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_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1584 spin_unlock(&rq
->lock
);
1586 spin_lock(&rq
->lock
);
1589 static void update_h_load(int cpu
)
1591 walk_tg_tree(tg_load_down
, tg_nop
, cpu
, NULL
);
1594 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1596 cfs_rq
->shares
= shares
;
1601 static inline void update_shares(struct sched_domain
*sd
)
1605 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1613 #include "sched_stats.h"
1614 #include "sched_idletask.c"
1615 #include "sched_fair.c"
1616 #include "sched_rt.c"
1617 #ifdef CONFIG_SCHED_DEBUG
1618 # include "sched_debug.c"
1621 #define sched_class_highest (&rt_sched_class)
1622 #define for_each_class(class) \
1623 for (class = sched_class_highest; class; class = class->next)
1625 static void inc_nr_running(struct rq
*rq
)
1630 static void dec_nr_running(struct rq
*rq
)
1635 static void set_load_weight(struct task_struct
*p
)
1637 if (task_has_rt_policy(p
)) {
1638 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1639 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1644 * SCHED_IDLE tasks get minimal weight:
1646 if (p
->policy
== SCHED_IDLE
) {
1647 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1648 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1652 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1653 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1656 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1658 sched_info_queued(p
);
1659 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1663 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1665 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1670 * __normal_prio - return the priority that is based on the static prio
1672 static inline int __normal_prio(struct task_struct
*p
)
1674 return p
->static_prio
;
1678 * Calculate the expected normal priority: i.e. priority
1679 * without taking RT-inheritance into account. Might be
1680 * boosted by interactivity modifiers. Changes upon fork,
1681 * setprio syscalls, and whenever the interactivity
1682 * estimator recalculates.
1684 static inline int normal_prio(struct task_struct
*p
)
1688 if (task_has_rt_policy(p
))
1689 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1691 prio
= __normal_prio(p
);
1696 * Calculate the current priority, i.e. the priority
1697 * taken into account by the scheduler. This value might
1698 * be boosted by RT tasks, or might be boosted by
1699 * interactivity modifiers. Will be RT if the task got
1700 * RT-boosted. If not then it returns p->normal_prio.
1702 static int effective_prio(struct task_struct
*p
)
1704 p
->normal_prio
= normal_prio(p
);
1706 * If we are RT tasks or we were boosted to RT priority,
1707 * keep the priority unchanged. Otherwise, update priority
1708 * to the normal priority:
1710 if (!rt_prio(p
->prio
))
1711 return p
->normal_prio
;
1716 * activate_task - move a task to the runqueue.
1718 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1720 if (task_contributes_to_load(p
))
1721 rq
->nr_uninterruptible
--;
1723 enqueue_task(rq
, p
, wakeup
);
1728 * deactivate_task - remove a task from the runqueue.
1730 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1732 if (task_contributes_to_load(p
))
1733 rq
->nr_uninterruptible
++;
1735 dequeue_task(rq
, p
, sleep
);
1740 * task_curr - is this task currently executing on a CPU?
1741 * @p: the task in question.
1743 inline int task_curr(const struct task_struct
*p
)
1745 return cpu_curr(task_cpu(p
)) == p
;
1748 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1750 set_task_rq(p
, cpu
);
1753 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1754 * successfuly executed on another CPU. We must ensure that updates of
1755 * per-task data have been completed by this moment.
1758 task_thread_info(p
)->cpu
= cpu
;
1762 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1763 const struct sched_class
*prev_class
,
1764 int oldprio
, int running
)
1766 if (prev_class
!= p
->sched_class
) {
1767 if (prev_class
->switched_from
)
1768 prev_class
->switched_from(rq
, p
, running
);
1769 p
->sched_class
->switched_to(rq
, p
, running
);
1771 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1776 /* Used instead of source_load when we know the type == 0 */
1777 static unsigned long weighted_cpuload(const int cpu
)
1779 return cpu_rq(cpu
)->load
.weight
;
1783 * Is this task likely cache-hot:
1786 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1791 * Buddy candidates are cache hot:
1793 if (sched_feat(CACHE_HOT_BUDDY
) && (&p
->se
== cfs_rq_of(&p
->se
)->next
))
1796 if (p
->sched_class
!= &fair_sched_class
)
1799 if (sysctl_sched_migration_cost
== -1)
1801 if (sysctl_sched_migration_cost
== 0)
1804 delta
= now
- p
->se
.exec_start
;
1806 return delta
< (s64
)sysctl_sched_migration_cost
;
1810 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1812 int old_cpu
= task_cpu(p
);
1813 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1814 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1815 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1818 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1820 #ifdef CONFIG_SCHEDSTATS
1821 if (p
->se
.wait_start
)
1822 p
->se
.wait_start
-= clock_offset
;
1823 if (p
->se
.sleep_start
)
1824 p
->se
.sleep_start
-= clock_offset
;
1825 if (p
->se
.block_start
)
1826 p
->se
.block_start
-= clock_offset
;
1827 if (old_cpu
!= new_cpu
) {
1828 schedstat_inc(p
, se
.nr_migrations
);
1829 if (task_hot(p
, old_rq
->clock
, NULL
))
1830 schedstat_inc(p
, se
.nr_forced2_migrations
);
1833 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1834 new_cfsrq
->min_vruntime
;
1836 __set_task_cpu(p
, new_cpu
);
1839 struct migration_req
{
1840 struct list_head list
;
1842 struct task_struct
*task
;
1845 struct completion done
;
1849 * The task's runqueue lock must be held.
1850 * Returns true if you have to wait for migration thread.
1853 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1855 struct rq
*rq
= task_rq(p
);
1858 * If the task is not on a runqueue (and not running), then
1859 * it is sufficient to simply update the task's cpu field.
1861 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1862 set_task_cpu(p
, dest_cpu
);
1866 init_completion(&req
->done
);
1868 req
->dest_cpu
= dest_cpu
;
1869 list_add(&req
->list
, &rq
->migration_queue
);
1875 * wait_task_inactive - wait for a thread to unschedule.
1877 * The caller must ensure that the task *will* unschedule sometime soon,
1878 * else this function might spin for a *long* time. This function can't
1879 * be called with interrupts off, or it may introduce deadlock with
1880 * smp_call_function() if an IPI is sent by the same process we are
1881 * waiting to become inactive.
1883 void wait_task_inactive(struct task_struct
*p
)
1885 unsigned long flags
;
1891 * We do the initial early heuristics without holding
1892 * any task-queue locks at all. We'll only try to get
1893 * the runqueue lock when things look like they will
1899 * If the task is actively running on another CPU
1900 * still, just relax and busy-wait without holding
1903 * NOTE! Since we don't hold any locks, it's not
1904 * even sure that "rq" stays as the right runqueue!
1905 * But we don't care, since "task_running()" will
1906 * return false if the runqueue has changed and p
1907 * is actually now running somewhere else!
1909 while (task_running(rq
, p
))
1913 * Ok, time to look more closely! We need the rq
1914 * lock now, to be *sure*. If we're wrong, we'll
1915 * just go back and repeat.
1917 rq
= task_rq_lock(p
, &flags
);
1918 running
= task_running(rq
, p
);
1919 on_rq
= p
->se
.on_rq
;
1920 task_rq_unlock(rq
, &flags
);
1923 * Was it really running after all now that we
1924 * checked with the proper locks actually held?
1926 * Oops. Go back and try again..
1928 if (unlikely(running
)) {
1934 * It's not enough that it's not actively running,
1935 * it must be off the runqueue _entirely_, and not
1938 * So if it wa still runnable (but just not actively
1939 * running right now), it's preempted, and we should
1940 * yield - it could be a while.
1942 if (unlikely(on_rq
)) {
1943 schedule_timeout_uninterruptible(1);
1948 * Ahh, all good. It wasn't running, and it wasn't
1949 * runnable, which means that it will never become
1950 * running in the future either. We're all done!
1957 * kick_process - kick a running thread to enter/exit the kernel
1958 * @p: the to-be-kicked thread
1960 * Cause a process which is running on another CPU to enter
1961 * kernel-mode, without any delay. (to get signals handled.)
1963 * NOTE: this function doesnt have to take the runqueue lock,
1964 * because all it wants to ensure is that the remote task enters
1965 * the kernel. If the IPI races and the task has been migrated
1966 * to another CPU then no harm is done and the purpose has been
1969 void kick_process(struct task_struct
*p
)
1975 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1976 smp_send_reschedule(cpu
);
1981 * Return a low guess at the load of a migration-source cpu weighted
1982 * according to the scheduling class and "nice" value.
1984 * We want to under-estimate the load of migration sources, to
1985 * balance conservatively.
1987 static unsigned long source_load(int cpu
, int type
)
1989 struct rq
*rq
= cpu_rq(cpu
);
1990 unsigned long total
= weighted_cpuload(cpu
);
1995 return min(rq
->cpu_load
[type
-1], total
);
1999 * Return a high guess at the load of a migration-target cpu weighted
2000 * according to the scheduling class and "nice" value.
2002 static unsigned long target_load(int cpu
, int type
)
2004 struct rq
*rq
= cpu_rq(cpu
);
2005 unsigned long total
= weighted_cpuload(cpu
);
2010 return max(rq
->cpu_load
[type
-1], total
);
2014 * Return the average load per task on the cpu's run queue
2016 static unsigned long cpu_avg_load_per_task(int cpu
)
2018 struct rq
*rq
= cpu_rq(cpu
);
2019 unsigned long total
= weighted_cpuload(cpu
);
2020 unsigned long n
= rq
->nr_running
;
2022 return n
? total
/ n
: SCHED_LOAD_SCALE
;
2026 * find_idlest_group finds and returns the least busy CPU group within the
2029 static struct sched_group
*
2030 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
2032 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2033 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
2034 int load_idx
= sd
->forkexec_idx
;
2035 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
2038 unsigned long load
, avg_load
;
2042 /* Skip over this group if it has no CPUs allowed */
2043 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
2046 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2048 /* Tally up the load of all CPUs in the group */
2051 for_each_cpu_mask(i
, group
->cpumask
) {
2052 /* Bias balancing toward cpus of our domain */
2054 load
= source_load(i
, load_idx
);
2056 load
= target_load(i
, load_idx
);
2061 /* Adjust by relative CPU power of the group */
2062 avg_load
= sg_div_cpu_power(group
,
2063 avg_load
* SCHED_LOAD_SCALE
);
2066 this_load
= avg_load
;
2068 } else if (avg_load
< min_load
) {
2069 min_load
= avg_load
;
2072 } while (group
= group
->next
, group
!= sd
->groups
);
2074 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
2080 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2083 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
,
2086 unsigned long load
, min_load
= ULONG_MAX
;
2090 /* Traverse only the allowed CPUs */
2091 cpus_and(*tmp
, group
->cpumask
, p
->cpus_allowed
);
2093 for_each_cpu_mask(i
, *tmp
) {
2094 load
= weighted_cpuload(i
);
2096 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2106 * sched_balance_self: balance the current task (running on cpu) in domains
2107 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2110 * Balance, ie. select the least loaded group.
2112 * Returns the target CPU number, or the same CPU if no balancing is needed.
2114 * preempt must be disabled.
2116 static int sched_balance_self(int cpu
, int flag
)
2118 struct task_struct
*t
= current
;
2119 struct sched_domain
*tmp
, *sd
= NULL
;
2121 for_each_domain(cpu
, tmp
) {
2123 * If power savings logic is enabled for a domain, stop there.
2125 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2127 if (tmp
->flags
& flag
)
2132 cpumask_t span
, tmpmask
;
2133 struct sched_group
*group
;
2134 int new_cpu
, weight
;
2136 if (!(sd
->flags
& flag
)) {
2142 group
= find_idlest_group(sd
, t
, cpu
);
2148 new_cpu
= find_idlest_cpu(group
, t
, cpu
, &tmpmask
);
2149 if (new_cpu
== -1 || new_cpu
== cpu
) {
2150 /* Now try balancing at a lower domain level of cpu */
2155 /* Now try balancing at a lower domain level of new_cpu */
2158 weight
= cpus_weight(span
);
2159 for_each_domain(cpu
, tmp
) {
2160 if (weight
<= cpus_weight(tmp
->span
))
2162 if (tmp
->flags
& flag
)
2165 /* while loop will break here if sd == NULL */
2171 #endif /* CONFIG_SMP */
2174 * try_to_wake_up - wake up a thread
2175 * @p: the to-be-woken-up thread
2176 * @state: the mask of task states that can be woken
2177 * @sync: do a synchronous wakeup?
2179 * Put it on the run-queue if it's not already there. The "current"
2180 * thread is always on the run-queue (except when the actual
2181 * re-schedule is in progress), and as such you're allowed to do
2182 * the simpler "current->state = TASK_RUNNING" to mark yourself
2183 * runnable without the overhead of this.
2185 * returns failure only if the task is already active.
2187 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
2189 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2190 unsigned long flags
;
2194 if (!sched_feat(SYNC_WAKEUPS
))
2198 rq
= task_rq_lock(p
, &flags
);
2199 old_state
= p
->state
;
2200 if (!(old_state
& state
))
2208 this_cpu
= smp_processor_id();
2211 if (unlikely(task_running(rq
, p
)))
2214 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2215 if (cpu
!= orig_cpu
) {
2216 set_task_cpu(p
, cpu
);
2217 task_rq_unlock(rq
, &flags
);
2218 /* might preempt at this point */
2219 rq
= task_rq_lock(p
, &flags
);
2220 old_state
= p
->state
;
2221 if (!(old_state
& state
))
2226 this_cpu
= smp_processor_id();
2230 #ifdef CONFIG_SCHEDSTATS
2231 schedstat_inc(rq
, ttwu_count
);
2232 if (cpu
== this_cpu
)
2233 schedstat_inc(rq
, ttwu_local
);
2235 struct sched_domain
*sd
;
2236 for_each_domain(this_cpu
, sd
) {
2237 if (cpu_isset(cpu
, sd
->span
)) {
2238 schedstat_inc(sd
, ttwu_wake_remote
);
2243 #endif /* CONFIG_SCHEDSTATS */
2246 #endif /* CONFIG_SMP */
2247 schedstat_inc(p
, se
.nr_wakeups
);
2249 schedstat_inc(p
, se
.nr_wakeups_sync
);
2250 if (orig_cpu
!= cpu
)
2251 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2252 if (cpu
== this_cpu
)
2253 schedstat_inc(p
, se
.nr_wakeups_local
);
2255 schedstat_inc(p
, se
.nr_wakeups_remote
);
2256 update_rq_clock(rq
);
2257 activate_task(rq
, p
, 1);
2261 check_preempt_curr(rq
, p
);
2263 p
->state
= TASK_RUNNING
;
2265 if (p
->sched_class
->task_wake_up
)
2266 p
->sched_class
->task_wake_up(rq
, p
);
2269 task_rq_unlock(rq
, &flags
);
2274 int wake_up_process(struct task_struct
*p
)
2276 return try_to_wake_up(p
, TASK_ALL
, 0);
2278 EXPORT_SYMBOL(wake_up_process
);
2280 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2282 return try_to_wake_up(p
, state
, 0);
2286 * Perform scheduler related setup for a newly forked process p.
2287 * p is forked by current.
2289 * __sched_fork() is basic setup used by init_idle() too:
2291 static void __sched_fork(struct task_struct
*p
)
2293 p
->se
.exec_start
= 0;
2294 p
->se
.sum_exec_runtime
= 0;
2295 p
->se
.prev_sum_exec_runtime
= 0;
2296 p
->se
.last_wakeup
= 0;
2297 p
->se
.avg_overlap
= 0;
2299 #ifdef CONFIG_SCHEDSTATS
2300 p
->se
.wait_start
= 0;
2301 p
->se
.sum_sleep_runtime
= 0;
2302 p
->se
.sleep_start
= 0;
2303 p
->se
.block_start
= 0;
2304 p
->se
.sleep_max
= 0;
2305 p
->se
.block_max
= 0;
2307 p
->se
.slice_max
= 0;
2311 INIT_LIST_HEAD(&p
->rt
.run_list
);
2313 INIT_LIST_HEAD(&p
->se
.group_node
);
2315 #ifdef CONFIG_PREEMPT_NOTIFIERS
2316 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2320 * We mark the process as running here, but have not actually
2321 * inserted it onto the runqueue yet. This guarantees that
2322 * nobody will actually run it, and a signal or other external
2323 * event cannot wake it up and insert it on the runqueue either.
2325 p
->state
= TASK_RUNNING
;
2329 * fork()/clone()-time setup:
2331 void sched_fork(struct task_struct
*p
, int clone_flags
)
2333 int cpu
= get_cpu();
2338 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2340 set_task_cpu(p
, cpu
);
2343 * Make sure we do not leak PI boosting priority to the child:
2345 p
->prio
= current
->normal_prio
;
2346 if (!rt_prio(p
->prio
))
2347 p
->sched_class
= &fair_sched_class
;
2349 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2350 if (likely(sched_info_on()))
2351 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2353 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2356 #ifdef CONFIG_PREEMPT
2357 /* Want to start with kernel preemption disabled. */
2358 task_thread_info(p
)->preempt_count
= 1;
2364 * wake_up_new_task - wake up a newly created task for the first time.
2366 * This function will do some initial scheduler statistics housekeeping
2367 * that must be done for every newly created context, then puts the task
2368 * on the runqueue and wakes it.
2370 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2372 unsigned long flags
;
2375 rq
= task_rq_lock(p
, &flags
);
2376 BUG_ON(p
->state
!= TASK_RUNNING
);
2377 update_rq_clock(rq
);
2379 p
->prio
= effective_prio(p
);
2381 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2382 activate_task(rq
, p
, 0);
2385 * Let the scheduling class do new task startup
2386 * management (if any):
2388 p
->sched_class
->task_new(rq
, p
);
2391 check_preempt_curr(rq
, p
);
2393 if (p
->sched_class
->task_wake_up
)
2394 p
->sched_class
->task_wake_up(rq
, p
);
2396 task_rq_unlock(rq
, &flags
);
2399 #ifdef CONFIG_PREEMPT_NOTIFIERS
2402 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2403 * @notifier: notifier struct to register
2405 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2407 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2409 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2412 * preempt_notifier_unregister - no longer interested in preemption notifications
2413 * @notifier: notifier struct to unregister
2415 * This is safe to call from within a preemption notifier.
2417 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2419 hlist_del(¬ifier
->link
);
2421 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2423 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
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_in(notifier
, raw_smp_processor_id());
2433 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2434 struct task_struct
*next
)
2436 struct preempt_notifier
*notifier
;
2437 struct hlist_node
*node
;
2439 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2440 notifier
->ops
->sched_out(notifier
, next
);
2443 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2445 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2450 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2451 struct task_struct
*next
)
2455 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2458 * prepare_task_switch - prepare to switch tasks
2459 * @rq: the runqueue preparing to switch
2460 * @prev: the current task that is being switched out
2461 * @next: the task we are going to switch to.
2463 * This is called with the rq lock held and interrupts off. It must
2464 * be paired with a subsequent finish_task_switch after the context
2467 * prepare_task_switch sets up locking and calls architecture specific
2471 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2472 struct task_struct
*next
)
2474 fire_sched_out_preempt_notifiers(prev
, next
);
2475 prepare_lock_switch(rq
, next
);
2476 prepare_arch_switch(next
);
2480 * finish_task_switch - clean up after a task-switch
2481 * @rq: runqueue associated with task-switch
2482 * @prev: the thread we just switched away from.
2484 * finish_task_switch must be called after the context switch, paired
2485 * with a prepare_task_switch call before the context switch.
2486 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2487 * and do any other architecture-specific cleanup actions.
2489 * Note that we may have delayed dropping an mm in context_switch(). If
2490 * so, we finish that here outside of the runqueue lock. (Doing it
2491 * with the lock held can cause deadlocks; see schedule() for
2494 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2495 __releases(rq
->lock
)
2497 struct mm_struct
*mm
= rq
->prev_mm
;
2503 * A task struct has one reference for the use as "current".
2504 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2505 * schedule one last time. The schedule call will never return, and
2506 * the scheduled task must drop that reference.
2507 * The test for TASK_DEAD must occur while the runqueue locks are
2508 * still held, otherwise prev could be scheduled on another cpu, die
2509 * there before we look at prev->state, and then the reference would
2511 * Manfred Spraul <manfred@colorfullife.com>
2513 prev_state
= prev
->state
;
2514 finish_arch_switch(prev
);
2515 finish_lock_switch(rq
, prev
);
2517 if (current
->sched_class
->post_schedule
)
2518 current
->sched_class
->post_schedule(rq
);
2521 fire_sched_in_preempt_notifiers(current
);
2524 if (unlikely(prev_state
== TASK_DEAD
)) {
2526 * Remove function-return probe instances associated with this
2527 * task and put them back on the free list.
2529 kprobe_flush_task(prev
);
2530 put_task_struct(prev
);
2535 * schedule_tail - first thing a freshly forked thread must call.
2536 * @prev: the thread we just switched away from.
2538 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2539 __releases(rq
->lock
)
2541 struct rq
*rq
= this_rq();
2543 finish_task_switch(rq
, prev
);
2544 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2545 /* In this case, finish_task_switch does not reenable preemption */
2548 if (current
->set_child_tid
)
2549 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2553 * context_switch - switch to the new MM and the new
2554 * thread's register state.
2557 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2558 struct task_struct
*next
)
2560 struct mm_struct
*mm
, *oldmm
;
2562 prepare_task_switch(rq
, prev
, next
);
2564 oldmm
= prev
->active_mm
;
2566 * For paravirt, this is coupled with an exit in switch_to to
2567 * combine the page table reload and the switch backend into
2570 arch_enter_lazy_cpu_mode();
2572 if (unlikely(!mm
)) {
2573 next
->active_mm
= oldmm
;
2574 atomic_inc(&oldmm
->mm_count
);
2575 enter_lazy_tlb(oldmm
, next
);
2577 switch_mm(oldmm
, mm
, next
);
2579 if (unlikely(!prev
->mm
)) {
2580 prev
->active_mm
= NULL
;
2581 rq
->prev_mm
= oldmm
;
2584 * Since the runqueue lock will be released by the next
2585 * task (which is an invalid locking op but in the case
2586 * of the scheduler it's an obvious special-case), so we
2587 * do an early lockdep release here:
2589 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2590 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2593 /* Here we just switch the register state and the stack. */
2594 switch_to(prev
, next
, prev
);
2598 * this_rq must be evaluated again because prev may have moved
2599 * CPUs since it called schedule(), thus the 'rq' on its stack
2600 * frame will be invalid.
2602 finish_task_switch(this_rq(), prev
);
2606 * nr_running, nr_uninterruptible and nr_context_switches:
2608 * externally visible scheduler statistics: current number of runnable
2609 * threads, current number of uninterruptible-sleeping threads, total
2610 * number of context switches performed since bootup.
2612 unsigned long nr_running(void)
2614 unsigned long i
, sum
= 0;
2616 for_each_online_cpu(i
)
2617 sum
+= cpu_rq(i
)->nr_running
;
2622 unsigned long nr_uninterruptible(void)
2624 unsigned long i
, sum
= 0;
2626 for_each_possible_cpu(i
)
2627 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2630 * Since we read the counters lockless, it might be slightly
2631 * inaccurate. Do not allow it to go below zero though:
2633 if (unlikely((long)sum
< 0))
2639 unsigned long long nr_context_switches(void)
2642 unsigned long long sum
= 0;
2644 for_each_possible_cpu(i
)
2645 sum
+= cpu_rq(i
)->nr_switches
;
2650 unsigned long nr_iowait(void)
2652 unsigned long i
, sum
= 0;
2654 for_each_possible_cpu(i
)
2655 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2660 unsigned long nr_active(void)
2662 unsigned long i
, running
= 0, uninterruptible
= 0;
2664 for_each_online_cpu(i
) {
2665 running
+= cpu_rq(i
)->nr_running
;
2666 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2669 if (unlikely((long)uninterruptible
< 0))
2670 uninterruptible
= 0;
2672 return running
+ uninterruptible
;
2676 * Update rq->cpu_load[] statistics. This function is usually called every
2677 * scheduler tick (TICK_NSEC).
2679 static void update_cpu_load(struct rq
*this_rq
)
2681 unsigned long this_load
= this_rq
->load
.weight
;
2684 this_rq
->nr_load_updates
++;
2686 /* Update our load: */
2687 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2688 unsigned long old_load
, new_load
;
2690 /* scale is effectively 1 << i now, and >> i divides by scale */
2692 old_load
= this_rq
->cpu_load
[i
];
2693 new_load
= this_load
;
2695 * Round up the averaging division if load is increasing. This
2696 * prevents us from getting stuck on 9 if the load is 10, for
2699 if (new_load
> old_load
)
2700 new_load
+= scale
-1;
2701 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2708 * double_rq_lock - safely lock two runqueues
2710 * Note this does not disable interrupts like task_rq_lock,
2711 * you need to do so manually before calling.
2713 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2714 __acquires(rq1
->lock
)
2715 __acquires(rq2
->lock
)
2717 BUG_ON(!irqs_disabled());
2719 spin_lock(&rq1
->lock
);
2720 __acquire(rq2
->lock
); /* Fake it out ;) */
2723 spin_lock(&rq1
->lock
);
2724 spin_lock(&rq2
->lock
);
2726 spin_lock(&rq2
->lock
);
2727 spin_lock(&rq1
->lock
);
2730 update_rq_clock(rq1
);
2731 update_rq_clock(rq2
);
2735 * double_rq_unlock - safely unlock two runqueues
2737 * Note this does not restore interrupts like task_rq_unlock,
2738 * you need to do so manually after calling.
2740 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2741 __releases(rq1
->lock
)
2742 __releases(rq2
->lock
)
2744 spin_unlock(&rq1
->lock
);
2746 spin_unlock(&rq2
->lock
);
2748 __release(rq2
->lock
);
2752 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2754 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2755 __releases(this_rq
->lock
)
2756 __acquires(busiest
->lock
)
2757 __acquires(this_rq
->lock
)
2761 if (unlikely(!irqs_disabled())) {
2762 /* printk() doesn't work good under rq->lock */
2763 spin_unlock(&this_rq
->lock
);
2766 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2767 if (busiest
< this_rq
) {
2768 spin_unlock(&this_rq
->lock
);
2769 spin_lock(&busiest
->lock
);
2770 spin_lock(&this_rq
->lock
);
2773 spin_lock(&busiest
->lock
);
2779 * If dest_cpu is allowed for this process, migrate the task to it.
2780 * This is accomplished by forcing the cpu_allowed mask to only
2781 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2782 * the cpu_allowed mask is restored.
2784 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2786 struct migration_req req
;
2787 unsigned long flags
;
2790 rq
= task_rq_lock(p
, &flags
);
2791 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2792 || unlikely(cpu_is_offline(dest_cpu
)))
2795 /* force the process onto the specified CPU */
2796 if (migrate_task(p
, dest_cpu
, &req
)) {
2797 /* Need to wait for migration thread (might exit: take ref). */
2798 struct task_struct
*mt
= rq
->migration_thread
;
2800 get_task_struct(mt
);
2801 task_rq_unlock(rq
, &flags
);
2802 wake_up_process(mt
);
2803 put_task_struct(mt
);
2804 wait_for_completion(&req
.done
);
2809 task_rq_unlock(rq
, &flags
);
2813 * sched_exec - execve() is a valuable balancing opportunity, because at
2814 * this point the task has the smallest effective memory and cache footprint.
2816 void sched_exec(void)
2818 int new_cpu
, this_cpu
= get_cpu();
2819 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2821 if (new_cpu
!= this_cpu
)
2822 sched_migrate_task(current
, new_cpu
);
2826 * pull_task - move a task from a remote runqueue to the local runqueue.
2827 * Both runqueues must be locked.
2829 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2830 struct rq
*this_rq
, int this_cpu
)
2832 deactivate_task(src_rq
, p
, 0);
2833 set_task_cpu(p
, this_cpu
);
2834 activate_task(this_rq
, p
, 0);
2836 * Note that idle threads have a prio of MAX_PRIO, for this test
2837 * to be always true for them.
2839 check_preempt_curr(this_rq
, p
);
2843 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2846 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2847 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2851 * We do not migrate tasks that are:
2852 * 1) running (obviously), or
2853 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2854 * 3) are cache-hot on their current CPU.
2856 if (!cpu_isset(this_cpu
, p
->cpus_allowed
)) {
2857 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2862 if (task_running(rq
, p
)) {
2863 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2868 * Aggressive migration if:
2869 * 1) task is cache cold, or
2870 * 2) too many balance attempts have failed.
2873 if (!task_hot(p
, rq
->clock
, sd
) ||
2874 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2875 #ifdef CONFIG_SCHEDSTATS
2876 if (task_hot(p
, rq
->clock
, sd
)) {
2877 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2878 schedstat_inc(p
, se
.nr_forced_migrations
);
2884 if (task_hot(p
, rq
->clock
, sd
)) {
2885 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
2891 static unsigned long
2892 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2893 unsigned long max_load_move
, struct sched_domain
*sd
,
2894 enum cpu_idle_type idle
, int *all_pinned
,
2895 int *this_best_prio
, struct rq_iterator
*iterator
)
2897 int loops
= 0, pulled
= 0, pinned
= 0, skip_for_load
;
2898 struct task_struct
*p
;
2899 long rem_load_move
= max_load_move
;
2901 if (max_load_move
== 0)
2907 * Start the load-balancing iterator:
2909 p
= iterator
->start(iterator
->arg
);
2911 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
2914 * To help distribute high priority tasks across CPUs we don't
2915 * skip a task if it will be the highest priority task (i.e. smallest
2916 * prio value) on its new queue regardless of its load weight
2918 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2919 SCHED_LOAD_SCALE_FUZZ
;
2920 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2921 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2922 p
= iterator
->next(iterator
->arg
);
2926 pull_task(busiest
, p
, this_rq
, this_cpu
);
2928 rem_load_move
-= p
->se
.load
.weight
;
2931 * We only want to steal up to the prescribed amount of weighted load.
2933 if (rem_load_move
> 0) {
2934 if (p
->prio
< *this_best_prio
)
2935 *this_best_prio
= p
->prio
;
2936 p
= iterator
->next(iterator
->arg
);
2941 * Right now, this is one of only two places pull_task() is called,
2942 * so we can safely collect pull_task() stats here rather than
2943 * inside pull_task().
2945 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2948 *all_pinned
= pinned
;
2950 return max_load_move
- rem_load_move
;
2954 * move_tasks tries to move up to max_load_move weighted load from busiest to
2955 * this_rq, as part of a balancing operation within domain "sd".
2956 * Returns 1 if successful and 0 otherwise.
2958 * Called with both runqueues locked.
2960 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2961 unsigned long max_load_move
,
2962 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2965 const struct sched_class
*class = sched_class_highest
;
2966 unsigned long total_load_moved
= 0;
2967 int this_best_prio
= this_rq
->curr
->prio
;
2971 class->load_balance(this_rq
, this_cpu
, busiest
,
2972 max_load_move
- total_load_moved
,
2973 sd
, idle
, all_pinned
, &this_best_prio
);
2974 class = class->next
;
2975 } while (class && max_load_move
> total_load_moved
);
2977 return total_load_moved
> 0;
2981 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2982 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2983 struct rq_iterator
*iterator
)
2985 struct task_struct
*p
= iterator
->start(iterator
->arg
);
2989 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2990 pull_task(busiest
, p
, this_rq
, this_cpu
);
2992 * Right now, this is only the second place pull_task()
2993 * is called, so we can safely collect pull_task()
2994 * stats here rather than inside pull_task().
2996 schedstat_inc(sd
, lb_gained
[idle
]);
3000 p
= iterator
->next(iterator
->arg
);
3007 * move_one_task tries to move exactly one task from busiest to this_rq, as
3008 * part of active balancing operations within "domain".
3009 * Returns 1 if successful and 0 otherwise.
3011 * Called with both runqueues locked.
3013 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3014 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3016 const struct sched_class
*class;
3018 for (class = sched_class_highest
; class; class = class->next
)
3019 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3026 * find_busiest_group finds and returns the busiest CPU group within the
3027 * domain. It calculates and returns the amount of weighted load which
3028 * should be moved to restore balance via the imbalance parameter.
3030 static struct sched_group
*
3031 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
3032 unsigned long *imbalance
, enum cpu_idle_type idle
,
3033 int *sd_idle
, const cpumask_t
*cpus
, int *balance
)
3035 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
3036 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
3037 unsigned long max_pull
;
3038 unsigned long busiest_load_per_task
, busiest_nr_running
;
3039 unsigned long this_load_per_task
, this_nr_running
;
3040 int load_idx
, group_imb
= 0;
3041 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3042 int power_savings_balance
= 1;
3043 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
3044 unsigned long min_nr_running
= ULONG_MAX
;
3045 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
3048 max_load
= this_load
= total_load
= total_pwr
= 0;
3049 busiest_load_per_task
= busiest_nr_running
= 0;
3050 this_load_per_task
= this_nr_running
= 0;
3051 if (idle
== CPU_NOT_IDLE
)
3052 load_idx
= sd
->busy_idx
;
3053 else if (idle
== CPU_NEWLY_IDLE
)
3054 load_idx
= sd
->newidle_idx
;
3056 load_idx
= sd
->idle_idx
;
3059 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
3062 int __group_imb
= 0;
3063 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3064 unsigned long sum_nr_running
, sum_weighted_load
;
3066 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
3069 balance_cpu
= first_cpu(group
->cpumask
);
3071 /* Tally up the load of all CPUs in the group */
3072 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
3074 min_cpu_load
= ~0UL;
3076 for_each_cpu_mask(i
, group
->cpumask
) {
3079 if (!cpu_isset(i
, *cpus
))
3084 if (*sd_idle
&& rq
->nr_running
)
3087 /* Bias balancing toward cpus of our domain */
3089 if (idle_cpu(i
) && !first_idle_cpu
) {
3094 load
= target_load(i
, load_idx
);
3096 load
= source_load(i
, load_idx
);
3097 if (load
> max_cpu_load
)
3098 max_cpu_load
= load
;
3099 if (min_cpu_load
> load
)
3100 min_cpu_load
= load
;
3104 sum_nr_running
+= rq
->nr_running
;
3105 sum_weighted_load
+= weighted_cpuload(i
);
3109 * First idle cpu or the first cpu(busiest) in this sched group
3110 * is eligible for doing load balancing at this and above
3111 * domains. In the newly idle case, we will allow all the cpu's
3112 * to do the newly idle load balance.
3114 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3115 balance_cpu
!= this_cpu
&& balance
) {
3120 total_load
+= avg_load
;
3121 total_pwr
+= group
->__cpu_power
;
3123 /* Adjust by relative CPU power of the group */
3124 avg_load
= sg_div_cpu_power(group
,
3125 avg_load
* SCHED_LOAD_SCALE
);
3127 if ((max_cpu_load
- min_cpu_load
) > SCHED_LOAD_SCALE
)
3130 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
3133 this_load
= avg_load
;
3135 this_nr_running
= sum_nr_running
;
3136 this_load_per_task
= sum_weighted_load
;
3137 } else if (avg_load
> max_load
&&
3138 (sum_nr_running
> group_capacity
|| __group_imb
)) {
3139 max_load
= avg_load
;
3141 busiest_nr_running
= sum_nr_running
;
3142 busiest_load_per_task
= sum_weighted_load
;
3143 group_imb
= __group_imb
;
3146 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3148 * Busy processors will not participate in power savings
3151 if (idle
== CPU_NOT_IDLE
||
3152 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3156 * If the local group is idle or completely loaded
3157 * no need to do power savings balance at this domain
3159 if (local_group
&& (this_nr_running
>= group_capacity
||
3161 power_savings_balance
= 0;
3164 * If a group is already running at full capacity or idle,
3165 * don't include that group in power savings calculations
3167 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
3172 * Calculate the group which has the least non-idle load.
3173 * This is the group from where we need to pick up the load
3176 if ((sum_nr_running
< min_nr_running
) ||
3177 (sum_nr_running
== min_nr_running
&&
3178 first_cpu(group
->cpumask
) <
3179 first_cpu(group_min
->cpumask
))) {
3181 min_nr_running
= sum_nr_running
;
3182 min_load_per_task
= sum_weighted_load
/
3187 * Calculate the group which is almost near its
3188 * capacity but still has some space to pick up some load
3189 * from other group and save more power
3191 if (sum_nr_running
<= group_capacity
- 1) {
3192 if (sum_nr_running
> leader_nr_running
||
3193 (sum_nr_running
== leader_nr_running
&&
3194 first_cpu(group
->cpumask
) >
3195 first_cpu(group_leader
->cpumask
))) {
3196 group_leader
= group
;
3197 leader_nr_running
= sum_nr_running
;
3202 group
= group
->next
;
3203 } while (group
!= sd
->groups
);
3205 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
3208 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
3210 if (this_load
>= avg_load
||
3211 100*max_load
<= sd
->imbalance_pct
*this_load
)
3214 busiest_load_per_task
/= busiest_nr_running
;
3216 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
3219 * We're trying to get all the cpus to the average_load, so we don't
3220 * want to push ourselves above the average load, nor do we wish to
3221 * reduce the max loaded cpu below the average load, as either of these
3222 * actions would just result in more rebalancing later, and ping-pong
3223 * tasks around. Thus we look for the minimum possible imbalance.
3224 * Negative imbalances (*we* are more loaded than anyone else) will
3225 * be counted as no imbalance for these purposes -- we can't fix that
3226 * by pulling tasks to us. Be careful of negative numbers as they'll
3227 * appear as very large values with unsigned longs.
3229 if (max_load
<= busiest_load_per_task
)
3233 * In the presence of smp nice balancing, certain scenarios can have
3234 * max load less than avg load(as we skip the groups at or below
3235 * its cpu_power, while calculating max_load..)
3237 if (max_load
< avg_load
) {
3239 goto small_imbalance
;
3242 /* Don't want to pull so many tasks that a group would go idle */
3243 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
3245 /* How much load to actually move to equalise the imbalance */
3246 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
3247 (avg_load
- this_load
) * this->__cpu_power
)
3251 * if *imbalance is less than the average load per runnable task
3252 * there is no gaurantee that any tasks will be moved so we'll have
3253 * a think about bumping its value to force at least one task to be
3256 if (*imbalance
< busiest_load_per_task
) {
3257 unsigned long tmp
, pwr_now
, pwr_move
;
3261 pwr_move
= pwr_now
= 0;
3263 if (this_nr_running
) {
3264 this_load_per_task
/= this_nr_running
;
3265 if (busiest_load_per_task
> this_load_per_task
)
3268 this_load_per_task
= SCHED_LOAD_SCALE
;
3270 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
3271 busiest_load_per_task
* imbn
) {
3272 *imbalance
= busiest_load_per_task
;
3277 * OK, we don't have enough imbalance to justify moving tasks,
3278 * however we may be able to increase total CPU power used by
3282 pwr_now
+= busiest
->__cpu_power
*
3283 min(busiest_load_per_task
, max_load
);
3284 pwr_now
+= this->__cpu_power
*
3285 min(this_load_per_task
, this_load
);
3286 pwr_now
/= SCHED_LOAD_SCALE
;
3288 /* Amount of load we'd subtract */
3289 tmp
= sg_div_cpu_power(busiest
,
3290 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3292 pwr_move
+= busiest
->__cpu_power
*
3293 min(busiest_load_per_task
, max_load
- tmp
);
3295 /* Amount of load we'd add */
3296 if (max_load
* busiest
->__cpu_power
<
3297 busiest_load_per_task
* SCHED_LOAD_SCALE
)
3298 tmp
= sg_div_cpu_power(this,
3299 max_load
* busiest
->__cpu_power
);
3301 tmp
= sg_div_cpu_power(this,
3302 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3303 pwr_move
+= this->__cpu_power
*
3304 min(this_load_per_task
, this_load
+ tmp
);
3305 pwr_move
/= SCHED_LOAD_SCALE
;
3307 /* Move if we gain throughput */
3308 if (pwr_move
> pwr_now
)
3309 *imbalance
= busiest_load_per_task
;
3315 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3316 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3319 if (this == group_leader
&& group_leader
!= group_min
) {
3320 *imbalance
= min_load_per_task
;
3330 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3333 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
3334 unsigned long imbalance
, const cpumask_t
*cpus
)
3336 struct rq
*busiest
= NULL
, *rq
;
3337 unsigned long max_load
= 0;
3340 for_each_cpu_mask(i
, group
->cpumask
) {
3343 if (!cpu_isset(i
, *cpus
))
3347 wl
= weighted_cpuload(i
);
3349 if (rq
->nr_running
== 1 && wl
> imbalance
)
3352 if (wl
> max_load
) {
3362 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3363 * so long as it is large enough.
3365 #define MAX_PINNED_INTERVAL 512
3368 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3369 * tasks if there is an imbalance.
3371 static int load_balance(int this_cpu
, struct rq
*this_rq
,
3372 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3373 int *balance
, cpumask_t
*cpus
)
3375 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
3376 struct sched_group
*group
;
3377 unsigned long imbalance
;
3379 unsigned long flags
;
3384 * When power savings policy is enabled for the parent domain, idle
3385 * sibling can pick up load irrespective of busy siblings. In this case,
3386 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3387 * portraying it as CPU_NOT_IDLE.
3389 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3390 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3393 schedstat_inc(sd
, lb_count
[idle
]);
3397 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3404 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3408 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
3410 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3414 BUG_ON(busiest
== this_rq
);
3416 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3419 if (busiest
->nr_running
> 1) {
3421 * Attempt to move tasks. If find_busiest_group has found
3422 * an imbalance but busiest->nr_running <= 1, the group is
3423 * still unbalanced. ld_moved simply stays zero, so it is
3424 * correctly treated as an imbalance.
3426 local_irq_save(flags
);
3427 double_rq_lock(this_rq
, busiest
);
3428 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3429 imbalance
, sd
, idle
, &all_pinned
);
3430 double_rq_unlock(this_rq
, busiest
);
3431 local_irq_restore(flags
);
3434 * some other cpu did the load balance for us.
3436 if (ld_moved
&& this_cpu
!= smp_processor_id())
3437 resched_cpu(this_cpu
);
3439 /* All tasks on this runqueue were pinned by CPU affinity */
3440 if (unlikely(all_pinned
)) {
3441 cpu_clear(cpu_of(busiest
), *cpus
);
3442 if (!cpus_empty(*cpus
))
3449 schedstat_inc(sd
, lb_failed
[idle
]);
3450 sd
->nr_balance_failed
++;
3452 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3454 spin_lock_irqsave(&busiest
->lock
, flags
);
3456 /* don't kick the migration_thread, if the curr
3457 * task on busiest cpu can't be moved to this_cpu
3459 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
3460 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3462 goto out_one_pinned
;
3465 if (!busiest
->active_balance
) {
3466 busiest
->active_balance
= 1;
3467 busiest
->push_cpu
= this_cpu
;
3470 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3472 wake_up_process(busiest
->migration_thread
);
3475 * We've kicked active balancing, reset the failure
3478 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3481 sd
->nr_balance_failed
= 0;
3483 if (likely(!active_balance
)) {
3484 /* We were unbalanced, so reset the balancing interval */
3485 sd
->balance_interval
= sd
->min_interval
;
3488 * If we've begun active balancing, start to back off. This
3489 * case may not be covered by the all_pinned logic if there
3490 * is only 1 task on the busy runqueue (because we don't call
3493 if (sd
->balance_interval
< sd
->max_interval
)
3494 sd
->balance_interval
*= 2;
3497 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3498 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3504 schedstat_inc(sd
, lb_balanced
[idle
]);
3506 sd
->nr_balance_failed
= 0;
3509 /* tune up the balancing interval */
3510 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3511 (sd
->balance_interval
< sd
->max_interval
))
3512 sd
->balance_interval
*= 2;
3514 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3515 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3526 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3527 * tasks if there is an imbalance.
3529 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3530 * this_rq is locked.
3533 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
,
3536 struct sched_group
*group
;
3537 struct rq
*busiest
= NULL
;
3538 unsigned long imbalance
;
3546 * When power savings policy is enabled for the parent domain, idle
3547 * sibling can pick up load irrespective of busy siblings. In this case,
3548 * let the state of idle sibling percolate up as IDLE, instead of
3549 * portraying it as CPU_NOT_IDLE.
3551 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
3552 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3555 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
3557 update_shares_locked(this_rq
, sd
);
3558 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
3559 &sd_idle
, cpus
, NULL
);
3561 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
3565 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
3567 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
3571 BUG_ON(busiest
== this_rq
);
3573 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
3576 if (busiest
->nr_running
> 1) {
3577 /* Attempt to move tasks */
3578 double_lock_balance(this_rq
, busiest
);
3579 /* this_rq->clock is already updated */
3580 update_rq_clock(busiest
);
3581 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3582 imbalance
, sd
, CPU_NEWLY_IDLE
,
3584 spin_unlock(&busiest
->lock
);
3586 if (unlikely(all_pinned
)) {
3587 cpu_clear(cpu_of(busiest
), *cpus
);
3588 if (!cpus_empty(*cpus
))
3594 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
3595 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3596 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3599 sd
->nr_balance_failed
= 0;
3601 update_shares_locked(this_rq
, sd
);
3605 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
3606 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3607 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3609 sd
->nr_balance_failed
= 0;
3615 * idle_balance is called by schedule() if this_cpu is about to become
3616 * idle. Attempts to pull tasks from other CPUs.
3618 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
3620 struct sched_domain
*sd
;
3621 int pulled_task
= -1;
3622 unsigned long next_balance
= jiffies
+ HZ
;
3625 for_each_domain(this_cpu
, sd
) {
3626 unsigned long interval
;
3628 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3631 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
3632 /* If we've pulled tasks over stop searching: */
3633 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
3636 interval
= msecs_to_jiffies(sd
->balance_interval
);
3637 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3638 next_balance
= sd
->last_balance
+ interval
;
3642 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3644 * We are going idle. next_balance may be set based on
3645 * a busy processor. So reset next_balance.
3647 this_rq
->next_balance
= next_balance
;
3652 * active_load_balance is run by migration threads. It pushes running tasks
3653 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3654 * running on each physical CPU where possible, and avoids physical /
3655 * logical imbalances.
3657 * Called with busiest_rq locked.
3659 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
3661 int target_cpu
= busiest_rq
->push_cpu
;
3662 struct sched_domain
*sd
;
3663 struct rq
*target_rq
;
3665 /* Is there any task to move? */
3666 if (busiest_rq
->nr_running
<= 1)
3669 target_rq
= cpu_rq(target_cpu
);
3672 * This condition is "impossible", if it occurs
3673 * we need to fix it. Originally reported by
3674 * Bjorn Helgaas on a 128-cpu setup.
3676 BUG_ON(busiest_rq
== target_rq
);
3678 /* move a task from busiest_rq to target_rq */
3679 double_lock_balance(busiest_rq
, target_rq
);
3680 update_rq_clock(busiest_rq
);
3681 update_rq_clock(target_rq
);
3683 /* Search for an sd spanning us and the target CPU. */
3684 for_each_domain(target_cpu
, sd
) {
3685 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3686 cpu_isset(busiest_cpu
, sd
->span
))
3691 schedstat_inc(sd
, alb_count
);
3693 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3695 schedstat_inc(sd
, alb_pushed
);
3697 schedstat_inc(sd
, alb_failed
);
3699 spin_unlock(&target_rq
->lock
);
3704 atomic_t load_balancer
;
3706 } nohz ____cacheline_aligned
= {
3707 .load_balancer
= ATOMIC_INIT(-1),
3708 .cpu_mask
= CPU_MASK_NONE
,
3712 * This routine will try to nominate the ilb (idle load balancing)
3713 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3714 * load balancing on behalf of all those cpus. If all the cpus in the system
3715 * go into this tickless mode, then there will be no ilb owner (as there is
3716 * no need for one) and all the cpus will sleep till the next wakeup event
3719 * For the ilb owner, tick is not stopped. And this tick will be used
3720 * for idle load balancing. ilb owner will still be part of
3723 * While stopping the tick, this cpu will become the ilb owner if there
3724 * is no other owner. And will be the owner till that cpu becomes busy
3725 * or if all cpus in the system stop their ticks at which point
3726 * there is no need for ilb owner.
3728 * When the ilb owner becomes busy, it nominates another owner, during the
3729 * next busy scheduler_tick()
3731 int select_nohz_load_balancer(int stop_tick
)
3733 int cpu
= smp_processor_id();
3736 cpu_set(cpu
, nohz
.cpu_mask
);
3737 cpu_rq(cpu
)->in_nohz_recently
= 1;
3740 * If we are going offline and still the leader, give up!
3742 if (cpu_is_offline(cpu
) &&
3743 atomic_read(&nohz
.load_balancer
) == cpu
) {
3744 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3749 /* time for ilb owner also to sleep */
3750 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3751 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3752 atomic_set(&nohz
.load_balancer
, -1);
3756 if (atomic_read(&nohz
.load_balancer
) == -1) {
3757 /* make me the ilb owner */
3758 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3760 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3763 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3766 cpu_clear(cpu
, nohz
.cpu_mask
);
3768 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3769 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3776 static DEFINE_SPINLOCK(balancing
);
3779 * It checks each scheduling domain to see if it is due to be balanced,
3780 * and initiates a balancing operation if so.
3782 * Balancing parameters are set up in arch_init_sched_domains.
3784 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3787 struct rq
*rq
= cpu_rq(cpu
);
3788 unsigned long interval
;
3789 struct sched_domain
*sd
;
3790 /* Earliest time when we have to do rebalance again */
3791 unsigned long next_balance
= jiffies
+ 60*HZ
;
3792 int update_next_balance
= 0;
3796 for_each_domain(cpu
, sd
) {
3797 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3800 interval
= sd
->balance_interval
;
3801 if (idle
!= CPU_IDLE
)
3802 interval
*= sd
->busy_factor
;
3804 /* scale ms to jiffies */
3805 interval
= msecs_to_jiffies(interval
);
3806 if (unlikely(!interval
))
3808 if (interval
> HZ
*NR_CPUS
/10)
3809 interval
= HZ
*NR_CPUS
/10;
3811 need_serialize
= sd
->flags
& SD_SERIALIZE
;
3813 if (need_serialize
) {
3814 if (!spin_trylock(&balancing
))
3818 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3819 if (load_balance(cpu
, rq
, sd
, idle
, &balance
, &tmp
)) {
3821 * We've pulled tasks over so either we're no
3822 * longer idle, or one of our SMT siblings is
3825 idle
= CPU_NOT_IDLE
;
3827 sd
->last_balance
= jiffies
;
3830 spin_unlock(&balancing
);
3832 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3833 next_balance
= sd
->last_balance
+ interval
;
3834 update_next_balance
= 1;
3838 * Stop the load balance at this level. There is another
3839 * CPU in our sched group which is doing load balancing more
3847 * next_balance will be updated only when there is a need.
3848 * When the cpu is attached to null domain for ex, it will not be
3851 if (likely(update_next_balance
))
3852 rq
->next_balance
= next_balance
;
3856 * run_rebalance_domains is triggered when needed from the scheduler tick.
3857 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3858 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3860 static void run_rebalance_domains(struct softirq_action
*h
)
3862 int this_cpu
= smp_processor_id();
3863 struct rq
*this_rq
= cpu_rq(this_cpu
);
3864 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3865 CPU_IDLE
: CPU_NOT_IDLE
;
3867 rebalance_domains(this_cpu
, idle
);
3871 * If this cpu is the owner for idle load balancing, then do the
3872 * balancing on behalf of the other idle cpus whose ticks are
3875 if (this_rq
->idle_at_tick
&&
3876 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3877 cpumask_t cpus
= nohz
.cpu_mask
;
3881 cpu_clear(this_cpu
, cpus
);
3882 for_each_cpu_mask(balance_cpu
, cpus
) {
3884 * If this cpu gets work to do, stop the load balancing
3885 * work being done for other cpus. Next load
3886 * balancing owner will pick it up.
3891 rebalance_domains(balance_cpu
, CPU_IDLE
);
3893 rq
= cpu_rq(balance_cpu
);
3894 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3895 this_rq
->next_balance
= rq
->next_balance
;
3902 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3904 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3905 * idle load balancing owner or decide to stop the periodic load balancing,
3906 * if the whole system is idle.
3908 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3912 * If we were in the nohz mode recently and busy at the current
3913 * scheduler tick, then check if we need to nominate new idle
3916 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3917 rq
->in_nohz_recently
= 0;
3919 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3920 cpu_clear(cpu
, nohz
.cpu_mask
);
3921 atomic_set(&nohz
.load_balancer
, -1);
3924 if (atomic_read(&nohz
.load_balancer
) == -1) {
3926 * simple selection for now: Nominate the
3927 * first cpu in the nohz list to be the next
3930 * TBD: Traverse the sched domains and nominate
3931 * the nearest cpu in the nohz.cpu_mask.
3933 int ilb
= first_cpu(nohz
.cpu_mask
);
3935 if (ilb
< nr_cpu_ids
)
3941 * If this cpu is idle and doing idle load balancing for all the
3942 * cpus with ticks stopped, is it time for that to stop?
3944 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3945 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3951 * If this cpu is idle and the idle load balancing is done by
3952 * someone else, then no need raise the SCHED_SOFTIRQ
3954 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3955 cpu_isset(cpu
, nohz
.cpu_mask
))
3958 if (time_after_eq(jiffies
, rq
->next_balance
))
3959 raise_softirq(SCHED_SOFTIRQ
);
3962 #else /* CONFIG_SMP */
3965 * on UP we do not need to balance between CPUs:
3967 static inline void idle_balance(int cpu
, struct rq
*rq
)
3973 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3975 EXPORT_PER_CPU_SYMBOL(kstat
);
3978 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3979 * that have not yet been banked in case the task is currently running.
3981 unsigned long long task_sched_runtime(struct task_struct
*p
)
3983 unsigned long flags
;
3987 rq
= task_rq_lock(p
, &flags
);
3988 ns
= p
->se
.sum_exec_runtime
;
3989 if (task_current(rq
, p
)) {
3990 update_rq_clock(rq
);
3991 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3992 if ((s64
)delta_exec
> 0)
3995 task_rq_unlock(rq
, &flags
);
4001 * Account user cpu time to a process.
4002 * @p: the process that the cpu time gets accounted to
4003 * @cputime: the cpu time spent in user space since the last update
4005 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
4007 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4010 p
->utime
= cputime_add(p
->utime
, cputime
);
4012 /* Add user time to cpustat. */
4013 tmp
= cputime_to_cputime64(cputime
);
4014 if (TASK_NICE(p
) > 0)
4015 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
4017 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4021 * Account guest cpu time to a process.
4022 * @p: the process that the cpu time gets accounted to
4023 * @cputime: the cpu time spent in virtual machine since the last update
4025 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
)
4028 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4030 tmp
= cputime_to_cputime64(cputime
);
4032 p
->utime
= cputime_add(p
->utime
, cputime
);
4033 p
->gtime
= cputime_add(p
->gtime
, cputime
);
4035 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4036 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
4040 * Account scaled user cpu time to a process.
4041 * @p: the process that the cpu time gets accounted to
4042 * @cputime: the cpu time spent in user space since the last update
4044 void account_user_time_scaled(struct task_struct
*p
, cputime_t cputime
)
4046 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime
);
4050 * Account system cpu time to a process.
4051 * @p: the process that the cpu time gets accounted to
4052 * @hardirq_offset: the offset to subtract from hardirq_count()
4053 * @cputime: the cpu time spent in kernel space since the last update
4055 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
4058 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4059 struct rq
*rq
= this_rq();
4062 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
4063 account_guest_time(p
, cputime
);
4067 p
->stime
= cputime_add(p
->stime
, cputime
);
4069 /* Add system time to cpustat. */
4070 tmp
= cputime_to_cputime64(cputime
);
4071 if (hardirq_count() - hardirq_offset
)
4072 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
4073 else if (softirq_count())
4074 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
4075 else if (p
!= rq
->idle
)
4076 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
4077 else if (atomic_read(&rq
->nr_iowait
) > 0)
4078 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
4080 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
4081 /* Account for system time used */
4082 acct_update_integrals(p
);
4086 * Account scaled system cpu time to a process.
4087 * @p: the process that the cpu time gets accounted to
4088 * @hardirq_offset: the offset to subtract from hardirq_count()
4089 * @cputime: the cpu time spent in kernel space since the last update
4091 void account_system_time_scaled(struct task_struct
*p
, cputime_t cputime
)
4093 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime
);
4097 * Account for involuntary wait time.
4098 * @p: the process from which the cpu time has been stolen
4099 * @steal: the cpu time spent in involuntary wait
4101 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
4103 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4104 cputime64_t tmp
= cputime_to_cputime64(steal
);
4105 struct rq
*rq
= this_rq();
4107 if (p
== rq
->idle
) {
4108 p
->stime
= cputime_add(p
->stime
, steal
);
4109 if (atomic_read(&rq
->nr_iowait
) > 0)
4110 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
4112 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
4114 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
4118 * This function gets called by the timer code, with HZ frequency.
4119 * We call it with interrupts disabled.
4121 * It also gets called by the fork code, when changing the parent's
4124 void scheduler_tick(void)
4126 int cpu
= smp_processor_id();
4127 struct rq
*rq
= cpu_rq(cpu
);
4128 struct task_struct
*curr
= rq
->curr
;
4132 spin_lock(&rq
->lock
);
4133 update_rq_clock(rq
);
4134 update_cpu_load(rq
);
4135 curr
->sched_class
->task_tick(rq
, curr
, 0);
4136 spin_unlock(&rq
->lock
);
4139 rq
->idle_at_tick
= idle_cpu(cpu
);
4140 trigger_load_balance(rq
, cpu
);
4144 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
4146 void __kprobes
add_preempt_count(int val
)
4151 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4153 preempt_count() += val
;
4155 * Spinlock count overflowing soon?
4157 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4160 EXPORT_SYMBOL(add_preempt_count
);
4162 void __kprobes
sub_preempt_count(int val
)
4167 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
4170 * Is the spinlock portion underflowing?
4172 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4173 !(preempt_count() & PREEMPT_MASK
)))
4176 preempt_count() -= val
;
4178 EXPORT_SYMBOL(sub_preempt_count
);
4183 * Print scheduling while atomic bug:
4185 static noinline
void __schedule_bug(struct task_struct
*prev
)
4187 struct pt_regs
*regs
= get_irq_regs();
4189 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4190 prev
->comm
, prev
->pid
, preempt_count());
4192 debug_show_held_locks(prev
);
4194 if (irqs_disabled())
4195 print_irqtrace_events(prev
);
4204 * Various schedule()-time debugging checks and statistics:
4206 static inline void schedule_debug(struct task_struct
*prev
)
4209 * Test if we are atomic. Since do_exit() needs to call into
4210 * schedule() atomically, we ignore that path for now.
4211 * Otherwise, whine if we are scheduling when we should not be.
4213 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
4214 __schedule_bug(prev
);
4216 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4218 schedstat_inc(this_rq(), sched_count
);
4219 #ifdef CONFIG_SCHEDSTATS
4220 if (unlikely(prev
->lock_depth
>= 0)) {
4221 schedstat_inc(this_rq(), bkl_count
);
4222 schedstat_inc(prev
, sched_info
.bkl_count
);
4228 * Pick up the highest-prio task:
4230 static inline struct task_struct
*
4231 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
4233 const struct sched_class
*class;
4234 struct task_struct
*p
;
4237 * Optimization: we know that if all tasks are in
4238 * the fair class we can call that function directly:
4240 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
4241 p
= fair_sched_class
.pick_next_task(rq
);
4246 class = sched_class_highest
;
4248 p
= class->pick_next_task(rq
);
4252 * Will never be NULL as the idle class always
4253 * returns a non-NULL p:
4255 class = class->next
;
4260 * schedule() is the main scheduler function.
4262 asmlinkage
void __sched
schedule(void)
4264 struct task_struct
*prev
, *next
;
4265 unsigned long *switch_count
;
4267 int cpu
, hrtick
= sched_feat(HRTICK
);
4271 cpu
= smp_processor_id();
4275 switch_count
= &prev
->nivcsw
;
4277 release_kernel_lock(prev
);
4278 need_resched_nonpreemptible
:
4280 schedule_debug(prev
);
4286 * Do the rq-clock update outside the rq lock:
4288 local_irq_disable();
4289 update_rq_clock(rq
);
4290 spin_lock(&rq
->lock
);
4291 clear_tsk_need_resched(prev
);
4293 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
4294 if (unlikely(signal_pending_state(prev
->state
, prev
)))
4295 prev
->state
= TASK_RUNNING
;
4297 deactivate_task(rq
, prev
, 1);
4298 switch_count
= &prev
->nvcsw
;
4302 if (prev
->sched_class
->pre_schedule
)
4303 prev
->sched_class
->pre_schedule(rq
, prev
);
4306 if (unlikely(!rq
->nr_running
))
4307 idle_balance(cpu
, rq
);
4309 prev
->sched_class
->put_prev_task(rq
, prev
);
4310 next
= pick_next_task(rq
, prev
);
4312 if (likely(prev
!= next
)) {
4313 sched_info_switch(prev
, next
);
4319 context_switch(rq
, prev
, next
); /* unlocks the rq */
4321 * the context switch might have flipped the stack from under
4322 * us, hence refresh the local variables.
4324 cpu
= smp_processor_id();
4327 spin_unlock_irq(&rq
->lock
);
4332 if (unlikely(reacquire_kernel_lock(current
) < 0))
4333 goto need_resched_nonpreemptible
;
4335 preempt_enable_no_resched();
4336 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
4339 EXPORT_SYMBOL(schedule
);
4341 #ifdef CONFIG_PREEMPT
4343 * this is the entry point to schedule() from in-kernel preemption
4344 * off of preempt_enable. Kernel preemptions off return from interrupt
4345 * occur there and call schedule directly.
4347 asmlinkage
void __sched
preempt_schedule(void)
4349 struct thread_info
*ti
= current_thread_info();
4352 * If there is a non-zero preempt_count or interrupts are disabled,
4353 * we do not want to preempt the current task. Just return..
4355 if (likely(ti
->preempt_count
|| irqs_disabled()))
4359 add_preempt_count(PREEMPT_ACTIVE
);
4361 sub_preempt_count(PREEMPT_ACTIVE
);
4364 * Check again in case we missed a preemption opportunity
4365 * between schedule and now.
4368 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4370 EXPORT_SYMBOL(preempt_schedule
);
4373 * this is the entry point to schedule() from kernel preemption
4374 * off of irq context.
4375 * Note, that this is called and return with irqs disabled. This will
4376 * protect us against recursive calling from irq.
4378 asmlinkage
void __sched
preempt_schedule_irq(void)
4380 struct thread_info
*ti
= current_thread_info();
4382 /* Catch callers which need to be fixed */
4383 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
4386 add_preempt_count(PREEMPT_ACTIVE
);
4389 local_irq_disable();
4390 sub_preempt_count(PREEMPT_ACTIVE
);
4393 * Check again in case we missed a preemption opportunity
4394 * between schedule and now.
4397 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4400 #endif /* CONFIG_PREEMPT */
4402 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
4405 return try_to_wake_up(curr
->private, mode
, sync
);
4407 EXPORT_SYMBOL(default_wake_function
);
4410 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4411 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4412 * number) then we wake all the non-exclusive tasks and one exclusive task.
4414 * There are circumstances in which we can try to wake a task which has already
4415 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4416 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4418 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4419 int nr_exclusive
, int sync
, void *key
)
4421 wait_queue_t
*curr
, *next
;
4423 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4424 unsigned flags
= curr
->flags
;
4426 if (curr
->func(curr
, mode
, sync
, key
) &&
4427 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4433 * __wake_up - wake up threads blocked on a waitqueue.
4435 * @mode: which threads
4436 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4437 * @key: is directly passed to the wakeup function
4439 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4440 int nr_exclusive
, void *key
)
4442 unsigned long flags
;
4444 spin_lock_irqsave(&q
->lock
, flags
);
4445 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4446 spin_unlock_irqrestore(&q
->lock
, flags
);
4448 EXPORT_SYMBOL(__wake_up
);
4451 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4453 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4455 __wake_up_common(q
, mode
, 1, 0, NULL
);
4459 * __wake_up_sync - wake up threads blocked on a waitqueue.
4461 * @mode: which threads
4462 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4464 * The sync wakeup differs that the waker knows that it will schedule
4465 * away soon, so while the target thread will be woken up, it will not
4466 * be migrated to another CPU - ie. the two threads are 'synchronized'
4467 * with each other. This can prevent needless bouncing between CPUs.
4469 * On UP it can prevent extra preemption.
4472 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4474 unsigned long flags
;
4480 if (unlikely(!nr_exclusive
))
4483 spin_lock_irqsave(&q
->lock
, flags
);
4484 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
4485 spin_unlock_irqrestore(&q
->lock
, flags
);
4487 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4489 void complete(struct completion
*x
)
4491 unsigned long flags
;
4493 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4495 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4496 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4498 EXPORT_SYMBOL(complete
);
4500 void complete_all(struct completion
*x
)
4502 unsigned long flags
;
4504 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4505 x
->done
+= UINT_MAX
/2;
4506 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4507 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4509 EXPORT_SYMBOL(complete_all
);
4511 static inline long __sched
4512 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4515 DECLARE_WAITQUEUE(wait
, current
);
4517 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4518 __add_wait_queue_tail(&x
->wait
, &wait
);
4520 if ((state
== TASK_INTERRUPTIBLE
&&
4521 signal_pending(current
)) ||
4522 (state
== TASK_KILLABLE
&&
4523 fatal_signal_pending(current
))) {
4524 timeout
= -ERESTARTSYS
;
4527 __set_current_state(state
);
4528 spin_unlock_irq(&x
->wait
.lock
);
4529 timeout
= schedule_timeout(timeout
);
4530 spin_lock_irq(&x
->wait
.lock
);
4531 } while (!x
->done
&& timeout
);
4532 __remove_wait_queue(&x
->wait
, &wait
);
4537 return timeout
?: 1;
4541 wait_for_common(struct completion
*x
, long timeout
, int state
)
4545 spin_lock_irq(&x
->wait
.lock
);
4546 timeout
= do_wait_for_common(x
, timeout
, state
);
4547 spin_unlock_irq(&x
->wait
.lock
);
4551 void __sched
wait_for_completion(struct completion
*x
)
4553 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4555 EXPORT_SYMBOL(wait_for_completion
);
4557 unsigned long __sched
4558 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4560 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4562 EXPORT_SYMBOL(wait_for_completion_timeout
);
4564 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4566 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4567 if (t
== -ERESTARTSYS
)
4571 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4573 unsigned long __sched
4574 wait_for_completion_interruptible_timeout(struct completion
*x
,
4575 unsigned long timeout
)
4577 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4579 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4581 int __sched
wait_for_completion_killable(struct completion
*x
)
4583 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4584 if (t
== -ERESTARTSYS
)
4588 EXPORT_SYMBOL(wait_for_completion_killable
);
4591 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4593 unsigned long flags
;
4596 init_waitqueue_entry(&wait
, current
);
4598 __set_current_state(state
);
4600 spin_lock_irqsave(&q
->lock
, flags
);
4601 __add_wait_queue(q
, &wait
);
4602 spin_unlock(&q
->lock
);
4603 timeout
= schedule_timeout(timeout
);
4604 spin_lock_irq(&q
->lock
);
4605 __remove_wait_queue(q
, &wait
);
4606 spin_unlock_irqrestore(&q
->lock
, flags
);
4611 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4613 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4615 EXPORT_SYMBOL(interruptible_sleep_on
);
4618 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4620 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4622 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4624 void __sched
sleep_on(wait_queue_head_t
*q
)
4626 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4628 EXPORT_SYMBOL(sleep_on
);
4630 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4632 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4634 EXPORT_SYMBOL(sleep_on_timeout
);
4636 #ifdef CONFIG_RT_MUTEXES
4639 * rt_mutex_setprio - set the current priority of a task
4641 * @prio: prio value (kernel-internal form)
4643 * This function changes the 'effective' priority of a task. It does
4644 * not touch ->normal_prio like __setscheduler().
4646 * Used by the rt_mutex code to implement priority inheritance logic.
4648 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4650 unsigned long flags
;
4651 int oldprio
, on_rq
, running
;
4653 const struct sched_class
*prev_class
= p
->sched_class
;
4655 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4657 rq
= task_rq_lock(p
, &flags
);
4658 update_rq_clock(rq
);
4661 on_rq
= p
->se
.on_rq
;
4662 running
= task_current(rq
, p
);
4664 dequeue_task(rq
, p
, 0);
4666 p
->sched_class
->put_prev_task(rq
, p
);
4669 p
->sched_class
= &rt_sched_class
;
4671 p
->sched_class
= &fair_sched_class
;
4676 p
->sched_class
->set_curr_task(rq
);
4678 enqueue_task(rq
, p
, 0);
4680 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4682 task_rq_unlock(rq
, &flags
);
4687 void set_user_nice(struct task_struct
*p
, long nice
)
4689 int old_prio
, delta
, on_rq
;
4690 unsigned long flags
;
4693 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4696 * We have to be careful, if called from sys_setpriority(),
4697 * the task might be in the middle of scheduling on another CPU.
4699 rq
= task_rq_lock(p
, &flags
);
4700 update_rq_clock(rq
);
4702 * The RT priorities are set via sched_setscheduler(), but we still
4703 * allow the 'normal' nice value to be set - but as expected
4704 * it wont have any effect on scheduling until the task is
4705 * SCHED_FIFO/SCHED_RR:
4707 if (task_has_rt_policy(p
)) {
4708 p
->static_prio
= NICE_TO_PRIO(nice
);
4711 on_rq
= p
->se
.on_rq
;
4713 dequeue_task(rq
, p
, 0);
4715 p
->static_prio
= NICE_TO_PRIO(nice
);
4718 p
->prio
= effective_prio(p
);
4719 delta
= p
->prio
- old_prio
;
4722 enqueue_task(rq
, p
, 0);
4724 * If the task increased its priority or is running and
4725 * lowered its priority, then reschedule its CPU:
4727 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4728 resched_task(rq
->curr
);
4731 task_rq_unlock(rq
, &flags
);
4733 EXPORT_SYMBOL(set_user_nice
);
4736 * can_nice - check if a task can reduce its nice value
4740 int can_nice(const struct task_struct
*p
, const int nice
)
4742 /* convert nice value [19,-20] to rlimit style value [1,40] */
4743 int nice_rlim
= 20 - nice
;
4745 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4746 capable(CAP_SYS_NICE
));
4749 #ifdef __ARCH_WANT_SYS_NICE
4752 * sys_nice - change the priority of the current process.
4753 * @increment: priority increment
4755 * sys_setpriority is a more generic, but much slower function that
4756 * does similar things.
4758 asmlinkage
long sys_nice(int increment
)
4763 * Setpriority might change our priority at the same moment.
4764 * We don't have to worry. Conceptually one call occurs first
4765 * and we have a single winner.
4767 if (increment
< -40)
4772 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4778 if (increment
< 0 && !can_nice(current
, nice
))
4781 retval
= security_task_setnice(current
, nice
);
4785 set_user_nice(current
, nice
);
4792 * task_prio - return the priority value of a given task.
4793 * @p: the task in question.
4795 * This is the priority value as seen by users in /proc.
4796 * RT tasks are offset by -200. Normal tasks are centered
4797 * around 0, value goes from -16 to +15.
4799 int task_prio(const struct task_struct
*p
)
4801 return p
->prio
- MAX_RT_PRIO
;
4805 * task_nice - return the nice value of a given task.
4806 * @p: the task in question.
4808 int task_nice(const struct task_struct
*p
)
4810 return TASK_NICE(p
);
4812 EXPORT_SYMBOL(task_nice
);
4815 * idle_cpu - is a given cpu idle currently?
4816 * @cpu: the processor in question.
4818 int idle_cpu(int cpu
)
4820 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4824 * idle_task - return the idle task for a given cpu.
4825 * @cpu: the processor in question.
4827 struct task_struct
*idle_task(int cpu
)
4829 return cpu_rq(cpu
)->idle
;
4833 * find_process_by_pid - find a process with a matching PID value.
4834 * @pid: the pid in question.
4836 static struct task_struct
*find_process_by_pid(pid_t pid
)
4838 return pid
? find_task_by_vpid(pid
) : current
;
4841 /* Actually do priority change: must hold rq lock. */
4843 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4845 BUG_ON(p
->se
.on_rq
);
4848 switch (p
->policy
) {
4852 p
->sched_class
= &fair_sched_class
;
4856 p
->sched_class
= &rt_sched_class
;
4860 p
->rt_priority
= prio
;
4861 p
->normal_prio
= normal_prio(p
);
4862 /* we are holding p->pi_lock already */
4863 p
->prio
= rt_mutex_getprio(p
);
4868 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4869 * @p: the task in question.
4870 * @policy: new policy.
4871 * @param: structure containing the new RT priority.
4873 * NOTE that the task may be already dead.
4875 int sched_setscheduler(struct task_struct
*p
, int policy
,
4876 struct sched_param
*param
)
4878 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4879 unsigned long flags
;
4880 const struct sched_class
*prev_class
= p
->sched_class
;
4883 /* may grab non-irq protected spin_locks */
4884 BUG_ON(in_interrupt());
4886 /* double check policy once rq lock held */
4888 policy
= oldpolicy
= p
->policy
;
4889 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4890 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4891 policy
!= SCHED_IDLE
)
4894 * Valid priorities for SCHED_FIFO and SCHED_RR are
4895 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4896 * SCHED_BATCH and SCHED_IDLE is 0.
4898 if (param
->sched_priority
< 0 ||
4899 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4900 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4902 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4906 * Allow unprivileged RT tasks to decrease priority:
4908 if (!capable(CAP_SYS_NICE
)) {
4909 if (rt_policy(policy
)) {
4910 unsigned long rlim_rtprio
;
4912 if (!lock_task_sighand(p
, &flags
))
4914 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4915 unlock_task_sighand(p
, &flags
);
4917 /* can't set/change the rt policy */
4918 if (policy
!= p
->policy
&& !rlim_rtprio
)
4921 /* can't increase priority */
4922 if (param
->sched_priority
> p
->rt_priority
&&
4923 param
->sched_priority
> rlim_rtprio
)
4927 * Like positive nice levels, dont allow tasks to
4928 * move out of SCHED_IDLE either:
4930 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4933 /* can't change other user's priorities */
4934 if ((current
->euid
!= p
->euid
) &&
4935 (current
->euid
!= p
->uid
))
4939 #ifdef CONFIG_RT_GROUP_SCHED
4941 * Do not allow realtime tasks into groups that have no runtime
4944 if (rt_policy(policy
) && task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
4948 retval
= security_task_setscheduler(p
, policy
, param
);
4952 * make sure no PI-waiters arrive (or leave) while we are
4953 * changing the priority of the task:
4955 spin_lock_irqsave(&p
->pi_lock
, flags
);
4957 * To be able to change p->policy safely, the apropriate
4958 * runqueue lock must be held.
4960 rq
= __task_rq_lock(p
);
4961 /* recheck policy now with rq lock held */
4962 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4963 policy
= oldpolicy
= -1;
4964 __task_rq_unlock(rq
);
4965 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4968 update_rq_clock(rq
);
4969 on_rq
= p
->se
.on_rq
;
4970 running
= task_current(rq
, p
);
4972 deactivate_task(rq
, p
, 0);
4974 p
->sched_class
->put_prev_task(rq
, p
);
4977 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4980 p
->sched_class
->set_curr_task(rq
);
4982 activate_task(rq
, p
, 0);
4984 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4986 __task_rq_unlock(rq
);
4987 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4989 rt_mutex_adjust_pi(p
);
4993 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4996 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4998 struct sched_param lparam
;
4999 struct task_struct
*p
;
5002 if (!param
|| pid
< 0)
5004 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
5009 p
= find_process_by_pid(pid
);
5011 retval
= sched_setscheduler(p
, policy
, &lparam
);
5018 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5019 * @pid: the pid in question.
5020 * @policy: new policy.
5021 * @param: structure containing the new RT priority.
5024 sys_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5026 /* negative values for policy are not valid */
5030 return do_sched_setscheduler(pid
, policy
, param
);
5034 * sys_sched_setparam - set/change the RT priority of a thread
5035 * @pid: the pid in question.
5036 * @param: structure containing the new RT priority.
5038 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
5040 return do_sched_setscheduler(pid
, -1, param
);
5044 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5045 * @pid: the pid in question.
5047 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
5049 struct task_struct
*p
;
5056 read_lock(&tasklist_lock
);
5057 p
= find_process_by_pid(pid
);
5059 retval
= security_task_getscheduler(p
);
5063 read_unlock(&tasklist_lock
);
5068 * sys_sched_getscheduler - get the RT priority of a thread
5069 * @pid: the pid in question.
5070 * @param: structure containing the RT priority.
5072 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
5074 struct sched_param lp
;
5075 struct task_struct
*p
;
5078 if (!param
|| pid
< 0)
5081 read_lock(&tasklist_lock
);
5082 p
= find_process_by_pid(pid
);
5087 retval
= security_task_getscheduler(p
);
5091 lp
.sched_priority
= p
->rt_priority
;
5092 read_unlock(&tasklist_lock
);
5095 * This one might sleep, we cannot do it with a spinlock held ...
5097 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
5102 read_unlock(&tasklist_lock
);
5106 long sched_setaffinity(pid_t pid
, const cpumask_t
*in_mask
)
5108 cpumask_t cpus_allowed
;
5109 cpumask_t new_mask
= *in_mask
;
5110 struct task_struct
*p
;
5114 read_lock(&tasklist_lock
);
5116 p
= find_process_by_pid(pid
);
5118 read_unlock(&tasklist_lock
);
5124 * It is not safe to call set_cpus_allowed with the
5125 * tasklist_lock held. We will bump the task_struct's
5126 * usage count and then drop tasklist_lock.
5129 read_unlock(&tasklist_lock
);
5132 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
5133 !capable(CAP_SYS_NICE
))
5136 retval
= security_task_setscheduler(p
, 0, NULL
);
5140 cpuset_cpus_allowed(p
, &cpus_allowed
);
5141 cpus_and(new_mask
, new_mask
, cpus_allowed
);
5143 retval
= set_cpus_allowed_ptr(p
, &new_mask
);
5146 cpuset_cpus_allowed(p
, &cpus_allowed
);
5147 if (!cpus_subset(new_mask
, cpus_allowed
)) {
5149 * We must have raced with a concurrent cpuset
5150 * update. Just reset the cpus_allowed to the
5151 * cpuset's cpus_allowed
5153 new_mask
= cpus_allowed
;
5163 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5164 cpumask_t
*new_mask
)
5166 if (len
< sizeof(cpumask_t
)) {
5167 memset(new_mask
, 0, sizeof(cpumask_t
));
5168 } else if (len
> sizeof(cpumask_t
)) {
5169 len
= sizeof(cpumask_t
);
5171 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
5175 * sys_sched_setaffinity - set the cpu affinity of a process
5176 * @pid: pid of the process
5177 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5178 * @user_mask_ptr: user-space pointer to the new cpu mask
5180 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
5181 unsigned long __user
*user_mask_ptr
)
5186 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
5190 return sched_setaffinity(pid
, &new_mask
);
5193 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
5195 struct task_struct
*p
;
5199 read_lock(&tasklist_lock
);
5202 p
= find_process_by_pid(pid
);
5206 retval
= security_task_getscheduler(p
);
5210 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
5213 read_unlock(&tasklist_lock
);
5220 * sys_sched_getaffinity - get the cpu affinity of a process
5221 * @pid: pid of the process
5222 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5223 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5225 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
5226 unsigned long __user
*user_mask_ptr
)
5231 if (len
< sizeof(cpumask_t
))
5234 ret
= sched_getaffinity(pid
, &mask
);
5238 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
5241 return sizeof(cpumask_t
);
5245 * sys_sched_yield - yield the current processor to other threads.
5247 * This function yields the current CPU to other tasks. If there are no
5248 * other threads running on this CPU then this function will return.
5250 asmlinkage
long sys_sched_yield(void)
5252 struct rq
*rq
= this_rq_lock();
5254 schedstat_inc(rq
, yld_count
);
5255 current
->sched_class
->yield_task(rq
);
5258 * Since we are going to call schedule() anyway, there's
5259 * no need to preempt or enable interrupts:
5261 __release(rq
->lock
);
5262 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5263 _raw_spin_unlock(&rq
->lock
);
5264 preempt_enable_no_resched();
5271 static void __cond_resched(void)
5273 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5274 __might_sleep(__FILE__
, __LINE__
);
5277 * The BKS might be reacquired before we have dropped
5278 * PREEMPT_ACTIVE, which could trigger a second
5279 * cond_resched() call.
5282 add_preempt_count(PREEMPT_ACTIVE
);
5284 sub_preempt_count(PREEMPT_ACTIVE
);
5285 } while (need_resched());
5288 int __sched
_cond_resched(void)
5290 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
5291 system_state
== SYSTEM_RUNNING
) {
5297 EXPORT_SYMBOL(_cond_resched
);
5300 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5301 * call schedule, and on return reacquire the lock.
5303 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5304 * operations here to prevent schedule() from being called twice (once via
5305 * spin_unlock(), once by hand).
5307 int cond_resched_lock(spinlock_t
*lock
)
5309 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
5312 if (spin_needbreak(lock
) || resched
) {
5314 if (resched
&& need_resched())
5323 EXPORT_SYMBOL(cond_resched_lock
);
5325 int __sched
cond_resched_softirq(void)
5327 BUG_ON(!in_softirq());
5329 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
5337 EXPORT_SYMBOL(cond_resched_softirq
);
5340 * yield - yield the current processor to other threads.
5342 * This is a shortcut for kernel-space yielding - it marks the
5343 * thread runnable and calls sys_sched_yield().
5345 void __sched
yield(void)
5347 set_current_state(TASK_RUNNING
);
5350 EXPORT_SYMBOL(yield
);
5353 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5354 * that process accounting knows that this is a task in IO wait state.
5356 * But don't do that if it is a deliberate, throttling IO wait (this task
5357 * has set its backing_dev_info: the queue against which it should throttle)
5359 void __sched
io_schedule(void)
5361 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5363 delayacct_blkio_start();
5364 atomic_inc(&rq
->nr_iowait
);
5366 atomic_dec(&rq
->nr_iowait
);
5367 delayacct_blkio_end();
5369 EXPORT_SYMBOL(io_schedule
);
5371 long __sched
io_schedule_timeout(long timeout
)
5373 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5376 delayacct_blkio_start();
5377 atomic_inc(&rq
->nr_iowait
);
5378 ret
= schedule_timeout(timeout
);
5379 atomic_dec(&rq
->nr_iowait
);
5380 delayacct_blkio_end();
5385 * sys_sched_get_priority_max - return maximum RT priority.
5386 * @policy: scheduling class.
5388 * this syscall returns the maximum rt_priority that can be used
5389 * by a given scheduling class.
5391 asmlinkage
long sys_sched_get_priority_max(int policy
)
5398 ret
= MAX_USER_RT_PRIO
-1;
5410 * sys_sched_get_priority_min - return minimum RT priority.
5411 * @policy: scheduling class.
5413 * this syscall returns the minimum rt_priority that can be used
5414 * by a given scheduling class.
5416 asmlinkage
long sys_sched_get_priority_min(int policy
)
5434 * sys_sched_rr_get_interval - return the default timeslice of a process.
5435 * @pid: pid of the process.
5436 * @interval: userspace pointer to the timeslice value.
5438 * this syscall writes the default timeslice value of a given process
5439 * into the user-space timespec buffer. A value of '0' means infinity.
5442 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
5444 struct task_struct
*p
;
5445 unsigned int time_slice
;
5453 read_lock(&tasklist_lock
);
5454 p
= find_process_by_pid(pid
);
5458 retval
= security_task_getscheduler(p
);
5463 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5464 * tasks that are on an otherwise idle runqueue:
5467 if (p
->policy
== SCHED_RR
) {
5468 time_slice
= DEF_TIMESLICE
;
5469 } else if (p
->policy
!= SCHED_FIFO
) {
5470 struct sched_entity
*se
= &p
->se
;
5471 unsigned long flags
;
5474 rq
= task_rq_lock(p
, &flags
);
5475 if (rq
->cfs
.load
.weight
)
5476 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
5477 task_rq_unlock(rq
, &flags
);
5479 read_unlock(&tasklist_lock
);
5480 jiffies_to_timespec(time_slice
, &t
);
5481 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5485 read_unlock(&tasklist_lock
);
5489 static const char stat_nam
[] = "RSDTtZX";
5491 void sched_show_task(struct task_struct
*p
)
5493 unsigned long free
= 0;
5496 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5497 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5498 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5499 #if BITS_PER_LONG == 32
5500 if (state
== TASK_RUNNING
)
5501 printk(KERN_CONT
" running ");
5503 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5505 if (state
== TASK_RUNNING
)
5506 printk(KERN_CONT
" running task ");
5508 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5510 #ifdef CONFIG_DEBUG_STACK_USAGE
5512 unsigned long *n
= end_of_stack(p
);
5515 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
5518 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
5519 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
5521 show_stack(p
, NULL
);
5524 void show_state_filter(unsigned long state_filter
)
5526 struct task_struct
*g
, *p
;
5528 #if BITS_PER_LONG == 32
5530 " task PC stack pid father\n");
5533 " task PC stack pid father\n");
5535 read_lock(&tasklist_lock
);
5536 do_each_thread(g
, p
) {
5538 * reset the NMI-timeout, listing all files on a slow
5539 * console might take alot of time:
5541 touch_nmi_watchdog();
5542 if (!state_filter
|| (p
->state
& state_filter
))
5544 } while_each_thread(g
, p
);
5546 touch_all_softlockup_watchdogs();
5548 #ifdef CONFIG_SCHED_DEBUG
5549 sysrq_sched_debug_show();
5551 read_unlock(&tasklist_lock
);
5553 * Only show locks if all tasks are dumped:
5555 if (state_filter
== -1)
5556 debug_show_all_locks();
5559 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5561 idle
->sched_class
= &idle_sched_class
;
5565 * init_idle - set up an idle thread for a given CPU
5566 * @idle: task in question
5567 * @cpu: cpu the idle task belongs to
5569 * NOTE: this function does not set the idle thread's NEED_RESCHED
5570 * flag, to make booting more robust.
5572 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5574 struct rq
*rq
= cpu_rq(cpu
);
5575 unsigned long flags
;
5578 idle
->se
.exec_start
= sched_clock();
5580 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
5581 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
5582 __set_task_cpu(idle
, cpu
);
5584 spin_lock_irqsave(&rq
->lock
, flags
);
5585 rq
->curr
= rq
->idle
= idle
;
5586 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5589 spin_unlock_irqrestore(&rq
->lock
, flags
);
5591 /* Set the preempt count _outside_ the spinlocks! */
5592 #if defined(CONFIG_PREEMPT)
5593 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5595 task_thread_info(idle
)->preempt_count
= 0;
5598 * The idle tasks have their own, simple scheduling class:
5600 idle
->sched_class
= &idle_sched_class
;
5604 * In a system that switches off the HZ timer nohz_cpu_mask
5605 * indicates which cpus entered this state. This is used
5606 * in the rcu update to wait only for active cpus. For system
5607 * which do not switch off the HZ timer nohz_cpu_mask should
5608 * always be CPU_MASK_NONE.
5610 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
5613 * Increase the granularity value when there are more CPUs,
5614 * because with more CPUs the 'effective latency' as visible
5615 * to users decreases. But the relationship is not linear,
5616 * so pick a second-best guess by going with the log2 of the
5619 * This idea comes from the SD scheduler of Con Kolivas:
5621 static inline void sched_init_granularity(void)
5623 unsigned int factor
= 1 + ilog2(num_online_cpus());
5624 const unsigned long limit
= 200000000;
5626 sysctl_sched_min_granularity
*= factor
;
5627 if (sysctl_sched_min_granularity
> limit
)
5628 sysctl_sched_min_granularity
= limit
;
5630 sysctl_sched_latency
*= factor
;
5631 if (sysctl_sched_latency
> limit
)
5632 sysctl_sched_latency
= limit
;
5634 sysctl_sched_wakeup_granularity
*= factor
;
5639 * This is how migration works:
5641 * 1) we queue a struct migration_req structure in the source CPU's
5642 * runqueue and wake up that CPU's migration thread.
5643 * 2) we down() the locked semaphore => thread blocks.
5644 * 3) migration thread wakes up (implicitly it forces the migrated
5645 * thread off the CPU)
5646 * 4) it gets the migration request and checks whether the migrated
5647 * task is still in the wrong runqueue.
5648 * 5) if it's in the wrong runqueue then the migration thread removes
5649 * it and puts it into the right queue.
5650 * 6) migration thread up()s the semaphore.
5651 * 7) we wake up and the migration is done.
5655 * Change a given task's CPU affinity. Migrate the thread to a
5656 * proper CPU and schedule it away if the CPU it's executing on
5657 * is removed from the allowed bitmask.
5659 * NOTE: the caller must have a valid reference to the task, the
5660 * task must not exit() & deallocate itself prematurely. The
5661 * call is not atomic; no spinlocks may be held.
5663 int set_cpus_allowed_ptr(struct task_struct
*p
, const cpumask_t
*new_mask
)
5665 struct migration_req req
;
5666 unsigned long flags
;
5670 rq
= task_rq_lock(p
, &flags
);
5671 if (!cpus_intersects(*new_mask
, cpu_online_map
)) {
5676 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
5677 !cpus_equal(p
->cpus_allowed
, *new_mask
))) {
5682 if (p
->sched_class
->set_cpus_allowed
)
5683 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
5685 p
->cpus_allowed
= *new_mask
;
5686 p
->rt
.nr_cpus_allowed
= cpus_weight(*new_mask
);
5689 /* Can the task run on the task's current CPU? If so, we're done */
5690 if (cpu_isset(task_cpu(p
), *new_mask
))
5693 if (migrate_task(p
, any_online_cpu(*new_mask
), &req
)) {
5694 /* Need help from migration thread: drop lock and wait. */
5695 task_rq_unlock(rq
, &flags
);
5696 wake_up_process(rq
->migration_thread
);
5697 wait_for_completion(&req
.done
);
5698 tlb_migrate_finish(p
->mm
);
5702 task_rq_unlock(rq
, &flags
);
5706 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5709 * Move (not current) task off this cpu, onto dest cpu. We're doing
5710 * this because either it can't run here any more (set_cpus_allowed()
5711 * away from this CPU, or CPU going down), or because we're
5712 * attempting to rebalance this task on exec (sched_exec).
5714 * So we race with normal scheduler movements, but that's OK, as long
5715 * as the task is no longer on this CPU.
5717 * Returns non-zero if task was successfully migrated.
5719 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5721 struct rq
*rq_dest
, *rq_src
;
5724 if (unlikely(cpu_is_offline(dest_cpu
)))
5727 rq_src
= cpu_rq(src_cpu
);
5728 rq_dest
= cpu_rq(dest_cpu
);
5730 double_rq_lock(rq_src
, rq_dest
);
5731 /* Already moved. */
5732 if (task_cpu(p
) != src_cpu
)
5734 /* Affinity changed (again). */
5735 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
5738 on_rq
= p
->se
.on_rq
;
5740 deactivate_task(rq_src
, p
, 0);
5742 set_task_cpu(p
, dest_cpu
);
5744 activate_task(rq_dest
, p
, 0);
5745 check_preempt_curr(rq_dest
, p
);
5749 double_rq_unlock(rq_src
, rq_dest
);
5754 * migration_thread - this is a highprio system thread that performs
5755 * thread migration by bumping thread off CPU then 'pushing' onto
5758 static int migration_thread(void *data
)
5760 int cpu
= (long)data
;
5764 BUG_ON(rq
->migration_thread
!= current
);
5766 set_current_state(TASK_INTERRUPTIBLE
);
5767 while (!kthread_should_stop()) {
5768 struct migration_req
*req
;
5769 struct list_head
*head
;
5771 spin_lock_irq(&rq
->lock
);
5773 if (cpu_is_offline(cpu
)) {
5774 spin_unlock_irq(&rq
->lock
);
5778 if (rq
->active_balance
) {
5779 active_load_balance(rq
, cpu
);
5780 rq
->active_balance
= 0;
5783 head
= &rq
->migration_queue
;
5785 if (list_empty(head
)) {
5786 spin_unlock_irq(&rq
->lock
);
5788 set_current_state(TASK_INTERRUPTIBLE
);
5791 req
= list_entry(head
->next
, struct migration_req
, list
);
5792 list_del_init(head
->next
);
5794 spin_unlock(&rq
->lock
);
5795 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5798 complete(&req
->done
);
5800 __set_current_state(TASK_RUNNING
);
5804 /* Wait for kthread_stop */
5805 set_current_state(TASK_INTERRUPTIBLE
);
5806 while (!kthread_should_stop()) {
5808 set_current_state(TASK_INTERRUPTIBLE
);
5810 __set_current_state(TASK_RUNNING
);
5814 #ifdef CONFIG_HOTPLUG_CPU
5816 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5820 local_irq_disable();
5821 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
5827 * Figure out where task on dead CPU should go, use force if necessary.
5828 * NOTE: interrupts should be disabled by the caller
5830 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5832 unsigned long flags
;
5839 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5840 cpus_and(mask
, mask
, p
->cpus_allowed
);
5841 dest_cpu
= any_online_cpu(mask
);
5843 /* On any allowed CPU? */
5844 if (dest_cpu
>= nr_cpu_ids
)
5845 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5847 /* No more Mr. Nice Guy. */
5848 if (dest_cpu
>= nr_cpu_ids
) {
5849 cpumask_t cpus_allowed
;
5851 cpuset_cpus_allowed_locked(p
, &cpus_allowed
);
5853 * Try to stay on the same cpuset, where the
5854 * current cpuset may be a subset of all cpus.
5855 * The cpuset_cpus_allowed_locked() variant of
5856 * cpuset_cpus_allowed() will not block. It must be
5857 * called within calls to cpuset_lock/cpuset_unlock.
5859 rq
= task_rq_lock(p
, &flags
);
5860 p
->cpus_allowed
= cpus_allowed
;
5861 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5862 task_rq_unlock(rq
, &flags
);
5865 * Don't tell them about moving exiting tasks or
5866 * kernel threads (both mm NULL), since they never
5869 if (p
->mm
&& printk_ratelimit()) {
5870 printk(KERN_INFO
"process %d (%s) no "
5871 "longer affine to cpu%d\n",
5872 task_pid_nr(p
), p
->comm
, dead_cpu
);
5875 } while (!__migrate_task_irq(p
, dead_cpu
, dest_cpu
));
5879 * While a dead CPU has no uninterruptible tasks queued at this point,
5880 * it might still have a nonzero ->nr_uninterruptible counter, because
5881 * for performance reasons the counter is not stricly tracking tasks to
5882 * their home CPUs. So we just add the counter to another CPU's counter,
5883 * to keep the global sum constant after CPU-down:
5885 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5887 struct rq
*rq_dest
= cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR
));
5888 unsigned long flags
;
5890 local_irq_save(flags
);
5891 double_rq_lock(rq_src
, rq_dest
);
5892 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5893 rq_src
->nr_uninterruptible
= 0;
5894 double_rq_unlock(rq_src
, rq_dest
);
5895 local_irq_restore(flags
);
5898 /* Run through task list and migrate tasks from the dead cpu. */
5899 static void migrate_live_tasks(int src_cpu
)
5901 struct task_struct
*p
, *t
;
5903 read_lock(&tasklist_lock
);
5905 do_each_thread(t
, p
) {
5909 if (task_cpu(p
) == src_cpu
)
5910 move_task_off_dead_cpu(src_cpu
, p
);
5911 } while_each_thread(t
, p
);
5913 read_unlock(&tasklist_lock
);
5917 * Schedules idle task to be the next runnable task on current CPU.
5918 * It does so by boosting its priority to highest possible.
5919 * Used by CPU offline code.
5921 void sched_idle_next(void)
5923 int this_cpu
= smp_processor_id();
5924 struct rq
*rq
= cpu_rq(this_cpu
);
5925 struct task_struct
*p
= rq
->idle
;
5926 unsigned long flags
;
5928 /* cpu has to be offline */
5929 BUG_ON(cpu_online(this_cpu
));
5932 * Strictly not necessary since rest of the CPUs are stopped by now
5933 * and interrupts disabled on the current cpu.
5935 spin_lock_irqsave(&rq
->lock
, flags
);
5937 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5939 update_rq_clock(rq
);
5940 activate_task(rq
, p
, 0);
5942 spin_unlock_irqrestore(&rq
->lock
, flags
);
5946 * Ensures that the idle task is using init_mm right before its cpu goes
5949 void idle_task_exit(void)
5951 struct mm_struct
*mm
= current
->active_mm
;
5953 BUG_ON(cpu_online(smp_processor_id()));
5956 switch_mm(mm
, &init_mm
, current
);
5960 /* called under rq->lock with disabled interrupts */
5961 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5963 struct rq
*rq
= cpu_rq(dead_cpu
);
5965 /* Must be exiting, otherwise would be on tasklist. */
5966 BUG_ON(!p
->exit_state
);
5968 /* Cannot have done final schedule yet: would have vanished. */
5969 BUG_ON(p
->state
== TASK_DEAD
);
5974 * Drop lock around migration; if someone else moves it,
5975 * that's OK. No task can be added to this CPU, so iteration is
5978 spin_unlock_irq(&rq
->lock
);
5979 move_task_off_dead_cpu(dead_cpu
, p
);
5980 spin_lock_irq(&rq
->lock
);
5985 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5986 static void migrate_dead_tasks(unsigned int dead_cpu
)
5988 struct rq
*rq
= cpu_rq(dead_cpu
);
5989 struct task_struct
*next
;
5992 if (!rq
->nr_running
)
5994 update_rq_clock(rq
);
5995 next
= pick_next_task(rq
, rq
->curr
);
5998 migrate_dead(dead_cpu
, next
);
6002 #endif /* CONFIG_HOTPLUG_CPU */
6004 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6006 static struct ctl_table sd_ctl_dir
[] = {
6008 .procname
= "sched_domain",
6014 static struct ctl_table sd_ctl_root
[] = {
6016 .ctl_name
= CTL_KERN
,
6017 .procname
= "kernel",
6019 .child
= sd_ctl_dir
,
6024 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
6026 struct ctl_table
*entry
=
6027 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
6032 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
6034 struct ctl_table
*entry
;
6037 * In the intermediate directories, both the child directory and
6038 * procname are dynamically allocated and could fail but the mode
6039 * will always be set. In the lowest directory the names are
6040 * static strings and all have proc handlers.
6042 for (entry
= *tablep
; entry
->mode
; entry
++) {
6044 sd_free_ctl_entry(&entry
->child
);
6045 if (entry
->proc_handler
== NULL
)
6046 kfree(entry
->procname
);
6054 set_table_entry(struct ctl_table
*entry
,
6055 const char *procname
, void *data
, int maxlen
,
6056 mode_t mode
, proc_handler
*proc_handler
)
6058 entry
->procname
= procname
;
6060 entry
->maxlen
= maxlen
;
6062 entry
->proc_handler
= proc_handler
;
6065 static struct ctl_table
*
6066 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
6068 struct ctl_table
*table
= sd_alloc_ctl_entry(12);
6073 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
6074 sizeof(long), 0644, proc_doulongvec_minmax
);
6075 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
6076 sizeof(long), 0644, proc_doulongvec_minmax
);
6077 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
6078 sizeof(int), 0644, proc_dointvec_minmax
);
6079 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
6080 sizeof(int), 0644, proc_dointvec_minmax
);
6081 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
6082 sizeof(int), 0644, proc_dointvec_minmax
);
6083 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
6084 sizeof(int), 0644, proc_dointvec_minmax
);
6085 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
6086 sizeof(int), 0644, proc_dointvec_minmax
);
6087 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
6088 sizeof(int), 0644, proc_dointvec_minmax
);
6089 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
6090 sizeof(int), 0644, proc_dointvec_minmax
);
6091 set_table_entry(&table
[9], "cache_nice_tries",
6092 &sd
->cache_nice_tries
,
6093 sizeof(int), 0644, proc_dointvec_minmax
);
6094 set_table_entry(&table
[10], "flags", &sd
->flags
,
6095 sizeof(int), 0644, proc_dointvec_minmax
);
6096 /* &table[11] is terminator */
6101 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
6103 struct ctl_table
*entry
, *table
;
6104 struct sched_domain
*sd
;
6105 int domain_num
= 0, i
;
6108 for_each_domain(cpu
, sd
)
6110 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
6115 for_each_domain(cpu
, sd
) {
6116 snprintf(buf
, 32, "domain%d", i
);
6117 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6119 entry
->child
= sd_alloc_ctl_domain_table(sd
);
6126 static struct ctl_table_header
*sd_sysctl_header
;
6127 static void register_sched_domain_sysctl(void)
6129 int i
, cpu_num
= num_online_cpus();
6130 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
6133 WARN_ON(sd_ctl_dir
[0].child
);
6134 sd_ctl_dir
[0].child
= entry
;
6139 for_each_online_cpu(i
) {
6140 snprintf(buf
, 32, "cpu%d", i
);
6141 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6143 entry
->child
= sd_alloc_ctl_cpu_table(i
);
6147 WARN_ON(sd_sysctl_header
);
6148 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
6151 /* may be called multiple times per register */
6152 static void unregister_sched_domain_sysctl(void)
6154 if (sd_sysctl_header
)
6155 unregister_sysctl_table(sd_sysctl_header
);
6156 sd_sysctl_header
= NULL
;
6157 if (sd_ctl_dir
[0].child
)
6158 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
6161 static void register_sched_domain_sysctl(void)
6164 static void unregister_sched_domain_sysctl(void)
6169 static void set_rq_online(struct rq
*rq
)
6172 const struct sched_class
*class;
6174 cpu_set(rq
->cpu
, rq
->rd
->online
);
6177 for_each_class(class) {
6178 if (class->rq_online
)
6179 class->rq_online(rq
);
6184 static void set_rq_offline(struct rq
*rq
)
6187 const struct sched_class
*class;
6189 for_each_class(class) {
6190 if (class->rq_offline
)
6191 class->rq_offline(rq
);
6194 cpu_clear(rq
->cpu
, rq
->rd
->online
);
6200 * migration_call - callback that gets triggered when a CPU is added.
6201 * Here we can start up the necessary migration thread for the new CPU.
6203 static int __cpuinit
6204 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
6206 struct task_struct
*p
;
6207 int cpu
= (long)hcpu
;
6208 unsigned long flags
;
6213 case CPU_UP_PREPARE
:
6214 case CPU_UP_PREPARE_FROZEN
:
6215 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
6218 kthread_bind(p
, cpu
);
6219 /* Must be high prio: stop_machine expects to yield to it. */
6220 rq
= task_rq_lock(p
, &flags
);
6221 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6222 task_rq_unlock(rq
, &flags
);
6223 cpu_rq(cpu
)->migration_thread
= p
;
6227 case CPU_ONLINE_FROZEN
:
6228 /* Strictly unnecessary, as first user will wake it. */
6229 wake_up_process(cpu_rq(cpu
)->migration_thread
);
6231 /* Update our root-domain */
6233 spin_lock_irqsave(&rq
->lock
, flags
);
6235 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6239 spin_unlock_irqrestore(&rq
->lock
, flags
);
6242 #ifdef CONFIG_HOTPLUG_CPU
6243 case CPU_UP_CANCELED
:
6244 case CPU_UP_CANCELED_FROZEN
:
6245 if (!cpu_rq(cpu
)->migration_thread
)
6247 /* Unbind it from offline cpu so it can run. Fall thru. */
6248 kthread_bind(cpu_rq(cpu
)->migration_thread
,
6249 any_online_cpu(cpu_online_map
));
6250 kthread_stop(cpu_rq(cpu
)->migration_thread
);
6251 cpu_rq(cpu
)->migration_thread
= NULL
;
6255 case CPU_DEAD_FROZEN
:
6256 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6257 migrate_live_tasks(cpu
);
6259 kthread_stop(rq
->migration_thread
);
6260 rq
->migration_thread
= NULL
;
6261 /* Idle task back to normal (off runqueue, low prio) */
6262 spin_lock_irq(&rq
->lock
);
6263 update_rq_clock(rq
);
6264 deactivate_task(rq
, rq
->idle
, 0);
6265 rq
->idle
->static_prio
= MAX_PRIO
;
6266 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
6267 rq
->idle
->sched_class
= &idle_sched_class
;
6268 migrate_dead_tasks(cpu
);
6269 spin_unlock_irq(&rq
->lock
);
6271 migrate_nr_uninterruptible(rq
);
6272 BUG_ON(rq
->nr_running
!= 0);
6275 * No need to migrate the tasks: it was best-effort if
6276 * they didn't take sched_hotcpu_mutex. Just wake up
6279 spin_lock_irq(&rq
->lock
);
6280 while (!list_empty(&rq
->migration_queue
)) {
6281 struct migration_req
*req
;
6283 req
= list_entry(rq
->migration_queue
.next
,
6284 struct migration_req
, list
);
6285 list_del_init(&req
->list
);
6286 complete(&req
->done
);
6288 spin_unlock_irq(&rq
->lock
);
6292 case CPU_DYING_FROZEN
:
6293 /* Update our root-domain */
6295 spin_lock_irqsave(&rq
->lock
, flags
);
6297 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6300 spin_unlock_irqrestore(&rq
->lock
, flags
);
6307 /* Register at highest priority so that task migration (migrate_all_tasks)
6308 * happens before everything else.
6310 static struct notifier_block __cpuinitdata migration_notifier
= {
6311 .notifier_call
= migration_call
,
6315 void __init
migration_init(void)
6317 void *cpu
= (void *)(long)smp_processor_id();
6320 /* Start one for the boot CPU: */
6321 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6322 BUG_ON(err
== NOTIFY_BAD
);
6323 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6324 register_cpu_notifier(&migration_notifier
);
6330 #ifdef CONFIG_SCHED_DEBUG
6332 static inline const char *sd_level_to_string(enum sched_domain_level lvl
)
6345 case SD_LV_ALLNODES
:
6354 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6355 cpumask_t
*groupmask
)
6357 struct sched_group
*group
= sd
->groups
;
6360 cpulist_scnprintf(str
, sizeof(str
), sd
->span
);
6361 cpus_clear(*groupmask
);
6363 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6365 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6366 printk("does not load-balance\n");
6368 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6373 printk(KERN_CONT
"span %s level %s\n",
6374 str
, sd_level_to_string(sd
->level
));
6376 if (!cpu_isset(cpu
, sd
->span
)) {
6377 printk(KERN_ERR
"ERROR: domain->span does not contain "
6380 if (!cpu_isset(cpu
, group
->cpumask
)) {
6381 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6385 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6389 printk(KERN_ERR
"ERROR: group is NULL\n");
6393 if (!group
->__cpu_power
) {
6394 printk(KERN_CONT
"\n");
6395 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6400 if (!cpus_weight(group
->cpumask
)) {
6401 printk(KERN_CONT
"\n");
6402 printk(KERN_ERR
"ERROR: empty group\n");
6406 if (cpus_intersects(*groupmask
, group
->cpumask
)) {
6407 printk(KERN_CONT
"\n");
6408 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6412 cpus_or(*groupmask
, *groupmask
, group
->cpumask
);
6414 cpulist_scnprintf(str
, sizeof(str
), group
->cpumask
);
6415 printk(KERN_CONT
" %s", str
);
6417 group
= group
->next
;
6418 } while (group
!= sd
->groups
);
6419 printk(KERN_CONT
"\n");
6421 if (!cpus_equal(sd
->span
, *groupmask
))
6422 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6424 if (sd
->parent
&& !cpus_subset(*groupmask
, sd
->parent
->span
))
6425 printk(KERN_ERR
"ERROR: parent span is not a superset "
6426 "of domain->span\n");
6430 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6432 cpumask_t
*groupmask
;
6436 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6440 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6442 groupmask
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
6444 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6449 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6458 #else /* !CONFIG_SCHED_DEBUG */
6459 # define sched_domain_debug(sd, cpu) do { } while (0)
6460 #endif /* CONFIG_SCHED_DEBUG */
6462 static int sd_degenerate(struct sched_domain
*sd
)
6464 if (cpus_weight(sd
->span
) == 1)
6467 /* Following flags need at least 2 groups */
6468 if (sd
->flags
& (SD_LOAD_BALANCE
|
6469 SD_BALANCE_NEWIDLE
|
6473 SD_SHARE_PKG_RESOURCES
)) {
6474 if (sd
->groups
!= sd
->groups
->next
)
6478 /* Following flags don't use groups */
6479 if (sd
->flags
& (SD_WAKE_IDLE
|
6488 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6490 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6492 if (sd_degenerate(parent
))
6495 if (!cpus_equal(sd
->span
, parent
->span
))
6498 /* Does parent contain flags not in child? */
6499 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6500 if (cflags
& SD_WAKE_AFFINE
)
6501 pflags
&= ~SD_WAKE_BALANCE
;
6502 /* Flags needing groups don't count if only 1 group in parent */
6503 if (parent
->groups
== parent
->groups
->next
) {
6504 pflags
&= ~(SD_LOAD_BALANCE
|
6505 SD_BALANCE_NEWIDLE
|
6509 SD_SHARE_PKG_RESOURCES
);
6511 if (~cflags
& pflags
)
6517 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6519 unsigned long flags
;
6521 spin_lock_irqsave(&rq
->lock
, flags
);
6524 struct root_domain
*old_rd
= rq
->rd
;
6526 if (cpu_isset(rq
->cpu
, old_rd
->online
))
6529 cpu_clear(rq
->cpu
, old_rd
->span
);
6531 if (atomic_dec_and_test(&old_rd
->refcount
))
6535 atomic_inc(&rd
->refcount
);
6538 cpu_set(rq
->cpu
, rd
->span
);
6539 if (cpu_isset(rq
->cpu
, cpu_online_map
))
6542 spin_unlock_irqrestore(&rq
->lock
, flags
);
6545 static void init_rootdomain(struct root_domain
*rd
)
6547 memset(rd
, 0, sizeof(*rd
));
6549 cpus_clear(rd
->span
);
6550 cpus_clear(rd
->online
);
6552 cpupri_init(&rd
->cpupri
);
6555 static void init_defrootdomain(void)
6557 init_rootdomain(&def_root_domain
);
6558 atomic_set(&def_root_domain
.refcount
, 1);
6561 static struct root_domain
*alloc_rootdomain(void)
6563 struct root_domain
*rd
;
6565 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6569 init_rootdomain(rd
);
6575 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6576 * hold the hotplug lock.
6579 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6581 struct rq
*rq
= cpu_rq(cpu
);
6582 struct sched_domain
*tmp
;
6584 /* Remove the sched domains which do not contribute to scheduling. */
6585 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
6586 struct sched_domain
*parent
= tmp
->parent
;
6589 if (sd_parent_degenerate(tmp
, parent
)) {
6590 tmp
->parent
= parent
->parent
;
6592 parent
->parent
->child
= tmp
;
6596 if (sd
&& sd_degenerate(sd
)) {
6602 sched_domain_debug(sd
, cpu
);
6604 rq_attach_root(rq
, rd
);
6605 rcu_assign_pointer(rq
->sd
, sd
);
6608 /* cpus with isolated domains */
6609 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
6611 /* Setup the mask of cpus configured for isolated domains */
6612 static int __init
isolated_cpu_setup(char *str
)
6614 int ints
[NR_CPUS
], i
;
6616 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
6617 cpus_clear(cpu_isolated_map
);
6618 for (i
= 1; i
<= ints
[0]; i
++)
6619 if (ints
[i
] < NR_CPUS
)
6620 cpu_set(ints
[i
], cpu_isolated_map
);
6624 __setup("isolcpus=", isolated_cpu_setup
);
6627 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6628 * to a function which identifies what group(along with sched group) a CPU
6629 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6630 * (due to the fact that we keep track of groups covered with a cpumask_t).
6632 * init_sched_build_groups will build a circular linked list of the groups
6633 * covered by the given span, and will set each group's ->cpumask correctly,
6634 * and ->cpu_power to 0.
6637 init_sched_build_groups(const cpumask_t
*span
, const cpumask_t
*cpu_map
,
6638 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
6639 struct sched_group
**sg
,
6640 cpumask_t
*tmpmask
),
6641 cpumask_t
*covered
, cpumask_t
*tmpmask
)
6643 struct sched_group
*first
= NULL
, *last
= NULL
;
6646 cpus_clear(*covered
);
6648 for_each_cpu_mask(i
, *span
) {
6649 struct sched_group
*sg
;
6650 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6653 if (cpu_isset(i
, *covered
))
6656 cpus_clear(sg
->cpumask
);
6657 sg
->__cpu_power
= 0;
6659 for_each_cpu_mask(j
, *span
) {
6660 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6663 cpu_set(j
, *covered
);
6664 cpu_set(j
, sg
->cpumask
);
6675 #define SD_NODES_PER_DOMAIN 16
6680 * find_next_best_node - find the next node to include in a sched_domain
6681 * @node: node whose sched_domain we're building
6682 * @used_nodes: nodes already in the sched_domain
6684 * Find the next node to include in a given scheduling domain. Simply
6685 * finds the closest node not already in the @used_nodes map.
6687 * Should use nodemask_t.
6689 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
6691 int i
, n
, val
, min_val
, best_node
= 0;
6695 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6696 /* Start at @node */
6697 n
= (node
+ i
) % MAX_NUMNODES
;
6699 if (!nr_cpus_node(n
))
6702 /* Skip already used nodes */
6703 if (node_isset(n
, *used_nodes
))
6706 /* Simple min distance search */
6707 val
= node_distance(node
, n
);
6709 if (val
< min_val
) {
6715 node_set(best_node
, *used_nodes
);
6720 * sched_domain_node_span - get a cpumask for a node's sched_domain
6721 * @node: node whose cpumask we're constructing
6722 * @span: resulting cpumask
6724 * Given a node, construct a good cpumask for its sched_domain to span. It
6725 * should be one that prevents unnecessary balancing, but also spreads tasks
6728 static void sched_domain_node_span(int node
, cpumask_t
*span
)
6730 nodemask_t used_nodes
;
6731 node_to_cpumask_ptr(nodemask
, node
);
6735 nodes_clear(used_nodes
);
6737 cpus_or(*span
, *span
, *nodemask
);
6738 node_set(node
, used_nodes
);
6740 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6741 int next_node
= find_next_best_node(node
, &used_nodes
);
6743 node_to_cpumask_ptr_next(nodemask
, next_node
);
6744 cpus_or(*span
, *span
, *nodemask
);
6747 #endif /* CONFIG_NUMA */
6749 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6752 * SMT sched-domains:
6754 #ifdef CONFIG_SCHED_SMT
6755 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
6756 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
6759 cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
6763 *sg
= &per_cpu(sched_group_cpus
, cpu
);
6766 #endif /* CONFIG_SCHED_SMT */
6769 * multi-core sched-domains:
6771 #ifdef CONFIG_SCHED_MC
6772 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
6773 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
6774 #endif /* CONFIG_SCHED_MC */
6776 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6778 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
6783 *mask
= per_cpu(cpu_sibling_map
, cpu
);
6784 cpus_and(*mask
, *mask
, *cpu_map
);
6785 group
= first_cpu(*mask
);
6787 *sg
= &per_cpu(sched_group_core
, group
);
6790 #elif defined(CONFIG_SCHED_MC)
6792 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
6796 *sg
= &per_cpu(sched_group_core
, cpu
);
6801 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
6802 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
6805 cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
6809 #ifdef CONFIG_SCHED_MC
6810 *mask
= cpu_coregroup_map(cpu
);
6811 cpus_and(*mask
, *mask
, *cpu_map
);
6812 group
= first_cpu(*mask
);
6813 #elif defined(CONFIG_SCHED_SMT)
6814 *mask
= per_cpu(cpu_sibling_map
, cpu
);
6815 cpus_and(*mask
, *mask
, *cpu_map
);
6816 group
= first_cpu(*mask
);
6821 *sg
= &per_cpu(sched_group_phys
, group
);
6827 * The init_sched_build_groups can't handle what we want to do with node
6828 * groups, so roll our own. Now each node has its own list of groups which
6829 * gets dynamically allocated.
6831 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
6832 static struct sched_group
***sched_group_nodes_bycpu
;
6834 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
6835 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
6837 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
6838 struct sched_group
**sg
, cpumask_t
*nodemask
)
6842 *nodemask
= node_to_cpumask(cpu_to_node(cpu
));
6843 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
6844 group
= first_cpu(*nodemask
);
6847 *sg
= &per_cpu(sched_group_allnodes
, group
);
6851 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6853 struct sched_group
*sg
= group_head
;
6859 for_each_cpu_mask(j
, sg
->cpumask
) {
6860 struct sched_domain
*sd
;
6862 sd
= &per_cpu(phys_domains
, j
);
6863 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
6865 * Only add "power" once for each
6871 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
6874 } while (sg
!= group_head
);
6876 #endif /* CONFIG_NUMA */
6879 /* Free memory allocated for various sched_group structures */
6880 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
6884 for_each_cpu_mask(cpu
, *cpu_map
) {
6885 struct sched_group
**sched_group_nodes
6886 = sched_group_nodes_bycpu
[cpu
];
6888 if (!sched_group_nodes
)
6891 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6892 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6894 *nodemask
= node_to_cpumask(i
);
6895 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
6896 if (cpus_empty(*nodemask
))
6906 if (oldsg
!= sched_group_nodes
[i
])
6909 kfree(sched_group_nodes
);
6910 sched_group_nodes_bycpu
[cpu
] = NULL
;
6913 #else /* !CONFIG_NUMA */
6914 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
6917 #endif /* CONFIG_NUMA */
6920 * Initialize sched groups cpu_power.
6922 * cpu_power indicates the capacity of sched group, which is used while
6923 * distributing the load between different sched groups in a sched domain.
6924 * Typically cpu_power for all the groups in a sched domain will be same unless
6925 * there are asymmetries in the topology. If there are asymmetries, group
6926 * having more cpu_power will pickup more load compared to the group having
6929 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6930 * the maximum number of tasks a group can handle in the presence of other idle
6931 * or lightly loaded groups in the same sched domain.
6933 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6935 struct sched_domain
*child
;
6936 struct sched_group
*group
;
6938 WARN_ON(!sd
|| !sd
->groups
);
6940 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
6945 sd
->groups
->__cpu_power
= 0;
6948 * For perf policy, if the groups in child domain share resources
6949 * (for example cores sharing some portions of the cache hierarchy
6950 * or SMT), then set this domain groups cpu_power such that each group
6951 * can handle only one task, when there are other idle groups in the
6952 * same sched domain.
6954 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
6956 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
6957 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
6962 * add cpu_power of each child group to this groups cpu_power
6964 group
= child
->groups
;
6966 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
6967 group
= group
->next
;
6968 } while (group
!= child
->groups
);
6972 * Initializers for schedule domains
6973 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6976 #define SD_INIT(sd, type) sd_init_##type(sd)
6977 #define SD_INIT_FUNC(type) \
6978 static noinline void sd_init_##type(struct sched_domain *sd) \
6980 memset(sd, 0, sizeof(*sd)); \
6981 *sd = SD_##type##_INIT; \
6982 sd->level = SD_LV_##type; \
6987 SD_INIT_FUNC(ALLNODES
)
6990 #ifdef CONFIG_SCHED_SMT
6991 SD_INIT_FUNC(SIBLING
)
6993 #ifdef CONFIG_SCHED_MC
6998 * To minimize stack usage kmalloc room for cpumasks and share the
6999 * space as the usage in build_sched_domains() dictates. Used only
7000 * if the amount of space is significant.
7003 cpumask_t tmpmask
; /* make this one first */
7006 cpumask_t this_sibling_map
;
7007 cpumask_t this_core_map
;
7009 cpumask_t send_covered
;
7012 cpumask_t domainspan
;
7014 cpumask_t notcovered
;
7019 #define SCHED_CPUMASK_ALLOC 1
7020 #define SCHED_CPUMASK_FREE(v) kfree(v)
7021 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7023 #define SCHED_CPUMASK_ALLOC 0
7024 #define SCHED_CPUMASK_FREE(v)
7025 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7028 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7029 ((unsigned long)(a) + offsetof(struct allmasks, v))
7031 static int default_relax_domain_level
= -1;
7033 static int __init
setup_relax_domain_level(char *str
)
7037 val
= simple_strtoul(str
, NULL
, 0);
7038 if (val
< SD_LV_MAX
)
7039 default_relax_domain_level
= val
;
7043 __setup("relax_domain_level=", setup_relax_domain_level
);
7045 static void set_domain_attribute(struct sched_domain
*sd
,
7046 struct sched_domain_attr
*attr
)
7050 if (!attr
|| attr
->relax_domain_level
< 0) {
7051 if (default_relax_domain_level
< 0)
7054 request
= default_relax_domain_level
;
7056 request
= attr
->relax_domain_level
;
7057 if (request
< sd
->level
) {
7058 /* turn off idle balance on this domain */
7059 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
7061 /* turn on idle balance on this domain */
7062 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
7067 * Build sched domains for a given set of cpus and attach the sched domains
7068 * to the individual cpus
7070 static int __build_sched_domains(const cpumask_t
*cpu_map
,
7071 struct sched_domain_attr
*attr
)
7074 struct root_domain
*rd
;
7075 SCHED_CPUMASK_DECLARE(allmasks
);
7078 struct sched_group
**sched_group_nodes
= NULL
;
7079 int sd_allnodes
= 0;
7082 * Allocate the per-node list of sched groups
7084 sched_group_nodes
= kcalloc(MAX_NUMNODES
, sizeof(struct sched_group
*),
7086 if (!sched_group_nodes
) {
7087 printk(KERN_WARNING
"Can not alloc sched group node list\n");
7092 rd
= alloc_rootdomain();
7094 printk(KERN_WARNING
"Cannot alloc root domain\n");
7096 kfree(sched_group_nodes
);
7101 #if SCHED_CPUMASK_ALLOC
7102 /* get space for all scratch cpumask variables */
7103 allmasks
= kmalloc(sizeof(*allmasks
), GFP_KERNEL
);
7105 printk(KERN_WARNING
"Cannot alloc cpumask array\n");
7108 kfree(sched_group_nodes
);
7113 tmpmask
= (cpumask_t
*)allmasks
;
7117 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
7121 * Set up domains for cpus specified by the cpu_map.
7123 for_each_cpu_mask(i
, *cpu_map
) {
7124 struct sched_domain
*sd
= NULL
, *p
;
7125 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7127 *nodemask
= node_to_cpumask(cpu_to_node(i
));
7128 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7131 if (cpus_weight(*cpu_map
) >
7132 SD_NODES_PER_DOMAIN
*cpus_weight(*nodemask
)) {
7133 sd
= &per_cpu(allnodes_domains
, i
);
7134 SD_INIT(sd
, ALLNODES
);
7135 set_domain_attribute(sd
, attr
);
7136 sd
->span
= *cpu_map
;
7137 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7143 sd
= &per_cpu(node_domains
, i
);
7145 set_domain_attribute(sd
, attr
);
7146 sched_domain_node_span(cpu_to_node(i
), &sd
->span
);
7150 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7154 sd
= &per_cpu(phys_domains
, i
);
7156 set_domain_attribute(sd
, attr
);
7157 sd
->span
= *nodemask
;
7161 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7163 #ifdef CONFIG_SCHED_MC
7165 sd
= &per_cpu(core_domains
, i
);
7167 set_domain_attribute(sd
, attr
);
7168 sd
->span
= cpu_coregroup_map(i
);
7169 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7172 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7175 #ifdef CONFIG_SCHED_SMT
7177 sd
= &per_cpu(cpu_domains
, i
);
7178 SD_INIT(sd
, SIBLING
);
7179 set_domain_attribute(sd
, attr
);
7180 sd
->span
= per_cpu(cpu_sibling_map
, i
);
7181 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7184 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7188 #ifdef CONFIG_SCHED_SMT
7189 /* Set up CPU (sibling) groups */
7190 for_each_cpu_mask(i
, *cpu_map
) {
7191 SCHED_CPUMASK_VAR(this_sibling_map
, allmasks
);
7192 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7194 *this_sibling_map
= per_cpu(cpu_sibling_map
, i
);
7195 cpus_and(*this_sibling_map
, *this_sibling_map
, *cpu_map
);
7196 if (i
!= first_cpu(*this_sibling_map
))
7199 init_sched_build_groups(this_sibling_map
, cpu_map
,
7201 send_covered
, tmpmask
);
7205 #ifdef CONFIG_SCHED_MC
7206 /* Set up multi-core groups */
7207 for_each_cpu_mask(i
, *cpu_map
) {
7208 SCHED_CPUMASK_VAR(this_core_map
, allmasks
);
7209 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7211 *this_core_map
= cpu_coregroup_map(i
);
7212 cpus_and(*this_core_map
, *this_core_map
, *cpu_map
);
7213 if (i
!= first_cpu(*this_core_map
))
7216 init_sched_build_groups(this_core_map
, cpu_map
,
7218 send_covered
, tmpmask
);
7222 /* Set up physical groups */
7223 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
7224 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7225 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7227 *nodemask
= node_to_cpumask(i
);
7228 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7229 if (cpus_empty(*nodemask
))
7232 init_sched_build_groups(nodemask
, cpu_map
,
7234 send_covered
, tmpmask
);
7238 /* Set up node groups */
7240 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7242 init_sched_build_groups(cpu_map
, cpu_map
,
7243 &cpu_to_allnodes_group
,
7244 send_covered
, tmpmask
);
7247 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
7248 /* Set up node groups */
7249 struct sched_group
*sg
, *prev
;
7250 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7251 SCHED_CPUMASK_VAR(domainspan
, allmasks
);
7252 SCHED_CPUMASK_VAR(covered
, allmasks
);
7255 *nodemask
= node_to_cpumask(i
);
7256 cpus_clear(*covered
);
7258 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7259 if (cpus_empty(*nodemask
)) {
7260 sched_group_nodes
[i
] = NULL
;
7264 sched_domain_node_span(i
, domainspan
);
7265 cpus_and(*domainspan
, *domainspan
, *cpu_map
);
7267 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
7269 printk(KERN_WARNING
"Can not alloc domain group for "
7273 sched_group_nodes
[i
] = sg
;
7274 for_each_cpu_mask(j
, *nodemask
) {
7275 struct sched_domain
*sd
;
7277 sd
= &per_cpu(node_domains
, j
);
7280 sg
->__cpu_power
= 0;
7281 sg
->cpumask
= *nodemask
;
7283 cpus_or(*covered
, *covered
, *nodemask
);
7286 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
7287 SCHED_CPUMASK_VAR(notcovered
, allmasks
);
7288 int n
= (i
+ j
) % MAX_NUMNODES
;
7289 node_to_cpumask_ptr(pnodemask
, n
);
7291 cpus_complement(*notcovered
, *covered
);
7292 cpus_and(*tmpmask
, *notcovered
, *cpu_map
);
7293 cpus_and(*tmpmask
, *tmpmask
, *domainspan
);
7294 if (cpus_empty(*tmpmask
))
7297 cpus_and(*tmpmask
, *tmpmask
, *pnodemask
);
7298 if (cpus_empty(*tmpmask
))
7301 sg
= kmalloc_node(sizeof(struct sched_group
),
7305 "Can not alloc domain group for node %d\n", j
);
7308 sg
->__cpu_power
= 0;
7309 sg
->cpumask
= *tmpmask
;
7310 sg
->next
= prev
->next
;
7311 cpus_or(*covered
, *covered
, *tmpmask
);
7318 /* Calculate CPU power for physical packages and nodes */
7319 #ifdef CONFIG_SCHED_SMT
7320 for_each_cpu_mask(i
, *cpu_map
) {
7321 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
7323 init_sched_groups_power(i
, sd
);
7326 #ifdef CONFIG_SCHED_MC
7327 for_each_cpu_mask(i
, *cpu_map
) {
7328 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
7330 init_sched_groups_power(i
, sd
);
7334 for_each_cpu_mask(i
, *cpu_map
) {
7335 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
7337 init_sched_groups_power(i
, sd
);
7341 for (i
= 0; i
< MAX_NUMNODES
; i
++)
7342 init_numa_sched_groups_power(sched_group_nodes
[i
]);
7345 struct sched_group
*sg
;
7347 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
,
7349 init_numa_sched_groups_power(sg
);
7353 /* Attach the domains */
7354 for_each_cpu_mask(i
, *cpu_map
) {
7355 struct sched_domain
*sd
;
7356 #ifdef CONFIG_SCHED_SMT
7357 sd
= &per_cpu(cpu_domains
, i
);
7358 #elif defined(CONFIG_SCHED_MC)
7359 sd
= &per_cpu(core_domains
, i
);
7361 sd
= &per_cpu(phys_domains
, i
);
7363 cpu_attach_domain(sd
, rd
, i
);
7366 SCHED_CPUMASK_FREE((void *)allmasks
);
7371 free_sched_groups(cpu_map
, tmpmask
);
7372 SCHED_CPUMASK_FREE((void *)allmasks
);
7377 static int build_sched_domains(const cpumask_t
*cpu_map
)
7379 return __build_sched_domains(cpu_map
, NULL
);
7382 static cpumask_t
*doms_cur
; /* current sched domains */
7383 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7384 static struct sched_domain_attr
*dattr_cur
;
7385 /* attribues of custom domains in 'doms_cur' */
7388 * Special case: If a kmalloc of a doms_cur partition (array of
7389 * cpumask_t) fails, then fallback to a single sched domain,
7390 * as determined by the single cpumask_t fallback_doms.
7392 static cpumask_t fallback_doms
;
7394 void __attribute__((weak
)) arch_update_cpu_topology(void)
7399 * Free current domain masks.
7400 * Called after all cpus are attached to NULL domain.
7402 static void free_sched_domains(void)
7405 if (doms_cur
!= &fallback_doms
)
7407 doms_cur
= &fallback_doms
;
7411 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7412 * For now this just excludes isolated cpus, but could be used to
7413 * exclude other special cases in the future.
7415 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
7419 arch_update_cpu_topology();
7421 doms_cur
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
7423 doms_cur
= &fallback_doms
;
7424 cpus_andnot(*doms_cur
, *cpu_map
, cpu_isolated_map
);
7426 err
= build_sched_domains(doms_cur
);
7427 register_sched_domain_sysctl();
7432 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
,
7435 free_sched_groups(cpu_map
, tmpmask
);
7439 * Detach sched domains from a group of cpus specified in cpu_map
7440 * These cpus will now be attached to the NULL domain
7442 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
7447 unregister_sched_domain_sysctl();
7449 for_each_cpu_mask(i
, *cpu_map
)
7450 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7451 synchronize_sched();
7452 arch_destroy_sched_domains(cpu_map
, &tmpmask
);
7455 /* handle null as "default" */
7456 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7457 struct sched_domain_attr
*new, int idx_new
)
7459 struct sched_domain_attr tmp
;
7466 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7467 new ? (new + idx_new
) : &tmp
,
7468 sizeof(struct sched_domain_attr
));
7472 * Partition sched domains as specified by the 'ndoms_new'
7473 * cpumasks in the array doms_new[] of cpumasks. This compares
7474 * doms_new[] to the current sched domain partitioning, doms_cur[].
7475 * It destroys each deleted domain and builds each new domain.
7477 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7478 * The masks don't intersect (don't overlap.) We should setup one
7479 * sched domain for each mask. CPUs not in any of the cpumasks will
7480 * not be load balanced. If the same cpumask appears both in the
7481 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7484 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7485 * ownership of it and will kfree it when done with it. If the caller
7486 * failed the kmalloc call, then it can pass in doms_new == NULL,
7487 * and partition_sched_domains() will fallback to the single partition
7490 * Call with hotplug lock held
7492 void partition_sched_domains(int ndoms_new
, cpumask_t
*doms_new
,
7493 struct sched_domain_attr
*dattr_new
)
7497 mutex_lock(&sched_domains_mutex
);
7499 /* always unregister in case we don't destroy any domains */
7500 unregister_sched_domain_sysctl();
7502 if (doms_new
== NULL
) {
7504 doms_new
= &fallback_doms
;
7505 cpus_andnot(doms_new
[0], cpu_online_map
, cpu_isolated_map
);
7509 /* Destroy deleted domains */
7510 for (i
= 0; i
< ndoms_cur
; i
++) {
7511 for (j
= 0; j
< ndoms_new
; j
++) {
7512 if (cpus_equal(doms_cur
[i
], doms_new
[j
])
7513 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7516 /* no match - a current sched domain not in new doms_new[] */
7517 detach_destroy_domains(doms_cur
+ i
);
7522 /* Build new domains */
7523 for (i
= 0; i
< ndoms_new
; i
++) {
7524 for (j
= 0; j
< ndoms_cur
; j
++) {
7525 if (cpus_equal(doms_new
[i
], doms_cur
[j
])
7526 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7529 /* no match - add a new doms_new */
7530 __build_sched_domains(doms_new
+ i
,
7531 dattr_new
? dattr_new
+ i
: NULL
);
7536 /* Remember the new sched domains */
7537 if (doms_cur
!= &fallback_doms
)
7539 kfree(dattr_cur
); /* kfree(NULL) is safe */
7540 doms_cur
= doms_new
;
7541 dattr_cur
= dattr_new
;
7542 ndoms_cur
= ndoms_new
;
7544 register_sched_domain_sysctl();
7546 mutex_unlock(&sched_domains_mutex
);
7549 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7550 int arch_reinit_sched_domains(void)
7555 mutex_lock(&sched_domains_mutex
);
7556 detach_destroy_domains(&cpu_online_map
);
7557 free_sched_domains();
7558 err
= arch_init_sched_domains(&cpu_online_map
);
7559 mutex_unlock(&sched_domains_mutex
);
7565 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7569 if (buf
[0] != '0' && buf
[0] != '1')
7573 sched_smt_power_savings
= (buf
[0] == '1');
7575 sched_mc_power_savings
= (buf
[0] == '1');
7577 ret
= arch_reinit_sched_domains();
7579 return ret
? ret
: count
;
7582 #ifdef CONFIG_SCHED_MC
7583 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
7585 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7587 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
7588 const char *buf
, size_t count
)
7590 return sched_power_savings_store(buf
, count
, 0);
7592 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
7593 sched_mc_power_savings_store
);
7596 #ifdef CONFIG_SCHED_SMT
7597 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
7599 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7601 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
7602 const char *buf
, size_t count
)
7604 return sched_power_savings_store(buf
, count
, 1);
7606 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
7607 sched_smt_power_savings_store
);
7610 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7614 #ifdef CONFIG_SCHED_SMT
7616 err
= sysfs_create_file(&cls
->kset
.kobj
,
7617 &attr_sched_smt_power_savings
.attr
);
7619 #ifdef CONFIG_SCHED_MC
7620 if (!err
&& mc_capable())
7621 err
= sysfs_create_file(&cls
->kset
.kobj
,
7622 &attr_sched_mc_power_savings
.attr
);
7626 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7629 * Force a reinitialization of the sched domains hierarchy. The domains
7630 * and groups cannot be updated in place without racing with the balancing
7631 * code, so we temporarily attach all running cpus to the NULL domain
7632 * which will prevent rebalancing while the sched domains are recalculated.
7634 static int update_sched_domains(struct notifier_block
*nfb
,
7635 unsigned long action
, void *hcpu
)
7637 int cpu
= (int)(long)hcpu
;
7640 case CPU_DOWN_PREPARE
:
7641 case CPU_DOWN_PREPARE_FROZEN
:
7642 disable_runtime(cpu_rq(cpu
));
7644 case CPU_UP_PREPARE
:
7645 case CPU_UP_PREPARE_FROZEN
:
7646 detach_destroy_domains(&cpu_online_map
);
7647 free_sched_domains();
7651 case CPU_DOWN_FAILED
:
7652 case CPU_DOWN_FAILED_FROZEN
:
7654 case CPU_ONLINE_FROZEN
:
7655 enable_runtime(cpu_rq(cpu
));
7657 case CPU_UP_CANCELED
:
7658 case CPU_UP_CANCELED_FROZEN
:
7660 case CPU_DEAD_FROZEN
:
7662 * Fall through and re-initialise the domains.
7669 #ifndef CONFIG_CPUSETS
7671 * Create default domain partitioning if cpusets are disabled.
7672 * Otherwise we let cpusets rebuild the domains based on the
7676 /* The hotplug lock is already held by cpu_up/cpu_down */
7677 arch_init_sched_domains(&cpu_online_map
);
7683 void __init
sched_init_smp(void)
7685 cpumask_t non_isolated_cpus
;
7687 #if defined(CONFIG_NUMA)
7688 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
7690 BUG_ON(sched_group_nodes_bycpu
== NULL
);
7693 mutex_lock(&sched_domains_mutex
);
7694 arch_init_sched_domains(&cpu_online_map
);
7695 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
7696 if (cpus_empty(non_isolated_cpus
))
7697 cpu_set(smp_processor_id(), non_isolated_cpus
);
7698 mutex_unlock(&sched_domains_mutex
);
7700 /* XXX: Theoretical race here - CPU may be hotplugged now */
7701 hotcpu_notifier(update_sched_domains
, 0);
7704 /* Move init over to a non-isolated CPU */
7705 if (set_cpus_allowed_ptr(current
, &non_isolated_cpus
) < 0)
7707 sched_init_granularity();
7710 void __init
sched_init_smp(void)
7712 sched_init_granularity();
7714 #endif /* CONFIG_SMP */
7716 int in_sched_functions(unsigned long addr
)
7718 return in_lock_functions(addr
) ||
7719 (addr
>= (unsigned long)__sched_text_start
7720 && addr
< (unsigned long)__sched_text_end
);
7723 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7725 cfs_rq
->tasks_timeline
= RB_ROOT
;
7726 INIT_LIST_HEAD(&cfs_rq
->tasks
);
7727 #ifdef CONFIG_FAIR_GROUP_SCHED
7730 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7733 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7735 struct rt_prio_array
*array
;
7738 array
= &rt_rq
->active
;
7739 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
7740 INIT_LIST_HEAD(array
->queue
+ i
);
7741 __clear_bit(i
, array
->bitmap
);
7743 /* delimiter for bitsearch: */
7744 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
7746 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7747 rt_rq
->highest_prio
= MAX_RT_PRIO
;
7750 rt_rq
->rt_nr_migratory
= 0;
7751 rt_rq
->overloaded
= 0;
7755 rt_rq
->rt_throttled
= 0;
7756 rt_rq
->rt_runtime
= 0;
7757 spin_lock_init(&rt_rq
->rt_runtime_lock
);
7759 #ifdef CONFIG_RT_GROUP_SCHED
7760 rt_rq
->rt_nr_boosted
= 0;
7765 #ifdef CONFIG_FAIR_GROUP_SCHED
7766 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7767 struct sched_entity
*se
, int cpu
, int add
,
7768 struct sched_entity
*parent
)
7770 struct rq
*rq
= cpu_rq(cpu
);
7771 tg
->cfs_rq
[cpu
] = cfs_rq
;
7772 init_cfs_rq(cfs_rq
, rq
);
7775 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7778 /* se could be NULL for init_task_group */
7783 se
->cfs_rq
= &rq
->cfs
;
7785 se
->cfs_rq
= parent
->my_q
;
7788 se
->load
.weight
= tg
->shares
;
7789 se
->load
.inv_weight
= 0;
7790 se
->parent
= parent
;
7794 #ifdef CONFIG_RT_GROUP_SCHED
7795 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
7796 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
7797 struct sched_rt_entity
*parent
)
7799 struct rq
*rq
= cpu_rq(cpu
);
7801 tg
->rt_rq
[cpu
] = rt_rq
;
7802 init_rt_rq(rt_rq
, rq
);
7804 rt_rq
->rt_se
= rt_se
;
7805 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7807 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
7809 tg
->rt_se
[cpu
] = rt_se
;
7814 rt_se
->rt_rq
= &rq
->rt
;
7816 rt_se
->rt_rq
= parent
->my_q
;
7818 rt_se
->my_q
= rt_rq
;
7819 rt_se
->parent
= parent
;
7820 INIT_LIST_HEAD(&rt_se
->run_list
);
7824 void __init
sched_init(void)
7827 unsigned long alloc_size
= 0, ptr
;
7829 #ifdef CONFIG_FAIR_GROUP_SCHED
7830 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7832 #ifdef CONFIG_RT_GROUP_SCHED
7833 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7835 #ifdef CONFIG_USER_SCHED
7839 * As sched_init() is called before page_alloc is setup,
7840 * we use alloc_bootmem().
7843 ptr
= (unsigned long)alloc_bootmem(alloc_size
);
7845 #ifdef CONFIG_FAIR_GROUP_SCHED
7846 init_task_group
.se
= (struct sched_entity
**)ptr
;
7847 ptr
+= nr_cpu_ids
* sizeof(void **);
7849 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7850 ptr
+= nr_cpu_ids
* sizeof(void **);
7852 #ifdef CONFIG_USER_SCHED
7853 root_task_group
.se
= (struct sched_entity
**)ptr
;
7854 ptr
+= nr_cpu_ids
* sizeof(void **);
7856 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7857 ptr
+= nr_cpu_ids
* sizeof(void **);
7858 #endif /* CONFIG_USER_SCHED */
7859 #endif /* CONFIG_FAIR_GROUP_SCHED */
7860 #ifdef CONFIG_RT_GROUP_SCHED
7861 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7862 ptr
+= nr_cpu_ids
* sizeof(void **);
7864 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7865 ptr
+= nr_cpu_ids
* sizeof(void **);
7867 #ifdef CONFIG_USER_SCHED
7868 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7869 ptr
+= nr_cpu_ids
* sizeof(void **);
7871 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7872 ptr
+= nr_cpu_ids
* sizeof(void **);
7873 #endif /* CONFIG_USER_SCHED */
7874 #endif /* CONFIG_RT_GROUP_SCHED */
7878 init_defrootdomain();
7881 init_rt_bandwidth(&def_rt_bandwidth
,
7882 global_rt_period(), global_rt_runtime());
7884 #ifdef CONFIG_RT_GROUP_SCHED
7885 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
7886 global_rt_period(), global_rt_runtime());
7887 #ifdef CONFIG_USER_SCHED
7888 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
7889 global_rt_period(), RUNTIME_INF
);
7890 #endif /* CONFIG_USER_SCHED */
7891 #endif /* CONFIG_RT_GROUP_SCHED */
7893 #ifdef CONFIG_GROUP_SCHED
7894 list_add(&init_task_group
.list
, &task_groups
);
7895 INIT_LIST_HEAD(&init_task_group
.children
);
7897 #ifdef CONFIG_USER_SCHED
7898 INIT_LIST_HEAD(&root_task_group
.children
);
7899 init_task_group
.parent
= &root_task_group
;
7900 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
7901 #endif /* CONFIG_USER_SCHED */
7902 #endif /* CONFIG_GROUP_SCHED */
7904 for_each_possible_cpu(i
) {
7908 spin_lock_init(&rq
->lock
);
7909 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
7911 init_cfs_rq(&rq
->cfs
, rq
);
7912 init_rt_rq(&rq
->rt
, rq
);
7913 #ifdef CONFIG_FAIR_GROUP_SCHED
7914 init_task_group
.shares
= init_task_group_load
;
7915 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7916 #ifdef CONFIG_CGROUP_SCHED
7918 * How much cpu bandwidth does init_task_group get?
7920 * In case of task-groups formed thr' the cgroup filesystem, it
7921 * gets 100% of the cpu resources in the system. This overall
7922 * system cpu resource is divided among the tasks of
7923 * init_task_group and its child task-groups in a fair manner,
7924 * based on each entity's (task or task-group's) weight
7925 * (se->load.weight).
7927 * In other words, if init_task_group has 10 tasks of weight
7928 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7929 * then A0's share of the cpu resource is:
7931 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7933 * We achieve this by letting init_task_group's tasks sit
7934 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7936 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
7937 #elif defined CONFIG_USER_SCHED
7938 root_task_group
.shares
= NICE_0_LOAD
;
7939 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
7941 * In case of task-groups formed thr' the user id of tasks,
7942 * init_task_group represents tasks belonging to root user.
7943 * Hence it forms a sibling of all subsequent groups formed.
7944 * In this case, init_task_group gets only a fraction of overall
7945 * system cpu resource, based on the weight assigned to root
7946 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
7947 * by letting tasks of init_task_group sit in a separate cfs_rq
7948 * (init_cfs_rq) and having one entity represent this group of
7949 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
7951 init_tg_cfs_entry(&init_task_group
,
7952 &per_cpu(init_cfs_rq
, i
),
7953 &per_cpu(init_sched_entity
, i
), i
, 1,
7954 root_task_group
.se
[i
]);
7957 #endif /* CONFIG_FAIR_GROUP_SCHED */
7959 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7960 #ifdef CONFIG_RT_GROUP_SCHED
7961 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
7962 #ifdef CONFIG_CGROUP_SCHED
7963 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
7964 #elif defined CONFIG_USER_SCHED
7965 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
7966 init_tg_rt_entry(&init_task_group
,
7967 &per_cpu(init_rt_rq
, i
),
7968 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
7969 root_task_group
.rt_se
[i
]);
7973 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7974 rq
->cpu_load
[j
] = 0;
7978 rq
->active_balance
= 0;
7979 rq
->next_balance
= jiffies
;
7983 rq
->migration_thread
= NULL
;
7984 INIT_LIST_HEAD(&rq
->migration_queue
);
7985 rq_attach_root(rq
, &def_root_domain
);
7988 atomic_set(&rq
->nr_iowait
, 0);
7991 set_load_weight(&init_task
);
7993 #ifdef CONFIG_PREEMPT_NOTIFIERS
7994 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7998 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
8001 #ifdef CONFIG_RT_MUTEXES
8002 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
8006 * The boot idle thread does lazy MMU switching as well:
8008 atomic_inc(&init_mm
.mm_count
);
8009 enter_lazy_tlb(&init_mm
, current
);
8012 * Make us the idle thread. Technically, schedule() should not be
8013 * called from this thread, however somewhere below it might be,
8014 * but because we are the idle thread, we just pick up running again
8015 * when this runqueue becomes "idle".
8017 init_idle(current
, smp_processor_id());
8019 * During early bootup we pretend to be a normal task:
8021 current
->sched_class
= &fair_sched_class
;
8023 scheduler_running
= 1;
8026 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8027 void __might_sleep(char *file
, int line
)
8030 static unsigned long prev_jiffy
; /* ratelimiting */
8032 if ((in_atomic() || irqs_disabled()) &&
8033 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
8034 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
8036 prev_jiffy
= jiffies
;
8037 printk(KERN_ERR
"BUG: sleeping function called from invalid"
8038 " context at %s:%d\n", file
, line
);
8039 printk("in_atomic():%d, irqs_disabled():%d\n",
8040 in_atomic(), irqs_disabled());
8041 debug_show_held_locks(current
);
8042 if (irqs_disabled())
8043 print_irqtrace_events(current
);
8048 EXPORT_SYMBOL(__might_sleep
);
8051 #ifdef CONFIG_MAGIC_SYSRQ
8052 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
8056 update_rq_clock(rq
);
8057 on_rq
= p
->se
.on_rq
;
8059 deactivate_task(rq
, p
, 0);
8060 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
8062 activate_task(rq
, p
, 0);
8063 resched_task(rq
->curr
);
8067 void normalize_rt_tasks(void)
8069 struct task_struct
*g
, *p
;
8070 unsigned long flags
;
8073 read_lock_irqsave(&tasklist_lock
, flags
);
8074 do_each_thread(g
, p
) {
8076 * Only normalize user tasks:
8081 p
->se
.exec_start
= 0;
8082 #ifdef CONFIG_SCHEDSTATS
8083 p
->se
.wait_start
= 0;
8084 p
->se
.sleep_start
= 0;
8085 p
->se
.block_start
= 0;
8090 * Renice negative nice level userspace
8093 if (TASK_NICE(p
) < 0 && p
->mm
)
8094 set_user_nice(p
, 0);
8098 spin_lock(&p
->pi_lock
);
8099 rq
= __task_rq_lock(p
);
8101 normalize_task(rq
, p
);
8103 __task_rq_unlock(rq
);
8104 spin_unlock(&p
->pi_lock
);
8105 } while_each_thread(g
, p
);
8107 read_unlock_irqrestore(&tasklist_lock
, flags
);
8110 #endif /* CONFIG_MAGIC_SYSRQ */
8114 * These functions are only useful for the IA64 MCA handling.
8116 * They can only be called when the whole system has been
8117 * stopped - every CPU needs to be quiescent, and no scheduling
8118 * activity can take place. Using them for anything else would
8119 * be a serious bug, and as a result, they aren't even visible
8120 * under any other configuration.
8124 * curr_task - return the current task for a given cpu.
8125 * @cpu: the processor in question.
8127 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8129 struct task_struct
*curr_task(int cpu
)
8131 return cpu_curr(cpu
);
8135 * set_curr_task - set the current task for a given cpu.
8136 * @cpu: the processor in question.
8137 * @p: the task pointer to set.
8139 * Description: This function must only be used when non-maskable interrupts
8140 * are serviced on a separate stack. It allows the architecture to switch the
8141 * notion of the current task on a cpu in a non-blocking manner. This function
8142 * must be called with all CPU's synchronized, and interrupts disabled, the
8143 * and caller must save the original value of the current task (see
8144 * curr_task() above) and restore that value before reenabling interrupts and
8145 * re-starting the system.
8147 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8149 void set_curr_task(int cpu
, struct task_struct
*p
)
8156 #ifdef CONFIG_FAIR_GROUP_SCHED
8157 static void free_fair_sched_group(struct task_group
*tg
)
8161 for_each_possible_cpu(i
) {
8163 kfree(tg
->cfs_rq
[i
]);
8173 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8175 struct cfs_rq
*cfs_rq
;
8176 struct sched_entity
*se
, *parent_se
;
8180 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8183 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8187 tg
->shares
= NICE_0_LOAD
;
8189 for_each_possible_cpu(i
) {
8192 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
),
8193 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8197 se
= kmalloc_node(sizeof(struct sched_entity
),
8198 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8202 parent_se
= parent
? parent
->se
[i
] : NULL
;
8203 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent_se
);
8212 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8214 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
8215 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
8218 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8220 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
8222 #else /* !CONFG_FAIR_GROUP_SCHED */
8223 static inline void free_fair_sched_group(struct task_group
*tg
)
8228 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8233 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8237 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8240 #endif /* CONFIG_FAIR_GROUP_SCHED */
8242 #ifdef CONFIG_RT_GROUP_SCHED
8243 static void free_rt_sched_group(struct task_group
*tg
)
8247 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8249 for_each_possible_cpu(i
) {
8251 kfree(tg
->rt_rq
[i
]);
8253 kfree(tg
->rt_se
[i
]);
8261 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8263 struct rt_rq
*rt_rq
;
8264 struct sched_rt_entity
*rt_se
, *parent_se
;
8268 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8271 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8275 init_rt_bandwidth(&tg
->rt_bandwidth
,
8276 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8278 for_each_possible_cpu(i
) {
8281 rt_rq
= kmalloc_node(sizeof(struct rt_rq
),
8282 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8286 rt_se
= kmalloc_node(sizeof(struct sched_rt_entity
),
8287 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8291 parent_se
= parent
? parent
->rt_se
[i
] : NULL
;
8292 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent_se
);
8301 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8303 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8304 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8307 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8309 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8311 #else /* !CONFIG_RT_GROUP_SCHED */
8312 static inline void free_rt_sched_group(struct task_group
*tg
)
8317 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8322 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8326 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8329 #endif /* CONFIG_RT_GROUP_SCHED */
8331 #ifdef CONFIG_GROUP_SCHED
8332 static void free_sched_group(struct task_group
*tg
)
8334 free_fair_sched_group(tg
);
8335 free_rt_sched_group(tg
);
8339 /* allocate runqueue etc for a new task group */
8340 struct task_group
*sched_create_group(struct task_group
*parent
)
8342 struct task_group
*tg
;
8343 unsigned long flags
;
8346 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8348 return ERR_PTR(-ENOMEM
);
8350 if (!alloc_fair_sched_group(tg
, parent
))
8353 if (!alloc_rt_sched_group(tg
, parent
))
8356 spin_lock_irqsave(&task_group_lock
, flags
);
8357 for_each_possible_cpu(i
) {
8358 register_fair_sched_group(tg
, i
);
8359 register_rt_sched_group(tg
, i
);
8361 list_add_rcu(&tg
->list
, &task_groups
);
8363 WARN_ON(!parent
); /* root should already exist */
8365 tg
->parent
= parent
;
8366 list_add_rcu(&tg
->siblings
, &parent
->children
);
8367 INIT_LIST_HEAD(&tg
->children
);
8368 spin_unlock_irqrestore(&task_group_lock
, flags
);
8373 free_sched_group(tg
);
8374 return ERR_PTR(-ENOMEM
);
8377 /* rcu callback to free various structures associated with a task group */
8378 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8380 /* now it should be safe to free those cfs_rqs */
8381 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8384 /* Destroy runqueue etc associated with a task group */
8385 void sched_destroy_group(struct task_group
*tg
)
8387 unsigned long flags
;
8390 spin_lock_irqsave(&task_group_lock
, flags
);
8391 for_each_possible_cpu(i
) {
8392 unregister_fair_sched_group(tg
, i
);
8393 unregister_rt_sched_group(tg
, i
);
8395 list_del_rcu(&tg
->list
);
8396 list_del_rcu(&tg
->siblings
);
8397 spin_unlock_irqrestore(&task_group_lock
, flags
);
8399 /* wait for possible concurrent references to cfs_rqs complete */
8400 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8403 /* change task's runqueue when it moves between groups.
8404 * The caller of this function should have put the task in its new group
8405 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8406 * reflect its new group.
8408 void sched_move_task(struct task_struct
*tsk
)
8411 unsigned long flags
;
8414 rq
= task_rq_lock(tsk
, &flags
);
8416 update_rq_clock(rq
);
8418 running
= task_current(rq
, tsk
);
8419 on_rq
= tsk
->se
.on_rq
;
8422 dequeue_task(rq
, tsk
, 0);
8423 if (unlikely(running
))
8424 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8426 set_task_rq(tsk
, task_cpu(tsk
));
8428 #ifdef CONFIG_FAIR_GROUP_SCHED
8429 if (tsk
->sched_class
->moved_group
)
8430 tsk
->sched_class
->moved_group(tsk
);
8433 if (unlikely(running
))
8434 tsk
->sched_class
->set_curr_task(rq
);
8436 enqueue_task(rq
, tsk
, 0);
8438 task_rq_unlock(rq
, &flags
);
8440 #endif /* CONFIG_GROUP_SCHED */
8442 #ifdef CONFIG_FAIR_GROUP_SCHED
8443 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8445 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8450 dequeue_entity(cfs_rq
, se
, 0);
8452 se
->load
.weight
= shares
;
8453 se
->load
.inv_weight
= 0;
8456 enqueue_entity(cfs_rq
, se
, 0);
8459 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8461 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8462 struct rq
*rq
= cfs_rq
->rq
;
8463 unsigned long flags
;
8465 spin_lock_irqsave(&rq
->lock
, flags
);
8466 __set_se_shares(se
, shares
);
8467 spin_unlock_irqrestore(&rq
->lock
, flags
);
8470 static DEFINE_MUTEX(shares_mutex
);
8472 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8475 unsigned long flags
;
8478 * We can't change the weight of the root cgroup.
8483 if (shares
< MIN_SHARES
)
8484 shares
= MIN_SHARES
;
8485 else if (shares
> MAX_SHARES
)
8486 shares
= MAX_SHARES
;
8488 mutex_lock(&shares_mutex
);
8489 if (tg
->shares
== shares
)
8492 spin_lock_irqsave(&task_group_lock
, flags
);
8493 for_each_possible_cpu(i
)
8494 unregister_fair_sched_group(tg
, i
);
8495 list_del_rcu(&tg
->siblings
);
8496 spin_unlock_irqrestore(&task_group_lock
, flags
);
8498 /* wait for any ongoing reference to this group to finish */
8499 synchronize_sched();
8502 * Now we are free to modify the group's share on each cpu
8503 * w/o tripping rebalance_share or load_balance_fair.
8505 tg
->shares
= shares
;
8506 for_each_possible_cpu(i
) {
8510 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
8511 set_se_shares(tg
->se
[i
], shares
);
8515 * Enable load balance activity on this group, by inserting it back on
8516 * each cpu's rq->leaf_cfs_rq_list.
8518 spin_lock_irqsave(&task_group_lock
, flags
);
8519 for_each_possible_cpu(i
)
8520 register_fair_sched_group(tg
, i
);
8521 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
8522 spin_unlock_irqrestore(&task_group_lock
, flags
);
8524 mutex_unlock(&shares_mutex
);
8528 unsigned long sched_group_shares(struct task_group
*tg
)
8534 #ifdef CONFIG_RT_GROUP_SCHED
8536 * Ensure that the real time constraints are schedulable.
8538 static DEFINE_MUTEX(rt_constraints_mutex
);
8540 static unsigned long to_ratio(u64 period
, u64 runtime
)
8542 if (runtime
== RUNTIME_INF
)
8545 return div64_u64(runtime
<< 16, period
);
8548 #ifdef CONFIG_CGROUP_SCHED
8549 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8551 struct task_group
*tgi
, *parent
= tg
->parent
;
8552 unsigned long total
= 0;
8555 if (global_rt_period() < period
)
8558 return to_ratio(period
, runtime
) <
8559 to_ratio(global_rt_period(), global_rt_runtime());
8562 if (ktime_to_ns(parent
->rt_bandwidth
.rt_period
) < period
)
8566 list_for_each_entry_rcu(tgi
, &parent
->children
, siblings
) {
8570 total
+= to_ratio(ktime_to_ns(tgi
->rt_bandwidth
.rt_period
),
8571 tgi
->rt_bandwidth
.rt_runtime
);
8575 return total
+ to_ratio(period
, runtime
) <=
8576 to_ratio(ktime_to_ns(parent
->rt_bandwidth
.rt_period
),
8577 parent
->rt_bandwidth
.rt_runtime
);
8579 #elif defined CONFIG_USER_SCHED
8580 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8582 struct task_group
*tgi
;
8583 unsigned long total
= 0;
8584 unsigned long global_ratio
=
8585 to_ratio(global_rt_period(), global_rt_runtime());
8588 list_for_each_entry_rcu(tgi
, &task_groups
, list
) {
8592 total
+= to_ratio(ktime_to_ns(tgi
->rt_bandwidth
.rt_period
),
8593 tgi
->rt_bandwidth
.rt_runtime
);
8597 return total
+ to_ratio(period
, runtime
) < global_ratio
;
8601 /* Must be called with tasklist_lock held */
8602 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8604 struct task_struct
*g
, *p
;
8605 do_each_thread(g
, p
) {
8606 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8608 } while_each_thread(g
, p
);
8612 static int tg_set_bandwidth(struct task_group
*tg
,
8613 u64 rt_period
, u64 rt_runtime
)
8617 mutex_lock(&rt_constraints_mutex
);
8618 read_lock(&tasklist_lock
);
8619 if (rt_runtime
== 0 && tg_has_rt_tasks(tg
)) {
8623 if (!__rt_schedulable(tg
, rt_period
, rt_runtime
)) {
8628 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8629 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8630 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8632 for_each_possible_cpu(i
) {
8633 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8635 spin_lock(&rt_rq
->rt_runtime_lock
);
8636 rt_rq
->rt_runtime
= rt_runtime
;
8637 spin_unlock(&rt_rq
->rt_runtime_lock
);
8639 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8641 read_unlock(&tasklist_lock
);
8642 mutex_unlock(&rt_constraints_mutex
);
8647 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8649 u64 rt_runtime
, rt_period
;
8651 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8652 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8653 if (rt_runtime_us
< 0)
8654 rt_runtime
= RUNTIME_INF
;
8656 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8659 long sched_group_rt_runtime(struct task_group
*tg
)
8663 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8666 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8667 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8668 return rt_runtime_us
;
8671 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8673 u64 rt_runtime
, rt_period
;
8675 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8676 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8678 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8681 long sched_group_rt_period(struct task_group
*tg
)
8685 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8686 do_div(rt_period_us
, NSEC_PER_USEC
);
8687 return rt_period_us
;
8690 static int sched_rt_global_constraints(void)
8692 struct task_group
*tg
= &root_task_group
;
8693 u64 rt_runtime
, rt_period
;
8696 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8697 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8699 mutex_lock(&rt_constraints_mutex
);
8700 if (!__rt_schedulable(tg
, rt_period
, rt_runtime
))
8702 mutex_unlock(&rt_constraints_mutex
);
8706 #else /* !CONFIG_RT_GROUP_SCHED */
8707 static int sched_rt_global_constraints(void)
8709 unsigned long flags
;
8712 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8713 for_each_possible_cpu(i
) {
8714 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
8716 spin_lock(&rt_rq
->rt_runtime_lock
);
8717 rt_rq
->rt_runtime
= global_rt_runtime();
8718 spin_unlock(&rt_rq
->rt_runtime_lock
);
8720 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8724 #endif /* CONFIG_RT_GROUP_SCHED */
8726 int sched_rt_handler(struct ctl_table
*table
, int write
,
8727 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
8731 int old_period
, old_runtime
;
8732 static DEFINE_MUTEX(mutex
);
8735 old_period
= sysctl_sched_rt_period
;
8736 old_runtime
= sysctl_sched_rt_runtime
;
8738 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
8740 if (!ret
&& write
) {
8741 ret
= sched_rt_global_constraints();
8743 sysctl_sched_rt_period
= old_period
;
8744 sysctl_sched_rt_runtime
= old_runtime
;
8746 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8747 def_rt_bandwidth
.rt_period
=
8748 ns_to_ktime(global_rt_period());
8751 mutex_unlock(&mutex
);
8756 #ifdef CONFIG_CGROUP_SCHED
8758 /* return corresponding task_group object of a cgroup */
8759 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
8761 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
8762 struct task_group
, css
);
8765 static struct cgroup_subsys_state
*
8766 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8768 struct task_group
*tg
, *parent
;
8770 if (!cgrp
->parent
) {
8771 /* This is early initialization for the top cgroup */
8772 init_task_group
.css
.cgroup
= cgrp
;
8773 return &init_task_group
.css
;
8776 parent
= cgroup_tg(cgrp
->parent
);
8777 tg
= sched_create_group(parent
);
8779 return ERR_PTR(-ENOMEM
);
8781 /* Bind the cgroup to task_group object we just created */
8782 tg
->css
.cgroup
= cgrp
;
8788 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8790 struct task_group
*tg
= cgroup_tg(cgrp
);
8792 sched_destroy_group(tg
);
8796 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8797 struct task_struct
*tsk
)
8799 #ifdef CONFIG_RT_GROUP_SCHED
8800 /* Don't accept realtime tasks when there is no way for them to run */
8801 if (rt_task(tsk
) && cgroup_tg(cgrp
)->rt_bandwidth
.rt_runtime
== 0)
8804 /* We don't support RT-tasks being in separate groups */
8805 if (tsk
->sched_class
!= &fair_sched_class
)
8813 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8814 struct cgroup
*old_cont
, struct task_struct
*tsk
)
8816 sched_move_task(tsk
);
8819 #ifdef CONFIG_FAIR_GROUP_SCHED
8820 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
8823 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
8826 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
8828 struct task_group
*tg
= cgroup_tg(cgrp
);
8830 return (u64
) tg
->shares
;
8832 #endif /* CONFIG_FAIR_GROUP_SCHED */
8834 #ifdef CONFIG_RT_GROUP_SCHED
8835 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
8838 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
8841 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8843 return sched_group_rt_runtime(cgroup_tg(cgrp
));
8846 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
8849 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
8852 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
8854 return sched_group_rt_period(cgroup_tg(cgrp
));
8856 #endif /* CONFIG_RT_GROUP_SCHED */
8858 static struct cftype cpu_files
[] = {
8859 #ifdef CONFIG_FAIR_GROUP_SCHED
8862 .read_u64
= cpu_shares_read_u64
,
8863 .write_u64
= cpu_shares_write_u64
,
8866 #ifdef CONFIG_RT_GROUP_SCHED
8868 .name
= "rt_runtime_us",
8869 .read_s64
= cpu_rt_runtime_read
,
8870 .write_s64
= cpu_rt_runtime_write
,
8873 .name
= "rt_period_us",
8874 .read_u64
= cpu_rt_period_read_uint
,
8875 .write_u64
= cpu_rt_period_write_uint
,
8880 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8882 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
8885 struct cgroup_subsys cpu_cgroup_subsys
= {
8887 .create
= cpu_cgroup_create
,
8888 .destroy
= cpu_cgroup_destroy
,
8889 .can_attach
= cpu_cgroup_can_attach
,
8890 .attach
= cpu_cgroup_attach
,
8891 .populate
= cpu_cgroup_populate
,
8892 .subsys_id
= cpu_cgroup_subsys_id
,
8896 #endif /* CONFIG_CGROUP_SCHED */
8898 #ifdef CONFIG_CGROUP_CPUACCT
8901 * CPU accounting code for task groups.
8903 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8904 * (balbir@in.ibm.com).
8907 /* track cpu usage of a group of tasks */
8909 struct cgroup_subsys_state css
;
8910 /* cpuusage holds pointer to a u64-type object on every cpu */
8914 struct cgroup_subsys cpuacct_subsys
;
8916 /* return cpu accounting group corresponding to this container */
8917 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
8919 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
8920 struct cpuacct
, css
);
8923 /* return cpu accounting group to which this task belongs */
8924 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
8926 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
8927 struct cpuacct
, css
);
8930 /* create a new cpu accounting group */
8931 static struct cgroup_subsys_state
*cpuacct_create(
8932 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8934 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
8937 return ERR_PTR(-ENOMEM
);
8939 ca
->cpuusage
= alloc_percpu(u64
);
8940 if (!ca
->cpuusage
) {
8942 return ERR_PTR(-ENOMEM
);
8948 /* destroy an existing cpu accounting group */
8950 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8952 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8954 free_percpu(ca
->cpuusage
);
8958 /* return total cpu usage (in nanoseconds) of a group */
8959 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8961 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8962 u64 totalcpuusage
= 0;
8965 for_each_possible_cpu(i
) {
8966 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
8969 * Take rq->lock to make 64-bit addition safe on 32-bit
8972 spin_lock_irq(&cpu_rq(i
)->lock
);
8973 totalcpuusage
+= *cpuusage
;
8974 spin_unlock_irq(&cpu_rq(i
)->lock
);
8977 return totalcpuusage
;
8980 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
8983 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8992 for_each_possible_cpu(i
) {
8993 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
8995 spin_lock_irq(&cpu_rq(i
)->lock
);
8997 spin_unlock_irq(&cpu_rq(i
)->lock
);
9003 static struct cftype files
[] = {
9006 .read_u64
= cpuusage_read
,
9007 .write_u64
= cpuusage_write
,
9011 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9013 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
9017 * charge this task's execution time to its accounting group.
9019 * called with rq->lock held.
9021 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
9025 if (!cpuacct_subsys
.active
)
9030 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, task_cpu(tsk
));
9032 *cpuusage
+= cputime
;
9036 struct cgroup_subsys cpuacct_subsys
= {
9038 .create
= cpuacct_create
,
9039 .destroy
= cpuacct_destroy
,
9040 .populate
= cpuacct_populate
,
9041 .subsys_id
= cpuacct_subsys_id
,
9043 #endif /* CONFIG_CGROUP_CPUACCT */