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 queue
[MAX_RT_PRIO
];
159 struct rt_bandwidth
{
160 /* nests inside the rq lock: */
161 spinlock_t rt_runtime_lock
;
164 struct hrtimer rt_period_timer
;
167 static struct rt_bandwidth def_rt_bandwidth
;
169 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
171 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
173 struct rt_bandwidth
*rt_b
=
174 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
180 now
= hrtimer_cb_get_time(timer
);
181 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
186 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
189 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
193 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
195 rt_b
->rt_period
= ns_to_ktime(period
);
196 rt_b
->rt_runtime
= runtime
;
198 spin_lock_init(&rt_b
->rt_runtime_lock
);
200 hrtimer_init(&rt_b
->rt_period_timer
,
201 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
202 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
203 rt_b
->rt_period_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_NO_SOFTIRQ
;
206 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
210 if (rt_b
->rt_runtime
== RUNTIME_INF
)
213 if (hrtimer_active(&rt_b
->rt_period_timer
))
216 spin_lock(&rt_b
->rt_runtime_lock
);
218 if (hrtimer_active(&rt_b
->rt_period_timer
))
221 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
222 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
223 hrtimer_start(&rt_b
->rt_period_timer
,
224 rt_b
->rt_period_timer
.expires
,
227 spin_unlock(&rt_b
->rt_runtime_lock
);
230 #ifdef CONFIG_RT_GROUP_SCHED
231 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
233 hrtimer_cancel(&rt_b
->rt_period_timer
);
238 * sched_domains_mutex serializes calls to arch_init_sched_domains,
239 * detach_destroy_domains and partition_sched_domains.
241 static DEFINE_MUTEX(sched_domains_mutex
);
243 #ifdef CONFIG_GROUP_SCHED
245 #include <linux/cgroup.h>
249 static LIST_HEAD(task_groups
);
251 /* task group related information */
253 #ifdef CONFIG_CGROUP_SCHED
254 struct cgroup_subsys_state css
;
257 #ifdef CONFIG_FAIR_GROUP_SCHED
258 /* schedulable entities of this group on each cpu */
259 struct sched_entity
**se
;
260 /* runqueue "owned" by this group on each cpu */
261 struct cfs_rq
**cfs_rq
;
262 unsigned long shares
;
265 #ifdef CONFIG_RT_GROUP_SCHED
266 struct sched_rt_entity
**rt_se
;
267 struct rt_rq
**rt_rq
;
269 struct rt_bandwidth rt_bandwidth
;
273 struct list_head list
;
275 struct task_group
*parent
;
276 struct list_head siblings
;
277 struct list_head children
;
280 #ifdef CONFIG_USER_SCHED
284 * Every UID task group (including init_task_group aka UID-0) will
285 * be a child to this group.
287 struct task_group root_task_group
;
289 #ifdef CONFIG_FAIR_GROUP_SCHED
290 /* Default task group's sched entity on each cpu */
291 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
292 /* Default task group's cfs_rq on each cpu */
293 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
294 #endif /* CONFIG_FAIR_GROUP_SCHED */
296 #ifdef CONFIG_RT_GROUP_SCHED
297 static DEFINE_PER_CPU(struct sched_rt_entity
, init_sched_rt_entity
);
298 static DEFINE_PER_CPU(struct rt_rq
, init_rt_rq
) ____cacheline_aligned_in_smp
;
299 #endif /* CONFIG_RT_GROUP_SCHED */
300 #else /* !CONFIG_FAIR_GROUP_SCHED */
301 #define root_task_group init_task_group
302 #endif /* CONFIG_FAIR_GROUP_SCHED */
304 /* task_group_lock serializes add/remove of task groups and also changes to
305 * a task group's cpu shares.
307 static DEFINE_SPINLOCK(task_group_lock
);
309 #ifdef CONFIG_FAIR_GROUP_SCHED
310 #ifdef CONFIG_USER_SCHED
311 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
312 #else /* !CONFIG_USER_SCHED */
313 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
314 #endif /* CONFIG_USER_SCHED */
317 * A weight of 0 or 1 can cause arithmetics problems.
318 * A weight of a cfs_rq is the sum of weights of which entities
319 * are queued on this cfs_rq, so a weight of a entity should not be
320 * too large, so as the shares value of a task group.
321 * (The default weight is 1024 - so there's no practical
322 * limitation from this.)
325 #define MAX_SHARES (1UL << 18)
327 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
330 /* Default task group.
331 * Every task in system belong to this group at bootup.
333 struct task_group init_task_group
;
335 /* return group to which a task belongs */
336 static inline struct task_group
*task_group(struct task_struct
*p
)
338 struct task_group
*tg
;
340 #ifdef CONFIG_USER_SCHED
342 #elif defined(CONFIG_CGROUP_SCHED)
343 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
344 struct task_group
, css
);
346 tg
= &init_task_group
;
351 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
352 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
354 #ifdef CONFIG_FAIR_GROUP_SCHED
355 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
356 p
->se
.parent
= task_group(p
)->se
[cpu
];
359 #ifdef CONFIG_RT_GROUP_SCHED
360 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
361 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
367 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
369 #endif /* CONFIG_GROUP_SCHED */
371 /* CFS-related fields in a runqueue */
373 struct load_weight load
;
374 unsigned long nr_running
;
380 struct rb_root tasks_timeline
;
381 struct rb_node
*rb_leftmost
;
383 struct list_head tasks
;
384 struct list_head
*balance_iterator
;
387 * 'curr' points to currently running entity on this cfs_rq.
388 * It is set to NULL otherwise (i.e when none are currently running).
390 struct sched_entity
*curr
, *next
;
392 unsigned long nr_spread_over
;
394 #ifdef CONFIG_FAIR_GROUP_SCHED
395 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
398 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
399 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
400 * (like users, containers etc.)
402 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
403 * list is used during load balance.
405 struct list_head leaf_cfs_rq_list
;
406 struct task_group
*tg
; /* group that "owns" this runqueue */
410 * the part of load.weight contributed by tasks
412 unsigned long task_weight
;
415 * h_load = weight * f(tg)
417 * Where f(tg) is the recursive weight fraction assigned to
420 unsigned long h_load
;
423 * this cpu's part of tg->shares
425 unsigned long shares
;
430 /* Real-Time classes' related field in a runqueue: */
432 struct rt_prio_array active
;
433 unsigned long rt_nr_running
;
434 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
435 int highest_prio
; /* highest queued rt task prio */
438 unsigned long rt_nr_migratory
;
444 /* Nests inside the rq lock: */
445 spinlock_t rt_runtime_lock
;
447 #ifdef CONFIG_RT_GROUP_SCHED
448 unsigned long rt_nr_boosted
;
451 struct list_head leaf_rt_rq_list
;
452 struct task_group
*tg
;
453 struct sched_rt_entity
*rt_se
;
460 * We add the notion of a root-domain which will be used to define per-domain
461 * variables. Each exclusive cpuset essentially defines an island domain by
462 * fully partitioning the member cpus from any other cpuset. Whenever a new
463 * exclusive cpuset is created, we also create and attach a new root-domain
473 * The "RT overload" flag: it gets set if a CPU has more than
474 * one runnable RT task.
479 struct cpupri cpupri
;
484 * By default the system creates a single root-domain with all cpus as
485 * members (mimicking the global state we have today).
487 static struct root_domain def_root_domain
;
492 * This is the main, per-CPU runqueue data structure.
494 * Locking rule: those places that want to lock multiple runqueues
495 * (such as the load balancing or the thread migration code), lock
496 * acquire operations must be ordered by ascending &runqueue.
503 * nr_running and cpu_load should be in the same cacheline because
504 * remote CPUs use both these fields when doing load calculation.
506 unsigned long nr_running
;
507 #define CPU_LOAD_IDX_MAX 5
508 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
509 unsigned char idle_at_tick
;
511 unsigned long last_tick_seen
;
512 unsigned char in_nohz_recently
;
514 /* capture load from *all* tasks on this cpu: */
515 struct load_weight load
;
516 unsigned long nr_load_updates
;
522 #ifdef CONFIG_FAIR_GROUP_SCHED
523 /* list of leaf cfs_rq on this cpu: */
524 struct list_head leaf_cfs_rq_list
;
526 #ifdef CONFIG_RT_GROUP_SCHED
527 struct list_head leaf_rt_rq_list
;
531 * This is part of a global counter where only the total sum
532 * over all CPUs matters. A task can increase this counter on
533 * one CPU and if it got migrated afterwards it may decrease
534 * it on another CPU. Always updated under the runqueue lock:
536 unsigned long nr_uninterruptible
;
538 struct task_struct
*curr
, *idle
;
539 unsigned long next_balance
;
540 struct mm_struct
*prev_mm
;
547 struct root_domain
*rd
;
548 struct sched_domain
*sd
;
550 /* For active balancing */
553 /* cpu of this runqueue: */
557 struct task_struct
*migration_thread
;
558 struct list_head migration_queue
;
561 #ifdef CONFIG_SCHED_HRTICK
562 unsigned long hrtick_flags
;
563 ktime_t hrtick_expire
;
564 struct hrtimer hrtick_timer
;
567 #ifdef CONFIG_SCHEDSTATS
569 struct sched_info rq_sched_info
;
571 /* sys_sched_yield() stats */
572 unsigned int yld_exp_empty
;
573 unsigned int yld_act_empty
;
574 unsigned int yld_both_empty
;
575 unsigned int yld_count
;
577 /* schedule() stats */
578 unsigned int sched_switch
;
579 unsigned int sched_count
;
580 unsigned int sched_goidle
;
582 /* try_to_wake_up() stats */
583 unsigned int ttwu_count
;
584 unsigned int ttwu_local
;
587 unsigned int bkl_count
;
589 struct lock_class_key rq_lock_key
;
592 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
594 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
)
596 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
);
599 static inline int cpu_of(struct rq
*rq
)
609 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
610 * See detach_destroy_domains: synchronize_sched for details.
612 * The domain tree of any CPU may only be accessed from within
613 * preempt-disabled sections.
615 #define for_each_domain(cpu, __sd) \
616 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
618 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
619 #define this_rq() (&__get_cpu_var(runqueues))
620 #define task_rq(p) cpu_rq(task_cpu(p))
621 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
623 static inline void update_rq_clock(struct rq
*rq
)
625 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
629 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
631 #ifdef CONFIG_SCHED_DEBUG
632 # define const_debug __read_mostly
634 # define const_debug static const
638 * Debugging: various feature bits
641 #define SCHED_FEAT(name, enabled) \
642 __SCHED_FEAT_##name ,
645 #include "sched_features.h"
650 #define SCHED_FEAT(name, enabled) \
651 (1UL << __SCHED_FEAT_##name) * enabled |
653 const_debug
unsigned int sysctl_sched_features
=
654 #include "sched_features.h"
659 #ifdef CONFIG_SCHED_DEBUG
660 #define SCHED_FEAT(name, enabled) \
663 static __read_mostly
char *sched_feat_names
[] = {
664 #include "sched_features.h"
670 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
672 filp
->private_data
= inode
->i_private
;
677 sched_feat_read(struct file
*filp
, char __user
*ubuf
,
678 size_t cnt
, loff_t
*ppos
)
685 for (i
= 0; sched_feat_names
[i
]; i
++) {
686 len
+= strlen(sched_feat_names
[i
]);
690 buf
= kmalloc(len
+ 2, GFP_KERNEL
);
694 for (i
= 0; sched_feat_names
[i
]; i
++) {
695 if (sysctl_sched_features
& (1UL << i
))
696 r
+= sprintf(buf
+ r
, "%s ", sched_feat_names
[i
]);
698 r
+= sprintf(buf
+ r
, "NO_%s ", sched_feat_names
[i
]);
701 r
+= sprintf(buf
+ r
, "\n");
702 WARN_ON(r
>= len
+ 2);
704 r
= simple_read_from_buffer(ubuf
, cnt
, ppos
, buf
, r
);
712 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
713 size_t cnt
, loff_t
*ppos
)
723 if (copy_from_user(&buf
, ubuf
, cnt
))
728 if (strncmp(buf
, "NO_", 3) == 0) {
733 for (i
= 0; sched_feat_names
[i
]; i
++) {
734 int len
= strlen(sched_feat_names
[i
]);
736 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
738 sysctl_sched_features
&= ~(1UL << i
);
740 sysctl_sched_features
|= (1UL << i
);
745 if (!sched_feat_names
[i
])
753 static struct file_operations sched_feat_fops
= {
754 .open
= sched_feat_open
,
755 .read
= sched_feat_read
,
756 .write
= sched_feat_write
,
759 static __init
int sched_init_debug(void)
761 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
766 late_initcall(sched_init_debug
);
770 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
773 * Number of tasks to iterate in a single balance run.
774 * Limited because this is done with IRQs disabled.
776 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
779 * period over which we measure -rt task cpu usage in us.
782 unsigned int sysctl_sched_rt_period
= 1000000;
784 static __read_mostly
int scheduler_running
;
787 * part of the period that we allow rt tasks to run in us.
790 int sysctl_sched_rt_runtime
= 950000;
792 static inline u64
global_rt_period(void)
794 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
797 static inline u64
global_rt_runtime(void)
799 if (sysctl_sched_rt_period
< 0)
802 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
805 #ifndef prepare_arch_switch
806 # define prepare_arch_switch(next) do { } while (0)
808 #ifndef finish_arch_switch
809 # define finish_arch_switch(prev) do { } while (0)
812 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
814 return rq
->curr
== p
;
817 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
818 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
820 return task_current(rq
, p
);
823 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
827 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
829 #ifdef CONFIG_DEBUG_SPINLOCK
830 /* this is a valid case when another task releases the spinlock */
831 rq
->lock
.owner
= current
;
834 * If we are tracking spinlock dependencies then we have to
835 * fix up the runqueue lock - which gets 'carried over' from
838 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
840 spin_unlock_irq(&rq
->lock
);
843 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
844 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
849 return task_current(rq
, p
);
853 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
857 * We can optimise this out completely for !SMP, because the
858 * SMP rebalancing from interrupt is the only thing that cares
863 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
864 spin_unlock_irq(&rq
->lock
);
866 spin_unlock(&rq
->lock
);
870 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
874 * After ->oncpu is cleared, the task can be moved to a different CPU.
875 * We must ensure this doesn't happen until the switch is completely
881 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
885 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
888 * __task_rq_lock - lock the runqueue a given task resides on.
889 * Must be called interrupts disabled.
891 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
895 struct rq
*rq
= task_rq(p
);
896 spin_lock(&rq
->lock
);
897 if (likely(rq
== task_rq(p
)))
899 spin_unlock(&rq
->lock
);
904 * task_rq_lock - lock the runqueue a given task resides on and disable
905 * interrupts. Note the ordering: we can safely lookup the task_rq without
906 * explicitly disabling preemption.
908 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
914 local_irq_save(*flags
);
916 spin_lock(&rq
->lock
);
917 if (likely(rq
== task_rq(p
)))
919 spin_unlock_irqrestore(&rq
->lock
, *flags
);
923 static void __task_rq_unlock(struct rq
*rq
)
926 spin_unlock(&rq
->lock
);
929 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
932 spin_unlock_irqrestore(&rq
->lock
, *flags
);
936 * this_rq_lock - lock this runqueue and disable interrupts.
938 static struct rq
*this_rq_lock(void)
945 spin_lock(&rq
->lock
);
950 static void __resched_task(struct task_struct
*p
, int tif_bit
);
952 static inline void resched_task(struct task_struct
*p
)
954 __resched_task(p
, TIF_NEED_RESCHED
);
957 #ifdef CONFIG_SCHED_HRTICK
959 * Use HR-timers to deliver accurate preemption points.
961 * Its all a bit involved since we cannot program an hrt while holding the
962 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
965 * When we get rescheduled we reprogram the hrtick_timer outside of the
968 static inline void resched_hrt(struct task_struct
*p
)
970 __resched_task(p
, TIF_HRTICK_RESCHED
);
973 static inline void resched_rq(struct rq
*rq
)
977 spin_lock_irqsave(&rq
->lock
, flags
);
978 resched_task(rq
->curr
);
979 spin_unlock_irqrestore(&rq
->lock
, flags
);
983 HRTICK_SET
, /* re-programm hrtick_timer */
984 HRTICK_RESET
, /* not a new slice */
985 HRTICK_BLOCK
, /* stop hrtick operations */
990 * - enabled by features
991 * - hrtimer is actually high res
993 static inline int hrtick_enabled(struct rq
*rq
)
995 if (!sched_feat(HRTICK
))
997 if (unlikely(test_bit(HRTICK_BLOCK
, &rq
->hrtick_flags
)))
999 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1003 * Called to set the hrtick timer state.
1005 * called with rq->lock held and irqs disabled
1007 static void hrtick_start(struct rq
*rq
, u64 delay
, int reset
)
1009 assert_spin_locked(&rq
->lock
);
1012 * preempt at: now + delay
1015 ktime_add_ns(rq
->hrtick_timer
.base
->get_time(), delay
);
1017 * indicate we need to program the timer
1019 __set_bit(HRTICK_SET
, &rq
->hrtick_flags
);
1021 __set_bit(HRTICK_RESET
, &rq
->hrtick_flags
);
1024 * New slices are called from the schedule path and don't need a
1025 * forced reschedule.
1028 resched_hrt(rq
->curr
);
1031 static void hrtick_clear(struct rq
*rq
)
1033 if (hrtimer_active(&rq
->hrtick_timer
))
1034 hrtimer_cancel(&rq
->hrtick_timer
);
1038 * Update the timer from the possible pending state.
1040 static void hrtick_set(struct rq
*rq
)
1044 unsigned long flags
;
1046 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1048 spin_lock_irqsave(&rq
->lock
, flags
);
1049 set
= __test_and_clear_bit(HRTICK_SET
, &rq
->hrtick_flags
);
1050 reset
= __test_and_clear_bit(HRTICK_RESET
, &rq
->hrtick_flags
);
1051 time
= rq
->hrtick_expire
;
1052 clear_thread_flag(TIF_HRTICK_RESCHED
);
1053 spin_unlock_irqrestore(&rq
->lock
, flags
);
1056 hrtimer_start(&rq
->hrtick_timer
, time
, HRTIMER_MODE_ABS
);
1057 if (reset
&& !hrtimer_active(&rq
->hrtick_timer
))
1064 * High-resolution timer tick.
1065 * Runs from hardirq context with interrupts disabled.
1067 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1069 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1071 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1073 spin_lock(&rq
->lock
);
1074 update_rq_clock(rq
);
1075 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1076 spin_unlock(&rq
->lock
);
1078 return HRTIMER_NORESTART
;
1082 static void hotplug_hrtick_disable(int cpu
)
1084 struct rq
*rq
= cpu_rq(cpu
);
1085 unsigned long flags
;
1087 spin_lock_irqsave(&rq
->lock
, flags
);
1088 rq
->hrtick_flags
= 0;
1089 __set_bit(HRTICK_BLOCK
, &rq
->hrtick_flags
);
1090 spin_unlock_irqrestore(&rq
->lock
, flags
);
1095 static void hotplug_hrtick_enable(int cpu
)
1097 struct rq
*rq
= cpu_rq(cpu
);
1098 unsigned long flags
;
1100 spin_lock_irqsave(&rq
->lock
, flags
);
1101 __clear_bit(HRTICK_BLOCK
, &rq
->hrtick_flags
);
1102 spin_unlock_irqrestore(&rq
->lock
, flags
);
1106 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1108 int cpu
= (int)(long)hcpu
;
1111 case CPU_UP_CANCELED
:
1112 case CPU_UP_CANCELED_FROZEN
:
1113 case CPU_DOWN_PREPARE
:
1114 case CPU_DOWN_PREPARE_FROZEN
:
1116 case CPU_DEAD_FROZEN
:
1117 hotplug_hrtick_disable(cpu
);
1120 case CPU_UP_PREPARE
:
1121 case CPU_UP_PREPARE_FROZEN
:
1122 case CPU_DOWN_FAILED
:
1123 case CPU_DOWN_FAILED_FROZEN
:
1125 case CPU_ONLINE_FROZEN
:
1126 hotplug_hrtick_enable(cpu
);
1133 static void init_hrtick(void)
1135 hotcpu_notifier(hotplug_hrtick
, 0);
1137 #endif /* CONFIG_SMP */
1139 static void init_rq_hrtick(struct rq
*rq
)
1141 rq
->hrtick_flags
= 0;
1142 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1143 rq
->hrtick_timer
.function
= hrtick
;
1144 rq
->hrtick_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_NO_SOFTIRQ
;
1147 void hrtick_resched(void)
1150 unsigned long flags
;
1152 if (!test_thread_flag(TIF_HRTICK_RESCHED
))
1155 local_irq_save(flags
);
1156 rq
= cpu_rq(smp_processor_id());
1158 local_irq_restore(flags
);
1161 static inline void hrtick_clear(struct rq
*rq
)
1165 static inline void hrtick_set(struct rq
*rq
)
1169 static inline void init_rq_hrtick(struct rq
*rq
)
1173 void hrtick_resched(void)
1177 static inline void init_hrtick(void)
1183 * resched_task - mark a task 'to be rescheduled now'.
1185 * On UP this means the setting of the need_resched flag, on SMP it
1186 * might also involve a cross-CPU call to trigger the scheduler on
1191 #ifndef tsk_is_polling
1192 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1195 static void __resched_task(struct task_struct
*p
, int tif_bit
)
1199 assert_spin_locked(&task_rq(p
)->lock
);
1201 if (unlikely(test_tsk_thread_flag(p
, tif_bit
)))
1204 set_tsk_thread_flag(p
, tif_bit
);
1207 if (cpu
== smp_processor_id())
1210 /* NEED_RESCHED must be visible before we test polling */
1212 if (!tsk_is_polling(p
))
1213 smp_send_reschedule(cpu
);
1216 static void resched_cpu(int cpu
)
1218 struct rq
*rq
= cpu_rq(cpu
);
1219 unsigned long flags
;
1221 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1223 resched_task(cpu_curr(cpu
));
1224 spin_unlock_irqrestore(&rq
->lock
, flags
);
1229 * When add_timer_on() enqueues a timer into the timer wheel of an
1230 * idle CPU then this timer might expire before the next timer event
1231 * which is scheduled to wake up that CPU. In case of a completely
1232 * idle system the next event might even be infinite time into the
1233 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1234 * leaves the inner idle loop so the newly added timer is taken into
1235 * account when the CPU goes back to idle and evaluates the timer
1236 * wheel for the next timer event.
1238 void wake_up_idle_cpu(int cpu
)
1240 struct rq
*rq
= cpu_rq(cpu
);
1242 if (cpu
== smp_processor_id())
1246 * This is safe, as this function is called with the timer
1247 * wheel base lock of (cpu) held. When the CPU is on the way
1248 * to idle and has not yet set rq->curr to idle then it will
1249 * be serialized on the timer wheel base lock and take the new
1250 * timer into account automatically.
1252 if (rq
->curr
!= rq
->idle
)
1256 * We can set TIF_RESCHED on the idle task of the other CPU
1257 * lockless. The worst case is that the other CPU runs the
1258 * idle task through an additional NOOP schedule()
1260 set_tsk_thread_flag(rq
->idle
, TIF_NEED_RESCHED
);
1262 /* NEED_RESCHED must be visible before we test polling */
1264 if (!tsk_is_polling(rq
->idle
))
1265 smp_send_reschedule(cpu
);
1267 #endif /* CONFIG_NO_HZ */
1269 #else /* !CONFIG_SMP */
1270 static void __resched_task(struct task_struct
*p
, int tif_bit
)
1272 assert_spin_locked(&task_rq(p
)->lock
);
1273 set_tsk_thread_flag(p
, tif_bit
);
1275 #endif /* CONFIG_SMP */
1277 #if BITS_PER_LONG == 32
1278 # define WMULT_CONST (~0UL)
1280 # define WMULT_CONST (1UL << 32)
1283 #define WMULT_SHIFT 32
1286 * Shift right and round:
1288 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1291 * delta *= weight / lw
1293 static unsigned long
1294 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1295 struct load_weight
*lw
)
1299 if (!lw
->inv_weight
) {
1300 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1303 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1307 tmp
= (u64
)delta_exec
* weight
;
1309 * Check whether we'd overflow the 64-bit multiplication:
1311 if (unlikely(tmp
> WMULT_CONST
))
1312 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1315 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1317 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1320 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1326 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1333 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1334 * of tasks with abnormal "nice" values across CPUs the contribution that
1335 * each task makes to its run queue's load is weighted according to its
1336 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1337 * scaled version of the new time slice allocation that they receive on time
1341 #define WEIGHT_IDLEPRIO 2
1342 #define WMULT_IDLEPRIO (1 << 31)
1345 * Nice levels are multiplicative, with a gentle 10% change for every
1346 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1347 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1348 * that remained on nice 0.
1350 * The "10% effect" is relative and cumulative: from _any_ nice level,
1351 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1352 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1353 * If a task goes up by ~10% and another task goes down by ~10% then
1354 * the relative distance between them is ~25%.)
1356 static const int prio_to_weight
[40] = {
1357 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1358 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1359 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1360 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1361 /* 0 */ 1024, 820, 655, 526, 423,
1362 /* 5 */ 335, 272, 215, 172, 137,
1363 /* 10 */ 110, 87, 70, 56, 45,
1364 /* 15 */ 36, 29, 23, 18, 15,
1368 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1370 * In cases where the weight does not change often, we can use the
1371 * precalculated inverse to speed up arithmetics by turning divisions
1372 * into multiplications:
1374 static const u32 prio_to_wmult
[40] = {
1375 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1376 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1377 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1378 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1379 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1380 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1381 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1382 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1385 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1388 * runqueue iterator, to support SMP load-balancing between different
1389 * scheduling classes, without having to expose their internal data
1390 * structures to the load-balancing proper:
1392 struct rq_iterator
{
1394 struct task_struct
*(*start
)(void *);
1395 struct task_struct
*(*next
)(void *);
1399 static unsigned long
1400 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1401 unsigned long max_load_move
, struct sched_domain
*sd
,
1402 enum cpu_idle_type idle
, int *all_pinned
,
1403 int *this_best_prio
, struct rq_iterator
*iterator
);
1406 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1407 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1408 struct rq_iterator
*iterator
);
1411 #ifdef CONFIG_CGROUP_CPUACCT
1412 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1414 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1417 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1419 update_load_add(&rq
->load
, load
);
1422 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1424 update_load_sub(&rq
->load
, load
);
1428 static unsigned long source_load(int cpu
, int type
);
1429 static unsigned long target_load(int cpu
, int type
);
1430 static unsigned long cpu_avg_load_per_task(int cpu
);
1431 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1433 #ifdef CONFIG_FAIR_GROUP_SCHED
1435 typedef void (*tg_visitor
)(struct task_group
*, int, struct sched_domain
*);
1438 * Iterate the full tree, calling @down when first entering a node and @up when
1439 * leaving it for the final time.
1442 walk_tg_tree(tg_visitor down
, tg_visitor up
, int cpu
, struct sched_domain
*sd
)
1444 struct task_group
*parent
, *child
;
1447 parent
= &root_task_group
;
1449 (*down
)(parent
, cpu
, sd
);
1450 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1457 (*up
)(parent
, cpu
, sd
);
1460 parent
= parent
->parent
;
1466 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1469 * Calculate and set the cpu's group shares.
1472 __update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1473 unsigned long sd_shares
, unsigned long sd_rq_weight
)
1476 unsigned long shares
;
1477 unsigned long rq_weight
;
1482 rq_weight
= tg
->cfs_rq
[cpu
]->load
.weight
;
1485 * If there are currently no tasks on the cpu pretend there is one of
1486 * average load so that when a new task gets to run here it will not
1487 * get delayed by group starvation.
1491 rq_weight
= NICE_0_LOAD
;
1494 if (unlikely(rq_weight
> sd_rq_weight
))
1495 rq_weight
= sd_rq_weight
;
1498 * \Sum shares * rq_weight
1499 * shares = -----------------------
1503 shares
= (sd_shares
* rq_weight
) / (sd_rq_weight
+ 1);
1506 * record the actual number of shares, not the boosted amount.
1508 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1510 if (shares
< MIN_SHARES
)
1511 shares
= MIN_SHARES
;
1512 else if (shares
> MAX_SHARES
)
1513 shares
= MAX_SHARES
;
1515 __set_se_shares(tg
->se
[cpu
], shares
);
1519 * Re-compute the task group their per cpu shares over the given domain.
1520 * This needs to be done in a bottom-up fashion because the rq weight of a
1521 * parent group depends on the shares of its child groups.
1524 tg_shares_up(struct task_group
*tg
, int cpu
, struct sched_domain
*sd
)
1526 unsigned long rq_weight
= 0;
1527 unsigned long shares
= 0;
1530 for_each_cpu_mask(i
, sd
->span
) {
1531 rq_weight
+= tg
->cfs_rq
[i
]->load
.weight
;
1532 shares
+= tg
->cfs_rq
[i
]->shares
;
1535 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1536 shares
= tg
->shares
;
1538 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1539 shares
= tg
->shares
;
1541 for_each_cpu_mask(i
, sd
->span
) {
1542 struct rq
*rq
= cpu_rq(i
);
1543 unsigned long flags
;
1545 spin_lock_irqsave(&rq
->lock
, flags
);
1546 __update_group_shares_cpu(tg
, i
, shares
, rq_weight
);
1547 spin_unlock_irqrestore(&rq
->lock
, flags
);
1552 * Compute the cpu's hierarchical load factor for each task group.
1553 * This needs to be done in a top-down fashion because the load of a child
1554 * group is a fraction of its parents load.
1557 tg_load_down(struct task_group
*tg
, int cpu
, struct sched_domain
*sd
)
1562 load
= cpu_rq(cpu
)->load
.weight
;
1564 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1565 load
*= tg
->cfs_rq
[cpu
]->shares
;
1566 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1569 tg
->cfs_rq
[cpu
]->h_load
= load
;
1573 tg_nop(struct task_group
*tg
, int cpu
, struct sched_domain
*sd
)
1577 static void update_shares(struct sched_domain
*sd
)
1579 walk_tg_tree(tg_nop
, tg_shares_up
, 0, sd
);
1582 static void update_h_load(int cpu
)
1584 walk_tg_tree(tg_load_down
, tg_nop
, cpu
, NULL
);
1587 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1589 cfs_rq
->shares
= shares
;
1594 static inline void update_shares(struct sched_domain
*sd
)
1602 #include "sched_stats.h"
1603 #include "sched_idletask.c"
1604 #include "sched_fair.c"
1605 #include "sched_rt.c"
1606 #ifdef CONFIG_SCHED_DEBUG
1607 # include "sched_debug.c"
1610 #define sched_class_highest (&rt_sched_class)
1611 #define for_each_class(class) \
1612 for (class = sched_class_highest; class; class = class->next)
1614 static void inc_nr_running(struct rq
*rq
)
1619 static void dec_nr_running(struct rq
*rq
)
1624 static void set_load_weight(struct task_struct
*p
)
1626 if (task_has_rt_policy(p
)) {
1627 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1628 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1633 * SCHED_IDLE tasks get minimal weight:
1635 if (p
->policy
== SCHED_IDLE
) {
1636 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1637 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1641 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1642 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1645 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1647 sched_info_queued(p
);
1648 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1652 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1654 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1659 * __normal_prio - return the priority that is based on the static prio
1661 static inline int __normal_prio(struct task_struct
*p
)
1663 return p
->static_prio
;
1667 * Calculate the expected normal priority: i.e. priority
1668 * without taking RT-inheritance into account. Might be
1669 * boosted by interactivity modifiers. Changes upon fork,
1670 * setprio syscalls, and whenever the interactivity
1671 * estimator recalculates.
1673 static inline int normal_prio(struct task_struct
*p
)
1677 if (task_has_rt_policy(p
))
1678 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1680 prio
= __normal_prio(p
);
1685 * Calculate the current priority, i.e. the priority
1686 * taken into account by the scheduler. This value might
1687 * be boosted by RT tasks, or might be boosted by
1688 * interactivity modifiers. Will be RT if the task got
1689 * RT-boosted. If not then it returns p->normal_prio.
1691 static int effective_prio(struct task_struct
*p
)
1693 p
->normal_prio
= normal_prio(p
);
1695 * If we are RT tasks or we were boosted to RT priority,
1696 * keep the priority unchanged. Otherwise, update priority
1697 * to the normal priority:
1699 if (!rt_prio(p
->prio
))
1700 return p
->normal_prio
;
1705 * activate_task - move a task to the runqueue.
1707 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1709 if (task_contributes_to_load(p
))
1710 rq
->nr_uninterruptible
--;
1712 enqueue_task(rq
, p
, wakeup
);
1717 * deactivate_task - remove a task from the runqueue.
1719 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1721 if (task_contributes_to_load(p
))
1722 rq
->nr_uninterruptible
++;
1724 dequeue_task(rq
, p
, sleep
);
1729 * task_curr - is this task currently executing on a CPU?
1730 * @p: the task in question.
1732 inline int task_curr(const struct task_struct
*p
)
1734 return cpu_curr(task_cpu(p
)) == p
;
1737 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1739 set_task_rq(p
, cpu
);
1742 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1743 * successfuly executed on another CPU. We must ensure that updates of
1744 * per-task data have been completed by this moment.
1747 task_thread_info(p
)->cpu
= cpu
;
1751 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1752 const struct sched_class
*prev_class
,
1753 int oldprio
, int running
)
1755 if (prev_class
!= p
->sched_class
) {
1756 if (prev_class
->switched_from
)
1757 prev_class
->switched_from(rq
, p
, running
);
1758 p
->sched_class
->switched_to(rq
, p
, running
);
1760 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1765 /* Used instead of source_load when we know the type == 0 */
1766 static unsigned long weighted_cpuload(const int cpu
)
1768 return cpu_rq(cpu
)->load
.weight
;
1772 * Is this task likely cache-hot:
1775 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1780 * Buddy candidates are cache hot:
1782 if (sched_feat(CACHE_HOT_BUDDY
) && (&p
->se
== cfs_rq_of(&p
->se
)->next
))
1785 if (p
->sched_class
!= &fair_sched_class
)
1788 if (sysctl_sched_migration_cost
== -1)
1790 if (sysctl_sched_migration_cost
== 0)
1793 delta
= now
- p
->se
.exec_start
;
1795 return delta
< (s64
)sysctl_sched_migration_cost
;
1799 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1801 int old_cpu
= task_cpu(p
);
1802 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1803 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1804 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1807 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1809 #ifdef CONFIG_SCHEDSTATS
1810 if (p
->se
.wait_start
)
1811 p
->se
.wait_start
-= clock_offset
;
1812 if (p
->se
.sleep_start
)
1813 p
->se
.sleep_start
-= clock_offset
;
1814 if (p
->se
.block_start
)
1815 p
->se
.block_start
-= clock_offset
;
1816 if (old_cpu
!= new_cpu
) {
1817 schedstat_inc(p
, se
.nr_migrations
);
1818 if (task_hot(p
, old_rq
->clock
, NULL
))
1819 schedstat_inc(p
, se
.nr_forced2_migrations
);
1822 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1823 new_cfsrq
->min_vruntime
;
1825 __set_task_cpu(p
, new_cpu
);
1828 struct migration_req
{
1829 struct list_head list
;
1831 struct task_struct
*task
;
1834 struct completion done
;
1838 * The task's runqueue lock must be held.
1839 * Returns true if you have to wait for migration thread.
1842 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1844 struct rq
*rq
= task_rq(p
);
1847 * If the task is not on a runqueue (and not running), then
1848 * it is sufficient to simply update the task's cpu field.
1850 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1851 set_task_cpu(p
, dest_cpu
);
1855 init_completion(&req
->done
);
1857 req
->dest_cpu
= dest_cpu
;
1858 list_add(&req
->list
, &rq
->migration_queue
);
1864 * wait_task_inactive - wait for a thread to unschedule.
1866 * The caller must ensure that the task *will* unschedule sometime soon,
1867 * else this function might spin for a *long* time. This function can't
1868 * be called with interrupts off, or it may introduce deadlock with
1869 * smp_call_function() if an IPI is sent by the same process we are
1870 * waiting to become inactive.
1872 void wait_task_inactive(struct task_struct
*p
)
1874 unsigned long flags
;
1880 * We do the initial early heuristics without holding
1881 * any task-queue locks at all. We'll only try to get
1882 * the runqueue lock when things look like they will
1888 * If the task is actively running on another CPU
1889 * still, just relax and busy-wait without holding
1892 * NOTE! Since we don't hold any locks, it's not
1893 * even sure that "rq" stays as the right runqueue!
1894 * But we don't care, since "task_running()" will
1895 * return false if the runqueue has changed and p
1896 * is actually now running somewhere else!
1898 while (task_running(rq
, p
))
1902 * Ok, time to look more closely! We need the rq
1903 * lock now, to be *sure*. If we're wrong, we'll
1904 * just go back and repeat.
1906 rq
= task_rq_lock(p
, &flags
);
1907 running
= task_running(rq
, p
);
1908 on_rq
= p
->se
.on_rq
;
1909 task_rq_unlock(rq
, &flags
);
1912 * Was it really running after all now that we
1913 * checked with the proper locks actually held?
1915 * Oops. Go back and try again..
1917 if (unlikely(running
)) {
1923 * It's not enough that it's not actively running,
1924 * it must be off the runqueue _entirely_, and not
1927 * So if it wa still runnable (but just not actively
1928 * running right now), it's preempted, and we should
1929 * yield - it could be a while.
1931 if (unlikely(on_rq
)) {
1932 schedule_timeout_uninterruptible(1);
1937 * Ahh, all good. It wasn't running, and it wasn't
1938 * runnable, which means that it will never become
1939 * running in the future either. We're all done!
1946 * kick_process - kick a running thread to enter/exit the kernel
1947 * @p: the to-be-kicked thread
1949 * Cause a process which is running on another CPU to enter
1950 * kernel-mode, without any delay. (to get signals handled.)
1952 * NOTE: this function doesnt have to take the runqueue lock,
1953 * because all it wants to ensure is that the remote task enters
1954 * the kernel. If the IPI races and the task has been migrated
1955 * to another CPU then no harm is done and the purpose has been
1958 void kick_process(struct task_struct
*p
)
1964 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1965 smp_send_reschedule(cpu
);
1970 * Return a low guess at the load of a migration-source cpu weighted
1971 * according to the scheduling class and "nice" value.
1973 * We want to under-estimate the load of migration sources, to
1974 * balance conservatively.
1976 static unsigned long source_load(int cpu
, int type
)
1978 struct rq
*rq
= cpu_rq(cpu
);
1979 unsigned long total
= weighted_cpuload(cpu
);
1984 return min(rq
->cpu_load
[type
-1], total
);
1988 * Return a high guess at the load of a migration-target cpu weighted
1989 * according to the scheduling class and "nice" value.
1991 static unsigned long target_load(int cpu
, int type
)
1993 struct rq
*rq
= cpu_rq(cpu
);
1994 unsigned long total
= weighted_cpuload(cpu
);
1999 return max(rq
->cpu_load
[type
-1], total
);
2003 * Return the average load per task on the cpu's run queue
2005 static unsigned long cpu_avg_load_per_task(int cpu
)
2007 struct rq
*rq
= cpu_rq(cpu
);
2008 unsigned long total
= weighted_cpuload(cpu
);
2009 unsigned long n
= rq
->nr_running
;
2011 return n
? total
/ n
: SCHED_LOAD_SCALE
;
2015 * find_idlest_group finds and returns the least busy CPU group within the
2018 static struct sched_group
*
2019 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
2021 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2022 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
2023 int load_idx
= sd
->forkexec_idx
;
2024 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
2027 unsigned long load
, avg_load
;
2031 /* Skip over this group if it has no CPUs allowed */
2032 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
2035 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2037 /* Tally up the load of all CPUs in the group */
2040 for_each_cpu_mask(i
, group
->cpumask
) {
2041 /* Bias balancing toward cpus of our domain */
2043 load
= source_load(i
, load_idx
);
2045 load
= target_load(i
, load_idx
);
2050 /* Adjust by relative CPU power of the group */
2051 avg_load
= sg_div_cpu_power(group
,
2052 avg_load
* SCHED_LOAD_SCALE
);
2055 this_load
= avg_load
;
2057 } else if (avg_load
< min_load
) {
2058 min_load
= avg_load
;
2061 } while (group
= group
->next
, group
!= sd
->groups
);
2063 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
2069 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2072 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
,
2075 unsigned long load
, min_load
= ULONG_MAX
;
2079 /* Traverse only the allowed CPUs */
2080 cpus_and(*tmp
, group
->cpumask
, p
->cpus_allowed
);
2082 for_each_cpu_mask(i
, *tmp
) {
2083 load
= weighted_cpuload(i
);
2085 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2095 * sched_balance_self: balance the current task (running on cpu) in domains
2096 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2099 * Balance, ie. select the least loaded group.
2101 * Returns the target CPU number, or the same CPU if no balancing is needed.
2103 * preempt must be disabled.
2105 static int sched_balance_self(int cpu
, int flag
)
2107 struct task_struct
*t
= current
;
2108 struct sched_domain
*tmp
, *sd
= NULL
;
2110 for_each_domain(cpu
, tmp
) {
2112 * If power savings logic is enabled for a domain, stop there.
2114 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2116 if (tmp
->flags
& flag
)
2121 cpumask_t span
, tmpmask
;
2122 struct sched_group
*group
;
2123 int new_cpu
, weight
;
2125 if (!(sd
->flags
& flag
)) {
2131 group
= find_idlest_group(sd
, t
, cpu
);
2137 new_cpu
= find_idlest_cpu(group
, t
, cpu
, &tmpmask
);
2138 if (new_cpu
== -1 || new_cpu
== cpu
) {
2139 /* Now try balancing at a lower domain level of cpu */
2144 /* Now try balancing at a lower domain level of new_cpu */
2147 weight
= cpus_weight(span
);
2148 for_each_domain(cpu
, tmp
) {
2149 if (weight
<= cpus_weight(tmp
->span
))
2151 if (tmp
->flags
& flag
)
2154 /* while loop will break here if sd == NULL */
2160 #endif /* CONFIG_SMP */
2163 * try_to_wake_up - wake up a thread
2164 * @p: the to-be-woken-up thread
2165 * @state: the mask of task states that can be woken
2166 * @sync: do a synchronous wakeup?
2168 * Put it on the run-queue if it's not already there. The "current"
2169 * thread is always on the run-queue (except when the actual
2170 * re-schedule is in progress), and as such you're allowed to do
2171 * the simpler "current->state = TASK_RUNNING" to mark yourself
2172 * runnable without the overhead of this.
2174 * returns failure only if the task is already active.
2176 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
2178 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2179 unsigned long flags
;
2183 if (!sched_feat(SYNC_WAKEUPS
))
2187 rq
= task_rq_lock(p
, &flags
);
2188 old_state
= p
->state
;
2189 if (!(old_state
& state
))
2197 this_cpu
= smp_processor_id();
2200 if (unlikely(task_running(rq
, p
)))
2203 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2204 if (cpu
!= orig_cpu
) {
2205 set_task_cpu(p
, cpu
);
2206 task_rq_unlock(rq
, &flags
);
2207 /* might preempt at this point */
2208 rq
= task_rq_lock(p
, &flags
);
2209 old_state
= p
->state
;
2210 if (!(old_state
& state
))
2215 this_cpu
= smp_processor_id();
2219 #ifdef CONFIG_SCHEDSTATS
2220 schedstat_inc(rq
, ttwu_count
);
2221 if (cpu
== this_cpu
)
2222 schedstat_inc(rq
, ttwu_local
);
2224 struct sched_domain
*sd
;
2225 for_each_domain(this_cpu
, sd
) {
2226 if (cpu_isset(cpu
, sd
->span
)) {
2227 schedstat_inc(sd
, ttwu_wake_remote
);
2232 #endif /* CONFIG_SCHEDSTATS */
2235 #endif /* CONFIG_SMP */
2236 schedstat_inc(p
, se
.nr_wakeups
);
2238 schedstat_inc(p
, se
.nr_wakeups_sync
);
2239 if (orig_cpu
!= cpu
)
2240 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2241 if (cpu
== this_cpu
)
2242 schedstat_inc(p
, se
.nr_wakeups_local
);
2244 schedstat_inc(p
, se
.nr_wakeups_remote
);
2245 update_rq_clock(rq
);
2246 activate_task(rq
, p
, 1);
2250 check_preempt_curr(rq
, p
);
2252 p
->state
= TASK_RUNNING
;
2254 if (p
->sched_class
->task_wake_up
)
2255 p
->sched_class
->task_wake_up(rq
, p
);
2258 task_rq_unlock(rq
, &flags
);
2263 int wake_up_process(struct task_struct
*p
)
2265 return try_to_wake_up(p
, TASK_ALL
, 0);
2267 EXPORT_SYMBOL(wake_up_process
);
2269 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2271 return try_to_wake_up(p
, state
, 0);
2275 * Perform scheduler related setup for a newly forked process p.
2276 * p is forked by current.
2278 * __sched_fork() is basic setup used by init_idle() too:
2280 static void __sched_fork(struct task_struct
*p
)
2282 p
->se
.exec_start
= 0;
2283 p
->se
.sum_exec_runtime
= 0;
2284 p
->se
.prev_sum_exec_runtime
= 0;
2285 p
->se
.last_wakeup
= 0;
2286 p
->se
.avg_overlap
= 0;
2288 #ifdef CONFIG_SCHEDSTATS
2289 p
->se
.wait_start
= 0;
2290 p
->se
.sum_sleep_runtime
= 0;
2291 p
->se
.sleep_start
= 0;
2292 p
->se
.block_start
= 0;
2293 p
->se
.sleep_max
= 0;
2294 p
->se
.block_max
= 0;
2296 p
->se
.slice_max
= 0;
2300 INIT_LIST_HEAD(&p
->rt
.run_list
);
2302 INIT_LIST_HEAD(&p
->se
.group_node
);
2304 #ifdef CONFIG_PREEMPT_NOTIFIERS
2305 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2309 * We mark the process as running here, but have not actually
2310 * inserted it onto the runqueue yet. This guarantees that
2311 * nobody will actually run it, and a signal or other external
2312 * event cannot wake it up and insert it on the runqueue either.
2314 p
->state
= TASK_RUNNING
;
2318 * fork()/clone()-time setup:
2320 void sched_fork(struct task_struct
*p
, int clone_flags
)
2322 int cpu
= get_cpu();
2327 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2329 set_task_cpu(p
, cpu
);
2332 * Make sure we do not leak PI boosting priority to the child:
2334 p
->prio
= current
->normal_prio
;
2335 if (!rt_prio(p
->prio
))
2336 p
->sched_class
= &fair_sched_class
;
2338 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2339 if (likely(sched_info_on()))
2340 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2342 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2345 #ifdef CONFIG_PREEMPT
2346 /* Want to start with kernel preemption disabled. */
2347 task_thread_info(p
)->preempt_count
= 1;
2353 * wake_up_new_task - wake up a newly created task for the first time.
2355 * This function will do some initial scheduler statistics housekeeping
2356 * that must be done for every newly created context, then puts the task
2357 * on the runqueue and wakes it.
2359 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2361 unsigned long flags
;
2364 rq
= task_rq_lock(p
, &flags
);
2365 BUG_ON(p
->state
!= TASK_RUNNING
);
2366 update_rq_clock(rq
);
2368 p
->prio
= effective_prio(p
);
2370 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2371 activate_task(rq
, p
, 0);
2374 * Let the scheduling class do new task startup
2375 * management (if any):
2377 p
->sched_class
->task_new(rq
, p
);
2380 check_preempt_curr(rq
, p
);
2382 if (p
->sched_class
->task_wake_up
)
2383 p
->sched_class
->task_wake_up(rq
, p
);
2385 task_rq_unlock(rq
, &flags
);
2388 #ifdef CONFIG_PREEMPT_NOTIFIERS
2391 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2392 * @notifier: notifier struct to register
2394 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2396 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2398 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2401 * preempt_notifier_unregister - no longer interested in preemption notifications
2402 * @notifier: notifier struct to unregister
2404 * This is safe to call from within a preemption notifier.
2406 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2408 hlist_del(¬ifier
->link
);
2410 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2412 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2414 struct preempt_notifier
*notifier
;
2415 struct hlist_node
*node
;
2417 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2418 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2422 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2423 struct task_struct
*next
)
2425 struct preempt_notifier
*notifier
;
2426 struct hlist_node
*node
;
2428 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2429 notifier
->ops
->sched_out(notifier
, next
);
2432 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2434 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2439 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2440 struct task_struct
*next
)
2444 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2447 * prepare_task_switch - prepare to switch tasks
2448 * @rq: the runqueue preparing to switch
2449 * @prev: the current task that is being switched out
2450 * @next: the task we are going to switch to.
2452 * This is called with the rq lock held and interrupts off. It must
2453 * be paired with a subsequent finish_task_switch after the context
2456 * prepare_task_switch sets up locking and calls architecture specific
2460 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2461 struct task_struct
*next
)
2463 fire_sched_out_preempt_notifiers(prev
, next
);
2464 prepare_lock_switch(rq
, next
);
2465 prepare_arch_switch(next
);
2469 * finish_task_switch - clean up after a task-switch
2470 * @rq: runqueue associated with task-switch
2471 * @prev: the thread we just switched away from.
2473 * finish_task_switch must be called after the context switch, paired
2474 * with a prepare_task_switch call before the context switch.
2475 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2476 * and do any other architecture-specific cleanup actions.
2478 * Note that we may have delayed dropping an mm in context_switch(). If
2479 * so, we finish that here outside of the runqueue lock. (Doing it
2480 * with the lock held can cause deadlocks; see schedule() for
2483 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2484 __releases(rq
->lock
)
2486 struct mm_struct
*mm
= rq
->prev_mm
;
2492 * A task struct has one reference for the use as "current".
2493 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2494 * schedule one last time. The schedule call will never return, and
2495 * the scheduled task must drop that reference.
2496 * The test for TASK_DEAD must occur while the runqueue locks are
2497 * still held, otherwise prev could be scheduled on another cpu, die
2498 * there before we look at prev->state, and then the reference would
2500 * Manfred Spraul <manfred@colorfullife.com>
2502 prev_state
= prev
->state
;
2503 finish_arch_switch(prev
);
2504 finish_lock_switch(rq
, prev
);
2506 if (current
->sched_class
->post_schedule
)
2507 current
->sched_class
->post_schedule(rq
);
2510 fire_sched_in_preempt_notifiers(current
);
2513 if (unlikely(prev_state
== TASK_DEAD
)) {
2515 * Remove function-return probe instances associated with this
2516 * task and put them back on the free list.
2518 kprobe_flush_task(prev
);
2519 put_task_struct(prev
);
2524 * schedule_tail - first thing a freshly forked thread must call.
2525 * @prev: the thread we just switched away from.
2527 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2528 __releases(rq
->lock
)
2530 struct rq
*rq
= this_rq();
2532 finish_task_switch(rq
, prev
);
2533 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2534 /* In this case, finish_task_switch does not reenable preemption */
2537 if (current
->set_child_tid
)
2538 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2542 * context_switch - switch to the new MM and the new
2543 * thread's register state.
2546 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2547 struct task_struct
*next
)
2549 struct mm_struct
*mm
, *oldmm
;
2551 prepare_task_switch(rq
, prev
, next
);
2553 oldmm
= prev
->active_mm
;
2555 * For paravirt, this is coupled with an exit in switch_to to
2556 * combine the page table reload and the switch backend into
2559 arch_enter_lazy_cpu_mode();
2561 if (unlikely(!mm
)) {
2562 next
->active_mm
= oldmm
;
2563 atomic_inc(&oldmm
->mm_count
);
2564 enter_lazy_tlb(oldmm
, next
);
2566 switch_mm(oldmm
, mm
, next
);
2568 if (unlikely(!prev
->mm
)) {
2569 prev
->active_mm
= NULL
;
2570 rq
->prev_mm
= oldmm
;
2573 * Since the runqueue lock will be released by the next
2574 * task (which is an invalid locking op but in the case
2575 * of the scheduler it's an obvious special-case), so we
2576 * do an early lockdep release here:
2578 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2579 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2582 /* Here we just switch the register state and the stack. */
2583 switch_to(prev
, next
, prev
);
2587 * this_rq must be evaluated again because prev may have moved
2588 * CPUs since it called schedule(), thus the 'rq' on its stack
2589 * frame will be invalid.
2591 finish_task_switch(this_rq(), prev
);
2595 * nr_running, nr_uninterruptible and nr_context_switches:
2597 * externally visible scheduler statistics: current number of runnable
2598 * threads, current number of uninterruptible-sleeping threads, total
2599 * number of context switches performed since bootup.
2601 unsigned long nr_running(void)
2603 unsigned long i
, sum
= 0;
2605 for_each_online_cpu(i
)
2606 sum
+= cpu_rq(i
)->nr_running
;
2611 unsigned long nr_uninterruptible(void)
2613 unsigned long i
, sum
= 0;
2615 for_each_possible_cpu(i
)
2616 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2619 * Since we read the counters lockless, it might be slightly
2620 * inaccurate. Do not allow it to go below zero though:
2622 if (unlikely((long)sum
< 0))
2628 unsigned long long nr_context_switches(void)
2631 unsigned long long sum
= 0;
2633 for_each_possible_cpu(i
)
2634 sum
+= cpu_rq(i
)->nr_switches
;
2639 unsigned long nr_iowait(void)
2641 unsigned long i
, sum
= 0;
2643 for_each_possible_cpu(i
)
2644 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2649 unsigned long nr_active(void)
2651 unsigned long i
, running
= 0, uninterruptible
= 0;
2653 for_each_online_cpu(i
) {
2654 running
+= cpu_rq(i
)->nr_running
;
2655 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2658 if (unlikely((long)uninterruptible
< 0))
2659 uninterruptible
= 0;
2661 return running
+ uninterruptible
;
2665 * Update rq->cpu_load[] statistics. This function is usually called every
2666 * scheduler tick (TICK_NSEC).
2668 static void update_cpu_load(struct rq
*this_rq
)
2670 unsigned long this_load
= this_rq
->load
.weight
;
2673 this_rq
->nr_load_updates
++;
2675 /* Update our load: */
2676 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2677 unsigned long old_load
, new_load
;
2679 /* scale is effectively 1 << i now, and >> i divides by scale */
2681 old_load
= this_rq
->cpu_load
[i
];
2682 new_load
= this_load
;
2684 * Round up the averaging division if load is increasing. This
2685 * prevents us from getting stuck on 9 if the load is 10, for
2688 if (new_load
> old_load
)
2689 new_load
+= scale
-1;
2690 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2697 * double_rq_lock - safely lock two runqueues
2699 * Note this does not disable interrupts like task_rq_lock,
2700 * you need to do so manually before calling.
2702 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2703 __acquires(rq1
->lock
)
2704 __acquires(rq2
->lock
)
2706 BUG_ON(!irqs_disabled());
2708 spin_lock(&rq1
->lock
);
2709 __acquire(rq2
->lock
); /* Fake it out ;) */
2712 spin_lock(&rq1
->lock
);
2713 spin_lock(&rq2
->lock
);
2715 spin_lock(&rq2
->lock
);
2716 spin_lock(&rq1
->lock
);
2719 update_rq_clock(rq1
);
2720 update_rq_clock(rq2
);
2724 * double_rq_unlock - safely unlock two runqueues
2726 * Note this does not restore interrupts like task_rq_unlock,
2727 * you need to do so manually after calling.
2729 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2730 __releases(rq1
->lock
)
2731 __releases(rq2
->lock
)
2733 spin_unlock(&rq1
->lock
);
2735 spin_unlock(&rq2
->lock
);
2737 __release(rq2
->lock
);
2741 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2743 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2744 __releases(this_rq
->lock
)
2745 __acquires(busiest
->lock
)
2746 __acquires(this_rq
->lock
)
2750 if (unlikely(!irqs_disabled())) {
2751 /* printk() doesn't work good under rq->lock */
2752 spin_unlock(&this_rq
->lock
);
2755 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2756 if (busiest
< this_rq
) {
2757 spin_unlock(&this_rq
->lock
);
2758 spin_lock(&busiest
->lock
);
2759 spin_lock(&this_rq
->lock
);
2762 spin_lock(&busiest
->lock
);
2768 * If dest_cpu is allowed for this process, migrate the task to it.
2769 * This is accomplished by forcing the cpu_allowed mask to only
2770 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2771 * the cpu_allowed mask is restored.
2773 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2775 struct migration_req req
;
2776 unsigned long flags
;
2779 rq
= task_rq_lock(p
, &flags
);
2780 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2781 || unlikely(cpu_is_offline(dest_cpu
)))
2784 /* force the process onto the specified CPU */
2785 if (migrate_task(p
, dest_cpu
, &req
)) {
2786 /* Need to wait for migration thread (might exit: take ref). */
2787 struct task_struct
*mt
= rq
->migration_thread
;
2789 get_task_struct(mt
);
2790 task_rq_unlock(rq
, &flags
);
2791 wake_up_process(mt
);
2792 put_task_struct(mt
);
2793 wait_for_completion(&req
.done
);
2798 task_rq_unlock(rq
, &flags
);
2802 * sched_exec - execve() is a valuable balancing opportunity, because at
2803 * this point the task has the smallest effective memory and cache footprint.
2805 void sched_exec(void)
2807 int new_cpu
, this_cpu
= get_cpu();
2808 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2810 if (new_cpu
!= this_cpu
)
2811 sched_migrate_task(current
, new_cpu
);
2815 * pull_task - move a task from a remote runqueue to the local runqueue.
2816 * Both runqueues must be locked.
2818 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2819 struct rq
*this_rq
, int this_cpu
)
2821 deactivate_task(src_rq
, p
, 0);
2822 set_task_cpu(p
, this_cpu
);
2823 activate_task(this_rq
, p
, 0);
2825 * Note that idle threads have a prio of MAX_PRIO, for this test
2826 * to be always true for them.
2828 check_preempt_curr(this_rq
, p
);
2832 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2835 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2836 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2840 * We do not migrate tasks that are:
2841 * 1) running (obviously), or
2842 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2843 * 3) are cache-hot on their current CPU.
2845 if (!cpu_isset(this_cpu
, p
->cpus_allowed
)) {
2846 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2851 if (task_running(rq
, p
)) {
2852 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2857 * Aggressive migration if:
2858 * 1) task is cache cold, or
2859 * 2) too many balance attempts have failed.
2862 if (!task_hot(p
, rq
->clock
, sd
) ||
2863 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2864 #ifdef CONFIG_SCHEDSTATS
2865 if (task_hot(p
, rq
->clock
, sd
)) {
2866 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2867 schedstat_inc(p
, se
.nr_forced_migrations
);
2873 if (task_hot(p
, rq
->clock
, sd
)) {
2874 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
2880 static unsigned long
2881 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2882 unsigned long max_load_move
, struct sched_domain
*sd
,
2883 enum cpu_idle_type idle
, int *all_pinned
,
2884 int *this_best_prio
, struct rq_iterator
*iterator
)
2886 int loops
= 0, pulled
= 0, pinned
= 0, skip_for_load
;
2887 struct task_struct
*p
;
2888 long rem_load_move
= max_load_move
;
2890 if (max_load_move
== 0)
2896 * Start the load-balancing iterator:
2898 p
= iterator
->start(iterator
->arg
);
2900 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
2903 * To help distribute high priority tasks across CPUs we don't
2904 * skip a task if it will be the highest priority task (i.e. smallest
2905 * prio value) on its new queue regardless of its load weight
2907 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2908 SCHED_LOAD_SCALE_FUZZ
;
2909 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2910 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2911 p
= iterator
->next(iterator
->arg
);
2915 pull_task(busiest
, p
, this_rq
, this_cpu
);
2917 rem_load_move
-= p
->se
.load
.weight
;
2920 * We only want to steal up to the prescribed amount of weighted load.
2922 if (rem_load_move
> 0) {
2923 if (p
->prio
< *this_best_prio
)
2924 *this_best_prio
= p
->prio
;
2925 p
= iterator
->next(iterator
->arg
);
2930 * Right now, this is one of only two places pull_task() is called,
2931 * so we can safely collect pull_task() stats here rather than
2932 * inside pull_task().
2934 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2937 *all_pinned
= pinned
;
2939 return max_load_move
- rem_load_move
;
2943 * move_tasks tries to move up to max_load_move weighted load from busiest to
2944 * this_rq, as part of a balancing operation within domain "sd".
2945 * Returns 1 if successful and 0 otherwise.
2947 * Called with both runqueues locked.
2949 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2950 unsigned long max_load_move
,
2951 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2954 const struct sched_class
*class = sched_class_highest
;
2955 unsigned long total_load_moved
= 0;
2956 int this_best_prio
= this_rq
->curr
->prio
;
2960 class->load_balance(this_rq
, this_cpu
, busiest
,
2961 max_load_move
- total_load_moved
,
2962 sd
, idle
, all_pinned
, &this_best_prio
);
2963 class = class->next
;
2964 } while (class && max_load_move
> total_load_moved
);
2966 return total_load_moved
> 0;
2970 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2971 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2972 struct rq_iterator
*iterator
)
2974 struct task_struct
*p
= iterator
->start(iterator
->arg
);
2978 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2979 pull_task(busiest
, p
, this_rq
, this_cpu
);
2981 * Right now, this is only the second place pull_task()
2982 * is called, so we can safely collect pull_task()
2983 * stats here rather than inside pull_task().
2985 schedstat_inc(sd
, lb_gained
[idle
]);
2989 p
= iterator
->next(iterator
->arg
);
2996 * move_one_task tries to move exactly one task from busiest to this_rq, as
2997 * part of active balancing operations within "domain".
2998 * Returns 1 if successful and 0 otherwise.
3000 * Called with both runqueues locked.
3002 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3003 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3005 const struct sched_class
*class;
3007 for (class = sched_class_highest
; class; class = class->next
)
3008 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3015 * find_busiest_group finds and returns the busiest CPU group within the
3016 * domain. It calculates and returns the amount of weighted load which
3017 * should be moved to restore balance via the imbalance parameter.
3019 static struct sched_group
*
3020 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
3021 unsigned long *imbalance
, enum cpu_idle_type idle
,
3022 int *sd_idle
, const cpumask_t
*cpus
, int *balance
)
3024 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
3025 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
3026 unsigned long max_pull
;
3027 unsigned long busiest_load_per_task
, busiest_nr_running
;
3028 unsigned long this_load_per_task
, this_nr_running
;
3029 int load_idx
, group_imb
= 0;
3030 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3031 int power_savings_balance
= 1;
3032 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
3033 unsigned long min_nr_running
= ULONG_MAX
;
3034 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
3037 max_load
= this_load
= total_load
= total_pwr
= 0;
3038 busiest_load_per_task
= busiest_nr_running
= 0;
3039 this_load_per_task
= this_nr_running
= 0;
3040 if (idle
== CPU_NOT_IDLE
)
3041 load_idx
= sd
->busy_idx
;
3042 else if (idle
== CPU_NEWLY_IDLE
)
3043 load_idx
= sd
->newidle_idx
;
3045 load_idx
= sd
->idle_idx
;
3048 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
3051 int __group_imb
= 0;
3052 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3053 unsigned long sum_nr_running
, sum_weighted_load
;
3055 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
3058 balance_cpu
= first_cpu(group
->cpumask
);
3060 /* Tally up the load of all CPUs in the group */
3061 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
3063 min_cpu_load
= ~0UL;
3065 for_each_cpu_mask(i
, group
->cpumask
) {
3068 if (!cpu_isset(i
, *cpus
))
3073 if (*sd_idle
&& rq
->nr_running
)
3076 /* Bias balancing toward cpus of our domain */
3078 if (idle_cpu(i
) && !first_idle_cpu
) {
3083 load
= target_load(i
, load_idx
);
3085 load
= source_load(i
, load_idx
);
3086 if (load
> max_cpu_load
)
3087 max_cpu_load
= load
;
3088 if (min_cpu_load
> load
)
3089 min_cpu_load
= load
;
3093 sum_nr_running
+= rq
->nr_running
;
3094 sum_weighted_load
+= weighted_cpuload(i
);
3098 * First idle cpu or the first cpu(busiest) in this sched group
3099 * is eligible for doing load balancing at this and above
3100 * domains. In the newly idle case, we will allow all the cpu's
3101 * to do the newly idle load balance.
3103 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3104 balance_cpu
!= this_cpu
&& balance
) {
3109 total_load
+= avg_load
;
3110 total_pwr
+= group
->__cpu_power
;
3112 /* Adjust by relative CPU power of the group */
3113 avg_load
= sg_div_cpu_power(group
,
3114 avg_load
* SCHED_LOAD_SCALE
);
3116 if ((max_cpu_load
- min_cpu_load
) > SCHED_LOAD_SCALE
)
3119 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
3122 this_load
= avg_load
;
3124 this_nr_running
= sum_nr_running
;
3125 this_load_per_task
= sum_weighted_load
;
3126 } else if (avg_load
> max_load
&&
3127 (sum_nr_running
> group_capacity
|| __group_imb
)) {
3128 max_load
= avg_load
;
3130 busiest_nr_running
= sum_nr_running
;
3131 busiest_load_per_task
= sum_weighted_load
;
3132 group_imb
= __group_imb
;
3135 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3137 * Busy processors will not participate in power savings
3140 if (idle
== CPU_NOT_IDLE
||
3141 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3145 * If the local group is idle or completely loaded
3146 * no need to do power savings balance at this domain
3148 if (local_group
&& (this_nr_running
>= group_capacity
||
3150 power_savings_balance
= 0;
3153 * If a group is already running at full capacity or idle,
3154 * don't include that group in power savings calculations
3156 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
3161 * Calculate the group which has the least non-idle load.
3162 * This is the group from where we need to pick up the load
3165 if ((sum_nr_running
< min_nr_running
) ||
3166 (sum_nr_running
== min_nr_running
&&
3167 first_cpu(group
->cpumask
) <
3168 first_cpu(group_min
->cpumask
))) {
3170 min_nr_running
= sum_nr_running
;
3171 min_load_per_task
= sum_weighted_load
/
3176 * Calculate the group which is almost near its
3177 * capacity but still has some space to pick up some load
3178 * from other group and save more power
3180 if (sum_nr_running
<= group_capacity
- 1) {
3181 if (sum_nr_running
> leader_nr_running
||
3182 (sum_nr_running
== leader_nr_running
&&
3183 first_cpu(group
->cpumask
) >
3184 first_cpu(group_leader
->cpumask
))) {
3185 group_leader
= group
;
3186 leader_nr_running
= sum_nr_running
;
3191 group
= group
->next
;
3192 } while (group
!= sd
->groups
);
3194 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
3197 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
3199 if (this_load
>= avg_load
||
3200 100*max_load
<= sd
->imbalance_pct
*this_load
)
3203 busiest_load_per_task
/= busiest_nr_running
;
3205 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
3208 * We're trying to get all the cpus to the average_load, so we don't
3209 * want to push ourselves above the average load, nor do we wish to
3210 * reduce the max loaded cpu below the average load, as either of these
3211 * actions would just result in more rebalancing later, and ping-pong
3212 * tasks around. Thus we look for the minimum possible imbalance.
3213 * Negative imbalances (*we* are more loaded than anyone else) will
3214 * be counted as no imbalance for these purposes -- we can't fix that
3215 * by pulling tasks to us. Be careful of negative numbers as they'll
3216 * appear as very large values with unsigned longs.
3218 if (max_load
<= busiest_load_per_task
)
3222 * In the presence of smp nice balancing, certain scenarios can have
3223 * max load less than avg load(as we skip the groups at or below
3224 * its cpu_power, while calculating max_load..)
3226 if (max_load
< avg_load
) {
3228 goto small_imbalance
;
3231 /* Don't want to pull so many tasks that a group would go idle */
3232 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
3234 /* How much load to actually move to equalise the imbalance */
3235 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
3236 (avg_load
- this_load
) * this->__cpu_power
)
3240 * if *imbalance is less than the average load per runnable task
3241 * there is no gaurantee that any tasks will be moved so we'll have
3242 * a think about bumping its value to force at least one task to be
3245 if (*imbalance
< busiest_load_per_task
) {
3246 unsigned long tmp
, pwr_now
, pwr_move
;
3250 pwr_move
= pwr_now
= 0;
3252 if (this_nr_running
) {
3253 this_load_per_task
/= this_nr_running
;
3254 if (busiest_load_per_task
> this_load_per_task
)
3257 this_load_per_task
= SCHED_LOAD_SCALE
;
3259 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
3260 busiest_load_per_task
* imbn
) {
3261 *imbalance
= busiest_load_per_task
;
3266 * OK, we don't have enough imbalance to justify moving tasks,
3267 * however we may be able to increase total CPU power used by
3271 pwr_now
+= busiest
->__cpu_power
*
3272 min(busiest_load_per_task
, max_load
);
3273 pwr_now
+= this->__cpu_power
*
3274 min(this_load_per_task
, this_load
);
3275 pwr_now
/= SCHED_LOAD_SCALE
;
3277 /* Amount of load we'd subtract */
3278 tmp
= sg_div_cpu_power(busiest
,
3279 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3281 pwr_move
+= busiest
->__cpu_power
*
3282 min(busiest_load_per_task
, max_load
- tmp
);
3284 /* Amount of load we'd add */
3285 if (max_load
* busiest
->__cpu_power
<
3286 busiest_load_per_task
* SCHED_LOAD_SCALE
)
3287 tmp
= sg_div_cpu_power(this,
3288 max_load
* busiest
->__cpu_power
);
3290 tmp
= sg_div_cpu_power(this,
3291 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3292 pwr_move
+= this->__cpu_power
*
3293 min(this_load_per_task
, this_load
+ tmp
);
3294 pwr_move
/= SCHED_LOAD_SCALE
;
3296 /* Move if we gain throughput */
3297 if (pwr_move
> pwr_now
)
3298 *imbalance
= busiest_load_per_task
;
3304 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3305 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3308 if (this == group_leader
&& group_leader
!= group_min
) {
3309 *imbalance
= min_load_per_task
;
3319 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3322 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
3323 unsigned long imbalance
, const cpumask_t
*cpus
)
3325 struct rq
*busiest
= NULL
, *rq
;
3326 unsigned long max_load
= 0;
3329 for_each_cpu_mask(i
, group
->cpumask
) {
3332 if (!cpu_isset(i
, *cpus
))
3336 wl
= weighted_cpuload(i
);
3338 if (rq
->nr_running
== 1 && wl
> imbalance
)
3341 if (wl
> max_load
) {
3351 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3352 * so long as it is large enough.
3354 #define MAX_PINNED_INTERVAL 512
3357 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3358 * tasks if there is an imbalance.
3360 static int load_balance(int this_cpu
, struct rq
*this_rq
,
3361 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3362 int *balance
, cpumask_t
*cpus
)
3364 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
3365 struct sched_group
*group
;
3366 unsigned long imbalance
;
3368 unsigned long flags
;
3373 * When power savings policy is enabled for the parent domain, idle
3374 * sibling can pick up load irrespective of busy siblings. In this case,
3375 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3376 * portraying it as CPU_NOT_IDLE.
3378 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3379 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3382 schedstat_inc(sd
, lb_count
[idle
]);
3386 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3393 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3397 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
3399 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3403 BUG_ON(busiest
== this_rq
);
3405 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3408 if (busiest
->nr_running
> 1) {
3410 * Attempt to move tasks. If find_busiest_group has found
3411 * an imbalance but busiest->nr_running <= 1, the group is
3412 * still unbalanced. ld_moved simply stays zero, so it is
3413 * correctly treated as an imbalance.
3415 local_irq_save(flags
);
3416 double_rq_lock(this_rq
, busiest
);
3417 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3418 imbalance
, sd
, idle
, &all_pinned
);
3419 double_rq_unlock(this_rq
, busiest
);
3420 local_irq_restore(flags
);
3423 * some other cpu did the load balance for us.
3425 if (ld_moved
&& this_cpu
!= smp_processor_id())
3426 resched_cpu(this_cpu
);
3428 /* All tasks on this runqueue were pinned by CPU affinity */
3429 if (unlikely(all_pinned
)) {
3430 cpu_clear(cpu_of(busiest
), *cpus
);
3431 if (!cpus_empty(*cpus
))
3438 schedstat_inc(sd
, lb_failed
[idle
]);
3439 sd
->nr_balance_failed
++;
3441 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3443 spin_lock_irqsave(&busiest
->lock
, flags
);
3445 /* don't kick the migration_thread, if the curr
3446 * task on busiest cpu can't be moved to this_cpu
3448 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
3449 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3451 goto out_one_pinned
;
3454 if (!busiest
->active_balance
) {
3455 busiest
->active_balance
= 1;
3456 busiest
->push_cpu
= this_cpu
;
3459 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3461 wake_up_process(busiest
->migration_thread
);
3464 * We've kicked active balancing, reset the failure
3467 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3470 sd
->nr_balance_failed
= 0;
3472 if (likely(!active_balance
)) {
3473 /* We were unbalanced, so reset the balancing interval */
3474 sd
->balance_interval
= sd
->min_interval
;
3477 * If we've begun active balancing, start to back off. This
3478 * case may not be covered by the all_pinned logic if there
3479 * is only 1 task on the busy runqueue (because we don't call
3482 if (sd
->balance_interval
< sd
->max_interval
)
3483 sd
->balance_interval
*= 2;
3486 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3487 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3493 schedstat_inc(sd
, lb_balanced
[idle
]);
3495 sd
->nr_balance_failed
= 0;
3498 /* tune up the balancing interval */
3499 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3500 (sd
->balance_interval
< sd
->max_interval
))
3501 sd
->balance_interval
*= 2;
3503 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3504 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3515 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3516 * tasks if there is an imbalance.
3518 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3519 * this_rq is locked.
3522 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
,
3525 struct sched_group
*group
;
3526 struct rq
*busiest
= NULL
;
3527 unsigned long imbalance
;
3535 * When power savings policy is enabled for the parent domain, idle
3536 * sibling can pick up load irrespective of busy siblings. In this case,
3537 * let the state of idle sibling percolate up as IDLE, instead of
3538 * portraying it as CPU_NOT_IDLE.
3540 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
3541 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3544 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
3546 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
3547 &sd_idle
, cpus
, NULL
);
3549 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
3553 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
3555 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
3559 BUG_ON(busiest
== this_rq
);
3561 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
3564 if (busiest
->nr_running
> 1) {
3565 /* Attempt to move tasks */
3566 double_lock_balance(this_rq
, busiest
);
3567 /* this_rq->clock is already updated */
3568 update_rq_clock(busiest
);
3569 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3570 imbalance
, sd
, CPU_NEWLY_IDLE
,
3572 spin_unlock(&busiest
->lock
);
3574 if (unlikely(all_pinned
)) {
3575 cpu_clear(cpu_of(busiest
), *cpus
);
3576 if (!cpus_empty(*cpus
))
3582 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
3583 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3584 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3587 sd
->nr_balance_failed
= 0;
3592 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
3593 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3594 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3596 sd
->nr_balance_failed
= 0;
3602 * idle_balance is called by schedule() if this_cpu is about to become
3603 * idle. Attempts to pull tasks from other CPUs.
3605 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
3607 struct sched_domain
*sd
;
3608 int pulled_task
= -1;
3609 unsigned long next_balance
= jiffies
+ HZ
;
3612 for_each_domain(this_cpu
, sd
) {
3613 unsigned long interval
;
3615 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3618 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
3619 /* If we've pulled tasks over stop searching: */
3620 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
3623 interval
= msecs_to_jiffies(sd
->balance_interval
);
3624 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3625 next_balance
= sd
->last_balance
+ interval
;
3629 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3631 * We are going idle. next_balance may be set based on
3632 * a busy processor. So reset next_balance.
3634 this_rq
->next_balance
= next_balance
;
3639 * active_load_balance is run by migration threads. It pushes running tasks
3640 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3641 * running on each physical CPU where possible, and avoids physical /
3642 * logical imbalances.
3644 * Called with busiest_rq locked.
3646 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
3648 int target_cpu
= busiest_rq
->push_cpu
;
3649 struct sched_domain
*sd
;
3650 struct rq
*target_rq
;
3652 /* Is there any task to move? */
3653 if (busiest_rq
->nr_running
<= 1)
3656 target_rq
= cpu_rq(target_cpu
);
3659 * This condition is "impossible", if it occurs
3660 * we need to fix it. Originally reported by
3661 * Bjorn Helgaas on a 128-cpu setup.
3663 BUG_ON(busiest_rq
== target_rq
);
3665 /* move a task from busiest_rq to target_rq */
3666 double_lock_balance(busiest_rq
, target_rq
);
3667 update_rq_clock(busiest_rq
);
3668 update_rq_clock(target_rq
);
3670 /* Search for an sd spanning us and the target CPU. */
3671 for_each_domain(target_cpu
, sd
) {
3672 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3673 cpu_isset(busiest_cpu
, sd
->span
))
3678 schedstat_inc(sd
, alb_count
);
3680 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3682 schedstat_inc(sd
, alb_pushed
);
3684 schedstat_inc(sd
, alb_failed
);
3686 spin_unlock(&target_rq
->lock
);
3691 atomic_t load_balancer
;
3693 } nohz ____cacheline_aligned
= {
3694 .load_balancer
= ATOMIC_INIT(-1),
3695 .cpu_mask
= CPU_MASK_NONE
,
3699 * This routine will try to nominate the ilb (idle load balancing)
3700 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3701 * load balancing on behalf of all those cpus. If all the cpus in the system
3702 * go into this tickless mode, then there will be no ilb owner (as there is
3703 * no need for one) and all the cpus will sleep till the next wakeup event
3706 * For the ilb owner, tick is not stopped. And this tick will be used
3707 * for idle load balancing. ilb owner will still be part of
3710 * While stopping the tick, this cpu will become the ilb owner if there
3711 * is no other owner. And will be the owner till that cpu becomes busy
3712 * or if all cpus in the system stop their ticks at which point
3713 * there is no need for ilb owner.
3715 * When the ilb owner becomes busy, it nominates another owner, during the
3716 * next busy scheduler_tick()
3718 int select_nohz_load_balancer(int stop_tick
)
3720 int cpu
= smp_processor_id();
3723 cpu_set(cpu
, nohz
.cpu_mask
);
3724 cpu_rq(cpu
)->in_nohz_recently
= 1;
3727 * If we are going offline and still the leader, give up!
3729 if (cpu_is_offline(cpu
) &&
3730 atomic_read(&nohz
.load_balancer
) == cpu
) {
3731 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3736 /* time for ilb owner also to sleep */
3737 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3738 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3739 atomic_set(&nohz
.load_balancer
, -1);
3743 if (atomic_read(&nohz
.load_balancer
) == -1) {
3744 /* make me the ilb owner */
3745 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3747 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3750 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3753 cpu_clear(cpu
, nohz
.cpu_mask
);
3755 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3756 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3763 static DEFINE_SPINLOCK(balancing
);
3766 * It checks each scheduling domain to see if it is due to be balanced,
3767 * and initiates a balancing operation if so.
3769 * Balancing parameters are set up in arch_init_sched_domains.
3771 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3774 struct rq
*rq
= cpu_rq(cpu
);
3775 unsigned long interval
;
3776 struct sched_domain
*sd
;
3777 /* Earliest time when we have to do rebalance again */
3778 unsigned long next_balance
= jiffies
+ 60*HZ
;
3779 int update_next_balance
= 0;
3783 for_each_domain(cpu
, sd
) {
3784 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3787 interval
= sd
->balance_interval
;
3788 if (idle
!= CPU_IDLE
)
3789 interval
*= sd
->busy_factor
;
3791 /* scale ms to jiffies */
3792 interval
= msecs_to_jiffies(interval
);
3793 if (unlikely(!interval
))
3795 if (interval
> HZ
*NR_CPUS
/10)
3796 interval
= HZ
*NR_CPUS
/10;
3798 need_serialize
= sd
->flags
& SD_SERIALIZE
;
3800 if (need_serialize
) {
3801 if (!spin_trylock(&balancing
))
3805 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3806 if (load_balance(cpu
, rq
, sd
, idle
, &balance
, &tmp
)) {
3808 * We've pulled tasks over so either we're no
3809 * longer idle, or one of our SMT siblings is
3812 idle
= CPU_NOT_IDLE
;
3814 sd
->last_balance
= jiffies
;
3817 spin_unlock(&balancing
);
3819 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3820 next_balance
= sd
->last_balance
+ interval
;
3821 update_next_balance
= 1;
3825 * Stop the load balance at this level. There is another
3826 * CPU in our sched group which is doing load balancing more
3834 * next_balance will be updated only when there is a need.
3835 * When the cpu is attached to null domain for ex, it will not be
3838 if (likely(update_next_balance
))
3839 rq
->next_balance
= next_balance
;
3843 * run_rebalance_domains is triggered when needed from the scheduler tick.
3844 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3845 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3847 static void run_rebalance_domains(struct softirq_action
*h
)
3849 int this_cpu
= smp_processor_id();
3850 struct rq
*this_rq
= cpu_rq(this_cpu
);
3851 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3852 CPU_IDLE
: CPU_NOT_IDLE
;
3854 rebalance_domains(this_cpu
, idle
);
3858 * If this cpu is the owner for idle load balancing, then do the
3859 * balancing on behalf of the other idle cpus whose ticks are
3862 if (this_rq
->idle_at_tick
&&
3863 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3864 cpumask_t cpus
= nohz
.cpu_mask
;
3868 cpu_clear(this_cpu
, cpus
);
3869 for_each_cpu_mask(balance_cpu
, cpus
) {
3871 * If this cpu gets work to do, stop the load balancing
3872 * work being done for other cpus. Next load
3873 * balancing owner will pick it up.
3878 rebalance_domains(balance_cpu
, CPU_IDLE
);
3880 rq
= cpu_rq(balance_cpu
);
3881 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3882 this_rq
->next_balance
= rq
->next_balance
;
3889 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3891 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3892 * idle load balancing owner or decide to stop the periodic load balancing,
3893 * if the whole system is idle.
3895 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3899 * If we were in the nohz mode recently and busy at the current
3900 * scheduler tick, then check if we need to nominate new idle
3903 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3904 rq
->in_nohz_recently
= 0;
3906 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3907 cpu_clear(cpu
, nohz
.cpu_mask
);
3908 atomic_set(&nohz
.load_balancer
, -1);
3911 if (atomic_read(&nohz
.load_balancer
) == -1) {
3913 * simple selection for now: Nominate the
3914 * first cpu in the nohz list to be the next
3917 * TBD: Traverse the sched domains and nominate
3918 * the nearest cpu in the nohz.cpu_mask.
3920 int ilb
= first_cpu(nohz
.cpu_mask
);
3922 if (ilb
< nr_cpu_ids
)
3928 * If this cpu is idle and doing idle load balancing for all the
3929 * cpus with ticks stopped, is it time for that to stop?
3931 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3932 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3938 * If this cpu is idle and the idle load balancing is done by
3939 * someone else, then no need raise the SCHED_SOFTIRQ
3941 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3942 cpu_isset(cpu
, nohz
.cpu_mask
))
3945 if (time_after_eq(jiffies
, rq
->next_balance
))
3946 raise_softirq(SCHED_SOFTIRQ
);
3949 #else /* CONFIG_SMP */
3952 * on UP we do not need to balance between CPUs:
3954 static inline void idle_balance(int cpu
, struct rq
*rq
)
3960 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3962 EXPORT_PER_CPU_SYMBOL(kstat
);
3965 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3966 * that have not yet been banked in case the task is currently running.
3968 unsigned long long task_sched_runtime(struct task_struct
*p
)
3970 unsigned long flags
;
3974 rq
= task_rq_lock(p
, &flags
);
3975 ns
= p
->se
.sum_exec_runtime
;
3976 if (task_current(rq
, p
)) {
3977 update_rq_clock(rq
);
3978 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3979 if ((s64
)delta_exec
> 0)
3982 task_rq_unlock(rq
, &flags
);
3988 * Account user cpu time to a process.
3989 * @p: the process that the cpu time gets accounted to
3990 * @cputime: the cpu time spent in user space since the last update
3992 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3994 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3997 p
->utime
= cputime_add(p
->utime
, cputime
);
3999 /* Add user time to cpustat. */
4000 tmp
= cputime_to_cputime64(cputime
);
4001 if (TASK_NICE(p
) > 0)
4002 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
4004 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4008 * Account guest cpu time to a process.
4009 * @p: the process that the cpu time gets accounted to
4010 * @cputime: the cpu time spent in virtual machine since the last update
4012 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
)
4015 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4017 tmp
= cputime_to_cputime64(cputime
);
4019 p
->utime
= cputime_add(p
->utime
, cputime
);
4020 p
->gtime
= cputime_add(p
->gtime
, cputime
);
4022 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4023 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
4027 * Account scaled user cpu time to a process.
4028 * @p: the process that the cpu time gets accounted to
4029 * @cputime: the cpu time spent in user space since the last update
4031 void account_user_time_scaled(struct task_struct
*p
, cputime_t cputime
)
4033 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime
);
4037 * Account system cpu time to a process.
4038 * @p: the process that the cpu time gets accounted to
4039 * @hardirq_offset: the offset to subtract from hardirq_count()
4040 * @cputime: the cpu time spent in kernel space since the last update
4042 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
4045 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4046 struct rq
*rq
= this_rq();
4049 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
4050 account_guest_time(p
, cputime
);
4054 p
->stime
= cputime_add(p
->stime
, cputime
);
4056 /* Add system time to cpustat. */
4057 tmp
= cputime_to_cputime64(cputime
);
4058 if (hardirq_count() - hardirq_offset
)
4059 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
4060 else if (softirq_count())
4061 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
4062 else if (p
!= rq
->idle
)
4063 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
4064 else if (atomic_read(&rq
->nr_iowait
) > 0)
4065 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
4067 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
4068 /* Account for system time used */
4069 acct_update_integrals(p
);
4073 * Account scaled system cpu time to a process.
4074 * @p: the process that the cpu time gets accounted to
4075 * @hardirq_offset: the offset to subtract from hardirq_count()
4076 * @cputime: the cpu time spent in kernel space since the last update
4078 void account_system_time_scaled(struct task_struct
*p
, cputime_t cputime
)
4080 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime
);
4084 * Account for involuntary wait time.
4085 * @p: the process from which the cpu time has been stolen
4086 * @steal: the cpu time spent in involuntary wait
4088 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
4090 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4091 cputime64_t tmp
= cputime_to_cputime64(steal
);
4092 struct rq
*rq
= this_rq();
4094 if (p
== rq
->idle
) {
4095 p
->stime
= cputime_add(p
->stime
, steal
);
4096 if (atomic_read(&rq
->nr_iowait
) > 0)
4097 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
4099 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
4101 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
4105 * This function gets called by the timer code, with HZ frequency.
4106 * We call it with interrupts disabled.
4108 * It also gets called by the fork code, when changing the parent's
4111 void scheduler_tick(void)
4113 int cpu
= smp_processor_id();
4114 struct rq
*rq
= cpu_rq(cpu
);
4115 struct task_struct
*curr
= rq
->curr
;
4119 spin_lock(&rq
->lock
);
4120 update_rq_clock(rq
);
4121 update_cpu_load(rq
);
4122 curr
->sched_class
->task_tick(rq
, curr
, 0);
4123 spin_unlock(&rq
->lock
);
4126 rq
->idle_at_tick
= idle_cpu(cpu
);
4127 trigger_load_balance(rq
, cpu
);
4131 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
4133 void __kprobes
add_preempt_count(int val
)
4138 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4140 preempt_count() += val
;
4142 * Spinlock count overflowing soon?
4144 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4147 EXPORT_SYMBOL(add_preempt_count
);
4149 void __kprobes
sub_preempt_count(int val
)
4154 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
4157 * Is the spinlock portion underflowing?
4159 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4160 !(preempt_count() & PREEMPT_MASK
)))
4163 preempt_count() -= val
;
4165 EXPORT_SYMBOL(sub_preempt_count
);
4170 * Print scheduling while atomic bug:
4172 static noinline
void __schedule_bug(struct task_struct
*prev
)
4174 struct pt_regs
*regs
= get_irq_regs();
4176 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4177 prev
->comm
, prev
->pid
, preempt_count());
4179 debug_show_held_locks(prev
);
4181 if (irqs_disabled())
4182 print_irqtrace_events(prev
);
4191 * Various schedule()-time debugging checks and statistics:
4193 static inline void schedule_debug(struct task_struct
*prev
)
4196 * Test if we are atomic. Since do_exit() needs to call into
4197 * schedule() atomically, we ignore that path for now.
4198 * Otherwise, whine if we are scheduling when we should not be.
4200 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
4201 __schedule_bug(prev
);
4203 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4205 schedstat_inc(this_rq(), sched_count
);
4206 #ifdef CONFIG_SCHEDSTATS
4207 if (unlikely(prev
->lock_depth
>= 0)) {
4208 schedstat_inc(this_rq(), bkl_count
);
4209 schedstat_inc(prev
, sched_info
.bkl_count
);
4215 * Pick up the highest-prio task:
4217 static inline struct task_struct
*
4218 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
4220 const struct sched_class
*class;
4221 struct task_struct
*p
;
4224 * Optimization: we know that if all tasks are in
4225 * the fair class we can call that function directly:
4227 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
4228 p
= fair_sched_class
.pick_next_task(rq
);
4233 class = sched_class_highest
;
4235 p
= class->pick_next_task(rq
);
4239 * Will never be NULL as the idle class always
4240 * returns a non-NULL p:
4242 class = class->next
;
4247 * schedule() is the main scheduler function.
4249 asmlinkage
void __sched
schedule(void)
4251 struct task_struct
*prev
, *next
;
4252 unsigned long *switch_count
;
4254 int cpu
, hrtick
= sched_feat(HRTICK
);
4258 cpu
= smp_processor_id();
4262 switch_count
= &prev
->nivcsw
;
4264 release_kernel_lock(prev
);
4265 need_resched_nonpreemptible
:
4267 schedule_debug(prev
);
4273 * Do the rq-clock update outside the rq lock:
4275 local_irq_disable();
4276 update_rq_clock(rq
);
4277 spin_lock(&rq
->lock
);
4278 clear_tsk_need_resched(prev
);
4280 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
4281 if (unlikely(signal_pending_state(prev
->state
, prev
)))
4282 prev
->state
= TASK_RUNNING
;
4284 deactivate_task(rq
, prev
, 1);
4285 switch_count
= &prev
->nvcsw
;
4289 if (prev
->sched_class
->pre_schedule
)
4290 prev
->sched_class
->pre_schedule(rq
, prev
);
4293 if (unlikely(!rq
->nr_running
))
4294 idle_balance(cpu
, rq
);
4296 prev
->sched_class
->put_prev_task(rq
, prev
);
4297 next
= pick_next_task(rq
, prev
);
4299 if (likely(prev
!= next
)) {
4300 sched_info_switch(prev
, next
);
4306 context_switch(rq
, prev
, next
); /* unlocks the rq */
4308 * the context switch might have flipped the stack from under
4309 * us, hence refresh the local variables.
4311 cpu
= smp_processor_id();
4314 spin_unlock_irq(&rq
->lock
);
4319 if (unlikely(reacquire_kernel_lock(current
) < 0))
4320 goto need_resched_nonpreemptible
;
4322 preempt_enable_no_resched();
4323 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
4326 EXPORT_SYMBOL(schedule
);
4328 #ifdef CONFIG_PREEMPT
4330 * this is the entry point to schedule() from in-kernel preemption
4331 * off of preempt_enable. Kernel preemptions off return from interrupt
4332 * occur there and call schedule directly.
4334 asmlinkage
void __sched
preempt_schedule(void)
4336 struct thread_info
*ti
= current_thread_info();
4339 * If there is a non-zero preempt_count or interrupts are disabled,
4340 * we do not want to preempt the current task. Just return..
4342 if (likely(ti
->preempt_count
|| irqs_disabled()))
4346 add_preempt_count(PREEMPT_ACTIVE
);
4348 sub_preempt_count(PREEMPT_ACTIVE
);
4351 * Check again in case we missed a preemption opportunity
4352 * between schedule and now.
4355 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4357 EXPORT_SYMBOL(preempt_schedule
);
4360 * this is the entry point to schedule() from kernel preemption
4361 * off of irq context.
4362 * Note, that this is called and return with irqs disabled. This will
4363 * protect us against recursive calling from irq.
4365 asmlinkage
void __sched
preempt_schedule_irq(void)
4367 struct thread_info
*ti
= current_thread_info();
4369 /* Catch callers which need to be fixed */
4370 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
4373 add_preempt_count(PREEMPT_ACTIVE
);
4376 local_irq_disable();
4377 sub_preempt_count(PREEMPT_ACTIVE
);
4380 * Check again in case we missed a preemption opportunity
4381 * between schedule and now.
4384 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4387 #endif /* CONFIG_PREEMPT */
4389 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
4392 return try_to_wake_up(curr
->private, mode
, sync
);
4394 EXPORT_SYMBOL(default_wake_function
);
4397 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4398 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4399 * number) then we wake all the non-exclusive tasks and one exclusive task.
4401 * There are circumstances in which we can try to wake a task which has already
4402 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4403 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4405 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4406 int nr_exclusive
, int sync
, void *key
)
4408 wait_queue_t
*curr
, *next
;
4410 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4411 unsigned flags
= curr
->flags
;
4413 if (curr
->func(curr
, mode
, sync
, key
) &&
4414 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4420 * __wake_up - wake up threads blocked on a waitqueue.
4422 * @mode: which threads
4423 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4424 * @key: is directly passed to the wakeup function
4426 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4427 int nr_exclusive
, void *key
)
4429 unsigned long flags
;
4431 spin_lock_irqsave(&q
->lock
, flags
);
4432 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4433 spin_unlock_irqrestore(&q
->lock
, flags
);
4435 EXPORT_SYMBOL(__wake_up
);
4438 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4440 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4442 __wake_up_common(q
, mode
, 1, 0, NULL
);
4446 * __wake_up_sync - wake up threads blocked on a waitqueue.
4448 * @mode: which threads
4449 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4451 * The sync wakeup differs that the waker knows that it will schedule
4452 * away soon, so while the target thread will be woken up, it will not
4453 * be migrated to another CPU - ie. the two threads are 'synchronized'
4454 * with each other. This can prevent needless bouncing between CPUs.
4456 * On UP it can prevent extra preemption.
4459 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4461 unsigned long flags
;
4467 if (unlikely(!nr_exclusive
))
4470 spin_lock_irqsave(&q
->lock
, flags
);
4471 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
4472 spin_unlock_irqrestore(&q
->lock
, flags
);
4474 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4476 void complete(struct completion
*x
)
4478 unsigned long flags
;
4480 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4482 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4483 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4485 EXPORT_SYMBOL(complete
);
4487 void complete_all(struct completion
*x
)
4489 unsigned long flags
;
4491 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4492 x
->done
+= UINT_MAX
/2;
4493 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4494 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4496 EXPORT_SYMBOL(complete_all
);
4498 static inline long __sched
4499 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4502 DECLARE_WAITQUEUE(wait
, current
);
4504 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4505 __add_wait_queue_tail(&x
->wait
, &wait
);
4507 if ((state
== TASK_INTERRUPTIBLE
&&
4508 signal_pending(current
)) ||
4509 (state
== TASK_KILLABLE
&&
4510 fatal_signal_pending(current
))) {
4511 timeout
= -ERESTARTSYS
;
4514 __set_current_state(state
);
4515 spin_unlock_irq(&x
->wait
.lock
);
4516 timeout
= schedule_timeout(timeout
);
4517 spin_lock_irq(&x
->wait
.lock
);
4518 } while (!x
->done
&& timeout
);
4519 __remove_wait_queue(&x
->wait
, &wait
);
4524 return timeout
?: 1;
4528 wait_for_common(struct completion
*x
, long timeout
, int state
)
4532 spin_lock_irq(&x
->wait
.lock
);
4533 timeout
= do_wait_for_common(x
, timeout
, state
);
4534 spin_unlock_irq(&x
->wait
.lock
);
4538 void __sched
wait_for_completion(struct completion
*x
)
4540 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4542 EXPORT_SYMBOL(wait_for_completion
);
4544 unsigned long __sched
4545 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4547 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4549 EXPORT_SYMBOL(wait_for_completion_timeout
);
4551 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4553 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4554 if (t
== -ERESTARTSYS
)
4558 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4560 unsigned long __sched
4561 wait_for_completion_interruptible_timeout(struct completion
*x
,
4562 unsigned long timeout
)
4564 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4566 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4568 int __sched
wait_for_completion_killable(struct completion
*x
)
4570 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4571 if (t
== -ERESTARTSYS
)
4575 EXPORT_SYMBOL(wait_for_completion_killable
);
4578 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4580 unsigned long flags
;
4583 init_waitqueue_entry(&wait
, current
);
4585 __set_current_state(state
);
4587 spin_lock_irqsave(&q
->lock
, flags
);
4588 __add_wait_queue(q
, &wait
);
4589 spin_unlock(&q
->lock
);
4590 timeout
= schedule_timeout(timeout
);
4591 spin_lock_irq(&q
->lock
);
4592 __remove_wait_queue(q
, &wait
);
4593 spin_unlock_irqrestore(&q
->lock
, flags
);
4598 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4600 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4602 EXPORT_SYMBOL(interruptible_sleep_on
);
4605 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4607 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4609 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4611 void __sched
sleep_on(wait_queue_head_t
*q
)
4613 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4615 EXPORT_SYMBOL(sleep_on
);
4617 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4619 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4621 EXPORT_SYMBOL(sleep_on_timeout
);
4623 #ifdef CONFIG_RT_MUTEXES
4626 * rt_mutex_setprio - set the current priority of a task
4628 * @prio: prio value (kernel-internal form)
4630 * This function changes the 'effective' priority of a task. It does
4631 * not touch ->normal_prio like __setscheduler().
4633 * Used by the rt_mutex code to implement priority inheritance logic.
4635 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4637 unsigned long flags
;
4638 int oldprio
, on_rq
, running
;
4640 const struct sched_class
*prev_class
= p
->sched_class
;
4642 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4644 rq
= task_rq_lock(p
, &flags
);
4645 update_rq_clock(rq
);
4648 on_rq
= p
->se
.on_rq
;
4649 running
= task_current(rq
, p
);
4651 dequeue_task(rq
, p
, 0);
4653 p
->sched_class
->put_prev_task(rq
, p
);
4656 p
->sched_class
= &rt_sched_class
;
4658 p
->sched_class
= &fair_sched_class
;
4663 p
->sched_class
->set_curr_task(rq
);
4665 enqueue_task(rq
, p
, 0);
4667 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4669 task_rq_unlock(rq
, &flags
);
4674 void set_user_nice(struct task_struct
*p
, long nice
)
4676 int old_prio
, delta
, on_rq
;
4677 unsigned long flags
;
4680 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4683 * We have to be careful, if called from sys_setpriority(),
4684 * the task might be in the middle of scheduling on another CPU.
4686 rq
= task_rq_lock(p
, &flags
);
4687 update_rq_clock(rq
);
4689 * The RT priorities are set via sched_setscheduler(), but we still
4690 * allow the 'normal' nice value to be set - but as expected
4691 * it wont have any effect on scheduling until the task is
4692 * SCHED_FIFO/SCHED_RR:
4694 if (task_has_rt_policy(p
)) {
4695 p
->static_prio
= NICE_TO_PRIO(nice
);
4698 on_rq
= p
->se
.on_rq
;
4700 dequeue_task(rq
, p
, 0);
4702 p
->static_prio
= NICE_TO_PRIO(nice
);
4705 p
->prio
= effective_prio(p
);
4706 delta
= p
->prio
- old_prio
;
4709 enqueue_task(rq
, p
, 0);
4711 * If the task increased its priority or is running and
4712 * lowered its priority, then reschedule its CPU:
4714 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4715 resched_task(rq
->curr
);
4718 task_rq_unlock(rq
, &flags
);
4720 EXPORT_SYMBOL(set_user_nice
);
4723 * can_nice - check if a task can reduce its nice value
4727 int can_nice(const struct task_struct
*p
, const int nice
)
4729 /* convert nice value [19,-20] to rlimit style value [1,40] */
4730 int nice_rlim
= 20 - nice
;
4732 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4733 capable(CAP_SYS_NICE
));
4736 #ifdef __ARCH_WANT_SYS_NICE
4739 * sys_nice - change the priority of the current process.
4740 * @increment: priority increment
4742 * sys_setpriority is a more generic, but much slower function that
4743 * does similar things.
4745 asmlinkage
long sys_nice(int increment
)
4750 * Setpriority might change our priority at the same moment.
4751 * We don't have to worry. Conceptually one call occurs first
4752 * and we have a single winner.
4754 if (increment
< -40)
4759 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4765 if (increment
< 0 && !can_nice(current
, nice
))
4768 retval
= security_task_setnice(current
, nice
);
4772 set_user_nice(current
, nice
);
4779 * task_prio - return the priority value of a given task.
4780 * @p: the task in question.
4782 * This is the priority value as seen by users in /proc.
4783 * RT tasks are offset by -200. Normal tasks are centered
4784 * around 0, value goes from -16 to +15.
4786 int task_prio(const struct task_struct
*p
)
4788 return p
->prio
- MAX_RT_PRIO
;
4792 * task_nice - return the nice value of a given task.
4793 * @p: the task in question.
4795 int task_nice(const struct task_struct
*p
)
4797 return TASK_NICE(p
);
4799 EXPORT_SYMBOL(task_nice
);
4802 * idle_cpu - is a given cpu idle currently?
4803 * @cpu: the processor in question.
4805 int idle_cpu(int cpu
)
4807 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4811 * idle_task - return the idle task for a given cpu.
4812 * @cpu: the processor in question.
4814 struct task_struct
*idle_task(int cpu
)
4816 return cpu_rq(cpu
)->idle
;
4820 * find_process_by_pid - find a process with a matching PID value.
4821 * @pid: the pid in question.
4823 static struct task_struct
*find_process_by_pid(pid_t pid
)
4825 return pid
? find_task_by_vpid(pid
) : current
;
4828 /* Actually do priority change: must hold rq lock. */
4830 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4832 BUG_ON(p
->se
.on_rq
);
4835 switch (p
->policy
) {
4839 p
->sched_class
= &fair_sched_class
;
4843 p
->sched_class
= &rt_sched_class
;
4847 p
->rt_priority
= prio
;
4848 p
->normal_prio
= normal_prio(p
);
4849 /* we are holding p->pi_lock already */
4850 p
->prio
= rt_mutex_getprio(p
);
4855 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4856 * @p: the task in question.
4857 * @policy: new policy.
4858 * @param: structure containing the new RT priority.
4860 * NOTE that the task may be already dead.
4862 int sched_setscheduler(struct task_struct
*p
, int policy
,
4863 struct sched_param
*param
)
4865 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4866 unsigned long flags
;
4867 const struct sched_class
*prev_class
= p
->sched_class
;
4870 /* may grab non-irq protected spin_locks */
4871 BUG_ON(in_interrupt());
4873 /* double check policy once rq lock held */
4875 policy
= oldpolicy
= p
->policy
;
4876 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4877 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4878 policy
!= SCHED_IDLE
)
4881 * Valid priorities for SCHED_FIFO and SCHED_RR are
4882 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4883 * SCHED_BATCH and SCHED_IDLE is 0.
4885 if (param
->sched_priority
< 0 ||
4886 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4887 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4889 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4893 * Allow unprivileged RT tasks to decrease priority:
4895 if (!capable(CAP_SYS_NICE
)) {
4896 if (rt_policy(policy
)) {
4897 unsigned long rlim_rtprio
;
4899 if (!lock_task_sighand(p
, &flags
))
4901 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4902 unlock_task_sighand(p
, &flags
);
4904 /* can't set/change the rt policy */
4905 if (policy
!= p
->policy
&& !rlim_rtprio
)
4908 /* can't increase priority */
4909 if (param
->sched_priority
> p
->rt_priority
&&
4910 param
->sched_priority
> rlim_rtprio
)
4914 * Like positive nice levels, dont allow tasks to
4915 * move out of SCHED_IDLE either:
4917 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4920 /* can't change other user's priorities */
4921 if ((current
->euid
!= p
->euid
) &&
4922 (current
->euid
!= p
->uid
))
4926 #ifdef CONFIG_RT_GROUP_SCHED
4928 * Do not allow realtime tasks into groups that have no runtime
4931 if (rt_policy(policy
) && task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
4935 retval
= security_task_setscheduler(p
, policy
, param
);
4939 * make sure no PI-waiters arrive (or leave) while we are
4940 * changing the priority of the task:
4942 spin_lock_irqsave(&p
->pi_lock
, flags
);
4944 * To be able to change p->policy safely, the apropriate
4945 * runqueue lock must be held.
4947 rq
= __task_rq_lock(p
);
4948 /* recheck policy now with rq lock held */
4949 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4950 policy
= oldpolicy
= -1;
4951 __task_rq_unlock(rq
);
4952 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4955 update_rq_clock(rq
);
4956 on_rq
= p
->se
.on_rq
;
4957 running
= task_current(rq
, p
);
4959 deactivate_task(rq
, p
, 0);
4961 p
->sched_class
->put_prev_task(rq
, p
);
4964 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4967 p
->sched_class
->set_curr_task(rq
);
4969 activate_task(rq
, p
, 0);
4971 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4973 __task_rq_unlock(rq
);
4974 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4976 rt_mutex_adjust_pi(p
);
4980 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4983 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4985 struct sched_param lparam
;
4986 struct task_struct
*p
;
4989 if (!param
|| pid
< 0)
4991 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4996 p
= find_process_by_pid(pid
);
4998 retval
= sched_setscheduler(p
, policy
, &lparam
);
5005 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5006 * @pid: the pid in question.
5007 * @policy: new policy.
5008 * @param: structure containing the new RT priority.
5011 sys_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5013 /* negative values for policy are not valid */
5017 return do_sched_setscheduler(pid
, policy
, param
);
5021 * sys_sched_setparam - set/change the RT priority of a thread
5022 * @pid: the pid in question.
5023 * @param: structure containing the new RT priority.
5025 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
5027 return do_sched_setscheduler(pid
, -1, param
);
5031 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5032 * @pid: the pid in question.
5034 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
5036 struct task_struct
*p
;
5043 read_lock(&tasklist_lock
);
5044 p
= find_process_by_pid(pid
);
5046 retval
= security_task_getscheduler(p
);
5050 read_unlock(&tasklist_lock
);
5055 * sys_sched_getscheduler - get the RT priority of a thread
5056 * @pid: the pid in question.
5057 * @param: structure containing the RT priority.
5059 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
5061 struct sched_param lp
;
5062 struct task_struct
*p
;
5065 if (!param
|| pid
< 0)
5068 read_lock(&tasklist_lock
);
5069 p
= find_process_by_pid(pid
);
5074 retval
= security_task_getscheduler(p
);
5078 lp
.sched_priority
= p
->rt_priority
;
5079 read_unlock(&tasklist_lock
);
5082 * This one might sleep, we cannot do it with a spinlock held ...
5084 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
5089 read_unlock(&tasklist_lock
);
5093 long sched_setaffinity(pid_t pid
, const cpumask_t
*in_mask
)
5095 cpumask_t cpus_allowed
;
5096 cpumask_t new_mask
= *in_mask
;
5097 struct task_struct
*p
;
5101 read_lock(&tasklist_lock
);
5103 p
= find_process_by_pid(pid
);
5105 read_unlock(&tasklist_lock
);
5111 * It is not safe to call set_cpus_allowed with the
5112 * tasklist_lock held. We will bump the task_struct's
5113 * usage count and then drop tasklist_lock.
5116 read_unlock(&tasklist_lock
);
5119 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
5120 !capable(CAP_SYS_NICE
))
5123 retval
= security_task_setscheduler(p
, 0, NULL
);
5127 cpuset_cpus_allowed(p
, &cpus_allowed
);
5128 cpus_and(new_mask
, new_mask
, cpus_allowed
);
5130 retval
= set_cpus_allowed_ptr(p
, &new_mask
);
5133 cpuset_cpus_allowed(p
, &cpus_allowed
);
5134 if (!cpus_subset(new_mask
, cpus_allowed
)) {
5136 * We must have raced with a concurrent cpuset
5137 * update. Just reset the cpus_allowed to the
5138 * cpuset's cpus_allowed
5140 new_mask
= cpus_allowed
;
5150 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5151 cpumask_t
*new_mask
)
5153 if (len
< sizeof(cpumask_t
)) {
5154 memset(new_mask
, 0, sizeof(cpumask_t
));
5155 } else if (len
> sizeof(cpumask_t
)) {
5156 len
= sizeof(cpumask_t
);
5158 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
5162 * sys_sched_setaffinity - set the cpu affinity of a process
5163 * @pid: pid of the process
5164 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5165 * @user_mask_ptr: user-space pointer to the new cpu mask
5167 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
5168 unsigned long __user
*user_mask_ptr
)
5173 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
5177 return sched_setaffinity(pid
, &new_mask
);
5180 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
5182 struct task_struct
*p
;
5186 read_lock(&tasklist_lock
);
5189 p
= find_process_by_pid(pid
);
5193 retval
= security_task_getscheduler(p
);
5197 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
5200 read_unlock(&tasklist_lock
);
5207 * sys_sched_getaffinity - get the cpu affinity of a process
5208 * @pid: pid of the process
5209 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5210 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5212 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
5213 unsigned long __user
*user_mask_ptr
)
5218 if (len
< sizeof(cpumask_t
))
5221 ret
= sched_getaffinity(pid
, &mask
);
5225 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
5228 return sizeof(cpumask_t
);
5232 * sys_sched_yield - yield the current processor to other threads.
5234 * This function yields the current CPU to other tasks. If there are no
5235 * other threads running on this CPU then this function will return.
5237 asmlinkage
long sys_sched_yield(void)
5239 struct rq
*rq
= this_rq_lock();
5241 schedstat_inc(rq
, yld_count
);
5242 current
->sched_class
->yield_task(rq
);
5245 * Since we are going to call schedule() anyway, there's
5246 * no need to preempt or enable interrupts:
5248 __release(rq
->lock
);
5249 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5250 _raw_spin_unlock(&rq
->lock
);
5251 preempt_enable_no_resched();
5258 static void __cond_resched(void)
5260 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5261 __might_sleep(__FILE__
, __LINE__
);
5264 * The BKS might be reacquired before we have dropped
5265 * PREEMPT_ACTIVE, which could trigger a second
5266 * cond_resched() call.
5269 add_preempt_count(PREEMPT_ACTIVE
);
5271 sub_preempt_count(PREEMPT_ACTIVE
);
5272 } while (need_resched());
5275 int __sched
_cond_resched(void)
5277 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
5278 system_state
== SYSTEM_RUNNING
) {
5284 EXPORT_SYMBOL(_cond_resched
);
5287 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5288 * call schedule, and on return reacquire the lock.
5290 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5291 * operations here to prevent schedule() from being called twice (once via
5292 * spin_unlock(), once by hand).
5294 int cond_resched_lock(spinlock_t
*lock
)
5296 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
5299 if (spin_needbreak(lock
) || resched
) {
5301 if (resched
&& need_resched())
5310 EXPORT_SYMBOL(cond_resched_lock
);
5312 int __sched
cond_resched_softirq(void)
5314 BUG_ON(!in_softirq());
5316 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
5324 EXPORT_SYMBOL(cond_resched_softirq
);
5327 * yield - yield the current processor to other threads.
5329 * This is a shortcut for kernel-space yielding - it marks the
5330 * thread runnable and calls sys_sched_yield().
5332 void __sched
yield(void)
5334 set_current_state(TASK_RUNNING
);
5337 EXPORT_SYMBOL(yield
);
5340 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5341 * that process accounting knows that this is a task in IO wait state.
5343 * But don't do that if it is a deliberate, throttling IO wait (this task
5344 * has set its backing_dev_info: the queue against which it should throttle)
5346 void __sched
io_schedule(void)
5348 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5350 delayacct_blkio_start();
5351 atomic_inc(&rq
->nr_iowait
);
5353 atomic_dec(&rq
->nr_iowait
);
5354 delayacct_blkio_end();
5356 EXPORT_SYMBOL(io_schedule
);
5358 long __sched
io_schedule_timeout(long timeout
)
5360 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5363 delayacct_blkio_start();
5364 atomic_inc(&rq
->nr_iowait
);
5365 ret
= schedule_timeout(timeout
);
5366 atomic_dec(&rq
->nr_iowait
);
5367 delayacct_blkio_end();
5372 * sys_sched_get_priority_max - return maximum RT priority.
5373 * @policy: scheduling class.
5375 * this syscall returns the maximum rt_priority that can be used
5376 * by a given scheduling class.
5378 asmlinkage
long sys_sched_get_priority_max(int policy
)
5385 ret
= MAX_USER_RT_PRIO
-1;
5397 * sys_sched_get_priority_min - return minimum RT priority.
5398 * @policy: scheduling class.
5400 * this syscall returns the minimum rt_priority that can be used
5401 * by a given scheduling class.
5403 asmlinkage
long sys_sched_get_priority_min(int policy
)
5421 * sys_sched_rr_get_interval - return the default timeslice of a process.
5422 * @pid: pid of the process.
5423 * @interval: userspace pointer to the timeslice value.
5425 * this syscall writes the default timeslice value of a given process
5426 * into the user-space timespec buffer. A value of '0' means infinity.
5429 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
5431 struct task_struct
*p
;
5432 unsigned int time_slice
;
5440 read_lock(&tasklist_lock
);
5441 p
= find_process_by_pid(pid
);
5445 retval
= security_task_getscheduler(p
);
5450 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5451 * tasks that are on an otherwise idle runqueue:
5454 if (p
->policy
== SCHED_RR
) {
5455 time_slice
= DEF_TIMESLICE
;
5456 } else if (p
->policy
!= SCHED_FIFO
) {
5457 struct sched_entity
*se
= &p
->se
;
5458 unsigned long flags
;
5461 rq
= task_rq_lock(p
, &flags
);
5462 if (rq
->cfs
.load
.weight
)
5463 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
5464 task_rq_unlock(rq
, &flags
);
5466 read_unlock(&tasklist_lock
);
5467 jiffies_to_timespec(time_slice
, &t
);
5468 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5472 read_unlock(&tasklist_lock
);
5476 static const char stat_nam
[] = "RSDTtZX";
5478 void sched_show_task(struct task_struct
*p
)
5480 unsigned long free
= 0;
5483 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5484 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5485 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5486 #if BITS_PER_LONG == 32
5487 if (state
== TASK_RUNNING
)
5488 printk(KERN_CONT
" running ");
5490 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5492 if (state
== TASK_RUNNING
)
5493 printk(KERN_CONT
" running task ");
5495 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5497 #ifdef CONFIG_DEBUG_STACK_USAGE
5499 unsigned long *n
= end_of_stack(p
);
5502 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
5505 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
5506 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
5508 show_stack(p
, NULL
);
5511 void show_state_filter(unsigned long state_filter
)
5513 struct task_struct
*g
, *p
;
5515 #if BITS_PER_LONG == 32
5517 " task PC stack pid father\n");
5520 " task PC stack pid father\n");
5522 read_lock(&tasklist_lock
);
5523 do_each_thread(g
, p
) {
5525 * reset the NMI-timeout, listing all files on a slow
5526 * console might take alot of time:
5528 touch_nmi_watchdog();
5529 if (!state_filter
|| (p
->state
& state_filter
))
5531 } while_each_thread(g
, p
);
5533 touch_all_softlockup_watchdogs();
5535 #ifdef CONFIG_SCHED_DEBUG
5536 sysrq_sched_debug_show();
5538 read_unlock(&tasklist_lock
);
5540 * Only show locks if all tasks are dumped:
5542 if (state_filter
== -1)
5543 debug_show_all_locks();
5546 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5548 idle
->sched_class
= &idle_sched_class
;
5552 * init_idle - set up an idle thread for a given CPU
5553 * @idle: task in question
5554 * @cpu: cpu the idle task belongs to
5556 * NOTE: this function does not set the idle thread's NEED_RESCHED
5557 * flag, to make booting more robust.
5559 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5561 struct rq
*rq
= cpu_rq(cpu
);
5562 unsigned long flags
;
5565 idle
->se
.exec_start
= sched_clock();
5567 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
5568 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
5569 __set_task_cpu(idle
, cpu
);
5571 spin_lock_irqsave(&rq
->lock
, flags
);
5572 rq
->curr
= rq
->idle
= idle
;
5573 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5576 spin_unlock_irqrestore(&rq
->lock
, flags
);
5578 /* Set the preempt count _outside_ the spinlocks! */
5579 #if defined(CONFIG_PREEMPT)
5580 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5582 task_thread_info(idle
)->preempt_count
= 0;
5585 * The idle tasks have their own, simple scheduling class:
5587 idle
->sched_class
= &idle_sched_class
;
5591 * In a system that switches off the HZ timer nohz_cpu_mask
5592 * indicates which cpus entered this state. This is used
5593 * in the rcu update to wait only for active cpus. For system
5594 * which do not switch off the HZ timer nohz_cpu_mask should
5595 * always be CPU_MASK_NONE.
5597 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
5600 * Increase the granularity value when there are more CPUs,
5601 * because with more CPUs the 'effective latency' as visible
5602 * to users decreases. But the relationship is not linear,
5603 * so pick a second-best guess by going with the log2 of the
5606 * This idea comes from the SD scheduler of Con Kolivas:
5608 static inline void sched_init_granularity(void)
5610 unsigned int factor
= 1 + ilog2(num_online_cpus());
5611 const unsigned long limit
= 200000000;
5613 sysctl_sched_min_granularity
*= factor
;
5614 if (sysctl_sched_min_granularity
> limit
)
5615 sysctl_sched_min_granularity
= limit
;
5617 sysctl_sched_latency
*= factor
;
5618 if (sysctl_sched_latency
> limit
)
5619 sysctl_sched_latency
= limit
;
5621 sysctl_sched_wakeup_granularity
*= factor
;
5626 * This is how migration works:
5628 * 1) we queue a struct migration_req structure in the source CPU's
5629 * runqueue and wake up that CPU's migration thread.
5630 * 2) we down() the locked semaphore => thread blocks.
5631 * 3) migration thread wakes up (implicitly it forces the migrated
5632 * thread off the CPU)
5633 * 4) it gets the migration request and checks whether the migrated
5634 * task is still in the wrong runqueue.
5635 * 5) if it's in the wrong runqueue then the migration thread removes
5636 * it and puts it into the right queue.
5637 * 6) migration thread up()s the semaphore.
5638 * 7) we wake up and the migration is done.
5642 * Change a given task's CPU affinity. Migrate the thread to a
5643 * proper CPU and schedule it away if the CPU it's executing on
5644 * is removed from the allowed bitmask.
5646 * NOTE: the caller must have a valid reference to the task, the
5647 * task must not exit() & deallocate itself prematurely. The
5648 * call is not atomic; no spinlocks may be held.
5650 int set_cpus_allowed_ptr(struct task_struct
*p
, const cpumask_t
*new_mask
)
5652 struct migration_req req
;
5653 unsigned long flags
;
5657 rq
= task_rq_lock(p
, &flags
);
5658 if (!cpus_intersects(*new_mask
, cpu_online_map
)) {
5663 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
5664 !cpus_equal(p
->cpus_allowed
, *new_mask
))) {
5669 if (p
->sched_class
->set_cpus_allowed
)
5670 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
5672 p
->cpus_allowed
= *new_mask
;
5673 p
->rt
.nr_cpus_allowed
= cpus_weight(*new_mask
);
5676 /* Can the task run on the task's current CPU? If so, we're done */
5677 if (cpu_isset(task_cpu(p
), *new_mask
))
5680 if (migrate_task(p
, any_online_cpu(*new_mask
), &req
)) {
5681 /* Need help from migration thread: drop lock and wait. */
5682 task_rq_unlock(rq
, &flags
);
5683 wake_up_process(rq
->migration_thread
);
5684 wait_for_completion(&req
.done
);
5685 tlb_migrate_finish(p
->mm
);
5689 task_rq_unlock(rq
, &flags
);
5693 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5696 * Move (not current) task off this cpu, onto dest cpu. We're doing
5697 * this because either it can't run here any more (set_cpus_allowed()
5698 * away from this CPU, or CPU going down), or because we're
5699 * attempting to rebalance this task on exec (sched_exec).
5701 * So we race with normal scheduler movements, but that's OK, as long
5702 * as the task is no longer on this CPU.
5704 * Returns non-zero if task was successfully migrated.
5706 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5708 struct rq
*rq_dest
, *rq_src
;
5711 if (unlikely(cpu_is_offline(dest_cpu
)))
5714 rq_src
= cpu_rq(src_cpu
);
5715 rq_dest
= cpu_rq(dest_cpu
);
5717 double_rq_lock(rq_src
, rq_dest
);
5718 /* Already moved. */
5719 if (task_cpu(p
) != src_cpu
)
5721 /* Affinity changed (again). */
5722 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
5725 on_rq
= p
->se
.on_rq
;
5727 deactivate_task(rq_src
, p
, 0);
5729 set_task_cpu(p
, dest_cpu
);
5731 activate_task(rq_dest
, p
, 0);
5732 check_preempt_curr(rq_dest
, p
);
5736 double_rq_unlock(rq_src
, rq_dest
);
5741 * migration_thread - this is a highprio system thread that performs
5742 * thread migration by bumping thread off CPU then 'pushing' onto
5745 static int migration_thread(void *data
)
5747 int cpu
= (long)data
;
5751 BUG_ON(rq
->migration_thread
!= current
);
5753 set_current_state(TASK_INTERRUPTIBLE
);
5754 while (!kthread_should_stop()) {
5755 struct migration_req
*req
;
5756 struct list_head
*head
;
5758 spin_lock_irq(&rq
->lock
);
5760 if (cpu_is_offline(cpu
)) {
5761 spin_unlock_irq(&rq
->lock
);
5765 if (rq
->active_balance
) {
5766 active_load_balance(rq
, cpu
);
5767 rq
->active_balance
= 0;
5770 head
= &rq
->migration_queue
;
5772 if (list_empty(head
)) {
5773 spin_unlock_irq(&rq
->lock
);
5775 set_current_state(TASK_INTERRUPTIBLE
);
5778 req
= list_entry(head
->next
, struct migration_req
, list
);
5779 list_del_init(head
->next
);
5781 spin_unlock(&rq
->lock
);
5782 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5785 complete(&req
->done
);
5787 __set_current_state(TASK_RUNNING
);
5791 /* Wait for kthread_stop */
5792 set_current_state(TASK_INTERRUPTIBLE
);
5793 while (!kthread_should_stop()) {
5795 set_current_state(TASK_INTERRUPTIBLE
);
5797 __set_current_state(TASK_RUNNING
);
5801 #ifdef CONFIG_HOTPLUG_CPU
5803 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5807 local_irq_disable();
5808 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
5814 * Figure out where task on dead CPU should go, use force if necessary.
5815 * NOTE: interrupts should be disabled by the caller
5817 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5819 unsigned long flags
;
5826 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5827 cpus_and(mask
, mask
, p
->cpus_allowed
);
5828 dest_cpu
= any_online_cpu(mask
);
5830 /* On any allowed CPU? */
5831 if (dest_cpu
>= nr_cpu_ids
)
5832 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5834 /* No more Mr. Nice Guy. */
5835 if (dest_cpu
>= nr_cpu_ids
) {
5836 cpumask_t cpus_allowed
;
5838 cpuset_cpus_allowed_locked(p
, &cpus_allowed
);
5840 * Try to stay on the same cpuset, where the
5841 * current cpuset may be a subset of all cpus.
5842 * The cpuset_cpus_allowed_locked() variant of
5843 * cpuset_cpus_allowed() will not block. It must be
5844 * called within calls to cpuset_lock/cpuset_unlock.
5846 rq
= task_rq_lock(p
, &flags
);
5847 p
->cpus_allowed
= cpus_allowed
;
5848 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5849 task_rq_unlock(rq
, &flags
);
5852 * Don't tell them about moving exiting tasks or
5853 * kernel threads (both mm NULL), since they never
5856 if (p
->mm
&& printk_ratelimit()) {
5857 printk(KERN_INFO
"process %d (%s) no "
5858 "longer affine to cpu%d\n",
5859 task_pid_nr(p
), p
->comm
, dead_cpu
);
5862 } while (!__migrate_task_irq(p
, dead_cpu
, dest_cpu
));
5866 * While a dead CPU has no uninterruptible tasks queued at this point,
5867 * it might still have a nonzero ->nr_uninterruptible counter, because
5868 * for performance reasons the counter is not stricly tracking tasks to
5869 * their home CPUs. So we just add the counter to another CPU's counter,
5870 * to keep the global sum constant after CPU-down:
5872 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5874 struct rq
*rq_dest
= cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR
));
5875 unsigned long flags
;
5877 local_irq_save(flags
);
5878 double_rq_lock(rq_src
, rq_dest
);
5879 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5880 rq_src
->nr_uninterruptible
= 0;
5881 double_rq_unlock(rq_src
, rq_dest
);
5882 local_irq_restore(flags
);
5885 /* Run through task list and migrate tasks from the dead cpu. */
5886 static void migrate_live_tasks(int src_cpu
)
5888 struct task_struct
*p
, *t
;
5890 read_lock(&tasklist_lock
);
5892 do_each_thread(t
, p
) {
5896 if (task_cpu(p
) == src_cpu
)
5897 move_task_off_dead_cpu(src_cpu
, p
);
5898 } while_each_thread(t
, p
);
5900 read_unlock(&tasklist_lock
);
5904 * Schedules idle task to be the next runnable task on current CPU.
5905 * It does so by boosting its priority to highest possible.
5906 * Used by CPU offline code.
5908 void sched_idle_next(void)
5910 int this_cpu
= smp_processor_id();
5911 struct rq
*rq
= cpu_rq(this_cpu
);
5912 struct task_struct
*p
= rq
->idle
;
5913 unsigned long flags
;
5915 /* cpu has to be offline */
5916 BUG_ON(cpu_online(this_cpu
));
5919 * Strictly not necessary since rest of the CPUs are stopped by now
5920 * and interrupts disabled on the current cpu.
5922 spin_lock_irqsave(&rq
->lock
, flags
);
5924 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5926 update_rq_clock(rq
);
5927 activate_task(rq
, p
, 0);
5929 spin_unlock_irqrestore(&rq
->lock
, flags
);
5933 * Ensures that the idle task is using init_mm right before its cpu goes
5936 void idle_task_exit(void)
5938 struct mm_struct
*mm
= current
->active_mm
;
5940 BUG_ON(cpu_online(smp_processor_id()));
5943 switch_mm(mm
, &init_mm
, current
);
5947 /* called under rq->lock with disabled interrupts */
5948 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5950 struct rq
*rq
= cpu_rq(dead_cpu
);
5952 /* Must be exiting, otherwise would be on tasklist. */
5953 BUG_ON(!p
->exit_state
);
5955 /* Cannot have done final schedule yet: would have vanished. */
5956 BUG_ON(p
->state
== TASK_DEAD
);
5961 * Drop lock around migration; if someone else moves it,
5962 * that's OK. No task can be added to this CPU, so iteration is
5965 spin_unlock_irq(&rq
->lock
);
5966 move_task_off_dead_cpu(dead_cpu
, p
);
5967 spin_lock_irq(&rq
->lock
);
5972 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5973 static void migrate_dead_tasks(unsigned int dead_cpu
)
5975 struct rq
*rq
= cpu_rq(dead_cpu
);
5976 struct task_struct
*next
;
5979 if (!rq
->nr_running
)
5981 update_rq_clock(rq
);
5982 next
= pick_next_task(rq
, rq
->curr
);
5985 migrate_dead(dead_cpu
, next
);
5989 #endif /* CONFIG_HOTPLUG_CPU */
5991 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5993 static struct ctl_table sd_ctl_dir
[] = {
5995 .procname
= "sched_domain",
6001 static struct ctl_table sd_ctl_root
[] = {
6003 .ctl_name
= CTL_KERN
,
6004 .procname
= "kernel",
6006 .child
= sd_ctl_dir
,
6011 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
6013 struct ctl_table
*entry
=
6014 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
6019 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
6021 struct ctl_table
*entry
;
6024 * In the intermediate directories, both the child directory and
6025 * procname are dynamically allocated and could fail but the mode
6026 * will always be set. In the lowest directory the names are
6027 * static strings and all have proc handlers.
6029 for (entry
= *tablep
; entry
->mode
; entry
++) {
6031 sd_free_ctl_entry(&entry
->child
);
6032 if (entry
->proc_handler
== NULL
)
6033 kfree(entry
->procname
);
6041 set_table_entry(struct ctl_table
*entry
,
6042 const char *procname
, void *data
, int maxlen
,
6043 mode_t mode
, proc_handler
*proc_handler
)
6045 entry
->procname
= procname
;
6047 entry
->maxlen
= maxlen
;
6049 entry
->proc_handler
= proc_handler
;
6052 static struct ctl_table
*
6053 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
6055 struct ctl_table
*table
= sd_alloc_ctl_entry(12);
6060 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
6061 sizeof(long), 0644, proc_doulongvec_minmax
);
6062 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
6063 sizeof(long), 0644, proc_doulongvec_minmax
);
6064 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
6065 sizeof(int), 0644, proc_dointvec_minmax
);
6066 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
6067 sizeof(int), 0644, proc_dointvec_minmax
);
6068 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
6069 sizeof(int), 0644, proc_dointvec_minmax
);
6070 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
6071 sizeof(int), 0644, proc_dointvec_minmax
);
6072 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
6073 sizeof(int), 0644, proc_dointvec_minmax
);
6074 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
6075 sizeof(int), 0644, proc_dointvec_minmax
);
6076 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
6077 sizeof(int), 0644, proc_dointvec_minmax
);
6078 set_table_entry(&table
[9], "cache_nice_tries",
6079 &sd
->cache_nice_tries
,
6080 sizeof(int), 0644, proc_dointvec_minmax
);
6081 set_table_entry(&table
[10], "flags", &sd
->flags
,
6082 sizeof(int), 0644, proc_dointvec_minmax
);
6083 /* &table[11] is terminator */
6088 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
6090 struct ctl_table
*entry
, *table
;
6091 struct sched_domain
*sd
;
6092 int domain_num
= 0, i
;
6095 for_each_domain(cpu
, sd
)
6097 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
6102 for_each_domain(cpu
, sd
) {
6103 snprintf(buf
, 32, "domain%d", i
);
6104 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6106 entry
->child
= sd_alloc_ctl_domain_table(sd
);
6113 static struct ctl_table_header
*sd_sysctl_header
;
6114 static void register_sched_domain_sysctl(void)
6116 int i
, cpu_num
= num_online_cpus();
6117 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
6120 WARN_ON(sd_ctl_dir
[0].child
);
6121 sd_ctl_dir
[0].child
= entry
;
6126 for_each_online_cpu(i
) {
6127 snprintf(buf
, 32, "cpu%d", i
);
6128 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6130 entry
->child
= sd_alloc_ctl_cpu_table(i
);
6134 WARN_ON(sd_sysctl_header
);
6135 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
6138 /* may be called multiple times per register */
6139 static void unregister_sched_domain_sysctl(void)
6141 if (sd_sysctl_header
)
6142 unregister_sysctl_table(sd_sysctl_header
);
6143 sd_sysctl_header
= NULL
;
6144 if (sd_ctl_dir
[0].child
)
6145 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
6148 static void register_sched_domain_sysctl(void)
6151 static void unregister_sched_domain_sysctl(void)
6156 static void set_rq_online(struct rq
*rq
)
6159 const struct sched_class
*class;
6161 cpu_set(rq
->cpu
, rq
->rd
->online
);
6164 for_each_class(class) {
6165 if (class->rq_online
)
6166 class->rq_online(rq
);
6171 static void set_rq_offline(struct rq
*rq
)
6174 const struct sched_class
*class;
6176 for_each_class(class) {
6177 if (class->rq_offline
)
6178 class->rq_offline(rq
);
6181 cpu_clear(rq
->cpu
, rq
->rd
->online
);
6187 * migration_call - callback that gets triggered when a CPU is added.
6188 * Here we can start up the necessary migration thread for the new CPU.
6190 static int __cpuinit
6191 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
6193 struct task_struct
*p
;
6194 int cpu
= (long)hcpu
;
6195 unsigned long flags
;
6200 case CPU_UP_PREPARE
:
6201 case CPU_UP_PREPARE_FROZEN
:
6202 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
6205 kthread_bind(p
, cpu
);
6206 /* Must be high prio: stop_machine expects to yield to it. */
6207 rq
= task_rq_lock(p
, &flags
);
6208 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6209 task_rq_unlock(rq
, &flags
);
6210 cpu_rq(cpu
)->migration_thread
= p
;
6214 case CPU_ONLINE_FROZEN
:
6215 /* Strictly unnecessary, as first user will wake it. */
6216 wake_up_process(cpu_rq(cpu
)->migration_thread
);
6218 /* Update our root-domain */
6220 spin_lock_irqsave(&rq
->lock
, flags
);
6222 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6226 spin_unlock_irqrestore(&rq
->lock
, flags
);
6229 #ifdef CONFIG_HOTPLUG_CPU
6230 case CPU_UP_CANCELED
:
6231 case CPU_UP_CANCELED_FROZEN
:
6232 if (!cpu_rq(cpu
)->migration_thread
)
6234 /* Unbind it from offline cpu so it can run. Fall thru. */
6235 kthread_bind(cpu_rq(cpu
)->migration_thread
,
6236 any_online_cpu(cpu_online_map
));
6237 kthread_stop(cpu_rq(cpu
)->migration_thread
);
6238 cpu_rq(cpu
)->migration_thread
= NULL
;
6242 case CPU_DEAD_FROZEN
:
6243 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6244 migrate_live_tasks(cpu
);
6246 kthread_stop(rq
->migration_thread
);
6247 rq
->migration_thread
= NULL
;
6248 /* Idle task back to normal (off runqueue, low prio) */
6249 spin_lock_irq(&rq
->lock
);
6250 update_rq_clock(rq
);
6251 deactivate_task(rq
, rq
->idle
, 0);
6252 rq
->idle
->static_prio
= MAX_PRIO
;
6253 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
6254 rq
->idle
->sched_class
= &idle_sched_class
;
6255 migrate_dead_tasks(cpu
);
6256 spin_unlock_irq(&rq
->lock
);
6258 migrate_nr_uninterruptible(rq
);
6259 BUG_ON(rq
->nr_running
!= 0);
6262 * No need to migrate the tasks: it was best-effort if
6263 * they didn't take sched_hotcpu_mutex. Just wake up
6266 spin_lock_irq(&rq
->lock
);
6267 while (!list_empty(&rq
->migration_queue
)) {
6268 struct migration_req
*req
;
6270 req
= list_entry(rq
->migration_queue
.next
,
6271 struct migration_req
, list
);
6272 list_del_init(&req
->list
);
6273 complete(&req
->done
);
6275 spin_unlock_irq(&rq
->lock
);
6279 case CPU_DYING_FROZEN
:
6280 /* Update our root-domain */
6282 spin_lock_irqsave(&rq
->lock
, flags
);
6284 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6287 spin_unlock_irqrestore(&rq
->lock
, flags
);
6294 /* Register at highest priority so that task migration (migrate_all_tasks)
6295 * happens before everything else.
6297 static struct notifier_block __cpuinitdata migration_notifier
= {
6298 .notifier_call
= migration_call
,
6302 void __init
migration_init(void)
6304 void *cpu
= (void *)(long)smp_processor_id();
6307 /* Start one for the boot CPU: */
6308 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6309 BUG_ON(err
== NOTIFY_BAD
);
6310 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6311 register_cpu_notifier(&migration_notifier
);
6317 #ifdef CONFIG_SCHED_DEBUG
6319 static inline const char *sd_level_to_string(enum sched_domain_level lvl
)
6332 case SD_LV_ALLNODES
:
6341 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6342 cpumask_t
*groupmask
)
6344 struct sched_group
*group
= sd
->groups
;
6347 cpulist_scnprintf(str
, sizeof(str
), sd
->span
);
6348 cpus_clear(*groupmask
);
6350 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6352 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6353 printk("does not load-balance\n");
6355 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6360 printk(KERN_CONT
"span %s level %s\n",
6361 str
, sd_level_to_string(sd
->level
));
6363 if (!cpu_isset(cpu
, sd
->span
)) {
6364 printk(KERN_ERR
"ERROR: domain->span does not contain "
6367 if (!cpu_isset(cpu
, group
->cpumask
)) {
6368 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6372 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6376 printk(KERN_ERR
"ERROR: group is NULL\n");
6380 if (!group
->__cpu_power
) {
6381 printk(KERN_CONT
"\n");
6382 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6387 if (!cpus_weight(group
->cpumask
)) {
6388 printk(KERN_CONT
"\n");
6389 printk(KERN_ERR
"ERROR: empty group\n");
6393 if (cpus_intersects(*groupmask
, group
->cpumask
)) {
6394 printk(KERN_CONT
"\n");
6395 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6399 cpus_or(*groupmask
, *groupmask
, group
->cpumask
);
6401 cpulist_scnprintf(str
, sizeof(str
), group
->cpumask
);
6402 printk(KERN_CONT
" %s", str
);
6404 group
= group
->next
;
6405 } while (group
!= sd
->groups
);
6406 printk(KERN_CONT
"\n");
6408 if (!cpus_equal(sd
->span
, *groupmask
))
6409 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6411 if (sd
->parent
&& !cpus_subset(*groupmask
, sd
->parent
->span
))
6412 printk(KERN_ERR
"ERROR: parent span is not a superset "
6413 "of domain->span\n");
6417 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6419 cpumask_t
*groupmask
;
6423 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6427 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6429 groupmask
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
6431 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6436 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6445 #else /* !CONFIG_SCHED_DEBUG */
6446 # define sched_domain_debug(sd, cpu) do { } while (0)
6447 #endif /* CONFIG_SCHED_DEBUG */
6449 static int sd_degenerate(struct sched_domain
*sd
)
6451 if (cpus_weight(sd
->span
) == 1)
6454 /* Following flags need at least 2 groups */
6455 if (sd
->flags
& (SD_LOAD_BALANCE
|
6456 SD_BALANCE_NEWIDLE
|
6460 SD_SHARE_PKG_RESOURCES
)) {
6461 if (sd
->groups
!= sd
->groups
->next
)
6465 /* Following flags don't use groups */
6466 if (sd
->flags
& (SD_WAKE_IDLE
|
6475 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6477 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6479 if (sd_degenerate(parent
))
6482 if (!cpus_equal(sd
->span
, parent
->span
))
6485 /* Does parent contain flags not in child? */
6486 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6487 if (cflags
& SD_WAKE_AFFINE
)
6488 pflags
&= ~SD_WAKE_BALANCE
;
6489 /* Flags needing groups don't count if only 1 group in parent */
6490 if (parent
->groups
== parent
->groups
->next
) {
6491 pflags
&= ~(SD_LOAD_BALANCE
|
6492 SD_BALANCE_NEWIDLE
|
6496 SD_SHARE_PKG_RESOURCES
);
6498 if (~cflags
& pflags
)
6504 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6506 unsigned long flags
;
6508 spin_lock_irqsave(&rq
->lock
, flags
);
6511 struct root_domain
*old_rd
= rq
->rd
;
6513 if (cpu_isset(rq
->cpu
, old_rd
->online
))
6516 cpu_clear(rq
->cpu
, old_rd
->span
);
6518 if (atomic_dec_and_test(&old_rd
->refcount
))
6522 atomic_inc(&rd
->refcount
);
6525 cpu_set(rq
->cpu
, rd
->span
);
6526 if (cpu_isset(rq
->cpu
, cpu_online_map
))
6529 spin_unlock_irqrestore(&rq
->lock
, flags
);
6532 static void init_rootdomain(struct root_domain
*rd
)
6534 memset(rd
, 0, sizeof(*rd
));
6536 cpus_clear(rd
->span
);
6537 cpus_clear(rd
->online
);
6539 cpupri_init(&rd
->cpupri
);
6542 static void init_defrootdomain(void)
6544 init_rootdomain(&def_root_domain
);
6545 atomic_set(&def_root_domain
.refcount
, 1);
6548 static struct root_domain
*alloc_rootdomain(void)
6550 struct root_domain
*rd
;
6552 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6556 init_rootdomain(rd
);
6562 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6563 * hold the hotplug lock.
6566 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6568 struct rq
*rq
= cpu_rq(cpu
);
6569 struct sched_domain
*tmp
;
6571 /* Remove the sched domains which do not contribute to scheduling. */
6572 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
6573 struct sched_domain
*parent
= tmp
->parent
;
6576 if (sd_parent_degenerate(tmp
, parent
)) {
6577 tmp
->parent
= parent
->parent
;
6579 parent
->parent
->child
= tmp
;
6583 if (sd
&& sd_degenerate(sd
)) {
6589 sched_domain_debug(sd
, cpu
);
6591 rq_attach_root(rq
, rd
);
6592 rcu_assign_pointer(rq
->sd
, sd
);
6595 /* cpus with isolated domains */
6596 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
6598 /* Setup the mask of cpus configured for isolated domains */
6599 static int __init
isolated_cpu_setup(char *str
)
6601 int ints
[NR_CPUS
], i
;
6603 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
6604 cpus_clear(cpu_isolated_map
);
6605 for (i
= 1; i
<= ints
[0]; i
++)
6606 if (ints
[i
] < NR_CPUS
)
6607 cpu_set(ints
[i
], cpu_isolated_map
);
6611 __setup("isolcpus=", isolated_cpu_setup
);
6614 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6615 * to a function which identifies what group(along with sched group) a CPU
6616 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6617 * (due to the fact that we keep track of groups covered with a cpumask_t).
6619 * init_sched_build_groups will build a circular linked list of the groups
6620 * covered by the given span, and will set each group's ->cpumask correctly,
6621 * and ->cpu_power to 0.
6624 init_sched_build_groups(const cpumask_t
*span
, const cpumask_t
*cpu_map
,
6625 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
6626 struct sched_group
**sg
,
6627 cpumask_t
*tmpmask
),
6628 cpumask_t
*covered
, cpumask_t
*tmpmask
)
6630 struct sched_group
*first
= NULL
, *last
= NULL
;
6633 cpus_clear(*covered
);
6635 for_each_cpu_mask(i
, *span
) {
6636 struct sched_group
*sg
;
6637 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6640 if (cpu_isset(i
, *covered
))
6643 cpus_clear(sg
->cpumask
);
6644 sg
->__cpu_power
= 0;
6646 for_each_cpu_mask(j
, *span
) {
6647 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6650 cpu_set(j
, *covered
);
6651 cpu_set(j
, sg
->cpumask
);
6662 #define SD_NODES_PER_DOMAIN 16
6667 * find_next_best_node - find the next node to include in a sched_domain
6668 * @node: node whose sched_domain we're building
6669 * @used_nodes: nodes already in the sched_domain
6671 * Find the next node to include in a given scheduling domain. Simply
6672 * finds the closest node not already in the @used_nodes map.
6674 * Should use nodemask_t.
6676 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
6678 int i
, n
, val
, min_val
, best_node
= 0;
6682 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6683 /* Start at @node */
6684 n
= (node
+ i
) % MAX_NUMNODES
;
6686 if (!nr_cpus_node(n
))
6689 /* Skip already used nodes */
6690 if (node_isset(n
, *used_nodes
))
6693 /* Simple min distance search */
6694 val
= node_distance(node
, n
);
6696 if (val
< min_val
) {
6702 node_set(best_node
, *used_nodes
);
6707 * sched_domain_node_span - get a cpumask for a node's sched_domain
6708 * @node: node whose cpumask we're constructing
6709 * @span: resulting cpumask
6711 * Given a node, construct a good cpumask for its sched_domain to span. It
6712 * should be one that prevents unnecessary balancing, but also spreads tasks
6715 static void sched_domain_node_span(int node
, cpumask_t
*span
)
6717 nodemask_t used_nodes
;
6718 node_to_cpumask_ptr(nodemask
, node
);
6722 nodes_clear(used_nodes
);
6724 cpus_or(*span
, *span
, *nodemask
);
6725 node_set(node
, used_nodes
);
6727 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6728 int next_node
= find_next_best_node(node
, &used_nodes
);
6730 node_to_cpumask_ptr_next(nodemask
, next_node
);
6731 cpus_or(*span
, *span
, *nodemask
);
6734 #endif /* CONFIG_NUMA */
6736 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6739 * SMT sched-domains:
6741 #ifdef CONFIG_SCHED_SMT
6742 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
6743 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
6746 cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
6750 *sg
= &per_cpu(sched_group_cpus
, cpu
);
6753 #endif /* CONFIG_SCHED_SMT */
6756 * multi-core sched-domains:
6758 #ifdef CONFIG_SCHED_MC
6759 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
6760 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
6761 #endif /* CONFIG_SCHED_MC */
6763 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6765 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
6770 *mask
= per_cpu(cpu_sibling_map
, cpu
);
6771 cpus_and(*mask
, *mask
, *cpu_map
);
6772 group
= first_cpu(*mask
);
6774 *sg
= &per_cpu(sched_group_core
, group
);
6777 #elif defined(CONFIG_SCHED_MC)
6779 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
6783 *sg
= &per_cpu(sched_group_core
, cpu
);
6788 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
6789 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
6792 cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
6796 #ifdef CONFIG_SCHED_MC
6797 *mask
= cpu_coregroup_map(cpu
);
6798 cpus_and(*mask
, *mask
, *cpu_map
);
6799 group
= first_cpu(*mask
);
6800 #elif defined(CONFIG_SCHED_SMT)
6801 *mask
= per_cpu(cpu_sibling_map
, cpu
);
6802 cpus_and(*mask
, *mask
, *cpu_map
);
6803 group
= first_cpu(*mask
);
6808 *sg
= &per_cpu(sched_group_phys
, group
);
6814 * The init_sched_build_groups can't handle what we want to do with node
6815 * groups, so roll our own. Now each node has its own list of groups which
6816 * gets dynamically allocated.
6818 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
6819 static struct sched_group
***sched_group_nodes_bycpu
;
6821 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
6822 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
6824 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
6825 struct sched_group
**sg
, cpumask_t
*nodemask
)
6829 *nodemask
= node_to_cpumask(cpu_to_node(cpu
));
6830 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
6831 group
= first_cpu(*nodemask
);
6834 *sg
= &per_cpu(sched_group_allnodes
, group
);
6838 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6840 struct sched_group
*sg
= group_head
;
6846 for_each_cpu_mask(j
, sg
->cpumask
) {
6847 struct sched_domain
*sd
;
6849 sd
= &per_cpu(phys_domains
, j
);
6850 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
6852 * Only add "power" once for each
6858 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
6861 } while (sg
!= group_head
);
6863 #endif /* CONFIG_NUMA */
6866 /* Free memory allocated for various sched_group structures */
6867 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
6871 for_each_cpu_mask(cpu
, *cpu_map
) {
6872 struct sched_group
**sched_group_nodes
6873 = sched_group_nodes_bycpu
[cpu
];
6875 if (!sched_group_nodes
)
6878 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6879 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6881 *nodemask
= node_to_cpumask(i
);
6882 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
6883 if (cpus_empty(*nodemask
))
6893 if (oldsg
!= sched_group_nodes
[i
])
6896 kfree(sched_group_nodes
);
6897 sched_group_nodes_bycpu
[cpu
] = NULL
;
6900 #else /* !CONFIG_NUMA */
6901 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
6904 #endif /* CONFIG_NUMA */
6907 * Initialize sched groups cpu_power.
6909 * cpu_power indicates the capacity of sched group, which is used while
6910 * distributing the load between different sched groups in a sched domain.
6911 * Typically cpu_power for all the groups in a sched domain will be same unless
6912 * there are asymmetries in the topology. If there are asymmetries, group
6913 * having more cpu_power will pickup more load compared to the group having
6916 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6917 * the maximum number of tasks a group can handle in the presence of other idle
6918 * or lightly loaded groups in the same sched domain.
6920 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6922 struct sched_domain
*child
;
6923 struct sched_group
*group
;
6925 WARN_ON(!sd
|| !sd
->groups
);
6927 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
6932 sd
->groups
->__cpu_power
= 0;
6935 * For perf policy, if the groups in child domain share resources
6936 * (for example cores sharing some portions of the cache hierarchy
6937 * or SMT), then set this domain groups cpu_power such that each group
6938 * can handle only one task, when there are other idle groups in the
6939 * same sched domain.
6941 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
6943 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
6944 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
6949 * add cpu_power of each child group to this groups cpu_power
6951 group
= child
->groups
;
6953 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
6954 group
= group
->next
;
6955 } while (group
!= child
->groups
);
6959 * Initializers for schedule domains
6960 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6963 #define SD_INIT(sd, type) sd_init_##type(sd)
6964 #define SD_INIT_FUNC(type) \
6965 static noinline void sd_init_##type(struct sched_domain *sd) \
6967 memset(sd, 0, sizeof(*sd)); \
6968 *sd = SD_##type##_INIT; \
6969 sd->level = SD_LV_##type; \
6974 SD_INIT_FUNC(ALLNODES
)
6977 #ifdef CONFIG_SCHED_SMT
6978 SD_INIT_FUNC(SIBLING
)
6980 #ifdef CONFIG_SCHED_MC
6985 * To minimize stack usage kmalloc room for cpumasks and share the
6986 * space as the usage in build_sched_domains() dictates. Used only
6987 * if the amount of space is significant.
6990 cpumask_t tmpmask
; /* make this one first */
6993 cpumask_t this_sibling_map
;
6994 cpumask_t this_core_map
;
6996 cpumask_t send_covered
;
6999 cpumask_t domainspan
;
7001 cpumask_t notcovered
;
7006 #define SCHED_CPUMASK_ALLOC 1
7007 #define SCHED_CPUMASK_FREE(v) kfree(v)
7008 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7010 #define SCHED_CPUMASK_ALLOC 0
7011 #define SCHED_CPUMASK_FREE(v)
7012 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7015 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7016 ((unsigned long)(a) + offsetof(struct allmasks, v))
7018 static int default_relax_domain_level
= -1;
7020 static int __init
setup_relax_domain_level(char *str
)
7024 val
= simple_strtoul(str
, NULL
, 0);
7025 if (val
< SD_LV_MAX
)
7026 default_relax_domain_level
= val
;
7030 __setup("relax_domain_level=", setup_relax_domain_level
);
7032 static void set_domain_attribute(struct sched_domain
*sd
,
7033 struct sched_domain_attr
*attr
)
7037 if (!attr
|| attr
->relax_domain_level
< 0) {
7038 if (default_relax_domain_level
< 0)
7041 request
= default_relax_domain_level
;
7043 request
= attr
->relax_domain_level
;
7044 if (request
< sd
->level
) {
7045 /* turn off idle balance on this domain */
7046 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
7048 /* turn on idle balance on this domain */
7049 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
7054 * Build sched domains for a given set of cpus and attach the sched domains
7055 * to the individual cpus
7057 static int __build_sched_domains(const cpumask_t
*cpu_map
,
7058 struct sched_domain_attr
*attr
)
7061 struct root_domain
*rd
;
7062 SCHED_CPUMASK_DECLARE(allmasks
);
7065 struct sched_group
**sched_group_nodes
= NULL
;
7066 int sd_allnodes
= 0;
7069 * Allocate the per-node list of sched groups
7071 sched_group_nodes
= kcalloc(MAX_NUMNODES
, sizeof(struct sched_group
*),
7073 if (!sched_group_nodes
) {
7074 printk(KERN_WARNING
"Can not alloc sched group node list\n");
7079 rd
= alloc_rootdomain();
7081 printk(KERN_WARNING
"Cannot alloc root domain\n");
7083 kfree(sched_group_nodes
);
7088 #if SCHED_CPUMASK_ALLOC
7089 /* get space for all scratch cpumask variables */
7090 allmasks
= kmalloc(sizeof(*allmasks
), GFP_KERNEL
);
7092 printk(KERN_WARNING
"Cannot alloc cpumask array\n");
7095 kfree(sched_group_nodes
);
7100 tmpmask
= (cpumask_t
*)allmasks
;
7104 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
7108 * Set up domains for cpus specified by the cpu_map.
7110 for_each_cpu_mask(i
, *cpu_map
) {
7111 struct sched_domain
*sd
= NULL
, *p
;
7112 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7114 *nodemask
= node_to_cpumask(cpu_to_node(i
));
7115 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7118 if (cpus_weight(*cpu_map
) >
7119 SD_NODES_PER_DOMAIN
*cpus_weight(*nodemask
)) {
7120 sd
= &per_cpu(allnodes_domains
, i
);
7121 SD_INIT(sd
, ALLNODES
);
7122 set_domain_attribute(sd
, attr
);
7123 sd
->span
= *cpu_map
;
7124 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7130 sd
= &per_cpu(node_domains
, i
);
7132 set_domain_attribute(sd
, attr
);
7133 sched_domain_node_span(cpu_to_node(i
), &sd
->span
);
7137 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7141 sd
= &per_cpu(phys_domains
, i
);
7143 set_domain_attribute(sd
, attr
);
7144 sd
->span
= *nodemask
;
7148 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7150 #ifdef CONFIG_SCHED_MC
7152 sd
= &per_cpu(core_domains
, i
);
7154 set_domain_attribute(sd
, attr
);
7155 sd
->span
= cpu_coregroup_map(i
);
7156 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7159 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7162 #ifdef CONFIG_SCHED_SMT
7164 sd
= &per_cpu(cpu_domains
, i
);
7165 SD_INIT(sd
, SIBLING
);
7166 set_domain_attribute(sd
, attr
);
7167 sd
->span
= per_cpu(cpu_sibling_map
, i
);
7168 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7171 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7175 #ifdef CONFIG_SCHED_SMT
7176 /* Set up CPU (sibling) groups */
7177 for_each_cpu_mask(i
, *cpu_map
) {
7178 SCHED_CPUMASK_VAR(this_sibling_map
, allmasks
);
7179 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7181 *this_sibling_map
= per_cpu(cpu_sibling_map
, i
);
7182 cpus_and(*this_sibling_map
, *this_sibling_map
, *cpu_map
);
7183 if (i
!= first_cpu(*this_sibling_map
))
7186 init_sched_build_groups(this_sibling_map
, cpu_map
,
7188 send_covered
, tmpmask
);
7192 #ifdef CONFIG_SCHED_MC
7193 /* Set up multi-core groups */
7194 for_each_cpu_mask(i
, *cpu_map
) {
7195 SCHED_CPUMASK_VAR(this_core_map
, allmasks
);
7196 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7198 *this_core_map
= cpu_coregroup_map(i
);
7199 cpus_and(*this_core_map
, *this_core_map
, *cpu_map
);
7200 if (i
!= first_cpu(*this_core_map
))
7203 init_sched_build_groups(this_core_map
, cpu_map
,
7205 send_covered
, tmpmask
);
7209 /* Set up physical groups */
7210 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
7211 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7212 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7214 *nodemask
= node_to_cpumask(i
);
7215 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7216 if (cpus_empty(*nodemask
))
7219 init_sched_build_groups(nodemask
, cpu_map
,
7221 send_covered
, tmpmask
);
7225 /* Set up node groups */
7227 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7229 init_sched_build_groups(cpu_map
, cpu_map
,
7230 &cpu_to_allnodes_group
,
7231 send_covered
, tmpmask
);
7234 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
7235 /* Set up node groups */
7236 struct sched_group
*sg
, *prev
;
7237 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7238 SCHED_CPUMASK_VAR(domainspan
, allmasks
);
7239 SCHED_CPUMASK_VAR(covered
, allmasks
);
7242 *nodemask
= node_to_cpumask(i
);
7243 cpus_clear(*covered
);
7245 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7246 if (cpus_empty(*nodemask
)) {
7247 sched_group_nodes
[i
] = NULL
;
7251 sched_domain_node_span(i
, domainspan
);
7252 cpus_and(*domainspan
, *domainspan
, *cpu_map
);
7254 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
7256 printk(KERN_WARNING
"Can not alloc domain group for "
7260 sched_group_nodes
[i
] = sg
;
7261 for_each_cpu_mask(j
, *nodemask
) {
7262 struct sched_domain
*sd
;
7264 sd
= &per_cpu(node_domains
, j
);
7267 sg
->__cpu_power
= 0;
7268 sg
->cpumask
= *nodemask
;
7270 cpus_or(*covered
, *covered
, *nodemask
);
7273 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
7274 SCHED_CPUMASK_VAR(notcovered
, allmasks
);
7275 int n
= (i
+ j
) % MAX_NUMNODES
;
7276 node_to_cpumask_ptr(pnodemask
, n
);
7278 cpus_complement(*notcovered
, *covered
);
7279 cpus_and(*tmpmask
, *notcovered
, *cpu_map
);
7280 cpus_and(*tmpmask
, *tmpmask
, *domainspan
);
7281 if (cpus_empty(*tmpmask
))
7284 cpus_and(*tmpmask
, *tmpmask
, *pnodemask
);
7285 if (cpus_empty(*tmpmask
))
7288 sg
= kmalloc_node(sizeof(struct sched_group
),
7292 "Can not alloc domain group for node %d\n", j
);
7295 sg
->__cpu_power
= 0;
7296 sg
->cpumask
= *tmpmask
;
7297 sg
->next
= prev
->next
;
7298 cpus_or(*covered
, *covered
, *tmpmask
);
7305 /* Calculate CPU power for physical packages and nodes */
7306 #ifdef CONFIG_SCHED_SMT
7307 for_each_cpu_mask(i
, *cpu_map
) {
7308 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
7310 init_sched_groups_power(i
, sd
);
7313 #ifdef CONFIG_SCHED_MC
7314 for_each_cpu_mask(i
, *cpu_map
) {
7315 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
7317 init_sched_groups_power(i
, sd
);
7321 for_each_cpu_mask(i
, *cpu_map
) {
7322 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
7324 init_sched_groups_power(i
, sd
);
7328 for (i
= 0; i
< MAX_NUMNODES
; i
++)
7329 init_numa_sched_groups_power(sched_group_nodes
[i
]);
7332 struct sched_group
*sg
;
7334 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
,
7336 init_numa_sched_groups_power(sg
);
7340 /* Attach the domains */
7341 for_each_cpu_mask(i
, *cpu_map
) {
7342 struct sched_domain
*sd
;
7343 #ifdef CONFIG_SCHED_SMT
7344 sd
= &per_cpu(cpu_domains
, i
);
7345 #elif defined(CONFIG_SCHED_MC)
7346 sd
= &per_cpu(core_domains
, i
);
7348 sd
= &per_cpu(phys_domains
, i
);
7350 cpu_attach_domain(sd
, rd
, i
);
7353 SCHED_CPUMASK_FREE((void *)allmasks
);
7358 free_sched_groups(cpu_map
, tmpmask
);
7359 SCHED_CPUMASK_FREE((void *)allmasks
);
7364 static int build_sched_domains(const cpumask_t
*cpu_map
)
7366 return __build_sched_domains(cpu_map
, NULL
);
7369 static cpumask_t
*doms_cur
; /* current sched domains */
7370 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7371 static struct sched_domain_attr
*dattr_cur
;
7372 /* attribues of custom domains in 'doms_cur' */
7375 * Special case: If a kmalloc of a doms_cur partition (array of
7376 * cpumask_t) fails, then fallback to a single sched domain,
7377 * as determined by the single cpumask_t fallback_doms.
7379 static cpumask_t fallback_doms
;
7381 void __attribute__((weak
)) arch_update_cpu_topology(void)
7386 * Free current domain masks.
7387 * Called after all cpus are attached to NULL domain.
7389 static void free_sched_domains(void)
7392 if (doms_cur
!= &fallback_doms
)
7394 doms_cur
= &fallback_doms
;
7398 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7399 * For now this just excludes isolated cpus, but could be used to
7400 * exclude other special cases in the future.
7402 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
7406 arch_update_cpu_topology();
7408 doms_cur
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
7410 doms_cur
= &fallback_doms
;
7411 cpus_andnot(*doms_cur
, *cpu_map
, cpu_isolated_map
);
7413 err
= build_sched_domains(doms_cur
);
7414 register_sched_domain_sysctl();
7419 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
,
7422 free_sched_groups(cpu_map
, tmpmask
);
7426 * Detach sched domains from a group of cpus specified in cpu_map
7427 * These cpus will now be attached to the NULL domain
7429 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
7434 unregister_sched_domain_sysctl();
7436 for_each_cpu_mask(i
, *cpu_map
)
7437 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7438 synchronize_sched();
7439 arch_destroy_sched_domains(cpu_map
, &tmpmask
);
7442 /* handle null as "default" */
7443 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7444 struct sched_domain_attr
*new, int idx_new
)
7446 struct sched_domain_attr tmp
;
7453 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7454 new ? (new + idx_new
) : &tmp
,
7455 sizeof(struct sched_domain_attr
));
7459 * Partition sched domains as specified by the 'ndoms_new'
7460 * cpumasks in the array doms_new[] of cpumasks. This compares
7461 * doms_new[] to the current sched domain partitioning, doms_cur[].
7462 * It destroys each deleted domain and builds each new domain.
7464 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7465 * The masks don't intersect (don't overlap.) We should setup one
7466 * sched domain for each mask. CPUs not in any of the cpumasks will
7467 * not be load balanced. If the same cpumask appears both in the
7468 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7471 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7472 * ownership of it and will kfree it when done with it. If the caller
7473 * failed the kmalloc call, then it can pass in doms_new == NULL,
7474 * and partition_sched_domains() will fallback to the single partition
7477 * Call with hotplug lock held
7479 void partition_sched_domains(int ndoms_new
, cpumask_t
*doms_new
,
7480 struct sched_domain_attr
*dattr_new
)
7484 mutex_lock(&sched_domains_mutex
);
7486 /* always unregister in case we don't destroy any domains */
7487 unregister_sched_domain_sysctl();
7489 if (doms_new
== NULL
) {
7491 doms_new
= &fallback_doms
;
7492 cpus_andnot(doms_new
[0], cpu_online_map
, cpu_isolated_map
);
7496 /* Destroy deleted domains */
7497 for (i
= 0; i
< ndoms_cur
; i
++) {
7498 for (j
= 0; j
< ndoms_new
; j
++) {
7499 if (cpus_equal(doms_cur
[i
], doms_new
[j
])
7500 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7503 /* no match - a current sched domain not in new doms_new[] */
7504 detach_destroy_domains(doms_cur
+ i
);
7509 /* Build new domains */
7510 for (i
= 0; i
< ndoms_new
; i
++) {
7511 for (j
= 0; j
< ndoms_cur
; j
++) {
7512 if (cpus_equal(doms_new
[i
], doms_cur
[j
])
7513 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7516 /* no match - add a new doms_new */
7517 __build_sched_domains(doms_new
+ i
,
7518 dattr_new
? dattr_new
+ i
: NULL
);
7523 /* Remember the new sched domains */
7524 if (doms_cur
!= &fallback_doms
)
7526 kfree(dattr_cur
); /* kfree(NULL) is safe */
7527 doms_cur
= doms_new
;
7528 dattr_cur
= dattr_new
;
7529 ndoms_cur
= ndoms_new
;
7531 register_sched_domain_sysctl();
7533 mutex_unlock(&sched_domains_mutex
);
7536 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7537 int arch_reinit_sched_domains(void)
7542 mutex_lock(&sched_domains_mutex
);
7543 detach_destroy_domains(&cpu_online_map
);
7544 free_sched_domains();
7545 err
= arch_init_sched_domains(&cpu_online_map
);
7546 mutex_unlock(&sched_domains_mutex
);
7552 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7556 if (buf
[0] != '0' && buf
[0] != '1')
7560 sched_smt_power_savings
= (buf
[0] == '1');
7562 sched_mc_power_savings
= (buf
[0] == '1');
7564 ret
= arch_reinit_sched_domains();
7566 return ret
? ret
: count
;
7569 #ifdef CONFIG_SCHED_MC
7570 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
7572 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7574 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
7575 const char *buf
, size_t count
)
7577 return sched_power_savings_store(buf
, count
, 0);
7579 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
7580 sched_mc_power_savings_store
);
7583 #ifdef CONFIG_SCHED_SMT
7584 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
7586 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7588 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
7589 const char *buf
, size_t count
)
7591 return sched_power_savings_store(buf
, count
, 1);
7593 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
7594 sched_smt_power_savings_store
);
7597 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7601 #ifdef CONFIG_SCHED_SMT
7603 err
= sysfs_create_file(&cls
->kset
.kobj
,
7604 &attr_sched_smt_power_savings
.attr
);
7606 #ifdef CONFIG_SCHED_MC
7607 if (!err
&& mc_capable())
7608 err
= sysfs_create_file(&cls
->kset
.kobj
,
7609 &attr_sched_mc_power_savings
.attr
);
7613 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7616 * Force a reinitialization of the sched domains hierarchy. The domains
7617 * and groups cannot be updated in place without racing with the balancing
7618 * code, so we temporarily attach all running cpus to the NULL domain
7619 * which will prevent rebalancing while the sched domains are recalculated.
7621 static int update_sched_domains(struct notifier_block
*nfb
,
7622 unsigned long action
, void *hcpu
)
7624 int cpu
= (int)(long)hcpu
;
7627 case CPU_DOWN_PREPARE
:
7628 case CPU_DOWN_PREPARE_FROZEN
:
7629 disable_runtime(cpu_rq(cpu
));
7631 case CPU_UP_PREPARE
:
7632 case CPU_UP_PREPARE_FROZEN
:
7633 detach_destroy_domains(&cpu_online_map
);
7634 free_sched_domains();
7638 case CPU_DOWN_FAILED
:
7639 case CPU_DOWN_FAILED_FROZEN
:
7641 case CPU_ONLINE_FROZEN
:
7642 enable_runtime(cpu_rq(cpu
));
7644 case CPU_UP_CANCELED
:
7645 case CPU_UP_CANCELED_FROZEN
:
7647 case CPU_DEAD_FROZEN
:
7649 * Fall through and re-initialise the domains.
7656 #ifndef CONFIG_CPUSETS
7658 * Create default domain partitioning if cpusets are disabled.
7659 * Otherwise we let cpusets rebuild the domains based on the
7663 /* The hotplug lock is already held by cpu_up/cpu_down */
7664 arch_init_sched_domains(&cpu_online_map
);
7670 void __init
sched_init_smp(void)
7672 cpumask_t non_isolated_cpus
;
7674 #if defined(CONFIG_NUMA)
7675 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
7677 BUG_ON(sched_group_nodes_bycpu
== NULL
);
7680 mutex_lock(&sched_domains_mutex
);
7681 arch_init_sched_domains(&cpu_online_map
);
7682 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
7683 if (cpus_empty(non_isolated_cpus
))
7684 cpu_set(smp_processor_id(), non_isolated_cpus
);
7685 mutex_unlock(&sched_domains_mutex
);
7687 /* XXX: Theoretical race here - CPU may be hotplugged now */
7688 hotcpu_notifier(update_sched_domains
, 0);
7691 /* Move init over to a non-isolated CPU */
7692 if (set_cpus_allowed_ptr(current
, &non_isolated_cpus
) < 0)
7694 sched_init_granularity();
7697 void __init
sched_init_smp(void)
7699 sched_init_granularity();
7701 #endif /* CONFIG_SMP */
7703 int in_sched_functions(unsigned long addr
)
7705 return in_lock_functions(addr
) ||
7706 (addr
>= (unsigned long)__sched_text_start
7707 && addr
< (unsigned long)__sched_text_end
);
7710 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7712 cfs_rq
->tasks_timeline
= RB_ROOT
;
7713 INIT_LIST_HEAD(&cfs_rq
->tasks
);
7714 #ifdef CONFIG_FAIR_GROUP_SCHED
7717 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7720 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7722 struct rt_prio_array
*array
;
7725 array
= &rt_rq
->active
;
7726 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
7727 INIT_LIST_HEAD(array
->queue
+ i
);
7728 __clear_bit(i
, array
->bitmap
);
7730 /* delimiter for bitsearch: */
7731 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
7733 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7734 rt_rq
->highest_prio
= MAX_RT_PRIO
;
7737 rt_rq
->rt_nr_migratory
= 0;
7738 rt_rq
->overloaded
= 0;
7742 rt_rq
->rt_throttled
= 0;
7743 rt_rq
->rt_runtime
= 0;
7744 spin_lock_init(&rt_rq
->rt_runtime_lock
);
7746 #ifdef CONFIG_RT_GROUP_SCHED
7747 rt_rq
->rt_nr_boosted
= 0;
7752 #ifdef CONFIG_FAIR_GROUP_SCHED
7753 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7754 struct sched_entity
*se
, int cpu
, int add
,
7755 struct sched_entity
*parent
)
7757 struct rq
*rq
= cpu_rq(cpu
);
7758 tg
->cfs_rq
[cpu
] = cfs_rq
;
7759 init_cfs_rq(cfs_rq
, rq
);
7762 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7765 /* se could be NULL for init_task_group */
7770 se
->cfs_rq
= &rq
->cfs
;
7772 se
->cfs_rq
= parent
->my_q
;
7775 se
->load
.weight
= tg
->shares
;
7776 se
->load
.inv_weight
= 0;
7777 se
->parent
= parent
;
7781 #ifdef CONFIG_RT_GROUP_SCHED
7782 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
7783 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
7784 struct sched_rt_entity
*parent
)
7786 struct rq
*rq
= cpu_rq(cpu
);
7788 tg
->rt_rq
[cpu
] = rt_rq
;
7789 init_rt_rq(rt_rq
, rq
);
7791 rt_rq
->rt_se
= rt_se
;
7792 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7794 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
7796 tg
->rt_se
[cpu
] = rt_se
;
7801 rt_se
->rt_rq
= &rq
->rt
;
7803 rt_se
->rt_rq
= parent
->my_q
;
7805 rt_se
->my_q
= rt_rq
;
7806 rt_se
->parent
= parent
;
7807 INIT_LIST_HEAD(&rt_se
->run_list
);
7811 void __init
sched_init(void)
7814 unsigned long alloc_size
= 0, ptr
;
7816 #ifdef CONFIG_FAIR_GROUP_SCHED
7817 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7819 #ifdef CONFIG_RT_GROUP_SCHED
7820 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7822 #ifdef CONFIG_USER_SCHED
7826 * As sched_init() is called before page_alloc is setup,
7827 * we use alloc_bootmem().
7830 ptr
= (unsigned long)alloc_bootmem(alloc_size
);
7832 #ifdef CONFIG_FAIR_GROUP_SCHED
7833 init_task_group
.se
= (struct sched_entity
**)ptr
;
7834 ptr
+= nr_cpu_ids
* sizeof(void **);
7836 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7837 ptr
+= nr_cpu_ids
* sizeof(void **);
7839 #ifdef CONFIG_USER_SCHED
7840 root_task_group
.se
= (struct sched_entity
**)ptr
;
7841 ptr
+= nr_cpu_ids
* sizeof(void **);
7843 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7844 ptr
+= nr_cpu_ids
* sizeof(void **);
7845 #endif /* CONFIG_USER_SCHED */
7846 #endif /* CONFIG_FAIR_GROUP_SCHED */
7847 #ifdef CONFIG_RT_GROUP_SCHED
7848 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7849 ptr
+= nr_cpu_ids
* sizeof(void **);
7851 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7852 ptr
+= nr_cpu_ids
* sizeof(void **);
7854 #ifdef CONFIG_USER_SCHED
7855 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7856 ptr
+= nr_cpu_ids
* sizeof(void **);
7858 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7859 ptr
+= nr_cpu_ids
* sizeof(void **);
7860 #endif /* CONFIG_USER_SCHED */
7861 #endif /* CONFIG_RT_GROUP_SCHED */
7865 init_defrootdomain();
7868 init_rt_bandwidth(&def_rt_bandwidth
,
7869 global_rt_period(), global_rt_runtime());
7871 #ifdef CONFIG_RT_GROUP_SCHED
7872 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
7873 global_rt_period(), global_rt_runtime());
7874 #ifdef CONFIG_USER_SCHED
7875 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
7876 global_rt_period(), RUNTIME_INF
);
7877 #endif /* CONFIG_USER_SCHED */
7878 #endif /* CONFIG_RT_GROUP_SCHED */
7880 #ifdef CONFIG_GROUP_SCHED
7881 list_add(&init_task_group
.list
, &task_groups
);
7882 INIT_LIST_HEAD(&init_task_group
.children
);
7884 #ifdef CONFIG_USER_SCHED
7885 INIT_LIST_HEAD(&root_task_group
.children
);
7886 init_task_group
.parent
= &root_task_group
;
7887 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
7888 #endif /* CONFIG_USER_SCHED */
7889 #endif /* CONFIG_GROUP_SCHED */
7891 for_each_possible_cpu(i
) {
7895 spin_lock_init(&rq
->lock
);
7896 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
7898 init_cfs_rq(&rq
->cfs
, rq
);
7899 init_rt_rq(&rq
->rt
, rq
);
7900 #ifdef CONFIG_FAIR_GROUP_SCHED
7901 init_task_group
.shares
= init_task_group_load
;
7902 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7903 #ifdef CONFIG_CGROUP_SCHED
7905 * How much cpu bandwidth does init_task_group get?
7907 * In case of task-groups formed thr' the cgroup filesystem, it
7908 * gets 100% of the cpu resources in the system. This overall
7909 * system cpu resource is divided among the tasks of
7910 * init_task_group and its child task-groups in a fair manner,
7911 * based on each entity's (task or task-group's) weight
7912 * (se->load.weight).
7914 * In other words, if init_task_group has 10 tasks of weight
7915 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7916 * then A0's share of the cpu resource is:
7918 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7920 * We achieve this by letting init_task_group's tasks sit
7921 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7923 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
7924 #elif defined CONFIG_USER_SCHED
7925 root_task_group
.shares
= NICE_0_LOAD
;
7926 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
7928 * In case of task-groups formed thr' the user id of tasks,
7929 * init_task_group represents tasks belonging to root user.
7930 * Hence it forms a sibling of all subsequent groups formed.
7931 * In this case, init_task_group gets only a fraction of overall
7932 * system cpu resource, based on the weight assigned to root
7933 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
7934 * by letting tasks of init_task_group sit in a separate cfs_rq
7935 * (init_cfs_rq) and having one entity represent this group of
7936 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
7938 init_tg_cfs_entry(&init_task_group
,
7939 &per_cpu(init_cfs_rq
, i
),
7940 &per_cpu(init_sched_entity
, i
), i
, 1,
7941 root_task_group
.se
[i
]);
7944 #endif /* CONFIG_FAIR_GROUP_SCHED */
7946 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7947 #ifdef CONFIG_RT_GROUP_SCHED
7948 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
7949 #ifdef CONFIG_CGROUP_SCHED
7950 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
7951 #elif defined CONFIG_USER_SCHED
7952 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
7953 init_tg_rt_entry(&init_task_group
,
7954 &per_cpu(init_rt_rq
, i
),
7955 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
7956 root_task_group
.rt_se
[i
]);
7960 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7961 rq
->cpu_load
[j
] = 0;
7965 rq
->active_balance
= 0;
7966 rq
->next_balance
= jiffies
;
7970 rq
->migration_thread
= NULL
;
7971 INIT_LIST_HEAD(&rq
->migration_queue
);
7972 rq_attach_root(rq
, &def_root_domain
);
7975 atomic_set(&rq
->nr_iowait
, 0);
7978 set_load_weight(&init_task
);
7980 #ifdef CONFIG_PREEMPT_NOTIFIERS
7981 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7985 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
7988 #ifdef CONFIG_RT_MUTEXES
7989 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
7993 * The boot idle thread does lazy MMU switching as well:
7995 atomic_inc(&init_mm
.mm_count
);
7996 enter_lazy_tlb(&init_mm
, current
);
7999 * Make us the idle thread. Technically, schedule() should not be
8000 * called from this thread, however somewhere below it might be,
8001 * but because we are the idle thread, we just pick up running again
8002 * when this runqueue becomes "idle".
8004 init_idle(current
, smp_processor_id());
8006 * During early bootup we pretend to be a normal task:
8008 current
->sched_class
= &fair_sched_class
;
8010 scheduler_running
= 1;
8013 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8014 void __might_sleep(char *file
, int line
)
8017 static unsigned long prev_jiffy
; /* ratelimiting */
8019 if ((in_atomic() || irqs_disabled()) &&
8020 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
8021 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
8023 prev_jiffy
= jiffies
;
8024 printk(KERN_ERR
"BUG: sleeping function called from invalid"
8025 " context at %s:%d\n", file
, line
);
8026 printk("in_atomic():%d, irqs_disabled():%d\n",
8027 in_atomic(), irqs_disabled());
8028 debug_show_held_locks(current
);
8029 if (irqs_disabled())
8030 print_irqtrace_events(current
);
8035 EXPORT_SYMBOL(__might_sleep
);
8038 #ifdef CONFIG_MAGIC_SYSRQ
8039 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
8043 update_rq_clock(rq
);
8044 on_rq
= p
->se
.on_rq
;
8046 deactivate_task(rq
, p
, 0);
8047 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
8049 activate_task(rq
, p
, 0);
8050 resched_task(rq
->curr
);
8054 void normalize_rt_tasks(void)
8056 struct task_struct
*g
, *p
;
8057 unsigned long flags
;
8060 read_lock_irqsave(&tasklist_lock
, flags
);
8061 do_each_thread(g
, p
) {
8063 * Only normalize user tasks:
8068 p
->se
.exec_start
= 0;
8069 #ifdef CONFIG_SCHEDSTATS
8070 p
->se
.wait_start
= 0;
8071 p
->se
.sleep_start
= 0;
8072 p
->se
.block_start
= 0;
8077 * Renice negative nice level userspace
8080 if (TASK_NICE(p
) < 0 && p
->mm
)
8081 set_user_nice(p
, 0);
8085 spin_lock(&p
->pi_lock
);
8086 rq
= __task_rq_lock(p
);
8088 normalize_task(rq
, p
);
8090 __task_rq_unlock(rq
);
8091 spin_unlock(&p
->pi_lock
);
8092 } while_each_thread(g
, p
);
8094 read_unlock_irqrestore(&tasklist_lock
, flags
);
8097 #endif /* CONFIG_MAGIC_SYSRQ */
8101 * These functions are only useful for the IA64 MCA handling.
8103 * They can only be called when the whole system has been
8104 * stopped - every CPU needs to be quiescent, and no scheduling
8105 * activity can take place. Using them for anything else would
8106 * be a serious bug, and as a result, they aren't even visible
8107 * under any other configuration.
8111 * curr_task - return the current task for a given cpu.
8112 * @cpu: the processor in question.
8114 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8116 struct task_struct
*curr_task(int cpu
)
8118 return cpu_curr(cpu
);
8122 * set_curr_task - set the current task for a given cpu.
8123 * @cpu: the processor in question.
8124 * @p: the task pointer to set.
8126 * Description: This function must only be used when non-maskable interrupts
8127 * are serviced on a separate stack. It allows the architecture to switch the
8128 * notion of the current task on a cpu in a non-blocking manner. This function
8129 * must be called with all CPU's synchronized, and interrupts disabled, the
8130 * and caller must save the original value of the current task (see
8131 * curr_task() above) and restore that value before reenabling interrupts and
8132 * re-starting the system.
8134 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8136 void set_curr_task(int cpu
, struct task_struct
*p
)
8143 #ifdef CONFIG_FAIR_GROUP_SCHED
8144 static void free_fair_sched_group(struct task_group
*tg
)
8148 for_each_possible_cpu(i
) {
8150 kfree(tg
->cfs_rq
[i
]);
8160 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8162 struct cfs_rq
*cfs_rq
;
8163 struct sched_entity
*se
, *parent_se
;
8167 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8170 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8174 tg
->shares
= NICE_0_LOAD
;
8176 for_each_possible_cpu(i
) {
8179 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
),
8180 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8184 se
= kmalloc_node(sizeof(struct sched_entity
),
8185 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8189 parent_se
= parent
? parent
->se
[i
] : NULL
;
8190 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent_se
);
8199 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8201 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
8202 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
8205 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8207 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
8209 #else /* !CONFG_FAIR_GROUP_SCHED */
8210 static inline void free_fair_sched_group(struct task_group
*tg
)
8215 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8220 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8224 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8227 #endif /* CONFIG_FAIR_GROUP_SCHED */
8229 #ifdef CONFIG_RT_GROUP_SCHED
8230 static void free_rt_sched_group(struct task_group
*tg
)
8234 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8236 for_each_possible_cpu(i
) {
8238 kfree(tg
->rt_rq
[i
]);
8240 kfree(tg
->rt_se
[i
]);
8248 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8250 struct rt_rq
*rt_rq
;
8251 struct sched_rt_entity
*rt_se
, *parent_se
;
8255 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8258 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8262 init_rt_bandwidth(&tg
->rt_bandwidth
,
8263 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8265 for_each_possible_cpu(i
) {
8268 rt_rq
= kmalloc_node(sizeof(struct rt_rq
),
8269 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8273 rt_se
= kmalloc_node(sizeof(struct sched_rt_entity
),
8274 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8278 parent_se
= parent
? parent
->rt_se
[i
] : NULL
;
8279 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent_se
);
8288 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8290 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8291 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8294 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8296 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8298 #else /* !CONFIG_RT_GROUP_SCHED */
8299 static inline void free_rt_sched_group(struct task_group
*tg
)
8304 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8309 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8313 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8316 #endif /* CONFIG_RT_GROUP_SCHED */
8318 #ifdef CONFIG_GROUP_SCHED
8319 static void free_sched_group(struct task_group
*tg
)
8321 free_fair_sched_group(tg
);
8322 free_rt_sched_group(tg
);
8326 /* allocate runqueue etc for a new task group */
8327 struct task_group
*sched_create_group(struct task_group
*parent
)
8329 struct task_group
*tg
;
8330 unsigned long flags
;
8333 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8335 return ERR_PTR(-ENOMEM
);
8337 if (!alloc_fair_sched_group(tg
, parent
))
8340 if (!alloc_rt_sched_group(tg
, parent
))
8343 spin_lock_irqsave(&task_group_lock
, flags
);
8344 for_each_possible_cpu(i
) {
8345 register_fair_sched_group(tg
, i
);
8346 register_rt_sched_group(tg
, i
);
8348 list_add_rcu(&tg
->list
, &task_groups
);
8350 WARN_ON(!parent
); /* root should already exist */
8352 tg
->parent
= parent
;
8353 list_add_rcu(&tg
->siblings
, &parent
->children
);
8354 INIT_LIST_HEAD(&tg
->children
);
8355 spin_unlock_irqrestore(&task_group_lock
, flags
);
8360 free_sched_group(tg
);
8361 return ERR_PTR(-ENOMEM
);
8364 /* rcu callback to free various structures associated with a task group */
8365 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8367 /* now it should be safe to free those cfs_rqs */
8368 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8371 /* Destroy runqueue etc associated with a task group */
8372 void sched_destroy_group(struct task_group
*tg
)
8374 unsigned long flags
;
8377 spin_lock_irqsave(&task_group_lock
, flags
);
8378 for_each_possible_cpu(i
) {
8379 unregister_fair_sched_group(tg
, i
);
8380 unregister_rt_sched_group(tg
, i
);
8382 list_del_rcu(&tg
->list
);
8383 list_del_rcu(&tg
->siblings
);
8384 spin_unlock_irqrestore(&task_group_lock
, flags
);
8386 /* wait for possible concurrent references to cfs_rqs complete */
8387 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8390 /* change task's runqueue when it moves between groups.
8391 * The caller of this function should have put the task in its new group
8392 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8393 * reflect its new group.
8395 void sched_move_task(struct task_struct
*tsk
)
8398 unsigned long flags
;
8401 rq
= task_rq_lock(tsk
, &flags
);
8403 update_rq_clock(rq
);
8405 running
= task_current(rq
, tsk
);
8406 on_rq
= tsk
->se
.on_rq
;
8409 dequeue_task(rq
, tsk
, 0);
8410 if (unlikely(running
))
8411 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8413 set_task_rq(tsk
, task_cpu(tsk
));
8415 #ifdef CONFIG_FAIR_GROUP_SCHED
8416 if (tsk
->sched_class
->moved_group
)
8417 tsk
->sched_class
->moved_group(tsk
);
8420 if (unlikely(running
))
8421 tsk
->sched_class
->set_curr_task(rq
);
8423 enqueue_task(rq
, tsk
, 0);
8425 task_rq_unlock(rq
, &flags
);
8427 #endif /* CONFIG_GROUP_SCHED */
8429 #ifdef CONFIG_FAIR_GROUP_SCHED
8430 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8432 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8437 dequeue_entity(cfs_rq
, se
, 0);
8439 se
->load
.weight
= shares
;
8440 se
->load
.inv_weight
= 0;
8443 enqueue_entity(cfs_rq
, se
, 0);
8446 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8448 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8449 struct rq
*rq
= cfs_rq
->rq
;
8450 unsigned long flags
;
8452 spin_lock_irqsave(&rq
->lock
, flags
);
8453 __set_se_shares(se
, shares
);
8454 spin_unlock_irqrestore(&rq
->lock
, flags
);
8457 static DEFINE_MUTEX(shares_mutex
);
8459 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8462 unsigned long flags
;
8465 * We can't change the weight of the root cgroup.
8470 if (shares
< MIN_SHARES
)
8471 shares
= MIN_SHARES
;
8472 else if (shares
> MAX_SHARES
)
8473 shares
= MAX_SHARES
;
8475 mutex_lock(&shares_mutex
);
8476 if (tg
->shares
== shares
)
8479 spin_lock_irqsave(&task_group_lock
, flags
);
8480 for_each_possible_cpu(i
)
8481 unregister_fair_sched_group(tg
, i
);
8482 list_del_rcu(&tg
->siblings
);
8483 spin_unlock_irqrestore(&task_group_lock
, flags
);
8485 /* wait for any ongoing reference to this group to finish */
8486 synchronize_sched();
8489 * Now we are free to modify the group's share on each cpu
8490 * w/o tripping rebalance_share or load_balance_fair.
8492 tg
->shares
= shares
;
8493 for_each_possible_cpu(i
) {
8497 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
8498 set_se_shares(tg
->se
[i
], shares
);
8502 * Enable load balance activity on this group, by inserting it back on
8503 * each cpu's rq->leaf_cfs_rq_list.
8505 spin_lock_irqsave(&task_group_lock
, flags
);
8506 for_each_possible_cpu(i
)
8507 register_fair_sched_group(tg
, i
);
8508 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
8509 spin_unlock_irqrestore(&task_group_lock
, flags
);
8511 mutex_unlock(&shares_mutex
);
8515 unsigned long sched_group_shares(struct task_group
*tg
)
8521 #ifdef CONFIG_RT_GROUP_SCHED
8523 * Ensure that the real time constraints are schedulable.
8525 static DEFINE_MUTEX(rt_constraints_mutex
);
8527 static unsigned long to_ratio(u64 period
, u64 runtime
)
8529 if (runtime
== RUNTIME_INF
)
8532 return div64_u64(runtime
<< 16, period
);
8535 #ifdef CONFIG_CGROUP_SCHED
8536 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8538 struct task_group
*tgi
, *parent
= tg
->parent
;
8539 unsigned long total
= 0;
8542 if (global_rt_period() < period
)
8545 return to_ratio(period
, runtime
) <
8546 to_ratio(global_rt_period(), global_rt_runtime());
8549 if (ktime_to_ns(parent
->rt_bandwidth
.rt_period
) < period
)
8553 list_for_each_entry_rcu(tgi
, &parent
->children
, siblings
) {
8557 total
+= to_ratio(ktime_to_ns(tgi
->rt_bandwidth
.rt_period
),
8558 tgi
->rt_bandwidth
.rt_runtime
);
8562 return total
+ to_ratio(period
, runtime
) <=
8563 to_ratio(ktime_to_ns(parent
->rt_bandwidth
.rt_period
),
8564 parent
->rt_bandwidth
.rt_runtime
);
8566 #elif defined CONFIG_USER_SCHED
8567 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8569 struct task_group
*tgi
;
8570 unsigned long total
= 0;
8571 unsigned long global_ratio
=
8572 to_ratio(global_rt_period(), global_rt_runtime());
8575 list_for_each_entry_rcu(tgi
, &task_groups
, list
) {
8579 total
+= to_ratio(ktime_to_ns(tgi
->rt_bandwidth
.rt_period
),
8580 tgi
->rt_bandwidth
.rt_runtime
);
8584 return total
+ to_ratio(period
, runtime
) < global_ratio
;
8588 /* Must be called with tasklist_lock held */
8589 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8591 struct task_struct
*g
, *p
;
8592 do_each_thread(g
, p
) {
8593 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8595 } while_each_thread(g
, p
);
8599 static int tg_set_bandwidth(struct task_group
*tg
,
8600 u64 rt_period
, u64 rt_runtime
)
8604 mutex_lock(&rt_constraints_mutex
);
8605 read_lock(&tasklist_lock
);
8606 if (rt_runtime
== 0 && tg_has_rt_tasks(tg
)) {
8610 if (!__rt_schedulable(tg
, rt_period
, rt_runtime
)) {
8615 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8616 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8617 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8619 for_each_possible_cpu(i
) {
8620 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8622 spin_lock(&rt_rq
->rt_runtime_lock
);
8623 rt_rq
->rt_runtime
= rt_runtime
;
8624 spin_unlock(&rt_rq
->rt_runtime_lock
);
8626 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8628 read_unlock(&tasklist_lock
);
8629 mutex_unlock(&rt_constraints_mutex
);
8634 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8636 u64 rt_runtime
, rt_period
;
8638 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8639 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8640 if (rt_runtime_us
< 0)
8641 rt_runtime
= RUNTIME_INF
;
8643 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8646 long sched_group_rt_runtime(struct task_group
*tg
)
8650 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8653 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8654 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8655 return rt_runtime_us
;
8658 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8660 u64 rt_runtime
, rt_period
;
8662 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8663 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8665 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8668 long sched_group_rt_period(struct task_group
*tg
)
8672 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8673 do_div(rt_period_us
, NSEC_PER_USEC
);
8674 return rt_period_us
;
8677 static int sched_rt_global_constraints(void)
8679 struct task_group
*tg
= &root_task_group
;
8680 u64 rt_runtime
, rt_period
;
8683 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8684 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8686 mutex_lock(&rt_constraints_mutex
);
8687 if (!__rt_schedulable(tg
, rt_period
, rt_runtime
))
8689 mutex_unlock(&rt_constraints_mutex
);
8693 #else /* !CONFIG_RT_GROUP_SCHED */
8694 static int sched_rt_global_constraints(void)
8696 unsigned long flags
;
8699 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8700 for_each_possible_cpu(i
) {
8701 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
8703 spin_lock(&rt_rq
->rt_runtime_lock
);
8704 rt_rq
->rt_runtime
= global_rt_runtime();
8705 spin_unlock(&rt_rq
->rt_runtime_lock
);
8707 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8711 #endif /* CONFIG_RT_GROUP_SCHED */
8713 int sched_rt_handler(struct ctl_table
*table
, int write
,
8714 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
8718 int old_period
, old_runtime
;
8719 static DEFINE_MUTEX(mutex
);
8722 old_period
= sysctl_sched_rt_period
;
8723 old_runtime
= sysctl_sched_rt_runtime
;
8725 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
8727 if (!ret
&& write
) {
8728 ret
= sched_rt_global_constraints();
8730 sysctl_sched_rt_period
= old_period
;
8731 sysctl_sched_rt_runtime
= old_runtime
;
8733 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8734 def_rt_bandwidth
.rt_period
=
8735 ns_to_ktime(global_rt_period());
8738 mutex_unlock(&mutex
);
8743 #ifdef CONFIG_CGROUP_SCHED
8745 /* return corresponding task_group object of a cgroup */
8746 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
8748 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
8749 struct task_group
, css
);
8752 static struct cgroup_subsys_state
*
8753 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8755 struct task_group
*tg
, *parent
;
8757 if (!cgrp
->parent
) {
8758 /* This is early initialization for the top cgroup */
8759 init_task_group
.css
.cgroup
= cgrp
;
8760 return &init_task_group
.css
;
8763 parent
= cgroup_tg(cgrp
->parent
);
8764 tg
= sched_create_group(parent
);
8766 return ERR_PTR(-ENOMEM
);
8768 /* Bind the cgroup to task_group object we just created */
8769 tg
->css
.cgroup
= cgrp
;
8775 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8777 struct task_group
*tg
= cgroup_tg(cgrp
);
8779 sched_destroy_group(tg
);
8783 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8784 struct task_struct
*tsk
)
8786 #ifdef CONFIG_RT_GROUP_SCHED
8787 /* Don't accept realtime tasks when there is no way for them to run */
8788 if (rt_task(tsk
) && cgroup_tg(cgrp
)->rt_bandwidth
.rt_runtime
== 0)
8791 /* We don't support RT-tasks being in separate groups */
8792 if (tsk
->sched_class
!= &fair_sched_class
)
8800 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8801 struct cgroup
*old_cont
, struct task_struct
*tsk
)
8803 sched_move_task(tsk
);
8806 #ifdef CONFIG_FAIR_GROUP_SCHED
8807 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
8810 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
8813 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
8815 struct task_group
*tg
= cgroup_tg(cgrp
);
8817 return (u64
) tg
->shares
;
8819 #endif /* CONFIG_FAIR_GROUP_SCHED */
8821 #ifdef CONFIG_RT_GROUP_SCHED
8822 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
8825 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
8828 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8830 return sched_group_rt_runtime(cgroup_tg(cgrp
));
8833 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
8836 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
8839 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
8841 return sched_group_rt_period(cgroup_tg(cgrp
));
8843 #endif /* CONFIG_RT_GROUP_SCHED */
8845 static struct cftype cpu_files
[] = {
8846 #ifdef CONFIG_FAIR_GROUP_SCHED
8849 .read_u64
= cpu_shares_read_u64
,
8850 .write_u64
= cpu_shares_write_u64
,
8853 #ifdef CONFIG_RT_GROUP_SCHED
8855 .name
= "rt_runtime_us",
8856 .read_s64
= cpu_rt_runtime_read
,
8857 .write_s64
= cpu_rt_runtime_write
,
8860 .name
= "rt_period_us",
8861 .read_u64
= cpu_rt_period_read_uint
,
8862 .write_u64
= cpu_rt_period_write_uint
,
8867 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8869 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
8872 struct cgroup_subsys cpu_cgroup_subsys
= {
8874 .create
= cpu_cgroup_create
,
8875 .destroy
= cpu_cgroup_destroy
,
8876 .can_attach
= cpu_cgroup_can_attach
,
8877 .attach
= cpu_cgroup_attach
,
8878 .populate
= cpu_cgroup_populate
,
8879 .subsys_id
= cpu_cgroup_subsys_id
,
8883 #endif /* CONFIG_CGROUP_SCHED */
8885 #ifdef CONFIG_CGROUP_CPUACCT
8888 * CPU accounting code for task groups.
8890 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8891 * (balbir@in.ibm.com).
8894 /* track cpu usage of a group of tasks */
8896 struct cgroup_subsys_state css
;
8897 /* cpuusage holds pointer to a u64-type object on every cpu */
8901 struct cgroup_subsys cpuacct_subsys
;
8903 /* return cpu accounting group corresponding to this container */
8904 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
8906 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
8907 struct cpuacct
, css
);
8910 /* return cpu accounting group to which this task belongs */
8911 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
8913 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
8914 struct cpuacct
, css
);
8917 /* create a new cpu accounting group */
8918 static struct cgroup_subsys_state
*cpuacct_create(
8919 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8921 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
8924 return ERR_PTR(-ENOMEM
);
8926 ca
->cpuusage
= alloc_percpu(u64
);
8927 if (!ca
->cpuusage
) {
8929 return ERR_PTR(-ENOMEM
);
8935 /* destroy an existing cpu accounting group */
8937 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8939 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8941 free_percpu(ca
->cpuusage
);
8945 /* return total cpu usage (in nanoseconds) of a group */
8946 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8948 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8949 u64 totalcpuusage
= 0;
8952 for_each_possible_cpu(i
) {
8953 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
8956 * Take rq->lock to make 64-bit addition safe on 32-bit
8959 spin_lock_irq(&cpu_rq(i
)->lock
);
8960 totalcpuusage
+= *cpuusage
;
8961 spin_unlock_irq(&cpu_rq(i
)->lock
);
8964 return totalcpuusage
;
8967 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
8970 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8979 for_each_possible_cpu(i
) {
8980 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
8982 spin_lock_irq(&cpu_rq(i
)->lock
);
8984 spin_unlock_irq(&cpu_rq(i
)->lock
);
8990 static struct cftype files
[] = {
8993 .read_u64
= cpuusage_read
,
8994 .write_u64
= cpuusage_write
,
8998 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9000 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
9004 * charge this task's execution time to its accounting group.
9006 * called with rq->lock held.
9008 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
9012 if (!cpuacct_subsys
.active
)
9017 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, task_cpu(tsk
));
9019 *cpuusage
+= cputime
;
9023 struct cgroup_subsys cpuacct_subsys
= {
9025 .create
= cpuacct_create
,
9026 .destroy
= cpuacct_destroy
,
9027 .populate
= cpuacct_populate
,
9028 .subsys_id
= cpuacct_subsys_id
,
9030 #endif /* CONFIG_CGROUP_CPUACCT */