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/debug_locks.h>
34 #include <linux/security.h>
35 #include <linux/notifier.h>
36 #include <linux/profile.h>
37 #include <linux/freezer.h>
38 #include <linux/vmalloc.h>
39 #include <linux/blkdev.h>
40 #include <linux/delay.h>
41 #include <linux/smp.h>
42 #include <linux/threads.h>
43 #include <linux/timer.h>
44 #include <linux/rcupdate.h>
45 #include <linux/cpu.h>
46 #include <linux/cpuset.h>
47 #include <linux/percpu.h>
48 #include <linux/kthread.h>
49 #include <linux/seq_file.h>
50 #include <linux/syscalls.h>
51 #include <linux/times.h>
52 #include <linux/tsacct_kern.h>
53 #include <linux/kprobes.h>
54 #include <linux/delayacct.h>
57 #include <asm/unistd.h>
60 * Convert user-nice values [ -20 ... 0 ... 19 ]
61 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
64 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
65 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
66 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
69 * 'User priority' is the nice value converted to something we
70 * can work with better when scaling various scheduler parameters,
71 * it's a [ 0 ... 39 ] range.
73 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
74 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
75 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
78 * Some helpers for converting nanosecond timing to jiffy resolution
80 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
81 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
84 * These are the 'tuning knobs' of the scheduler:
86 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
87 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
88 * Timeslices get refilled after they expire.
90 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
91 #define DEF_TIMESLICE (100 * HZ / 1000)
92 #define ON_RUNQUEUE_WEIGHT 30
93 #define CHILD_PENALTY 95
94 #define PARENT_PENALTY 100
96 #define PRIO_BONUS_RATIO 25
97 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
98 #define INTERACTIVE_DELTA 2
99 #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
100 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
101 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
104 * If a task is 'interactive' then we reinsert it in the active
105 * array after it has expired its current timeslice. (it will not
106 * continue to run immediately, it will still roundrobin with
107 * other interactive tasks.)
109 * This part scales the interactivity limit depending on niceness.
111 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
112 * Here are a few examples of different nice levels:
114 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
115 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
116 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
117 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
118 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
120 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
121 * priority range a task can explore, a value of '1' means the
122 * task is rated interactive.)
124 * Ie. nice +19 tasks can never get 'interactive' enough to be
125 * reinserted into the active array. And only heavily CPU-hog nice -20
126 * tasks will be expired. Default nice 0 tasks are somewhere between,
127 * it takes some effort for them to get interactive, but it's not
131 #define CURRENT_BONUS(p) \
132 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
135 #define GRANULARITY (10 * HZ / 1000 ? : 1)
138 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
139 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
142 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
143 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
146 #define SCALE(v1,v1_max,v2_max) \
147 (v1) * (v2_max) / (v1_max)
150 (SCALE(TASK_NICE(p) + 20, 40, MAX_BONUS) - 20 * MAX_BONUS / 40 + \
153 #define TASK_INTERACTIVE(p) \
154 ((p)->prio <= (p)->static_prio - DELTA(p))
156 #define INTERACTIVE_SLEEP(p) \
157 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
158 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
160 #define TASK_PREEMPTS_CURR(p, rq) \
161 ((p)->prio < (rq)->curr->prio)
163 #define SCALE_PRIO(x, prio) \
164 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
166 static unsigned int static_prio_timeslice(int static_prio
)
168 if (static_prio
< NICE_TO_PRIO(0))
169 return SCALE_PRIO(DEF_TIMESLICE
* 4, static_prio
);
171 return SCALE_PRIO(DEF_TIMESLICE
, static_prio
);
175 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
176 * to time slice values: [800ms ... 100ms ... 5ms]
178 * The higher a thread's priority, the bigger timeslices
179 * it gets during one round of execution. But even the lowest
180 * priority thread gets MIN_TIMESLICE worth of execution time.
183 static inline unsigned int task_timeslice(struct task_struct
*p
)
185 return static_prio_timeslice(p
->static_prio
);
189 * These are the runqueue data structures:
193 unsigned int nr_active
;
194 DECLARE_BITMAP(bitmap
, MAX_PRIO
+1); /* include 1 bit for delimiter */
195 struct list_head queue
[MAX_PRIO
];
199 * This is the main, per-CPU runqueue data structure.
201 * Locking rule: those places that want to lock multiple runqueues
202 * (such as the load balancing or the thread migration code), lock
203 * acquire operations must be ordered by ascending &runqueue.
209 * nr_running and cpu_load should be in the same cacheline because
210 * remote CPUs use both these fields when doing load calculation.
212 unsigned long nr_running
;
213 unsigned long raw_weighted_load
;
215 unsigned long cpu_load
[3];
217 unsigned long long nr_switches
;
220 * This is part of a global counter where only the total sum
221 * over all CPUs matters. A task can increase this counter on
222 * one CPU and if it got migrated afterwards it may decrease
223 * it on another CPU. Always updated under the runqueue lock:
225 unsigned long nr_uninterruptible
;
227 unsigned long expired_timestamp
;
228 unsigned long long timestamp_last_tick
;
229 struct task_struct
*curr
, *idle
;
230 struct mm_struct
*prev_mm
;
231 struct prio_array
*active
, *expired
, arrays
[2];
232 int best_expired_prio
;
236 struct sched_domain
*sd
;
238 /* For active balancing */
241 int cpu
; /* cpu of this runqueue */
243 struct task_struct
*migration_thread
;
244 struct list_head migration_queue
;
247 #ifdef CONFIG_SCHEDSTATS
249 struct sched_info rq_sched_info
;
251 /* sys_sched_yield() stats */
252 unsigned long yld_exp_empty
;
253 unsigned long yld_act_empty
;
254 unsigned long yld_both_empty
;
255 unsigned long yld_cnt
;
257 /* schedule() stats */
258 unsigned long sched_switch
;
259 unsigned long sched_cnt
;
260 unsigned long sched_goidle
;
262 /* try_to_wake_up() stats */
263 unsigned long ttwu_cnt
;
264 unsigned long ttwu_local
;
266 struct lock_class_key rq_lock_key
;
269 static DEFINE_PER_CPU(struct rq
, runqueues
);
271 static inline int cpu_of(struct rq
*rq
)
281 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
282 * See detach_destroy_domains: synchronize_sched for details.
284 * The domain tree of any CPU may only be accessed from within
285 * preempt-disabled sections.
287 #define for_each_domain(cpu, __sd) \
288 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
290 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
291 #define this_rq() (&__get_cpu_var(runqueues))
292 #define task_rq(p) cpu_rq(task_cpu(p))
293 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
295 #ifndef prepare_arch_switch
296 # define prepare_arch_switch(next) do { } while (0)
298 #ifndef finish_arch_switch
299 # define finish_arch_switch(prev) do { } while (0)
302 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
303 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
305 return rq
->curr
== p
;
308 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
312 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
314 #ifdef CONFIG_DEBUG_SPINLOCK
315 /* this is a valid case when another task releases the spinlock */
316 rq
->lock
.owner
= current
;
319 * If we are tracking spinlock dependencies then we have to
320 * fix up the runqueue lock - which gets 'carried over' from
323 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
325 spin_unlock_irq(&rq
->lock
);
328 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
329 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
334 return rq
->curr
== p
;
338 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
342 * We can optimise this out completely for !SMP, because the
343 * SMP rebalancing from interrupt is the only thing that cares
348 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
349 spin_unlock_irq(&rq
->lock
);
351 spin_unlock(&rq
->lock
);
355 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
359 * After ->oncpu is cleared, the task can be moved to a different CPU.
360 * We must ensure this doesn't happen until the switch is completely
366 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
370 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
373 * __task_rq_lock - lock the runqueue a given task resides on.
374 * Must be called interrupts disabled.
376 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
383 spin_lock(&rq
->lock
);
384 if (unlikely(rq
!= task_rq(p
))) {
385 spin_unlock(&rq
->lock
);
386 goto repeat_lock_task
;
392 * task_rq_lock - lock the runqueue a given task resides on and disable
393 * interrupts. Note the ordering: we can safely lookup the task_rq without
394 * explicitly disabling preemption.
396 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
402 local_irq_save(*flags
);
404 spin_lock(&rq
->lock
);
405 if (unlikely(rq
!= task_rq(p
))) {
406 spin_unlock_irqrestore(&rq
->lock
, *flags
);
407 goto repeat_lock_task
;
412 static inline void __task_rq_unlock(struct rq
*rq
)
415 spin_unlock(&rq
->lock
);
418 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
421 spin_unlock_irqrestore(&rq
->lock
, *flags
);
424 #ifdef CONFIG_SCHEDSTATS
426 * bump this up when changing the output format or the meaning of an existing
427 * format, so that tools can adapt (or abort)
429 #define SCHEDSTAT_VERSION 12
431 static int show_schedstat(struct seq_file
*seq
, void *v
)
435 seq_printf(seq
, "version %d\n", SCHEDSTAT_VERSION
);
436 seq_printf(seq
, "timestamp %lu\n", jiffies
);
437 for_each_online_cpu(cpu
) {
438 struct rq
*rq
= cpu_rq(cpu
);
440 struct sched_domain
*sd
;
444 /* runqueue-specific stats */
446 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
447 cpu
, rq
->yld_both_empty
,
448 rq
->yld_act_empty
, rq
->yld_exp_empty
, rq
->yld_cnt
,
449 rq
->sched_switch
, rq
->sched_cnt
, rq
->sched_goidle
,
450 rq
->ttwu_cnt
, rq
->ttwu_local
,
451 rq
->rq_sched_info
.cpu_time
,
452 rq
->rq_sched_info
.run_delay
, rq
->rq_sched_info
.pcnt
);
454 seq_printf(seq
, "\n");
457 /* domain-specific stats */
459 for_each_domain(cpu
, sd
) {
460 enum idle_type itype
;
461 char mask_str
[NR_CPUS
];
463 cpumask_scnprintf(mask_str
, NR_CPUS
, sd
->span
);
464 seq_printf(seq
, "domain%d %s", dcnt
++, mask_str
);
465 for (itype
= SCHED_IDLE
; itype
< MAX_IDLE_TYPES
;
467 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu %lu",
469 sd
->lb_balanced
[itype
],
470 sd
->lb_failed
[itype
],
471 sd
->lb_imbalance
[itype
],
472 sd
->lb_gained
[itype
],
473 sd
->lb_hot_gained
[itype
],
474 sd
->lb_nobusyq
[itype
],
475 sd
->lb_nobusyg
[itype
]);
477 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
478 sd
->alb_cnt
, sd
->alb_failed
, sd
->alb_pushed
,
479 sd
->sbe_cnt
, sd
->sbe_balanced
, sd
->sbe_pushed
,
480 sd
->sbf_cnt
, sd
->sbf_balanced
, sd
->sbf_pushed
,
481 sd
->ttwu_wake_remote
, sd
->ttwu_move_affine
, sd
->ttwu_move_balance
);
489 static int schedstat_open(struct inode
*inode
, struct file
*file
)
491 unsigned int size
= PAGE_SIZE
* (1 + num_online_cpus() / 32);
492 char *buf
= kmalloc(size
, GFP_KERNEL
);
498 res
= single_open(file
, show_schedstat
, NULL
);
500 m
= file
->private_data
;
508 const struct file_operations proc_schedstat_operations
= {
509 .open
= schedstat_open
,
512 .release
= single_release
,
516 * Expects runqueue lock to be held for atomicity of update
519 rq_sched_info_arrive(struct rq
*rq
, unsigned long delta_jiffies
)
522 rq
->rq_sched_info
.run_delay
+= delta_jiffies
;
523 rq
->rq_sched_info
.pcnt
++;
528 * Expects runqueue lock to be held for atomicity of update
531 rq_sched_info_depart(struct rq
*rq
, unsigned long delta_jiffies
)
534 rq
->rq_sched_info
.cpu_time
+= delta_jiffies
;
536 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
537 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
538 #else /* !CONFIG_SCHEDSTATS */
540 rq_sched_info_arrive(struct rq
*rq
, unsigned long delta_jiffies
)
543 rq_sched_info_depart(struct rq
*rq
, unsigned long delta_jiffies
)
545 # define schedstat_inc(rq, field) do { } while (0)
546 # define schedstat_add(rq, field, amt) do { } while (0)
550 * this_rq_lock - lock this runqueue and disable interrupts.
552 static inline struct rq
*this_rq_lock(void)
559 spin_lock(&rq
->lock
);
564 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
566 * Called when a process is dequeued from the active array and given
567 * the cpu. We should note that with the exception of interactive
568 * tasks, the expired queue will become the active queue after the active
569 * queue is empty, without explicitly dequeuing and requeuing tasks in the
570 * expired queue. (Interactive tasks may be requeued directly to the
571 * active queue, thus delaying tasks in the expired queue from running;
572 * see scheduler_tick()).
574 * This function is only called from sched_info_arrive(), rather than
575 * dequeue_task(). Even though a task may be queued and dequeued multiple
576 * times as it is shuffled about, we're really interested in knowing how
577 * long it was from the *first* time it was queued to the time that it
580 static inline void sched_info_dequeued(struct task_struct
*t
)
582 t
->sched_info
.last_queued
= 0;
586 * Called when a task finally hits the cpu. We can now calculate how
587 * long it was waiting to run. We also note when it began so that we
588 * can keep stats on how long its timeslice is.
590 static void sched_info_arrive(struct task_struct
*t
)
592 unsigned long now
= jiffies
, delta_jiffies
= 0;
594 if (t
->sched_info
.last_queued
)
595 delta_jiffies
= now
- t
->sched_info
.last_queued
;
596 sched_info_dequeued(t
);
597 t
->sched_info
.run_delay
+= delta_jiffies
;
598 t
->sched_info
.last_arrival
= now
;
599 t
->sched_info
.pcnt
++;
601 rq_sched_info_arrive(task_rq(t
), delta_jiffies
);
605 * Called when a process is queued into either the active or expired
606 * array. The time is noted and later used to determine how long we
607 * had to wait for us to reach the cpu. Since the expired queue will
608 * become the active queue after active queue is empty, without dequeuing
609 * and requeuing any tasks, we are interested in queuing to either. It
610 * is unusual but not impossible for tasks to be dequeued and immediately
611 * requeued in the same or another array: this can happen in sched_yield(),
612 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
615 * This function is only called from enqueue_task(), but also only updates
616 * the timestamp if it is already not set. It's assumed that
617 * sched_info_dequeued() will clear that stamp when appropriate.
619 static inline void sched_info_queued(struct task_struct
*t
)
621 if (unlikely(sched_info_on()))
622 if (!t
->sched_info
.last_queued
)
623 t
->sched_info
.last_queued
= jiffies
;
627 * Called when a process ceases being the active-running process, either
628 * voluntarily or involuntarily. Now we can calculate how long we ran.
630 static inline void sched_info_depart(struct task_struct
*t
)
632 unsigned long delta_jiffies
= jiffies
- t
->sched_info
.last_arrival
;
634 t
->sched_info
.cpu_time
+= delta_jiffies
;
635 rq_sched_info_depart(task_rq(t
), delta_jiffies
);
639 * Called when tasks are switched involuntarily due, typically, to expiring
640 * their time slice. (This may also be called when switching to or from
641 * the idle task.) We are only called when prev != next.
644 __sched_info_switch(struct task_struct
*prev
, struct task_struct
*next
)
646 struct rq
*rq
= task_rq(prev
);
649 * prev now departs the cpu. It's not interesting to record
650 * stats about how efficient we were at scheduling the idle
653 if (prev
!= rq
->idle
)
654 sched_info_depart(prev
);
656 if (next
!= rq
->idle
)
657 sched_info_arrive(next
);
660 sched_info_switch(struct task_struct
*prev
, struct task_struct
*next
)
662 if (unlikely(sched_info_on()))
663 __sched_info_switch(prev
, next
);
666 #define sched_info_queued(t) do { } while (0)
667 #define sched_info_switch(t, next) do { } while (0)
668 #endif /* CONFIG_SCHEDSTATS || CONFIG_TASK_DELAY_ACCT */
671 * Adding/removing a task to/from a priority array:
673 static void dequeue_task(struct task_struct
*p
, struct prio_array
*array
)
676 list_del(&p
->run_list
);
677 if (list_empty(array
->queue
+ p
->prio
))
678 __clear_bit(p
->prio
, array
->bitmap
);
681 static void enqueue_task(struct task_struct
*p
, struct prio_array
*array
)
683 sched_info_queued(p
);
684 list_add_tail(&p
->run_list
, array
->queue
+ p
->prio
);
685 __set_bit(p
->prio
, array
->bitmap
);
691 * Put task to the end of the run list without the overhead of dequeue
692 * followed by enqueue.
694 static void requeue_task(struct task_struct
*p
, struct prio_array
*array
)
696 list_move_tail(&p
->run_list
, array
->queue
+ p
->prio
);
700 enqueue_task_head(struct task_struct
*p
, struct prio_array
*array
)
702 list_add(&p
->run_list
, array
->queue
+ p
->prio
);
703 __set_bit(p
->prio
, array
->bitmap
);
709 * __normal_prio - return the priority that is based on the static
710 * priority but is modified by bonuses/penalties.
712 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
713 * into the -5 ... 0 ... +5 bonus/penalty range.
715 * We use 25% of the full 0...39 priority range so that:
717 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
718 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
720 * Both properties are important to certain workloads.
723 static inline int __normal_prio(struct task_struct
*p
)
727 bonus
= CURRENT_BONUS(p
) - MAX_BONUS
/ 2;
729 prio
= p
->static_prio
- bonus
;
730 if (prio
< MAX_RT_PRIO
)
732 if (prio
> MAX_PRIO
-1)
738 * To aid in avoiding the subversion of "niceness" due to uneven distribution
739 * of tasks with abnormal "nice" values across CPUs the contribution that
740 * each task makes to its run queue's load is weighted according to its
741 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
742 * scaled version of the new time slice allocation that they receive on time
747 * Assume: static_prio_timeslice(NICE_TO_PRIO(0)) == DEF_TIMESLICE
748 * If static_prio_timeslice() is ever changed to break this assumption then
749 * this code will need modification
751 #define TIME_SLICE_NICE_ZERO DEF_TIMESLICE
752 #define LOAD_WEIGHT(lp) \
753 (((lp) * SCHED_LOAD_SCALE) / TIME_SLICE_NICE_ZERO)
754 #define PRIO_TO_LOAD_WEIGHT(prio) \
755 LOAD_WEIGHT(static_prio_timeslice(prio))
756 #define RTPRIO_TO_LOAD_WEIGHT(rp) \
757 (PRIO_TO_LOAD_WEIGHT(MAX_RT_PRIO) + LOAD_WEIGHT(rp))
759 static void set_load_weight(struct task_struct
*p
)
761 if (has_rt_policy(p
)) {
763 if (p
== task_rq(p
)->migration_thread
)
765 * The migration thread does the actual balancing.
766 * Giving its load any weight will skew balancing
772 p
->load_weight
= RTPRIO_TO_LOAD_WEIGHT(p
->rt_priority
);
774 p
->load_weight
= PRIO_TO_LOAD_WEIGHT(p
->static_prio
);
778 inc_raw_weighted_load(struct rq
*rq
, const struct task_struct
*p
)
780 rq
->raw_weighted_load
+= p
->load_weight
;
784 dec_raw_weighted_load(struct rq
*rq
, const struct task_struct
*p
)
786 rq
->raw_weighted_load
-= p
->load_weight
;
789 static inline void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
792 inc_raw_weighted_load(rq
, p
);
795 static inline void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
798 dec_raw_weighted_load(rq
, p
);
802 * Calculate the expected normal priority: i.e. priority
803 * without taking RT-inheritance into account. Might be
804 * boosted by interactivity modifiers. Changes upon fork,
805 * setprio syscalls, and whenever the interactivity
806 * estimator recalculates.
808 static inline int normal_prio(struct task_struct
*p
)
812 if (has_rt_policy(p
))
813 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
815 prio
= __normal_prio(p
);
820 * Calculate the current priority, i.e. the priority
821 * taken into account by the scheduler. This value might
822 * be boosted by RT tasks, or might be boosted by
823 * interactivity modifiers. Will be RT if the task got
824 * RT-boosted. If not then it returns p->normal_prio.
826 static int effective_prio(struct task_struct
*p
)
828 p
->normal_prio
= normal_prio(p
);
830 * If we are RT tasks or we were boosted to RT priority,
831 * keep the priority unchanged. Otherwise, update priority
832 * to the normal priority:
834 if (!rt_prio(p
->prio
))
835 return p
->normal_prio
;
840 * __activate_task - move a task to the runqueue.
842 static void __activate_task(struct task_struct
*p
, struct rq
*rq
)
844 struct prio_array
*target
= rq
->active
;
847 target
= rq
->expired
;
848 enqueue_task(p
, target
);
849 inc_nr_running(p
, rq
);
853 * __activate_idle_task - move idle task to the _front_ of runqueue.
855 static inline void __activate_idle_task(struct task_struct
*p
, struct rq
*rq
)
857 enqueue_task_head(p
, rq
->active
);
858 inc_nr_running(p
, rq
);
862 * Recalculate p->normal_prio and p->prio after having slept,
863 * updating the sleep-average too:
865 static int recalc_task_prio(struct task_struct
*p
, unsigned long long now
)
867 /* Caller must always ensure 'now >= p->timestamp' */
868 unsigned long sleep_time
= now
- p
->timestamp
;
873 if (likely(sleep_time
> 0)) {
875 * This ceiling is set to the lowest priority that would allow
876 * a task to be reinserted into the active array on timeslice
879 unsigned long ceiling
= INTERACTIVE_SLEEP(p
);
881 if (p
->mm
&& sleep_time
> ceiling
&& p
->sleep_avg
< ceiling
) {
883 * Prevents user tasks from achieving best priority
884 * with one single large enough sleep.
886 p
->sleep_avg
= ceiling
;
888 * Using INTERACTIVE_SLEEP() as a ceiling places a
889 * nice(0) task 1ms sleep away from promotion, and
890 * gives it 700ms to round-robin with no chance of
891 * being demoted. This is more than generous, so
892 * mark this sleep as non-interactive to prevent the
893 * on-runqueue bonus logic from intervening should
894 * this task not receive cpu immediately.
896 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
899 * Tasks waking from uninterruptible sleep are
900 * limited in their sleep_avg rise as they
901 * are likely to be waiting on I/O
903 if (p
->sleep_type
== SLEEP_NONINTERACTIVE
&& p
->mm
) {
904 if (p
->sleep_avg
>= ceiling
)
906 else if (p
->sleep_avg
+ sleep_time
>=
908 p
->sleep_avg
= ceiling
;
914 * This code gives a bonus to interactive tasks.
916 * The boost works by updating the 'average sleep time'
917 * value here, based on ->timestamp. The more time a
918 * task spends sleeping, the higher the average gets -
919 * and the higher the priority boost gets as well.
921 p
->sleep_avg
+= sleep_time
;
924 if (p
->sleep_avg
> NS_MAX_SLEEP_AVG
)
925 p
->sleep_avg
= NS_MAX_SLEEP_AVG
;
928 return effective_prio(p
);
932 * activate_task - move a task to the runqueue and do priority recalculation
934 * Update all the scheduling statistics stuff. (sleep average
935 * calculation, priority modifiers, etc.)
937 static void activate_task(struct task_struct
*p
, struct rq
*rq
, int local
)
939 unsigned long long now
;
944 /* Compensate for drifting sched_clock */
945 struct rq
*this_rq
= this_rq();
946 now
= (now
- this_rq
->timestamp_last_tick
)
947 + rq
->timestamp_last_tick
;
952 * Sleep time is in units of nanosecs, so shift by 20 to get a
953 * milliseconds-range estimation of the amount of time that the task
956 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
957 if (p
->state
== TASK_UNINTERRUPTIBLE
)
958 profile_hits(SLEEP_PROFILING
, (void *)get_wchan(p
),
959 (now
- p
->timestamp
) >> 20);
963 p
->prio
= recalc_task_prio(p
, now
);
966 * This checks to make sure it's not an uninterruptible task
967 * that is now waking up.
969 if (p
->sleep_type
== SLEEP_NORMAL
) {
971 * Tasks which were woken up by interrupts (ie. hw events)
972 * are most likely of interactive nature. So we give them
973 * the credit of extending their sleep time to the period
974 * of time they spend on the runqueue, waiting for execution
975 * on a CPU, first time around:
978 p
->sleep_type
= SLEEP_INTERRUPTED
;
981 * Normal first-time wakeups get a credit too for
982 * on-runqueue time, but it will be weighted down:
984 p
->sleep_type
= SLEEP_INTERACTIVE
;
989 __activate_task(p
, rq
);
993 * deactivate_task - remove a task from the runqueue.
995 static void deactivate_task(struct task_struct
*p
, struct rq
*rq
)
997 dec_nr_running(p
, rq
);
998 dequeue_task(p
, p
->array
);
1003 * resched_task - mark a task 'to be rescheduled now'.
1005 * On UP this means the setting of the need_resched flag, on SMP it
1006 * might also involve a cross-CPU call to trigger the scheduler on
1011 #ifndef tsk_is_polling
1012 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1015 static void resched_task(struct task_struct
*p
)
1019 assert_spin_locked(&task_rq(p
)->lock
);
1021 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
1024 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
1027 if (cpu
== smp_processor_id())
1030 /* NEED_RESCHED must be visible before we test polling */
1032 if (!tsk_is_polling(p
))
1033 smp_send_reschedule(cpu
);
1036 static inline void resched_task(struct task_struct
*p
)
1038 assert_spin_locked(&task_rq(p
)->lock
);
1039 set_tsk_need_resched(p
);
1044 * task_curr - is this task currently executing on a CPU?
1045 * @p: the task in question.
1047 inline int task_curr(const struct task_struct
*p
)
1049 return cpu_curr(task_cpu(p
)) == p
;
1052 /* Used instead of source_load when we know the type == 0 */
1053 unsigned long weighted_cpuload(const int cpu
)
1055 return cpu_rq(cpu
)->raw_weighted_load
;
1059 struct migration_req
{
1060 struct list_head list
;
1062 struct task_struct
*task
;
1065 struct completion done
;
1069 * The task's runqueue lock must be held.
1070 * Returns true if you have to wait for migration thread.
1073 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1075 struct rq
*rq
= task_rq(p
);
1078 * If the task is not on a runqueue (and not running), then
1079 * it is sufficient to simply update the task's cpu field.
1081 if (!p
->array
&& !task_running(rq
, p
)) {
1082 set_task_cpu(p
, dest_cpu
);
1086 init_completion(&req
->done
);
1088 req
->dest_cpu
= dest_cpu
;
1089 list_add(&req
->list
, &rq
->migration_queue
);
1095 * wait_task_inactive - wait for a thread to unschedule.
1097 * The caller must ensure that the task *will* unschedule sometime soon,
1098 * else this function might spin for a *long* time. This function can't
1099 * be called with interrupts off, or it may introduce deadlock with
1100 * smp_call_function() if an IPI is sent by the same process we are
1101 * waiting to become inactive.
1103 void wait_task_inactive(struct task_struct
*p
)
1105 unsigned long flags
;
1110 rq
= task_rq_lock(p
, &flags
);
1111 /* Must be off runqueue entirely, not preempted. */
1112 if (unlikely(p
->array
|| task_running(rq
, p
))) {
1113 /* If it's preempted, we yield. It could be a while. */
1114 preempted
= !task_running(rq
, p
);
1115 task_rq_unlock(rq
, &flags
);
1121 task_rq_unlock(rq
, &flags
);
1125 * kick_process - kick a running thread to enter/exit the kernel
1126 * @p: the to-be-kicked thread
1128 * Cause a process which is running on another CPU to enter
1129 * kernel-mode, without any delay. (to get signals handled.)
1131 * NOTE: this function doesnt have to take the runqueue lock,
1132 * because all it wants to ensure is that the remote task enters
1133 * the kernel. If the IPI races and the task has been migrated
1134 * to another CPU then no harm is done and the purpose has been
1137 void kick_process(struct task_struct
*p
)
1143 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1144 smp_send_reschedule(cpu
);
1149 * Return a low guess at the load of a migration-source cpu weighted
1150 * according to the scheduling class and "nice" value.
1152 * We want to under-estimate the load of migration sources, to
1153 * balance conservatively.
1155 static inline unsigned long source_load(int cpu
, int type
)
1157 struct rq
*rq
= cpu_rq(cpu
);
1160 return rq
->raw_weighted_load
;
1162 return min(rq
->cpu_load
[type
-1], rq
->raw_weighted_load
);
1166 * Return a high guess at the load of a migration-target cpu weighted
1167 * according to the scheduling class and "nice" value.
1169 static inline unsigned long target_load(int cpu
, int type
)
1171 struct rq
*rq
= cpu_rq(cpu
);
1174 return rq
->raw_weighted_load
;
1176 return max(rq
->cpu_load
[type
-1], rq
->raw_weighted_load
);
1180 * Return the average load per task on the cpu's run queue
1182 static inline unsigned long cpu_avg_load_per_task(int cpu
)
1184 struct rq
*rq
= cpu_rq(cpu
);
1185 unsigned long n
= rq
->nr_running
;
1187 return n
? rq
->raw_weighted_load
/ n
: SCHED_LOAD_SCALE
;
1191 * find_idlest_group finds and returns the least busy CPU group within the
1194 static struct sched_group
*
1195 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1197 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1198 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1199 int load_idx
= sd
->forkexec_idx
;
1200 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1203 unsigned long load
, avg_load
;
1207 /* Skip over this group if it has no CPUs allowed */
1208 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1211 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1213 /* Tally up the load of all CPUs in the group */
1216 for_each_cpu_mask(i
, group
->cpumask
) {
1217 /* Bias balancing toward cpus of our domain */
1219 load
= source_load(i
, load_idx
);
1221 load
= target_load(i
, load_idx
);
1226 /* Adjust by relative CPU power of the group */
1227 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
1230 this_load
= avg_load
;
1232 } else if (avg_load
< min_load
) {
1233 min_load
= avg_load
;
1237 group
= group
->next
;
1238 } while (group
!= sd
->groups
);
1240 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1246 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1249 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1252 unsigned long load
, min_load
= ULONG_MAX
;
1256 /* Traverse only the allowed CPUs */
1257 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1259 for_each_cpu_mask(i
, tmp
) {
1260 load
= weighted_cpuload(i
);
1262 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1272 * sched_balance_self: balance the current task (running on cpu) in domains
1273 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1276 * Balance, ie. select the least loaded group.
1278 * Returns the target CPU number, or the same CPU if no balancing is needed.
1280 * preempt must be disabled.
1282 static int sched_balance_self(int cpu
, int flag
)
1284 struct task_struct
*t
= current
;
1285 struct sched_domain
*tmp
, *sd
= NULL
;
1287 for_each_domain(cpu
, tmp
) {
1289 * If power savings logic is enabled for a domain, stop there.
1291 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1293 if (tmp
->flags
& flag
)
1299 struct sched_group
*group
;
1300 int new_cpu
, weight
;
1302 if (!(sd
->flags
& flag
)) {
1308 group
= find_idlest_group(sd
, t
, cpu
);
1314 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1315 if (new_cpu
== -1 || new_cpu
== cpu
) {
1316 /* Now try balancing at a lower domain level of cpu */
1321 /* Now try balancing at a lower domain level of new_cpu */
1324 weight
= cpus_weight(span
);
1325 for_each_domain(cpu
, tmp
) {
1326 if (weight
<= cpus_weight(tmp
->span
))
1328 if (tmp
->flags
& flag
)
1331 /* while loop will break here if sd == NULL */
1337 #endif /* CONFIG_SMP */
1340 * wake_idle() will wake a task on an idle cpu if task->cpu is
1341 * not idle and an idle cpu is available. The span of cpus to
1342 * search starts with cpus closest then further out as needed,
1343 * so we always favor a closer, idle cpu.
1345 * Returns the CPU we should wake onto.
1347 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1348 static int wake_idle(int cpu
, struct task_struct
*p
)
1351 struct sched_domain
*sd
;
1357 for_each_domain(cpu
, sd
) {
1358 if (sd
->flags
& SD_WAKE_IDLE
) {
1359 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1360 for_each_cpu_mask(i
, tmp
) {
1371 static inline int wake_idle(int cpu
, struct task_struct
*p
)
1378 * try_to_wake_up - wake up a thread
1379 * @p: the to-be-woken-up thread
1380 * @state: the mask of task states that can be woken
1381 * @sync: do a synchronous wakeup?
1383 * Put it on the run-queue if it's not already there. The "current"
1384 * thread is always on the run-queue (except when the actual
1385 * re-schedule is in progress), and as such you're allowed to do
1386 * the simpler "current->state = TASK_RUNNING" to mark yourself
1387 * runnable without the overhead of this.
1389 * returns failure only if the task is already active.
1391 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1393 int cpu
, this_cpu
, success
= 0;
1394 unsigned long flags
;
1398 struct sched_domain
*sd
, *this_sd
= NULL
;
1399 unsigned long load
, this_load
;
1403 rq
= task_rq_lock(p
, &flags
);
1404 old_state
= p
->state
;
1405 if (!(old_state
& state
))
1412 this_cpu
= smp_processor_id();
1415 if (unlikely(task_running(rq
, p
)))
1420 schedstat_inc(rq
, ttwu_cnt
);
1421 if (cpu
== this_cpu
) {
1422 schedstat_inc(rq
, ttwu_local
);
1426 for_each_domain(this_cpu
, sd
) {
1427 if (cpu_isset(cpu
, sd
->span
)) {
1428 schedstat_inc(sd
, ttwu_wake_remote
);
1434 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1438 * Check for affine wakeup and passive balancing possibilities.
1441 int idx
= this_sd
->wake_idx
;
1442 unsigned int imbalance
;
1444 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1446 load
= source_load(cpu
, idx
);
1447 this_load
= target_load(this_cpu
, idx
);
1449 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1451 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1452 unsigned long tl
= this_load
;
1453 unsigned long tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1456 * If sync wakeup then subtract the (maximum possible)
1457 * effect of the currently running task from the load
1458 * of the current CPU:
1461 tl
-= current
->load_weight
;
1464 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1465 100*(tl
+ p
->load_weight
) <= imbalance
*load
) {
1467 * This domain has SD_WAKE_AFFINE and
1468 * p is cache cold in this domain, and
1469 * there is no bad imbalance.
1471 schedstat_inc(this_sd
, ttwu_move_affine
);
1477 * Start passive balancing when half the imbalance_pct
1480 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1481 if (imbalance
*this_load
<= 100*load
) {
1482 schedstat_inc(this_sd
, ttwu_move_balance
);
1488 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1490 new_cpu
= wake_idle(new_cpu
, p
);
1491 if (new_cpu
!= cpu
) {
1492 set_task_cpu(p
, new_cpu
);
1493 task_rq_unlock(rq
, &flags
);
1494 /* might preempt at this point */
1495 rq
= task_rq_lock(p
, &flags
);
1496 old_state
= p
->state
;
1497 if (!(old_state
& state
))
1502 this_cpu
= smp_processor_id();
1507 #endif /* CONFIG_SMP */
1508 if (old_state
== TASK_UNINTERRUPTIBLE
) {
1509 rq
->nr_uninterruptible
--;
1511 * Tasks on involuntary sleep don't earn
1512 * sleep_avg beyond just interactive state.
1514 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
1518 * Tasks that have marked their sleep as noninteractive get
1519 * woken up with their sleep average not weighted in an
1522 if (old_state
& TASK_NONINTERACTIVE
)
1523 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
1526 activate_task(p
, rq
, cpu
== this_cpu
);
1528 * Sync wakeups (i.e. those types of wakeups where the waker
1529 * has indicated that it will leave the CPU in short order)
1530 * don't trigger a preemption, if the woken up task will run on
1531 * this cpu. (in this case the 'I will reschedule' promise of
1532 * the waker guarantees that the freshly woken up task is going
1533 * to be considered on this CPU.)
1535 if (!sync
|| cpu
!= this_cpu
) {
1536 if (TASK_PREEMPTS_CURR(p
, rq
))
1537 resched_task(rq
->curr
);
1542 p
->state
= TASK_RUNNING
;
1544 task_rq_unlock(rq
, &flags
);
1549 int fastcall
wake_up_process(struct task_struct
*p
)
1551 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1552 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1554 EXPORT_SYMBOL(wake_up_process
);
1556 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1558 return try_to_wake_up(p
, state
, 0);
1562 * Perform scheduler related setup for a newly forked process p.
1563 * p is forked by current.
1565 void fastcall
sched_fork(struct task_struct
*p
, int clone_flags
)
1567 int cpu
= get_cpu();
1570 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1572 set_task_cpu(p
, cpu
);
1575 * We mark the process as running here, but have not actually
1576 * inserted it onto the runqueue yet. This guarantees that
1577 * nobody will actually run it, and a signal or other external
1578 * event cannot wake it up and insert it on the runqueue either.
1580 p
->state
= TASK_RUNNING
;
1583 * Make sure we do not leak PI boosting priority to the child:
1585 p
->prio
= current
->normal_prio
;
1587 INIT_LIST_HEAD(&p
->run_list
);
1589 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1590 if (unlikely(sched_info_on()))
1591 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1593 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1596 #ifdef CONFIG_PREEMPT
1597 /* Want to start with kernel preemption disabled. */
1598 task_thread_info(p
)->preempt_count
= 1;
1601 * Share the timeslice between parent and child, thus the
1602 * total amount of pending timeslices in the system doesn't change,
1603 * resulting in more scheduling fairness.
1605 local_irq_disable();
1606 p
->time_slice
= (current
->time_slice
+ 1) >> 1;
1608 * The remainder of the first timeslice might be recovered by
1609 * the parent if the child exits early enough.
1611 p
->first_time_slice
= 1;
1612 current
->time_slice
>>= 1;
1613 p
->timestamp
= sched_clock();
1614 if (unlikely(!current
->time_slice
)) {
1616 * This case is rare, it happens when the parent has only
1617 * a single jiffy left from its timeslice. Taking the
1618 * runqueue lock is not a problem.
1620 current
->time_slice
= 1;
1628 * wake_up_new_task - wake up a newly created task for the first time.
1630 * This function will do some initial scheduler statistics housekeeping
1631 * that must be done for every newly created context, then puts the task
1632 * on the runqueue and wakes it.
1634 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1636 struct rq
*rq
, *this_rq
;
1637 unsigned long flags
;
1640 rq
= task_rq_lock(p
, &flags
);
1641 BUG_ON(p
->state
!= TASK_RUNNING
);
1642 this_cpu
= smp_processor_id();
1646 * We decrease the sleep average of forking parents
1647 * and children as well, to keep max-interactive tasks
1648 * from forking tasks that are max-interactive. The parent
1649 * (current) is done further down, under its lock.
1651 p
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(p
) *
1652 CHILD_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1654 p
->prio
= effective_prio(p
);
1656 if (likely(cpu
== this_cpu
)) {
1657 if (!(clone_flags
& CLONE_VM
)) {
1659 * The VM isn't cloned, so we're in a good position to
1660 * do child-runs-first in anticipation of an exec. This
1661 * usually avoids a lot of COW overhead.
1663 if (unlikely(!current
->array
))
1664 __activate_task(p
, rq
);
1666 p
->prio
= current
->prio
;
1667 p
->normal_prio
= current
->normal_prio
;
1668 list_add_tail(&p
->run_list
, ¤t
->run_list
);
1669 p
->array
= current
->array
;
1670 p
->array
->nr_active
++;
1671 inc_nr_running(p
, rq
);
1675 /* Run child last */
1676 __activate_task(p
, rq
);
1678 * We skip the following code due to cpu == this_cpu
1680 * task_rq_unlock(rq, &flags);
1681 * this_rq = task_rq_lock(current, &flags);
1685 this_rq
= cpu_rq(this_cpu
);
1688 * Not the local CPU - must adjust timestamp. This should
1689 * get optimised away in the !CONFIG_SMP case.
1691 p
->timestamp
= (p
->timestamp
- this_rq
->timestamp_last_tick
)
1692 + rq
->timestamp_last_tick
;
1693 __activate_task(p
, rq
);
1694 if (TASK_PREEMPTS_CURR(p
, rq
))
1695 resched_task(rq
->curr
);
1698 * Parent and child are on different CPUs, now get the
1699 * parent runqueue to update the parent's ->sleep_avg:
1701 task_rq_unlock(rq
, &flags
);
1702 this_rq
= task_rq_lock(current
, &flags
);
1704 current
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(current
) *
1705 PARENT_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1706 task_rq_unlock(this_rq
, &flags
);
1710 * Potentially available exiting-child timeslices are
1711 * retrieved here - this way the parent does not get
1712 * penalized for creating too many threads.
1714 * (this cannot be used to 'generate' timeslices
1715 * artificially, because any timeslice recovered here
1716 * was given away by the parent in the first place.)
1718 void fastcall
sched_exit(struct task_struct
*p
)
1720 unsigned long flags
;
1724 * If the child was a (relative-) CPU hog then decrease
1725 * the sleep_avg of the parent as well.
1727 rq
= task_rq_lock(p
->parent
, &flags
);
1728 if (p
->first_time_slice
&& task_cpu(p
) == task_cpu(p
->parent
)) {
1729 p
->parent
->time_slice
+= p
->time_slice
;
1730 if (unlikely(p
->parent
->time_slice
> task_timeslice(p
)))
1731 p
->parent
->time_slice
= task_timeslice(p
);
1733 if (p
->sleep_avg
< p
->parent
->sleep_avg
)
1734 p
->parent
->sleep_avg
= p
->parent
->sleep_avg
/
1735 (EXIT_WEIGHT
+ 1) * EXIT_WEIGHT
+ p
->sleep_avg
/
1737 task_rq_unlock(rq
, &flags
);
1741 * prepare_task_switch - prepare to switch tasks
1742 * @rq: the runqueue preparing to switch
1743 * @next: the task we are going to switch to.
1745 * This is called with the rq lock held and interrupts off. It must
1746 * be paired with a subsequent finish_task_switch after the context
1749 * prepare_task_switch sets up locking and calls architecture specific
1752 static inline void prepare_task_switch(struct rq
*rq
, struct task_struct
*next
)
1754 prepare_lock_switch(rq
, next
);
1755 prepare_arch_switch(next
);
1759 * finish_task_switch - clean up after a task-switch
1760 * @rq: runqueue associated with task-switch
1761 * @prev: the thread we just switched away from.
1763 * finish_task_switch must be called after the context switch, paired
1764 * with a prepare_task_switch call before the context switch.
1765 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1766 * and do any other architecture-specific cleanup actions.
1768 * Note that we may have delayed dropping an mm in context_switch(). If
1769 * so, we finish that here outside of the runqueue lock. (Doing it
1770 * with the lock held can cause deadlocks; see schedule() for
1773 static inline void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1774 __releases(rq
->lock
)
1776 struct mm_struct
*mm
= rq
->prev_mm
;
1782 * A task struct has one reference for the use as "current".
1783 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1784 * schedule one last time. The schedule call will never return, and
1785 * the scheduled task must drop that reference.
1786 * The test for TASK_DEAD must occur while the runqueue locks are
1787 * still held, otherwise prev could be scheduled on another cpu, die
1788 * there before we look at prev->state, and then the reference would
1790 * Manfred Spraul <manfred@colorfullife.com>
1792 prev_state
= prev
->state
;
1793 finish_arch_switch(prev
);
1794 finish_lock_switch(rq
, prev
);
1797 if (unlikely(prev_state
== TASK_DEAD
)) {
1799 * Remove function-return probe instances associated with this
1800 * task and put them back on the free list.
1802 kprobe_flush_task(prev
);
1803 put_task_struct(prev
);
1808 * schedule_tail - first thing a freshly forked thread must call.
1809 * @prev: the thread we just switched away from.
1811 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1812 __releases(rq
->lock
)
1814 struct rq
*rq
= this_rq();
1816 finish_task_switch(rq
, prev
);
1817 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1818 /* In this case, finish_task_switch does not reenable preemption */
1821 if (current
->set_child_tid
)
1822 put_user(current
->pid
, current
->set_child_tid
);
1826 * context_switch - switch to the new MM and the new
1827 * thread's register state.
1829 static inline struct task_struct
*
1830 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1831 struct task_struct
*next
)
1833 struct mm_struct
*mm
= next
->mm
;
1834 struct mm_struct
*oldmm
= prev
->active_mm
;
1837 next
->active_mm
= oldmm
;
1838 atomic_inc(&oldmm
->mm_count
);
1839 enter_lazy_tlb(oldmm
, next
);
1841 switch_mm(oldmm
, mm
, next
);
1844 prev
->active_mm
= NULL
;
1845 WARN_ON(rq
->prev_mm
);
1846 rq
->prev_mm
= oldmm
;
1849 * Since the runqueue lock will be released by the next
1850 * task (which is an invalid locking op but in the case
1851 * of the scheduler it's an obvious special-case), so we
1852 * do an early lockdep release here:
1854 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1855 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1858 /* Here we just switch the register state and the stack. */
1859 switch_to(prev
, next
, prev
);
1865 * nr_running, nr_uninterruptible and nr_context_switches:
1867 * externally visible scheduler statistics: current number of runnable
1868 * threads, current number of uninterruptible-sleeping threads, total
1869 * number of context switches performed since bootup.
1871 unsigned long nr_running(void)
1873 unsigned long i
, sum
= 0;
1875 for_each_online_cpu(i
)
1876 sum
+= cpu_rq(i
)->nr_running
;
1881 unsigned long nr_uninterruptible(void)
1883 unsigned long i
, sum
= 0;
1885 for_each_possible_cpu(i
)
1886 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1889 * Since we read the counters lockless, it might be slightly
1890 * inaccurate. Do not allow it to go below zero though:
1892 if (unlikely((long)sum
< 0))
1898 unsigned long long nr_context_switches(void)
1901 unsigned long long sum
= 0;
1903 for_each_possible_cpu(i
)
1904 sum
+= cpu_rq(i
)->nr_switches
;
1909 unsigned long nr_iowait(void)
1911 unsigned long i
, sum
= 0;
1913 for_each_possible_cpu(i
)
1914 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1919 unsigned long nr_active(void)
1921 unsigned long i
, running
= 0, uninterruptible
= 0;
1923 for_each_online_cpu(i
) {
1924 running
+= cpu_rq(i
)->nr_running
;
1925 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
1928 if (unlikely((long)uninterruptible
< 0))
1929 uninterruptible
= 0;
1931 return running
+ uninterruptible
;
1937 * Is this task likely cache-hot:
1940 task_hot(struct task_struct
*p
, unsigned long long now
, struct sched_domain
*sd
)
1942 return (long long)(now
- p
->last_ran
) < (long long)sd
->cache_hot_time
;
1946 * double_rq_lock - safely lock two runqueues
1948 * Note this does not disable interrupts like task_rq_lock,
1949 * you need to do so manually before calling.
1951 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1952 __acquires(rq1
->lock
)
1953 __acquires(rq2
->lock
)
1955 BUG_ON(!irqs_disabled());
1957 spin_lock(&rq1
->lock
);
1958 __acquire(rq2
->lock
); /* Fake it out ;) */
1961 spin_lock(&rq1
->lock
);
1962 spin_lock(&rq2
->lock
);
1964 spin_lock(&rq2
->lock
);
1965 spin_lock(&rq1
->lock
);
1971 * double_rq_unlock - safely unlock two runqueues
1973 * Note this does not restore interrupts like task_rq_unlock,
1974 * you need to do so manually after calling.
1976 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
1977 __releases(rq1
->lock
)
1978 __releases(rq2
->lock
)
1980 spin_unlock(&rq1
->lock
);
1982 spin_unlock(&rq2
->lock
);
1984 __release(rq2
->lock
);
1988 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1990 static void double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1991 __releases(this_rq
->lock
)
1992 __acquires(busiest
->lock
)
1993 __acquires(this_rq
->lock
)
1995 if (unlikely(!irqs_disabled())) {
1996 /* printk() doesn't work good under rq->lock */
1997 spin_unlock(&this_rq
->lock
);
2000 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2001 if (busiest
< this_rq
) {
2002 spin_unlock(&this_rq
->lock
);
2003 spin_lock(&busiest
->lock
);
2004 spin_lock(&this_rq
->lock
);
2006 spin_lock(&busiest
->lock
);
2011 * If dest_cpu is allowed for this process, migrate the task to it.
2012 * This is accomplished by forcing the cpu_allowed mask to only
2013 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2014 * the cpu_allowed mask is restored.
2016 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2018 struct migration_req req
;
2019 unsigned long flags
;
2022 rq
= task_rq_lock(p
, &flags
);
2023 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2024 || unlikely(cpu_is_offline(dest_cpu
)))
2027 /* force the process onto the specified CPU */
2028 if (migrate_task(p
, dest_cpu
, &req
)) {
2029 /* Need to wait for migration thread (might exit: take ref). */
2030 struct task_struct
*mt
= rq
->migration_thread
;
2032 get_task_struct(mt
);
2033 task_rq_unlock(rq
, &flags
);
2034 wake_up_process(mt
);
2035 put_task_struct(mt
);
2036 wait_for_completion(&req
.done
);
2041 task_rq_unlock(rq
, &flags
);
2045 * sched_exec - execve() is a valuable balancing opportunity, because at
2046 * this point the task has the smallest effective memory and cache footprint.
2048 void sched_exec(void)
2050 int new_cpu
, this_cpu
= get_cpu();
2051 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2053 if (new_cpu
!= this_cpu
)
2054 sched_migrate_task(current
, new_cpu
);
2058 * pull_task - move a task from a remote runqueue to the local runqueue.
2059 * Both runqueues must be locked.
2061 static void pull_task(struct rq
*src_rq
, struct prio_array
*src_array
,
2062 struct task_struct
*p
, struct rq
*this_rq
,
2063 struct prio_array
*this_array
, int this_cpu
)
2065 dequeue_task(p
, src_array
);
2066 dec_nr_running(p
, src_rq
);
2067 set_task_cpu(p
, this_cpu
);
2068 inc_nr_running(p
, this_rq
);
2069 enqueue_task(p
, this_array
);
2070 p
->timestamp
= (p
->timestamp
- src_rq
->timestamp_last_tick
)
2071 + this_rq
->timestamp_last_tick
;
2073 * Note that idle threads have a prio of MAX_PRIO, for this test
2074 * to be always true for them.
2076 if (TASK_PREEMPTS_CURR(p
, this_rq
))
2077 resched_task(this_rq
->curr
);
2081 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2084 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2085 struct sched_domain
*sd
, enum idle_type idle
,
2089 * We do not migrate tasks that are:
2090 * 1) running (obviously), or
2091 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2092 * 3) are cache-hot on their current CPU.
2094 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
2098 if (task_running(rq
, p
))
2102 * Aggressive migration if:
2103 * 1) task is cache cold, or
2104 * 2) too many balance attempts have failed.
2107 if (sd
->nr_balance_failed
> sd
->cache_nice_tries
)
2110 if (task_hot(p
, rq
->timestamp_last_tick
, sd
))
2115 #define rq_best_prio(rq) min((rq)->curr->prio, (rq)->best_expired_prio)
2118 * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted
2119 * load from busiest to this_rq, as part of a balancing operation within
2120 * "domain". Returns the number of tasks moved.
2122 * Called with both runqueues locked.
2124 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2125 unsigned long max_nr_move
, unsigned long max_load_move
,
2126 struct sched_domain
*sd
, enum idle_type idle
,
2129 int idx
, pulled
= 0, pinned
= 0, this_best_prio
, best_prio
,
2130 best_prio_seen
, skip_for_load
;
2131 struct prio_array
*array
, *dst_array
;
2132 struct list_head
*head
, *curr
;
2133 struct task_struct
*tmp
;
2136 if (max_nr_move
== 0 || max_load_move
== 0)
2139 rem_load_move
= max_load_move
;
2141 this_best_prio
= rq_best_prio(this_rq
);
2142 best_prio
= rq_best_prio(busiest
);
2144 * Enable handling of the case where there is more than one task
2145 * with the best priority. If the current running task is one
2146 * of those with prio==best_prio we know it won't be moved
2147 * and therefore it's safe to override the skip (based on load) of
2148 * any task we find with that prio.
2150 best_prio_seen
= best_prio
== busiest
->curr
->prio
;
2153 * We first consider expired tasks. Those will likely not be
2154 * executed in the near future, and they are most likely to
2155 * be cache-cold, thus switching CPUs has the least effect
2158 if (busiest
->expired
->nr_active
) {
2159 array
= busiest
->expired
;
2160 dst_array
= this_rq
->expired
;
2162 array
= busiest
->active
;
2163 dst_array
= this_rq
->active
;
2167 /* Start searching at priority 0: */
2171 idx
= sched_find_first_bit(array
->bitmap
);
2173 idx
= find_next_bit(array
->bitmap
, MAX_PRIO
, idx
);
2174 if (idx
>= MAX_PRIO
) {
2175 if (array
== busiest
->expired
&& busiest
->active
->nr_active
) {
2176 array
= busiest
->active
;
2177 dst_array
= this_rq
->active
;
2183 head
= array
->queue
+ idx
;
2186 tmp
= list_entry(curr
, struct task_struct
, run_list
);
2191 * To help distribute high priority tasks accross CPUs we don't
2192 * skip a task if it will be the highest priority task (i.e. smallest
2193 * prio value) on its new queue regardless of its load weight
2195 skip_for_load
= tmp
->load_weight
> rem_load_move
;
2196 if (skip_for_load
&& idx
< this_best_prio
)
2197 skip_for_load
= !best_prio_seen
&& idx
== best_prio
;
2198 if (skip_for_load
||
2199 !can_migrate_task(tmp
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2201 best_prio_seen
|= idx
== best_prio
;
2208 #ifdef CONFIG_SCHEDSTATS
2209 if (task_hot(tmp
, busiest
->timestamp_last_tick
, sd
))
2210 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2213 pull_task(busiest
, array
, tmp
, this_rq
, dst_array
, this_cpu
);
2215 rem_load_move
-= tmp
->load_weight
;
2218 * We only want to steal up to the prescribed number of tasks
2219 * and the prescribed amount of weighted load.
2221 if (pulled
< max_nr_move
&& rem_load_move
> 0) {
2222 if (idx
< this_best_prio
)
2223 this_best_prio
= idx
;
2231 * Right now, this is the only place pull_task() is called,
2232 * so we can safely collect pull_task() stats here rather than
2233 * inside pull_task().
2235 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2238 *all_pinned
= pinned
;
2243 * find_busiest_group finds and returns the busiest CPU group within the
2244 * domain. It calculates and returns the amount of weighted load which
2245 * should be moved to restore balance via the imbalance parameter.
2247 static struct sched_group
*
2248 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2249 unsigned long *imbalance
, enum idle_type idle
, int *sd_idle
,
2252 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2253 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2254 unsigned long max_pull
;
2255 unsigned long busiest_load_per_task
, busiest_nr_running
;
2256 unsigned long this_load_per_task
, this_nr_running
;
2258 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2259 int power_savings_balance
= 1;
2260 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2261 unsigned long min_nr_running
= ULONG_MAX
;
2262 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2265 max_load
= this_load
= total_load
= total_pwr
= 0;
2266 busiest_load_per_task
= busiest_nr_running
= 0;
2267 this_load_per_task
= this_nr_running
= 0;
2268 if (idle
== NOT_IDLE
)
2269 load_idx
= sd
->busy_idx
;
2270 else if (idle
== NEWLY_IDLE
)
2271 load_idx
= sd
->newidle_idx
;
2273 load_idx
= sd
->idle_idx
;
2276 unsigned long load
, group_capacity
;
2279 unsigned long sum_nr_running
, sum_weighted_load
;
2281 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2283 /* Tally up the load of all CPUs in the group */
2284 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2286 for_each_cpu_mask(i
, group
->cpumask
) {
2289 if (!cpu_isset(i
, *cpus
))
2294 if (*sd_idle
&& !idle_cpu(i
))
2297 /* Bias balancing toward cpus of our domain */
2299 load
= target_load(i
, load_idx
);
2301 load
= source_load(i
, load_idx
);
2304 sum_nr_running
+= rq
->nr_running
;
2305 sum_weighted_load
+= rq
->raw_weighted_load
;
2308 total_load
+= avg_load
;
2309 total_pwr
+= group
->cpu_power
;
2311 /* Adjust by relative CPU power of the group */
2312 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
2314 group_capacity
= group
->cpu_power
/ SCHED_LOAD_SCALE
;
2317 this_load
= avg_load
;
2319 this_nr_running
= sum_nr_running
;
2320 this_load_per_task
= sum_weighted_load
;
2321 } else if (avg_load
> max_load
&&
2322 sum_nr_running
> group_capacity
) {
2323 max_load
= avg_load
;
2325 busiest_nr_running
= sum_nr_running
;
2326 busiest_load_per_task
= sum_weighted_load
;
2329 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2331 * Busy processors will not participate in power savings
2334 if (idle
== NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2338 * If the local group is idle or completely loaded
2339 * no need to do power savings balance at this domain
2341 if (local_group
&& (this_nr_running
>= group_capacity
||
2343 power_savings_balance
= 0;
2346 * If a group is already running at full capacity or idle,
2347 * don't include that group in power savings calculations
2349 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2354 * Calculate the group which has the least non-idle load.
2355 * This is the group from where we need to pick up the load
2358 if ((sum_nr_running
< min_nr_running
) ||
2359 (sum_nr_running
== min_nr_running
&&
2360 first_cpu(group
->cpumask
) <
2361 first_cpu(group_min
->cpumask
))) {
2363 min_nr_running
= sum_nr_running
;
2364 min_load_per_task
= sum_weighted_load
/
2369 * Calculate the group which is almost near its
2370 * capacity but still has some space to pick up some load
2371 * from other group and save more power
2373 if (sum_nr_running
<= group_capacity
- 1) {
2374 if (sum_nr_running
> leader_nr_running
||
2375 (sum_nr_running
== leader_nr_running
&&
2376 first_cpu(group
->cpumask
) >
2377 first_cpu(group_leader
->cpumask
))) {
2378 group_leader
= group
;
2379 leader_nr_running
= sum_nr_running
;
2384 group
= group
->next
;
2385 } while (group
!= sd
->groups
);
2387 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2390 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2392 if (this_load
>= avg_load
||
2393 100*max_load
<= sd
->imbalance_pct
*this_load
)
2396 busiest_load_per_task
/= busiest_nr_running
;
2398 * We're trying to get all the cpus to the average_load, so we don't
2399 * want to push ourselves above the average load, nor do we wish to
2400 * reduce the max loaded cpu below the average load, as either of these
2401 * actions would just result in more rebalancing later, and ping-pong
2402 * tasks around. Thus we look for the minimum possible imbalance.
2403 * Negative imbalances (*we* are more loaded than anyone else) will
2404 * be counted as no imbalance for these purposes -- we can't fix that
2405 * by pulling tasks to us. Be careful of negative numbers as they'll
2406 * appear as very large values with unsigned longs.
2408 if (max_load
<= busiest_load_per_task
)
2412 * In the presence of smp nice balancing, certain scenarios can have
2413 * max load less than avg load(as we skip the groups at or below
2414 * its cpu_power, while calculating max_load..)
2416 if (max_load
< avg_load
) {
2418 goto small_imbalance
;
2421 /* Don't want to pull so many tasks that a group would go idle */
2422 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2424 /* How much load to actually move to equalise the imbalance */
2425 *imbalance
= min(max_pull
* busiest
->cpu_power
,
2426 (avg_load
- this_load
) * this->cpu_power
)
2430 * if *imbalance is less than the average load per runnable task
2431 * there is no gaurantee that any tasks will be moved so we'll have
2432 * a think about bumping its value to force at least one task to be
2435 if (*imbalance
< busiest_load_per_task
) {
2436 unsigned long tmp
, pwr_now
, pwr_move
;
2440 pwr_move
= pwr_now
= 0;
2442 if (this_nr_running
) {
2443 this_load_per_task
/= this_nr_running
;
2444 if (busiest_load_per_task
> this_load_per_task
)
2447 this_load_per_task
= SCHED_LOAD_SCALE
;
2449 if (max_load
- this_load
>= busiest_load_per_task
* imbn
) {
2450 *imbalance
= busiest_load_per_task
;
2455 * OK, we don't have enough imbalance to justify moving tasks,
2456 * however we may be able to increase total CPU power used by
2460 pwr_now
+= busiest
->cpu_power
*
2461 min(busiest_load_per_task
, max_load
);
2462 pwr_now
+= this->cpu_power
*
2463 min(this_load_per_task
, this_load
);
2464 pwr_now
/= SCHED_LOAD_SCALE
;
2466 /* Amount of load we'd subtract */
2467 tmp
= busiest_load_per_task
*SCHED_LOAD_SCALE
/busiest
->cpu_power
;
2469 pwr_move
+= busiest
->cpu_power
*
2470 min(busiest_load_per_task
, max_load
- tmp
);
2472 /* Amount of load we'd add */
2473 if (max_load
*busiest
->cpu_power
<
2474 busiest_load_per_task
*SCHED_LOAD_SCALE
)
2475 tmp
= max_load
*busiest
->cpu_power
/this->cpu_power
;
2477 tmp
= busiest_load_per_task
*SCHED_LOAD_SCALE
/this->cpu_power
;
2478 pwr_move
+= this->cpu_power
*min(this_load_per_task
, this_load
+ tmp
);
2479 pwr_move
/= SCHED_LOAD_SCALE
;
2481 /* Move if we gain throughput */
2482 if (pwr_move
<= pwr_now
)
2485 *imbalance
= busiest_load_per_task
;
2491 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2492 if (idle
== NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2495 if (this == group_leader
&& group_leader
!= group_min
) {
2496 *imbalance
= min_load_per_task
;
2506 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2509 find_busiest_queue(struct sched_group
*group
, enum idle_type idle
,
2510 unsigned long imbalance
, cpumask_t
*cpus
)
2512 struct rq
*busiest
= NULL
, *rq
;
2513 unsigned long max_load
= 0;
2516 for_each_cpu_mask(i
, group
->cpumask
) {
2518 if (!cpu_isset(i
, *cpus
))
2523 if (rq
->nr_running
== 1 && rq
->raw_weighted_load
> imbalance
)
2526 if (rq
->raw_weighted_load
> max_load
) {
2527 max_load
= rq
->raw_weighted_load
;
2536 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2537 * so long as it is large enough.
2539 #define MAX_PINNED_INTERVAL 512
2541 static inline unsigned long minus_1_or_zero(unsigned long n
)
2543 return n
> 0 ? n
- 1 : 0;
2547 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2548 * tasks if there is an imbalance.
2550 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2551 struct sched_domain
*sd
, enum idle_type idle
)
2553 int nr_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2554 struct sched_group
*group
;
2555 unsigned long imbalance
;
2557 cpumask_t cpus
= CPU_MASK_ALL
;
2558 unsigned long flags
;
2561 * When power savings policy is enabled for the parent domain, idle
2562 * sibling can pick up load irrespective of busy siblings. In this case,
2563 * let the state of idle sibling percolate up as IDLE, instead of
2564 * portraying it as NOT_IDLE.
2566 if (idle
!= NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2567 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2570 schedstat_inc(sd
, lb_cnt
[idle
]);
2573 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2576 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2580 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2582 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2586 BUG_ON(busiest
== this_rq
);
2588 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2591 if (busiest
->nr_running
> 1) {
2593 * Attempt to move tasks. If find_busiest_group has found
2594 * an imbalance but busiest->nr_running <= 1, the group is
2595 * still unbalanced. nr_moved simply stays zero, so it is
2596 * correctly treated as an imbalance.
2598 local_irq_save(flags
);
2599 double_rq_lock(this_rq
, busiest
);
2600 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2601 minus_1_or_zero(busiest
->nr_running
),
2602 imbalance
, sd
, idle
, &all_pinned
);
2603 double_rq_unlock(this_rq
, busiest
);
2604 local_irq_restore(flags
);
2606 /* All tasks on this runqueue were pinned by CPU affinity */
2607 if (unlikely(all_pinned
)) {
2608 cpu_clear(cpu_of(busiest
), cpus
);
2609 if (!cpus_empty(cpus
))
2616 schedstat_inc(sd
, lb_failed
[idle
]);
2617 sd
->nr_balance_failed
++;
2619 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2621 spin_lock_irqsave(&busiest
->lock
, flags
);
2623 /* don't kick the migration_thread, if the curr
2624 * task on busiest cpu can't be moved to this_cpu
2626 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2627 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2629 goto out_one_pinned
;
2632 if (!busiest
->active_balance
) {
2633 busiest
->active_balance
= 1;
2634 busiest
->push_cpu
= this_cpu
;
2637 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2639 wake_up_process(busiest
->migration_thread
);
2642 * We've kicked active balancing, reset the failure
2645 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2648 sd
->nr_balance_failed
= 0;
2650 if (likely(!active_balance
)) {
2651 /* We were unbalanced, so reset the balancing interval */
2652 sd
->balance_interval
= sd
->min_interval
;
2655 * If we've begun active balancing, start to back off. This
2656 * case may not be covered by the all_pinned logic if there
2657 * is only 1 task on the busy runqueue (because we don't call
2660 if (sd
->balance_interval
< sd
->max_interval
)
2661 sd
->balance_interval
*= 2;
2664 if (!nr_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2665 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2670 schedstat_inc(sd
, lb_balanced
[idle
]);
2672 sd
->nr_balance_failed
= 0;
2675 /* tune up the balancing interval */
2676 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2677 (sd
->balance_interval
< sd
->max_interval
))
2678 sd
->balance_interval
*= 2;
2680 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2681 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2687 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2688 * tasks if there is an imbalance.
2690 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2691 * this_rq is locked.
2694 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
2696 struct sched_group
*group
;
2697 struct rq
*busiest
= NULL
;
2698 unsigned long imbalance
;
2701 cpumask_t cpus
= CPU_MASK_ALL
;
2704 * When power savings policy is enabled for the parent domain, idle
2705 * sibling can pick up load irrespective of busy siblings. In this case,
2706 * let the state of idle sibling percolate up as IDLE, instead of
2707 * portraying it as NOT_IDLE.
2709 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
2710 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2713 schedstat_inc(sd
, lb_cnt
[NEWLY_IDLE
]);
2715 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, NEWLY_IDLE
,
2718 schedstat_inc(sd
, lb_nobusyg
[NEWLY_IDLE
]);
2722 busiest
= find_busiest_queue(group
, NEWLY_IDLE
, imbalance
,
2725 schedstat_inc(sd
, lb_nobusyq
[NEWLY_IDLE
]);
2729 BUG_ON(busiest
== this_rq
);
2731 schedstat_add(sd
, lb_imbalance
[NEWLY_IDLE
], imbalance
);
2734 if (busiest
->nr_running
> 1) {
2735 /* Attempt to move tasks */
2736 double_lock_balance(this_rq
, busiest
);
2737 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2738 minus_1_or_zero(busiest
->nr_running
),
2739 imbalance
, sd
, NEWLY_IDLE
, NULL
);
2740 spin_unlock(&busiest
->lock
);
2743 cpu_clear(cpu_of(busiest
), cpus
);
2744 if (!cpus_empty(cpus
))
2750 schedstat_inc(sd
, lb_failed
[NEWLY_IDLE
]);
2751 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2752 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2755 sd
->nr_balance_failed
= 0;
2760 schedstat_inc(sd
, lb_balanced
[NEWLY_IDLE
]);
2761 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2762 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2764 sd
->nr_balance_failed
= 0;
2770 * idle_balance is called by schedule() if this_cpu is about to become
2771 * idle. Attempts to pull tasks from other CPUs.
2773 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
2775 struct sched_domain
*sd
;
2777 for_each_domain(this_cpu
, sd
) {
2778 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
2779 /* If we've pulled tasks over stop searching: */
2780 if (load_balance_newidle(this_cpu
, this_rq
, sd
))
2787 * active_load_balance is run by migration threads. It pushes running tasks
2788 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2789 * running on each physical CPU where possible, and avoids physical /
2790 * logical imbalances.
2792 * Called with busiest_rq locked.
2794 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
2796 int target_cpu
= busiest_rq
->push_cpu
;
2797 struct sched_domain
*sd
;
2798 struct rq
*target_rq
;
2800 /* Is there any task to move? */
2801 if (busiest_rq
->nr_running
<= 1)
2804 target_rq
= cpu_rq(target_cpu
);
2807 * This condition is "impossible", if it occurs
2808 * we need to fix it. Originally reported by
2809 * Bjorn Helgaas on a 128-cpu setup.
2811 BUG_ON(busiest_rq
== target_rq
);
2813 /* move a task from busiest_rq to target_rq */
2814 double_lock_balance(busiest_rq
, target_rq
);
2816 /* Search for an sd spanning us and the target CPU. */
2817 for_each_domain(target_cpu
, sd
) {
2818 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2819 cpu_isset(busiest_cpu
, sd
->span
))
2824 schedstat_inc(sd
, alb_cnt
);
2826 if (move_tasks(target_rq
, target_cpu
, busiest_rq
, 1,
2827 RTPRIO_TO_LOAD_WEIGHT(100), sd
, SCHED_IDLE
,
2829 schedstat_inc(sd
, alb_pushed
);
2831 schedstat_inc(sd
, alb_failed
);
2833 spin_unlock(&target_rq
->lock
);
2837 * rebalance_tick will get called every timer tick, on every CPU.
2839 * It checks each scheduling domain to see if it is due to be balanced,
2840 * and initiates a balancing operation if so.
2842 * Balancing parameters are set up in arch_init_sched_domains.
2846 rebalance_tick(int this_cpu
, struct rq
*this_rq
, enum idle_type idle
)
2848 unsigned long this_load
, interval
;
2849 struct sched_domain
*sd
;
2852 this_load
= this_rq
->raw_weighted_load
;
2854 /* Update our load: */
2855 for (i
= 0, scale
= 1; i
< 3; i
++, scale
<<= 1) {
2856 unsigned long old_load
, new_load
;
2858 old_load
= this_rq
->cpu_load
[i
];
2859 new_load
= this_load
;
2861 * Round up the averaging division if load is increasing. This
2862 * prevents us from getting stuck on 9 if the load is 10, for
2865 if (new_load
> old_load
)
2866 new_load
+= scale
-1;
2867 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) / scale
;
2870 for_each_domain(this_cpu
, sd
) {
2871 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2874 interval
= sd
->balance_interval
;
2875 if (idle
!= SCHED_IDLE
)
2876 interval
*= sd
->busy_factor
;
2878 /* scale ms to jiffies */
2879 interval
= msecs_to_jiffies(interval
);
2880 if (unlikely(!interval
))
2883 if (jiffies
- sd
->last_balance
>= interval
) {
2884 if (load_balance(this_cpu
, this_rq
, sd
, idle
)) {
2886 * We've pulled tasks over so either we're no
2887 * longer idle, or one of our SMT siblings is
2892 sd
->last_balance
+= interval
;
2898 * on UP we do not need to balance between CPUs:
2900 static inline void rebalance_tick(int cpu
, struct rq
*rq
, enum idle_type idle
)
2903 static inline void idle_balance(int cpu
, struct rq
*rq
)
2908 static inline int wake_priority_sleeper(struct rq
*rq
)
2912 #ifdef CONFIG_SCHED_SMT
2913 if (!rq
->nr_running
)
2916 spin_lock(&rq
->lock
);
2918 * If an SMT sibling task has been put to sleep for priority
2919 * reasons reschedule the idle task to see if it can now run.
2921 if (rq
->nr_running
) {
2922 resched_task(rq
->idle
);
2925 spin_unlock(&rq
->lock
);
2930 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2932 EXPORT_PER_CPU_SYMBOL(kstat
);
2935 * This is called on clock ticks and on context switches.
2936 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2939 update_cpu_clock(struct task_struct
*p
, struct rq
*rq
, unsigned long long now
)
2941 p
->sched_time
+= now
- max(p
->timestamp
, rq
->timestamp_last_tick
);
2945 * Return current->sched_time plus any more ns on the sched_clock
2946 * that have not yet been banked.
2948 unsigned long long current_sched_time(const struct task_struct
*p
)
2950 unsigned long long ns
;
2951 unsigned long flags
;
2953 local_irq_save(flags
);
2954 ns
= max(p
->timestamp
, task_rq(p
)->timestamp_last_tick
);
2955 ns
= p
->sched_time
+ sched_clock() - ns
;
2956 local_irq_restore(flags
);
2962 * We place interactive tasks back into the active array, if possible.
2964 * To guarantee that this does not starve expired tasks we ignore the
2965 * interactivity of a task if the first expired task had to wait more
2966 * than a 'reasonable' amount of time. This deadline timeout is
2967 * load-dependent, as the frequency of array switched decreases with
2968 * increasing number of running tasks. We also ignore the interactivity
2969 * if a better static_prio task has expired:
2971 static inline int expired_starving(struct rq
*rq
)
2973 if (rq
->curr
->static_prio
> rq
->best_expired_prio
)
2975 if (!STARVATION_LIMIT
|| !rq
->expired_timestamp
)
2977 if (jiffies
- rq
->expired_timestamp
> STARVATION_LIMIT
* rq
->nr_running
)
2983 * Account user cpu time to a process.
2984 * @p: the process that the cpu time gets accounted to
2985 * @hardirq_offset: the offset to subtract from hardirq_count()
2986 * @cputime: the cpu time spent in user space since the last update
2988 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
2990 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2993 p
->utime
= cputime_add(p
->utime
, cputime
);
2995 /* Add user time to cpustat. */
2996 tmp
= cputime_to_cputime64(cputime
);
2997 if (TASK_NICE(p
) > 0)
2998 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3000 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3004 * Account system cpu time to a process.
3005 * @p: the process that the cpu time gets accounted to
3006 * @hardirq_offset: the offset to subtract from hardirq_count()
3007 * @cputime: the cpu time spent in kernel space since the last update
3009 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3012 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3013 struct rq
*rq
= this_rq();
3016 p
->stime
= cputime_add(p
->stime
, cputime
);
3018 /* Add system time to cpustat. */
3019 tmp
= cputime_to_cputime64(cputime
);
3020 if (hardirq_count() - hardirq_offset
)
3021 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3022 else if (softirq_count())
3023 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3024 else if (p
!= rq
->idle
)
3025 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3026 else if (atomic_read(&rq
->nr_iowait
) > 0)
3027 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3029 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3030 /* Account for system time used */
3031 acct_update_integrals(p
);
3035 * Account for involuntary wait time.
3036 * @p: the process from which the cpu time has been stolen
3037 * @steal: the cpu time spent in involuntary wait
3039 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3041 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3042 cputime64_t tmp
= cputime_to_cputime64(steal
);
3043 struct rq
*rq
= this_rq();
3045 if (p
== rq
->idle
) {
3046 p
->stime
= cputime_add(p
->stime
, steal
);
3047 if (atomic_read(&rq
->nr_iowait
) > 0)
3048 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3050 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3052 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3056 * This function gets called by the timer code, with HZ frequency.
3057 * We call it with interrupts disabled.
3059 * It also gets called by the fork code, when changing the parent's
3062 void scheduler_tick(void)
3064 unsigned long long now
= sched_clock();
3065 struct task_struct
*p
= current
;
3066 int cpu
= smp_processor_id();
3067 struct rq
*rq
= cpu_rq(cpu
);
3069 update_cpu_clock(p
, rq
, now
);
3071 rq
->timestamp_last_tick
= now
;
3073 if (p
== rq
->idle
) {
3074 if (wake_priority_sleeper(rq
))
3076 rebalance_tick(cpu
, rq
, SCHED_IDLE
);
3080 /* Task might have expired already, but not scheduled off yet */
3081 if (p
->array
!= rq
->active
) {
3082 set_tsk_need_resched(p
);
3085 spin_lock(&rq
->lock
);
3087 * The task was running during this tick - update the
3088 * time slice counter. Note: we do not update a thread's
3089 * priority until it either goes to sleep or uses up its
3090 * timeslice. This makes it possible for interactive tasks
3091 * to use up their timeslices at their highest priority levels.
3095 * RR tasks need a special form of timeslice management.
3096 * FIFO tasks have no timeslices.
3098 if ((p
->policy
== SCHED_RR
) && !--p
->time_slice
) {
3099 p
->time_slice
= task_timeslice(p
);
3100 p
->first_time_slice
= 0;
3101 set_tsk_need_resched(p
);
3103 /* put it at the end of the queue: */
3104 requeue_task(p
, rq
->active
);
3108 if (!--p
->time_slice
) {
3109 dequeue_task(p
, rq
->active
);
3110 set_tsk_need_resched(p
);
3111 p
->prio
= effective_prio(p
);
3112 p
->time_slice
= task_timeslice(p
);
3113 p
->first_time_slice
= 0;
3115 if (!rq
->expired_timestamp
)
3116 rq
->expired_timestamp
= jiffies
;
3117 if (!TASK_INTERACTIVE(p
) || expired_starving(rq
)) {
3118 enqueue_task(p
, rq
->expired
);
3119 if (p
->static_prio
< rq
->best_expired_prio
)
3120 rq
->best_expired_prio
= p
->static_prio
;
3122 enqueue_task(p
, rq
->active
);
3125 * Prevent a too long timeslice allowing a task to monopolize
3126 * the CPU. We do this by splitting up the timeslice into
3129 * Note: this does not mean the task's timeslices expire or
3130 * get lost in any way, they just might be preempted by
3131 * another task of equal priority. (one with higher
3132 * priority would have preempted this task already.) We
3133 * requeue this task to the end of the list on this priority
3134 * level, which is in essence a round-robin of tasks with
3137 * This only applies to tasks in the interactive
3138 * delta range with at least TIMESLICE_GRANULARITY to requeue.
3140 if (TASK_INTERACTIVE(p
) && !((task_timeslice(p
) -
3141 p
->time_slice
) % TIMESLICE_GRANULARITY(p
)) &&
3142 (p
->time_slice
>= TIMESLICE_GRANULARITY(p
)) &&
3143 (p
->array
== rq
->active
)) {
3145 requeue_task(p
, rq
->active
);
3146 set_tsk_need_resched(p
);
3150 spin_unlock(&rq
->lock
);
3152 rebalance_tick(cpu
, rq
, NOT_IDLE
);
3155 #ifdef CONFIG_SCHED_SMT
3156 static inline void wakeup_busy_runqueue(struct rq
*rq
)
3158 /* If an SMT runqueue is sleeping due to priority reasons wake it up */
3159 if (rq
->curr
== rq
->idle
&& rq
->nr_running
)
3160 resched_task(rq
->idle
);
3164 * Called with interrupt disabled and this_rq's runqueue locked.
3166 static void wake_sleeping_dependent(int this_cpu
)
3168 struct sched_domain
*tmp
, *sd
= NULL
;
3171 for_each_domain(this_cpu
, tmp
) {
3172 if (tmp
->flags
& SD_SHARE_CPUPOWER
) {
3181 for_each_cpu_mask(i
, sd
->span
) {
3182 struct rq
*smt_rq
= cpu_rq(i
);
3186 if (unlikely(!spin_trylock(&smt_rq
->lock
)))
3189 wakeup_busy_runqueue(smt_rq
);
3190 spin_unlock(&smt_rq
->lock
);
3195 * number of 'lost' timeslices this task wont be able to fully
3196 * utilize, if another task runs on a sibling. This models the
3197 * slowdown effect of other tasks running on siblings:
3199 static inline unsigned long
3200 smt_slice(struct task_struct
*p
, struct sched_domain
*sd
)
3202 return p
->time_slice
* (100 - sd
->per_cpu_gain
) / 100;
3206 * To minimise lock contention and not have to drop this_rq's runlock we only
3207 * trylock the sibling runqueues and bypass those runqueues if we fail to
3208 * acquire their lock. As we only trylock the normal locking order does not
3209 * need to be obeyed.
3212 dependent_sleeper(int this_cpu
, struct rq
*this_rq
, struct task_struct
*p
)
3214 struct sched_domain
*tmp
, *sd
= NULL
;
3217 /* kernel/rt threads do not participate in dependent sleeping */
3218 if (!p
->mm
|| rt_task(p
))
3221 for_each_domain(this_cpu
, tmp
) {
3222 if (tmp
->flags
& SD_SHARE_CPUPOWER
) {
3231 for_each_cpu_mask(i
, sd
->span
) {
3232 struct task_struct
*smt_curr
;
3239 if (unlikely(!spin_trylock(&smt_rq
->lock
)))
3242 smt_curr
= smt_rq
->curr
;
3248 * If a user task with lower static priority than the
3249 * running task on the SMT sibling is trying to schedule,
3250 * delay it till there is proportionately less timeslice
3251 * left of the sibling task to prevent a lower priority
3252 * task from using an unfair proportion of the
3253 * physical cpu's resources. -ck
3255 if (rt_task(smt_curr
)) {
3257 * With real time tasks we run non-rt tasks only
3258 * per_cpu_gain% of the time.
3260 if ((jiffies
% DEF_TIMESLICE
) >
3261 (sd
->per_cpu_gain
* DEF_TIMESLICE
/ 100))
3264 if (smt_curr
->static_prio
< p
->static_prio
&&
3265 !TASK_PREEMPTS_CURR(p
, smt_rq
) &&
3266 smt_slice(smt_curr
, sd
) > task_timeslice(p
))
3270 spin_unlock(&smt_rq
->lock
);
3275 static inline void wake_sleeping_dependent(int this_cpu
)
3279 dependent_sleeper(int this_cpu
, struct rq
*this_rq
, struct task_struct
*p
)
3285 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3287 void fastcall
add_preempt_count(int val
)
3292 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3294 preempt_count() += val
;
3296 * Spinlock count overflowing soon?
3298 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >= PREEMPT_MASK
-10);
3300 EXPORT_SYMBOL(add_preempt_count
);
3302 void fastcall
sub_preempt_count(int val
)
3307 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3310 * Is the spinlock portion underflowing?
3312 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3313 !(preempt_count() & PREEMPT_MASK
)))
3316 preempt_count() -= val
;
3318 EXPORT_SYMBOL(sub_preempt_count
);
3322 static inline int interactive_sleep(enum sleep_type sleep_type
)
3324 return (sleep_type
== SLEEP_INTERACTIVE
||
3325 sleep_type
== SLEEP_INTERRUPTED
);
3329 * schedule() is the main scheduler function.
3331 asmlinkage
void __sched
schedule(void)
3333 struct task_struct
*prev
, *next
;
3334 struct prio_array
*array
;
3335 struct list_head
*queue
;
3336 unsigned long long now
;
3337 unsigned long run_time
;
3338 int cpu
, idx
, new_prio
;
3343 * Test if we are atomic. Since do_exit() needs to call into
3344 * schedule() atomically, we ignore that path for now.
3345 * Otherwise, whine if we are scheduling when we should not be.
3347 if (unlikely(in_atomic() && !current
->exit_state
)) {
3348 printk(KERN_ERR
"BUG: scheduling while atomic: "
3350 current
->comm
, preempt_count(), current
->pid
);
3351 debug_show_held_locks(current
);
3354 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3359 release_kernel_lock(prev
);
3360 need_resched_nonpreemptible
:
3364 * The idle thread is not allowed to schedule!
3365 * Remove this check after it has been exercised a bit.
3367 if (unlikely(prev
== rq
->idle
) && prev
->state
!= TASK_RUNNING
) {
3368 printk(KERN_ERR
"bad: scheduling from the idle thread!\n");
3372 schedstat_inc(rq
, sched_cnt
);
3373 now
= sched_clock();
3374 if (likely((long long)(now
- prev
->timestamp
) < NS_MAX_SLEEP_AVG
)) {
3375 run_time
= now
- prev
->timestamp
;
3376 if (unlikely((long long)(now
- prev
->timestamp
) < 0))
3379 run_time
= NS_MAX_SLEEP_AVG
;
3382 * Tasks charged proportionately less run_time at high sleep_avg to
3383 * delay them losing their interactive status
3385 run_time
/= (CURRENT_BONUS(prev
) ? : 1);
3387 spin_lock_irq(&rq
->lock
);
3389 switch_count
= &prev
->nivcsw
;
3390 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3391 switch_count
= &prev
->nvcsw
;
3392 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3393 unlikely(signal_pending(prev
))))
3394 prev
->state
= TASK_RUNNING
;
3396 if (prev
->state
== TASK_UNINTERRUPTIBLE
)
3397 rq
->nr_uninterruptible
++;
3398 deactivate_task(prev
, rq
);
3402 cpu
= smp_processor_id();
3403 if (unlikely(!rq
->nr_running
)) {
3404 idle_balance(cpu
, rq
);
3405 if (!rq
->nr_running
) {
3407 rq
->expired_timestamp
= 0;
3408 wake_sleeping_dependent(cpu
);
3414 if (unlikely(!array
->nr_active
)) {
3416 * Switch the active and expired arrays.
3418 schedstat_inc(rq
, sched_switch
);
3419 rq
->active
= rq
->expired
;
3420 rq
->expired
= array
;
3422 rq
->expired_timestamp
= 0;
3423 rq
->best_expired_prio
= MAX_PRIO
;
3426 idx
= sched_find_first_bit(array
->bitmap
);
3427 queue
= array
->queue
+ idx
;
3428 next
= list_entry(queue
->next
, struct task_struct
, run_list
);
3430 if (!rt_task(next
) && interactive_sleep(next
->sleep_type
)) {
3431 unsigned long long delta
= now
- next
->timestamp
;
3432 if (unlikely((long long)(now
- next
->timestamp
) < 0))
3435 if (next
->sleep_type
== SLEEP_INTERACTIVE
)
3436 delta
= delta
* (ON_RUNQUEUE_WEIGHT
* 128 / 100) / 128;
3438 array
= next
->array
;
3439 new_prio
= recalc_task_prio(next
, next
->timestamp
+ delta
);
3441 if (unlikely(next
->prio
!= new_prio
)) {
3442 dequeue_task(next
, array
);
3443 next
->prio
= new_prio
;
3444 enqueue_task(next
, array
);
3447 next
->sleep_type
= SLEEP_NORMAL
;
3448 if (dependent_sleeper(cpu
, rq
, next
))
3451 if (next
== rq
->idle
)
3452 schedstat_inc(rq
, sched_goidle
);
3454 prefetch_stack(next
);
3455 clear_tsk_need_resched(prev
);
3456 rcu_qsctr_inc(task_cpu(prev
));
3458 update_cpu_clock(prev
, rq
, now
);
3460 prev
->sleep_avg
-= run_time
;
3461 if ((long)prev
->sleep_avg
<= 0)
3462 prev
->sleep_avg
= 0;
3463 prev
->timestamp
= prev
->last_ran
= now
;
3465 sched_info_switch(prev
, next
);
3466 if (likely(prev
!= next
)) {
3467 next
->timestamp
= now
;
3472 prepare_task_switch(rq
, next
);
3473 prev
= context_switch(rq
, prev
, next
);
3476 * this_rq must be evaluated again because prev may have moved
3477 * CPUs since it called schedule(), thus the 'rq' on its stack
3478 * frame will be invalid.
3480 finish_task_switch(this_rq(), prev
);
3482 spin_unlock_irq(&rq
->lock
);
3485 if (unlikely(reacquire_kernel_lock(prev
) < 0))
3486 goto need_resched_nonpreemptible
;
3487 preempt_enable_no_resched();
3488 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3491 EXPORT_SYMBOL(schedule
);
3493 #ifdef CONFIG_PREEMPT
3495 * this is the entry point to schedule() from in-kernel preemption
3496 * off of preempt_enable. Kernel preemptions off return from interrupt
3497 * occur there and call schedule directly.
3499 asmlinkage
void __sched
preempt_schedule(void)
3501 struct thread_info
*ti
= current_thread_info();
3502 #ifdef CONFIG_PREEMPT_BKL
3503 struct task_struct
*task
= current
;
3504 int saved_lock_depth
;
3507 * If there is a non-zero preempt_count or interrupts are disabled,
3508 * we do not want to preempt the current task. Just return..
3510 if (likely(ti
->preempt_count
|| irqs_disabled()))
3514 add_preempt_count(PREEMPT_ACTIVE
);
3516 * We keep the big kernel semaphore locked, but we
3517 * clear ->lock_depth so that schedule() doesnt
3518 * auto-release the semaphore:
3520 #ifdef CONFIG_PREEMPT_BKL
3521 saved_lock_depth
= task
->lock_depth
;
3522 task
->lock_depth
= -1;
3525 #ifdef CONFIG_PREEMPT_BKL
3526 task
->lock_depth
= saved_lock_depth
;
3528 sub_preempt_count(PREEMPT_ACTIVE
);
3530 /* we could miss a preemption opportunity between schedule and now */
3532 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3535 EXPORT_SYMBOL(preempt_schedule
);
3538 * this is the entry point to schedule() from kernel preemption
3539 * off of irq context.
3540 * Note, that this is called and return with irqs disabled. This will
3541 * protect us against recursive calling from irq.
3543 asmlinkage
void __sched
preempt_schedule_irq(void)
3545 struct thread_info
*ti
= current_thread_info();
3546 #ifdef CONFIG_PREEMPT_BKL
3547 struct task_struct
*task
= current
;
3548 int saved_lock_depth
;
3550 /* Catch callers which need to be fixed */
3551 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3554 add_preempt_count(PREEMPT_ACTIVE
);
3556 * We keep the big kernel semaphore locked, but we
3557 * clear ->lock_depth so that schedule() doesnt
3558 * auto-release the semaphore:
3560 #ifdef CONFIG_PREEMPT_BKL
3561 saved_lock_depth
= task
->lock_depth
;
3562 task
->lock_depth
= -1;
3566 local_irq_disable();
3567 #ifdef CONFIG_PREEMPT_BKL
3568 task
->lock_depth
= saved_lock_depth
;
3570 sub_preempt_count(PREEMPT_ACTIVE
);
3572 /* we could miss a preemption opportunity between schedule and now */
3574 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3578 #endif /* CONFIG_PREEMPT */
3580 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3583 return try_to_wake_up(curr
->private, mode
, sync
);
3585 EXPORT_SYMBOL(default_wake_function
);
3588 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3589 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3590 * number) then we wake all the non-exclusive tasks and one exclusive task.
3592 * There are circumstances in which we can try to wake a task which has already
3593 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3594 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3596 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3597 int nr_exclusive
, int sync
, void *key
)
3599 struct list_head
*tmp
, *next
;
3601 list_for_each_safe(tmp
, next
, &q
->task_list
) {
3602 wait_queue_t
*curr
= list_entry(tmp
, wait_queue_t
, task_list
);
3603 unsigned flags
= curr
->flags
;
3605 if (curr
->func(curr
, mode
, sync
, key
) &&
3606 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3612 * __wake_up - wake up threads blocked on a waitqueue.
3614 * @mode: which threads
3615 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3616 * @key: is directly passed to the wakeup function
3618 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3619 int nr_exclusive
, void *key
)
3621 unsigned long flags
;
3623 spin_lock_irqsave(&q
->lock
, flags
);
3624 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3625 spin_unlock_irqrestore(&q
->lock
, flags
);
3627 EXPORT_SYMBOL(__wake_up
);
3630 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3632 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3634 __wake_up_common(q
, mode
, 1, 0, NULL
);
3638 * __wake_up_sync - wake up threads blocked on a waitqueue.
3640 * @mode: which threads
3641 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3643 * The sync wakeup differs that the waker knows that it will schedule
3644 * away soon, so while the target thread will be woken up, it will not
3645 * be migrated to another CPU - ie. the two threads are 'synchronized'
3646 * with each other. This can prevent needless bouncing between CPUs.
3648 * On UP it can prevent extra preemption.
3651 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3653 unsigned long flags
;
3659 if (unlikely(!nr_exclusive
))
3662 spin_lock_irqsave(&q
->lock
, flags
);
3663 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3664 spin_unlock_irqrestore(&q
->lock
, flags
);
3666 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3668 void fastcall
complete(struct completion
*x
)
3670 unsigned long flags
;
3672 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3674 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3676 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3678 EXPORT_SYMBOL(complete
);
3680 void fastcall
complete_all(struct completion
*x
)
3682 unsigned long flags
;
3684 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3685 x
->done
+= UINT_MAX
/2;
3686 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3688 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3690 EXPORT_SYMBOL(complete_all
);
3692 void fastcall __sched
wait_for_completion(struct completion
*x
)
3696 spin_lock_irq(&x
->wait
.lock
);
3698 DECLARE_WAITQUEUE(wait
, current
);
3700 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3701 __add_wait_queue_tail(&x
->wait
, &wait
);
3703 __set_current_state(TASK_UNINTERRUPTIBLE
);
3704 spin_unlock_irq(&x
->wait
.lock
);
3706 spin_lock_irq(&x
->wait
.lock
);
3708 __remove_wait_queue(&x
->wait
, &wait
);
3711 spin_unlock_irq(&x
->wait
.lock
);
3713 EXPORT_SYMBOL(wait_for_completion
);
3715 unsigned long fastcall __sched
3716 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3720 spin_lock_irq(&x
->wait
.lock
);
3722 DECLARE_WAITQUEUE(wait
, current
);
3724 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3725 __add_wait_queue_tail(&x
->wait
, &wait
);
3727 __set_current_state(TASK_UNINTERRUPTIBLE
);
3728 spin_unlock_irq(&x
->wait
.lock
);
3729 timeout
= schedule_timeout(timeout
);
3730 spin_lock_irq(&x
->wait
.lock
);
3732 __remove_wait_queue(&x
->wait
, &wait
);
3736 __remove_wait_queue(&x
->wait
, &wait
);
3740 spin_unlock_irq(&x
->wait
.lock
);
3743 EXPORT_SYMBOL(wait_for_completion_timeout
);
3745 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3751 spin_lock_irq(&x
->wait
.lock
);
3753 DECLARE_WAITQUEUE(wait
, current
);
3755 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3756 __add_wait_queue_tail(&x
->wait
, &wait
);
3758 if (signal_pending(current
)) {
3760 __remove_wait_queue(&x
->wait
, &wait
);
3763 __set_current_state(TASK_INTERRUPTIBLE
);
3764 spin_unlock_irq(&x
->wait
.lock
);
3766 spin_lock_irq(&x
->wait
.lock
);
3768 __remove_wait_queue(&x
->wait
, &wait
);
3772 spin_unlock_irq(&x
->wait
.lock
);
3776 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3778 unsigned long fastcall __sched
3779 wait_for_completion_interruptible_timeout(struct completion
*x
,
3780 unsigned long timeout
)
3784 spin_lock_irq(&x
->wait
.lock
);
3786 DECLARE_WAITQUEUE(wait
, current
);
3788 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3789 __add_wait_queue_tail(&x
->wait
, &wait
);
3791 if (signal_pending(current
)) {
3792 timeout
= -ERESTARTSYS
;
3793 __remove_wait_queue(&x
->wait
, &wait
);
3796 __set_current_state(TASK_INTERRUPTIBLE
);
3797 spin_unlock_irq(&x
->wait
.lock
);
3798 timeout
= schedule_timeout(timeout
);
3799 spin_lock_irq(&x
->wait
.lock
);
3801 __remove_wait_queue(&x
->wait
, &wait
);
3805 __remove_wait_queue(&x
->wait
, &wait
);
3809 spin_unlock_irq(&x
->wait
.lock
);
3812 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3815 #define SLEEP_ON_VAR \
3816 unsigned long flags; \
3817 wait_queue_t wait; \
3818 init_waitqueue_entry(&wait, current);
3820 #define SLEEP_ON_HEAD \
3821 spin_lock_irqsave(&q->lock,flags); \
3822 __add_wait_queue(q, &wait); \
3823 spin_unlock(&q->lock);
3825 #define SLEEP_ON_TAIL \
3826 spin_lock_irq(&q->lock); \
3827 __remove_wait_queue(q, &wait); \
3828 spin_unlock_irqrestore(&q->lock, flags);
3830 void fastcall __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3834 current
->state
= TASK_INTERRUPTIBLE
;
3840 EXPORT_SYMBOL(interruptible_sleep_on
);
3842 long fastcall __sched
3843 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3847 current
->state
= TASK_INTERRUPTIBLE
;
3850 timeout
= schedule_timeout(timeout
);
3855 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3857 void fastcall __sched
sleep_on(wait_queue_head_t
*q
)
3861 current
->state
= TASK_UNINTERRUPTIBLE
;
3867 EXPORT_SYMBOL(sleep_on
);
3869 long fastcall __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3873 current
->state
= TASK_UNINTERRUPTIBLE
;
3876 timeout
= schedule_timeout(timeout
);
3882 EXPORT_SYMBOL(sleep_on_timeout
);
3884 #ifdef CONFIG_RT_MUTEXES
3887 * rt_mutex_setprio - set the current priority of a task
3889 * @prio: prio value (kernel-internal form)
3891 * This function changes the 'effective' priority of a task. It does
3892 * not touch ->normal_prio like __setscheduler().
3894 * Used by the rt_mutex code to implement priority inheritance logic.
3896 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3898 struct prio_array
*array
;
3899 unsigned long flags
;
3903 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3905 rq
= task_rq_lock(p
, &flags
);
3910 dequeue_task(p
, array
);
3915 * If changing to an RT priority then queue it
3916 * in the active array!
3920 enqueue_task(p
, array
);
3922 * Reschedule if we are currently running on this runqueue and
3923 * our priority decreased, or if we are not currently running on
3924 * this runqueue and our priority is higher than the current's
3926 if (task_running(rq
, p
)) {
3927 if (p
->prio
> oldprio
)
3928 resched_task(rq
->curr
);
3929 } else if (TASK_PREEMPTS_CURR(p
, rq
))
3930 resched_task(rq
->curr
);
3932 task_rq_unlock(rq
, &flags
);
3937 void set_user_nice(struct task_struct
*p
, long nice
)
3939 struct prio_array
*array
;
3940 int old_prio
, delta
;
3941 unsigned long flags
;
3944 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3947 * We have to be careful, if called from sys_setpriority(),
3948 * the task might be in the middle of scheduling on another CPU.
3950 rq
= task_rq_lock(p
, &flags
);
3952 * The RT priorities are set via sched_setscheduler(), but we still
3953 * allow the 'normal' nice value to be set - but as expected
3954 * it wont have any effect on scheduling until the task is
3955 * not SCHED_NORMAL/SCHED_BATCH:
3957 if (has_rt_policy(p
)) {
3958 p
->static_prio
= NICE_TO_PRIO(nice
);
3963 dequeue_task(p
, array
);
3964 dec_raw_weighted_load(rq
, p
);
3967 p
->static_prio
= NICE_TO_PRIO(nice
);
3970 p
->prio
= effective_prio(p
);
3971 delta
= p
->prio
- old_prio
;
3974 enqueue_task(p
, array
);
3975 inc_raw_weighted_load(rq
, p
);
3977 * If the task increased its priority or is running and
3978 * lowered its priority, then reschedule its CPU:
3980 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3981 resched_task(rq
->curr
);
3984 task_rq_unlock(rq
, &flags
);
3986 EXPORT_SYMBOL(set_user_nice
);
3989 * can_nice - check if a task can reduce its nice value
3993 int can_nice(const struct task_struct
*p
, const int nice
)
3995 /* convert nice value [19,-20] to rlimit style value [1,40] */
3996 int nice_rlim
= 20 - nice
;
3998 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
3999 capable(CAP_SYS_NICE
));
4002 #ifdef __ARCH_WANT_SYS_NICE
4005 * sys_nice - change the priority of the current process.
4006 * @increment: priority increment
4008 * sys_setpriority is a more generic, but much slower function that
4009 * does similar things.
4011 asmlinkage
long sys_nice(int increment
)
4016 * Setpriority might change our priority at the same moment.
4017 * We don't have to worry. Conceptually one call occurs first
4018 * and we have a single winner.
4020 if (increment
< -40)
4025 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4031 if (increment
< 0 && !can_nice(current
, nice
))
4034 retval
= security_task_setnice(current
, nice
);
4038 set_user_nice(current
, nice
);
4045 * task_prio - return the priority value of a given task.
4046 * @p: the task in question.
4048 * This is the priority value as seen by users in /proc.
4049 * RT tasks are offset by -200. Normal tasks are centered
4050 * around 0, value goes from -16 to +15.
4052 int task_prio(const struct task_struct
*p
)
4054 return p
->prio
- MAX_RT_PRIO
;
4058 * task_nice - return the nice value of a given task.
4059 * @p: the task in question.
4061 int task_nice(const struct task_struct
*p
)
4063 return TASK_NICE(p
);
4065 EXPORT_SYMBOL_GPL(task_nice
);
4068 * idle_cpu - is a given cpu idle currently?
4069 * @cpu: the processor in question.
4071 int idle_cpu(int cpu
)
4073 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4077 * idle_task - return the idle task for a given cpu.
4078 * @cpu: the processor in question.
4080 struct task_struct
*idle_task(int cpu
)
4082 return cpu_rq(cpu
)->idle
;
4086 * find_process_by_pid - find a process with a matching PID value.
4087 * @pid: the pid in question.
4089 static inline struct task_struct
*find_process_by_pid(pid_t pid
)
4091 return pid
? find_task_by_pid(pid
) : current
;
4094 /* Actually do priority change: must hold rq lock. */
4095 static void __setscheduler(struct task_struct
*p
, int policy
, int prio
)
4100 p
->rt_priority
= prio
;
4101 p
->normal_prio
= normal_prio(p
);
4102 /* we are holding p->pi_lock already */
4103 p
->prio
= rt_mutex_getprio(p
);
4105 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
4107 if (policy
== SCHED_BATCH
)
4113 * sched_setscheduler - change the scheduling policy and/or RT priority of
4115 * @p: the task in question.
4116 * @policy: new policy.
4117 * @param: structure containing the new RT priority.
4119 * NOTE: the task may be already dead
4121 int sched_setscheduler(struct task_struct
*p
, int policy
,
4122 struct sched_param
*param
)
4124 int retval
, oldprio
, oldpolicy
= -1;
4125 struct prio_array
*array
;
4126 unsigned long flags
;
4129 /* may grab non-irq protected spin_locks */
4130 BUG_ON(in_interrupt());
4132 /* double check policy once rq lock held */
4134 policy
= oldpolicy
= p
->policy
;
4135 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4136 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
)
4139 * Valid priorities for SCHED_FIFO and SCHED_RR are
4140 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
4143 if (param
->sched_priority
< 0 ||
4144 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4145 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4147 if (is_rt_policy(policy
) != (param
->sched_priority
!= 0))
4151 * Allow unprivileged RT tasks to decrease priority:
4153 if (!capable(CAP_SYS_NICE
)) {
4154 if (is_rt_policy(policy
)) {
4155 unsigned long rlim_rtprio
;
4156 unsigned long flags
;
4158 if (!lock_task_sighand(p
, &flags
))
4160 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4161 unlock_task_sighand(p
, &flags
);
4163 /* can't set/change the rt policy */
4164 if (policy
!= p
->policy
&& !rlim_rtprio
)
4167 /* can't increase priority */
4168 if (param
->sched_priority
> p
->rt_priority
&&
4169 param
->sched_priority
> rlim_rtprio
)
4173 /* can't change other user's priorities */
4174 if ((current
->euid
!= p
->euid
) &&
4175 (current
->euid
!= p
->uid
))
4179 retval
= security_task_setscheduler(p
, policy
, param
);
4183 * make sure no PI-waiters arrive (or leave) while we are
4184 * changing the priority of the task:
4186 spin_lock_irqsave(&p
->pi_lock
, flags
);
4188 * To be able to change p->policy safely, the apropriate
4189 * runqueue lock must be held.
4191 rq
= __task_rq_lock(p
);
4192 /* recheck policy now with rq lock held */
4193 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4194 policy
= oldpolicy
= -1;
4195 __task_rq_unlock(rq
);
4196 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4201 deactivate_task(p
, rq
);
4203 __setscheduler(p
, policy
, param
->sched_priority
);
4205 __activate_task(p
, rq
);
4207 * Reschedule if we are currently running on this runqueue and
4208 * our priority decreased, or if we are not currently running on
4209 * this runqueue and our priority is higher than the current's
4211 if (task_running(rq
, p
)) {
4212 if (p
->prio
> oldprio
)
4213 resched_task(rq
->curr
);
4214 } else if (TASK_PREEMPTS_CURR(p
, rq
))
4215 resched_task(rq
->curr
);
4217 __task_rq_unlock(rq
);
4218 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4220 rt_mutex_adjust_pi(p
);
4224 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4227 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4229 struct sched_param lparam
;
4230 struct task_struct
*p
;
4233 if (!param
|| pid
< 0)
4235 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4240 p
= find_process_by_pid(pid
);
4242 retval
= sched_setscheduler(p
, policy
, &lparam
);
4249 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4250 * @pid: the pid in question.
4251 * @policy: new policy.
4252 * @param: structure containing the new RT priority.
4254 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
4255 struct sched_param __user
*param
)
4257 /* negative values for policy are not valid */
4261 return do_sched_setscheduler(pid
, policy
, param
);
4265 * sys_sched_setparam - set/change the RT priority of a thread
4266 * @pid: the pid in question.
4267 * @param: structure containing the new RT priority.
4269 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4271 return do_sched_setscheduler(pid
, -1, param
);
4275 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4276 * @pid: the pid in question.
4278 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4280 struct task_struct
*p
;
4281 int retval
= -EINVAL
;
4287 read_lock(&tasklist_lock
);
4288 p
= find_process_by_pid(pid
);
4290 retval
= security_task_getscheduler(p
);
4294 read_unlock(&tasklist_lock
);
4301 * sys_sched_getscheduler - get the RT priority of a thread
4302 * @pid: the pid in question.
4303 * @param: structure containing the RT priority.
4305 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4307 struct sched_param lp
;
4308 struct task_struct
*p
;
4309 int retval
= -EINVAL
;
4311 if (!param
|| pid
< 0)
4314 read_lock(&tasklist_lock
);
4315 p
= find_process_by_pid(pid
);
4320 retval
= security_task_getscheduler(p
);
4324 lp
.sched_priority
= p
->rt_priority
;
4325 read_unlock(&tasklist_lock
);
4328 * This one might sleep, we cannot do it with a spinlock held ...
4330 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4336 read_unlock(&tasklist_lock
);
4340 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4342 cpumask_t cpus_allowed
;
4343 struct task_struct
*p
;
4347 read_lock(&tasklist_lock
);
4349 p
= find_process_by_pid(pid
);
4351 read_unlock(&tasklist_lock
);
4352 unlock_cpu_hotplug();
4357 * It is not safe to call set_cpus_allowed with the
4358 * tasklist_lock held. We will bump the task_struct's
4359 * usage count and then drop tasklist_lock.
4362 read_unlock(&tasklist_lock
);
4365 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4366 !capable(CAP_SYS_NICE
))
4369 retval
= security_task_setscheduler(p
, 0, NULL
);
4373 cpus_allowed
= cpuset_cpus_allowed(p
);
4374 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4375 retval
= set_cpus_allowed(p
, new_mask
);
4379 unlock_cpu_hotplug();
4383 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4384 cpumask_t
*new_mask
)
4386 if (len
< sizeof(cpumask_t
)) {
4387 memset(new_mask
, 0, sizeof(cpumask_t
));
4388 } else if (len
> sizeof(cpumask_t
)) {
4389 len
= sizeof(cpumask_t
);
4391 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4395 * sys_sched_setaffinity - set the cpu affinity of a process
4396 * @pid: pid of the process
4397 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4398 * @user_mask_ptr: user-space pointer to the new cpu mask
4400 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4401 unsigned long __user
*user_mask_ptr
)
4406 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4410 return sched_setaffinity(pid
, new_mask
);
4414 * Represents all cpu's present in the system
4415 * In systems capable of hotplug, this map could dynamically grow
4416 * as new cpu's are detected in the system via any platform specific
4417 * method, such as ACPI for e.g.
4420 cpumask_t cpu_present_map __read_mostly
;
4421 EXPORT_SYMBOL(cpu_present_map
);
4424 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4425 EXPORT_SYMBOL(cpu_online_map
);
4427 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4428 EXPORT_SYMBOL(cpu_possible_map
);
4431 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4433 struct task_struct
*p
;
4437 read_lock(&tasklist_lock
);
4440 p
= find_process_by_pid(pid
);
4444 retval
= security_task_getscheduler(p
);
4448 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4451 read_unlock(&tasklist_lock
);
4452 unlock_cpu_hotplug();
4460 * sys_sched_getaffinity - get the cpu affinity of a process
4461 * @pid: pid of the process
4462 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4463 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4465 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4466 unsigned long __user
*user_mask_ptr
)
4471 if (len
< sizeof(cpumask_t
))
4474 ret
= sched_getaffinity(pid
, &mask
);
4478 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4481 return sizeof(cpumask_t
);
4485 * sys_sched_yield - yield the current processor to other threads.
4487 * this function yields the current CPU by moving the calling thread
4488 * to the expired array. If there are no other threads running on this
4489 * CPU then this function will return.
4491 asmlinkage
long sys_sched_yield(void)
4493 struct rq
*rq
= this_rq_lock();
4494 struct prio_array
*array
= current
->array
, *target
= rq
->expired
;
4496 schedstat_inc(rq
, yld_cnt
);
4498 * We implement yielding by moving the task into the expired
4501 * (special rule: RT tasks will just roundrobin in the active
4504 if (rt_task(current
))
4505 target
= rq
->active
;
4507 if (array
->nr_active
== 1) {
4508 schedstat_inc(rq
, yld_act_empty
);
4509 if (!rq
->expired
->nr_active
)
4510 schedstat_inc(rq
, yld_both_empty
);
4511 } else if (!rq
->expired
->nr_active
)
4512 schedstat_inc(rq
, yld_exp_empty
);
4514 if (array
!= target
) {
4515 dequeue_task(current
, array
);
4516 enqueue_task(current
, target
);
4519 * requeue_task is cheaper so perform that if possible.
4521 requeue_task(current
, array
);
4524 * Since we are going to call schedule() anyway, there's
4525 * no need to preempt or enable interrupts:
4527 __release(rq
->lock
);
4528 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4529 _raw_spin_unlock(&rq
->lock
);
4530 preempt_enable_no_resched();
4537 static inline int __resched_legal(int expected_preempt_count
)
4539 if (unlikely(preempt_count() != expected_preempt_count
))
4541 if (unlikely(system_state
!= SYSTEM_RUNNING
))
4546 static void __cond_resched(void)
4548 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4549 __might_sleep(__FILE__
, __LINE__
);
4552 * The BKS might be reacquired before we have dropped
4553 * PREEMPT_ACTIVE, which could trigger a second
4554 * cond_resched() call.
4557 add_preempt_count(PREEMPT_ACTIVE
);
4559 sub_preempt_count(PREEMPT_ACTIVE
);
4560 } while (need_resched());
4563 int __sched
cond_resched(void)
4565 if (need_resched() && __resched_legal(0)) {
4571 EXPORT_SYMBOL(cond_resched
);
4574 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4575 * call schedule, and on return reacquire the lock.
4577 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4578 * operations here to prevent schedule() from being called twice (once via
4579 * spin_unlock(), once by hand).
4581 int cond_resched_lock(spinlock_t
*lock
)
4585 if (need_lockbreak(lock
)) {
4591 if (need_resched() && __resched_legal(1)) {
4592 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4593 _raw_spin_unlock(lock
);
4594 preempt_enable_no_resched();
4601 EXPORT_SYMBOL(cond_resched_lock
);
4603 int __sched
cond_resched_softirq(void)
4605 BUG_ON(!in_softirq());
4607 if (need_resched() && __resched_legal(0)) {
4608 raw_local_irq_disable();
4610 raw_local_irq_enable();
4617 EXPORT_SYMBOL(cond_resched_softirq
);
4620 * yield - yield the current processor to other threads.
4622 * this is a shortcut for kernel-space yielding - it marks the
4623 * thread runnable and calls sys_sched_yield().
4625 void __sched
yield(void)
4627 set_current_state(TASK_RUNNING
);
4630 EXPORT_SYMBOL(yield
);
4633 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4634 * that process accounting knows that this is a task in IO wait state.
4636 * But don't do that if it is a deliberate, throttling IO wait (this task
4637 * has set its backing_dev_info: the queue against which it should throttle)
4639 void __sched
io_schedule(void)
4641 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4643 delayacct_blkio_start();
4644 atomic_inc(&rq
->nr_iowait
);
4646 atomic_dec(&rq
->nr_iowait
);
4647 delayacct_blkio_end();
4649 EXPORT_SYMBOL(io_schedule
);
4651 long __sched
io_schedule_timeout(long timeout
)
4653 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4656 delayacct_blkio_start();
4657 atomic_inc(&rq
->nr_iowait
);
4658 ret
= schedule_timeout(timeout
);
4659 atomic_dec(&rq
->nr_iowait
);
4660 delayacct_blkio_end();
4665 * sys_sched_get_priority_max - return maximum RT priority.
4666 * @policy: scheduling class.
4668 * this syscall returns the maximum rt_priority that can be used
4669 * by a given scheduling class.
4671 asmlinkage
long sys_sched_get_priority_max(int policy
)
4678 ret
= MAX_USER_RT_PRIO
-1;
4689 * sys_sched_get_priority_min - return minimum RT priority.
4690 * @policy: scheduling class.
4692 * this syscall returns the minimum rt_priority that can be used
4693 * by a given scheduling class.
4695 asmlinkage
long sys_sched_get_priority_min(int policy
)
4712 * sys_sched_rr_get_interval - return the default timeslice of a process.
4713 * @pid: pid of the process.
4714 * @interval: userspace pointer to the timeslice value.
4716 * this syscall writes the default timeslice value of a given process
4717 * into the user-space timespec buffer. A value of '0' means infinity.
4720 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4722 struct task_struct
*p
;
4723 int retval
= -EINVAL
;
4730 read_lock(&tasklist_lock
);
4731 p
= find_process_by_pid(pid
);
4735 retval
= security_task_getscheduler(p
);
4739 jiffies_to_timespec(p
->policy
== SCHED_FIFO
?
4740 0 : task_timeslice(p
), &t
);
4741 read_unlock(&tasklist_lock
);
4742 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4746 read_unlock(&tasklist_lock
);
4750 static inline struct task_struct
*eldest_child(struct task_struct
*p
)
4752 if (list_empty(&p
->children
))
4754 return list_entry(p
->children
.next
,struct task_struct
,sibling
);
4757 static inline struct task_struct
*older_sibling(struct task_struct
*p
)
4759 if (p
->sibling
.prev
==&p
->parent
->children
)
4761 return list_entry(p
->sibling
.prev
,struct task_struct
,sibling
);
4764 static inline struct task_struct
*younger_sibling(struct task_struct
*p
)
4766 if (p
->sibling
.next
==&p
->parent
->children
)
4768 return list_entry(p
->sibling
.next
,struct task_struct
,sibling
);
4771 static const char stat_nam
[] = "RSDTtZX";
4773 static void show_task(struct task_struct
*p
)
4775 struct task_struct
*relative
;
4776 unsigned long free
= 0;
4779 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4780 printk("%-13.13s %c", p
->comm
,
4781 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4782 #if (BITS_PER_LONG == 32)
4783 if (state
== TASK_RUNNING
)
4784 printk(" running ");
4786 printk(" %08lX ", thread_saved_pc(p
));
4788 if (state
== TASK_RUNNING
)
4789 printk(" running task ");
4791 printk(" %016lx ", thread_saved_pc(p
));
4793 #ifdef CONFIG_DEBUG_STACK_USAGE
4795 unsigned long *n
= end_of_stack(p
);
4798 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4801 printk("%5lu %5d %6d ", free
, p
->pid
, p
->parent
->pid
);
4802 if ((relative
= eldest_child(p
)))
4803 printk("%5d ", relative
->pid
);
4806 if ((relative
= younger_sibling(p
)))
4807 printk("%7d", relative
->pid
);
4810 if ((relative
= older_sibling(p
)))
4811 printk(" %5d", relative
->pid
);
4815 printk(" (L-TLB)\n");
4817 printk(" (NOTLB)\n");
4819 if (state
!= TASK_RUNNING
)
4820 show_stack(p
, NULL
);
4823 void show_state_filter(unsigned long state_filter
)
4825 struct task_struct
*g
, *p
;
4827 #if (BITS_PER_LONG == 32)
4830 printk(" task PC stack pid father child younger older\n");
4834 printk(" task PC stack pid father child younger older\n");
4836 read_lock(&tasklist_lock
);
4837 do_each_thread(g
, p
) {
4839 * reset the NMI-timeout, listing all files on a slow
4840 * console might take alot of time:
4842 touch_nmi_watchdog();
4843 if (p
->state
& state_filter
)
4845 } while_each_thread(g
, p
);
4847 read_unlock(&tasklist_lock
);
4849 * Only show locks if all tasks are dumped:
4851 if (state_filter
== -1)
4852 debug_show_all_locks();
4856 * init_idle - set up an idle thread for a given CPU
4857 * @idle: task in question
4858 * @cpu: cpu the idle task belongs to
4860 * NOTE: this function does not set the idle thread's NEED_RESCHED
4861 * flag, to make booting more robust.
4863 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4865 struct rq
*rq
= cpu_rq(cpu
);
4866 unsigned long flags
;
4868 idle
->timestamp
= sched_clock();
4869 idle
->sleep_avg
= 0;
4871 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
4872 idle
->state
= TASK_RUNNING
;
4873 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4874 set_task_cpu(idle
, cpu
);
4876 spin_lock_irqsave(&rq
->lock
, flags
);
4877 rq
->curr
= rq
->idle
= idle
;
4878 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4881 spin_unlock_irqrestore(&rq
->lock
, flags
);
4883 /* Set the preempt count _outside_ the spinlocks! */
4884 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4885 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4887 task_thread_info(idle
)->preempt_count
= 0;
4892 * In a system that switches off the HZ timer nohz_cpu_mask
4893 * indicates which cpus entered this state. This is used
4894 * in the rcu update to wait only for active cpus. For system
4895 * which do not switch off the HZ timer nohz_cpu_mask should
4896 * always be CPU_MASK_NONE.
4898 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4902 * This is how migration works:
4904 * 1) we queue a struct migration_req structure in the source CPU's
4905 * runqueue and wake up that CPU's migration thread.
4906 * 2) we down() the locked semaphore => thread blocks.
4907 * 3) migration thread wakes up (implicitly it forces the migrated
4908 * thread off the CPU)
4909 * 4) it gets the migration request and checks whether the migrated
4910 * task is still in the wrong runqueue.
4911 * 5) if it's in the wrong runqueue then the migration thread removes
4912 * it and puts it into the right queue.
4913 * 6) migration thread up()s the semaphore.
4914 * 7) we wake up and the migration is done.
4918 * Change a given task's CPU affinity. Migrate the thread to a
4919 * proper CPU and schedule it away if the CPU it's executing on
4920 * is removed from the allowed bitmask.
4922 * NOTE: the caller must have a valid reference to the task, the
4923 * task must not exit() & deallocate itself prematurely. The
4924 * call is not atomic; no spinlocks may be held.
4926 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
4928 struct migration_req req
;
4929 unsigned long flags
;
4933 rq
= task_rq_lock(p
, &flags
);
4934 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4939 p
->cpus_allowed
= new_mask
;
4940 /* Can the task run on the task's current CPU? If so, we're done */
4941 if (cpu_isset(task_cpu(p
), new_mask
))
4944 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4945 /* Need help from migration thread: drop lock and wait. */
4946 task_rq_unlock(rq
, &flags
);
4947 wake_up_process(rq
->migration_thread
);
4948 wait_for_completion(&req
.done
);
4949 tlb_migrate_finish(p
->mm
);
4953 task_rq_unlock(rq
, &flags
);
4957 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4960 * Move (not current) task off this cpu, onto dest cpu. We're doing
4961 * this because either it can't run here any more (set_cpus_allowed()
4962 * away from this CPU, or CPU going down), or because we're
4963 * attempting to rebalance this task on exec (sched_exec).
4965 * So we race with normal scheduler movements, but that's OK, as long
4966 * as the task is no longer on this CPU.
4968 * Returns non-zero if task was successfully migrated.
4970 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4972 struct rq
*rq_dest
, *rq_src
;
4975 if (unlikely(cpu_is_offline(dest_cpu
)))
4978 rq_src
= cpu_rq(src_cpu
);
4979 rq_dest
= cpu_rq(dest_cpu
);
4981 double_rq_lock(rq_src
, rq_dest
);
4982 /* Already moved. */
4983 if (task_cpu(p
) != src_cpu
)
4985 /* Affinity changed (again). */
4986 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
4989 set_task_cpu(p
, dest_cpu
);
4992 * Sync timestamp with rq_dest's before activating.
4993 * The same thing could be achieved by doing this step
4994 * afterwards, and pretending it was a local activate.
4995 * This way is cleaner and logically correct.
4997 p
->timestamp
= p
->timestamp
- rq_src
->timestamp_last_tick
4998 + rq_dest
->timestamp_last_tick
;
4999 deactivate_task(p
, rq_src
);
5000 __activate_task(p
, rq_dest
);
5001 if (TASK_PREEMPTS_CURR(p
, rq_dest
))
5002 resched_task(rq_dest
->curr
);
5006 double_rq_unlock(rq_src
, rq_dest
);
5011 * migration_thread - this is a highprio system thread that performs
5012 * thread migration by bumping thread off CPU then 'pushing' onto
5015 static int migration_thread(void *data
)
5017 int cpu
= (long)data
;
5021 BUG_ON(rq
->migration_thread
!= current
);
5023 set_current_state(TASK_INTERRUPTIBLE
);
5024 while (!kthread_should_stop()) {
5025 struct migration_req
*req
;
5026 struct list_head
*head
;
5030 spin_lock_irq(&rq
->lock
);
5032 if (cpu_is_offline(cpu
)) {
5033 spin_unlock_irq(&rq
->lock
);
5037 if (rq
->active_balance
) {
5038 active_load_balance(rq
, cpu
);
5039 rq
->active_balance
= 0;
5042 head
= &rq
->migration_queue
;
5044 if (list_empty(head
)) {
5045 spin_unlock_irq(&rq
->lock
);
5047 set_current_state(TASK_INTERRUPTIBLE
);
5050 req
= list_entry(head
->next
, struct migration_req
, list
);
5051 list_del_init(head
->next
);
5053 spin_unlock(&rq
->lock
);
5054 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5057 complete(&req
->done
);
5059 __set_current_state(TASK_RUNNING
);
5063 /* Wait for kthread_stop */
5064 set_current_state(TASK_INTERRUPTIBLE
);
5065 while (!kthread_should_stop()) {
5067 set_current_state(TASK_INTERRUPTIBLE
);
5069 __set_current_state(TASK_RUNNING
);
5073 #ifdef CONFIG_HOTPLUG_CPU
5075 * Figure out where task on dead CPU should go, use force if neccessary.
5076 * NOTE: interrupts should be disabled by the caller
5078 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5080 unsigned long flags
;
5087 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5088 cpus_and(mask
, mask
, p
->cpus_allowed
);
5089 dest_cpu
= any_online_cpu(mask
);
5091 /* On any allowed CPU? */
5092 if (dest_cpu
== NR_CPUS
)
5093 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5095 /* No more Mr. Nice Guy. */
5096 if (dest_cpu
== NR_CPUS
) {
5097 rq
= task_rq_lock(p
, &flags
);
5098 cpus_setall(p
->cpus_allowed
);
5099 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5100 task_rq_unlock(rq
, &flags
);
5103 * Don't tell them about moving exiting tasks or
5104 * kernel threads (both mm NULL), since they never
5107 if (p
->mm
&& printk_ratelimit())
5108 printk(KERN_INFO
"process %d (%s) no "
5109 "longer affine to cpu%d\n",
5110 p
->pid
, p
->comm
, dead_cpu
);
5112 if (!__migrate_task(p
, dead_cpu
, dest_cpu
))
5117 * While a dead CPU has no uninterruptible tasks queued at this point,
5118 * it might still have a nonzero ->nr_uninterruptible counter, because
5119 * for performance reasons the counter is not stricly tracking tasks to
5120 * their home CPUs. So we just add the counter to another CPU's counter,
5121 * to keep the global sum constant after CPU-down:
5123 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5125 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5126 unsigned long flags
;
5128 local_irq_save(flags
);
5129 double_rq_lock(rq_src
, rq_dest
);
5130 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5131 rq_src
->nr_uninterruptible
= 0;
5132 double_rq_unlock(rq_src
, rq_dest
);
5133 local_irq_restore(flags
);
5136 /* Run through task list and migrate tasks from the dead cpu. */
5137 static void migrate_live_tasks(int src_cpu
)
5139 struct task_struct
*p
, *t
;
5141 write_lock_irq(&tasklist_lock
);
5143 do_each_thread(t
, p
) {
5147 if (task_cpu(p
) == src_cpu
)
5148 move_task_off_dead_cpu(src_cpu
, p
);
5149 } while_each_thread(t
, p
);
5151 write_unlock_irq(&tasklist_lock
);
5154 /* Schedules idle task to be the next runnable task on current CPU.
5155 * It does so by boosting its priority to highest possible and adding it to
5156 * the _front_ of the runqueue. Used by CPU offline code.
5158 void sched_idle_next(void)
5160 int this_cpu
= smp_processor_id();
5161 struct rq
*rq
= cpu_rq(this_cpu
);
5162 struct task_struct
*p
= rq
->idle
;
5163 unsigned long flags
;
5165 /* cpu has to be offline */
5166 BUG_ON(cpu_online(this_cpu
));
5169 * Strictly not necessary since rest of the CPUs are stopped by now
5170 * and interrupts disabled on the current cpu.
5172 spin_lock_irqsave(&rq
->lock
, flags
);
5174 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5176 /* Add idle task to the _front_ of its priority queue: */
5177 __activate_idle_task(p
, rq
);
5179 spin_unlock_irqrestore(&rq
->lock
, flags
);
5183 * Ensures that the idle task is using init_mm right before its cpu goes
5186 void idle_task_exit(void)
5188 struct mm_struct
*mm
= current
->active_mm
;
5190 BUG_ON(cpu_online(smp_processor_id()));
5193 switch_mm(mm
, &init_mm
, current
);
5197 /* called under rq->lock with disabled interrupts */
5198 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5200 struct rq
*rq
= cpu_rq(dead_cpu
);
5202 /* Must be exiting, otherwise would be on tasklist. */
5203 BUG_ON(p
->exit_state
!= EXIT_ZOMBIE
&& p
->exit_state
!= EXIT_DEAD
);
5205 /* Cannot have done final schedule yet: would have vanished. */
5206 BUG_ON(p
->state
== TASK_DEAD
);
5211 * Drop lock around migration; if someone else moves it,
5212 * that's OK. No task can be added to this CPU, so iteration is
5214 * NOTE: interrupts should be left disabled --dev@
5216 spin_unlock(&rq
->lock
);
5217 move_task_off_dead_cpu(dead_cpu
, p
);
5218 spin_lock(&rq
->lock
);
5223 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5224 static void migrate_dead_tasks(unsigned int dead_cpu
)
5226 struct rq
*rq
= cpu_rq(dead_cpu
);
5227 unsigned int arr
, i
;
5229 for (arr
= 0; arr
< 2; arr
++) {
5230 for (i
= 0; i
< MAX_PRIO
; i
++) {
5231 struct list_head
*list
= &rq
->arrays
[arr
].queue
[i
];
5233 while (!list_empty(list
))
5234 migrate_dead(dead_cpu
, list_entry(list
->next
,
5235 struct task_struct
, run_list
));
5239 #endif /* CONFIG_HOTPLUG_CPU */
5242 * migration_call - callback that gets triggered when a CPU is added.
5243 * Here we can start up the necessary migration thread for the new CPU.
5245 static int __cpuinit
5246 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5248 struct task_struct
*p
;
5249 int cpu
= (long)hcpu
;
5250 unsigned long flags
;
5254 case CPU_UP_PREPARE
:
5255 p
= kthread_create(migration_thread
, hcpu
, "migration/%d",cpu
);
5258 p
->flags
|= PF_NOFREEZE
;
5259 kthread_bind(p
, cpu
);
5260 /* Must be high prio: stop_machine expects to yield to it. */
5261 rq
= task_rq_lock(p
, &flags
);
5262 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5263 task_rq_unlock(rq
, &flags
);
5264 cpu_rq(cpu
)->migration_thread
= p
;
5268 /* Strictly unneccessary, as first user will wake it. */
5269 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5272 #ifdef CONFIG_HOTPLUG_CPU
5273 case CPU_UP_CANCELED
:
5274 if (!cpu_rq(cpu
)->migration_thread
)
5276 /* Unbind it from offline cpu so it can run. Fall thru. */
5277 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5278 any_online_cpu(cpu_online_map
));
5279 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5280 cpu_rq(cpu
)->migration_thread
= NULL
;
5284 migrate_live_tasks(cpu
);
5286 kthread_stop(rq
->migration_thread
);
5287 rq
->migration_thread
= NULL
;
5288 /* Idle task back to normal (off runqueue, low prio) */
5289 rq
= task_rq_lock(rq
->idle
, &flags
);
5290 deactivate_task(rq
->idle
, rq
);
5291 rq
->idle
->static_prio
= MAX_PRIO
;
5292 __setscheduler(rq
->idle
, SCHED_NORMAL
, 0);
5293 migrate_dead_tasks(cpu
);
5294 task_rq_unlock(rq
, &flags
);
5295 migrate_nr_uninterruptible(rq
);
5296 BUG_ON(rq
->nr_running
!= 0);
5298 /* No need to migrate the tasks: it was best-effort if
5299 * they didn't do lock_cpu_hotplug(). Just wake up
5300 * the requestors. */
5301 spin_lock_irq(&rq
->lock
);
5302 while (!list_empty(&rq
->migration_queue
)) {
5303 struct migration_req
*req
;
5305 req
= list_entry(rq
->migration_queue
.next
,
5306 struct migration_req
, list
);
5307 list_del_init(&req
->list
);
5308 complete(&req
->done
);
5310 spin_unlock_irq(&rq
->lock
);
5317 /* Register at highest priority so that task migration (migrate_all_tasks)
5318 * happens before everything else.
5320 static struct notifier_block __cpuinitdata migration_notifier
= {
5321 .notifier_call
= migration_call
,
5325 int __init
migration_init(void)
5327 void *cpu
= (void *)(long)smp_processor_id();
5330 /* Start one for the boot CPU: */
5331 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5332 BUG_ON(err
== NOTIFY_BAD
);
5333 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5334 register_cpu_notifier(&migration_notifier
);
5341 #undef SCHED_DOMAIN_DEBUG
5342 #ifdef SCHED_DOMAIN_DEBUG
5343 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5348 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5352 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5357 struct sched_group
*group
= sd
->groups
;
5358 cpumask_t groupmask
;
5360 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5361 cpus_clear(groupmask
);
5364 for (i
= 0; i
< level
+ 1; i
++)
5366 printk("domain %d: ", level
);
5368 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5369 printk("does not load-balance\n");
5371 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain has parent");
5375 printk("span %s\n", str
);
5377 if (!cpu_isset(cpu
, sd
->span
))
5378 printk(KERN_ERR
"ERROR: domain->span does not contain CPU%d\n", cpu
);
5379 if (!cpu_isset(cpu
, group
->cpumask
))
5380 printk(KERN_ERR
"ERROR: domain->groups does not contain CPU%d\n", cpu
);
5383 for (i
= 0; i
< level
+ 2; i
++)
5389 printk(KERN_ERR
"ERROR: group is NULL\n");
5393 if (!group
->cpu_power
) {
5395 printk(KERN_ERR
"ERROR: domain->cpu_power not set\n");
5398 if (!cpus_weight(group
->cpumask
)) {
5400 printk(KERN_ERR
"ERROR: empty group\n");
5403 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5405 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5408 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5410 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5413 group
= group
->next
;
5414 } while (group
!= sd
->groups
);
5417 if (!cpus_equal(sd
->span
, groupmask
))
5418 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5424 if (!cpus_subset(groupmask
, sd
->span
))
5425 printk(KERN_ERR
"ERROR: parent span is not a superset of domain->span\n");
5431 # define sched_domain_debug(sd, cpu) do { } while (0)
5434 static int sd_degenerate(struct sched_domain
*sd
)
5436 if (cpus_weight(sd
->span
) == 1)
5439 /* Following flags need at least 2 groups */
5440 if (sd
->flags
& (SD_LOAD_BALANCE
|
5441 SD_BALANCE_NEWIDLE
|
5445 SD_SHARE_PKG_RESOURCES
)) {
5446 if (sd
->groups
!= sd
->groups
->next
)
5450 /* Following flags don't use groups */
5451 if (sd
->flags
& (SD_WAKE_IDLE
|
5460 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5462 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5464 if (sd_degenerate(parent
))
5467 if (!cpus_equal(sd
->span
, parent
->span
))
5470 /* Does parent contain flags not in child? */
5471 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5472 if (cflags
& SD_WAKE_AFFINE
)
5473 pflags
&= ~SD_WAKE_BALANCE
;
5474 /* Flags needing groups don't count if only 1 group in parent */
5475 if (parent
->groups
== parent
->groups
->next
) {
5476 pflags
&= ~(SD_LOAD_BALANCE
|
5477 SD_BALANCE_NEWIDLE
|
5481 SD_SHARE_PKG_RESOURCES
);
5483 if (~cflags
& pflags
)
5490 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5491 * hold the hotplug lock.
5493 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5495 struct rq
*rq
= cpu_rq(cpu
);
5496 struct sched_domain
*tmp
;
5498 /* Remove the sched domains which do not contribute to scheduling. */
5499 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5500 struct sched_domain
*parent
= tmp
->parent
;
5503 if (sd_parent_degenerate(tmp
, parent
)) {
5504 tmp
->parent
= parent
->parent
;
5506 parent
->parent
->child
= tmp
;
5510 if (sd
&& sd_degenerate(sd
)) {
5516 sched_domain_debug(sd
, cpu
);
5518 rcu_assign_pointer(rq
->sd
, sd
);
5521 /* cpus with isolated domains */
5522 static cpumask_t __cpuinitdata cpu_isolated_map
= CPU_MASK_NONE
;
5524 /* Setup the mask of cpus configured for isolated domains */
5525 static int __init
isolated_cpu_setup(char *str
)
5527 int ints
[NR_CPUS
], i
;
5529 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5530 cpus_clear(cpu_isolated_map
);
5531 for (i
= 1; i
<= ints
[0]; i
++)
5532 if (ints
[i
] < NR_CPUS
)
5533 cpu_set(ints
[i
], cpu_isolated_map
);
5537 __setup ("isolcpus=", isolated_cpu_setup
);
5540 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5541 * to a function which identifies what group(along with sched group) a CPU
5542 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5543 * (due to the fact that we keep track of groups covered with a cpumask_t).
5545 * init_sched_build_groups will build a circular linked list of the groups
5546 * covered by the given span, and will set each group's ->cpumask correctly,
5547 * and ->cpu_power to 0.
5550 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
5551 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
5552 struct sched_group
**sg
))
5554 struct sched_group
*first
= NULL
, *last
= NULL
;
5555 cpumask_t covered
= CPU_MASK_NONE
;
5558 for_each_cpu_mask(i
, span
) {
5559 struct sched_group
*sg
;
5560 int group
= group_fn(i
, cpu_map
, &sg
);
5563 if (cpu_isset(i
, covered
))
5566 sg
->cpumask
= CPU_MASK_NONE
;
5569 for_each_cpu_mask(j
, span
) {
5570 if (group_fn(j
, cpu_map
, NULL
) != group
)
5573 cpu_set(j
, covered
);
5574 cpu_set(j
, sg
->cpumask
);
5585 #define SD_NODES_PER_DOMAIN 16
5588 * Self-tuning task migration cost measurement between source and target CPUs.
5590 * This is done by measuring the cost of manipulating buffers of varying
5591 * sizes. For a given buffer-size here are the steps that are taken:
5593 * 1) the source CPU reads+dirties a shared buffer
5594 * 2) the target CPU reads+dirties the same shared buffer
5596 * We measure how long they take, in the following 4 scenarios:
5598 * - source: CPU1, target: CPU2 | cost1
5599 * - source: CPU2, target: CPU1 | cost2
5600 * - source: CPU1, target: CPU1 | cost3
5601 * - source: CPU2, target: CPU2 | cost4
5603 * We then calculate the cost3+cost4-cost1-cost2 difference - this is
5604 * the cost of migration.
5606 * We then start off from a small buffer-size and iterate up to larger
5607 * buffer sizes, in 5% steps - measuring each buffer-size separately, and
5608 * doing a maximum search for the cost. (The maximum cost for a migration
5609 * normally occurs when the working set size is around the effective cache
5612 #define SEARCH_SCOPE 2
5613 #define MIN_CACHE_SIZE (64*1024U)
5614 #define DEFAULT_CACHE_SIZE (5*1024*1024U)
5615 #define ITERATIONS 1
5616 #define SIZE_THRESH 130
5617 #define COST_THRESH 130
5620 * The migration cost is a function of 'domain distance'. Domain
5621 * distance is the number of steps a CPU has to iterate down its
5622 * domain tree to share a domain with the other CPU. The farther
5623 * two CPUs are from each other, the larger the distance gets.
5625 * Note that we use the distance only to cache measurement results,
5626 * the distance value is not used numerically otherwise. When two
5627 * CPUs have the same distance it is assumed that the migration
5628 * cost is the same. (this is a simplification but quite practical)
5630 #define MAX_DOMAIN_DISTANCE 32
5632 static unsigned long long migration_cost
[MAX_DOMAIN_DISTANCE
] =
5633 { [ 0 ... MAX_DOMAIN_DISTANCE
-1 ] =
5635 * Architectures may override the migration cost and thus avoid
5636 * boot-time calibration. Unit is nanoseconds. Mostly useful for
5637 * virtualized hardware:
5639 #ifdef CONFIG_DEFAULT_MIGRATION_COST
5640 CONFIG_DEFAULT_MIGRATION_COST
5647 * Allow override of migration cost - in units of microseconds.
5648 * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost
5649 * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs:
5651 static int __init
migration_cost_setup(char *str
)
5653 int ints
[MAX_DOMAIN_DISTANCE
+1], i
;
5655 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5657 printk("#ints: %d\n", ints
[0]);
5658 for (i
= 1; i
<= ints
[0]; i
++) {
5659 migration_cost
[i
-1] = (unsigned long long)ints
[i
]*1000;
5660 printk("migration_cost[%d]: %Ld\n", i
-1, migration_cost
[i
-1]);
5665 __setup ("migration_cost=", migration_cost_setup
);
5668 * Global multiplier (divisor) for migration-cutoff values,
5669 * in percentiles. E.g. use a value of 150 to get 1.5 times
5670 * longer cache-hot cutoff times.
5672 * (We scale it from 100 to 128 to long long handling easier.)
5675 #define MIGRATION_FACTOR_SCALE 128
5677 static unsigned int migration_factor
= MIGRATION_FACTOR_SCALE
;
5679 static int __init
setup_migration_factor(char *str
)
5681 get_option(&str
, &migration_factor
);
5682 migration_factor
= migration_factor
* MIGRATION_FACTOR_SCALE
/ 100;
5686 __setup("migration_factor=", setup_migration_factor
);
5689 * Estimated distance of two CPUs, measured via the number of domains
5690 * we have to pass for the two CPUs to be in the same span:
5692 static unsigned long domain_distance(int cpu1
, int cpu2
)
5694 unsigned long distance
= 0;
5695 struct sched_domain
*sd
;
5697 for_each_domain(cpu1
, sd
) {
5698 WARN_ON(!cpu_isset(cpu1
, sd
->span
));
5699 if (cpu_isset(cpu2
, sd
->span
))
5703 if (distance
>= MAX_DOMAIN_DISTANCE
) {
5705 distance
= MAX_DOMAIN_DISTANCE
-1;
5711 static unsigned int migration_debug
;
5713 static int __init
setup_migration_debug(char *str
)
5715 get_option(&str
, &migration_debug
);
5719 __setup("migration_debug=", setup_migration_debug
);
5722 * Maximum cache-size that the scheduler should try to measure.
5723 * Architectures with larger caches should tune this up during
5724 * bootup. Gets used in the domain-setup code (i.e. during SMP
5727 unsigned int max_cache_size
;
5729 static int __init
setup_max_cache_size(char *str
)
5731 get_option(&str
, &max_cache_size
);
5735 __setup("max_cache_size=", setup_max_cache_size
);
5738 * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This
5739 * is the operation that is timed, so we try to generate unpredictable
5740 * cachemisses that still end up filling the L2 cache:
5742 static void touch_cache(void *__cache
, unsigned long __size
)
5744 unsigned long size
= __size
/sizeof(long), chunk1
= size
/3,
5746 unsigned long *cache
= __cache
;
5749 for (i
= 0; i
< size
/6; i
+= 8) {
5752 case 1: cache
[size
-1-i
]++;
5753 case 2: cache
[chunk1
-i
]++;
5754 case 3: cache
[chunk1
+i
]++;
5755 case 4: cache
[chunk2
-i
]++;
5756 case 5: cache
[chunk2
+i
]++;
5762 * Measure the cache-cost of one task migration. Returns in units of nsec.
5764 static unsigned long long
5765 measure_one(void *cache
, unsigned long size
, int source
, int target
)
5767 cpumask_t mask
, saved_mask
;
5768 unsigned long long t0
, t1
, t2
, t3
, cost
;
5770 saved_mask
= current
->cpus_allowed
;
5773 * Flush source caches to RAM and invalidate them:
5778 * Migrate to the source CPU:
5780 mask
= cpumask_of_cpu(source
);
5781 set_cpus_allowed(current
, mask
);
5782 WARN_ON(smp_processor_id() != source
);
5785 * Dirty the working set:
5788 touch_cache(cache
, size
);
5792 * Migrate to the target CPU, dirty the L2 cache and access
5793 * the shared buffer. (which represents the working set
5794 * of a migrated task.)
5796 mask
= cpumask_of_cpu(target
);
5797 set_cpus_allowed(current
, mask
);
5798 WARN_ON(smp_processor_id() != target
);
5801 touch_cache(cache
, size
);
5804 cost
= t1
-t0
+ t3
-t2
;
5806 if (migration_debug
>= 2)
5807 printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n",
5808 source
, target
, t1
-t0
, t1
-t0
, t3
-t2
, cost
);
5810 * Flush target caches to RAM and invalidate them:
5814 set_cpus_allowed(current
, saved_mask
);
5820 * Measure a series of task migrations and return the average
5821 * result. Since this code runs early during bootup the system
5822 * is 'undisturbed' and the average latency makes sense.
5824 * The algorithm in essence auto-detects the relevant cache-size,
5825 * so it will properly detect different cachesizes for different
5826 * cache-hierarchies, depending on how the CPUs are connected.
5828 * Architectures can prime the upper limit of the search range via
5829 * max_cache_size, otherwise the search range defaults to 20MB...64K.
5831 static unsigned long long
5832 measure_cost(int cpu1
, int cpu2
, void *cache
, unsigned int size
)
5834 unsigned long long cost1
, cost2
;
5838 * Measure the migration cost of 'size' bytes, over an
5839 * average of 10 runs:
5841 * (We perturb the cache size by a small (0..4k)
5842 * value to compensate size/alignment related artifacts.
5843 * We also subtract the cost of the operation done on
5849 * dry run, to make sure we start off cache-cold on cpu1,
5850 * and to get any vmalloc pagefaults in advance:
5852 measure_one(cache
, size
, cpu1
, cpu2
);
5853 for (i
= 0; i
< ITERATIONS
; i
++)
5854 cost1
+= measure_one(cache
, size
- i
*1024, cpu1
, cpu2
);
5856 measure_one(cache
, size
, cpu2
, cpu1
);
5857 for (i
= 0; i
< ITERATIONS
; i
++)
5858 cost1
+= measure_one(cache
, size
- i
*1024, cpu2
, cpu1
);
5861 * (We measure the non-migrating [cached] cost on both
5862 * cpu1 and cpu2, to handle CPUs with different speeds)
5866 measure_one(cache
, size
, cpu1
, cpu1
);
5867 for (i
= 0; i
< ITERATIONS
; i
++)
5868 cost2
+= measure_one(cache
, size
- i
*1024, cpu1
, cpu1
);
5870 measure_one(cache
, size
, cpu2
, cpu2
);
5871 for (i
= 0; i
< ITERATIONS
; i
++)
5872 cost2
+= measure_one(cache
, size
- i
*1024, cpu2
, cpu2
);
5875 * Get the per-iteration migration cost:
5877 do_div(cost1
, 2*ITERATIONS
);
5878 do_div(cost2
, 2*ITERATIONS
);
5880 return cost1
- cost2
;
5883 static unsigned long long measure_migration_cost(int cpu1
, int cpu2
)
5885 unsigned long long max_cost
= 0, fluct
= 0, avg_fluct
= 0;
5886 unsigned int max_size
, size
, size_found
= 0;
5887 long long cost
= 0, prev_cost
;
5891 * Search from max_cache_size*5 down to 64K - the real relevant
5892 * cachesize has to lie somewhere inbetween.
5894 if (max_cache_size
) {
5895 max_size
= max(max_cache_size
* SEARCH_SCOPE
, MIN_CACHE_SIZE
);
5896 size
= max(max_cache_size
/ SEARCH_SCOPE
, MIN_CACHE_SIZE
);
5899 * Since we have no estimation about the relevant
5902 max_size
= DEFAULT_CACHE_SIZE
* SEARCH_SCOPE
;
5903 size
= MIN_CACHE_SIZE
;
5906 if (!cpu_online(cpu1
) || !cpu_online(cpu2
)) {
5907 printk("cpu %d and %d not both online!\n", cpu1
, cpu2
);
5912 * Allocate the working set:
5914 cache
= vmalloc(max_size
);
5916 printk("could not vmalloc %d bytes for cache!\n", 2*max_size
);
5917 return 1000000; /* return 1 msec on very small boxen */
5920 while (size
<= max_size
) {
5922 cost
= measure_cost(cpu1
, cpu2
, cache
, size
);
5928 if (max_cost
< cost
) {
5934 * Calculate average fluctuation, we use this to prevent
5935 * noise from triggering an early break out of the loop:
5937 fluct
= abs(cost
- prev_cost
);
5938 avg_fluct
= (avg_fluct
+ fluct
)/2;
5940 if (migration_debug
)
5941 printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): (%8Ld %8Ld)\n",
5943 (long)cost
/ 1000000,
5944 ((long)cost
/ 100000) % 10,
5945 (long)max_cost
/ 1000000,
5946 ((long)max_cost
/ 100000) % 10,
5947 domain_distance(cpu1
, cpu2
),
5951 * If we iterated at least 20% past the previous maximum,
5952 * and the cost has dropped by more than 20% already,
5953 * (taking fluctuations into account) then we assume to
5954 * have found the maximum and break out of the loop early:
5956 if (size_found
&& (size
*100 > size_found
*SIZE_THRESH
))
5957 if (cost
+avg_fluct
<= 0 ||
5958 max_cost
*100 > (cost
+avg_fluct
)*COST_THRESH
) {
5960 if (migration_debug
)
5961 printk("-> found max.\n");
5965 * Increase the cachesize in 10% steps:
5967 size
= size
* 10 / 9;
5970 if (migration_debug
)
5971 printk("[%d][%d] working set size found: %d, cost: %Ld\n",
5972 cpu1
, cpu2
, size_found
, max_cost
);
5977 * A task is considered 'cache cold' if at least 2 times
5978 * the worst-case cost of migration has passed.
5980 * (this limit is only listened to if the load-balancing
5981 * situation is 'nice' - if there is a large imbalance we
5982 * ignore it for the sake of CPU utilization and
5983 * processing fairness.)
5985 return 2 * max_cost
* migration_factor
/ MIGRATION_FACTOR_SCALE
;
5988 static void calibrate_migration_costs(const cpumask_t
*cpu_map
)
5990 int cpu1
= -1, cpu2
= -1, cpu
, orig_cpu
= raw_smp_processor_id();
5991 unsigned long j0
, j1
, distance
, max_distance
= 0;
5992 struct sched_domain
*sd
;
5997 * First pass - calculate the cacheflush times:
5999 for_each_cpu_mask(cpu1
, *cpu_map
) {
6000 for_each_cpu_mask(cpu2
, *cpu_map
) {
6003 distance
= domain_distance(cpu1
, cpu2
);
6004 max_distance
= max(max_distance
, distance
);
6006 * No result cached yet?
6008 if (migration_cost
[distance
] == -1LL)
6009 migration_cost
[distance
] =
6010 measure_migration_cost(cpu1
, cpu2
);
6014 * Second pass - update the sched domain hierarchy with
6015 * the new cache-hot-time estimations:
6017 for_each_cpu_mask(cpu
, *cpu_map
) {
6019 for_each_domain(cpu
, sd
) {
6020 sd
->cache_hot_time
= migration_cost
[distance
];
6027 if (migration_debug
)
6028 printk("migration: max_cache_size: %d, cpu: %d MHz:\n",
6036 if (system_state
== SYSTEM_BOOTING
) {
6037 if (num_online_cpus() > 1) {
6038 printk("migration_cost=");
6039 for (distance
= 0; distance
<= max_distance
; distance
++) {
6042 printk("%ld", (long)migration_cost
[distance
] / 1000);
6048 if (migration_debug
)
6049 printk("migration: %ld seconds\n", (j1
-j0
)/HZ
);
6052 * Move back to the original CPU. NUMA-Q gets confused
6053 * if we migrate to another quad during bootup.
6055 if (raw_smp_processor_id() != orig_cpu
) {
6056 cpumask_t mask
= cpumask_of_cpu(orig_cpu
),
6057 saved_mask
= current
->cpus_allowed
;
6059 set_cpus_allowed(current
, mask
);
6060 set_cpus_allowed(current
, saved_mask
);
6067 * find_next_best_node - find the next node to include in a sched_domain
6068 * @node: node whose sched_domain we're building
6069 * @used_nodes: nodes already in the sched_domain
6071 * Find the next node to include in a given scheduling domain. Simply
6072 * finds the closest node not already in the @used_nodes map.
6074 * Should use nodemask_t.
6076 static int find_next_best_node(int node
, unsigned long *used_nodes
)
6078 int i
, n
, val
, min_val
, best_node
= 0;
6082 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6083 /* Start at @node */
6084 n
= (node
+ i
) % MAX_NUMNODES
;
6086 if (!nr_cpus_node(n
))
6089 /* Skip already used nodes */
6090 if (test_bit(n
, used_nodes
))
6093 /* Simple min distance search */
6094 val
= node_distance(node
, n
);
6096 if (val
< min_val
) {
6102 set_bit(best_node
, used_nodes
);
6107 * sched_domain_node_span - get a cpumask for a node's sched_domain
6108 * @node: node whose cpumask we're constructing
6109 * @size: number of nodes to include in this span
6111 * Given a node, construct a good cpumask for its sched_domain to span. It
6112 * should be one that prevents unnecessary balancing, but also spreads tasks
6115 static cpumask_t
sched_domain_node_span(int node
)
6117 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
6118 cpumask_t span
, nodemask
;
6122 bitmap_zero(used_nodes
, MAX_NUMNODES
);
6124 nodemask
= node_to_cpumask(node
);
6125 cpus_or(span
, span
, nodemask
);
6126 set_bit(node
, used_nodes
);
6128 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6129 int next_node
= find_next_best_node(node
, used_nodes
);
6131 nodemask
= node_to_cpumask(next_node
);
6132 cpus_or(span
, span
, nodemask
);
6139 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6142 * SMT sched-domains:
6144 #ifdef CONFIG_SCHED_SMT
6145 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
6146 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
6148 static int cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
,
6149 struct sched_group
**sg
)
6152 *sg
= &per_cpu(sched_group_cpus
, cpu
);
6158 * multi-core sched-domains:
6160 #ifdef CONFIG_SCHED_MC
6161 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
6162 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
6165 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6166 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
6167 struct sched_group
**sg
)
6170 cpumask_t mask
= cpu_sibling_map
[cpu
];
6171 cpus_and(mask
, mask
, *cpu_map
);
6172 group
= first_cpu(mask
);
6174 *sg
= &per_cpu(sched_group_core
, group
);
6177 #elif defined(CONFIG_SCHED_MC)
6178 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
6179 struct sched_group
**sg
)
6182 *sg
= &per_cpu(sched_group_core
, cpu
);
6187 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
6188 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
6190 static int cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
,
6191 struct sched_group
**sg
)
6194 #ifdef CONFIG_SCHED_MC
6195 cpumask_t mask
= cpu_coregroup_map(cpu
);
6196 cpus_and(mask
, mask
, *cpu_map
);
6197 group
= first_cpu(mask
);
6198 #elif defined(CONFIG_SCHED_SMT)
6199 cpumask_t mask
= cpu_sibling_map
[cpu
];
6200 cpus_and(mask
, mask
, *cpu_map
);
6201 group
= first_cpu(mask
);
6206 *sg
= &per_cpu(sched_group_phys
, group
);
6212 * The init_sched_build_groups can't handle what we want to do with node
6213 * groups, so roll our own. Now each node has its own list of groups which
6214 * gets dynamically allocated.
6216 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
6217 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
6219 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
6220 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
6222 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
6223 struct sched_group
**sg
)
6225 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
6228 cpus_and(nodemask
, nodemask
, *cpu_map
);
6229 group
= first_cpu(nodemask
);
6232 *sg
= &per_cpu(sched_group_allnodes
, group
);
6236 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6238 struct sched_group
*sg
= group_head
;
6244 for_each_cpu_mask(j
, sg
->cpumask
) {
6245 struct sched_domain
*sd
;
6247 sd
= &per_cpu(phys_domains
, j
);
6248 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
6250 * Only add "power" once for each
6256 sg
->cpu_power
+= sd
->groups
->cpu_power
;
6259 if (sg
!= group_head
)
6265 /* Free memory allocated for various sched_group structures */
6266 static void free_sched_groups(const cpumask_t
*cpu_map
)
6270 for_each_cpu_mask(cpu
, *cpu_map
) {
6271 struct sched_group
**sched_group_nodes
6272 = sched_group_nodes_bycpu
[cpu
];
6274 if (!sched_group_nodes
)
6277 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6278 cpumask_t nodemask
= node_to_cpumask(i
);
6279 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6281 cpus_and(nodemask
, nodemask
, *cpu_map
);
6282 if (cpus_empty(nodemask
))
6292 if (oldsg
!= sched_group_nodes
[i
])
6295 kfree(sched_group_nodes
);
6296 sched_group_nodes_bycpu
[cpu
] = NULL
;
6300 static void free_sched_groups(const cpumask_t
*cpu_map
)
6306 * Initialize sched groups cpu_power.
6308 * cpu_power indicates the capacity of sched group, which is used while
6309 * distributing the load between different sched groups in a sched domain.
6310 * Typically cpu_power for all the groups in a sched domain will be same unless
6311 * there are asymmetries in the topology. If there are asymmetries, group
6312 * having more cpu_power will pickup more load compared to the group having
6315 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6316 * the maximum number of tasks a group can handle in the presence of other idle
6317 * or lightly loaded groups in the same sched domain.
6319 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6321 struct sched_domain
*child
;
6322 struct sched_group
*group
;
6324 WARN_ON(!sd
|| !sd
->groups
);
6326 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
6332 * For perf policy, if the groups in child domain share resources
6333 * (for example cores sharing some portions of the cache hierarchy
6334 * or SMT), then set this domain groups cpu_power such that each group
6335 * can handle only one task, when there are other idle groups in the
6336 * same sched domain.
6338 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
6340 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
6341 sd
->groups
->cpu_power
= SCHED_LOAD_SCALE
;
6345 sd
->groups
->cpu_power
= 0;
6348 * add cpu_power of each child group to this groups cpu_power
6350 group
= child
->groups
;
6352 sd
->groups
->cpu_power
+= group
->cpu_power
;
6353 group
= group
->next
;
6354 } while (group
!= child
->groups
);
6358 * Build sched domains for a given set of cpus and attach the sched domains
6359 * to the individual cpus
6361 static int build_sched_domains(const cpumask_t
*cpu_map
)
6364 struct sched_domain
*sd
;
6366 struct sched_group
**sched_group_nodes
= NULL
;
6367 int sd_allnodes
= 0;
6370 * Allocate the per-node list of sched groups
6372 sched_group_nodes
= kzalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
6374 if (!sched_group_nodes
) {
6375 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6378 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6382 * Set up domains for cpus specified by the cpu_map.
6384 for_each_cpu_mask(i
, *cpu_map
) {
6385 struct sched_domain
*sd
= NULL
, *p
;
6386 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6388 cpus_and(nodemask
, nodemask
, *cpu_map
);
6391 if (cpus_weight(*cpu_map
)
6392 > SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6393 sd
= &per_cpu(allnodes_domains
, i
);
6394 *sd
= SD_ALLNODES_INIT
;
6395 sd
->span
= *cpu_map
;
6396 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6402 sd
= &per_cpu(node_domains
, i
);
6404 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6408 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6412 sd
= &per_cpu(phys_domains
, i
);
6414 sd
->span
= nodemask
;
6418 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6420 #ifdef CONFIG_SCHED_MC
6422 sd
= &per_cpu(core_domains
, i
);
6424 sd
->span
= cpu_coregroup_map(i
);
6425 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6428 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6431 #ifdef CONFIG_SCHED_SMT
6433 sd
= &per_cpu(cpu_domains
, i
);
6434 *sd
= SD_SIBLING_INIT
;
6435 sd
->span
= cpu_sibling_map
[i
];
6436 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6439 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6443 #ifdef CONFIG_SCHED_SMT
6444 /* Set up CPU (sibling) groups */
6445 for_each_cpu_mask(i
, *cpu_map
) {
6446 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
6447 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6448 if (i
!= first_cpu(this_sibling_map
))
6451 init_sched_build_groups(this_sibling_map
, cpu_map
, &cpu_to_cpu_group
);
6455 #ifdef CONFIG_SCHED_MC
6456 /* Set up multi-core groups */
6457 for_each_cpu_mask(i
, *cpu_map
) {
6458 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6459 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6460 if (i
!= first_cpu(this_core_map
))
6462 init_sched_build_groups(this_core_map
, cpu_map
, &cpu_to_core_group
);
6467 /* Set up physical groups */
6468 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6469 cpumask_t nodemask
= node_to_cpumask(i
);
6471 cpus_and(nodemask
, nodemask
, *cpu_map
);
6472 if (cpus_empty(nodemask
))
6475 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6479 /* Set up node groups */
6481 init_sched_build_groups(*cpu_map
, cpu_map
, &cpu_to_allnodes_group
);
6483 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6484 /* Set up node groups */
6485 struct sched_group
*sg
, *prev
;
6486 cpumask_t nodemask
= node_to_cpumask(i
);
6487 cpumask_t domainspan
;
6488 cpumask_t covered
= CPU_MASK_NONE
;
6491 cpus_and(nodemask
, nodemask
, *cpu_map
);
6492 if (cpus_empty(nodemask
)) {
6493 sched_group_nodes
[i
] = NULL
;
6497 domainspan
= sched_domain_node_span(i
);
6498 cpus_and(domainspan
, domainspan
, *cpu_map
);
6500 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6502 printk(KERN_WARNING
"Can not alloc domain group for "
6506 sched_group_nodes
[i
] = sg
;
6507 for_each_cpu_mask(j
, nodemask
) {
6508 struct sched_domain
*sd
;
6509 sd
= &per_cpu(node_domains
, j
);
6513 sg
->cpumask
= nodemask
;
6515 cpus_or(covered
, covered
, nodemask
);
6518 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6519 cpumask_t tmp
, notcovered
;
6520 int n
= (i
+ j
) % MAX_NUMNODES
;
6522 cpus_complement(notcovered
, covered
);
6523 cpus_and(tmp
, notcovered
, *cpu_map
);
6524 cpus_and(tmp
, tmp
, domainspan
);
6525 if (cpus_empty(tmp
))
6528 nodemask
= node_to_cpumask(n
);
6529 cpus_and(tmp
, tmp
, nodemask
);
6530 if (cpus_empty(tmp
))
6533 sg
= kmalloc_node(sizeof(struct sched_group
),
6537 "Can not alloc domain group for node %d\n", j
);
6542 sg
->next
= prev
->next
;
6543 cpus_or(covered
, covered
, tmp
);
6550 /* Calculate CPU power for physical packages and nodes */
6551 #ifdef CONFIG_SCHED_SMT
6552 for_each_cpu_mask(i
, *cpu_map
) {
6553 sd
= &per_cpu(cpu_domains
, i
);
6554 init_sched_groups_power(i
, sd
);
6557 #ifdef CONFIG_SCHED_MC
6558 for_each_cpu_mask(i
, *cpu_map
) {
6559 sd
= &per_cpu(core_domains
, i
);
6560 init_sched_groups_power(i
, sd
);
6564 for_each_cpu_mask(i
, *cpu_map
) {
6565 sd
= &per_cpu(phys_domains
, i
);
6566 init_sched_groups_power(i
, sd
);
6570 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6571 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6574 struct sched_group
*sg
;
6576 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6577 init_numa_sched_groups_power(sg
);
6581 /* Attach the domains */
6582 for_each_cpu_mask(i
, *cpu_map
) {
6583 struct sched_domain
*sd
;
6584 #ifdef CONFIG_SCHED_SMT
6585 sd
= &per_cpu(cpu_domains
, i
);
6586 #elif defined(CONFIG_SCHED_MC)
6587 sd
= &per_cpu(core_domains
, i
);
6589 sd
= &per_cpu(phys_domains
, i
);
6591 cpu_attach_domain(sd
, i
);
6594 * Tune cache-hot values:
6596 calibrate_migration_costs(cpu_map
);
6602 free_sched_groups(cpu_map
);
6607 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6609 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6611 cpumask_t cpu_default_map
;
6615 * Setup mask for cpus without special case scheduling requirements.
6616 * For now this just excludes isolated cpus, but could be used to
6617 * exclude other special cases in the future.
6619 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
6621 err
= build_sched_domains(&cpu_default_map
);
6626 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6628 free_sched_groups(cpu_map
);
6632 * Detach sched domains from a group of cpus specified in cpu_map
6633 * These cpus will now be attached to the NULL domain
6635 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6639 for_each_cpu_mask(i
, *cpu_map
)
6640 cpu_attach_domain(NULL
, i
);
6641 synchronize_sched();
6642 arch_destroy_sched_domains(cpu_map
);
6646 * Partition sched domains as specified by the cpumasks below.
6647 * This attaches all cpus from the cpumasks to the NULL domain,
6648 * waits for a RCU quiescent period, recalculates sched
6649 * domain information and then attaches them back to the
6650 * correct sched domains
6651 * Call with hotplug lock held
6653 int partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
6655 cpumask_t change_map
;
6658 cpus_and(*partition1
, *partition1
, cpu_online_map
);
6659 cpus_and(*partition2
, *partition2
, cpu_online_map
);
6660 cpus_or(change_map
, *partition1
, *partition2
);
6662 /* Detach sched domains from all of the affected cpus */
6663 detach_destroy_domains(&change_map
);
6664 if (!cpus_empty(*partition1
))
6665 err
= build_sched_domains(partition1
);
6666 if (!err
&& !cpus_empty(*partition2
))
6667 err
= build_sched_domains(partition2
);
6672 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6673 int arch_reinit_sched_domains(void)
6678 detach_destroy_domains(&cpu_online_map
);
6679 err
= arch_init_sched_domains(&cpu_online_map
);
6680 unlock_cpu_hotplug();
6685 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6689 if (buf
[0] != '0' && buf
[0] != '1')
6693 sched_smt_power_savings
= (buf
[0] == '1');
6695 sched_mc_power_savings
= (buf
[0] == '1');
6697 ret
= arch_reinit_sched_domains();
6699 return ret
? ret
: count
;
6702 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6706 #ifdef CONFIG_SCHED_SMT
6708 err
= sysfs_create_file(&cls
->kset
.kobj
,
6709 &attr_sched_smt_power_savings
.attr
);
6711 #ifdef CONFIG_SCHED_MC
6712 if (!err
&& mc_capable())
6713 err
= sysfs_create_file(&cls
->kset
.kobj
,
6714 &attr_sched_mc_power_savings
.attr
);
6720 #ifdef CONFIG_SCHED_MC
6721 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6723 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6725 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6726 const char *buf
, size_t count
)
6728 return sched_power_savings_store(buf
, count
, 0);
6730 SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6731 sched_mc_power_savings_store
);
6734 #ifdef CONFIG_SCHED_SMT
6735 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6737 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6739 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6740 const char *buf
, size_t count
)
6742 return sched_power_savings_store(buf
, count
, 1);
6744 SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6745 sched_smt_power_savings_store
);
6749 * Force a reinitialization of the sched domains hierarchy. The domains
6750 * and groups cannot be updated in place without racing with the balancing
6751 * code, so we temporarily attach all running cpus to the NULL domain
6752 * which will prevent rebalancing while the sched domains are recalculated.
6754 static int update_sched_domains(struct notifier_block
*nfb
,
6755 unsigned long action
, void *hcpu
)
6758 case CPU_UP_PREPARE
:
6759 case CPU_DOWN_PREPARE
:
6760 detach_destroy_domains(&cpu_online_map
);
6763 case CPU_UP_CANCELED
:
6764 case CPU_DOWN_FAILED
:
6768 * Fall through and re-initialise the domains.
6775 /* The hotplug lock is already held by cpu_up/cpu_down */
6776 arch_init_sched_domains(&cpu_online_map
);
6781 void __init
sched_init_smp(void)
6783 cpumask_t non_isolated_cpus
;
6786 arch_init_sched_domains(&cpu_online_map
);
6787 cpus_andnot(non_isolated_cpus
, cpu_online_map
, cpu_isolated_map
);
6788 if (cpus_empty(non_isolated_cpus
))
6789 cpu_set(smp_processor_id(), non_isolated_cpus
);
6790 unlock_cpu_hotplug();
6791 /* XXX: Theoretical race here - CPU may be hotplugged now */
6792 hotcpu_notifier(update_sched_domains
, 0);
6794 /* Move init over to a non-isolated CPU */
6795 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
6799 void __init
sched_init_smp(void)
6802 #endif /* CONFIG_SMP */
6804 int in_sched_functions(unsigned long addr
)
6806 /* Linker adds these: start and end of __sched functions */
6807 extern char __sched_text_start
[], __sched_text_end
[];
6809 return in_lock_functions(addr
) ||
6810 (addr
>= (unsigned long)__sched_text_start
6811 && addr
< (unsigned long)__sched_text_end
);
6814 void __init
sched_init(void)
6818 for_each_possible_cpu(i
) {
6819 struct prio_array
*array
;
6823 spin_lock_init(&rq
->lock
);
6824 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
6826 rq
->active
= rq
->arrays
;
6827 rq
->expired
= rq
->arrays
+ 1;
6828 rq
->best_expired_prio
= MAX_PRIO
;
6832 for (j
= 1; j
< 3; j
++)
6833 rq
->cpu_load
[j
] = 0;
6834 rq
->active_balance
= 0;
6837 rq
->migration_thread
= NULL
;
6838 INIT_LIST_HEAD(&rq
->migration_queue
);
6840 atomic_set(&rq
->nr_iowait
, 0);
6842 for (j
= 0; j
< 2; j
++) {
6843 array
= rq
->arrays
+ j
;
6844 for (k
= 0; k
< MAX_PRIO
; k
++) {
6845 INIT_LIST_HEAD(array
->queue
+ k
);
6846 __clear_bit(k
, array
->bitmap
);
6848 // delimiter for bitsearch
6849 __set_bit(MAX_PRIO
, array
->bitmap
);
6853 set_load_weight(&init_task
);
6855 #ifdef CONFIG_RT_MUTEXES
6856 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
6860 * The boot idle thread does lazy MMU switching as well:
6862 atomic_inc(&init_mm
.mm_count
);
6863 enter_lazy_tlb(&init_mm
, current
);
6866 * Make us the idle thread. Technically, schedule() should not be
6867 * called from this thread, however somewhere below it might be,
6868 * but because we are the idle thread, we just pick up running again
6869 * when this runqueue becomes "idle".
6871 init_idle(current
, smp_processor_id());
6874 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6875 void __might_sleep(char *file
, int line
)
6878 static unsigned long prev_jiffy
; /* ratelimiting */
6880 if ((in_atomic() || irqs_disabled()) &&
6881 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6882 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6884 prev_jiffy
= jiffies
;
6885 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6886 " context at %s:%d\n", file
, line
);
6887 printk("in_atomic():%d, irqs_disabled():%d\n",
6888 in_atomic(), irqs_disabled());
6889 debug_show_held_locks(current
);
6894 EXPORT_SYMBOL(__might_sleep
);
6897 #ifdef CONFIG_MAGIC_SYSRQ
6898 void normalize_rt_tasks(void)
6900 struct prio_array
*array
;
6901 struct task_struct
*p
;
6902 unsigned long flags
;
6905 read_lock_irq(&tasklist_lock
);
6906 for_each_process(p
) {
6910 spin_lock_irqsave(&p
->pi_lock
, flags
);
6911 rq
= __task_rq_lock(p
);
6915 deactivate_task(p
, task_rq(p
));
6916 __setscheduler(p
, SCHED_NORMAL
, 0);
6918 __activate_task(p
, task_rq(p
));
6919 resched_task(rq
->curr
);
6922 __task_rq_unlock(rq
);
6923 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6925 read_unlock_irq(&tasklist_lock
);
6928 #endif /* CONFIG_MAGIC_SYSRQ */
6932 * These functions are only useful for the IA64 MCA handling.
6934 * They can only be called when the whole system has been
6935 * stopped - every CPU needs to be quiescent, and no scheduling
6936 * activity can take place. Using them for anything else would
6937 * be a serious bug, and as a result, they aren't even visible
6938 * under any other configuration.
6942 * curr_task - return the current task for a given cpu.
6943 * @cpu: the processor in question.
6945 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6947 struct task_struct
*curr_task(int cpu
)
6949 return cpu_curr(cpu
);
6953 * set_curr_task - set the current task for a given cpu.
6954 * @cpu: the processor in question.
6955 * @p: the task pointer to set.
6957 * Description: This function must only be used when non-maskable interrupts
6958 * are serviced on a separate stack. It allows the architecture to switch the
6959 * notion of the current task on a cpu in a non-blocking manner. This function
6960 * must be called with all CPU's synchronized, and interrupts disabled, the
6961 * and caller must save the original value of the current task (see
6962 * curr_task() above) and restore that value before reenabling interrupts and
6963 * re-starting the system.
6965 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6967 void set_curr_task(int cpu
, struct task_struct
*p
)