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 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[3];
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)
271 #ifndef prepare_arch_switch
272 # define prepare_arch_switch(next) do { } while (0)
274 #ifndef finish_arch_switch
275 # define finish_arch_switch(prev) do { } while (0)
278 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
279 static inline int task_running(runqueue_t *rq, task_t *p)
281 return rq->curr == p;
284 static inline void prepare_lock_switch(runqueue_t *rq, task_t *next)
288 static inline void finish_lock_switch(runqueue_t *rq, task_t *prev)
290 spin_unlock_irq(&rq->lock);
293 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
294 static inline int task_running(runqueue_t *rq, task_t *p)
299 return rq->curr == p;
303 static inline void prepare_lock_switch(runqueue_t *rq, task_t *next)
307 * We can optimise this out completely for !SMP, because the
308 * SMP rebalancing from interrupt is the only thing that cares
313 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
314 spin_unlock_irq(&rq->lock);
316 spin_unlock(&rq->lock);
320 static inline void finish_lock_switch(runqueue_t *rq, task_t *prev)
324 * After ->oncpu is cleared, the task can be moved to a different CPU.
325 * We must ensure this doesn't happen until the switch is completely
331 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
335 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
338 * task_rq_lock - lock the runqueue a given task resides on and disable
339 * interrupts. Note the ordering: we can safely lookup the task_rq without
340 * explicitly disabling preemption.
342 static inline runqueue_t *task_rq_lock(task_t *p, unsigned long *flags)
348 local_irq_save(*flags);
350 spin_lock(&rq->lock);
351 if (unlikely(rq != task_rq(p))) {
352 spin_unlock_irqrestore(&rq->lock, *flags);
353 goto repeat_lock_task;
358 static inline void task_rq_unlock(runqueue_t *rq, unsigned long *flags)
361 spin_unlock_irqrestore(&rq->lock, *flags);
364 #ifdef CONFIG_SCHEDSTATS
366 * bump this up when changing the output format or the meaning of an existing
367 * format, so that tools can adapt (or abort)
369 #define SCHEDSTAT_VERSION 12
371 static int show_schedstat(struct seq_file *seq, void *v)
375 seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION);
376 seq_printf(seq, "timestamp %lu\n", jiffies);
377 for_each_online_cpu(cpu) {
378 runqueue_t *rq = cpu_rq(cpu);
380 struct sched_domain *sd;
384 /* runqueue-specific stats */
386 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
387 cpu, rq->yld_both_empty,
388 rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt,
389 rq->sched_switch, rq->sched_cnt, rq->sched_goidle,
390 rq->ttwu_cnt, rq->ttwu_local,
391 rq->rq_sched_info.cpu_time,
392 rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt);
394 seq_printf(seq, "\n");
397 /* domain-specific stats */
398 for_each_domain(cpu, sd) {
399 enum idle_type itype;
400 char mask_str[NR_CPUS];
402 cpumask_scnprintf(mask_str, NR_CPUS, sd->span);
403 seq_printf(seq, "domain%d %s", dcnt++, mask_str);
404 for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES;
406 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu",
408 sd->lb_balanced[itype],
409 sd->lb_failed[itype],
410 sd->lb_imbalance[itype],
411 sd->lb_gained[itype],
412 sd->lb_hot_gained[itype],
413 sd->lb_nobusyq[itype],
414 sd->lb_nobusyg[itype]);
416 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
417 sd->alb_cnt, sd->alb_failed, sd->alb_pushed,
418 sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed,
419 sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed,
420 sd->ttwu_wake_remote, sd->ttwu_move_affine, sd->ttwu_move_balance);
427 static int schedstat_open(struct inode *inode, struct file *file)
429 unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32);
430 char *buf = kmalloc(size, GFP_KERNEL);
436 res = single_open(file, show_schedstat, NULL);
438 m = file->private_data;
446 struct file_operations proc_schedstat_operations = {
447 .open = schedstat_open,
450 .release = single_release,
453 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
454 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
455 #else /* !CONFIG_SCHEDSTATS */
456 # define schedstat_inc(rq, field) do { } while (0)
457 # define schedstat_add(rq, field, amt) do { } while (0)
461 * rq_lock - lock a given runqueue and disable interrupts.
463 static inline runqueue_t *this_rq_lock(void)
470 spin_lock(&rq->lock);
475 #ifdef CONFIG_SCHEDSTATS
477 * Called when a process is dequeued from the active array and given
478 * the cpu. We should note that with the exception of interactive
479 * tasks, the expired queue will become the active queue after the active
480 * queue is empty, without explicitly dequeuing and requeuing tasks in the
481 * expired queue. (Interactive tasks may be requeued directly to the
482 * active queue, thus delaying tasks in the expired queue from running;
483 * see scheduler_tick()).
485 * This function is only called from sched_info_arrive(), rather than
486 * dequeue_task(). Even though a task may be queued and dequeued multiple
487 * times as it is shuffled about, we're really interested in knowing how
488 * long it was from the *first* time it was queued to the time that it
491 static inline void sched_info_dequeued(task_t *t)
493 t->sched_info.last_queued = 0;
497 * Called when a task finally hits the cpu. We can now calculate how
498 * long it was waiting to run. We also note when it began so that we
499 * can keep stats on how long its timeslice is.
501 static inline void sched_info_arrive(task_t *t)
503 unsigned long now = jiffies, diff = 0;
504 struct runqueue *rq = task_rq(t);
506 if (t->sched_info.last_queued)
507 diff = now - t->sched_info.last_queued;
508 sched_info_dequeued(t);
509 t->sched_info.run_delay += diff;
510 t->sched_info.last_arrival = now;
511 t->sched_info.pcnt++;
516 rq->rq_sched_info.run_delay += diff;
517 rq->rq_sched_info.pcnt++;
521 * Called when a process is queued into either the active or expired
522 * array. The time is noted and later used to determine how long we
523 * had to wait for us to reach the cpu. Since the expired queue will
524 * become the active queue after active queue is empty, without dequeuing
525 * and requeuing any tasks, we are interested in queuing to either. It
526 * is unusual but not impossible for tasks to be dequeued and immediately
527 * requeued in the same or another array: this can happen in sched_yield(),
528 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
531 * This function is only called from enqueue_task(), but also only updates
532 * the timestamp if it is already not set. It's assumed that
533 * sched_info_dequeued() will clear that stamp when appropriate.
535 static inline void sched_info_queued(task_t *t)
537 if (!t->sched_info.last_queued)
538 t->sched_info.last_queued = jiffies;
542 * Called when a process ceases being the active-running process, either
543 * voluntarily or involuntarily. Now we can calculate how long we ran.
545 static inline void sched_info_depart(task_t *t)
547 struct runqueue *rq = task_rq(t);
548 unsigned long diff = jiffies - t->sched_info.last_arrival;
550 t->sched_info.cpu_time += diff;
553 rq->rq_sched_info.cpu_time += diff;
557 * Called when tasks are switched involuntarily due, typically, to expiring
558 * their time slice. (This may also be called when switching to or from
559 * the idle task.) We are only called when prev != next.
561 static inline void sched_info_switch(task_t *prev, task_t *next)
563 struct runqueue *rq = task_rq(prev);
566 * prev now departs the cpu. It's not interesting to record
567 * stats about how efficient we were at scheduling the idle
570 if (prev != rq->idle)
571 sched_info_depart(prev);
573 if (next != rq->idle)
574 sched_info_arrive(next);
577 #define sched_info_queued(t) do { } while (0)
578 #define sched_info_switch(t, next) do { } while (0)
579 #endif /* CONFIG_SCHEDSTATS */
582 * Adding/removing a task to/from a priority array:
584 static void dequeue_task(struct task_struct *p, prio_array_t *array)
587 list_del(&p->run_list);
588 if (list_empty(array->queue + p->prio))
589 __clear_bit(p->prio, array->bitmap);
592 static void enqueue_task(struct task_struct *p, prio_array_t *array)
594 sched_info_queued(p);
595 list_add_tail(&p->run_list, array->queue + p->prio);
596 __set_bit(p->prio, array->bitmap);
602 * Put task to the end of the run list without the overhead of dequeue
603 * followed by enqueue.
605 static void requeue_task(struct task_struct *p, prio_array_t *array)
607 list_move_tail(&p->run_list, array->queue + p->prio);
610 static inline void enqueue_task_head(struct task_struct *p, prio_array_t *array)
612 list_add(&p->run_list, array->queue + p->prio);
613 __set_bit(p->prio, array->bitmap);
619 * effective_prio - return the priority that is based on the static
620 * priority but is modified by bonuses/penalties.
622 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
623 * into the -5 ... 0 ... +5 bonus/penalty range.
625 * We use 25% of the full 0...39 priority range so that:
627 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
628 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
630 * Both properties are important to certain workloads.
632 static int effective_prio(task_t *p)
639 bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
641 prio = p->static_prio - bonus;
642 if (prio < MAX_RT_PRIO)
644 if (prio > MAX_PRIO-1)
650 * __activate_task - move a task to the runqueue.
652 static inline void __activate_task(task_t *p, runqueue_t *rq)
654 enqueue_task(p, rq->active);
659 * __activate_idle_task - move idle task to the _front_ of runqueue.
661 static inline void __activate_idle_task(task_t *p, runqueue_t *rq)
663 enqueue_task_head(p, rq->active);
667 static void recalc_task_prio(task_t *p, unsigned long long now)
669 /* Caller must always ensure 'now >= p->timestamp' */
670 unsigned long long __sleep_time = now - p->timestamp;
671 unsigned long sleep_time;
673 if (__sleep_time > NS_MAX_SLEEP_AVG)
674 sleep_time = NS_MAX_SLEEP_AVG;
676 sleep_time = (unsigned long)__sleep_time;
678 if (likely(sleep_time > 0)) {
680 * User tasks that sleep a long time are categorised as
681 * idle and will get just interactive status to stay active &
682 * prevent them suddenly becoming cpu hogs and starving
685 if (p->mm && p->activated != -1 &&
686 sleep_time > INTERACTIVE_SLEEP(p)) {
687 p->sleep_avg = JIFFIES_TO_NS(MAX_SLEEP_AVG -
691 * The lower the sleep avg a task has the more
692 * rapidly it will rise with sleep time.
694 sleep_time *= (MAX_BONUS - CURRENT_BONUS(p)) ? : 1;
697 * Tasks waking from uninterruptible sleep are
698 * limited in their sleep_avg rise as they
699 * are likely to be waiting on I/O
701 if (p->activated == -1 && p->mm) {
702 if (p->sleep_avg >= INTERACTIVE_SLEEP(p))
704 else if (p->sleep_avg + sleep_time >=
705 INTERACTIVE_SLEEP(p)) {
706 p->sleep_avg = INTERACTIVE_SLEEP(p);
712 * This code gives a bonus to interactive tasks.
714 * The boost works by updating the 'average sleep time'
715 * value here, based on ->timestamp. The more time a
716 * task spends sleeping, the higher the average gets -
717 * and the higher the priority boost gets as well.
719 p->sleep_avg += sleep_time;
721 if (p->sleep_avg > NS_MAX_SLEEP_AVG)
722 p->sleep_avg = NS_MAX_SLEEP_AVG;
726 p->prio = effective_prio(p);
730 * activate_task - move a task to the runqueue and do priority recalculation
732 * Update all the scheduling statistics stuff. (sleep average
733 * calculation, priority modifiers, etc.)
735 static void activate_task(task_t *p, runqueue_t *rq, int local)
737 unsigned long long now;
742 /* Compensate for drifting sched_clock */
743 runqueue_t *this_rq = this_rq();
744 now = (now - this_rq->timestamp_last_tick)
745 + rq->timestamp_last_tick;
749 recalc_task_prio(p, now);
752 * This checks to make sure it's not an uninterruptible task
753 * that is now waking up.
757 * Tasks which were woken up by interrupts (ie. hw events)
758 * are most likely of interactive nature. So we give them
759 * the credit of extending their sleep time to the period
760 * of time they spend on the runqueue, waiting for execution
761 * on a CPU, first time around:
767 * Normal first-time wakeups get a credit too for
768 * on-runqueue time, but it will be weighted down:
775 __activate_task(p, rq);
779 * deactivate_task - remove a task from the runqueue.
781 static void deactivate_task(struct task_struct *p, runqueue_t *rq)
784 dequeue_task(p, p->array);
789 * resched_task - mark a task 'to be rescheduled now'.
791 * On UP this means the setting of the need_resched flag, on SMP it
792 * might also involve a cross-CPU call to trigger the scheduler on
796 static void resched_task(task_t *p)
798 int need_resched, nrpolling;
800 assert_spin_locked(&task_rq(p)->lock);
802 /* minimise the chance of sending an interrupt to poll_idle() */
803 nrpolling = test_tsk_thread_flag(p,TIF_POLLING_NRFLAG);
804 need_resched = test_and_set_tsk_thread_flag(p,TIF_NEED_RESCHED);
805 nrpolling |= test_tsk_thread_flag(p,TIF_POLLING_NRFLAG);
807 if (!need_resched && !nrpolling && (task_cpu(p) != smp_processor_id()))
808 smp_send_reschedule(task_cpu(p));
811 static inline void resched_task(task_t *p)
813 set_tsk_need_resched(p);
818 * task_curr - is this task currently executing on a CPU?
819 * @p: the task in question.
821 inline int task_curr(const task_t *p)
823 return cpu_curr(task_cpu(p)) == p;
833 struct list_head list;
834 enum request_type type;
836 /* For REQ_MOVE_TASK */
840 /* For REQ_SET_DOMAIN */
841 struct sched_domain *sd;
843 struct completion done;
847 * The task's runqueue lock must be held.
848 * Returns true if you have to wait for migration thread.
850 static int migrate_task(task_t *p, int dest_cpu, migration_req_t *req)
852 runqueue_t *rq = task_rq(p);
855 * If the task is not on a runqueue (and not running), then
856 * it is sufficient to simply update the task's cpu field.
858 if (!p->array && !task_running(rq, p)) {
859 set_task_cpu(p, dest_cpu);
863 init_completion(&req->done);
864 req->type = REQ_MOVE_TASK;
866 req->dest_cpu = dest_cpu;
867 list_add(&req->list, &rq->migration_queue);
872 * wait_task_inactive - wait for a thread to unschedule.
874 * The caller must ensure that the task *will* unschedule sometime soon,
875 * else this function might spin for a *long* time. This function can't
876 * be called with interrupts off, or it may introduce deadlock with
877 * smp_call_function() if an IPI is sent by the same process we are
878 * waiting to become inactive.
880 void wait_task_inactive(task_t * p)
887 rq = task_rq_lock(p, &flags);
888 /* Must be off runqueue entirely, not preempted. */
889 if (unlikely(p->array || task_running(rq, p))) {
890 /* If it's preempted, we yield. It could be a while. */
891 preempted = !task_running(rq, p);
892 task_rq_unlock(rq, &flags);
898 task_rq_unlock(rq, &flags);
902 * kick_process - kick a running thread to enter/exit the kernel
903 * @p: the to-be-kicked thread
905 * Cause a process which is running on another CPU to enter
906 * kernel-mode, without any delay. (to get signals handled.)
908 * NOTE: this function doesnt have to take the runqueue lock,
909 * because all it wants to ensure is that the remote task enters
910 * the kernel. If the IPI races and the task has been migrated
911 * to another CPU then no harm is done and the purpose has been
914 void kick_process(task_t *p)
920 if ((cpu != smp_processor_id()) && task_curr(p))
921 smp_send_reschedule(cpu);
926 * Return a low guess at the load of a migration-source cpu.
928 * We want to under-estimate the load of migration sources, to
929 * balance conservatively.
931 static inline unsigned long source_load(int cpu, int type)
933 runqueue_t *rq = cpu_rq(cpu);
934 unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE;
938 return min(rq->cpu_load[type-1], load_now);
942 * Return a high guess at the load of a migration-target cpu
944 static inline unsigned long target_load(int cpu, int type)
946 runqueue_t *rq = cpu_rq(cpu);
947 unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE;
951 return max(rq->cpu_load[type-1], load_now);
955 * find_idlest_group finds and returns the least busy CPU group within the
958 static struct sched_group *
959 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
961 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
962 unsigned long min_load = ULONG_MAX, this_load = 0;
963 int load_idx = sd->forkexec_idx;
964 int imbalance = 100 + (sd->imbalance_pct-100)/2;
967 unsigned long load, avg_load;
971 local_group = cpu_isset(this_cpu, group->cpumask);
972 /* XXX: put a cpus allowed check */
974 /* Tally up the load of all CPUs in the group */
977 for_each_cpu_mask(i, group->cpumask) {
978 /* Bias balancing toward cpus of our domain */
980 load = source_load(i, load_idx);
982 load = target_load(i, load_idx);
987 /* Adjust by relative CPU power of the group */
988 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
991 this_load = avg_load;
993 } else if (avg_load < min_load) {
998 } while (group != sd->groups);
1000 if (!idlest || 100*this_load < imbalance*min_load)
1006 * find_idlest_queue - find the idlest runqueue among the cpus in group.
1008 static int find_idlest_cpu(struct sched_group *group, int this_cpu)
1010 unsigned long load, min_load = ULONG_MAX;
1014 for_each_cpu_mask(i, group->cpumask) {
1015 load = source_load(i, 0);
1017 if (load < min_load || (load == min_load && i == this_cpu)) {
1030 * wake_idle() will wake a task on an idle cpu if task->cpu is
1031 * not idle and an idle cpu is available. The span of cpus to
1032 * search starts with cpus closest then further out as needed,
1033 * so we always favor a closer, idle cpu.
1035 * Returns the CPU we should wake onto.
1037 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1038 static int wake_idle(int cpu, task_t *p)
1041 struct sched_domain *sd;
1047 for_each_domain(cpu, sd) {
1048 if (sd->flags & SD_WAKE_IDLE) {
1049 cpus_and(tmp, sd->span, p->cpus_allowed);
1050 for_each_cpu_mask(i, tmp) {
1061 static inline int wake_idle(int cpu, task_t *p)
1068 * try_to_wake_up - wake up a thread
1069 * @p: the to-be-woken-up thread
1070 * @state: the mask of task states that can be woken
1071 * @sync: do a synchronous wakeup?
1073 * Put it on the run-queue if it's not already there. The "current"
1074 * thread is always on the run-queue (except when the actual
1075 * re-schedule is in progress), and as such you're allowed to do
1076 * the simpler "current->state = TASK_RUNNING" to mark yourself
1077 * runnable without the overhead of this.
1079 * returns failure only if the task is already active.
1081 static int try_to_wake_up(task_t * p, unsigned int state, int sync)
1083 int cpu, this_cpu, success = 0;
1084 unsigned long flags;
1088 unsigned long load, this_load;
1089 struct sched_domain *sd, *this_sd = NULL;
1093 rq = task_rq_lock(p, &flags);
1094 old_state = p->state;
1095 if (!(old_state & state))
1102 this_cpu = smp_processor_id();
1105 if (unlikely(task_running(rq, p)))
1110 schedstat_inc(rq, ttwu_cnt);
1111 if (cpu == this_cpu) {
1112 schedstat_inc(rq, ttwu_local);
1116 for_each_domain(this_cpu, sd) {
1117 if (cpu_isset(cpu, sd->span)) {
1118 schedstat_inc(sd, ttwu_wake_remote);
1124 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1128 * Check for affine wakeup and passive balancing possibilities.
1131 int idx = this_sd->wake_idx;
1132 unsigned int imbalance;
1134 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1136 load = source_load(cpu, idx);
1137 this_load = target_load(this_cpu, idx);
1139 new_cpu = this_cpu; /* Wake to this CPU if we can */
1141 if (this_sd->flags & SD_WAKE_AFFINE) {
1142 unsigned long tl = this_load;
1144 * If sync wakeup then subtract the (maximum possible)
1145 * effect of the currently running task from the load
1146 * of the current CPU:
1149 tl -= SCHED_LOAD_SCALE;
1152 tl + target_load(cpu, idx) <= SCHED_LOAD_SCALE) ||
1153 100*(tl + SCHED_LOAD_SCALE) <= imbalance*load) {
1155 * This domain has SD_WAKE_AFFINE and
1156 * p is cache cold in this domain, and
1157 * there is no bad imbalance.
1159 schedstat_inc(this_sd, ttwu_move_affine);
1165 * Start passive balancing when half the imbalance_pct
1168 if (this_sd->flags & SD_WAKE_BALANCE) {
1169 if (imbalance*this_load <= 100*load) {
1170 schedstat_inc(this_sd, ttwu_move_balance);
1176 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1178 new_cpu = wake_idle(new_cpu, p);
1179 if (new_cpu != cpu) {
1180 set_task_cpu(p, new_cpu);
1181 task_rq_unlock(rq, &flags);
1182 /* might preempt at this point */
1183 rq = task_rq_lock(p, &flags);
1184 old_state = p->state;
1185 if (!(old_state & state))
1190 this_cpu = smp_processor_id();
1195 #endif /* CONFIG_SMP */
1196 if (old_state == TASK_UNINTERRUPTIBLE) {
1197 rq->nr_uninterruptible--;
1199 * Tasks on involuntary sleep don't earn
1200 * sleep_avg beyond just interactive state.
1206 * Sync wakeups (i.e. those types of wakeups where the waker
1207 * has indicated that it will leave the CPU in short order)
1208 * don't trigger a preemption, if the woken up task will run on
1209 * this cpu. (in this case the 'I will reschedule' promise of
1210 * the waker guarantees that the freshly woken up task is going
1211 * to be considered on this CPU.)
1213 activate_task(p, rq, cpu == this_cpu);
1214 if (!sync || cpu != this_cpu) {
1215 if (TASK_PREEMPTS_CURR(p, rq))
1216 resched_task(rq->curr);
1221 p->state = TASK_RUNNING;
1223 task_rq_unlock(rq, &flags);
1228 int fastcall wake_up_process(task_t * p)
1230 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1231 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1234 EXPORT_SYMBOL(wake_up_process);
1236 int fastcall wake_up_state(task_t *p, unsigned int state)
1238 return try_to_wake_up(p, state, 0);
1242 * Perform scheduler related setup for a newly forked process p.
1243 * p is forked by current.
1245 void fastcall sched_fork(task_t *p)
1248 * We mark the process as running here, but have not actually
1249 * inserted it onto the runqueue yet. This guarantees that
1250 * nobody will actually run it, and a signal or other external
1251 * event cannot wake it up and insert it on the runqueue either.
1253 p->state = TASK_RUNNING;
1254 INIT_LIST_HEAD(&p->run_list);
1256 #ifdef CONFIG_SCHEDSTATS
1257 memset(&p->sched_info, 0, sizeof(p->sched_info));
1259 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1262 #ifdef CONFIG_PREEMPT
1263 /* Want to start with kernel preemption disabled. */
1264 p->thread_info->preempt_count = 1;
1267 * Share the timeslice between parent and child, thus the
1268 * total amount of pending timeslices in the system doesn't change,
1269 * resulting in more scheduling fairness.
1271 local_irq_disable();
1272 p->time_slice = (current->time_slice + 1) >> 1;
1274 * The remainder of the first timeslice might be recovered by
1275 * the parent if the child exits early enough.
1277 p->first_time_slice = 1;
1278 current->time_slice >>= 1;
1279 p->timestamp = sched_clock();
1280 if (unlikely(!current->time_slice)) {
1282 * This case is rare, it happens when the parent has only
1283 * a single jiffy left from its timeslice. Taking the
1284 * runqueue lock is not a problem.
1286 current->time_slice = 1;
1296 * wake_up_new_task - wake up a newly created task for the first time.
1298 * This function will do some initial scheduler statistics housekeeping
1299 * that must be done for every newly created context, then puts the task
1300 * on the runqueue and wakes it.
1302 void fastcall wake_up_new_task(task_t * p, unsigned long clone_flags)
1304 unsigned long flags;
1306 runqueue_t *rq, *this_rq;
1308 struct sched_domain *tmp, *sd = NULL;
1311 rq = task_rq_lock(p, &flags);
1312 BUG_ON(p->state != TASK_RUNNING);
1313 this_cpu = smp_processor_id();
1317 for_each_domain(cpu, tmp)
1318 if (tmp->flags & SD_BALANCE_FORK)
1323 struct sched_group *group;
1325 schedstat_inc(sd, sbf_cnt);
1327 group = find_idlest_group(sd, p, cpu);
1329 schedstat_inc(sd, sbf_balanced);
1330 goto no_forkbalance;
1333 new_cpu = find_idlest_cpu(group, cpu);
1334 if (new_cpu == -1 || new_cpu == cpu) {
1335 schedstat_inc(sd, sbf_balanced);
1336 goto no_forkbalance;
1339 if (cpu_isset(new_cpu, p->cpus_allowed)) {
1340 schedstat_inc(sd, sbf_pushed);
1341 set_task_cpu(p, new_cpu);
1342 task_rq_unlock(rq, &flags);
1343 rq = task_rq_lock(p, &flags);
1351 * We decrease the sleep average of forking parents
1352 * and children as well, to keep max-interactive tasks
1353 * from forking tasks that are max-interactive. The parent
1354 * (current) is done further down, under its lock.
1356 p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1357 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1359 p->prio = effective_prio(p);
1361 if (likely(cpu == this_cpu)) {
1362 if (!(clone_flags & CLONE_VM)) {
1364 * The VM isn't cloned, so we're in a good position to
1365 * do child-runs-first in anticipation of an exec. This
1366 * usually avoids a lot of COW overhead.
1368 if (unlikely(!current->array))
1369 __activate_task(p, rq);
1371 p->prio = current->prio;
1372 list_add_tail(&p->run_list, ¤t->run_list);
1373 p->array = current->array;
1374 p->array->nr_active++;
1379 /* Run child last */
1380 __activate_task(p, rq);
1382 * We skip the following code due to cpu == this_cpu
1384 * task_rq_unlock(rq, &flags);
1385 * this_rq = task_rq_lock(current, &flags);
1389 this_rq = cpu_rq(this_cpu);
1392 * Not the local CPU - must adjust timestamp. This should
1393 * get optimised away in the !CONFIG_SMP case.
1395 p->timestamp = (p->timestamp - this_rq->timestamp_last_tick)
1396 + rq->timestamp_last_tick;
1397 __activate_task(p, rq);
1398 if (TASK_PREEMPTS_CURR(p, rq))
1399 resched_task(rq->curr);
1402 * Parent and child are on different CPUs, now get the
1403 * parent runqueue to update the parent's ->sleep_avg:
1405 task_rq_unlock(rq, &flags);
1406 this_rq = task_rq_lock(current, &flags);
1408 current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1409 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1410 task_rq_unlock(this_rq, &flags);
1414 * Potentially available exiting-child timeslices are
1415 * retrieved here - this way the parent does not get
1416 * penalized for creating too many threads.
1418 * (this cannot be used to 'generate' timeslices
1419 * artificially, because any timeslice recovered here
1420 * was given away by the parent in the first place.)
1422 void fastcall sched_exit(task_t * p)
1424 unsigned long flags;
1428 * If the child was a (relative-) CPU hog then decrease
1429 * the sleep_avg of the parent as well.
1431 rq = task_rq_lock(p->parent, &flags);
1432 if (p->first_time_slice) {
1433 p->parent->time_slice += p->time_slice;
1434 if (unlikely(p->parent->time_slice > task_timeslice(p)))
1435 p->parent->time_slice = task_timeslice(p);
1437 if (p->sleep_avg < p->parent->sleep_avg)
1438 p->parent->sleep_avg = p->parent->sleep_avg /
1439 (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg /
1441 task_rq_unlock(rq, &flags);
1445 * prepare_task_switch - prepare to switch tasks
1446 * @rq: the runqueue preparing to switch
1447 * @next: the task we are going to switch to.
1449 * This is called with the rq lock held and interrupts off. It must
1450 * be paired with a subsequent finish_task_switch after the context
1453 * prepare_task_switch sets up locking and calls architecture specific
1456 static inline void prepare_task_switch(runqueue_t *rq, task_t *next)
1458 prepare_lock_switch(rq, next);
1459 prepare_arch_switch(next);
1463 * finish_task_switch - clean up after a task-switch
1464 * @prev: the thread we just switched away from.
1466 * finish_task_switch must be called after the context switch, paired
1467 * with a prepare_task_switch call before the context switch.
1468 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1469 * and do any other architecture-specific cleanup actions.
1471 * Note that we may have delayed dropping an mm in context_switch(). If
1472 * so, we finish that here outside of the runqueue lock. (Doing it
1473 * with the lock held can cause deadlocks; see schedule() for
1476 static inline void finish_task_switch(runqueue_t *rq, task_t *prev)
1477 __releases(rq->lock)
1479 struct mm_struct *mm = rq->prev_mm;
1480 unsigned long prev_task_flags;
1485 * A task struct has one reference for the use as "current".
1486 * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1487 * calls schedule one last time. The schedule call will never return,
1488 * and the scheduled task must drop that reference.
1489 * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1490 * still held, otherwise prev could be scheduled on another cpu, die
1491 * there before we look at prev->state, and then the reference would
1493 * Manfred Spraul <manfred@colorfullife.com>
1495 prev_task_flags = prev->flags;
1496 finish_arch_switch(prev);
1497 finish_lock_switch(rq, prev);
1500 if (unlikely(prev_task_flags & PF_DEAD))
1501 put_task_struct(prev);
1505 * schedule_tail - first thing a freshly forked thread must call.
1506 * @prev: the thread we just switched away from.
1508 asmlinkage void schedule_tail(task_t *prev)
1509 __releases(rq->lock)
1511 runqueue_t *rq = this_rq();
1512 finish_task_switch(rq, prev);
1513 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1514 /* In this case, finish_task_switch does not reenable preemption */
1517 if (current->set_child_tid)
1518 put_user(current->pid, current->set_child_tid);
1522 * context_switch - switch to the new MM and the new
1523 * thread's register state.
1526 task_t * context_switch(runqueue_t *rq, task_t *prev, task_t *next)
1528 struct mm_struct *mm = next->mm;
1529 struct mm_struct *oldmm = prev->active_mm;
1531 if (unlikely(!mm)) {
1532 next->active_mm = oldmm;
1533 atomic_inc(&oldmm->mm_count);
1534 enter_lazy_tlb(oldmm, next);
1536 switch_mm(oldmm, mm, next);
1538 if (unlikely(!prev->mm)) {
1539 prev->active_mm = NULL;
1540 WARN_ON(rq->prev_mm);
1541 rq->prev_mm = oldmm;
1544 /* Here we just switch the register state and the stack. */
1545 switch_to(prev, next, prev);
1551 * nr_running, nr_uninterruptible and nr_context_switches:
1553 * externally visible scheduler statistics: current number of runnable
1554 * threads, current number of uninterruptible-sleeping threads, total
1555 * number of context switches performed since bootup.
1557 unsigned long nr_running(void)
1559 unsigned long i, sum = 0;
1561 for_each_online_cpu(i)
1562 sum += cpu_rq(i)->nr_running;
1567 unsigned long nr_uninterruptible(void)
1569 unsigned long i, sum = 0;
1572 sum += cpu_rq(i)->nr_uninterruptible;
1575 * Since we read the counters lockless, it might be slightly
1576 * inaccurate. Do not allow it to go below zero though:
1578 if (unlikely((long)sum < 0))
1584 unsigned long long nr_context_switches(void)
1586 unsigned long long i, sum = 0;
1589 sum += cpu_rq(i)->nr_switches;
1594 unsigned long nr_iowait(void)
1596 unsigned long i, sum = 0;
1599 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1607 * double_rq_lock - safely lock two runqueues
1609 * Note this does not disable interrupts like task_rq_lock,
1610 * you need to do so manually before calling.
1612 static void double_rq_lock(runqueue_t *rq1, runqueue_t *rq2)
1613 __acquires(rq1->lock)
1614 __acquires(rq2->lock)
1617 spin_lock(&rq1->lock);
1618 __acquire(rq2->lock); /* Fake it out ;) */
1621 spin_lock(&rq1->lock);
1622 spin_lock(&rq2->lock);
1624 spin_lock(&rq2->lock);
1625 spin_lock(&rq1->lock);
1631 * double_rq_unlock - safely unlock two runqueues
1633 * Note this does not restore interrupts like task_rq_unlock,
1634 * you need to do so manually after calling.
1636 static void double_rq_unlock(runqueue_t *rq1, runqueue_t *rq2)
1637 __releases(rq1->lock)
1638 __releases(rq2->lock)
1640 spin_unlock(&rq1->lock);
1642 spin_unlock(&rq2->lock);
1644 __release(rq2->lock);
1648 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1650 static void double_lock_balance(runqueue_t *this_rq, runqueue_t *busiest)
1651 __releases(this_rq->lock)
1652 __acquires(busiest->lock)
1653 __acquires(this_rq->lock)
1655 if (unlikely(!spin_trylock(&busiest->lock))) {
1656 if (busiest < this_rq) {
1657 spin_unlock(&this_rq->lock);
1658 spin_lock(&busiest->lock);
1659 spin_lock(&this_rq->lock);
1661 spin_lock(&busiest->lock);
1666 * If dest_cpu is allowed for this process, migrate the task to it.
1667 * This is accomplished by forcing the cpu_allowed mask to only
1668 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
1669 * the cpu_allowed mask is restored.
1671 static void sched_migrate_task(task_t *p, int dest_cpu)
1673 migration_req_t req;
1675 unsigned long flags;
1677 rq = task_rq_lock(p, &flags);
1678 if (!cpu_isset(dest_cpu, p->cpus_allowed)
1679 || unlikely(cpu_is_offline(dest_cpu)))
1682 /* force the process onto the specified CPU */
1683 if (migrate_task(p, dest_cpu, &req)) {
1684 /* Need to wait for migration thread (might exit: take ref). */
1685 struct task_struct *mt = rq->migration_thread;
1686 get_task_struct(mt);
1687 task_rq_unlock(rq, &flags);
1688 wake_up_process(mt);
1689 put_task_struct(mt);
1690 wait_for_completion(&req.done);
1694 task_rq_unlock(rq, &flags);
1698 * sched_exec(): find the highest-level, exec-balance-capable
1699 * domain and try to migrate the task to the least loaded CPU.
1701 * execve() is a valuable balancing opportunity, because at this point
1702 * the task has the smallest effective memory and cache footprint.
1704 void sched_exec(void)
1706 struct sched_domain *tmp, *sd = NULL;
1707 int new_cpu, this_cpu = get_cpu();
1709 for_each_domain(this_cpu, tmp)
1710 if (tmp->flags & SD_BALANCE_EXEC)
1714 struct sched_group *group;
1715 schedstat_inc(sd, sbe_cnt);
1716 group = find_idlest_group(sd, current, this_cpu);
1718 schedstat_inc(sd, sbe_balanced);
1721 new_cpu = find_idlest_cpu(group, this_cpu);
1722 if (new_cpu == -1 || new_cpu == this_cpu) {
1723 schedstat_inc(sd, sbe_balanced);
1727 schedstat_inc(sd, sbe_pushed);
1729 sched_migrate_task(current, new_cpu);
1737 * pull_task - move a task from a remote runqueue to the local runqueue.
1738 * Both runqueues must be locked.
1741 void pull_task(runqueue_t *src_rq, prio_array_t *src_array, task_t *p,
1742 runqueue_t *this_rq, prio_array_t *this_array, int this_cpu)
1744 dequeue_task(p, src_array);
1745 src_rq->nr_running--;
1746 set_task_cpu(p, this_cpu);
1747 this_rq->nr_running++;
1748 enqueue_task(p, this_array);
1749 p->timestamp = (p->timestamp - src_rq->timestamp_last_tick)
1750 + this_rq->timestamp_last_tick;
1752 * Note that idle threads have a prio of MAX_PRIO, for this test
1753 * to be always true for them.
1755 if (TASK_PREEMPTS_CURR(p, this_rq))
1756 resched_task(this_rq->curr);
1760 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1763 int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu,
1764 struct sched_domain *sd, enum idle_type idle, int *all_pinned)
1767 * We do not migrate tasks that are:
1768 * 1) running (obviously), or
1769 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1770 * 3) are cache-hot on their current CPU.
1772 if (!cpu_isset(this_cpu, p->cpus_allowed))
1776 if (task_running(rq, p))
1780 * Aggressive migration if:
1781 * 1) task is cache cold, or
1782 * 2) too many balance attempts have failed.
1785 if (sd->nr_balance_failed > sd->cache_nice_tries)
1788 if (task_hot(p, rq->timestamp_last_tick, sd))
1794 * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq,
1795 * as part of a balancing operation within "domain". Returns the number of
1798 * Called with both runqueues locked.
1800 static int move_tasks(runqueue_t *this_rq, int this_cpu, runqueue_t *busiest,
1801 unsigned long max_nr_move, struct sched_domain *sd,
1802 enum idle_type idle, int *all_pinned)
1804 prio_array_t *array, *dst_array;
1805 struct list_head *head, *curr;
1806 int idx, pulled = 0, pinned = 0;
1809 if (max_nr_move == 0)
1815 * We first consider expired tasks. Those will likely not be
1816 * executed in the near future, and they are most likely to
1817 * be cache-cold, thus switching CPUs has the least effect
1820 if (busiest->expired->nr_active) {
1821 array = busiest->expired;
1822 dst_array = this_rq->expired;
1824 array = busiest->active;
1825 dst_array = this_rq->active;
1829 /* Start searching at priority 0: */
1833 idx = sched_find_first_bit(array->bitmap);
1835 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
1836 if (idx >= MAX_PRIO) {
1837 if (array == busiest->expired && busiest->active->nr_active) {
1838 array = busiest->active;
1839 dst_array = this_rq->active;
1845 head = array->queue + idx;
1848 tmp = list_entry(curr, task_t, run_list);
1852 if (!can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
1859 #ifdef CONFIG_SCHEDSTATS
1860 if (task_hot(tmp, busiest->timestamp_last_tick, sd))
1861 schedstat_inc(sd, lb_hot_gained[idle]);
1864 pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
1867 /* We only want to steal up to the prescribed number of tasks. */
1868 if (pulled < max_nr_move) {
1876 * Right now, this is the only place pull_task() is called,
1877 * so we can safely collect pull_task() stats here rather than
1878 * inside pull_task().
1880 schedstat_add(sd, lb_gained[idle], pulled);
1883 *all_pinned = pinned;
1888 * find_busiest_group finds and returns the busiest CPU group within the
1889 * domain. It calculates and returns the number of tasks which should be
1890 * moved to restore balance via the imbalance parameter.
1892 static struct sched_group *
1893 find_busiest_group(struct sched_domain *sd, int this_cpu,
1894 unsigned long *imbalance, enum idle_type idle)
1896 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
1897 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
1900 max_load = this_load = total_load = total_pwr = 0;
1901 if (idle == NOT_IDLE)
1902 load_idx = sd->busy_idx;
1903 else if (idle == NEWLY_IDLE)
1904 load_idx = sd->newidle_idx;
1906 load_idx = sd->idle_idx;
1913 local_group = cpu_isset(this_cpu, group->cpumask);
1915 /* Tally up the load of all CPUs in the group */
1918 for_each_cpu_mask(i, group->cpumask) {
1919 /* Bias balancing toward cpus of our domain */
1921 load = target_load(i, load_idx);
1923 load = source_load(i, load_idx);
1928 total_load += avg_load;
1929 total_pwr += group->cpu_power;
1931 /* Adjust by relative CPU power of the group */
1932 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1935 this_load = avg_load;
1937 } else if (avg_load > max_load) {
1938 max_load = avg_load;
1941 group = group->next;
1942 } while (group != sd->groups);
1944 if (!busiest || this_load >= max_load)
1947 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
1949 if (this_load >= avg_load ||
1950 100*max_load <= sd->imbalance_pct*this_load)
1954 * We're trying to get all the cpus to the average_load, so we don't
1955 * want to push ourselves above the average load, nor do we wish to
1956 * reduce the max loaded cpu below the average load, as either of these
1957 * actions would just result in more rebalancing later, and ping-pong
1958 * tasks around. Thus we look for the minimum possible imbalance.
1959 * Negative imbalances (*we* are more loaded than anyone else) will
1960 * be counted as no imbalance for these purposes -- we can't fix that
1961 * by pulling tasks to us. Be careful of negative numbers as they'll
1962 * appear as very large values with unsigned longs.
1964 /* How much load to actually move to equalise the imbalance */
1965 *imbalance = min((max_load - avg_load) * busiest->cpu_power,
1966 (avg_load - this_load) * this->cpu_power)
1969 if (*imbalance < SCHED_LOAD_SCALE) {
1970 unsigned long pwr_now = 0, pwr_move = 0;
1973 if (max_load - this_load >= SCHED_LOAD_SCALE*2) {
1979 * OK, we don't have enough imbalance to justify moving tasks,
1980 * however we may be able to increase total CPU power used by
1984 pwr_now += busiest->cpu_power*min(SCHED_LOAD_SCALE, max_load);
1985 pwr_now += this->cpu_power*min(SCHED_LOAD_SCALE, this_load);
1986 pwr_now /= SCHED_LOAD_SCALE;
1988 /* Amount of load we'd subtract */
1989 tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/busiest->cpu_power;
1991 pwr_move += busiest->cpu_power*min(SCHED_LOAD_SCALE,
1994 /* Amount of load we'd add */
1995 if (max_load*busiest->cpu_power <
1996 SCHED_LOAD_SCALE*SCHED_LOAD_SCALE)
1997 tmp = max_load*busiest->cpu_power/this->cpu_power;
1999 tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/this->cpu_power;
2000 pwr_move += this->cpu_power*min(SCHED_LOAD_SCALE, this_load + tmp);
2001 pwr_move /= SCHED_LOAD_SCALE;
2003 /* Move if we gain throughput */
2004 if (pwr_move <= pwr_now)
2011 /* Get rid of the scaling factor, rounding down as we divide */
2012 *imbalance = *imbalance / SCHED_LOAD_SCALE;
2022 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2024 static runqueue_t *find_busiest_queue(struct sched_group *group)
2026 unsigned long load, max_load = 0;
2027 runqueue_t *busiest = NULL;
2030 for_each_cpu_mask(i, group->cpumask) {
2031 load = source_load(i, 0);
2033 if (load > max_load) {
2035 busiest = cpu_rq(i);
2043 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2044 * tasks if there is an imbalance.
2046 * Called with this_rq unlocked.
2048 static int load_balance(int this_cpu, runqueue_t *this_rq,
2049 struct sched_domain *sd, enum idle_type idle)
2051 struct sched_group *group;
2052 runqueue_t *busiest;
2053 unsigned long imbalance;
2054 int nr_moved, all_pinned;
2055 int active_balance = 0;
2057 spin_lock(&this_rq->lock);
2058 schedstat_inc(sd, lb_cnt[idle]);
2060 group = find_busiest_group(sd, this_cpu, &imbalance, idle);
2062 schedstat_inc(sd, lb_nobusyg[idle]);
2066 busiest = find_busiest_queue(group);
2068 schedstat_inc(sd, lb_nobusyq[idle]);
2072 BUG_ON(busiest == this_rq);
2074 schedstat_add(sd, lb_imbalance[idle], imbalance);
2077 if (busiest->nr_running > 1) {
2079 * Attempt to move tasks. If find_busiest_group has found
2080 * an imbalance but busiest->nr_running <= 1, the group is
2081 * still unbalanced. nr_moved simply stays zero, so it is
2082 * correctly treated as an imbalance.
2084 double_lock_balance(this_rq, busiest);
2085 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2086 imbalance, sd, idle,
2088 spin_unlock(&busiest->lock);
2090 /* All tasks on this runqueue were pinned by CPU affinity */
2091 if (unlikely(all_pinned))
2095 spin_unlock(&this_rq->lock);
2098 schedstat_inc(sd, lb_failed[idle]);
2099 sd->nr_balance_failed++;
2101 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2103 spin_lock(&busiest->lock);
2104 if (!busiest->active_balance) {
2105 busiest->active_balance = 1;
2106 busiest->push_cpu = this_cpu;
2109 spin_unlock(&busiest->lock);
2111 wake_up_process(busiest->migration_thread);
2114 * We've kicked active balancing, reset the failure
2117 sd->nr_balance_failed = sd->cache_nice_tries+1;
2120 sd->nr_balance_failed = 0;
2122 if (likely(!active_balance)) {
2123 /* We were unbalanced, so reset the balancing interval */
2124 sd->balance_interval = sd->min_interval;
2127 * If we've begun active balancing, start to back off. This
2128 * case may not be covered by the all_pinned logic if there
2129 * is only 1 task on the busy runqueue (because we don't call
2132 if (sd->balance_interval < sd->max_interval)
2133 sd->balance_interval *= 2;
2139 spin_unlock(&this_rq->lock);
2141 schedstat_inc(sd, lb_balanced[idle]);
2143 sd->nr_balance_failed = 0;
2144 /* tune up the balancing interval */
2145 if (sd->balance_interval < sd->max_interval)
2146 sd->balance_interval *= 2;
2152 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2153 * tasks if there is an imbalance.
2155 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2156 * this_rq is locked.
2158 static int load_balance_newidle(int this_cpu, runqueue_t *this_rq,
2159 struct sched_domain *sd)
2161 struct sched_group *group;
2162 runqueue_t *busiest = NULL;
2163 unsigned long imbalance;
2166 schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
2167 group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE);
2169 schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]);
2173 busiest = find_busiest_queue(group);
2175 schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]);
2179 BUG_ON(busiest == this_rq);
2181 /* Attempt to move tasks */
2182 double_lock_balance(this_rq, busiest);
2184 schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
2185 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2186 imbalance, sd, NEWLY_IDLE, NULL);
2188 schedstat_inc(sd, lb_failed[NEWLY_IDLE]);
2190 sd->nr_balance_failed = 0;
2192 spin_unlock(&busiest->lock);
2196 schedstat_inc(sd, lb_balanced[NEWLY_IDLE]);
2197 sd->nr_balance_failed = 0;
2202 * idle_balance is called by schedule() if this_cpu is about to become
2203 * idle. Attempts to pull tasks from other CPUs.
2205 static inline void idle_balance(int this_cpu, runqueue_t *this_rq)
2207 struct sched_domain *sd;
2209 for_each_domain(this_cpu, sd) {
2210 if (sd->flags & SD_BALANCE_NEWIDLE) {
2211 if (load_balance_newidle(this_cpu, this_rq, sd)) {
2212 /* We've pulled tasks over so stop searching */
2220 * active_load_balance is run by migration threads. It pushes running tasks
2221 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2222 * running on each physical CPU where possible, and avoids physical /
2223 * logical imbalances.
2225 * Called with busiest_rq locked.
2227 static void active_load_balance(runqueue_t *busiest_rq, int busiest_cpu)
2229 struct sched_domain *sd;
2230 runqueue_t *target_rq;
2231 int target_cpu = busiest_rq->push_cpu;
2233 if (busiest_rq->nr_running <= 1)
2234 /* no task to move */
2237 target_rq = cpu_rq(target_cpu);
2240 * This condition is "impossible", if it occurs
2241 * we need to fix it. Originally reported by
2242 * Bjorn Helgaas on a 128-cpu setup.
2244 BUG_ON(busiest_rq == target_rq);
2246 /* move a task from busiest_rq to target_rq */
2247 double_lock_balance(busiest_rq, target_rq);
2249 /* Search for an sd spanning us and the target CPU. */
2250 for_each_domain(target_cpu, sd)
2251 if ((sd->flags & SD_LOAD_BALANCE) &&
2252 cpu_isset(busiest_cpu, sd->span))
2255 if (unlikely(sd == NULL))
2258 schedstat_inc(sd, alb_cnt);
2260 if (move_tasks(target_rq, target_cpu, busiest_rq, 1, sd, SCHED_IDLE, NULL))
2261 schedstat_inc(sd, alb_pushed);
2263 schedstat_inc(sd, alb_failed);
2265 spin_unlock(&target_rq->lock);
2269 * rebalance_tick will get called every timer tick, on every CPU.
2271 * It checks each scheduling domain to see if it is due to be balanced,
2272 * and initiates a balancing operation if so.
2274 * Balancing parameters are set up in arch_init_sched_domains.
2277 /* Don't have all balancing operations going off at once */
2278 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2280 static void rebalance_tick(int this_cpu, runqueue_t *this_rq,
2281 enum idle_type idle)
2283 unsigned long old_load, this_load;
2284 unsigned long j = jiffies + CPU_OFFSET(this_cpu);
2285 struct sched_domain *sd;
2288 this_load = this_rq->nr_running * SCHED_LOAD_SCALE;
2289 /* Update our load */
2290 for (i = 0; i < 3; i++) {
2291 unsigned long new_load = this_load;
2293 old_load = this_rq->cpu_load[i];
2295 * Round up the averaging division if load is increasing. This
2296 * prevents us from getting stuck on 9 if the load is 10, for
2299 if (new_load > old_load)
2300 new_load += scale-1;
2301 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale;
2304 for_each_domain(this_cpu, sd) {
2305 unsigned long interval;
2307 if (!(sd->flags & SD_LOAD_BALANCE))
2310 interval = sd->balance_interval;
2311 if (idle != SCHED_IDLE)
2312 interval *= sd->busy_factor;
2314 /* scale ms to jiffies */
2315 interval = msecs_to_jiffies(interval);
2316 if (unlikely(!interval))
2319 if (j - sd->last_balance >= interval) {
2320 if (load_balance(this_cpu, this_rq, sd, idle)) {
2321 /* We've pulled tasks over so no longer idle */
2324 sd->last_balance += interval;
2330 * on UP we do not need to balance between CPUs:
2332 static inline void rebalance_tick(int cpu, runqueue_t *rq, enum idle_type idle)
2335 static inline void idle_balance(int cpu, runqueue_t *rq)
2340 static inline int wake_priority_sleeper(runqueue_t *rq)
2343 #ifdef CONFIG_SCHED_SMT
2344 spin_lock(&rq->lock);
2346 * If an SMT sibling task has been put to sleep for priority
2347 * reasons reschedule the idle task to see if it can now run.
2349 if (rq->nr_running) {
2350 resched_task(rq->idle);
2353 spin_unlock(&rq->lock);
2358 DEFINE_PER_CPU(struct kernel_stat, kstat);
2360 EXPORT_PER_CPU_SYMBOL(kstat);
2363 * This is called on clock ticks and on context switches.
2364 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2366 static inline void update_cpu_clock(task_t *p, runqueue_t *rq,
2367 unsigned long long now)
2369 unsigned long long last = max(p->timestamp, rq->timestamp_last_tick);
2370 p->sched_time += now - last;
2374 * Return current->sched_time plus any more ns on the sched_clock
2375 * that have not yet been banked.
2377 unsigned long long current_sched_time(const task_t *tsk)
2379 unsigned long long ns;
2380 unsigned long flags;
2381 local_irq_save(flags);
2382 ns = max(tsk->timestamp, task_rq(tsk)->timestamp_last_tick);
2383 ns = tsk->sched_time + (sched_clock() - ns);
2384 local_irq_restore(flags);
2389 * We place interactive tasks back into the active array, if possible.
2391 * To guarantee that this does not starve expired tasks we ignore the
2392 * interactivity of a task if the first expired task had to wait more
2393 * than a 'reasonable' amount of time. This deadline timeout is
2394 * load-dependent, as the frequency of array switched decreases with
2395 * increasing number of running tasks. We also ignore the interactivity
2396 * if a better static_prio task has expired:
2398 #define EXPIRED_STARVING(rq) \
2399 ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2400 (jiffies - (rq)->expired_timestamp >= \
2401 STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2402 ((rq)->curr->static_prio > (rq)->best_expired_prio))
2405 * Account user cpu time to a process.
2406 * @p: the process that the cpu time gets accounted to
2407 * @hardirq_offset: the offset to subtract from hardirq_count()
2408 * @cputime: the cpu time spent in user space since the last update
2410 void account_user_time(struct task_struct *p, cputime_t cputime)
2412 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2415 p->utime = cputime_add(p->utime, cputime);
2417 /* Add user time to cpustat. */
2418 tmp = cputime_to_cputime64(cputime);
2419 if (TASK_NICE(p) > 0)
2420 cpustat->nice = cputime64_add(cpustat->nice, tmp);
2422 cpustat->user = cputime64_add(cpustat->user, tmp);
2426 * Account system cpu time to a process.
2427 * @p: the process that the cpu time gets accounted to
2428 * @hardirq_offset: the offset to subtract from hardirq_count()
2429 * @cputime: the cpu time spent in kernel space since the last update
2431 void account_system_time(struct task_struct *p, int hardirq_offset,
2434 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2435 runqueue_t *rq = this_rq();
2438 p->stime = cputime_add(p->stime, cputime);
2440 /* Add system time to cpustat. */
2441 tmp = cputime_to_cputime64(cputime);
2442 if (hardirq_count() - hardirq_offset)
2443 cpustat->irq = cputime64_add(cpustat->irq, tmp);
2444 else if (softirq_count())
2445 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
2446 else if (p != rq->idle)
2447 cpustat->system = cputime64_add(cpustat->system, tmp);
2448 else if (atomic_read(&rq->nr_iowait) > 0)
2449 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2451 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2452 /* Account for system time used */
2453 acct_update_integrals(p);
2454 /* Update rss highwater mark */
2455 update_mem_hiwater(p);
2459 * Account for involuntary wait time.
2460 * @p: the process from which the cpu time has been stolen
2461 * @steal: the cpu time spent in involuntary wait
2463 void account_steal_time(struct task_struct *p, cputime_t steal)
2465 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2466 cputime64_t tmp = cputime_to_cputime64(steal);
2467 runqueue_t *rq = this_rq();
2469 if (p == rq->idle) {
2470 p->stime = cputime_add(p->stime, steal);
2471 if (atomic_read(&rq->nr_iowait) > 0)
2472 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2474 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2476 cpustat->steal = cputime64_add(cpustat->steal, tmp);
2480 * This function gets called by the timer code, with HZ frequency.
2481 * We call it with interrupts disabled.
2483 * It also gets called by the fork code, when changing the parent's
2486 void scheduler_tick(void)
2488 int cpu = smp_processor_id();
2489 runqueue_t *rq = this_rq();
2490 task_t *p = current;
2491 unsigned long long now = sched_clock();
2493 update_cpu_clock(p, rq, now);
2495 rq->timestamp_last_tick = now;
2497 if (p == rq->idle) {
2498 if (wake_priority_sleeper(rq))
2500 rebalance_tick(cpu, rq, SCHED_IDLE);
2504 /* Task might have expired already, but not scheduled off yet */
2505 if (p->array != rq->active) {
2506 set_tsk_need_resched(p);
2509 spin_lock(&rq->lock);
2511 * The task was running during this tick - update the
2512 * time slice counter. Note: we do not update a thread's
2513 * priority until it either goes to sleep or uses up its
2514 * timeslice. This makes it possible for interactive tasks
2515 * to use up their timeslices at their highest priority levels.
2519 * RR tasks need a special form of timeslice management.
2520 * FIFO tasks have no timeslices.
2522 if ((p->policy == SCHED_RR) && !--p->time_slice) {
2523 p->time_slice = task_timeslice(p);
2524 p->first_time_slice = 0;
2525 set_tsk_need_resched(p);
2527 /* put it at the end of the queue: */
2528 requeue_task(p, rq->active);
2532 if (!--p->time_slice) {
2533 dequeue_task(p, rq->active);
2534 set_tsk_need_resched(p);
2535 p->prio = effective_prio(p);
2536 p->time_slice = task_timeslice(p);
2537 p->first_time_slice = 0;
2539 if (!rq->expired_timestamp)
2540 rq->expired_timestamp = jiffies;
2541 if (!TASK_INTERACTIVE(p) || EXPIRED_STARVING(rq)) {
2542 enqueue_task(p, rq->expired);
2543 if (p->static_prio < rq->best_expired_prio)
2544 rq->best_expired_prio = p->static_prio;
2546 enqueue_task(p, rq->active);
2549 * Prevent a too long timeslice allowing a task to monopolize
2550 * the CPU. We do this by splitting up the timeslice into
2553 * Note: this does not mean the task's timeslices expire or
2554 * get lost in any way, they just might be preempted by
2555 * another task of equal priority. (one with higher
2556 * priority would have preempted this task already.) We
2557 * requeue this task to the end of the list on this priority
2558 * level, which is in essence a round-robin of tasks with
2561 * This only applies to tasks in the interactive
2562 * delta range with at least TIMESLICE_GRANULARITY to requeue.
2564 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
2565 p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
2566 (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
2567 (p->array == rq->active)) {
2569 requeue_task(p, rq->active);
2570 set_tsk_need_resched(p);
2574 spin_unlock(&rq->lock);
2576 rebalance_tick(cpu, rq, NOT_IDLE);
2579 #ifdef CONFIG_SCHED_SMT
2580 static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2582 struct sched_domain *tmp, *sd = NULL;
2583 cpumask_t sibling_map;
2586 for_each_domain(this_cpu, tmp)
2587 if (tmp->flags & SD_SHARE_CPUPOWER)
2594 * Unlock the current runqueue because we have to lock in
2595 * CPU order to avoid deadlocks. Caller knows that we might
2596 * unlock. We keep IRQs disabled.
2598 spin_unlock(&this_rq->lock);
2600 sibling_map = sd->span;
2602 for_each_cpu_mask(i, sibling_map)
2603 spin_lock(&cpu_rq(i)->lock);
2605 * We clear this CPU from the mask. This both simplifies the
2606 * inner loop and keps this_rq locked when we exit:
2608 cpu_clear(this_cpu, sibling_map);
2610 for_each_cpu_mask(i, sibling_map) {
2611 runqueue_t *smt_rq = cpu_rq(i);
2614 * If an SMT sibling task is sleeping due to priority
2615 * reasons wake it up now.
2617 if (smt_rq->curr == smt_rq->idle && smt_rq->nr_running)
2618 resched_task(smt_rq->idle);
2621 for_each_cpu_mask(i, sibling_map)
2622 spin_unlock(&cpu_rq(i)->lock);
2624 * We exit with this_cpu's rq still held and IRQs
2629 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2631 struct sched_domain *tmp, *sd = NULL;
2632 cpumask_t sibling_map;
2633 prio_array_t *array;
2637 for_each_domain(this_cpu, tmp)
2638 if (tmp->flags & SD_SHARE_CPUPOWER)
2645 * The same locking rules and details apply as for
2646 * wake_sleeping_dependent():
2648 spin_unlock(&this_rq->lock);
2649 sibling_map = sd->span;
2650 for_each_cpu_mask(i, sibling_map)
2651 spin_lock(&cpu_rq(i)->lock);
2652 cpu_clear(this_cpu, sibling_map);
2655 * Establish next task to be run - it might have gone away because
2656 * we released the runqueue lock above:
2658 if (!this_rq->nr_running)
2660 array = this_rq->active;
2661 if (!array->nr_active)
2662 array = this_rq->expired;
2663 BUG_ON(!array->nr_active);
2665 p = list_entry(array->queue[sched_find_first_bit(array->bitmap)].next,
2668 for_each_cpu_mask(i, sibling_map) {
2669 runqueue_t *smt_rq = cpu_rq(i);
2670 task_t *smt_curr = smt_rq->curr;
2673 * If a user task with lower static priority than the
2674 * running task on the SMT sibling is trying to schedule,
2675 * delay it till there is proportionately less timeslice
2676 * left of the sibling task to prevent a lower priority
2677 * task from using an unfair proportion of the
2678 * physical cpu's resources. -ck
2680 if (((smt_curr->time_slice * (100 - sd->per_cpu_gain) / 100) >
2681 task_timeslice(p) || rt_task(smt_curr)) &&
2682 p->mm && smt_curr->mm && !rt_task(p))
2686 * Reschedule a lower priority task on the SMT sibling,
2687 * or wake it up if it has been put to sleep for priority
2690 if ((((p->time_slice * (100 - sd->per_cpu_gain) / 100) >
2691 task_timeslice(smt_curr) || rt_task(p)) &&
2692 smt_curr->mm && p->mm && !rt_task(smt_curr)) ||
2693 (smt_curr == smt_rq->idle && smt_rq->nr_running))
2694 resched_task(smt_curr);
2697 for_each_cpu_mask(i, sibling_map)
2698 spin_unlock(&cpu_rq(i)->lock);
2702 static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2706 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2712 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
2714 void fastcall add_preempt_count(int val)
2719 BUG_ON((preempt_count() < 0));
2720 preempt_count() += val;
2722 * Spinlock count overflowing soon?
2724 BUG_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10);
2726 EXPORT_SYMBOL(add_preempt_count);
2728 void fastcall sub_preempt_count(int val)
2733 BUG_ON(val > preempt_count());
2735 * Is the spinlock portion underflowing?
2737 BUG_ON((val < PREEMPT_MASK) && !(preempt_count() & PREEMPT_MASK));
2738 preempt_count() -= val;
2740 EXPORT_SYMBOL(sub_preempt_count);
2745 * schedule() is the main scheduler function.
2747 asmlinkage void __sched schedule(void)
2750 task_t *prev, *next;
2752 prio_array_t *array;
2753 struct list_head *queue;
2754 unsigned long long now;
2755 unsigned long run_time;
2759 * Test if we are atomic. Since do_exit() needs to call into
2760 * schedule() atomically, we ignore that path for now.
2761 * Otherwise, whine if we are scheduling when we should not be.
2763 if (likely(!current->exit_state)) {
2764 if (unlikely(in_atomic())) {
2765 printk(KERN_ERR "scheduling while atomic: "
2767 current->comm, preempt_count(), current->pid);
2771 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2776 release_kernel_lock(prev);
2777 need_resched_nonpreemptible:
2781 * The idle thread is not allowed to schedule!
2782 * Remove this check after it has been exercised a bit.
2784 if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
2785 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
2789 schedstat_inc(rq, sched_cnt);
2790 now = sched_clock();
2791 if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
2792 run_time = now - prev->timestamp;
2793 if (unlikely((long long)(now - prev->timestamp) < 0))
2796 run_time = NS_MAX_SLEEP_AVG;
2799 * Tasks charged proportionately less run_time at high sleep_avg to
2800 * delay them losing their interactive status
2802 run_time /= (CURRENT_BONUS(prev) ? : 1);
2804 spin_lock_irq(&rq->lock);
2806 if (unlikely(prev->flags & PF_DEAD))
2807 prev->state = EXIT_DEAD;
2809 switch_count = &prev->nivcsw;
2810 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2811 switch_count = &prev->nvcsw;
2812 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
2813 unlikely(signal_pending(prev))))
2814 prev->state = TASK_RUNNING;
2816 if (prev->state == TASK_UNINTERRUPTIBLE)
2817 rq->nr_uninterruptible++;
2818 deactivate_task(prev, rq);
2822 cpu = smp_processor_id();
2823 if (unlikely(!rq->nr_running)) {
2825 idle_balance(cpu, rq);
2826 if (!rq->nr_running) {
2828 rq->expired_timestamp = 0;
2829 wake_sleeping_dependent(cpu, rq);
2831 * wake_sleeping_dependent() might have released
2832 * the runqueue, so break out if we got new
2835 if (!rq->nr_running)
2839 if (dependent_sleeper(cpu, rq)) {
2844 * dependent_sleeper() releases and reacquires the runqueue
2845 * lock, hence go into the idle loop if the rq went
2848 if (unlikely(!rq->nr_running))
2853 if (unlikely(!array->nr_active)) {
2855 * Switch the active and expired arrays.
2857 schedstat_inc(rq, sched_switch);
2858 rq->active = rq->expired;
2859 rq->expired = array;
2861 rq->expired_timestamp = 0;
2862 rq->best_expired_prio = MAX_PRIO;
2865 idx = sched_find_first_bit(array->bitmap);
2866 queue = array->queue + idx;
2867 next = list_entry(queue->next, task_t, run_list);
2869 if (!rt_task(next) && next->activated > 0) {
2870 unsigned long long delta = now - next->timestamp;
2871 if (unlikely((long long)(now - next->timestamp) < 0))
2874 if (next->activated == 1)
2875 delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
2877 array = next->array;
2878 dequeue_task(next, array);
2879 recalc_task_prio(next, next->timestamp + delta);
2880 enqueue_task(next, array);
2882 next->activated = 0;
2884 if (next == rq->idle)
2885 schedstat_inc(rq, sched_goidle);
2887 clear_tsk_need_resched(prev);
2888 rcu_qsctr_inc(task_cpu(prev));
2890 update_cpu_clock(prev, rq, now);
2892 prev->sleep_avg -= run_time;
2893 if ((long)prev->sleep_avg <= 0)
2894 prev->sleep_avg = 0;
2895 prev->timestamp = prev->last_ran = now;
2897 sched_info_switch(prev, next);
2898 if (likely(prev != next)) {
2899 next->timestamp = now;
2904 prepare_task_switch(rq, next);
2905 prev = context_switch(rq, prev, next);
2908 * this_rq must be evaluated again because prev may have moved
2909 * CPUs since it called schedule(), thus the 'rq' on its stack
2910 * frame will be invalid.
2912 finish_task_switch(this_rq(), prev);
2914 spin_unlock_irq(&rq->lock);
2917 if (unlikely(reacquire_kernel_lock(prev) < 0))
2918 goto need_resched_nonpreemptible;
2919 preempt_enable_no_resched();
2920 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
2924 EXPORT_SYMBOL(schedule);
2926 #ifdef CONFIG_PREEMPT
2928 * this is is the entry point to schedule() from in-kernel preemption
2929 * off of preempt_enable. Kernel preemptions off return from interrupt
2930 * occur there and call schedule directly.
2932 asmlinkage void __sched preempt_schedule(void)
2934 struct thread_info *ti = current_thread_info();
2935 #ifdef CONFIG_PREEMPT_BKL
2936 struct task_struct *task = current;
2937 int saved_lock_depth;
2940 * If there is a non-zero preempt_count or interrupts are disabled,
2941 * we do not want to preempt the current task. Just return..
2943 if (unlikely(ti->preempt_count || irqs_disabled()))
2947 add_preempt_count(PREEMPT_ACTIVE);
2949 * We keep the big kernel semaphore locked, but we
2950 * clear ->lock_depth so that schedule() doesnt
2951 * auto-release the semaphore:
2953 #ifdef CONFIG_PREEMPT_BKL
2954 saved_lock_depth = task->lock_depth;
2955 task->lock_depth = -1;
2958 #ifdef CONFIG_PREEMPT_BKL
2959 task->lock_depth = saved_lock_depth;
2961 sub_preempt_count(PREEMPT_ACTIVE);
2963 /* we could miss a preemption opportunity between schedule and now */
2965 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
2969 EXPORT_SYMBOL(preempt_schedule);
2972 * this is is the entry point to schedule() from kernel preemption
2973 * off of irq context.
2974 * Note, that this is called and return with irqs disabled. This will
2975 * protect us against recursive calling from irq.
2977 asmlinkage void __sched preempt_schedule_irq(void)
2979 struct thread_info *ti = current_thread_info();
2980 #ifdef CONFIG_PREEMPT_BKL
2981 struct task_struct *task = current;
2982 int saved_lock_depth;
2984 /* Catch callers which need to be fixed*/
2985 BUG_ON(ti->preempt_count || !irqs_disabled());
2988 add_preempt_count(PREEMPT_ACTIVE);
2990 * We keep the big kernel semaphore locked, but we
2991 * clear ->lock_depth so that schedule() doesnt
2992 * auto-release the semaphore:
2994 #ifdef CONFIG_PREEMPT_BKL
2995 saved_lock_depth = task->lock_depth;
2996 task->lock_depth = -1;
3000 local_irq_disable();
3001 #ifdef CONFIG_PREEMPT_BKL
3002 task->lock_depth = saved_lock_depth;
3004 sub_preempt_count(PREEMPT_ACTIVE);
3006 /* we could miss a preemption opportunity between schedule and now */
3008 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3012 #endif /* CONFIG_PREEMPT */
3014 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync, void *key)
3016 task_t *p = curr->private;
3017 return try_to_wake_up(p, mode, sync);
3020 EXPORT_SYMBOL(default_wake_function);
3023 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3024 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3025 * number) then we wake all the non-exclusive tasks and one exclusive task.
3027 * There are circumstances in which we can try to wake a task which has already
3028 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3029 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3031 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3032 int nr_exclusive, int sync, void *key)
3034 struct list_head *tmp, *next;
3036 list_for_each_safe(tmp, next, &q->task_list) {
3039 curr = list_entry(tmp, wait_queue_t, task_list);
3040 flags = curr->flags;
3041 if (curr->func(curr, mode, sync, key) &&
3042 (flags & WQ_FLAG_EXCLUSIVE) &&
3049 * __wake_up - wake up threads blocked on a waitqueue.
3051 * @mode: which threads
3052 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3053 * @key: is directly passed to the wakeup function
3055 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3056 int nr_exclusive, void *key)
3058 unsigned long flags;
3060 spin_lock_irqsave(&q->lock, flags);
3061 __wake_up_common(q, mode, nr_exclusive, 0, key);
3062 spin_unlock_irqrestore(&q->lock, flags);
3065 EXPORT_SYMBOL(__wake_up);
3068 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3070 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3072 __wake_up_common(q, mode, 1, 0, NULL);
3076 * __wake_up_sync - wake up threads blocked on a waitqueue.
3078 * @mode: which threads
3079 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3081 * The sync wakeup differs that the waker knows that it will schedule
3082 * away soon, so while the target thread will be woken up, it will not
3083 * be migrated to another CPU - ie. the two threads are 'synchronized'
3084 * with each other. This can prevent needless bouncing between CPUs.
3086 * On UP it can prevent extra preemption.
3088 void fastcall __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3090 unsigned long flags;
3096 if (unlikely(!nr_exclusive))
3099 spin_lock_irqsave(&q->lock, flags);
3100 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3101 spin_unlock_irqrestore(&q->lock, flags);
3103 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3105 void fastcall complete(struct completion *x)
3107 unsigned long flags;
3109 spin_lock_irqsave(&x->wait.lock, flags);
3111 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3113 spin_unlock_irqrestore(&x->wait.lock, flags);
3115 EXPORT_SYMBOL(complete);
3117 void fastcall complete_all(struct completion *x)
3119 unsigned long flags;
3121 spin_lock_irqsave(&x->wait.lock, flags);
3122 x->done += UINT_MAX/2;
3123 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3125 spin_unlock_irqrestore(&x->wait.lock, flags);
3127 EXPORT_SYMBOL(complete_all);
3129 void fastcall __sched wait_for_completion(struct completion *x)
3132 spin_lock_irq(&x->wait.lock);
3134 DECLARE_WAITQUEUE(wait, current);
3136 wait.flags |= WQ_FLAG_EXCLUSIVE;
3137 __add_wait_queue_tail(&x->wait, &wait);
3139 __set_current_state(TASK_UNINTERRUPTIBLE);
3140 spin_unlock_irq(&x->wait.lock);
3142 spin_lock_irq(&x->wait.lock);
3144 __remove_wait_queue(&x->wait, &wait);
3147 spin_unlock_irq(&x->wait.lock);
3149 EXPORT_SYMBOL(wait_for_completion);
3151 unsigned long fastcall __sched
3152 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3156 spin_lock_irq(&x->wait.lock);
3158 DECLARE_WAITQUEUE(wait, current);
3160 wait.flags |= WQ_FLAG_EXCLUSIVE;
3161 __add_wait_queue_tail(&x->wait, &wait);
3163 __set_current_state(TASK_UNINTERRUPTIBLE);
3164 spin_unlock_irq(&x->wait.lock);
3165 timeout = schedule_timeout(timeout);
3166 spin_lock_irq(&x->wait.lock);
3168 __remove_wait_queue(&x->wait, &wait);
3172 __remove_wait_queue(&x->wait, &wait);
3176 spin_unlock_irq(&x->wait.lock);
3179 EXPORT_SYMBOL(wait_for_completion_timeout);
3181 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3187 spin_lock_irq(&x->wait.lock);
3189 DECLARE_WAITQUEUE(wait, current);
3191 wait.flags |= WQ_FLAG_EXCLUSIVE;
3192 __add_wait_queue_tail(&x->wait, &wait);
3194 if (signal_pending(current)) {
3196 __remove_wait_queue(&x->wait, &wait);
3199 __set_current_state(TASK_INTERRUPTIBLE);
3200 spin_unlock_irq(&x->wait.lock);
3202 spin_lock_irq(&x->wait.lock);
3204 __remove_wait_queue(&x->wait, &wait);
3208 spin_unlock_irq(&x->wait.lock);
3212 EXPORT_SYMBOL(wait_for_completion_interruptible);
3214 unsigned long fastcall __sched
3215 wait_for_completion_interruptible_timeout(struct completion *x,
3216 unsigned long timeout)
3220 spin_lock_irq(&x->wait.lock);
3222 DECLARE_WAITQUEUE(wait, current);
3224 wait.flags |= WQ_FLAG_EXCLUSIVE;
3225 __add_wait_queue_tail(&x->wait, &wait);
3227 if (signal_pending(current)) {
3228 timeout = -ERESTARTSYS;
3229 __remove_wait_queue(&x->wait, &wait);
3232 __set_current_state(TASK_INTERRUPTIBLE);
3233 spin_unlock_irq(&x->wait.lock);
3234 timeout = schedule_timeout(timeout);
3235 spin_lock_irq(&x->wait.lock);
3237 __remove_wait_queue(&x->wait, &wait);
3241 __remove_wait_queue(&x->wait, &wait);
3245 spin_unlock_irq(&x->wait.lock);
3248 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3251 #define SLEEP_ON_VAR \
3252 unsigned long flags; \
3253 wait_queue_t wait; \
3254 init_waitqueue_entry(&wait, current);
3256 #define SLEEP_ON_HEAD \
3257 spin_lock_irqsave(&q->lock,flags); \
3258 __add_wait_queue(q, &wait); \
3259 spin_unlock(&q->lock);
3261 #define SLEEP_ON_TAIL \
3262 spin_lock_irq(&q->lock); \
3263 __remove_wait_queue(q, &wait); \
3264 spin_unlock_irqrestore(&q->lock, flags);
3266 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
3270 current->state = TASK_INTERRUPTIBLE;
3277 EXPORT_SYMBOL(interruptible_sleep_on);
3279 long fastcall __sched interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3283 current->state = TASK_INTERRUPTIBLE;
3286 timeout = schedule_timeout(timeout);
3292 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3294 void fastcall __sched sleep_on(wait_queue_head_t *q)
3298 current->state = TASK_UNINTERRUPTIBLE;
3305 EXPORT_SYMBOL(sleep_on);
3307 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3311 current->state = TASK_UNINTERRUPTIBLE;
3314 timeout = schedule_timeout(timeout);
3320 EXPORT_SYMBOL(sleep_on_timeout);
3322 void set_user_nice(task_t *p, long nice)
3324 unsigned long flags;
3325 prio_array_t *array;
3327 int old_prio, new_prio, delta;
3329 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3332 * We have to be careful, if called from sys_setpriority(),
3333 * the task might be in the middle of scheduling on another CPU.
3335 rq = task_rq_lock(p, &flags);
3337 * The RT priorities are set via sched_setscheduler(), but we still
3338 * allow the 'normal' nice value to be set - but as expected
3339 * it wont have any effect on scheduling until the task is
3343 p->static_prio = NICE_TO_PRIO(nice);
3348 dequeue_task(p, array);
3351 new_prio = NICE_TO_PRIO(nice);
3352 delta = new_prio - old_prio;
3353 p->static_prio = NICE_TO_PRIO(nice);
3357 enqueue_task(p, array);
3359 * If the task increased its priority or is running and
3360 * lowered its priority, then reschedule its CPU:
3362 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3363 resched_task(rq->curr);
3366 task_rq_unlock(rq, &flags);
3369 EXPORT_SYMBOL(set_user_nice);
3372 * can_nice - check if a task can reduce its nice value
3376 int can_nice(const task_t *p, const int nice)
3378 /* convert nice value [19,-20] to rlimit style value [0,39] */
3379 int nice_rlim = 19 - nice;
3380 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
3381 capable(CAP_SYS_NICE));
3384 #ifdef __ARCH_WANT_SYS_NICE
3387 * sys_nice - change the priority of the current process.
3388 * @increment: priority increment
3390 * sys_setpriority is a more generic, but much slower function that
3391 * does similar things.
3393 asmlinkage long sys_nice(int increment)
3399 * Setpriority might change our priority at the same moment.
3400 * We don't have to worry. Conceptually one call occurs first
3401 * and we have a single winner.
3403 if (increment < -40)
3408 nice = PRIO_TO_NICE(current->static_prio) + increment;
3414 if (increment < 0 && !can_nice(current, nice))
3417 retval = security_task_setnice(current, nice);
3421 set_user_nice(current, nice);
3428 * task_prio - return the priority value of a given task.
3429 * @p: the task in question.
3431 * This is the priority value as seen by users in /proc.
3432 * RT tasks are offset by -200. Normal tasks are centered
3433 * around 0, value goes from -16 to +15.
3435 int task_prio(const task_t *p)
3437 return p->prio - MAX_RT_PRIO;
3441 * task_nice - return the nice value of a given task.
3442 * @p: the task in question.
3444 int task_nice(const task_t *p)
3446 return TASK_NICE(p);
3450 * The only users of task_nice are binfmt_elf and binfmt_elf32.
3451 * binfmt_elf is no longer modular, but binfmt_elf32 still is.
3452 * Therefore, task_nice is needed if there is a compat_mode.
3454 #ifdef CONFIG_COMPAT
3455 EXPORT_SYMBOL_GPL(task_nice);
3459 * idle_cpu - is a given cpu idle currently?
3460 * @cpu: the processor in question.
3462 int idle_cpu(int cpu)
3464 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
3467 EXPORT_SYMBOL_GPL(idle_cpu);
3470 * idle_task - return the idle task for a given cpu.
3471 * @cpu: the processor in question.
3473 task_t *idle_task(int cpu)
3475 return cpu_rq(cpu)->idle;
3479 * find_process_by_pid - find a process with a matching PID value.
3480 * @pid: the pid in question.
3482 static inline task_t *find_process_by_pid(pid_t pid)
3484 return pid ? find_task_by_pid(pid) : current;
3487 /* Actually do priority change: must hold rq lock. */
3488 static void __setscheduler(struct task_struct *p, int policy, int prio)
3492 p->rt_priority = prio;
3493 if (policy != SCHED_NORMAL)
3494 p->prio = MAX_USER_RT_PRIO-1 - p->rt_priority;
3496 p->prio = p->static_prio;
3500 * sched_setscheduler - change the scheduling policy and/or RT priority of
3502 * @p: the task in question.
3503 * @policy: new policy.
3504 * @param: structure containing the new RT priority.
3506 int sched_setscheduler(struct task_struct *p, int policy, struct sched_param *param)
3509 int oldprio, oldpolicy = -1;
3510 prio_array_t *array;
3511 unsigned long flags;
3515 /* double check policy once rq lock held */
3517 policy = oldpolicy = p->policy;
3518 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
3519 policy != SCHED_NORMAL)
3522 * Valid priorities for SCHED_FIFO and SCHED_RR are
3523 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL is 0.
3525 if (param->sched_priority < 0 ||
3526 param->sched_priority > MAX_USER_RT_PRIO-1)
3528 if ((policy == SCHED_NORMAL) != (param->sched_priority == 0))
3531 if ((policy == SCHED_FIFO || policy == SCHED_RR) &&
3532 param->sched_priority > p->signal->rlim[RLIMIT_RTPRIO].rlim_cur &&
3533 !capable(CAP_SYS_NICE))
3535 if ((current->euid != p->euid) && (current->euid != p->uid) &&
3536 !capable(CAP_SYS_NICE))
3539 retval = security_task_setscheduler(p, policy, param);
3543 * To be able to change p->policy safely, the apropriate
3544 * runqueue lock must be held.
3546 rq = task_rq_lock(p, &flags);
3547 /* recheck policy now with rq lock held */
3548 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3549 policy = oldpolicy = -1;
3550 task_rq_unlock(rq, &flags);
3555 deactivate_task(p, rq);
3557 __setscheduler(p, policy, param->sched_priority);
3559 __activate_task(p, rq);
3561 * Reschedule if we are currently running on this runqueue and
3562 * our priority decreased, or if we are not currently running on
3563 * this runqueue and our priority is higher than the current's
3565 if (task_running(rq, p)) {
3566 if (p->prio > oldprio)
3567 resched_task(rq->curr);
3568 } else if (TASK_PREEMPTS_CURR(p, rq))
3569 resched_task(rq->curr);
3571 task_rq_unlock(rq, &flags);
3574 EXPORT_SYMBOL_GPL(sched_setscheduler);
3576 static int do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3579 struct sched_param lparam;
3580 struct task_struct *p;
3582 if (!param || pid < 0)
3584 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3586 read_lock_irq(&tasklist_lock);
3587 p = find_process_by_pid(pid);
3589 read_unlock_irq(&tasklist_lock);
3592 retval = sched_setscheduler(p, policy, &lparam);
3593 read_unlock_irq(&tasklist_lock);
3598 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3599 * @pid: the pid in question.
3600 * @policy: new policy.
3601 * @param: structure containing the new RT priority.
3603 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
3604 struct sched_param __user *param)
3606 return do_sched_setscheduler(pid, policy, param);
3610 * sys_sched_setparam - set/change the RT priority of a thread
3611 * @pid: the pid in question.
3612 * @param: structure containing the new RT priority.
3614 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
3616 return do_sched_setscheduler(pid, -1, param);
3620 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3621 * @pid: the pid in question.
3623 asmlinkage long sys_sched_getscheduler(pid_t pid)
3625 int retval = -EINVAL;
3632 read_lock(&tasklist_lock);
3633 p = find_process_by_pid(pid);
3635 retval = security_task_getscheduler(p);
3639 read_unlock(&tasklist_lock);
3646 * sys_sched_getscheduler - get the RT priority of a thread
3647 * @pid: the pid in question.
3648 * @param: structure containing the RT priority.
3650 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
3652 struct sched_param lp;
3653 int retval = -EINVAL;
3656 if (!param || pid < 0)
3659 read_lock(&tasklist_lock);
3660 p = find_process_by_pid(pid);
3665 retval = security_task_getscheduler(p);
3669 lp.sched_priority = p->rt_priority;
3670 read_unlock(&tasklist_lock);
3673 * This one might sleep, we cannot do it with a spinlock held ...
3675 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3681 read_unlock(&tasklist_lock);
3685 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
3689 cpumask_t cpus_allowed;
3692 read_lock(&tasklist_lock);
3694 p = find_process_by_pid(pid);
3696 read_unlock(&tasklist_lock);
3697 unlock_cpu_hotplug();
3702 * It is not safe to call set_cpus_allowed with the
3703 * tasklist_lock held. We will bump the task_struct's
3704 * usage count and then drop tasklist_lock.
3707 read_unlock(&tasklist_lock);
3710 if ((current->euid != p->euid) && (current->euid != p->uid) &&
3711 !capable(CAP_SYS_NICE))
3714 cpus_allowed = cpuset_cpus_allowed(p);
3715 cpus_and(new_mask, new_mask, cpus_allowed);
3716 retval = set_cpus_allowed(p, new_mask);
3720 unlock_cpu_hotplug();
3724 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3725 cpumask_t *new_mask)
3727 if (len < sizeof(cpumask_t)) {
3728 memset(new_mask, 0, sizeof(cpumask_t));
3729 } else if (len > sizeof(cpumask_t)) {
3730 len = sizeof(cpumask_t);
3732 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3736 * sys_sched_setaffinity - set the cpu affinity of a process
3737 * @pid: pid of the process
3738 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3739 * @user_mask_ptr: user-space pointer to the new cpu mask
3741 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
3742 unsigned long __user *user_mask_ptr)
3747 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
3751 return sched_setaffinity(pid, new_mask);
3755 * Represents all cpu's present in the system
3756 * In systems capable of hotplug, this map could dynamically grow
3757 * as new cpu's are detected in the system via any platform specific
3758 * method, such as ACPI for e.g.
3761 cpumask_t cpu_present_map;
3762 EXPORT_SYMBOL(cpu_present_map);
3765 cpumask_t cpu_online_map = CPU_MASK_ALL;
3766 cpumask_t cpu_possible_map = CPU_MASK_ALL;
3769 long sched_getaffinity(pid_t pid, cpumask_t *mask)
3775 read_lock(&tasklist_lock);
3778 p = find_process_by_pid(pid);
3783 cpus_and(*mask, p->cpus_allowed, cpu_possible_map);
3786 read_unlock(&tasklist_lock);
3787 unlock_cpu_hotplug();
3795 * sys_sched_getaffinity - get the cpu affinity of a process
3796 * @pid: pid of the process
3797 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3798 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3800 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
3801 unsigned long __user *user_mask_ptr)
3806 if (len < sizeof(cpumask_t))
3809 ret = sched_getaffinity(pid, &mask);
3813 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
3816 return sizeof(cpumask_t);
3820 * sys_sched_yield - yield the current processor to other threads.
3822 * this function yields the current CPU by moving the calling thread
3823 * to the expired array. If there are no other threads running on this
3824 * CPU then this function will return.
3826 asmlinkage long sys_sched_yield(void)
3828 runqueue_t *rq = this_rq_lock();
3829 prio_array_t *array = current->array;
3830 prio_array_t *target = rq->expired;
3832 schedstat_inc(rq, yld_cnt);
3834 * We implement yielding by moving the task into the expired
3837 * (special rule: RT tasks will just roundrobin in the active
3840 if (rt_task(current))
3841 target = rq->active;
3843 if (current->array->nr_active == 1) {
3844 schedstat_inc(rq, yld_act_empty);
3845 if (!rq->expired->nr_active)
3846 schedstat_inc(rq, yld_both_empty);
3847 } else if (!rq->expired->nr_active)
3848 schedstat_inc(rq, yld_exp_empty);
3850 if (array != target) {
3851 dequeue_task(current, array);
3852 enqueue_task(current, target);
3855 * requeue_task is cheaper so perform that if possible.
3857 requeue_task(current, array);
3860 * Since we are going to call schedule() anyway, there's
3861 * no need to preempt or enable interrupts:
3863 __release(rq->lock);
3864 _raw_spin_unlock(&rq->lock);
3865 preempt_enable_no_resched();
3872 static inline void __cond_resched(void)
3875 add_preempt_count(PREEMPT_ACTIVE);
3877 sub_preempt_count(PREEMPT_ACTIVE);
3878 } while (need_resched());
3881 int __sched cond_resched(void)
3883 if (need_resched()) {
3890 EXPORT_SYMBOL(cond_resched);
3893 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
3894 * call schedule, and on return reacquire the lock.
3896 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
3897 * operations here to prevent schedule() from being called twice (once via
3898 * spin_unlock(), once by hand).
3900 int cond_resched_lock(spinlock_t * lock)
3904 if (need_lockbreak(lock)) {
3910 if (need_resched()) {
3911 _raw_spin_unlock(lock);
3912 preempt_enable_no_resched();
3920 EXPORT_SYMBOL(cond_resched_lock);
3922 int __sched cond_resched_softirq(void)
3924 BUG_ON(!in_softirq());
3926 if (need_resched()) {
3927 __local_bh_enable();
3935 EXPORT_SYMBOL(cond_resched_softirq);
3939 * yield - yield the current processor to other threads.
3941 * this is a shortcut for kernel-space yielding - it marks the
3942 * thread runnable and calls sys_sched_yield().
3944 void __sched yield(void)
3946 set_current_state(TASK_RUNNING);
3950 EXPORT_SYMBOL(yield);
3953 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
3954 * that process accounting knows that this is a task in IO wait state.
3956 * But don't do that if it is a deliberate, throttling IO wait (this task
3957 * has set its backing_dev_info: the queue against which it should throttle)
3959 void __sched io_schedule(void)
3961 struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id());
3963 atomic_inc(&rq->nr_iowait);
3965 atomic_dec(&rq->nr_iowait);
3968 EXPORT_SYMBOL(io_schedule);
3970 long __sched io_schedule_timeout(long timeout)
3972 struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id());
3975 atomic_inc(&rq->nr_iowait);
3976 ret = schedule_timeout(timeout);
3977 atomic_dec(&rq->nr_iowait);
3982 * sys_sched_get_priority_max - return maximum RT priority.
3983 * @policy: scheduling class.
3985 * this syscall returns the maximum rt_priority that can be used
3986 * by a given scheduling class.
3988 asmlinkage long sys_sched_get_priority_max(int policy)
3995 ret = MAX_USER_RT_PRIO-1;
4005 * sys_sched_get_priority_min - return minimum RT priority.
4006 * @policy: scheduling class.
4008 * this syscall returns the minimum rt_priority that can be used
4009 * by a given scheduling class.
4011 asmlinkage long sys_sched_get_priority_min(int policy)
4027 * sys_sched_rr_get_interval - return the default timeslice of a process.
4028 * @pid: pid of the process.
4029 * @interval: userspace pointer to the timeslice value.
4031 * this syscall writes the default timeslice value of a given process
4032 * into the user-space timespec buffer. A value of '0' means infinity.
4035 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4037 int retval = -EINVAL;
4045 read_lock(&tasklist_lock);
4046 p = find_process_by_pid(pid);
4050 retval = security_task_getscheduler(p);
4054 jiffies_to_timespec(p->policy & SCHED_FIFO ?
4055 0 : task_timeslice(p), &t);
4056 read_unlock(&tasklist_lock);
4057 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4061 read_unlock(&tasklist_lock);
4065 static inline struct task_struct *eldest_child(struct task_struct *p)
4067 if (list_empty(&p->children)) return NULL;
4068 return list_entry(p->children.next,struct task_struct,sibling);
4071 static inline struct task_struct *older_sibling(struct task_struct *p)
4073 if (p->sibling.prev==&p->parent->children) return NULL;
4074 return list_entry(p->sibling.prev,struct task_struct,sibling);
4077 static inline struct task_struct *younger_sibling(struct task_struct *p)
4079 if (p->sibling.next==&p->parent->children) return NULL;
4080 return list_entry(p->sibling.next,struct task_struct,sibling);
4083 static void show_task(task_t * p)
4087 unsigned long free = 0;
4088 static const char *stat_nam[] = { "R", "S", "D", "T", "t", "Z", "X" };
4090 printk("%-13.13s ", p->comm);
4091 state = p->state ? __ffs(p->state) + 1 : 0;
4092 if (state < ARRAY_SIZE(stat_nam))
4093 printk(stat_nam[state]);
4096 #if (BITS_PER_LONG == 32)
4097 if (state == TASK_RUNNING)
4098 printk(" running ");
4100 printk(" %08lX ", thread_saved_pc(p));
4102 if (state == TASK_RUNNING)
4103 printk(" running task ");
4105 printk(" %016lx ", thread_saved_pc(p));
4107 #ifdef CONFIG_DEBUG_STACK_USAGE
4109 unsigned long * n = (unsigned long *) (p->thread_info+1);
4112 free = (unsigned long) n - (unsigned long)(p->thread_info+1);
4115 printk("%5lu %5d %6d ", free, p->pid, p->parent->pid);
4116 if ((relative = eldest_child(p)))
4117 printk("%5d ", relative->pid);
4120 if ((relative = younger_sibling(p)))
4121 printk("%7d", relative->pid);
4124 if ((relative = older_sibling(p)))
4125 printk(" %5d", relative->pid);
4129 printk(" (L-TLB)\n");
4131 printk(" (NOTLB)\n");
4133 if (state != TASK_RUNNING)
4134 show_stack(p, NULL);
4137 void show_state(void)
4141 #if (BITS_PER_LONG == 32)
4144 printk(" task PC pid father child younger older\n");
4148 printk(" task PC pid father child younger older\n");
4150 read_lock(&tasklist_lock);
4151 do_each_thread(g, p) {
4153 * reset the NMI-timeout, listing all files on a slow
4154 * console might take alot of time:
4156 touch_nmi_watchdog();
4158 } while_each_thread(g, p);
4160 read_unlock(&tasklist_lock);
4163 void __devinit init_idle(task_t *idle, int cpu)
4165 runqueue_t *rq = cpu_rq(cpu);
4166 unsigned long flags;
4168 idle->sleep_avg = 0;
4170 idle->prio = MAX_PRIO;
4171 idle->state = TASK_RUNNING;
4172 idle->cpus_allowed = cpumask_of_cpu(cpu);
4173 set_task_cpu(idle, cpu);
4175 spin_lock_irqsave(&rq->lock, flags);
4176 rq->curr = rq->idle = idle;
4177 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4180 set_tsk_need_resched(idle);
4181 spin_unlock_irqrestore(&rq->lock, flags);
4183 /* Set the preempt count _outside_ the spinlocks! */
4184 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4185 idle->thread_info->preempt_count = (idle->lock_depth >= 0);
4187 idle->thread_info->preempt_count = 0;
4192 * In a system that switches off the HZ timer nohz_cpu_mask
4193 * indicates which cpus entered this state. This is used
4194 * in the rcu update to wait only for active cpus. For system
4195 * which do not switch off the HZ timer nohz_cpu_mask should
4196 * always be CPU_MASK_NONE.
4198 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4202 * This is how migration works:
4204 * 1) we queue a migration_req_t structure in the source CPU's
4205 * runqueue and wake up that CPU's migration thread.
4206 * 2) we down() the locked semaphore => thread blocks.
4207 * 3) migration thread wakes up (implicitly it forces the migrated
4208 * thread off the CPU)
4209 * 4) it gets the migration request and checks whether the migrated
4210 * task is still in the wrong runqueue.
4211 * 5) if it's in the wrong runqueue then the migration thread removes
4212 * it and puts it into the right queue.
4213 * 6) migration thread up()s the semaphore.
4214 * 7) we wake up and the migration is done.
4218 * Change a given task's CPU affinity. Migrate the thread to a
4219 * proper CPU and schedule it away if the CPU it's executing on
4220 * is removed from the allowed bitmask.
4222 * NOTE: the caller must have a valid reference to the task, the
4223 * task must not exit() & deallocate itself prematurely. The
4224 * call is not atomic; no spinlocks may be held.
4226 int set_cpus_allowed(task_t *p, cpumask_t new_mask)
4228 unsigned long flags;
4230 migration_req_t req;
4233 rq = task_rq_lock(p, &flags);
4234 if (!cpus_intersects(new_mask, cpu_online_map)) {
4239 p->cpus_allowed = new_mask;
4240 /* Can the task run on the task's current CPU? If so, we're done */
4241 if (cpu_isset(task_cpu(p), new_mask))
4244 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4245 /* Need help from migration thread: drop lock and wait. */
4246 task_rq_unlock(rq, &flags);
4247 wake_up_process(rq->migration_thread);
4248 wait_for_completion(&req.done);
4249 tlb_migrate_finish(p->mm);
4253 task_rq_unlock(rq, &flags);
4257 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4260 * Move (not current) task off this cpu, onto dest cpu. We're doing
4261 * this because either it can't run here any more (set_cpus_allowed()
4262 * away from this CPU, or CPU going down), or because we're
4263 * attempting to rebalance this task on exec (sched_exec).
4265 * So we race with normal scheduler movements, but that's OK, as long
4266 * as the task is no longer on this CPU.
4268 static void __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4270 runqueue_t *rq_dest, *rq_src;
4272 if (unlikely(cpu_is_offline(dest_cpu)))
4275 rq_src = cpu_rq(src_cpu);
4276 rq_dest = cpu_rq(dest_cpu);
4278 double_rq_lock(rq_src, rq_dest);
4279 /* Already moved. */
4280 if (task_cpu(p) != src_cpu)
4282 /* Affinity changed (again). */
4283 if (!cpu_isset(dest_cpu, p->cpus_allowed))
4286 set_task_cpu(p, dest_cpu);
4289 * Sync timestamp with rq_dest's before activating.
4290 * The same thing could be achieved by doing this step
4291 * afterwards, and pretending it was a local activate.
4292 * This way is cleaner and logically correct.
4294 p->timestamp = p->timestamp - rq_src->timestamp_last_tick
4295 + rq_dest->timestamp_last_tick;
4296 deactivate_task(p, rq_src);
4297 activate_task(p, rq_dest, 0);
4298 if (TASK_PREEMPTS_CURR(p, rq_dest))
4299 resched_task(rq_dest->curr);
4303 double_rq_unlock(rq_src, rq_dest);
4307 * migration_thread - this is a highprio system thread that performs
4308 * thread migration by bumping thread off CPU then 'pushing' onto
4311 static int migration_thread(void * data)
4314 int cpu = (long)data;
4317 BUG_ON(rq->migration_thread != current);
4319 set_current_state(TASK_INTERRUPTIBLE);
4320 while (!kthread_should_stop()) {
4321 struct list_head *head;
4322 migration_req_t *req;
4324 if (current->flags & PF_FREEZE)
4325 refrigerator(PF_FREEZE);
4327 spin_lock_irq(&rq->lock);
4329 if (cpu_is_offline(cpu)) {
4330 spin_unlock_irq(&rq->lock);
4334 if (rq->active_balance) {
4335 active_load_balance(rq, cpu);
4336 rq->active_balance = 0;
4339 head = &rq->migration_queue;
4341 if (list_empty(head)) {
4342 spin_unlock_irq(&rq->lock);
4344 set_current_state(TASK_INTERRUPTIBLE);
4347 req = list_entry(head->next, migration_req_t, list);
4348 list_del_init(head->next);
4350 if (req->type == REQ_MOVE_TASK) {
4351 spin_unlock(&rq->lock);
4352 __migrate_task(req->task, cpu, req->dest_cpu);
4354 } else if (req->type == REQ_SET_DOMAIN) {
4356 spin_unlock_irq(&rq->lock);
4358 spin_unlock_irq(&rq->lock);
4362 complete(&req->done);
4364 __set_current_state(TASK_RUNNING);
4368 /* Wait for kthread_stop */
4369 set_current_state(TASK_INTERRUPTIBLE);
4370 while (!kthread_should_stop()) {
4372 set_current_state(TASK_INTERRUPTIBLE);
4374 __set_current_state(TASK_RUNNING);
4378 #ifdef CONFIG_HOTPLUG_CPU
4379 /* Figure out where task on dead CPU should go, use force if neccessary. */
4380 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *tsk)
4386 mask = node_to_cpumask(cpu_to_node(dead_cpu));
4387 cpus_and(mask, mask, tsk->cpus_allowed);
4388 dest_cpu = any_online_cpu(mask);
4390 /* On any allowed CPU? */
4391 if (dest_cpu == NR_CPUS)
4392 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4394 /* No more Mr. Nice Guy. */
4395 if (dest_cpu == NR_CPUS) {
4396 cpus_setall(tsk->cpus_allowed);
4397 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4400 * Don't tell them about moving exiting tasks or
4401 * kernel threads (both mm NULL), since they never
4404 if (tsk->mm && printk_ratelimit())
4405 printk(KERN_INFO "process %d (%s) no "
4406 "longer affine to cpu%d\n",
4407 tsk->pid, tsk->comm, dead_cpu);
4409 __migrate_task(tsk, dead_cpu, dest_cpu);
4413 * While a dead CPU has no uninterruptible tasks queued at this point,
4414 * it might still have a nonzero ->nr_uninterruptible counter, because
4415 * for performance reasons the counter is not stricly tracking tasks to
4416 * their home CPUs. So we just add the counter to another CPU's counter,
4417 * to keep the global sum constant after CPU-down:
4419 static void migrate_nr_uninterruptible(runqueue_t *rq_src)
4421 runqueue_t *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
4422 unsigned long flags;
4424 local_irq_save(flags);
4425 double_rq_lock(rq_src, rq_dest);
4426 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
4427 rq_src->nr_uninterruptible = 0;
4428 double_rq_unlock(rq_src, rq_dest);
4429 local_irq_restore(flags);
4432 /* Run through task list and migrate tasks from the dead cpu. */
4433 static void migrate_live_tasks(int src_cpu)
4435 struct task_struct *tsk, *t;
4437 write_lock_irq(&tasklist_lock);
4439 do_each_thread(t, tsk) {
4443 if (task_cpu(tsk) == src_cpu)
4444 move_task_off_dead_cpu(src_cpu, tsk);
4445 } while_each_thread(t, tsk);
4447 write_unlock_irq(&tasklist_lock);
4450 /* Schedules idle task to be the next runnable task on current CPU.
4451 * It does so by boosting its priority to highest possible and adding it to
4452 * the _front_ of runqueue. Used by CPU offline code.
4454 void sched_idle_next(void)
4456 int cpu = smp_processor_id();
4457 runqueue_t *rq = this_rq();
4458 struct task_struct *p = rq->idle;
4459 unsigned long flags;
4461 /* cpu has to be offline */
4462 BUG_ON(cpu_online(cpu));
4464 /* Strictly not necessary since rest of the CPUs are stopped by now
4465 * and interrupts disabled on current cpu.
4467 spin_lock_irqsave(&rq->lock, flags);
4469 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4470 /* Add idle task to _front_ of it's priority queue */
4471 __activate_idle_task(p, rq);
4473 spin_unlock_irqrestore(&rq->lock, flags);
4476 /* Ensures that the idle task is using init_mm right before its cpu goes
4479 void idle_task_exit(void)
4481 struct mm_struct *mm = current->active_mm;
4483 BUG_ON(cpu_online(smp_processor_id()));
4486 switch_mm(mm, &init_mm, current);
4490 static void migrate_dead(unsigned int dead_cpu, task_t *tsk)
4492 struct runqueue *rq = cpu_rq(dead_cpu);
4494 /* Must be exiting, otherwise would be on tasklist. */
4495 BUG_ON(tsk->exit_state != EXIT_ZOMBIE && tsk->exit_state != EXIT_DEAD);
4497 /* Cannot have done final schedule yet: would have vanished. */
4498 BUG_ON(tsk->flags & PF_DEAD);
4500 get_task_struct(tsk);
4503 * Drop lock around migration; if someone else moves it,
4504 * that's OK. No task can be added to this CPU, so iteration is
4507 spin_unlock_irq(&rq->lock);
4508 move_task_off_dead_cpu(dead_cpu, tsk);
4509 spin_lock_irq(&rq->lock);
4511 put_task_struct(tsk);
4514 /* release_task() removes task from tasklist, so we won't find dead tasks. */
4515 static void migrate_dead_tasks(unsigned int dead_cpu)
4518 struct runqueue *rq = cpu_rq(dead_cpu);
4520 for (arr = 0; arr < 2; arr++) {
4521 for (i = 0; i < MAX_PRIO; i++) {
4522 struct list_head *list = &rq->arrays[arr].queue[i];
4523 while (!list_empty(list))
4524 migrate_dead(dead_cpu,
4525 list_entry(list->next, task_t,
4530 #endif /* CONFIG_HOTPLUG_CPU */
4533 * migration_call - callback that gets triggered when a CPU is added.
4534 * Here we can start up the necessary migration thread for the new CPU.
4536 static int migration_call(struct notifier_block *nfb, unsigned long action,
4539 int cpu = (long)hcpu;
4540 struct task_struct *p;
4541 struct runqueue *rq;
4542 unsigned long flags;
4545 case CPU_UP_PREPARE:
4546 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
4549 p->flags |= PF_NOFREEZE;
4550 kthread_bind(p, cpu);
4551 /* Must be high prio: stop_machine expects to yield to it. */
4552 rq = task_rq_lock(p, &flags);
4553 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4554 task_rq_unlock(rq, &flags);
4555 cpu_rq(cpu)->migration_thread = p;
4558 /* Strictly unneccessary, as first user will wake it. */
4559 wake_up_process(cpu_rq(cpu)->migration_thread);
4561 #ifdef CONFIG_HOTPLUG_CPU
4562 case CPU_UP_CANCELED:
4563 /* Unbind it from offline cpu so it can run. Fall thru. */
4564 kthread_bind(cpu_rq(cpu)->migration_thread,smp_processor_id());
4565 kthread_stop(cpu_rq(cpu)->migration_thread);
4566 cpu_rq(cpu)->migration_thread = NULL;
4569 migrate_live_tasks(cpu);
4571 kthread_stop(rq->migration_thread);
4572 rq->migration_thread = NULL;
4573 /* Idle task back to normal (off runqueue, low prio) */
4574 rq = task_rq_lock(rq->idle, &flags);
4575 deactivate_task(rq->idle, rq);
4576 rq->idle->static_prio = MAX_PRIO;
4577 __setscheduler(rq->idle, SCHED_NORMAL, 0);
4578 migrate_dead_tasks(cpu);
4579 task_rq_unlock(rq, &flags);
4580 migrate_nr_uninterruptible(rq);
4581 BUG_ON(rq->nr_running != 0);
4583 /* No need to migrate the tasks: it was best-effort if
4584 * they didn't do lock_cpu_hotplug(). Just wake up
4585 * the requestors. */
4586 spin_lock_irq(&rq->lock);
4587 while (!list_empty(&rq->migration_queue)) {
4588 migration_req_t *req;
4589 req = list_entry(rq->migration_queue.next,
4590 migration_req_t, list);
4591 BUG_ON(req->type != REQ_MOVE_TASK);
4592 list_del_init(&req->list);
4593 complete(&req->done);
4595 spin_unlock_irq(&rq->lock);
4602 /* Register at highest priority so that task migration (migrate_all_tasks)
4603 * happens before everything else.
4605 static struct notifier_block __devinitdata migration_notifier = {
4606 .notifier_call = migration_call,
4610 int __init migration_init(void)
4612 void *cpu = (void *)(long)smp_processor_id();
4613 /* Start one for boot CPU. */
4614 migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
4615 migration_call(&migration_notifier, CPU_ONLINE, cpu);
4616 register_cpu_notifier(&migration_notifier);
4622 #define SCHED_DOMAIN_DEBUG
4623 #ifdef SCHED_DOMAIN_DEBUG
4624 static void sched_domain_debug(struct sched_domain *sd, int cpu)
4629 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
4633 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
4638 struct sched_group *group = sd->groups;
4639 cpumask_t groupmask;
4641 cpumask_scnprintf(str, NR_CPUS, sd->span);
4642 cpus_clear(groupmask);
4645 for (i = 0; i < level + 1; i++)
4647 printk("domain %d: ", level);
4649 if (!(sd->flags & SD_LOAD_BALANCE)) {
4650 printk("does not load-balance\n");
4652 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent");
4656 printk("span %s\n", str);
4658 if (!cpu_isset(cpu, sd->span))
4659 printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
4660 if (!cpu_isset(cpu, group->cpumask))
4661 printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
4664 for (i = 0; i < level + 2; i++)
4670 printk(KERN_ERR "ERROR: group is NULL\n");
4674 if (!group->cpu_power) {
4676 printk(KERN_ERR "ERROR: domain->cpu_power not set\n");
4679 if (!cpus_weight(group->cpumask)) {
4681 printk(KERN_ERR "ERROR: empty group\n");
4684 if (cpus_intersects(groupmask, group->cpumask)) {
4686 printk(KERN_ERR "ERROR: repeated CPUs\n");
4689 cpus_or(groupmask, groupmask, group->cpumask);
4691 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
4694 group = group->next;
4695 } while (group != sd->groups);
4698 if (!cpus_equal(sd->span, groupmask))
4699 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
4705 if (!cpus_subset(groupmask, sd->span))
4706 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
4712 #define sched_domain_debug(sd, cpu) {}
4715 static int __devinit sd_degenerate(struct sched_domain *sd)
4717 if (cpus_weight(sd->span) == 1)
4720 /* Following flags need at least 2 groups */
4721 if (sd->flags & (SD_LOAD_BALANCE |
4722 SD_BALANCE_NEWIDLE |
4725 if (sd->groups != sd->groups->next)
4729 /* Following flags don't use groups */
4730 if (sd->flags & (SD_WAKE_IDLE |
4738 static int __devinit sd_parent_degenerate(struct sched_domain *sd,
4739 struct sched_domain *parent)
4741 unsigned long cflags = sd->flags, pflags = parent->flags;
4743 if (sd_degenerate(parent))
4746 if (!cpus_equal(sd->span, parent->span))
4749 /* Does parent contain flags not in child? */
4750 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
4751 if (cflags & SD_WAKE_AFFINE)
4752 pflags &= ~SD_WAKE_BALANCE;
4753 /* Flags needing groups don't count if only 1 group in parent */
4754 if (parent->groups == parent->groups->next) {
4755 pflags &= ~(SD_LOAD_BALANCE |
4756 SD_BALANCE_NEWIDLE |
4760 if (~cflags & pflags)
4767 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
4768 * hold the hotplug lock.
4770 void __devinit cpu_attach_domain(struct sched_domain *sd, int cpu)
4772 migration_req_t req;
4773 unsigned long flags;
4774 runqueue_t *rq = cpu_rq(cpu);
4776 struct sched_domain *tmp;
4778 /* Remove the sched domains which do not contribute to scheduling. */
4779 for (tmp = sd; tmp; tmp = tmp->parent) {
4780 struct sched_domain *parent = tmp->parent;
4783 if (sd_parent_degenerate(tmp, parent))
4784 tmp->parent = parent->parent;
4787 if (sd && sd_degenerate(sd))
4790 sched_domain_debug(sd, cpu);
4792 spin_lock_irqsave(&rq->lock, flags);
4794 if (cpu == smp_processor_id() || !cpu_online(cpu)) {
4797 init_completion(&req.done);
4798 req.type = REQ_SET_DOMAIN;
4800 list_add(&req.list, &rq->migration_queue);
4804 spin_unlock_irqrestore(&rq->lock, flags);
4807 wake_up_process(rq->migration_thread);
4808 wait_for_completion(&req.done);
4812 /* cpus with isolated domains */
4813 cpumask_t __devinitdata cpu_isolated_map = CPU_MASK_NONE;
4815 /* Setup the mask of cpus configured for isolated domains */
4816 static int __init isolated_cpu_setup(char *str)
4818 int ints[NR_CPUS], i;
4820 str = get_options(str, ARRAY_SIZE(ints), ints);
4821 cpus_clear(cpu_isolated_map);
4822 for (i = 1; i <= ints[0]; i++)
4823 if (ints[i] < NR_CPUS)
4824 cpu_set(ints[i], cpu_isolated_map);
4828 __setup ("isolcpus=", isolated_cpu_setup);
4831 * init_sched_build_groups takes an array of groups, the cpumask we wish
4832 * to span, and a pointer to a function which identifies what group a CPU
4833 * belongs to. The return value of group_fn must be a valid index into the
4834 * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
4835 * keep track of groups covered with a cpumask_t).
4837 * init_sched_build_groups will build a circular linked list of the groups
4838 * covered by the given span, and will set each group's ->cpumask correctly,
4839 * and ->cpu_power to 0.
4841 void __devinit init_sched_build_groups(struct sched_group groups[],
4842 cpumask_t span, int (*group_fn)(int cpu))
4844 struct sched_group *first = NULL, *last = NULL;
4845 cpumask_t covered = CPU_MASK_NONE;
4848 for_each_cpu_mask(i, span) {
4849 int group = group_fn(i);
4850 struct sched_group *sg = &groups[group];
4853 if (cpu_isset(i, covered))
4856 sg->cpumask = CPU_MASK_NONE;
4859 for_each_cpu_mask(j, span) {
4860 if (group_fn(j) != group)
4863 cpu_set(j, covered);
4864 cpu_set(j, sg->cpumask);
4876 #ifdef ARCH_HAS_SCHED_DOMAIN
4877 extern void __devinit arch_init_sched_domains(void);
4878 extern void __devinit arch_destroy_sched_domains(void);
4880 #ifdef CONFIG_SCHED_SMT
4881 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
4882 static struct sched_group sched_group_cpus[NR_CPUS];
4883 static int __devinit cpu_to_cpu_group(int cpu)
4889 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
4890 static struct sched_group sched_group_phys[NR_CPUS];
4891 static int __devinit cpu_to_phys_group(int cpu)
4893 #ifdef CONFIG_SCHED_SMT
4894 return first_cpu(cpu_sibling_map[cpu]);
4902 static DEFINE_PER_CPU(struct sched_domain, node_domains);
4903 static struct sched_group sched_group_nodes[MAX_NUMNODES];
4904 static int __devinit cpu_to_node_group(int cpu)
4906 return cpu_to_node(cpu);
4910 #if defined(CONFIG_SCHED_SMT) && defined(CONFIG_NUMA)
4912 * The domains setup code relies on siblings not spanning
4913 * multiple nodes. Make sure the architecture has a proper
4916 static void check_sibling_maps(void)
4920 for_each_online_cpu(i) {
4921 for_each_cpu_mask(j, cpu_sibling_map[i]) {
4922 if (cpu_to_node(i) != cpu_to_node(j)) {
4923 printk(KERN_INFO "warning: CPU %d siblings map "
4924 "to different node - isolating "
4926 cpu_sibling_map[i] = cpumask_of_cpu(i);
4935 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
4937 static void __devinit arch_init_sched_domains(void)
4940 cpumask_t cpu_default_map;
4942 #if defined(CONFIG_SCHED_SMT) && defined(CONFIG_NUMA)
4943 check_sibling_maps();
4946 * Setup mask for cpus without special case scheduling requirements.
4947 * For now this just excludes isolated cpus, but could be used to
4948 * exclude other special cases in the future.
4950 cpus_complement(cpu_default_map, cpu_isolated_map);
4951 cpus_and(cpu_default_map, cpu_default_map, cpu_online_map);
4954 * Set up domains. Isolated domains just stay on the NULL domain.
4956 for_each_cpu_mask(i, cpu_default_map) {
4958 struct sched_domain *sd = NULL, *p;
4959 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
4961 cpus_and(nodemask, nodemask, cpu_default_map);
4964 sd = &per_cpu(node_domains, i);
4965 group = cpu_to_node_group(i);
4967 sd->span = cpu_default_map;
4968 sd->groups = &sched_group_nodes[group];
4972 sd = &per_cpu(phys_domains, i);
4973 group = cpu_to_phys_group(i);
4975 sd->span = nodemask;
4977 sd->groups = &sched_group_phys[group];
4979 #ifdef CONFIG_SCHED_SMT
4981 sd = &per_cpu(cpu_domains, i);
4982 group = cpu_to_cpu_group(i);
4983 *sd = SD_SIBLING_INIT;
4984 sd->span = cpu_sibling_map[i];
4985 cpus_and(sd->span, sd->span, cpu_default_map);
4987 sd->groups = &sched_group_cpus[group];
4991 #ifdef CONFIG_SCHED_SMT
4992 /* Set up CPU (sibling) groups */
4993 for_each_online_cpu(i) {
4994 cpumask_t this_sibling_map = cpu_sibling_map[i];
4995 cpus_and(this_sibling_map, this_sibling_map, cpu_default_map);
4996 if (i != first_cpu(this_sibling_map))
4999 init_sched_build_groups(sched_group_cpus, this_sibling_map,
5004 /* Set up physical groups */
5005 for (i = 0; i < MAX_NUMNODES; i++) {
5006 cpumask_t nodemask = node_to_cpumask(i);
5008 cpus_and(nodemask, nodemask, cpu_default_map);
5009 if (cpus_empty(nodemask))
5012 init_sched_build_groups(sched_group_phys, nodemask,
5013 &cpu_to_phys_group);
5017 /* Set up node groups */
5018 init_sched_build_groups(sched_group_nodes, cpu_default_map,
5019 &cpu_to_node_group);
5022 /* Calculate CPU power for physical packages and nodes */
5023 for_each_cpu_mask(i, cpu_default_map) {
5025 struct sched_domain *sd;
5026 #ifdef CONFIG_SCHED_SMT
5027 sd = &per_cpu(cpu_domains, i);
5028 power = SCHED_LOAD_SCALE;
5029 sd->groups->cpu_power = power;
5032 sd = &per_cpu(phys_domains, i);
5033 power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
5034 (cpus_weight(sd->groups->cpumask)-1) / 10;
5035 sd->groups->cpu_power = power;
5038 if (i == first_cpu(sd->groups->cpumask)) {
5039 /* Only add "power" once for each physical package. */
5040 sd = &per_cpu(node_domains, i);
5041 sd->groups->cpu_power += power;
5046 /* Attach the domains */
5047 for_each_online_cpu(i) {
5048 struct sched_domain *sd;
5049 #ifdef CONFIG_SCHED_SMT
5050 sd = &per_cpu(cpu_domains, i);
5052 sd = &per_cpu(phys_domains, i);
5054 cpu_attach_domain(sd, i);
5058 #ifdef CONFIG_HOTPLUG_CPU
5059 static void __devinit arch_destroy_sched_domains(void)
5061 /* Do nothing: everything is statically allocated. */
5065 #endif /* ARCH_HAS_SCHED_DOMAIN */
5067 #ifdef CONFIG_HOTPLUG_CPU
5069 * Force a reinitialization of the sched domains hierarchy. The domains
5070 * and groups cannot be updated in place without racing with the balancing
5071 * code, so we temporarily attach all running cpus to the NULL domain
5072 * which will prevent rebalancing while the sched domains are recalculated.
5074 static int update_sched_domains(struct notifier_block *nfb,
5075 unsigned long action, void *hcpu)
5080 case CPU_UP_PREPARE:
5081 case CPU_DOWN_PREPARE:
5082 for_each_online_cpu(i)
5083 cpu_attach_domain(NULL, i);
5084 arch_destroy_sched_domains();
5087 case CPU_UP_CANCELED:
5088 case CPU_DOWN_FAILED:
5092 * Fall through and re-initialise the domains.
5099 /* The hotplug lock is already held by cpu_up/cpu_down */
5100 arch_init_sched_domains();
5106 void __init sched_init_smp(void)
5109 arch_init_sched_domains();
5110 unlock_cpu_hotplug();
5111 /* XXX: Theoretical race here - CPU may be hotplugged now */
5112 hotcpu_notifier(update_sched_domains, 0);
5115 void __init sched_init_smp(void)
5118 #endif /* CONFIG_SMP */
5120 int in_sched_functions(unsigned long addr)
5122 /* Linker adds these: start and end of __sched functions */
5123 extern char __sched_text_start[], __sched_text_end[];
5124 return in_lock_functions(addr) ||
5125 (addr >= (unsigned long)__sched_text_start
5126 && addr < (unsigned long)__sched_text_end);
5129 void __init sched_init(void)
5134 for (i = 0; i < NR_CPUS; i++) {
5135 prio_array_t *array;
5138 spin_lock_init(&rq->lock);
5140 rq->active = rq->arrays;
5141 rq->expired = rq->arrays + 1;
5142 rq->best_expired_prio = MAX_PRIO;
5146 for (j = 1; j < 3; j++)
5147 rq->cpu_load[j] = 0;
5148 rq->active_balance = 0;
5150 rq->migration_thread = NULL;
5151 INIT_LIST_HEAD(&rq->migration_queue);
5153 atomic_set(&rq->nr_iowait, 0);
5155 for (j = 0; j < 2; j++) {
5156 array = rq->arrays + j;
5157 for (k = 0; k < MAX_PRIO; k++) {
5158 INIT_LIST_HEAD(array->queue + k);
5159 __clear_bit(k, array->bitmap);
5161 // delimiter for bitsearch
5162 __set_bit(MAX_PRIO, array->bitmap);
5167 * The boot idle thread does lazy MMU switching as well:
5169 atomic_inc(&init_mm.mm_count);
5170 enter_lazy_tlb(&init_mm, current);
5173 * Make us the idle thread. Technically, schedule() should not be
5174 * called from this thread, however somewhere below it might be,
5175 * but because we are the idle thread, we just pick up running again
5176 * when this runqueue becomes "idle".
5178 init_idle(current, smp_processor_id());
5181 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5182 void __might_sleep(char *file, int line)
5184 #if defined(in_atomic)
5185 static unsigned long prev_jiffy; /* ratelimiting */
5187 if ((in_atomic() || irqs_disabled()) &&
5188 system_state == SYSTEM_RUNNING && !oops_in_progress) {
5189 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
5191 prev_jiffy = jiffies;
5192 printk(KERN_ERR "Debug: sleeping function called from invalid"
5193 " context at %s:%d\n", file, line);
5194 printk("in_atomic():%d, irqs_disabled():%d\n",
5195 in_atomic(), irqs_disabled());
5200 EXPORT_SYMBOL(__might_sleep);
5203 #ifdef CONFIG_MAGIC_SYSRQ
5204 void normalize_rt_tasks(void)
5206 struct task_struct *p;
5207 prio_array_t *array;
5208 unsigned long flags;
5211 read_lock_irq(&tasklist_lock);
5212 for_each_process (p) {
5216 rq = task_rq_lock(p, &flags);
5220 deactivate_task(p, task_rq(p));
5221 __setscheduler(p, SCHED_NORMAL, 0);
5223 __activate_task(p, task_rq(p));
5224 resched_task(rq->curr);
5227 task_rq_unlock(rq, &flags);
5229 read_unlock_irq(&tasklist_lock);
5232 #endif /* CONFIG_MAGIC_SYSRQ */