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);
264 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
265 * See update_sched_domains: synchronize_kernel for details.
267 * The domain tree of any CPU may only be accessed from within
268 * preempt-disabled sections.
270 #define for_each_domain(cpu, domain) \
271 for (domain = rcu_dereference(cpu_rq(cpu)->sd); domain; domain = domain->parent)
273 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
274 #define this_rq() (&__get_cpu_var(runqueues))
275 #define task_rq(p) cpu_rq(task_cpu(p))
276 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
278 #ifndef prepare_arch_switch
279 # define prepare_arch_switch(next) do { } while (0)
281 #ifndef finish_arch_switch
282 # define finish_arch_switch(prev) do { } while (0)
285 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
286 static inline int task_running(runqueue_t *rq, task_t *p)
288 return rq->curr == p;
291 static inline void prepare_lock_switch(runqueue_t *rq, task_t *next)
295 static inline void finish_lock_switch(runqueue_t *rq, task_t *prev)
297 spin_unlock_irq(&rq->lock);
300 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
301 static inline int task_running(runqueue_t *rq, task_t *p)
306 return rq->curr == p;
310 static inline void prepare_lock_switch(runqueue_t *rq, task_t *next)
314 * We can optimise this out completely for !SMP, because the
315 * SMP rebalancing from interrupt is the only thing that cares
320 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
321 spin_unlock_irq(&rq->lock);
323 spin_unlock(&rq->lock);
327 static inline void finish_lock_switch(runqueue_t *rq, task_t *prev)
331 * After ->oncpu is cleared, the task can be moved to a different CPU.
332 * We must ensure this doesn't happen until the switch is completely
338 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
342 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
345 * task_rq_lock - lock the runqueue a given task resides on and disable
346 * interrupts. Note the ordering: we can safely lookup the task_rq without
347 * explicitly disabling preemption.
349 static inline runqueue_t *task_rq_lock(task_t *p, unsigned long *flags)
355 local_irq_save(*flags);
357 spin_lock(&rq->lock);
358 if (unlikely(rq != task_rq(p))) {
359 spin_unlock_irqrestore(&rq->lock, *flags);
360 goto repeat_lock_task;
365 static inline void task_rq_unlock(runqueue_t *rq, unsigned long *flags)
368 spin_unlock_irqrestore(&rq->lock, *flags);
371 #ifdef CONFIG_SCHEDSTATS
373 * bump this up when changing the output format or the meaning of an existing
374 * format, so that tools can adapt (or abort)
376 #define SCHEDSTAT_VERSION 12
378 static int show_schedstat(struct seq_file *seq, void *v)
382 seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION);
383 seq_printf(seq, "timestamp %lu\n", jiffies);
384 for_each_online_cpu(cpu) {
385 runqueue_t *rq = cpu_rq(cpu);
387 struct sched_domain *sd;
391 /* runqueue-specific stats */
393 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
394 cpu, rq->yld_both_empty,
395 rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt,
396 rq->sched_switch, rq->sched_cnt, rq->sched_goidle,
397 rq->ttwu_cnt, rq->ttwu_local,
398 rq->rq_sched_info.cpu_time,
399 rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt);
401 seq_printf(seq, "\n");
404 /* domain-specific stats */
406 for_each_domain(cpu, sd) {
407 enum idle_type itype;
408 char mask_str[NR_CPUS];
410 cpumask_scnprintf(mask_str, NR_CPUS, sd->span);
411 seq_printf(seq, "domain%d %s", dcnt++, mask_str);
412 for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES;
414 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu",
416 sd->lb_balanced[itype],
417 sd->lb_failed[itype],
418 sd->lb_imbalance[itype],
419 sd->lb_gained[itype],
420 sd->lb_hot_gained[itype],
421 sd->lb_nobusyq[itype],
422 sd->lb_nobusyg[itype]);
424 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
425 sd->alb_cnt, sd->alb_failed, sd->alb_pushed,
426 sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed,
427 sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed,
428 sd->ttwu_wake_remote, sd->ttwu_move_affine, sd->ttwu_move_balance);
436 static int schedstat_open(struct inode *inode, struct file *file)
438 unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32);
439 char *buf = kmalloc(size, GFP_KERNEL);
445 res = single_open(file, show_schedstat, NULL);
447 m = file->private_data;
455 struct file_operations proc_schedstat_operations = {
456 .open = schedstat_open,
459 .release = single_release,
462 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
463 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
464 #else /* !CONFIG_SCHEDSTATS */
465 # define schedstat_inc(rq, field) do { } while (0)
466 # define schedstat_add(rq, field, amt) do { } while (0)
470 * rq_lock - lock a given runqueue and disable interrupts.
472 static inline runqueue_t *this_rq_lock(void)
479 spin_lock(&rq->lock);
484 #ifdef CONFIG_SCHEDSTATS
486 * Called when a process is dequeued from the active array and given
487 * the cpu. We should note that with the exception of interactive
488 * tasks, the expired queue will become the active queue after the active
489 * queue is empty, without explicitly dequeuing and requeuing tasks in the
490 * expired queue. (Interactive tasks may be requeued directly to the
491 * active queue, thus delaying tasks in the expired queue from running;
492 * see scheduler_tick()).
494 * This function is only called from sched_info_arrive(), rather than
495 * dequeue_task(). Even though a task may be queued and dequeued multiple
496 * times as it is shuffled about, we're really interested in knowing how
497 * long it was from the *first* time it was queued to the time that it
500 static inline void sched_info_dequeued(task_t *t)
502 t->sched_info.last_queued = 0;
506 * Called when a task finally hits the cpu. We can now calculate how
507 * long it was waiting to run. We also note when it began so that we
508 * can keep stats on how long its timeslice is.
510 static inline void sched_info_arrive(task_t *t)
512 unsigned long now = jiffies, diff = 0;
513 struct runqueue *rq = task_rq(t);
515 if (t->sched_info.last_queued)
516 diff = now - t->sched_info.last_queued;
517 sched_info_dequeued(t);
518 t->sched_info.run_delay += diff;
519 t->sched_info.last_arrival = now;
520 t->sched_info.pcnt++;
525 rq->rq_sched_info.run_delay += diff;
526 rq->rq_sched_info.pcnt++;
530 * Called when a process is queued into either the active or expired
531 * array. The time is noted and later used to determine how long we
532 * had to wait for us to reach the cpu. Since the expired queue will
533 * become the active queue after active queue is empty, without dequeuing
534 * and requeuing any tasks, we are interested in queuing to either. It
535 * is unusual but not impossible for tasks to be dequeued and immediately
536 * requeued in the same or another array: this can happen in sched_yield(),
537 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
540 * This function is only called from enqueue_task(), but also only updates
541 * the timestamp if it is already not set. It's assumed that
542 * sched_info_dequeued() will clear that stamp when appropriate.
544 static inline void sched_info_queued(task_t *t)
546 if (!t->sched_info.last_queued)
547 t->sched_info.last_queued = jiffies;
551 * Called when a process ceases being the active-running process, either
552 * voluntarily or involuntarily. Now we can calculate how long we ran.
554 static inline void sched_info_depart(task_t *t)
556 struct runqueue *rq = task_rq(t);
557 unsigned long diff = jiffies - t->sched_info.last_arrival;
559 t->sched_info.cpu_time += diff;
562 rq->rq_sched_info.cpu_time += diff;
566 * Called when tasks are switched involuntarily due, typically, to expiring
567 * their time slice. (This may also be called when switching to or from
568 * the idle task.) We are only called when prev != next.
570 static inline void sched_info_switch(task_t *prev, task_t *next)
572 struct runqueue *rq = task_rq(prev);
575 * prev now departs the cpu. It's not interesting to record
576 * stats about how efficient we were at scheduling the idle
579 if (prev != rq->idle)
580 sched_info_depart(prev);
582 if (next != rq->idle)
583 sched_info_arrive(next);
586 #define sched_info_queued(t) do { } while (0)
587 #define sched_info_switch(t, next) do { } while (0)
588 #endif /* CONFIG_SCHEDSTATS */
591 * Adding/removing a task to/from a priority array:
593 static void dequeue_task(struct task_struct *p, prio_array_t *array)
596 list_del(&p->run_list);
597 if (list_empty(array->queue + p->prio))
598 __clear_bit(p->prio, array->bitmap);
601 static void enqueue_task(struct task_struct *p, prio_array_t *array)
603 sched_info_queued(p);
604 list_add_tail(&p->run_list, array->queue + p->prio);
605 __set_bit(p->prio, array->bitmap);
611 * Put task to the end of the run list without the overhead of dequeue
612 * followed by enqueue.
614 static void requeue_task(struct task_struct *p, prio_array_t *array)
616 list_move_tail(&p->run_list, array->queue + p->prio);
619 static inline void enqueue_task_head(struct task_struct *p, prio_array_t *array)
621 list_add(&p->run_list, array->queue + p->prio);
622 __set_bit(p->prio, array->bitmap);
628 * effective_prio - return the priority that is based on the static
629 * priority but is modified by bonuses/penalties.
631 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
632 * into the -5 ... 0 ... +5 bonus/penalty range.
634 * We use 25% of the full 0...39 priority range so that:
636 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
637 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
639 * Both properties are important to certain workloads.
641 static int effective_prio(task_t *p)
648 bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
650 prio = p->static_prio - bonus;
651 if (prio < MAX_RT_PRIO)
653 if (prio > MAX_PRIO-1)
659 * __activate_task - move a task to the runqueue.
661 static inline void __activate_task(task_t *p, runqueue_t *rq)
663 enqueue_task(p, rq->active);
668 * __activate_idle_task - move idle task to the _front_ of runqueue.
670 static inline void __activate_idle_task(task_t *p, runqueue_t *rq)
672 enqueue_task_head(p, rq->active);
676 static void recalc_task_prio(task_t *p, unsigned long long now)
678 /* Caller must always ensure 'now >= p->timestamp' */
679 unsigned long long __sleep_time = now - p->timestamp;
680 unsigned long sleep_time;
682 if (__sleep_time > NS_MAX_SLEEP_AVG)
683 sleep_time = NS_MAX_SLEEP_AVG;
685 sleep_time = (unsigned long)__sleep_time;
687 if (likely(sleep_time > 0)) {
689 * User tasks that sleep a long time are categorised as
690 * idle and will get just interactive status to stay active &
691 * prevent them suddenly becoming cpu hogs and starving
694 if (p->mm && p->activated != -1 &&
695 sleep_time > INTERACTIVE_SLEEP(p)) {
696 p->sleep_avg = JIFFIES_TO_NS(MAX_SLEEP_AVG -
700 * The lower the sleep avg a task has the more
701 * rapidly it will rise with sleep time.
703 sleep_time *= (MAX_BONUS - CURRENT_BONUS(p)) ? : 1;
706 * Tasks waking from uninterruptible sleep are
707 * limited in their sleep_avg rise as they
708 * are likely to be waiting on I/O
710 if (p->activated == -1 && p->mm) {
711 if (p->sleep_avg >= INTERACTIVE_SLEEP(p))
713 else if (p->sleep_avg + sleep_time >=
714 INTERACTIVE_SLEEP(p)) {
715 p->sleep_avg = INTERACTIVE_SLEEP(p);
721 * This code gives a bonus to interactive tasks.
723 * The boost works by updating the 'average sleep time'
724 * value here, based on ->timestamp. The more time a
725 * task spends sleeping, the higher the average gets -
726 * and the higher the priority boost gets as well.
728 p->sleep_avg += sleep_time;
730 if (p->sleep_avg > NS_MAX_SLEEP_AVG)
731 p->sleep_avg = NS_MAX_SLEEP_AVG;
735 p->prio = effective_prio(p);
739 * activate_task - move a task to the runqueue and do priority recalculation
741 * Update all the scheduling statistics stuff. (sleep average
742 * calculation, priority modifiers, etc.)
744 static void activate_task(task_t *p, runqueue_t *rq, int local)
746 unsigned long long now;
751 /* Compensate for drifting sched_clock */
752 runqueue_t *this_rq = this_rq();
753 now = (now - this_rq->timestamp_last_tick)
754 + rq->timestamp_last_tick;
758 recalc_task_prio(p, now);
761 * This checks to make sure it's not an uninterruptible task
762 * that is now waking up.
766 * Tasks which were woken up by interrupts (ie. hw events)
767 * are most likely of interactive nature. So we give them
768 * the credit of extending their sleep time to the period
769 * of time they spend on the runqueue, waiting for execution
770 * on a CPU, first time around:
776 * Normal first-time wakeups get a credit too for
777 * on-runqueue time, but it will be weighted down:
784 __activate_task(p, rq);
788 * deactivate_task - remove a task from the runqueue.
790 static void deactivate_task(struct task_struct *p, runqueue_t *rq)
793 dequeue_task(p, p->array);
798 * resched_task - mark a task 'to be rescheduled now'.
800 * On UP this means the setting of the need_resched flag, on SMP it
801 * might also involve a cross-CPU call to trigger the scheduler on
805 static void resched_task(task_t *p)
807 int need_resched, nrpolling;
809 assert_spin_locked(&task_rq(p)->lock);
811 /* minimise the chance of sending an interrupt to poll_idle() */
812 nrpolling = test_tsk_thread_flag(p,TIF_POLLING_NRFLAG);
813 need_resched = test_and_set_tsk_thread_flag(p,TIF_NEED_RESCHED);
814 nrpolling |= test_tsk_thread_flag(p,TIF_POLLING_NRFLAG);
816 if (!need_resched && !nrpolling && (task_cpu(p) != smp_processor_id()))
817 smp_send_reschedule(task_cpu(p));
820 static inline void resched_task(task_t *p)
822 set_tsk_need_resched(p);
827 * task_curr - is this task currently executing on a CPU?
828 * @p: the task in question.
830 inline int task_curr(const task_t *p)
832 return cpu_curr(task_cpu(p)) == p;
837 struct list_head list;
842 struct completion done;
846 * The task's runqueue lock must be held.
847 * Returns true if you have to wait for migration thread.
849 static int migrate_task(task_t *p, int dest_cpu, migration_req_t *req)
851 runqueue_t *rq = task_rq(p);
854 * If the task is not on a runqueue (and not running), then
855 * it is sufficient to simply update the task's cpu field.
857 if (!p->array && !task_running(rq, p)) {
858 set_task_cpu(p, dest_cpu);
862 init_completion(&req->done);
864 req->dest_cpu = dest_cpu;
865 list_add(&req->list, &rq->migration_queue);
870 * wait_task_inactive - wait for a thread to unschedule.
872 * The caller must ensure that the task *will* unschedule sometime soon,
873 * else this function might spin for a *long* time. This function can't
874 * be called with interrupts off, or it may introduce deadlock with
875 * smp_call_function() if an IPI is sent by the same process we are
876 * waiting to become inactive.
878 void wait_task_inactive(task_t * p)
885 rq = task_rq_lock(p, &flags);
886 /* Must be off runqueue entirely, not preempted. */
887 if (unlikely(p->array || task_running(rq, p))) {
888 /* If it's preempted, we yield. It could be a while. */
889 preempted = !task_running(rq, p);
890 task_rq_unlock(rq, &flags);
896 task_rq_unlock(rq, &flags);
900 * kick_process - kick a running thread to enter/exit the kernel
901 * @p: the to-be-kicked thread
903 * Cause a process which is running on another CPU to enter
904 * kernel-mode, without any delay. (to get signals handled.)
906 * NOTE: this function doesnt have to take the runqueue lock,
907 * because all it wants to ensure is that the remote task enters
908 * the kernel. If the IPI races and the task has been migrated
909 * to another CPU then no harm is done and the purpose has been
912 void kick_process(task_t *p)
918 if ((cpu != smp_processor_id()) && task_curr(p))
919 smp_send_reschedule(cpu);
924 * Return a low guess at the load of a migration-source cpu.
926 * We want to under-estimate the load of migration sources, to
927 * balance conservatively.
929 static inline unsigned long source_load(int cpu, int type)
931 runqueue_t *rq = cpu_rq(cpu);
932 unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE;
936 return min(rq->cpu_load[type-1], load_now);
940 * Return a high guess at the load of a migration-target cpu
942 static inline unsigned long target_load(int cpu, int type)
944 runqueue_t *rq = cpu_rq(cpu);
945 unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE;
949 return max(rq->cpu_load[type-1], load_now);
953 * find_idlest_group finds and returns the least busy CPU group within the
956 static struct sched_group *
957 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
959 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
960 unsigned long min_load = ULONG_MAX, this_load = 0;
961 int load_idx = sd->forkexec_idx;
962 int imbalance = 100 + (sd->imbalance_pct-100)/2;
965 unsigned long load, avg_load;
969 local_group = cpu_isset(this_cpu, group->cpumask);
970 /* XXX: put a cpus allowed check */
972 /* Tally up the load of all CPUs in the group */
975 for_each_cpu_mask(i, group->cpumask) {
976 /* Bias balancing toward cpus of our domain */
978 load = source_load(i, load_idx);
980 load = target_load(i, load_idx);
985 /* Adjust by relative CPU power of the group */
986 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
989 this_load = avg_load;
991 } else if (avg_load < min_load) {
996 } while (group != sd->groups);
998 if (!idlest || 100*this_load < imbalance*min_load)
1004 * find_idlest_queue - find the idlest runqueue among the cpus in group.
1006 static int find_idlest_cpu(struct sched_group *group, int this_cpu)
1008 unsigned long load, min_load = ULONG_MAX;
1012 for_each_cpu_mask(i, group->cpumask) {
1013 load = source_load(i, 0);
1015 if (load < min_load || (load == min_load && i == this_cpu)) {
1028 * wake_idle() will wake a task on an idle cpu if task->cpu is
1029 * not idle and an idle cpu is available. The span of cpus to
1030 * search starts with cpus closest then further out as needed,
1031 * so we always favor a closer, idle cpu.
1033 * Returns the CPU we should wake onto.
1035 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1036 static int wake_idle(int cpu, task_t *p)
1039 struct sched_domain *sd;
1045 for_each_domain(cpu, sd) {
1046 if (sd->flags & SD_WAKE_IDLE) {
1047 cpus_and(tmp, sd->span, p->cpus_allowed);
1048 for_each_cpu_mask(i, tmp) {
1059 static inline int wake_idle(int cpu, task_t *p)
1066 * try_to_wake_up - wake up a thread
1067 * @p: the to-be-woken-up thread
1068 * @state: the mask of task states that can be woken
1069 * @sync: do a synchronous wakeup?
1071 * Put it on the run-queue if it's not already there. The "current"
1072 * thread is always on the run-queue (except when the actual
1073 * re-schedule is in progress), and as such you're allowed to do
1074 * the simpler "current->state = TASK_RUNNING" to mark yourself
1075 * runnable without the overhead of this.
1077 * returns failure only if the task is already active.
1079 static int try_to_wake_up(task_t * p, unsigned int state, int sync)
1081 int cpu, this_cpu, success = 0;
1082 unsigned long flags;
1086 unsigned long load, this_load;
1087 struct sched_domain *sd, *this_sd = NULL;
1091 rq = task_rq_lock(p, &flags);
1092 old_state = p->state;
1093 if (!(old_state & state))
1100 this_cpu = smp_processor_id();
1103 if (unlikely(task_running(rq, p)))
1108 schedstat_inc(rq, ttwu_cnt);
1109 if (cpu == this_cpu) {
1110 schedstat_inc(rq, ttwu_local);
1114 for_each_domain(this_cpu, sd) {
1115 if (cpu_isset(cpu, sd->span)) {
1116 schedstat_inc(sd, ttwu_wake_remote);
1122 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1126 * Check for affine wakeup and passive balancing possibilities.
1129 int idx = this_sd->wake_idx;
1130 unsigned int imbalance;
1132 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1134 load = source_load(cpu, idx);
1135 this_load = target_load(this_cpu, idx);
1137 new_cpu = this_cpu; /* Wake to this CPU if we can */
1139 if (this_sd->flags & SD_WAKE_AFFINE) {
1140 unsigned long tl = this_load;
1142 * If sync wakeup then subtract the (maximum possible)
1143 * effect of the currently running task from the load
1144 * of the current CPU:
1147 tl -= SCHED_LOAD_SCALE;
1150 tl + target_load(cpu, idx) <= SCHED_LOAD_SCALE) ||
1151 100*(tl + SCHED_LOAD_SCALE) <= imbalance*load) {
1153 * This domain has SD_WAKE_AFFINE and
1154 * p is cache cold in this domain, and
1155 * there is no bad imbalance.
1157 schedstat_inc(this_sd, ttwu_move_affine);
1163 * Start passive balancing when half the imbalance_pct
1166 if (this_sd->flags & SD_WAKE_BALANCE) {
1167 if (imbalance*this_load <= 100*load) {
1168 schedstat_inc(this_sd, ttwu_move_balance);
1174 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1176 new_cpu = wake_idle(new_cpu, p);
1177 if (new_cpu != cpu) {
1178 set_task_cpu(p, new_cpu);
1179 task_rq_unlock(rq, &flags);
1180 /* might preempt at this point */
1181 rq = task_rq_lock(p, &flags);
1182 old_state = p->state;
1183 if (!(old_state & state))
1188 this_cpu = smp_processor_id();
1193 #endif /* CONFIG_SMP */
1194 if (old_state == TASK_UNINTERRUPTIBLE) {
1195 rq->nr_uninterruptible--;
1197 * Tasks on involuntary sleep don't earn
1198 * sleep_avg beyond just interactive state.
1204 * Sync wakeups (i.e. those types of wakeups where the waker
1205 * has indicated that it will leave the CPU in short order)
1206 * don't trigger a preemption, if the woken up task will run on
1207 * this cpu. (in this case the 'I will reschedule' promise of
1208 * the waker guarantees that the freshly woken up task is going
1209 * to be considered on this CPU.)
1211 activate_task(p, rq, cpu == this_cpu);
1212 if (!sync || cpu != this_cpu) {
1213 if (TASK_PREEMPTS_CURR(p, rq))
1214 resched_task(rq->curr);
1219 p->state = TASK_RUNNING;
1221 task_rq_unlock(rq, &flags);
1226 int fastcall wake_up_process(task_t * p)
1228 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1229 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1232 EXPORT_SYMBOL(wake_up_process);
1234 int fastcall wake_up_state(task_t *p, unsigned int state)
1236 return try_to_wake_up(p, state, 0);
1240 * Perform scheduler related setup for a newly forked process p.
1241 * p is forked by current.
1243 void fastcall sched_fork(task_t *p)
1246 * We mark the process as running here, but have not actually
1247 * inserted it onto the runqueue yet. This guarantees that
1248 * nobody will actually run it, and a signal or other external
1249 * event cannot wake it up and insert it on the runqueue either.
1251 p->state = TASK_RUNNING;
1252 INIT_LIST_HEAD(&p->run_list);
1254 #ifdef CONFIG_SCHEDSTATS
1255 memset(&p->sched_info, 0, sizeof(p->sched_info));
1257 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1260 #ifdef CONFIG_PREEMPT
1261 /* Want to start with kernel preemption disabled. */
1262 p->thread_info->preempt_count = 1;
1265 * Share the timeslice between parent and child, thus the
1266 * total amount of pending timeslices in the system doesn't change,
1267 * resulting in more scheduling fairness.
1269 local_irq_disable();
1270 p->time_slice = (current->time_slice + 1) >> 1;
1272 * The remainder of the first timeslice might be recovered by
1273 * the parent if the child exits early enough.
1275 p->first_time_slice = 1;
1276 current->time_slice >>= 1;
1277 p->timestamp = sched_clock();
1278 if (unlikely(!current->time_slice)) {
1280 * This case is rare, it happens when the parent has only
1281 * a single jiffy left from its timeslice. Taking the
1282 * runqueue lock is not a problem.
1284 current->time_slice = 1;
1294 * wake_up_new_task - wake up a newly created task for the first time.
1296 * This function will do some initial scheduler statistics housekeeping
1297 * that must be done for every newly created context, then puts the task
1298 * on the runqueue and wakes it.
1300 void fastcall wake_up_new_task(task_t * p, unsigned long clone_flags)
1302 unsigned long flags;
1304 runqueue_t *rq, *this_rq;
1306 struct sched_domain *tmp, *sd = NULL;
1309 rq = task_rq_lock(p, &flags);
1310 BUG_ON(p->state != TASK_RUNNING);
1311 this_cpu = smp_processor_id();
1315 for_each_domain(cpu, tmp)
1316 if (tmp->flags & SD_BALANCE_FORK)
1322 struct sched_group *group;
1325 schedstat_inc(sd, sbf_cnt);
1328 group = find_idlest_group(sd, p, cpu);
1330 schedstat_inc(sd, sbf_balanced);
1334 new_cpu = find_idlest_cpu(group, cpu);
1335 if (new_cpu == -1 || new_cpu == cpu) {
1336 schedstat_inc(sd, sbf_balanced);
1340 if (cpu_isset(new_cpu, p->cpus_allowed)) {
1341 schedstat_inc(sd, sbf_pushed);
1342 set_task_cpu(p, new_cpu);
1343 task_rq_unlock(rq, &flags);
1344 rq = task_rq_lock(p, &flags);
1348 /* Now try balancing at a lower domain level */
1351 for_each_domain(cpu, tmp) {
1352 if (cpus_subset(span, tmp->span))
1354 if (tmp->flags & SD_BALANCE_FORK)
1364 * We decrease the sleep average of forking parents
1365 * and children as well, to keep max-interactive tasks
1366 * from forking tasks that are max-interactive. The parent
1367 * (current) is done further down, under its lock.
1369 p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1370 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1372 p->prio = effective_prio(p);
1374 if (likely(cpu == this_cpu)) {
1375 if (!(clone_flags & CLONE_VM)) {
1377 * The VM isn't cloned, so we're in a good position to
1378 * do child-runs-first in anticipation of an exec. This
1379 * usually avoids a lot of COW overhead.
1381 if (unlikely(!current->array))
1382 __activate_task(p, rq);
1384 p->prio = current->prio;
1385 list_add_tail(&p->run_list, ¤t->run_list);
1386 p->array = current->array;
1387 p->array->nr_active++;
1392 /* Run child last */
1393 __activate_task(p, rq);
1395 * We skip the following code due to cpu == this_cpu
1397 * task_rq_unlock(rq, &flags);
1398 * this_rq = task_rq_lock(current, &flags);
1402 this_rq = cpu_rq(this_cpu);
1405 * Not the local CPU - must adjust timestamp. This should
1406 * get optimised away in the !CONFIG_SMP case.
1408 p->timestamp = (p->timestamp - this_rq->timestamp_last_tick)
1409 + rq->timestamp_last_tick;
1410 __activate_task(p, rq);
1411 if (TASK_PREEMPTS_CURR(p, rq))
1412 resched_task(rq->curr);
1415 * Parent and child are on different CPUs, now get the
1416 * parent runqueue to update the parent's ->sleep_avg:
1418 task_rq_unlock(rq, &flags);
1419 this_rq = task_rq_lock(current, &flags);
1421 current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1422 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1423 task_rq_unlock(this_rq, &flags);
1427 * Potentially available exiting-child timeslices are
1428 * retrieved here - this way the parent does not get
1429 * penalized for creating too many threads.
1431 * (this cannot be used to 'generate' timeslices
1432 * artificially, because any timeslice recovered here
1433 * was given away by the parent in the first place.)
1435 void fastcall sched_exit(task_t * p)
1437 unsigned long flags;
1441 * If the child was a (relative-) CPU hog then decrease
1442 * the sleep_avg of the parent as well.
1444 rq = task_rq_lock(p->parent, &flags);
1445 if (p->first_time_slice) {
1446 p->parent->time_slice += p->time_slice;
1447 if (unlikely(p->parent->time_slice > task_timeslice(p)))
1448 p->parent->time_slice = task_timeslice(p);
1450 if (p->sleep_avg < p->parent->sleep_avg)
1451 p->parent->sleep_avg = p->parent->sleep_avg /
1452 (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg /
1454 task_rq_unlock(rq, &flags);
1458 * prepare_task_switch - prepare to switch tasks
1459 * @rq: the runqueue preparing to switch
1460 * @next: the task we are going to switch to.
1462 * This is called with the rq lock held and interrupts off. It must
1463 * be paired with a subsequent finish_task_switch after the context
1466 * prepare_task_switch sets up locking and calls architecture specific
1469 static inline void prepare_task_switch(runqueue_t *rq, task_t *next)
1471 prepare_lock_switch(rq, next);
1472 prepare_arch_switch(next);
1476 * finish_task_switch - clean up after a task-switch
1477 * @prev: the thread we just switched away from.
1479 * finish_task_switch must be called after the context switch, paired
1480 * with a prepare_task_switch call before the context switch.
1481 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1482 * and do any other architecture-specific cleanup actions.
1484 * Note that we may have delayed dropping an mm in context_switch(). If
1485 * so, we finish that here outside of the runqueue lock. (Doing it
1486 * with the lock held can cause deadlocks; see schedule() for
1489 static inline void finish_task_switch(runqueue_t *rq, task_t *prev)
1490 __releases(rq->lock)
1492 struct mm_struct *mm = rq->prev_mm;
1493 unsigned long prev_task_flags;
1498 * A task struct has one reference for the use as "current".
1499 * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1500 * calls schedule one last time. The schedule call will never return,
1501 * and the scheduled task must drop that reference.
1502 * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1503 * still held, otherwise prev could be scheduled on another cpu, die
1504 * there before we look at prev->state, and then the reference would
1506 * Manfred Spraul <manfred@colorfullife.com>
1508 prev_task_flags = prev->flags;
1509 finish_arch_switch(prev);
1510 finish_lock_switch(rq, prev);
1513 if (unlikely(prev_task_flags & PF_DEAD))
1514 put_task_struct(prev);
1518 * schedule_tail - first thing a freshly forked thread must call.
1519 * @prev: the thread we just switched away from.
1521 asmlinkage void schedule_tail(task_t *prev)
1522 __releases(rq->lock)
1524 runqueue_t *rq = this_rq();
1525 finish_task_switch(rq, prev);
1526 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1527 /* In this case, finish_task_switch does not reenable preemption */
1530 if (current->set_child_tid)
1531 put_user(current->pid, current->set_child_tid);
1535 * context_switch - switch to the new MM and the new
1536 * thread's register state.
1539 task_t * context_switch(runqueue_t *rq, task_t *prev, task_t *next)
1541 struct mm_struct *mm = next->mm;
1542 struct mm_struct *oldmm = prev->active_mm;
1544 if (unlikely(!mm)) {
1545 next->active_mm = oldmm;
1546 atomic_inc(&oldmm->mm_count);
1547 enter_lazy_tlb(oldmm, next);
1549 switch_mm(oldmm, mm, next);
1551 if (unlikely(!prev->mm)) {
1552 prev->active_mm = NULL;
1553 WARN_ON(rq->prev_mm);
1554 rq->prev_mm = oldmm;
1557 /* Here we just switch the register state and the stack. */
1558 switch_to(prev, next, prev);
1564 * nr_running, nr_uninterruptible and nr_context_switches:
1566 * externally visible scheduler statistics: current number of runnable
1567 * threads, current number of uninterruptible-sleeping threads, total
1568 * number of context switches performed since bootup.
1570 unsigned long nr_running(void)
1572 unsigned long i, sum = 0;
1574 for_each_online_cpu(i)
1575 sum += cpu_rq(i)->nr_running;
1580 unsigned long nr_uninterruptible(void)
1582 unsigned long i, sum = 0;
1585 sum += cpu_rq(i)->nr_uninterruptible;
1588 * Since we read the counters lockless, it might be slightly
1589 * inaccurate. Do not allow it to go below zero though:
1591 if (unlikely((long)sum < 0))
1597 unsigned long long nr_context_switches(void)
1599 unsigned long long i, sum = 0;
1602 sum += cpu_rq(i)->nr_switches;
1607 unsigned long nr_iowait(void)
1609 unsigned long i, sum = 0;
1612 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1620 * double_rq_lock - safely lock two runqueues
1622 * Note this does not disable interrupts like task_rq_lock,
1623 * you need to do so manually before calling.
1625 static void double_rq_lock(runqueue_t *rq1, runqueue_t *rq2)
1626 __acquires(rq1->lock)
1627 __acquires(rq2->lock)
1630 spin_lock(&rq1->lock);
1631 __acquire(rq2->lock); /* Fake it out ;) */
1634 spin_lock(&rq1->lock);
1635 spin_lock(&rq2->lock);
1637 spin_lock(&rq2->lock);
1638 spin_lock(&rq1->lock);
1644 * double_rq_unlock - safely unlock two runqueues
1646 * Note this does not restore interrupts like task_rq_unlock,
1647 * you need to do so manually after calling.
1649 static void double_rq_unlock(runqueue_t *rq1, runqueue_t *rq2)
1650 __releases(rq1->lock)
1651 __releases(rq2->lock)
1653 spin_unlock(&rq1->lock);
1655 spin_unlock(&rq2->lock);
1657 __release(rq2->lock);
1661 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1663 static void double_lock_balance(runqueue_t *this_rq, runqueue_t *busiest)
1664 __releases(this_rq->lock)
1665 __acquires(busiest->lock)
1666 __acquires(this_rq->lock)
1668 if (unlikely(!spin_trylock(&busiest->lock))) {
1669 if (busiest < this_rq) {
1670 spin_unlock(&this_rq->lock);
1671 spin_lock(&busiest->lock);
1672 spin_lock(&this_rq->lock);
1674 spin_lock(&busiest->lock);
1679 * If dest_cpu is allowed for this process, migrate the task to it.
1680 * This is accomplished by forcing the cpu_allowed mask to only
1681 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
1682 * the cpu_allowed mask is restored.
1684 static void sched_migrate_task(task_t *p, int dest_cpu)
1686 migration_req_t req;
1688 unsigned long flags;
1690 rq = task_rq_lock(p, &flags);
1691 if (!cpu_isset(dest_cpu, p->cpus_allowed)
1692 || unlikely(cpu_is_offline(dest_cpu)))
1695 /* force the process onto the specified CPU */
1696 if (migrate_task(p, dest_cpu, &req)) {
1697 /* Need to wait for migration thread (might exit: take ref). */
1698 struct task_struct *mt = rq->migration_thread;
1699 get_task_struct(mt);
1700 task_rq_unlock(rq, &flags);
1701 wake_up_process(mt);
1702 put_task_struct(mt);
1703 wait_for_completion(&req.done);
1707 task_rq_unlock(rq, &flags);
1711 * sched_exec(): find the highest-level, exec-balance-capable
1712 * domain and try to migrate the task to the least loaded CPU.
1714 * execve() is a valuable balancing opportunity, because at this point
1715 * the task has the smallest effective memory and cache footprint.
1717 void sched_exec(void)
1719 struct sched_domain *tmp, *sd = NULL;
1720 int new_cpu, this_cpu = get_cpu();
1722 for_each_domain(this_cpu, tmp)
1723 if (tmp->flags & SD_BALANCE_EXEC)
1728 struct sched_group *group;
1730 schedstat_inc(sd, sbe_cnt);
1732 group = find_idlest_group(sd, current, this_cpu);
1734 schedstat_inc(sd, sbe_balanced);
1737 new_cpu = find_idlest_cpu(group, this_cpu);
1738 if (new_cpu == -1 || new_cpu == this_cpu) {
1739 schedstat_inc(sd, sbe_balanced);
1743 schedstat_inc(sd, sbe_pushed);
1745 sched_migrate_task(current, new_cpu);
1747 /* Now try balancing at a lower domain level */
1748 this_cpu = get_cpu();
1751 for_each_domain(this_cpu, tmp) {
1752 if (cpus_subset(span, tmp->span))
1754 if (tmp->flags & SD_BALANCE_EXEC)
1766 * pull_task - move a task from a remote runqueue to the local runqueue.
1767 * Both runqueues must be locked.
1770 void pull_task(runqueue_t *src_rq, prio_array_t *src_array, task_t *p,
1771 runqueue_t *this_rq, prio_array_t *this_array, int this_cpu)
1773 dequeue_task(p, src_array);
1774 src_rq->nr_running--;
1775 set_task_cpu(p, this_cpu);
1776 this_rq->nr_running++;
1777 enqueue_task(p, this_array);
1778 p->timestamp = (p->timestamp - src_rq->timestamp_last_tick)
1779 + this_rq->timestamp_last_tick;
1781 * Note that idle threads have a prio of MAX_PRIO, for this test
1782 * to be always true for them.
1784 if (TASK_PREEMPTS_CURR(p, this_rq))
1785 resched_task(this_rq->curr);
1789 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1792 int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu,
1793 struct sched_domain *sd, enum idle_type idle, int *all_pinned)
1796 * We do not migrate tasks that are:
1797 * 1) running (obviously), or
1798 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1799 * 3) are cache-hot on their current CPU.
1801 if (!cpu_isset(this_cpu, p->cpus_allowed))
1805 if (task_running(rq, p))
1809 * Aggressive migration if:
1810 * 1) task is cache cold, or
1811 * 2) too many balance attempts have failed.
1814 if (sd->nr_balance_failed > sd->cache_nice_tries)
1817 if (task_hot(p, rq->timestamp_last_tick, sd))
1823 * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq,
1824 * as part of a balancing operation within "domain". Returns the number of
1827 * Called with both runqueues locked.
1829 static int move_tasks(runqueue_t *this_rq, int this_cpu, runqueue_t *busiest,
1830 unsigned long max_nr_move, struct sched_domain *sd,
1831 enum idle_type idle, int *all_pinned)
1833 prio_array_t *array, *dst_array;
1834 struct list_head *head, *curr;
1835 int idx, pulled = 0, pinned = 0;
1838 if (max_nr_move == 0)
1844 * We first consider expired tasks. Those will likely not be
1845 * executed in the near future, and they are most likely to
1846 * be cache-cold, thus switching CPUs has the least effect
1849 if (busiest->expired->nr_active) {
1850 array = busiest->expired;
1851 dst_array = this_rq->expired;
1853 array = busiest->active;
1854 dst_array = this_rq->active;
1858 /* Start searching at priority 0: */
1862 idx = sched_find_first_bit(array->bitmap);
1864 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
1865 if (idx >= MAX_PRIO) {
1866 if (array == busiest->expired && busiest->active->nr_active) {
1867 array = busiest->active;
1868 dst_array = this_rq->active;
1874 head = array->queue + idx;
1877 tmp = list_entry(curr, task_t, run_list);
1881 if (!can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
1888 #ifdef CONFIG_SCHEDSTATS
1889 if (task_hot(tmp, busiest->timestamp_last_tick, sd))
1890 schedstat_inc(sd, lb_hot_gained[idle]);
1893 pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
1896 /* We only want to steal up to the prescribed number of tasks. */
1897 if (pulled < max_nr_move) {
1905 * Right now, this is the only place pull_task() is called,
1906 * so we can safely collect pull_task() stats here rather than
1907 * inside pull_task().
1909 schedstat_add(sd, lb_gained[idle], pulled);
1912 *all_pinned = pinned;
1917 * find_busiest_group finds and returns the busiest CPU group within the
1918 * domain. It calculates and returns the number of tasks which should be
1919 * moved to restore balance via the imbalance parameter.
1921 static struct sched_group *
1922 find_busiest_group(struct sched_domain *sd, int this_cpu,
1923 unsigned long *imbalance, enum idle_type idle)
1925 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
1926 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
1929 max_load = this_load = total_load = total_pwr = 0;
1930 if (idle == NOT_IDLE)
1931 load_idx = sd->busy_idx;
1932 else if (idle == NEWLY_IDLE)
1933 load_idx = sd->newidle_idx;
1935 load_idx = sd->idle_idx;
1942 local_group = cpu_isset(this_cpu, group->cpumask);
1944 /* Tally up the load of all CPUs in the group */
1947 for_each_cpu_mask(i, group->cpumask) {
1948 /* Bias balancing toward cpus of our domain */
1950 load = target_load(i, load_idx);
1952 load = source_load(i, load_idx);
1957 total_load += avg_load;
1958 total_pwr += group->cpu_power;
1960 /* Adjust by relative CPU power of the group */
1961 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1964 this_load = avg_load;
1966 } else if (avg_load > max_load) {
1967 max_load = avg_load;
1970 group = group->next;
1971 } while (group != sd->groups);
1973 if (!busiest || this_load >= max_load)
1976 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
1978 if (this_load >= avg_load ||
1979 100*max_load <= sd->imbalance_pct*this_load)
1983 * We're trying to get all the cpus to the average_load, so we don't
1984 * want to push ourselves above the average load, nor do we wish to
1985 * reduce the max loaded cpu below the average load, as either of these
1986 * actions would just result in more rebalancing later, and ping-pong
1987 * tasks around. Thus we look for the minimum possible imbalance.
1988 * Negative imbalances (*we* are more loaded than anyone else) will
1989 * be counted as no imbalance for these purposes -- we can't fix that
1990 * by pulling tasks to us. Be careful of negative numbers as they'll
1991 * appear as very large values with unsigned longs.
1993 /* How much load to actually move to equalise the imbalance */
1994 *imbalance = min((max_load - avg_load) * busiest->cpu_power,
1995 (avg_load - this_load) * this->cpu_power)
1998 if (*imbalance < SCHED_LOAD_SCALE) {
1999 unsigned long pwr_now = 0, pwr_move = 0;
2002 if (max_load - this_load >= SCHED_LOAD_SCALE*2) {
2008 * OK, we don't have enough imbalance to justify moving tasks,
2009 * however we may be able to increase total CPU power used by
2013 pwr_now += busiest->cpu_power*min(SCHED_LOAD_SCALE, max_load);
2014 pwr_now += this->cpu_power*min(SCHED_LOAD_SCALE, this_load);
2015 pwr_now /= SCHED_LOAD_SCALE;
2017 /* Amount of load we'd subtract */
2018 tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/busiest->cpu_power;
2020 pwr_move += busiest->cpu_power*min(SCHED_LOAD_SCALE,
2023 /* Amount of load we'd add */
2024 if (max_load*busiest->cpu_power <
2025 SCHED_LOAD_SCALE*SCHED_LOAD_SCALE)
2026 tmp = max_load*busiest->cpu_power/this->cpu_power;
2028 tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/this->cpu_power;
2029 pwr_move += this->cpu_power*min(SCHED_LOAD_SCALE, this_load + tmp);
2030 pwr_move /= SCHED_LOAD_SCALE;
2032 /* Move if we gain throughput */
2033 if (pwr_move <= pwr_now)
2040 /* Get rid of the scaling factor, rounding down as we divide */
2041 *imbalance = *imbalance / SCHED_LOAD_SCALE;
2051 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2053 static runqueue_t *find_busiest_queue(struct sched_group *group)
2055 unsigned long load, max_load = 0;
2056 runqueue_t *busiest = NULL;
2059 for_each_cpu_mask(i, group->cpumask) {
2060 load = source_load(i, 0);
2062 if (load > max_load) {
2064 busiest = cpu_rq(i);
2072 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2073 * tasks if there is an imbalance.
2075 * Called with this_rq unlocked.
2077 static int load_balance(int this_cpu, runqueue_t *this_rq,
2078 struct sched_domain *sd, enum idle_type idle)
2080 struct sched_group *group;
2081 runqueue_t *busiest;
2082 unsigned long imbalance;
2083 int nr_moved, all_pinned;
2084 int active_balance = 0;
2086 spin_lock(&this_rq->lock);
2087 schedstat_inc(sd, lb_cnt[idle]);
2089 group = find_busiest_group(sd, this_cpu, &imbalance, idle);
2091 schedstat_inc(sd, lb_nobusyg[idle]);
2095 busiest = find_busiest_queue(group);
2097 schedstat_inc(sd, lb_nobusyq[idle]);
2101 BUG_ON(busiest == this_rq);
2103 schedstat_add(sd, lb_imbalance[idle], imbalance);
2106 if (busiest->nr_running > 1) {
2108 * Attempt to move tasks. If find_busiest_group has found
2109 * an imbalance but busiest->nr_running <= 1, the group is
2110 * still unbalanced. nr_moved simply stays zero, so it is
2111 * correctly treated as an imbalance.
2113 double_lock_balance(this_rq, busiest);
2114 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2115 imbalance, sd, idle,
2117 spin_unlock(&busiest->lock);
2119 /* All tasks on this runqueue were pinned by CPU affinity */
2120 if (unlikely(all_pinned))
2124 spin_unlock(&this_rq->lock);
2127 schedstat_inc(sd, lb_failed[idle]);
2128 sd->nr_balance_failed++;
2130 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2132 spin_lock(&busiest->lock);
2133 if (!busiest->active_balance) {
2134 busiest->active_balance = 1;
2135 busiest->push_cpu = this_cpu;
2138 spin_unlock(&busiest->lock);
2140 wake_up_process(busiest->migration_thread);
2143 * We've kicked active balancing, reset the failure
2146 sd->nr_balance_failed = sd->cache_nice_tries+1;
2149 sd->nr_balance_failed = 0;
2151 if (likely(!active_balance)) {
2152 /* We were unbalanced, so reset the balancing interval */
2153 sd->balance_interval = sd->min_interval;
2156 * If we've begun active balancing, start to back off. This
2157 * case may not be covered by the all_pinned logic if there
2158 * is only 1 task on the busy runqueue (because we don't call
2161 if (sd->balance_interval < sd->max_interval)
2162 sd->balance_interval *= 2;
2168 spin_unlock(&this_rq->lock);
2170 schedstat_inc(sd, lb_balanced[idle]);
2172 sd->nr_balance_failed = 0;
2173 /* tune up the balancing interval */
2174 if (sd->balance_interval < sd->max_interval)
2175 sd->balance_interval *= 2;
2181 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2182 * tasks if there is an imbalance.
2184 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2185 * this_rq is locked.
2187 static int load_balance_newidle(int this_cpu, runqueue_t *this_rq,
2188 struct sched_domain *sd)
2190 struct sched_group *group;
2191 runqueue_t *busiest = NULL;
2192 unsigned long imbalance;
2195 schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
2196 group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE);
2198 schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]);
2202 busiest = find_busiest_queue(group);
2204 schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]);
2208 BUG_ON(busiest == this_rq);
2210 /* Attempt to move tasks */
2211 double_lock_balance(this_rq, busiest);
2213 schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
2214 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2215 imbalance, sd, NEWLY_IDLE, NULL);
2217 schedstat_inc(sd, lb_failed[NEWLY_IDLE]);
2219 sd->nr_balance_failed = 0;
2221 spin_unlock(&busiest->lock);
2225 schedstat_inc(sd, lb_balanced[NEWLY_IDLE]);
2226 sd->nr_balance_failed = 0;
2231 * idle_balance is called by schedule() if this_cpu is about to become
2232 * idle. Attempts to pull tasks from other CPUs.
2234 static inline void idle_balance(int this_cpu, runqueue_t *this_rq)
2236 struct sched_domain *sd;
2238 for_each_domain(this_cpu, sd) {
2239 if (sd->flags & SD_BALANCE_NEWIDLE) {
2240 if (load_balance_newidle(this_cpu, this_rq, sd)) {
2241 /* We've pulled tasks over so stop searching */
2249 * active_load_balance is run by migration threads. It pushes running tasks
2250 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2251 * running on each physical CPU where possible, and avoids physical /
2252 * logical imbalances.
2254 * Called with busiest_rq locked.
2256 static void active_load_balance(runqueue_t *busiest_rq, int busiest_cpu)
2258 struct sched_domain *sd;
2259 runqueue_t *target_rq;
2260 int target_cpu = busiest_rq->push_cpu;
2262 if (busiest_rq->nr_running <= 1)
2263 /* no task to move */
2266 target_rq = cpu_rq(target_cpu);
2269 * This condition is "impossible", if it occurs
2270 * we need to fix it. Originally reported by
2271 * Bjorn Helgaas on a 128-cpu setup.
2273 BUG_ON(busiest_rq == target_rq);
2275 /* move a task from busiest_rq to target_rq */
2276 double_lock_balance(busiest_rq, target_rq);
2278 /* Search for an sd spanning us and the target CPU. */
2279 for_each_domain(target_cpu, sd)
2280 if ((sd->flags & SD_LOAD_BALANCE) &&
2281 cpu_isset(busiest_cpu, sd->span))
2284 if (unlikely(sd == NULL))
2287 schedstat_inc(sd, alb_cnt);
2289 if (move_tasks(target_rq, target_cpu, busiest_rq, 1, sd, SCHED_IDLE, NULL))
2290 schedstat_inc(sd, alb_pushed);
2292 schedstat_inc(sd, alb_failed);
2294 spin_unlock(&target_rq->lock);
2298 * rebalance_tick will get called every timer tick, on every CPU.
2300 * It checks each scheduling domain to see if it is due to be balanced,
2301 * and initiates a balancing operation if so.
2303 * Balancing parameters are set up in arch_init_sched_domains.
2306 /* Don't have all balancing operations going off at once */
2307 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2309 static void rebalance_tick(int this_cpu, runqueue_t *this_rq,
2310 enum idle_type idle)
2312 unsigned long old_load, this_load;
2313 unsigned long j = jiffies + CPU_OFFSET(this_cpu);
2314 struct sched_domain *sd;
2317 this_load = this_rq->nr_running * SCHED_LOAD_SCALE;
2318 /* Update our load */
2319 for (i = 0; i < 3; i++) {
2320 unsigned long new_load = this_load;
2322 old_load = this_rq->cpu_load[i];
2324 * Round up the averaging division if load is increasing. This
2325 * prevents us from getting stuck on 9 if the load is 10, for
2328 if (new_load > old_load)
2329 new_load += scale-1;
2330 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale;
2333 for_each_domain(this_cpu, sd) {
2334 unsigned long interval;
2336 if (!(sd->flags & SD_LOAD_BALANCE))
2339 interval = sd->balance_interval;
2340 if (idle != SCHED_IDLE)
2341 interval *= sd->busy_factor;
2343 /* scale ms to jiffies */
2344 interval = msecs_to_jiffies(interval);
2345 if (unlikely(!interval))
2348 if (j - sd->last_balance >= interval) {
2349 if (load_balance(this_cpu, this_rq, sd, idle)) {
2350 /* We've pulled tasks over so no longer idle */
2353 sd->last_balance += interval;
2359 * on UP we do not need to balance between CPUs:
2361 static inline void rebalance_tick(int cpu, runqueue_t *rq, enum idle_type idle)
2364 static inline void idle_balance(int cpu, runqueue_t *rq)
2369 static inline int wake_priority_sleeper(runqueue_t *rq)
2372 #ifdef CONFIG_SCHED_SMT
2373 spin_lock(&rq->lock);
2375 * If an SMT sibling task has been put to sleep for priority
2376 * reasons reschedule the idle task to see if it can now run.
2378 if (rq->nr_running) {
2379 resched_task(rq->idle);
2382 spin_unlock(&rq->lock);
2387 DEFINE_PER_CPU(struct kernel_stat, kstat);
2389 EXPORT_PER_CPU_SYMBOL(kstat);
2392 * This is called on clock ticks and on context switches.
2393 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2395 static inline void update_cpu_clock(task_t *p, runqueue_t *rq,
2396 unsigned long long now)
2398 unsigned long long last = max(p->timestamp, rq->timestamp_last_tick);
2399 p->sched_time += now - last;
2403 * Return current->sched_time plus any more ns on the sched_clock
2404 * that have not yet been banked.
2406 unsigned long long current_sched_time(const task_t *tsk)
2408 unsigned long long ns;
2409 unsigned long flags;
2410 local_irq_save(flags);
2411 ns = max(tsk->timestamp, task_rq(tsk)->timestamp_last_tick);
2412 ns = tsk->sched_time + (sched_clock() - ns);
2413 local_irq_restore(flags);
2418 * We place interactive tasks back into the active array, if possible.
2420 * To guarantee that this does not starve expired tasks we ignore the
2421 * interactivity of a task if the first expired task had to wait more
2422 * than a 'reasonable' amount of time. This deadline timeout is
2423 * load-dependent, as the frequency of array switched decreases with
2424 * increasing number of running tasks. We also ignore the interactivity
2425 * if a better static_prio task has expired:
2427 #define EXPIRED_STARVING(rq) \
2428 ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2429 (jiffies - (rq)->expired_timestamp >= \
2430 STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2431 ((rq)->curr->static_prio > (rq)->best_expired_prio))
2434 * Account user cpu time to a process.
2435 * @p: the process that the cpu time gets accounted to
2436 * @hardirq_offset: the offset to subtract from hardirq_count()
2437 * @cputime: the cpu time spent in user space since the last update
2439 void account_user_time(struct task_struct *p, cputime_t cputime)
2441 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2444 p->utime = cputime_add(p->utime, cputime);
2446 /* Add user time to cpustat. */
2447 tmp = cputime_to_cputime64(cputime);
2448 if (TASK_NICE(p) > 0)
2449 cpustat->nice = cputime64_add(cpustat->nice, tmp);
2451 cpustat->user = cputime64_add(cpustat->user, tmp);
2455 * Account system cpu time to a process.
2456 * @p: the process that the cpu time gets accounted to
2457 * @hardirq_offset: the offset to subtract from hardirq_count()
2458 * @cputime: the cpu time spent in kernel space since the last update
2460 void account_system_time(struct task_struct *p, int hardirq_offset,
2463 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2464 runqueue_t *rq = this_rq();
2467 p->stime = cputime_add(p->stime, cputime);
2469 /* Add system time to cpustat. */
2470 tmp = cputime_to_cputime64(cputime);
2471 if (hardirq_count() - hardirq_offset)
2472 cpustat->irq = cputime64_add(cpustat->irq, tmp);
2473 else if (softirq_count())
2474 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
2475 else if (p != rq->idle)
2476 cpustat->system = cputime64_add(cpustat->system, tmp);
2477 else if (atomic_read(&rq->nr_iowait) > 0)
2478 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2480 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2481 /* Account for system time used */
2482 acct_update_integrals(p);
2483 /* Update rss highwater mark */
2484 update_mem_hiwater(p);
2488 * Account for involuntary wait time.
2489 * @p: the process from which the cpu time has been stolen
2490 * @steal: the cpu time spent in involuntary wait
2492 void account_steal_time(struct task_struct *p, cputime_t steal)
2494 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2495 cputime64_t tmp = cputime_to_cputime64(steal);
2496 runqueue_t *rq = this_rq();
2498 if (p == rq->idle) {
2499 p->stime = cputime_add(p->stime, steal);
2500 if (atomic_read(&rq->nr_iowait) > 0)
2501 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2503 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2505 cpustat->steal = cputime64_add(cpustat->steal, tmp);
2509 * This function gets called by the timer code, with HZ frequency.
2510 * We call it with interrupts disabled.
2512 * It also gets called by the fork code, when changing the parent's
2515 void scheduler_tick(void)
2517 int cpu = smp_processor_id();
2518 runqueue_t *rq = this_rq();
2519 task_t *p = current;
2520 unsigned long long now = sched_clock();
2522 update_cpu_clock(p, rq, now);
2524 rq->timestamp_last_tick = now;
2526 if (p == rq->idle) {
2527 if (wake_priority_sleeper(rq))
2529 rebalance_tick(cpu, rq, SCHED_IDLE);
2533 /* Task might have expired already, but not scheduled off yet */
2534 if (p->array != rq->active) {
2535 set_tsk_need_resched(p);
2538 spin_lock(&rq->lock);
2540 * The task was running during this tick - update the
2541 * time slice counter. Note: we do not update a thread's
2542 * priority until it either goes to sleep or uses up its
2543 * timeslice. This makes it possible for interactive tasks
2544 * to use up their timeslices at their highest priority levels.
2548 * RR tasks need a special form of timeslice management.
2549 * FIFO tasks have no timeslices.
2551 if ((p->policy == SCHED_RR) && !--p->time_slice) {
2552 p->time_slice = task_timeslice(p);
2553 p->first_time_slice = 0;
2554 set_tsk_need_resched(p);
2556 /* put it at the end of the queue: */
2557 requeue_task(p, rq->active);
2561 if (!--p->time_slice) {
2562 dequeue_task(p, rq->active);
2563 set_tsk_need_resched(p);
2564 p->prio = effective_prio(p);
2565 p->time_slice = task_timeslice(p);
2566 p->first_time_slice = 0;
2568 if (!rq->expired_timestamp)
2569 rq->expired_timestamp = jiffies;
2570 if (!TASK_INTERACTIVE(p) || EXPIRED_STARVING(rq)) {
2571 enqueue_task(p, rq->expired);
2572 if (p->static_prio < rq->best_expired_prio)
2573 rq->best_expired_prio = p->static_prio;
2575 enqueue_task(p, rq->active);
2578 * Prevent a too long timeslice allowing a task to monopolize
2579 * the CPU. We do this by splitting up the timeslice into
2582 * Note: this does not mean the task's timeslices expire or
2583 * get lost in any way, they just might be preempted by
2584 * another task of equal priority. (one with higher
2585 * priority would have preempted this task already.) We
2586 * requeue this task to the end of the list on this priority
2587 * level, which is in essence a round-robin of tasks with
2590 * This only applies to tasks in the interactive
2591 * delta range with at least TIMESLICE_GRANULARITY to requeue.
2593 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
2594 p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
2595 (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
2596 (p->array == rq->active)) {
2598 requeue_task(p, rq->active);
2599 set_tsk_need_resched(p);
2603 spin_unlock(&rq->lock);
2605 rebalance_tick(cpu, rq, NOT_IDLE);
2608 #ifdef CONFIG_SCHED_SMT
2609 static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2611 struct sched_domain *tmp, *sd = NULL;
2612 cpumask_t sibling_map;
2615 for_each_domain(this_cpu, tmp)
2616 if (tmp->flags & SD_SHARE_CPUPOWER)
2623 * Unlock the current runqueue because we have to lock in
2624 * CPU order to avoid deadlocks. Caller knows that we might
2625 * unlock. We keep IRQs disabled.
2627 spin_unlock(&this_rq->lock);
2629 sibling_map = sd->span;
2631 for_each_cpu_mask(i, sibling_map)
2632 spin_lock(&cpu_rq(i)->lock);
2634 * We clear this CPU from the mask. This both simplifies the
2635 * inner loop and keps this_rq locked when we exit:
2637 cpu_clear(this_cpu, sibling_map);
2639 for_each_cpu_mask(i, sibling_map) {
2640 runqueue_t *smt_rq = cpu_rq(i);
2643 * If an SMT sibling task is sleeping due to priority
2644 * reasons wake it up now.
2646 if (smt_rq->curr == smt_rq->idle && smt_rq->nr_running)
2647 resched_task(smt_rq->idle);
2650 for_each_cpu_mask(i, sibling_map)
2651 spin_unlock(&cpu_rq(i)->lock);
2653 * We exit with this_cpu's rq still held and IRQs
2658 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2660 struct sched_domain *tmp, *sd = NULL;
2661 cpumask_t sibling_map;
2662 prio_array_t *array;
2666 for_each_domain(this_cpu, tmp)
2667 if (tmp->flags & SD_SHARE_CPUPOWER)
2674 * The same locking rules and details apply as for
2675 * wake_sleeping_dependent():
2677 spin_unlock(&this_rq->lock);
2678 sibling_map = sd->span;
2679 for_each_cpu_mask(i, sibling_map)
2680 spin_lock(&cpu_rq(i)->lock);
2681 cpu_clear(this_cpu, sibling_map);
2684 * Establish next task to be run - it might have gone away because
2685 * we released the runqueue lock above:
2687 if (!this_rq->nr_running)
2689 array = this_rq->active;
2690 if (!array->nr_active)
2691 array = this_rq->expired;
2692 BUG_ON(!array->nr_active);
2694 p = list_entry(array->queue[sched_find_first_bit(array->bitmap)].next,
2697 for_each_cpu_mask(i, sibling_map) {
2698 runqueue_t *smt_rq = cpu_rq(i);
2699 task_t *smt_curr = smt_rq->curr;
2702 * If a user task with lower static priority than the
2703 * running task on the SMT sibling is trying to schedule,
2704 * delay it till there is proportionately less timeslice
2705 * left of the sibling task to prevent a lower priority
2706 * task from using an unfair proportion of the
2707 * physical cpu's resources. -ck
2709 if (((smt_curr->time_slice * (100 - sd->per_cpu_gain) / 100) >
2710 task_timeslice(p) || rt_task(smt_curr)) &&
2711 p->mm && smt_curr->mm && !rt_task(p))
2715 * Reschedule a lower priority task on the SMT sibling,
2716 * or wake it up if it has been put to sleep for priority
2719 if ((((p->time_slice * (100 - sd->per_cpu_gain) / 100) >
2720 task_timeslice(smt_curr) || rt_task(p)) &&
2721 smt_curr->mm && p->mm && !rt_task(smt_curr)) ||
2722 (smt_curr == smt_rq->idle && smt_rq->nr_running))
2723 resched_task(smt_curr);
2726 for_each_cpu_mask(i, sibling_map)
2727 spin_unlock(&cpu_rq(i)->lock);
2731 static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2735 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2741 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
2743 void fastcall add_preempt_count(int val)
2748 BUG_ON((preempt_count() < 0));
2749 preempt_count() += val;
2751 * Spinlock count overflowing soon?
2753 BUG_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10);
2755 EXPORT_SYMBOL(add_preempt_count);
2757 void fastcall sub_preempt_count(int val)
2762 BUG_ON(val > preempt_count());
2764 * Is the spinlock portion underflowing?
2766 BUG_ON((val < PREEMPT_MASK) && !(preempt_count() & PREEMPT_MASK));
2767 preempt_count() -= val;
2769 EXPORT_SYMBOL(sub_preempt_count);
2774 * schedule() is the main scheduler function.
2776 asmlinkage void __sched schedule(void)
2779 task_t *prev, *next;
2781 prio_array_t *array;
2782 struct list_head *queue;
2783 unsigned long long now;
2784 unsigned long run_time;
2788 * Test if we are atomic. Since do_exit() needs to call into
2789 * schedule() atomically, we ignore that path for now.
2790 * Otherwise, whine if we are scheduling when we should not be.
2792 if (likely(!current->exit_state)) {
2793 if (unlikely(in_atomic())) {
2794 printk(KERN_ERR "scheduling while atomic: "
2796 current->comm, preempt_count(), current->pid);
2800 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2805 release_kernel_lock(prev);
2806 need_resched_nonpreemptible:
2810 * The idle thread is not allowed to schedule!
2811 * Remove this check after it has been exercised a bit.
2813 if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
2814 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
2818 schedstat_inc(rq, sched_cnt);
2819 now = sched_clock();
2820 if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
2821 run_time = now - prev->timestamp;
2822 if (unlikely((long long)(now - prev->timestamp) < 0))
2825 run_time = NS_MAX_SLEEP_AVG;
2828 * Tasks charged proportionately less run_time at high sleep_avg to
2829 * delay them losing their interactive status
2831 run_time /= (CURRENT_BONUS(prev) ? : 1);
2833 spin_lock_irq(&rq->lock);
2835 if (unlikely(prev->flags & PF_DEAD))
2836 prev->state = EXIT_DEAD;
2838 switch_count = &prev->nivcsw;
2839 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2840 switch_count = &prev->nvcsw;
2841 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
2842 unlikely(signal_pending(prev))))
2843 prev->state = TASK_RUNNING;
2845 if (prev->state == TASK_UNINTERRUPTIBLE)
2846 rq->nr_uninterruptible++;
2847 deactivate_task(prev, rq);
2851 cpu = smp_processor_id();
2852 if (unlikely(!rq->nr_running)) {
2854 idle_balance(cpu, rq);
2855 if (!rq->nr_running) {
2857 rq->expired_timestamp = 0;
2858 wake_sleeping_dependent(cpu, rq);
2860 * wake_sleeping_dependent() might have released
2861 * the runqueue, so break out if we got new
2864 if (!rq->nr_running)
2868 if (dependent_sleeper(cpu, rq)) {
2873 * dependent_sleeper() releases and reacquires the runqueue
2874 * lock, hence go into the idle loop if the rq went
2877 if (unlikely(!rq->nr_running))
2882 if (unlikely(!array->nr_active)) {
2884 * Switch the active and expired arrays.
2886 schedstat_inc(rq, sched_switch);
2887 rq->active = rq->expired;
2888 rq->expired = array;
2890 rq->expired_timestamp = 0;
2891 rq->best_expired_prio = MAX_PRIO;
2894 idx = sched_find_first_bit(array->bitmap);
2895 queue = array->queue + idx;
2896 next = list_entry(queue->next, task_t, run_list);
2898 if (!rt_task(next) && next->activated > 0) {
2899 unsigned long long delta = now - next->timestamp;
2900 if (unlikely((long long)(now - next->timestamp) < 0))
2903 if (next->activated == 1)
2904 delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
2906 array = next->array;
2907 dequeue_task(next, array);
2908 recalc_task_prio(next, next->timestamp + delta);
2909 enqueue_task(next, array);
2911 next->activated = 0;
2913 if (next == rq->idle)
2914 schedstat_inc(rq, sched_goidle);
2916 clear_tsk_need_resched(prev);
2917 rcu_qsctr_inc(task_cpu(prev));
2919 update_cpu_clock(prev, rq, now);
2921 prev->sleep_avg -= run_time;
2922 if ((long)prev->sleep_avg <= 0)
2923 prev->sleep_avg = 0;
2924 prev->timestamp = prev->last_ran = now;
2926 sched_info_switch(prev, next);
2927 if (likely(prev != next)) {
2928 next->timestamp = now;
2933 prepare_task_switch(rq, next);
2934 prev = context_switch(rq, prev, next);
2937 * this_rq must be evaluated again because prev may have moved
2938 * CPUs since it called schedule(), thus the 'rq' on its stack
2939 * frame will be invalid.
2941 finish_task_switch(this_rq(), prev);
2943 spin_unlock_irq(&rq->lock);
2946 if (unlikely(reacquire_kernel_lock(prev) < 0))
2947 goto need_resched_nonpreemptible;
2948 preempt_enable_no_resched();
2949 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
2953 EXPORT_SYMBOL(schedule);
2955 #ifdef CONFIG_PREEMPT
2957 * this is is the entry point to schedule() from in-kernel preemption
2958 * off of preempt_enable. Kernel preemptions off return from interrupt
2959 * occur there and call schedule directly.
2961 asmlinkage void __sched preempt_schedule(void)
2963 struct thread_info *ti = current_thread_info();
2964 #ifdef CONFIG_PREEMPT_BKL
2965 struct task_struct *task = current;
2966 int saved_lock_depth;
2969 * If there is a non-zero preempt_count or interrupts are disabled,
2970 * we do not want to preempt the current task. Just return..
2972 if (unlikely(ti->preempt_count || irqs_disabled()))
2976 add_preempt_count(PREEMPT_ACTIVE);
2978 * We keep the big kernel semaphore locked, but we
2979 * clear ->lock_depth so that schedule() doesnt
2980 * auto-release the semaphore:
2982 #ifdef CONFIG_PREEMPT_BKL
2983 saved_lock_depth = task->lock_depth;
2984 task->lock_depth = -1;
2987 #ifdef CONFIG_PREEMPT_BKL
2988 task->lock_depth = saved_lock_depth;
2990 sub_preempt_count(PREEMPT_ACTIVE);
2992 /* we could miss a preemption opportunity between schedule and now */
2994 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
2998 EXPORT_SYMBOL(preempt_schedule);
3001 * this is is the entry point to schedule() from kernel preemption
3002 * off of irq context.
3003 * Note, that this is called and return with irqs disabled. This will
3004 * protect us against recursive calling from irq.
3006 asmlinkage void __sched preempt_schedule_irq(void)
3008 struct thread_info *ti = current_thread_info();
3009 #ifdef CONFIG_PREEMPT_BKL
3010 struct task_struct *task = current;
3011 int saved_lock_depth;
3013 /* Catch callers which need to be fixed*/
3014 BUG_ON(ti->preempt_count || !irqs_disabled());
3017 add_preempt_count(PREEMPT_ACTIVE);
3019 * We keep the big kernel semaphore locked, but we
3020 * clear ->lock_depth so that schedule() doesnt
3021 * auto-release the semaphore:
3023 #ifdef CONFIG_PREEMPT_BKL
3024 saved_lock_depth = task->lock_depth;
3025 task->lock_depth = -1;
3029 local_irq_disable();
3030 #ifdef CONFIG_PREEMPT_BKL
3031 task->lock_depth = saved_lock_depth;
3033 sub_preempt_count(PREEMPT_ACTIVE);
3035 /* we could miss a preemption opportunity between schedule and now */
3037 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3041 #endif /* CONFIG_PREEMPT */
3043 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync, void *key)
3045 task_t *p = curr->private;
3046 return try_to_wake_up(p, mode, sync);
3049 EXPORT_SYMBOL(default_wake_function);
3052 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3053 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3054 * number) then we wake all the non-exclusive tasks and one exclusive task.
3056 * There are circumstances in which we can try to wake a task which has already
3057 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3058 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3060 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3061 int nr_exclusive, int sync, void *key)
3063 struct list_head *tmp, *next;
3065 list_for_each_safe(tmp, next, &q->task_list) {
3068 curr = list_entry(tmp, wait_queue_t, task_list);
3069 flags = curr->flags;
3070 if (curr->func(curr, mode, sync, key) &&
3071 (flags & WQ_FLAG_EXCLUSIVE) &&
3078 * __wake_up - wake up threads blocked on a waitqueue.
3080 * @mode: which threads
3081 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3082 * @key: is directly passed to the wakeup function
3084 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3085 int nr_exclusive, void *key)
3087 unsigned long flags;
3089 spin_lock_irqsave(&q->lock, flags);
3090 __wake_up_common(q, mode, nr_exclusive, 0, key);
3091 spin_unlock_irqrestore(&q->lock, flags);
3094 EXPORT_SYMBOL(__wake_up);
3097 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3099 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3101 __wake_up_common(q, mode, 1, 0, NULL);
3105 * __wake_up_sync - wake up threads blocked on a waitqueue.
3107 * @mode: which threads
3108 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3110 * The sync wakeup differs that the waker knows that it will schedule
3111 * away soon, so while the target thread will be woken up, it will not
3112 * be migrated to another CPU - ie. the two threads are 'synchronized'
3113 * with each other. This can prevent needless bouncing between CPUs.
3115 * On UP it can prevent extra preemption.
3117 void fastcall __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3119 unsigned long flags;
3125 if (unlikely(!nr_exclusive))
3128 spin_lock_irqsave(&q->lock, flags);
3129 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3130 spin_unlock_irqrestore(&q->lock, flags);
3132 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3134 void fastcall complete(struct completion *x)
3136 unsigned long flags;
3138 spin_lock_irqsave(&x->wait.lock, flags);
3140 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3142 spin_unlock_irqrestore(&x->wait.lock, flags);
3144 EXPORT_SYMBOL(complete);
3146 void fastcall complete_all(struct completion *x)
3148 unsigned long flags;
3150 spin_lock_irqsave(&x->wait.lock, flags);
3151 x->done += UINT_MAX/2;
3152 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3154 spin_unlock_irqrestore(&x->wait.lock, flags);
3156 EXPORT_SYMBOL(complete_all);
3158 void fastcall __sched wait_for_completion(struct completion *x)
3161 spin_lock_irq(&x->wait.lock);
3163 DECLARE_WAITQUEUE(wait, current);
3165 wait.flags |= WQ_FLAG_EXCLUSIVE;
3166 __add_wait_queue_tail(&x->wait, &wait);
3168 __set_current_state(TASK_UNINTERRUPTIBLE);
3169 spin_unlock_irq(&x->wait.lock);
3171 spin_lock_irq(&x->wait.lock);
3173 __remove_wait_queue(&x->wait, &wait);
3176 spin_unlock_irq(&x->wait.lock);
3178 EXPORT_SYMBOL(wait_for_completion);
3180 unsigned long fastcall __sched
3181 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3185 spin_lock_irq(&x->wait.lock);
3187 DECLARE_WAITQUEUE(wait, current);
3189 wait.flags |= WQ_FLAG_EXCLUSIVE;
3190 __add_wait_queue_tail(&x->wait, &wait);
3192 __set_current_state(TASK_UNINTERRUPTIBLE);
3193 spin_unlock_irq(&x->wait.lock);
3194 timeout = schedule_timeout(timeout);
3195 spin_lock_irq(&x->wait.lock);
3197 __remove_wait_queue(&x->wait, &wait);
3201 __remove_wait_queue(&x->wait, &wait);
3205 spin_unlock_irq(&x->wait.lock);
3208 EXPORT_SYMBOL(wait_for_completion_timeout);
3210 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3216 spin_lock_irq(&x->wait.lock);
3218 DECLARE_WAITQUEUE(wait, current);
3220 wait.flags |= WQ_FLAG_EXCLUSIVE;
3221 __add_wait_queue_tail(&x->wait, &wait);
3223 if (signal_pending(current)) {
3225 __remove_wait_queue(&x->wait, &wait);
3228 __set_current_state(TASK_INTERRUPTIBLE);
3229 spin_unlock_irq(&x->wait.lock);
3231 spin_lock_irq(&x->wait.lock);
3233 __remove_wait_queue(&x->wait, &wait);
3237 spin_unlock_irq(&x->wait.lock);
3241 EXPORT_SYMBOL(wait_for_completion_interruptible);
3243 unsigned long fastcall __sched
3244 wait_for_completion_interruptible_timeout(struct completion *x,
3245 unsigned long timeout)
3249 spin_lock_irq(&x->wait.lock);
3251 DECLARE_WAITQUEUE(wait, current);
3253 wait.flags |= WQ_FLAG_EXCLUSIVE;
3254 __add_wait_queue_tail(&x->wait, &wait);
3256 if (signal_pending(current)) {
3257 timeout = -ERESTARTSYS;
3258 __remove_wait_queue(&x->wait, &wait);
3261 __set_current_state(TASK_INTERRUPTIBLE);
3262 spin_unlock_irq(&x->wait.lock);
3263 timeout = schedule_timeout(timeout);
3264 spin_lock_irq(&x->wait.lock);
3266 __remove_wait_queue(&x->wait, &wait);
3270 __remove_wait_queue(&x->wait, &wait);
3274 spin_unlock_irq(&x->wait.lock);
3277 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3280 #define SLEEP_ON_VAR \
3281 unsigned long flags; \
3282 wait_queue_t wait; \
3283 init_waitqueue_entry(&wait, current);
3285 #define SLEEP_ON_HEAD \
3286 spin_lock_irqsave(&q->lock,flags); \
3287 __add_wait_queue(q, &wait); \
3288 spin_unlock(&q->lock);
3290 #define SLEEP_ON_TAIL \
3291 spin_lock_irq(&q->lock); \
3292 __remove_wait_queue(q, &wait); \
3293 spin_unlock_irqrestore(&q->lock, flags);
3295 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
3299 current->state = TASK_INTERRUPTIBLE;
3306 EXPORT_SYMBOL(interruptible_sleep_on);
3308 long fastcall __sched interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3312 current->state = TASK_INTERRUPTIBLE;
3315 timeout = schedule_timeout(timeout);
3321 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3323 void fastcall __sched sleep_on(wait_queue_head_t *q)
3327 current->state = TASK_UNINTERRUPTIBLE;
3334 EXPORT_SYMBOL(sleep_on);
3336 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3340 current->state = TASK_UNINTERRUPTIBLE;
3343 timeout = schedule_timeout(timeout);
3349 EXPORT_SYMBOL(sleep_on_timeout);
3351 void set_user_nice(task_t *p, long nice)
3353 unsigned long flags;
3354 prio_array_t *array;
3356 int old_prio, new_prio, delta;
3358 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3361 * We have to be careful, if called from sys_setpriority(),
3362 * the task might be in the middle of scheduling on another CPU.
3364 rq = task_rq_lock(p, &flags);
3366 * The RT priorities are set via sched_setscheduler(), but we still
3367 * allow the 'normal' nice value to be set - but as expected
3368 * it wont have any effect on scheduling until the task is
3372 p->static_prio = NICE_TO_PRIO(nice);
3377 dequeue_task(p, array);
3380 new_prio = NICE_TO_PRIO(nice);
3381 delta = new_prio - old_prio;
3382 p->static_prio = NICE_TO_PRIO(nice);
3386 enqueue_task(p, array);
3388 * If the task increased its priority or is running and
3389 * lowered its priority, then reschedule its CPU:
3391 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3392 resched_task(rq->curr);
3395 task_rq_unlock(rq, &flags);
3398 EXPORT_SYMBOL(set_user_nice);
3401 * can_nice - check if a task can reduce its nice value
3405 int can_nice(const task_t *p, const int nice)
3407 /* convert nice value [19,-20] to rlimit style value [0,39] */
3408 int nice_rlim = 19 - nice;
3409 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
3410 capable(CAP_SYS_NICE));
3413 #ifdef __ARCH_WANT_SYS_NICE
3416 * sys_nice - change the priority of the current process.
3417 * @increment: priority increment
3419 * sys_setpriority is a more generic, but much slower function that
3420 * does similar things.
3422 asmlinkage long sys_nice(int increment)
3428 * Setpriority might change our priority at the same moment.
3429 * We don't have to worry. Conceptually one call occurs first
3430 * and we have a single winner.
3432 if (increment < -40)
3437 nice = PRIO_TO_NICE(current->static_prio) + increment;
3443 if (increment < 0 && !can_nice(current, nice))
3446 retval = security_task_setnice(current, nice);
3450 set_user_nice(current, nice);
3457 * task_prio - return the priority value of a given task.
3458 * @p: the task in question.
3460 * This is the priority value as seen by users in /proc.
3461 * RT tasks are offset by -200. Normal tasks are centered
3462 * around 0, value goes from -16 to +15.
3464 int task_prio(const task_t *p)
3466 return p->prio - MAX_RT_PRIO;
3470 * task_nice - return the nice value of a given task.
3471 * @p: the task in question.
3473 int task_nice(const task_t *p)
3475 return TASK_NICE(p);
3479 * The only users of task_nice are binfmt_elf and binfmt_elf32.
3480 * binfmt_elf is no longer modular, but binfmt_elf32 still is.
3481 * Therefore, task_nice is needed if there is a compat_mode.
3483 #ifdef CONFIG_COMPAT
3484 EXPORT_SYMBOL_GPL(task_nice);
3488 * idle_cpu - is a given cpu idle currently?
3489 * @cpu: the processor in question.
3491 int idle_cpu(int cpu)
3493 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
3496 EXPORT_SYMBOL_GPL(idle_cpu);
3499 * idle_task - return the idle task for a given cpu.
3500 * @cpu: the processor in question.
3502 task_t *idle_task(int cpu)
3504 return cpu_rq(cpu)->idle;
3508 * find_process_by_pid - find a process with a matching PID value.
3509 * @pid: the pid in question.
3511 static inline task_t *find_process_by_pid(pid_t pid)
3513 return pid ? find_task_by_pid(pid) : current;
3516 /* Actually do priority change: must hold rq lock. */
3517 static void __setscheduler(struct task_struct *p, int policy, int prio)
3521 p->rt_priority = prio;
3522 if (policy != SCHED_NORMAL)
3523 p->prio = MAX_USER_RT_PRIO-1 - p->rt_priority;
3525 p->prio = p->static_prio;
3529 * sched_setscheduler - change the scheduling policy and/or RT priority of
3531 * @p: the task in question.
3532 * @policy: new policy.
3533 * @param: structure containing the new RT priority.
3535 int sched_setscheduler(struct task_struct *p, int policy, struct sched_param *param)
3538 int oldprio, oldpolicy = -1;
3539 prio_array_t *array;
3540 unsigned long flags;
3544 /* double check policy once rq lock held */
3546 policy = oldpolicy = p->policy;
3547 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
3548 policy != SCHED_NORMAL)
3551 * Valid priorities for SCHED_FIFO and SCHED_RR are
3552 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL is 0.
3554 if (param->sched_priority < 0 ||
3555 param->sched_priority > MAX_USER_RT_PRIO-1)
3557 if ((policy == SCHED_NORMAL) != (param->sched_priority == 0))
3560 if ((policy == SCHED_FIFO || policy == SCHED_RR) &&
3561 param->sched_priority > p->signal->rlim[RLIMIT_RTPRIO].rlim_cur &&
3562 !capable(CAP_SYS_NICE))
3564 if ((current->euid != p->euid) && (current->euid != p->uid) &&
3565 !capable(CAP_SYS_NICE))
3568 retval = security_task_setscheduler(p, policy, param);
3572 * To be able to change p->policy safely, the apropriate
3573 * runqueue lock must be held.
3575 rq = task_rq_lock(p, &flags);
3576 /* recheck policy now with rq lock held */
3577 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3578 policy = oldpolicy = -1;
3579 task_rq_unlock(rq, &flags);
3584 deactivate_task(p, rq);
3586 __setscheduler(p, policy, param->sched_priority);
3588 __activate_task(p, rq);
3590 * Reschedule if we are currently running on this runqueue and
3591 * our priority decreased, or if we are not currently running on
3592 * this runqueue and our priority is higher than the current's
3594 if (task_running(rq, p)) {
3595 if (p->prio > oldprio)
3596 resched_task(rq->curr);
3597 } else if (TASK_PREEMPTS_CURR(p, rq))
3598 resched_task(rq->curr);
3600 task_rq_unlock(rq, &flags);
3603 EXPORT_SYMBOL_GPL(sched_setscheduler);
3605 static int do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3608 struct sched_param lparam;
3609 struct task_struct *p;
3611 if (!param || pid < 0)
3613 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3615 read_lock_irq(&tasklist_lock);
3616 p = find_process_by_pid(pid);
3618 read_unlock_irq(&tasklist_lock);
3621 retval = sched_setscheduler(p, policy, &lparam);
3622 read_unlock_irq(&tasklist_lock);
3627 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3628 * @pid: the pid in question.
3629 * @policy: new policy.
3630 * @param: structure containing the new RT priority.
3632 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
3633 struct sched_param __user *param)
3635 return do_sched_setscheduler(pid, policy, param);
3639 * sys_sched_setparam - set/change the RT priority of a thread
3640 * @pid: the pid in question.
3641 * @param: structure containing the new RT priority.
3643 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
3645 return do_sched_setscheduler(pid, -1, param);
3649 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3650 * @pid: the pid in question.
3652 asmlinkage long sys_sched_getscheduler(pid_t pid)
3654 int retval = -EINVAL;
3661 read_lock(&tasklist_lock);
3662 p = find_process_by_pid(pid);
3664 retval = security_task_getscheduler(p);
3668 read_unlock(&tasklist_lock);
3675 * sys_sched_getscheduler - get the RT priority of a thread
3676 * @pid: the pid in question.
3677 * @param: structure containing the RT priority.
3679 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
3681 struct sched_param lp;
3682 int retval = -EINVAL;
3685 if (!param || pid < 0)
3688 read_lock(&tasklist_lock);
3689 p = find_process_by_pid(pid);
3694 retval = security_task_getscheduler(p);
3698 lp.sched_priority = p->rt_priority;
3699 read_unlock(&tasklist_lock);
3702 * This one might sleep, we cannot do it with a spinlock held ...
3704 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3710 read_unlock(&tasklist_lock);
3714 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
3718 cpumask_t cpus_allowed;
3721 read_lock(&tasklist_lock);
3723 p = find_process_by_pid(pid);
3725 read_unlock(&tasklist_lock);
3726 unlock_cpu_hotplug();
3731 * It is not safe to call set_cpus_allowed with the
3732 * tasklist_lock held. We will bump the task_struct's
3733 * usage count and then drop tasklist_lock.
3736 read_unlock(&tasklist_lock);
3739 if ((current->euid != p->euid) && (current->euid != p->uid) &&
3740 !capable(CAP_SYS_NICE))
3743 cpus_allowed = cpuset_cpus_allowed(p);
3744 cpus_and(new_mask, new_mask, cpus_allowed);
3745 retval = set_cpus_allowed(p, new_mask);
3749 unlock_cpu_hotplug();
3753 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3754 cpumask_t *new_mask)
3756 if (len < sizeof(cpumask_t)) {
3757 memset(new_mask, 0, sizeof(cpumask_t));
3758 } else if (len > sizeof(cpumask_t)) {
3759 len = sizeof(cpumask_t);
3761 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3765 * sys_sched_setaffinity - set the cpu affinity of a process
3766 * @pid: pid of the process
3767 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3768 * @user_mask_ptr: user-space pointer to the new cpu mask
3770 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
3771 unsigned long __user *user_mask_ptr)
3776 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
3780 return sched_setaffinity(pid, new_mask);
3784 * Represents all cpu's present in the system
3785 * In systems capable of hotplug, this map could dynamically grow
3786 * as new cpu's are detected in the system via any platform specific
3787 * method, such as ACPI for e.g.
3790 cpumask_t cpu_present_map;
3791 EXPORT_SYMBOL(cpu_present_map);
3794 cpumask_t cpu_online_map = CPU_MASK_ALL;
3795 cpumask_t cpu_possible_map = CPU_MASK_ALL;
3798 long sched_getaffinity(pid_t pid, cpumask_t *mask)
3804 read_lock(&tasklist_lock);
3807 p = find_process_by_pid(pid);
3812 cpus_and(*mask, p->cpus_allowed, cpu_possible_map);
3815 read_unlock(&tasklist_lock);
3816 unlock_cpu_hotplug();
3824 * sys_sched_getaffinity - get the cpu affinity of a process
3825 * @pid: pid of the process
3826 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3827 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3829 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
3830 unsigned long __user *user_mask_ptr)
3835 if (len < sizeof(cpumask_t))
3838 ret = sched_getaffinity(pid, &mask);
3842 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
3845 return sizeof(cpumask_t);
3849 * sys_sched_yield - yield the current processor to other threads.
3851 * this function yields the current CPU by moving the calling thread
3852 * to the expired array. If there are no other threads running on this
3853 * CPU then this function will return.
3855 asmlinkage long sys_sched_yield(void)
3857 runqueue_t *rq = this_rq_lock();
3858 prio_array_t *array = current->array;
3859 prio_array_t *target = rq->expired;
3861 schedstat_inc(rq, yld_cnt);
3863 * We implement yielding by moving the task into the expired
3866 * (special rule: RT tasks will just roundrobin in the active
3869 if (rt_task(current))
3870 target = rq->active;
3872 if (current->array->nr_active == 1) {
3873 schedstat_inc(rq, yld_act_empty);
3874 if (!rq->expired->nr_active)
3875 schedstat_inc(rq, yld_both_empty);
3876 } else if (!rq->expired->nr_active)
3877 schedstat_inc(rq, yld_exp_empty);
3879 if (array != target) {
3880 dequeue_task(current, array);
3881 enqueue_task(current, target);
3884 * requeue_task is cheaper so perform that if possible.
3886 requeue_task(current, array);
3889 * Since we are going to call schedule() anyway, there's
3890 * no need to preempt or enable interrupts:
3892 __release(rq->lock);
3893 _raw_spin_unlock(&rq->lock);
3894 preempt_enable_no_resched();
3901 static inline void __cond_resched(void)
3904 add_preempt_count(PREEMPT_ACTIVE);
3906 sub_preempt_count(PREEMPT_ACTIVE);
3907 } while (need_resched());
3910 int __sched cond_resched(void)
3912 if (need_resched()) {
3919 EXPORT_SYMBOL(cond_resched);
3922 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
3923 * call schedule, and on return reacquire the lock.
3925 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
3926 * operations here to prevent schedule() from being called twice (once via
3927 * spin_unlock(), once by hand).
3929 int cond_resched_lock(spinlock_t * lock)
3933 if (need_lockbreak(lock)) {
3939 if (need_resched()) {
3940 _raw_spin_unlock(lock);
3941 preempt_enable_no_resched();
3949 EXPORT_SYMBOL(cond_resched_lock);
3951 int __sched cond_resched_softirq(void)
3953 BUG_ON(!in_softirq());
3955 if (need_resched()) {
3956 __local_bh_enable();
3964 EXPORT_SYMBOL(cond_resched_softirq);
3968 * yield - yield the current processor to other threads.
3970 * this is a shortcut for kernel-space yielding - it marks the
3971 * thread runnable and calls sys_sched_yield().
3973 void __sched yield(void)
3975 set_current_state(TASK_RUNNING);
3979 EXPORT_SYMBOL(yield);
3982 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
3983 * that process accounting knows that this is a task in IO wait state.
3985 * But don't do that if it is a deliberate, throttling IO wait (this task
3986 * has set its backing_dev_info: the queue against which it should throttle)
3988 void __sched io_schedule(void)
3990 struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id());
3992 atomic_inc(&rq->nr_iowait);
3994 atomic_dec(&rq->nr_iowait);
3997 EXPORT_SYMBOL(io_schedule);
3999 long __sched io_schedule_timeout(long timeout)
4001 struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id());
4004 atomic_inc(&rq->nr_iowait);
4005 ret = schedule_timeout(timeout);
4006 atomic_dec(&rq->nr_iowait);
4011 * sys_sched_get_priority_max - return maximum RT priority.
4012 * @policy: scheduling class.
4014 * this syscall returns the maximum rt_priority that can be used
4015 * by a given scheduling class.
4017 asmlinkage long sys_sched_get_priority_max(int policy)
4024 ret = MAX_USER_RT_PRIO-1;
4034 * sys_sched_get_priority_min - return minimum RT priority.
4035 * @policy: scheduling class.
4037 * this syscall returns the minimum rt_priority that can be used
4038 * by a given scheduling class.
4040 asmlinkage long sys_sched_get_priority_min(int policy)
4056 * sys_sched_rr_get_interval - return the default timeslice of a process.
4057 * @pid: pid of the process.
4058 * @interval: userspace pointer to the timeslice value.
4060 * this syscall writes the default timeslice value of a given process
4061 * into the user-space timespec buffer. A value of '0' means infinity.
4064 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4066 int retval = -EINVAL;
4074 read_lock(&tasklist_lock);
4075 p = find_process_by_pid(pid);
4079 retval = security_task_getscheduler(p);
4083 jiffies_to_timespec(p->policy & SCHED_FIFO ?
4084 0 : task_timeslice(p), &t);
4085 read_unlock(&tasklist_lock);
4086 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4090 read_unlock(&tasklist_lock);
4094 static inline struct task_struct *eldest_child(struct task_struct *p)
4096 if (list_empty(&p->children)) return NULL;
4097 return list_entry(p->children.next,struct task_struct,sibling);
4100 static inline struct task_struct *older_sibling(struct task_struct *p)
4102 if (p->sibling.prev==&p->parent->children) return NULL;
4103 return list_entry(p->sibling.prev,struct task_struct,sibling);
4106 static inline struct task_struct *younger_sibling(struct task_struct *p)
4108 if (p->sibling.next==&p->parent->children) return NULL;
4109 return list_entry(p->sibling.next,struct task_struct,sibling);
4112 static void show_task(task_t * p)
4116 unsigned long free = 0;
4117 static const char *stat_nam[] = { "R", "S", "D", "T", "t", "Z", "X" };
4119 printk("%-13.13s ", p->comm);
4120 state = p->state ? __ffs(p->state) + 1 : 0;
4121 if (state < ARRAY_SIZE(stat_nam))
4122 printk(stat_nam[state]);
4125 #if (BITS_PER_LONG == 32)
4126 if (state == TASK_RUNNING)
4127 printk(" running ");
4129 printk(" %08lX ", thread_saved_pc(p));
4131 if (state == TASK_RUNNING)
4132 printk(" running task ");
4134 printk(" %016lx ", thread_saved_pc(p));
4136 #ifdef CONFIG_DEBUG_STACK_USAGE
4138 unsigned long * n = (unsigned long *) (p->thread_info+1);
4141 free = (unsigned long) n - (unsigned long)(p->thread_info+1);
4144 printk("%5lu %5d %6d ", free, p->pid, p->parent->pid);
4145 if ((relative = eldest_child(p)))
4146 printk("%5d ", relative->pid);
4149 if ((relative = younger_sibling(p)))
4150 printk("%7d", relative->pid);
4153 if ((relative = older_sibling(p)))
4154 printk(" %5d", relative->pid);
4158 printk(" (L-TLB)\n");
4160 printk(" (NOTLB)\n");
4162 if (state != TASK_RUNNING)
4163 show_stack(p, NULL);
4166 void show_state(void)
4170 #if (BITS_PER_LONG == 32)
4173 printk(" task PC pid father child younger older\n");
4177 printk(" task PC pid father child younger older\n");
4179 read_lock(&tasklist_lock);
4180 do_each_thread(g, p) {
4182 * reset the NMI-timeout, listing all files on a slow
4183 * console might take alot of time:
4185 touch_nmi_watchdog();
4187 } while_each_thread(g, p);
4189 read_unlock(&tasklist_lock);
4192 void __devinit init_idle(task_t *idle, int cpu)
4194 runqueue_t *rq = cpu_rq(cpu);
4195 unsigned long flags;
4197 idle->sleep_avg = 0;
4199 idle->prio = MAX_PRIO;
4200 idle->state = TASK_RUNNING;
4201 idle->cpus_allowed = cpumask_of_cpu(cpu);
4202 set_task_cpu(idle, cpu);
4204 spin_lock_irqsave(&rq->lock, flags);
4205 rq->curr = rq->idle = idle;
4206 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4209 set_tsk_need_resched(idle);
4210 spin_unlock_irqrestore(&rq->lock, flags);
4212 /* Set the preempt count _outside_ the spinlocks! */
4213 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4214 idle->thread_info->preempt_count = (idle->lock_depth >= 0);
4216 idle->thread_info->preempt_count = 0;
4221 * In a system that switches off the HZ timer nohz_cpu_mask
4222 * indicates which cpus entered this state. This is used
4223 * in the rcu update to wait only for active cpus. For system
4224 * which do not switch off the HZ timer nohz_cpu_mask should
4225 * always be CPU_MASK_NONE.
4227 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4231 * This is how migration works:
4233 * 1) we queue a migration_req_t structure in the source CPU's
4234 * runqueue and wake up that CPU's migration thread.
4235 * 2) we down() the locked semaphore => thread blocks.
4236 * 3) migration thread wakes up (implicitly it forces the migrated
4237 * thread off the CPU)
4238 * 4) it gets the migration request and checks whether the migrated
4239 * task is still in the wrong runqueue.
4240 * 5) if it's in the wrong runqueue then the migration thread removes
4241 * it and puts it into the right queue.
4242 * 6) migration thread up()s the semaphore.
4243 * 7) we wake up and the migration is done.
4247 * Change a given task's CPU affinity. Migrate the thread to a
4248 * proper CPU and schedule it away if the CPU it's executing on
4249 * is removed from the allowed bitmask.
4251 * NOTE: the caller must have a valid reference to the task, the
4252 * task must not exit() & deallocate itself prematurely. The
4253 * call is not atomic; no spinlocks may be held.
4255 int set_cpus_allowed(task_t *p, cpumask_t new_mask)
4257 unsigned long flags;
4259 migration_req_t req;
4262 rq = task_rq_lock(p, &flags);
4263 if (!cpus_intersects(new_mask, cpu_online_map)) {
4268 p->cpus_allowed = new_mask;
4269 /* Can the task run on the task's current CPU? If so, we're done */
4270 if (cpu_isset(task_cpu(p), new_mask))
4273 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4274 /* Need help from migration thread: drop lock and wait. */
4275 task_rq_unlock(rq, &flags);
4276 wake_up_process(rq->migration_thread);
4277 wait_for_completion(&req.done);
4278 tlb_migrate_finish(p->mm);
4282 task_rq_unlock(rq, &flags);
4286 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4289 * Move (not current) task off this cpu, onto dest cpu. We're doing
4290 * this because either it can't run here any more (set_cpus_allowed()
4291 * away from this CPU, or CPU going down), or because we're
4292 * attempting to rebalance this task on exec (sched_exec).
4294 * So we race with normal scheduler movements, but that's OK, as long
4295 * as the task is no longer on this CPU.
4297 static void __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4299 runqueue_t *rq_dest, *rq_src;
4301 if (unlikely(cpu_is_offline(dest_cpu)))
4304 rq_src = cpu_rq(src_cpu);
4305 rq_dest = cpu_rq(dest_cpu);
4307 double_rq_lock(rq_src, rq_dest);
4308 /* Already moved. */
4309 if (task_cpu(p) != src_cpu)
4311 /* Affinity changed (again). */
4312 if (!cpu_isset(dest_cpu, p->cpus_allowed))
4315 set_task_cpu(p, dest_cpu);
4318 * Sync timestamp with rq_dest's before activating.
4319 * The same thing could be achieved by doing this step
4320 * afterwards, and pretending it was a local activate.
4321 * This way is cleaner and logically correct.
4323 p->timestamp = p->timestamp - rq_src->timestamp_last_tick
4324 + rq_dest->timestamp_last_tick;
4325 deactivate_task(p, rq_src);
4326 activate_task(p, rq_dest, 0);
4327 if (TASK_PREEMPTS_CURR(p, rq_dest))
4328 resched_task(rq_dest->curr);
4332 double_rq_unlock(rq_src, rq_dest);
4336 * migration_thread - this is a highprio system thread that performs
4337 * thread migration by bumping thread off CPU then 'pushing' onto
4340 static int migration_thread(void * data)
4343 int cpu = (long)data;
4346 BUG_ON(rq->migration_thread != current);
4348 set_current_state(TASK_INTERRUPTIBLE);
4349 while (!kthread_should_stop()) {
4350 struct list_head *head;
4351 migration_req_t *req;
4353 if (current->flags & PF_FREEZE)
4354 refrigerator(PF_FREEZE);
4356 spin_lock_irq(&rq->lock);
4358 if (cpu_is_offline(cpu)) {
4359 spin_unlock_irq(&rq->lock);
4363 if (rq->active_balance) {
4364 active_load_balance(rq, cpu);
4365 rq->active_balance = 0;
4368 head = &rq->migration_queue;
4370 if (list_empty(head)) {
4371 spin_unlock_irq(&rq->lock);
4373 set_current_state(TASK_INTERRUPTIBLE);
4376 req = list_entry(head->next, migration_req_t, list);
4377 list_del_init(head->next);
4379 spin_unlock(&rq->lock);
4380 __migrate_task(req->task, cpu, req->dest_cpu);
4383 complete(&req->done);
4385 __set_current_state(TASK_RUNNING);
4389 /* Wait for kthread_stop */
4390 set_current_state(TASK_INTERRUPTIBLE);
4391 while (!kthread_should_stop()) {
4393 set_current_state(TASK_INTERRUPTIBLE);
4395 __set_current_state(TASK_RUNNING);
4399 #ifdef CONFIG_HOTPLUG_CPU
4400 /* Figure out where task on dead CPU should go, use force if neccessary. */
4401 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *tsk)
4407 mask = node_to_cpumask(cpu_to_node(dead_cpu));
4408 cpus_and(mask, mask, tsk->cpus_allowed);
4409 dest_cpu = any_online_cpu(mask);
4411 /* On any allowed CPU? */
4412 if (dest_cpu == NR_CPUS)
4413 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4415 /* No more Mr. Nice Guy. */
4416 if (dest_cpu == NR_CPUS) {
4417 cpus_setall(tsk->cpus_allowed);
4418 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4421 * Don't tell them about moving exiting tasks or
4422 * kernel threads (both mm NULL), since they never
4425 if (tsk->mm && printk_ratelimit())
4426 printk(KERN_INFO "process %d (%s) no "
4427 "longer affine to cpu%d\n",
4428 tsk->pid, tsk->comm, dead_cpu);
4430 __migrate_task(tsk, dead_cpu, dest_cpu);
4434 * While a dead CPU has no uninterruptible tasks queued at this point,
4435 * it might still have a nonzero ->nr_uninterruptible counter, because
4436 * for performance reasons the counter is not stricly tracking tasks to
4437 * their home CPUs. So we just add the counter to another CPU's counter,
4438 * to keep the global sum constant after CPU-down:
4440 static void migrate_nr_uninterruptible(runqueue_t *rq_src)
4442 runqueue_t *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
4443 unsigned long flags;
4445 local_irq_save(flags);
4446 double_rq_lock(rq_src, rq_dest);
4447 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
4448 rq_src->nr_uninterruptible = 0;
4449 double_rq_unlock(rq_src, rq_dest);
4450 local_irq_restore(flags);
4453 /* Run through task list and migrate tasks from the dead cpu. */
4454 static void migrate_live_tasks(int src_cpu)
4456 struct task_struct *tsk, *t;
4458 write_lock_irq(&tasklist_lock);
4460 do_each_thread(t, tsk) {
4464 if (task_cpu(tsk) == src_cpu)
4465 move_task_off_dead_cpu(src_cpu, tsk);
4466 } while_each_thread(t, tsk);
4468 write_unlock_irq(&tasklist_lock);
4471 /* Schedules idle task to be the next runnable task on current CPU.
4472 * It does so by boosting its priority to highest possible and adding it to
4473 * the _front_ of runqueue. Used by CPU offline code.
4475 void sched_idle_next(void)
4477 int cpu = smp_processor_id();
4478 runqueue_t *rq = this_rq();
4479 struct task_struct *p = rq->idle;
4480 unsigned long flags;
4482 /* cpu has to be offline */
4483 BUG_ON(cpu_online(cpu));
4485 /* Strictly not necessary since rest of the CPUs are stopped by now
4486 * and interrupts disabled on current cpu.
4488 spin_lock_irqsave(&rq->lock, flags);
4490 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4491 /* Add idle task to _front_ of it's priority queue */
4492 __activate_idle_task(p, rq);
4494 spin_unlock_irqrestore(&rq->lock, flags);
4497 /* Ensures that the idle task is using init_mm right before its cpu goes
4500 void idle_task_exit(void)
4502 struct mm_struct *mm = current->active_mm;
4504 BUG_ON(cpu_online(smp_processor_id()));
4507 switch_mm(mm, &init_mm, current);
4511 static void migrate_dead(unsigned int dead_cpu, task_t *tsk)
4513 struct runqueue *rq = cpu_rq(dead_cpu);
4515 /* Must be exiting, otherwise would be on tasklist. */
4516 BUG_ON(tsk->exit_state != EXIT_ZOMBIE && tsk->exit_state != EXIT_DEAD);
4518 /* Cannot have done final schedule yet: would have vanished. */
4519 BUG_ON(tsk->flags & PF_DEAD);
4521 get_task_struct(tsk);
4524 * Drop lock around migration; if someone else moves it,
4525 * that's OK. No task can be added to this CPU, so iteration is
4528 spin_unlock_irq(&rq->lock);
4529 move_task_off_dead_cpu(dead_cpu, tsk);
4530 spin_lock_irq(&rq->lock);
4532 put_task_struct(tsk);
4535 /* release_task() removes task from tasklist, so we won't find dead tasks. */
4536 static void migrate_dead_tasks(unsigned int dead_cpu)
4539 struct runqueue *rq = cpu_rq(dead_cpu);
4541 for (arr = 0; arr < 2; arr++) {
4542 for (i = 0; i < MAX_PRIO; i++) {
4543 struct list_head *list = &rq->arrays[arr].queue[i];
4544 while (!list_empty(list))
4545 migrate_dead(dead_cpu,
4546 list_entry(list->next, task_t,
4551 #endif /* CONFIG_HOTPLUG_CPU */
4554 * migration_call - callback that gets triggered when a CPU is added.
4555 * Here we can start up the necessary migration thread for the new CPU.
4557 static int migration_call(struct notifier_block *nfb, unsigned long action,
4560 int cpu = (long)hcpu;
4561 struct task_struct *p;
4562 struct runqueue *rq;
4563 unsigned long flags;
4566 case CPU_UP_PREPARE:
4567 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
4570 p->flags |= PF_NOFREEZE;
4571 kthread_bind(p, cpu);
4572 /* Must be high prio: stop_machine expects to yield to it. */
4573 rq = task_rq_lock(p, &flags);
4574 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4575 task_rq_unlock(rq, &flags);
4576 cpu_rq(cpu)->migration_thread = p;
4579 /* Strictly unneccessary, as first user will wake it. */
4580 wake_up_process(cpu_rq(cpu)->migration_thread);
4582 #ifdef CONFIG_HOTPLUG_CPU
4583 case CPU_UP_CANCELED:
4584 /* Unbind it from offline cpu so it can run. Fall thru. */
4585 kthread_bind(cpu_rq(cpu)->migration_thread,smp_processor_id());
4586 kthread_stop(cpu_rq(cpu)->migration_thread);
4587 cpu_rq(cpu)->migration_thread = NULL;
4590 migrate_live_tasks(cpu);
4592 kthread_stop(rq->migration_thread);
4593 rq->migration_thread = NULL;
4594 /* Idle task back to normal (off runqueue, low prio) */
4595 rq = task_rq_lock(rq->idle, &flags);
4596 deactivate_task(rq->idle, rq);
4597 rq->idle->static_prio = MAX_PRIO;
4598 __setscheduler(rq->idle, SCHED_NORMAL, 0);
4599 migrate_dead_tasks(cpu);
4600 task_rq_unlock(rq, &flags);
4601 migrate_nr_uninterruptible(rq);
4602 BUG_ON(rq->nr_running != 0);
4604 /* No need to migrate the tasks: it was best-effort if
4605 * they didn't do lock_cpu_hotplug(). Just wake up
4606 * the requestors. */
4607 spin_lock_irq(&rq->lock);
4608 while (!list_empty(&rq->migration_queue)) {
4609 migration_req_t *req;
4610 req = list_entry(rq->migration_queue.next,
4611 migration_req_t, list);
4612 list_del_init(&req->list);
4613 complete(&req->done);
4615 spin_unlock_irq(&rq->lock);
4622 /* Register at highest priority so that task migration (migrate_all_tasks)
4623 * happens before everything else.
4625 static struct notifier_block __devinitdata migration_notifier = {
4626 .notifier_call = migration_call,
4630 int __init migration_init(void)
4632 void *cpu = (void *)(long)smp_processor_id();
4633 /* Start one for boot CPU. */
4634 migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
4635 migration_call(&migration_notifier, CPU_ONLINE, cpu);
4636 register_cpu_notifier(&migration_notifier);
4642 #define SCHED_DOMAIN_DEBUG
4643 #ifdef SCHED_DOMAIN_DEBUG
4644 static void sched_domain_debug(struct sched_domain *sd, int cpu)
4649 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
4653 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
4658 struct sched_group *group = sd->groups;
4659 cpumask_t groupmask;
4661 cpumask_scnprintf(str, NR_CPUS, sd->span);
4662 cpus_clear(groupmask);
4665 for (i = 0; i < level + 1; i++)
4667 printk("domain %d: ", level);
4669 if (!(sd->flags & SD_LOAD_BALANCE)) {
4670 printk("does not load-balance\n");
4672 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent");
4676 printk("span %s\n", str);
4678 if (!cpu_isset(cpu, sd->span))
4679 printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
4680 if (!cpu_isset(cpu, group->cpumask))
4681 printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
4684 for (i = 0; i < level + 2; i++)
4690 printk(KERN_ERR "ERROR: group is NULL\n");
4694 if (!group->cpu_power) {
4696 printk(KERN_ERR "ERROR: domain->cpu_power not set\n");
4699 if (!cpus_weight(group->cpumask)) {
4701 printk(KERN_ERR "ERROR: empty group\n");
4704 if (cpus_intersects(groupmask, group->cpumask)) {
4706 printk(KERN_ERR "ERROR: repeated CPUs\n");
4709 cpus_or(groupmask, groupmask, group->cpumask);
4711 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
4714 group = group->next;
4715 } while (group != sd->groups);
4718 if (!cpus_equal(sd->span, groupmask))
4719 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
4725 if (!cpus_subset(groupmask, sd->span))
4726 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
4732 #define sched_domain_debug(sd, cpu) {}
4735 static int __devinit sd_degenerate(struct sched_domain *sd)
4737 if (cpus_weight(sd->span) == 1)
4740 /* Following flags need at least 2 groups */
4741 if (sd->flags & (SD_LOAD_BALANCE |
4742 SD_BALANCE_NEWIDLE |
4745 if (sd->groups != sd->groups->next)
4749 /* Following flags don't use groups */
4750 if (sd->flags & (SD_WAKE_IDLE |
4758 static int __devinit sd_parent_degenerate(struct sched_domain *sd,
4759 struct sched_domain *parent)
4761 unsigned long cflags = sd->flags, pflags = parent->flags;
4763 if (sd_degenerate(parent))
4766 if (!cpus_equal(sd->span, parent->span))
4769 /* Does parent contain flags not in child? */
4770 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
4771 if (cflags & SD_WAKE_AFFINE)
4772 pflags &= ~SD_WAKE_BALANCE;
4773 /* Flags needing groups don't count if only 1 group in parent */
4774 if (parent->groups == parent->groups->next) {
4775 pflags &= ~(SD_LOAD_BALANCE |
4776 SD_BALANCE_NEWIDLE |
4780 if (~cflags & pflags)
4787 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
4788 * hold the hotplug lock.
4790 void __devinit cpu_attach_domain(struct sched_domain *sd, int cpu)
4792 runqueue_t *rq = cpu_rq(cpu);
4793 struct sched_domain *tmp;
4795 /* Remove the sched domains which do not contribute to scheduling. */
4796 for (tmp = sd; tmp; tmp = tmp->parent) {
4797 struct sched_domain *parent = tmp->parent;
4800 if (sd_parent_degenerate(tmp, parent))
4801 tmp->parent = parent->parent;
4804 if (sd && sd_degenerate(sd))
4807 sched_domain_debug(sd, cpu);
4809 rcu_assign_pointer(rq->sd, sd);
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 synchronize_kernel();
5085 arch_destroy_sched_domains();
5088 case CPU_UP_CANCELED:
5089 case CPU_DOWN_FAILED:
5093 * Fall through and re-initialise the domains.
5100 /* The hotplug lock is already held by cpu_up/cpu_down */
5101 arch_init_sched_domains();
5107 void __init sched_init_smp(void)
5110 arch_init_sched_domains();
5111 unlock_cpu_hotplug();
5112 /* XXX: Theoretical race here - CPU may be hotplugged now */
5113 hotcpu_notifier(update_sched_domains, 0);
5116 void __init sched_init_smp(void)
5119 #endif /* CONFIG_SMP */
5121 int in_sched_functions(unsigned long addr)
5123 /* Linker adds these: start and end of __sched functions */
5124 extern char __sched_text_start[], __sched_text_end[];
5125 return in_lock_functions(addr) ||
5126 (addr >= (unsigned long)__sched_text_start
5127 && addr < (unsigned long)__sched_text_end);
5130 void __init sched_init(void)
5135 for (i = 0; i < NR_CPUS; i++) {
5136 prio_array_t *array;
5139 spin_lock_init(&rq->lock);
5141 rq->active = rq->arrays;
5142 rq->expired = rq->arrays + 1;
5143 rq->best_expired_prio = MAX_PRIO;
5147 for (j = 1; j < 3; j++)
5148 rq->cpu_load[j] = 0;
5149 rq->active_balance = 0;
5151 rq->migration_thread = NULL;
5152 INIT_LIST_HEAD(&rq->migration_queue);
5154 atomic_set(&rq->nr_iowait, 0);
5156 for (j = 0; j < 2; j++) {
5157 array = rq->arrays + j;
5158 for (k = 0; k < MAX_PRIO; k++) {
5159 INIT_LIST_HEAD(array->queue + k);
5160 __clear_bit(k, array->bitmap);
5162 // delimiter for bitsearch
5163 __set_bit(MAX_PRIO, array->bitmap);
5168 * The boot idle thread does lazy MMU switching as well:
5170 atomic_inc(&init_mm.mm_count);
5171 enter_lazy_tlb(&init_mm, current);
5174 * Make us the idle thread. Technically, schedule() should not be
5175 * called from this thread, however somewhere below it might be,
5176 * but because we are the idle thread, we just pick up running again
5177 * when this runqueue becomes "idle".
5179 init_idle(current, smp_processor_id());
5182 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5183 void __might_sleep(char *file, int line)
5185 #if defined(in_atomic)
5186 static unsigned long prev_jiffy; /* ratelimiting */
5188 if ((in_atomic() || irqs_disabled()) &&
5189 system_state == SYSTEM_RUNNING && !oops_in_progress) {
5190 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
5192 prev_jiffy = jiffies;
5193 printk(KERN_ERR "Debug: sleeping function called from invalid"
5194 " context at %s:%d\n", file, line);
5195 printk("in_atomic():%d, irqs_disabled():%d\n",
5196 in_atomic(), irqs_disabled());
5201 EXPORT_SYMBOL(__might_sleep);
5204 #ifdef CONFIG_MAGIC_SYSRQ
5205 void normalize_rt_tasks(void)
5207 struct task_struct *p;
5208 prio_array_t *array;
5209 unsigned long flags;
5212 read_lock_irq(&tasklist_lock);
5213 for_each_process (p) {
5217 rq = task_rq_lock(p, &flags);
5221 deactivate_task(p, task_rq(p));
5222 __setscheduler(p, SCHED_NORMAL, 0);
5224 __activate_task(p, task_rq(p));
5225 resched_task(rq->curr);
5228 task_rq_unlock(rq, &flags);
5230 read_unlock_irq(&tasklist_lock);
5233 #endif /* CONFIG_MAGIC_SYSRQ */