4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
22 #include <linux/module.h>
23 #include <linux/nmi.h>
24 #include <linux/init.h>
25 #include <asm/uaccess.h>
26 #include <linux/highmem.h>
27 #include <linux/smp_lock.h>
28 #include <asm/mmu_context.h>
29 #include <linux/interrupt.h>
30 #include <linux/completion.h>
31 #include <linux/kernel_stat.h>
32 #include <linux/security.h>
33 #include <linux/notifier.h>
34 #include <linux/profile.h>
35 #include <linux/suspend.h>
36 #include <linux/blkdev.h>
37 #include <linux/delay.h>
38 #include <linux/smp.h>
39 #include <linux/threads.h>
40 #include <linux/timer.h>
41 #include <linux/rcupdate.h>
42 #include <linux/cpu.h>
43 #include <linux/cpuset.h>
44 #include <linux/percpu.h>
45 #include <linux/kthread.h>
46 #include <linux/seq_file.h>
47 #include <linux/syscalls.h>
48 #include <linux/times.h>
49 #include <linux/acct.h>
52 #include <asm/unistd.h>
55 * Convert user-nice values [ -20 ... 0 ... 19 ]
56 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
59 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
60 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
61 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
64 * 'User priority' is the nice value converted to something we
65 * can work with better when scaling various scheduler parameters,
66 * it's a [ 0 ... 39 ] range.
68 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
69 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
70 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
73 * Some helpers for converting nanosecond timing to jiffy resolution
75 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
76 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
79 * These are the 'tuning knobs' of the scheduler:
81 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
82 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
83 * Timeslices get refilled after they expire.
85 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
86 #define DEF_TIMESLICE (100 * HZ / 1000)
87 #define ON_RUNQUEUE_WEIGHT 30
88 #define CHILD_PENALTY 95
89 #define PARENT_PENALTY 100
91 #define PRIO_BONUS_RATIO 25
92 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
93 #define INTERACTIVE_DELTA 2
94 #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
95 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
96 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
99 * If a task is 'interactive' then we reinsert it in the active
100 * array after it has expired its current timeslice. (it will not
101 * continue to run immediately, it will still roundrobin with
102 * other interactive tasks.)
104 * This part scales the interactivity limit depending on niceness.
106 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
107 * Here are a few examples of different nice levels:
109 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
110 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
111 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
112 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
113 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
115 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
116 * priority range a task can explore, a value of '1' means the
117 * task is rated interactive.)
119 * Ie. nice +19 tasks can never get 'interactive' enough to be
120 * reinserted into the active array. And only heavily CPU-hog nice -20
121 * tasks will be expired. Default nice 0 tasks are somewhere between,
122 * it takes some effort for them to get interactive, but it's not
126 #define CURRENT_BONUS(p) \
127 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
130 #define GRANULARITY (10 * HZ / 1000 ? : 1)
133 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
134 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
137 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
138 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
141 #define SCALE(v1,v1_max,v2_max) \
142 (v1) * (v2_max) / (v1_max)
145 (SCALE(TASK_NICE(p), 40, MAX_BONUS) + INTERACTIVE_DELTA)
147 #define TASK_INTERACTIVE(p) \
148 ((p)->prio <= (p)->static_prio - DELTA(p))
150 #define INTERACTIVE_SLEEP(p) \
151 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
152 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
154 #define TASK_PREEMPTS_CURR(p, rq) \
155 ((p)->prio < (rq)->curr->prio)
158 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
159 * to time slice values: [800ms ... 100ms ... 5ms]
161 * The higher a thread's priority, the bigger timeslices
162 * it gets during one round of execution. But even the lowest
163 * priority thread gets MIN_TIMESLICE worth of execution time.
166 #define SCALE_PRIO(x, prio) \
167 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO/2), MIN_TIMESLICE)
169 static inline unsigned int task_timeslice(task_t *p)
171 if (p->static_prio < NICE_TO_PRIO(0))
172 return SCALE_PRIO(DEF_TIMESLICE*4, p->static_prio);
174 return SCALE_PRIO(DEF_TIMESLICE, p->static_prio);
176 #define task_hot(p, now, sd) ((long long) ((now) - (p)->last_ran) \
177 < (long long) (sd)->cache_hot_time)
180 * These are the runqueue data structures:
183 #define BITMAP_SIZE ((((MAX_PRIO+1+7)/8)+sizeof(long)-1)/sizeof(long))
185 typedef struct runqueue runqueue_t;
188 unsigned int nr_active;
189 unsigned long bitmap[BITMAP_SIZE];
190 struct list_head queue[MAX_PRIO];
194 * This is the main, per-CPU runqueue data structure.
196 * Locking rule: those places that want to lock multiple runqueues
197 * (such as the load balancing or the thread migration code), lock
198 * acquire operations must be ordered by ascending &runqueue.
204 * nr_running and cpu_load should be in the same cacheline because
205 * remote CPUs use both these fields when doing load calculation.
207 unsigned long nr_running;
209 unsigned long cpu_load;
211 unsigned long long nr_switches;
214 * This is part of a global counter where only the total sum
215 * over all CPUs matters. A task can increase this counter on
216 * one CPU and if it got migrated afterwards it may decrease
217 * it on another CPU. Always updated under the runqueue lock:
219 unsigned long nr_uninterruptible;
221 unsigned long expired_timestamp;
222 unsigned long long timestamp_last_tick;
224 struct mm_struct *prev_mm;
225 prio_array_t *active, *expired, arrays[2];
226 int best_expired_prio;
230 struct sched_domain *sd;
232 /* For active balancing */
236 task_t *migration_thread;
237 struct list_head migration_queue;
240 #ifdef CONFIG_SCHEDSTATS
242 struct sched_info rq_sched_info;
244 /* sys_sched_yield() stats */
245 unsigned long yld_exp_empty;
246 unsigned long yld_act_empty;
247 unsigned long yld_both_empty;
248 unsigned long yld_cnt;
250 /* schedule() stats */
251 unsigned long sched_switch;
252 unsigned long sched_cnt;
253 unsigned long sched_goidle;
255 /* try_to_wake_up() stats */
256 unsigned long ttwu_cnt;
257 unsigned long ttwu_local;
261 static DEFINE_PER_CPU(struct runqueue, runqueues);
263 #define for_each_domain(cpu, domain) \
264 for (domain = cpu_rq(cpu)->sd; domain; domain = domain->parent)
266 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
267 #define this_rq() (&__get_cpu_var(runqueues))
268 #define task_rq(p) cpu_rq(task_cpu(p))
269 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
272 * Default context-switch locking:
274 #ifndef prepare_arch_switch
275 # define prepare_arch_switch(rq, next) do { } while (0)
276 # define finish_arch_switch(rq, next) spin_unlock_irq(&(rq)->lock)
277 # define task_running(rq, p) ((rq)->curr == (p))
281 * task_rq_lock - lock the runqueue a given task resides on and disable
282 * interrupts. Note the ordering: we can safely lookup the task_rq without
283 * explicitly disabling preemption.
285 static inline runqueue_t *task_rq_lock(task_t *p, unsigned long *flags)
291 local_irq_save(*flags);
293 spin_lock(&rq->lock);
294 if (unlikely(rq != task_rq(p))) {
295 spin_unlock_irqrestore(&rq->lock, *flags);
296 goto repeat_lock_task;
301 static inline void task_rq_unlock(runqueue_t *rq, unsigned long *flags)
304 spin_unlock_irqrestore(&rq->lock, *flags);
307 #ifdef CONFIG_SCHEDSTATS
309 * bump this up when changing the output format or the meaning of an existing
310 * format, so that tools can adapt (or abort)
312 #define SCHEDSTAT_VERSION 11
314 static int show_schedstat(struct seq_file *seq, void *v)
318 seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION);
319 seq_printf(seq, "timestamp %lu\n", jiffies);
320 for_each_online_cpu(cpu) {
321 runqueue_t *rq = cpu_rq(cpu);
323 struct sched_domain *sd;
327 /* runqueue-specific stats */
329 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
330 cpu, rq->yld_both_empty,
331 rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt,
332 rq->sched_switch, rq->sched_cnt, rq->sched_goidle,
333 rq->ttwu_cnt, rq->ttwu_local,
334 rq->rq_sched_info.cpu_time,
335 rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt);
337 seq_printf(seq, "\n");
340 /* domain-specific stats */
341 for_each_domain(cpu, sd) {
342 enum idle_type itype;
343 char mask_str[NR_CPUS];
345 cpumask_scnprintf(mask_str, NR_CPUS, sd->span);
346 seq_printf(seq, "domain%d %s", dcnt++, mask_str);
347 for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES;
349 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu",
351 sd->lb_balanced[itype],
352 sd->lb_failed[itype],
353 sd->lb_imbalance[itype],
354 sd->lb_gained[itype],
355 sd->lb_hot_gained[itype],
356 sd->lb_nobusyq[itype],
357 sd->lb_nobusyg[itype]);
359 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu\n",
360 sd->alb_cnt, sd->alb_failed, sd->alb_pushed,
361 sd->sbe_pushed, sd->sbe_attempts,
362 sd->ttwu_wake_remote, sd->ttwu_move_affine, sd->ttwu_move_balance);
369 static int schedstat_open(struct inode *inode, struct file *file)
371 unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32);
372 char *buf = kmalloc(size, GFP_KERNEL);
378 res = single_open(file, show_schedstat, NULL);
380 m = file->private_data;
388 struct file_operations proc_schedstat_operations = {
389 .open = schedstat_open,
392 .release = single_release,
395 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
396 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
397 #else /* !CONFIG_SCHEDSTATS */
398 # define schedstat_inc(rq, field) do { } while (0)
399 # define schedstat_add(rq, field, amt) do { } while (0)
403 * rq_lock - lock a given runqueue and disable interrupts.
405 static inline runqueue_t *this_rq_lock(void)
412 spin_lock(&rq->lock);
417 #ifdef CONFIG_SCHED_SMT
418 static int cpu_and_siblings_are_idle(int cpu)
421 for_each_cpu_mask(sib, cpu_sibling_map[cpu]) {
430 #define cpu_and_siblings_are_idle(A) idle_cpu(A)
433 #ifdef CONFIG_SCHEDSTATS
435 * Called when a process is dequeued from the active array and given
436 * the cpu. We should note that with the exception of interactive
437 * tasks, the expired queue will become the active queue after the active
438 * queue is empty, without explicitly dequeuing and requeuing tasks in the
439 * expired queue. (Interactive tasks may be requeued directly to the
440 * active queue, thus delaying tasks in the expired queue from running;
441 * see scheduler_tick()).
443 * This function is only called from sched_info_arrive(), rather than
444 * dequeue_task(). Even though a task may be queued and dequeued multiple
445 * times as it is shuffled about, we're really interested in knowing how
446 * long it was from the *first* time it was queued to the time that it
449 static inline void sched_info_dequeued(task_t *t)
451 t->sched_info.last_queued = 0;
455 * Called when a task finally hits the cpu. We can now calculate how
456 * long it was waiting to run. We also note when it began so that we
457 * can keep stats on how long its timeslice is.
459 static inline void sched_info_arrive(task_t *t)
461 unsigned long now = jiffies, diff = 0;
462 struct runqueue *rq = task_rq(t);
464 if (t->sched_info.last_queued)
465 diff = now - t->sched_info.last_queued;
466 sched_info_dequeued(t);
467 t->sched_info.run_delay += diff;
468 t->sched_info.last_arrival = now;
469 t->sched_info.pcnt++;
474 rq->rq_sched_info.run_delay += diff;
475 rq->rq_sched_info.pcnt++;
479 * Called when a process is queued into either the active or expired
480 * array. The time is noted and later used to determine how long we
481 * had to wait for us to reach the cpu. Since the expired queue will
482 * become the active queue after active queue is empty, without dequeuing
483 * and requeuing any tasks, we are interested in queuing to either. It
484 * is unusual but not impossible for tasks to be dequeued and immediately
485 * requeued in the same or another array: this can happen in sched_yield(),
486 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
489 * This function is only called from enqueue_task(), but also only updates
490 * the timestamp if it is already not set. It's assumed that
491 * sched_info_dequeued() will clear that stamp when appropriate.
493 static inline void sched_info_queued(task_t *t)
495 if (!t->sched_info.last_queued)
496 t->sched_info.last_queued = jiffies;
500 * Called when a process ceases being the active-running process, either
501 * voluntarily or involuntarily. Now we can calculate how long we ran.
503 static inline void sched_info_depart(task_t *t)
505 struct runqueue *rq = task_rq(t);
506 unsigned long diff = jiffies - t->sched_info.last_arrival;
508 t->sched_info.cpu_time += diff;
511 rq->rq_sched_info.cpu_time += diff;
515 * Called when tasks are switched involuntarily due, typically, to expiring
516 * their time slice. (This may also be called when switching to or from
517 * the idle task.) We are only called when prev != next.
519 static inline void sched_info_switch(task_t *prev, task_t *next)
521 struct runqueue *rq = task_rq(prev);
524 * prev now departs the cpu. It's not interesting to record
525 * stats about how efficient we were at scheduling the idle
528 if (prev != rq->idle)
529 sched_info_depart(prev);
531 if (next != rq->idle)
532 sched_info_arrive(next);
535 #define sched_info_queued(t) do { } while (0)
536 #define sched_info_switch(t, next) do { } while (0)
537 #endif /* CONFIG_SCHEDSTATS */
540 * Adding/removing a task to/from a priority array:
542 static void dequeue_task(struct task_struct *p, prio_array_t *array)
545 list_del(&p->run_list);
546 if (list_empty(array->queue + p->prio))
547 __clear_bit(p->prio, array->bitmap);
550 static void enqueue_task(struct task_struct *p, prio_array_t *array)
552 sched_info_queued(p);
553 list_add_tail(&p->run_list, array->queue + p->prio);
554 __set_bit(p->prio, array->bitmap);
560 * Put task to the end of the run list without the overhead of dequeue
561 * followed by enqueue.
563 static void requeue_task(struct task_struct *p, prio_array_t *array)
565 list_move_tail(&p->run_list, array->queue + p->prio);
568 static inline void enqueue_task_head(struct task_struct *p, prio_array_t *array)
570 list_add(&p->run_list, array->queue + p->prio);
571 __set_bit(p->prio, array->bitmap);
577 * effective_prio - return the priority that is based on the static
578 * priority but is modified by bonuses/penalties.
580 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
581 * into the -5 ... 0 ... +5 bonus/penalty range.
583 * We use 25% of the full 0...39 priority range so that:
585 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
586 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
588 * Both properties are important to certain workloads.
590 static int effective_prio(task_t *p)
597 bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
599 prio = p->static_prio - bonus;
600 if (prio < MAX_RT_PRIO)
602 if (prio > MAX_PRIO-1)
608 * __activate_task - move a task to the runqueue.
610 static inline void __activate_task(task_t *p, runqueue_t *rq)
612 enqueue_task(p, rq->active);
617 * __activate_idle_task - move idle task to the _front_ of runqueue.
619 static inline void __activate_idle_task(task_t *p, runqueue_t *rq)
621 enqueue_task_head(p, rq->active);
625 static void recalc_task_prio(task_t *p, unsigned long long now)
627 /* Caller must always ensure 'now >= p->timestamp' */
628 unsigned long long __sleep_time = now - p->timestamp;
629 unsigned long sleep_time;
631 if (__sleep_time > NS_MAX_SLEEP_AVG)
632 sleep_time = NS_MAX_SLEEP_AVG;
634 sleep_time = (unsigned long)__sleep_time;
636 if (likely(sleep_time > 0)) {
638 * User tasks that sleep a long time are categorised as
639 * idle and will get just interactive status to stay active &
640 * prevent them suddenly becoming cpu hogs and starving
643 if (p->mm && p->activated != -1 &&
644 sleep_time > INTERACTIVE_SLEEP(p)) {
645 p->sleep_avg = JIFFIES_TO_NS(MAX_SLEEP_AVG -
649 * The lower the sleep avg a task has the more
650 * rapidly it will rise with sleep time.
652 sleep_time *= (MAX_BONUS - CURRENT_BONUS(p)) ? : 1;
655 * Tasks waking from uninterruptible sleep are
656 * limited in their sleep_avg rise as they
657 * are likely to be waiting on I/O
659 if (p->activated == -1 && p->mm) {
660 if (p->sleep_avg >= INTERACTIVE_SLEEP(p))
662 else if (p->sleep_avg + sleep_time >=
663 INTERACTIVE_SLEEP(p)) {
664 p->sleep_avg = INTERACTIVE_SLEEP(p);
670 * This code gives a bonus to interactive tasks.
672 * The boost works by updating the 'average sleep time'
673 * value here, based on ->timestamp. The more time a
674 * task spends sleeping, the higher the average gets -
675 * and the higher the priority boost gets as well.
677 p->sleep_avg += sleep_time;
679 if (p->sleep_avg > NS_MAX_SLEEP_AVG)
680 p->sleep_avg = NS_MAX_SLEEP_AVG;
684 p->prio = effective_prio(p);
688 * activate_task - move a task to the runqueue and do priority recalculation
690 * Update all the scheduling statistics stuff. (sleep average
691 * calculation, priority modifiers, etc.)
693 static void activate_task(task_t *p, runqueue_t *rq, int local)
695 unsigned long long now;
700 /* Compensate for drifting sched_clock */
701 runqueue_t *this_rq = this_rq();
702 now = (now - this_rq->timestamp_last_tick)
703 + rq->timestamp_last_tick;
707 recalc_task_prio(p, now);
710 * This checks to make sure it's not an uninterruptible task
711 * that is now waking up.
715 * Tasks which were woken up by interrupts (ie. hw events)
716 * are most likely of interactive nature. So we give them
717 * the credit of extending their sleep time to the period
718 * of time they spend on the runqueue, waiting for execution
719 * on a CPU, first time around:
725 * Normal first-time wakeups get a credit too for
726 * on-runqueue time, but it will be weighted down:
733 __activate_task(p, rq);
737 * deactivate_task - remove a task from the runqueue.
739 static void deactivate_task(struct task_struct *p, runqueue_t *rq)
742 dequeue_task(p, p->array);
747 * resched_task - mark a task 'to be rescheduled now'.
749 * On UP this means the setting of the need_resched flag, on SMP it
750 * might also involve a cross-CPU call to trigger the scheduler on
754 static void resched_task(task_t *p)
756 int need_resched, nrpolling;
758 assert_spin_locked(&task_rq(p)->lock);
760 /* minimise the chance of sending an interrupt to poll_idle() */
761 nrpolling = test_tsk_thread_flag(p,TIF_POLLING_NRFLAG);
762 need_resched = test_and_set_tsk_thread_flag(p,TIF_NEED_RESCHED);
763 nrpolling |= test_tsk_thread_flag(p,TIF_POLLING_NRFLAG);
765 if (!need_resched && !nrpolling && (task_cpu(p) != smp_processor_id()))
766 smp_send_reschedule(task_cpu(p));
769 static inline void resched_task(task_t *p)
771 set_tsk_need_resched(p);
776 * task_curr - is this task currently executing on a CPU?
777 * @p: the task in question.
779 inline int task_curr(const task_t *p)
781 return cpu_curr(task_cpu(p)) == p;
791 struct list_head list;
792 enum request_type type;
794 /* For REQ_MOVE_TASK */
798 /* For REQ_SET_DOMAIN */
799 struct sched_domain *sd;
801 struct completion done;
805 * The task's runqueue lock must be held.
806 * Returns true if you have to wait for migration thread.
808 static int migrate_task(task_t *p, int dest_cpu, migration_req_t *req)
810 runqueue_t *rq = task_rq(p);
813 * If the task is not on a runqueue (and not running), then
814 * it is sufficient to simply update the task's cpu field.
816 if (!p->array && !task_running(rq, p)) {
817 set_task_cpu(p, dest_cpu);
821 init_completion(&req->done);
822 req->type = REQ_MOVE_TASK;
824 req->dest_cpu = dest_cpu;
825 list_add(&req->list, &rq->migration_queue);
830 * wait_task_inactive - wait for a thread to unschedule.
832 * The caller must ensure that the task *will* unschedule sometime soon,
833 * else this function might spin for a *long* time. This function can't
834 * be called with interrupts off, or it may introduce deadlock with
835 * smp_call_function() if an IPI is sent by the same process we are
836 * waiting to become inactive.
838 void wait_task_inactive(task_t * p)
845 rq = task_rq_lock(p, &flags);
846 /* Must be off runqueue entirely, not preempted. */
847 if (unlikely(p->array || task_running(rq, p))) {
848 /* If it's preempted, we yield. It could be a while. */
849 preempted = !task_running(rq, p);
850 task_rq_unlock(rq, &flags);
856 task_rq_unlock(rq, &flags);
860 * kick_process - kick a running thread to enter/exit the kernel
861 * @p: the to-be-kicked thread
863 * Cause a process which is running on another CPU to enter
864 * kernel-mode, without any delay. (to get signals handled.)
866 * NOTE: this function doesnt have to take the runqueue lock,
867 * because all it wants to ensure is that the remote task enters
868 * the kernel. If the IPI races and the task has been migrated
869 * to another CPU then no harm is done and the purpose has been
872 void kick_process(task_t *p)
878 if ((cpu != smp_processor_id()) && task_curr(p))
879 smp_send_reschedule(cpu);
884 * Return a low guess at the load of a migration-source cpu.
886 * We want to under-estimate the load of migration sources, to
887 * balance conservatively.
889 static inline unsigned long source_load(int cpu)
891 runqueue_t *rq = cpu_rq(cpu);
892 unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE;
894 return min(rq->cpu_load, load_now);
898 * Return a high guess at the load of a migration-target cpu
900 static inline unsigned long target_load(int cpu)
902 runqueue_t *rq = cpu_rq(cpu);
903 unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE;
905 return max(rq->cpu_load, load_now);
911 * wake_idle() will wake a task on an idle cpu if task->cpu is
912 * not idle and an idle cpu is available. The span of cpus to
913 * search starts with cpus closest then further out as needed,
914 * so we always favor a closer, idle cpu.
916 * Returns the CPU we should wake onto.
918 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
919 static int wake_idle(int cpu, task_t *p)
922 struct sched_domain *sd;
928 for_each_domain(cpu, sd) {
929 if (sd->flags & SD_WAKE_IDLE) {
930 cpus_and(tmp, sd->span, p->cpus_allowed);
931 for_each_cpu_mask(i, tmp) {
942 static inline int wake_idle(int cpu, task_t *p)
949 * try_to_wake_up - wake up a thread
950 * @p: the to-be-woken-up thread
951 * @state: the mask of task states that can be woken
952 * @sync: do a synchronous wakeup?
954 * Put it on the run-queue if it's not already there. The "current"
955 * thread is always on the run-queue (except when the actual
956 * re-schedule is in progress), and as such you're allowed to do
957 * the simpler "current->state = TASK_RUNNING" to mark yourself
958 * runnable without the overhead of this.
960 * returns failure only if the task is already active.
962 static int try_to_wake_up(task_t * p, unsigned int state, int sync)
964 int cpu, this_cpu, success = 0;
969 unsigned long load, this_load;
970 struct sched_domain *sd;
974 rq = task_rq_lock(p, &flags);
975 old_state = p->state;
976 if (!(old_state & state))
983 this_cpu = smp_processor_id();
986 if (unlikely(task_running(rq, p)))
989 #ifdef CONFIG_SCHEDSTATS
990 schedstat_inc(rq, ttwu_cnt);
991 if (cpu == this_cpu) {
992 schedstat_inc(rq, ttwu_local);
994 for_each_domain(this_cpu, sd) {
995 if (cpu_isset(cpu, sd->span)) {
996 schedstat_inc(sd, ttwu_wake_remote);
1004 if (cpu == this_cpu || unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1007 load = source_load(cpu);
1008 this_load = target_load(this_cpu);
1011 * If sync wakeup then subtract the (maximum possible) effect of
1012 * the currently running task from the load of the current CPU:
1015 this_load -= SCHED_LOAD_SCALE;
1017 /* Don't pull the task off an idle CPU to a busy one */
1018 if (load < SCHED_LOAD_SCALE/2 && this_load > SCHED_LOAD_SCALE/2)
1021 new_cpu = this_cpu; /* Wake to this CPU if we can */
1024 * Scan domains for affine wakeup and passive balancing
1027 for_each_domain(this_cpu, sd) {
1028 unsigned int imbalance;
1030 * Start passive balancing when half the imbalance_pct
1033 imbalance = sd->imbalance_pct + (sd->imbalance_pct - 100) / 2;
1035 if ((sd->flags & SD_WAKE_AFFINE) &&
1036 !task_hot(p, rq->timestamp_last_tick, sd)) {
1038 * This domain has SD_WAKE_AFFINE and p is cache cold
1041 if (cpu_isset(cpu, sd->span)) {
1042 schedstat_inc(sd, ttwu_move_affine);
1045 } else if ((sd->flags & SD_WAKE_BALANCE) &&
1046 imbalance*this_load <= 100*load) {
1048 * This domain has SD_WAKE_BALANCE and there is
1051 if (cpu_isset(cpu, sd->span)) {
1052 schedstat_inc(sd, ttwu_move_balance);
1058 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1060 new_cpu = wake_idle(new_cpu, p);
1061 if (new_cpu != cpu) {
1062 set_task_cpu(p, new_cpu);
1063 task_rq_unlock(rq, &flags);
1064 /* might preempt at this point */
1065 rq = task_rq_lock(p, &flags);
1066 old_state = p->state;
1067 if (!(old_state & state))
1072 this_cpu = smp_processor_id();
1077 #endif /* CONFIG_SMP */
1078 if (old_state == TASK_UNINTERRUPTIBLE) {
1079 rq->nr_uninterruptible--;
1081 * Tasks on involuntary sleep don't earn
1082 * sleep_avg beyond just interactive state.
1088 * Sync wakeups (i.e. those types of wakeups where the waker
1089 * has indicated that it will leave the CPU in short order)
1090 * don't trigger a preemption, if the woken up task will run on
1091 * this cpu. (in this case the 'I will reschedule' promise of
1092 * the waker guarantees that the freshly woken up task is going
1093 * to be considered on this CPU.)
1095 activate_task(p, rq, cpu == this_cpu);
1096 if (!sync || cpu != this_cpu) {
1097 if (TASK_PREEMPTS_CURR(p, rq))
1098 resched_task(rq->curr);
1103 p->state = TASK_RUNNING;
1105 task_rq_unlock(rq, &flags);
1110 int fastcall wake_up_process(task_t * p)
1112 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1113 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1116 EXPORT_SYMBOL(wake_up_process);
1118 int fastcall wake_up_state(task_t *p, unsigned int state)
1120 return try_to_wake_up(p, state, 0);
1124 static int find_idlest_cpu(struct task_struct *p, int this_cpu,
1125 struct sched_domain *sd);
1129 * Perform scheduler related setup for a newly forked process p.
1130 * p is forked by current.
1132 void fastcall sched_fork(task_t *p)
1135 * We mark the process as running here, but have not actually
1136 * inserted it onto the runqueue yet. This guarantees that
1137 * nobody will actually run it, and a signal or other external
1138 * event cannot wake it up and insert it on the runqueue either.
1140 p->state = TASK_RUNNING;
1141 INIT_LIST_HEAD(&p->run_list);
1143 spin_lock_init(&p->switch_lock);
1144 #ifdef CONFIG_SCHEDSTATS
1145 memset(&p->sched_info, 0, sizeof(p->sched_info));
1147 #ifdef CONFIG_PREEMPT
1149 * During context-switch we hold precisely one spinlock, which
1150 * schedule_tail drops. (in the common case it's this_rq()->lock,
1151 * but it also can be p->switch_lock.) So we compensate with a count
1152 * of 1. Also, we want to start with kernel preemption disabled.
1154 p->thread_info->preempt_count = 1;
1157 * Share the timeslice between parent and child, thus the
1158 * total amount of pending timeslices in the system doesn't change,
1159 * resulting in more scheduling fairness.
1161 local_irq_disable();
1162 p->time_slice = (current->time_slice + 1) >> 1;
1164 * The remainder of the first timeslice might be recovered by
1165 * the parent if the child exits early enough.
1167 p->first_time_slice = 1;
1168 current->time_slice >>= 1;
1169 p->timestamp = sched_clock();
1170 if (unlikely(!current->time_slice)) {
1172 * This case is rare, it happens when the parent has only
1173 * a single jiffy left from its timeslice. Taking the
1174 * runqueue lock is not a problem.
1176 current->time_slice = 1;
1186 * wake_up_new_task - wake up a newly created task for the first time.
1188 * This function will do some initial scheduler statistics housekeeping
1189 * that must be done for every newly created context, then puts the task
1190 * on the runqueue and wakes it.
1192 void fastcall wake_up_new_task(task_t * p, unsigned long clone_flags)
1194 unsigned long flags;
1196 runqueue_t *rq, *this_rq;
1198 rq = task_rq_lock(p, &flags);
1200 this_cpu = smp_processor_id();
1202 BUG_ON(p->state != TASK_RUNNING);
1205 * We decrease the sleep average of forking parents
1206 * and children as well, to keep max-interactive tasks
1207 * from forking tasks that are max-interactive. The parent
1208 * (current) is done further down, under its lock.
1210 p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1211 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1213 p->prio = effective_prio(p);
1215 if (likely(cpu == this_cpu)) {
1216 if (!(clone_flags & CLONE_VM)) {
1218 * The VM isn't cloned, so we're in a good position to
1219 * do child-runs-first in anticipation of an exec. This
1220 * usually avoids a lot of COW overhead.
1222 if (unlikely(!current->array))
1223 __activate_task(p, rq);
1225 p->prio = current->prio;
1226 list_add_tail(&p->run_list, ¤t->run_list);
1227 p->array = current->array;
1228 p->array->nr_active++;
1233 /* Run child last */
1234 __activate_task(p, rq);
1236 * We skip the following code due to cpu == this_cpu
1238 * task_rq_unlock(rq, &flags);
1239 * this_rq = task_rq_lock(current, &flags);
1243 this_rq = cpu_rq(this_cpu);
1246 * Not the local CPU - must adjust timestamp. This should
1247 * get optimised away in the !CONFIG_SMP case.
1249 p->timestamp = (p->timestamp - this_rq->timestamp_last_tick)
1250 + rq->timestamp_last_tick;
1251 __activate_task(p, rq);
1252 if (TASK_PREEMPTS_CURR(p, rq))
1253 resched_task(rq->curr);
1256 * Parent and child are on different CPUs, now get the
1257 * parent runqueue to update the parent's ->sleep_avg:
1259 task_rq_unlock(rq, &flags);
1260 this_rq = task_rq_lock(current, &flags);
1262 current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1263 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1264 task_rq_unlock(this_rq, &flags);
1268 * Potentially available exiting-child timeslices are
1269 * retrieved here - this way the parent does not get
1270 * penalized for creating too many threads.
1272 * (this cannot be used to 'generate' timeslices
1273 * artificially, because any timeslice recovered here
1274 * was given away by the parent in the first place.)
1276 void fastcall sched_exit(task_t * p)
1278 unsigned long flags;
1282 * If the child was a (relative-) CPU hog then decrease
1283 * the sleep_avg of the parent as well.
1285 rq = task_rq_lock(p->parent, &flags);
1286 if (p->first_time_slice) {
1287 p->parent->time_slice += p->time_slice;
1288 if (unlikely(p->parent->time_slice > task_timeslice(p)))
1289 p->parent->time_slice = task_timeslice(p);
1291 if (p->sleep_avg < p->parent->sleep_avg)
1292 p->parent->sleep_avg = p->parent->sleep_avg /
1293 (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg /
1295 task_rq_unlock(rq, &flags);
1299 * finish_task_switch - clean up after a task-switch
1300 * @prev: the thread we just switched away from.
1302 * We enter this with the runqueue still locked, and finish_arch_switch()
1303 * will unlock it along with doing any other architecture-specific cleanup
1306 * Note that we may have delayed dropping an mm in context_switch(). If
1307 * so, we finish that here outside of the runqueue lock. (Doing it
1308 * with the lock held can cause deadlocks; see schedule() for
1311 static inline void finish_task_switch(task_t *prev)
1312 __releases(rq->lock)
1314 runqueue_t *rq = this_rq();
1315 struct mm_struct *mm = rq->prev_mm;
1316 unsigned long prev_task_flags;
1321 * A task struct has one reference for the use as "current".
1322 * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1323 * calls schedule one last time. The schedule call will never return,
1324 * and the scheduled task must drop that reference.
1325 * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1326 * still held, otherwise prev could be scheduled on another cpu, die
1327 * there before we look at prev->state, and then the reference would
1329 * Manfred Spraul <manfred@colorfullife.com>
1331 prev_task_flags = prev->flags;
1332 finish_arch_switch(rq, prev);
1335 if (unlikely(prev_task_flags & PF_DEAD))
1336 put_task_struct(prev);
1340 * schedule_tail - first thing a freshly forked thread must call.
1341 * @prev: the thread we just switched away from.
1343 asmlinkage void schedule_tail(task_t *prev)
1344 __releases(rq->lock)
1346 finish_task_switch(prev);
1348 if (current->set_child_tid)
1349 put_user(current->pid, current->set_child_tid);
1353 * context_switch - switch to the new MM and the new
1354 * thread's register state.
1357 task_t * context_switch(runqueue_t *rq, task_t *prev, task_t *next)
1359 struct mm_struct *mm = next->mm;
1360 struct mm_struct *oldmm = prev->active_mm;
1362 if (unlikely(!mm)) {
1363 next->active_mm = oldmm;
1364 atomic_inc(&oldmm->mm_count);
1365 enter_lazy_tlb(oldmm, next);
1367 switch_mm(oldmm, mm, next);
1369 if (unlikely(!prev->mm)) {
1370 prev->active_mm = NULL;
1371 WARN_ON(rq->prev_mm);
1372 rq->prev_mm = oldmm;
1375 /* Here we just switch the register state and the stack. */
1376 switch_to(prev, next, prev);
1382 * nr_running, nr_uninterruptible and nr_context_switches:
1384 * externally visible scheduler statistics: current number of runnable
1385 * threads, current number of uninterruptible-sleeping threads, total
1386 * number of context switches performed since bootup.
1388 unsigned long nr_running(void)
1390 unsigned long i, sum = 0;
1392 for_each_online_cpu(i)
1393 sum += cpu_rq(i)->nr_running;
1398 unsigned long nr_uninterruptible(void)
1400 unsigned long i, sum = 0;
1403 sum += cpu_rq(i)->nr_uninterruptible;
1406 * Since we read the counters lockless, it might be slightly
1407 * inaccurate. Do not allow it to go below zero though:
1409 if (unlikely((long)sum < 0))
1415 unsigned long long nr_context_switches(void)
1417 unsigned long long i, sum = 0;
1420 sum += cpu_rq(i)->nr_switches;
1425 unsigned long nr_iowait(void)
1427 unsigned long i, sum = 0;
1430 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1438 * double_rq_lock - safely lock two runqueues
1440 * Note this does not disable interrupts like task_rq_lock,
1441 * you need to do so manually before calling.
1443 static void double_rq_lock(runqueue_t *rq1, runqueue_t *rq2)
1444 __acquires(rq1->lock)
1445 __acquires(rq2->lock)
1448 spin_lock(&rq1->lock);
1449 __acquire(rq2->lock); /* Fake it out ;) */
1452 spin_lock(&rq1->lock);
1453 spin_lock(&rq2->lock);
1455 spin_lock(&rq2->lock);
1456 spin_lock(&rq1->lock);
1462 * double_rq_unlock - safely unlock two runqueues
1464 * Note this does not restore interrupts like task_rq_unlock,
1465 * you need to do so manually after calling.
1467 static void double_rq_unlock(runqueue_t *rq1, runqueue_t *rq2)
1468 __releases(rq1->lock)
1469 __releases(rq2->lock)
1471 spin_unlock(&rq1->lock);
1473 spin_unlock(&rq2->lock);
1475 __release(rq2->lock);
1479 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1481 static void double_lock_balance(runqueue_t *this_rq, runqueue_t *busiest)
1482 __releases(this_rq->lock)
1483 __acquires(busiest->lock)
1484 __acquires(this_rq->lock)
1486 if (unlikely(!spin_trylock(&busiest->lock))) {
1487 if (busiest < this_rq) {
1488 spin_unlock(&this_rq->lock);
1489 spin_lock(&busiest->lock);
1490 spin_lock(&this_rq->lock);
1492 spin_lock(&busiest->lock);
1497 * find_idlest_cpu - find the least busy runqueue.
1499 static int find_idlest_cpu(struct task_struct *p, int this_cpu,
1500 struct sched_domain *sd)
1502 unsigned long load, min_load, this_load;
1507 min_load = ULONG_MAX;
1509 cpus_and(mask, sd->span, p->cpus_allowed);
1511 for_each_cpu_mask(i, mask) {
1512 load = target_load(i);
1514 if (load < min_load) {
1518 /* break out early on an idle CPU: */
1524 /* add +1 to account for the new task */
1525 this_load = source_load(this_cpu) + SCHED_LOAD_SCALE;
1528 * Would with the addition of the new task to the
1529 * current CPU there be an imbalance between this
1530 * CPU and the idlest CPU?
1532 * Use half of the balancing threshold - new-context is
1533 * a good opportunity to balance.
1535 if (min_load*(100 + (sd->imbalance_pct-100)/2) < this_load*100)
1542 * If dest_cpu is allowed for this process, migrate the task to it.
1543 * This is accomplished by forcing the cpu_allowed mask to only
1544 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
1545 * the cpu_allowed mask is restored.
1547 static void sched_migrate_task(task_t *p, int dest_cpu)
1549 migration_req_t req;
1551 unsigned long flags;
1553 rq = task_rq_lock(p, &flags);
1554 if (!cpu_isset(dest_cpu, p->cpus_allowed)
1555 || unlikely(cpu_is_offline(dest_cpu)))
1558 /* force the process onto the specified CPU */
1559 if (migrate_task(p, dest_cpu, &req)) {
1560 /* Need to wait for migration thread (might exit: take ref). */
1561 struct task_struct *mt = rq->migration_thread;
1562 get_task_struct(mt);
1563 task_rq_unlock(rq, &flags);
1564 wake_up_process(mt);
1565 put_task_struct(mt);
1566 wait_for_completion(&req.done);
1570 task_rq_unlock(rq, &flags);
1574 * sched_exec(): find the highest-level, exec-balance-capable
1575 * domain and try to migrate the task to the least loaded CPU.
1577 * execve() is a valuable balancing opportunity, because at this point
1578 * the task has the smallest effective memory and cache footprint.
1580 void sched_exec(void)
1582 struct sched_domain *tmp, *sd = NULL;
1583 int new_cpu, this_cpu = get_cpu();
1585 /* Prefer the current CPU if there's only this task running */
1586 if (this_rq()->nr_running <= 1)
1589 for_each_domain(this_cpu, tmp)
1590 if (tmp->flags & SD_BALANCE_EXEC)
1594 schedstat_inc(sd, sbe_attempts);
1595 new_cpu = find_idlest_cpu(current, this_cpu, sd);
1596 if (new_cpu != this_cpu) {
1597 schedstat_inc(sd, sbe_pushed);
1599 sched_migrate_task(current, new_cpu);
1608 * pull_task - move a task from a remote runqueue to the local runqueue.
1609 * Both runqueues must be locked.
1612 void pull_task(runqueue_t *src_rq, prio_array_t *src_array, task_t *p,
1613 runqueue_t *this_rq, prio_array_t *this_array, int this_cpu)
1615 dequeue_task(p, src_array);
1616 src_rq->nr_running--;
1617 set_task_cpu(p, this_cpu);
1618 this_rq->nr_running++;
1619 enqueue_task(p, this_array);
1620 p->timestamp = (p->timestamp - src_rq->timestamp_last_tick)
1621 + this_rq->timestamp_last_tick;
1623 * Note that idle threads have a prio of MAX_PRIO, for this test
1624 * to be always true for them.
1626 if (TASK_PREEMPTS_CURR(p, this_rq))
1627 resched_task(this_rq->curr);
1631 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1634 int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu,
1635 struct sched_domain *sd, enum idle_type idle, int *all_pinned)
1638 * We do not migrate tasks that are:
1639 * 1) running (obviously), or
1640 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1641 * 3) are cache-hot on their current CPU.
1643 if (!cpu_isset(this_cpu, p->cpus_allowed))
1647 if (task_running(rq, p))
1651 * Aggressive migration if:
1652 * 1) the [whole] cpu is idle, or
1653 * 2) too many balance attempts have failed.
1656 if (cpu_and_siblings_are_idle(this_cpu) || \
1657 sd->nr_balance_failed > sd->cache_nice_tries)
1660 if (task_hot(p, rq->timestamp_last_tick, sd))
1666 * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq,
1667 * as part of a balancing operation within "domain". Returns the number of
1670 * Called with both runqueues locked.
1672 static int move_tasks(runqueue_t *this_rq, int this_cpu, runqueue_t *busiest,
1673 unsigned long max_nr_move, struct sched_domain *sd,
1674 enum idle_type idle, int *all_pinned)
1676 prio_array_t *array, *dst_array;
1677 struct list_head *head, *curr;
1678 int idx, pulled = 0, pinned = 0;
1681 if (max_nr_move == 0)
1687 * We first consider expired tasks. Those will likely not be
1688 * executed in the near future, and they are most likely to
1689 * be cache-cold, thus switching CPUs has the least effect
1692 if (busiest->expired->nr_active) {
1693 array = busiest->expired;
1694 dst_array = this_rq->expired;
1696 array = busiest->active;
1697 dst_array = this_rq->active;
1701 /* Start searching at priority 0: */
1705 idx = sched_find_first_bit(array->bitmap);
1707 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
1708 if (idx >= MAX_PRIO) {
1709 if (array == busiest->expired && busiest->active->nr_active) {
1710 array = busiest->active;
1711 dst_array = this_rq->active;
1717 head = array->queue + idx;
1720 tmp = list_entry(curr, task_t, run_list);
1724 if (!can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
1731 #ifdef CONFIG_SCHEDSTATS
1732 if (task_hot(tmp, busiest->timestamp_last_tick, sd))
1733 schedstat_inc(sd, lb_hot_gained[idle]);
1736 pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
1739 /* We only want to steal up to the prescribed number of tasks. */
1740 if (pulled < max_nr_move) {
1748 * Right now, this is the only place pull_task() is called,
1749 * so we can safely collect pull_task() stats here rather than
1750 * inside pull_task().
1752 schedstat_add(sd, lb_gained[idle], pulled);
1755 *all_pinned = pinned;
1760 * find_busiest_group finds and returns the busiest CPU group within the
1761 * domain. It calculates and returns the number of tasks which should be
1762 * moved to restore balance via the imbalance parameter.
1764 static struct sched_group *
1765 find_busiest_group(struct sched_domain *sd, int this_cpu,
1766 unsigned long *imbalance, enum idle_type idle)
1768 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
1769 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
1771 max_load = this_load = total_load = total_pwr = 0;
1778 local_group = cpu_isset(this_cpu, group->cpumask);
1780 /* Tally up the load of all CPUs in the group */
1783 for_each_cpu_mask(i, group->cpumask) {
1784 /* Bias balancing toward cpus of our domain */
1786 load = target_load(i);
1788 load = source_load(i);
1793 total_load += avg_load;
1794 total_pwr += group->cpu_power;
1796 /* Adjust by relative CPU power of the group */
1797 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1800 this_load = avg_load;
1803 } else if (avg_load > max_load) {
1804 max_load = avg_load;
1808 group = group->next;
1809 } while (group != sd->groups);
1811 if (!busiest || this_load >= max_load)
1814 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
1816 if (this_load >= avg_load ||
1817 100*max_load <= sd->imbalance_pct*this_load)
1821 * We're trying to get all the cpus to the average_load, so we don't
1822 * want to push ourselves above the average load, nor do we wish to
1823 * reduce the max loaded cpu below the average load, as either of these
1824 * actions would just result in more rebalancing later, and ping-pong
1825 * tasks around. Thus we look for the minimum possible imbalance.
1826 * Negative imbalances (*we* are more loaded than anyone else) will
1827 * be counted as no imbalance for these purposes -- we can't fix that
1828 * by pulling tasks to us. Be careful of negative numbers as they'll
1829 * appear as very large values with unsigned longs.
1831 /* How much load to actually move to equalise the imbalance */
1832 *imbalance = min((max_load - avg_load) * busiest->cpu_power,
1833 (avg_load - this_load) * this->cpu_power)
1836 if (*imbalance < SCHED_LOAD_SCALE) {
1837 unsigned long pwr_now = 0, pwr_move = 0;
1840 if (max_load - this_load >= SCHED_LOAD_SCALE*2) {
1846 * OK, we don't have enough imbalance to justify moving tasks,
1847 * however we may be able to increase total CPU power used by
1851 pwr_now += busiest->cpu_power*min(SCHED_LOAD_SCALE, max_load);
1852 pwr_now += this->cpu_power*min(SCHED_LOAD_SCALE, this_load);
1853 pwr_now /= SCHED_LOAD_SCALE;
1855 /* Amount of load we'd subtract */
1856 tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/busiest->cpu_power;
1858 pwr_move += busiest->cpu_power*min(SCHED_LOAD_SCALE,
1861 /* Amount of load we'd add */
1862 if (max_load*busiest->cpu_power <
1863 SCHED_LOAD_SCALE*SCHED_LOAD_SCALE)
1864 tmp = max_load*busiest->cpu_power/this->cpu_power;
1866 tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/this->cpu_power;
1867 pwr_move += this->cpu_power*min(SCHED_LOAD_SCALE, this_load + tmp);
1868 pwr_move /= SCHED_LOAD_SCALE;
1870 /* Move if we gain throughput */
1871 if (pwr_move <= pwr_now)
1878 /* Get rid of the scaling factor, rounding down as we divide */
1879 *imbalance = *imbalance / SCHED_LOAD_SCALE;
1884 if (busiest && (idle == NEWLY_IDLE ||
1885 (idle == SCHED_IDLE && max_load > SCHED_LOAD_SCALE)) ) {
1895 * find_busiest_queue - find the busiest runqueue among the cpus in group.
1897 static runqueue_t *find_busiest_queue(struct sched_group *group)
1899 unsigned long load, max_load = 0;
1900 runqueue_t *busiest = NULL;
1903 for_each_cpu_mask(i, group->cpumask) {
1904 load = source_load(i);
1906 if (load > max_load) {
1908 busiest = cpu_rq(i);
1916 * Check this_cpu to ensure it is balanced within domain. Attempt to move
1917 * tasks if there is an imbalance.
1919 * Called with this_rq unlocked.
1921 static int load_balance(int this_cpu, runqueue_t *this_rq,
1922 struct sched_domain *sd, enum idle_type idle)
1924 struct sched_group *group;
1925 runqueue_t *busiest;
1926 unsigned long imbalance;
1927 int nr_moved, all_pinned;
1928 int active_balance = 0;
1930 spin_lock(&this_rq->lock);
1931 schedstat_inc(sd, lb_cnt[idle]);
1933 group = find_busiest_group(sd, this_cpu, &imbalance, idle);
1935 schedstat_inc(sd, lb_nobusyg[idle]);
1939 busiest = find_busiest_queue(group);
1941 schedstat_inc(sd, lb_nobusyq[idle]);
1946 * This should be "impossible", but since load
1947 * balancing is inherently racy and statistical,
1948 * it could happen in theory.
1950 if (unlikely(busiest == this_rq)) {
1955 schedstat_add(sd, lb_imbalance[idle], imbalance);
1958 if (busiest->nr_running > 1) {
1960 * Attempt to move tasks. If find_busiest_group has found
1961 * an imbalance but busiest->nr_running <= 1, the group is
1962 * still unbalanced. nr_moved simply stays zero, so it is
1963 * correctly treated as an imbalance.
1965 double_lock_balance(this_rq, busiest);
1966 nr_moved = move_tasks(this_rq, this_cpu, busiest,
1967 imbalance, sd, idle,
1969 spin_unlock(&busiest->lock);
1971 /* All tasks on this runqueue were pinned by CPU affinity */
1972 if (unlikely(all_pinned))
1976 spin_unlock(&this_rq->lock);
1979 schedstat_inc(sd, lb_failed[idle]);
1980 sd->nr_balance_failed++;
1982 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
1984 spin_lock(&busiest->lock);
1985 if (!busiest->active_balance) {
1986 busiest->active_balance = 1;
1987 busiest->push_cpu = this_cpu;
1990 spin_unlock(&busiest->lock);
1992 wake_up_process(busiest->migration_thread);
1995 * We've kicked active balancing, reset the failure
1998 sd->nr_balance_failed = sd->cache_nice_tries;
2001 sd->nr_balance_failed = 0;
2003 if (likely(!active_balance)) {
2004 /* We were unbalanced, so reset the balancing interval */
2005 sd->balance_interval = sd->min_interval;
2008 * If we've begun active balancing, start to back off. This
2009 * case may not be covered by the all_pinned logic if there
2010 * is only 1 task on the busy runqueue (because we don't call
2013 if (sd->balance_interval < sd->max_interval)
2014 sd->balance_interval *= 2;
2020 spin_unlock(&this_rq->lock);
2022 schedstat_inc(sd, lb_balanced[idle]);
2024 sd->nr_balance_failed = 0;
2025 /* tune up the balancing interval */
2026 if (sd->balance_interval < sd->max_interval)
2027 sd->balance_interval *= 2;
2033 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2034 * tasks if there is an imbalance.
2036 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2037 * this_rq is locked.
2039 static int load_balance_newidle(int this_cpu, runqueue_t *this_rq,
2040 struct sched_domain *sd)
2042 struct sched_group *group;
2043 runqueue_t *busiest = NULL;
2044 unsigned long imbalance;
2047 schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
2048 group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE);
2050 schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]);
2054 busiest = find_busiest_queue(group);
2055 if (!busiest || busiest == this_rq) {
2056 schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]);
2060 /* Attempt to move tasks */
2061 double_lock_balance(this_rq, busiest);
2063 schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
2064 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2065 imbalance, sd, NEWLY_IDLE, NULL);
2067 schedstat_inc(sd, lb_failed[NEWLY_IDLE]);
2069 sd->nr_balance_failed = 0;
2071 spin_unlock(&busiest->lock);
2075 schedstat_inc(sd, lb_balanced[NEWLY_IDLE]);
2076 sd->nr_balance_failed = 0;
2081 * idle_balance is called by schedule() if this_cpu is about to become
2082 * idle. Attempts to pull tasks from other CPUs.
2084 static inline void idle_balance(int this_cpu, runqueue_t *this_rq)
2086 struct sched_domain *sd;
2088 for_each_domain(this_cpu, sd) {
2089 if (sd->flags & SD_BALANCE_NEWIDLE) {
2090 if (load_balance_newidle(this_cpu, this_rq, sd)) {
2091 /* We've pulled tasks over so stop searching */
2099 * active_load_balance is run by migration threads. It pushes running tasks
2100 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2101 * running on each physical CPU where possible, and avoids physical /
2102 * logical imbalances.
2104 * Called with busiest_rq locked.
2106 static void active_load_balance(runqueue_t *busiest_rq, int busiest_cpu)
2108 struct sched_domain *sd;
2109 struct sched_group *cpu_group;
2110 runqueue_t *target_rq;
2111 cpumask_t visited_cpus;
2115 * Search for suitable CPUs to push tasks to in successively higher
2116 * domains with SD_LOAD_BALANCE set.
2118 visited_cpus = CPU_MASK_NONE;
2119 for_each_domain(busiest_cpu, sd) {
2120 if (!(sd->flags & SD_LOAD_BALANCE))
2121 /* no more domains to search */
2124 schedstat_inc(sd, alb_cnt);
2126 cpu_group = sd->groups;
2128 for_each_cpu_mask(cpu, cpu_group->cpumask) {
2129 if (busiest_rq->nr_running <= 1)
2130 /* no more tasks left to move */
2132 if (cpu_isset(cpu, visited_cpus))
2134 cpu_set(cpu, visited_cpus);
2135 if (!cpu_and_siblings_are_idle(cpu) || cpu == busiest_cpu)
2138 target_rq = cpu_rq(cpu);
2140 * This condition is "impossible", if it occurs
2141 * we need to fix it. Originally reported by
2142 * Bjorn Helgaas on a 128-cpu setup.
2144 BUG_ON(busiest_rq == target_rq);
2146 /* move a task from busiest_rq to target_rq */
2147 double_lock_balance(busiest_rq, target_rq);
2148 if (move_tasks(target_rq, cpu, busiest_rq,
2149 1, sd, SCHED_IDLE, NULL)) {
2150 schedstat_inc(sd, alb_pushed);
2152 schedstat_inc(sd, alb_failed);
2154 spin_unlock(&target_rq->lock);
2156 cpu_group = cpu_group->next;
2157 } while (cpu_group != sd->groups);
2162 * rebalance_tick will get called every timer tick, on every CPU.
2164 * It checks each scheduling domain to see if it is due to be balanced,
2165 * and initiates a balancing operation if so.
2167 * Balancing parameters are set up in arch_init_sched_domains.
2170 /* Don't have all balancing operations going off at once */
2171 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2173 static void rebalance_tick(int this_cpu, runqueue_t *this_rq,
2174 enum idle_type idle)
2176 unsigned long old_load, this_load;
2177 unsigned long j = jiffies + CPU_OFFSET(this_cpu);
2178 struct sched_domain *sd;
2180 /* Update our load */
2181 old_load = this_rq->cpu_load;
2182 this_load = this_rq->nr_running * SCHED_LOAD_SCALE;
2184 * Round up the averaging division if load is increasing. This
2185 * prevents us from getting stuck on 9 if the load is 10, for
2188 if (this_load > old_load)
2190 this_rq->cpu_load = (old_load + this_load) / 2;
2192 for_each_domain(this_cpu, sd) {
2193 unsigned long interval;
2195 if (!(sd->flags & SD_LOAD_BALANCE))
2198 interval = sd->balance_interval;
2199 if (idle != SCHED_IDLE)
2200 interval *= sd->busy_factor;
2202 /* scale ms to jiffies */
2203 interval = msecs_to_jiffies(interval);
2204 if (unlikely(!interval))
2207 if (j - sd->last_balance >= interval) {
2208 if (load_balance(this_cpu, this_rq, sd, idle)) {
2209 /* We've pulled tasks over so no longer idle */
2212 sd->last_balance += interval;
2218 * on UP we do not need to balance between CPUs:
2220 static inline void rebalance_tick(int cpu, runqueue_t *rq, enum idle_type idle)
2223 static inline void idle_balance(int cpu, runqueue_t *rq)
2228 static inline int wake_priority_sleeper(runqueue_t *rq)
2231 #ifdef CONFIG_SCHED_SMT
2232 spin_lock(&rq->lock);
2234 * If an SMT sibling task has been put to sleep for priority
2235 * reasons reschedule the idle task to see if it can now run.
2237 if (rq->nr_running) {
2238 resched_task(rq->idle);
2241 spin_unlock(&rq->lock);
2246 DEFINE_PER_CPU(struct kernel_stat, kstat);
2248 EXPORT_PER_CPU_SYMBOL(kstat);
2251 * This is called on clock ticks and on context switches.
2252 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2254 static inline void update_cpu_clock(task_t *p, runqueue_t *rq,
2255 unsigned long long now)
2257 unsigned long long last = max(p->timestamp, rq->timestamp_last_tick);
2258 p->sched_time += now - last;
2262 * Return current->sched_time plus any more ns on the sched_clock
2263 * that have not yet been banked.
2265 unsigned long long current_sched_time(const task_t *tsk)
2267 unsigned long long ns;
2268 unsigned long flags;
2269 local_irq_save(flags);
2270 ns = max(tsk->timestamp, task_rq(tsk)->timestamp_last_tick);
2271 ns = tsk->sched_time + (sched_clock() - ns);
2272 local_irq_restore(flags);
2277 * We place interactive tasks back into the active array, if possible.
2279 * To guarantee that this does not starve expired tasks we ignore the
2280 * interactivity of a task if the first expired task had to wait more
2281 * than a 'reasonable' amount of time. This deadline timeout is
2282 * load-dependent, as the frequency of array switched decreases with
2283 * increasing number of running tasks. We also ignore the interactivity
2284 * if a better static_prio task has expired:
2286 #define EXPIRED_STARVING(rq) \
2287 ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2288 (jiffies - (rq)->expired_timestamp >= \
2289 STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2290 ((rq)->curr->static_prio > (rq)->best_expired_prio))
2293 * Account user cpu time to a process.
2294 * @p: the process that the cpu time gets accounted to
2295 * @hardirq_offset: the offset to subtract from hardirq_count()
2296 * @cputime: the cpu time spent in user space since the last update
2298 void account_user_time(struct task_struct *p, cputime_t cputime)
2300 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2303 p->utime = cputime_add(p->utime, cputime);
2305 /* Add user time to cpustat. */
2306 tmp = cputime_to_cputime64(cputime);
2307 if (TASK_NICE(p) > 0)
2308 cpustat->nice = cputime64_add(cpustat->nice, tmp);
2310 cpustat->user = cputime64_add(cpustat->user, tmp);
2314 * Account system cpu time to a process.
2315 * @p: the process that the cpu time gets accounted to
2316 * @hardirq_offset: the offset to subtract from hardirq_count()
2317 * @cputime: the cpu time spent in kernel space since the last update
2319 void account_system_time(struct task_struct *p, int hardirq_offset,
2322 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2323 runqueue_t *rq = this_rq();
2326 p->stime = cputime_add(p->stime, cputime);
2328 /* Add system time to cpustat. */
2329 tmp = cputime_to_cputime64(cputime);
2330 if (hardirq_count() - hardirq_offset)
2331 cpustat->irq = cputime64_add(cpustat->irq, tmp);
2332 else if (softirq_count())
2333 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
2334 else if (p != rq->idle)
2335 cpustat->system = cputime64_add(cpustat->system, tmp);
2336 else if (atomic_read(&rq->nr_iowait) > 0)
2337 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2339 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2340 /* Account for system time used */
2341 acct_update_integrals(p);
2342 /* Update rss highwater mark */
2343 update_mem_hiwater(p);
2347 * Account for involuntary wait time.
2348 * @p: the process from which the cpu time has been stolen
2349 * @steal: the cpu time spent in involuntary wait
2351 void account_steal_time(struct task_struct *p, cputime_t steal)
2353 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2354 cputime64_t tmp = cputime_to_cputime64(steal);
2355 runqueue_t *rq = this_rq();
2357 if (p == rq->idle) {
2358 p->stime = cputime_add(p->stime, steal);
2359 if (atomic_read(&rq->nr_iowait) > 0)
2360 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2362 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2364 cpustat->steal = cputime64_add(cpustat->steal, tmp);
2368 * This function gets called by the timer code, with HZ frequency.
2369 * We call it with interrupts disabled.
2371 * It also gets called by the fork code, when changing the parent's
2374 void scheduler_tick(void)
2376 int cpu = smp_processor_id();
2377 runqueue_t *rq = this_rq();
2378 task_t *p = current;
2379 unsigned long long now = sched_clock();
2381 update_cpu_clock(p, rq, now);
2383 rq->timestamp_last_tick = now;
2385 if (p == rq->idle) {
2386 if (wake_priority_sleeper(rq))
2388 rebalance_tick(cpu, rq, SCHED_IDLE);
2392 /* Task might have expired already, but not scheduled off yet */
2393 if (p->array != rq->active) {
2394 set_tsk_need_resched(p);
2397 spin_lock(&rq->lock);
2399 * The task was running during this tick - update the
2400 * time slice counter. Note: we do not update a thread's
2401 * priority until it either goes to sleep or uses up its
2402 * timeslice. This makes it possible for interactive tasks
2403 * to use up their timeslices at their highest priority levels.
2407 * RR tasks need a special form of timeslice management.
2408 * FIFO tasks have no timeslices.
2410 if ((p->policy == SCHED_RR) && !--p->time_slice) {
2411 p->time_slice = task_timeslice(p);
2412 p->first_time_slice = 0;
2413 set_tsk_need_resched(p);
2415 /* put it at the end of the queue: */
2416 requeue_task(p, rq->active);
2420 if (!--p->time_slice) {
2421 dequeue_task(p, rq->active);
2422 set_tsk_need_resched(p);
2423 p->prio = effective_prio(p);
2424 p->time_slice = task_timeslice(p);
2425 p->first_time_slice = 0;
2427 if (!rq->expired_timestamp)
2428 rq->expired_timestamp = jiffies;
2429 if (!TASK_INTERACTIVE(p) || EXPIRED_STARVING(rq)) {
2430 enqueue_task(p, rq->expired);
2431 if (p->static_prio < rq->best_expired_prio)
2432 rq->best_expired_prio = p->static_prio;
2434 enqueue_task(p, rq->active);
2437 * Prevent a too long timeslice allowing a task to monopolize
2438 * the CPU. We do this by splitting up the timeslice into
2441 * Note: this does not mean the task's timeslices expire or
2442 * get lost in any way, they just might be preempted by
2443 * another task of equal priority. (one with higher
2444 * priority would have preempted this task already.) We
2445 * requeue this task to the end of the list on this priority
2446 * level, which is in essence a round-robin of tasks with
2449 * This only applies to tasks in the interactive
2450 * delta range with at least TIMESLICE_GRANULARITY to requeue.
2452 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
2453 p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
2454 (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
2455 (p->array == rq->active)) {
2457 requeue_task(p, rq->active);
2458 set_tsk_need_resched(p);
2462 spin_unlock(&rq->lock);
2464 rebalance_tick(cpu, rq, NOT_IDLE);
2467 #ifdef CONFIG_SCHED_SMT
2468 static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2470 struct sched_domain *sd = this_rq->sd;
2471 cpumask_t sibling_map;
2474 if (!(sd->flags & SD_SHARE_CPUPOWER))
2478 * Unlock the current runqueue because we have to lock in
2479 * CPU order to avoid deadlocks. Caller knows that we might
2480 * unlock. We keep IRQs disabled.
2482 spin_unlock(&this_rq->lock);
2484 sibling_map = sd->span;
2486 for_each_cpu_mask(i, sibling_map)
2487 spin_lock(&cpu_rq(i)->lock);
2489 * We clear this CPU from the mask. This both simplifies the
2490 * inner loop and keps this_rq locked when we exit:
2492 cpu_clear(this_cpu, sibling_map);
2494 for_each_cpu_mask(i, sibling_map) {
2495 runqueue_t *smt_rq = cpu_rq(i);
2498 * If an SMT sibling task is sleeping due to priority
2499 * reasons wake it up now.
2501 if (smt_rq->curr == smt_rq->idle && smt_rq->nr_running)
2502 resched_task(smt_rq->idle);
2505 for_each_cpu_mask(i, sibling_map)
2506 spin_unlock(&cpu_rq(i)->lock);
2508 * We exit with this_cpu's rq still held and IRQs
2513 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2515 struct sched_domain *sd = this_rq->sd;
2516 cpumask_t sibling_map;
2517 prio_array_t *array;
2521 if (!(sd->flags & SD_SHARE_CPUPOWER))
2525 * The same locking rules and details apply as for
2526 * wake_sleeping_dependent():
2528 spin_unlock(&this_rq->lock);
2529 sibling_map = sd->span;
2530 for_each_cpu_mask(i, sibling_map)
2531 spin_lock(&cpu_rq(i)->lock);
2532 cpu_clear(this_cpu, sibling_map);
2535 * Establish next task to be run - it might have gone away because
2536 * we released the runqueue lock above:
2538 if (!this_rq->nr_running)
2540 array = this_rq->active;
2541 if (!array->nr_active)
2542 array = this_rq->expired;
2543 BUG_ON(!array->nr_active);
2545 p = list_entry(array->queue[sched_find_first_bit(array->bitmap)].next,
2548 for_each_cpu_mask(i, sibling_map) {
2549 runqueue_t *smt_rq = cpu_rq(i);
2550 task_t *smt_curr = smt_rq->curr;
2553 * If a user task with lower static priority than the
2554 * running task on the SMT sibling is trying to schedule,
2555 * delay it till there is proportionately less timeslice
2556 * left of the sibling task to prevent a lower priority
2557 * task from using an unfair proportion of the
2558 * physical cpu's resources. -ck
2560 if (((smt_curr->time_slice * (100 - sd->per_cpu_gain) / 100) >
2561 task_timeslice(p) || rt_task(smt_curr)) &&
2562 p->mm && smt_curr->mm && !rt_task(p))
2566 * Reschedule a lower priority task on the SMT sibling,
2567 * or wake it up if it has been put to sleep for priority
2570 if ((((p->time_slice * (100 - sd->per_cpu_gain) / 100) >
2571 task_timeslice(smt_curr) || rt_task(p)) &&
2572 smt_curr->mm && p->mm && !rt_task(smt_curr)) ||
2573 (smt_curr == smt_rq->idle && smt_rq->nr_running))
2574 resched_task(smt_curr);
2577 for_each_cpu_mask(i, sibling_map)
2578 spin_unlock(&cpu_rq(i)->lock);
2582 static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2586 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2592 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
2594 void fastcall add_preempt_count(int val)
2599 BUG_ON((preempt_count() < 0));
2600 preempt_count() += val;
2602 * Spinlock count overflowing soon?
2604 BUG_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10);
2606 EXPORT_SYMBOL(add_preempt_count);
2608 void fastcall sub_preempt_count(int val)
2613 BUG_ON(val > preempt_count());
2615 * Is the spinlock portion underflowing?
2617 BUG_ON((val < PREEMPT_MASK) && !(preempt_count() & PREEMPT_MASK));
2618 preempt_count() -= val;
2620 EXPORT_SYMBOL(sub_preempt_count);
2625 * schedule() is the main scheduler function.
2627 asmlinkage void __sched schedule(void)
2630 task_t *prev, *next;
2632 prio_array_t *array;
2633 struct list_head *queue;
2634 unsigned long long now;
2635 unsigned long run_time;
2639 * Test if we are atomic. Since do_exit() needs to call into
2640 * schedule() atomically, we ignore that path for now.
2641 * Otherwise, whine if we are scheduling when we should not be.
2643 if (likely(!current->exit_state)) {
2644 if (unlikely(in_atomic())) {
2645 printk(KERN_ERR "scheduling while atomic: "
2647 current->comm, preempt_count(), current->pid);
2651 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2656 release_kernel_lock(prev);
2657 need_resched_nonpreemptible:
2661 * The idle thread is not allowed to schedule!
2662 * Remove this check after it has been exercised a bit.
2664 if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
2665 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
2669 schedstat_inc(rq, sched_cnt);
2670 now = sched_clock();
2671 if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
2672 run_time = now - prev->timestamp;
2673 if (unlikely((long long)(now - prev->timestamp) < 0))
2676 run_time = NS_MAX_SLEEP_AVG;
2679 * Tasks charged proportionately less run_time at high sleep_avg to
2680 * delay them losing their interactive status
2682 run_time /= (CURRENT_BONUS(prev) ? : 1);
2684 spin_lock_irq(&rq->lock);
2686 if (unlikely(prev->flags & PF_DEAD))
2687 prev->state = EXIT_DEAD;
2689 switch_count = &prev->nivcsw;
2690 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2691 switch_count = &prev->nvcsw;
2692 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
2693 unlikely(signal_pending(prev))))
2694 prev->state = TASK_RUNNING;
2696 if (prev->state == TASK_UNINTERRUPTIBLE)
2697 rq->nr_uninterruptible++;
2698 deactivate_task(prev, rq);
2702 cpu = smp_processor_id();
2703 if (unlikely(!rq->nr_running)) {
2705 idle_balance(cpu, rq);
2706 if (!rq->nr_running) {
2708 rq->expired_timestamp = 0;
2709 wake_sleeping_dependent(cpu, rq);
2711 * wake_sleeping_dependent() might have released
2712 * the runqueue, so break out if we got new
2715 if (!rq->nr_running)
2719 if (dependent_sleeper(cpu, rq)) {
2724 * dependent_sleeper() releases and reacquires the runqueue
2725 * lock, hence go into the idle loop if the rq went
2728 if (unlikely(!rq->nr_running))
2733 if (unlikely(!array->nr_active)) {
2735 * Switch the active and expired arrays.
2737 schedstat_inc(rq, sched_switch);
2738 rq->active = rq->expired;
2739 rq->expired = array;
2741 rq->expired_timestamp = 0;
2742 rq->best_expired_prio = MAX_PRIO;
2745 idx = sched_find_first_bit(array->bitmap);
2746 queue = array->queue + idx;
2747 next = list_entry(queue->next, task_t, run_list);
2749 if (!rt_task(next) && next->activated > 0) {
2750 unsigned long long delta = now - next->timestamp;
2751 if (unlikely((long long)(now - next->timestamp) < 0))
2754 if (next->activated == 1)
2755 delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
2757 array = next->array;
2758 dequeue_task(next, array);
2759 recalc_task_prio(next, next->timestamp + delta);
2760 enqueue_task(next, array);
2762 next->activated = 0;
2764 if (next == rq->idle)
2765 schedstat_inc(rq, sched_goidle);
2767 clear_tsk_need_resched(prev);
2768 rcu_qsctr_inc(task_cpu(prev));
2770 update_cpu_clock(prev, rq, now);
2772 prev->sleep_avg -= run_time;
2773 if ((long)prev->sleep_avg <= 0)
2774 prev->sleep_avg = 0;
2775 prev->timestamp = prev->last_ran = now;
2777 sched_info_switch(prev, next);
2778 if (likely(prev != next)) {
2779 next->timestamp = now;
2784 prepare_arch_switch(rq, next);
2785 prev = context_switch(rq, prev, next);
2788 finish_task_switch(prev);
2790 spin_unlock_irq(&rq->lock);
2793 if (unlikely(reacquire_kernel_lock(prev) < 0))
2794 goto need_resched_nonpreemptible;
2795 preempt_enable_no_resched();
2796 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
2800 EXPORT_SYMBOL(schedule);
2802 #ifdef CONFIG_PREEMPT
2804 * this is is the entry point to schedule() from in-kernel preemption
2805 * off of preempt_enable. Kernel preemptions off return from interrupt
2806 * occur there and call schedule directly.
2808 asmlinkage void __sched preempt_schedule(void)
2810 struct thread_info *ti = current_thread_info();
2811 #ifdef CONFIG_PREEMPT_BKL
2812 struct task_struct *task = current;
2813 int saved_lock_depth;
2816 * If there is a non-zero preempt_count or interrupts are disabled,
2817 * we do not want to preempt the current task. Just return..
2819 if (unlikely(ti->preempt_count || irqs_disabled()))
2823 add_preempt_count(PREEMPT_ACTIVE);
2825 * We keep the big kernel semaphore locked, but we
2826 * clear ->lock_depth so that schedule() doesnt
2827 * auto-release the semaphore:
2829 #ifdef CONFIG_PREEMPT_BKL
2830 saved_lock_depth = task->lock_depth;
2831 task->lock_depth = -1;
2834 #ifdef CONFIG_PREEMPT_BKL
2835 task->lock_depth = saved_lock_depth;
2837 sub_preempt_count(PREEMPT_ACTIVE);
2839 /* we could miss a preemption opportunity between schedule and now */
2841 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
2845 EXPORT_SYMBOL(preempt_schedule);
2848 * this is is the entry point to schedule() from kernel preemption
2849 * off of irq context.
2850 * Note, that this is called and return with irqs disabled. This will
2851 * protect us against recursive calling from irq.
2853 asmlinkage void __sched preempt_schedule_irq(void)
2855 struct thread_info *ti = current_thread_info();
2856 #ifdef CONFIG_PREEMPT_BKL
2857 struct task_struct *task = current;
2858 int saved_lock_depth;
2860 /* Catch callers which need to be fixed*/
2861 BUG_ON(ti->preempt_count || !irqs_disabled());
2864 add_preempt_count(PREEMPT_ACTIVE);
2866 * We keep the big kernel semaphore locked, but we
2867 * clear ->lock_depth so that schedule() doesnt
2868 * auto-release the semaphore:
2870 #ifdef CONFIG_PREEMPT_BKL
2871 saved_lock_depth = task->lock_depth;
2872 task->lock_depth = -1;
2876 local_irq_disable();
2877 #ifdef CONFIG_PREEMPT_BKL
2878 task->lock_depth = saved_lock_depth;
2880 sub_preempt_count(PREEMPT_ACTIVE);
2882 /* we could miss a preemption opportunity between schedule and now */
2884 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
2888 #endif /* CONFIG_PREEMPT */
2890 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync, void *key)
2892 task_t *p = curr->private;
2893 return try_to_wake_up(p, mode, sync);
2896 EXPORT_SYMBOL(default_wake_function);
2899 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
2900 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
2901 * number) then we wake all the non-exclusive tasks and one exclusive task.
2903 * There are circumstances in which we can try to wake a task which has already
2904 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
2905 * zero in this (rare) case, and we handle it by continuing to scan the queue.
2907 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
2908 int nr_exclusive, int sync, void *key)
2910 struct list_head *tmp, *next;
2912 list_for_each_safe(tmp, next, &q->task_list) {
2915 curr = list_entry(tmp, wait_queue_t, task_list);
2916 flags = curr->flags;
2917 if (curr->func(curr, mode, sync, key) &&
2918 (flags & WQ_FLAG_EXCLUSIVE) &&
2925 * __wake_up - wake up threads blocked on a waitqueue.
2927 * @mode: which threads
2928 * @nr_exclusive: how many wake-one or wake-many threads to wake up
2929 * @key: is directly passed to the wakeup function
2931 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
2932 int nr_exclusive, void *key)
2934 unsigned long flags;
2936 spin_lock_irqsave(&q->lock, flags);
2937 __wake_up_common(q, mode, nr_exclusive, 0, key);
2938 spin_unlock_irqrestore(&q->lock, flags);
2941 EXPORT_SYMBOL(__wake_up);
2944 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
2946 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
2948 __wake_up_common(q, mode, 1, 0, NULL);
2952 * __wake_up_sync - wake up threads blocked on a waitqueue.
2954 * @mode: which threads
2955 * @nr_exclusive: how many wake-one or wake-many threads to wake up
2957 * The sync wakeup differs that the waker knows that it will schedule
2958 * away soon, so while the target thread will be woken up, it will not
2959 * be migrated to another CPU - ie. the two threads are 'synchronized'
2960 * with each other. This can prevent needless bouncing between CPUs.
2962 * On UP it can prevent extra preemption.
2964 void fastcall __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
2966 unsigned long flags;
2972 if (unlikely(!nr_exclusive))
2975 spin_lock_irqsave(&q->lock, flags);
2976 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
2977 spin_unlock_irqrestore(&q->lock, flags);
2979 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
2981 void fastcall complete(struct completion *x)
2983 unsigned long flags;
2985 spin_lock_irqsave(&x->wait.lock, flags);
2987 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
2989 spin_unlock_irqrestore(&x->wait.lock, flags);
2991 EXPORT_SYMBOL(complete);
2993 void fastcall complete_all(struct completion *x)
2995 unsigned long flags;
2997 spin_lock_irqsave(&x->wait.lock, flags);
2998 x->done += UINT_MAX/2;
2999 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3001 spin_unlock_irqrestore(&x->wait.lock, flags);
3003 EXPORT_SYMBOL(complete_all);
3005 void fastcall __sched wait_for_completion(struct completion *x)
3008 spin_lock_irq(&x->wait.lock);
3010 DECLARE_WAITQUEUE(wait, current);
3012 wait.flags |= WQ_FLAG_EXCLUSIVE;
3013 __add_wait_queue_tail(&x->wait, &wait);
3015 __set_current_state(TASK_UNINTERRUPTIBLE);
3016 spin_unlock_irq(&x->wait.lock);
3018 spin_lock_irq(&x->wait.lock);
3020 __remove_wait_queue(&x->wait, &wait);
3023 spin_unlock_irq(&x->wait.lock);
3025 EXPORT_SYMBOL(wait_for_completion);
3027 unsigned long fastcall __sched
3028 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3032 spin_lock_irq(&x->wait.lock);
3034 DECLARE_WAITQUEUE(wait, current);
3036 wait.flags |= WQ_FLAG_EXCLUSIVE;
3037 __add_wait_queue_tail(&x->wait, &wait);
3039 __set_current_state(TASK_UNINTERRUPTIBLE);
3040 spin_unlock_irq(&x->wait.lock);
3041 timeout = schedule_timeout(timeout);
3042 spin_lock_irq(&x->wait.lock);
3044 __remove_wait_queue(&x->wait, &wait);
3048 __remove_wait_queue(&x->wait, &wait);
3052 spin_unlock_irq(&x->wait.lock);
3055 EXPORT_SYMBOL(wait_for_completion_timeout);
3057 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3063 spin_lock_irq(&x->wait.lock);
3065 DECLARE_WAITQUEUE(wait, current);
3067 wait.flags |= WQ_FLAG_EXCLUSIVE;
3068 __add_wait_queue_tail(&x->wait, &wait);
3070 if (signal_pending(current)) {
3072 __remove_wait_queue(&x->wait, &wait);
3075 __set_current_state(TASK_INTERRUPTIBLE);
3076 spin_unlock_irq(&x->wait.lock);
3078 spin_lock_irq(&x->wait.lock);
3080 __remove_wait_queue(&x->wait, &wait);
3084 spin_unlock_irq(&x->wait.lock);
3088 EXPORT_SYMBOL(wait_for_completion_interruptible);
3090 unsigned long fastcall __sched
3091 wait_for_completion_interruptible_timeout(struct completion *x,
3092 unsigned long timeout)
3096 spin_lock_irq(&x->wait.lock);
3098 DECLARE_WAITQUEUE(wait, current);
3100 wait.flags |= WQ_FLAG_EXCLUSIVE;
3101 __add_wait_queue_tail(&x->wait, &wait);
3103 if (signal_pending(current)) {
3104 timeout = -ERESTARTSYS;
3105 __remove_wait_queue(&x->wait, &wait);
3108 __set_current_state(TASK_INTERRUPTIBLE);
3109 spin_unlock_irq(&x->wait.lock);
3110 timeout = schedule_timeout(timeout);
3111 spin_lock_irq(&x->wait.lock);
3113 __remove_wait_queue(&x->wait, &wait);
3117 __remove_wait_queue(&x->wait, &wait);
3121 spin_unlock_irq(&x->wait.lock);
3124 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3127 #define SLEEP_ON_VAR \
3128 unsigned long flags; \
3129 wait_queue_t wait; \
3130 init_waitqueue_entry(&wait, current);
3132 #define SLEEP_ON_HEAD \
3133 spin_lock_irqsave(&q->lock,flags); \
3134 __add_wait_queue(q, &wait); \
3135 spin_unlock(&q->lock);
3137 #define SLEEP_ON_TAIL \
3138 spin_lock_irq(&q->lock); \
3139 __remove_wait_queue(q, &wait); \
3140 spin_unlock_irqrestore(&q->lock, flags);
3142 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
3146 current->state = TASK_INTERRUPTIBLE;
3153 EXPORT_SYMBOL(interruptible_sleep_on);
3155 long fastcall __sched interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3159 current->state = TASK_INTERRUPTIBLE;
3162 timeout = schedule_timeout(timeout);
3168 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3170 void fastcall __sched sleep_on(wait_queue_head_t *q)
3174 current->state = TASK_UNINTERRUPTIBLE;
3181 EXPORT_SYMBOL(sleep_on);
3183 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3187 current->state = TASK_UNINTERRUPTIBLE;
3190 timeout = schedule_timeout(timeout);
3196 EXPORT_SYMBOL(sleep_on_timeout);
3198 void set_user_nice(task_t *p, long nice)
3200 unsigned long flags;
3201 prio_array_t *array;
3203 int old_prio, new_prio, delta;
3205 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3208 * We have to be careful, if called from sys_setpriority(),
3209 * the task might be in the middle of scheduling on another CPU.
3211 rq = task_rq_lock(p, &flags);
3213 * The RT priorities are set via sched_setscheduler(), but we still
3214 * allow the 'normal' nice value to be set - but as expected
3215 * it wont have any effect on scheduling until the task is
3219 p->static_prio = NICE_TO_PRIO(nice);
3224 dequeue_task(p, array);
3227 new_prio = NICE_TO_PRIO(nice);
3228 delta = new_prio - old_prio;
3229 p->static_prio = NICE_TO_PRIO(nice);
3233 enqueue_task(p, array);
3235 * If the task increased its priority or is running and
3236 * lowered its priority, then reschedule its CPU:
3238 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3239 resched_task(rq->curr);
3242 task_rq_unlock(rq, &flags);
3245 EXPORT_SYMBOL(set_user_nice);
3248 * can_nice - check if a task can reduce its nice value
3252 int can_nice(const task_t *p, const int nice)
3254 /* convert nice value [19,-20] to rlimit style value [0,39] */
3255 int nice_rlim = 19 - nice;
3256 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
3257 capable(CAP_SYS_NICE));
3260 #ifdef __ARCH_WANT_SYS_NICE
3263 * sys_nice - change the priority of the current process.
3264 * @increment: priority increment
3266 * sys_setpriority is a more generic, but much slower function that
3267 * does similar things.
3269 asmlinkage long sys_nice(int increment)
3275 * Setpriority might change our priority at the same moment.
3276 * We don't have to worry. Conceptually one call occurs first
3277 * and we have a single winner.
3279 if (increment < -40)
3284 nice = PRIO_TO_NICE(current->static_prio) + increment;
3290 if (increment < 0 && !can_nice(current, nice))
3293 retval = security_task_setnice(current, nice);
3297 set_user_nice(current, nice);
3304 * task_prio - return the priority value of a given task.
3305 * @p: the task in question.
3307 * This is the priority value as seen by users in /proc.
3308 * RT tasks are offset by -200. Normal tasks are centered
3309 * around 0, value goes from -16 to +15.
3311 int task_prio(const task_t *p)
3313 return p->prio - MAX_RT_PRIO;
3317 * task_nice - return the nice value of a given task.
3318 * @p: the task in question.
3320 int task_nice(const task_t *p)
3322 return TASK_NICE(p);
3326 * The only users of task_nice are binfmt_elf and binfmt_elf32.
3327 * binfmt_elf is no longer modular, but binfmt_elf32 still is.
3328 * Therefore, task_nice is needed if there is a compat_mode.
3330 #ifdef CONFIG_COMPAT
3331 EXPORT_SYMBOL_GPL(task_nice);
3335 * idle_cpu - is a given cpu idle currently?
3336 * @cpu: the processor in question.
3338 int idle_cpu(int cpu)
3340 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
3343 EXPORT_SYMBOL_GPL(idle_cpu);
3346 * idle_task - return the idle task for a given cpu.
3347 * @cpu: the processor in question.
3349 task_t *idle_task(int cpu)
3351 return cpu_rq(cpu)->idle;
3355 * find_process_by_pid - find a process with a matching PID value.
3356 * @pid: the pid in question.
3358 static inline task_t *find_process_by_pid(pid_t pid)
3360 return pid ? find_task_by_pid(pid) : current;
3363 /* Actually do priority change: must hold rq lock. */
3364 static void __setscheduler(struct task_struct *p, int policy, int prio)
3368 p->rt_priority = prio;
3369 if (policy != SCHED_NORMAL)
3370 p->prio = MAX_USER_RT_PRIO-1 - p->rt_priority;
3372 p->prio = p->static_prio;
3376 * sched_setscheduler - change the scheduling policy and/or RT priority of
3378 * @p: the task in question.
3379 * @policy: new policy.
3380 * @param: structure containing the new RT priority.
3382 int sched_setscheduler(struct task_struct *p, int policy, struct sched_param *param)
3385 int oldprio, oldpolicy = -1;
3386 prio_array_t *array;
3387 unsigned long flags;
3391 /* double check policy once rq lock held */
3393 policy = oldpolicy = p->policy;
3394 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
3395 policy != SCHED_NORMAL)
3398 * Valid priorities for SCHED_FIFO and SCHED_RR are
3399 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL is 0.
3401 if (param->sched_priority < 0 ||
3402 param->sched_priority > MAX_USER_RT_PRIO-1)
3404 if ((policy == SCHED_NORMAL) != (param->sched_priority == 0))
3407 if ((policy == SCHED_FIFO || policy == SCHED_RR) &&
3408 param->sched_priority > p->signal->rlim[RLIMIT_RTPRIO].rlim_cur &&
3409 !capable(CAP_SYS_NICE))
3411 if ((current->euid != p->euid) && (current->euid != p->uid) &&
3412 !capable(CAP_SYS_NICE))
3415 retval = security_task_setscheduler(p, policy, param);
3419 * To be able to change p->policy safely, the apropriate
3420 * runqueue lock must be held.
3422 rq = task_rq_lock(p, &flags);
3423 /* recheck policy now with rq lock held */
3424 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3425 policy = oldpolicy = -1;
3426 task_rq_unlock(rq, &flags);
3431 deactivate_task(p, rq);
3433 __setscheduler(p, policy, param->sched_priority);
3435 __activate_task(p, rq);
3437 * Reschedule if we are currently running on this runqueue and
3438 * our priority decreased, or if we are not currently running on
3439 * this runqueue and our priority is higher than the current's
3441 if (task_running(rq, p)) {
3442 if (p->prio > oldprio)
3443 resched_task(rq->curr);
3444 } else if (TASK_PREEMPTS_CURR(p, rq))
3445 resched_task(rq->curr);
3447 task_rq_unlock(rq, &flags);
3450 EXPORT_SYMBOL_GPL(sched_setscheduler);
3452 static int do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3455 struct sched_param lparam;
3456 struct task_struct *p;
3458 if (!param || pid < 0)
3460 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3462 read_lock_irq(&tasklist_lock);
3463 p = find_process_by_pid(pid);
3465 read_unlock_irq(&tasklist_lock);
3468 retval = sched_setscheduler(p, policy, &lparam);
3469 read_unlock_irq(&tasklist_lock);
3474 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3475 * @pid: the pid in question.
3476 * @policy: new policy.
3477 * @param: structure containing the new RT priority.
3479 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
3480 struct sched_param __user *param)
3482 return do_sched_setscheduler(pid, policy, param);
3486 * sys_sched_setparam - set/change the RT priority of a thread
3487 * @pid: the pid in question.
3488 * @param: structure containing the new RT priority.
3490 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
3492 return do_sched_setscheduler(pid, -1, param);
3496 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3497 * @pid: the pid in question.
3499 asmlinkage long sys_sched_getscheduler(pid_t pid)
3501 int retval = -EINVAL;
3508 read_lock(&tasklist_lock);
3509 p = find_process_by_pid(pid);
3511 retval = security_task_getscheduler(p);
3515 read_unlock(&tasklist_lock);
3522 * sys_sched_getscheduler - get the RT priority of a thread
3523 * @pid: the pid in question.
3524 * @param: structure containing the RT priority.
3526 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
3528 struct sched_param lp;
3529 int retval = -EINVAL;
3532 if (!param || pid < 0)
3535 read_lock(&tasklist_lock);
3536 p = find_process_by_pid(pid);
3541 retval = security_task_getscheduler(p);
3545 lp.sched_priority = p->rt_priority;
3546 read_unlock(&tasklist_lock);
3549 * This one might sleep, we cannot do it with a spinlock held ...
3551 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3557 read_unlock(&tasklist_lock);
3561 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
3565 cpumask_t cpus_allowed;
3568 read_lock(&tasklist_lock);
3570 p = find_process_by_pid(pid);
3572 read_unlock(&tasklist_lock);
3573 unlock_cpu_hotplug();
3578 * It is not safe to call set_cpus_allowed with the
3579 * tasklist_lock held. We will bump the task_struct's
3580 * usage count and then drop tasklist_lock.
3583 read_unlock(&tasklist_lock);
3586 if ((current->euid != p->euid) && (current->euid != p->uid) &&
3587 !capable(CAP_SYS_NICE))
3590 cpus_allowed = cpuset_cpus_allowed(p);
3591 cpus_and(new_mask, new_mask, cpus_allowed);
3592 retval = set_cpus_allowed(p, new_mask);
3596 unlock_cpu_hotplug();
3600 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3601 cpumask_t *new_mask)
3603 if (len < sizeof(cpumask_t)) {
3604 memset(new_mask, 0, sizeof(cpumask_t));
3605 } else if (len > sizeof(cpumask_t)) {
3606 len = sizeof(cpumask_t);
3608 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3612 * sys_sched_setaffinity - set the cpu affinity of a process
3613 * @pid: pid of the process
3614 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3615 * @user_mask_ptr: user-space pointer to the new cpu mask
3617 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
3618 unsigned long __user *user_mask_ptr)
3623 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
3627 return sched_setaffinity(pid, new_mask);
3631 * Represents all cpu's present in the system
3632 * In systems capable of hotplug, this map could dynamically grow
3633 * as new cpu's are detected in the system via any platform specific
3634 * method, such as ACPI for e.g.
3637 cpumask_t cpu_present_map;
3638 EXPORT_SYMBOL(cpu_present_map);
3641 cpumask_t cpu_online_map = CPU_MASK_ALL;
3642 cpumask_t cpu_possible_map = CPU_MASK_ALL;
3645 long sched_getaffinity(pid_t pid, cpumask_t *mask)
3651 read_lock(&tasklist_lock);
3654 p = find_process_by_pid(pid);
3659 cpus_and(*mask, p->cpus_allowed, cpu_possible_map);
3662 read_unlock(&tasklist_lock);
3663 unlock_cpu_hotplug();
3671 * sys_sched_getaffinity - get the cpu affinity of a process
3672 * @pid: pid of the process
3673 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3674 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3676 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
3677 unsigned long __user *user_mask_ptr)
3682 if (len < sizeof(cpumask_t))
3685 ret = sched_getaffinity(pid, &mask);
3689 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
3692 return sizeof(cpumask_t);
3696 * sys_sched_yield - yield the current processor to other threads.
3698 * this function yields the current CPU by moving the calling thread
3699 * to the expired array. If there are no other threads running on this
3700 * CPU then this function will return.
3702 asmlinkage long sys_sched_yield(void)
3704 runqueue_t *rq = this_rq_lock();
3705 prio_array_t *array = current->array;
3706 prio_array_t *target = rq->expired;
3708 schedstat_inc(rq, yld_cnt);
3710 * We implement yielding by moving the task into the expired
3713 * (special rule: RT tasks will just roundrobin in the active
3716 if (rt_task(current))
3717 target = rq->active;
3719 if (current->array->nr_active == 1) {
3720 schedstat_inc(rq, yld_act_empty);
3721 if (!rq->expired->nr_active)
3722 schedstat_inc(rq, yld_both_empty);
3723 } else if (!rq->expired->nr_active)
3724 schedstat_inc(rq, yld_exp_empty);
3726 if (array != target) {
3727 dequeue_task(current, array);
3728 enqueue_task(current, target);
3731 * requeue_task is cheaper so perform that if possible.
3733 requeue_task(current, array);
3736 * Since we are going to call schedule() anyway, there's
3737 * no need to preempt or enable interrupts:
3739 __release(rq->lock);
3740 _raw_spin_unlock(&rq->lock);
3741 preempt_enable_no_resched();
3748 static inline void __cond_resched(void)
3751 add_preempt_count(PREEMPT_ACTIVE);
3753 sub_preempt_count(PREEMPT_ACTIVE);
3754 } while (need_resched());
3757 int __sched cond_resched(void)
3759 if (need_resched()) {
3766 EXPORT_SYMBOL(cond_resched);
3769 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
3770 * call schedule, and on return reacquire the lock.
3772 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
3773 * operations here to prevent schedule() from being called twice (once via
3774 * spin_unlock(), once by hand).
3776 int cond_resched_lock(spinlock_t * lock)
3780 if (need_lockbreak(lock)) {
3786 if (need_resched()) {
3787 _raw_spin_unlock(lock);
3788 preempt_enable_no_resched();
3796 EXPORT_SYMBOL(cond_resched_lock);
3798 int __sched cond_resched_softirq(void)
3800 BUG_ON(!in_softirq());
3802 if (need_resched()) {
3803 __local_bh_enable();
3811 EXPORT_SYMBOL(cond_resched_softirq);
3815 * yield - yield the current processor to other threads.
3817 * this is a shortcut for kernel-space yielding - it marks the
3818 * thread runnable and calls sys_sched_yield().
3820 void __sched yield(void)
3822 set_current_state(TASK_RUNNING);
3826 EXPORT_SYMBOL(yield);
3829 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
3830 * that process accounting knows that this is a task in IO wait state.
3832 * But don't do that if it is a deliberate, throttling IO wait (this task
3833 * has set its backing_dev_info: the queue against which it should throttle)
3835 void __sched io_schedule(void)
3837 struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id());
3839 atomic_inc(&rq->nr_iowait);
3841 atomic_dec(&rq->nr_iowait);
3844 EXPORT_SYMBOL(io_schedule);
3846 long __sched io_schedule_timeout(long timeout)
3848 struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id());
3851 atomic_inc(&rq->nr_iowait);
3852 ret = schedule_timeout(timeout);
3853 atomic_dec(&rq->nr_iowait);
3858 * sys_sched_get_priority_max - return maximum RT priority.
3859 * @policy: scheduling class.
3861 * this syscall returns the maximum rt_priority that can be used
3862 * by a given scheduling class.
3864 asmlinkage long sys_sched_get_priority_max(int policy)
3871 ret = MAX_USER_RT_PRIO-1;
3881 * sys_sched_get_priority_min - return minimum RT priority.
3882 * @policy: scheduling class.
3884 * this syscall returns the minimum rt_priority that can be used
3885 * by a given scheduling class.
3887 asmlinkage long sys_sched_get_priority_min(int policy)
3903 * sys_sched_rr_get_interval - return the default timeslice of a process.
3904 * @pid: pid of the process.
3905 * @interval: userspace pointer to the timeslice value.
3907 * this syscall writes the default timeslice value of a given process
3908 * into the user-space timespec buffer. A value of '0' means infinity.
3911 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
3913 int retval = -EINVAL;
3921 read_lock(&tasklist_lock);
3922 p = find_process_by_pid(pid);
3926 retval = security_task_getscheduler(p);
3930 jiffies_to_timespec(p->policy & SCHED_FIFO ?
3931 0 : task_timeslice(p), &t);
3932 read_unlock(&tasklist_lock);
3933 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
3937 read_unlock(&tasklist_lock);
3941 static inline struct task_struct *eldest_child(struct task_struct *p)
3943 if (list_empty(&p->children)) return NULL;
3944 return list_entry(p->children.next,struct task_struct,sibling);
3947 static inline struct task_struct *older_sibling(struct task_struct *p)
3949 if (p->sibling.prev==&p->parent->children) return NULL;
3950 return list_entry(p->sibling.prev,struct task_struct,sibling);
3953 static inline struct task_struct *younger_sibling(struct task_struct *p)
3955 if (p->sibling.next==&p->parent->children) return NULL;
3956 return list_entry(p->sibling.next,struct task_struct,sibling);
3959 static void show_task(task_t * p)
3963 unsigned long free = 0;
3964 static const char *stat_nam[] = { "R", "S", "D", "T", "t", "Z", "X" };
3966 printk("%-13.13s ", p->comm);
3967 state = p->state ? __ffs(p->state) + 1 : 0;
3968 if (state < ARRAY_SIZE(stat_nam))
3969 printk(stat_nam[state]);
3972 #if (BITS_PER_LONG == 32)
3973 if (state == TASK_RUNNING)
3974 printk(" running ");
3976 printk(" %08lX ", thread_saved_pc(p));
3978 if (state == TASK_RUNNING)
3979 printk(" running task ");
3981 printk(" %016lx ", thread_saved_pc(p));
3983 #ifdef CONFIG_DEBUG_STACK_USAGE
3985 unsigned long * n = (unsigned long *) (p->thread_info+1);
3988 free = (unsigned long) n - (unsigned long)(p->thread_info+1);
3991 printk("%5lu %5d %6d ", free, p->pid, p->parent->pid);
3992 if ((relative = eldest_child(p)))
3993 printk("%5d ", relative->pid);
3996 if ((relative = younger_sibling(p)))
3997 printk("%7d", relative->pid);
4000 if ((relative = older_sibling(p)))
4001 printk(" %5d", relative->pid);
4005 printk(" (L-TLB)\n");
4007 printk(" (NOTLB)\n");
4009 if (state != TASK_RUNNING)
4010 show_stack(p, NULL);
4013 void show_state(void)
4017 #if (BITS_PER_LONG == 32)
4020 printk(" task PC pid father child younger older\n");
4024 printk(" task PC pid father child younger older\n");
4026 read_lock(&tasklist_lock);
4027 do_each_thread(g, p) {
4029 * reset the NMI-timeout, listing all files on a slow
4030 * console might take alot of time:
4032 touch_nmi_watchdog();
4034 } while_each_thread(g, p);
4036 read_unlock(&tasklist_lock);
4039 void __devinit init_idle(task_t *idle, int cpu)
4041 runqueue_t *rq = cpu_rq(cpu);
4042 unsigned long flags;
4044 idle->sleep_avg = 0;
4046 idle->prio = MAX_PRIO;
4047 idle->state = TASK_RUNNING;
4048 idle->cpus_allowed = cpumask_of_cpu(cpu);
4049 set_task_cpu(idle, cpu);
4051 spin_lock_irqsave(&rq->lock, flags);
4052 rq->curr = rq->idle = idle;
4053 set_tsk_need_resched(idle);
4054 spin_unlock_irqrestore(&rq->lock, flags);
4056 /* Set the preempt count _outside_ the spinlocks! */
4057 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4058 idle->thread_info->preempt_count = (idle->lock_depth >= 0);
4060 idle->thread_info->preempt_count = 0;
4065 * In a system that switches off the HZ timer nohz_cpu_mask
4066 * indicates which cpus entered this state. This is used
4067 * in the rcu update to wait only for active cpus. For system
4068 * which do not switch off the HZ timer nohz_cpu_mask should
4069 * always be CPU_MASK_NONE.
4071 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4075 * This is how migration works:
4077 * 1) we queue a migration_req_t structure in the source CPU's
4078 * runqueue and wake up that CPU's migration thread.
4079 * 2) we down() the locked semaphore => thread blocks.
4080 * 3) migration thread wakes up (implicitly it forces the migrated
4081 * thread off the CPU)
4082 * 4) it gets the migration request and checks whether the migrated
4083 * task is still in the wrong runqueue.
4084 * 5) if it's in the wrong runqueue then the migration thread removes
4085 * it and puts it into the right queue.
4086 * 6) migration thread up()s the semaphore.
4087 * 7) we wake up and the migration is done.
4091 * Change a given task's CPU affinity. Migrate the thread to a
4092 * proper CPU and schedule it away if the CPU it's executing on
4093 * is removed from the allowed bitmask.
4095 * NOTE: the caller must have a valid reference to the task, the
4096 * task must not exit() & deallocate itself prematurely. The
4097 * call is not atomic; no spinlocks may be held.
4099 int set_cpus_allowed(task_t *p, cpumask_t new_mask)
4101 unsigned long flags;
4103 migration_req_t req;
4106 rq = task_rq_lock(p, &flags);
4107 if (!cpus_intersects(new_mask, cpu_online_map)) {
4112 p->cpus_allowed = new_mask;
4113 /* Can the task run on the task's current CPU? If so, we're done */
4114 if (cpu_isset(task_cpu(p), new_mask))
4117 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4118 /* Need help from migration thread: drop lock and wait. */
4119 task_rq_unlock(rq, &flags);
4120 wake_up_process(rq->migration_thread);
4121 wait_for_completion(&req.done);
4122 tlb_migrate_finish(p->mm);
4126 task_rq_unlock(rq, &flags);
4130 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4133 * Move (not current) task off this cpu, onto dest cpu. We're doing
4134 * this because either it can't run here any more (set_cpus_allowed()
4135 * away from this CPU, or CPU going down), or because we're
4136 * attempting to rebalance this task on exec (sched_exec).
4138 * So we race with normal scheduler movements, but that's OK, as long
4139 * as the task is no longer on this CPU.
4141 static void __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4143 runqueue_t *rq_dest, *rq_src;
4145 if (unlikely(cpu_is_offline(dest_cpu)))
4148 rq_src = cpu_rq(src_cpu);
4149 rq_dest = cpu_rq(dest_cpu);
4151 double_rq_lock(rq_src, rq_dest);
4152 /* Already moved. */
4153 if (task_cpu(p) != src_cpu)
4155 /* Affinity changed (again). */
4156 if (!cpu_isset(dest_cpu, p->cpus_allowed))
4159 set_task_cpu(p, dest_cpu);
4162 * Sync timestamp with rq_dest's before activating.
4163 * The same thing could be achieved by doing this step
4164 * afterwards, and pretending it was a local activate.
4165 * This way is cleaner and logically correct.
4167 p->timestamp = p->timestamp - rq_src->timestamp_last_tick
4168 + rq_dest->timestamp_last_tick;
4169 deactivate_task(p, rq_src);
4170 activate_task(p, rq_dest, 0);
4171 if (TASK_PREEMPTS_CURR(p, rq_dest))
4172 resched_task(rq_dest->curr);
4176 double_rq_unlock(rq_src, rq_dest);
4180 * migration_thread - this is a highprio system thread that performs
4181 * thread migration by bumping thread off CPU then 'pushing' onto
4184 static int migration_thread(void * data)
4187 int cpu = (long)data;
4190 BUG_ON(rq->migration_thread != current);
4192 set_current_state(TASK_INTERRUPTIBLE);
4193 while (!kthread_should_stop()) {
4194 struct list_head *head;
4195 migration_req_t *req;
4197 if (current->flags & PF_FREEZE)
4198 refrigerator(PF_FREEZE);
4200 spin_lock_irq(&rq->lock);
4202 if (cpu_is_offline(cpu)) {
4203 spin_unlock_irq(&rq->lock);
4207 if (rq->active_balance) {
4208 active_load_balance(rq, cpu);
4209 rq->active_balance = 0;
4212 head = &rq->migration_queue;
4214 if (list_empty(head)) {
4215 spin_unlock_irq(&rq->lock);
4217 set_current_state(TASK_INTERRUPTIBLE);
4220 req = list_entry(head->next, migration_req_t, list);
4221 list_del_init(head->next);
4223 if (req->type == REQ_MOVE_TASK) {
4224 spin_unlock(&rq->lock);
4225 __migrate_task(req->task, cpu, req->dest_cpu);
4227 } else if (req->type == REQ_SET_DOMAIN) {
4229 spin_unlock_irq(&rq->lock);
4231 spin_unlock_irq(&rq->lock);
4235 complete(&req->done);
4237 __set_current_state(TASK_RUNNING);
4241 /* Wait for kthread_stop */
4242 set_current_state(TASK_INTERRUPTIBLE);
4243 while (!kthread_should_stop()) {
4245 set_current_state(TASK_INTERRUPTIBLE);
4247 __set_current_state(TASK_RUNNING);
4251 #ifdef CONFIG_HOTPLUG_CPU
4252 /* Figure out where task on dead CPU should go, use force if neccessary. */
4253 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *tsk)
4259 mask = node_to_cpumask(cpu_to_node(dead_cpu));
4260 cpus_and(mask, mask, tsk->cpus_allowed);
4261 dest_cpu = any_online_cpu(mask);
4263 /* On any allowed CPU? */
4264 if (dest_cpu == NR_CPUS)
4265 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4267 /* No more Mr. Nice Guy. */
4268 if (dest_cpu == NR_CPUS) {
4269 cpus_setall(tsk->cpus_allowed);
4270 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4273 * Don't tell them about moving exiting tasks or
4274 * kernel threads (both mm NULL), since they never
4277 if (tsk->mm && printk_ratelimit())
4278 printk(KERN_INFO "process %d (%s) no "
4279 "longer affine to cpu%d\n",
4280 tsk->pid, tsk->comm, dead_cpu);
4282 __migrate_task(tsk, dead_cpu, dest_cpu);
4286 * While a dead CPU has no uninterruptible tasks queued at this point,
4287 * it might still have a nonzero ->nr_uninterruptible counter, because
4288 * for performance reasons the counter is not stricly tracking tasks to
4289 * their home CPUs. So we just add the counter to another CPU's counter,
4290 * to keep the global sum constant after CPU-down:
4292 static void migrate_nr_uninterruptible(runqueue_t *rq_src)
4294 runqueue_t *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
4295 unsigned long flags;
4297 local_irq_save(flags);
4298 double_rq_lock(rq_src, rq_dest);
4299 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
4300 rq_src->nr_uninterruptible = 0;
4301 double_rq_unlock(rq_src, rq_dest);
4302 local_irq_restore(flags);
4305 /* Run through task list and migrate tasks from the dead cpu. */
4306 static void migrate_live_tasks(int src_cpu)
4308 struct task_struct *tsk, *t;
4310 write_lock_irq(&tasklist_lock);
4312 do_each_thread(t, tsk) {
4316 if (task_cpu(tsk) == src_cpu)
4317 move_task_off_dead_cpu(src_cpu, tsk);
4318 } while_each_thread(t, tsk);
4320 write_unlock_irq(&tasklist_lock);
4323 /* Schedules idle task to be the next runnable task on current CPU.
4324 * It does so by boosting its priority to highest possible and adding it to
4325 * the _front_ of runqueue. Used by CPU offline code.
4327 void sched_idle_next(void)
4329 int cpu = smp_processor_id();
4330 runqueue_t *rq = this_rq();
4331 struct task_struct *p = rq->idle;
4332 unsigned long flags;
4334 /* cpu has to be offline */
4335 BUG_ON(cpu_online(cpu));
4337 /* Strictly not necessary since rest of the CPUs are stopped by now
4338 * and interrupts disabled on current cpu.
4340 spin_lock_irqsave(&rq->lock, flags);
4342 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4343 /* Add idle task to _front_ of it's priority queue */
4344 __activate_idle_task(p, rq);
4346 spin_unlock_irqrestore(&rq->lock, flags);
4349 /* Ensures that the idle task is using init_mm right before its cpu goes
4352 void idle_task_exit(void)
4354 struct mm_struct *mm = current->active_mm;
4356 BUG_ON(cpu_online(smp_processor_id()));
4359 switch_mm(mm, &init_mm, current);
4363 static void migrate_dead(unsigned int dead_cpu, task_t *tsk)
4365 struct runqueue *rq = cpu_rq(dead_cpu);
4367 /* Must be exiting, otherwise would be on tasklist. */
4368 BUG_ON(tsk->exit_state != EXIT_ZOMBIE && tsk->exit_state != EXIT_DEAD);
4370 /* Cannot have done final schedule yet: would have vanished. */
4371 BUG_ON(tsk->flags & PF_DEAD);
4373 get_task_struct(tsk);
4376 * Drop lock around migration; if someone else moves it,
4377 * that's OK. No task can be added to this CPU, so iteration is
4380 spin_unlock_irq(&rq->lock);
4381 move_task_off_dead_cpu(dead_cpu, tsk);
4382 spin_lock_irq(&rq->lock);
4384 put_task_struct(tsk);
4387 /* release_task() removes task from tasklist, so we won't find dead tasks. */
4388 static void migrate_dead_tasks(unsigned int dead_cpu)
4391 struct runqueue *rq = cpu_rq(dead_cpu);
4393 for (arr = 0; arr < 2; arr++) {
4394 for (i = 0; i < MAX_PRIO; i++) {
4395 struct list_head *list = &rq->arrays[arr].queue[i];
4396 while (!list_empty(list))
4397 migrate_dead(dead_cpu,
4398 list_entry(list->next, task_t,
4403 #endif /* CONFIG_HOTPLUG_CPU */
4406 * migration_call - callback that gets triggered when a CPU is added.
4407 * Here we can start up the necessary migration thread for the new CPU.
4409 static int migration_call(struct notifier_block *nfb, unsigned long action,
4412 int cpu = (long)hcpu;
4413 struct task_struct *p;
4414 struct runqueue *rq;
4415 unsigned long flags;
4418 case CPU_UP_PREPARE:
4419 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
4422 p->flags |= PF_NOFREEZE;
4423 kthread_bind(p, cpu);
4424 /* Must be high prio: stop_machine expects to yield to it. */
4425 rq = task_rq_lock(p, &flags);
4426 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4427 task_rq_unlock(rq, &flags);
4428 cpu_rq(cpu)->migration_thread = p;
4431 /* Strictly unneccessary, as first user will wake it. */
4432 wake_up_process(cpu_rq(cpu)->migration_thread);
4434 #ifdef CONFIG_HOTPLUG_CPU
4435 case CPU_UP_CANCELED:
4436 /* Unbind it from offline cpu so it can run. Fall thru. */
4437 kthread_bind(cpu_rq(cpu)->migration_thread,smp_processor_id());
4438 kthread_stop(cpu_rq(cpu)->migration_thread);
4439 cpu_rq(cpu)->migration_thread = NULL;
4442 migrate_live_tasks(cpu);
4444 kthread_stop(rq->migration_thread);
4445 rq->migration_thread = NULL;
4446 /* Idle task back to normal (off runqueue, low prio) */
4447 rq = task_rq_lock(rq->idle, &flags);
4448 deactivate_task(rq->idle, rq);
4449 rq->idle->static_prio = MAX_PRIO;
4450 __setscheduler(rq->idle, SCHED_NORMAL, 0);
4451 migrate_dead_tasks(cpu);
4452 task_rq_unlock(rq, &flags);
4453 migrate_nr_uninterruptible(rq);
4454 BUG_ON(rq->nr_running != 0);
4456 /* No need to migrate the tasks: it was best-effort if
4457 * they didn't do lock_cpu_hotplug(). Just wake up
4458 * the requestors. */
4459 spin_lock_irq(&rq->lock);
4460 while (!list_empty(&rq->migration_queue)) {
4461 migration_req_t *req;
4462 req = list_entry(rq->migration_queue.next,
4463 migration_req_t, list);
4464 BUG_ON(req->type != REQ_MOVE_TASK);
4465 list_del_init(&req->list);
4466 complete(&req->done);
4468 spin_unlock_irq(&rq->lock);
4475 /* Register at highest priority so that task migration (migrate_all_tasks)
4476 * happens before everything else.
4478 static struct notifier_block __devinitdata migration_notifier = {
4479 .notifier_call = migration_call,
4483 int __init migration_init(void)
4485 void *cpu = (void *)(long)smp_processor_id();
4486 /* Start one for boot CPU. */
4487 migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
4488 migration_call(&migration_notifier, CPU_ONLINE, cpu);
4489 register_cpu_notifier(&migration_notifier);
4495 #define SCHED_DOMAIN_DEBUG
4496 #ifdef SCHED_DOMAIN_DEBUG
4497 static void sched_domain_debug(struct sched_domain *sd, int cpu)
4501 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
4506 struct sched_group *group = sd->groups;
4507 cpumask_t groupmask;
4509 cpumask_scnprintf(str, NR_CPUS, sd->span);
4510 cpus_clear(groupmask);
4513 for (i = 0; i < level + 1; i++)
4515 printk("domain %d: ", level);
4517 if (!(sd->flags & SD_LOAD_BALANCE)) {
4518 printk("does not load-balance\n");
4520 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent");
4524 printk("span %s\n", str);
4526 if (!cpu_isset(cpu, sd->span))
4527 printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
4528 if (!cpu_isset(cpu, group->cpumask))
4529 printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
4532 for (i = 0; i < level + 2; i++)
4538 printk(KERN_ERR "ERROR: group is NULL\n");
4542 if (!group->cpu_power) {
4544 printk(KERN_ERR "ERROR: domain->cpu_power not set\n");
4547 if (!cpus_weight(group->cpumask)) {
4549 printk(KERN_ERR "ERROR: empty group\n");
4552 if (cpus_intersects(groupmask, group->cpumask)) {
4554 printk(KERN_ERR "ERROR: repeated CPUs\n");
4557 cpus_or(groupmask, groupmask, group->cpumask);
4559 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
4562 group = group->next;
4563 } while (group != sd->groups);
4566 if (!cpus_equal(sd->span, groupmask))
4567 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
4573 if (!cpus_subset(groupmask, sd->span))
4574 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
4580 #define sched_domain_debug(sd, cpu) {}
4584 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
4585 * hold the hotplug lock.
4587 void __devinit cpu_attach_domain(struct sched_domain *sd, int cpu)
4589 migration_req_t req;
4590 unsigned long flags;
4591 runqueue_t *rq = cpu_rq(cpu);
4594 sched_domain_debug(sd, cpu);
4596 spin_lock_irqsave(&rq->lock, flags);
4598 if (cpu == smp_processor_id() || !cpu_online(cpu)) {
4601 init_completion(&req.done);
4602 req.type = REQ_SET_DOMAIN;
4604 list_add(&req.list, &rq->migration_queue);
4608 spin_unlock_irqrestore(&rq->lock, flags);
4611 wake_up_process(rq->migration_thread);
4612 wait_for_completion(&req.done);
4616 /* cpus with isolated domains */
4617 cpumask_t __devinitdata cpu_isolated_map = CPU_MASK_NONE;
4619 /* Setup the mask of cpus configured for isolated domains */
4620 static int __init isolated_cpu_setup(char *str)
4622 int ints[NR_CPUS], i;
4624 str = get_options(str, ARRAY_SIZE(ints), ints);
4625 cpus_clear(cpu_isolated_map);
4626 for (i = 1; i <= ints[0]; i++)
4627 if (ints[i] < NR_CPUS)
4628 cpu_set(ints[i], cpu_isolated_map);
4632 __setup ("isolcpus=", isolated_cpu_setup);
4635 * init_sched_build_groups takes an array of groups, the cpumask we wish
4636 * to span, and a pointer to a function which identifies what group a CPU
4637 * belongs to. The return value of group_fn must be a valid index into the
4638 * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
4639 * keep track of groups covered with a cpumask_t).
4641 * init_sched_build_groups will build a circular linked list of the groups
4642 * covered by the given span, and will set each group's ->cpumask correctly,
4643 * and ->cpu_power to 0.
4645 void __devinit init_sched_build_groups(struct sched_group groups[],
4646 cpumask_t span, int (*group_fn)(int cpu))
4648 struct sched_group *first = NULL, *last = NULL;
4649 cpumask_t covered = CPU_MASK_NONE;
4652 for_each_cpu_mask(i, span) {
4653 int group = group_fn(i);
4654 struct sched_group *sg = &groups[group];
4657 if (cpu_isset(i, covered))
4660 sg->cpumask = CPU_MASK_NONE;
4663 for_each_cpu_mask(j, span) {
4664 if (group_fn(j) != group)
4667 cpu_set(j, covered);
4668 cpu_set(j, sg->cpumask);
4680 #ifdef ARCH_HAS_SCHED_DOMAIN
4681 extern void __devinit arch_init_sched_domains(void);
4682 extern void __devinit arch_destroy_sched_domains(void);
4684 #ifdef CONFIG_SCHED_SMT
4685 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
4686 static struct sched_group sched_group_cpus[NR_CPUS];
4687 static int __devinit cpu_to_cpu_group(int cpu)
4693 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
4694 static struct sched_group sched_group_phys[NR_CPUS];
4695 static int __devinit cpu_to_phys_group(int cpu)
4697 #ifdef CONFIG_SCHED_SMT
4698 return first_cpu(cpu_sibling_map[cpu]);
4706 static DEFINE_PER_CPU(struct sched_domain, node_domains);
4707 static struct sched_group sched_group_nodes[MAX_NUMNODES];
4708 static int __devinit cpu_to_node_group(int cpu)
4710 return cpu_to_node(cpu);
4714 #if defined(CONFIG_SCHED_SMT) && defined(CONFIG_NUMA)
4716 * The domains setup code relies on siblings not spanning
4717 * multiple nodes. Make sure the architecture has a proper
4720 static void check_sibling_maps(void)
4724 for_each_online_cpu(i) {
4725 for_each_cpu_mask(j, cpu_sibling_map[i]) {
4726 if (cpu_to_node(i) != cpu_to_node(j)) {
4727 printk(KERN_INFO "warning: CPU %d siblings map "
4728 "to different node - isolating "
4730 cpu_sibling_map[i] = cpumask_of_cpu(i);
4739 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
4741 static void __devinit arch_init_sched_domains(void)
4744 cpumask_t cpu_default_map;
4746 #if defined(CONFIG_SCHED_SMT) && defined(CONFIG_NUMA)
4747 check_sibling_maps();
4750 * Setup mask for cpus without special case scheduling requirements.
4751 * For now this just excludes isolated cpus, but could be used to
4752 * exclude other special cases in the future.
4754 cpus_complement(cpu_default_map, cpu_isolated_map);
4755 cpus_and(cpu_default_map, cpu_default_map, cpu_online_map);
4758 * Set up domains. Isolated domains just stay on the dummy domain.
4760 for_each_cpu_mask(i, cpu_default_map) {
4762 struct sched_domain *sd = NULL, *p;
4763 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
4765 cpus_and(nodemask, nodemask, cpu_default_map);
4768 sd = &per_cpu(node_domains, i);
4769 group = cpu_to_node_group(i);
4771 sd->span = cpu_default_map;
4772 sd->groups = &sched_group_nodes[group];
4776 sd = &per_cpu(phys_domains, i);
4777 group = cpu_to_phys_group(i);
4779 sd->span = nodemask;
4781 sd->groups = &sched_group_phys[group];
4783 #ifdef CONFIG_SCHED_SMT
4785 sd = &per_cpu(cpu_domains, i);
4786 group = cpu_to_cpu_group(i);
4787 *sd = SD_SIBLING_INIT;
4788 sd->span = cpu_sibling_map[i];
4789 cpus_and(sd->span, sd->span, cpu_default_map);
4791 sd->groups = &sched_group_cpus[group];
4795 #ifdef CONFIG_SCHED_SMT
4796 /* Set up CPU (sibling) groups */
4797 for_each_online_cpu(i) {
4798 cpumask_t this_sibling_map = cpu_sibling_map[i];
4799 cpus_and(this_sibling_map, this_sibling_map, cpu_default_map);
4800 if (i != first_cpu(this_sibling_map))
4803 init_sched_build_groups(sched_group_cpus, this_sibling_map,
4808 /* Set up physical groups */
4809 for (i = 0; i < MAX_NUMNODES; i++) {
4810 cpumask_t nodemask = node_to_cpumask(i);
4812 cpus_and(nodemask, nodemask, cpu_default_map);
4813 if (cpus_empty(nodemask))
4816 init_sched_build_groups(sched_group_phys, nodemask,
4817 &cpu_to_phys_group);
4821 /* Set up node groups */
4822 init_sched_build_groups(sched_group_nodes, cpu_default_map,
4823 &cpu_to_node_group);
4826 /* Calculate CPU power for physical packages and nodes */
4827 for_each_cpu_mask(i, cpu_default_map) {
4829 struct sched_domain *sd;
4830 #ifdef CONFIG_SCHED_SMT
4831 sd = &per_cpu(cpu_domains, i);
4832 power = SCHED_LOAD_SCALE;
4833 sd->groups->cpu_power = power;
4836 sd = &per_cpu(phys_domains, i);
4837 power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
4838 (cpus_weight(sd->groups->cpumask)-1) / 10;
4839 sd->groups->cpu_power = power;
4842 if (i == first_cpu(sd->groups->cpumask)) {
4843 /* Only add "power" once for each physical package. */
4844 sd = &per_cpu(node_domains, i);
4845 sd->groups->cpu_power += power;
4850 /* Attach the domains */
4851 for_each_online_cpu(i) {
4852 struct sched_domain *sd;
4853 #ifdef CONFIG_SCHED_SMT
4854 sd = &per_cpu(cpu_domains, i);
4856 sd = &per_cpu(phys_domains, i);
4858 cpu_attach_domain(sd, i);
4862 #ifdef CONFIG_HOTPLUG_CPU
4863 static void __devinit arch_destroy_sched_domains(void)
4865 /* Do nothing: everything is statically allocated. */
4869 #endif /* ARCH_HAS_SCHED_DOMAIN */
4872 * Initial dummy domain for early boot and for hotplug cpu. Being static,
4873 * it is initialized to zero, so all balancing flags are cleared which is
4876 static struct sched_domain sched_domain_dummy;
4878 #ifdef CONFIG_HOTPLUG_CPU
4880 * Force a reinitialization of the sched domains hierarchy. The domains
4881 * and groups cannot be updated in place without racing with the balancing
4882 * code, so we temporarily attach all running cpus to a "dummy" domain
4883 * which will prevent rebalancing while the sched domains are recalculated.
4885 static int update_sched_domains(struct notifier_block *nfb,
4886 unsigned long action, void *hcpu)
4891 case CPU_UP_PREPARE:
4892 case CPU_DOWN_PREPARE:
4893 for_each_online_cpu(i)
4894 cpu_attach_domain(&sched_domain_dummy, i);
4895 arch_destroy_sched_domains();
4898 case CPU_UP_CANCELED:
4899 case CPU_DOWN_FAILED:
4903 * Fall through and re-initialise the domains.
4910 /* The hotplug lock is already held by cpu_up/cpu_down */
4911 arch_init_sched_domains();
4917 void __init sched_init_smp(void)
4920 arch_init_sched_domains();
4921 unlock_cpu_hotplug();
4922 /* XXX: Theoretical race here - CPU may be hotplugged now */
4923 hotcpu_notifier(update_sched_domains, 0);
4926 void __init sched_init_smp(void)
4929 #endif /* CONFIG_SMP */
4931 int in_sched_functions(unsigned long addr)
4933 /* Linker adds these: start and end of __sched functions */
4934 extern char __sched_text_start[], __sched_text_end[];
4935 return in_lock_functions(addr) ||
4936 (addr >= (unsigned long)__sched_text_start
4937 && addr < (unsigned long)__sched_text_end);
4940 void __init sched_init(void)
4945 for (i = 0; i < NR_CPUS; i++) {
4946 prio_array_t *array;
4949 spin_lock_init(&rq->lock);
4950 rq->active = rq->arrays;
4951 rq->expired = rq->arrays + 1;
4952 rq->best_expired_prio = MAX_PRIO;
4955 rq->sd = &sched_domain_dummy;
4957 rq->active_balance = 0;
4959 rq->migration_thread = NULL;
4960 INIT_LIST_HEAD(&rq->migration_queue);
4962 atomic_set(&rq->nr_iowait, 0);
4964 for (j = 0; j < 2; j++) {
4965 array = rq->arrays + j;
4966 for (k = 0; k < MAX_PRIO; k++) {
4967 INIT_LIST_HEAD(array->queue + k);
4968 __clear_bit(k, array->bitmap);
4970 // delimiter for bitsearch
4971 __set_bit(MAX_PRIO, array->bitmap);
4976 * The boot idle thread does lazy MMU switching as well:
4978 atomic_inc(&init_mm.mm_count);
4979 enter_lazy_tlb(&init_mm, current);
4982 * Make us the idle thread. Technically, schedule() should not be
4983 * called from this thread, however somewhere below it might be,
4984 * but because we are the idle thread, we just pick up running again
4985 * when this runqueue becomes "idle".
4987 init_idle(current, smp_processor_id());
4990 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4991 void __might_sleep(char *file, int line)
4993 #if defined(in_atomic)
4994 static unsigned long prev_jiffy; /* ratelimiting */
4996 if ((in_atomic() || irqs_disabled()) &&
4997 system_state == SYSTEM_RUNNING && !oops_in_progress) {
4998 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
5000 prev_jiffy = jiffies;
5001 printk(KERN_ERR "Debug: sleeping function called from invalid"
5002 " context at %s:%d\n", file, line);
5003 printk("in_atomic():%d, irqs_disabled():%d\n",
5004 in_atomic(), irqs_disabled());
5009 EXPORT_SYMBOL(__might_sleep);
5012 #ifdef CONFIG_MAGIC_SYSRQ
5013 void normalize_rt_tasks(void)
5015 struct task_struct *p;
5016 prio_array_t *array;
5017 unsigned long flags;
5020 read_lock_irq(&tasklist_lock);
5021 for_each_process (p) {
5025 rq = task_rq_lock(p, &flags);
5029 deactivate_task(p, task_rq(p));
5030 __setscheduler(p, SCHED_NORMAL, 0);
5032 __activate_task(p, task_rq(p));
5033 resched_task(rq->curr);
5036 task_rq_unlock(rq, &flags);
5038 read_unlock_irq(&tasklist_lock);
5041 #endif /* CONFIG_MAGIC_SYSRQ */