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
) { }
368 static inline struct task_group
*task_group(struct task_struct
*p
)
373 #endif /* CONFIG_GROUP_SCHED */
375 /* CFS-related fields in a runqueue */
377 struct load_weight load
;
378 unsigned long nr_running
;
384 struct rb_root tasks_timeline
;
385 struct rb_node
*rb_leftmost
;
387 struct list_head tasks
;
388 struct list_head
*balance_iterator
;
391 * 'curr' points to currently running entity on this cfs_rq.
392 * It is set to NULL otherwise (i.e when none are currently running).
394 struct sched_entity
*curr
, *next
;
396 unsigned long nr_spread_over
;
398 #ifdef CONFIG_FAIR_GROUP_SCHED
399 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
402 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
403 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
404 * (like users, containers etc.)
406 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
407 * list is used during load balance.
409 struct list_head leaf_cfs_rq_list
;
410 struct task_group
*tg
; /* group that "owns" this runqueue */
414 * the part of load.weight contributed by tasks
416 unsigned long task_weight
;
419 * h_load = weight * f(tg)
421 * Where f(tg) is the recursive weight fraction assigned to
424 unsigned long h_load
;
427 * this cpu's part of tg->shares
429 unsigned long shares
;
432 * load.weight at the time we set shares
434 unsigned long rq_weight
;
439 /* Real-Time classes' related field in a runqueue: */
441 struct rt_prio_array active
;
442 unsigned long rt_nr_running
;
443 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
444 int highest_prio
; /* highest queued rt task prio */
447 unsigned long rt_nr_migratory
;
453 /* Nests inside the rq lock: */
454 spinlock_t rt_runtime_lock
;
456 #ifdef CONFIG_RT_GROUP_SCHED
457 unsigned long rt_nr_boosted
;
460 struct list_head leaf_rt_rq_list
;
461 struct task_group
*tg
;
462 struct sched_rt_entity
*rt_se
;
469 * We add the notion of a root-domain which will be used to define per-domain
470 * variables. Each exclusive cpuset essentially defines an island domain by
471 * fully partitioning the member cpus from any other cpuset. Whenever a new
472 * exclusive cpuset is created, we also create and attach a new root-domain
482 * The "RT overload" flag: it gets set if a CPU has more than
483 * one runnable RT task.
488 struct cpupri cpupri
;
493 * By default the system creates a single root-domain with all cpus as
494 * members (mimicking the global state we have today).
496 static struct root_domain def_root_domain
;
501 * This is the main, per-CPU runqueue data structure.
503 * Locking rule: those places that want to lock multiple runqueues
504 * (such as the load balancing or the thread migration code), lock
505 * acquire operations must be ordered by ascending &runqueue.
512 * nr_running and cpu_load should be in the same cacheline because
513 * remote CPUs use both these fields when doing load calculation.
515 unsigned long nr_running
;
516 #define CPU_LOAD_IDX_MAX 5
517 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
518 unsigned char idle_at_tick
;
520 unsigned long last_tick_seen
;
521 unsigned char in_nohz_recently
;
523 /* capture load from *all* tasks on this cpu: */
524 struct load_weight load
;
525 unsigned long nr_load_updates
;
531 #ifdef CONFIG_FAIR_GROUP_SCHED
532 /* list of leaf cfs_rq on this cpu: */
533 struct list_head leaf_cfs_rq_list
;
535 #ifdef CONFIG_RT_GROUP_SCHED
536 struct list_head leaf_rt_rq_list
;
540 * This is part of a global counter where only the total sum
541 * over all CPUs matters. A task can increase this counter on
542 * one CPU and if it got migrated afterwards it may decrease
543 * it on another CPU. Always updated under the runqueue lock:
545 unsigned long nr_uninterruptible
;
547 struct task_struct
*curr
, *idle
;
548 unsigned long next_balance
;
549 struct mm_struct
*prev_mm
;
556 struct root_domain
*rd
;
557 struct sched_domain
*sd
;
559 /* For active balancing */
562 /* cpu of this runqueue: */
566 unsigned long avg_load_per_task
;
568 struct task_struct
*migration_thread
;
569 struct list_head migration_queue
;
572 #ifdef CONFIG_SCHED_HRTICK
573 unsigned long hrtick_flags
;
574 ktime_t hrtick_expire
;
575 struct hrtimer hrtick_timer
;
578 #ifdef CONFIG_SCHEDSTATS
580 struct sched_info rq_sched_info
;
582 /* sys_sched_yield() stats */
583 unsigned int yld_exp_empty
;
584 unsigned int yld_act_empty
;
585 unsigned int yld_both_empty
;
586 unsigned int yld_count
;
588 /* schedule() stats */
589 unsigned int sched_switch
;
590 unsigned int sched_count
;
591 unsigned int sched_goidle
;
593 /* try_to_wake_up() stats */
594 unsigned int ttwu_count
;
595 unsigned int ttwu_local
;
598 unsigned int bkl_count
;
600 struct lock_class_key rq_lock_key
;
603 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
605 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
)
607 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
);
610 static inline int cpu_of(struct rq
*rq
)
620 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
621 * See detach_destroy_domains: synchronize_sched for details.
623 * The domain tree of any CPU may only be accessed from within
624 * preempt-disabled sections.
626 #define for_each_domain(cpu, __sd) \
627 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
629 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
630 #define this_rq() (&__get_cpu_var(runqueues))
631 #define task_rq(p) cpu_rq(task_cpu(p))
632 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
634 static inline void update_rq_clock(struct rq
*rq
)
636 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
640 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
642 #ifdef CONFIG_SCHED_DEBUG
643 # define const_debug __read_mostly
645 # define const_debug static const
649 * Debugging: various feature bits
652 #define SCHED_FEAT(name, enabled) \
653 __SCHED_FEAT_##name ,
656 #include "sched_features.h"
661 #define SCHED_FEAT(name, enabled) \
662 (1UL << __SCHED_FEAT_##name) * enabled |
664 const_debug
unsigned int sysctl_sched_features
=
665 #include "sched_features.h"
670 #ifdef CONFIG_SCHED_DEBUG
671 #define SCHED_FEAT(name, enabled) \
674 static __read_mostly
char *sched_feat_names
[] = {
675 #include "sched_features.h"
681 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
683 filp
->private_data
= inode
->i_private
;
688 sched_feat_read(struct file
*filp
, char __user
*ubuf
,
689 size_t cnt
, loff_t
*ppos
)
696 for (i
= 0; sched_feat_names
[i
]; i
++) {
697 len
+= strlen(sched_feat_names
[i
]);
701 buf
= kmalloc(len
+ 2, GFP_KERNEL
);
705 for (i
= 0; sched_feat_names
[i
]; i
++) {
706 if (sysctl_sched_features
& (1UL << i
))
707 r
+= sprintf(buf
+ r
, "%s ", sched_feat_names
[i
]);
709 r
+= sprintf(buf
+ r
, "NO_%s ", sched_feat_names
[i
]);
712 r
+= sprintf(buf
+ r
, "\n");
713 WARN_ON(r
>= len
+ 2);
715 r
= simple_read_from_buffer(ubuf
, cnt
, ppos
, buf
, r
);
723 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
724 size_t cnt
, loff_t
*ppos
)
734 if (copy_from_user(&buf
, ubuf
, cnt
))
739 if (strncmp(buf
, "NO_", 3) == 0) {
744 for (i
= 0; sched_feat_names
[i
]; i
++) {
745 int len
= strlen(sched_feat_names
[i
]);
747 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
749 sysctl_sched_features
&= ~(1UL << i
);
751 sysctl_sched_features
|= (1UL << i
);
756 if (!sched_feat_names
[i
])
764 static struct file_operations sched_feat_fops
= {
765 .open
= sched_feat_open
,
766 .read
= sched_feat_read
,
767 .write
= sched_feat_write
,
770 static __init
int sched_init_debug(void)
772 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
777 late_initcall(sched_init_debug
);
781 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
784 * Number of tasks to iterate in a single balance run.
785 * Limited because this is done with IRQs disabled.
787 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
790 * ratelimit for updating the group shares.
793 const_debug
unsigned int sysctl_sched_shares_ratelimit
= 500000;
796 * period over which we measure -rt task cpu usage in us.
799 unsigned int sysctl_sched_rt_period
= 1000000;
801 static __read_mostly
int scheduler_running
;
804 * part of the period that we allow rt tasks to run in us.
807 int sysctl_sched_rt_runtime
= 950000;
809 static inline u64
global_rt_period(void)
811 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
814 static inline u64
global_rt_runtime(void)
816 if (sysctl_sched_rt_period
< 0)
819 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
822 #ifndef prepare_arch_switch
823 # define prepare_arch_switch(next) do { } while (0)
825 #ifndef finish_arch_switch
826 # define finish_arch_switch(prev) do { } while (0)
829 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
831 return rq
->curr
== p
;
834 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
835 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
837 return task_current(rq
, p
);
840 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
844 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
846 #ifdef CONFIG_DEBUG_SPINLOCK
847 /* this is a valid case when another task releases the spinlock */
848 rq
->lock
.owner
= current
;
851 * If we are tracking spinlock dependencies then we have to
852 * fix up the runqueue lock - which gets 'carried over' from
855 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
857 spin_unlock_irq(&rq
->lock
);
860 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
861 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
866 return task_current(rq
, p
);
870 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
874 * We can optimise this out completely for !SMP, because the
875 * SMP rebalancing from interrupt is the only thing that cares
880 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
881 spin_unlock_irq(&rq
->lock
);
883 spin_unlock(&rq
->lock
);
887 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
891 * After ->oncpu is cleared, the task can be moved to a different CPU.
892 * We must ensure this doesn't happen until the switch is completely
898 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
902 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
905 * __task_rq_lock - lock the runqueue a given task resides on.
906 * Must be called interrupts disabled.
908 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
912 struct rq
*rq
= task_rq(p
);
913 spin_lock(&rq
->lock
);
914 if (likely(rq
== task_rq(p
)))
916 spin_unlock(&rq
->lock
);
921 * task_rq_lock - lock the runqueue a given task resides on and disable
922 * interrupts. Note the ordering: we can safely lookup the task_rq without
923 * explicitly disabling preemption.
925 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
931 local_irq_save(*flags
);
933 spin_lock(&rq
->lock
);
934 if (likely(rq
== task_rq(p
)))
936 spin_unlock_irqrestore(&rq
->lock
, *flags
);
940 static void __task_rq_unlock(struct rq
*rq
)
943 spin_unlock(&rq
->lock
);
946 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
949 spin_unlock_irqrestore(&rq
->lock
, *flags
);
953 * this_rq_lock - lock this runqueue and disable interrupts.
955 static struct rq
*this_rq_lock(void)
962 spin_lock(&rq
->lock
);
967 static void __resched_task(struct task_struct
*p
, int tif_bit
);
969 static inline void resched_task(struct task_struct
*p
)
971 __resched_task(p
, TIF_NEED_RESCHED
);
974 #ifdef CONFIG_SCHED_HRTICK
976 * Use HR-timers to deliver accurate preemption points.
978 * Its all a bit involved since we cannot program an hrt while holding the
979 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
982 * When we get rescheduled we reprogram the hrtick_timer outside of the
985 static inline void resched_hrt(struct task_struct
*p
)
987 __resched_task(p
, TIF_HRTICK_RESCHED
);
990 static inline void resched_rq(struct rq
*rq
)
994 spin_lock_irqsave(&rq
->lock
, flags
);
995 resched_task(rq
->curr
);
996 spin_unlock_irqrestore(&rq
->lock
, flags
);
1000 HRTICK_SET
, /* re-programm hrtick_timer */
1001 HRTICK_RESET
, /* not a new slice */
1002 HRTICK_BLOCK
, /* stop hrtick operations */
1007 * - enabled by features
1008 * - hrtimer is actually high res
1010 static inline int hrtick_enabled(struct rq
*rq
)
1012 if (!sched_feat(HRTICK
))
1014 if (unlikely(test_bit(HRTICK_BLOCK
, &rq
->hrtick_flags
)))
1016 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1020 * Called to set the hrtick timer state.
1022 * called with rq->lock held and irqs disabled
1024 static void hrtick_start(struct rq
*rq
, u64 delay
, int reset
)
1026 assert_spin_locked(&rq
->lock
);
1029 * preempt at: now + delay
1032 ktime_add_ns(rq
->hrtick_timer
.base
->get_time(), delay
);
1034 * indicate we need to program the timer
1036 __set_bit(HRTICK_SET
, &rq
->hrtick_flags
);
1038 __set_bit(HRTICK_RESET
, &rq
->hrtick_flags
);
1041 * New slices are called from the schedule path and don't need a
1042 * forced reschedule.
1045 resched_hrt(rq
->curr
);
1048 static void hrtick_clear(struct rq
*rq
)
1050 if (hrtimer_active(&rq
->hrtick_timer
))
1051 hrtimer_cancel(&rq
->hrtick_timer
);
1055 * Update the timer from the possible pending state.
1057 static void hrtick_set(struct rq
*rq
)
1061 unsigned long flags
;
1063 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1065 spin_lock_irqsave(&rq
->lock
, flags
);
1066 set
= __test_and_clear_bit(HRTICK_SET
, &rq
->hrtick_flags
);
1067 reset
= __test_and_clear_bit(HRTICK_RESET
, &rq
->hrtick_flags
);
1068 time
= rq
->hrtick_expire
;
1069 clear_thread_flag(TIF_HRTICK_RESCHED
);
1070 spin_unlock_irqrestore(&rq
->lock
, flags
);
1073 hrtimer_start(&rq
->hrtick_timer
, time
, HRTIMER_MODE_ABS
);
1074 if (reset
&& !hrtimer_active(&rq
->hrtick_timer
))
1081 * High-resolution timer tick.
1082 * Runs from hardirq context with interrupts disabled.
1084 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1086 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1088 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1090 spin_lock(&rq
->lock
);
1091 update_rq_clock(rq
);
1092 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1093 spin_unlock(&rq
->lock
);
1095 return HRTIMER_NORESTART
;
1099 static void hotplug_hrtick_disable(int cpu
)
1101 struct rq
*rq
= cpu_rq(cpu
);
1102 unsigned long flags
;
1104 spin_lock_irqsave(&rq
->lock
, flags
);
1105 rq
->hrtick_flags
= 0;
1106 __set_bit(HRTICK_BLOCK
, &rq
->hrtick_flags
);
1107 spin_unlock_irqrestore(&rq
->lock
, flags
);
1112 static void hotplug_hrtick_enable(int cpu
)
1114 struct rq
*rq
= cpu_rq(cpu
);
1115 unsigned long flags
;
1117 spin_lock_irqsave(&rq
->lock
, flags
);
1118 __clear_bit(HRTICK_BLOCK
, &rq
->hrtick_flags
);
1119 spin_unlock_irqrestore(&rq
->lock
, flags
);
1123 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1125 int cpu
= (int)(long)hcpu
;
1128 case CPU_UP_CANCELED
:
1129 case CPU_UP_CANCELED_FROZEN
:
1130 case CPU_DOWN_PREPARE
:
1131 case CPU_DOWN_PREPARE_FROZEN
:
1133 case CPU_DEAD_FROZEN
:
1134 hotplug_hrtick_disable(cpu
);
1137 case CPU_UP_PREPARE
:
1138 case CPU_UP_PREPARE_FROZEN
:
1139 case CPU_DOWN_FAILED
:
1140 case CPU_DOWN_FAILED_FROZEN
:
1142 case CPU_ONLINE_FROZEN
:
1143 hotplug_hrtick_enable(cpu
);
1150 static void init_hrtick(void)
1152 hotcpu_notifier(hotplug_hrtick
, 0);
1154 #endif /* CONFIG_SMP */
1156 static void init_rq_hrtick(struct rq
*rq
)
1158 rq
->hrtick_flags
= 0;
1159 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1160 rq
->hrtick_timer
.function
= hrtick
;
1161 rq
->hrtick_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_NO_SOFTIRQ
;
1164 void hrtick_resched(void)
1167 unsigned long flags
;
1169 if (!test_thread_flag(TIF_HRTICK_RESCHED
))
1172 local_irq_save(flags
);
1173 rq
= cpu_rq(smp_processor_id());
1175 local_irq_restore(flags
);
1178 static inline void hrtick_clear(struct rq
*rq
)
1182 static inline void hrtick_set(struct rq
*rq
)
1186 static inline void init_rq_hrtick(struct rq
*rq
)
1190 void hrtick_resched(void)
1194 static inline void init_hrtick(void)
1200 * resched_task - mark a task 'to be rescheduled now'.
1202 * On UP this means the setting of the need_resched flag, on SMP it
1203 * might also involve a cross-CPU call to trigger the scheduler on
1208 #ifndef tsk_is_polling
1209 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1212 static void __resched_task(struct task_struct
*p
, int tif_bit
)
1216 assert_spin_locked(&task_rq(p
)->lock
);
1218 if (unlikely(test_tsk_thread_flag(p
, tif_bit
)))
1221 set_tsk_thread_flag(p
, tif_bit
);
1224 if (cpu
== smp_processor_id())
1227 /* NEED_RESCHED must be visible before we test polling */
1229 if (!tsk_is_polling(p
))
1230 smp_send_reschedule(cpu
);
1233 static void resched_cpu(int cpu
)
1235 struct rq
*rq
= cpu_rq(cpu
);
1236 unsigned long flags
;
1238 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1240 resched_task(cpu_curr(cpu
));
1241 spin_unlock_irqrestore(&rq
->lock
, flags
);
1246 * When add_timer_on() enqueues a timer into the timer wheel of an
1247 * idle CPU then this timer might expire before the next timer event
1248 * which is scheduled to wake up that CPU. In case of a completely
1249 * idle system the next event might even be infinite time into the
1250 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1251 * leaves the inner idle loop so the newly added timer is taken into
1252 * account when the CPU goes back to idle and evaluates the timer
1253 * wheel for the next timer event.
1255 void wake_up_idle_cpu(int cpu
)
1257 struct rq
*rq
= cpu_rq(cpu
);
1259 if (cpu
== smp_processor_id())
1263 * This is safe, as this function is called with the timer
1264 * wheel base lock of (cpu) held. When the CPU is on the way
1265 * to idle and has not yet set rq->curr to idle then it will
1266 * be serialized on the timer wheel base lock and take the new
1267 * timer into account automatically.
1269 if (rq
->curr
!= rq
->idle
)
1273 * We can set TIF_RESCHED on the idle task of the other CPU
1274 * lockless. The worst case is that the other CPU runs the
1275 * idle task through an additional NOOP schedule()
1277 set_tsk_thread_flag(rq
->idle
, TIF_NEED_RESCHED
);
1279 /* NEED_RESCHED must be visible before we test polling */
1281 if (!tsk_is_polling(rq
->idle
))
1282 smp_send_reschedule(cpu
);
1284 #endif /* CONFIG_NO_HZ */
1286 #else /* !CONFIG_SMP */
1287 static void __resched_task(struct task_struct
*p
, int tif_bit
)
1289 assert_spin_locked(&task_rq(p
)->lock
);
1290 set_tsk_thread_flag(p
, tif_bit
);
1292 #endif /* CONFIG_SMP */
1294 #if BITS_PER_LONG == 32
1295 # define WMULT_CONST (~0UL)
1297 # define WMULT_CONST (1UL << 32)
1300 #define WMULT_SHIFT 32
1303 * Shift right and round:
1305 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1308 * delta *= weight / lw
1310 static unsigned long
1311 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1312 struct load_weight
*lw
)
1316 if (!lw
->inv_weight
) {
1317 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1320 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1324 tmp
= (u64
)delta_exec
* weight
;
1326 * Check whether we'd overflow the 64-bit multiplication:
1328 if (unlikely(tmp
> WMULT_CONST
))
1329 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1332 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1334 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1337 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1343 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1350 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1351 * of tasks with abnormal "nice" values across CPUs the contribution that
1352 * each task makes to its run queue's load is weighted according to its
1353 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1354 * scaled version of the new time slice allocation that they receive on time
1358 #define WEIGHT_IDLEPRIO 2
1359 #define WMULT_IDLEPRIO (1 << 31)
1362 * Nice levels are multiplicative, with a gentle 10% change for every
1363 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1364 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1365 * that remained on nice 0.
1367 * The "10% effect" is relative and cumulative: from _any_ nice level,
1368 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1369 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1370 * If a task goes up by ~10% and another task goes down by ~10% then
1371 * the relative distance between them is ~25%.)
1373 static const int prio_to_weight
[40] = {
1374 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1375 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1376 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1377 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1378 /* 0 */ 1024, 820, 655, 526, 423,
1379 /* 5 */ 335, 272, 215, 172, 137,
1380 /* 10 */ 110, 87, 70, 56, 45,
1381 /* 15 */ 36, 29, 23, 18, 15,
1385 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1387 * In cases where the weight does not change often, we can use the
1388 * precalculated inverse to speed up arithmetics by turning divisions
1389 * into multiplications:
1391 static const u32 prio_to_wmult
[40] = {
1392 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1393 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1394 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1395 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1396 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1397 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1398 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1399 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1402 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1405 * runqueue iterator, to support SMP load-balancing between different
1406 * scheduling classes, without having to expose their internal data
1407 * structures to the load-balancing proper:
1409 struct rq_iterator
{
1411 struct task_struct
*(*start
)(void *);
1412 struct task_struct
*(*next
)(void *);
1416 static unsigned long
1417 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1418 unsigned long max_load_move
, struct sched_domain
*sd
,
1419 enum cpu_idle_type idle
, int *all_pinned
,
1420 int *this_best_prio
, struct rq_iterator
*iterator
);
1423 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1424 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1425 struct rq_iterator
*iterator
);
1428 #ifdef CONFIG_CGROUP_CPUACCT
1429 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1431 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1434 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1436 update_load_add(&rq
->load
, load
);
1439 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1441 update_load_sub(&rq
->load
, load
);
1445 static unsigned long source_load(int cpu
, int type
);
1446 static unsigned long target_load(int cpu
, int type
);
1447 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1449 static unsigned long cpu_avg_load_per_task(int cpu
)
1451 struct rq
*rq
= cpu_rq(cpu
);
1454 rq
->avg_load_per_task
= rq
->load
.weight
/ rq
->nr_running
;
1456 return rq
->avg_load_per_task
;
1459 #ifdef CONFIG_FAIR_GROUP_SCHED
1461 typedef void (*tg_visitor
)(struct task_group
*, int, struct sched_domain
*);
1464 * Iterate the full tree, calling @down when first entering a node and @up when
1465 * leaving it for the final time.
1468 walk_tg_tree(tg_visitor down
, tg_visitor up
, int cpu
, struct sched_domain
*sd
)
1470 struct task_group
*parent
, *child
;
1473 parent
= &root_task_group
;
1475 (*down
)(parent
, cpu
, sd
);
1476 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1483 (*up
)(parent
, cpu
, sd
);
1486 parent
= parent
->parent
;
1492 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1495 * Calculate and set the cpu's group shares.
1498 __update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1499 unsigned long sd_shares
, unsigned long sd_rq_weight
)
1502 unsigned long shares
;
1503 unsigned long rq_weight
;
1508 rq_weight
= tg
->cfs_rq
[cpu
]->load
.weight
;
1511 * If there are currently no tasks on the cpu pretend there is one of
1512 * average load so that when a new task gets to run here it will not
1513 * get delayed by group starvation.
1517 rq_weight
= NICE_0_LOAD
;
1520 if (unlikely(rq_weight
> sd_rq_weight
))
1521 rq_weight
= sd_rq_weight
;
1524 * \Sum shares * rq_weight
1525 * shares = -----------------------
1529 shares
= (sd_shares
* rq_weight
) / (sd_rq_weight
+ 1);
1532 * record the actual number of shares, not the boosted amount.
1534 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1535 tg
->cfs_rq
[cpu
]->rq_weight
= rq_weight
;
1537 if (shares
< MIN_SHARES
)
1538 shares
= MIN_SHARES
;
1539 else if (shares
> MAX_SHARES
)
1540 shares
= MAX_SHARES
;
1542 __set_se_shares(tg
->se
[cpu
], shares
);
1546 * Re-compute the task group their per cpu shares over the given domain.
1547 * This needs to be done in a bottom-up fashion because the rq weight of a
1548 * parent group depends on the shares of its child groups.
1551 tg_shares_up(struct task_group
*tg
, int cpu
, struct sched_domain
*sd
)
1553 unsigned long rq_weight
= 0;
1554 unsigned long shares
= 0;
1557 for_each_cpu_mask(i
, sd
->span
) {
1558 rq_weight
+= tg
->cfs_rq
[i
]->load
.weight
;
1559 shares
+= tg
->cfs_rq
[i
]->shares
;
1562 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1563 shares
= tg
->shares
;
1565 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1566 shares
= tg
->shares
;
1569 rq_weight
= cpus_weight(sd
->span
) * NICE_0_LOAD
;
1571 for_each_cpu_mask(i
, sd
->span
) {
1572 struct rq
*rq
= cpu_rq(i
);
1573 unsigned long flags
;
1575 spin_lock_irqsave(&rq
->lock
, flags
);
1576 __update_group_shares_cpu(tg
, i
, shares
, rq_weight
);
1577 spin_unlock_irqrestore(&rq
->lock
, flags
);
1582 * Compute the cpu's hierarchical load factor for each task group.
1583 * This needs to be done in a top-down fashion because the load of a child
1584 * group is a fraction of its parents load.
1587 tg_load_down(struct task_group
*tg
, int cpu
, struct sched_domain
*sd
)
1592 load
= cpu_rq(cpu
)->load
.weight
;
1594 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1595 load
*= tg
->cfs_rq
[cpu
]->shares
;
1596 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1599 tg
->cfs_rq
[cpu
]->h_load
= load
;
1603 tg_nop(struct task_group
*tg
, int cpu
, struct sched_domain
*sd
)
1607 static void update_shares(struct sched_domain
*sd
)
1609 u64 now
= cpu_clock(raw_smp_processor_id());
1610 s64 elapsed
= now
- sd
->last_update
;
1612 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1613 sd
->last_update
= now
;
1614 walk_tg_tree(tg_nop
, tg_shares_up
, 0, sd
);
1618 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1620 spin_unlock(&rq
->lock
);
1622 spin_lock(&rq
->lock
);
1625 static void update_h_load(int cpu
)
1627 walk_tg_tree(tg_load_down
, tg_nop
, cpu
, NULL
);
1632 static inline void update_shares(struct sched_domain
*sd
)
1636 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1644 #ifdef CONFIG_FAIR_GROUP_SCHED
1645 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1648 cfs_rq
->shares
= shares
;
1653 #include "sched_stats.h"
1654 #include "sched_idletask.c"
1655 #include "sched_fair.c"
1656 #include "sched_rt.c"
1657 #ifdef CONFIG_SCHED_DEBUG
1658 # include "sched_debug.c"
1661 #define sched_class_highest (&rt_sched_class)
1662 #define for_each_class(class) \
1663 for (class = sched_class_highest; class; class = class->next)
1665 static void inc_nr_running(struct rq
*rq
)
1670 static void dec_nr_running(struct rq
*rq
)
1675 static void set_load_weight(struct task_struct
*p
)
1677 if (task_has_rt_policy(p
)) {
1678 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1679 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1684 * SCHED_IDLE tasks get minimal weight:
1686 if (p
->policy
== SCHED_IDLE
) {
1687 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1688 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1692 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1693 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1696 static void update_avg(u64
*avg
, u64 sample
)
1698 s64 diff
= sample
- *avg
;
1702 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1704 sched_info_queued(p
);
1705 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1709 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1711 if (sleep
&& p
->se
.last_wakeup
) {
1712 update_avg(&p
->se
.avg_overlap
,
1713 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1714 p
->se
.last_wakeup
= 0;
1717 sched_info_dequeued(p
);
1718 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1723 * __normal_prio - return the priority that is based on the static prio
1725 static inline int __normal_prio(struct task_struct
*p
)
1727 return p
->static_prio
;
1731 * Calculate the expected normal priority: i.e. priority
1732 * without taking RT-inheritance into account. Might be
1733 * boosted by interactivity modifiers. Changes upon fork,
1734 * setprio syscalls, and whenever the interactivity
1735 * estimator recalculates.
1737 static inline int normal_prio(struct task_struct
*p
)
1741 if (task_has_rt_policy(p
))
1742 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1744 prio
= __normal_prio(p
);
1749 * Calculate the current priority, i.e. the priority
1750 * taken into account by the scheduler. This value might
1751 * be boosted by RT tasks, or might be boosted by
1752 * interactivity modifiers. Will be RT if the task got
1753 * RT-boosted. If not then it returns p->normal_prio.
1755 static int effective_prio(struct task_struct
*p
)
1757 p
->normal_prio
= normal_prio(p
);
1759 * If we are RT tasks or we were boosted to RT priority,
1760 * keep the priority unchanged. Otherwise, update priority
1761 * to the normal priority:
1763 if (!rt_prio(p
->prio
))
1764 return p
->normal_prio
;
1769 * activate_task - move a task to the runqueue.
1771 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1773 if (task_contributes_to_load(p
))
1774 rq
->nr_uninterruptible
--;
1776 enqueue_task(rq
, p
, wakeup
);
1781 * deactivate_task - remove a task from the runqueue.
1783 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1785 if (task_contributes_to_load(p
))
1786 rq
->nr_uninterruptible
++;
1788 dequeue_task(rq
, p
, sleep
);
1793 * task_curr - is this task currently executing on a CPU?
1794 * @p: the task in question.
1796 inline int task_curr(const struct task_struct
*p
)
1798 return cpu_curr(task_cpu(p
)) == p
;
1801 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1803 set_task_rq(p
, cpu
);
1806 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1807 * successfuly executed on another CPU. We must ensure that updates of
1808 * per-task data have been completed by this moment.
1811 task_thread_info(p
)->cpu
= cpu
;
1815 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1816 const struct sched_class
*prev_class
,
1817 int oldprio
, int running
)
1819 if (prev_class
!= p
->sched_class
) {
1820 if (prev_class
->switched_from
)
1821 prev_class
->switched_from(rq
, p
, running
);
1822 p
->sched_class
->switched_to(rq
, p
, running
);
1824 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1829 /* Used instead of source_load when we know the type == 0 */
1830 static unsigned long weighted_cpuload(const int cpu
)
1832 return cpu_rq(cpu
)->load
.weight
;
1836 * Is this task likely cache-hot:
1839 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1844 * Buddy candidates are cache hot:
1846 if (sched_feat(CACHE_HOT_BUDDY
) && (&p
->se
== cfs_rq_of(&p
->se
)->next
))
1849 if (p
->sched_class
!= &fair_sched_class
)
1852 if (sysctl_sched_migration_cost
== -1)
1854 if (sysctl_sched_migration_cost
== 0)
1857 delta
= now
- p
->se
.exec_start
;
1859 return delta
< (s64
)sysctl_sched_migration_cost
;
1863 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1865 int old_cpu
= task_cpu(p
);
1866 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1867 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1868 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1871 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1873 #ifdef CONFIG_SCHEDSTATS
1874 if (p
->se
.wait_start
)
1875 p
->se
.wait_start
-= clock_offset
;
1876 if (p
->se
.sleep_start
)
1877 p
->se
.sleep_start
-= clock_offset
;
1878 if (p
->se
.block_start
)
1879 p
->se
.block_start
-= clock_offset
;
1880 if (old_cpu
!= new_cpu
) {
1881 schedstat_inc(p
, se
.nr_migrations
);
1882 if (task_hot(p
, old_rq
->clock
, NULL
))
1883 schedstat_inc(p
, se
.nr_forced2_migrations
);
1886 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1887 new_cfsrq
->min_vruntime
;
1889 __set_task_cpu(p
, new_cpu
);
1892 struct migration_req
{
1893 struct list_head list
;
1895 struct task_struct
*task
;
1898 struct completion done
;
1902 * The task's runqueue lock must be held.
1903 * Returns true if you have to wait for migration thread.
1906 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1908 struct rq
*rq
= task_rq(p
);
1911 * If the task is not on a runqueue (and not running), then
1912 * it is sufficient to simply update the task's cpu field.
1914 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1915 set_task_cpu(p
, dest_cpu
);
1919 init_completion(&req
->done
);
1921 req
->dest_cpu
= dest_cpu
;
1922 list_add(&req
->list
, &rq
->migration_queue
);
1928 * wait_task_inactive - wait for a thread to unschedule.
1930 * The caller must ensure that the task *will* unschedule sometime soon,
1931 * else this function might spin for a *long* time. This function can't
1932 * be called with interrupts off, or it may introduce deadlock with
1933 * smp_call_function() if an IPI is sent by the same process we are
1934 * waiting to become inactive.
1936 void wait_task_inactive(struct task_struct
*p
)
1938 unsigned long flags
;
1944 * We do the initial early heuristics without holding
1945 * any task-queue locks at all. We'll only try to get
1946 * the runqueue lock when things look like they will
1952 * If the task is actively running on another CPU
1953 * still, just relax and busy-wait without holding
1956 * NOTE! Since we don't hold any locks, it's not
1957 * even sure that "rq" stays as the right runqueue!
1958 * But we don't care, since "task_running()" will
1959 * return false if the runqueue has changed and p
1960 * is actually now running somewhere else!
1962 while (task_running(rq
, p
))
1966 * Ok, time to look more closely! We need the rq
1967 * lock now, to be *sure*. If we're wrong, we'll
1968 * just go back and repeat.
1970 rq
= task_rq_lock(p
, &flags
);
1971 running
= task_running(rq
, p
);
1972 on_rq
= p
->se
.on_rq
;
1973 task_rq_unlock(rq
, &flags
);
1976 * Was it really running after all now that we
1977 * checked with the proper locks actually held?
1979 * Oops. Go back and try again..
1981 if (unlikely(running
)) {
1987 * It's not enough that it's not actively running,
1988 * it must be off the runqueue _entirely_, and not
1991 * So if it wa still runnable (but just not actively
1992 * running right now), it's preempted, and we should
1993 * yield - it could be a while.
1995 if (unlikely(on_rq
)) {
1996 schedule_timeout_uninterruptible(1);
2001 * Ahh, all good. It wasn't running, and it wasn't
2002 * runnable, which means that it will never become
2003 * running in the future either. We're all done!
2010 * kick_process - kick a running thread to enter/exit the kernel
2011 * @p: the to-be-kicked thread
2013 * Cause a process which is running on another CPU to enter
2014 * kernel-mode, without any delay. (to get signals handled.)
2016 * NOTE: this function doesnt have to take the runqueue lock,
2017 * because all it wants to ensure is that the remote task enters
2018 * the kernel. If the IPI races and the task has been migrated
2019 * to another CPU then no harm is done and the purpose has been
2022 void kick_process(struct task_struct
*p
)
2028 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2029 smp_send_reschedule(cpu
);
2034 * Return a low guess at the load of a migration-source cpu weighted
2035 * according to the scheduling class and "nice" value.
2037 * We want to under-estimate the load of migration sources, to
2038 * balance conservatively.
2040 static unsigned long source_load(int cpu
, int type
)
2042 struct rq
*rq
= cpu_rq(cpu
);
2043 unsigned long total
= weighted_cpuload(cpu
);
2045 if (type
== 0 || !sched_feat(LB_BIAS
))
2048 return min(rq
->cpu_load
[type
-1], total
);
2052 * Return a high guess at the load of a migration-target cpu weighted
2053 * according to the scheduling class and "nice" value.
2055 static unsigned long target_load(int cpu
, int type
)
2057 struct rq
*rq
= cpu_rq(cpu
);
2058 unsigned long total
= weighted_cpuload(cpu
);
2060 if (type
== 0 || !sched_feat(LB_BIAS
))
2063 return max(rq
->cpu_load
[type
-1], total
);
2067 * find_idlest_group finds and returns the least busy CPU group within the
2070 static struct sched_group
*
2071 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
2073 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2074 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
2075 int load_idx
= sd
->forkexec_idx
;
2076 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
2079 unsigned long load
, avg_load
;
2083 /* Skip over this group if it has no CPUs allowed */
2084 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
2087 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2089 /* Tally up the load of all CPUs in the group */
2092 for_each_cpu_mask(i
, group
->cpumask
) {
2093 /* Bias balancing toward cpus of our domain */
2095 load
= source_load(i
, load_idx
);
2097 load
= target_load(i
, load_idx
);
2102 /* Adjust by relative CPU power of the group */
2103 avg_load
= sg_div_cpu_power(group
,
2104 avg_load
* SCHED_LOAD_SCALE
);
2107 this_load
= avg_load
;
2109 } else if (avg_load
< min_load
) {
2110 min_load
= avg_load
;
2113 } while (group
= group
->next
, group
!= sd
->groups
);
2115 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
2121 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2124 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
,
2127 unsigned long load
, min_load
= ULONG_MAX
;
2131 /* Traverse only the allowed CPUs */
2132 cpus_and(*tmp
, group
->cpumask
, p
->cpus_allowed
);
2134 for_each_cpu_mask(i
, *tmp
) {
2135 load
= weighted_cpuload(i
);
2137 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2147 * sched_balance_self: balance the current task (running on cpu) in domains
2148 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2151 * Balance, ie. select the least loaded group.
2153 * Returns the target CPU number, or the same CPU if no balancing is needed.
2155 * preempt must be disabled.
2157 static int sched_balance_self(int cpu
, int flag
)
2159 struct task_struct
*t
= current
;
2160 struct sched_domain
*tmp
, *sd
= NULL
;
2162 for_each_domain(cpu
, tmp
) {
2164 * If power savings logic is enabled for a domain, stop there.
2166 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2168 if (tmp
->flags
& flag
)
2176 cpumask_t span
, tmpmask
;
2177 struct sched_group
*group
;
2178 int new_cpu
, weight
;
2180 if (!(sd
->flags
& flag
)) {
2186 group
= find_idlest_group(sd
, t
, cpu
);
2192 new_cpu
= find_idlest_cpu(group
, t
, cpu
, &tmpmask
);
2193 if (new_cpu
== -1 || new_cpu
== cpu
) {
2194 /* Now try balancing at a lower domain level of cpu */
2199 /* Now try balancing at a lower domain level of new_cpu */
2202 weight
= cpus_weight(span
);
2203 for_each_domain(cpu
, tmp
) {
2204 if (weight
<= cpus_weight(tmp
->span
))
2206 if (tmp
->flags
& flag
)
2209 /* while loop will break here if sd == NULL */
2215 #endif /* CONFIG_SMP */
2218 * try_to_wake_up - wake up a thread
2219 * @p: the to-be-woken-up thread
2220 * @state: the mask of task states that can be woken
2221 * @sync: do a synchronous wakeup?
2223 * Put it on the run-queue if it's not already there. The "current"
2224 * thread is always on the run-queue (except when the actual
2225 * re-schedule is in progress), and as such you're allowed to do
2226 * the simpler "current->state = TASK_RUNNING" to mark yourself
2227 * runnable without the overhead of this.
2229 * returns failure only if the task is already active.
2231 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
2233 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2234 unsigned long flags
;
2238 if (!sched_feat(SYNC_WAKEUPS
))
2242 if (sched_feat(LB_WAKEUP_UPDATE
)) {
2243 struct sched_domain
*sd
;
2245 this_cpu
= raw_smp_processor_id();
2248 for_each_domain(this_cpu
, sd
) {
2249 if (cpu_isset(cpu
, sd
->span
)) {
2258 rq
= task_rq_lock(p
, &flags
);
2259 old_state
= p
->state
;
2260 if (!(old_state
& state
))
2268 this_cpu
= smp_processor_id();
2271 if (unlikely(task_running(rq
, p
)))
2274 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2275 if (cpu
!= orig_cpu
) {
2276 set_task_cpu(p
, cpu
);
2277 task_rq_unlock(rq
, &flags
);
2278 /* might preempt at this point */
2279 rq
= task_rq_lock(p
, &flags
);
2280 old_state
= p
->state
;
2281 if (!(old_state
& state
))
2286 this_cpu
= smp_processor_id();
2290 #ifdef CONFIG_SCHEDSTATS
2291 schedstat_inc(rq
, ttwu_count
);
2292 if (cpu
== this_cpu
)
2293 schedstat_inc(rq
, ttwu_local
);
2295 struct sched_domain
*sd
;
2296 for_each_domain(this_cpu
, sd
) {
2297 if (cpu_isset(cpu
, sd
->span
)) {
2298 schedstat_inc(sd
, ttwu_wake_remote
);
2303 #endif /* CONFIG_SCHEDSTATS */
2306 #endif /* CONFIG_SMP */
2307 schedstat_inc(p
, se
.nr_wakeups
);
2309 schedstat_inc(p
, se
.nr_wakeups_sync
);
2310 if (orig_cpu
!= cpu
)
2311 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2312 if (cpu
== this_cpu
)
2313 schedstat_inc(p
, se
.nr_wakeups_local
);
2315 schedstat_inc(p
, se
.nr_wakeups_remote
);
2316 update_rq_clock(rq
);
2317 activate_task(rq
, p
, 1);
2321 check_preempt_curr(rq
, p
);
2323 p
->state
= TASK_RUNNING
;
2325 if (p
->sched_class
->task_wake_up
)
2326 p
->sched_class
->task_wake_up(rq
, p
);
2329 current
->se
.last_wakeup
= current
->se
.sum_exec_runtime
;
2331 task_rq_unlock(rq
, &flags
);
2336 int wake_up_process(struct task_struct
*p
)
2338 return try_to_wake_up(p
, TASK_ALL
, 0);
2340 EXPORT_SYMBOL(wake_up_process
);
2342 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2344 return try_to_wake_up(p
, state
, 0);
2348 * Perform scheduler related setup for a newly forked process p.
2349 * p is forked by current.
2351 * __sched_fork() is basic setup used by init_idle() too:
2353 static void __sched_fork(struct task_struct
*p
)
2355 p
->se
.exec_start
= 0;
2356 p
->se
.sum_exec_runtime
= 0;
2357 p
->se
.prev_sum_exec_runtime
= 0;
2358 p
->se
.last_wakeup
= 0;
2359 p
->se
.avg_overlap
= 0;
2361 #ifdef CONFIG_SCHEDSTATS
2362 p
->se
.wait_start
= 0;
2363 p
->se
.sum_sleep_runtime
= 0;
2364 p
->se
.sleep_start
= 0;
2365 p
->se
.block_start
= 0;
2366 p
->se
.sleep_max
= 0;
2367 p
->se
.block_max
= 0;
2369 p
->se
.slice_max
= 0;
2373 INIT_LIST_HEAD(&p
->rt
.run_list
);
2375 INIT_LIST_HEAD(&p
->se
.group_node
);
2377 #ifdef CONFIG_PREEMPT_NOTIFIERS
2378 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2382 * We mark the process as running here, but have not actually
2383 * inserted it onto the runqueue yet. This guarantees that
2384 * nobody will actually run it, and a signal or other external
2385 * event cannot wake it up and insert it on the runqueue either.
2387 p
->state
= TASK_RUNNING
;
2391 * fork()/clone()-time setup:
2393 void sched_fork(struct task_struct
*p
, int clone_flags
)
2395 int cpu
= get_cpu();
2400 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2402 set_task_cpu(p
, cpu
);
2405 * Make sure we do not leak PI boosting priority to the child:
2407 p
->prio
= current
->normal_prio
;
2408 if (!rt_prio(p
->prio
))
2409 p
->sched_class
= &fair_sched_class
;
2411 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2412 if (likely(sched_info_on()))
2413 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2415 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2418 #ifdef CONFIG_PREEMPT
2419 /* Want to start with kernel preemption disabled. */
2420 task_thread_info(p
)->preempt_count
= 1;
2426 * wake_up_new_task - wake up a newly created task for the first time.
2428 * This function will do some initial scheduler statistics housekeeping
2429 * that must be done for every newly created context, then puts the task
2430 * on the runqueue and wakes it.
2432 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2434 unsigned long flags
;
2437 rq
= task_rq_lock(p
, &flags
);
2438 BUG_ON(p
->state
!= TASK_RUNNING
);
2439 update_rq_clock(rq
);
2441 p
->prio
= effective_prio(p
);
2443 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2444 activate_task(rq
, p
, 0);
2447 * Let the scheduling class do new task startup
2448 * management (if any):
2450 p
->sched_class
->task_new(rq
, p
);
2453 check_preempt_curr(rq
, p
);
2455 if (p
->sched_class
->task_wake_up
)
2456 p
->sched_class
->task_wake_up(rq
, p
);
2458 task_rq_unlock(rq
, &flags
);
2461 #ifdef CONFIG_PREEMPT_NOTIFIERS
2464 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2465 * @notifier: notifier struct to register
2467 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2469 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2471 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2474 * preempt_notifier_unregister - no longer interested in preemption notifications
2475 * @notifier: notifier struct to unregister
2477 * This is safe to call from within a preemption notifier.
2479 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2481 hlist_del(¬ifier
->link
);
2483 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2485 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2487 struct preempt_notifier
*notifier
;
2488 struct hlist_node
*node
;
2490 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2491 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2495 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2496 struct task_struct
*next
)
2498 struct preempt_notifier
*notifier
;
2499 struct hlist_node
*node
;
2501 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2502 notifier
->ops
->sched_out(notifier
, next
);
2505 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2507 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2512 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2513 struct task_struct
*next
)
2517 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2520 * prepare_task_switch - prepare to switch tasks
2521 * @rq: the runqueue preparing to switch
2522 * @prev: the current task that is being switched out
2523 * @next: the task we are going to switch to.
2525 * This is called with the rq lock held and interrupts off. It must
2526 * be paired with a subsequent finish_task_switch after the context
2529 * prepare_task_switch sets up locking and calls architecture specific
2533 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2534 struct task_struct
*next
)
2536 fire_sched_out_preempt_notifiers(prev
, next
);
2537 prepare_lock_switch(rq
, next
);
2538 prepare_arch_switch(next
);
2542 * finish_task_switch - clean up after a task-switch
2543 * @rq: runqueue associated with task-switch
2544 * @prev: the thread we just switched away from.
2546 * finish_task_switch must be called after the context switch, paired
2547 * with a prepare_task_switch call before the context switch.
2548 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2549 * and do any other architecture-specific cleanup actions.
2551 * Note that we may have delayed dropping an mm in context_switch(). If
2552 * so, we finish that here outside of the runqueue lock. (Doing it
2553 * with the lock held can cause deadlocks; see schedule() for
2556 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2557 __releases(rq
->lock
)
2559 struct mm_struct
*mm
= rq
->prev_mm
;
2565 * A task struct has one reference for the use as "current".
2566 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2567 * schedule one last time. The schedule call will never return, and
2568 * the scheduled task must drop that reference.
2569 * The test for TASK_DEAD must occur while the runqueue locks are
2570 * still held, otherwise prev could be scheduled on another cpu, die
2571 * there before we look at prev->state, and then the reference would
2573 * Manfred Spraul <manfred@colorfullife.com>
2575 prev_state
= prev
->state
;
2576 finish_arch_switch(prev
);
2577 finish_lock_switch(rq
, prev
);
2579 if (current
->sched_class
->post_schedule
)
2580 current
->sched_class
->post_schedule(rq
);
2583 fire_sched_in_preempt_notifiers(current
);
2586 if (unlikely(prev_state
== TASK_DEAD
)) {
2588 * Remove function-return probe instances associated with this
2589 * task and put them back on the free list.
2591 kprobe_flush_task(prev
);
2592 put_task_struct(prev
);
2597 * schedule_tail - first thing a freshly forked thread must call.
2598 * @prev: the thread we just switched away from.
2600 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2601 __releases(rq
->lock
)
2603 struct rq
*rq
= this_rq();
2605 finish_task_switch(rq
, prev
);
2606 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2607 /* In this case, finish_task_switch does not reenable preemption */
2610 if (current
->set_child_tid
)
2611 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2615 * context_switch - switch to the new MM and the new
2616 * thread's register state.
2619 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2620 struct task_struct
*next
)
2622 struct mm_struct
*mm
, *oldmm
;
2624 prepare_task_switch(rq
, prev
, next
);
2626 oldmm
= prev
->active_mm
;
2628 * For paravirt, this is coupled with an exit in switch_to to
2629 * combine the page table reload and the switch backend into
2632 arch_enter_lazy_cpu_mode();
2634 if (unlikely(!mm
)) {
2635 next
->active_mm
= oldmm
;
2636 atomic_inc(&oldmm
->mm_count
);
2637 enter_lazy_tlb(oldmm
, next
);
2639 switch_mm(oldmm
, mm
, next
);
2641 if (unlikely(!prev
->mm
)) {
2642 prev
->active_mm
= NULL
;
2643 rq
->prev_mm
= oldmm
;
2646 * Since the runqueue lock will be released by the next
2647 * task (which is an invalid locking op but in the case
2648 * of the scheduler it's an obvious special-case), so we
2649 * do an early lockdep release here:
2651 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2652 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2655 /* Here we just switch the register state and the stack. */
2656 switch_to(prev
, next
, prev
);
2660 * this_rq must be evaluated again because prev may have moved
2661 * CPUs since it called schedule(), thus the 'rq' on its stack
2662 * frame will be invalid.
2664 finish_task_switch(this_rq(), prev
);
2668 * nr_running, nr_uninterruptible and nr_context_switches:
2670 * externally visible scheduler statistics: current number of runnable
2671 * threads, current number of uninterruptible-sleeping threads, total
2672 * number of context switches performed since bootup.
2674 unsigned long nr_running(void)
2676 unsigned long i
, sum
= 0;
2678 for_each_online_cpu(i
)
2679 sum
+= cpu_rq(i
)->nr_running
;
2684 unsigned long nr_uninterruptible(void)
2686 unsigned long i
, sum
= 0;
2688 for_each_possible_cpu(i
)
2689 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2692 * Since we read the counters lockless, it might be slightly
2693 * inaccurate. Do not allow it to go below zero though:
2695 if (unlikely((long)sum
< 0))
2701 unsigned long long nr_context_switches(void)
2704 unsigned long long sum
= 0;
2706 for_each_possible_cpu(i
)
2707 sum
+= cpu_rq(i
)->nr_switches
;
2712 unsigned long nr_iowait(void)
2714 unsigned long i
, sum
= 0;
2716 for_each_possible_cpu(i
)
2717 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2722 unsigned long nr_active(void)
2724 unsigned long i
, running
= 0, uninterruptible
= 0;
2726 for_each_online_cpu(i
) {
2727 running
+= cpu_rq(i
)->nr_running
;
2728 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2731 if (unlikely((long)uninterruptible
< 0))
2732 uninterruptible
= 0;
2734 return running
+ uninterruptible
;
2738 * Update rq->cpu_load[] statistics. This function is usually called every
2739 * scheduler tick (TICK_NSEC).
2741 static void update_cpu_load(struct rq
*this_rq
)
2743 unsigned long this_load
= this_rq
->load
.weight
;
2746 this_rq
->nr_load_updates
++;
2748 /* Update our load: */
2749 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2750 unsigned long old_load
, new_load
;
2752 /* scale is effectively 1 << i now, and >> i divides by scale */
2754 old_load
= this_rq
->cpu_load
[i
];
2755 new_load
= this_load
;
2757 * Round up the averaging division if load is increasing. This
2758 * prevents us from getting stuck on 9 if the load is 10, for
2761 if (new_load
> old_load
)
2762 new_load
+= scale
-1;
2763 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2770 * double_rq_lock - safely lock two runqueues
2772 * Note this does not disable interrupts like task_rq_lock,
2773 * you need to do so manually before calling.
2775 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2776 __acquires(rq1
->lock
)
2777 __acquires(rq2
->lock
)
2779 BUG_ON(!irqs_disabled());
2781 spin_lock(&rq1
->lock
);
2782 __acquire(rq2
->lock
); /* Fake it out ;) */
2785 spin_lock(&rq1
->lock
);
2786 spin_lock(&rq2
->lock
);
2788 spin_lock(&rq2
->lock
);
2789 spin_lock(&rq1
->lock
);
2792 update_rq_clock(rq1
);
2793 update_rq_clock(rq2
);
2797 * double_rq_unlock - safely unlock two runqueues
2799 * Note this does not restore interrupts like task_rq_unlock,
2800 * you need to do so manually after calling.
2802 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2803 __releases(rq1
->lock
)
2804 __releases(rq2
->lock
)
2806 spin_unlock(&rq1
->lock
);
2808 spin_unlock(&rq2
->lock
);
2810 __release(rq2
->lock
);
2814 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2816 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2817 __releases(this_rq
->lock
)
2818 __acquires(busiest
->lock
)
2819 __acquires(this_rq
->lock
)
2823 if (unlikely(!irqs_disabled())) {
2824 /* printk() doesn't work good under rq->lock */
2825 spin_unlock(&this_rq
->lock
);
2828 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2829 if (busiest
< this_rq
) {
2830 spin_unlock(&this_rq
->lock
);
2831 spin_lock(&busiest
->lock
);
2832 spin_lock(&this_rq
->lock
);
2835 spin_lock(&busiest
->lock
);
2841 * If dest_cpu is allowed for this process, migrate the task to it.
2842 * This is accomplished by forcing the cpu_allowed mask to only
2843 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2844 * the cpu_allowed mask is restored.
2846 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2848 struct migration_req req
;
2849 unsigned long flags
;
2852 rq
= task_rq_lock(p
, &flags
);
2853 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2854 || unlikely(cpu_is_offline(dest_cpu
)))
2857 /* force the process onto the specified CPU */
2858 if (migrate_task(p
, dest_cpu
, &req
)) {
2859 /* Need to wait for migration thread (might exit: take ref). */
2860 struct task_struct
*mt
= rq
->migration_thread
;
2862 get_task_struct(mt
);
2863 task_rq_unlock(rq
, &flags
);
2864 wake_up_process(mt
);
2865 put_task_struct(mt
);
2866 wait_for_completion(&req
.done
);
2871 task_rq_unlock(rq
, &flags
);
2875 * sched_exec - execve() is a valuable balancing opportunity, because at
2876 * this point the task has the smallest effective memory and cache footprint.
2878 void sched_exec(void)
2880 int new_cpu
, this_cpu
= get_cpu();
2881 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2883 if (new_cpu
!= this_cpu
)
2884 sched_migrate_task(current
, new_cpu
);
2888 * pull_task - move a task from a remote runqueue to the local runqueue.
2889 * Both runqueues must be locked.
2891 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2892 struct rq
*this_rq
, int this_cpu
)
2894 deactivate_task(src_rq
, p
, 0);
2895 set_task_cpu(p
, this_cpu
);
2896 activate_task(this_rq
, p
, 0);
2898 * Note that idle threads have a prio of MAX_PRIO, for this test
2899 * to be always true for them.
2901 check_preempt_curr(this_rq
, p
);
2905 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2908 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2909 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2913 * We do not migrate tasks that are:
2914 * 1) running (obviously), or
2915 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2916 * 3) are cache-hot on their current CPU.
2918 if (!cpu_isset(this_cpu
, p
->cpus_allowed
)) {
2919 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2924 if (task_running(rq
, p
)) {
2925 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2930 * Aggressive migration if:
2931 * 1) task is cache cold, or
2932 * 2) too many balance attempts have failed.
2935 if (!task_hot(p
, rq
->clock
, sd
) ||
2936 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2937 #ifdef CONFIG_SCHEDSTATS
2938 if (task_hot(p
, rq
->clock
, sd
)) {
2939 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2940 schedstat_inc(p
, se
.nr_forced_migrations
);
2946 if (task_hot(p
, rq
->clock
, sd
)) {
2947 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
2953 static unsigned long
2954 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2955 unsigned long max_load_move
, struct sched_domain
*sd
,
2956 enum cpu_idle_type idle
, int *all_pinned
,
2957 int *this_best_prio
, struct rq_iterator
*iterator
)
2959 int loops
= 0, pulled
= 0, pinned
= 0;
2960 struct task_struct
*p
;
2961 long rem_load_move
= max_load_move
;
2963 if (max_load_move
== 0)
2969 * Start the load-balancing iterator:
2971 p
= iterator
->start(iterator
->arg
);
2973 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
2976 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
2977 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2978 p
= iterator
->next(iterator
->arg
);
2982 pull_task(busiest
, p
, this_rq
, this_cpu
);
2984 rem_load_move
-= p
->se
.load
.weight
;
2987 * We only want to steal up to the prescribed amount of weighted load.
2989 if (rem_load_move
> 0) {
2990 if (p
->prio
< *this_best_prio
)
2991 *this_best_prio
= p
->prio
;
2992 p
= iterator
->next(iterator
->arg
);
2997 * Right now, this is one of only two places pull_task() is called,
2998 * so we can safely collect pull_task() stats here rather than
2999 * inside pull_task().
3001 schedstat_add(sd
, lb_gained
[idle
], pulled
);
3004 *all_pinned
= pinned
;
3006 return max_load_move
- rem_load_move
;
3010 * move_tasks tries to move up to max_load_move weighted load from busiest to
3011 * this_rq, as part of a balancing operation within domain "sd".
3012 * Returns 1 if successful and 0 otherwise.
3014 * Called with both runqueues locked.
3016 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3017 unsigned long max_load_move
,
3018 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3021 const struct sched_class
*class = sched_class_highest
;
3022 unsigned long total_load_moved
= 0;
3023 int this_best_prio
= this_rq
->curr
->prio
;
3027 class->load_balance(this_rq
, this_cpu
, busiest
,
3028 max_load_move
- total_load_moved
,
3029 sd
, idle
, all_pinned
, &this_best_prio
);
3030 class = class->next
;
3032 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3035 } while (class && max_load_move
> total_load_moved
);
3037 return total_load_moved
> 0;
3041 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3042 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3043 struct rq_iterator
*iterator
)
3045 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3049 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3050 pull_task(busiest
, p
, this_rq
, this_cpu
);
3052 * Right now, this is only the second place pull_task()
3053 * is called, so we can safely collect pull_task()
3054 * stats here rather than inside pull_task().
3056 schedstat_inc(sd
, lb_gained
[idle
]);
3060 p
= iterator
->next(iterator
->arg
);
3067 * move_one_task tries to move exactly one task from busiest to this_rq, as
3068 * part of active balancing operations within "domain".
3069 * Returns 1 if successful and 0 otherwise.
3071 * Called with both runqueues locked.
3073 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3074 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3076 const struct sched_class
*class;
3078 for (class = sched_class_highest
; class; class = class->next
)
3079 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3086 * find_busiest_group finds and returns the busiest CPU group within the
3087 * domain. It calculates and returns the amount of weighted load which
3088 * should be moved to restore balance via the imbalance parameter.
3090 static struct sched_group
*
3091 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
3092 unsigned long *imbalance
, enum cpu_idle_type idle
,
3093 int *sd_idle
, const cpumask_t
*cpus
, int *balance
)
3095 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
3096 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
3097 unsigned long max_pull
;
3098 unsigned long busiest_load_per_task
, busiest_nr_running
;
3099 unsigned long this_load_per_task
, this_nr_running
;
3100 int load_idx
, group_imb
= 0;
3101 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3102 int power_savings_balance
= 1;
3103 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
3104 unsigned long min_nr_running
= ULONG_MAX
;
3105 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
3108 max_load
= this_load
= total_load
= total_pwr
= 0;
3109 busiest_load_per_task
= busiest_nr_running
= 0;
3110 this_load_per_task
= this_nr_running
= 0;
3112 if (idle
== CPU_NOT_IDLE
)
3113 load_idx
= sd
->busy_idx
;
3114 else if (idle
== CPU_NEWLY_IDLE
)
3115 load_idx
= sd
->newidle_idx
;
3117 load_idx
= sd
->idle_idx
;
3120 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
3123 int __group_imb
= 0;
3124 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3125 unsigned long sum_nr_running
, sum_weighted_load
;
3126 unsigned long sum_avg_load_per_task
;
3127 unsigned long avg_load_per_task
;
3129 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
3132 balance_cpu
= first_cpu(group
->cpumask
);
3134 /* Tally up the load of all CPUs in the group */
3135 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
3136 sum_avg_load_per_task
= avg_load_per_task
= 0;
3139 min_cpu_load
= ~0UL;
3141 for_each_cpu_mask(i
, group
->cpumask
) {
3144 if (!cpu_isset(i
, *cpus
))
3149 if (*sd_idle
&& rq
->nr_running
)
3152 /* Bias balancing toward cpus of our domain */
3154 if (idle_cpu(i
) && !first_idle_cpu
) {
3159 load
= target_load(i
, load_idx
);
3161 load
= source_load(i
, load_idx
);
3162 if (load
> max_cpu_load
)
3163 max_cpu_load
= load
;
3164 if (min_cpu_load
> load
)
3165 min_cpu_load
= load
;
3169 sum_nr_running
+= rq
->nr_running
;
3170 sum_weighted_load
+= weighted_cpuload(i
);
3172 sum_avg_load_per_task
+= cpu_avg_load_per_task(i
);
3176 * First idle cpu or the first cpu(busiest) in this sched group
3177 * is eligible for doing load balancing at this and above
3178 * domains. In the newly idle case, we will allow all the cpu's
3179 * to do the newly idle load balance.
3181 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3182 balance_cpu
!= this_cpu
&& balance
) {
3187 total_load
+= avg_load
;
3188 total_pwr
+= group
->__cpu_power
;
3190 /* Adjust by relative CPU power of the group */
3191 avg_load
= sg_div_cpu_power(group
,
3192 avg_load
* SCHED_LOAD_SCALE
);
3196 * Consider the group unbalanced when the imbalance is larger
3197 * than the average weight of two tasks.
3199 * APZ: with cgroup the avg task weight can vary wildly and
3200 * might not be a suitable number - should we keep a
3201 * normalized nr_running number somewhere that negates
3204 avg_load_per_task
= sg_div_cpu_power(group
,
3205 sum_avg_load_per_task
* SCHED_LOAD_SCALE
);
3207 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3210 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
3213 this_load
= avg_load
;
3215 this_nr_running
= sum_nr_running
;
3216 this_load_per_task
= sum_weighted_load
;
3217 } else if (avg_load
> max_load
&&
3218 (sum_nr_running
> group_capacity
|| __group_imb
)) {
3219 max_load
= avg_load
;
3221 busiest_nr_running
= sum_nr_running
;
3222 busiest_load_per_task
= sum_weighted_load
;
3223 group_imb
= __group_imb
;
3226 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3228 * Busy processors will not participate in power savings
3231 if (idle
== CPU_NOT_IDLE
||
3232 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3236 * If the local group is idle or completely loaded
3237 * no need to do power savings balance at this domain
3239 if (local_group
&& (this_nr_running
>= group_capacity
||
3241 power_savings_balance
= 0;
3244 * If a group is already running at full capacity or idle,
3245 * don't include that group in power savings calculations
3247 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
3252 * Calculate the group which has the least non-idle load.
3253 * This is the group from where we need to pick up the load
3256 if ((sum_nr_running
< min_nr_running
) ||
3257 (sum_nr_running
== min_nr_running
&&
3258 first_cpu(group
->cpumask
) <
3259 first_cpu(group_min
->cpumask
))) {
3261 min_nr_running
= sum_nr_running
;
3262 min_load_per_task
= sum_weighted_load
/
3267 * Calculate the group which is almost near its
3268 * capacity but still has some space to pick up some load
3269 * from other group and save more power
3271 if (sum_nr_running
<= group_capacity
- 1) {
3272 if (sum_nr_running
> leader_nr_running
||
3273 (sum_nr_running
== leader_nr_running
&&
3274 first_cpu(group
->cpumask
) >
3275 first_cpu(group_leader
->cpumask
))) {
3276 group_leader
= group
;
3277 leader_nr_running
= sum_nr_running
;
3282 group
= group
->next
;
3283 } while (group
!= sd
->groups
);
3285 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
3288 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
3290 if (this_load
>= avg_load
||
3291 100*max_load
<= sd
->imbalance_pct
*this_load
)
3294 busiest_load_per_task
/= busiest_nr_running
;
3296 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
3299 * We're trying to get all the cpus to the average_load, so we don't
3300 * want to push ourselves above the average load, nor do we wish to
3301 * reduce the max loaded cpu below the average load, as either of these
3302 * actions would just result in more rebalancing later, and ping-pong
3303 * tasks around. Thus we look for the minimum possible imbalance.
3304 * Negative imbalances (*we* are more loaded than anyone else) will
3305 * be counted as no imbalance for these purposes -- we can't fix that
3306 * by pulling tasks to us. Be careful of negative numbers as they'll
3307 * appear as very large values with unsigned longs.
3309 if (max_load
<= busiest_load_per_task
)
3313 * In the presence of smp nice balancing, certain scenarios can have
3314 * max load less than avg load(as we skip the groups at or below
3315 * its cpu_power, while calculating max_load..)
3317 if (max_load
< avg_load
) {
3319 goto small_imbalance
;
3322 /* Don't want to pull so many tasks that a group would go idle */
3323 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
3325 /* How much load to actually move to equalise the imbalance */
3326 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
3327 (avg_load
- this_load
) * this->__cpu_power
)
3331 * if *imbalance is less than the average load per runnable task
3332 * there is no gaurantee that any tasks will be moved so we'll have
3333 * a think about bumping its value to force at least one task to be
3336 if (*imbalance
< busiest_load_per_task
) {
3337 unsigned long tmp
, pwr_now
, pwr_move
;
3341 pwr_move
= pwr_now
= 0;
3343 if (this_nr_running
) {
3344 this_load_per_task
/= this_nr_running
;
3345 if (busiest_load_per_task
> this_load_per_task
)
3348 this_load_per_task
= cpu_avg_load_per_task(this_cpu
);
3350 if (max_load
- this_load
+ 2*busiest_load_per_task
>=
3351 busiest_load_per_task
* imbn
) {
3352 *imbalance
= busiest_load_per_task
;
3357 * OK, we don't have enough imbalance to justify moving tasks,
3358 * however we may be able to increase total CPU power used by
3362 pwr_now
+= busiest
->__cpu_power
*
3363 min(busiest_load_per_task
, max_load
);
3364 pwr_now
+= this->__cpu_power
*
3365 min(this_load_per_task
, this_load
);
3366 pwr_now
/= SCHED_LOAD_SCALE
;
3368 /* Amount of load we'd subtract */
3369 tmp
= sg_div_cpu_power(busiest
,
3370 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3372 pwr_move
+= busiest
->__cpu_power
*
3373 min(busiest_load_per_task
, max_load
- tmp
);
3375 /* Amount of load we'd add */
3376 if (max_load
* busiest
->__cpu_power
<
3377 busiest_load_per_task
* SCHED_LOAD_SCALE
)
3378 tmp
= sg_div_cpu_power(this,
3379 max_load
* busiest
->__cpu_power
);
3381 tmp
= sg_div_cpu_power(this,
3382 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3383 pwr_move
+= this->__cpu_power
*
3384 min(this_load_per_task
, this_load
+ tmp
);
3385 pwr_move
/= SCHED_LOAD_SCALE
;
3387 /* Move if we gain throughput */
3388 if (pwr_move
> pwr_now
)
3389 *imbalance
= busiest_load_per_task
;
3395 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3396 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3399 if (this == group_leader
&& group_leader
!= group_min
) {
3400 *imbalance
= min_load_per_task
;
3410 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3413 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
3414 unsigned long imbalance
, const cpumask_t
*cpus
)
3416 struct rq
*busiest
= NULL
, *rq
;
3417 unsigned long max_load
= 0;
3420 for_each_cpu_mask(i
, group
->cpumask
) {
3423 if (!cpu_isset(i
, *cpus
))
3427 wl
= weighted_cpuload(i
);
3429 if (rq
->nr_running
== 1 && wl
> imbalance
)
3432 if (wl
> max_load
) {
3442 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3443 * so long as it is large enough.
3445 #define MAX_PINNED_INTERVAL 512
3448 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3449 * tasks if there is an imbalance.
3451 static int load_balance(int this_cpu
, struct rq
*this_rq
,
3452 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3453 int *balance
, cpumask_t
*cpus
)
3455 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
3456 struct sched_group
*group
;
3457 unsigned long imbalance
;
3459 unsigned long flags
;
3464 * When power savings policy is enabled for the parent domain, idle
3465 * sibling can pick up load irrespective of busy siblings. In this case,
3466 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3467 * portraying it as CPU_NOT_IDLE.
3469 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3470 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3473 schedstat_inc(sd
, lb_count
[idle
]);
3477 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3484 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3488 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
3490 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3494 BUG_ON(busiest
== this_rq
);
3496 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3499 if (busiest
->nr_running
> 1) {
3501 * Attempt to move tasks. If find_busiest_group has found
3502 * an imbalance but busiest->nr_running <= 1, the group is
3503 * still unbalanced. ld_moved simply stays zero, so it is
3504 * correctly treated as an imbalance.
3506 local_irq_save(flags
);
3507 double_rq_lock(this_rq
, busiest
);
3508 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3509 imbalance
, sd
, idle
, &all_pinned
);
3510 double_rq_unlock(this_rq
, busiest
);
3511 local_irq_restore(flags
);
3514 * some other cpu did the load balance for us.
3516 if (ld_moved
&& this_cpu
!= smp_processor_id())
3517 resched_cpu(this_cpu
);
3519 /* All tasks on this runqueue were pinned by CPU affinity */
3520 if (unlikely(all_pinned
)) {
3521 cpu_clear(cpu_of(busiest
), *cpus
);
3522 if (!cpus_empty(*cpus
))
3529 schedstat_inc(sd
, lb_failed
[idle
]);
3530 sd
->nr_balance_failed
++;
3532 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3534 spin_lock_irqsave(&busiest
->lock
, flags
);
3536 /* don't kick the migration_thread, if the curr
3537 * task on busiest cpu can't be moved to this_cpu
3539 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
3540 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3542 goto out_one_pinned
;
3545 if (!busiest
->active_balance
) {
3546 busiest
->active_balance
= 1;
3547 busiest
->push_cpu
= this_cpu
;
3550 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3552 wake_up_process(busiest
->migration_thread
);
3555 * We've kicked active balancing, reset the failure
3558 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3561 sd
->nr_balance_failed
= 0;
3563 if (likely(!active_balance
)) {
3564 /* We were unbalanced, so reset the balancing interval */
3565 sd
->balance_interval
= sd
->min_interval
;
3568 * If we've begun active balancing, start to back off. This
3569 * case may not be covered by the all_pinned logic if there
3570 * is only 1 task on the busy runqueue (because we don't call
3573 if (sd
->balance_interval
< sd
->max_interval
)
3574 sd
->balance_interval
*= 2;
3577 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3578 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3584 schedstat_inc(sd
, lb_balanced
[idle
]);
3586 sd
->nr_balance_failed
= 0;
3589 /* tune up the balancing interval */
3590 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3591 (sd
->balance_interval
< sd
->max_interval
))
3592 sd
->balance_interval
*= 2;
3594 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3595 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3606 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3607 * tasks if there is an imbalance.
3609 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3610 * this_rq is locked.
3613 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
,
3616 struct sched_group
*group
;
3617 struct rq
*busiest
= NULL
;
3618 unsigned long imbalance
;
3626 * When power savings policy is enabled for the parent domain, idle
3627 * sibling can pick up load irrespective of busy siblings. In this case,
3628 * let the state of idle sibling percolate up as IDLE, instead of
3629 * portraying it as CPU_NOT_IDLE.
3631 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
3632 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3635 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
3637 update_shares_locked(this_rq
, sd
);
3638 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
3639 &sd_idle
, cpus
, NULL
);
3641 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
3645 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
3647 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
3651 BUG_ON(busiest
== this_rq
);
3653 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
3656 if (busiest
->nr_running
> 1) {
3657 /* Attempt to move tasks */
3658 double_lock_balance(this_rq
, busiest
);
3659 /* this_rq->clock is already updated */
3660 update_rq_clock(busiest
);
3661 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3662 imbalance
, sd
, CPU_NEWLY_IDLE
,
3664 spin_unlock(&busiest
->lock
);
3666 if (unlikely(all_pinned
)) {
3667 cpu_clear(cpu_of(busiest
), *cpus
);
3668 if (!cpus_empty(*cpus
))
3674 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
3675 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3676 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3679 sd
->nr_balance_failed
= 0;
3681 update_shares_locked(this_rq
, sd
);
3685 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
3686 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3687 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3689 sd
->nr_balance_failed
= 0;
3695 * idle_balance is called by schedule() if this_cpu is about to become
3696 * idle. Attempts to pull tasks from other CPUs.
3698 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
3700 struct sched_domain
*sd
;
3701 int pulled_task
= -1;
3702 unsigned long next_balance
= jiffies
+ HZ
;
3705 for_each_domain(this_cpu
, sd
) {
3706 unsigned long interval
;
3708 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3711 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
3712 /* If we've pulled tasks over stop searching: */
3713 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
3716 interval
= msecs_to_jiffies(sd
->balance_interval
);
3717 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3718 next_balance
= sd
->last_balance
+ interval
;
3722 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3724 * We are going idle. next_balance may be set based on
3725 * a busy processor. So reset next_balance.
3727 this_rq
->next_balance
= next_balance
;
3732 * active_load_balance is run by migration threads. It pushes running tasks
3733 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3734 * running on each physical CPU where possible, and avoids physical /
3735 * logical imbalances.
3737 * Called with busiest_rq locked.
3739 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
3741 int target_cpu
= busiest_rq
->push_cpu
;
3742 struct sched_domain
*sd
;
3743 struct rq
*target_rq
;
3745 /* Is there any task to move? */
3746 if (busiest_rq
->nr_running
<= 1)
3749 target_rq
= cpu_rq(target_cpu
);
3752 * This condition is "impossible", if it occurs
3753 * we need to fix it. Originally reported by
3754 * Bjorn Helgaas on a 128-cpu setup.
3756 BUG_ON(busiest_rq
== target_rq
);
3758 /* move a task from busiest_rq to target_rq */
3759 double_lock_balance(busiest_rq
, target_rq
);
3760 update_rq_clock(busiest_rq
);
3761 update_rq_clock(target_rq
);
3763 /* Search for an sd spanning us and the target CPU. */
3764 for_each_domain(target_cpu
, sd
) {
3765 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3766 cpu_isset(busiest_cpu
, sd
->span
))
3771 schedstat_inc(sd
, alb_count
);
3773 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3775 schedstat_inc(sd
, alb_pushed
);
3777 schedstat_inc(sd
, alb_failed
);
3779 spin_unlock(&target_rq
->lock
);
3784 atomic_t load_balancer
;
3786 } nohz ____cacheline_aligned
= {
3787 .load_balancer
= ATOMIC_INIT(-1),
3788 .cpu_mask
= CPU_MASK_NONE
,
3792 * This routine will try to nominate the ilb (idle load balancing)
3793 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3794 * load balancing on behalf of all those cpus. If all the cpus in the system
3795 * go into this tickless mode, then there will be no ilb owner (as there is
3796 * no need for one) and all the cpus will sleep till the next wakeup event
3799 * For the ilb owner, tick is not stopped. And this tick will be used
3800 * for idle load balancing. ilb owner will still be part of
3803 * While stopping the tick, this cpu will become the ilb owner if there
3804 * is no other owner. And will be the owner till that cpu becomes busy
3805 * or if all cpus in the system stop their ticks at which point
3806 * there is no need for ilb owner.
3808 * When the ilb owner becomes busy, it nominates another owner, during the
3809 * next busy scheduler_tick()
3811 int select_nohz_load_balancer(int stop_tick
)
3813 int cpu
= smp_processor_id();
3816 cpu_set(cpu
, nohz
.cpu_mask
);
3817 cpu_rq(cpu
)->in_nohz_recently
= 1;
3820 * If we are going offline and still the leader, give up!
3822 if (cpu_is_offline(cpu
) &&
3823 atomic_read(&nohz
.load_balancer
) == cpu
) {
3824 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3829 /* time for ilb owner also to sleep */
3830 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3831 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3832 atomic_set(&nohz
.load_balancer
, -1);
3836 if (atomic_read(&nohz
.load_balancer
) == -1) {
3837 /* make me the ilb owner */
3838 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3840 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3843 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3846 cpu_clear(cpu
, nohz
.cpu_mask
);
3848 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3849 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3856 static DEFINE_SPINLOCK(balancing
);
3859 * It checks each scheduling domain to see if it is due to be balanced,
3860 * and initiates a balancing operation if so.
3862 * Balancing parameters are set up in arch_init_sched_domains.
3864 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3867 struct rq
*rq
= cpu_rq(cpu
);
3868 unsigned long interval
;
3869 struct sched_domain
*sd
;
3870 /* Earliest time when we have to do rebalance again */
3871 unsigned long next_balance
= jiffies
+ 60*HZ
;
3872 int update_next_balance
= 0;
3876 for_each_domain(cpu
, sd
) {
3877 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3880 interval
= sd
->balance_interval
;
3881 if (idle
!= CPU_IDLE
)
3882 interval
*= sd
->busy_factor
;
3884 /* scale ms to jiffies */
3885 interval
= msecs_to_jiffies(interval
);
3886 if (unlikely(!interval
))
3888 if (interval
> HZ
*NR_CPUS
/10)
3889 interval
= HZ
*NR_CPUS
/10;
3891 need_serialize
= sd
->flags
& SD_SERIALIZE
;
3893 if (need_serialize
) {
3894 if (!spin_trylock(&balancing
))
3898 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3899 if (load_balance(cpu
, rq
, sd
, idle
, &balance
, &tmp
)) {
3901 * We've pulled tasks over so either we're no
3902 * longer idle, or one of our SMT siblings is
3905 idle
= CPU_NOT_IDLE
;
3907 sd
->last_balance
= jiffies
;
3910 spin_unlock(&balancing
);
3912 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3913 next_balance
= sd
->last_balance
+ interval
;
3914 update_next_balance
= 1;
3918 * Stop the load balance at this level. There is another
3919 * CPU in our sched group which is doing load balancing more
3927 * next_balance will be updated only when there is a need.
3928 * When the cpu is attached to null domain for ex, it will not be
3931 if (likely(update_next_balance
))
3932 rq
->next_balance
= next_balance
;
3936 * run_rebalance_domains is triggered when needed from the scheduler tick.
3937 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3938 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3940 static void run_rebalance_domains(struct softirq_action
*h
)
3942 int this_cpu
= smp_processor_id();
3943 struct rq
*this_rq
= cpu_rq(this_cpu
);
3944 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3945 CPU_IDLE
: CPU_NOT_IDLE
;
3947 rebalance_domains(this_cpu
, idle
);
3951 * If this cpu is the owner for idle load balancing, then do the
3952 * balancing on behalf of the other idle cpus whose ticks are
3955 if (this_rq
->idle_at_tick
&&
3956 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3957 cpumask_t cpus
= nohz
.cpu_mask
;
3961 cpu_clear(this_cpu
, cpus
);
3962 for_each_cpu_mask(balance_cpu
, cpus
) {
3964 * If this cpu gets work to do, stop the load balancing
3965 * work being done for other cpus. Next load
3966 * balancing owner will pick it up.
3971 rebalance_domains(balance_cpu
, CPU_IDLE
);
3973 rq
= cpu_rq(balance_cpu
);
3974 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3975 this_rq
->next_balance
= rq
->next_balance
;
3982 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3984 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3985 * idle load balancing owner or decide to stop the periodic load balancing,
3986 * if the whole system is idle.
3988 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3992 * If we were in the nohz mode recently and busy at the current
3993 * scheduler tick, then check if we need to nominate new idle
3996 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3997 rq
->in_nohz_recently
= 0;
3999 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4000 cpu_clear(cpu
, nohz
.cpu_mask
);
4001 atomic_set(&nohz
.load_balancer
, -1);
4004 if (atomic_read(&nohz
.load_balancer
) == -1) {
4006 * simple selection for now: Nominate the
4007 * first cpu in the nohz list to be the next
4010 * TBD: Traverse the sched domains and nominate
4011 * the nearest cpu in the nohz.cpu_mask.
4013 int ilb
= first_cpu(nohz
.cpu_mask
);
4015 if (ilb
< nr_cpu_ids
)
4021 * If this cpu is idle and doing idle load balancing for all the
4022 * cpus with ticks stopped, is it time for that to stop?
4024 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4025 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4031 * If this cpu is idle and the idle load balancing is done by
4032 * someone else, then no need raise the SCHED_SOFTIRQ
4034 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4035 cpu_isset(cpu
, nohz
.cpu_mask
))
4038 if (time_after_eq(jiffies
, rq
->next_balance
))
4039 raise_softirq(SCHED_SOFTIRQ
);
4042 #else /* CONFIG_SMP */
4045 * on UP we do not need to balance between CPUs:
4047 static inline void idle_balance(int cpu
, struct rq
*rq
)
4053 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4055 EXPORT_PER_CPU_SYMBOL(kstat
);
4058 * Return p->sum_exec_runtime plus any more ns on the sched_clock
4059 * that have not yet been banked in case the task is currently running.
4061 unsigned long long task_sched_runtime(struct task_struct
*p
)
4063 unsigned long flags
;
4067 rq
= task_rq_lock(p
, &flags
);
4068 ns
= p
->se
.sum_exec_runtime
;
4069 if (task_current(rq
, p
)) {
4070 update_rq_clock(rq
);
4071 delta_exec
= rq
->clock
- p
->se
.exec_start
;
4072 if ((s64
)delta_exec
> 0)
4075 task_rq_unlock(rq
, &flags
);
4081 * Account user cpu time to a process.
4082 * @p: the process that the cpu time gets accounted to
4083 * @cputime: the cpu time spent in user space since the last update
4085 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
4087 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4090 p
->utime
= cputime_add(p
->utime
, cputime
);
4092 /* Add user time to cpustat. */
4093 tmp
= cputime_to_cputime64(cputime
);
4094 if (TASK_NICE(p
) > 0)
4095 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
4097 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4101 * Account guest cpu time to a process.
4102 * @p: the process that the cpu time gets accounted to
4103 * @cputime: the cpu time spent in virtual machine since the last update
4105 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
)
4108 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4110 tmp
= cputime_to_cputime64(cputime
);
4112 p
->utime
= cputime_add(p
->utime
, cputime
);
4113 p
->gtime
= cputime_add(p
->gtime
, cputime
);
4115 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4116 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
4120 * Account scaled user cpu time to a process.
4121 * @p: the process that the cpu time gets accounted to
4122 * @cputime: the cpu time spent in user space since the last update
4124 void account_user_time_scaled(struct task_struct
*p
, cputime_t cputime
)
4126 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime
);
4130 * Account system cpu time to a process.
4131 * @p: the process that the cpu time gets accounted to
4132 * @hardirq_offset: the offset to subtract from hardirq_count()
4133 * @cputime: the cpu time spent in kernel space since the last update
4135 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
4138 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4139 struct rq
*rq
= this_rq();
4142 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
4143 account_guest_time(p
, cputime
);
4147 p
->stime
= cputime_add(p
->stime
, cputime
);
4149 /* Add system time to cpustat. */
4150 tmp
= cputime_to_cputime64(cputime
);
4151 if (hardirq_count() - hardirq_offset
)
4152 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
4153 else if (softirq_count())
4154 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
4155 else if (p
!= rq
->idle
)
4156 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
4157 else if (atomic_read(&rq
->nr_iowait
) > 0)
4158 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
4160 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
4161 /* Account for system time used */
4162 acct_update_integrals(p
);
4166 * Account scaled system cpu time to a process.
4167 * @p: the process that the cpu time gets accounted to
4168 * @hardirq_offset: the offset to subtract from hardirq_count()
4169 * @cputime: the cpu time spent in kernel space since the last update
4171 void account_system_time_scaled(struct task_struct
*p
, cputime_t cputime
)
4173 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime
);
4177 * Account for involuntary wait time.
4178 * @p: the process from which the cpu time has been stolen
4179 * @steal: the cpu time spent in involuntary wait
4181 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
4183 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4184 cputime64_t tmp
= cputime_to_cputime64(steal
);
4185 struct rq
*rq
= this_rq();
4187 if (p
== rq
->idle
) {
4188 p
->stime
= cputime_add(p
->stime
, steal
);
4189 if (atomic_read(&rq
->nr_iowait
) > 0)
4190 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
4192 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
4194 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
4198 * This function gets called by the timer code, with HZ frequency.
4199 * We call it with interrupts disabled.
4201 * It also gets called by the fork code, when changing the parent's
4204 void scheduler_tick(void)
4206 int cpu
= smp_processor_id();
4207 struct rq
*rq
= cpu_rq(cpu
);
4208 struct task_struct
*curr
= rq
->curr
;
4212 spin_lock(&rq
->lock
);
4213 update_rq_clock(rq
);
4214 update_cpu_load(rq
);
4215 curr
->sched_class
->task_tick(rq
, curr
, 0);
4216 spin_unlock(&rq
->lock
);
4219 rq
->idle_at_tick
= idle_cpu(cpu
);
4220 trigger_load_balance(rq
, cpu
);
4224 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
4226 void __kprobes
add_preempt_count(int val
)
4231 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4233 preempt_count() += val
;
4235 * Spinlock count overflowing soon?
4237 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4240 EXPORT_SYMBOL(add_preempt_count
);
4242 void __kprobes
sub_preempt_count(int val
)
4247 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
4250 * Is the spinlock portion underflowing?
4252 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4253 !(preempt_count() & PREEMPT_MASK
)))
4256 preempt_count() -= val
;
4258 EXPORT_SYMBOL(sub_preempt_count
);
4263 * Print scheduling while atomic bug:
4265 static noinline
void __schedule_bug(struct task_struct
*prev
)
4267 struct pt_regs
*regs
= get_irq_regs();
4269 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4270 prev
->comm
, prev
->pid
, preempt_count());
4272 debug_show_held_locks(prev
);
4274 if (irqs_disabled())
4275 print_irqtrace_events(prev
);
4284 * Various schedule()-time debugging checks and statistics:
4286 static inline void schedule_debug(struct task_struct
*prev
)
4289 * Test if we are atomic. Since do_exit() needs to call into
4290 * schedule() atomically, we ignore that path for now.
4291 * Otherwise, whine if we are scheduling when we should not be.
4293 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
4294 __schedule_bug(prev
);
4296 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4298 schedstat_inc(this_rq(), sched_count
);
4299 #ifdef CONFIG_SCHEDSTATS
4300 if (unlikely(prev
->lock_depth
>= 0)) {
4301 schedstat_inc(this_rq(), bkl_count
);
4302 schedstat_inc(prev
, sched_info
.bkl_count
);
4308 * Pick up the highest-prio task:
4310 static inline struct task_struct
*
4311 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
4313 const struct sched_class
*class;
4314 struct task_struct
*p
;
4317 * Optimization: we know that if all tasks are in
4318 * the fair class we can call that function directly:
4320 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
4321 p
= fair_sched_class
.pick_next_task(rq
);
4326 class = sched_class_highest
;
4328 p
= class->pick_next_task(rq
);
4332 * Will never be NULL as the idle class always
4333 * returns a non-NULL p:
4335 class = class->next
;
4340 * schedule() is the main scheduler function.
4342 asmlinkage
void __sched
schedule(void)
4344 struct task_struct
*prev
, *next
;
4345 unsigned long *switch_count
;
4347 int cpu
, hrtick
= sched_feat(HRTICK
);
4351 cpu
= smp_processor_id();
4355 switch_count
= &prev
->nivcsw
;
4357 release_kernel_lock(prev
);
4358 need_resched_nonpreemptible
:
4360 schedule_debug(prev
);
4366 * Do the rq-clock update outside the rq lock:
4368 local_irq_disable();
4369 update_rq_clock(rq
);
4370 spin_lock(&rq
->lock
);
4371 clear_tsk_need_resched(prev
);
4373 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
4374 if (unlikely(signal_pending_state(prev
->state
, prev
)))
4375 prev
->state
= TASK_RUNNING
;
4377 deactivate_task(rq
, prev
, 1);
4378 switch_count
= &prev
->nvcsw
;
4382 if (prev
->sched_class
->pre_schedule
)
4383 prev
->sched_class
->pre_schedule(rq
, prev
);
4386 if (unlikely(!rq
->nr_running
))
4387 idle_balance(cpu
, rq
);
4389 prev
->sched_class
->put_prev_task(rq
, prev
);
4390 next
= pick_next_task(rq
, prev
);
4392 if (likely(prev
!= next
)) {
4393 sched_info_switch(prev
, next
);
4399 context_switch(rq
, prev
, next
); /* unlocks the rq */
4401 * the context switch might have flipped the stack from under
4402 * us, hence refresh the local variables.
4404 cpu
= smp_processor_id();
4407 spin_unlock_irq(&rq
->lock
);
4412 if (unlikely(reacquire_kernel_lock(current
) < 0))
4413 goto need_resched_nonpreemptible
;
4415 preempt_enable_no_resched();
4416 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
4419 EXPORT_SYMBOL(schedule
);
4421 #ifdef CONFIG_PREEMPT
4423 * this is the entry point to schedule() from in-kernel preemption
4424 * off of preempt_enable. Kernel preemptions off return from interrupt
4425 * occur there and call schedule directly.
4427 asmlinkage
void __sched
preempt_schedule(void)
4429 struct thread_info
*ti
= current_thread_info();
4432 * If there is a non-zero preempt_count or interrupts are disabled,
4433 * we do not want to preempt the current task. Just return..
4435 if (likely(ti
->preempt_count
|| irqs_disabled()))
4439 add_preempt_count(PREEMPT_ACTIVE
);
4441 sub_preempt_count(PREEMPT_ACTIVE
);
4444 * Check again in case we missed a preemption opportunity
4445 * between schedule and now.
4448 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4450 EXPORT_SYMBOL(preempt_schedule
);
4453 * this is the entry point to schedule() from kernel preemption
4454 * off of irq context.
4455 * Note, that this is called and return with irqs disabled. This will
4456 * protect us against recursive calling from irq.
4458 asmlinkage
void __sched
preempt_schedule_irq(void)
4460 struct thread_info
*ti
= current_thread_info();
4462 /* Catch callers which need to be fixed */
4463 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
4466 add_preempt_count(PREEMPT_ACTIVE
);
4469 local_irq_disable();
4470 sub_preempt_count(PREEMPT_ACTIVE
);
4473 * Check again in case we missed a preemption opportunity
4474 * between schedule and now.
4477 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4480 #endif /* CONFIG_PREEMPT */
4482 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
4485 return try_to_wake_up(curr
->private, mode
, sync
);
4487 EXPORT_SYMBOL(default_wake_function
);
4490 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4491 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4492 * number) then we wake all the non-exclusive tasks and one exclusive task.
4494 * There are circumstances in which we can try to wake a task which has already
4495 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4496 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4498 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4499 int nr_exclusive
, int sync
, void *key
)
4501 wait_queue_t
*curr
, *next
;
4503 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4504 unsigned flags
= curr
->flags
;
4506 if (curr
->func(curr
, mode
, sync
, key
) &&
4507 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4513 * __wake_up - wake up threads blocked on a waitqueue.
4515 * @mode: which threads
4516 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4517 * @key: is directly passed to the wakeup function
4519 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4520 int nr_exclusive
, void *key
)
4522 unsigned long flags
;
4524 spin_lock_irqsave(&q
->lock
, flags
);
4525 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4526 spin_unlock_irqrestore(&q
->lock
, flags
);
4528 EXPORT_SYMBOL(__wake_up
);
4531 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4533 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4535 __wake_up_common(q
, mode
, 1, 0, NULL
);
4539 * __wake_up_sync - wake up threads blocked on a waitqueue.
4541 * @mode: which threads
4542 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4544 * The sync wakeup differs that the waker knows that it will schedule
4545 * away soon, so while the target thread will be woken up, it will not
4546 * be migrated to another CPU - ie. the two threads are 'synchronized'
4547 * with each other. This can prevent needless bouncing between CPUs.
4549 * On UP it can prevent extra preemption.
4552 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4554 unsigned long flags
;
4560 if (unlikely(!nr_exclusive
))
4563 spin_lock_irqsave(&q
->lock
, flags
);
4564 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
4565 spin_unlock_irqrestore(&q
->lock
, flags
);
4567 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4569 void complete(struct completion
*x
)
4571 unsigned long flags
;
4573 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4575 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4576 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4578 EXPORT_SYMBOL(complete
);
4580 void complete_all(struct completion
*x
)
4582 unsigned long flags
;
4584 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4585 x
->done
+= UINT_MAX
/2;
4586 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4587 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4589 EXPORT_SYMBOL(complete_all
);
4591 static inline long __sched
4592 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4595 DECLARE_WAITQUEUE(wait
, current
);
4597 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4598 __add_wait_queue_tail(&x
->wait
, &wait
);
4600 if ((state
== TASK_INTERRUPTIBLE
&&
4601 signal_pending(current
)) ||
4602 (state
== TASK_KILLABLE
&&
4603 fatal_signal_pending(current
))) {
4604 timeout
= -ERESTARTSYS
;
4607 __set_current_state(state
);
4608 spin_unlock_irq(&x
->wait
.lock
);
4609 timeout
= schedule_timeout(timeout
);
4610 spin_lock_irq(&x
->wait
.lock
);
4611 } while (!x
->done
&& timeout
);
4612 __remove_wait_queue(&x
->wait
, &wait
);
4617 return timeout
?: 1;
4621 wait_for_common(struct completion
*x
, long timeout
, int state
)
4625 spin_lock_irq(&x
->wait
.lock
);
4626 timeout
= do_wait_for_common(x
, timeout
, state
);
4627 spin_unlock_irq(&x
->wait
.lock
);
4631 void __sched
wait_for_completion(struct completion
*x
)
4633 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4635 EXPORT_SYMBOL(wait_for_completion
);
4637 unsigned long __sched
4638 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4640 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4642 EXPORT_SYMBOL(wait_for_completion_timeout
);
4644 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4646 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4647 if (t
== -ERESTARTSYS
)
4651 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4653 unsigned long __sched
4654 wait_for_completion_interruptible_timeout(struct completion
*x
,
4655 unsigned long timeout
)
4657 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4659 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4661 int __sched
wait_for_completion_killable(struct completion
*x
)
4663 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4664 if (t
== -ERESTARTSYS
)
4668 EXPORT_SYMBOL(wait_for_completion_killable
);
4671 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4673 unsigned long flags
;
4676 init_waitqueue_entry(&wait
, current
);
4678 __set_current_state(state
);
4680 spin_lock_irqsave(&q
->lock
, flags
);
4681 __add_wait_queue(q
, &wait
);
4682 spin_unlock(&q
->lock
);
4683 timeout
= schedule_timeout(timeout
);
4684 spin_lock_irq(&q
->lock
);
4685 __remove_wait_queue(q
, &wait
);
4686 spin_unlock_irqrestore(&q
->lock
, flags
);
4691 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4693 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4695 EXPORT_SYMBOL(interruptible_sleep_on
);
4698 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4700 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4702 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4704 void __sched
sleep_on(wait_queue_head_t
*q
)
4706 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4708 EXPORT_SYMBOL(sleep_on
);
4710 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4712 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4714 EXPORT_SYMBOL(sleep_on_timeout
);
4716 #ifdef CONFIG_RT_MUTEXES
4719 * rt_mutex_setprio - set the current priority of a task
4721 * @prio: prio value (kernel-internal form)
4723 * This function changes the 'effective' priority of a task. It does
4724 * not touch ->normal_prio like __setscheduler().
4726 * Used by the rt_mutex code to implement priority inheritance logic.
4728 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4730 unsigned long flags
;
4731 int oldprio
, on_rq
, running
;
4733 const struct sched_class
*prev_class
= p
->sched_class
;
4735 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4737 rq
= task_rq_lock(p
, &flags
);
4738 update_rq_clock(rq
);
4741 on_rq
= p
->se
.on_rq
;
4742 running
= task_current(rq
, p
);
4744 dequeue_task(rq
, p
, 0);
4746 p
->sched_class
->put_prev_task(rq
, p
);
4749 p
->sched_class
= &rt_sched_class
;
4751 p
->sched_class
= &fair_sched_class
;
4756 p
->sched_class
->set_curr_task(rq
);
4758 enqueue_task(rq
, p
, 0);
4760 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4762 task_rq_unlock(rq
, &flags
);
4767 void set_user_nice(struct task_struct
*p
, long nice
)
4769 int old_prio
, delta
, on_rq
;
4770 unsigned long flags
;
4773 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4776 * We have to be careful, if called from sys_setpriority(),
4777 * the task might be in the middle of scheduling on another CPU.
4779 rq
= task_rq_lock(p
, &flags
);
4780 update_rq_clock(rq
);
4782 * The RT priorities are set via sched_setscheduler(), but we still
4783 * allow the 'normal' nice value to be set - but as expected
4784 * it wont have any effect on scheduling until the task is
4785 * SCHED_FIFO/SCHED_RR:
4787 if (task_has_rt_policy(p
)) {
4788 p
->static_prio
= NICE_TO_PRIO(nice
);
4791 on_rq
= p
->se
.on_rq
;
4793 dequeue_task(rq
, p
, 0);
4795 p
->static_prio
= NICE_TO_PRIO(nice
);
4798 p
->prio
= effective_prio(p
);
4799 delta
= p
->prio
- old_prio
;
4802 enqueue_task(rq
, p
, 0);
4804 * If the task increased its priority or is running and
4805 * lowered its priority, then reschedule its CPU:
4807 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4808 resched_task(rq
->curr
);
4811 task_rq_unlock(rq
, &flags
);
4813 EXPORT_SYMBOL(set_user_nice
);
4816 * can_nice - check if a task can reduce its nice value
4820 int can_nice(const struct task_struct
*p
, const int nice
)
4822 /* convert nice value [19,-20] to rlimit style value [1,40] */
4823 int nice_rlim
= 20 - nice
;
4825 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4826 capable(CAP_SYS_NICE
));
4829 #ifdef __ARCH_WANT_SYS_NICE
4832 * sys_nice - change the priority of the current process.
4833 * @increment: priority increment
4835 * sys_setpriority is a more generic, but much slower function that
4836 * does similar things.
4838 asmlinkage
long sys_nice(int increment
)
4843 * Setpriority might change our priority at the same moment.
4844 * We don't have to worry. Conceptually one call occurs first
4845 * and we have a single winner.
4847 if (increment
< -40)
4852 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4858 if (increment
< 0 && !can_nice(current
, nice
))
4861 retval
= security_task_setnice(current
, nice
);
4865 set_user_nice(current
, nice
);
4872 * task_prio - return the priority value of a given task.
4873 * @p: the task in question.
4875 * This is the priority value as seen by users in /proc.
4876 * RT tasks are offset by -200. Normal tasks are centered
4877 * around 0, value goes from -16 to +15.
4879 int task_prio(const struct task_struct
*p
)
4881 return p
->prio
- MAX_RT_PRIO
;
4885 * task_nice - return the nice value of a given task.
4886 * @p: the task in question.
4888 int task_nice(const struct task_struct
*p
)
4890 return TASK_NICE(p
);
4892 EXPORT_SYMBOL(task_nice
);
4895 * idle_cpu - is a given cpu idle currently?
4896 * @cpu: the processor in question.
4898 int idle_cpu(int cpu
)
4900 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4904 * idle_task - return the idle task for a given cpu.
4905 * @cpu: the processor in question.
4907 struct task_struct
*idle_task(int cpu
)
4909 return cpu_rq(cpu
)->idle
;
4913 * find_process_by_pid - find a process with a matching PID value.
4914 * @pid: the pid in question.
4916 static struct task_struct
*find_process_by_pid(pid_t pid
)
4918 return pid
? find_task_by_vpid(pid
) : current
;
4921 /* Actually do priority change: must hold rq lock. */
4923 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4925 BUG_ON(p
->se
.on_rq
);
4928 switch (p
->policy
) {
4932 p
->sched_class
= &fair_sched_class
;
4936 p
->sched_class
= &rt_sched_class
;
4940 p
->rt_priority
= prio
;
4941 p
->normal_prio
= normal_prio(p
);
4942 /* we are holding p->pi_lock already */
4943 p
->prio
= rt_mutex_getprio(p
);
4948 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4949 * @p: the task in question.
4950 * @policy: new policy.
4951 * @param: structure containing the new RT priority.
4953 * NOTE that the task may be already dead.
4955 int sched_setscheduler(struct task_struct
*p
, int policy
,
4956 struct sched_param
*param
)
4958 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4959 unsigned long flags
;
4960 const struct sched_class
*prev_class
= p
->sched_class
;
4963 /* may grab non-irq protected spin_locks */
4964 BUG_ON(in_interrupt());
4966 /* double check policy once rq lock held */
4968 policy
= oldpolicy
= p
->policy
;
4969 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4970 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4971 policy
!= SCHED_IDLE
)
4974 * Valid priorities for SCHED_FIFO and SCHED_RR are
4975 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4976 * SCHED_BATCH and SCHED_IDLE is 0.
4978 if (param
->sched_priority
< 0 ||
4979 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4980 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4982 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4986 * Allow unprivileged RT tasks to decrease priority:
4988 if (!capable(CAP_SYS_NICE
)) {
4989 if (rt_policy(policy
)) {
4990 unsigned long rlim_rtprio
;
4992 if (!lock_task_sighand(p
, &flags
))
4994 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4995 unlock_task_sighand(p
, &flags
);
4997 /* can't set/change the rt policy */
4998 if (policy
!= p
->policy
&& !rlim_rtprio
)
5001 /* can't increase priority */
5002 if (param
->sched_priority
> p
->rt_priority
&&
5003 param
->sched_priority
> rlim_rtprio
)
5007 * Like positive nice levels, dont allow tasks to
5008 * move out of SCHED_IDLE either:
5010 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
5013 /* can't change other user's priorities */
5014 if ((current
->euid
!= p
->euid
) &&
5015 (current
->euid
!= p
->uid
))
5019 #ifdef CONFIG_RT_GROUP_SCHED
5021 * Do not allow realtime tasks into groups that have no runtime
5024 if (rt_policy(policy
) && task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
5028 retval
= security_task_setscheduler(p
, policy
, param
);
5032 * make sure no PI-waiters arrive (or leave) while we are
5033 * changing the priority of the task:
5035 spin_lock_irqsave(&p
->pi_lock
, flags
);
5037 * To be able to change p->policy safely, the apropriate
5038 * runqueue lock must be held.
5040 rq
= __task_rq_lock(p
);
5041 /* recheck policy now with rq lock held */
5042 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
5043 policy
= oldpolicy
= -1;
5044 __task_rq_unlock(rq
);
5045 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5048 update_rq_clock(rq
);
5049 on_rq
= p
->se
.on_rq
;
5050 running
= task_current(rq
, p
);
5052 deactivate_task(rq
, p
, 0);
5054 p
->sched_class
->put_prev_task(rq
, p
);
5057 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
5060 p
->sched_class
->set_curr_task(rq
);
5062 activate_task(rq
, p
, 0);
5064 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5066 __task_rq_unlock(rq
);
5067 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5069 rt_mutex_adjust_pi(p
);
5073 EXPORT_SYMBOL_GPL(sched_setscheduler
);
5076 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5078 struct sched_param lparam
;
5079 struct task_struct
*p
;
5082 if (!param
|| pid
< 0)
5084 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
5089 p
= find_process_by_pid(pid
);
5091 retval
= sched_setscheduler(p
, policy
, &lparam
);
5098 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5099 * @pid: the pid in question.
5100 * @policy: new policy.
5101 * @param: structure containing the new RT priority.
5104 sys_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5106 /* negative values for policy are not valid */
5110 return do_sched_setscheduler(pid
, policy
, param
);
5114 * sys_sched_setparam - set/change the RT priority of a thread
5115 * @pid: the pid in question.
5116 * @param: structure containing the new RT priority.
5118 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
5120 return do_sched_setscheduler(pid
, -1, param
);
5124 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5125 * @pid: the pid in question.
5127 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
5129 struct task_struct
*p
;
5136 read_lock(&tasklist_lock
);
5137 p
= find_process_by_pid(pid
);
5139 retval
= security_task_getscheduler(p
);
5143 read_unlock(&tasklist_lock
);
5148 * sys_sched_getscheduler - get the RT priority of a thread
5149 * @pid: the pid in question.
5150 * @param: structure containing the RT priority.
5152 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
5154 struct sched_param lp
;
5155 struct task_struct
*p
;
5158 if (!param
|| pid
< 0)
5161 read_lock(&tasklist_lock
);
5162 p
= find_process_by_pid(pid
);
5167 retval
= security_task_getscheduler(p
);
5171 lp
.sched_priority
= p
->rt_priority
;
5172 read_unlock(&tasklist_lock
);
5175 * This one might sleep, we cannot do it with a spinlock held ...
5177 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
5182 read_unlock(&tasklist_lock
);
5186 long sched_setaffinity(pid_t pid
, const cpumask_t
*in_mask
)
5188 cpumask_t cpus_allowed
;
5189 cpumask_t new_mask
= *in_mask
;
5190 struct task_struct
*p
;
5194 read_lock(&tasklist_lock
);
5196 p
= find_process_by_pid(pid
);
5198 read_unlock(&tasklist_lock
);
5204 * It is not safe to call set_cpus_allowed with the
5205 * tasklist_lock held. We will bump the task_struct's
5206 * usage count and then drop tasklist_lock.
5209 read_unlock(&tasklist_lock
);
5212 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
5213 !capable(CAP_SYS_NICE
))
5216 retval
= security_task_setscheduler(p
, 0, NULL
);
5220 cpuset_cpus_allowed(p
, &cpus_allowed
);
5221 cpus_and(new_mask
, new_mask
, cpus_allowed
);
5223 retval
= set_cpus_allowed_ptr(p
, &new_mask
);
5226 cpuset_cpus_allowed(p
, &cpus_allowed
);
5227 if (!cpus_subset(new_mask
, cpus_allowed
)) {
5229 * We must have raced with a concurrent cpuset
5230 * update. Just reset the cpus_allowed to the
5231 * cpuset's cpus_allowed
5233 new_mask
= cpus_allowed
;
5243 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5244 cpumask_t
*new_mask
)
5246 if (len
< sizeof(cpumask_t
)) {
5247 memset(new_mask
, 0, sizeof(cpumask_t
));
5248 } else if (len
> sizeof(cpumask_t
)) {
5249 len
= sizeof(cpumask_t
);
5251 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
5255 * sys_sched_setaffinity - set the cpu affinity of a process
5256 * @pid: pid of the process
5257 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5258 * @user_mask_ptr: user-space pointer to the new cpu mask
5260 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
5261 unsigned long __user
*user_mask_ptr
)
5266 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
5270 return sched_setaffinity(pid
, &new_mask
);
5273 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
5275 struct task_struct
*p
;
5279 read_lock(&tasklist_lock
);
5282 p
= find_process_by_pid(pid
);
5286 retval
= security_task_getscheduler(p
);
5290 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
5293 read_unlock(&tasklist_lock
);
5300 * sys_sched_getaffinity - get the cpu affinity of a process
5301 * @pid: pid of the process
5302 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5303 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5305 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
5306 unsigned long __user
*user_mask_ptr
)
5311 if (len
< sizeof(cpumask_t
))
5314 ret
= sched_getaffinity(pid
, &mask
);
5318 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
5321 return sizeof(cpumask_t
);
5325 * sys_sched_yield - yield the current processor to other threads.
5327 * This function yields the current CPU to other tasks. If there are no
5328 * other threads running on this CPU then this function will return.
5330 asmlinkage
long sys_sched_yield(void)
5332 struct rq
*rq
= this_rq_lock();
5334 schedstat_inc(rq
, yld_count
);
5335 current
->sched_class
->yield_task(rq
);
5338 * Since we are going to call schedule() anyway, there's
5339 * no need to preempt or enable interrupts:
5341 __release(rq
->lock
);
5342 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5343 _raw_spin_unlock(&rq
->lock
);
5344 preempt_enable_no_resched();
5351 static void __cond_resched(void)
5353 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5354 __might_sleep(__FILE__
, __LINE__
);
5357 * The BKS might be reacquired before we have dropped
5358 * PREEMPT_ACTIVE, which could trigger a second
5359 * cond_resched() call.
5362 add_preempt_count(PREEMPT_ACTIVE
);
5364 sub_preempt_count(PREEMPT_ACTIVE
);
5365 } while (need_resched());
5368 int __sched
_cond_resched(void)
5370 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
5371 system_state
== SYSTEM_RUNNING
) {
5377 EXPORT_SYMBOL(_cond_resched
);
5380 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5381 * call schedule, and on return reacquire the lock.
5383 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5384 * operations here to prevent schedule() from being called twice (once via
5385 * spin_unlock(), once by hand).
5387 int cond_resched_lock(spinlock_t
*lock
)
5389 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
5392 if (spin_needbreak(lock
) || resched
) {
5394 if (resched
&& need_resched())
5403 EXPORT_SYMBOL(cond_resched_lock
);
5405 int __sched
cond_resched_softirq(void)
5407 BUG_ON(!in_softirq());
5409 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
5417 EXPORT_SYMBOL(cond_resched_softirq
);
5420 * yield - yield the current processor to other threads.
5422 * This is a shortcut for kernel-space yielding - it marks the
5423 * thread runnable and calls sys_sched_yield().
5425 void __sched
yield(void)
5427 set_current_state(TASK_RUNNING
);
5430 EXPORT_SYMBOL(yield
);
5433 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5434 * that process accounting knows that this is a task in IO wait state.
5436 * But don't do that if it is a deliberate, throttling IO wait (this task
5437 * has set its backing_dev_info: the queue against which it should throttle)
5439 void __sched
io_schedule(void)
5441 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5443 delayacct_blkio_start();
5444 atomic_inc(&rq
->nr_iowait
);
5446 atomic_dec(&rq
->nr_iowait
);
5447 delayacct_blkio_end();
5449 EXPORT_SYMBOL(io_schedule
);
5451 long __sched
io_schedule_timeout(long timeout
)
5453 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5456 delayacct_blkio_start();
5457 atomic_inc(&rq
->nr_iowait
);
5458 ret
= schedule_timeout(timeout
);
5459 atomic_dec(&rq
->nr_iowait
);
5460 delayacct_blkio_end();
5465 * sys_sched_get_priority_max - return maximum RT priority.
5466 * @policy: scheduling class.
5468 * this syscall returns the maximum rt_priority that can be used
5469 * by a given scheduling class.
5471 asmlinkage
long sys_sched_get_priority_max(int policy
)
5478 ret
= MAX_USER_RT_PRIO
-1;
5490 * sys_sched_get_priority_min - return minimum RT priority.
5491 * @policy: scheduling class.
5493 * this syscall returns the minimum rt_priority that can be used
5494 * by a given scheduling class.
5496 asmlinkage
long sys_sched_get_priority_min(int policy
)
5514 * sys_sched_rr_get_interval - return the default timeslice of a process.
5515 * @pid: pid of the process.
5516 * @interval: userspace pointer to the timeslice value.
5518 * this syscall writes the default timeslice value of a given process
5519 * into the user-space timespec buffer. A value of '0' means infinity.
5522 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
5524 struct task_struct
*p
;
5525 unsigned int time_slice
;
5533 read_lock(&tasklist_lock
);
5534 p
= find_process_by_pid(pid
);
5538 retval
= security_task_getscheduler(p
);
5543 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5544 * tasks that are on an otherwise idle runqueue:
5547 if (p
->policy
== SCHED_RR
) {
5548 time_slice
= DEF_TIMESLICE
;
5549 } else if (p
->policy
!= SCHED_FIFO
) {
5550 struct sched_entity
*se
= &p
->se
;
5551 unsigned long flags
;
5554 rq
= task_rq_lock(p
, &flags
);
5555 if (rq
->cfs
.load
.weight
)
5556 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
5557 task_rq_unlock(rq
, &flags
);
5559 read_unlock(&tasklist_lock
);
5560 jiffies_to_timespec(time_slice
, &t
);
5561 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5565 read_unlock(&tasklist_lock
);
5569 static const char stat_nam
[] = "RSDTtZX";
5571 void sched_show_task(struct task_struct
*p
)
5573 unsigned long free
= 0;
5576 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5577 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5578 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5579 #if BITS_PER_LONG == 32
5580 if (state
== TASK_RUNNING
)
5581 printk(KERN_CONT
" running ");
5583 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5585 if (state
== TASK_RUNNING
)
5586 printk(KERN_CONT
" running task ");
5588 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5590 #ifdef CONFIG_DEBUG_STACK_USAGE
5592 unsigned long *n
= end_of_stack(p
);
5595 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
5598 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
5599 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
5601 show_stack(p
, NULL
);
5604 void show_state_filter(unsigned long state_filter
)
5606 struct task_struct
*g
, *p
;
5608 #if BITS_PER_LONG == 32
5610 " task PC stack pid father\n");
5613 " task PC stack pid father\n");
5615 read_lock(&tasklist_lock
);
5616 do_each_thread(g
, p
) {
5618 * reset the NMI-timeout, listing all files on a slow
5619 * console might take alot of time:
5621 touch_nmi_watchdog();
5622 if (!state_filter
|| (p
->state
& state_filter
))
5624 } while_each_thread(g
, p
);
5626 touch_all_softlockup_watchdogs();
5628 #ifdef CONFIG_SCHED_DEBUG
5629 sysrq_sched_debug_show();
5631 read_unlock(&tasklist_lock
);
5633 * Only show locks if all tasks are dumped:
5635 if (state_filter
== -1)
5636 debug_show_all_locks();
5639 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5641 idle
->sched_class
= &idle_sched_class
;
5645 * init_idle - set up an idle thread for a given CPU
5646 * @idle: task in question
5647 * @cpu: cpu the idle task belongs to
5649 * NOTE: this function does not set the idle thread's NEED_RESCHED
5650 * flag, to make booting more robust.
5652 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5654 struct rq
*rq
= cpu_rq(cpu
);
5655 unsigned long flags
;
5658 idle
->se
.exec_start
= sched_clock();
5660 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
5661 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
5662 __set_task_cpu(idle
, cpu
);
5664 spin_lock_irqsave(&rq
->lock
, flags
);
5665 rq
->curr
= rq
->idle
= idle
;
5666 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5669 spin_unlock_irqrestore(&rq
->lock
, flags
);
5671 /* Set the preempt count _outside_ the spinlocks! */
5672 #if defined(CONFIG_PREEMPT)
5673 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5675 task_thread_info(idle
)->preempt_count
= 0;
5678 * The idle tasks have their own, simple scheduling class:
5680 idle
->sched_class
= &idle_sched_class
;
5684 * In a system that switches off the HZ timer nohz_cpu_mask
5685 * indicates which cpus entered this state. This is used
5686 * in the rcu update to wait only for active cpus. For system
5687 * which do not switch off the HZ timer nohz_cpu_mask should
5688 * always be CPU_MASK_NONE.
5690 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
5693 * Increase the granularity value when there are more CPUs,
5694 * because with more CPUs the 'effective latency' as visible
5695 * to users decreases. But the relationship is not linear,
5696 * so pick a second-best guess by going with the log2 of the
5699 * This idea comes from the SD scheduler of Con Kolivas:
5701 static inline void sched_init_granularity(void)
5703 unsigned int factor
= 1 + ilog2(num_online_cpus());
5704 const unsigned long limit
= 200000000;
5706 sysctl_sched_min_granularity
*= factor
;
5707 if (sysctl_sched_min_granularity
> limit
)
5708 sysctl_sched_min_granularity
= limit
;
5710 sysctl_sched_latency
*= factor
;
5711 if (sysctl_sched_latency
> limit
)
5712 sysctl_sched_latency
= limit
;
5714 sysctl_sched_wakeup_granularity
*= factor
;
5719 * This is how migration works:
5721 * 1) we queue a struct migration_req structure in the source CPU's
5722 * runqueue and wake up that CPU's migration thread.
5723 * 2) we down() the locked semaphore => thread blocks.
5724 * 3) migration thread wakes up (implicitly it forces the migrated
5725 * thread off the CPU)
5726 * 4) it gets the migration request and checks whether the migrated
5727 * task is still in the wrong runqueue.
5728 * 5) if it's in the wrong runqueue then the migration thread removes
5729 * it and puts it into the right queue.
5730 * 6) migration thread up()s the semaphore.
5731 * 7) we wake up and the migration is done.
5735 * Change a given task's CPU affinity. Migrate the thread to a
5736 * proper CPU and schedule it away if the CPU it's executing on
5737 * is removed from the allowed bitmask.
5739 * NOTE: the caller must have a valid reference to the task, the
5740 * task must not exit() & deallocate itself prematurely. The
5741 * call is not atomic; no spinlocks may be held.
5743 int set_cpus_allowed_ptr(struct task_struct
*p
, const cpumask_t
*new_mask
)
5745 struct migration_req req
;
5746 unsigned long flags
;
5750 rq
= task_rq_lock(p
, &flags
);
5751 if (!cpus_intersects(*new_mask
, cpu_online_map
)) {
5756 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
5757 !cpus_equal(p
->cpus_allowed
, *new_mask
))) {
5762 if (p
->sched_class
->set_cpus_allowed
)
5763 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
5765 p
->cpus_allowed
= *new_mask
;
5766 p
->rt
.nr_cpus_allowed
= cpus_weight(*new_mask
);
5769 /* Can the task run on the task's current CPU? If so, we're done */
5770 if (cpu_isset(task_cpu(p
), *new_mask
))
5773 if (migrate_task(p
, any_online_cpu(*new_mask
), &req
)) {
5774 /* Need help from migration thread: drop lock and wait. */
5775 task_rq_unlock(rq
, &flags
);
5776 wake_up_process(rq
->migration_thread
);
5777 wait_for_completion(&req
.done
);
5778 tlb_migrate_finish(p
->mm
);
5782 task_rq_unlock(rq
, &flags
);
5786 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5789 * Move (not current) task off this cpu, onto dest cpu. We're doing
5790 * this because either it can't run here any more (set_cpus_allowed()
5791 * away from this CPU, or CPU going down), or because we're
5792 * attempting to rebalance this task on exec (sched_exec).
5794 * So we race with normal scheduler movements, but that's OK, as long
5795 * as the task is no longer on this CPU.
5797 * Returns non-zero if task was successfully migrated.
5799 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5801 struct rq
*rq_dest
, *rq_src
;
5804 if (unlikely(cpu_is_offline(dest_cpu
)))
5807 rq_src
= cpu_rq(src_cpu
);
5808 rq_dest
= cpu_rq(dest_cpu
);
5810 double_rq_lock(rq_src
, rq_dest
);
5811 /* Already moved. */
5812 if (task_cpu(p
) != src_cpu
)
5814 /* Affinity changed (again). */
5815 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
5818 on_rq
= p
->se
.on_rq
;
5820 deactivate_task(rq_src
, p
, 0);
5822 set_task_cpu(p
, dest_cpu
);
5824 activate_task(rq_dest
, p
, 0);
5825 check_preempt_curr(rq_dest
, p
);
5829 double_rq_unlock(rq_src
, rq_dest
);
5834 * migration_thread - this is a highprio system thread that performs
5835 * thread migration by bumping thread off CPU then 'pushing' onto
5838 static int migration_thread(void *data
)
5840 int cpu
= (long)data
;
5844 BUG_ON(rq
->migration_thread
!= current
);
5846 set_current_state(TASK_INTERRUPTIBLE
);
5847 while (!kthread_should_stop()) {
5848 struct migration_req
*req
;
5849 struct list_head
*head
;
5851 spin_lock_irq(&rq
->lock
);
5853 if (cpu_is_offline(cpu
)) {
5854 spin_unlock_irq(&rq
->lock
);
5858 if (rq
->active_balance
) {
5859 active_load_balance(rq
, cpu
);
5860 rq
->active_balance
= 0;
5863 head
= &rq
->migration_queue
;
5865 if (list_empty(head
)) {
5866 spin_unlock_irq(&rq
->lock
);
5868 set_current_state(TASK_INTERRUPTIBLE
);
5871 req
= list_entry(head
->next
, struct migration_req
, list
);
5872 list_del_init(head
->next
);
5874 spin_unlock(&rq
->lock
);
5875 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5878 complete(&req
->done
);
5880 __set_current_state(TASK_RUNNING
);
5884 /* Wait for kthread_stop */
5885 set_current_state(TASK_INTERRUPTIBLE
);
5886 while (!kthread_should_stop()) {
5888 set_current_state(TASK_INTERRUPTIBLE
);
5890 __set_current_state(TASK_RUNNING
);
5894 #ifdef CONFIG_HOTPLUG_CPU
5896 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5900 local_irq_disable();
5901 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
5907 * Figure out where task on dead CPU should go, use force if necessary.
5908 * NOTE: interrupts should be disabled by the caller
5910 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5912 unsigned long flags
;
5919 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5920 cpus_and(mask
, mask
, p
->cpus_allowed
);
5921 dest_cpu
= any_online_cpu(mask
);
5923 /* On any allowed CPU? */
5924 if (dest_cpu
>= nr_cpu_ids
)
5925 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5927 /* No more Mr. Nice Guy. */
5928 if (dest_cpu
>= nr_cpu_ids
) {
5929 cpumask_t cpus_allowed
;
5931 cpuset_cpus_allowed_locked(p
, &cpus_allowed
);
5933 * Try to stay on the same cpuset, where the
5934 * current cpuset may be a subset of all cpus.
5935 * The cpuset_cpus_allowed_locked() variant of
5936 * cpuset_cpus_allowed() will not block. It must be
5937 * called within calls to cpuset_lock/cpuset_unlock.
5939 rq
= task_rq_lock(p
, &flags
);
5940 p
->cpus_allowed
= cpus_allowed
;
5941 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5942 task_rq_unlock(rq
, &flags
);
5945 * Don't tell them about moving exiting tasks or
5946 * kernel threads (both mm NULL), since they never
5949 if (p
->mm
&& printk_ratelimit()) {
5950 printk(KERN_INFO
"process %d (%s) no "
5951 "longer affine to cpu%d\n",
5952 task_pid_nr(p
), p
->comm
, dead_cpu
);
5955 } while (!__migrate_task_irq(p
, dead_cpu
, dest_cpu
));
5959 * While a dead CPU has no uninterruptible tasks queued at this point,
5960 * it might still have a nonzero ->nr_uninterruptible counter, because
5961 * for performance reasons the counter is not stricly tracking tasks to
5962 * their home CPUs. So we just add the counter to another CPU's counter,
5963 * to keep the global sum constant after CPU-down:
5965 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5967 struct rq
*rq_dest
= cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR
));
5968 unsigned long flags
;
5970 local_irq_save(flags
);
5971 double_rq_lock(rq_src
, rq_dest
);
5972 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5973 rq_src
->nr_uninterruptible
= 0;
5974 double_rq_unlock(rq_src
, rq_dest
);
5975 local_irq_restore(flags
);
5978 /* Run through task list and migrate tasks from the dead cpu. */
5979 static void migrate_live_tasks(int src_cpu
)
5981 struct task_struct
*p
, *t
;
5983 read_lock(&tasklist_lock
);
5985 do_each_thread(t
, p
) {
5989 if (task_cpu(p
) == src_cpu
)
5990 move_task_off_dead_cpu(src_cpu
, p
);
5991 } while_each_thread(t
, p
);
5993 read_unlock(&tasklist_lock
);
5997 * Schedules idle task to be the next runnable task on current CPU.
5998 * It does so by boosting its priority to highest possible.
5999 * Used by CPU offline code.
6001 void sched_idle_next(void)
6003 int this_cpu
= smp_processor_id();
6004 struct rq
*rq
= cpu_rq(this_cpu
);
6005 struct task_struct
*p
= rq
->idle
;
6006 unsigned long flags
;
6008 /* cpu has to be offline */
6009 BUG_ON(cpu_online(this_cpu
));
6012 * Strictly not necessary since rest of the CPUs are stopped by now
6013 * and interrupts disabled on the current cpu.
6015 spin_lock_irqsave(&rq
->lock
, flags
);
6017 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6019 update_rq_clock(rq
);
6020 activate_task(rq
, p
, 0);
6022 spin_unlock_irqrestore(&rq
->lock
, flags
);
6026 * Ensures that the idle task is using init_mm right before its cpu goes
6029 void idle_task_exit(void)
6031 struct mm_struct
*mm
= current
->active_mm
;
6033 BUG_ON(cpu_online(smp_processor_id()));
6036 switch_mm(mm
, &init_mm
, current
);
6040 /* called under rq->lock with disabled interrupts */
6041 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
6043 struct rq
*rq
= cpu_rq(dead_cpu
);
6045 /* Must be exiting, otherwise would be on tasklist. */
6046 BUG_ON(!p
->exit_state
);
6048 /* Cannot have done final schedule yet: would have vanished. */
6049 BUG_ON(p
->state
== TASK_DEAD
);
6054 * Drop lock around migration; if someone else moves it,
6055 * that's OK. No task can be added to this CPU, so iteration is
6058 spin_unlock_irq(&rq
->lock
);
6059 move_task_off_dead_cpu(dead_cpu
, p
);
6060 spin_lock_irq(&rq
->lock
);
6065 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6066 static void migrate_dead_tasks(unsigned int dead_cpu
)
6068 struct rq
*rq
= cpu_rq(dead_cpu
);
6069 struct task_struct
*next
;
6072 if (!rq
->nr_running
)
6074 update_rq_clock(rq
);
6075 next
= pick_next_task(rq
, rq
->curr
);
6078 next
->sched_class
->put_prev_task(rq
, next
);
6079 migrate_dead(dead_cpu
, next
);
6083 #endif /* CONFIG_HOTPLUG_CPU */
6085 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6087 static struct ctl_table sd_ctl_dir
[] = {
6089 .procname
= "sched_domain",
6095 static struct ctl_table sd_ctl_root
[] = {
6097 .ctl_name
= CTL_KERN
,
6098 .procname
= "kernel",
6100 .child
= sd_ctl_dir
,
6105 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
6107 struct ctl_table
*entry
=
6108 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
6113 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
6115 struct ctl_table
*entry
;
6118 * In the intermediate directories, both the child directory and
6119 * procname are dynamically allocated and could fail but the mode
6120 * will always be set. In the lowest directory the names are
6121 * static strings and all have proc handlers.
6123 for (entry
= *tablep
; entry
->mode
; entry
++) {
6125 sd_free_ctl_entry(&entry
->child
);
6126 if (entry
->proc_handler
== NULL
)
6127 kfree(entry
->procname
);
6135 set_table_entry(struct ctl_table
*entry
,
6136 const char *procname
, void *data
, int maxlen
,
6137 mode_t mode
, proc_handler
*proc_handler
)
6139 entry
->procname
= procname
;
6141 entry
->maxlen
= maxlen
;
6143 entry
->proc_handler
= proc_handler
;
6146 static struct ctl_table
*
6147 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
6149 struct ctl_table
*table
= sd_alloc_ctl_entry(12);
6154 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
6155 sizeof(long), 0644, proc_doulongvec_minmax
);
6156 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
6157 sizeof(long), 0644, proc_doulongvec_minmax
);
6158 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
6159 sizeof(int), 0644, proc_dointvec_minmax
);
6160 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
6161 sizeof(int), 0644, proc_dointvec_minmax
);
6162 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
6163 sizeof(int), 0644, proc_dointvec_minmax
);
6164 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
6165 sizeof(int), 0644, proc_dointvec_minmax
);
6166 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
6167 sizeof(int), 0644, proc_dointvec_minmax
);
6168 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
6169 sizeof(int), 0644, proc_dointvec_minmax
);
6170 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
6171 sizeof(int), 0644, proc_dointvec_minmax
);
6172 set_table_entry(&table
[9], "cache_nice_tries",
6173 &sd
->cache_nice_tries
,
6174 sizeof(int), 0644, proc_dointvec_minmax
);
6175 set_table_entry(&table
[10], "flags", &sd
->flags
,
6176 sizeof(int), 0644, proc_dointvec_minmax
);
6177 /* &table[11] is terminator */
6182 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
6184 struct ctl_table
*entry
, *table
;
6185 struct sched_domain
*sd
;
6186 int domain_num
= 0, i
;
6189 for_each_domain(cpu
, sd
)
6191 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
6196 for_each_domain(cpu
, sd
) {
6197 snprintf(buf
, 32, "domain%d", i
);
6198 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6200 entry
->child
= sd_alloc_ctl_domain_table(sd
);
6207 static struct ctl_table_header
*sd_sysctl_header
;
6208 static void register_sched_domain_sysctl(void)
6210 int i
, cpu_num
= num_online_cpus();
6211 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
6214 WARN_ON(sd_ctl_dir
[0].child
);
6215 sd_ctl_dir
[0].child
= entry
;
6220 for_each_online_cpu(i
) {
6221 snprintf(buf
, 32, "cpu%d", i
);
6222 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6224 entry
->child
= sd_alloc_ctl_cpu_table(i
);
6228 WARN_ON(sd_sysctl_header
);
6229 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
6232 /* may be called multiple times per register */
6233 static void unregister_sched_domain_sysctl(void)
6235 if (sd_sysctl_header
)
6236 unregister_sysctl_table(sd_sysctl_header
);
6237 sd_sysctl_header
= NULL
;
6238 if (sd_ctl_dir
[0].child
)
6239 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
6242 static void register_sched_domain_sysctl(void)
6245 static void unregister_sched_domain_sysctl(void)
6250 static void set_rq_online(struct rq
*rq
)
6253 const struct sched_class
*class;
6255 cpu_set(rq
->cpu
, rq
->rd
->online
);
6258 for_each_class(class) {
6259 if (class->rq_online
)
6260 class->rq_online(rq
);
6265 static void set_rq_offline(struct rq
*rq
)
6268 const struct sched_class
*class;
6270 for_each_class(class) {
6271 if (class->rq_offline
)
6272 class->rq_offline(rq
);
6275 cpu_clear(rq
->cpu
, rq
->rd
->online
);
6281 * migration_call - callback that gets triggered when a CPU is added.
6282 * Here we can start up the necessary migration thread for the new CPU.
6284 static int __cpuinit
6285 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
6287 struct task_struct
*p
;
6288 int cpu
= (long)hcpu
;
6289 unsigned long flags
;
6294 case CPU_UP_PREPARE
:
6295 case CPU_UP_PREPARE_FROZEN
:
6296 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
6299 kthread_bind(p
, cpu
);
6300 /* Must be high prio: stop_machine expects to yield to it. */
6301 rq
= task_rq_lock(p
, &flags
);
6302 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6303 task_rq_unlock(rq
, &flags
);
6304 cpu_rq(cpu
)->migration_thread
= p
;
6308 case CPU_ONLINE_FROZEN
:
6309 /* Strictly unnecessary, as first user will wake it. */
6310 wake_up_process(cpu_rq(cpu
)->migration_thread
);
6312 /* Update our root-domain */
6314 spin_lock_irqsave(&rq
->lock
, flags
);
6316 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6320 spin_unlock_irqrestore(&rq
->lock
, flags
);
6323 #ifdef CONFIG_HOTPLUG_CPU
6324 case CPU_UP_CANCELED
:
6325 case CPU_UP_CANCELED_FROZEN
:
6326 if (!cpu_rq(cpu
)->migration_thread
)
6328 /* Unbind it from offline cpu so it can run. Fall thru. */
6329 kthread_bind(cpu_rq(cpu
)->migration_thread
,
6330 any_online_cpu(cpu_online_map
));
6331 kthread_stop(cpu_rq(cpu
)->migration_thread
);
6332 cpu_rq(cpu
)->migration_thread
= NULL
;
6336 case CPU_DEAD_FROZEN
:
6337 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6338 migrate_live_tasks(cpu
);
6340 kthread_stop(rq
->migration_thread
);
6341 rq
->migration_thread
= NULL
;
6342 /* Idle task back to normal (off runqueue, low prio) */
6343 spin_lock_irq(&rq
->lock
);
6344 update_rq_clock(rq
);
6345 deactivate_task(rq
, rq
->idle
, 0);
6346 rq
->idle
->static_prio
= MAX_PRIO
;
6347 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
6348 rq
->idle
->sched_class
= &idle_sched_class
;
6349 migrate_dead_tasks(cpu
);
6350 spin_unlock_irq(&rq
->lock
);
6352 migrate_nr_uninterruptible(rq
);
6353 BUG_ON(rq
->nr_running
!= 0);
6356 * No need to migrate the tasks: it was best-effort if
6357 * they didn't take sched_hotcpu_mutex. Just wake up
6360 spin_lock_irq(&rq
->lock
);
6361 while (!list_empty(&rq
->migration_queue
)) {
6362 struct migration_req
*req
;
6364 req
= list_entry(rq
->migration_queue
.next
,
6365 struct migration_req
, list
);
6366 list_del_init(&req
->list
);
6367 complete(&req
->done
);
6369 spin_unlock_irq(&rq
->lock
);
6373 case CPU_DYING_FROZEN
:
6374 /* Update our root-domain */
6376 spin_lock_irqsave(&rq
->lock
, flags
);
6378 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6381 spin_unlock_irqrestore(&rq
->lock
, flags
);
6388 /* Register at highest priority so that task migration (migrate_all_tasks)
6389 * happens before everything else.
6391 static struct notifier_block __cpuinitdata migration_notifier
= {
6392 .notifier_call
= migration_call
,
6396 void __init
migration_init(void)
6398 void *cpu
= (void *)(long)smp_processor_id();
6401 /* Start one for the boot CPU: */
6402 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6403 BUG_ON(err
== NOTIFY_BAD
);
6404 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6405 register_cpu_notifier(&migration_notifier
);
6411 #ifdef CONFIG_SCHED_DEBUG
6413 static inline const char *sd_level_to_string(enum sched_domain_level lvl
)
6426 case SD_LV_ALLNODES
:
6435 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6436 cpumask_t
*groupmask
)
6438 struct sched_group
*group
= sd
->groups
;
6441 cpulist_scnprintf(str
, sizeof(str
), sd
->span
);
6442 cpus_clear(*groupmask
);
6444 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6446 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6447 printk("does not load-balance\n");
6449 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6454 printk(KERN_CONT
"span %s level %s\n",
6455 str
, sd_level_to_string(sd
->level
));
6457 if (!cpu_isset(cpu
, sd
->span
)) {
6458 printk(KERN_ERR
"ERROR: domain->span does not contain "
6461 if (!cpu_isset(cpu
, group
->cpumask
)) {
6462 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6466 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6470 printk(KERN_ERR
"ERROR: group is NULL\n");
6474 if (!group
->__cpu_power
) {
6475 printk(KERN_CONT
"\n");
6476 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6481 if (!cpus_weight(group
->cpumask
)) {
6482 printk(KERN_CONT
"\n");
6483 printk(KERN_ERR
"ERROR: empty group\n");
6487 if (cpus_intersects(*groupmask
, group
->cpumask
)) {
6488 printk(KERN_CONT
"\n");
6489 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6493 cpus_or(*groupmask
, *groupmask
, group
->cpumask
);
6495 cpulist_scnprintf(str
, sizeof(str
), group
->cpumask
);
6496 printk(KERN_CONT
" %s", str
);
6498 group
= group
->next
;
6499 } while (group
!= sd
->groups
);
6500 printk(KERN_CONT
"\n");
6502 if (!cpus_equal(sd
->span
, *groupmask
))
6503 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6505 if (sd
->parent
&& !cpus_subset(*groupmask
, sd
->parent
->span
))
6506 printk(KERN_ERR
"ERROR: parent span is not a superset "
6507 "of domain->span\n");
6511 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6513 cpumask_t
*groupmask
;
6517 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6521 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6523 groupmask
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
6525 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6530 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6539 #else /* !CONFIG_SCHED_DEBUG */
6540 # define sched_domain_debug(sd, cpu) do { } while (0)
6541 #endif /* CONFIG_SCHED_DEBUG */
6543 static int sd_degenerate(struct sched_domain
*sd
)
6545 if (cpus_weight(sd
->span
) == 1)
6548 /* Following flags need at least 2 groups */
6549 if (sd
->flags
& (SD_LOAD_BALANCE
|
6550 SD_BALANCE_NEWIDLE
|
6554 SD_SHARE_PKG_RESOURCES
)) {
6555 if (sd
->groups
!= sd
->groups
->next
)
6559 /* Following flags don't use groups */
6560 if (sd
->flags
& (SD_WAKE_IDLE
|
6569 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6571 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6573 if (sd_degenerate(parent
))
6576 if (!cpus_equal(sd
->span
, parent
->span
))
6579 /* Does parent contain flags not in child? */
6580 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6581 if (cflags
& SD_WAKE_AFFINE
)
6582 pflags
&= ~SD_WAKE_BALANCE
;
6583 /* Flags needing groups don't count if only 1 group in parent */
6584 if (parent
->groups
== parent
->groups
->next
) {
6585 pflags
&= ~(SD_LOAD_BALANCE
|
6586 SD_BALANCE_NEWIDLE
|
6590 SD_SHARE_PKG_RESOURCES
);
6592 if (~cflags
& pflags
)
6598 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6600 unsigned long flags
;
6602 spin_lock_irqsave(&rq
->lock
, flags
);
6605 struct root_domain
*old_rd
= rq
->rd
;
6607 if (cpu_isset(rq
->cpu
, old_rd
->online
))
6610 cpu_clear(rq
->cpu
, old_rd
->span
);
6612 if (atomic_dec_and_test(&old_rd
->refcount
))
6616 atomic_inc(&rd
->refcount
);
6619 cpu_set(rq
->cpu
, rd
->span
);
6620 if (cpu_isset(rq
->cpu
, cpu_online_map
))
6623 spin_unlock_irqrestore(&rq
->lock
, flags
);
6626 static void init_rootdomain(struct root_domain
*rd
)
6628 memset(rd
, 0, sizeof(*rd
));
6630 cpus_clear(rd
->span
);
6631 cpus_clear(rd
->online
);
6633 cpupri_init(&rd
->cpupri
);
6636 static void init_defrootdomain(void)
6638 init_rootdomain(&def_root_domain
);
6639 atomic_set(&def_root_domain
.refcount
, 1);
6642 static struct root_domain
*alloc_rootdomain(void)
6644 struct root_domain
*rd
;
6646 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6650 init_rootdomain(rd
);
6656 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6657 * hold the hotplug lock.
6660 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6662 struct rq
*rq
= cpu_rq(cpu
);
6663 struct sched_domain
*tmp
;
6665 /* Remove the sched domains which do not contribute to scheduling. */
6666 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
6667 struct sched_domain
*parent
= tmp
->parent
;
6670 if (sd_parent_degenerate(tmp
, parent
)) {
6671 tmp
->parent
= parent
->parent
;
6673 parent
->parent
->child
= tmp
;
6677 if (sd
&& sd_degenerate(sd
)) {
6683 sched_domain_debug(sd
, cpu
);
6685 rq_attach_root(rq
, rd
);
6686 rcu_assign_pointer(rq
->sd
, sd
);
6689 /* cpus with isolated domains */
6690 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
6692 /* Setup the mask of cpus configured for isolated domains */
6693 static int __init
isolated_cpu_setup(char *str
)
6695 int ints
[NR_CPUS
], i
;
6697 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
6698 cpus_clear(cpu_isolated_map
);
6699 for (i
= 1; i
<= ints
[0]; i
++)
6700 if (ints
[i
] < NR_CPUS
)
6701 cpu_set(ints
[i
], cpu_isolated_map
);
6705 __setup("isolcpus=", isolated_cpu_setup
);
6708 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6709 * to a function which identifies what group(along with sched group) a CPU
6710 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6711 * (due to the fact that we keep track of groups covered with a cpumask_t).
6713 * init_sched_build_groups will build a circular linked list of the groups
6714 * covered by the given span, and will set each group's ->cpumask correctly,
6715 * and ->cpu_power to 0.
6718 init_sched_build_groups(const cpumask_t
*span
, const cpumask_t
*cpu_map
,
6719 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
6720 struct sched_group
**sg
,
6721 cpumask_t
*tmpmask
),
6722 cpumask_t
*covered
, cpumask_t
*tmpmask
)
6724 struct sched_group
*first
= NULL
, *last
= NULL
;
6727 cpus_clear(*covered
);
6729 for_each_cpu_mask(i
, *span
) {
6730 struct sched_group
*sg
;
6731 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6734 if (cpu_isset(i
, *covered
))
6737 cpus_clear(sg
->cpumask
);
6738 sg
->__cpu_power
= 0;
6740 for_each_cpu_mask(j
, *span
) {
6741 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6744 cpu_set(j
, *covered
);
6745 cpu_set(j
, sg
->cpumask
);
6756 #define SD_NODES_PER_DOMAIN 16
6761 * find_next_best_node - find the next node to include in a sched_domain
6762 * @node: node whose sched_domain we're building
6763 * @used_nodes: nodes already in the sched_domain
6765 * Find the next node to include in a given scheduling domain. Simply
6766 * finds the closest node not already in the @used_nodes map.
6768 * Should use nodemask_t.
6770 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
6772 int i
, n
, val
, min_val
, best_node
= 0;
6776 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6777 /* Start at @node */
6778 n
= (node
+ i
) % MAX_NUMNODES
;
6780 if (!nr_cpus_node(n
))
6783 /* Skip already used nodes */
6784 if (node_isset(n
, *used_nodes
))
6787 /* Simple min distance search */
6788 val
= node_distance(node
, n
);
6790 if (val
< min_val
) {
6796 node_set(best_node
, *used_nodes
);
6801 * sched_domain_node_span - get a cpumask for a node's sched_domain
6802 * @node: node whose cpumask we're constructing
6803 * @span: resulting cpumask
6805 * Given a node, construct a good cpumask for its sched_domain to span. It
6806 * should be one that prevents unnecessary balancing, but also spreads tasks
6809 static void sched_domain_node_span(int node
, cpumask_t
*span
)
6811 nodemask_t used_nodes
;
6812 node_to_cpumask_ptr(nodemask
, node
);
6816 nodes_clear(used_nodes
);
6818 cpus_or(*span
, *span
, *nodemask
);
6819 node_set(node
, used_nodes
);
6821 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6822 int next_node
= find_next_best_node(node
, &used_nodes
);
6824 node_to_cpumask_ptr_next(nodemask
, next_node
);
6825 cpus_or(*span
, *span
, *nodemask
);
6828 #endif /* CONFIG_NUMA */
6830 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6833 * SMT sched-domains:
6835 #ifdef CONFIG_SCHED_SMT
6836 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
6837 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
6840 cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
6844 *sg
= &per_cpu(sched_group_cpus
, cpu
);
6847 #endif /* CONFIG_SCHED_SMT */
6850 * multi-core sched-domains:
6852 #ifdef CONFIG_SCHED_MC
6853 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
6854 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
6855 #endif /* CONFIG_SCHED_MC */
6857 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6859 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
6864 *mask
= per_cpu(cpu_sibling_map
, cpu
);
6865 cpus_and(*mask
, *mask
, *cpu_map
);
6866 group
= first_cpu(*mask
);
6868 *sg
= &per_cpu(sched_group_core
, group
);
6871 #elif defined(CONFIG_SCHED_MC)
6873 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
6877 *sg
= &per_cpu(sched_group_core
, cpu
);
6882 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
6883 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
6886 cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
6890 #ifdef CONFIG_SCHED_MC
6891 *mask
= cpu_coregroup_map(cpu
);
6892 cpus_and(*mask
, *mask
, *cpu_map
);
6893 group
= first_cpu(*mask
);
6894 #elif defined(CONFIG_SCHED_SMT)
6895 *mask
= per_cpu(cpu_sibling_map
, cpu
);
6896 cpus_and(*mask
, *mask
, *cpu_map
);
6897 group
= first_cpu(*mask
);
6902 *sg
= &per_cpu(sched_group_phys
, group
);
6908 * The init_sched_build_groups can't handle what we want to do with node
6909 * groups, so roll our own. Now each node has its own list of groups which
6910 * gets dynamically allocated.
6912 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
6913 static struct sched_group
***sched_group_nodes_bycpu
;
6915 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
6916 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
6918 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
6919 struct sched_group
**sg
, cpumask_t
*nodemask
)
6923 *nodemask
= node_to_cpumask(cpu_to_node(cpu
));
6924 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
6925 group
= first_cpu(*nodemask
);
6928 *sg
= &per_cpu(sched_group_allnodes
, group
);
6932 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6934 struct sched_group
*sg
= group_head
;
6940 for_each_cpu_mask(j
, sg
->cpumask
) {
6941 struct sched_domain
*sd
;
6943 sd
= &per_cpu(phys_domains
, j
);
6944 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
6946 * Only add "power" once for each
6952 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
6955 } while (sg
!= group_head
);
6957 #endif /* CONFIG_NUMA */
6960 /* Free memory allocated for various sched_group structures */
6961 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
6965 for_each_cpu_mask(cpu
, *cpu_map
) {
6966 struct sched_group
**sched_group_nodes
6967 = sched_group_nodes_bycpu
[cpu
];
6969 if (!sched_group_nodes
)
6972 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6973 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6975 *nodemask
= node_to_cpumask(i
);
6976 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
6977 if (cpus_empty(*nodemask
))
6987 if (oldsg
!= sched_group_nodes
[i
])
6990 kfree(sched_group_nodes
);
6991 sched_group_nodes_bycpu
[cpu
] = NULL
;
6994 #else /* !CONFIG_NUMA */
6995 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
6998 #endif /* CONFIG_NUMA */
7001 * Initialize sched groups cpu_power.
7003 * cpu_power indicates the capacity of sched group, which is used while
7004 * distributing the load between different sched groups in a sched domain.
7005 * Typically cpu_power for all the groups in a sched domain will be same unless
7006 * there are asymmetries in the topology. If there are asymmetries, group
7007 * having more cpu_power will pickup more load compared to the group having
7010 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7011 * the maximum number of tasks a group can handle in the presence of other idle
7012 * or lightly loaded groups in the same sched domain.
7014 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
7016 struct sched_domain
*child
;
7017 struct sched_group
*group
;
7019 WARN_ON(!sd
|| !sd
->groups
);
7021 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
7026 sd
->groups
->__cpu_power
= 0;
7029 * For perf policy, if the groups in child domain share resources
7030 * (for example cores sharing some portions of the cache hierarchy
7031 * or SMT), then set this domain groups cpu_power such that each group
7032 * can handle only one task, when there are other idle groups in the
7033 * same sched domain.
7035 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
7037 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
7038 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
7043 * add cpu_power of each child group to this groups cpu_power
7045 group
= child
->groups
;
7047 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
7048 group
= group
->next
;
7049 } while (group
!= child
->groups
);
7053 * Initializers for schedule domains
7054 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7057 #define SD_INIT(sd, type) sd_init_##type(sd)
7058 #define SD_INIT_FUNC(type) \
7059 static noinline void sd_init_##type(struct sched_domain *sd) \
7061 memset(sd, 0, sizeof(*sd)); \
7062 *sd = SD_##type##_INIT; \
7063 sd->level = SD_LV_##type; \
7068 SD_INIT_FUNC(ALLNODES
)
7071 #ifdef CONFIG_SCHED_SMT
7072 SD_INIT_FUNC(SIBLING
)
7074 #ifdef CONFIG_SCHED_MC
7079 * To minimize stack usage kmalloc room for cpumasks and share the
7080 * space as the usage in build_sched_domains() dictates. Used only
7081 * if the amount of space is significant.
7084 cpumask_t tmpmask
; /* make this one first */
7087 cpumask_t this_sibling_map
;
7088 cpumask_t this_core_map
;
7090 cpumask_t send_covered
;
7093 cpumask_t domainspan
;
7095 cpumask_t notcovered
;
7100 #define SCHED_CPUMASK_ALLOC 1
7101 #define SCHED_CPUMASK_FREE(v) kfree(v)
7102 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7104 #define SCHED_CPUMASK_ALLOC 0
7105 #define SCHED_CPUMASK_FREE(v)
7106 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7109 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7110 ((unsigned long)(a) + offsetof(struct allmasks, v))
7112 static int default_relax_domain_level
= -1;
7114 static int __init
setup_relax_domain_level(char *str
)
7118 val
= simple_strtoul(str
, NULL
, 0);
7119 if (val
< SD_LV_MAX
)
7120 default_relax_domain_level
= val
;
7124 __setup("relax_domain_level=", setup_relax_domain_level
);
7126 static void set_domain_attribute(struct sched_domain
*sd
,
7127 struct sched_domain_attr
*attr
)
7131 if (!attr
|| attr
->relax_domain_level
< 0) {
7132 if (default_relax_domain_level
< 0)
7135 request
= default_relax_domain_level
;
7137 request
= attr
->relax_domain_level
;
7138 if (request
< sd
->level
) {
7139 /* turn off idle balance on this domain */
7140 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
7142 /* turn on idle balance on this domain */
7143 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
7148 * Build sched domains for a given set of cpus and attach the sched domains
7149 * to the individual cpus
7151 static int __build_sched_domains(const cpumask_t
*cpu_map
,
7152 struct sched_domain_attr
*attr
)
7155 struct root_domain
*rd
;
7156 SCHED_CPUMASK_DECLARE(allmasks
);
7159 struct sched_group
**sched_group_nodes
= NULL
;
7160 int sd_allnodes
= 0;
7163 * Allocate the per-node list of sched groups
7165 sched_group_nodes
= kcalloc(MAX_NUMNODES
, sizeof(struct sched_group
*),
7167 if (!sched_group_nodes
) {
7168 printk(KERN_WARNING
"Can not alloc sched group node list\n");
7173 rd
= alloc_rootdomain();
7175 printk(KERN_WARNING
"Cannot alloc root domain\n");
7177 kfree(sched_group_nodes
);
7182 #if SCHED_CPUMASK_ALLOC
7183 /* get space for all scratch cpumask variables */
7184 allmasks
= kmalloc(sizeof(*allmasks
), GFP_KERNEL
);
7186 printk(KERN_WARNING
"Cannot alloc cpumask array\n");
7189 kfree(sched_group_nodes
);
7194 tmpmask
= (cpumask_t
*)allmasks
;
7198 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
7202 * Set up domains for cpus specified by the cpu_map.
7204 for_each_cpu_mask(i
, *cpu_map
) {
7205 struct sched_domain
*sd
= NULL
, *p
;
7206 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7208 *nodemask
= node_to_cpumask(cpu_to_node(i
));
7209 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7212 if (cpus_weight(*cpu_map
) >
7213 SD_NODES_PER_DOMAIN
*cpus_weight(*nodemask
)) {
7214 sd
= &per_cpu(allnodes_domains
, i
);
7215 SD_INIT(sd
, ALLNODES
);
7216 set_domain_attribute(sd
, attr
);
7217 sd
->span
= *cpu_map
;
7218 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7224 sd
= &per_cpu(node_domains
, i
);
7226 set_domain_attribute(sd
, attr
);
7227 sched_domain_node_span(cpu_to_node(i
), &sd
->span
);
7231 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7235 sd
= &per_cpu(phys_domains
, i
);
7237 set_domain_attribute(sd
, attr
);
7238 sd
->span
= *nodemask
;
7242 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7244 #ifdef CONFIG_SCHED_MC
7246 sd
= &per_cpu(core_domains
, i
);
7248 set_domain_attribute(sd
, attr
);
7249 sd
->span
= cpu_coregroup_map(i
);
7250 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7253 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7256 #ifdef CONFIG_SCHED_SMT
7258 sd
= &per_cpu(cpu_domains
, i
);
7259 SD_INIT(sd
, SIBLING
);
7260 set_domain_attribute(sd
, attr
);
7261 sd
->span
= per_cpu(cpu_sibling_map
, i
);
7262 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7265 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7269 #ifdef CONFIG_SCHED_SMT
7270 /* Set up CPU (sibling) groups */
7271 for_each_cpu_mask(i
, *cpu_map
) {
7272 SCHED_CPUMASK_VAR(this_sibling_map
, allmasks
);
7273 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7275 *this_sibling_map
= per_cpu(cpu_sibling_map
, i
);
7276 cpus_and(*this_sibling_map
, *this_sibling_map
, *cpu_map
);
7277 if (i
!= first_cpu(*this_sibling_map
))
7280 init_sched_build_groups(this_sibling_map
, cpu_map
,
7282 send_covered
, tmpmask
);
7286 #ifdef CONFIG_SCHED_MC
7287 /* Set up multi-core groups */
7288 for_each_cpu_mask(i
, *cpu_map
) {
7289 SCHED_CPUMASK_VAR(this_core_map
, allmasks
);
7290 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7292 *this_core_map
= cpu_coregroup_map(i
);
7293 cpus_and(*this_core_map
, *this_core_map
, *cpu_map
);
7294 if (i
!= first_cpu(*this_core_map
))
7297 init_sched_build_groups(this_core_map
, cpu_map
,
7299 send_covered
, tmpmask
);
7303 /* Set up physical groups */
7304 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
7305 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7306 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7308 *nodemask
= node_to_cpumask(i
);
7309 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7310 if (cpus_empty(*nodemask
))
7313 init_sched_build_groups(nodemask
, cpu_map
,
7315 send_covered
, tmpmask
);
7319 /* Set up node groups */
7321 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7323 init_sched_build_groups(cpu_map
, cpu_map
,
7324 &cpu_to_allnodes_group
,
7325 send_covered
, tmpmask
);
7328 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
7329 /* Set up node groups */
7330 struct sched_group
*sg
, *prev
;
7331 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7332 SCHED_CPUMASK_VAR(domainspan
, allmasks
);
7333 SCHED_CPUMASK_VAR(covered
, allmasks
);
7336 *nodemask
= node_to_cpumask(i
);
7337 cpus_clear(*covered
);
7339 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7340 if (cpus_empty(*nodemask
)) {
7341 sched_group_nodes
[i
] = NULL
;
7345 sched_domain_node_span(i
, domainspan
);
7346 cpus_and(*domainspan
, *domainspan
, *cpu_map
);
7348 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
7350 printk(KERN_WARNING
"Can not alloc domain group for "
7354 sched_group_nodes
[i
] = sg
;
7355 for_each_cpu_mask(j
, *nodemask
) {
7356 struct sched_domain
*sd
;
7358 sd
= &per_cpu(node_domains
, j
);
7361 sg
->__cpu_power
= 0;
7362 sg
->cpumask
= *nodemask
;
7364 cpus_or(*covered
, *covered
, *nodemask
);
7367 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
7368 SCHED_CPUMASK_VAR(notcovered
, allmasks
);
7369 int n
= (i
+ j
) % MAX_NUMNODES
;
7370 node_to_cpumask_ptr(pnodemask
, n
);
7372 cpus_complement(*notcovered
, *covered
);
7373 cpus_and(*tmpmask
, *notcovered
, *cpu_map
);
7374 cpus_and(*tmpmask
, *tmpmask
, *domainspan
);
7375 if (cpus_empty(*tmpmask
))
7378 cpus_and(*tmpmask
, *tmpmask
, *pnodemask
);
7379 if (cpus_empty(*tmpmask
))
7382 sg
= kmalloc_node(sizeof(struct sched_group
),
7386 "Can not alloc domain group for node %d\n", j
);
7389 sg
->__cpu_power
= 0;
7390 sg
->cpumask
= *tmpmask
;
7391 sg
->next
= prev
->next
;
7392 cpus_or(*covered
, *covered
, *tmpmask
);
7399 /* Calculate CPU power for physical packages and nodes */
7400 #ifdef CONFIG_SCHED_SMT
7401 for_each_cpu_mask(i
, *cpu_map
) {
7402 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
7404 init_sched_groups_power(i
, sd
);
7407 #ifdef CONFIG_SCHED_MC
7408 for_each_cpu_mask(i
, *cpu_map
) {
7409 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
7411 init_sched_groups_power(i
, sd
);
7415 for_each_cpu_mask(i
, *cpu_map
) {
7416 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
7418 init_sched_groups_power(i
, sd
);
7422 for (i
= 0; i
< MAX_NUMNODES
; i
++)
7423 init_numa_sched_groups_power(sched_group_nodes
[i
]);
7426 struct sched_group
*sg
;
7428 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
,
7430 init_numa_sched_groups_power(sg
);
7434 /* Attach the domains */
7435 for_each_cpu_mask(i
, *cpu_map
) {
7436 struct sched_domain
*sd
;
7437 #ifdef CONFIG_SCHED_SMT
7438 sd
= &per_cpu(cpu_domains
, i
);
7439 #elif defined(CONFIG_SCHED_MC)
7440 sd
= &per_cpu(core_domains
, i
);
7442 sd
= &per_cpu(phys_domains
, i
);
7444 cpu_attach_domain(sd
, rd
, i
);
7447 SCHED_CPUMASK_FREE((void *)allmasks
);
7452 free_sched_groups(cpu_map
, tmpmask
);
7453 SCHED_CPUMASK_FREE((void *)allmasks
);
7458 static int build_sched_domains(const cpumask_t
*cpu_map
)
7460 return __build_sched_domains(cpu_map
, NULL
);
7463 static cpumask_t
*doms_cur
; /* current sched domains */
7464 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7465 static struct sched_domain_attr
*dattr_cur
;
7466 /* attribues of custom domains in 'doms_cur' */
7469 * Special case: If a kmalloc of a doms_cur partition (array of
7470 * cpumask_t) fails, then fallback to a single sched domain,
7471 * as determined by the single cpumask_t fallback_doms.
7473 static cpumask_t fallback_doms
;
7475 void __attribute__((weak
)) arch_update_cpu_topology(void)
7480 * Free current domain masks.
7481 * Called after all cpus are attached to NULL domain.
7483 static void free_sched_domains(void)
7486 if (doms_cur
!= &fallback_doms
)
7488 doms_cur
= &fallback_doms
;
7492 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7493 * For now this just excludes isolated cpus, but could be used to
7494 * exclude other special cases in the future.
7496 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
7500 arch_update_cpu_topology();
7502 doms_cur
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
7504 doms_cur
= &fallback_doms
;
7505 cpus_andnot(*doms_cur
, *cpu_map
, cpu_isolated_map
);
7507 err
= build_sched_domains(doms_cur
);
7508 register_sched_domain_sysctl();
7513 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
,
7516 free_sched_groups(cpu_map
, tmpmask
);
7520 * Detach sched domains from a group of cpus specified in cpu_map
7521 * These cpus will now be attached to the NULL domain
7523 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
7528 unregister_sched_domain_sysctl();
7530 for_each_cpu_mask(i
, *cpu_map
)
7531 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7532 synchronize_sched();
7533 arch_destroy_sched_domains(cpu_map
, &tmpmask
);
7536 /* handle null as "default" */
7537 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7538 struct sched_domain_attr
*new, int idx_new
)
7540 struct sched_domain_attr tmp
;
7547 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7548 new ? (new + idx_new
) : &tmp
,
7549 sizeof(struct sched_domain_attr
));
7553 * Partition sched domains as specified by the 'ndoms_new'
7554 * cpumasks in the array doms_new[] of cpumasks. This compares
7555 * doms_new[] to the current sched domain partitioning, doms_cur[].
7556 * It destroys each deleted domain and builds each new domain.
7558 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7559 * The masks don't intersect (don't overlap.) We should setup one
7560 * sched domain for each mask. CPUs not in any of the cpumasks will
7561 * not be load balanced. If the same cpumask appears both in the
7562 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7565 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7566 * ownership of it and will kfree it when done with it. If the caller
7567 * failed the kmalloc call, then it can pass in doms_new == NULL,
7568 * and partition_sched_domains() will fallback to the single partition
7571 * Call with hotplug lock held
7573 void partition_sched_domains(int ndoms_new
, cpumask_t
*doms_new
,
7574 struct sched_domain_attr
*dattr_new
)
7578 mutex_lock(&sched_domains_mutex
);
7580 /* always unregister in case we don't destroy any domains */
7581 unregister_sched_domain_sysctl();
7583 if (doms_new
== NULL
) {
7585 doms_new
= &fallback_doms
;
7586 cpus_andnot(doms_new
[0], cpu_online_map
, cpu_isolated_map
);
7590 /* Destroy deleted domains */
7591 for (i
= 0; i
< ndoms_cur
; i
++) {
7592 for (j
= 0; j
< ndoms_new
; j
++) {
7593 if (cpus_equal(doms_cur
[i
], doms_new
[j
])
7594 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7597 /* no match - a current sched domain not in new doms_new[] */
7598 detach_destroy_domains(doms_cur
+ i
);
7603 /* Build new domains */
7604 for (i
= 0; i
< ndoms_new
; i
++) {
7605 for (j
= 0; j
< ndoms_cur
; j
++) {
7606 if (cpus_equal(doms_new
[i
], doms_cur
[j
])
7607 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7610 /* no match - add a new doms_new */
7611 __build_sched_domains(doms_new
+ i
,
7612 dattr_new
? dattr_new
+ i
: NULL
);
7617 /* Remember the new sched domains */
7618 if (doms_cur
!= &fallback_doms
)
7620 kfree(dattr_cur
); /* kfree(NULL) is safe */
7621 doms_cur
= doms_new
;
7622 dattr_cur
= dattr_new
;
7623 ndoms_cur
= ndoms_new
;
7625 register_sched_domain_sysctl();
7627 mutex_unlock(&sched_domains_mutex
);
7630 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7631 int arch_reinit_sched_domains(void)
7636 mutex_lock(&sched_domains_mutex
);
7637 detach_destroy_domains(&cpu_online_map
);
7638 free_sched_domains();
7639 err
= arch_init_sched_domains(&cpu_online_map
);
7640 mutex_unlock(&sched_domains_mutex
);
7646 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7650 if (buf
[0] != '0' && buf
[0] != '1')
7654 sched_smt_power_savings
= (buf
[0] == '1');
7656 sched_mc_power_savings
= (buf
[0] == '1');
7658 ret
= arch_reinit_sched_domains();
7660 return ret
? ret
: count
;
7663 #ifdef CONFIG_SCHED_MC
7664 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
7666 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7668 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
7669 const char *buf
, size_t count
)
7671 return sched_power_savings_store(buf
, count
, 0);
7673 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
7674 sched_mc_power_savings_store
);
7677 #ifdef CONFIG_SCHED_SMT
7678 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
7680 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7682 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
7683 const char *buf
, size_t count
)
7685 return sched_power_savings_store(buf
, count
, 1);
7687 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
7688 sched_smt_power_savings_store
);
7691 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7695 #ifdef CONFIG_SCHED_SMT
7697 err
= sysfs_create_file(&cls
->kset
.kobj
,
7698 &attr_sched_smt_power_savings
.attr
);
7700 #ifdef CONFIG_SCHED_MC
7701 if (!err
&& mc_capable())
7702 err
= sysfs_create_file(&cls
->kset
.kobj
,
7703 &attr_sched_mc_power_savings
.attr
);
7707 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7710 * Force a reinitialization of the sched domains hierarchy. The domains
7711 * and groups cannot be updated in place without racing with the balancing
7712 * code, so we temporarily attach all running cpus to the NULL domain
7713 * which will prevent rebalancing while the sched domains are recalculated.
7715 static int update_sched_domains(struct notifier_block
*nfb
,
7716 unsigned long action
, void *hcpu
)
7718 int cpu
= (int)(long)hcpu
;
7721 case CPU_DOWN_PREPARE
:
7722 case CPU_DOWN_PREPARE_FROZEN
:
7723 disable_runtime(cpu_rq(cpu
));
7725 case CPU_UP_PREPARE
:
7726 case CPU_UP_PREPARE_FROZEN
:
7727 detach_destroy_domains(&cpu_online_map
);
7728 free_sched_domains();
7732 case CPU_DOWN_FAILED
:
7733 case CPU_DOWN_FAILED_FROZEN
:
7735 case CPU_ONLINE_FROZEN
:
7736 enable_runtime(cpu_rq(cpu
));
7738 case CPU_UP_CANCELED
:
7739 case CPU_UP_CANCELED_FROZEN
:
7741 case CPU_DEAD_FROZEN
:
7743 * Fall through and re-initialise the domains.
7750 #ifndef CONFIG_CPUSETS
7752 * Create default domain partitioning if cpusets are disabled.
7753 * Otherwise we let cpusets rebuild the domains based on the
7757 /* The hotplug lock is already held by cpu_up/cpu_down */
7758 arch_init_sched_domains(&cpu_online_map
);
7764 void __init
sched_init_smp(void)
7766 cpumask_t non_isolated_cpus
;
7768 #if defined(CONFIG_NUMA)
7769 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
7771 BUG_ON(sched_group_nodes_bycpu
== NULL
);
7774 mutex_lock(&sched_domains_mutex
);
7775 arch_init_sched_domains(&cpu_online_map
);
7776 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
7777 if (cpus_empty(non_isolated_cpus
))
7778 cpu_set(smp_processor_id(), non_isolated_cpus
);
7779 mutex_unlock(&sched_domains_mutex
);
7781 /* XXX: Theoretical race here - CPU may be hotplugged now */
7782 hotcpu_notifier(update_sched_domains
, 0);
7785 /* Move init over to a non-isolated CPU */
7786 if (set_cpus_allowed_ptr(current
, &non_isolated_cpus
) < 0)
7788 sched_init_granularity();
7791 void __init
sched_init_smp(void)
7793 sched_init_granularity();
7795 #endif /* CONFIG_SMP */
7797 int in_sched_functions(unsigned long addr
)
7799 return in_lock_functions(addr
) ||
7800 (addr
>= (unsigned long)__sched_text_start
7801 && addr
< (unsigned long)__sched_text_end
);
7804 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7806 cfs_rq
->tasks_timeline
= RB_ROOT
;
7807 INIT_LIST_HEAD(&cfs_rq
->tasks
);
7808 #ifdef CONFIG_FAIR_GROUP_SCHED
7811 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7814 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7816 struct rt_prio_array
*array
;
7819 array
= &rt_rq
->active
;
7820 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
7821 INIT_LIST_HEAD(array
->queue
+ i
);
7822 __clear_bit(i
, array
->bitmap
);
7824 /* delimiter for bitsearch: */
7825 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
7827 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7828 rt_rq
->highest_prio
= MAX_RT_PRIO
;
7831 rt_rq
->rt_nr_migratory
= 0;
7832 rt_rq
->overloaded
= 0;
7836 rt_rq
->rt_throttled
= 0;
7837 rt_rq
->rt_runtime
= 0;
7838 spin_lock_init(&rt_rq
->rt_runtime_lock
);
7840 #ifdef CONFIG_RT_GROUP_SCHED
7841 rt_rq
->rt_nr_boosted
= 0;
7846 #ifdef CONFIG_FAIR_GROUP_SCHED
7847 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7848 struct sched_entity
*se
, int cpu
, int add
,
7849 struct sched_entity
*parent
)
7851 struct rq
*rq
= cpu_rq(cpu
);
7852 tg
->cfs_rq
[cpu
] = cfs_rq
;
7853 init_cfs_rq(cfs_rq
, rq
);
7856 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7859 /* se could be NULL for init_task_group */
7864 se
->cfs_rq
= &rq
->cfs
;
7866 se
->cfs_rq
= parent
->my_q
;
7869 se
->load
.weight
= tg
->shares
;
7870 se
->load
.inv_weight
= 0;
7871 se
->parent
= parent
;
7875 #ifdef CONFIG_RT_GROUP_SCHED
7876 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
7877 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
7878 struct sched_rt_entity
*parent
)
7880 struct rq
*rq
= cpu_rq(cpu
);
7882 tg
->rt_rq
[cpu
] = rt_rq
;
7883 init_rt_rq(rt_rq
, rq
);
7885 rt_rq
->rt_se
= rt_se
;
7886 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7888 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
7890 tg
->rt_se
[cpu
] = rt_se
;
7895 rt_se
->rt_rq
= &rq
->rt
;
7897 rt_se
->rt_rq
= parent
->my_q
;
7899 rt_se
->my_q
= rt_rq
;
7900 rt_se
->parent
= parent
;
7901 INIT_LIST_HEAD(&rt_se
->run_list
);
7905 void __init
sched_init(void)
7908 unsigned long alloc_size
= 0, ptr
;
7910 #ifdef CONFIG_FAIR_GROUP_SCHED
7911 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7913 #ifdef CONFIG_RT_GROUP_SCHED
7914 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7916 #ifdef CONFIG_USER_SCHED
7920 * As sched_init() is called before page_alloc is setup,
7921 * we use alloc_bootmem().
7924 ptr
= (unsigned long)alloc_bootmem(alloc_size
);
7926 #ifdef CONFIG_FAIR_GROUP_SCHED
7927 init_task_group
.se
= (struct sched_entity
**)ptr
;
7928 ptr
+= nr_cpu_ids
* sizeof(void **);
7930 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7931 ptr
+= nr_cpu_ids
* sizeof(void **);
7933 #ifdef CONFIG_USER_SCHED
7934 root_task_group
.se
= (struct sched_entity
**)ptr
;
7935 ptr
+= nr_cpu_ids
* sizeof(void **);
7937 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7938 ptr
+= nr_cpu_ids
* sizeof(void **);
7939 #endif /* CONFIG_USER_SCHED */
7940 #endif /* CONFIG_FAIR_GROUP_SCHED */
7941 #ifdef CONFIG_RT_GROUP_SCHED
7942 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7943 ptr
+= nr_cpu_ids
* sizeof(void **);
7945 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7946 ptr
+= nr_cpu_ids
* sizeof(void **);
7948 #ifdef CONFIG_USER_SCHED
7949 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7950 ptr
+= nr_cpu_ids
* sizeof(void **);
7952 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7953 ptr
+= nr_cpu_ids
* sizeof(void **);
7954 #endif /* CONFIG_USER_SCHED */
7955 #endif /* CONFIG_RT_GROUP_SCHED */
7959 init_defrootdomain();
7962 init_rt_bandwidth(&def_rt_bandwidth
,
7963 global_rt_period(), global_rt_runtime());
7965 #ifdef CONFIG_RT_GROUP_SCHED
7966 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
7967 global_rt_period(), global_rt_runtime());
7968 #ifdef CONFIG_USER_SCHED
7969 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
7970 global_rt_period(), RUNTIME_INF
);
7971 #endif /* CONFIG_USER_SCHED */
7972 #endif /* CONFIG_RT_GROUP_SCHED */
7974 #ifdef CONFIG_GROUP_SCHED
7975 list_add(&init_task_group
.list
, &task_groups
);
7976 INIT_LIST_HEAD(&init_task_group
.children
);
7978 #ifdef CONFIG_USER_SCHED
7979 INIT_LIST_HEAD(&root_task_group
.children
);
7980 init_task_group
.parent
= &root_task_group
;
7981 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
7982 #endif /* CONFIG_USER_SCHED */
7983 #endif /* CONFIG_GROUP_SCHED */
7985 for_each_possible_cpu(i
) {
7989 spin_lock_init(&rq
->lock
);
7990 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
7992 init_cfs_rq(&rq
->cfs
, rq
);
7993 init_rt_rq(&rq
->rt
, rq
);
7994 #ifdef CONFIG_FAIR_GROUP_SCHED
7995 init_task_group
.shares
= init_task_group_load
;
7996 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7997 #ifdef CONFIG_CGROUP_SCHED
7999 * How much cpu bandwidth does init_task_group get?
8001 * In case of task-groups formed thr' the cgroup filesystem, it
8002 * gets 100% of the cpu resources in the system. This overall
8003 * system cpu resource is divided among the tasks of
8004 * init_task_group and its child task-groups in a fair manner,
8005 * based on each entity's (task or task-group's) weight
8006 * (se->load.weight).
8008 * In other words, if init_task_group has 10 tasks of weight
8009 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8010 * then A0's share of the cpu resource is:
8012 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8014 * We achieve this by letting init_task_group's tasks sit
8015 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8017 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
8018 #elif defined CONFIG_USER_SCHED
8019 root_task_group
.shares
= NICE_0_LOAD
;
8020 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
8022 * In case of task-groups formed thr' the user id of tasks,
8023 * init_task_group represents tasks belonging to root user.
8024 * Hence it forms a sibling of all subsequent groups formed.
8025 * In this case, init_task_group gets only a fraction of overall
8026 * system cpu resource, based on the weight assigned to root
8027 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8028 * by letting tasks of init_task_group sit in a separate cfs_rq
8029 * (init_cfs_rq) and having one entity represent this group of
8030 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8032 init_tg_cfs_entry(&init_task_group
,
8033 &per_cpu(init_cfs_rq
, i
),
8034 &per_cpu(init_sched_entity
, i
), i
, 1,
8035 root_task_group
.se
[i
]);
8038 #endif /* CONFIG_FAIR_GROUP_SCHED */
8040 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
8041 #ifdef CONFIG_RT_GROUP_SCHED
8042 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
8043 #ifdef CONFIG_CGROUP_SCHED
8044 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
8045 #elif defined CONFIG_USER_SCHED
8046 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
8047 init_tg_rt_entry(&init_task_group
,
8048 &per_cpu(init_rt_rq
, i
),
8049 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
8050 root_task_group
.rt_se
[i
]);
8054 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
8055 rq
->cpu_load
[j
] = 0;
8059 rq
->active_balance
= 0;
8060 rq
->next_balance
= jiffies
;
8064 rq
->migration_thread
= NULL
;
8065 INIT_LIST_HEAD(&rq
->migration_queue
);
8066 rq_attach_root(rq
, &def_root_domain
);
8069 atomic_set(&rq
->nr_iowait
, 0);
8072 set_load_weight(&init_task
);
8074 #ifdef CONFIG_PREEMPT_NOTIFIERS
8075 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
8079 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
8082 #ifdef CONFIG_RT_MUTEXES
8083 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
8087 * The boot idle thread does lazy MMU switching as well:
8089 atomic_inc(&init_mm
.mm_count
);
8090 enter_lazy_tlb(&init_mm
, current
);
8093 * Make us the idle thread. Technically, schedule() should not be
8094 * called from this thread, however somewhere below it might be,
8095 * but because we are the idle thread, we just pick up running again
8096 * when this runqueue becomes "idle".
8098 init_idle(current
, smp_processor_id());
8100 * During early bootup we pretend to be a normal task:
8102 current
->sched_class
= &fair_sched_class
;
8104 scheduler_running
= 1;
8107 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8108 void __might_sleep(char *file
, int line
)
8111 static unsigned long prev_jiffy
; /* ratelimiting */
8113 if ((in_atomic() || irqs_disabled()) &&
8114 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
8115 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
8117 prev_jiffy
= jiffies
;
8118 printk(KERN_ERR
"BUG: sleeping function called from invalid"
8119 " context at %s:%d\n", file
, line
);
8120 printk("in_atomic():%d, irqs_disabled():%d\n",
8121 in_atomic(), irqs_disabled());
8122 debug_show_held_locks(current
);
8123 if (irqs_disabled())
8124 print_irqtrace_events(current
);
8129 EXPORT_SYMBOL(__might_sleep
);
8132 #ifdef CONFIG_MAGIC_SYSRQ
8133 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
8137 update_rq_clock(rq
);
8138 on_rq
= p
->se
.on_rq
;
8140 deactivate_task(rq
, p
, 0);
8141 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
8143 activate_task(rq
, p
, 0);
8144 resched_task(rq
->curr
);
8148 void normalize_rt_tasks(void)
8150 struct task_struct
*g
, *p
;
8151 unsigned long flags
;
8154 read_lock_irqsave(&tasklist_lock
, flags
);
8155 do_each_thread(g
, p
) {
8157 * Only normalize user tasks:
8162 p
->se
.exec_start
= 0;
8163 #ifdef CONFIG_SCHEDSTATS
8164 p
->se
.wait_start
= 0;
8165 p
->se
.sleep_start
= 0;
8166 p
->se
.block_start
= 0;
8171 * Renice negative nice level userspace
8174 if (TASK_NICE(p
) < 0 && p
->mm
)
8175 set_user_nice(p
, 0);
8179 spin_lock(&p
->pi_lock
);
8180 rq
= __task_rq_lock(p
);
8182 normalize_task(rq
, p
);
8184 __task_rq_unlock(rq
);
8185 spin_unlock(&p
->pi_lock
);
8186 } while_each_thread(g
, p
);
8188 read_unlock_irqrestore(&tasklist_lock
, flags
);
8191 #endif /* CONFIG_MAGIC_SYSRQ */
8195 * These functions are only useful for the IA64 MCA handling.
8197 * They can only be called when the whole system has been
8198 * stopped - every CPU needs to be quiescent, and no scheduling
8199 * activity can take place. Using them for anything else would
8200 * be a serious bug, and as a result, they aren't even visible
8201 * under any other configuration.
8205 * curr_task - return the current task for a given cpu.
8206 * @cpu: the processor in question.
8208 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8210 struct task_struct
*curr_task(int cpu
)
8212 return cpu_curr(cpu
);
8216 * set_curr_task - set the current task for a given cpu.
8217 * @cpu: the processor in question.
8218 * @p: the task pointer to set.
8220 * Description: This function must only be used when non-maskable interrupts
8221 * are serviced on a separate stack. It allows the architecture to switch the
8222 * notion of the current task on a cpu in a non-blocking manner. This function
8223 * must be called with all CPU's synchronized, and interrupts disabled, the
8224 * and caller must save the original value of the current task (see
8225 * curr_task() above) and restore that value before reenabling interrupts and
8226 * re-starting the system.
8228 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8230 void set_curr_task(int cpu
, struct task_struct
*p
)
8237 #ifdef CONFIG_FAIR_GROUP_SCHED
8238 static void free_fair_sched_group(struct task_group
*tg
)
8242 for_each_possible_cpu(i
) {
8244 kfree(tg
->cfs_rq
[i
]);
8254 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8256 struct cfs_rq
*cfs_rq
;
8257 struct sched_entity
*se
, *parent_se
;
8261 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8264 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8268 tg
->shares
= NICE_0_LOAD
;
8270 for_each_possible_cpu(i
) {
8273 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
),
8274 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8278 se
= kmalloc_node(sizeof(struct sched_entity
),
8279 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8283 parent_se
= parent
? parent
->se
[i
] : NULL
;
8284 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent_se
);
8293 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8295 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
8296 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
8299 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8301 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
8303 #else /* !CONFG_FAIR_GROUP_SCHED */
8304 static inline void free_fair_sched_group(struct task_group
*tg
)
8309 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8314 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8318 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8321 #endif /* CONFIG_FAIR_GROUP_SCHED */
8323 #ifdef CONFIG_RT_GROUP_SCHED
8324 static void free_rt_sched_group(struct task_group
*tg
)
8328 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8330 for_each_possible_cpu(i
) {
8332 kfree(tg
->rt_rq
[i
]);
8334 kfree(tg
->rt_se
[i
]);
8342 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8344 struct rt_rq
*rt_rq
;
8345 struct sched_rt_entity
*rt_se
, *parent_se
;
8349 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8352 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8356 init_rt_bandwidth(&tg
->rt_bandwidth
,
8357 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8359 for_each_possible_cpu(i
) {
8362 rt_rq
= kmalloc_node(sizeof(struct rt_rq
),
8363 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8367 rt_se
= kmalloc_node(sizeof(struct sched_rt_entity
),
8368 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8372 parent_se
= parent
? parent
->rt_se
[i
] : NULL
;
8373 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent_se
);
8382 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8384 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8385 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8388 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8390 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8392 #else /* !CONFIG_RT_GROUP_SCHED */
8393 static inline void free_rt_sched_group(struct task_group
*tg
)
8398 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8403 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8407 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8410 #endif /* CONFIG_RT_GROUP_SCHED */
8412 #ifdef CONFIG_GROUP_SCHED
8413 static void free_sched_group(struct task_group
*tg
)
8415 free_fair_sched_group(tg
);
8416 free_rt_sched_group(tg
);
8420 /* allocate runqueue etc for a new task group */
8421 struct task_group
*sched_create_group(struct task_group
*parent
)
8423 struct task_group
*tg
;
8424 unsigned long flags
;
8427 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8429 return ERR_PTR(-ENOMEM
);
8431 if (!alloc_fair_sched_group(tg
, parent
))
8434 if (!alloc_rt_sched_group(tg
, parent
))
8437 spin_lock_irqsave(&task_group_lock
, flags
);
8438 for_each_possible_cpu(i
) {
8439 register_fair_sched_group(tg
, i
);
8440 register_rt_sched_group(tg
, i
);
8442 list_add_rcu(&tg
->list
, &task_groups
);
8444 WARN_ON(!parent
); /* root should already exist */
8446 tg
->parent
= parent
;
8447 list_add_rcu(&tg
->siblings
, &parent
->children
);
8448 INIT_LIST_HEAD(&tg
->children
);
8449 spin_unlock_irqrestore(&task_group_lock
, flags
);
8454 free_sched_group(tg
);
8455 return ERR_PTR(-ENOMEM
);
8458 /* rcu callback to free various structures associated with a task group */
8459 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8461 /* now it should be safe to free those cfs_rqs */
8462 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8465 /* Destroy runqueue etc associated with a task group */
8466 void sched_destroy_group(struct task_group
*tg
)
8468 unsigned long flags
;
8471 spin_lock_irqsave(&task_group_lock
, flags
);
8472 for_each_possible_cpu(i
) {
8473 unregister_fair_sched_group(tg
, i
);
8474 unregister_rt_sched_group(tg
, i
);
8476 list_del_rcu(&tg
->list
);
8477 list_del_rcu(&tg
->siblings
);
8478 spin_unlock_irqrestore(&task_group_lock
, flags
);
8480 /* wait for possible concurrent references to cfs_rqs complete */
8481 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8484 /* change task's runqueue when it moves between groups.
8485 * The caller of this function should have put the task in its new group
8486 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8487 * reflect its new group.
8489 void sched_move_task(struct task_struct
*tsk
)
8492 unsigned long flags
;
8495 rq
= task_rq_lock(tsk
, &flags
);
8497 update_rq_clock(rq
);
8499 running
= task_current(rq
, tsk
);
8500 on_rq
= tsk
->se
.on_rq
;
8503 dequeue_task(rq
, tsk
, 0);
8504 if (unlikely(running
))
8505 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8507 set_task_rq(tsk
, task_cpu(tsk
));
8509 #ifdef CONFIG_FAIR_GROUP_SCHED
8510 if (tsk
->sched_class
->moved_group
)
8511 tsk
->sched_class
->moved_group(tsk
);
8514 if (unlikely(running
))
8515 tsk
->sched_class
->set_curr_task(rq
);
8517 enqueue_task(rq
, tsk
, 0);
8519 task_rq_unlock(rq
, &flags
);
8521 #endif /* CONFIG_GROUP_SCHED */
8523 #ifdef CONFIG_FAIR_GROUP_SCHED
8524 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8526 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8531 dequeue_entity(cfs_rq
, se
, 0);
8533 se
->load
.weight
= shares
;
8534 se
->load
.inv_weight
= 0;
8537 enqueue_entity(cfs_rq
, se
, 0);
8540 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8542 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8543 struct rq
*rq
= cfs_rq
->rq
;
8544 unsigned long flags
;
8546 spin_lock_irqsave(&rq
->lock
, flags
);
8547 __set_se_shares(se
, shares
);
8548 spin_unlock_irqrestore(&rq
->lock
, flags
);
8551 static DEFINE_MUTEX(shares_mutex
);
8553 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8556 unsigned long flags
;
8559 * We can't change the weight of the root cgroup.
8564 if (shares
< MIN_SHARES
)
8565 shares
= MIN_SHARES
;
8566 else if (shares
> MAX_SHARES
)
8567 shares
= MAX_SHARES
;
8569 mutex_lock(&shares_mutex
);
8570 if (tg
->shares
== shares
)
8573 spin_lock_irqsave(&task_group_lock
, flags
);
8574 for_each_possible_cpu(i
)
8575 unregister_fair_sched_group(tg
, i
);
8576 list_del_rcu(&tg
->siblings
);
8577 spin_unlock_irqrestore(&task_group_lock
, flags
);
8579 /* wait for any ongoing reference to this group to finish */
8580 synchronize_sched();
8583 * Now we are free to modify the group's share on each cpu
8584 * w/o tripping rebalance_share or load_balance_fair.
8586 tg
->shares
= shares
;
8587 for_each_possible_cpu(i
) {
8591 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
8592 set_se_shares(tg
->se
[i
], shares
);
8596 * Enable load balance activity on this group, by inserting it back on
8597 * each cpu's rq->leaf_cfs_rq_list.
8599 spin_lock_irqsave(&task_group_lock
, flags
);
8600 for_each_possible_cpu(i
)
8601 register_fair_sched_group(tg
, i
);
8602 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
8603 spin_unlock_irqrestore(&task_group_lock
, flags
);
8605 mutex_unlock(&shares_mutex
);
8609 unsigned long sched_group_shares(struct task_group
*tg
)
8615 #ifdef CONFIG_RT_GROUP_SCHED
8617 * Ensure that the real time constraints are schedulable.
8619 static DEFINE_MUTEX(rt_constraints_mutex
);
8621 static unsigned long to_ratio(u64 period
, u64 runtime
)
8623 if (runtime
== RUNTIME_INF
)
8626 return div64_u64(runtime
<< 16, period
);
8629 #ifdef CONFIG_CGROUP_SCHED
8630 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8632 struct task_group
*tgi
, *parent
= tg
->parent
;
8633 unsigned long total
= 0;
8636 if (global_rt_period() < period
)
8639 return to_ratio(period
, runtime
) <
8640 to_ratio(global_rt_period(), global_rt_runtime());
8643 if (ktime_to_ns(parent
->rt_bandwidth
.rt_period
) < period
)
8647 list_for_each_entry_rcu(tgi
, &parent
->children
, siblings
) {
8651 total
+= to_ratio(ktime_to_ns(tgi
->rt_bandwidth
.rt_period
),
8652 tgi
->rt_bandwidth
.rt_runtime
);
8656 return total
+ to_ratio(period
, runtime
) <=
8657 to_ratio(ktime_to_ns(parent
->rt_bandwidth
.rt_period
),
8658 parent
->rt_bandwidth
.rt_runtime
);
8660 #elif defined CONFIG_USER_SCHED
8661 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8663 struct task_group
*tgi
;
8664 unsigned long total
= 0;
8665 unsigned long global_ratio
=
8666 to_ratio(global_rt_period(), global_rt_runtime());
8669 list_for_each_entry_rcu(tgi
, &task_groups
, list
) {
8673 total
+= to_ratio(ktime_to_ns(tgi
->rt_bandwidth
.rt_period
),
8674 tgi
->rt_bandwidth
.rt_runtime
);
8678 return total
+ to_ratio(period
, runtime
) < global_ratio
;
8682 /* Must be called with tasklist_lock held */
8683 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8685 struct task_struct
*g
, *p
;
8686 do_each_thread(g
, p
) {
8687 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8689 } while_each_thread(g
, p
);
8693 static int tg_set_bandwidth(struct task_group
*tg
,
8694 u64 rt_period
, u64 rt_runtime
)
8698 mutex_lock(&rt_constraints_mutex
);
8699 read_lock(&tasklist_lock
);
8700 if (rt_runtime
== 0 && tg_has_rt_tasks(tg
)) {
8704 if (!__rt_schedulable(tg
, rt_period
, rt_runtime
)) {
8709 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8710 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8711 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8713 for_each_possible_cpu(i
) {
8714 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8716 spin_lock(&rt_rq
->rt_runtime_lock
);
8717 rt_rq
->rt_runtime
= rt_runtime
;
8718 spin_unlock(&rt_rq
->rt_runtime_lock
);
8720 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8722 read_unlock(&tasklist_lock
);
8723 mutex_unlock(&rt_constraints_mutex
);
8728 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8730 u64 rt_runtime
, rt_period
;
8732 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8733 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8734 if (rt_runtime_us
< 0)
8735 rt_runtime
= RUNTIME_INF
;
8737 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8740 long sched_group_rt_runtime(struct task_group
*tg
)
8744 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8747 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8748 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8749 return rt_runtime_us
;
8752 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8754 u64 rt_runtime
, rt_period
;
8756 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8757 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8762 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8765 long sched_group_rt_period(struct task_group
*tg
)
8769 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8770 do_div(rt_period_us
, NSEC_PER_USEC
);
8771 return rt_period_us
;
8774 static int sched_rt_global_constraints(void)
8776 struct task_group
*tg
= &root_task_group
;
8777 u64 rt_runtime
, rt_period
;
8780 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8781 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8783 mutex_lock(&rt_constraints_mutex
);
8784 if (!__rt_schedulable(tg
, rt_period
, rt_runtime
))
8786 mutex_unlock(&rt_constraints_mutex
);
8790 #else /* !CONFIG_RT_GROUP_SCHED */
8791 static int sched_rt_global_constraints(void)
8793 unsigned long flags
;
8796 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8797 for_each_possible_cpu(i
) {
8798 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
8800 spin_lock(&rt_rq
->rt_runtime_lock
);
8801 rt_rq
->rt_runtime
= global_rt_runtime();
8802 spin_unlock(&rt_rq
->rt_runtime_lock
);
8804 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8808 #endif /* CONFIG_RT_GROUP_SCHED */
8810 int sched_rt_handler(struct ctl_table
*table
, int write
,
8811 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
8815 int old_period
, old_runtime
;
8816 static DEFINE_MUTEX(mutex
);
8819 old_period
= sysctl_sched_rt_period
;
8820 old_runtime
= sysctl_sched_rt_runtime
;
8822 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
8824 if (!ret
&& write
) {
8825 ret
= sched_rt_global_constraints();
8827 sysctl_sched_rt_period
= old_period
;
8828 sysctl_sched_rt_runtime
= old_runtime
;
8830 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8831 def_rt_bandwidth
.rt_period
=
8832 ns_to_ktime(global_rt_period());
8835 mutex_unlock(&mutex
);
8840 #ifdef CONFIG_CGROUP_SCHED
8842 /* return corresponding task_group object of a cgroup */
8843 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
8845 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
8846 struct task_group
, css
);
8849 static struct cgroup_subsys_state
*
8850 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8852 struct task_group
*tg
, *parent
;
8854 if (!cgrp
->parent
) {
8855 /* This is early initialization for the top cgroup */
8856 init_task_group
.css
.cgroup
= cgrp
;
8857 return &init_task_group
.css
;
8860 parent
= cgroup_tg(cgrp
->parent
);
8861 tg
= sched_create_group(parent
);
8863 return ERR_PTR(-ENOMEM
);
8865 /* Bind the cgroup to task_group object we just created */
8866 tg
->css
.cgroup
= cgrp
;
8872 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8874 struct task_group
*tg
= cgroup_tg(cgrp
);
8876 sched_destroy_group(tg
);
8880 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8881 struct task_struct
*tsk
)
8883 #ifdef CONFIG_RT_GROUP_SCHED
8884 /* Don't accept realtime tasks when there is no way for them to run */
8885 if (rt_task(tsk
) && cgroup_tg(cgrp
)->rt_bandwidth
.rt_runtime
== 0)
8888 /* We don't support RT-tasks being in separate groups */
8889 if (tsk
->sched_class
!= &fair_sched_class
)
8897 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8898 struct cgroup
*old_cont
, struct task_struct
*tsk
)
8900 sched_move_task(tsk
);
8903 #ifdef CONFIG_FAIR_GROUP_SCHED
8904 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
8907 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
8910 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
8912 struct task_group
*tg
= cgroup_tg(cgrp
);
8914 return (u64
) tg
->shares
;
8916 #endif /* CONFIG_FAIR_GROUP_SCHED */
8918 #ifdef CONFIG_RT_GROUP_SCHED
8919 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
8922 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
8925 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8927 return sched_group_rt_runtime(cgroup_tg(cgrp
));
8930 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
8933 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
8936 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
8938 return sched_group_rt_period(cgroup_tg(cgrp
));
8940 #endif /* CONFIG_RT_GROUP_SCHED */
8942 static struct cftype cpu_files
[] = {
8943 #ifdef CONFIG_FAIR_GROUP_SCHED
8946 .read_u64
= cpu_shares_read_u64
,
8947 .write_u64
= cpu_shares_write_u64
,
8950 #ifdef CONFIG_RT_GROUP_SCHED
8952 .name
= "rt_runtime_us",
8953 .read_s64
= cpu_rt_runtime_read
,
8954 .write_s64
= cpu_rt_runtime_write
,
8957 .name
= "rt_period_us",
8958 .read_u64
= cpu_rt_period_read_uint
,
8959 .write_u64
= cpu_rt_period_write_uint
,
8964 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8966 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
8969 struct cgroup_subsys cpu_cgroup_subsys
= {
8971 .create
= cpu_cgroup_create
,
8972 .destroy
= cpu_cgroup_destroy
,
8973 .can_attach
= cpu_cgroup_can_attach
,
8974 .attach
= cpu_cgroup_attach
,
8975 .populate
= cpu_cgroup_populate
,
8976 .subsys_id
= cpu_cgroup_subsys_id
,
8980 #endif /* CONFIG_CGROUP_SCHED */
8982 #ifdef CONFIG_CGROUP_CPUACCT
8985 * CPU accounting code for task groups.
8987 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8988 * (balbir@in.ibm.com).
8991 /* track cpu usage of a group of tasks */
8993 struct cgroup_subsys_state css
;
8994 /* cpuusage holds pointer to a u64-type object on every cpu */
8998 struct cgroup_subsys cpuacct_subsys
;
9000 /* return cpu accounting group corresponding to this container */
9001 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
9003 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
9004 struct cpuacct
, css
);
9007 /* return cpu accounting group to which this task belongs */
9008 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
9010 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
9011 struct cpuacct
, css
);
9014 /* create a new cpu accounting group */
9015 static struct cgroup_subsys_state
*cpuacct_create(
9016 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9018 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
9021 return ERR_PTR(-ENOMEM
);
9023 ca
->cpuusage
= alloc_percpu(u64
);
9024 if (!ca
->cpuusage
) {
9026 return ERR_PTR(-ENOMEM
);
9032 /* destroy an existing cpu accounting group */
9034 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9036 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9038 free_percpu(ca
->cpuusage
);
9042 /* return total cpu usage (in nanoseconds) of a group */
9043 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9045 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9046 u64 totalcpuusage
= 0;
9049 for_each_possible_cpu(i
) {
9050 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
9053 * Take rq->lock to make 64-bit addition safe on 32-bit
9056 spin_lock_irq(&cpu_rq(i
)->lock
);
9057 totalcpuusage
+= *cpuusage
;
9058 spin_unlock_irq(&cpu_rq(i
)->lock
);
9061 return totalcpuusage
;
9064 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
9067 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9076 for_each_possible_cpu(i
) {
9077 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
9079 spin_lock_irq(&cpu_rq(i
)->lock
);
9081 spin_unlock_irq(&cpu_rq(i
)->lock
);
9087 static struct cftype files
[] = {
9090 .read_u64
= cpuusage_read
,
9091 .write_u64
= cpuusage_write
,
9095 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9097 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
9101 * charge this task's execution time to its accounting group.
9103 * called with rq->lock held.
9105 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
9109 if (!cpuacct_subsys
.active
)
9114 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, task_cpu(tsk
));
9116 *cpuusage
+= cputime
;
9120 struct cgroup_subsys cpuacct_subsys
= {
9122 .create
= cpuacct_create
,
9123 .destroy
= cpuacct_destroy
,
9124 .populate
= cpuacct_populate
,
9125 .subsys_id
= cpuacct_subsys_id
,
9127 #endif /* CONFIG_CGROUP_CPUACCT */