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 unsigned long next_balance
;
231 struct mm_struct
*prev_mm
;
232 struct prio_array
*active
, *expired
, arrays
[2];
233 int best_expired_prio
;
237 struct sched_domain
*sd
;
239 /* For active balancing */
242 int cpu
; /* cpu of this runqueue */
244 struct task_struct
*migration_thread
;
245 struct list_head migration_queue
;
248 #ifdef CONFIG_SCHEDSTATS
250 struct sched_info rq_sched_info
;
252 /* sys_sched_yield() stats */
253 unsigned long yld_exp_empty
;
254 unsigned long yld_act_empty
;
255 unsigned long yld_both_empty
;
256 unsigned long yld_cnt
;
258 /* schedule() stats */
259 unsigned long sched_switch
;
260 unsigned long sched_cnt
;
261 unsigned long sched_goidle
;
263 /* try_to_wake_up() stats */
264 unsigned long ttwu_cnt
;
265 unsigned long ttwu_local
;
267 struct lock_class_key rq_lock_key
;
270 static DEFINE_PER_CPU(struct rq
, runqueues
);
272 static inline int cpu_of(struct rq
*rq
)
282 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
283 * See detach_destroy_domains: synchronize_sched for details.
285 * The domain tree of any CPU may only be accessed from within
286 * preempt-disabled sections.
288 #define for_each_domain(cpu, __sd) \
289 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
291 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
292 #define this_rq() (&__get_cpu_var(runqueues))
293 #define task_rq(p) cpu_rq(task_cpu(p))
294 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
296 #ifndef prepare_arch_switch
297 # define prepare_arch_switch(next) do { } while (0)
299 #ifndef finish_arch_switch
300 # define finish_arch_switch(prev) do { } while (0)
303 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
304 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
306 return rq
->curr
== p
;
309 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
313 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
315 #ifdef CONFIG_DEBUG_SPINLOCK
316 /* this is a valid case when another task releases the spinlock */
317 rq
->lock
.owner
= current
;
320 * If we are tracking spinlock dependencies then we have to
321 * fix up the runqueue lock - which gets 'carried over' from
324 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
326 spin_unlock_irq(&rq
->lock
);
329 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
330 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
335 return rq
->curr
== p
;
339 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
343 * We can optimise this out completely for !SMP, because the
344 * SMP rebalancing from interrupt is the only thing that cares
349 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
350 spin_unlock_irq(&rq
->lock
);
352 spin_unlock(&rq
->lock
);
356 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
360 * After ->oncpu is cleared, the task can be moved to a different CPU.
361 * We must ensure this doesn't happen until the switch is completely
367 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
371 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
374 * __task_rq_lock - lock the runqueue a given task resides on.
375 * Must be called interrupts disabled.
377 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
384 spin_lock(&rq
->lock
);
385 if (unlikely(rq
!= task_rq(p
))) {
386 spin_unlock(&rq
->lock
);
387 goto repeat_lock_task
;
393 * task_rq_lock - lock the runqueue a given task resides on and disable
394 * interrupts. Note the ordering: we can safely lookup the task_rq without
395 * explicitly disabling preemption.
397 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
403 local_irq_save(*flags
);
405 spin_lock(&rq
->lock
);
406 if (unlikely(rq
!= task_rq(p
))) {
407 spin_unlock_irqrestore(&rq
->lock
, *flags
);
408 goto repeat_lock_task
;
413 static inline void __task_rq_unlock(struct rq
*rq
)
416 spin_unlock(&rq
->lock
);
419 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
422 spin_unlock_irqrestore(&rq
->lock
, *flags
);
425 #ifdef CONFIG_SCHEDSTATS
427 * bump this up when changing the output format or the meaning of an existing
428 * format, so that tools can adapt (or abort)
430 #define SCHEDSTAT_VERSION 12
432 static int show_schedstat(struct seq_file
*seq
, void *v
)
436 seq_printf(seq
, "version %d\n", SCHEDSTAT_VERSION
);
437 seq_printf(seq
, "timestamp %lu\n", jiffies
);
438 for_each_online_cpu(cpu
) {
439 struct rq
*rq
= cpu_rq(cpu
);
441 struct sched_domain
*sd
;
445 /* runqueue-specific stats */
447 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
448 cpu
, rq
->yld_both_empty
,
449 rq
->yld_act_empty
, rq
->yld_exp_empty
, rq
->yld_cnt
,
450 rq
->sched_switch
, rq
->sched_cnt
, rq
->sched_goidle
,
451 rq
->ttwu_cnt
, rq
->ttwu_local
,
452 rq
->rq_sched_info
.cpu_time
,
453 rq
->rq_sched_info
.run_delay
, rq
->rq_sched_info
.pcnt
);
455 seq_printf(seq
, "\n");
458 /* domain-specific stats */
460 for_each_domain(cpu
, sd
) {
461 enum idle_type itype
;
462 char mask_str
[NR_CPUS
];
464 cpumask_scnprintf(mask_str
, NR_CPUS
, sd
->span
);
465 seq_printf(seq
, "domain%d %s", dcnt
++, mask_str
);
466 for (itype
= SCHED_IDLE
; itype
< MAX_IDLE_TYPES
;
468 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu %lu",
470 sd
->lb_balanced
[itype
],
471 sd
->lb_failed
[itype
],
472 sd
->lb_imbalance
[itype
],
473 sd
->lb_gained
[itype
],
474 sd
->lb_hot_gained
[itype
],
475 sd
->lb_nobusyq
[itype
],
476 sd
->lb_nobusyg
[itype
]);
478 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
479 sd
->alb_cnt
, sd
->alb_failed
, sd
->alb_pushed
,
480 sd
->sbe_cnt
, sd
->sbe_balanced
, sd
->sbe_pushed
,
481 sd
->sbf_cnt
, sd
->sbf_balanced
, sd
->sbf_pushed
,
482 sd
->ttwu_wake_remote
, sd
->ttwu_move_affine
, sd
->ttwu_move_balance
);
490 static int schedstat_open(struct inode
*inode
, struct file
*file
)
492 unsigned int size
= PAGE_SIZE
* (1 + num_online_cpus() / 32);
493 char *buf
= kmalloc(size
, GFP_KERNEL
);
499 res
= single_open(file
, show_schedstat
, NULL
);
501 m
= file
->private_data
;
509 const struct file_operations proc_schedstat_operations
= {
510 .open
= schedstat_open
,
513 .release
= single_release
,
517 * Expects runqueue lock to be held for atomicity of update
520 rq_sched_info_arrive(struct rq
*rq
, unsigned long delta_jiffies
)
523 rq
->rq_sched_info
.run_delay
+= delta_jiffies
;
524 rq
->rq_sched_info
.pcnt
++;
529 * Expects runqueue lock to be held for atomicity of update
532 rq_sched_info_depart(struct rq
*rq
, unsigned long delta_jiffies
)
535 rq
->rq_sched_info
.cpu_time
+= delta_jiffies
;
537 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
538 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
539 #else /* !CONFIG_SCHEDSTATS */
541 rq_sched_info_arrive(struct rq
*rq
, unsigned long delta_jiffies
)
544 rq_sched_info_depart(struct rq
*rq
, unsigned long delta_jiffies
)
546 # define schedstat_inc(rq, field) do { } while (0)
547 # define schedstat_add(rq, field, amt) do { } while (0)
551 * this_rq_lock - lock this runqueue and disable interrupts.
553 static inline struct rq
*this_rq_lock(void)
560 spin_lock(&rq
->lock
);
565 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
567 * Called when a process is dequeued from the active array and given
568 * the cpu. We should note that with the exception of interactive
569 * tasks, the expired queue will become the active queue after the active
570 * queue is empty, without explicitly dequeuing and requeuing tasks in the
571 * expired queue. (Interactive tasks may be requeued directly to the
572 * active queue, thus delaying tasks in the expired queue from running;
573 * see scheduler_tick()).
575 * This function is only called from sched_info_arrive(), rather than
576 * dequeue_task(). Even though a task may be queued and dequeued multiple
577 * times as it is shuffled about, we're really interested in knowing how
578 * long it was from the *first* time it was queued to the time that it
581 static inline void sched_info_dequeued(struct task_struct
*t
)
583 t
->sched_info
.last_queued
= 0;
587 * Called when a task finally hits the cpu. We can now calculate how
588 * long it was waiting to run. We also note when it began so that we
589 * can keep stats on how long its timeslice is.
591 static void sched_info_arrive(struct task_struct
*t
)
593 unsigned long now
= jiffies
, delta_jiffies
= 0;
595 if (t
->sched_info
.last_queued
)
596 delta_jiffies
= now
- t
->sched_info
.last_queued
;
597 sched_info_dequeued(t
);
598 t
->sched_info
.run_delay
+= delta_jiffies
;
599 t
->sched_info
.last_arrival
= now
;
600 t
->sched_info
.pcnt
++;
602 rq_sched_info_arrive(task_rq(t
), delta_jiffies
);
606 * Called when a process is queued into either the active or expired
607 * array. The time is noted and later used to determine how long we
608 * had to wait for us to reach the cpu. Since the expired queue will
609 * become the active queue after active queue is empty, without dequeuing
610 * and requeuing any tasks, we are interested in queuing to either. It
611 * is unusual but not impossible for tasks to be dequeued and immediately
612 * requeued in the same or another array: this can happen in sched_yield(),
613 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
616 * This function is only called from enqueue_task(), but also only updates
617 * the timestamp if it is already not set. It's assumed that
618 * sched_info_dequeued() will clear that stamp when appropriate.
620 static inline void sched_info_queued(struct task_struct
*t
)
622 if (unlikely(sched_info_on()))
623 if (!t
->sched_info
.last_queued
)
624 t
->sched_info
.last_queued
= jiffies
;
628 * Called when a process ceases being the active-running process, either
629 * voluntarily or involuntarily. Now we can calculate how long we ran.
631 static inline void sched_info_depart(struct task_struct
*t
)
633 unsigned long delta_jiffies
= jiffies
- t
->sched_info
.last_arrival
;
635 t
->sched_info
.cpu_time
+= delta_jiffies
;
636 rq_sched_info_depart(task_rq(t
), delta_jiffies
);
640 * Called when tasks are switched involuntarily due, typically, to expiring
641 * their time slice. (This may also be called when switching to or from
642 * the idle task.) We are only called when prev != next.
645 __sched_info_switch(struct task_struct
*prev
, struct task_struct
*next
)
647 struct rq
*rq
= task_rq(prev
);
650 * prev now departs the cpu. It's not interesting to record
651 * stats about how efficient we were at scheduling the idle
654 if (prev
!= rq
->idle
)
655 sched_info_depart(prev
);
657 if (next
!= rq
->idle
)
658 sched_info_arrive(next
);
661 sched_info_switch(struct task_struct
*prev
, struct task_struct
*next
)
663 if (unlikely(sched_info_on()))
664 __sched_info_switch(prev
, next
);
667 #define sched_info_queued(t) do { } while (0)
668 #define sched_info_switch(t, next) do { } while (0)
669 #endif /* CONFIG_SCHEDSTATS || CONFIG_TASK_DELAY_ACCT */
672 * Adding/removing a task to/from a priority array:
674 static void dequeue_task(struct task_struct
*p
, struct prio_array
*array
)
677 list_del(&p
->run_list
);
678 if (list_empty(array
->queue
+ p
->prio
))
679 __clear_bit(p
->prio
, array
->bitmap
);
682 static void enqueue_task(struct task_struct
*p
, struct prio_array
*array
)
684 sched_info_queued(p
);
685 list_add_tail(&p
->run_list
, array
->queue
+ p
->prio
);
686 __set_bit(p
->prio
, array
->bitmap
);
692 * Put task to the end of the run list without the overhead of dequeue
693 * followed by enqueue.
695 static void requeue_task(struct task_struct
*p
, struct prio_array
*array
)
697 list_move_tail(&p
->run_list
, array
->queue
+ p
->prio
);
701 enqueue_task_head(struct task_struct
*p
, struct prio_array
*array
)
703 list_add(&p
->run_list
, array
->queue
+ p
->prio
);
704 __set_bit(p
->prio
, array
->bitmap
);
710 * __normal_prio - return the priority that is based on the static
711 * priority but is modified by bonuses/penalties.
713 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
714 * into the -5 ... 0 ... +5 bonus/penalty range.
716 * We use 25% of the full 0...39 priority range so that:
718 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
719 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
721 * Both properties are important to certain workloads.
724 static inline int __normal_prio(struct task_struct
*p
)
728 bonus
= CURRENT_BONUS(p
) - MAX_BONUS
/ 2;
730 prio
= p
->static_prio
- bonus
;
731 if (prio
< MAX_RT_PRIO
)
733 if (prio
> MAX_PRIO
-1)
739 * To aid in avoiding the subversion of "niceness" due to uneven distribution
740 * of tasks with abnormal "nice" values across CPUs the contribution that
741 * each task makes to its run queue's load is weighted according to its
742 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
743 * scaled version of the new time slice allocation that they receive on time
748 * Assume: static_prio_timeslice(NICE_TO_PRIO(0)) == DEF_TIMESLICE
749 * If static_prio_timeslice() is ever changed to break this assumption then
750 * this code will need modification
752 #define TIME_SLICE_NICE_ZERO DEF_TIMESLICE
753 #define LOAD_WEIGHT(lp) \
754 (((lp) * SCHED_LOAD_SCALE) / TIME_SLICE_NICE_ZERO)
755 #define PRIO_TO_LOAD_WEIGHT(prio) \
756 LOAD_WEIGHT(static_prio_timeslice(prio))
757 #define RTPRIO_TO_LOAD_WEIGHT(rp) \
758 (PRIO_TO_LOAD_WEIGHT(MAX_RT_PRIO) + LOAD_WEIGHT(rp))
760 static void set_load_weight(struct task_struct
*p
)
762 if (has_rt_policy(p
)) {
764 if (p
== task_rq(p
)->migration_thread
)
766 * The migration thread does the actual balancing.
767 * Giving its load any weight will skew balancing
773 p
->load_weight
= RTPRIO_TO_LOAD_WEIGHT(p
->rt_priority
);
775 p
->load_weight
= PRIO_TO_LOAD_WEIGHT(p
->static_prio
);
779 inc_raw_weighted_load(struct rq
*rq
, const struct task_struct
*p
)
781 rq
->raw_weighted_load
+= p
->load_weight
;
785 dec_raw_weighted_load(struct rq
*rq
, const struct task_struct
*p
)
787 rq
->raw_weighted_load
-= p
->load_weight
;
790 static inline void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
793 inc_raw_weighted_load(rq
, p
);
796 static inline void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
799 dec_raw_weighted_load(rq
, p
);
803 * Calculate the expected normal priority: i.e. priority
804 * without taking RT-inheritance into account. Might be
805 * boosted by interactivity modifiers. Changes upon fork,
806 * setprio syscalls, and whenever the interactivity
807 * estimator recalculates.
809 static inline int normal_prio(struct task_struct
*p
)
813 if (has_rt_policy(p
))
814 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
816 prio
= __normal_prio(p
);
821 * Calculate the current priority, i.e. the priority
822 * taken into account by the scheduler. This value might
823 * be boosted by RT tasks, or might be boosted by
824 * interactivity modifiers. Will be RT if the task got
825 * RT-boosted. If not then it returns p->normal_prio.
827 static int effective_prio(struct task_struct
*p
)
829 p
->normal_prio
= normal_prio(p
);
831 * If we are RT tasks or we were boosted to RT priority,
832 * keep the priority unchanged. Otherwise, update priority
833 * to the normal priority:
835 if (!rt_prio(p
->prio
))
836 return p
->normal_prio
;
841 * __activate_task - move a task to the runqueue.
843 static void __activate_task(struct task_struct
*p
, struct rq
*rq
)
845 struct prio_array
*target
= rq
->active
;
848 target
= rq
->expired
;
849 enqueue_task(p
, target
);
850 inc_nr_running(p
, rq
);
854 * __activate_idle_task - move idle task to the _front_ of runqueue.
856 static inline void __activate_idle_task(struct task_struct
*p
, struct rq
*rq
)
858 enqueue_task_head(p
, rq
->active
);
859 inc_nr_running(p
, rq
);
863 * Recalculate p->normal_prio and p->prio after having slept,
864 * updating the sleep-average too:
866 static int recalc_task_prio(struct task_struct
*p
, unsigned long long now
)
868 /* Caller must always ensure 'now >= p->timestamp' */
869 unsigned long sleep_time
= now
- p
->timestamp
;
874 if (likely(sleep_time
> 0)) {
876 * This ceiling is set to the lowest priority that would allow
877 * a task to be reinserted into the active array on timeslice
880 unsigned long ceiling
= INTERACTIVE_SLEEP(p
);
882 if (p
->mm
&& sleep_time
> ceiling
&& p
->sleep_avg
< ceiling
) {
884 * Prevents user tasks from achieving best priority
885 * with one single large enough sleep.
887 p
->sleep_avg
= ceiling
;
889 * Using INTERACTIVE_SLEEP() as a ceiling places a
890 * nice(0) task 1ms sleep away from promotion, and
891 * gives it 700ms to round-robin with no chance of
892 * being demoted. This is more than generous, so
893 * mark this sleep as non-interactive to prevent the
894 * on-runqueue bonus logic from intervening should
895 * this task not receive cpu immediately.
897 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
900 * Tasks waking from uninterruptible sleep are
901 * limited in their sleep_avg rise as they
902 * are likely to be waiting on I/O
904 if (p
->sleep_type
== SLEEP_NONINTERACTIVE
&& p
->mm
) {
905 if (p
->sleep_avg
>= ceiling
)
907 else if (p
->sleep_avg
+ sleep_time
>=
909 p
->sleep_avg
= ceiling
;
915 * This code gives a bonus to interactive tasks.
917 * The boost works by updating the 'average sleep time'
918 * value here, based on ->timestamp. The more time a
919 * task spends sleeping, the higher the average gets -
920 * and the higher the priority boost gets as well.
922 p
->sleep_avg
+= sleep_time
;
925 if (p
->sleep_avg
> NS_MAX_SLEEP_AVG
)
926 p
->sleep_avg
= NS_MAX_SLEEP_AVG
;
929 return effective_prio(p
);
933 * activate_task - move a task to the runqueue and do priority recalculation
935 * Update all the scheduling statistics stuff. (sleep average
936 * calculation, priority modifiers, etc.)
938 static void activate_task(struct task_struct
*p
, struct rq
*rq
, int local
)
940 unsigned long long now
;
945 /* Compensate for drifting sched_clock */
946 struct rq
*this_rq
= this_rq();
947 now
= (now
- this_rq
->timestamp_last_tick
)
948 + rq
->timestamp_last_tick
;
953 * Sleep time is in units of nanosecs, so shift by 20 to get a
954 * milliseconds-range estimation of the amount of time that the task
957 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
958 if (p
->state
== TASK_UNINTERRUPTIBLE
)
959 profile_hits(SLEEP_PROFILING
, (void *)get_wchan(p
),
960 (now
- p
->timestamp
) >> 20);
964 p
->prio
= recalc_task_prio(p
, now
);
967 * This checks to make sure it's not an uninterruptible task
968 * that is now waking up.
970 if (p
->sleep_type
== SLEEP_NORMAL
) {
972 * Tasks which were woken up by interrupts (ie. hw events)
973 * are most likely of interactive nature. So we give them
974 * the credit of extending their sleep time to the period
975 * of time they spend on the runqueue, waiting for execution
976 * on a CPU, first time around:
979 p
->sleep_type
= SLEEP_INTERRUPTED
;
982 * Normal first-time wakeups get a credit too for
983 * on-runqueue time, but it will be weighted down:
985 p
->sleep_type
= SLEEP_INTERACTIVE
;
990 __activate_task(p
, rq
);
994 * deactivate_task - remove a task from the runqueue.
996 static void deactivate_task(struct task_struct
*p
, struct rq
*rq
)
998 dec_nr_running(p
, rq
);
999 dequeue_task(p
, p
->array
);
1004 * resched_task - mark a task 'to be rescheduled now'.
1006 * On UP this means the setting of the need_resched flag, on SMP it
1007 * might also involve a cross-CPU call to trigger the scheduler on
1012 #ifndef tsk_is_polling
1013 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1016 static void resched_task(struct task_struct
*p
)
1020 assert_spin_locked(&task_rq(p
)->lock
);
1022 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
1025 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
1028 if (cpu
== smp_processor_id())
1031 /* NEED_RESCHED must be visible before we test polling */
1033 if (!tsk_is_polling(p
))
1034 smp_send_reschedule(cpu
);
1037 static inline void resched_task(struct task_struct
*p
)
1039 assert_spin_locked(&task_rq(p
)->lock
);
1040 set_tsk_need_resched(p
);
1045 * task_curr - is this task currently executing on a CPU?
1046 * @p: the task in question.
1048 inline int task_curr(const struct task_struct
*p
)
1050 return cpu_curr(task_cpu(p
)) == p
;
1053 /* Used instead of source_load when we know the type == 0 */
1054 unsigned long weighted_cpuload(const int cpu
)
1056 return cpu_rq(cpu
)->raw_weighted_load
;
1060 struct migration_req
{
1061 struct list_head list
;
1063 struct task_struct
*task
;
1066 struct completion done
;
1070 * The task's runqueue lock must be held.
1071 * Returns true if you have to wait for migration thread.
1074 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1076 struct rq
*rq
= task_rq(p
);
1079 * If the task is not on a runqueue (and not running), then
1080 * it is sufficient to simply update the task's cpu field.
1082 if (!p
->array
&& !task_running(rq
, p
)) {
1083 set_task_cpu(p
, dest_cpu
);
1087 init_completion(&req
->done
);
1089 req
->dest_cpu
= dest_cpu
;
1090 list_add(&req
->list
, &rq
->migration_queue
);
1096 * wait_task_inactive - wait for a thread to unschedule.
1098 * The caller must ensure that the task *will* unschedule sometime soon,
1099 * else this function might spin for a *long* time. This function can't
1100 * be called with interrupts off, or it may introduce deadlock with
1101 * smp_call_function() if an IPI is sent by the same process we are
1102 * waiting to become inactive.
1104 void wait_task_inactive(struct task_struct
*p
)
1106 unsigned long flags
;
1111 rq
= task_rq_lock(p
, &flags
);
1112 /* Must be off runqueue entirely, not preempted. */
1113 if (unlikely(p
->array
|| task_running(rq
, p
))) {
1114 /* If it's preempted, we yield. It could be a while. */
1115 preempted
= !task_running(rq
, p
);
1116 task_rq_unlock(rq
, &flags
);
1122 task_rq_unlock(rq
, &flags
);
1126 * kick_process - kick a running thread to enter/exit the kernel
1127 * @p: the to-be-kicked thread
1129 * Cause a process which is running on another CPU to enter
1130 * kernel-mode, without any delay. (to get signals handled.)
1132 * NOTE: this function doesnt have to take the runqueue lock,
1133 * because all it wants to ensure is that the remote task enters
1134 * the kernel. If the IPI races and the task has been migrated
1135 * to another CPU then no harm is done and the purpose has been
1138 void kick_process(struct task_struct
*p
)
1144 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1145 smp_send_reschedule(cpu
);
1150 * Return a low guess at the load of a migration-source cpu weighted
1151 * according to the scheduling class and "nice" value.
1153 * We want to under-estimate the load of migration sources, to
1154 * balance conservatively.
1156 static inline unsigned long source_load(int cpu
, int type
)
1158 struct rq
*rq
= cpu_rq(cpu
);
1161 return rq
->raw_weighted_load
;
1163 return min(rq
->cpu_load
[type
-1], rq
->raw_weighted_load
);
1167 * Return a high guess at the load of a migration-target cpu weighted
1168 * according to the scheduling class and "nice" value.
1170 static inline unsigned long target_load(int cpu
, int type
)
1172 struct rq
*rq
= cpu_rq(cpu
);
1175 return rq
->raw_weighted_load
;
1177 return max(rq
->cpu_load
[type
-1], rq
->raw_weighted_load
);
1181 * Return the average load per task on the cpu's run queue
1183 static inline unsigned long cpu_avg_load_per_task(int cpu
)
1185 struct rq
*rq
= cpu_rq(cpu
);
1186 unsigned long n
= rq
->nr_running
;
1188 return n
? rq
->raw_weighted_load
/ n
: SCHED_LOAD_SCALE
;
1192 * find_idlest_group finds and returns the least busy CPU group within the
1195 static struct sched_group
*
1196 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1198 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1199 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1200 int load_idx
= sd
->forkexec_idx
;
1201 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1204 unsigned long load
, avg_load
;
1208 /* Skip over this group if it has no CPUs allowed */
1209 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1212 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1214 /* Tally up the load of all CPUs in the group */
1217 for_each_cpu_mask(i
, group
->cpumask
) {
1218 /* Bias balancing toward cpus of our domain */
1220 load
= source_load(i
, load_idx
);
1222 load
= target_load(i
, load_idx
);
1227 /* Adjust by relative CPU power of the group */
1228 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
1231 this_load
= avg_load
;
1233 } else if (avg_load
< min_load
) {
1234 min_load
= avg_load
;
1238 group
= group
->next
;
1239 } while (group
!= sd
->groups
);
1241 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1247 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1250 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1253 unsigned long load
, min_load
= ULONG_MAX
;
1257 /* Traverse only the allowed CPUs */
1258 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1260 for_each_cpu_mask(i
, tmp
) {
1261 load
= weighted_cpuload(i
);
1263 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1273 * sched_balance_self: balance the current task (running on cpu) in domains
1274 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1277 * Balance, ie. select the least loaded group.
1279 * Returns the target CPU number, or the same CPU if no balancing is needed.
1281 * preempt must be disabled.
1283 static int sched_balance_self(int cpu
, int flag
)
1285 struct task_struct
*t
= current
;
1286 struct sched_domain
*tmp
, *sd
= NULL
;
1288 for_each_domain(cpu
, tmp
) {
1290 * If power savings logic is enabled for a domain, stop there.
1292 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1294 if (tmp
->flags
& flag
)
1300 struct sched_group
*group
;
1301 int new_cpu
, weight
;
1303 if (!(sd
->flags
& flag
)) {
1309 group
= find_idlest_group(sd
, t
, cpu
);
1315 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1316 if (new_cpu
== -1 || new_cpu
== cpu
) {
1317 /* Now try balancing at a lower domain level of cpu */
1322 /* Now try balancing at a lower domain level of new_cpu */
1325 weight
= cpus_weight(span
);
1326 for_each_domain(cpu
, tmp
) {
1327 if (weight
<= cpus_weight(tmp
->span
))
1329 if (tmp
->flags
& flag
)
1332 /* while loop will break here if sd == NULL */
1338 #endif /* CONFIG_SMP */
1341 * wake_idle() will wake a task on an idle cpu if task->cpu is
1342 * not idle and an idle cpu is available. The span of cpus to
1343 * search starts with cpus closest then further out as needed,
1344 * so we always favor a closer, idle cpu.
1346 * Returns the CPU we should wake onto.
1348 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1349 static int wake_idle(int cpu
, struct task_struct
*p
)
1352 struct sched_domain
*sd
;
1358 for_each_domain(cpu
, sd
) {
1359 if (sd
->flags
& SD_WAKE_IDLE
) {
1360 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1361 for_each_cpu_mask(i
, tmp
) {
1372 static inline int wake_idle(int cpu
, struct task_struct
*p
)
1379 * try_to_wake_up - wake up a thread
1380 * @p: the to-be-woken-up thread
1381 * @state: the mask of task states that can be woken
1382 * @sync: do a synchronous wakeup?
1384 * Put it on the run-queue if it's not already there. The "current"
1385 * thread is always on the run-queue (except when the actual
1386 * re-schedule is in progress), and as such you're allowed to do
1387 * the simpler "current->state = TASK_RUNNING" to mark yourself
1388 * runnable without the overhead of this.
1390 * returns failure only if the task is already active.
1392 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1394 int cpu
, this_cpu
, success
= 0;
1395 unsigned long flags
;
1399 struct sched_domain
*sd
, *this_sd
= NULL
;
1400 unsigned long load
, this_load
;
1404 rq
= task_rq_lock(p
, &flags
);
1405 old_state
= p
->state
;
1406 if (!(old_state
& state
))
1413 this_cpu
= smp_processor_id();
1416 if (unlikely(task_running(rq
, p
)))
1421 schedstat_inc(rq
, ttwu_cnt
);
1422 if (cpu
== this_cpu
) {
1423 schedstat_inc(rq
, ttwu_local
);
1427 for_each_domain(this_cpu
, sd
) {
1428 if (cpu_isset(cpu
, sd
->span
)) {
1429 schedstat_inc(sd
, ttwu_wake_remote
);
1435 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1439 * Check for affine wakeup and passive balancing possibilities.
1442 int idx
= this_sd
->wake_idx
;
1443 unsigned int imbalance
;
1445 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1447 load
= source_load(cpu
, idx
);
1448 this_load
= target_load(this_cpu
, idx
);
1450 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1452 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1453 unsigned long tl
= this_load
;
1454 unsigned long tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1457 * If sync wakeup then subtract the (maximum possible)
1458 * effect of the currently running task from the load
1459 * of the current CPU:
1462 tl
-= current
->load_weight
;
1465 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1466 100*(tl
+ p
->load_weight
) <= imbalance
*load
) {
1468 * This domain has SD_WAKE_AFFINE and
1469 * p is cache cold in this domain, and
1470 * there is no bad imbalance.
1472 schedstat_inc(this_sd
, ttwu_move_affine
);
1478 * Start passive balancing when half the imbalance_pct
1481 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1482 if (imbalance
*this_load
<= 100*load
) {
1483 schedstat_inc(this_sd
, ttwu_move_balance
);
1489 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1491 new_cpu
= wake_idle(new_cpu
, p
);
1492 if (new_cpu
!= cpu
) {
1493 set_task_cpu(p
, new_cpu
);
1494 task_rq_unlock(rq
, &flags
);
1495 /* might preempt at this point */
1496 rq
= task_rq_lock(p
, &flags
);
1497 old_state
= p
->state
;
1498 if (!(old_state
& state
))
1503 this_cpu
= smp_processor_id();
1508 #endif /* CONFIG_SMP */
1509 if (old_state
== TASK_UNINTERRUPTIBLE
) {
1510 rq
->nr_uninterruptible
--;
1512 * Tasks on involuntary sleep don't earn
1513 * sleep_avg beyond just interactive state.
1515 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
1519 * Tasks that have marked their sleep as noninteractive get
1520 * woken up with their sleep average not weighted in an
1523 if (old_state
& TASK_NONINTERACTIVE
)
1524 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
1527 activate_task(p
, rq
, cpu
== this_cpu
);
1529 * Sync wakeups (i.e. those types of wakeups where the waker
1530 * has indicated that it will leave the CPU in short order)
1531 * don't trigger a preemption, if the woken up task will run on
1532 * this cpu. (in this case the 'I will reschedule' promise of
1533 * the waker guarantees that the freshly woken up task is going
1534 * to be considered on this CPU.)
1536 if (!sync
|| cpu
!= this_cpu
) {
1537 if (TASK_PREEMPTS_CURR(p
, rq
))
1538 resched_task(rq
->curr
);
1543 p
->state
= TASK_RUNNING
;
1545 task_rq_unlock(rq
, &flags
);
1550 int fastcall
wake_up_process(struct task_struct
*p
)
1552 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1553 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1555 EXPORT_SYMBOL(wake_up_process
);
1557 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1559 return try_to_wake_up(p
, state
, 0);
1563 * Perform scheduler related setup for a newly forked process p.
1564 * p is forked by current.
1566 void fastcall
sched_fork(struct task_struct
*p
, int clone_flags
)
1568 int cpu
= get_cpu();
1571 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1573 set_task_cpu(p
, cpu
);
1576 * We mark the process as running here, but have not actually
1577 * inserted it onto the runqueue yet. This guarantees that
1578 * nobody will actually run it, and a signal or other external
1579 * event cannot wake it up and insert it on the runqueue either.
1581 p
->state
= TASK_RUNNING
;
1584 * Make sure we do not leak PI boosting priority to the child:
1586 p
->prio
= current
->normal_prio
;
1588 INIT_LIST_HEAD(&p
->run_list
);
1590 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1591 if (unlikely(sched_info_on()))
1592 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1594 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1597 #ifdef CONFIG_PREEMPT
1598 /* Want to start with kernel preemption disabled. */
1599 task_thread_info(p
)->preempt_count
= 1;
1602 * Share the timeslice between parent and child, thus the
1603 * total amount of pending timeslices in the system doesn't change,
1604 * resulting in more scheduling fairness.
1606 local_irq_disable();
1607 p
->time_slice
= (current
->time_slice
+ 1) >> 1;
1609 * The remainder of the first timeslice might be recovered by
1610 * the parent if the child exits early enough.
1612 p
->first_time_slice
= 1;
1613 current
->time_slice
>>= 1;
1614 p
->timestamp
= sched_clock();
1615 if (unlikely(!current
->time_slice
)) {
1617 * This case is rare, it happens when the parent has only
1618 * a single jiffy left from its timeslice. Taking the
1619 * runqueue lock is not a problem.
1621 current
->time_slice
= 1;
1629 * wake_up_new_task - wake up a newly created task for the first time.
1631 * This function will do some initial scheduler statistics housekeeping
1632 * that must be done for every newly created context, then puts the task
1633 * on the runqueue and wakes it.
1635 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1637 struct rq
*rq
, *this_rq
;
1638 unsigned long flags
;
1641 rq
= task_rq_lock(p
, &flags
);
1642 BUG_ON(p
->state
!= TASK_RUNNING
);
1643 this_cpu
= smp_processor_id();
1647 * We decrease the sleep average of forking parents
1648 * and children as well, to keep max-interactive tasks
1649 * from forking tasks that are max-interactive. The parent
1650 * (current) is done further down, under its lock.
1652 p
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(p
) *
1653 CHILD_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1655 p
->prio
= effective_prio(p
);
1657 if (likely(cpu
== this_cpu
)) {
1658 if (!(clone_flags
& CLONE_VM
)) {
1660 * The VM isn't cloned, so we're in a good position to
1661 * do child-runs-first in anticipation of an exec. This
1662 * usually avoids a lot of COW overhead.
1664 if (unlikely(!current
->array
))
1665 __activate_task(p
, rq
);
1667 p
->prio
= current
->prio
;
1668 p
->normal_prio
= current
->normal_prio
;
1669 list_add_tail(&p
->run_list
, ¤t
->run_list
);
1670 p
->array
= current
->array
;
1671 p
->array
->nr_active
++;
1672 inc_nr_running(p
, rq
);
1676 /* Run child last */
1677 __activate_task(p
, rq
);
1679 * We skip the following code due to cpu == this_cpu
1681 * task_rq_unlock(rq, &flags);
1682 * this_rq = task_rq_lock(current, &flags);
1686 this_rq
= cpu_rq(this_cpu
);
1689 * Not the local CPU - must adjust timestamp. This should
1690 * get optimised away in the !CONFIG_SMP case.
1692 p
->timestamp
= (p
->timestamp
- this_rq
->timestamp_last_tick
)
1693 + rq
->timestamp_last_tick
;
1694 __activate_task(p
, rq
);
1695 if (TASK_PREEMPTS_CURR(p
, rq
))
1696 resched_task(rq
->curr
);
1699 * Parent and child are on different CPUs, now get the
1700 * parent runqueue to update the parent's ->sleep_avg:
1702 task_rq_unlock(rq
, &flags
);
1703 this_rq
= task_rq_lock(current
, &flags
);
1705 current
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(current
) *
1706 PARENT_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1707 task_rq_unlock(this_rq
, &flags
);
1711 * Potentially available exiting-child timeslices are
1712 * retrieved here - this way the parent does not get
1713 * penalized for creating too many threads.
1715 * (this cannot be used to 'generate' timeslices
1716 * artificially, because any timeslice recovered here
1717 * was given away by the parent in the first place.)
1719 void fastcall
sched_exit(struct task_struct
*p
)
1721 unsigned long flags
;
1725 * If the child was a (relative-) CPU hog then decrease
1726 * the sleep_avg of the parent as well.
1728 rq
= task_rq_lock(p
->parent
, &flags
);
1729 if (p
->first_time_slice
&& task_cpu(p
) == task_cpu(p
->parent
)) {
1730 p
->parent
->time_slice
+= p
->time_slice
;
1731 if (unlikely(p
->parent
->time_slice
> task_timeslice(p
)))
1732 p
->parent
->time_slice
= task_timeslice(p
);
1734 if (p
->sleep_avg
< p
->parent
->sleep_avg
)
1735 p
->parent
->sleep_avg
= p
->parent
->sleep_avg
/
1736 (EXIT_WEIGHT
+ 1) * EXIT_WEIGHT
+ p
->sleep_avg
/
1738 task_rq_unlock(rq
, &flags
);
1742 * prepare_task_switch - prepare to switch tasks
1743 * @rq: the runqueue preparing to switch
1744 * @next: the task we are going to switch to.
1746 * This is called with the rq lock held and interrupts off. It must
1747 * be paired with a subsequent finish_task_switch after the context
1750 * prepare_task_switch sets up locking and calls architecture specific
1753 static inline void prepare_task_switch(struct rq
*rq
, struct task_struct
*next
)
1755 prepare_lock_switch(rq
, next
);
1756 prepare_arch_switch(next
);
1760 * finish_task_switch - clean up after a task-switch
1761 * @rq: runqueue associated with task-switch
1762 * @prev: the thread we just switched away from.
1764 * finish_task_switch must be called after the context switch, paired
1765 * with a prepare_task_switch call before the context switch.
1766 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1767 * and do any other architecture-specific cleanup actions.
1769 * Note that we may have delayed dropping an mm in context_switch(). If
1770 * so, we finish that here outside of the runqueue lock. (Doing it
1771 * with the lock held can cause deadlocks; see schedule() for
1774 static inline void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1775 __releases(rq
->lock
)
1777 struct mm_struct
*mm
= rq
->prev_mm
;
1783 * A task struct has one reference for the use as "current".
1784 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1785 * schedule one last time. The schedule call will never return, and
1786 * the scheduled task must drop that reference.
1787 * The test for TASK_DEAD must occur while the runqueue locks are
1788 * still held, otherwise prev could be scheduled on another cpu, die
1789 * there before we look at prev->state, and then the reference would
1791 * Manfred Spraul <manfred@colorfullife.com>
1793 prev_state
= prev
->state
;
1794 finish_arch_switch(prev
);
1795 finish_lock_switch(rq
, prev
);
1798 if (unlikely(prev_state
== TASK_DEAD
)) {
1800 * Remove function-return probe instances associated with this
1801 * task and put them back on the free list.
1803 kprobe_flush_task(prev
);
1804 put_task_struct(prev
);
1809 * schedule_tail - first thing a freshly forked thread must call.
1810 * @prev: the thread we just switched away from.
1812 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1813 __releases(rq
->lock
)
1815 struct rq
*rq
= this_rq();
1817 finish_task_switch(rq
, prev
);
1818 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1819 /* In this case, finish_task_switch does not reenable preemption */
1822 if (current
->set_child_tid
)
1823 put_user(current
->pid
, current
->set_child_tid
);
1827 * context_switch - switch to the new MM and the new
1828 * thread's register state.
1830 static inline struct task_struct
*
1831 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1832 struct task_struct
*next
)
1834 struct mm_struct
*mm
= next
->mm
;
1835 struct mm_struct
*oldmm
= prev
->active_mm
;
1838 next
->active_mm
= oldmm
;
1839 atomic_inc(&oldmm
->mm_count
);
1840 enter_lazy_tlb(oldmm
, next
);
1842 switch_mm(oldmm
, mm
, next
);
1845 prev
->active_mm
= NULL
;
1846 WARN_ON(rq
->prev_mm
);
1847 rq
->prev_mm
= oldmm
;
1850 * Since the runqueue lock will be released by the next
1851 * task (which is an invalid locking op but in the case
1852 * of the scheduler it's an obvious special-case), so we
1853 * do an early lockdep release here:
1855 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1856 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1859 /* Here we just switch the register state and the stack. */
1860 switch_to(prev
, next
, prev
);
1866 * nr_running, nr_uninterruptible and nr_context_switches:
1868 * externally visible scheduler statistics: current number of runnable
1869 * threads, current number of uninterruptible-sleeping threads, total
1870 * number of context switches performed since bootup.
1872 unsigned long nr_running(void)
1874 unsigned long i
, sum
= 0;
1876 for_each_online_cpu(i
)
1877 sum
+= cpu_rq(i
)->nr_running
;
1882 unsigned long nr_uninterruptible(void)
1884 unsigned long i
, sum
= 0;
1886 for_each_possible_cpu(i
)
1887 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1890 * Since we read the counters lockless, it might be slightly
1891 * inaccurate. Do not allow it to go below zero though:
1893 if (unlikely((long)sum
< 0))
1899 unsigned long long nr_context_switches(void)
1902 unsigned long long sum
= 0;
1904 for_each_possible_cpu(i
)
1905 sum
+= cpu_rq(i
)->nr_switches
;
1910 unsigned long nr_iowait(void)
1912 unsigned long i
, sum
= 0;
1914 for_each_possible_cpu(i
)
1915 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1920 unsigned long nr_active(void)
1922 unsigned long i
, running
= 0, uninterruptible
= 0;
1924 for_each_online_cpu(i
) {
1925 running
+= cpu_rq(i
)->nr_running
;
1926 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
1929 if (unlikely((long)uninterruptible
< 0))
1930 uninterruptible
= 0;
1932 return running
+ uninterruptible
;
1938 * Is this task likely cache-hot:
1941 task_hot(struct task_struct
*p
, unsigned long long now
, struct sched_domain
*sd
)
1943 return (long long)(now
- p
->last_ran
) < (long long)sd
->cache_hot_time
;
1947 * double_rq_lock - safely lock two runqueues
1949 * Note this does not disable interrupts like task_rq_lock,
1950 * you need to do so manually before calling.
1952 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1953 __acquires(rq1
->lock
)
1954 __acquires(rq2
->lock
)
1956 BUG_ON(!irqs_disabled());
1958 spin_lock(&rq1
->lock
);
1959 __acquire(rq2
->lock
); /* Fake it out ;) */
1962 spin_lock(&rq1
->lock
);
1963 spin_lock(&rq2
->lock
);
1965 spin_lock(&rq2
->lock
);
1966 spin_lock(&rq1
->lock
);
1972 * double_rq_unlock - safely unlock two runqueues
1974 * Note this does not restore interrupts like task_rq_unlock,
1975 * you need to do so manually after calling.
1977 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
1978 __releases(rq1
->lock
)
1979 __releases(rq2
->lock
)
1981 spin_unlock(&rq1
->lock
);
1983 spin_unlock(&rq2
->lock
);
1985 __release(rq2
->lock
);
1989 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1991 static void double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1992 __releases(this_rq
->lock
)
1993 __acquires(busiest
->lock
)
1994 __acquires(this_rq
->lock
)
1996 if (unlikely(!irqs_disabled())) {
1997 /* printk() doesn't work good under rq->lock */
1998 spin_unlock(&this_rq
->lock
);
2001 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2002 if (busiest
< this_rq
) {
2003 spin_unlock(&this_rq
->lock
);
2004 spin_lock(&busiest
->lock
);
2005 spin_lock(&this_rq
->lock
);
2007 spin_lock(&busiest
->lock
);
2012 * If dest_cpu is allowed for this process, migrate the task to it.
2013 * This is accomplished by forcing the cpu_allowed mask to only
2014 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2015 * the cpu_allowed mask is restored.
2017 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2019 struct migration_req req
;
2020 unsigned long flags
;
2023 rq
= task_rq_lock(p
, &flags
);
2024 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2025 || unlikely(cpu_is_offline(dest_cpu
)))
2028 /* force the process onto the specified CPU */
2029 if (migrate_task(p
, dest_cpu
, &req
)) {
2030 /* Need to wait for migration thread (might exit: take ref). */
2031 struct task_struct
*mt
= rq
->migration_thread
;
2033 get_task_struct(mt
);
2034 task_rq_unlock(rq
, &flags
);
2035 wake_up_process(mt
);
2036 put_task_struct(mt
);
2037 wait_for_completion(&req
.done
);
2042 task_rq_unlock(rq
, &flags
);
2046 * sched_exec - execve() is a valuable balancing opportunity, because at
2047 * this point the task has the smallest effective memory and cache footprint.
2049 void sched_exec(void)
2051 int new_cpu
, this_cpu
= get_cpu();
2052 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2054 if (new_cpu
!= this_cpu
)
2055 sched_migrate_task(current
, new_cpu
);
2059 * pull_task - move a task from a remote runqueue to the local runqueue.
2060 * Both runqueues must be locked.
2062 static void pull_task(struct rq
*src_rq
, struct prio_array
*src_array
,
2063 struct task_struct
*p
, struct rq
*this_rq
,
2064 struct prio_array
*this_array
, int this_cpu
)
2066 dequeue_task(p
, src_array
);
2067 dec_nr_running(p
, src_rq
);
2068 set_task_cpu(p
, this_cpu
);
2069 inc_nr_running(p
, this_rq
);
2070 enqueue_task(p
, this_array
);
2071 p
->timestamp
= (p
->timestamp
- src_rq
->timestamp_last_tick
)
2072 + this_rq
->timestamp_last_tick
;
2074 * Note that idle threads have a prio of MAX_PRIO, for this test
2075 * to be always true for them.
2077 if (TASK_PREEMPTS_CURR(p
, this_rq
))
2078 resched_task(this_rq
->curr
);
2082 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2085 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2086 struct sched_domain
*sd
, enum idle_type idle
,
2090 * We do not migrate tasks that are:
2091 * 1) running (obviously), or
2092 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2093 * 3) are cache-hot on their current CPU.
2095 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
2099 if (task_running(rq
, p
))
2103 * Aggressive migration if:
2104 * 1) task is cache cold, or
2105 * 2) too many balance attempts have failed.
2108 if (sd
->nr_balance_failed
> sd
->cache_nice_tries
)
2111 if (task_hot(p
, rq
->timestamp_last_tick
, sd
))
2116 #define rq_best_prio(rq) min((rq)->curr->prio, (rq)->best_expired_prio)
2119 * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted
2120 * load from busiest to this_rq, as part of a balancing operation within
2121 * "domain". Returns the number of tasks moved.
2123 * Called with both runqueues locked.
2125 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2126 unsigned long max_nr_move
, unsigned long max_load_move
,
2127 struct sched_domain
*sd
, enum idle_type idle
,
2130 int idx
, pulled
= 0, pinned
= 0, this_best_prio
, best_prio
,
2131 best_prio_seen
, skip_for_load
;
2132 struct prio_array
*array
, *dst_array
;
2133 struct list_head
*head
, *curr
;
2134 struct task_struct
*tmp
;
2137 if (max_nr_move
== 0 || max_load_move
== 0)
2140 rem_load_move
= max_load_move
;
2142 this_best_prio
= rq_best_prio(this_rq
);
2143 best_prio
= rq_best_prio(busiest
);
2145 * Enable handling of the case where there is more than one task
2146 * with the best priority. If the current running task is one
2147 * of those with prio==best_prio we know it won't be moved
2148 * and therefore it's safe to override the skip (based on load) of
2149 * any task we find with that prio.
2151 best_prio_seen
= best_prio
== busiest
->curr
->prio
;
2154 * We first consider expired tasks. Those will likely not be
2155 * executed in the near future, and they are most likely to
2156 * be cache-cold, thus switching CPUs has the least effect
2159 if (busiest
->expired
->nr_active
) {
2160 array
= busiest
->expired
;
2161 dst_array
= this_rq
->expired
;
2163 array
= busiest
->active
;
2164 dst_array
= this_rq
->active
;
2168 /* Start searching at priority 0: */
2172 idx
= sched_find_first_bit(array
->bitmap
);
2174 idx
= find_next_bit(array
->bitmap
, MAX_PRIO
, idx
);
2175 if (idx
>= MAX_PRIO
) {
2176 if (array
== busiest
->expired
&& busiest
->active
->nr_active
) {
2177 array
= busiest
->active
;
2178 dst_array
= this_rq
->active
;
2184 head
= array
->queue
+ idx
;
2187 tmp
= list_entry(curr
, struct task_struct
, run_list
);
2192 * To help distribute high priority tasks accross CPUs we don't
2193 * skip a task if it will be the highest priority task (i.e. smallest
2194 * prio value) on its new queue regardless of its load weight
2196 skip_for_load
= tmp
->load_weight
> rem_load_move
;
2197 if (skip_for_load
&& idx
< this_best_prio
)
2198 skip_for_load
= !best_prio_seen
&& idx
== best_prio
;
2199 if (skip_for_load
||
2200 !can_migrate_task(tmp
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2202 best_prio_seen
|= idx
== best_prio
;
2209 #ifdef CONFIG_SCHEDSTATS
2210 if (task_hot(tmp
, busiest
->timestamp_last_tick
, sd
))
2211 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2214 pull_task(busiest
, array
, tmp
, this_rq
, dst_array
, this_cpu
);
2216 rem_load_move
-= tmp
->load_weight
;
2219 * We only want to steal up to the prescribed number of tasks
2220 * and the prescribed amount of weighted load.
2222 if (pulled
< max_nr_move
&& rem_load_move
> 0) {
2223 if (idx
< this_best_prio
)
2224 this_best_prio
= idx
;
2232 * Right now, this is the only place pull_task() is called,
2233 * so we can safely collect pull_task() stats here rather than
2234 * inside pull_task().
2236 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2239 *all_pinned
= pinned
;
2244 * find_busiest_group finds and returns the busiest CPU group within the
2245 * domain. It calculates and returns the amount of weighted load which
2246 * should be moved to restore balance via the imbalance parameter.
2248 static struct sched_group
*
2249 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2250 unsigned long *imbalance
, enum idle_type idle
, int *sd_idle
,
2253 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2254 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2255 unsigned long max_pull
;
2256 unsigned long busiest_load_per_task
, busiest_nr_running
;
2257 unsigned long this_load_per_task
, this_nr_running
;
2259 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2260 int power_savings_balance
= 1;
2261 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2262 unsigned long min_nr_running
= ULONG_MAX
;
2263 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2266 max_load
= this_load
= total_load
= total_pwr
= 0;
2267 busiest_load_per_task
= busiest_nr_running
= 0;
2268 this_load_per_task
= this_nr_running
= 0;
2269 if (idle
== NOT_IDLE
)
2270 load_idx
= sd
->busy_idx
;
2271 else if (idle
== NEWLY_IDLE
)
2272 load_idx
= sd
->newidle_idx
;
2274 load_idx
= sd
->idle_idx
;
2277 unsigned long load
, group_capacity
;
2280 unsigned long sum_nr_running
, sum_weighted_load
;
2282 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2284 /* Tally up the load of all CPUs in the group */
2285 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2287 for_each_cpu_mask(i
, group
->cpumask
) {
2290 if (!cpu_isset(i
, *cpus
))
2295 if (*sd_idle
&& !idle_cpu(i
))
2298 /* Bias balancing toward cpus of our domain */
2300 load
= target_load(i
, load_idx
);
2302 load
= source_load(i
, load_idx
);
2305 sum_nr_running
+= rq
->nr_running
;
2306 sum_weighted_load
+= rq
->raw_weighted_load
;
2309 total_load
+= avg_load
;
2310 total_pwr
+= group
->cpu_power
;
2312 /* Adjust by relative CPU power of the group */
2313 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
2315 group_capacity
= group
->cpu_power
/ SCHED_LOAD_SCALE
;
2318 this_load
= avg_load
;
2320 this_nr_running
= sum_nr_running
;
2321 this_load_per_task
= sum_weighted_load
;
2322 } else if (avg_load
> max_load
&&
2323 sum_nr_running
> group_capacity
) {
2324 max_load
= avg_load
;
2326 busiest_nr_running
= sum_nr_running
;
2327 busiest_load_per_task
= sum_weighted_load
;
2330 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2332 * Busy processors will not participate in power savings
2335 if (idle
== NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2339 * If the local group is idle or completely loaded
2340 * no need to do power savings balance at this domain
2342 if (local_group
&& (this_nr_running
>= group_capacity
||
2344 power_savings_balance
= 0;
2347 * If a group is already running at full capacity or idle,
2348 * don't include that group in power savings calculations
2350 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2355 * Calculate the group which has the least non-idle load.
2356 * This is the group from where we need to pick up the load
2359 if ((sum_nr_running
< min_nr_running
) ||
2360 (sum_nr_running
== min_nr_running
&&
2361 first_cpu(group
->cpumask
) <
2362 first_cpu(group_min
->cpumask
))) {
2364 min_nr_running
= sum_nr_running
;
2365 min_load_per_task
= sum_weighted_load
/
2370 * Calculate the group which is almost near its
2371 * capacity but still has some space to pick up some load
2372 * from other group and save more power
2374 if (sum_nr_running
<= group_capacity
- 1) {
2375 if (sum_nr_running
> leader_nr_running
||
2376 (sum_nr_running
== leader_nr_running
&&
2377 first_cpu(group
->cpumask
) >
2378 first_cpu(group_leader
->cpumask
))) {
2379 group_leader
= group
;
2380 leader_nr_running
= sum_nr_running
;
2385 group
= group
->next
;
2386 } while (group
!= sd
->groups
);
2388 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2391 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2393 if (this_load
>= avg_load
||
2394 100*max_load
<= sd
->imbalance_pct
*this_load
)
2397 busiest_load_per_task
/= busiest_nr_running
;
2399 * We're trying to get all the cpus to the average_load, so we don't
2400 * want to push ourselves above the average load, nor do we wish to
2401 * reduce the max loaded cpu below the average load, as either of these
2402 * actions would just result in more rebalancing later, and ping-pong
2403 * tasks around. Thus we look for the minimum possible imbalance.
2404 * Negative imbalances (*we* are more loaded than anyone else) will
2405 * be counted as no imbalance for these purposes -- we can't fix that
2406 * by pulling tasks to us. Be careful of negative numbers as they'll
2407 * appear as very large values with unsigned longs.
2409 if (max_load
<= busiest_load_per_task
)
2413 * In the presence of smp nice balancing, certain scenarios can have
2414 * max load less than avg load(as we skip the groups at or below
2415 * its cpu_power, while calculating max_load..)
2417 if (max_load
< avg_load
) {
2419 goto small_imbalance
;
2422 /* Don't want to pull so many tasks that a group would go idle */
2423 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2425 /* How much load to actually move to equalise the imbalance */
2426 *imbalance
= min(max_pull
* busiest
->cpu_power
,
2427 (avg_load
- this_load
) * this->cpu_power
)
2431 * if *imbalance is less than the average load per runnable task
2432 * there is no gaurantee that any tasks will be moved so we'll have
2433 * a think about bumping its value to force at least one task to be
2436 if (*imbalance
< busiest_load_per_task
) {
2437 unsigned long tmp
, pwr_now
, pwr_move
;
2441 pwr_move
= pwr_now
= 0;
2443 if (this_nr_running
) {
2444 this_load_per_task
/= this_nr_running
;
2445 if (busiest_load_per_task
> this_load_per_task
)
2448 this_load_per_task
= SCHED_LOAD_SCALE
;
2450 if (max_load
- this_load
>= busiest_load_per_task
* imbn
) {
2451 *imbalance
= busiest_load_per_task
;
2456 * OK, we don't have enough imbalance to justify moving tasks,
2457 * however we may be able to increase total CPU power used by
2461 pwr_now
+= busiest
->cpu_power
*
2462 min(busiest_load_per_task
, max_load
);
2463 pwr_now
+= this->cpu_power
*
2464 min(this_load_per_task
, this_load
);
2465 pwr_now
/= SCHED_LOAD_SCALE
;
2467 /* Amount of load we'd subtract */
2468 tmp
= busiest_load_per_task
*SCHED_LOAD_SCALE
/busiest
->cpu_power
;
2470 pwr_move
+= busiest
->cpu_power
*
2471 min(busiest_load_per_task
, max_load
- tmp
);
2473 /* Amount of load we'd add */
2474 if (max_load
*busiest
->cpu_power
<
2475 busiest_load_per_task
*SCHED_LOAD_SCALE
)
2476 tmp
= max_load
*busiest
->cpu_power
/this->cpu_power
;
2478 tmp
= busiest_load_per_task
*SCHED_LOAD_SCALE
/this->cpu_power
;
2479 pwr_move
+= this->cpu_power
*min(this_load_per_task
, this_load
+ tmp
);
2480 pwr_move
/= SCHED_LOAD_SCALE
;
2482 /* Move if we gain throughput */
2483 if (pwr_move
<= pwr_now
)
2486 *imbalance
= busiest_load_per_task
;
2492 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2493 if (idle
== NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2496 if (this == group_leader
&& group_leader
!= group_min
) {
2497 *imbalance
= min_load_per_task
;
2507 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2510 find_busiest_queue(struct sched_group
*group
, enum idle_type idle
,
2511 unsigned long imbalance
, cpumask_t
*cpus
)
2513 struct rq
*busiest
= NULL
, *rq
;
2514 unsigned long max_load
= 0;
2517 for_each_cpu_mask(i
, group
->cpumask
) {
2519 if (!cpu_isset(i
, *cpus
))
2524 if (rq
->nr_running
== 1 && rq
->raw_weighted_load
> imbalance
)
2527 if (rq
->raw_weighted_load
> max_load
) {
2528 max_load
= rq
->raw_weighted_load
;
2537 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2538 * so long as it is large enough.
2540 #define MAX_PINNED_INTERVAL 512
2542 static inline unsigned long minus_1_or_zero(unsigned long n
)
2544 return n
> 0 ? n
- 1 : 0;
2548 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2549 * tasks if there is an imbalance.
2551 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2552 struct sched_domain
*sd
, enum idle_type idle
)
2554 int nr_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2555 struct sched_group
*group
;
2556 unsigned long imbalance
;
2558 cpumask_t cpus
= CPU_MASK_ALL
;
2559 unsigned long flags
;
2562 * When power savings policy is enabled for the parent domain, idle
2563 * sibling can pick up load irrespective of busy siblings. In this case,
2564 * let the state of idle sibling percolate up as IDLE, instead of
2565 * portraying it as NOT_IDLE.
2567 if (idle
!= NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2568 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2571 schedstat_inc(sd
, lb_cnt
[idle
]);
2574 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2577 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2581 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2583 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2587 BUG_ON(busiest
== this_rq
);
2589 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2592 if (busiest
->nr_running
> 1) {
2594 * Attempt to move tasks. If find_busiest_group has found
2595 * an imbalance but busiest->nr_running <= 1, the group is
2596 * still unbalanced. nr_moved simply stays zero, so it is
2597 * correctly treated as an imbalance.
2599 local_irq_save(flags
);
2600 double_rq_lock(this_rq
, busiest
);
2601 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2602 minus_1_or_zero(busiest
->nr_running
),
2603 imbalance
, sd
, idle
, &all_pinned
);
2604 double_rq_unlock(this_rq
, busiest
);
2605 local_irq_restore(flags
);
2607 /* All tasks on this runqueue were pinned by CPU affinity */
2608 if (unlikely(all_pinned
)) {
2609 cpu_clear(cpu_of(busiest
), cpus
);
2610 if (!cpus_empty(cpus
))
2617 schedstat_inc(sd
, lb_failed
[idle
]);
2618 sd
->nr_balance_failed
++;
2620 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2622 spin_lock_irqsave(&busiest
->lock
, flags
);
2624 /* don't kick the migration_thread, if the curr
2625 * task on busiest cpu can't be moved to this_cpu
2627 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2628 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2630 goto out_one_pinned
;
2633 if (!busiest
->active_balance
) {
2634 busiest
->active_balance
= 1;
2635 busiest
->push_cpu
= this_cpu
;
2638 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2640 wake_up_process(busiest
->migration_thread
);
2643 * We've kicked active balancing, reset the failure
2646 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2649 sd
->nr_balance_failed
= 0;
2651 if (likely(!active_balance
)) {
2652 /* We were unbalanced, so reset the balancing interval */
2653 sd
->balance_interval
= sd
->min_interval
;
2656 * If we've begun active balancing, start to back off. This
2657 * case may not be covered by the all_pinned logic if there
2658 * is only 1 task on the busy runqueue (because we don't call
2661 if (sd
->balance_interval
< sd
->max_interval
)
2662 sd
->balance_interval
*= 2;
2665 if (!nr_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2666 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2671 schedstat_inc(sd
, lb_balanced
[idle
]);
2673 sd
->nr_balance_failed
= 0;
2676 /* tune up the balancing interval */
2677 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2678 (sd
->balance_interval
< sd
->max_interval
))
2679 sd
->balance_interval
*= 2;
2681 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2682 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2688 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2689 * tasks if there is an imbalance.
2691 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2692 * this_rq is locked.
2695 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
2697 struct sched_group
*group
;
2698 struct rq
*busiest
= NULL
;
2699 unsigned long imbalance
;
2702 cpumask_t cpus
= CPU_MASK_ALL
;
2705 * When power savings policy is enabled for the parent domain, idle
2706 * sibling can pick up load irrespective of busy siblings. In this case,
2707 * let the state of idle sibling percolate up as IDLE, instead of
2708 * portraying it as NOT_IDLE.
2710 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
2711 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2714 schedstat_inc(sd
, lb_cnt
[NEWLY_IDLE
]);
2716 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, NEWLY_IDLE
,
2719 schedstat_inc(sd
, lb_nobusyg
[NEWLY_IDLE
]);
2723 busiest
= find_busiest_queue(group
, NEWLY_IDLE
, imbalance
,
2726 schedstat_inc(sd
, lb_nobusyq
[NEWLY_IDLE
]);
2730 BUG_ON(busiest
== this_rq
);
2732 schedstat_add(sd
, lb_imbalance
[NEWLY_IDLE
], imbalance
);
2735 if (busiest
->nr_running
> 1) {
2736 /* Attempt to move tasks */
2737 double_lock_balance(this_rq
, busiest
);
2738 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2739 minus_1_or_zero(busiest
->nr_running
),
2740 imbalance
, sd
, NEWLY_IDLE
, NULL
);
2741 spin_unlock(&busiest
->lock
);
2744 cpu_clear(cpu_of(busiest
), cpus
);
2745 if (!cpus_empty(cpus
))
2751 schedstat_inc(sd
, lb_failed
[NEWLY_IDLE
]);
2752 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2753 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2756 sd
->nr_balance_failed
= 0;
2761 schedstat_inc(sd
, lb_balanced
[NEWLY_IDLE
]);
2762 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2763 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2765 sd
->nr_balance_failed
= 0;
2771 * idle_balance is called by schedule() if this_cpu is about to become
2772 * idle. Attempts to pull tasks from other CPUs.
2774 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
2776 struct sched_domain
*sd
;
2777 int pulled_task
= 0;
2778 unsigned long next_balance
= jiffies
+ 60 * HZ
;
2780 for_each_domain(this_cpu
, sd
) {
2781 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
2782 /* If we've pulled tasks over stop searching: */
2783 pulled_task
= load_balance_newidle(this_cpu
,
2785 if (time_after(next_balance
,
2786 sd
->last_balance
+ sd
->balance_interval
))
2787 next_balance
= sd
->last_balance
2788 + sd
->balance_interval
;
2795 * We are going idle. next_balance may be set based on
2796 * a busy processor. So reset next_balance.
2798 this_rq
->next_balance
= next_balance
;
2802 * active_load_balance is run by migration threads. It pushes running tasks
2803 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2804 * running on each physical CPU where possible, and avoids physical /
2805 * logical imbalances.
2807 * Called with busiest_rq locked.
2809 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
2811 int target_cpu
= busiest_rq
->push_cpu
;
2812 struct sched_domain
*sd
;
2813 struct rq
*target_rq
;
2815 /* Is there any task to move? */
2816 if (busiest_rq
->nr_running
<= 1)
2819 target_rq
= cpu_rq(target_cpu
);
2822 * This condition is "impossible", if it occurs
2823 * we need to fix it. Originally reported by
2824 * Bjorn Helgaas on a 128-cpu setup.
2826 BUG_ON(busiest_rq
== target_rq
);
2828 /* move a task from busiest_rq to target_rq */
2829 double_lock_balance(busiest_rq
, target_rq
);
2831 /* Search for an sd spanning us and the target CPU. */
2832 for_each_domain(target_cpu
, sd
) {
2833 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2834 cpu_isset(busiest_cpu
, sd
->span
))
2839 schedstat_inc(sd
, alb_cnt
);
2841 if (move_tasks(target_rq
, target_cpu
, busiest_rq
, 1,
2842 RTPRIO_TO_LOAD_WEIGHT(100), sd
, SCHED_IDLE
,
2844 schedstat_inc(sd
, alb_pushed
);
2846 schedstat_inc(sd
, alb_failed
);
2848 spin_unlock(&target_rq
->lock
);
2851 static void update_load(struct rq
*this_rq
)
2853 unsigned long this_load
;
2856 this_load
= this_rq
->raw_weighted_load
;
2858 /* Update our load: */
2859 for (i
= 0, scale
= 1; i
< 3; i
++, scale
<<= 1) {
2860 unsigned long old_load
, new_load
;
2862 old_load
= this_rq
->cpu_load
[i
];
2863 new_load
= this_load
;
2865 * Round up the averaging division if load is increasing. This
2866 * prevents us from getting stuck on 9 if the load is 10, for
2869 if (new_load
> old_load
)
2870 new_load
+= scale
-1;
2871 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) / scale
;
2876 * run_rebalance_domains is triggered when needed from the scheduler tick.
2878 * It checks each scheduling domain to see if it is due to be balanced,
2879 * and initiates a balancing operation if so.
2881 * Balancing parameters are set up in arch_init_sched_domains.
2883 static DEFINE_SPINLOCK(balancing
);
2885 static void run_rebalance_domains(struct softirq_action
*h
)
2887 int this_cpu
= smp_processor_id();
2888 struct rq
*this_rq
= cpu_rq(this_cpu
);
2889 unsigned long interval
;
2890 struct sched_domain
*sd
;
2892 * We are idle if there are no processes running. This
2893 * is valid even if we are the idle process (SMT).
2895 enum idle_type idle
= !this_rq
->nr_running
?
2896 SCHED_IDLE
: NOT_IDLE
;
2897 /* Earliest time when we have to call run_rebalance_domains again */
2898 unsigned long next_balance
= jiffies
+ 60*HZ
;
2900 for_each_domain(this_cpu
, sd
) {
2901 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2904 interval
= sd
->balance_interval
;
2905 if (idle
!= SCHED_IDLE
)
2906 interval
*= sd
->busy_factor
;
2908 /* scale ms to jiffies */
2909 interval
= msecs_to_jiffies(interval
);
2910 if (unlikely(!interval
))
2913 if (sd
->flags
& SD_SERIALIZE
) {
2914 if (!spin_trylock(&balancing
))
2918 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
2919 if (load_balance(this_cpu
, this_rq
, sd
, idle
)) {
2921 * We've pulled tasks over so either we're no
2922 * longer idle, or one of our SMT siblings is
2927 sd
->last_balance
= jiffies
;
2929 if (sd
->flags
& SD_SERIALIZE
)
2930 spin_unlock(&balancing
);
2932 if (time_after(next_balance
, sd
->last_balance
+ interval
))
2933 next_balance
= sd
->last_balance
+ interval
;
2935 this_rq
->next_balance
= next_balance
;
2939 * on UP we do not need to balance between CPUs:
2941 static inline void idle_balance(int cpu
, struct rq
*rq
)
2946 static inline void wake_priority_sleeper(struct rq
*rq
)
2948 #ifdef CONFIG_SCHED_SMT
2949 if (!rq
->nr_running
)
2952 spin_lock(&rq
->lock
);
2954 * If an SMT sibling task has been put to sleep for priority
2955 * reasons reschedule the idle task to see if it can now run.
2958 resched_task(rq
->idle
);
2959 spin_unlock(&rq
->lock
);
2963 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2965 EXPORT_PER_CPU_SYMBOL(kstat
);
2968 * This is called on clock ticks and on context switches.
2969 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2972 update_cpu_clock(struct task_struct
*p
, struct rq
*rq
, unsigned long long now
)
2974 p
->sched_time
+= now
- max(p
->timestamp
, rq
->timestamp_last_tick
);
2978 * Return current->sched_time plus any more ns on the sched_clock
2979 * that have not yet been banked.
2981 unsigned long long current_sched_time(const struct task_struct
*p
)
2983 unsigned long long ns
;
2984 unsigned long flags
;
2986 local_irq_save(flags
);
2987 ns
= max(p
->timestamp
, task_rq(p
)->timestamp_last_tick
);
2988 ns
= p
->sched_time
+ sched_clock() - ns
;
2989 local_irq_restore(flags
);
2995 * We place interactive tasks back into the active array, if possible.
2997 * To guarantee that this does not starve expired tasks we ignore the
2998 * interactivity of a task if the first expired task had to wait more
2999 * than a 'reasonable' amount of time. This deadline timeout is
3000 * load-dependent, as the frequency of array switched decreases with
3001 * increasing number of running tasks. We also ignore the interactivity
3002 * if a better static_prio task has expired:
3004 static inline int expired_starving(struct rq
*rq
)
3006 if (rq
->curr
->static_prio
> rq
->best_expired_prio
)
3008 if (!STARVATION_LIMIT
|| !rq
->expired_timestamp
)
3010 if (jiffies
- rq
->expired_timestamp
> STARVATION_LIMIT
* rq
->nr_running
)
3016 * Account user cpu time to a process.
3017 * @p: the process that the cpu time gets accounted to
3018 * @hardirq_offset: the offset to subtract from hardirq_count()
3019 * @cputime: the cpu time spent in user space since the last update
3021 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3023 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3026 p
->utime
= cputime_add(p
->utime
, cputime
);
3028 /* Add user time to cpustat. */
3029 tmp
= cputime_to_cputime64(cputime
);
3030 if (TASK_NICE(p
) > 0)
3031 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3033 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3037 * Account system cpu time to a process.
3038 * @p: the process that the cpu time gets accounted to
3039 * @hardirq_offset: the offset to subtract from hardirq_count()
3040 * @cputime: the cpu time spent in kernel space since the last update
3042 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3045 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3046 struct rq
*rq
= this_rq();
3049 p
->stime
= cputime_add(p
->stime
, cputime
);
3051 /* Add system time to cpustat. */
3052 tmp
= cputime_to_cputime64(cputime
);
3053 if (hardirq_count() - hardirq_offset
)
3054 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3055 else if (softirq_count())
3056 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3057 else if (p
!= rq
->idle
)
3058 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3059 else if (atomic_read(&rq
->nr_iowait
) > 0)
3060 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3062 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3063 /* Account for system time used */
3064 acct_update_integrals(p
);
3068 * Account for involuntary wait time.
3069 * @p: the process from which the cpu time has been stolen
3070 * @steal: the cpu time spent in involuntary wait
3072 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3074 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3075 cputime64_t tmp
= cputime_to_cputime64(steal
);
3076 struct rq
*rq
= this_rq();
3078 if (p
== rq
->idle
) {
3079 p
->stime
= cputime_add(p
->stime
, steal
);
3080 if (atomic_read(&rq
->nr_iowait
) > 0)
3081 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3083 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3085 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3088 static void task_running_tick(struct rq
*rq
, struct task_struct
*p
)
3090 if (p
->array
!= rq
->active
) {
3091 /* Task has expired but was not scheduled yet */
3092 set_tsk_need_resched(p
);
3095 spin_lock(&rq
->lock
);
3097 * The task was running during this tick - update the
3098 * time slice counter. Note: we do not update a thread's
3099 * priority until it either goes to sleep or uses up its
3100 * timeslice. This makes it possible for interactive tasks
3101 * to use up their timeslices at their highest priority levels.
3105 * RR tasks need a special form of timeslice management.
3106 * FIFO tasks have no timeslices.
3108 if ((p
->policy
== SCHED_RR
) && !--p
->time_slice
) {
3109 p
->time_slice
= task_timeslice(p
);
3110 p
->first_time_slice
= 0;
3111 set_tsk_need_resched(p
);
3113 /* put it at the end of the queue: */
3114 requeue_task(p
, rq
->active
);
3118 if (!--p
->time_slice
) {
3119 dequeue_task(p
, rq
->active
);
3120 set_tsk_need_resched(p
);
3121 p
->prio
= effective_prio(p
);
3122 p
->time_slice
= task_timeslice(p
);
3123 p
->first_time_slice
= 0;
3125 if (!rq
->expired_timestamp
)
3126 rq
->expired_timestamp
= jiffies
;
3127 if (!TASK_INTERACTIVE(p
) || expired_starving(rq
)) {
3128 enqueue_task(p
, rq
->expired
);
3129 if (p
->static_prio
< rq
->best_expired_prio
)
3130 rq
->best_expired_prio
= p
->static_prio
;
3132 enqueue_task(p
, rq
->active
);
3135 * Prevent a too long timeslice allowing a task to monopolize
3136 * the CPU. We do this by splitting up the timeslice into
3139 * Note: this does not mean the task's timeslices expire or
3140 * get lost in any way, they just might be preempted by
3141 * another task of equal priority. (one with higher
3142 * priority would have preempted this task already.) We
3143 * requeue this task to the end of the list on this priority
3144 * level, which is in essence a round-robin of tasks with
3147 * This only applies to tasks in the interactive
3148 * delta range with at least TIMESLICE_GRANULARITY to requeue.
3150 if (TASK_INTERACTIVE(p
) && !((task_timeslice(p
) -
3151 p
->time_slice
) % TIMESLICE_GRANULARITY(p
)) &&
3152 (p
->time_slice
>= TIMESLICE_GRANULARITY(p
)) &&
3153 (p
->array
== rq
->active
)) {
3155 requeue_task(p
, rq
->active
);
3156 set_tsk_need_resched(p
);
3160 spin_unlock(&rq
->lock
);
3164 * This function gets called by the timer code, with HZ frequency.
3165 * We call it with interrupts disabled.
3167 * It also gets called by the fork code, when changing the parent's
3170 void scheduler_tick(void)
3172 unsigned long long now
= sched_clock();
3173 struct task_struct
*p
= current
;
3174 int cpu
= smp_processor_id();
3175 struct rq
*rq
= cpu_rq(cpu
);
3177 update_cpu_clock(p
, rq
, now
);
3179 rq
->timestamp_last_tick
= now
;
3182 /* Task on the idle queue */
3183 wake_priority_sleeper(rq
);
3185 task_running_tick(rq
, p
);
3188 if (time_after_eq(jiffies
, rq
->next_balance
))
3189 raise_softirq(SCHED_SOFTIRQ
);
3193 #ifdef CONFIG_SCHED_SMT
3194 static inline void wakeup_busy_runqueue(struct rq
*rq
)
3196 /* If an SMT runqueue is sleeping due to priority reasons wake it up */
3197 if (rq
->curr
== rq
->idle
&& rq
->nr_running
)
3198 resched_task(rq
->idle
);
3202 * Called with interrupt disabled and this_rq's runqueue locked.
3204 static void wake_sleeping_dependent(int this_cpu
)
3206 struct sched_domain
*tmp
, *sd
= NULL
;
3209 for_each_domain(this_cpu
, tmp
) {
3210 if (tmp
->flags
& SD_SHARE_CPUPOWER
) {
3219 for_each_cpu_mask(i
, sd
->span
) {
3220 struct rq
*smt_rq
= cpu_rq(i
);
3224 if (unlikely(!spin_trylock(&smt_rq
->lock
)))
3227 wakeup_busy_runqueue(smt_rq
);
3228 spin_unlock(&smt_rq
->lock
);
3233 * number of 'lost' timeslices this task wont be able to fully
3234 * utilize, if another task runs on a sibling. This models the
3235 * slowdown effect of other tasks running on siblings:
3237 static inline unsigned long
3238 smt_slice(struct task_struct
*p
, struct sched_domain
*sd
)
3240 return p
->time_slice
* (100 - sd
->per_cpu_gain
) / 100;
3244 * To minimise lock contention and not have to drop this_rq's runlock we only
3245 * trylock the sibling runqueues and bypass those runqueues if we fail to
3246 * acquire their lock. As we only trylock the normal locking order does not
3247 * need to be obeyed.
3250 dependent_sleeper(int this_cpu
, struct rq
*this_rq
, struct task_struct
*p
)
3252 struct sched_domain
*tmp
, *sd
= NULL
;
3255 /* kernel/rt threads do not participate in dependent sleeping */
3256 if (!p
->mm
|| rt_task(p
))
3259 for_each_domain(this_cpu
, tmp
) {
3260 if (tmp
->flags
& SD_SHARE_CPUPOWER
) {
3269 for_each_cpu_mask(i
, sd
->span
) {
3270 struct task_struct
*smt_curr
;
3277 if (unlikely(!spin_trylock(&smt_rq
->lock
)))
3280 smt_curr
= smt_rq
->curr
;
3286 * If a user task with lower static priority than the
3287 * running task on the SMT sibling is trying to schedule,
3288 * delay it till there is proportionately less timeslice
3289 * left of the sibling task to prevent a lower priority
3290 * task from using an unfair proportion of the
3291 * physical cpu's resources. -ck
3293 if (rt_task(smt_curr
)) {
3295 * With real time tasks we run non-rt tasks only
3296 * per_cpu_gain% of the time.
3298 if ((jiffies
% DEF_TIMESLICE
) >
3299 (sd
->per_cpu_gain
* DEF_TIMESLICE
/ 100))
3302 if (smt_curr
->static_prio
< p
->static_prio
&&
3303 !TASK_PREEMPTS_CURR(p
, smt_rq
) &&
3304 smt_slice(smt_curr
, sd
) > task_timeslice(p
))
3308 spin_unlock(&smt_rq
->lock
);
3313 static inline void wake_sleeping_dependent(int this_cpu
)
3317 dependent_sleeper(int this_cpu
, struct rq
*this_rq
, struct task_struct
*p
)
3323 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3325 void fastcall
add_preempt_count(int val
)
3330 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3332 preempt_count() += val
;
3334 * Spinlock count overflowing soon?
3336 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >= PREEMPT_MASK
-10);
3338 EXPORT_SYMBOL(add_preempt_count
);
3340 void fastcall
sub_preempt_count(int val
)
3345 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3348 * Is the spinlock portion underflowing?
3350 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3351 !(preempt_count() & PREEMPT_MASK
)))
3354 preempt_count() -= val
;
3356 EXPORT_SYMBOL(sub_preempt_count
);
3360 static inline int interactive_sleep(enum sleep_type sleep_type
)
3362 return (sleep_type
== SLEEP_INTERACTIVE
||
3363 sleep_type
== SLEEP_INTERRUPTED
);
3367 * schedule() is the main scheduler function.
3369 asmlinkage
void __sched
schedule(void)
3371 struct task_struct
*prev
, *next
;
3372 struct prio_array
*array
;
3373 struct list_head
*queue
;
3374 unsigned long long now
;
3375 unsigned long run_time
;
3376 int cpu
, idx
, new_prio
;
3381 * Test if we are atomic. Since do_exit() needs to call into
3382 * schedule() atomically, we ignore that path for now.
3383 * Otherwise, whine if we are scheduling when we should not be.
3385 if (unlikely(in_atomic() && !current
->exit_state
)) {
3386 printk(KERN_ERR
"BUG: scheduling while atomic: "
3388 current
->comm
, preempt_count(), current
->pid
);
3389 debug_show_held_locks(current
);
3392 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3397 release_kernel_lock(prev
);
3398 need_resched_nonpreemptible
:
3402 * The idle thread is not allowed to schedule!
3403 * Remove this check after it has been exercised a bit.
3405 if (unlikely(prev
== rq
->idle
) && prev
->state
!= TASK_RUNNING
) {
3406 printk(KERN_ERR
"bad: scheduling from the idle thread!\n");
3410 schedstat_inc(rq
, sched_cnt
);
3411 now
= sched_clock();
3412 if (likely((long long)(now
- prev
->timestamp
) < NS_MAX_SLEEP_AVG
)) {
3413 run_time
= now
- prev
->timestamp
;
3414 if (unlikely((long long)(now
- prev
->timestamp
) < 0))
3417 run_time
= NS_MAX_SLEEP_AVG
;
3420 * Tasks charged proportionately less run_time at high sleep_avg to
3421 * delay them losing their interactive status
3423 run_time
/= (CURRENT_BONUS(prev
) ? : 1);
3425 spin_lock_irq(&rq
->lock
);
3427 switch_count
= &prev
->nivcsw
;
3428 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3429 switch_count
= &prev
->nvcsw
;
3430 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3431 unlikely(signal_pending(prev
))))
3432 prev
->state
= TASK_RUNNING
;
3434 if (prev
->state
== TASK_UNINTERRUPTIBLE
)
3435 rq
->nr_uninterruptible
++;
3436 deactivate_task(prev
, rq
);
3440 cpu
= smp_processor_id();
3441 if (unlikely(!rq
->nr_running
)) {
3442 idle_balance(cpu
, rq
);
3443 if (!rq
->nr_running
) {
3445 rq
->expired_timestamp
= 0;
3446 wake_sleeping_dependent(cpu
);
3452 if (unlikely(!array
->nr_active
)) {
3454 * Switch the active and expired arrays.
3456 schedstat_inc(rq
, sched_switch
);
3457 rq
->active
= rq
->expired
;
3458 rq
->expired
= array
;
3460 rq
->expired_timestamp
= 0;
3461 rq
->best_expired_prio
= MAX_PRIO
;
3464 idx
= sched_find_first_bit(array
->bitmap
);
3465 queue
= array
->queue
+ idx
;
3466 next
= list_entry(queue
->next
, struct task_struct
, run_list
);
3468 if (!rt_task(next
) && interactive_sleep(next
->sleep_type
)) {
3469 unsigned long long delta
= now
- next
->timestamp
;
3470 if (unlikely((long long)(now
- next
->timestamp
) < 0))
3473 if (next
->sleep_type
== SLEEP_INTERACTIVE
)
3474 delta
= delta
* (ON_RUNQUEUE_WEIGHT
* 128 / 100) / 128;
3476 array
= next
->array
;
3477 new_prio
= recalc_task_prio(next
, next
->timestamp
+ delta
);
3479 if (unlikely(next
->prio
!= new_prio
)) {
3480 dequeue_task(next
, array
);
3481 next
->prio
= new_prio
;
3482 enqueue_task(next
, array
);
3485 next
->sleep_type
= SLEEP_NORMAL
;
3486 if (dependent_sleeper(cpu
, rq
, next
))
3489 if (next
== rq
->idle
)
3490 schedstat_inc(rq
, sched_goidle
);
3492 prefetch_stack(next
);
3493 clear_tsk_need_resched(prev
);
3494 rcu_qsctr_inc(task_cpu(prev
));
3496 update_cpu_clock(prev
, rq
, now
);
3498 prev
->sleep_avg
-= run_time
;
3499 if ((long)prev
->sleep_avg
<= 0)
3500 prev
->sleep_avg
= 0;
3501 prev
->timestamp
= prev
->last_ran
= now
;
3503 sched_info_switch(prev
, next
);
3504 if (likely(prev
!= next
)) {
3505 next
->timestamp
= now
;
3510 prepare_task_switch(rq
, next
);
3511 prev
= context_switch(rq
, prev
, next
);
3514 * this_rq must be evaluated again because prev may have moved
3515 * CPUs since it called schedule(), thus the 'rq' on its stack
3516 * frame will be invalid.
3518 finish_task_switch(this_rq(), prev
);
3520 spin_unlock_irq(&rq
->lock
);
3523 if (unlikely(reacquire_kernel_lock(prev
) < 0))
3524 goto need_resched_nonpreemptible
;
3525 preempt_enable_no_resched();
3526 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3529 EXPORT_SYMBOL(schedule
);
3531 #ifdef CONFIG_PREEMPT
3533 * this is the entry point to schedule() from in-kernel preemption
3534 * off of preempt_enable. Kernel preemptions off return from interrupt
3535 * occur there and call schedule directly.
3537 asmlinkage
void __sched
preempt_schedule(void)
3539 struct thread_info
*ti
= current_thread_info();
3540 #ifdef CONFIG_PREEMPT_BKL
3541 struct task_struct
*task
= current
;
3542 int saved_lock_depth
;
3545 * If there is a non-zero preempt_count or interrupts are disabled,
3546 * we do not want to preempt the current task. Just return..
3548 if (likely(ti
->preempt_count
|| irqs_disabled()))
3552 add_preempt_count(PREEMPT_ACTIVE
);
3554 * We keep the big kernel semaphore locked, but we
3555 * clear ->lock_depth so that schedule() doesnt
3556 * auto-release the semaphore:
3558 #ifdef CONFIG_PREEMPT_BKL
3559 saved_lock_depth
= task
->lock_depth
;
3560 task
->lock_depth
= -1;
3563 #ifdef CONFIG_PREEMPT_BKL
3564 task
->lock_depth
= saved_lock_depth
;
3566 sub_preempt_count(PREEMPT_ACTIVE
);
3568 /* we could miss a preemption opportunity between schedule and now */
3570 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3573 EXPORT_SYMBOL(preempt_schedule
);
3576 * this is the entry point to schedule() from kernel preemption
3577 * off of irq context.
3578 * Note, that this is called and return with irqs disabled. This will
3579 * protect us against recursive calling from irq.
3581 asmlinkage
void __sched
preempt_schedule_irq(void)
3583 struct thread_info
*ti
= current_thread_info();
3584 #ifdef CONFIG_PREEMPT_BKL
3585 struct task_struct
*task
= current
;
3586 int saved_lock_depth
;
3588 /* Catch callers which need to be fixed */
3589 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3592 add_preempt_count(PREEMPT_ACTIVE
);
3594 * We keep the big kernel semaphore locked, but we
3595 * clear ->lock_depth so that schedule() doesnt
3596 * auto-release the semaphore:
3598 #ifdef CONFIG_PREEMPT_BKL
3599 saved_lock_depth
= task
->lock_depth
;
3600 task
->lock_depth
= -1;
3604 local_irq_disable();
3605 #ifdef CONFIG_PREEMPT_BKL
3606 task
->lock_depth
= saved_lock_depth
;
3608 sub_preempt_count(PREEMPT_ACTIVE
);
3610 /* we could miss a preemption opportunity between schedule and now */
3612 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3616 #endif /* CONFIG_PREEMPT */
3618 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3621 return try_to_wake_up(curr
->private, mode
, sync
);
3623 EXPORT_SYMBOL(default_wake_function
);
3626 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3627 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3628 * number) then we wake all the non-exclusive tasks and one exclusive task.
3630 * There are circumstances in which we can try to wake a task which has already
3631 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3632 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3634 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3635 int nr_exclusive
, int sync
, void *key
)
3637 struct list_head
*tmp
, *next
;
3639 list_for_each_safe(tmp
, next
, &q
->task_list
) {
3640 wait_queue_t
*curr
= list_entry(tmp
, wait_queue_t
, task_list
);
3641 unsigned flags
= curr
->flags
;
3643 if (curr
->func(curr
, mode
, sync
, key
) &&
3644 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3650 * __wake_up - wake up threads blocked on a waitqueue.
3652 * @mode: which threads
3653 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3654 * @key: is directly passed to the wakeup function
3656 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3657 int nr_exclusive
, void *key
)
3659 unsigned long flags
;
3661 spin_lock_irqsave(&q
->lock
, flags
);
3662 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3663 spin_unlock_irqrestore(&q
->lock
, flags
);
3665 EXPORT_SYMBOL(__wake_up
);
3668 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3670 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3672 __wake_up_common(q
, mode
, 1, 0, NULL
);
3676 * __wake_up_sync - wake up threads blocked on a waitqueue.
3678 * @mode: which threads
3679 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3681 * The sync wakeup differs that the waker knows that it will schedule
3682 * away soon, so while the target thread will be woken up, it will not
3683 * be migrated to another CPU - ie. the two threads are 'synchronized'
3684 * with each other. This can prevent needless bouncing between CPUs.
3686 * On UP it can prevent extra preemption.
3689 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3691 unsigned long flags
;
3697 if (unlikely(!nr_exclusive
))
3700 spin_lock_irqsave(&q
->lock
, flags
);
3701 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3702 spin_unlock_irqrestore(&q
->lock
, flags
);
3704 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3706 void fastcall
complete(struct completion
*x
)
3708 unsigned long flags
;
3710 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3712 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3714 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3716 EXPORT_SYMBOL(complete
);
3718 void fastcall
complete_all(struct completion
*x
)
3720 unsigned long flags
;
3722 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3723 x
->done
+= UINT_MAX
/2;
3724 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3726 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3728 EXPORT_SYMBOL(complete_all
);
3730 void fastcall __sched
wait_for_completion(struct completion
*x
)
3734 spin_lock_irq(&x
->wait
.lock
);
3736 DECLARE_WAITQUEUE(wait
, current
);
3738 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3739 __add_wait_queue_tail(&x
->wait
, &wait
);
3741 __set_current_state(TASK_UNINTERRUPTIBLE
);
3742 spin_unlock_irq(&x
->wait
.lock
);
3744 spin_lock_irq(&x
->wait
.lock
);
3746 __remove_wait_queue(&x
->wait
, &wait
);
3749 spin_unlock_irq(&x
->wait
.lock
);
3751 EXPORT_SYMBOL(wait_for_completion
);
3753 unsigned long fastcall __sched
3754 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3758 spin_lock_irq(&x
->wait
.lock
);
3760 DECLARE_WAITQUEUE(wait
, current
);
3762 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3763 __add_wait_queue_tail(&x
->wait
, &wait
);
3765 __set_current_state(TASK_UNINTERRUPTIBLE
);
3766 spin_unlock_irq(&x
->wait
.lock
);
3767 timeout
= schedule_timeout(timeout
);
3768 spin_lock_irq(&x
->wait
.lock
);
3770 __remove_wait_queue(&x
->wait
, &wait
);
3774 __remove_wait_queue(&x
->wait
, &wait
);
3778 spin_unlock_irq(&x
->wait
.lock
);
3781 EXPORT_SYMBOL(wait_for_completion_timeout
);
3783 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3789 spin_lock_irq(&x
->wait
.lock
);
3791 DECLARE_WAITQUEUE(wait
, current
);
3793 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3794 __add_wait_queue_tail(&x
->wait
, &wait
);
3796 if (signal_pending(current
)) {
3798 __remove_wait_queue(&x
->wait
, &wait
);
3801 __set_current_state(TASK_INTERRUPTIBLE
);
3802 spin_unlock_irq(&x
->wait
.lock
);
3804 spin_lock_irq(&x
->wait
.lock
);
3806 __remove_wait_queue(&x
->wait
, &wait
);
3810 spin_unlock_irq(&x
->wait
.lock
);
3814 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3816 unsigned long fastcall __sched
3817 wait_for_completion_interruptible_timeout(struct completion
*x
,
3818 unsigned long timeout
)
3822 spin_lock_irq(&x
->wait
.lock
);
3824 DECLARE_WAITQUEUE(wait
, current
);
3826 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3827 __add_wait_queue_tail(&x
->wait
, &wait
);
3829 if (signal_pending(current
)) {
3830 timeout
= -ERESTARTSYS
;
3831 __remove_wait_queue(&x
->wait
, &wait
);
3834 __set_current_state(TASK_INTERRUPTIBLE
);
3835 spin_unlock_irq(&x
->wait
.lock
);
3836 timeout
= schedule_timeout(timeout
);
3837 spin_lock_irq(&x
->wait
.lock
);
3839 __remove_wait_queue(&x
->wait
, &wait
);
3843 __remove_wait_queue(&x
->wait
, &wait
);
3847 spin_unlock_irq(&x
->wait
.lock
);
3850 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3853 #define SLEEP_ON_VAR \
3854 unsigned long flags; \
3855 wait_queue_t wait; \
3856 init_waitqueue_entry(&wait, current);
3858 #define SLEEP_ON_HEAD \
3859 spin_lock_irqsave(&q->lock,flags); \
3860 __add_wait_queue(q, &wait); \
3861 spin_unlock(&q->lock);
3863 #define SLEEP_ON_TAIL \
3864 spin_lock_irq(&q->lock); \
3865 __remove_wait_queue(q, &wait); \
3866 spin_unlock_irqrestore(&q->lock, flags);
3868 void fastcall __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3872 current
->state
= TASK_INTERRUPTIBLE
;
3878 EXPORT_SYMBOL(interruptible_sleep_on
);
3880 long fastcall __sched
3881 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3885 current
->state
= TASK_INTERRUPTIBLE
;
3888 timeout
= schedule_timeout(timeout
);
3893 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3895 void fastcall __sched
sleep_on(wait_queue_head_t
*q
)
3899 current
->state
= TASK_UNINTERRUPTIBLE
;
3905 EXPORT_SYMBOL(sleep_on
);
3907 long fastcall __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3911 current
->state
= TASK_UNINTERRUPTIBLE
;
3914 timeout
= schedule_timeout(timeout
);
3920 EXPORT_SYMBOL(sleep_on_timeout
);
3922 #ifdef CONFIG_RT_MUTEXES
3925 * rt_mutex_setprio - set the current priority of a task
3927 * @prio: prio value (kernel-internal form)
3929 * This function changes the 'effective' priority of a task. It does
3930 * not touch ->normal_prio like __setscheduler().
3932 * Used by the rt_mutex code to implement priority inheritance logic.
3934 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3936 struct prio_array
*array
;
3937 unsigned long flags
;
3941 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3943 rq
= task_rq_lock(p
, &flags
);
3948 dequeue_task(p
, array
);
3953 * If changing to an RT priority then queue it
3954 * in the active array!
3958 enqueue_task(p
, array
);
3960 * Reschedule if we are currently running on this runqueue and
3961 * our priority decreased, or if we are not currently running on
3962 * this runqueue and our priority is higher than the current's
3964 if (task_running(rq
, p
)) {
3965 if (p
->prio
> oldprio
)
3966 resched_task(rq
->curr
);
3967 } else if (TASK_PREEMPTS_CURR(p
, rq
))
3968 resched_task(rq
->curr
);
3970 task_rq_unlock(rq
, &flags
);
3975 void set_user_nice(struct task_struct
*p
, long nice
)
3977 struct prio_array
*array
;
3978 int old_prio
, delta
;
3979 unsigned long flags
;
3982 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3985 * We have to be careful, if called from sys_setpriority(),
3986 * the task might be in the middle of scheduling on another CPU.
3988 rq
= task_rq_lock(p
, &flags
);
3990 * The RT priorities are set via sched_setscheduler(), but we still
3991 * allow the 'normal' nice value to be set - but as expected
3992 * it wont have any effect on scheduling until the task is
3993 * not SCHED_NORMAL/SCHED_BATCH:
3995 if (has_rt_policy(p
)) {
3996 p
->static_prio
= NICE_TO_PRIO(nice
);
4001 dequeue_task(p
, array
);
4002 dec_raw_weighted_load(rq
, p
);
4005 p
->static_prio
= NICE_TO_PRIO(nice
);
4008 p
->prio
= effective_prio(p
);
4009 delta
= p
->prio
- old_prio
;
4012 enqueue_task(p
, array
);
4013 inc_raw_weighted_load(rq
, p
);
4015 * If the task increased its priority or is running and
4016 * lowered its priority, then reschedule its CPU:
4018 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4019 resched_task(rq
->curr
);
4022 task_rq_unlock(rq
, &flags
);
4024 EXPORT_SYMBOL(set_user_nice
);
4027 * can_nice - check if a task can reduce its nice value
4031 int can_nice(const struct task_struct
*p
, const int nice
)
4033 /* convert nice value [19,-20] to rlimit style value [1,40] */
4034 int nice_rlim
= 20 - nice
;
4036 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4037 capable(CAP_SYS_NICE
));
4040 #ifdef __ARCH_WANT_SYS_NICE
4043 * sys_nice - change the priority of the current process.
4044 * @increment: priority increment
4046 * sys_setpriority is a more generic, but much slower function that
4047 * does similar things.
4049 asmlinkage
long sys_nice(int increment
)
4054 * Setpriority might change our priority at the same moment.
4055 * We don't have to worry. Conceptually one call occurs first
4056 * and we have a single winner.
4058 if (increment
< -40)
4063 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4069 if (increment
< 0 && !can_nice(current
, nice
))
4072 retval
= security_task_setnice(current
, nice
);
4076 set_user_nice(current
, nice
);
4083 * task_prio - return the priority value of a given task.
4084 * @p: the task in question.
4086 * This is the priority value as seen by users in /proc.
4087 * RT tasks are offset by -200. Normal tasks are centered
4088 * around 0, value goes from -16 to +15.
4090 int task_prio(const struct task_struct
*p
)
4092 return p
->prio
- MAX_RT_PRIO
;
4096 * task_nice - return the nice value of a given task.
4097 * @p: the task in question.
4099 int task_nice(const struct task_struct
*p
)
4101 return TASK_NICE(p
);
4103 EXPORT_SYMBOL_GPL(task_nice
);
4106 * idle_cpu - is a given cpu idle currently?
4107 * @cpu: the processor in question.
4109 int idle_cpu(int cpu
)
4111 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4115 * idle_task - return the idle task for a given cpu.
4116 * @cpu: the processor in question.
4118 struct task_struct
*idle_task(int cpu
)
4120 return cpu_rq(cpu
)->idle
;
4124 * find_process_by_pid - find a process with a matching PID value.
4125 * @pid: the pid in question.
4127 static inline struct task_struct
*find_process_by_pid(pid_t pid
)
4129 return pid
? find_task_by_pid(pid
) : current
;
4132 /* Actually do priority change: must hold rq lock. */
4133 static void __setscheduler(struct task_struct
*p
, int policy
, int prio
)
4138 p
->rt_priority
= prio
;
4139 p
->normal_prio
= normal_prio(p
);
4140 /* we are holding p->pi_lock already */
4141 p
->prio
= rt_mutex_getprio(p
);
4143 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
4145 if (policy
== SCHED_BATCH
)
4151 * sched_setscheduler - change the scheduling policy and/or RT priority of
4153 * @p: the task in question.
4154 * @policy: new policy.
4155 * @param: structure containing the new RT priority.
4157 * NOTE: the task may be already dead
4159 int sched_setscheduler(struct task_struct
*p
, int policy
,
4160 struct sched_param
*param
)
4162 int retval
, oldprio
, oldpolicy
= -1;
4163 struct prio_array
*array
;
4164 unsigned long flags
;
4167 /* may grab non-irq protected spin_locks */
4168 BUG_ON(in_interrupt());
4170 /* double check policy once rq lock held */
4172 policy
= oldpolicy
= p
->policy
;
4173 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4174 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
)
4177 * Valid priorities for SCHED_FIFO and SCHED_RR are
4178 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
4181 if (param
->sched_priority
< 0 ||
4182 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4183 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4185 if (is_rt_policy(policy
) != (param
->sched_priority
!= 0))
4189 * Allow unprivileged RT tasks to decrease priority:
4191 if (!capable(CAP_SYS_NICE
)) {
4192 if (is_rt_policy(policy
)) {
4193 unsigned long rlim_rtprio
;
4194 unsigned long flags
;
4196 if (!lock_task_sighand(p
, &flags
))
4198 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4199 unlock_task_sighand(p
, &flags
);
4201 /* can't set/change the rt policy */
4202 if (policy
!= p
->policy
&& !rlim_rtprio
)
4205 /* can't increase priority */
4206 if (param
->sched_priority
> p
->rt_priority
&&
4207 param
->sched_priority
> rlim_rtprio
)
4211 /* can't change other user's priorities */
4212 if ((current
->euid
!= p
->euid
) &&
4213 (current
->euid
!= p
->uid
))
4217 retval
= security_task_setscheduler(p
, policy
, param
);
4221 * make sure no PI-waiters arrive (or leave) while we are
4222 * changing the priority of the task:
4224 spin_lock_irqsave(&p
->pi_lock
, flags
);
4226 * To be able to change p->policy safely, the apropriate
4227 * runqueue lock must be held.
4229 rq
= __task_rq_lock(p
);
4230 /* recheck policy now with rq lock held */
4231 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4232 policy
= oldpolicy
= -1;
4233 __task_rq_unlock(rq
);
4234 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4239 deactivate_task(p
, rq
);
4241 __setscheduler(p
, policy
, param
->sched_priority
);
4243 __activate_task(p
, rq
);
4245 * Reschedule if we are currently running on this runqueue and
4246 * our priority decreased, or if we are not currently running on
4247 * this runqueue and our priority is higher than the current's
4249 if (task_running(rq
, p
)) {
4250 if (p
->prio
> oldprio
)
4251 resched_task(rq
->curr
);
4252 } else if (TASK_PREEMPTS_CURR(p
, rq
))
4253 resched_task(rq
->curr
);
4255 __task_rq_unlock(rq
);
4256 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4258 rt_mutex_adjust_pi(p
);
4262 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4265 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4267 struct sched_param lparam
;
4268 struct task_struct
*p
;
4271 if (!param
|| pid
< 0)
4273 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4278 p
= find_process_by_pid(pid
);
4280 retval
= sched_setscheduler(p
, policy
, &lparam
);
4287 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4288 * @pid: the pid in question.
4289 * @policy: new policy.
4290 * @param: structure containing the new RT priority.
4292 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
4293 struct sched_param __user
*param
)
4295 /* negative values for policy are not valid */
4299 return do_sched_setscheduler(pid
, policy
, param
);
4303 * sys_sched_setparam - set/change the RT priority of a thread
4304 * @pid: the pid in question.
4305 * @param: structure containing the new RT priority.
4307 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4309 return do_sched_setscheduler(pid
, -1, param
);
4313 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4314 * @pid: the pid in question.
4316 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4318 struct task_struct
*p
;
4319 int retval
= -EINVAL
;
4325 read_lock(&tasklist_lock
);
4326 p
= find_process_by_pid(pid
);
4328 retval
= security_task_getscheduler(p
);
4332 read_unlock(&tasklist_lock
);
4339 * sys_sched_getscheduler - get the RT priority of a thread
4340 * @pid: the pid in question.
4341 * @param: structure containing the RT priority.
4343 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4345 struct sched_param lp
;
4346 struct task_struct
*p
;
4347 int retval
= -EINVAL
;
4349 if (!param
|| pid
< 0)
4352 read_lock(&tasklist_lock
);
4353 p
= find_process_by_pid(pid
);
4358 retval
= security_task_getscheduler(p
);
4362 lp
.sched_priority
= p
->rt_priority
;
4363 read_unlock(&tasklist_lock
);
4366 * This one might sleep, we cannot do it with a spinlock held ...
4368 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4374 read_unlock(&tasklist_lock
);
4378 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4380 cpumask_t cpus_allowed
;
4381 struct task_struct
*p
;
4385 read_lock(&tasklist_lock
);
4387 p
= find_process_by_pid(pid
);
4389 read_unlock(&tasklist_lock
);
4390 unlock_cpu_hotplug();
4395 * It is not safe to call set_cpus_allowed with the
4396 * tasklist_lock held. We will bump the task_struct's
4397 * usage count and then drop tasklist_lock.
4400 read_unlock(&tasklist_lock
);
4403 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4404 !capable(CAP_SYS_NICE
))
4407 retval
= security_task_setscheduler(p
, 0, NULL
);
4411 cpus_allowed
= cpuset_cpus_allowed(p
);
4412 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4413 retval
= set_cpus_allowed(p
, new_mask
);
4417 unlock_cpu_hotplug();
4421 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4422 cpumask_t
*new_mask
)
4424 if (len
< sizeof(cpumask_t
)) {
4425 memset(new_mask
, 0, sizeof(cpumask_t
));
4426 } else if (len
> sizeof(cpumask_t
)) {
4427 len
= sizeof(cpumask_t
);
4429 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4433 * sys_sched_setaffinity - set the cpu affinity of a process
4434 * @pid: pid of the process
4435 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4436 * @user_mask_ptr: user-space pointer to the new cpu mask
4438 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4439 unsigned long __user
*user_mask_ptr
)
4444 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4448 return sched_setaffinity(pid
, new_mask
);
4452 * Represents all cpu's present in the system
4453 * In systems capable of hotplug, this map could dynamically grow
4454 * as new cpu's are detected in the system via any platform specific
4455 * method, such as ACPI for e.g.
4458 cpumask_t cpu_present_map __read_mostly
;
4459 EXPORT_SYMBOL(cpu_present_map
);
4462 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4463 EXPORT_SYMBOL(cpu_online_map
);
4465 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4466 EXPORT_SYMBOL(cpu_possible_map
);
4469 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4471 struct task_struct
*p
;
4475 read_lock(&tasklist_lock
);
4478 p
= find_process_by_pid(pid
);
4482 retval
= security_task_getscheduler(p
);
4486 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4489 read_unlock(&tasklist_lock
);
4490 unlock_cpu_hotplug();
4498 * sys_sched_getaffinity - get the cpu affinity of a process
4499 * @pid: pid of the process
4500 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4501 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4503 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4504 unsigned long __user
*user_mask_ptr
)
4509 if (len
< sizeof(cpumask_t
))
4512 ret
= sched_getaffinity(pid
, &mask
);
4516 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4519 return sizeof(cpumask_t
);
4523 * sys_sched_yield - yield the current processor to other threads.
4525 * this function yields the current CPU by moving the calling thread
4526 * to the expired array. If there are no other threads running on this
4527 * CPU then this function will return.
4529 asmlinkage
long sys_sched_yield(void)
4531 struct rq
*rq
= this_rq_lock();
4532 struct prio_array
*array
= current
->array
, *target
= rq
->expired
;
4534 schedstat_inc(rq
, yld_cnt
);
4536 * We implement yielding by moving the task into the expired
4539 * (special rule: RT tasks will just roundrobin in the active
4542 if (rt_task(current
))
4543 target
= rq
->active
;
4545 if (array
->nr_active
== 1) {
4546 schedstat_inc(rq
, yld_act_empty
);
4547 if (!rq
->expired
->nr_active
)
4548 schedstat_inc(rq
, yld_both_empty
);
4549 } else if (!rq
->expired
->nr_active
)
4550 schedstat_inc(rq
, yld_exp_empty
);
4552 if (array
!= target
) {
4553 dequeue_task(current
, array
);
4554 enqueue_task(current
, target
);
4557 * requeue_task is cheaper so perform that if possible.
4559 requeue_task(current
, array
);
4562 * Since we are going to call schedule() anyway, there's
4563 * no need to preempt or enable interrupts:
4565 __release(rq
->lock
);
4566 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4567 _raw_spin_unlock(&rq
->lock
);
4568 preempt_enable_no_resched();
4575 static inline int __resched_legal(int expected_preempt_count
)
4577 if (unlikely(preempt_count() != expected_preempt_count
))
4579 if (unlikely(system_state
!= SYSTEM_RUNNING
))
4584 static void __cond_resched(void)
4586 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4587 __might_sleep(__FILE__
, __LINE__
);
4590 * The BKS might be reacquired before we have dropped
4591 * PREEMPT_ACTIVE, which could trigger a second
4592 * cond_resched() call.
4595 add_preempt_count(PREEMPT_ACTIVE
);
4597 sub_preempt_count(PREEMPT_ACTIVE
);
4598 } while (need_resched());
4601 int __sched
cond_resched(void)
4603 if (need_resched() && __resched_legal(0)) {
4609 EXPORT_SYMBOL(cond_resched
);
4612 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4613 * call schedule, and on return reacquire the lock.
4615 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4616 * operations here to prevent schedule() from being called twice (once via
4617 * spin_unlock(), once by hand).
4619 int cond_resched_lock(spinlock_t
*lock
)
4623 if (need_lockbreak(lock
)) {
4629 if (need_resched() && __resched_legal(1)) {
4630 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4631 _raw_spin_unlock(lock
);
4632 preempt_enable_no_resched();
4639 EXPORT_SYMBOL(cond_resched_lock
);
4641 int __sched
cond_resched_softirq(void)
4643 BUG_ON(!in_softirq());
4645 if (need_resched() && __resched_legal(0)) {
4646 raw_local_irq_disable();
4648 raw_local_irq_enable();
4655 EXPORT_SYMBOL(cond_resched_softirq
);
4658 * yield - yield the current processor to other threads.
4660 * this is a shortcut for kernel-space yielding - it marks the
4661 * thread runnable and calls sys_sched_yield().
4663 void __sched
yield(void)
4665 set_current_state(TASK_RUNNING
);
4668 EXPORT_SYMBOL(yield
);
4671 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4672 * that process accounting knows that this is a task in IO wait state.
4674 * But don't do that if it is a deliberate, throttling IO wait (this task
4675 * has set its backing_dev_info: the queue against which it should throttle)
4677 void __sched
io_schedule(void)
4679 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4681 delayacct_blkio_start();
4682 atomic_inc(&rq
->nr_iowait
);
4684 atomic_dec(&rq
->nr_iowait
);
4685 delayacct_blkio_end();
4687 EXPORT_SYMBOL(io_schedule
);
4689 long __sched
io_schedule_timeout(long timeout
)
4691 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4694 delayacct_blkio_start();
4695 atomic_inc(&rq
->nr_iowait
);
4696 ret
= schedule_timeout(timeout
);
4697 atomic_dec(&rq
->nr_iowait
);
4698 delayacct_blkio_end();
4703 * sys_sched_get_priority_max - return maximum RT priority.
4704 * @policy: scheduling class.
4706 * this syscall returns the maximum rt_priority that can be used
4707 * by a given scheduling class.
4709 asmlinkage
long sys_sched_get_priority_max(int policy
)
4716 ret
= MAX_USER_RT_PRIO
-1;
4727 * sys_sched_get_priority_min - return minimum RT priority.
4728 * @policy: scheduling class.
4730 * this syscall returns the minimum rt_priority that can be used
4731 * by a given scheduling class.
4733 asmlinkage
long sys_sched_get_priority_min(int policy
)
4750 * sys_sched_rr_get_interval - return the default timeslice of a process.
4751 * @pid: pid of the process.
4752 * @interval: userspace pointer to the timeslice value.
4754 * this syscall writes the default timeslice value of a given process
4755 * into the user-space timespec buffer. A value of '0' means infinity.
4758 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4760 struct task_struct
*p
;
4761 int retval
= -EINVAL
;
4768 read_lock(&tasklist_lock
);
4769 p
= find_process_by_pid(pid
);
4773 retval
= security_task_getscheduler(p
);
4777 jiffies_to_timespec(p
->policy
== SCHED_FIFO
?
4778 0 : task_timeslice(p
), &t
);
4779 read_unlock(&tasklist_lock
);
4780 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4784 read_unlock(&tasklist_lock
);
4788 static inline struct task_struct
*eldest_child(struct task_struct
*p
)
4790 if (list_empty(&p
->children
))
4792 return list_entry(p
->children
.next
,struct task_struct
,sibling
);
4795 static inline struct task_struct
*older_sibling(struct task_struct
*p
)
4797 if (p
->sibling
.prev
==&p
->parent
->children
)
4799 return list_entry(p
->sibling
.prev
,struct task_struct
,sibling
);
4802 static inline struct task_struct
*younger_sibling(struct task_struct
*p
)
4804 if (p
->sibling
.next
==&p
->parent
->children
)
4806 return list_entry(p
->sibling
.next
,struct task_struct
,sibling
);
4809 static const char stat_nam
[] = "RSDTtZX";
4811 static void show_task(struct task_struct
*p
)
4813 struct task_struct
*relative
;
4814 unsigned long free
= 0;
4817 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4818 printk("%-13.13s %c", p
->comm
,
4819 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4820 #if (BITS_PER_LONG == 32)
4821 if (state
== TASK_RUNNING
)
4822 printk(" running ");
4824 printk(" %08lX ", thread_saved_pc(p
));
4826 if (state
== TASK_RUNNING
)
4827 printk(" running task ");
4829 printk(" %016lx ", thread_saved_pc(p
));
4831 #ifdef CONFIG_DEBUG_STACK_USAGE
4833 unsigned long *n
= end_of_stack(p
);
4836 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4839 printk("%5lu %5d %6d ", free
, p
->pid
, p
->parent
->pid
);
4840 if ((relative
= eldest_child(p
)))
4841 printk("%5d ", relative
->pid
);
4844 if ((relative
= younger_sibling(p
)))
4845 printk("%7d", relative
->pid
);
4848 if ((relative
= older_sibling(p
)))
4849 printk(" %5d", relative
->pid
);
4853 printk(" (L-TLB)\n");
4855 printk(" (NOTLB)\n");
4857 if (state
!= TASK_RUNNING
)
4858 show_stack(p
, NULL
);
4861 void show_state_filter(unsigned long state_filter
)
4863 struct task_struct
*g
, *p
;
4865 #if (BITS_PER_LONG == 32)
4868 printk(" task PC stack pid father child younger older\n");
4872 printk(" task PC stack pid father child younger older\n");
4874 read_lock(&tasklist_lock
);
4875 do_each_thread(g
, p
) {
4877 * reset the NMI-timeout, listing all files on a slow
4878 * console might take alot of time:
4880 touch_nmi_watchdog();
4881 if (p
->state
& state_filter
)
4883 } while_each_thread(g
, p
);
4885 read_unlock(&tasklist_lock
);
4887 * Only show locks if all tasks are dumped:
4889 if (state_filter
== -1)
4890 debug_show_all_locks();
4894 * init_idle - set up an idle thread for a given CPU
4895 * @idle: task in question
4896 * @cpu: cpu the idle task belongs to
4898 * NOTE: this function does not set the idle thread's NEED_RESCHED
4899 * flag, to make booting more robust.
4901 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4903 struct rq
*rq
= cpu_rq(cpu
);
4904 unsigned long flags
;
4906 idle
->timestamp
= sched_clock();
4907 idle
->sleep_avg
= 0;
4909 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
4910 idle
->state
= TASK_RUNNING
;
4911 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4912 set_task_cpu(idle
, cpu
);
4914 spin_lock_irqsave(&rq
->lock
, flags
);
4915 rq
->curr
= rq
->idle
= idle
;
4916 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4919 spin_unlock_irqrestore(&rq
->lock
, flags
);
4921 /* Set the preempt count _outside_ the spinlocks! */
4922 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4923 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4925 task_thread_info(idle
)->preempt_count
= 0;
4930 * In a system that switches off the HZ timer nohz_cpu_mask
4931 * indicates which cpus entered this state. This is used
4932 * in the rcu update to wait only for active cpus. For system
4933 * which do not switch off the HZ timer nohz_cpu_mask should
4934 * always be CPU_MASK_NONE.
4936 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4940 * This is how migration works:
4942 * 1) we queue a struct migration_req structure in the source CPU's
4943 * runqueue and wake up that CPU's migration thread.
4944 * 2) we down() the locked semaphore => thread blocks.
4945 * 3) migration thread wakes up (implicitly it forces the migrated
4946 * thread off the CPU)
4947 * 4) it gets the migration request and checks whether the migrated
4948 * task is still in the wrong runqueue.
4949 * 5) if it's in the wrong runqueue then the migration thread removes
4950 * it and puts it into the right queue.
4951 * 6) migration thread up()s the semaphore.
4952 * 7) we wake up and the migration is done.
4956 * Change a given task's CPU affinity. Migrate the thread to a
4957 * proper CPU and schedule it away if the CPU it's executing on
4958 * is removed from the allowed bitmask.
4960 * NOTE: the caller must have a valid reference to the task, the
4961 * task must not exit() & deallocate itself prematurely. The
4962 * call is not atomic; no spinlocks may be held.
4964 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
4966 struct migration_req req
;
4967 unsigned long flags
;
4971 rq
= task_rq_lock(p
, &flags
);
4972 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4977 p
->cpus_allowed
= new_mask
;
4978 /* Can the task run on the task's current CPU? If so, we're done */
4979 if (cpu_isset(task_cpu(p
), new_mask
))
4982 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4983 /* Need help from migration thread: drop lock and wait. */
4984 task_rq_unlock(rq
, &flags
);
4985 wake_up_process(rq
->migration_thread
);
4986 wait_for_completion(&req
.done
);
4987 tlb_migrate_finish(p
->mm
);
4991 task_rq_unlock(rq
, &flags
);
4995 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4998 * Move (not current) task off this cpu, onto dest cpu. We're doing
4999 * this because either it can't run here any more (set_cpus_allowed()
5000 * away from this CPU, or CPU going down), or because we're
5001 * attempting to rebalance this task on exec (sched_exec).
5003 * So we race with normal scheduler movements, but that's OK, as long
5004 * as the task is no longer on this CPU.
5006 * Returns non-zero if task was successfully migrated.
5008 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5010 struct rq
*rq_dest
, *rq_src
;
5013 if (unlikely(cpu_is_offline(dest_cpu
)))
5016 rq_src
= cpu_rq(src_cpu
);
5017 rq_dest
= cpu_rq(dest_cpu
);
5019 double_rq_lock(rq_src
, rq_dest
);
5020 /* Already moved. */
5021 if (task_cpu(p
) != src_cpu
)
5023 /* Affinity changed (again). */
5024 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
5027 set_task_cpu(p
, dest_cpu
);
5030 * Sync timestamp with rq_dest's before activating.
5031 * The same thing could be achieved by doing this step
5032 * afterwards, and pretending it was a local activate.
5033 * This way is cleaner and logically correct.
5035 p
->timestamp
= p
->timestamp
- rq_src
->timestamp_last_tick
5036 + rq_dest
->timestamp_last_tick
;
5037 deactivate_task(p
, rq_src
);
5038 __activate_task(p
, rq_dest
);
5039 if (TASK_PREEMPTS_CURR(p
, rq_dest
))
5040 resched_task(rq_dest
->curr
);
5044 double_rq_unlock(rq_src
, rq_dest
);
5049 * migration_thread - this is a highprio system thread that performs
5050 * thread migration by bumping thread off CPU then 'pushing' onto
5053 static int migration_thread(void *data
)
5055 int cpu
= (long)data
;
5059 BUG_ON(rq
->migration_thread
!= current
);
5061 set_current_state(TASK_INTERRUPTIBLE
);
5062 while (!kthread_should_stop()) {
5063 struct migration_req
*req
;
5064 struct list_head
*head
;
5068 spin_lock_irq(&rq
->lock
);
5070 if (cpu_is_offline(cpu
)) {
5071 spin_unlock_irq(&rq
->lock
);
5075 if (rq
->active_balance
) {
5076 active_load_balance(rq
, cpu
);
5077 rq
->active_balance
= 0;
5080 head
= &rq
->migration_queue
;
5082 if (list_empty(head
)) {
5083 spin_unlock_irq(&rq
->lock
);
5085 set_current_state(TASK_INTERRUPTIBLE
);
5088 req
= list_entry(head
->next
, struct migration_req
, list
);
5089 list_del_init(head
->next
);
5091 spin_unlock(&rq
->lock
);
5092 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5095 complete(&req
->done
);
5097 __set_current_state(TASK_RUNNING
);
5101 /* Wait for kthread_stop */
5102 set_current_state(TASK_INTERRUPTIBLE
);
5103 while (!kthread_should_stop()) {
5105 set_current_state(TASK_INTERRUPTIBLE
);
5107 __set_current_state(TASK_RUNNING
);
5111 #ifdef CONFIG_HOTPLUG_CPU
5113 * Figure out where task on dead CPU should go, use force if neccessary.
5114 * NOTE: interrupts should be disabled by the caller
5116 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5118 unsigned long flags
;
5125 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5126 cpus_and(mask
, mask
, p
->cpus_allowed
);
5127 dest_cpu
= any_online_cpu(mask
);
5129 /* On any allowed CPU? */
5130 if (dest_cpu
== NR_CPUS
)
5131 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5133 /* No more Mr. Nice Guy. */
5134 if (dest_cpu
== NR_CPUS
) {
5135 rq
= task_rq_lock(p
, &flags
);
5136 cpus_setall(p
->cpus_allowed
);
5137 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5138 task_rq_unlock(rq
, &flags
);
5141 * Don't tell them about moving exiting tasks or
5142 * kernel threads (both mm NULL), since they never
5145 if (p
->mm
&& printk_ratelimit())
5146 printk(KERN_INFO
"process %d (%s) no "
5147 "longer affine to cpu%d\n",
5148 p
->pid
, p
->comm
, dead_cpu
);
5150 if (!__migrate_task(p
, dead_cpu
, dest_cpu
))
5155 * While a dead CPU has no uninterruptible tasks queued at this point,
5156 * it might still have a nonzero ->nr_uninterruptible counter, because
5157 * for performance reasons the counter is not stricly tracking tasks to
5158 * their home CPUs. So we just add the counter to another CPU's counter,
5159 * to keep the global sum constant after CPU-down:
5161 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5163 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5164 unsigned long flags
;
5166 local_irq_save(flags
);
5167 double_rq_lock(rq_src
, rq_dest
);
5168 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5169 rq_src
->nr_uninterruptible
= 0;
5170 double_rq_unlock(rq_src
, rq_dest
);
5171 local_irq_restore(flags
);
5174 /* Run through task list and migrate tasks from the dead cpu. */
5175 static void migrate_live_tasks(int src_cpu
)
5177 struct task_struct
*p
, *t
;
5179 write_lock_irq(&tasklist_lock
);
5181 do_each_thread(t
, p
) {
5185 if (task_cpu(p
) == src_cpu
)
5186 move_task_off_dead_cpu(src_cpu
, p
);
5187 } while_each_thread(t
, p
);
5189 write_unlock_irq(&tasklist_lock
);
5192 /* Schedules idle task to be the next runnable task on current CPU.
5193 * It does so by boosting its priority to highest possible and adding it to
5194 * the _front_ of the runqueue. Used by CPU offline code.
5196 void sched_idle_next(void)
5198 int this_cpu
= smp_processor_id();
5199 struct rq
*rq
= cpu_rq(this_cpu
);
5200 struct task_struct
*p
= rq
->idle
;
5201 unsigned long flags
;
5203 /* cpu has to be offline */
5204 BUG_ON(cpu_online(this_cpu
));
5207 * Strictly not necessary since rest of the CPUs are stopped by now
5208 * and interrupts disabled on the current cpu.
5210 spin_lock_irqsave(&rq
->lock
, flags
);
5212 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5214 /* Add idle task to the _front_ of its priority queue: */
5215 __activate_idle_task(p
, rq
);
5217 spin_unlock_irqrestore(&rq
->lock
, flags
);
5221 * Ensures that the idle task is using init_mm right before its cpu goes
5224 void idle_task_exit(void)
5226 struct mm_struct
*mm
= current
->active_mm
;
5228 BUG_ON(cpu_online(smp_processor_id()));
5231 switch_mm(mm
, &init_mm
, current
);
5235 /* called under rq->lock with disabled interrupts */
5236 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5238 struct rq
*rq
= cpu_rq(dead_cpu
);
5240 /* Must be exiting, otherwise would be on tasklist. */
5241 BUG_ON(p
->exit_state
!= EXIT_ZOMBIE
&& p
->exit_state
!= EXIT_DEAD
);
5243 /* Cannot have done final schedule yet: would have vanished. */
5244 BUG_ON(p
->state
== TASK_DEAD
);
5249 * Drop lock around migration; if someone else moves it,
5250 * that's OK. No task can be added to this CPU, so iteration is
5252 * NOTE: interrupts should be left disabled --dev@
5254 spin_unlock(&rq
->lock
);
5255 move_task_off_dead_cpu(dead_cpu
, p
);
5256 spin_lock(&rq
->lock
);
5261 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5262 static void migrate_dead_tasks(unsigned int dead_cpu
)
5264 struct rq
*rq
= cpu_rq(dead_cpu
);
5265 unsigned int arr
, i
;
5267 for (arr
= 0; arr
< 2; arr
++) {
5268 for (i
= 0; i
< MAX_PRIO
; i
++) {
5269 struct list_head
*list
= &rq
->arrays
[arr
].queue
[i
];
5271 while (!list_empty(list
))
5272 migrate_dead(dead_cpu
, list_entry(list
->next
,
5273 struct task_struct
, run_list
));
5277 #endif /* CONFIG_HOTPLUG_CPU */
5280 * migration_call - callback that gets triggered when a CPU is added.
5281 * Here we can start up the necessary migration thread for the new CPU.
5283 static int __cpuinit
5284 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5286 struct task_struct
*p
;
5287 int cpu
= (long)hcpu
;
5288 unsigned long flags
;
5292 case CPU_UP_PREPARE
:
5293 p
= kthread_create(migration_thread
, hcpu
, "migration/%d",cpu
);
5296 p
->flags
|= PF_NOFREEZE
;
5297 kthread_bind(p
, cpu
);
5298 /* Must be high prio: stop_machine expects to yield to it. */
5299 rq
= task_rq_lock(p
, &flags
);
5300 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5301 task_rq_unlock(rq
, &flags
);
5302 cpu_rq(cpu
)->migration_thread
= p
;
5306 /* Strictly unneccessary, as first user will wake it. */
5307 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5310 #ifdef CONFIG_HOTPLUG_CPU
5311 case CPU_UP_CANCELED
:
5312 if (!cpu_rq(cpu
)->migration_thread
)
5314 /* Unbind it from offline cpu so it can run. Fall thru. */
5315 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5316 any_online_cpu(cpu_online_map
));
5317 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5318 cpu_rq(cpu
)->migration_thread
= NULL
;
5322 migrate_live_tasks(cpu
);
5324 kthread_stop(rq
->migration_thread
);
5325 rq
->migration_thread
= NULL
;
5326 /* Idle task back to normal (off runqueue, low prio) */
5327 rq
= task_rq_lock(rq
->idle
, &flags
);
5328 deactivate_task(rq
->idle
, rq
);
5329 rq
->idle
->static_prio
= MAX_PRIO
;
5330 __setscheduler(rq
->idle
, SCHED_NORMAL
, 0);
5331 migrate_dead_tasks(cpu
);
5332 task_rq_unlock(rq
, &flags
);
5333 migrate_nr_uninterruptible(rq
);
5334 BUG_ON(rq
->nr_running
!= 0);
5336 /* No need to migrate the tasks: it was best-effort if
5337 * they didn't do lock_cpu_hotplug(). Just wake up
5338 * the requestors. */
5339 spin_lock_irq(&rq
->lock
);
5340 while (!list_empty(&rq
->migration_queue
)) {
5341 struct migration_req
*req
;
5343 req
= list_entry(rq
->migration_queue
.next
,
5344 struct migration_req
, list
);
5345 list_del_init(&req
->list
);
5346 complete(&req
->done
);
5348 spin_unlock_irq(&rq
->lock
);
5355 /* Register at highest priority so that task migration (migrate_all_tasks)
5356 * happens before everything else.
5358 static struct notifier_block __cpuinitdata migration_notifier
= {
5359 .notifier_call
= migration_call
,
5363 int __init
migration_init(void)
5365 void *cpu
= (void *)(long)smp_processor_id();
5368 /* Start one for the boot CPU: */
5369 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5370 BUG_ON(err
== NOTIFY_BAD
);
5371 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5372 register_cpu_notifier(&migration_notifier
);
5379 #undef SCHED_DOMAIN_DEBUG
5380 #ifdef SCHED_DOMAIN_DEBUG
5381 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5386 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5390 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5395 struct sched_group
*group
= sd
->groups
;
5396 cpumask_t groupmask
;
5398 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5399 cpus_clear(groupmask
);
5402 for (i
= 0; i
< level
+ 1; i
++)
5404 printk("domain %d: ", level
);
5406 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5407 printk("does not load-balance\n");
5409 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain has parent");
5413 printk("span %s\n", str
);
5415 if (!cpu_isset(cpu
, sd
->span
))
5416 printk(KERN_ERR
"ERROR: domain->span does not contain CPU%d\n", cpu
);
5417 if (!cpu_isset(cpu
, group
->cpumask
))
5418 printk(KERN_ERR
"ERROR: domain->groups does not contain CPU%d\n", cpu
);
5421 for (i
= 0; i
< level
+ 2; i
++)
5427 printk(KERN_ERR
"ERROR: group is NULL\n");
5431 if (!group
->cpu_power
) {
5433 printk(KERN_ERR
"ERROR: domain->cpu_power not set\n");
5436 if (!cpus_weight(group
->cpumask
)) {
5438 printk(KERN_ERR
"ERROR: empty group\n");
5441 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5443 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5446 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5448 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5451 group
= group
->next
;
5452 } while (group
!= sd
->groups
);
5455 if (!cpus_equal(sd
->span
, groupmask
))
5456 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5462 if (!cpus_subset(groupmask
, sd
->span
))
5463 printk(KERN_ERR
"ERROR: parent span is not a superset of domain->span\n");
5469 # define sched_domain_debug(sd, cpu) do { } while (0)
5472 static int sd_degenerate(struct sched_domain
*sd
)
5474 if (cpus_weight(sd
->span
) == 1)
5477 /* Following flags need at least 2 groups */
5478 if (sd
->flags
& (SD_LOAD_BALANCE
|
5479 SD_BALANCE_NEWIDLE
|
5483 SD_SHARE_PKG_RESOURCES
)) {
5484 if (sd
->groups
!= sd
->groups
->next
)
5488 /* Following flags don't use groups */
5489 if (sd
->flags
& (SD_WAKE_IDLE
|
5498 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5500 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5502 if (sd_degenerate(parent
))
5505 if (!cpus_equal(sd
->span
, parent
->span
))
5508 /* Does parent contain flags not in child? */
5509 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5510 if (cflags
& SD_WAKE_AFFINE
)
5511 pflags
&= ~SD_WAKE_BALANCE
;
5512 /* Flags needing groups don't count if only 1 group in parent */
5513 if (parent
->groups
== parent
->groups
->next
) {
5514 pflags
&= ~(SD_LOAD_BALANCE
|
5515 SD_BALANCE_NEWIDLE
|
5519 SD_SHARE_PKG_RESOURCES
);
5521 if (~cflags
& pflags
)
5528 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5529 * hold the hotplug lock.
5531 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5533 struct rq
*rq
= cpu_rq(cpu
);
5534 struct sched_domain
*tmp
;
5536 /* Remove the sched domains which do not contribute to scheduling. */
5537 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5538 struct sched_domain
*parent
= tmp
->parent
;
5541 if (sd_parent_degenerate(tmp
, parent
)) {
5542 tmp
->parent
= parent
->parent
;
5544 parent
->parent
->child
= tmp
;
5548 if (sd
&& sd_degenerate(sd
)) {
5554 sched_domain_debug(sd
, cpu
);
5556 rcu_assign_pointer(rq
->sd
, sd
);
5559 /* cpus with isolated domains */
5560 static cpumask_t __cpuinitdata cpu_isolated_map
= CPU_MASK_NONE
;
5562 /* Setup the mask of cpus configured for isolated domains */
5563 static int __init
isolated_cpu_setup(char *str
)
5565 int ints
[NR_CPUS
], i
;
5567 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5568 cpus_clear(cpu_isolated_map
);
5569 for (i
= 1; i
<= ints
[0]; i
++)
5570 if (ints
[i
] < NR_CPUS
)
5571 cpu_set(ints
[i
], cpu_isolated_map
);
5575 __setup ("isolcpus=", isolated_cpu_setup
);
5578 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5579 * to a function which identifies what group(along with sched group) a CPU
5580 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5581 * (due to the fact that we keep track of groups covered with a cpumask_t).
5583 * init_sched_build_groups will build a circular linked list of the groups
5584 * covered by the given span, and will set each group's ->cpumask correctly,
5585 * and ->cpu_power to 0.
5588 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
5589 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
5590 struct sched_group
**sg
))
5592 struct sched_group
*first
= NULL
, *last
= NULL
;
5593 cpumask_t covered
= CPU_MASK_NONE
;
5596 for_each_cpu_mask(i
, span
) {
5597 struct sched_group
*sg
;
5598 int group
= group_fn(i
, cpu_map
, &sg
);
5601 if (cpu_isset(i
, covered
))
5604 sg
->cpumask
= CPU_MASK_NONE
;
5607 for_each_cpu_mask(j
, span
) {
5608 if (group_fn(j
, cpu_map
, NULL
) != group
)
5611 cpu_set(j
, covered
);
5612 cpu_set(j
, sg
->cpumask
);
5623 #define SD_NODES_PER_DOMAIN 16
5626 * Self-tuning task migration cost measurement between source and target CPUs.
5628 * This is done by measuring the cost of manipulating buffers of varying
5629 * sizes. For a given buffer-size here are the steps that are taken:
5631 * 1) the source CPU reads+dirties a shared buffer
5632 * 2) the target CPU reads+dirties the same shared buffer
5634 * We measure how long they take, in the following 4 scenarios:
5636 * - source: CPU1, target: CPU2 | cost1
5637 * - source: CPU2, target: CPU1 | cost2
5638 * - source: CPU1, target: CPU1 | cost3
5639 * - source: CPU2, target: CPU2 | cost4
5641 * We then calculate the cost3+cost4-cost1-cost2 difference - this is
5642 * the cost of migration.
5644 * We then start off from a small buffer-size and iterate up to larger
5645 * buffer sizes, in 5% steps - measuring each buffer-size separately, and
5646 * doing a maximum search for the cost. (The maximum cost for a migration
5647 * normally occurs when the working set size is around the effective cache
5650 #define SEARCH_SCOPE 2
5651 #define MIN_CACHE_SIZE (64*1024U)
5652 #define DEFAULT_CACHE_SIZE (5*1024*1024U)
5653 #define ITERATIONS 1
5654 #define SIZE_THRESH 130
5655 #define COST_THRESH 130
5658 * The migration cost is a function of 'domain distance'. Domain
5659 * distance is the number of steps a CPU has to iterate down its
5660 * domain tree to share a domain with the other CPU. The farther
5661 * two CPUs are from each other, the larger the distance gets.
5663 * Note that we use the distance only to cache measurement results,
5664 * the distance value is not used numerically otherwise. When two
5665 * CPUs have the same distance it is assumed that the migration
5666 * cost is the same. (this is a simplification but quite practical)
5668 #define MAX_DOMAIN_DISTANCE 32
5670 static unsigned long long migration_cost
[MAX_DOMAIN_DISTANCE
] =
5671 { [ 0 ... MAX_DOMAIN_DISTANCE
-1 ] =
5673 * Architectures may override the migration cost and thus avoid
5674 * boot-time calibration. Unit is nanoseconds. Mostly useful for
5675 * virtualized hardware:
5677 #ifdef CONFIG_DEFAULT_MIGRATION_COST
5678 CONFIG_DEFAULT_MIGRATION_COST
5685 * Allow override of migration cost - in units of microseconds.
5686 * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost
5687 * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs:
5689 static int __init
migration_cost_setup(char *str
)
5691 int ints
[MAX_DOMAIN_DISTANCE
+1], i
;
5693 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5695 printk("#ints: %d\n", ints
[0]);
5696 for (i
= 1; i
<= ints
[0]; i
++) {
5697 migration_cost
[i
-1] = (unsigned long long)ints
[i
]*1000;
5698 printk("migration_cost[%d]: %Ld\n", i
-1, migration_cost
[i
-1]);
5703 __setup ("migration_cost=", migration_cost_setup
);
5706 * Global multiplier (divisor) for migration-cutoff values,
5707 * in percentiles. E.g. use a value of 150 to get 1.5 times
5708 * longer cache-hot cutoff times.
5710 * (We scale it from 100 to 128 to long long handling easier.)
5713 #define MIGRATION_FACTOR_SCALE 128
5715 static unsigned int migration_factor
= MIGRATION_FACTOR_SCALE
;
5717 static int __init
setup_migration_factor(char *str
)
5719 get_option(&str
, &migration_factor
);
5720 migration_factor
= migration_factor
* MIGRATION_FACTOR_SCALE
/ 100;
5724 __setup("migration_factor=", setup_migration_factor
);
5727 * Estimated distance of two CPUs, measured via the number of domains
5728 * we have to pass for the two CPUs to be in the same span:
5730 static unsigned long domain_distance(int cpu1
, int cpu2
)
5732 unsigned long distance
= 0;
5733 struct sched_domain
*sd
;
5735 for_each_domain(cpu1
, sd
) {
5736 WARN_ON(!cpu_isset(cpu1
, sd
->span
));
5737 if (cpu_isset(cpu2
, sd
->span
))
5741 if (distance
>= MAX_DOMAIN_DISTANCE
) {
5743 distance
= MAX_DOMAIN_DISTANCE
-1;
5749 static unsigned int migration_debug
;
5751 static int __init
setup_migration_debug(char *str
)
5753 get_option(&str
, &migration_debug
);
5757 __setup("migration_debug=", setup_migration_debug
);
5760 * Maximum cache-size that the scheduler should try to measure.
5761 * Architectures with larger caches should tune this up during
5762 * bootup. Gets used in the domain-setup code (i.e. during SMP
5765 unsigned int max_cache_size
;
5767 static int __init
setup_max_cache_size(char *str
)
5769 get_option(&str
, &max_cache_size
);
5773 __setup("max_cache_size=", setup_max_cache_size
);
5776 * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This
5777 * is the operation that is timed, so we try to generate unpredictable
5778 * cachemisses that still end up filling the L2 cache:
5780 static void touch_cache(void *__cache
, unsigned long __size
)
5782 unsigned long size
= __size
/sizeof(long), chunk1
= size
/3,
5784 unsigned long *cache
= __cache
;
5787 for (i
= 0; i
< size
/6; i
+= 8) {
5790 case 1: cache
[size
-1-i
]++;
5791 case 2: cache
[chunk1
-i
]++;
5792 case 3: cache
[chunk1
+i
]++;
5793 case 4: cache
[chunk2
-i
]++;
5794 case 5: cache
[chunk2
+i
]++;
5800 * Measure the cache-cost of one task migration. Returns in units of nsec.
5802 static unsigned long long
5803 measure_one(void *cache
, unsigned long size
, int source
, int target
)
5805 cpumask_t mask
, saved_mask
;
5806 unsigned long long t0
, t1
, t2
, t3
, cost
;
5808 saved_mask
= current
->cpus_allowed
;
5811 * Flush source caches to RAM and invalidate them:
5816 * Migrate to the source CPU:
5818 mask
= cpumask_of_cpu(source
);
5819 set_cpus_allowed(current
, mask
);
5820 WARN_ON(smp_processor_id() != source
);
5823 * Dirty the working set:
5826 touch_cache(cache
, size
);
5830 * Migrate to the target CPU, dirty the L2 cache and access
5831 * the shared buffer. (which represents the working set
5832 * of a migrated task.)
5834 mask
= cpumask_of_cpu(target
);
5835 set_cpus_allowed(current
, mask
);
5836 WARN_ON(smp_processor_id() != target
);
5839 touch_cache(cache
, size
);
5842 cost
= t1
-t0
+ t3
-t2
;
5844 if (migration_debug
>= 2)
5845 printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n",
5846 source
, target
, t1
-t0
, t1
-t0
, t3
-t2
, cost
);
5848 * Flush target caches to RAM and invalidate them:
5852 set_cpus_allowed(current
, saved_mask
);
5858 * Measure a series of task migrations and return the average
5859 * result. Since this code runs early during bootup the system
5860 * is 'undisturbed' and the average latency makes sense.
5862 * The algorithm in essence auto-detects the relevant cache-size,
5863 * so it will properly detect different cachesizes for different
5864 * cache-hierarchies, depending on how the CPUs are connected.
5866 * Architectures can prime the upper limit of the search range via
5867 * max_cache_size, otherwise the search range defaults to 20MB...64K.
5869 static unsigned long long
5870 measure_cost(int cpu1
, int cpu2
, void *cache
, unsigned int size
)
5872 unsigned long long cost1
, cost2
;
5876 * Measure the migration cost of 'size' bytes, over an
5877 * average of 10 runs:
5879 * (We perturb the cache size by a small (0..4k)
5880 * value to compensate size/alignment related artifacts.
5881 * We also subtract the cost of the operation done on
5887 * dry run, to make sure we start off cache-cold on cpu1,
5888 * and to get any vmalloc pagefaults in advance:
5890 measure_one(cache
, size
, cpu1
, cpu2
);
5891 for (i
= 0; i
< ITERATIONS
; i
++)
5892 cost1
+= measure_one(cache
, size
- i
*1024, cpu1
, cpu2
);
5894 measure_one(cache
, size
, cpu2
, cpu1
);
5895 for (i
= 0; i
< ITERATIONS
; i
++)
5896 cost1
+= measure_one(cache
, size
- i
*1024, cpu2
, cpu1
);
5899 * (We measure the non-migrating [cached] cost on both
5900 * cpu1 and cpu2, to handle CPUs with different speeds)
5904 measure_one(cache
, size
, cpu1
, cpu1
);
5905 for (i
= 0; i
< ITERATIONS
; i
++)
5906 cost2
+= measure_one(cache
, size
- i
*1024, cpu1
, cpu1
);
5908 measure_one(cache
, size
, cpu2
, cpu2
);
5909 for (i
= 0; i
< ITERATIONS
; i
++)
5910 cost2
+= measure_one(cache
, size
- i
*1024, cpu2
, cpu2
);
5913 * Get the per-iteration migration cost:
5915 do_div(cost1
, 2*ITERATIONS
);
5916 do_div(cost2
, 2*ITERATIONS
);
5918 return cost1
- cost2
;
5921 static unsigned long long measure_migration_cost(int cpu1
, int cpu2
)
5923 unsigned long long max_cost
= 0, fluct
= 0, avg_fluct
= 0;
5924 unsigned int max_size
, size
, size_found
= 0;
5925 long long cost
= 0, prev_cost
;
5929 * Search from max_cache_size*5 down to 64K - the real relevant
5930 * cachesize has to lie somewhere inbetween.
5932 if (max_cache_size
) {
5933 max_size
= max(max_cache_size
* SEARCH_SCOPE
, MIN_CACHE_SIZE
);
5934 size
= max(max_cache_size
/ SEARCH_SCOPE
, MIN_CACHE_SIZE
);
5937 * Since we have no estimation about the relevant
5940 max_size
= DEFAULT_CACHE_SIZE
* SEARCH_SCOPE
;
5941 size
= MIN_CACHE_SIZE
;
5944 if (!cpu_online(cpu1
) || !cpu_online(cpu2
)) {
5945 printk("cpu %d and %d not both online!\n", cpu1
, cpu2
);
5950 * Allocate the working set:
5952 cache
= vmalloc(max_size
);
5954 printk("could not vmalloc %d bytes for cache!\n", 2*max_size
);
5955 return 1000000; /* return 1 msec on very small boxen */
5958 while (size
<= max_size
) {
5960 cost
= measure_cost(cpu1
, cpu2
, cache
, size
);
5966 if (max_cost
< cost
) {
5972 * Calculate average fluctuation, we use this to prevent
5973 * noise from triggering an early break out of the loop:
5975 fluct
= abs(cost
- prev_cost
);
5976 avg_fluct
= (avg_fluct
+ fluct
)/2;
5978 if (migration_debug
)
5979 printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): (%8Ld %8Ld)\n",
5981 (long)cost
/ 1000000,
5982 ((long)cost
/ 100000) % 10,
5983 (long)max_cost
/ 1000000,
5984 ((long)max_cost
/ 100000) % 10,
5985 domain_distance(cpu1
, cpu2
),
5989 * If we iterated at least 20% past the previous maximum,
5990 * and the cost has dropped by more than 20% already,
5991 * (taking fluctuations into account) then we assume to
5992 * have found the maximum and break out of the loop early:
5994 if (size_found
&& (size
*100 > size_found
*SIZE_THRESH
))
5995 if (cost
+avg_fluct
<= 0 ||
5996 max_cost
*100 > (cost
+avg_fluct
)*COST_THRESH
) {
5998 if (migration_debug
)
5999 printk("-> found max.\n");
6003 * Increase the cachesize in 10% steps:
6005 size
= size
* 10 / 9;
6008 if (migration_debug
)
6009 printk("[%d][%d] working set size found: %d, cost: %Ld\n",
6010 cpu1
, cpu2
, size_found
, max_cost
);
6015 * A task is considered 'cache cold' if at least 2 times
6016 * the worst-case cost of migration has passed.
6018 * (this limit is only listened to if the load-balancing
6019 * situation is 'nice' - if there is a large imbalance we
6020 * ignore it for the sake of CPU utilization and
6021 * processing fairness.)
6023 return 2 * max_cost
* migration_factor
/ MIGRATION_FACTOR_SCALE
;
6026 static void calibrate_migration_costs(const cpumask_t
*cpu_map
)
6028 int cpu1
= -1, cpu2
= -1, cpu
, orig_cpu
= raw_smp_processor_id();
6029 unsigned long j0
, j1
, distance
, max_distance
= 0;
6030 struct sched_domain
*sd
;
6035 * First pass - calculate the cacheflush times:
6037 for_each_cpu_mask(cpu1
, *cpu_map
) {
6038 for_each_cpu_mask(cpu2
, *cpu_map
) {
6041 distance
= domain_distance(cpu1
, cpu2
);
6042 max_distance
= max(max_distance
, distance
);
6044 * No result cached yet?
6046 if (migration_cost
[distance
] == -1LL)
6047 migration_cost
[distance
] =
6048 measure_migration_cost(cpu1
, cpu2
);
6052 * Second pass - update the sched domain hierarchy with
6053 * the new cache-hot-time estimations:
6055 for_each_cpu_mask(cpu
, *cpu_map
) {
6057 for_each_domain(cpu
, sd
) {
6058 sd
->cache_hot_time
= migration_cost
[distance
];
6065 if (migration_debug
)
6066 printk("migration: max_cache_size: %d, cpu: %d MHz:\n",
6074 if (system_state
== SYSTEM_BOOTING
) {
6075 if (num_online_cpus() > 1) {
6076 printk("migration_cost=");
6077 for (distance
= 0; distance
<= max_distance
; distance
++) {
6080 printk("%ld", (long)migration_cost
[distance
] / 1000);
6086 if (migration_debug
)
6087 printk("migration: %ld seconds\n", (j1
-j0
)/HZ
);
6090 * Move back to the original CPU. NUMA-Q gets confused
6091 * if we migrate to another quad during bootup.
6093 if (raw_smp_processor_id() != orig_cpu
) {
6094 cpumask_t mask
= cpumask_of_cpu(orig_cpu
),
6095 saved_mask
= current
->cpus_allowed
;
6097 set_cpus_allowed(current
, mask
);
6098 set_cpus_allowed(current
, saved_mask
);
6105 * find_next_best_node - find the next node to include in a sched_domain
6106 * @node: node whose sched_domain we're building
6107 * @used_nodes: nodes already in the sched_domain
6109 * Find the next node to include in a given scheduling domain. Simply
6110 * finds the closest node not already in the @used_nodes map.
6112 * Should use nodemask_t.
6114 static int find_next_best_node(int node
, unsigned long *used_nodes
)
6116 int i
, n
, val
, min_val
, best_node
= 0;
6120 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6121 /* Start at @node */
6122 n
= (node
+ i
) % MAX_NUMNODES
;
6124 if (!nr_cpus_node(n
))
6127 /* Skip already used nodes */
6128 if (test_bit(n
, used_nodes
))
6131 /* Simple min distance search */
6132 val
= node_distance(node
, n
);
6134 if (val
< min_val
) {
6140 set_bit(best_node
, used_nodes
);
6145 * sched_domain_node_span - get a cpumask for a node's sched_domain
6146 * @node: node whose cpumask we're constructing
6147 * @size: number of nodes to include in this span
6149 * Given a node, construct a good cpumask for its sched_domain to span. It
6150 * should be one that prevents unnecessary balancing, but also spreads tasks
6153 static cpumask_t
sched_domain_node_span(int node
)
6155 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
6156 cpumask_t span
, nodemask
;
6160 bitmap_zero(used_nodes
, MAX_NUMNODES
);
6162 nodemask
= node_to_cpumask(node
);
6163 cpus_or(span
, span
, nodemask
);
6164 set_bit(node
, used_nodes
);
6166 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6167 int next_node
= find_next_best_node(node
, used_nodes
);
6169 nodemask
= node_to_cpumask(next_node
);
6170 cpus_or(span
, span
, nodemask
);
6177 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6180 * SMT sched-domains:
6182 #ifdef CONFIG_SCHED_SMT
6183 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
6184 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
6186 static int cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
,
6187 struct sched_group
**sg
)
6190 *sg
= &per_cpu(sched_group_cpus
, cpu
);
6196 * multi-core sched-domains:
6198 #ifdef CONFIG_SCHED_MC
6199 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
6200 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
6203 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6204 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
6205 struct sched_group
**sg
)
6208 cpumask_t mask
= cpu_sibling_map
[cpu
];
6209 cpus_and(mask
, mask
, *cpu_map
);
6210 group
= first_cpu(mask
);
6212 *sg
= &per_cpu(sched_group_core
, group
);
6215 #elif defined(CONFIG_SCHED_MC)
6216 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
6217 struct sched_group
**sg
)
6220 *sg
= &per_cpu(sched_group_core
, cpu
);
6225 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
6226 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
6228 static int cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
,
6229 struct sched_group
**sg
)
6232 #ifdef CONFIG_SCHED_MC
6233 cpumask_t mask
= cpu_coregroup_map(cpu
);
6234 cpus_and(mask
, mask
, *cpu_map
);
6235 group
= first_cpu(mask
);
6236 #elif defined(CONFIG_SCHED_SMT)
6237 cpumask_t mask
= cpu_sibling_map
[cpu
];
6238 cpus_and(mask
, mask
, *cpu_map
);
6239 group
= first_cpu(mask
);
6244 *sg
= &per_cpu(sched_group_phys
, group
);
6250 * The init_sched_build_groups can't handle what we want to do with node
6251 * groups, so roll our own. Now each node has its own list of groups which
6252 * gets dynamically allocated.
6254 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
6255 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
6257 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
6258 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
6260 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
6261 struct sched_group
**sg
)
6263 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
6266 cpus_and(nodemask
, nodemask
, *cpu_map
);
6267 group
= first_cpu(nodemask
);
6270 *sg
= &per_cpu(sched_group_allnodes
, group
);
6274 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6276 struct sched_group
*sg
= group_head
;
6282 for_each_cpu_mask(j
, sg
->cpumask
) {
6283 struct sched_domain
*sd
;
6285 sd
= &per_cpu(phys_domains
, j
);
6286 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
6288 * Only add "power" once for each
6294 sg
->cpu_power
+= sd
->groups
->cpu_power
;
6297 if (sg
!= group_head
)
6303 /* Free memory allocated for various sched_group structures */
6304 static void free_sched_groups(const cpumask_t
*cpu_map
)
6308 for_each_cpu_mask(cpu
, *cpu_map
) {
6309 struct sched_group
**sched_group_nodes
6310 = sched_group_nodes_bycpu
[cpu
];
6312 if (!sched_group_nodes
)
6315 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6316 cpumask_t nodemask
= node_to_cpumask(i
);
6317 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6319 cpus_and(nodemask
, nodemask
, *cpu_map
);
6320 if (cpus_empty(nodemask
))
6330 if (oldsg
!= sched_group_nodes
[i
])
6333 kfree(sched_group_nodes
);
6334 sched_group_nodes_bycpu
[cpu
] = NULL
;
6338 static void free_sched_groups(const cpumask_t
*cpu_map
)
6344 * Initialize sched groups cpu_power.
6346 * cpu_power indicates the capacity of sched group, which is used while
6347 * distributing the load between different sched groups in a sched domain.
6348 * Typically cpu_power for all the groups in a sched domain will be same unless
6349 * there are asymmetries in the topology. If there are asymmetries, group
6350 * having more cpu_power will pickup more load compared to the group having
6353 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6354 * the maximum number of tasks a group can handle in the presence of other idle
6355 * or lightly loaded groups in the same sched domain.
6357 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6359 struct sched_domain
*child
;
6360 struct sched_group
*group
;
6362 WARN_ON(!sd
|| !sd
->groups
);
6364 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
6370 * For perf policy, if the groups in child domain share resources
6371 * (for example cores sharing some portions of the cache hierarchy
6372 * or SMT), then set this domain groups cpu_power such that each group
6373 * can handle only one task, when there are other idle groups in the
6374 * same sched domain.
6376 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
6378 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
6379 sd
->groups
->cpu_power
= SCHED_LOAD_SCALE
;
6383 sd
->groups
->cpu_power
= 0;
6386 * add cpu_power of each child group to this groups cpu_power
6388 group
= child
->groups
;
6390 sd
->groups
->cpu_power
+= group
->cpu_power
;
6391 group
= group
->next
;
6392 } while (group
!= child
->groups
);
6396 * Build sched domains for a given set of cpus and attach the sched domains
6397 * to the individual cpus
6399 static int build_sched_domains(const cpumask_t
*cpu_map
)
6402 struct sched_domain
*sd
;
6404 struct sched_group
**sched_group_nodes
= NULL
;
6405 int sd_allnodes
= 0;
6408 * Allocate the per-node list of sched groups
6410 sched_group_nodes
= kzalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
6412 if (!sched_group_nodes
) {
6413 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6416 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6420 * Set up domains for cpus specified by the cpu_map.
6422 for_each_cpu_mask(i
, *cpu_map
) {
6423 struct sched_domain
*sd
= NULL
, *p
;
6424 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6426 cpus_and(nodemask
, nodemask
, *cpu_map
);
6429 if (cpus_weight(*cpu_map
)
6430 > SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6431 sd
= &per_cpu(allnodes_domains
, i
);
6432 *sd
= SD_ALLNODES_INIT
;
6433 sd
->span
= *cpu_map
;
6434 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6440 sd
= &per_cpu(node_domains
, i
);
6442 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6446 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6450 sd
= &per_cpu(phys_domains
, i
);
6452 sd
->span
= nodemask
;
6456 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6458 #ifdef CONFIG_SCHED_MC
6460 sd
= &per_cpu(core_domains
, i
);
6462 sd
->span
= cpu_coregroup_map(i
);
6463 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6466 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6469 #ifdef CONFIG_SCHED_SMT
6471 sd
= &per_cpu(cpu_domains
, i
);
6472 *sd
= SD_SIBLING_INIT
;
6473 sd
->span
= cpu_sibling_map
[i
];
6474 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6477 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6481 #ifdef CONFIG_SCHED_SMT
6482 /* Set up CPU (sibling) groups */
6483 for_each_cpu_mask(i
, *cpu_map
) {
6484 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
6485 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6486 if (i
!= first_cpu(this_sibling_map
))
6489 init_sched_build_groups(this_sibling_map
, cpu_map
, &cpu_to_cpu_group
);
6493 #ifdef CONFIG_SCHED_MC
6494 /* Set up multi-core groups */
6495 for_each_cpu_mask(i
, *cpu_map
) {
6496 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6497 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6498 if (i
!= first_cpu(this_core_map
))
6500 init_sched_build_groups(this_core_map
, cpu_map
, &cpu_to_core_group
);
6505 /* Set up physical groups */
6506 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6507 cpumask_t nodemask
= node_to_cpumask(i
);
6509 cpus_and(nodemask
, nodemask
, *cpu_map
);
6510 if (cpus_empty(nodemask
))
6513 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6517 /* Set up node groups */
6519 init_sched_build_groups(*cpu_map
, cpu_map
, &cpu_to_allnodes_group
);
6521 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6522 /* Set up node groups */
6523 struct sched_group
*sg
, *prev
;
6524 cpumask_t nodemask
= node_to_cpumask(i
);
6525 cpumask_t domainspan
;
6526 cpumask_t covered
= CPU_MASK_NONE
;
6529 cpus_and(nodemask
, nodemask
, *cpu_map
);
6530 if (cpus_empty(nodemask
)) {
6531 sched_group_nodes
[i
] = NULL
;
6535 domainspan
= sched_domain_node_span(i
);
6536 cpus_and(domainspan
, domainspan
, *cpu_map
);
6538 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6540 printk(KERN_WARNING
"Can not alloc domain group for "
6544 sched_group_nodes
[i
] = sg
;
6545 for_each_cpu_mask(j
, nodemask
) {
6546 struct sched_domain
*sd
;
6547 sd
= &per_cpu(node_domains
, j
);
6551 sg
->cpumask
= nodemask
;
6553 cpus_or(covered
, covered
, nodemask
);
6556 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6557 cpumask_t tmp
, notcovered
;
6558 int n
= (i
+ j
) % MAX_NUMNODES
;
6560 cpus_complement(notcovered
, covered
);
6561 cpus_and(tmp
, notcovered
, *cpu_map
);
6562 cpus_and(tmp
, tmp
, domainspan
);
6563 if (cpus_empty(tmp
))
6566 nodemask
= node_to_cpumask(n
);
6567 cpus_and(tmp
, tmp
, nodemask
);
6568 if (cpus_empty(tmp
))
6571 sg
= kmalloc_node(sizeof(struct sched_group
),
6575 "Can not alloc domain group for node %d\n", j
);
6580 sg
->next
= prev
->next
;
6581 cpus_or(covered
, covered
, tmp
);
6588 /* Calculate CPU power for physical packages and nodes */
6589 #ifdef CONFIG_SCHED_SMT
6590 for_each_cpu_mask(i
, *cpu_map
) {
6591 sd
= &per_cpu(cpu_domains
, i
);
6592 init_sched_groups_power(i
, sd
);
6595 #ifdef CONFIG_SCHED_MC
6596 for_each_cpu_mask(i
, *cpu_map
) {
6597 sd
= &per_cpu(core_domains
, i
);
6598 init_sched_groups_power(i
, sd
);
6602 for_each_cpu_mask(i
, *cpu_map
) {
6603 sd
= &per_cpu(phys_domains
, i
);
6604 init_sched_groups_power(i
, sd
);
6608 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6609 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6612 struct sched_group
*sg
;
6614 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6615 init_numa_sched_groups_power(sg
);
6619 /* Attach the domains */
6620 for_each_cpu_mask(i
, *cpu_map
) {
6621 struct sched_domain
*sd
;
6622 #ifdef CONFIG_SCHED_SMT
6623 sd
= &per_cpu(cpu_domains
, i
);
6624 #elif defined(CONFIG_SCHED_MC)
6625 sd
= &per_cpu(core_domains
, i
);
6627 sd
= &per_cpu(phys_domains
, i
);
6629 cpu_attach_domain(sd
, i
);
6632 * Tune cache-hot values:
6634 calibrate_migration_costs(cpu_map
);
6640 free_sched_groups(cpu_map
);
6645 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6647 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6649 cpumask_t cpu_default_map
;
6653 * Setup mask for cpus without special case scheduling requirements.
6654 * For now this just excludes isolated cpus, but could be used to
6655 * exclude other special cases in the future.
6657 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
6659 err
= build_sched_domains(&cpu_default_map
);
6664 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6666 free_sched_groups(cpu_map
);
6670 * Detach sched domains from a group of cpus specified in cpu_map
6671 * These cpus will now be attached to the NULL domain
6673 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6677 for_each_cpu_mask(i
, *cpu_map
)
6678 cpu_attach_domain(NULL
, i
);
6679 synchronize_sched();
6680 arch_destroy_sched_domains(cpu_map
);
6684 * Partition sched domains as specified by the cpumasks below.
6685 * This attaches all cpus from the cpumasks to the NULL domain,
6686 * waits for a RCU quiescent period, recalculates sched
6687 * domain information and then attaches them back to the
6688 * correct sched domains
6689 * Call with hotplug lock held
6691 int partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
6693 cpumask_t change_map
;
6696 cpus_and(*partition1
, *partition1
, cpu_online_map
);
6697 cpus_and(*partition2
, *partition2
, cpu_online_map
);
6698 cpus_or(change_map
, *partition1
, *partition2
);
6700 /* Detach sched domains from all of the affected cpus */
6701 detach_destroy_domains(&change_map
);
6702 if (!cpus_empty(*partition1
))
6703 err
= build_sched_domains(partition1
);
6704 if (!err
&& !cpus_empty(*partition2
))
6705 err
= build_sched_domains(partition2
);
6710 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6711 int arch_reinit_sched_domains(void)
6716 detach_destroy_domains(&cpu_online_map
);
6717 err
= arch_init_sched_domains(&cpu_online_map
);
6718 unlock_cpu_hotplug();
6723 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6727 if (buf
[0] != '0' && buf
[0] != '1')
6731 sched_smt_power_savings
= (buf
[0] == '1');
6733 sched_mc_power_savings
= (buf
[0] == '1');
6735 ret
= arch_reinit_sched_domains();
6737 return ret
? ret
: count
;
6740 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6744 #ifdef CONFIG_SCHED_SMT
6746 err
= sysfs_create_file(&cls
->kset
.kobj
,
6747 &attr_sched_smt_power_savings
.attr
);
6749 #ifdef CONFIG_SCHED_MC
6750 if (!err
&& mc_capable())
6751 err
= sysfs_create_file(&cls
->kset
.kobj
,
6752 &attr_sched_mc_power_savings
.attr
);
6758 #ifdef CONFIG_SCHED_MC
6759 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6761 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6763 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6764 const char *buf
, size_t count
)
6766 return sched_power_savings_store(buf
, count
, 0);
6768 SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6769 sched_mc_power_savings_store
);
6772 #ifdef CONFIG_SCHED_SMT
6773 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6775 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6777 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6778 const char *buf
, size_t count
)
6780 return sched_power_savings_store(buf
, count
, 1);
6782 SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6783 sched_smt_power_savings_store
);
6787 * Force a reinitialization of the sched domains hierarchy. The domains
6788 * and groups cannot be updated in place without racing with the balancing
6789 * code, so we temporarily attach all running cpus to the NULL domain
6790 * which will prevent rebalancing while the sched domains are recalculated.
6792 static int update_sched_domains(struct notifier_block
*nfb
,
6793 unsigned long action
, void *hcpu
)
6796 case CPU_UP_PREPARE
:
6797 case CPU_DOWN_PREPARE
:
6798 detach_destroy_domains(&cpu_online_map
);
6801 case CPU_UP_CANCELED
:
6802 case CPU_DOWN_FAILED
:
6806 * Fall through and re-initialise the domains.
6813 /* The hotplug lock is already held by cpu_up/cpu_down */
6814 arch_init_sched_domains(&cpu_online_map
);
6819 void __init
sched_init_smp(void)
6821 cpumask_t non_isolated_cpus
;
6824 arch_init_sched_domains(&cpu_online_map
);
6825 cpus_andnot(non_isolated_cpus
, cpu_online_map
, cpu_isolated_map
);
6826 if (cpus_empty(non_isolated_cpus
))
6827 cpu_set(smp_processor_id(), non_isolated_cpus
);
6828 unlock_cpu_hotplug();
6829 /* XXX: Theoretical race here - CPU may be hotplugged now */
6830 hotcpu_notifier(update_sched_domains
, 0);
6832 /* Move init over to a non-isolated CPU */
6833 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
6837 void __init
sched_init_smp(void)
6840 #endif /* CONFIG_SMP */
6842 int in_sched_functions(unsigned long addr
)
6844 /* Linker adds these: start and end of __sched functions */
6845 extern char __sched_text_start
[], __sched_text_end
[];
6847 return in_lock_functions(addr
) ||
6848 (addr
>= (unsigned long)__sched_text_start
6849 && addr
< (unsigned long)__sched_text_end
);
6852 void __init
sched_init(void)
6856 for_each_possible_cpu(i
) {
6857 struct prio_array
*array
;
6861 spin_lock_init(&rq
->lock
);
6862 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
6864 rq
->active
= rq
->arrays
;
6865 rq
->expired
= rq
->arrays
+ 1;
6866 rq
->best_expired_prio
= MAX_PRIO
;
6870 for (j
= 1; j
< 3; j
++)
6871 rq
->cpu_load
[j
] = 0;
6872 rq
->active_balance
= 0;
6875 rq
->migration_thread
= NULL
;
6876 INIT_LIST_HEAD(&rq
->migration_queue
);
6878 atomic_set(&rq
->nr_iowait
, 0);
6880 for (j
= 0; j
< 2; j
++) {
6881 array
= rq
->arrays
+ j
;
6882 for (k
= 0; k
< MAX_PRIO
; k
++) {
6883 INIT_LIST_HEAD(array
->queue
+ k
);
6884 __clear_bit(k
, array
->bitmap
);
6886 // delimiter for bitsearch
6887 __set_bit(MAX_PRIO
, array
->bitmap
);
6891 set_load_weight(&init_task
);
6894 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
6897 #ifdef CONFIG_RT_MUTEXES
6898 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
6902 * The boot idle thread does lazy MMU switching as well:
6904 atomic_inc(&init_mm
.mm_count
);
6905 enter_lazy_tlb(&init_mm
, current
);
6908 * Make us the idle thread. Technically, schedule() should not be
6909 * called from this thread, however somewhere below it might be,
6910 * but because we are the idle thread, we just pick up running again
6911 * when this runqueue becomes "idle".
6913 init_idle(current
, smp_processor_id());
6916 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6917 void __might_sleep(char *file
, int line
)
6920 static unsigned long prev_jiffy
; /* ratelimiting */
6922 if ((in_atomic() || irqs_disabled()) &&
6923 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6924 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6926 prev_jiffy
= jiffies
;
6927 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6928 " context at %s:%d\n", file
, line
);
6929 printk("in_atomic():%d, irqs_disabled():%d\n",
6930 in_atomic(), irqs_disabled());
6931 debug_show_held_locks(current
);
6936 EXPORT_SYMBOL(__might_sleep
);
6939 #ifdef CONFIG_MAGIC_SYSRQ
6940 void normalize_rt_tasks(void)
6942 struct prio_array
*array
;
6943 struct task_struct
*p
;
6944 unsigned long flags
;
6947 read_lock_irq(&tasklist_lock
);
6948 for_each_process(p
) {
6952 spin_lock_irqsave(&p
->pi_lock
, flags
);
6953 rq
= __task_rq_lock(p
);
6957 deactivate_task(p
, task_rq(p
));
6958 __setscheduler(p
, SCHED_NORMAL
, 0);
6960 __activate_task(p
, task_rq(p
));
6961 resched_task(rq
->curr
);
6964 __task_rq_unlock(rq
);
6965 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6967 read_unlock_irq(&tasklist_lock
);
6970 #endif /* CONFIG_MAGIC_SYSRQ */
6974 * These functions are only useful for the IA64 MCA handling.
6976 * They can only be called when the whole system has been
6977 * stopped - every CPU needs to be quiescent, and no scheduling
6978 * activity can take place. Using them for anything else would
6979 * be a serious bug, and as a result, they aren't even visible
6980 * under any other configuration.
6984 * curr_task - return the current task for a given cpu.
6985 * @cpu: the processor in question.
6987 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6989 struct task_struct
*curr_task(int cpu
)
6991 return cpu_curr(cpu
);
6995 * set_curr_task - set the current task for a given cpu.
6996 * @cpu: the processor in question.
6997 * @p: the task pointer to set.
6999 * Description: This function must only be used when non-maskable interrupts
7000 * are serviced on a separate stack. It allows the architecture to switch the
7001 * notion of the current task on a cpu in a non-blocking manner. This function
7002 * must be called with all CPU's synchronized, and interrupts disabled, the
7003 * and caller must save the original value of the current task (see
7004 * curr_task() above) and restore that value before reenabling interrupts and
7005 * re-starting the system.
7007 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7009 void set_curr_task(int cpu
, struct task_struct
*p
)