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>
52 #include <linux/kprobes.h>
55 #include <asm/unistd.h>
58 * Convert user-nice values [ -20 ... 0 ... 19 ]
59 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
62 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
63 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
64 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
67 * 'User priority' is the nice value converted to something we
68 * can work with better when scaling various scheduler parameters,
69 * it's a [ 0 ... 39 ] range.
71 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
72 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
73 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
76 * Some helpers for converting nanosecond timing to jiffy resolution
78 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
79 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
82 * These are the 'tuning knobs' of the scheduler:
84 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
85 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
86 * Timeslices get refilled after they expire.
88 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
89 #define DEF_TIMESLICE (100 * HZ / 1000)
90 #define ON_RUNQUEUE_WEIGHT 30
91 #define CHILD_PENALTY 95
92 #define PARENT_PENALTY 100
94 #define PRIO_BONUS_RATIO 25
95 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
96 #define INTERACTIVE_DELTA 2
97 #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
98 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
99 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
102 * If a task is 'interactive' then we reinsert it in the active
103 * array after it has expired its current timeslice. (it will not
104 * continue to run immediately, it will still roundrobin with
105 * other interactive tasks.)
107 * This part scales the interactivity limit depending on niceness.
109 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
110 * Here are a few examples of different nice levels:
112 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
113 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
114 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
115 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
116 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
118 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
119 * priority range a task can explore, a value of '1' means the
120 * task is rated interactive.)
122 * Ie. nice +19 tasks can never get 'interactive' enough to be
123 * reinserted into the active array. And only heavily CPU-hog nice -20
124 * tasks will be expired. Default nice 0 tasks are somewhere between,
125 * it takes some effort for them to get interactive, but it's not
129 #define CURRENT_BONUS(p) \
130 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
133 #define GRANULARITY (10 * HZ / 1000 ? : 1)
136 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
137 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
140 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
141 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
144 #define SCALE(v1,v1_max,v2_max) \
145 (v1) * (v2_max) / (v1_max)
148 (SCALE(TASK_NICE(p) + 20, 40, MAX_BONUS) - 20 * MAX_BONUS / 40 + \
151 #define TASK_INTERACTIVE(p) \
152 ((p)->prio <= (p)->static_prio - DELTA(p))
154 #define INTERACTIVE_SLEEP(p) \
155 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
156 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
158 #define TASK_PREEMPTS_CURR(p, rq) \
159 ((p)->prio < (rq)->curr->prio)
162 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
163 * to time slice values: [800ms ... 100ms ... 5ms]
165 * The higher a thread's priority, the bigger timeslices
166 * it gets during one round of execution. But even the lowest
167 * priority thread gets MIN_TIMESLICE worth of execution time.
170 #define SCALE_PRIO(x, prio) \
171 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO/2), MIN_TIMESLICE)
173 static unsigned int task_timeslice(task_t
*p
)
175 if (p
->static_prio
< NICE_TO_PRIO(0))
176 return SCALE_PRIO(DEF_TIMESLICE
*4, p
->static_prio
);
178 return SCALE_PRIO(DEF_TIMESLICE
, p
->static_prio
);
180 #define task_hot(p, now, sd) ((long long) ((now) - (p)->last_ran) \
181 < (long long) (sd)->cache_hot_time)
184 * These are the runqueue data structures:
187 #define BITMAP_SIZE ((((MAX_PRIO+1+7)/8)+sizeof(long)-1)/sizeof(long))
189 typedef struct runqueue runqueue_t
;
192 unsigned int nr_active
;
193 unsigned long bitmap
[BITMAP_SIZE
];
194 struct list_head queue
[MAX_PRIO
];
198 * This is the main, per-CPU runqueue data structure.
200 * Locking rule: those places that want to lock multiple runqueues
201 * (such as the load balancing or the thread migration code), lock
202 * acquire operations must be ordered by ascending &runqueue.
208 * nr_running and cpu_load should be in the same cacheline because
209 * remote CPUs use both these fields when doing load calculation.
211 unsigned long nr_running
;
213 unsigned long cpu_load
[3];
215 unsigned long long nr_switches
;
218 * This is part of a global counter where only the total sum
219 * over all CPUs matters. A task can increase this counter on
220 * one CPU and if it got migrated afterwards it may decrease
221 * it on another CPU. Always updated under the runqueue lock:
223 unsigned long nr_uninterruptible
;
225 unsigned long expired_timestamp
;
226 unsigned long long timestamp_last_tick
;
228 struct mm_struct
*prev_mm
;
229 prio_array_t
*active
, *expired
, arrays
[2];
230 int best_expired_prio
;
234 struct sched_domain
*sd
;
236 /* For active balancing */
240 task_t
*migration_thread
;
241 struct list_head migration_queue
;
245 #ifdef CONFIG_SCHEDSTATS
247 struct sched_info rq_sched_info
;
249 /* sys_sched_yield() stats */
250 unsigned long yld_exp_empty
;
251 unsigned long yld_act_empty
;
252 unsigned long yld_both_empty
;
253 unsigned long yld_cnt
;
255 /* schedule() stats */
256 unsigned long sched_switch
;
257 unsigned long sched_cnt
;
258 unsigned long sched_goidle
;
260 /* try_to_wake_up() stats */
261 unsigned long ttwu_cnt
;
262 unsigned long ttwu_local
;
266 static DEFINE_PER_CPU(struct runqueue
, runqueues
);
269 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
270 * See detach_destroy_domains: synchronize_sched for details.
272 * The domain tree of any CPU may only be accessed from within
273 * preempt-disabled sections.
275 #define for_each_domain(cpu, domain) \
276 for (domain = rcu_dereference(cpu_rq(cpu)->sd); domain; domain = domain->parent)
278 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
279 #define this_rq() (&__get_cpu_var(runqueues))
280 #define task_rq(p) cpu_rq(task_cpu(p))
281 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
283 #ifndef prepare_arch_switch
284 # define prepare_arch_switch(next) do { } while (0)
286 #ifndef finish_arch_switch
287 # define finish_arch_switch(prev) do { } while (0)
290 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
291 static inline int task_running(runqueue_t
*rq
, task_t
*p
)
293 return rq
->curr
== p
;
296 static inline void prepare_lock_switch(runqueue_t
*rq
, task_t
*next
)
300 static inline void finish_lock_switch(runqueue_t
*rq
, task_t
*prev
)
302 #ifdef CONFIG_DEBUG_SPINLOCK
303 /* this is a valid case when another task releases the spinlock */
304 rq
->lock
.owner
= current
;
306 spin_unlock_irq(&rq
->lock
);
309 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
310 static inline int task_running(runqueue_t
*rq
, task_t
*p
)
315 return rq
->curr
== p
;
319 static inline void prepare_lock_switch(runqueue_t
*rq
, task_t
*next
)
323 * We can optimise this out completely for !SMP, because the
324 * SMP rebalancing from interrupt is the only thing that cares
329 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
330 spin_unlock_irq(&rq
->lock
);
332 spin_unlock(&rq
->lock
);
336 static inline void finish_lock_switch(runqueue_t
*rq
, task_t
*prev
)
340 * After ->oncpu is cleared, the task can be moved to a different CPU.
341 * We must ensure this doesn't happen until the switch is completely
347 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
351 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
354 * task_rq_lock - lock the runqueue a given task resides on and disable
355 * interrupts. Note the ordering: we can safely lookup the task_rq without
356 * explicitly disabling preemption.
358 static inline runqueue_t
*task_rq_lock(task_t
*p
, unsigned long *flags
)
364 local_irq_save(*flags
);
366 spin_lock(&rq
->lock
);
367 if (unlikely(rq
!= task_rq(p
))) {
368 spin_unlock_irqrestore(&rq
->lock
, *flags
);
369 goto repeat_lock_task
;
374 static inline void task_rq_unlock(runqueue_t
*rq
, unsigned long *flags
)
377 spin_unlock_irqrestore(&rq
->lock
, *flags
);
380 #ifdef CONFIG_SCHEDSTATS
382 * bump this up when changing the output format or the meaning of an existing
383 * format, so that tools can adapt (or abort)
385 #define SCHEDSTAT_VERSION 12
387 static int show_schedstat(struct seq_file
*seq
, void *v
)
391 seq_printf(seq
, "version %d\n", SCHEDSTAT_VERSION
);
392 seq_printf(seq
, "timestamp %lu\n", jiffies
);
393 for_each_online_cpu(cpu
) {
394 runqueue_t
*rq
= cpu_rq(cpu
);
396 struct sched_domain
*sd
;
400 /* runqueue-specific stats */
402 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
403 cpu
, rq
->yld_both_empty
,
404 rq
->yld_act_empty
, rq
->yld_exp_empty
, rq
->yld_cnt
,
405 rq
->sched_switch
, rq
->sched_cnt
, rq
->sched_goidle
,
406 rq
->ttwu_cnt
, rq
->ttwu_local
,
407 rq
->rq_sched_info
.cpu_time
,
408 rq
->rq_sched_info
.run_delay
, rq
->rq_sched_info
.pcnt
);
410 seq_printf(seq
, "\n");
413 /* domain-specific stats */
415 for_each_domain(cpu
, sd
) {
416 enum idle_type itype
;
417 char mask_str
[NR_CPUS
];
419 cpumask_scnprintf(mask_str
, NR_CPUS
, sd
->span
);
420 seq_printf(seq
, "domain%d %s", dcnt
++, mask_str
);
421 for (itype
= SCHED_IDLE
; itype
< MAX_IDLE_TYPES
;
423 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu %lu",
425 sd
->lb_balanced
[itype
],
426 sd
->lb_failed
[itype
],
427 sd
->lb_imbalance
[itype
],
428 sd
->lb_gained
[itype
],
429 sd
->lb_hot_gained
[itype
],
430 sd
->lb_nobusyq
[itype
],
431 sd
->lb_nobusyg
[itype
]);
433 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
434 sd
->alb_cnt
, sd
->alb_failed
, sd
->alb_pushed
,
435 sd
->sbe_cnt
, sd
->sbe_balanced
, sd
->sbe_pushed
,
436 sd
->sbf_cnt
, sd
->sbf_balanced
, sd
->sbf_pushed
,
437 sd
->ttwu_wake_remote
, sd
->ttwu_move_affine
, sd
->ttwu_move_balance
);
445 static int schedstat_open(struct inode
*inode
, struct file
*file
)
447 unsigned int size
= PAGE_SIZE
* (1 + num_online_cpus() / 32);
448 char *buf
= kmalloc(size
, GFP_KERNEL
);
454 res
= single_open(file
, show_schedstat
, NULL
);
456 m
= file
->private_data
;
464 struct file_operations proc_schedstat_operations
= {
465 .open
= schedstat_open
,
468 .release
= single_release
,
471 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
472 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
473 #else /* !CONFIG_SCHEDSTATS */
474 # define schedstat_inc(rq, field) do { } while (0)
475 # define schedstat_add(rq, field, amt) do { } while (0)
479 * rq_lock - lock a given runqueue and disable interrupts.
481 static inline runqueue_t
*this_rq_lock(void)
488 spin_lock(&rq
->lock
);
493 #ifdef CONFIG_SCHEDSTATS
495 * Called when a process is dequeued from the active array and given
496 * the cpu. We should note that with the exception of interactive
497 * tasks, the expired queue will become the active queue after the active
498 * queue is empty, without explicitly dequeuing and requeuing tasks in the
499 * expired queue. (Interactive tasks may be requeued directly to the
500 * active queue, thus delaying tasks in the expired queue from running;
501 * see scheduler_tick()).
503 * This function is only called from sched_info_arrive(), rather than
504 * dequeue_task(). Even though a task may be queued and dequeued multiple
505 * times as it is shuffled about, we're really interested in knowing how
506 * long it was from the *first* time it was queued to the time that it
509 static inline void sched_info_dequeued(task_t
*t
)
511 t
->sched_info
.last_queued
= 0;
515 * Called when a task finally hits the cpu. We can now calculate how
516 * long it was waiting to run. We also note when it began so that we
517 * can keep stats on how long its timeslice is.
519 static void sched_info_arrive(task_t
*t
)
521 unsigned long now
= jiffies
, diff
= 0;
522 struct runqueue
*rq
= task_rq(t
);
524 if (t
->sched_info
.last_queued
)
525 diff
= now
- t
->sched_info
.last_queued
;
526 sched_info_dequeued(t
);
527 t
->sched_info
.run_delay
+= diff
;
528 t
->sched_info
.last_arrival
= now
;
529 t
->sched_info
.pcnt
++;
534 rq
->rq_sched_info
.run_delay
+= diff
;
535 rq
->rq_sched_info
.pcnt
++;
539 * Called when a process is queued into either the active or expired
540 * array. The time is noted and later used to determine how long we
541 * had to wait for us to reach the cpu. Since the expired queue will
542 * become the active queue after active queue is empty, without dequeuing
543 * and requeuing any tasks, we are interested in queuing to either. It
544 * is unusual but not impossible for tasks to be dequeued and immediately
545 * requeued in the same or another array: this can happen in sched_yield(),
546 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
549 * This function is only called from enqueue_task(), but also only updates
550 * the timestamp if it is already not set. It's assumed that
551 * sched_info_dequeued() will clear that stamp when appropriate.
553 static inline void sched_info_queued(task_t
*t
)
555 if (!t
->sched_info
.last_queued
)
556 t
->sched_info
.last_queued
= jiffies
;
560 * Called when a process ceases being the active-running process, either
561 * voluntarily or involuntarily. Now we can calculate how long we ran.
563 static inline void sched_info_depart(task_t
*t
)
565 struct runqueue
*rq
= task_rq(t
);
566 unsigned long diff
= jiffies
- t
->sched_info
.last_arrival
;
568 t
->sched_info
.cpu_time
+= diff
;
571 rq
->rq_sched_info
.cpu_time
+= diff
;
575 * Called when tasks are switched involuntarily due, typically, to expiring
576 * their time slice. (This may also be called when switching to or from
577 * the idle task.) We are only called when prev != next.
579 static inline void sched_info_switch(task_t
*prev
, task_t
*next
)
581 struct runqueue
*rq
= task_rq(prev
);
584 * prev now departs the cpu. It's not interesting to record
585 * stats about how efficient we were at scheduling the idle
588 if (prev
!= rq
->idle
)
589 sched_info_depart(prev
);
591 if (next
!= rq
->idle
)
592 sched_info_arrive(next
);
595 #define sched_info_queued(t) do { } while (0)
596 #define sched_info_switch(t, next) do { } while (0)
597 #endif /* CONFIG_SCHEDSTATS */
600 * Adding/removing a task to/from a priority array:
602 static void dequeue_task(struct task_struct
*p
, prio_array_t
*array
)
605 list_del(&p
->run_list
);
606 if (list_empty(array
->queue
+ p
->prio
))
607 __clear_bit(p
->prio
, array
->bitmap
);
610 static void enqueue_task(struct task_struct
*p
, prio_array_t
*array
)
612 sched_info_queued(p
);
613 list_add_tail(&p
->run_list
, array
->queue
+ p
->prio
);
614 __set_bit(p
->prio
, array
->bitmap
);
620 * Put task to the end of the run list without the overhead of dequeue
621 * followed by enqueue.
623 static void requeue_task(struct task_struct
*p
, prio_array_t
*array
)
625 list_move_tail(&p
->run_list
, array
->queue
+ p
->prio
);
628 static inline void enqueue_task_head(struct task_struct
*p
, prio_array_t
*array
)
630 list_add(&p
->run_list
, array
->queue
+ p
->prio
);
631 __set_bit(p
->prio
, array
->bitmap
);
637 * effective_prio - return the priority that is based on the static
638 * priority but is modified by bonuses/penalties.
640 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
641 * into the -5 ... 0 ... +5 bonus/penalty range.
643 * We use 25% of the full 0...39 priority range so that:
645 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
646 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
648 * Both properties are important to certain workloads.
650 static int effective_prio(task_t
*p
)
657 bonus
= CURRENT_BONUS(p
) - MAX_BONUS
/ 2;
659 prio
= p
->static_prio
- bonus
;
660 if (prio
< MAX_RT_PRIO
)
662 if (prio
> MAX_PRIO
-1)
668 * __activate_task - move a task to the runqueue.
670 static inline void __activate_task(task_t
*p
, runqueue_t
*rq
)
672 enqueue_task(p
, rq
->active
);
677 * __activate_idle_task - move idle task to the _front_ of runqueue.
679 static inline void __activate_idle_task(task_t
*p
, runqueue_t
*rq
)
681 enqueue_task_head(p
, rq
->active
);
685 static int recalc_task_prio(task_t
*p
, unsigned long long now
)
687 /* Caller must always ensure 'now >= p->timestamp' */
688 unsigned long long __sleep_time
= now
- p
->timestamp
;
689 unsigned long sleep_time
;
691 if (unlikely(p
->policy
== SCHED_BATCH
))
694 if (__sleep_time
> NS_MAX_SLEEP_AVG
)
695 sleep_time
= NS_MAX_SLEEP_AVG
;
697 sleep_time
= (unsigned long)__sleep_time
;
700 if (likely(sleep_time
> 0)) {
702 * User tasks that sleep a long time are categorised as
703 * idle and will get just interactive status to stay active &
704 * prevent them suddenly becoming cpu hogs and starving
707 if (p
->mm
&& p
->activated
!= -1 &&
708 sleep_time
> INTERACTIVE_SLEEP(p
)) {
709 p
->sleep_avg
= JIFFIES_TO_NS(MAX_SLEEP_AVG
-
713 * Tasks waking from uninterruptible sleep are
714 * limited in their sleep_avg rise as they
715 * are likely to be waiting on I/O
717 if (p
->activated
== -1 && p
->mm
) {
718 if (p
->sleep_avg
>= INTERACTIVE_SLEEP(p
))
720 else if (p
->sleep_avg
+ sleep_time
>=
721 INTERACTIVE_SLEEP(p
)) {
722 p
->sleep_avg
= INTERACTIVE_SLEEP(p
);
728 * This code gives a bonus to interactive tasks.
730 * The boost works by updating the 'average sleep time'
731 * value here, based on ->timestamp. The more time a
732 * task spends sleeping, the higher the average gets -
733 * and the higher the priority boost gets as well.
735 p
->sleep_avg
+= sleep_time
;
737 if (p
->sleep_avg
> NS_MAX_SLEEP_AVG
)
738 p
->sleep_avg
= NS_MAX_SLEEP_AVG
;
742 return effective_prio(p
);
746 * activate_task - move a task to the runqueue and do priority recalculation
748 * Update all the scheduling statistics stuff. (sleep average
749 * calculation, priority modifiers, etc.)
751 static void activate_task(task_t
*p
, runqueue_t
*rq
, int local
)
753 unsigned long long now
;
758 /* Compensate for drifting sched_clock */
759 runqueue_t
*this_rq
= this_rq();
760 now
= (now
- this_rq
->timestamp_last_tick
)
761 + rq
->timestamp_last_tick
;
766 p
->prio
= recalc_task_prio(p
, now
);
769 * This checks to make sure it's not an uninterruptible task
770 * that is now waking up.
774 * Tasks which were woken up by interrupts (ie. hw events)
775 * are most likely of interactive nature. So we give them
776 * the credit of extending their sleep time to the period
777 * of time they spend on the runqueue, waiting for execution
778 * on a CPU, first time around:
784 * Normal first-time wakeups get a credit too for
785 * on-runqueue time, but it will be weighted down:
792 __activate_task(p
, rq
);
796 * deactivate_task - remove a task from the runqueue.
798 static void deactivate_task(struct task_struct
*p
, runqueue_t
*rq
)
801 dequeue_task(p
, p
->array
);
806 * resched_task - mark a task 'to be rescheduled now'.
808 * On UP this means the setting of the need_resched flag, on SMP it
809 * might also involve a cross-CPU call to trigger the scheduler on
813 static void resched_task(task_t
*p
)
817 assert_spin_locked(&task_rq(p
)->lock
);
819 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
822 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
825 if (cpu
== smp_processor_id())
828 /* NEED_RESCHED must be visible before we test POLLING_NRFLAG */
830 if (!test_tsk_thread_flag(p
, TIF_POLLING_NRFLAG
))
831 smp_send_reschedule(cpu
);
834 static inline void resched_task(task_t
*p
)
836 assert_spin_locked(&task_rq(p
)->lock
);
837 set_tsk_need_resched(p
);
842 * task_curr - is this task currently executing on a CPU?
843 * @p: the task in question.
845 inline int task_curr(const task_t
*p
)
847 return cpu_curr(task_cpu(p
)) == p
;
852 struct list_head list
;
857 struct completion done
;
861 * The task's runqueue lock must be held.
862 * Returns true if you have to wait for migration thread.
864 static int migrate_task(task_t
*p
, int dest_cpu
, migration_req_t
*req
)
866 runqueue_t
*rq
= task_rq(p
);
869 * If the task is not on a runqueue (and not running), then
870 * it is sufficient to simply update the task's cpu field.
872 if (!p
->array
&& !task_running(rq
, p
)) {
873 set_task_cpu(p
, dest_cpu
);
877 init_completion(&req
->done
);
879 req
->dest_cpu
= dest_cpu
;
880 list_add(&req
->list
, &rq
->migration_queue
);
885 * wait_task_inactive - wait for a thread to unschedule.
887 * The caller must ensure that the task *will* unschedule sometime soon,
888 * else this function might spin for a *long* time. This function can't
889 * be called with interrupts off, or it may introduce deadlock with
890 * smp_call_function() if an IPI is sent by the same process we are
891 * waiting to become inactive.
893 void wait_task_inactive(task_t
*p
)
900 rq
= task_rq_lock(p
, &flags
);
901 /* Must be off runqueue entirely, not preempted. */
902 if (unlikely(p
->array
|| task_running(rq
, p
))) {
903 /* If it's preempted, we yield. It could be a while. */
904 preempted
= !task_running(rq
, p
);
905 task_rq_unlock(rq
, &flags
);
911 task_rq_unlock(rq
, &flags
);
915 * kick_process - kick a running thread to enter/exit the kernel
916 * @p: the to-be-kicked thread
918 * Cause a process which is running on another CPU to enter
919 * kernel-mode, without any delay. (to get signals handled.)
921 * NOTE: this function doesnt have to take the runqueue lock,
922 * because all it wants to ensure is that the remote task enters
923 * the kernel. If the IPI races and the task has been migrated
924 * to another CPU then no harm is done and the purpose has been
927 void kick_process(task_t
*p
)
933 if ((cpu
!= smp_processor_id()) && task_curr(p
))
934 smp_send_reschedule(cpu
);
939 * Return a low guess at the load of a migration-source cpu.
941 * We want to under-estimate the load of migration sources, to
942 * balance conservatively.
944 static inline unsigned long source_load(int cpu
, int type
)
946 runqueue_t
*rq
= cpu_rq(cpu
);
947 unsigned long load_now
= rq
->nr_running
* SCHED_LOAD_SCALE
;
951 return min(rq
->cpu_load
[type
-1], load_now
);
955 * Return a high guess at the load of a migration-target cpu
957 static inline unsigned long target_load(int cpu
, int type
)
959 runqueue_t
*rq
= cpu_rq(cpu
);
960 unsigned long load_now
= rq
->nr_running
* SCHED_LOAD_SCALE
;
964 return max(rq
->cpu_load
[type
-1], load_now
);
968 * find_idlest_group finds and returns the least busy CPU group within the
971 static struct sched_group
*
972 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
974 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
975 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
976 int load_idx
= sd
->forkexec_idx
;
977 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
980 unsigned long load
, avg_load
;
984 /* Skip over this group if it has no CPUs allowed */
985 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
988 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
990 /* Tally up the load of all CPUs in the group */
993 for_each_cpu_mask(i
, group
->cpumask
) {
994 /* Bias balancing toward cpus of our domain */
996 load
= source_load(i
, load_idx
);
998 load
= target_load(i
, load_idx
);
1003 /* Adjust by relative CPU power of the group */
1004 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
1007 this_load
= avg_load
;
1009 } else if (avg_load
< min_load
) {
1010 min_load
= avg_load
;
1014 group
= group
->next
;
1015 } while (group
!= sd
->groups
);
1017 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1023 * find_idlest_queue - find the idlest runqueue among the cpus in group.
1026 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1029 unsigned long load
, min_load
= ULONG_MAX
;
1033 /* Traverse only the allowed CPUs */
1034 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1036 for_each_cpu_mask(i
, tmp
) {
1037 load
= source_load(i
, 0);
1039 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1049 * sched_balance_self: balance the current task (running on cpu) in domains
1050 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1053 * Balance, ie. select the least loaded group.
1055 * Returns the target CPU number, or the same CPU if no balancing is needed.
1057 * preempt must be disabled.
1059 static int sched_balance_self(int cpu
, int flag
)
1061 struct task_struct
*t
= current
;
1062 struct sched_domain
*tmp
, *sd
= NULL
;
1064 for_each_domain(cpu
, tmp
)
1065 if (tmp
->flags
& flag
)
1070 struct sched_group
*group
;
1075 group
= find_idlest_group(sd
, t
, cpu
);
1079 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1080 if (new_cpu
== -1 || new_cpu
== cpu
)
1083 /* Now try balancing at a lower domain level */
1087 weight
= cpus_weight(span
);
1088 for_each_domain(cpu
, tmp
) {
1089 if (weight
<= cpus_weight(tmp
->span
))
1091 if (tmp
->flags
& flag
)
1094 /* while loop will break here if sd == NULL */
1100 #endif /* CONFIG_SMP */
1103 * wake_idle() will wake a task on an idle cpu if task->cpu is
1104 * not idle and an idle cpu is available. The span of cpus to
1105 * search starts with cpus closest then further out as needed,
1106 * so we always favor a closer, idle cpu.
1108 * Returns the CPU we should wake onto.
1110 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1111 static int wake_idle(int cpu
, task_t
*p
)
1114 struct sched_domain
*sd
;
1120 for_each_domain(cpu
, sd
) {
1121 if (sd
->flags
& SD_WAKE_IDLE
) {
1122 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1123 for_each_cpu_mask(i
, tmp
) {
1134 static inline int wake_idle(int cpu
, task_t
*p
)
1141 * try_to_wake_up - wake up a thread
1142 * @p: the to-be-woken-up thread
1143 * @state: the mask of task states that can be woken
1144 * @sync: do a synchronous wakeup?
1146 * Put it on the run-queue if it's not already there. The "current"
1147 * thread is always on the run-queue (except when the actual
1148 * re-schedule is in progress), and as such you're allowed to do
1149 * the simpler "current->state = TASK_RUNNING" to mark yourself
1150 * runnable without the overhead of this.
1152 * returns failure only if the task is already active.
1154 static int try_to_wake_up(task_t
*p
, unsigned int state
, int sync
)
1156 int cpu
, this_cpu
, success
= 0;
1157 unsigned long flags
;
1161 unsigned long load
, this_load
;
1162 struct sched_domain
*sd
, *this_sd
= NULL
;
1166 rq
= task_rq_lock(p
, &flags
);
1167 old_state
= p
->state
;
1168 if (!(old_state
& state
))
1175 this_cpu
= smp_processor_id();
1178 if (unlikely(task_running(rq
, p
)))
1183 schedstat_inc(rq
, ttwu_cnt
);
1184 if (cpu
== this_cpu
) {
1185 schedstat_inc(rq
, ttwu_local
);
1189 for_each_domain(this_cpu
, sd
) {
1190 if (cpu_isset(cpu
, sd
->span
)) {
1191 schedstat_inc(sd
, ttwu_wake_remote
);
1197 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1201 * Check for affine wakeup and passive balancing possibilities.
1204 int idx
= this_sd
->wake_idx
;
1205 unsigned int imbalance
;
1207 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1209 load
= source_load(cpu
, idx
);
1210 this_load
= target_load(this_cpu
, idx
);
1212 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1214 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1215 unsigned long tl
= this_load
;
1217 * If sync wakeup then subtract the (maximum possible)
1218 * effect of the currently running task from the load
1219 * of the current CPU:
1222 tl
-= SCHED_LOAD_SCALE
;
1225 tl
+ target_load(cpu
, idx
) <= SCHED_LOAD_SCALE
) ||
1226 100*(tl
+ SCHED_LOAD_SCALE
) <= imbalance
*load
) {
1228 * This domain has SD_WAKE_AFFINE and
1229 * p is cache cold in this domain, and
1230 * there is no bad imbalance.
1232 schedstat_inc(this_sd
, ttwu_move_affine
);
1238 * Start passive balancing when half the imbalance_pct
1241 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1242 if (imbalance
*this_load
<= 100*load
) {
1243 schedstat_inc(this_sd
, ttwu_move_balance
);
1249 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1251 new_cpu
= wake_idle(new_cpu
, p
);
1252 if (new_cpu
!= cpu
) {
1253 set_task_cpu(p
, new_cpu
);
1254 task_rq_unlock(rq
, &flags
);
1255 /* might preempt at this point */
1256 rq
= task_rq_lock(p
, &flags
);
1257 old_state
= p
->state
;
1258 if (!(old_state
& state
))
1263 this_cpu
= smp_processor_id();
1268 #endif /* CONFIG_SMP */
1269 if (old_state
== TASK_UNINTERRUPTIBLE
) {
1270 rq
->nr_uninterruptible
--;
1272 * Tasks on involuntary sleep don't earn
1273 * sleep_avg beyond just interactive state.
1279 * Tasks that have marked their sleep as noninteractive get
1280 * woken up without updating their sleep average. (i.e. their
1281 * sleep is handled in a priority-neutral manner, no priority
1282 * boost and no penalty.)
1284 if (old_state
& TASK_NONINTERACTIVE
)
1285 __activate_task(p
, rq
);
1287 activate_task(p
, rq
, cpu
== this_cpu
);
1289 * Sync wakeups (i.e. those types of wakeups where the waker
1290 * has indicated that it will leave the CPU in short order)
1291 * don't trigger a preemption, if the woken up task will run on
1292 * this cpu. (in this case the 'I will reschedule' promise of
1293 * the waker guarantees that the freshly woken up task is going
1294 * to be considered on this CPU.)
1296 if (!sync
|| cpu
!= this_cpu
) {
1297 if (TASK_PREEMPTS_CURR(p
, rq
))
1298 resched_task(rq
->curr
);
1303 p
->state
= TASK_RUNNING
;
1305 task_rq_unlock(rq
, &flags
);
1310 int fastcall
wake_up_process(task_t
*p
)
1312 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1313 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1316 EXPORT_SYMBOL(wake_up_process
);
1318 int fastcall
wake_up_state(task_t
*p
, unsigned int state
)
1320 return try_to_wake_up(p
, state
, 0);
1324 * Perform scheduler related setup for a newly forked process p.
1325 * p is forked by current.
1327 void fastcall
sched_fork(task_t
*p
, int clone_flags
)
1329 int cpu
= get_cpu();
1332 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1334 set_task_cpu(p
, cpu
);
1337 * We mark the process as running here, but have not actually
1338 * inserted it onto the runqueue yet. This guarantees that
1339 * nobody will actually run it, and a signal or other external
1340 * event cannot wake it up and insert it on the runqueue either.
1342 p
->state
= TASK_RUNNING
;
1343 INIT_LIST_HEAD(&p
->run_list
);
1345 #ifdef CONFIG_SCHEDSTATS
1346 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1348 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1351 #ifdef CONFIG_PREEMPT
1352 /* Want to start with kernel preemption disabled. */
1353 task_thread_info(p
)->preempt_count
= 1;
1356 * Share the timeslice between parent and child, thus the
1357 * total amount of pending timeslices in the system doesn't change,
1358 * resulting in more scheduling fairness.
1360 local_irq_disable();
1361 p
->time_slice
= (current
->time_slice
+ 1) >> 1;
1363 * The remainder of the first timeslice might be recovered by
1364 * the parent if the child exits early enough.
1366 p
->first_time_slice
= 1;
1367 current
->time_slice
>>= 1;
1368 p
->timestamp
= sched_clock();
1369 if (unlikely(!current
->time_slice
)) {
1371 * This case is rare, it happens when the parent has only
1372 * a single jiffy left from its timeslice. Taking the
1373 * runqueue lock is not a problem.
1375 current
->time_slice
= 1;
1383 * wake_up_new_task - wake up a newly created task for the first time.
1385 * This function will do some initial scheduler statistics housekeeping
1386 * that must be done for every newly created context, then puts the task
1387 * on the runqueue and wakes it.
1389 void fastcall
wake_up_new_task(task_t
*p
, unsigned long clone_flags
)
1391 unsigned long flags
;
1393 runqueue_t
*rq
, *this_rq
;
1395 rq
= task_rq_lock(p
, &flags
);
1396 BUG_ON(p
->state
!= TASK_RUNNING
);
1397 this_cpu
= smp_processor_id();
1401 * We decrease the sleep average of forking parents
1402 * and children as well, to keep max-interactive tasks
1403 * from forking tasks that are max-interactive. The parent
1404 * (current) is done further down, under its lock.
1406 p
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(p
) *
1407 CHILD_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1409 p
->prio
= effective_prio(p
);
1411 if (likely(cpu
== this_cpu
)) {
1412 if (!(clone_flags
& CLONE_VM
)) {
1414 * The VM isn't cloned, so we're in a good position to
1415 * do child-runs-first in anticipation of an exec. This
1416 * usually avoids a lot of COW overhead.
1418 if (unlikely(!current
->array
))
1419 __activate_task(p
, rq
);
1421 p
->prio
= current
->prio
;
1422 list_add_tail(&p
->run_list
, ¤t
->run_list
);
1423 p
->array
= current
->array
;
1424 p
->array
->nr_active
++;
1429 /* Run child last */
1430 __activate_task(p
, rq
);
1432 * We skip the following code due to cpu == this_cpu
1434 * task_rq_unlock(rq, &flags);
1435 * this_rq = task_rq_lock(current, &flags);
1439 this_rq
= cpu_rq(this_cpu
);
1442 * Not the local CPU - must adjust timestamp. This should
1443 * get optimised away in the !CONFIG_SMP case.
1445 p
->timestamp
= (p
->timestamp
- this_rq
->timestamp_last_tick
)
1446 + rq
->timestamp_last_tick
;
1447 __activate_task(p
, rq
);
1448 if (TASK_PREEMPTS_CURR(p
, rq
))
1449 resched_task(rq
->curr
);
1452 * Parent and child are on different CPUs, now get the
1453 * parent runqueue to update the parent's ->sleep_avg:
1455 task_rq_unlock(rq
, &flags
);
1456 this_rq
= task_rq_lock(current
, &flags
);
1458 current
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(current
) *
1459 PARENT_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1460 task_rq_unlock(this_rq
, &flags
);
1464 * Potentially available exiting-child timeslices are
1465 * retrieved here - this way the parent does not get
1466 * penalized for creating too many threads.
1468 * (this cannot be used to 'generate' timeslices
1469 * artificially, because any timeslice recovered here
1470 * was given away by the parent in the first place.)
1472 void fastcall
sched_exit(task_t
*p
)
1474 unsigned long flags
;
1478 * If the child was a (relative-) CPU hog then decrease
1479 * the sleep_avg of the parent as well.
1481 rq
= task_rq_lock(p
->parent
, &flags
);
1482 if (p
->first_time_slice
&& task_cpu(p
) == task_cpu(p
->parent
)) {
1483 p
->parent
->time_slice
+= p
->time_slice
;
1484 if (unlikely(p
->parent
->time_slice
> task_timeslice(p
)))
1485 p
->parent
->time_slice
= task_timeslice(p
);
1487 if (p
->sleep_avg
< p
->parent
->sleep_avg
)
1488 p
->parent
->sleep_avg
= p
->parent
->sleep_avg
/
1489 (EXIT_WEIGHT
+ 1) * EXIT_WEIGHT
+ p
->sleep_avg
/
1491 task_rq_unlock(rq
, &flags
);
1495 * prepare_task_switch - prepare to switch tasks
1496 * @rq: the runqueue preparing to switch
1497 * @next: the task we are going to switch to.
1499 * This is called with the rq lock held and interrupts off. It must
1500 * be paired with a subsequent finish_task_switch after the context
1503 * prepare_task_switch sets up locking and calls architecture specific
1506 static inline void prepare_task_switch(runqueue_t
*rq
, task_t
*next
)
1508 prepare_lock_switch(rq
, next
);
1509 prepare_arch_switch(next
);
1513 * finish_task_switch - clean up after a task-switch
1514 * @rq: runqueue associated with task-switch
1515 * @prev: the thread we just switched away from.
1517 * finish_task_switch must be called after the context switch, paired
1518 * with a prepare_task_switch call before the context switch.
1519 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1520 * and do any other architecture-specific cleanup actions.
1522 * Note that we may have delayed dropping an mm in context_switch(). If
1523 * so, we finish that here outside of the runqueue lock. (Doing it
1524 * with the lock held can cause deadlocks; see schedule() for
1527 static inline void finish_task_switch(runqueue_t
*rq
, task_t
*prev
)
1528 __releases(rq
->lock
)
1530 struct mm_struct
*mm
= rq
->prev_mm
;
1531 unsigned long prev_task_flags
;
1536 * A task struct has one reference for the use as "current".
1537 * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1538 * calls schedule one last time. The schedule call will never return,
1539 * and the scheduled task must drop that reference.
1540 * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1541 * still held, otherwise prev could be scheduled on another cpu, die
1542 * there before we look at prev->state, and then the reference would
1544 * Manfred Spraul <manfred@colorfullife.com>
1546 prev_task_flags
= prev
->flags
;
1547 finish_arch_switch(prev
);
1548 finish_lock_switch(rq
, prev
);
1551 if (unlikely(prev_task_flags
& PF_DEAD
)) {
1553 * Remove function-return probe instances associated with this
1554 * task and put them back on the free list.
1556 kprobe_flush_task(prev
);
1557 put_task_struct(prev
);
1562 * schedule_tail - first thing a freshly forked thread must call.
1563 * @prev: the thread we just switched away from.
1565 asmlinkage
void schedule_tail(task_t
*prev
)
1566 __releases(rq
->lock
)
1568 runqueue_t
*rq
= this_rq();
1569 finish_task_switch(rq
, prev
);
1570 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1571 /* In this case, finish_task_switch does not reenable preemption */
1574 if (current
->set_child_tid
)
1575 put_user(current
->pid
, current
->set_child_tid
);
1579 * context_switch - switch to the new MM and the new
1580 * thread's register state.
1583 task_t
* context_switch(runqueue_t
*rq
, task_t
*prev
, task_t
*next
)
1585 struct mm_struct
*mm
= next
->mm
;
1586 struct mm_struct
*oldmm
= prev
->active_mm
;
1588 if (unlikely(!mm
)) {
1589 next
->active_mm
= oldmm
;
1590 atomic_inc(&oldmm
->mm_count
);
1591 enter_lazy_tlb(oldmm
, next
);
1593 switch_mm(oldmm
, mm
, next
);
1595 if (unlikely(!prev
->mm
)) {
1596 prev
->active_mm
= NULL
;
1597 WARN_ON(rq
->prev_mm
);
1598 rq
->prev_mm
= oldmm
;
1601 /* Here we just switch the register state and the stack. */
1602 switch_to(prev
, next
, prev
);
1608 * nr_running, nr_uninterruptible and nr_context_switches:
1610 * externally visible scheduler statistics: current number of runnable
1611 * threads, current number of uninterruptible-sleeping threads, total
1612 * number of context switches performed since bootup.
1614 unsigned long nr_running(void)
1616 unsigned long i
, sum
= 0;
1618 for_each_online_cpu(i
)
1619 sum
+= cpu_rq(i
)->nr_running
;
1624 unsigned long nr_uninterruptible(void)
1626 unsigned long i
, sum
= 0;
1628 for_each_possible_cpu(i
)
1629 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1632 * Since we read the counters lockless, it might be slightly
1633 * inaccurate. Do not allow it to go below zero though:
1635 if (unlikely((long)sum
< 0))
1641 unsigned long long nr_context_switches(void)
1643 unsigned long long i
, sum
= 0;
1645 for_each_possible_cpu(i
)
1646 sum
+= cpu_rq(i
)->nr_switches
;
1651 unsigned long nr_iowait(void)
1653 unsigned long i
, sum
= 0;
1655 for_each_possible_cpu(i
)
1656 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1664 * double_rq_lock - safely lock two runqueues
1666 * We must take them in cpu order to match code in
1667 * dependent_sleeper and wake_dependent_sleeper.
1669 * Note this does not disable interrupts like task_rq_lock,
1670 * you need to do so manually before calling.
1672 static void double_rq_lock(runqueue_t
*rq1
, runqueue_t
*rq2
)
1673 __acquires(rq1
->lock
)
1674 __acquires(rq2
->lock
)
1677 spin_lock(&rq1
->lock
);
1678 __acquire(rq2
->lock
); /* Fake it out ;) */
1680 if (rq1
->cpu
< rq2
->cpu
) {
1681 spin_lock(&rq1
->lock
);
1682 spin_lock(&rq2
->lock
);
1684 spin_lock(&rq2
->lock
);
1685 spin_lock(&rq1
->lock
);
1691 * double_rq_unlock - safely unlock two runqueues
1693 * Note this does not restore interrupts like task_rq_unlock,
1694 * you need to do so manually after calling.
1696 static void double_rq_unlock(runqueue_t
*rq1
, runqueue_t
*rq2
)
1697 __releases(rq1
->lock
)
1698 __releases(rq2
->lock
)
1700 spin_unlock(&rq1
->lock
);
1702 spin_unlock(&rq2
->lock
);
1704 __release(rq2
->lock
);
1708 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1710 static void double_lock_balance(runqueue_t
*this_rq
, runqueue_t
*busiest
)
1711 __releases(this_rq
->lock
)
1712 __acquires(busiest
->lock
)
1713 __acquires(this_rq
->lock
)
1715 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1716 if (busiest
->cpu
< this_rq
->cpu
) {
1717 spin_unlock(&this_rq
->lock
);
1718 spin_lock(&busiest
->lock
);
1719 spin_lock(&this_rq
->lock
);
1721 spin_lock(&busiest
->lock
);
1726 * If dest_cpu is allowed for this process, migrate the task to it.
1727 * This is accomplished by forcing the cpu_allowed mask to only
1728 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
1729 * the cpu_allowed mask is restored.
1731 static void sched_migrate_task(task_t
*p
, int dest_cpu
)
1733 migration_req_t req
;
1735 unsigned long flags
;
1737 rq
= task_rq_lock(p
, &flags
);
1738 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
1739 || unlikely(cpu_is_offline(dest_cpu
)))
1742 /* force the process onto the specified CPU */
1743 if (migrate_task(p
, dest_cpu
, &req
)) {
1744 /* Need to wait for migration thread (might exit: take ref). */
1745 struct task_struct
*mt
= rq
->migration_thread
;
1746 get_task_struct(mt
);
1747 task_rq_unlock(rq
, &flags
);
1748 wake_up_process(mt
);
1749 put_task_struct(mt
);
1750 wait_for_completion(&req
.done
);
1754 task_rq_unlock(rq
, &flags
);
1758 * sched_exec - execve() is a valuable balancing opportunity, because at
1759 * this point the task has the smallest effective memory and cache footprint.
1761 void sched_exec(void)
1763 int new_cpu
, this_cpu
= get_cpu();
1764 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
1766 if (new_cpu
!= this_cpu
)
1767 sched_migrate_task(current
, new_cpu
);
1771 * pull_task - move a task from a remote runqueue to the local runqueue.
1772 * Both runqueues must be locked.
1775 void pull_task(runqueue_t
*src_rq
, prio_array_t
*src_array
, task_t
*p
,
1776 runqueue_t
*this_rq
, prio_array_t
*this_array
, int this_cpu
)
1778 dequeue_task(p
, src_array
);
1779 src_rq
->nr_running
--;
1780 set_task_cpu(p
, this_cpu
);
1781 this_rq
->nr_running
++;
1782 enqueue_task(p
, this_array
);
1783 p
->timestamp
= (p
->timestamp
- src_rq
->timestamp_last_tick
)
1784 + this_rq
->timestamp_last_tick
;
1786 * Note that idle threads have a prio of MAX_PRIO, for this test
1787 * to be always true for them.
1789 if (TASK_PREEMPTS_CURR(p
, this_rq
))
1790 resched_task(this_rq
->curr
);
1794 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1797 int can_migrate_task(task_t
*p
, runqueue_t
*rq
, int this_cpu
,
1798 struct sched_domain
*sd
, enum idle_type idle
,
1802 * We do not migrate tasks that are:
1803 * 1) running (obviously), or
1804 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1805 * 3) are cache-hot on their current CPU.
1807 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
1811 if (task_running(rq
, p
))
1815 * Aggressive migration if:
1816 * 1) task is cache cold, or
1817 * 2) too many balance attempts have failed.
1820 if (sd
->nr_balance_failed
> sd
->cache_nice_tries
)
1823 if (task_hot(p
, rq
->timestamp_last_tick
, sd
))
1829 * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq,
1830 * as part of a balancing operation within "domain". Returns the number of
1833 * Called with both runqueues locked.
1835 static int move_tasks(runqueue_t
*this_rq
, int this_cpu
, runqueue_t
*busiest
,
1836 unsigned long max_nr_move
, struct sched_domain
*sd
,
1837 enum idle_type idle
, int *all_pinned
)
1839 prio_array_t
*array
, *dst_array
;
1840 struct list_head
*head
, *curr
;
1841 int idx
, pulled
= 0, pinned
= 0;
1844 if (max_nr_move
== 0)
1850 * We first consider expired tasks. Those will likely not be
1851 * executed in the near future, and they are most likely to
1852 * be cache-cold, thus switching CPUs has the least effect
1855 if (busiest
->expired
->nr_active
) {
1856 array
= busiest
->expired
;
1857 dst_array
= this_rq
->expired
;
1859 array
= busiest
->active
;
1860 dst_array
= this_rq
->active
;
1864 /* Start searching at priority 0: */
1868 idx
= sched_find_first_bit(array
->bitmap
);
1870 idx
= find_next_bit(array
->bitmap
, MAX_PRIO
, idx
);
1871 if (idx
>= MAX_PRIO
) {
1872 if (array
== busiest
->expired
&& busiest
->active
->nr_active
) {
1873 array
= busiest
->active
;
1874 dst_array
= this_rq
->active
;
1880 head
= array
->queue
+ idx
;
1883 tmp
= list_entry(curr
, task_t
, run_list
);
1887 if (!can_migrate_task(tmp
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
1894 #ifdef CONFIG_SCHEDSTATS
1895 if (task_hot(tmp
, busiest
->timestamp_last_tick
, sd
))
1896 schedstat_inc(sd
, lb_hot_gained
[idle
]);
1899 pull_task(busiest
, array
, tmp
, this_rq
, dst_array
, this_cpu
);
1902 /* We only want to steal up to the prescribed number of tasks. */
1903 if (pulled
< max_nr_move
) {
1911 * Right now, this is the only place pull_task() is called,
1912 * so we can safely collect pull_task() stats here rather than
1913 * inside pull_task().
1915 schedstat_add(sd
, lb_gained
[idle
], pulled
);
1918 *all_pinned
= pinned
;
1923 * find_busiest_group finds and returns the busiest CPU group within the
1924 * domain. It calculates and returns the number of tasks which should be
1925 * moved to restore balance via the imbalance parameter.
1927 static struct sched_group
*
1928 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
1929 unsigned long *imbalance
, enum idle_type idle
, int *sd_idle
)
1931 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1932 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
1933 unsigned long max_pull
;
1936 max_load
= this_load
= total_load
= total_pwr
= 0;
1937 if (idle
== NOT_IDLE
)
1938 load_idx
= sd
->busy_idx
;
1939 else if (idle
== NEWLY_IDLE
)
1940 load_idx
= sd
->newidle_idx
;
1942 load_idx
= sd
->idle_idx
;
1949 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1951 /* Tally up the load of all CPUs in the group */
1954 for_each_cpu_mask(i
, group
->cpumask
) {
1955 if (*sd_idle
&& !idle_cpu(i
))
1958 /* Bias balancing toward cpus of our domain */
1960 load
= target_load(i
, load_idx
);
1962 load
= source_load(i
, load_idx
);
1967 total_load
+= avg_load
;
1968 total_pwr
+= group
->cpu_power
;
1970 /* Adjust by relative CPU power of the group */
1971 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
1974 this_load
= avg_load
;
1976 } else if (avg_load
> max_load
) {
1977 max_load
= avg_load
;
1980 group
= group
->next
;
1981 } while (group
!= sd
->groups
);
1983 if (!busiest
|| this_load
>= max_load
|| max_load
<= SCHED_LOAD_SCALE
)
1986 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
1988 if (this_load
>= avg_load
||
1989 100*max_load
<= sd
->imbalance_pct
*this_load
)
1993 * We're trying to get all the cpus to the average_load, so we don't
1994 * want to push ourselves above the average load, nor do we wish to
1995 * reduce the max loaded cpu below the average load, as either of these
1996 * actions would just result in more rebalancing later, and ping-pong
1997 * tasks around. Thus we look for the minimum possible imbalance.
1998 * Negative imbalances (*we* are more loaded than anyone else) will
1999 * be counted as no imbalance for these purposes -- we can't fix that
2000 * by pulling tasks to us. Be careful of negative numbers as they'll
2001 * appear as very large values with unsigned longs.
2004 /* Don't want to pull so many tasks that a group would go idle */
2005 max_pull
= min(max_load
- avg_load
, max_load
- SCHED_LOAD_SCALE
);
2007 /* How much load to actually move to equalise the imbalance */
2008 *imbalance
= min(max_pull
* busiest
->cpu_power
,
2009 (avg_load
- this_load
) * this->cpu_power
)
2012 if (*imbalance
< SCHED_LOAD_SCALE
) {
2013 unsigned long pwr_now
= 0, pwr_move
= 0;
2016 if (max_load
- this_load
>= SCHED_LOAD_SCALE
*2) {
2022 * OK, we don't have enough imbalance to justify moving tasks,
2023 * however we may be able to increase total CPU power used by
2027 pwr_now
+= busiest
->cpu_power
*min(SCHED_LOAD_SCALE
, max_load
);
2028 pwr_now
+= this->cpu_power
*min(SCHED_LOAD_SCALE
, this_load
);
2029 pwr_now
/= SCHED_LOAD_SCALE
;
2031 /* Amount of load we'd subtract */
2032 tmp
= SCHED_LOAD_SCALE
*SCHED_LOAD_SCALE
/busiest
->cpu_power
;
2034 pwr_move
+= busiest
->cpu_power
*min(SCHED_LOAD_SCALE
,
2037 /* Amount of load we'd add */
2038 if (max_load
*busiest
->cpu_power
<
2039 SCHED_LOAD_SCALE
*SCHED_LOAD_SCALE
)
2040 tmp
= max_load
*busiest
->cpu_power
/this->cpu_power
;
2042 tmp
= SCHED_LOAD_SCALE
*SCHED_LOAD_SCALE
/this->cpu_power
;
2043 pwr_move
+= this->cpu_power
*min(SCHED_LOAD_SCALE
, this_load
+ tmp
);
2044 pwr_move
/= SCHED_LOAD_SCALE
;
2046 /* Move if we gain throughput */
2047 if (pwr_move
<= pwr_now
)
2054 /* Get rid of the scaling factor, rounding down as we divide */
2055 *imbalance
= *imbalance
/ SCHED_LOAD_SCALE
;
2065 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2067 static runqueue_t
*find_busiest_queue(struct sched_group
*group
,
2068 enum idle_type idle
)
2070 unsigned long load
, max_load
= 0;
2071 runqueue_t
*busiest
= NULL
;
2074 for_each_cpu_mask(i
, group
->cpumask
) {
2075 load
= source_load(i
, 0);
2077 if (load
> max_load
) {
2079 busiest
= cpu_rq(i
);
2087 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2088 * so long as it is large enough.
2090 #define MAX_PINNED_INTERVAL 512
2093 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2094 * tasks if there is an imbalance.
2096 * Called with this_rq unlocked.
2098 static int load_balance(int this_cpu
, runqueue_t
*this_rq
,
2099 struct sched_domain
*sd
, enum idle_type idle
)
2101 struct sched_group
*group
;
2102 runqueue_t
*busiest
;
2103 unsigned long imbalance
;
2104 int nr_moved
, all_pinned
= 0;
2105 int active_balance
= 0;
2108 if (idle
!= NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2111 schedstat_inc(sd
, lb_cnt
[idle
]);
2113 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
);
2115 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2119 busiest
= find_busiest_queue(group
, idle
);
2121 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2125 BUG_ON(busiest
== this_rq
);
2127 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2130 if (busiest
->nr_running
> 1) {
2132 * Attempt to move tasks. If find_busiest_group has found
2133 * an imbalance but busiest->nr_running <= 1, the group is
2134 * still unbalanced. nr_moved simply stays zero, so it is
2135 * correctly treated as an imbalance.
2137 double_rq_lock(this_rq
, busiest
);
2138 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2139 imbalance
, sd
, idle
, &all_pinned
);
2140 double_rq_unlock(this_rq
, busiest
);
2142 /* All tasks on this runqueue were pinned by CPU affinity */
2143 if (unlikely(all_pinned
))
2148 schedstat_inc(sd
, lb_failed
[idle
]);
2149 sd
->nr_balance_failed
++;
2151 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2153 spin_lock(&busiest
->lock
);
2155 /* don't kick the migration_thread, if the curr
2156 * task on busiest cpu can't be moved to this_cpu
2158 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2159 spin_unlock(&busiest
->lock
);
2161 goto out_one_pinned
;
2164 if (!busiest
->active_balance
) {
2165 busiest
->active_balance
= 1;
2166 busiest
->push_cpu
= this_cpu
;
2169 spin_unlock(&busiest
->lock
);
2171 wake_up_process(busiest
->migration_thread
);
2174 * We've kicked active balancing, reset the failure
2177 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2180 sd
->nr_balance_failed
= 0;
2182 if (likely(!active_balance
)) {
2183 /* We were unbalanced, so reset the balancing interval */
2184 sd
->balance_interval
= sd
->min_interval
;
2187 * If we've begun active balancing, start to back off. This
2188 * case may not be covered by the all_pinned logic if there
2189 * is only 1 task on the busy runqueue (because we don't call
2192 if (sd
->balance_interval
< sd
->max_interval
)
2193 sd
->balance_interval
*= 2;
2196 if (!nr_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2201 schedstat_inc(sd
, lb_balanced
[idle
]);
2203 sd
->nr_balance_failed
= 0;
2206 /* tune up the balancing interval */
2207 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2208 (sd
->balance_interval
< sd
->max_interval
))
2209 sd
->balance_interval
*= 2;
2211 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2217 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2218 * tasks if there is an imbalance.
2220 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2221 * this_rq is locked.
2223 static int load_balance_newidle(int this_cpu
, runqueue_t
*this_rq
,
2224 struct sched_domain
*sd
)
2226 struct sched_group
*group
;
2227 runqueue_t
*busiest
= NULL
;
2228 unsigned long imbalance
;
2232 if (sd
->flags
& SD_SHARE_CPUPOWER
)
2235 schedstat_inc(sd
, lb_cnt
[NEWLY_IDLE
]);
2236 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, NEWLY_IDLE
, &sd_idle
);
2238 schedstat_inc(sd
, lb_nobusyg
[NEWLY_IDLE
]);
2242 busiest
= find_busiest_queue(group
, NEWLY_IDLE
);
2244 schedstat_inc(sd
, lb_nobusyq
[NEWLY_IDLE
]);
2248 BUG_ON(busiest
== this_rq
);
2250 schedstat_add(sd
, lb_imbalance
[NEWLY_IDLE
], imbalance
);
2253 if (busiest
->nr_running
> 1) {
2254 /* Attempt to move tasks */
2255 double_lock_balance(this_rq
, busiest
);
2256 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2257 imbalance
, sd
, NEWLY_IDLE
, NULL
);
2258 spin_unlock(&busiest
->lock
);
2262 schedstat_inc(sd
, lb_failed
[NEWLY_IDLE
]);
2263 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2266 sd
->nr_balance_failed
= 0;
2271 schedstat_inc(sd
, lb_balanced
[NEWLY_IDLE
]);
2272 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2274 sd
->nr_balance_failed
= 0;
2279 * idle_balance is called by schedule() if this_cpu is about to become
2280 * idle. Attempts to pull tasks from other CPUs.
2282 static void idle_balance(int this_cpu
, runqueue_t
*this_rq
)
2284 struct sched_domain
*sd
;
2286 for_each_domain(this_cpu
, sd
) {
2287 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
2288 if (load_balance_newidle(this_cpu
, this_rq
, sd
)) {
2289 /* We've pulled tasks over so stop searching */
2297 * active_load_balance is run by migration threads. It pushes running tasks
2298 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2299 * running on each physical CPU where possible, and avoids physical /
2300 * logical imbalances.
2302 * Called with busiest_rq locked.
2304 static void active_load_balance(runqueue_t
*busiest_rq
, int busiest_cpu
)
2306 struct sched_domain
*sd
;
2307 runqueue_t
*target_rq
;
2308 int target_cpu
= busiest_rq
->push_cpu
;
2310 if (busiest_rq
->nr_running
<= 1)
2311 /* no task to move */
2314 target_rq
= cpu_rq(target_cpu
);
2317 * This condition is "impossible", if it occurs
2318 * we need to fix it. Originally reported by
2319 * Bjorn Helgaas on a 128-cpu setup.
2321 BUG_ON(busiest_rq
== target_rq
);
2323 /* move a task from busiest_rq to target_rq */
2324 double_lock_balance(busiest_rq
, target_rq
);
2326 /* Search for an sd spanning us and the target CPU. */
2327 for_each_domain(target_cpu
, sd
)
2328 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2329 cpu_isset(busiest_cpu
, sd
->span
))
2332 if (unlikely(sd
== NULL
))
2335 schedstat_inc(sd
, alb_cnt
);
2337 if (move_tasks(target_rq
, target_cpu
, busiest_rq
, 1, sd
, SCHED_IDLE
, NULL
))
2338 schedstat_inc(sd
, alb_pushed
);
2340 schedstat_inc(sd
, alb_failed
);
2342 spin_unlock(&target_rq
->lock
);
2346 * rebalance_tick will get called every timer tick, on every CPU.
2348 * It checks each scheduling domain to see if it is due to be balanced,
2349 * and initiates a balancing operation if so.
2351 * Balancing parameters are set up in arch_init_sched_domains.
2354 /* Don't have all balancing operations going off at once */
2355 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2357 static void rebalance_tick(int this_cpu
, runqueue_t
*this_rq
,
2358 enum idle_type idle
)
2360 unsigned long old_load
, this_load
;
2361 unsigned long j
= jiffies
+ CPU_OFFSET(this_cpu
);
2362 struct sched_domain
*sd
;
2365 this_load
= this_rq
->nr_running
* SCHED_LOAD_SCALE
;
2366 /* Update our load */
2367 for (i
= 0; i
< 3; i
++) {
2368 unsigned long new_load
= this_load
;
2370 old_load
= this_rq
->cpu_load
[i
];
2372 * Round up the averaging division if load is increasing. This
2373 * prevents us from getting stuck on 9 if the load is 10, for
2376 if (new_load
> old_load
)
2377 new_load
+= scale
-1;
2378 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) / scale
;
2381 for_each_domain(this_cpu
, sd
) {
2382 unsigned long interval
;
2384 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2387 interval
= sd
->balance_interval
;
2388 if (idle
!= SCHED_IDLE
)
2389 interval
*= sd
->busy_factor
;
2391 /* scale ms to jiffies */
2392 interval
= msecs_to_jiffies(interval
);
2393 if (unlikely(!interval
))
2396 if (j
- sd
->last_balance
>= interval
) {
2397 if (load_balance(this_cpu
, this_rq
, sd
, idle
)) {
2399 * We've pulled tasks over so either we're no
2400 * longer idle, or one of our SMT siblings is
2405 sd
->last_balance
+= interval
;
2411 * on UP we do not need to balance between CPUs:
2413 static inline void rebalance_tick(int cpu
, runqueue_t
*rq
, enum idle_type idle
)
2416 static inline void idle_balance(int cpu
, runqueue_t
*rq
)
2421 static inline int wake_priority_sleeper(runqueue_t
*rq
)
2424 #ifdef CONFIG_SCHED_SMT
2425 spin_lock(&rq
->lock
);
2427 * If an SMT sibling task has been put to sleep for priority
2428 * reasons reschedule the idle task to see if it can now run.
2430 if (rq
->nr_running
) {
2431 resched_task(rq
->idle
);
2434 spin_unlock(&rq
->lock
);
2439 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2441 EXPORT_PER_CPU_SYMBOL(kstat
);
2444 * This is called on clock ticks and on context switches.
2445 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2447 static inline void update_cpu_clock(task_t
*p
, runqueue_t
*rq
,
2448 unsigned long long now
)
2450 unsigned long long last
= max(p
->timestamp
, rq
->timestamp_last_tick
);
2451 p
->sched_time
+= now
- last
;
2455 * Return current->sched_time plus any more ns on the sched_clock
2456 * that have not yet been banked.
2458 unsigned long long current_sched_time(const task_t
*tsk
)
2460 unsigned long long ns
;
2461 unsigned long flags
;
2462 local_irq_save(flags
);
2463 ns
= max(tsk
->timestamp
, task_rq(tsk
)->timestamp_last_tick
);
2464 ns
= tsk
->sched_time
+ (sched_clock() - ns
);
2465 local_irq_restore(flags
);
2470 * We place interactive tasks back into the active array, if possible.
2472 * To guarantee that this does not starve expired tasks we ignore the
2473 * interactivity of a task if the first expired task had to wait more
2474 * than a 'reasonable' amount of time. This deadline timeout is
2475 * load-dependent, as the frequency of array switched decreases with
2476 * increasing number of running tasks. We also ignore the interactivity
2477 * if a better static_prio task has expired:
2479 #define EXPIRED_STARVING(rq) \
2480 ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2481 (jiffies - (rq)->expired_timestamp >= \
2482 STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2483 ((rq)->curr->static_prio > (rq)->best_expired_prio))
2486 * Account user cpu time to a process.
2487 * @p: the process that the cpu time gets accounted to
2488 * @hardirq_offset: the offset to subtract from hardirq_count()
2489 * @cputime: the cpu time spent in user space since the last update
2491 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
2493 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2496 p
->utime
= cputime_add(p
->utime
, cputime
);
2498 /* Add user time to cpustat. */
2499 tmp
= cputime_to_cputime64(cputime
);
2500 if (TASK_NICE(p
) > 0)
2501 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
2503 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
2507 * Account system cpu time to a process.
2508 * @p: the process that the cpu time gets accounted to
2509 * @hardirq_offset: the offset to subtract from hardirq_count()
2510 * @cputime: the cpu time spent in kernel space since the last update
2512 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
2515 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2516 runqueue_t
*rq
= this_rq();
2519 p
->stime
= cputime_add(p
->stime
, cputime
);
2521 /* Add system time to cpustat. */
2522 tmp
= cputime_to_cputime64(cputime
);
2523 if (hardirq_count() - hardirq_offset
)
2524 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
2525 else if (softirq_count())
2526 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
2527 else if (p
!= rq
->idle
)
2528 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
2529 else if (atomic_read(&rq
->nr_iowait
) > 0)
2530 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
2532 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
2533 /* Account for system time used */
2534 acct_update_integrals(p
);
2538 * Account for involuntary wait time.
2539 * @p: the process from which the cpu time has been stolen
2540 * @steal: the cpu time spent in involuntary wait
2542 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
2544 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2545 cputime64_t tmp
= cputime_to_cputime64(steal
);
2546 runqueue_t
*rq
= this_rq();
2548 if (p
== rq
->idle
) {
2549 p
->stime
= cputime_add(p
->stime
, steal
);
2550 if (atomic_read(&rq
->nr_iowait
) > 0)
2551 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
2553 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
2555 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
2559 * This function gets called by the timer code, with HZ frequency.
2560 * We call it with interrupts disabled.
2562 * It also gets called by the fork code, when changing the parent's
2565 void scheduler_tick(void)
2567 int cpu
= smp_processor_id();
2568 runqueue_t
*rq
= this_rq();
2569 task_t
*p
= current
;
2570 unsigned long long now
= sched_clock();
2572 update_cpu_clock(p
, rq
, now
);
2574 rq
->timestamp_last_tick
= now
;
2576 if (p
== rq
->idle
) {
2577 if (wake_priority_sleeper(rq
))
2579 rebalance_tick(cpu
, rq
, SCHED_IDLE
);
2583 /* Task might have expired already, but not scheduled off yet */
2584 if (p
->array
!= rq
->active
) {
2585 set_tsk_need_resched(p
);
2588 spin_lock(&rq
->lock
);
2590 * The task was running during this tick - update the
2591 * time slice counter. Note: we do not update a thread's
2592 * priority until it either goes to sleep or uses up its
2593 * timeslice. This makes it possible for interactive tasks
2594 * to use up their timeslices at their highest priority levels.
2598 * RR tasks need a special form of timeslice management.
2599 * FIFO tasks have no timeslices.
2601 if ((p
->policy
== SCHED_RR
) && !--p
->time_slice
) {
2602 p
->time_slice
= task_timeslice(p
);
2603 p
->first_time_slice
= 0;
2604 set_tsk_need_resched(p
);
2606 /* put it at the end of the queue: */
2607 requeue_task(p
, rq
->active
);
2611 if (!--p
->time_slice
) {
2612 dequeue_task(p
, rq
->active
);
2613 set_tsk_need_resched(p
);
2614 p
->prio
= effective_prio(p
);
2615 p
->time_slice
= task_timeslice(p
);
2616 p
->first_time_slice
= 0;
2618 if (!rq
->expired_timestamp
)
2619 rq
->expired_timestamp
= jiffies
;
2620 if (!TASK_INTERACTIVE(p
) || EXPIRED_STARVING(rq
)) {
2621 enqueue_task(p
, rq
->expired
);
2622 if (p
->static_prio
< rq
->best_expired_prio
)
2623 rq
->best_expired_prio
= p
->static_prio
;
2625 enqueue_task(p
, rq
->active
);
2628 * Prevent a too long timeslice allowing a task to monopolize
2629 * the CPU. We do this by splitting up the timeslice into
2632 * Note: this does not mean the task's timeslices expire or
2633 * get lost in any way, they just might be preempted by
2634 * another task of equal priority. (one with higher
2635 * priority would have preempted this task already.) We
2636 * requeue this task to the end of the list on this priority
2637 * level, which is in essence a round-robin of tasks with
2640 * This only applies to tasks in the interactive
2641 * delta range with at least TIMESLICE_GRANULARITY to requeue.
2643 if (TASK_INTERACTIVE(p
) && !((task_timeslice(p
) -
2644 p
->time_slice
) % TIMESLICE_GRANULARITY(p
)) &&
2645 (p
->time_slice
>= TIMESLICE_GRANULARITY(p
)) &&
2646 (p
->array
== rq
->active
)) {
2648 requeue_task(p
, rq
->active
);
2649 set_tsk_need_resched(p
);
2653 spin_unlock(&rq
->lock
);
2655 rebalance_tick(cpu
, rq
, NOT_IDLE
);
2658 #ifdef CONFIG_SCHED_SMT
2659 static inline void wakeup_busy_runqueue(runqueue_t
*rq
)
2661 /* If an SMT runqueue is sleeping due to priority reasons wake it up */
2662 if (rq
->curr
== rq
->idle
&& rq
->nr_running
)
2663 resched_task(rq
->idle
);
2666 static void wake_sleeping_dependent(int this_cpu
, runqueue_t
*this_rq
)
2668 struct sched_domain
*tmp
, *sd
= NULL
;
2669 cpumask_t sibling_map
;
2672 for_each_domain(this_cpu
, tmp
)
2673 if (tmp
->flags
& SD_SHARE_CPUPOWER
)
2680 * Unlock the current runqueue because we have to lock in
2681 * CPU order to avoid deadlocks. Caller knows that we might
2682 * unlock. We keep IRQs disabled.
2684 spin_unlock(&this_rq
->lock
);
2686 sibling_map
= sd
->span
;
2688 for_each_cpu_mask(i
, sibling_map
)
2689 spin_lock(&cpu_rq(i
)->lock
);
2691 * We clear this CPU from the mask. This both simplifies the
2692 * inner loop and keps this_rq locked when we exit:
2694 cpu_clear(this_cpu
, sibling_map
);
2696 for_each_cpu_mask(i
, sibling_map
) {
2697 runqueue_t
*smt_rq
= cpu_rq(i
);
2699 wakeup_busy_runqueue(smt_rq
);
2702 for_each_cpu_mask(i
, sibling_map
)
2703 spin_unlock(&cpu_rq(i
)->lock
);
2705 * We exit with this_cpu's rq still held and IRQs
2711 * number of 'lost' timeslices this task wont be able to fully
2712 * utilize, if another task runs on a sibling. This models the
2713 * slowdown effect of other tasks running on siblings:
2715 static inline unsigned long smt_slice(task_t
*p
, struct sched_domain
*sd
)
2717 return p
->time_slice
* (100 - sd
->per_cpu_gain
) / 100;
2720 static int dependent_sleeper(int this_cpu
, runqueue_t
*this_rq
)
2722 struct sched_domain
*tmp
, *sd
= NULL
;
2723 cpumask_t sibling_map
;
2724 prio_array_t
*array
;
2728 for_each_domain(this_cpu
, tmp
)
2729 if (tmp
->flags
& SD_SHARE_CPUPOWER
)
2736 * The same locking rules and details apply as for
2737 * wake_sleeping_dependent():
2739 spin_unlock(&this_rq
->lock
);
2740 sibling_map
= sd
->span
;
2741 for_each_cpu_mask(i
, sibling_map
)
2742 spin_lock(&cpu_rq(i
)->lock
);
2743 cpu_clear(this_cpu
, sibling_map
);
2746 * Establish next task to be run - it might have gone away because
2747 * we released the runqueue lock above:
2749 if (!this_rq
->nr_running
)
2751 array
= this_rq
->active
;
2752 if (!array
->nr_active
)
2753 array
= this_rq
->expired
;
2754 BUG_ON(!array
->nr_active
);
2756 p
= list_entry(array
->queue
[sched_find_first_bit(array
->bitmap
)].next
,
2759 for_each_cpu_mask(i
, sibling_map
) {
2760 runqueue_t
*smt_rq
= cpu_rq(i
);
2761 task_t
*smt_curr
= smt_rq
->curr
;
2763 /* Kernel threads do not participate in dependent sleeping */
2764 if (!p
->mm
|| !smt_curr
->mm
|| rt_task(p
))
2765 goto check_smt_task
;
2768 * If a user task with lower static priority than the
2769 * running task on the SMT sibling is trying to schedule,
2770 * delay it till there is proportionately less timeslice
2771 * left of the sibling task to prevent a lower priority
2772 * task from using an unfair proportion of the
2773 * physical cpu's resources. -ck
2775 if (rt_task(smt_curr
)) {
2777 * With real time tasks we run non-rt tasks only
2778 * per_cpu_gain% of the time.
2780 if ((jiffies
% DEF_TIMESLICE
) >
2781 (sd
->per_cpu_gain
* DEF_TIMESLICE
/ 100))
2784 if (smt_curr
->static_prio
< p
->static_prio
&&
2785 !TASK_PREEMPTS_CURR(p
, smt_rq
) &&
2786 smt_slice(smt_curr
, sd
) > task_timeslice(p
))
2790 if ((!smt_curr
->mm
&& smt_curr
!= smt_rq
->idle
) ||
2794 wakeup_busy_runqueue(smt_rq
);
2799 * Reschedule a lower priority task on the SMT sibling for
2800 * it to be put to sleep, or wake it up if it has been put to
2801 * sleep for priority reasons to see if it should run now.
2804 if ((jiffies
% DEF_TIMESLICE
) >
2805 (sd
->per_cpu_gain
* DEF_TIMESLICE
/ 100))
2806 resched_task(smt_curr
);
2808 if (TASK_PREEMPTS_CURR(p
, smt_rq
) &&
2809 smt_slice(p
, sd
) > task_timeslice(smt_curr
))
2810 resched_task(smt_curr
);
2812 wakeup_busy_runqueue(smt_rq
);
2816 for_each_cpu_mask(i
, sibling_map
)
2817 spin_unlock(&cpu_rq(i
)->lock
);
2821 static inline void wake_sleeping_dependent(int this_cpu
, runqueue_t
*this_rq
)
2825 static inline int dependent_sleeper(int this_cpu
, runqueue_t
*this_rq
)
2831 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
2833 void fastcall
add_preempt_count(int val
)
2838 BUG_ON((preempt_count() < 0));
2839 preempt_count() += val
;
2841 * Spinlock count overflowing soon?
2843 BUG_ON((preempt_count() & PREEMPT_MASK
) >= PREEMPT_MASK
-10);
2845 EXPORT_SYMBOL(add_preempt_count
);
2847 void fastcall
sub_preempt_count(int val
)
2852 BUG_ON(val
> preempt_count());
2854 * Is the spinlock portion underflowing?
2856 BUG_ON((val
< PREEMPT_MASK
) && !(preempt_count() & PREEMPT_MASK
));
2857 preempt_count() -= val
;
2859 EXPORT_SYMBOL(sub_preempt_count
);
2864 * schedule() is the main scheduler function.
2866 asmlinkage
void __sched
schedule(void)
2869 task_t
*prev
, *next
;
2871 prio_array_t
*array
;
2872 struct list_head
*queue
;
2873 unsigned long long now
;
2874 unsigned long run_time
;
2875 int cpu
, idx
, new_prio
;
2878 * Test if we are atomic. Since do_exit() needs to call into
2879 * schedule() atomically, we ignore that path for now.
2880 * Otherwise, whine if we are scheduling when we should not be.
2882 if (unlikely(in_atomic() && !current
->exit_state
)) {
2883 printk(KERN_ERR
"BUG: scheduling while atomic: "
2885 current
->comm
, preempt_count(), current
->pid
);
2888 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2893 release_kernel_lock(prev
);
2894 need_resched_nonpreemptible
:
2898 * The idle thread is not allowed to schedule!
2899 * Remove this check after it has been exercised a bit.
2901 if (unlikely(prev
== rq
->idle
) && prev
->state
!= TASK_RUNNING
) {
2902 printk(KERN_ERR
"bad: scheduling from the idle thread!\n");
2906 schedstat_inc(rq
, sched_cnt
);
2907 now
= sched_clock();
2908 if (likely((long long)(now
- prev
->timestamp
) < NS_MAX_SLEEP_AVG
)) {
2909 run_time
= now
- prev
->timestamp
;
2910 if (unlikely((long long)(now
- prev
->timestamp
) < 0))
2913 run_time
= NS_MAX_SLEEP_AVG
;
2916 * Tasks charged proportionately less run_time at high sleep_avg to
2917 * delay them losing their interactive status
2919 run_time
/= (CURRENT_BONUS(prev
) ? : 1);
2921 spin_lock_irq(&rq
->lock
);
2923 if (unlikely(prev
->flags
& PF_DEAD
))
2924 prev
->state
= EXIT_DEAD
;
2926 switch_count
= &prev
->nivcsw
;
2927 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
2928 switch_count
= &prev
->nvcsw
;
2929 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
2930 unlikely(signal_pending(prev
))))
2931 prev
->state
= TASK_RUNNING
;
2933 if (prev
->state
== TASK_UNINTERRUPTIBLE
)
2934 rq
->nr_uninterruptible
++;
2935 deactivate_task(prev
, rq
);
2939 cpu
= smp_processor_id();
2940 if (unlikely(!rq
->nr_running
)) {
2942 idle_balance(cpu
, rq
);
2943 if (!rq
->nr_running
) {
2945 rq
->expired_timestamp
= 0;
2946 wake_sleeping_dependent(cpu
, rq
);
2948 * wake_sleeping_dependent() might have released
2949 * the runqueue, so break out if we got new
2952 if (!rq
->nr_running
)
2956 if (dependent_sleeper(cpu
, rq
)) {
2961 * dependent_sleeper() releases and reacquires the runqueue
2962 * lock, hence go into the idle loop if the rq went
2965 if (unlikely(!rq
->nr_running
))
2970 if (unlikely(!array
->nr_active
)) {
2972 * Switch the active and expired arrays.
2974 schedstat_inc(rq
, sched_switch
);
2975 rq
->active
= rq
->expired
;
2976 rq
->expired
= array
;
2978 rq
->expired_timestamp
= 0;
2979 rq
->best_expired_prio
= MAX_PRIO
;
2982 idx
= sched_find_first_bit(array
->bitmap
);
2983 queue
= array
->queue
+ idx
;
2984 next
= list_entry(queue
->next
, task_t
, run_list
);
2986 if (!rt_task(next
) && next
->activated
> 0) {
2987 unsigned long long delta
= now
- next
->timestamp
;
2988 if (unlikely((long long)(now
- next
->timestamp
) < 0))
2991 if (next
->activated
== 1)
2992 delta
= delta
* (ON_RUNQUEUE_WEIGHT
* 128 / 100) / 128;
2994 array
= next
->array
;
2995 new_prio
= recalc_task_prio(next
, next
->timestamp
+ delta
);
2997 if (unlikely(next
->prio
!= new_prio
)) {
2998 dequeue_task(next
, array
);
2999 next
->prio
= new_prio
;
3000 enqueue_task(next
, array
);
3002 requeue_task(next
, array
);
3004 next
->activated
= 0;
3006 if (next
== rq
->idle
)
3007 schedstat_inc(rq
, sched_goidle
);
3009 prefetch_stack(next
);
3010 clear_tsk_need_resched(prev
);
3011 rcu_qsctr_inc(task_cpu(prev
));
3013 update_cpu_clock(prev
, rq
, now
);
3015 prev
->sleep_avg
-= run_time
;
3016 if ((long)prev
->sleep_avg
<= 0)
3017 prev
->sleep_avg
= 0;
3018 prev
->timestamp
= prev
->last_ran
= now
;
3020 sched_info_switch(prev
, next
);
3021 if (likely(prev
!= next
)) {
3022 next
->timestamp
= now
;
3027 prepare_task_switch(rq
, next
);
3028 prev
= context_switch(rq
, prev
, next
);
3031 * this_rq must be evaluated again because prev may have moved
3032 * CPUs since it called schedule(), thus the 'rq' on its stack
3033 * frame will be invalid.
3035 finish_task_switch(this_rq(), prev
);
3037 spin_unlock_irq(&rq
->lock
);
3040 if (unlikely(reacquire_kernel_lock(prev
) < 0))
3041 goto need_resched_nonpreemptible
;
3042 preempt_enable_no_resched();
3043 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3047 EXPORT_SYMBOL(schedule
);
3049 #ifdef CONFIG_PREEMPT
3051 * this is is the entry point to schedule() from in-kernel preemption
3052 * off of preempt_enable. Kernel preemptions off return from interrupt
3053 * occur there and call schedule directly.
3055 asmlinkage
void __sched
preempt_schedule(void)
3057 struct thread_info
*ti
= current_thread_info();
3058 #ifdef CONFIG_PREEMPT_BKL
3059 struct task_struct
*task
= current
;
3060 int saved_lock_depth
;
3063 * If there is a non-zero preempt_count or interrupts are disabled,
3064 * we do not want to preempt the current task. Just return..
3066 if (unlikely(ti
->preempt_count
|| irqs_disabled()))
3070 add_preempt_count(PREEMPT_ACTIVE
);
3072 * We keep the big kernel semaphore locked, but we
3073 * clear ->lock_depth so that schedule() doesnt
3074 * auto-release the semaphore:
3076 #ifdef CONFIG_PREEMPT_BKL
3077 saved_lock_depth
= task
->lock_depth
;
3078 task
->lock_depth
= -1;
3081 #ifdef CONFIG_PREEMPT_BKL
3082 task
->lock_depth
= saved_lock_depth
;
3084 sub_preempt_count(PREEMPT_ACTIVE
);
3086 /* we could miss a preemption opportunity between schedule and now */
3088 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3092 EXPORT_SYMBOL(preempt_schedule
);
3095 * this is is the entry point to schedule() from kernel preemption
3096 * off of irq context.
3097 * Note, that this is called and return with irqs disabled. This will
3098 * protect us against recursive calling from irq.
3100 asmlinkage
void __sched
preempt_schedule_irq(void)
3102 struct thread_info
*ti
= current_thread_info();
3103 #ifdef CONFIG_PREEMPT_BKL
3104 struct task_struct
*task
= current
;
3105 int saved_lock_depth
;
3107 /* Catch callers which need to be fixed*/
3108 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3111 add_preempt_count(PREEMPT_ACTIVE
);
3113 * We keep the big kernel semaphore locked, but we
3114 * clear ->lock_depth so that schedule() doesnt
3115 * auto-release the semaphore:
3117 #ifdef CONFIG_PREEMPT_BKL
3118 saved_lock_depth
= task
->lock_depth
;
3119 task
->lock_depth
= -1;
3123 local_irq_disable();
3124 #ifdef CONFIG_PREEMPT_BKL
3125 task
->lock_depth
= saved_lock_depth
;
3127 sub_preempt_count(PREEMPT_ACTIVE
);
3129 /* we could miss a preemption opportunity between schedule and now */
3131 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3135 #endif /* CONFIG_PREEMPT */
3137 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3140 task_t
*p
= curr
->private;
3141 return try_to_wake_up(p
, mode
, sync
);
3144 EXPORT_SYMBOL(default_wake_function
);
3147 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3148 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3149 * number) then we wake all the non-exclusive tasks and one exclusive task.
3151 * There are circumstances in which we can try to wake a task which has already
3152 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3153 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3155 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3156 int nr_exclusive
, int sync
, void *key
)
3158 struct list_head
*tmp
, *next
;
3160 list_for_each_safe(tmp
, next
, &q
->task_list
) {
3163 curr
= list_entry(tmp
, wait_queue_t
, task_list
);
3164 flags
= curr
->flags
;
3165 if (curr
->func(curr
, mode
, sync
, key
) &&
3166 (flags
& WQ_FLAG_EXCLUSIVE
) &&
3173 * __wake_up - wake up threads blocked on a waitqueue.
3175 * @mode: which threads
3176 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3177 * @key: is directly passed to the wakeup function
3179 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3180 int nr_exclusive
, void *key
)
3182 unsigned long flags
;
3184 spin_lock_irqsave(&q
->lock
, flags
);
3185 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3186 spin_unlock_irqrestore(&q
->lock
, flags
);
3189 EXPORT_SYMBOL(__wake_up
);
3192 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3194 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3196 __wake_up_common(q
, mode
, 1, 0, NULL
);
3200 * __wake_up_sync - wake up threads blocked on a waitqueue.
3202 * @mode: which threads
3203 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3205 * The sync wakeup differs that the waker knows that it will schedule
3206 * away soon, so while the target thread will be woken up, it will not
3207 * be migrated to another CPU - ie. the two threads are 'synchronized'
3208 * with each other. This can prevent needless bouncing between CPUs.
3210 * On UP it can prevent extra preemption.
3213 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3215 unsigned long flags
;
3221 if (unlikely(!nr_exclusive
))
3224 spin_lock_irqsave(&q
->lock
, flags
);
3225 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3226 spin_unlock_irqrestore(&q
->lock
, flags
);
3228 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3230 void fastcall
complete(struct completion
*x
)
3232 unsigned long flags
;
3234 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3236 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3238 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3240 EXPORT_SYMBOL(complete
);
3242 void fastcall
complete_all(struct completion
*x
)
3244 unsigned long flags
;
3246 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3247 x
->done
+= UINT_MAX
/2;
3248 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3250 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3252 EXPORT_SYMBOL(complete_all
);
3254 void fastcall __sched
wait_for_completion(struct completion
*x
)
3257 spin_lock_irq(&x
->wait
.lock
);
3259 DECLARE_WAITQUEUE(wait
, current
);
3261 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3262 __add_wait_queue_tail(&x
->wait
, &wait
);
3264 __set_current_state(TASK_UNINTERRUPTIBLE
);
3265 spin_unlock_irq(&x
->wait
.lock
);
3267 spin_lock_irq(&x
->wait
.lock
);
3269 __remove_wait_queue(&x
->wait
, &wait
);
3272 spin_unlock_irq(&x
->wait
.lock
);
3274 EXPORT_SYMBOL(wait_for_completion
);
3276 unsigned long fastcall __sched
3277 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3281 spin_lock_irq(&x
->wait
.lock
);
3283 DECLARE_WAITQUEUE(wait
, current
);
3285 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3286 __add_wait_queue_tail(&x
->wait
, &wait
);
3288 __set_current_state(TASK_UNINTERRUPTIBLE
);
3289 spin_unlock_irq(&x
->wait
.lock
);
3290 timeout
= schedule_timeout(timeout
);
3291 spin_lock_irq(&x
->wait
.lock
);
3293 __remove_wait_queue(&x
->wait
, &wait
);
3297 __remove_wait_queue(&x
->wait
, &wait
);
3301 spin_unlock_irq(&x
->wait
.lock
);
3304 EXPORT_SYMBOL(wait_for_completion_timeout
);
3306 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3312 spin_lock_irq(&x
->wait
.lock
);
3314 DECLARE_WAITQUEUE(wait
, current
);
3316 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3317 __add_wait_queue_tail(&x
->wait
, &wait
);
3319 if (signal_pending(current
)) {
3321 __remove_wait_queue(&x
->wait
, &wait
);
3324 __set_current_state(TASK_INTERRUPTIBLE
);
3325 spin_unlock_irq(&x
->wait
.lock
);
3327 spin_lock_irq(&x
->wait
.lock
);
3329 __remove_wait_queue(&x
->wait
, &wait
);
3333 spin_unlock_irq(&x
->wait
.lock
);
3337 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3339 unsigned long fastcall __sched
3340 wait_for_completion_interruptible_timeout(struct completion
*x
,
3341 unsigned long timeout
)
3345 spin_lock_irq(&x
->wait
.lock
);
3347 DECLARE_WAITQUEUE(wait
, current
);
3349 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3350 __add_wait_queue_tail(&x
->wait
, &wait
);
3352 if (signal_pending(current
)) {
3353 timeout
= -ERESTARTSYS
;
3354 __remove_wait_queue(&x
->wait
, &wait
);
3357 __set_current_state(TASK_INTERRUPTIBLE
);
3358 spin_unlock_irq(&x
->wait
.lock
);
3359 timeout
= schedule_timeout(timeout
);
3360 spin_lock_irq(&x
->wait
.lock
);
3362 __remove_wait_queue(&x
->wait
, &wait
);
3366 __remove_wait_queue(&x
->wait
, &wait
);
3370 spin_unlock_irq(&x
->wait
.lock
);
3373 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3376 #define SLEEP_ON_VAR \
3377 unsigned long flags; \
3378 wait_queue_t wait; \
3379 init_waitqueue_entry(&wait, current);
3381 #define SLEEP_ON_HEAD \
3382 spin_lock_irqsave(&q->lock,flags); \
3383 __add_wait_queue(q, &wait); \
3384 spin_unlock(&q->lock);
3386 #define SLEEP_ON_TAIL \
3387 spin_lock_irq(&q->lock); \
3388 __remove_wait_queue(q, &wait); \
3389 spin_unlock_irqrestore(&q->lock, flags);
3391 void fastcall __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3395 current
->state
= TASK_INTERRUPTIBLE
;
3402 EXPORT_SYMBOL(interruptible_sleep_on
);
3404 long fastcall __sched
3405 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3409 current
->state
= TASK_INTERRUPTIBLE
;
3412 timeout
= schedule_timeout(timeout
);
3418 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3420 void fastcall __sched
sleep_on(wait_queue_head_t
*q
)
3424 current
->state
= TASK_UNINTERRUPTIBLE
;
3431 EXPORT_SYMBOL(sleep_on
);
3433 long fastcall __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3437 current
->state
= TASK_UNINTERRUPTIBLE
;
3440 timeout
= schedule_timeout(timeout
);
3446 EXPORT_SYMBOL(sleep_on_timeout
);
3448 void set_user_nice(task_t
*p
, long nice
)
3450 unsigned long flags
;
3451 prio_array_t
*array
;
3453 int old_prio
, new_prio
, delta
;
3455 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3458 * We have to be careful, if called from sys_setpriority(),
3459 * the task might be in the middle of scheduling on another CPU.
3461 rq
= task_rq_lock(p
, &flags
);
3463 * The RT priorities are set via sched_setscheduler(), but we still
3464 * allow the 'normal' nice value to be set - but as expected
3465 * it wont have any effect on scheduling until the task is
3466 * not SCHED_NORMAL/SCHED_BATCH:
3469 p
->static_prio
= NICE_TO_PRIO(nice
);
3474 dequeue_task(p
, array
);
3477 new_prio
= NICE_TO_PRIO(nice
);
3478 delta
= new_prio
- old_prio
;
3479 p
->static_prio
= NICE_TO_PRIO(nice
);
3483 enqueue_task(p
, array
);
3485 * If the task increased its priority or is running and
3486 * lowered its priority, then reschedule its CPU:
3488 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3489 resched_task(rq
->curr
);
3492 task_rq_unlock(rq
, &flags
);
3495 EXPORT_SYMBOL(set_user_nice
);
3498 * can_nice - check if a task can reduce its nice value
3502 int can_nice(const task_t
*p
, const int nice
)
3504 /* convert nice value [19,-20] to rlimit style value [1,40] */
3505 int nice_rlim
= 20 - nice
;
3506 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
3507 capable(CAP_SYS_NICE
));
3510 #ifdef __ARCH_WANT_SYS_NICE
3513 * sys_nice - change the priority of the current process.
3514 * @increment: priority increment
3516 * sys_setpriority is a more generic, but much slower function that
3517 * does similar things.
3519 asmlinkage
long sys_nice(int increment
)
3525 * Setpriority might change our priority at the same moment.
3526 * We don't have to worry. Conceptually one call occurs first
3527 * and we have a single winner.
3529 if (increment
< -40)
3534 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
3540 if (increment
< 0 && !can_nice(current
, nice
))
3543 retval
= security_task_setnice(current
, nice
);
3547 set_user_nice(current
, nice
);
3554 * task_prio - return the priority value of a given task.
3555 * @p: the task in question.
3557 * This is the priority value as seen by users in /proc.
3558 * RT tasks are offset by -200. Normal tasks are centered
3559 * around 0, value goes from -16 to +15.
3561 int task_prio(const task_t
*p
)
3563 return p
->prio
- MAX_RT_PRIO
;
3567 * task_nice - return the nice value of a given task.
3568 * @p: the task in question.
3570 int task_nice(const task_t
*p
)
3572 return TASK_NICE(p
);
3574 EXPORT_SYMBOL_GPL(task_nice
);
3577 * idle_cpu - is a given cpu idle currently?
3578 * @cpu: the processor in question.
3580 int idle_cpu(int cpu
)
3582 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
3586 * idle_task - return the idle task for a given cpu.
3587 * @cpu: the processor in question.
3589 task_t
*idle_task(int cpu
)
3591 return cpu_rq(cpu
)->idle
;
3595 * find_process_by_pid - find a process with a matching PID value.
3596 * @pid: the pid in question.
3598 static inline task_t
*find_process_by_pid(pid_t pid
)
3600 return pid
? find_task_by_pid(pid
) : current
;
3603 /* Actually do priority change: must hold rq lock. */
3604 static void __setscheduler(struct task_struct
*p
, int policy
, int prio
)
3608 p
->rt_priority
= prio
;
3609 if (policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
) {
3610 p
->prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
3612 p
->prio
= p
->static_prio
;
3614 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
3616 if (policy
== SCHED_BATCH
)
3622 * sched_setscheduler - change the scheduling policy and/or RT priority of
3624 * @p: the task in question.
3625 * @policy: new policy.
3626 * @param: structure containing the new RT priority.
3628 int sched_setscheduler(struct task_struct
*p
, int policy
,
3629 struct sched_param
*param
)
3632 int oldprio
, oldpolicy
= -1;
3633 prio_array_t
*array
;
3634 unsigned long flags
;
3638 /* double check policy once rq lock held */
3640 policy
= oldpolicy
= p
->policy
;
3641 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3642 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
)
3645 * Valid priorities for SCHED_FIFO and SCHED_RR are
3646 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
3649 if (param
->sched_priority
< 0 ||
3650 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3651 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
3653 if ((policy
== SCHED_NORMAL
|| policy
== SCHED_BATCH
)
3654 != (param
->sched_priority
== 0))
3658 * Allow unprivileged RT tasks to decrease priority:
3660 if (!capable(CAP_SYS_NICE
)) {
3662 * can't change policy, except between SCHED_NORMAL
3665 if (((policy
!= SCHED_NORMAL
&& p
->policy
!= SCHED_BATCH
) &&
3666 (policy
!= SCHED_BATCH
&& p
->policy
!= SCHED_NORMAL
)) &&
3667 !p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
)
3669 /* can't increase priority */
3670 if ((policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
) &&
3671 param
->sched_priority
> p
->rt_priority
&&
3672 param
->sched_priority
>
3673 p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
)
3675 /* can't change other user's priorities */
3676 if ((current
->euid
!= p
->euid
) &&
3677 (current
->euid
!= p
->uid
))
3681 retval
= security_task_setscheduler(p
, policy
, param
);
3685 * To be able to change p->policy safely, the apropriate
3686 * runqueue lock must be held.
3688 rq
= task_rq_lock(p
, &flags
);
3689 /* recheck policy now with rq lock held */
3690 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3691 policy
= oldpolicy
= -1;
3692 task_rq_unlock(rq
, &flags
);
3697 deactivate_task(p
, rq
);
3699 __setscheduler(p
, policy
, param
->sched_priority
);
3701 __activate_task(p
, rq
);
3703 * Reschedule if we are currently running on this runqueue and
3704 * our priority decreased, or if we are not currently running on
3705 * this runqueue and our priority is higher than the current's
3707 if (task_running(rq
, p
)) {
3708 if (p
->prio
> oldprio
)
3709 resched_task(rq
->curr
);
3710 } else if (TASK_PREEMPTS_CURR(p
, rq
))
3711 resched_task(rq
->curr
);
3713 task_rq_unlock(rq
, &flags
);
3716 EXPORT_SYMBOL_GPL(sched_setscheduler
);
3719 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
3722 struct sched_param lparam
;
3723 struct task_struct
*p
;
3725 if (!param
|| pid
< 0)
3727 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
3729 read_lock_irq(&tasklist_lock
);
3730 p
= find_process_by_pid(pid
);
3732 read_unlock_irq(&tasklist_lock
);
3735 retval
= sched_setscheduler(p
, policy
, &lparam
);
3736 read_unlock_irq(&tasklist_lock
);
3741 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3742 * @pid: the pid in question.
3743 * @policy: new policy.
3744 * @param: structure containing the new RT priority.
3746 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
3747 struct sched_param __user
*param
)
3749 /* negative values for policy are not valid */
3753 return do_sched_setscheduler(pid
, policy
, param
);
3757 * sys_sched_setparam - set/change the RT priority of a thread
3758 * @pid: the pid in question.
3759 * @param: structure containing the new RT priority.
3761 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
3763 return do_sched_setscheduler(pid
, -1, param
);
3767 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3768 * @pid: the pid in question.
3770 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
3772 int retval
= -EINVAL
;
3779 read_lock(&tasklist_lock
);
3780 p
= find_process_by_pid(pid
);
3782 retval
= security_task_getscheduler(p
);
3786 read_unlock(&tasklist_lock
);
3793 * sys_sched_getscheduler - get the RT priority of a thread
3794 * @pid: the pid in question.
3795 * @param: structure containing the RT priority.
3797 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
3799 struct sched_param lp
;
3800 int retval
= -EINVAL
;
3803 if (!param
|| pid
< 0)
3806 read_lock(&tasklist_lock
);
3807 p
= find_process_by_pid(pid
);
3812 retval
= security_task_getscheduler(p
);
3816 lp
.sched_priority
= p
->rt_priority
;
3817 read_unlock(&tasklist_lock
);
3820 * This one might sleep, we cannot do it with a spinlock held ...
3822 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
3828 read_unlock(&tasklist_lock
);
3832 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
3836 cpumask_t cpus_allowed
;
3839 read_lock(&tasklist_lock
);
3841 p
= find_process_by_pid(pid
);
3843 read_unlock(&tasklist_lock
);
3844 unlock_cpu_hotplug();
3849 * It is not safe to call set_cpus_allowed with the
3850 * tasklist_lock held. We will bump the task_struct's
3851 * usage count and then drop tasklist_lock.
3854 read_unlock(&tasklist_lock
);
3857 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
3858 !capable(CAP_SYS_NICE
))
3861 cpus_allowed
= cpuset_cpus_allowed(p
);
3862 cpus_and(new_mask
, new_mask
, cpus_allowed
);
3863 retval
= set_cpus_allowed(p
, new_mask
);
3867 unlock_cpu_hotplug();
3871 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
3872 cpumask_t
*new_mask
)
3874 if (len
< sizeof(cpumask_t
)) {
3875 memset(new_mask
, 0, sizeof(cpumask_t
));
3876 } else if (len
> sizeof(cpumask_t
)) {
3877 len
= sizeof(cpumask_t
);
3879 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
3883 * sys_sched_setaffinity - set the cpu affinity of a process
3884 * @pid: pid of the process
3885 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3886 * @user_mask_ptr: user-space pointer to the new cpu mask
3888 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
3889 unsigned long __user
*user_mask_ptr
)
3894 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
3898 return sched_setaffinity(pid
, new_mask
);
3902 * Represents all cpu's present in the system
3903 * In systems capable of hotplug, this map could dynamically grow
3904 * as new cpu's are detected in the system via any platform specific
3905 * method, such as ACPI for e.g.
3908 cpumask_t cpu_present_map __read_mostly
;
3909 EXPORT_SYMBOL(cpu_present_map
);
3912 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
3913 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
3916 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
3922 read_lock(&tasklist_lock
);
3925 p
= find_process_by_pid(pid
);
3930 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
3933 read_unlock(&tasklist_lock
);
3934 unlock_cpu_hotplug();
3942 * sys_sched_getaffinity - get the cpu affinity of a process
3943 * @pid: pid of the process
3944 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3945 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3947 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
3948 unsigned long __user
*user_mask_ptr
)
3953 if (len
< sizeof(cpumask_t
))
3956 ret
= sched_getaffinity(pid
, &mask
);
3960 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
3963 return sizeof(cpumask_t
);
3967 * sys_sched_yield - yield the current processor to other threads.
3969 * this function yields the current CPU by moving the calling thread
3970 * to the expired array. If there are no other threads running on this
3971 * CPU then this function will return.
3973 asmlinkage
long sys_sched_yield(void)
3975 runqueue_t
*rq
= this_rq_lock();
3976 prio_array_t
*array
= current
->array
;
3977 prio_array_t
*target
= rq
->expired
;
3979 schedstat_inc(rq
, yld_cnt
);
3981 * We implement yielding by moving the task into the expired
3984 * (special rule: RT tasks will just roundrobin in the active
3987 if (rt_task(current
))
3988 target
= rq
->active
;
3990 if (array
->nr_active
== 1) {
3991 schedstat_inc(rq
, yld_act_empty
);
3992 if (!rq
->expired
->nr_active
)
3993 schedstat_inc(rq
, yld_both_empty
);
3994 } else if (!rq
->expired
->nr_active
)
3995 schedstat_inc(rq
, yld_exp_empty
);
3997 if (array
!= target
) {
3998 dequeue_task(current
, array
);
3999 enqueue_task(current
, target
);
4002 * requeue_task is cheaper so perform that if possible.
4004 requeue_task(current
, array
);
4007 * Since we are going to call schedule() anyway, there's
4008 * no need to preempt or enable interrupts:
4010 __release(rq
->lock
);
4011 _raw_spin_unlock(&rq
->lock
);
4012 preempt_enable_no_resched();
4019 static inline void __cond_resched(void)
4022 * The BKS might be reacquired before we have dropped
4023 * PREEMPT_ACTIVE, which could trigger a second
4024 * cond_resched() call.
4026 if (unlikely(preempt_count()))
4028 if (unlikely(system_state
!= SYSTEM_RUNNING
))
4031 add_preempt_count(PREEMPT_ACTIVE
);
4033 sub_preempt_count(PREEMPT_ACTIVE
);
4034 } while (need_resched());
4037 int __sched
cond_resched(void)
4039 if (need_resched()) {
4046 EXPORT_SYMBOL(cond_resched
);
4049 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4050 * call schedule, and on return reacquire the lock.
4052 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4053 * operations here to prevent schedule() from being called twice (once via
4054 * spin_unlock(), once by hand).
4056 int cond_resched_lock(spinlock_t
*lock
)
4060 if (need_lockbreak(lock
)) {
4066 if (need_resched()) {
4067 _raw_spin_unlock(lock
);
4068 preempt_enable_no_resched();
4076 EXPORT_SYMBOL(cond_resched_lock
);
4078 int __sched
cond_resched_softirq(void)
4080 BUG_ON(!in_softirq());
4082 if (need_resched()) {
4083 __local_bh_enable();
4091 EXPORT_SYMBOL(cond_resched_softirq
);
4095 * yield - yield the current processor to other threads.
4097 * this is a shortcut for kernel-space yielding - it marks the
4098 * thread runnable and calls sys_sched_yield().
4100 void __sched
yield(void)
4102 set_current_state(TASK_RUNNING
);
4106 EXPORT_SYMBOL(yield
);
4109 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4110 * that process accounting knows that this is a task in IO wait state.
4112 * But don't do that if it is a deliberate, throttling IO wait (this task
4113 * has set its backing_dev_info: the queue against which it should throttle)
4115 void __sched
io_schedule(void)
4117 struct runqueue
*rq
= &per_cpu(runqueues
, raw_smp_processor_id());
4119 atomic_inc(&rq
->nr_iowait
);
4121 atomic_dec(&rq
->nr_iowait
);
4124 EXPORT_SYMBOL(io_schedule
);
4126 long __sched
io_schedule_timeout(long timeout
)
4128 struct runqueue
*rq
= &per_cpu(runqueues
, raw_smp_processor_id());
4131 atomic_inc(&rq
->nr_iowait
);
4132 ret
= schedule_timeout(timeout
);
4133 atomic_dec(&rq
->nr_iowait
);
4138 * sys_sched_get_priority_max - return maximum RT priority.
4139 * @policy: scheduling class.
4141 * this syscall returns the maximum rt_priority that can be used
4142 * by a given scheduling class.
4144 asmlinkage
long sys_sched_get_priority_max(int policy
)
4151 ret
= MAX_USER_RT_PRIO
-1;
4162 * sys_sched_get_priority_min - return minimum RT priority.
4163 * @policy: scheduling class.
4165 * this syscall returns the minimum rt_priority that can be used
4166 * by a given scheduling class.
4168 asmlinkage
long sys_sched_get_priority_min(int policy
)
4185 * sys_sched_rr_get_interval - return the default timeslice of a process.
4186 * @pid: pid of the process.
4187 * @interval: userspace pointer to the timeslice value.
4189 * this syscall writes the default timeslice value of a given process
4190 * into the user-space timespec buffer. A value of '0' means infinity.
4193 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4195 int retval
= -EINVAL
;
4203 read_lock(&tasklist_lock
);
4204 p
= find_process_by_pid(pid
);
4208 retval
= security_task_getscheduler(p
);
4212 jiffies_to_timespec(p
->policy
& SCHED_FIFO
?
4213 0 : task_timeslice(p
), &t
);
4214 read_unlock(&tasklist_lock
);
4215 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4219 read_unlock(&tasklist_lock
);
4223 static inline struct task_struct
*eldest_child(struct task_struct
*p
)
4225 if (list_empty(&p
->children
)) return NULL
;
4226 return list_entry(p
->children
.next
,struct task_struct
,sibling
);
4229 static inline struct task_struct
*older_sibling(struct task_struct
*p
)
4231 if (p
->sibling
.prev
==&p
->parent
->children
) return NULL
;
4232 return list_entry(p
->sibling
.prev
,struct task_struct
,sibling
);
4235 static inline struct task_struct
*younger_sibling(struct task_struct
*p
)
4237 if (p
->sibling
.next
==&p
->parent
->children
) return NULL
;
4238 return list_entry(p
->sibling
.next
,struct task_struct
,sibling
);
4241 static void show_task(task_t
*p
)
4245 unsigned long free
= 0;
4246 static const char *stat_nam
[] = { "R", "S", "D", "T", "t", "Z", "X" };
4248 printk("%-13.13s ", p
->comm
);
4249 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4250 if (state
< ARRAY_SIZE(stat_nam
))
4251 printk(stat_nam
[state
]);
4254 #if (BITS_PER_LONG == 32)
4255 if (state
== TASK_RUNNING
)
4256 printk(" running ");
4258 printk(" %08lX ", thread_saved_pc(p
));
4260 if (state
== TASK_RUNNING
)
4261 printk(" running task ");
4263 printk(" %016lx ", thread_saved_pc(p
));
4265 #ifdef CONFIG_DEBUG_STACK_USAGE
4267 unsigned long *n
= end_of_stack(p
);
4270 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4273 printk("%5lu %5d %6d ", free
, p
->pid
, p
->parent
->pid
);
4274 if ((relative
= eldest_child(p
)))
4275 printk("%5d ", relative
->pid
);
4278 if ((relative
= younger_sibling(p
)))
4279 printk("%7d", relative
->pid
);
4282 if ((relative
= older_sibling(p
)))
4283 printk(" %5d", relative
->pid
);
4287 printk(" (L-TLB)\n");
4289 printk(" (NOTLB)\n");
4291 if (state
!= TASK_RUNNING
)
4292 show_stack(p
, NULL
);
4295 void show_state(void)
4299 #if (BITS_PER_LONG == 32)
4302 printk(" task PC pid father child younger older\n");
4306 printk(" task PC pid father child younger older\n");
4308 read_lock(&tasklist_lock
);
4309 do_each_thread(g
, p
) {
4311 * reset the NMI-timeout, listing all files on a slow
4312 * console might take alot of time:
4314 touch_nmi_watchdog();
4316 } while_each_thread(g
, p
);
4318 read_unlock(&tasklist_lock
);
4319 mutex_debug_show_all_locks();
4323 * init_idle - set up an idle thread for a given CPU
4324 * @idle: task in question
4325 * @cpu: cpu the idle task belongs to
4327 * NOTE: this function does not set the idle thread's NEED_RESCHED
4328 * flag, to make booting more robust.
4330 void __devinit
init_idle(task_t
*idle
, int cpu
)
4332 runqueue_t
*rq
= cpu_rq(cpu
);
4333 unsigned long flags
;
4335 idle
->timestamp
= sched_clock();
4336 idle
->sleep_avg
= 0;
4338 idle
->prio
= MAX_PRIO
;
4339 idle
->state
= TASK_RUNNING
;
4340 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4341 set_task_cpu(idle
, cpu
);
4343 spin_lock_irqsave(&rq
->lock
, flags
);
4344 rq
->curr
= rq
->idle
= idle
;
4345 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4348 spin_unlock_irqrestore(&rq
->lock
, flags
);
4350 /* Set the preempt count _outside_ the spinlocks! */
4351 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4352 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4354 task_thread_info(idle
)->preempt_count
= 0;
4359 * In a system that switches off the HZ timer nohz_cpu_mask
4360 * indicates which cpus entered this state. This is used
4361 * in the rcu update to wait only for active cpus. For system
4362 * which do not switch off the HZ timer nohz_cpu_mask should
4363 * always be CPU_MASK_NONE.
4365 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4369 * This is how migration works:
4371 * 1) we queue a migration_req_t structure in the source CPU's
4372 * runqueue and wake up that CPU's migration thread.
4373 * 2) we down() the locked semaphore => thread blocks.
4374 * 3) migration thread wakes up (implicitly it forces the migrated
4375 * thread off the CPU)
4376 * 4) it gets the migration request and checks whether the migrated
4377 * task is still in the wrong runqueue.
4378 * 5) if it's in the wrong runqueue then the migration thread removes
4379 * it and puts it into the right queue.
4380 * 6) migration thread up()s the semaphore.
4381 * 7) we wake up and the migration is done.
4385 * Change a given task's CPU affinity. Migrate the thread to a
4386 * proper CPU and schedule it away if the CPU it's executing on
4387 * is removed from the allowed bitmask.
4389 * NOTE: the caller must have a valid reference to the task, the
4390 * task must not exit() & deallocate itself prematurely. The
4391 * call is not atomic; no spinlocks may be held.
4393 int set_cpus_allowed(task_t
*p
, cpumask_t new_mask
)
4395 unsigned long flags
;
4397 migration_req_t req
;
4400 rq
= task_rq_lock(p
, &flags
);
4401 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4406 p
->cpus_allowed
= new_mask
;
4407 /* Can the task run on the task's current CPU? If so, we're done */
4408 if (cpu_isset(task_cpu(p
), new_mask
))
4411 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4412 /* Need help from migration thread: drop lock and wait. */
4413 task_rq_unlock(rq
, &flags
);
4414 wake_up_process(rq
->migration_thread
);
4415 wait_for_completion(&req
.done
);
4416 tlb_migrate_finish(p
->mm
);
4420 task_rq_unlock(rq
, &flags
);
4424 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4427 * Move (not current) task off this cpu, onto dest cpu. We're doing
4428 * this because either it can't run here any more (set_cpus_allowed()
4429 * away from this CPU, or CPU going down), or because we're
4430 * attempting to rebalance this task on exec (sched_exec).
4432 * So we race with normal scheduler movements, but that's OK, as long
4433 * as the task is no longer on this CPU.
4435 static void __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4437 runqueue_t
*rq_dest
, *rq_src
;
4439 if (unlikely(cpu_is_offline(dest_cpu
)))
4442 rq_src
= cpu_rq(src_cpu
);
4443 rq_dest
= cpu_rq(dest_cpu
);
4445 double_rq_lock(rq_src
, rq_dest
);
4446 /* Already moved. */
4447 if (task_cpu(p
) != src_cpu
)
4449 /* Affinity changed (again). */
4450 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
4453 set_task_cpu(p
, dest_cpu
);
4456 * Sync timestamp with rq_dest's before activating.
4457 * The same thing could be achieved by doing this step
4458 * afterwards, and pretending it was a local activate.
4459 * This way is cleaner and logically correct.
4461 p
->timestamp
= p
->timestamp
- rq_src
->timestamp_last_tick
4462 + rq_dest
->timestamp_last_tick
;
4463 deactivate_task(p
, rq_src
);
4464 activate_task(p
, rq_dest
, 0);
4465 if (TASK_PREEMPTS_CURR(p
, rq_dest
))
4466 resched_task(rq_dest
->curr
);
4470 double_rq_unlock(rq_src
, rq_dest
);
4474 * migration_thread - this is a highprio system thread that performs
4475 * thread migration by bumping thread off CPU then 'pushing' onto
4478 static int migration_thread(void *data
)
4481 int cpu
= (long)data
;
4484 BUG_ON(rq
->migration_thread
!= current
);
4486 set_current_state(TASK_INTERRUPTIBLE
);
4487 while (!kthread_should_stop()) {
4488 struct list_head
*head
;
4489 migration_req_t
*req
;
4493 spin_lock_irq(&rq
->lock
);
4495 if (cpu_is_offline(cpu
)) {
4496 spin_unlock_irq(&rq
->lock
);
4500 if (rq
->active_balance
) {
4501 active_load_balance(rq
, cpu
);
4502 rq
->active_balance
= 0;
4505 head
= &rq
->migration_queue
;
4507 if (list_empty(head
)) {
4508 spin_unlock_irq(&rq
->lock
);
4510 set_current_state(TASK_INTERRUPTIBLE
);
4513 req
= list_entry(head
->next
, migration_req_t
, list
);
4514 list_del_init(head
->next
);
4516 spin_unlock(&rq
->lock
);
4517 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
4520 complete(&req
->done
);
4522 __set_current_state(TASK_RUNNING
);
4526 /* Wait for kthread_stop */
4527 set_current_state(TASK_INTERRUPTIBLE
);
4528 while (!kthread_should_stop()) {
4530 set_current_state(TASK_INTERRUPTIBLE
);
4532 __set_current_state(TASK_RUNNING
);
4536 #ifdef CONFIG_HOTPLUG_CPU
4537 /* Figure out where task on dead CPU should go, use force if neccessary. */
4538 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*tsk
)
4544 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
4545 cpus_and(mask
, mask
, tsk
->cpus_allowed
);
4546 dest_cpu
= any_online_cpu(mask
);
4548 /* On any allowed CPU? */
4549 if (dest_cpu
== NR_CPUS
)
4550 dest_cpu
= any_online_cpu(tsk
->cpus_allowed
);
4552 /* No more Mr. Nice Guy. */
4553 if (dest_cpu
== NR_CPUS
) {
4554 cpus_setall(tsk
->cpus_allowed
);
4555 dest_cpu
= any_online_cpu(tsk
->cpus_allowed
);
4558 * Don't tell them about moving exiting tasks or
4559 * kernel threads (both mm NULL), since they never
4562 if (tsk
->mm
&& printk_ratelimit())
4563 printk(KERN_INFO
"process %d (%s) no "
4564 "longer affine to cpu%d\n",
4565 tsk
->pid
, tsk
->comm
, dead_cpu
);
4567 __migrate_task(tsk
, dead_cpu
, dest_cpu
);
4571 * While a dead CPU has no uninterruptible tasks queued at this point,
4572 * it might still have a nonzero ->nr_uninterruptible counter, because
4573 * for performance reasons the counter is not stricly tracking tasks to
4574 * their home CPUs. So we just add the counter to another CPU's counter,
4575 * to keep the global sum constant after CPU-down:
4577 static void migrate_nr_uninterruptible(runqueue_t
*rq_src
)
4579 runqueue_t
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
4580 unsigned long flags
;
4582 local_irq_save(flags
);
4583 double_rq_lock(rq_src
, rq_dest
);
4584 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
4585 rq_src
->nr_uninterruptible
= 0;
4586 double_rq_unlock(rq_src
, rq_dest
);
4587 local_irq_restore(flags
);
4590 /* Run through task list and migrate tasks from the dead cpu. */
4591 static void migrate_live_tasks(int src_cpu
)
4593 struct task_struct
*tsk
, *t
;
4595 write_lock_irq(&tasklist_lock
);
4597 do_each_thread(t
, tsk
) {
4601 if (task_cpu(tsk
) == src_cpu
)
4602 move_task_off_dead_cpu(src_cpu
, tsk
);
4603 } while_each_thread(t
, tsk
);
4605 write_unlock_irq(&tasklist_lock
);
4608 /* Schedules idle task to be the next runnable task on current CPU.
4609 * It does so by boosting its priority to highest possible and adding it to
4610 * the _front_ of runqueue. Used by CPU offline code.
4612 void sched_idle_next(void)
4614 int cpu
= smp_processor_id();
4615 runqueue_t
*rq
= this_rq();
4616 struct task_struct
*p
= rq
->idle
;
4617 unsigned long flags
;
4619 /* cpu has to be offline */
4620 BUG_ON(cpu_online(cpu
));
4622 /* Strictly not necessary since rest of the CPUs are stopped by now
4623 * and interrupts disabled on current cpu.
4625 spin_lock_irqsave(&rq
->lock
, flags
);
4627 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
4628 /* Add idle task to _front_ of it's priority queue */
4629 __activate_idle_task(p
, rq
);
4631 spin_unlock_irqrestore(&rq
->lock
, flags
);
4634 /* Ensures that the idle task is using init_mm right before its cpu goes
4637 void idle_task_exit(void)
4639 struct mm_struct
*mm
= current
->active_mm
;
4641 BUG_ON(cpu_online(smp_processor_id()));
4644 switch_mm(mm
, &init_mm
, current
);
4648 static void migrate_dead(unsigned int dead_cpu
, task_t
*tsk
)
4650 struct runqueue
*rq
= cpu_rq(dead_cpu
);
4652 /* Must be exiting, otherwise would be on tasklist. */
4653 BUG_ON(tsk
->exit_state
!= EXIT_ZOMBIE
&& tsk
->exit_state
!= EXIT_DEAD
);
4655 /* Cannot have done final schedule yet: would have vanished. */
4656 BUG_ON(tsk
->flags
& PF_DEAD
);
4658 get_task_struct(tsk
);
4661 * Drop lock around migration; if someone else moves it,
4662 * that's OK. No task can be added to this CPU, so iteration is
4665 spin_unlock_irq(&rq
->lock
);
4666 move_task_off_dead_cpu(dead_cpu
, tsk
);
4667 spin_lock_irq(&rq
->lock
);
4669 put_task_struct(tsk
);
4672 /* release_task() removes task from tasklist, so we won't find dead tasks. */
4673 static void migrate_dead_tasks(unsigned int dead_cpu
)
4676 struct runqueue
*rq
= cpu_rq(dead_cpu
);
4678 for (arr
= 0; arr
< 2; arr
++) {
4679 for (i
= 0; i
< MAX_PRIO
; i
++) {
4680 struct list_head
*list
= &rq
->arrays
[arr
].queue
[i
];
4681 while (!list_empty(list
))
4682 migrate_dead(dead_cpu
,
4683 list_entry(list
->next
, task_t
,
4688 #endif /* CONFIG_HOTPLUG_CPU */
4691 * migration_call - callback that gets triggered when a CPU is added.
4692 * Here we can start up the necessary migration thread for the new CPU.
4694 static int migration_call(struct notifier_block
*nfb
, unsigned long action
,
4697 int cpu
= (long)hcpu
;
4698 struct task_struct
*p
;
4699 struct runqueue
*rq
;
4700 unsigned long flags
;
4703 case CPU_UP_PREPARE
:
4704 p
= kthread_create(migration_thread
, hcpu
, "migration/%d",cpu
);
4707 p
->flags
|= PF_NOFREEZE
;
4708 kthread_bind(p
, cpu
);
4709 /* Must be high prio: stop_machine expects to yield to it. */
4710 rq
= task_rq_lock(p
, &flags
);
4711 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
4712 task_rq_unlock(rq
, &flags
);
4713 cpu_rq(cpu
)->migration_thread
= p
;
4716 /* Strictly unneccessary, as first user will wake it. */
4717 wake_up_process(cpu_rq(cpu
)->migration_thread
);
4719 #ifdef CONFIG_HOTPLUG_CPU
4720 case CPU_UP_CANCELED
:
4721 /* Unbind it from offline cpu so it can run. Fall thru. */
4722 kthread_bind(cpu_rq(cpu
)->migration_thread
,
4723 any_online_cpu(cpu_online_map
));
4724 kthread_stop(cpu_rq(cpu
)->migration_thread
);
4725 cpu_rq(cpu
)->migration_thread
= NULL
;
4728 migrate_live_tasks(cpu
);
4730 kthread_stop(rq
->migration_thread
);
4731 rq
->migration_thread
= NULL
;
4732 /* Idle task back to normal (off runqueue, low prio) */
4733 rq
= task_rq_lock(rq
->idle
, &flags
);
4734 deactivate_task(rq
->idle
, rq
);
4735 rq
->idle
->static_prio
= MAX_PRIO
;
4736 __setscheduler(rq
->idle
, SCHED_NORMAL
, 0);
4737 migrate_dead_tasks(cpu
);
4738 task_rq_unlock(rq
, &flags
);
4739 migrate_nr_uninterruptible(rq
);
4740 BUG_ON(rq
->nr_running
!= 0);
4742 /* No need to migrate the tasks: it was best-effort if
4743 * they didn't do lock_cpu_hotplug(). Just wake up
4744 * the requestors. */
4745 spin_lock_irq(&rq
->lock
);
4746 while (!list_empty(&rq
->migration_queue
)) {
4747 migration_req_t
*req
;
4748 req
= list_entry(rq
->migration_queue
.next
,
4749 migration_req_t
, list
);
4750 list_del_init(&req
->list
);
4751 complete(&req
->done
);
4753 spin_unlock_irq(&rq
->lock
);
4760 /* Register at highest priority so that task migration (migrate_all_tasks)
4761 * happens before everything else.
4763 static struct notifier_block __devinitdata migration_notifier
= {
4764 .notifier_call
= migration_call
,
4768 int __init
migration_init(void)
4770 void *cpu
= (void *)(long)smp_processor_id();
4771 /* Start one for boot CPU. */
4772 migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
4773 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
4774 register_cpu_notifier(&migration_notifier
);
4780 #undef SCHED_DOMAIN_DEBUG
4781 #ifdef SCHED_DOMAIN_DEBUG
4782 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
4787 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
4791 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
4796 struct sched_group
*group
= sd
->groups
;
4797 cpumask_t groupmask
;
4799 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
4800 cpus_clear(groupmask
);
4803 for (i
= 0; i
< level
+ 1; i
++)
4805 printk("domain %d: ", level
);
4807 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
4808 printk("does not load-balance\n");
4810 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain has parent");
4814 printk("span %s\n", str
);
4816 if (!cpu_isset(cpu
, sd
->span
))
4817 printk(KERN_ERR
"ERROR: domain->span does not contain CPU%d\n", cpu
);
4818 if (!cpu_isset(cpu
, group
->cpumask
))
4819 printk(KERN_ERR
"ERROR: domain->groups does not contain CPU%d\n", cpu
);
4822 for (i
= 0; i
< level
+ 2; i
++)
4828 printk(KERN_ERR
"ERROR: group is NULL\n");
4832 if (!group
->cpu_power
) {
4834 printk(KERN_ERR
"ERROR: domain->cpu_power not set\n");
4837 if (!cpus_weight(group
->cpumask
)) {
4839 printk(KERN_ERR
"ERROR: empty group\n");
4842 if (cpus_intersects(groupmask
, group
->cpumask
)) {
4844 printk(KERN_ERR
"ERROR: repeated CPUs\n");
4847 cpus_or(groupmask
, groupmask
, group
->cpumask
);
4849 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
4852 group
= group
->next
;
4853 } while (group
!= sd
->groups
);
4856 if (!cpus_equal(sd
->span
, groupmask
))
4857 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
4863 if (!cpus_subset(groupmask
, sd
->span
))
4864 printk(KERN_ERR
"ERROR: parent span is not a superset of domain->span\n");
4870 #define sched_domain_debug(sd, cpu) {}
4873 static int sd_degenerate(struct sched_domain
*sd
)
4875 if (cpus_weight(sd
->span
) == 1)
4878 /* Following flags need at least 2 groups */
4879 if (sd
->flags
& (SD_LOAD_BALANCE
|
4880 SD_BALANCE_NEWIDLE
|
4883 if (sd
->groups
!= sd
->groups
->next
)
4887 /* Following flags don't use groups */
4888 if (sd
->flags
& (SD_WAKE_IDLE
|
4896 static int sd_parent_degenerate(struct sched_domain
*sd
,
4897 struct sched_domain
*parent
)
4899 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
4901 if (sd_degenerate(parent
))
4904 if (!cpus_equal(sd
->span
, parent
->span
))
4907 /* Does parent contain flags not in child? */
4908 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
4909 if (cflags
& SD_WAKE_AFFINE
)
4910 pflags
&= ~SD_WAKE_BALANCE
;
4911 /* Flags needing groups don't count if only 1 group in parent */
4912 if (parent
->groups
== parent
->groups
->next
) {
4913 pflags
&= ~(SD_LOAD_BALANCE
|
4914 SD_BALANCE_NEWIDLE
|
4918 if (~cflags
& pflags
)
4925 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
4926 * hold the hotplug lock.
4928 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
4930 runqueue_t
*rq
= cpu_rq(cpu
);
4931 struct sched_domain
*tmp
;
4933 /* Remove the sched domains which do not contribute to scheduling. */
4934 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
4935 struct sched_domain
*parent
= tmp
->parent
;
4938 if (sd_parent_degenerate(tmp
, parent
))
4939 tmp
->parent
= parent
->parent
;
4942 if (sd
&& sd_degenerate(sd
))
4945 sched_domain_debug(sd
, cpu
);
4947 rcu_assign_pointer(rq
->sd
, sd
);
4950 /* cpus with isolated domains */
4951 static cpumask_t __devinitdata cpu_isolated_map
= CPU_MASK_NONE
;
4953 /* Setup the mask of cpus configured for isolated domains */
4954 static int __init
isolated_cpu_setup(char *str
)
4956 int ints
[NR_CPUS
], i
;
4958 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
4959 cpus_clear(cpu_isolated_map
);
4960 for (i
= 1; i
<= ints
[0]; i
++)
4961 if (ints
[i
] < NR_CPUS
)
4962 cpu_set(ints
[i
], cpu_isolated_map
);
4966 __setup ("isolcpus=", isolated_cpu_setup
);
4969 * init_sched_build_groups takes an array of groups, the cpumask we wish
4970 * to span, and a pointer to a function which identifies what group a CPU
4971 * belongs to. The return value of group_fn must be a valid index into the
4972 * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
4973 * keep track of groups covered with a cpumask_t).
4975 * init_sched_build_groups will build a circular linked list of the groups
4976 * covered by the given span, and will set each group's ->cpumask correctly,
4977 * and ->cpu_power to 0.
4979 static void init_sched_build_groups(struct sched_group groups
[], cpumask_t span
,
4980 int (*group_fn
)(int cpu
))
4982 struct sched_group
*first
= NULL
, *last
= NULL
;
4983 cpumask_t covered
= CPU_MASK_NONE
;
4986 for_each_cpu_mask(i
, span
) {
4987 int group
= group_fn(i
);
4988 struct sched_group
*sg
= &groups
[group
];
4991 if (cpu_isset(i
, covered
))
4994 sg
->cpumask
= CPU_MASK_NONE
;
4997 for_each_cpu_mask(j
, span
) {
4998 if (group_fn(j
) != group
)
5001 cpu_set(j
, covered
);
5002 cpu_set(j
, sg
->cpumask
);
5013 #define SD_NODES_PER_DOMAIN 16
5016 * Self-tuning task migration cost measurement between source and target CPUs.
5018 * This is done by measuring the cost of manipulating buffers of varying
5019 * sizes. For a given buffer-size here are the steps that are taken:
5021 * 1) the source CPU reads+dirties a shared buffer
5022 * 2) the target CPU reads+dirties the same shared buffer
5024 * We measure how long they take, in the following 4 scenarios:
5026 * - source: CPU1, target: CPU2 | cost1
5027 * - source: CPU2, target: CPU1 | cost2
5028 * - source: CPU1, target: CPU1 | cost3
5029 * - source: CPU2, target: CPU2 | cost4
5031 * We then calculate the cost3+cost4-cost1-cost2 difference - this is
5032 * the cost of migration.
5034 * We then start off from a small buffer-size and iterate up to larger
5035 * buffer sizes, in 5% steps - measuring each buffer-size separately, and
5036 * doing a maximum search for the cost. (The maximum cost for a migration
5037 * normally occurs when the working set size is around the effective cache
5040 #define SEARCH_SCOPE 2
5041 #define MIN_CACHE_SIZE (64*1024U)
5042 #define DEFAULT_CACHE_SIZE (5*1024*1024U)
5043 #define ITERATIONS 1
5044 #define SIZE_THRESH 130
5045 #define COST_THRESH 130
5048 * The migration cost is a function of 'domain distance'. Domain
5049 * distance is the number of steps a CPU has to iterate down its
5050 * domain tree to share a domain with the other CPU. The farther
5051 * two CPUs are from each other, the larger the distance gets.
5053 * Note that we use the distance only to cache measurement results,
5054 * the distance value is not used numerically otherwise. When two
5055 * CPUs have the same distance it is assumed that the migration
5056 * cost is the same. (this is a simplification but quite practical)
5058 #define MAX_DOMAIN_DISTANCE 32
5060 static unsigned long long migration_cost
[MAX_DOMAIN_DISTANCE
] =
5061 { [ 0 ... MAX_DOMAIN_DISTANCE
-1 ] =
5063 * Architectures may override the migration cost and thus avoid
5064 * boot-time calibration. Unit is nanoseconds. Mostly useful for
5065 * virtualized hardware:
5067 #ifdef CONFIG_DEFAULT_MIGRATION_COST
5068 CONFIG_DEFAULT_MIGRATION_COST
5075 * Allow override of migration cost - in units of microseconds.
5076 * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost
5077 * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs:
5079 static int __init
migration_cost_setup(char *str
)
5081 int ints
[MAX_DOMAIN_DISTANCE
+1], i
;
5083 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5085 printk("#ints: %d\n", ints
[0]);
5086 for (i
= 1; i
<= ints
[0]; i
++) {
5087 migration_cost
[i
-1] = (unsigned long long)ints
[i
]*1000;
5088 printk("migration_cost[%d]: %Ld\n", i
-1, migration_cost
[i
-1]);
5093 __setup ("migration_cost=", migration_cost_setup
);
5096 * Global multiplier (divisor) for migration-cutoff values,
5097 * in percentiles. E.g. use a value of 150 to get 1.5 times
5098 * longer cache-hot cutoff times.
5100 * (We scale it from 100 to 128 to long long handling easier.)
5103 #define MIGRATION_FACTOR_SCALE 128
5105 static unsigned int migration_factor
= MIGRATION_FACTOR_SCALE
;
5107 static int __init
setup_migration_factor(char *str
)
5109 get_option(&str
, &migration_factor
);
5110 migration_factor
= migration_factor
* MIGRATION_FACTOR_SCALE
/ 100;
5114 __setup("migration_factor=", setup_migration_factor
);
5117 * Estimated distance of two CPUs, measured via the number of domains
5118 * we have to pass for the two CPUs to be in the same span:
5120 static unsigned long domain_distance(int cpu1
, int cpu2
)
5122 unsigned long distance
= 0;
5123 struct sched_domain
*sd
;
5125 for_each_domain(cpu1
, sd
) {
5126 WARN_ON(!cpu_isset(cpu1
, sd
->span
));
5127 if (cpu_isset(cpu2
, sd
->span
))
5131 if (distance
>= MAX_DOMAIN_DISTANCE
) {
5133 distance
= MAX_DOMAIN_DISTANCE
-1;
5139 static unsigned int migration_debug
;
5141 static int __init
setup_migration_debug(char *str
)
5143 get_option(&str
, &migration_debug
);
5147 __setup("migration_debug=", setup_migration_debug
);
5150 * Maximum cache-size that the scheduler should try to measure.
5151 * Architectures with larger caches should tune this up during
5152 * bootup. Gets used in the domain-setup code (i.e. during SMP
5155 unsigned int max_cache_size
;
5157 static int __init
setup_max_cache_size(char *str
)
5159 get_option(&str
, &max_cache_size
);
5163 __setup("max_cache_size=", setup_max_cache_size
);
5166 * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This
5167 * is the operation that is timed, so we try to generate unpredictable
5168 * cachemisses that still end up filling the L2 cache:
5170 static void touch_cache(void *__cache
, unsigned long __size
)
5172 unsigned long size
= __size
/sizeof(long), chunk1
= size
/3,
5174 unsigned long *cache
= __cache
;
5177 for (i
= 0; i
< size
/6; i
+= 8) {
5180 case 1: cache
[size
-1-i
]++;
5181 case 2: cache
[chunk1
-i
]++;
5182 case 3: cache
[chunk1
+i
]++;
5183 case 4: cache
[chunk2
-i
]++;
5184 case 5: cache
[chunk2
+i
]++;
5190 * Measure the cache-cost of one task migration. Returns in units of nsec.
5192 static unsigned long long measure_one(void *cache
, unsigned long size
,
5193 int source
, int target
)
5195 cpumask_t mask
, saved_mask
;
5196 unsigned long long t0
, t1
, t2
, t3
, cost
;
5198 saved_mask
= current
->cpus_allowed
;
5201 * Flush source caches to RAM and invalidate them:
5206 * Migrate to the source CPU:
5208 mask
= cpumask_of_cpu(source
);
5209 set_cpus_allowed(current
, mask
);
5210 WARN_ON(smp_processor_id() != source
);
5213 * Dirty the working set:
5216 touch_cache(cache
, size
);
5220 * Migrate to the target CPU, dirty the L2 cache and access
5221 * the shared buffer. (which represents the working set
5222 * of a migrated task.)
5224 mask
= cpumask_of_cpu(target
);
5225 set_cpus_allowed(current
, mask
);
5226 WARN_ON(smp_processor_id() != target
);
5229 touch_cache(cache
, size
);
5232 cost
= t1
-t0
+ t3
-t2
;
5234 if (migration_debug
>= 2)
5235 printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n",
5236 source
, target
, t1
-t0
, t1
-t0
, t3
-t2
, cost
);
5238 * Flush target caches to RAM and invalidate them:
5242 set_cpus_allowed(current
, saved_mask
);
5248 * Measure a series of task migrations and return the average
5249 * result. Since this code runs early during bootup the system
5250 * is 'undisturbed' and the average latency makes sense.
5252 * The algorithm in essence auto-detects the relevant cache-size,
5253 * so it will properly detect different cachesizes for different
5254 * cache-hierarchies, depending on how the CPUs are connected.
5256 * Architectures can prime the upper limit of the search range via
5257 * max_cache_size, otherwise the search range defaults to 20MB...64K.
5259 static unsigned long long
5260 measure_cost(int cpu1
, int cpu2
, void *cache
, unsigned int size
)
5262 unsigned long long cost1
, cost2
;
5266 * Measure the migration cost of 'size' bytes, over an
5267 * average of 10 runs:
5269 * (We perturb the cache size by a small (0..4k)
5270 * value to compensate size/alignment related artifacts.
5271 * We also subtract the cost of the operation done on
5277 * dry run, to make sure we start off cache-cold on cpu1,
5278 * and to get any vmalloc pagefaults in advance:
5280 measure_one(cache
, size
, cpu1
, cpu2
);
5281 for (i
= 0; i
< ITERATIONS
; i
++)
5282 cost1
+= measure_one(cache
, size
- i
*1024, cpu1
, cpu2
);
5284 measure_one(cache
, size
, cpu2
, cpu1
);
5285 for (i
= 0; i
< ITERATIONS
; i
++)
5286 cost1
+= measure_one(cache
, size
- i
*1024, cpu2
, cpu1
);
5289 * (We measure the non-migrating [cached] cost on both
5290 * cpu1 and cpu2, to handle CPUs with different speeds)
5294 measure_one(cache
, size
, cpu1
, cpu1
);
5295 for (i
= 0; i
< ITERATIONS
; i
++)
5296 cost2
+= measure_one(cache
, size
- i
*1024, cpu1
, cpu1
);
5298 measure_one(cache
, size
, cpu2
, cpu2
);
5299 for (i
= 0; i
< ITERATIONS
; i
++)
5300 cost2
+= measure_one(cache
, size
- i
*1024, cpu2
, cpu2
);
5303 * Get the per-iteration migration cost:
5305 do_div(cost1
, 2*ITERATIONS
);
5306 do_div(cost2
, 2*ITERATIONS
);
5308 return cost1
- cost2
;
5311 static unsigned long long measure_migration_cost(int cpu1
, int cpu2
)
5313 unsigned long long max_cost
= 0, fluct
= 0, avg_fluct
= 0;
5314 unsigned int max_size
, size
, size_found
= 0;
5315 long long cost
= 0, prev_cost
;
5319 * Search from max_cache_size*5 down to 64K - the real relevant
5320 * cachesize has to lie somewhere inbetween.
5322 if (max_cache_size
) {
5323 max_size
= max(max_cache_size
* SEARCH_SCOPE
, MIN_CACHE_SIZE
);
5324 size
= max(max_cache_size
/ SEARCH_SCOPE
, MIN_CACHE_SIZE
);
5327 * Since we have no estimation about the relevant
5330 max_size
= DEFAULT_CACHE_SIZE
* SEARCH_SCOPE
;
5331 size
= MIN_CACHE_SIZE
;
5334 if (!cpu_online(cpu1
) || !cpu_online(cpu2
)) {
5335 printk("cpu %d and %d not both online!\n", cpu1
, cpu2
);
5340 * Allocate the working set:
5342 cache
= vmalloc(max_size
);
5344 printk("could not vmalloc %d bytes for cache!\n", 2*max_size
);
5345 return 1000000; // return 1 msec on very small boxen
5348 while (size
<= max_size
) {
5350 cost
= measure_cost(cpu1
, cpu2
, cache
, size
);
5356 if (max_cost
< cost
) {
5362 * Calculate average fluctuation, we use this to prevent
5363 * noise from triggering an early break out of the loop:
5365 fluct
= abs(cost
- prev_cost
);
5366 avg_fluct
= (avg_fluct
+ fluct
)/2;
5368 if (migration_debug
)
5369 printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): (%8Ld %8Ld)\n",
5371 (long)cost
/ 1000000,
5372 ((long)cost
/ 100000) % 10,
5373 (long)max_cost
/ 1000000,
5374 ((long)max_cost
/ 100000) % 10,
5375 domain_distance(cpu1
, cpu2
),
5379 * If we iterated at least 20% past the previous maximum,
5380 * and the cost has dropped by more than 20% already,
5381 * (taking fluctuations into account) then we assume to
5382 * have found the maximum and break out of the loop early:
5384 if (size_found
&& (size
*100 > size_found
*SIZE_THRESH
))
5385 if (cost
+avg_fluct
<= 0 ||
5386 max_cost
*100 > (cost
+avg_fluct
)*COST_THRESH
) {
5388 if (migration_debug
)
5389 printk("-> found max.\n");
5393 * Increase the cachesize in 10% steps:
5395 size
= size
* 10 / 9;
5398 if (migration_debug
)
5399 printk("[%d][%d] working set size found: %d, cost: %Ld\n",
5400 cpu1
, cpu2
, size_found
, max_cost
);
5405 * A task is considered 'cache cold' if at least 2 times
5406 * the worst-case cost of migration has passed.
5408 * (this limit is only listened to if the load-balancing
5409 * situation is 'nice' - if there is a large imbalance we
5410 * ignore it for the sake of CPU utilization and
5411 * processing fairness.)
5413 return 2 * max_cost
* migration_factor
/ MIGRATION_FACTOR_SCALE
;
5416 static void calibrate_migration_costs(const cpumask_t
*cpu_map
)
5418 int cpu1
= -1, cpu2
= -1, cpu
, orig_cpu
= raw_smp_processor_id();
5419 unsigned long j0
, j1
, distance
, max_distance
= 0;
5420 struct sched_domain
*sd
;
5425 * First pass - calculate the cacheflush times:
5427 for_each_cpu_mask(cpu1
, *cpu_map
) {
5428 for_each_cpu_mask(cpu2
, *cpu_map
) {
5431 distance
= domain_distance(cpu1
, cpu2
);
5432 max_distance
= max(max_distance
, distance
);
5434 * No result cached yet?
5436 if (migration_cost
[distance
] == -1LL)
5437 migration_cost
[distance
] =
5438 measure_migration_cost(cpu1
, cpu2
);
5442 * Second pass - update the sched domain hierarchy with
5443 * the new cache-hot-time estimations:
5445 for_each_cpu_mask(cpu
, *cpu_map
) {
5447 for_each_domain(cpu
, sd
) {
5448 sd
->cache_hot_time
= migration_cost
[distance
];
5455 if (migration_debug
)
5456 printk("migration: max_cache_size: %d, cpu: %d MHz:\n",
5464 if (system_state
== SYSTEM_BOOTING
) {
5465 printk("migration_cost=");
5466 for (distance
= 0; distance
<= max_distance
; distance
++) {
5469 printk("%ld", (long)migration_cost
[distance
] / 1000);
5474 if (migration_debug
)
5475 printk("migration: %ld seconds\n", (j1
-j0
)/HZ
);
5478 * Move back to the original CPU. NUMA-Q gets confused
5479 * if we migrate to another quad during bootup.
5481 if (raw_smp_processor_id() != orig_cpu
) {
5482 cpumask_t mask
= cpumask_of_cpu(orig_cpu
),
5483 saved_mask
= current
->cpus_allowed
;
5485 set_cpus_allowed(current
, mask
);
5486 set_cpus_allowed(current
, saved_mask
);
5493 * find_next_best_node - find the next node to include in a sched_domain
5494 * @node: node whose sched_domain we're building
5495 * @used_nodes: nodes already in the sched_domain
5497 * Find the next node to include in a given scheduling domain. Simply
5498 * finds the closest node not already in the @used_nodes map.
5500 * Should use nodemask_t.
5502 static int find_next_best_node(int node
, unsigned long *used_nodes
)
5504 int i
, n
, val
, min_val
, best_node
= 0;
5508 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5509 /* Start at @node */
5510 n
= (node
+ i
) % MAX_NUMNODES
;
5512 if (!nr_cpus_node(n
))
5515 /* Skip already used nodes */
5516 if (test_bit(n
, used_nodes
))
5519 /* Simple min distance search */
5520 val
= node_distance(node
, n
);
5522 if (val
< min_val
) {
5528 set_bit(best_node
, used_nodes
);
5533 * sched_domain_node_span - get a cpumask for a node's sched_domain
5534 * @node: node whose cpumask we're constructing
5535 * @size: number of nodes to include in this span
5537 * Given a node, construct a good cpumask for its sched_domain to span. It
5538 * should be one that prevents unnecessary balancing, but also spreads tasks
5541 static cpumask_t
sched_domain_node_span(int node
)
5544 cpumask_t span
, nodemask
;
5545 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
5548 bitmap_zero(used_nodes
, MAX_NUMNODES
);
5550 nodemask
= node_to_cpumask(node
);
5551 cpus_or(span
, span
, nodemask
);
5552 set_bit(node
, used_nodes
);
5554 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5555 int next_node
= find_next_best_node(node
, used_nodes
);
5556 nodemask
= node_to_cpumask(next_node
);
5557 cpus_or(span
, span
, nodemask
);
5565 * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we
5566 * can switch it on easily if needed.
5568 #ifdef CONFIG_SCHED_SMT
5569 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
5570 static struct sched_group sched_group_cpus
[NR_CPUS
];
5571 static int cpu_to_cpu_group(int cpu
)
5577 #ifdef CONFIG_SCHED_MC
5578 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
5579 static struct sched_group sched_group_core
[NR_CPUS
];
5582 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5583 static int cpu_to_core_group(int cpu
)
5585 return first_cpu(cpu_sibling_map
[cpu
]);
5587 #elif defined(CONFIG_SCHED_MC)
5588 static int cpu_to_core_group(int cpu
)
5594 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
5595 static struct sched_group sched_group_phys
[NR_CPUS
];
5596 static int cpu_to_phys_group(int cpu
)
5598 #if defined(CONFIG_SCHED_MC)
5599 cpumask_t mask
= cpu_coregroup_map(cpu
);
5600 return first_cpu(mask
);
5601 #elif defined(CONFIG_SCHED_SMT)
5602 return first_cpu(cpu_sibling_map
[cpu
]);
5610 * The init_sched_build_groups can't handle what we want to do with node
5611 * groups, so roll our own. Now each node has its own list of groups which
5612 * gets dynamically allocated.
5614 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
5615 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
5617 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
5618 static struct sched_group
*sched_group_allnodes_bycpu
[NR_CPUS
];
5620 static int cpu_to_allnodes_group(int cpu
)
5622 return cpu_to_node(cpu
);
5624 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
5626 struct sched_group
*sg
= group_head
;
5632 for_each_cpu_mask(j
, sg
->cpumask
) {
5633 struct sched_domain
*sd
;
5635 sd
= &per_cpu(phys_domains
, j
);
5636 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
5638 * Only add "power" once for each
5644 sg
->cpu_power
+= sd
->groups
->cpu_power
;
5647 if (sg
!= group_head
)
5653 * Build sched domains for a given set of cpus and attach the sched domains
5654 * to the individual cpus
5656 void build_sched_domains(const cpumask_t
*cpu_map
)
5660 struct sched_group
**sched_group_nodes
= NULL
;
5661 struct sched_group
*sched_group_allnodes
= NULL
;
5664 * Allocate the per-node list of sched groups
5666 sched_group_nodes
= kmalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
5668 if (!sched_group_nodes
) {
5669 printk(KERN_WARNING
"Can not alloc sched group node list\n");
5672 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
5676 * Set up domains for cpus specified by the cpu_map.
5678 for_each_cpu_mask(i
, *cpu_map
) {
5680 struct sched_domain
*sd
= NULL
, *p
;
5681 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
5683 cpus_and(nodemask
, nodemask
, *cpu_map
);
5686 if (cpus_weight(*cpu_map
)
5687 > SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
5688 if (!sched_group_allnodes
) {
5689 sched_group_allnodes
5690 = kmalloc(sizeof(struct sched_group
)
5693 if (!sched_group_allnodes
) {
5695 "Can not alloc allnodes sched group\n");
5698 sched_group_allnodes_bycpu
[i
]
5699 = sched_group_allnodes
;
5701 sd
= &per_cpu(allnodes_domains
, i
);
5702 *sd
= SD_ALLNODES_INIT
;
5703 sd
->span
= *cpu_map
;
5704 group
= cpu_to_allnodes_group(i
);
5705 sd
->groups
= &sched_group_allnodes
[group
];
5710 sd
= &per_cpu(node_domains
, i
);
5712 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
5714 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
5718 sd
= &per_cpu(phys_domains
, i
);
5719 group
= cpu_to_phys_group(i
);
5721 sd
->span
= nodemask
;
5723 sd
->groups
= &sched_group_phys
[group
];
5725 #ifdef CONFIG_SCHED_MC
5727 sd
= &per_cpu(core_domains
, i
);
5728 group
= cpu_to_core_group(i
);
5730 sd
->span
= cpu_coregroup_map(i
);
5731 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
5733 sd
->groups
= &sched_group_core
[group
];
5736 #ifdef CONFIG_SCHED_SMT
5738 sd
= &per_cpu(cpu_domains
, i
);
5739 group
= cpu_to_cpu_group(i
);
5740 *sd
= SD_SIBLING_INIT
;
5741 sd
->span
= cpu_sibling_map
[i
];
5742 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
5744 sd
->groups
= &sched_group_cpus
[group
];
5748 #ifdef CONFIG_SCHED_SMT
5749 /* Set up CPU (sibling) groups */
5750 for_each_cpu_mask(i
, *cpu_map
) {
5751 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
5752 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
5753 if (i
!= first_cpu(this_sibling_map
))
5756 init_sched_build_groups(sched_group_cpus
, this_sibling_map
,
5761 #ifdef CONFIG_SCHED_MC
5762 /* Set up multi-core groups */
5763 for_each_cpu_mask(i
, *cpu_map
) {
5764 cpumask_t this_core_map
= cpu_coregroup_map(i
);
5765 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
5766 if (i
!= first_cpu(this_core_map
))
5768 init_sched_build_groups(sched_group_core
, this_core_map
,
5769 &cpu_to_core_group
);
5774 /* Set up physical groups */
5775 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5776 cpumask_t nodemask
= node_to_cpumask(i
);
5778 cpus_and(nodemask
, nodemask
, *cpu_map
);
5779 if (cpus_empty(nodemask
))
5782 init_sched_build_groups(sched_group_phys
, nodemask
,
5783 &cpu_to_phys_group
);
5787 /* Set up node groups */
5788 if (sched_group_allnodes
)
5789 init_sched_build_groups(sched_group_allnodes
, *cpu_map
,
5790 &cpu_to_allnodes_group
);
5792 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5793 /* Set up node groups */
5794 struct sched_group
*sg
, *prev
;
5795 cpumask_t nodemask
= node_to_cpumask(i
);
5796 cpumask_t domainspan
;
5797 cpumask_t covered
= CPU_MASK_NONE
;
5800 cpus_and(nodemask
, nodemask
, *cpu_map
);
5801 if (cpus_empty(nodemask
)) {
5802 sched_group_nodes
[i
] = NULL
;
5806 domainspan
= sched_domain_node_span(i
);
5807 cpus_and(domainspan
, domainspan
, *cpu_map
);
5809 sg
= kmalloc(sizeof(struct sched_group
), GFP_KERNEL
);
5810 sched_group_nodes
[i
] = sg
;
5811 for_each_cpu_mask(j
, nodemask
) {
5812 struct sched_domain
*sd
;
5813 sd
= &per_cpu(node_domains
, j
);
5815 if (sd
->groups
== NULL
) {
5816 /* Turn off balancing if we have no groups */
5822 "Can not alloc domain group for node %d\n", i
);
5826 sg
->cpumask
= nodemask
;
5827 cpus_or(covered
, covered
, nodemask
);
5830 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
5831 cpumask_t tmp
, notcovered
;
5832 int n
= (i
+ j
) % MAX_NUMNODES
;
5834 cpus_complement(notcovered
, covered
);
5835 cpus_and(tmp
, notcovered
, *cpu_map
);
5836 cpus_and(tmp
, tmp
, domainspan
);
5837 if (cpus_empty(tmp
))
5840 nodemask
= node_to_cpumask(n
);
5841 cpus_and(tmp
, tmp
, nodemask
);
5842 if (cpus_empty(tmp
))
5845 sg
= kmalloc(sizeof(struct sched_group
), GFP_KERNEL
);
5848 "Can not alloc domain group for node %d\n", j
);
5853 cpus_or(covered
, covered
, tmp
);
5857 prev
->next
= sched_group_nodes
[i
];
5861 /* Calculate CPU power for physical packages and nodes */
5862 for_each_cpu_mask(i
, *cpu_map
) {
5864 struct sched_domain
*sd
;
5865 #ifdef CONFIG_SCHED_SMT
5866 sd
= &per_cpu(cpu_domains
, i
);
5867 power
= SCHED_LOAD_SCALE
;
5868 sd
->groups
->cpu_power
= power
;
5870 #ifdef CONFIG_SCHED_MC
5871 sd
= &per_cpu(core_domains
, i
);
5872 power
= SCHED_LOAD_SCALE
+ (cpus_weight(sd
->groups
->cpumask
)-1)
5873 * SCHED_LOAD_SCALE
/ 10;
5874 sd
->groups
->cpu_power
= power
;
5876 sd
= &per_cpu(phys_domains
, i
);
5879 * This has to be < 2 * SCHED_LOAD_SCALE
5880 * Lets keep it SCHED_LOAD_SCALE, so that
5881 * while calculating NUMA group's cpu_power
5883 * numa_group->cpu_power += phys_group->cpu_power;
5885 * See "only add power once for each physical pkg"
5888 sd
->groups
->cpu_power
= SCHED_LOAD_SCALE
;
5890 sd
= &per_cpu(phys_domains
, i
);
5891 power
= SCHED_LOAD_SCALE
+ SCHED_LOAD_SCALE
*
5892 (cpus_weight(sd
->groups
->cpumask
)-1) / 10;
5893 sd
->groups
->cpu_power
= power
;
5898 for (i
= 0; i
< MAX_NUMNODES
; i
++)
5899 init_numa_sched_groups_power(sched_group_nodes
[i
]);
5901 init_numa_sched_groups_power(sched_group_allnodes
);
5904 /* Attach the domains */
5905 for_each_cpu_mask(i
, *cpu_map
) {
5906 struct sched_domain
*sd
;
5907 #ifdef CONFIG_SCHED_SMT
5908 sd
= &per_cpu(cpu_domains
, i
);
5909 #elif defined(CONFIG_SCHED_MC)
5910 sd
= &per_cpu(core_domains
, i
);
5912 sd
= &per_cpu(phys_domains
, i
);
5914 cpu_attach_domain(sd
, i
);
5917 * Tune cache-hot values:
5919 calibrate_migration_costs(cpu_map
);
5922 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
5924 static void arch_init_sched_domains(const cpumask_t
*cpu_map
)
5926 cpumask_t cpu_default_map
;
5929 * Setup mask for cpus without special case scheduling requirements.
5930 * For now this just excludes isolated cpus, but could be used to
5931 * exclude other special cases in the future.
5933 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
5935 build_sched_domains(&cpu_default_map
);
5938 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
5944 for_each_cpu_mask(cpu
, *cpu_map
) {
5945 struct sched_group
*sched_group_allnodes
5946 = sched_group_allnodes_bycpu
[cpu
];
5947 struct sched_group
**sched_group_nodes
5948 = sched_group_nodes_bycpu
[cpu
];
5950 if (sched_group_allnodes
) {
5951 kfree(sched_group_allnodes
);
5952 sched_group_allnodes_bycpu
[cpu
] = NULL
;
5955 if (!sched_group_nodes
)
5958 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5959 cpumask_t nodemask
= node_to_cpumask(i
);
5960 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
5962 cpus_and(nodemask
, nodemask
, *cpu_map
);
5963 if (cpus_empty(nodemask
))
5973 if (oldsg
!= sched_group_nodes
[i
])
5976 kfree(sched_group_nodes
);
5977 sched_group_nodes_bycpu
[cpu
] = NULL
;
5983 * Detach sched domains from a group of cpus specified in cpu_map
5984 * These cpus will now be attached to the NULL domain
5986 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
5990 for_each_cpu_mask(i
, *cpu_map
)
5991 cpu_attach_domain(NULL
, i
);
5992 synchronize_sched();
5993 arch_destroy_sched_domains(cpu_map
);
5997 * Partition sched domains as specified by the cpumasks below.
5998 * This attaches all cpus from the cpumasks to the NULL domain,
5999 * waits for a RCU quiescent period, recalculates sched
6000 * domain information and then attaches them back to the
6001 * correct sched domains
6002 * Call with hotplug lock held
6004 void partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
6006 cpumask_t change_map
;
6008 cpus_and(*partition1
, *partition1
, cpu_online_map
);
6009 cpus_and(*partition2
, *partition2
, cpu_online_map
);
6010 cpus_or(change_map
, *partition1
, *partition2
);
6012 /* Detach sched domains from all of the affected cpus */
6013 detach_destroy_domains(&change_map
);
6014 if (!cpus_empty(*partition1
))
6015 build_sched_domains(partition1
);
6016 if (!cpus_empty(*partition2
))
6017 build_sched_domains(partition2
);
6020 #ifdef CONFIG_HOTPLUG_CPU
6022 * Force a reinitialization of the sched domains hierarchy. The domains
6023 * and groups cannot be updated in place without racing with the balancing
6024 * code, so we temporarily attach all running cpus to the NULL domain
6025 * which will prevent rebalancing while the sched domains are recalculated.
6027 static int update_sched_domains(struct notifier_block
*nfb
,
6028 unsigned long action
, void *hcpu
)
6031 case CPU_UP_PREPARE
:
6032 case CPU_DOWN_PREPARE
:
6033 detach_destroy_domains(&cpu_online_map
);
6036 case CPU_UP_CANCELED
:
6037 case CPU_DOWN_FAILED
:
6041 * Fall through and re-initialise the domains.
6048 /* The hotplug lock is already held by cpu_up/cpu_down */
6049 arch_init_sched_domains(&cpu_online_map
);
6055 void __init
sched_init_smp(void)
6058 arch_init_sched_domains(&cpu_online_map
);
6059 unlock_cpu_hotplug();
6060 /* XXX: Theoretical race here - CPU may be hotplugged now */
6061 hotcpu_notifier(update_sched_domains
, 0);
6064 void __init
sched_init_smp(void)
6067 #endif /* CONFIG_SMP */
6069 int in_sched_functions(unsigned long addr
)
6071 /* Linker adds these: start and end of __sched functions */
6072 extern char __sched_text_start
[], __sched_text_end
[];
6073 return in_lock_functions(addr
) ||
6074 (addr
>= (unsigned long)__sched_text_start
6075 && addr
< (unsigned long)__sched_text_end
);
6078 void __init
sched_init(void)
6083 for_each_possible_cpu(i
) {
6084 prio_array_t
*array
;
6087 spin_lock_init(&rq
->lock
);
6089 rq
->active
= rq
->arrays
;
6090 rq
->expired
= rq
->arrays
+ 1;
6091 rq
->best_expired_prio
= MAX_PRIO
;
6095 for (j
= 1; j
< 3; j
++)
6096 rq
->cpu_load
[j
] = 0;
6097 rq
->active_balance
= 0;
6099 rq
->migration_thread
= NULL
;
6100 INIT_LIST_HEAD(&rq
->migration_queue
);
6103 atomic_set(&rq
->nr_iowait
, 0);
6105 for (j
= 0; j
< 2; j
++) {
6106 array
= rq
->arrays
+ j
;
6107 for (k
= 0; k
< MAX_PRIO
; k
++) {
6108 INIT_LIST_HEAD(array
->queue
+ k
);
6109 __clear_bit(k
, array
->bitmap
);
6111 // delimiter for bitsearch
6112 __set_bit(MAX_PRIO
, array
->bitmap
);
6117 * The boot idle thread does lazy MMU switching as well:
6119 atomic_inc(&init_mm
.mm_count
);
6120 enter_lazy_tlb(&init_mm
, current
);
6123 * Make us the idle thread. Technically, schedule() should not be
6124 * called from this thread, however somewhere below it might be,
6125 * but because we are the idle thread, we just pick up running again
6126 * when this runqueue becomes "idle".
6128 init_idle(current
, smp_processor_id());
6131 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6132 void __might_sleep(char *file
, int line
)
6134 #if defined(in_atomic)
6135 static unsigned long prev_jiffy
; /* ratelimiting */
6137 if ((in_atomic() || irqs_disabled()) &&
6138 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6139 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6141 prev_jiffy
= jiffies
;
6142 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6143 " context at %s:%d\n", file
, line
);
6144 printk("in_atomic():%d, irqs_disabled():%d\n",
6145 in_atomic(), irqs_disabled());
6150 EXPORT_SYMBOL(__might_sleep
);
6153 #ifdef CONFIG_MAGIC_SYSRQ
6154 void normalize_rt_tasks(void)
6156 struct task_struct
*p
;
6157 prio_array_t
*array
;
6158 unsigned long flags
;
6161 read_lock_irq(&tasklist_lock
);
6162 for_each_process (p
) {
6166 rq
= task_rq_lock(p
, &flags
);
6170 deactivate_task(p
, task_rq(p
));
6171 __setscheduler(p
, SCHED_NORMAL
, 0);
6173 __activate_task(p
, task_rq(p
));
6174 resched_task(rq
->curr
);
6177 task_rq_unlock(rq
, &flags
);
6179 read_unlock_irq(&tasklist_lock
);
6182 #endif /* CONFIG_MAGIC_SYSRQ */
6186 * These functions are only useful for the IA64 MCA handling.
6188 * They can only be called when the whole system has been
6189 * stopped - every CPU needs to be quiescent, and no scheduling
6190 * activity can take place. Using them for anything else would
6191 * be a serious bug, and as a result, they aren't even visible
6192 * under any other configuration.
6196 * curr_task - return the current task for a given cpu.
6197 * @cpu: the processor in question.
6199 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6201 task_t
*curr_task(int cpu
)
6203 return cpu_curr(cpu
);
6207 * set_curr_task - set the current task for a given cpu.
6208 * @cpu: the processor in question.
6209 * @p: the task pointer to set.
6211 * Description: This function must only be used when non-maskable interrupts
6212 * are serviced on a separate stack. It allows the architecture to switch the
6213 * notion of the current task on a cpu in a non-blocking manner. This function
6214 * must be called with all CPU's synchronized, and interrupts disabled, the
6215 * and caller must save the original value of the current task (see
6216 * curr_task() above) and restore that value before reenabling interrupts and
6217 * re-starting the system.
6219 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6221 void set_curr_task(int cpu
, task_t
*p
)