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/capability.h>
31 #include <linux/completion.h>
32 #include <linux/kernel_stat.h>
33 #include <linux/security.h>
34 #include <linux/notifier.h>
35 #include <linux/profile.h>
36 #include <linux/suspend.h>
37 #include <linux/vmalloc.h>
38 #include <linux/blkdev.h>
39 #include <linux/delay.h>
40 #include <linux/smp.h>
41 #include <linux/threads.h>
42 #include <linux/timer.h>
43 #include <linux/rcupdate.h>
44 #include <linux/cpu.h>
45 #include <linux/cpuset.h>
46 #include <linux/percpu.h>
47 #include <linux/kthread.h>
48 #include <linux/seq_file.h>
49 #include <linux/syscalls.h>
50 #include <linux/times.h>
51 #include <linux/acct.h>
54 #include <asm/unistd.h>
57 * Convert user-nice values [ -20 ... 0 ... 19 ]
58 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
61 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
62 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
63 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
66 * 'User priority' is the nice value converted to something we
67 * can work with better when scaling various scheduler parameters,
68 * it's a [ 0 ... 39 ] range.
70 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
71 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
72 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
75 * Some helpers for converting nanosecond timing to jiffy resolution
77 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
78 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
81 * These are the 'tuning knobs' of the scheduler:
83 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
84 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
85 * Timeslices get refilled after they expire.
87 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
88 #define DEF_TIMESLICE (100 * HZ / 1000)
89 #define ON_RUNQUEUE_WEIGHT 30
90 #define CHILD_PENALTY 95
91 #define PARENT_PENALTY 100
93 #define PRIO_BONUS_RATIO 25
94 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
95 #define INTERACTIVE_DELTA 2
96 #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
97 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
98 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
101 * If a task is 'interactive' then we reinsert it in the active
102 * array after it has expired its current timeslice. (it will not
103 * continue to run immediately, it will still roundrobin with
104 * other interactive tasks.)
106 * This part scales the interactivity limit depending on niceness.
108 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
109 * Here are a few examples of different nice levels:
111 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
112 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
113 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
114 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
115 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
117 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
118 * priority range a task can explore, a value of '1' means the
119 * task is rated interactive.)
121 * Ie. nice +19 tasks can never get 'interactive' enough to be
122 * reinserted into the active array. And only heavily CPU-hog nice -20
123 * tasks will be expired. Default nice 0 tasks are somewhere between,
124 * it takes some effort for them to get interactive, but it's not
128 #define CURRENT_BONUS(p) \
129 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
132 #define GRANULARITY (10 * HZ / 1000 ? : 1)
135 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
136 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
139 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
140 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
143 #define SCALE(v1,v1_max,v2_max) \
144 (v1) * (v2_max) / (v1_max)
147 (SCALE(TASK_NICE(p), 40, MAX_BONUS) + INTERACTIVE_DELTA)
149 #define TASK_INTERACTIVE(p) \
150 ((p)->prio <= (p)->static_prio - DELTA(p))
152 #define INTERACTIVE_SLEEP(p) \
153 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
154 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
156 #define TASK_PREEMPTS_CURR(p, rq) \
157 ((p)->prio < (rq)->curr->prio)
160 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
161 * to time slice values: [800ms ... 100ms ... 5ms]
163 * The higher a thread's priority, the bigger timeslices
164 * it gets during one round of execution. But even the lowest
165 * priority thread gets MIN_TIMESLICE worth of execution time.
168 #define SCALE_PRIO(x, prio) \
169 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO/2), MIN_TIMESLICE)
171 static unsigned int task_timeslice(task_t *p)
173 if (p->static_prio < NICE_TO_PRIO(0))
174 return SCALE_PRIO(DEF_TIMESLICE*4, p->static_prio);
176 return SCALE_PRIO(DEF_TIMESLICE, p->static_prio);
178 #define task_hot(p, now, sd) ((long long) ((now) - (p)->last_ran) \
179 < (long long) (sd)->cache_hot_time)
181 void __put_task_struct_cb(struct rcu_head *rhp)
183 __put_task_struct(container_of(rhp, struct task_struct, rcu));
186 EXPORT_SYMBOL_GPL(__put_task_struct_cb);
189 * These are the runqueue data structures:
192 #define BITMAP_SIZE ((((MAX_PRIO+1+7)/8)+sizeof(long)-1)/sizeof(long))
194 typedef struct runqueue runqueue_t;
197 unsigned int nr_active;
198 unsigned long bitmap[BITMAP_SIZE];
199 struct list_head queue[MAX_PRIO];
203 * This is the main, per-CPU runqueue data structure.
205 * Locking rule: those places that want to lock multiple runqueues
206 * (such as the load balancing or the thread migration code), lock
207 * acquire operations must be ordered by ascending &runqueue.
213 * nr_running and cpu_load should be in the same cacheline because
214 * remote CPUs use both these fields when doing load calculation.
216 unsigned long nr_running;
218 unsigned long prio_bias;
219 unsigned long cpu_load[3];
221 unsigned long long nr_switches;
224 * This is part of a global counter where only the total sum
225 * over all CPUs matters. A task can increase this counter on
226 * one CPU and if it got migrated afterwards it may decrease
227 * it on another CPU. Always updated under the runqueue lock:
229 unsigned long nr_uninterruptible;
231 unsigned long expired_timestamp;
232 unsigned long long timestamp_last_tick;
234 struct mm_struct *prev_mm;
235 prio_array_t *active, *expired, arrays[2];
236 int best_expired_prio;
240 struct sched_domain *sd;
242 /* For active balancing */
246 task_t *migration_thread;
247 struct list_head migration_queue;
250 #ifdef CONFIG_SCHEDSTATS
252 struct sched_info rq_sched_info;
254 /* sys_sched_yield() stats */
255 unsigned long yld_exp_empty;
256 unsigned long yld_act_empty;
257 unsigned long yld_both_empty;
258 unsigned long yld_cnt;
260 /* schedule() stats */
261 unsigned long sched_switch;
262 unsigned long sched_cnt;
263 unsigned long sched_goidle;
265 /* try_to_wake_up() stats */
266 unsigned long ttwu_cnt;
267 unsigned long ttwu_local;
271 static DEFINE_PER_CPU(struct runqueue, runqueues);
274 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
275 * See detach_destroy_domains: synchronize_sched for details.
277 * The domain tree of any CPU may only be accessed from within
278 * preempt-disabled sections.
280 #define for_each_domain(cpu, domain) \
281 for (domain = rcu_dereference(cpu_rq(cpu)->sd); domain; domain = domain->parent)
283 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
284 #define this_rq() (&__get_cpu_var(runqueues))
285 #define task_rq(p) cpu_rq(task_cpu(p))
286 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
288 #ifndef prepare_arch_switch
289 # define prepare_arch_switch(next) do { } while (0)
291 #ifndef finish_arch_switch
292 # define finish_arch_switch(prev) do { } while (0)
295 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
296 static inline int task_running(runqueue_t *rq, task_t *p)
298 return rq->curr == p;
301 static inline void prepare_lock_switch(runqueue_t *rq, task_t *next)
305 static inline void finish_lock_switch(runqueue_t *rq, task_t *prev)
307 #ifdef CONFIG_DEBUG_SPINLOCK
308 /* this is a valid case when another task releases the spinlock */
309 rq->lock.owner = current;
311 spin_unlock_irq(&rq->lock);
314 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
315 static inline int task_running(runqueue_t *rq, task_t *p)
320 return rq->curr == p;
324 static inline void prepare_lock_switch(runqueue_t *rq, task_t *next)
328 * We can optimise this out completely for !SMP, because the
329 * SMP rebalancing from interrupt is the only thing that cares
334 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
335 spin_unlock_irq(&rq->lock);
337 spin_unlock(&rq->lock);
341 static inline void finish_lock_switch(runqueue_t *rq, task_t *prev)
345 * After ->oncpu is cleared, the task can be moved to a different CPU.
346 * We must ensure this doesn't happen until the switch is completely
352 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
356 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
359 * task_rq_lock - lock the runqueue a given task resides on and disable
360 * interrupts. Note the ordering: we can safely lookup the task_rq without
361 * explicitly disabling preemption.
363 static inline runqueue_t *task_rq_lock(task_t *p, unsigned long *flags)
369 local_irq_save(*flags);
371 spin_lock(&rq->lock);
372 if (unlikely(rq != task_rq(p))) {
373 spin_unlock_irqrestore(&rq->lock, *flags);
374 goto repeat_lock_task;
379 static inline void task_rq_unlock(runqueue_t *rq, unsigned long *flags)
382 spin_unlock_irqrestore(&rq->lock, *flags);
385 #ifdef CONFIG_SCHEDSTATS
387 * bump this up when changing the output format or the meaning of an existing
388 * format, so that tools can adapt (or abort)
390 #define SCHEDSTAT_VERSION 12
392 static int show_schedstat(struct seq_file *seq, void *v)
396 seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION);
397 seq_printf(seq, "timestamp %lu\n", jiffies);
398 for_each_online_cpu(cpu) {
399 runqueue_t *rq = cpu_rq(cpu);
401 struct sched_domain *sd;
405 /* runqueue-specific stats */
407 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
408 cpu, rq->yld_both_empty,
409 rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt,
410 rq->sched_switch, rq->sched_cnt, rq->sched_goidle,
411 rq->ttwu_cnt, rq->ttwu_local,
412 rq->rq_sched_info.cpu_time,
413 rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt);
415 seq_printf(seq, "\n");
418 /* domain-specific stats */
420 for_each_domain(cpu, sd) {
421 enum idle_type itype;
422 char mask_str[NR_CPUS];
424 cpumask_scnprintf(mask_str, NR_CPUS, sd->span);
425 seq_printf(seq, "domain%d %s", dcnt++, mask_str);
426 for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES;
428 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu",
430 sd->lb_balanced[itype],
431 sd->lb_failed[itype],
432 sd->lb_imbalance[itype],
433 sd->lb_gained[itype],
434 sd->lb_hot_gained[itype],
435 sd->lb_nobusyq[itype],
436 sd->lb_nobusyg[itype]);
438 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
439 sd->alb_cnt, sd->alb_failed, sd->alb_pushed,
440 sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed,
441 sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed,
442 sd->ttwu_wake_remote, sd->ttwu_move_affine, sd->ttwu_move_balance);
450 static int schedstat_open(struct inode *inode, struct file *file)
452 unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32);
453 char *buf = kmalloc(size, GFP_KERNEL);
459 res = single_open(file, show_schedstat, NULL);
461 m = file->private_data;
469 struct file_operations proc_schedstat_operations = {
470 .open = schedstat_open,
473 .release = single_release,
476 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
477 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
478 #else /* !CONFIG_SCHEDSTATS */
479 # define schedstat_inc(rq, field) do { } while (0)
480 # define schedstat_add(rq, field, amt) do { } while (0)
484 * rq_lock - lock a given runqueue and disable interrupts.
486 static inline runqueue_t *this_rq_lock(void)
493 spin_lock(&rq->lock);
498 #ifdef CONFIG_SCHEDSTATS
500 * Called when a process is dequeued from the active array and given
501 * the cpu. We should note that with the exception of interactive
502 * tasks, the expired queue will become the active queue after the active
503 * queue is empty, without explicitly dequeuing and requeuing tasks in the
504 * expired queue. (Interactive tasks may be requeued directly to the
505 * active queue, thus delaying tasks in the expired queue from running;
506 * see scheduler_tick()).
508 * This function is only called from sched_info_arrive(), rather than
509 * dequeue_task(). Even though a task may be queued and dequeued multiple
510 * times as it is shuffled about, we're really interested in knowing how
511 * long it was from the *first* time it was queued to the time that it
514 static inline void sched_info_dequeued(task_t *t)
516 t->sched_info.last_queued = 0;
520 * Called when a task finally hits the cpu. We can now calculate how
521 * long it was waiting to run. We also note when it began so that we
522 * can keep stats on how long its timeslice is.
524 static inline void sched_info_arrive(task_t *t)
526 unsigned long now = jiffies, diff = 0;
527 struct runqueue *rq = task_rq(t);
529 if (t->sched_info.last_queued)
530 diff = now - t->sched_info.last_queued;
531 sched_info_dequeued(t);
532 t->sched_info.run_delay += diff;
533 t->sched_info.last_arrival = now;
534 t->sched_info.pcnt++;
539 rq->rq_sched_info.run_delay += diff;
540 rq->rq_sched_info.pcnt++;
544 * Called when a process is queued into either the active or expired
545 * array. The time is noted and later used to determine how long we
546 * had to wait for us to reach the cpu. Since the expired queue will
547 * become the active queue after active queue is empty, without dequeuing
548 * and requeuing any tasks, we are interested in queuing to either. It
549 * is unusual but not impossible for tasks to be dequeued and immediately
550 * requeued in the same or another array: this can happen in sched_yield(),
551 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
554 * This function is only called from enqueue_task(), but also only updates
555 * the timestamp if it is already not set. It's assumed that
556 * sched_info_dequeued() will clear that stamp when appropriate.
558 static inline void sched_info_queued(task_t *t)
560 if (!t->sched_info.last_queued)
561 t->sched_info.last_queued = jiffies;
565 * Called when a process ceases being the active-running process, either
566 * voluntarily or involuntarily. Now we can calculate how long we ran.
568 static inline void sched_info_depart(task_t *t)
570 struct runqueue *rq = task_rq(t);
571 unsigned long diff = jiffies - t->sched_info.last_arrival;
573 t->sched_info.cpu_time += diff;
576 rq->rq_sched_info.cpu_time += diff;
580 * Called when tasks are switched involuntarily due, typically, to expiring
581 * their time slice. (This may also be called when switching to or from
582 * the idle task.) We are only called when prev != next.
584 static inline void sched_info_switch(task_t *prev, task_t *next)
586 struct runqueue *rq = task_rq(prev);
589 * prev now departs the cpu. It's not interesting to record
590 * stats about how efficient we were at scheduling the idle
593 if (prev != rq->idle)
594 sched_info_depart(prev);
596 if (next != rq->idle)
597 sched_info_arrive(next);
600 #define sched_info_queued(t) do { } while (0)
601 #define sched_info_switch(t, next) do { } while (0)
602 #endif /* CONFIG_SCHEDSTATS */
605 * Adding/removing a task to/from a priority array:
607 static void dequeue_task(struct task_struct *p, prio_array_t *array)
610 list_del(&p->run_list);
611 if (list_empty(array->queue + p->prio))
612 __clear_bit(p->prio, array->bitmap);
615 static void enqueue_task(struct task_struct *p, prio_array_t *array)
617 sched_info_queued(p);
618 list_add_tail(&p->run_list, array->queue + p->prio);
619 __set_bit(p->prio, array->bitmap);
625 * Put task to the end of the run list without the overhead of dequeue
626 * followed by enqueue.
628 static void requeue_task(struct task_struct *p, prio_array_t *array)
630 list_move_tail(&p->run_list, array->queue + p->prio);
633 static inline void enqueue_task_head(struct task_struct *p, prio_array_t *array)
635 list_add(&p->run_list, array->queue + p->prio);
636 __set_bit(p->prio, array->bitmap);
642 * effective_prio - return the priority that is based on the static
643 * priority but is modified by bonuses/penalties.
645 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
646 * into the -5 ... 0 ... +5 bonus/penalty range.
648 * We use 25% of the full 0...39 priority range so that:
650 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
651 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
653 * Both properties are important to certain workloads.
655 static int effective_prio(task_t *p)
662 bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
664 prio = p->static_prio - bonus;
665 if (prio < MAX_RT_PRIO)
667 if (prio > MAX_PRIO-1)
673 static inline void inc_prio_bias(runqueue_t *rq, int prio)
675 rq->prio_bias += MAX_PRIO - prio;
678 static inline void dec_prio_bias(runqueue_t *rq, int prio)
680 rq->prio_bias -= MAX_PRIO - prio;
683 static inline void inc_nr_running(task_t *p, runqueue_t *rq)
687 if (p != rq->migration_thread)
689 * The migration thread does the actual balancing. Do
690 * not bias by its priority as the ultra high priority
691 * will skew balancing adversely.
693 inc_prio_bias(rq, p->prio);
695 inc_prio_bias(rq, p->static_prio);
698 static inline void dec_nr_running(task_t *p, runqueue_t *rq)
702 if (p != rq->migration_thread)
703 dec_prio_bias(rq, p->prio);
705 dec_prio_bias(rq, p->static_prio);
708 static inline void inc_prio_bias(runqueue_t *rq, int prio)
712 static inline void dec_prio_bias(runqueue_t *rq, int prio)
716 static inline void inc_nr_running(task_t *p, runqueue_t *rq)
721 static inline void dec_nr_running(task_t *p, runqueue_t *rq)
728 * __activate_task - move a task to the runqueue.
730 static inline void __activate_task(task_t *p, runqueue_t *rq)
732 enqueue_task(p, rq->active);
733 inc_nr_running(p, rq);
737 * __activate_idle_task - move idle task to the _front_ of runqueue.
739 static inline void __activate_idle_task(task_t *p, runqueue_t *rq)
741 enqueue_task_head(p, rq->active);
742 inc_nr_running(p, rq);
745 static int recalc_task_prio(task_t *p, unsigned long long now)
747 /* Caller must always ensure 'now >= p->timestamp' */
748 unsigned long long __sleep_time = now - p->timestamp;
749 unsigned long sleep_time;
751 if (__sleep_time > NS_MAX_SLEEP_AVG)
752 sleep_time = NS_MAX_SLEEP_AVG;
754 sleep_time = (unsigned long)__sleep_time;
756 if (likely(sleep_time > 0)) {
758 * User tasks that sleep a long time are categorised as
759 * idle and will get just interactive status to stay active &
760 * prevent them suddenly becoming cpu hogs and starving
763 if (p->mm && p->activated != -1 &&
764 sleep_time > INTERACTIVE_SLEEP(p)) {
765 p->sleep_avg = JIFFIES_TO_NS(MAX_SLEEP_AVG -
769 * The lower the sleep avg a task has the more
770 * rapidly it will rise with sleep time.
772 sleep_time *= (MAX_BONUS - CURRENT_BONUS(p)) ? : 1;
775 * Tasks waking from uninterruptible sleep are
776 * limited in their sleep_avg rise as they
777 * are likely to be waiting on I/O
779 if (p->activated == -1 && p->mm) {
780 if (p->sleep_avg >= INTERACTIVE_SLEEP(p))
782 else if (p->sleep_avg + sleep_time >=
783 INTERACTIVE_SLEEP(p)) {
784 p->sleep_avg = INTERACTIVE_SLEEP(p);
790 * This code gives a bonus to interactive tasks.
792 * The boost works by updating the 'average sleep time'
793 * value here, based on ->timestamp. The more time a
794 * task spends sleeping, the higher the average gets -
795 * and the higher the priority boost gets as well.
797 p->sleep_avg += sleep_time;
799 if (p->sleep_avg > NS_MAX_SLEEP_AVG)
800 p->sleep_avg = NS_MAX_SLEEP_AVG;
804 return effective_prio(p);
808 * activate_task - move a task to the runqueue and do priority recalculation
810 * Update all the scheduling statistics stuff. (sleep average
811 * calculation, priority modifiers, etc.)
813 static void activate_task(task_t *p, runqueue_t *rq, int local)
815 unsigned long long now;
820 /* Compensate for drifting sched_clock */
821 runqueue_t *this_rq = this_rq();
822 now = (now - this_rq->timestamp_last_tick)
823 + rq->timestamp_last_tick;
828 p->prio = recalc_task_prio(p, now);
831 * This checks to make sure it's not an uninterruptible task
832 * that is now waking up.
836 * Tasks which were woken up by interrupts (ie. hw events)
837 * are most likely of interactive nature. So we give them
838 * the credit of extending their sleep time to the period
839 * of time they spend on the runqueue, waiting for execution
840 * on a CPU, first time around:
846 * Normal first-time wakeups get a credit too for
847 * on-runqueue time, but it will be weighted down:
854 __activate_task(p, rq);
858 * deactivate_task - remove a task from the runqueue.
860 static void deactivate_task(struct task_struct *p, runqueue_t *rq)
862 dec_nr_running(p, rq);
863 dequeue_task(p, p->array);
868 * resched_task - mark a task 'to be rescheduled now'.
870 * On UP this means the setting of the need_resched flag, on SMP it
871 * might also involve a cross-CPU call to trigger the scheduler on
875 static void resched_task(task_t *p)
879 assert_spin_locked(&task_rq(p)->lock);
881 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
884 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
887 if (cpu == smp_processor_id())
890 /* NEED_RESCHED must be visible before we test POLLING_NRFLAG */
892 if (!test_tsk_thread_flag(p, TIF_POLLING_NRFLAG))
893 smp_send_reschedule(cpu);
896 static inline void resched_task(task_t *p)
898 assert_spin_locked(&task_rq(p)->lock);
899 set_tsk_need_resched(p);
904 * task_curr - is this task currently executing on a CPU?
905 * @p: the task in question.
907 inline int task_curr(const task_t *p)
909 return cpu_curr(task_cpu(p)) == p;
914 struct list_head list;
919 struct completion done;
923 * The task's runqueue lock must be held.
924 * Returns true if you have to wait for migration thread.
926 static int migrate_task(task_t *p, int dest_cpu, migration_req_t *req)
928 runqueue_t *rq = task_rq(p);
931 * If the task is not on a runqueue (and not running), then
932 * it is sufficient to simply update the task's cpu field.
934 if (!p->array && !task_running(rq, p)) {
935 set_task_cpu(p, dest_cpu);
939 init_completion(&req->done);
941 req->dest_cpu = dest_cpu;
942 list_add(&req->list, &rq->migration_queue);
947 * wait_task_inactive - wait for a thread to unschedule.
949 * The caller must ensure that the task *will* unschedule sometime soon,
950 * else this function might spin for a *long* time. This function can't
951 * be called with interrupts off, or it may introduce deadlock with
952 * smp_call_function() if an IPI is sent by the same process we are
953 * waiting to become inactive.
955 void wait_task_inactive(task_t *p)
962 rq = task_rq_lock(p, &flags);
963 /* Must be off runqueue entirely, not preempted. */
964 if (unlikely(p->array || task_running(rq, p))) {
965 /* If it's preempted, we yield. It could be a while. */
966 preempted = !task_running(rq, p);
967 task_rq_unlock(rq, &flags);
973 task_rq_unlock(rq, &flags);
977 * kick_process - kick a running thread to enter/exit the kernel
978 * @p: the to-be-kicked thread
980 * Cause a process which is running on another CPU to enter
981 * kernel-mode, without any delay. (to get signals handled.)
983 * NOTE: this function doesnt have to take the runqueue lock,
984 * because all it wants to ensure is that the remote task enters
985 * the kernel. If the IPI races and the task has been migrated
986 * to another CPU then no harm is done and the purpose has been
989 void kick_process(task_t *p)
995 if ((cpu != smp_processor_id()) && task_curr(p))
996 smp_send_reschedule(cpu);
1001 * Return a low guess at the load of a migration-source cpu.
1003 * We want to under-estimate the load of migration sources, to
1004 * balance conservatively.
1006 static inline unsigned long __source_load(int cpu, int type, enum idle_type idle)
1008 runqueue_t *rq = cpu_rq(cpu);
1009 unsigned long running = rq->nr_running;
1010 unsigned long source_load, cpu_load = rq->cpu_load[type-1],
1011 load_now = running * SCHED_LOAD_SCALE;
1014 source_load = load_now;
1016 source_load = min(cpu_load, load_now);
1018 if (running > 1 || (idle == NOT_IDLE && running))
1020 * If we are busy rebalancing the load is biased by
1021 * priority to create 'nice' support across cpus. When
1022 * idle rebalancing we should only bias the source_load if
1023 * there is more than one task running on that queue to
1024 * prevent idle rebalance from trying to pull tasks from a
1025 * queue with only one running task.
1027 source_load = source_load * rq->prio_bias / running;
1032 static inline unsigned long source_load(int cpu, int type)
1034 return __source_load(cpu, type, NOT_IDLE);
1038 * Return a high guess at the load of a migration-target cpu
1040 static inline unsigned long __target_load(int cpu, int type, enum idle_type idle)
1042 runqueue_t *rq = cpu_rq(cpu);
1043 unsigned long running = rq->nr_running;
1044 unsigned long target_load, cpu_load = rq->cpu_load[type-1],
1045 load_now = running * SCHED_LOAD_SCALE;
1048 target_load = load_now;
1050 target_load = max(cpu_load, load_now);
1052 if (running > 1 || (idle == NOT_IDLE && running))
1053 target_load = target_load * rq->prio_bias / running;
1058 static inline unsigned long target_load(int cpu, int type)
1060 return __target_load(cpu, type, NOT_IDLE);
1064 * find_idlest_group finds and returns the least busy CPU group within the
1067 static struct sched_group *
1068 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1070 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1071 unsigned long min_load = ULONG_MAX, this_load = 0;
1072 int load_idx = sd->forkexec_idx;
1073 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1076 unsigned long load, avg_load;
1080 /* Skip over this group if it has no CPUs allowed */
1081 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1084 local_group = cpu_isset(this_cpu, group->cpumask);
1086 /* Tally up the load of all CPUs in the group */
1089 for_each_cpu_mask(i, group->cpumask) {
1090 /* Bias balancing toward cpus of our domain */
1092 load = source_load(i, load_idx);
1094 load = target_load(i, load_idx);
1099 /* Adjust by relative CPU power of the group */
1100 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1103 this_load = avg_load;
1105 } else if (avg_load < min_load) {
1106 min_load = avg_load;
1110 group = group->next;
1111 } while (group != sd->groups);
1113 if (!idlest || 100*this_load < imbalance*min_load)
1119 * find_idlest_queue - find the idlest runqueue among the cpus in group.
1122 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1125 unsigned long load, min_load = ULONG_MAX;
1129 /* Traverse only the allowed CPUs */
1130 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1132 for_each_cpu_mask(i, tmp) {
1133 load = source_load(i, 0);
1135 if (load < min_load || (load == min_load && i == this_cpu)) {
1145 * sched_balance_self: balance the current task (running on cpu) in domains
1146 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1149 * Balance, ie. select the least loaded group.
1151 * Returns the target CPU number, or the same CPU if no balancing is needed.
1153 * preempt must be disabled.
1155 static int sched_balance_self(int cpu, int flag)
1157 struct task_struct *t = current;
1158 struct sched_domain *tmp, *sd = NULL;
1160 for_each_domain(cpu, tmp)
1161 if (tmp->flags & flag)
1166 struct sched_group *group;
1171 group = find_idlest_group(sd, t, cpu);
1175 new_cpu = find_idlest_cpu(group, t, cpu);
1176 if (new_cpu == -1 || new_cpu == cpu)
1179 /* Now try balancing at a lower domain level */
1183 weight = cpus_weight(span);
1184 for_each_domain(cpu, tmp) {
1185 if (weight <= cpus_weight(tmp->span))
1187 if (tmp->flags & flag)
1190 /* while loop will break here if sd == NULL */
1196 #endif /* CONFIG_SMP */
1199 * wake_idle() will wake a task on an idle cpu if task->cpu is
1200 * not idle and an idle cpu is available. The span of cpus to
1201 * search starts with cpus closest then further out as needed,
1202 * so we always favor a closer, idle cpu.
1204 * Returns the CPU we should wake onto.
1206 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1207 static int wake_idle(int cpu, task_t *p)
1210 struct sched_domain *sd;
1216 for_each_domain(cpu, sd) {
1217 if (sd->flags & SD_WAKE_IDLE) {
1218 cpus_and(tmp, sd->span, p->cpus_allowed);
1219 for_each_cpu_mask(i, tmp) {
1230 static inline int wake_idle(int cpu, task_t *p)
1237 * try_to_wake_up - wake up a thread
1238 * @p: the to-be-woken-up thread
1239 * @state: the mask of task states that can be woken
1240 * @sync: do a synchronous wakeup?
1242 * Put it on the run-queue if it's not already there. The "current"
1243 * thread is always on the run-queue (except when the actual
1244 * re-schedule is in progress), and as such you're allowed to do
1245 * the simpler "current->state = TASK_RUNNING" to mark yourself
1246 * runnable without the overhead of this.
1248 * returns failure only if the task is already active.
1250 static int try_to_wake_up(task_t *p, unsigned int state, int sync)
1252 int cpu, this_cpu, success = 0;
1253 unsigned long flags;
1257 unsigned long load, this_load;
1258 struct sched_domain *sd, *this_sd = NULL;
1262 rq = task_rq_lock(p, &flags);
1263 old_state = p->state;
1264 if (!(old_state & state))
1271 this_cpu = smp_processor_id();
1274 if (unlikely(task_running(rq, p)))
1279 schedstat_inc(rq, ttwu_cnt);
1280 if (cpu == this_cpu) {
1281 schedstat_inc(rq, ttwu_local);
1285 for_each_domain(this_cpu, sd) {
1286 if (cpu_isset(cpu, sd->span)) {
1287 schedstat_inc(sd, ttwu_wake_remote);
1293 if (p->last_waker_cpu != this_cpu)
1296 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1300 * Check for affine wakeup and passive balancing possibilities.
1303 int idx = this_sd->wake_idx;
1304 unsigned int imbalance;
1306 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1308 load = source_load(cpu, idx);
1309 this_load = target_load(this_cpu, idx);
1311 new_cpu = this_cpu; /* Wake to this CPU if we can */
1313 if (this_sd->flags & SD_WAKE_AFFINE) {
1314 unsigned long tl = this_load;
1316 * If sync wakeup then subtract the (maximum possible)
1317 * effect of the currently running task from the load
1318 * of the current CPU:
1321 tl -= SCHED_LOAD_SCALE;
1324 tl + target_load(cpu, idx) <= SCHED_LOAD_SCALE) ||
1325 100*(tl + SCHED_LOAD_SCALE) <= imbalance*load) {
1327 * This domain has SD_WAKE_AFFINE and
1328 * p is cache cold in this domain, and
1329 * there is no bad imbalance.
1331 schedstat_inc(this_sd, ttwu_move_affine);
1337 * Start passive balancing when half the imbalance_pct
1340 if (this_sd->flags & SD_WAKE_BALANCE) {
1341 if (imbalance*this_load <= 100*load) {
1342 schedstat_inc(this_sd, ttwu_move_balance);
1348 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1350 new_cpu = wake_idle(new_cpu, p);
1351 if (new_cpu != cpu) {
1352 set_task_cpu(p, new_cpu);
1353 task_rq_unlock(rq, &flags);
1354 /* might preempt at this point */
1355 rq = task_rq_lock(p, &flags);
1356 old_state = p->state;
1357 if (!(old_state & state))
1362 this_cpu = smp_processor_id();
1366 p->last_waker_cpu = this_cpu;
1369 #endif /* CONFIG_SMP */
1370 if (old_state == TASK_UNINTERRUPTIBLE) {
1371 rq->nr_uninterruptible--;
1373 * Tasks on involuntary sleep don't earn
1374 * sleep_avg beyond just interactive state.
1380 * Tasks that have marked their sleep as noninteractive get
1381 * woken up without updating their sleep average. (i.e. their
1382 * sleep is handled in a priority-neutral manner, no priority
1383 * boost and no penalty.)
1385 if (old_state & TASK_NONINTERACTIVE)
1386 __activate_task(p, rq);
1388 activate_task(p, rq, cpu == this_cpu);
1390 * Sync wakeups (i.e. those types of wakeups where the waker
1391 * has indicated that it will leave the CPU in short order)
1392 * don't trigger a preemption, if the woken up task will run on
1393 * this cpu. (in this case the 'I will reschedule' promise of
1394 * the waker guarantees that the freshly woken up task is going
1395 * to be considered on this CPU.)
1397 if (!sync || cpu != this_cpu) {
1398 if (TASK_PREEMPTS_CURR(p, rq))
1399 resched_task(rq->curr);
1404 p->state = TASK_RUNNING;
1406 task_rq_unlock(rq, &flags);
1411 int fastcall wake_up_process(task_t *p)
1413 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1414 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1417 EXPORT_SYMBOL(wake_up_process);
1419 int fastcall wake_up_state(task_t *p, unsigned int state)
1421 return try_to_wake_up(p, state, 0);
1425 * Perform scheduler related setup for a newly forked process p.
1426 * p is forked by current.
1428 void fastcall sched_fork(task_t *p, int clone_flags)
1430 int cpu = get_cpu();
1433 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1435 set_task_cpu(p, cpu);
1438 * We mark the process as running here, but have not actually
1439 * inserted it onto the runqueue yet. This guarantees that
1440 * nobody will actually run it, and a signal or other external
1441 * event cannot wake it up and insert it on the runqueue either.
1443 p->state = TASK_RUNNING;
1444 INIT_LIST_HEAD(&p->run_list);
1446 #ifdef CONFIG_SCHEDSTATS
1447 memset(&p->sched_info, 0, sizeof(p->sched_info));
1449 #if defined(CONFIG_SMP)
1450 p->last_waker_cpu = cpu;
1451 #if defined(__ARCH_WANT_UNLOCKED_CTXSW)
1455 #ifdef CONFIG_PREEMPT
1456 /* Want to start with kernel preemption disabled. */
1457 task_thread_info(p)->preempt_count = 1;
1460 * Share the timeslice between parent and child, thus the
1461 * total amount of pending timeslices in the system doesn't change,
1462 * resulting in more scheduling fairness.
1464 local_irq_disable();
1465 p->time_slice = (current->time_slice + 1) >> 1;
1467 * The remainder of the first timeslice might be recovered by
1468 * the parent if the child exits early enough.
1470 p->first_time_slice = 1;
1471 current->time_slice >>= 1;
1472 p->timestamp = sched_clock();
1473 if (unlikely(!current->time_slice)) {
1475 * This case is rare, it happens when the parent has only
1476 * a single jiffy left from its timeslice. Taking the
1477 * runqueue lock is not a problem.
1479 current->time_slice = 1;
1487 * wake_up_new_task - wake up a newly created task for the first time.
1489 * This function will do some initial scheduler statistics housekeeping
1490 * that must be done for every newly created context, then puts the task
1491 * on the runqueue and wakes it.
1493 void fastcall wake_up_new_task(task_t *p, unsigned long clone_flags)
1495 unsigned long flags;
1497 runqueue_t *rq, *this_rq;
1499 rq = task_rq_lock(p, &flags);
1500 BUG_ON(p->state != TASK_RUNNING);
1501 this_cpu = smp_processor_id();
1505 * We decrease the sleep average of forking parents
1506 * and children as well, to keep max-interactive tasks
1507 * from forking tasks that are max-interactive. The parent
1508 * (current) is done further down, under its lock.
1510 p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1511 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1513 p->prio = effective_prio(p);
1515 if (likely(cpu == this_cpu)) {
1516 if (!(clone_flags & CLONE_VM)) {
1518 * The VM isn't cloned, so we're in a good position to
1519 * do child-runs-first in anticipation of an exec. This
1520 * usually avoids a lot of COW overhead.
1522 if (unlikely(!current->array))
1523 __activate_task(p, rq);
1525 p->prio = current->prio;
1526 list_add_tail(&p->run_list, ¤t->run_list);
1527 p->array = current->array;
1528 p->array->nr_active++;
1529 inc_nr_running(p, rq);
1533 /* Run child last */
1534 __activate_task(p, rq);
1536 * We skip the following code due to cpu == this_cpu
1538 * task_rq_unlock(rq, &flags);
1539 * this_rq = task_rq_lock(current, &flags);
1543 this_rq = cpu_rq(this_cpu);
1546 * Not the local CPU - must adjust timestamp. This should
1547 * get optimised away in the !CONFIG_SMP case.
1549 p->timestamp = (p->timestamp - this_rq->timestamp_last_tick)
1550 + rq->timestamp_last_tick;
1551 __activate_task(p, rq);
1552 if (TASK_PREEMPTS_CURR(p, rq))
1553 resched_task(rq->curr);
1556 * Parent and child are on different CPUs, now get the
1557 * parent runqueue to update the parent's ->sleep_avg:
1559 task_rq_unlock(rq, &flags);
1560 this_rq = task_rq_lock(current, &flags);
1562 current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1563 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1564 task_rq_unlock(this_rq, &flags);
1568 * Potentially available exiting-child timeslices are
1569 * retrieved here - this way the parent does not get
1570 * penalized for creating too many threads.
1572 * (this cannot be used to 'generate' timeslices
1573 * artificially, because any timeslice recovered here
1574 * was given away by the parent in the first place.)
1576 void fastcall sched_exit(task_t *p)
1578 unsigned long flags;
1582 * If the child was a (relative-) CPU hog then decrease
1583 * the sleep_avg of the parent as well.
1585 rq = task_rq_lock(p->parent, &flags);
1586 if (p->first_time_slice && task_cpu(p) == task_cpu(p->parent)) {
1587 p->parent->time_slice += p->time_slice;
1588 if (unlikely(p->parent->time_slice > task_timeslice(p)))
1589 p->parent->time_slice = task_timeslice(p);
1591 if (p->sleep_avg < p->parent->sleep_avg)
1592 p->parent->sleep_avg = p->parent->sleep_avg /
1593 (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg /
1595 task_rq_unlock(rq, &flags);
1599 * prepare_task_switch - prepare to switch tasks
1600 * @rq: the runqueue preparing to switch
1601 * @next: the task we are going to switch to.
1603 * This is called with the rq lock held and interrupts off. It must
1604 * be paired with a subsequent finish_task_switch after the context
1607 * prepare_task_switch sets up locking and calls architecture specific
1610 static inline void prepare_task_switch(runqueue_t *rq, task_t *next)
1612 prepare_lock_switch(rq, next);
1613 prepare_arch_switch(next);
1617 * finish_task_switch - clean up after a task-switch
1618 * @rq: runqueue associated with task-switch
1619 * @prev: the thread we just switched away from.
1621 * finish_task_switch must be called after the context switch, paired
1622 * with a prepare_task_switch call before the context switch.
1623 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1624 * and do any other architecture-specific cleanup actions.
1626 * Note that we may have delayed dropping an mm in context_switch(). If
1627 * so, we finish that here outside of the runqueue lock. (Doing it
1628 * with the lock held can cause deadlocks; see schedule() for
1631 static inline void finish_task_switch(runqueue_t *rq, task_t *prev)
1632 __releases(rq->lock)
1634 struct mm_struct *mm = rq->prev_mm;
1635 unsigned long prev_task_flags;
1640 * A task struct has one reference for the use as "current".
1641 * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1642 * calls schedule one last time. The schedule call will never return,
1643 * and the scheduled task must drop that reference.
1644 * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1645 * still held, otherwise prev could be scheduled on another cpu, die
1646 * there before we look at prev->state, and then the reference would
1648 * Manfred Spraul <manfred@colorfullife.com>
1650 prev_task_flags = prev->flags;
1651 finish_arch_switch(prev);
1652 finish_lock_switch(rq, prev);
1655 if (unlikely(prev_task_flags & PF_DEAD))
1656 put_task_struct(prev);
1660 * schedule_tail - first thing a freshly forked thread must call.
1661 * @prev: the thread we just switched away from.
1663 asmlinkage void schedule_tail(task_t *prev)
1664 __releases(rq->lock)
1666 runqueue_t *rq = this_rq();
1667 finish_task_switch(rq, prev);
1668 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1669 /* In this case, finish_task_switch does not reenable preemption */
1672 if (current->set_child_tid)
1673 put_user(current->pid, current->set_child_tid);
1677 * context_switch - switch to the new MM and the new
1678 * thread's register state.
1681 task_t * context_switch(runqueue_t *rq, task_t *prev, task_t *next)
1683 struct mm_struct *mm = next->mm;
1684 struct mm_struct *oldmm = prev->active_mm;
1686 if (unlikely(!mm)) {
1687 next->active_mm = oldmm;
1688 atomic_inc(&oldmm->mm_count);
1689 enter_lazy_tlb(oldmm, next);
1691 switch_mm(oldmm, mm, next);
1693 if (unlikely(!prev->mm)) {
1694 prev->active_mm = NULL;
1695 WARN_ON(rq->prev_mm);
1696 rq->prev_mm = oldmm;
1699 /* Here we just switch the register state and the stack. */
1700 switch_to(prev, next, prev);
1706 * nr_running, nr_uninterruptible and nr_context_switches:
1708 * externally visible scheduler statistics: current number of runnable
1709 * threads, current number of uninterruptible-sleeping threads, total
1710 * number of context switches performed since bootup.
1712 unsigned long nr_running(void)
1714 unsigned long i, sum = 0;
1716 for_each_online_cpu(i)
1717 sum += cpu_rq(i)->nr_running;
1722 unsigned long nr_uninterruptible(void)
1724 unsigned long i, sum = 0;
1727 sum += cpu_rq(i)->nr_uninterruptible;
1730 * Since we read the counters lockless, it might be slightly
1731 * inaccurate. Do not allow it to go below zero though:
1733 if (unlikely((long)sum < 0))
1739 unsigned long long nr_context_switches(void)
1741 unsigned long long i, sum = 0;
1744 sum += cpu_rq(i)->nr_switches;
1749 unsigned long nr_iowait(void)
1751 unsigned long i, sum = 0;
1754 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1762 * double_rq_lock - safely lock two runqueues
1764 * Note this does not disable interrupts like task_rq_lock,
1765 * you need to do so manually before calling.
1767 static void double_rq_lock(runqueue_t *rq1, runqueue_t *rq2)
1768 __acquires(rq1->lock)
1769 __acquires(rq2->lock)
1772 spin_lock(&rq1->lock);
1773 __acquire(rq2->lock); /* Fake it out ;) */
1776 spin_lock(&rq1->lock);
1777 spin_lock(&rq2->lock);
1779 spin_lock(&rq2->lock);
1780 spin_lock(&rq1->lock);
1786 * double_rq_unlock - safely unlock two runqueues
1788 * Note this does not restore interrupts like task_rq_unlock,
1789 * you need to do so manually after calling.
1791 static void double_rq_unlock(runqueue_t *rq1, runqueue_t *rq2)
1792 __releases(rq1->lock)
1793 __releases(rq2->lock)
1795 spin_unlock(&rq1->lock);
1797 spin_unlock(&rq2->lock);
1799 __release(rq2->lock);
1803 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1805 static void double_lock_balance(runqueue_t *this_rq, runqueue_t *busiest)
1806 __releases(this_rq->lock)
1807 __acquires(busiest->lock)
1808 __acquires(this_rq->lock)
1810 if (unlikely(!spin_trylock(&busiest->lock))) {
1811 if (busiest < this_rq) {
1812 spin_unlock(&this_rq->lock);
1813 spin_lock(&busiest->lock);
1814 spin_lock(&this_rq->lock);
1816 spin_lock(&busiest->lock);
1821 * If dest_cpu is allowed for this process, migrate the task to it.
1822 * This is accomplished by forcing the cpu_allowed mask to only
1823 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
1824 * the cpu_allowed mask is restored.
1826 static void sched_migrate_task(task_t *p, int dest_cpu)
1828 migration_req_t req;
1830 unsigned long flags;
1832 rq = task_rq_lock(p, &flags);
1833 if (!cpu_isset(dest_cpu, p->cpus_allowed)
1834 || unlikely(cpu_is_offline(dest_cpu)))
1837 /* force the process onto the specified CPU */
1838 if (migrate_task(p, dest_cpu, &req)) {
1839 /* Need to wait for migration thread (might exit: take ref). */
1840 struct task_struct *mt = rq->migration_thread;
1841 get_task_struct(mt);
1842 task_rq_unlock(rq, &flags);
1843 wake_up_process(mt);
1844 put_task_struct(mt);
1845 wait_for_completion(&req.done);
1849 task_rq_unlock(rq, &flags);
1853 * sched_exec - execve() is a valuable balancing opportunity, because at
1854 * this point the task has the smallest effective memory and cache footprint.
1856 void sched_exec(void)
1858 int new_cpu, this_cpu = get_cpu();
1859 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
1861 if (new_cpu != this_cpu)
1862 sched_migrate_task(current, new_cpu);
1866 * pull_task - move a task from a remote runqueue to the local runqueue.
1867 * Both runqueues must be locked.
1870 void pull_task(runqueue_t *src_rq, prio_array_t *src_array, task_t *p,
1871 runqueue_t *this_rq, prio_array_t *this_array, int this_cpu)
1873 dequeue_task(p, src_array);
1874 dec_nr_running(p, src_rq);
1875 set_task_cpu(p, this_cpu);
1876 inc_nr_running(p, this_rq);
1877 enqueue_task(p, this_array);
1878 p->timestamp = (p->timestamp - src_rq->timestamp_last_tick)
1879 + this_rq->timestamp_last_tick;
1881 * Note that idle threads have a prio of MAX_PRIO, for this test
1882 * to be always true for them.
1884 if (TASK_PREEMPTS_CURR(p, this_rq))
1885 resched_task(this_rq->curr);
1889 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1892 int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu,
1893 struct sched_domain *sd, enum idle_type idle,
1897 * We do not migrate tasks that are:
1898 * 1) running (obviously), or
1899 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1900 * 3) are cache-hot on their current CPU.
1902 if (!cpu_isset(this_cpu, p->cpus_allowed))
1906 if (task_running(rq, p))
1910 * Aggressive migration if:
1911 * 1) task is cache cold, or
1912 * 2) too many balance attempts have failed.
1915 if (sd->nr_balance_failed > sd->cache_nice_tries)
1918 if (task_hot(p, rq->timestamp_last_tick, sd))
1924 * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq,
1925 * as part of a balancing operation within "domain". Returns the number of
1928 * Called with both runqueues locked.
1930 static int move_tasks(runqueue_t *this_rq, int this_cpu, runqueue_t *busiest,
1931 unsigned long max_nr_move, struct sched_domain *sd,
1932 enum idle_type idle, int *all_pinned)
1934 prio_array_t *array, *dst_array;
1935 struct list_head *head, *curr;
1936 int idx, pulled = 0, pinned = 0;
1939 if (max_nr_move == 0)
1945 * We first consider expired tasks. Those will likely not be
1946 * executed in the near future, and they are most likely to
1947 * be cache-cold, thus switching CPUs has the least effect
1950 if (busiest->expired->nr_active) {
1951 array = busiest->expired;
1952 dst_array = this_rq->expired;
1954 array = busiest->active;
1955 dst_array = this_rq->active;
1959 /* Start searching at priority 0: */
1963 idx = sched_find_first_bit(array->bitmap);
1965 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
1966 if (idx >= MAX_PRIO) {
1967 if (array == busiest->expired && busiest->active->nr_active) {
1968 array = busiest->active;
1969 dst_array = this_rq->active;
1975 head = array->queue + idx;
1978 tmp = list_entry(curr, task_t, run_list);
1982 if (!can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
1989 #ifdef CONFIG_SCHEDSTATS
1990 if (task_hot(tmp, busiest->timestamp_last_tick, sd))
1991 schedstat_inc(sd, lb_hot_gained[idle]);
1994 pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
1997 /* We only want to steal up to the prescribed number of tasks. */
1998 if (pulled < max_nr_move) {
2006 * Right now, this is the only place pull_task() is called,
2007 * so we can safely collect pull_task() stats here rather than
2008 * inside pull_task().
2010 schedstat_add(sd, lb_gained[idle], pulled);
2013 *all_pinned = pinned;
2018 * find_busiest_group finds and returns the busiest CPU group within the
2019 * domain. It calculates and returns the number of tasks which should be
2020 * moved to restore balance via the imbalance parameter.
2022 static struct sched_group *
2023 find_busiest_group(struct sched_domain *sd, int this_cpu,
2024 unsigned long *imbalance, enum idle_type idle, int *sd_idle)
2026 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2027 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2028 unsigned long max_pull;
2031 max_load = this_load = total_load = total_pwr = 0;
2032 if (idle == NOT_IDLE)
2033 load_idx = sd->busy_idx;
2034 else if (idle == NEWLY_IDLE)
2035 load_idx = sd->newidle_idx;
2037 load_idx = sd->idle_idx;
2044 local_group = cpu_isset(this_cpu, group->cpumask);
2046 /* Tally up the load of all CPUs in the group */
2049 for_each_cpu_mask(i, group->cpumask) {
2050 if (*sd_idle && !idle_cpu(i))
2053 /* Bias balancing toward cpus of our domain */
2055 load = __target_load(i, load_idx, idle);
2057 load = __source_load(i, load_idx, idle);
2062 total_load += avg_load;
2063 total_pwr += group->cpu_power;
2065 /* Adjust by relative CPU power of the group */
2066 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
2069 this_load = avg_load;
2071 } else if (avg_load > max_load) {
2072 max_load = avg_load;
2075 group = group->next;
2076 } while (group != sd->groups);
2078 if (!busiest || this_load >= max_load || max_load <= SCHED_LOAD_SCALE)
2081 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2083 if (this_load >= avg_load ||
2084 100*max_load <= sd->imbalance_pct*this_load)
2088 * We're trying to get all the cpus to the average_load, so we don't
2089 * want to push ourselves above the average load, nor do we wish to
2090 * reduce the max loaded cpu below the average load, as either of these
2091 * actions would just result in more rebalancing later, and ping-pong
2092 * tasks around. Thus we look for the minimum possible imbalance.
2093 * Negative imbalances (*we* are more loaded than anyone else) will
2094 * be counted as no imbalance for these purposes -- we can't fix that
2095 * by pulling tasks to us. Be careful of negative numbers as they'll
2096 * appear as very large values with unsigned longs.
2099 /* Don't want to pull so many tasks that a group would go idle */
2100 max_pull = min(max_load - avg_load, max_load - SCHED_LOAD_SCALE);
2102 /* How much load to actually move to equalise the imbalance */
2103 *imbalance = min(max_pull * busiest->cpu_power,
2104 (avg_load - this_load) * this->cpu_power)
2107 if (*imbalance < SCHED_LOAD_SCALE) {
2108 unsigned long pwr_now = 0, pwr_move = 0;
2111 if (max_load - this_load >= SCHED_LOAD_SCALE*2) {
2117 * OK, we don't have enough imbalance to justify moving tasks,
2118 * however we may be able to increase total CPU power used by
2122 pwr_now += busiest->cpu_power*min(SCHED_LOAD_SCALE, max_load);
2123 pwr_now += this->cpu_power*min(SCHED_LOAD_SCALE, this_load);
2124 pwr_now /= SCHED_LOAD_SCALE;
2126 /* Amount of load we'd subtract */
2127 tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/busiest->cpu_power;
2129 pwr_move += busiest->cpu_power*min(SCHED_LOAD_SCALE,
2132 /* Amount of load we'd add */
2133 if (max_load*busiest->cpu_power <
2134 SCHED_LOAD_SCALE*SCHED_LOAD_SCALE)
2135 tmp = max_load*busiest->cpu_power/this->cpu_power;
2137 tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/this->cpu_power;
2138 pwr_move += this->cpu_power*min(SCHED_LOAD_SCALE, this_load + tmp);
2139 pwr_move /= SCHED_LOAD_SCALE;
2141 /* Move if we gain throughput */
2142 if (pwr_move <= pwr_now)
2149 /* Get rid of the scaling factor, rounding down as we divide */
2150 *imbalance = *imbalance / SCHED_LOAD_SCALE;
2160 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2162 static runqueue_t *find_busiest_queue(struct sched_group *group,
2163 enum idle_type idle)
2165 unsigned long load, max_load = 0;
2166 runqueue_t *busiest = NULL;
2169 for_each_cpu_mask(i, group->cpumask) {
2170 load = __source_load(i, 0, idle);
2172 if (load > max_load) {
2174 busiest = cpu_rq(i);
2182 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2183 * so long as it is large enough.
2185 #define MAX_PINNED_INTERVAL 512
2188 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2189 * tasks if there is an imbalance.
2191 * Called with this_rq unlocked.
2193 static int load_balance(int this_cpu, runqueue_t *this_rq,
2194 struct sched_domain *sd, enum idle_type idle)
2196 struct sched_group *group;
2197 runqueue_t *busiest;
2198 unsigned long imbalance;
2199 int nr_moved, all_pinned = 0;
2200 int active_balance = 0;
2203 if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER)
2206 schedstat_inc(sd, lb_cnt[idle]);
2208 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle);
2210 schedstat_inc(sd, lb_nobusyg[idle]);
2214 busiest = find_busiest_queue(group, idle);
2216 schedstat_inc(sd, lb_nobusyq[idle]);
2220 BUG_ON(busiest == this_rq);
2222 schedstat_add(sd, lb_imbalance[idle], imbalance);
2225 if (busiest->nr_running > 1) {
2227 * Attempt to move tasks. If find_busiest_group has found
2228 * an imbalance but busiest->nr_running <= 1, the group is
2229 * still unbalanced. nr_moved simply stays zero, so it is
2230 * correctly treated as an imbalance.
2232 double_rq_lock(this_rq, busiest);
2233 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2234 imbalance, sd, idle, &all_pinned);
2235 double_rq_unlock(this_rq, busiest);
2237 /* All tasks on this runqueue were pinned by CPU affinity */
2238 if (unlikely(all_pinned))
2243 schedstat_inc(sd, lb_failed[idle]);
2244 sd->nr_balance_failed++;
2246 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2248 spin_lock(&busiest->lock);
2250 /* don't kick the migration_thread, if the curr
2251 * task on busiest cpu can't be moved to this_cpu
2253 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2254 spin_unlock(&busiest->lock);
2256 goto out_one_pinned;
2259 if (!busiest->active_balance) {
2260 busiest->active_balance = 1;
2261 busiest->push_cpu = this_cpu;
2264 spin_unlock(&busiest->lock);
2266 wake_up_process(busiest->migration_thread);
2269 * We've kicked active balancing, reset the failure
2272 sd->nr_balance_failed = sd->cache_nice_tries+1;
2275 sd->nr_balance_failed = 0;
2277 if (likely(!active_balance)) {
2278 /* We were unbalanced, so reset the balancing interval */
2279 sd->balance_interval = sd->min_interval;
2282 * If we've begun active balancing, start to back off. This
2283 * case may not be covered by the all_pinned logic if there
2284 * is only 1 task on the busy runqueue (because we don't call
2287 if (sd->balance_interval < sd->max_interval)
2288 sd->balance_interval *= 2;
2291 if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2296 schedstat_inc(sd, lb_balanced[idle]);
2298 sd->nr_balance_failed = 0;
2301 /* tune up the balancing interval */
2302 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2303 (sd->balance_interval < sd->max_interval))
2304 sd->balance_interval *= 2;
2306 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2312 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2313 * tasks if there is an imbalance.
2315 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2316 * this_rq is locked.
2318 static int load_balance_newidle(int this_cpu, runqueue_t *this_rq,
2319 struct sched_domain *sd)
2321 struct sched_group *group;
2322 runqueue_t *busiest = NULL;
2323 unsigned long imbalance;
2327 if (sd->flags & SD_SHARE_CPUPOWER)
2330 schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
2331 group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE, &sd_idle);
2333 schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]);
2337 busiest = find_busiest_queue(group, NEWLY_IDLE);
2339 schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]);
2343 BUG_ON(busiest == this_rq);
2345 schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
2348 if (busiest->nr_running > 1) {
2349 /* Attempt to move tasks */
2350 double_lock_balance(this_rq, busiest);
2351 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2352 imbalance, sd, NEWLY_IDLE, NULL);
2353 spin_unlock(&busiest->lock);
2357 schedstat_inc(sd, lb_failed[NEWLY_IDLE]);
2358 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2361 sd->nr_balance_failed = 0;
2366 schedstat_inc(sd, lb_balanced[NEWLY_IDLE]);
2367 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2369 sd->nr_balance_failed = 0;
2374 * idle_balance is called by schedule() if this_cpu is about to become
2375 * idle. Attempts to pull tasks from other CPUs.
2377 static inline void idle_balance(int this_cpu, runqueue_t *this_rq)
2379 struct sched_domain *sd;
2381 for_each_domain(this_cpu, sd) {
2382 if (sd->flags & SD_BALANCE_NEWIDLE) {
2383 if (load_balance_newidle(this_cpu, this_rq, sd)) {
2384 /* We've pulled tasks over so stop searching */
2392 * active_load_balance is run by migration threads. It pushes running tasks
2393 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2394 * running on each physical CPU where possible, and avoids physical /
2395 * logical imbalances.
2397 * Called with busiest_rq locked.
2399 static void active_load_balance(runqueue_t *busiest_rq, int busiest_cpu)
2401 struct sched_domain *sd;
2402 runqueue_t *target_rq;
2403 int target_cpu = busiest_rq->push_cpu;
2405 if (busiest_rq->nr_running <= 1)
2406 /* no task to move */
2409 target_rq = cpu_rq(target_cpu);
2412 * This condition is "impossible", if it occurs
2413 * we need to fix it. Originally reported by
2414 * Bjorn Helgaas on a 128-cpu setup.
2416 BUG_ON(busiest_rq == target_rq);
2418 /* move a task from busiest_rq to target_rq */
2419 double_lock_balance(busiest_rq, target_rq);
2421 /* Search for an sd spanning us and the target CPU. */
2422 for_each_domain(target_cpu, sd)
2423 if ((sd->flags & SD_LOAD_BALANCE) &&
2424 cpu_isset(busiest_cpu, sd->span))
2427 if (unlikely(sd == NULL))
2430 schedstat_inc(sd, alb_cnt);
2432 if (move_tasks(target_rq, target_cpu, busiest_rq, 1, sd, SCHED_IDLE, NULL))
2433 schedstat_inc(sd, alb_pushed);
2435 schedstat_inc(sd, alb_failed);
2437 spin_unlock(&target_rq->lock);
2441 * rebalance_tick will get called every timer tick, on every CPU.
2443 * It checks each scheduling domain to see if it is due to be balanced,
2444 * and initiates a balancing operation if so.
2446 * Balancing parameters are set up in arch_init_sched_domains.
2449 /* Don't have all balancing operations going off at once */
2450 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2452 static void rebalance_tick(int this_cpu, runqueue_t *this_rq,
2453 enum idle_type idle)
2455 unsigned long old_load, this_load;
2456 unsigned long j = jiffies + CPU_OFFSET(this_cpu);
2457 struct sched_domain *sd;
2460 this_load = this_rq->nr_running * SCHED_LOAD_SCALE;
2461 /* Update our load */
2462 for (i = 0; i < 3; i++) {
2463 unsigned long new_load = this_load;
2465 old_load = this_rq->cpu_load[i];
2467 * Round up the averaging division if load is increasing. This
2468 * prevents us from getting stuck on 9 if the load is 10, for
2471 if (new_load > old_load)
2472 new_load += scale-1;
2473 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale;
2476 for_each_domain(this_cpu, sd) {
2477 unsigned long interval;
2479 if (!(sd->flags & SD_LOAD_BALANCE))
2482 interval = sd->balance_interval;
2483 if (idle != SCHED_IDLE)
2484 interval *= sd->busy_factor;
2486 /* scale ms to jiffies */
2487 interval = msecs_to_jiffies(interval);
2488 if (unlikely(!interval))
2491 if (j - sd->last_balance >= interval) {
2492 if (load_balance(this_cpu, this_rq, sd, idle)) {
2494 * We've pulled tasks over so either we're no
2495 * longer idle, or one of our SMT siblings is
2500 sd->last_balance += interval;
2506 * on UP we do not need to balance between CPUs:
2508 static inline void rebalance_tick(int cpu, runqueue_t *rq, enum idle_type idle)
2511 static inline void idle_balance(int cpu, runqueue_t *rq)
2516 static inline int wake_priority_sleeper(runqueue_t *rq)
2519 #ifdef CONFIG_SCHED_SMT
2520 spin_lock(&rq->lock);
2522 * If an SMT sibling task has been put to sleep for priority
2523 * reasons reschedule the idle task to see if it can now run.
2525 if (rq->nr_running) {
2526 resched_task(rq->idle);
2529 spin_unlock(&rq->lock);
2534 DEFINE_PER_CPU(struct kernel_stat, kstat);
2536 EXPORT_PER_CPU_SYMBOL(kstat);
2539 * This is called on clock ticks and on context switches.
2540 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2542 static inline void update_cpu_clock(task_t *p, runqueue_t *rq,
2543 unsigned long long now)
2545 unsigned long long last = max(p->timestamp, rq->timestamp_last_tick);
2546 p->sched_time += now - last;
2550 * Return current->sched_time plus any more ns on the sched_clock
2551 * that have not yet been banked.
2553 unsigned long long current_sched_time(const task_t *tsk)
2555 unsigned long long ns;
2556 unsigned long flags;
2557 local_irq_save(flags);
2558 ns = max(tsk->timestamp, task_rq(tsk)->timestamp_last_tick);
2559 ns = tsk->sched_time + (sched_clock() - ns);
2560 local_irq_restore(flags);
2565 * We place interactive tasks back into the active array, if possible.
2567 * To guarantee that this does not starve expired tasks we ignore the
2568 * interactivity of a task if the first expired task had to wait more
2569 * than a 'reasonable' amount of time. This deadline timeout is
2570 * load-dependent, as the frequency of array switched decreases with
2571 * increasing number of running tasks. We also ignore the interactivity
2572 * if a better static_prio task has expired:
2574 #define EXPIRED_STARVING(rq) \
2575 ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2576 (jiffies - (rq)->expired_timestamp >= \
2577 STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2578 ((rq)->curr->static_prio > (rq)->best_expired_prio))
2581 * Account user cpu time to a process.
2582 * @p: the process that the cpu time gets accounted to
2583 * @hardirq_offset: the offset to subtract from hardirq_count()
2584 * @cputime: the cpu time spent in user space since the last update
2586 void account_user_time(struct task_struct *p, cputime_t cputime)
2588 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2591 p->utime = cputime_add(p->utime, cputime);
2593 /* Add user time to cpustat. */
2594 tmp = cputime_to_cputime64(cputime);
2595 if (TASK_NICE(p) > 0)
2596 cpustat->nice = cputime64_add(cpustat->nice, tmp);
2598 cpustat->user = cputime64_add(cpustat->user, tmp);
2602 * Account system cpu time to a process.
2603 * @p: the process that the cpu time gets accounted to
2604 * @hardirq_offset: the offset to subtract from hardirq_count()
2605 * @cputime: the cpu time spent in kernel space since the last update
2607 void account_system_time(struct task_struct *p, int hardirq_offset,
2610 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2611 runqueue_t *rq = this_rq();
2614 p->stime = cputime_add(p->stime, cputime);
2616 /* Add system time to cpustat. */
2617 tmp = cputime_to_cputime64(cputime);
2618 if (hardirq_count() - hardirq_offset)
2619 cpustat->irq = cputime64_add(cpustat->irq, tmp);
2620 else if (softirq_count())
2621 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
2622 else if (p != rq->idle)
2623 cpustat->system = cputime64_add(cpustat->system, tmp);
2624 else if (atomic_read(&rq->nr_iowait) > 0)
2625 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2627 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2628 /* Account for system time used */
2629 acct_update_integrals(p);
2633 * Account for involuntary wait time.
2634 * @p: the process from which the cpu time has been stolen
2635 * @steal: the cpu time spent in involuntary wait
2637 void account_steal_time(struct task_struct *p, cputime_t steal)
2639 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2640 cputime64_t tmp = cputime_to_cputime64(steal);
2641 runqueue_t *rq = this_rq();
2643 if (p == rq->idle) {
2644 p->stime = cputime_add(p->stime, steal);
2645 if (atomic_read(&rq->nr_iowait) > 0)
2646 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2648 cpustat->idle = cputime64_add(cpustat->idle, tmp);
2650 cpustat->steal = cputime64_add(cpustat->steal, tmp);
2654 * This function gets called by the timer code, with HZ frequency.
2655 * We call it with interrupts disabled.
2657 * It also gets called by the fork code, when changing the parent's
2660 void scheduler_tick(void)
2662 int cpu = smp_processor_id();
2663 runqueue_t *rq = this_rq();
2664 task_t *p = current;
2665 unsigned long long now = sched_clock();
2667 update_cpu_clock(p, rq, now);
2669 rq->timestamp_last_tick = now;
2671 if (p == rq->idle) {
2672 if (wake_priority_sleeper(rq))
2674 rebalance_tick(cpu, rq, SCHED_IDLE);
2678 /* Task might have expired already, but not scheduled off yet */
2679 if (p->array != rq->active) {
2680 set_tsk_need_resched(p);
2683 spin_lock(&rq->lock);
2685 * The task was running during this tick - update the
2686 * time slice counter. Note: we do not update a thread's
2687 * priority until it either goes to sleep or uses up its
2688 * timeslice. This makes it possible for interactive tasks
2689 * to use up their timeslices at their highest priority levels.
2693 * RR tasks need a special form of timeslice management.
2694 * FIFO tasks have no timeslices.
2696 if ((p->policy == SCHED_RR) && !--p->time_slice) {
2697 p->time_slice = task_timeslice(p);
2698 p->first_time_slice = 0;
2699 set_tsk_need_resched(p);
2701 /* put it at the end of the queue: */
2702 requeue_task(p, rq->active);
2706 if (!--p->time_slice) {
2707 dequeue_task(p, rq->active);
2708 set_tsk_need_resched(p);
2709 p->prio = effective_prio(p);
2710 p->time_slice = task_timeslice(p);
2711 p->first_time_slice = 0;
2713 if (!rq->expired_timestamp)
2714 rq->expired_timestamp = jiffies;
2715 if (!TASK_INTERACTIVE(p) || EXPIRED_STARVING(rq)) {
2716 enqueue_task(p, rq->expired);
2717 if (p->static_prio < rq->best_expired_prio)
2718 rq->best_expired_prio = p->static_prio;
2720 enqueue_task(p, rq->active);
2723 * Prevent a too long timeslice allowing a task to monopolize
2724 * the CPU. We do this by splitting up the timeslice into
2727 * Note: this does not mean the task's timeslices expire or
2728 * get lost in any way, they just might be preempted by
2729 * another task of equal priority. (one with higher
2730 * priority would have preempted this task already.) We
2731 * requeue this task to the end of the list on this priority
2732 * level, which is in essence a round-robin of tasks with
2735 * This only applies to tasks in the interactive
2736 * delta range with at least TIMESLICE_GRANULARITY to requeue.
2738 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
2739 p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
2740 (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
2741 (p->array == rq->active)) {
2743 requeue_task(p, rq->active);
2744 set_tsk_need_resched(p);
2748 spin_unlock(&rq->lock);
2750 rebalance_tick(cpu, rq, NOT_IDLE);
2753 #ifdef CONFIG_SCHED_SMT
2754 static inline void wakeup_busy_runqueue(runqueue_t *rq)
2756 /* If an SMT runqueue is sleeping due to priority reasons wake it up */
2757 if (rq->curr == rq->idle && rq->nr_running)
2758 resched_task(rq->idle);
2761 static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2763 struct sched_domain *tmp, *sd = NULL;
2764 cpumask_t sibling_map;
2767 for_each_domain(this_cpu, tmp)
2768 if (tmp->flags & SD_SHARE_CPUPOWER)
2775 * Unlock the current runqueue because we have to lock in
2776 * CPU order to avoid deadlocks. Caller knows that we might
2777 * unlock. We keep IRQs disabled.
2779 spin_unlock(&this_rq->lock);
2781 sibling_map = sd->span;
2783 for_each_cpu_mask(i, sibling_map)
2784 spin_lock(&cpu_rq(i)->lock);
2786 * We clear this CPU from the mask. This both simplifies the
2787 * inner loop and keps this_rq locked when we exit:
2789 cpu_clear(this_cpu, sibling_map);
2791 for_each_cpu_mask(i, sibling_map) {
2792 runqueue_t *smt_rq = cpu_rq(i);
2794 wakeup_busy_runqueue(smt_rq);
2797 for_each_cpu_mask(i, sibling_map)
2798 spin_unlock(&cpu_rq(i)->lock);
2800 * We exit with this_cpu's rq still held and IRQs
2806 * number of 'lost' timeslices this task wont be able to fully
2807 * utilize, if another task runs on a sibling. This models the
2808 * slowdown effect of other tasks running on siblings:
2810 static inline unsigned long smt_slice(task_t *p, struct sched_domain *sd)
2812 return p->time_slice * (100 - sd->per_cpu_gain) / 100;
2815 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2817 struct sched_domain *tmp, *sd = NULL;
2818 cpumask_t sibling_map;
2819 prio_array_t *array;
2823 for_each_domain(this_cpu, tmp)
2824 if (tmp->flags & SD_SHARE_CPUPOWER)
2831 * The same locking rules and details apply as for
2832 * wake_sleeping_dependent():
2834 spin_unlock(&this_rq->lock);
2835 sibling_map = sd->span;
2836 for_each_cpu_mask(i, sibling_map)
2837 spin_lock(&cpu_rq(i)->lock);
2838 cpu_clear(this_cpu, sibling_map);
2841 * Establish next task to be run - it might have gone away because
2842 * we released the runqueue lock above:
2844 if (!this_rq->nr_running)
2846 array = this_rq->active;
2847 if (!array->nr_active)
2848 array = this_rq->expired;
2849 BUG_ON(!array->nr_active);
2851 p = list_entry(array->queue[sched_find_first_bit(array->bitmap)].next,
2854 for_each_cpu_mask(i, sibling_map) {
2855 runqueue_t *smt_rq = cpu_rq(i);
2856 task_t *smt_curr = smt_rq->curr;
2858 /* Kernel threads do not participate in dependent sleeping */
2859 if (!p->mm || !smt_curr->mm || rt_task(p))
2860 goto check_smt_task;
2863 * If a user task with lower static priority than the
2864 * running task on the SMT sibling is trying to schedule,
2865 * delay it till there is proportionately less timeslice
2866 * left of the sibling task to prevent a lower priority
2867 * task from using an unfair proportion of the
2868 * physical cpu's resources. -ck
2870 if (rt_task(smt_curr)) {
2872 * With real time tasks we run non-rt tasks only
2873 * per_cpu_gain% of the time.
2875 if ((jiffies % DEF_TIMESLICE) >
2876 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
2879 if (smt_curr->static_prio < p->static_prio &&
2880 !TASK_PREEMPTS_CURR(p, smt_rq) &&
2881 smt_slice(smt_curr, sd) > task_timeslice(p))
2885 if ((!smt_curr->mm && smt_curr != smt_rq->idle) ||
2889 wakeup_busy_runqueue(smt_rq);
2894 * Reschedule a lower priority task on the SMT sibling for
2895 * it to be put to sleep, or wake it up if it has been put to
2896 * sleep for priority reasons to see if it should run now.
2899 if ((jiffies % DEF_TIMESLICE) >
2900 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
2901 resched_task(smt_curr);
2903 if (TASK_PREEMPTS_CURR(p, smt_rq) &&
2904 smt_slice(p, sd) > task_timeslice(smt_curr))
2905 resched_task(smt_curr);
2907 wakeup_busy_runqueue(smt_rq);
2911 for_each_cpu_mask(i, sibling_map)
2912 spin_unlock(&cpu_rq(i)->lock);
2916 static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
2920 static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
2926 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
2928 void fastcall add_preempt_count(int val)
2933 BUG_ON((preempt_count() < 0));
2934 preempt_count() += val;
2936 * Spinlock count overflowing soon?
2938 BUG_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10);
2940 EXPORT_SYMBOL(add_preempt_count);
2942 void fastcall sub_preempt_count(int val)
2947 BUG_ON(val > preempt_count());
2949 * Is the spinlock portion underflowing?
2951 BUG_ON((val < PREEMPT_MASK) && !(preempt_count() & PREEMPT_MASK));
2952 preempt_count() -= val;
2954 EXPORT_SYMBOL(sub_preempt_count);
2959 * schedule() is the main scheduler function.
2961 asmlinkage void __sched schedule(void)
2964 task_t *prev, *next;
2966 prio_array_t *array;
2967 struct list_head *queue;
2968 unsigned long long now;
2969 unsigned long run_time;
2970 int cpu, idx, new_prio;
2973 * Test if we are atomic. Since do_exit() needs to call into
2974 * schedule() atomically, we ignore that path for now.
2975 * Otherwise, whine if we are scheduling when we should not be.
2977 if (likely(!current->exit_state)) {
2978 if (unlikely(in_atomic())) {
2979 printk(KERN_ERR "scheduling while atomic: "
2981 current->comm, preempt_count(), current->pid);
2985 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2990 release_kernel_lock(prev);
2991 need_resched_nonpreemptible:
2995 * The idle thread is not allowed to schedule!
2996 * Remove this check after it has been exercised a bit.
2998 if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
2999 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
3003 schedstat_inc(rq, sched_cnt);
3004 now = sched_clock();
3005 if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
3006 run_time = now - prev->timestamp;
3007 if (unlikely((long long)(now - prev->timestamp) < 0))
3010 run_time = NS_MAX_SLEEP_AVG;
3013 * Tasks charged proportionately less run_time at high sleep_avg to
3014 * delay them losing their interactive status
3016 run_time /= (CURRENT_BONUS(prev) ? : 1);
3018 spin_lock_irq(&rq->lock);
3020 if (unlikely(prev->flags & PF_DEAD))
3021 prev->state = EXIT_DEAD;
3023 switch_count = &prev->nivcsw;
3024 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3025 switch_count = &prev->nvcsw;
3026 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3027 unlikely(signal_pending(prev))))
3028 prev->state = TASK_RUNNING;
3030 if (prev->state == TASK_UNINTERRUPTIBLE)
3031 rq->nr_uninterruptible++;
3032 deactivate_task(prev, rq);
3036 cpu = smp_processor_id();
3037 if (unlikely(!rq->nr_running)) {
3039 idle_balance(cpu, rq);
3040 if (!rq->nr_running) {
3042 rq->expired_timestamp = 0;
3043 wake_sleeping_dependent(cpu, rq);
3045 * wake_sleeping_dependent() might have released
3046 * the runqueue, so break out if we got new
3049 if (!rq->nr_running)
3053 if (dependent_sleeper(cpu, rq)) {
3058 * dependent_sleeper() releases and reacquires the runqueue
3059 * lock, hence go into the idle loop if the rq went
3062 if (unlikely(!rq->nr_running))
3067 if (unlikely(!array->nr_active)) {
3069 * Switch the active and expired arrays.
3071 schedstat_inc(rq, sched_switch);
3072 rq->active = rq->expired;
3073 rq->expired = array;
3075 rq->expired_timestamp = 0;
3076 rq->best_expired_prio = MAX_PRIO;
3079 idx = sched_find_first_bit(array->bitmap);
3080 queue = array->queue + idx;
3081 next = list_entry(queue->next, task_t, run_list);
3083 if (!rt_task(next) && next->activated > 0) {
3084 unsigned long long delta = now - next->timestamp;
3085 if (unlikely((long long)(now - next->timestamp) < 0))
3088 if (next->activated == 1)
3089 delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
3091 array = next->array;
3092 new_prio = recalc_task_prio(next, next->timestamp + delta);
3094 if (unlikely(next->prio != new_prio)) {
3095 dequeue_task(next, array);
3096 next->prio = new_prio;
3097 enqueue_task(next, array);
3099 requeue_task(next, array);
3101 next->activated = 0;
3103 if (next == rq->idle)
3104 schedstat_inc(rq, sched_goidle);
3106 prefetch_stack(next);
3107 clear_tsk_need_resched(prev);
3108 rcu_qsctr_inc(task_cpu(prev));
3110 update_cpu_clock(prev, rq, now);
3112 prev->sleep_avg -= run_time;
3113 if ((long)prev->sleep_avg <= 0)
3114 prev->sleep_avg = 0;
3115 prev->timestamp = prev->last_ran = now;
3117 sched_info_switch(prev, next);
3118 if (likely(prev != next)) {
3119 next->timestamp = now;
3124 prepare_task_switch(rq, next);
3125 prev = context_switch(rq, prev, next);
3128 * this_rq must be evaluated again because prev may have moved
3129 * CPUs since it called schedule(), thus the 'rq' on its stack
3130 * frame will be invalid.
3132 finish_task_switch(this_rq(), prev);
3134 spin_unlock_irq(&rq->lock);
3137 if (unlikely(reacquire_kernel_lock(prev) < 0))
3138 goto need_resched_nonpreemptible;
3139 preempt_enable_no_resched();
3140 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3144 EXPORT_SYMBOL(schedule);
3146 #ifdef CONFIG_PREEMPT
3148 * this is is the entry point to schedule() from in-kernel preemption
3149 * off of preempt_enable. Kernel preemptions off return from interrupt
3150 * occur there and call schedule directly.
3152 asmlinkage void __sched preempt_schedule(void)
3154 struct thread_info *ti = current_thread_info();
3155 #ifdef CONFIG_PREEMPT_BKL
3156 struct task_struct *task = current;
3157 int saved_lock_depth;
3160 * If there is a non-zero preempt_count or interrupts are disabled,
3161 * we do not want to preempt the current task. Just return..
3163 if (unlikely(ti->preempt_count || irqs_disabled()))
3167 add_preempt_count(PREEMPT_ACTIVE);
3169 * We keep the big kernel semaphore locked, but we
3170 * clear ->lock_depth so that schedule() doesnt
3171 * auto-release the semaphore:
3173 #ifdef CONFIG_PREEMPT_BKL
3174 saved_lock_depth = task->lock_depth;
3175 task->lock_depth = -1;
3178 #ifdef CONFIG_PREEMPT_BKL
3179 task->lock_depth = saved_lock_depth;
3181 sub_preempt_count(PREEMPT_ACTIVE);
3183 /* we could miss a preemption opportunity between schedule and now */
3185 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3189 EXPORT_SYMBOL(preempt_schedule);
3192 * this is is the entry point to schedule() from kernel preemption
3193 * off of irq context.
3194 * Note, that this is called and return with irqs disabled. This will
3195 * protect us against recursive calling from irq.
3197 asmlinkage void __sched preempt_schedule_irq(void)
3199 struct thread_info *ti = current_thread_info();
3200 #ifdef CONFIG_PREEMPT_BKL
3201 struct task_struct *task = current;
3202 int saved_lock_depth;
3204 /* Catch callers which need to be fixed*/
3205 BUG_ON(ti->preempt_count || !irqs_disabled());
3208 add_preempt_count(PREEMPT_ACTIVE);
3210 * We keep the big kernel semaphore locked, but we
3211 * clear ->lock_depth so that schedule() doesnt
3212 * auto-release the semaphore:
3214 #ifdef CONFIG_PREEMPT_BKL
3215 saved_lock_depth = task->lock_depth;
3216 task->lock_depth = -1;
3220 local_irq_disable();
3221 #ifdef CONFIG_PREEMPT_BKL
3222 task->lock_depth = saved_lock_depth;
3224 sub_preempt_count(PREEMPT_ACTIVE);
3226 /* we could miss a preemption opportunity between schedule and now */
3228 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3232 #endif /* CONFIG_PREEMPT */
3234 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3237 task_t *p = curr->private;
3238 return try_to_wake_up(p, mode, sync);
3241 EXPORT_SYMBOL(default_wake_function);
3244 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3245 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3246 * number) then we wake all the non-exclusive tasks and one exclusive task.
3248 * There are circumstances in which we can try to wake a task which has already
3249 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3250 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3252 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3253 int nr_exclusive, int sync, void *key)
3255 struct list_head *tmp, *next;
3257 list_for_each_safe(tmp, next, &q->task_list) {
3260 curr = list_entry(tmp, wait_queue_t, task_list);
3261 flags = curr->flags;
3262 if (curr->func(curr, mode, sync, key) &&
3263 (flags & WQ_FLAG_EXCLUSIVE) &&
3270 * __wake_up - wake up threads blocked on a waitqueue.
3272 * @mode: which threads
3273 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3274 * @key: is directly passed to the wakeup function
3276 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3277 int nr_exclusive, void *key)
3279 unsigned long flags;
3281 spin_lock_irqsave(&q->lock, flags);
3282 __wake_up_common(q, mode, nr_exclusive, 0, key);
3283 spin_unlock_irqrestore(&q->lock, flags);
3286 EXPORT_SYMBOL(__wake_up);
3289 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3291 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3293 __wake_up_common(q, mode, 1, 0, NULL);
3297 * __wake_up_sync - wake up threads blocked on a waitqueue.
3299 * @mode: which threads
3300 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3302 * The sync wakeup differs that the waker knows that it will schedule
3303 * away soon, so while the target thread will be woken up, it will not
3304 * be migrated to another CPU - ie. the two threads are 'synchronized'
3305 * with each other. This can prevent needless bouncing between CPUs.
3307 * On UP it can prevent extra preemption.
3310 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3312 unsigned long flags;
3318 if (unlikely(!nr_exclusive))
3321 spin_lock_irqsave(&q->lock, flags);
3322 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3323 spin_unlock_irqrestore(&q->lock, flags);
3325 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3327 void fastcall complete(struct completion *x)
3329 unsigned long flags;
3331 spin_lock_irqsave(&x->wait.lock, flags);
3333 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3335 spin_unlock_irqrestore(&x->wait.lock, flags);
3337 EXPORT_SYMBOL(complete);
3339 void fastcall complete_all(struct completion *x)
3341 unsigned long flags;
3343 spin_lock_irqsave(&x->wait.lock, flags);
3344 x->done += UINT_MAX/2;
3345 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3347 spin_unlock_irqrestore(&x->wait.lock, flags);
3349 EXPORT_SYMBOL(complete_all);
3351 void fastcall __sched wait_for_completion(struct completion *x)
3354 spin_lock_irq(&x->wait.lock);
3356 DECLARE_WAITQUEUE(wait, current);
3358 wait.flags |= WQ_FLAG_EXCLUSIVE;
3359 __add_wait_queue_tail(&x->wait, &wait);
3361 __set_current_state(TASK_UNINTERRUPTIBLE);
3362 spin_unlock_irq(&x->wait.lock);
3364 spin_lock_irq(&x->wait.lock);
3366 __remove_wait_queue(&x->wait, &wait);
3369 spin_unlock_irq(&x->wait.lock);
3371 EXPORT_SYMBOL(wait_for_completion);
3373 unsigned long fastcall __sched
3374 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3378 spin_lock_irq(&x->wait.lock);
3380 DECLARE_WAITQUEUE(wait, current);
3382 wait.flags |= WQ_FLAG_EXCLUSIVE;
3383 __add_wait_queue_tail(&x->wait, &wait);
3385 __set_current_state(TASK_UNINTERRUPTIBLE);
3386 spin_unlock_irq(&x->wait.lock);
3387 timeout = schedule_timeout(timeout);
3388 spin_lock_irq(&x->wait.lock);
3390 __remove_wait_queue(&x->wait, &wait);
3394 __remove_wait_queue(&x->wait, &wait);
3398 spin_unlock_irq(&x->wait.lock);
3401 EXPORT_SYMBOL(wait_for_completion_timeout);
3403 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3409 spin_lock_irq(&x->wait.lock);
3411 DECLARE_WAITQUEUE(wait, current);
3413 wait.flags |= WQ_FLAG_EXCLUSIVE;
3414 __add_wait_queue_tail(&x->wait, &wait);
3416 if (signal_pending(current)) {
3418 __remove_wait_queue(&x->wait, &wait);
3421 __set_current_state(TASK_INTERRUPTIBLE);
3422 spin_unlock_irq(&x->wait.lock);
3424 spin_lock_irq(&x->wait.lock);
3426 __remove_wait_queue(&x->wait, &wait);
3430 spin_unlock_irq(&x->wait.lock);
3434 EXPORT_SYMBOL(wait_for_completion_interruptible);
3436 unsigned long fastcall __sched
3437 wait_for_completion_interruptible_timeout(struct completion *x,
3438 unsigned long timeout)
3442 spin_lock_irq(&x->wait.lock);
3444 DECLARE_WAITQUEUE(wait, current);
3446 wait.flags |= WQ_FLAG_EXCLUSIVE;
3447 __add_wait_queue_tail(&x->wait, &wait);
3449 if (signal_pending(current)) {
3450 timeout = -ERESTARTSYS;
3451 __remove_wait_queue(&x->wait, &wait);
3454 __set_current_state(TASK_INTERRUPTIBLE);
3455 spin_unlock_irq(&x->wait.lock);
3456 timeout = schedule_timeout(timeout);
3457 spin_lock_irq(&x->wait.lock);
3459 __remove_wait_queue(&x->wait, &wait);
3463 __remove_wait_queue(&x->wait, &wait);
3467 spin_unlock_irq(&x->wait.lock);
3470 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3473 #define SLEEP_ON_VAR \
3474 unsigned long flags; \
3475 wait_queue_t wait; \
3476 init_waitqueue_entry(&wait, current);
3478 #define SLEEP_ON_HEAD \
3479 spin_lock_irqsave(&q->lock,flags); \
3480 __add_wait_queue(q, &wait); \
3481 spin_unlock(&q->lock);
3483 #define SLEEP_ON_TAIL \
3484 spin_lock_irq(&q->lock); \
3485 __remove_wait_queue(q, &wait); \
3486 spin_unlock_irqrestore(&q->lock, flags);
3488 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
3492 current->state = TASK_INTERRUPTIBLE;
3499 EXPORT_SYMBOL(interruptible_sleep_on);
3501 long fastcall __sched
3502 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3506 current->state = TASK_INTERRUPTIBLE;
3509 timeout = schedule_timeout(timeout);
3515 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3517 void fastcall __sched sleep_on(wait_queue_head_t *q)
3521 current->state = TASK_UNINTERRUPTIBLE;
3528 EXPORT_SYMBOL(sleep_on);
3530 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3534 current->state = TASK_UNINTERRUPTIBLE;
3537 timeout = schedule_timeout(timeout);
3543 EXPORT_SYMBOL(sleep_on_timeout);
3545 void set_user_nice(task_t *p, long nice)
3547 unsigned long flags;
3548 prio_array_t *array;
3550 int old_prio, new_prio, delta;
3552 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3555 * We have to be careful, if called from sys_setpriority(),
3556 * the task might be in the middle of scheduling on another CPU.
3558 rq = task_rq_lock(p, &flags);
3560 * The RT priorities are set via sched_setscheduler(), but we still
3561 * allow the 'normal' nice value to be set - but as expected
3562 * it wont have any effect on scheduling until the task is
3566 p->static_prio = NICE_TO_PRIO(nice);
3571 dequeue_task(p, array);
3572 dec_prio_bias(rq, p->static_prio);
3576 new_prio = NICE_TO_PRIO(nice);
3577 delta = new_prio - old_prio;
3578 p->static_prio = NICE_TO_PRIO(nice);
3582 enqueue_task(p, array);
3583 inc_prio_bias(rq, p->static_prio);
3585 * If the task increased its priority or is running and
3586 * lowered its priority, then reschedule its CPU:
3588 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3589 resched_task(rq->curr);
3592 task_rq_unlock(rq, &flags);
3595 EXPORT_SYMBOL(set_user_nice);
3598 * can_nice - check if a task can reduce its nice value
3602 int can_nice(const task_t *p, const int nice)
3604 /* convert nice value [19,-20] to rlimit style value [1,40] */
3605 int nice_rlim = 20 - nice;
3606 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
3607 capable(CAP_SYS_NICE));
3610 #ifdef __ARCH_WANT_SYS_NICE
3613 * sys_nice - change the priority of the current process.
3614 * @increment: priority increment
3616 * sys_setpriority is a more generic, but much slower function that
3617 * does similar things.
3619 asmlinkage long sys_nice(int increment)
3625 * Setpriority might change our priority at the same moment.
3626 * We don't have to worry. Conceptually one call occurs first
3627 * and we have a single winner.
3629 if (increment < -40)
3634 nice = PRIO_TO_NICE(current->static_prio) + increment;
3640 if (increment < 0 && !can_nice(current, nice))
3643 retval = security_task_setnice(current, nice);
3647 set_user_nice(current, nice);
3654 * task_prio - return the priority value of a given task.
3655 * @p: the task in question.
3657 * This is the priority value as seen by users in /proc.
3658 * RT tasks are offset by -200. Normal tasks are centered
3659 * around 0, value goes from -16 to +15.
3661 int task_prio(const task_t *p)
3663 return p->prio - MAX_RT_PRIO;
3667 * task_nice - return the nice value of a given task.
3668 * @p: the task in question.
3670 int task_nice(const task_t *p)
3672 return TASK_NICE(p);
3674 EXPORT_SYMBOL_GPL(task_nice);
3677 * idle_cpu - is a given cpu idle currently?
3678 * @cpu: the processor in question.
3680 int idle_cpu(int cpu)
3682 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
3686 * idle_task - return the idle task for a given cpu.
3687 * @cpu: the processor in question.
3689 task_t *idle_task(int cpu)
3691 return cpu_rq(cpu)->idle;
3695 * find_process_by_pid - find a process with a matching PID value.
3696 * @pid: the pid in question.
3698 static inline task_t *find_process_by_pid(pid_t pid)
3700 return pid ? find_task_by_pid(pid) : current;
3703 /* Actually do priority change: must hold rq lock. */
3704 static void __setscheduler(struct task_struct *p, int policy, int prio)
3708 p->rt_priority = prio;
3709 if (policy != SCHED_NORMAL)
3710 p->prio = MAX_RT_PRIO-1 - p->rt_priority;
3712 p->prio = p->static_prio;
3716 * sched_setscheduler - change the scheduling policy and/or RT priority of
3718 * @p: the task in question.
3719 * @policy: new policy.
3720 * @param: structure containing the new RT priority.
3722 int sched_setscheduler(struct task_struct *p, int policy,
3723 struct sched_param *param)
3726 int oldprio, oldpolicy = -1;
3727 prio_array_t *array;
3728 unsigned long flags;
3732 /* double check policy once rq lock held */
3734 policy = oldpolicy = p->policy;
3735 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
3736 policy != SCHED_NORMAL)
3739 * Valid priorities for SCHED_FIFO and SCHED_RR are
3740 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL is 0.
3742 if (param->sched_priority < 0 ||
3743 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3744 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3746 if ((policy == SCHED_NORMAL) != (param->sched_priority == 0))
3750 * Allow unprivileged RT tasks to decrease priority:
3752 if (!capable(CAP_SYS_NICE)) {
3753 /* can't change policy */
3754 if (policy != p->policy &&
3755 !p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
3757 /* can't increase priority */
3758 if (policy != SCHED_NORMAL &&
3759 param->sched_priority > p->rt_priority &&
3760 param->sched_priority >
3761 p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
3763 /* can't change other user's priorities */
3764 if ((current->euid != p->euid) &&
3765 (current->euid != p->uid))
3769 retval = security_task_setscheduler(p, policy, param);
3773 * To be able to change p->policy safely, the apropriate
3774 * runqueue lock must be held.
3776 rq = task_rq_lock(p, &flags);
3777 /* recheck policy now with rq lock held */
3778 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3779 policy = oldpolicy = -1;
3780 task_rq_unlock(rq, &flags);
3785 deactivate_task(p, rq);
3787 __setscheduler(p, policy, param->sched_priority);
3789 __activate_task(p, rq);
3791 * Reschedule if we are currently running on this runqueue and
3792 * our priority decreased, or if we are not currently running on
3793 * this runqueue and our priority is higher than the current's
3795 if (task_running(rq, p)) {
3796 if (p->prio > oldprio)
3797 resched_task(rq->curr);
3798 } else if (TASK_PREEMPTS_CURR(p, rq))
3799 resched_task(rq->curr);
3801 task_rq_unlock(rq, &flags);
3804 EXPORT_SYMBOL_GPL(sched_setscheduler);
3807 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3810 struct sched_param lparam;
3811 struct task_struct *p;
3813 if (!param || pid < 0)
3815 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3817 read_lock_irq(&tasklist_lock);
3818 p = find_process_by_pid(pid);
3820 read_unlock_irq(&tasklist_lock);
3823 retval = sched_setscheduler(p, policy, &lparam);
3824 read_unlock_irq(&tasklist_lock);
3829 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3830 * @pid: the pid in question.
3831 * @policy: new policy.
3832 * @param: structure containing the new RT priority.
3834 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
3835 struct sched_param __user *param)
3837 return do_sched_setscheduler(pid, policy, param);
3841 * sys_sched_setparam - set/change the RT priority of a thread
3842 * @pid: the pid in question.
3843 * @param: structure containing the new RT priority.
3845 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
3847 return do_sched_setscheduler(pid, -1, param);
3851 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3852 * @pid: the pid in question.
3854 asmlinkage long sys_sched_getscheduler(pid_t pid)
3856 int retval = -EINVAL;
3863 read_lock(&tasklist_lock);
3864 p = find_process_by_pid(pid);
3866 retval = security_task_getscheduler(p);
3870 read_unlock(&tasklist_lock);
3877 * sys_sched_getscheduler - get the RT priority of a thread
3878 * @pid: the pid in question.
3879 * @param: structure containing the RT priority.
3881 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
3883 struct sched_param lp;
3884 int retval = -EINVAL;
3887 if (!param || pid < 0)
3890 read_lock(&tasklist_lock);
3891 p = find_process_by_pid(pid);
3896 retval = security_task_getscheduler(p);
3900 lp.sched_priority = p->rt_priority;
3901 read_unlock(&tasklist_lock);
3904 * This one might sleep, we cannot do it with a spinlock held ...
3906 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3912 read_unlock(&tasklist_lock);
3916 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
3920 cpumask_t cpus_allowed;
3923 read_lock(&tasklist_lock);
3925 p = find_process_by_pid(pid);
3927 read_unlock(&tasklist_lock);
3928 unlock_cpu_hotplug();
3933 * It is not safe to call set_cpus_allowed with the
3934 * tasklist_lock held. We will bump the task_struct's
3935 * usage count and then drop tasklist_lock.
3938 read_unlock(&tasklist_lock);
3941 if ((current->euid != p->euid) && (current->euid != p->uid) &&
3942 !capable(CAP_SYS_NICE))
3945 cpus_allowed = cpuset_cpus_allowed(p);
3946 cpus_and(new_mask, new_mask, cpus_allowed);
3947 retval = set_cpus_allowed(p, new_mask);
3951 unlock_cpu_hotplug();
3955 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3956 cpumask_t *new_mask)
3958 if (len < sizeof(cpumask_t)) {
3959 memset(new_mask, 0, sizeof(cpumask_t));
3960 } else if (len > sizeof(cpumask_t)) {
3961 len = sizeof(cpumask_t);
3963 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3967 * sys_sched_setaffinity - set the cpu affinity of a process
3968 * @pid: pid of the process
3969 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3970 * @user_mask_ptr: user-space pointer to the new cpu mask
3972 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
3973 unsigned long __user *user_mask_ptr)
3978 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
3982 return sched_setaffinity(pid, new_mask);
3986 * Represents all cpu's present in the system
3987 * In systems capable of hotplug, this map could dynamically grow
3988 * as new cpu's are detected in the system via any platform specific
3989 * method, such as ACPI for e.g.
3992 cpumask_t cpu_present_map __read_mostly;
3993 EXPORT_SYMBOL(cpu_present_map);
3996 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
3997 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4000 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4006 read_lock(&tasklist_lock);
4009 p = find_process_by_pid(pid);
4014 cpus_and(*mask, p->cpus_allowed, cpu_possible_map);
4017 read_unlock(&tasklist_lock);
4018 unlock_cpu_hotplug();
4026 * sys_sched_getaffinity - get the cpu affinity of a process
4027 * @pid: pid of the process
4028 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4029 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4031 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4032 unsigned long __user *user_mask_ptr)
4037 if (len < sizeof(cpumask_t))
4040 ret = sched_getaffinity(pid, &mask);
4044 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4047 return sizeof(cpumask_t);
4051 * sys_sched_yield - yield the current processor to other threads.
4053 * this function yields the current CPU by moving the calling thread
4054 * to the expired array. If there are no other threads running on this
4055 * CPU then this function will return.
4057 asmlinkage long sys_sched_yield(void)
4059 runqueue_t *rq = this_rq_lock();
4060 prio_array_t *array = current->array;
4061 prio_array_t *target = rq->expired;
4063 schedstat_inc(rq, yld_cnt);
4065 * We implement yielding by moving the task into the expired
4068 * (special rule: RT tasks will just roundrobin in the active
4071 if (rt_task(current))
4072 target = rq->active;
4074 if (array->nr_active == 1) {
4075 schedstat_inc(rq, yld_act_empty);
4076 if (!rq->expired->nr_active)
4077 schedstat_inc(rq, yld_both_empty);
4078 } else if (!rq->expired->nr_active)
4079 schedstat_inc(rq, yld_exp_empty);
4081 if (array != target) {
4082 dequeue_task(current, array);
4083 enqueue_task(current, target);
4086 * requeue_task is cheaper so perform that if possible.
4088 requeue_task(current, array);
4091 * Since we are going to call schedule() anyway, there's
4092 * no need to preempt or enable interrupts:
4094 __release(rq->lock);
4095 _raw_spin_unlock(&rq->lock);
4096 preempt_enable_no_resched();
4103 static inline void __cond_resched(void)
4106 * The BKS might be reacquired before we have dropped
4107 * PREEMPT_ACTIVE, which could trigger a second
4108 * cond_resched() call.
4110 if (unlikely(preempt_count()))
4113 add_preempt_count(PREEMPT_ACTIVE);
4115 sub_preempt_count(PREEMPT_ACTIVE);
4116 } while (need_resched());
4119 int __sched cond_resched(void)
4121 if (need_resched()) {
4128 EXPORT_SYMBOL(cond_resched);
4131 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4132 * call schedule, and on return reacquire the lock.
4134 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4135 * operations here to prevent schedule() from being called twice (once via
4136 * spin_unlock(), once by hand).
4138 int cond_resched_lock(spinlock_t *lock)
4142 if (need_lockbreak(lock)) {
4148 if (need_resched()) {
4149 _raw_spin_unlock(lock);
4150 preempt_enable_no_resched();
4158 EXPORT_SYMBOL(cond_resched_lock);
4160 int __sched cond_resched_softirq(void)
4162 BUG_ON(!in_softirq());
4164 if (need_resched()) {
4165 __local_bh_enable();
4173 EXPORT_SYMBOL(cond_resched_softirq);
4177 * yield - yield the current processor to other threads.
4179 * this is a shortcut for kernel-space yielding - it marks the
4180 * thread runnable and calls sys_sched_yield().
4182 void __sched yield(void)
4184 set_current_state(TASK_RUNNING);
4188 EXPORT_SYMBOL(yield);
4191 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4192 * that process accounting knows that this is a task in IO wait state.
4194 * But don't do that if it is a deliberate, throttling IO wait (this task
4195 * has set its backing_dev_info: the queue against which it should throttle)
4197 void __sched io_schedule(void)
4199 struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id());
4201 atomic_inc(&rq->nr_iowait);
4203 atomic_dec(&rq->nr_iowait);
4206 EXPORT_SYMBOL(io_schedule);
4208 long __sched io_schedule_timeout(long timeout)
4210 struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id());
4213 atomic_inc(&rq->nr_iowait);
4214 ret = schedule_timeout(timeout);
4215 atomic_dec(&rq->nr_iowait);
4220 * sys_sched_get_priority_max - return maximum RT priority.
4221 * @policy: scheduling class.
4223 * this syscall returns the maximum rt_priority that can be used
4224 * by a given scheduling class.
4226 asmlinkage long sys_sched_get_priority_max(int policy)
4233 ret = MAX_USER_RT_PRIO-1;
4243 * sys_sched_get_priority_min - return minimum RT priority.
4244 * @policy: scheduling class.
4246 * this syscall returns the minimum rt_priority that can be used
4247 * by a given scheduling class.
4249 asmlinkage long sys_sched_get_priority_min(int policy)
4265 * sys_sched_rr_get_interval - return the default timeslice of a process.
4266 * @pid: pid of the process.
4267 * @interval: userspace pointer to the timeslice value.
4269 * this syscall writes the default timeslice value of a given process
4270 * into the user-space timespec buffer. A value of '0' means infinity.
4273 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4275 int retval = -EINVAL;
4283 read_lock(&tasklist_lock);
4284 p = find_process_by_pid(pid);
4288 retval = security_task_getscheduler(p);
4292 jiffies_to_timespec(p->policy & SCHED_FIFO ?
4293 0 : task_timeslice(p), &t);
4294 read_unlock(&tasklist_lock);
4295 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4299 read_unlock(&tasklist_lock);
4303 static inline struct task_struct *eldest_child(struct task_struct *p)
4305 if (list_empty(&p->children)) return NULL;
4306 return list_entry(p->children.next,struct task_struct,sibling);
4309 static inline struct task_struct *older_sibling(struct task_struct *p)
4311 if (p->sibling.prev==&p->parent->children) return NULL;
4312 return list_entry(p->sibling.prev,struct task_struct,sibling);
4315 static inline struct task_struct *younger_sibling(struct task_struct *p)
4317 if (p->sibling.next==&p->parent->children) return NULL;
4318 return list_entry(p->sibling.next,struct task_struct,sibling);
4321 static void show_task(task_t *p)
4325 unsigned long free = 0;
4326 static const char *stat_nam[] = { "R", "S", "D", "T", "t", "Z", "X" };
4328 printk("%-13.13s ", p->comm);
4329 state = p->state ? __ffs(p->state) + 1 : 0;
4330 if (state < ARRAY_SIZE(stat_nam))
4331 printk(stat_nam[state]);
4334 #if (BITS_PER_LONG == 32)
4335 if (state == TASK_RUNNING)
4336 printk(" running ");
4338 printk(" %08lX ", thread_saved_pc(p));
4340 if (state == TASK_RUNNING)
4341 printk(" running task ");
4343 printk(" %016lx ", thread_saved_pc(p));
4345 #ifdef CONFIG_DEBUG_STACK_USAGE
4347 unsigned long *n = end_of_stack(p);
4350 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4353 printk("%5lu %5d %6d ", free, p->pid, p->parent->pid);
4354 if ((relative = eldest_child(p)))
4355 printk("%5d ", relative->pid);
4358 if ((relative = younger_sibling(p)))
4359 printk("%7d", relative->pid);
4362 if ((relative = older_sibling(p)))
4363 printk(" %5d", relative->pid);
4367 printk(" (L-TLB)\n");
4369 printk(" (NOTLB)\n");
4371 if (state != TASK_RUNNING)
4372 show_stack(p, NULL);
4375 void show_state(void)
4379 #if (BITS_PER_LONG == 32)
4382 printk(" task PC pid father child younger older\n");
4386 printk(" task PC pid father child younger older\n");
4388 read_lock(&tasklist_lock);
4389 do_each_thread(g, p) {
4391 * reset the NMI-timeout, listing all files on a slow
4392 * console might take alot of time:
4394 touch_nmi_watchdog();
4396 } while_each_thread(g, p);
4398 read_unlock(&tasklist_lock);
4399 mutex_debug_show_all_locks();
4403 * init_idle - set up an idle thread for a given CPU
4404 * @idle: task in question
4405 * @cpu: cpu the idle task belongs to
4407 * NOTE: this function does not set the idle thread's NEED_RESCHED
4408 * flag, to make booting more robust.
4410 void __devinit init_idle(task_t *idle, int cpu)
4412 runqueue_t *rq = cpu_rq(cpu);
4413 unsigned long flags;
4415 idle->sleep_avg = 0;
4417 idle->prio = MAX_PRIO;
4418 idle->state = TASK_RUNNING;
4419 idle->cpus_allowed = cpumask_of_cpu(cpu);
4420 set_task_cpu(idle, cpu);
4422 spin_lock_irqsave(&rq->lock, flags);
4423 rq->curr = rq->idle = idle;
4424 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4427 spin_unlock_irqrestore(&rq->lock, flags);
4429 /* Set the preempt count _outside_ the spinlocks! */
4430 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4431 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4433 task_thread_info(idle)->preempt_count = 0;
4438 * In a system that switches off the HZ timer nohz_cpu_mask
4439 * indicates which cpus entered this state. This is used
4440 * in the rcu update to wait only for active cpus. For system
4441 * which do not switch off the HZ timer nohz_cpu_mask should
4442 * always be CPU_MASK_NONE.
4444 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4448 * This is how migration works:
4450 * 1) we queue a migration_req_t structure in the source CPU's
4451 * runqueue and wake up that CPU's migration thread.
4452 * 2) we down() the locked semaphore => thread blocks.
4453 * 3) migration thread wakes up (implicitly it forces the migrated
4454 * thread off the CPU)
4455 * 4) it gets the migration request and checks whether the migrated
4456 * task is still in the wrong runqueue.
4457 * 5) if it's in the wrong runqueue then the migration thread removes
4458 * it and puts it into the right queue.
4459 * 6) migration thread up()s the semaphore.
4460 * 7) we wake up and the migration is done.
4464 * Change a given task's CPU affinity. Migrate the thread to a
4465 * proper CPU and schedule it away if the CPU it's executing on
4466 * is removed from the allowed bitmask.
4468 * NOTE: the caller must have a valid reference to the task, the
4469 * task must not exit() & deallocate itself prematurely. The
4470 * call is not atomic; no spinlocks may be held.
4472 int set_cpus_allowed(task_t *p, cpumask_t new_mask)
4474 unsigned long flags;
4476 migration_req_t req;
4479 rq = task_rq_lock(p, &flags);
4480 if (!cpus_intersects(new_mask, cpu_online_map)) {
4485 p->cpus_allowed = new_mask;
4486 /* Can the task run on the task's current CPU? If so, we're done */
4487 if (cpu_isset(task_cpu(p), new_mask))
4490 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4491 /* Need help from migration thread: drop lock and wait. */
4492 task_rq_unlock(rq, &flags);
4493 wake_up_process(rq->migration_thread);
4494 wait_for_completion(&req.done);
4495 tlb_migrate_finish(p->mm);
4499 task_rq_unlock(rq, &flags);
4503 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4506 * Move (not current) task off this cpu, onto dest cpu. We're doing
4507 * this because either it can't run here any more (set_cpus_allowed()
4508 * away from this CPU, or CPU going down), or because we're
4509 * attempting to rebalance this task on exec (sched_exec).
4511 * So we race with normal scheduler movements, but that's OK, as long
4512 * as the task is no longer on this CPU.
4514 static void __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4516 runqueue_t *rq_dest, *rq_src;
4518 if (unlikely(cpu_is_offline(dest_cpu)))
4521 rq_src = cpu_rq(src_cpu);
4522 rq_dest = cpu_rq(dest_cpu);
4524 double_rq_lock(rq_src, rq_dest);
4525 /* Already moved. */
4526 if (task_cpu(p) != src_cpu)
4528 /* Affinity changed (again). */
4529 if (!cpu_isset(dest_cpu, p->cpus_allowed))
4532 set_task_cpu(p, dest_cpu);
4535 * Sync timestamp with rq_dest's before activating.
4536 * The same thing could be achieved by doing this step
4537 * afterwards, and pretending it was a local activate.
4538 * This way is cleaner and logically correct.
4540 p->timestamp = p->timestamp - rq_src->timestamp_last_tick
4541 + rq_dest->timestamp_last_tick;
4542 deactivate_task(p, rq_src);
4543 activate_task(p, rq_dest, 0);
4544 if (TASK_PREEMPTS_CURR(p, rq_dest))
4545 resched_task(rq_dest->curr);
4549 double_rq_unlock(rq_src, rq_dest);
4553 * migration_thread - this is a highprio system thread that performs
4554 * thread migration by bumping thread off CPU then 'pushing' onto
4557 static int migration_thread(void *data)
4560 int cpu = (long)data;
4563 BUG_ON(rq->migration_thread != current);
4565 set_current_state(TASK_INTERRUPTIBLE);
4566 while (!kthread_should_stop()) {
4567 struct list_head *head;
4568 migration_req_t *req;
4572 spin_lock_irq(&rq->lock);
4574 if (cpu_is_offline(cpu)) {
4575 spin_unlock_irq(&rq->lock);
4579 if (rq->active_balance) {
4580 active_load_balance(rq, cpu);
4581 rq->active_balance = 0;
4584 head = &rq->migration_queue;
4586 if (list_empty(head)) {
4587 spin_unlock_irq(&rq->lock);
4589 set_current_state(TASK_INTERRUPTIBLE);
4592 req = list_entry(head->next, migration_req_t, list);
4593 list_del_init(head->next);
4595 spin_unlock(&rq->lock);
4596 __migrate_task(req->task, cpu, req->dest_cpu);
4599 complete(&req->done);
4601 __set_current_state(TASK_RUNNING);
4605 /* Wait for kthread_stop */
4606 set_current_state(TASK_INTERRUPTIBLE);
4607 while (!kthread_should_stop()) {
4609 set_current_state(TASK_INTERRUPTIBLE);
4611 __set_current_state(TASK_RUNNING);
4615 #ifdef CONFIG_HOTPLUG_CPU
4616 /* Figure out where task on dead CPU should go, use force if neccessary. */
4617 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *tsk)
4623 mask = node_to_cpumask(cpu_to_node(dead_cpu));
4624 cpus_and(mask, mask, tsk->cpus_allowed);
4625 dest_cpu = any_online_cpu(mask);
4627 /* On any allowed CPU? */
4628 if (dest_cpu == NR_CPUS)
4629 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4631 /* No more Mr. Nice Guy. */
4632 if (dest_cpu == NR_CPUS) {
4633 cpus_setall(tsk->cpus_allowed);
4634 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4637 * Don't tell them about moving exiting tasks or
4638 * kernel threads (both mm NULL), since they never
4641 if (tsk->mm && printk_ratelimit())
4642 printk(KERN_INFO "process %d (%s) no "
4643 "longer affine to cpu%d\n",
4644 tsk->pid, tsk->comm, dead_cpu);
4646 __migrate_task(tsk, dead_cpu, dest_cpu);
4650 * While a dead CPU has no uninterruptible tasks queued at this point,
4651 * it might still have a nonzero ->nr_uninterruptible counter, because
4652 * for performance reasons the counter is not stricly tracking tasks to
4653 * their home CPUs. So we just add the counter to another CPU's counter,
4654 * to keep the global sum constant after CPU-down:
4656 static void migrate_nr_uninterruptible(runqueue_t *rq_src)
4658 runqueue_t *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
4659 unsigned long flags;
4661 local_irq_save(flags);
4662 double_rq_lock(rq_src, rq_dest);
4663 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
4664 rq_src->nr_uninterruptible = 0;
4665 double_rq_unlock(rq_src, rq_dest);
4666 local_irq_restore(flags);
4669 /* Run through task list and migrate tasks from the dead cpu. */
4670 static void migrate_live_tasks(int src_cpu)
4672 struct task_struct *tsk, *t;
4674 write_lock_irq(&tasklist_lock);
4676 do_each_thread(t, tsk) {
4680 if (task_cpu(tsk) == src_cpu)
4681 move_task_off_dead_cpu(src_cpu, tsk);
4682 } while_each_thread(t, tsk);
4684 write_unlock_irq(&tasklist_lock);
4687 /* Schedules idle task to be the next runnable task on current CPU.
4688 * It does so by boosting its priority to highest possible and adding it to
4689 * the _front_ of runqueue. Used by CPU offline code.
4691 void sched_idle_next(void)
4693 int cpu = smp_processor_id();
4694 runqueue_t *rq = this_rq();
4695 struct task_struct *p = rq->idle;
4696 unsigned long flags;
4698 /* cpu has to be offline */
4699 BUG_ON(cpu_online(cpu));
4701 /* Strictly not necessary since rest of the CPUs are stopped by now
4702 * and interrupts disabled on current cpu.
4704 spin_lock_irqsave(&rq->lock, flags);
4706 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4707 /* Add idle task to _front_ of it's priority queue */
4708 __activate_idle_task(p, rq);
4710 spin_unlock_irqrestore(&rq->lock, flags);
4713 /* Ensures that the idle task is using init_mm right before its cpu goes
4716 void idle_task_exit(void)
4718 struct mm_struct *mm = current->active_mm;
4720 BUG_ON(cpu_online(smp_processor_id()));
4723 switch_mm(mm, &init_mm, current);
4727 static void migrate_dead(unsigned int dead_cpu, task_t *tsk)
4729 struct runqueue *rq = cpu_rq(dead_cpu);
4731 /* Must be exiting, otherwise would be on tasklist. */
4732 BUG_ON(tsk->exit_state != EXIT_ZOMBIE && tsk->exit_state != EXIT_DEAD);
4734 /* Cannot have done final schedule yet: would have vanished. */
4735 BUG_ON(tsk->flags & PF_DEAD);
4737 get_task_struct(tsk);
4740 * Drop lock around migration; if someone else moves it,
4741 * that's OK. No task can be added to this CPU, so iteration is
4744 spin_unlock_irq(&rq->lock);
4745 move_task_off_dead_cpu(dead_cpu, tsk);
4746 spin_lock_irq(&rq->lock);
4748 put_task_struct(tsk);
4751 /* release_task() removes task from tasklist, so we won't find dead tasks. */
4752 static void migrate_dead_tasks(unsigned int dead_cpu)
4755 struct runqueue *rq = cpu_rq(dead_cpu);
4757 for (arr = 0; arr < 2; arr++) {
4758 for (i = 0; i < MAX_PRIO; i++) {
4759 struct list_head *list = &rq->arrays[arr].queue[i];
4760 while (!list_empty(list))
4761 migrate_dead(dead_cpu,
4762 list_entry(list->next, task_t,
4767 #endif /* CONFIG_HOTPLUG_CPU */
4770 * migration_call - callback that gets triggered when a CPU is added.
4771 * Here we can start up the necessary migration thread for the new CPU.
4773 static int migration_call(struct notifier_block *nfb, unsigned long action,
4776 int cpu = (long)hcpu;
4777 struct task_struct *p;
4778 struct runqueue *rq;
4779 unsigned long flags;
4782 case CPU_UP_PREPARE:
4783 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
4786 p->flags |= PF_NOFREEZE;
4787 kthread_bind(p, cpu);
4788 /* Must be high prio: stop_machine expects to yield to it. */
4789 rq = task_rq_lock(p, &flags);
4790 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4791 task_rq_unlock(rq, &flags);
4792 cpu_rq(cpu)->migration_thread = p;
4795 /* Strictly unneccessary, as first user will wake it. */
4796 wake_up_process(cpu_rq(cpu)->migration_thread);
4798 #ifdef CONFIG_HOTPLUG_CPU
4799 case CPU_UP_CANCELED:
4800 /* Unbind it from offline cpu so it can run. Fall thru. */
4801 kthread_bind(cpu_rq(cpu)->migration_thread,
4802 any_online_cpu(cpu_online_map));
4803 kthread_stop(cpu_rq(cpu)->migration_thread);
4804 cpu_rq(cpu)->migration_thread = NULL;
4807 migrate_live_tasks(cpu);
4809 kthread_stop(rq->migration_thread);
4810 rq->migration_thread = NULL;
4811 /* Idle task back to normal (off runqueue, low prio) */
4812 rq = task_rq_lock(rq->idle, &flags);
4813 deactivate_task(rq->idle, rq);
4814 rq->idle->static_prio = MAX_PRIO;
4815 __setscheduler(rq->idle, SCHED_NORMAL, 0);
4816 migrate_dead_tasks(cpu);
4817 task_rq_unlock(rq, &flags);
4818 migrate_nr_uninterruptible(rq);
4819 BUG_ON(rq->nr_running != 0);
4821 /* No need to migrate the tasks: it was best-effort if
4822 * they didn't do lock_cpu_hotplug(). Just wake up
4823 * the requestors. */
4824 spin_lock_irq(&rq->lock);
4825 while (!list_empty(&rq->migration_queue)) {
4826 migration_req_t *req;
4827 req = list_entry(rq->migration_queue.next,
4828 migration_req_t, list);
4829 list_del_init(&req->list);
4830 complete(&req->done);
4832 spin_unlock_irq(&rq->lock);
4839 /* Register at highest priority so that task migration (migrate_all_tasks)
4840 * happens before everything else.
4842 static struct notifier_block __devinitdata migration_notifier = {
4843 .notifier_call = migration_call,
4847 int __init migration_init(void)
4849 void *cpu = (void *)(long)smp_processor_id();
4850 /* Start one for boot CPU. */
4851 migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
4852 migration_call(&migration_notifier, CPU_ONLINE, cpu);
4853 register_cpu_notifier(&migration_notifier);
4859 #undef SCHED_DOMAIN_DEBUG
4860 #ifdef SCHED_DOMAIN_DEBUG
4861 static void sched_domain_debug(struct sched_domain *sd, int cpu)
4866 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
4870 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
4875 struct sched_group *group = sd->groups;
4876 cpumask_t groupmask;
4878 cpumask_scnprintf(str, NR_CPUS, sd->span);
4879 cpus_clear(groupmask);
4882 for (i = 0; i < level + 1; i++)
4884 printk("domain %d: ", level);
4886 if (!(sd->flags & SD_LOAD_BALANCE)) {
4887 printk("does not load-balance\n");
4889 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent");
4893 printk("span %s\n", str);
4895 if (!cpu_isset(cpu, sd->span))
4896 printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
4897 if (!cpu_isset(cpu, group->cpumask))
4898 printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
4901 for (i = 0; i < level + 2; i++)
4907 printk(KERN_ERR "ERROR: group is NULL\n");
4911 if (!group->cpu_power) {
4913 printk(KERN_ERR "ERROR: domain->cpu_power not set\n");
4916 if (!cpus_weight(group->cpumask)) {
4918 printk(KERN_ERR "ERROR: empty group\n");
4921 if (cpus_intersects(groupmask, group->cpumask)) {
4923 printk(KERN_ERR "ERROR: repeated CPUs\n");
4926 cpus_or(groupmask, groupmask, group->cpumask);
4928 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
4931 group = group->next;
4932 } while (group != sd->groups);
4935 if (!cpus_equal(sd->span, groupmask))
4936 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
4942 if (!cpus_subset(groupmask, sd->span))
4943 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
4949 #define sched_domain_debug(sd, cpu) {}
4952 static int sd_degenerate(struct sched_domain *sd)
4954 if (cpus_weight(sd->span) == 1)
4957 /* Following flags need at least 2 groups */
4958 if (sd->flags & (SD_LOAD_BALANCE |
4959 SD_BALANCE_NEWIDLE |
4962 if (sd->groups != sd->groups->next)
4966 /* Following flags don't use groups */
4967 if (sd->flags & (SD_WAKE_IDLE |
4975 static int sd_parent_degenerate(struct sched_domain *sd,
4976 struct sched_domain *parent)
4978 unsigned long cflags = sd->flags, pflags = parent->flags;
4980 if (sd_degenerate(parent))
4983 if (!cpus_equal(sd->span, parent->span))
4986 /* Does parent contain flags not in child? */
4987 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
4988 if (cflags & SD_WAKE_AFFINE)
4989 pflags &= ~SD_WAKE_BALANCE;
4990 /* Flags needing groups don't count if only 1 group in parent */
4991 if (parent->groups == parent->groups->next) {
4992 pflags &= ~(SD_LOAD_BALANCE |
4993 SD_BALANCE_NEWIDLE |
4997 if (~cflags & pflags)
5004 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5005 * hold the hotplug lock.
5007 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5009 runqueue_t *rq = cpu_rq(cpu);
5010 struct sched_domain *tmp;
5012 /* Remove the sched domains which do not contribute to scheduling. */
5013 for (tmp = sd; tmp; tmp = tmp->parent) {
5014 struct sched_domain *parent = tmp->parent;
5017 if (sd_parent_degenerate(tmp, parent))
5018 tmp->parent = parent->parent;
5021 if (sd && sd_degenerate(sd))
5024 sched_domain_debug(sd, cpu);
5026 rcu_assign_pointer(rq->sd, sd);
5029 /* cpus with isolated domains */
5030 static cpumask_t __devinitdata cpu_isolated_map = CPU_MASK_NONE;
5032 /* Setup the mask of cpus configured for isolated domains */
5033 static int __init isolated_cpu_setup(char *str)
5035 int ints[NR_CPUS], i;
5037 str = get_options(str, ARRAY_SIZE(ints), ints);
5038 cpus_clear(cpu_isolated_map);
5039 for (i = 1; i <= ints[0]; i++)
5040 if (ints[i] < NR_CPUS)
5041 cpu_set(ints[i], cpu_isolated_map);
5045 __setup ("isolcpus=", isolated_cpu_setup);
5048 * init_sched_build_groups takes an array of groups, the cpumask we wish
5049 * to span, and a pointer to a function which identifies what group a CPU
5050 * belongs to. The return value of group_fn must be a valid index into the
5051 * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
5052 * keep track of groups covered with a cpumask_t).
5054 * init_sched_build_groups will build a circular linked list of the groups
5055 * covered by the given span, and will set each group's ->cpumask correctly,
5056 * and ->cpu_power to 0.
5058 static void init_sched_build_groups(struct sched_group groups[], cpumask_t span,
5059 int (*group_fn)(int cpu))
5061 struct sched_group *first = NULL, *last = NULL;
5062 cpumask_t covered = CPU_MASK_NONE;
5065 for_each_cpu_mask(i, span) {
5066 int group = group_fn(i);
5067 struct sched_group *sg = &groups[group];
5070 if (cpu_isset(i, covered))
5073 sg->cpumask = CPU_MASK_NONE;
5076 for_each_cpu_mask(j, span) {
5077 if (group_fn(j) != group)
5080 cpu_set(j, covered);
5081 cpu_set(j, sg->cpumask);
5092 #define SD_NODES_PER_DOMAIN 16
5095 * Self-tuning task migration cost measurement between source and target CPUs.
5097 * This is done by measuring the cost of manipulating buffers of varying
5098 * sizes. For a given buffer-size here are the steps that are taken:
5100 * 1) the source CPU reads+dirties a shared buffer
5101 * 2) the target CPU reads+dirties the same shared buffer
5103 * We measure how long they take, in the following 4 scenarios:
5105 * - source: CPU1, target: CPU2 | cost1
5106 * - source: CPU2, target: CPU1 | cost2
5107 * - source: CPU1, target: CPU1 | cost3
5108 * - source: CPU2, target: CPU2 | cost4
5110 * We then calculate the cost3+cost4-cost1-cost2 difference - this is
5111 * the cost of migration.
5113 * We then start off from a small buffer-size and iterate up to larger
5114 * buffer sizes, in 5% steps - measuring each buffer-size separately, and
5115 * doing a maximum search for the cost. (The maximum cost for a migration
5116 * normally occurs when the working set size is around the effective cache
5119 #define SEARCH_SCOPE 2
5120 #define MIN_CACHE_SIZE (64*1024U)
5121 #define DEFAULT_CACHE_SIZE (5*1024*1024U)
5122 #define ITERATIONS 2
5123 #define SIZE_THRESH 130
5124 #define COST_THRESH 130
5127 * The migration cost is a function of 'domain distance'. Domain
5128 * distance is the number of steps a CPU has to iterate down its
5129 * domain tree to share a domain with the other CPU. The farther
5130 * two CPUs are from each other, the larger the distance gets.
5132 * Note that we use the distance only to cache measurement results,
5133 * the distance value is not used numerically otherwise. When two
5134 * CPUs have the same distance it is assumed that the migration
5135 * cost is the same. (this is a simplification but quite practical)
5137 #define MAX_DOMAIN_DISTANCE 32
5139 static unsigned long long migration_cost[MAX_DOMAIN_DISTANCE] =
5140 { [ 0 ... MAX_DOMAIN_DISTANCE-1 ] = -1LL };
5143 * Allow override of migration cost - in units of microseconds.
5144 * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost
5145 * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs:
5147 static int __init migration_cost_setup(char *str)
5149 int ints[MAX_DOMAIN_DISTANCE+1], i;
5151 str = get_options(str, ARRAY_SIZE(ints), ints);
5153 printk("#ints: %d\n", ints[0]);
5154 for (i = 1; i <= ints[0]; i++) {
5155 migration_cost[i-1] = (unsigned long long)ints[i]*1000;
5156 printk("migration_cost[%d]: %Ld\n", i-1, migration_cost[i-1]);
5161 __setup ("migration_cost=", migration_cost_setup);
5164 * Global multiplier (divisor) for migration-cutoff values,
5165 * in percentiles. E.g. use a value of 150 to get 1.5 times
5166 * longer cache-hot cutoff times.
5168 * (We scale it from 100 to 128 to long long handling easier.)
5171 #define MIGRATION_FACTOR_SCALE 128
5173 static unsigned int migration_factor = MIGRATION_FACTOR_SCALE;
5175 static int __init setup_migration_factor(char *str)
5177 get_option(&str, &migration_factor);
5178 migration_factor = migration_factor * MIGRATION_FACTOR_SCALE / 100;
5182 __setup("migration_factor=", setup_migration_factor);
5185 * Estimated distance of two CPUs, measured via the number of domains
5186 * we have to pass for the two CPUs to be in the same span:
5188 static unsigned long domain_distance(int cpu1, int cpu2)
5190 unsigned long distance = 0;
5191 struct sched_domain *sd;
5193 for_each_domain(cpu1, sd) {
5194 WARN_ON(!cpu_isset(cpu1, sd->span));
5195 if (cpu_isset(cpu2, sd->span))
5199 if (distance >= MAX_DOMAIN_DISTANCE) {
5201 distance = MAX_DOMAIN_DISTANCE-1;
5207 static unsigned int migration_debug;
5209 static int __init setup_migration_debug(char *str)
5211 get_option(&str, &migration_debug);
5215 __setup("migration_debug=", setup_migration_debug);
5218 * Maximum cache-size that the scheduler should try to measure.
5219 * Architectures with larger caches should tune this up during
5220 * bootup. Gets used in the domain-setup code (i.e. during SMP
5223 unsigned int max_cache_size;
5225 static int __init setup_max_cache_size(char *str)
5227 get_option(&str, &max_cache_size);
5231 __setup("max_cache_size=", setup_max_cache_size);
5234 * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This
5235 * is the operation that is timed, so we try to generate unpredictable
5236 * cachemisses that still end up filling the L2 cache:
5238 static void touch_cache(void *__cache, unsigned long __size)
5240 unsigned long size = __size/sizeof(long), chunk1 = size/3,
5242 unsigned long *cache = __cache;
5245 for (i = 0; i < size/6; i += 8) {
5248 case 1: cache[size-1-i]++;
5249 case 2: cache[chunk1-i]++;
5250 case 3: cache[chunk1+i]++;
5251 case 4: cache[chunk2-i]++;
5252 case 5: cache[chunk2+i]++;
5258 * Measure the cache-cost of one task migration. Returns in units of nsec.
5260 static unsigned long long measure_one(void *cache, unsigned long size,
5261 int source, int target)
5263 cpumask_t mask, saved_mask;
5264 unsigned long long t0, t1, t2, t3, cost;
5266 saved_mask = current->cpus_allowed;
5269 * Flush source caches to RAM and invalidate them:
5274 * Migrate to the source CPU:
5276 mask = cpumask_of_cpu(source);
5277 set_cpus_allowed(current, mask);
5278 WARN_ON(smp_processor_id() != source);
5281 * Dirty the working set:
5284 touch_cache(cache, size);
5288 * Migrate to the target CPU, dirty the L2 cache and access
5289 * the shared buffer. (which represents the working set
5290 * of a migrated task.)
5292 mask = cpumask_of_cpu(target);
5293 set_cpus_allowed(current, mask);
5294 WARN_ON(smp_processor_id() != target);
5297 touch_cache(cache, size);
5300 cost = t1-t0 + t3-t2;
5302 if (migration_debug >= 2)
5303 printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n",
5304 source, target, t1-t0, t1-t0, t3-t2, cost);
5306 * Flush target caches to RAM and invalidate them:
5310 set_cpus_allowed(current, saved_mask);
5316 * Measure a series of task migrations and return the average
5317 * result. Since this code runs early during bootup the system
5318 * is 'undisturbed' and the average latency makes sense.
5320 * The algorithm in essence auto-detects the relevant cache-size,
5321 * so it will properly detect different cachesizes for different
5322 * cache-hierarchies, depending on how the CPUs are connected.
5324 * Architectures can prime the upper limit of the search range via
5325 * max_cache_size, otherwise the search range defaults to 20MB...64K.
5327 static unsigned long long
5328 measure_cost(int cpu1, int cpu2, void *cache, unsigned int size)
5330 unsigned long long cost1, cost2;
5334 * Measure the migration cost of 'size' bytes, over an
5335 * average of 10 runs:
5337 * (We perturb the cache size by a small (0..4k)
5338 * value to compensate size/alignment related artifacts.
5339 * We also subtract the cost of the operation done on
5345 * dry run, to make sure we start off cache-cold on cpu1,
5346 * and to get any vmalloc pagefaults in advance:
5348 measure_one(cache, size, cpu1, cpu2);
5349 for (i = 0; i < ITERATIONS; i++)
5350 cost1 += measure_one(cache, size - i*1024, cpu1, cpu2);
5352 measure_one(cache, size, cpu2, cpu1);
5353 for (i = 0; i < ITERATIONS; i++)
5354 cost1 += measure_one(cache, size - i*1024, cpu2, cpu1);
5357 * (We measure the non-migrating [cached] cost on both
5358 * cpu1 and cpu2, to handle CPUs with different speeds)
5362 measure_one(cache, size, cpu1, cpu1);
5363 for (i = 0; i < ITERATIONS; i++)
5364 cost2 += measure_one(cache, size - i*1024, cpu1, cpu1);
5366 measure_one(cache, size, cpu2, cpu2);
5367 for (i = 0; i < ITERATIONS; i++)
5368 cost2 += measure_one(cache, size - i*1024, cpu2, cpu2);
5371 * Get the per-iteration migration cost:
5373 do_div(cost1, 2*ITERATIONS);
5374 do_div(cost2, 2*ITERATIONS);
5376 return cost1 - cost2;
5379 static unsigned long long measure_migration_cost(int cpu1, int cpu2)
5381 unsigned long long max_cost = 0, fluct = 0, avg_fluct = 0;
5382 unsigned int max_size, size, size_found = 0;
5383 long long cost = 0, prev_cost;
5387 * Search from max_cache_size*5 down to 64K - the real relevant
5388 * cachesize has to lie somewhere inbetween.
5390 if (max_cache_size) {
5391 max_size = max(max_cache_size * SEARCH_SCOPE, MIN_CACHE_SIZE);
5392 size = max(max_cache_size / SEARCH_SCOPE, MIN_CACHE_SIZE);
5395 * Since we have no estimation about the relevant
5398 max_size = DEFAULT_CACHE_SIZE * SEARCH_SCOPE;
5399 size = MIN_CACHE_SIZE;
5402 if (!cpu_online(cpu1) || !cpu_online(cpu2)) {
5403 printk("cpu %d and %d not both online!\n", cpu1, cpu2);
5408 * Allocate the working set:
5410 cache = vmalloc(max_size);
5412 printk("could not vmalloc %d bytes for cache!\n", 2*max_size);
5413 return 1000000; // return 1 msec on very small boxen
5416 while (size <= max_size) {
5418 cost = measure_cost(cpu1, cpu2, cache, size);
5424 if (max_cost < cost) {
5430 * Calculate average fluctuation, we use this to prevent
5431 * noise from triggering an early break out of the loop:
5433 fluct = abs(cost - prev_cost);
5434 avg_fluct = (avg_fluct + fluct)/2;
5436 if (migration_debug)
5437 printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): (%8Ld %8Ld)\n",
5439 (long)cost / 1000000,
5440 ((long)cost / 100000) % 10,
5441 (long)max_cost / 1000000,
5442 ((long)max_cost / 100000) % 10,
5443 domain_distance(cpu1, cpu2),
5447 * If we iterated at least 20% past the previous maximum,
5448 * and the cost has dropped by more than 20% already,
5449 * (taking fluctuations into account) then we assume to
5450 * have found the maximum and break out of the loop early:
5452 if (size_found && (size*100 > size_found*SIZE_THRESH))
5453 if (cost+avg_fluct <= 0 ||
5454 max_cost*100 > (cost+avg_fluct)*COST_THRESH) {
5456 if (migration_debug)
5457 printk("-> found max.\n");
5461 * Increase the cachesize in 5% steps:
5463 size = size * 20 / 19;
5466 if (migration_debug)
5467 printk("[%d][%d] working set size found: %d, cost: %Ld\n",
5468 cpu1, cpu2, size_found, max_cost);
5473 * A task is considered 'cache cold' if at least 2 times
5474 * the worst-case cost of migration has passed.
5476 * (this limit is only listened to if the load-balancing
5477 * situation is 'nice' - if there is a large imbalance we
5478 * ignore it for the sake of CPU utilization and
5479 * processing fairness.)
5481 return 2 * max_cost * migration_factor / MIGRATION_FACTOR_SCALE;
5484 static void calibrate_migration_costs(const cpumask_t *cpu_map)
5486 int cpu1 = -1, cpu2 = -1, cpu, orig_cpu = raw_smp_processor_id();
5487 unsigned long j0, j1, distance, max_distance = 0;
5488 struct sched_domain *sd;
5493 * First pass - calculate the cacheflush times:
5495 for_each_cpu_mask(cpu1, *cpu_map) {
5496 for_each_cpu_mask(cpu2, *cpu_map) {
5499 distance = domain_distance(cpu1, cpu2);
5500 max_distance = max(max_distance, distance);
5502 * No result cached yet?
5504 if (migration_cost[distance] == -1LL)
5505 migration_cost[distance] =
5506 measure_migration_cost(cpu1, cpu2);
5510 * Second pass - update the sched domain hierarchy with
5511 * the new cache-hot-time estimations:
5513 for_each_cpu_mask(cpu, *cpu_map) {
5515 for_each_domain(cpu, sd) {
5516 sd->cache_hot_time = migration_cost[distance];
5523 if (migration_debug)
5524 printk("migration: max_cache_size: %d, cpu: %d MHz:\n",
5532 printk("migration_cost=");
5533 for (distance = 0; distance <= max_distance; distance++) {
5536 printk("%ld", (long)migration_cost[distance] / 1000);
5540 if (migration_debug)
5541 printk("migration: %ld seconds\n", (j1-j0)/HZ);
5544 * Move back to the original CPU. NUMA-Q gets confused
5545 * if we migrate to another quad during bootup.
5547 if (raw_smp_processor_id() != orig_cpu) {
5548 cpumask_t mask = cpumask_of_cpu(orig_cpu),
5549 saved_mask = current->cpus_allowed;
5551 set_cpus_allowed(current, mask);
5552 set_cpus_allowed(current, saved_mask);
5559 * find_next_best_node - find the next node to include in a sched_domain
5560 * @node: node whose sched_domain we're building
5561 * @used_nodes: nodes already in the sched_domain
5563 * Find the next node to include in a given scheduling domain. Simply
5564 * finds the closest node not already in the @used_nodes map.
5566 * Should use nodemask_t.
5568 static int find_next_best_node(int node, unsigned long *used_nodes)
5570 int i, n, val, min_val, best_node = 0;
5574 for (i = 0; i < MAX_NUMNODES; i++) {
5575 /* Start at @node */
5576 n = (node + i) % MAX_NUMNODES;
5578 if (!nr_cpus_node(n))
5581 /* Skip already used nodes */
5582 if (test_bit(n, used_nodes))
5585 /* Simple min distance search */
5586 val = node_distance(node, n);
5588 if (val < min_val) {
5594 set_bit(best_node, used_nodes);
5599 * sched_domain_node_span - get a cpumask for a node's sched_domain
5600 * @node: node whose cpumask we're constructing
5601 * @size: number of nodes to include in this span
5603 * Given a node, construct a good cpumask for its sched_domain to span. It
5604 * should be one that prevents unnecessary balancing, but also spreads tasks
5607 static cpumask_t sched_domain_node_span(int node)
5610 cpumask_t span, nodemask;
5611 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5614 bitmap_zero(used_nodes, MAX_NUMNODES);
5616 nodemask = node_to_cpumask(node);
5617 cpus_or(span, span, nodemask);
5618 set_bit(node, used_nodes);
5620 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5621 int next_node = find_next_best_node(node, used_nodes);
5622 nodemask = node_to_cpumask(next_node);
5623 cpus_or(span, span, nodemask);
5631 * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we
5632 * can switch it on easily if needed.
5634 #ifdef CONFIG_SCHED_SMT
5635 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5636 static struct sched_group sched_group_cpus[NR_CPUS];
5637 static int cpu_to_cpu_group(int cpu)
5643 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5644 static struct sched_group sched_group_phys[NR_CPUS];
5645 static int cpu_to_phys_group(int cpu)
5647 #ifdef CONFIG_SCHED_SMT
5648 return first_cpu(cpu_sibling_map[cpu]);
5656 * The init_sched_build_groups can't handle what we want to do with node
5657 * groups, so roll our own. Now each node has its own list of groups which
5658 * gets dynamically allocated.
5660 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5661 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5663 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5664 static struct sched_group *sched_group_allnodes_bycpu[NR_CPUS];
5666 static int cpu_to_allnodes_group(int cpu)
5668 return cpu_to_node(cpu);
5673 * Build sched domains for a given set of cpus and attach the sched domains
5674 * to the individual cpus
5676 void build_sched_domains(const cpumask_t *cpu_map)
5680 struct sched_group **sched_group_nodes = NULL;
5681 struct sched_group *sched_group_allnodes = NULL;
5684 * Allocate the per-node list of sched groups
5686 sched_group_nodes = kmalloc(sizeof(struct sched_group*)*MAX_NUMNODES,
5688 if (!sched_group_nodes) {
5689 printk(KERN_WARNING "Can not alloc sched group node list\n");
5692 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
5696 * Set up domains for cpus specified by the cpu_map.
5698 for_each_cpu_mask(i, *cpu_map) {
5700 struct sched_domain *sd = NULL, *p;
5701 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
5703 cpus_and(nodemask, nodemask, *cpu_map);
5706 if (cpus_weight(*cpu_map)
5707 > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
5708 if (!sched_group_allnodes) {
5709 sched_group_allnodes
5710 = kmalloc(sizeof(struct sched_group)
5713 if (!sched_group_allnodes) {
5715 "Can not alloc allnodes sched group\n");
5718 sched_group_allnodes_bycpu[i]
5719 = sched_group_allnodes;
5721 sd = &per_cpu(allnodes_domains, i);
5722 *sd = SD_ALLNODES_INIT;
5723 sd->span = *cpu_map;
5724 group = cpu_to_allnodes_group(i);
5725 sd->groups = &sched_group_allnodes[group];
5730 sd = &per_cpu(node_domains, i);
5732 sd->span = sched_domain_node_span(cpu_to_node(i));
5734 cpus_and(sd->span, sd->span, *cpu_map);
5738 sd = &per_cpu(phys_domains, i);
5739 group = cpu_to_phys_group(i);
5741 sd->span = nodemask;
5743 sd->groups = &sched_group_phys[group];
5745 #ifdef CONFIG_SCHED_SMT
5747 sd = &per_cpu(cpu_domains, i);
5748 group = cpu_to_cpu_group(i);
5749 *sd = SD_SIBLING_INIT;
5750 sd->span = cpu_sibling_map[i];
5751 cpus_and(sd->span, sd->span, *cpu_map);
5753 sd->groups = &sched_group_cpus[group];
5757 #ifdef CONFIG_SCHED_SMT
5758 /* Set up CPU (sibling) groups */
5759 for_each_cpu_mask(i, *cpu_map) {
5760 cpumask_t this_sibling_map = cpu_sibling_map[i];
5761 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
5762 if (i != first_cpu(this_sibling_map))
5765 init_sched_build_groups(sched_group_cpus, this_sibling_map,
5770 /* Set up physical groups */
5771 for (i = 0; i < MAX_NUMNODES; i++) {
5772 cpumask_t nodemask = node_to_cpumask(i);
5774 cpus_and(nodemask, nodemask, *cpu_map);
5775 if (cpus_empty(nodemask))
5778 init_sched_build_groups(sched_group_phys, nodemask,
5779 &cpu_to_phys_group);
5783 /* Set up node groups */
5784 if (sched_group_allnodes)
5785 init_sched_build_groups(sched_group_allnodes, *cpu_map,
5786 &cpu_to_allnodes_group);
5788 for (i = 0; i < MAX_NUMNODES; i++) {
5789 /* Set up node groups */
5790 struct sched_group *sg, *prev;
5791 cpumask_t nodemask = node_to_cpumask(i);
5792 cpumask_t domainspan;
5793 cpumask_t covered = CPU_MASK_NONE;
5796 cpus_and(nodemask, nodemask, *cpu_map);
5797 if (cpus_empty(nodemask)) {
5798 sched_group_nodes[i] = NULL;
5802 domainspan = sched_domain_node_span(i);
5803 cpus_and(domainspan, domainspan, *cpu_map);
5805 sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL);
5806 sched_group_nodes[i] = sg;
5807 for_each_cpu_mask(j, nodemask) {
5808 struct sched_domain *sd;
5809 sd = &per_cpu(node_domains, j);
5811 if (sd->groups == NULL) {
5812 /* Turn off balancing if we have no groups */
5818 "Can not alloc domain group for node %d\n", i);
5822 sg->cpumask = nodemask;
5823 cpus_or(covered, covered, nodemask);
5826 for (j = 0; j < MAX_NUMNODES; j++) {
5827 cpumask_t tmp, notcovered;
5828 int n = (i + j) % MAX_NUMNODES;
5830 cpus_complement(notcovered, covered);
5831 cpus_and(tmp, notcovered, *cpu_map);
5832 cpus_and(tmp, tmp, domainspan);
5833 if (cpus_empty(tmp))
5836 nodemask = node_to_cpumask(n);
5837 cpus_and(tmp, tmp, nodemask);
5838 if (cpus_empty(tmp))
5841 sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL);
5844 "Can not alloc domain group for node %d\n", j);
5849 cpus_or(covered, covered, tmp);
5853 prev->next = sched_group_nodes[i];
5857 /* Calculate CPU power for physical packages and nodes */
5858 for_each_cpu_mask(i, *cpu_map) {
5860 struct sched_domain *sd;
5861 #ifdef CONFIG_SCHED_SMT
5862 sd = &per_cpu(cpu_domains, i);
5863 power = SCHED_LOAD_SCALE;
5864 sd->groups->cpu_power = power;
5867 sd = &per_cpu(phys_domains, i);
5868 power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
5869 (cpus_weight(sd->groups->cpumask)-1) / 10;
5870 sd->groups->cpu_power = power;
5873 sd = &per_cpu(allnodes_domains, i);
5875 power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
5876 (cpus_weight(sd->groups->cpumask)-1) / 10;
5877 sd->groups->cpu_power = power;
5883 for (i = 0; i < MAX_NUMNODES; i++) {
5884 struct sched_group *sg = sched_group_nodes[i];
5890 for_each_cpu_mask(j, sg->cpumask) {
5891 struct sched_domain *sd;
5894 sd = &per_cpu(phys_domains, j);
5895 if (j != first_cpu(sd->groups->cpumask)) {
5897 * Only add "power" once for each
5902 power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE *
5903 (cpus_weight(sd->groups->cpumask)-1) / 10;
5905 sg->cpu_power += power;
5908 if (sg != sched_group_nodes[i])
5913 /* Attach the domains */
5914 for_each_cpu_mask(i, *cpu_map) {
5915 struct sched_domain *sd;
5916 #ifdef CONFIG_SCHED_SMT
5917 sd = &per_cpu(cpu_domains, i);
5919 sd = &per_cpu(phys_domains, i);
5921 cpu_attach_domain(sd, i);
5924 * Tune cache-hot values:
5926 calibrate_migration_costs(cpu_map);
5929 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
5931 static void arch_init_sched_domains(const cpumask_t *cpu_map)
5933 cpumask_t cpu_default_map;
5936 * Setup mask for cpus without special case scheduling requirements.
5937 * For now this just excludes isolated cpus, but could be used to
5938 * exclude other special cases in the future.
5940 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
5942 build_sched_domains(&cpu_default_map);
5945 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
5951 for_each_cpu_mask(cpu, *cpu_map) {
5952 struct sched_group *sched_group_allnodes
5953 = sched_group_allnodes_bycpu[cpu];
5954 struct sched_group **sched_group_nodes
5955 = sched_group_nodes_bycpu[cpu];
5957 if (sched_group_allnodes) {
5958 kfree(sched_group_allnodes);
5959 sched_group_allnodes_bycpu[cpu] = NULL;
5962 if (!sched_group_nodes)
5965 for (i = 0; i < MAX_NUMNODES; i++) {
5966 cpumask_t nodemask = node_to_cpumask(i);
5967 struct sched_group *oldsg, *sg = sched_group_nodes[i];
5969 cpus_and(nodemask, nodemask, *cpu_map);
5970 if (cpus_empty(nodemask))
5980 if (oldsg != sched_group_nodes[i])
5983 kfree(sched_group_nodes);
5984 sched_group_nodes_bycpu[cpu] = NULL;
5990 * Detach sched domains from a group of cpus specified in cpu_map
5991 * These cpus will now be attached to the NULL domain
5993 static inline void detach_destroy_domains(const cpumask_t *cpu_map)
5997 for_each_cpu_mask(i, *cpu_map)
5998 cpu_attach_domain(NULL, i);
5999 synchronize_sched();
6000 arch_destroy_sched_domains(cpu_map);
6004 * Partition sched domains as specified by the cpumasks below.
6005 * This attaches all cpus from the cpumasks to the NULL domain,
6006 * waits for a RCU quiescent period, recalculates sched
6007 * domain information and then attaches them back to the
6008 * correct sched domains
6009 * Call with hotplug lock held
6011 void partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6013 cpumask_t change_map;
6015 cpus_and(*partition1, *partition1, cpu_online_map);
6016 cpus_and(*partition2, *partition2, cpu_online_map);
6017 cpus_or(change_map, *partition1, *partition2);
6019 /* Detach sched domains from all of the affected cpus */
6020 detach_destroy_domains(&change_map);
6021 if (!cpus_empty(*partition1))
6022 build_sched_domains(partition1);
6023 if (!cpus_empty(*partition2))
6024 build_sched_domains(partition2);
6027 #ifdef CONFIG_HOTPLUG_CPU
6029 * Force a reinitialization of the sched domains hierarchy. The domains
6030 * and groups cannot be updated in place without racing with the balancing
6031 * code, so we temporarily attach all running cpus to the NULL domain
6032 * which will prevent rebalancing while the sched domains are recalculated.
6034 static int update_sched_domains(struct notifier_block *nfb,
6035 unsigned long action, void *hcpu)
6038 case CPU_UP_PREPARE:
6039 case CPU_DOWN_PREPARE:
6040 detach_destroy_domains(&cpu_online_map);
6043 case CPU_UP_CANCELED:
6044 case CPU_DOWN_FAILED:
6048 * Fall through and re-initialise the domains.
6055 /* The hotplug lock is already held by cpu_up/cpu_down */
6056 arch_init_sched_domains(&cpu_online_map);
6062 void __init sched_init_smp(void)
6065 arch_init_sched_domains(&cpu_online_map);
6066 unlock_cpu_hotplug();
6067 /* XXX: Theoretical race here - CPU may be hotplugged now */
6068 hotcpu_notifier(update_sched_domains, 0);
6071 void __init sched_init_smp(void)
6074 #endif /* CONFIG_SMP */
6076 int in_sched_functions(unsigned long addr)
6078 /* Linker adds these: start and end of __sched functions */
6079 extern char __sched_text_start[], __sched_text_end[];
6080 return in_lock_functions(addr) ||
6081 (addr >= (unsigned long)__sched_text_start
6082 && addr < (unsigned long)__sched_text_end);
6085 void __init sched_init(void)
6090 for (i = 0; i < NR_CPUS; i++) {
6091 prio_array_t *array;
6094 spin_lock_init(&rq->lock);
6096 rq->active = rq->arrays;
6097 rq->expired = rq->arrays + 1;
6098 rq->best_expired_prio = MAX_PRIO;
6102 for (j = 1; j < 3; j++)
6103 rq->cpu_load[j] = 0;
6104 rq->active_balance = 0;
6106 rq->migration_thread = NULL;
6107 INIT_LIST_HEAD(&rq->migration_queue);
6109 atomic_set(&rq->nr_iowait, 0);
6111 for (j = 0; j < 2; j++) {
6112 array = rq->arrays + j;
6113 for (k = 0; k < MAX_PRIO; k++) {
6114 INIT_LIST_HEAD(array->queue + k);
6115 __clear_bit(k, array->bitmap);
6117 // delimiter for bitsearch
6118 __set_bit(MAX_PRIO, array->bitmap);
6123 * The boot idle thread does lazy MMU switching as well:
6125 atomic_inc(&init_mm.mm_count);
6126 enter_lazy_tlb(&init_mm, current);
6129 * Make us the idle thread. Technically, schedule() should not be
6130 * called from this thread, however somewhere below it might be,
6131 * but because we are the idle thread, we just pick up running again
6132 * when this runqueue becomes "idle".
6134 init_idle(current, smp_processor_id());
6137 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6138 void __might_sleep(char *file, int line)
6140 #if defined(in_atomic)
6141 static unsigned long prev_jiffy; /* ratelimiting */
6143 if ((in_atomic() || irqs_disabled()) &&
6144 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6145 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6147 prev_jiffy = jiffies;
6148 printk(KERN_ERR "Debug: sleeping function called from invalid"
6149 " context at %s:%d\n", file, line);
6150 printk("in_atomic():%d, irqs_disabled():%d\n",
6151 in_atomic(), irqs_disabled());
6156 EXPORT_SYMBOL(__might_sleep);
6159 #ifdef CONFIG_MAGIC_SYSRQ
6160 void normalize_rt_tasks(void)
6162 struct task_struct *p;
6163 prio_array_t *array;
6164 unsigned long flags;
6167 read_lock_irq(&tasklist_lock);
6168 for_each_process (p) {
6172 rq = task_rq_lock(p, &flags);
6176 deactivate_task(p, task_rq(p));
6177 __setscheduler(p, SCHED_NORMAL, 0);
6179 __activate_task(p, task_rq(p));
6180 resched_task(rq->curr);
6183 task_rq_unlock(rq, &flags);
6185 read_unlock_irq(&tasklist_lock);
6188 #endif /* CONFIG_MAGIC_SYSRQ */
6192 * These functions are only useful for the IA64 MCA handling.
6194 * They can only be called when the whole system has been
6195 * stopped - every CPU needs to be quiescent, and no scheduling
6196 * activity can take place. Using them for anything else would
6197 * be a serious bug, and as a result, they aren't even visible
6198 * under any other configuration.
6202 * curr_task - return the current task for a given cpu.
6203 * @cpu: the processor in question.
6205 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6207 task_t *curr_task(int cpu)
6209 return cpu_curr(cpu);
6213 * set_curr_task - set the current task for a given cpu.
6214 * @cpu: the processor in question.
6215 * @p: the task pointer to set.
6217 * Description: This function must only be used when non-maskable interrupts
6218 * are serviced on a separate stack. It allows the architecture to switch the
6219 * notion of the current task on a cpu in a non-blocking manner. This function
6220 * must be called with all CPU's synchronized, and interrupts disabled, the
6221 * and caller must save the original value of the current task (see
6222 * curr_task() above) and restore that value before reenabling interrupts and
6223 * re-starting the system.
6225 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6227 void set_curr_task(int cpu, task_t *p)