239 /* For active balancing */
243 task_t *migration_thread;
244 struct list_head migration_queue;
247 #ifdef CONFIG_VSERVER_HARDCPU
248 struct list_head hold_queue;
252 #ifdef CONFIG_SCHEDSTATS
254 struct sched_info rq_sched_info;
256 /* sys_sched_yield() stats */
257 unsigned long yld_exp_empty;
258 unsigned long yld_act_empty;
259 unsigned long yld_both_empty;
260 unsigned long yld_cnt;
262 /* schedule() stats */
263 unsigned long sched_switch;
264 unsigned long sched_cnt;
265 unsigned long sched_goidle;
267 /* try_to_wake_up() stats */
268 unsigned long ttwu_cnt;
269 unsigned long ttwu_local;
273 static DEFINE_PER_CPU(struct runqueue, runqueues);
276 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
277 * See detach_destroy_domains: synchronize_sched for details.
279 * The domain tree of any CPU may only be accessed from within
280 * preempt-disabled sections.
282 #define for_each_domain(cpu, domain) \
283 for (domain = rcu_dereference(cpu_rq(cpu)->sd); domain; domain = domain->parent)
285 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
286 #define this_rq() (&__get_cpu_var(runqueues))
287 #define task_rq(p) cpu_rq(task_cpu(p))
288 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
290 #ifndef prepare_arch_switch
291 # define prepare_arch_switch(next) do { } while (0)
293 #ifndef finish_arch_switch
294 # define finish_arch_switch(prev) do { } while (0)
297 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
298 static inline int task_running(runqueue_t *rq, task_t *p)
300 return rq->curr == p;
303 static inline void prepare_lock_switch(runqueue_t *rq, task_t *next)
307 static inline void finish_lock_switch(runqueue_t *rq, task_t *prev)
309 #ifdef CONFIG_DEBUG_SPINLOCK
310 /* this is a valid case when another task releases the spinlock */
311 rq->lock.owner = current;
313 spin_unlock_irq(&rq->lock);
316 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
317 static inline int task_running(runqueue_t *rq, task_t *p)
322 return rq->curr == p;
326 static inline void prepare_lock_switch(runqueue_t *rq, task_t *next)
330 * We can optimise this out completely for !SMP, because the
331 * SMP rebalancing from interrupt is the only thing that cares
336 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
337 spin_unlock_irq(&rq->lock);
339 spin_unlock(&rq->lock);
343 static inline void finish_lock_switch(runqueue_t *rq, task_t *prev)
347 * After ->oncpu is cleared, the task can be moved to a different CPU.
348 * We must ensure this doesn't happen until the switch is completely
354 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
358 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
361 * task_rq_lock - lock the runqueue a given task resides on and disable
362 * interrupts. Note the ordering: we can safely lookup the task_rq without
363 * explicitly disabling preemption.
365 static inline runqueue_t *task_rq_lock(task_t *p, unsigned long *flags)
371 local_irq_save(*flags);
373 spin_lock(&rq->lock);
374 if (unlikely(rq != task_rq(p))) {
375 spin_unlock_irqrestore(&rq->lock, *flags);
376 goto repeat_lock_task;
381 static inline void task_rq_unlock(runqueue_t *rq, unsigned long *flags)
384 spin_unlock_irqrestore(&rq->lock, *flags);
387 #ifdef CONFIG_SCHEDSTATS
389 * bump this up when changing the output format or the meaning of an existing
390 * format, so that tools can adapt (or abort)
392 #define SCHEDSTAT_VERSION 12
394 static int show_schedstat(struct seq_file *seq, void *v)
398 seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION);
399 seq_printf(seq, "timestamp %lu\n", jiffies);
400 for_each_online_cpu(cpu) {
401 runqueue_t *rq = cpu_rq(cpu);
403 struct sched_domain *sd;
407 /* runqueue-specific stats */
409 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
410 cpu, rq->yld_both_empty,
411 rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt,
412 rq->sched_switch, rq->sched_cnt, rq->sched_goidle,
413 rq->ttwu_cnt, rq->ttwu_local,
414 rq->rq_sched_info.cpu_time,
415 rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt);
417 seq_printf(seq, "\n");
420 /* domain-specific stats */
422 for_each_domain(cpu, sd) {
423 enum idle_type itype;
424 char mask_str[NR_CPUS];
426 cpumask_scnprintf(mask_str, NR_CPUS, sd->span);
427 seq_printf(seq, "domain%d %s", dcnt++, mask_str);
428 for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES;
430 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu",
432 sd->lb_balanced[itype],
433 sd->lb_failed[itype],
434 sd->lb_imbalance[itype],
435 sd->lb_gained[itype],
436 sd->lb_hot_gained[itype],
437 sd->lb_nobusyq[itype],
438 sd->lb_nobusyg[itype]);
440 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
441 sd->alb_cnt, sd->alb_failed, sd->alb_pushed,
442 sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed,
443 sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed,
444 sd->ttwu_wake_remote, sd->ttwu_move_affine, sd->ttwu_move_balance);
452 static int schedstat_open(struct inode *inode, struct file *file)
454 unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32);
455 char *buf = kmalloc(size, GFP_KERNEL);
461 res = single_open(file, show_schedstat, NULL);
463 m = file->private_data;
471 struct file_operations proc_schedstat_operations = {
472 .open = schedstat_open,
475 .release = single_release,
478 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
479 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
480 #else /* !CONFIG_SCHEDSTATS */
481 # define schedstat_inc(rq, field) do { } while (0)
482 # define schedstat_add(rq, field, amt) do { } while (0)
486 * rq_lock - lock a given runqueue and disable interrupts.
488 static inline runqueue_t *this_rq_lock(void)
495 spin_lock(&rq->lock);
500 #ifdef CONFIG_SCHEDSTATS
502 * Called when a process is dequeued from the active array and given
503 * the cpu. We should note that with the exception of interactive
504 * tasks, the expired queue will become the active queue after the active
505 * queue is empty, without explicitly dequeuing and requeuing tasks in the
506 * expired queue. (Interactive tasks may be requeued directly to the
507 * active queue, thus delaying tasks in the expired queue from running;
508 * see scheduler_tick()).
510 * This function is only called from sched_info_arrive(), rather than
511 * dequeue_task(). Even though a task may be queued and dequeued multiple
512 * times as it is shuffled about, we're really interested in knowing how
513 * long it was from the *first* time it was queued to the time that it
516 static inline void sched_info_dequeued(task_t *t)
518 t->sched_info.last_queued = 0;
522 * Called when a task finally hits the cpu. We can now calculate how
523 * long it was waiting to run. We also note when it began so that we
524 * can keep stats on how long its timeslice is.
526 static void sched_info_arrive(task_t *t)
528 unsigned long now = jiffies, diff = 0;
529 struct runqueue *rq = task_rq(t);
531 if (t->sched_info.last_queued)
532 diff = now - t->sched_info.last_queued;
533 sched_info_dequeued(t);
534 t->sched_info.run_delay += diff;
535 t->sched_info.last_arrival = now;
536 t->sched_info.pcnt++;
541 rq->rq_sched_info.run_delay += diff;
542 rq->rq_sched_info.pcnt++;
546 * Called when a process is queued into either the active or expired
547 * array. The time is noted and later used to determine how long we
548 * had to wait for us to reach the cpu. Since the expired queue will
549 * become the active queue after active queue is empty, without dequeuing
550 * and requeuing any tasks, we are interested in queuing to either. It
551 * is unusual but not impossible for tasks to be dequeued and immediately
552 * requeued in the same or another array: this can happen in sched_yield(),
553 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
556 * This function is only called from enqueue_task(), but also only updates
557 * the timestamp if it is already not set. It's assumed that
558 * sched_info_dequeued() will clear that stamp when appropriate.
560 static inline void sched_info_queued(task_t *t)
562 if (!t->sched_info.last_queued)
563 t->sched_info.last_queued = jiffies;
567 * Called when a process ceases being the active-running process, either
568 * voluntarily or involuntarily. Now we can calculate how long we ran.
570 static inline void sched_info_depart(task_t *t)
572 struct runqueue *rq = task_rq(t);
573 unsigned long diff = jiffies - t->sched_info.last_arrival;
575 t->sched_info.cpu_time += diff;
578 rq->rq_sched_info.cpu_time += diff;
582 * Called when tasks are switched involuntarily due, typically, to expiring
583 * their time slice. (This may also be called when switching to or from
584 * the idle task.) We are only called when prev != next.
586 static inline void sched_info_switch(task_t *prev, task_t *next)
588 struct runqueue *rq = task_rq(prev);
591 * prev now departs the cpu. It's not interesting to record
592 * stats about how efficient we were at scheduling the idle
595 if (prev != rq->idle)
596 sched_info_depart(prev);
598 if (next != rq->idle)
599 sched_info_arrive(next);
602 #define sched_info_queued(t) do { } while (0)
603 #define sched_info_switch(t, next) do { } while (0)
604 #endif /* CONFIG_SCHEDSTATS */
607 * Adding/removing a task to/from a priority array:
609 static void dequeue_task(struct task_struct *p, prio_array_t *array)
611 BUG_ON(p->state & TASK_ONHOLD);
613 list_del(&p->run_list);
614 if (list_empty(array->queue + p->prio))
615 __clear_bit(p->prio, array->bitmap);
618 static void enqueue_task(struct task_struct *p, prio_array_t *array)
620 BUG_ON(p->state & TASK_ONHOLD);
621 sched_info_queued(p);
622 list_add_tail(&p->run_list, array->queue + p->prio);
623 __set_bit(p->prio, array->bitmap);
629 * Put task to the end of the run list without the overhead of dequeue
630 * followed by enqueue.
632 static void requeue_task(struct task_struct *p, prio_array_t *array)
634 BUG_ON(p->state & TASK_ONHOLD);
635 list_move_tail(&p->run_list, array->queue + p->prio);
638 static inline void enqueue_task_head(struct task_struct *p, prio_array_t *array)
640 BUG_ON(p->state & TASK_ONHOLD);
641 list_add(&p->run_list, array->queue + p->prio);
642 __set_bit(p->prio, array->bitmap);
648 * effective_prio - return the priority that is based on the static
649 * priority but is modified by bonuses/penalties.
651 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
652 * into the -5 ... 0 ... +5 bonus/penalty range.
654 * We use 25% of the full 0...39 priority range so that:
656 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
657 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
659 * Both properties are important to certain workloads.
661 static int effective_prio(task_t *p)
669 bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
671 prio = p->static_prio - bonus;
673 if ((vxi = p->vx_info) &&
674 vx_info_flags(vxi, VXF_SCHED_PRIO, 0))
675 prio += vx_effective_vavavoom(vxi, MAX_USER_PRIO);
677 if (prio < MAX_RT_PRIO)
679 if (prio > MAX_PRIO-1)
685 * __activate_task - move a task to the runqueue.
687 static void __activate_task(task_t *p, runqueue_t *rq)
689 prio_array_t *target = rq->active;
692 target = rq->expired;
693 enqueue_task(p, target);
698 * __activate_idle_task - move idle task to the _front_ of runqueue.
700 static inline void __activate_idle_task(task_t *p, runqueue_t *rq)
702 enqueue_task_head(p, rq->active);
706 static int recalc_task_prio(task_t *p, unsigned long long now)
708 /* Caller must always ensure 'now >= p->timestamp' */
709 unsigned long long __sleep_time = now - p->timestamp;
710 unsigned long sleep_time;
715 if (__sleep_time > NS_MAX_SLEEP_AVG)
716 sleep_time = NS_MAX_SLEEP_AVG;
718 sleep_time = (unsigned long)__sleep_time;
721 if (likely(sleep_time > 0)) {
723 * User tasks that sleep a long time are categorised as
724 * idle. They will only have their sleep_avg increased to a
725 * level that makes them just interactive priority to stay
726 * active yet prevent them suddenly becoming cpu hogs and
727 * starving other processes.
729 if (p->mm && sleep_time > INTERACTIVE_SLEEP(p)) {
730 unsigned long ceiling;
732 ceiling = JIFFIES_TO_NS(MAX_SLEEP_AVG -
734 if (p->sleep_avg < ceiling)
735 p->sleep_avg = ceiling;
738 * Tasks waking from uninterruptible sleep are
739 * limited in their sleep_avg rise as they
740 * are likely to be waiting on I/O
742 if (p->sleep_type == SLEEP_NONINTERACTIVE && p->mm) {
743 if (p->sleep_avg >= INTERACTIVE_SLEEP(p))
745 else if (p->sleep_avg + sleep_time >=
746 INTERACTIVE_SLEEP(p)) {
747 p->sleep_avg = INTERACTIVE_SLEEP(p);
753 * This code gives a bonus to interactive tasks.
755 * The boost works by updating the 'average sleep time'
756 * value here, based on ->timestamp. The more time a
757 * task spends sleeping, the higher the average gets -
758 * and the higher the priority boost gets as well.
760 p->sleep_avg += sleep_time;
762 if (p->sleep_avg > NS_MAX_SLEEP_AVG)
763 p->sleep_avg = NS_MAX_SLEEP_AVG;
767 return effective_prio(p);
771 * activate_task - move a task to the runqueue and do priority recalculation
773 * Update all the scheduling statistics stuff. (sleep average
774 * calculation, priority modifiers, etc.)
776 static void activate_task(task_t *p, runqueue_t *rq, int local)
778 unsigned long long now;
783 /* Compensate for drifting sched_clock */
784 runqueue_t *this_rq = this_rq();
785 now = (now - this_rq->timestamp_last_tick)
786 + rq->timestamp_last_tick;
791 p->prio = recalc_task_prio(p, now);
794 * This checks to make sure it's not an uninterruptible task
795 * that is now waking up.
797 if (p->sleep_type == SLEEP_NORMAL) {
799 * Tasks which were woken up by interrupts (ie. hw events)
800 * are most likely of interactive nature. So we give them
801 * the credit of extending their sleep time to the period
802 * of time they spend on the runqueue, waiting for execution
803 * on a CPU, first time around:
806 p->sleep_type = SLEEP_INTERRUPTED;
809 * Normal first-time wakeups get a credit too for
810 * on-runqueue time, but it will be weighted down:
812 p->sleep_type = SLEEP_INTERACTIVE;
818 __activate_task(p, rq);
822 * deactivate_task - remove a task from the runqueue.
824 static void __deactivate_task(struct task_struct *p, runqueue_t *rq)
827 dequeue_task(p, p->array);
832 void deactivate_task(struct task_struct *p, runqueue_t *rq)
834 vx_deactivate_task(p);
835 __deactivate_task(p, rq);
839 #ifdef CONFIG_VSERVER_HARDCPU
841 * vx_hold_task - put a task on the hold queue
844 void vx_hold_task(struct vx_info *vxi,
845 struct task_struct *p, runqueue_t *rq)
847 __deactivate_task(p, rq);
848 p->state |= TASK_ONHOLD;
849 /* a new one on hold */
851 list_add_tail(&p->run_list, &rq->hold_queue);
855 * vx_unhold_task - put a task back to the runqueue
858 void vx_unhold_task(struct vx_info *vxi,
859 struct task_struct *p, runqueue_t *rq)
861 list_del(&p->run_list);
862 /* one less waiting */
864 p->state &= ~TASK_ONHOLD;
865 enqueue_task(p, rq->expired);
868 if (p->static_prio < rq->best_expired_prio)
869 rq->best_expired_prio = p->static_prio;
873 void vx_hold_task(struct vx_info *vxi,
874 struct task_struct *p, runqueue_t *rq)
880 void vx_unhold_task(struct vx_info *vxi,
881 struct task_struct *p, runqueue_t *rq)
885 #endif /* CONFIG_VSERVER_HARDCPU */
889 * resched_task - mark a task 'to be rescheduled now'.
891 * On UP this means the setting of the need_resched flag, on SMP it
892 * might also involve a cross-CPU call to trigger the scheduler on
896 static void resched_task(task_t *p)
900 assert_spin_locked(&task_rq(p)->lock);
902 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
905 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
908 if (cpu == smp_processor_id())
911 /* NEED_RESCHED must be visible before we test POLLING_NRFLAG */
913 if (!test_tsk_thread_flag(p, TIF_POLLING_NRFLAG))
914 smp_send_reschedule(cpu);
917 static inline void resched_task(task_t *p)
919 assert_spin_locked(&task_rq(p)->lock);
920 set_tsk_need_resched(p);
925 * task_curr - is this task currently executing on a CPU?
926 * @p: the task in question.
928 inline int task_curr(const task_t *p)
930 return cpu_curr(task_cpu(p)) == p;
935 struct list_head list;
940 struct completion done;
944 * The task's runqueue lock must be held.
945 * Returns true if you have to wait for migration thread.
947 static int migrate_task(task_t *p, int dest_cpu, migration_req_t *req)
949 runqueue_t *rq = task_rq(p);
952 * If the task is not on a runqueue (and not running), then
953 * it is sufficient to simply update the task's cpu field.
955 if (!p->array && !task_running(rq, p)) {
956 set_task_cpu(p, dest_cpu);
960 init_completion(&req->done);
962 req->dest_cpu = dest_cpu;
963 list_add(&req->list, &rq->migration_queue);
968 * wait_task_inactive - wait for a thread to unschedule.
970 * The caller must ensure that the task *will* unschedule sometime soon,
971 * else this function might spin for a *long* time. This function can't
972 * be called with interrupts off, or it may introduce deadlock with
973 * smp_call_function() if an IPI is sent by the same process we are
974 * waiting to become inactive.
976 void wait_task_inactive(task_t *p)
983 rq = task_rq_lock(p, &flags);
984 /* Must be off runqueue entirely, not preempted. */
985 if (unlikely(p->array || task_running(rq, p))) {
986 /* If it's preempted, we yield. It could be a while. */
987 preempted = !task_running(rq, p);
988 task_rq_unlock(rq, &flags);
994 task_rq_unlock(rq, &flags);
998 * kick_process - kick a running thread to enter/exit the kernel
999 * @p: the to-be-kicked thread
1001 * Cause a process which is running on another CPU to enter
1002 * kernel-mode, without any delay. (to get signals handled.)
1004 * NOTE: this function doesnt have to take the runqueue lock,
1005 * because all it wants to ensure is that the remote task enters
1006 * the kernel. If the IPI races and the task has been migrated
1007 * to another CPU then no harm is done and the purpose has been
1010 void kick_process(task_t *p)
1016 if ((cpu != smp_processor_id()) && task_curr(p))
1017 smp_send_reschedule(cpu);
1022 * Return a low guess at the load of a migration-source cpu.
1024 * We want to under-estimate the load of migration sources, to
1025 * balance conservatively.
1027 static inline unsigned long source_load(int cpu, int type)
1029 runqueue_t *rq = cpu_rq(cpu);
1030 unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE;
1034 return min(rq->cpu_load[type-1], load_now);
1038 * Return a high guess at the load of a migration-target cpu
1040 static inline unsigned long target_load(int cpu, int type)
1042 runqueue_t *rq = cpu_rq(cpu);
1043 unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE;
1047 return max(rq->cpu_load[type-1], load_now);
1051 * find_idlest_group finds and returns the least busy CPU group within the
1054 static struct sched_group *
1055 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1057 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1058 unsigned long min_load = ULONG_MAX, this_load = 0;
1059 int load_idx = sd->forkexec_idx;
1060 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1063 unsigned long load, avg_load;
1067 /* Skip over this group if it has no CPUs allowed */
1068 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1071 local_group = cpu_isset(this_cpu, group->cpumask);
1073 /* Tally up the load of all CPUs in the group */
1076 for_each_cpu_mask(i, group->cpumask) {
1077 /* Bias balancing toward cpus of our domain */
1079 load = source_load(i, load_idx);
1081 load = target_load(i, load_idx);
1086 /* Adjust by relative CPU power of the group */
1087 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1090 this_load = avg_load;
1092 } else if (avg_load < min_load) {
1093 min_load = avg_load;
1097 group = group->next;
1098 } while (group != sd->groups);
1100 if (!idlest || 100*this_load < imbalance*min_load)
1106 * find_idlest_queue - find the idlest runqueue among the cpus in group.
1109 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1112 unsigned long load, min_load = ULONG_MAX;
1116 /* Traverse only the allowed CPUs */
1117 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1119 for_each_cpu_mask(i, tmp) {
1120 load = source_load(i, 0);
1122 if (load < min_load || (load == min_load && i == this_cpu)) {
1132 * sched_balance_self: balance the current task (running on cpu) in domains
1133 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1136 * Balance, ie. select the least loaded group.
1138 * Returns the target CPU number, or the same CPU if no balancing is needed.
1140 * preempt must be disabled.
1142 static int sched_balance_self(int cpu, int flag)
1144 struct task_struct *t = current;
1145 struct sched_domain *tmp, *sd = NULL;
1147 for_each_domain(cpu, tmp)
1148 if (tmp->flags & flag)
1153 struct sched_group *group;
1158 group = find_idlest_group(sd, t, cpu);
1162 new_cpu = find_idlest_cpu(group, t, cpu);
1163 if (new_cpu == -1 || new_cpu == cpu)
1166 /* Now try balancing at a lower domain level */
1170 weight = cpus_weight(span);
1171 for_each_domain(cpu, tmp) {
1172 if (weight <= cpus_weight(tmp->span))
1174 if (tmp->flags & flag)
1177 /* while loop will break here if sd == NULL */
1183 #endif /* CONFIG_SMP */
1186 * wake_idle() will wake a task on an idle cpu if task->cpu is
1187 * not idle and an idle cpu is available. The span of cpus to
1188 * search starts with cpus closest then further out as needed,
1189 * so we always favor a closer, idle cpu.
1191 * Returns the CPU we should wake onto.
1193 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1194 static int wake_idle(int cpu, task_t *p)
1197 struct sched_domain *sd;
1203 for_each_domain(cpu, sd) {
1204 if (sd->flags & SD_WAKE_IDLE) {
1205 cpus_and(tmp, sd->span, p->cpus_allowed);
1206 for_each_cpu_mask(i, tmp) {
1217 static inline int wake_idle(int cpu, task_t *p)
1224 * try_to_wake_up - wake up a thread
1225 * @p: the to-be-woken-up thread
1226 * @state: the mask of task states that can be woken
1227 * @sync: do a synchronous wakeup?
1229 * Put it on the run-queue if it's not already there. The "current"
1230 * thread is always on the run-queue (except when the actual
1231 * re-schedule is in progress), and as such you're allowed to do
1232 * the simpler "current->state = TASK_RUNNING" to mark yourself
1233 * runnable without the overhead of this.
1235 * returns failure only if the task is already active.
1237 static int try_to_wake_up(task_t *p, unsigned int state, int sync)
1239 int cpu, this_cpu, success = 0;
1240 unsigned long flags;
1244 unsigned long load, this_load;
1245 struct sched_domain *sd, *this_sd = NULL;
1249 rq = task_rq_lock(p, &flags);
1250 old_state = p->state;
1252 /* we need to unhold suspended tasks */
1253 if (old_state & TASK_ONHOLD) {
1254 vx_unhold_task(p->vx_info, p, rq);
1255 old_state = p->state;
1257 if (!(old_state & state))
1264 this_cpu = smp_processor_id();
1267 if (unlikely(task_running(rq, p)))
1272 schedstat_inc(rq, ttwu_cnt);
1273 if (cpu == this_cpu) {
1274 schedstat_inc(rq, ttwu_local);
1278 for_each_domain(this_cpu, sd) {
1279 if (cpu_isset(cpu, sd->span)) {
1280 schedstat_inc(sd, ttwu_wake_remote);
1286 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1290 * Check for affine wakeup and passive balancing possibilities.
1293 int idx = this_sd->wake_idx;
1294 unsigned int imbalance;
1296 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1298 load = source_load(cpu, idx);
1299 this_load = target_load(this_cpu, idx);
1301 new_cpu = this_cpu; /* Wake to this CPU if we can */
1303 if (this_sd->flags & SD_WAKE_AFFINE) {
1304 unsigned long tl = this_load;
1306 * If sync wakeup then subtract the (maximum possible)
1307 * effect of the currently running task from the load
1308 * of the current CPU:
1311 tl -= SCHED_LOAD_SCALE;
1314 tl + target_load(cpu, idx) <= SCHED_LOAD_SCALE) ||
1315 100*(tl + SCHED_LOAD_SCALE) <= imbalance*load) {
1317 * This domain has SD_WAKE_AFFINE and
1318 * p is cache cold in this domain, and
1319 * there is no bad imbalance.
1321 schedstat_inc(this_sd, ttwu_move_affine);
1327 * Start passive balancing when half the imbalance_pct
1330 if (this_sd->flags & SD_WAKE_BALANCE) {
1331 if (imbalance*this_load <= 100*load) {
1332 schedstat_inc(this_sd, ttwu_move_balance);
1338 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1340 new_cpu = wake_idle(new_cpu, p);
1341 if (new_cpu != cpu) {
1342 set_task_cpu(p, new_cpu);
1343 task_rq_unlock(rq, &flags);
1344 /* might preempt at this point */
1345 rq = task_rq_lock(p, &flags);
1346 old_state = p->state;
1347 if (!(old_state & state))
1352 this_cpu = smp_processor_id();
1357 #endif /* CONFIG_SMP */
1358 if (old_state == TASK_UNINTERRUPTIBLE) {
1359 rq->nr_uninterruptible--;
1360 vx_uninterruptible_dec(p);
1362 * Tasks on involuntary sleep don't earn
1363 * sleep_avg beyond just interactive state.
1365 p->sleep_type = SLEEP_NONINTERACTIVE;
1369 * Tasks that have marked their sleep as noninteractive get
1370 * woken up with their sleep average not weighted in an
1373 if (old_state & TASK_NONINTERACTIVE)
1374 p->sleep_type = SLEEP_NONINTERACTIVE;
1377 activate_task(p, rq, cpu == this_cpu);
1379 * Sync wakeups (i.e. those types of wakeups where the waker
1380 * has indicated that it will leave the CPU in short order)
1381 * don't trigger a preemption, if the woken up task will run on
1382 * this cpu. (in this case the 'I will reschedule' promise of
1383 * the waker guarantees that the freshly woken up task is going
1384 * to be considered on this CPU.)
1386 if (!sync || cpu != this_cpu) {
1387 if (TASK_PREEMPTS_CURR(p, rq))
1388 resched_task(rq->curr);
1393 p->state = TASK_RUNNING;
1395 task_rq_unlock(rq, &flags);
1400 int fastcall wake_up_process(task_t *p)
1402 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1403 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1406 EXPORT_SYMBOL(wake_up_process);
1408 int fastcall wake_up_state(task_t *p, unsigned int state)
1410 return try_to_wake_up(p, state, 0);
1414 * Perform scheduler related setup for a newly forked process p.
1415 * p is forked by current.
1417 void fastcall sched_fork(task_t *p, int clone_flags)
1419 int cpu = get_cpu();
1422 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1424 set_task_cpu(p, cpu);
1427 * We mark the process as running here, but have not actually
1428 * inserted it onto the runqueue yet. This guarantees that
1429 * nobody will actually run it, and a signal or other external
1430 * event cannot wake it up and insert it on the runqueue either.
1432 p->state = TASK_RUNNING;
1433 INIT_LIST_HEAD(&p->run_list);
1435 #ifdef CONFIG_SCHEDSTATS
1436 memset(&p->sched_info, 0, sizeof(p->sched_info));
1438 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1441 #ifdef CONFIG_PREEMPT
1442 /* Want to start with kernel preemption disabled. */
1443 task_thread_info(p)->preempt_count = 1;
1446 * Share the timeslice between parent and child, thus the
1447 * total amount of pending timeslices in the system doesn't change,
1448 * resulting in more scheduling fairness.
1450 local_irq_disable();
1451 p->time_slice = (current->time_slice + 1) >> 1;
1453 * The remainder of the first timeslice might be recovered by
1454 * the parent if the child exits early enough.
1456 p->first_time_slice = 1;
1457 current->time_slice >>= 1;
1458 p->timestamp = sched_clock();
1459 if (unlikely(!current->time_slice)) {
1461 * This case is rare, it happens when the parent has only
1462 * a single jiffy left from its timeslice. Taking the
1463 * runqueue lock is not a problem.
1465 current->time_slice = 1;
1473 * wake_up_new_task - wake up a newly created task for the first time.
1475 * This function will do some initial scheduler statistics housekeeping
1476 * that must be done for every newly created context, then puts the task
1477 * on the runqueue and wakes it.
1479 void fastcall wake_up_new_task(task_t *p, unsigned long clone_flags)
1481 unsigned long flags;
1483 runqueue_t *rq, *this_rq;
1485 rq = task_rq_lock(p, &flags);
1486 BUG_ON(p->state != TASK_RUNNING);
1487 this_cpu = smp_processor_id();
1491 * We decrease the sleep average of forking parents
1492 * and children as well, to keep max-interactive tasks
1493 * from forking tasks that are max-interactive. The parent
1494 * (current) is done further down, under its lock.
1496 p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1497 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1499 p->prio = effective_prio(p);
1501 vx_activate_task(p);
1502 if (likely(cpu == this_cpu)) {
1503 if (!(clone_flags & CLONE_VM)) {
1505 * The VM isn't cloned, so we're in a good position to
1506 * do child-runs-first in anticipation of an exec. This
1507 * usually avoids a lot of COW overhead.
1509 if (unlikely(!current->array))
1510 __activate_task(p, rq);
1512 p->prio = current->prio;
1513 BUG_ON(p->state & TASK_ONHOLD);
1514 list_add_tail(&p->run_list, ¤t->run_list);
1515 p->array = current->array;
1516 p->array->nr_active++;
1521 /* Run child last */
1522 __activate_task(p, rq);
1524 * We skip the following code due to cpu == this_cpu
1526 * task_rq_unlock(rq, &flags);
1527 * this_rq = task_rq_lock(current, &flags);
1531 this_rq = cpu_rq(this_cpu);
1534 * Not the local CPU - must adjust timestamp. This should
1535 * get optimised away in the !CONFIG_SMP case.
1537 p->timestamp = (p->timestamp - this_rq->timestamp_last_tick)
1538 + rq->timestamp_last_tick;
1539 __activate_task(p, rq);
1540 if (TASK_PREEMPTS_CURR(p, rq))
1541 resched_task(rq->curr);
1544 * Parent and child are on different CPUs, now get the
1545 * parent runqueue to update the parent's ->sleep_avg:
1547 task_rq_unlock(rq, &flags);
1548 this_rq = task_rq_lock(current, &flags);
1550 current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1551 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1552 task_rq_unlock(this_rq, &flags);
1556 * Potentially available exiting-child timeslices are
1557 * retrieved here - this way the parent does not get
1558 * penalized for creating too many threads.
1560 * (this cannot be used to 'generate' timeslices
1561 * artificially, because any timeslice recovered here
1562 * was given away by the parent in the first place.)
1564 void fastcall sched_exit(task_t *p)
1566 unsigned long flags;
1570 * If the child was a (relative-) CPU hog then decrease
1571 * the sleep_avg of the parent as well.
1573 rq = task_rq_lock(p->parent, &flags);
1574 if (p->first_time_slice && task_cpu(p) == task_cpu(p->parent)) {
1575 p->parent->time_slice += p->time_slice;
1576 if (unlikely(p->parent->time_slice > task_timeslice(p)))
1577 p->parent->time_slice = task_timeslice(p);
1579 if (p->sleep_avg < p->parent->sleep_avg)
1580 p->parent->sleep_avg = p->parent->sleep_avg /
1581 (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg /
1583 task_rq_unlock(rq, &flags);
1587 * prepare_task_switch - prepare to switch tasks
1588 * @rq: the runqueue preparing to switch
1589 * @next: the task we are going to switch to.
1591 * This is called with the rq lock held and interrupts off. It must
1592 * be paired with a subsequent finish_task_switch after the context
1595 * prepare_task_switch sets up locking and calls architecture specific
1598 static inline void prepare_task_switch(runqueue_t *rq, task_t *next)
1600 prepare_lock_switch(rq, next);
1601 prepare_arch_switch(next);
1605 * finish_task_switch - clean up after a task-switch
1606 * @rq: runqueue associated with task-switch
1607 * @prev: the thread we just switched away from.
1609 * finish_task_switch must be called after the context switch, paired
1610 * with a prepare_task_switch call before the context switch.
1611 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1612 * and do any other architecture-specific cleanup actions.
1614 * Note that we may have delayed dropping an mm in context_switch(). If
1615 * so, we finish that here outside of the runqueue lock. (Doing it
1616 * with the lock held can cause deadlocks; see schedule() for
1619 static inline void finish_task_switch(runqueue_t *rq, task_t *prev)
1620 __releases(rq->lock)
1622 struct mm_struct *mm = rq->prev_mm;
1623 unsigned long prev_task_flags;
1628 * A task struct has one reference for the use as "current".
1629 * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1630 * calls schedule one last time. The schedule call will never return,
1631 * and the scheduled task must drop that reference.
1632 * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1633 * still held, otherwise prev could be scheduled on another cpu, die
1634 * there before we look at prev->state, and then the reference would
1636 * Manfred Spraul <manfred@colorfullife.com>
1638 prev_task_flags = prev->flags;
1639 finish_arch_switch(prev);
1640 finish_lock_switch(rq, prev);
1643 if (unlikely(prev_task_flags & PF_DEAD)) {
1645 * Remove function-return probe instances associated with this
1646 * task and put them back on the free list.
1648 kprobe_flush_task(prev);
1649 put_task_struct(prev);
1654 * schedule_tail - first thing a freshly forked thread must call.
1655 * @prev: the thread we just switched away from.
1657 asmlinkage void schedule_tail(task_t *prev)
1658 __releases(rq->lock)
1660 runqueue_t *rq = this_rq();
1661 finish_task_switch(rq, prev);
1662 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1663 /* In this case, finish_task_switch does not reenable preemption */
1666 if (current->set_child_tid)
1667 put_user(current->pid, current->set_child_tid);
1671 * context_switch - switch to the new MM and the new
1672 * thread's register state.
1675 task_t * context_switch(runqueue_t *rq, task_t *prev, task_t *next)
1677 struct mm_struct *mm = next->mm;
1678 struct mm_struct *oldmm = prev->active_mm;
1680 if (unlikely(!mm)) {
1681 next->active_mm = oldmm;
1682 atomic_inc(&oldmm->mm_count);
1683 enter_lazy_tlb(oldmm, next);
1685 switch_mm(oldmm, mm, next);
1687 if (unlikely(!prev->mm)) {
1688 prev->active_mm = NULL;
1689 WARN_ON(rq->prev_mm);
1690 rq->prev_mm = oldmm;
1693 /* Here we just switch the register state and the stack. */
1694 switch_to(prev, next, prev);
1700 * nr_running, nr_uninterruptible and nr_context_switches:
1702 * externally visible scheduler statistics: current number of runnable
1703 * threads, current number of uninterruptible-sleeping threads, total
1704 * number of context switches performed since bootup.
1706 unsigned long nr_running(void)
1708 unsigned long i, sum = 0;
1710 for_each_online_cpu(i)
1711 sum += cpu_rq(i)->nr_running;
1716 unsigned long nr_uninterruptible(void)
1718 unsigned long i, sum = 0;
1720 for_each_possible_cpu(i)
1721 sum += cpu_rq(i)->nr_uninterruptible;
1724 * Since we read the counters lockless, it might be slightly
1725 * inaccurate. Do not allow it to go below zero though:
1727 if (unlikely((long)sum < 0))
1733 unsigned long long nr_context_switches(void)
1735 unsigned long long i, sum = 0;
1737 for_each_possible_cpu(i)
1738 sum += cpu_rq(i)->nr_switches;
1743 unsigned long nr_iowait(void)
1745 unsigned long i, sum = 0;
1747 for_each_possible_cpu(i)
1748 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1753 unsigned long nr_active(void)
1755 unsigned long i, running = 0, uninterruptible = 0;
1757 for_each_online_cpu(i) {
1758 running += cpu_rq(i)->nr_running;
1759 uninterruptible += cpu_rq(i)->nr_uninterruptible;
1762 if (unlikely((long)uninterruptible < 0))
1763 uninterruptible = 0;
1765 return running + uninterruptible;
1771 * double_rq_lock - safely lock two runqueues
1773 * We must take them in cpu order to match code in
1774 * dependent_sleeper and wake_dependent_sleeper.
1776 * Note this does not disable interrupts like task_rq_lock,
1777 * you need to do so manually before calling.
1779 static void double_rq_lock(runqueue_t *rq1, runqueue_t *rq2)
1780 __acquires(rq1->lock)
1781 __acquires(rq2->lock)
1784 spin_lock(&rq1->lock);
1785 __acquire(rq2->lock); /* Fake it out ;) */
1787 if (rq1->cpu < rq2->cpu) {
1788 spin_lock(&rq1->lock);
1789 spin_lock(&rq2->lock);
1791 spin_lock(&rq2->lock);
1792 spin_lock(&rq1->lock);
1798 * double_rq_unlock - safely unlock two runqueues
1800 * Note this does not restore interrupts like task_rq_unlock,
1801 * you need to do so manually after calling.
1803 static void double_rq_unlock(runqueue_t *rq1, runqueue_t *rq2)
1804 __releases(rq1->lock)
1805 __releases(rq2->lock)
1807 spin_unlock(&rq1->lock);
1809 spin_unlock(&rq2->lock);
1811 __release(rq2->lock);
1815 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1817 static void double_lock_balance(runqueue_t *this_rq, runqueue_t *busiest)
1818 __releases(this_rq->lock)
1819 __acquires(busiest->lock)
1820 __acquires(this_rq->lock)
1822 if (unlikely(!spin_trylock(&busiest->lock))) {
1823 if (busiest->cpu < this_rq->cpu) {
1824 spin_unlock(&this_rq->lock);
1825 spin_lock(&busiest->lock);
1826 spin_lock(&this_rq->lock);
1828 spin_lock(&busiest->lock);
1833 * If dest_cpu is allowed for this process, migrate the task to it.
1834 * This is accomplished by forcing the cpu_allowed mask to only
1835 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
1836 * the cpu_allowed mask is restored.
1838 static void sched_migrate_task(task_t *p, int dest_cpu)
1840 migration_req_t req;
1842 unsigned long flags;
1844 rq = task_rq_lock(p, &flags);
1845 if (!cpu_isset(dest_cpu, p->cpus_allowed)
1846 || unlikely(cpu_is_offline(dest_cpu)))
1849 /* force the process onto the specified CPU */
1850 if (migrate_task(p, dest_cpu, &req)) {
1851 /* Need to wait for migration thread (might exit: take ref). */
1852 struct task_struct *mt = rq->migration_thread;
1853 get_task_struct(mt);
1854 task_rq_unlock(rq, &flags);
1855 wake_up_process(mt);
1856 put_task_struct(mt);
1857 wait_for_completion(&req.done);
1861 task_rq_unlock(rq, &flags);
1865 * sched_exec - execve() is a valuable balancing opportunity, because at
1866 * this point the task has the smallest effective memory and cache footprint.
1868 void sched_exec(void)
1870 int new_cpu, this_cpu = get_cpu();
1871 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
1873 if (new_cpu != this_cpu)
1874 sched_migrate_task(current, new_cpu);
1878 * pull_task - move a task from a remote runqueue to the local runqueue.
1879 * Both runqueues must be locked.
1882 void pull_task(runqueue_t *src_rq, prio_array_t *src_array, task_t *p,
1883 runqueue_t *this_rq, prio_array_t *this_array, int this_cpu)
1885 dequeue_task(p, src_array);
1886 src_rq->nr_running--;
1887 set_task_cpu(p, this_cpu);
1888 this_rq->nr_running++;
1889 enqueue_task(p, this_array);
1890 p->timestamp = (p->timestamp - src_rq->timestamp_last_tick)
1891 + this_rq->timestamp_last_tick;
1893 * Note that idle threads have a prio of MAX_PRIO, for this test
1894 * to be always true for them.
1896 if (TASK_PREEMPTS_CURR(p, this_rq))
1897 resched_task(this_rq->curr);
1901 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1904 int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu,
1905 struct sched_domain *sd, enum idle_type idle,
1909 * We do not migrate tasks that are:
1910 * 1) running (obviously), or
1911 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1912 * 3) are cache-hot on their current CPU.
1914 if (!cpu_isset(this_cpu, p->cpus_allowed))
1918 if (task_running(rq, p))
1922 * Aggressive migration if:
1923 * 1) task is cache cold, or
1924 * 2) too many balance attempts have failed.
1927 if (sd->nr_balance_failed > sd->cache_nice_tries)
1930 if (task_hot(p, rq->timestamp_last_tick, sd))
1936 * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq,
1937 * as part of a balancing operation within "domain". Returns the number of
1940 * Called with both runqueues locked.
1942 static int move_tasks(runqueue_t *this_rq, int this_cpu, runqueue_t *busiest,
1943 unsigned long max_nr_move, struct sched_domain *sd,
1944 enum idle_type idle, int *all_pinned)
1946 prio_array_t *array, *dst_array;
1947 struct list_head *head, *curr;
1948 int idx, pulled = 0, pinned = 0;
1951 if (max_nr_move == 0)
1957 * We first consider expired tasks. Those will likely not be
1958 * executed in the near future, and they are most likely to
1959 * be cache-cold, thus switching CPUs has the least effect
1962 if (busiest->expired->nr_active) {
1963 array = busiest->expired;
1964 dst_array = this_rq->expired;
1966 array = busiest->active;
1967 dst_array = this_rq->active;
1971 /* Start searching at priority 0: */
1975 idx = sched_find_first_bit(array->bitmap);
1977 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
1978 if (idx >= MAX_PRIO) {
1979 if (array == busiest->expired && busiest->active->nr_active) {
1980 array = busiest->active;
1981 dst_array = this_rq->active;
1987 head = array->queue + idx;
1990 tmp = list_entry(curr, task_t, run_list);
1994 if (!can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
2001 #ifdef CONFIG_SCHEDSTATS
2002 if (task_hot(tmp, busiest->timestamp_last_tick, sd))
2003 schedstat_inc(sd, lb_hot_gained[idle]);
2006 pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
2009 /* We only want to steal up to the prescribed number of tasks. */
2010 if (pulled < max_nr_move) {
2018 * Right now, this is the only place pull_task() is called,
2019 * so we can safely collect pull_task() stats here rather than
2020 * inside pull_task().
2022 schedstat_add(sd, lb_gained[idle], pulled);
2025 *all_pinned = pinned;
2030 * find_busiest_group finds and returns the busiest CPU group within the
2031 * domain. It calculates and returns the number of tasks which should be
2032 * moved to restore balance via the imbalance parameter.
2034 static struct sched_group *
2035 find_busiest_group(struct sched_domain *sd, int this_cpu,
2036 unsigned long *imbalance, enum idle_type idle, int *sd_idle)
2038 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2039 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2040 unsigned long max_pull;
2043 max_load = this_load = total_load = total_pwr = 0;
2044 if (idle == NOT_IDLE)
2045 load_idx = sd->busy_idx;
2046 else if (idle == NEWLY_IDLE)
2047 load_idx = sd->newidle_idx;
2049 load_idx = sd->idle_idx;
2056 local_group = cpu_isset(this_cpu, group->cpumask);
2058 /* Tally up the load of all CPUs in the group */
2061 for_each_cpu_mask(i, group->cpumask) {
2062 if (*sd_idle && !idle_cpu(i))
2065 /* Bias balancing toward cpus of our domain */
2067 load = target_load(i, load_idx);
2069 load = source_load(i, load_idx);
2074 total_load += avg_load;
2075 total_pwr += group->cpu_power;
2077 /* Adjust by relative CPU power of the group */
2078 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
2081 this_load = avg_load;
2083 } else if (avg_load > max_load) {
2084 max_load = avg_load;
2087 group = group->next;
2088 } while (group != sd->groups);
2090 if (!busiest || this_load >= max_load || max_load <= SCHED_LOAD_SCALE)
2093 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2095 if (this_load >= avg_load ||
2096 100*max_load <= sd->imbalance_pct*this_load)
2100 * We're trying to get all the cpus to the average_load, so we don't
2101 * want to push ourselves above the average load, nor do we wish to
2102 * reduce the max loaded cpu below the average load, as either of these
2103 * actions would just result in more rebalancing later, and ping-pong
2104 * tasks around. Thus we look for the minimum possible imbalance.
2105 * Negative imbalances (*we* are more loaded than anyone else) will
2106 * be counted as no imbalance for these purposes -- we can't fix that
2107 * by pulling tasks to us. Be careful of negative numbers as they'll
2108 * appear as very large values with unsigned longs.
2111 /* Don't want to pull so many tasks that a group would go idle */
2112 max_pull = min(max_load - avg_load, max_load - SCHED_LOAD_SCALE);
2114 /* How much load to actually move to equalise the imbalance */
2115 *imbalance = min(max_pull * busiest->cpu_power,
2116 (avg_load - this_load) * this->cpu_power)
2119 if (*imbalance < SCHED_LOAD_SCALE) {
2120 unsigned long pwr_now = 0, pwr_move = 0;
2123 if (max_load - this_load >= SCHED_LOAD_SCALE*2) {
2129 * OK, we don't have enough imbalance to justify moving tasks,
2130 * however we may be able to increase total CPU power used by
2134 pwr_now += busiest->cpu_power*min(SCHED_LOAD_SCALE, max_load);
2135 pwr_now += this->cpu_power*min(SCHED_LOAD_SCALE, this_load);
2136 pwr_now /= SCHED_LOAD_SCALE;
2138 /* Amount of load we'd subtract */
2139 tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/busiest->cpu_power;
2141 pwr_move += busiest->cpu_power*min(SCHED_LOAD_SCALE,
2144 /* Amount of load we'd add */
2145 if (max_load*busiest->cpu_power <
2146 SCHED_LOAD_SCALE*SCHED_LOAD_SCALE)
2147 tmp = max_load*busiest->cpu_power/this->cpu_power;
2149 tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/this->cpu_power;
2150 pwr_move += this->cpu_power*min(SCHED_LOAD_SCALE, this_load + tmp);
2151 pwr_move /= SCHED_LOAD_SCALE;
2153 /* Move if we gain throughput */
2154 if (pwr_move <= pwr_now)
2161 /* Get rid of the scaling factor, rounding down as we divide */
2162 *imbalance = *imbalance / SCHED_LOAD_SCALE;
2172 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2174 static runqueue_t *find_busiest_queue(struct sched_group *group,
2175 enum idle_type idle)
2177 unsigned long load, max_load = 0;
2178 runqueue_t *busiest = NULL;
2181 for_each_cpu_mask(i, group->cpumask) {
2182 load = source_load(i, 0);
2184 if (load > max_load) {
2186 busiest = cpu_rq(i);
2194 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2195 * so long as it is large enough.
2197 #define MAX_PINNED_INTERVAL 512
2200 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2201 * tasks if there is an imbalance.
2203 * Called with this_rq unlocked.
2205 static int load_balance(int this_cpu, runqueue_t *this_rq,
2206 struct sched_domain *sd, enum idle_type idle)
2208 struct sched_group *group;
2209 runqueue_t *busiest;
2210 unsigned long imbalance;
2211 int nr_moved, all_pinned = 0;
2212 int active_balance = 0;
2215 if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER)
2218 schedstat_inc(sd, lb_cnt[idle]);
2220 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle);
2222 schedstat_inc(sd, lb_nobusyg[idle]);
2226 busiest = find_busiest_queue(group, idle);
2228 schedstat_inc(sd, lb_nobusyq[idle]);
2232 BUG_ON(busiest == this_rq);
2234 schedstat_add(sd, lb_imbalance[idle], imbalance);
2237 if (busiest->nr_running > 1) {
2239 * Attempt to move tasks. If find_busiest_group has found
2240 * an imbalance but busiest->nr_running <= 1, the group is
2241 * still unbalanced. nr_moved simply stays zero, so it is
2242 * correctly treated as an imbalance.
2244 double_rq_lock(this_rq, busiest);
2245 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2246 imbalance, sd, idle, &all_pinned);
2247 double_rq_unlock(this_rq, busiest);
2249 /* All tasks on this runqueue were pinned by CPU affinity */
2250 if (unlikely(all_pinned))
2255 schedstat_inc(sd, lb_failed[idle]);
2256 sd->nr_balance_failed++;
2258 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2260 spin_lock(&busiest->lock);
2262 /* don't kick the migration_thread, if the curr
2263 * task on busiest cpu can't be moved to this_cpu
2265 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2266 spin_unlock(&busiest->lock);
2268 goto out_one_pinned;
2271 if (!busiest->active_balance) {
2272 busiest->active_balance = 1;
2273 busiest->push_cpu = this_cpu;
2276 spin_unlock(&busiest->lock);
2278 wake_up_process(busiest->migration_thread);
2281 * We've kicked active balancing, reset the failure
2284 sd->nr_balance_failed = sd->cache_nice_tries+1;
2287 sd->nr_balance_failed = 0;
2289 if (likely(!active_balance)) {
2290 /* We were unbalanced, so reset the balancing interval */
2291 sd->balance_interval = sd->min_interval;
2294 * If we've begun active balancing, start to back off. This
2295 * case may not be covered by the all_pinned logic if there
2296 * is only 1 task on the busy runqueue (because we don't call
2299 if (sd->balance_interval < sd->max_interval)
2300 sd->balance_interval *= 2;
2303 if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2308 schedstat_inc(sd, lb_balanced[idle]);
2310 sd->nr_balance_failed = 0;
2313 /* tune up the balancing interval */
2314 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2315 (sd->balance_interval < sd->max_interval))
2316 sd->balance_interval *= 2;
2318 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2324 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2325 * tasks if there is an imbalance.
2327 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2328 * this_rq is locked.
2330 static int load_balance_newidle(int this_cpu, runqueue_t *this_rq,
2331 struct sched_domain *sd)
2333 struct sched_group *group;
2334 runqueue_t *busiest = NULL;
2335 unsigned long imbalance;
2339 if (sd->flags & SD_SHARE_CPUPOWER)
2342 schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
2343 group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE, &sd_idle);
2345 schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]);
2349 busiest = find_busiest_queue(group, NEWLY_IDLE);
2351 schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]);
2355 BUG_ON(busiest == this_rq);
2357 schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
2360 if (busiest->nr_running > 1) {
2361 /* Attempt to move tasks */
2362 double_lock_balance(this_rq, busiest);
2363 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2364 imbalance, sd, NEWLY_IDLE, NULL);
2365 spin_unlock(&busiest->lock);
2369 schedstat_inc(sd, lb_failed[NEWLY_IDLE]);
2370 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2373 sd->nr_balance_failed = 0;
2378 schedstat_inc(sd, lb_balanced[NEWLY_IDLE]);
2379 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2381 sd->nr_balance_failed = 0;
2386 * idle_balance is called by schedule() if this_cpu is about to become
2387 * idle. Attempts to pull tasks from other CPUs.
2389 static void idle_balance(int this_cpu, runqueue_t *this_rq)
2391 struct sched_domain *sd;
2393 for_each_domain(this_cpu, sd) {
2394 if (sd->flags & SD_BALANCE_NEWIDLE) {
2395 if (load_balance_newidle(this_cpu, this_rq, sd)) {
2396 /* We've pulled tasks over so stop searching */
2404 * active_load_balance is run by migration threads. It pushes running tasks
2405 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2406 * running on each physical CPU where possible, and avoids physical /
2407 * logical imbalances.
2409 * Called with busiest_rq locked.
2411 static void active_load_balance(runqueue_t *busiest_rq, int busiest_cpu)
2413 struct sched_domain *sd;
2414 runqueue_t *target_rq;
2415 int target_cpu = busiest_rq->push_cpu;
2417 if (busiest_rq->nr_running <= 1)
2418 /* no task to move */
2421 target_rq = cpu_rq(target_cpu);
2424 * This condition is "impossible", if it occurs
2425 * we need to fix it. Originally reported by
2426 * Bjorn Helgaas on a 128-cpu setup.
2428 BUG_ON(busiest_rq == target_rq);
2430 /* move a task from busiest_rq to target_rq */
2431 double_lock_balance(busiest_rq, target_rq);
2433 /* Search for an sd spanning us and the target CPU. */
2434 for_each_domain(target_cpu, sd)
2435 if ((sd->flags & SD_LOAD_BALANCE) &&
2436 cpu_isset(busiest_cpu, sd->span))
2439 if (unlikely(sd == NULL))
2442 schedstat_inc(sd, alb_cnt);
2444 if (move_tasks(target_rq, target_cpu, busiest_rq, 1, sd, SCHED_IDLE, NULL))
2445 schedstat_inc(sd, alb_pushed);
2447 schedstat_inc(sd, alb_failed);
2449 spin_unlock(&target_rq->lock);
2453 * rebalance_tick will get called every timer tick, on every CPU.
2455 * It checks each scheduling domain to see if it is due to be balanced,
2456 * and initiates a balancing operation if so.
2458 * Balancing parameters are set up in arch_init_sched_domains.
2461 /* Don't have all balancing operations going off at once */
2462 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2464 static void rebalance_tick(int this_cpu, runqueue_t *this_rq,
2465 enum idle_type idle)
2467 unsigned long old_load, this_load;
2468 unsigned long j = jiffies + CPU_OFFSET(this_cpu);
2469 struct sched_domain *sd;
2472 this_load = this_rq->nr_running * SCHED_LOAD_SCALE;
2473 /* Update our load */
2474 for (i = 0; i < 3; i++) {
2475 unsigned long new_load = this_load;
2477 old_load = this_rq->cpu_load[i];
2479 * Round up the averaging division if load is increasing. This
2480 * prevents us from getting stuck on 9 if the load is 10, for
2483 if (new_load > old_load)
2484 new_load += scale-1;
2485 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale;
2488 for_each_domain(this_cpu, sd) {
2489 unsigned long interval;
2491 if (!(sd->flags & SD_LOAD_BALANCE))
2494 interval = sd->balance_interval;
2495 if (idle != SCHED_IDLE)
2496 interval *= sd->busy_factor;
2498 /* scale ms to jiffies */
2499 interval = msecs_to_jiffies(interval);
2500 if (unlikely(!interval))
2503 if (j - sd->last_balance >= interval) {
2504 if (load_balance(this_cpu, this_rq, sd, idle)) {
2506 * We've pulled tasks over so either we're no
2507 * longer idle, or one of our SMT siblings is
2512 sd->last_balance += interval;
2518 * on UP we do not need to balance between CPUs:
2520 static inline void rebalance_tick(int cpu, runqueue_t *rq, enum idle_type idle)
2523 static inline void idle_balance(int cpu, runqueue_t *rq)
2528 static inline int wake_priority_sleeper(runqueue_t *rq)
2531 #ifdef CONFIG_SCHED_SMT
2532 spin_lock(&rq->lock);
2534 * If an SMT sibling task has been put to sleep for priority
2535 * reasons reschedule the idle task to see if it can now run.
2537 if (rq->nr_running) {
2538 resched_task(rq->idle);
2541 spin_unlock(&rq->lock);
2546 DEFINE_PER_CPU(struct kernel_stat, kstat);
2548 EXPORT_PER_CPU_SYMBOL(kstat);
2551 * This is called on clock ticks and on context switches.
2552 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2554 static inline void update_cpu_clock(task_t *p, runqueue_t *rq,
2555 unsigned long long now)
2557 unsigned long long last = max(p->timestamp, rq->timestamp_last_tick);
2558 p->sched_time += now - last;
2562 * Return current->sched_time plus any more ns on the sched_clock
2563 * that have not yet been banked.
2565 unsigned long long current_sched_time(const task_t *tsk)
2567 unsigned long long ns;
2568 unsigned long flags;
2569 local_irq_save(flags);
2570 ns = max(tsk->timestamp, task_rq(tsk)->timestamp_last_tick);
2571 ns = tsk->sched_time + (sched_clock() - ns);
2572 local_irq_restore(flags);
2577 * We place interactive tasks back into the active array, if possible.
2579 * To guarantee that this does not starve expired tasks we ignore the
2580 * interactivity of a task if the first expired task had to wait more
2581 * than a 'reasonable' amount of time. This deadline timeout is
2582 * load-dependent, as the frequency of array switched decreases with
2583 * increasing number of running tasks. We also ignore the interactivity
2584 * if a better static_prio task has expired:
2586 #define EXPIRED_STARVING(rq) \
2587 ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2588 (jiffies - (rq)->expired_timestamp >= \
2589 STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2590 ((rq)->curr->static_prio > (rq)->best_expired_prio))
2593 * Account user cpu time to a process.
2594 * @p: the process that the cpu time gets accounted to
2595 * @hardirq_offset: the offset to subtract from hardirq_count()
2596 * @cputime: the cpu time spent in user space since the last update
2598 void account_user_time(struct task_struct *p, cputime_t cputime)
2600 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2601 struct vx_info *vxi = p->vx_info; /* p is _always_ current */
2603 int nice = (TASK_NICE(p) > 0);
2605 p->utime = cputime_add(p->utime, cputime);
2606 vx_account_user(vxi, cputime, nice);
2608 /* Add user time to cpustat. */
2609 tmp = cputime_to_cputime64(cputime);
2611 cpustat->nice = cputime64_add(cpustat->nice, tmp);
2613 cpustat->user = cputime64_add(cpustat->user, tmp);
2617 * Account system cpu time to a process.
2618 * @p: the process that the cpu time gets accounted to
2619 * @hardirq_offset: the offset to subtract from hardirq_count()
2620 * @cputime: the cpu time spent in kernel space since the last update
2622 void account_system_time(struct task_struct *p, int hardirq_offset,
2625 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2626 struct vx_info *vxi = p->vx_info; /* p is _always_ current */
2627 runqueue_t *rq = this_rq();
2630 p->stime = cputime_add(p->stime, cputime);
2631 vx_account_system(vxi, cputime, (p == rq->idle));
2633 /* Add system time to cpustat. */
2634 tmp = cputime_to_cputime64(cputime);
2635 if (hardirq_count() - hardirq_offset)
2636 cpustat->irq = cputime64_add(cpustat->irq, tmp);
2637 else if (softirq_count())
2638 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
2639 else if (p != rq->idle)
2640 cpustat->system = cputime64_add(cpustat->system, tmp);
2641 else if (atomic_read(&rq->nr_iowait) > 0)
2642 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2644 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2645 /* Account for system time used */
2646 acct_update_integrals(p);
2650 * Account for involuntary wait time.
2651 * @p: the process from which the cpu time has been stolen
2652 * @steal: the cpu time spent in involuntary wait
2654 void account_steal_time(struct task_struct *p, cputime_t steal)
2656 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2657 cputime64_t tmp = cputime_to_cputime64(steal);
2658 runqueue_t *rq = this_rq();
2660 if (p == rq->idle) {
2661 p->stime = cputime_add(p->stime, steal);
2662 if (atomic_read(&rq->nr_iowait) > 0)
2663 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2665 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2667 cpustat->steal = cputime64_add(cpustat->steal, tmp);
2671 * This function gets called by the timer code, with HZ frequency.
2672 * We call it with interrupts disabled.
2674 * It also gets called by the fork code, when changing the parent's
2677 void scheduler_tick(void)
2679 int cpu = smp_processor_id();
2680 runqueue_t *rq = this_rq();
2681 task_t *p = current;
2682 unsigned long long now = sched_clock();
2684 update_cpu_clock(p, rq, now);
2686 rq->timestamp_last_tick = now;
2688 #if defined(CONFIG_VSERVER_HARDCPU) && defined(CONFIG_VSERVER_ACB_SCHED)
2689 vx_scheduler_tick();
2692 if (p == rq->idle) {
2693 if (wake_priority_sleeper(rq))
2695 #ifdef CONFIG_VSERVER_HARDCPU_IDLE
2696 if (!--rq->idle_tokens && !list_empty(&rq->hold_queue))
2699 rebalance_tick(cpu, rq, SCHED_IDLE);
2703 /* Task might have expired already, but not scheduled off yet */
2704 if (p->array != rq->active) {
2705 set_tsk_need_resched(p);
2708 spin_lock(&rq->lock);
2710 * The task was running during this tick - update the
2711 * time slice counter. Note: we do not update a thread's
2712 * priority until it either goes to sleep or uses up its
2713 * timeslice. This makes it possible for interactive tasks
2714 * to use up their timeslices at their highest priority levels.
2718 * RR tasks need a special form of timeslice management.
2719 * FIFO tasks have no timeslices.
2721 if ((p->policy == SCHED_RR) && !--p->time_slice) {
2722 p->time_slice = task_timeslice(p);
2723 p->first_time_slice = 0;
2724 set_tsk_need_resched(p);
2726 /* put it at the end of the queue: */
2727 requeue_task(p, rq->active);
2731 if (vx_need_resched(p)) {
2732 dequeue_task(p, rq->active);
2733 set_tsk_need_resched(p);
2734 p->prio = effective_prio(p);
2735 p->time_slice = task_timeslice(p);
2736 p->first_time_slice = 0;
2738 if (!rq->expired_timestamp)
2739 rq->expired_timestamp = jiffies;
2740 if (!TASK_INTERACTIVE(p) || EXPIRED_STARVING(rq)) {
2741 enqueue_task(p, rq->expired);
2742 if (p->static_prio < rq->best_expired_prio)
2743 rq->best_expired_prio = p->static_prio;
2745 enqueue_task(p, rq->active);
2748 * Prevent a too long timeslice allowing a task to monopolize
2749 * the CPU. We do this by splitting up the timeslice into
2752 * Note: this does not mean the task's timeslices expire or
2753 * get lost in any way, they just might be preempted by
2754 * another task of equal priority. (one with higher
2755 * priority would have preempted this task already.) We
2756 * requeue this task to the end of the list on this priority
2757 * level, which is in essence a round-robin of tasks with
2760 * This only applies to tasks in the interactive
2761 * delta range with at least TIMESLICE_GRANULARITY to requeue.
2763 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
2764 p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
2765 (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
2766 (p->array == rq->active)) {
2768 requeue_task(p, rq->active);
2769 set_tsk_need_resched(p);
2773 spin_unlock(&rq->lock);
2775 rebalance_tick(cpu, rq, NOT_IDLE);
2778 #ifdef CONFIG_SCHED_SMT
2779 static inline void wakeup_busy_runqueue(runqueue_t *rq)
2781 /* If an SMT runqueue is sleeping due to priority reasons wake it up */
2782 if (rq->curr == rq->idle && rq->nr_running)
2783 resched_task(rq->idle);
2786 static void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2788 struct sched_domain *tmp, *sd = NULL;
2789 cpumask_t sibling_map;
2792 for_each_domain(this_cpu, tmp)
2793 if (tmp->flags & SD_SHARE_CPUPOWER)
2800 * Unlock the current runqueue because we have to lock in
2801 * CPU order to avoid deadlocks. Caller knows that we might
2802 * unlock. We keep IRQs disabled.
2804 spin_unlock(&this_rq->lock);
2806 sibling_map = sd->span;
2808 for_each_cpu_mask(i, sibling_map)
2809 spin_lock(&cpu_rq(i)->lock);
2811 * We clear this CPU from the mask. This both simplifies the
2812 * inner loop and keps this_rq locked when we exit:
2814 cpu_clear(this_cpu, sibling_map);
2816 for_each_cpu_mask(i, sibling_map) {
2817 runqueue_t *smt_rq = cpu_rq(i);
2819 wakeup_busy_runqueue(smt_rq);
2822 for_each_cpu_mask(i, sibling_map)
2823 spin_unlock(&cpu_rq(i)->lock);
2825 * We exit with this_cpu's rq still held and IRQs
2831 * number of 'lost' timeslices this task wont be able to fully
2832 * utilize, if another task runs on a sibling. This models the
2833 * slowdown effect of other tasks running on siblings:
2835 static inline unsigned long smt_slice(task_t *p, struct sched_domain *sd)
2837 return p->time_slice * (100 - sd->per_cpu_gain) / 100;
2840 static int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2842 struct sched_domain *tmp, *sd = NULL;
2843 cpumask_t sibling_map;
2844 prio_array_t *array;
2848 for_each_domain(this_cpu, tmp)
2849 if (tmp->flags & SD_SHARE_CPUPOWER)
2856 * The same locking rules and details apply as for
2857 * wake_sleeping_dependent():
2859 spin_unlock(&this_rq->lock);
2860 sibling_map = sd->span;
2861 for_each_cpu_mask(i, sibling_map)
2862 spin_lock(&cpu_rq(i)->lock);
2863 cpu_clear(this_cpu, sibling_map);
2866 * Establish next task to be run - it might have gone away because
2867 * we released the runqueue lock above:
2869 if (!this_rq->nr_running)
2871 array = this_rq->active;
2872 if (!array->nr_active)
2873 array = this_rq->expired;
2874 BUG_ON(!array->nr_active);
2876 p = list_entry(array->queue[sched_find_first_bit(array->bitmap)].next,
2879 for_each_cpu_mask(i, sibling_map) {
2880 runqueue_t *smt_rq = cpu_rq(i);
2881 task_t *smt_curr = smt_rq->curr;
2883 /* Kernel threads do not participate in dependent sleeping */
2884 if (!p->mm || !smt_curr->mm || rt_task(p))
2885 goto check_smt_task;
2888 * If a user task with lower static priority than the
2889 * running task on the SMT sibling is trying to schedule,
2890 * delay it till there is proportionately less timeslice
2891 * left of the sibling task to prevent a lower priority
2892 * task from using an unfair proportion of the
2893 * physical cpu's resources. -ck
2895 if (rt_task(smt_curr)) {
2897 * With real time tasks we run non-rt tasks only
2898 * per_cpu_gain% of the time.
2900 if ((jiffies % DEF_TIMESLICE) >
2901 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
2904 if (smt_curr->static_prio < p->static_prio &&
2905 !TASK_PREEMPTS_CURR(p, smt_rq) &&
2906 smt_slice(smt_curr, sd) > task_timeslice(p))
2910 if ((!smt_curr->mm && smt_curr != smt_rq->idle) ||
2914 wakeup_busy_runqueue(smt_rq);
2919 * Reschedule a lower priority task on the SMT sibling for
2920 * it to be put to sleep, or wake it up if it has been put to
2921 * sleep for priority reasons to see if it should run now.
2924 if ((jiffies % DEF_TIMESLICE) >
2925 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
2926 resched_task(smt_curr);
2928 if (TASK_PREEMPTS_CURR(p, smt_rq) &&
2929 smt_slice(p, sd) > task_timeslice(smt_curr))
2930 resched_task(smt_curr);
2932 wakeup_busy_runqueue(smt_rq);
2936 for_each_cpu_mask(i, sibling_map)
2937 spin_unlock(&cpu_rq(i)->lock);
2941 static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2945 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2951 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
2953 void fastcall add_preempt_count(int val)
2958 BUG_ON((preempt_count() < 0));
2959 preempt_count() += val;
2961 * Spinlock count overflowing soon?
2963 BUG_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10);
2965 EXPORT_SYMBOL(add_preempt_count);
2967 void fastcall sub_preempt_count(int val)
2972 BUG_ON(val > preempt_count());
2974 * Is the spinlock portion underflowing?
2976 BUG_ON((val < PREEMPT_MASK) && !(preempt_count() & PREEMPT_MASK));
2977 preempt_count() -= val;
2979 EXPORT_SYMBOL(sub_preempt_count);
2983 static inline int interactive_sleep(enum sleep_type sleep_type)
2985 return (sleep_type == SLEEP_INTERACTIVE ||
2986 sleep_type == SLEEP_INTERRUPTED);
2990 * schedule() is the main scheduler function.
2992 asmlinkage void __sched schedule(void)
2995 task_t *prev, *next;
2997 prio_array_t *array;
2998 struct list_head *queue;
2999 unsigned long long now;
3000 unsigned long run_time;
3001 int cpu, idx, new_prio;
3002 struct vx_info *vxi;
3003 #ifdef CONFIG_VSERVER_HARDCPU
3005 # ifdef CONFIG_VSERVER_ACB_SCHED
3006 int min_guarantee_ticks = VX_INVALID_TICKS;
3007 int min_best_effort_ticks = VX_INVALID_TICKS;
3012 * Test if we are atomic. Since do_exit() needs to call into
3013 * schedule() atomically, we ignore that path for now.
3014 * Otherwise, whine if we are scheduling when we should not be.
3016 if (unlikely(in_atomic() && !current->exit_state)) {
3017 printk(KERN_ERR "BUG: scheduling while atomic: "
3019 current->comm, preempt_count(), current->pid);
3022 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3027 release_kernel_lock(prev);
3028 need_resched_nonpreemptible:
3032 * The idle thread is not allowed to schedule!
3033 * Remove this check after it has been exercised a bit.
3035 if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
3036 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
3040 schedstat_inc(rq, sched_cnt);
3041 now = sched_clock();
3042 if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
3043 run_time = now - prev->timestamp;
3044 if (unlikely((long long)(now - prev->timestamp) < 0))
3047 run_time = NS_MAX_SLEEP_AVG;
3050 * Tasks charged proportionately less run_time at high sleep_avg to
3051 * delay them losing their interactive status
3053 run_time /= (CURRENT_BONUS(prev) ? : 1);
3055 spin_lock_irq(&rq->lock);
3057 if (unlikely(prev->flags & PF_DEAD))
3058 prev->state = EXIT_DEAD;
3060 switch_count = &prev->nivcsw;
3061 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3062 switch_count = &prev->nvcsw;
3063 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3064 unlikely(signal_pending(prev))))
3065 prev->state = TASK_RUNNING;
3067 if (prev->state == TASK_UNINTERRUPTIBLE) {
3068 rq->nr_uninterruptible++;
3069 vx_uninterruptible_inc(prev);
3071 deactivate_task(prev, rq);
3075 #ifdef CONFIG_VSERVER_HARDCPU
3076 # ifdef CONFIG_VSERVER_ACB_SCHED
3079 min_guarantee_ticks = VX_INVALID_TICKS;
3080 min_best_effort_ticks = VX_INVALID_TICKS;
3083 if (!list_empty(&rq->hold_queue)) {
3084 struct list_head *l, *n;
3088 list_for_each_safe(l, n, &rq->hold_queue) {
3089 next = list_entry(l, task_t, run_list);
3090 if (vxi == next->vx_info)
3093 vxi = next->vx_info;
3094 ret = vx_tokens_recalc(vxi);
3097 vx_unhold_task(vxi, next, rq);
3100 if ((ret < 0) && (maxidle < ret))
3102 # ifdef CONFIG_VSERVER_ACB_SCHED
3104 if (IS_BEST_EFFORT(vxi)) {
3105 if (min_best_effort_ticks < ret)
3106 min_best_effort_ticks = ret;
3108 if (min_guarantee_ticks < ret)
3109 min_guarantee_ticks = ret;
3115 rq->idle_tokens = -maxidle;
3120 cpu = smp_processor_id();
3121 if (unlikely(!rq->nr_running)) {
3123 idle_balance(cpu, rq);
3124 if (!rq->nr_running) {
3126 rq->expired_timestamp = 0;
3127 wake_sleeping_dependent(cpu, rq);
3129 * wake_sleeping_dependent() might have released
3130 * the runqueue, so break out if we got new
3133 if (!rq->nr_running)
3137 if (dependent_sleeper(cpu, rq)) {
3142 * dependent_sleeper() releases and reacquires the runqueue
3143 * lock, hence go into the idle loop if the rq went
3146 if (unlikely(!rq->nr_running))
3151 if (unlikely(!array->nr_active)) {
3153 * Switch the active and expired arrays.
3155 schedstat_inc(rq, sched_switch);
3156 rq->active = rq->expired;
3157 rq->expired = array;
3159 rq->expired_timestamp = 0;
3160 rq->best_expired_prio = MAX_PRIO;
3163 idx = sched_find_first_bit(array->bitmap);
3164 queue = array->queue + idx;
3165 next = list_entry(queue->next, task_t, run_list);
3167 vxi = next->vx_info;
3168 #ifdef CONFIG_VSERVER_HARDCPU
3169 if (vx_info_flags(vxi, VXF_SCHED_PAUSE|VXF_SCHED_HARD, 0)) {
3170 int ret = vx_tokens_recalc(vxi);
3172 if (unlikely(ret <= 0)) {
3174 if ((rq->idle_tokens > -ret))
3175 rq->idle_tokens = -ret;
3176 # ifdef CONFIG_VSERVER_ACB_SCHED
3177 if (IS_BEST_EFFORT(vxi)) {
3178 if (min_best_effort_ticks < ret)
3179 min_best_effort_ticks = ret;
3181 if (min_guarantee_ticks < ret)
3182 min_guarantee_ticks = ret;
3186 vx_hold_task(vxi, next, rq);
3189 } else /* well, looks ugly but not as ugly as the ifdef-ed version */
3191 if (vx_info_flags(vxi, VXF_SCHED_PRIO, 0))
3192 vx_tokens_recalc(vxi);
3194 if (!rt_task(next) && interactive_sleep(next->sleep_type)) {
3195 unsigned long long delta = now - next->timestamp;
3196 if (unlikely((long long)(now - next->timestamp) < 0))
3199 if (next->sleep_type == SLEEP_INTERACTIVE)
3200 delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
3202 array = next->array;
3203 new_prio = recalc_task_prio(next, next->timestamp + delta);
3205 if (unlikely(next->prio != new_prio)) {
3206 dequeue_task(next, array);
3207 next->prio = new_prio;
3208 enqueue_task(next, array);
3211 next->sleep_type = SLEEP_NORMAL;
3213 #if defined(CONFIG_VSERVER_HARDCPU) && defined(CONFIG_VSERVER_ACB_SCHED)
3214 if (next == rq->idle && !list_empty(&rq->hold_queue)) {
3215 if (min_best_effort_ticks != VX_INVALID_TICKS) {
3216 vx_advance_best_effort_ticks(-min_best_effort_ticks);
3217 goto drain_hold_queue;
3219 if (min_guarantee_ticks != VX_INVALID_TICKS) {
3220 vx_advance_guaranteed_ticks(-min_guarantee_ticks);
3221 goto drain_hold_queue;
3225 if (next == rq->idle)
3226 schedstat_inc(rq, sched_goidle);
3228 prefetch_stack(next);
3229 clear_tsk_need_resched(prev);
3230 rcu_qsctr_inc(task_cpu(prev));
3232 update_cpu_clock(prev, rq, now);
3234 prev->sleep_avg -= run_time;
3235 if ((long)prev->sleep_avg <= 0)
3236 prev->sleep_avg = 0;
3237 prev->timestamp = prev->last_ran = now;
3239 sched_info_switch(prev, next);
3240 if (likely(prev != next)) {
3241 next->timestamp = now;
3246 prepare_task_switch(rq, next);
3247 prev = context_switch(rq, prev, next);
3250 * this_rq must be evaluated again because prev may have moved
3251 * CPUs since it called schedule(), thus the 'rq' on its stack
3252 * frame will be invalid.
3254 finish_task_switch(this_rq(), prev);
3256 spin_unlock_irq(&rq->lock);
3259 if (unlikely(reacquire_kernel_lock(prev) < 0))
3260 goto need_resched_nonpreemptible;
3261 preempt_enable_no_resched();
3262 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3266 EXPORT_SYMBOL(schedule);
3268 #ifdef CONFIG_PREEMPT
3270 * this is is the entry point to schedule() from in-kernel preemption
3271 * off of preempt_enable. Kernel preemptions off return from interrupt
3272 * occur there and call schedule directly.
3274 asmlinkage void __sched preempt_schedule(void)
3276 struct thread_info *ti = current_thread_info();
3277 #ifdef CONFIG_PREEMPT_BKL
3278 struct task_struct *task = current;
3279 int saved_lock_depth;
3282 * If there is a non-zero preempt_count or interrupts are disabled,
3283 * we do not want to preempt the current task. Just return..
3285 if (unlikely(ti->preempt_count || irqs_disabled()))
3289 add_preempt_count(PREEMPT_ACTIVE);
3291 * We keep the big kernel semaphore locked, but we
3292 * clear ->lock_depth so that schedule() doesnt
3293 * auto-release the semaphore:
3295 #ifdef CONFIG_PREEMPT_BKL
3296 saved_lock_depth = task->lock_depth;
3297 task->lock_depth = -1;
3300 #ifdef CONFIG_PREEMPT_BKL
3301 task->lock_depth = saved_lock_depth;
3303 sub_preempt_count(PREEMPT_ACTIVE);
3305 /* we could miss a preemption opportunity between schedule and now */
3307 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3311 EXPORT_SYMBOL(preempt_schedule);
3314 * this is is the entry point to schedule() from kernel preemption
3315 * off of irq context.
3316 * Note, that this is called and return with irqs disabled. This will
3317 * protect us against recursive calling from irq.
3319 asmlinkage void __sched preempt_schedule_irq(void)
3321 struct thread_info *ti = current_thread_info();
3322 #ifdef CONFIG_PREEMPT_BKL
3323 struct task_struct *task = current;
3324 int saved_lock_depth;
3326 /* Catch callers which need to be fixed*/
3327 BUG_ON(ti->preempt_count || !irqs_disabled());
3330 add_preempt_count(PREEMPT_ACTIVE);
3332 * We keep the big kernel semaphore locked, but we
3333 * clear ->lock_depth so that schedule() doesnt
3334 * auto-release the semaphore:
3336 #ifdef CONFIG_PREEMPT_BKL
3337 saved_lock_depth = task->lock_depth;
3338 task->lock_depth = -1;
3342 local_irq_disable();
3343 #ifdef CONFIG_PREEMPT_BKL
3344 task->lock_depth = saved_lock_depth;
3346 sub_preempt_count(PREEMPT_ACTIVE);
3348 /* we could miss a preemption opportunity between schedule and now */
3350 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3354 #endif /* CONFIG_PREEMPT */
3356 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3359 task_t *p = curr->private;
3360 return try_to_wake_up(p, mode, sync);
3363 EXPORT_SYMBOL(default_wake_function);
3366 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3367 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3368 * number) then we wake all the non-exclusive tasks and one exclusive task.
3370 * There are circumstances in which we can try to wake a task which has already
3371 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3372 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3374 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3375 int nr_exclusive, int sync, void *key)
3377 struct list_head *tmp, *next;
3379 list_for_each_safe(tmp, next, &q->task_list) {
3382 curr = list_entry(tmp, wait_queue_t, task_list);
3383 flags = curr->flags;
3384 if (curr->func(curr, mode, sync, key) &&
3385 (flags & WQ_FLAG_EXCLUSIVE) &&
3392 * __wake_up - wake up threads blocked on a waitqueue.
3394 * @mode: which threads
3395 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3396 * @key: is directly passed to the wakeup function
3398 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3399 int nr_exclusive, void *key)
3401 unsigned long flags;
3403 spin_lock_irqsave(&q->lock, flags);
3404 __wake_up_common(q, mode, nr_exclusive, 0, key);
3405 spin_unlock_irqrestore(&q->lock, flags);
3408 EXPORT_SYMBOL(__wake_up);
3411 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3413 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3415 __wake_up_common(q, mode, 1, 0, NULL);
3419 * __wake_up_sync - wake up threads blocked on a waitqueue.
3421 * @mode: which threads
3422 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3424 * The sync wakeup differs that the waker knows that it will schedule
3425 * away soon, so while the target thread will be woken up, it will not
3426 * be migrated to another CPU - ie. the two threads are 'synchronized'
3427 * with each other. This can prevent needless bouncing between CPUs.
3429 * On UP it can prevent extra preemption.
3432 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3434 unsigned long flags;
3440 if (unlikely(!nr_exclusive))
3443 spin_lock_irqsave(&q->lock, flags);
3444 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3445 spin_unlock_irqrestore(&q->lock, flags);
3447 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3449 void fastcall complete(struct completion *x)
3451 unsigned long flags;
3453 spin_lock_irqsave(&x->wait.lock, flags);
3455 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3457 spin_unlock_irqrestore(&x->wait.lock, flags);
3459 EXPORT_SYMBOL(complete);
3461 void fastcall complete_all(struct completion *x)
3463 unsigned long flags;
3465 spin_lock_irqsave(&x->wait.lock, flags);
3466 x->done += UINT_MAX/2;
3467 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3469 spin_unlock_irqrestore(&x->wait.lock, flags);
3471 EXPORT_SYMBOL(complete_all);
3473 void fastcall __sched wait_for_completion(struct completion *x)
3476 spin_lock_irq(&x->wait.lock);
3478 DECLARE_WAITQUEUE(wait, current);
3480 wait.flags |= WQ_FLAG_EXCLUSIVE;
3481 __add_wait_queue_tail(&x->wait, &wait);
3483 __set_current_state(TASK_UNINTERRUPTIBLE);
3484 spin_unlock_irq(&x->wait.lock);
3486 spin_lock_irq(&x->wait.lock);
3488 __remove_wait_queue(&x->wait, &wait);
3491 spin_unlock_irq(&x->wait.lock);
3493 EXPORT_SYMBOL(wait_for_completion);
3495 unsigned long fastcall __sched
3496 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3500 spin_lock_irq(&x->wait.lock);
3502 DECLARE_WAITQUEUE(wait, current);
3504 wait.flags |= WQ_FLAG_EXCLUSIVE;
3505 __add_wait_queue_tail(&x->wait, &wait);
3507 __set_current_state(TASK_UNINTERRUPTIBLE);
3508 spin_unlock_irq(&x->wait.lock);
3509 timeout = schedule_timeout(timeout);
3510 spin_lock_irq(&x->wait.lock);
3512 __remove_wait_queue(&x->wait, &wait);
3516 __remove_wait_queue(&x->wait, &wait);
3520 spin_unlock_irq(&x->wait.lock);
3523 EXPORT_SYMBOL(wait_for_completion_timeout);
3525 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3531 spin_lock_irq(&x->wait.lock);
3533 DECLARE_WAITQUEUE(wait, current);
3535 wait.flags |= WQ_FLAG_EXCLUSIVE;
3536 __add_wait_queue_tail(&x->wait, &wait);
3538 if (signal_pending(current)) {
3540 __remove_wait_queue(&x->wait, &wait);
3543 __set_current_state(TASK_INTERRUPTIBLE);
3544 spin_unlock_irq(&x->wait.lock);
3546 spin_lock_irq(&x->wait.lock);
3548 __remove_wait_queue(&x->wait, &wait);
3552 spin_unlock_irq(&x->wait.lock);
3556 EXPORT_SYMBOL(wait_for_completion_interruptible);
3558 unsigned long fastcall __sched
3559 wait_for_completion_interruptible_timeout(struct completion *x,
3560 unsigned long timeout)
3564 spin_lock_irq(&x->wait.lock);
3566 DECLARE_WAITQUEUE(wait, current);
3568 wait.flags |= WQ_FLAG_EXCLUSIVE;
3569 __add_wait_queue_tail(&x->wait, &wait);
3571 if (signal_pending(current)) {
3572 timeout = -ERESTARTSYS;
3573 __remove_wait_queue(&x->wait, &wait);
3576 __set_current_state(TASK_INTERRUPTIBLE);
3577 spin_unlock_irq(&x->wait.lock);
3578 timeout = schedule_timeout(timeout);
3579 spin_lock_irq(&x->wait.lock);
3581 __remove_wait_queue(&x->wait, &wait);
3585 __remove_wait_queue(&x->wait, &wait);
3589 spin_unlock_irq(&x->wait.lock);
3592 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3595 #define SLEEP_ON_VAR \
3596 unsigned long flags; \
3597 wait_queue_t wait; \
3598 init_waitqueue_entry(&wait, current);
3600 #define SLEEP_ON_HEAD \
3601 spin_lock_irqsave(&q->lock,flags); \
3602 __add_wait_queue(q, &wait); \
3603 spin_unlock(&q->lock);
3605 #define SLEEP_ON_TAIL \
3606 spin_lock_irq(&q->lock); \
3607 __remove_wait_queue(q, &wait); \
3608 spin_unlock_irqrestore(&q->lock, flags);
3610 #define SLEEP_ON_BKLCHECK \
3611 if (unlikely(!kernel_locked()) && \
3612 sleep_on_bkl_warnings < 10) { \
3613 sleep_on_bkl_warnings++; \
3617 static int sleep_on_bkl_warnings;
3619 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
3625 current->state = TASK_INTERRUPTIBLE;
3632 EXPORT_SYMBOL(interruptible_sleep_on);
3634 long fastcall __sched
3635 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3641 current->state = TASK_INTERRUPTIBLE;
3644 timeout = schedule_timeout(timeout);
3650 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3652 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3658 current->state = TASK_UNINTERRUPTIBLE;
3661 timeout = schedule_timeout(timeout);
3667 EXPORT_SYMBOL(sleep_on_timeout);
3669 void set_user_nice(task_t *p, long nice)
3671 unsigned long flags;
3672 prio_array_t *array;
3674 int old_prio, new_prio, delta;
3676 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3679 * We have to be careful, if called from sys_setpriority(),
3680 * the task might be in the middle of scheduling on another CPU.
3682 rq = task_rq_lock(p, &flags);
3684 * The RT priorities are set via sched_setscheduler(), but we still
3685 * allow the 'normal' nice value to be set - but as expected
3686 * it wont have any effect on scheduling until the task is
3687 * not SCHED_NORMAL/SCHED_BATCH:
3690 p->static_prio = NICE_TO_PRIO(nice);
3695 dequeue_task(p, array);
3698 new_prio = NICE_TO_PRIO(nice);
3699 delta = new_prio - old_prio;
3700 p->static_prio = NICE_TO_PRIO(nice);
3704 enqueue_task(p, array);
3706 * If the task increased its priority or is running and
3707 * lowered its priority, then reschedule its CPU:
3709 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3710 resched_task(rq->curr);
3713 task_rq_unlock(rq, &flags);
3716 EXPORT_SYMBOL(set_user_nice);
3719 * can_nice - check if a task can reduce its nice value
3723 int can_nice(const task_t *p, const int nice)
3725 /* convert nice value [19,-20] to rlimit style value [1,40] */
3726 int nice_rlim = 20 - nice;
3727 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
3728 capable(CAP_SYS_NICE));
3731 #ifdef __ARCH_WANT_SYS_NICE
3734 * sys_nice - change the priority of the current process.
3735 * @increment: priority increment
3737 * sys_setpriority is a more generic, but much slower function that
3738 * does similar things.
3740 asmlinkage long sys_nice(int increment)
3746 * Setpriority might change our priority at the same moment.
3747 * We don't have to worry. Conceptually one call occurs first
3748 * and we have a single winner.
3750 if (increment < -40)
3755 nice = PRIO_TO_NICE(current->static_prio) + increment;
3761 if (increment < 0 && !can_nice(current, nice))
3762 return vx_flags(VXF_IGNEG_NICE, 0) ? 0 : -EPERM;
3764 retval = security_task_setnice(current, nice);
3768 set_user_nice(current, nice);
3775 * task_prio - return the priority value of a given task.
3776 * @p: the task in question.
3778 * This is the priority value as seen by users in /proc.
3779 * RT tasks are offset by -200. Normal tasks are centered
3780 * around 0, value goes from -16 to +15.
3782 int task_prio(const task_t *p)
3784 return p->prio - MAX_RT_PRIO;
3788 * task_nice - return the nice value of a given task.
3789 * @p: the task in question.
3791 int task_nice(const task_t *p)
3793 return TASK_NICE(p);
3795 EXPORT_SYMBOL_GPL(task_nice);
3798 * idle_cpu - is a given cpu idle currently?
3799 * @cpu: the processor in question.
3801 int idle_cpu(int cpu)
3803 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
3807 * idle_task - return the idle task for a given cpu.
3808 * @cpu: the processor in question.
3810 task_t *idle_task(int cpu)
3812 return cpu_rq(cpu)->idle;
3816 * find_process_by_pid - find a process with a matching PID value.
3817 * @pid: the pid in question.
3819 static inline task_t *find_process_by_pid(pid_t pid)
3821 return pid ? find_task_by_pid(pid) : current;
3824 /* Actually do priority change: must hold rq lock. */
3825 static void __setscheduler(struct task_struct *p, int policy, int prio)
3829 p->rt_priority = prio;
3830 if (policy != SCHED_NORMAL && policy != SCHED_BATCH) {
3831 p->prio = MAX_RT_PRIO-1 - p->rt_priority;
3833 p->prio = p->static_prio;
3835 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
3837 if (policy == SCHED_BATCH)
3843 * sched_setscheduler - change the scheduling policy and/or RT priority of
3845 * @p: the task in question.
3846 * @policy: new policy.
3847 * @param: structure containing the new RT priority.
3849 int sched_setscheduler(struct task_struct *p, int policy,
3850 struct sched_param *param)
3853 int oldprio, oldpolicy = -1;
3854 prio_array_t *array;
3855 unsigned long flags;
3859 /* double check policy once rq lock held */
3861 policy = oldpolicy = p->policy;
3862 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
3863 policy != SCHED_NORMAL && policy != SCHED_BATCH)
3866 * Valid priorities for SCHED_FIFO and SCHED_RR are
3867 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
3870 if (param->sched_priority < 0 ||
3871 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3872 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3874 if ((policy == SCHED_NORMAL || policy == SCHED_BATCH)
3875 != (param->sched_priority == 0))
3879 * Allow unprivileged RT tasks to decrease priority:
3881 if (!capable(CAP_SYS_NICE)) {
3883 * can't change policy, except between SCHED_NORMAL
3886 if (((policy != SCHED_NORMAL && p->policy != SCHED_BATCH) &&
3887 (policy != SCHED_BATCH && p->policy != SCHED_NORMAL)) &&
3888 !p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
3890 /* can't increase priority */
3891 if ((policy != SCHED_NORMAL && policy != SCHED_BATCH) &&
3892 param->sched_priority > p->rt_priority &&
3893 param->sched_priority >
3894 p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
3896 /* can't change other user's priorities */
3897 if ((current->euid != p->euid) &&
3898 (current->euid != p->uid))
3902 retval = security_task_setscheduler(p, policy, param);
3906 * To be able to change p->policy safely, the apropriate
3907 * runqueue lock must be held.
3909 rq = task_rq_lock(p, &flags);
3910 /* recheck policy now with rq lock held */
3911 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3912 policy = oldpolicy = -1;
3913 task_rq_unlock(rq, &flags);
3918 deactivate_task(p, rq);
3920 __setscheduler(p, policy, param->sched_priority);
3922 vx_activate_task(p);
3923 __activate_task(p, rq);
3925 * Reschedule if we are currently running on this runqueue and
3926 * our priority decreased, or if we are not currently running on
3927 * this runqueue and our priority is higher than the current's
3929 if (task_running(rq, p)) {
3930 if (p->prio > oldprio)
3931 resched_task(rq->curr);
3932 } else if (TASK_PREEMPTS_CURR(p, rq))
3933 resched_task(rq->curr);
3935 task_rq_unlock(rq, &flags);
3938 EXPORT_SYMBOL_GPL(sched_setscheduler);
3941 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3944 struct sched_param lparam;
3945 struct task_struct *p;
3947 if (!param || pid < 0)
3949 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3951 read_lock_irq(&tasklist_lock);
3952 p = find_process_by_pid(pid);
3954 read_unlock_irq(&tasklist_lock);
3957 retval = sched_setscheduler(p, policy, &lparam);
3958 read_unlock_irq(&tasklist_lock);
3963 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3964 * @pid: the pid in question.
3965 * @policy: new policy.
3966 * @param: structure containing the new RT priority.
3968 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
3969 struct sched_param __user *param)
3971 /* negative values for policy are not valid */
3975 return do_sched_setscheduler(pid, policy, param);
3979 * sys_sched_setparam - set/change the RT priority of a thread
3980 * @pid: the pid in question.
3981 * @param: structure containing the new RT priority.
3983 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
3985 return do_sched_setscheduler(pid, -1, param);
3989 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3990 * @pid: the pid in question.
3992 asmlinkage long sys_sched_getscheduler(pid_t pid)
3994 int retval = -EINVAL;
4001 read_lock(&tasklist_lock);
4002 p = find_process_by_pid(pid);
4004 retval = security_task_getscheduler(p);
4008 read_unlock(&tasklist_lock);
4015 * sys_sched_getscheduler - get the RT priority of a thread
4016 * @pid: the pid in question.
4017 * @param: structure containing the RT priority.
4019 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4021 struct sched_param lp;
4022 int retval = -EINVAL;
4025 if (!param || pid < 0)
4028 read_lock(&tasklist_lock);
4029 p = find_process_by_pid(pid);
4034 retval = security_task_getscheduler(p);
4038 lp.sched_priority = p->rt_priority;
4039 read_unlock(&tasklist_lock);
4042 * This one might sleep, we cannot do it with a spinlock held ...
4044 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4050 read_unlock(&tasklist_lock);
4054 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4058 cpumask_t cpus_allowed;
4061 read_lock(&tasklist_lock);
4063 p = find_process_by_pid(pid);
4065 read_unlock(&tasklist_lock);
4066 unlock_cpu_hotplug();
4071 * It is not safe to call set_cpus_allowed with the
4072 * tasklist_lock held. We will bump the task_struct's
4073 * usage count and then drop tasklist_lock.
4076 read_unlock(&tasklist_lock);
4079 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4080 !capable(CAP_SYS_NICE))
4083 cpus_allowed = cpuset_cpus_allowed(p);
4084 cpus_and(new_mask, new_mask, cpus_allowed);
4085 retval = set_cpus_allowed(p, new_mask);
4089 unlock_cpu_hotplug();
4093 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4094 cpumask_t *new_mask)
4096 if (len < sizeof(cpumask_t)) {
4097 memset(new_mask, 0, sizeof(cpumask_t));
4098 } else if (len > sizeof(cpumask_t)) {
4099 len = sizeof(cpumask_t);
4101 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4105 * sys_sched_setaffinity - set the cpu affinity of a process
4106 * @pid: pid of the process
4107 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4108 * @user_mask_ptr: user-space pointer to the new cpu mask
4110 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4111 unsigned long __user *user_mask_ptr)
4116 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4120 return sched_setaffinity(pid, new_mask);
4124 * Represents all cpu's present in the system
4125 * In systems capable of hotplug, this map could dynamically grow
4126 * as new cpu's are detected in the system via any platform specific
4127 * method, such as ACPI for e.g.
4130 cpumask_t cpu_present_map __read_mostly;
4131 EXPORT_SYMBOL(cpu_present_map);
4134 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4135 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4138 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4144 read_lock(&tasklist_lock);
4147 p = find_process_by_pid(pid);
4152 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4155 read_unlock(&tasklist_lock);
4156 unlock_cpu_hotplug();
4164 * sys_sched_getaffinity - get the cpu affinity of a process
4165 * @pid: pid of the process
4166 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4167 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4169 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4170 unsigned long __user *user_mask_ptr)
4175 if (len < sizeof(cpumask_t))
4178 ret = sched_getaffinity(pid, &mask);
4182 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4185 return sizeof(cpumask_t);
4189 * sys_sched_yield - yield the current processor to other threads.
4191 * this function yields the current CPU by moving the calling thread
4192 * to the expired array. If there are no other threads running on this
4193 * CPU then this function will return.
4195 asmlinkage long sys_sched_yield(void)
4197 runqueue_t *rq = this_rq_lock();
4198 prio_array_t *array = current->array;
4199 prio_array_t *target = rq->expired;
4201 schedstat_inc(rq, yld_cnt);
4203 * We implement yielding by moving the task into the expired
4206 * (special rule: RT tasks will just roundrobin in the active
4209 if (rt_task(current))
4210 target = rq->active;
4212 if (array->nr_active == 1) {
4213 schedstat_inc(rq, yld_act_empty);
4214 if (!rq->expired->nr_active)
4215 schedstat_inc(rq, yld_both_empty);
4216 } else if (!rq->expired->nr_active)
4217 schedstat_inc(rq, yld_exp_empty);
4219 if (array != target) {
4220 dequeue_task(current, array);
4221 enqueue_task(current, target);
4224 * requeue_task is cheaper so perform that if possible.
4226 requeue_task(current, array);
4229 * Since we are going to call schedule() anyway, there's
4230 * no need to preempt or enable interrupts:
4232 __release(rq->lock);
4233 _raw_spin_unlock(&rq->lock);
4234 preempt_enable_no_resched();
4241 static inline int __resched_legal(int expected_preempt_count)
4243 if (unlikely(preempt_count() != expected_preempt_count))
4245 if (unlikely(system_state != SYSTEM_RUNNING))
4250 static void __cond_resched(void)
4253 * The BKS might be reacquired before we have dropped
4254 * PREEMPT_ACTIVE, which could trigger a second
4255 * cond_resched() call.
4258 add_preempt_count(PREEMPT_ACTIVE);
4260 sub_preempt_count(PREEMPT_ACTIVE);
4261 } while (need_resched());
4264 int __sched cond_resched(void)
4266 if (need_resched() && __resched_legal(0)) {
4272 EXPORT_SYMBOL(cond_resched);
4275 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4276 * call schedule, and on return reacquire the lock.
4278 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4279 * operations here to prevent schedule() from being called twice (once via
4280 * spin_unlock(), once by hand).
4282 int cond_resched_lock(spinlock_t *lock)
4286 if (need_lockbreak(lock)) {
4292 if (need_resched() && __resched_legal(1)) {
4293 _raw_spin_unlock(lock);
4294 preempt_enable_no_resched();
4301 EXPORT_SYMBOL(cond_resched_lock);
4303 int __sched cond_resched_softirq(void)
4305 BUG_ON(!in_softirq());
4307 if (need_resched() && __resched_legal(0)) {
4308 __local_bh_enable();
4315 EXPORT_SYMBOL(cond_resched_softirq);
4318 * yield - yield the current processor to other threads.
4320 * this is a shortcut for kernel-space yielding - it marks the
4321 * thread runnable and calls sys_sched_yield().
4323 void __sched yield(void)
4325 set_current_state(TASK_RUNNING);
4329 EXPORT_SYMBOL(yield);
4332 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4333 * that process accounting knows that this is a task in IO wait state.
4335 * But don't do that if it is a deliberate, throttling IO wait (this task
4336 * has set its backing_dev_info: the queue against which it should throttle)
4338 void __sched io_schedule(void)
4340 struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id());
4342 atomic_inc(&rq->nr_iowait);
4344 atomic_dec(&rq->nr_iowait);
4347 EXPORT_SYMBOL(io_schedule);
4349 long __sched io_schedule_timeout(long timeout)
4351 struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id());
4354 atomic_inc(&rq->nr_iowait);
4355 ret = schedule_timeout(timeout);
4356 atomic_dec(&rq->nr_iowait);
4361 * sys_sched_get_priority_max - return maximum RT priority.
4362 * @policy: scheduling class.
4364 * this syscall returns the maximum rt_priority that can be used
4365 * by a given scheduling class.
4367 asmlinkage long sys_sched_get_priority_max(int policy)
4374 ret = MAX_USER_RT_PRIO-1;
4385 * sys_sched_get_priority_min - return minimum RT priority.
4386 * @policy: scheduling class.
4388 * this syscall returns the minimum rt_priority that can be used
4389 * by a given scheduling class.
4391 asmlinkage long sys_sched_get_priority_min(int policy)
4408 * sys_sched_rr_get_interval - return the default timeslice of a process.
4409 * @pid: pid of the process.
4410 * @interval: userspace pointer to the timeslice value.
4412 * this syscall writes the default timeslice value of a given process
4413 * into the user-space timespec buffer. A value of '0' means infinity.
4416 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4418 int retval = -EINVAL;
4426 read_lock(&tasklist_lock);
4427 p = find_process_by_pid(pid);
4431 retval = security_task_getscheduler(p);
4435 jiffies_to_timespec(p->policy & SCHED_FIFO ?
4436 0 : task_timeslice(p), &t);
4437 read_unlock(&tasklist_lock);
4438 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4442 read_unlock(&tasklist_lock);
4446 static inline struct task_struct *eldest_child(struct task_struct *p)
4448 if (list_empty(&p->children)) return NULL;
4449 return list_entry(p->children.next,struct task_struct,sibling);
4452 static inline struct task_struct *older_sibling(struct task_struct *p)
4454 if (p->sibling.prev==&p->parent->children) return NULL;
4455 return list_entry(p->sibling.prev,struct task_struct,sibling);
4458 static inline struct task_struct *younger_sibling(struct task_struct *p)
4460 if (p->sibling.next==&p->parent->children) return NULL;
4461 return list_entry(p->sibling.next,struct task_struct,sibling);
4464 static void show_task(task_t *p)
4468 unsigned long free = 0;
4469 static const char *stat_nam[] = { "R", "S", "D", "T", "t", "Z", "X" };
4471 printk("%-13.13s ", p->comm);
4472 state = p->state ? __ffs(p->state) + 1 : 0;
4473 if (state < ARRAY_SIZE(stat_nam))
4474 printk(stat_nam[state]);
4477 #if (BITS_PER_LONG == 32)
4478 if (state == TASK_RUNNING)
4479 printk(" running ");
4481 printk(" %08lX ", thread_saved_pc(p));
4483 if (state == TASK_RUNNING)
4484 printk(" running task ");
4486 printk(" %016lx ", thread_saved_pc(p));
4488 #ifdef CONFIG_DEBUG_STACK_USAGE
4490 unsigned long *n = end_of_stack(p);
4493 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4496 printk("%5lu %5d %6d ", free, p->pid, p->parent->pid);
4497 if ((relative = eldest_child(p)))
4498 printk("%5d ", relative->pid);
4501 if ((relative = younger_sibling(p)))
4502 printk("%7d", relative->pid);
4505 if ((relative = older_sibling(p)))
4506 printk(" %5d", relative->pid);
4510 printk(" (L-TLB)\n");
4512 printk(" (NOTLB)\n");
4514 if (state != TASK_RUNNING)
4515 show_stack(p, NULL);
4518 void show_state(void)
4522 #if (BITS_PER_LONG == 32)
4525 printk(" task PC pid father child younger older\n");
4529 printk(" task PC pid father child younger older\n");
4531 read_lock(&tasklist_lock);
4532 do_each_thread(g, p) {
4534 * reset the NMI-timeout, listing all files on a slow
4535 * console might take alot of time:
4537 touch_nmi_watchdog();
4539 } while_each_thread(g, p);
4541 read_unlock(&tasklist_lock);
4542 mutex_debug_show_all_locks();
4546 * init_idle - set up an idle thread for a given CPU
4547 * @idle: task in question
4548 * @cpu: cpu the idle task belongs to
4550 * NOTE: this function does not set the idle thread's NEED_RESCHED
4551 * flag, to make booting more robust.
4553 void __devinit init_idle(task_t *idle, int cpu)
4555 runqueue_t *rq = cpu_rq(cpu);
4556 unsigned long flags;
4558 idle->timestamp = sched_clock();
4559 idle->sleep_avg = 0;
4561 idle->prio = MAX_PRIO;
4562 idle->state = TASK_RUNNING;
4563 idle->cpus_allowed = cpumask_of_cpu(cpu);
4564 set_task_cpu(idle, cpu);
4566 spin_lock_irqsave(&rq->lock, flags);
4567 rq->curr = rq->idle = idle;
4568 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4571 spin_unlock_irqrestore(&rq->lock, flags);
4573 /* Set the preempt count _outside_ the spinlocks! */
4574 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4575 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4577 task_thread_info(idle)->preempt_count = 0;
4582 * In a system that switches off the HZ timer nohz_cpu_mask
4583 * indicates which cpus entered this state. This is used
4584 * in the rcu update to wait only for active cpus. For system
4585 * which do not switch off the HZ timer nohz_cpu_mask should
4586 * always be CPU_MASK_NONE.
4588 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4592 * This is how migration works:
4594 * 1) we queue a migration_req_t structure in the source CPU's
4595 * runqueue and wake up that CPU's migration thread.
4596 * 2) we down() the locked semaphore => thread blocks.
4597 * 3) migration thread wakes up (implicitly it forces the migrated
4598 * thread off the CPU)
4599 * 4) it gets the migration request and checks whether the migrated
4600 * task is still in the wrong runqueue.
4601 * 5) if it's in the wrong runqueue then the migration thread removes
4602 * it and puts it into the right queue.
4603 * 6) migration thread up()s the semaphore.
4604 * 7) we wake up and the migration is done.
4608 * Change a given task's CPU affinity. Migrate the thread to a
4609 * proper CPU and schedule it away if the CPU it's executing on
4610 * is removed from the allowed bitmask.
4612 * NOTE: the caller must have a valid reference to the task, the
4613 * task must not exit() & deallocate itself prematurely. The
4614 * call is not atomic; no spinlocks may be held.
4616 int set_cpus_allowed(task_t *p, cpumask_t new_mask)
4618 unsigned long flags;
4620 migration_req_t req;
4623 rq = task_rq_lock(p, &flags);
4624 if (!cpus_intersects(new_mask, cpu_online_map)) {
4629 p->cpus_allowed = new_mask;
4630 /* Can the task run on the task's current CPU? If so, we're done */
4631 if (cpu_isset(task_cpu(p), new_mask))
4634 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4635 /* Need help from migration thread: drop lock and wait. */
4636 task_rq_unlock(rq, &flags);
4637 wake_up_process(rq->migration_thread);
4638 wait_for_completion(&req.done);
4639 tlb_migrate_finish(p->mm);
4643 task_rq_unlock(rq, &flags);
4647 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4650 * Move (not current) task off this cpu, onto dest cpu. We're doing
4651 * this because either it can't run here any more (set_cpus_allowed()
4652 * away from this CPU, or CPU going down), or because we're
4653 * attempting to rebalance this task on exec (sched_exec).
4655 * So we race with normal scheduler movements, but that's OK, as long
4656 * as the task is no longer on this CPU.
4658 static void __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4660 runqueue_t *rq_dest, *rq_src;
4662 if (unlikely(cpu_is_offline(dest_cpu)))
4665 rq_src = cpu_rq(src_cpu);
4666 rq_dest = cpu_rq(dest_cpu);
4668 double_rq_lock(rq_src, rq_dest);
4669 /* Already moved. */
4670 if (task_cpu(p) != src_cpu)
4672 /* Affinity changed (again). */
4673 if (!cpu_isset(dest_cpu, p->cpus_allowed))
4676 set_task_cpu(p, dest_cpu);
4679 * Sync timestamp with rq_dest's before activating.
4680 * The same thing could be achieved by doing this step
4681 * afterwards, and pretending it was a local activate.
4682 * This way is cleaner and logically correct.
4684 p->timestamp = p->timestamp - rq_src->timestamp_last_tick
4685 + rq_dest->timestamp_last_tick;
4686 deactivate_task(p, rq_src);
4687 activate_task(p, rq_dest, 0);
4688 if (TASK_PREEMPTS_CURR(p, rq_dest))
4689 resched_task(rq_dest->curr);
4693 double_rq_unlock(rq_src, rq_dest);
4697 * migration_thread - this is a highprio system thread that performs
4698 * thread migration by bumping thread off CPU then 'pushing' onto
4701 static int migration_thread(void *data)
4704 int cpu = (long)data;
4707 BUG_ON(rq->migration_thread != current);
4709 set_current_state(TASK_INTERRUPTIBLE);
4710 while (!kthread_should_stop()) {
4711 struct list_head *head;
4712 migration_req_t *req;
4716 spin_lock_irq(&rq->lock);
4718 if (cpu_is_offline(cpu)) {
4719 spin_unlock_irq(&rq->lock);
4723 if (rq->active_balance) {
4724 active_load_balance(rq, cpu);
4725 rq->active_balance = 0;
4728 head = &rq->migration_queue;
4730 if (list_empty(head)) {
4731 spin_unlock_irq(&rq->lock);
4733 set_current_state(TASK_INTERRUPTIBLE);
4736 req = list_entry(head->next, migration_req_t, list);
4737 list_del_init(head->next);
4739 spin_unlock(&rq->lock);
4740 __migrate_task(req->task, cpu, req->dest_cpu);
4743 complete(&req->done);
4745 __set_current_state(TASK_RUNNING);
4749 /* Wait for kthread_stop */
4750 set_current_state(TASK_INTERRUPTIBLE);
4751 while (!kthread_should_stop()) {
4753 set_current_state(TASK_INTERRUPTIBLE);
4755 __set_current_state(TASK_RUNNING);
4759 #ifdef CONFIG_HOTPLUG_CPU
4760 /* Figure out where task on dead CPU should go, use force if neccessary. */
4761 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *tsk)
4767 mask = node_to_cpumask(cpu_to_node(dead_cpu));
4768 cpus_and(mask, mask, tsk->cpus_allowed);
4769 dest_cpu = any_online_cpu(mask);
4771 /* On any allowed CPU? */
4772 if (dest_cpu == NR_CPUS)
4773 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4775 /* No more Mr. Nice Guy. */
4776 if (dest_cpu == NR_CPUS) {
4777 cpus_setall(tsk->cpus_allowed);
4778 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4781 * Don't tell them about moving exiting tasks or
4782 * kernel threads (both mm NULL), since they never
4785 if (tsk->mm && printk_ratelimit())
4786 printk(KERN_INFO "process %d (%s) no "
4787 "longer affine to cpu%d\n",
4788 tsk->pid, tsk->comm, dead_cpu);
4790 __migrate_task(tsk, dead_cpu, dest_cpu);
4794 * While a dead CPU has no uninterruptible tasks queued at this point,
4795 * it might still have a nonzero ->nr_uninterruptible counter, because
4796 * for performance reasons the counter is not stricly tracking tasks to
4797 * their home CPUs. So we just add the counter to another CPU's counter,
4798 * to keep the global sum constant after CPU-down:
4800 static void migrate_nr_uninterruptible(runqueue_t *rq_src)
4802 runqueue_t *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
4803 unsigned long flags;
4805 local_irq_save(flags);
4806 double_rq_lock(rq_src, rq_dest);
4807 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
4808 rq_src->nr_uninterruptible = 0;
4809 double_rq_unlock(rq_src, rq_dest);
4810 local_irq_restore(flags);
4813 /* Run through task list and migrate tasks from the dead cpu. */
4814 static void migrate_live_tasks(int src_cpu)
4816 struct task_struct *tsk, *t;
4818 write_lock_irq(&tasklist_lock);
4820 do_each_thread(t, tsk) {
4824 if (task_cpu(tsk) == src_cpu)
4825 move_task_off_dead_cpu(src_cpu, tsk);
4826 } while_each_thread(t, tsk);
4828 write_unlock_irq(&tasklist_lock);
4831 /* Schedules idle task to be the next runnable task on current CPU.
4832 * It does so by boosting its priority to highest possible and adding it to
4833 * the _front_ of runqueue. Used by CPU offline code.
4835 void sched_idle_next(void)
4837 int cpu = smp_processor_id();
4838 runqueue_t *rq = this_rq();
4839 struct task_struct *p = rq->idle;
4840 unsigned long flags;
4842 /* cpu has to be offline */
4843 BUG_ON(cpu_online(cpu));
4845 /* Strictly not necessary since rest of the CPUs are stopped by now
4846 * and interrupts disabled on current cpu.
4848 spin_lock_irqsave(&rq->lock, flags);
4850 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4851 /* Add idle task to _front_ of it's priority queue */
4852 __activate_idle_task(p, rq);
4854 spin_unlock_irqrestore(&rq->lock, flags);
4857 /* Ensures that the idle task is using init_mm right before its cpu goes
4860 void idle_task_exit(void)
4862 struct mm_struct *mm = current->active_mm;
4864 BUG_ON(cpu_online(smp_processor_id()));
4867 switch_mm(mm, &init_mm, current);
4871 static void migrate_dead(unsigned int dead_cpu, task_t *tsk)
4873 struct runqueue *rq = cpu_rq(dead_cpu);
4875 /* Must be exiting, otherwise would be on tasklist. */
4876 BUG_ON(tsk->exit_state != EXIT_ZOMBIE && tsk->exit_state != EXIT_DEAD);
4878 /* Cannot have done final schedule yet: would have vanished. */
4879 BUG_ON(tsk->flags & PF_DEAD);
4881 get_task_struct(tsk);
4884 * Drop lock around migration; if someone else moves it,
4885 * that's OK. No task can be added to this CPU, so iteration is
4888 spin_unlock_irq(&rq->lock);
4889 move_task_off_dead_cpu(dead_cpu, tsk);
4890 spin_lock_irq(&rq->lock);
4892 put_task_struct(tsk);
4895 /* release_task() removes task from tasklist, so we won't find dead tasks. */
4896 static void migrate_dead_tasks(unsigned int dead_cpu)
4899 struct runqueue *rq = cpu_rq(dead_cpu);
4901 for (arr = 0; arr < 2; arr++) {
4902 for (i = 0; i < MAX_PRIO; i++) {
4903 struct list_head *list = &rq->arrays[arr].queue[i];
4904 while (!list_empty(list))
4905 migrate_dead(dead_cpu,
4906 list_entry(list->next, task_t,
4911 #endif /* CONFIG_HOTPLUG_CPU */
4914 * migration_call - callback that gets triggered when a CPU is added.
4915 * Here we can start up the necessary migration thread for the new CPU.
4917 static int migration_call(struct notifier_block *nfb, unsigned long action,
4920 int cpu = (long)hcpu;
4921 struct task_struct *p;
4922 struct runqueue *rq;
4923 unsigned long flags;
4926 case CPU_UP_PREPARE:
4927 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
4930 p->flags |= PF_NOFREEZE;
4931 kthread_bind(p, cpu);
4932 /* Must be high prio: stop_machine expects to yield to it. */
4933 rq = task_rq_lock(p, &flags);
4934 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4935 task_rq_unlock(rq, &flags);
4936 cpu_rq(cpu)->migration_thread = p;
4939 /* Strictly unneccessary, as first user will wake it. */
4940 wake_up_process(cpu_rq(cpu)->migration_thread);
4942 #ifdef CONFIG_HOTPLUG_CPU
4943 case CPU_UP_CANCELED:
4944 /* Unbind it from offline cpu so it can run. Fall thru. */
4945 kthread_bind(cpu_rq(cpu)->migration_thread,
4946 any_online_cpu(cpu_online_map));
4947 kthread_stop(cpu_rq(cpu)->migration_thread);
4948 cpu_rq(cpu)->migration_thread = NULL;
4951 migrate_live_tasks(cpu);
4953 kthread_stop(rq->migration_thread);
4954 rq->migration_thread = NULL;
4955 /* Idle task back to normal (off runqueue, low prio) */
4956 rq = task_rq_lock(rq->idle, &flags);
4957 deactivate_task(rq->idle, rq);
4958 rq->idle->static_prio = MAX_PRIO;
4959 __setscheduler(rq->idle, SCHED_NORMAL, 0);
4960 migrate_dead_tasks(cpu);
4961 task_rq_unlock(rq, &flags);
4962 migrate_nr_uninterruptible(rq);
4963 BUG_ON(rq->nr_running != 0);
4965 /* No need to migrate the tasks: it was best-effort if
4966 * they didn't do lock_cpu_hotplug(). Just wake up
4967 * the requestors. */
4968 spin_lock_irq(&rq->lock);
4969 while (!list_empty(&rq->migration_queue)) {
4970 migration_req_t *req;
4971 req = list_entry(rq->migration_queue.next,
4972 migration_req_t, list);
4973 list_del_init(&req->list);
4974 complete(&req->done);
4976 spin_unlock_irq(&rq->lock);
4983 /* Register at highest priority so that task migration (migrate_all_tasks)
4984 * happens before everything else.
4986 static struct notifier_block migration_notifier = {
4987 .notifier_call = migration_call,
4991 int __init migration_init(void)
4993 void *cpu = (void *)(long)smp_processor_id();
4994 /* Start one for boot CPU. */
4995 migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
4996 migration_call(&migration_notifier, CPU_ONLINE, cpu);
4997 register_cpu_notifier(&migration_notifier);
5003 #undef SCHED_DOMAIN_DEBUG
5004 #ifdef SCHED_DOMAIN_DEBUG
5005 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5010 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5014 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5019 struct sched_group *group = sd->groups;
5020 cpumask_t groupmask;
5022 cpumask_scnprintf(str, NR_CPUS, sd->span);
5023 cpus_clear(groupmask);
5026 for (i = 0; i < level + 1; i++)
5028 printk("domain %d: ", level);
5030 if (!(sd->flags & SD_LOAD_BALANCE)) {
5031 printk("does not load-balance\n");
5033 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent");
5037 printk("span %s\n", str);
5039 if (!cpu_isset(cpu, sd->span))
5040 printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
5041 if (!cpu_isset(cpu, group->cpumask))
5042 printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
5045 for (i = 0; i < level + 2; i++)
5051 printk(KERN_ERR "ERROR: group is NULL\n");
5055 if (!group->cpu_power) {
5057 printk(KERN_ERR "ERROR: domain->cpu_power not set\n");
5060 if (!cpus_weight(group->cpumask)) {
5062 printk(KERN_ERR "ERROR: empty group\n");
5065 if (cpus_intersects(groupmask, group->cpumask)) {
5067 printk(KERN_ERR "ERROR: repeated CPUs\n");
5070 cpus_or(groupmask, groupmask, group->cpumask);
5072 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5075 group = group->next;
5076 } while (group != sd->groups);
5079 if (!cpus_equal(sd->span, groupmask))
5080 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5086 if (!cpus_subset(groupmask, sd->span))
5087 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
5093 #define sched_domain_debug(sd, cpu) {}
5096 static int sd_degenerate(struct sched_domain *sd)
5098 if (cpus_weight(sd->span) == 1)
5101 /* Following flags need at least 2 groups */
5102 if (sd->flags & (SD_LOAD_BALANCE |
5103 SD_BALANCE_NEWIDLE |
5106 if (sd->groups != sd->groups->next)
5110 /* Following flags don't use groups */
5111 if (sd->flags & (SD_WAKE_IDLE |
5119 static int sd_parent_degenerate(struct sched_domain *sd,
5120 struct sched_domain *parent)
5122 unsigned long cflags = sd->flags, pflags = parent->flags;
5124 if (sd_degenerate(parent))
5127 if (!cpus_equal(sd->span, parent->span))
5130 /* Does parent contain flags not in child? */
5131 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5132 if (cflags & SD_WAKE_AFFINE)
5133 pflags &= ~SD_WAKE_BALANCE;
5134 /* Flags needing groups don't count if only 1 group in parent */
5135 if (parent->groups == parent->groups->next) {
5136 pflags &= ~(SD_LOAD_BALANCE |
5137 SD_BALANCE_NEWIDLE |
5141 if (~cflags & pflags)
5148 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5149 * hold the hotplug lock.
5151 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5153 runqueue_t *rq = cpu_rq(cpu);
5154 struct sched_domain *tmp;
5156 /* Remove the sched domains which do not contribute to scheduling. */
5157 for (tmp = sd; tmp; tmp = tmp->parent) {
5158 struct sched_domain *parent = tmp->parent;
5161 if (sd_parent_degenerate(tmp, parent))
5162 tmp->parent = parent->parent;
5165 if (sd && sd_degenerate(sd))
5168 sched_domain_debug(sd, cpu);
5170 rcu_assign_pointer(rq->sd, sd);
5173 /* cpus with isolated domains */
5174 static cpumask_t __devinitdata cpu_isolated_map = CPU_MASK_NONE;
5176 /* Setup the mask of cpus configured for isolated domains */
5177 static int __init isolated_cpu_setup(char *str)
5179 int ints[NR_CPUS], i;
5181 str = get_options(str, ARRAY_SIZE(ints), ints);
5182 cpus_clear(cpu_isolated_map);
5183 for (i = 1; i <= ints[0]; i++)
5184 if (ints[i] < NR_CPUS)
5185 cpu_set(ints[i], cpu_isolated_map);
5189 __setup ("isolcpus=", isolated_cpu_setup);
5192 * init_sched_build_groups takes an array of groups, the cpumask we wish
5193 * to span, and a pointer to a function which identifies what group a CPU
5194 * belongs to. The return value of group_fn must be a valid index into the
5195 * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
5196 * keep track of groups covered with a cpumask_t).
5198 * init_sched_build_groups will build a circular linked list of the groups
5199 * covered by the given span, and will set each group's ->cpumask correctly,
5200 * and ->cpu_power to 0.
5202 static void init_sched_build_groups(struct sched_group groups[], cpumask_t span,
5203 int (*group_fn)(int cpu))
5205 struct sched_group *first = NULL, *last = NULL;
5206 cpumask_t covered = CPU_MASK_NONE;
5209 for_each_cpu_mask(i, span) {
5210 int group = group_fn(i);
5211 struct sched_group *sg = &groups[group];
5214 if (cpu_isset(i, covered))
5217 sg->cpumask = CPU_MASK_NONE;
5220 for_each_cpu_mask(j, span) {
5221 if (group_fn(j) != group)
5224 cpu_set(j, covered);
5225 cpu_set(j, sg->cpumask);
5236 #define SD_NODES_PER_DOMAIN 16
5239 * Self-tuning task migration cost measurement between source and target CPUs.
5241 * This is done by measuring the cost of manipulating buffers of varying
5242 * sizes. For a given buffer-size here are the steps that are taken:
5244 * 1) the source CPU reads+dirties a shared buffer
5245 * 2) the target CPU reads+dirties the same shared buffer
5247 * We measure how long they take, in the following 4 scenarios:
5249 * - source: CPU1, target: CPU2 | cost1
5250 * - source: CPU2, target: CPU1 | cost2
5251 * - source: CPU1, target: CPU1 | cost3
5252 * - source: CPU2, target: CPU2 | cost4
5254 * We then calculate the cost3+cost4-cost1-cost2 difference - this is
5255 * the cost of migration.
5257 * We then start off from a small buffer-size and iterate up to larger
5258 * buffer sizes, in 5% steps - measuring each buffer-size separately, and
5259 * doing a maximum search for the cost. (The maximum cost for a migration
5260 * normally occurs when the working set size is around the effective cache
5263 #define SEARCH_SCOPE 2
5264 #define MIN_CACHE_SIZE (64*1024U)
5265 #define DEFAULT_CACHE_SIZE (5*1024*1024U)
5266 #define ITERATIONS 1
5267 #define SIZE_THRESH 130
5268 #define COST_THRESH 130
5271 * The migration cost is a function of 'domain distance'. Domain
5272 * distance is the number of steps a CPU has to iterate down its
5273 * domain tree to share a domain with the other CPU. The farther
5274 * two CPUs are from each other, the larger the distance gets.
5276 * Note that we use the distance only to cache measurement results,
5277 * the distance value is not used numerically otherwise. When two
5278 * CPUs have the same distance it is assumed that the migration
5279 * cost is the same. (this is a simplification but quite practical)
5281 #define MAX_DOMAIN_DISTANCE 32
5283 static unsigned long long migration_cost[MAX_DOMAIN_DISTANCE] =
5284 { [ 0 ... MAX_DOMAIN_DISTANCE-1 ] =
5286 * Architectures may override the migration cost and thus avoid
5287 * boot-time calibration. Unit is nanoseconds. Mostly useful for
5288 * virtualized hardware:
5290 #ifdef CONFIG_DEFAULT_MIGRATION_COST
5291 CONFIG_DEFAULT_MIGRATION_COST
5298 * Allow override of migration cost - in units of microseconds.
5299 * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost
5300 * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs:
5302 static int __init migration_cost_setup(char *str)
5304 int ints[MAX_DOMAIN_DISTANCE+1], i;
5306 str = get_options(str, ARRAY_SIZE(ints), ints);
5308 printk("#ints: %d\n", ints[0]);
5309 for (i = 1; i <= ints[0]; i++) {
5310 migration_cost[i-1] = (unsigned long long)ints[i]*1000;
5311 printk("migration_cost[%d]: %Ld\n", i-1, migration_cost[i-1]);
5316 __setup ("migration_cost=", migration_cost_setup);
5319 * Global multiplier (divisor) for migration-cutoff values,
5320 * in percentiles. E.g. use a value of 150 to get 1.5 times
5321 * longer cache-hot cutoff times.
5323 * (We scale it from 100 to 128 to long long handling easier.)
5326 #define MIGRATION_FACTOR_SCALE 128
5328 static unsigned int migration_factor = MIGRATION_FACTOR_SCALE;
5330 static int __init setup_migration_factor(char *str)
5332 get_option(&str, &migration_factor);
5333 migration_factor = migration_factor * MIGRATION_FACTOR_SCALE / 100;
5337 __setup("migration_factor=", setup_migration_factor);
5340 * Estimated distance of two CPUs, measured via the number of domains
5341 * we have to pass for the two CPUs to be in the same span:
5343 static unsigned long domain_distance(int cpu1, int cpu2)
5345 unsigned long distance = 0;
5346 struct sched_domain *sd;
5348 for_each_domain(cpu1, sd) {
5349 WARN_ON(!cpu_isset(cpu1, sd->span));
5350 if (cpu_isset(cpu2, sd->span))
5354 if (distance >= MAX_DOMAIN_DISTANCE) {
5356 distance = MAX_DOMAIN_DISTANCE-1;
5362 static unsigned int migration_debug;
5364 static int __init setup_migration_debug(char *str)
5366 get_option(&str, &migration_debug);
5370 __setup("migration_debug=", setup_migration_debug);
5373 * Maximum cache-size that the scheduler should try to measure.
5374 * Architectures with larger caches should tune this up during
5375 * bootup. Gets used in the domain-setup code (i.e. during SMP
5378 unsigned int max_cache_size;
5380 static int __init setup_max_cache_size(char *str)
5382 get_option(&str, &max_cache_size);
5386 __setup("max_cache_size=", setup_max_cache_size);
5389 * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This
5390 * is the operation that is timed, so we try to generate unpredictable
5391 * cachemisses that still end up filling the L2 cache:
5393 static void touch_cache(void *__cache, unsigned long __size)
5395 unsigned long size = __size/sizeof(long), chunk1 = size/3,
5397 unsigned long *cache = __cache;
5400 for (i = 0; i < size/6; i += 8) {
5403 case 1: cache[size-1-i]++;
5404 case 2: cache[chunk1-i]++;
5405 case 3: cache[chunk1+i]++;
5406 case 4: cache[chunk2-i]++;
5407 case 5: cache[chunk2+i]++;
5413 * Measure the cache-cost of one task migration. Returns in units of nsec.
5415 static unsigned long long measure_one(void *cache, unsigned long size,
5416 int source, int target)
5418 cpumask_t mask, saved_mask;
5419 unsigned long long t0, t1, t2, t3, cost;
5421 saved_mask = current->cpus_allowed;
5424 * Flush source caches to RAM and invalidate them:
5429 * Migrate to the source CPU:
5431 mask = cpumask_of_cpu(source);
5432 set_cpus_allowed(current, mask);
5433 WARN_ON(smp_processor_id() != source);
5436 * Dirty the working set:
5439 touch_cache(cache, size);
5443 * Migrate to the target CPU, dirty the L2 cache and access
5444 * the shared buffer. (which represents the working set
5445 * of a migrated task.)
5447 mask = cpumask_of_cpu(target);
5448 set_cpus_allowed(current, mask);
5449 WARN_ON(smp_processor_id() != target);
5452 touch_cache(cache, size);
5455 cost = t1-t0 + t3-t2;
5457 if (migration_debug >= 2)
5458 printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n",
5459 source, target, t1-t0, t1-t0, t3-t2, cost);
5461 * Flush target caches to RAM and invalidate them:
5465 set_cpus_allowed(current, saved_mask);
5471 * Measure a series of task migrations and return the average
5472 * result. Since this code runs early during bootup the system
5473 * is 'undisturbed' and the average latency makes sense.
5475 * The algorithm in essence auto-detects the relevant cache-size,
5476 * so it will properly detect different cachesizes for different
5477 * cache-hierarchies, depending on how the CPUs are connected.
5479 * Architectures can prime the upper limit of the search range via
5480 * max_cache_size, otherwise the search range defaults to 20MB...64K.
5482 static unsigned long long
5483 measure_cost(int cpu1, int cpu2, void *cache, unsigned int size)
5485 unsigned long long cost1, cost2;
5489 * Measure the migration cost of 'size' bytes, over an
5490 * average of 10 runs:
5492 * (We perturb the cache size by a small (0..4k)
5493 * value to compensate size/alignment related artifacts.
5494 * We also subtract the cost of the operation done on
5500 * dry run, to make sure we start off cache-cold on cpu1,
5501 * and to get any vmalloc pagefaults in advance:
5503 measure_one(cache, size, cpu1, cpu2);
5504 for (i = 0; i < ITERATIONS; i++)
5505 cost1 += measure_one(cache, size - i*1024, cpu1, cpu2);
5507 measure_one(cache, size, cpu2, cpu1);
5508 for (i = 0; i < ITERATIONS; i++)
5509 cost1 += measure_one(cache, size - i*1024, cpu2, cpu1);
5512 * (We measure the non-migrating [cached] cost on both
5513 * cpu1 and cpu2, to handle CPUs with different speeds)
5517 measure_one(cache, size, cpu1, cpu1);
5518 for (i = 0; i < ITERATIONS; i++)
5519 cost2 += measure_one(cache, size - i*1024, cpu1, cpu1);
5521 measure_one(cache, size, cpu2, cpu2);
5522 for (i = 0; i < ITERATIONS; i++)
5523 cost2 += measure_one(cache, size - i*1024, cpu2, cpu2);
5526 * Get the per-iteration migration cost:
5528 do_div(cost1, 2*ITERATIONS);
5529 do_div(cost2, 2*ITERATIONS);
5531 return cost1 - cost2;
5534 static unsigned long long measure_migration_cost(int cpu1, int cpu2)
5536 unsigned long long max_cost = 0, fluct = 0, avg_fluct = 0;
5537 unsigned int max_size, size, size_found = 0;
5538 long long cost = 0, prev_cost;
5542 * Search from max_cache_size*5 down to 64K - the real relevant
5543 * cachesize has to lie somewhere inbetween.
5545 if (max_cache_size) {
5546 max_size = max(max_cache_size * SEARCH_SCOPE, MIN_CACHE_SIZE);
5547 size = max(max_cache_size / SEARCH_SCOPE, MIN_CACHE_SIZE);
5550 * Since we have no estimation about the relevant
5553 max_size = DEFAULT_CACHE_SIZE * SEARCH_SCOPE;
5554 size = MIN_CACHE_SIZE;
5557 if (!cpu_online(cpu1) || !cpu_online(cpu2)) {
5558 printk("cpu %d and %d not both online!\n", cpu1, cpu2);
5563 * Allocate the working set:
5565 cache = vmalloc(max_size);
5567 printk("could not vmalloc %d bytes for cache!\n", 2*max_size);
5568 return 1000000; // return 1 msec on very small boxen
5571 while (size <= max_size) {
5573 cost = measure_cost(cpu1, cpu2, cache, size);
5579 if (max_cost < cost) {
5585 * Calculate average fluctuation, we use this to prevent
5586 * noise from triggering an early break out of the loop:
5588 fluct = abs(cost - prev_cost);
5589 avg_fluct = (avg_fluct + fluct)/2;
5591 if (migration_debug)
5592 printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): (%8Ld %8Ld)\n",
5594 (long)cost / 1000000,
5595 ((long)cost / 100000) % 10,
5596 (long)max_cost / 1000000,
5597 ((long)max_cost / 100000) % 10,
5598 domain_distance(cpu1, cpu2),
5602 * If we iterated at least 20% past the previous maximum,
5603 * and the cost has dropped by more than 20% already,
5604 * (taking fluctuations into account) then we assume to
5605 * have found the maximum and break out of the loop early:
5607 if (size_found && (size*100 > size_found*SIZE_THRESH))
5608 if (cost+avg_fluct <= 0 ||
5609 max_cost*100 > (cost+avg_fluct)*COST_THRESH) {
5611 if (migration_debug)
5612 printk("-> found max.\n");
5616 * Increase the cachesize in 10% steps:
5618 size = size * 10 / 9;
5621 if (migration_debug)
5622 printk("[%d][%d] working set size found: %d, cost: %Ld\n",
5623 cpu1, cpu2, size_found, max_cost);
5628 * A task is considered 'cache cold' if at least 2 times
5629 * the worst-case cost of migration has passed.
5631 * (this limit is only listened to if the load-balancing
5632 * situation is 'nice' - if there is a large imbalance we
5633 * ignore it for the sake of CPU utilization and
5634 * processing fairness.)
5636 return 2 * max_cost * migration_factor / MIGRATION_FACTOR_SCALE;
5639 static void calibrate_migration_costs(const cpumask_t *cpu_map)
5641 int cpu1 = -1, cpu2 = -1, cpu, orig_cpu = raw_smp_processor_id();
5642 unsigned long j0, j1, distance, max_distance = 0;
5643 struct sched_domain *sd;
5648 * First pass - calculate the cacheflush times:
5650 for_each_cpu_mask(cpu1, *cpu_map) {
5651 for_each_cpu_mask(cpu2, *cpu_map) {
5654 distance = domain_distance(cpu1, cpu2);
5655 max_distance = max(max_distance, distance);
5657 * No result cached yet?
5659 if (migration_cost[distance] == -1LL)
5660 migration_cost[distance] =
5661 measure_migration_cost(cpu1, cpu2);
5665 * Second pass - update the sched domain hierarchy with
5666 * the new cache-hot-time estimations:
5668 for_each_cpu_mask(cpu, *cpu_map) {
5670 for_each_domain(cpu, sd) {
5671 sd->cache_hot_time = migration_cost[distance];
5678 if (migration_debug)
5679 printk("migration: max_cache_size: %d, cpu: %d MHz:\n",
5687 if (system_state == SYSTEM_BOOTING) {
5688 if (num_online_cpus() > 1) {
5689 printk("migration_cost=");
5690 for (distance = 0; distance <= max_distance; distance++) {
5693 printk("%ld", (long)migration_cost[distance] / 1000);
5699 if (migration_debug)
5700 printk("migration: %ld seconds\n", (j1-j0)/HZ);
5703 * Move back to the original CPU. NUMA-Q gets confused
5704 * if we migrate to another quad during bootup.
5706 if (raw_smp_processor_id() != orig_cpu) {
5707 cpumask_t mask = cpumask_of_cpu(orig_cpu),
5708 saved_mask = current->cpus_allowed;
5710 set_cpus_allowed(current, mask);
5711 set_cpus_allowed(current, saved_mask);
5718 * find_next_best_node - find the next node to include in a sched_domain
5719 * @node: node whose sched_domain we're building
5720 * @used_nodes: nodes already in the sched_domain
5722 * Find the next node to include in a given scheduling domain. Simply
5723 * finds the closest node not already in the @used_nodes map.
5725 * Should use nodemask_t.
5727 static int find_next_best_node(int node, unsigned long *used_nodes)
5729 int i, n, val, min_val, best_node = 0;
5733 for (i = 0; i < MAX_NUMNODES; i++) {
5734 /* Start at @node */
5735 n = (node + i) % MAX_NUMNODES;
5737 if (!nr_cpus_node(n))
5740 /* Skip already used nodes */
5741 if (test_bit(n, used_nodes))
5744 /* Simple min distance search */
5745 val = node_distance(node, n);
5747 if (val < min_val) {
5753 set_bit(best_node, used_nodes);
5758 * sched_domain_node_span - get a cpumask for a node's sched_domain
5759 * @node: node whose cpumask we're constructing
5760 * @size: number of nodes to include in this span
5762 * Given a node, construct a good cpumask for its sched_domain to span. It
5763 * should be one that prevents unnecessary balancing, but also spreads tasks
5766 static cpumask_t sched_domain_node_span(int node)
5769 cpumask_t span, nodemask;
5770 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5773 bitmap_zero(used_nodes, MAX_NUMNODES);
5775 nodemask = node_to_cpumask(node);
5776 cpus_or(span, span, nodemask);
5777 set_bit(node, used_nodes);
5779 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5780 int next_node = find_next_best_node(node, used_nodes);
5781 nodemask = node_to_cpumask(next_node);
5782 cpus_or(span, span, nodemask);
5790 * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we
5791 * can switch it on easily if needed.
5793 #ifdef CONFIG_SCHED_SMT
5794 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5795 static struct sched_group sched_group_cpus[NR_CPUS];
5796 static int cpu_to_cpu_group(int cpu)
5802 #ifdef CONFIG_SCHED_MC
5803 static DEFINE_PER_CPU(struct sched_domain, core_domains);
5804 static struct sched_group sched_group_core[NR_CPUS];
5807 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5808 static int cpu_to_core_group(int cpu)
5810 return first_cpu(cpu_sibling_map[cpu]);
5812 #elif defined(CONFIG_SCHED_MC)
5813 static int cpu_to_core_group(int cpu)
5819 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5820 static struct sched_group sched_group_phys[NR_CPUS];
5821 static int cpu_to_phys_group(int cpu)
5823 #if defined(CONFIG_SCHED_MC)
5824 cpumask_t mask = cpu_coregroup_map(cpu);
5825 return first_cpu(mask);
5826 #elif defined(CONFIG_SCHED_SMT)
5827 return first_cpu(cpu_sibling_map[cpu]);
5835 * The init_sched_build_groups can't handle what we want to do with node
5836 * groups, so roll our own. Now each node has its own list of groups which
5837 * gets dynamically allocated.
5839 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5840 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5842 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5843 static struct sched_group *sched_group_allnodes_bycpu[NR_CPUS];
5845 static int cpu_to_allnodes_group(int cpu)
5847 return cpu_to_node(cpu);
5849 static void init_numa_sched_groups_power(struct sched_group *group_head)
5851 struct sched_group *sg = group_head;
5857 for_each_cpu_mask(j, sg->cpumask) {
5858 struct sched_domain *sd;
5860 sd = &per_cpu(phys_domains, j);
5861 if (j != first_cpu(sd->groups->cpumask)) {
5863 * Only add "power" once for each
5869 sg->cpu_power += sd->groups->cpu_power;
5872 if (sg != group_head)
5878 * Build sched domains for a given set of cpus and attach the sched domains
5879 * to the individual cpus
5881 void build_sched_domains(const cpumask_t *cpu_map)
5885 struct sched_group **sched_group_nodes = NULL;
5886 struct sched_group *sched_group_allnodes = NULL;
5889 * Allocate the per-node list of sched groups
5891 sched_group_nodes = kmalloc(sizeof(struct sched_group*)*MAX_NUMNODES,
5893 if (!sched_group_nodes) {
5894 printk(KERN_WARNING "Can not alloc sched group node list\n");
5897 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
5901 * Set up domains for cpus specified by the cpu_map.
5903 for_each_cpu_mask(i, *cpu_map) {
5905 struct sched_domain *sd = NULL, *p;
5906 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
5908 cpus_and(nodemask, nodemask, *cpu_map);
5911 if (cpus_weight(*cpu_map)
5912 > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
5913 if (!sched_group_allnodes) {
5914 sched_group_allnodes
5915 = kmalloc(sizeof(struct sched_group)
5918 if (!sched_group_allnodes) {
5920 "Can not alloc allnodes sched group\n");
5923 sched_group_allnodes_bycpu[i]
5924 = sched_group_allnodes;
5926 sd = &per_cpu(allnodes_domains, i);
5927 *sd = SD_ALLNODES_INIT;
5928 sd->span = *cpu_map;
5929 group = cpu_to_allnodes_group(i);
5930 sd->groups = &sched_group_allnodes[group];
5935 sd = &per_cpu(node_domains, i);
5937 sd->span = sched_domain_node_span(cpu_to_node(i));
5939 cpus_and(sd->span, sd->span, *cpu_map);
5943 sd = &per_cpu(phys_domains, i);
5944 group = cpu_to_phys_group(i);
5946 sd->span = nodemask;
5948 sd->groups = &sched_group_phys[group];
5950 #ifdef CONFIG_SCHED_MC
5952 sd = &per_cpu(core_domains, i);
5953 group = cpu_to_core_group(i);
5955 sd->span = cpu_coregroup_map(i);
5956 cpus_and(sd->span, sd->span, *cpu_map);
5958 sd->groups = &sched_group_core[group];
5961 #ifdef CONFIG_SCHED_SMT
5963 sd = &per_cpu(cpu_domains, i);
5964 group = cpu_to_cpu_group(i);
5965 *sd = SD_SIBLING_INIT;
5966 sd->span = cpu_sibling_map[i];
5967 cpus_and(sd->span, sd->span, *cpu_map);
5969 sd->groups = &sched_group_cpus[group];
5973 #ifdef CONFIG_SCHED_SMT
5974 /* Set up CPU (sibling) groups */
5975 for_each_cpu_mask(i, *cpu_map) {
5976 cpumask_t this_sibling_map = cpu_sibling_map[i];
5977 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
5978 if (i != first_cpu(this_sibling_map))
5981 init_sched_build_groups(sched_group_cpus, this_sibling_map,
5986 #ifdef CONFIG_SCHED_MC
5987 /* Set up multi-core groups */
5988 for_each_cpu_mask(i, *cpu_map) {
5989 cpumask_t this_core_map = cpu_coregroup_map(i);
5990 cpus_and(this_core_map, this_core_map, *cpu_map);
5991 if (i != first_cpu(this_core_map))
5993 init_sched_build_groups(sched_group_core, this_core_map,
5994 &cpu_to_core_group);
5999 /* Set up physical groups */
6000 for (i = 0; i < MAX_NUMNODES; i++) {
6001 cpumask_t nodemask = node_to_cpumask(i);
6003 cpus_and(nodemask, nodemask, *cpu_map);
6004 if (cpus_empty(nodemask))
6007 init_sched_build_groups(sched_group_phys, nodemask,
6008 &cpu_to_phys_group);
6012 /* Set up node groups */
6013 if (sched_group_allnodes)
6014 init_sched_build_groups(sched_group_allnodes, *cpu_map,
6015 &cpu_to_allnodes_group);
6017 for (i = 0; i < MAX_NUMNODES; i++) {
6018 /* Set up node groups */
6019 struct sched_group *sg, *prev;
6020 cpumask_t nodemask = node_to_cpumask(i);
6021 cpumask_t domainspan;
6022 cpumask_t covered = CPU_MASK_NONE;
6025 cpus_and(nodemask, nodemask, *cpu_map);
6026 if (cpus_empty(nodemask)) {
6027 sched_group_nodes[i] = NULL;
6031 domainspan = sched_domain_node_span(i);
6032 cpus_and(domainspan, domainspan, *cpu_map);
6034 sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL);
6035 sched_group_nodes[i] = sg;
6036 for_each_cpu_mask(j, nodemask) {
6037 struct sched_domain *sd;
6038 sd = &per_cpu(node_domains, j);
6040 if (sd->groups == NULL) {
6041 /* Turn off balancing if we have no groups */
6047 "Can not alloc domain group for node %d\n", i);
6051 sg->cpumask = nodemask;
6052 cpus_or(covered, covered, nodemask);
6055 for (j = 0; j < MAX_NUMNODES; j++) {
6056 cpumask_t tmp, notcovered;
6057 int n = (i + j) % MAX_NUMNODES;
6059 cpus_complement(notcovered, covered);
6060 cpus_and(tmp, notcovered, *cpu_map);
6061 cpus_and(tmp, tmp, domainspan);
6062 if (cpus_empty(tmp))
6065 nodemask = node_to_cpumask(n);
6066 cpus_and(tmp, tmp, nodemask);
6067 if (cpus_empty(tmp))
6070 sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL);
6073 "Can not alloc domain group for node %d\n", j);
6078 cpus_or(covered, covered, tmp);
6082 prev->next = sched_group_nodes[i];
6086 /* Calculate CPU power for physical packages and nodes */
6087 for_each_cpu_mask(i, *cpu_map) {
6089 struct sched_domain *sd;
6090 #ifdef CONFIG_SCHED_SMT
6091 sd = &per_cpu(cpu_domains, i);
6092 power = SCHED_LOAD_SCALE;
6093 sd->groups->cpu_power = power;
6095 #ifdef CONFIG_SCHED_MC
6096 sd = &per_cpu(core_domains, i);
6097 power = SCHED_LOAD_SCALE + (cpus_weight(sd->groups->cpumask)-1)
6098 * SCHED_LOAD_SCALE / 10;
6099 sd->groups->cpu_power = power;
6101 sd = &per_cpu(phys_domains, i);
6104 * This has to be < 2 * SCHED_LOAD_SCALE
6105 * Lets keep it SCHED_LOAD_SCALE, so that
6106 * while calculating NUMA group's cpu_power
6108 * numa_group->cpu_power += phys_group->cpu_power;
6110 * See "only add power once for each physical pkg"
6113 sd->groups->cpu_power = SCHED_LOAD_SCALE;
6115 sd = &per_cpu(phys_domains, i);
6116 power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
6117 (cpus_weight(sd->groups->cpumask)-1) / 10;
6118 sd->groups->cpu_power = power;
6123 for (i = 0; i < MAX_NUMNODES; i++)
6124 init_numa_sched_groups_power(sched_group_nodes[i]);
6126 init_numa_sched_groups_power(sched_group_allnodes);
6129 /* Attach the domains */
6130 for_each_cpu_mask(i, *cpu_map) {
6131 struct sched_domain *sd;
6132 #ifdef CONFIG_SCHED_SMT
6133 sd = &per_cpu(cpu_domains, i);
6134 #elif defined(CONFIG_SCHED_MC)
6135 sd = &per_cpu(core_domains, i);
6137 sd = &per_cpu(phys_domains, i);
6139 cpu_attach_domain(sd, i);
6142 * Tune cache-hot values:
6144 calibrate_migration_costs(cpu_map);
6147 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6149 static void arch_init_sched_domains(const cpumask_t *cpu_map)
6151 cpumask_t cpu_default_map;
6154 * Setup mask for cpus without special case scheduling requirements.
6155 * For now this just excludes isolated cpus, but could be used to
6156 * exclude other special cases in the future.
6158 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6160 build_sched_domains(&cpu_default_map);
6163 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6169 for_each_cpu_mask(cpu, *cpu_map) {
6170 struct sched_group *sched_group_allnodes
6171 = sched_group_allnodes_bycpu[cpu];
6172 struct sched_group **sched_group_nodes
6173 = sched_group_nodes_bycpu[cpu];
6175 if (sched_group_allnodes) {
6176 kfree(sched_group_allnodes);
6177 sched_group_allnodes_bycpu[cpu] = NULL;
6180 if (!sched_group_nodes)
6183 for (i = 0; i < MAX_NUMNODES; i++) {
6184 cpumask_t nodemask = node_to_cpumask(i);
6185 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6187 cpus_and(nodemask, nodemask, *cpu_map);
6188 if (cpus_empty(nodemask))
6198 if (oldsg != sched_group_nodes[i])
6201 kfree(sched_group_nodes);
6202 sched_group_nodes_bycpu[cpu] = NULL;
6208 * Detach sched domains from a group of cpus specified in cpu_map
6209 * These cpus will now be attached to the NULL domain
6211 static void detach_destroy_domains(const cpumask_t *cpu_map)
6215 for_each_cpu_mask(i, *cpu_map)
6216 cpu_attach_domain(NULL, i);
6217 synchronize_sched();
6218 arch_destroy_sched_domains(cpu_map);
6222 * Partition sched domains as specified by the cpumasks below.
6223 * This attaches all cpus from the cpumasks to the NULL domain,
6224 * waits for a RCU quiescent period, recalculates sched
6225 * domain information and then attaches them back to the
6226 * correct sched domains
6227 * Call with hotplug lock held
6229 void partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6231 cpumask_t change_map;
6233 cpus_and(*partition1, *partition1, cpu_online_map);
6234 cpus_and(*partition2, *partition2, cpu_online_map);
6235 cpus_or(change_map, *partition1, *partition2);
6237 /* Detach sched domains from all of the affected cpus */
6238 detach_destroy_domains(&change_map);
6239 if (!cpus_empty(*partition1))
6240 build_sched_domains(partition1);
6241 if (!cpus_empty(*partition2))
6242 build_sched_domains(partition2);
6245 #ifdef CONFIG_HOTPLUG_CPU
6247 * Force a reinitialization of the sched domains hierarchy. The domains
6248 * and groups cannot be updated in place without racing with the balancing
6249 * code, so we temporarily attach all running cpus to the NULL domain
6250 * which will prevent rebalancing while the sched domains are recalculated.
6252 static int update_sched_domains(struct notifier_block *nfb,
6253 unsigned long action, void *hcpu)
6256 case CPU_UP_PREPARE:
6257 case CPU_DOWN_PREPARE:
6258 detach_destroy_domains(&cpu_online_map);
6261 case CPU_UP_CANCELED:
6262 case CPU_DOWN_FAILED:
6266 * Fall through and re-initialise the domains.
6273 /* The hotplug lock is already held by cpu_up/cpu_down */
6274 arch_init_sched_domains(&cpu_online_map);
6280 void __init sched_init_smp(void)
6283 arch_init_sched_domains(&cpu_online_map);
6284 unlock_cpu_hotplug();
6285 /* XXX: Theoretical race here - CPU may be hotplugged now */
6286 hotcpu_notifier(update_sched_domains, 0);
6289 void __init sched_init_smp(void)
6292 #endif /* CONFIG_SMP */
6294 int in_sched_functions(unsigned long addr)
6296 /* Linker adds these: start and end of __sched functions */
6297 extern char __sched_text_start[], __sched_text_end[];
6298 return in_lock_functions(addr) ||
6299 (addr >= (unsigned long)__sched_text_start
6300 && addr < (unsigned long)__sched_text_end);
6303 void __init sched_init(void)
6308 for_each_possible_cpu(i) {
6309 prio_array_t *array;
6312 spin_lock_init(&rq->lock);
6314 rq->active = rq->arrays;
6315 rq->expired = rq->arrays + 1;
6316 rq->best_expired_prio = MAX_PRIO;
6320 for (j = 1; j < 3; j++)
6321 rq->cpu_load[j] = 0;
6322 rq->active_balance = 0;
6324 rq->migration_thread = NULL;
6325 INIT_LIST_HEAD(&rq->migration_queue);
6328 atomic_set(&rq->nr_iowait, 0);
6329 #ifdef CONFIG_VSERVER_HARDCPU
6330 INIT_LIST_HEAD(&rq->hold_queue);
6333 for (j = 0; j < 2; j++) {
6334 array = rq->arrays + j;
6335 for (k = 0; k < MAX_PRIO; k++) {
6336 INIT_LIST_HEAD(array->queue + k);
6337 __clear_bit(k, array->bitmap);
6339 // delimiter for bitsearch
6340 __set_bit(MAX_PRIO, array->bitmap);
6345 * The boot idle thread does lazy MMU switching as well:
6347 atomic_inc(&init_mm.mm_count);
6348 enter_lazy_tlb(&init_mm, current);
6351 * Make us the idle thread. Technically, schedule() should not be
6352 * called from this thread, however somewhere below it might be,
6353 * but because we are the idle thread, we just pick up running again
6354 * when this runqueue becomes "idle".
6356 init_idle(current, smp_processor_id());
6359 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6360 void __might_sleep(char *file, int line)
6362 #if defined(in_atomic)
6363 static unsigned long prev_jiffy; /* ratelimiting */
6365 if ((in_atomic() || irqs_disabled()) &&
6366 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6367 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6369 prev_jiffy = jiffies;
6370 printk(KERN_ERR "BUG: sleeping function called from invalid"
6371 " context at %s:%d\n", file, line);
6372 printk("in_atomic():%d, irqs_disabled():%d\n",
6373 in_atomic(), irqs_disabled());
6378 EXPORT_SYMBOL(__might_sleep);
6381 #ifdef CONFIG_MAGIC_SYSRQ
6382 void normalize_rt_tasks(void)
6384 struct task_struct *p;
6385 prio_array_t *array;
6386 unsigned long flags;
6389 read_lock_irq(&tasklist_lock);
6390 for_each_process (p) {
6394 rq = task_rq_lock(p, &flags);
6398 deactivate_task(p, task_rq(p));
6399 __setscheduler(p, SCHED_NORMAL, 0);
6401 vx_activate_task(p);
6402 __activate_task(p, task_rq(p));
6403 resched_task(rq->curr);
6406 task_rq_unlock(rq, &flags);
6408 read_unlock_irq(&tasklist_lock);
6411 #endif /* CONFIG_MAGIC_SYSRQ */
6415 * These functions are only useful for the IA64 MCA handling.
6417 * They can only be called when the whole system has been
6418 * stopped - every CPU needs to be quiescent, and no scheduling
6419 * activity can take place. Using them for anything else would
6420 * be a serious bug, and as a result, they aren't even visible
6421 * under any other configuration.
6425 * curr_task - return the current task for a given cpu.
6426 * @cpu: the processor in question.
6428 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6430 task_t *curr_task(int cpu)
6432 return cpu_curr(cpu);
6436 * set_curr_task - set the current task for a given cpu.
6437 * @cpu: the processor in question.
6438 * @p: the task pointer to set.
6440 * Description: This function must only be used when non-maskable interrupts
6441 * are serviced on a separate stack. It allows the architecture to switch the
6442 * notion of the current task on a cpu in a non-blocking manner. This function
6443 * must be called with all CPU's synchronized, and interrupts disabled, the
6444 * and caller must save the original value of the current task (see
6445 * curr_task() above) and restore that value before reenabling interrupts and
6446 * re-starting the system.
6448 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6450 void set_curr_task(int cpu, task_t *p)