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 static_prio_timeslice(int static_prio
)
175 if (static_prio
< NICE_TO_PRIO(0))
176 return SCALE_PRIO(DEF_TIMESLICE
* 4, static_prio
);
178 return SCALE_PRIO(DEF_TIMESLICE
, static_prio
);
181 static inline unsigned int task_timeslice(task_t
*p
)
183 return static_prio_timeslice(p
->static_prio
);
186 #define task_hot(p, now, sd) ((long long) ((now) - (p)->last_ran) \
187 < (long long) (sd)->cache_hot_time)
190 * These are the runqueue data structures:
193 typedef struct runqueue runqueue_t
;
196 unsigned int nr_active
;
197 DECLARE_BITMAP(bitmap
, MAX_PRIO
+1); /* include 1 bit for delimiter */
198 struct list_head queue
[MAX_PRIO
];
202 * This is the main, per-CPU runqueue data structure.
204 * Locking rule: those places that want to lock multiple runqueues
205 * (such as the load balancing or the thread migration code), lock
206 * acquire operations must be ordered by ascending &runqueue.
212 * nr_running and cpu_load should be in the same cacheline because
213 * remote CPUs use both these fields when doing load calculation.
215 unsigned long nr_running
;
216 unsigned long raw_weighted_load
;
218 unsigned long cpu_load
[3];
220 unsigned long long nr_switches
;
223 * This is part of a global counter where only the total sum
224 * over all CPUs matters. A task can increase this counter on
225 * one CPU and if it got migrated afterwards it may decrease
226 * it on another CPU. Always updated under the runqueue lock:
228 unsigned long nr_uninterruptible
;
230 unsigned long expired_timestamp
;
231 unsigned long long timestamp_last_tick
;
233 struct mm_struct
*prev_mm
;
234 prio_array_t
*active
, *expired
, arrays
[2];
235 int best_expired_prio
;
239 struct sched_domain
*sd
;
241 /* For active balancing */
245 task_t
*migration_thread
;
246 struct list_head migration_queue
;
249 #ifdef CONFIG_SCHEDSTATS
251 struct sched_info rq_sched_info
;
253 /* sys_sched_yield() stats */
254 unsigned long yld_exp_empty
;
255 unsigned long yld_act_empty
;
256 unsigned long yld_both_empty
;
257 unsigned long yld_cnt
;
259 /* schedule() stats */
260 unsigned long sched_switch
;
261 unsigned long sched_cnt
;
262 unsigned long sched_goidle
;
264 /* try_to_wake_up() stats */
265 unsigned long ttwu_cnt
;
266 unsigned long ttwu_local
;
270 static DEFINE_PER_CPU(struct runqueue
, runqueues
);
273 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
274 * See detach_destroy_domains: synchronize_sched for details.
276 * The domain tree of any CPU may only be accessed from within
277 * preempt-disabled sections.
279 #define for_each_domain(cpu, domain) \
280 for (domain = rcu_dereference(cpu_rq(cpu)->sd); domain; domain = domain->parent)
282 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
283 #define this_rq() (&__get_cpu_var(runqueues))
284 #define task_rq(p) cpu_rq(task_cpu(p))
285 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
287 #ifndef prepare_arch_switch
288 # define prepare_arch_switch(next) do { } while (0)
290 #ifndef finish_arch_switch
291 # define finish_arch_switch(prev) do { } while (0)
294 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
295 static inline int task_running(runqueue_t
*rq
, task_t
*p
)
297 return rq
->curr
== p
;
300 static inline void prepare_lock_switch(runqueue_t
*rq
, task_t
*next
)
304 static inline void finish_lock_switch(runqueue_t
*rq
, task_t
*prev
)
306 #ifdef CONFIG_DEBUG_SPINLOCK
307 /* this is a valid case when another task releases the spinlock */
308 rq
->lock
.owner
= current
;
310 spin_unlock_irq(&rq
->lock
);
313 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
314 static inline int task_running(runqueue_t
*rq
, task_t
*p
)
319 return rq
->curr
== p
;
323 static inline void prepare_lock_switch(runqueue_t
*rq
, task_t
*next
)
327 * We can optimise this out completely for !SMP, because the
328 * SMP rebalancing from interrupt is the only thing that cares
333 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
334 spin_unlock_irq(&rq
->lock
);
336 spin_unlock(&rq
->lock
);
340 static inline void finish_lock_switch(runqueue_t
*rq
, task_t
*prev
)
344 * After ->oncpu is cleared, the task can be moved to a different CPU.
345 * We must ensure this doesn't happen until the switch is completely
351 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
355 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
358 * task_rq_lock - lock the runqueue a given task resides on and disable
359 * interrupts. Note the ordering: we can safely lookup the task_rq without
360 * explicitly disabling preemption.
362 static runqueue_t
*task_rq_lock(task_t
*p
, unsigned long *flags
)
368 local_irq_save(*flags
);
370 spin_lock(&rq
->lock
);
371 if (unlikely(rq
!= task_rq(p
))) {
372 spin_unlock_irqrestore(&rq
->lock
, *flags
);
373 goto repeat_lock_task
;
378 static inline void task_rq_unlock(runqueue_t
*rq
, unsigned long *flags
)
381 spin_unlock_irqrestore(&rq
->lock
, *flags
);
384 #ifdef CONFIG_SCHEDSTATS
386 * bump this up when changing the output format or the meaning of an existing
387 * format, so that tools can adapt (or abort)
389 #define SCHEDSTAT_VERSION 12
391 static int show_schedstat(struct seq_file
*seq
, void *v
)
395 seq_printf(seq
, "version %d\n", SCHEDSTAT_VERSION
);
396 seq_printf(seq
, "timestamp %lu\n", jiffies
);
397 for_each_online_cpu(cpu
) {
398 runqueue_t
*rq
= cpu_rq(cpu
);
400 struct sched_domain
*sd
;
404 /* runqueue-specific stats */
406 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
407 cpu
, rq
->yld_both_empty
,
408 rq
->yld_act_empty
, rq
->yld_exp_empty
, rq
->yld_cnt
,
409 rq
->sched_switch
, rq
->sched_cnt
, rq
->sched_goidle
,
410 rq
->ttwu_cnt
, rq
->ttwu_local
,
411 rq
->rq_sched_info
.cpu_time
,
412 rq
->rq_sched_info
.run_delay
, rq
->rq_sched_info
.pcnt
);
414 seq_printf(seq
, "\n");
417 /* domain-specific stats */
419 for_each_domain(cpu
, sd
) {
420 enum idle_type itype
;
421 char mask_str
[NR_CPUS
];
423 cpumask_scnprintf(mask_str
, NR_CPUS
, sd
->span
);
424 seq_printf(seq
, "domain%d %s", dcnt
++, mask_str
);
425 for (itype
= SCHED_IDLE
; itype
< MAX_IDLE_TYPES
;
427 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu %lu",
429 sd
->lb_balanced
[itype
],
430 sd
->lb_failed
[itype
],
431 sd
->lb_imbalance
[itype
],
432 sd
->lb_gained
[itype
],
433 sd
->lb_hot_gained
[itype
],
434 sd
->lb_nobusyq
[itype
],
435 sd
->lb_nobusyg
[itype
]);
437 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
438 sd
->alb_cnt
, sd
->alb_failed
, sd
->alb_pushed
,
439 sd
->sbe_cnt
, sd
->sbe_balanced
, sd
->sbe_pushed
,
440 sd
->sbf_cnt
, sd
->sbf_balanced
, sd
->sbf_pushed
,
441 sd
->ttwu_wake_remote
, sd
->ttwu_move_affine
, sd
->ttwu_move_balance
);
449 static int schedstat_open(struct inode
*inode
, struct file
*file
)
451 unsigned int size
= PAGE_SIZE
* (1 + num_online_cpus() / 32);
452 char *buf
= kmalloc(size
, GFP_KERNEL
);
458 res
= single_open(file
, show_schedstat
, NULL
);
460 m
= file
->private_data
;
468 struct file_operations proc_schedstat_operations
= {
469 .open
= schedstat_open
,
472 .release
= single_release
,
475 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
476 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
477 #else /* !CONFIG_SCHEDSTATS */
478 # define schedstat_inc(rq, field) do { } while (0)
479 # define schedstat_add(rq, field, amt) do { } while (0)
483 * rq_lock - lock a given runqueue and disable interrupts.
485 static inline runqueue_t
*this_rq_lock(void)
492 spin_lock(&rq
->lock
);
497 #ifdef CONFIG_SCHEDSTATS
499 * Called when a process is dequeued from the active array and given
500 * the cpu. We should note that with the exception of interactive
501 * tasks, the expired queue will become the active queue after the active
502 * queue is empty, without explicitly dequeuing and requeuing tasks in the
503 * expired queue. (Interactive tasks may be requeued directly to the
504 * active queue, thus delaying tasks in the expired queue from running;
505 * see scheduler_tick()).
507 * This function is only called from sched_info_arrive(), rather than
508 * dequeue_task(). Even though a task may be queued and dequeued multiple
509 * times as it is shuffled about, we're really interested in knowing how
510 * long it was from the *first* time it was queued to the time that it
513 static inline void sched_info_dequeued(task_t
*t
)
515 t
->sched_info
.last_queued
= 0;
519 * Called when a task finally hits the cpu. We can now calculate how
520 * long it was waiting to run. We also note when it began so that we
521 * can keep stats on how long its timeslice is.
523 static void sched_info_arrive(task_t
*t
)
525 unsigned long now
= jiffies
, diff
= 0;
526 struct runqueue
*rq
= task_rq(t
);
528 if (t
->sched_info
.last_queued
)
529 diff
= now
- t
->sched_info
.last_queued
;
530 sched_info_dequeued(t
);
531 t
->sched_info
.run_delay
+= diff
;
532 t
->sched_info
.last_arrival
= now
;
533 t
->sched_info
.pcnt
++;
538 rq
->rq_sched_info
.run_delay
+= diff
;
539 rq
->rq_sched_info
.pcnt
++;
543 * Called when a process is queued into either the active or expired
544 * array. The time is noted and later used to determine how long we
545 * had to wait for us to reach the cpu. Since the expired queue will
546 * become the active queue after active queue is empty, without dequeuing
547 * and requeuing any tasks, we are interested in queuing to either. It
548 * is unusual but not impossible for tasks to be dequeued and immediately
549 * requeued in the same or another array: this can happen in sched_yield(),
550 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
553 * This function is only called from enqueue_task(), but also only updates
554 * the timestamp if it is already not set. It's assumed that
555 * sched_info_dequeued() will clear that stamp when appropriate.
557 static inline void sched_info_queued(task_t
*t
)
559 if (!t
->sched_info
.last_queued
)
560 t
->sched_info
.last_queued
= jiffies
;
564 * Called when a process ceases being the active-running process, either
565 * voluntarily or involuntarily. Now we can calculate how long we ran.
567 static inline void sched_info_depart(task_t
*t
)
569 struct runqueue
*rq
= task_rq(t
);
570 unsigned long diff
= jiffies
- t
->sched_info
.last_arrival
;
572 t
->sched_info
.cpu_time
+= diff
;
575 rq
->rq_sched_info
.cpu_time
+= diff
;
579 * Called when tasks are switched involuntarily due, typically, to expiring
580 * their time slice. (This may also be called when switching to or from
581 * the idle task.) We are only called when prev != next.
583 static inline void sched_info_switch(task_t
*prev
, task_t
*next
)
585 struct runqueue
*rq
= task_rq(prev
);
588 * prev now departs the cpu. It's not interesting to record
589 * stats about how efficient we were at scheduling the idle
592 if (prev
!= rq
->idle
)
593 sched_info_depart(prev
);
595 if (next
!= rq
->idle
)
596 sched_info_arrive(next
);
599 #define sched_info_queued(t) do { } while (0)
600 #define sched_info_switch(t, next) do { } while (0)
601 #endif /* CONFIG_SCHEDSTATS */
604 * Adding/removing a task to/from a priority array:
606 static void dequeue_task(struct task_struct
*p
, prio_array_t
*array
)
609 list_del(&p
->run_list
);
610 if (list_empty(array
->queue
+ p
->prio
))
611 __clear_bit(p
->prio
, array
->bitmap
);
614 static void enqueue_task(struct task_struct
*p
, prio_array_t
*array
)
616 sched_info_queued(p
);
617 list_add_tail(&p
->run_list
, array
->queue
+ p
->prio
);
618 __set_bit(p
->prio
, array
->bitmap
);
624 * Put task to the end of the run list without the overhead of dequeue
625 * followed by enqueue.
627 static void requeue_task(struct task_struct
*p
, prio_array_t
*array
)
629 list_move_tail(&p
->run_list
, array
->queue
+ p
->prio
);
632 static inline void enqueue_task_head(struct task_struct
*p
, prio_array_t
*array
)
634 list_add(&p
->run_list
, array
->queue
+ p
->prio
);
635 __set_bit(p
->prio
, array
->bitmap
);
641 * effective_prio - return the priority that is based on the static
642 * priority but is modified by bonuses/penalties.
644 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
645 * into the -5 ... 0 ... +5 bonus/penalty range.
647 * We use 25% of the full 0...39 priority range so that:
649 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
650 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
652 * Both properties are important to certain workloads.
654 static int effective_prio(task_t
*p
)
661 bonus
= CURRENT_BONUS(p
) - MAX_BONUS
/ 2;
663 prio
= p
->static_prio
- bonus
;
664 if (prio
< MAX_RT_PRIO
)
666 if (prio
> MAX_PRIO
-1)
672 * To aid in avoiding the subversion of "niceness" due to uneven distribution
673 * of tasks with abnormal "nice" values across CPUs the contribution that
674 * each task makes to its run queue's load is weighted according to its
675 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
676 * scaled version of the new time slice allocation that they receive on time
681 * Assume: static_prio_timeslice(NICE_TO_PRIO(0)) == DEF_TIMESLICE
682 * If static_prio_timeslice() is ever changed to break this assumption then
683 * this code will need modification
685 #define TIME_SLICE_NICE_ZERO DEF_TIMESLICE
686 #define LOAD_WEIGHT(lp) \
687 (((lp) * SCHED_LOAD_SCALE) / TIME_SLICE_NICE_ZERO)
688 #define PRIO_TO_LOAD_WEIGHT(prio) \
689 LOAD_WEIGHT(static_prio_timeslice(prio))
690 #define RTPRIO_TO_LOAD_WEIGHT(rp) \
691 (PRIO_TO_LOAD_WEIGHT(MAX_RT_PRIO) + LOAD_WEIGHT(rp))
693 static void set_load_weight(task_t
*p
)
697 if (p
== task_rq(p
)->migration_thread
)
699 * The migration thread does the actual balancing.
700 * Giving its load any weight will skew balancing
706 p
->load_weight
= RTPRIO_TO_LOAD_WEIGHT(p
->rt_priority
);
708 p
->load_weight
= PRIO_TO_LOAD_WEIGHT(p
->static_prio
);
711 static inline void inc_raw_weighted_load(runqueue_t
*rq
, const task_t
*p
)
713 rq
->raw_weighted_load
+= p
->load_weight
;
716 static inline void dec_raw_weighted_load(runqueue_t
*rq
, const task_t
*p
)
718 rq
->raw_weighted_load
-= p
->load_weight
;
721 static inline void inc_nr_running(task_t
*p
, runqueue_t
*rq
)
724 inc_raw_weighted_load(rq
, p
);
727 static inline void dec_nr_running(task_t
*p
, runqueue_t
*rq
)
730 dec_raw_weighted_load(rq
, p
);
734 * __activate_task - move a task to the runqueue.
736 static void __activate_task(task_t
*p
, runqueue_t
*rq
)
738 prio_array_t
*target
= rq
->active
;
741 target
= rq
->expired
;
742 enqueue_task(p
, target
);
743 inc_nr_running(p
, rq
);
747 * __activate_idle_task - move idle task to the _front_ of runqueue.
749 static inline void __activate_idle_task(task_t
*p
, runqueue_t
*rq
)
751 enqueue_task_head(p
, rq
->active
);
752 inc_nr_running(p
, rq
);
755 static int recalc_task_prio(task_t
*p
, unsigned long long now
)
757 /* Caller must always ensure 'now >= p->timestamp' */
758 unsigned long sleep_time
= now
- p
->timestamp
;
763 if (likely(sleep_time
> 0)) {
765 * This ceiling is set to the lowest priority that would allow
766 * a task to be reinserted into the active array on timeslice
769 unsigned long ceiling
= INTERACTIVE_SLEEP(p
);
771 if (p
->mm
&& sleep_time
> ceiling
&& p
->sleep_avg
< ceiling
) {
773 * Prevents user tasks from achieving best priority
774 * with one single large enough sleep.
776 p
->sleep_avg
= ceiling
;
778 * Using INTERACTIVE_SLEEP() as a ceiling places a
779 * nice(0) task 1ms sleep away from promotion, and
780 * gives it 700ms to round-robin with no chance of
781 * being demoted. This is more than generous, so
782 * mark this sleep as non-interactive to prevent the
783 * on-runqueue bonus logic from intervening should
784 * this task not receive cpu immediately.
786 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
789 * Tasks waking from uninterruptible sleep are
790 * limited in their sleep_avg rise as they
791 * are likely to be waiting on I/O
793 if (p
->sleep_type
== SLEEP_NONINTERACTIVE
&& p
->mm
) {
794 if (p
->sleep_avg
>= ceiling
)
796 else if (p
->sleep_avg
+ sleep_time
>=
798 p
->sleep_avg
= ceiling
;
804 * This code gives a bonus to interactive tasks.
806 * The boost works by updating the 'average sleep time'
807 * value here, based on ->timestamp. The more time a
808 * task spends sleeping, the higher the average gets -
809 * and the higher the priority boost gets as well.
811 p
->sleep_avg
+= sleep_time
;
814 if (p
->sleep_avg
> NS_MAX_SLEEP_AVG
)
815 p
->sleep_avg
= NS_MAX_SLEEP_AVG
;
818 return effective_prio(p
);
822 * activate_task - move a task to the runqueue and do priority recalculation
824 * Update all the scheduling statistics stuff. (sleep average
825 * calculation, priority modifiers, etc.)
827 static void activate_task(task_t
*p
, runqueue_t
*rq
, int local
)
829 unsigned long long now
;
834 /* Compensate for drifting sched_clock */
835 runqueue_t
*this_rq
= this_rq();
836 now
= (now
- this_rq
->timestamp_last_tick
)
837 + rq
->timestamp_last_tick
;
842 p
->prio
= recalc_task_prio(p
, now
);
845 * This checks to make sure it's not an uninterruptible task
846 * that is now waking up.
848 if (p
->sleep_type
== SLEEP_NORMAL
) {
850 * Tasks which were woken up by interrupts (ie. hw events)
851 * are most likely of interactive nature. So we give them
852 * the credit of extending their sleep time to the period
853 * of time they spend on the runqueue, waiting for execution
854 * on a CPU, first time around:
857 p
->sleep_type
= SLEEP_INTERRUPTED
;
860 * Normal first-time wakeups get a credit too for
861 * on-runqueue time, but it will be weighted down:
863 p
->sleep_type
= SLEEP_INTERACTIVE
;
868 __activate_task(p
, rq
);
872 * deactivate_task - remove a task from the runqueue.
874 static void deactivate_task(struct task_struct
*p
, runqueue_t
*rq
)
876 dec_nr_running(p
, rq
);
877 dequeue_task(p
, p
->array
);
882 * resched_task - mark a task 'to be rescheduled now'.
884 * On UP this means the setting of the need_resched flag, on SMP it
885 * might also involve a cross-CPU call to trigger the scheduler on
890 #ifndef tsk_is_polling
891 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
894 static void resched_task(task_t
*p
)
898 assert_spin_locked(&task_rq(p
)->lock
);
900 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
903 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
906 if (cpu
== smp_processor_id())
909 /* NEED_RESCHED must be visible before we test polling */
911 if (!tsk_is_polling(p
))
912 smp_send_reschedule(cpu
);
915 static inline void resched_task(task_t
*p
)
917 assert_spin_locked(&task_rq(p
)->lock
);
918 set_tsk_need_resched(p
);
923 * task_curr - is this task currently executing on a CPU?
924 * @p: the task in question.
926 inline int task_curr(const task_t
*p
)
928 return cpu_curr(task_cpu(p
)) == p
;
931 /* Used instead of source_load when we know the type == 0 */
932 unsigned long weighted_cpuload(const int cpu
)
934 return cpu_rq(cpu
)->raw_weighted_load
;
939 struct list_head list
;
944 struct completion done
;
948 * The task's runqueue lock must be held.
949 * Returns true if you have to wait for migration thread.
951 static int migrate_task(task_t
*p
, int dest_cpu
, migration_req_t
*req
)
953 runqueue_t
*rq
= task_rq(p
);
956 * If the task is not on a runqueue (and not running), then
957 * it is sufficient to simply update the task's cpu field.
959 if (!p
->array
&& !task_running(rq
, p
)) {
960 set_task_cpu(p
, dest_cpu
);
964 init_completion(&req
->done
);
966 req
->dest_cpu
= dest_cpu
;
967 list_add(&req
->list
, &rq
->migration_queue
);
972 * wait_task_inactive - wait for a thread to unschedule.
974 * The caller must ensure that the task *will* unschedule sometime soon,
975 * else this function might spin for a *long* time. This function can't
976 * be called with interrupts off, or it may introduce deadlock with
977 * smp_call_function() if an IPI is sent by the same process we are
978 * waiting to become inactive.
980 void wait_task_inactive(task_t
*p
)
987 rq
= task_rq_lock(p
, &flags
);
988 /* Must be off runqueue entirely, not preempted. */
989 if (unlikely(p
->array
|| task_running(rq
, p
))) {
990 /* If it's preempted, we yield. It could be a while. */
991 preempted
= !task_running(rq
, p
);
992 task_rq_unlock(rq
, &flags
);
998 task_rq_unlock(rq
, &flags
);
1002 * kick_process - kick a running thread to enter/exit the kernel
1003 * @p: the to-be-kicked thread
1005 * Cause a process which is running on another CPU to enter
1006 * kernel-mode, without any delay. (to get signals handled.)
1008 * NOTE: this function doesnt have to take the runqueue lock,
1009 * because all it wants to ensure is that the remote task enters
1010 * the kernel. If the IPI races and the task has been migrated
1011 * to another CPU then no harm is done and the purpose has been
1014 void kick_process(task_t
*p
)
1020 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1021 smp_send_reschedule(cpu
);
1026 * Return a low guess at the load of a migration-source cpu weighted
1027 * according to the scheduling class and "nice" value.
1029 * We want to under-estimate the load of migration sources, to
1030 * balance conservatively.
1032 static inline unsigned long source_load(int cpu
, int type
)
1034 runqueue_t
*rq
= cpu_rq(cpu
);
1037 return rq
->raw_weighted_load
;
1039 return min(rq
->cpu_load
[type
-1], rq
->raw_weighted_load
);
1043 * Return a high guess at the load of a migration-target cpu weighted
1044 * according to the scheduling class and "nice" value.
1046 static inline unsigned long target_load(int cpu
, int type
)
1048 runqueue_t
*rq
= cpu_rq(cpu
);
1051 return rq
->raw_weighted_load
;
1053 return max(rq
->cpu_load
[type
-1], rq
->raw_weighted_load
);
1057 * Return the average load per task on the cpu's run queue
1059 static inline unsigned long cpu_avg_load_per_task(int cpu
)
1061 runqueue_t
*rq
= cpu_rq(cpu
);
1062 unsigned long n
= rq
->nr_running
;
1064 return n
? rq
->raw_weighted_load
/ n
: SCHED_LOAD_SCALE
;
1068 * find_idlest_group finds and returns the least busy CPU group within the
1071 static struct sched_group
*
1072 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1074 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1075 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1076 int load_idx
= sd
->forkexec_idx
;
1077 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1080 unsigned long load
, avg_load
;
1084 /* Skip over this group if it has no CPUs allowed */
1085 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1088 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1090 /* Tally up the load of all CPUs in the group */
1093 for_each_cpu_mask(i
, group
->cpumask
) {
1094 /* Bias balancing toward cpus of our domain */
1096 load
= source_load(i
, load_idx
);
1098 load
= target_load(i
, load_idx
);
1103 /* Adjust by relative CPU power of the group */
1104 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
1107 this_load
= avg_load
;
1109 } else if (avg_load
< min_load
) {
1110 min_load
= avg_load
;
1114 group
= group
->next
;
1115 } while (group
!= sd
->groups
);
1117 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1123 * find_idlest_queue - find the idlest runqueue among the cpus in group.
1126 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1129 unsigned long load
, min_load
= ULONG_MAX
;
1133 /* Traverse only the allowed CPUs */
1134 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1136 for_each_cpu_mask(i
, tmp
) {
1137 load
= weighted_cpuload(i
);
1139 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1149 * sched_balance_self: balance the current task (running on cpu) in domains
1150 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1153 * Balance, ie. select the least loaded group.
1155 * Returns the target CPU number, or the same CPU if no balancing is needed.
1157 * preempt must be disabled.
1159 static int sched_balance_self(int cpu
, int flag
)
1161 struct task_struct
*t
= current
;
1162 struct sched_domain
*tmp
, *sd
= NULL
;
1164 for_each_domain(cpu
, tmp
) {
1166 * If power savings logic is enabled for a domain, stop there.
1168 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1170 if (tmp
->flags
& flag
)
1176 struct sched_group
*group
;
1181 group
= find_idlest_group(sd
, t
, cpu
);
1185 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1186 if (new_cpu
== -1 || new_cpu
== cpu
)
1189 /* Now try balancing at a lower domain level */
1193 weight
= cpus_weight(span
);
1194 for_each_domain(cpu
, tmp
) {
1195 if (weight
<= cpus_weight(tmp
->span
))
1197 if (tmp
->flags
& flag
)
1200 /* while loop will break here if sd == NULL */
1206 #endif /* CONFIG_SMP */
1209 * wake_idle() will wake a task on an idle cpu if task->cpu is
1210 * not idle and an idle cpu is available. The span of cpus to
1211 * search starts with cpus closest then further out as needed,
1212 * so we always favor a closer, idle cpu.
1214 * Returns the CPU we should wake onto.
1216 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1217 static int wake_idle(int cpu
, task_t
*p
)
1220 struct sched_domain
*sd
;
1226 for_each_domain(cpu
, sd
) {
1227 if (sd
->flags
& SD_WAKE_IDLE
) {
1228 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1229 for_each_cpu_mask(i
, tmp
) {
1240 static inline int wake_idle(int cpu
, task_t
*p
)
1247 * try_to_wake_up - wake up a thread
1248 * @p: the to-be-woken-up thread
1249 * @state: the mask of task states that can be woken
1250 * @sync: do a synchronous wakeup?
1252 * Put it on the run-queue if it's not already there. The "current"
1253 * thread is always on the run-queue (except when the actual
1254 * re-schedule is in progress), and as such you're allowed to do
1255 * the simpler "current->state = TASK_RUNNING" to mark yourself
1256 * runnable without the overhead of this.
1258 * returns failure only if the task is already active.
1260 static int try_to_wake_up(task_t
*p
, unsigned int state
, int sync
)
1262 int cpu
, this_cpu
, success
= 0;
1263 unsigned long flags
;
1267 unsigned long load
, this_load
;
1268 struct sched_domain
*sd
, *this_sd
= NULL
;
1272 rq
= task_rq_lock(p
, &flags
);
1273 old_state
= p
->state
;
1274 if (!(old_state
& state
))
1281 this_cpu
= smp_processor_id();
1284 if (unlikely(task_running(rq
, p
)))
1289 schedstat_inc(rq
, ttwu_cnt
);
1290 if (cpu
== this_cpu
) {
1291 schedstat_inc(rq
, ttwu_local
);
1295 for_each_domain(this_cpu
, sd
) {
1296 if (cpu_isset(cpu
, sd
->span
)) {
1297 schedstat_inc(sd
, ttwu_wake_remote
);
1303 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1307 * Check for affine wakeup and passive balancing possibilities.
1310 int idx
= this_sd
->wake_idx
;
1311 unsigned int imbalance
;
1313 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1315 load
= source_load(cpu
, idx
);
1316 this_load
= target_load(this_cpu
, idx
);
1318 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1320 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1321 unsigned long tl
= this_load
;
1322 unsigned long tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1325 * If sync wakeup then subtract the (maximum possible)
1326 * effect of the currently running task from the load
1327 * of the current CPU:
1330 tl
-= current
->load_weight
;
1333 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1334 100*(tl
+ p
->load_weight
) <= imbalance
*load
) {
1336 * This domain has SD_WAKE_AFFINE and
1337 * p is cache cold in this domain, and
1338 * there is no bad imbalance.
1340 schedstat_inc(this_sd
, ttwu_move_affine
);
1346 * Start passive balancing when half the imbalance_pct
1349 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1350 if (imbalance
*this_load
<= 100*load
) {
1351 schedstat_inc(this_sd
, ttwu_move_balance
);
1357 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1359 new_cpu
= wake_idle(new_cpu
, p
);
1360 if (new_cpu
!= cpu
) {
1361 set_task_cpu(p
, new_cpu
);
1362 task_rq_unlock(rq
, &flags
);
1363 /* might preempt at this point */
1364 rq
= task_rq_lock(p
, &flags
);
1365 old_state
= p
->state
;
1366 if (!(old_state
& state
))
1371 this_cpu
= smp_processor_id();
1376 #endif /* CONFIG_SMP */
1377 if (old_state
== TASK_UNINTERRUPTIBLE
) {
1378 rq
->nr_uninterruptible
--;
1380 * Tasks on involuntary sleep don't earn
1381 * sleep_avg beyond just interactive state.
1383 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
1387 * Tasks that have marked their sleep as noninteractive get
1388 * woken up with their sleep average not weighted in an
1391 if (old_state
& TASK_NONINTERACTIVE
)
1392 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
1395 activate_task(p
, rq
, cpu
== this_cpu
);
1397 * Sync wakeups (i.e. those types of wakeups where the waker
1398 * has indicated that it will leave the CPU in short order)
1399 * don't trigger a preemption, if the woken up task will run on
1400 * this cpu. (in this case the 'I will reschedule' promise of
1401 * the waker guarantees that the freshly woken up task is going
1402 * to be considered on this CPU.)
1404 if (!sync
|| cpu
!= this_cpu
) {
1405 if (TASK_PREEMPTS_CURR(p
, rq
))
1406 resched_task(rq
->curr
);
1411 p
->state
= TASK_RUNNING
;
1413 task_rq_unlock(rq
, &flags
);
1418 int fastcall
wake_up_process(task_t
*p
)
1420 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1421 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1424 EXPORT_SYMBOL(wake_up_process
);
1426 int fastcall
wake_up_state(task_t
*p
, unsigned int state
)
1428 return try_to_wake_up(p
, state
, 0);
1432 * Perform scheduler related setup for a newly forked process p.
1433 * p is forked by current.
1435 void fastcall
sched_fork(task_t
*p
, int clone_flags
)
1437 int cpu
= get_cpu();
1440 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1442 set_task_cpu(p
, cpu
);
1445 * We mark the process as running here, but have not actually
1446 * inserted it onto the runqueue yet. This guarantees that
1447 * nobody will actually run it, and a signal or other external
1448 * event cannot wake it up and insert it on the runqueue either.
1450 p
->state
= TASK_RUNNING
;
1451 INIT_LIST_HEAD(&p
->run_list
);
1453 #ifdef CONFIG_SCHEDSTATS
1454 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1456 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1459 #ifdef CONFIG_PREEMPT
1460 /* Want to start with kernel preemption disabled. */
1461 task_thread_info(p
)->preempt_count
= 1;
1464 * Share the timeslice between parent and child, thus the
1465 * total amount of pending timeslices in the system doesn't change,
1466 * resulting in more scheduling fairness.
1468 local_irq_disable();
1469 p
->time_slice
= (current
->time_slice
+ 1) >> 1;
1471 * The remainder of the first timeslice might be recovered by
1472 * the parent if the child exits early enough.
1474 p
->first_time_slice
= 1;
1475 current
->time_slice
>>= 1;
1476 p
->timestamp
= sched_clock();
1477 if (unlikely(!current
->time_slice
)) {
1479 * This case is rare, it happens when the parent has only
1480 * a single jiffy left from its timeslice. Taking the
1481 * runqueue lock is not a problem.
1483 current
->time_slice
= 1;
1491 * wake_up_new_task - wake up a newly created task for the first time.
1493 * This function will do some initial scheduler statistics housekeeping
1494 * that must be done for every newly created context, then puts the task
1495 * on the runqueue and wakes it.
1497 void fastcall
wake_up_new_task(task_t
*p
, unsigned long clone_flags
)
1499 unsigned long flags
;
1501 runqueue_t
*rq
, *this_rq
;
1503 rq
= task_rq_lock(p
, &flags
);
1504 BUG_ON(p
->state
!= TASK_RUNNING
);
1505 this_cpu
= smp_processor_id();
1509 * We decrease the sleep average of forking parents
1510 * and children as well, to keep max-interactive tasks
1511 * from forking tasks that are max-interactive. The parent
1512 * (current) is done further down, under its lock.
1514 p
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(p
) *
1515 CHILD_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1517 p
->prio
= effective_prio(p
);
1519 if (likely(cpu
== this_cpu
)) {
1520 if (!(clone_flags
& CLONE_VM
)) {
1522 * The VM isn't cloned, so we're in a good position to
1523 * do child-runs-first in anticipation of an exec. This
1524 * usually avoids a lot of COW overhead.
1526 if (unlikely(!current
->array
))
1527 __activate_task(p
, rq
);
1529 p
->prio
= current
->prio
;
1530 list_add_tail(&p
->run_list
, ¤t
->run_list
);
1531 p
->array
= current
->array
;
1532 p
->array
->nr_active
++;
1533 inc_nr_running(p
, rq
);
1537 /* Run child last */
1538 __activate_task(p
, rq
);
1540 * We skip the following code due to cpu == this_cpu
1542 * task_rq_unlock(rq, &flags);
1543 * this_rq = task_rq_lock(current, &flags);
1547 this_rq
= cpu_rq(this_cpu
);
1550 * Not the local CPU - must adjust timestamp. This should
1551 * get optimised away in the !CONFIG_SMP case.
1553 p
->timestamp
= (p
->timestamp
- this_rq
->timestamp_last_tick
)
1554 + rq
->timestamp_last_tick
;
1555 __activate_task(p
, rq
);
1556 if (TASK_PREEMPTS_CURR(p
, rq
))
1557 resched_task(rq
->curr
);
1560 * Parent and child are on different CPUs, now get the
1561 * parent runqueue to update the parent's ->sleep_avg:
1563 task_rq_unlock(rq
, &flags
);
1564 this_rq
= task_rq_lock(current
, &flags
);
1566 current
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(current
) *
1567 PARENT_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1568 task_rq_unlock(this_rq
, &flags
);
1572 * Potentially available exiting-child timeslices are
1573 * retrieved here - this way the parent does not get
1574 * penalized for creating too many threads.
1576 * (this cannot be used to 'generate' timeslices
1577 * artificially, because any timeslice recovered here
1578 * was given away by the parent in the first place.)
1580 void fastcall
sched_exit(task_t
*p
)
1582 unsigned long flags
;
1586 * If the child was a (relative-) CPU hog then decrease
1587 * the sleep_avg of the parent as well.
1589 rq
= task_rq_lock(p
->parent
, &flags
);
1590 if (p
->first_time_slice
&& task_cpu(p
) == task_cpu(p
->parent
)) {
1591 p
->parent
->time_slice
+= p
->time_slice
;
1592 if (unlikely(p
->parent
->time_slice
> task_timeslice(p
)))
1593 p
->parent
->time_slice
= task_timeslice(p
);
1595 if (p
->sleep_avg
< p
->parent
->sleep_avg
)
1596 p
->parent
->sleep_avg
= p
->parent
->sleep_avg
/
1597 (EXIT_WEIGHT
+ 1) * EXIT_WEIGHT
+ p
->sleep_avg
/
1599 task_rq_unlock(rq
, &flags
);
1603 * prepare_task_switch - prepare to switch tasks
1604 * @rq: the runqueue preparing to switch
1605 * @next: the task we are going to switch to.
1607 * This is called with the rq lock held and interrupts off. It must
1608 * be paired with a subsequent finish_task_switch after the context
1611 * prepare_task_switch sets up locking and calls architecture specific
1614 static inline void prepare_task_switch(runqueue_t
*rq
, task_t
*next
)
1616 prepare_lock_switch(rq
, next
);
1617 prepare_arch_switch(next
);
1621 * finish_task_switch - clean up after a task-switch
1622 * @rq: runqueue associated with task-switch
1623 * @prev: the thread we just switched away from.
1625 * finish_task_switch must be called after the context switch, paired
1626 * with a prepare_task_switch call before the context switch.
1627 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1628 * and do any other architecture-specific cleanup actions.
1630 * Note that we may have delayed dropping an mm in context_switch(). If
1631 * so, we finish that here outside of the runqueue lock. (Doing it
1632 * with the lock held can cause deadlocks; see schedule() for
1635 static inline void finish_task_switch(runqueue_t
*rq
, task_t
*prev
)
1636 __releases(rq
->lock
)
1638 struct mm_struct
*mm
= rq
->prev_mm
;
1639 unsigned long prev_task_flags
;
1644 * A task struct has one reference for the use as "current".
1645 * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1646 * calls schedule one last time. The schedule call will never return,
1647 * and the scheduled task must drop that reference.
1648 * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1649 * still held, otherwise prev could be scheduled on another cpu, die
1650 * there before we look at prev->state, and then the reference would
1652 * Manfred Spraul <manfred@colorfullife.com>
1654 prev_task_flags
= prev
->flags
;
1655 finish_arch_switch(prev
);
1656 finish_lock_switch(rq
, prev
);
1659 if (unlikely(prev_task_flags
& PF_DEAD
)) {
1661 * Remove function-return probe instances associated with this
1662 * task and put them back on the free list.
1664 kprobe_flush_task(prev
);
1665 put_task_struct(prev
);
1670 * schedule_tail - first thing a freshly forked thread must call.
1671 * @prev: the thread we just switched away from.
1673 asmlinkage
void schedule_tail(task_t
*prev
)
1674 __releases(rq
->lock
)
1676 runqueue_t
*rq
= this_rq();
1677 finish_task_switch(rq
, prev
);
1678 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1679 /* In this case, finish_task_switch does not reenable preemption */
1682 if (current
->set_child_tid
)
1683 put_user(current
->pid
, current
->set_child_tid
);
1687 * context_switch - switch to the new MM and the new
1688 * thread's register state.
1691 task_t
* context_switch(runqueue_t
*rq
, task_t
*prev
, task_t
*next
)
1693 struct mm_struct
*mm
= next
->mm
;
1694 struct mm_struct
*oldmm
= prev
->active_mm
;
1696 if (unlikely(!mm
)) {
1697 next
->active_mm
= oldmm
;
1698 atomic_inc(&oldmm
->mm_count
);
1699 enter_lazy_tlb(oldmm
, next
);
1701 switch_mm(oldmm
, mm
, next
);
1703 if (unlikely(!prev
->mm
)) {
1704 prev
->active_mm
= NULL
;
1705 WARN_ON(rq
->prev_mm
);
1706 rq
->prev_mm
= oldmm
;
1709 /* Here we just switch the register state and the stack. */
1710 switch_to(prev
, next
, prev
);
1716 * nr_running, nr_uninterruptible and nr_context_switches:
1718 * externally visible scheduler statistics: current number of runnable
1719 * threads, current number of uninterruptible-sleeping threads, total
1720 * number of context switches performed since bootup.
1722 unsigned long nr_running(void)
1724 unsigned long i
, sum
= 0;
1726 for_each_online_cpu(i
)
1727 sum
+= cpu_rq(i
)->nr_running
;
1732 unsigned long nr_uninterruptible(void)
1734 unsigned long i
, sum
= 0;
1736 for_each_possible_cpu(i
)
1737 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1740 * Since we read the counters lockless, it might be slightly
1741 * inaccurate. Do not allow it to go below zero though:
1743 if (unlikely((long)sum
< 0))
1749 unsigned long long nr_context_switches(void)
1752 unsigned long long sum
= 0;
1754 for_each_possible_cpu(i
)
1755 sum
+= cpu_rq(i
)->nr_switches
;
1760 unsigned long nr_iowait(void)
1762 unsigned long i
, sum
= 0;
1764 for_each_possible_cpu(i
)
1765 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1770 unsigned long nr_active(void)
1772 unsigned long i
, running
= 0, uninterruptible
= 0;
1774 for_each_online_cpu(i
) {
1775 running
+= cpu_rq(i
)->nr_running
;
1776 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
1779 if (unlikely((long)uninterruptible
< 0))
1780 uninterruptible
= 0;
1782 return running
+ uninterruptible
;
1788 * double_rq_lock - safely lock two runqueues
1790 * Note this does not disable interrupts like task_rq_lock,
1791 * you need to do so manually before calling.
1793 static void double_rq_lock(runqueue_t
*rq1
, runqueue_t
*rq2
)
1794 __acquires(rq1
->lock
)
1795 __acquires(rq2
->lock
)
1798 spin_lock(&rq1
->lock
);
1799 __acquire(rq2
->lock
); /* Fake it out ;) */
1802 spin_lock(&rq1
->lock
);
1803 spin_lock(&rq2
->lock
);
1805 spin_lock(&rq2
->lock
);
1806 spin_lock(&rq1
->lock
);
1812 * double_rq_unlock - safely unlock two runqueues
1814 * Note this does not restore interrupts like task_rq_unlock,
1815 * you need to do so manually after calling.
1817 static void double_rq_unlock(runqueue_t
*rq1
, runqueue_t
*rq2
)
1818 __releases(rq1
->lock
)
1819 __releases(rq2
->lock
)
1821 spin_unlock(&rq1
->lock
);
1823 spin_unlock(&rq2
->lock
);
1825 __release(rq2
->lock
);
1829 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1831 static void double_lock_balance(runqueue_t
*this_rq
, runqueue_t
*busiest
)
1832 __releases(this_rq
->lock
)
1833 __acquires(busiest
->lock
)
1834 __acquires(this_rq
->lock
)
1836 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1837 if (busiest
< this_rq
) {
1838 spin_unlock(&this_rq
->lock
);
1839 spin_lock(&busiest
->lock
);
1840 spin_lock(&this_rq
->lock
);
1842 spin_lock(&busiest
->lock
);
1847 * If dest_cpu is allowed for this process, migrate the task to it.
1848 * This is accomplished by forcing the cpu_allowed mask to only
1849 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
1850 * the cpu_allowed mask is restored.
1852 static void sched_migrate_task(task_t
*p
, int dest_cpu
)
1854 migration_req_t req
;
1856 unsigned long flags
;
1858 rq
= task_rq_lock(p
, &flags
);
1859 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
1860 || unlikely(cpu_is_offline(dest_cpu
)))
1863 /* force the process onto the specified CPU */
1864 if (migrate_task(p
, dest_cpu
, &req
)) {
1865 /* Need to wait for migration thread (might exit: take ref). */
1866 struct task_struct
*mt
= rq
->migration_thread
;
1867 get_task_struct(mt
);
1868 task_rq_unlock(rq
, &flags
);
1869 wake_up_process(mt
);
1870 put_task_struct(mt
);
1871 wait_for_completion(&req
.done
);
1875 task_rq_unlock(rq
, &flags
);
1879 * sched_exec - execve() is a valuable balancing opportunity, because at
1880 * this point the task has the smallest effective memory and cache footprint.
1882 void sched_exec(void)
1884 int new_cpu
, this_cpu
= get_cpu();
1885 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
1887 if (new_cpu
!= this_cpu
)
1888 sched_migrate_task(current
, new_cpu
);
1892 * pull_task - move a task from a remote runqueue to the local runqueue.
1893 * Both runqueues must be locked.
1896 void pull_task(runqueue_t
*src_rq
, prio_array_t
*src_array
, task_t
*p
,
1897 runqueue_t
*this_rq
, prio_array_t
*this_array
, int this_cpu
)
1899 dequeue_task(p
, src_array
);
1900 dec_nr_running(p
, src_rq
);
1901 set_task_cpu(p
, this_cpu
);
1902 inc_nr_running(p
, this_rq
);
1903 enqueue_task(p
, this_array
);
1904 p
->timestamp
= (p
->timestamp
- src_rq
->timestamp_last_tick
)
1905 + this_rq
->timestamp_last_tick
;
1907 * Note that idle threads have a prio of MAX_PRIO, for this test
1908 * to be always true for them.
1910 if (TASK_PREEMPTS_CURR(p
, this_rq
))
1911 resched_task(this_rq
->curr
);
1915 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1918 int can_migrate_task(task_t
*p
, runqueue_t
*rq
, int this_cpu
,
1919 struct sched_domain
*sd
, enum idle_type idle
,
1923 * We do not migrate tasks that are:
1924 * 1) running (obviously), or
1925 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1926 * 3) are cache-hot on their current CPU.
1928 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
1932 if (task_running(rq
, p
))
1936 * Aggressive migration if:
1937 * 1) task is cache cold, or
1938 * 2) too many balance attempts have failed.
1941 if (sd
->nr_balance_failed
> sd
->cache_nice_tries
)
1944 if (task_hot(p
, rq
->timestamp_last_tick
, sd
))
1949 #define rq_best_prio(rq) min((rq)->curr->prio, (rq)->best_expired_prio)
1951 * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted
1952 * load from busiest to this_rq, as part of a balancing operation within
1953 * "domain". Returns the number of tasks moved.
1955 * Called with both runqueues locked.
1957 static int move_tasks(runqueue_t
*this_rq
, int this_cpu
, runqueue_t
*busiest
,
1958 unsigned long max_nr_move
, unsigned long max_load_move
,
1959 struct sched_domain
*sd
, enum idle_type idle
,
1962 prio_array_t
*array
, *dst_array
;
1963 struct list_head
*head
, *curr
;
1964 int idx
, pulled
= 0, pinned
= 0, this_best_prio
, busiest_best_prio
;
1965 int busiest_best_prio_seen
;
1966 int skip_for_load
; /* skip the task based on weighted load issues */
1970 if (max_nr_move
== 0 || max_load_move
== 0)
1973 rem_load_move
= max_load_move
;
1975 this_best_prio
= rq_best_prio(this_rq
);
1976 busiest_best_prio
= rq_best_prio(busiest
);
1978 * Enable handling of the case where there is more than one task
1979 * with the best priority. If the current running task is one
1980 * of those with prio==busiest_best_prio we know it won't be moved
1981 * and therefore it's safe to override the skip (based on load) of
1982 * any task we find with that prio.
1984 busiest_best_prio_seen
= busiest_best_prio
== busiest
->curr
->prio
;
1987 * We first consider expired tasks. Those will likely not be
1988 * executed in the near future, and they are most likely to
1989 * be cache-cold, thus switching CPUs has the least effect
1992 if (busiest
->expired
->nr_active
) {
1993 array
= busiest
->expired
;
1994 dst_array
= this_rq
->expired
;
1996 array
= busiest
->active
;
1997 dst_array
= this_rq
->active
;
2001 /* Start searching at priority 0: */
2005 idx
= sched_find_first_bit(array
->bitmap
);
2007 idx
= find_next_bit(array
->bitmap
, MAX_PRIO
, idx
);
2008 if (idx
>= MAX_PRIO
) {
2009 if (array
== busiest
->expired
&& busiest
->active
->nr_active
) {
2010 array
= busiest
->active
;
2011 dst_array
= this_rq
->active
;
2017 head
= array
->queue
+ idx
;
2020 tmp
= list_entry(curr
, task_t
, run_list
);
2025 * To help distribute high priority tasks accross CPUs we don't
2026 * skip a task if it will be the highest priority task (i.e. smallest
2027 * prio value) on its new queue regardless of its load weight
2029 skip_for_load
= tmp
->load_weight
> rem_load_move
;
2030 if (skip_for_load
&& idx
< this_best_prio
)
2031 skip_for_load
= !busiest_best_prio_seen
&& idx
== busiest_best_prio
;
2032 if (skip_for_load
||
2033 !can_migrate_task(tmp
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2034 busiest_best_prio_seen
|= idx
== busiest_best_prio
;
2041 #ifdef CONFIG_SCHEDSTATS
2042 if (task_hot(tmp
, busiest
->timestamp_last_tick
, sd
))
2043 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2046 pull_task(busiest
, array
, tmp
, this_rq
, dst_array
, this_cpu
);
2048 rem_load_move
-= tmp
->load_weight
;
2051 * We only want to steal up to the prescribed number of tasks
2052 * and the prescribed amount of weighted load.
2054 if (pulled
< max_nr_move
&& rem_load_move
> 0) {
2055 if (idx
< this_best_prio
)
2056 this_best_prio
= idx
;
2064 * Right now, this is the only place pull_task() is called,
2065 * so we can safely collect pull_task() stats here rather than
2066 * inside pull_task().
2068 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2071 *all_pinned
= pinned
;
2076 * find_busiest_group finds and returns the busiest CPU group within the
2077 * domain. It calculates and returns the amount of weighted load which should be
2078 * moved to restore balance via the imbalance parameter.
2080 static struct sched_group
*
2081 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2082 unsigned long *imbalance
, enum idle_type idle
, int *sd_idle
)
2084 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2085 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2086 unsigned long max_pull
;
2087 unsigned long busiest_load_per_task
, busiest_nr_running
;
2088 unsigned long this_load_per_task
, this_nr_running
;
2090 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2091 int power_savings_balance
= 1;
2092 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2093 unsigned long min_nr_running
= ULONG_MAX
;
2094 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2097 max_load
= this_load
= total_load
= total_pwr
= 0;
2098 busiest_load_per_task
= busiest_nr_running
= 0;
2099 this_load_per_task
= this_nr_running
= 0;
2100 if (idle
== NOT_IDLE
)
2101 load_idx
= sd
->busy_idx
;
2102 else if (idle
== NEWLY_IDLE
)
2103 load_idx
= sd
->newidle_idx
;
2105 load_idx
= sd
->idle_idx
;
2108 unsigned long load
, group_capacity
;
2111 unsigned long sum_nr_running
, sum_weighted_load
;
2113 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2115 /* Tally up the load of all CPUs in the group */
2116 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2118 for_each_cpu_mask(i
, group
->cpumask
) {
2119 runqueue_t
*rq
= cpu_rq(i
);
2121 if (*sd_idle
&& !idle_cpu(i
))
2124 /* Bias balancing toward cpus of our domain */
2126 load
= target_load(i
, load_idx
);
2128 load
= source_load(i
, load_idx
);
2131 sum_nr_running
+= rq
->nr_running
;
2132 sum_weighted_load
+= rq
->raw_weighted_load
;
2135 total_load
+= avg_load
;
2136 total_pwr
+= group
->cpu_power
;
2138 /* Adjust by relative CPU power of the group */
2139 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
2141 group_capacity
= group
->cpu_power
/ SCHED_LOAD_SCALE
;
2144 this_load
= avg_load
;
2146 this_nr_running
= sum_nr_running
;
2147 this_load_per_task
= sum_weighted_load
;
2148 } else if (avg_load
> max_load
&&
2149 sum_nr_running
> group_capacity
) {
2150 max_load
= avg_load
;
2152 busiest_nr_running
= sum_nr_running
;
2153 busiest_load_per_task
= sum_weighted_load
;
2156 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2158 * Busy processors will not participate in power savings
2161 if (idle
== NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2165 * If the local group is idle or completely loaded
2166 * no need to do power savings balance at this domain
2168 if (local_group
&& (this_nr_running
>= group_capacity
||
2170 power_savings_balance
= 0;
2173 * If a group is already running at full capacity or idle,
2174 * don't include that group in power savings calculations
2176 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2181 * Calculate the group which has the least non-idle load.
2182 * This is the group from where we need to pick up the load
2185 if ((sum_nr_running
< min_nr_running
) ||
2186 (sum_nr_running
== min_nr_running
&&
2187 first_cpu(group
->cpumask
) <
2188 first_cpu(group_min
->cpumask
))) {
2190 min_nr_running
= sum_nr_running
;
2191 min_load_per_task
= sum_weighted_load
/
2196 * Calculate the group which is almost near its
2197 * capacity but still has some space to pick up some load
2198 * from other group and save more power
2200 if (sum_nr_running
<= group_capacity
- 1)
2201 if (sum_nr_running
> leader_nr_running
||
2202 (sum_nr_running
== leader_nr_running
&&
2203 first_cpu(group
->cpumask
) >
2204 first_cpu(group_leader
->cpumask
))) {
2205 group_leader
= group
;
2206 leader_nr_running
= sum_nr_running
;
2211 group
= group
->next
;
2212 } while (group
!= sd
->groups
);
2214 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2217 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2219 if (this_load
>= avg_load
||
2220 100*max_load
<= sd
->imbalance_pct
*this_load
)
2223 busiest_load_per_task
/= busiest_nr_running
;
2225 * We're trying to get all the cpus to the average_load, so we don't
2226 * want to push ourselves above the average load, nor do we wish to
2227 * reduce the max loaded cpu below the average load, as either of these
2228 * actions would just result in more rebalancing later, and ping-pong
2229 * tasks around. Thus we look for the minimum possible imbalance.
2230 * Negative imbalances (*we* are more loaded than anyone else) will
2231 * be counted as no imbalance for these purposes -- we can't fix that
2232 * by pulling tasks to us. Be careful of negative numbers as they'll
2233 * appear as very large values with unsigned longs.
2235 if (max_load
<= busiest_load_per_task
)
2239 * In the presence of smp nice balancing, certain scenarios can have
2240 * max load less than avg load(as we skip the groups at or below
2241 * its cpu_power, while calculating max_load..)
2243 if (max_load
< avg_load
) {
2245 goto small_imbalance
;
2248 /* Don't want to pull so many tasks that a group would go idle */
2249 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2251 /* How much load to actually move to equalise the imbalance */
2252 *imbalance
= min(max_pull
* busiest
->cpu_power
,
2253 (avg_load
- this_load
) * this->cpu_power
)
2257 * if *imbalance is less than the average load per runnable task
2258 * there is no gaurantee that any tasks will be moved so we'll have
2259 * a think about bumping its value to force at least one task to be
2262 if (*imbalance
< busiest_load_per_task
) {
2263 unsigned long pwr_now
, pwr_move
;
2268 pwr_move
= pwr_now
= 0;
2270 if (this_nr_running
) {
2271 this_load_per_task
/= this_nr_running
;
2272 if (busiest_load_per_task
> this_load_per_task
)
2275 this_load_per_task
= SCHED_LOAD_SCALE
;
2277 if (max_load
- this_load
>= busiest_load_per_task
* imbn
) {
2278 *imbalance
= busiest_load_per_task
;
2283 * OK, we don't have enough imbalance to justify moving tasks,
2284 * however we may be able to increase total CPU power used by
2288 pwr_now
+= busiest
->cpu_power
*
2289 min(busiest_load_per_task
, max_load
);
2290 pwr_now
+= this->cpu_power
*
2291 min(this_load_per_task
, this_load
);
2292 pwr_now
/= SCHED_LOAD_SCALE
;
2294 /* Amount of load we'd subtract */
2295 tmp
= busiest_load_per_task
*SCHED_LOAD_SCALE
/busiest
->cpu_power
;
2297 pwr_move
+= busiest
->cpu_power
*
2298 min(busiest_load_per_task
, max_load
- tmp
);
2300 /* Amount of load we'd add */
2301 if (max_load
*busiest
->cpu_power
<
2302 busiest_load_per_task
*SCHED_LOAD_SCALE
)
2303 tmp
= max_load
*busiest
->cpu_power
/this->cpu_power
;
2305 tmp
= busiest_load_per_task
*SCHED_LOAD_SCALE
/this->cpu_power
;
2306 pwr_move
+= this->cpu_power
*min(this_load_per_task
, this_load
+ tmp
);
2307 pwr_move
/= SCHED_LOAD_SCALE
;
2309 /* Move if we gain throughput */
2310 if (pwr_move
<= pwr_now
)
2313 *imbalance
= busiest_load_per_task
;
2319 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2320 if (idle
== NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2323 if (this == group_leader
&& group_leader
!= group_min
) {
2324 *imbalance
= min_load_per_task
;
2334 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2336 static runqueue_t
*find_busiest_queue(struct sched_group
*group
,
2337 enum idle_type idle
, unsigned long imbalance
)
2339 unsigned long max_load
= 0;
2340 runqueue_t
*busiest
= NULL
, *rqi
;
2343 for_each_cpu_mask(i
, group
->cpumask
) {
2346 if (rqi
->nr_running
== 1 && rqi
->raw_weighted_load
> imbalance
)
2349 if (rqi
->raw_weighted_load
> max_load
) {
2350 max_load
= rqi
->raw_weighted_load
;
2359 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2360 * so long as it is large enough.
2362 #define MAX_PINNED_INTERVAL 512
2364 #define minus_1_or_zero(n) ((n) > 0 ? (n) - 1 : 0)
2366 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2367 * tasks if there is an imbalance.
2369 * Called with this_rq unlocked.
2371 static int load_balance(int this_cpu
, runqueue_t
*this_rq
,
2372 struct sched_domain
*sd
, enum idle_type idle
)
2374 struct sched_group
*group
;
2375 runqueue_t
*busiest
;
2376 unsigned long imbalance
;
2377 int nr_moved
, all_pinned
= 0;
2378 int active_balance
= 0;
2381 if (idle
!= NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2382 !sched_smt_power_savings
)
2385 schedstat_inc(sd
, lb_cnt
[idle
]);
2387 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
);
2389 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2393 busiest
= find_busiest_queue(group
, idle
, imbalance
);
2395 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2399 BUG_ON(busiest
== this_rq
);
2401 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2404 if (busiest
->nr_running
> 1) {
2406 * Attempt to move tasks. If find_busiest_group has found
2407 * an imbalance but busiest->nr_running <= 1, the group is
2408 * still unbalanced. nr_moved simply stays zero, so it is
2409 * correctly treated as an imbalance.
2411 double_rq_lock(this_rq
, busiest
);
2412 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2413 minus_1_or_zero(busiest
->nr_running
),
2414 imbalance
, sd
, idle
, &all_pinned
);
2415 double_rq_unlock(this_rq
, busiest
);
2417 /* All tasks on this runqueue were pinned by CPU affinity */
2418 if (unlikely(all_pinned
))
2423 schedstat_inc(sd
, lb_failed
[idle
]);
2424 sd
->nr_balance_failed
++;
2426 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2428 spin_lock(&busiest
->lock
);
2430 /* don't kick the migration_thread, if the curr
2431 * task on busiest cpu can't be moved to this_cpu
2433 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2434 spin_unlock(&busiest
->lock
);
2436 goto out_one_pinned
;
2439 if (!busiest
->active_balance
) {
2440 busiest
->active_balance
= 1;
2441 busiest
->push_cpu
= this_cpu
;
2444 spin_unlock(&busiest
->lock
);
2446 wake_up_process(busiest
->migration_thread
);
2449 * We've kicked active balancing, reset the failure
2452 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2455 sd
->nr_balance_failed
= 0;
2457 if (likely(!active_balance
)) {
2458 /* We were unbalanced, so reset the balancing interval */
2459 sd
->balance_interval
= sd
->min_interval
;
2462 * If we've begun active balancing, start to back off. This
2463 * case may not be covered by the all_pinned logic if there
2464 * is only 1 task on the busy runqueue (because we don't call
2467 if (sd
->balance_interval
< sd
->max_interval
)
2468 sd
->balance_interval
*= 2;
2471 if (!nr_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2472 !sched_smt_power_savings
)
2477 schedstat_inc(sd
, lb_balanced
[idle
]);
2479 sd
->nr_balance_failed
= 0;
2482 /* tune up the balancing interval */
2483 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2484 (sd
->balance_interval
< sd
->max_interval
))
2485 sd
->balance_interval
*= 2;
2487 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&& !sched_smt_power_savings
)
2493 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2494 * tasks if there is an imbalance.
2496 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2497 * this_rq is locked.
2499 static int load_balance_newidle(int this_cpu
, runqueue_t
*this_rq
,
2500 struct sched_domain
*sd
)
2502 struct sched_group
*group
;
2503 runqueue_t
*busiest
= NULL
;
2504 unsigned long imbalance
;
2508 if (sd
->flags
& SD_SHARE_CPUPOWER
&& !sched_smt_power_savings
)
2511 schedstat_inc(sd
, lb_cnt
[NEWLY_IDLE
]);
2512 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, NEWLY_IDLE
, &sd_idle
);
2514 schedstat_inc(sd
, lb_nobusyg
[NEWLY_IDLE
]);
2518 busiest
= find_busiest_queue(group
, NEWLY_IDLE
, imbalance
);
2520 schedstat_inc(sd
, lb_nobusyq
[NEWLY_IDLE
]);
2524 BUG_ON(busiest
== this_rq
);
2526 schedstat_add(sd
, lb_imbalance
[NEWLY_IDLE
], imbalance
);
2529 if (busiest
->nr_running
> 1) {
2530 /* Attempt to move tasks */
2531 double_lock_balance(this_rq
, busiest
);
2532 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2533 minus_1_or_zero(busiest
->nr_running
),
2534 imbalance
, sd
, NEWLY_IDLE
, NULL
);
2535 spin_unlock(&busiest
->lock
);
2539 schedstat_inc(sd
, lb_failed
[NEWLY_IDLE
]);
2540 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2543 sd
->nr_balance_failed
= 0;
2548 schedstat_inc(sd
, lb_balanced
[NEWLY_IDLE
]);
2549 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&& !sched_smt_power_savings
)
2551 sd
->nr_balance_failed
= 0;
2556 * idle_balance is called by schedule() if this_cpu is about to become
2557 * idle. Attempts to pull tasks from other CPUs.
2559 static void idle_balance(int this_cpu
, runqueue_t
*this_rq
)
2561 struct sched_domain
*sd
;
2563 for_each_domain(this_cpu
, sd
) {
2564 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
2565 if (load_balance_newidle(this_cpu
, this_rq
, sd
)) {
2566 /* We've pulled tasks over so stop searching */
2574 * active_load_balance is run by migration threads. It pushes running tasks
2575 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2576 * running on each physical CPU where possible, and avoids physical /
2577 * logical imbalances.
2579 * Called with busiest_rq locked.
2581 static void active_load_balance(runqueue_t
*busiest_rq
, int busiest_cpu
)
2583 struct sched_domain
*sd
;
2584 runqueue_t
*target_rq
;
2585 int target_cpu
= busiest_rq
->push_cpu
;
2587 if (busiest_rq
->nr_running
<= 1)
2588 /* no task to move */
2591 target_rq
= cpu_rq(target_cpu
);
2594 * This condition is "impossible", if it occurs
2595 * we need to fix it. Originally reported by
2596 * Bjorn Helgaas on a 128-cpu setup.
2598 BUG_ON(busiest_rq
== target_rq
);
2600 /* move a task from busiest_rq to target_rq */
2601 double_lock_balance(busiest_rq
, target_rq
);
2603 /* Search for an sd spanning us and the target CPU. */
2604 for_each_domain(target_cpu
, sd
) {
2605 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2606 cpu_isset(busiest_cpu
, sd
->span
))
2610 if (unlikely(sd
== NULL
))
2613 schedstat_inc(sd
, alb_cnt
);
2615 if (move_tasks(target_rq
, target_cpu
, busiest_rq
, 1,
2616 RTPRIO_TO_LOAD_WEIGHT(100), sd
, SCHED_IDLE
, NULL
))
2617 schedstat_inc(sd
, alb_pushed
);
2619 schedstat_inc(sd
, alb_failed
);
2621 spin_unlock(&target_rq
->lock
);
2625 * rebalance_tick will get called every timer tick, on every CPU.
2627 * It checks each scheduling domain to see if it is due to be balanced,
2628 * and initiates a balancing operation if so.
2630 * Balancing parameters are set up in arch_init_sched_domains.
2633 /* Don't have all balancing operations going off at once */
2634 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2636 static void rebalance_tick(int this_cpu
, runqueue_t
*this_rq
,
2637 enum idle_type idle
)
2639 unsigned long old_load
, this_load
;
2640 unsigned long j
= jiffies
+ CPU_OFFSET(this_cpu
);
2641 struct sched_domain
*sd
;
2644 this_load
= this_rq
->raw_weighted_load
;
2645 /* Update our load */
2646 for (i
= 0; i
< 3; i
++) {
2647 unsigned long new_load
= this_load
;
2649 old_load
= this_rq
->cpu_load
[i
];
2651 * Round up the averaging division if load is increasing. This
2652 * prevents us from getting stuck on 9 if the load is 10, for
2655 if (new_load
> old_load
)
2656 new_load
+= scale
-1;
2657 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) / scale
;
2660 for_each_domain(this_cpu
, sd
) {
2661 unsigned long interval
;
2663 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2666 interval
= sd
->balance_interval
;
2667 if (idle
!= SCHED_IDLE
)
2668 interval
*= sd
->busy_factor
;
2670 /* scale ms to jiffies */
2671 interval
= msecs_to_jiffies(interval
);
2672 if (unlikely(!interval
))
2675 if (j
- sd
->last_balance
>= interval
) {
2676 if (load_balance(this_cpu
, this_rq
, sd
, idle
)) {
2678 * We've pulled tasks over so either we're no
2679 * longer idle, or one of our SMT siblings is
2684 sd
->last_balance
+= interval
;
2690 * on UP we do not need to balance between CPUs:
2692 static inline void rebalance_tick(int cpu
, runqueue_t
*rq
, enum idle_type idle
)
2695 static inline void idle_balance(int cpu
, runqueue_t
*rq
)
2700 static inline int wake_priority_sleeper(runqueue_t
*rq
)
2703 #ifdef CONFIG_SCHED_SMT
2704 spin_lock(&rq
->lock
);
2706 * If an SMT sibling task has been put to sleep for priority
2707 * reasons reschedule the idle task to see if it can now run.
2709 if (rq
->nr_running
) {
2710 resched_task(rq
->idle
);
2713 spin_unlock(&rq
->lock
);
2718 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2720 EXPORT_PER_CPU_SYMBOL(kstat
);
2723 * This is called on clock ticks and on context switches.
2724 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2726 static inline void update_cpu_clock(task_t
*p
, runqueue_t
*rq
,
2727 unsigned long long now
)
2729 unsigned long long last
= max(p
->timestamp
, rq
->timestamp_last_tick
);
2730 p
->sched_time
+= now
- last
;
2734 * Return current->sched_time plus any more ns on the sched_clock
2735 * that have not yet been banked.
2737 unsigned long long current_sched_time(const task_t
*tsk
)
2739 unsigned long long ns
;
2740 unsigned long flags
;
2741 local_irq_save(flags
);
2742 ns
= max(tsk
->timestamp
, task_rq(tsk
)->timestamp_last_tick
);
2743 ns
= tsk
->sched_time
+ (sched_clock() - ns
);
2744 local_irq_restore(flags
);
2749 * We place interactive tasks back into the active array, if possible.
2751 * To guarantee that this does not starve expired tasks we ignore the
2752 * interactivity of a task if the first expired task had to wait more
2753 * than a 'reasonable' amount of time. This deadline timeout is
2754 * load-dependent, as the frequency of array switched decreases with
2755 * increasing number of running tasks. We also ignore the interactivity
2756 * if a better static_prio task has expired:
2758 #define EXPIRED_STARVING(rq) \
2759 ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2760 (jiffies - (rq)->expired_timestamp >= \
2761 STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2762 ((rq)->curr->static_prio > (rq)->best_expired_prio))
2765 * Account user cpu time to a process.
2766 * @p: the process that the cpu time gets accounted to
2767 * @hardirq_offset: the offset to subtract from hardirq_count()
2768 * @cputime: the cpu time spent in user space since the last update
2770 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
2772 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2775 p
->utime
= cputime_add(p
->utime
, cputime
);
2777 /* Add user time to cpustat. */
2778 tmp
= cputime_to_cputime64(cputime
);
2779 if (TASK_NICE(p
) > 0)
2780 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
2782 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
2786 * Account system cpu time to a process.
2787 * @p: the process that the cpu time gets accounted to
2788 * @hardirq_offset: the offset to subtract from hardirq_count()
2789 * @cputime: the cpu time spent in kernel space since the last update
2791 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
2794 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2795 runqueue_t
*rq
= this_rq();
2798 p
->stime
= cputime_add(p
->stime
, cputime
);
2800 /* Add system time to cpustat. */
2801 tmp
= cputime_to_cputime64(cputime
);
2802 if (hardirq_count() - hardirq_offset
)
2803 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
2804 else if (softirq_count())
2805 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
2806 else if (p
!= rq
->idle
)
2807 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
2808 else if (atomic_read(&rq
->nr_iowait
) > 0)
2809 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
2811 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
2812 /* Account for system time used */
2813 acct_update_integrals(p
);
2817 * Account for involuntary wait time.
2818 * @p: the process from which the cpu time has been stolen
2819 * @steal: the cpu time spent in involuntary wait
2821 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
2823 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2824 cputime64_t tmp
= cputime_to_cputime64(steal
);
2825 runqueue_t
*rq
= this_rq();
2827 if (p
== rq
->idle
) {
2828 p
->stime
= cputime_add(p
->stime
, steal
);
2829 if (atomic_read(&rq
->nr_iowait
) > 0)
2830 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
2832 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
2834 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
2838 * This function gets called by the timer code, with HZ frequency.
2839 * We call it with interrupts disabled.
2841 * It also gets called by the fork code, when changing the parent's
2844 void scheduler_tick(void)
2846 int cpu
= smp_processor_id();
2847 runqueue_t
*rq
= this_rq();
2848 task_t
*p
= current
;
2849 unsigned long long now
= sched_clock();
2851 update_cpu_clock(p
, rq
, now
);
2853 rq
->timestamp_last_tick
= now
;
2855 if (p
== rq
->idle
) {
2856 if (wake_priority_sleeper(rq
))
2858 rebalance_tick(cpu
, rq
, SCHED_IDLE
);
2862 /* Task might have expired already, but not scheduled off yet */
2863 if (p
->array
!= rq
->active
) {
2864 set_tsk_need_resched(p
);
2867 spin_lock(&rq
->lock
);
2869 * The task was running during this tick - update the
2870 * time slice counter. Note: we do not update a thread's
2871 * priority until it either goes to sleep or uses up its
2872 * timeslice. This makes it possible for interactive tasks
2873 * to use up their timeslices at their highest priority levels.
2877 * RR tasks need a special form of timeslice management.
2878 * FIFO tasks have no timeslices.
2880 if ((p
->policy
== SCHED_RR
) && !--p
->time_slice
) {
2881 p
->time_slice
= task_timeslice(p
);
2882 p
->first_time_slice
= 0;
2883 set_tsk_need_resched(p
);
2885 /* put it at the end of the queue: */
2886 requeue_task(p
, rq
->active
);
2890 if (!--p
->time_slice
) {
2891 dequeue_task(p
, rq
->active
);
2892 set_tsk_need_resched(p
);
2893 p
->prio
= effective_prio(p
);
2894 p
->time_slice
= task_timeslice(p
);
2895 p
->first_time_slice
= 0;
2897 if (!rq
->expired_timestamp
)
2898 rq
->expired_timestamp
= jiffies
;
2899 if (!TASK_INTERACTIVE(p
) || EXPIRED_STARVING(rq
)) {
2900 enqueue_task(p
, rq
->expired
);
2901 if (p
->static_prio
< rq
->best_expired_prio
)
2902 rq
->best_expired_prio
= p
->static_prio
;
2904 enqueue_task(p
, rq
->active
);
2907 * Prevent a too long timeslice allowing a task to monopolize
2908 * the CPU. We do this by splitting up the timeslice into
2911 * Note: this does not mean the task's timeslices expire or
2912 * get lost in any way, they just might be preempted by
2913 * another task of equal priority. (one with higher
2914 * priority would have preempted this task already.) We
2915 * requeue this task to the end of the list on this priority
2916 * level, which is in essence a round-robin of tasks with
2919 * This only applies to tasks in the interactive
2920 * delta range with at least TIMESLICE_GRANULARITY to requeue.
2922 if (TASK_INTERACTIVE(p
) && !((task_timeslice(p
) -
2923 p
->time_slice
) % TIMESLICE_GRANULARITY(p
)) &&
2924 (p
->time_slice
>= TIMESLICE_GRANULARITY(p
)) &&
2925 (p
->array
== rq
->active
)) {
2927 requeue_task(p
, rq
->active
);
2928 set_tsk_need_resched(p
);
2932 spin_unlock(&rq
->lock
);
2934 rebalance_tick(cpu
, rq
, NOT_IDLE
);
2937 #ifdef CONFIG_SCHED_SMT
2938 static inline void wakeup_busy_runqueue(runqueue_t
*rq
)
2940 /* If an SMT runqueue is sleeping due to priority reasons wake it up */
2941 if (rq
->curr
== rq
->idle
&& rq
->nr_running
)
2942 resched_task(rq
->idle
);
2946 * Called with interrupt disabled and this_rq's runqueue locked.
2948 static void wake_sleeping_dependent(int this_cpu
)
2950 struct sched_domain
*tmp
, *sd
= NULL
;
2953 for_each_domain(this_cpu
, tmp
) {
2954 if (tmp
->flags
& SD_SHARE_CPUPOWER
) {
2963 for_each_cpu_mask(i
, sd
->span
) {
2964 runqueue_t
*smt_rq
= cpu_rq(i
);
2968 if (unlikely(!spin_trylock(&smt_rq
->lock
)))
2971 wakeup_busy_runqueue(smt_rq
);
2972 spin_unlock(&smt_rq
->lock
);
2977 * number of 'lost' timeslices this task wont be able to fully
2978 * utilize, if another task runs on a sibling. This models the
2979 * slowdown effect of other tasks running on siblings:
2981 static inline unsigned long smt_slice(task_t
*p
, struct sched_domain
*sd
)
2983 return p
->time_slice
* (100 - sd
->per_cpu_gain
) / 100;
2987 * To minimise lock contention and not have to drop this_rq's runlock we only
2988 * trylock the sibling runqueues and bypass those runqueues if we fail to
2989 * acquire their lock. As we only trylock the normal locking order does not
2990 * need to be obeyed.
2992 static int dependent_sleeper(int this_cpu
, runqueue_t
*this_rq
, task_t
*p
)
2994 struct sched_domain
*tmp
, *sd
= NULL
;
2997 /* kernel/rt threads do not participate in dependent sleeping */
2998 if (!p
->mm
|| rt_task(p
))
3001 for_each_domain(this_cpu
, tmp
) {
3002 if (tmp
->flags
& SD_SHARE_CPUPOWER
) {
3011 for_each_cpu_mask(i
, sd
->span
) {
3019 if (unlikely(!spin_trylock(&smt_rq
->lock
)))
3022 smt_curr
= smt_rq
->curr
;
3028 * If a user task with lower static priority than the
3029 * running task on the SMT sibling is trying to schedule,
3030 * delay it till there is proportionately less timeslice
3031 * left of the sibling task to prevent a lower priority
3032 * task from using an unfair proportion of the
3033 * physical cpu's resources. -ck
3035 if (rt_task(smt_curr
)) {
3037 * With real time tasks we run non-rt tasks only
3038 * per_cpu_gain% of the time.
3040 if ((jiffies
% DEF_TIMESLICE
) >
3041 (sd
->per_cpu_gain
* DEF_TIMESLICE
/ 100))
3044 if (smt_curr
->static_prio
< p
->static_prio
&&
3045 !TASK_PREEMPTS_CURR(p
, smt_rq
) &&
3046 smt_slice(smt_curr
, sd
) > task_timeslice(p
))
3050 spin_unlock(&smt_rq
->lock
);
3055 static inline void wake_sleeping_dependent(int this_cpu
)
3059 static inline int dependent_sleeper(int this_cpu
, runqueue_t
*this_rq
,
3066 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3068 void fastcall
add_preempt_count(int val
)
3073 BUG_ON((preempt_count() < 0));
3074 preempt_count() += val
;
3076 * Spinlock count overflowing soon?
3078 BUG_ON((preempt_count() & PREEMPT_MASK
) >= PREEMPT_MASK
-10);
3080 EXPORT_SYMBOL(add_preempt_count
);
3082 void fastcall
sub_preempt_count(int val
)
3087 BUG_ON(val
> preempt_count());
3089 * Is the spinlock portion underflowing?
3091 BUG_ON((val
< PREEMPT_MASK
) && !(preempt_count() & PREEMPT_MASK
));
3092 preempt_count() -= val
;
3094 EXPORT_SYMBOL(sub_preempt_count
);
3098 static inline int interactive_sleep(enum sleep_type sleep_type
)
3100 return (sleep_type
== SLEEP_INTERACTIVE
||
3101 sleep_type
== SLEEP_INTERRUPTED
);
3105 * schedule() is the main scheduler function.
3107 asmlinkage
void __sched
schedule(void)
3110 task_t
*prev
, *next
;
3112 prio_array_t
*array
;
3113 struct list_head
*queue
;
3114 unsigned long long now
;
3115 unsigned long run_time
;
3116 int cpu
, idx
, new_prio
;
3119 * Test if we are atomic. Since do_exit() needs to call into
3120 * schedule() atomically, we ignore that path for now.
3121 * Otherwise, whine if we are scheduling when we should not be.
3123 if (unlikely(in_atomic() && !current
->exit_state
)) {
3124 printk(KERN_ERR
"BUG: scheduling while atomic: "
3126 current
->comm
, preempt_count(), current
->pid
);
3129 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3134 release_kernel_lock(prev
);
3135 need_resched_nonpreemptible
:
3139 * The idle thread is not allowed to schedule!
3140 * Remove this check after it has been exercised a bit.
3142 if (unlikely(prev
== rq
->idle
) && prev
->state
!= TASK_RUNNING
) {
3143 printk(KERN_ERR
"bad: scheduling from the idle thread!\n");
3147 schedstat_inc(rq
, sched_cnt
);
3148 now
= sched_clock();
3149 if (likely((long long)(now
- prev
->timestamp
) < NS_MAX_SLEEP_AVG
)) {
3150 run_time
= now
- prev
->timestamp
;
3151 if (unlikely((long long)(now
- prev
->timestamp
) < 0))
3154 run_time
= NS_MAX_SLEEP_AVG
;
3157 * Tasks charged proportionately less run_time at high sleep_avg to
3158 * delay them losing their interactive status
3160 run_time
/= (CURRENT_BONUS(prev
) ? : 1);
3162 spin_lock_irq(&rq
->lock
);
3164 if (unlikely(prev
->flags
& PF_DEAD
))
3165 prev
->state
= EXIT_DEAD
;
3167 switch_count
= &prev
->nivcsw
;
3168 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3169 switch_count
= &prev
->nvcsw
;
3170 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3171 unlikely(signal_pending(prev
))))
3172 prev
->state
= TASK_RUNNING
;
3174 if (prev
->state
== TASK_UNINTERRUPTIBLE
)
3175 rq
->nr_uninterruptible
++;
3176 deactivate_task(prev
, rq
);
3180 cpu
= smp_processor_id();
3181 if (unlikely(!rq
->nr_running
)) {
3182 idle_balance(cpu
, rq
);
3183 if (!rq
->nr_running
) {
3185 rq
->expired_timestamp
= 0;
3186 wake_sleeping_dependent(cpu
);
3192 if (unlikely(!array
->nr_active
)) {
3194 * Switch the active and expired arrays.
3196 schedstat_inc(rq
, sched_switch
);
3197 rq
->active
= rq
->expired
;
3198 rq
->expired
= array
;
3200 rq
->expired_timestamp
= 0;
3201 rq
->best_expired_prio
= MAX_PRIO
;
3204 idx
= sched_find_first_bit(array
->bitmap
);
3205 queue
= array
->queue
+ idx
;
3206 next
= list_entry(queue
->next
, task_t
, run_list
);
3208 if (!rt_task(next
) && interactive_sleep(next
->sleep_type
)) {
3209 unsigned long long delta
= now
- next
->timestamp
;
3210 if (unlikely((long long)(now
- next
->timestamp
) < 0))
3213 if (next
->sleep_type
== SLEEP_INTERACTIVE
)
3214 delta
= delta
* (ON_RUNQUEUE_WEIGHT
* 128 / 100) / 128;
3216 array
= next
->array
;
3217 new_prio
= recalc_task_prio(next
, next
->timestamp
+ delta
);
3219 if (unlikely(next
->prio
!= new_prio
)) {
3220 dequeue_task(next
, array
);
3221 next
->prio
= new_prio
;
3222 enqueue_task(next
, array
);
3225 next
->sleep_type
= SLEEP_NORMAL
;
3226 if (dependent_sleeper(cpu
, rq
, next
))
3229 if (next
== rq
->idle
)
3230 schedstat_inc(rq
, sched_goidle
);
3232 prefetch_stack(next
);
3233 clear_tsk_need_resched(prev
);
3234 rcu_qsctr_inc(task_cpu(prev
));
3236 update_cpu_clock(prev
, rq
, now
);
3238 prev
->sleep_avg
-= run_time
;
3239 if ((long)prev
->sleep_avg
<= 0)
3240 prev
->sleep_avg
= 0;
3241 prev
->timestamp
= prev
->last_ran
= now
;
3243 sched_info_switch(prev
, next
);
3244 if (likely(prev
!= next
)) {
3245 next
->timestamp
= now
;
3250 prepare_task_switch(rq
, next
);
3251 prev
= context_switch(rq
, prev
, next
);
3254 * this_rq must be evaluated again because prev may have moved
3255 * CPUs since it called schedule(), thus the 'rq' on its stack
3256 * frame will be invalid.
3258 finish_task_switch(this_rq(), prev
);
3260 spin_unlock_irq(&rq
->lock
);
3263 if (unlikely(reacquire_kernel_lock(prev
) < 0))
3264 goto need_resched_nonpreemptible
;
3265 preempt_enable_no_resched();
3266 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3270 EXPORT_SYMBOL(schedule
);
3272 #ifdef CONFIG_PREEMPT
3274 * this is is the entry point to schedule() from in-kernel preemption
3275 * off of preempt_enable. Kernel preemptions off return from interrupt
3276 * occur there and call schedule directly.
3278 asmlinkage
void __sched
preempt_schedule(void)
3280 struct thread_info
*ti
= current_thread_info();
3281 #ifdef CONFIG_PREEMPT_BKL
3282 struct task_struct
*task
= current
;
3283 int saved_lock_depth
;
3286 * If there is a non-zero preempt_count or interrupts are disabled,
3287 * we do not want to preempt the current task. Just return..
3289 if (unlikely(ti
->preempt_count
|| irqs_disabled()))
3293 add_preempt_count(PREEMPT_ACTIVE
);
3295 * We keep the big kernel semaphore locked, but we
3296 * clear ->lock_depth so that schedule() doesnt
3297 * auto-release the semaphore:
3299 #ifdef CONFIG_PREEMPT_BKL
3300 saved_lock_depth
= task
->lock_depth
;
3301 task
->lock_depth
= -1;
3304 #ifdef CONFIG_PREEMPT_BKL
3305 task
->lock_depth
= saved_lock_depth
;
3307 sub_preempt_count(PREEMPT_ACTIVE
);
3309 /* we could miss a preemption opportunity between schedule and now */
3311 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3315 EXPORT_SYMBOL(preempt_schedule
);
3318 * this is is the entry point to schedule() from kernel preemption
3319 * off of irq context.
3320 * Note, that this is called and return with irqs disabled. This will
3321 * protect us against recursive calling from irq.
3323 asmlinkage
void __sched
preempt_schedule_irq(void)
3325 struct thread_info
*ti
= current_thread_info();
3326 #ifdef CONFIG_PREEMPT_BKL
3327 struct task_struct
*task
= current
;
3328 int saved_lock_depth
;
3330 /* Catch callers which need to be fixed*/
3331 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3334 add_preempt_count(PREEMPT_ACTIVE
);
3336 * We keep the big kernel semaphore locked, but we
3337 * clear ->lock_depth so that schedule() doesnt
3338 * auto-release the semaphore:
3340 #ifdef CONFIG_PREEMPT_BKL
3341 saved_lock_depth
= task
->lock_depth
;
3342 task
->lock_depth
= -1;
3346 local_irq_disable();
3347 #ifdef CONFIG_PREEMPT_BKL
3348 task
->lock_depth
= saved_lock_depth
;
3350 sub_preempt_count(PREEMPT_ACTIVE
);
3352 /* we could miss a preemption opportunity between schedule and now */
3354 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3358 #endif /* CONFIG_PREEMPT */
3360 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3363 task_t
*p
= curr
->private;
3364 return try_to_wake_up(p
, mode
, sync
);
3367 EXPORT_SYMBOL(default_wake_function
);
3370 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3371 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3372 * number) then we wake all the non-exclusive tasks and one exclusive task.
3374 * There are circumstances in which we can try to wake a task which has already
3375 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3376 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3378 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3379 int nr_exclusive
, int sync
, void *key
)
3381 struct list_head
*tmp
, *next
;
3383 list_for_each_safe(tmp
, next
, &q
->task_list
) {
3386 curr
= list_entry(tmp
, wait_queue_t
, task_list
);
3387 flags
= curr
->flags
;
3388 if (curr
->func(curr
, mode
, sync
, key
) &&
3389 (flags
& WQ_FLAG_EXCLUSIVE
) &&
3396 * __wake_up - wake up threads blocked on a waitqueue.
3398 * @mode: which threads
3399 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3400 * @key: is directly passed to the wakeup function
3402 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3403 int nr_exclusive
, void *key
)
3405 unsigned long flags
;
3407 spin_lock_irqsave(&q
->lock
, flags
);
3408 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3409 spin_unlock_irqrestore(&q
->lock
, flags
);
3412 EXPORT_SYMBOL(__wake_up
);
3415 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3417 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3419 __wake_up_common(q
, mode
, 1, 0, NULL
);
3423 * __wake_up_sync - wake up threads blocked on a waitqueue.
3425 * @mode: which threads
3426 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3428 * The sync wakeup differs that the waker knows that it will schedule
3429 * away soon, so while the target thread will be woken up, it will not
3430 * be migrated to another CPU - ie. the two threads are 'synchronized'
3431 * with each other. This can prevent needless bouncing between CPUs.
3433 * On UP it can prevent extra preemption.
3436 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3438 unsigned long flags
;
3444 if (unlikely(!nr_exclusive
))
3447 spin_lock_irqsave(&q
->lock
, flags
);
3448 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3449 spin_unlock_irqrestore(&q
->lock
, flags
);
3451 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3453 void fastcall
complete(struct completion
*x
)
3455 unsigned long flags
;
3457 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3459 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3461 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3463 EXPORT_SYMBOL(complete
);
3465 void fastcall
complete_all(struct completion
*x
)
3467 unsigned long flags
;
3469 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3470 x
->done
+= UINT_MAX
/2;
3471 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3473 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3475 EXPORT_SYMBOL(complete_all
);
3477 void fastcall __sched
wait_for_completion(struct completion
*x
)
3480 spin_lock_irq(&x
->wait
.lock
);
3482 DECLARE_WAITQUEUE(wait
, current
);
3484 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3485 __add_wait_queue_tail(&x
->wait
, &wait
);
3487 __set_current_state(TASK_UNINTERRUPTIBLE
);
3488 spin_unlock_irq(&x
->wait
.lock
);
3490 spin_lock_irq(&x
->wait
.lock
);
3492 __remove_wait_queue(&x
->wait
, &wait
);
3495 spin_unlock_irq(&x
->wait
.lock
);
3497 EXPORT_SYMBOL(wait_for_completion
);
3499 unsigned long fastcall __sched
3500 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3504 spin_lock_irq(&x
->wait
.lock
);
3506 DECLARE_WAITQUEUE(wait
, current
);
3508 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3509 __add_wait_queue_tail(&x
->wait
, &wait
);
3511 __set_current_state(TASK_UNINTERRUPTIBLE
);
3512 spin_unlock_irq(&x
->wait
.lock
);
3513 timeout
= schedule_timeout(timeout
);
3514 spin_lock_irq(&x
->wait
.lock
);
3516 __remove_wait_queue(&x
->wait
, &wait
);
3520 __remove_wait_queue(&x
->wait
, &wait
);
3524 spin_unlock_irq(&x
->wait
.lock
);
3527 EXPORT_SYMBOL(wait_for_completion_timeout
);
3529 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3535 spin_lock_irq(&x
->wait
.lock
);
3537 DECLARE_WAITQUEUE(wait
, current
);
3539 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3540 __add_wait_queue_tail(&x
->wait
, &wait
);
3542 if (signal_pending(current
)) {
3544 __remove_wait_queue(&x
->wait
, &wait
);
3547 __set_current_state(TASK_INTERRUPTIBLE
);
3548 spin_unlock_irq(&x
->wait
.lock
);
3550 spin_lock_irq(&x
->wait
.lock
);
3552 __remove_wait_queue(&x
->wait
, &wait
);
3556 spin_unlock_irq(&x
->wait
.lock
);
3560 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3562 unsigned long fastcall __sched
3563 wait_for_completion_interruptible_timeout(struct completion
*x
,
3564 unsigned long timeout
)
3568 spin_lock_irq(&x
->wait
.lock
);
3570 DECLARE_WAITQUEUE(wait
, current
);
3572 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3573 __add_wait_queue_tail(&x
->wait
, &wait
);
3575 if (signal_pending(current
)) {
3576 timeout
= -ERESTARTSYS
;
3577 __remove_wait_queue(&x
->wait
, &wait
);
3580 __set_current_state(TASK_INTERRUPTIBLE
);
3581 spin_unlock_irq(&x
->wait
.lock
);
3582 timeout
= schedule_timeout(timeout
);
3583 spin_lock_irq(&x
->wait
.lock
);
3585 __remove_wait_queue(&x
->wait
, &wait
);
3589 __remove_wait_queue(&x
->wait
, &wait
);
3593 spin_unlock_irq(&x
->wait
.lock
);
3596 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3599 #define SLEEP_ON_VAR \
3600 unsigned long flags; \
3601 wait_queue_t wait; \
3602 init_waitqueue_entry(&wait, current);
3604 #define SLEEP_ON_HEAD \
3605 spin_lock_irqsave(&q->lock,flags); \
3606 __add_wait_queue(q, &wait); \
3607 spin_unlock(&q->lock);
3609 #define SLEEP_ON_TAIL \
3610 spin_lock_irq(&q->lock); \
3611 __remove_wait_queue(q, &wait); \
3612 spin_unlock_irqrestore(&q->lock, flags);
3614 void fastcall __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3618 current
->state
= TASK_INTERRUPTIBLE
;
3625 EXPORT_SYMBOL(interruptible_sleep_on
);
3627 long fastcall __sched
3628 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3632 current
->state
= TASK_INTERRUPTIBLE
;
3635 timeout
= schedule_timeout(timeout
);
3641 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3643 void fastcall __sched
sleep_on(wait_queue_head_t
*q
)
3647 current
->state
= TASK_UNINTERRUPTIBLE
;
3654 EXPORT_SYMBOL(sleep_on
);
3656 long fastcall __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3660 current
->state
= TASK_UNINTERRUPTIBLE
;
3663 timeout
= schedule_timeout(timeout
);
3669 EXPORT_SYMBOL(sleep_on_timeout
);
3671 void set_user_nice(task_t
*p
, long nice
)
3673 unsigned long flags
;
3674 prio_array_t
*array
;
3676 int old_prio
, new_prio
, delta
;
3678 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3681 * We have to be careful, if called from sys_setpriority(),
3682 * the task might be in the middle of scheduling on another CPU.
3684 rq
= task_rq_lock(p
, &flags
);
3686 * The RT priorities are set via sched_setscheduler(), but we still
3687 * allow the 'normal' nice value to be set - but as expected
3688 * it wont have any effect on scheduling until the task is
3689 * not SCHED_NORMAL/SCHED_BATCH:
3692 p
->static_prio
= NICE_TO_PRIO(nice
);
3697 dequeue_task(p
, array
);
3698 dec_raw_weighted_load(rq
, p
);
3702 new_prio
= NICE_TO_PRIO(nice
);
3703 delta
= new_prio
- old_prio
;
3704 p
->static_prio
= NICE_TO_PRIO(nice
);
3709 enqueue_task(p
, array
);
3710 inc_raw_weighted_load(rq
, p
);
3712 * If the task increased its priority or is running and
3713 * lowered its priority, then reschedule its CPU:
3715 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3716 resched_task(rq
->curr
);
3719 task_rq_unlock(rq
, &flags
);
3722 EXPORT_SYMBOL(set_user_nice
);
3725 * can_nice - check if a task can reduce its nice value
3729 int can_nice(const task_t
*p
, const int nice
)
3731 /* convert nice value [19,-20] to rlimit style value [1,40] */
3732 int nice_rlim
= 20 - nice
;
3733 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
3734 capable(CAP_SYS_NICE
));
3737 #ifdef __ARCH_WANT_SYS_NICE
3740 * sys_nice - change the priority of the current process.
3741 * @increment: priority increment
3743 * sys_setpriority is a more generic, but much slower function that
3744 * does similar things.
3746 asmlinkage
long sys_nice(int increment
)
3752 * Setpriority might change our priority at the same moment.
3753 * We don't have to worry. Conceptually one call occurs first
3754 * and we have a single winner.
3756 if (increment
< -40)
3761 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
3767 if (increment
< 0 && !can_nice(current
, nice
))
3770 retval
= security_task_setnice(current
, nice
);
3774 set_user_nice(current
, nice
);
3781 * task_prio - return the priority value of a given task.
3782 * @p: the task in question.
3784 * This is the priority value as seen by users in /proc.
3785 * RT tasks are offset by -200. Normal tasks are centered
3786 * around 0, value goes from -16 to +15.
3788 int task_prio(const task_t
*p
)
3790 return p
->prio
- MAX_RT_PRIO
;
3794 * task_nice - return the nice value of a given task.
3795 * @p: the task in question.
3797 int task_nice(const task_t
*p
)
3799 return TASK_NICE(p
);
3801 EXPORT_SYMBOL_GPL(task_nice
);
3804 * idle_cpu - is a given cpu idle currently?
3805 * @cpu: the processor in question.
3807 int idle_cpu(int cpu
)
3809 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
3813 * idle_task - return the idle task for a given cpu.
3814 * @cpu: the processor in question.
3816 task_t
*idle_task(int cpu
)
3818 return cpu_rq(cpu
)->idle
;
3822 * find_process_by_pid - find a process with a matching PID value.
3823 * @pid: the pid in question.
3825 static inline task_t
*find_process_by_pid(pid_t pid
)
3827 return pid
? find_task_by_pid(pid
) : current
;
3830 /* Actually do priority change: must hold rq lock. */
3831 static void __setscheduler(struct task_struct
*p
, int policy
, int prio
)
3835 p
->rt_priority
= prio
;
3836 if (policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
) {
3837 p
->prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
3839 p
->prio
= p
->static_prio
;
3841 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
3843 if (policy
== SCHED_BATCH
)
3850 * sched_setscheduler - change the scheduling policy and/or RT priority of
3852 * @p: the task in question.
3853 * @policy: new policy.
3854 * @param: structure containing the new RT priority.
3856 int sched_setscheduler(struct task_struct
*p
, int policy
,
3857 struct sched_param
*param
)
3860 int oldprio
, oldpolicy
= -1;
3861 prio_array_t
*array
;
3862 unsigned long flags
;
3865 /* may grab non-irq protected spin_locks */
3866 BUG_ON(in_interrupt());
3868 /* double check policy once rq lock held */
3870 policy
= oldpolicy
= p
->policy
;
3871 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3872 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
)
3875 * Valid priorities for SCHED_FIFO and SCHED_RR are
3876 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
3879 if (param
->sched_priority
< 0 ||
3880 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3881 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
3883 if ((policy
== SCHED_NORMAL
|| policy
== SCHED_BATCH
)
3884 != (param
->sched_priority
== 0))
3888 * Allow unprivileged RT tasks to decrease priority:
3890 if (!capable(CAP_SYS_NICE
)) {
3892 * can't change policy, except between SCHED_NORMAL
3895 if (((policy
!= SCHED_NORMAL
&& p
->policy
!= SCHED_BATCH
) &&
3896 (policy
!= SCHED_BATCH
&& p
->policy
!= SCHED_NORMAL
)) &&
3897 !p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
)
3899 /* can't increase priority */
3900 if ((policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
) &&
3901 param
->sched_priority
> p
->rt_priority
&&
3902 param
->sched_priority
>
3903 p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
)
3905 /* can't change other user's priorities */
3906 if ((current
->euid
!= p
->euid
) &&
3907 (current
->euid
!= p
->uid
))
3911 retval
= security_task_setscheduler(p
, policy
, param
);
3915 * To be able to change p->policy safely, the apropriate
3916 * runqueue lock must be held.
3918 rq
= task_rq_lock(p
, &flags
);
3919 /* recheck policy now with rq lock held */
3920 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3921 policy
= oldpolicy
= -1;
3922 task_rq_unlock(rq
, &flags
);
3927 deactivate_task(p
, rq
);
3929 __setscheduler(p
, policy
, param
->sched_priority
);
3931 __activate_task(p
, rq
);
3933 * Reschedule if we are currently running on this runqueue and
3934 * our priority decreased, or if we are not currently running on
3935 * this runqueue and our priority is higher than the current's
3937 if (task_running(rq
, p
)) {
3938 if (p
->prio
> oldprio
)
3939 resched_task(rq
->curr
);
3940 } else if (TASK_PREEMPTS_CURR(p
, rq
))
3941 resched_task(rq
->curr
);
3943 task_rq_unlock(rq
, &flags
);
3946 EXPORT_SYMBOL_GPL(sched_setscheduler
);
3949 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
3952 struct sched_param lparam
;
3953 struct task_struct
*p
;
3955 if (!param
|| pid
< 0)
3957 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
3959 read_lock_irq(&tasklist_lock
);
3960 p
= find_process_by_pid(pid
);
3962 read_unlock_irq(&tasklist_lock
);
3965 retval
= sched_setscheduler(p
, policy
, &lparam
);
3966 read_unlock_irq(&tasklist_lock
);
3971 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3972 * @pid: the pid in question.
3973 * @policy: new policy.
3974 * @param: structure containing the new RT priority.
3976 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
3977 struct sched_param __user
*param
)
3979 /* negative values for policy are not valid */
3983 return do_sched_setscheduler(pid
, policy
, param
);
3987 * sys_sched_setparam - set/change the RT priority of a thread
3988 * @pid: the pid in question.
3989 * @param: structure containing the new RT priority.
3991 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
3993 return do_sched_setscheduler(pid
, -1, param
);
3997 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3998 * @pid: the pid in question.
4000 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4002 int retval
= -EINVAL
;
4009 read_lock(&tasklist_lock
);
4010 p
= find_process_by_pid(pid
);
4012 retval
= security_task_getscheduler(p
);
4016 read_unlock(&tasklist_lock
);
4023 * sys_sched_getscheduler - get the RT priority of a thread
4024 * @pid: the pid in question.
4025 * @param: structure containing the RT priority.
4027 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4029 struct sched_param lp
;
4030 int retval
= -EINVAL
;
4033 if (!param
|| pid
< 0)
4036 read_lock(&tasklist_lock
);
4037 p
= find_process_by_pid(pid
);
4042 retval
= security_task_getscheduler(p
);
4046 lp
.sched_priority
= p
->rt_priority
;
4047 read_unlock(&tasklist_lock
);
4050 * This one might sleep, we cannot do it with a spinlock held ...
4052 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4058 read_unlock(&tasklist_lock
);
4062 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4066 cpumask_t cpus_allowed
;
4069 read_lock(&tasklist_lock
);
4071 p
= find_process_by_pid(pid
);
4073 read_unlock(&tasklist_lock
);
4074 unlock_cpu_hotplug();
4079 * It is not safe to call set_cpus_allowed with the
4080 * tasklist_lock held. We will bump the task_struct's
4081 * usage count and then drop tasklist_lock.
4084 read_unlock(&tasklist_lock
);
4087 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4088 !capable(CAP_SYS_NICE
))
4091 retval
= security_task_setscheduler(p
, 0, NULL
);
4095 cpus_allowed
= cpuset_cpus_allowed(p
);
4096 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4097 retval
= set_cpus_allowed(p
, new_mask
);
4101 unlock_cpu_hotplug();
4105 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4106 cpumask_t
*new_mask
)
4108 if (len
< sizeof(cpumask_t
)) {
4109 memset(new_mask
, 0, sizeof(cpumask_t
));
4110 } else if (len
> sizeof(cpumask_t
)) {
4111 len
= sizeof(cpumask_t
);
4113 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4117 * sys_sched_setaffinity - set the cpu affinity of a process
4118 * @pid: pid of the process
4119 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4120 * @user_mask_ptr: user-space pointer to the new cpu mask
4122 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4123 unsigned long __user
*user_mask_ptr
)
4128 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4132 return sched_setaffinity(pid
, new_mask
);
4136 * Represents all cpu's present in the system
4137 * In systems capable of hotplug, this map could dynamically grow
4138 * as new cpu's are detected in the system via any platform specific
4139 * method, such as ACPI for e.g.
4142 cpumask_t cpu_present_map __read_mostly
;
4143 EXPORT_SYMBOL(cpu_present_map
);
4146 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4147 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4150 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4156 read_lock(&tasklist_lock
);
4159 p
= find_process_by_pid(pid
);
4163 retval
= security_task_getscheduler(p
);
4167 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4170 read_unlock(&tasklist_lock
);
4171 unlock_cpu_hotplug();
4179 * sys_sched_getaffinity - get the cpu affinity of a process
4180 * @pid: pid of the process
4181 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4182 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4184 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4185 unsigned long __user
*user_mask_ptr
)
4190 if (len
< sizeof(cpumask_t
))
4193 ret
= sched_getaffinity(pid
, &mask
);
4197 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4200 return sizeof(cpumask_t
);
4204 * sys_sched_yield - yield the current processor to other threads.
4206 * this function yields the current CPU by moving the calling thread
4207 * to the expired array. If there are no other threads running on this
4208 * CPU then this function will return.
4210 asmlinkage
long sys_sched_yield(void)
4212 runqueue_t
*rq
= this_rq_lock();
4213 prio_array_t
*array
= current
->array
;
4214 prio_array_t
*target
= rq
->expired
;
4216 schedstat_inc(rq
, yld_cnt
);
4218 * We implement yielding by moving the task into the expired
4221 * (special rule: RT tasks will just roundrobin in the active
4224 if (rt_task(current
))
4225 target
= rq
->active
;
4227 if (array
->nr_active
== 1) {
4228 schedstat_inc(rq
, yld_act_empty
);
4229 if (!rq
->expired
->nr_active
)
4230 schedstat_inc(rq
, yld_both_empty
);
4231 } else if (!rq
->expired
->nr_active
)
4232 schedstat_inc(rq
, yld_exp_empty
);
4234 if (array
!= target
) {
4235 dequeue_task(current
, array
);
4236 enqueue_task(current
, target
);
4239 * requeue_task is cheaper so perform that if possible.
4241 requeue_task(current
, array
);
4244 * Since we are going to call schedule() anyway, there's
4245 * no need to preempt or enable interrupts:
4247 __release(rq
->lock
);
4248 _raw_spin_unlock(&rq
->lock
);
4249 preempt_enable_no_resched();
4256 static inline void __cond_resched(void)
4258 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4259 __might_sleep(__FILE__
, __LINE__
);
4262 * The BKS might be reacquired before we have dropped
4263 * PREEMPT_ACTIVE, which could trigger a second
4264 * cond_resched() call.
4266 if (unlikely(preempt_count()))
4268 if (unlikely(system_state
!= SYSTEM_RUNNING
))
4271 add_preempt_count(PREEMPT_ACTIVE
);
4273 sub_preempt_count(PREEMPT_ACTIVE
);
4274 } while (need_resched());
4277 int __sched
cond_resched(void)
4279 if (need_resched()) {
4286 EXPORT_SYMBOL(cond_resched
);
4289 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4290 * call schedule, and on return reacquire the lock.
4292 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4293 * operations here to prevent schedule() from being called twice (once via
4294 * spin_unlock(), once by hand).
4296 int cond_resched_lock(spinlock_t
*lock
)
4300 if (need_lockbreak(lock
)) {
4306 if (need_resched()) {
4307 _raw_spin_unlock(lock
);
4308 preempt_enable_no_resched();
4316 EXPORT_SYMBOL(cond_resched_lock
);
4318 int __sched
cond_resched_softirq(void)
4320 BUG_ON(!in_softirq());
4322 if (need_resched()) {
4323 __local_bh_enable();
4331 EXPORT_SYMBOL(cond_resched_softirq
);
4335 * yield - yield the current processor to other threads.
4337 * this is a shortcut for kernel-space yielding - it marks the
4338 * thread runnable and calls sys_sched_yield().
4340 void __sched
yield(void)
4342 set_current_state(TASK_RUNNING
);
4346 EXPORT_SYMBOL(yield
);
4349 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4350 * that process accounting knows that this is a task in IO wait state.
4352 * But don't do that if it is a deliberate, throttling IO wait (this task
4353 * has set its backing_dev_info: the queue against which it should throttle)
4355 void __sched
io_schedule(void)
4357 struct runqueue
*rq
= &__raw_get_cpu_var(runqueues
);
4359 atomic_inc(&rq
->nr_iowait
);
4361 atomic_dec(&rq
->nr_iowait
);
4364 EXPORT_SYMBOL(io_schedule
);
4366 long __sched
io_schedule_timeout(long timeout
)
4368 struct runqueue
*rq
= &__raw_get_cpu_var(runqueues
);
4371 atomic_inc(&rq
->nr_iowait
);
4372 ret
= schedule_timeout(timeout
);
4373 atomic_dec(&rq
->nr_iowait
);
4378 * sys_sched_get_priority_max - return maximum RT priority.
4379 * @policy: scheduling class.
4381 * this syscall returns the maximum rt_priority that can be used
4382 * by a given scheduling class.
4384 asmlinkage
long sys_sched_get_priority_max(int policy
)
4391 ret
= MAX_USER_RT_PRIO
-1;
4402 * sys_sched_get_priority_min - return minimum RT priority.
4403 * @policy: scheduling class.
4405 * this syscall returns the minimum rt_priority that can be used
4406 * by a given scheduling class.
4408 asmlinkage
long sys_sched_get_priority_min(int policy
)
4425 * sys_sched_rr_get_interval - return the default timeslice of a process.
4426 * @pid: pid of the process.
4427 * @interval: userspace pointer to the timeslice value.
4429 * this syscall writes the default timeslice value of a given process
4430 * into the user-space timespec buffer. A value of '0' means infinity.
4433 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4435 int retval
= -EINVAL
;
4443 read_lock(&tasklist_lock
);
4444 p
= find_process_by_pid(pid
);
4448 retval
= security_task_getscheduler(p
);
4452 jiffies_to_timespec(p
->policy
== SCHED_FIFO
?
4453 0 : task_timeslice(p
), &t
);
4454 read_unlock(&tasklist_lock
);
4455 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4459 read_unlock(&tasklist_lock
);
4463 static inline struct task_struct
*eldest_child(struct task_struct
*p
)
4465 if (list_empty(&p
->children
)) return NULL
;
4466 return list_entry(p
->children
.next
,struct task_struct
,sibling
);
4469 static inline struct task_struct
*older_sibling(struct task_struct
*p
)
4471 if (p
->sibling
.prev
==&p
->parent
->children
) return NULL
;
4472 return list_entry(p
->sibling
.prev
,struct task_struct
,sibling
);
4475 static inline struct task_struct
*younger_sibling(struct task_struct
*p
)
4477 if (p
->sibling
.next
==&p
->parent
->children
) return NULL
;
4478 return list_entry(p
->sibling
.next
,struct task_struct
,sibling
);
4481 static void show_task(task_t
*p
)
4485 unsigned long free
= 0;
4486 static const char *stat_nam
[] = { "R", "S", "D", "T", "t", "Z", "X" };
4488 printk("%-13.13s ", p
->comm
);
4489 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4490 if (state
< ARRAY_SIZE(stat_nam
))
4491 printk(stat_nam
[state
]);
4494 #if (BITS_PER_LONG == 32)
4495 if (state
== TASK_RUNNING
)
4496 printk(" running ");
4498 printk(" %08lX ", thread_saved_pc(p
));
4500 if (state
== TASK_RUNNING
)
4501 printk(" running task ");
4503 printk(" %016lx ", thread_saved_pc(p
));
4505 #ifdef CONFIG_DEBUG_STACK_USAGE
4507 unsigned long *n
= end_of_stack(p
);
4510 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4513 printk("%5lu %5d %6d ", free
, p
->pid
, p
->parent
->pid
);
4514 if ((relative
= eldest_child(p
)))
4515 printk("%5d ", relative
->pid
);
4518 if ((relative
= younger_sibling(p
)))
4519 printk("%7d", relative
->pid
);
4522 if ((relative
= older_sibling(p
)))
4523 printk(" %5d", relative
->pid
);
4527 printk(" (L-TLB)\n");
4529 printk(" (NOTLB)\n");
4531 if (state
!= TASK_RUNNING
)
4532 show_stack(p
, NULL
);
4535 void show_state(void)
4539 #if (BITS_PER_LONG == 32)
4542 printk(" task PC pid father child younger older\n");
4546 printk(" task PC pid father child younger older\n");
4548 read_lock(&tasklist_lock
);
4549 do_each_thread(g
, p
) {
4551 * reset the NMI-timeout, listing all files on a slow
4552 * console might take alot of time:
4554 touch_nmi_watchdog();
4556 } while_each_thread(g
, p
);
4558 read_unlock(&tasklist_lock
);
4559 mutex_debug_show_all_locks();
4563 * init_idle - set up an idle thread for a given CPU
4564 * @idle: task in question
4565 * @cpu: cpu the idle task belongs to
4567 * NOTE: this function does not set the idle thread's NEED_RESCHED
4568 * flag, to make booting more robust.
4570 void __devinit
init_idle(task_t
*idle
, int cpu
)
4572 runqueue_t
*rq
= cpu_rq(cpu
);
4573 unsigned long flags
;
4575 idle
->timestamp
= sched_clock();
4576 idle
->sleep_avg
= 0;
4578 idle
->prio
= MAX_PRIO
;
4579 idle
->state
= TASK_RUNNING
;
4580 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4581 set_task_cpu(idle
, cpu
);
4583 spin_lock_irqsave(&rq
->lock
, flags
);
4584 rq
->curr
= rq
->idle
= idle
;
4585 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4588 spin_unlock_irqrestore(&rq
->lock
, flags
);
4590 /* Set the preempt count _outside_ the spinlocks! */
4591 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4592 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4594 task_thread_info(idle
)->preempt_count
= 0;
4599 * In a system that switches off the HZ timer nohz_cpu_mask
4600 * indicates which cpus entered this state. This is used
4601 * in the rcu update to wait only for active cpus. For system
4602 * which do not switch off the HZ timer nohz_cpu_mask should
4603 * always be CPU_MASK_NONE.
4605 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4609 * This is how migration works:
4611 * 1) we queue a migration_req_t structure in the source CPU's
4612 * runqueue and wake up that CPU's migration thread.
4613 * 2) we down() the locked semaphore => thread blocks.
4614 * 3) migration thread wakes up (implicitly it forces the migrated
4615 * thread off the CPU)
4616 * 4) it gets the migration request and checks whether the migrated
4617 * task is still in the wrong runqueue.
4618 * 5) if it's in the wrong runqueue then the migration thread removes
4619 * it and puts it into the right queue.
4620 * 6) migration thread up()s the semaphore.
4621 * 7) we wake up and the migration is done.
4625 * Change a given task's CPU affinity. Migrate the thread to a
4626 * proper CPU and schedule it away if the CPU it's executing on
4627 * is removed from the allowed bitmask.
4629 * NOTE: the caller must have a valid reference to the task, the
4630 * task must not exit() & deallocate itself prematurely. The
4631 * call is not atomic; no spinlocks may be held.
4633 int set_cpus_allowed(task_t
*p
, cpumask_t new_mask
)
4635 unsigned long flags
;
4637 migration_req_t req
;
4640 rq
= task_rq_lock(p
, &flags
);
4641 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4646 p
->cpus_allowed
= new_mask
;
4647 /* Can the task run on the task's current CPU? If so, we're done */
4648 if (cpu_isset(task_cpu(p
), new_mask
))
4651 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4652 /* Need help from migration thread: drop lock and wait. */
4653 task_rq_unlock(rq
, &flags
);
4654 wake_up_process(rq
->migration_thread
);
4655 wait_for_completion(&req
.done
);
4656 tlb_migrate_finish(p
->mm
);
4660 task_rq_unlock(rq
, &flags
);
4664 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4667 * Move (not current) task off this cpu, onto dest cpu. We're doing
4668 * this because either it can't run here any more (set_cpus_allowed()
4669 * away from this CPU, or CPU going down), or because we're
4670 * attempting to rebalance this task on exec (sched_exec).
4672 * So we race with normal scheduler movements, but that's OK, as long
4673 * as the task is no longer on this CPU.
4675 * Returns non-zero if task was successfully migrated.
4677 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4679 runqueue_t
*rq_dest
, *rq_src
;
4682 if (unlikely(cpu_is_offline(dest_cpu
)))
4685 rq_src
= cpu_rq(src_cpu
);
4686 rq_dest
= cpu_rq(dest_cpu
);
4688 double_rq_lock(rq_src
, rq_dest
);
4689 /* Already moved. */
4690 if (task_cpu(p
) != src_cpu
)
4692 /* Affinity changed (again). */
4693 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
4696 set_task_cpu(p
, dest_cpu
);
4699 * Sync timestamp with rq_dest's before activating.
4700 * The same thing could be achieved by doing this step
4701 * afterwards, and pretending it was a local activate.
4702 * This way is cleaner and logically correct.
4704 p
->timestamp
= p
->timestamp
- rq_src
->timestamp_last_tick
4705 + rq_dest
->timestamp_last_tick
;
4706 deactivate_task(p
, rq_src
);
4707 activate_task(p
, rq_dest
, 0);
4708 if (TASK_PREEMPTS_CURR(p
, rq_dest
))
4709 resched_task(rq_dest
->curr
);
4713 double_rq_unlock(rq_src
, rq_dest
);
4718 * migration_thread - this is a highprio system thread that performs
4719 * thread migration by bumping thread off CPU then 'pushing' onto
4722 static int migration_thread(void *data
)
4725 int cpu
= (long)data
;
4728 BUG_ON(rq
->migration_thread
!= current
);
4730 set_current_state(TASK_INTERRUPTIBLE
);
4731 while (!kthread_should_stop()) {
4732 struct list_head
*head
;
4733 migration_req_t
*req
;
4737 spin_lock_irq(&rq
->lock
);
4739 if (cpu_is_offline(cpu
)) {
4740 spin_unlock_irq(&rq
->lock
);
4744 if (rq
->active_balance
) {
4745 active_load_balance(rq
, cpu
);
4746 rq
->active_balance
= 0;
4749 head
= &rq
->migration_queue
;
4751 if (list_empty(head
)) {
4752 spin_unlock_irq(&rq
->lock
);
4754 set_current_state(TASK_INTERRUPTIBLE
);
4757 req
= list_entry(head
->next
, migration_req_t
, list
);
4758 list_del_init(head
->next
);
4760 spin_unlock(&rq
->lock
);
4761 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
4764 complete(&req
->done
);
4766 __set_current_state(TASK_RUNNING
);
4770 /* Wait for kthread_stop */
4771 set_current_state(TASK_INTERRUPTIBLE
);
4772 while (!kthread_should_stop()) {
4774 set_current_state(TASK_INTERRUPTIBLE
);
4776 __set_current_state(TASK_RUNNING
);
4780 #ifdef CONFIG_HOTPLUG_CPU
4781 /* Figure out where task on dead CPU should go, use force if neccessary. */
4782 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*tsk
)
4785 unsigned long flags
;
4791 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
4792 cpus_and(mask
, mask
, tsk
->cpus_allowed
);
4793 dest_cpu
= any_online_cpu(mask
);
4795 /* On any allowed CPU? */
4796 if (dest_cpu
== NR_CPUS
)
4797 dest_cpu
= any_online_cpu(tsk
->cpus_allowed
);
4799 /* No more Mr. Nice Guy. */
4800 if (dest_cpu
== NR_CPUS
) {
4801 rq
= task_rq_lock(tsk
, &flags
);
4802 cpus_setall(tsk
->cpus_allowed
);
4803 dest_cpu
= any_online_cpu(tsk
->cpus_allowed
);
4804 task_rq_unlock(rq
, &flags
);
4807 * Don't tell them about moving exiting tasks or
4808 * kernel threads (both mm NULL), since they never
4811 if (tsk
->mm
&& printk_ratelimit())
4812 printk(KERN_INFO
"process %d (%s) no "
4813 "longer affine to cpu%d\n",
4814 tsk
->pid
, tsk
->comm
, dead_cpu
);
4816 if (!__migrate_task(tsk
, dead_cpu
, dest_cpu
))
4821 * While a dead CPU has no uninterruptible tasks queued at this point,
4822 * it might still have a nonzero ->nr_uninterruptible counter, because
4823 * for performance reasons the counter is not stricly tracking tasks to
4824 * their home CPUs. So we just add the counter to another CPU's counter,
4825 * to keep the global sum constant after CPU-down:
4827 static void migrate_nr_uninterruptible(runqueue_t
*rq_src
)
4829 runqueue_t
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
4830 unsigned long flags
;
4832 local_irq_save(flags
);
4833 double_rq_lock(rq_src
, rq_dest
);
4834 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
4835 rq_src
->nr_uninterruptible
= 0;
4836 double_rq_unlock(rq_src
, rq_dest
);
4837 local_irq_restore(flags
);
4840 /* Run through task list and migrate tasks from the dead cpu. */
4841 static void migrate_live_tasks(int src_cpu
)
4843 struct task_struct
*tsk
, *t
;
4845 write_lock_irq(&tasklist_lock
);
4847 do_each_thread(t
, tsk
) {
4851 if (task_cpu(tsk
) == src_cpu
)
4852 move_task_off_dead_cpu(src_cpu
, tsk
);
4853 } while_each_thread(t
, tsk
);
4855 write_unlock_irq(&tasklist_lock
);
4858 /* Schedules idle task to be the next runnable task on current CPU.
4859 * It does so by boosting its priority to highest possible and adding it to
4860 * the _front_ of runqueue. Used by CPU offline code.
4862 void sched_idle_next(void)
4864 int cpu
= smp_processor_id();
4865 runqueue_t
*rq
= this_rq();
4866 struct task_struct
*p
= rq
->idle
;
4867 unsigned long flags
;
4869 /* cpu has to be offline */
4870 BUG_ON(cpu_online(cpu
));
4872 /* Strictly not necessary since rest of the CPUs are stopped by now
4873 * and interrupts disabled on current cpu.
4875 spin_lock_irqsave(&rq
->lock
, flags
);
4877 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
4878 /* Add idle task to _front_ of it's priority queue */
4879 __activate_idle_task(p
, rq
);
4881 spin_unlock_irqrestore(&rq
->lock
, flags
);
4884 /* Ensures that the idle task is using init_mm right before its cpu goes
4887 void idle_task_exit(void)
4889 struct mm_struct
*mm
= current
->active_mm
;
4891 BUG_ON(cpu_online(smp_processor_id()));
4894 switch_mm(mm
, &init_mm
, current
);
4898 static void migrate_dead(unsigned int dead_cpu
, task_t
*tsk
)
4900 struct runqueue
*rq
= cpu_rq(dead_cpu
);
4902 /* Must be exiting, otherwise would be on tasklist. */
4903 BUG_ON(tsk
->exit_state
!= EXIT_ZOMBIE
&& tsk
->exit_state
!= EXIT_DEAD
);
4905 /* Cannot have done final schedule yet: would have vanished. */
4906 BUG_ON(tsk
->flags
& PF_DEAD
);
4908 get_task_struct(tsk
);
4911 * Drop lock around migration; if someone else moves it,
4912 * that's OK. No task can be added to this CPU, so iteration is
4915 spin_unlock_irq(&rq
->lock
);
4916 move_task_off_dead_cpu(dead_cpu
, tsk
);
4917 spin_lock_irq(&rq
->lock
);
4919 put_task_struct(tsk
);
4922 /* release_task() removes task from tasklist, so we won't find dead tasks. */
4923 static void migrate_dead_tasks(unsigned int dead_cpu
)
4926 struct runqueue
*rq
= cpu_rq(dead_cpu
);
4928 for (arr
= 0; arr
< 2; arr
++) {
4929 for (i
= 0; i
< MAX_PRIO
; i
++) {
4930 struct list_head
*list
= &rq
->arrays
[arr
].queue
[i
];
4931 while (!list_empty(list
))
4932 migrate_dead(dead_cpu
,
4933 list_entry(list
->next
, task_t
,
4938 #endif /* CONFIG_HOTPLUG_CPU */
4941 * migration_call - callback that gets triggered when a CPU is added.
4942 * Here we can start up the necessary migration thread for the new CPU.
4944 static int __cpuinit
migration_call(struct notifier_block
*nfb
,
4945 unsigned long action
,
4948 int cpu
= (long)hcpu
;
4949 struct task_struct
*p
;
4950 struct runqueue
*rq
;
4951 unsigned long flags
;
4954 case CPU_UP_PREPARE
:
4955 p
= kthread_create(migration_thread
, hcpu
, "migration/%d",cpu
);
4958 p
->flags
|= PF_NOFREEZE
;
4959 kthread_bind(p
, cpu
);
4960 /* Must be high prio: stop_machine expects to yield to it. */
4961 rq
= task_rq_lock(p
, &flags
);
4962 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
4963 task_rq_unlock(rq
, &flags
);
4964 cpu_rq(cpu
)->migration_thread
= p
;
4967 /* Strictly unneccessary, as first user will wake it. */
4968 wake_up_process(cpu_rq(cpu
)->migration_thread
);
4970 #ifdef CONFIG_HOTPLUG_CPU
4971 case CPU_UP_CANCELED
:
4972 if (!cpu_rq(cpu
)->migration_thread
)
4974 /* Unbind it from offline cpu so it can run. Fall thru. */
4975 kthread_bind(cpu_rq(cpu
)->migration_thread
,
4976 any_online_cpu(cpu_online_map
));
4977 kthread_stop(cpu_rq(cpu
)->migration_thread
);
4978 cpu_rq(cpu
)->migration_thread
= NULL
;
4981 migrate_live_tasks(cpu
);
4983 kthread_stop(rq
->migration_thread
);
4984 rq
->migration_thread
= NULL
;
4985 /* Idle task back to normal (off runqueue, low prio) */
4986 rq
= task_rq_lock(rq
->idle
, &flags
);
4987 deactivate_task(rq
->idle
, rq
);
4988 rq
->idle
->static_prio
= MAX_PRIO
;
4989 __setscheduler(rq
->idle
, SCHED_NORMAL
, 0);
4990 migrate_dead_tasks(cpu
);
4991 task_rq_unlock(rq
, &flags
);
4992 migrate_nr_uninterruptible(rq
);
4993 BUG_ON(rq
->nr_running
!= 0);
4995 /* No need to migrate the tasks: it was best-effort if
4996 * they didn't do lock_cpu_hotplug(). Just wake up
4997 * the requestors. */
4998 spin_lock_irq(&rq
->lock
);
4999 while (!list_empty(&rq
->migration_queue
)) {
5000 migration_req_t
*req
;
5001 req
= list_entry(rq
->migration_queue
.next
,
5002 migration_req_t
, list
);
5003 list_del_init(&req
->list
);
5004 complete(&req
->done
);
5006 spin_unlock_irq(&rq
->lock
);
5013 /* Register at highest priority so that task migration (migrate_all_tasks)
5014 * happens before everything else.
5016 static struct notifier_block __cpuinitdata migration_notifier
= {
5017 .notifier_call
= migration_call
,
5021 int __init
migration_init(void)
5023 void *cpu
= (void *)(long)smp_processor_id();
5024 /* Start one for boot CPU. */
5025 migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5026 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5027 register_cpu_notifier(&migration_notifier
);
5033 #undef SCHED_DOMAIN_DEBUG
5034 #ifdef SCHED_DOMAIN_DEBUG
5035 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5040 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5044 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5049 struct sched_group
*group
= sd
->groups
;
5050 cpumask_t groupmask
;
5052 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5053 cpus_clear(groupmask
);
5056 for (i
= 0; i
< level
+ 1; i
++)
5058 printk("domain %d: ", level
);
5060 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5061 printk("does not load-balance\n");
5063 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain has parent");
5067 printk("span %s\n", str
);
5069 if (!cpu_isset(cpu
, sd
->span
))
5070 printk(KERN_ERR
"ERROR: domain->span does not contain CPU%d\n", cpu
);
5071 if (!cpu_isset(cpu
, group
->cpumask
))
5072 printk(KERN_ERR
"ERROR: domain->groups does not contain CPU%d\n", cpu
);
5075 for (i
= 0; i
< level
+ 2; i
++)
5081 printk(KERN_ERR
"ERROR: group is NULL\n");
5085 if (!group
->cpu_power
) {
5087 printk(KERN_ERR
"ERROR: domain->cpu_power not set\n");
5090 if (!cpus_weight(group
->cpumask
)) {
5092 printk(KERN_ERR
"ERROR: empty group\n");
5095 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5097 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5100 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5102 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5105 group
= group
->next
;
5106 } while (group
!= sd
->groups
);
5109 if (!cpus_equal(sd
->span
, groupmask
))
5110 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5116 if (!cpus_subset(groupmask
, sd
->span
))
5117 printk(KERN_ERR
"ERROR: parent span is not a superset of domain->span\n");
5123 #define sched_domain_debug(sd, cpu) {}
5126 static int sd_degenerate(struct sched_domain
*sd
)
5128 if (cpus_weight(sd
->span
) == 1)
5131 /* Following flags need at least 2 groups */
5132 if (sd
->flags
& (SD_LOAD_BALANCE
|
5133 SD_BALANCE_NEWIDLE
|
5136 if (sd
->groups
!= sd
->groups
->next
)
5140 /* Following flags don't use groups */
5141 if (sd
->flags
& (SD_WAKE_IDLE
|
5149 static int sd_parent_degenerate(struct sched_domain
*sd
,
5150 struct sched_domain
*parent
)
5152 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5154 if (sd_degenerate(parent
))
5157 if (!cpus_equal(sd
->span
, parent
->span
))
5160 /* Does parent contain flags not in child? */
5161 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5162 if (cflags
& SD_WAKE_AFFINE
)
5163 pflags
&= ~SD_WAKE_BALANCE
;
5164 /* Flags needing groups don't count if only 1 group in parent */
5165 if (parent
->groups
== parent
->groups
->next
) {
5166 pflags
&= ~(SD_LOAD_BALANCE
|
5167 SD_BALANCE_NEWIDLE
|
5171 if (~cflags
& pflags
)
5178 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5179 * hold the hotplug lock.
5181 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5183 runqueue_t
*rq
= cpu_rq(cpu
);
5184 struct sched_domain
*tmp
;
5186 /* Remove the sched domains which do not contribute to scheduling. */
5187 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5188 struct sched_domain
*parent
= tmp
->parent
;
5191 if (sd_parent_degenerate(tmp
, parent
))
5192 tmp
->parent
= parent
->parent
;
5195 if (sd
&& sd_degenerate(sd
))
5198 sched_domain_debug(sd
, cpu
);
5200 rcu_assign_pointer(rq
->sd
, sd
);
5203 /* cpus with isolated domains */
5204 static cpumask_t __devinitdata cpu_isolated_map
= CPU_MASK_NONE
;
5206 /* Setup the mask of cpus configured for isolated domains */
5207 static int __init
isolated_cpu_setup(char *str
)
5209 int ints
[NR_CPUS
], i
;
5211 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5212 cpus_clear(cpu_isolated_map
);
5213 for (i
= 1; i
<= ints
[0]; i
++)
5214 if (ints
[i
] < NR_CPUS
)
5215 cpu_set(ints
[i
], cpu_isolated_map
);
5219 __setup ("isolcpus=", isolated_cpu_setup
);
5222 * init_sched_build_groups takes an array of groups, the cpumask we wish
5223 * to span, and a pointer to a function which identifies what group a CPU
5224 * belongs to. The return value of group_fn must be a valid index into the
5225 * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
5226 * keep track of groups covered with a cpumask_t).
5228 * init_sched_build_groups will build a circular linked list of the groups
5229 * covered by the given span, and will set each group's ->cpumask correctly,
5230 * and ->cpu_power to 0.
5232 static void init_sched_build_groups(struct sched_group groups
[], cpumask_t span
,
5233 int (*group_fn
)(int cpu
))
5235 struct sched_group
*first
= NULL
, *last
= NULL
;
5236 cpumask_t covered
= CPU_MASK_NONE
;
5239 for_each_cpu_mask(i
, span
) {
5240 int group
= group_fn(i
);
5241 struct sched_group
*sg
= &groups
[group
];
5244 if (cpu_isset(i
, covered
))
5247 sg
->cpumask
= CPU_MASK_NONE
;
5250 for_each_cpu_mask(j
, span
) {
5251 if (group_fn(j
) != group
)
5254 cpu_set(j
, covered
);
5255 cpu_set(j
, sg
->cpumask
);
5266 #define SD_NODES_PER_DOMAIN 16
5269 * Self-tuning task migration cost measurement between source and target CPUs.
5271 * This is done by measuring the cost of manipulating buffers of varying
5272 * sizes. For a given buffer-size here are the steps that are taken:
5274 * 1) the source CPU reads+dirties a shared buffer
5275 * 2) the target CPU reads+dirties the same shared buffer
5277 * We measure how long they take, in the following 4 scenarios:
5279 * - source: CPU1, target: CPU2 | cost1
5280 * - source: CPU2, target: CPU1 | cost2
5281 * - source: CPU1, target: CPU1 | cost3
5282 * - source: CPU2, target: CPU2 | cost4
5284 * We then calculate the cost3+cost4-cost1-cost2 difference - this is
5285 * the cost of migration.
5287 * We then start off from a small buffer-size and iterate up to larger
5288 * buffer sizes, in 5% steps - measuring each buffer-size separately, and
5289 * doing a maximum search for the cost. (The maximum cost for a migration
5290 * normally occurs when the working set size is around the effective cache
5293 #define SEARCH_SCOPE 2
5294 #define MIN_CACHE_SIZE (64*1024U)
5295 #define DEFAULT_CACHE_SIZE (5*1024*1024U)
5296 #define ITERATIONS 1
5297 #define SIZE_THRESH 130
5298 #define COST_THRESH 130
5301 * The migration cost is a function of 'domain distance'. Domain
5302 * distance is the number of steps a CPU has to iterate down its
5303 * domain tree to share a domain with the other CPU. The farther
5304 * two CPUs are from each other, the larger the distance gets.
5306 * Note that we use the distance only to cache measurement results,
5307 * the distance value is not used numerically otherwise. When two
5308 * CPUs have the same distance it is assumed that the migration
5309 * cost is the same. (this is a simplification but quite practical)
5311 #define MAX_DOMAIN_DISTANCE 32
5313 static unsigned long long migration_cost
[MAX_DOMAIN_DISTANCE
] =
5314 { [ 0 ... MAX_DOMAIN_DISTANCE
-1 ] =
5316 * Architectures may override the migration cost and thus avoid
5317 * boot-time calibration. Unit is nanoseconds. Mostly useful for
5318 * virtualized hardware:
5320 #ifdef CONFIG_DEFAULT_MIGRATION_COST
5321 CONFIG_DEFAULT_MIGRATION_COST
5328 * Allow override of migration cost - in units of microseconds.
5329 * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost
5330 * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs:
5332 static int __init
migration_cost_setup(char *str
)
5334 int ints
[MAX_DOMAIN_DISTANCE
+1], i
;
5336 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5338 printk("#ints: %d\n", ints
[0]);
5339 for (i
= 1; i
<= ints
[0]; i
++) {
5340 migration_cost
[i
-1] = (unsigned long long)ints
[i
]*1000;
5341 printk("migration_cost[%d]: %Ld\n", i
-1, migration_cost
[i
-1]);
5346 __setup ("migration_cost=", migration_cost_setup
);
5349 * Global multiplier (divisor) for migration-cutoff values,
5350 * in percentiles. E.g. use a value of 150 to get 1.5 times
5351 * longer cache-hot cutoff times.
5353 * (We scale it from 100 to 128 to long long handling easier.)
5356 #define MIGRATION_FACTOR_SCALE 128
5358 static unsigned int migration_factor
= MIGRATION_FACTOR_SCALE
;
5360 static int __init
setup_migration_factor(char *str
)
5362 get_option(&str
, &migration_factor
);
5363 migration_factor
= migration_factor
* MIGRATION_FACTOR_SCALE
/ 100;
5367 __setup("migration_factor=", setup_migration_factor
);
5370 * Estimated distance of two CPUs, measured via the number of domains
5371 * we have to pass for the two CPUs to be in the same span:
5373 static unsigned long domain_distance(int cpu1
, int cpu2
)
5375 unsigned long distance
= 0;
5376 struct sched_domain
*sd
;
5378 for_each_domain(cpu1
, sd
) {
5379 WARN_ON(!cpu_isset(cpu1
, sd
->span
));
5380 if (cpu_isset(cpu2
, sd
->span
))
5384 if (distance
>= MAX_DOMAIN_DISTANCE
) {
5386 distance
= MAX_DOMAIN_DISTANCE
-1;
5392 static unsigned int migration_debug
;
5394 static int __init
setup_migration_debug(char *str
)
5396 get_option(&str
, &migration_debug
);
5400 __setup("migration_debug=", setup_migration_debug
);
5403 * Maximum cache-size that the scheduler should try to measure.
5404 * Architectures with larger caches should tune this up during
5405 * bootup. Gets used in the domain-setup code (i.e. during SMP
5408 unsigned int max_cache_size
;
5410 static int __init
setup_max_cache_size(char *str
)
5412 get_option(&str
, &max_cache_size
);
5416 __setup("max_cache_size=", setup_max_cache_size
);
5419 * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This
5420 * is the operation that is timed, so we try to generate unpredictable
5421 * cachemisses that still end up filling the L2 cache:
5423 static void touch_cache(void *__cache
, unsigned long __size
)
5425 unsigned long size
= __size
/sizeof(long), chunk1
= size
/3,
5427 unsigned long *cache
= __cache
;
5430 for (i
= 0; i
< size
/6; i
+= 8) {
5433 case 1: cache
[size
-1-i
]++;
5434 case 2: cache
[chunk1
-i
]++;
5435 case 3: cache
[chunk1
+i
]++;
5436 case 4: cache
[chunk2
-i
]++;
5437 case 5: cache
[chunk2
+i
]++;
5443 * Measure the cache-cost of one task migration. Returns in units of nsec.
5445 static unsigned long long measure_one(void *cache
, unsigned long size
,
5446 int source
, int target
)
5448 cpumask_t mask
, saved_mask
;
5449 unsigned long long t0
, t1
, t2
, t3
, cost
;
5451 saved_mask
= current
->cpus_allowed
;
5454 * Flush source caches to RAM and invalidate them:
5459 * Migrate to the source CPU:
5461 mask
= cpumask_of_cpu(source
);
5462 set_cpus_allowed(current
, mask
);
5463 WARN_ON(smp_processor_id() != source
);
5466 * Dirty the working set:
5469 touch_cache(cache
, size
);
5473 * Migrate to the target CPU, dirty the L2 cache and access
5474 * the shared buffer. (which represents the working set
5475 * of a migrated task.)
5477 mask
= cpumask_of_cpu(target
);
5478 set_cpus_allowed(current
, mask
);
5479 WARN_ON(smp_processor_id() != target
);
5482 touch_cache(cache
, size
);
5485 cost
= t1
-t0
+ t3
-t2
;
5487 if (migration_debug
>= 2)
5488 printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n",
5489 source
, target
, t1
-t0
, t1
-t0
, t3
-t2
, cost
);
5491 * Flush target caches to RAM and invalidate them:
5495 set_cpus_allowed(current
, saved_mask
);
5501 * Measure a series of task migrations and return the average
5502 * result. Since this code runs early during bootup the system
5503 * is 'undisturbed' and the average latency makes sense.
5505 * The algorithm in essence auto-detects the relevant cache-size,
5506 * so it will properly detect different cachesizes for different
5507 * cache-hierarchies, depending on how the CPUs are connected.
5509 * Architectures can prime the upper limit of the search range via
5510 * max_cache_size, otherwise the search range defaults to 20MB...64K.
5512 static unsigned long long
5513 measure_cost(int cpu1
, int cpu2
, void *cache
, unsigned int size
)
5515 unsigned long long cost1
, cost2
;
5519 * Measure the migration cost of 'size' bytes, over an
5520 * average of 10 runs:
5522 * (We perturb the cache size by a small (0..4k)
5523 * value to compensate size/alignment related artifacts.
5524 * We also subtract the cost of the operation done on
5530 * dry run, to make sure we start off cache-cold on cpu1,
5531 * and to get any vmalloc pagefaults in advance:
5533 measure_one(cache
, size
, cpu1
, cpu2
);
5534 for (i
= 0; i
< ITERATIONS
; i
++)
5535 cost1
+= measure_one(cache
, size
- i
*1024, cpu1
, cpu2
);
5537 measure_one(cache
, size
, cpu2
, cpu1
);
5538 for (i
= 0; i
< ITERATIONS
; i
++)
5539 cost1
+= measure_one(cache
, size
- i
*1024, cpu2
, cpu1
);
5542 * (We measure the non-migrating [cached] cost on both
5543 * cpu1 and cpu2, to handle CPUs with different speeds)
5547 measure_one(cache
, size
, cpu1
, cpu1
);
5548 for (i
= 0; i
< ITERATIONS
; i
++)
5549 cost2
+= measure_one(cache
, size
- i
*1024, cpu1
, cpu1
);
5551 measure_one(cache
, size
, cpu2
, cpu2
);
5552 for (i
= 0; i
< ITERATIONS
; i
++)
5553 cost2
+= measure_one(cache
, size
- i
*1024, cpu2
, cpu2
);
5556 * Get the per-iteration migration cost:
5558 do_div(cost1
, 2*ITERATIONS
);
5559 do_div(cost2
, 2*ITERATIONS
);
5561 return cost1
- cost2
;
5564 static unsigned long long measure_migration_cost(int cpu1
, int cpu2
)
5566 unsigned long long max_cost
= 0, fluct
= 0, avg_fluct
= 0;
5567 unsigned int max_size
, size
, size_found
= 0;
5568 long long cost
= 0, prev_cost
;
5572 * Search from max_cache_size*5 down to 64K - the real relevant
5573 * cachesize has to lie somewhere inbetween.
5575 if (max_cache_size
) {
5576 max_size
= max(max_cache_size
* SEARCH_SCOPE
, MIN_CACHE_SIZE
);
5577 size
= max(max_cache_size
/ SEARCH_SCOPE
, MIN_CACHE_SIZE
);
5580 * Since we have no estimation about the relevant
5583 max_size
= DEFAULT_CACHE_SIZE
* SEARCH_SCOPE
;
5584 size
= MIN_CACHE_SIZE
;
5587 if (!cpu_online(cpu1
) || !cpu_online(cpu2
)) {
5588 printk("cpu %d and %d not both online!\n", cpu1
, cpu2
);
5593 * Allocate the working set:
5595 cache
= vmalloc(max_size
);
5597 printk("could not vmalloc %d bytes for cache!\n", 2*max_size
);
5598 return 1000000; // return 1 msec on very small boxen
5601 while (size
<= max_size
) {
5603 cost
= measure_cost(cpu1
, cpu2
, cache
, size
);
5609 if (max_cost
< cost
) {
5615 * Calculate average fluctuation, we use this to prevent
5616 * noise from triggering an early break out of the loop:
5618 fluct
= abs(cost
- prev_cost
);
5619 avg_fluct
= (avg_fluct
+ fluct
)/2;
5621 if (migration_debug
)
5622 printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): (%8Ld %8Ld)\n",
5624 (long)cost
/ 1000000,
5625 ((long)cost
/ 100000) % 10,
5626 (long)max_cost
/ 1000000,
5627 ((long)max_cost
/ 100000) % 10,
5628 domain_distance(cpu1
, cpu2
),
5632 * If we iterated at least 20% past the previous maximum,
5633 * and the cost has dropped by more than 20% already,
5634 * (taking fluctuations into account) then we assume to
5635 * have found the maximum and break out of the loop early:
5637 if (size_found
&& (size
*100 > size_found
*SIZE_THRESH
))
5638 if (cost
+avg_fluct
<= 0 ||
5639 max_cost
*100 > (cost
+avg_fluct
)*COST_THRESH
) {
5641 if (migration_debug
)
5642 printk("-> found max.\n");
5646 * Increase the cachesize in 10% steps:
5648 size
= size
* 10 / 9;
5651 if (migration_debug
)
5652 printk("[%d][%d] working set size found: %d, cost: %Ld\n",
5653 cpu1
, cpu2
, size_found
, max_cost
);
5658 * A task is considered 'cache cold' if at least 2 times
5659 * the worst-case cost of migration has passed.
5661 * (this limit is only listened to if the load-balancing
5662 * situation is 'nice' - if there is a large imbalance we
5663 * ignore it for the sake of CPU utilization and
5664 * processing fairness.)
5666 return 2 * max_cost
* migration_factor
/ MIGRATION_FACTOR_SCALE
;
5669 static void calibrate_migration_costs(const cpumask_t
*cpu_map
)
5671 int cpu1
= -1, cpu2
= -1, cpu
, orig_cpu
= raw_smp_processor_id();
5672 unsigned long j0
, j1
, distance
, max_distance
= 0;
5673 struct sched_domain
*sd
;
5678 * First pass - calculate the cacheflush times:
5680 for_each_cpu_mask(cpu1
, *cpu_map
) {
5681 for_each_cpu_mask(cpu2
, *cpu_map
) {
5684 distance
= domain_distance(cpu1
, cpu2
);
5685 max_distance
= max(max_distance
, distance
);
5687 * No result cached yet?
5689 if (migration_cost
[distance
] == -1LL)
5690 migration_cost
[distance
] =
5691 measure_migration_cost(cpu1
, cpu2
);
5695 * Second pass - update the sched domain hierarchy with
5696 * the new cache-hot-time estimations:
5698 for_each_cpu_mask(cpu
, *cpu_map
) {
5700 for_each_domain(cpu
, sd
) {
5701 sd
->cache_hot_time
= migration_cost
[distance
];
5708 if (migration_debug
)
5709 printk("migration: max_cache_size: %d, cpu: %d MHz:\n",
5717 if (system_state
== SYSTEM_BOOTING
) {
5718 printk("migration_cost=");
5719 for (distance
= 0; distance
<= max_distance
; distance
++) {
5722 printk("%ld", (long)migration_cost
[distance
] / 1000);
5727 if (migration_debug
)
5728 printk("migration: %ld seconds\n", (j1
-j0
)/HZ
);
5731 * Move back to the original CPU. NUMA-Q gets confused
5732 * if we migrate to another quad during bootup.
5734 if (raw_smp_processor_id() != orig_cpu
) {
5735 cpumask_t mask
= cpumask_of_cpu(orig_cpu
),
5736 saved_mask
= current
->cpus_allowed
;
5738 set_cpus_allowed(current
, mask
);
5739 set_cpus_allowed(current
, saved_mask
);
5746 * find_next_best_node - find the next node to include in a sched_domain
5747 * @node: node whose sched_domain we're building
5748 * @used_nodes: nodes already in the sched_domain
5750 * Find the next node to include in a given scheduling domain. Simply
5751 * finds the closest node not already in the @used_nodes map.
5753 * Should use nodemask_t.
5755 static int find_next_best_node(int node
, unsigned long *used_nodes
)
5757 int i
, n
, val
, min_val
, best_node
= 0;
5761 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5762 /* Start at @node */
5763 n
= (node
+ i
) % MAX_NUMNODES
;
5765 if (!nr_cpus_node(n
))
5768 /* Skip already used nodes */
5769 if (test_bit(n
, used_nodes
))
5772 /* Simple min distance search */
5773 val
= node_distance(node
, n
);
5775 if (val
< min_val
) {
5781 set_bit(best_node
, used_nodes
);
5786 * sched_domain_node_span - get a cpumask for a node's sched_domain
5787 * @node: node whose cpumask we're constructing
5788 * @size: number of nodes to include in this span
5790 * Given a node, construct a good cpumask for its sched_domain to span. It
5791 * should be one that prevents unnecessary balancing, but also spreads tasks
5794 static cpumask_t
sched_domain_node_span(int node
)
5797 cpumask_t span
, nodemask
;
5798 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
5801 bitmap_zero(used_nodes
, MAX_NUMNODES
);
5803 nodemask
= node_to_cpumask(node
);
5804 cpus_or(span
, span
, nodemask
);
5805 set_bit(node
, used_nodes
);
5807 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5808 int next_node
= find_next_best_node(node
, used_nodes
);
5809 nodemask
= node_to_cpumask(next_node
);
5810 cpus_or(span
, span
, nodemask
);
5817 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
5819 * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we
5820 * can switch it on easily if needed.
5822 #ifdef CONFIG_SCHED_SMT
5823 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
5824 static struct sched_group sched_group_cpus
[NR_CPUS
];
5825 static int cpu_to_cpu_group(int cpu
)
5831 #ifdef CONFIG_SCHED_MC
5832 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
5833 static struct sched_group
*sched_group_core_bycpu
[NR_CPUS
];
5836 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5837 static int cpu_to_core_group(int cpu
)
5839 return first_cpu(cpu_sibling_map
[cpu
]);
5841 #elif defined(CONFIG_SCHED_MC)
5842 static int cpu_to_core_group(int cpu
)
5848 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
5849 static struct sched_group
*sched_group_phys_bycpu
[NR_CPUS
];
5850 static int cpu_to_phys_group(int cpu
)
5852 #if defined(CONFIG_SCHED_MC)
5853 cpumask_t mask
= cpu_coregroup_map(cpu
);
5854 return first_cpu(mask
);
5855 #elif defined(CONFIG_SCHED_SMT)
5856 return first_cpu(cpu_sibling_map
[cpu
]);
5864 * The init_sched_build_groups can't handle what we want to do with node
5865 * groups, so roll our own. Now each node has its own list of groups which
5866 * gets dynamically allocated.
5868 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
5869 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
5871 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
5872 static struct sched_group
*sched_group_allnodes_bycpu
[NR_CPUS
];
5874 static int cpu_to_allnodes_group(int cpu
)
5876 return cpu_to_node(cpu
);
5878 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
5880 struct sched_group
*sg
= group_head
;
5886 for_each_cpu_mask(j
, sg
->cpumask
) {
5887 struct sched_domain
*sd
;
5889 sd
= &per_cpu(phys_domains
, j
);
5890 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
5892 * Only add "power" once for each
5898 sg
->cpu_power
+= sd
->groups
->cpu_power
;
5901 if (sg
!= group_head
)
5906 /* Free memory allocated for various sched_group structures */
5907 static void free_sched_groups(const cpumask_t
*cpu_map
)
5913 for_each_cpu_mask(cpu
, *cpu_map
) {
5914 struct sched_group
*sched_group_allnodes
5915 = sched_group_allnodes_bycpu
[cpu
];
5916 struct sched_group
**sched_group_nodes
5917 = sched_group_nodes_bycpu
[cpu
];
5919 if (sched_group_allnodes
) {
5920 kfree(sched_group_allnodes
);
5921 sched_group_allnodes_bycpu
[cpu
] = NULL
;
5924 if (!sched_group_nodes
)
5927 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5928 cpumask_t nodemask
= node_to_cpumask(i
);
5929 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
5931 cpus_and(nodemask
, nodemask
, *cpu_map
);
5932 if (cpus_empty(nodemask
))
5942 if (oldsg
!= sched_group_nodes
[i
])
5945 kfree(sched_group_nodes
);
5946 sched_group_nodes_bycpu
[cpu
] = NULL
;
5949 for_each_cpu_mask(cpu
, *cpu_map
) {
5950 if (sched_group_phys_bycpu
[cpu
]) {
5951 kfree(sched_group_phys_bycpu
[cpu
]);
5952 sched_group_phys_bycpu
[cpu
] = NULL
;
5954 #ifdef CONFIG_SCHED_MC
5955 if (sched_group_core_bycpu
[cpu
]) {
5956 kfree(sched_group_core_bycpu
[cpu
]);
5957 sched_group_core_bycpu
[cpu
] = NULL
;
5964 * Build sched domains for a given set of cpus and attach the sched domains
5965 * to the individual cpus
5967 static int build_sched_domains(const cpumask_t
*cpu_map
)
5970 struct sched_group
*sched_group_phys
= NULL
;
5971 #ifdef CONFIG_SCHED_MC
5972 struct sched_group
*sched_group_core
= NULL
;
5975 struct sched_group
**sched_group_nodes
= NULL
;
5976 struct sched_group
*sched_group_allnodes
= NULL
;
5979 * Allocate the per-node list of sched groups
5981 sched_group_nodes
= kzalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
5983 if (!sched_group_nodes
) {
5984 printk(KERN_WARNING
"Can not alloc sched group node list\n");
5987 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
5991 * Set up domains for cpus specified by the cpu_map.
5993 for_each_cpu_mask(i
, *cpu_map
) {
5995 struct sched_domain
*sd
= NULL
, *p
;
5996 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
5998 cpus_and(nodemask
, nodemask
, *cpu_map
);
6001 if (cpus_weight(*cpu_map
)
6002 > SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6003 if (!sched_group_allnodes
) {
6004 sched_group_allnodes
6005 = kmalloc(sizeof(struct sched_group
)
6008 if (!sched_group_allnodes
) {
6010 "Can not alloc allnodes sched group\n");
6013 sched_group_allnodes_bycpu
[i
]
6014 = sched_group_allnodes
;
6016 sd
= &per_cpu(allnodes_domains
, i
);
6017 *sd
= SD_ALLNODES_INIT
;
6018 sd
->span
= *cpu_map
;
6019 group
= cpu_to_allnodes_group(i
);
6020 sd
->groups
= &sched_group_allnodes
[group
];
6025 sd
= &per_cpu(node_domains
, i
);
6027 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6029 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6032 if (!sched_group_phys
) {
6034 = kmalloc(sizeof(struct sched_group
) * NR_CPUS
,
6036 if (!sched_group_phys
) {
6037 printk (KERN_WARNING
"Can not alloc phys sched"
6041 sched_group_phys_bycpu
[i
] = sched_group_phys
;
6045 sd
= &per_cpu(phys_domains
, i
);
6046 group
= cpu_to_phys_group(i
);
6048 sd
->span
= nodemask
;
6050 sd
->groups
= &sched_group_phys
[group
];
6052 #ifdef CONFIG_SCHED_MC
6053 if (!sched_group_core
) {
6055 = kmalloc(sizeof(struct sched_group
) * NR_CPUS
,
6057 if (!sched_group_core
) {
6058 printk (KERN_WARNING
"Can not alloc core sched"
6062 sched_group_core_bycpu
[i
] = sched_group_core
;
6066 sd
= &per_cpu(core_domains
, i
);
6067 group
= cpu_to_core_group(i
);
6069 sd
->span
= cpu_coregroup_map(i
);
6070 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6072 sd
->groups
= &sched_group_core
[group
];
6075 #ifdef CONFIG_SCHED_SMT
6077 sd
= &per_cpu(cpu_domains
, i
);
6078 group
= cpu_to_cpu_group(i
);
6079 *sd
= SD_SIBLING_INIT
;
6080 sd
->span
= cpu_sibling_map
[i
];
6081 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6083 sd
->groups
= &sched_group_cpus
[group
];
6087 #ifdef CONFIG_SCHED_SMT
6088 /* Set up CPU (sibling) groups */
6089 for_each_cpu_mask(i
, *cpu_map
) {
6090 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
6091 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6092 if (i
!= first_cpu(this_sibling_map
))
6095 init_sched_build_groups(sched_group_cpus
, this_sibling_map
,
6100 #ifdef CONFIG_SCHED_MC
6101 /* Set up multi-core groups */
6102 for_each_cpu_mask(i
, *cpu_map
) {
6103 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6104 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6105 if (i
!= first_cpu(this_core_map
))
6107 init_sched_build_groups(sched_group_core
, this_core_map
,
6108 &cpu_to_core_group
);
6113 /* Set up physical groups */
6114 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6115 cpumask_t nodemask
= node_to_cpumask(i
);
6117 cpus_and(nodemask
, nodemask
, *cpu_map
);
6118 if (cpus_empty(nodemask
))
6121 init_sched_build_groups(sched_group_phys
, nodemask
,
6122 &cpu_to_phys_group
);
6126 /* Set up node groups */
6127 if (sched_group_allnodes
)
6128 init_sched_build_groups(sched_group_allnodes
, *cpu_map
,
6129 &cpu_to_allnodes_group
);
6131 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6132 /* Set up node groups */
6133 struct sched_group
*sg
, *prev
;
6134 cpumask_t nodemask
= node_to_cpumask(i
);
6135 cpumask_t domainspan
;
6136 cpumask_t covered
= CPU_MASK_NONE
;
6139 cpus_and(nodemask
, nodemask
, *cpu_map
);
6140 if (cpus_empty(nodemask
)) {
6141 sched_group_nodes
[i
] = NULL
;
6145 domainspan
= sched_domain_node_span(i
);
6146 cpus_and(domainspan
, domainspan
, *cpu_map
);
6148 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6150 printk(KERN_WARNING
"Can not alloc domain group for "
6154 sched_group_nodes
[i
] = sg
;
6155 for_each_cpu_mask(j
, nodemask
) {
6156 struct sched_domain
*sd
;
6157 sd
= &per_cpu(node_domains
, j
);
6161 sg
->cpumask
= nodemask
;
6163 cpus_or(covered
, covered
, nodemask
);
6166 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6167 cpumask_t tmp
, notcovered
;
6168 int n
= (i
+ j
) % MAX_NUMNODES
;
6170 cpus_complement(notcovered
, covered
);
6171 cpus_and(tmp
, notcovered
, *cpu_map
);
6172 cpus_and(tmp
, tmp
, domainspan
);
6173 if (cpus_empty(tmp
))
6176 nodemask
= node_to_cpumask(n
);
6177 cpus_and(tmp
, tmp
, nodemask
);
6178 if (cpus_empty(tmp
))
6181 sg
= kmalloc_node(sizeof(struct sched_group
),
6185 "Can not alloc domain group for node %d\n", j
);
6190 sg
->next
= prev
->next
;
6191 cpus_or(covered
, covered
, tmp
);
6198 /* Calculate CPU power for physical packages and nodes */
6199 #ifdef CONFIG_SCHED_SMT
6200 for_each_cpu_mask(i
, *cpu_map
) {
6201 struct sched_domain
*sd
;
6202 sd
= &per_cpu(cpu_domains
, i
);
6203 sd
->groups
->cpu_power
= SCHED_LOAD_SCALE
;
6206 #ifdef CONFIG_SCHED_MC
6207 for_each_cpu_mask(i
, *cpu_map
) {
6209 struct sched_domain
*sd
;
6210 sd
= &per_cpu(core_domains
, i
);
6211 if (sched_smt_power_savings
)
6212 power
= SCHED_LOAD_SCALE
* cpus_weight(sd
->groups
->cpumask
);
6214 power
= SCHED_LOAD_SCALE
+ (cpus_weight(sd
->groups
->cpumask
)-1)
6215 * SCHED_LOAD_SCALE
/ 10;
6216 sd
->groups
->cpu_power
= power
;
6220 for_each_cpu_mask(i
, *cpu_map
) {
6221 struct sched_domain
*sd
;
6222 #ifdef CONFIG_SCHED_MC
6223 sd
= &per_cpu(phys_domains
, i
);
6224 if (i
!= first_cpu(sd
->groups
->cpumask
))
6227 sd
->groups
->cpu_power
= 0;
6228 if (sched_mc_power_savings
|| sched_smt_power_savings
) {
6231 for_each_cpu_mask(j
, sd
->groups
->cpumask
) {
6232 struct sched_domain
*sd1
;
6233 sd1
= &per_cpu(core_domains
, j
);
6235 * for each core we will add once
6236 * to the group in physical domain
6238 if (j
!= first_cpu(sd1
->groups
->cpumask
))
6241 if (sched_smt_power_savings
)
6242 sd
->groups
->cpu_power
+= sd1
->groups
->cpu_power
;
6244 sd
->groups
->cpu_power
+= SCHED_LOAD_SCALE
;
6248 * This has to be < 2 * SCHED_LOAD_SCALE
6249 * Lets keep it SCHED_LOAD_SCALE, so that
6250 * while calculating NUMA group's cpu_power
6252 * numa_group->cpu_power += phys_group->cpu_power;
6254 * See "only add power once for each physical pkg"
6257 sd
->groups
->cpu_power
= SCHED_LOAD_SCALE
;
6260 sd
= &per_cpu(phys_domains
, i
);
6261 if (sched_smt_power_savings
)
6262 power
= SCHED_LOAD_SCALE
* cpus_weight(sd
->groups
->cpumask
);
6264 power
= SCHED_LOAD_SCALE
;
6265 sd
->groups
->cpu_power
= power
;
6270 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6271 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6273 init_numa_sched_groups_power(sched_group_allnodes
);
6276 /* Attach the domains */
6277 for_each_cpu_mask(i
, *cpu_map
) {
6278 struct sched_domain
*sd
;
6279 #ifdef CONFIG_SCHED_SMT
6280 sd
= &per_cpu(cpu_domains
, i
);
6281 #elif defined(CONFIG_SCHED_MC)
6282 sd
= &per_cpu(core_domains
, i
);
6284 sd
= &per_cpu(phys_domains
, i
);
6286 cpu_attach_domain(sd
, i
);
6289 * Tune cache-hot values:
6291 calibrate_migration_costs(cpu_map
);
6296 free_sched_groups(cpu_map
);
6300 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6302 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6304 cpumask_t cpu_default_map
;
6308 * Setup mask for cpus without special case scheduling requirements.
6309 * For now this just excludes isolated cpus, but could be used to
6310 * exclude other special cases in the future.
6312 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
6314 err
= build_sched_domains(&cpu_default_map
);
6319 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6321 free_sched_groups(cpu_map
);
6325 * Detach sched domains from a group of cpus specified in cpu_map
6326 * These cpus will now be attached to the NULL domain
6328 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6332 for_each_cpu_mask(i
, *cpu_map
)
6333 cpu_attach_domain(NULL
, i
);
6334 synchronize_sched();
6335 arch_destroy_sched_domains(cpu_map
);
6339 * Partition sched domains as specified by the cpumasks below.
6340 * This attaches all cpus from the cpumasks to the NULL domain,
6341 * waits for a RCU quiescent period, recalculates sched
6342 * domain information and then attaches them back to the
6343 * correct sched domains
6344 * Call with hotplug lock held
6346 int partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
6348 cpumask_t change_map
;
6351 cpus_and(*partition1
, *partition1
, cpu_online_map
);
6352 cpus_and(*partition2
, *partition2
, cpu_online_map
);
6353 cpus_or(change_map
, *partition1
, *partition2
);
6355 /* Detach sched domains from all of the affected cpus */
6356 detach_destroy_domains(&change_map
);
6357 if (!cpus_empty(*partition1
))
6358 err
= build_sched_domains(partition1
);
6359 if (!err
&& !cpus_empty(*partition2
))
6360 err
= build_sched_domains(partition2
);
6365 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6366 int arch_reinit_sched_domains(void)
6371 detach_destroy_domains(&cpu_online_map
);
6372 err
= arch_init_sched_domains(&cpu_online_map
);
6373 unlock_cpu_hotplug();
6378 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6382 if (buf
[0] != '0' && buf
[0] != '1')
6386 sched_smt_power_savings
= (buf
[0] == '1');
6388 sched_mc_power_savings
= (buf
[0] == '1');
6390 ret
= arch_reinit_sched_domains();
6392 return ret
? ret
: count
;
6395 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6398 #ifdef CONFIG_SCHED_SMT
6400 err
= sysfs_create_file(&cls
->kset
.kobj
,
6401 &attr_sched_smt_power_savings
.attr
);
6403 #ifdef CONFIG_SCHED_MC
6404 if (!err
&& mc_capable())
6405 err
= sysfs_create_file(&cls
->kset
.kobj
,
6406 &attr_sched_mc_power_savings
.attr
);
6412 #ifdef CONFIG_SCHED_MC
6413 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6415 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6417 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
, const char *buf
, size_t count
)
6419 return sched_power_savings_store(buf
, count
, 0);
6421 SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6422 sched_mc_power_savings_store
);
6425 #ifdef CONFIG_SCHED_SMT
6426 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6428 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6430 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
, const char *buf
, size_t count
)
6432 return sched_power_savings_store(buf
, count
, 1);
6434 SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6435 sched_smt_power_savings_store
);
6439 #ifdef CONFIG_HOTPLUG_CPU
6441 * Force a reinitialization of the sched domains hierarchy. The domains
6442 * and groups cannot be updated in place without racing with the balancing
6443 * code, so we temporarily attach all running cpus to the NULL domain
6444 * which will prevent rebalancing while the sched domains are recalculated.
6446 static int update_sched_domains(struct notifier_block
*nfb
,
6447 unsigned long action
, void *hcpu
)
6450 case CPU_UP_PREPARE
:
6451 case CPU_DOWN_PREPARE
:
6452 detach_destroy_domains(&cpu_online_map
);
6455 case CPU_UP_CANCELED
:
6456 case CPU_DOWN_FAILED
:
6460 * Fall through and re-initialise the domains.
6467 /* The hotplug lock is already held by cpu_up/cpu_down */
6468 arch_init_sched_domains(&cpu_online_map
);
6474 void __init
sched_init_smp(void)
6477 arch_init_sched_domains(&cpu_online_map
);
6478 unlock_cpu_hotplug();
6479 /* XXX: Theoretical race here - CPU may be hotplugged now */
6480 hotcpu_notifier(update_sched_domains
, 0);
6483 void __init
sched_init_smp(void)
6486 #endif /* CONFIG_SMP */
6488 int in_sched_functions(unsigned long addr
)
6490 /* Linker adds these: start and end of __sched functions */
6491 extern char __sched_text_start
[], __sched_text_end
[];
6492 return in_lock_functions(addr
) ||
6493 (addr
>= (unsigned long)__sched_text_start
6494 && addr
< (unsigned long)__sched_text_end
);
6497 void __init
sched_init(void)
6502 for_each_possible_cpu(i
) {
6503 prio_array_t
*array
;
6506 spin_lock_init(&rq
->lock
);
6508 rq
->active
= rq
->arrays
;
6509 rq
->expired
= rq
->arrays
+ 1;
6510 rq
->best_expired_prio
= MAX_PRIO
;
6514 for (j
= 1; j
< 3; j
++)
6515 rq
->cpu_load
[j
] = 0;
6516 rq
->active_balance
= 0;
6518 rq
->migration_thread
= NULL
;
6519 INIT_LIST_HEAD(&rq
->migration_queue
);
6521 atomic_set(&rq
->nr_iowait
, 0);
6523 for (j
= 0; j
< 2; j
++) {
6524 array
= rq
->arrays
+ j
;
6525 for (k
= 0; k
< MAX_PRIO
; k
++) {
6526 INIT_LIST_HEAD(array
->queue
+ k
);
6527 __clear_bit(k
, array
->bitmap
);
6529 // delimiter for bitsearch
6530 __set_bit(MAX_PRIO
, array
->bitmap
);
6534 set_load_weight(&init_task
);
6536 * The boot idle thread does lazy MMU switching as well:
6538 atomic_inc(&init_mm
.mm_count
);
6539 enter_lazy_tlb(&init_mm
, current
);
6542 * Make us the idle thread. Technically, schedule() should not be
6543 * called from this thread, however somewhere below it might be,
6544 * but because we are the idle thread, we just pick up running again
6545 * when this runqueue becomes "idle".
6547 init_idle(current
, smp_processor_id());
6550 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6551 void __might_sleep(char *file
, int line
)
6553 #if defined(in_atomic)
6554 static unsigned long prev_jiffy
; /* ratelimiting */
6556 if ((in_atomic() || irqs_disabled()) &&
6557 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6558 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6560 prev_jiffy
= jiffies
;
6561 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6562 " context at %s:%d\n", file
, line
);
6563 printk("in_atomic():%d, irqs_disabled():%d\n",
6564 in_atomic(), irqs_disabled());
6569 EXPORT_SYMBOL(__might_sleep
);
6572 #ifdef CONFIG_MAGIC_SYSRQ
6573 void normalize_rt_tasks(void)
6575 struct task_struct
*p
;
6576 prio_array_t
*array
;
6577 unsigned long flags
;
6580 read_lock_irq(&tasklist_lock
);
6581 for_each_process(p
) {
6585 rq
= task_rq_lock(p
, &flags
);
6589 deactivate_task(p
, task_rq(p
));
6590 __setscheduler(p
, SCHED_NORMAL
, 0);
6592 __activate_task(p
, task_rq(p
));
6593 resched_task(rq
->curr
);
6596 task_rq_unlock(rq
, &flags
);
6598 read_unlock_irq(&tasklist_lock
);
6601 #endif /* CONFIG_MAGIC_SYSRQ */
6605 * These functions are only useful for the IA64 MCA handling.
6607 * They can only be called when the whole system has been
6608 * stopped - every CPU needs to be quiescent, and no scheduling
6609 * activity can take place. Using them for anything else would
6610 * be a serious bug, and as a result, they aren't even visible
6611 * under any other configuration.
6615 * curr_task - return the current task for a given cpu.
6616 * @cpu: the processor in question.
6618 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6620 task_t
*curr_task(int cpu
)
6622 return cpu_curr(cpu
);
6626 * set_curr_task - set the current task for a given cpu.
6627 * @cpu: the processor in question.
6628 * @p: the task pointer to set.
6630 * Description: This function must only be used when non-maskable interrupts
6631 * are serviced on a separate stack. It allows the architecture to switch the
6632 * notion of the current task on a cpu in a non-blocking manner. This function
6633 * must be called with all CPU's synchronized, and interrupts disabled, the
6634 * and caller must save the original value of the current task (see
6635 * curr_task() above) and restore that value before reenabling interrupts and
6636 * re-starting the system.
6638 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6640 void set_curr_task(int cpu
, task_t
*p
)