[SPARC64]: Fix 2 bugs in huge page support.
[linux-2.6/verdex.git] / kernel / sched.c
blob4d46e90f59c32fcbe88f320af11794c43af9e55a
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/security.h>
34 #include <linux/notifier.h>
35 #include <linux/profile.h>
36 #include <linux/suspend.h>
37 #include <linux/vmalloc.h>
38 #include <linux/blkdev.h>
39 #include <linux/delay.h>
40 #include <linux/smp.h>
41 #include <linux/threads.h>
42 #include <linux/timer.h>
43 #include <linux/rcupdate.h>
44 #include <linux/cpu.h>
45 #include <linux/cpuset.h>
46 #include <linux/percpu.h>
47 #include <linux/kthread.h>
48 #include <linux/seq_file.h>
49 #include <linux/syscalls.h>
50 #include <linux/times.h>
51 #include <linux/acct.h>
52 #include <asm/tlb.h>
54 #include <asm/unistd.h>
57 * Convert user-nice values [ -20 ... 0 ... 19 ]
58 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
59 * and back.
61 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
62 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
63 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
66 * 'User priority' is the nice value converted to something we
67 * can work with better when scaling various scheduler parameters,
68 * it's a [ 0 ... 39 ] range.
70 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
71 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
72 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
75 * Some helpers for converting nanosecond timing to jiffy resolution
77 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
78 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
81 * These are the 'tuning knobs' of the scheduler:
83 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
84 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
85 * Timeslices get refilled after they expire.
87 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
88 #define DEF_TIMESLICE (100 * HZ / 1000)
89 #define ON_RUNQUEUE_WEIGHT 30
90 #define CHILD_PENALTY 95
91 #define PARENT_PENALTY 100
92 #define EXIT_WEIGHT 3
93 #define PRIO_BONUS_RATIO 25
94 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
95 #define INTERACTIVE_DELTA 2
96 #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
97 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
98 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
101 * If a task is 'interactive' then we reinsert it in the active
102 * array after it has expired its current timeslice. (it will not
103 * continue to run immediately, it will still roundrobin with
104 * other interactive tasks.)
106 * This part scales the interactivity limit depending on niceness.
108 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
109 * Here are a few examples of different nice levels:
111 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
112 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
113 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
114 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
115 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
117 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
118 * priority range a task can explore, a value of '1' means the
119 * task is rated interactive.)
121 * Ie. nice +19 tasks can never get 'interactive' enough to be
122 * reinserted into the active array. And only heavily CPU-hog nice -20
123 * tasks will be expired. Default nice 0 tasks are somewhere between,
124 * it takes some effort for them to get interactive, but it's not
125 * too hard.
128 #define CURRENT_BONUS(p) \
129 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
130 MAX_SLEEP_AVG)
132 #define GRANULARITY (10 * HZ / 1000 ? : 1)
134 #ifdef CONFIG_SMP
135 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
136 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
137 num_online_cpus())
138 #else
139 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
140 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
141 #endif
143 #define SCALE(v1,v1_max,v2_max) \
144 (v1) * (v2_max) / (v1_max)
146 #define DELTA(p) \
147 (SCALE(TASK_NICE(p), 40, MAX_BONUS) + INTERACTIVE_DELTA)
149 #define TASK_INTERACTIVE(p) \
150 ((p)->prio <= (p)->static_prio - DELTA(p))
152 #define INTERACTIVE_SLEEP(p) \
153 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
154 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
156 #define TASK_PREEMPTS_CURR(p, rq) \
157 ((p)->prio < (rq)->curr->prio)
160 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
161 * to time slice values: [800ms ... 100ms ... 5ms]
163 * The higher a thread's priority, the bigger timeslices
164 * it gets during one round of execution. But even the lowest
165 * priority thread gets MIN_TIMESLICE worth of execution time.
168 #define SCALE_PRIO(x, prio) \
169 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO/2), MIN_TIMESLICE)
171 static unsigned int task_timeslice(task_t *p)
173 if (p->static_prio < NICE_TO_PRIO(0))
174 return SCALE_PRIO(DEF_TIMESLICE*4, p->static_prio);
175 else
176 return SCALE_PRIO(DEF_TIMESLICE, p->static_prio);
178 #define task_hot(p, now, sd) ((long long) ((now) - (p)->last_ran) \
179 < (long long) (sd)->cache_hot_time)
182 * These are the runqueue data structures:
185 #define BITMAP_SIZE ((((MAX_PRIO+1+7)/8)+sizeof(long)-1)/sizeof(long))
187 typedef struct runqueue runqueue_t;
189 struct prio_array {
190 unsigned int nr_active;
191 unsigned long bitmap[BITMAP_SIZE];
192 struct list_head queue[MAX_PRIO];
196 * This is the main, per-CPU runqueue data structure.
198 * Locking rule: those places that want to lock multiple runqueues
199 * (such as the load balancing or the thread migration code), lock
200 * acquire operations must be ordered by ascending &runqueue.
202 struct runqueue {
203 spinlock_t lock;
206 * nr_running and cpu_load should be in the same cacheline because
207 * remote CPUs use both these fields when doing load calculation.
209 unsigned long nr_running;
210 #ifdef CONFIG_SMP
211 unsigned long cpu_load[3];
212 #endif
213 unsigned long long nr_switches;
216 * This is part of a global counter where only the total sum
217 * over all CPUs matters. A task can increase this counter on
218 * one CPU and if it got migrated afterwards it may decrease
219 * it on another CPU. Always updated under the runqueue lock:
221 unsigned long nr_uninterruptible;
223 unsigned long expired_timestamp;
224 unsigned long long timestamp_last_tick;
225 task_t *curr, *idle;
226 struct mm_struct *prev_mm;
227 prio_array_t *active, *expired, arrays[2];
228 int best_expired_prio;
229 atomic_t nr_iowait;
231 #ifdef CONFIG_SMP
232 struct sched_domain *sd;
234 /* For active balancing */
235 int active_balance;
236 int push_cpu;
238 task_t *migration_thread;
239 struct list_head migration_queue;
240 #endif
242 #ifdef CONFIG_SCHEDSTATS
243 /* latency stats */
244 struct sched_info rq_sched_info;
246 /* sys_sched_yield() stats */
247 unsigned long yld_exp_empty;
248 unsigned long yld_act_empty;
249 unsigned long yld_both_empty;
250 unsigned long yld_cnt;
252 /* schedule() stats */
253 unsigned long sched_switch;
254 unsigned long sched_cnt;
255 unsigned long sched_goidle;
257 /* try_to_wake_up() stats */
258 unsigned long ttwu_cnt;
259 unsigned long ttwu_local;
260 #endif
263 static DEFINE_PER_CPU(struct runqueue, runqueues);
266 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
267 * See detach_destroy_domains: synchronize_sched for details.
269 * The domain tree of any CPU may only be accessed from within
270 * preempt-disabled sections.
272 #define for_each_domain(cpu, domain) \
273 for (domain = rcu_dereference(cpu_rq(cpu)->sd); domain; domain = domain->parent)
275 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
276 #define this_rq() (&__get_cpu_var(runqueues))
277 #define task_rq(p) cpu_rq(task_cpu(p))
278 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
280 #ifndef prepare_arch_switch
281 # define prepare_arch_switch(next) do { } while (0)
282 #endif
283 #ifndef finish_arch_switch
284 # define finish_arch_switch(prev) do { } while (0)
285 #endif
287 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
288 static inline int task_running(runqueue_t *rq, task_t *p)
290 return rq->curr == p;
293 static inline void prepare_lock_switch(runqueue_t *rq, task_t *next)
297 static inline void finish_lock_switch(runqueue_t *rq, task_t *prev)
299 #ifdef CONFIG_DEBUG_SPINLOCK
300 /* this is a valid case when another task releases the spinlock */
301 rq->lock.owner = current;
302 #endif
303 spin_unlock_irq(&rq->lock);
306 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
307 static inline int task_running(runqueue_t *rq, task_t *p)
309 #ifdef CONFIG_SMP
310 return p->oncpu;
311 #else
312 return rq->curr == p;
313 #endif
316 static inline void prepare_lock_switch(runqueue_t *rq, task_t *next)
318 #ifdef CONFIG_SMP
320 * We can optimise this out completely for !SMP, because the
321 * SMP rebalancing from interrupt is the only thing that cares
322 * here.
324 next->oncpu = 1;
325 #endif
326 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
327 spin_unlock_irq(&rq->lock);
328 #else
329 spin_unlock(&rq->lock);
330 #endif
333 static inline void finish_lock_switch(runqueue_t *rq, task_t *prev)
335 #ifdef CONFIG_SMP
337 * After ->oncpu is cleared, the task can be moved to a different CPU.
338 * We must ensure this doesn't happen until the switch is completely
339 * finished.
341 smp_wmb();
342 prev->oncpu = 0;
343 #endif
344 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
345 local_irq_enable();
346 #endif
348 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
351 * task_rq_lock - lock the runqueue a given task resides on and disable
352 * interrupts. Note the ordering: we can safely lookup the task_rq without
353 * explicitly disabling preemption.
355 static inline runqueue_t *task_rq_lock(task_t *p, unsigned long *flags)
356 __acquires(rq->lock)
358 struct runqueue *rq;
360 repeat_lock_task:
361 local_irq_save(*flags);
362 rq = task_rq(p);
363 spin_lock(&rq->lock);
364 if (unlikely(rq != task_rq(p))) {
365 spin_unlock_irqrestore(&rq->lock, *flags);
366 goto repeat_lock_task;
368 return rq;
371 static inline void task_rq_unlock(runqueue_t *rq, unsigned long *flags)
372 __releases(rq->lock)
374 spin_unlock_irqrestore(&rq->lock, *flags);
377 #ifdef CONFIG_SCHEDSTATS
379 * bump this up when changing the output format or the meaning of an existing
380 * format, so that tools can adapt (or abort)
382 #define SCHEDSTAT_VERSION 12
384 static int show_schedstat(struct seq_file *seq, void *v)
386 int cpu;
388 seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION);
389 seq_printf(seq, "timestamp %lu\n", jiffies);
390 for_each_online_cpu(cpu) {
391 runqueue_t *rq = cpu_rq(cpu);
392 #ifdef CONFIG_SMP
393 struct sched_domain *sd;
394 int dcnt = 0;
395 #endif
397 /* runqueue-specific stats */
398 seq_printf(seq,
399 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
400 cpu, rq->yld_both_empty,
401 rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt,
402 rq->sched_switch, rq->sched_cnt, rq->sched_goidle,
403 rq->ttwu_cnt, rq->ttwu_local,
404 rq->rq_sched_info.cpu_time,
405 rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt);
407 seq_printf(seq, "\n");
409 #ifdef CONFIG_SMP
410 /* domain-specific stats */
411 preempt_disable();
412 for_each_domain(cpu, sd) {
413 enum idle_type itype;
414 char mask_str[NR_CPUS];
416 cpumask_scnprintf(mask_str, NR_CPUS, sd->span);
417 seq_printf(seq, "domain%d %s", dcnt++, mask_str);
418 for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES;
419 itype++) {
420 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu",
421 sd->lb_cnt[itype],
422 sd->lb_balanced[itype],
423 sd->lb_failed[itype],
424 sd->lb_imbalance[itype],
425 sd->lb_gained[itype],
426 sd->lb_hot_gained[itype],
427 sd->lb_nobusyq[itype],
428 sd->lb_nobusyg[itype]);
430 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
431 sd->alb_cnt, sd->alb_failed, sd->alb_pushed,
432 sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed,
433 sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed,
434 sd->ttwu_wake_remote, sd->ttwu_move_affine, sd->ttwu_move_balance);
436 preempt_enable();
437 #endif
439 return 0;
442 static int schedstat_open(struct inode *inode, struct file *file)
444 unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32);
445 char *buf = kmalloc(size, GFP_KERNEL);
446 struct seq_file *m;
447 int res;
449 if (!buf)
450 return -ENOMEM;
451 res = single_open(file, show_schedstat, NULL);
452 if (!res) {
453 m = file->private_data;
454 m->buf = buf;
455 m->size = size;
456 } else
457 kfree(buf);
458 return res;
461 struct file_operations proc_schedstat_operations = {
462 .open = schedstat_open,
463 .read = seq_read,
464 .llseek = seq_lseek,
465 .release = single_release,
468 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
469 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
470 #else /* !CONFIG_SCHEDSTATS */
471 # define schedstat_inc(rq, field) do { } while (0)
472 # define schedstat_add(rq, field, amt) do { } while (0)
473 #endif
476 * rq_lock - lock a given runqueue and disable interrupts.
478 static inline runqueue_t *this_rq_lock(void)
479 __acquires(rq->lock)
481 runqueue_t *rq;
483 local_irq_disable();
484 rq = this_rq();
485 spin_lock(&rq->lock);
487 return rq;
490 #ifdef CONFIG_SCHEDSTATS
492 * Called when a process is dequeued from the active array and given
493 * the cpu. We should note that with the exception of interactive
494 * tasks, the expired queue will become the active queue after the active
495 * queue is empty, without explicitly dequeuing and requeuing tasks in the
496 * expired queue. (Interactive tasks may be requeued directly to the
497 * active queue, thus delaying tasks in the expired queue from running;
498 * see scheduler_tick()).
500 * This function is only called from sched_info_arrive(), rather than
501 * dequeue_task(). Even though a task may be queued and dequeued multiple
502 * times as it is shuffled about, we're really interested in knowing how
503 * long it was from the *first* time it was queued to the time that it
504 * finally hit a cpu.
506 static inline void sched_info_dequeued(task_t *t)
508 t->sched_info.last_queued = 0;
512 * Called when a task finally hits the cpu. We can now calculate how
513 * long it was waiting to run. We also note when it began so that we
514 * can keep stats on how long its timeslice is.
516 static void sched_info_arrive(task_t *t)
518 unsigned long now = jiffies, diff = 0;
519 struct runqueue *rq = task_rq(t);
521 if (t->sched_info.last_queued)
522 diff = now - t->sched_info.last_queued;
523 sched_info_dequeued(t);
524 t->sched_info.run_delay += diff;
525 t->sched_info.last_arrival = now;
526 t->sched_info.pcnt++;
528 if (!rq)
529 return;
531 rq->rq_sched_info.run_delay += diff;
532 rq->rq_sched_info.pcnt++;
536 * Called when a process is queued into either the active or expired
537 * array. The time is noted and later used to determine how long we
538 * had to wait for us to reach the cpu. Since the expired queue will
539 * become the active queue after active queue is empty, without dequeuing
540 * and requeuing any tasks, we are interested in queuing to either. It
541 * is unusual but not impossible for tasks to be dequeued and immediately
542 * requeued in the same or another array: this can happen in sched_yield(),
543 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
544 * to runqueue.
546 * This function is only called from enqueue_task(), but also only updates
547 * the timestamp if it is already not set. It's assumed that
548 * sched_info_dequeued() will clear that stamp when appropriate.
550 static inline void sched_info_queued(task_t *t)
552 if (!t->sched_info.last_queued)
553 t->sched_info.last_queued = jiffies;
557 * Called when a process ceases being the active-running process, either
558 * voluntarily or involuntarily. Now we can calculate how long we ran.
560 static inline void sched_info_depart(task_t *t)
562 struct runqueue *rq = task_rq(t);
563 unsigned long diff = jiffies - t->sched_info.last_arrival;
565 t->sched_info.cpu_time += diff;
567 if (rq)
568 rq->rq_sched_info.cpu_time += diff;
572 * Called when tasks are switched involuntarily due, typically, to expiring
573 * their time slice. (This may also be called when switching to or from
574 * the idle task.) We are only called when prev != next.
576 static inline void sched_info_switch(task_t *prev, task_t *next)
578 struct runqueue *rq = task_rq(prev);
581 * prev now departs the cpu. It's not interesting to record
582 * stats about how efficient we were at scheduling the idle
583 * process, however.
585 if (prev != rq->idle)
586 sched_info_depart(prev);
588 if (next != rq->idle)
589 sched_info_arrive(next);
591 #else
592 #define sched_info_queued(t) do { } while (0)
593 #define sched_info_switch(t, next) do { } while (0)
594 #endif /* CONFIG_SCHEDSTATS */
597 * Adding/removing a task to/from a priority array:
599 static void dequeue_task(struct task_struct *p, prio_array_t *array)
601 array->nr_active--;
602 list_del(&p->run_list);
603 if (list_empty(array->queue + p->prio))
604 __clear_bit(p->prio, array->bitmap);
607 static void enqueue_task(struct task_struct *p, prio_array_t *array)
609 sched_info_queued(p);
610 list_add_tail(&p->run_list, array->queue + p->prio);
611 __set_bit(p->prio, array->bitmap);
612 array->nr_active++;
613 p->array = array;
617 * Put task to the end of the run list without the overhead of dequeue
618 * followed by enqueue.
620 static void requeue_task(struct task_struct *p, prio_array_t *array)
622 list_move_tail(&p->run_list, array->queue + p->prio);
625 static inline void enqueue_task_head(struct task_struct *p, prio_array_t *array)
627 list_add(&p->run_list, array->queue + p->prio);
628 __set_bit(p->prio, array->bitmap);
629 array->nr_active++;
630 p->array = array;
634 * effective_prio - return the priority that is based on the static
635 * priority but is modified by bonuses/penalties.
637 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
638 * into the -5 ... 0 ... +5 bonus/penalty range.
640 * We use 25% of the full 0...39 priority range so that:
642 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
643 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
645 * Both properties are important to certain workloads.
647 static int effective_prio(task_t *p)
649 int bonus, prio;
651 if (rt_task(p))
652 return p->prio;
654 bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
656 prio = p->static_prio - bonus;
657 if (prio < MAX_RT_PRIO)
658 prio = MAX_RT_PRIO;
659 if (prio > MAX_PRIO-1)
660 prio = MAX_PRIO-1;
661 return prio;
665 * __activate_task - move a task to the runqueue.
667 static inline void __activate_task(task_t *p, runqueue_t *rq)
669 enqueue_task(p, rq->active);
670 rq->nr_running++;
674 * __activate_idle_task - move idle task to the _front_ of runqueue.
676 static inline void __activate_idle_task(task_t *p, runqueue_t *rq)
678 enqueue_task_head(p, rq->active);
679 rq->nr_running++;
682 static int recalc_task_prio(task_t *p, unsigned long long now)
684 /* Caller must always ensure 'now >= p->timestamp' */
685 unsigned long long __sleep_time = now - p->timestamp;
686 unsigned long sleep_time;
688 if (unlikely(p->policy == SCHED_BATCH))
689 sleep_time = 0;
690 else {
691 if (__sleep_time > NS_MAX_SLEEP_AVG)
692 sleep_time = NS_MAX_SLEEP_AVG;
693 else
694 sleep_time = (unsigned long)__sleep_time;
697 if (likely(sleep_time > 0)) {
699 * User tasks that sleep a long time are categorised as
700 * idle and will get just interactive status to stay active &
701 * prevent them suddenly becoming cpu hogs and starving
702 * other processes.
704 if (p->mm && p->activated != -1 &&
705 sleep_time > INTERACTIVE_SLEEP(p)) {
706 p->sleep_avg = JIFFIES_TO_NS(MAX_SLEEP_AVG -
707 DEF_TIMESLICE);
708 } else {
710 * The lower the sleep avg a task has the more
711 * rapidly it will rise with sleep time.
713 sleep_time *= (MAX_BONUS - CURRENT_BONUS(p)) ? : 1;
716 * Tasks waking from uninterruptible sleep are
717 * limited in their sleep_avg rise as they
718 * are likely to be waiting on I/O
720 if (p->activated == -1 && p->mm) {
721 if (p->sleep_avg >= INTERACTIVE_SLEEP(p))
722 sleep_time = 0;
723 else if (p->sleep_avg + sleep_time >=
724 INTERACTIVE_SLEEP(p)) {
725 p->sleep_avg = INTERACTIVE_SLEEP(p);
726 sleep_time = 0;
731 * This code gives a bonus to interactive tasks.
733 * The boost works by updating the 'average sleep time'
734 * value here, based on ->timestamp. The more time a
735 * task spends sleeping, the higher the average gets -
736 * and the higher the priority boost gets as well.
738 p->sleep_avg += sleep_time;
740 if (p->sleep_avg > NS_MAX_SLEEP_AVG)
741 p->sleep_avg = NS_MAX_SLEEP_AVG;
745 return effective_prio(p);
749 * activate_task - move a task to the runqueue and do priority recalculation
751 * Update all the scheduling statistics stuff. (sleep average
752 * calculation, priority modifiers, etc.)
754 static void activate_task(task_t *p, runqueue_t *rq, int local)
756 unsigned long long now;
758 now = sched_clock();
759 #ifdef CONFIG_SMP
760 if (!local) {
761 /* Compensate for drifting sched_clock */
762 runqueue_t *this_rq = this_rq();
763 now = (now - this_rq->timestamp_last_tick)
764 + rq->timestamp_last_tick;
766 #endif
768 if (!rt_task(p))
769 p->prio = recalc_task_prio(p, now);
772 * This checks to make sure it's not an uninterruptible task
773 * that is now waking up.
775 if (!p->activated) {
777 * Tasks which were woken up by interrupts (ie. hw events)
778 * are most likely of interactive nature. So we give them
779 * the credit of extending their sleep time to the period
780 * of time they spend on the runqueue, waiting for execution
781 * on a CPU, first time around:
783 if (in_interrupt())
784 p->activated = 2;
785 else {
787 * Normal first-time wakeups get a credit too for
788 * on-runqueue time, but it will be weighted down:
790 p->activated = 1;
793 p->timestamp = now;
795 __activate_task(p, rq);
799 * deactivate_task - remove a task from the runqueue.
801 static void deactivate_task(struct task_struct *p, runqueue_t *rq)
803 rq->nr_running--;
804 dequeue_task(p, p->array);
805 p->array = NULL;
809 * resched_task - mark a task 'to be rescheduled now'.
811 * On UP this means the setting of the need_resched flag, on SMP it
812 * might also involve a cross-CPU call to trigger the scheduler on
813 * the target CPU.
815 #ifdef CONFIG_SMP
816 static void resched_task(task_t *p)
818 int cpu;
820 assert_spin_locked(&task_rq(p)->lock);
822 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
823 return;
825 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
827 cpu = task_cpu(p);
828 if (cpu == smp_processor_id())
829 return;
831 /* NEED_RESCHED must be visible before we test POLLING_NRFLAG */
832 smp_mb();
833 if (!test_tsk_thread_flag(p, TIF_POLLING_NRFLAG))
834 smp_send_reschedule(cpu);
836 #else
837 static inline void resched_task(task_t *p)
839 assert_spin_locked(&task_rq(p)->lock);
840 set_tsk_need_resched(p);
842 #endif
845 * task_curr - is this task currently executing on a CPU?
846 * @p: the task in question.
848 inline int task_curr(const task_t *p)
850 return cpu_curr(task_cpu(p)) == p;
853 #ifdef CONFIG_SMP
854 typedef struct {
855 struct list_head list;
857 task_t *task;
858 int dest_cpu;
860 struct completion done;
861 } migration_req_t;
864 * The task's runqueue lock must be held.
865 * Returns true if you have to wait for migration thread.
867 static int migrate_task(task_t *p, int dest_cpu, migration_req_t *req)
869 runqueue_t *rq = task_rq(p);
872 * If the task is not on a runqueue (and not running), then
873 * it is sufficient to simply update the task's cpu field.
875 if (!p->array && !task_running(rq, p)) {
876 set_task_cpu(p, dest_cpu);
877 return 0;
880 init_completion(&req->done);
881 req->task = p;
882 req->dest_cpu = dest_cpu;
883 list_add(&req->list, &rq->migration_queue);
884 return 1;
888 * wait_task_inactive - wait for a thread to unschedule.
890 * The caller must ensure that the task *will* unschedule sometime soon,
891 * else this function might spin for a *long* time. This function can't
892 * be called with interrupts off, or it may introduce deadlock with
893 * smp_call_function() if an IPI is sent by the same process we are
894 * waiting to become inactive.
896 void wait_task_inactive(task_t *p)
898 unsigned long flags;
899 runqueue_t *rq;
900 int preempted;
902 repeat:
903 rq = task_rq_lock(p, &flags);
904 /* Must be off runqueue entirely, not preempted. */
905 if (unlikely(p->array || task_running(rq, p))) {
906 /* If it's preempted, we yield. It could be a while. */
907 preempted = !task_running(rq, p);
908 task_rq_unlock(rq, &flags);
909 cpu_relax();
910 if (preempted)
911 yield();
912 goto repeat;
914 task_rq_unlock(rq, &flags);
917 /***
918 * kick_process - kick a running thread to enter/exit the kernel
919 * @p: the to-be-kicked thread
921 * Cause a process which is running on another CPU to enter
922 * kernel-mode, without any delay. (to get signals handled.)
924 * NOTE: this function doesnt have to take the runqueue lock,
925 * because all it wants to ensure is that the remote task enters
926 * the kernel. If the IPI races and the task has been migrated
927 * to another CPU then no harm is done and the purpose has been
928 * achieved as well.
930 void kick_process(task_t *p)
932 int cpu;
934 preempt_disable();
935 cpu = task_cpu(p);
936 if ((cpu != smp_processor_id()) && task_curr(p))
937 smp_send_reschedule(cpu);
938 preempt_enable();
942 * Return a low guess at the load of a migration-source cpu.
944 * We want to under-estimate the load of migration sources, to
945 * balance conservatively.
947 static inline unsigned long source_load(int cpu, int type)
949 runqueue_t *rq = cpu_rq(cpu);
950 unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE;
951 if (type == 0)
952 return load_now;
954 return min(rq->cpu_load[type-1], load_now);
958 * Return a high guess at the load of a migration-target cpu
960 static inline unsigned long target_load(int cpu, int type)
962 runqueue_t *rq = cpu_rq(cpu);
963 unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE;
964 if (type == 0)
965 return load_now;
967 return max(rq->cpu_load[type-1], load_now);
971 * find_idlest_group finds and returns the least busy CPU group within the
972 * domain.
974 static struct sched_group *
975 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
977 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
978 unsigned long min_load = ULONG_MAX, this_load = 0;
979 int load_idx = sd->forkexec_idx;
980 int imbalance = 100 + (sd->imbalance_pct-100)/2;
982 do {
983 unsigned long load, avg_load;
984 int local_group;
985 int i;
987 /* Skip over this group if it has no CPUs allowed */
988 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
989 goto nextgroup;
991 local_group = cpu_isset(this_cpu, group->cpumask);
993 /* Tally up the load of all CPUs in the group */
994 avg_load = 0;
996 for_each_cpu_mask(i, group->cpumask) {
997 /* Bias balancing toward cpus of our domain */
998 if (local_group)
999 load = source_load(i, load_idx);
1000 else
1001 load = target_load(i, load_idx);
1003 avg_load += load;
1006 /* Adjust by relative CPU power of the group */
1007 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1009 if (local_group) {
1010 this_load = avg_load;
1011 this = group;
1012 } else if (avg_load < min_load) {
1013 min_load = avg_load;
1014 idlest = group;
1016 nextgroup:
1017 group = group->next;
1018 } while (group != sd->groups);
1020 if (!idlest || 100*this_load < imbalance*min_load)
1021 return NULL;
1022 return idlest;
1026 * find_idlest_queue - find the idlest runqueue among the cpus in group.
1028 static int
1029 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1031 cpumask_t tmp;
1032 unsigned long load, min_load = ULONG_MAX;
1033 int idlest = -1;
1034 int i;
1036 /* Traverse only the allowed CPUs */
1037 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1039 for_each_cpu_mask(i, tmp) {
1040 load = source_load(i, 0);
1042 if (load < min_load || (load == min_load && i == this_cpu)) {
1043 min_load = load;
1044 idlest = i;
1048 return idlest;
1052 * sched_balance_self: balance the current task (running on cpu) in domains
1053 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1054 * SD_BALANCE_EXEC.
1056 * Balance, ie. select the least loaded group.
1058 * Returns the target CPU number, or the same CPU if no balancing is needed.
1060 * preempt must be disabled.
1062 static int sched_balance_self(int cpu, int flag)
1064 struct task_struct *t = current;
1065 struct sched_domain *tmp, *sd = NULL;
1067 for_each_domain(cpu, tmp)
1068 if (tmp->flags & flag)
1069 sd = tmp;
1071 while (sd) {
1072 cpumask_t span;
1073 struct sched_group *group;
1074 int new_cpu;
1075 int weight;
1077 span = sd->span;
1078 group = find_idlest_group(sd, t, cpu);
1079 if (!group)
1080 goto nextlevel;
1082 new_cpu = find_idlest_cpu(group, t, cpu);
1083 if (new_cpu == -1 || new_cpu == cpu)
1084 goto nextlevel;
1086 /* Now try balancing at a lower domain level */
1087 cpu = new_cpu;
1088 nextlevel:
1089 sd = NULL;
1090 weight = cpus_weight(span);
1091 for_each_domain(cpu, tmp) {
1092 if (weight <= cpus_weight(tmp->span))
1093 break;
1094 if (tmp->flags & flag)
1095 sd = tmp;
1097 /* while loop will break here if sd == NULL */
1100 return cpu;
1103 #endif /* CONFIG_SMP */
1106 * wake_idle() will wake a task on an idle cpu if task->cpu is
1107 * not idle and an idle cpu is available. The span of cpus to
1108 * search starts with cpus closest then further out as needed,
1109 * so we always favor a closer, idle cpu.
1111 * Returns the CPU we should wake onto.
1113 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1114 static int wake_idle(int cpu, task_t *p)
1116 cpumask_t tmp;
1117 struct sched_domain *sd;
1118 int i;
1120 if (idle_cpu(cpu))
1121 return cpu;
1123 for_each_domain(cpu, sd) {
1124 if (sd->flags & SD_WAKE_IDLE) {
1125 cpus_and(tmp, sd->span, p->cpus_allowed);
1126 for_each_cpu_mask(i, tmp) {
1127 if (idle_cpu(i))
1128 return i;
1131 else
1132 break;
1134 return cpu;
1136 #else
1137 static inline int wake_idle(int cpu, task_t *p)
1139 return cpu;
1141 #endif
1143 /***
1144 * try_to_wake_up - wake up a thread
1145 * @p: the to-be-woken-up thread
1146 * @state: the mask of task states that can be woken
1147 * @sync: do a synchronous wakeup?
1149 * Put it on the run-queue if it's not already there. The "current"
1150 * thread is always on the run-queue (except when the actual
1151 * re-schedule is in progress), and as such you're allowed to do
1152 * the simpler "current->state = TASK_RUNNING" to mark yourself
1153 * runnable without the overhead of this.
1155 * returns failure only if the task is already active.
1157 static int try_to_wake_up(task_t *p, unsigned int state, int sync)
1159 int cpu, this_cpu, success = 0;
1160 unsigned long flags;
1161 long old_state;
1162 runqueue_t *rq;
1163 #ifdef CONFIG_SMP
1164 unsigned long load, this_load;
1165 struct sched_domain *sd, *this_sd = NULL;
1166 int new_cpu;
1167 #endif
1169 rq = task_rq_lock(p, &flags);
1170 old_state = p->state;
1171 if (!(old_state & state))
1172 goto out;
1174 if (p->array)
1175 goto out_running;
1177 cpu = task_cpu(p);
1178 this_cpu = smp_processor_id();
1180 #ifdef CONFIG_SMP
1181 if (unlikely(task_running(rq, p)))
1182 goto out_activate;
1184 new_cpu = cpu;
1186 schedstat_inc(rq, ttwu_cnt);
1187 if (cpu == this_cpu) {
1188 schedstat_inc(rq, ttwu_local);
1189 goto out_set_cpu;
1192 for_each_domain(this_cpu, sd) {
1193 if (cpu_isset(cpu, sd->span)) {
1194 schedstat_inc(sd, ttwu_wake_remote);
1195 this_sd = sd;
1196 break;
1200 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1201 goto out_set_cpu;
1204 * Check for affine wakeup and passive balancing possibilities.
1206 if (this_sd) {
1207 int idx = this_sd->wake_idx;
1208 unsigned int imbalance;
1210 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1212 load = source_load(cpu, idx);
1213 this_load = target_load(this_cpu, idx);
1215 new_cpu = this_cpu; /* Wake to this CPU if we can */
1217 if (this_sd->flags & SD_WAKE_AFFINE) {
1218 unsigned long tl = this_load;
1220 * If sync wakeup then subtract the (maximum possible)
1221 * effect of the currently running task from the load
1222 * of the current CPU:
1224 if (sync)
1225 tl -= SCHED_LOAD_SCALE;
1227 if ((tl <= load &&
1228 tl + target_load(cpu, idx) <= SCHED_LOAD_SCALE) ||
1229 100*(tl + SCHED_LOAD_SCALE) <= imbalance*load) {
1231 * This domain has SD_WAKE_AFFINE and
1232 * p is cache cold in this domain, and
1233 * there is no bad imbalance.
1235 schedstat_inc(this_sd, ttwu_move_affine);
1236 goto out_set_cpu;
1241 * Start passive balancing when half the imbalance_pct
1242 * limit is reached.
1244 if (this_sd->flags & SD_WAKE_BALANCE) {
1245 if (imbalance*this_load <= 100*load) {
1246 schedstat_inc(this_sd, ttwu_move_balance);
1247 goto out_set_cpu;
1252 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1253 out_set_cpu:
1254 new_cpu = wake_idle(new_cpu, p);
1255 if (new_cpu != cpu) {
1256 set_task_cpu(p, new_cpu);
1257 task_rq_unlock(rq, &flags);
1258 /* might preempt at this point */
1259 rq = task_rq_lock(p, &flags);
1260 old_state = p->state;
1261 if (!(old_state & state))
1262 goto out;
1263 if (p->array)
1264 goto out_running;
1266 this_cpu = smp_processor_id();
1267 cpu = task_cpu(p);
1270 out_activate:
1271 #endif /* CONFIG_SMP */
1272 if (old_state == TASK_UNINTERRUPTIBLE) {
1273 rq->nr_uninterruptible--;
1275 * Tasks on involuntary sleep don't earn
1276 * sleep_avg beyond just interactive state.
1278 p->activated = -1;
1282 * Tasks that have marked their sleep as noninteractive get
1283 * woken up without updating their sleep average. (i.e. their
1284 * sleep is handled in a priority-neutral manner, no priority
1285 * boost and no penalty.)
1287 if (old_state & TASK_NONINTERACTIVE)
1288 __activate_task(p, rq);
1289 else
1290 activate_task(p, rq, cpu == this_cpu);
1292 * Sync wakeups (i.e. those types of wakeups where the waker
1293 * has indicated that it will leave the CPU in short order)
1294 * don't trigger a preemption, if the woken up task will run on
1295 * this cpu. (in this case the 'I will reschedule' promise of
1296 * the waker guarantees that the freshly woken up task is going
1297 * to be considered on this CPU.)
1299 if (!sync || cpu != this_cpu) {
1300 if (TASK_PREEMPTS_CURR(p, rq))
1301 resched_task(rq->curr);
1303 success = 1;
1305 out_running:
1306 p->state = TASK_RUNNING;
1307 out:
1308 task_rq_unlock(rq, &flags);
1310 return success;
1313 int fastcall wake_up_process(task_t *p)
1315 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1316 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1319 EXPORT_SYMBOL(wake_up_process);
1321 int fastcall wake_up_state(task_t *p, unsigned int state)
1323 return try_to_wake_up(p, state, 0);
1327 * Perform scheduler related setup for a newly forked process p.
1328 * p is forked by current.
1330 void fastcall sched_fork(task_t *p, int clone_flags)
1332 int cpu = get_cpu();
1334 #ifdef CONFIG_SMP
1335 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1336 #endif
1337 set_task_cpu(p, cpu);
1340 * We mark the process as running here, but have not actually
1341 * inserted it onto the runqueue yet. This guarantees that
1342 * nobody will actually run it, and a signal or other external
1343 * event cannot wake it up and insert it on the runqueue either.
1345 p->state = TASK_RUNNING;
1346 INIT_LIST_HEAD(&p->run_list);
1347 p->array = NULL;
1348 #ifdef CONFIG_SCHEDSTATS
1349 memset(&p->sched_info, 0, sizeof(p->sched_info));
1350 #endif
1351 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1352 p->oncpu = 0;
1353 #endif
1354 #ifdef CONFIG_PREEMPT
1355 /* Want to start with kernel preemption disabled. */
1356 task_thread_info(p)->preempt_count = 1;
1357 #endif
1359 * Share the timeslice between parent and child, thus the
1360 * total amount of pending timeslices in the system doesn't change,
1361 * resulting in more scheduling fairness.
1363 local_irq_disable();
1364 p->time_slice = (current->time_slice + 1) >> 1;
1366 * The remainder of the first timeslice might be recovered by
1367 * the parent if the child exits early enough.
1369 p->first_time_slice = 1;
1370 current->time_slice >>= 1;
1371 p->timestamp = sched_clock();
1372 if (unlikely(!current->time_slice)) {
1374 * This case is rare, it happens when the parent has only
1375 * a single jiffy left from its timeslice. Taking the
1376 * runqueue lock is not a problem.
1378 current->time_slice = 1;
1379 scheduler_tick();
1381 local_irq_enable();
1382 put_cpu();
1386 * wake_up_new_task - wake up a newly created task for the first time.
1388 * This function will do some initial scheduler statistics housekeeping
1389 * that must be done for every newly created context, then puts the task
1390 * on the runqueue and wakes it.
1392 void fastcall wake_up_new_task(task_t *p, unsigned long clone_flags)
1394 unsigned long flags;
1395 int this_cpu, cpu;
1396 runqueue_t *rq, *this_rq;
1398 rq = task_rq_lock(p, &flags);
1399 BUG_ON(p->state != TASK_RUNNING);
1400 this_cpu = smp_processor_id();
1401 cpu = task_cpu(p);
1404 * We decrease the sleep average of forking parents
1405 * and children as well, to keep max-interactive tasks
1406 * from forking tasks that are max-interactive. The parent
1407 * (current) is done further down, under its lock.
1409 p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1410 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1412 p->prio = effective_prio(p);
1414 if (likely(cpu == this_cpu)) {
1415 if (!(clone_flags & CLONE_VM)) {
1417 * The VM isn't cloned, so we're in a good position to
1418 * do child-runs-first in anticipation of an exec. This
1419 * usually avoids a lot of COW overhead.
1421 if (unlikely(!current->array))
1422 __activate_task(p, rq);
1423 else {
1424 p->prio = current->prio;
1425 list_add_tail(&p->run_list, &current->run_list);
1426 p->array = current->array;
1427 p->array->nr_active++;
1428 rq->nr_running++;
1430 set_need_resched();
1431 } else
1432 /* Run child last */
1433 __activate_task(p, rq);
1435 * We skip the following code due to cpu == this_cpu
1437 * task_rq_unlock(rq, &flags);
1438 * this_rq = task_rq_lock(current, &flags);
1440 this_rq = rq;
1441 } else {
1442 this_rq = cpu_rq(this_cpu);
1445 * Not the local CPU - must adjust timestamp. This should
1446 * get optimised away in the !CONFIG_SMP case.
1448 p->timestamp = (p->timestamp - this_rq->timestamp_last_tick)
1449 + rq->timestamp_last_tick;
1450 __activate_task(p, rq);
1451 if (TASK_PREEMPTS_CURR(p, rq))
1452 resched_task(rq->curr);
1455 * Parent and child are on different CPUs, now get the
1456 * parent runqueue to update the parent's ->sleep_avg:
1458 task_rq_unlock(rq, &flags);
1459 this_rq = task_rq_lock(current, &flags);
1461 current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1462 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1463 task_rq_unlock(this_rq, &flags);
1467 * Potentially available exiting-child timeslices are
1468 * retrieved here - this way the parent does not get
1469 * penalized for creating too many threads.
1471 * (this cannot be used to 'generate' timeslices
1472 * artificially, because any timeslice recovered here
1473 * was given away by the parent in the first place.)
1475 void fastcall sched_exit(task_t *p)
1477 unsigned long flags;
1478 runqueue_t *rq;
1481 * If the child was a (relative-) CPU hog then decrease
1482 * the sleep_avg of the parent as well.
1484 rq = task_rq_lock(p->parent, &flags);
1485 if (p->first_time_slice && task_cpu(p) == task_cpu(p->parent)) {
1486 p->parent->time_slice += p->time_slice;
1487 if (unlikely(p->parent->time_slice > task_timeslice(p)))
1488 p->parent->time_slice = task_timeslice(p);
1490 if (p->sleep_avg < p->parent->sleep_avg)
1491 p->parent->sleep_avg = p->parent->sleep_avg /
1492 (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg /
1493 (EXIT_WEIGHT + 1);
1494 task_rq_unlock(rq, &flags);
1498 * prepare_task_switch - prepare to switch tasks
1499 * @rq: the runqueue preparing to switch
1500 * @next: the task we are going to switch to.
1502 * This is called with the rq lock held and interrupts off. It must
1503 * be paired with a subsequent finish_task_switch after the context
1504 * switch.
1506 * prepare_task_switch sets up locking and calls architecture specific
1507 * hooks.
1509 static inline void prepare_task_switch(runqueue_t *rq, task_t *next)
1511 prepare_lock_switch(rq, next);
1512 prepare_arch_switch(next);
1516 * finish_task_switch - clean up after a task-switch
1517 * @rq: runqueue associated with task-switch
1518 * @prev: the thread we just switched away from.
1520 * finish_task_switch must be called after the context switch, paired
1521 * with a prepare_task_switch call before the context switch.
1522 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1523 * and do any other architecture-specific cleanup actions.
1525 * Note that we may have delayed dropping an mm in context_switch(). If
1526 * so, we finish that here outside of the runqueue lock. (Doing it
1527 * with the lock held can cause deadlocks; see schedule() for
1528 * details.)
1530 static inline void finish_task_switch(runqueue_t *rq, task_t *prev)
1531 __releases(rq->lock)
1533 struct mm_struct *mm = rq->prev_mm;
1534 unsigned long prev_task_flags;
1536 rq->prev_mm = NULL;
1539 * A task struct has one reference for the use as "current".
1540 * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1541 * calls schedule one last time. The schedule call will never return,
1542 * and the scheduled task must drop that reference.
1543 * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1544 * still held, otherwise prev could be scheduled on another cpu, die
1545 * there before we look at prev->state, and then the reference would
1546 * be dropped twice.
1547 * Manfred Spraul <manfred@colorfullife.com>
1549 prev_task_flags = prev->flags;
1550 finish_arch_switch(prev);
1551 finish_lock_switch(rq, prev);
1552 if (mm)
1553 mmdrop(mm);
1554 if (unlikely(prev_task_flags & PF_DEAD))
1555 put_task_struct(prev);
1559 * schedule_tail - first thing a freshly forked thread must call.
1560 * @prev: the thread we just switched away from.
1562 asmlinkage void schedule_tail(task_t *prev)
1563 __releases(rq->lock)
1565 runqueue_t *rq = this_rq();
1566 finish_task_switch(rq, prev);
1567 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1568 /* In this case, finish_task_switch does not reenable preemption */
1569 preempt_enable();
1570 #endif
1571 if (current->set_child_tid)
1572 put_user(current->pid, current->set_child_tid);
1576 * context_switch - switch to the new MM and the new
1577 * thread's register state.
1579 static inline
1580 task_t * context_switch(runqueue_t *rq, task_t *prev, task_t *next)
1582 struct mm_struct *mm = next->mm;
1583 struct mm_struct *oldmm = prev->active_mm;
1585 if (unlikely(!mm)) {
1586 next->active_mm = oldmm;
1587 atomic_inc(&oldmm->mm_count);
1588 enter_lazy_tlb(oldmm, next);
1589 } else
1590 switch_mm(oldmm, mm, next);
1592 if (unlikely(!prev->mm)) {
1593 prev->active_mm = NULL;
1594 WARN_ON(rq->prev_mm);
1595 rq->prev_mm = oldmm;
1598 /* Here we just switch the register state and the stack. */
1599 switch_to(prev, next, prev);
1601 return prev;
1605 * nr_running, nr_uninterruptible and nr_context_switches:
1607 * externally visible scheduler statistics: current number of runnable
1608 * threads, current number of uninterruptible-sleeping threads, total
1609 * number of context switches performed since bootup.
1611 unsigned long nr_running(void)
1613 unsigned long i, sum = 0;
1615 for_each_online_cpu(i)
1616 sum += cpu_rq(i)->nr_running;
1618 return sum;
1621 unsigned long nr_uninterruptible(void)
1623 unsigned long i, sum = 0;
1625 for_each_cpu(i)
1626 sum += cpu_rq(i)->nr_uninterruptible;
1629 * Since we read the counters lockless, it might be slightly
1630 * inaccurate. Do not allow it to go below zero though:
1632 if (unlikely((long)sum < 0))
1633 sum = 0;
1635 return sum;
1638 unsigned long long nr_context_switches(void)
1640 unsigned long long i, sum = 0;
1642 for_each_cpu(i)
1643 sum += cpu_rq(i)->nr_switches;
1645 return sum;
1648 unsigned long nr_iowait(void)
1650 unsigned long i, sum = 0;
1652 for_each_cpu(i)
1653 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1655 return sum;
1658 #ifdef CONFIG_SMP
1661 * double_rq_lock - safely lock two runqueues
1663 * Note this does not disable interrupts like task_rq_lock,
1664 * you need to do so manually before calling.
1666 static void double_rq_lock(runqueue_t *rq1, runqueue_t *rq2)
1667 __acquires(rq1->lock)
1668 __acquires(rq2->lock)
1670 if (rq1 == rq2) {
1671 spin_lock(&rq1->lock);
1672 __acquire(rq2->lock); /* Fake it out ;) */
1673 } else {
1674 if (rq1 < rq2) {
1675 spin_lock(&rq1->lock);
1676 spin_lock(&rq2->lock);
1677 } else {
1678 spin_lock(&rq2->lock);
1679 spin_lock(&rq1->lock);
1685 * double_rq_unlock - safely unlock two runqueues
1687 * Note this does not restore interrupts like task_rq_unlock,
1688 * you need to do so manually after calling.
1690 static void double_rq_unlock(runqueue_t *rq1, runqueue_t *rq2)
1691 __releases(rq1->lock)
1692 __releases(rq2->lock)
1694 spin_unlock(&rq1->lock);
1695 if (rq1 != rq2)
1696 spin_unlock(&rq2->lock);
1697 else
1698 __release(rq2->lock);
1702 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1704 static void double_lock_balance(runqueue_t *this_rq, runqueue_t *busiest)
1705 __releases(this_rq->lock)
1706 __acquires(busiest->lock)
1707 __acquires(this_rq->lock)
1709 if (unlikely(!spin_trylock(&busiest->lock))) {
1710 if (busiest < this_rq) {
1711 spin_unlock(&this_rq->lock);
1712 spin_lock(&busiest->lock);
1713 spin_lock(&this_rq->lock);
1714 } else
1715 spin_lock(&busiest->lock);
1720 * If dest_cpu is allowed for this process, migrate the task to it.
1721 * This is accomplished by forcing the cpu_allowed mask to only
1722 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
1723 * the cpu_allowed mask is restored.
1725 static void sched_migrate_task(task_t *p, int dest_cpu)
1727 migration_req_t req;
1728 runqueue_t *rq;
1729 unsigned long flags;
1731 rq = task_rq_lock(p, &flags);
1732 if (!cpu_isset(dest_cpu, p->cpus_allowed)
1733 || unlikely(cpu_is_offline(dest_cpu)))
1734 goto out;
1736 /* force the process onto the specified CPU */
1737 if (migrate_task(p, dest_cpu, &req)) {
1738 /* Need to wait for migration thread (might exit: take ref). */
1739 struct task_struct *mt = rq->migration_thread;
1740 get_task_struct(mt);
1741 task_rq_unlock(rq, &flags);
1742 wake_up_process(mt);
1743 put_task_struct(mt);
1744 wait_for_completion(&req.done);
1745 return;
1747 out:
1748 task_rq_unlock(rq, &flags);
1752 * sched_exec - execve() is a valuable balancing opportunity, because at
1753 * this point the task has the smallest effective memory and cache footprint.
1755 void sched_exec(void)
1757 int new_cpu, this_cpu = get_cpu();
1758 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
1759 put_cpu();
1760 if (new_cpu != this_cpu)
1761 sched_migrate_task(current, new_cpu);
1765 * pull_task - move a task from a remote runqueue to the local runqueue.
1766 * Both runqueues must be locked.
1768 static
1769 void pull_task(runqueue_t *src_rq, prio_array_t *src_array, task_t *p,
1770 runqueue_t *this_rq, prio_array_t *this_array, int this_cpu)
1772 dequeue_task(p, src_array);
1773 src_rq->nr_running--;
1774 set_task_cpu(p, this_cpu);
1775 this_rq->nr_running++;
1776 enqueue_task(p, this_array);
1777 p->timestamp = (p->timestamp - src_rq->timestamp_last_tick)
1778 + this_rq->timestamp_last_tick;
1780 * Note that idle threads have a prio of MAX_PRIO, for this test
1781 * to be always true for them.
1783 if (TASK_PREEMPTS_CURR(p, this_rq))
1784 resched_task(this_rq->curr);
1788 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1790 static
1791 int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu,
1792 struct sched_domain *sd, enum idle_type idle,
1793 int *all_pinned)
1796 * We do not migrate tasks that are:
1797 * 1) running (obviously), or
1798 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1799 * 3) are cache-hot on their current CPU.
1801 if (!cpu_isset(this_cpu, p->cpus_allowed))
1802 return 0;
1803 *all_pinned = 0;
1805 if (task_running(rq, p))
1806 return 0;
1809 * Aggressive migration if:
1810 * 1) task is cache cold, or
1811 * 2) too many balance attempts have failed.
1814 if (sd->nr_balance_failed > sd->cache_nice_tries)
1815 return 1;
1817 if (task_hot(p, rq->timestamp_last_tick, sd))
1818 return 0;
1819 return 1;
1823 * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq,
1824 * as part of a balancing operation within "domain". Returns the number of
1825 * tasks moved.
1827 * Called with both runqueues locked.
1829 static int move_tasks(runqueue_t *this_rq, int this_cpu, runqueue_t *busiest,
1830 unsigned long max_nr_move, struct sched_domain *sd,
1831 enum idle_type idle, int *all_pinned)
1833 prio_array_t *array, *dst_array;
1834 struct list_head *head, *curr;
1835 int idx, pulled = 0, pinned = 0;
1836 task_t *tmp;
1838 if (max_nr_move == 0)
1839 goto out;
1841 pinned = 1;
1844 * We first consider expired tasks. Those will likely not be
1845 * executed in the near future, and they are most likely to
1846 * be cache-cold, thus switching CPUs has the least effect
1847 * on them.
1849 if (busiest->expired->nr_active) {
1850 array = busiest->expired;
1851 dst_array = this_rq->expired;
1852 } else {
1853 array = busiest->active;
1854 dst_array = this_rq->active;
1857 new_array:
1858 /* Start searching at priority 0: */
1859 idx = 0;
1860 skip_bitmap:
1861 if (!idx)
1862 idx = sched_find_first_bit(array->bitmap);
1863 else
1864 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
1865 if (idx >= MAX_PRIO) {
1866 if (array == busiest->expired && busiest->active->nr_active) {
1867 array = busiest->active;
1868 dst_array = this_rq->active;
1869 goto new_array;
1871 goto out;
1874 head = array->queue + idx;
1875 curr = head->prev;
1876 skip_queue:
1877 tmp = list_entry(curr, task_t, run_list);
1879 curr = curr->prev;
1881 if (!can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
1882 if (curr != head)
1883 goto skip_queue;
1884 idx++;
1885 goto skip_bitmap;
1888 #ifdef CONFIG_SCHEDSTATS
1889 if (task_hot(tmp, busiest->timestamp_last_tick, sd))
1890 schedstat_inc(sd, lb_hot_gained[idle]);
1891 #endif
1893 pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
1894 pulled++;
1896 /* We only want to steal up to the prescribed number of tasks. */
1897 if (pulled < max_nr_move) {
1898 if (curr != head)
1899 goto skip_queue;
1900 idx++;
1901 goto skip_bitmap;
1903 out:
1905 * Right now, this is the only place pull_task() is called,
1906 * so we can safely collect pull_task() stats here rather than
1907 * inside pull_task().
1909 schedstat_add(sd, lb_gained[idle], pulled);
1911 if (all_pinned)
1912 *all_pinned = pinned;
1913 return pulled;
1917 * find_busiest_group finds and returns the busiest CPU group within the
1918 * domain. It calculates and returns the number of tasks which should be
1919 * moved to restore balance via the imbalance parameter.
1921 static struct sched_group *
1922 find_busiest_group(struct sched_domain *sd, int this_cpu,
1923 unsigned long *imbalance, enum idle_type idle, int *sd_idle)
1925 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
1926 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
1927 unsigned long max_pull;
1928 int load_idx;
1930 max_load = this_load = total_load = total_pwr = 0;
1931 if (idle == NOT_IDLE)
1932 load_idx = sd->busy_idx;
1933 else if (idle == NEWLY_IDLE)
1934 load_idx = sd->newidle_idx;
1935 else
1936 load_idx = sd->idle_idx;
1938 do {
1939 unsigned long load;
1940 int local_group;
1941 int i;
1943 local_group = cpu_isset(this_cpu, group->cpumask);
1945 /* Tally up the load of all CPUs in the group */
1946 avg_load = 0;
1948 for_each_cpu_mask(i, group->cpumask) {
1949 if (*sd_idle && !idle_cpu(i))
1950 *sd_idle = 0;
1952 /* Bias balancing toward cpus of our domain */
1953 if (local_group)
1954 load = target_load(i, load_idx);
1955 else
1956 load = source_load(i, load_idx);
1958 avg_load += load;
1961 total_load += avg_load;
1962 total_pwr += group->cpu_power;
1964 /* Adjust by relative CPU power of the group */
1965 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1967 if (local_group) {
1968 this_load = avg_load;
1969 this = group;
1970 } else if (avg_load > max_load) {
1971 max_load = avg_load;
1972 busiest = group;
1974 group = group->next;
1975 } while (group != sd->groups);
1977 if (!busiest || this_load >= max_load || max_load <= SCHED_LOAD_SCALE)
1978 goto out_balanced;
1980 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
1982 if (this_load >= avg_load ||
1983 100*max_load <= sd->imbalance_pct*this_load)
1984 goto out_balanced;
1987 * We're trying to get all the cpus to the average_load, so we don't
1988 * want to push ourselves above the average load, nor do we wish to
1989 * reduce the max loaded cpu below the average load, as either of these
1990 * actions would just result in more rebalancing later, and ping-pong
1991 * tasks around. Thus we look for the minimum possible imbalance.
1992 * Negative imbalances (*we* are more loaded than anyone else) will
1993 * be counted as no imbalance for these purposes -- we can't fix that
1994 * by pulling tasks to us. Be careful of negative numbers as they'll
1995 * appear as very large values with unsigned longs.
1998 /* Don't want to pull so many tasks that a group would go idle */
1999 max_pull = min(max_load - avg_load, max_load - SCHED_LOAD_SCALE);
2001 /* How much load to actually move to equalise the imbalance */
2002 *imbalance = min(max_pull * busiest->cpu_power,
2003 (avg_load - this_load) * this->cpu_power)
2004 / SCHED_LOAD_SCALE;
2006 if (*imbalance < SCHED_LOAD_SCALE) {
2007 unsigned long pwr_now = 0, pwr_move = 0;
2008 unsigned long tmp;
2010 if (max_load - this_load >= SCHED_LOAD_SCALE*2) {
2011 *imbalance = 1;
2012 return busiest;
2016 * OK, we don't have enough imbalance to justify moving tasks,
2017 * however we may be able to increase total CPU power used by
2018 * moving them.
2021 pwr_now += busiest->cpu_power*min(SCHED_LOAD_SCALE, max_load);
2022 pwr_now += this->cpu_power*min(SCHED_LOAD_SCALE, this_load);
2023 pwr_now /= SCHED_LOAD_SCALE;
2025 /* Amount of load we'd subtract */
2026 tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/busiest->cpu_power;
2027 if (max_load > tmp)
2028 pwr_move += busiest->cpu_power*min(SCHED_LOAD_SCALE,
2029 max_load - tmp);
2031 /* Amount of load we'd add */
2032 if (max_load*busiest->cpu_power <
2033 SCHED_LOAD_SCALE*SCHED_LOAD_SCALE)
2034 tmp = max_load*busiest->cpu_power/this->cpu_power;
2035 else
2036 tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/this->cpu_power;
2037 pwr_move += this->cpu_power*min(SCHED_LOAD_SCALE, this_load + tmp);
2038 pwr_move /= SCHED_LOAD_SCALE;
2040 /* Move if we gain throughput */
2041 if (pwr_move <= pwr_now)
2042 goto out_balanced;
2044 *imbalance = 1;
2045 return busiest;
2048 /* Get rid of the scaling factor, rounding down as we divide */
2049 *imbalance = *imbalance / SCHED_LOAD_SCALE;
2050 return busiest;
2052 out_balanced:
2054 *imbalance = 0;
2055 return NULL;
2059 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2061 static runqueue_t *find_busiest_queue(struct sched_group *group,
2062 enum idle_type idle)
2064 unsigned long load, max_load = 0;
2065 runqueue_t *busiest = NULL;
2066 int i;
2068 for_each_cpu_mask(i, group->cpumask) {
2069 load = source_load(i, 0);
2071 if (load > max_load) {
2072 max_load = load;
2073 busiest = cpu_rq(i);
2077 return busiest;
2081 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2082 * so long as it is large enough.
2084 #define MAX_PINNED_INTERVAL 512
2087 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2088 * tasks if there is an imbalance.
2090 * Called with this_rq unlocked.
2092 static int load_balance(int this_cpu, runqueue_t *this_rq,
2093 struct sched_domain *sd, enum idle_type idle)
2095 struct sched_group *group;
2096 runqueue_t *busiest;
2097 unsigned long imbalance;
2098 int nr_moved, all_pinned = 0;
2099 int active_balance = 0;
2100 int sd_idle = 0;
2102 if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER)
2103 sd_idle = 1;
2105 schedstat_inc(sd, lb_cnt[idle]);
2107 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle);
2108 if (!group) {
2109 schedstat_inc(sd, lb_nobusyg[idle]);
2110 goto out_balanced;
2113 busiest = find_busiest_queue(group, idle);
2114 if (!busiest) {
2115 schedstat_inc(sd, lb_nobusyq[idle]);
2116 goto out_balanced;
2119 BUG_ON(busiest == this_rq);
2121 schedstat_add(sd, lb_imbalance[idle], imbalance);
2123 nr_moved = 0;
2124 if (busiest->nr_running > 1) {
2126 * Attempt to move tasks. If find_busiest_group has found
2127 * an imbalance but busiest->nr_running <= 1, the group is
2128 * still unbalanced. nr_moved simply stays zero, so it is
2129 * correctly treated as an imbalance.
2131 double_rq_lock(this_rq, busiest);
2132 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2133 imbalance, sd, idle, &all_pinned);
2134 double_rq_unlock(this_rq, busiest);
2136 /* All tasks on this runqueue were pinned by CPU affinity */
2137 if (unlikely(all_pinned))
2138 goto out_balanced;
2141 if (!nr_moved) {
2142 schedstat_inc(sd, lb_failed[idle]);
2143 sd->nr_balance_failed++;
2145 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2147 spin_lock(&busiest->lock);
2149 /* don't kick the migration_thread, if the curr
2150 * task on busiest cpu can't be moved to this_cpu
2152 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2153 spin_unlock(&busiest->lock);
2154 all_pinned = 1;
2155 goto out_one_pinned;
2158 if (!busiest->active_balance) {
2159 busiest->active_balance = 1;
2160 busiest->push_cpu = this_cpu;
2161 active_balance = 1;
2163 spin_unlock(&busiest->lock);
2164 if (active_balance)
2165 wake_up_process(busiest->migration_thread);
2168 * We've kicked active balancing, reset the failure
2169 * counter.
2171 sd->nr_balance_failed = sd->cache_nice_tries+1;
2173 } else
2174 sd->nr_balance_failed = 0;
2176 if (likely(!active_balance)) {
2177 /* We were unbalanced, so reset the balancing interval */
2178 sd->balance_interval = sd->min_interval;
2179 } else {
2181 * If we've begun active balancing, start to back off. This
2182 * case may not be covered by the all_pinned logic if there
2183 * is only 1 task on the busy runqueue (because we don't call
2184 * move_tasks).
2186 if (sd->balance_interval < sd->max_interval)
2187 sd->balance_interval *= 2;
2190 if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2191 return -1;
2192 return nr_moved;
2194 out_balanced:
2195 schedstat_inc(sd, lb_balanced[idle]);
2197 sd->nr_balance_failed = 0;
2199 out_one_pinned:
2200 /* tune up the balancing interval */
2201 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2202 (sd->balance_interval < sd->max_interval))
2203 sd->balance_interval *= 2;
2205 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2206 return -1;
2207 return 0;
2211 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2212 * tasks if there is an imbalance.
2214 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2215 * this_rq is locked.
2217 static int load_balance_newidle(int this_cpu, runqueue_t *this_rq,
2218 struct sched_domain *sd)
2220 struct sched_group *group;
2221 runqueue_t *busiest = NULL;
2222 unsigned long imbalance;
2223 int nr_moved = 0;
2224 int sd_idle = 0;
2226 if (sd->flags & SD_SHARE_CPUPOWER)
2227 sd_idle = 1;
2229 schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
2230 group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE, &sd_idle);
2231 if (!group) {
2232 schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]);
2233 goto out_balanced;
2236 busiest = find_busiest_queue(group, NEWLY_IDLE);
2237 if (!busiest) {
2238 schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]);
2239 goto out_balanced;
2242 BUG_ON(busiest == this_rq);
2244 schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
2246 nr_moved = 0;
2247 if (busiest->nr_running > 1) {
2248 /* Attempt to move tasks */
2249 double_lock_balance(this_rq, busiest);
2250 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2251 imbalance, sd, NEWLY_IDLE, NULL);
2252 spin_unlock(&busiest->lock);
2255 if (!nr_moved) {
2256 schedstat_inc(sd, lb_failed[NEWLY_IDLE]);
2257 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2258 return -1;
2259 } else
2260 sd->nr_balance_failed = 0;
2262 return nr_moved;
2264 out_balanced:
2265 schedstat_inc(sd, lb_balanced[NEWLY_IDLE]);
2266 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2267 return -1;
2268 sd->nr_balance_failed = 0;
2269 return 0;
2273 * idle_balance is called by schedule() if this_cpu is about to become
2274 * idle. Attempts to pull tasks from other CPUs.
2276 static void idle_balance(int this_cpu, runqueue_t *this_rq)
2278 struct sched_domain *sd;
2280 for_each_domain(this_cpu, sd) {
2281 if (sd->flags & SD_BALANCE_NEWIDLE) {
2282 if (load_balance_newidle(this_cpu, this_rq, sd)) {
2283 /* We've pulled tasks over so stop searching */
2284 break;
2291 * active_load_balance is run by migration threads. It pushes running tasks
2292 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2293 * running on each physical CPU where possible, and avoids physical /
2294 * logical imbalances.
2296 * Called with busiest_rq locked.
2298 static void active_load_balance(runqueue_t *busiest_rq, int busiest_cpu)
2300 struct sched_domain *sd;
2301 runqueue_t *target_rq;
2302 int target_cpu = busiest_rq->push_cpu;
2304 if (busiest_rq->nr_running <= 1)
2305 /* no task to move */
2306 return;
2308 target_rq = cpu_rq(target_cpu);
2311 * This condition is "impossible", if it occurs
2312 * we need to fix it. Originally reported by
2313 * Bjorn Helgaas on a 128-cpu setup.
2315 BUG_ON(busiest_rq == target_rq);
2317 /* move a task from busiest_rq to target_rq */
2318 double_lock_balance(busiest_rq, target_rq);
2320 /* Search for an sd spanning us and the target CPU. */
2321 for_each_domain(target_cpu, sd)
2322 if ((sd->flags & SD_LOAD_BALANCE) &&
2323 cpu_isset(busiest_cpu, sd->span))
2324 break;
2326 if (unlikely(sd == NULL))
2327 goto out;
2329 schedstat_inc(sd, alb_cnt);
2331 if (move_tasks(target_rq, target_cpu, busiest_rq, 1, sd, SCHED_IDLE, NULL))
2332 schedstat_inc(sd, alb_pushed);
2333 else
2334 schedstat_inc(sd, alb_failed);
2335 out:
2336 spin_unlock(&target_rq->lock);
2340 * rebalance_tick will get called every timer tick, on every CPU.
2342 * It checks each scheduling domain to see if it is due to be balanced,
2343 * and initiates a balancing operation if so.
2345 * Balancing parameters are set up in arch_init_sched_domains.
2348 /* Don't have all balancing operations going off at once */
2349 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2351 static void rebalance_tick(int this_cpu, runqueue_t *this_rq,
2352 enum idle_type idle)
2354 unsigned long old_load, this_load;
2355 unsigned long j = jiffies + CPU_OFFSET(this_cpu);
2356 struct sched_domain *sd;
2357 int i;
2359 this_load = this_rq->nr_running * SCHED_LOAD_SCALE;
2360 /* Update our load */
2361 for (i = 0; i < 3; i++) {
2362 unsigned long new_load = this_load;
2363 int scale = 1 << i;
2364 old_load = this_rq->cpu_load[i];
2366 * Round up the averaging division if load is increasing. This
2367 * prevents us from getting stuck on 9 if the load is 10, for
2368 * example.
2370 if (new_load > old_load)
2371 new_load += scale-1;
2372 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale;
2375 for_each_domain(this_cpu, sd) {
2376 unsigned long interval;
2378 if (!(sd->flags & SD_LOAD_BALANCE))
2379 continue;
2381 interval = sd->balance_interval;
2382 if (idle != SCHED_IDLE)
2383 interval *= sd->busy_factor;
2385 /* scale ms to jiffies */
2386 interval = msecs_to_jiffies(interval);
2387 if (unlikely(!interval))
2388 interval = 1;
2390 if (j - sd->last_balance >= interval) {
2391 if (load_balance(this_cpu, this_rq, sd, idle)) {
2393 * We've pulled tasks over so either we're no
2394 * longer idle, or one of our SMT siblings is
2395 * not idle.
2397 idle = NOT_IDLE;
2399 sd->last_balance += interval;
2403 #else
2405 * on UP we do not need to balance between CPUs:
2407 static inline void rebalance_tick(int cpu, runqueue_t *rq, enum idle_type idle)
2410 static inline void idle_balance(int cpu, runqueue_t *rq)
2413 #endif
2415 static inline int wake_priority_sleeper(runqueue_t *rq)
2417 int ret = 0;
2418 #ifdef CONFIG_SCHED_SMT
2419 spin_lock(&rq->lock);
2421 * If an SMT sibling task has been put to sleep for priority
2422 * reasons reschedule the idle task to see if it can now run.
2424 if (rq->nr_running) {
2425 resched_task(rq->idle);
2426 ret = 1;
2428 spin_unlock(&rq->lock);
2429 #endif
2430 return ret;
2433 DEFINE_PER_CPU(struct kernel_stat, kstat);
2435 EXPORT_PER_CPU_SYMBOL(kstat);
2438 * This is called on clock ticks and on context switches.
2439 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2441 static inline void update_cpu_clock(task_t *p, runqueue_t *rq,
2442 unsigned long long now)
2444 unsigned long long last = max(p->timestamp, rq->timestamp_last_tick);
2445 p->sched_time += now - last;
2449 * Return current->sched_time plus any more ns on the sched_clock
2450 * that have not yet been banked.
2452 unsigned long long current_sched_time(const task_t *tsk)
2454 unsigned long long ns;
2455 unsigned long flags;
2456 local_irq_save(flags);
2457 ns = max(tsk->timestamp, task_rq(tsk)->timestamp_last_tick);
2458 ns = tsk->sched_time + (sched_clock() - ns);
2459 local_irq_restore(flags);
2460 return ns;
2464 * We place interactive tasks back into the active array, if possible.
2466 * To guarantee that this does not starve expired tasks we ignore the
2467 * interactivity of a task if the first expired task had to wait more
2468 * than a 'reasonable' amount of time. This deadline timeout is
2469 * load-dependent, as the frequency of array switched decreases with
2470 * increasing number of running tasks. We also ignore the interactivity
2471 * if a better static_prio task has expired:
2473 #define EXPIRED_STARVING(rq) \
2474 ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2475 (jiffies - (rq)->expired_timestamp >= \
2476 STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2477 ((rq)->curr->static_prio > (rq)->best_expired_prio))
2480 * Account user cpu time to a process.
2481 * @p: the process that the cpu time gets accounted to
2482 * @hardirq_offset: the offset to subtract from hardirq_count()
2483 * @cputime: the cpu time spent in user space since the last update
2485 void account_user_time(struct task_struct *p, cputime_t cputime)
2487 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2488 cputime64_t tmp;
2490 p->utime = cputime_add(p->utime, cputime);
2492 /* Add user time to cpustat. */
2493 tmp = cputime_to_cputime64(cputime);
2494 if (TASK_NICE(p) > 0)
2495 cpustat->nice = cputime64_add(cpustat->nice, tmp);
2496 else
2497 cpustat->user = cputime64_add(cpustat->user, tmp);
2501 * Account system cpu time to a process.
2502 * @p: the process that the cpu time gets accounted to
2503 * @hardirq_offset: the offset to subtract from hardirq_count()
2504 * @cputime: the cpu time spent in kernel space since the last update
2506 void account_system_time(struct task_struct *p, int hardirq_offset,
2507 cputime_t cputime)
2509 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2510 runqueue_t *rq = this_rq();
2511 cputime64_t tmp;
2513 p->stime = cputime_add(p->stime, cputime);
2515 /* Add system time to cpustat. */
2516 tmp = cputime_to_cputime64(cputime);
2517 if (hardirq_count() - hardirq_offset)
2518 cpustat->irq = cputime64_add(cpustat->irq, tmp);
2519 else if (softirq_count())
2520 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
2521 else if (p != rq->idle)
2522 cpustat->system = cputime64_add(cpustat->system, tmp);
2523 else if (atomic_read(&rq->nr_iowait) > 0)
2524 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2525 else
2526 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2527 /* Account for system time used */
2528 acct_update_integrals(p);
2532 * Account for involuntary wait time.
2533 * @p: the process from which the cpu time has been stolen
2534 * @steal: the cpu time spent in involuntary wait
2536 void account_steal_time(struct task_struct *p, cputime_t steal)
2538 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2539 cputime64_t tmp = cputime_to_cputime64(steal);
2540 runqueue_t *rq = this_rq();
2542 if (p == rq->idle) {
2543 p->stime = cputime_add(p->stime, steal);
2544 if (atomic_read(&rq->nr_iowait) > 0)
2545 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2546 else
2547 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2548 } else
2549 cpustat->steal = cputime64_add(cpustat->steal, tmp);
2553 * This function gets called by the timer code, with HZ frequency.
2554 * We call it with interrupts disabled.
2556 * It also gets called by the fork code, when changing the parent's
2557 * timeslices.
2559 void scheduler_tick(void)
2561 int cpu = smp_processor_id();
2562 runqueue_t *rq = this_rq();
2563 task_t *p = current;
2564 unsigned long long now = sched_clock();
2566 update_cpu_clock(p, rq, now);
2568 rq->timestamp_last_tick = now;
2570 if (p == rq->idle) {
2571 if (wake_priority_sleeper(rq))
2572 goto out;
2573 rebalance_tick(cpu, rq, SCHED_IDLE);
2574 return;
2577 /* Task might have expired already, but not scheduled off yet */
2578 if (p->array != rq->active) {
2579 set_tsk_need_resched(p);
2580 goto out;
2582 spin_lock(&rq->lock);
2584 * The task was running during this tick - update the
2585 * time slice counter. Note: we do not update a thread's
2586 * priority until it either goes to sleep or uses up its
2587 * timeslice. This makes it possible for interactive tasks
2588 * to use up their timeslices at their highest priority levels.
2590 if (rt_task(p)) {
2592 * RR tasks need a special form of timeslice management.
2593 * FIFO tasks have no timeslices.
2595 if ((p->policy == SCHED_RR) && !--p->time_slice) {
2596 p->time_slice = task_timeslice(p);
2597 p->first_time_slice = 0;
2598 set_tsk_need_resched(p);
2600 /* put it at the end of the queue: */
2601 requeue_task(p, rq->active);
2603 goto out_unlock;
2605 if (!--p->time_slice) {
2606 dequeue_task(p, rq->active);
2607 set_tsk_need_resched(p);
2608 p->prio = effective_prio(p);
2609 p->time_slice = task_timeslice(p);
2610 p->first_time_slice = 0;
2612 if (!rq->expired_timestamp)
2613 rq->expired_timestamp = jiffies;
2614 if (!TASK_INTERACTIVE(p) || EXPIRED_STARVING(rq)) {
2615 enqueue_task(p, rq->expired);
2616 if (p->static_prio < rq->best_expired_prio)
2617 rq->best_expired_prio = p->static_prio;
2618 } else
2619 enqueue_task(p, rq->active);
2620 } else {
2622 * Prevent a too long timeslice allowing a task to monopolize
2623 * the CPU. We do this by splitting up the timeslice into
2624 * smaller pieces.
2626 * Note: this does not mean the task's timeslices expire or
2627 * get lost in any way, they just might be preempted by
2628 * another task of equal priority. (one with higher
2629 * priority would have preempted this task already.) We
2630 * requeue this task to the end of the list on this priority
2631 * level, which is in essence a round-robin of tasks with
2632 * equal priority.
2634 * This only applies to tasks in the interactive
2635 * delta range with at least TIMESLICE_GRANULARITY to requeue.
2637 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
2638 p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
2639 (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
2640 (p->array == rq->active)) {
2642 requeue_task(p, rq->active);
2643 set_tsk_need_resched(p);
2646 out_unlock:
2647 spin_unlock(&rq->lock);
2648 out:
2649 rebalance_tick(cpu, rq, NOT_IDLE);
2652 #ifdef CONFIG_SCHED_SMT
2653 static inline void wakeup_busy_runqueue(runqueue_t *rq)
2655 /* If an SMT runqueue is sleeping due to priority reasons wake it up */
2656 if (rq->curr == rq->idle && rq->nr_running)
2657 resched_task(rq->idle);
2660 static void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2662 struct sched_domain *tmp, *sd = NULL;
2663 cpumask_t sibling_map;
2664 int i;
2666 for_each_domain(this_cpu, tmp)
2667 if (tmp->flags & SD_SHARE_CPUPOWER)
2668 sd = tmp;
2670 if (!sd)
2671 return;
2674 * Unlock the current runqueue because we have to lock in
2675 * CPU order to avoid deadlocks. Caller knows that we might
2676 * unlock. We keep IRQs disabled.
2678 spin_unlock(&this_rq->lock);
2680 sibling_map = sd->span;
2682 for_each_cpu_mask(i, sibling_map)
2683 spin_lock(&cpu_rq(i)->lock);
2685 * We clear this CPU from the mask. This both simplifies the
2686 * inner loop and keps this_rq locked when we exit:
2688 cpu_clear(this_cpu, sibling_map);
2690 for_each_cpu_mask(i, sibling_map) {
2691 runqueue_t *smt_rq = cpu_rq(i);
2693 wakeup_busy_runqueue(smt_rq);
2696 for_each_cpu_mask(i, sibling_map)
2697 spin_unlock(&cpu_rq(i)->lock);
2699 * We exit with this_cpu's rq still held and IRQs
2700 * still disabled:
2705 * number of 'lost' timeslices this task wont be able to fully
2706 * utilize, if another task runs on a sibling. This models the
2707 * slowdown effect of other tasks running on siblings:
2709 static inline unsigned long smt_slice(task_t *p, struct sched_domain *sd)
2711 return p->time_slice * (100 - sd->per_cpu_gain) / 100;
2714 static int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2716 struct sched_domain *tmp, *sd = NULL;
2717 cpumask_t sibling_map;
2718 prio_array_t *array;
2719 int ret = 0, i;
2720 task_t *p;
2722 for_each_domain(this_cpu, tmp)
2723 if (tmp->flags & SD_SHARE_CPUPOWER)
2724 sd = tmp;
2726 if (!sd)
2727 return 0;
2730 * The same locking rules and details apply as for
2731 * wake_sleeping_dependent():
2733 spin_unlock(&this_rq->lock);
2734 sibling_map = sd->span;
2735 for_each_cpu_mask(i, sibling_map)
2736 spin_lock(&cpu_rq(i)->lock);
2737 cpu_clear(this_cpu, sibling_map);
2740 * Establish next task to be run - it might have gone away because
2741 * we released the runqueue lock above:
2743 if (!this_rq->nr_running)
2744 goto out_unlock;
2745 array = this_rq->active;
2746 if (!array->nr_active)
2747 array = this_rq->expired;
2748 BUG_ON(!array->nr_active);
2750 p = list_entry(array->queue[sched_find_first_bit(array->bitmap)].next,
2751 task_t, run_list);
2753 for_each_cpu_mask(i, sibling_map) {
2754 runqueue_t *smt_rq = cpu_rq(i);
2755 task_t *smt_curr = smt_rq->curr;
2757 /* Kernel threads do not participate in dependent sleeping */
2758 if (!p->mm || !smt_curr->mm || rt_task(p))
2759 goto check_smt_task;
2762 * If a user task with lower static priority than the
2763 * running task on the SMT sibling is trying to schedule,
2764 * delay it till there is proportionately less timeslice
2765 * left of the sibling task to prevent a lower priority
2766 * task from using an unfair proportion of the
2767 * physical cpu's resources. -ck
2769 if (rt_task(smt_curr)) {
2771 * With real time tasks we run non-rt tasks only
2772 * per_cpu_gain% of the time.
2774 if ((jiffies % DEF_TIMESLICE) >
2775 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
2776 ret = 1;
2777 } else
2778 if (smt_curr->static_prio < p->static_prio &&
2779 !TASK_PREEMPTS_CURR(p, smt_rq) &&
2780 smt_slice(smt_curr, sd) > task_timeslice(p))
2781 ret = 1;
2783 check_smt_task:
2784 if ((!smt_curr->mm && smt_curr != smt_rq->idle) ||
2785 rt_task(smt_curr))
2786 continue;
2787 if (!p->mm) {
2788 wakeup_busy_runqueue(smt_rq);
2789 continue;
2793 * Reschedule a lower priority task on the SMT sibling for
2794 * it to be put to sleep, or wake it up if it has been put to
2795 * sleep for priority reasons to see if it should run now.
2797 if (rt_task(p)) {
2798 if ((jiffies % DEF_TIMESLICE) >
2799 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
2800 resched_task(smt_curr);
2801 } else {
2802 if (TASK_PREEMPTS_CURR(p, smt_rq) &&
2803 smt_slice(p, sd) > task_timeslice(smt_curr))
2804 resched_task(smt_curr);
2805 else
2806 wakeup_busy_runqueue(smt_rq);
2809 out_unlock:
2810 for_each_cpu_mask(i, sibling_map)
2811 spin_unlock(&cpu_rq(i)->lock);
2812 return ret;
2814 #else
2815 static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2819 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2821 return 0;
2823 #endif
2825 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
2827 void fastcall add_preempt_count(int val)
2830 * Underflow?
2832 BUG_ON((preempt_count() < 0));
2833 preempt_count() += val;
2835 * Spinlock count overflowing soon?
2837 BUG_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10);
2839 EXPORT_SYMBOL(add_preempt_count);
2841 void fastcall sub_preempt_count(int val)
2844 * Underflow?
2846 BUG_ON(val > preempt_count());
2848 * Is the spinlock portion underflowing?
2850 BUG_ON((val < PREEMPT_MASK) && !(preempt_count() & PREEMPT_MASK));
2851 preempt_count() -= val;
2853 EXPORT_SYMBOL(sub_preempt_count);
2855 #endif
2858 * schedule() is the main scheduler function.
2860 asmlinkage void __sched schedule(void)
2862 long *switch_count;
2863 task_t *prev, *next;
2864 runqueue_t *rq;
2865 prio_array_t *array;
2866 struct list_head *queue;
2867 unsigned long long now;
2868 unsigned long run_time;
2869 int cpu, idx, new_prio;
2872 * Test if we are atomic. Since do_exit() needs to call into
2873 * schedule() atomically, we ignore that path for now.
2874 * Otherwise, whine if we are scheduling when we should not be.
2876 if (likely(!current->exit_state)) {
2877 if (unlikely(in_atomic())) {
2878 printk(KERN_ERR "scheduling while atomic: "
2879 "%s/0x%08x/%d\n",
2880 current->comm, preempt_count(), current->pid);
2881 dump_stack();
2884 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2886 need_resched:
2887 preempt_disable();
2888 prev = current;
2889 release_kernel_lock(prev);
2890 need_resched_nonpreemptible:
2891 rq = this_rq();
2894 * The idle thread is not allowed to schedule!
2895 * Remove this check after it has been exercised a bit.
2897 if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
2898 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
2899 dump_stack();
2902 schedstat_inc(rq, sched_cnt);
2903 now = sched_clock();
2904 if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
2905 run_time = now - prev->timestamp;
2906 if (unlikely((long long)(now - prev->timestamp) < 0))
2907 run_time = 0;
2908 } else
2909 run_time = NS_MAX_SLEEP_AVG;
2912 * Tasks charged proportionately less run_time at high sleep_avg to
2913 * delay them losing their interactive status
2915 run_time /= (CURRENT_BONUS(prev) ? : 1);
2917 spin_lock_irq(&rq->lock);
2919 if (unlikely(prev->flags & PF_DEAD))
2920 prev->state = EXIT_DEAD;
2922 switch_count = &prev->nivcsw;
2923 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2924 switch_count = &prev->nvcsw;
2925 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
2926 unlikely(signal_pending(prev))))
2927 prev->state = TASK_RUNNING;
2928 else {
2929 if (prev->state == TASK_UNINTERRUPTIBLE)
2930 rq->nr_uninterruptible++;
2931 deactivate_task(prev, rq);
2935 cpu = smp_processor_id();
2936 if (unlikely(!rq->nr_running)) {
2937 go_idle:
2938 idle_balance(cpu, rq);
2939 if (!rq->nr_running) {
2940 next = rq->idle;
2941 rq->expired_timestamp = 0;
2942 wake_sleeping_dependent(cpu, rq);
2944 * wake_sleeping_dependent() might have released
2945 * the runqueue, so break out if we got new
2946 * tasks meanwhile:
2948 if (!rq->nr_running)
2949 goto switch_tasks;
2951 } else {
2952 if (dependent_sleeper(cpu, rq)) {
2953 next = rq->idle;
2954 goto switch_tasks;
2957 * dependent_sleeper() releases and reacquires the runqueue
2958 * lock, hence go into the idle loop if the rq went
2959 * empty meanwhile:
2961 if (unlikely(!rq->nr_running))
2962 goto go_idle;
2965 array = rq->active;
2966 if (unlikely(!array->nr_active)) {
2968 * Switch the active and expired arrays.
2970 schedstat_inc(rq, sched_switch);
2971 rq->active = rq->expired;
2972 rq->expired = array;
2973 array = rq->active;
2974 rq->expired_timestamp = 0;
2975 rq->best_expired_prio = MAX_PRIO;
2978 idx = sched_find_first_bit(array->bitmap);
2979 queue = array->queue + idx;
2980 next = list_entry(queue->next, task_t, run_list);
2982 if (!rt_task(next) && next->activated > 0) {
2983 unsigned long long delta = now - next->timestamp;
2984 if (unlikely((long long)(now - next->timestamp) < 0))
2985 delta = 0;
2987 if (next->activated == 1)
2988 delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
2990 array = next->array;
2991 new_prio = recalc_task_prio(next, next->timestamp + delta);
2993 if (unlikely(next->prio != new_prio)) {
2994 dequeue_task(next, array);
2995 next->prio = new_prio;
2996 enqueue_task(next, array);
2997 } else
2998 requeue_task(next, array);
3000 next->activated = 0;
3001 switch_tasks:
3002 if (next == rq->idle)
3003 schedstat_inc(rq, sched_goidle);
3004 prefetch(next);
3005 prefetch_stack(next);
3006 clear_tsk_need_resched(prev);
3007 rcu_qsctr_inc(task_cpu(prev));
3009 update_cpu_clock(prev, rq, now);
3011 prev->sleep_avg -= run_time;
3012 if ((long)prev->sleep_avg <= 0)
3013 prev->sleep_avg = 0;
3014 prev->timestamp = prev->last_ran = now;
3016 sched_info_switch(prev, next);
3017 if (likely(prev != next)) {
3018 next->timestamp = now;
3019 rq->nr_switches++;
3020 rq->curr = next;
3021 ++*switch_count;
3023 prepare_task_switch(rq, next);
3024 prev = context_switch(rq, prev, next);
3025 barrier();
3027 * this_rq must be evaluated again because prev may have moved
3028 * CPUs since it called schedule(), thus the 'rq' on its stack
3029 * frame will be invalid.
3031 finish_task_switch(this_rq(), prev);
3032 } else
3033 spin_unlock_irq(&rq->lock);
3035 prev = current;
3036 if (unlikely(reacquire_kernel_lock(prev) < 0))
3037 goto need_resched_nonpreemptible;
3038 preempt_enable_no_resched();
3039 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3040 goto need_resched;
3043 EXPORT_SYMBOL(schedule);
3045 #ifdef CONFIG_PREEMPT
3047 * this is is the entry point to schedule() from in-kernel preemption
3048 * off of preempt_enable. Kernel preemptions off return from interrupt
3049 * occur there and call schedule directly.
3051 asmlinkage void __sched preempt_schedule(void)
3053 struct thread_info *ti = current_thread_info();
3054 #ifdef CONFIG_PREEMPT_BKL
3055 struct task_struct *task = current;
3056 int saved_lock_depth;
3057 #endif
3059 * If there is a non-zero preempt_count or interrupts are disabled,
3060 * we do not want to preempt the current task. Just return..
3062 if (unlikely(ti->preempt_count || irqs_disabled()))
3063 return;
3065 need_resched:
3066 add_preempt_count(PREEMPT_ACTIVE);
3068 * We keep the big kernel semaphore locked, but we
3069 * clear ->lock_depth so that schedule() doesnt
3070 * auto-release the semaphore:
3072 #ifdef CONFIG_PREEMPT_BKL
3073 saved_lock_depth = task->lock_depth;
3074 task->lock_depth = -1;
3075 #endif
3076 schedule();
3077 #ifdef CONFIG_PREEMPT_BKL
3078 task->lock_depth = saved_lock_depth;
3079 #endif
3080 sub_preempt_count(PREEMPT_ACTIVE);
3082 /* we could miss a preemption opportunity between schedule and now */
3083 barrier();
3084 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3085 goto need_resched;
3088 EXPORT_SYMBOL(preempt_schedule);
3091 * this is is the entry point to schedule() from kernel preemption
3092 * off of irq context.
3093 * Note, that this is called and return with irqs disabled. This will
3094 * protect us against recursive calling from irq.
3096 asmlinkage void __sched preempt_schedule_irq(void)
3098 struct thread_info *ti = current_thread_info();
3099 #ifdef CONFIG_PREEMPT_BKL
3100 struct task_struct *task = current;
3101 int saved_lock_depth;
3102 #endif
3103 /* Catch callers which need to be fixed*/
3104 BUG_ON(ti->preempt_count || !irqs_disabled());
3106 need_resched:
3107 add_preempt_count(PREEMPT_ACTIVE);
3109 * We keep the big kernel semaphore locked, but we
3110 * clear ->lock_depth so that schedule() doesnt
3111 * auto-release the semaphore:
3113 #ifdef CONFIG_PREEMPT_BKL
3114 saved_lock_depth = task->lock_depth;
3115 task->lock_depth = -1;
3116 #endif
3117 local_irq_enable();
3118 schedule();
3119 local_irq_disable();
3120 #ifdef CONFIG_PREEMPT_BKL
3121 task->lock_depth = saved_lock_depth;
3122 #endif
3123 sub_preempt_count(PREEMPT_ACTIVE);
3125 /* we could miss a preemption opportunity between schedule and now */
3126 barrier();
3127 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3128 goto need_resched;
3131 #endif /* CONFIG_PREEMPT */
3133 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3134 void *key)
3136 task_t *p = curr->private;
3137 return try_to_wake_up(p, mode, sync);
3140 EXPORT_SYMBOL(default_wake_function);
3143 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3144 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3145 * number) then we wake all the non-exclusive tasks and one exclusive task.
3147 * There are circumstances in which we can try to wake a task which has already
3148 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3149 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3151 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3152 int nr_exclusive, int sync, void *key)
3154 struct list_head *tmp, *next;
3156 list_for_each_safe(tmp, next, &q->task_list) {
3157 wait_queue_t *curr;
3158 unsigned flags;
3159 curr = list_entry(tmp, wait_queue_t, task_list);
3160 flags = curr->flags;
3161 if (curr->func(curr, mode, sync, key) &&
3162 (flags & WQ_FLAG_EXCLUSIVE) &&
3163 !--nr_exclusive)
3164 break;
3169 * __wake_up - wake up threads blocked on a waitqueue.
3170 * @q: the waitqueue
3171 * @mode: which threads
3172 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3173 * @key: is directly passed to the wakeup function
3175 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3176 int nr_exclusive, void *key)
3178 unsigned long flags;
3180 spin_lock_irqsave(&q->lock, flags);
3181 __wake_up_common(q, mode, nr_exclusive, 0, key);
3182 spin_unlock_irqrestore(&q->lock, flags);
3185 EXPORT_SYMBOL(__wake_up);
3188 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3190 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3192 __wake_up_common(q, mode, 1, 0, NULL);
3196 * __wake_up_sync - wake up threads blocked on a waitqueue.
3197 * @q: the waitqueue
3198 * @mode: which threads
3199 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3201 * The sync wakeup differs that the waker knows that it will schedule
3202 * away soon, so while the target thread will be woken up, it will not
3203 * be migrated to another CPU - ie. the two threads are 'synchronized'
3204 * with each other. This can prevent needless bouncing between CPUs.
3206 * On UP it can prevent extra preemption.
3208 void fastcall
3209 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3211 unsigned long flags;
3212 int sync = 1;
3214 if (unlikely(!q))
3215 return;
3217 if (unlikely(!nr_exclusive))
3218 sync = 0;
3220 spin_lock_irqsave(&q->lock, flags);
3221 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3222 spin_unlock_irqrestore(&q->lock, flags);
3224 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3226 void fastcall complete(struct completion *x)
3228 unsigned long flags;
3230 spin_lock_irqsave(&x->wait.lock, flags);
3231 x->done++;
3232 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3233 1, 0, NULL);
3234 spin_unlock_irqrestore(&x->wait.lock, flags);
3236 EXPORT_SYMBOL(complete);
3238 void fastcall complete_all(struct completion *x)
3240 unsigned long flags;
3242 spin_lock_irqsave(&x->wait.lock, flags);
3243 x->done += UINT_MAX/2;
3244 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3245 0, 0, NULL);
3246 spin_unlock_irqrestore(&x->wait.lock, flags);
3248 EXPORT_SYMBOL(complete_all);
3250 void fastcall __sched wait_for_completion(struct completion *x)
3252 might_sleep();
3253 spin_lock_irq(&x->wait.lock);
3254 if (!x->done) {
3255 DECLARE_WAITQUEUE(wait, current);
3257 wait.flags |= WQ_FLAG_EXCLUSIVE;
3258 __add_wait_queue_tail(&x->wait, &wait);
3259 do {
3260 __set_current_state(TASK_UNINTERRUPTIBLE);
3261 spin_unlock_irq(&x->wait.lock);
3262 schedule();
3263 spin_lock_irq(&x->wait.lock);
3264 } while (!x->done);
3265 __remove_wait_queue(&x->wait, &wait);
3267 x->done--;
3268 spin_unlock_irq(&x->wait.lock);
3270 EXPORT_SYMBOL(wait_for_completion);
3272 unsigned long fastcall __sched
3273 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3275 might_sleep();
3277 spin_lock_irq(&x->wait.lock);
3278 if (!x->done) {
3279 DECLARE_WAITQUEUE(wait, current);
3281 wait.flags |= WQ_FLAG_EXCLUSIVE;
3282 __add_wait_queue_tail(&x->wait, &wait);
3283 do {
3284 __set_current_state(TASK_UNINTERRUPTIBLE);
3285 spin_unlock_irq(&x->wait.lock);
3286 timeout = schedule_timeout(timeout);
3287 spin_lock_irq(&x->wait.lock);
3288 if (!timeout) {
3289 __remove_wait_queue(&x->wait, &wait);
3290 goto out;
3292 } while (!x->done);
3293 __remove_wait_queue(&x->wait, &wait);
3295 x->done--;
3296 out:
3297 spin_unlock_irq(&x->wait.lock);
3298 return timeout;
3300 EXPORT_SYMBOL(wait_for_completion_timeout);
3302 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3304 int ret = 0;
3306 might_sleep();
3308 spin_lock_irq(&x->wait.lock);
3309 if (!x->done) {
3310 DECLARE_WAITQUEUE(wait, current);
3312 wait.flags |= WQ_FLAG_EXCLUSIVE;
3313 __add_wait_queue_tail(&x->wait, &wait);
3314 do {
3315 if (signal_pending(current)) {
3316 ret = -ERESTARTSYS;
3317 __remove_wait_queue(&x->wait, &wait);
3318 goto out;
3320 __set_current_state(TASK_INTERRUPTIBLE);
3321 spin_unlock_irq(&x->wait.lock);
3322 schedule();
3323 spin_lock_irq(&x->wait.lock);
3324 } while (!x->done);
3325 __remove_wait_queue(&x->wait, &wait);
3327 x->done--;
3328 out:
3329 spin_unlock_irq(&x->wait.lock);
3331 return ret;
3333 EXPORT_SYMBOL(wait_for_completion_interruptible);
3335 unsigned long fastcall __sched
3336 wait_for_completion_interruptible_timeout(struct completion *x,
3337 unsigned long timeout)
3339 might_sleep();
3341 spin_lock_irq(&x->wait.lock);
3342 if (!x->done) {
3343 DECLARE_WAITQUEUE(wait, current);
3345 wait.flags |= WQ_FLAG_EXCLUSIVE;
3346 __add_wait_queue_tail(&x->wait, &wait);
3347 do {
3348 if (signal_pending(current)) {
3349 timeout = -ERESTARTSYS;
3350 __remove_wait_queue(&x->wait, &wait);
3351 goto out;
3353 __set_current_state(TASK_INTERRUPTIBLE);
3354 spin_unlock_irq(&x->wait.lock);
3355 timeout = schedule_timeout(timeout);
3356 spin_lock_irq(&x->wait.lock);
3357 if (!timeout) {
3358 __remove_wait_queue(&x->wait, &wait);
3359 goto out;
3361 } while (!x->done);
3362 __remove_wait_queue(&x->wait, &wait);
3364 x->done--;
3365 out:
3366 spin_unlock_irq(&x->wait.lock);
3367 return timeout;
3369 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3372 #define SLEEP_ON_VAR \
3373 unsigned long flags; \
3374 wait_queue_t wait; \
3375 init_waitqueue_entry(&wait, current);
3377 #define SLEEP_ON_HEAD \
3378 spin_lock_irqsave(&q->lock,flags); \
3379 __add_wait_queue(q, &wait); \
3380 spin_unlock(&q->lock);
3382 #define SLEEP_ON_TAIL \
3383 spin_lock_irq(&q->lock); \
3384 __remove_wait_queue(q, &wait); \
3385 spin_unlock_irqrestore(&q->lock, flags);
3387 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
3389 SLEEP_ON_VAR
3391 current->state = TASK_INTERRUPTIBLE;
3393 SLEEP_ON_HEAD
3394 schedule();
3395 SLEEP_ON_TAIL
3398 EXPORT_SYMBOL(interruptible_sleep_on);
3400 long fastcall __sched
3401 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3403 SLEEP_ON_VAR
3405 current->state = TASK_INTERRUPTIBLE;
3407 SLEEP_ON_HEAD
3408 timeout = schedule_timeout(timeout);
3409 SLEEP_ON_TAIL
3411 return timeout;
3414 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3416 void fastcall __sched sleep_on(wait_queue_head_t *q)
3418 SLEEP_ON_VAR
3420 current->state = TASK_UNINTERRUPTIBLE;
3422 SLEEP_ON_HEAD
3423 schedule();
3424 SLEEP_ON_TAIL
3427 EXPORT_SYMBOL(sleep_on);
3429 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3431 SLEEP_ON_VAR
3433 current->state = TASK_UNINTERRUPTIBLE;
3435 SLEEP_ON_HEAD
3436 timeout = schedule_timeout(timeout);
3437 SLEEP_ON_TAIL
3439 return timeout;
3442 EXPORT_SYMBOL(sleep_on_timeout);
3444 void set_user_nice(task_t *p, long nice)
3446 unsigned long flags;
3447 prio_array_t *array;
3448 runqueue_t *rq;
3449 int old_prio, new_prio, delta;
3451 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3452 return;
3454 * We have to be careful, if called from sys_setpriority(),
3455 * the task might be in the middle of scheduling on another CPU.
3457 rq = task_rq_lock(p, &flags);
3459 * The RT priorities are set via sched_setscheduler(), but we still
3460 * allow the 'normal' nice value to be set - but as expected
3461 * it wont have any effect on scheduling until the task is
3462 * not SCHED_NORMAL/SCHED_BATCH:
3464 if (rt_task(p)) {
3465 p->static_prio = NICE_TO_PRIO(nice);
3466 goto out_unlock;
3468 array = p->array;
3469 if (array)
3470 dequeue_task(p, array);
3472 old_prio = p->prio;
3473 new_prio = NICE_TO_PRIO(nice);
3474 delta = new_prio - old_prio;
3475 p->static_prio = NICE_TO_PRIO(nice);
3476 p->prio += delta;
3478 if (array) {
3479 enqueue_task(p, array);
3481 * If the task increased its priority or is running and
3482 * lowered its priority, then reschedule its CPU:
3484 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3485 resched_task(rq->curr);
3487 out_unlock:
3488 task_rq_unlock(rq, &flags);
3491 EXPORT_SYMBOL(set_user_nice);
3494 * can_nice - check if a task can reduce its nice value
3495 * @p: task
3496 * @nice: nice value
3498 int can_nice(const task_t *p, const int nice)
3500 /* convert nice value [19,-20] to rlimit style value [1,40] */
3501 int nice_rlim = 20 - nice;
3502 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
3503 capable(CAP_SYS_NICE));
3506 #ifdef __ARCH_WANT_SYS_NICE
3509 * sys_nice - change the priority of the current process.
3510 * @increment: priority increment
3512 * sys_setpriority is a more generic, but much slower function that
3513 * does similar things.
3515 asmlinkage long sys_nice(int increment)
3517 int retval;
3518 long nice;
3521 * Setpriority might change our priority at the same moment.
3522 * We don't have to worry. Conceptually one call occurs first
3523 * and we have a single winner.
3525 if (increment < -40)
3526 increment = -40;
3527 if (increment > 40)
3528 increment = 40;
3530 nice = PRIO_TO_NICE(current->static_prio) + increment;
3531 if (nice < -20)
3532 nice = -20;
3533 if (nice > 19)
3534 nice = 19;
3536 if (increment < 0 && !can_nice(current, nice))
3537 return -EPERM;
3539 retval = security_task_setnice(current, nice);
3540 if (retval)
3541 return retval;
3543 set_user_nice(current, nice);
3544 return 0;
3547 #endif
3550 * task_prio - return the priority value of a given task.
3551 * @p: the task in question.
3553 * This is the priority value as seen by users in /proc.
3554 * RT tasks are offset by -200. Normal tasks are centered
3555 * around 0, value goes from -16 to +15.
3557 int task_prio(const task_t *p)
3559 return p->prio - MAX_RT_PRIO;
3563 * task_nice - return the nice value of a given task.
3564 * @p: the task in question.
3566 int task_nice(const task_t *p)
3568 return TASK_NICE(p);
3570 EXPORT_SYMBOL_GPL(task_nice);
3573 * idle_cpu - is a given cpu idle currently?
3574 * @cpu: the processor in question.
3576 int idle_cpu(int cpu)
3578 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
3582 * idle_task - return the idle task for a given cpu.
3583 * @cpu: the processor in question.
3585 task_t *idle_task(int cpu)
3587 return cpu_rq(cpu)->idle;
3591 * find_process_by_pid - find a process with a matching PID value.
3592 * @pid: the pid in question.
3594 static inline task_t *find_process_by_pid(pid_t pid)
3596 return pid ? find_task_by_pid(pid) : current;
3599 /* Actually do priority change: must hold rq lock. */
3600 static void __setscheduler(struct task_struct *p, int policy, int prio)
3602 BUG_ON(p->array);
3603 p->policy = policy;
3604 p->rt_priority = prio;
3605 if (policy != SCHED_NORMAL && policy != SCHED_BATCH) {
3606 p->prio = MAX_RT_PRIO-1 - p->rt_priority;
3607 } else {
3608 p->prio = p->static_prio;
3610 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
3612 if (policy == SCHED_BATCH)
3613 p->sleep_avg = 0;
3618 * sched_setscheduler - change the scheduling policy and/or RT priority of
3619 * a thread.
3620 * @p: the task in question.
3621 * @policy: new policy.
3622 * @param: structure containing the new RT priority.
3624 int sched_setscheduler(struct task_struct *p, int policy,
3625 struct sched_param *param)
3627 int retval;
3628 int oldprio, oldpolicy = -1;
3629 prio_array_t *array;
3630 unsigned long flags;
3631 runqueue_t *rq;
3633 recheck:
3634 /* double check policy once rq lock held */
3635 if (policy < 0)
3636 policy = oldpolicy = p->policy;
3637 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
3638 policy != SCHED_NORMAL && policy != SCHED_BATCH)
3639 return -EINVAL;
3641 * Valid priorities for SCHED_FIFO and SCHED_RR are
3642 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
3643 * SCHED_BATCH is 0.
3645 if (param->sched_priority < 0 ||
3646 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3647 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3648 return -EINVAL;
3649 if ((policy == SCHED_NORMAL || policy == SCHED_BATCH)
3650 != (param->sched_priority == 0))
3651 return -EINVAL;
3654 * Allow unprivileged RT tasks to decrease priority:
3656 if (!capable(CAP_SYS_NICE)) {
3658 * can't change policy, except between SCHED_NORMAL
3659 * and SCHED_BATCH:
3661 if (((policy != SCHED_NORMAL && p->policy != SCHED_BATCH) &&
3662 (policy != SCHED_BATCH && p->policy != SCHED_NORMAL)) &&
3663 !p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
3664 return -EPERM;
3665 /* can't increase priority */
3666 if ((policy != SCHED_NORMAL && policy != SCHED_BATCH) &&
3667 param->sched_priority > p->rt_priority &&
3668 param->sched_priority >
3669 p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
3670 return -EPERM;
3671 /* can't change other user's priorities */
3672 if ((current->euid != p->euid) &&
3673 (current->euid != p->uid))
3674 return -EPERM;
3677 retval = security_task_setscheduler(p, policy, param);
3678 if (retval)
3679 return retval;
3681 * To be able to change p->policy safely, the apropriate
3682 * runqueue lock must be held.
3684 rq = task_rq_lock(p, &flags);
3685 /* recheck policy now with rq lock held */
3686 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3687 policy = oldpolicy = -1;
3688 task_rq_unlock(rq, &flags);
3689 goto recheck;
3691 array = p->array;
3692 if (array)
3693 deactivate_task(p, rq);
3694 oldprio = p->prio;
3695 __setscheduler(p, policy, param->sched_priority);
3696 if (array) {
3697 __activate_task(p, rq);
3699 * Reschedule if we are currently running on this runqueue and
3700 * our priority decreased, or if we are not currently running on
3701 * this runqueue and our priority is higher than the current's
3703 if (task_running(rq, p)) {
3704 if (p->prio > oldprio)
3705 resched_task(rq->curr);
3706 } else if (TASK_PREEMPTS_CURR(p, rq))
3707 resched_task(rq->curr);
3709 task_rq_unlock(rq, &flags);
3710 return 0;
3712 EXPORT_SYMBOL_GPL(sched_setscheduler);
3714 static int
3715 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3717 int retval;
3718 struct sched_param lparam;
3719 struct task_struct *p;
3721 if (!param || pid < 0)
3722 return -EINVAL;
3723 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3724 return -EFAULT;
3725 read_lock_irq(&tasklist_lock);
3726 p = find_process_by_pid(pid);
3727 if (!p) {
3728 read_unlock_irq(&tasklist_lock);
3729 return -ESRCH;
3731 retval = sched_setscheduler(p, policy, &lparam);
3732 read_unlock_irq(&tasklist_lock);
3733 return retval;
3737 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3738 * @pid: the pid in question.
3739 * @policy: new policy.
3740 * @param: structure containing the new RT priority.
3742 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
3743 struct sched_param __user *param)
3745 /* negative values for policy are not valid */
3746 if (policy < 0)
3747 return -EINVAL;
3749 return do_sched_setscheduler(pid, policy, param);
3753 * sys_sched_setparam - set/change the RT priority of a thread
3754 * @pid: the pid in question.
3755 * @param: structure containing the new RT priority.
3757 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
3759 return do_sched_setscheduler(pid, -1, param);
3763 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3764 * @pid: the pid in question.
3766 asmlinkage long sys_sched_getscheduler(pid_t pid)
3768 int retval = -EINVAL;
3769 task_t *p;
3771 if (pid < 0)
3772 goto out_nounlock;
3774 retval = -ESRCH;
3775 read_lock(&tasklist_lock);
3776 p = find_process_by_pid(pid);
3777 if (p) {
3778 retval = security_task_getscheduler(p);
3779 if (!retval)
3780 retval = p->policy;
3782 read_unlock(&tasklist_lock);
3784 out_nounlock:
3785 return retval;
3789 * sys_sched_getscheduler - get the RT priority of a thread
3790 * @pid: the pid in question.
3791 * @param: structure containing the RT priority.
3793 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
3795 struct sched_param lp;
3796 int retval = -EINVAL;
3797 task_t *p;
3799 if (!param || pid < 0)
3800 goto out_nounlock;
3802 read_lock(&tasklist_lock);
3803 p = find_process_by_pid(pid);
3804 retval = -ESRCH;
3805 if (!p)
3806 goto out_unlock;
3808 retval = security_task_getscheduler(p);
3809 if (retval)
3810 goto out_unlock;
3812 lp.sched_priority = p->rt_priority;
3813 read_unlock(&tasklist_lock);
3816 * This one might sleep, we cannot do it with a spinlock held ...
3818 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3820 out_nounlock:
3821 return retval;
3823 out_unlock:
3824 read_unlock(&tasklist_lock);
3825 return retval;
3828 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
3830 task_t *p;
3831 int retval;
3832 cpumask_t cpus_allowed;
3834 lock_cpu_hotplug();
3835 read_lock(&tasklist_lock);
3837 p = find_process_by_pid(pid);
3838 if (!p) {
3839 read_unlock(&tasklist_lock);
3840 unlock_cpu_hotplug();
3841 return -ESRCH;
3845 * It is not safe to call set_cpus_allowed with the
3846 * tasklist_lock held. We will bump the task_struct's
3847 * usage count and then drop tasklist_lock.
3849 get_task_struct(p);
3850 read_unlock(&tasklist_lock);
3852 retval = -EPERM;
3853 if ((current->euid != p->euid) && (current->euid != p->uid) &&
3854 !capable(CAP_SYS_NICE))
3855 goto out_unlock;
3857 cpus_allowed = cpuset_cpus_allowed(p);
3858 cpus_and(new_mask, new_mask, cpus_allowed);
3859 retval = set_cpus_allowed(p, new_mask);
3861 out_unlock:
3862 put_task_struct(p);
3863 unlock_cpu_hotplug();
3864 return retval;
3867 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3868 cpumask_t *new_mask)
3870 if (len < sizeof(cpumask_t)) {
3871 memset(new_mask, 0, sizeof(cpumask_t));
3872 } else if (len > sizeof(cpumask_t)) {
3873 len = sizeof(cpumask_t);
3875 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3879 * sys_sched_setaffinity - set the cpu affinity of a process
3880 * @pid: pid of the process
3881 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3882 * @user_mask_ptr: user-space pointer to the new cpu mask
3884 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
3885 unsigned long __user *user_mask_ptr)
3887 cpumask_t new_mask;
3888 int retval;
3890 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
3891 if (retval)
3892 return retval;
3894 return sched_setaffinity(pid, new_mask);
3898 * Represents all cpu's present in the system
3899 * In systems capable of hotplug, this map could dynamically grow
3900 * as new cpu's are detected in the system via any platform specific
3901 * method, such as ACPI for e.g.
3904 cpumask_t cpu_present_map __read_mostly;
3905 EXPORT_SYMBOL(cpu_present_map);
3907 #ifndef CONFIG_SMP
3908 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
3909 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
3910 #endif
3912 long sched_getaffinity(pid_t pid, cpumask_t *mask)
3914 int retval;
3915 task_t *p;
3917 lock_cpu_hotplug();
3918 read_lock(&tasklist_lock);
3920 retval = -ESRCH;
3921 p = find_process_by_pid(pid);
3922 if (!p)
3923 goto out_unlock;
3925 retval = 0;
3926 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
3928 out_unlock:
3929 read_unlock(&tasklist_lock);
3930 unlock_cpu_hotplug();
3931 if (retval)
3932 return retval;
3934 return 0;
3938 * sys_sched_getaffinity - get the cpu affinity of a process
3939 * @pid: pid of the process
3940 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3941 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3943 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
3944 unsigned long __user *user_mask_ptr)
3946 int ret;
3947 cpumask_t mask;
3949 if (len < sizeof(cpumask_t))
3950 return -EINVAL;
3952 ret = sched_getaffinity(pid, &mask);
3953 if (ret < 0)
3954 return ret;
3956 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
3957 return -EFAULT;
3959 return sizeof(cpumask_t);
3963 * sys_sched_yield - yield the current processor to other threads.
3965 * this function yields the current CPU by moving the calling thread
3966 * to the expired array. If there are no other threads running on this
3967 * CPU then this function will return.
3969 asmlinkage long sys_sched_yield(void)
3971 runqueue_t *rq = this_rq_lock();
3972 prio_array_t *array = current->array;
3973 prio_array_t *target = rq->expired;
3975 schedstat_inc(rq, yld_cnt);
3977 * We implement yielding by moving the task into the expired
3978 * queue.
3980 * (special rule: RT tasks will just roundrobin in the active
3981 * array.)
3983 if (rt_task(current))
3984 target = rq->active;
3986 if (array->nr_active == 1) {
3987 schedstat_inc(rq, yld_act_empty);
3988 if (!rq->expired->nr_active)
3989 schedstat_inc(rq, yld_both_empty);
3990 } else if (!rq->expired->nr_active)
3991 schedstat_inc(rq, yld_exp_empty);
3993 if (array != target) {
3994 dequeue_task(current, array);
3995 enqueue_task(current, target);
3996 } else
3998 * requeue_task is cheaper so perform that if possible.
4000 requeue_task(current, array);
4003 * Since we are going to call schedule() anyway, there's
4004 * no need to preempt or enable interrupts:
4006 __release(rq->lock);
4007 _raw_spin_unlock(&rq->lock);
4008 preempt_enable_no_resched();
4010 schedule();
4012 return 0;
4015 static inline void __cond_resched(void)
4018 * The BKS might be reacquired before we have dropped
4019 * PREEMPT_ACTIVE, which could trigger a second
4020 * cond_resched() call.
4022 if (unlikely(preempt_count()))
4023 return;
4024 if (unlikely(system_state != SYSTEM_RUNNING))
4025 return;
4026 do {
4027 add_preempt_count(PREEMPT_ACTIVE);
4028 schedule();
4029 sub_preempt_count(PREEMPT_ACTIVE);
4030 } while (need_resched());
4033 int __sched cond_resched(void)
4035 if (need_resched()) {
4036 __cond_resched();
4037 return 1;
4039 return 0;
4042 EXPORT_SYMBOL(cond_resched);
4045 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4046 * call schedule, and on return reacquire the lock.
4048 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4049 * operations here to prevent schedule() from being called twice (once via
4050 * spin_unlock(), once by hand).
4052 int cond_resched_lock(spinlock_t *lock)
4054 int ret = 0;
4056 if (need_lockbreak(lock)) {
4057 spin_unlock(lock);
4058 cpu_relax();
4059 ret = 1;
4060 spin_lock(lock);
4062 if (need_resched()) {
4063 _raw_spin_unlock(lock);
4064 preempt_enable_no_resched();
4065 __cond_resched();
4066 ret = 1;
4067 spin_lock(lock);
4069 return ret;
4072 EXPORT_SYMBOL(cond_resched_lock);
4074 int __sched cond_resched_softirq(void)
4076 BUG_ON(!in_softirq());
4078 if (need_resched()) {
4079 __local_bh_enable();
4080 __cond_resched();
4081 local_bh_disable();
4082 return 1;
4084 return 0;
4087 EXPORT_SYMBOL(cond_resched_softirq);
4091 * yield - yield the current processor to other threads.
4093 * this is a shortcut for kernel-space yielding - it marks the
4094 * thread runnable and calls sys_sched_yield().
4096 void __sched yield(void)
4098 set_current_state(TASK_RUNNING);
4099 sys_sched_yield();
4102 EXPORT_SYMBOL(yield);
4105 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4106 * that process accounting knows that this is a task in IO wait state.
4108 * But don't do that if it is a deliberate, throttling IO wait (this task
4109 * has set its backing_dev_info: the queue against which it should throttle)
4111 void __sched io_schedule(void)
4113 struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id());
4115 atomic_inc(&rq->nr_iowait);
4116 schedule();
4117 atomic_dec(&rq->nr_iowait);
4120 EXPORT_SYMBOL(io_schedule);
4122 long __sched io_schedule_timeout(long timeout)
4124 struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id());
4125 long ret;
4127 atomic_inc(&rq->nr_iowait);
4128 ret = schedule_timeout(timeout);
4129 atomic_dec(&rq->nr_iowait);
4130 return ret;
4134 * sys_sched_get_priority_max - return maximum RT priority.
4135 * @policy: scheduling class.
4137 * this syscall returns the maximum rt_priority that can be used
4138 * by a given scheduling class.
4140 asmlinkage long sys_sched_get_priority_max(int policy)
4142 int ret = -EINVAL;
4144 switch (policy) {
4145 case SCHED_FIFO:
4146 case SCHED_RR:
4147 ret = MAX_USER_RT_PRIO-1;
4148 break;
4149 case SCHED_NORMAL:
4150 case SCHED_BATCH:
4151 ret = 0;
4152 break;
4154 return ret;
4158 * sys_sched_get_priority_min - return minimum RT priority.
4159 * @policy: scheduling class.
4161 * this syscall returns the minimum rt_priority that can be used
4162 * by a given scheduling class.
4164 asmlinkage long sys_sched_get_priority_min(int policy)
4166 int ret = -EINVAL;
4168 switch (policy) {
4169 case SCHED_FIFO:
4170 case SCHED_RR:
4171 ret = 1;
4172 break;
4173 case SCHED_NORMAL:
4174 case SCHED_BATCH:
4175 ret = 0;
4177 return ret;
4181 * sys_sched_rr_get_interval - return the default timeslice of a process.
4182 * @pid: pid of the process.
4183 * @interval: userspace pointer to the timeslice value.
4185 * this syscall writes the default timeslice value of a given process
4186 * into the user-space timespec buffer. A value of '0' means infinity.
4188 asmlinkage
4189 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4191 int retval = -EINVAL;
4192 struct timespec t;
4193 task_t *p;
4195 if (pid < 0)
4196 goto out_nounlock;
4198 retval = -ESRCH;
4199 read_lock(&tasklist_lock);
4200 p = find_process_by_pid(pid);
4201 if (!p)
4202 goto out_unlock;
4204 retval = security_task_getscheduler(p);
4205 if (retval)
4206 goto out_unlock;
4208 jiffies_to_timespec(p->policy & SCHED_FIFO ?
4209 0 : task_timeslice(p), &t);
4210 read_unlock(&tasklist_lock);
4211 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4212 out_nounlock:
4213 return retval;
4214 out_unlock:
4215 read_unlock(&tasklist_lock);
4216 return retval;
4219 static inline struct task_struct *eldest_child(struct task_struct *p)
4221 if (list_empty(&p->children)) return NULL;
4222 return list_entry(p->children.next,struct task_struct,sibling);
4225 static inline struct task_struct *older_sibling(struct task_struct *p)
4227 if (p->sibling.prev==&p->parent->children) return NULL;
4228 return list_entry(p->sibling.prev,struct task_struct,sibling);
4231 static inline struct task_struct *younger_sibling(struct task_struct *p)
4233 if (p->sibling.next==&p->parent->children) return NULL;
4234 return list_entry(p->sibling.next,struct task_struct,sibling);
4237 static void show_task(task_t *p)
4239 task_t *relative;
4240 unsigned state;
4241 unsigned long free = 0;
4242 static const char *stat_nam[] = { "R", "S", "D", "T", "t", "Z", "X" };
4244 printk("%-13.13s ", p->comm);
4245 state = p->state ? __ffs(p->state) + 1 : 0;
4246 if (state < ARRAY_SIZE(stat_nam))
4247 printk(stat_nam[state]);
4248 else
4249 printk("?");
4250 #if (BITS_PER_LONG == 32)
4251 if (state == TASK_RUNNING)
4252 printk(" running ");
4253 else
4254 printk(" %08lX ", thread_saved_pc(p));
4255 #else
4256 if (state == TASK_RUNNING)
4257 printk(" running task ");
4258 else
4259 printk(" %016lx ", thread_saved_pc(p));
4260 #endif
4261 #ifdef CONFIG_DEBUG_STACK_USAGE
4263 unsigned long *n = end_of_stack(p);
4264 while (!*n)
4265 n++;
4266 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4268 #endif
4269 printk("%5lu %5d %6d ", free, p->pid, p->parent->pid);
4270 if ((relative = eldest_child(p)))
4271 printk("%5d ", relative->pid);
4272 else
4273 printk(" ");
4274 if ((relative = younger_sibling(p)))
4275 printk("%7d", relative->pid);
4276 else
4277 printk(" ");
4278 if ((relative = older_sibling(p)))
4279 printk(" %5d", relative->pid);
4280 else
4281 printk(" ");
4282 if (!p->mm)
4283 printk(" (L-TLB)\n");
4284 else
4285 printk(" (NOTLB)\n");
4287 if (state != TASK_RUNNING)
4288 show_stack(p, NULL);
4291 void show_state(void)
4293 task_t *g, *p;
4295 #if (BITS_PER_LONG == 32)
4296 printk("\n"
4297 " sibling\n");
4298 printk(" task PC pid father child younger older\n");
4299 #else
4300 printk("\n"
4301 " sibling\n");
4302 printk(" task PC pid father child younger older\n");
4303 #endif
4304 read_lock(&tasklist_lock);
4305 do_each_thread(g, p) {
4307 * reset the NMI-timeout, listing all files on a slow
4308 * console might take alot of time:
4310 touch_nmi_watchdog();
4311 show_task(p);
4312 } while_each_thread(g, p);
4314 read_unlock(&tasklist_lock);
4315 mutex_debug_show_all_locks();
4319 * init_idle - set up an idle thread for a given CPU
4320 * @idle: task in question
4321 * @cpu: cpu the idle task belongs to
4323 * NOTE: this function does not set the idle thread's NEED_RESCHED
4324 * flag, to make booting more robust.
4326 void __devinit init_idle(task_t *idle, int cpu)
4328 runqueue_t *rq = cpu_rq(cpu);
4329 unsigned long flags;
4331 idle->timestamp = sched_clock();
4332 idle->sleep_avg = 0;
4333 idle->array = NULL;
4334 idle->prio = MAX_PRIO;
4335 idle->state = TASK_RUNNING;
4336 idle->cpus_allowed = cpumask_of_cpu(cpu);
4337 set_task_cpu(idle, cpu);
4339 spin_lock_irqsave(&rq->lock, flags);
4340 rq->curr = rq->idle = idle;
4341 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4342 idle->oncpu = 1;
4343 #endif
4344 spin_unlock_irqrestore(&rq->lock, flags);
4346 /* Set the preempt count _outside_ the spinlocks! */
4347 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4348 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4349 #else
4350 task_thread_info(idle)->preempt_count = 0;
4351 #endif
4355 * In a system that switches off the HZ timer nohz_cpu_mask
4356 * indicates which cpus entered this state. This is used
4357 * in the rcu update to wait only for active cpus. For system
4358 * which do not switch off the HZ timer nohz_cpu_mask should
4359 * always be CPU_MASK_NONE.
4361 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4363 #ifdef CONFIG_SMP
4365 * This is how migration works:
4367 * 1) we queue a migration_req_t structure in the source CPU's
4368 * runqueue and wake up that CPU's migration thread.
4369 * 2) we down() the locked semaphore => thread blocks.
4370 * 3) migration thread wakes up (implicitly it forces the migrated
4371 * thread off the CPU)
4372 * 4) it gets the migration request and checks whether the migrated
4373 * task is still in the wrong runqueue.
4374 * 5) if it's in the wrong runqueue then the migration thread removes
4375 * it and puts it into the right queue.
4376 * 6) migration thread up()s the semaphore.
4377 * 7) we wake up and the migration is done.
4381 * Change a given task's CPU affinity. Migrate the thread to a
4382 * proper CPU and schedule it away if the CPU it's executing on
4383 * is removed from the allowed bitmask.
4385 * NOTE: the caller must have a valid reference to the task, the
4386 * task must not exit() & deallocate itself prematurely. The
4387 * call is not atomic; no spinlocks may be held.
4389 int set_cpus_allowed(task_t *p, cpumask_t new_mask)
4391 unsigned long flags;
4392 int ret = 0;
4393 migration_req_t req;
4394 runqueue_t *rq;
4396 rq = task_rq_lock(p, &flags);
4397 if (!cpus_intersects(new_mask, cpu_online_map)) {
4398 ret = -EINVAL;
4399 goto out;
4402 p->cpus_allowed = new_mask;
4403 /* Can the task run on the task's current CPU? If so, we're done */
4404 if (cpu_isset(task_cpu(p), new_mask))
4405 goto out;
4407 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4408 /* Need help from migration thread: drop lock and wait. */
4409 task_rq_unlock(rq, &flags);
4410 wake_up_process(rq->migration_thread);
4411 wait_for_completion(&req.done);
4412 tlb_migrate_finish(p->mm);
4413 return 0;
4415 out:
4416 task_rq_unlock(rq, &flags);
4417 return ret;
4420 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4423 * Move (not current) task off this cpu, onto dest cpu. We're doing
4424 * this because either it can't run here any more (set_cpus_allowed()
4425 * away from this CPU, or CPU going down), or because we're
4426 * attempting to rebalance this task on exec (sched_exec).
4428 * So we race with normal scheduler movements, but that's OK, as long
4429 * as the task is no longer on this CPU.
4431 static void __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4433 runqueue_t *rq_dest, *rq_src;
4435 if (unlikely(cpu_is_offline(dest_cpu)))
4436 return;
4438 rq_src = cpu_rq(src_cpu);
4439 rq_dest = cpu_rq(dest_cpu);
4441 double_rq_lock(rq_src, rq_dest);
4442 /* Already moved. */
4443 if (task_cpu(p) != src_cpu)
4444 goto out;
4445 /* Affinity changed (again). */
4446 if (!cpu_isset(dest_cpu, p->cpus_allowed))
4447 goto out;
4449 set_task_cpu(p, dest_cpu);
4450 if (p->array) {
4452 * Sync timestamp with rq_dest's before activating.
4453 * The same thing could be achieved by doing this step
4454 * afterwards, and pretending it was a local activate.
4455 * This way is cleaner and logically correct.
4457 p->timestamp = p->timestamp - rq_src->timestamp_last_tick
4458 + rq_dest->timestamp_last_tick;
4459 deactivate_task(p, rq_src);
4460 activate_task(p, rq_dest, 0);
4461 if (TASK_PREEMPTS_CURR(p, rq_dest))
4462 resched_task(rq_dest->curr);
4465 out:
4466 double_rq_unlock(rq_src, rq_dest);
4470 * migration_thread - this is a highprio system thread that performs
4471 * thread migration by bumping thread off CPU then 'pushing' onto
4472 * another runqueue.
4474 static int migration_thread(void *data)
4476 runqueue_t *rq;
4477 int cpu = (long)data;
4479 rq = cpu_rq(cpu);
4480 BUG_ON(rq->migration_thread != current);
4482 set_current_state(TASK_INTERRUPTIBLE);
4483 while (!kthread_should_stop()) {
4484 struct list_head *head;
4485 migration_req_t *req;
4487 try_to_freeze();
4489 spin_lock_irq(&rq->lock);
4491 if (cpu_is_offline(cpu)) {
4492 spin_unlock_irq(&rq->lock);
4493 goto wait_to_die;
4496 if (rq->active_balance) {
4497 active_load_balance(rq, cpu);
4498 rq->active_balance = 0;
4501 head = &rq->migration_queue;
4503 if (list_empty(head)) {
4504 spin_unlock_irq(&rq->lock);
4505 schedule();
4506 set_current_state(TASK_INTERRUPTIBLE);
4507 continue;
4509 req = list_entry(head->next, migration_req_t, list);
4510 list_del_init(head->next);
4512 spin_unlock(&rq->lock);
4513 __migrate_task(req->task, cpu, req->dest_cpu);
4514 local_irq_enable();
4516 complete(&req->done);
4518 __set_current_state(TASK_RUNNING);
4519 return 0;
4521 wait_to_die:
4522 /* Wait for kthread_stop */
4523 set_current_state(TASK_INTERRUPTIBLE);
4524 while (!kthread_should_stop()) {
4525 schedule();
4526 set_current_state(TASK_INTERRUPTIBLE);
4528 __set_current_state(TASK_RUNNING);
4529 return 0;
4532 #ifdef CONFIG_HOTPLUG_CPU
4533 /* Figure out where task on dead CPU should go, use force if neccessary. */
4534 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *tsk)
4536 int dest_cpu;
4537 cpumask_t mask;
4539 /* On same node? */
4540 mask = node_to_cpumask(cpu_to_node(dead_cpu));
4541 cpus_and(mask, mask, tsk->cpus_allowed);
4542 dest_cpu = any_online_cpu(mask);
4544 /* On any allowed CPU? */
4545 if (dest_cpu == NR_CPUS)
4546 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4548 /* No more Mr. Nice Guy. */
4549 if (dest_cpu == NR_CPUS) {
4550 cpus_setall(tsk->cpus_allowed);
4551 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4554 * Don't tell them about moving exiting tasks or
4555 * kernel threads (both mm NULL), since they never
4556 * leave kernel.
4558 if (tsk->mm && printk_ratelimit())
4559 printk(KERN_INFO "process %d (%s) no "
4560 "longer affine to cpu%d\n",
4561 tsk->pid, tsk->comm, dead_cpu);
4563 __migrate_task(tsk, dead_cpu, dest_cpu);
4567 * While a dead CPU has no uninterruptible tasks queued at this point,
4568 * it might still have a nonzero ->nr_uninterruptible counter, because
4569 * for performance reasons the counter is not stricly tracking tasks to
4570 * their home CPUs. So we just add the counter to another CPU's counter,
4571 * to keep the global sum constant after CPU-down:
4573 static void migrate_nr_uninterruptible(runqueue_t *rq_src)
4575 runqueue_t *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
4576 unsigned long flags;
4578 local_irq_save(flags);
4579 double_rq_lock(rq_src, rq_dest);
4580 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
4581 rq_src->nr_uninterruptible = 0;
4582 double_rq_unlock(rq_src, rq_dest);
4583 local_irq_restore(flags);
4586 /* Run through task list and migrate tasks from the dead cpu. */
4587 static void migrate_live_tasks(int src_cpu)
4589 struct task_struct *tsk, *t;
4591 write_lock_irq(&tasklist_lock);
4593 do_each_thread(t, tsk) {
4594 if (tsk == current)
4595 continue;
4597 if (task_cpu(tsk) == src_cpu)
4598 move_task_off_dead_cpu(src_cpu, tsk);
4599 } while_each_thread(t, tsk);
4601 write_unlock_irq(&tasklist_lock);
4604 /* Schedules idle task to be the next runnable task on current CPU.
4605 * It does so by boosting its priority to highest possible and adding it to
4606 * the _front_ of runqueue. Used by CPU offline code.
4608 void sched_idle_next(void)
4610 int cpu = smp_processor_id();
4611 runqueue_t *rq = this_rq();
4612 struct task_struct *p = rq->idle;
4613 unsigned long flags;
4615 /* cpu has to be offline */
4616 BUG_ON(cpu_online(cpu));
4618 /* Strictly not necessary since rest of the CPUs are stopped by now
4619 * and interrupts disabled on current cpu.
4621 spin_lock_irqsave(&rq->lock, flags);
4623 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4624 /* Add idle task to _front_ of it's priority queue */
4625 __activate_idle_task(p, rq);
4627 spin_unlock_irqrestore(&rq->lock, flags);
4630 /* Ensures that the idle task is using init_mm right before its cpu goes
4631 * offline.
4633 void idle_task_exit(void)
4635 struct mm_struct *mm = current->active_mm;
4637 BUG_ON(cpu_online(smp_processor_id()));
4639 if (mm != &init_mm)
4640 switch_mm(mm, &init_mm, current);
4641 mmdrop(mm);
4644 static void migrate_dead(unsigned int dead_cpu, task_t *tsk)
4646 struct runqueue *rq = cpu_rq(dead_cpu);
4648 /* Must be exiting, otherwise would be on tasklist. */
4649 BUG_ON(tsk->exit_state != EXIT_ZOMBIE && tsk->exit_state != EXIT_DEAD);
4651 /* Cannot have done final schedule yet: would have vanished. */
4652 BUG_ON(tsk->flags & PF_DEAD);
4654 get_task_struct(tsk);
4657 * Drop lock around migration; if someone else moves it,
4658 * that's OK. No task can be added to this CPU, so iteration is
4659 * fine.
4661 spin_unlock_irq(&rq->lock);
4662 move_task_off_dead_cpu(dead_cpu, tsk);
4663 spin_lock_irq(&rq->lock);
4665 put_task_struct(tsk);
4668 /* release_task() removes task from tasklist, so we won't find dead tasks. */
4669 static void migrate_dead_tasks(unsigned int dead_cpu)
4671 unsigned arr, i;
4672 struct runqueue *rq = cpu_rq(dead_cpu);
4674 for (arr = 0; arr < 2; arr++) {
4675 for (i = 0; i < MAX_PRIO; i++) {
4676 struct list_head *list = &rq->arrays[arr].queue[i];
4677 while (!list_empty(list))
4678 migrate_dead(dead_cpu,
4679 list_entry(list->next, task_t,
4680 run_list));
4684 #endif /* CONFIG_HOTPLUG_CPU */
4687 * migration_call - callback that gets triggered when a CPU is added.
4688 * Here we can start up the necessary migration thread for the new CPU.
4690 static int migration_call(struct notifier_block *nfb, unsigned long action,
4691 void *hcpu)
4693 int cpu = (long)hcpu;
4694 struct task_struct *p;
4695 struct runqueue *rq;
4696 unsigned long flags;
4698 switch (action) {
4699 case CPU_UP_PREPARE:
4700 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
4701 if (IS_ERR(p))
4702 return NOTIFY_BAD;
4703 p->flags |= PF_NOFREEZE;
4704 kthread_bind(p, cpu);
4705 /* Must be high prio: stop_machine expects to yield to it. */
4706 rq = task_rq_lock(p, &flags);
4707 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4708 task_rq_unlock(rq, &flags);
4709 cpu_rq(cpu)->migration_thread = p;
4710 break;
4711 case CPU_ONLINE:
4712 /* Strictly unneccessary, as first user will wake it. */
4713 wake_up_process(cpu_rq(cpu)->migration_thread);
4714 break;
4715 #ifdef CONFIG_HOTPLUG_CPU
4716 case CPU_UP_CANCELED:
4717 /* Unbind it from offline cpu so it can run. Fall thru. */
4718 kthread_bind(cpu_rq(cpu)->migration_thread,
4719 any_online_cpu(cpu_online_map));
4720 kthread_stop(cpu_rq(cpu)->migration_thread);
4721 cpu_rq(cpu)->migration_thread = NULL;
4722 break;
4723 case CPU_DEAD:
4724 migrate_live_tasks(cpu);
4725 rq = cpu_rq(cpu);
4726 kthread_stop(rq->migration_thread);
4727 rq->migration_thread = NULL;
4728 /* Idle task back to normal (off runqueue, low prio) */
4729 rq = task_rq_lock(rq->idle, &flags);
4730 deactivate_task(rq->idle, rq);
4731 rq->idle->static_prio = MAX_PRIO;
4732 __setscheduler(rq->idle, SCHED_NORMAL, 0);
4733 migrate_dead_tasks(cpu);
4734 task_rq_unlock(rq, &flags);
4735 migrate_nr_uninterruptible(rq);
4736 BUG_ON(rq->nr_running != 0);
4738 /* No need to migrate the tasks: it was best-effort if
4739 * they didn't do lock_cpu_hotplug(). Just wake up
4740 * the requestors. */
4741 spin_lock_irq(&rq->lock);
4742 while (!list_empty(&rq->migration_queue)) {
4743 migration_req_t *req;
4744 req = list_entry(rq->migration_queue.next,
4745 migration_req_t, list);
4746 list_del_init(&req->list);
4747 complete(&req->done);
4749 spin_unlock_irq(&rq->lock);
4750 break;
4751 #endif
4753 return NOTIFY_OK;
4756 /* Register at highest priority so that task migration (migrate_all_tasks)
4757 * happens before everything else.
4759 static struct notifier_block __devinitdata migration_notifier = {
4760 .notifier_call = migration_call,
4761 .priority = 10
4764 int __init migration_init(void)
4766 void *cpu = (void *)(long)smp_processor_id();
4767 /* Start one for boot CPU. */
4768 migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
4769 migration_call(&migration_notifier, CPU_ONLINE, cpu);
4770 register_cpu_notifier(&migration_notifier);
4771 return 0;
4773 #endif
4775 #ifdef CONFIG_SMP
4776 #undef SCHED_DOMAIN_DEBUG
4777 #ifdef SCHED_DOMAIN_DEBUG
4778 static void sched_domain_debug(struct sched_domain *sd, int cpu)
4780 int level = 0;
4782 if (!sd) {
4783 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
4784 return;
4787 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
4789 do {
4790 int i;
4791 char str[NR_CPUS];
4792 struct sched_group *group = sd->groups;
4793 cpumask_t groupmask;
4795 cpumask_scnprintf(str, NR_CPUS, sd->span);
4796 cpus_clear(groupmask);
4798 printk(KERN_DEBUG);
4799 for (i = 0; i < level + 1; i++)
4800 printk(" ");
4801 printk("domain %d: ", level);
4803 if (!(sd->flags & SD_LOAD_BALANCE)) {
4804 printk("does not load-balance\n");
4805 if (sd->parent)
4806 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent");
4807 break;
4810 printk("span %s\n", str);
4812 if (!cpu_isset(cpu, sd->span))
4813 printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
4814 if (!cpu_isset(cpu, group->cpumask))
4815 printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
4817 printk(KERN_DEBUG);
4818 for (i = 0; i < level + 2; i++)
4819 printk(" ");
4820 printk("groups:");
4821 do {
4822 if (!group) {
4823 printk("\n");
4824 printk(KERN_ERR "ERROR: group is NULL\n");
4825 break;
4828 if (!group->cpu_power) {
4829 printk("\n");
4830 printk(KERN_ERR "ERROR: domain->cpu_power not set\n");
4833 if (!cpus_weight(group->cpumask)) {
4834 printk("\n");
4835 printk(KERN_ERR "ERROR: empty group\n");
4838 if (cpus_intersects(groupmask, group->cpumask)) {
4839 printk("\n");
4840 printk(KERN_ERR "ERROR: repeated CPUs\n");
4843 cpus_or(groupmask, groupmask, group->cpumask);
4845 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
4846 printk(" %s", str);
4848 group = group->next;
4849 } while (group != sd->groups);
4850 printk("\n");
4852 if (!cpus_equal(sd->span, groupmask))
4853 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
4855 level++;
4856 sd = sd->parent;
4858 if (sd) {
4859 if (!cpus_subset(groupmask, sd->span))
4860 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
4863 } while (sd);
4865 #else
4866 #define sched_domain_debug(sd, cpu) {}
4867 #endif
4869 static int sd_degenerate(struct sched_domain *sd)
4871 if (cpus_weight(sd->span) == 1)
4872 return 1;
4874 /* Following flags need at least 2 groups */
4875 if (sd->flags & (SD_LOAD_BALANCE |
4876 SD_BALANCE_NEWIDLE |
4877 SD_BALANCE_FORK |
4878 SD_BALANCE_EXEC)) {
4879 if (sd->groups != sd->groups->next)
4880 return 0;
4883 /* Following flags don't use groups */
4884 if (sd->flags & (SD_WAKE_IDLE |
4885 SD_WAKE_AFFINE |
4886 SD_WAKE_BALANCE))
4887 return 0;
4889 return 1;
4892 static int sd_parent_degenerate(struct sched_domain *sd,
4893 struct sched_domain *parent)
4895 unsigned long cflags = sd->flags, pflags = parent->flags;
4897 if (sd_degenerate(parent))
4898 return 1;
4900 if (!cpus_equal(sd->span, parent->span))
4901 return 0;
4903 /* Does parent contain flags not in child? */
4904 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
4905 if (cflags & SD_WAKE_AFFINE)
4906 pflags &= ~SD_WAKE_BALANCE;
4907 /* Flags needing groups don't count if only 1 group in parent */
4908 if (parent->groups == parent->groups->next) {
4909 pflags &= ~(SD_LOAD_BALANCE |
4910 SD_BALANCE_NEWIDLE |
4911 SD_BALANCE_FORK |
4912 SD_BALANCE_EXEC);
4914 if (~cflags & pflags)
4915 return 0;
4917 return 1;
4921 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
4922 * hold the hotplug lock.
4924 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
4926 runqueue_t *rq = cpu_rq(cpu);
4927 struct sched_domain *tmp;
4929 /* Remove the sched domains which do not contribute to scheduling. */
4930 for (tmp = sd; tmp; tmp = tmp->parent) {
4931 struct sched_domain *parent = tmp->parent;
4932 if (!parent)
4933 break;
4934 if (sd_parent_degenerate(tmp, parent))
4935 tmp->parent = parent->parent;
4938 if (sd && sd_degenerate(sd))
4939 sd = sd->parent;
4941 sched_domain_debug(sd, cpu);
4943 rcu_assign_pointer(rq->sd, sd);
4946 /* cpus with isolated domains */
4947 static cpumask_t __devinitdata cpu_isolated_map = CPU_MASK_NONE;
4949 /* Setup the mask of cpus configured for isolated domains */
4950 static int __init isolated_cpu_setup(char *str)
4952 int ints[NR_CPUS], i;
4954 str = get_options(str, ARRAY_SIZE(ints), ints);
4955 cpus_clear(cpu_isolated_map);
4956 for (i = 1; i <= ints[0]; i++)
4957 if (ints[i] < NR_CPUS)
4958 cpu_set(ints[i], cpu_isolated_map);
4959 return 1;
4962 __setup ("isolcpus=", isolated_cpu_setup);
4965 * init_sched_build_groups takes an array of groups, the cpumask we wish
4966 * to span, and a pointer to a function which identifies what group a CPU
4967 * belongs to. The return value of group_fn must be a valid index into the
4968 * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
4969 * keep track of groups covered with a cpumask_t).
4971 * init_sched_build_groups will build a circular linked list of the groups
4972 * covered by the given span, and will set each group's ->cpumask correctly,
4973 * and ->cpu_power to 0.
4975 static void init_sched_build_groups(struct sched_group groups[], cpumask_t span,
4976 int (*group_fn)(int cpu))
4978 struct sched_group *first = NULL, *last = NULL;
4979 cpumask_t covered = CPU_MASK_NONE;
4980 int i;
4982 for_each_cpu_mask(i, span) {
4983 int group = group_fn(i);
4984 struct sched_group *sg = &groups[group];
4985 int j;
4987 if (cpu_isset(i, covered))
4988 continue;
4990 sg->cpumask = CPU_MASK_NONE;
4991 sg->cpu_power = 0;
4993 for_each_cpu_mask(j, span) {
4994 if (group_fn(j) != group)
4995 continue;
4997 cpu_set(j, covered);
4998 cpu_set(j, sg->cpumask);
5000 if (!first)
5001 first = sg;
5002 if (last)
5003 last->next = sg;
5004 last = sg;
5006 last->next = first;
5009 #define SD_NODES_PER_DOMAIN 16
5012 * Self-tuning task migration cost measurement between source and target CPUs.
5014 * This is done by measuring the cost of manipulating buffers of varying
5015 * sizes. For a given buffer-size here are the steps that are taken:
5017 * 1) the source CPU reads+dirties a shared buffer
5018 * 2) the target CPU reads+dirties the same shared buffer
5020 * We measure how long they take, in the following 4 scenarios:
5022 * - source: CPU1, target: CPU2 | cost1
5023 * - source: CPU2, target: CPU1 | cost2
5024 * - source: CPU1, target: CPU1 | cost3
5025 * - source: CPU2, target: CPU2 | cost4
5027 * We then calculate the cost3+cost4-cost1-cost2 difference - this is
5028 * the cost of migration.
5030 * We then start off from a small buffer-size and iterate up to larger
5031 * buffer sizes, in 5% steps - measuring each buffer-size separately, and
5032 * doing a maximum search for the cost. (The maximum cost for a migration
5033 * normally occurs when the working set size is around the effective cache
5034 * size.)
5036 #define SEARCH_SCOPE 2
5037 #define MIN_CACHE_SIZE (64*1024U)
5038 #define DEFAULT_CACHE_SIZE (5*1024*1024U)
5039 #define ITERATIONS 1
5040 #define SIZE_THRESH 130
5041 #define COST_THRESH 130
5044 * The migration cost is a function of 'domain distance'. Domain
5045 * distance is the number of steps a CPU has to iterate down its
5046 * domain tree to share a domain with the other CPU. The farther
5047 * two CPUs are from each other, the larger the distance gets.
5049 * Note that we use the distance only to cache measurement results,
5050 * the distance value is not used numerically otherwise. When two
5051 * CPUs have the same distance it is assumed that the migration
5052 * cost is the same. (this is a simplification but quite practical)
5054 #define MAX_DOMAIN_DISTANCE 32
5056 static unsigned long long migration_cost[MAX_DOMAIN_DISTANCE] =
5057 { [ 0 ... MAX_DOMAIN_DISTANCE-1 ] =
5059 * Architectures may override the migration cost and thus avoid
5060 * boot-time calibration. Unit is nanoseconds. Mostly useful for
5061 * virtualized hardware:
5063 #ifdef CONFIG_DEFAULT_MIGRATION_COST
5064 CONFIG_DEFAULT_MIGRATION_COST
5065 #else
5066 -1LL
5067 #endif
5071 * Allow override of migration cost - in units of microseconds.
5072 * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost
5073 * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs:
5075 static int __init migration_cost_setup(char *str)
5077 int ints[MAX_DOMAIN_DISTANCE+1], i;
5079 str = get_options(str, ARRAY_SIZE(ints), ints);
5081 printk("#ints: %d\n", ints[0]);
5082 for (i = 1; i <= ints[0]; i++) {
5083 migration_cost[i-1] = (unsigned long long)ints[i]*1000;
5084 printk("migration_cost[%d]: %Ld\n", i-1, migration_cost[i-1]);
5086 return 1;
5089 __setup ("migration_cost=", migration_cost_setup);
5092 * Global multiplier (divisor) for migration-cutoff values,
5093 * in percentiles. E.g. use a value of 150 to get 1.5 times
5094 * longer cache-hot cutoff times.
5096 * (We scale it from 100 to 128 to long long handling easier.)
5099 #define MIGRATION_FACTOR_SCALE 128
5101 static unsigned int migration_factor = MIGRATION_FACTOR_SCALE;
5103 static int __init setup_migration_factor(char *str)
5105 get_option(&str, &migration_factor);
5106 migration_factor = migration_factor * MIGRATION_FACTOR_SCALE / 100;
5107 return 1;
5110 __setup("migration_factor=", setup_migration_factor);
5113 * Estimated distance of two CPUs, measured via the number of domains
5114 * we have to pass for the two CPUs to be in the same span:
5116 static unsigned long domain_distance(int cpu1, int cpu2)
5118 unsigned long distance = 0;
5119 struct sched_domain *sd;
5121 for_each_domain(cpu1, sd) {
5122 WARN_ON(!cpu_isset(cpu1, sd->span));
5123 if (cpu_isset(cpu2, sd->span))
5124 return distance;
5125 distance++;
5127 if (distance >= MAX_DOMAIN_DISTANCE) {
5128 WARN_ON(1);
5129 distance = MAX_DOMAIN_DISTANCE-1;
5132 return distance;
5135 static unsigned int migration_debug;
5137 static int __init setup_migration_debug(char *str)
5139 get_option(&str, &migration_debug);
5140 return 1;
5143 __setup("migration_debug=", setup_migration_debug);
5146 * Maximum cache-size that the scheduler should try to measure.
5147 * Architectures with larger caches should tune this up during
5148 * bootup. Gets used in the domain-setup code (i.e. during SMP
5149 * bootup).
5151 unsigned int max_cache_size;
5153 static int __init setup_max_cache_size(char *str)
5155 get_option(&str, &max_cache_size);
5156 return 1;
5159 __setup("max_cache_size=", setup_max_cache_size);
5162 * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This
5163 * is the operation that is timed, so we try to generate unpredictable
5164 * cachemisses that still end up filling the L2 cache:
5166 static void touch_cache(void *__cache, unsigned long __size)
5168 unsigned long size = __size/sizeof(long), chunk1 = size/3,
5169 chunk2 = 2*size/3;
5170 unsigned long *cache = __cache;
5171 int i;
5173 for (i = 0; i < size/6; i += 8) {
5174 switch (i % 6) {
5175 case 0: cache[i]++;
5176 case 1: cache[size-1-i]++;
5177 case 2: cache[chunk1-i]++;
5178 case 3: cache[chunk1+i]++;
5179 case 4: cache[chunk2-i]++;
5180 case 5: cache[chunk2+i]++;
5186 * Measure the cache-cost of one task migration. Returns in units of nsec.
5188 static unsigned long long measure_one(void *cache, unsigned long size,
5189 int source, int target)
5191 cpumask_t mask, saved_mask;
5192 unsigned long long t0, t1, t2, t3, cost;
5194 saved_mask = current->cpus_allowed;
5197 * Flush source caches to RAM and invalidate them:
5199 sched_cacheflush();
5202 * Migrate to the source CPU:
5204 mask = cpumask_of_cpu(source);
5205 set_cpus_allowed(current, mask);
5206 WARN_ON(smp_processor_id() != source);
5209 * Dirty the working set:
5211 t0 = sched_clock();
5212 touch_cache(cache, size);
5213 t1 = sched_clock();
5216 * Migrate to the target CPU, dirty the L2 cache and access
5217 * the shared buffer. (which represents the working set
5218 * of a migrated task.)
5220 mask = cpumask_of_cpu(target);
5221 set_cpus_allowed(current, mask);
5222 WARN_ON(smp_processor_id() != target);
5224 t2 = sched_clock();
5225 touch_cache(cache, size);
5226 t3 = sched_clock();
5228 cost = t1-t0 + t3-t2;
5230 if (migration_debug >= 2)
5231 printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n",
5232 source, target, t1-t0, t1-t0, t3-t2, cost);
5234 * Flush target caches to RAM and invalidate them:
5236 sched_cacheflush();
5238 set_cpus_allowed(current, saved_mask);
5240 return cost;
5244 * Measure a series of task migrations and return the average
5245 * result. Since this code runs early during bootup the system
5246 * is 'undisturbed' and the average latency makes sense.
5248 * The algorithm in essence auto-detects the relevant cache-size,
5249 * so it will properly detect different cachesizes for different
5250 * cache-hierarchies, depending on how the CPUs are connected.
5252 * Architectures can prime the upper limit of the search range via
5253 * max_cache_size, otherwise the search range defaults to 20MB...64K.
5255 static unsigned long long
5256 measure_cost(int cpu1, int cpu2, void *cache, unsigned int size)
5258 unsigned long long cost1, cost2;
5259 int i;
5262 * Measure the migration cost of 'size' bytes, over an
5263 * average of 10 runs:
5265 * (We perturb the cache size by a small (0..4k)
5266 * value to compensate size/alignment related artifacts.
5267 * We also subtract the cost of the operation done on
5268 * the same CPU.)
5270 cost1 = 0;
5273 * dry run, to make sure we start off cache-cold on cpu1,
5274 * and to get any vmalloc pagefaults in advance:
5276 measure_one(cache, size, cpu1, cpu2);
5277 for (i = 0; i < ITERATIONS; i++)
5278 cost1 += measure_one(cache, size - i*1024, cpu1, cpu2);
5280 measure_one(cache, size, cpu2, cpu1);
5281 for (i = 0; i < ITERATIONS; i++)
5282 cost1 += measure_one(cache, size - i*1024, cpu2, cpu1);
5285 * (We measure the non-migrating [cached] cost on both
5286 * cpu1 and cpu2, to handle CPUs with different speeds)
5288 cost2 = 0;
5290 measure_one(cache, size, cpu1, cpu1);
5291 for (i = 0; i < ITERATIONS; i++)
5292 cost2 += measure_one(cache, size - i*1024, cpu1, cpu1);
5294 measure_one(cache, size, cpu2, cpu2);
5295 for (i = 0; i < ITERATIONS; i++)
5296 cost2 += measure_one(cache, size - i*1024, cpu2, cpu2);
5299 * Get the per-iteration migration cost:
5301 do_div(cost1, 2*ITERATIONS);
5302 do_div(cost2, 2*ITERATIONS);
5304 return cost1 - cost2;
5307 static unsigned long long measure_migration_cost(int cpu1, int cpu2)
5309 unsigned long long max_cost = 0, fluct = 0, avg_fluct = 0;
5310 unsigned int max_size, size, size_found = 0;
5311 long long cost = 0, prev_cost;
5312 void *cache;
5315 * Search from max_cache_size*5 down to 64K - the real relevant
5316 * cachesize has to lie somewhere inbetween.
5318 if (max_cache_size) {
5319 max_size = max(max_cache_size * SEARCH_SCOPE, MIN_CACHE_SIZE);
5320 size = max(max_cache_size / SEARCH_SCOPE, MIN_CACHE_SIZE);
5321 } else {
5323 * Since we have no estimation about the relevant
5324 * search range
5326 max_size = DEFAULT_CACHE_SIZE * SEARCH_SCOPE;
5327 size = MIN_CACHE_SIZE;
5330 if (!cpu_online(cpu1) || !cpu_online(cpu2)) {
5331 printk("cpu %d and %d not both online!\n", cpu1, cpu2);
5332 return 0;
5336 * Allocate the working set:
5338 cache = vmalloc(max_size);
5339 if (!cache) {
5340 printk("could not vmalloc %d bytes for cache!\n", 2*max_size);
5341 return 1000000; // return 1 msec on very small boxen
5344 while (size <= max_size) {
5345 prev_cost = cost;
5346 cost = measure_cost(cpu1, cpu2, cache, size);
5349 * Update the max:
5351 if (cost > 0) {
5352 if (max_cost < cost) {
5353 max_cost = cost;
5354 size_found = size;
5358 * Calculate average fluctuation, we use this to prevent
5359 * noise from triggering an early break out of the loop:
5361 fluct = abs(cost - prev_cost);
5362 avg_fluct = (avg_fluct + fluct)/2;
5364 if (migration_debug)
5365 printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): (%8Ld %8Ld)\n",
5366 cpu1, cpu2, size,
5367 (long)cost / 1000000,
5368 ((long)cost / 100000) % 10,
5369 (long)max_cost / 1000000,
5370 ((long)max_cost / 100000) % 10,
5371 domain_distance(cpu1, cpu2),
5372 cost, avg_fluct);
5375 * If we iterated at least 20% past the previous maximum,
5376 * and the cost has dropped by more than 20% already,
5377 * (taking fluctuations into account) then we assume to
5378 * have found the maximum and break out of the loop early:
5380 if (size_found && (size*100 > size_found*SIZE_THRESH))
5381 if (cost+avg_fluct <= 0 ||
5382 max_cost*100 > (cost+avg_fluct)*COST_THRESH) {
5384 if (migration_debug)
5385 printk("-> found max.\n");
5386 break;
5389 * Increase the cachesize in 10% steps:
5391 size = size * 10 / 9;
5394 if (migration_debug)
5395 printk("[%d][%d] working set size found: %d, cost: %Ld\n",
5396 cpu1, cpu2, size_found, max_cost);
5398 vfree(cache);
5401 * A task is considered 'cache cold' if at least 2 times
5402 * the worst-case cost of migration has passed.
5404 * (this limit is only listened to if the load-balancing
5405 * situation is 'nice' - if there is a large imbalance we
5406 * ignore it for the sake of CPU utilization and
5407 * processing fairness.)
5409 return 2 * max_cost * migration_factor / MIGRATION_FACTOR_SCALE;
5412 static void calibrate_migration_costs(const cpumask_t *cpu_map)
5414 int cpu1 = -1, cpu2 = -1, cpu, orig_cpu = raw_smp_processor_id();
5415 unsigned long j0, j1, distance, max_distance = 0;
5416 struct sched_domain *sd;
5418 j0 = jiffies;
5421 * First pass - calculate the cacheflush times:
5423 for_each_cpu_mask(cpu1, *cpu_map) {
5424 for_each_cpu_mask(cpu2, *cpu_map) {
5425 if (cpu1 == cpu2)
5426 continue;
5427 distance = domain_distance(cpu1, cpu2);
5428 max_distance = max(max_distance, distance);
5430 * No result cached yet?
5432 if (migration_cost[distance] == -1LL)
5433 migration_cost[distance] =
5434 measure_migration_cost(cpu1, cpu2);
5438 * Second pass - update the sched domain hierarchy with
5439 * the new cache-hot-time estimations:
5441 for_each_cpu_mask(cpu, *cpu_map) {
5442 distance = 0;
5443 for_each_domain(cpu, sd) {
5444 sd->cache_hot_time = migration_cost[distance];
5445 distance++;
5449 * Print the matrix:
5451 if (migration_debug)
5452 printk("migration: max_cache_size: %d, cpu: %d MHz:\n",
5453 max_cache_size,
5454 #ifdef CONFIG_X86
5455 cpu_khz/1000
5456 #else
5458 #endif
5460 if (system_state == SYSTEM_BOOTING) {
5461 printk("migration_cost=");
5462 for (distance = 0; distance <= max_distance; distance++) {
5463 if (distance)
5464 printk(",");
5465 printk("%ld", (long)migration_cost[distance] / 1000);
5467 printk("\n");
5469 j1 = jiffies;
5470 if (migration_debug)
5471 printk("migration: %ld seconds\n", (j1-j0)/HZ);
5474 * Move back to the original CPU. NUMA-Q gets confused
5475 * if we migrate to another quad during bootup.
5477 if (raw_smp_processor_id() != orig_cpu) {
5478 cpumask_t mask = cpumask_of_cpu(orig_cpu),
5479 saved_mask = current->cpus_allowed;
5481 set_cpus_allowed(current, mask);
5482 set_cpus_allowed(current, saved_mask);
5486 #ifdef CONFIG_NUMA
5489 * find_next_best_node - find the next node to include in a sched_domain
5490 * @node: node whose sched_domain we're building
5491 * @used_nodes: nodes already in the sched_domain
5493 * Find the next node to include in a given scheduling domain. Simply
5494 * finds the closest node not already in the @used_nodes map.
5496 * Should use nodemask_t.
5498 static int find_next_best_node(int node, unsigned long *used_nodes)
5500 int i, n, val, min_val, best_node = 0;
5502 min_val = INT_MAX;
5504 for (i = 0; i < MAX_NUMNODES; i++) {
5505 /* Start at @node */
5506 n = (node + i) % MAX_NUMNODES;
5508 if (!nr_cpus_node(n))
5509 continue;
5511 /* Skip already used nodes */
5512 if (test_bit(n, used_nodes))
5513 continue;
5515 /* Simple min distance search */
5516 val = node_distance(node, n);
5518 if (val < min_val) {
5519 min_val = val;
5520 best_node = n;
5524 set_bit(best_node, used_nodes);
5525 return best_node;
5529 * sched_domain_node_span - get a cpumask for a node's sched_domain
5530 * @node: node whose cpumask we're constructing
5531 * @size: number of nodes to include in this span
5533 * Given a node, construct a good cpumask for its sched_domain to span. It
5534 * should be one that prevents unnecessary balancing, but also spreads tasks
5535 * out optimally.
5537 static cpumask_t sched_domain_node_span(int node)
5539 int i;
5540 cpumask_t span, nodemask;
5541 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5543 cpus_clear(span);
5544 bitmap_zero(used_nodes, MAX_NUMNODES);
5546 nodemask = node_to_cpumask(node);
5547 cpus_or(span, span, nodemask);
5548 set_bit(node, used_nodes);
5550 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5551 int next_node = find_next_best_node(node, used_nodes);
5552 nodemask = node_to_cpumask(next_node);
5553 cpus_or(span, span, nodemask);
5556 return span;
5558 #endif
5561 * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we
5562 * can switch it on easily if needed.
5564 #ifdef CONFIG_SCHED_SMT
5565 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5566 static struct sched_group sched_group_cpus[NR_CPUS];
5567 static int cpu_to_cpu_group(int cpu)
5569 return cpu;
5571 #endif
5573 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5574 static struct sched_group sched_group_phys[NR_CPUS];
5575 static int cpu_to_phys_group(int cpu)
5577 #ifdef CONFIG_SCHED_SMT
5578 return first_cpu(cpu_sibling_map[cpu]);
5579 #else
5580 return cpu;
5581 #endif
5584 #ifdef CONFIG_NUMA
5586 * The init_sched_build_groups can't handle what we want to do with node
5587 * groups, so roll our own. Now each node has its own list of groups which
5588 * gets dynamically allocated.
5590 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5591 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5593 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5594 static struct sched_group *sched_group_allnodes_bycpu[NR_CPUS];
5596 static int cpu_to_allnodes_group(int cpu)
5598 return cpu_to_node(cpu);
5600 #endif
5603 * Build sched domains for a given set of cpus and attach the sched domains
5604 * to the individual cpus
5606 void build_sched_domains(const cpumask_t *cpu_map)
5608 int i;
5609 #ifdef CONFIG_NUMA
5610 struct sched_group **sched_group_nodes = NULL;
5611 struct sched_group *sched_group_allnodes = NULL;
5614 * Allocate the per-node list of sched groups
5616 sched_group_nodes = kmalloc(sizeof(struct sched_group*)*MAX_NUMNODES,
5617 GFP_ATOMIC);
5618 if (!sched_group_nodes) {
5619 printk(KERN_WARNING "Can not alloc sched group node list\n");
5620 return;
5622 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
5623 #endif
5626 * Set up domains for cpus specified by the cpu_map.
5628 for_each_cpu_mask(i, *cpu_map) {
5629 int group;
5630 struct sched_domain *sd = NULL, *p;
5631 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
5633 cpus_and(nodemask, nodemask, *cpu_map);
5635 #ifdef CONFIG_NUMA
5636 if (cpus_weight(*cpu_map)
5637 > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
5638 if (!sched_group_allnodes) {
5639 sched_group_allnodes
5640 = kmalloc(sizeof(struct sched_group)
5641 * MAX_NUMNODES,
5642 GFP_KERNEL);
5643 if (!sched_group_allnodes) {
5644 printk(KERN_WARNING
5645 "Can not alloc allnodes sched group\n");
5646 break;
5648 sched_group_allnodes_bycpu[i]
5649 = sched_group_allnodes;
5651 sd = &per_cpu(allnodes_domains, i);
5652 *sd = SD_ALLNODES_INIT;
5653 sd->span = *cpu_map;
5654 group = cpu_to_allnodes_group(i);
5655 sd->groups = &sched_group_allnodes[group];
5656 p = sd;
5657 } else
5658 p = NULL;
5660 sd = &per_cpu(node_domains, i);
5661 *sd = SD_NODE_INIT;
5662 sd->span = sched_domain_node_span(cpu_to_node(i));
5663 sd->parent = p;
5664 cpus_and(sd->span, sd->span, *cpu_map);
5665 #endif
5667 p = sd;
5668 sd = &per_cpu(phys_domains, i);
5669 group = cpu_to_phys_group(i);
5670 *sd = SD_CPU_INIT;
5671 sd->span = nodemask;
5672 sd->parent = p;
5673 sd->groups = &sched_group_phys[group];
5675 #ifdef CONFIG_SCHED_SMT
5676 p = sd;
5677 sd = &per_cpu(cpu_domains, i);
5678 group = cpu_to_cpu_group(i);
5679 *sd = SD_SIBLING_INIT;
5680 sd->span = cpu_sibling_map[i];
5681 cpus_and(sd->span, sd->span, *cpu_map);
5682 sd->parent = p;
5683 sd->groups = &sched_group_cpus[group];
5684 #endif
5687 #ifdef CONFIG_SCHED_SMT
5688 /* Set up CPU (sibling) groups */
5689 for_each_cpu_mask(i, *cpu_map) {
5690 cpumask_t this_sibling_map = cpu_sibling_map[i];
5691 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
5692 if (i != first_cpu(this_sibling_map))
5693 continue;
5695 init_sched_build_groups(sched_group_cpus, this_sibling_map,
5696 &cpu_to_cpu_group);
5698 #endif
5700 /* Set up physical groups */
5701 for (i = 0; i < MAX_NUMNODES; i++) {
5702 cpumask_t nodemask = node_to_cpumask(i);
5704 cpus_and(nodemask, nodemask, *cpu_map);
5705 if (cpus_empty(nodemask))
5706 continue;
5708 init_sched_build_groups(sched_group_phys, nodemask,
5709 &cpu_to_phys_group);
5712 #ifdef CONFIG_NUMA
5713 /* Set up node groups */
5714 if (sched_group_allnodes)
5715 init_sched_build_groups(sched_group_allnodes, *cpu_map,
5716 &cpu_to_allnodes_group);
5718 for (i = 0; i < MAX_NUMNODES; i++) {
5719 /* Set up node groups */
5720 struct sched_group *sg, *prev;
5721 cpumask_t nodemask = node_to_cpumask(i);
5722 cpumask_t domainspan;
5723 cpumask_t covered = CPU_MASK_NONE;
5724 int j;
5726 cpus_and(nodemask, nodemask, *cpu_map);
5727 if (cpus_empty(nodemask)) {
5728 sched_group_nodes[i] = NULL;
5729 continue;
5732 domainspan = sched_domain_node_span(i);
5733 cpus_and(domainspan, domainspan, *cpu_map);
5735 sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL);
5736 sched_group_nodes[i] = sg;
5737 for_each_cpu_mask(j, nodemask) {
5738 struct sched_domain *sd;
5739 sd = &per_cpu(node_domains, j);
5740 sd->groups = sg;
5741 if (sd->groups == NULL) {
5742 /* Turn off balancing if we have no groups */
5743 sd->flags = 0;
5746 if (!sg) {
5747 printk(KERN_WARNING
5748 "Can not alloc domain group for node %d\n", i);
5749 continue;
5751 sg->cpu_power = 0;
5752 sg->cpumask = nodemask;
5753 cpus_or(covered, covered, nodemask);
5754 prev = sg;
5756 for (j = 0; j < MAX_NUMNODES; j++) {
5757 cpumask_t tmp, notcovered;
5758 int n = (i + j) % MAX_NUMNODES;
5760 cpus_complement(notcovered, covered);
5761 cpus_and(tmp, notcovered, *cpu_map);
5762 cpus_and(tmp, tmp, domainspan);
5763 if (cpus_empty(tmp))
5764 break;
5766 nodemask = node_to_cpumask(n);
5767 cpus_and(tmp, tmp, nodemask);
5768 if (cpus_empty(tmp))
5769 continue;
5771 sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL);
5772 if (!sg) {
5773 printk(KERN_WARNING
5774 "Can not alloc domain group for node %d\n", j);
5775 break;
5777 sg->cpu_power = 0;
5778 sg->cpumask = tmp;
5779 cpus_or(covered, covered, tmp);
5780 prev->next = sg;
5781 prev = sg;
5783 prev->next = sched_group_nodes[i];
5785 #endif
5787 /* Calculate CPU power for physical packages and nodes */
5788 for_each_cpu_mask(i, *cpu_map) {
5789 int power;
5790 struct sched_domain *sd;
5791 #ifdef CONFIG_SCHED_SMT
5792 sd = &per_cpu(cpu_domains, i);
5793 power = SCHED_LOAD_SCALE;
5794 sd->groups->cpu_power = power;
5795 #endif
5797 sd = &per_cpu(phys_domains, i);
5798 power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
5799 (cpus_weight(sd->groups->cpumask)-1) / 10;
5800 sd->groups->cpu_power = power;
5802 #ifdef CONFIG_NUMA
5803 sd = &per_cpu(allnodes_domains, i);
5804 if (sd->groups) {
5805 power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
5806 (cpus_weight(sd->groups->cpumask)-1) / 10;
5807 sd->groups->cpu_power = power;
5809 #endif
5812 #ifdef CONFIG_NUMA
5813 for (i = 0; i < MAX_NUMNODES; i++) {
5814 struct sched_group *sg = sched_group_nodes[i];
5815 int j;
5817 if (sg == NULL)
5818 continue;
5819 next_sg:
5820 for_each_cpu_mask(j, sg->cpumask) {
5821 struct sched_domain *sd;
5822 int power;
5824 sd = &per_cpu(phys_domains, j);
5825 if (j != first_cpu(sd->groups->cpumask)) {
5827 * Only add "power" once for each
5828 * physical package.
5830 continue;
5832 power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
5833 (cpus_weight(sd->groups->cpumask)-1) / 10;
5835 sg->cpu_power += power;
5837 sg = sg->next;
5838 if (sg != sched_group_nodes[i])
5839 goto next_sg;
5841 #endif
5843 /* Attach the domains */
5844 for_each_cpu_mask(i, *cpu_map) {
5845 struct sched_domain *sd;
5846 #ifdef CONFIG_SCHED_SMT
5847 sd = &per_cpu(cpu_domains, i);
5848 #else
5849 sd = &per_cpu(phys_domains, i);
5850 #endif
5851 cpu_attach_domain(sd, i);
5854 * Tune cache-hot values:
5856 calibrate_migration_costs(cpu_map);
5859 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
5861 static void arch_init_sched_domains(const cpumask_t *cpu_map)
5863 cpumask_t cpu_default_map;
5866 * Setup mask for cpus without special case scheduling requirements.
5867 * For now this just excludes isolated cpus, but could be used to
5868 * exclude other special cases in the future.
5870 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
5872 build_sched_domains(&cpu_default_map);
5875 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
5877 #ifdef CONFIG_NUMA
5878 int i;
5879 int cpu;
5881 for_each_cpu_mask(cpu, *cpu_map) {
5882 struct sched_group *sched_group_allnodes
5883 = sched_group_allnodes_bycpu[cpu];
5884 struct sched_group **sched_group_nodes
5885 = sched_group_nodes_bycpu[cpu];
5887 if (sched_group_allnodes) {
5888 kfree(sched_group_allnodes);
5889 sched_group_allnodes_bycpu[cpu] = NULL;
5892 if (!sched_group_nodes)
5893 continue;
5895 for (i = 0; i < MAX_NUMNODES; i++) {
5896 cpumask_t nodemask = node_to_cpumask(i);
5897 struct sched_group *oldsg, *sg = sched_group_nodes[i];
5899 cpus_and(nodemask, nodemask, *cpu_map);
5900 if (cpus_empty(nodemask))
5901 continue;
5903 if (sg == NULL)
5904 continue;
5905 sg = sg->next;
5906 next_sg:
5907 oldsg = sg;
5908 sg = sg->next;
5909 kfree(oldsg);
5910 if (oldsg != sched_group_nodes[i])
5911 goto next_sg;
5913 kfree(sched_group_nodes);
5914 sched_group_nodes_bycpu[cpu] = NULL;
5916 #endif
5920 * Detach sched domains from a group of cpus specified in cpu_map
5921 * These cpus will now be attached to the NULL domain
5923 static void detach_destroy_domains(const cpumask_t *cpu_map)
5925 int i;
5927 for_each_cpu_mask(i, *cpu_map)
5928 cpu_attach_domain(NULL, i);
5929 synchronize_sched();
5930 arch_destroy_sched_domains(cpu_map);
5934 * Partition sched domains as specified by the cpumasks below.
5935 * This attaches all cpus from the cpumasks to the NULL domain,
5936 * waits for a RCU quiescent period, recalculates sched
5937 * domain information and then attaches them back to the
5938 * correct sched domains
5939 * Call with hotplug lock held
5941 void partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
5943 cpumask_t change_map;
5945 cpus_and(*partition1, *partition1, cpu_online_map);
5946 cpus_and(*partition2, *partition2, cpu_online_map);
5947 cpus_or(change_map, *partition1, *partition2);
5949 /* Detach sched domains from all of the affected cpus */
5950 detach_destroy_domains(&change_map);
5951 if (!cpus_empty(*partition1))
5952 build_sched_domains(partition1);
5953 if (!cpus_empty(*partition2))
5954 build_sched_domains(partition2);
5957 #ifdef CONFIG_HOTPLUG_CPU
5959 * Force a reinitialization of the sched domains hierarchy. The domains
5960 * and groups cannot be updated in place without racing with the balancing
5961 * code, so we temporarily attach all running cpus to the NULL domain
5962 * which will prevent rebalancing while the sched domains are recalculated.
5964 static int update_sched_domains(struct notifier_block *nfb,
5965 unsigned long action, void *hcpu)
5967 switch (action) {
5968 case CPU_UP_PREPARE:
5969 case CPU_DOWN_PREPARE:
5970 detach_destroy_domains(&cpu_online_map);
5971 return NOTIFY_OK;
5973 case CPU_UP_CANCELED:
5974 case CPU_DOWN_FAILED:
5975 case CPU_ONLINE:
5976 case CPU_DEAD:
5978 * Fall through and re-initialise the domains.
5980 break;
5981 default:
5982 return NOTIFY_DONE;
5985 /* The hotplug lock is already held by cpu_up/cpu_down */
5986 arch_init_sched_domains(&cpu_online_map);
5988 return NOTIFY_OK;
5990 #endif
5992 void __init sched_init_smp(void)
5994 lock_cpu_hotplug();
5995 arch_init_sched_domains(&cpu_online_map);
5996 unlock_cpu_hotplug();
5997 /* XXX: Theoretical race here - CPU may be hotplugged now */
5998 hotcpu_notifier(update_sched_domains, 0);
6000 #else
6001 void __init sched_init_smp(void)
6004 #endif /* CONFIG_SMP */
6006 int in_sched_functions(unsigned long addr)
6008 /* Linker adds these: start and end of __sched functions */
6009 extern char __sched_text_start[], __sched_text_end[];
6010 return in_lock_functions(addr) ||
6011 (addr >= (unsigned long)__sched_text_start
6012 && addr < (unsigned long)__sched_text_end);
6015 void __init sched_init(void)
6017 runqueue_t *rq;
6018 int i, j, k;
6020 for_each_cpu(i) {
6021 prio_array_t *array;
6023 rq = cpu_rq(i);
6024 spin_lock_init(&rq->lock);
6025 rq->nr_running = 0;
6026 rq->active = rq->arrays;
6027 rq->expired = rq->arrays + 1;
6028 rq->best_expired_prio = MAX_PRIO;
6030 #ifdef CONFIG_SMP
6031 rq->sd = NULL;
6032 for (j = 1; j < 3; j++)
6033 rq->cpu_load[j] = 0;
6034 rq->active_balance = 0;
6035 rq->push_cpu = 0;
6036 rq->migration_thread = NULL;
6037 INIT_LIST_HEAD(&rq->migration_queue);
6038 #endif
6039 atomic_set(&rq->nr_iowait, 0);
6041 for (j = 0; j < 2; j++) {
6042 array = rq->arrays + j;
6043 for (k = 0; k < MAX_PRIO; k++) {
6044 INIT_LIST_HEAD(array->queue + k);
6045 __clear_bit(k, array->bitmap);
6047 // delimiter for bitsearch
6048 __set_bit(MAX_PRIO, array->bitmap);
6053 * The boot idle thread does lazy MMU switching as well:
6055 atomic_inc(&init_mm.mm_count);
6056 enter_lazy_tlb(&init_mm, current);
6059 * Make us the idle thread. Technically, schedule() should not be
6060 * called from this thread, however somewhere below it might be,
6061 * but because we are the idle thread, we just pick up running again
6062 * when this runqueue becomes "idle".
6064 init_idle(current, smp_processor_id());
6067 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6068 void __might_sleep(char *file, int line)
6070 #if defined(in_atomic)
6071 static unsigned long prev_jiffy; /* ratelimiting */
6073 if ((in_atomic() || irqs_disabled()) &&
6074 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6075 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6076 return;
6077 prev_jiffy = jiffies;
6078 printk(KERN_ERR "Debug: sleeping function called from invalid"
6079 " context at %s:%d\n", file, line);
6080 printk("in_atomic():%d, irqs_disabled():%d\n",
6081 in_atomic(), irqs_disabled());
6082 dump_stack();
6084 #endif
6086 EXPORT_SYMBOL(__might_sleep);
6087 #endif
6089 #ifdef CONFIG_MAGIC_SYSRQ
6090 void normalize_rt_tasks(void)
6092 struct task_struct *p;
6093 prio_array_t *array;
6094 unsigned long flags;
6095 runqueue_t *rq;
6097 read_lock_irq(&tasklist_lock);
6098 for_each_process (p) {
6099 if (!rt_task(p))
6100 continue;
6102 rq = task_rq_lock(p, &flags);
6104 array = p->array;
6105 if (array)
6106 deactivate_task(p, task_rq(p));
6107 __setscheduler(p, SCHED_NORMAL, 0);
6108 if (array) {
6109 __activate_task(p, task_rq(p));
6110 resched_task(rq->curr);
6113 task_rq_unlock(rq, &flags);
6115 read_unlock_irq(&tasklist_lock);
6118 #endif /* CONFIG_MAGIC_SYSRQ */
6120 #ifdef CONFIG_IA64
6122 * These functions are only useful for the IA64 MCA handling.
6124 * They can only be called when the whole system has been
6125 * stopped - every CPU needs to be quiescent, and no scheduling
6126 * activity can take place. Using them for anything else would
6127 * be a serious bug, and as a result, they aren't even visible
6128 * under any other configuration.
6132 * curr_task - return the current task for a given cpu.
6133 * @cpu: the processor in question.
6135 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6137 task_t *curr_task(int cpu)
6139 return cpu_curr(cpu);
6143 * set_curr_task - set the current task for a given cpu.
6144 * @cpu: the processor in question.
6145 * @p: the task pointer to set.
6147 * Description: This function must only be used when non-maskable interrupts
6148 * are serviced on a separate stack. It allows the architecture to switch the
6149 * notion of the current task on a cpu in a non-blocking manner. This function
6150 * must be called with all CPU's synchronized, and interrupts disabled, the
6151 * and caller must save the original value of the current task (see
6152 * curr_task() above) and restore that value before reenabling interrupts and
6153 * re-starting the system.
6155 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6157 void set_curr_task(int cpu, task_t *p)
6159 cpu_curr(cpu) = p;
6162 #endif