USB: Sierra Wireless auto set D0
[linux-2.6/mini2440.git] / kernel / sched.c
blobcca93cc0dd7d8ba7272b1ea4782e87c97bfe1a09
1 /*
2 * kernel/sched.c
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
11 * by Andrea Arcangeli
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
21 #include <linux/mm.h>
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/freezer.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/tsacct_kern.h>
53 #include <linux/kprobes.h>
54 #include <linux/delayacct.h>
55 #include <asm/tlb.h>
57 #include <asm/unistd.h>
60 * Convert user-nice values [ -20 ... 0 ... 19 ]
61 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
62 * and back.
64 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
65 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
66 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
69 * 'User priority' is the nice value converted to something we
70 * can work with better when scaling various scheduler parameters,
71 * it's a [ 0 ... 39 ] range.
73 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
74 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
75 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
78 * Some helpers for converting nanosecond timing to jiffy resolution
80 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
81 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
84 * These are the 'tuning knobs' of the scheduler:
86 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
87 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
88 * Timeslices get refilled after they expire.
90 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
91 #define DEF_TIMESLICE (100 * HZ / 1000)
92 #define ON_RUNQUEUE_WEIGHT 30
93 #define CHILD_PENALTY 95
94 #define PARENT_PENALTY 100
95 #define EXIT_WEIGHT 3
96 #define PRIO_BONUS_RATIO 25
97 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
98 #define INTERACTIVE_DELTA 2
99 #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
100 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
101 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
104 * If a task is 'interactive' then we reinsert it in the active
105 * array after it has expired its current timeslice. (it will not
106 * continue to run immediately, it will still roundrobin with
107 * other interactive tasks.)
109 * This part scales the interactivity limit depending on niceness.
111 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
112 * Here are a few examples of different nice levels:
114 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
115 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
116 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
117 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
118 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
120 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
121 * priority range a task can explore, a value of '1' means the
122 * task is rated interactive.)
124 * Ie. nice +19 tasks can never get 'interactive' enough to be
125 * reinserted into the active array. And only heavily CPU-hog nice -20
126 * tasks will be expired. Default nice 0 tasks are somewhere between,
127 * it takes some effort for them to get interactive, but it's not
128 * too hard.
131 #define CURRENT_BONUS(p) \
132 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
133 MAX_SLEEP_AVG)
135 #define GRANULARITY (10 * HZ / 1000 ? : 1)
137 #ifdef CONFIG_SMP
138 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
139 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
140 num_online_cpus())
141 #else
142 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
143 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
144 #endif
146 #define SCALE(v1,v1_max,v2_max) \
147 (v1) * (v2_max) / (v1_max)
149 #define DELTA(p) \
150 (SCALE(TASK_NICE(p) + 20, 40, MAX_BONUS) - 20 * MAX_BONUS / 40 + \
151 INTERACTIVE_DELTA)
153 #define TASK_INTERACTIVE(p) \
154 ((p)->prio <= (p)->static_prio - DELTA(p))
156 #define INTERACTIVE_SLEEP(p) \
157 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
158 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
160 #define TASK_PREEMPTS_CURR(p, rq) \
161 ((p)->prio < (rq)->curr->prio)
163 #define SCALE_PRIO(x, prio) \
164 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
166 static unsigned int static_prio_timeslice(int static_prio)
168 if (static_prio < NICE_TO_PRIO(0))
169 return SCALE_PRIO(DEF_TIMESLICE * 4, static_prio);
170 else
171 return SCALE_PRIO(DEF_TIMESLICE, static_prio);
175 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
176 * to time slice values: [800ms ... 100ms ... 5ms]
178 * The higher a thread's priority, the bigger timeslices
179 * it gets during one round of execution. But even the lowest
180 * priority thread gets MIN_TIMESLICE worth of execution time.
183 static inline unsigned int task_timeslice(struct task_struct *p)
185 return static_prio_timeslice(p->static_prio);
189 * These are the runqueue data structures:
192 struct prio_array {
193 unsigned int nr_active;
194 DECLARE_BITMAP(bitmap, MAX_PRIO+1); /* include 1 bit for delimiter */
195 struct list_head queue[MAX_PRIO];
199 * This is the main, per-CPU runqueue data structure.
201 * Locking rule: those places that want to lock multiple runqueues
202 * (such as the load balancing or the thread migration code), lock
203 * acquire operations must be ordered by ascending &runqueue.
205 struct rq {
206 spinlock_t lock;
209 * nr_running and cpu_load should be in the same cacheline because
210 * remote CPUs use both these fields when doing load calculation.
212 unsigned long nr_running;
213 unsigned long raw_weighted_load;
214 #ifdef CONFIG_SMP
215 unsigned long cpu_load[3];
216 #endif
217 unsigned long long nr_switches;
220 * This is part of a global counter where only the total sum
221 * over all CPUs matters. A task can increase this counter on
222 * one CPU and if it got migrated afterwards it may decrease
223 * it on another CPU. Always updated under the runqueue lock:
225 unsigned long nr_uninterruptible;
227 unsigned long expired_timestamp;
228 /* Cached timestamp set by update_cpu_clock() */
229 unsigned long long most_recent_timestamp;
230 struct task_struct *curr, *idle;
231 unsigned long next_balance;
232 struct mm_struct *prev_mm;
233 struct prio_array *active, *expired, arrays[2];
234 int best_expired_prio;
235 atomic_t nr_iowait;
237 #ifdef CONFIG_SMP
238 struct sched_domain *sd;
240 /* For active balancing */
241 int active_balance;
242 int push_cpu;
243 int cpu; /* cpu of this runqueue */
245 struct task_struct *migration_thread;
246 struct list_head migration_queue;
247 #endif
249 #ifdef CONFIG_SCHEDSTATS
250 /* latency stats */
251 struct sched_info rq_sched_info;
253 /* sys_sched_yield() stats */
254 unsigned long yld_exp_empty;
255 unsigned long yld_act_empty;
256 unsigned long yld_both_empty;
257 unsigned long yld_cnt;
259 /* schedule() stats */
260 unsigned long sched_switch;
261 unsigned long sched_cnt;
262 unsigned long sched_goidle;
264 /* try_to_wake_up() stats */
265 unsigned long ttwu_cnt;
266 unsigned long ttwu_local;
267 #endif
268 struct lock_class_key rq_lock_key;
271 static DEFINE_PER_CPU(struct rq, runqueues);
273 static inline int cpu_of(struct rq *rq)
275 #ifdef CONFIG_SMP
276 return rq->cpu;
277 #else
278 return 0;
279 #endif
283 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
284 * See detach_destroy_domains: synchronize_sched for details.
286 * The domain tree of any CPU may only be accessed from within
287 * preempt-disabled sections.
289 #define for_each_domain(cpu, __sd) \
290 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
292 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
293 #define this_rq() (&__get_cpu_var(runqueues))
294 #define task_rq(p) cpu_rq(task_cpu(p))
295 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
297 #ifndef prepare_arch_switch
298 # define prepare_arch_switch(next) do { } while (0)
299 #endif
300 #ifndef finish_arch_switch
301 # define finish_arch_switch(prev) do { } while (0)
302 #endif
304 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
305 static inline int task_running(struct rq *rq, struct task_struct *p)
307 return rq->curr == p;
310 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
314 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
316 #ifdef CONFIG_DEBUG_SPINLOCK
317 /* this is a valid case when another task releases the spinlock */
318 rq->lock.owner = current;
319 #endif
321 * If we are tracking spinlock dependencies then we have to
322 * fix up the runqueue lock - which gets 'carried over' from
323 * prev into current:
325 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
327 spin_unlock_irq(&rq->lock);
330 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
331 static inline int task_running(struct rq *rq, struct task_struct *p)
333 #ifdef CONFIG_SMP
334 return p->oncpu;
335 #else
336 return rq->curr == p;
337 #endif
340 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
342 #ifdef CONFIG_SMP
344 * We can optimise this out completely for !SMP, because the
345 * SMP rebalancing from interrupt is the only thing that cares
346 * here.
348 next->oncpu = 1;
349 #endif
350 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
351 spin_unlock_irq(&rq->lock);
352 #else
353 spin_unlock(&rq->lock);
354 #endif
357 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
359 #ifdef CONFIG_SMP
361 * After ->oncpu is cleared, the task can be moved to a different CPU.
362 * We must ensure this doesn't happen until the switch is completely
363 * finished.
365 smp_wmb();
366 prev->oncpu = 0;
367 #endif
368 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
369 local_irq_enable();
370 #endif
372 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
375 * __task_rq_lock - lock the runqueue a given task resides on.
376 * Must be called interrupts disabled.
378 static inline struct rq *__task_rq_lock(struct task_struct *p)
379 __acquires(rq->lock)
381 struct rq *rq;
383 repeat_lock_task:
384 rq = task_rq(p);
385 spin_lock(&rq->lock);
386 if (unlikely(rq != task_rq(p))) {
387 spin_unlock(&rq->lock);
388 goto repeat_lock_task;
390 return rq;
394 * task_rq_lock - lock the runqueue a given task resides on and disable
395 * interrupts. Note the ordering: we can safely lookup the task_rq without
396 * explicitly disabling preemption.
398 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
399 __acquires(rq->lock)
401 struct rq *rq;
403 repeat_lock_task:
404 local_irq_save(*flags);
405 rq = task_rq(p);
406 spin_lock(&rq->lock);
407 if (unlikely(rq != task_rq(p))) {
408 spin_unlock_irqrestore(&rq->lock, *flags);
409 goto repeat_lock_task;
411 return rq;
414 static inline void __task_rq_unlock(struct rq *rq)
415 __releases(rq->lock)
417 spin_unlock(&rq->lock);
420 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
421 __releases(rq->lock)
423 spin_unlock_irqrestore(&rq->lock, *flags);
426 #ifdef CONFIG_SCHEDSTATS
428 * bump this up when changing the output format or the meaning of an existing
429 * format, so that tools can adapt (or abort)
431 #define SCHEDSTAT_VERSION 14
433 static int show_schedstat(struct seq_file *seq, void *v)
435 int cpu;
437 seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION);
438 seq_printf(seq, "timestamp %lu\n", jiffies);
439 for_each_online_cpu(cpu) {
440 struct rq *rq = cpu_rq(cpu);
441 #ifdef CONFIG_SMP
442 struct sched_domain *sd;
443 int dcnt = 0;
444 #endif
446 /* runqueue-specific stats */
447 seq_printf(seq,
448 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
449 cpu, rq->yld_both_empty,
450 rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt,
451 rq->sched_switch, rq->sched_cnt, rq->sched_goidle,
452 rq->ttwu_cnt, rq->ttwu_local,
453 rq->rq_sched_info.cpu_time,
454 rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt);
456 seq_printf(seq, "\n");
458 #ifdef CONFIG_SMP
459 /* domain-specific stats */
460 preempt_disable();
461 for_each_domain(cpu, sd) {
462 enum idle_type itype;
463 char mask_str[NR_CPUS];
465 cpumask_scnprintf(mask_str, NR_CPUS, sd->span);
466 seq_printf(seq, "domain%d %s", dcnt++, mask_str);
467 for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES;
468 itype++) {
469 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu "
470 "%lu",
471 sd->lb_cnt[itype],
472 sd->lb_balanced[itype],
473 sd->lb_failed[itype],
474 sd->lb_imbalance[itype],
475 sd->lb_gained[itype],
476 sd->lb_hot_gained[itype],
477 sd->lb_nobusyq[itype],
478 sd->lb_nobusyg[itype]);
480 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu"
481 " %lu %lu %lu\n",
482 sd->alb_cnt, sd->alb_failed, sd->alb_pushed,
483 sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed,
484 sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed,
485 sd->ttwu_wake_remote, sd->ttwu_move_affine,
486 sd->ttwu_move_balance);
488 preempt_enable();
489 #endif
491 return 0;
494 static int schedstat_open(struct inode *inode, struct file *file)
496 unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32);
497 char *buf = kmalloc(size, GFP_KERNEL);
498 struct seq_file *m;
499 int res;
501 if (!buf)
502 return -ENOMEM;
503 res = single_open(file, show_schedstat, NULL);
504 if (!res) {
505 m = file->private_data;
506 m->buf = buf;
507 m->size = size;
508 } else
509 kfree(buf);
510 return res;
513 const struct file_operations proc_schedstat_operations = {
514 .open = schedstat_open,
515 .read = seq_read,
516 .llseek = seq_lseek,
517 .release = single_release,
521 * Expects runqueue lock to be held for atomicity of update
523 static inline void
524 rq_sched_info_arrive(struct rq *rq, unsigned long delta_jiffies)
526 if (rq) {
527 rq->rq_sched_info.run_delay += delta_jiffies;
528 rq->rq_sched_info.pcnt++;
533 * Expects runqueue lock to be held for atomicity of update
535 static inline void
536 rq_sched_info_depart(struct rq *rq, unsigned long delta_jiffies)
538 if (rq)
539 rq->rq_sched_info.cpu_time += delta_jiffies;
541 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
542 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
543 #else /* !CONFIG_SCHEDSTATS */
544 static inline void
545 rq_sched_info_arrive(struct rq *rq, unsigned long delta_jiffies)
547 static inline void
548 rq_sched_info_depart(struct rq *rq, unsigned long delta_jiffies)
550 # define schedstat_inc(rq, field) do { } while (0)
551 # define schedstat_add(rq, field, amt) do { } while (0)
552 #endif
555 * this_rq_lock - lock this runqueue and disable interrupts.
557 static inline struct rq *this_rq_lock(void)
558 __acquires(rq->lock)
560 struct rq *rq;
562 local_irq_disable();
563 rq = this_rq();
564 spin_lock(&rq->lock);
566 return rq;
569 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
571 * Called when a process is dequeued from the active array and given
572 * the cpu. We should note that with the exception of interactive
573 * tasks, the expired queue will become the active queue after the active
574 * queue is empty, without explicitly dequeuing and requeuing tasks in the
575 * expired queue. (Interactive tasks may be requeued directly to the
576 * active queue, thus delaying tasks in the expired queue from running;
577 * see scheduler_tick()).
579 * This function is only called from sched_info_arrive(), rather than
580 * dequeue_task(). Even though a task may be queued and dequeued multiple
581 * times as it is shuffled about, we're really interested in knowing how
582 * long it was from the *first* time it was queued to the time that it
583 * finally hit a cpu.
585 static inline void sched_info_dequeued(struct task_struct *t)
587 t->sched_info.last_queued = 0;
591 * Called when a task finally hits the cpu. We can now calculate how
592 * long it was waiting to run. We also note when it began so that we
593 * can keep stats on how long its timeslice is.
595 static void sched_info_arrive(struct task_struct *t)
597 unsigned long now = jiffies, delta_jiffies = 0;
599 if (t->sched_info.last_queued)
600 delta_jiffies = now - t->sched_info.last_queued;
601 sched_info_dequeued(t);
602 t->sched_info.run_delay += delta_jiffies;
603 t->sched_info.last_arrival = now;
604 t->sched_info.pcnt++;
606 rq_sched_info_arrive(task_rq(t), delta_jiffies);
610 * Called when a process is queued into either the active or expired
611 * array. The time is noted and later used to determine how long we
612 * had to wait for us to reach the cpu. Since the expired queue will
613 * become the active queue after active queue is empty, without dequeuing
614 * and requeuing any tasks, we are interested in queuing to either. It
615 * is unusual but not impossible for tasks to be dequeued and immediately
616 * requeued in the same or another array: this can happen in sched_yield(),
617 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
618 * to runqueue.
620 * This function is only called from enqueue_task(), but also only updates
621 * the timestamp if it is already not set. It's assumed that
622 * sched_info_dequeued() will clear that stamp when appropriate.
624 static inline void sched_info_queued(struct task_struct *t)
626 if (unlikely(sched_info_on()))
627 if (!t->sched_info.last_queued)
628 t->sched_info.last_queued = jiffies;
632 * Called when a process ceases being the active-running process, either
633 * voluntarily or involuntarily. Now we can calculate how long we ran.
635 static inline void sched_info_depart(struct task_struct *t)
637 unsigned long delta_jiffies = jiffies - t->sched_info.last_arrival;
639 t->sched_info.cpu_time += delta_jiffies;
640 rq_sched_info_depart(task_rq(t), delta_jiffies);
644 * Called when tasks are switched involuntarily due, typically, to expiring
645 * their time slice. (This may also be called when switching to or from
646 * the idle task.) We are only called when prev != next.
648 static inline void
649 __sched_info_switch(struct task_struct *prev, struct task_struct *next)
651 struct rq *rq = task_rq(prev);
654 * prev now departs the cpu. It's not interesting to record
655 * stats about how efficient we were at scheduling the idle
656 * process, however.
658 if (prev != rq->idle)
659 sched_info_depart(prev);
661 if (next != rq->idle)
662 sched_info_arrive(next);
664 static inline void
665 sched_info_switch(struct task_struct *prev, struct task_struct *next)
667 if (unlikely(sched_info_on()))
668 __sched_info_switch(prev, next);
670 #else
671 #define sched_info_queued(t) do { } while (0)
672 #define sched_info_switch(t, next) do { } while (0)
673 #endif /* CONFIG_SCHEDSTATS || CONFIG_TASK_DELAY_ACCT */
676 * Adding/removing a task to/from a priority array:
678 static void dequeue_task(struct task_struct *p, struct prio_array *array)
680 array->nr_active--;
681 list_del(&p->run_list);
682 if (list_empty(array->queue + p->prio))
683 __clear_bit(p->prio, array->bitmap);
686 static void enqueue_task(struct task_struct *p, struct prio_array *array)
688 sched_info_queued(p);
689 list_add_tail(&p->run_list, array->queue + p->prio);
690 __set_bit(p->prio, array->bitmap);
691 array->nr_active++;
692 p->array = array;
696 * Put task to the end of the run list without the overhead of dequeue
697 * followed by enqueue.
699 static void requeue_task(struct task_struct *p, struct prio_array *array)
701 list_move_tail(&p->run_list, array->queue + p->prio);
704 static inline void
705 enqueue_task_head(struct task_struct *p, struct prio_array *array)
707 list_add(&p->run_list, array->queue + p->prio);
708 __set_bit(p->prio, array->bitmap);
709 array->nr_active++;
710 p->array = array;
714 * __normal_prio - return the priority that is based on the static
715 * priority but is modified by bonuses/penalties.
717 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
718 * into the -5 ... 0 ... +5 bonus/penalty range.
720 * We use 25% of the full 0...39 priority range so that:
722 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
723 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
725 * Both properties are important to certain workloads.
728 static inline int __normal_prio(struct task_struct *p)
730 int bonus, prio;
732 bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
734 prio = p->static_prio - bonus;
735 if (prio < MAX_RT_PRIO)
736 prio = MAX_RT_PRIO;
737 if (prio > MAX_PRIO-1)
738 prio = MAX_PRIO-1;
739 return prio;
743 * To aid in avoiding the subversion of "niceness" due to uneven distribution
744 * of tasks with abnormal "nice" values across CPUs the contribution that
745 * each task makes to its run queue's load is weighted according to its
746 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
747 * scaled version of the new time slice allocation that they receive on time
748 * slice expiry etc.
752 * Assume: static_prio_timeslice(NICE_TO_PRIO(0)) == DEF_TIMESLICE
753 * If static_prio_timeslice() is ever changed to break this assumption then
754 * this code will need modification
756 #define TIME_SLICE_NICE_ZERO DEF_TIMESLICE
757 #define LOAD_WEIGHT(lp) \
758 (((lp) * SCHED_LOAD_SCALE) / TIME_SLICE_NICE_ZERO)
759 #define PRIO_TO_LOAD_WEIGHT(prio) \
760 LOAD_WEIGHT(static_prio_timeslice(prio))
761 #define RTPRIO_TO_LOAD_WEIGHT(rp) \
762 (PRIO_TO_LOAD_WEIGHT(MAX_RT_PRIO) + LOAD_WEIGHT(rp))
764 static void set_load_weight(struct task_struct *p)
766 if (has_rt_policy(p)) {
767 #ifdef CONFIG_SMP
768 if (p == task_rq(p)->migration_thread)
770 * The migration thread does the actual balancing.
771 * Giving its load any weight will skew balancing
772 * adversely.
774 p->load_weight = 0;
775 else
776 #endif
777 p->load_weight = RTPRIO_TO_LOAD_WEIGHT(p->rt_priority);
778 } else
779 p->load_weight = PRIO_TO_LOAD_WEIGHT(p->static_prio);
782 static inline void
783 inc_raw_weighted_load(struct rq *rq, const struct task_struct *p)
785 rq->raw_weighted_load += p->load_weight;
788 static inline void
789 dec_raw_weighted_load(struct rq *rq, const struct task_struct *p)
791 rq->raw_weighted_load -= p->load_weight;
794 static inline void inc_nr_running(struct task_struct *p, struct rq *rq)
796 rq->nr_running++;
797 inc_raw_weighted_load(rq, p);
800 static inline void dec_nr_running(struct task_struct *p, struct rq *rq)
802 rq->nr_running--;
803 dec_raw_weighted_load(rq, p);
807 * Calculate the expected normal priority: i.e. priority
808 * without taking RT-inheritance into account. Might be
809 * boosted by interactivity modifiers. Changes upon fork,
810 * setprio syscalls, and whenever the interactivity
811 * estimator recalculates.
813 static inline int normal_prio(struct task_struct *p)
815 int prio;
817 if (has_rt_policy(p))
818 prio = MAX_RT_PRIO-1 - p->rt_priority;
819 else
820 prio = __normal_prio(p);
821 return prio;
825 * Calculate the current priority, i.e. the priority
826 * taken into account by the scheduler. This value might
827 * be boosted by RT tasks, or might be boosted by
828 * interactivity modifiers. Will be RT if the task got
829 * RT-boosted. If not then it returns p->normal_prio.
831 static int effective_prio(struct task_struct *p)
833 p->normal_prio = normal_prio(p);
835 * If we are RT tasks or we were boosted to RT priority,
836 * keep the priority unchanged. Otherwise, update priority
837 * to the normal priority:
839 if (!rt_prio(p->prio))
840 return p->normal_prio;
841 return p->prio;
845 * __activate_task - move a task to the runqueue.
847 static void __activate_task(struct task_struct *p, struct rq *rq)
849 struct prio_array *target = rq->active;
851 if (batch_task(p))
852 target = rq->expired;
853 enqueue_task(p, target);
854 inc_nr_running(p, rq);
858 * __activate_idle_task - move idle task to the _front_ of runqueue.
860 static inline void __activate_idle_task(struct task_struct *p, struct rq *rq)
862 enqueue_task_head(p, rq->active);
863 inc_nr_running(p, rq);
867 * Recalculate p->normal_prio and p->prio after having slept,
868 * updating the sleep-average too:
870 static int recalc_task_prio(struct task_struct *p, unsigned long long now)
872 /* Caller must always ensure 'now >= p->timestamp' */
873 unsigned long sleep_time = now - p->timestamp;
875 if (batch_task(p))
876 sleep_time = 0;
878 if (likely(sleep_time > 0)) {
880 * This ceiling is set to the lowest priority that would allow
881 * a task to be reinserted into the active array on timeslice
882 * completion.
884 unsigned long ceiling = INTERACTIVE_SLEEP(p);
886 if (p->mm && sleep_time > ceiling && p->sleep_avg < ceiling) {
888 * Prevents user tasks from achieving best priority
889 * with one single large enough sleep.
891 p->sleep_avg = ceiling;
893 * Using INTERACTIVE_SLEEP() as a ceiling places a
894 * nice(0) task 1ms sleep away from promotion, and
895 * gives it 700ms to round-robin with no chance of
896 * being demoted. This is more than generous, so
897 * mark this sleep as non-interactive to prevent the
898 * on-runqueue bonus logic from intervening should
899 * this task not receive cpu immediately.
901 p->sleep_type = SLEEP_NONINTERACTIVE;
902 } else {
904 * Tasks waking from uninterruptible sleep are
905 * limited in their sleep_avg rise as they
906 * are likely to be waiting on I/O
908 if (p->sleep_type == SLEEP_NONINTERACTIVE && p->mm) {
909 if (p->sleep_avg >= ceiling)
910 sleep_time = 0;
911 else if (p->sleep_avg + sleep_time >=
912 ceiling) {
913 p->sleep_avg = ceiling;
914 sleep_time = 0;
919 * This code gives a bonus to interactive tasks.
921 * The boost works by updating the 'average sleep time'
922 * value here, based on ->timestamp. The more time a
923 * task spends sleeping, the higher the average gets -
924 * and the higher the priority boost gets as well.
926 p->sleep_avg += sleep_time;
929 if (p->sleep_avg > NS_MAX_SLEEP_AVG)
930 p->sleep_avg = NS_MAX_SLEEP_AVG;
933 return effective_prio(p);
937 * activate_task - move a task to the runqueue and do priority recalculation
939 * Update all the scheduling statistics stuff. (sleep average
940 * calculation, priority modifiers, etc.)
942 static void activate_task(struct task_struct *p, struct rq *rq, int local)
944 unsigned long long now;
946 if (rt_task(p))
947 goto out;
949 now = sched_clock();
950 #ifdef CONFIG_SMP
951 if (!local) {
952 /* Compensate for drifting sched_clock */
953 struct rq *this_rq = this_rq();
954 now = (now - this_rq->most_recent_timestamp)
955 + rq->most_recent_timestamp;
957 #endif
960 * Sleep time is in units of nanosecs, so shift by 20 to get a
961 * milliseconds-range estimation of the amount of time that the task
962 * spent sleeping:
964 if (unlikely(prof_on == SLEEP_PROFILING)) {
965 if (p->state == TASK_UNINTERRUPTIBLE)
966 profile_hits(SLEEP_PROFILING, (void *)get_wchan(p),
967 (now - p->timestamp) >> 20);
970 p->prio = recalc_task_prio(p, now);
973 * This checks to make sure it's not an uninterruptible task
974 * that is now waking up.
976 if (p->sleep_type == SLEEP_NORMAL) {
978 * Tasks which were woken up by interrupts (ie. hw events)
979 * are most likely of interactive nature. So we give them
980 * the credit of extending their sleep time to the period
981 * of time they spend on the runqueue, waiting for execution
982 * on a CPU, first time around:
984 if (in_interrupt())
985 p->sleep_type = SLEEP_INTERRUPTED;
986 else {
988 * Normal first-time wakeups get a credit too for
989 * on-runqueue time, but it will be weighted down:
991 p->sleep_type = SLEEP_INTERACTIVE;
994 p->timestamp = now;
995 out:
996 __activate_task(p, rq);
1000 * deactivate_task - remove a task from the runqueue.
1002 static void deactivate_task(struct task_struct *p, struct rq *rq)
1004 dec_nr_running(p, rq);
1005 dequeue_task(p, p->array);
1006 p->array = NULL;
1010 * resched_task - mark a task 'to be rescheduled now'.
1012 * On UP this means the setting of the need_resched flag, on SMP it
1013 * might also involve a cross-CPU call to trigger the scheduler on
1014 * the target CPU.
1016 #ifdef CONFIG_SMP
1018 #ifndef tsk_is_polling
1019 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1020 #endif
1022 static void resched_task(struct task_struct *p)
1024 int cpu;
1026 assert_spin_locked(&task_rq(p)->lock);
1028 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1029 return;
1031 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1033 cpu = task_cpu(p);
1034 if (cpu == smp_processor_id())
1035 return;
1037 /* NEED_RESCHED must be visible before we test polling */
1038 smp_mb();
1039 if (!tsk_is_polling(p))
1040 smp_send_reschedule(cpu);
1042 #else
1043 static inline void resched_task(struct task_struct *p)
1045 assert_spin_locked(&task_rq(p)->lock);
1046 set_tsk_need_resched(p);
1048 #endif
1051 * task_curr - is this task currently executing on a CPU?
1052 * @p: the task in question.
1054 inline int task_curr(const struct task_struct *p)
1056 return cpu_curr(task_cpu(p)) == p;
1059 /* Used instead of source_load when we know the type == 0 */
1060 unsigned long weighted_cpuload(const int cpu)
1062 return cpu_rq(cpu)->raw_weighted_load;
1065 #ifdef CONFIG_SMP
1066 struct migration_req {
1067 struct list_head list;
1069 struct task_struct *task;
1070 int dest_cpu;
1072 struct completion done;
1076 * The task's runqueue lock must be held.
1077 * Returns true if you have to wait for migration thread.
1079 static int
1080 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1082 struct rq *rq = task_rq(p);
1085 * If the task is not on a runqueue (and not running), then
1086 * it is sufficient to simply update the task's cpu field.
1088 if (!p->array && !task_running(rq, p)) {
1089 set_task_cpu(p, dest_cpu);
1090 return 0;
1093 init_completion(&req->done);
1094 req->task = p;
1095 req->dest_cpu = dest_cpu;
1096 list_add(&req->list, &rq->migration_queue);
1098 return 1;
1102 * wait_task_inactive - wait for a thread to unschedule.
1104 * The caller must ensure that the task *will* unschedule sometime soon,
1105 * else this function might spin for a *long* time. This function can't
1106 * be called with interrupts off, or it may introduce deadlock with
1107 * smp_call_function() if an IPI is sent by the same process we are
1108 * waiting to become inactive.
1110 void wait_task_inactive(struct task_struct *p)
1112 unsigned long flags;
1113 struct rq *rq;
1114 int preempted;
1116 repeat:
1117 rq = task_rq_lock(p, &flags);
1118 /* Must be off runqueue entirely, not preempted. */
1119 if (unlikely(p->array || task_running(rq, p))) {
1120 /* If it's preempted, we yield. It could be a while. */
1121 preempted = !task_running(rq, p);
1122 task_rq_unlock(rq, &flags);
1123 cpu_relax();
1124 if (preempted)
1125 yield();
1126 goto repeat;
1128 task_rq_unlock(rq, &flags);
1131 /***
1132 * kick_process - kick a running thread to enter/exit the kernel
1133 * @p: the to-be-kicked thread
1135 * Cause a process which is running on another CPU to enter
1136 * kernel-mode, without any delay. (to get signals handled.)
1138 * NOTE: this function doesnt have to take the runqueue lock,
1139 * because all it wants to ensure is that the remote task enters
1140 * the kernel. If the IPI races and the task has been migrated
1141 * to another CPU then no harm is done and the purpose has been
1142 * achieved as well.
1144 void kick_process(struct task_struct *p)
1146 int cpu;
1148 preempt_disable();
1149 cpu = task_cpu(p);
1150 if ((cpu != smp_processor_id()) && task_curr(p))
1151 smp_send_reschedule(cpu);
1152 preempt_enable();
1156 * Return a low guess at the load of a migration-source cpu weighted
1157 * according to the scheduling class and "nice" value.
1159 * We want to under-estimate the load of migration sources, to
1160 * balance conservatively.
1162 static inline unsigned long source_load(int cpu, int type)
1164 struct rq *rq = cpu_rq(cpu);
1166 if (type == 0)
1167 return rq->raw_weighted_load;
1169 return min(rq->cpu_load[type-1], rq->raw_weighted_load);
1173 * Return a high guess at the load of a migration-target cpu weighted
1174 * according to the scheduling class and "nice" value.
1176 static inline unsigned long target_load(int cpu, int type)
1178 struct rq *rq = cpu_rq(cpu);
1180 if (type == 0)
1181 return rq->raw_weighted_load;
1183 return max(rq->cpu_load[type-1], rq->raw_weighted_load);
1187 * Return the average load per task on the cpu's run queue
1189 static inline unsigned long cpu_avg_load_per_task(int cpu)
1191 struct rq *rq = cpu_rq(cpu);
1192 unsigned long n = rq->nr_running;
1194 return n ? rq->raw_weighted_load / n : SCHED_LOAD_SCALE;
1198 * find_idlest_group finds and returns the least busy CPU group within the
1199 * domain.
1201 static struct sched_group *
1202 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1204 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1205 unsigned long min_load = ULONG_MAX, this_load = 0;
1206 int load_idx = sd->forkexec_idx;
1207 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1209 do {
1210 unsigned long load, avg_load;
1211 int local_group;
1212 int i;
1214 /* Skip over this group if it has no CPUs allowed */
1215 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1216 goto nextgroup;
1218 local_group = cpu_isset(this_cpu, group->cpumask);
1220 /* Tally up the load of all CPUs in the group */
1221 avg_load = 0;
1223 for_each_cpu_mask(i, group->cpumask) {
1224 /* Bias balancing toward cpus of our domain */
1225 if (local_group)
1226 load = source_load(i, load_idx);
1227 else
1228 load = target_load(i, load_idx);
1230 avg_load += load;
1233 /* Adjust by relative CPU power of the group */
1234 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1236 if (local_group) {
1237 this_load = avg_load;
1238 this = group;
1239 } else if (avg_load < min_load) {
1240 min_load = avg_load;
1241 idlest = group;
1243 nextgroup:
1244 group = group->next;
1245 } while (group != sd->groups);
1247 if (!idlest || 100*this_load < imbalance*min_load)
1248 return NULL;
1249 return idlest;
1253 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1255 static int
1256 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1258 cpumask_t tmp;
1259 unsigned long load, min_load = ULONG_MAX;
1260 int idlest = -1;
1261 int i;
1263 /* Traverse only the allowed CPUs */
1264 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1266 for_each_cpu_mask(i, tmp) {
1267 load = weighted_cpuload(i);
1269 if (load < min_load || (load == min_load && i == this_cpu)) {
1270 min_load = load;
1271 idlest = i;
1275 return idlest;
1279 * sched_balance_self: balance the current task (running on cpu) in domains
1280 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1281 * SD_BALANCE_EXEC.
1283 * Balance, ie. select the least loaded group.
1285 * Returns the target CPU number, or the same CPU if no balancing is needed.
1287 * preempt must be disabled.
1289 static int sched_balance_self(int cpu, int flag)
1291 struct task_struct *t = current;
1292 struct sched_domain *tmp, *sd = NULL;
1294 for_each_domain(cpu, tmp) {
1296 * If power savings logic is enabled for a domain, stop there.
1298 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1299 break;
1300 if (tmp->flags & flag)
1301 sd = tmp;
1304 while (sd) {
1305 cpumask_t span;
1306 struct sched_group *group;
1307 int new_cpu, weight;
1309 if (!(sd->flags & flag)) {
1310 sd = sd->child;
1311 continue;
1314 span = sd->span;
1315 group = find_idlest_group(sd, t, cpu);
1316 if (!group) {
1317 sd = sd->child;
1318 continue;
1321 new_cpu = find_idlest_cpu(group, t, cpu);
1322 if (new_cpu == -1 || new_cpu == cpu) {
1323 /* Now try balancing at a lower domain level of cpu */
1324 sd = sd->child;
1325 continue;
1328 /* Now try balancing at a lower domain level of new_cpu */
1329 cpu = new_cpu;
1330 sd = NULL;
1331 weight = cpus_weight(span);
1332 for_each_domain(cpu, tmp) {
1333 if (weight <= cpus_weight(tmp->span))
1334 break;
1335 if (tmp->flags & flag)
1336 sd = tmp;
1338 /* while loop will break here if sd == NULL */
1341 return cpu;
1344 #endif /* CONFIG_SMP */
1347 * wake_idle() will wake a task on an idle cpu if task->cpu is
1348 * not idle and an idle cpu is available. The span of cpus to
1349 * search starts with cpus closest then further out as needed,
1350 * so we always favor a closer, idle cpu.
1352 * Returns the CPU we should wake onto.
1354 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1355 static int wake_idle(int cpu, struct task_struct *p)
1357 cpumask_t tmp;
1358 struct sched_domain *sd;
1359 int i;
1361 if (idle_cpu(cpu))
1362 return cpu;
1364 for_each_domain(cpu, sd) {
1365 if (sd->flags & SD_WAKE_IDLE) {
1366 cpus_and(tmp, sd->span, p->cpus_allowed);
1367 for_each_cpu_mask(i, tmp) {
1368 if (idle_cpu(i))
1369 return i;
1372 else
1373 break;
1375 return cpu;
1377 #else
1378 static inline int wake_idle(int cpu, struct task_struct *p)
1380 return cpu;
1382 #endif
1384 /***
1385 * try_to_wake_up - wake up a thread
1386 * @p: the to-be-woken-up thread
1387 * @state: the mask of task states that can be woken
1388 * @sync: do a synchronous wakeup?
1390 * Put it on the run-queue if it's not already there. The "current"
1391 * thread is always on the run-queue (except when the actual
1392 * re-schedule is in progress), and as such you're allowed to do
1393 * the simpler "current->state = TASK_RUNNING" to mark yourself
1394 * runnable without the overhead of this.
1396 * returns failure only if the task is already active.
1398 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1400 int cpu, this_cpu, success = 0;
1401 unsigned long flags;
1402 long old_state;
1403 struct rq *rq;
1404 #ifdef CONFIG_SMP
1405 struct sched_domain *sd, *this_sd = NULL;
1406 unsigned long load, this_load;
1407 int new_cpu;
1408 #endif
1410 rq = task_rq_lock(p, &flags);
1411 old_state = p->state;
1412 if (!(old_state & state))
1413 goto out;
1415 if (p->array)
1416 goto out_running;
1418 cpu = task_cpu(p);
1419 this_cpu = smp_processor_id();
1421 #ifdef CONFIG_SMP
1422 if (unlikely(task_running(rq, p)))
1423 goto out_activate;
1425 new_cpu = cpu;
1427 schedstat_inc(rq, ttwu_cnt);
1428 if (cpu == this_cpu) {
1429 schedstat_inc(rq, ttwu_local);
1430 goto out_set_cpu;
1433 for_each_domain(this_cpu, sd) {
1434 if (cpu_isset(cpu, sd->span)) {
1435 schedstat_inc(sd, ttwu_wake_remote);
1436 this_sd = sd;
1437 break;
1441 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1442 goto out_set_cpu;
1445 * Check for affine wakeup and passive balancing possibilities.
1447 if (this_sd) {
1448 int idx = this_sd->wake_idx;
1449 unsigned int imbalance;
1451 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1453 load = source_load(cpu, idx);
1454 this_load = target_load(this_cpu, idx);
1456 new_cpu = this_cpu; /* Wake to this CPU if we can */
1458 if (this_sd->flags & SD_WAKE_AFFINE) {
1459 unsigned long tl = this_load;
1460 unsigned long tl_per_task;
1462 tl_per_task = cpu_avg_load_per_task(this_cpu);
1465 * If sync wakeup then subtract the (maximum possible)
1466 * effect of the currently running task from the load
1467 * of the current CPU:
1469 if (sync)
1470 tl -= current->load_weight;
1472 if ((tl <= load &&
1473 tl + target_load(cpu, idx) <= tl_per_task) ||
1474 100*(tl + p->load_weight) <= imbalance*load) {
1476 * This domain has SD_WAKE_AFFINE and
1477 * p is cache cold in this domain, and
1478 * there is no bad imbalance.
1480 schedstat_inc(this_sd, ttwu_move_affine);
1481 goto out_set_cpu;
1486 * Start passive balancing when half the imbalance_pct
1487 * limit is reached.
1489 if (this_sd->flags & SD_WAKE_BALANCE) {
1490 if (imbalance*this_load <= 100*load) {
1491 schedstat_inc(this_sd, ttwu_move_balance);
1492 goto out_set_cpu;
1497 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1498 out_set_cpu:
1499 new_cpu = wake_idle(new_cpu, p);
1500 if (new_cpu != cpu) {
1501 set_task_cpu(p, new_cpu);
1502 task_rq_unlock(rq, &flags);
1503 /* might preempt at this point */
1504 rq = task_rq_lock(p, &flags);
1505 old_state = p->state;
1506 if (!(old_state & state))
1507 goto out;
1508 if (p->array)
1509 goto out_running;
1511 this_cpu = smp_processor_id();
1512 cpu = task_cpu(p);
1515 out_activate:
1516 #endif /* CONFIG_SMP */
1517 if (old_state == TASK_UNINTERRUPTIBLE) {
1518 rq->nr_uninterruptible--;
1520 * Tasks on involuntary sleep don't earn
1521 * sleep_avg beyond just interactive state.
1523 p->sleep_type = SLEEP_NONINTERACTIVE;
1524 } else
1527 * Tasks that have marked their sleep as noninteractive get
1528 * woken up with their sleep average not weighted in an
1529 * interactive way.
1531 if (old_state & TASK_NONINTERACTIVE)
1532 p->sleep_type = SLEEP_NONINTERACTIVE;
1535 activate_task(p, rq, cpu == this_cpu);
1537 * Sync wakeups (i.e. those types of wakeups where the waker
1538 * has indicated that it will leave the CPU in short order)
1539 * don't trigger a preemption, if the woken up task will run on
1540 * this cpu. (in this case the 'I will reschedule' promise of
1541 * the waker guarantees that the freshly woken up task is going
1542 * to be considered on this CPU.)
1544 if (!sync || cpu != this_cpu) {
1545 if (TASK_PREEMPTS_CURR(p, rq))
1546 resched_task(rq->curr);
1548 success = 1;
1550 out_running:
1551 p->state = TASK_RUNNING;
1552 out:
1553 task_rq_unlock(rq, &flags);
1555 return success;
1558 int fastcall wake_up_process(struct task_struct *p)
1560 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1561 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1563 EXPORT_SYMBOL(wake_up_process);
1565 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1567 return try_to_wake_up(p, state, 0);
1570 static void task_running_tick(struct rq *rq, struct task_struct *p);
1572 * Perform scheduler related setup for a newly forked process p.
1573 * p is forked by current.
1575 void fastcall sched_fork(struct task_struct *p, int clone_flags)
1577 int cpu = get_cpu();
1579 #ifdef CONFIG_SMP
1580 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1581 #endif
1582 set_task_cpu(p, cpu);
1585 * We mark the process as running here, but have not actually
1586 * inserted it onto the runqueue yet. This guarantees that
1587 * nobody will actually run it, and a signal or other external
1588 * event cannot wake it up and insert it on the runqueue either.
1590 p->state = TASK_RUNNING;
1593 * Make sure we do not leak PI boosting priority to the child:
1595 p->prio = current->normal_prio;
1597 INIT_LIST_HEAD(&p->run_list);
1598 p->array = NULL;
1599 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1600 if (unlikely(sched_info_on()))
1601 memset(&p->sched_info, 0, sizeof(p->sched_info));
1602 #endif
1603 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1604 p->oncpu = 0;
1605 #endif
1606 #ifdef CONFIG_PREEMPT
1607 /* Want to start with kernel preemption disabled. */
1608 task_thread_info(p)->preempt_count = 1;
1609 #endif
1611 * Share the timeslice between parent and child, thus the
1612 * total amount of pending timeslices in the system doesn't change,
1613 * resulting in more scheduling fairness.
1615 local_irq_disable();
1616 p->time_slice = (current->time_slice + 1) >> 1;
1618 * The remainder of the first timeslice might be recovered by
1619 * the parent if the child exits early enough.
1621 p->first_time_slice = 1;
1622 current->time_slice >>= 1;
1623 p->timestamp = sched_clock();
1624 if (unlikely(!current->time_slice)) {
1626 * This case is rare, it happens when the parent has only
1627 * a single jiffy left from its timeslice. Taking the
1628 * runqueue lock is not a problem.
1630 current->time_slice = 1;
1631 task_running_tick(cpu_rq(cpu), current);
1633 local_irq_enable();
1634 put_cpu();
1638 * wake_up_new_task - wake up a newly created task for the first time.
1640 * This function will do some initial scheduler statistics housekeeping
1641 * that must be done for every newly created context, then puts the task
1642 * on the runqueue and wakes it.
1644 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1646 struct rq *rq, *this_rq;
1647 unsigned long flags;
1648 int this_cpu, cpu;
1650 rq = task_rq_lock(p, &flags);
1651 BUG_ON(p->state != TASK_RUNNING);
1652 this_cpu = smp_processor_id();
1653 cpu = task_cpu(p);
1656 * We decrease the sleep average of forking parents
1657 * and children as well, to keep max-interactive tasks
1658 * from forking tasks that are max-interactive. The parent
1659 * (current) is done further down, under its lock.
1661 p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1662 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1664 p->prio = effective_prio(p);
1666 if (likely(cpu == this_cpu)) {
1667 if (!(clone_flags & CLONE_VM)) {
1669 * The VM isn't cloned, so we're in a good position to
1670 * do child-runs-first in anticipation of an exec. This
1671 * usually avoids a lot of COW overhead.
1673 if (unlikely(!current->array))
1674 __activate_task(p, rq);
1675 else {
1676 p->prio = current->prio;
1677 p->normal_prio = current->normal_prio;
1678 list_add_tail(&p->run_list, &current->run_list);
1679 p->array = current->array;
1680 p->array->nr_active++;
1681 inc_nr_running(p, rq);
1683 set_need_resched();
1684 } else
1685 /* Run child last */
1686 __activate_task(p, rq);
1688 * We skip the following code due to cpu == this_cpu
1690 * task_rq_unlock(rq, &flags);
1691 * this_rq = task_rq_lock(current, &flags);
1693 this_rq = rq;
1694 } else {
1695 this_rq = cpu_rq(this_cpu);
1698 * Not the local CPU - must adjust timestamp. This should
1699 * get optimised away in the !CONFIG_SMP case.
1701 p->timestamp = (p->timestamp - this_rq->most_recent_timestamp)
1702 + rq->most_recent_timestamp;
1703 __activate_task(p, rq);
1704 if (TASK_PREEMPTS_CURR(p, rq))
1705 resched_task(rq->curr);
1708 * Parent and child are on different CPUs, now get the
1709 * parent runqueue to update the parent's ->sleep_avg:
1711 task_rq_unlock(rq, &flags);
1712 this_rq = task_rq_lock(current, &flags);
1714 current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1715 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1716 task_rq_unlock(this_rq, &flags);
1720 * Potentially available exiting-child timeslices are
1721 * retrieved here - this way the parent does not get
1722 * penalized for creating too many threads.
1724 * (this cannot be used to 'generate' timeslices
1725 * artificially, because any timeslice recovered here
1726 * was given away by the parent in the first place.)
1728 void fastcall sched_exit(struct task_struct *p)
1730 unsigned long flags;
1731 struct rq *rq;
1734 * If the child was a (relative-) CPU hog then decrease
1735 * the sleep_avg of the parent as well.
1737 rq = task_rq_lock(p->parent, &flags);
1738 if (p->first_time_slice && task_cpu(p) == task_cpu(p->parent)) {
1739 p->parent->time_slice += p->time_slice;
1740 if (unlikely(p->parent->time_slice > task_timeslice(p)))
1741 p->parent->time_slice = task_timeslice(p);
1743 if (p->sleep_avg < p->parent->sleep_avg)
1744 p->parent->sleep_avg = p->parent->sleep_avg /
1745 (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg /
1746 (EXIT_WEIGHT + 1);
1747 task_rq_unlock(rq, &flags);
1751 * prepare_task_switch - prepare to switch tasks
1752 * @rq: the runqueue preparing to switch
1753 * @next: the task we are going to switch to.
1755 * This is called with the rq lock held and interrupts off. It must
1756 * be paired with a subsequent finish_task_switch after the context
1757 * switch.
1759 * prepare_task_switch sets up locking and calls architecture specific
1760 * hooks.
1762 static inline void prepare_task_switch(struct rq *rq, struct task_struct *next)
1764 prepare_lock_switch(rq, next);
1765 prepare_arch_switch(next);
1769 * finish_task_switch - clean up after a task-switch
1770 * @rq: runqueue associated with task-switch
1771 * @prev: the thread we just switched away from.
1773 * finish_task_switch must be called after the context switch, paired
1774 * with a prepare_task_switch call before the context switch.
1775 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1776 * and do any other architecture-specific cleanup actions.
1778 * Note that we may have delayed dropping an mm in context_switch(). If
1779 * so, we finish that here outside of the runqueue lock. (Doing it
1780 * with the lock held can cause deadlocks; see schedule() for
1781 * details.)
1783 static inline void finish_task_switch(struct rq *rq, struct task_struct *prev)
1784 __releases(rq->lock)
1786 struct mm_struct *mm = rq->prev_mm;
1787 long prev_state;
1789 rq->prev_mm = NULL;
1792 * A task struct has one reference for the use as "current".
1793 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1794 * schedule one last time. The schedule call will never return, and
1795 * the scheduled task must drop that reference.
1796 * The test for TASK_DEAD must occur while the runqueue locks are
1797 * still held, otherwise prev could be scheduled on another cpu, die
1798 * there before we look at prev->state, and then the reference would
1799 * be dropped twice.
1800 * Manfred Spraul <manfred@colorfullife.com>
1802 prev_state = prev->state;
1803 finish_arch_switch(prev);
1804 finish_lock_switch(rq, prev);
1805 if (mm)
1806 mmdrop(mm);
1807 if (unlikely(prev_state == TASK_DEAD)) {
1809 * Remove function-return probe instances associated with this
1810 * task and put them back on the free list.
1812 kprobe_flush_task(prev);
1813 put_task_struct(prev);
1818 * schedule_tail - first thing a freshly forked thread must call.
1819 * @prev: the thread we just switched away from.
1821 asmlinkage void schedule_tail(struct task_struct *prev)
1822 __releases(rq->lock)
1824 struct rq *rq = this_rq();
1826 finish_task_switch(rq, prev);
1827 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1828 /* In this case, finish_task_switch does not reenable preemption */
1829 preempt_enable();
1830 #endif
1831 if (current->set_child_tid)
1832 put_user(current->pid, current->set_child_tid);
1836 * context_switch - switch to the new MM and the new
1837 * thread's register state.
1839 static inline struct task_struct *
1840 context_switch(struct rq *rq, struct task_struct *prev,
1841 struct task_struct *next)
1843 struct mm_struct *mm = next->mm;
1844 struct mm_struct *oldmm = prev->active_mm;
1846 if (!mm) {
1847 next->active_mm = oldmm;
1848 atomic_inc(&oldmm->mm_count);
1849 enter_lazy_tlb(oldmm, next);
1850 } else
1851 switch_mm(oldmm, mm, next);
1853 if (!prev->mm) {
1854 prev->active_mm = NULL;
1855 WARN_ON(rq->prev_mm);
1856 rq->prev_mm = oldmm;
1859 * Since the runqueue lock will be released by the next
1860 * task (which is an invalid locking op but in the case
1861 * of the scheduler it's an obvious special-case), so we
1862 * do an early lockdep release here:
1864 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1865 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1866 #endif
1868 /* Here we just switch the register state and the stack. */
1869 switch_to(prev, next, prev);
1871 return prev;
1875 * nr_running, nr_uninterruptible and nr_context_switches:
1877 * externally visible scheduler statistics: current number of runnable
1878 * threads, current number of uninterruptible-sleeping threads, total
1879 * number of context switches performed since bootup.
1881 unsigned long nr_running(void)
1883 unsigned long i, sum = 0;
1885 for_each_online_cpu(i)
1886 sum += cpu_rq(i)->nr_running;
1888 return sum;
1891 unsigned long nr_uninterruptible(void)
1893 unsigned long i, sum = 0;
1895 for_each_possible_cpu(i)
1896 sum += cpu_rq(i)->nr_uninterruptible;
1899 * Since we read the counters lockless, it might be slightly
1900 * inaccurate. Do not allow it to go below zero though:
1902 if (unlikely((long)sum < 0))
1903 sum = 0;
1905 return sum;
1908 unsigned long long nr_context_switches(void)
1910 int i;
1911 unsigned long long sum = 0;
1913 for_each_possible_cpu(i)
1914 sum += cpu_rq(i)->nr_switches;
1916 return sum;
1919 unsigned long nr_iowait(void)
1921 unsigned long i, sum = 0;
1923 for_each_possible_cpu(i)
1924 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1926 return sum;
1929 unsigned long nr_active(void)
1931 unsigned long i, running = 0, uninterruptible = 0;
1933 for_each_online_cpu(i) {
1934 running += cpu_rq(i)->nr_running;
1935 uninterruptible += cpu_rq(i)->nr_uninterruptible;
1938 if (unlikely((long)uninterruptible < 0))
1939 uninterruptible = 0;
1941 return running + uninterruptible;
1944 #ifdef CONFIG_SMP
1947 * Is this task likely cache-hot:
1949 static inline int
1950 task_hot(struct task_struct *p, unsigned long long now, struct sched_domain *sd)
1952 return (long long)(now - p->last_ran) < (long long)sd->cache_hot_time;
1956 * double_rq_lock - safely lock two runqueues
1958 * Note this does not disable interrupts like task_rq_lock,
1959 * you need to do so manually before calling.
1961 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1962 __acquires(rq1->lock)
1963 __acquires(rq2->lock)
1965 BUG_ON(!irqs_disabled());
1966 if (rq1 == rq2) {
1967 spin_lock(&rq1->lock);
1968 __acquire(rq2->lock); /* Fake it out ;) */
1969 } else {
1970 if (rq1 < rq2) {
1971 spin_lock(&rq1->lock);
1972 spin_lock(&rq2->lock);
1973 } else {
1974 spin_lock(&rq2->lock);
1975 spin_lock(&rq1->lock);
1981 * double_rq_unlock - safely unlock two runqueues
1983 * Note this does not restore interrupts like task_rq_unlock,
1984 * you need to do so manually after calling.
1986 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1987 __releases(rq1->lock)
1988 __releases(rq2->lock)
1990 spin_unlock(&rq1->lock);
1991 if (rq1 != rq2)
1992 spin_unlock(&rq2->lock);
1993 else
1994 __release(rq2->lock);
1998 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2000 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2001 __releases(this_rq->lock)
2002 __acquires(busiest->lock)
2003 __acquires(this_rq->lock)
2005 if (unlikely(!irqs_disabled())) {
2006 /* printk() doesn't work good under rq->lock */
2007 spin_unlock(&this_rq->lock);
2008 BUG_ON(1);
2010 if (unlikely(!spin_trylock(&busiest->lock))) {
2011 if (busiest < this_rq) {
2012 spin_unlock(&this_rq->lock);
2013 spin_lock(&busiest->lock);
2014 spin_lock(&this_rq->lock);
2015 } else
2016 spin_lock(&busiest->lock);
2021 * If dest_cpu is allowed for this process, migrate the task to it.
2022 * This is accomplished by forcing the cpu_allowed mask to only
2023 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2024 * the cpu_allowed mask is restored.
2026 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2028 struct migration_req req;
2029 unsigned long flags;
2030 struct rq *rq;
2032 rq = task_rq_lock(p, &flags);
2033 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2034 || unlikely(cpu_is_offline(dest_cpu)))
2035 goto out;
2037 /* force the process onto the specified CPU */
2038 if (migrate_task(p, dest_cpu, &req)) {
2039 /* Need to wait for migration thread (might exit: take ref). */
2040 struct task_struct *mt = rq->migration_thread;
2042 get_task_struct(mt);
2043 task_rq_unlock(rq, &flags);
2044 wake_up_process(mt);
2045 put_task_struct(mt);
2046 wait_for_completion(&req.done);
2048 return;
2050 out:
2051 task_rq_unlock(rq, &flags);
2055 * sched_exec - execve() is a valuable balancing opportunity, because at
2056 * this point the task has the smallest effective memory and cache footprint.
2058 void sched_exec(void)
2060 int new_cpu, this_cpu = get_cpu();
2061 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2062 put_cpu();
2063 if (new_cpu != this_cpu)
2064 sched_migrate_task(current, new_cpu);
2068 * pull_task - move a task from a remote runqueue to the local runqueue.
2069 * Both runqueues must be locked.
2071 static void pull_task(struct rq *src_rq, struct prio_array *src_array,
2072 struct task_struct *p, struct rq *this_rq,
2073 struct prio_array *this_array, int this_cpu)
2075 dequeue_task(p, src_array);
2076 dec_nr_running(p, src_rq);
2077 set_task_cpu(p, this_cpu);
2078 inc_nr_running(p, this_rq);
2079 enqueue_task(p, this_array);
2080 p->timestamp = (p->timestamp - src_rq->most_recent_timestamp)
2081 + this_rq->most_recent_timestamp;
2083 * Note that idle threads have a prio of MAX_PRIO, for this test
2084 * to be always true for them.
2086 if (TASK_PREEMPTS_CURR(p, this_rq))
2087 resched_task(this_rq->curr);
2091 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2093 static
2094 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2095 struct sched_domain *sd, enum idle_type idle,
2096 int *all_pinned)
2099 * We do not migrate tasks that are:
2100 * 1) running (obviously), or
2101 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2102 * 3) are cache-hot on their current CPU.
2104 if (!cpu_isset(this_cpu, p->cpus_allowed))
2105 return 0;
2106 *all_pinned = 0;
2108 if (task_running(rq, p))
2109 return 0;
2112 * Aggressive migration if:
2113 * 1) task is cache cold, or
2114 * 2) too many balance attempts have failed.
2117 if (sd->nr_balance_failed > sd->cache_nice_tries) {
2118 #ifdef CONFIG_SCHEDSTATS
2119 if (task_hot(p, rq->most_recent_timestamp, sd))
2120 schedstat_inc(sd, lb_hot_gained[idle]);
2121 #endif
2122 return 1;
2125 if (task_hot(p, rq->most_recent_timestamp, sd))
2126 return 0;
2127 return 1;
2130 #define rq_best_prio(rq) min((rq)->curr->prio, (rq)->best_expired_prio)
2133 * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted
2134 * load from busiest to this_rq, as part of a balancing operation within
2135 * "domain". Returns the number of tasks moved.
2137 * Called with both runqueues locked.
2139 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2140 unsigned long max_nr_move, unsigned long max_load_move,
2141 struct sched_domain *sd, enum idle_type idle,
2142 int *all_pinned)
2144 int idx, pulled = 0, pinned = 0, this_best_prio, best_prio,
2145 best_prio_seen, skip_for_load;
2146 struct prio_array *array, *dst_array;
2147 struct list_head *head, *curr;
2148 struct task_struct *tmp;
2149 long rem_load_move;
2151 if (max_nr_move == 0 || max_load_move == 0)
2152 goto out;
2154 rem_load_move = max_load_move;
2155 pinned = 1;
2156 this_best_prio = rq_best_prio(this_rq);
2157 best_prio = rq_best_prio(busiest);
2159 * Enable handling of the case where there is more than one task
2160 * with the best priority. If the current running task is one
2161 * of those with prio==best_prio we know it won't be moved
2162 * and therefore it's safe to override the skip (based on load) of
2163 * any task we find with that prio.
2165 best_prio_seen = best_prio == busiest->curr->prio;
2168 * We first consider expired tasks. Those will likely not be
2169 * executed in the near future, and they are most likely to
2170 * be cache-cold, thus switching CPUs has the least effect
2171 * on them.
2173 if (busiest->expired->nr_active) {
2174 array = busiest->expired;
2175 dst_array = this_rq->expired;
2176 } else {
2177 array = busiest->active;
2178 dst_array = this_rq->active;
2181 new_array:
2182 /* Start searching at priority 0: */
2183 idx = 0;
2184 skip_bitmap:
2185 if (!idx)
2186 idx = sched_find_first_bit(array->bitmap);
2187 else
2188 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
2189 if (idx >= MAX_PRIO) {
2190 if (array == busiest->expired && busiest->active->nr_active) {
2191 array = busiest->active;
2192 dst_array = this_rq->active;
2193 goto new_array;
2195 goto out;
2198 head = array->queue + idx;
2199 curr = head->prev;
2200 skip_queue:
2201 tmp = list_entry(curr, struct task_struct, run_list);
2203 curr = curr->prev;
2206 * To help distribute high priority tasks accross CPUs we don't
2207 * skip a task if it will be the highest priority task (i.e. smallest
2208 * prio value) on its new queue regardless of its load weight
2210 skip_for_load = tmp->load_weight > rem_load_move;
2211 if (skip_for_load && idx < this_best_prio)
2212 skip_for_load = !best_prio_seen && idx == best_prio;
2213 if (skip_for_load ||
2214 !can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
2216 best_prio_seen |= idx == best_prio;
2217 if (curr != head)
2218 goto skip_queue;
2219 idx++;
2220 goto skip_bitmap;
2223 pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
2224 pulled++;
2225 rem_load_move -= tmp->load_weight;
2228 * We only want to steal up to the prescribed number of tasks
2229 * and the prescribed amount of weighted load.
2231 if (pulled < max_nr_move && rem_load_move > 0) {
2232 if (idx < this_best_prio)
2233 this_best_prio = idx;
2234 if (curr != head)
2235 goto skip_queue;
2236 idx++;
2237 goto skip_bitmap;
2239 out:
2241 * Right now, this is the only place pull_task() is called,
2242 * so we can safely collect pull_task() stats here rather than
2243 * inside pull_task().
2245 schedstat_add(sd, lb_gained[idle], pulled);
2247 if (all_pinned)
2248 *all_pinned = pinned;
2249 return pulled;
2253 * find_busiest_group finds and returns the busiest CPU group within the
2254 * domain. It calculates and returns the amount of weighted load which
2255 * should be moved to restore balance via the imbalance parameter.
2257 static struct sched_group *
2258 find_busiest_group(struct sched_domain *sd, int this_cpu,
2259 unsigned long *imbalance, enum idle_type idle, int *sd_idle,
2260 cpumask_t *cpus, int *balance)
2262 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2263 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2264 unsigned long max_pull;
2265 unsigned long busiest_load_per_task, busiest_nr_running;
2266 unsigned long this_load_per_task, this_nr_running;
2267 int load_idx;
2268 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2269 int power_savings_balance = 1;
2270 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2271 unsigned long min_nr_running = ULONG_MAX;
2272 struct sched_group *group_min = NULL, *group_leader = NULL;
2273 #endif
2275 max_load = this_load = total_load = total_pwr = 0;
2276 busiest_load_per_task = busiest_nr_running = 0;
2277 this_load_per_task = this_nr_running = 0;
2278 if (idle == NOT_IDLE)
2279 load_idx = sd->busy_idx;
2280 else if (idle == NEWLY_IDLE)
2281 load_idx = sd->newidle_idx;
2282 else
2283 load_idx = sd->idle_idx;
2285 do {
2286 unsigned long load, group_capacity;
2287 int local_group;
2288 int i;
2289 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2290 unsigned long sum_nr_running, sum_weighted_load;
2292 local_group = cpu_isset(this_cpu, group->cpumask);
2294 if (local_group)
2295 balance_cpu = first_cpu(group->cpumask);
2297 /* Tally up the load of all CPUs in the group */
2298 sum_weighted_load = sum_nr_running = avg_load = 0;
2300 for_each_cpu_mask(i, group->cpumask) {
2301 struct rq *rq;
2303 if (!cpu_isset(i, *cpus))
2304 continue;
2306 rq = cpu_rq(i);
2308 if (*sd_idle && !idle_cpu(i))
2309 *sd_idle = 0;
2311 /* Bias balancing toward cpus of our domain */
2312 if (local_group) {
2313 if (idle_cpu(i) && !first_idle_cpu) {
2314 first_idle_cpu = 1;
2315 balance_cpu = i;
2318 load = target_load(i, load_idx);
2319 } else
2320 load = source_load(i, load_idx);
2322 avg_load += load;
2323 sum_nr_running += rq->nr_running;
2324 sum_weighted_load += rq->raw_weighted_load;
2328 * First idle cpu or the first cpu(busiest) in this sched group
2329 * is eligible for doing load balancing at this and above
2330 * domains.
2332 if (local_group && balance_cpu != this_cpu && balance) {
2333 *balance = 0;
2334 goto ret;
2337 total_load += avg_load;
2338 total_pwr += group->cpu_power;
2340 /* Adjust by relative CPU power of the group */
2341 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
2343 group_capacity = group->cpu_power / SCHED_LOAD_SCALE;
2345 if (local_group) {
2346 this_load = avg_load;
2347 this = group;
2348 this_nr_running = sum_nr_running;
2349 this_load_per_task = sum_weighted_load;
2350 } else if (avg_load > max_load &&
2351 sum_nr_running > group_capacity) {
2352 max_load = avg_load;
2353 busiest = group;
2354 busiest_nr_running = sum_nr_running;
2355 busiest_load_per_task = sum_weighted_load;
2358 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2360 * Busy processors will not participate in power savings
2361 * balance.
2363 if (idle == NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2364 goto group_next;
2367 * If the local group is idle or completely loaded
2368 * no need to do power savings balance at this domain
2370 if (local_group && (this_nr_running >= group_capacity ||
2371 !this_nr_running))
2372 power_savings_balance = 0;
2375 * If a group is already running at full capacity or idle,
2376 * don't include that group in power savings calculations
2378 if (!power_savings_balance || sum_nr_running >= group_capacity
2379 || !sum_nr_running)
2380 goto group_next;
2383 * Calculate the group which has the least non-idle load.
2384 * This is the group from where we need to pick up the load
2385 * for saving power
2387 if ((sum_nr_running < min_nr_running) ||
2388 (sum_nr_running == min_nr_running &&
2389 first_cpu(group->cpumask) <
2390 first_cpu(group_min->cpumask))) {
2391 group_min = group;
2392 min_nr_running = sum_nr_running;
2393 min_load_per_task = sum_weighted_load /
2394 sum_nr_running;
2398 * Calculate the group which is almost near its
2399 * capacity but still has some space to pick up some load
2400 * from other group and save more power
2402 if (sum_nr_running <= group_capacity - 1) {
2403 if (sum_nr_running > leader_nr_running ||
2404 (sum_nr_running == leader_nr_running &&
2405 first_cpu(group->cpumask) >
2406 first_cpu(group_leader->cpumask))) {
2407 group_leader = group;
2408 leader_nr_running = sum_nr_running;
2411 group_next:
2412 #endif
2413 group = group->next;
2414 } while (group != sd->groups);
2416 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2417 goto out_balanced;
2419 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2421 if (this_load >= avg_load ||
2422 100*max_load <= sd->imbalance_pct*this_load)
2423 goto out_balanced;
2425 busiest_load_per_task /= busiest_nr_running;
2427 * We're trying to get all the cpus to the average_load, so we don't
2428 * want to push ourselves above the average load, nor do we wish to
2429 * reduce the max loaded cpu below the average load, as either of these
2430 * actions would just result in more rebalancing later, and ping-pong
2431 * tasks around. Thus we look for the minimum possible imbalance.
2432 * Negative imbalances (*we* are more loaded than anyone else) will
2433 * be counted as no imbalance for these purposes -- we can't fix that
2434 * by pulling tasks to us. Be careful of negative numbers as they'll
2435 * appear as very large values with unsigned longs.
2437 if (max_load <= busiest_load_per_task)
2438 goto out_balanced;
2441 * In the presence of smp nice balancing, certain scenarios can have
2442 * max load less than avg load(as we skip the groups at or below
2443 * its cpu_power, while calculating max_load..)
2445 if (max_load < avg_load) {
2446 *imbalance = 0;
2447 goto small_imbalance;
2450 /* Don't want to pull so many tasks that a group would go idle */
2451 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2453 /* How much load to actually move to equalise the imbalance */
2454 *imbalance = min(max_pull * busiest->cpu_power,
2455 (avg_load - this_load) * this->cpu_power)
2456 / SCHED_LOAD_SCALE;
2459 * if *imbalance is less than the average load per runnable task
2460 * there is no gaurantee that any tasks will be moved so we'll have
2461 * a think about bumping its value to force at least one task to be
2462 * moved
2464 if (*imbalance < busiest_load_per_task) {
2465 unsigned long tmp, pwr_now, pwr_move;
2466 unsigned int imbn;
2468 small_imbalance:
2469 pwr_move = pwr_now = 0;
2470 imbn = 2;
2471 if (this_nr_running) {
2472 this_load_per_task /= this_nr_running;
2473 if (busiest_load_per_task > this_load_per_task)
2474 imbn = 1;
2475 } else
2476 this_load_per_task = SCHED_LOAD_SCALE;
2478 if (max_load - this_load >= busiest_load_per_task * imbn) {
2479 *imbalance = busiest_load_per_task;
2480 return busiest;
2484 * OK, we don't have enough imbalance to justify moving tasks,
2485 * however we may be able to increase total CPU power used by
2486 * moving them.
2489 pwr_now += busiest->cpu_power *
2490 min(busiest_load_per_task, max_load);
2491 pwr_now += this->cpu_power *
2492 min(this_load_per_task, this_load);
2493 pwr_now /= SCHED_LOAD_SCALE;
2495 /* Amount of load we'd subtract */
2496 tmp = busiest_load_per_task * SCHED_LOAD_SCALE /
2497 busiest->cpu_power;
2498 if (max_load > tmp)
2499 pwr_move += busiest->cpu_power *
2500 min(busiest_load_per_task, max_load - tmp);
2502 /* Amount of load we'd add */
2503 if (max_load * busiest->cpu_power <
2504 busiest_load_per_task * SCHED_LOAD_SCALE)
2505 tmp = max_load * busiest->cpu_power / this->cpu_power;
2506 else
2507 tmp = busiest_load_per_task * SCHED_LOAD_SCALE /
2508 this->cpu_power;
2509 pwr_move += this->cpu_power *
2510 min(this_load_per_task, this_load + tmp);
2511 pwr_move /= SCHED_LOAD_SCALE;
2513 /* Move if we gain throughput */
2514 if (pwr_move <= pwr_now)
2515 goto out_balanced;
2517 *imbalance = busiest_load_per_task;
2520 return busiest;
2522 out_balanced:
2523 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2524 if (idle == NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2525 goto ret;
2527 if (this == group_leader && group_leader != group_min) {
2528 *imbalance = min_load_per_task;
2529 return group_min;
2531 #endif
2532 ret:
2533 *imbalance = 0;
2534 return NULL;
2538 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2540 static struct rq *
2541 find_busiest_queue(struct sched_group *group, enum idle_type idle,
2542 unsigned long imbalance, cpumask_t *cpus)
2544 struct rq *busiest = NULL, *rq;
2545 unsigned long max_load = 0;
2546 int i;
2548 for_each_cpu_mask(i, group->cpumask) {
2550 if (!cpu_isset(i, *cpus))
2551 continue;
2553 rq = cpu_rq(i);
2555 if (rq->nr_running == 1 && rq->raw_weighted_load > imbalance)
2556 continue;
2558 if (rq->raw_weighted_load > max_load) {
2559 max_load = rq->raw_weighted_load;
2560 busiest = rq;
2564 return busiest;
2568 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2569 * so long as it is large enough.
2571 #define MAX_PINNED_INTERVAL 512
2573 static inline unsigned long minus_1_or_zero(unsigned long n)
2575 return n > 0 ? n - 1 : 0;
2579 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2580 * tasks if there is an imbalance.
2582 static int load_balance(int this_cpu, struct rq *this_rq,
2583 struct sched_domain *sd, enum idle_type idle,
2584 int *balance)
2586 int nr_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2587 struct sched_group *group;
2588 unsigned long imbalance;
2589 struct rq *busiest;
2590 cpumask_t cpus = CPU_MASK_ALL;
2591 unsigned long flags;
2594 * When power savings policy is enabled for the parent domain, idle
2595 * sibling can pick up load irrespective of busy siblings. In this case,
2596 * let the state of idle sibling percolate up as IDLE, instead of
2597 * portraying it as NOT_IDLE.
2599 if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2600 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2601 sd_idle = 1;
2603 schedstat_inc(sd, lb_cnt[idle]);
2605 redo:
2606 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2607 &cpus, balance);
2609 if (*balance == 0)
2610 goto out_balanced;
2612 if (!group) {
2613 schedstat_inc(sd, lb_nobusyg[idle]);
2614 goto out_balanced;
2617 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2618 if (!busiest) {
2619 schedstat_inc(sd, lb_nobusyq[idle]);
2620 goto out_balanced;
2623 BUG_ON(busiest == this_rq);
2625 schedstat_add(sd, lb_imbalance[idle], imbalance);
2627 nr_moved = 0;
2628 if (busiest->nr_running > 1) {
2630 * Attempt to move tasks. If find_busiest_group has found
2631 * an imbalance but busiest->nr_running <= 1, the group is
2632 * still unbalanced. nr_moved simply stays zero, so it is
2633 * correctly treated as an imbalance.
2635 local_irq_save(flags);
2636 double_rq_lock(this_rq, busiest);
2637 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2638 minus_1_or_zero(busiest->nr_running),
2639 imbalance, sd, idle, &all_pinned);
2640 double_rq_unlock(this_rq, busiest);
2641 local_irq_restore(flags);
2643 /* All tasks on this runqueue were pinned by CPU affinity */
2644 if (unlikely(all_pinned)) {
2645 cpu_clear(cpu_of(busiest), cpus);
2646 if (!cpus_empty(cpus))
2647 goto redo;
2648 goto out_balanced;
2652 if (!nr_moved) {
2653 schedstat_inc(sd, lb_failed[idle]);
2654 sd->nr_balance_failed++;
2656 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2658 spin_lock_irqsave(&busiest->lock, flags);
2660 /* don't kick the migration_thread, if the curr
2661 * task on busiest cpu can't be moved to this_cpu
2663 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2664 spin_unlock_irqrestore(&busiest->lock, flags);
2665 all_pinned = 1;
2666 goto out_one_pinned;
2669 if (!busiest->active_balance) {
2670 busiest->active_balance = 1;
2671 busiest->push_cpu = this_cpu;
2672 active_balance = 1;
2674 spin_unlock_irqrestore(&busiest->lock, flags);
2675 if (active_balance)
2676 wake_up_process(busiest->migration_thread);
2679 * We've kicked active balancing, reset the failure
2680 * counter.
2682 sd->nr_balance_failed = sd->cache_nice_tries+1;
2684 } else
2685 sd->nr_balance_failed = 0;
2687 if (likely(!active_balance)) {
2688 /* We were unbalanced, so reset the balancing interval */
2689 sd->balance_interval = sd->min_interval;
2690 } else {
2692 * If we've begun active balancing, start to back off. This
2693 * case may not be covered by the all_pinned logic if there
2694 * is only 1 task on the busy runqueue (because we don't call
2695 * move_tasks).
2697 if (sd->balance_interval < sd->max_interval)
2698 sd->balance_interval *= 2;
2701 if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2702 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2703 return -1;
2704 return nr_moved;
2706 out_balanced:
2707 schedstat_inc(sd, lb_balanced[idle]);
2709 sd->nr_balance_failed = 0;
2711 out_one_pinned:
2712 /* tune up the balancing interval */
2713 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2714 (sd->balance_interval < sd->max_interval))
2715 sd->balance_interval *= 2;
2717 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2718 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2719 return -1;
2720 return 0;
2724 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2725 * tasks if there is an imbalance.
2727 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2728 * this_rq is locked.
2730 static int
2731 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2733 struct sched_group *group;
2734 struct rq *busiest = NULL;
2735 unsigned long imbalance;
2736 int nr_moved = 0;
2737 int sd_idle = 0;
2738 cpumask_t cpus = CPU_MASK_ALL;
2741 * When power savings policy is enabled for the parent domain, idle
2742 * sibling can pick up load irrespective of busy siblings. In this case,
2743 * let the state of idle sibling percolate up as IDLE, instead of
2744 * portraying it as NOT_IDLE.
2746 if (sd->flags & SD_SHARE_CPUPOWER &&
2747 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2748 sd_idle = 1;
2750 schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
2751 redo:
2752 group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE,
2753 &sd_idle, &cpus, NULL);
2754 if (!group) {
2755 schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]);
2756 goto out_balanced;
2759 busiest = find_busiest_queue(group, NEWLY_IDLE, imbalance,
2760 &cpus);
2761 if (!busiest) {
2762 schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]);
2763 goto out_balanced;
2766 BUG_ON(busiest == this_rq);
2768 schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
2770 nr_moved = 0;
2771 if (busiest->nr_running > 1) {
2772 /* Attempt to move tasks */
2773 double_lock_balance(this_rq, busiest);
2774 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2775 minus_1_or_zero(busiest->nr_running),
2776 imbalance, sd, NEWLY_IDLE, NULL);
2777 spin_unlock(&busiest->lock);
2779 if (!nr_moved) {
2780 cpu_clear(cpu_of(busiest), cpus);
2781 if (!cpus_empty(cpus))
2782 goto redo;
2786 if (!nr_moved) {
2787 schedstat_inc(sd, lb_failed[NEWLY_IDLE]);
2788 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2789 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2790 return -1;
2791 } else
2792 sd->nr_balance_failed = 0;
2794 return nr_moved;
2796 out_balanced:
2797 schedstat_inc(sd, lb_balanced[NEWLY_IDLE]);
2798 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2799 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2800 return -1;
2801 sd->nr_balance_failed = 0;
2803 return 0;
2807 * idle_balance is called by schedule() if this_cpu is about to become
2808 * idle. Attempts to pull tasks from other CPUs.
2810 static void idle_balance(int this_cpu, struct rq *this_rq)
2812 struct sched_domain *sd;
2813 int pulled_task = 0;
2814 unsigned long next_balance = jiffies + 60 * HZ;
2816 for_each_domain(this_cpu, sd) {
2817 if (sd->flags & SD_BALANCE_NEWIDLE) {
2818 /* If we've pulled tasks over stop searching: */
2819 pulled_task = load_balance_newidle(this_cpu,
2820 this_rq, sd);
2821 if (time_after(next_balance,
2822 sd->last_balance + sd->balance_interval))
2823 next_balance = sd->last_balance
2824 + sd->balance_interval;
2825 if (pulled_task)
2826 break;
2829 if (!pulled_task)
2831 * We are going idle. next_balance may be set based on
2832 * a busy processor. So reset next_balance.
2834 this_rq->next_balance = next_balance;
2838 * active_load_balance is run by migration threads. It pushes running tasks
2839 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2840 * running on each physical CPU where possible, and avoids physical /
2841 * logical imbalances.
2843 * Called with busiest_rq locked.
2845 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2847 int target_cpu = busiest_rq->push_cpu;
2848 struct sched_domain *sd;
2849 struct rq *target_rq;
2851 /* Is there any task to move? */
2852 if (busiest_rq->nr_running <= 1)
2853 return;
2855 target_rq = cpu_rq(target_cpu);
2858 * This condition is "impossible", if it occurs
2859 * we need to fix it. Originally reported by
2860 * Bjorn Helgaas on a 128-cpu setup.
2862 BUG_ON(busiest_rq == target_rq);
2864 /* move a task from busiest_rq to target_rq */
2865 double_lock_balance(busiest_rq, target_rq);
2867 /* Search for an sd spanning us and the target CPU. */
2868 for_each_domain(target_cpu, sd) {
2869 if ((sd->flags & SD_LOAD_BALANCE) &&
2870 cpu_isset(busiest_cpu, sd->span))
2871 break;
2874 if (likely(sd)) {
2875 schedstat_inc(sd, alb_cnt);
2877 if (move_tasks(target_rq, target_cpu, busiest_rq, 1,
2878 RTPRIO_TO_LOAD_WEIGHT(100), sd, SCHED_IDLE,
2879 NULL))
2880 schedstat_inc(sd, alb_pushed);
2881 else
2882 schedstat_inc(sd, alb_failed);
2884 spin_unlock(&target_rq->lock);
2887 static void update_load(struct rq *this_rq)
2889 unsigned long this_load;
2890 int i, scale;
2892 this_load = this_rq->raw_weighted_load;
2894 /* Update our load: */
2895 for (i = 0, scale = 1; i < 3; i++, scale <<= 1) {
2896 unsigned long old_load, new_load;
2898 old_load = this_rq->cpu_load[i];
2899 new_load = this_load;
2901 * Round up the averaging division if load is increasing. This
2902 * prevents us from getting stuck on 9 if the load is 10, for
2903 * example.
2905 if (new_load > old_load)
2906 new_load += scale-1;
2907 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale;
2912 * run_rebalance_domains is triggered when needed from the scheduler tick.
2914 * It checks each scheduling domain to see if it is due to be balanced,
2915 * and initiates a balancing operation if so.
2917 * Balancing parameters are set up in arch_init_sched_domains.
2919 static DEFINE_SPINLOCK(balancing);
2921 static void run_rebalance_domains(struct softirq_action *h)
2923 int this_cpu = smp_processor_id(), balance = 1;
2924 struct rq *this_rq = cpu_rq(this_cpu);
2925 unsigned long interval;
2926 struct sched_domain *sd;
2928 * We are idle if there are no processes running. This
2929 * is valid even if we are the idle process (SMT).
2931 enum idle_type idle = !this_rq->nr_running ?
2932 SCHED_IDLE : NOT_IDLE;
2933 /* Earliest time when we have to call run_rebalance_domains again */
2934 unsigned long next_balance = jiffies + 60*HZ;
2936 for_each_domain(this_cpu, sd) {
2937 if (!(sd->flags & SD_LOAD_BALANCE))
2938 continue;
2940 interval = sd->balance_interval;
2941 if (idle != SCHED_IDLE)
2942 interval *= sd->busy_factor;
2944 /* scale ms to jiffies */
2945 interval = msecs_to_jiffies(interval);
2946 if (unlikely(!interval))
2947 interval = 1;
2949 if (sd->flags & SD_SERIALIZE) {
2950 if (!spin_trylock(&balancing))
2951 goto out;
2954 if (time_after_eq(jiffies, sd->last_balance + interval)) {
2955 if (load_balance(this_cpu, this_rq, sd, idle, &balance)) {
2957 * We've pulled tasks over so either we're no
2958 * longer idle, or one of our SMT siblings is
2959 * not idle.
2961 idle = NOT_IDLE;
2963 sd->last_balance = jiffies;
2965 if (sd->flags & SD_SERIALIZE)
2966 spin_unlock(&balancing);
2967 out:
2968 if (time_after(next_balance, sd->last_balance + interval))
2969 next_balance = sd->last_balance + interval;
2972 * Stop the load balance at this level. There is another
2973 * CPU in our sched group which is doing load balancing more
2974 * actively.
2976 if (!balance)
2977 break;
2979 this_rq->next_balance = next_balance;
2981 #else
2983 * on UP we do not need to balance between CPUs:
2985 static inline void idle_balance(int cpu, struct rq *rq)
2988 #endif
2990 static inline void wake_priority_sleeper(struct rq *rq)
2992 #ifdef CONFIG_SCHED_SMT
2993 if (!rq->nr_running)
2994 return;
2996 spin_lock(&rq->lock);
2998 * If an SMT sibling task has been put to sleep for priority
2999 * reasons reschedule the idle task to see if it can now run.
3001 if (rq->nr_running)
3002 resched_task(rq->idle);
3003 spin_unlock(&rq->lock);
3004 #endif
3007 DEFINE_PER_CPU(struct kernel_stat, kstat);
3009 EXPORT_PER_CPU_SYMBOL(kstat);
3012 * This is called on clock ticks and on context switches.
3013 * Bank in p->sched_time the ns elapsed since the last tick or switch.
3015 static inline void
3016 update_cpu_clock(struct task_struct *p, struct rq *rq, unsigned long long now)
3018 p->sched_time += now - p->last_ran;
3019 p->last_ran = rq->most_recent_timestamp = now;
3023 * Return current->sched_time plus any more ns on the sched_clock
3024 * that have not yet been banked.
3026 unsigned long long current_sched_time(const struct task_struct *p)
3028 unsigned long long ns;
3029 unsigned long flags;
3031 local_irq_save(flags);
3032 ns = p->sched_time + sched_clock() - p->last_ran;
3033 local_irq_restore(flags);
3035 return ns;
3039 * We place interactive tasks back into the active array, if possible.
3041 * To guarantee that this does not starve expired tasks we ignore the
3042 * interactivity of a task if the first expired task had to wait more
3043 * than a 'reasonable' amount of time. This deadline timeout is
3044 * load-dependent, as the frequency of array switched decreases with
3045 * increasing number of running tasks. We also ignore the interactivity
3046 * if a better static_prio task has expired:
3048 static inline int expired_starving(struct rq *rq)
3050 if (rq->curr->static_prio > rq->best_expired_prio)
3051 return 1;
3052 if (!STARVATION_LIMIT || !rq->expired_timestamp)
3053 return 0;
3054 if (jiffies - rq->expired_timestamp > STARVATION_LIMIT * rq->nr_running)
3055 return 1;
3056 return 0;
3060 * Account user cpu time to a process.
3061 * @p: the process that the cpu time gets accounted to
3062 * @hardirq_offset: the offset to subtract from hardirq_count()
3063 * @cputime: the cpu time spent in user space since the last update
3065 void account_user_time(struct task_struct *p, cputime_t cputime)
3067 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3068 cputime64_t tmp;
3070 p->utime = cputime_add(p->utime, cputime);
3072 /* Add user time to cpustat. */
3073 tmp = cputime_to_cputime64(cputime);
3074 if (TASK_NICE(p) > 0)
3075 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3076 else
3077 cpustat->user = cputime64_add(cpustat->user, tmp);
3081 * Account system cpu time to a process.
3082 * @p: the process that the cpu time gets accounted to
3083 * @hardirq_offset: the offset to subtract from hardirq_count()
3084 * @cputime: the cpu time spent in kernel space since the last update
3086 void account_system_time(struct task_struct *p, int hardirq_offset,
3087 cputime_t cputime)
3089 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3090 struct rq *rq = this_rq();
3091 cputime64_t tmp;
3093 p->stime = cputime_add(p->stime, cputime);
3095 /* Add system time to cpustat. */
3096 tmp = cputime_to_cputime64(cputime);
3097 if (hardirq_count() - hardirq_offset)
3098 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3099 else if (softirq_count())
3100 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3101 else if (p != rq->idle)
3102 cpustat->system = cputime64_add(cpustat->system, tmp);
3103 else if (atomic_read(&rq->nr_iowait) > 0)
3104 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3105 else
3106 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3107 /* Account for system time used */
3108 acct_update_integrals(p);
3112 * Account for involuntary wait time.
3113 * @p: the process from which the cpu time has been stolen
3114 * @steal: the cpu time spent in involuntary wait
3116 void account_steal_time(struct task_struct *p, cputime_t steal)
3118 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3119 cputime64_t tmp = cputime_to_cputime64(steal);
3120 struct rq *rq = this_rq();
3122 if (p == rq->idle) {
3123 p->stime = cputime_add(p->stime, steal);
3124 if (atomic_read(&rq->nr_iowait) > 0)
3125 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3126 else
3127 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3128 } else
3129 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3132 static void task_running_tick(struct rq *rq, struct task_struct *p)
3134 if (p->array != rq->active) {
3135 /* Task has expired but was not scheduled yet */
3136 set_tsk_need_resched(p);
3137 return;
3139 spin_lock(&rq->lock);
3141 * The task was running during this tick - update the
3142 * time slice counter. Note: we do not update a thread's
3143 * priority until it either goes to sleep or uses up its
3144 * timeslice. This makes it possible for interactive tasks
3145 * to use up their timeslices at their highest priority levels.
3147 if (rt_task(p)) {
3149 * RR tasks need a special form of timeslice management.
3150 * FIFO tasks have no timeslices.
3152 if ((p->policy == SCHED_RR) && !--p->time_slice) {
3153 p->time_slice = task_timeslice(p);
3154 p->first_time_slice = 0;
3155 set_tsk_need_resched(p);
3157 /* put it at the end of the queue: */
3158 requeue_task(p, rq->active);
3160 goto out_unlock;
3162 if (!--p->time_slice) {
3163 dequeue_task(p, rq->active);
3164 set_tsk_need_resched(p);
3165 p->prio = effective_prio(p);
3166 p->time_slice = task_timeslice(p);
3167 p->first_time_slice = 0;
3169 if (!rq->expired_timestamp)
3170 rq->expired_timestamp = jiffies;
3171 if (!TASK_INTERACTIVE(p) || expired_starving(rq)) {
3172 enqueue_task(p, rq->expired);
3173 if (p->static_prio < rq->best_expired_prio)
3174 rq->best_expired_prio = p->static_prio;
3175 } else
3176 enqueue_task(p, rq->active);
3177 } else {
3179 * Prevent a too long timeslice allowing a task to monopolize
3180 * the CPU. We do this by splitting up the timeslice into
3181 * smaller pieces.
3183 * Note: this does not mean the task's timeslices expire or
3184 * get lost in any way, they just might be preempted by
3185 * another task of equal priority. (one with higher
3186 * priority would have preempted this task already.) We
3187 * requeue this task to the end of the list on this priority
3188 * level, which is in essence a round-robin of tasks with
3189 * equal priority.
3191 * This only applies to tasks in the interactive
3192 * delta range with at least TIMESLICE_GRANULARITY to requeue.
3194 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
3195 p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
3196 (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
3197 (p->array == rq->active)) {
3199 requeue_task(p, rq->active);
3200 set_tsk_need_resched(p);
3203 out_unlock:
3204 spin_unlock(&rq->lock);
3208 * This function gets called by the timer code, with HZ frequency.
3209 * We call it with interrupts disabled.
3211 * It also gets called by the fork code, when changing the parent's
3212 * timeslices.
3214 void scheduler_tick(void)
3216 unsigned long long now = sched_clock();
3217 struct task_struct *p = current;
3218 int cpu = smp_processor_id();
3219 struct rq *rq = cpu_rq(cpu);
3221 update_cpu_clock(p, rq, now);
3223 if (p == rq->idle)
3224 /* Task on the idle queue */
3225 wake_priority_sleeper(rq);
3226 else
3227 task_running_tick(rq, p);
3228 #ifdef CONFIG_SMP
3229 update_load(rq);
3230 if (time_after_eq(jiffies, rq->next_balance))
3231 raise_softirq(SCHED_SOFTIRQ);
3232 #endif
3235 #ifdef CONFIG_SCHED_SMT
3236 static inline void wakeup_busy_runqueue(struct rq *rq)
3238 /* If an SMT runqueue is sleeping due to priority reasons wake it up */
3239 if (rq->curr == rq->idle && rq->nr_running)
3240 resched_task(rq->idle);
3244 * Called with interrupt disabled and this_rq's runqueue locked.
3246 static void wake_sleeping_dependent(int this_cpu)
3248 struct sched_domain *tmp, *sd = NULL;
3249 int i;
3251 for_each_domain(this_cpu, tmp) {
3252 if (tmp->flags & SD_SHARE_CPUPOWER) {
3253 sd = tmp;
3254 break;
3258 if (!sd)
3259 return;
3261 for_each_cpu_mask(i, sd->span) {
3262 struct rq *smt_rq = cpu_rq(i);
3264 if (i == this_cpu)
3265 continue;
3266 if (unlikely(!spin_trylock(&smt_rq->lock)))
3267 continue;
3269 wakeup_busy_runqueue(smt_rq);
3270 spin_unlock(&smt_rq->lock);
3275 * number of 'lost' timeslices this task wont be able to fully
3276 * utilize, if another task runs on a sibling. This models the
3277 * slowdown effect of other tasks running on siblings:
3279 static inline unsigned long
3280 smt_slice(struct task_struct *p, struct sched_domain *sd)
3282 return p->time_slice * (100 - sd->per_cpu_gain) / 100;
3286 * To minimise lock contention and not have to drop this_rq's runlock we only
3287 * trylock the sibling runqueues and bypass those runqueues if we fail to
3288 * acquire their lock. As we only trylock the normal locking order does not
3289 * need to be obeyed.
3291 static int
3292 dependent_sleeper(int this_cpu, struct rq *this_rq, struct task_struct *p)
3294 struct sched_domain *tmp, *sd = NULL;
3295 int ret = 0, i;
3297 /* kernel/rt threads do not participate in dependent sleeping */
3298 if (!p->mm || rt_task(p))
3299 return 0;
3301 for_each_domain(this_cpu, tmp) {
3302 if (tmp->flags & SD_SHARE_CPUPOWER) {
3303 sd = tmp;
3304 break;
3308 if (!sd)
3309 return 0;
3311 for_each_cpu_mask(i, sd->span) {
3312 struct task_struct *smt_curr;
3313 struct rq *smt_rq;
3315 if (i == this_cpu)
3316 continue;
3318 smt_rq = cpu_rq(i);
3319 if (unlikely(!spin_trylock(&smt_rq->lock)))
3320 continue;
3322 smt_curr = smt_rq->curr;
3324 if (!smt_curr->mm)
3325 goto unlock;
3328 * If a user task with lower static priority than the
3329 * running task on the SMT sibling is trying to schedule,
3330 * delay it till there is proportionately less timeslice
3331 * left of the sibling task to prevent a lower priority
3332 * task from using an unfair proportion of the
3333 * physical cpu's resources. -ck
3335 if (rt_task(smt_curr)) {
3337 * With real time tasks we run non-rt tasks only
3338 * per_cpu_gain% of the time.
3340 if ((jiffies % DEF_TIMESLICE) >
3341 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
3342 ret = 1;
3343 } else {
3344 if (smt_curr->static_prio < p->static_prio &&
3345 !TASK_PREEMPTS_CURR(p, smt_rq) &&
3346 smt_slice(smt_curr, sd) > task_timeslice(p))
3347 ret = 1;
3349 unlock:
3350 spin_unlock(&smt_rq->lock);
3352 return ret;
3354 #else
3355 static inline void wake_sleeping_dependent(int this_cpu)
3358 static inline int
3359 dependent_sleeper(int this_cpu, struct rq *this_rq, struct task_struct *p)
3361 return 0;
3363 #endif
3365 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3367 void fastcall add_preempt_count(int val)
3370 * Underflow?
3372 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3373 return;
3374 preempt_count() += val;
3376 * Spinlock count overflowing soon?
3378 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3379 PREEMPT_MASK - 10);
3381 EXPORT_SYMBOL(add_preempt_count);
3383 void fastcall sub_preempt_count(int val)
3386 * Underflow?
3388 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3389 return;
3391 * Is the spinlock portion underflowing?
3393 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3394 !(preempt_count() & PREEMPT_MASK)))
3395 return;
3397 preempt_count() -= val;
3399 EXPORT_SYMBOL(sub_preempt_count);
3401 #endif
3403 static inline int interactive_sleep(enum sleep_type sleep_type)
3405 return (sleep_type == SLEEP_INTERACTIVE ||
3406 sleep_type == SLEEP_INTERRUPTED);
3410 * schedule() is the main scheduler function.
3412 asmlinkage void __sched schedule(void)
3414 struct task_struct *prev, *next;
3415 struct prio_array *array;
3416 struct list_head *queue;
3417 unsigned long long now;
3418 unsigned long run_time;
3419 int cpu, idx, new_prio;
3420 long *switch_count;
3421 struct rq *rq;
3424 * Test if we are atomic. Since do_exit() needs to call into
3425 * schedule() atomically, we ignore that path for now.
3426 * Otherwise, whine if we are scheduling when we should not be.
3428 if (unlikely(in_atomic() && !current->exit_state)) {
3429 printk(KERN_ERR "BUG: scheduling while atomic: "
3430 "%s/0x%08x/%d\n",
3431 current->comm, preempt_count(), current->pid);
3432 debug_show_held_locks(current);
3433 if (irqs_disabled())
3434 print_irqtrace_events(current);
3435 dump_stack();
3437 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3439 need_resched:
3440 preempt_disable();
3441 prev = current;
3442 release_kernel_lock(prev);
3443 need_resched_nonpreemptible:
3444 rq = this_rq();
3447 * The idle thread is not allowed to schedule!
3448 * Remove this check after it has been exercised a bit.
3450 if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
3451 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
3452 dump_stack();
3455 schedstat_inc(rq, sched_cnt);
3456 now = sched_clock();
3457 if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
3458 run_time = now - prev->timestamp;
3459 if (unlikely((long long)(now - prev->timestamp) < 0))
3460 run_time = 0;
3461 } else
3462 run_time = NS_MAX_SLEEP_AVG;
3465 * Tasks charged proportionately less run_time at high sleep_avg to
3466 * delay them losing their interactive status
3468 run_time /= (CURRENT_BONUS(prev) ? : 1);
3470 spin_lock_irq(&rq->lock);
3472 switch_count = &prev->nivcsw;
3473 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3474 switch_count = &prev->nvcsw;
3475 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3476 unlikely(signal_pending(prev))))
3477 prev->state = TASK_RUNNING;
3478 else {
3479 if (prev->state == TASK_UNINTERRUPTIBLE)
3480 rq->nr_uninterruptible++;
3481 deactivate_task(prev, rq);
3485 cpu = smp_processor_id();
3486 if (unlikely(!rq->nr_running)) {
3487 idle_balance(cpu, rq);
3488 if (!rq->nr_running) {
3489 next = rq->idle;
3490 rq->expired_timestamp = 0;
3491 wake_sleeping_dependent(cpu);
3492 goto switch_tasks;
3496 array = rq->active;
3497 if (unlikely(!array->nr_active)) {
3499 * Switch the active and expired arrays.
3501 schedstat_inc(rq, sched_switch);
3502 rq->active = rq->expired;
3503 rq->expired = array;
3504 array = rq->active;
3505 rq->expired_timestamp = 0;
3506 rq->best_expired_prio = MAX_PRIO;
3509 idx = sched_find_first_bit(array->bitmap);
3510 queue = array->queue + idx;
3511 next = list_entry(queue->next, struct task_struct, run_list);
3513 if (!rt_task(next) && interactive_sleep(next->sleep_type)) {
3514 unsigned long long delta = now - next->timestamp;
3515 if (unlikely((long long)(now - next->timestamp) < 0))
3516 delta = 0;
3518 if (next->sleep_type == SLEEP_INTERACTIVE)
3519 delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
3521 array = next->array;
3522 new_prio = recalc_task_prio(next, next->timestamp + delta);
3524 if (unlikely(next->prio != new_prio)) {
3525 dequeue_task(next, array);
3526 next->prio = new_prio;
3527 enqueue_task(next, array);
3530 next->sleep_type = SLEEP_NORMAL;
3531 if (dependent_sleeper(cpu, rq, next))
3532 next = rq->idle;
3533 switch_tasks:
3534 if (next == rq->idle)
3535 schedstat_inc(rq, sched_goidle);
3536 prefetch(next);
3537 prefetch_stack(next);
3538 clear_tsk_need_resched(prev);
3539 rcu_qsctr_inc(task_cpu(prev));
3541 update_cpu_clock(prev, rq, now);
3543 prev->sleep_avg -= run_time;
3544 if ((long)prev->sleep_avg <= 0)
3545 prev->sleep_avg = 0;
3546 prev->timestamp = prev->last_ran = now;
3548 sched_info_switch(prev, next);
3549 if (likely(prev != next)) {
3550 next->timestamp = now;
3551 rq->nr_switches++;
3552 rq->curr = next;
3553 ++*switch_count;
3555 prepare_task_switch(rq, next);
3556 prev = context_switch(rq, prev, next);
3557 barrier();
3559 * this_rq must be evaluated again because prev may have moved
3560 * CPUs since it called schedule(), thus the 'rq' on its stack
3561 * frame will be invalid.
3563 finish_task_switch(this_rq(), prev);
3564 } else
3565 spin_unlock_irq(&rq->lock);
3567 prev = current;
3568 if (unlikely(reacquire_kernel_lock(prev) < 0))
3569 goto need_resched_nonpreemptible;
3570 preempt_enable_no_resched();
3571 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3572 goto need_resched;
3574 EXPORT_SYMBOL(schedule);
3576 #ifdef CONFIG_PREEMPT
3578 * this is the entry point to schedule() from in-kernel preemption
3579 * off of preempt_enable. Kernel preemptions off return from interrupt
3580 * occur there and call schedule directly.
3582 asmlinkage void __sched preempt_schedule(void)
3584 struct thread_info *ti = current_thread_info();
3585 #ifdef CONFIG_PREEMPT_BKL
3586 struct task_struct *task = current;
3587 int saved_lock_depth;
3588 #endif
3590 * If there is a non-zero preempt_count or interrupts are disabled,
3591 * we do not want to preempt the current task. Just return..
3593 if (likely(ti->preempt_count || irqs_disabled()))
3594 return;
3596 need_resched:
3597 add_preempt_count(PREEMPT_ACTIVE);
3599 * We keep the big kernel semaphore locked, but we
3600 * clear ->lock_depth so that schedule() doesnt
3601 * auto-release the semaphore:
3603 #ifdef CONFIG_PREEMPT_BKL
3604 saved_lock_depth = task->lock_depth;
3605 task->lock_depth = -1;
3606 #endif
3607 schedule();
3608 #ifdef CONFIG_PREEMPT_BKL
3609 task->lock_depth = saved_lock_depth;
3610 #endif
3611 sub_preempt_count(PREEMPT_ACTIVE);
3613 /* we could miss a preemption opportunity between schedule and now */
3614 barrier();
3615 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3616 goto need_resched;
3618 EXPORT_SYMBOL(preempt_schedule);
3621 * this is the entry point to schedule() from kernel preemption
3622 * off of irq context.
3623 * Note, that this is called and return with irqs disabled. This will
3624 * protect us against recursive calling from irq.
3626 asmlinkage void __sched preempt_schedule_irq(void)
3628 struct thread_info *ti = current_thread_info();
3629 #ifdef CONFIG_PREEMPT_BKL
3630 struct task_struct *task = current;
3631 int saved_lock_depth;
3632 #endif
3633 /* Catch callers which need to be fixed */
3634 BUG_ON(ti->preempt_count || !irqs_disabled());
3636 need_resched:
3637 add_preempt_count(PREEMPT_ACTIVE);
3639 * We keep the big kernel semaphore locked, but we
3640 * clear ->lock_depth so that schedule() doesnt
3641 * auto-release the semaphore:
3643 #ifdef CONFIG_PREEMPT_BKL
3644 saved_lock_depth = task->lock_depth;
3645 task->lock_depth = -1;
3646 #endif
3647 local_irq_enable();
3648 schedule();
3649 local_irq_disable();
3650 #ifdef CONFIG_PREEMPT_BKL
3651 task->lock_depth = saved_lock_depth;
3652 #endif
3653 sub_preempt_count(PREEMPT_ACTIVE);
3655 /* we could miss a preemption opportunity between schedule and now */
3656 barrier();
3657 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3658 goto need_resched;
3661 #endif /* CONFIG_PREEMPT */
3663 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3664 void *key)
3666 return try_to_wake_up(curr->private, mode, sync);
3668 EXPORT_SYMBOL(default_wake_function);
3671 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3672 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3673 * number) then we wake all the non-exclusive tasks and one exclusive task.
3675 * There are circumstances in which we can try to wake a task which has already
3676 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3677 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3679 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3680 int nr_exclusive, int sync, void *key)
3682 struct list_head *tmp, *next;
3684 list_for_each_safe(tmp, next, &q->task_list) {
3685 wait_queue_t *curr = list_entry(tmp, wait_queue_t, task_list);
3686 unsigned flags = curr->flags;
3688 if (curr->func(curr, mode, sync, key) &&
3689 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3690 break;
3695 * __wake_up - wake up threads blocked on a waitqueue.
3696 * @q: the waitqueue
3697 * @mode: which threads
3698 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3699 * @key: is directly passed to the wakeup function
3701 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3702 int nr_exclusive, void *key)
3704 unsigned long flags;
3706 spin_lock_irqsave(&q->lock, flags);
3707 __wake_up_common(q, mode, nr_exclusive, 0, key);
3708 spin_unlock_irqrestore(&q->lock, flags);
3710 EXPORT_SYMBOL(__wake_up);
3713 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3715 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3717 __wake_up_common(q, mode, 1, 0, NULL);
3721 * __wake_up_sync - wake up threads blocked on a waitqueue.
3722 * @q: the waitqueue
3723 * @mode: which threads
3724 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3726 * The sync wakeup differs that the waker knows that it will schedule
3727 * away soon, so while the target thread will be woken up, it will not
3728 * be migrated to another CPU - ie. the two threads are 'synchronized'
3729 * with each other. This can prevent needless bouncing between CPUs.
3731 * On UP it can prevent extra preemption.
3733 void fastcall
3734 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3736 unsigned long flags;
3737 int sync = 1;
3739 if (unlikely(!q))
3740 return;
3742 if (unlikely(!nr_exclusive))
3743 sync = 0;
3745 spin_lock_irqsave(&q->lock, flags);
3746 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3747 spin_unlock_irqrestore(&q->lock, flags);
3749 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3751 void fastcall complete(struct completion *x)
3753 unsigned long flags;
3755 spin_lock_irqsave(&x->wait.lock, flags);
3756 x->done++;
3757 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3758 1, 0, NULL);
3759 spin_unlock_irqrestore(&x->wait.lock, flags);
3761 EXPORT_SYMBOL(complete);
3763 void fastcall complete_all(struct completion *x)
3765 unsigned long flags;
3767 spin_lock_irqsave(&x->wait.lock, flags);
3768 x->done += UINT_MAX/2;
3769 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3770 0, 0, NULL);
3771 spin_unlock_irqrestore(&x->wait.lock, flags);
3773 EXPORT_SYMBOL(complete_all);
3775 void fastcall __sched wait_for_completion(struct completion *x)
3777 might_sleep();
3779 spin_lock_irq(&x->wait.lock);
3780 if (!x->done) {
3781 DECLARE_WAITQUEUE(wait, current);
3783 wait.flags |= WQ_FLAG_EXCLUSIVE;
3784 __add_wait_queue_tail(&x->wait, &wait);
3785 do {
3786 __set_current_state(TASK_UNINTERRUPTIBLE);
3787 spin_unlock_irq(&x->wait.lock);
3788 schedule();
3789 spin_lock_irq(&x->wait.lock);
3790 } while (!x->done);
3791 __remove_wait_queue(&x->wait, &wait);
3793 x->done--;
3794 spin_unlock_irq(&x->wait.lock);
3796 EXPORT_SYMBOL(wait_for_completion);
3798 unsigned long fastcall __sched
3799 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3801 might_sleep();
3803 spin_lock_irq(&x->wait.lock);
3804 if (!x->done) {
3805 DECLARE_WAITQUEUE(wait, current);
3807 wait.flags |= WQ_FLAG_EXCLUSIVE;
3808 __add_wait_queue_tail(&x->wait, &wait);
3809 do {
3810 __set_current_state(TASK_UNINTERRUPTIBLE);
3811 spin_unlock_irq(&x->wait.lock);
3812 timeout = schedule_timeout(timeout);
3813 spin_lock_irq(&x->wait.lock);
3814 if (!timeout) {
3815 __remove_wait_queue(&x->wait, &wait);
3816 goto out;
3818 } while (!x->done);
3819 __remove_wait_queue(&x->wait, &wait);
3821 x->done--;
3822 out:
3823 spin_unlock_irq(&x->wait.lock);
3824 return timeout;
3826 EXPORT_SYMBOL(wait_for_completion_timeout);
3828 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3830 int ret = 0;
3832 might_sleep();
3834 spin_lock_irq(&x->wait.lock);
3835 if (!x->done) {
3836 DECLARE_WAITQUEUE(wait, current);
3838 wait.flags |= WQ_FLAG_EXCLUSIVE;
3839 __add_wait_queue_tail(&x->wait, &wait);
3840 do {
3841 if (signal_pending(current)) {
3842 ret = -ERESTARTSYS;
3843 __remove_wait_queue(&x->wait, &wait);
3844 goto out;
3846 __set_current_state(TASK_INTERRUPTIBLE);
3847 spin_unlock_irq(&x->wait.lock);
3848 schedule();
3849 spin_lock_irq(&x->wait.lock);
3850 } while (!x->done);
3851 __remove_wait_queue(&x->wait, &wait);
3853 x->done--;
3854 out:
3855 spin_unlock_irq(&x->wait.lock);
3857 return ret;
3859 EXPORT_SYMBOL(wait_for_completion_interruptible);
3861 unsigned long fastcall __sched
3862 wait_for_completion_interruptible_timeout(struct completion *x,
3863 unsigned long timeout)
3865 might_sleep();
3867 spin_lock_irq(&x->wait.lock);
3868 if (!x->done) {
3869 DECLARE_WAITQUEUE(wait, current);
3871 wait.flags |= WQ_FLAG_EXCLUSIVE;
3872 __add_wait_queue_tail(&x->wait, &wait);
3873 do {
3874 if (signal_pending(current)) {
3875 timeout = -ERESTARTSYS;
3876 __remove_wait_queue(&x->wait, &wait);
3877 goto out;
3879 __set_current_state(TASK_INTERRUPTIBLE);
3880 spin_unlock_irq(&x->wait.lock);
3881 timeout = schedule_timeout(timeout);
3882 spin_lock_irq(&x->wait.lock);
3883 if (!timeout) {
3884 __remove_wait_queue(&x->wait, &wait);
3885 goto out;
3887 } while (!x->done);
3888 __remove_wait_queue(&x->wait, &wait);
3890 x->done--;
3891 out:
3892 spin_unlock_irq(&x->wait.lock);
3893 return timeout;
3895 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3898 #define SLEEP_ON_VAR \
3899 unsigned long flags; \
3900 wait_queue_t wait; \
3901 init_waitqueue_entry(&wait, current);
3903 #define SLEEP_ON_HEAD \
3904 spin_lock_irqsave(&q->lock,flags); \
3905 __add_wait_queue(q, &wait); \
3906 spin_unlock(&q->lock);
3908 #define SLEEP_ON_TAIL \
3909 spin_lock_irq(&q->lock); \
3910 __remove_wait_queue(q, &wait); \
3911 spin_unlock_irqrestore(&q->lock, flags);
3913 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
3915 SLEEP_ON_VAR
3917 current->state = TASK_INTERRUPTIBLE;
3919 SLEEP_ON_HEAD
3920 schedule();
3921 SLEEP_ON_TAIL
3923 EXPORT_SYMBOL(interruptible_sleep_on);
3925 long fastcall __sched
3926 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3928 SLEEP_ON_VAR
3930 current->state = TASK_INTERRUPTIBLE;
3932 SLEEP_ON_HEAD
3933 timeout = schedule_timeout(timeout);
3934 SLEEP_ON_TAIL
3936 return timeout;
3938 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3940 void fastcall __sched sleep_on(wait_queue_head_t *q)
3942 SLEEP_ON_VAR
3944 current->state = TASK_UNINTERRUPTIBLE;
3946 SLEEP_ON_HEAD
3947 schedule();
3948 SLEEP_ON_TAIL
3950 EXPORT_SYMBOL(sleep_on);
3952 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3954 SLEEP_ON_VAR
3956 current->state = TASK_UNINTERRUPTIBLE;
3958 SLEEP_ON_HEAD
3959 timeout = schedule_timeout(timeout);
3960 SLEEP_ON_TAIL
3962 return timeout;
3965 EXPORT_SYMBOL(sleep_on_timeout);
3967 #ifdef CONFIG_RT_MUTEXES
3970 * rt_mutex_setprio - set the current priority of a task
3971 * @p: task
3972 * @prio: prio value (kernel-internal form)
3974 * This function changes the 'effective' priority of a task. It does
3975 * not touch ->normal_prio like __setscheduler().
3977 * Used by the rt_mutex code to implement priority inheritance logic.
3979 void rt_mutex_setprio(struct task_struct *p, int prio)
3981 struct prio_array *array;
3982 unsigned long flags;
3983 struct rq *rq;
3984 int oldprio;
3986 BUG_ON(prio < 0 || prio > MAX_PRIO);
3988 rq = task_rq_lock(p, &flags);
3990 oldprio = p->prio;
3991 array = p->array;
3992 if (array)
3993 dequeue_task(p, array);
3994 p->prio = prio;
3996 if (array) {
3998 * If changing to an RT priority then queue it
3999 * in the active array!
4001 if (rt_task(p))
4002 array = rq->active;
4003 enqueue_task(p, array);
4005 * Reschedule if we are currently running on this runqueue and
4006 * our priority decreased, or if we are not currently running on
4007 * this runqueue and our priority is higher than the current's
4009 if (task_running(rq, p)) {
4010 if (p->prio > oldprio)
4011 resched_task(rq->curr);
4012 } else if (TASK_PREEMPTS_CURR(p, rq))
4013 resched_task(rq->curr);
4015 task_rq_unlock(rq, &flags);
4018 #endif
4020 void set_user_nice(struct task_struct *p, long nice)
4022 struct prio_array *array;
4023 int old_prio, delta;
4024 unsigned long flags;
4025 struct rq *rq;
4027 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4028 return;
4030 * We have to be careful, if called from sys_setpriority(),
4031 * the task might be in the middle of scheduling on another CPU.
4033 rq = task_rq_lock(p, &flags);
4035 * The RT priorities are set via sched_setscheduler(), but we still
4036 * allow the 'normal' nice value to be set - but as expected
4037 * it wont have any effect on scheduling until the task is
4038 * not SCHED_NORMAL/SCHED_BATCH:
4040 if (has_rt_policy(p)) {
4041 p->static_prio = NICE_TO_PRIO(nice);
4042 goto out_unlock;
4044 array = p->array;
4045 if (array) {
4046 dequeue_task(p, array);
4047 dec_raw_weighted_load(rq, p);
4050 p->static_prio = NICE_TO_PRIO(nice);
4051 set_load_weight(p);
4052 old_prio = p->prio;
4053 p->prio = effective_prio(p);
4054 delta = p->prio - old_prio;
4056 if (array) {
4057 enqueue_task(p, array);
4058 inc_raw_weighted_load(rq, p);
4060 * If the task increased its priority or is running and
4061 * lowered its priority, then reschedule its CPU:
4063 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4064 resched_task(rq->curr);
4066 out_unlock:
4067 task_rq_unlock(rq, &flags);
4069 EXPORT_SYMBOL(set_user_nice);
4072 * can_nice - check if a task can reduce its nice value
4073 * @p: task
4074 * @nice: nice value
4076 int can_nice(const struct task_struct *p, const int nice)
4078 /* convert nice value [19,-20] to rlimit style value [1,40] */
4079 int nice_rlim = 20 - nice;
4081 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4082 capable(CAP_SYS_NICE));
4085 #ifdef __ARCH_WANT_SYS_NICE
4088 * sys_nice - change the priority of the current process.
4089 * @increment: priority increment
4091 * sys_setpriority is a more generic, but much slower function that
4092 * does similar things.
4094 asmlinkage long sys_nice(int increment)
4096 long nice, retval;
4099 * Setpriority might change our priority at the same moment.
4100 * We don't have to worry. Conceptually one call occurs first
4101 * and we have a single winner.
4103 if (increment < -40)
4104 increment = -40;
4105 if (increment > 40)
4106 increment = 40;
4108 nice = PRIO_TO_NICE(current->static_prio) + increment;
4109 if (nice < -20)
4110 nice = -20;
4111 if (nice > 19)
4112 nice = 19;
4114 if (increment < 0 && !can_nice(current, nice))
4115 return -EPERM;
4117 retval = security_task_setnice(current, nice);
4118 if (retval)
4119 return retval;
4121 set_user_nice(current, nice);
4122 return 0;
4125 #endif
4128 * task_prio - return the priority value of a given task.
4129 * @p: the task in question.
4131 * This is the priority value as seen by users in /proc.
4132 * RT tasks are offset by -200. Normal tasks are centered
4133 * around 0, value goes from -16 to +15.
4135 int task_prio(const struct task_struct *p)
4137 return p->prio - MAX_RT_PRIO;
4141 * task_nice - return the nice value of a given task.
4142 * @p: the task in question.
4144 int task_nice(const struct task_struct *p)
4146 return TASK_NICE(p);
4148 EXPORT_SYMBOL_GPL(task_nice);
4151 * idle_cpu - is a given cpu idle currently?
4152 * @cpu: the processor in question.
4154 int idle_cpu(int cpu)
4156 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4160 * idle_task - return the idle task for a given cpu.
4161 * @cpu: the processor in question.
4163 struct task_struct *idle_task(int cpu)
4165 return cpu_rq(cpu)->idle;
4169 * find_process_by_pid - find a process with a matching PID value.
4170 * @pid: the pid in question.
4172 static inline struct task_struct *find_process_by_pid(pid_t pid)
4174 return pid ? find_task_by_pid(pid) : current;
4177 /* Actually do priority change: must hold rq lock. */
4178 static void __setscheduler(struct task_struct *p, int policy, int prio)
4180 BUG_ON(p->array);
4182 p->policy = policy;
4183 p->rt_priority = prio;
4184 p->normal_prio = normal_prio(p);
4185 /* we are holding p->pi_lock already */
4186 p->prio = rt_mutex_getprio(p);
4188 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
4190 if (policy == SCHED_BATCH)
4191 p->sleep_avg = 0;
4192 set_load_weight(p);
4196 * sched_setscheduler - change the scheduling policy and/or RT priority of
4197 * a thread.
4198 * @p: the task in question.
4199 * @policy: new policy.
4200 * @param: structure containing the new RT priority.
4202 * NOTE: the task may be already dead
4204 int sched_setscheduler(struct task_struct *p, int policy,
4205 struct sched_param *param)
4207 int retval, oldprio, oldpolicy = -1;
4208 struct prio_array *array;
4209 unsigned long flags;
4210 struct rq *rq;
4212 /* may grab non-irq protected spin_locks */
4213 BUG_ON(in_interrupt());
4214 recheck:
4215 /* double check policy once rq lock held */
4216 if (policy < 0)
4217 policy = oldpolicy = p->policy;
4218 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4219 policy != SCHED_NORMAL && policy != SCHED_BATCH)
4220 return -EINVAL;
4222 * Valid priorities for SCHED_FIFO and SCHED_RR are
4223 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
4224 * SCHED_BATCH is 0.
4226 if (param->sched_priority < 0 ||
4227 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4228 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4229 return -EINVAL;
4230 if (is_rt_policy(policy) != (param->sched_priority != 0))
4231 return -EINVAL;
4234 * Allow unprivileged RT tasks to decrease priority:
4236 if (!capable(CAP_SYS_NICE)) {
4237 if (is_rt_policy(policy)) {
4238 unsigned long rlim_rtprio;
4239 unsigned long flags;
4241 if (!lock_task_sighand(p, &flags))
4242 return -ESRCH;
4243 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4244 unlock_task_sighand(p, &flags);
4246 /* can't set/change the rt policy */
4247 if (policy != p->policy && !rlim_rtprio)
4248 return -EPERM;
4250 /* can't increase priority */
4251 if (param->sched_priority > p->rt_priority &&
4252 param->sched_priority > rlim_rtprio)
4253 return -EPERM;
4256 /* can't change other user's priorities */
4257 if ((current->euid != p->euid) &&
4258 (current->euid != p->uid))
4259 return -EPERM;
4262 retval = security_task_setscheduler(p, policy, param);
4263 if (retval)
4264 return retval;
4266 * make sure no PI-waiters arrive (or leave) while we are
4267 * changing the priority of the task:
4269 spin_lock_irqsave(&p->pi_lock, flags);
4271 * To be able to change p->policy safely, the apropriate
4272 * runqueue lock must be held.
4274 rq = __task_rq_lock(p);
4275 /* recheck policy now with rq lock held */
4276 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4277 policy = oldpolicy = -1;
4278 __task_rq_unlock(rq);
4279 spin_unlock_irqrestore(&p->pi_lock, flags);
4280 goto recheck;
4282 array = p->array;
4283 if (array)
4284 deactivate_task(p, rq);
4285 oldprio = p->prio;
4286 __setscheduler(p, policy, param->sched_priority);
4287 if (array) {
4288 __activate_task(p, rq);
4290 * Reschedule if we are currently running on this runqueue and
4291 * our priority decreased, or if we are not currently running on
4292 * this runqueue and our priority is higher than the current's
4294 if (task_running(rq, p)) {
4295 if (p->prio > oldprio)
4296 resched_task(rq->curr);
4297 } else if (TASK_PREEMPTS_CURR(p, rq))
4298 resched_task(rq->curr);
4300 __task_rq_unlock(rq);
4301 spin_unlock_irqrestore(&p->pi_lock, flags);
4303 rt_mutex_adjust_pi(p);
4305 return 0;
4307 EXPORT_SYMBOL_GPL(sched_setscheduler);
4309 static int
4310 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4312 struct sched_param lparam;
4313 struct task_struct *p;
4314 int retval;
4316 if (!param || pid < 0)
4317 return -EINVAL;
4318 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4319 return -EFAULT;
4321 rcu_read_lock();
4322 retval = -ESRCH;
4323 p = find_process_by_pid(pid);
4324 if (p != NULL)
4325 retval = sched_setscheduler(p, policy, &lparam);
4326 rcu_read_unlock();
4328 return retval;
4332 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4333 * @pid: the pid in question.
4334 * @policy: new policy.
4335 * @param: structure containing the new RT priority.
4337 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4338 struct sched_param __user *param)
4340 /* negative values for policy are not valid */
4341 if (policy < 0)
4342 return -EINVAL;
4344 return do_sched_setscheduler(pid, policy, param);
4348 * sys_sched_setparam - set/change the RT priority of a thread
4349 * @pid: the pid in question.
4350 * @param: structure containing the new RT priority.
4352 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4354 return do_sched_setscheduler(pid, -1, param);
4358 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4359 * @pid: the pid in question.
4361 asmlinkage long sys_sched_getscheduler(pid_t pid)
4363 struct task_struct *p;
4364 int retval = -EINVAL;
4366 if (pid < 0)
4367 goto out_nounlock;
4369 retval = -ESRCH;
4370 read_lock(&tasklist_lock);
4371 p = find_process_by_pid(pid);
4372 if (p) {
4373 retval = security_task_getscheduler(p);
4374 if (!retval)
4375 retval = p->policy;
4377 read_unlock(&tasklist_lock);
4379 out_nounlock:
4380 return retval;
4384 * sys_sched_getscheduler - get the RT priority of a thread
4385 * @pid: the pid in question.
4386 * @param: structure containing the RT priority.
4388 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4390 struct sched_param lp;
4391 struct task_struct *p;
4392 int retval = -EINVAL;
4394 if (!param || pid < 0)
4395 goto out_nounlock;
4397 read_lock(&tasklist_lock);
4398 p = find_process_by_pid(pid);
4399 retval = -ESRCH;
4400 if (!p)
4401 goto out_unlock;
4403 retval = security_task_getscheduler(p);
4404 if (retval)
4405 goto out_unlock;
4407 lp.sched_priority = p->rt_priority;
4408 read_unlock(&tasklist_lock);
4411 * This one might sleep, we cannot do it with a spinlock held ...
4413 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4415 out_nounlock:
4416 return retval;
4418 out_unlock:
4419 read_unlock(&tasklist_lock);
4420 return retval;
4423 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4425 cpumask_t cpus_allowed;
4426 struct task_struct *p;
4427 int retval;
4429 lock_cpu_hotplug();
4430 read_lock(&tasklist_lock);
4432 p = find_process_by_pid(pid);
4433 if (!p) {
4434 read_unlock(&tasklist_lock);
4435 unlock_cpu_hotplug();
4436 return -ESRCH;
4440 * It is not safe to call set_cpus_allowed with the
4441 * tasklist_lock held. We will bump the task_struct's
4442 * usage count and then drop tasklist_lock.
4444 get_task_struct(p);
4445 read_unlock(&tasklist_lock);
4447 retval = -EPERM;
4448 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4449 !capable(CAP_SYS_NICE))
4450 goto out_unlock;
4452 retval = security_task_setscheduler(p, 0, NULL);
4453 if (retval)
4454 goto out_unlock;
4456 cpus_allowed = cpuset_cpus_allowed(p);
4457 cpus_and(new_mask, new_mask, cpus_allowed);
4458 retval = set_cpus_allowed(p, new_mask);
4460 out_unlock:
4461 put_task_struct(p);
4462 unlock_cpu_hotplug();
4463 return retval;
4466 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4467 cpumask_t *new_mask)
4469 if (len < sizeof(cpumask_t)) {
4470 memset(new_mask, 0, sizeof(cpumask_t));
4471 } else if (len > sizeof(cpumask_t)) {
4472 len = sizeof(cpumask_t);
4474 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4478 * sys_sched_setaffinity - set the cpu affinity of a process
4479 * @pid: pid of the process
4480 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4481 * @user_mask_ptr: user-space pointer to the new cpu mask
4483 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4484 unsigned long __user *user_mask_ptr)
4486 cpumask_t new_mask;
4487 int retval;
4489 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4490 if (retval)
4491 return retval;
4493 return sched_setaffinity(pid, new_mask);
4497 * Represents all cpu's present in the system
4498 * In systems capable of hotplug, this map could dynamically grow
4499 * as new cpu's are detected in the system via any platform specific
4500 * method, such as ACPI for e.g.
4503 cpumask_t cpu_present_map __read_mostly;
4504 EXPORT_SYMBOL(cpu_present_map);
4506 #ifndef CONFIG_SMP
4507 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4508 EXPORT_SYMBOL(cpu_online_map);
4510 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4511 EXPORT_SYMBOL(cpu_possible_map);
4512 #endif
4514 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4516 struct task_struct *p;
4517 int retval;
4519 lock_cpu_hotplug();
4520 read_lock(&tasklist_lock);
4522 retval = -ESRCH;
4523 p = find_process_by_pid(pid);
4524 if (!p)
4525 goto out_unlock;
4527 retval = security_task_getscheduler(p);
4528 if (retval)
4529 goto out_unlock;
4531 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4533 out_unlock:
4534 read_unlock(&tasklist_lock);
4535 unlock_cpu_hotplug();
4536 if (retval)
4537 return retval;
4539 return 0;
4543 * sys_sched_getaffinity - get the cpu affinity of a process
4544 * @pid: pid of the process
4545 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4546 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4548 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4549 unsigned long __user *user_mask_ptr)
4551 int ret;
4552 cpumask_t mask;
4554 if (len < sizeof(cpumask_t))
4555 return -EINVAL;
4557 ret = sched_getaffinity(pid, &mask);
4558 if (ret < 0)
4559 return ret;
4561 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4562 return -EFAULT;
4564 return sizeof(cpumask_t);
4568 * sys_sched_yield - yield the current processor to other threads.
4570 * this function yields the current CPU by moving the calling thread
4571 * to the expired array. If there are no other threads running on this
4572 * CPU then this function will return.
4574 asmlinkage long sys_sched_yield(void)
4576 struct rq *rq = this_rq_lock();
4577 struct prio_array *array = current->array, *target = rq->expired;
4579 schedstat_inc(rq, yld_cnt);
4581 * We implement yielding by moving the task into the expired
4582 * queue.
4584 * (special rule: RT tasks will just roundrobin in the active
4585 * array.)
4587 if (rt_task(current))
4588 target = rq->active;
4590 if (array->nr_active == 1) {
4591 schedstat_inc(rq, yld_act_empty);
4592 if (!rq->expired->nr_active)
4593 schedstat_inc(rq, yld_both_empty);
4594 } else if (!rq->expired->nr_active)
4595 schedstat_inc(rq, yld_exp_empty);
4597 if (array != target) {
4598 dequeue_task(current, array);
4599 enqueue_task(current, target);
4600 } else
4602 * requeue_task is cheaper so perform that if possible.
4604 requeue_task(current, array);
4607 * Since we are going to call schedule() anyway, there's
4608 * no need to preempt or enable interrupts:
4610 __release(rq->lock);
4611 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4612 _raw_spin_unlock(&rq->lock);
4613 preempt_enable_no_resched();
4615 schedule();
4617 return 0;
4620 static void __cond_resched(void)
4622 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4623 __might_sleep(__FILE__, __LINE__);
4624 #endif
4626 * The BKS might be reacquired before we have dropped
4627 * PREEMPT_ACTIVE, which could trigger a second
4628 * cond_resched() call.
4630 do {
4631 add_preempt_count(PREEMPT_ACTIVE);
4632 schedule();
4633 sub_preempt_count(PREEMPT_ACTIVE);
4634 } while (need_resched());
4637 int __sched cond_resched(void)
4639 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4640 system_state == SYSTEM_RUNNING) {
4641 __cond_resched();
4642 return 1;
4644 return 0;
4646 EXPORT_SYMBOL(cond_resched);
4649 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4650 * call schedule, and on return reacquire the lock.
4652 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4653 * operations here to prevent schedule() from being called twice (once via
4654 * spin_unlock(), once by hand).
4656 int cond_resched_lock(spinlock_t *lock)
4658 int ret = 0;
4660 if (need_lockbreak(lock)) {
4661 spin_unlock(lock);
4662 cpu_relax();
4663 ret = 1;
4664 spin_lock(lock);
4666 if (need_resched() && system_state == SYSTEM_RUNNING) {
4667 spin_release(&lock->dep_map, 1, _THIS_IP_);
4668 _raw_spin_unlock(lock);
4669 preempt_enable_no_resched();
4670 __cond_resched();
4671 ret = 1;
4672 spin_lock(lock);
4674 return ret;
4676 EXPORT_SYMBOL(cond_resched_lock);
4678 int __sched cond_resched_softirq(void)
4680 BUG_ON(!in_softirq());
4682 if (need_resched() && system_state == SYSTEM_RUNNING) {
4683 raw_local_irq_disable();
4684 _local_bh_enable();
4685 raw_local_irq_enable();
4686 __cond_resched();
4687 local_bh_disable();
4688 return 1;
4690 return 0;
4692 EXPORT_SYMBOL(cond_resched_softirq);
4695 * yield - yield the current processor to other threads.
4697 * this is a shortcut for kernel-space yielding - it marks the
4698 * thread runnable and calls sys_sched_yield().
4700 void __sched yield(void)
4702 set_current_state(TASK_RUNNING);
4703 sys_sched_yield();
4705 EXPORT_SYMBOL(yield);
4708 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4709 * that process accounting knows that this is a task in IO wait state.
4711 * But don't do that if it is a deliberate, throttling IO wait (this task
4712 * has set its backing_dev_info: the queue against which it should throttle)
4714 void __sched io_schedule(void)
4716 struct rq *rq = &__raw_get_cpu_var(runqueues);
4718 delayacct_blkio_start();
4719 atomic_inc(&rq->nr_iowait);
4720 schedule();
4721 atomic_dec(&rq->nr_iowait);
4722 delayacct_blkio_end();
4724 EXPORT_SYMBOL(io_schedule);
4726 long __sched io_schedule_timeout(long timeout)
4728 struct rq *rq = &__raw_get_cpu_var(runqueues);
4729 long ret;
4731 delayacct_blkio_start();
4732 atomic_inc(&rq->nr_iowait);
4733 ret = schedule_timeout(timeout);
4734 atomic_dec(&rq->nr_iowait);
4735 delayacct_blkio_end();
4736 return ret;
4740 * sys_sched_get_priority_max - return maximum RT priority.
4741 * @policy: scheduling class.
4743 * this syscall returns the maximum rt_priority that can be used
4744 * by a given scheduling class.
4746 asmlinkage long sys_sched_get_priority_max(int policy)
4748 int ret = -EINVAL;
4750 switch (policy) {
4751 case SCHED_FIFO:
4752 case SCHED_RR:
4753 ret = MAX_USER_RT_PRIO-1;
4754 break;
4755 case SCHED_NORMAL:
4756 case SCHED_BATCH:
4757 ret = 0;
4758 break;
4760 return ret;
4764 * sys_sched_get_priority_min - return minimum RT priority.
4765 * @policy: scheduling class.
4767 * this syscall returns the minimum rt_priority that can be used
4768 * by a given scheduling class.
4770 asmlinkage long sys_sched_get_priority_min(int policy)
4772 int ret = -EINVAL;
4774 switch (policy) {
4775 case SCHED_FIFO:
4776 case SCHED_RR:
4777 ret = 1;
4778 break;
4779 case SCHED_NORMAL:
4780 case SCHED_BATCH:
4781 ret = 0;
4783 return ret;
4787 * sys_sched_rr_get_interval - return the default timeslice of a process.
4788 * @pid: pid of the process.
4789 * @interval: userspace pointer to the timeslice value.
4791 * this syscall writes the default timeslice value of a given process
4792 * into the user-space timespec buffer. A value of '0' means infinity.
4794 asmlinkage
4795 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4797 struct task_struct *p;
4798 int retval = -EINVAL;
4799 struct timespec t;
4801 if (pid < 0)
4802 goto out_nounlock;
4804 retval = -ESRCH;
4805 read_lock(&tasklist_lock);
4806 p = find_process_by_pid(pid);
4807 if (!p)
4808 goto out_unlock;
4810 retval = security_task_getscheduler(p);
4811 if (retval)
4812 goto out_unlock;
4814 jiffies_to_timespec(p->policy == SCHED_FIFO ?
4815 0 : task_timeslice(p), &t);
4816 read_unlock(&tasklist_lock);
4817 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4818 out_nounlock:
4819 return retval;
4820 out_unlock:
4821 read_unlock(&tasklist_lock);
4822 return retval;
4825 static inline struct task_struct *eldest_child(struct task_struct *p)
4827 if (list_empty(&p->children))
4828 return NULL;
4829 return list_entry(p->children.next,struct task_struct,sibling);
4832 static inline struct task_struct *older_sibling(struct task_struct *p)
4834 if (p->sibling.prev==&p->parent->children)
4835 return NULL;
4836 return list_entry(p->sibling.prev,struct task_struct,sibling);
4839 static inline struct task_struct *younger_sibling(struct task_struct *p)
4841 if (p->sibling.next==&p->parent->children)
4842 return NULL;
4843 return list_entry(p->sibling.next,struct task_struct,sibling);
4846 static const char stat_nam[] = "RSDTtZX";
4848 static void show_task(struct task_struct *p)
4850 struct task_struct *relative;
4851 unsigned long free = 0;
4852 unsigned state;
4854 state = p->state ? __ffs(p->state) + 1 : 0;
4855 printk("%-13.13s %c", p->comm,
4856 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4857 #if (BITS_PER_LONG == 32)
4858 if (state == TASK_RUNNING)
4859 printk(" running ");
4860 else
4861 printk(" %08lX ", thread_saved_pc(p));
4862 #else
4863 if (state == TASK_RUNNING)
4864 printk(" running task ");
4865 else
4866 printk(" %016lx ", thread_saved_pc(p));
4867 #endif
4868 #ifdef CONFIG_DEBUG_STACK_USAGE
4870 unsigned long *n = end_of_stack(p);
4871 while (!*n)
4872 n++;
4873 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4875 #endif
4876 printk("%5lu %5d %6d ", free, p->pid, p->parent->pid);
4877 if ((relative = eldest_child(p)))
4878 printk("%5d ", relative->pid);
4879 else
4880 printk(" ");
4881 if ((relative = younger_sibling(p)))
4882 printk("%7d", relative->pid);
4883 else
4884 printk(" ");
4885 if ((relative = older_sibling(p)))
4886 printk(" %5d", relative->pid);
4887 else
4888 printk(" ");
4889 if (!p->mm)
4890 printk(" (L-TLB)\n");
4891 else
4892 printk(" (NOTLB)\n");
4894 if (state != TASK_RUNNING)
4895 show_stack(p, NULL);
4898 void show_state_filter(unsigned long state_filter)
4900 struct task_struct *g, *p;
4902 #if (BITS_PER_LONG == 32)
4903 printk("\n"
4904 " free sibling\n");
4905 printk(" task PC stack pid father child younger older\n");
4906 #else
4907 printk("\n"
4908 " free sibling\n");
4909 printk(" task PC stack pid father child younger older\n");
4910 #endif
4911 read_lock(&tasklist_lock);
4912 do_each_thread(g, p) {
4914 * reset the NMI-timeout, listing all files on a slow
4915 * console might take alot of time:
4917 touch_nmi_watchdog();
4918 if (p->state & state_filter)
4919 show_task(p);
4920 } while_each_thread(g, p);
4922 read_unlock(&tasklist_lock);
4924 * Only show locks if all tasks are dumped:
4926 if (state_filter == -1)
4927 debug_show_all_locks();
4931 * init_idle - set up an idle thread for a given CPU
4932 * @idle: task in question
4933 * @cpu: cpu the idle task belongs to
4935 * NOTE: this function does not set the idle thread's NEED_RESCHED
4936 * flag, to make booting more robust.
4938 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4940 struct rq *rq = cpu_rq(cpu);
4941 unsigned long flags;
4943 idle->timestamp = sched_clock();
4944 idle->sleep_avg = 0;
4945 idle->array = NULL;
4946 idle->prio = idle->normal_prio = MAX_PRIO;
4947 idle->state = TASK_RUNNING;
4948 idle->cpus_allowed = cpumask_of_cpu(cpu);
4949 set_task_cpu(idle, cpu);
4951 spin_lock_irqsave(&rq->lock, flags);
4952 rq->curr = rq->idle = idle;
4953 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4954 idle->oncpu = 1;
4955 #endif
4956 spin_unlock_irqrestore(&rq->lock, flags);
4958 /* Set the preempt count _outside_ the spinlocks! */
4959 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4960 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4961 #else
4962 task_thread_info(idle)->preempt_count = 0;
4963 #endif
4967 * In a system that switches off the HZ timer nohz_cpu_mask
4968 * indicates which cpus entered this state. This is used
4969 * in the rcu update to wait only for active cpus. For system
4970 * which do not switch off the HZ timer nohz_cpu_mask should
4971 * always be CPU_MASK_NONE.
4973 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4975 #ifdef CONFIG_SMP
4977 * This is how migration works:
4979 * 1) we queue a struct migration_req structure in the source CPU's
4980 * runqueue and wake up that CPU's migration thread.
4981 * 2) we down() the locked semaphore => thread blocks.
4982 * 3) migration thread wakes up (implicitly it forces the migrated
4983 * thread off the CPU)
4984 * 4) it gets the migration request and checks whether the migrated
4985 * task is still in the wrong runqueue.
4986 * 5) if it's in the wrong runqueue then the migration thread removes
4987 * it and puts it into the right queue.
4988 * 6) migration thread up()s the semaphore.
4989 * 7) we wake up and the migration is done.
4993 * Change a given task's CPU affinity. Migrate the thread to a
4994 * proper CPU and schedule it away if the CPU it's executing on
4995 * is removed from the allowed bitmask.
4997 * NOTE: the caller must have a valid reference to the task, the
4998 * task must not exit() & deallocate itself prematurely. The
4999 * call is not atomic; no spinlocks may be held.
5001 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
5003 struct migration_req req;
5004 unsigned long flags;
5005 struct rq *rq;
5006 int ret = 0;
5008 rq = task_rq_lock(p, &flags);
5009 if (!cpus_intersects(new_mask, cpu_online_map)) {
5010 ret = -EINVAL;
5011 goto out;
5014 p->cpus_allowed = new_mask;
5015 /* Can the task run on the task's current CPU? If so, we're done */
5016 if (cpu_isset(task_cpu(p), new_mask))
5017 goto out;
5019 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5020 /* Need help from migration thread: drop lock and wait. */
5021 task_rq_unlock(rq, &flags);
5022 wake_up_process(rq->migration_thread);
5023 wait_for_completion(&req.done);
5024 tlb_migrate_finish(p->mm);
5025 return 0;
5027 out:
5028 task_rq_unlock(rq, &flags);
5030 return ret;
5032 EXPORT_SYMBOL_GPL(set_cpus_allowed);
5035 * Move (not current) task off this cpu, onto dest cpu. We're doing
5036 * this because either it can't run here any more (set_cpus_allowed()
5037 * away from this CPU, or CPU going down), or because we're
5038 * attempting to rebalance this task on exec (sched_exec).
5040 * So we race with normal scheduler movements, but that's OK, as long
5041 * as the task is no longer on this CPU.
5043 * Returns non-zero if task was successfully migrated.
5045 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5047 struct rq *rq_dest, *rq_src;
5048 int ret = 0;
5050 if (unlikely(cpu_is_offline(dest_cpu)))
5051 return ret;
5053 rq_src = cpu_rq(src_cpu);
5054 rq_dest = cpu_rq(dest_cpu);
5056 double_rq_lock(rq_src, rq_dest);
5057 /* Already moved. */
5058 if (task_cpu(p) != src_cpu)
5059 goto out;
5060 /* Affinity changed (again). */
5061 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5062 goto out;
5064 set_task_cpu(p, dest_cpu);
5065 if (p->array) {
5067 * Sync timestamp with rq_dest's before activating.
5068 * The same thing could be achieved by doing this step
5069 * afterwards, and pretending it was a local activate.
5070 * This way is cleaner and logically correct.
5072 p->timestamp = p->timestamp - rq_src->most_recent_timestamp
5073 + rq_dest->most_recent_timestamp;
5074 deactivate_task(p, rq_src);
5075 __activate_task(p, rq_dest);
5076 if (TASK_PREEMPTS_CURR(p, rq_dest))
5077 resched_task(rq_dest->curr);
5079 ret = 1;
5080 out:
5081 double_rq_unlock(rq_src, rq_dest);
5082 return ret;
5086 * migration_thread - this is a highprio system thread that performs
5087 * thread migration by bumping thread off CPU then 'pushing' onto
5088 * another runqueue.
5090 static int migration_thread(void *data)
5092 int cpu = (long)data;
5093 struct rq *rq;
5095 rq = cpu_rq(cpu);
5096 BUG_ON(rq->migration_thread != current);
5098 set_current_state(TASK_INTERRUPTIBLE);
5099 while (!kthread_should_stop()) {
5100 struct migration_req *req;
5101 struct list_head *head;
5103 try_to_freeze();
5105 spin_lock_irq(&rq->lock);
5107 if (cpu_is_offline(cpu)) {
5108 spin_unlock_irq(&rq->lock);
5109 goto wait_to_die;
5112 if (rq->active_balance) {
5113 active_load_balance(rq, cpu);
5114 rq->active_balance = 0;
5117 head = &rq->migration_queue;
5119 if (list_empty(head)) {
5120 spin_unlock_irq(&rq->lock);
5121 schedule();
5122 set_current_state(TASK_INTERRUPTIBLE);
5123 continue;
5125 req = list_entry(head->next, struct migration_req, list);
5126 list_del_init(head->next);
5128 spin_unlock(&rq->lock);
5129 __migrate_task(req->task, cpu, req->dest_cpu);
5130 local_irq_enable();
5132 complete(&req->done);
5134 __set_current_state(TASK_RUNNING);
5135 return 0;
5137 wait_to_die:
5138 /* Wait for kthread_stop */
5139 set_current_state(TASK_INTERRUPTIBLE);
5140 while (!kthread_should_stop()) {
5141 schedule();
5142 set_current_state(TASK_INTERRUPTIBLE);
5144 __set_current_state(TASK_RUNNING);
5145 return 0;
5148 #ifdef CONFIG_HOTPLUG_CPU
5150 * Figure out where task on dead CPU should go, use force if neccessary.
5151 * NOTE: interrupts should be disabled by the caller
5153 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5155 unsigned long flags;
5156 cpumask_t mask;
5157 struct rq *rq;
5158 int dest_cpu;
5160 restart:
5161 /* On same node? */
5162 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5163 cpus_and(mask, mask, p->cpus_allowed);
5164 dest_cpu = any_online_cpu(mask);
5166 /* On any allowed CPU? */
5167 if (dest_cpu == NR_CPUS)
5168 dest_cpu = any_online_cpu(p->cpus_allowed);
5170 /* No more Mr. Nice Guy. */
5171 if (dest_cpu == NR_CPUS) {
5172 rq = task_rq_lock(p, &flags);
5173 cpus_setall(p->cpus_allowed);
5174 dest_cpu = any_online_cpu(p->cpus_allowed);
5175 task_rq_unlock(rq, &flags);
5178 * Don't tell them about moving exiting tasks or
5179 * kernel threads (both mm NULL), since they never
5180 * leave kernel.
5182 if (p->mm && printk_ratelimit())
5183 printk(KERN_INFO "process %d (%s) no "
5184 "longer affine to cpu%d\n",
5185 p->pid, p->comm, dead_cpu);
5187 if (!__migrate_task(p, dead_cpu, dest_cpu))
5188 goto restart;
5192 * While a dead CPU has no uninterruptible tasks queued at this point,
5193 * it might still have a nonzero ->nr_uninterruptible counter, because
5194 * for performance reasons the counter is not stricly tracking tasks to
5195 * their home CPUs. So we just add the counter to another CPU's counter,
5196 * to keep the global sum constant after CPU-down:
5198 static void migrate_nr_uninterruptible(struct rq *rq_src)
5200 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5201 unsigned long flags;
5203 local_irq_save(flags);
5204 double_rq_lock(rq_src, rq_dest);
5205 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5206 rq_src->nr_uninterruptible = 0;
5207 double_rq_unlock(rq_src, rq_dest);
5208 local_irq_restore(flags);
5211 /* Run through task list and migrate tasks from the dead cpu. */
5212 static void migrate_live_tasks(int src_cpu)
5214 struct task_struct *p, *t;
5216 write_lock_irq(&tasklist_lock);
5218 do_each_thread(t, p) {
5219 if (p == current)
5220 continue;
5222 if (task_cpu(p) == src_cpu)
5223 move_task_off_dead_cpu(src_cpu, p);
5224 } while_each_thread(t, p);
5226 write_unlock_irq(&tasklist_lock);
5229 /* Schedules idle task to be the next runnable task on current CPU.
5230 * It does so by boosting its priority to highest possible and adding it to
5231 * the _front_ of the runqueue. Used by CPU offline code.
5233 void sched_idle_next(void)
5235 int this_cpu = smp_processor_id();
5236 struct rq *rq = cpu_rq(this_cpu);
5237 struct task_struct *p = rq->idle;
5238 unsigned long flags;
5240 /* cpu has to be offline */
5241 BUG_ON(cpu_online(this_cpu));
5244 * Strictly not necessary since rest of the CPUs are stopped by now
5245 * and interrupts disabled on the current cpu.
5247 spin_lock_irqsave(&rq->lock, flags);
5249 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
5251 /* Add idle task to the _front_ of its priority queue: */
5252 __activate_idle_task(p, rq);
5254 spin_unlock_irqrestore(&rq->lock, flags);
5258 * Ensures that the idle task is using init_mm right before its cpu goes
5259 * offline.
5261 void idle_task_exit(void)
5263 struct mm_struct *mm = current->active_mm;
5265 BUG_ON(cpu_online(smp_processor_id()));
5267 if (mm != &init_mm)
5268 switch_mm(mm, &init_mm, current);
5269 mmdrop(mm);
5272 /* called under rq->lock with disabled interrupts */
5273 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5275 struct rq *rq = cpu_rq(dead_cpu);
5277 /* Must be exiting, otherwise would be on tasklist. */
5278 BUG_ON(p->exit_state != EXIT_ZOMBIE && p->exit_state != EXIT_DEAD);
5280 /* Cannot have done final schedule yet: would have vanished. */
5281 BUG_ON(p->state == TASK_DEAD);
5283 get_task_struct(p);
5286 * Drop lock around migration; if someone else moves it,
5287 * that's OK. No task can be added to this CPU, so iteration is
5288 * fine.
5289 * NOTE: interrupts should be left disabled --dev@
5291 spin_unlock(&rq->lock);
5292 move_task_off_dead_cpu(dead_cpu, p);
5293 spin_lock(&rq->lock);
5295 put_task_struct(p);
5298 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5299 static void migrate_dead_tasks(unsigned int dead_cpu)
5301 struct rq *rq = cpu_rq(dead_cpu);
5302 unsigned int arr, i;
5304 for (arr = 0; arr < 2; arr++) {
5305 for (i = 0; i < MAX_PRIO; i++) {
5306 struct list_head *list = &rq->arrays[arr].queue[i];
5308 while (!list_empty(list))
5309 migrate_dead(dead_cpu, list_entry(list->next,
5310 struct task_struct, run_list));
5314 #endif /* CONFIG_HOTPLUG_CPU */
5317 * migration_call - callback that gets triggered when a CPU is added.
5318 * Here we can start up the necessary migration thread for the new CPU.
5320 static int __cpuinit
5321 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5323 struct task_struct *p;
5324 int cpu = (long)hcpu;
5325 unsigned long flags;
5326 struct rq *rq;
5328 switch (action) {
5329 case CPU_UP_PREPARE:
5330 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
5331 if (IS_ERR(p))
5332 return NOTIFY_BAD;
5333 p->flags |= PF_NOFREEZE;
5334 kthread_bind(p, cpu);
5335 /* Must be high prio: stop_machine expects to yield to it. */
5336 rq = task_rq_lock(p, &flags);
5337 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
5338 task_rq_unlock(rq, &flags);
5339 cpu_rq(cpu)->migration_thread = p;
5340 break;
5342 case CPU_ONLINE:
5343 /* Strictly unneccessary, as first user will wake it. */
5344 wake_up_process(cpu_rq(cpu)->migration_thread);
5345 break;
5347 #ifdef CONFIG_HOTPLUG_CPU
5348 case CPU_UP_CANCELED:
5349 if (!cpu_rq(cpu)->migration_thread)
5350 break;
5351 /* Unbind it from offline cpu so it can run. Fall thru. */
5352 kthread_bind(cpu_rq(cpu)->migration_thread,
5353 any_online_cpu(cpu_online_map));
5354 kthread_stop(cpu_rq(cpu)->migration_thread);
5355 cpu_rq(cpu)->migration_thread = NULL;
5356 break;
5358 case CPU_DEAD:
5359 migrate_live_tasks(cpu);
5360 rq = cpu_rq(cpu);
5361 kthread_stop(rq->migration_thread);
5362 rq->migration_thread = NULL;
5363 /* Idle task back to normal (off runqueue, low prio) */
5364 rq = task_rq_lock(rq->idle, &flags);
5365 deactivate_task(rq->idle, rq);
5366 rq->idle->static_prio = MAX_PRIO;
5367 __setscheduler(rq->idle, SCHED_NORMAL, 0);
5368 migrate_dead_tasks(cpu);
5369 task_rq_unlock(rq, &flags);
5370 migrate_nr_uninterruptible(rq);
5371 BUG_ON(rq->nr_running != 0);
5373 /* No need to migrate the tasks: it was best-effort if
5374 * they didn't do lock_cpu_hotplug(). Just wake up
5375 * the requestors. */
5376 spin_lock_irq(&rq->lock);
5377 while (!list_empty(&rq->migration_queue)) {
5378 struct migration_req *req;
5380 req = list_entry(rq->migration_queue.next,
5381 struct migration_req, list);
5382 list_del_init(&req->list);
5383 complete(&req->done);
5385 spin_unlock_irq(&rq->lock);
5386 break;
5387 #endif
5389 return NOTIFY_OK;
5392 /* Register at highest priority so that task migration (migrate_all_tasks)
5393 * happens before everything else.
5395 static struct notifier_block __cpuinitdata migration_notifier = {
5396 .notifier_call = migration_call,
5397 .priority = 10
5400 int __init migration_init(void)
5402 void *cpu = (void *)(long)smp_processor_id();
5403 int err;
5405 /* Start one for the boot CPU: */
5406 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5407 BUG_ON(err == NOTIFY_BAD);
5408 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5409 register_cpu_notifier(&migration_notifier);
5411 return 0;
5413 #endif
5415 #ifdef CONFIG_SMP
5416 #undef SCHED_DOMAIN_DEBUG
5417 #ifdef SCHED_DOMAIN_DEBUG
5418 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5420 int level = 0;
5422 if (!sd) {
5423 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5424 return;
5427 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5429 do {
5430 int i;
5431 char str[NR_CPUS];
5432 struct sched_group *group = sd->groups;
5433 cpumask_t groupmask;
5435 cpumask_scnprintf(str, NR_CPUS, sd->span);
5436 cpus_clear(groupmask);
5438 printk(KERN_DEBUG);
5439 for (i = 0; i < level + 1; i++)
5440 printk(" ");
5441 printk("domain %d: ", level);
5443 if (!(sd->flags & SD_LOAD_BALANCE)) {
5444 printk("does not load-balance\n");
5445 if (sd->parent)
5446 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5447 " has parent");
5448 break;
5451 printk("span %s\n", str);
5453 if (!cpu_isset(cpu, sd->span))
5454 printk(KERN_ERR "ERROR: domain->span does not contain "
5455 "CPU%d\n", cpu);
5456 if (!cpu_isset(cpu, group->cpumask))
5457 printk(KERN_ERR "ERROR: domain->groups does not contain"
5458 " CPU%d\n", cpu);
5460 printk(KERN_DEBUG);
5461 for (i = 0; i < level + 2; i++)
5462 printk(" ");
5463 printk("groups:");
5464 do {
5465 if (!group) {
5466 printk("\n");
5467 printk(KERN_ERR "ERROR: group is NULL\n");
5468 break;
5471 if (!group->cpu_power) {
5472 printk("\n");
5473 printk(KERN_ERR "ERROR: domain->cpu_power not "
5474 "set\n");
5477 if (!cpus_weight(group->cpumask)) {
5478 printk("\n");
5479 printk(KERN_ERR "ERROR: empty group\n");
5482 if (cpus_intersects(groupmask, group->cpumask)) {
5483 printk("\n");
5484 printk(KERN_ERR "ERROR: repeated CPUs\n");
5487 cpus_or(groupmask, groupmask, group->cpumask);
5489 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5490 printk(" %s", str);
5492 group = group->next;
5493 } while (group != sd->groups);
5494 printk("\n");
5496 if (!cpus_equal(sd->span, groupmask))
5497 printk(KERN_ERR "ERROR: groups don't span "
5498 "domain->span\n");
5500 level++;
5501 sd = sd->parent;
5502 if (!sd)
5503 continue;
5505 if (!cpus_subset(groupmask, sd->span))
5506 printk(KERN_ERR "ERROR: parent span is not a superset "
5507 "of domain->span\n");
5509 } while (sd);
5511 #else
5512 # define sched_domain_debug(sd, cpu) do { } while (0)
5513 #endif
5515 static int sd_degenerate(struct sched_domain *sd)
5517 if (cpus_weight(sd->span) == 1)
5518 return 1;
5520 /* Following flags need at least 2 groups */
5521 if (sd->flags & (SD_LOAD_BALANCE |
5522 SD_BALANCE_NEWIDLE |
5523 SD_BALANCE_FORK |
5524 SD_BALANCE_EXEC |
5525 SD_SHARE_CPUPOWER |
5526 SD_SHARE_PKG_RESOURCES)) {
5527 if (sd->groups != sd->groups->next)
5528 return 0;
5531 /* Following flags don't use groups */
5532 if (sd->flags & (SD_WAKE_IDLE |
5533 SD_WAKE_AFFINE |
5534 SD_WAKE_BALANCE))
5535 return 0;
5537 return 1;
5540 static int
5541 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5543 unsigned long cflags = sd->flags, pflags = parent->flags;
5545 if (sd_degenerate(parent))
5546 return 1;
5548 if (!cpus_equal(sd->span, parent->span))
5549 return 0;
5551 /* Does parent contain flags not in child? */
5552 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5553 if (cflags & SD_WAKE_AFFINE)
5554 pflags &= ~SD_WAKE_BALANCE;
5555 /* Flags needing groups don't count if only 1 group in parent */
5556 if (parent->groups == parent->groups->next) {
5557 pflags &= ~(SD_LOAD_BALANCE |
5558 SD_BALANCE_NEWIDLE |
5559 SD_BALANCE_FORK |
5560 SD_BALANCE_EXEC |
5561 SD_SHARE_CPUPOWER |
5562 SD_SHARE_PKG_RESOURCES);
5564 if (~cflags & pflags)
5565 return 0;
5567 return 1;
5571 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5572 * hold the hotplug lock.
5574 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5576 struct rq *rq = cpu_rq(cpu);
5577 struct sched_domain *tmp;
5579 /* Remove the sched domains which do not contribute to scheduling. */
5580 for (tmp = sd; tmp; tmp = tmp->parent) {
5581 struct sched_domain *parent = tmp->parent;
5582 if (!parent)
5583 break;
5584 if (sd_parent_degenerate(tmp, parent)) {
5585 tmp->parent = parent->parent;
5586 if (parent->parent)
5587 parent->parent->child = tmp;
5591 if (sd && sd_degenerate(sd)) {
5592 sd = sd->parent;
5593 if (sd)
5594 sd->child = NULL;
5597 sched_domain_debug(sd, cpu);
5599 rcu_assign_pointer(rq->sd, sd);
5602 /* cpus with isolated domains */
5603 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5605 /* Setup the mask of cpus configured for isolated domains */
5606 static int __init isolated_cpu_setup(char *str)
5608 int ints[NR_CPUS], i;
5610 str = get_options(str, ARRAY_SIZE(ints), ints);
5611 cpus_clear(cpu_isolated_map);
5612 for (i = 1; i <= ints[0]; i++)
5613 if (ints[i] < NR_CPUS)
5614 cpu_set(ints[i], cpu_isolated_map);
5615 return 1;
5618 __setup ("isolcpus=", isolated_cpu_setup);
5621 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5622 * to a function which identifies what group(along with sched group) a CPU
5623 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5624 * (due to the fact that we keep track of groups covered with a cpumask_t).
5626 * init_sched_build_groups will build a circular linked list of the groups
5627 * covered by the given span, and will set each group's ->cpumask correctly,
5628 * and ->cpu_power to 0.
5630 static void
5631 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5632 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5633 struct sched_group **sg))
5635 struct sched_group *first = NULL, *last = NULL;
5636 cpumask_t covered = CPU_MASK_NONE;
5637 int i;
5639 for_each_cpu_mask(i, span) {
5640 struct sched_group *sg;
5641 int group = group_fn(i, cpu_map, &sg);
5642 int j;
5644 if (cpu_isset(i, covered))
5645 continue;
5647 sg->cpumask = CPU_MASK_NONE;
5648 sg->cpu_power = 0;
5650 for_each_cpu_mask(j, span) {
5651 if (group_fn(j, cpu_map, NULL) != group)
5652 continue;
5654 cpu_set(j, covered);
5655 cpu_set(j, sg->cpumask);
5657 if (!first)
5658 first = sg;
5659 if (last)
5660 last->next = sg;
5661 last = sg;
5663 last->next = first;
5666 #define SD_NODES_PER_DOMAIN 16
5669 * Self-tuning task migration cost measurement between source and target CPUs.
5671 * This is done by measuring the cost of manipulating buffers of varying
5672 * sizes. For a given buffer-size here are the steps that are taken:
5674 * 1) the source CPU reads+dirties a shared buffer
5675 * 2) the target CPU reads+dirties the same shared buffer
5677 * We measure how long they take, in the following 4 scenarios:
5679 * - source: CPU1, target: CPU2 | cost1
5680 * - source: CPU2, target: CPU1 | cost2
5681 * - source: CPU1, target: CPU1 | cost3
5682 * - source: CPU2, target: CPU2 | cost4
5684 * We then calculate the cost3+cost4-cost1-cost2 difference - this is
5685 * the cost of migration.
5687 * We then start off from a small buffer-size and iterate up to larger
5688 * buffer sizes, in 5% steps - measuring each buffer-size separately, and
5689 * doing a maximum search for the cost. (The maximum cost for a migration
5690 * normally occurs when the working set size is around the effective cache
5691 * size.)
5693 #define SEARCH_SCOPE 2
5694 #define MIN_CACHE_SIZE (64*1024U)
5695 #define DEFAULT_CACHE_SIZE (5*1024*1024U)
5696 #define ITERATIONS 1
5697 #define SIZE_THRESH 130
5698 #define COST_THRESH 130
5701 * The migration cost is a function of 'domain distance'. Domain
5702 * distance is the number of steps a CPU has to iterate down its
5703 * domain tree to share a domain with the other CPU. The farther
5704 * two CPUs are from each other, the larger the distance gets.
5706 * Note that we use the distance only to cache measurement results,
5707 * the distance value is not used numerically otherwise. When two
5708 * CPUs have the same distance it is assumed that the migration
5709 * cost is the same. (this is a simplification but quite practical)
5711 #define MAX_DOMAIN_DISTANCE 32
5713 static unsigned long long migration_cost[MAX_DOMAIN_DISTANCE] =
5714 { [ 0 ... MAX_DOMAIN_DISTANCE-1 ] =
5716 * Architectures may override the migration cost and thus avoid
5717 * boot-time calibration. Unit is nanoseconds. Mostly useful for
5718 * virtualized hardware:
5720 #ifdef CONFIG_DEFAULT_MIGRATION_COST
5721 CONFIG_DEFAULT_MIGRATION_COST
5722 #else
5723 -1LL
5724 #endif
5728 * Allow override of migration cost - in units of microseconds.
5729 * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost
5730 * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs:
5732 static int __init migration_cost_setup(char *str)
5734 int ints[MAX_DOMAIN_DISTANCE+1], i;
5736 str = get_options(str, ARRAY_SIZE(ints), ints);
5738 printk("#ints: %d\n", ints[0]);
5739 for (i = 1; i <= ints[0]; i++) {
5740 migration_cost[i-1] = (unsigned long long)ints[i]*1000;
5741 printk("migration_cost[%d]: %Ld\n", i-1, migration_cost[i-1]);
5743 return 1;
5746 __setup ("migration_cost=", migration_cost_setup);
5749 * Global multiplier (divisor) for migration-cutoff values,
5750 * in percentiles. E.g. use a value of 150 to get 1.5 times
5751 * longer cache-hot cutoff times.
5753 * (We scale it from 100 to 128 to long long handling easier.)
5756 #define MIGRATION_FACTOR_SCALE 128
5758 static unsigned int migration_factor = MIGRATION_FACTOR_SCALE;
5760 static int __init setup_migration_factor(char *str)
5762 get_option(&str, &migration_factor);
5763 migration_factor = migration_factor * MIGRATION_FACTOR_SCALE / 100;
5764 return 1;
5767 __setup("migration_factor=", setup_migration_factor);
5770 * Estimated distance of two CPUs, measured via the number of domains
5771 * we have to pass for the two CPUs to be in the same span:
5773 static unsigned long domain_distance(int cpu1, int cpu2)
5775 unsigned long distance = 0;
5776 struct sched_domain *sd;
5778 for_each_domain(cpu1, sd) {
5779 WARN_ON(!cpu_isset(cpu1, sd->span));
5780 if (cpu_isset(cpu2, sd->span))
5781 return distance;
5782 distance++;
5784 if (distance >= MAX_DOMAIN_DISTANCE) {
5785 WARN_ON(1);
5786 distance = MAX_DOMAIN_DISTANCE-1;
5789 return distance;
5792 static unsigned int migration_debug;
5794 static int __init setup_migration_debug(char *str)
5796 get_option(&str, &migration_debug);
5797 return 1;
5800 __setup("migration_debug=", setup_migration_debug);
5803 * Maximum cache-size that the scheduler should try to measure.
5804 * Architectures with larger caches should tune this up during
5805 * bootup. Gets used in the domain-setup code (i.e. during SMP
5806 * bootup).
5808 unsigned int max_cache_size;
5810 static int __init setup_max_cache_size(char *str)
5812 get_option(&str, &max_cache_size);
5813 return 1;
5816 __setup("max_cache_size=", setup_max_cache_size);
5819 * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This
5820 * is the operation that is timed, so we try to generate unpredictable
5821 * cachemisses that still end up filling the L2 cache:
5823 static void touch_cache(void *__cache, unsigned long __size)
5825 unsigned long size = __size / sizeof(long);
5826 unsigned long chunk1 = size / 3;
5827 unsigned long chunk2 = 2 * size / 3;
5828 unsigned long *cache = __cache;
5829 int i;
5831 for (i = 0; i < size/6; i += 8) {
5832 switch (i % 6) {
5833 case 0: cache[i]++;
5834 case 1: cache[size-1-i]++;
5835 case 2: cache[chunk1-i]++;
5836 case 3: cache[chunk1+i]++;
5837 case 4: cache[chunk2-i]++;
5838 case 5: cache[chunk2+i]++;
5844 * Measure the cache-cost of one task migration. Returns in units of nsec.
5846 static unsigned long long
5847 measure_one(void *cache, unsigned long size, int source, int target)
5849 cpumask_t mask, saved_mask;
5850 unsigned long long t0, t1, t2, t3, cost;
5852 saved_mask = current->cpus_allowed;
5855 * Flush source caches to RAM and invalidate them:
5857 sched_cacheflush();
5860 * Migrate to the source CPU:
5862 mask = cpumask_of_cpu(source);
5863 set_cpus_allowed(current, mask);
5864 WARN_ON(smp_processor_id() != source);
5867 * Dirty the working set:
5869 t0 = sched_clock();
5870 touch_cache(cache, size);
5871 t1 = sched_clock();
5874 * Migrate to the target CPU, dirty the L2 cache and access
5875 * the shared buffer. (which represents the working set
5876 * of a migrated task.)
5878 mask = cpumask_of_cpu(target);
5879 set_cpus_allowed(current, mask);
5880 WARN_ON(smp_processor_id() != target);
5882 t2 = sched_clock();
5883 touch_cache(cache, size);
5884 t3 = sched_clock();
5886 cost = t1-t0 + t3-t2;
5888 if (migration_debug >= 2)
5889 printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n",
5890 source, target, t1-t0, t1-t0, t3-t2, cost);
5892 * Flush target caches to RAM and invalidate them:
5894 sched_cacheflush();
5896 set_cpus_allowed(current, saved_mask);
5898 return cost;
5902 * Measure a series of task migrations and return the average
5903 * result. Since this code runs early during bootup the system
5904 * is 'undisturbed' and the average latency makes sense.
5906 * The algorithm in essence auto-detects the relevant cache-size,
5907 * so it will properly detect different cachesizes for different
5908 * cache-hierarchies, depending on how the CPUs are connected.
5910 * Architectures can prime the upper limit of the search range via
5911 * max_cache_size, otherwise the search range defaults to 20MB...64K.
5913 static unsigned long long
5914 measure_cost(int cpu1, int cpu2, void *cache, unsigned int size)
5916 unsigned long long cost1, cost2;
5917 int i;
5920 * Measure the migration cost of 'size' bytes, over an
5921 * average of 10 runs:
5923 * (We perturb the cache size by a small (0..4k)
5924 * value to compensate size/alignment related artifacts.
5925 * We also subtract the cost of the operation done on
5926 * the same CPU.)
5928 cost1 = 0;
5931 * dry run, to make sure we start off cache-cold on cpu1,
5932 * and to get any vmalloc pagefaults in advance:
5934 measure_one(cache, size, cpu1, cpu2);
5935 for (i = 0; i < ITERATIONS; i++)
5936 cost1 += measure_one(cache, size - i * 1024, cpu1, cpu2);
5938 measure_one(cache, size, cpu2, cpu1);
5939 for (i = 0; i < ITERATIONS; i++)
5940 cost1 += measure_one(cache, size - i * 1024, cpu2, cpu1);
5943 * (We measure the non-migrating [cached] cost on both
5944 * cpu1 and cpu2, to handle CPUs with different speeds)
5946 cost2 = 0;
5948 measure_one(cache, size, cpu1, cpu1);
5949 for (i = 0; i < ITERATIONS; i++)
5950 cost2 += measure_one(cache, size - i * 1024, cpu1, cpu1);
5952 measure_one(cache, size, cpu2, cpu2);
5953 for (i = 0; i < ITERATIONS; i++)
5954 cost2 += measure_one(cache, size - i * 1024, cpu2, cpu2);
5957 * Get the per-iteration migration cost:
5959 do_div(cost1, 2 * ITERATIONS);
5960 do_div(cost2, 2 * ITERATIONS);
5962 return cost1 - cost2;
5965 static unsigned long long measure_migration_cost(int cpu1, int cpu2)
5967 unsigned long long max_cost = 0, fluct = 0, avg_fluct = 0;
5968 unsigned int max_size, size, size_found = 0;
5969 long long cost = 0, prev_cost;
5970 void *cache;
5973 * Search from max_cache_size*5 down to 64K - the real relevant
5974 * cachesize has to lie somewhere inbetween.
5976 if (max_cache_size) {
5977 max_size = max(max_cache_size * SEARCH_SCOPE, MIN_CACHE_SIZE);
5978 size = max(max_cache_size / SEARCH_SCOPE, MIN_CACHE_SIZE);
5979 } else {
5981 * Since we have no estimation about the relevant
5982 * search range
5984 max_size = DEFAULT_CACHE_SIZE * SEARCH_SCOPE;
5985 size = MIN_CACHE_SIZE;
5988 if (!cpu_online(cpu1) || !cpu_online(cpu2)) {
5989 printk("cpu %d and %d not both online!\n", cpu1, cpu2);
5990 return 0;
5994 * Allocate the working set:
5996 cache = vmalloc(max_size);
5997 if (!cache) {
5998 printk("could not vmalloc %d bytes for cache!\n", 2 * max_size);
5999 return 1000000; /* return 1 msec on very small boxen */
6002 while (size <= max_size) {
6003 prev_cost = cost;
6004 cost = measure_cost(cpu1, cpu2, cache, size);
6007 * Update the max:
6009 if (cost > 0) {
6010 if (max_cost < cost) {
6011 max_cost = cost;
6012 size_found = size;
6016 * Calculate average fluctuation, we use this to prevent
6017 * noise from triggering an early break out of the loop:
6019 fluct = abs(cost - prev_cost);
6020 avg_fluct = (avg_fluct + fluct)/2;
6022 if (migration_debug)
6023 printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): "
6024 "(%8Ld %8Ld)\n",
6025 cpu1, cpu2, size,
6026 (long)cost / 1000000,
6027 ((long)cost / 100000) % 10,
6028 (long)max_cost / 1000000,
6029 ((long)max_cost / 100000) % 10,
6030 domain_distance(cpu1, cpu2),
6031 cost, avg_fluct);
6034 * If we iterated at least 20% past the previous maximum,
6035 * and the cost has dropped by more than 20% already,
6036 * (taking fluctuations into account) then we assume to
6037 * have found the maximum and break out of the loop early:
6039 if (size_found && (size*100 > size_found*SIZE_THRESH))
6040 if (cost+avg_fluct <= 0 ||
6041 max_cost*100 > (cost+avg_fluct)*COST_THRESH) {
6043 if (migration_debug)
6044 printk("-> found max.\n");
6045 break;
6048 * Increase the cachesize in 10% steps:
6050 size = size * 10 / 9;
6053 if (migration_debug)
6054 printk("[%d][%d] working set size found: %d, cost: %Ld\n",
6055 cpu1, cpu2, size_found, max_cost);
6057 vfree(cache);
6060 * A task is considered 'cache cold' if at least 2 times
6061 * the worst-case cost of migration has passed.
6063 * (this limit is only listened to if the load-balancing
6064 * situation is 'nice' - if there is a large imbalance we
6065 * ignore it for the sake of CPU utilization and
6066 * processing fairness.)
6068 return 2 * max_cost * migration_factor / MIGRATION_FACTOR_SCALE;
6071 static void calibrate_migration_costs(const cpumask_t *cpu_map)
6073 int cpu1 = -1, cpu2 = -1, cpu, orig_cpu = raw_smp_processor_id();
6074 unsigned long j0, j1, distance, max_distance = 0;
6075 struct sched_domain *sd;
6077 j0 = jiffies;
6080 * First pass - calculate the cacheflush times:
6082 for_each_cpu_mask(cpu1, *cpu_map) {
6083 for_each_cpu_mask(cpu2, *cpu_map) {
6084 if (cpu1 == cpu2)
6085 continue;
6086 distance = domain_distance(cpu1, cpu2);
6087 max_distance = max(max_distance, distance);
6089 * No result cached yet?
6091 if (migration_cost[distance] == -1LL)
6092 migration_cost[distance] =
6093 measure_migration_cost(cpu1, cpu2);
6097 * Second pass - update the sched domain hierarchy with
6098 * the new cache-hot-time estimations:
6100 for_each_cpu_mask(cpu, *cpu_map) {
6101 distance = 0;
6102 for_each_domain(cpu, sd) {
6103 sd->cache_hot_time = migration_cost[distance];
6104 distance++;
6108 * Print the matrix:
6110 if (migration_debug)
6111 printk("migration: max_cache_size: %d, cpu: %d MHz:\n",
6112 max_cache_size,
6113 #ifdef CONFIG_X86
6114 cpu_khz/1000
6115 #else
6117 #endif
6119 if (system_state == SYSTEM_BOOTING && num_online_cpus() > 1) {
6120 printk("migration_cost=");
6121 for (distance = 0; distance <= max_distance; distance++) {
6122 if (distance)
6123 printk(",");
6124 printk("%ld", (long)migration_cost[distance] / 1000);
6126 printk("\n");
6128 j1 = jiffies;
6129 if (migration_debug)
6130 printk("migration: %ld seconds\n", (j1-j0) / HZ);
6133 * Move back to the original CPU. NUMA-Q gets confused
6134 * if we migrate to another quad during bootup.
6136 if (raw_smp_processor_id() != orig_cpu) {
6137 cpumask_t mask = cpumask_of_cpu(orig_cpu),
6138 saved_mask = current->cpus_allowed;
6140 set_cpus_allowed(current, mask);
6141 set_cpus_allowed(current, saved_mask);
6145 #ifdef CONFIG_NUMA
6148 * find_next_best_node - find the next node to include in a sched_domain
6149 * @node: node whose sched_domain we're building
6150 * @used_nodes: nodes already in the sched_domain
6152 * Find the next node to include in a given scheduling domain. Simply
6153 * finds the closest node not already in the @used_nodes map.
6155 * Should use nodemask_t.
6157 static int find_next_best_node(int node, unsigned long *used_nodes)
6159 int i, n, val, min_val, best_node = 0;
6161 min_val = INT_MAX;
6163 for (i = 0; i < MAX_NUMNODES; i++) {
6164 /* Start at @node */
6165 n = (node + i) % MAX_NUMNODES;
6167 if (!nr_cpus_node(n))
6168 continue;
6170 /* Skip already used nodes */
6171 if (test_bit(n, used_nodes))
6172 continue;
6174 /* Simple min distance search */
6175 val = node_distance(node, n);
6177 if (val < min_val) {
6178 min_val = val;
6179 best_node = n;
6183 set_bit(best_node, used_nodes);
6184 return best_node;
6188 * sched_domain_node_span - get a cpumask for a node's sched_domain
6189 * @node: node whose cpumask we're constructing
6190 * @size: number of nodes to include in this span
6192 * Given a node, construct a good cpumask for its sched_domain to span. It
6193 * should be one that prevents unnecessary balancing, but also spreads tasks
6194 * out optimally.
6196 static cpumask_t sched_domain_node_span(int node)
6198 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
6199 cpumask_t span, nodemask;
6200 int i;
6202 cpus_clear(span);
6203 bitmap_zero(used_nodes, MAX_NUMNODES);
6205 nodemask = node_to_cpumask(node);
6206 cpus_or(span, span, nodemask);
6207 set_bit(node, used_nodes);
6209 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6210 int next_node = find_next_best_node(node, used_nodes);
6212 nodemask = node_to_cpumask(next_node);
6213 cpus_or(span, span, nodemask);
6216 return span;
6218 #endif
6220 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6223 * SMT sched-domains:
6225 #ifdef CONFIG_SCHED_SMT
6226 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6227 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6229 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
6230 struct sched_group **sg)
6232 if (sg)
6233 *sg = &per_cpu(sched_group_cpus, cpu);
6234 return cpu;
6236 #endif
6239 * multi-core sched-domains:
6241 #ifdef CONFIG_SCHED_MC
6242 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6243 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6244 #endif
6246 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6247 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
6248 struct sched_group **sg)
6250 int group;
6251 cpumask_t mask = cpu_sibling_map[cpu];
6252 cpus_and(mask, mask, *cpu_map);
6253 group = first_cpu(mask);
6254 if (sg)
6255 *sg = &per_cpu(sched_group_core, group);
6256 return group;
6258 #elif defined(CONFIG_SCHED_MC)
6259 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
6260 struct sched_group **sg)
6262 if (sg)
6263 *sg = &per_cpu(sched_group_core, cpu);
6264 return cpu;
6266 #endif
6268 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6269 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6271 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
6272 struct sched_group **sg)
6274 int group;
6275 #ifdef CONFIG_SCHED_MC
6276 cpumask_t mask = cpu_coregroup_map(cpu);
6277 cpus_and(mask, mask, *cpu_map);
6278 group = first_cpu(mask);
6279 #elif defined(CONFIG_SCHED_SMT)
6280 cpumask_t mask = cpu_sibling_map[cpu];
6281 cpus_and(mask, mask, *cpu_map);
6282 group = first_cpu(mask);
6283 #else
6284 group = cpu;
6285 #endif
6286 if (sg)
6287 *sg = &per_cpu(sched_group_phys, group);
6288 return group;
6291 #ifdef CONFIG_NUMA
6293 * The init_sched_build_groups can't handle what we want to do with node
6294 * groups, so roll our own. Now each node has its own list of groups which
6295 * gets dynamically allocated.
6297 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6298 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
6300 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6301 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6303 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6304 struct sched_group **sg)
6306 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
6307 int group;
6309 cpus_and(nodemask, nodemask, *cpu_map);
6310 group = first_cpu(nodemask);
6312 if (sg)
6313 *sg = &per_cpu(sched_group_allnodes, group);
6314 return group;
6317 static void init_numa_sched_groups_power(struct sched_group *group_head)
6319 struct sched_group *sg = group_head;
6320 int j;
6322 if (!sg)
6323 return;
6324 next_sg:
6325 for_each_cpu_mask(j, sg->cpumask) {
6326 struct sched_domain *sd;
6328 sd = &per_cpu(phys_domains, j);
6329 if (j != first_cpu(sd->groups->cpumask)) {
6331 * Only add "power" once for each
6332 * physical package.
6334 continue;
6337 sg->cpu_power += sd->groups->cpu_power;
6339 sg = sg->next;
6340 if (sg != group_head)
6341 goto next_sg;
6343 #endif
6345 #ifdef CONFIG_NUMA
6346 /* Free memory allocated for various sched_group structures */
6347 static void free_sched_groups(const cpumask_t *cpu_map)
6349 int cpu, i;
6351 for_each_cpu_mask(cpu, *cpu_map) {
6352 struct sched_group **sched_group_nodes
6353 = sched_group_nodes_bycpu[cpu];
6355 if (!sched_group_nodes)
6356 continue;
6358 for (i = 0; i < MAX_NUMNODES; i++) {
6359 cpumask_t nodemask = node_to_cpumask(i);
6360 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6362 cpus_and(nodemask, nodemask, *cpu_map);
6363 if (cpus_empty(nodemask))
6364 continue;
6366 if (sg == NULL)
6367 continue;
6368 sg = sg->next;
6369 next_sg:
6370 oldsg = sg;
6371 sg = sg->next;
6372 kfree(oldsg);
6373 if (oldsg != sched_group_nodes[i])
6374 goto next_sg;
6376 kfree(sched_group_nodes);
6377 sched_group_nodes_bycpu[cpu] = NULL;
6380 #else
6381 static void free_sched_groups(const cpumask_t *cpu_map)
6384 #endif
6387 * Initialize sched groups cpu_power.
6389 * cpu_power indicates the capacity of sched group, which is used while
6390 * distributing the load between different sched groups in a sched domain.
6391 * Typically cpu_power for all the groups in a sched domain will be same unless
6392 * there are asymmetries in the topology. If there are asymmetries, group
6393 * having more cpu_power will pickup more load compared to the group having
6394 * less cpu_power.
6396 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6397 * the maximum number of tasks a group can handle in the presence of other idle
6398 * or lightly loaded groups in the same sched domain.
6400 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6402 struct sched_domain *child;
6403 struct sched_group *group;
6405 WARN_ON(!sd || !sd->groups);
6407 if (cpu != first_cpu(sd->groups->cpumask))
6408 return;
6410 child = sd->child;
6413 * For perf policy, if the groups in child domain share resources
6414 * (for example cores sharing some portions of the cache hierarchy
6415 * or SMT), then set this domain groups cpu_power such that each group
6416 * can handle only one task, when there are other idle groups in the
6417 * same sched domain.
6419 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6420 (child->flags &
6421 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6422 sd->groups->cpu_power = SCHED_LOAD_SCALE;
6423 return;
6426 sd->groups->cpu_power = 0;
6429 * add cpu_power of each child group to this groups cpu_power
6431 group = child->groups;
6432 do {
6433 sd->groups->cpu_power += group->cpu_power;
6434 group = group->next;
6435 } while (group != child->groups);
6439 * Build sched domains for a given set of cpus and attach the sched domains
6440 * to the individual cpus
6442 static int build_sched_domains(const cpumask_t *cpu_map)
6444 int i;
6445 struct sched_domain *sd;
6446 #ifdef CONFIG_NUMA
6447 struct sched_group **sched_group_nodes = NULL;
6448 int sd_allnodes = 0;
6451 * Allocate the per-node list of sched groups
6453 sched_group_nodes = kzalloc(sizeof(struct sched_group*)*MAX_NUMNODES,
6454 GFP_KERNEL);
6455 if (!sched_group_nodes) {
6456 printk(KERN_WARNING "Can not alloc sched group node list\n");
6457 return -ENOMEM;
6459 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6460 #endif
6463 * Set up domains for cpus specified by the cpu_map.
6465 for_each_cpu_mask(i, *cpu_map) {
6466 struct sched_domain *sd = NULL, *p;
6467 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6469 cpus_and(nodemask, nodemask, *cpu_map);
6471 #ifdef CONFIG_NUMA
6472 if (cpus_weight(*cpu_map)
6473 > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6474 sd = &per_cpu(allnodes_domains, i);
6475 *sd = SD_ALLNODES_INIT;
6476 sd->span = *cpu_map;
6477 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6478 p = sd;
6479 sd_allnodes = 1;
6480 } else
6481 p = NULL;
6483 sd = &per_cpu(node_domains, i);
6484 *sd = SD_NODE_INIT;
6485 sd->span = sched_domain_node_span(cpu_to_node(i));
6486 sd->parent = p;
6487 if (p)
6488 p->child = sd;
6489 cpus_and(sd->span, sd->span, *cpu_map);
6490 #endif
6492 p = sd;
6493 sd = &per_cpu(phys_domains, i);
6494 *sd = SD_CPU_INIT;
6495 sd->span = nodemask;
6496 sd->parent = p;
6497 if (p)
6498 p->child = sd;
6499 cpu_to_phys_group(i, cpu_map, &sd->groups);
6501 #ifdef CONFIG_SCHED_MC
6502 p = sd;
6503 sd = &per_cpu(core_domains, i);
6504 *sd = SD_MC_INIT;
6505 sd->span = cpu_coregroup_map(i);
6506 cpus_and(sd->span, sd->span, *cpu_map);
6507 sd->parent = p;
6508 p->child = sd;
6509 cpu_to_core_group(i, cpu_map, &sd->groups);
6510 #endif
6512 #ifdef CONFIG_SCHED_SMT
6513 p = sd;
6514 sd = &per_cpu(cpu_domains, i);
6515 *sd = SD_SIBLING_INIT;
6516 sd->span = cpu_sibling_map[i];
6517 cpus_and(sd->span, sd->span, *cpu_map);
6518 sd->parent = p;
6519 p->child = sd;
6520 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6521 #endif
6524 #ifdef CONFIG_SCHED_SMT
6525 /* Set up CPU (sibling) groups */
6526 for_each_cpu_mask(i, *cpu_map) {
6527 cpumask_t this_sibling_map = cpu_sibling_map[i];
6528 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6529 if (i != first_cpu(this_sibling_map))
6530 continue;
6532 init_sched_build_groups(this_sibling_map, cpu_map, &cpu_to_cpu_group);
6534 #endif
6536 #ifdef CONFIG_SCHED_MC
6537 /* Set up multi-core groups */
6538 for_each_cpu_mask(i, *cpu_map) {
6539 cpumask_t this_core_map = cpu_coregroup_map(i);
6540 cpus_and(this_core_map, this_core_map, *cpu_map);
6541 if (i != first_cpu(this_core_map))
6542 continue;
6543 init_sched_build_groups(this_core_map, cpu_map, &cpu_to_core_group);
6545 #endif
6548 /* Set up physical groups */
6549 for (i = 0; i < MAX_NUMNODES; i++) {
6550 cpumask_t nodemask = node_to_cpumask(i);
6552 cpus_and(nodemask, nodemask, *cpu_map);
6553 if (cpus_empty(nodemask))
6554 continue;
6556 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6559 #ifdef CONFIG_NUMA
6560 /* Set up node groups */
6561 if (sd_allnodes)
6562 init_sched_build_groups(*cpu_map, cpu_map, &cpu_to_allnodes_group);
6564 for (i = 0; i < MAX_NUMNODES; i++) {
6565 /* Set up node groups */
6566 struct sched_group *sg, *prev;
6567 cpumask_t nodemask = node_to_cpumask(i);
6568 cpumask_t domainspan;
6569 cpumask_t covered = CPU_MASK_NONE;
6570 int j;
6572 cpus_and(nodemask, nodemask, *cpu_map);
6573 if (cpus_empty(nodemask)) {
6574 sched_group_nodes[i] = NULL;
6575 continue;
6578 domainspan = sched_domain_node_span(i);
6579 cpus_and(domainspan, domainspan, *cpu_map);
6581 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6582 if (!sg) {
6583 printk(KERN_WARNING "Can not alloc domain group for "
6584 "node %d\n", i);
6585 goto error;
6587 sched_group_nodes[i] = sg;
6588 for_each_cpu_mask(j, nodemask) {
6589 struct sched_domain *sd;
6590 sd = &per_cpu(node_domains, j);
6591 sd->groups = sg;
6593 sg->cpu_power = 0;
6594 sg->cpumask = nodemask;
6595 sg->next = sg;
6596 cpus_or(covered, covered, nodemask);
6597 prev = sg;
6599 for (j = 0; j < MAX_NUMNODES; j++) {
6600 cpumask_t tmp, notcovered;
6601 int n = (i + j) % MAX_NUMNODES;
6603 cpus_complement(notcovered, covered);
6604 cpus_and(tmp, notcovered, *cpu_map);
6605 cpus_and(tmp, tmp, domainspan);
6606 if (cpus_empty(tmp))
6607 break;
6609 nodemask = node_to_cpumask(n);
6610 cpus_and(tmp, tmp, nodemask);
6611 if (cpus_empty(tmp))
6612 continue;
6614 sg = kmalloc_node(sizeof(struct sched_group),
6615 GFP_KERNEL, i);
6616 if (!sg) {
6617 printk(KERN_WARNING
6618 "Can not alloc domain group for node %d\n", j);
6619 goto error;
6621 sg->cpu_power = 0;
6622 sg->cpumask = tmp;
6623 sg->next = prev->next;
6624 cpus_or(covered, covered, tmp);
6625 prev->next = sg;
6626 prev = sg;
6629 #endif
6631 /* Calculate CPU power for physical packages and nodes */
6632 #ifdef CONFIG_SCHED_SMT
6633 for_each_cpu_mask(i, *cpu_map) {
6634 sd = &per_cpu(cpu_domains, i);
6635 init_sched_groups_power(i, sd);
6637 #endif
6638 #ifdef CONFIG_SCHED_MC
6639 for_each_cpu_mask(i, *cpu_map) {
6640 sd = &per_cpu(core_domains, i);
6641 init_sched_groups_power(i, sd);
6643 #endif
6645 for_each_cpu_mask(i, *cpu_map) {
6646 sd = &per_cpu(phys_domains, i);
6647 init_sched_groups_power(i, sd);
6650 #ifdef CONFIG_NUMA
6651 for (i = 0; i < MAX_NUMNODES; i++)
6652 init_numa_sched_groups_power(sched_group_nodes[i]);
6654 if (sd_allnodes) {
6655 struct sched_group *sg;
6657 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6658 init_numa_sched_groups_power(sg);
6660 #endif
6662 /* Attach the domains */
6663 for_each_cpu_mask(i, *cpu_map) {
6664 struct sched_domain *sd;
6665 #ifdef CONFIG_SCHED_SMT
6666 sd = &per_cpu(cpu_domains, i);
6667 #elif defined(CONFIG_SCHED_MC)
6668 sd = &per_cpu(core_domains, i);
6669 #else
6670 sd = &per_cpu(phys_domains, i);
6671 #endif
6672 cpu_attach_domain(sd, i);
6675 * Tune cache-hot values:
6677 calibrate_migration_costs(cpu_map);
6679 return 0;
6681 #ifdef CONFIG_NUMA
6682 error:
6683 free_sched_groups(cpu_map);
6684 return -ENOMEM;
6685 #endif
6688 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6690 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6692 cpumask_t cpu_default_map;
6693 int err;
6696 * Setup mask for cpus without special case scheduling requirements.
6697 * For now this just excludes isolated cpus, but could be used to
6698 * exclude other special cases in the future.
6700 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6702 err = build_sched_domains(&cpu_default_map);
6704 return err;
6707 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6709 free_sched_groups(cpu_map);
6713 * Detach sched domains from a group of cpus specified in cpu_map
6714 * These cpus will now be attached to the NULL domain
6716 static void detach_destroy_domains(const cpumask_t *cpu_map)
6718 int i;
6720 for_each_cpu_mask(i, *cpu_map)
6721 cpu_attach_domain(NULL, i);
6722 synchronize_sched();
6723 arch_destroy_sched_domains(cpu_map);
6727 * Partition sched domains as specified by the cpumasks below.
6728 * This attaches all cpus from the cpumasks to the NULL domain,
6729 * waits for a RCU quiescent period, recalculates sched
6730 * domain information and then attaches them back to the
6731 * correct sched domains
6732 * Call with hotplug lock held
6734 int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6736 cpumask_t change_map;
6737 int err = 0;
6739 cpus_and(*partition1, *partition1, cpu_online_map);
6740 cpus_and(*partition2, *partition2, cpu_online_map);
6741 cpus_or(change_map, *partition1, *partition2);
6743 /* Detach sched domains from all of the affected cpus */
6744 detach_destroy_domains(&change_map);
6745 if (!cpus_empty(*partition1))
6746 err = build_sched_domains(partition1);
6747 if (!err && !cpus_empty(*partition2))
6748 err = build_sched_domains(partition2);
6750 return err;
6753 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6754 int arch_reinit_sched_domains(void)
6756 int err;
6758 lock_cpu_hotplug();
6759 detach_destroy_domains(&cpu_online_map);
6760 err = arch_init_sched_domains(&cpu_online_map);
6761 unlock_cpu_hotplug();
6763 return err;
6766 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6768 int ret;
6770 if (buf[0] != '0' && buf[0] != '1')
6771 return -EINVAL;
6773 if (smt)
6774 sched_smt_power_savings = (buf[0] == '1');
6775 else
6776 sched_mc_power_savings = (buf[0] == '1');
6778 ret = arch_reinit_sched_domains();
6780 return ret ? ret : count;
6783 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6785 int err = 0;
6787 #ifdef CONFIG_SCHED_SMT
6788 if (smt_capable())
6789 err = sysfs_create_file(&cls->kset.kobj,
6790 &attr_sched_smt_power_savings.attr);
6791 #endif
6792 #ifdef CONFIG_SCHED_MC
6793 if (!err && mc_capable())
6794 err = sysfs_create_file(&cls->kset.kobj,
6795 &attr_sched_mc_power_savings.attr);
6796 #endif
6797 return err;
6799 #endif
6801 #ifdef CONFIG_SCHED_MC
6802 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6804 return sprintf(page, "%u\n", sched_mc_power_savings);
6806 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6807 const char *buf, size_t count)
6809 return sched_power_savings_store(buf, count, 0);
6811 SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6812 sched_mc_power_savings_store);
6813 #endif
6815 #ifdef CONFIG_SCHED_SMT
6816 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6818 return sprintf(page, "%u\n", sched_smt_power_savings);
6820 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6821 const char *buf, size_t count)
6823 return sched_power_savings_store(buf, count, 1);
6825 SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6826 sched_smt_power_savings_store);
6827 #endif
6830 * Force a reinitialization of the sched domains hierarchy. The domains
6831 * and groups cannot be updated in place without racing with the balancing
6832 * code, so we temporarily attach all running cpus to the NULL domain
6833 * which will prevent rebalancing while the sched domains are recalculated.
6835 static int update_sched_domains(struct notifier_block *nfb,
6836 unsigned long action, void *hcpu)
6838 switch (action) {
6839 case CPU_UP_PREPARE:
6840 case CPU_DOWN_PREPARE:
6841 detach_destroy_domains(&cpu_online_map);
6842 return NOTIFY_OK;
6844 case CPU_UP_CANCELED:
6845 case CPU_DOWN_FAILED:
6846 case CPU_ONLINE:
6847 case CPU_DEAD:
6849 * Fall through and re-initialise the domains.
6851 break;
6852 default:
6853 return NOTIFY_DONE;
6856 /* The hotplug lock is already held by cpu_up/cpu_down */
6857 arch_init_sched_domains(&cpu_online_map);
6859 return NOTIFY_OK;
6862 void __init sched_init_smp(void)
6864 cpumask_t non_isolated_cpus;
6866 lock_cpu_hotplug();
6867 arch_init_sched_domains(&cpu_online_map);
6868 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6869 if (cpus_empty(non_isolated_cpus))
6870 cpu_set(smp_processor_id(), non_isolated_cpus);
6871 unlock_cpu_hotplug();
6872 /* XXX: Theoretical race here - CPU may be hotplugged now */
6873 hotcpu_notifier(update_sched_domains, 0);
6875 /* Move init over to a non-isolated CPU */
6876 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6877 BUG();
6879 #else
6880 void __init sched_init_smp(void)
6883 #endif /* CONFIG_SMP */
6885 int in_sched_functions(unsigned long addr)
6887 /* Linker adds these: start and end of __sched functions */
6888 extern char __sched_text_start[], __sched_text_end[];
6890 return in_lock_functions(addr) ||
6891 (addr >= (unsigned long)__sched_text_start
6892 && addr < (unsigned long)__sched_text_end);
6895 void __init sched_init(void)
6897 int i, j, k;
6899 for_each_possible_cpu(i) {
6900 struct prio_array *array;
6901 struct rq *rq;
6903 rq = cpu_rq(i);
6904 spin_lock_init(&rq->lock);
6905 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6906 rq->nr_running = 0;
6907 rq->active = rq->arrays;
6908 rq->expired = rq->arrays + 1;
6909 rq->best_expired_prio = MAX_PRIO;
6911 #ifdef CONFIG_SMP
6912 rq->sd = NULL;
6913 for (j = 1; j < 3; j++)
6914 rq->cpu_load[j] = 0;
6915 rq->active_balance = 0;
6916 rq->push_cpu = 0;
6917 rq->cpu = i;
6918 rq->migration_thread = NULL;
6919 INIT_LIST_HEAD(&rq->migration_queue);
6920 #endif
6921 atomic_set(&rq->nr_iowait, 0);
6923 for (j = 0; j < 2; j++) {
6924 array = rq->arrays + j;
6925 for (k = 0; k < MAX_PRIO; k++) {
6926 INIT_LIST_HEAD(array->queue + k);
6927 __clear_bit(k, array->bitmap);
6929 // delimiter for bitsearch
6930 __set_bit(MAX_PRIO, array->bitmap);
6934 set_load_weight(&init_task);
6936 #ifdef CONFIG_SMP
6937 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6938 #endif
6940 #ifdef CONFIG_RT_MUTEXES
6941 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6942 #endif
6945 * The boot idle thread does lazy MMU switching as well:
6947 atomic_inc(&init_mm.mm_count);
6948 enter_lazy_tlb(&init_mm, current);
6951 * Make us the idle thread. Technically, schedule() should not be
6952 * called from this thread, however somewhere below it might be,
6953 * but because we are the idle thread, we just pick up running again
6954 * when this runqueue becomes "idle".
6956 init_idle(current, smp_processor_id());
6959 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6960 void __might_sleep(char *file, int line)
6962 #ifdef in_atomic
6963 static unsigned long prev_jiffy; /* ratelimiting */
6965 if ((in_atomic() || irqs_disabled()) &&
6966 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6967 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6968 return;
6969 prev_jiffy = jiffies;
6970 printk(KERN_ERR "BUG: sleeping function called from invalid"
6971 " context at %s:%d\n", file, line);
6972 printk("in_atomic():%d, irqs_disabled():%d\n",
6973 in_atomic(), irqs_disabled());
6974 debug_show_held_locks(current);
6975 if (irqs_disabled())
6976 print_irqtrace_events(current);
6977 dump_stack();
6979 #endif
6981 EXPORT_SYMBOL(__might_sleep);
6982 #endif
6984 #ifdef CONFIG_MAGIC_SYSRQ
6985 void normalize_rt_tasks(void)
6987 struct prio_array *array;
6988 struct task_struct *p;
6989 unsigned long flags;
6990 struct rq *rq;
6992 read_lock_irq(&tasklist_lock);
6993 for_each_process(p) {
6994 if (!rt_task(p))
6995 continue;
6997 spin_lock_irqsave(&p->pi_lock, flags);
6998 rq = __task_rq_lock(p);
7000 array = p->array;
7001 if (array)
7002 deactivate_task(p, task_rq(p));
7003 __setscheduler(p, SCHED_NORMAL, 0);
7004 if (array) {
7005 __activate_task(p, task_rq(p));
7006 resched_task(rq->curr);
7009 __task_rq_unlock(rq);
7010 spin_unlock_irqrestore(&p->pi_lock, flags);
7012 read_unlock_irq(&tasklist_lock);
7015 #endif /* CONFIG_MAGIC_SYSRQ */
7017 #ifdef CONFIG_IA64
7019 * These functions are only useful for the IA64 MCA handling.
7021 * They can only be called when the whole system has been
7022 * stopped - every CPU needs to be quiescent, and no scheduling
7023 * activity can take place. Using them for anything else would
7024 * be a serious bug, and as a result, they aren't even visible
7025 * under any other configuration.
7029 * curr_task - return the current task for a given cpu.
7030 * @cpu: the processor in question.
7032 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7034 struct task_struct *curr_task(int cpu)
7036 return cpu_curr(cpu);
7040 * set_curr_task - set the current task for a given cpu.
7041 * @cpu: the processor in question.
7042 * @p: the task pointer to set.
7044 * Description: This function must only be used when non-maskable interrupts
7045 * are serviced on a separate stack. It allows the architecture to switch the
7046 * notion of the current task on a cpu in a non-blocking manner. This function
7047 * must be called with all CPU's synchronized, and interrupts disabled, the
7048 * and caller must save the original value of the current task (see
7049 * curr_task() above) and restore that value before reenabling interrupts and
7050 * re-starting the system.
7052 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7054 void set_curr_task(int cpu, struct task_struct *p)
7056 cpu_curr(cpu) = p;
7059 #endif