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/debug_locks.h>
34 #include <linux/security.h>
35 #include <linux/notifier.h>
36 #include <linux/profile.h>
37 #include <linux/suspend.h>
38 #include <linux/vmalloc.h>
39 #include <linux/blkdev.h>
40 #include <linux/delay.h>
41 #include <linux/smp.h>
42 #include <linux/threads.h>
43 #include <linux/timer.h>
44 #include <linux/rcupdate.h>
45 #include <linux/cpu.h>
46 #include <linux/cpuset.h>
47 #include <linux/percpu.h>
48 #include <linux/kthread.h>
49 #include <linux/seq_file.h>
50 #include <linux/syscalls.h>
51 #include <linux/times.h>
52 #include <linux/acct.h>
53 #include <linux/kprobes.h>
54 #include <linux/delayacct.h>
55 #include <linux/vs_base.h>
56 #include <linux/vs_memory.h>
57 #include <linux/vs_context.h>
58 #include <linux/vs_cvirt.h>
59 #include <linux/vs_sched.h>
62 #include <asm/unistd.h>
65 * Convert user-nice values [ -20 ... 0 ... 19 ]
66 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
69 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
70 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
71 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
74 * 'User priority' is the nice value converted to something we
75 * can work with better when scaling various scheduler parameters,
76 * it's a [ 0 ... 39 ] range.
78 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
79 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
80 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
83 * Some helpers for converting nanosecond timing to jiffy resolution
85 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
86 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
89 * These are the 'tuning knobs' of the scheduler:
91 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
92 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
93 * Timeslices get refilled after they expire.
95 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
96 #define DEF_TIMESLICE (100 * HZ / 1000)
97 #define ON_RUNQUEUE_WEIGHT 30
98 #define CHILD_PENALTY 95
99 #define PARENT_PENALTY 100
100 #define EXIT_WEIGHT 3
101 #define PRIO_BONUS_RATIO 25
102 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
103 #define INTERACTIVE_DELTA 2
104 #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
105 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
106 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
109 * If a task is 'interactive' then we reinsert it in the active
110 * array after it has expired its current timeslice. (it will not
111 * continue to run immediately, it will still roundrobin with
112 * other interactive tasks.)
114 * This part scales the interactivity limit depending on niceness.
116 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
117 * Here are a few examples of different nice levels:
119 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
120 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
121 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
122 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
123 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
125 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
126 * priority range a task can explore, a value of '1' means the
127 * task is rated interactive.)
129 * Ie. nice +19 tasks can never get 'interactive' enough to be
130 * reinserted into the active array. And only heavily CPU-hog nice -20
131 * tasks will be expired. Default nice 0 tasks are somewhere between,
132 * it takes some effort for them to get interactive, but it's not
136 #define CURRENT_BONUS(p) \
137 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
140 #define GRANULARITY (10 * HZ / 1000 ? : 1)
143 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
144 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
147 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
148 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
151 #define SCALE(v1,v1_max,v2_max) \
152 (v1) * (v2_max) / (v1_max)
155 (SCALE(TASK_NICE(p) + 20, 40, MAX_BONUS) - 20 * MAX_BONUS / 40 + \
158 #define TASK_INTERACTIVE(p) \
159 ((p)->prio <= (p)->static_prio - DELTA(p))
161 #define INTERACTIVE_SLEEP(p) \
162 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
163 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
165 #define TASK_PREEMPTS_CURR(p, rq) \
166 ((p)->prio < (rq)->curr->prio)
169 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
170 * to time slice values: [800ms ... 100ms ... 5ms]
172 * The higher a thread's priority, the bigger timeslices
173 * it gets during one round of execution. But even the lowest
174 * priority thread gets MIN_TIMESLICE worth of execution time.
177 #define SCALE_PRIO(x, prio) \
178 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
180 static unsigned int static_prio_timeslice(int static_prio)
182 if (static_prio < NICE_TO_PRIO(0))
183 return SCALE_PRIO(DEF_TIMESLICE * 4, static_prio);
185 return SCALE_PRIO(DEF_TIMESLICE, static_prio);
188 static inline unsigned int task_timeslice(struct task_struct *p)
190 return static_prio_timeslice(p->static_prio);
194 * These are the runqueue data structures:
198 unsigned int nr_active;
199 DECLARE_BITMAP(bitmap, MAX_PRIO+1); /* include 1 bit for delimiter */
200 struct list_head queue[MAX_PRIO];
204 * This is the main, per-CPU runqueue data structure.
206 * Locking rule: those places that want to lock multiple runqueues
207 * (such as the load balancing or the thread migration code), lock
208 * acquire operations must be ordered by ascending &runqueue.
214 * nr_running and cpu_load should be in the same cacheline because
215 * remote CPUs use both these fields when doing load calculation.
217 unsigned long nr_running;
218 unsigned long raw_weighted_load;
220 unsigned long cpu_load[3];
222 unsigned long long nr_switches;
225 * This is part of a global counter where only the total sum
226 * over all CPUs matters. A task can increase this counter on
227 * one CPU and if it got migrated afterwards it may decrease
228 * it on another CPU. Always updated under the runqueue lock:
230 unsigned long nr_uninterruptible;
232 unsigned long expired_timestamp;
233 unsigned long long timestamp_last_tick;
234 struct task_struct *curr, *idle;
235 struct mm_struct *prev_mm;
236 struct prio_array *active, *expired, arrays[2];
237 int best_expired_prio;
241 struct sched_domain *sd;
243 /* For active balancing */
246 int cpu; /* cpu of this runqueue */
248 struct task_struct *migration_thread;
249 struct list_head migration_queue;
251 #ifdef CONFIG_VSERVER_HARDCPU
252 struct list_head hold_queue;
256 #ifdef CONFIG_SCHEDSTATS
258 struct sched_info rq_sched_info;
260 /* sys_sched_yield() stats */
261 unsigned long yld_exp_empty;
262 unsigned long yld_act_empty;
263 unsigned long yld_both_empty;
264 unsigned long yld_cnt;
266 /* schedule() stats */
267 unsigned long sched_switch;
268 unsigned long sched_cnt;
269 unsigned long sched_goidle;
271 /* try_to_wake_up() stats */
272 unsigned long ttwu_cnt;
273 unsigned long ttwu_local;
275 struct lock_class_key rq_lock_key;
278 static DEFINE_PER_CPU(struct rq, runqueues);
280 static inline int cpu_of(struct rq *rq)
290 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
291 * See detach_destroy_domains: synchronize_sched for details.
293 * The domain tree of any CPU may only be accessed from within
294 * preempt-disabled sections.
296 #define for_each_domain(cpu, __sd) \
297 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
299 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
300 #define this_rq() (&__get_cpu_var(runqueues))
301 #define task_rq(p) cpu_rq(task_cpu(p))
302 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
304 #ifndef prepare_arch_switch
305 # define prepare_arch_switch(next) do { } while (0)
307 #ifndef finish_arch_switch
308 # define finish_arch_switch(prev) do { } while (0)
311 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
312 static inline int task_running(struct rq *rq, struct task_struct *p)
314 return rq->curr == p;
317 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
321 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
323 #ifdef CONFIG_DEBUG_SPINLOCK
324 /* this is a valid case when another task releases the spinlock */
325 rq->lock.owner = current;
328 * If we are tracking spinlock dependencies then we have to
329 * fix up the runqueue lock - which gets 'carried over' from
332 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
334 spin_unlock_irq(&rq->lock);
337 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
338 static inline int task_running(struct rq *rq, struct task_struct *p)
343 return rq->curr == p;
347 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
351 * We can optimise this out completely for !SMP, because the
352 * SMP rebalancing from interrupt is the only thing that cares
357 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
358 spin_unlock_irq(&rq->lock);
360 spin_unlock(&rq->lock);
364 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
368 * After ->oncpu is cleared, the task can be moved to a different CPU.
369 * We must ensure this doesn't happen until the switch is completely
375 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
379 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
382 * __task_rq_lock - lock the runqueue a given task resides on.
383 * Must be called interrupts disabled.
385 static inline struct rq *__task_rq_lock(struct task_struct *p)
392 spin_lock(&rq->lock);
393 if (unlikely(rq != task_rq(p))) {
394 spin_unlock(&rq->lock);
395 goto repeat_lock_task;
401 * task_rq_lock - lock the runqueue a given task resides on and disable
402 * interrupts. Note the ordering: we can safely lookup the task_rq without
403 * explicitly disabling preemption.
405 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
411 local_irq_save(*flags);
413 spin_lock(&rq->lock);
414 if (unlikely(rq != task_rq(p))) {
415 spin_unlock_irqrestore(&rq->lock, *flags);
416 goto repeat_lock_task;
421 static inline void __task_rq_unlock(struct rq *rq)
424 spin_unlock(&rq->lock);
427 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
430 spin_unlock_irqrestore(&rq->lock, *flags);
433 #ifdef CONFIG_SCHEDSTATS
435 * bump this up when changing the output format or the meaning of an existing
436 * format, so that tools can adapt (or abort)
438 #define SCHEDSTAT_VERSION 12
440 static int show_schedstat(struct seq_file *seq, void *v)
444 seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION);
445 seq_printf(seq, "timestamp %lu\n", jiffies);
446 for_each_online_cpu(cpu) {
447 struct rq *rq = cpu_rq(cpu);
449 struct sched_domain *sd;
453 /* runqueue-specific stats */
455 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
456 cpu, rq->yld_both_empty,
457 rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt,
458 rq->sched_switch, rq->sched_cnt, rq->sched_goidle,
459 rq->ttwu_cnt, rq->ttwu_local,
460 rq->rq_sched_info.cpu_time,
461 rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt);
463 seq_printf(seq, "\n");
466 /* domain-specific stats */
468 for_each_domain(cpu, sd) {
469 enum idle_type itype;
470 char mask_str[NR_CPUS];
472 cpumask_scnprintf(mask_str, NR_CPUS, sd->span);
473 seq_printf(seq, "domain%d %s", dcnt++, mask_str);
474 for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES;
476 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu",
478 sd->lb_balanced[itype],
479 sd->lb_failed[itype],
480 sd->lb_imbalance[itype],
481 sd->lb_gained[itype],
482 sd->lb_hot_gained[itype],
483 sd->lb_nobusyq[itype],
484 sd->lb_nobusyg[itype]);
486 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
487 sd->alb_cnt, sd->alb_failed, sd->alb_pushed,
488 sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed,
489 sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed,
490 sd->ttwu_wake_remote, sd->ttwu_move_affine, sd->ttwu_move_balance);
498 static int schedstat_open(struct inode *inode, struct file *file)
500 unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32);
501 char *buf = kmalloc(size, GFP_KERNEL);
507 res = single_open(file, show_schedstat, NULL);
509 m = file->private_data;
517 struct file_operations proc_schedstat_operations = {
518 .open = schedstat_open,
521 .release = single_release,
525 * Expects runqueue lock to be held for atomicity of update
528 rq_sched_info_arrive(struct rq *rq, unsigned long delta_jiffies)
531 rq->rq_sched_info.run_delay += delta_jiffies;
532 rq->rq_sched_info.pcnt++;
537 * Expects runqueue lock to be held for atomicity of update
540 rq_sched_info_depart(struct rq *rq, unsigned long delta_jiffies)
543 rq->rq_sched_info.cpu_time += delta_jiffies;
545 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
546 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
547 #else /* !CONFIG_SCHEDSTATS */
549 rq_sched_info_arrive(struct rq *rq, unsigned long delta_jiffies)
552 rq_sched_info_depart(struct rq *rq, unsigned long delta_jiffies)
554 # define schedstat_inc(rq, field) do { } while (0)
555 # define schedstat_add(rq, field, amt) do { } while (0)
559 * rq_lock - lock a given runqueue and disable interrupts.
561 static inline struct rq *this_rq_lock(void)
568 spin_lock(&rq->lock);
573 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
575 * Called when a process is dequeued from the active array and given
576 * the cpu. We should note that with the exception of interactive
577 * tasks, the expired queue will become the active queue after the active
578 * queue is empty, without explicitly dequeuing and requeuing tasks in the
579 * expired queue. (Interactive tasks may be requeued directly to the
580 * active queue, thus delaying tasks in the expired queue from running;
581 * see scheduler_tick()).
583 * This function is only called from sched_info_arrive(), rather than
584 * dequeue_task(). Even though a task may be queued and dequeued multiple
585 * times as it is shuffled about, we're really interested in knowing how
586 * long it was from the *first* time it was queued to the time that it
589 static inline void sched_info_dequeued(struct task_struct *t)
591 t->sched_info.last_queued = 0;
595 * Called when a task finally hits the cpu. We can now calculate how
596 * long it was waiting to run. We also note when it began so that we
597 * can keep stats on how long its timeslice is.
599 static void sched_info_arrive(struct task_struct *t)
601 unsigned long now = jiffies, delta_jiffies = 0;
603 if (t->sched_info.last_queued)
604 delta_jiffies = now - t->sched_info.last_queued;
605 sched_info_dequeued(t);
606 t->sched_info.run_delay += delta_jiffies;
607 t->sched_info.last_arrival = now;
608 t->sched_info.pcnt++;
610 rq_sched_info_arrive(task_rq(t), delta_jiffies);
614 * Called when a process is queued into either the active or expired
615 * array. The time is noted and later used to determine how long we
616 * had to wait for us to reach the cpu. Since the expired queue will
617 * become the active queue after active queue is empty, without dequeuing
618 * and requeuing any tasks, we are interested in queuing to either. It
619 * is unusual but not impossible for tasks to be dequeued and immediately
620 * requeued in the same or another array: this can happen in sched_yield(),
621 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
624 * This function is only called from enqueue_task(), but also only updates
625 * the timestamp if it is already not set. It's assumed that
626 * sched_info_dequeued() will clear that stamp when appropriate.
628 static inline void sched_info_queued(struct task_struct *t)
630 if (unlikely(sched_info_on()))
631 if (!t->sched_info.last_queued)
632 t->sched_info.last_queued = jiffies;
636 * Called when a process ceases being the active-running process, either
637 * voluntarily or involuntarily. Now we can calculate how long we ran.
639 static inline void sched_info_depart(struct task_struct *t)
641 unsigned long delta_jiffies = jiffies - t->sched_info.last_arrival;
643 t->sched_info.cpu_time += delta_jiffies;
644 rq_sched_info_depart(task_rq(t), delta_jiffies);
648 * Called when tasks are switched involuntarily due, typically, to expiring
649 * their time slice. (This may also be called when switching to or from
650 * the idle task.) We are only called when prev != next.
653 __sched_info_switch(struct task_struct *prev, struct task_struct *next)
655 struct rq *rq = task_rq(prev);
658 * prev now departs the cpu. It's not interesting to record
659 * stats about how efficient we were at scheduling the idle
662 if (prev != rq->idle)
663 sched_info_depart(prev);
665 if (next != rq->idle)
666 sched_info_arrive(next);
669 sched_info_switch(struct task_struct *prev, struct task_struct *next)
671 if (unlikely(sched_info_on()))
672 __sched_info_switch(prev, next);
675 #define sched_info_queued(t) do { } while (0)
676 #define sched_info_switch(t, next) do { } while (0)
677 #endif /* CONFIG_SCHEDSTATS || CONFIG_TASK_DELAY_ACCT */
680 * Adding/removing a task to/from a priority array:
682 static void dequeue_task(struct task_struct *p, struct prio_array *array)
684 BUG_ON(p->state & TASK_ONHOLD);
686 list_del(&p->run_list);
687 if (list_empty(array->queue + p->prio))
688 __clear_bit(p->prio, array->bitmap);
691 static void enqueue_task(struct task_struct *p, struct prio_array *array)
693 BUG_ON(p->state & TASK_ONHOLD);
694 sched_info_queued(p);
695 list_add_tail(&p->run_list, array->queue + p->prio);
696 __set_bit(p->prio, array->bitmap);
702 * Put task to the end of the run list without the overhead of dequeue
703 * followed by enqueue.
705 static void requeue_task(struct task_struct *p, struct prio_array *array)
707 BUG_ON(p->state & TASK_ONHOLD);
708 list_move_tail(&p->run_list, array->queue + p->prio);
712 enqueue_task_head(struct task_struct *p, struct prio_array *array)
714 BUG_ON(p->state & TASK_ONHOLD);
715 list_add(&p->run_list, array->queue + p->prio);
716 __set_bit(p->prio, array->bitmap);
722 * __normal_prio - return the priority that is based on the static
723 * priority but is modified by bonuses/penalties.
725 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
726 * into the -5 ... 0 ... +5 bonus/penalty range.
728 * We use 25% of the full 0...39 priority range so that:
730 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
731 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
733 * Both properties are important to certain workloads.
736 static inline int __normal_prio(struct task_struct *p)
741 bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
743 prio = p->static_prio - bonus;
745 if ((vxi = p->vx_info) &&
746 vx_info_flags(vxi, VXF_SCHED_PRIO, 0))
747 prio += vx_effective_vavavoom(vxi, MAX_USER_PRIO);
749 if (prio < MAX_RT_PRIO)
751 if (prio > MAX_PRIO-1)
757 * To aid in avoiding the subversion of "niceness" due to uneven distribution
758 * of tasks with abnormal "nice" values across CPUs the contribution that
759 * each task makes to its run queue's load is weighted according to its
760 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
761 * scaled version of the new time slice allocation that they receive on time
766 * Assume: static_prio_timeslice(NICE_TO_PRIO(0)) == DEF_TIMESLICE
767 * If static_prio_timeslice() is ever changed to break this assumption then
768 * this code will need modification
770 #define TIME_SLICE_NICE_ZERO DEF_TIMESLICE
771 #define LOAD_WEIGHT(lp) \
772 (((lp) * SCHED_LOAD_SCALE) / TIME_SLICE_NICE_ZERO)
773 #define PRIO_TO_LOAD_WEIGHT(prio) \
774 LOAD_WEIGHT(static_prio_timeslice(prio))
775 #define RTPRIO_TO_LOAD_WEIGHT(rp) \
776 (PRIO_TO_LOAD_WEIGHT(MAX_RT_PRIO) + LOAD_WEIGHT(rp))
778 static void set_load_weight(struct task_struct *p)
780 if (has_rt_policy(p)) {
782 if (p == task_rq(p)->migration_thread)
784 * The migration thread does the actual balancing.
785 * Giving its load any weight will skew balancing
791 p->load_weight = RTPRIO_TO_LOAD_WEIGHT(p->rt_priority);
793 p->load_weight = PRIO_TO_LOAD_WEIGHT(p->static_prio);
797 inc_raw_weighted_load(struct rq *rq, const struct task_struct *p)
799 rq->raw_weighted_load += p->load_weight;
803 dec_raw_weighted_load(struct rq *rq, const struct task_struct *p)
805 rq->raw_weighted_load -= p->load_weight;
808 static inline void inc_nr_running(struct task_struct *p, struct rq *rq)
811 inc_raw_weighted_load(rq, p);
814 static inline void dec_nr_running(struct task_struct *p, struct rq *rq)
817 dec_raw_weighted_load(rq, p);
821 * Calculate the expected normal priority: i.e. priority
822 * without taking RT-inheritance into account. Might be
823 * boosted by interactivity modifiers. Changes upon fork,
824 * setprio syscalls, and whenever the interactivity
825 * estimator recalculates.
827 static inline int normal_prio(struct task_struct *p)
831 if (has_rt_policy(p))
832 prio = MAX_RT_PRIO-1 - p->rt_priority;
834 prio = __normal_prio(p);
839 * Calculate the current priority, i.e. the priority
840 * taken into account by the scheduler. This value might
841 * be boosted by RT tasks, or might be boosted by
842 * interactivity modifiers. Will be RT if the task got
843 * RT-boosted. If not then it returns p->normal_prio.
845 static int effective_prio(struct task_struct *p)
847 p->normal_prio = normal_prio(p);
849 * If we are RT tasks or we were boosted to RT priority,
850 * keep the priority unchanged. Otherwise, update priority
851 * to the normal priority:
853 if (!rt_prio(p->prio))
854 return p->normal_prio;
859 * __activate_task - move a task to the runqueue.
861 static void __activate_task(struct task_struct *p, struct rq *rq)
863 struct prio_array *target = rq->active;
866 target = rq->expired;
867 enqueue_task(p, target);
868 inc_nr_running(p, rq);
872 * __activate_idle_task - move idle task to the _front_ of runqueue.
874 static inline void __activate_idle_task(struct task_struct *p, struct rq *rq)
876 enqueue_task_head(p, rq->active);
877 inc_nr_running(p, rq);
881 * Recalculate p->normal_prio and p->prio after having slept,
882 * updating the sleep-average too:
884 static int recalc_task_prio(struct task_struct *p, unsigned long long now)
886 /* Caller must always ensure 'now >= p->timestamp' */
887 unsigned long sleep_time = now - p->timestamp;
892 if (likely(sleep_time > 0)) {
894 * This ceiling is set to the lowest priority that would allow
895 * a task to be reinserted into the active array on timeslice
898 unsigned long ceiling = INTERACTIVE_SLEEP(p);
900 if (p->mm && sleep_time > ceiling && p->sleep_avg < ceiling) {
902 * Prevents user tasks from achieving best priority
903 * with one single large enough sleep.
905 p->sleep_avg = ceiling;
907 * Using INTERACTIVE_SLEEP() as a ceiling places a
908 * nice(0) task 1ms sleep away from promotion, and
909 * gives it 700ms to round-robin with no chance of
910 * being demoted. This is more than generous, so
911 * mark this sleep as non-interactive to prevent the
912 * on-runqueue bonus logic from intervening should
913 * this task not receive cpu immediately.
915 p->sleep_type = SLEEP_NONINTERACTIVE;
918 * Tasks waking from uninterruptible sleep are
919 * limited in their sleep_avg rise as they
920 * are likely to be waiting on I/O
922 if (p->sleep_type == SLEEP_NONINTERACTIVE && p->mm) {
923 if (p->sleep_avg >= ceiling)
925 else if (p->sleep_avg + sleep_time >=
927 p->sleep_avg = ceiling;
933 * This code gives a bonus to interactive tasks.
935 * The boost works by updating the 'average sleep time'
936 * value here, based on ->timestamp. The more time a
937 * task spends sleeping, the higher the average gets -
938 * and the higher the priority boost gets as well.
940 p->sleep_avg += sleep_time;
943 if (p->sleep_avg > NS_MAX_SLEEP_AVG)
944 p->sleep_avg = NS_MAX_SLEEP_AVG;
947 return effective_prio(p);
951 * activate_task - move a task to the runqueue and do priority recalculation
953 * Update all the scheduling statistics stuff. (sleep average
954 * calculation, priority modifiers, etc.)
956 static void activate_task(struct task_struct *p, struct rq *rq, int local)
958 unsigned long long now;
963 /* Compensate for drifting sched_clock */
964 struct rq *this_rq = this_rq();
965 now = (now - this_rq->timestamp_last_tick)
966 + rq->timestamp_last_tick;
971 p->prio = recalc_task_prio(p, now);
974 * This checks to make sure it's not an uninterruptible task
975 * that is now waking up.
977 if (p->sleep_type == SLEEP_NORMAL) {
979 * Tasks which were woken up by interrupts (ie. hw events)
980 * are most likely of interactive nature. So we give them
981 * the credit of extending their sleep time to the period
982 * of time they spend on the runqueue, waiting for execution
983 * on a CPU, first time around:
986 p->sleep_type = SLEEP_INTERRUPTED;
989 * Normal first-time wakeups get a credit too for
990 * on-runqueue time, but it will be weighted down:
992 p->sleep_type = SLEEP_INTERACTIVE;
998 __activate_task(p, rq);
1002 * deactivate_task - remove a task from the runqueue.
1004 static void __deactivate_task(struct task_struct *p, struct rq *rq)
1006 dec_nr_running(p, rq);
1007 dequeue_task(p, p->array);
1012 void deactivate_task(struct task_struct *p, struct rq *rq)
1014 vx_deactivate_task(p);
1015 __deactivate_task(p, rq);
1019 #ifdef CONFIG_VSERVER_HARDCPU
1021 * vx_hold_task - put a task on the hold queue
1024 void vx_hold_task(struct vx_info *vxi,
1025 struct task_struct *p, struct rq *rq)
1027 __deactivate_task(p, rq);
1028 p->state |= TASK_ONHOLD;
1029 /* a new one on hold */
1031 list_add_tail(&p->run_list, &rq->hold_queue);
1035 * vx_unhold_task - put a task back to the runqueue
1038 void vx_unhold_task(struct vx_info *vxi,
1039 struct task_struct *p, struct rq *rq)
1041 list_del(&p->run_list);
1042 /* one less waiting */
1044 p->state &= ~TASK_ONHOLD;
1045 enqueue_task(p, rq->expired);
1046 inc_nr_running(p, rq);
1048 if (p->static_prio < rq->best_expired_prio)
1049 rq->best_expired_prio = p->static_prio;
1053 void vx_hold_task(struct vx_info *vxi,
1054 struct task_struct *p, struct rq *rq)
1060 void vx_unhold_task(struct vx_info *vxi,
1061 struct task_struct *p, struct rq *rq)
1065 #endif /* CONFIG_VSERVER_HARDCPU */
1069 * resched_task - mark a task 'to be rescheduled now'.
1071 * On UP this means the setting of the need_resched flag, on SMP it
1072 * might also involve a cross-CPU call to trigger the scheduler on
1077 #ifndef tsk_is_polling
1078 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1081 static void resched_task(struct task_struct *p)
1085 assert_spin_locked(&task_rq(p)->lock);
1087 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1090 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1093 if (cpu == smp_processor_id())
1096 /* NEED_RESCHED must be visible before we test polling */
1098 if (!tsk_is_polling(p))
1099 smp_send_reschedule(cpu);
1102 static inline void resched_task(struct task_struct *p)
1104 assert_spin_locked(&task_rq(p)->lock);
1105 set_tsk_need_resched(p);
1110 * task_curr - is this task currently executing on a CPU?
1111 * @p: the task in question.
1113 inline int task_curr(const struct task_struct *p)
1115 return cpu_curr(task_cpu(p)) == p;
1118 /* Used instead of source_load when we know the type == 0 */
1119 unsigned long weighted_cpuload(const int cpu)
1121 return cpu_rq(cpu)->raw_weighted_load;
1125 struct migration_req {
1126 struct list_head list;
1128 struct task_struct *task;
1131 struct completion done;
1135 * The task's runqueue lock must be held.
1136 * Returns true if you have to wait for migration thread.
1139 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1141 struct rq *rq = task_rq(p);
1144 * If the task is not on a runqueue (and not running), then
1145 * it is sufficient to simply update the task's cpu field.
1147 if (!p->array && !task_running(rq, p)) {
1148 set_task_cpu(p, dest_cpu);
1152 init_completion(&req->done);
1154 req->dest_cpu = dest_cpu;
1155 list_add(&req->list, &rq->migration_queue);
1161 * wait_task_inactive - wait for a thread to unschedule.
1163 * The caller must ensure that the task *will* unschedule sometime soon,
1164 * else this function might spin for a *long* time. This function can't
1165 * be called with interrupts off, or it may introduce deadlock with
1166 * smp_call_function() if an IPI is sent by the same process we are
1167 * waiting to become inactive.
1169 void wait_task_inactive(struct task_struct *p)
1171 unsigned long flags;
1176 rq = task_rq_lock(p, &flags);
1177 /* Must be off runqueue entirely, not preempted. */
1178 if (unlikely(p->array || task_running(rq, p))) {
1179 /* If it's preempted, we yield. It could be a while. */
1180 preempted = !task_running(rq, p);
1181 task_rq_unlock(rq, &flags);
1187 task_rq_unlock(rq, &flags);
1191 * kick_process - kick a running thread to enter/exit the kernel
1192 * @p: the to-be-kicked thread
1194 * Cause a process which is running on another CPU to enter
1195 * kernel-mode, without any delay. (to get signals handled.)
1197 * NOTE: this function doesnt have to take the runqueue lock,
1198 * because all it wants to ensure is that the remote task enters
1199 * the kernel. If the IPI races and the task has been migrated
1200 * to another CPU then no harm is done and the purpose has been
1203 void kick_process(struct task_struct *p)
1209 if ((cpu != smp_processor_id()) && task_curr(p))
1210 smp_send_reschedule(cpu);
1215 * Return a low guess at the load of a migration-source cpu weighted
1216 * according to the scheduling class and "nice" value.
1218 * We want to under-estimate the load of migration sources, to
1219 * balance conservatively.
1221 static inline unsigned long source_load(int cpu, int type)
1223 struct rq *rq = cpu_rq(cpu);
1226 return rq->raw_weighted_load;
1228 return min(rq->cpu_load[type-1], rq->raw_weighted_load);
1232 * Return a high guess at the load of a migration-target cpu weighted
1233 * according to the scheduling class and "nice" value.
1235 static inline unsigned long target_load(int cpu, int type)
1237 struct rq *rq = cpu_rq(cpu);
1240 return rq->raw_weighted_load;
1242 return max(rq->cpu_load[type-1], rq->raw_weighted_load);
1246 * Return the average load per task on the cpu's run queue
1248 static inline unsigned long cpu_avg_load_per_task(int cpu)
1250 struct rq *rq = cpu_rq(cpu);
1251 unsigned long n = rq->nr_running;
1253 return n ? rq->raw_weighted_load / n : SCHED_LOAD_SCALE;
1257 * find_idlest_group finds and returns the least busy CPU group within the
1260 static struct sched_group *
1261 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1263 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1264 unsigned long min_load = ULONG_MAX, this_load = 0;
1265 int load_idx = sd->forkexec_idx;
1266 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1269 unsigned long load, avg_load;
1273 /* Skip over this group if it has no CPUs allowed */
1274 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1277 local_group = cpu_isset(this_cpu, group->cpumask);
1279 /* Tally up the load of all CPUs in the group */
1282 for_each_cpu_mask(i, group->cpumask) {
1283 /* Bias balancing toward cpus of our domain */
1285 load = source_load(i, load_idx);
1287 load = target_load(i, load_idx);
1292 /* Adjust by relative CPU power of the group */
1293 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1296 this_load = avg_load;
1298 } else if (avg_load < min_load) {
1299 min_load = avg_load;
1303 group = group->next;
1304 } while (group != sd->groups);
1306 if (!idlest || 100*this_load < imbalance*min_load)
1312 * find_idlest_queue - find the idlest runqueue among the cpus in group.
1315 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1318 unsigned long load, min_load = ULONG_MAX;
1322 /* Traverse only the allowed CPUs */
1323 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1325 for_each_cpu_mask(i, tmp) {
1326 load = weighted_cpuload(i);
1328 if (load < min_load || (load == min_load && i == this_cpu)) {
1338 * sched_balance_self: balance the current task (running on cpu) in domains
1339 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1342 * Balance, ie. select the least loaded group.
1344 * Returns the target CPU number, or the same CPU if no balancing is needed.
1346 * preempt must be disabled.
1348 static int sched_balance_self(int cpu, int flag)
1350 struct task_struct *t = current;
1351 struct sched_domain *tmp, *sd = NULL;
1353 for_each_domain(cpu, tmp) {
1355 * If power savings logic is enabled for a domain, stop there.
1357 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1359 if (tmp->flags & flag)
1365 struct sched_group *group;
1370 group = find_idlest_group(sd, t, cpu);
1374 new_cpu = find_idlest_cpu(group, t, cpu);
1375 if (new_cpu == -1 || new_cpu == cpu)
1378 /* Now try balancing at a lower domain level */
1382 weight = cpus_weight(span);
1383 for_each_domain(cpu, tmp) {
1384 if (weight <= cpus_weight(tmp->span))
1386 if (tmp->flags & flag)
1389 /* while loop will break here if sd == NULL */
1395 #endif /* CONFIG_SMP */
1398 * wake_idle() will wake a task on an idle cpu if task->cpu is
1399 * not idle and an idle cpu is available. The span of cpus to
1400 * search starts with cpus closest then further out as needed,
1401 * so we always favor a closer, idle cpu.
1403 * Returns the CPU we should wake onto.
1405 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1406 static int wake_idle(int cpu, struct task_struct *p)
1409 struct sched_domain *sd;
1415 for_each_domain(cpu, sd) {
1416 if (sd->flags & SD_WAKE_IDLE) {
1417 cpus_and(tmp, sd->span, p->cpus_allowed);
1418 for_each_cpu_mask(i, tmp) {
1429 static inline int wake_idle(int cpu, struct task_struct *p)
1436 * try_to_wake_up - wake up a thread
1437 * @p: the to-be-woken-up thread
1438 * @state: the mask of task states that can be woken
1439 * @sync: do a synchronous wakeup?
1441 * Put it on the run-queue if it's not already there. The "current"
1442 * thread is always on the run-queue (except when the actual
1443 * re-schedule is in progress), and as such you're allowed to do
1444 * the simpler "current->state = TASK_RUNNING" to mark yourself
1445 * runnable without the overhead of this.
1447 * returns failure only if the task is already active.
1449 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1451 int cpu, this_cpu, success = 0;
1452 unsigned long flags;
1456 struct sched_domain *sd, *this_sd = NULL;
1457 unsigned long load, this_load;
1461 rq = task_rq_lock(p, &flags);
1462 old_state = p->state;
1464 /* we need to unhold suspended tasks */
1465 if (old_state & TASK_ONHOLD) {
1466 vx_unhold_task(p->vx_info, p, rq);
1467 old_state = p->state;
1469 if (!(old_state & state))
1476 this_cpu = smp_processor_id();
1479 if (unlikely(task_running(rq, p)))
1484 schedstat_inc(rq, ttwu_cnt);
1485 if (cpu == this_cpu) {
1486 schedstat_inc(rq, ttwu_local);
1490 for_each_domain(this_cpu, sd) {
1491 if (cpu_isset(cpu, sd->span)) {
1492 schedstat_inc(sd, ttwu_wake_remote);
1498 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1502 * Check for affine wakeup and passive balancing possibilities.
1505 int idx = this_sd->wake_idx;
1506 unsigned int imbalance;
1508 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1510 load = source_load(cpu, idx);
1511 this_load = target_load(this_cpu, idx);
1513 new_cpu = this_cpu; /* Wake to this CPU if we can */
1515 if (this_sd->flags & SD_WAKE_AFFINE) {
1516 unsigned long tl = this_load;
1517 unsigned long tl_per_task = cpu_avg_load_per_task(this_cpu);
1520 * If sync wakeup then subtract the (maximum possible)
1521 * effect of the currently running task from the load
1522 * of the current CPU:
1525 tl -= current->load_weight;
1528 tl + target_load(cpu, idx) <= tl_per_task) ||
1529 100*(tl + p->load_weight) <= imbalance*load) {
1531 * This domain has SD_WAKE_AFFINE and
1532 * p is cache cold in this domain, and
1533 * there is no bad imbalance.
1535 schedstat_inc(this_sd, ttwu_move_affine);
1541 * Start passive balancing when half the imbalance_pct
1544 if (this_sd->flags & SD_WAKE_BALANCE) {
1545 if (imbalance*this_load <= 100*load) {
1546 schedstat_inc(this_sd, ttwu_move_balance);
1552 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1554 new_cpu = wake_idle(new_cpu, p);
1555 if (new_cpu != cpu) {
1556 set_task_cpu(p, new_cpu);
1557 task_rq_unlock(rq, &flags);
1558 /* might preempt at this point */
1559 rq = task_rq_lock(p, &flags);
1560 old_state = p->state;
1561 if (!(old_state & state))
1566 this_cpu = smp_processor_id();
1571 #endif /* CONFIG_SMP */
1572 if (old_state == TASK_UNINTERRUPTIBLE) {
1573 rq->nr_uninterruptible--;
1574 vx_uninterruptible_dec(p);
1576 * Tasks on involuntary sleep don't earn
1577 * sleep_avg beyond just interactive state.
1579 p->sleep_type = SLEEP_NONINTERACTIVE;
1583 * Tasks that have marked their sleep as noninteractive get
1584 * woken up with their sleep average not weighted in an
1587 if (old_state & TASK_NONINTERACTIVE)
1588 p->sleep_type = SLEEP_NONINTERACTIVE;
1591 activate_task(p, rq, cpu == this_cpu);
1593 * Sync wakeups (i.e. those types of wakeups where the waker
1594 * has indicated that it will leave the CPU in short order)
1595 * don't trigger a preemption, if the woken up task will run on
1596 * this cpu. (in this case the 'I will reschedule' promise of
1597 * the waker guarantees that the freshly woken up task is going
1598 * to be considered on this CPU.)
1600 if (!sync || cpu != this_cpu) {
1601 if (TASK_PREEMPTS_CURR(p, rq))
1602 resched_task(rq->curr);
1607 p->state = TASK_RUNNING;
1609 task_rq_unlock(rq, &flags);
1614 int fastcall wake_up_process(struct task_struct *p)
1616 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1617 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1619 EXPORT_SYMBOL(wake_up_process);
1621 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1623 return try_to_wake_up(p, state, 0);
1627 * Perform scheduler related setup for a newly forked process p.
1628 * p is forked by current.
1630 void fastcall sched_fork(struct task_struct *p, int clone_flags)
1632 int cpu = get_cpu();
1635 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1637 set_task_cpu(p, cpu);
1640 * We mark the process as running here, but have not actually
1641 * inserted it onto the runqueue yet. This guarantees that
1642 * nobody will actually run it, and a signal or other external
1643 * event cannot wake it up and insert it on the runqueue either.
1645 p->state = TASK_RUNNING;
1648 * Make sure we do not leak PI boosting priority to the child:
1650 p->prio = current->normal_prio;
1652 INIT_LIST_HEAD(&p->run_list);
1654 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1655 if (unlikely(sched_info_on()))
1656 memset(&p->sched_info, 0, sizeof(p->sched_info));
1658 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1661 #ifdef CONFIG_PREEMPT
1662 /* Want to start with kernel preemption disabled. */
1663 task_thread_info(p)->preempt_count = 1;
1666 * Share the timeslice between parent and child, thus the
1667 * total amount of pending timeslices in the system doesn't change,
1668 * resulting in more scheduling fairness.
1670 local_irq_disable();
1671 p->time_slice = (current->time_slice + 1) >> 1;
1673 * The remainder of the first timeslice might be recovered by
1674 * the parent if the child exits early enough.
1676 p->first_time_slice = 1;
1677 current->time_slice >>= 1;
1678 p->timestamp = sched_clock();
1679 if (unlikely(!current->time_slice)) {
1681 * This case is rare, it happens when the parent has only
1682 * a single jiffy left from its timeslice. Taking the
1683 * runqueue lock is not a problem.
1685 current->time_slice = 1;
1693 * wake_up_new_task - wake up a newly created task for the first time.
1695 * This function will do some initial scheduler statistics housekeeping
1696 * that must be done for every newly created context, then puts the task
1697 * on the runqueue and wakes it.
1699 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1701 struct rq *rq, *this_rq;
1702 unsigned long flags;
1705 rq = task_rq_lock(p, &flags);
1706 BUG_ON(p->state != TASK_RUNNING);
1707 this_cpu = smp_processor_id();
1711 * We decrease the sleep average of forking parents
1712 * and children as well, to keep max-interactive tasks
1713 * from forking tasks that are max-interactive. The parent
1714 * (current) is done further down, under its lock.
1716 p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1717 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1719 p->prio = effective_prio(p);
1721 vx_activate_task(p);
1722 if (likely(cpu == this_cpu)) {
1723 if (!(clone_flags & CLONE_VM)) {
1725 * The VM isn't cloned, so we're in a good position to
1726 * do child-runs-first in anticipation of an exec. This
1727 * usually avoids a lot of COW overhead.
1729 if (unlikely(!current->array))
1730 __activate_task(p, rq);
1732 p->prio = current->prio;
1733 BUG_ON(p->state & TASK_ONHOLD);
1734 p->normal_prio = current->normal_prio;
1735 list_add_tail(&p->run_list, ¤t->run_list);
1736 p->array = current->array;
1737 p->array->nr_active++;
1738 inc_nr_running(p, rq);
1742 /* Run child last */
1743 __activate_task(p, rq);
1745 * We skip the following code due to cpu == this_cpu
1747 * task_rq_unlock(rq, &flags);
1748 * this_rq = task_rq_lock(current, &flags);
1752 this_rq = cpu_rq(this_cpu);
1755 * Not the local CPU - must adjust timestamp. This should
1756 * get optimised away in the !CONFIG_SMP case.
1758 p->timestamp = (p->timestamp - this_rq->timestamp_last_tick)
1759 + rq->timestamp_last_tick;
1760 __activate_task(p, rq);
1761 if (TASK_PREEMPTS_CURR(p, rq))
1762 resched_task(rq->curr);
1765 * Parent and child are on different CPUs, now get the
1766 * parent runqueue to update the parent's ->sleep_avg:
1768 task_rq_unlock(rq, &flags);
1769 this_rq = task_rq_lock(current, &flags);
1771 current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1772 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1773 task_rq_unlock(this_rq, &flags);
1777 * Potentially available exiting-child timeslices are
1778 * retrieved here - this way the parent does not get
1779 * penalized for creating too many threads.
1781 * (this cannot be used to 'generate' timeslices
1782 * artificially, because any timeslice recovered here
1783 * was given away by the parent in the first place.)
1785 void fastcall sched_exit(struct task_struct *p)
1787 unsigned long flags;
1791 * If the child was a (relative-) CPU hog then decrease
1792 * the sleep_avg of the parent as well.
1794 rq = task_rq_lock(p->parent, &flags);
1795 if (p->first_time_slice && task_cpu(p) == task_cpu(p->parent)) {
1796 p->parent->time_slice += p->time_slice;
1797 if (unlikely(p->parent->time_slice > task_timeslice(p)))
1798 p->parent->time_slice = task_timeslice(p);
1800 if (p->sleep_avg < p->parent->sleep_avg)
1801 p->parent->sleep_avg = p->parent->sleep_avg /
1802 (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg /
1804 task_rq_unlock(rq, &flags);
1808 * prepare_task_switch - prepare to switch tasks
1809 * @rq: the runqueue preparing to switch
1810 * @next: the task we are going to switch to.
1812 * This is called with the rq lock held and interrupts off. It must
1813 * be paired with a subsequent finish_task_switch after the context
1816 * prepare_task_switch sets up locking and calls architecture specific
1819 static inline void prepare_task_switch(struct rq *rq, struct task_struct *next)
1821 prepare_lock_switch(rq, next);
1822 prepare_arch_switch(next);
1826 * finish_task_switch - clean up after a task-switch
1827 * @rq: runqueue associated with task-switch
1828 * @prev: the thread we just switched away from.
1830 * finish_task_switch must be called after the context switch, paired
1831 * with a prepare_task_switch call before the context switch.
1832 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1833 * and do any other architecture-specific cleanup actions.
1835 * Note that we may have delayed dropping an mm in context_switch(). If
1836 * so, we finish that here outside of the runqueue lock. (Doing it
1837 * with the lock held can cause deadlocks; see schedule() for
1840 static inline void finish_task_switch(struct rq *rq, struct task_struct *prev)
1841 __releases(rq->lock)
1843 struct mm_struct *mm = rq->prev_mm;
1844 unsigned long prev_task_flags;
1849 * A task struct has one reference for the use as "current".
1850 * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1851 * calls schedule one last time. The schedule call will never return,
1852 * and the scheduled task must drop that reference.
1853 * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1854 * still held, otherwise prev could be scheduled on another cpu, die
1855 * there before we look at prev->state, and then the reference would
1857 * Manfred Spraul <manfred@colorfullife.com>
1859 prev_task_flags = prev->flags;
1860 finish_arch_switch(prev);
1861 finish_lock_switch(rq, prev);
1864 if (unlikely(prev_task_flags & PF_DEAD)) {
1866 * Remove function-return probe instances associated with this
1867 * task and put them back on the free list.
1869 kprobe_flush_task(prev);
1870 put_task_struct(prev);
1875 * schedule_tail - first thing a freshly forked thread must call.
1876 * @prev: the thread we just switched away from.
1878 asmlinkage void schedule_tail(struct task_struct *prev)
1879 __releases(rq->lock)
1881 struct rq *rq = this_rq();
1883 finish_task_switch(rq, prev);
1884 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1885 /* In this case, finish_task_switch does not reenable preemption */
1888 if (current->set_child_tid)
1889 put_user(current->pid, current->set_child_tid);
1893 * context_switch - switch to the new MM and the new
1894 * thread's register state.
1896 static inline struct task_struct *
1897 context_switch(struct rq *rq, struct task_struct *prev,
1898 struct task_struct *next)
1900 struct mm_struct *mm = next->mm;
1901 struct mm_struct *oldmm = prev->active_mm;
1903 if (unlikely(!mm)) {
1904 next->active_mm = oldmm;
1905 atomic_inc(&oldmm->mm_count);
1906 enter_lazy_tlb(oldmm, next);
1908 switch_mm(oldmm, mm, next);
1910 if (unlikely(!prev->mm)) {
1911 prev->active_mm = NULL;
1912 WARN_ON(rq->prev_mm);
1913 rq->prev_mm = oldmm;
1916 * Since the runqueue lock will be released by the next
1917 * task (which is an invalid locking op but in the case
1918 * of the scheduler it's an obvious special-case), so we
1919 * do an early lockdep release here:
1921 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1922 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1925 /* Here we just switch the register state and the stack. */
1926 switch_to(prev, next, prev);
1932 * nr_running, nr_uninterruptible and nr_context_switches:
1934 * externally visible scheduler statistics: current number of runnable
1935 * threads, current number of uninterruptible-sleeping threads, total
1936 * number of context switches performed since bootup.
1938 unsigned long nr_running(void)
1940 unsigned long i, sum = 0;
1942 for_each_online_cpu(i)
1943 sum += cpu_rq(i)->nr_running;
1948 unsigned long nr_uninterruptible(void)
1950 unsigned long i, sum = 0;
1952 for_each_possible_cpu(i)
1953 sum += cpu_rq(i)->nr_uninterruptible;
1956 * Since we read the counters lockless, it might be slightly
1957 * inaccurate. Do not allow it to go below zero though:
1959 if (unlikely((long)sum < 0))
1965 unsigned long long nr_context_switches(void)
1968 unsigned long long sum = 0;
1970 for_each_possible_cpu(i)
1971 sum += cpu_rq(i)->nr_switches;
1976 unsigned long nr_iowait(void)
1978 unsigned long i, sum = 0;
1980 for_each_possible_cpu(i)
1981 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1986 unsigned long nr_active(void)
1988 unsigned long i, running = 0, uninterruptible = 0;
1990 for_each_online_cpu(i) {
1991 running += cpu_rq(i)->nr_running;
1992 uninterruptible += cpu_rq(i)->nr_uninterruptible;
1995 if (unlikely((long)uninterruptible < 0))
1996 uninterruptible = 0;
1998 return running + uninterruptible;
2004 * Is this task likely cache-hot:
2007 task_hot(struct task_struct *p, unsigned long long now, struct sched_domain *sd)
2009 return (long long)(now - p->last_ran) < (long long)sd->cache_hot_time;
2013 * double_rq_lock - safely lock two runqueues
2015 * Note this does not disable interrupts like task_rq_lock,
2016 * you need to do so manually before calling.
2018 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2019 __acquires(rq1->lock)
2020 __acquires(rq2->lock)
2023 spin_lock(&rq1->lock);
2024 __acquire(rq2->lock); /* Fake it out ;) */
2027 spin_lock(&rq1->lock);
2028 spin_lock(&rq2->lock);
2030 spin_lock(&rq2->lock);
2031 spin_lock(&rq1->lock);
2037 * double_rq_unlock - safely unlock two runqueues
2039 * Note this does not restore interrupts like task_rq_unlock,
2040 * you need to do so manually after calling.
2042 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2043 __releases(rq1->lock)
2044 __releases(rq2->lock)
2046 spin_unlock(&rq1->lock);
2048 spin_unlock(&rq2->lock);
2050 __release(rq2->lock);
2054 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2056 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2057 __releases(this_rq->lock)
2058 __acquires(busiest->lock)
2059 __acquires(this_rq->lock)
2061 if (unlikely(!spin_trylock(&busiest->lock))) {
2062 if (busiest < this_rq) {
2063 spin_unlock(&this_rq->lock);
2064 spin_lock(&busiest->lock);
2065 spin_lock(&this_rq->lock);
2067 spin_lock(&busiest->lock);
2072 * If dest_cpu is allowed for this process, migrate the task to it.
2073 * This is accomplished by forcing the cpu_allowed mask to only
2074 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2075 * the cpu_allowed mask is restored.
2077 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2079 struct migration_req req;
2080 unsigned long flags;
2083 rq = task_rq_lock(p, &flags);
2084 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2085 || unlikely(cpu_is_offline(dest_cpu)))
2088 /* force the process onto the specified CPU */
2089 if (migrate_task(p, dest_cpu, &req)) {
2090 /* Need to wait for migration thread (might exit: take ref). */
2091 struct task_struct *mt = rq->migration_thread;
2093 get_task_struct(mt);
2094 task_rq_unlock(rq, &flags);
2095 wake_up_process(mt);
2096 put_task_struct(mt);
2097 wait_for_completion(&req.done);
2102 task_rq_unlock(rq, &flags);
2106 * sched_exec - execve() is a valuable balancing opportunity, because at
2107 * this point the task has the smallest effective memory and cache footprint.
2109 void sched_exec(void)
2111 int new_cpu, this_cpu = get_cpu();
2112 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2114 if (new_cpu != this_cpu)
2115 sched_migrate_task(current, new_cpu);
2119 * pull_task - move a task from a remote runqueue to the local runqueue.
2120 * Both runqueues must be locked.
2122 static void pull_task(struct rq *src_rq, struct prio_array *src_array,
2123 struct task_struct *p, struct rq *this_rq,
2124 struct prio_array *this_array, int this_cpu)
2126 dequeue_task(p, src_array);
2127 dec_nr_running(p, src_rq);
2128 set_task_cpu(p, this_cpu);
2129 inc_nr_running(p, this_rq);
2130 enqueue_task(p, this_array);
2131 p->timestamp = (p->timestamp - src_rq->timestamp_last_tick)
2132 + this_rq->timestamp_last_tick;
2134 * Note that idle threads have a prio of MAX_PRIO, for this test
2135 * to be always true for them.
2137 if (TASK_PREEMPTS_CURR(p, this_rq))
2138 resched_task(this_rq->curr);
2142 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2145 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2146 struct sched_domain *sd, enum idle_type idle,
2150 * We do not migrate tasks that are:
2151 * 1) running (obviously), or
2152 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2153 * 3) are cache-hot on their current CPU.
2155 if (!cpu_isset(this_cpu, p->cpus_allowed))
2159 if (task_running(rq, p))
2163 * Aggressive migration if:
2164 * 1) task is cache cold, or
2165 * 2) too many balance attempts have failed.
2168 if (sd->nr_balance_failed > sd->cache_nice_tries)
2171 if (task_hot(p, rq->timestamp_last_tick, sd))
2176 #define rq_best_prio(rq) min((rq)->curr->prio, (rq)->best_expired_prio)
2179 * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted
2180 * load from busiest to this_rq, as part of a balancing operation within
2181 * "domain". Returns the number of tasks moved.
2183 * Called with both runqueues locked.
2185 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2186 unsigned long max_nr_move, unsigned long max_load_move,
2187 struct sched_domain *sd, enum idle_type idle,
2190 int idx, pulled = 0, pinned = 0, this_best_prio, best_prio,
2191 best_prio_seen, skip_for_load;
2192 struct prio_array *array, *dst_array;
2193 struct list_head *head, *curr;
2194 struct task_struct *tmp;
2197 if (max_nr_move == 0 || max_load_move == 0)
2200 rem_load_move = max_load_move;
2202 this_best_prio = rq_best_prio(this_rq);
2203 best_prio = rq_best_prio(busiest);
2205 * Enable handling of the case where there is more than one task
2206 * with the best priority. If the current running task is one
2207 * of those with prio==best_prio we know it won't be moved
2208 * and therefore it's safe to override the skip (based on load) of
2209 * any task we find with that prio.
2211 best_prio_seen = best_prio == busiest->curr->prio;
2214 * We first consider expired tasks. Those will likely not be
2215 * executed in the near future, and they are most likely to
2216 * be cache-cold, thus switching CPUs has the least effect
2219 if (busiest->expired->nr_active) {
2220 array = busiest->expired;
2221 dst_array = this_rq->expired;
2223 array = busiest->active;
2224 dst_array = this_rq->active;
2228 /* Start searching at priority 0: */
2232 idx = sched_find_first_bit(array->bitmap);
2234 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
2235 if (idx >= MAX_PRIO) {
2236 if (array == busiest->expired && busiest->active->nr_active) {
2237 array = busiest->active;
2238 dst_array = this_rq->active;
2244 head = array->queue + idx;
2247 tmp = list_entry(curr, struct task_struct, run_list);
2252 * To help distribute high priority tasks accross CPUs we don't
2253 * skip a task if it will be the highest priority task (i.e. smallest
2254 * prio value) on its new queue regardless of its load weight
2256 skip_for_load = tmp->load_weight > rem_load_move;
2257 if (skip_for_load && idx < this_best_prio)
2258 skip_for_load = !best_prio_seen && idx == best_prio;
2259 if (skip_for_load ||
2260 !can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
2262 best_prio_seen |= idx == best_prio;
2269 #ifdef CONFIG_SCHEDSTATS
2270 if (task_hot(tmp, busiest->timestamp_last_tick, sd))
2271 schedstat_inc(sd, lb_hot_gained[idle]);
2274 pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
2276 rem_load_move -= tmp->load_weight;
2279 * We only want to steal up to the prescribed number of tasks
2280 * and the prescribed amount of weighted load.
2282 if (pulled < max_nr_move && rem_load_move > 0) {
2283 if (idx < this_best_prio)
2284 this_best_prio = idx;
2292 * Right now, this is the only place pull_task() is called,
2293 * so we can safely collect pull_task() stats here rather than
2294 * inside pull_task().
2296 schedstat_add(sd, lb_gained[idle], pulled);
2299 *all_pinned = pinned;
2304 * find_busiest_group finds and returns the busiest CPU group within the
2305 * domain. It calculates and returns the amount of weighted load which
2306 * should be moved to restore balance via the imbalance parameter.
2308 static struct sched_group *
2309 find_busiest_group(struct sched_domain *sd, int this_cpu,
2310 unsigned long *imbalance, enum idle_type idle, int *sd_idle,
2313 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2314 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2315 unsigned long max_pull;
2316 unsigned long busiest_load_per_task, busiest_nr_running;
2317 unsigned long this_load_per_task, this_nr_running;
2319 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2320 int power_savings_balance = 1;
2321 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2322 unsigned long min_nr_running = ULONG_MAX;
2323 struct sched_group *group_min = NULL, *group_leader = NULL;
2326 max_load = this_load = total_load = total_pwr = 0;
2327 busiest_load_per_task = busiest_nr_running = 0;
2328 this_load_per_task = this_nr_running = 0;
2329 if (idle == NOT_IDLE)
2330 load_idx = sd->busy_idx;
2331 else if (idle == NEWLY_IDLE)
2332 load_idx = sd->newidle_idx;
2334 load_idx = sd->idle_idx;
2337 unsigned long load, group_capacity;
2340 unsigned long sum_nr_running, sum_weighted_load;
2342 local_group = cpu_isset(this_cpu, group->cpumask);
2344 /* Tally up the load of all CPUs in the group */
2345 sum_weighted_load = sum_nr_running = avg_load = 0;
2347 for_each_cpu_mask(i, group->cpumask) {
2350 if (!cpu_isset(i, *cpus))
2355 if (*sd_idle && !idle_cpu(i))
2358 /* Bias balancing toward cpus of our domain */
2360 load = target_load(i, load_idx);
2362 load = source_load(i, load_idx);
2365 sum_nr_running += rq->nr_running;
2366 sum_weighted_load += rq->raw_weighted_load;
2369 total_load += avg_load;
2370 total_pwr += group->cpu_power;
2372 /* Adjust by relative CPU power of the group */
2373 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
2375 group_capacity = group->cpu_power / SCHED_LOAD_SCALE;
2378 this_load = avg_load;
2380 this_nr_running = sum_nr_running;
2381 this_load_per_task = sum_weighted_load;
2382 } else if (avg_load > max_load &&
2383 sum_nr_running > group_capacity) {
2384 max_load = avg_load;
2386 busiest_nr_running = sum_nr_running;
2387 busiest_load_per_task = sum_weighted_load;
2390 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2392 * Busy processors will not participate in power savings
2395 if (idle == NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2399 * If the local group is idle or completely loaded
2400 * no need to do power savings balance at this domain
2402 if (local_group && (this_nr_running >= group_capacity ||
2404 power_savings_balance = 0;
2407 * If a group is already running at full capacity or idle,
2408 * don't include that group in power savings calculations
2410 if (!power_savings_balance || sum_nr_running >= group_capacity
2415 * Calculate the group which has the least non-idle load.
2416 * This is the group from where we need to pick up the load
2419 if ((sum_nr_running < min_nr_running) ||
2420 (sum_nr_running == min_nr_running &&
2421 first_cpu(group->cpumask) <
2422 first_cpu(group_min->cpumask))) {
2424 min_nr_running = sum_nr_running;
2425 min_load_per_task = sum_weighted_load /
2430 * Calculate the group which is almost near its
2431 * capacity but still has some space to pick up some load
2432 * from other group and save more power
2434 if (sum_nr_running <= group_capacity - 1) {
2435 if (sum_nr_running > leader_nr_running ||
2436 (sum_nr_running == leader_nr_running &&
2437 first_cpu(group->cpumask) >
2438 first_cpu(group_leader->cpumask))) {
2439 group_leader = group;
2440 leader_nr_running = sum_nr_running;
2445 group = group->next;
2446 } while (group != sd->groups);
2448 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2451 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2453 if (this_load >= avg_load ||
2454 100*max_load <= sd->imbalance_pct*this_load)
2457 busiest_load_per_task /= busiest_nr_running;
2459 * We're trying to get all the cpus to the average_load, so we don't
2460 * want to push ourselves above the average load, nor do we wish to
2461 * reduce the max loaded cpu below the average load, as either of these
2462 * actions would just result in more rebalancing later, and ping-pong
2463 * tasks around. Thus we look for the minimum possible imbalance.
2464 * Negative imbalances (*we* are more loaded than anyone else) will
2465 * be counted as no imbalance for these purposes -- we can't fix that
2466 * by pulling tasks to us. Be careful of negative numbers as they'll
2467 * appear as very large values with unsigned longs.
2469 if (max_load <= busiest_load_per_task)
2473 * In the presence of smp nice balancing, certain scenarios can have
2474 * max load less than avg load(as we skip the groups at or below
2475 * its cpu_power, while calculating max_load..)
2477 if (max_load < avg_load) {
2479 goto small_imbalance;
2482 /* Don't want to pull so many tasks that a group would go idle */
2483 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2485 /* How much load to actually move to equalise the imbalance */
2486 *imbalance = min(max_pull * busiest->cpu_power,
2487 (avg_load - this_load) * this->cpu_power)
2491 * if *imbalance is less than the average load per runnable task
2492 * there is no gaurantee that any tasks will be moved so we'll have
2493 * a think about bumping its value to force at least one task to be
2496 if (*imbalance < busiest_load_per_task) {
2497 unsigned long tmp, pwr_now, pwr_move;
2501 pwr_move = pwr_now = 0;
2503 if (this_nr_running) {
2504 this_load_per_task /= this_nr_running;
2505 if (busiest_load_per_task > this_load_per_task)
2508 this_load_per_task = SCHED_LOAD_SCALE;
2510 if (max_load - this_load >= busiest_load_per_task * imbn) {
2511 *imbalance = busiest_load_per_task;
2516 * OK, we don't have enough imbalance to justify moving tasks,
2517 * however we may be able to increase total CPU power used by
2521 pwr_now += busiest->cpu_power *
2522 min(busiest_load_per_task, max_load);
2523 pwr_now += this->cpu_power *
2524 min(this_load_per_task, this_load);
2525 pwr_now /= SCHED_LOAD_SCALE;
2527 /* Amount of load we'd subtract */
2528 tmp = busiest_load_per_task*SCHED_LOAD_SCALE/busiest->cpu_power;
2530 pwr_move += busiest->cpu_power *
2531 min(busiest_load_per_task, max_load - tmp);
2533 /* Amount of load we'd add */
2534 if (max_load*busiest->cpu_power <
2535 busiest_load_per_task*SCHED_LOAD_SCALE)
2536 tmp = max_load*busiest->cpu_power/this->cpu_power;
2538 tmp = busiest_load_per_task*SCHED_LOAD_SCALE/this->cpu_power;
2539 pwr_move += this->cpu_power*min(this_load_per_task, this_load + tmp);
2540 pwr_move /= SCHED_LOAD_SCALE;
2542 /* Move if we gain throughput */
2543 if (pwr_move <= pwr_now)
2546 *imbalance = busiest_load_per_task;
2552 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2553 if (idle == NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2556 if (this == group_leader && group_leader != group_min) {
2557 *imbalance = min_load_per_task;
2567 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2570 find_busiest_queue(struct sched_group *group, enum idle_type idle,
2571 unsigned long imbalance, cpumask_t *cpus)
2573 struct rq *busiest = NULL, *rq;
2574 unsigned long max_load = 0;
2577 for_each_cpu_mask(i, group->cpumask) {
2579 if (!cpu_isset(i, *cpus))
2584 if (rq->nr_running == 1 && rq->raw_weighted_load > imbalance)
2587 if (rq->raw_weighted_load > max_load) {
2588 max_load = rq->raw_weighted_load;
2597 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2598 * so long as it is large enough.
2600 #define MAX_PINNED_INTERVAL 512
2602 static inline unsigned long minus_1_or_zero(unsigned long n)
2604 return n > 0 ? n - 1 : 0;
2608 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2609 * tasks if there is an imbalance.
2611 * Called with this_rq unlocked.
2613 static int load_balance(int this_cpu, struct rq *this_rq,
2614 struct sched_domain *sd, enum idle_type idle)
2616 int nr_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2617 struct sched_group *group;
2618 unsigned long imbalance;
2620 cpumask_t cpus = CPU_MASK_ALL;
2622 if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2623 !sched_smt_power_savings)
2626 schedstat_inc(sd, lb_cnt[idle]);
2629 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2632 schedstat_inc(sd, lb_nobusyg[idle]);
2636 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2638 schedstat_inc(sd, lb_nobusyq[idle]);
2642 BUG_ON(busiest == this_rq);
2644 schedstat_add(sd, lb_imbalance[idle], imbalance);
2647 if (busiest->nr_running > 1) {
2649 * Attempt to move tasks. If find_busiest_group has found
2650 * an imbalance but busiest->nr_running <= 1, the group is
2651 * still unbalanced. nr_moved simply stays zero, so it is
2652 * correctly treated as an imbalance.
2654 double_rq_lock(this_rq, busiest);
2655 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2656 minus_1_or_zero(busiest->nr_running),
2657 imbalance, sd, idle, &all_pinned);
2658 double_rq_unlock(this_rq, busiest);
2660 /* All tasks on this runqueue were pinned by CPU affinity */
2661 if (unlikely(all_pinned)) {
2662 cpu_clear(cpu_of(busiest), cpus);
2663 if (!cpus_empty(cpus))
2670 schedstat_inc(sd, lb_failed[idle]);
2671 sd->nr_balance_failed++;
2673 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2675 spin_lock(&busiest->lock);
2677 /* don't kick the migration_thread, if the curr
2678 * task on busiest cpu can't be moved to this_cpu
2680 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2681 spin_unlock(&busiest->lock);
2683 goto out_one_pinned;
2686 if (!busiest->active_balance) {
2687 busiest->active_balance = 1;
2688 busiest->push_cpu = this_cpu;
2691 spin_unlock(&busiest->lock);
2693 wake_up_process(busiest->migration_thread);
2696 * We've kicked active balancing, reset the failure
2699 sd->nr_balance_failed = sd->cache_nice_tries+1;
2702 sd->nr_balance_failed = 0;
2704 if (likely(!active_balance)) {
2705 /* We were unbalanced, so reset the balancing interval */
2706 sd->balance_interval = sd->min_interval;
2709 * If we've begun active balancing, start to back off. This
2710 * case may not be covered by the all_pinned logic if there
2711 * is only 1 task on the busy runqueue (because we don't call
2714 if (sd->balance_interval < sd->max_interval)
2715 sd->balance_interval *= 2;
2718 if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2719 !sched_smt_power_savings)
2724 schedstat_inc(sd, lb_balanced[idle]);
2726 sd->nr_balance_failed = 0;
2729 /* tune up the balancing interval */
2730 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2731 (sd->balance_interval < sd->max_interval))
2732 sd->balance_interval *= 2;
2734 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2735 !sched_smt_power_savings)
2741 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2742 * tasks if there is an imbalance.
2744 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2745 * this_rq is locked.
2748 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2750 struct sched_group *group;
2751 struct rq *busiest = NULL;
2752 unsigned long imbalance;
2755 cpumask_t cpus = CPU_MASK_ALL;
2757 if (sd->flags & SD_SHARE_CPUPOWER && !sched_smt_power_savings)
2760 schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
2762 group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE,
2765 schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]);
2769 busiest = find_busiest_queue(group, NEWLY_IDLE, imbalance,
2772 schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]);
2776 BUG_ON(busiest == this_rq);
2778 schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
2781 if (busiest->nr_running > 1) {
2782 /* Attempt to move tasks */
2783 double_lock_balance(this_rq, busiest);
2784 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2785 minus_1_or_zero(busiest->nr_running),
2786 imbalance, sd, NEWLY_IDLE, NULL);
2787 spin_unlock(&busiest->lock);
2790 cpu_clear(cpu_of(busiest), cpus);
2791 if (!cpus_empty(cpus))
2797 schedstat_inc(sd, lb_failed[NEWLY_IDLE]);
2798 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2801 sd->nr_balance_failed = 0;
2806 schedstat_inc(sd, lb_balanced[NEWLY_IDLE]);
2807 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2808 !sched_smt_power_savings)
2810 sd->nr_balance_failed = 0;
2816 * idle_balance is called by schedule() if this_cpu is about to become
2817 * idle. Attempts to pull tasks from other CPUs.
2819 static void idle_balance(int this_cpu, struct rq *this_rq)
2821 struct sched_domain *sd;
2823 for_each_domain(this_cpu, sd) {
2824 if (sd->flags & SD_BALANCE_NEWIDLE) {
2825 /* If we've pulled tasks over stop searching: */
2826 if (load_balance_newidle(this_cpu, this_rq, sd))
2833 * active_load_balance is run by migration threads. It pushes running tasks
2834 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2835 * running on each physical CPU where possible, and avoids physical /
2836 * logical imbalances.
2838 * Called with busiest_rq locked.
2840 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2842 int target_cpu = busiest_rq->push_cpu;
2843 struct sched_domain *sd;
2844 struct rq *target_rq;
2846 /* Is there any task to move? */
2847 if (busiest_rq->nr_running <= 1)
2850 target_rq = cpu_rq(target_cpu);
2853 * This condition is "impossible", if it occurs
2854 * we need to fix it. Originally reported by
2855 * Bjorn Helgaas on a 128-cpu setup.
2857 BUG_ON(busiest_rq == target_rq);
2859 /* move a task from busiest_rq to target_rq */
2860 double_lock_balance(busiest_rq, target_rq);
2862 /* Search for an sd spanning us and the target CPU. */
2863 for_each_domain(target_cpu, sd) {
2864 if ((sd->flags & SD_LOAD_BALANCE) &&
2865 cpu_isset(busiest_cpu, sd->span))
2870 schedstat_inc(sd, alb_cnt);
2872 if (move_tasks(target_rq, target_cpu, busiest_rq, 1,
2873 RTPRIO_TO_LOAD_WEIGHT(100), sd, SCHED_IDLE,
2875 schedstat_inc(sd, alb_pushed);
2877 schedstat_inc(sd, alb_failed);
2879 spin_unlock(&target_rq->lock);
2883 * rebalance_tick will get called every timer tick, on every CPU.
2885 * It checks each scheduling domain to see if it is due to be balanced,
2886 * and initiates a balancing operation if so.
2888 * Balancing parameters are set up in arch_init_sched_domains.
2891 /* Don't have all balancing operations going off at once: */
2892 static inline unsigned long cpu_offset(int cpu)
2894 return jiffies + cpu * HZ / NR_CPUS;
2898 rebalance_tick(int this_cpu, struct rq *this_rq, enum idle_type idle)
2900 unsigned long this_load, interval, j = cpu_offset(this_cpu);
2901 struct sched_domain *sd;
2904 this_load = this_rq->raw_weighted_load;
2906 /* Update our load: */
2907 for (i = 0, scale = 1; i < 3; i++, scale <<= 1) {
2908 unsigned long old_load, new_load;
2910 old_load = this_rq->cpu_load[i];
2911 new_load = this_load;
2913 * Round up the averaging division if load is increasing. This
2914 * prevents us from getting stuck on 9 if the load is 10, for
2917 if (new_load > old_load)
2918 new_load += scale-1;
2919 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale;
2922 for_each_domain(this_cpu, sd) {
2923 if (!(sd->flags & SD_LOAD_BALANCE))
2926 interval = sd->balance_interval;
2927 if (idle != SCHED_IDLE)
2928 interval *= sd->busy_factor;
2930 /* scale ms to jiffies */
2931 interval = msecs_to_jiffies(interval);
2932 if (unlikely(!interval))
2935 if (j - sd->last_balance >= interval) {
2936 if (load_balance(this_cpu, this_rq, sd, idle)) {
2938 * We've pulled tasks over so either we're no
2939 * longer idle, or one of our SMT siblings is
2944 sd->last_balance += interval;
2950 * on UP we do not need to balance between CPUs:
2952 static inline void rebalance_tick(int cpu, struct rq *rq, enum idle_type idle)
2955 static inline void idle_balance(int cpu, struct rq *rq)
2960 static inline int wake_priority_sleeper(struct rq *rq)
2964 #ifdef CONFIG_SCHED_SMT
2965 spin_lock(&rq->lock);
2967 * If an SMT sibling task has been put to sleep for priority
2968 * reasons reschedule the idle task to see if it can now run.
2970 if (rq->nr_running) {
2971 resched_task(rq->idle);
2974 spin_unlock(&rq->lock);
2979 DEFINE_PER_CPU(struct kernel_stat, kstat);
2981 EXPORT_PER_CPU_SYMBOL(kstat);
2984 * This is called on clock ticks and on context switches.
2985 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2988 update_cpu_clock(struct task_struct *p, struct rq *rq, unsigned long long now)
2990 p->sched_time += now - max(p->timestamp, rq->timestamp_last_tick);
2994 * Return current->sched_time plus any more ns on the sched_clock
2995 * that have not yet been banked.
2997 unsigned long long current_sched_time(const struct task_struct *p)
2999 unsigned long long ns;
3000 unsigned long flags;
3002 local_irq_save(flags);
3003 ns = max(p->timestamp, task_rq(p)->timestamp_last_tick);
3004 ns = p->sched_time + sched_clock() - ns;
3005 local_irq_restore(flags);
3011 * We place interactive tasks back into the active array, if possible.
3013 * To guarantee that this does not starve expired tasks we ignore the
3014 * interactivity of a task if the first expired task had to wait more
3015 * than a 'reasonable' amount of time. This deadline timeout is
3016 * load-dependent, as the frequency of array switched decreases with
3017 * increasing number of running tasks. We also ignore the interactivity
3018 * if a better static_prio task has expired:
3020 static inline int expired_starving(struct rq *rq)
3022 if (rq->curr->static_prio > rq->best_expired_prio)
3024 if (!STARVATION_LIMIT || !rq->expired_timestamp)
3026 if (jiffies - rq->expired_timestamp > STARVATION_LIMIT * rq->nr_running)
3032 * Account user cpu time to a process.
3033 * @p: the process that the cpu time gets accounted to
3034 * @hardirq_offset: the offset to subtract from hardirq_count()
3035 * @cputime: the cpu time spent in user space since the last update
3037 void account_user_time(struct task_struct *p, cputime_t cputime)
3039 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3040 struct vx_info *vxi = p->vx_info; /* p is _always_ current */
3042 int nice = (TASK_NICE(p) > 0);
3044 p->utime = cputime_add(p->utime, cputime);
3045 vx_account_user(vxi, cputime, nice);
3047 /* Add user time to cpustat. */
3048 tmp = cputime_to_cputime64(cputime);
3050 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3052 cpustat->user = cputime64_add(cpustat->user, tmp);
3056 * Account system cpu time to a process.
3057 * @p: the process that the cpu time gets accounted to
3058 * @hardirq_offset: the offset to subtract from hardirq_count()
3059 * @cputime: the cpu time spent in kernel space since the last update
3061 void account_system_time(struct task_struct *p, int hardirq_offset,
3064 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3065 struct vx_info *vxi = p->vx_info; /* p is _always_ current */
3066 struct rq *rq = this_rq();
3069 p->stime = cputime_add(p->stime, cputime);
3070 vx_account_system(vxi, cputime, (p == rq->idle));
3072 /* Add system time to cpustat. */
3073 tmp = cputime_to_cputime64(cputime);
3074 if (hardirq_count() - hardirq_offset)
3075 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3076 else if (softirq_count())
3077 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3078 else if (p != rq->idle)
3079 cpustat->system = cputime64_add(cpustat->system, tmp);
3080 else if (atomic_read(&rq->nr_iowait) > 0)
3081 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3083 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3084 /* Account for system time used */
3085 acct_update_integrals(p);
3089 * Account for involuntary wait time.
3090 * @p: the process from which the cpu time has been stolen
3091 * @steal: the cpu time spent in involuntary wait
3093 void account_steal_time(struct task_struct *p, cputime_t steal)
3095 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3096 cputime64_t tmp = cputime_to_cputime64(steal);
3097 struct rq *rq = this_rq();
3099 if (p == rq->idle) {
3100 p->stime = cputime_add(p->stime, steal);
3101 if (atomic_read(&rq->nr_iowait) > 0)
3102 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3104 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3106 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3110 * This function gets called by the timer code, with HZ frequency.
3111 * We call it with interrupts disabled.
3113 * It also gets called by the fork code, when changing the parent's
3116 void scheduler_tick(void)
3118 unsigned long long now = sched_clock();
3119 struct task_struct *p = current;
3120 int cpu = smp_processor_id();
3121 struct rq *rq = cpu_rq(cpu);
3123 update_cpu_clock(p, rq, now);
3125 rq->timestamp_last_tick = now;
3127 #if defined(CONFIG_VSERVER_HARDCPU) && defined(CONFIG_VSERVER_ACB_SCHED)
3128 vx_scheduler_tick();
3131 if (p == rq->idle) {
3132 if (wake_priority_sleeper(rq))
3134 #ifdef CONFIG_VSERVER_HARDCPU_IDLE
3135 if (!--rq->idle_tokens && !list_empty(&rq->hold_queue))
3138 rebalance_tick(cpu, rq, SCHED_IDLE);
3142 /* Task might have expired already, but not scheduled off yet */
3143 if (p->array != rq->active) {
3144 set_tsk_need_resched(p);
3147 spin_lock(&rq->lock);
3149 * The task was running during this tick - update the
3150 * time slice counter. Note: we do not update a thread's
3151 * priority until it either goes to sleep or uses up its
3152 * timeslice. This makes it possible for interactive tasks
3153 * to use up their timeslices at their highest priority levels.
3157 * RR tasks need a special form of timeslice management.
3158 * FIFO tasks have no timeslices.
3160 if ((p->policy == SCHED_RR) && vx_need_resched(p)) {
3161 p->time_slice = task_timeslice(p);
3162 p->first_time_slice = 0;
3163 set_tsk_need_resched(p);
3165 /* put it at the end of the queue: */
3166 requeue_task(p, rq->active);
3170 if (vx_need_resched(p)) {
3171 dequeue_task(p, rq->active);
3172 set_tsk_need_resched(p);
3173 p->prio = effective_prio(p);
3174 p->time_slice = task_timeslice(p);
3175 p->first_time_slice = 0;
3177 if (!rq->expired_timestamp)
3178 rq->expired_timestamp = jiffies;
3179 if (!TASK_INTERACTIVE(p) || expired_starving(rq)) {
3180 enqueue_task(p, rq->expired);
3181 if (p->static_prio < rq->best_expired_prio)
3182 rq->best_expired_prio = p->static_prio;
3184 enqueue_task(p, rq->active);
3187 * Prevent a too long timeslice allowing a task to monopolize
3188 * the CPU. We do this by splitting up the timeslice into
3191 * Note: this does not mean the task's timeslices expire or
3192 * get lost in any way, they just might be preempted by
3193 * another task of equal priority. (one with higher
3194 * priority would have preempted this task already.) We
3195 * requeue this task to the end of the list on this priority
3196 * level, which is in essence a round-robin of tasks with
3199 * This only applies to tasks in the interactive
3200 * delta range with at least TIMESLICE_GRANULARITY to requeue.
3202 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
3203 p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
3204 (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
3205 (p->array == rq->active)) {
3207 requeue_task(p, rq->active);
3208 set_tsk_need_resched(p);
3212 spin_unlock(&rq->lock);
3214 rebalance_tick(cpu, rq, NOT_IDLE);
3217 #ifdef CONFIG_SCHED_SMT
3218 static inline void wakeup_busy_runqueue(struct rq *rq)
3220 /* If an SMT runqueue is sleeping due to priority reasons wake it up */
3221 if (rq->curr == rq->idle && rq->nr_running)
3222 resched_task(rq->idle);
3226 * Called with interrupt disabled and this_rq's runqueue locked.
3228 static void wake_sleeping_dependent(int this_cpu)
3230 struct sched_domain *tmp, *sd = NULL;
3233 for_each_domain(this_cpu, tmp) {
3234 if (tmp->flags & SD_SHARE_CPUPOWER) {
3243 for_each_cpu_mask(i, sd->span) {
3244 struct rq *smt_rq = cpu_rq(i);
3248 if (unlikely(!spin_trylock(&smt_rq->lock)))
3251 wakeup_busy_runqueue(smt_rq);
3252 spin_unlock(&smt_rq->lock);
3257 * number of 'lost' timeslices this task wont be able to fully
3258 * utilize, if another task runs on a sibling. This models the
3259 * slowdown effect of other tasks running on siblings:
3261 static inline unsigned long
3262 smt_slice(struct task_struct *p, struct sched_domain *sd)
3264 return p->time_slice * (100 - sd->per_cpu_gain) / 100;
3268 * To minimise lock contention and not have to drop this_rq's runlock we only
3269 * trylock the sibling runqueues and bypass those runqueues if we fail to
3270 * acquire their lock. As we only trylock the normal locking order does not
3271 * need to be obeyed.
3274 dependent_sleeper(int this_cpu, struct rq *this_rq, struct task_struct *p)
3276 struct sched_domain *tmp, *sd = NULL;
3279 /* kernel/rt threads do not participate in dependent sleeping */
3280 if (!p->mm || rt_task(p))
3283 for_each_domain(this_cpu, tmp) {
3284 if (tmp->flags & SD_SHARE_CPUPOWER) {
3293 for_each_cpu_mask(i, sd->span) {
3294 struct task_struct *smt_curr;
3301 if (unlikely(!spin_trylock(&smt_rq->lock)))
3304 smt_curr = smt_rq->curr;
3310 * If a user task with lower static priority than the
3311 * running task on the SMT sibling is trying to schedule,
3312 * delay it till there is proportionately less timeslice
3313 * left of the sibling task to prevent a lower priority
3314 * task from using an unfair proportion of the
3315 * physical cpu's resources. -ck
3317 if (rt_task(smt_curr)) {
3319 * With real time tasks we run non-rt tasks only
3320 * per_cpu_gain% of the time.
3322 if ((jiffies % DEF_TIMESLICE) >
3323 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
3326 if (smt_curr->static_prio < p->static_prio &&
3327 !TASK_PREEMPTS_CURR(p, smt_rq) &&
3328 smt_slice(smt_curr, sd) > task_timeslice(p))
3332 spin_unlock(&smt_rq->lock);
3337 static inline void wake_sleeping_dependent(int this_cpu)
3341 dependent_sleeper(int this_cpu, struct rq *this_rq, struct task_struct *p)
3347 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3349 void fastcall add_preempt_count(int val)
3354 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3356 preempt_count() += val;
3358 * Spinlock count overflowing soon?
3360 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10);
3362 EXPORT_SYMBOL(add_preempt_count);
3364 void fastcall sub_preempt_count(int val)
3369 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3372 * Is the spinlock portion underflowing?
3374 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3375 !(preempt_count() & PREEMPT_MASK)))
3378 preempt_count() -= val;
3380 EXPORT_SYMBOL(sub_preempt_count);
3384 static inline int interactive_sleep(enum sleep_type sleep_type)
3386 return (sleep_type == SLEEP_INTERACTIVE ||
3387 sleep_type == SLEEP_INTERRUPTED);
3391 * schedule() is the main scheduler function.
3393 asmlinkage void __sched schedule(void)
3395 struct task_struct *prev, *next;
3396 struct prio_array *array;
3397 struct list_head *queue;
3398 unsigned long long now;
3399 unsigned long run_time;
3400 int cpu, idx, new_prio;
3403 struct vx_info *vxi;
3404 #ifdef CONFIG_VSERVER_HARDCPU
3406 # ifdef CONFIG_VSERVER_ACB_SCHED
3407 int min_guarantee_ticks = VX_INVALID_TICKS;
3408 int min_best_effort_ticks = VX_INVALID_TICKS;
3413 * Test if we are atomic. Since do_exit() needs to call into
3414 * schedule() atomically, we ignore that path for now.
3415 * Otherwise, whine if we are scheduling when we should not be.
3417 if (unlikely(in_atomic() && !current->exit_state)) {
3418 printk(KERN_ERR "BUG: scheduling while atomic: "
3420 current->comm, preempt_count(), current->pid);
3423 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3428 release_kernel_lock(prev);
3429 need_resched_nonpreemptible:
3433 * The idle thread is not allowed to schedule!
3434 * Remove this check after it has been exercised a bit.
3436 if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
3437 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
3441 schedstat_inc(rq, sched_cnt);
3442 now = sched_clock();
3443 if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
3444 run_time = now - prev->timestamp;
3445 if (unlikely((long long)(now - prev->timestamp) < 0))
3448 run_time = NS_MAX_SLEEP_AVG;
3451 * Tasks charged proportionately less run_time at high sleep_avg to
3452 * delay them losing their interactive status
3454 run_time /= (CURRENT_BONUS(prev) ? : 1);
3456 spin_lock_irq(&rq->lock);
3458 if (unlikely(prev->flags & PF_DEAD))
3459 prev->state = EXIT_DEAD;
3461 switch_count = &prev->nivcsw;
3462 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3463 switch_count = &prev->nvcsw;
3464 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3465 unlikely(signal_pending(prev))))
3466 prev->state = TASK_RUNNING;
3468 if (prev->state == TASK_UNINTERRUPTIBLE) {
3469 rq->nr_uninterruptible++;
3470 vx_uninterruptible_inc(prev);
3472 deactivate_task(prev, rq);
3476 #ifdef CONFIG_VSERVER_HARDCPU
3477 # ifdef CONFIG_VSERVER_ACB_SCHED
3480 min_guarantee_ticks = VX_INVALID_TICKS;
3481 min_best_effort_ticks = VX_INVALID_TICKS;
3484 if (!list_empty(&rq->hold_queue)) {
3485 struct list_head *l, *n;
3489 list_for_each_safe(l, n, &rq->hold_queue) {
3490 next = list_entry(l, struct task_struct, run_list);
3491 if (vxi == next->vx_info)
3494 vxi = next->vx_info;
3495 ret = vx_tokens_recalc(vxi);
3498 vx_unhold_task(vxi, next, rq);
3501 if ((ret < 0) && (maxidle < ret))
3503 # ifdef CONFIG_VSERVER_ACB_SCHED
3505 if (IS_BEST_EFFORT(vxi)) {
3506 if (min_best_effort_ticks < ret)
3507 min_best_effort_ticks = ret;
3509 if (min_guarantee_ticks < ret)
3510 min_guarantee_ticks = ret;
3516 rq->idle_tokens = -maxidle;
3521 cpu = smp_processor_id();
3522 if (unlikely(!rq->nr_running)) {
3523 idle_balance(cpu, rq);
3524 if (!rq->nr_running) {
3526 rq->expired_timestamp = 0;
3527 wake_sleeping_dependent(cpu);
3533 if (unlikely(!array->nr_active)) {
3535 * Switch the active and expired arrays.
3537 schedstat_inc(rq, sched_switch);
3538 rq->active = rq->expired;
3539 rq->expired = array;
3541 rq->expired_timestamp = 0;
3542 rq->best_expired_prio = MAX_PRIO;
3545 idx = sched_find_first_bit(array->bitmap);
3546 queue = array->queue + idx;
3547 next = list_entry(queue->next, struct task_struct, run_list);
3549 vxi = next->vx_info;
3550 #ifdef CONFIG_VSERVER_HARDCPU
3551 if (vx_info_flags(vxi, VXF_SCHED_PAUSE|VXF_SCHED_HARD, 0)) {
3552 int ret = vx_tokens_recalc(vxi);
3554 if (unlikely(ret <= 0)) {
3556 if ((rq->idle_tokens > -ret))
3557 rq->idle_tokens = -ret;
3558 # ifdef CONFIG_VSERVER_ACB_SCHED
3559 if (IS_BEST_EFFORT(vxi)) {
3560 if (min_best_effort_ticks < ret)
3561 min_best_effort_ticks = ret;
3563 if (min_guarantee_ticks < ret)
3564 min_guarantee_ticks = ret;
3568 vx_hold_task(vxi, next, rq);
3571 } else /* well, looks ugly but not as ugly as the ifdef-ed version */
3573 if (vx_info_flags(vxi, VXF_SCHED_PRIO, 0))
3574 vx_tokens_recalc(vxi);
3576 if (!rt_task(next) && interactive_sleep(next->sleep_type)) {
3577 unsigned long long delta = now - next->timestamp;
3578 if (unlikely((long long)(now - next->timestamp) < 0))
3581 if (next->sleep_type == SLEEP_INTERACTIVE)
3582 delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
3584 array = next->array;
3585 new_prio = recalc_task_prio(next, next->timestamp + delta);
3587 if (unlikely(next->prio != new_prio)) {
3588 dequeue_task(next, array);
3589 next->prio = new_prio;
3590 enqueue_task(next, array);
3593 next->sleep_type = SLEEP_NORMAL;
3594 if (dependent_sleeper(cpu, rq, next))
3597 #if defined(CONFIG_VSERVER_HARDCPU) && defined(CONFIG_VSERVER_ACB_SCHED)
3598 if (next == rq->idle && !list_empty(&rq->hold_queue)) {
3599 if (min_best_effort_ticks != VX_INVALID_TICKS) {
3600 vx_advance_best_effort_ticks(-min_best_effort_ticks);
3601 goto drain_hold_queue;
3603 if (min_guarantee_ticks != VX_INVALID_TICKS) {
3604 vx_advance_guaranteed_ticks(-min_guarantee_ticks);
3605 goto drain_hold_queue;
3609 if (next == rq->idle)
3610 schedstat_inc(rq, sched_goidle);
3612 prefetch_stack(next);
3613 clear_tsk_need_resched(prev);
3614 rcu_qsctr_inc(task_cpu(prev));
3616 update_cpu_clock(prev, rq, now);
3618 prev->sleep_avg -= run_time;
3619 if ((long)prev->sleep_avg <= 0)
3620 prev->sleep_avg = 0;
3621 prev->timestamp = prev->last_ran = now;
3623 sched_info_switch(prev, next);
3624 if (likely(prev != next)) {
3625 next->timestamp = now;
3630 prepare_task_switch(rq, next);
3631 prev = context_switch(rq, prev, next);
3634 * this_rq must be evaluated again because prev may have moved
3635 * CPUs since it called schedule(), thus the 'rq' on its stack
3636 * frame will be invalid.
3638 finish_task_switch(this_rq(), prev);
3640 spin_unlock_irq(&rq->lock);
3643 if (unlikely(reacquire_kernel_lock(prev) < 0))
3644 goto need_resched_nonpreemptible;
3645 preempt_enable_no_resched();
3646 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3649 EXPORT_SYMBOL(schedule);
3651 #ifdef CONFIG_PREEMPT
3653 * this is the entry point to schedule() from in-kernel preemption
3654 * off of preempt_enable. Kernel preemptions off return from interrupt
3655 * occur there and call schedule directly.
3657 asmlinkage void __sched preempt_schedule(void)
3659 struct thread_info *ti = current_thread_info();
3660 #ifdef CONFIG_PREEMPT_BKL
3661 struct task_struct *task = current;
3662 int saved_lock_depth;
3665 * If there is a non-zero preempt_count or interrupts are disabled,
3666 * we do not want to preempt the current task. Just return..
3668 if (unlikely(ti->preempt_count || irqs_disabled()))
3672 add_preempt_count(PREEMPT_ACTIVE);
3674 * We keep the big kernel semaphore locked, but we
3675 * clear ->lock_depth so that schedule() doesnt
3676 * auto-release the semaphore:
3678 #ifdef CONFIG_PREEMPT_BKL
3679 saved_lock_depth = task->lock_depth;
3680 task->lock_depth = -1;
3683 #ifdef CONFIG_PREEMPT_BKL
3684 task->lock_depth = saved_lock_depth;
3686 sub_preempt_count(PREEMPT_ACTIVE);
3688 /* we could miss a preemption opportunity between schedule and now */
3690 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3693 EXPORT_SYMBOL(preempt_schedule);
3696 * this is the entry point to schedule() from kernel preemption
3697 * off of irq context.
3698 * Note, that this is called and return with irqs disabled. This will
3699 * protect us against recursive calling from irq.
3701 asmlinkage void __sched preempt_schedule_irq(void)
3703 struct thread_info *ti = current_thread_info();
3704 #ifdef CONFIG_PREEMPT_BKL
3705 struct task_struct *task = current;
3706 int saved_lock_depth;
3708 /* Catch callers which need to be fixed */
3709 BUG_ON(ti->preempt_count || !irqs_disabled());
3712 add_preempt_count(PREEMPT_ACTIVE);
3714 * We keep the big kernel semaphore locked, but we
3715 * clear ->lock_depth so that schedule() doesnt
3716 * auto-release the semaphore:
3718 #ifdef CONFIG_PREEMPT_BKL
3719 saved_lock_depth = task->lock_depth;
3720 task->lock_depth = -1;
3724 local_irq_disable();
3725 #ifdef CONFIG_PREEMPT_BKL
3726 task->lock_depth = saved_lock_depth;
3728 sub_preempt_count(PREEMPT_ACTIVE);
3730 /* we could miss a preemption opportunity between schedule and now */
3732 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3736 #endif /* CONFIG_PREEMPT */
3738 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3741 return try_to_wake_up(curr->private, mode, sync);
3743 EXPORT_SYMBOL(default_wake_function);
3746 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3747 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3748 * number) then we wake all the non-exclusive tasks and one exclusive task.
3750 * There are circumstances in which we can try to wake a task which has already
3751 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3752 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3754 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3755 int nr_exclusive, int sync, void *key)
3757 struct list_head *tmp, *next;
3759 list_for_each_safe(tmp, next, &q->task_list) {
3760 wait_queue_t *curr = list_entry(tmp, wait_queue_t, task_list);
3761 unsigned flags = curr->flags;
3763 if (curr->func(curr, mode, sync, key) &&
3764 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3770 * __wake_up - wake up threads blocked on a waitqueue.
3772 * @mode: which threads
3773 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3774 * @key: is directly passed to the wakeup function
3776 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3777 int nr_exclusive, void *key)
3779 unsigned long flags;
3781 spin_lock_irqsave(&q->lock, flags);
3782 __wake_up_common(q, mode, nr_exclusive, 0, key);
3783 spin_unlock_irqrestore(&q->lock, flags);
3785 EXPORT_SYMBOL(__wake_up);
3788 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3790 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3792 __wake_up_common(q, mode, 1, 0, NULL);
3796 * __wake_up_sync - wake up threads blocked on a waitqueue.
3798 * @mode: which threads
3799 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3801 * The sync wakeup differs that the waker knows that it will schedule
3802 * away soon, so while the target thread will be woken up, it will not
3803 * be migrated to another CPU - ie. the two threads are 'synchronized'
3804 * with each other. This can prevent needless bouncing between CPUs.
3806 * On UP it can prevent extra preemption.
3809 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3811 unsigned long flags;
3817 if (unlikely(!nr_exclusive))
3820 spin_lock_irqsave(&q->lock, flags);
3821 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3822 spin_unlock_irqrestore(&q->lock, flags);
3824 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3826 void fastcall complete(struct completion *x)
3828 unsigned long flags;
3830 spin_lock_irqsave(&x->wait.lock, flags);
3832 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3834 spin_unlock_irqrestore(&x->wait.lock, flags);
3836 EXPORT_SYMBOL(complete);
3838 void fastcall complete_all(struct completion *x)
3840 unsigned long flags;
3842 spin_lock_irqsave(&x->wait.lock, flags);
3843 x->done += UINT_MAX/2;
3844 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3846 spin_unlock_irqrestore(&x->wait.lock, flags);
3848 EXPORT_SYMBOL(complete_all);
3850 void fastcall __sched wait_for_completion(struct completion *x)
3854 spin_lock_irq(&x->wait.lock);
3856 DECLARE_WAITQUEUE(wait, current);
3858 wait.flags |= WQ_FLAG_EXCLUSIVE;
3859 __add_wait_queue_tail(&x->wait, &wait);
3861 __set_current_state(TASK_UNINTERRUPTIBLE);
3862 spin_unlock_irq(&x->wait.lock);
3864 spin_lock_irq(&x->wait.lock);
3866 __remove_wait_queue(&x->wait, &wait);
3869 spin_unlock_irq(&x->wait.lock);
3871 EXPORT_SYMBOL(wait_for_completion);
3873 unsigned long fastcall __sched
3874 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3878 spin_lock_irq(&x->wait.lock);
3880 DECLARE_WAITQUEUE(wait, current);
3882 wait.flags |= WQ_FLAG_EXCLUSIVE;
3883 __add_wait_queue_tail(&x->wait, &wait);
3885 __set_current_state(TASK_UNINTERRUPTIBLE);
3886 spin_unlock_irq(&x->wait.lock);
3887 timeout = schedule_timeout(timeout);
3888 spin_lock_irq(&x->wait.lock);
3890 __remove_wait_queue(&x->wait, &wait);
3894 __remove_wait_queue(&x->wait, &wait);
3898 spin_unlock_irq(&x->wait.lock);
3901 EXPORT_SYMBOL(wait_for_completion_timeout);
3903 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3909 spin_lock_irq(&x->wait.lock);
3911 DECLARE_WAITQUEUE(wait, current);
3913 wait.flags |= WQ_FLAG_EXCLUSIVE;
3914 __add_wait_queue_tail(&x->wait, &wait);
3916 if (signal_pending(current)) {
3918 __remove_wait_queue(&x->wait, &wait);
3921 __set_current_state(TASK_INTERRUPTIBLE);
3922 spin_unlock_irq(&x->wait.lock);
3924 spin_lock_irq(&x->wait.lock);
3926 __remove_wait_queue(&x->wait, &wait);
3930 spin_unlock_irq(&x->wait.lock);
3934 EXPORT_SYMBOL(wait_for_completion_interruptible);
3936 unsigned long fastcall __sched
3937 wait_for_completion_interruptible_timeout(struct completion *x,
3938 unsigned long timeout)
3942 spin_lock_irq(&x->wait.lock);
3944 DECLARE_WAITQUEUE(wait, current);
3946 wait.flags |= WQ_FLAG_EXCLUSIVE;
3947 __add_wait_queue_tail(&x->wait, &wait);
3949 if (signal_pending(current)) {
3950 timeout = -ERESTARTSYS;
3951 __remove_wait_queue(&x->wait, &wait);
3954 __set_current_state(TASK_INTERRUPTIBLE);
3955 spin_unlock_irq(&x->wait.lock);
3956 timeout = schedule_timeout(timeout);
3957 spin_lock_irq(&x->wait.lock);
3959 __remove_wait_queue(&x->wait, &wait);
3963 __remove_wait_queue(&x->wait, &wait);
3967 spin_unlock_irq(&x->wait.lock);
3970 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3973 #define SLEEP_ON_VAR \
3974 unsigned long flags; \
3975 wait_queue_t wait; \
3976 init_waitqueue_entry(&wait, current);
3978 #define SLEEP_ON_HEAD \
3979 spin_lock_irqsave(&q->lock,flags); \
3980 __add_wait_queue(q, &wait); \
3981 spin_unlock(&q->lock);
3983 #define SLEEP_ON_TAIL \
3984 spin_lock_irq(&q->lock); \
3985 __remove_wait_queue(q, &wait); \
3986 spin_unlock_irqrestore(&q->lock, flags);
3988 #define SLEEP_ON_BKLCHECK \
3989 if (unlikely(!kernel_locked()) && \
3990 sleep_on_bkl_warnings < 10) { \
3991 sleep_on_bkl_warnings++; \
3995 static int sleep_on_bkl_warnings;
3997 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
4003 current->state = TASK_INTERRUPTIBLE;
4009 EXPORT_SYMBOL(interruptible_sleep_on);
4011 long fastcall __sched
4012 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4018 current->state = TASK_INTERRUPTIBLE;
4021 timeout = schedule_timeout(timeout);
4026 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4028 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4034 current->state = TASK_UNINTERRUPTIBLE;
4037 timeout = schedule_timeout(timeout);
4043 EXPORT_SYMBOL(sleep_on_timeout);
4045 #ifdef CONFIG_RT_MUTEXES
4048 * rt_mutex_setprio - set the current priority of a task
4050 * @prio: prio value (kernel-internal form)
4052 * This function changes the 'effective' priority of a task. It does
4053 * not touch ->normal_prio like __setscheduler().
4055 * Used by the rt_mutex code to implement priority inheritance logic.
4057 void rt_mutex_setprio(struct task_struct *p, int prio)
4059 struct prio_array *array;
4060 unsigned long flags;
4064 BUG_ON(prio < 0 || prio > MAX_PRIO);
4066 rq = task_rq_lock(p, &flags);
4071 dequeue_task(p, array);
4076 * If changing to an RT priority then queue it
4077 * in the active array!
4081 enqueue_task(p, array);
4083 * Reschedule if we are currently running on this runqueue and
4084 * our priority decreased, or if we are not currently running on
4085 * this runqueue and our priority is higher than the current's
4087 if (task_running(rq, p)) {
4088 if (p->prio > oldprio)
4089 resched_task(rq->curr);
4090 } else if (TASK_PREEMPTS_CURR(p, rq))
4091 resched_task(rq->curr);
4093 task_rq_unlock(rq, &flags);
4098 void set_user_nice(struct task_struct *p, long nice)
4100 struct prio_array *array;
4101 int old_prio, delta;
4102 unsigned long flags;
4105 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4108 * We have to be careful, if called from sys_setpriority(),
4109 * the task might be in the middle of scheduling on another CPU.
4111 rq = task_rq_lock(p, &flags);
4113 * The RT priorities are set via sched_setscheduler(), but we still
4114 * allow the 'normal' nice value to be set - but as expected
4115 * it wont have any effect on scheduling until the task is
4116 * not SCHED_NORMAL/SCHED_BATCH:
4118 if (has_rt_policy(p)) {
4119 p->static_prio = NICE_TO_PRIO(nice);
4124 dequeue_task(p, array);
4125 dec_raw_weighted_load(rq, p);
4128 p->static_prio = NICE_TO_PRIO(nice);
4131 p->prio = effective_prio(p);
4132 delta = p->prio - old_prio;
4135 enqueue_task(p, array);
4136 inc_raw_weighted_load(rq, p);
4138 * If the task increased its priority or is running and
4139 * lowered its priority, then reschedule its CPU:
4141 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4142 resched_task(rq->curr);
4145 task_rq_unlock(rq, &flags);
4147 EXPORT_SYMBOL(set_user_nice);
4150 * can_nice - check if a task can reduce its nice value
4154 int can_nice(const struct task_struct *p, const int nice)
4156 /* convert nice value [19,-20] to rlimit style value [1,40] */
4157 int nice_rlim = 20 - nice;
4159 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4160 capable(CAP_SYS_NICE));
4163 #ifdef __ARCH_WANT_SYS_NICE
4166 * sys_nice - change the priority of the current process.
4167 * @increment: priority increment
4169 * sys_setpriority is a more generic, but much slower function that
4170 * does similar things.
4172 asmlinkage long sys_nice(int increment)
4177 * Setpriority might change our priority at the same moment.
4178 * We don't have to worry. Conceptually one call occurs first
4179 * and we have a single winner.
4181 if (increment < -40)
4186 nice = PRIO_TO_NICE(current->static_prio) + increment;
4192 if (increment < 0 && !can_nice(current, nice))
4193 return vx_flags(VXF_IGNEG_NICE, 0) ? 0 : -EPERM;
4195 retval = security_task_setnice(current, nice);
4199 set_user_nice(current, nice);
4206 * task_prio - return the priority value of a given task.
4207 * @p: the task in question.
4209 * This is the priority value as seen by users in /proc.
4210 * RT tasks are offset by -200. Normal tasks are centered
4211 * around 0, value goes from -16 to +15.
4213 int task_prio(const struct task_struct *p)
4215 return p->prio - MAX_RT_PRIO;
4219 * task_nice - return the nice value of a given task.
4220 * @p: the task in question.
4222 int task_nice(const struct task_struct *p)
4224 return TASK_NICE(p);
4226 EXPORT_SYMBOL_GPL(task_nice);
4229 * idle_cpu - is a given cpu idle currently?
4230 * @cpu: the processor in question.
4232 int idle_cpu(int cpu)
4234 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4238 * idle_task - return the idle task for a given cpu.
4239 * @cpu: the processor in question.
4241 struct task_struct *idle_task(int cpu)
4243 return cpu_rq(cpu)->idle;
4247 * find_process_by_pid - find a process with a matching PID value.
4248 * @pid: the pid in question.
4250 static inline struct task_struct *find_process_by_pid(pid_t pid)
4252 return pid ? find_task_by_pid(pid) : current;
4255 /* Actually do priority change: must hold rq lock. */
4256 static void __setscheduler(struct task_struct *p, int policy, int prio)
4261 p->rt_priority = prio;
4262 p->normal_prio = normal_prio(p);
4263 /* we are holding p->pi_lock already */
4264 p->prio = rt_mutex_getprio(p);
4266 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
4268 if (policy == SCHED_BATCH)
4274 * sched_setscheduler - change the scheduling policy and/or RT priority of
4276 * @p: the task in question.
4277 * @policy: new policy.
4278 * @param: structure containing the new RT priority.
4280 int sched_setscheduler(struct task_struct *p, int policy,
4281 struct sched_param *param)
4283 int retval, oldprio, oldpolicy = -1;
4284 struct prio_array *array;
4285 unsigned long flags;
4288 /* may grab non-irq protected spin_locks */
4289 BUG_ON(in_interrupt());
4291 /* double check policy once rq lock held */
4293 policy = oldpolicy = p->policy;
4294 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4295 policy != SCHED_NORMAL && policy != SCHED_BATCH)
4298 * Valid priorities for SCHED_FIFO and SCHED_RR are
4299 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
4302 if (param->sched_priority < 0 ||
4303 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4304 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4306 if ((policy == SCHED_NORMAL || policy == SCHED_BATCH)
4307 != (param->sched_priority == 0))
4311 * Allow unprivileged RT tasks to decrease priority:
4313 if (!capable(CAP_SYS_NICE)) {
4315 * can't change policy, except between SCHED_NORMAL
4318 if (((policy != SCHED_NORMAL && p->policy != SCHED_BATCH) &&
4319 (policy != SCHED_BATCH && p->policy != SCHED_NORMAL)) &&
4320 !p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
4322 /* can't increase priority */
4323 if ((policy != SCHED_NORMAL && policy != SCHED_BATCH) &&
4324 param->sched_priority > p->rt_priority &&
4325 param->sched_priority >
4326 p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
4328 /* can't change other user's priorities */
4329 if ((current->euid != p->euid) &&
4330 (current->euid != p->uid))
4334 retval = security_task_setscheduler(p, policy, param);
4338 * make sure no PI-waiters arrive (or leave) while we are
4339 * changing the priority of the task:
4341 spin_lock_irqsave(&p->pi_lock, flags);
4343 * To be able to change p->policy safely, the apropriate
4344 * runqueue lock must be held.
4346 rq = __task_rq_lock(p);
4347 /* recheck policy now with rq lock held */
4348 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4349 policy = oldpolicy = -1;
4350 __task_rq_unlock(rq);
4351 spin_unlock_irqrestore(&p->pi_lock, flags);
4356 deactivate_task(p, rq);
4358 __setscheduler(p, policy, param->sched_priority);
4360 vx_activate_task(p);
4361 __activate_task(p, rq);
4363 * Reschedule if we are currently running on this runqueue and
4364 * our priority decreased, or if we are not currently running on
4365 * this runqueue and our priority is higher than the current's
4367 if (task_running(rq, p)) {
4368 if (p->prio > oldprio)
4369 resched_task(rq->curr);
4370 } else if (TASK_PREEMPTS_CURR(p, rq))
4371 resched_task(rq->curr);
4373 __task_rq_unlock(rq);
4374 spin_unlock_irqrestore(&p->pi_lock, flags);
4376 rt_mutex_adjust_pi(p);
4380 EXPORT_SYMBOL_GPL(sched_setscheduler);
4383 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4385 struct sched_param lparam;
4386 struct task_struct *p;
4389 if (!param || pid < 0)
4391 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4393 read_lock_irq(&tasklist_lock);
4394 p = find_process_by_pid(pid);
4396 read_unlock_irq(&tasklist_lock);
4399 retval = sched_setscheduler(p, policy, &lparam);
4400 read_unlock_irq(&tasklist_lock);
4406 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4407 * @pid: the pid in question.
4408 * @policy: new policy.
4409 * @param: structure containing the new RT priority.
4411 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4412 struct sched_param __user *param)
4414 /* negative values for policy are not valid */
4418 return do_sched_setscheduler(pid, policy, param);
4422 * sys_sched_setparam - set/change the RT priority of a thread
4423 * @pid: the pid in question.
4424 * @param: structure containing the new RT priority.
4426 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4428 return do_sched_setscheduler(pid, -1, param);
4432 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4433 * @pid: the pid in question.
4435 asmlinkage long sys_sched_getscheduler(pid_t pid)
4437 struct task_struct *p;
4438 int retval = -EINVAL;
4444 read_lock(&tasklist_lock);
4445 p = find_process_by_pid(pid);
4447 retval = security_task_getscheduler(p);
4451 read_unlock(&tasklist_lock);
4458 * sys_sched_getscheduler - get the RT priority of a thread
4459 * @pid: the pid in question.
4460 * @param: structure containing the RT priority.
4462 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4464 struct sched_param lp;
4465 struct task_struct *p;
4466 int retval = -EINVAL;
4468 if (!param || pid < 0)
4471 read_lock(&tasklist_lock);
4472 p = find_process_by_pid(pid);
4477 retval = security_task_getscheduler(p);
4481 lp.sched_priority = p->rt_priority;
4482 read_unlock(&tasklist_lock);
4485 * This one might sleep, we cannot do it with a spinlock held ...
4487 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4493 read_unlock(&tasklist_lock);
4497 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4499 cpumask_t cpus_allowed;
4500 struct task_struct *p;
4504 read_lock(&tasklist_lock);
4506 p = find_process_by_pid(pid);
4508 read_unlock(&tasklist_lock);
4509 unlock_cpu_hotplug();
4514 * It is not safe to call set_cpus_allowed with the
4515 * tasklist_lock held. We will bump the task_struct's
4516 * usage count and then drop tasklist_lock.
4519 read_unlock(&tasklist_lock);
4522 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4523 !capable(CAP_SYS_NICE))
4526 retval = security_task_setscheduler(p, 0, NULL);
4530 cpus_allowed = cpuset_cpus_allowed(p);
4531 cpus_and(new_mask, new_mask, cpus_allowed);
4532 retval = set_cpus_allowed(p, new_mask);
4536 unlock_cpu_hotplug();
4540 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4541 cpumask_t *new_mask)
4543 if (len < sizeof(cpumask_t)) {
4544 memset(new_mask, 0, sizeof(cpumask_t));
4545 } else if (len > sizeof(cpumask_t)) {
4546 len = sizeof(cpumask_t);
4548 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4552 * sys_sched_setaffinity - set the cpu affinity of a process
4553 * @pid: pid of the process
4554 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4555 * @user_mask_ptr: user-space pointer to the new cpu mask
4557 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4558 unsigned long __user *user_mask_ptr)
4563 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4567 return sched_setaffinity(pid, new_mask);
4571 * Represents all cpu's present in the system
4572 * In systems capable of hotplug, this map could dynamically grow
4573 * as new cpu's are detected in the system via any platform specific
4574 * method, such as ACPI for e.g.
4577 cpumask_t cpu_present_map __read_mostly;
4578 EXPORT_SYMBOL(cpu_present_map);
4581 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4582 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4585 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4587 struct task_struct *p;
4591 read_lock(&tasklist_lock);
4594 p = find_process_by_pid(pid);
4598 retval = security_task_getscheduler(p);
4602 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4605 read_unlock(&tasklist_lock);
4606 unlock_cpu_hotplug();
4614 * sys_sched_getaffinity - get the cpu affinity of a process
4615 * @pid: pid of the process
4616 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4617 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4619 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4620 unsigned long __user *user_mask_ptr)
4625 if (len < sizeof(cpumask_t))
4628 ret = sched_getaffinity(pid, &mask);
4632 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4635 return sizeof(cpumask_t);
4639 * sys_sched_yield - yield the current processor to other threads.
4641 * this function yields the current CPU by moving the calling thread
4642 * to the expired array. If there are no other threads running on this
4643 * CPU then this function will return.
4645 asmlinkage long sys_sched_yield(void)
4647 struct rq *rq = this_rq_lock();
4648 struct prio_array *array = current->array, *target = rq->expired;
4650 schedstat_inc(rq, yld_cnt);
4652 * We implement yielding by moving the task into the expired
4655 * (special rule: RT tasks will just roundrobin in the active
4658 if (rt_task(current))
4659 target = rq->active;
4661 if (array->nr_active == 1) {
4662 schedstat_inc(rq, yld_act_empty);
4663 if (!rq->expired->nr_active)
4664 schedstat_inc(rq, yld_both_empty);
4665 } else if (!rq->expired->nr_active)
4666 schedstat_inc(rq, yld_exp_empty);
4668 if (array != target) {
4669 dequeue_task(current, array);
4670 enqueue_task(current, target);
4673 * requeue_task is cheaper so perform that if possible.
4675 requeue_task(current, array);
4678 * Since we are going to call schedule() anyway, there's
4679 * no need to preempt or enable interrupts:
4681 __release(rq->lock);
4682 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4683 _raw_spin_unlock(&rq->lock);
4684 preempt_enable_no_resched();
4691 static inline int __resched_legal(int expected_preempt_count)
4693 if (unlikely(preempt_count() != expected_preempt_count))
4695 if (unlikely(system_state != SYSTEM_RUNNING))
4700 static void __cond_resched(void)
4702 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4703 __might_sleep(__FILE__, __LINE__);
4706 * The BKS might be reacquired before we have dropped
4707 * PREEMPT_ACTIVE, which could trigger a second
4708 * cond_resched() call.
4711 add_preempt_count(PREEMPT_ACTIVE);
4713 sub_preempt_count(PREEMPT_ACTIVE);
4714 } while (need_resched());
4717 int __sched cond_resched(void)
4719 if (need_resched() && __resched_legal(0)) {
4725 EXPORT_SYMBOL(cond_resched);
4728 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4729 * call schedule, and on return reacquire the lock.
4731 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4732 * operations here to prevent schedule() from being called twice (once via
4733 * spin_unlock(), once by hand).
4735 int cond_resched_lock(spinlock_t *lock)
4739 if (need_lockbreak(lock)) {
4745 if (need_resched() && __resched_legal(1)) {
4746 spin_release(&lock->dep_map, 1, _THIS_IP_);
4747 _raw_spin_unlock(lock);
4748 preempt_enable_no_resched();
4755 EXPORT_SYMBOL(cond_resched_lock);
4757 int __sched cond_resched_softirq(void)
4759 BUG_ON(!in_softirq());
4761 if (need_resched() && __resched_legal(0)) {
4762 raw_local_irq_disable();
4764 raw_local_irq_enable();
4771 EXPORT_SYMBOL(cond_resched_softirq);
4774 * yield - yield the current processor to other threads.
4776 * this is a shortcut for kernel-space yielding - it marks the
4777 * thread runnable and calls sys_sched_yield().
4779 void __sched yield(void)
4781 set_current_state(TASK_RUNNING);
4784 EXPORT_SYMBOL(yield);
4787 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4788 * that process accounting knows that this is a task in IO wait state.
4790 * But don't do that if it is a deliberate, throttling IO wait (this task
4791 * has set its backing_dev_info: the queue against which it should throttle)
4793 void __sched io_schedule(void)
4795 struct rq *rq = &__raw_get_cpu_var(runqueues);
4797 delayacct_blkio_start();
4798 atomic_inc(&rq->nr_iowait);
4800 atomic_dec(&rq->nr_iowait);
4801 delayacct_blkio_end();
4803 EXPORT_SYMBOL(io_schedule);
4805 long __sched io_schedule_timeout(long timeout)
4807 struct rq *rq = &__raw_get_cpu_var(runqueues);
4810 delayacct_blkio_start();
4811 atomic_inc(&rq->nr_iowait);
4812 ret = schedule_timeout(timeout);
4813 atomic_dec(&rq->nr_iowait);
4814 delayacct_blkio_end();
4819 * sys_sched_get_priority_max - return maximum RT priority.
4820 * @policy: scheduling class.
4822 * this syscall returns the maximum rt_priority that can be used
4823 * by a given scheduling class.
4825 asmlinkage long sys_sched_get_priority_max(int policy)
4832 ret = MAX_USER_RT_PRIO-1;
4843 * sys_sched_get_priority_min - return minimum RT priority.
4844 * @policy: scheduling class.
4846 * this syscall returns the minimum rt_priority that can be used
4847 * by a given scheduling class.
4849 asmlinkage long sys_sched_get_priority_min(int policy)
4866 * sys_sched_rr_get_interval - return the default timeslice of a process.
4867 * @pid: pid of the process.
4868 * @interval: userspace pointer to the timeslice value.
4870 * this syscall writes the default timeslice value of a given process
4871 * into the user-space timespec buffer. A value of '0' means infinity.
4874 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4876 struct task_struct *p;
4877 int retval = -EINVAL;
4884 read_lock(&tasklist_lock);
4885 p = find_process_by_pid(pid);
4889 retval = security_task_getscheduler(p);
4893 jiffies_to_timespec(p->policy == SCHED_FIFO ?
4894 0 : task_timeslice(p), &t);
4895 read_unlock(&tasklist_lock);
4896 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4900 read_unlock(&tasklist_lock);
4904 static inline struct task_struct *eldest_child(struct task_struct *p)
4906 if (list_empty(&p->children))
4908 return list_entry(p->children.next,struct task_struct,sibling);
4911 static inline struct task_struct *older_sibling(struct task_struct *p)
4913 if (p->sibling.prev==&p->parent->children)
4915 return list_entry(p->sibling.prev,struct task_struct,sibling);
4918 static inline struct task_struct *younger_sibling(struct task_struct *p)
4920 if (p->sibling.next==&p->parent->children)
4922 return list_entry(p->sibling.next,struct task_struct,sibling);
4925 static const char stat_nam[] = "RSDTtZX";
4927 static void show_task(struct task_struct *p)
4929 struct task_struct *relative;
4930 unsigned long free = 0;
4933 state = p->state ? __ffs(p->state) + 1 : 0;
4934 printk("%-13.13s %c", p->comm,
4935 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4936 #if (BITS_PER_LONG == 32)
4937 if (state == TASK_RUNNING)
4938 printk(" running ");
4940 printk(" %08lX ", thread_saved_pc(p));
4942 if (state == TASK_RUNNING)
4943 printk(" running task ");
4945 printk(" %016lx ", thread_saved_pc(p));
4947 #ifdef CONFIG_DEBUG_STACK_USAGE
4949 unsigned long *n = end_of_stack(p);
4952 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4955 printk("%5lu %5d %6d ", free, p->pid, p->parent->pid);
4956 if ((relative = eldest_child(p)))
4957 printk("%5d ", relative->pid);
4960 if ((relative = younger_sibling(p)))
4961 printk("%7d", relative->pid);
4964 if ((relative = older_sibling(p)))
4965 printk(" %5d", relative->pid);
4969 printk(" (L-TLB)\n");
4971 printk(" (NOTLB)\n");
4973 if (state != TASK_RUNNING)
4974 show_stack(p, NULL);
4977 void show_state(void)
4979 struct task_struct *g, *p;
4981 #if (BITS_PER_LONG == 32)
4984 printk(" task PC pid father child younger older\n");
4988 printk(" task PC pid father child younger older\n");
4990 read_lock(&tasklist_lock);
4991 do_each_thread(g, p) {
4993 * reset the NMI-timeout, listing all files on a slow
4994 * console might take alot of time:
4996 touch_nmi_watchdog();
4998 } while_each_thread(g, p);
5000 read_unlock(&tasklist_lock);
5001 debug_show_all_locks();
5005 * init_idle - set up an idle thread for a given CPU
5006 * @idle: task in question
5007 * @cpu: cpu the idle task belongs to
5009 * NOTE: this function does not set the idle thread's NEED_RESCHED
5010 * flag, to make booting more robust.
5012 void __devinit init_idle(struct task_struct *idle, int cpu)
5014 struct rq *rq = cpu_rq(cpu);
5015 unsigned long flags;
5017 idle->timestamp = sched_clock();
5018 idle->sleep_avg = 0;
5020 idle->prio = idle->normal_prio = MAX_PRIO;
5021 idle->state = TASK_RUNNING;
5022 idle->cpus_allowed = cpumask_of_cpu(cpu);
5023 set_task_cpu(idle, cpu);
5025 spin_lock_irqsave(&rq->lock, flags);
5026 rq->curr = rq->idle = idle;
5027 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5030 spin_unlock_irqrestore(&rq->lock, flags);
5032 /* Set the preempt count _outside_ the spinlocks! */
5033 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
5034 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5036 task_thread_info(idle)->preempt_count = 0;
5041 * In a system that switches off the HZ timer nohz_cpu_mask
5042 * indicates which cpus entered this state. This is used
5043 * in the rcu update to wait only for active cpus. For system
5044 * which do not switch off the HZ timer nohz_cpu_mask should
5045 * always be CPU_MASK_NONE.
5047 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5051 * This is how migration works:
5053 * 1) we queue a struct migration_req structure in the source CPU's
5054 * runqueue and wake up that CPU's migration thread.
5055 * 2) we down() the locked semaphore => thread blocks.
5056 * 3) migration thread wakes up (implicitly it forces the migrated
5057 * thread off the CPU)
5058 * 4) it gets the migration request and checks whether the migrated
5059 * task is still in the wrong runqueue.
5060 * 5) if it's in the wrong runqueue then the migration thread removes
5061 * it and puts it into the right queue.
5062 * 6) migration thread up()s the semaphore.
5063 * 7) we wake up and the migration is done.
5067 * Change a given task's CPU affinity. Migrate the thread to a
5068 * proper CPU and schedule it away if the CPU it's executing on
5069 * is removed from the allowed bitmask.
5071 * NOTE: the caller must have a valid reference to the task, the
5072 * task must not exit() & deallocate itself prematurely. The
5073 * call is not atomic; no spinlocks may be held.
5075 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
5077 struct migration_req req;
5078 unsigned long flags;
5082 rq = task_rq_lock(p, &flags);
5083 if (!cpus_intersects(new_mask, cpu_online_map)) {
5088 p->cpus_allowed = new_mask;
5089 /* Can the task run on the task's current CPU? If so, we're done */
5090 if (cpu_isset(task_cpu(p), new_mask))
5093 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5094 /* Need help from migration thread: drop lock and wait. */
5095 task_rq_unlock(rq, &flags);
5096 wake_up_process(rq->migration_thread);
5097 wait_for_completion(&req.done);
5098 tlb_migrate_finish(p->mm);
5102 task_rq_unlock(rq, &flags);
5106 EXPORT_SYMBOL_GPL(set_cpus_allowed);
5109 * Move (not current) task off this cpu, onto dest cpu. We're doing
5110 * this because either it can't run here any more (set_cpus_allowed()
5111 * away from this CPU, or CPU going down), or because we're
5112 * attempting to rebalance this task on exec (sched_exec).
5114 * So we race with normal scheduler movements, but that's OK, as long
5115 * as the task is no longer on this CPU.
5117 * Returns non-zero if task was successfully migrated.
5119 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5121 struct rq *rq_dest, *rq_src;
5124 if (unlikely(cpu_is_offline(dest_cpu)))
5127 rq_src = cpu_rq(src_cpu);
5128 rq_dest = cpu_rq(dest_cpu);
5130 double_rq_lock(rq_src, rq_dest);
5131 /* Already moved. */
5132 if (task_cpu(p) != src_cpu)
5134 /* Affinity changed (again). */
5135 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5138 set_task_cpu(p, dest_cpu);
5141 * Sync timestamp with rq_dest's before activating.
5142 * The same thing could be achieved by doing this step
5143 * afterwards, and pretending it was a local activate.
5144 * This way is cleaner and logically correct.
5146 p->timestamp = p->timestamp - rq_src->timestamp_last_tick
5147 + rq_dest->timestamp_last_tick;
5148 deactivate_task(p, rq_src);
5149 vx_activate_task(p);
5150 __activate_task(p, rq_dest);
5151 if (TASK_PREEMPTS_CURR(p, rq_dest))
5152 resched_task(rq_dest->curr);
5156 double_rq_unlock(rq_src, rq_dest);
5161 * migration_thread - this is a highprio system thread that performs
5162 * thread migration by bumping thread off CPU then 'pushing' onto
5165 static int migration_thread(void *data)
5167 int cpu = (long)data;
5171 BUG_ON(rq->migration_thread != current);
5173 set_current_state(TASK_INTERRUPTIBLE);
5174 while (!kthread_should_stop()) {
5175 struct migration_req *req;
5176 struct list_head *head;
5180 spin_lock_irq(&rq->lock);
5182 if (cpu_is_offline(cpu)) {
5183 spin_unlock_irq(&rq->lock);
5187 if (rq->active_balance) {
5188 active_load_balance(rq, cpu);
5189 rq->active_balance = 0;
5192 head = &rq->migration_queue;
5194 if (list_empty(head)) {
5195 spin_unlock_irq(&rq->lock);
5197 set_current_state(TASK_INTERRUPTIBLE);
5200 req = list_entry(head->next, struct migration_req, list);
5201 list_del_init(head->next);
5203 spin_unlock(&rq->lock);
5204 __migrate_task(req->task, cpu, req->dest_cpu);
5207 complete(&req->done);
5209 __set_current_state(TASK_RUNNING);
5213 /* Wait for kthread_stop */
5214 set_current_state(TASK_INTERRUPTIBLE);
5215 while (!kthread_should_stop()) {
5217 set_current_state(TASK_INTERRUPTIBLE);
5219 __set_current_state(TASK_RUNNING);
5223 #ifdef CONFIG_HOTPLUG_CPU
5224 /* Figure out where task on dead CPU should go, use force if neccessary. */
5225 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5227 unsigned long flags;
5234 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5235 cpus_and(mask, mask, p->cpus_allowed);
5236 dest_cpu = any_online_cpu(mask);
5238 /* On any allowed CPU? */
5239 if (dest_cpu == NR_CPUS)
5240 dest_cpu = any_online_cpu(p->cpus_allowed);
5242 /* No more Mr. Nice Guy. */
5243 if (dest_cpu == NR_CPUS) {
5244 rq = task_rq_lock(p, &flags);
5245 cpus_setall(p->cpus_allowed);
5246 dest_cpu = any_online_cpu(p->cpus_allowed);
5247 task_rq_unlock(rq, &flags);
5250 * Don't tell them about moving exiting tasks or
5251 * kernel threads (both mm NULL), since they never
5254 if (p->mm && printk_ratelimit())
5255 printk(KERN_INFO "process %d (%s) no "
5256 "longer affine to cpu%d\n",
5257 p->pid, p->comm, dead_cpu);
5259 if (!__migrate_task(p, dead_cpu, dest_cpu))
5264 * While a dead CPU has no uninterruptible tasks queued at this point,
5265 * it might still have a nonzero ->nr_uninterruptible counter, because
5266 * for performance reasons the counter is not stricly tracking tasks to
5267 * their home CPUs. So we just add the counter to another CPU's counter,
5268 * to keep the global sum constant after CPU-down:
5270 static void migrate_nr_uninterruptible(struct rq *rq_src)
5272 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5273 unsigned long flags;
5275 local_irq_save(flags);
5276 double_rq_lock(rq_src, rq_dest);
5277 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5278 rq_src->nr_uninterruptible = 0;
5279 double_rq_unlock(rq_src, rq_dest);
5280 local_irq_restore(flags);
5283 /* Run through task list and migrate tasks from the dead cpu. */
5284 static void migrate_live_tasks(int src_cpu)
5286 struct task_struct *p, *t;
5288 write_lock_irq(&tasklist_lock);
5290 do_each_thread(t, p) {
5294 if (task_cpu(p) == src_cpu)
5295 move_task_off_dead_cpu(src_cpu, p);
5296 } while_each_thread(t, p);
5298 write_unlock_irq(&tasklist_lock);
5301 /* Schedules idle task to be the next runnable task on current CPU.
5302 * It does so by boosting its priority to highest possible and adding it to
5303 * the _front_ of the runqueue. Used by CPU offline code.
5305 void sched_idle_next(void)
5307 int this_cpu = smp_processor_id();
5308 struct rq *rq = cpu_rq(this_cpu);
5309 struct task_struct *p = rq->idle;
5310 unsigned long flags;
5312 /* cpu has to be offline */
5313 BUG_ON(cpu_online(this_cpu));
5316 * Strictly not necessary since rest of the CPUs are stopped by now
5317 * and interrupts disabled on the current cpu.
5319 spin_lock_irqsave(&rq->lock, flags);
5321 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
5323 /* Add idle task to the _front_ of its priority queue: */
5324 __activate_idle_task(p, rq);
5326 spin_unlock_irqrestore(&rq->lock, flags);
5330 * Ensures that the idle task is using init_mm right before its cpu goes
5333 void idle_task_exit(void)
5335 struct mm_struct *mm = current->active_mm;
5337 BUG_ON(cpu_online(smp_processor_id()));
5340 switch_mm(mm, &init_mm, current);
5344 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5346 struct rq *rq = cpu_rq(dead_cpu);
5348 /* Must be exiting, otherwise would be on tasklist. */
5349 BUG_ON(p->exit_state != EXIT_ZOMBIE && p->exit_state != EXIT_DEAD);
5351 /* Cannot have done final schedule yet: would have vanished. */
5352 BUG_ON(p->flags & PF_DEAD);
5357 * Drop lock around migration; if someone else moves it,
5358 * that's OK. No task can be added to this CPU, so iteration is
5361 spin_unlock_irq(&rq->lock);
5362 move_task_off_dead_cpu(dead_cpu, p);
5363 spin_lock_irq(&rq->lock);
5368 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5369 static void migrate_dead_tasks(unsigned int dead_cpu)
5371 struct rq *rq = cpu_rq(dead_cpu);
5372 unsigned int arr, i;
5374 for (arr = 0; arr < 2; arr++) {
5375 for (i = 0; i < MAX_PRIO; i++) {
5376 struct list_head *list = &rq->arrays[arr].queue[i];
5378 while (!list_empty(list))
5379 migrate_dead(dead_cpu, list_entry(list->next,
5380 struct task_struct, run_list));
5384 #endif /* CONFIG_HOTPLUG_CPU */
5387 * migration_call - callback that gets triggered when a CPU is added.
5388 * Here we can start up the necessary migration thread for the new CPU.
5390 static int __cpuinit
5391 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5393 struct task_struct *p;
5394 int cpu = (long)hcpu;
5395 unsigned long flags;
5399 case CPU_UP_PREPARE:
5400 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
5403 p->flags |= PF_NOFREEZE;
5404 kthread_bind(p, cpu);
5405 /* Must be high prio: stop_machine expects to yield to it. */
5406 rq = task_rq_lock(p, &flags);
5407 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
5408 task_rq_unlock(rq, &flags);
5409 cpu_rq(cpu)->migration_thread = p;
5413 /* Strictly unneccessary, as first user will wake it. */
5414 wake_up_process(cpu_rq(cpu)->migration_thread);
5417 #ifdef CONFIG_HOTPLUG_CPU
5418 case CPU_UP_CANCELED:
5419 if (!cpu_rq(cpu)->migration_thread)
5421 /* Unbind it from offline cpu so it can run. Fall thru. */
5422 kthread_bind(cpu_rq(cpu)->migration_thread,
5423 any_online_cpu(cpu_online_map));
5424 kthread_stop(cpu_rq(cpu)->migration_thread);
5425 cpu_rq(cpu)->migration_thread = NULL;
5429 migrate_live_tasks(cpu);
5431 kthread_stop(rq->migration_thread);
5432 rq->migration_thread = NULL;
5433 /* Idle task back to normal (off runqueue, low prio) */
5434 rq = task_rq_lock(rq->idle, &flags);
5435 deactivate_task(rq->idle, rq);
5436 rq->idle->static_prio = MAX_PRIO;
5437 __setscheduler(rq->idle, SCHED_NORMAL, 0);
5438 migrate_dead_tasks(cpu);
5439 task_rq_unlock(rq, &flags);
5440 migrate_nr_uninterruptible(rq);
5441 BUG_ON(rq->nr_running != 0);
5443 /* No need to migrate the tasks: it was best-effort if
5444 * they didn't do lock_cpu_hotplug(). Just wake up
5445 * the requestors. */
5446 spin_lock_irq(&rq->lock);
5447 while (!list_empty(&rq->migration_queue)) {
5448 struct migration_req *req;
5450 req = list_entry(rq->migration_queue.next,
5451 struct migration_req, list);
5452 list_del_init(&req->list);
5453 complete(&req->done);
5455 spin_unlock_irq(&rq->lock);
5462 /* Register at highest priority so that task migration (migrate_all_tasks)
5463 * happens before everything else.
5465 static struct notifier_block __cpuinitdata migration_notifier = {
5466 .notifier_call = migration_call,
5470 int __init migration_init(void)
5472 void *cpu = (void *)(long)smp_processor_id();
5474 /* Start one for the boot CPU: */
5475 migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5476 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5477 register_cpu_notifier(&migration_notifier);
5484 #undef SCHED_DOMAIN_DEBUG
5485 #ifdef SCHED_DOMAIN_DEBUG
5486 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5491 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5495 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5500 struct sched_group *group = sd->groups;
5501 cpumask_t groupmask;
5503 cpumask_scnprintf(str, NR_CPUS, sd->span);
5504 cpus_clear(groupmask);
5507 for (i = 0; i < level + 1; i++)
5509 printk("domain %d: ", level);
5511 if (!(sd->flags & SD_LOAD_BALANCE)) {
5512 printk("does not load-balance\n");
5514 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent");
5518 printk("span %s\n", str);
5520 if (!cpu_isset(cpu, sd->span))
5521 printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
5522 if (!cpu_isset(cpu, group->cpumask))
5523 printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
5526 for (i = 0; i < level + 2; i++)
5532 printk(KERN_ERR "ERROR: group is NULL\n");
5536 if (!group->cpu_power) {
5538 printk(KERN_ERR "ERROR: domain->cpu_power not set\n");
5541 if (!cpus_weight(group->cpumask)) {
5543 printk(KERN_ERR "ERROR: empty group\n");
5546 if (cpus_intersects(groupmask, group->cpumask)) {
5548 printk(KERN_ERR "ERROR: repeated CPUs\n");
5551 cpus_or(groupmask, groupmask, group->cpumask);
5553 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5556 group = group->next;
5557 } while (group != sd->groups);
5560 if (!cpus_equal(sd->span, groupmask))
5561 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5567 if (!cpus_subset(groupmask, sd->span))
5568 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
5574 # define sched_domain_debug(sd, cpu) do { } while (0)
5577 static int sd_degenerate(struct sched_domain *sd)
5579 if (cpus_weight(sd->span) == 1)
5582 /* Following flags need at least 2 groups */
5583 if (sd->flags & (SD_LOAD_BALANCE |
5584 SD_BALANCE_NEWIDLE |
5587 if (sd->groups != sd->groups->next)
5591 /* Following flags don't use groups */
5592 if (sd->flags & (SD_WAKE_IDLE |
5601 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5603 unsigned long cflags = sd->flags, pflags = parent->flags;
5605 if (sd_degenerate(parent))
5608 if (!cpus_equal(sd->span, parent->span))
5611 /* Does parent contain flags not in child? */
5612 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5613 if (cflags & SD_WAKE_AFFINE)
5614 pflags &= ~SD_WAKE_BALANCE;
5615 /* Flags needing groups don't count if only 1 group in parent */
5616 if (parent->groups == parent->groups->next) {
5617 pflags &= ~(SD_LOAD_BALANCE |
5618 SD_BALANCE_NEWIDLE |
5622 if (~cflags & pflags)
5629 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5630 * hold the hotplug lock.
5632 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5634 struct rq *rq = cpu_rq(cpu);
5635 struct sched_domain *tmp;
5637 /* Remove the sched domains which do not contribute to scheduling. */
5638 for (tmp = sd; tmp; tmp = tmp->parent) {
5639 struct sched_domain *parent = tmp->parent;
5642 if (sd_parent_degenerate(tmp, parent))
5643 tmp->parent = parent->parent;
5646 if (sd && sd_degenerate(sd))
5649 sched_domain_debug(sd, cpu);
5651 rcu_assign_pointer(rq->sd, sd);
5654 /* cpus with isolated domains */
5655 static cpumask_t __devinitdata cpu_isolated_map = CPU_MASK_NONE;
5657 /* Setup the mask of cpus configured for isolated domains */
5658 static int __init isolated_cpu_setup(char *str)
5660 int ints[NR_CPUS], i;
5662 str = get_options(str, ARRAY_SIZE(ints), ints);
5663 cpus_clear(cpu_isolated_map);
5664 for (i = 1; i <= ints[0]; i++)
5665 if (ints[i] < NR_CPUS)
5666 cpu_set(ints[i], cpu_isolated_map);
5670 __setup ("isolcpus=", isolated_cpu_setup);
5673 * init_sched_build_groups takes an array of groups, the cpumask we wish
5674 * to span, and a pointer to a function which identifies what group a CPU
5675 * belongs to. The return value of group_fn must be a valid index into the
5676 * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
5677 * keep track of groups covered with a cpumask_t).
5679 * init_sched_build_groups will build a circular linked list of the groups
5680 * covered by the given span, and will set each group's ->cpumask correctly,
5681 * and ->cpu_power to 0.
5683 static void init_sched_build_groups(struct sched_group groups[], cpumask_t span,
5684 int (*group_fn)(int cpu))
5686 struct sched_group *first = NULL, *last = NULL;
5687 cpumask_t covered = CPU_MASK_NONE;
5690 for_each_cpu_mask(i, span) {
5691 int group = group_fn(i);
5692 struct sched_group *sg = &groups[group];
5695 if (cpu_isset(i, covered))
5698 sg->cpumask = CPU_MASK_NONE;
5701 for_each_cpu_mask(j, span) {
5702 if (group_fn(j) != group)
5705 cpu_set(j, covered);
5706 cpu_set(j, sg->cpumask);
5717 #define SD_NODES_PER_DOMAIN 16
5720 * Self-tuning task migration cost measurement between source and target CPUs.
5722 * This is done by measuring the cost of manipulating buffers of varying
5723 * sizes. For a given buffer-size here are the steps that are taken:
5725 * 1) the source CPU reads+dirties a shared buffer
5726 * 2) the target CPU reads+dirties the same shared buffer
5728 * We measure how long they take, in the following 4 scenarios:
5730 * - source: CPU1, target: CPU2 | cost1
5731 * - source: CPU2, target: CPU1 | cost2
5732 * - source: CPU1, target: CPU1 | cost3
5733 * - source: CPU2, target: CPU2 | cost4
5735 * We then calculate the cost3+cost4-cost1-cost2 difference - this is
5736 * the cost of migration.
5738 * We then start off from a small buffer-size and iterate up to larger
5739 * buffer sizes, in 5% steps - measuring each buffer-size separately, and
5740 * doing a maximum search for the cost. (The maximum cost for a migration
5741 * normally occurs when the working set size is around the effective cache
5744 #define SEARCH_SCOPE 2
5745 #define MIN_CACHE_SIZE (64*1024U)
5746 #define DEFAULT_CACHE_SIZE (5*1024*1024U)
5747 #define ITERATIONS 1
5748 #define SIZE_THRESH 130
5749 #define COST_THRESH 130
5752 * The migration cost is a function of 'domain distance'. Domain
5753 * distance is the number of steps a CPU has to iterate down its
5754 * domain tree to share a domain with the other CPU. The farther
5755 * two CPUs are from each other, the larger the distance gets.
5757 * Note that we use the distance only to cache measurement results,
5758 * the distance value is not used numerically otherwise. When two
5759 * CPUs have the same distance it is assumed that the migration
5760 * cost is the same. (this is a simplification but quite practical)
5762 #define MAX_DOMAIN_DISTANCE 32
5764 static unsigned long long migration_cost[MAX_DOMAIN_DISTANCE] =
5765 { [ 0 ... MAX_DOMAIN_DISTANCE-1 ] =
5767 * Architectures may override the migration cost and thus avoid
5768 * boot-time calibration. Unit is nanoseconds. Mostly useful for
5769 * virtualized hardware:
5771 #ifdef CONFIG_DEFAULT_MIGRATION_COST
5772 CONFIG_DEFAULT_MIGRATION_COST
5779 * Allow override of migration cost - in units of microseconds.
5780 * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost
5781 * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs:
5783 static int __init migration_cost_setup(char *str)
5785 int ints[MAX_DOMAIN_DISTANCE+1], i;
5787 str = get_options(str, ARRAY_SIZE(ints), ints);
5789 printk("#ints: %d\n", ints[0]);
5790 for (i = 1; i <= ints[0]; i++) {
5791 migration_cost[i-1] = (unsigned long long)ints[i]*1000;
5792 printk("migration_cost[%d]: %Ld\n", i-1, migration_cost[i-1]);
5797 __setup ("migration_cost=", migration_cost_setup);
5800 * Global multiplier (divisor) for migration-cutoff values,
5801 * in percentiles. E.g. use a value of 150 to get 1.5 times
5802 * longer cache-hot cutoff times.
5804 * (We scale it from 100 to 128 to long long handling easier.)
5807 #define MIGRATION_FACTOR_SCALE 128
5809 static unsigned int migration_factor = MIGRATION_FACTOR_SCALE;
5811 static int __init setup_migration_factor(char *str)
5813 get_option(&str, &migration_factor);
5814 migration_factor = migration_factor * MIGRATION_FACTOR_SCALE / 100;
5818 __setup("migration_factor=", setup_migration_factor);
5821 * Estimated distance of two CPUs, measured via the number of domains
5822 * we have to pass for the two CPUs to be in the same span:
5824 static unsigned long domain_distance(int cpu1, int cpu2)
5826 unsigned long distance = 0;
5827 struct sched_domain *sd;
5829 for_each_domain(cpu1, sd) {
5830 WARN_ON(!cpu_isset(cpu1, sd->span));
5831 if (cpu_isset(cpu2, sd->span))
5835 if (distance >= MAX_DOMAIN_DISTANCE) {
5837 distance = MAX_DOMAIN_DISTANCE-1;
5843 static unsigned int migration_debug;
5845 static int __init setup_migration_debug(char *str)
5847 get_option(&str, &migration_debug);
5851 __setup("migration_debug=", setup_migration_debug);
5854 * Maximum cache-size that the scheduler should try to measure.
5855 * Architectures with larger caches should tune this up during
5856 * bootup. Gets used in the domain-setup code (i.e. during SMP
5859 unsigned int max_cache_size;
5861 static int __init setup_max_cache_size(char *str)
5863 get_option(&str, &max_cache_size);
5867 __setup("max_cache_size=", setup_max_cache_size);
5870 * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This
5871 * is the operation that is timed, so we try to generate unpredictable
5872 * cachemisses that still end up filling the L2 cache:
5874 static void touch_cache(void *__cache, unsigned long __size)
5876 unsigned long size = __size/sizeof(long), chunk1 = size/3,
5878 unsigned long *cache = __cache;
5881 for (i = 0; i < size/6; i += 8) {
5884 case 1: cache[size-1-i]++;
5885 case 2: cache[chunk1-i]++;
5886 case 3: cache[chunk1+i]++;
5887 case 4: cache[chunk2-i]++;
5888 case 5: cache[chunk2+i]++;
5894 * Measure the cache-cost of one task migration. Returns in units of nsec.
5896 static unsigned long long
5897 measure_one(void *cache, unsigned long size, int source, int target)
5899 cpumask_t mask, saved_mask;
5900 unsigned long long t0, t1, t2, t3, cost;
5902 saved_mask = current->cpus_allowed;
5905 * Flush source caches to RAM and invalidate them:
5910 * Migrate to the source CPU:
5912 mask = cpumask_of_cpu(source);
5913 set_cpus_allowed(current, mask);
5914 WARN_ON(smp_processor_id() != source);
5917 * Dirty the working set:
5920 touch_cache(cache, size);
5924 * Migrate to the target CPU, dirty the L2 cache and access
5925 * the shared buffer. (which represents the working set
5926 * of a migrated task.)
5928 mask = cpumask_of_cpu(target);
5929 set_cpus_allowed(current, mask);
5930 WARN_ON(smp_processor_id() != target);
5933 touch_cache(cache, size);
5936 cost = t1-t0 + t3-t2;
5938 if (migration_debug >= 2)
5939 printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n",
5940 source, target, t1-t0, t1-t0, t3-t2, cost);
5942 * Flush target caches to RAM and invalidate them:
5946 set_cpus_allowed(current, saved_mask);
5952 * Measure a series of task migrations and return the average
5953 * result. Since this code runs early during bootup the system
5954 * is 'undisturbed' and the average latency makes sense.
5956 * The algorithm in essence auto-detects the relevant cache-size,
5957 * so it will properly detect different cachesizes for different
5958 * cache-hierarchies, depending on how the CPUs are connected.
5960 * Architectures can prime the upper limit of the search range via
5961 * max_cache_size, otherwise the search range defaults to 20MB...64K.
5963 static unsigned long long
5964 measure_cost(int cpu1, int cpu2, void *cache, unsigned int size)
5966 unsigned long long cost1, cost2;
5970 * Measure the migration cost of 'size' bytes, over an
5971 * average of 10 runs:
5973 * (We perturb the cache size by a small (0..4k)
5974 * value to compensate size/alignment related artifacts.
5975 * We also subtract the cost of the operation done on
5981 * dry run, to make sure we start off cache-cold on cpu1,
5982 * and to get any vmalloc pagefaults in advance:
5984 measure_one(cache, size, cpu1, cpu2);
5985 for (i = 0; i < ITERATIONS; i++)
5986 cost1 += measure_one(cache, size - i*1024, cpu1, cpu2);
5988 measure_one(cache, size, cpu2, cpu1);
5989 for (i = 0; i < ITERATIONS; i++)
5990 cost1 += measure_one(cache, size - i*1024, cpu2, cpu1);
5993 * (We measure the non-migrating [cached] cost on both
5994 * cpu1 and cpu2, to handle CPUs with different speeds)
5998 measure_one(cache, size, cpu1, cpu1);
5999 for (i = 0; i < ITERATIONS; i++)
6000 cost2 += measure_one(cache, size - i*1024, cpu1, cpu1);
6002 measure_one(cache, size, cpu2, cpu2);
6003 for (i = 0; i < ITERATIONS; i++)
6004 cost2 += measure_one(cache, size - i*1024, cpu2, cpu2);
6007 * Get the per-iteration migration cost:
6009 do_div(cost1, 2*ITERATIONS);
6010 do_div(cost2, 2*ITERATIONS);
6012 return cost1 - cost2;
6015 static unsigned long long measure_migration_cost(int cpu1, int cpu2)
6017 unsigned long long max_cost = 0, fluct = 0, avg_fluct = 0;
6018 unsigned int max_size, size, size_found = 0;
6019 long long cost = 0, prev_cost;
6023 * Search from max_cache_size*5 down to 64K - the real relevant
6024 * cachesize has to lie somewhere inbetween.
6026 if (max_cache_size) {
6027 max_size = max(max_cache_size * SEARCH_SCOPE, MIN_CACHE_SIZE);
6028 size = max(max_cache_size / SEARCH_SCOPE, MIN_CACHE_SIZE);
6031 * Since we have no estimation about the relevant
6034 max_size = DEFAULT_CACHE_SIZE * SEARCH_SCOPE;
6035 size = MIN_CACHE_SIZE;
6038 if (!cpu_online(cpu1) || !cpu_online(cpu2)) {
6039 printk("cpu %d and %d not both online!\n", cpu1, cpu2);
6044 * Allocate the working set:
6046 cache = vmalloc(max_size);
6048 printk("could not vmalloc %d bytes for cache!\n", 2*max_size);
6049 return 1000000; /* return 1 msec on very small boxen */
6052 while (size <= max_size) {
6054 cost = measure_cost(cpu1, cpu2, cache, size);
6060 if (max_cost < cost) {
6066 * Calculate average fluctuation, we use this to prevent
6067 * noise from triggering an early break out of the loop:
6069 fluct = abs(cost - prev_cost);
6070 avg_fluct = (avg_fluct + fluct)/2;
6072 if (migration_debug)
6073 printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): (%8Ld %8Ld)\n",
6075 (long)cost / 1000000,
6076 ((long)cost / 100000) % 10,
6077 (long)max_cost / 1000000,
6078 ((long)max_cost / 100000) % 10,
6079 domain_distance(cpu1, cpu2),
6083 * If we iterated at least 20% past the previous maximum,
6084 * and the cost has dropped by more than 20% already,
6085 * (taking fluctuations into account) then we assume to
6086 * have found the maximum and break out of the loop early:
6088 if (size_found && (size*100 > size_found*SIZE_THRESH))
6089 if (cost+avg_fluct <= 0 ||
6090 max_cost*100 > (cost+avg_fluct)*COST_THRESH) {
6092 if (migration_debug)
6093 printk("-> found max.\n");
6097 * Increase the cachesize in 10% steps:
6099 size = size * 10 / 9;
6102 if (migration_debug)
6103 printk("[%d][%d] working set size found: %d, cost: %Ld\n",
6104 cpu1, cpu2, size_found, max_cost);
6109 * A task is considered 'cache cold' if at least 2 times
6110 * the worst-case cost of migration has passed.
6112 * (this limit is only listened to if the load-balancing
6113 * situation is 'nice' - if there is a large imbalance we
6114 * ignore it for the sake of CPU utilization and
6115 * processing fairness.)
6117 return 2 * max_cost * migration_factor / MIGRATION_FACTOR_SCALE;
6120 static void calibrate_migration_costs(const cpumask_t *cpu_map)
6122 int cpu1 = -1, cpu2 = -1, cpu, orig_cpu = raw_smp_processor_id();
6123 unsigned long j0, j1, distance, max_distance = 0;
6124 struct sched_domain *sd;
6129 * First pass - calculate the cacheflush times:
6131 for_each_cpu_mask(cpu1, *cpu_map) {
6132 for_each_cpu_mask(cpu2, *cpu_map) {
6135 distance = domain_distance(cpu1, cpu2);
6136 max_distance = max(max_distance, distance);
6138 * No result cached yet?
6140 if (migration_cost[distance] == -1LL)
6141 migration_cost[distance] =
6142 measure_migration_cost(cpu1, cpu2);
6146 * Second pass - update the sched domain hierarchy with
6147 * the new cache-hot-time estimations:
6149 for_each_cpu_mask(cpu, *cpu_map) {
6151 for_each_domain(cpu, sd) {
6152 sd->cache_hot_time = migration_cost[distance];
6159 if (migration_debug)
6160 printk("migration: max_cache_size: %d, cpu: %d MHz:\n",
6168 if (system_state == SYSTEM_BOOTING) {
6169 if (num_online_cpus() > 1) {
6170 printk("migration_cost=");
6171 for (distance = 0; distance <= max_distance; distance++) {
6174 printk("%ld", (long)migration_cost[distance] / 1000);
6180 if (migration_debug)
6181 printk("migration: %ld seconds\n", (j1-j0)/HZ);
6184 * Move back to the original CPU. NUMA-Q gets confused
6185 * if we migrate to another quad during bootup.
6187 if (raw_smp_processor_id() != orig_cpu) {
6188 cpumask_t mask = cpumask_of_cpu(orig_cpu),
6189 saved_mask = current->cpus_allowed;
6191 set_cpus_allowed(current, mask);
6192 set_cpus_allowed(current, saved_mask);
6199 * find_next_best_node - find the next node to include in a sched_domain
6200 * @node: node whose sched_domain we're building
6201 * @used_nodes: nodes already in the sched_domain
6203 * Find the next node to include in a given scheduling domain. Simply
6204 * finds the closest node not already in the @used_nodes map.
6206 * Should use nodemask_t.
6208 static int find_next_best_node(int node, unsigned long *used_nodes)
6210 int i, n, val, min_val, best_node = 0;
6214 for (i = 0; i < MAX_NUMNODES; i++) {
6215 /* Start at @node */
6216 n = (node + i) % MAX_NUMNODES;
6218 if (!nr_cpus_node(n))
6221 /* Skip already used nodes */
6222 if (test_bit(n, used_nodes))
6225 /* Simple min distance search */
6226 val = node_distance(node, n);
6228 if (val < min_val) {
6234 set_bit(best_node, used_nodes);
6239 * sched_domain_node_span - get a cpumask for a node's sched_domain
6240 * @node: node whose cpumask we're constructing
6241 * @size: number of nodes to include in this span
6243 * Given a node, construct a good cpumask for its sched_domain to span. It
6244 * should be one that prevents unnecessary balancing, but also spreads tasks
6247 static cpumask_t sched_domain_node_span(int node)
6249 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
6250 cpumask_t span, nodemask;
6254 bitmap_zero(used_nodes, MAX_NUMNODES);
6256 nodemask = node_to_cpumask(node);
6257 cpus_or(span, span, nodemask);
6258 set_bit(node, used_nodes);
6260 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6261 int next_node = find_next_best_node(node, used_nodes);
6263 nodemask = node_to_cpumask(next_node);
6264 cpus_or(span, span, nodemask);
6271 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6274 * SMT sched-domains:
6276 #ifdef CONFIG_SCHED_SMT
6277 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6278 static struct sched_group sched_group_cpus[NR_CPUS];
6280 static int cpu_to_cpu_group(int cpu)
6287 * multi-core sched-domains:
6289 #ifdef CONFIG_SCHED_MC
6290 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6291 static struct sched_group *sched_group_core_bycpu[NR_CPUS];
6294 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6295 static int cpu_to_core_group(int cpu)
6297 return first_cpu(cpu_sibling_map[cpu]);
6299 #elif defined(CONFIG_SCHED_MC)
6300 static int cpu_to_core_group(int cpu)
6306 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6307 static struct sched_group *sched_group_phys_bycpu[NR_CPUS];
6309 static int cpu_to_phys_group(int cpu)
6311 #ifdef CONFIG_SCHED_MC
6312 cpumask_t mask = cpu_coregroup_map(cpu);
6313 return first_cpu(mask);
6314 #elif defined(CONFIG_SCHED_SMT)
6315 return first_cpu(cpu_sibling_map[cpu]);
6323 * The init_sched_build_groups can't handle what we want to do with node
6324 * groups, so roll our own. Now each node has its own list of groups which
6325 * gets dynamically allocated.
6327 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6328 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
6330 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6331 static struct sched_group *sched_group_allnodes_bycpu[NR_CPUS];
6333 static int cpu_to_allnodes_group(int cpu)
6335 return cpu_to_node(cpu);
6337 static void init_numa_sched_groups_power(struct sched_group *group_head)
6339 struct sched_group *sg = group_head;
6345 for_each_cpu_mask(j, sg->cpumask) {
6346 struct sched_domain *sd;
6348 sd = &per_cpu(phys_domains, j);
6349 if (j != first_cpu(sd->groups->cpumask)) {
6351 * Only add "power" once for each
6357 sg->cpu_power += sd->groups->cpu_power;
6360 if (sg != group_head)
6365 /* Free memory allocated for various sched_group structures */
6366 static void free_sched_groups(const cpumask_t *cpu_map)
6372 for_each_cpu_mask(cpu, *cpu_map) {
6373 struct sched_group *sched_group_allnodes
6374 = sched_group_allnodes_bycpu[cpu];
6375 struct sched_group **sched_group_nodes
6376 = sched_group_nodes_bycpu[cpu];
6378 if (sched_group_allnodes) {
6379 kfree(sched_group_allnodes);
6380 sched_group_allnodes_bycpu[cpu] = NULL;
6383 if (!sched_group_nodes)
6386 for (i = 0; i < MAX_NUMNODES; i++) {
6387 cpumask_t nodemask = node_to_cpumask(i);
6388 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6390 cpus_and(nodemask, nodemask, *cpu_map);
6391 if (cpus_empty(nodemask))
6401 if (oldsg != sched_group_nodes[i])
6404 kfree(sched_group_nodes);
6405 sched_group_nodes_bycpu[cpu] = NULL;
6408 for_each_cpu_mask(cpu, *cpu_map) {
6409 if (sched_group_phys_bycpu[cpu]) {
6410 kfree(sched_group_phys_bycpu[cpu]);
6411 sched_group_phys_bycpu[cpu] = NULL;
6413 #ifdef CONFIG_SCHED_MC
6414 if (sched_group_core_bycpu[cpu]) {
6415 kfree(sched_group_core_bycpu[cpu]);
6416 sched_group_core_bycpu[cpu] = NULL;
6423 * Build sched domains for a given set of cpus and attach the sched domains
6424 * to the individual cpus
6426 static int build_sched_domains(const cpumask_t *cpu_map)
6429 struct sched_group *sched_group_phys = NULL;
6430 #ifdef CONFIG_SCHED_MC
6431 struct sched_group *sched_group_core = NULL;
6434 struct sched_group **sched_group_nodes = NULL;
6435 struct sched_group *sched_group_allnodes = NULL;
6438 * Allocate the per-node list of sched groups
6440 sched_group_nodes = kzalloc(sizeof(struct sched_group*)*MAX_NUMNODES,
6442 if (!sched_group_nodes) {
6443 printk(KERN_WARNING "Can not alloc sched group node list\n");
6446 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6450 * Set up domains for cpus specified by the cpu_map.
6452 for_each_cpu_mask(i, *cpu_map) {
6454 struct sched_domain *sd = NULL, *p;
6455 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6457 cpus_and(nodemask, nodemask, *cpu_map);
6460 if (cpus_weight(*cpu_map)
6461 > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6462 if (!sched_group_allnodes) {
6463 sched_group_allnodes
6464 = kmalloc(sizeof(struct sched_group)
6467 if (!sched_group_allnodes) {
6469 "Can not alloc allnodes sched group\n");
6472 sched_group_allnodes_bycpu[i]
6473 = sched_group_allnodes;
6475 sd = &per_cpu(allnodes_domains, i);
6476 *sd = SD_ALLNODES_INIT;
6477 sd->span = *cpu_map;
6478 group = cpu_to_allnodes_group(i);
6479 sd->groups = &sched_group_allnodes[group];
6484 sd = &per_cpu(node_domains, i);
6486 sd->span = sched_domain_node_span(cpu_to_node(i));
6488 cpus_and(sd->span, sd->span, *cpu_map);
6491 if (!sched_group_phys) {
6493 = kmalloc(sizeof(struct sched_group) * NR_CPUS,
6495 if (!sched_group_phys) {
6496 printk (KERN_WARNING "Can not alloc phys sched"
6500 sched_group_phys_bycpu[i] = sched_group_phys;
6504 sd = &per_cpu(phys_domains, i);
6505 group = cpu_to_phys_group(i);
6507 sd->span = nodemask;
6509 sd->groups = &sched_group_phys[group];
6511 #ifdef CONFIG_SCHED_MC
6512 if (!sched_group_core) {
6514 = kmalloc(sizeof(struct sched_group) * NR_CPUS,
6516 if (!sched_group_core) {
6517 printk (KERN_WARNING "Can not alloc core sched"
6521 sched_group_core_bycpu[i] = sched_group_core;
6525 sd = &per_cpu(core_domains, i);
6526 group = cpu_to_core_group(i);
6528 sd->span = cpu_coregroup_map(i);
6529 cpus_and(sd->span, sd->span, *cpu_map);
6531 sd->groups = &sched_group_core[group];
6534 #ifdef CONFIG_SCHED_SMT
6536 sd = &per_cpu(cpu_domains, i);
6537 group = cpu_to_cpu_group(i);
6538 *sd = SD_SIBLING_INIT;
6539 sd->span = cpu_sibling_map[i];
6540 cpus_and(sd->span, sd->span, *cpu_map);
6542 sd->groups = &sched_group_cpus[group];
6546 #ifdef CONFIG_SCHED_SMT
6547 /* Set up CPU (sibling) groups */
6548 for_each_cpu_mask(i, *cpu_map) {
6549 cpumask_t this_sibling_map = cpu_sibling_map[i];
6550 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6551 if (i != first_cpu(this_sibling_map))
6554 init_sched_build_groups(sched_group_cpus, this_sibling_map,
6559 #ifdef CONFIG_SCHED_MC
6560 /* Set up multi-core groups */
6561 for_each_cpu_mask(i, *cpu_map) {
6562 cpumask_t this_core_map = cpu_coregroup_map(i);
6563 cpus_and(this_core_map, this_core_map, *cpu_map);
6564 if (i != first_cpu(this_core_map))
6566 init_sched_build_groups(sched_group_core, this_core_map,
6567 &cpu_to_core_group);
6572 /* Set up physical groups */
6573 for (i = 0; i < MAX_NUMNODES; i++) {
6574 cpumask_t nodemask = node_to_cpumask(i);
6576 cpus_and(nodemask, nodemask, *cpu_map);
6577 if (cpus_empty(nodemask))
6580 init_sched_build_groups(sched_group_phys, nodemask,
6581 &cpu_to_phys_group);
6585 /* Set up node groups */
6586 if (sched_group_allnodes)
6587 init_sched_build_groups(sched_group_allnodes, *cpu_map,
6588 &cpu_to_allnodes_group);
6590 for (i = 0; i < MAX_NUMNODES; i++) {
6591 /* Set up node groups */
6592 struct sched_group *sg, *prev;
6593 cpumask_t nodemask = node_to_cpumask(i);
6594 cpumask_t domainspan;
6595 cpumask_t covered = CPU_MASK_NONE;
6598 cpus_and(nodemask, nodemask, *cpu_map);
6599 if (cpus_empty(nodemask)) {
6600 sched_group_nodes[i] = NULL;
6604 domainspan = sched_domain_node_span(i);
6605 cpus_and(domainspan, domainspan, *cpu_map);
6607 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6609 printk(KERN_WARNING "Can not alloc domain group for "
6613 sched_group_nodes[i] = sg;
6614 for_each_cpu_mask(j, nodemask) {
6615 struct sched_domain *sd;
6616 sd = &per_cpu(node_domains, j);
6620 sg->cpumask = nodemask;
6622 cpus_or(covered, covered, nodemask);
6625 for (j = 0; j < MAX_NUMNODES; j++) {
6626 cpumask_t tmp, notcovered;
6627 int n = (i + j) % MAX_NUMNODES;
6629 cpus_complement(notcovered, covered);
6630 cpus_and(tmp, notcovered, *cpu_map);
6631 cpus_and(tmp, tmp, domainspan);
6632 if (cpus_empty(tmp))
6635 nodemask = node_to_cpumask(n);
6636 cpus_and(tmp, tmp, nodemask);
6637 if (cpus_empty(tmp))
6640 sg = kmalloc_node(sizeof(struct sched_group),
6644 "Can not alloc domain group for node %d\n", j);
6649 sg->next = prev->next;
6650 cpus_or(covered, covered, tmp);
6657 /* Calculate CPU power for physical packages and nodes */
6658 #ifdef CONFIG_SCHED_SMT
6659 for_each_cpu_mask(i, *cpu_map) {
6660 struct sched_domain *sd;
6661 sd = &per_cpu(cpu_domains, i);
6662 sd->groups->cpu_power = SCHED_LOAD_SCALE;
6665 #ifdef CONFIG_SCHED_MC
6666 for_each_cpu_mask(i, *cpu_map) {
6668 struct sched_domain *sd;
6669 sd = &per_cpu(core_domains, i);
6670 if (sched_smt_power_savings)
6671 power = SCHED_LOAD_SCALE * cpus_weight(sd->groups->cpumask);
6673 power = SCHED_LOAD_SCALE + (cpus_weight(sd->groups->cpumask)-1)
6674 * SCHED_LOAD_SCALE / 10;
6675 sd->groups->cpu_power = power;
6679 for_each_cpu_mask(i, *cpu_map) {
6680 struct sched_domain *sd;
6681 #ifdef CONFIG_SCHED_MC
6682 sd = &per_cpu(phys_domains, i);
6683 if (i != first_cpu(sd->groups->cpumask))
6686 sd->groups->cpu_power = 0;
6687 if (sched_mc_power_savings || sched_smt_power_savings) {
6690 for_each_cpu_mask(j, sd->groups->cpumask) {
6691 struct sched_domain *sd1;
6692 sd1 = &per_cpu(core_domains, j);
6694 * for each core we will add once
6695 * to the group in physical domain
6697 if (j != first_cpu(sd1->groups->cpumask))
6700 if (sched_smt_power_savings)
6701 sd->groups->cpu_power += sd1->groups->cpu_power;
6703 sd->groups->cpu_power += SCHED_LOAD_SCALE;
6707 * This has to be < 2 * SCHED_LOAD_SCALE
6708 * Lets keep it SCHED_LOAD_SCALE, so that
6709 * while calculating NUMA group's cpu_power
6711 * numa_group->cpu_power += phys_group->cpu_power;
6713 * See "only add power once for each physical pkg"
6716 sd->groups->cpu_power = SCHED_LOAD_SCALE;
6719 sd = &per_cpu(phys_domains, i);
6720 if (sched_smt_power_savings)
6721 power = SCHED_LOAD_SCALE * cpus_weight(sd->groups->cpumask);
6723 power = SCHED_LOAD_SCALE;
6724 sd->groups->cpu_power = power;
6729 for (i = 0; i < MAX_NUMNODES; i++)
6730 init_numa_sched_groups_power(sched_group_nodes[i]);
6732 if (sched_group_allnodes) {
6733 int group = cpu_to_allnodes_group(first_cpu(*cpu_map));
6734 struct sched_group *sg = &sched_group_allnodes[group];
6736 init_numa_sched_groups_power(sg);
6740 /* Attach the domains */
6741 for_each_cpu_mask(i, *cpu_map) {
6742 struct sched_domain *sd;
6743 #ifdef CONFIG_SCHED_SMT
6744 sd = &per_cpu(cpu_domains, i);
6745 #elif defined(CONFIG_SCHED_MC)
6746 sd = &per_cpu(core_domains, i);
6748 sd = &per_cpu(phys_domains, i);
6750 cpu_attach_domain(sd, i);
6753 * Tune cache-hot values:
6755 calibrate_migration_costs(cpu_map);
6760 free_sched_groups(cpu_map);
6764 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6766 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6768 cpumask_t cpu_default_map;
6772 * Setup mask for cpus without special case scheduling requirements.
6773 * For now this just excludes isolated cpus, but could be used to
6774 * exclude other special cases in the future.
6776 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6778 err = build_sched_domains(&cpu_default_map);
6783 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6785 free_sched_groups(cpu_map);
6789 * Detach sched domains from a group of cpus specified in cpu_map
6790 * These cpus will now be attached to the NULL domain
6792 static void detach_destroy_domains(const cpumask_t *cpu_map)
6796 for_each_cpu_mask(i, *cpu_map)
6797 cpu_attach_domain(NULL, i);
6798 synchronize_sched();
6799 arch_destroy_sched_domains(cpu_map);
6803 * Partition sched domains as specified by the cpumasks below.
6804 * This attaches all cpus from the cpumasks to the NULL domain,
6805 * waits for a RCU quiescent period, recalculates sched
6806 * domain information and then attaches them back to the
6807 * correct sched domains
6808 * Call with hotplug lock held
6810 int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6812 cpumask_t change_map;
6815 cpus_and(*partition1, *partition1, cpu_online_map);
6816 cpus_and(*partition2, *partition2, cpu_online_map);
6817 cpus_or(change_map, *partition1, *partition2);
6819 /* Detach sched domains from all of the affected cpus */
6820 detach_destroy_domains(&change_map);
6821 if (!cpus_empty(*partition1))
6822 err = build_sched_domains(partition1);
6823 if (!err && !cpus_empty(*partition2))
6824 err = build_sched_domains(partition2);
6829 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6830 int arch_reinit_sched_domains(void)
6835 detach_destroy_domains(&cpu_online_map);
6836 err = arch_init_sched_domains(&cpu_online_map);
6837 unlock_cpu_hotplug();
6842 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6846 if (buf[0] != '0' && buf[0] != '1')
6850 sched_smt_power_savings = (buf[0] == '1');
6852 sched_mc_power_savings = (buf[0] == '1');
6854 ret = arch_reinit_sched_domains();
6856 return ret ? ret : count;
6859 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6863 #ifdef CONFIG_SCHED_SMT
6865 err = sysfs_create_file(&cls->kset.kobj,
6866 &attr_sched_smt_power_savings.attr);
6868 #ifdef CONFIG_SCHED_MC
6869 if (!err && mc_capable())
6870 err = sysfs_create_file(&cls->kset.kobj,
6871 &attr_sched_mc_power_savings.attr);
6877 #ifdef CONFIG_SCHED_MC
6878 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6880 return sprintf(page, "%u\n", sched_mc_power_savings);
6882 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6883 const char *buf, size_t count)
6885 return sched_power_savings_store(buf, count, 0);
6887 SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6888 sched_mc_power_savings_store);
6891 #ifdef CONFIG_SCHED_SMT
6892 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6894 return sprintf(page, "%u\n", sched_smt_power_savings);
6896 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6897 const char *buf, size_t count)
6899 return sched_power_savings_store(buf, count, 1);
6901 SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6902 sched_smt_power_savings_store);
6906 #ifdef CONFIG_HOTPLUG_CPU
6908 * Force a reinitialization of the sched domains hierarchy. The domains
6909 * and groups cannot be updated in place without racing with the balancing
6910 * code, so we temporarily attach all running cpus to the NULL domain
6911 * which will prevent rebalancing while the sched domains are recalculated.
6913 static int update_sched_domains(struct notifier_block *nfb,
6914 unsigned long action, void *hcpu)
6917 case CPU_UP_PREPARE:
6918 case CPU_DOWN_PREPARE:
6919 detach_destroy_domains(&cpu_online_map);
6922 case CPU_UP_CANCELED:
6923 case CPU_DOWN_FAILED:
6927 * Fall through and re-initialise the domains.
6934 /* The hotplug lock is already held by cpu_up/cpu_down */
6935 arch_init_sched_domains(&cpu_online_map);
6941 void __init sched_init_smp(void)
6944 arch_init_sched_domains(&cpu_online_map);
6945 unlock_cpu_hotplug();
6946 /* XXX: Theoretical race here - CPU may be hotplugged now */
6947 hotcpu_notifier(update_sched_domains, 0);
6950 void __init sched_init_smp(void)
6953 #endif /* CONFIG_SMP */
6955 int in_sched_functions(unsigned long addr)
6957 /* Linker adds these: start and end of __sched functions */
6958 extern char __sched_text_start[], __sched_text_end[];
6960 return in_lock_functions(addr) ||
6961 (addr >= (unsigned long)__sched_text_start
6962 && addr < (unsigned long)__sched_text_end);
6965 void __init sched_init(void)
6969 for_each_possible_cpu(i) {
6970 struct prio_array *array;
6974 spin_lock_init(&rq->lock);
6975 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6977 rq->active = rq->arrays;
6978 rq->expired = rq->arrays + 1;
6979 rq->best_expired_prio = MAX_PRIO;
6983 for (j = 1; j < 3; j++)
6984 rq->cpu_load[j] = 0;
6985 rq->active_balance = 0;
6988 rq->migration_thread = NULL;
6989 INIT_LIST_HEAD(&rq->migration_queue);
6991 atomic_set(&rq->nr_iowait, 0);
6992 #ifdef CONFIG_VSERVER_HARDCPU
6993 INIT_LIST_HEAD(&rq->hold_queue);
6996 for (j = 0; j < 2; j++) {
6997 array = rq->arrays + j;
6998 for (k = 0; k < MAX_PRIO; k++) {
6999 INIT_LIST_HEAD(array->queue + k);
7000 __clear_bit(k, array->bitmap);
7002 // delimiter for bitsearch
7003 __set_bit(MAX_PRIO, array->bitmap);
7007 set_load_weight(&init_task);
7009 #ifdef CONFIG_RT_MUTEXES
7010 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
7014 * The boot idle thread does lazy MMU switching as well:
7016 atomic_inc(&init_mm.mm_count);
7017 enter_lazy_tlb(&init_mm, current);
7020 * Make us the idle thread. Technically, schedule() should not be
7021 * called from this thread, however somewhere below it might be,
7022 * but because we are the idle thread, we just pick up running again
7023 * when this runqueue becomes "idle".
7025 init_idle(current, smp_processor_id());
7028 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7029 void __might_sleep(char *file, int line)
7032 static unsigned long prev_jiffy; /* ratelimiting */
7034 if ((in_atomic() || irqs_disabled()) &&
7035 system_state == SYSTEM_RUNNING && !oops_in_progress) {
7036 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7038 prev_jiffy = jiffies;
7039 printk(KERN_ERR "BUG: sleeping function called from invalid"
7040 " context at %s:%d\n", file, line);
7041 printk("in_atomic():%d, irqs_disabled():%d\n",
7042 in_atomic(), irqs_disabled());
7047 EXPORT_SYMBOL(__might_sleep);
7050 #ifdef CONFIG_MAGIC_SYSRQ
7051 void normalize_rt_tasks(void)
7053 struct prio_array *array;
7054 struct task_struct *p;
7055 unsigned long flags;
7058 read_lock_irq(&tasklist_lock);
7059 for_each_process(p) {
7063 spin_lock_irqsave(&p->pi_lock, flags);
7064 rq = __task_rq_lock(p);
7068 deactivate_task(p, task_rq(p));
7069 __setscheduler(p, SCHED_NORMAL, 0);
7071 vx_activate_task(p);
7072 __activate_task(p, task_rq(p));
7073 resched_task(rq->curr);
7076 __task_rq_unlock(rq);
7077 spin_unlock_irqrestore(&p->pi_lock, flags);
7079 read_unlock_irq(&tasklist_lock);
7082 #endif /* CONFIG_MAGIC_SYSRQ */
7086 * These functions are only useful for the IA64 MCA handling.
7088 * They can only be called when the whole system has been
7089 * stopped - every CPU needs to be quiescent, and no scheduling
7090 * activity can take place. Using them for anything else would
7091 * be a serious bug, and as a result, they aren't even visible
7092 * under any other configuration.
7096 * curr_task - return the current task for a given cpu.
7097 * @cpu: the processor in question.
7099 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7101 struct task_struct *curr_task(int cpu)
7103 return cpu_curr(cpu);
7107 * set_curr_task - set the current task for a given cpu.
7108 * @cpu: the processor in question.
7109 * @p: the task pointer to set.
7111 * Description: This function must only be used when non-maskable interrupts
7112 * are serviced on a separate stack. It allows the architecture to switch the
7113 * notion of the current task on a cpu in a non-blocking manner. This function
7114 * must be called with all CPU's synchronized, and interrupts disabled, the
7115 * and caller must save the original value of the current task (see
7116 * curr_task() above) and restore that value before reenabling interrupts and
7117 * re-starting the system.
7119 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7121 void set_curr_task(int cpu, struct task_struct *p)