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
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/reciprocal_div.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/bootmem.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
75 #include <asm/irq_regs.h>
77 #include "sched_cpupri.h"
80 * Convert user-nice values [ -20 ... 0 ... 19 ]
81 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
84 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
85 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
86 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
89 * 'User priority' is the nice value converted to something we
90 * can work with better when scaling various scheduler parameters,
91 * it's a [ 0 ... 39 ] range.
93 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
94 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
95 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
98 * Helpers for converting nanosecond timing to jiffy resolution
100 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
102 #define NICE_0_LOAD SCHED_LOAD_SCALE
103 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
106 * These are the 'tuning knobs' of the scheduler:
108 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
109 * Timeslices get refilled after they expire.
111 #define DEF_TIMESLICE (100 * HZ / 1000)
114 * single value that denotes runtime == period, ie unlimited time.
116 #define RUNTIME_INF ((u64)~0ULL)
120 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
121 * Since cpu_power is a 'constant', we can use a reciprocal divide.
123 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
125 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
129 * Each time a sched group cpu_power is changed,
130 * we must compute its reciprocal value
132 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
134 sg
->__cpu_power
+= val
;
135 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
139 static inline int rt_policy(int policy
)
141 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
146 static inline int task_has_rt_policy(struct task_struct
*p
)
148 return rt_policy(p
->policy
);
152 * This is the priority-queue data structure of the RT scheduling class:
154 struct rt_prio_array
{
155 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
156 struct list_head xqueue
[MAX_RT_PRIO
]; /* exclusive queue */
157 struct list_head squeue
[MAX_RT_PRIO
]; /* shared queue */
160 struct rt_bandwidth
{
161 /* nests inside the rq lock: */
162 spinlock_t rt_runtime_lock
;
165 struct hrtimer rt_period_timer
;
168 static struct rt_bandwidth def_rt_bandwidth
;
170 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
172 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
174 struct rt_bandwidth
*rt_b
=
175 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
181 now
= hrtimer_cb_get_time(timer
);
182 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
187 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
190 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
194 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
196 rt_b
->rt_period
= ns_to_ktime(period
);
197 rt_b
->rt_runtime
= runtime
;
199 spin_lock_init(&rt_b
->rt_runtime_lock
);
201 hrtimer_init(&rt_b
->rt_period_timer
,
202 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
203 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
204 rt_b
->rt_period_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_NO_SOFTIRQ
;
207 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
211 if (rt_b
->rt_runtime
== RUNTIME_INF
)
214 if (hrtimer_active(&rt_b
->rt_period_timer
))
217 spin_lock(&rt_b
->rt_runtime_lock
);
219 if (hrtimer_active(&rt_b
->rt_period_timer
))
222 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
223 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
224 hrtimer_start(&rt_b
->rt_period_timer
,
225 rt_b
->rt_period_timer
.expires
,
228 spin_unlock(&rt_b
->rt_runtime_lock
);
231 #ifdef CONFIG_RT_GROUP_SCHED
232 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
234 hrtimer_cancel(&rt_b
->rt_period_timer
);
239 * sched_domains_mutex serializes calls to arch_init_sched_domains,
240 * detach_destroy_domains and partition_sched_domains.
242 static DEFINE_MUTEX(sched_domains_mutex
);
244 #ifdef CONFIG_GROUP_SCHED
246 #include <linux/cgroup.h>
250 static LIST_HEAD(task_groups
);
252 /* task group related information */
254 #ifdef CONFIG_CGROUP_SCHED
255 struct cgroup_subsys_state css
;
258 #ifdef CONFIG_FAIR_GROUP_SCHED
259 /* schedulable entities of this group on each cpu */
260 struct sched_entity
**se
;
261 /* runqueue "owned" by this group on each cpu */
262 struct cfs_rq
**cfs_rq
;
263 unsigned long shares
;
266 #ifdef CONFIG_RT_GROUP_SCHED
267 struct sched_rt_entity
**rt_se
;
268 struct rt_rq
**rt_rq
;
270 struct rt_bandwidth rt_bandwidth
;
274 struct list_head list
;
276 struct task_group
*parent
;
277 struct list_head siblings
;
278 struct list_head children
;
281 #ifdef CONFIG_USER_SCHED
285 * Every UID task group (including init_task_group aka UID-0) will
286 * be a child to this group.
288 struct task_group root_task_group
;
290 #ifdef CONFIG_FAIR_GROUP_SCHED
291 /* Default task group's sched entity on each cpu */
292 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
293 /* Default task group's cfs_rq on each cpu */
294 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
295 #endif /* CONFIG_FAIR_GROUP_SCHED */
297 #ifdef CONFIG_RT_GROUP_SCHED
298 static DEFINE_PER_CPU(struct sched_rt_entity
, init_sched_rt_entity
);
299 static DEFINE_PER_CPU(struct rt_rq
, init_rt_rq
) ____cacheline_aligned_in_smp
;
300 #endif /* CONFIG_RT_GROUP_SCHED */
301 #else /* !CONFIG_FAIR_GROUP_SCHED */
302 #define root_task_group init_task_group
303 #endif /* CONFIG_FAIR_GROUP_SCHED */
305 /* task_group_lock serializes add/remove of task groups and also changes to
306 * a task group's cpu shares.
308 static DEFINE_SPINLOCK(task_group_lock
);
310 #ifdef CONFIG_FAIR_GROUP_SCHED
311 #ifdef CONFIG_USER_SCHED
312 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
313 #else /* !CONFIG_USER_SCHED */
314 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
315 #endif /* CONFIG_USER_SCHED */
318 * A weight of 0, 1 or ULONG_MAX can cause arithmetics problems.
319 * (The default weight is 1024 - so there's no practical
320 * limitation from this.)
323 #define MAX_SHARES (ULONG_MAX - 1)
325 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
328 /* Default task group.
329 * Every task in system belong to this group at bootup.
331 struct task_group init_task_group
;
333 /* return group to which a task belongs */
334 static inline struct task_group
*task_group(struct task_struct
*p
)
336 struct task_group
*tg
;
338 #ifdef CONFIG_USER_SCHED
340 #elif defined(CONFIG_CGROUP_SCHED)
341 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
342 struct task_group
, css
);
344 tg
= &init_task_group
;
349 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
350 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
352 #ifdef CONFIG_FAIR_GROUP_SCHED
353 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
354 p
->se
.parent
= task_group(p
)->se
[cpu
];
357 #ifdef CONFIG_RT_GROUP_SCHED
358 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
359 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
365 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
367 #endif /* CONFIG_GROUP_SCHED */
369 /* CFS-related fields in a runqueue */
371 struct load_weight load
;
372 unsigned long nr_running
;
377 struct rb_root tasks_timeline
;
378 struct rb_node
*rb_leftmost
;
380 struct list_head tasks
;
381 struct list_head
*balance_iterator
;
384 * 'curr' points to currently running entity on this cfs_rq.
385 * It is set to NULL otherwise (i.e when none are currently running).
387 struct sched_entity
*curr
, *next
;
389 unsigned long nr_spread_over
;
391 #ifdef CONFIG_FAIR_GROUP_SCHED
392 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
395 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
396 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
397 * (like users, containers etc.)
399 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
400 * list is used during load balance.
402 struct list_head leaf_cfs_rq_list
;
403 struct task_group
*tg
; /* group that "owns" this runqueue */
407 /* Real-Time classes' related field in a runqueue: */
409 struct rt_prio_array active
;
410 unsigned long rt_nr_running
;
411 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
412 int highest_prio
; /* highest queued rt task prio */
415 unsigned long rt_nr_migratory
;
421 /* Nests inside the rq lock: */
422 spinlock_t rt_runtime_lock
;
424 #ifdef CONFIG_RT_GROUP_SCHED
425 unsigned long rt_nr_boosted
;
428 struct list_head leaf_rt_rq_list
;
429 struct task_group
*tg
;
430 struct sched_rt_entity
*rt_se
;
437 * We add the notion of a root-domain which will be used to define per-domain
438 * variables. Each exclusive cpuset essentially defines an island domain by
439 * fully partitioning the member cpus from any other cpuset. Whenever a new
440 * exclusive cpuset is created, we also create and attach a new root-domain
450 * The "RT overload" flag: it gets set if a CPU has more than
451 * one runnable RT task.
456 struct cpupri cpupri
;
461 * By default the system creates a single root-domain with all cpus as
462 * members (mimicking the global state we have today).
464 static struct root_domain def_root_domain
;
469 * This is the main, per-CPU runqueue data structure.
471 * Locking rule: those places that want to lock multiple runqueues
472 * (such as the load balancing or the thread migration code), lock
473 * acquire operations must be ordered by ascending &runqueue.
480 * nr_running and cpu_load should be in the same cacheline because
481 * remote CPUs use both these fields when doing load calculation.
483 unsigned long nr_running
;
484 #define CPU_LOAD_IDX_MAX 5
485 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
486 unsigned char idle_at_tick
;
488 unsigned long last_tick_seen
;
489 unsigned char in_nohz_recently
;
491 /* capture load from *all* tasks on this cpu: */
492 struct load_weight load
;
493 unsigned long nr_load_updates
;
499 #ifdef CONFIG_FAIR_GROUP_SCHED
500 /* list of leaf cfs_rq on this cpu: */
501 struct list_head leaf_cfs_rq_list
;
503 #ifdef CONFIG_RT_GROUP_SCHED
504 struct list_head leaf_rt_rq_list
;
508 * This is part of a global counter where only the total sum
509 * over all CPUs matters. A task can increase this counter on
510 * one CPU and if it got migrated afterwards it may decrease
511 * it on another CPU. Always updated under the runqueue lock:
513 unsigned long nr_uninterruptible
;
515 struct task_struct
*curr
, *idle
;
516 unsigned long next_balance
;
517 struct mm_struct
*prev_mm
;
524 struct root_domain
*rd
;
525 struct sched_domain
*sd
;
527 /* For active balancing */
530 /* cpu of this runqueue: */
534 struct task_struct
*migration_thread
;
535 struct list_head migration_queue
;
538 #ifdef CONFIG_SCHED_HRTICK
539 unsigned long hrtick_flags
;
540 ktime_t hrtick_expire
;
541 struct hrtimer hrtick_timer
;
544 #ifdef CONFIG_SCHEDSTATS
546 struct sched_info rq_sched_info
;
548 /* sys_sched_yield() stats */
549 unsigned int yld_exp_empty
;
550 unsigned int yld_act_empty
;
551 unsigned int yld_both_empty
;
552 unsigned int yld_count
;
554 /* schedule() stats */
555 unsigned int sched_switch
;
556 unsigned int sched_count
;
557 unsigned int sched_goidle
;
559 /* try_to_wake_up() stats */
560 unsigned int ttwu_count
;
561 unsigned int ttwu_local
;
564 unsigned int bkl_count
;
566 struct lock_class_key rq_lock_key
;
569 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
571 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
)
573 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
);
576 static inline int cpu_of(struct rq
*rq
)
586 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
587 * See detach_destroy_domains: synchronize_sched for details.
589 * The domain tree of any CPU may only be accessed from within
590 * preempt-disabled sections.
592 #define for_each_domain(cpu, __sd) \
593 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
595 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
596 #define this_rq() (&__get_cpu_var(runqueues))
597 #define task_rq(p) cpu_rq(task_cpu(p))
598 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
600 static inline void update_rq_clock(struct rq
*rq
)
602 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
606 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
608 #ifdef CONFIG_SCHED_DEBUG
609 # define const_debug __read_mostly
611 # define const_debug static const
615 * Debugging: various feature bits
618 #define SCHED_FEAT(name, enabled) \
619 __SCHED_FEAT_##name ,
622 #include "sched_features.h"
627 #define SCHED_FEAT(name, enabled) \
628 (1UL << __SCHED_FEAT_##name) * enabled |
630 const_debug
unsigned int sysctl_sched_features
=
631 #include "sched_features.h"
636 #ifdef CONFIG_SCHED_DEBUG
637 #define SCHED_FEAT(name, enabled) \
640 static __read_mostly
char *sched_feat_names
[] = {
641 #include "sched_features.h"
647 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
649 filp
->private_data
= inode
->i_private
;
654 sched_feat_read(struct file
*filp
, char __user
*ubuf
,
655 size_t cnt
, loff_t
*ppos
)
662 for (i
= 0; sched_feat_names
[i
]; i
++) {
663 len
+= strlen(sched_feat_names
[i
]);
667 buf
= kmalloc(len
+ 2, GFP_KERNEL
);
671 for (i
= 0; sched_feat_names
[i
]; i
++) {
672 if (sysctl_sched_features
& (1UL << i
))
673 r
+= sprintf(buf
+ r
, "%s ", sched_feat_names
[i
]);
675 r
+= sprintf(buf
+ r
, "NO_%s ", sched_feat_names
[i
]);
678 r
+= sprintf(buf
+ r
, "\n");
679 WARN_ON(r
>= len
+ 2);
681 r
= simple_read_from_buffer(ubuf
, cnt
, ppos
, buf
, r
);
689 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
690 size_t cnt
, loff_t
*ppos
)
700 if (copy_from_user(&buf
, ubuf
, cnt
))
705 if (strncmp(buf
, "NO_", 3) == 0) {
710 for (i
= 0; sched_feat_names
[i
]; i
++) {
711 int len
= strlen(sched_feat_names
[i
]);
713 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
715 sysctl_sched_features
&= ~(1UL << i
);
717 sysctl_sched_features
|= (1UL << i
);
722 if (!sched_feat_names
[i
])
730 static struct file_operations sched_feat_fops
= {
731 .open
= sched_feat_open
,
732 .read
= sched_feat_read
,
733 .write
= sched_feat_write
,
736 static __init
int sched_init_debug(void)
738 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
743 late_initcall(sched_init_debug
);
747 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
750 * Number of tasks to iterate in a single balance run.
751 * Limited because this is done with IRQs disabled.
753 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
756 * period over which we measure -rt task cpu usage in us.
759 unsigned int sysctl_sched_rt_period
= 1000000;
761 static __read_mostly
int scheduler_running
;
764 * part of the period that we allow rt tasks to run in us.
767 int sysctl_sched_rt_runtime
= 950000;
769 static inline u64
global_rt_period(void)
771 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
774 static inline u64
global_rt_runtime(void)
776 if (sysctl_sched_rt_period
< 0)
779 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
782 unsigned long long time_sync_thresh
= 100000;
784 static DEFINE_PER_CPU(unsigned long long, time_offset
);
785 static DEFINE_PER_CPU(unsigned long long, prev_cpu_time
);
788 * Global lock which we take every now and then to synchronize
789 * the CPUs time. This method is not warp-safe, but it's good
790 * enough to synchronize slowly diverging time sources and thus
791 * it's good enough for tracing:
793 static DEFINE_SPINLOCK(time_sync_lock
);
794 static unsigned long long prev_global_time
;
796 static unsigned long long __sync_cpu_clock(unsigned long long time
, int cpu
)
799 * We want this inlined, to not get tracer function calls
800 * in this critical section:
802 spin_acquire(&time_sync_lock
.dep_map
, 0, 0, _THIS_IP_
);
803 __raw_spin_lock(&time_sync_lock
.raw_lock
);
805 if (time
< prev_global_time
) {
806 per_cpu(time_offset
, cpu
) += prev_global_time
- time
;
807 time
= prev_global_time
;
809 prev_global_time
= time
;
812 __raw_spin_unlock(&time_sync_lock
.raw_lock
);
813 spin_release(&time_sync_lock
.dep_map
, 1, _THIS_IP_
);
818 static unsigned long long __cpu_clock(int cpu
)
820 unsigned long long now
;
823 * Only call sched_clock() if the scheduler has already been
824 * initialized (some code might call cpu_clock() very early):
826 if (unlikely(!scheduler_running
))
829 now
= sched_clock_cpu(cpu
);
835 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
836 * clock constructed from sched_clock():
838 unsigned long long cpu_clock(int cpu
)
840 unsigned long long prev_cpu_time
, time
, delta_time
;
843 local_irq_save(flags
);
844 prev_cpu_time
= per_cpu(prev_cpu_time
, cpu
);
845 time
= __cpu_clock(cpu
) + per_cpu(time_offset
, cpu
);
846 delta_time
= time
-prev_cpu_time
;
848 if (unlikely(delta_time
> time_sync_thresh
)) {
849 time
= __sync_cpu_clock(time
, cpu
);
850 per_cpu(prev_cpu_time
, cpu
) = time
;
852 local_irq_restore(flags
);
856 EXPORT_SYMBOL_GPL(cpu_clock
);
858 #ifndef prepare_arch_switch
859 # define prepare_arch_switch(next) do { } while (0)
861 #ifndef finish_arch_switch
862 # define finish_arch_switch(prev) do { } while (0)
865 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
867 return rq
->curr
== p
;
870 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
871 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
873 return task_current(rq
, p
);
876 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
880 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
882 #ifdef CONFIG_DEBUG_SPINLOCK
883 /* this is a valid case when another task releases the spinlock */
884 rq
->lock
.owner
= current
;
887 * If we are tracking spinlock dependencies then we have to
888 * fix up the runqueue lock - which gets 'carried over' from
891 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
893 spin_unlock_irq(&rq
->lock
);
896 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
897 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
902 return task_current(rq
, p
);
906 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
910 * We can optimise this out completely for !SMP, because the
911 * SMP rebalancing from interrupt is the only thing that cares
916 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
917 spin_unlock_irq(&rq
->lock
);
919 spin_unlock(&rq
->lock
);
923 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
927 * After ->oncpu is cleared, the task can be moved to a different CPU.
928 * We must ensure this doesn't happen until the switch is completely
934 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
938 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
941 * __task_rq_lock - lock the runqueue a given task resides on.
942 * Must be called interrupts disabled.
944 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
948 struct rq
*rq
= task_rq(p
);
949 spin_lock(&rq
->lock
);
950 if (likely(rq
== task_rq(p
)))
952 spin_unlock(&rq
->lock
);
957 * task_rq_lock - lock the runqueue a given task resides on and disable
958 * interrupts. Note the ordering: we can safely lookup the task_rq without
959 * explicitly disabling preemption.
961 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
967 local_irq_save(*flags
);
969 spin_lock(&rq
->lock
);
970 if (likely(rq
== task_rq(p
)))
972 spin_unlock_irqrestore(&rq
->lock
, *flags
);
976 static void __task_rq_unlock(struct rq
*rq
)
979 spin_unlock(&rq
->lock
);
982 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
985 spin_unlock_irqrestore(&rq
->lock
, *flags
);
989 * this_rq_lock - lock this runqueue and disable interrupts.
991 static struct rq
*this_rq_lock(void)
998 spin_lock(&rq
->lock
);
1003 static void __resched_task(struct task_struct
*p
, int tif_bit
);
1005 static inline void resched_task(struct task_struct
*p
)
1007 __resched_task(p
, TIF_NEED_RESCHED
);
1010 #ifdef CONFIG_SCHED_HRTICK
1012 * Use HR-timers to deliver accurate preemption points.
1014 * Its all a bit involved since we cannot program an hrt while holding the
1015 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1018 * When we get rescheduled we reprogram the hrtick_timer outside of the
1021 static inline void resched_hrt(struct task_struct
*p
)
1023 __resched_task(p
, TIF_HRTICK_RESCHED
);
1026 static inline void resched_rq(struct rq
*rq
)
1028 unsigned long flags
;
1030 spin_lock_irqsave(&rq
->lock
, flags
);
1031 resched_task(rq
->curr
);
1032 spin_unlock_irqrestore(&rq
->lock
, flags
);
1036 HRTICK_SET
, /* re-programm hrtick_timer */
1037 HRTICK_RESET
, /* not a new slice */
1038 HRTICK_BLOCK
, /* stop hrtick operations */
1043 * - enabled by features
1044 * - hrtimer is actually high res
1046 static inline int hrtick_enabled(struct rq
*rq
)
1048 if (!sched_feat(HRTICK
))
1050 if (unlikely(test_bit(HRTICK_BLOCK
, &rq
->hrtick_flags
)))
1052 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1056 * Called to set the hrtick timer state.
1058 * called with rq->lock held and irqs disabled
1060 static void hrtick_start(struct rq
*rq
, u64 delay
, int reset
)
1062 assert_spin_locked(&rq
->lock
);
1065 * preempt at: now + delay
1068 ktime_add_ns(rq
->hrtick_timer
.base
->get_time(), delay
);
1070 * indicate we need to program the timer
1072 __set_bit(HRTICK_SET
, &rq
->hrtick_flags
);
1074 __set_bit(HRTICK_RESET
, &rq
->hrtick_flags
);
1077 * New slices are called from the schedule path and don't need a
1078 * forced reschedule.
1081 resched_hrt(rq
->curr
);
1084 static void hrtick_clear(struct rq
*rq
)
1086 if (hrtimer_active(&rq
->hrtick_timer
))
1087 hrtimer_cancel(&rq
->hrtick_timer
);
1091 * Update the timer from the possible pending state.
1093 static void hrtick_set(struct rq
*rq
)
1097 unsigned long flags
;
1099 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1101 spin_lock_irqsave(&rq
->lock
, flags
);
1102 set
= __test_and_clear_bit(HRTICK_SET
, &rq
->hrtick_flags
);
1103 reset
= __test_and_clear_bit(HRTICK_RESET
, &rq
->hrtick_flags
);
1104 time
= rq
->hrtick_expire
;
1105 clear_thread_flag(TIF_HRTICK_RESCHED
);
1106 spin_unlock_irqrestore(&rq
->lock
, flags
);
1109 hrtimer_start(&rq
->hrtick_timer
, time
, HRTIMER_MODE_ABS
);
1110 if (reset
&& !hrtimer_active(&rq
->hrtick_timer
))
1117 * High-resolution timer tick.
1118 * Runs from hardirq context with interrupts disabled.
1120 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1122 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1124 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1126 spin_lock(&rq
->lock
);
1127 update_rq_clock(rq
);
1128 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1129 spin_unlock(&rq
->lock
);
1131 return HRTIMER_NORESTART
;
1135 static void hotplug_hrtick_disable(int cpu
)
1137 struct rq
*rq
= cpu_rq(cpu
);
1138 unsigned long flags
;
1140 spin_lock_irqsave(&rq
->lock
, flags
);
1141 rq
->hrtick_flags
= 0;
1142 __set_bit(HRTICK_BLOCK
, &rq
->hrtick_flags
);
1143 spin_unlock_irqrestore(&rq
->lock
, flags
);
1148 static void hotplug_hrtick_enable(int cpu
)
1150 struct rq
*rq
= cpu_rq(cpu
);
1151 unsigned long flags
;
1153 spin_lock_irqsave(&rq
->lock
, flags
);
1154 __clear_bit(HRTICK_BLOCK
, &rq
->hrtick_flags
);
1155 spin_unlock_irqrestore(&rq
->lock
, flags
);
1159 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1161 int cpu
= (int)(long)hcpu
;
1164 case CPU_UP_CANCELED
:
1165 case CPU_UP_CANCELED_FROZEN
:
1166 case CPU_DOWN_PREPARE
:
1167 case CPU_DOWN_PREPARE_FROZEN
:
1169 case CPU_DEAD_FROZEN
:
1170 hotplug_hrtick_disable(cpu
);
1173 case CPU_UP_PREPARE
:
1174 case CPU_UP_PREPARE_FROZEN
:
1175 case CPU_DOWN_FAILED
:
1176 case CPU_DOWN_FAILED_FROZEN
:
1178 case CPU_ONLINE_FROZEN
:
1179 hotplug_hrtick_enable(cpu
);
1186 static void init_hrtick(void)
1188 hotcpu_notifier(hotplug_hrtick
, 0);
1190 #endif /* CONFIG_SMP */
1192 static void init_rq_hrtick(struct rq
*rq
)
1194 rq
->hrtick_flags
= 0;
1195 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1196 rq
->hrtick_timer
.function
= hrtick
;
1197 rq
->hrtick_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_NO_SOFTIRQ
;
1200 void hrtick_resched(void)
1203 unsigned long flags
;
1205 if (!test_thread_flag(TIF_HRTICK_RESCHED
))
1208 local_irq_save(flags
);
1209 rq
= cpu_rq(smp_processor_id());
1211 local_irq_restore(flags
);
1214 static inline void hrtick_clear(struct rq
*rq
)
1218 static inline void hrtick_set(struct rq
*rq
)
1222 static inline void init_rq_hrtick(struct rq
*rq
)
1226 void hrtick_resched(void)
1230 static inline void init_hrtick(void)
1236 * resched_task - mark a task 'to be rescheduled now'.
1238 * On UP this means the setting of the need_resched flag, on SMP it
1239 * might also involve a cross-CPU call to trigger the scheduler on
1244 #ifndef tsk_is_polling
1245 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1248 static void __resched_task(struct task_struct
*p
, int tif_bit
)
1252 assert_spin_locked(&task_rq(p
)->lock
);
1254 if (unlikely(test_tsk_thread_flag(p
, tif_bit
)))
1257 set_tsk_thread_flag(p
, tif_bit
);
1260 if (cpu
== smp_processor_id())
1263 /* NEED_RESCHED must be visible before we test polling */
1265 if (!tsk_is_polling(p
))
1266 smp_send_reschedule(cpu
);
1269 static void resched_cpu(int cpu
)
1271 struct rq
*rq
= cpu_rq(cpu
);
1272 unsigned long flags
;
1274 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1276 resched_task(cpu_curr(cpu
));
1277 spin_unlock_irqrestore(&rq
->lock
, flags
);
1282 * When add_timer_on() enqueues a timer into the timer wheel of an
1283 * idle CPU then this timer might expire before the next timer event
1284 * which is scheduled to wake up that CPU. In case of a completely
1285 * idle system the next event might even be infinite time into the
1286 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1287 * leaves the inner idle loop so the newly added timer is taken into
1288 * account when the CPU goes back to idle and evaluates the timer
1289 * wheel for the next timer event.
1291 void wake_up_idle_cpu(int cpu
)
1293 struct rq
*rq
= cpu_rq(cpu
);
1295 if (cpu
== smp_processor_id())
1299 * This is safe, as this function is called with the timer
1300 * wheel base lock of (cpu) held. When the CPU is on the way
1301 * to idle and has not yet set rq->curr to idle then it will
1302 * be serialized on the timer wheel base lock and take the new
1303 * timer into account automatically.
1305 if (rq
->curr
!= rq
->idle
)
1309 * We can set TIF_RESCHED on the idle task of the other CPU
1310 * lockless. The worst case is that the other CPU runs the
1311 * idle task through an additional NOOP schedule()
1313 set_tsk_thread_flag(rq
->idle
, TIF_NEED_RESCHED
);
1315 /* NEED_RESCHED must be visible before we test polling */
1317 if (!tsk_is_polling(rq
->idle
))
1318 smp_send_reschedule(cpu
);
1320 #endif /* CONFIG_NO_HZ */
1322 #else /* !CONFIG_SMP */
1323 static void __resched_task(struct task_struct
*p
, int tif_bit
)
1325 assert_spin_locked(&task_rq(p
)->lock
);
1326 set_tsk_thread_flag(p
, tif_bit
);
1328 #endif /* CONFIG_SMP */
1330 #if BITS_PER_LONG == 32
1331 # define WMULT_CONST (~0UL)
1333 # define WMULT_CONST (1UL << 32)
1336 #define WMULT_SHIFT 32
1339 * Shift right and round:
1341 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1343 static unsigned long
1344 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1345 struct load_weight
*lw
)
1349 if (!lw
->inv_weight
)
1350 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)/(lw
->weight
+1);
1352 tmp
= (u64
)delta_exec
* weight
;
1354 * Check whether we'd overflow the 64-bit multiplication:
1356 if (unlikely(tmp
> WMULT_CONST
))
1357 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1360 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1362 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1365 static inline unsigned long
1366 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
1368 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
1371 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1377 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1384 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1385 * of tasks with abnormal "nice" values across CPUs the contribution that
1386 * each task makes to its run queue's load is weighted according to its
1387 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1388 * scaled version of the new time slice allocation that they receive on time
1392 #define WEIGHT_IDLEPRIO 2
1393 #define WMULT_IDLEPRIO (1 << 31)
1396 * Nice levels are multiplicative, with a gentle 10% change for every
1397 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1398 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1399 * that remained on nice 0.
1401 * The "10% effect" is relative and cumulative: from _any_ nice level,
1402 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1403 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1404 * If a task goes up by ~10% and another task goes down by ~10% then
1405 * the relative distance between them is ~25%.)
1407 static const int prio_to_weight
[40] = {
1408 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1409 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1410 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1411 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1412 /* 0 */ 1024, 820, 655, 526, 423,
1413 /* 5 */ 335, 272, 215, 172, 137,
1414 /* 10 */ 110, 87, 70, 56, 45,
1415 /* 15 */ 36, 29, 23, 18, 15,
1419 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1421 * In cases where the weight does not change often, we can use the
1422 * precalculated inverse to speed up arithmetics by turning divisions
1423 * into multiplications:
1425 static const u32 prio_to_wmult
[40] = {
1426 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1427 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1428 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1429 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1430 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1431 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1432 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1433 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1436 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1439 * runqueue iterator, to support SMP load-balancing between different
1440 * scheduling classes, without having to expose their internal data
1441 * structures to the load-balancing proper:
1443 struct rq_iterator
{
1445 struct task_struct
*(*start
)(void *);
1446 struct task_struct
*(*next
)(void *);
1450 static unsigned long
1451 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1452 unsigned long max_load_move
, struct sched_domain
*sd
,
1453 enum cpu_idle_type idle
, int *all_pinned
,
1454 int *this_best_prio
, struct rq_iterator
*iterator
);
1457 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1458 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1459 struct rq_iterator
*iterator
);
1462 #ifdef CONFIG_CGROUP_CPUACCT
1463 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1465 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1468 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1470 update_load_add(&rq
->load
, load
);
1473 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1475 update_load_sub(&rq
->load
, load
);
1479 static unsigned long source_load(int cpu
, int type
);
1480 static unsigned long target_load(int cpu
, int type
);
1481 static unsigned long cpu_avg_load_per_task(int cpu
);
1482 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1483 #else /* CONFIG_SMP */
1485 #ifdef CONFIG_FAIR_GROUP_SCHED
1486 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1491 #endif /* CONFIG_SMP */
1493 #include "sched_stats.h"
1494 #include "sched_idletask.c"
1495 #include "sched_fair.c"
1496 #include "sched_rt.c"
1497 #ifdef CONFIG_SCHED_DEBUG
1498 # include "sched_debug.c"
1501 #define sched_class_highest (&rt_sched_class)
1502 #define for_each_class(class) \
1503 for (class = sched_class_highest; class; class = class->next)
1505 static inline void inc_load(struct rq
*rq
, const struct task_struct
*p
)
1507 update_load_add(&rq
->load
, p
->se
.load
.weight
);
1510 static inline void dec_load(struct rq
*rq
, const struct task_struct
*p
)
1512 update_load_sub(&rq
->load
, p
->se
.load
.weight
);
1515 static void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
1521 static void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
1527 static void set_load_weight(struct task_struct
*p
)
1529 if (task_has_rt_policy(p
)) {
1530 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1531 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1536 * SCHED_IDLE tasks get minimal weight:
1538 if (p
->policy
== SCHED_IDLE
) {
1539 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1540 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1544 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1545 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1548 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1550 sched_info_queued(p
);
1551 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1555 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1557 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1562 * __normal_prio - return the priority that is based on the static prio
1564 static inline int __normal_prio(struct task_struct
*p
)
1566 return p
->static_prio
;
1570 * Calculate the expected normal priority: i.e. priority
1571 * without taking RT-inheritance into account. Might be
1572 * boosted by interactivity modifiers. Changes upon fork,
1573 * setprio syscalls, and whenever the interactivity
1574 * estimator recalculates.
1576 static inline int normal_prio(struct task_struct
*p
)
1580 if (task_has_rt_policy(p
))
1581 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1583 prio
= __normal_prio(p
);
1588 * Calculate the current priority, i.e. the priority
1589 * taken into account by the scheduler. This value might
1590 * be boosted by RT tasks, or might be boosted by
1591 * interactivity modifiers. Will be RT if the task got
1592 * RT-boosted. If not then it returns p->normal_prio.
1594 static int effective_prio(struct task_struct
*p
)
1596 p
->normal_prio
= normal_prio(p
);
1598 * If we are RT tasks or we were boosted to RT priority,
1599 * keep the priority unchanged. Otherwise, update priority
1600 * to the normal priority:
1602 if (!rt_prio(p
->prio
))
1603 return p
->normal_prio
;
1608 * activate_task - move a task to the runqueue.
1610 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1612 if (task_contributes_to_load(p
))
1613 rq
->nr_uninterruptible
--;
1615 enqueue_task(rq
, p
, wakeup
);
1616 inc_nr_running(p
, rq
);
1620 * deactivate_task - remove a task from the runqueue.
1622 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1624 if (task_contributes_to_load(p
))
1625 rq
->nr_uninterruptible
++;
1627 dequeue_task(rq
, p
, sleep
);
1628 dec_nr_running(p
, rq
);
1632 * task_curr - is this task currently executing on a CPU?
1633 * @p: the task in question.
1635 inline int task_curr(const struct task_struct
*p
)
1637 return cpu_curr(task_cpu(p
)) == p
;
1640 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1642 set_task_rq(p
, cpu
);
1645 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1646 * successfuly executed on another CPU. We must ensure that updates of
1647 * per-task data have been completed by this moment.
1650 task_thread_info(p
)->cpu
= cpu
;
1654 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1655 const struct sched_class
*prev_class
,
1656 int oldprio
, int running
)
1658 if (prev_class
!= p
->sched_class
) {
1659 if (prev_class
->switched_from
)
1660 prev_class
->switched_from(rq
, p
, running
);
1661 p
->sched_class
->switched_to(rq
, p
, running
);
1663 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1668 /* Used instead of source_load when we know the type == 0 */
1669 static unsigned long weighted_cpuload(const int cpu
)
1671 return cpu_rq(cpu
)->load
.weight
;
1675 * Is this task likely cache-hot:
1678 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1683 * Buddy candidates are cache hot:
1685 if (sched_feat(CACHE_HOT_BUDDY
) && (&p
->se
== cfs_rq_of(&p
->se
)->next
))
1688 if (p
->sched_class
!= &fair_sched_class
)
1691 if (sysctl_sched_migration_cost
== -1)
1693 if (sysctl_sched_migration_cost
== 0)
1696 delta
= now
- p
->se
.exec_start
;
1698 return delta
< (s64
)sysctl_sched_migration_cost
;
1702 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1704 int old_cpu
= task_cpu(p
);
1705 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1706 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1707 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1710 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1712 #ifdef CONFIG_SCHEDSTATS
1713 if (p
->se
.wait_start
)
1714 p
->se
.wait_start
-= clock_offset
;
1715 if (p
->se
.sleep_start
)
1716 p
->se
.sleep_start
-= clock_offset
;
1717 if (p
->se
.block_start
)
1718 p
->se
.block_start
-= clock_offset
;
1719 if (old_cpu
!= new_cpu
) {
1720 schedstat_inc(p
, se
.nr_migrations
);
1721 if (task_hot(p
, old_rq
->clock
, NULL
))
1722 schedstat_inc(p
, se
.nr_forced2_migrations
);
1725 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1726 new_cfsrq
->min_vruntime
;
1728 __set_task_cpu(p
, new_cpu
);
1731 struct migration_req
{
1732 struct list_head list
;
1734 struct task_struct
*task
;
1737 struct completion done
;
1741 * The task's runqueue lock must be held.
1742 * Returns true if you have to wait for migration thread.
1745 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1747 struct rq
*rq
= task_rq(p
);
1750 * If the task is not on a runqueue (and not running), then
1751 * it is sufficient to simply update the task's cpu field.
1753 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1754 set_task_cpu(p
, dest_cpu
);
1758 init_completion(&req
->done
);
1760 req
->dest_cpu
= dest_cpu
;
1761 list_add(&req
->list
, &rq
->migration_queue
);
1767 * wait_task_inactive - wait for a thread to unschedule.
1769 * The caller must ensure that the task *will* unschedule sometime soon,
1770 * else this function might spin for a *long* time. This function can't
1771 * be called with interrupts off, or it may introduce deadlock with
1772 * smp_call_function() if an IPI is sent by the same process we are
1773 * waiting to become inactive.
1775 void wait_task_inactive(struct task_struct
*p
)
1777 unsigned long flags
;
1783 * We do the initial early heuristics without holding
1784 * any task-queue locks at all. We'll only try to get
1785 * the runqueue lock when things look like they will
1791 * If the task is actively running on another CPU
1792 * still, just relax and busy-wait without holding
1795 * NOTE! Since we don't hold any locks, it's not
1796 * even sure that "rq" stays as the right runqueue!
1797 * But we don't care, since "task_running()" will
1798 * return false if the runqueue has changed and p
1799 * is actually now running somewhere else!
1801 while (task_running(rq
, p
))
1805 * Ok, time to look more closely! We need the rq
1806 * lock now, to be *sure*. If we're wrong, we'll
1807 * just go back and repeat.
1809 rq
= task_rq_lock(p
, &flags
);
1810 running
= task_running(rq
, p
);
1811 on_rq
= p
->se
.on_rq
;
1812 task_rq_unlock(rq
, &flags
);
1815 * Was it really running after all now that we
1816 * checked with the proper locks actually held?
1818 * Oops. Go back and try again..
1820 if (unlikely(running
)) {
1826 * It's not enough that it's not actively running,
1827 * it must be off the runqueue _entirely_, and not
1830 * So if it wa still runnable (but just not actively
1831 * running right now), it's preempted, and we should
1832 * yield - it could be a while.
1834 if (unlikely(on_rq
)) {
1835 schedule_timeout_uninterruptible(1);
1840 * Ahh, all good. It wasn't running, and it wasn't
1841 * runnable, which means that it will never become
1842 * running in the future either. We're all done!
1849 * kick_process - kick a running thread to enter/exit the kernel
1850 * @p: the to-be-kicked thread
1852 * Cause a process which is running on another CPU to enter
1853 * kernel-mode, without any delay. (to get signals handled.)
1855 * NOTE: this function doesnt have to take the runqueue lock,
1856 * because all it wants to ensure is that the remote task enters
1857 * the kernel. If the IPI races and the task has been migrated
1858 * to another CPU then no harm is done and the purpose has been
1861 void kick_process(struct task_struct
*p
)
1867 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1868 smp_send_reschedule(cpu
);
1873 * Return a low guess at the load of a migration-source cpu weighted
1874 * according to the scheduling class and "nice" value.
1876 * We want to under-estimate the load of migration sources, to
1877 * balance conservatively.
1879 static unsigned long source_load(int cpu
, int type
)
1881 struct rq
*rq
= cpu_rq(cpu
);
1882 unsigned long total
= weighted_cpuload(cpu
);
1887 return min(rq
->cpu_load
[type
-1], total
);
1891 * Return a high guess at the load of a migration-target cpu weighted
1892 * according to the scheduling class and "nice" value.
1894 static unsigned long target_load(int cpu
, int type
)
1896 struct rq
*rq
= cpu_rq(cpu
);
1897 unsigned long total
= weighted_cpuload(cpu
);
1902 return max(rq
->cpu_load
[type
-1], total
);
1906 * Return the average load per task on the cpu's run queue
1908 static unsigned long cpu_avg_load_per_task(int cpu
)
1910 struct rq
*rq
= cpu_rq(cpu
);
1911 unsigned long total
= weighted_cpuload(cpu
);
1912 unsigned long n
= rq
->nr_running
;
1914 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1918 * find_idlest_group finds and returns the least busy CPU group within the
1921 static struct sched_group
*
1922 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1924 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1925 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1926 int load_idx
= sd
->forkexec_idx
;
1927 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1930 unsigned long load
, avg_load
;
1934 /* Skip over this group if it has no CPUs allowed */
1935 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1938 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1940 /* Tally up the load of all CPUs in the group */
1943 for_each_cpu_mask(i
, group
->cpumask
) {
1944 /* Bias balancing toward cpus of our domain */
1946 load
= source_load(i
, load_idx
);
1948 load
= target_load(i
, load_idx
);
1953 /* Adjust by relative CPU power of the group */
1954 avg_load
= sg_div_cpu_power(group
,
1955 avg_load
* SCHED_LOAD_SCALE
);
1958 this_load
= avg_load
;
1960 } else if (avg_load
< min_load
) {
1961 min_load
= avg_load
;
1964 } while (group
= group
->next
, group
!= sd
->groups
);
1966 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1972 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1975 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
,
1978 unsigned long load
, min_load
= ULONG_MAX
;
1982 /* Traverse only the allowed CPUs */
1983 cpus_and(*tmp
, group
->cpumask
, p
->cpus_allowed
);
1985 for_each_cpu_mask(i
, *tmp
) {
1986 load
= weighted_cpuload(i
);
1988 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1998 * sched_balance_self: balance the current task (running on cpu) in domains
1999 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2002 * Balance, ie. select the least loaded group.
2004 * Returns the target CPU number, or the same CPU if no balancing is needed.
2006 * preempt must be disabled.
2008 static int sched_balance_self(int cpu
, int flag
)
2010 struct task_struct
*t
= current
;
2011 struct sched_domain
*tmp
, *sd
= NULL
;
2013 for_each_domain(cpu
, tmp
) {
2015 * If power savings logic is enabled for a domain, stop there.
2017 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2019 if (tmp
->flags
& flag
)
2024 cpumask_t span
, tmpmask
;
2025 struct sched_group
*group
;
2026 int new_cpu
, weight
;
2028 if (!(sd
->flags
& flag
)) {
2034 group
= find_idlest_group(sd
, t
, cpu
);
2040 new_cpu
= find_idlest_cpu(group
, t
, cpu
, &tmpmask
);
2041 if (new_cpu
== -1 || new_cpu
== cpu
) {
2042 /* Now try balancing at a lower domain level of cpu */
2047 /* Now try balancing at a lower domain level of new_cpu */
2050 weight
= cpus_weight(span
);
2051 for_each_domain(cpu
, tmp
) {
2052 if (weight
<= cpus_weight(tmp
->span
))
2054 if (tmp
->flags
& flag
)
2057 /* while loop will break here if sd == NULL */
2063 #endif /* CONFIG_SMP */
2066 * try_to_wake_up - wake up a thread
2067 * @p: the to-be-woken-up thread
2068 * @state: the mask of task states that can be woken
2069 * @sync: do a synchronous wakeup?
2071 * Put it on the run-queue if it's not already there. The "current"
2072 * thread is always on the run-queue (except when the actual
2073 * re-schedule is in progress), and as such you're allowed to do
2074 * the simpler "current->state = TASK_RUNNING" to mark yourself
2075 * runnable without the overhead of this.
2077 * returns failure only if the task is already active.
2079 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
2081 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2082 unsigned long flags
;
2086 if (!sched_feat(SYNC_WAKEUPS
))
2090 rq
= task_rq_lock(p
, &flags
);
2091 old_state
= p
->state
;
2092 if (!(old_state
& state
))
2100 this_cpu
= smp_processor_id();
2103 if (unlikely(task_running(rq
, p
)))
2106 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2107 if (cpu
!= orig_cpu
) {
2108 set_task_cpu(p
, cpu
);
2109 task_rq_unlock(rq
, &flags
);
2110 /* might preempt at this point */
2111 rq
= task_rq_lock(p
, &flags
);
2112 old_state
= p
->state
;
2113 if (!(old_state
& state
))
2118 this_cpu
= smp_processor_id();
2122 #ifdef CONFIG_SCHEDSTATS
2123 schedstat_inc(rq
, ttwu_count
);
2124 if (cpu
== this_cpu
)
2125 schedstat_inc(rq
, ttwu_local
);
2127 struct sched_domain
*sd
;
2128 for_each_domain(this_cpu
, sd
) {
2129 if (cpu_isset(cpu
, sd
->span
)) {
2130 schedstat_inc(sd
, ttwu_wake_remote
);
2135 #endif /* CONFIG_SCHEDSTATS */
2138 #endif /* CONFIG_SMP */
2139 schedstat_inc(p
, se
.nr_wakeups
);
2141 schedstat_inc(p
, se
.nr_wakeups_sync
);
2142 if (orig_cpu
!= cpu
)
2143 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2144 if (cpu
== this_cpu
)
2145 schedstat_inc(p
, se
.nr_wakeups_local
);
2147 schedstat_inc(p
, se
.nr_wakeups_remote
);
2148 update_rq_clock(rq
);
2149 activate_task(rq
, p
, 1);
2153 check_preempt_curr(rq
, p
);
2155 p
->state
= TASK_RUNNING
;
2157 if (p
->sched_class
->task_wake_up
)
2158 p
->sched_class
->task_wake_up(rq
, p
);
2161 task_rq_unlock(rq
, &flags
);
2166 int wake_up_process(struct task_struct
*p
)
2168 return try_to_wake_up(p
, TASK_ALL
, 0);
2170 EXPORT_SYMBOL(wake_up_process
);
2172 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2174 return try_to_wake_up(p
, state
, 0);
2178 * Perform scheduler related setup for a newly forked process p.
2179 * p is forked by current.
2181 * __sched_fork() is basic setup used by init_idle() too:
2183 static void __sched_fork(struct task_struct
*p
)
2185 p
->se
.exec_start
= 0;
2186 p
->se
.sum_exec_runtime
= 0;
2187 p
->se
.prev_sum_exec_runtime
= 0;
2188 p
->se
.last_wakeup
= 0;
2189 p
->se
.avg_overlap
= 0;
2191 #ifdef CONFIG_SCHEDSTATS
2192 p
->se
.wait_start
= 0;
2193 p
->se
.sum_sleep_runtime
= 0;
2194 p
->se
.sleep_start
= 0;
2195 p
->se
.block_start
= 0;
2196 p
->se
.sleep_max
= 0;
2197 p
->se
.block_max
= 0;
2199 p
->se
.slice_max
= 0;
2203 INIT_LIST_HEAD(&p
->rt
.run_list
);
2205 INIT_LIST_HEAD(&p
->se
.group_node
);
2207 #ifdef CONFIG_PREEMPT_NOTIFIERS
2208 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2212 * We mark the process as running here, but have not actually
2213 * inserted it onto the runqueue yet. This guarantees that
2214 * nobody will actually run it, and a signal or other external
2215 * event cannot wake it up and insert it on the runqueue either.
2217 p
->state
= TASK_RUNNING
;
2221 * fork()/clone()-time setup:
2223 void sched_fork(struct task_struct
*p
, int clone_flags
)
2225 int cpu
= get_cpu();
2230 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2232 set_task_cpu(p
, cpu
);
2235 * Make sure we do not leak PI boosting priority to the child:
2237 p
->prio
= current
->normal_prio
;
2238 if (!rt_prio(p
->prio
))
2239 p
->sched_class
= &fair_sched_class
;
2241 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2242 if (likely(sched_info_on()))
2243 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2245 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2248 #ifdef CONFIG_PREEMPT
2249 /* Want to start with kernel preemption disabled. */
2250 task_thread_info(p
)->preempt_count
= 1;
2256 * wake_up_new_task - wake up a newly created task for the first time.
2258 * This function will do some initial scheduler statistics housekeeping
2259 * that must be done for every newly created context, then puts the task
2260 * on the runqueue and wakes it.
2262 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2264 unsigned long flags
;
2267 rq
= task_rq_lock(p
, &flags
);
2268 BUG_ON(p
->state
!= TASK_RUNNING
);
2269 update_rq_clock(rq
);
2271 p
->prio
= effective_prio(p
);
2273 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2274 activate_task(rq
, p
, 0);
2277 * Let the scheduling class do new task startup
2278 * management (if any):
2280 p
->sched_class
->task_new(rq
, p
);
2281 inc_nr_running(p
, rq
);
2283 check_preempt_curr(rq
, p
);
2285 if (p
->sched_class
->task_wake_up
)
2286 p
->sched_class
->task_wake_up(rq
, p
);
2288 task_rq_unlock(rq
, &flags
);
2291 #ifdef CONFIG_PREEMPT_NOTIFIERS
2294 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2295 * @notifier: notifier struct to register
2297 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2299 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2301 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2304 * preempt_notifier_unregister - no longer interested in preemption notifications
2305 * @notifier: notifier struct to unregister
2307 * This is safe to call from within a preemption notifier.
2309 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2311 hlist_del(¬ifier
->link
);
2313 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2315 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2317 struct preempt_notifier
*notifier
;
2318 struct hlist_node
*node
;
2320 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2321 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2325 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2326 struct task_struct
*next
)
2328 struct preempt_notifier
*notifier
;
2329 struct hlist_node
*node
;
2331 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2332 notifier
->ops
->sched_out(notifier
, next
);
2335 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2337 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2342 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2343 struct task_struct
*next
)
2347 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2350 * prepare_task_switch - prepare to switch tasks
2351 * @rq: the runqueue preparing to switch
2352 * @prev: the current task that is being switched out
2353 * @next: the task we are going to switch to.
2355 * This is called with the rq lock held and interrupts off. It must
2356 * be paired with a subsequent finish_task_switch after the context
2359 * prepare_task_switch sets up locking and calls architecture specific
2363 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2364 struct task_struct
*next
)
2366 fire_sched_out_preempt_notifiers(prev
, next
);
2367 prepare_lock_switch(rq
, next
);
2368 prepare_arch_switch(next
);
2372 * finish_task_switch - clean up after a task-switch
2373 * @rq: runqueue associated with task-switch
2374 * @prev: the thread we just switched away from.
2376 * finish_task_switch must be called after the context switch, paired
2377 * with a prepare_task_switch call before the context switch.
2378 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2379 * and do any other architecture-specific cleanup actions.
2381 * Note that we may have delayed dropping an mm in context_switch(). If
2382 * so, we finish that here outside of the runqueue lock. (Doing it
2383 * with the lock held can cause deadlocks; see schedule() for
2386 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2387 __releases(rq
->lock
)
2389 struct mm_struct
*mm
= rq
->prev_mm
;
2395 * A task struct has one reference for the use as "current".
2396 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2397 * schedule one last time. The schedule call will never return, and
2398 * the scheduled task must drop that reference.
2399 * The test for TASK_DEAD must occur while the runqueue locks are
2400 * still held, otherwise prev could be scheduled on another cpu, die
2401 * there before we look at prev->state, and then the reference would
2403 * Manfred Spraul <manfred@colorfullife.com>
2405 prev_state
= prev
->state
;
2406 finish_arch_switch(prev
);
2407 finish_lock_switch(rq
, prev
);
2409 if (current
->sched_class
->post_schedule
)
2410 current
->sched_class
->post_schedule(rq
);
2413 fire_sched_in_preempt_notifiers(current
);
2416 if (unlikely(prev_state
== TASK_DEAD
)) {
2418 * Remove function-return probe instances associated with this
2419 * task and put them back on the free list.
2421 kprobe_flush_task(prev
);
2422 put_task_struct(prev
);
2427 * schedule_tail - first thing a freshly forked thread must call.
2428 * @prev: the thread we just switched away from.
2430 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2431 __releases(rq
->lock
)
2433 struct rq
*rq
= this_rq();
2435 finish_task_switch(rq
, prev
);
2436 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2437 /* In this case, finish_task_switch does not reenable preemption */
2440 if (current
->set_child_tid
)
2441 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2445 * context_switch - switch to the new MM and the new
2446 * thread's register state.
2449 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2450 struct task_struct
*next
)
2452 struct mm_struct
*mm
, *oldmm
;
2454 prepare_task_switch(rq
, prev
, next
);
2456 oldmm
= prev
->active_mm
;
2458 * For paravirt, this is coupled with an exit in switch_to to
2459 * combine the page table reload and the switch backend into
2462 arch_enter_lazy_cpu_mode();
2464 if (unlikely(!mm
)) {
2465 next
->active_mm
= oldmm
;
2466 atomic_inc(&oldmm
->mm_count
);
2467 enter_lazy_tlb(oldmm
, next
);
2469 switch_mm(oldmm
, mm
, next
);
2471 if (unlikely(!prev
->mm
)) {
2472 prev
->active_mm
= NULL
;
2473 rq
->prev_mm
= oldmm
;
2476 * Since the runqueue lock will be released by the next
2477 * task (which is an invalid locking op but in the case
2478 * of the scheduler it's an obvious special-case), so we
2479 * do an early lockdep release here:
2481 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2482 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2485 /* Here we just switch the register state and the stack. */
2486 switch_to(prev
, next
, prev
);
2490 * this_rq must be evaluated again because prev may have moved
2491 * CPUs since it called schedule(), thus the 'rq' on its stack
2492 * frame will be invalid.
2494 finish_task_switch(this_rq(), prev
);
2498 * nr_running, nr_uninterruptible and nr_context_switches:
2500 * externally visible scheduler statistics: current number of runnable
2501 * threads, current number of uninterruptible-sleeping threads, total
2502 * number of context switches performed since bootup.
2504 unsigned long nr_running(void)
2506 unsigned long i
, sum
= 0;
2508 for_each_online_cpu(i
)
2509 sum
+= cpu_rq(i
)->nr_running
;
2514 unsigned long nr_uninterruptible(void)
2516 unsigned long i
, sum
= 0;
2518 for_each_possible_cpu(i
)
2519 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2522 * Since we read the counters lockless, it might be slightly
2523 * inaccurate. Do not allow it to go below zero though:
2525 if (unlikely((long)sum
< 0))
2531 unsigned long long nr_context_switches(void)
2534 unsigned long long sum
= 0;
2536 for_each_possible_cpu(i
)
2537 sum
+= cpu_rq(i
)->nr_switches
;
2542 unsigned long nr_iowait(void)
2544 unsigned long i
, sum
= 0;
2546 for_each_possible_cpu(i
)
2547 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2552 unsigned long nr_active(void)
2554 unsigned long i
, running
= 0, uninterruptible
= 0;
2556 for_each_online_cpu(i
) {
2557 running
+= cpu_rq(i
)->nr_running
;
2558 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2561 if (unlikely((long)uninterruptible
< 0))
2562 uninterruptible
= 0;
2564 return running
+ uninterruptible
;
2568 * Update rq->cpu_load[] statistics. This function is usually called every
2569 * scheduler tick (TICK_NSEC).
2571 static void update_cpu_load(struct rq
*this_rq
)
2573 unsigned long this_load
= this_rq
->load
.weight
;
2576 this_rq
->nr_load_updates
++;
2578 /* Update our load: */
2579 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2580 unsigned long old_load
, new_load
;
2582 /* scale is effectively 1 << i now, and >> i divides by scale */
2584 old_load
= this_rq
->cpu_load
[i
];
2585 new_load
= this_load
;
2587 * Round up the averaging division if load is increasing. This
2588 * prevents us from getting stuck on 9 if the load is 10, for
2591 if (new_load
> old_load
)
2592 new_load
+= scale
-1;
2593 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2600 * double_rq_lock - safely lock two runqueues
2602 * Note this does not disable interrupts like task_rq_lock,
2603 * you need to do so manually before calling.
2605 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2606 __acquires(rq1
->lock
)
2607 __acquires(rq2
->lock
)
2609 BUG_ON(!irqs_disabled());
2611 spin_lock(&rq1
->lock
);
2612 __acquire(rq2
->lock
); /* Fake it out ;) */
2615 spin_lock(&rq1
->lock
);
2616 spin_lock(&rq2
->lock
);
2618 spin_lock(&rq2
->lock
);
2619 spin_lock(&rq1
->lock
);
2622 update_rq_clock(rq1
);
2623 update_rq_clock(rq2
);
2627 * double_rq_unlock - safely unlock two runqueues
2629 * Note this does not restore interrupts like task_rq_unlock,
2630 * you need to do so manually after calling.
2632 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2633 __releases(rq1
->lock
)
2634 __releases(rq2
->lock
)
2636 spin_unlock(&rq1
->lock
);
2638 spin_unlock(&rq2
->lock
);
2640 __release(rq2
->lock
);
2644 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2646 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2647 __releases(this_rq
->lock
)
2648 __acquires(busiest
->lock
)
2649 __acquires(this_rq
->lock
)
2653 if (unlikely(!irqs_disabled())) {
2654 /* printk() doesn't work good under rq->lock */
2655 spin_unlock(&this_rq
->lock
);
2658 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2659 if (busiest
< this_rq
) {
2660 spin_unlock(&this_rq
->lock
);
2661 spin_lock(&busiest
->lock
);
2662 spin_lock(&this_rq
->lock
);
2665 spin_lock(&busiest
->lock
);
2671 * If dest_cpu is allowed for this process, migrate the task to it.
2672 * This is accomplished by forcing the cpu_allowed mask to only
2673 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2674 * the cpu_allowed mask is restored.
2676 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2678 struct migration_req req
;
2679 unsigned long flags
;
2682 rq
= task_rq_lock(p
, &flags
);
2683 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2684 || unlikely(cpu_is_offline(dest_cpu
)))
2687 /* force the process onto the specified CPU */
2688 if (migrate_task(p
, dest_cpu
, &req
)) {
2689 /* Need to wait for migration thread (might exit: take ref). */
2690 struct task_struct
*mt
= rq
->migration_thread
;
2692 get_task_struct(mt
);
2693 task_rq_unlock(rq
, &flags
);
2694 wake_up_process(mt
);
2695 put_task_struct(mt
);
2696 wait_for_completion(&req
.done
);
2701 task_rq_unlock(rq
, &flags
);
2705 * sched_exec - execve() is a valuable balancing opportunity, because at
2706 * this point the task has the smallest effective memory and cache footprint.
2708 void sched_exec(void)
2710 int new_cpu
, this_cpu
= get_cpu();
2711 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2713 if (new_cpu
!= this_cpu
)
2714 sched_migrate_task(current
, new_cpu
);
2718 * pull_task - move a task from a remote runqueue to the local runqueue.
2719 * Both runqueues must be locked.
2721 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2722 struct rq
*this_rq
, int this_cpu
)
2724 deactivate_task(src_rq
, p
, 0);
2725 set_task_cpu(p
, this_cpu
);
2726 activate_task(this_rq
, p
, 0);
2728 * Note that idle threads have a prio of MAX_PRIO, for this test
2729 * to be always true for them.
2731 check_preempt_curr(this_rq
, p
);
2735 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2738 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2739 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2743 * We do not migrate tasks that are:
2744 * 1) running (obviously), or
2745 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2746 * 3) are cache-hot on their current CPU.
2748 if (!cpu_isset(this_cpu
, p
->cpus_allowed
)) {
2749 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2754 if (task_running(rq
, p
)) {
2755 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2760 * Aggressive migration if:
2761 * 1) task is cache cold, or
2762 * 2) too many balance attempts have failed.
2765 if (!task_hot(p
, rq
->clock
, sd
) ||
2766 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2767 #ifdef CONFIG_SCHEDSTATS
2768 if (task_hot(p
, rq
->clock
, sd
)) {
2769 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2770 schedstat_inc(p
, se
.nr_forced_migrations
);
2776 if (task_hot(p
, rq
->clock
, sd
)) {
2777 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
2783 static unsigned long
2784 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2785 unsigned long max_load_move
, struct sched_domain
*sd
,
2786 enum cpu_idle_type idle
, int *all_pinned
,
2787 int *this_best_prio
, struct rq_iterator
*iterator
)
2789 int loops
= 0, pulled
= 0, pinned
= 0, skip_for_load
;
2790 struct task_struct
*p
;
2791 long rem_load_move
= max_load_move
;
2793 if (max_load_move
== 0)
2799 * Start the load-balancing iterator:
2801 p
= iterator
->start(iterator
->arg
);
2803 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
2806 * To help distribute high priority tasks across CPUs we don't
2807 * skip a task if it will be the highest priority task (i.e. smallest
2808 * prio value) on its new queue regardless of its load weight
2810 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2811 SCHED_LOAD_SCALE_FUZZ
;
2812 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2813 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2814 p
= iterator
->next(iterator
->arg
);
2818 pull_task(busiest
, p
, this_rq
, this_cpu
);
2820 rem_load_move
-= p
->se
.load
.weight
;
2823 * We only want to steal up to the prescribed amount of weighted load.
2825 if (rem_load_move
> 0) {
2826 if (p
->prio
< *this_best_prio
)
2827 *this_best_prio
= p
->prio
;
2828 p
= iterator
->next(iterator
->arg
);
2833 * Right now, this is one of only two places pull_task() is called,
2834 * so we can safely collect pull_task() stats here rather than
2835 * inside pull_task().
2837 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2840 *all_pinned
= pinned
;
2842 return max_load_move
- rem_load_move
;
2846 * move_tasks tries to move up to max_load_move weighted load from busiest to
2847 * this_rq, as part of a balancing operation within domain "sd".
2848 * Returns 1 if successful and 0 otherwise.
2850 * Called with both runqueues locked.
2852 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2853 unsigned long max_load_move
,
2854 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2857 const struct sched_class
*class = sched_class_highest
;
2858 unsigned long total_load_moved
= 0;
2859 int this_best_prio
= this_rq
->curr
->prio
;
2863 class->load_balance(this_rq
, this_cpu
, busiest
,
2864 max_load_move
- total_load_moved
,
2865 sd
, idle
, all_pinned
, &this_best_prio
);
2866 class = class->next
;
2867 } while (class && max_load_move
> total_load_moved
);
2869 return total_load_moved
> 0;
2873 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2874 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2875 struct rq_iterator
*iterator
)
2877 struct task_struct
*p
= iterator
->start(iterator
->arg
);
2881 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2882 pull_task(busiest
, p
, this_rq
, this_cpu
);
2884 * Right now, this is only the second place pull_task()
2885 * is called, so we can safely collect pull_task()
2886 * stats here rather than inside pull_task().
2888 schedstat_inc(sd
, lb_gained
[idle
]);
2892 p
= iterator
->next(iterator
->arg
);
2899 * move_one_task tries to move exactly one task from busiest to this_rq, as
2900 * part of active balancing operations within "domain".
2901 * Returns 1 if successful and 0 otherwise.
2903 * Called with both runqueues locked.
2905 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2906 struct sched_domain
*sd
, enum cpu_idle_type idle
)
2908 const struct sched_class
*class;
2910 for (class = sched_class_highest
; class; class = class->next
)
2911 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
2918 * find_busiest_group finds and returns the busiest CPU group within the
2919 * domain. It calculates and returns the amount of weighted load which
2920 * should be moved to restore balance via the imbalance parameter.
2922 static struct sched_group
*
2923 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2924 unsigned long *imbalance
, enum cpu_idle_type idle
,
2925 int *sd_idle
, const cpumask_t
*cpus
, int *balance
)
2927 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2928 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2929 unsigned long max_pull
;
2930 unsigned long busiest_load_per_task
, busiest_nr_running
;
2931 unsigned long this_load_per_task
, this_nr_running
;
2932 int load_idx
, group_imb
= 0;
2933 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2934 int power_savings_balance
= 1;
2935 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2936 unsigned long min_nr_running
= ULONG_MAX
;
2937 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2940 max_load
= this_load
= total_load
= total_pwr
= 0;
2941 busiest_load_per_task
= busiest_nr_running
= 0;
2942 this_load_per_task
= this_nr_running
= 0;
2943 if (idle
== CPU_NOT_IDLE
)
2944 load_idx
= sd
->busy_idx
;
2945 else if (idle
== CPU_NEWLY_IDLE
)
2946 load_idx
= sd
->newidle_idx
;
2948 load_idx
= sd
->idle_idx
;
2951 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
2954 int __group_imb
= 0;
2955 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2956 unsigned long sum_nr_running
, sum_weighted_load
;
2958 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2961 balance_cpu
= first_cpu(group
->cpumask
);
2963 /* Tally up the load of all CPUs in the group */
2964 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2966 min_cpu_load
= ~0UL;
2968 for_each_cpu_mask(i
, group
->cpumask
) {
2971 if (!cpu_isset(i
, *cpus
))
2976 if (*sd_idle
&& rq
->nr_running
)
2979 /* Bias balancing toward cpus of our domain */
2981 if (idle_cpu(i
) && !first_idle_cpu
) {
2986 load
= target_load(i
, load_idx
);
2988 load
= source_load(i
, load_idx
);
2989 if (load
> max_cpu_load
)
2990 max_cpu_load
= load
;
2991 if (min_cpu_load
> load
)
2992 min_cpu_load
= load
;
2996 sum_nr_running
+= rq
->nr_running
;
2997 sum_weighted_load
+= weighted_cpuload(i
);
3001 * First idle cpu or the first cpu(busiest) in this sched group
3002 * is eligible for doing load balancing at this and above
3003 * domains. In the newly idle case, we will allow all the cpu's
3004 * to do the newly idle load balance.
3006 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3007 balance_cpu
!= this_cpu
&& balance
) {
3012 total_load
+= avg_load
;
3013 total_pwr
+= group
->__cpu_power
;
3015 /* Adjust by relative CPU power of the group */
3016 avg_load
= sg_div_cpu_power(group
,
3017 avg_load
* SCHED_LOAD_SCALE
);
3019 if ((max_cpu_load
- min_cpu_load
) > SCHED_LOAD_SCALE
)
3022 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
3025 this_load
= avg_load
;
3027 this_nr_running
= sum_nr_running
;
3028 this_load_per_task
= sum_weighted_load
;
3029 } else if (avg_load
> max_load
&&
3030 (sum_nr_running
> group_capacity
|| __group_imb
)) {
3031 max_load
= avg_load
;
3033 busiest_nr_running
= sum_nr_running
;
3034 busiest_load_per_task
= sum_weighted_load
;
3035 group_imb
= __group_imb
;
3038 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3040 * Busy processors will not participate in power savings
3043 if (idle
== CPU_NOT_IDLE
||
3044 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3048 * If the local group is idle or completely loaded
3049 * no need to do power savings balance at this domain
3051 if (local_group
&& (this_nr_running
>= group_capacity
||
3053 power_savings_balance
= 0;
3056 * If a group is already running at full capacity or idle,
3057 * don't include that group in power savings calculations
3059 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
3064 * Calculate the group which has the least non-idle load.
3065 * This is the group from where we need to pick up the load
3068 if ((sum_nr_running
< min_nr_running
) ||
3069 (sum_nr_running
== min_nr_running
&&
3070 first_cpu(group
->cpumask
) <
3071 first_cpu(group_min
->cpumask
))) {
3073 min_nr_running
= sum_nr_running
;
3074 min_load_per_task
= sum_weighted_load
/
3079 * Calculate the group which is almost near its
3080 * capacity but still has some space to pick up some load
3081 * from other group and save more power
3083 if (sum_nr_running
<= group_capacity
- 1) {
3084 if (sum_nr_running
> leader_nr_running
||
3085 (sum_nr_running
== leader_nr_running
&&
3086 first_cpu(group
->cpumask
) >
3087 first_cpu(group_leader
->cpumask
))) {
3088 group_leader
= group
;
3089 leader_nr_running
= sum_nr_running
;
3094 group
= group
->next
;
3095 } while (group
!= sd
->groups
);
3097 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
3100 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
3102 if (this_load
>= avg_load
||
3103 100*max_load
<= sd
->imbalance_pct
*this_load
)
3106 busiest_load_per_task
/= busiest_nr_running
;
3108 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
3111 * We're trying to get all the cpus to the average_load, so we don't
3112 * want to push ourselves above the average load, nor do we wish to
3113 * reduce the max loaded cpu below the average load, as either of these
3114 * actions would just result in more rebalancing later, and ping-pong
3115 * tasks around. Thus we look for the minimum possible imbalance.
3116 * Negative imbalances (*we* are more loaded than anyone else) will
3117 * be counted as no imbalance for these purposes -- we can't fix that
3118 * by pulling tasks to us. Be careful of negative numbers as they'll
3119 * appear as very large values with unsigned longs.
3121 if (max_load
<= busiest_load_per_task
)
3125 * In the presence of smp nice balancing, certain scenarios can have
3126 * max load less than avg load(as we skip the groups at or below
3127 * its cpu_power, while calculating max_load..)
3129 if (max_load
< avg_load
) {
3131 goto small_imbalance
;
3134 /* Don't want to pull so many tasks that a group would go idle */
3135 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
3137 /* How much load to actually move to equalise the imbalance */
3138 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
3139 (avg_load
- this_load
) * this->__cpu_power
)
3143 * if *imbalance is less than the average load per runnable task
3144 * there is no gaurantee that any tasks will be moved so we'll have
3145 * a think about bumping its value to force at least one task to be
3148 if (*imbalance
< busiest_load_per_task
) {
3149 unsigned long tmp
, pwr_now
, pwr_move
;
3153 pwr_move
= pwr_now
= 0;
3155 if (this_nr_running
) {
3156 this_load_per_task
/= this_nr_running
;
3157 if (busiest_load_per_task
> this_load_per_task
)
3160 this_load_per_task
= SCHED_LOAD_SCALE
;
3162 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
3163 busiest_load_per_task
* imbn
) {
3164 *imbalance
= busiest_load_per_task
;
3169 * OK, we don't have enough imbalance to justify moving tasks,
3170 * however we may be able to increase total CPU power used by
3174 pwr_now
+= busiest
->__cpu_power
*
3175 min(busiest_load_per_task
, max_load
);
3176 pwr_now
+= this->__cpu_power
*
3177 min(this_load_per_task
, this_load
);
3178 pwr_now
/= SCHED_LOAD_SCALE
;
3180 /* Amount of load we'd subtract */
3181 tmp
= sg_div_cpu_power(busiest
,
3182 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3184 pwr_move
+= busiest
->__cpu_power
*
3185 min(busiest_load_per_task
, max_load
- tmp
);
3187 /* Amount of load we'd add */
3188 if (max_load
* busiest
->__cpu_power
<
3189 busiest_load_per_task
* SCHED_LOAD_SCALE
)
3190 tmp
= sg_div_cpu_power(this,
3191 max_load
* busiest
->__cpu_power
);
3193 tmp
= sg_div_cpu_power(this,
3194 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3195 pwr_move
+= this->__cpu_power
*
3196 min(this_load_per_task
, this_load
+ tmp
);
3197 pwr_move
/= SCHED_LOAD_SCALE
;
3199 /* Move if we gain throughput */
3200 if (pwr_move
> pwr_now
)
3201 *imbalance
= busiest_load_per_task
;
3207 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3208 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3211 if (this == group_leader
&& group_leader
!= group_min
) {
3212 *imbalance
= min_load_per_task
;
3222 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3225 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
3226 unsigned long imbalance
, const cpumask_t
*cpus
)
3228 struct rq
*busiest
= NULL
, *rq
;
3229 unsigned long max_load
= 0;
3232 for_each_cpu_mask(i
, group
->cpumask
) {
3235 if (!cpu_isset(i
, *cpus
))
3239 wl
= weighted_cpuload(i
);
3241 if (rq
->nr_running
== 1 && wl
> imbalance
)
3244 if (wl
> max_load
) {
3254 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3255 * so long as it is large enough.
3257 #define MAX_PINNED_INTERVAL 512
3260 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3261 * tasks if there is an imbalance.
3263 static int load_balance(int this_cpu
, struct rq
*this_rq
,
3264 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3265 int *balance
, cpumask_t
*cpus
)
3267 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
3268 struct sched_group
*group
;
3269 unsigned long imbalance
;
3271 unsigned long flags
;
3276 * When power savings policy is enabled for the parent domain, idle
3277 * sibling can pick up load irrespective of busy siblings. In this case,
3278 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3279 * portraying it as CPU_NOT_IDLE.
3281 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3282 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3285 schedstat_inc(sd
, lb_count
[idle
]);
3288 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3295 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3299 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
3301 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3305 BUG_ON(busiest
== this_rq
);
3307 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3310 if (busiest
->nr_running
> 1) {
3312 * Attempt to move tasks. If find_busiest_group has found
3313 * an imbalance but busiest->nr_running <= 1, the group is
3314 * still unbalanced. ld_moved simply stays zero, so it is
3315 * correctly treated as an imbalance.
3317 local_irq_save(flags
);
3318 double_rq_lock(this_rq
, busiest
);
3319 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3320 imbalance
, sd
, idle
, &all_pinned
);
3321 double_rq_unlock(this_rq
, busiest
);
3322 local_irq_restore(flags
);
3325 * some other cpu did the load balance for us.
3327 if (ld_moved
&& this_cpu
!= smp_processor_id())
3328 resched_cpu(this_cpu
);
3330 /* All tasks on this runqueue were pinned by CPU affinity */
3331 if (unlikely(all_pinned
)) {
3332 cpu_clear(cpu_of(busiest
), *cpus
);
3333 if (!cpus_empty(*cpus
))
3340 schedstat_inc(sd
, lb_failed
[idle
]);
3341 sd
->nr_balance_failed
++;
3343 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3345 spin_lock_irqsave(&busiest
->lock
, flags
);
3347 /* don't kick the migration_thread, if the curr
3348 * task on busiest cpu can't be moved to this_cpu
3350 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
3351 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3353 goto out_one_pinned
;
3356 if (!busiest
->active_balance
) {
3357 busiest
->active_balance
= 1;
3358 busiest
->push_cpu
= this_cpu
;
3361 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3363 wake_up_process(busiest
->migration_thread
);
3366 * We've kicked active balancing, reset the failure
3369 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3372 sd
->nr_balance_failed
= 0;
3374 if (likely(!active_balance
)) {
3375 /* We were unbalanced, so reset the balancing interval */
3376 sd
->balance_interval
= sd
->min_interval
;
3379 * If we've begun active balancing, start to back off. This
3380 * case may not be covered by the all_pinned logic if there
3381 * is only 1 task on the busy runqueue (because we don't call
3384 if (sd
->balance_interval
< sd
->max_interval
)
3385 sd
->balance_interval
*= 2;
3388 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3389 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3394 schedstat_inc(sd
, lb_balanced
[idle
]);
3396 sd
->nr_balance_failed
= 0;
3399 /* tune up the balancing interval */
3400 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3401 (sd
->balance_interval
< sd
->max_interval
))
3402 sd
->balance_interval
*= 2;
3404 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3405 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3411 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3412 * tasks if there is an imbalance.
3414 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3415 * this_rq is locked.
3418 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
,
3421 struct sched_group
*group
;
3422 struct rq
*busiest
= NULL
;
3423 unsigned long imbalance
;
3431 * When power savings policy is enabled for the parent domain, idle
3432 * sibling can pick up load irrespective of busy siblings. In this case,
3433 * let the state of idle sibling percolate up as IDLE, instead of
3434 * portraying it as CPU_NOT_IDLE.
3436 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
3437 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3440 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
3442 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
3443 &sd_idle
, cpus
, NULL
);
3445 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
3449 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
3451 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
3455 BUG_ON(busiest
== this_rq
);
3457 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
3460 if (busiest
->nr_running
> 1) {
3461 /* Attempt to move tasks */
3462 double_lock_balance(this_rq
, busiest
);
3463 /* this_rq->clock is already updated */
3464 update_rq_clock(busiest
);
3465 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3466 imbalance
, sd
, CPU_NEWLY_IDLE
,
3468 spin_unlock(&busiest
->lock
);
3470 if (unlikely(all_pinned
)) {
3471 cpu_clear(cpu_of(busiest
), *cpus
);
3472 if (!cpus_empty(*cpus
))
3478 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
3479 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3480 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3483 sd
->nr_balance_failed
= 0;
3488 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
3489 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3490 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3492 sd
->nr_balance_failed
= 0;
3498 * idle_balance is called by schedule() if this_cpu is about to become
3499 * idle. Attempts to pull tasks from other CPUs.
3501 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
3503 struct sched_domain
*sd
;
3504 int pulled_task
= -1;
3505 unsigned long next_balance
= jiffies
+ HZ
;
3508 for_each_domain(this_cpu
, sd
) {
3509 unsigned long interval
;
3511 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3514 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
3515 /* If we've pulled tasks over stop searching: */
3516 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
3519 interval
= msecs_to_jiffies(sd
->balance_interval
);
3520 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3521 next_balance
= sd
->last_balance
+ interval
;
3525 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3527 * We are going idle. next_balance may be set based on
3528 * a busy processor. So reset next_balance.
3530 this_rq
->next_balance
= next_balance
;
3535 * active_load_balance is run by migration threads. It pushes running tasks
3536 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3537 * running on each physical CPU where possible, and avoids physical /
3538 * logical imbalances.
3540 * Called with busiest_rq locked.
3542 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
3544 int target_cpu
= busiest_rq
->push_cpu
;
3545 struct sched_domain
*sd
;
3546 struct rq
*target_rq
;
3548 /* Is there any task to move? */
3549 if (busiest_rq
->nr_running
<= 1)
3552 target_rq
= cpu_rq(target_cpu
);
3555 * This condition is "impossible", if it occurs
3556 * we need to fix it. Originally reported by
3557 * Bjorn Helgaas on a 128-cpu setup.
3559 BUG_ON(busiest_rq
== target_rq
);
3561 /* move a task from busiest_rq to target_rq */
3562 double_lock_balance(busiest_rq
, target_rq
);
3563 update_rq_clock(busiest_rq
);
3564 update_rq_clock(target_rq
);
3566 /* Search for an sd spanning us and the target CPU. */
3567 for_each_domain(target_cpu
, sd
) {
3568 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3569 cpu_isset(busiest_cpu
, sd
->span
))
3574 schedstat_inc(sd
, alb_count
);
3576 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3578 schedstat_inc(sd
, alb_pushed
);
3580 schedstat_inc(sd
, alb_failed
);
3582 spin_unlock(&target_rq
->lock
);
3587 atomic_t load_balancer
;
3589 } nohz ____cacheline_aligned
= {
3590 .load_balancer
= ATOMIC_INIT(-1),
3591 .cpu_mask
= CPU_MASK_NONE
,
3595 * This routine will try to nominate the ilb (idle load balancing)
3596 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3597 * load balancing on behalf of all those cpus. If all the cpus in the system
3598 * go into this tickless mode, then there will be no ilb owner (as there is
3599 * no need for one) and all the cpus will sleep till the next wakeup event
3602 * For the ilb owner, tick is not stopped. And this tick will be used
3603 * for idle load balancing. ilb owner will still be part of
3606 * While stopping the tick, this cpu will become the ilb owner if there
3607 * is no other owner. And will be the owner till that cpu becomes busy
3608 * or if all cpus in the system stop their ticks at which point
3609 * there is no need for ilb owner.
3611 * When the ilb owner becomes busy, it nominates another owner, during the
3612 * next busy scheduler_tick()
3614 int select_nohz_load_balancer(int stop_tick
)
3616 int cpu
= smp_processor_id();
3619 cpu_set(cpu
, nohz
.cpu_mask
);
3620 cpu_rq(cpu
)->in_nohz_recently
= 1;
3623 * If we are going offline and still the leader, give up!
3625 if (cpu_is_offline(cpu
) &&
3626 atomic_read(&nohz
.load_balancer
) == cpu
) {
3627 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3632 /* time for ilb owner also to sleep */
3633 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3634 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3635 atomic_set(&nohz
.load_balancer
, -1);
3639 if (atomic_read(&nohz
.load_balancer
) == -1) {
3640 /* make me the ilb owner */
3641 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3643 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3646 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3649 cpu_clear(cpu
, nohz
.cpu_mask
);
3651 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3652 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3659 static DEFINE_SPINLOCK(balancing
);
3662 * It checks each scheduling domain to see if it is due to be balanced,
3663 * and initiates a balancing operation if so.
3665 * Balancing parameters are set up in arch_init_sched_domains.
3667 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3670 struct rq
*rq
= cpu_rq(cpu
);
3671 unsigned long interval
;
3672 struct sched_domain
*sd
;
3673 /* Earliest time when we have to do rebalance again */
3674 unsigned long next_balance
= jiffies
+ 60*HZ
;
3675 int update_next_balance
= 0;
3679 for_each_domain(cpu
, sd
) {
3680 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3683 interval
= sd
->balance_interval
;
3684 if (idle
!= CPU_IDLE
)
3685 interval
*= sd
->busy_factor
;
3687 /* scale ms to jiffies */
3688 interval
= msecs_to_jiffies(interval
);
3689 if (unlikely(!interval
))
3691 if (interval
> HZ
*NR_CPUS
/10)
3692 interval
= HZ
*NR_CPUS
/10;
3694 need_serialize
= sd
->flags
& SD_SERIALIZE
;
3696 if (need_serialize
) {
3697 if (!spin_trylock(&balancing
))
3701 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3702 if (load_balance(cpu
, rq
, sd
, idle
, &balance
, &tmp
)) {
3704 * We've pulled tasks over so either we're no
3705 * longer idle, or one of our SMT siblings is
3708 idle
= CPU_NOT_IDLE
;
3710 sd
->last_balance
= jiffies
;
3713 spin_unlock(&balancing
);
3715 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3716 next_balance
= sd
->last_balance
+ interval
;
3717 update_next_balance
= 1;
3721 * Stop the load balance at this level. There is another
3722 * CPU in our sched group which is doing load balancing more
3730 * next_balance will be updated only when there is a need.
3731 * When the cpu is attached to null domain for ex, it will not be
3734 if (likely(update_next_balance
))
3735 rq
->next_balance
= next_balance
;
3739 * run_rebalance_domains is triggered when needed from the scheduler tick.
3740 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3741 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3743 static void run_rebalance_domains(struct softirq_action
*h
)
3745 int this_cpu
= smp_processor_id();
3746 struct rq
*this_rq
= cpu_rq(this_cpu
);
3747 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3748 CPU_IDLE
: CPU_NOT_IDLE
;
3750 rebalance_domains(this_cpu
, idle
);
3754 * If this cpu is the owner for idle load balancing, then do the
3755 * balancing on behalf of the other idle cpus whose ticks are
3758 if (this_rq
->idle_at_tick
&&
3759 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3760 cpumask_t cpus
= nohz
.cpu_mask
;
3764 cpu_clear(this_cpu
, cpus
);
3765 for_each_cpu_mask(balance_cpu
, cpus
) {
3767 * If this cpu gets work to do, stop the load balancing
3768 * work being done for other cpus. Next load
3769 * balancing owner will pick it up.
3774 rebalance_domains(balance_cpu
, CPU_IDLE
);
3776 rq
= cpu_rq(balance_cpu
);
3777 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3778 this_rq
->next_balance
= rq
->next_balance
;
3785 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3787 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3788 * idle load balancing owner or decide to stop the periodic load balancing,
3789 * if the whole system is idle.
3791 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3795 * If we were in the nohz mode recently and busy at the current
3796 * scheduler tick, then check if we need to nominate new idle
3799 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3800 rq
->in_nohz_recently
= 0;
3802 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3803 cpu_clear(cpu
, nohz
.cpu_mask
);
3804 atomic_set(&nohz
.load_balancer
, -1);
3807 if (atomic_read(&nohz
.load_balancer
) == -1) {
3809 * simple selection for now: Nominate the
3810 * first cpu in the nohz list to be the next
3813 * TBD: Traverse the sched domains and nominate
3814 * the nearest cpu in the nohz.cpu_mask.
3816 int ilb
= first_cpu(nohz
.cpu_mask
);
3818 if (ilb
< nr_cpu_ids
)
3824 * If this cpu is idle and doing idle load balancing for all the
3825 * cpus with ticks stopped, is it time for that to stop?
3827 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3828 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3834 * If this cpu is idle and the idle load balancing is done by
3835 * someone else, then no need raise the SCHED_SOFTIRQ
3837 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3838 cpu_isset(cpu
, nohz
.cpu_mask
))
3841 if (time_after_eq(jiffies
, rq
->next_balance
))
3842 raise_softirq(SCHED_SOFTIRQ
);
3845 #else /* CONFIG_SMP */
3848 * on UP we do not need to balance between CPUs:
3850 static inline void idle_balance(int cpu
, struct rq
*rq
)
3856 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3858 EXPORT_PER_CPU_SYMBOL(kstat
);
3861 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3862 * that have not yet been banked in case the task is currently running.
3864 unsigned long long task_sched_runtime(struct task_struct
*p
)
3866 unsigned long flags
;
3870 rq
= task_rq_lock(p
, &flags
);
3871 ns
= p
->se
.sum_exec_runtime
;
3872 if (task_current(rq
, p
)) {
3873 update_rq_clock(rq
);
3874 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3875 if ((s64
)delta_exec
> 0)
3878 task_rq_unlock(rq
, &flags
);
3884 * Account user cpu time to a process.
3885 * @p: the process that the cpu time gets accounted to
3886 * @cputime: the cpu time spent in user space since the last update
3888 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3890 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3893 p
->utime
= cputime_add(p
->utime
, cputime
);
3895 /* Add user time to cpustat. */
3896 tmp
= cputime_to_cputime64(cputime
);
3897 if (TASK_NICE(p
) > 0)
3898 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3900 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3904 * Account guest cpu time to a process.
3905 * @p: the process that the cpu time gets accounted to
3906 * @cputime: the cpu time spent in virtual machine since the last update
3908 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
)
3911 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3913 tmp
= cputime_to_cputime64(cputime
);
3915 p
->utime
= cputime_add(p
->utime
, cputime
);
3916 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3918 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3919 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3923 * Account scaled user cpu time to a process.
3924 * @p: the process that the cpu time gets accounted to
3925 * @cputime: the cpu time spent in user space since the last update
3927 void account_user_time_scaled(struct task_struct
*p
, cputime_t cputime
)
3929 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime
);
3933 * Account system cpu time to a process.
3934 * @p: the process that the cpu time gets accounted to
3935 * @hardirq_offset: the offset to subtract from hardirq_count()
3936 * @cputime: the cpu time spent in kernel space since the last update
3938 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3941 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3942 struct rq
*rq
= this_rq();
3945 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
3946 account_guest_time(p
, cputime
);
3950 p
->stime
= cputime_add(p
->stime
, cputime
);
3952 /* Add system time to cpustat. */
3953 tmp
= cputime_to_cputime64(cputime
);
3954 if (hardirq_count() - hardirq_offset
)
3955 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3956 else if (softirq_count())
3957 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3958 else if (p
!= rq
->idle
)
3959 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3960 else if (atomic_read(&rq
->nr_iowait
) > 0)
3961 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3963 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3964 /* Account for system time used */
3965 acct_update_integrals(p
);
3969 * Account scaled system cpu time to a process.
3970 * @p: the process that the cpu time gets accounted to
3971 * @hardirq_offset: the offset to subtract from hardirq_count()
3972 * @cputime: the cpu time spent in kernel space since the last update
3974 void account_system_time_scaled(struct task_struct
*p
, cputime_t cputime
)
3976 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime
);
3980 * Account for involuntary wait time.
3981 * @p: the process from which the cpu time has been stolen
3982 * @steal: the cpu time spent in involuntary wait
3984 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3986 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3987 cputime64_t tmp
= cputime_to_cputime64(steal
);
3988 struct rq
*rq
= this_rq();
3990 if (p
== rq
->idle
) {
3991 p
->stime
= cputime_add(p
->stime
, steal
);
3992 if (atomic_read(&rq
->nr_iowait
) > 0)
3993 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3995 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3997 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
4001 * This function gets called by the timer code, with HZ frequency.
4002 * We call it with interrupts disabled.
4004 * It also gets called by the fork code, when changing the parent's
4007 void scheduler_tick(void)
4009 int cpu
= smp_processor_id();
4010 struct rq
*rq
= cpu_rq(cpu
);
4011 struct task_struct
*curr
= rq
->curr
;
4015 spin_lock(&rq
->lock
);
4016 update_rq_clock(rq
);
4017 update_cpu_load(rq
);
4018 curr
->sched_class
->task_tick(rq
, curr
, 0);
4019 spin_unlock(&rq
->lock
);
4022 rq
->idle_at_tick
= idle_cpu(cpu
);
4023 trigger_load_balance(rq
, cpu
);
4027 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
4029 void __kprobes
add_preempt_count(int val
)
4034 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4036 preempt_count() += val
;
4038 * Spinlock count overflowing soon?
4040 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4043 EXPORT_SYMBOL(add_preempt_count
);
4045 void __kprobes
sub_preempt_count(int val
)
4050 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
4053 * Is the spinlock portion underflowing?
4055 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4056 !(preempt_count() & PREEMPT_MASK
)))
4059 preempt_count() -= val
;
4061 EXPORT_SYMBOL(sub_preempt_count
);
4066 * Print scheduling while atomic bug:
4068 static noinline
void __schedule_bug(struct task_struct
*prev
)
4070 struct pt_regs
*regs
= get_irq_regs();
4072 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4073 prev
->comm
, prev
->pid
, preempt_count());
4075 debug_show_held_locks(prev
);
4077 if (irqs_disabled())
4078 print_irqtrace_events(prev
);
4087 * Various schedule()-time debugging checks and statistics:
4089 static inline void schedule_debug(struct task_struct
*prev
)
4092 * Test if we are atomic. Since do_exit() needs to call into
4093 * schedule() atomically, we ignore that path for now.
4094 * Otherwise, whine if we are scheduling when we should not be.
4096 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
4097 __schedule_bug(prev
);
4099 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4101 schedstat_inc(this_rq(), sched_count
);
4102 #ifdef CONFIG_SCHEDSTATS
4103 if (unlikely(prev
->lock_depth
>= 0)) {
4104 schedstat_inc(this_rq(), bkl_count
);
4105 schedstat_inc(prev
, sched_info
.bkl_count
);
4111 * Pick up the highest-prio task:
4113 static inline struct task_struct
*
4114 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
4116 const struct sched_class
*class;
4117 struct task_struct
*p
;
4120 * Optimization: we know that if all tasks are in
4121 * the fair class we can call that function directly:
4123 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
4124 p
= fair_sched_class
.pick_next_task(rq
);
4129 class = sched_class_highest
;
4131 p
= class->pick_next_task(rq
);
4135 * Will never be NULL as the idle class always
4136 * returns a non-NULL p:
4138 class = class->next
;
4143 * schedule() is the main scheduler function.
4145 asmlinkage
void __sched
schedule(void)
4147 struct task_struct
*prev
, *next
;
4148 unsigned long *switch_count
;
4150 int cpu
, hrtick
= sched_feat(HRTICK
);
4154 cpu
= smp_processor_id();
4158 switch_count
= &prev
->nivcsw
;
4160 release_kernel_lock(prev
);
4161 need_resched_nonpreemptible
:
4163 schedule_debug(prev
);
4169 * Do the rq-clock update outside the rq lock:
4171 local_irq_disable();
4172 update_rq_clock(rq
);
4173 spin_lock(&rq
->lock
);
4174 clear_tsk_need_resched(prev
);
4176 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
4177 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
4178 signal_pending(prev
))) {
4179 prev
->state
= TASK_RUNNING
;
4181 deactivate_task(rq
, prev
, 1);
4183 switch_count
= &prev
->nvcsw
;
4187 if (prev
->sched_class
->pre_schedule
)
4188 prev
->sched_class
->pre_schedule(rq
, prev
);
4191 if (unlikely(!rq
->nr_running
))
4192 idle_balance(cpu
, rq
);
4194 prev
->sched_class
->put_prev_task(rq
, prev
);
4195 next
= pick_next_task(rq
, prev
);
4197 if (likely(prev
!= next
)) {
4198 sched_info_switch(prev
, next
);
4204 context_switch(rq
, prev
, next
); /* unlocks the rq */
4206 * the context switch might have flipped the stack from under
4207 * us, hence refresh the local variables.
4209 cpu
= smp_processor_id();
4212 spin_unlock_irq(&rq
->lock
);
4217 if (unlikely(reacquire_kernel_lock(current
) < 0))
4218 goto need_resched_nonpreemptible
;
4220 preempt_enable_no_resched();
4221 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
4224 EXPORT_SYMBOL(schedule
);
4226 #ifdef CONFIG_PREEMPT
4228 * this is the entry point to schedule() from in-kernel preemption
4229 * off of preempt_enable. Kernel preemptions off return from interrupt
4230 * occur there and call schedule directly.
4232 asmlinkage
void __sched
preempt_schedule(void)
4234 struct thread_info
*ti
= current_thread_info();
4237 * If there is a non-zero preempt_count or interrupts are disabled,
4238 * we do not want to preempt the current task. Just return..
4240 if (likely(ti
->preempt_count
|| irqs_disabled()))
4244 add_preempt_count(PREEMPT_ACTIVE
);
4246 sub_preempt_count(PREEMPT_ACTIVE
);
4249 * Check again in case we missed a preemption opportunity
4250 * between schedule and now.
4253 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4255 EXPORT_SYMBOL(preempt_schedule
);
4258 * this is the entry point to schedule() from kernel preemption
4259 * off of irq context.
4260 * Note, that this is called and return with irqs disabled. This will
4261 * protect us against recursive calling from irq.
4263 asmlinkage
void __sched
preempt_schedule_irq(void)
4265 struct thread_info
*ti
= current_thread_info();
4267 /* Catch callers which need to be fixed */
4268 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
4271 add_preempt_count(PREEMPT_ACTIVE
);
4274 local_irq_disable();
4275 sub_preempt_count(PREEMPT_ACTIVE
);
4278 * Check again in case we missed a preemption opportunity
4279 * between schedule and now.
4282 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4285 #endif /* CONFIG_PREEMPT */
4287 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
4290 return try_to_wake_up(curr
->private, mode
, sync
);
4292 EXPORT_SYMBOL(default_wake_function
);
4295 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4296 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4297 * number) then we wake all the non-exclusive tasks and one exclusive task.
4299 * There are circumstances in which we can try to wake a task which has already
4300 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4301 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4303 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4304 int nr_exclusive
, int sync
, void *key
)
4306 wait_queue_t
*curr
, *next
;
4308 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4309 unsigned flags
= curr
->flags
;
4311 if (curr
->func(curr
, mode
, sync
, key
) &&
4312 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4318 * __wake_up - wake up threads blocked on a waitqueue.
4320 * @mode: which threads
4321 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4322 * @key: is directly passed to the wakeup function
4324 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4325 int nr_exclusive
, void *key
)
4327 unsigned long flags
;
4329 spin_lock_irqsave(&q
->lock
, flags
);
4330 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4331 spin_unlock_irqrestore(&q
->lock
, flags
);
4333 EXPORT_SYMBOL(__wake_up
);
4336 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4338 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4340 __wake_up_common(q
, mode
, 1, 0, NULL
);
4344 * __wake_up_sync - wake up threads blocked on a waitqueue.
4346 * @mode: which threads
4347 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4349 * The sync wakeup differs that the waker knows that it will schedule
4350 * away soon, so while the target thread will be woken up, it will not
4351 * be migrated to another CPU - ie. the two threads are 'synchronized'
4352 * with each other. This can prevent needless bouncing between CPUs.
4354 * On UP it can prevent extra preemption.
4357 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4359 unsigned long flags
;
4365 if (unlikely(!nr_exclusive
))
4368 spin_lock_irqsave(&q
->lock
, flags
);
4369 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
4370 spin_unlock_irqrestore(&q
->lock
, flags
);
4372 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4374 void complete(struct completion
*x
)
4376 unsigned long flags
;
4378 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4380 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4381 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4383 EXPORT_SYMBOL(complete
);
4385 void complete_all(struct completion
*x
)
4387 unsigned long flags
;
4389 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4390 x
->done
+= UINT_MAX
/2;
4391 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4392 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4394 EXPORT_SYMBOL(complete_all
);
4396 static inline long __sched
4397 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4400 DECLARE_WAITQUEUE(wait
, current
);
4402 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4403 __add_wait_queue_tail(&x
->wait
, &wait
);
4405 if ((state
== TASK_INTERRUPTIBLE
&&
4406 signal_pending(current
)) ||
4407 (state
== TASK_KILLABLE
&&
4408 fatal_signal_pending(current
))) {
4409 __remove_wait_queue(&x
->wait
, &wait
);
4410 return -ERESTARTSYS
;
4412 __set_current_state(state
);
4413 spin_unlock_irq(&x
->wait
.lock
);
4414 timeout
= schedule_timeout(timeout
);
4415 spin_lock_irq(&x
->wait
.lock
);
4417 __remove_wait_queue(&x
->wait
, &wait
);
4421 __remove_wait_queue(&x
->wait
, &wait
);
4428 wait_for_common(struct completion
*x
, long timeout
, int state
)
4432 spin_lock_irq(&x
->wait
.lock
);
4433 timeout
= do_wait_for_common(x
, timeout
, state
);
4434 spin_unlock_irq(&x
->wait
.lock
);
4438 void __sched
wait_for_completion(struct completion
*x
)
4440 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4442 EXPORT_SYMBOL(wait_for_completion
);
4444 unsigned long __sched
4445 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4447 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4449 EXPORT_SYMBOL(wait_for_completion_timeout
);
4451 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4453 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4454 if (t
== -ERESTARTSYS
)
4458 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4460 unsigned long __sched
4461 wait_for_completion_interruptible_timeout(struct completion
*x
,
4462 unsigned long timeout
)
4464 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4466 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4468 int __sched
wait_for_completion_killable(struct completion
*x
)
4470 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4471 if (t
== -ERESTARTSYS
)
4475 EXPORT_SYMBOL(wait_for_completion_killable
);
4478 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4480 unsigned long flags
;
4483 init_waitqueue_entry(&wait
, current
);
4485 __set_current_state(state
);
4487 spin_lock_irqsave(&q
->lock
, flags
);
4488 __add_wait_queue(q
, &wait
);
4489 spin_unlock(&q
->lock
);
4490 timeout
= schedule_timeout(timeout
);
4491 spin_lock_irq(&q
->lock
);
4492 __remove_wait_queue(q
, &wait
);
4493 spin_unlock_irqrestore(&q
->lock
, flags
);
4498 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4500 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4502 EXPORT_SYMBOL(interruptible_sleep_on
);
4505 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4507 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4509 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4511 void __sched
sleep_on(wait_queue_head_t
*q
)
4513 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4515 EXPORT_SYMBOL(sleep_on
);
4517 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4519 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4521 EXPORT_SYMBOL(sleep_on_timeout
);
4523 #ifdef CONFIG_RT_MUTEXES
4526 * rt_mutex_setprio - set the current priority of a task
4528 * @prio: prio value (kernel-internal form)
4530 * This function changes the 'effective' priority of a task. It does
4531 * not touch ->normal_prio like __setscheduler().
4533 * Used by the rt_mutex code to implement priority inheritance logic.
4535 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4537 unsigned long flags
;
4538 int oldprio
, on_rq
, running
;
4540 const struct sched_class
*prev_class
= p
->sched_class
;
4542 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4544 rq
= task_rq_lock(p
, &flags
);
4545 update_rq_clock(rq
);
4548 on_rq
= p
->se
.on_rq
;
4549 running
= task_current(rq
, p
);
4551 dequeue_task(rq
, p
, 0);
4553 p
->sched_class
->put_prev_task(rq
, p
);
4556 p
->sched_class
= &rt_sched_class
;
4558 p
->sched_class
= &fair_sched_class
;
4563 p
->sched_class
->set_curr_task(rq
);
4565 enqueue_task(rq
, p
, 0);
4567 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4569 task_rq_unlock(rq
, &flags
);
4574 void set_user_nice(struct task_struct
*p
, long nice
)
4576 int old_prio
, delta
, on_rq
;
4577 unsigned long flags
;
4580 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4583 * We have to be careful, if called from sys_setpriority(),
4584 * the task might be in the middle of scheduling on another CPU.
4586 rq
= task_rq_lock(p
, &flags
);
4587 update_rq_clock(rq
);
4589 * The RT priorities are set via sched_setscheduler(), but we still
4590 * allow the 'normal' nice value to be set - but as expected
4591 * it wont have any effect on scheduling until the task is
4592 * SCHED_FIFO/SCHED_RR:
4594 if (task_has_rt_policy(p
)) {
4595 p
->static_prio
= NICE_TO_PRIO(nice
);
4598 on_rq
= p
->se
.on_rq
;
4600 dequeue_task(rq
, p
, 0);
4604 p
->static_prio
= NICE_TO_PRIO(nice
);
4607 p
->prio
= effective_prio(p
);
4608 delta
= p
->prio
- old_prio
;
4611 enqueue_task(rq
, p
, 0);
4614 * If the task increased its priority or is running and
4615 * lowered its priority, then reschedule its CPU:
4617 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4618 resched_task(rq
->curr
);
4621 task_rq_unlock(rq
, &flags
);
4623 EXPORT_SYMBOL(set_user_nice
);
4626 * can_nice - check if a task can reduce its nice value
4630 int can_nice(const struct task_struct
*p
, const int nice
)
4632 /* convert nice value [19,-20] to rlimit style value [1,40] */
4633 int nice_rlim
= 20 - nice
;
4635 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4636 capable(CAP_SYS_NICE
));
4639 #ifdef __ARCH_WANT_SYS_NICE
4642 * sys_nice - change the priority of the current process.
4643 * @increment: priority increment
4645 * sys_setpriority is a more generic, but much slower function that
4646 * does similar things.
4648 asmlinkage
long sys_nice(int increment
)
4653 * Setpriority might change our priority at the same moment.
4654 * We don't have to worry. Conceptually one call occurs first
4655 * and we have a single winner.
4657 if (increment
< -40)
4662 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4668 if (increment
< 0 && !can_nice(current
, nice
))
4671 retval
= security_task_setnice(current
, nice
);
4675 set_user_nice(current
, nice
);
4682 * task_prio - return the priority value of a given task.
4683 * @p: the task in question.
4685 * This is the priority value as seen by users in /proc.
4686 * RT tasks are offset by -200. Normal tasks are centered
4687 * around 0, value goes from -16 to +15.
4689 int task_prio(const struct task_struct
*p
)
4691 return p
->prio
- MAX_RT_PRIO
;
4695 * task_nice - return the nice value of a given task.
4696 * @p: the task in question.
4698 int task_nice(const struct task_struct
*p
)
4700 return TASK_NICE(p
);
4702 EXPORT_SYMBOL(task_nice
);
4705 * idle_cpu - is a given cpu idle currently?
4706 * @cpu: the processor in question.
4708 int idle_cpu(int cpu
)
4710 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4714 * idle_task - return the idle task for a given cpu.
4715 * @cpu: the processor in question.
4717 struct task_struct
*idle_task(int cpu
)
4719 return cpu_rq(cpu
)->idle
;
4723 * find_process_by_pid - find a process with a matching PID value.
4724 * @pid: the pid in question.
4726 static struct task_struct
*find_process_by_pid(pid_t pid
)
4728 return pid
? find_task_by_vpid(pid
) : current
;
4731 /* Actually do priority change: must hold rq lock. */
4733 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4735 BUG_ON(p
->se
.on_rq
);
4738 switch (p
->policy
) {
4742 p
->sched_class
= &fair_sched_class
;
4746 p
->sched_class
= &rt_sched_class
;
4750 p
->rt_priority
= prio
;
4751 p
->normal_prio
= normal_prio(p
);
4752 /* we are holding p->pi_lock already */
4753 p
->prio
= rt_mutex_getprio(p
);
4758 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4759 * @p: the task in question.
4760 * @policy: new policy.
4761 * @param: structure containing the new RT priority.
4763 * NOTE that the task may be already dead.
4765 int sched_setscheduler(struct task_struct
*p
, int policy
,
4766 struct sched_param
*param
)
4768 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4769 unsigned long flags
;
4770 const struct sched_class
*prev_class
= p
->sched_class
;
4773 /* may grab non-irq protected spin_locks */
4774 BUG_ON(in_interrupt());
4776 /* double check policy once rq lock held */
4778 policy
= oldpolicy
= p
->policy
;
4779 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4780 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4781 policy
!= SCHED_IDLE
)
4784 * Valid priorities for SCHED_FIFO and SCHED_RR are
4785 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4786 * SCHED_BATCH and SCHED_IDLE is 0.
4788 if (param
->sched_priority
< 0 ||
4789 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4790 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4792 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4796 * Allow unprivileged RT tasks to decrease priority:
4798 if (!capable(CAP_SYS_NICE
)) {
4799 if (rt_policy(policy
)) {
4800 unsigned long rlim_rtprio
;
4802 if (!lock_task_sighand(p
, &flags
))
4804 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4805 unlock_task_sighand(p
, &flags
);
4807 /* can't set/change the rt policy */
4808 if (policy
!= p
->policy
&& !rlim_rtprio
)
4811 /* can't increase priority */
4812 if (param
->sched_priority
> p
->rt_priority
&&
4813 param
->sched_priority
> rlim_rtprio
)
4817 * Like positive nice levels, dont allow tasks to
4818 * move out of SCHED_IDLE either:
4820 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4823 /* can't change other user's priorities */
4824 if ((current
->euid
!= p
->euid
) &&
4825 (current
->euid
!= p
->uid
))
4829 #ifdef CONFIG_RT_GROUP_SCHED
4831 * Do not allow realtime tasks into groups that have no runtime
4834 if (rt_policy(policy
) && task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
4838 retval
= security_task_setscheduler(p
, policy
, param
);
4842 * make sure no PI-waiters arrive (or leave) while we are
4843 * changing the priority of the task:
4845 spin_lock_irqsave(&p
->pi_lock
, flags
);
4847 * To be able to change p->policy safely, the apropriate
4848 * runqueue lock must be held.
4850 rq
= __task_rq_lock(p
);
4851 /* recheck policy now with rq lock held */
4852 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4853 policy
= oldpolicy
= -1;
4854 __task_rq_unlock(rq
);
4855 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4858 update_rq_clock(rq
);
4859 on_rq
= p
->se
.on_rq
;
4860 running
= task_current(rq
, p
);
4862 deactivate_task(rq
, p
, 0);
4864 p
->sched_class
->put_prev_task(rq
, p
);
4867 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4870 p
->sched_class
->set_curr_task(rq
);
4872 activate_task(rq
, p
, 0);
4874 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4876 __task_rq_unlock(rq
);
4877 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4879 rt_mutex_adjust_pi(p
);
4883 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4886 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4888 struct sched_param lparam
;
4889 struct task_struct
*p
;
4892 if (!param
|| pid
< 0)
4894 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4899 p
= find_process_by_pid(pid
);
4901 retval
= sched_setscheduler(p
, policy
, &lparam
);
4908 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4909 * @pid: the pid in question.
4910 * @policy: new policy.
4911 * @param: structure containing the new RT priority.
4914 sys_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4916 /* negative values for policy are not valid */
4920 return do_sched_setscheduler(pid
, policy
, param
);
4924 * sys_sched_setparam - set/change the RT priority of a thread
4925 * @pid: the pid in question.
4926 * @param: structure containing the new RT priority.
4928 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4930 return do_sched_setscheduler(pid
, -1, param
);
4934 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4935 * @pid: the pid in question.
4937 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4939 struct task_struct
*p
;
4946 read_lock(&tasklist_lock
);
4947 p
= find_process_by_pid(pid
);
4949 retval
= security_task_getscheduler(p
);
4953 read_unlock(&tasklist_lock
);
4958 * sys_sched_getscheduler - get the RT priority of a thread
4959 * @pid: the pid in question.
4960 * @param: structure containing the RT priority.
4962 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4964 struct sched_param lp
;
4965 struct task_struct
*p
;
4968 if (!param
|| pid
< 0)
4971 read_lock(&tasklist_lock
);
4972 p
= find_process_by_pid(pid
);
4977 retval
= security_task_getscheduler(p
);
4981 lp
.sched_priority
= p
->rt_priority
;
4982 read_unlock(&tasklist_lock
);
4985 * This one might sleep, we cannot do it with a spinlock held ...
4987 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4992 read_unlock(&tasklist_lock
);
4996 long sched_setaffinity(pid_t pid
, const cpumask_t
*in_mask
)
4998 cpumask_t cpus_allowed
;
4999 cpumask_t new_mask
= *in_mask
;
5000 struct task_struct
*p
;
5004 read_lock(&tasklist_lock
);
5006 p
= find_process_by_pid(pid
);
5008 read_unlock(&tasklist_lock
);
5014 * It is not safe to call set_cpus_allowed with the
5015 * tasklist_lock held. We will bump the task_struct's
5016 * usage count and then drop tasklist_lock.
5019 read_unlock(&tasklist_lock
);
5022 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
5023 !capable(CAP_SYS_NICE
))
5026 retval
= security_task_setscheduler(p
, 0, NULL
);
5030 cpuset_cpus_allowed(p
, &cpus_allowed
);
5031 cpus_and(new_mask
, new_mask
, cpus_allowed
);
5033 retval
= set_cpus_allowed_ptr(p
, &new_mask
);
5036 cpuset_cpus_allowed(p
, &cpus_allowed
);
5037 if (!cpus_subset(new_mask
, cpus_allowed
)) {
5039 * We must have raced with a concurrent cpuset
5040 * update. Just reset the cpus_allowed to the
5041 * cpuset's cpus_allowed
5043 new_mask
= cpus_allowed
;
5053 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5054 cpumask_t
*new_mask
)
5056 if (len
< sizeof(cpumask_t
)) {
5057 memset(new_mask
, 0, sizeof(cpumask_t
));
5058 } else if (len
> sizeof(cpumask_t
)) {
5059 len
= sizeof(cpumask_t
);
5061 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
5065 * sys_sched_setaffinity - set the cpu affinity of a process
5066 * @pid: pid of the process
5067 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5068 * @user_mask_ptr: user-space pointer to the new cpu mask
5070 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
5071 unsigned long __user
*user_mask_ptr
)
5076 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
5080 return sched_setaffinity(pid
, &new_mask
);
5083 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
5085 struct task_struct
*p
;
5089 read_lock(&tasklist_lock
);
5092 p
= find_process_by_pid(pid
);
5096 retval
= security_task_getscheduler(p
);
5100 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
5103 read_unlock(&tasklist_lock
);
5110 * sys_sched_getaffinity - get the cpu affinity of a process
5111 * @pid: pid of the process
5112 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5113 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5115 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
5116 unsigned long __user
*user_mask_ptr
)
5121 if (len
< sizeof(cpumask_t
))
5124 ret
= sched_getaffinity(pid
, &mask
);
5128 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
5131 return sizeof(cpumask_t
);
5135 * sys_sched_yield - yield the current processor to other threads.
5137 * This function yields the current CPU to other tasks. If there are no
5138 * other threads running on this CPU then this function will return.
5140 asmlinkage
long sys_sched_yield(void)
5142 struct rq
*rq
= this_rq_lock();
5144 schedstat_inc(rq
, yld_count
);
5145 current
->sched_class
->yield_task(rq
);
5148 * Since we are going to call schedule() anyway, there's
5149 * no need to preempt or enable interrupts:
5151 __release(rq
->lock
);
5152 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5153 _raw_spin_unlock(&rq
->lock
);
5154 preempt_enable_no_resched();
5161 static void __cond_resched(void)
5163 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5164 __might_sleep(__FILE__
, __LINE__
);
5167 * The BKS might be reacquired before we have dropped
5168 * PREEMPT_ACTIVE, which could trigger a second
5169 * cond_resched() call.
5172 add_preempt_count(PREEMPT_ACTIVE
);
5174 sub_preempt_count(PREEMPT_ACTIVE
);
5175 } while (need_resched());
5178 int __sched
_cond_resched(void)
5180 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
5181 system_state
== SYSTEM_RUNNING
) {
5187 EXPORT_SYMBOL(_cond_resched
);
5190 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5191 * call schedule, and on return reacquire the lock.
5193 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5194 * operations here to prevent schedule() from being called twice (once via
5195 * spin_unlock(), once by hand).
5197 int cond_resched_lock(spinlock_t
*lock
)
5199 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
5202 if (spin_needbreak(lock
) || resched
) {
5204 if (resched
&& need_resched())
5213 EXPORT_SYMBOL(cond_resched_lock
);
5215 int __sched
cond_resched_softirq(void)
5217 BUG_ON(!in_softirq());
5219 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
5227 EXPORT_SYMBOL(cond_resched_softirq
);
5230 * yield - yield the current processor to other threads.
5232 * This is a shortcut for kernel-space yielding - it marks the
5233 * thread runnable and calls sys_sched_yield().
5235 void __sched
yield(void)
5237 set_current_state(TASK_RUNNING
);
5240 EXPORT_SYMBOL(yield
);
5243 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5244 * that process accounting knows that this is a task in IO wait state.
5246 * But don't do that if it is a deliberate, throttling IO wait (this task
5247 * has set its backing_dev_info: the queue against which it should throttle)
5249 void __sched
io_schedule(void)
5251 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5253 delayacct_blkio_start();
5254 atomic_inc(&rq
->nr_iowait
);
5256 atomic_dec(&rq
->nr_iowait
);
5257 delayacct_blkio_end();
5259 EXPORT_SYMBOL(io_schedule
);
5261 long __sched
io_schedule_timeout(long timeout
)
5263 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5266 delayacct_blkio_start();
5267 atomic_inc(&rq
->nr_iowait
);
5268 ret
= schedule_timeout(timeout
);
5269 atomic_dec(&rq
->nr_iowait
);
5270 delayacct_blkio_end();
5275 * sys_sched_get_priority_max - return maximum RT priority.
5276 * @policy: scheduling class.
5278 * this syscall returns the maximum rt_priority that can be used
5279 * by a given scheduling class.
5281 asmlinkage
long sys_sched_get_priority_max(int policy
)
5288 ret
= MAX_USER_RT_PRIO
-1;
5300 * sys_sched_get_priority_min - return minimum RT priority.
5301 * @policy: scheduling class.
5303 * this syscall returns the minimum rt_priority that can be used
5304 * by a given scheduling class.
5306 asmlinkage
long sys_sched_get_priority_min(int policy
)
5324 * sys_sched_rr_get_interval - return the default timeslice of a process.
5325 * @pid: pid of the process.
5326 * @interval: userspace pointer to the timeslice value.
5328 * this syscall writes the default timeslice value of a given process
5329 * into the user-space timespec buffer. A value of '0' means infinity.
5332 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
5334 struct task_struct
*p
;
5335 unsigned int time_slice
;
5343 read_lock(&tasklist_lock
);
5344 p
= find_process_by_pid(pid
);
5348 retval
= security_task_getscheduler(p
);
5353 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5354 * tasks that are on an otherwise idle runqueue:
5357 if (p
->policy
== SCHED_RR
) {
5358 time_slice
= DEF_TIMESLICE
;
5359 } else if (p
->policy
!= SCHED_FIFO
) {
5360 struct sched_entity
*se
= &p
->se
;
5361 unsigned long flags
;
5364 rq
= task_rq_lock(p
, &flags
);
5365 if (rq
->cfs
.load
.weight
)
5366 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
5367 task_rq_unlock(rq
, &flags
);
5369 read_unlock(&tasklist_lock
);
5370 jiffies_to_timespec(time_slice
, &t
);
5371 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5375 read_unlock(&tasklist_lock
);
5379 static const char stat_nam
[] = "RSDTtZX";
5381 void sched_show_task(struct task_struct
*p
)
5383 unsigned long free
= 0;
5386 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5387 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5388 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5389 #if BITS_PER_LONG == 32
5390 if (state
== TASK_RUNNING
)
5391 printk(KERN_CONT
" running ");
5393 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5395 if (state
== TASK_RUNNING
)
5396 printk(KERN_CONT
" running task ");
5398 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5400 #ifdef CONFIG_DEBUG_STACK_USAGE
5402 unsigned long *n
= end_of_stack(p
);
5405 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
5408 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
5409 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
5411 show_stack(p
, NULL
);
5414 void show_state_filter(unsigned long state_filter
)
5416 struct task_struct
*g
, *p
;
5418 #if BITS_PER_LONG == 32
5420 " task PC stack pid father\n");
5423 " task PC stack pid father\n");
5425 read_lock(&tasklist_lock
);
5426 do_each_thread(g
, p
) {
5428 * reset the NMI-timeout, listing all files on a slow
5429 * console might take alot of time:
5431 touch_nmi_watchdog();
5432 if (!state_filter
|| (p
->state
& state_filter
))
5434 } while_each_thread(g
, p
);
5436 touch_all_softlockup_watchdogs();
5438 #ifdef CONFIG_SCHED_DEBUG
5439 sysrq_sched_debug_show();
5441 read_unlock(&tasklist_lock
);
5443 * Only show locks if all tasks are dumped:
5445 if (state_filter
== -1)
5446 debug_show_all_locks();
5449 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5451 idle
->sched_class
= &idle_sched_class
;
5455 * init_idle - set up an idle thread for a given CPU
5456 * @idle: task in question
5457 * @cpu: cpu the idle task belongs to
5459 * NOTE: this function does not set the idle thread's NEED_RESCHED
5460 * flag, to make booting more robust.
5462 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5464 struct rq
*rq
= cpu_rq(cpu
);
5465 unsigned long flags
;
5468 idle
->se
.exec_start
= sched_clock();
5470 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
5471 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
5472 __set_task_cpu(idle
, cpu
);
5474 spin_lock_irqsave(&rq
->lock
, flags
);
5475 rq
->curr
= rq
->idle
= idle
;
5476 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5479 spin_unlock_irqrestore(&rq
->lock
, flags
);
5481 /* Set the preempt count _outside_ the spinlocks! */
5482 #if defined(CONFIG_PREEMPT)
5483 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5485 task_thread_info(idle
)->preempt_count
= 0;
5488 * The idle tasks have their own, simple scheduling class:
5490 idle
->sched_class
= &idle_sched_class
;
5494 * In a system that switches off the HZ timer nohz_cpu_mask
5495 * indicates which cpus entered this state. This is used
5496 * in the rcu update to wait only for active cpus. For system
5497 * which do not switch off the HZ timer nohz_cpu_mask should
5498 * always be CPU_MASK_NONE.
5500 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
5503 * Increase the granularity value when there are more CPUs,
5504 * because with more CPUs the 'effective latency' as visible
5505 * to users decreases. But the relationship is not linear,
5506 * so pick a second-best guess by going with the log2 of the
5509 * This idea comes from the SD scheduler of Con Kolivas:
5511 static inline void sched_init_granularity(void)
5513 unsigned int factor
= 1 + ilog2(num_online_cpus());
5514 const unsigned long limit
= 200000000;
5516 sysctl_sched_min_granularity
*= factor
;
5517 if (sysctl_sched_min_granularity
> limit
)
5518 sysctl_sched_min_granularity
= limit
;
5520 sysctl_sched_latency
*= factor
;
5521 if (sysctl_sched_latency
> limit
)
5522 sysctl_sched_latency
= limit
;
5524 sysctl_sched_wakeup_granularity
*= factor
;
5529 * This is how migration works:
5531 * 1) we queue a struct migration_req structure in the source CPU's
5532 * runqueue and wake up that CPU's migration thread.
5533 * 2) we down() the locked semaphore => thread blocks.
5534 * 3) migration thread wakes up (implicitly it forces the migrated
5535 * thread off the CPU)
5536 * 4) it gets the migration request and checks whether the migrated
5537 * task is still in the wrong runqueue.
5538 * 5) if it's in the wrong runqueue then the migration thread removes
5539 * it and puts it into the right queue.
5540 * 6) migration thread up()s the semaphore.
5541 * 7) we wake up and the migration is done.
5545 * Change a given task's CPU affinity. Migrate the thread to a
5546 * proper CPU and schedule it away if the CPU it's executing on
5547 * is removed from the allowed bitmask.
5549 * NOTE: the caller must have a valid reference to the task, the
5550 * task must not exit() & deallocate itself prematurely. The
5551 * call is not atomic; no spinlocks may be held.
5553 int set_cpus_allowed_ptr(struct task_struct
*p
, const cpumask_t
*new_mask
)
5555 struct migration_req req
;
5556 unsigned long flags
;
5560 rq
= task_rq_lock(p
, &flags
);
5561 if (!cpus_intersects(*new_mask
, cpu_online_map
)) {
5566 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
5567 !cpus_equal(p
->cpus_allowed
, *new_mask
))) {
5572 if (p
->sched_class
->set_cpus_allowed
)
5573 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
5575 p
->cpus_allowed
= *new_mask
;
5576 p
->rt
.nr_cpus_allowed
= cpus_weight(*new_mask
);
5579 /* Can the task run on the task's current CPU? If so, we're done */
5580 if (cpu_isset(task_cpu(p
), *new_mask
))
5583 if (migrate_task(p
, any_online_cpu(*new_mask
), &req
)) {
5584 /* Need help from migration thread: drop lock and wait. */
5585 task_rq_unlock(rq
, &flags
);
5586 wake_up_process(rq
->migration_thread
);
5587 wait_for_completion(&req
.done
);
5588 tlb_migrate_finish(p
->mm
);
5592 task_rq_unlock(rq
, &flags
);
5596 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5599 * Move (not current) task off this cpu, onto dest cpu. We're doing
5600 * this because either it can't run here any more (set_cpus_allowed()
5601 * away from this CPU, or CPU going down), or because we're
5602 * attempting to rebalance this task on exec (sched_exec).
5604 * So we race with normal scheduler movements, but that's OK, as long
5605 * as the task is no longer on this CPU.
5607 * Returns non-zero if task was successfully migrated.
5609 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5611 struct rq
*rq_dest
, *rq_src
;
5614 if (unlikely(cpu_is_offline(dest_cpu
)))
5617 rq_src
= cpu_rq(src_cpu
);
5618 rq_dest
= cpu_rq(dest_cpu
);
5620 double_rq_lock(rq_src
, rq_dest
);
5621 /* Already moved. */
5622 if (task_cpu(p
) != src_cpu
)
5624 /* Affinity changed (again). */
5625 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
5628 on_rq
= p
->se
.on_rq
;
5630 deactivate_task(rq_src
, p
, 0);
5632 set_task_cpu(p
, dest_cpu
);
5634 activate_task(rq_dest
, p
, 0);
5635 check_preempt_curr(rq_dest
, p
);
5639 double_rq_unlock(rq_src
, rq_dest
);
5644 * migration_thread - this is a highprio system thread that performs
5645 * thread migration by bumping thread off CPU then 'pushing' onto
5648 static int migration_thread(void *data
)
5650 int cpu
= (long)data
;
5654 BUG_ON(rq
->migration_thread
!= current
);
5656 set_current_state(TASK_INTERRUPTIBLE
);
5657 while (!kthread_should_stop()) {
5658 struct migration_req
*req
;
5659 struct list_head
*head
;
5661 spin_lock_irq(&rq
->lock
);
5663 if (cpu_is_offline(cpu
)) {
5664 spin_unlock_irq(&rq
->lock
);
5668 if (rq
->active_balance
) {
5669 active_load_balance(rq
, cpu
);
5670 rq
->active_balance
= 0;
5673 head
= &rq
->migration_queue
;
5675 if (list_empty(head
)) {
5676 spin_unlock_irq(&rq
->lock
);
5678 set_current_state(TASK_INTERRUPTIBLE
);
5681 req
= list_entry(head
->next
, struct migration_req
, list
);
5682 list_del_init(head
->next
);
5684 spin_unlock(&rq
->lock
);
5685 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5688 complete(&req
->done
);
5690 __set_current_state(TASK_RUNNING
);
5694 /* Wait for kthread_stop */
5695 set_current_state(TASK_INTERRUPTIBLE
);
5696 while (!kthread_should_stop()) {
5698 set_current_state(TASK_INTERRUPTIBLE
);
5700 __set_current_state(TASK_RUNNING
);
5704 #ifdef CONFIG_HOTPLUG_CPU
5706 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5710 local_irq_disable();
5711 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
5717 * Figure out where task on dead CPU should go, use force if necessary.
5718 * NOTE: interrupts should be disabled by the caller
5720 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5722 unsigned long flags
;
5729 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5730 cpus_and(mask
, mask
, p
->cpus_allowed
);
5731 dest_cpu
= any_online_cpu(mask
);
5733 /* On any allowed CPU? */
5734 if (dest_cpu
>= nr_cpu_ids
)
5735 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5737 /* No more Mr. Nice Guy. */
5738 if (dest_cpu
>= nr_cpu_ids
) {
5739 cpumask_t cpus_allowed
;
5741 cpuset_cpus_allowed_locked(p
, &cpus_allowed
);
5743 * Try to stay on the same cpuset, where the
5744 * current cpuset may be a subset of all cpus.
5745 * The cpuset_cpus_allowed_locked() variant of
5746 * cpuset_cpus_allowed() will not block. It must be
5747 * called within calls to cpuset_lock/cpuset_unlock.
5749 rq
= task_rq_lock(p
, &flags
);
5750 p
->cpus_allowed
= cpus_allowed
;
5751 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5752 task_rq_unlock(rq
, &flags
);
5755 * Don't tell them about moving exiting tasks or
5756 * kernel threads (both mm NULL), since they never
5759 if (p
->mm
&& printk_ratelimit()) {
5760 printk(KERN_INFO
"process %d (%s) no "
5761 "longer affine to cpu%d\n",
5762 task_pid_nr(p
), p
->comm
, dead_cpu
);
5765 } while (!__migrate_task_irq(p
, dead_cpu
, dest_cpu
));
5769 * While a dead CPU has no uninterruptible tasks queued at this point,
5770 * it might still have a nonzero ->nr_uninterruptible counter, because
5771 * for performance reasons the counter is not stricly tracking tasks to
5772 * their home CPUs. So we just add the counter to another CPU's counter,
5773 * to keep the global sum constant after CPU-down:
5775 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5777 struct rq
*rq_dest
= cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR
));
5778 unsigned long flags
;
5780 local_irq_save(flags
);
5781 double_rq_lock(rq_src
, rq_dest
);
5782 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5783 rq_src
->nr_uninterruptible
= 0;
5784 double_rq_unlock(rq_src
, rq_dest
);
5785 local_irq_restore(flags
);
5788 /* Run through task list and migrate tasks from the dead cpu. */
5789 static void migrate_live_tasks(int src_cpu
)
5791 struct task_struct
*p
, *t
;
5793 read_lock(&tasklist_lock
);
5795 do_each_thread(t
, p
) {
5799 if (task_cpu(p
) == src_cpu
)
5800 move_task_off_dead_cpu(src_cpu
, p
);
5801 } while_each_thread(t
, p
);
5803 read_unlock(&tasklist_lock
);
5807 * Schedules idle task to be the next runnable task on current CPU.
5808 * It does so by boosting its priority to highest possible.
5809 * Used by CPU offline code.
5811 void sched_idle_next(void)
5813 int this_cpu
= smp_processor_id();
5814 struct rq
*rq
= cpu_rq(this_cpu
);
5815 struct task_struct
*p
= rq
->idle
;
5816 unsigned long flags
;
5818 /* cpu has to be offline */
5819 BUG_ON(cpu_online(this_cpu
));
5822 * Strictly not necessary since rest of the CPUs are stopped by now
5823 * and interrupts disabled on the current cpu.
5825 spin_lock_irqsave(&rq
->lock
, flags
);
5827 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5829 update_rq_clock(rq
);
5830 activate_task(rq
, p
, 0);
5832 spin_unlock_irqrestore(&rq
->lock
, flags
);
5836 * Ensures that the idle task is using init_mm right before its cpu goes
5839 void idle_task_exit(void)
5841 struct mm_struct
*mm
= current
->active_mm
;
5843 BUG_ON(cpu_online(smp_processor_id()));
5846 switch_mm(mm
, &init_mm
, current
);
5850 /* called under rq->lock with disabled interrupts */
5851 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5853 struct rq
*rq
= cpu_rq(dead_cpu
);
5855 /* Must be exiting, otherwise would be on tasklist. */
5856 BUG_ON(!p
->exit_state
);
5858 /* Cannot have done final schedule yet: would have vanished. */
5859 BUG_ON(p
->state
== TASK_DEAD
);
5864 * Drop lock around migration; if someone else moves it,
5865 * that's OK. No task can be added to this CPU, so iteration is
5868 spin_unlock_irq(&rq
->lock
);
5869 move_task_off_dead_cpu(dead_cpu
, p
);
5870 spin_lock_irq(&rq
->lock
);
5875 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5876 static void migrate_dead_tasks(unsigned int dead_cpu
)
5878 struct rq
*rq
= cpu_rq(dead_cpu
);
5879 struct task_struct
*next
;
5882 if (!rq
->nr_running
)
5884 update_rq_clock(rq
);
5885 next
= pick_next_task(rq
, rq
->curr
);
5888 migrate_dead(dead_cpu
, next
);
5892 #endif /* CONFIG_HOTPLUG_CPU */
5894 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5896 static struct ctl_table sd_ctl_dir
[] = {
5898 .procname
= "sched_domain",
5904 static struct ctl_table sd_ctl_root
[] = {
5906 .ctl_name
= CTL_KERN
,
5907 .procname
= "kernel",
5909 .child
= sd_ctl_dir
,
5914 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5916 struct ctl_table
*entry
=
5917 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5922 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5924 struct ctl_table
*entry
;
5927 * In the intermediate directories, both the child directory and
5928 * procname are dynamically allocated and could fail but the mode
5929 * will always be set. In the lowest directory the names are
5930 * static strings and all have proc handlers.
5932 for (entry
= *tablep
; entry
->mode
; entry
++) {
5934 sd_free_ctl_entry(&entry
->child
);
5935 if (entry
->proc_handler
== NULL
)
5936 kfree(entry
->procname
);
5944 set_table_entry(struct ctl_table
*entry
,
5945 const char *procname
, void *data
, int maxlen
,
5946 mode_t mode
, proc_handler
*proc_handler
)
5948 entry
->procname
= procname
;
5950 entry
->maxlen
= maxlen
;
5952 entry
->proc_handler
= proc_handler
;
5955 static struct ctl_table
*
5956 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5958 struct ctl_table
*table
= sd_alloc_ctl_entry(12);
5963 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5964 sizeof(long), 0644, proc_doulongvec_minmax
);
5965 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5966 sizeof(long), 0644, proc_doulongvec_minmax
);
5967 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5968 sizeof(int), 0644, proc_dointvec_minmax
);
5969 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5970 sizeof(int), 0644, proc_dointvec_minmax
);
5971 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5972 sizeof(int), 0644, proc_dointvec_minmax
);
5973 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5974 sizeof(int), 0644, proc_dointvec_minmax
);
5975 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5976 sizeof(int), 0644, proc_dointvec_minmax
);
5977 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5978 sizeof(int), 0644, proc_dointvec_minmax
);
5979 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5980 sizeof(int), 0644, proc_dointvec_minmax
);
5981 set_table_entry(&table
[9], "cache_nice_tries",
5982 &sd
->cache_nice_tries
,
5983 sizeof(int), 0644, proc_dointvec_minmax
);
5984 set_table_entry(&table
[10], "flags", &sd
->flags
,
5985 sizeof(int), 0644, proc_dointvec_minmax
);
5986 /* &table[11] is terminator */
5991 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5993 struct ctl_table
*entry
, *table
;
5994 struct sched_domain
*sd
;
5995 int domain_num
= 0, i
;
5998 for_each_domain(cpu
, sd
)
6000 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
6005 for_each_domain(cpu
, sd
) {
6006 snprintf(buf
, 32, "domain%d", i
);
6007 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6009 entry
->child
= sd_alloc_ctl_domain_table(sd
);
6016 static struct ctl_table_header
*sd_sysctl_header
;
6017 static void register_sched_domain_sysctl(void)
6019 int i
, cpu_num
= num_online_cpus();
6020 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
6023 WARN_ON(sd_ctl_dir
[0].child
);
6024 sd_ctl_dir
[0].child
= entry
;
6029 for_each_online_cpu(i
) {
6030 snprintf(buf
, 32, "cpu%d", i
);
6031 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6033 entry
->child
= sd_alloc_ctl_cpu_table(i
);
6037 WARN_ON(sd_sysctl_header
);
6038 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
6041 /* may be called multiple times per register */
6042 static void unregister_sched_domain_sysctl(void)
6044 if (sd_sysctl_header
)
6045 unregister_sysctl_table(sd_sysctl_header
);
6046 sd_sysctl_header
= NULL
;
6047 if (sd_ctl_dir
[0].child
)
6048 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
6051 static void register_sched_domain_sysctl(void)
6054 static void unregister_sched_domain_sysctl(void)
6059 static void set_rq_online(struct rq
*rq
)
6062 const struct sched_class
*class;
6064 cpu_set(rq
->cpu
, rq
->rd
->online
);
6067 for_each_class(class) {
6068 if (class->rq_online
)
6069 class->rq_online(rq
);
6074 static void set_rq_offline(struct rq
*rq
)
6077 const struct sched_class
*class;
6079 for_each_class(class) {
6080 if (class->rq_offline
)
6081 class->rq_offline(rq
);
6084 cpu_clear(rq
->cpu
, rq
->rd
->online
);
6090 * migration_call - callback that gets triggered when a CPU is added.
6091 * Here we can start up the necessary migration thread for the new CPU.
6093 static int __cpuinit
6094 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
6096 struct task_struct
*p
;
6097 int cpu
= (long)hcpu
;
6098 unsigned long flags
;
6103 case CPU_UP_PREPARE
:
6104 case CPU_UP_PREPARE_FROZEN
:
6105 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
6108 kthread_bind(p
, cpu
);
6109 /* Must be high prio: stop_machine expects to yield to it. */
6110 rq
= task_rq_lock(p
, &flags
);
6111 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6112 task_rq_unlock(rq
, &flags
);
6113 cpu_rq(cpu
)->migration_thread
= p
;
6117 case CPU_ONLINE_FROZEN
:
6118 /* Strictly unnecessary, as first user will wake it. */
6119 wake_up_process(cpu_rq(cpu
)->migration_thread
);
6121 /* Update our root-domain */
6123 spin_lock_irqsave(&rq
->lock
, flags
);
6125 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6129 spin_unlock_irqrestore(&rq
->lock
, flags
);
6132 #ifdef CONFIG_HOTPLUG_CPU
6133 case CPU_UP_CANCELED
:
6134 case CPU_UP_CANCELED_FROZEN
:
6135 if (!cpu_rq(cpu
)->migration_thread
)
6137 /* Unbind it from offline cpu so it can run. Fall thru. */
6138 kthread_bind(cpu_rq(cpu
)->migration_thread
,
6139 any_online_cpu(cpu_online_map
));
6140 kthread_stop(cpu_rq(cpu
)->migration_thread
);
6141 cpu_rq(cpu
)->migration_thread
= NULL
;
6145 case CPU_DEAD_FROZEN
:
6146 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6147 migrate_live_tasks(cpu
);
6149 kthread_stop(rq
->migration_thread
);
6150 rq
->migration_thread
= NULL
;
6151 /* Idle task back to normal (off runqueue, low prio) */
6152 spin_lock_irq(&rq
->lock
);
6153 update_rq_clock(rq
);
6154 deactivate_task(rq
, rq
->idle
, 0);
6155 rq
->idle
->static_prio
= MAX_PRIO
;
6156 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
6157 rq
->idle
->sched_class
= &idle_sched_class
;
6158 migrate_dead_tasks(cpu
);
6159 spin_unlock_irq(&rq
->lock
);
6161 migrate_nr_uninterruptible(rq
);
6162 BUG_ON(rq
->nr_running
!= 0);
6165 * No need to migrate the tasks: it was best-effort if
6166 * they didn't take sched_hotcpu_mutex. Just wake up
6169 spin_lock_irq(&rq
->lock
);
6170 while (!list_empty(&rq
->migration_queue
)) {
6171 struct migration_req
*req
;
6173 req
= list_entry(rq
->migration_queue
.next
,
6174 struct migration_req
, list
);
6175 list_del_init(&req
->list
);
6176 complete(&req
->done
);
6178 spin_unlock_irq(&rq
->lock
);
6182 case CPU_DYING_FROZEN
:
6183 /* Update our root-domain */
6185 spin_lock_irqsave(&rq
->lock
, flags
);
6187 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6190 spin_unlock_irqrestore(&rq
->lock
, flags
);
6197 /* Register at highest priority so that task migration (migrate_all_tasks)
6198 * happens before everything else.
6200 static struct notifier_block __cpuinitdata migration_notifier
= {
6201 .notifier_call
= migration_call
,
6205 void __init
migration_init(void)
6207 void *cpu
= (void *)(long)smp_processor_id();
6210 /* Start one for the boot CPU: */
6211 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6212 BUG_ON(err
== NOTIFY_BAD
);
6213 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6214 register_cpu_notifier(&migration_notifier
);
6220 #ifdef CONFIG_SCHED_DEBUG
6222 static inline const char *sd_level_to_string(enum sched_domain_level lvl
)
6235 case SD_LV_ALLNODES
:
6244 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6245 cpumask_t
*groupmask
)
6247 struct sched_group
*group
= sd
->groups
;
6250 cpulist_scnprintf(str
, sizeof(str
), sd
->span
);
6251 cpus_clear(*groupmask
);
6253 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6255 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6256 printk("does not load-balance\n");
6258 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6263 printk(KERN_CONT
"span %s level %s\n",
6264 str
, sd_level_to_string(sd
->level
));
6266 if (!cpu_isset(cpu
, sd
->span
)) {
6267 printk(KERN_ERR
"ERROR: domain->span does not contain "
6270 if (!cpu_isset(cpu
, group
->cpumask
)) {
6271 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6275 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6279 printk(KERN_ERR
"ERROR: group is NULL\n");
6283 if (!group
->__cpu_power
) {
6284 printk(KERN_CONT
"\n");
6285 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6290 if (!cpus_weight(group
->cpumask
)) {
6291 printk(KERN_CONT
"\n");
6292 printk(KERN_ERR
"ERROR: empty group\n");
6296 if (cpus_intersects(*groupmask
, group
->cpumask
)) {
6297 printk(KERN_CONT
"\n");
6298 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6302 cpus_or(*groupmask
, *groupmask
, group
->cpumask
);
6304 cpulist_scnprintf(str
, sizeof(str
), group
->cpumask
);
6305 printk(KERN_CONT
" %s", str
);
6307 group
= group
->next
;
6308 } while (group
!= sd
->groups
);
6309 printk(KERN_CONT
"\n");
6311 if (!cpus_equal(sd
->span
, *groupmask
))
6312 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6314 if (sd
->parent
&& !cpus_subset(*groupmask
, sd
->parent
->span
))
6315 printk(KERN_ERR
"ERROR: parent span is not a superset "
6316 "of domain->span\n");
6320 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6322 cpumask_t
*groupmask
;
6326 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6330 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6332 groupmask
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
6334 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6339 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6348 #else /* !CONFIG_SCHED_DEBUG */
6349 # define sched_domain_debug(sd, cpu) do { } while (0)
6350 #endif /* CONFIG_SCHED_DEBUG */
6352 static int sd_degenerate(struct sched_domain
*sd
)
6354 if (cpus_weight(sd
->span
) == 1)
6357 /* Following flags need at least 2 groups */
6358 if (sd
->flags
& (SD_LOAD_BALANCE
|
6359 SD_BALANCE_NEWIDLE
|
6363 SD_SHARE_PKG_RESOURCES
)) {
6364 if (sd
->groups
!= sd
->groups
->next
)
6368 /* Following flags don't use groups */
6369 if (sd
->flags
& (SD_WAKE_IDLE
|
6378 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6380 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6382 if (sd_degenerate(parent
))
6385 if (!cpus_equal(sd
->span
, parent
->span
))
6388 /* Does parent contain flags not in child? */
6389 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6390 if (cflags
& SD_WAKE_AFFINE
)
6391 pflags
&= ~SD_WAKE_BALANCE
;
6392 /* Flags needing groups don't count if only 1 group in parent */
6393 if (parent
->groups
== parent
->groups
->next
) {
6394 pflags
&= ~(SD_LOAD_BALANCE
|
6395 SD_BALANCE_NEWIDLE
|
6399 SD_SHARE_PKG_RESOURCES
);
6401 if (~cflags
& pflags
)
6407 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6409 unsigned long flags
;
6411 spin_lock_irqsave(&rq
->lock
, flags
);
6414 struct root_domain
*old_rd
= rq
->rd
;
6416 if (cpu_isset(rq
->cpu
, old_rd
->online
))
6419 cpu_clear(rq
->cpu
, old_rd
->span
);
6421 if (atomic_dec_and_test(&old_rd
->refcount
))
6425 atomic_inc(&rd
->refcount
);
6428 cpu_set(rq
->cpu
, rd
->span
);
6429 if (cpu_isset(rq
->cpu
, cpu_online_map
))
6432 spin_unlock_irqrestore(&rq
->lock
, flags
);
6435 static void init_rootdomain(struct root_domain
*rd
)
6437 memset(rd
, 0, sizeof(*rd
));
6439 cpus_clear(rd
->span
);
6440 cpus_clear(rd
->online
);
6442 cpupri_init(&rd
->cpupri
);
6445 static void init_defrootdomain(void)
6447 init_rootdomain(&def_root_domain
);
6448 atomic_set(&def_root_domain
.refcount
, 1);
6451 static struct root_domain
*alloc_rootdomain(void)
6453 struct root_domain
*rd
;
6455 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6459 init_rootdomain(rd
);
6465 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6466 * hold the hotplug lock.
6469 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6471 struct rq
*rq
= cpu_rq(cpu
);
6472 struct sched_domain
*tmp
;
6474 /* Remove the sched domains which do not contribute to scheduling. */
6475 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
6476 struct sched_domain
*parent
= tmp
->parent
;
6479 if (sd_parent_degenerate(tmp
, parent
)) {
6480 tmp
->parent
= parent
->parent
;
6482 parent
->parent
->child
= tmp
;
6486 if (sd
&& sd_degenerate(sd
)) {
6492 sched_domain_debug(sd
, cpu
);
6494 rq_attach_root(rq
, rd
);
6495 rcu_assign_pointer(rq
->sd
, sd
);
6498 /* cpus with isolated domains */
6499 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
6501 /* Setup the mask of cpus configured for isolated domains */
6502 static int __init
isolated_cpu_setup(char *str
)
6504 int ints
[NR_CPUS
], i
;
6506 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
6507 cpus_clear(cpu_isolated_map
);
6508 for (i
= 1; i
<= ints
[0]; i
++)
6509 if (ints
[i
] < NR_CPUS
)
6510 cpu_set(ints
[i
], cpu_isolated_map
);
6514 __setup("isolcpus=", isolated_cpu_setup
);
6517 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6518 * to a function which identifies what group(along with sched group) a CPU
6519 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6520 * (due to the fact that we keep track of groups covered with a cpumask_t).
6522 * init_sched_build_groups will build a circular linked list of the groups
6523 * covered by the given span, and will set each group's ->cpumask correctly,
6524 * and ->cpu_power to 0.
6527 init_sched_build_groups(const cpumask_t
*span
, const cpumask_t
*cpu_map
,
6528 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
6529 struct sched_group
**sg
,
6530 cpumask_t
*tmpmask
),
6531 cpumask_t
*covered
, cpumask_t
*tmpmask
)
6533 struct sched_group
*first
= NULL
, *last
= NULL
;
6536 cpus_clear(*covered
);
6538 for_each_cpu_mask(i
, *span
) {
6539 struct sched_group
*sg
;
6540 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6543 if (cpu_isset(i
, *covered
))
6546 cpus_clear(sg
->cpumask
);
6547 sg
->__cpu_power
= 0;
6549 for_each_cpu_mask(j
, *span
) {
6550 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6553 cpu_set(j
, *covered
);
6554 cpu_set(j
, sg
->cpumask
);
6565 #define SD_NODES_PER_DOMAIN 16
6570 * find_next_best_node - find the next node to include in a sched_domain
6571 * @node: node whose sched_domain we're building
6572 * @used_nodes: nodes already in the sched_domain
6574 * Find the next node to include in a given scheduling domain. Simply
6575 * finds the closest node not already in the @used_nodes map.
6577 * Should use nodemask_t.
6579 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
6581 int i
, n
, val
, min_val
, best_node
= 0;
6585 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6586 /* Start at @node */
6587 n
= (node
+ i
) % MAX_NUMNODES
;
6589 if (!nr_cpus_node(n
))
6592 /* Skip already used nodes */
6593 if (node_isset(n
, *used_nodes
))
6596 /* Simple min distance search */
6597 val
= node_distance(node
, n
);
6599 if (val
< min_val
) {
6605 node_set(best_node
, *used_nodes
);
6610 * sched_domain_node_span - get a cpumask for a node's sched_domain
6611 * @node: node whose cpumask we're constructing
6612 * @span: resulting cpumask
6614 * Given a node, construct a good cpumask for its sched_domain to span. It
6615 * should be one that prevents unnecessary balancing, but also spreads tasks
6618 static void sched_domain_node_span(int node
, cpumask_t
*span
)
6620 nodemask_t used_nodes
;
6621 node_to_cpumask_ptr(nodemask
, node
);
6625 nodes_clear(used_nodes
);
6627 cpus_or(*span
, *span
, *nodemask
);
6628 node_set(node
, used_nodes
);
6630 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6631 int next_node
= find_next_best_node(node
, &used_nodes
);
6633 node_to_cpumask_ptr_next(nodemask
, next_node
);
6634 cpus_or(*span
, *span
, *nodemask
);
6637 #endif /* CONFIG_NUMA */
6639 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6642 * SMT sched-domains:
6644 #ifdef CONFIG_SCHED_SMT
6645 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
6646 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
6649 cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
6653 *sg
= &per_cpu(sched_group_cpus
, cpu
);
6656 #endif /* CONFIG_SCHED_SMT */
6659 * multi-core sched-domains:
6661 #ifdef CONFIG_SCHED_MC
6662 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
6663 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
6664 #endif /* CONFIG_SCHED_MC */
6666 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6668 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
6673 *mask
= per_cpu(cpu_sibling_map
, cpu
);
6674 cpus_and(*mask
, *mask
, *cpu_map
);
6675 group
= first_cpu(*mask
);
6677 *sg
= &per_cpu(sched_group_core
, group
);
6680 #elif defined(CONFIG_SCHED_MC)
6682 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
6686 *sg
= &per_cpu(sched_group_core
, cpu
);
6691 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
6692 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
6695 cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
6699 #ifdef CONFIG_SCHED_MC
6700 *mask
= cpu_coregroup_map(cpu
);
6701 cpus_and(*mask
, *mask
, *cpu_map
);
6702 group
= first_cpu(*mask
);
6703 #elif defined(CONFIG_SCHED_SMT)
6704 *mask
= per_cpu(cpu_sibling_map
, cpu
);
6705 cpus_and(*mask
, *mask
, *cpu_map
);
6706 group
= first_cpu(*mask
);
6711 *sg
= &per_cpu(sched_group_phys
, group
);
6717 * The init_sched_build_groups can't handle what we want to do with node
6718 * groups, so roll our own. Now each node has its own list of groups which
6719 * gets dynamically allocated.
6721 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
6722 static struct sched_group
***sched_group_nodes_bycpu
;
6724 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
6725 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
6727 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
6728 struct sched_group
**sg
, cpumask_t
*nodemask
)
6732 *nodemask
= node_to_cpumask(cpu_to_node(cpu
));
6733 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
6734 group
= first_cpu(*nodemask
);
6737 *sg
= &per_cpu(sched_group_allnodes
, group
);
6741 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6743 struct sched_group
*sg
= group_head
;
6749 for_each_cpu_mask(j
, sg
->cpumask
) {
6750 struct sched_domain
*sd
;
6752 sd
= &per_cpu(phys_domains
, j
);
6753 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
6755 * Only add "power" once for each
6761 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
6764 } while (sg
!= group_head
);
6766 #endif /* CONFIG_NUMA */
6769 /* Free memory allocated for various sched_group structures */
6770 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
6774 for_each_cpu_mask(cpu
, *cpu_map
) {
6775 struct sched_group
**sched_group_nodes
6776 = sched_group_nodes_bycpu
[cpu
];
6778 if (!sched_group_nodes
)
6781 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6782 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6784 *nodemask
= node_to_cpumask(i
);
6785 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
6786 if (cpus_empty(*nodemask
))
6796 if (oldsg
!= sched_group_nodes
[i
])
6799 kfree(sched_group_nodes
);
6800 sched_group_nodes_bycpu
[cpu
] = NULL
;
6803 #else /* !CONFIG_NUMA */
6804 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
6807 #endif /* CONFIG_NUMA */
6810 * Initialize sched groups cpu_power.
6812 * cpu_power indicates the capacity of sched group, which is used while
6813 * distributing the load between different sched groups in a sched domain.
6814 * Typically cpu_power for all the groups in a sched domain will be same unless
6815 * there are asymmetries in the topology. If there are asymmetries, group
6816 * having more cpu_power will pickup more load compared to the group having
6819 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6820 * the maximum number of tasks a group can handle in the presence of other idle
6821 * or lightly loaded groups in the same sched domain.
6823 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6825 struct sched_domain
*child
;
6826 struct sched_group
*group
;
6828 WARN_ON(!sd
|| !sd
->groups
);
6830 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
6835 sd
->groups
->__cpu_power
= 0;
6838 * For perf policy, if the groups in child domain share resources
6839 * (for example cores sharing some portions of the cache hierarchy
6840 * or SMT), then set this domain groups cpu_power such that each group
6841 * can handle only one task, when there are other idle groups in the
6842 * same sched domain.
6844 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
6846 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
6847 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
6852 * add cpu_power of each child group to this groups cpu_power
6854 group
= child
->groups
;
6856 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
6857 group
= group
->next
;
6858 } while (group
!= child
->groups
);
6862 * Initializers for schedule domains
6863 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6866 #define SD_INIT(sd, type) sd_init_##type(sd)
6867 #define SD_INIT_FUNC(type) \
6868 static noinline void sd_init_##type(struct sched_domain *sd) \
6870 memset(sd, 0, sizeof(*sd)); \
6871 *sd = SD_##type##_INIT; \
6872 sd->level = SD_LV_##type; \
6877 SD_INIT_FUNC(ALLNODES
)
6880 #ifdef CONFIG_SCHED_SMT
6881 SD_INIT_FUNC(SIBLING
)
6883 #ifdef CONFIG_SCHED_MC
6888 * To minimize stack usage kmalloc room for cpumasks and share the
6889 * space as the usage in build_sched_domains() dictates. Used only
6890 * if the amount of space is significant.
6893 cpumask_t tmpmask
; /* make this one first */
6896 cpumask_t this_sibling_map
;
6897 cpumask_t this_core_map
;
6899 cpumask_t send_covered
;
6902 cpumask_t domainspan
;
6904 cpumask_t notcovered
;
6909 #define SCHED_CPUMASK_ALLOC 1
6910 #define SCHED_CPUMASK_FREE(v) kfree(v)
6911 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
6913 #define SCHED_CPUMASK_ALLOC 0
6914 #define SCHED_CPUMASK_FREE(v)
6915 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
6918 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
6919 ((unsigned long)(a) + offsetof(struct allmasks, v))
6921 static int default_relax_domain_level
= -1;
6923 static int __init
setup_relax_domain_level(char *str
)
6925 default_relax_domain_level
= simple_strtoul(str
, NULL
, 0);
6928 __setup("relax_domain_level=", setup_relax_domain_level
);
6930 static void set_domain_attribute(struct sched_domain
*sd
,
6931 struct sched_domain_attr
*attr
)
6935 if (!attr
|| attr
->relax_domain_level
< 0) {
6936 if (default_relax_domain_level
< 0)
6939 request
= default_relax_domain_level
;
6941 request
= attr
->relax_domain_level
;
6942 if (request
< sd
->level
) {
6943 /* turn off idle balance on this domain */
6944 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
6946 /* turn on idle balance on this domain */
6947 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
6952 * Build sched domains for a given set of cpus and attach the sched domains
6953 * to the individual cpus
6955 static int __build_sched_domains(const cpumask_t
*cpu_map
,
6956 struct sched_domain_attr
*attr
)
6959 struct root_domain
*rd
;
6960 SCHED_CPUMASK_DECLARE(allmasks
);
6963 struct sched_group
**sched_group_nodes
= NULL
;
6964 int sd_allnodes
= 0;
6967 * Allocate the per-node list of sched groups
6969 sched_group_nodes
= kcalloc(MAX_NUMNODES
, sizeof(struct sched_group
*),
6971 if (!sched_group_nodes
) {
6972 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6977 rd
= alloc_rootdomain();
6979 printk(KERN_WARNING
"Cannot alloc root domain\n");
6981 kfree(sched_group_nodes
);
6986 #if SCHED_CPUMASK_ALLOC
6987 /* get space for all scratch cpumask variables */
6988 allmasks
= kmalloc(sizeof(*allmasks
), GFP_KERNEL
);
6990 printk(KERN_WARNING
"Cannot alloc cpumask array\n");
6993 kfree(sched_group_nodes
);
6998 tmpmask
= (cpumask_t
*)allmasks
;
7002 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
7006 * Set up domains for cpus specified by the cpu_map.
7008 for_each_cpu_mask(i
, *cpu_map
) {
7009 struct sched_domain
*sd
= NULL
, *p
;
7010 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7012 *nodemask
= node_to_cpumask(cpu_to_node(i
));
7013 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7016 if (cpus_weight(*cpu_map
) >
7017 SD_NODES_PER_DOMAIN
*cpus_weight(*nodemask
)) {
7018 sd
= &per_cpu(allnodes_domains
, i
);
7019 SD_INIT(sd
, ALLNODES
);
7020 set_domain_attribute(sd
, attr
);
7021 sd
->span
= *cpu_map
;
7022 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7028 sd
= &per_cpu(node_domains
, i
);
7030 set_domain_attribute(sd
, attr
);
7031 sched_domain_node_span(cpu_to_node(i
), &sd
->span
);
7035 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7039 sd
= &per_cpu(phys_domains
, i
);
7041 set_domain_attribute(sd
, attr
);
7042 sd
->span
= *nodemask
;
7046 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7048 #ifdef CONFIG_SCHED_MC
7050 sd
= &per_cpu(core_domains
, i
);
7052 set_domain_attribute(sd
, attr
);
7053 sd
->span
= cpu_coregroup_map(i
);
7054 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7057 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7060 #ifdef CONFIG_SCHED_SMT
7062 sd
= &per_cpu(cpu_domains
, i
);
7063 SD_INIT(sd
, SIBLING
);
7064 set_domain_attribute(sd
, attr
);
7065 sd
->span
= per_cpu(cpu_sibling_map
, i
);
7066 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7069 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7073 #ifdef CONFIG_SCHED_SMT
7074 /* Set up CPU (sibling) groups */
7075 for_each_cpu_mask(i
, *cpu_map
) {
7076 SCHED_CPUMASK_VAR(this_sibling_map
, allmasks
);
7077 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7079 *this_sibling_map
= per_cpu(cpu_sibling_map
, i
);
7080 cpus_and(*this_sibling_map
, *this_sibling_map
, *cpu_map
);
7081 if (i
!= first_cpu(*this_sibling_map
))
7084 init_sched_build_groups(this_sibling_map
, cpu_map
,
7086 send_covered
, tmpmask
);
7090 #ifdef CONFIG_SCHED_MC
7091 /* Set up multi-core groups */
7092 for_each_cpu_mask(i
, *cpu_map
) {
7093 SCHED_CPUMASK_VAR(this_core_map
, allmasks
);
7094 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7096 *this_core_map
= cpu_coregroup_map(i
);
7097 cpus_and(*this_core_map
, *this_core_map
, *cpu_map
);
7098 if (i
!= first_cpu(*this_core_map
))
7101 init_sched_build_groups(this_core_map
, cpu_map
,
7103 send_covered
, tmpmask
);
7107 /* Set up physical groups */
7108 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
7109 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7110 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7112 *nodemask
= node_to_cpumask(i
);
7113 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7114 if (cpus_empty(*nodemask
))
7117 init_sched_build_groups(nodemask
, cpu_map
,
7119 send_covered
, tmpmask
);
7123 /* Set up node groups */
7125 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7127 init_sched_build_groups(cpu_map
, cpu_map
,
7128 &cpu_to_allnodes_group
,
7129 send_covered
, tmpmask
);
7132 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
7133 /* Set up node groups */
7134 struct sched_group
*sg
, *prev
;
7135 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7136 SCHED_CPUMASK_VAR(domainspan
, allmasks
);
7137 SCHED_CPUMASK_VAR(covered
, allmasks
);
7140 *nodemask
= node_to_cpumask(i
);
7141 cpus_clear(*covered
);
7143 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7144 if (cpus_empty(*nodemask
)) {
7145 sched_group_nodes
[i
] = NULL
;
7149 sched_domain_node_span(i
, domainspan
);
7150 cpus_and(*domainspan
, *domainspan
, *cpu_map
);
7152 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
7154 printk(KERN_WARNING
"Can not alloc domain group for "
7158 sched_group_nodes
[i
] = sg
;
7159 for_each_cpu_mask(j
, *nodemask
) {
7160 struct sched_domain
*sd
;
7162 sd
= &per_cpu(node_domains
, j
);
7165 sg
->__cpu_power
= 0;
7166 sg
->cpumask
= *nodemask
;
7168 cpus_or(*covered
, *covered
, *nodemask
);
7171 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
7172 SCHED_CPUMASK_VAR(notcovered
, allmasks
);
7173 int n
= (i
+ j
) % MAX_NUMNODES
;
7174 node_to_cpumask_ptr(pnodemask
, n
);
7176 cpus_complement(*notcovered
, *covered
);
7177 cpus_and(*tmpmask
, *notcovered
, *cpu_map
);
7178 cpus_and(*tmpmask
, *tmpmask
, *domainspan
);
7179 if (cpus_empty(*tmpmask
))
7182 cpus_and(*tmpmask
, *tmpmask
, *pnodemask
);
7183 if (cpus_empty(*tmpmask
))
7186 sg
= kmalloc_node(sizeof(struct sched_group
),
7190 "Can not alloc domain group for node %d\n", j
);
7193 sg
->__cpu_power
= 0;
7194 sg
->cpumask
= *tmpmask
;
7195 sg
->next
= prev
->next
;
7196 cpus_or(*covered
, *covered
, *tmpmask
);
7203 /* Calculate CPU power for physical packages and nodes */
7204 #ifdef CONFIG_SCHED_SMT
7205 for_each_cpu_mask(i
, *cpu_map
) {
7206 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
7208 init_sched_groups_power(i
, sd
);
7211 #ifdef CONFIG_SCHED_MC
7212 for_each_cpu_mask(i
, *cpu_map
) {
7213 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
7215 init_sched_groups_power(i
, sd
);
7219 for_each_cpu_mask(i
, *cpu_map
) {
7220 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
7222 init_sched_groups_power(i
, sd
);
7226 for (i
= 0; i
< MAX_NUMNODES
; i
++)
7227 init_numa_sched_groups_power(sched_group_nodes
[i
]);
7230 struct sched_group
*sg
;
7232 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
,
7234 init_numa_sched_groups_power(sg
);
7238 /* Attach the domains */
7239 for_each_cpu_mask(i
, *cpu_map
) {
7240 struct sched_domain
*sd
;
7241 #ifdef CONFIG_SCHED_SMT
7242 sd
= &per_cpu(cpu_domains
, i
);
7243 #elif defined(CONFIG_SCHED_MC)
7244 sd
= &per_cpu(core_domains
, i
);
7246 sd
= &per_cpu(phys_domains
, i
);
7248 cpu_attach_domain(sd
, rd
, i
);
7251 SCHED_CPUMASK_FREE((void *)allmasks
);
7256 free_sched_groups(cpu_map
, tmpmask
);
7257 SCHED_CPUMASK_FREE((void *)allmasks
);
7262 static int build_sched_domains(const cpumask_t
*cpu_map
)
7264 return __build_sched_domains(cpu_map
, NULL
);
7267 static cpumask_t
*doms_cur
; /* current sched domains */
7268 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7269 static struct sched_domain_attr
*dattr_cur
;
7270 /* attribues of custom domains in 'doms_cur' */
7273 * Special case: If a kmalloc of a doms_cur partition (array of
7274 * cpumask_t) fails, then fallback to a single sched domain,
7275 * as determined by the single cpumask_t fallback_doms.
7277 static cpumask_t fallback_doms
;
7279 void __attribute__((weak
)) arch_update_cpu_topology(void)
7284 * Free current domain masks.
7285 * Called after all cpus are attached to NULL domain.
7287 static void free_sched_domains(void)
7290 if (doms_cur
!= &fallback_doms
)
7292 doms_cur
= &fallback_doms
;
7296 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7297 * For now this just excludes isolated cpus, but could be used to
7298 * exclude other special cases in the future.
7300 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
7304 arch_update_cpu_topology();
7306 doms_cur
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
7308 doms_cur
= &fallback_doms
;
7309 cpus_andnot(*doms_cur
, *cpu_map
, cpu_isolated_map
);
7311 err
= build_sched_domains(doms_cur
);
7312 register_sched_domain_sysctl();
7317 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
,
7320 free_sched_groups(cpu_map
, tmpmask
);
7324 * Detach sched domains from a group of cpus specified in cpu_map
7325 * These cpus will now be attached to the NULL domain
7327 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
7332 unregister_sched_domain_sysctl();
7334 for_each_cpu_mask(i
, *cpu_map
)
7335 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7336 synchronize_sched();
7337 arch_destroy_sched_domains(cpu_map
, &tmpmask
);
7340 /* handle null as "default" */
7341 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7342 struct sched_domain_attr
*new, int idx_new
)
7344 struct sched_domain_attr tmp
;
7351 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7352 new ? (new + idx_new
) : &tmp
,
7353 sizeof(struct sched_domain_attr
));
7357 * Partition sched domains as specified by the 'ndoms_new'
7358 * cpumasks in the array doms_new[] of cpumasks. This compares
7359 * doms_new[] to the current sched domain partitioning, doms_cur[].
7360 * It destroys each deleted domain and builds each new domain.
7362 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7363 * The masks don't intersect (don't overlap.) We should setup one
7364 * sched domain for each mask. CPUs not in any of the cpumasks will
7365 * not be load balanced. If the same cpumask appears both in the
7366 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7369 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7370 * ownership of it and will kfree it when done with it. If the caller
7371 * failed the kmalloc call, then it can pass in doms_new == NULL,
7372 * and partition_sched_domains() will fallback to the single partition
7375 * Call with hotplug lock held
7377 void partition_sched_domains(int ndoms_new
, cpumask_t
*doms_new
,
7378 struct sched_domain_attr
*dattr_new
)
7382 mutex_lock(&sched_domains_mutex
);
7384 /* always unregister in case we don't destroy any domains */
7385 unregister_sched_domain_sysctl();
7387 if (doms_new
== NULL
) {
7389 doms_new
= &fallback_doms
;
7390 cpus_andnot(doms_new
[0], cpu_online_map
, cpu_isolated_map
);
7394 /* Destroy deleted domains */
7395 for (i
= 0; i
< ndoms_cur
; i
++) {
7396 for (j
= 0; j
< ndoms_new
; j
++) {
7397 if (cpus_equal(doms_cur
[i
], doms_new
[j
])
7398 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7401 /* no match - a current sched domain not in new doms_new[] */
7402 detach_destroy_domains(doms_cur
+ i
);
7407 /* Build new domains */
7408 for (i
= 0; i
< ndoms_new
; i
++) {
7409 for (j
= 0; j
< ndoms_cur
; j
++) {
7410 if (cpus_equal(doms_new
[i
], doms_cur
[j
])
7411 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7414 /* no match - add a new doms_new */
7415 __build_sched_domains(doms_new
+ i
,
7416 dattr_new
? dattr_new
+ i
: NULL
);
7421 /* Remember the new sched domains */
7422 if (doms_cur
!= &fallback_doms
)
7424 kfree(dattr_cur
); /* kfree(NULL) is safe */
7425 doms_cur
= doms_new
;
7426 dattr_cur
= dattr_new
;
7427 ndoms_cur
= ndoms_new
;
7429 register_sched_domain_sysctl();
7431 mutex_unlock(&sched_domains_mutex
);
7434 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7435 int arch_reinit_sched_domains(void)
7440 mutex_lock(&sched_domains_mutex
);
7441 detach_destroy_domains(&cpu_online_map
);
7442 free_sched_domains();
7443 err
= arch_init_sched_domains(&cpu_online_map
);
7444 mutex_unlock(&sched_domains_mutex
);
7450 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7454 if (buf
[0] != '0' && buf
[0] != '1')
7458 sched_smt_power_savings
= (buf
[0] == '1');
7460 sched_mc_power_savings
= (buf
[0] == '1');
7462 ret
= arch_reinit_sched_domains();
7464 return ret
? ret
: count
;
7467 #ifdef CONFIG_SCHED_MC
7468 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
7470 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7472 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
7473 const char *buf
, size_t count
)
7475 return sched_power_savings_store(buf
, count
, 0);
7477 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
7478 sched_mc_power_savings_store
);
7481 #ifdef CONFIG_SCHED_SMT
7482 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
7484 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7486 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
7487 const char *buf
, size_t count
)
7489 return sched_power_savings_store(buf
, count
, 1);
7491 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
7492 sched_smt_power_savings_store
);
7495 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7499 #ifdef CONFIG_SCHED_SMT
7501 err
= sysfs_create_file(&cls
->kset
.kobj
,
7502 &attr_sched_smt_power_savings
.attr
);
7504 #ifdef CONFIG_SCHED_MC
7505 if (!err
&& mc_capable())
7506 err
= sysfs_create_file(&cls
->kset
.kobj
,
7507 &attr_sched_mc_power_savings
.attr
);
7511 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7514 * Force a reinitialization of the sched domains hierarchy. The domains
7515 * and groups cannot be updated in place without racing with the balancing
7516 * code, so we temporarily attach all running cpus to the NULL domain
7517 * which will prevent rebalancing while the sched domains are recalculated.
7519 static int update_sched_domains(struct notifier_block
*nfb
,
7520 unsigned long action
, void *hcpu
)
7522 int cpu
= (int)(long)hcpu
;
7525 case CPU_DOWN_PREPARE
:
7526 case CPU_DOWN_PREPARE_FROZEN
:
7527 disable_runtime(cpu_rq(cpu
));
7529 case CPU_UP_PREPARE
:
7530 case CPU_UP_PREPARE_FROZEN
:
7531 detach_destroy_domains(&cpu_online_map
);
7532 free_sched_domains();
7536 case CPU_DOWN_FAILED
:
7537 case CPU_DOWN_FAILED_FROZEN
:
7539 case CPU_ONLINE_FROZEN
:
7540 enable_runtime(cpu_rq(cpu
));
7542 case CPU_UP_CANCELED
:
7543 case CPU_UP_CANCELED_FROZEN
:
7545 case CPU_DEAD_FROZEN
:
7547 * Fall through and re-initialise the domains.
7554 #ifndef CONFIG_CPUSETS
7556 * Create default domain partitioning if cpusets are disabled.
7557 * Otherwise we let cpusets rebuild the domains based on the
7561 /* The hotplug lock is already held by cpu_up/cpu_down */
7562 arch_init_sched_domains(&cpu_online_map
);
7568 void __init
sched_init_smp(void)
7570 cpumask_t non_isolated_cpus
;
7572 #if defined(CONFIG_NUMA)
7573 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
7575 BUG_ON(sched_group_nodes_bycpu
== NULL
);
7578 mutex_lock(&sched_domains_mutex
);
7579 arch_init_sched_domains(&cpu_online_map
);
7580 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
7581 if (cpus_empty(non_isolated_cpus
))
7582 cpu_set(smp_processor_id(), non_isolated_cpus
);
7583 mutex_unlock(&sched_domains_mutex
);
7585 /* XXX: Theoretical race here - CPU may be hotplugged now */
7586 hotcpu_notifier(update_sched_domains
, 0);
7589 /* Move init over to a non-isolated CPU */
7590 if (set_cpus_allowed_ptr(current
, &non_isolated_cpus
) < 0)
7592 sched_init_granularity();
7595 void __init
sched_init_smp(void)
7597 sched_init_granularity();
7599 #endif /* CONFIG_SMP */
7601 int in_sched_functions(unsigned long addr
)
7603 return in_lock_functions(addr
) ||
7604 (addr
>= (unsigned long)__sched_text_start
7605 && addr
< (unsigned long)__sched_text_end
);
7608 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7610 cfs_rq
->tasks_timeline
= RB_ROOT
;
7611 INIT_LIST_HEAD(&cfs_rq
->tasks
);
7612 #ifdef CONFIG_FAIR_GROUP_SCHED
7615 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7618 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7620 struct rt_prio_array
*array
;
7623 array
= &rt_rq
->active
;
7624 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
7625 INIT_LIST_HEAD(array
->xqueue
+ i
);
7626 INIT_LIST_HEAD(array
->squeue
+ i
);
7627 __clear_bit(i
, array
->bitmap
);
7629 /* delimiter for bitsearch: */
7630 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
7632 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7633 rt_rq
->highest_prio
= MAX_RT_PRIO
;
7636 rt_rq
->rt_nr_migratory
= 0;
7637 rt_rq
->overloaded
= 0;
7641 rt_rq
->rt_throttled
= 0;
7642 rt_rq
->rt_runtime
= 0;
7643 spin_lock_init(&rt_rq
->rt_runtime_lock
);
7645 #ifdef CONFIG_RT_GROUP_SCHED
7646 rt_rq
->rt_nr_boosted
= 0;
7651 #ifdef CONFIG_FAIR_GROUP_SCHED
7652 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7653 struct sched_entity
*se
, int cpu
, int add
,
7654 struct sched_entity
*parent
)
7656 struct rq
*rq
= cpu_rq(cpu
);
7657 tg
->cfs_rq
[cpu
] = cfs_rq
;
7658 init_cfs_rq(cfs_rq
, rq
);
7661 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7664 /* se could be NULL for init_task_group */
7669 se
->cfs_rq
= &rq
->cfs
;
7671 se
->cfs_rq
= parent
->my_q
;
7674 se
->load
.weight
= tg
->shares
;
7675 se
->load
.inv_weight
= 0;
7676 se
->parent
= parent
;
7680 #ifdef CONFIG_RT_GROUP_SCHED
7681 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
7682 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
7683 struct sched_rt_entity
*parent
)
7685 struct rq
*rq
= cpu_rq(cpu
);
7687 tg
->rt_rq
[cpu
] = rt_rq
;
7688 init_rt_rq(rt_rq
, rq
);
7690 rt_rq
->rt_se
= rt_se
;
7691 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7693 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
7695 tg
->rt_se
[cpu
] = rt_se
;
7700 rt_se
->rt_rq
= &rq
->rt
;
7702 rt_se
->rt_rq
= parent
->my_q
;
7704 rt_se
->rt_rq
= &rq
->rt
;
7705 rt_se
->my_q
= rt_rq
;
7706 rt_se
->parent
= parent
;
7707 INIT_LIST_HEAD(&rt_se
->run_list
);
7711 void __init
sched_init(void)
7714 unsigned long alloc_size
= 0, ptr
;
7716 #ifdef CONFIG_FAIR_GROUP_SCHED
7717 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7719 #ifdef CONFIG_RT_GROUP_SCHED
7720 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7722 #ifdef CONFIG_USER_SCHED
7726 * As sched_init() is called before page_alloc is setup,
7727 * we use alloc_bootmem().
7730 ptr
= (unsigned long)alloc_bootmem(alloc_size
);
7732 #ifdef CONFIG_FAIR_GROUP_SCHED
7733 init_task_group
.se
= (struct sched_entity
**)ptr
;
7734 ptr
+= nr_cpu_ids
* sizeof(void **);
7736 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7737 ptr
+= nr_cpu_ids
* sizeof(void **);
7739 #ifdef CONFIG_USER_SCHED
7740 root_task_group
.se
= (struct sched_entity
**)ptr
;
7741 ptr
+= nr_cpu_ids
* sizeof(void **);
7743 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7744 ptr
+= nr_cpu_ids
* sizeof(void **);
7745 #endif /* CONFIG_USER_SCHED */
7746 #endif /* CONFIG_FAIR_GROUP_SCHED */
7747 #ifdef CONFIG_RT_GROUP_SCHED
7748 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7749 ptr
+= nr_cpu_ids
* sizeof(void **);
7751 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7752 ptr
+= nr_cpu_ids
* sizeof(void **);
7754 #ifdef CONFIG_USER_SCHED
7755 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7756 ptr
+= nr_cpu_ids
* sizeof(void **);
7758 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7759 ptr
+= nr_cpu_ids
* sizeof(void **);
7760 #endif /* CONFIG_USER_SCHED */
7761 #endif /* CONFIG_RT_GROUP_SCHED */
7765 init_defrootdomain();
7768 init_rt_bandwidth(&def_rt_bandwidth
,
7769 global_rt_period(), global_rt_runtime());
7771 #ifdef CONFIG_RT_GROUP_SCHED
7772 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
7773 global_rt_period(), global_rt_runtime());
7774 #ifdef CONFIG_USER_SCHED
7775 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
7776 global_rt_period(), RUNTIME_INF
);
7777 #endif /* CONFIG_USER_SCHED */
7778 #endif /* CONFIG_RT_GROUP_SCHED */
7780 #ifdef CONFIG_GROUP_SCHED
7781 list_add(&init_task_group
.list
, &task_groups
);
7782 INIT_LIST_HEAD(&init_task_group
.children
);
7784 #ifdef CONFIG_USER_SCHED
7785 INIT_LIST_HEAD(&root_task_group
.children
);
7786 init_task_group
.parent
= &root_task_group
;
7787 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
7788 #endif /* CONFIG_USER_SCHED */
7789 #endif /* CONFIG_GROUP_SCHED */
7791 for_each_possible_cpu(i
) {
7795 spin_lock_init(&rq
->lock
);
7796 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
7798 init_cfs_rq(&rq
->cfs
, rq
);
7799 init_rt_rq(&rq
->rt
, rq
);
7800 #ifdef CONFIG_FAIR_GROUP_SCHED
7801 init_task_group
.shares
= init_task_group_load
;
7802 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7803 #ifdef CONFIG_CGROUP_SCHED
7805 * How much cpu bandwidth does init_task_group get?
7807 * In case of task-groups formed thr' the cgroup filesystem, it
7808 * gets 100% of the cpu resources in the system. This overall
7809 * system cpu resource is divided among the tasks of
7810 * init_task_group and its child task-groups in a fair manner,
7811 * based on each entity's (task or task-group's) weight
7812 * (se->load.weight).
7814 * In other words, if init_task_group has 10 tasks of weight
7815 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7816 * then A0's share of the cpu resource is:
7818 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7820 * We achieve this by letting init_task_group's tasks sit
7821 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7823 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
7824 #elif defined CONFIG_USER_SCHED
7825 root_task_group
.shares
= NICE_0_LOAD
;
7826 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
7828 * In case of task-groups formed thr' the user id of tasks,
7829 * init_task_group represents tasks belonging to root user.
7830 * Hence it forms a sibling of all subsequent groups formed.
7831 * In this case, init_task_group gets only a fraction of overall
7832 * system cpu resource, based on the weight assigned to root
7833 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
7834 * by letting tasks of init_task_group sit in a separate cfs_rq
7835 * (init_cfs_rq) and having one entity represent this group of
7836 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
7838 init_tg_cfs_entry(&init_task_group
,
7839 &per_cpu(init_cfs_rq
, i
),
7840 &per_cpu(init_sched_entity
, i
), i
, 1,
7841 root_task_group
.se
[i
]);
7844 #endif /* CONFIG_FAIR_GROUP_SCHED */
7846 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7847 #ifdef CONFIG_RT_GROUP_SCHED
7848 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
7849 #ifdef CONFIG_CGROUP_SCHED
7850 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
7851 #elif defined CONFIG_USER_SCHED
7852 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
7853 init_tg_rt_entry(&init_task_group
,
7854 &per_cpu(init_rt_rq
, i
),
7855 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
7856 root_task_group
.rt_se
[i
]);
7860 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7861 rq
->cpu_load
[j
] = 0;
7865 rq
->active_balance
= 0;
7866 rq
->next_balance
= jiffies
;
7870 rq
->migration_thread
= NULL
;
7871 INIT_LIST_HEAD(&rq
->migration_queue
);
7872 rq_attach_root(rq
, &def_root_domain
);
7875 atomic_set(&rq
->nr_iowait
, 0);
7878 set_load_weight(&init_task
);
7880 #ifdef CONFIG_PREEMPT_NOTIFIERS
7881 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7885 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
7888 #ifdef CONFIG_RT_MUTEXES
7889 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
7893 * The boot idle thread does lazy MMU switching as well:
7895 atomic_inc(&init_mm
.mm_count
);
7896 enter_lazy_tlb(&init_mm
, current
);
7899 * Make us the idle thread. Technically, schedule() should not be
7900 * called from this thread, however somewhere below it might be,
7901 * but because we are the idle thread, we just pick up running again
7902 * when this runqueue becomes "idle".
7904 init_idle(current
, smp_processor_id());
7906 * During early bootup we pretend to be a normal task:
7908 current
->sched_class
= &fair_sched_class
;
7910 scheduler_running
= 1;
7913 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7914 void __might_sleep(char *file
, int line
)
7917 static unsigned long prev_jiffy
; /* ratelimiting */
7919 if ((in_atomic() || irqs_disabled()) &&
7920 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
7921 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7923 prev_jiffy
= jiffies
;
7924 printk(KERN_ERR
"BUG: sleeping function called from invalid"
7925 " context at %s:%d\n", file
, line
);
7926 printk("in_atomic():%d, irqs_disabled():%d\n",
7927 in_atomic(), irqs_disabled());
7928 debug_show_held_locks(current
);
7929 if (irqs_disabled())
7930 print_irqtrace_events(current
);
7935 EXPORT_SYMBOL(__might_sleep
);
7938 #ifdef CONFIG_MAGIC_SYSRQ
7939 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7943 update_rq_clock(rq
);
7944 on_rq
= p
->se
.on_rq
;
7946 deactivate_task(rq
, p
, 0);
7947 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7949 activate_task(rq
, p
, 0);
7950 resched_task(rq
->curr
);
7954 void normalize_rt_tasks(void)
7956 struct task_struct
*g
, *p
;
7957 unsigned long flags
;
7960 read_lock_irqsave(&tasklist_lock
, flags
);
7961 do_each_thread(g
, p
) {
7963 * Only normalize user tasks:
7968 p
->se
.exec_start
= 0;
7969 #ifdef CONFIG_SCHEDSTATS
7970 p
->se
.wait_start
= 0;
7971 p
->se
.sleep_start
= 0;
7972 p
->se
.block_start
= 0;
7977 * Renice negative nice level userspace
7980 if (TASK_NICE(p
) < 0 && p
->mm
)
7981 set_user_nice(p
, 0);
7985 spin_lock(&p
->pi_lock
);
7986 rq
= __task_rq_lock(p
);
7988 normalize_task(rq
, p
);
7990 __task_rq_unlock(rq
);
7991 spin_unlock(&p
->pi_lock
);
7992 } while_each_thread(g
, p
);
7994 read_unlock_irqrestore(&tasklist_lock
, flags
);
7997 #endif /* CONFIG_MAGIC_SYSRQ */
8001 * These functions are only useful for the IA64 MCA handling.
8003 * They can only be called when the whole system has been
8004 * stopped - every CPU needs to be quiescent, and no scheduling
8005 * activity can take place. Using them for anything else would
8006 * be a serious bug, and as a result, they aren't even visible
8007 * under any other configuration.
8011 * curr_task - return the current task for a given cpu.
8012 * @cpu: the processor in question.
8014 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8016 struct task_struct
*curr_task(int cpu
)
8018 return cpu_curr(cpu
);
8022 * set_curr_task - set the current task for a given cpu.
8023 * @cpu: the processor in question.
8024 * @p: the task pointer to set.
8026 * Description: This function must only be used when non-maskable interrupts
8027 * are serviced on a separate stack. It allows the architecture to switch the
8028 * notion of the current task on a cpu in a non-blocking manner. This function
8029 * must be called with all CPU's synchronized, and interrupts disabled, the
8030 * and caller must save the original value of the current task (see
8031 * curr_task() above) and restore that value before reenabling interrupts and
8032 * re-starting the system.
8034 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8036 void set_curr_task(int cpu
, struct task_struct
*p
)
8043 #ifdef CONFIG_FAIR_GROUP_SCHED
8044 static void free_fair_sched_group(struct task_group
*tg
)
8048 for_each_possible_cpu(i
) {
8050 kfree(tg
->cfs_rq
[i
]);
8060 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8062 struct cfs_rq
*cfs_rq
;
8063 struct sched_entity
*se
, *parent_se
;
8067 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8070 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8074 tg
->shares
= NICE_0_LOAD
;
8076 for_each_possible_cpu(i
) {
8079 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
),
8080 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8084 se
= kmalloc_node(sizeof(struct sched_entity
),
8085 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8089 parent_se
= parent
? parent
->se
[i
] : NULL
;
8090 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent_se
);
8099 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8101 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
8102 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
8105 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8107 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
8109 #else /* !CONFG_FAIR_GROUP_SCHED */
8110 static inline void free_fair_sched_group(struct task_group
*tg
)
8115 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8120 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8124 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8127 #endif /* CONFIG_FAIR_GROUP_SCHED */
8129 #ifdef CONFIG_RT_GROUP_SCHED
8130 static void free_rt_sched_group(struct task_group
*tg
)
8134 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8136 for_each_possible_cpu(i
) {
8138 kfree(tg
->rt_rq
[i
]);
8140 kfree(tg
->rt_se
[i
]);
8148 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8150 struct rt_rq
*rt_rq
;
8151 struct sched_rt_entity
*rt_se
, *parent_se
;
8155 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8158 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8162 init_rt_bandwidth(&tg
->rt_bandwidth
,
8163 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8165 for_each_possible_cpu(i
) {
8168 rt_rq
= kmalloc_node(sizeof(struct rt_rq
),
8169 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8173 rt_se
= kmalloc_node(sizeof(struct sched_rt_entity
),
8174 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8178 parent_se
= parent
? parent
->rt_se
[i
] : NULL
;
8179 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent_se
);
8188 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8190 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8191 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8194 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8196 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8198 #else /* !CONFIG_RT_GROUP_SCHED */
8199 static inline void free_rt_sched_group(struct task_group
*tg
)
8204 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8209 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8213 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8216 #endif /* CONFIG_RT_GROUP_SCHED */
8218 #ifdef CONFIG_GROUP_SCHED
8219 static void free_sched_group(struct task_group
*tg
)
8221 free_fair_sched_group(tg
);
8222 free_rt_sched_group(tg
);
8226 /* allocate runqueue etc for a new task group */
8227 struct task_group
*sched_create_group(struct task_group
*parent
)
8229 struct task_group
*tg
;
8230 unsigned long flags
;
8233 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8235 return ERR_PTR(-ENOMEM
);
8237 if (!alloc_fair_sched_group(tg
, parent
))
8240 if (!alloc_rt_sched_group(tg
, parent
))
8243 spin_lock_irqsave(&task_group_lock
, flags
);
8244 for_each_possible_cpu(i
) {
8245 register_fair_sched_group(tg
, i
);
8246 register_rt_sched_group(tg
, i
);
8248 list_add_rcu(&tg
->list
, &task_groups
);
8250 WARN_ON(!parent
); /* root should already exist */
8252 tg
->parent
= parent
;
8253 list_add_rcu(&tg
->siblings
, &parent
->children
);
8254 INIT_LIST_HEAD(&tg
->children
);
8255 spin_unlock_irqrestore(&task_group_lock
, flags
);
8260 free_sched_group(tg
);
8261 return ERR_PTR(-ENOMEM
);
8264 /* rcu callback to free various structures associated with a task group */
8265 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8267 /* now it should be safe to free those cfs_rqs */
8268 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8271 /* Destroy runqueue etc associated with a task group */
8272 void sched_destroy_group(struct task_group
*tg
)
8274 unsigned long flags
;
8277 spin_lock_irqsave(&task_group_lock
, flags
);
8278 for_each_possible_cpu(i
) {
8279 unregister_fair_sched_group(tg
, i
);
8280 unregister_rt_sched_group(tg
, i
);
8282 list_del_rcu(&tg
->list
);
8283 list_del_rcu(&tg
->siblings
);
8284 spin_unlock_irqrestore(&task_group_lock
, flags
);
8286 /* wait for possible concurrent references to cfs_rqs complete */
8287 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8290 /* change task's runqueue when it moves between groups.
8291 * The caller of this function should have put the task in its new group
8292 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8293 * reflect its new group.
8295 void sched_move_task(struct task_struct
*tsk
)
8298 unsigned long flags
;
8301 rq
= task_rq_lock(tsk
, &flags
);
8303 update_rq_clock(rq
);
8305 running
= task_current(rq
, tsk
);
8306 on_rq
= tsk
->se
.on_rq
;
8309 dequeue_task(rq
, tsk
, 0);
8310 if (unlikely(running
))
8311 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8313 set_task_rq(tsk
, task_cpu(tsk
));
8315 #ifdef CONFIG_FAIR_GROUP_SCHED
8316 if (tsk
->sched_class
->moved_group
)
8317 tsk
->sched_class
->moved_group(tsk
);
8320 if (unlikely(running
))
8321 tsk
->sched_class
->set_curr_task(rq
);
8323 enqueue_task(rq
, tsk
, 0);
8325 task_rq_unlock(rq
, &flags
);
8327 #endif /* CONFIG_GROUP_SCHED */
8329 #ifdef CONFIG_FAIR_GROUP_SCHED
8330 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8332 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8333 struct rq
*rq
= cfs_rq
->rq
;
8336 spin_lock_irq(&rq
->lock
);
8340 dequeue_entity(cfs_rq
, se
, 0);
8342 se
->load
.weight
= shares
;
8343 se
->load
.inv_weight
= 0;
8346 enqueue_entity(cfs_rq
, se
, 0);
8348 spin_unlock_irq(&rq
->lock
);
8351 static DEFINE_MUTEX(shares_mutex
);
8353 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8356 unsigned long flags
;
8359 * We can't change the weight of the root cgroup.
8364 if (shares
< MIN_SHARES
)
8365 shares
= MIN_SHARES
;
8366 else if (shares
> MAX_SHARES
)
8367 shares
= MAX_SHARES
;
8369 mutex_lock(&shares_mutex
);
8370 if (tg
->shares
== shares
)
8373 spin_lock_irqsave(&task_group_lock
, flags
);
8374 for_each_possible_cpu(i
)
8375 unregister_fair_sched_group(tg
, i
);
8376 list_del_rcu(&tg
->siblings
);
8377 spin_unlock_irqrestore(&task_group_lock
, flags
);
8379 /* wait for any ongoing reference to this group to finish */
8380 synchronize_sched();
8383 * Now we are free to modify the group's share on each cpu
8384 * w/o tripping rebalance_share or load_balance_fair.
8386 tg
->shares
= shares
;
8387 for_each_possible_cpu(i
)
8388 set_se_shares(tg
->se
[i
], shares
);
8391 * Enable load balance activity on this group, by inserting it back on
8392 * each cpu's rq->leaf_cfs_rq_list.
8394 spin_lock_irqsave(&task_group_lock
, flags
);
8395 for_each_possible_cpu(i
)
8396 register_fair_sched_group(tg
, i
);
8397 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
8398 spin_unlock_irqrestore(&task_group_lock
, flags
);
8400 mutex_unlock(&shares_mutex
);
8404 unsigned long sched_group_shares(struct task_group
*tg
)
8410 #ifdef CONFIG_RT_GROUP_SCHED
8412 * Ensure that the real time constraints are schedulable.
8414 static DEFINE_MUTEX(rt_constraints_mutex
);
8416 static unsigned long to_ratio(u64 period
, u64 runtime
)
8418 if (runtime
== RUNTIME_INF
)
8421 return div64_u64(runtime
<< 16, period
);
8424 #ifdef CONFIG_CGROUP_SCHED
8425 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8427 struct task_group
*tgi
, *parent
= tg
->parent
;
8428 unsigned long total
= 0;
8431 if (global_rt_period() < period
)
8434 return to_ratio(period
, runtime
) <
8435 to_ratio(global_rt_period(), global_rt_runtime());
8438 if (ktime_to_ns(parent
->rt_bandwidth
.rt_period
) < period
)
8442 list_for_each_entry_rcu(tgi
, &parent
->children
, siblings
) {
8446 total
+= to_ratio(ktime_to_ns(tgi
->rt_bandwidth
.rt_period
),
8447 tgi
->rt_bandwidth
.rt_runtime
);
8451 return total
+ to_ratio(period
, runtime
) <
8452 to_ratio(ktime_to_ns(parent
->rt_bandwidth
.rt_period
),
8453 parent
->rt_bandwidth
.rt_runtime
);
8455 #elif defined CONFIG_USER_SCHED
8456 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8458 struct task_group
*tgi
;
8459 unsigned long total
= 0;
8460 unsigned long global_ratio
=
8461 to_ratio(global_rt_period(), global_rt_runtime());
8464 list_for_each_entry_rcu(tgi
, &task_groups
, list
) {
8468 total
+= to_ratio(ktime_to_ns(tgi
->rt_bandwidth
.rt_period
),
8469 tgi
->rt_bandwidth
.rt_runtime
);
8473 return total
+ to_ratio(period
, runtime
) < global_ratio
;
8477 /* Must be called with tasklist_lock held */
8478 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8480 struct task_struct
*g
, *p
;
8481 do_each_thread(g
, p
) {
8482 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8484 } while_each_thread(g
, p
);
8488 static int tg_set_bandwidth(struct task_group
*tg
,
8489 u64 rt_period
, u64 rt_runtime
)
8493 mutex_lock(&rt_constraints_mutex
);
8494 read_lock(&tasklist_lock
);
8495 if (rt_runtime
== 0 && tg_has_rt_tasks(tg
)) {
8499 if (!__rt_schedulable(tg
, rt_period
, rt_runtime
)) {
8504 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8505 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8506 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8508 for_each_possible_cpu(i
) {
8509 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8511 spin_lock(&rt_rq
->rt_runtime_lock
);
8512 rt_rq
->rt_runtime
= rt_runtime
;
8513 spin_unlock(&rt_rq
->rt_runtime_lock
);
8515 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8517 read_unlock(&tasklist_lock
);
8518 mutex_unlock(&rt_constraints_mutex
);
8523 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8525 u64 rt_runtime
, rt_period
;
8527 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8528 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8529 if (rt_runtime_us
< 0)
8530 rt_runtime
= RUNTIME_INF
;
8532 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8535 long sched_group_rt_runtime(struct task_group
*tg
)
8539 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8542 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8543 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8544 return rt_runtime_us
;
8547 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8549 u64 rt_runtime
, rt_period
;
8551 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8552 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8554 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8557 long sched_group_rt_period(struct task_group
*tg
)
8561 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8562 do_div(rt_period_us
, NSEC_PER_USEC
);
8563 return rt_period_us
;
8566 static int sched_rt_global_constraints(void)
8570 mutex_lock(&rt_constraints_mutex
);
8571 if (!__rt_schedulable(NULL
, 1, 0))
8573 mutex_unlock(&rt_constraints_mutex
);
8577 #else /* !CONFIG_RT_GROUP_SCHED */
8578 static int sched_rt_global_constraints(void)
8580 unsigned long flags
;
8583 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8584 for_each_possible_cpu(i
) {
8585 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
8587 spin_lock(&rt_rq
->rt_runtime_lock
);
8588 rt_rq
->rt_runtime
= global_rt_runtime();
8589 spin_unlock(&rt_rq
->rt_runtime_lock
);
8591 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8595 #endif /* CONFIG_RT_GROUP_SCHED */
8597 int sched_rt_handler(struct ctl_table
*table
, int write
,
8598 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
8602 int old_period
, old_runtime
;
8603 static DEFINE_MUTEX(mutex
);
8606 old_period
= sysctl_sched_rt_period
;
8607 old_runtime
= sysctl_sched_rt_runtime
;
8609 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
8611 if (!ret
&& write
) {
8612 ret
= sched_rt_global_constraints();
8614 sysctl_sched_rt_period
= old_period
;
8615 sysctl_sched_rt_runtime
= old_runtime
;
8617 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8618 def_rt_bandwidth
.rt_period
=
8619 ns_to_ktime(global_rt_period());
8622 mutex_unlock(&mutex
);
8627 #ifdef CONFIG_CGROUP_SCHED
8629 /* return corresponding task_group object of a cgroup */
8630 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
8632 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
8633 struct task_group
, css
);
8636 static struct cgroup_subsys_state
*
8637 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8639 struct task_group
*tg
, *parent
;
8641 if (!cgrp
->parent
) {
8642 /* This is early initialization for the top cgroup */
8643 init_task_group
.css
.cgroup
= cgrp
;
8644 return &init_task_group
.css
;
8647 parent
= cgroup_tg(cgrp
->parent
);
8648 tg
= sched_create_group(parent
);
8650 return ERR_PTR(-ENOMEM
);
8652 /* Bind the cgroup to task_group object we just created */
8653 tg
->css
.cgroup
= cgrp
;
8659 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8661 struct task_group
*tg
= cgroup_tg(cgrp
);
8663 sched_destroy_group(tg
);
8667 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8668 struct task_struct
*tsk
)
8670 #ifdef CONFIG_RT_GROUP_SCHED
8671 /* Don't accept realtime tasks when there is no way for them to run */
8672 if (rt_task(tsk
) && cgroup_tg(cgrp
)->rt_bandwidth
.rt_runtime
== 0)
8675 /* We don't support RT-tasks being in separate groups */
8676 if (tsk
->sched_class
!= &fair_sched_class
)
8684 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8685 struct cgroup
*old_cont
, struct task_struct
*tsk
)
8687 sched_move_task(tsk
);
8690 #ifdef CONFIG_FAIR_GROUP_SCHED
8691 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
8694 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
8697 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
8699 struct task_group
*tg
= cgroup_tg(cgrp
);
8701 return (u64
) tg
->shares
;
8703 #endif /* CONFIG_FAIR_GROUP_SCHED */
8705 #ifdef CONFIG_RT_GROUP_SCHED
8706 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
8709 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
8712 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8714 return sched_group_rt_runtime(cgroup_tg(cgrp
));
8717 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
8720 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
8723 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
8725 return sched_group_rt_period(cgroup_tg(cgrp
));
8727 #endif /* CONFIG_RT_GROUP_SCHED */
8729 static struct cftype cpu_files
[] = {
8730 #ifdef CONFIG_FAIR_GROUP_SCHED
8733 .read_u64
= cpu_shares_read_u64
,
8734 .write_u64
= cpu_shares_write_u64
,
8737 #ifdef CONFIG_RT_GROUP_SCHED
8739 .name
= "rt_runtime_us",
8740 .read_s64
= cpu_rt_runtime_read
,
8741 .write_s64
= cpu_rt_runtime_write
,
8744 .name
= "rt_period_us",
8745 .read_u64
= cpu_rt_period_read_uint
,
8746 .write_u64
= cpu_rt_period_write_uint
,
8751 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8753 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
8756 struct cgroup_subsys cpu_cgroup_subsys
= {
8758 .create
= cpu_cgroup_create
,
8759 .destroy
= cpu_cgroup_destroy
,
8760 .can_attach
= cpu_cgroup_can_attach
,
8761 .attach
= cpu_cgroup_attach
,
8762 .populate
= cpu_cgroup_populate
,
8763 .subsys_id
= cpu_cgroup_subsys_id
,
8767 #endif /* CONFIG_CGROUP_SCHED */
8769 #ifdef CONFIG_CGROUP_CPUACCT
8772 * CPU accounting code for task groups.
8774 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8775 * (balbir@in.ibm.com).
8778 /* track cpu usage of a group of tasks */
8780 struct cgroup_subsys_state css
;
8781 /* cpuusage holds pointer to a u64-type object on every cpu */
8785 struct cgroup_subsys cpuacct_subsys
;
8787 /* return cpu accounting group corresponding to this container */
8788 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
8790 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
8791 struct cpuacct
, css
);
8794 /* return cpu accounting group to which this task belongs */
8795 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
8797 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
8798 struct cpuacct
, css
);
8801 /* create a new cpu accounting group */
8802 static struct cgroup_subsys_state
*cpuacct_create(
8803 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8805 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
8808 return ERR_PTR(-ENOMEM
);
8810 ca
->cpuusage
= alloc_percpu(u64
);
8811 if (!ca
->cpuusage
) {
8813 return ERR_PTR(-ENOMEM
);
8819 /* destroy an existing cpu accounting group */
8821 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8823 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8825 free_percpu(ca
->cpuusage
);
8829 /* return total cpu usage (in nanoseconds) of a group */
8830 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8832 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8833 u64 totalcpuusage
= 0;
8836 for_each_possible_cpu(i
) {
8837 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
8840 * Take rq->lock to make 64-bit addition safe on 32-bit
8843 spin_lock_irq(&cpu_rq(i
)->lock
);
8844 totalcpuusage
+= *cpuusage
;
8845 spin_unlock_irq(&cpu_rq(i
)->lock
);
8848 return totalcpuusage
;
8851 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
8854 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8863 for_each_possible_cpu(i
) {
8864 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
8866 spin_lock_irq(&cpu_rq(i
)->lock
);
8868 spin_unlock_irq(&cpu_rq(i
)->lock
);
8874 static struct cftype files
[] = {
8877 .read_u64
= cpuusage_read
,
8878 .write_u64
= cpuusage_write
,
8882 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8884 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
8888 * charge this task's execution time to its accounting group.
8890 * called with rq->lock held.
8892 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
8896 if (!cpuacct_subsys
.active
)
8901 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, task_cpu(tsk
));
8903 *cpuusage
+= cputime
;
8907 struct cgroup_subsys cpuacct_subsys
= {
8909 .create
= cpuacct_create
,
8910 .destroy
= cpuacct_destroy
,
8911 .populate
= cpuacct_populate
,
8912 .subsys_id
= cpuacct_subsys_id
,
8914 #endif /* CONFIG_CGROUP_CPUACCT */