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/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/reciprocal_div.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/bootmem.h>
72 #include <linux/debugfs.h>
73 #include <linux/ctype.h>
74 #include <linux/ftrace.h>
75 #include <trace/sched.h>
78 #include <asm/irq_regs.h>
80 #include "sched_cpupri.h"
83 * Convert user-nice values [ -20 ... 0 ... 19 ]
84 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
87 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
88 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
89 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
92 * 'User priority' is the nice value converted to something we
93 * can work with better when scaling various scheduler parameters,
94 * it's a [ 0 ... 39 ] range.
96 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
97 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
98 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
101 * Helpers for converting nanosecond timing to jiffy resolution
103 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
105 #define NICE_0_LOAD SCHED_LOAD_SCALE
106 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
109 * These are the 'tuning knobs' of the scheduler:
111 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
112 * Timeslices get refilled after they expire.
114 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * single value that denotes runtime == period, ie unlimited time.
119 #define RUNTIME_INF ((u64)~0ULL)
123 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
124 * Since cpu_power is a 'constant', we can use a reciprocal divide.
126 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
128 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
132 * Each time a sched group cpu_power is changed,
133 * we must compute its reciprocal value
135 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
137 sg
->__cpu_power
+= val
;
138 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
142 static inline int rt_policy(int policy
)
144 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
149 static inline int task_has_rt_policy(struct task_struct
*p
)
151 return rt_policy(p
->policy
);
155 * This is the priority-queue data structure of the RT scheduling class:
157 struct rt_prio_array
{
158 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
159 struct list_head queue
[MAX_RT_PRIO
];
162 struct rt_bandwidth
{
163 /* nests inside the rq lock: */
164 spinlock_t rt_runtime_lock
;
167 struct hrtimer rt_period_timer
;
170 static struct rt_bandwidth def_rt_bandwidth
;
172 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
174 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
176 struct rt_bandwidth
*rt_b
=
177 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
183 now
= hrtimer_cb_get_time(timer
);
184 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
189 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
192 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
196 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
198 rt_b
->rt_period
= ns_to_ktime(period
);
199 rt_b
->rt_runtime
= runtime
;
201 spin_lock_init(&rt_b
->rt_runtime_lock
);
203 hrtimer_init(&rt_b
->rt_period_timer
,
204 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
205 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
206 rt_b
->rt_period_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_UNLOCKED
;
209 static inline int rt_bandwidth_enabled(void)
211 return sysctl_sched_rt_runtime
>= 0;
214 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
218 if (rt_bandwidth_enabled() && rt_b
->rt_runtime
== RUNTIME_INF
)
221 if (hrtimer_active(&rt_b
->rt_period_timer
))
224 spin_lock(&rt_b
->rt_runtime_lock
);
226 if (hrtimer_active(&rt_b
->rt_period_timer
))
229 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
230 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
231 hrtimer_start_expires(&rt_b
->rt_period_timer
,
234 spin_unlock(&rt_b
->rt_runtime_lock
);
237 #ifdef CONFIG_RT_GROUP_SCHED
238 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
240 hrtimer_cancel(&rt_b
->rt_period_timer
);
245 * sched_domains_mutex serializes calls to arch_init_sched_domains,
246 * detach_destroy_domains and partition_sched_domains.
248 static DEFINE_MUTEX(sched_domains_mutex
);
250 #ifdef CONFIG_GROUP_SCHED
252 #include <linux/cgroup.h>
256 static LIST_HEAD(task_groups
);
258 /* task group related information */
260 #ifdef CONFIG_CGROUP_SCHED
261 struct cgroup_subsys_state css
;
264 #ifdef CONFIG_FAIR_GROUP_SCHED
265 /* schedulable entities of this group on each cpu */
266 struct sched_entity
**se
;
267 /* runqueue "owned" by this group on each cpu */
268 struct cfs_rq
**cfs_rq
;
269 unsigned long shares
;
272 #ifdef CONFIG_RT_GROUP_SCHED
273 struct sched_rt_entity
**rt_se
;
274 struct rt_rq
**rt_rq
;
276 struct rt_bandwidth rt_bandwidth
;
280 struct list_head list
;
282 struct task_group
*parent
;
283 struct list_head siblings
;
284 struct list_head children
;
287 #ifdef CONFIG_USER_SCHED
291 * Every UID task group (including init_task_group aka UID-0) will
292 * be a child to this group.
294 struct task_group root_task_group
;
296 #ifdef CONFIG_FAIR_GROUP_SCHED
297 /* Default task group's sched entity on each cpu */
298 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
299 /* Default task group's cfs_rq on each cpu */
300 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
301 #endif /* CONFIG_FAIR_GROUP_SCHED */
303 #ifdef CONFIG_RT_GROUP_SCHED
304 static DEFINE_PER_CPU(struct sched_rt_entity
, init_sched_rt_entity
);
305 static DEFINE_PER_CPU(struct rt_rq
, init_rt_rq
) ____cacheline_aligned_in_smp
;
306 #endif /* CONFIG_RT_GROUP_SCHED */
307 #else /* !CONFIG_USER_SCHED */
308 #define root_task_group init_task_group
309 #endif /* CONFIG_USER_SCHED */
311 /* task_group_lock serializes add/remove of task groups and also changes to
312 * a task group's cpu shares.
314 static DEFINE_SPINLOCK(task_group_lock
);
316 #ifdef CONFIG_FAIR_GROUP_SCHED
317 #ifdef CONFIG_USER_SCHED
318 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
319 #else /* !CONFIG_USER_SCHED */
320 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
321 #endif /* CONFIG_USER_SCHED */
324 * A weight of 0 or 1 can cause arithmetics problems.
325 * A weight of a cfs_rq is the sum of weights of which entities
326 * are queued on this cfs_rq, so a weight of a entity should not be
327 * too large, so as the shares value of a task group.
328 * (The default weight is 1024 - so there's no practical
329 * limitation from this.)
332 #define MAX_SHARES (1UL << 18)
334 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
337 /* Default task group.
338 * Every task in system belong to this group at bootup.
340 struct task_group init_task_group
;
342 /* return group to which a task belongs */
343 static inline struct task_group
*task_group(struct task_struct
*p
)
345 struct task_group
*tg
;
347 #ifdef CONFIG_USER_SCHED
349 #elif defined(CONFIG_CGROUP_SCHED)
350 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
351 struct task_group
, css
);
353 tg
= &init_task_group
;
358 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
359 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
361 #ifdef CONFIG_FAIR_GROUP_SCHED
362 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
363 p
->se
.parent
= task_group(p
)->se
[cpu
];
366 #ifdef CONFIG_RT_GROUP_SCHED
367 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
368 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
374 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
375 static inline struct task_group
*task_group(struct task_struct
*p
)
380 #endif /* CONFIG_GROUP_SCHED */
382 /* CFS-related fields in a runqueue */
384 struct load_weight load
;
385 unsigned long nr_running
;
390 struct rb_root tasks_timeline
;
391 struct rb_node
*rb_leftmost
;
393 struct list_head tasks
;
394 struct list_head
*balance_iterator
;
397 * 'curr' points to currently running entity on this cfs_rq.
398 * It is set to NULL otherwise (i.e when none are currently running).
400 struct sched_entity
*curr
, *next
;
402 unsigned long nr_spread_over
;
404 #ifdef CONFIG_FAIR_GROUP_SCHED
405 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
408 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
409 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
410 * (like users, containers etc.)
412 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
413 * list is used during load balance.
415 struct list_head leaf_cfs_rq_list
;
416 struct task_group
*tg
; /* group that "owns" this runqueue */
420 * the part of load.weight contributed by tasks
422 unsigned long task_weight
;
425 * h_load = weight * f(tg)
427 * Where f(tg) is the recursive weight fraction assigned to
430 unsigned long h_load
;
433 * this cpu's part of tg->shares
435 unsigned long shares
;
438 * load.weight at the time we set shares
440 unsigned long rq_weight
;
445 /* Real-Time classes' related field in a runqueue: */
447 struct rt_prio_array active
;
448 unsigned long rt_nr_running
;
449 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
450 int highest_prio
; /* highest queued rt task prio */
453 unsigned long rt_nr_migratory
;
459 /* Nests inside the rq lock: */
460 spinlock_t rt_runtime_lock
;
462 #ifdef CONFIG_RT_GROUP_SCHED
463 unsigned long rt_nr_boosted
;
466 struct list_head leaf_rt_rq_list
;
467 struct task_group
*tg
;
468 struct sched_rt_entity
*rt_se
;
475 * We add the notion of a root-domain which will be used to define per-domain
476 * variables. Each exclusive cpuset essentially defines an island domain by
477 * fully partitioning the member cpus from any other cpuset. Whenever a new
478 * exclusive cpuset is created, we also create and attach a new root-domain
488 * The "RT overload" flag: it gets set if a CPU has more than
489 * one runnable RT task.
494 struct cpupri cpupri
;
499 * By default the system creates a single root-domain with all cpus as
500 * members (mimicking the global state we have today).
502 static struct root_domain def_root_domain
;
507 * This is the main, per-CPU runqueue data structure.
509 * Locking rule: those places that want to lock multiple runqueues
510 * (such as the load balancing or the thread migration code), lock
511 * acquire operations must be ordered by ascending &runqueue.
518 * nr_running and cpu_load should be in the same cacheline because
519 * remote CPUs use both these fields when doing load calculation.
521 unsigned long nr_running
;
522 #define CPU_LOAD_IDX_MAX 5
523 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
524 unsigned char idle_at_tick
;
526 unsigned long last_tick_seen
;
527 unsigned char in_nohz_recently
;
529 /* capture load from *all* tasks on this cpu: */
530 struct load_weight load
;
531 unsigned long nr_load_updates
;
537 #ifdef CONFIG_FAIR_GROUP_SCHED
538 /* list of leaf cfs_rq on this cpu: */
539 struct list_head leaf_cfs_rq_list
;
541 #ifdef CONFIG_RT_GROUP_SCHED
542 struct list_head leaf_rt_rq_list
;
546 * This is part of a global counter where only the total sum
547 * over all CPUs matters. A task can increase this counter on
548 * one CPU and if it got migrated afterwards it may decrease
549 * it on another CPU. Always updated under the runqueue lock:
551 unsigned long nr_uninterruptible
;
553 struct task_struct
*curr
, *idle
;
554 unsigned long next_balance
;
555 struct mm_struct
*prev_mm
;
562 struct root_domain
*rd
;
563 struct sched_domain
*sd
;
565 /* For active balancing */
568 /* cpu of this runqueue: */
572 unsigned long avg_load_per_task
;
574 struct task_struct
*migration_thread
;
575 struct list_head migration_queue
;
578 #ifdef CONFIG_SCHED_HRTICK
580 int hrtick_csd_pending
;
581 struct call_single_data hrtick_csd
;
583 struct hrtimer hrtick_timer
;
586 #ifdef CONFIG_SCHEDSTATS
588 struct sched_info rq_sched_info
;
590 /* sys_sched_yield() stats */
591 unsigned int yld_exp_empty
;
592 unsigned int yld_act_empty
;
593 unsigned int yld_both_empty
;
594 unsigned int yld_count
;
596 /* schedule() stats */
597 unsigned int sched_switch
;
598 unsigned int sched_count
;
599 unsigned int sched_goidle
;
601 /* try_to_wake_up() stats */
602 unsigned int ttwu_count
;
603 unsigned int ttwu_local
;
606 unsigned int bkl_count
;
610 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
612 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int sync
)
614 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, sync
);
617 static inline int cpu_of(struct rq
*rq
)
627 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
628 * See detach_destroy_domains: synchronize_sched for details.
630 * The domain tree of any CPU may only be accessed from within
631 * preempt-disabled sections.
633 #define for_each_domain(cpu, __sd) \
634 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
636 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
637 #define this_rq() (&__get_cpu_var(runqueues))
638 #define task_rq(p) cpu_rq(task_cpu(p))
639 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
641 static inline void update_rq_clock(struct rq
*rq
)
643 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
647 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
649 #ifdef CONFIG_SCHED_DEBUG
650 # define const_debug __read_mostly
652 # define const_debug static const
658 * Returns true if the current cpu runqueue is locked.
659 * This interface allows printk to be called with the runqueue lock
660 * held and know whether or not it is OK to wake up the klogd.
662 int runqueue_is_locked(void)
665 struct rq
*rq
= cpu_rq(cpu
);
668 ret
= spin_is_locked(&rq
->lock
);
674 * Debugging: various feature bits
677 #define SCHED_FEAT(name, enabled) \
678 __SCHED_FEAT_##name ,
681 #include "sched_features.h"
686 #define SCHED_FEAT(name, enabled) \
687 (1UL << __SCHED_FEAT_##name) * enabled |
689 const_debug
unsigned int sysctl_sched_features
=
690 #include "sched_features.h"
695 #ifdef CONFIG_SCHED_DEBUG
696 #define SCHED_FEAT(name, enabled) \
699 static __read_mostly
char *sched_feat_names
[] = {
700 #include "sched_features.h"
706 static int sched_feat_show(struct seq_file
*m
, void *v
)
710 for (i
= 0; sched_feat_names
[i
]; i
++) {
711 if (!(sysctl_sched_features
& (1UL << i
)))
713 seq_printf(m
, "%s ", sched_feat_names
[i
]);
721 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
722 size_t cnt
, loff_t
*ppos
)
732 if (copy_from_user(&buf
, ubuf
, cnt
))
737 if (strncmp(buf
, "NO_", 3) == 0) {
742 for (i
= 0; sched_feat_names
[i
]; i
++) {
743 int len
= strlen(sched_feat_names
[i
]);
745 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
747 sysctl_sched_features
&= ~(1UL << i
);
749 sysctl_sched_features
|= (1UL << i
);
754 if (!sched_feat_names
[i
])
762 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
764 return single_open(filp
, sched_feat_show
, NULL
);
767 static struct file_operations sched_feat_fops
= {
768 .open
= sched_feat_open
,
769 .write
= sched_feat_write
,
772 .release
= single_release
,
775 static __init
int sched_init_debug(void)
777 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
782 late_initcall(sched_init_debug
);
786 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
789 * Number of tasks to iterate in a single balance run.
790 * Limited because this is done with IRQs disabled.
792 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
795 * ratelimit for updating the group shares.
798 unsigned int sysctl_sched_shares_ratelimit
= 250000;
801 * Inject some fuzzyness into changing the per-cpu group shares
802 * this avoids remote rq-locks at the expense of fairness.
805 unsigned int sysctl_sched_shares_thresh
= 4;
808 * period over which we measure -rt task cpu usage in us.
811 unsigned int sysctl_sched_rt_period
= 1000000;
813 static __read_mostly
int scheduler_running
;
816 * part of the period that we allow rt tasks to run in us.
819 int sysctl_sched_rt_runtime
= 950000;
821 static inline u64
global_rt_period(void)
823 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
826 static inline u64
global_rt_runtime(void)
828 if (sysctl_sched_rt_runtime
< 0)
831 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
834 #ifndef prepare_arch_switch
835 # define prepare_arch_switch(next) do { } while (0)
837 #ifndef finish_arch_switch
838 # define finish_arch_switch(prev) do { } while (0)
841 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
843 return rq
->curr
== p
;
846 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
847 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
849 return task_current(rq
, p
);
852 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
856 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
858 #ifdef CONFIG_DEBUG_SPINLOCK
859 /* this is a valid case when another task releases the spinlock */
860 rq
->lock
.owner
= current
;
863 * If we are tracking spinlock dependencies then we have to
864 * fix up the runqueue lock - which gets 'carried over' from
867 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
869 spin_unlock_irq(&rq
->lock
);
872 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
873 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
878 return task_current(rq
, p
);
882 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
886 * We can optimise this out completely for !SMP, because the
887 * SMP rebalancing from interrupt is the only thing that cares
892 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
893 spin_unlock_irq(&rq
->lock
);
895 spin_unlock(&rq
->lock
);
899 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
903 * After ->oncpu is cleared, the task can be moved to a different CPU.
904 * We must ensure this doesn't happen until the switch is completely
910 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
914 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
917 * __task_rq_lock - lock the runqueue a given task resides on.
918 * Must be called interrupts disabled.
920 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
924 struct rq
*rq
= task_rq(p
);
925 spin_lock(&rq
->lock
);
926 if (likely(rq
== task_rq(p
)))
928 spin_unlock(&rq
->lock
);
933 * task_rq_lock - lock the runqueue a given task resides on and disable
934 * interrupts. Note the ordering: we can safely lookup the task_rq without
935 * explicitly disabling preemption.
937 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
943 local_irq_save(*flags
);
945 spin_lock(&rq
->lock
);
946 if (likely(rq
== task_rq(p
)))
948 spin_unlock_irqrestore(&rq
->lock
, *flags
);
952 static void __task_rq_unlock(struct rq
*rq
)
955 spin_unlock(&rq
->lock
);
958 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
961 spin_unlock_irqrestore(&rq
->lock
, *flags
);
965 * this_rq_lock - lock this runqueue and disable interrupts.
967 static struct rq
*this_rq_lock(void)
974 spin_lock(&rq
->lock
);
979 #ifdef CONFIG_SCHED_HRTICK
981 * Use HR-timers to deliver accurate preemption points.
983 * Its all a bit involved since we cannot program an hrt while holding the
984 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
987 * When we get rescheduled we reprogram the hrtick_timer outside of the
993 * - enabled by features
994 * - hrtimer is actually high res
996 static inline int hrtick_enabled(struct rq
*rq
)
998 if (!sched_feat(HRTICK
))
1000 if (!cpu_active(cpu_of(rq
)))
1002 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1005 static void hrtick_clear(struct rq
*rq
)
1007 if (hrtimer_active(&rq
->hrtick_timer
))
1008 hrtimer_cancel(&rq
->hrtick_timer
);
1012 * High-resolution timer tick.
1013 * Runs from hardirq context with interrupts disabled.
1015 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1017 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1019 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1021 spin_lock(&rq
->lock
);
1022 update_rq_clock(rq
);
1023 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1024 spin_unlock(&rq
->lock
);
1026 return HRTIMER_NORESTART
;
1031 * called from hardirq (IPI) context
1033 static void __hrtick_start(void *arg
)
1035 struct rq
*rq
= arg
;
1037 spin_lock(&rq
->lock
);
1038 hrtimer_restart(&rq
->hrtick_timer
);
1039 rq
->hrtick_csd_pending
= 0;
1040 spin_unlock(&rq
->lock
);
1044 * Called to set the hrtick timer state.
1046 * called with rq->lock held and irqs disabled
1048 static void hrtick_start(struct rq
*rq
, u64 delay
)
1050 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1051 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1053 hrtimer_set_expires(timer
, time
);
1055 if (rq
== this_rq()) {
1056 hrtimer_restart(timer
);
1057 } else if (!rq
->hrtick_csd_pending
) {
1058 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
);
1059 rq
->hrtick_csd_pending
= 1;
1064 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1066 int cpu
= (int)(long)hcpu
;
1069 case CPU_UP_CANCELED
:
1070 case CPU_UP_CANCELED_FROZEN
:
1071 case CPU_DOWN_PREPARE
:
1072 case CPU_DOWN_PREPARE_FROZEN
:
1074 case CPU_DEAD_FROZEN
:
1075 hrtick_clear(cpu_rq(cpu
));
1082 static __init
void init_hrtick(void)
1084 hotcpu_notifier(hotplug_hrtick
, 0);
1088 * Called to set the hrtick timer state.
1090 * called with rq->lock held and irqs disabled
1092 static void hrtick_start(struct rq
*rq
, u64 delay
)
1094 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
), HRTIMER_MODE_REL
);
1097 static inline void init_hrtick(void)
1100 #endif /* CONFIG_SMP */
1102 static void init_rq_hrtick(struct rq
*rq
)
1105 rq
->hrtick_csd_pending
= 0;
1107 rq
->hrtick_csd
.flags
= 0;
1108 rq
->hrtick_csd
.func
= __hrtick_start
;
1109 rq
->hrtick_csd
.info
= rq
;
1112 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1113 rq
->hrtick_timer
.function
= hrtick
;
1114 rq
->hrtick_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_PERCPU
;
1116 #else /* CONFIG_SCHED_HRTICK */
1117 static inline void hrtick_clear(struct rq
*rq
)
1121 static inline void init_rq_hrtick(struct rq
*rq
)
1125 static inline void init_hrtick(void)
1128 #endif /* CONFIG_SCHED_HRTICK */
1131 * resched_task - mark a task 'to be rescheduled now'.
1133 * On UP this means the setting of the need_resched flag, on SMP it
1134 * might also involve a cross-CPU call to trigger the scheduler on
1139 #ifndef tsk_is_polling
1140 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1143 static void resched_task(struct task_struct
*p
)
1147 assert_spin_locked(&task_rq(p
)->lock
);
1149 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
1152 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
1155 if (cpu
== smp_processor_id())
1158 /* NEED_RESCHED must be visible before we test polling */
1160 if (!tsk_is_polling(p
))
1161 smp_send_reschedule(cpu
);
1164 static void resched_cpu(int cpu
)
1166 struct rq
*rq
= cpu_rq(cpu
);
1167 unsigned long flags
;
1169 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1171 resched_task(cpu_curr(cpu
));
1172 spin_unlock_irqrestore(&rq
->lock
, flags
);
1177 * When add_timer_on() enqueues a timer into the timer wheel of an
1178 * idle CPU then this timer might expire before the next timer event
1179 * which is scheduled to wake up that CPU. In case of a completely
1180 * idle system the next event might even be infinite time into the
1181 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1182 * leaves the inner idle loop so the newly added timer is taken into
1183 * account when the CPU goes back to idle and evaluates the timer
1184 * wheel for the next timer event.
1186 void wake_up_idle_cpu(int cpu
)
1188 struct rq
*rq
= cpu_rq(cpu
);
1190 if (cpu
== smp_processor_id())
1194 * This is safe, as this function is called with the timer
1195 * wheel base lock of (cpu) held. When the CPU is on the way
1196 * to idle and has not yet set rq->curr to idle then it will
1197 * be serialized on the timer wheel base lock and take the new
1198 * timer into account automatically.
1200 if (rq
->curr
!= rq
->idle
)
1204 * We can set TIF_RESCHED on the idle task of the other CPU
1205 * lockless. The worst case is that the other CPU runs the
1206 * idle task through an additional NOOP schedule()
1208 set_tsk_thread_flag(rq
->idle
, TIF_NEED_RESCHED
);
1210 /* NEED_RESCHED must be visible before we test polling */
1212 if (!tsk_is_polling(rq
->idle
))
1213 smp_send_reschedule(cpu
);
1215 #endif /* CONFIG_NO_HZ */
1217 #else /* !CONFIG_SMP */
1218 static void resched_task(struct task_struct
*p
)
1220 assert_spin_locked(&task_rq(p
)->lock
);
1221 set_tsk_need_resched(p
);
1223 #endif /* CONFIG_SMP */
1225 #if BITS_PER_LONG == 32
1226 # define WMULT_CONST (~0UL)
1228 # define WMULT_CONST (1UL << 32)
1231 #define WMULT_SHIFT 32
1234 * Shift right and round:
1236 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1239 * delta *= weight / lw
1241 static unsigned long
1242 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1243 struct load_weight
*lw
)
1247 if (!lw
->inv_weight
) {
1248 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1251 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1255 tmp
= (u64
)delta_exec
* weight
;
1257 * Check whether we'd overflow the 64-bit multiplication:
1259 if (unlikely(tmp
> WMULT_CONST
))
1260 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1263 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1265 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1268 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1274 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1281 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1282 * of tasks with abnormal "nice" values across CPUs the contribution that
1283 * each task makes to its run queue's load is weighted according to its
1284 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1285 * scaled version of the new time slice allocation that they receive on time
1289 #define WEIGHT_IDLEPRIO 2
1290 #define WMULT_IDLEPRIO (1 << 31)
1293 * Nice levels are multiplicative, with a gentle 10% change for every
1294 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1295 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1296 * that remained on nice 0.
1298 * The "10% effect" is relative and cumulative: from _any_ nice level,
1299 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1300 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1301 * If a task goes up by ~10% and another task goes down by ~10% then
1302 * the relative distance between them is ~25%.)
1304 static const int prio_to_weight
[40] = {
1305 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1306 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1307 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1308 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1309 /* 0 */ 1024, 820, 655, 526, 423,
1310 /* 5 */ 335, 272, 215, 172, 137,
1311 /* 10 */ 110, 87, 70, 56, 45,
1312 /* 15 */ 36, 29, 23, 18, 15,
1316 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1318 * In cases where the weight does not change often, we can use the
1319 * precalculated inverse to speed up arithmetics by turning divisions
1320 * into multiplications:
1322 static const u32 prio_to_wmult
[40] = {
1323 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1324 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1325 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1326 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1327 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1328 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1329 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1330 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1333 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1336 * runqueue iterator, to support SMP load-balancing between different
1337 * scheduling classes, without having to expose their internal data
1338 * structures to the load-balancing proper:
1340 struct rq_iterator
{
1342 struct task_struct
*(*start
)(void *);
1343 struct task_struct
*(*next
)(void *);
1347 static unsigned long
1348 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1349 unsigned long max_load_move
, struct sched_domain
*sd
,
1350 enum cpu_idle_type idle
, int *all_pinned
,
1351 int *this_best_prio
, struct rq_iterator
*iterator
);
1354 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1355 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1356 struct rq_iterator
*iterator
);
1359 #ifdef CONFIG_CGROUP_CPUACCT
1360 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1362 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1365 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1367 update_load_add(&rq
->load
, load
);
1370 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1372 update_load_sub(&rq
->load
, load
);
1375 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1376 typedef int (*tg_visitor
)(struct task_group
*, void *);
1379 * Iterate the full tree, calling @down when first entering a node and @up when
1380 * leaving it for the final time.
1382 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1384 struct task_group
*parent
, *child
;
1388 parent
= &root_task_group
;
1390 ret
= (*down
)(parent
, data
);
1393 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1400 ret
= (*up
)(parent
, data
);
1405 parent
= parent
->parent
;
1414 static int tg_nop(struct task_group
*tg
, void *data
)
1421 static unsigned long source_load(int cpu
, int type
);
1422 static unsigned long target_load(int cpu
, int type
);
1423 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1425 static unsigned long cpu_avg_load_per_task(int cpu
)
1427 struct rq
*rq
= cpu_rq(cpu
);
1430 rq
->avg_load_per_task
= rq
->load
.weight
/ rq
->nr_running
;
1432 return rq
->avg_load_per_task
;
1435 #ifdef CONFIG_FAIR_GROUP_SCHED
1437 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1440 * Calculate and set the cpu's group shares.
1443 update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1444 unsigned long sd_shares
, unsigned long sd_rq_weight
)
1447 unsigned long shares
;
1448 unsigned long rq_weight
;
1453 rq_weight
= tg
->cfs_rq
[cpu
]->load
.weight
;
1456 * If there are currently no tasks on the cpu pretend there is one of
1457 * average load so that when a new task gets to run here it will not
1458 * get delayed by group starvation.
1462 rq_weight
= NICE_0_LOAD
;
1465 if (unlikely(rq_weight
> sd_rq_weight
))
1466 rq_weight
= sd_rq_weight
;
1469 * \Sum shares * rq_weight
1470 * shares = -----------------------
1474 shares
= (sd_shares
* rq_weight
) / (sd_rq_weight
+ 1);
1475 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1477 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1478 sysctl_sched_shares_thresh
) {
1479 struct rq
*rq
= cpu_rq(cpu
);
1480 unsigned long flags
;
1482 spin_lock_irqsave(&rq
->lock
, flags
);
1484 * record the actual number of shares, not the boosted amount.
1486 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1487 tg
->cfs_rq
[cpu
]->rq_weight
= rq_weight
;
1489 __set_se_shares(tg
->se
[cpu
], shares
);
1490 spin_unlock_irqrestore(&rq
->lock
, flags
);
1495 * Re-compute the task group their per cpu shares over the given domain.
1496 * This needs to be done in a bottom-up fashion because the rq weight of a
1497 * parent group depends on the shares of its child groups.
1499 static int tg_shares_up(struct task_group
*tg
, void *data
)
1501 unsigned long rq_weight
= 0;
1502 unsigned long shares
= 0;
1503 struct sched_domain
*sd
= data
;
1506 for_each_cpu_mask(i
, sd
->span
) {
1507 rq_weight
+= tg
->cfs_rq
[i
]->load
.weight
;
1508 shares
+= tg
->cfs_rq
[i
]->shares
;
1511 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1512 shares
= tg
->shares
;
1514 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1515 shares
= tg
->shares
;
1518 rq_weight
= cpus_weight(sd
->span
) * NICE_0_LOAD
;
1520 for_each_cpu_mask(i
, sd
->span
)
1521 update_group_shares_cpu(tg
, i
, shares
, rq_weight
);
1527 * Compute the cpu's hierarchical load factor for each task group.
1528 * This needs to be done in a top-down fashion because the load of a child
1529 * group is a fraction of its parents load.
1531 static int tg_load_down(struct task_group
*tg
, void *data
)
1534 long cpu
= (long)data
;
1537 load
= cpu_rq(cpu
)->load
.weight
;
1539 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1540 load
*= tg
->cfs_rq
[cpu
]->shares
;
1541 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1544 tg
->cfs_rq
[cpu
]->h_load
= load
;
1549 static void update_shares(struct sched_domain
*sd
)
1551 u64 now
= cpu_clock(raw_smp_processor_id());
1552 s64 elapsed
= now
- sd
->last_update
;
1554 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1555 sd
->last_update
= now
;
1556 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1560 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1562 spin_unlock(&rq
->lock
);
1564 spin_lock(&rq
->lock
);
1567 static void update_h_load(long cpu
)
1569 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1574 static inline void update_shares(struct sched_domain
*sd
)
1578 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1586 #ifdef CONFIG_FAIR_GROUP_SCHED
1587 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1590 cfs_rq
->shares
= shares
;
1595 #include "sched_stats.h"
1596 #include "sched_idletask.c"
1597 #include "sched_fair.c"
1598 #include "sched_rt.c"
1599 #ifdef CONFIG_SCHED_DEBUG
1600 # include "sched_debug.c"
1603 #define sched_class_highest (&rt_sched_class)
1604 #define for_each_class(class) \
1605 for (class = sched_class_highest; class; class = class->next)
1607 static void inc_nr_running(struct rq
*rq
)
1612 static void dec_nr_running(struct rq
*rq
)
1617 static void set_load_weight(struct task_struct
*p
)
1619 if (task_has_rt_policy(p
)) {
1620 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1621 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1626 * SCHED_IDLE tasks get minimal weight:
1628 if (p
->policy
== SCHED_IDLE
) {
1629 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1630 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1634 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1635 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1638 static void update_avg(u64
*avg
, u64 sample
)
1640 s64 diff
= sample
- *avg
;
1644 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1646 sched_info_queued(p
);
1647 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1651 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1653 if (sleep
&& p
->se
.last_wakeup
) {
1654 update_avg(&p
->se
.avg_overlap
,
1655 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1656 p
->se
.last_wakeup
= 0;
1659 sched_info_dequeued(p
);
1660 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1665 * __normal_prio - return the priority that is based on the static prio
1667 static inline int __normal_prio(struct task_struct
*p
)
1669 return p
->static_prio
;
1673 * Calculate the expected normal priority: i.e. priority
1674 * without taking RT-inheritance into account. Might be
1675 * boosted by interactivity modifiers. Changes upon fork,
1676 * setprio syscalls, and whenever the interactivity
1677 * estimator recalculates.
1679 static inline int normal_prio(struct task_struct
*p
)
1683 if (task_has_rt_policy(p
))
1684 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1686 prio
= __normal_prio(p
);
1691 * Calculate the current priority, i.e. the priority
1692 * taken into account by the scheduler. This value might
1693 * be boosted by RT tasks, or might be boosted by
1694 * interactivity modifiers. Will be RT if the task got
1695 * RT-boosted. If not then it returns p->normal_prio.
1697 static int effective_prio(struct task_struct
*p
)
1699 p
->normal_prio
= normal_prio(p
);
1701 * If we are RT tasks or we were boosted to RT priority,
1702 * keep the priority unchanged. Otherwise, update priority
1703 * to the normal priority:
1705 if (!rt_prio(p
->prio
))
1706 return p
->normal_prio
;
1711 * activate_task - move a task to the runqueue.
1713 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1715 if (task_contributes_to_load(p
))
1716 rq
->nr_uninterruptible
--;
1718 enqueue_task(rq
, p
, wakeup
);
1723 * deactivate_task - remove a task from the runqueue.
1725 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1727 if (task_contributes_to_load(p
))
1728 rq
->nr_uninterruptible
++;
1730 dequeue_task(rq
, p
, sleep
);
1735 * task_curr - is this task currently executing on a CPU?
1736 * @p: the task in question.
1738 inline int task_curr(const struct task_struct
*p
)
1740 return cpu_curr(task_cpu(p
)) == p
;
1743 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1745 set_task_rq(p
, cpu
);
1748 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1749 * successfuly executed on another CPU. We must ensure that updates of
1750 * per-task data have been completed by this moment.
1753 task_thread_info(p
)->cpu
= cpu
;
1757 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1758 const struct sched_class
*prev_class
,
1759 int oldprio
, int running
)
1761 if (prev_class
!= p
->sched_class
) {
1762 if (prev_class
->switched_from
)
1763 prev_class
->switched_from(rq
, p
, running
);
1764 p
->sched_class
->switched_to(rq
, p
, running
);
1766 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1771 /* Used instead of source_load when we know the type == 0 */
1772 static unsigned long weighted_cpuload(const int cpu
)
1774 return cpu_rq(cpu
)->load
.weight
;
1778 * Is this task likely cache-hot:
1781 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1786 * Buddy candidates are cache hot:
1788 if (sched_feat(CACHE_HOT_BUDDY
) && (&p
->se
== cfs_rq_of(&p
->se
)->next
))
1791 if (p
->sched_class
!= &fair_sched_class
)
1794 if (sysctl_sched_migration_cost
== -1)
1796 if (sysctl_sched_migration_cost
== 0)
1799 delta
= now
- p
->se
.exec_start
;
1801 return delta
< (s64
)sysctl_sched_migration_cost
;
1805 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1807 int old_cpu
= task_cpu(p
);
1808 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1809 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1810 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1813 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1815 #ifdef CONFIG_SCHEDSTATS
1816 if (p
->se
.wait_start
)
1817 p
->se
.wait_start
-= clock_offset
;
1818 if (p
->se
.sleep_start
)
1819 p
->se
.sleep_start
-= clock_offset
;
1820 if (p
->se
.block_start
)
1821 p
->se
.block_start
-= clock_offset
;
1822 if (old_cpu
!= new_cpu
) {
1823 schedstat_inc(p
, se
.nr_migrations
);
1824 if (task_hot(p
, old_rq
->clock
, NULL
))
1825 schedstat_inc(p
, se
.nr_forced2_migrations
);
1828 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1829 new_cfsrq
->min_vruntime
;
1831 __set_task_cpu(p
, new_cpu
);
1834 struct migration_req
{
1835 struct list_head list
;
1837 struct task_struct
*task
;
1840 struct completion done
;
1844 * The task's runqueue lock must be held.
1845 * Returns true if you have to wait for migration thread.
1848 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1850 struct rq
*rq
= task_rq(p
);
1853 * If the task is not on a runqueue (and not running), then
1854 * it is sufficient to simply update the task's cpu field.
1856 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1857 set_task_cpu(p
, dest_cpu
);
1861 init_completion(&req
->done
);
1863 req
->dest_cpu
= dest_cpu
;
1864 list_add(&req
->list
, &rq
->migration_queue
);
1870 * wait_task_inactive - wait for a thread to unschedule.
1872 * If @match_state is nonzero, it's the @p->state value just checked and
1873 * not expected to change. If it changes, i.e. @p might have woken up,
1874 * then return zero. When we succeed in waiting for @p to be off its CPU,
1875 * we return a positive number (its total switch count). If a second call
1876 * a short while later returns the same number, the caller can be sure that
1877 * @p has remained unscheduled the whole time.
1879 * The caller must ensure that the task *will* unschedule sometime soon,
1880 * else this function might spin for a *long* time. This function can't
1881 * be called with interrupts off, or it may introduce deadlock with
1882 * smp_call_function() if an IPI is sent by the same process we are
1883 * waiting to become inactive.
1885 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1887 unsigned long flags
;
1894 * We do the initial early heuristics without holding
1895 * any task-queue locks at all. We'll only try to get
1896 * the runqueue lock when things look like they will
1902 * If the task is actively running on another CPU
1903 * still, just relax and busy-wait without holding
1906 * NOTE! Since we don't hold any locks, it's not
1907 * even sure that "rq" stays as the right runqueue!
1908 * But we don't care, since "task_running()" will
1909 * return false if the runqueue has changed and p
1910 * is actually now running somewhere else!
1912 while (task_running(rq
, p
)) {
1913 if (match_state
&& unlikely(p
->state
!= match_state
))
1919 * Ok, time to look more closely! We need the rq
1920 * lock now, to be *sure*. If we're wrong, we'll
1921 * just go back and repeat.
1923 rq
= task_rq_lock(p
, &flags
);
1924 trace_sched_wait_task(rq
, p
);
1925 running
= task_running(rq
, p
);
1926 on_rq
= p
->se
.on_rq
;
1928 if (!match_state
|| p
->state
== match_state
)
1929 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1930 task_rq_unlock(rq
, &flags
);
1933 * If it changed from the expected state, bail out now.
1935 if (unlikely(!ncsw
))
1939 * Was it really running after all now that we
1940 * checked with the proper locks actually held?
1942 * Oops. Go back and try again..
1944 if (unlikely(running
)) {
1950 * It's not enough that it's not actively running,
1951 * it must be off the runqueue _entirely_, and not
1954 * So if it wa still runnable (but just not actively
1955 * running right now), it's preempted, and we should
1956 * yield - it could be a while.
1958 if (unlikely(on_rq
)) {
1959 schedule_timeout_uninterruptible(1);
1964 * Ahh, all good. It wasn't running, and it wasn't
1965 * runnable, which means that it will never become
1966 * running in the future either. We're all done!
1975 * kick_process - kick a running thread to enter/exit the kernel
1976 * @p: the to-be-kicked thread
1978 * Cause a process which is running on another CPU to enter
1979 * kernel-mode, without any delay. (to get signals handled.)
1981 * NOTE: this function doesnt have to take the runqueue lock,
1982 * because all it wants to ensure is that the remote task enters
1983 * the kernel. If the IPI races and the task has been migrated
1984 * to another CPU then no harm is done and the purpose has been
1987 void kick_process(struct task_struct
*p
)
1993 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1994 smp_send_reschedule(cpu
);
1999 * Return a low guess at the load of a migration-source cpu weighted
2000 * according to the scheduling class and "nice" value.
2002 * We want to under-estimate the load of migration sources, to
2003 * balance conservatively.
2005 static unsigned long source_load(int cpu
, int type
)
2007 struct rq
*rq
= cpu_rq(cpu
);
2008 unsigned long total
= weighted_cpuload(cpu
);
2010 if (type
== 0 || !sched_feat(LB_BIAS
))
2013 return min(rq
->cpu_load
[type
-1], total
);
2017 * Return a high guess at the load of a migration-target cpu weighted
2018 * according to the scheduling class and "nice" value.
2020 static unsigned long target_load(int cpu
, int type
)
2022 struct rq
*rq
= cpu_rq(cpu
);
2023 unsigned long total
= weighted_cpuload(cpu
);
2025 if (type
== 0 || !sched_feat(LB_BIAS
))
2028 return max(rq
->cpu_load
[type
-1], total
);
2032 * find_idlest_group finds and returns the least busy CPU group within the
2035 static struct sched_group
*
2036 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
2038 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2039 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
2040 int load_idx
= sd
->forkexec_idx
;
2041 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
2044 unsigned long load
, avg_load
;
2048 /* Skip over this group if it has no CPUs allowed */
2049 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
2052 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2054 /* Tally up the load of all CPUs in the group */
2057 for_each_cpu_mask_nr(i
, group
->cpumask
) {
2058 /* Bias balancing toward cpus of our domain */
2060 load
= source_load(i
, load_idx
);
2062 load
= target_load(i
, load_idx
);
2067 /* Adjust by relative CPU power of the group */
2068 avg_load
= sg_div_cpu_power(group
,
2069 avg_load
* SCHED_LOAD_SCALE
);
2072 this_load
= avg_load
;
2074 } else if (avg_load
< min_load
) {
2075 min_load
= avg_load
;
2078 } while (group
= group
->next
, group
!= sd
->groups
);
2080 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
2086 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2089 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
,
2092 unsigned long load
, min_load
= ULONG_MAX
;
2096 /* Traverse only the allowed CPUs */
2097 cpus_and(*tmp
, group
->cpumask
, p
->cpus_allowed
);
2099 for_each_cpu_mask_nr(i
, *tmp
) {
2100 load
= weighted_cpuload(i
);
2102 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2112 * sched_balance_self: balance the current task (running on cpu) in domains
2113 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2116 * Balance, ie. select the least loaded group.
2118 * Returns the target CPU number, or the same CPU if no balancing is needed.
2120 * preempt must be disabled.
2122 static int sched_balance_self(int cpu
, int flag
)
2124 struct task_struct
*t
= current
;
2125 struct sched_domain
*tmp
, *sd
= NULL
;
2127 for_each_domain(cpu
, tmp
) {
2129 * If power savings logic is enabled for a domain, stop there.
2131 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2133 if (tmp
->flags
& flag
)
2141 cpumask_t span
, tmpmask
;
2142 struct sched_group
*group
;
2143 int new_cpu
, weight
;
2145 if (!(sd
->flags
& flag
)) {
2151 group
= find_idlest_group(sd
, t
, cpu
);
2157 new_cpu
= find_idlest_cpu(group
, t
, cpu
, &tmpmask
);
2158 if (new_cpu
== -1 || new_cpu
== cpu
) {
2159 /* Now try balancing at a lower domain level of cpu */
2164 /* Now try balancing at a lower domain level of new_cpu */
2167 weight
= cpus_weight(span
);
2168 for_each_domain(cpu
, tmp
) {
2169 if (weight
<= cpus_weight(tmp
->span
))
2171 if (tmp
->flags
& flag
)
2174 /* while loop will break here if sd == NULL */
2180 #endif /* CONFIG_SMP */
2183 * try_to_wake_up - wake up a thread
2184 * @p: the to-be-woken-up thread
2185 * @state: the mask of task states that can be woken
2186 * @sync: do a synchronous wakeup?
2188 * Put it on the run-queue if it's not already there. The "current"
2189 * thread is always on the run-queue (except when the actual
2190 * re-schedule is in progress), and as such you're allowed to do
2191 * the simpler "current->state = TASK_RUNNING" to mark yourself
2192 * runnable without the overhead of this.
2194 * returns failure only if the task is already active.
2196 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
2198 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2199 unsigned long flags
;
2203 if (!sched_feat(SYNC_WAKEUPS
))
2207 if (sched_feat(LB_WAKEUP_UPDATE
)) {
2208 struct sched_domain
*sd
;
2210 this_cpu
= raw_smp_processor_id();
2213 for_each_domain(this_cpu
, sd
) {
2214 if (cpu_isset(cpu
, sd
->span
)) {
2223 rq
= task_rq_lock(p
, &flags
);
2224 old_state
= p
->state
;
2225 if (!(old_state
& state
))
2233 this_cpu
= smp_processor_id();
2236 if (unlikely(task_running(rq
, p
)))
2239 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2240 if (cpu
!= orig_cpu
) {
2241 set_task_cpu(p
, cpu
);
2242 task_rq_unlock(rq
, &flags
);
2243 /* might preempt at this point */
2244 rq
= task_rq_lock(p
, &flags
);
2245 old_state
= p
->state
;
2246 if (!(old_state
& state
))
2251 this_cpu
= smp_processor_id();
2255 #ifdef CONFIG_SCHEDSTATS
2256 schedstat_inc(rq
, ttwu_count
);
2257 if (cpu
== this_cpu
)
2258 schedstat_inc(rq
, ttwu_local
);
2260 struct sched_domain
*sd
;
2261 for_each_domain(this_cpu
, sd
) {
2262 if (cpu_isset(cpu
, sd
->span
)) {
2263 schedstat_inc(sd
, ttwu_wake_remote
);
2268 #endif /* CONFIG_SCHEDSTATS */
2271 #endif /* CONFIG_SMP */
2272 schedstat_inc(p
, se
.nr_wakeups
);
2274 schedstat_inc(p
, se
.nr_wakeups_sync
);
2275 if (orig_cpu
!= cpu
)
2276 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2277 if (cpu
== this_cpu
)
2278 schedstat_inc(p
, se
.nr_wakeups_local
);
2280 schedstat_inc(p
, se
.nr_wakeups_remote
);
2281 update_rq_clock(rq
);
2282 activate_task(rq
, p
, 1);
2286 trace_sched_wakeup(rq
, p
);
2287 check_preempt_curr(rq
, p
, sync
);
2289 p
->state
= TASK_RUNNING
;
2291 if (p
->sched_class
->task_wake_up
)
2292 p
->sched_class
->task_wake_up(rq
, p
);
2295 current
->se
.last_wakeup
= current
->se
.sum_exec_runtime
;
2297 task_rq_unlock(rq
, &flags
);
2302 int wake_up_process(struct task_struct
*p
)
2304 return try_to_wake_up(p
, TASK_ALL
, 0);
2306 EXPORT_SYMBOL(wake_up_process
);
2308 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2310 return try_to_wake_up(p
, state
, 0);
2314 * Perform scheduler related setup for a newly forked process p.
2315 * p is forked by current.
2317 * __sched_fork() is basic setup used by init_idle() too:
2319 static void __sched_fork(struct task_struct
*p
)
2321 p
->se
.exec_start
= 0;
2322 p
->se
.sum_exec_runtime
= 0;
2323 p
->se
.prev_sum_exec_runtime
= 0;
2324 p
->se
.last_wakeup
= 0;
2325 p
->se
.avg_overlap
= 0;
2327 #ifdef CONFIG_SCHEDSTATS
2328 p
->se
.wait_start
= 0;
2329 p
->se
.sum_sleep_runtime
= 0;
2330 p
->se
.sleep_start
= 0;
2331 p
->se
.block_start
= 0;
2332 p
->se
.sleep_max
= 0;
2333 p
->se
.block_max
= 0;
2335 p
->se
.slice_max
= 0;
2339 INIT_LIST_HEAD(&p
->rt
.run_list
);
2341 INIT_LIST_HEAD(&p
->se
.group_node
);
2343 #ifdef CONFIG_PREEMPT_NOTIFIERS
2344 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2348 * We mark the process as running here, but have not actually
2349 * inserted it onto the runqueue yet. This guarantees that
2350 * nobody will actually run it, and a signal or other external
2351 * event cannot wake it up and insert it on the runqueue either.
2353 p
->state
= TASK_RUNNING
;
2357 * fork()/clone()-time setup:
2359 void sched_fork(struct task_struct
*p
, int clone_flags
)
2361 int cpu
= get_cpu();
2366 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2368 set_task_cpu(p
, cpu
);
2371 * Make sure we do not leak PI boosting priority to the child:
2373 p
->prio
= current
->normal_prio
;
2374 if (!rt_prio(p
->prio
))
2375 p
->sched_class
= &fair_sched_class
;
2377 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2378 if (likely(sched_info_on()))
2379 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2381 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2384 #ifdef CONFIG_PREEMPT
2385 /* Want to start with kernel preemption disabled. */
2386 task_thread_info(p
)->preempt_count
= 1;
2392 * wake_up_new_task - wake up a newly created task for the first time.
2394 * This function will do some initial scheduler statistics housekeeping
2395 * that must be done for every newly created context, then puts the task
2396 * on the runqueue and wakes it.
2398 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2400 unsigned long flags
;
2403 rq
= task_rq_lock(p
, &flags
);
2404 BUG_ON(p
->state
!= TASK_RUNNING
);
2405 update_rq_clock(rq
);
2407 p
->prio
= effective_prio(p
);
2409 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2410 activate_task(rq
, p
, 0);
2413 * Let the scheduling class do new task startup
2414 * management (if any):
2416 p
->sched_class
->task_new(rq
, p
);
2419 trace_sched_wakeup_new(rq
, p
);
2420 check_preempt_curr(rq
, p
, 0);
2422 if (p
->sched_class
->task_wake_up
)
2423 p
->sched_class
->task_wake_up(rq
, p
);
2425 task_rq_unlock(rq
, &flags
);
2428 #ifdef CONFIG_PREEMPT_NOTIFIERS
2431 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2432 * @notifier: notifier struct to register
2434 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2436 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2438 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2441 * preempt_notifier_unregister - no longer interested in preemption notifications
2442 * @notifier: notifier struct to unregister
2444 * This is safe to call from within a preemption notifier.
2446 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2448 hlist_del(¬ifier
->link
);
2450 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2452 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2454 struct preempt_notifier
*notifier
;
2455 struct hlist_node
*node
;
2457 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2458 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2462 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2463 struct task_struct
*next
)
2465 struct preempt_notifier
*notifier
;
2466 struct hlist_node
*node
;
2468 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2469 notifier
->ops
->sched_out(notifier
, next
);
2472 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2474 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2479 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2480 struct task_struct
*next
)
2484 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2487 * prepare_task_switch - prepare to switch tasks
2488 * @rq: the runqueue preparing to switch
2489 * @prev: the current task that is being switched out
2490 * @next: the task we are going to switch to.
2492 * This is called with the rq lock held and interrupts off. It must
2493 * be paired with a subsequent finish_task_switch after the context
2496 * prepare_task_switch sets up locking and calls architecture specific
2500 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2501 struct task_struct
*next
)
2503 fire_sched_out_preempt_notifiers(prev
, next
);
2504 prepare_lock_switch(rq
, next
);
2505 prepare_arch_switch(next
);
2509 * finish_task_switch - clean up after a task-switch
2510 * @rq: runqueue associated with task-switch
2511 * @prev: the thread we just switched away from.
2513 * finish_task_switch must be called after the context switch, paired
2514 * with a prepare_task_switch call before the context switch.
2515 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2516 * and do any other architecture-specific cleanup actions.
2518 * Note that we may have delayed dropping an mm in context_switch(). If
2519 * so, we finish that here outside of the runqueue lock. (Doing it
2520 * with the lock held can cause deadlocks; see schedule() for
2523 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2524 __releases(rq
->lock
)
2526 struct mm_struct
*mm
= rq
->prev_mm
;
2532 * A task struct has one reference for the use as "current".
2533 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2534 * schedule one last time. The schedule call will never return, and
2535 * the scheduled task must drop that reference.
2536 * The test for TASK_DEAD must occur while the runqueue locks are
2537 * still held, otherwise prev could be scheduled on another cpu, die
2538 * there before we look at prev->state, and then the reference would
2540 * Manfred Spraul <manfred@colorfullife.com>
2542 prev_state
= prev
->state
;
2543 finish_arch_switch(prev
);
2544 finish_lock_switch(rq
, prev
);
2546 if (current
->sched_class
->post_schedule
)
2547 current
->sched_class
->post_schedule(rq
);
2550 fire_sched_in_preempt_notifiers(current
);
2553 if (unlikely(prev_state
== TASK_DEAD
)) {
2555 * Remove function-return probe instances associated with this
2556 * task and put them back on the free list.
2558 kprobe_flush_task(prev
);
2559 put_task_struct(prev
);
2564 * schedule_tail - first thing a freshly forked thread must call.
2565 * @prev: the thread we just switched away from.
2567 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2568 __releases(rq
->lock
)
2570 struct rq
*rq
= this_rq();
2572 finish_task_switch(rq
, prev
);
2573 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2574 /* In this case, finish_task_switch does not reenable preemption */
2577 if (current
->set_child_tid
)
2578 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2582 * context_switch - switch to the new MM and the new
2583 * thread's register state.
2586 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2587 struct task_struct
*next
)
2589 struct mm_struct
*mm
, *oldmm
;
2591 prepare_task_switch(rq
, prev
, next
);
2592 trace_sched_switch(rq
, prev
, next
);
2594 oldmm
= prev
->active_mm
;
2596 * For paravirt, this is coupled with an exit in switch_to to
2597 * combine the page table reload and the switch backend into
2600 arch_enter_lazy_cpu_mode();
2602 if (unlikely(!mm
)) {
2603 next
->active_mm
= oldmm
;
2604 atomic_inc(&oldmm
->mm_count
);
2605 enter_lazy_tlb(oldmm
, next
);
2607 switch_mm(oldmm
, mm
, next
);
2609 if (unlikely(!prev
->mm
)) {
2610 prev
->active_mm
= NULL
;
2611 rq
->prev_mm
= oldmm
;
2614 * Since the runqueue lock will be released by the next
2615 * task (which is an invalid locking op but in the case
2616 * of the scheduler it's an obvious special-case), so we
2617 * do an early lockdep release here:
2619 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2620 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2623 /* Here we just switch the register state and the stack. */
2624 switch_to(prev
, next
, prev
);
2628 * this_rq must be evaluated again because prev may have moved
2629 * CPUs since it called schedule(), thus the 'rq' on its stack
2630 * frame will be invalid.
2632 finish_task_switch(this_rq(), prev
);
2636 * nr_running, nr_uninterruptible and nr_context_switches:
2638 * externally visible scheduler statistics: current number of runnable
2639 * threads, current number of uninterruptible-sleeping threads, total
2640 * number of context switches performed since bootup.
2642 unsigned long nr_running(void)
2644 unsigned long i
, sum
= 0;
2646 for_each_online_cpu(i
)
2647 sum
+= cpu_rq(i
)->nr_running
;
2652 unsigned long nr_uninterruptible(void)
2654 unsigned long i
, sum
= 0;
2656 for_each_possible_cpu(i
)
2657 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2660 * Since we read the counters lockless, it might be slightly
2661 * inaccurate. Do not allow it to go below zero though:
2663 if (unlikely((long)sum
< 0))
2669 unsigned long long nr_context_switches(void)
2672 unsigned long long sum
= 0;
2674 for_each_possible_cpu(i
)
2675 sum
+= cpu_rq(i
)->nr_switches
;
2680 unsigned long nr_iowait(void)
2682 unsigned long i
, sum
= 0;
2684 for_each_possible_cpu(i
)
2685 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2690 unsigned long nr_active(void)
2692 unsigned long i
, running
= 0, uninterruptible
= 0;
2694 for_each_online_cpu(i
) {
2695 running
+= cpu_rq(i
)->nr_running
;
2696 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2699 if (unlikely((long)uninterruptible
< 0))
2700 uninterruptible
= 0;
2702 return running
+ uninterruptible
;
2706 * Update rq->cpu_load[] statistics. This function is usually called every
2707 * scheduler tick (TICK_NSEC).
2709 static void update_cpu_load(struct rq
*this_rq
)
2711 unsigned long this_load
= this_rq
->load
.weight
;
2714 this_rq
->nr_load_updates
++;
2716 /* Update our load: */
2717 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2718 unsigned long old_load
, new_load
;
2720 /* scale is effectively 1 << i now, and >> i divides by scale */
2722 old_load
= this_rq
->cpu_load
[i
];
2723 new_load
= this_load
;
2725 * Round up the averaging division if load is increasing. This
2726 * prevents us from getting stuck on 9 if the load is 10, for
2729 if (new_load
> old_load
)
2730 new_load
+= scale
-1;
2731 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2738 * double_rq_lock - safely lock two runqueues
2740 * Note this does not disable interrupts like task_rq_lock,
2741 * you need to do so manually before calling.
2743 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2744 __acquires(rq1
->lock
)
2745 __acquires(rq2
->lock
)
2747 BUG_ON(!irqs_disabled());
2749 spin_lock(&rq1
->lock
);
2750 __acquire(rq2
->lock
); /* Fake it out ;) */
2753 spin_lock(&rq1
->lock
);
2754 spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
2756 spin_lock(&rq2
->lock
);
2757 spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
2760 update_rq_clock(rq1
);
2761 update_rq_clock(rq2
);
2765 * double_rq_unlock - safely unlock two runqueues
2767 * Note this does not restore interrupts like task_rq_unlock,
2768 * you need to do so manually after calling.
2770 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2771 __releases(rq1
->lock
)
2772 __releases(rq2
->lock
)
2774 spin_unlock(&rq1
->lock
);
2776 spin_unlock(&rq2
->lock
);
2778 __release(rq2
->lock
);
2782 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2784 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2785 __releases(this_rq
->lock
)
2786 __acquires(busiest
->lock
)
2787 __acquires(this_rq
->lock
)
2791 if (unlikely(!irqs_disabled())) {
2792 /* printk() doesn't work good under rq->lock */
2793 spin_unlock(&this_rq
->lock
);
2796 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2797 if (busiest
< this_rq
) {
2798 spin_unlock(&this_rq
->lock
);
2799 spin_lock(&busiest
->lock
);
2800 spin_lock_nested(&this_rq
->lock
, SINGLE_DEPTH_NESTING
);
2803 spin_lock_nested(&busiest
->lock
, SINGLE_DEPTH_NESTING
);
2808 static void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2809 __releases(busiest
->lock
)
2811 spin_unlock(&busiest
->lock
);
2812 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
2816 * If dest_cpu is allowed for this process, migrate the task to it.
2817 * This is accomplished by forcing the cpu_allowed mask to only
2818 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2819 * the cpu_allowed mask is restored.
2821 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2823 struct migration_req req
;
2824 unsigned long flags
;
2827 rq
= task_rq_lock(p
, &flags
);
2828 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2829 || unlikely(!cpu_active(dest_cpu
)))
2832 trace_sched_migrate_task(rq
, p
, dest_cpu
);
2833 /* force the process onto the specified CPU */
2834 if (migrate_task(p
, dest_cpu
, &req
)) {
2835 /* Need to wait for migration thread (might exit: take ref). */
2836 struct task_struct
*mt
= rq
->migration_thread
;
2838 get_task_struct(mt
);
2839 task_rq_unlock(rq
, &flags
);
2840 wake_up_process(mt
);
2841 put_task_struct(mt
);
2842 wait_for_completion(&req
.done
);
2847 task_rq_unlock(rq
, &flags
);
2851 * sched_exec - execve() is a valuable balancing opportunity, because at
2852 * this point the task has the smallest effective memory and cache footprint.
2854 void sched_exec(void)
2856 int new_cpu
, this_cpu
= get_cpu();
2857 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2859 if (new_cpu
!= this_cpu
)
2860 sched_migrate_task(current
, new_cpu
);
2864 * pull_task - move a task from a remote runqueue to the local runqueue.
2865 * Both runqueues must be locked.
2867 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2868 struct rq
*this_rq
, int this_cpu
)
2870 deactivate_task(src_rq
, p
, 0);
2871 set_task_cpu(p
, this_cpu
);
2872 activate_task(this_rq
, p
, 0);
2874 * Note that idle threads have a prio of MAX_PRIO, for this test
2875 * to be always true for them.
2877 check_preempt_curr(this_rq
, p
, 0);
2881 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2884 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2885 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2889 * We do not migrate tasks that are:
2890 * 1) running (obviously), or
2891 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2892 * 3) are cache-hot on their current CPU.
2894 if (!cpu_isset(this_cpu
, p
->cpus_allowed
)) {
2895 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2900 if (task_running(rq
, p
)) {
2901 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2906 * Aggressive migration if:
2907 * 1) task is cache cold, or
2908 * 2) too many balance attempts have failed.
2911 if (!task_hot(p
, rq
->clock
, sd
) ||
2912 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2913 #ifdef CONFIG_SCHEDSTATS
2914 if (task_hot(p
, rq
->clock
, sd
)) {
2915 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2916 schedstat_inc(p
, se
.nr_forced_migrations
);
2922 if (task_hot(p
, rq
->clock
, sd
)) {
2923 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
2929 static unsigned long
2930 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2931 unsigned long max_load_move
, struct sched_domain
*sd
,
2932 enum cpu_idle_type idle
, int *all_pinned
,
2933 int *this_best_prio
, struct rq_iterator
*iterator
)
2935 int loops
= 0, pulled
= 0, pinned
= 0;
2936 struct task_struct
*p
;
2937 long rem_load_move
= max_load_move
;
2939 if (max_load_move
== 0)
2945 * Start the load-balancing iterator:
2947 p
= iterator
->start(iterator
->arg
);
2949 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
2952 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
2953 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2954 p
= iterator
->next(iterator
->arg
);
2958 pull_task(busiest
, p
, this_rq
, this_cpu
);
2960 rem_load_move
-= p
->se
.load
.weight
;
2963 * We only want to steal up to the prescribed amount of weighted load.
2965 if (rem_load_move
> 0) {
2966 if (p
->prio
< *this_best_prio
)
2967 *this_best_prio
= p
->prio
;
2968 p
= iterator
->next(iterator
->arg
);
2973 * Right now, this is one of only two places pull_task() is called,
2974 * so we can safely collect pull_task() stats here rather than
2975 * inside pull_task().
2977 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2980 *all_pinned
= pinned
;
2982 return max_load_move
- rem_load_move
;
2986 * move_tasks tries to move up to max_load_move weighted load from busiest to
2987 * this_rq, as part of a balancing operation within domain "sd".
2988 * Returns 1 if successful and 0 otherwise.
2990 * Called with both runqueues locked.
2992 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2993 unsigned long max_load_move
,
2994 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2997 const struct sched_class
*class = sched_class_highest
;
2998 unsigned long total_load_moved
= 0;
2999 int this_best_prio
= this_rq
->curr
->prio
;
3003 class->load_balance(this_rq
, this_cpu
, busiest
,
3004 max_load_move
- total_load_moved
,
3005 sd
, idle
, all_pinned
, &this_best_prio
);
3006 class = class->next
;
3008 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3011 } while (class && max_load_move
> total_load_moved
);
3013 return total_load_moved
> 0;
3017 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3018 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3019 struct rq_iterator
*iterator
)
3021 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3025 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3026 pull_task(busiest
, p
, this_rq
, this_cpu
);
3028 * Right now, this is only the second place pull_task()
3029 * is called, so we can safely collect pull_task()
3030 * stats here rather than inside pull_task().
3032 schedstat_inc(sd
, lb_gained
[idle
]);
3036 p
= iterator
->next(iterator
->arg
);
3043 * move_one_task tries to move exactly one task from busiest to this_rq, as
3044 * part of active balancing operations within "domain".
3045 * Returns 1 if successful and 0 otherwise.
3047 * Called with both runqueues locked.
3049 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3050 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3052 const struct sched_class
*class;
3054 for (class = sched_class_highest
; class; class = class->next
)
3055 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3062 * find_busiest_group finds and returns the busiest CPU group within the
3063 * domain. It calculates and returns the amount of weighted load which
3064 * should be moved to restore balance via the imbalance parameter.
3066 static struct sched_group
*
3067 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
3068 unsigned long *imbalance
, enum cpu_idle_type idle
,
3069 int *sd_idle
, const cpumask_t
*cpus
, int *balance
)
3071 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
3072 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
3073 unsigned long max_pull
;
3074 unsigned long busiest_load_per_task
, busiest_nr_running
;
3075 unsigned long this_load_per_task
, this_nr_running
;
3076 int load_idx
, group_imb
= 0;
3077 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3078 int power_savings_balance
= 1;
3079 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
3080 unsigned long min_nr_running
= ULONG_MAX
;
3081 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
3084 max_load
= this_load
= total_load
= total_pwr
= 0;
3085 busiest_load_per_task
= busiest_nr_running
= 0;
3086 this_load_per_task
= this_nr_running
= 0;
3088 if (idle
== CPU_NOT_IDLE
)
3089 load_idx
= sd
->busy_idx
;
3090 else if (idle
== CPU_NEWLY_IDLE
)
3091 load_idx
= sd
->newidle_idx
;
3093 load_idx
= sd
->idle_idx
;
3096 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
3099 int __group_imb
= 0;
3100 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3101 unsigned long sum_nr_running
, sum_weighted_load
;
3102 unsigned long sum_avg_load_per_task
;
3103 unsigned long avg_load_per_task
;
3105 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
3108 balance_cpu
= first_cpu(group
->cpumask
);
3110 /* Tally up the load of all CPUs in the group */
3111 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
3112 sum_avg_load_per_task
= avg_load_per_task
= 0;
3115 min_cpu_load
= ~0UL;
3117 for_each_cpu_mask_nr(i
, group
->cpumask
) {
3120 if (!cpu_isset(i
, *cpus
))
3125 if (*sd_idle
&& rq
->nr_running
)
3128 /* Bias balancing toward cpus of our domain */
3130 if (idle_cpu(i
) && !first_idle_cpu
) {
3135 load
= target_load(i
, load_idx
);
3137 load
= source_load(i
, load_idx
);
3138 if (load
> max_cpu_load
)
3139 max_cpu_load
= load
;
3140 if (min_cpu_load
> load
)
3141 min_cpu_load
= load
;
3145 sum_nr_running
+= rq
->nr_running
;
3146 sum_weighted_load
+= weighted_cpuload(i
);
3148 sum_avg_load_per_task
+= cpu_avg_load_per_task(i
);
3152 * First idle cpu or the first cpu(busiest) in this sched group
3153 * is eligible for doing load balancing at this and above
3154 * domains. In the newly idle case, we will allow all the cpu's
3155 * to do the newly idle load balance.
3157 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3158 balance_cpu
!= this_cpu
&& balance
) {
3163 total_load
+= avg_load
;
3164 total_pwr
+= group
->__cpu_power
;
3166 /* Adjust by relative CPU power of the group */
3167 avg_load
= sg_div_cpu_power(group
,
3168 avg_load
* SCHED_LOAD_SCALE
);
3172 * Consider the group unbalanced when the imbalance is larger
3173 * than the average weight of two tasks.
3175 * APZ: with cgroup the avg task weight can vary wildly and
3176 * might not be a suitable number - should we keep a
3177 * normalized nr_running number somewhere that negates
3180 avg_load_per_task
= sg_div_cpu_power(group
,
3181 sum_avg_load_per_task
* SCHED_LOAD_SCALE
);
3183 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3186 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
3189 this_load
= avg_load
;
3191 this_nr_running
= sum_nr_running
;
3192 this_load_per_task
= sum_weighted_load
;
3193 } else if (avg_load
> max_load
&&
3194 (sum_nr_running
> group_capacity
|| __group_imb
)) {
3195 max_load
= avg_load
;
3197 busiest_nr_running
= sum_nr_running
;
3198 busiest_load_per_task
= sum_weighted_load
;
3199 group_imb
= __group_imb
;
3202 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3204 * Busy processors will not participate in power savings
3207 if (idle
== CPU_NOT_IDLE
||
3208 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3212 * If the local group is idle or completely loaded
3213 * no need to do power savings balance at this domain
3215 if (local_group
&& (this_nr_running
>= group_capacity
||
3217 power_savings_balance
= 0;
3220 * If a group is already running at full capacity or idle,
3221 * don't include that group in power savings calculations
3223 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
3228 * Calculate the group which has the least non-idle load.
3229 * This is the group from where we need to pick up the load
3232 if ((sum_nr_running
< min_nr_running
) ||
3233 (sum_nr_running
== min_nr_running
&&
3234 first_cpu(group
->cpumask
) <
3235 first_cpu(group_min
->cpumask
))) {
3237 min_nr_running
= sum_nr_running
;
3238 min_load_per_task
= sum_weighted_load
/
3243 * Calculate the group which is almost near its
3244 * capacity but still has some space to pick up some load
3245 * from other group and save more power
3247 if (sum_nr_running
<= group_capacity
- 1) {
3248 if (sum_nr_running
> leader_nr_running
||
3249 (sum_nr_running
== leader_nr_running
&&
3250 first_cpu(group
->cpumask
) >
3251 first_cpu(group_leader
->cpumask
))) {
3252 group_leader
= group
;
3253 leader_nr_running
= sum_nr_running
;
3258 group
= group
->next
;
3259 } while (group
!= sd
->groups
);
3261 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
3264 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
3266 if (this_load
>= avg_load
||
3267 100*max_load
<= sd
->imbalance_pct
*this_load
)
3270 busiest_load_per_task
/= busiest_nr_running
;
3272 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
3275 * We're trying to get all the cpus to the average_load, so we don't
3276 * want to push ourselves above the average load, nor do we wish to
3277 * reduce the max loaded cpu below the average load, as either of these
3278 * actions would just result in more rebalancing later, and ping-pong
3279 * tasks around. Thus we look for the minimum possible imbalance.
3280 * Negative imbalances (*we* are more loaded than anyone else) will
3281 * be counted as no imbalance for these purposes -- we can't fix that
3282 * by pulling tasks to us. Be careful of negative numbers as they'll
3283 * appear as very large values with unsigned longs.
3285 if (max_load
<= busiest_load_per_task
)
3289 * In the presence of smp nice balancing, certain scenarios can have
3290 * max load less than avg load(as we skip the groups at or below
3291 * its cpu_power, while calculating max_load..)
3293 if (max_load
< avg_load
) {
3295 goto small_imbalance
;
3298 /* Don't want to pull so many tasks that a group would go idle */
3299 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
3301 /* How much load to actually move to equalise the imbalance */
3302 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
3303 (avg_load
- this_load
) * this->__cpu_power
)
3307 * if *imbalance is less than the average load per runnable task
3308 * there is no gaurantee that any tasks will be moved so we'll have
3309 * a think about bumping its value to force at least one task to be
3312 if (*imbalance
< busiest_load_per_task
) {
3313 unsigned long tmp
, pwr_now
, pwr_move
;
3317 pwr_move
= pwr_now
= 0;
3319 if (this_nr_running
) {
3320 this_load_per_task
/= this_nr_running
;
3321 if (busiest_load_per_task
> this_load_per_task
)
3324 this_load_per_task
= cpu_avg_load_per_task(this_cpu
);
3326 if (max_load
- this_load
+ busiest_load_per_task
>=
3327 busiest_load_per_task
* imbn
) {
3328 *imbalance
= busiest_load_per_task
;
3333 * OK, we don't have enough imbalance to justify moving tasks,
3334 * however we may be able to increase total CPU power used by
3338 pwr_now
+= busiest
->__cpu_power
*
3339 min(busiest_load_per_task
, max_load
);
3340 pwr_now
+= this->__cpu_power
*
3341 min(this_load_per_task
, this_load
);
3342 pwr_now
/= SCHED_LOAD_SCALE
;
3344 /* Amount of load we'd subtract */
3345 tmp
= sg_div_cpu_power(busiest
,
3346 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3348 pwr_move
+= busiest
->__cpu_power
*
3349 min(busiest_load_per_task
, max_load
- tmp
);
3351 /* Amount of load we'd add */
3352 if (max_load
* busiest
->__cpu_power
<
3353 busiest_load_per_task
* SCHED_LOAD_SCALE
)
3354 tmp
= sg_div_cpu_power(this,
3355 max_load
* busiest
->__cpu_power
);
3357 tmp
= sg_div_cpu_power(this,
3358 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3359 pwr_move
+= this->__cpu_power
*
3360 min(this_load_per_task
, this_load
+ tmp
);
3361 pwr_move
/= SCHED_LOAD_SCALE
;
3363 /* Move if we gain throughput */
3364 if (pwr_move
> pwr_now
)
3365 *imbalance
= busiest_load_per_task
;
3371 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3372 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3375 if (this == group_leader
&& group_leader
!= group_min
) {
3376 *imbalance
= min_load_per_task
;
3386 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3389 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
3390 unsigned long imbalance
, const cpumask_t
*cpus
)
3392 struct rq
*busiest
= NULL
, *rq
;
3393 unsigned long max_load
= 0;
3396 for_each_cpu_mask_nr(i
, group
->cpumask
) {
3399 if (!cpu_isset(i
, *cpus
))
3403 wl
= weighted_cpuload(i
);
3405 if (rq
->nr_running
== 1 && wl
> imbalance
)
3408 if (wl
> max_load
) {
3418 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3419 * so long as it is large enough.
3421 #define MAX_PINNED_INTERVAL 512
3424 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3425 * tasks if there is an imbalance.
3427 static int load_balance(int this_cpu
, struct rq
*this_rq
,
3428 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3429 int *balance
, cpumask_t
*cpus
)
3431 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
3432 struct sched_group
*group
;
3433 unsigned long imbalance
;
3435 unsigned long flags
;
3440 * When power savings policy is enabled for the parent domain, idle
3441 * sibling can pick up load irrespective of busy siblings. In this case,
3442 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3443 * portraying it as CPU_NOT_IDLE.
3445 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3446 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3449 schedstat_inc(sd
, lb_count
[idle
]);
3453 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3460 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3464 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
3466 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3470 BUG_ON(busiest
== this_rq
);
3472 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3475 if (busiest
->nr_running
> 1) {
3477 * Attempt to move tasks. If find_busiest_group has found
3478 * an imbalance but busiest->nr_running <= 1, the group is
3479 * still unbalanced. ld_moved simply stays zero, so it is
3480 * correctly treated as an imbalance.
3482 local_irq_save(flags
);
3483 double_rq_lock(this_rq
, busiest
);
3484 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3485 imbalance
, sd
, idle
, &all_pinned
);
3486 double_rq_unlock(this_rq
, busiest
);
3487 local_irq_restore(flags
);
3490 * some other cpu did the load balance for us.
3492 if (ld_moved
&& this_cpu
!= smp_processor_id())
3493 resched_cpu(this_cpu
);
3495 /* All tasks on this runqueue were pinned by CPU affinity */
3496 if (unlikely(all_pinned
)) {
3497 cpu_clear(cpu_of(busiest
), *cpus
);
3498 if (!cpus_empty(*cpus
))
3505 schedstat_inc(sd
, lb_failed
[idle
]);
3506 sd
->nr_balance_failed
++;
3508 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3510 spin_lock_irqsave(&busiest
->lock
, flags
);
3512 /* don't kick the migration_thread, if the curr
3513 * task on busiest cpu can't be moved to this_cpu
3515 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
3516 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3518 goto out_one_pinned
;
3521 if (!busiest
->active_balance
) {
3522 busiest
->active_balance
= 1;
3523 busiest
->push_cpu
= this_cpu
;
3526 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3528 wake_up_process(busiest
->migration_thread
);
3531 * We've kicked active balancing, reset the failure
3534 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3537 sd
->nr_balance_failed
= 0;
3539 if (likely(!active_balance
)) {
3540 /* We were unbalanced, so reset the balancing interval */
3541 sd
->balance_interval
= sd
->min_interval
;
3544 * If we've begun active balancing, start to back off. This
3545 * case may not be covered by the all_pinned logic if there
3546 * is only 1 task on the busy runqueue (because we don't call
3549 if (sd
->balance_interval
< sd
->max_interval
)
3550 sd
->balance_interval
*= 2;
3553 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3554 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3560 schedstat_inc(sd
, lb_balanced
[idle
]);
3562 sd
->nr_balance_failed
= 0;
3565 /* tune up the balancing interval */
3566 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3567 (sd
->balance_interval
< sd
->max_interval
))
3568 sd
->balance_interval
*= 2;
3570 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3571 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3582 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3583 * tasks if there is an imbalance.
3585 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3586 * this_rq is locked.
3589 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
,
3592 struct sched_group
*group
;
3593 struct rq
*busiest
= NULL
;
3594 unsigned long imbalance
;
3602 * When power savings policy is enabled for the parent domain, idle
3603 * sibling can pick up load irrespective of busy siblings. In this case,
3604 * let the state of idle sibling percolate up as IDLE, instead of
3605 * portraying it as CPU_NOT_IDLE.
3607 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
3608 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3611 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
3613 update_shares_locked(this_rq
, sd
);
3614 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
3615 &sd_idle
, cpus
, NULL
);
3617 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
3621 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
3623 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
3627 BUG_ON(busiest
== this_rq
);
3629 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
3632 if (busiest
->nr_running
> 1) {
3633 /* Attempt to move tasks */
3634 double_lock_balance(this_rq
, busiest
);
3635 /* this_rq->clock is already updated */
3636 update_rq_clock(busiest
);
3637 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3638 imbalance
, sd
, CPU_NEWLY_IDLE
,
3640 double_unlock_balance(this_rq
, busiest
);
3642 if (unlikely(all_pinned
)) {
3643 cpu_clear(cpu_of(busiest
), *cpus
);
3644 if (!cpus_empty(*cpus
))
3650 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
3651 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3652 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3655 sd
->nr_balance_failed
= 0;
3657 update_shares_locked(this_rq
, sd
);
3661 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
3662 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3663 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3665 sd
->nr_balance_failed
= 0;
3671 * idle_balance is called by schedule() if this_cpu is about to become
3672 * idle. Attempts to pull tasks from other CPUs.
3674 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
3676 struct sched_domain
*sd
;
3677 int pulled_task
= -1;
3678 unsigned long next_balance
= jiffies
+ HZ
;
3681 for_each_domain(this_cpu
, sd
) {
3682 unsigned long interval
;
3684 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3687 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
3688 /* If we've pulled tasks over stop searching: */
3689 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
3692 interval
= msecs_to_jiffies(sd
->balance_interval
);
3693 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3694 next_balance
= sd
->last_balance
+ interval
;
3698 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3700 * We are going idle. next_balance may be set based on
3701 * a busy processor. So reset next_balance.
3703 this_rq
->next_balance
= next_balance
;
3708 * active_load_balance is run by migration threads. It pushes running tasks
3709 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3710 * running on each physical CPU where possible, and avoids physical /
3711 * logical imbalances.
3713 * Called with busiest_rq locked.
3715 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
3717 int target_cpu
= busiest_rq
->push_cpu
;
3718 struct sched_domain
*sd
;
3719 struct rq
*target_rq
;
3721 /* Is there any task to move? */
3722 if (busiest_rq
->nr_running
<= 1)
3725 target_rq
= cpu_rq(target_cpu
);
3728 * This condition is "impossible", if it occurs
3729 * we need to fix it. Originally reported by
3730 * Bjorn Helgaas on a 128-cpu setup.
3732 BUG_ON(busiest_rq
== target_rq
);
3734 /* move a task from busiest_rq to target_rq */
3735 double_lock_balance(busiest_rq
, target_rq
);
3736 update_rq_clock(busiest_rq
);
3737 update_rq_clock(target_rq
);
3739 /* Search for an sd spanning us and the target CPU. */
3740 for_each_domain(target_cpu
, sd
) {
3741 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3742 cpu_isset(busiest_cpu
, sd
->span
))
3747 schedstat_inc(sd
, alb_count
);
3749 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3751 schedstat_inc(sd
, alb_pushed
);
3753 schedstat_inc(sd
, alb_failed
);
3755 double_unlock_balance(busiest_rq
, target_rq
);
3760 atomic_t load_balancer
;
3762 } nohz ____cacheline_aligned
= {
3763 .load_balancer
= ATOMIC_INIT(-1),
3764 .cpu_mask
= CPU_MASK_NONE
,
3768 * This routine will try to nominate the ilb (idle load balancing)
3769 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3770 * load balancing on behalf of all those cpus. If all the cpus in the system
3771 * go into this tickless mode, then there will be no ilb owner (as there is
3772 * no need for one) and all the cpus will sleep till the next wakeup event
3775 * For the ilb owner, tick is not stopped. And this tick will be used
3776 * for idle load balancing. ilb owner will still be part of
3779 * While stopping the tick, this cpu will become the ilb owner if there
3780 * is no other owner. And will be the owner till that cpu becomes busy
3781 * or if all cpus in the system stop their ticks at which point
3782 * there is no need for ilb owner.
3784 * When the ilb owner becomes busy, it nominates another owner, during the
3785 * next busy scheduler_tick()
3787 int select_nohz_load_balancer(int stop_tick
)
3789 int cpu
= smp_processor_id();
3792 cpu_set(cpu
, nohz
.cpu_mask
);
3793 cpu_rq(cpu
)->in_nohz_recently
= 1;
3796 * If we are going offline and still the leader, give up!
3798 if (!cpu_active(cpu
) &&
3799 atomic_read(&nohz
.load_balancer
) == cpu
) {
3800 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3805 /* time for ilb owner also to sleep */
3806 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3807 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3808 atomic_set(&nohz
.load_balancer
, -1);
3812 if (atomic_read(&nohz
.load_balancer
) == -1) {
3813 /* make me the ilb owner */
3814 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3816 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3819 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3822 cpu_clear(cpu
, nohz
.cpu_mask
);
3824 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3825 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3832 static DEFINE_SPINLOCK(balancing
);
3835 * It checks each scheduling domain to see if it is due to be balanced,
3836 * and initiates a balancing operation if so.
3838 * Balancing parameters are set up in arch_init_sched_domains.
3840 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3843 struct rq
*rq
= cpu_rq(cpu
);
3844 unsigned long interval
;
3845 struct sched_domain
*sd
;
3846 /* Earliest time when we have to do rebalance again */
3847 unsigned long next_balance
= jiffies
+ 60*HZ
;
3848 int update_next_balance
= 0;
3852 for_each_domain(cpu
, sd
) {
3853 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3856 interval
= sd
->balance_interval
;
3857 if (idle
!= CPU_IDLE
)
3858 interval
*= sd
->busy_factor
;
3860 /* scale ms to jiffies */
3861 interval
= msecs_to_jiffies(interval
);
3862 if (unlikely(!interval
))
3864 if (interval
> HZ
*NR_CPUS
/10)
3865 interval
= HZ
*NR_CPUS
/10;
3867 need_serialize
= sd
->flags
& SD_SERIALIZE
;
3869 if (need_serialize
) {
3870 if (!spin_trylock(&balancing
))
3874 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3875 if (load_balance(cpu
, rq
, sd
, idle
, &balance
, &tmp
)) {
3877 * We've pulled tasks over so either we're no
3878 * longer idle, or one of our SMT siblings is
3881 idle
= CPU_NOT_IDLE
;
3883 sd
->last_balance
= jiffies
;
3886 spin_unlock(&balancing
);
3888 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3889 next_balance
= sd
->last_balance
+ interval
;
3890 update_next_balance
= 1;
3894 * Stop the load balance at this level. There is another
3895 * CPU in our sched group which is doing load balancing more
3903 * next_balance will be updated only when there is a need.
3904 * When the cpu is attached to null domain for ex, it will not be
3907 if (likely(update_next_balance
))
3908 rq
->next_balance
= next_balance
;
3912 * run_rebalance_domains is triggered when needed from the scheduler tick.
3913 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3914 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3916 static void run_rebalance_domains(struct softirq_action
*h
)
3918 int this_cpu
= smp_processor_id();
3919 struct rq
*this_rq
= cpu_rq(this_cpu
);
3920 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3921 CPU_IDLE
: CPU_NOT_IDLE
;
3923 rebalance_domains(this_cpu
, idle
);
3927 * If this cpu is the owner for idle load balancing, then do the
3928 * balancing on behalf of the other idle cpus whose ticks are
3931 if (this_rq
->idle_at_tick
&&
3932 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3933 cpumask_t cpus
= nohz
.cpu_mask
;
3937 cpu_clear(this_cpu
, cpus
);
3938 for_each_cpu_mask_nr(balance_cpu
, cpus
) {
3940 * If this cpu gets work to do, stop the load balancing
3941 * work being done for other cpus. Next load
3942 * balancing owner will pick it up.
3947 rebalance_domains(balance_cpu
, CPU_IDLE
);
3949 rq
= cpu_rq(balance_cpu
);
3950 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3951 this_rq
->next_balance
= rq
->next_balance
;
3958 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3960 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3961 * idle load balancing owner or decide to stop the periodic load balancing,
3962 * if the whole system is idle.
3964 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3968 * If we were in the nohz mode recently and busy at the current
3969 * scheduler tick, then check if we need to nominate new idle
3972 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3973 rq
->in_nohz_recently
= 0;
3975 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3976 cpu_clear(cpu
, nohz
.cpu_mask
);
3977 atomic_set(&nohz
.load_balancer
, -1);
3980 if (atomic_read(&nohz
.load_balancer
) == -1) {
3982 * simple selection for now: Nominate the
3983 * first cpu in the nohz list to be the next
3986 * TBD: Traverse the sched domains and nominate
3987 * the nearest cpu in the nohz.cpu_mask.
3989 int ilb
= first_cpu(nohz
.cpu_mask
);
3991 if (ilb
< nr_cpu_ids
)
3997 * If this cpu is idle and doing idle load balancing for all the
3998 * cpus with ticks stopped, is it time for that to stop?
4000 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4001 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4007 * If this cpu is idle and the idle load balancing is done by
4008 * someone else, then no need raise the SCHED_SOFTIRQ
4010 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4011 cpu_isset(cpu
, nohz
.cpu_mask
))
4014 if (time_after_eq(jiffies
, rq
->next_balance
))
4015 raise_softirq(SCHED_SOFTIRQ
);
4018 #else /* CONFIG_SMP */
4021 * on UP we do not need to balance between CPUs:
4023 static inline void idle_balance(int cpu
, struct rq
*rq
)
4029 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4031 EXPORT_PER_CPU_SYMBOL(kstat
);
4034 * Return any ns on the sched_clock that have not yet been banked in
4035 * @p in case that task is currently running.
4037 unsigned long long task_delta_exec(struct task_struct
*p
)
4039 unsigned long flags
;
4043 rq
= task_rq_lock(p
, &flags
);
4045 if (task_current(rq
, p
)) {
4048 update_rq_clock(rq
);
4049 delta_exec
= rq
->clock
- p
->se
.exec_start
;
4050 if ((s64
)delta_exec
> 0)
4054 task_rq_unlock(rq
, &flags
);
4060 * Account user cpu time to a process.
4061 * @p: the process that the cpu time gets accounted to
4062 * @cputime: the cpu time spent in user space since the last update
4064 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
4066 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4069 p
->utime
= cputime_add(p
->utime
, cputime
);
4070 account_group_user_time(p
, cputime
);
4072 /* Add user time to cpustat. */
4073 tmp
= cputime_to_cputime64(cputime
);
4074 if (TASK_NICE(p
) > 0)
4075 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
4077 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4078 /* Account for user time used */
4079 acct_update_integrals(p
);
4083 * Account guest cpu time to a process.
4084 * @p: the process that the cpu time gets accounted to
4085 * @cputime: the cpu time spent in virtual machine since the last update
4087 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
)
4090 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4092 tmp
= cputime_to_cputime64(cputime
);
4094 p
->utime
= cputime_add(p
->utime
, cputime
);
4095 account_group_user_time(p
, cputime
);
4096 p
->gtime
= cputime_add(p
->gtime
, cputime
);
4098 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4099 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
4103 * Account scaled user cpu time to a process.
4104 * @p: the process that the cpu time gets accounted to
4105 * @cputime: the cpu time spent in user space since the last update
4107 void account_user_time_scaled(struct task_struct
*p
, cputime_t cputime
)
4109 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime
);
4113 * Account system cpu time to a process.
4114 * @p: the process that the cpu time gets accounted to
4115 * @hardirq_offset: the offset to subtract from hardirq_count()
4116 * @cputime: the cpu time spent in kernel space since the last update
4118 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
4121 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4122 struct rq
*rq
= this_rq();
4125 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
4126 account_guest_time(p
, cputime
);
4130 p
->stime
= cputime_add(p
->stime
, cputime
);
4131 account_group_system_time(p
, cputime
);
4133 /* Add system time to cpustat. */
4134 tmp
= cputime_to_cputime64(cputime
);
4135 if (hardirq_count() - hardirq_offset
)
4136 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
4137 else if (softirq_count())
4138 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
4139 else if (p
!= rq
->idle
)
4140 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
4141 else if (atomic_read(&rq
->nr_iowait
) > 0)
4142 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
4144 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
4145 /* Account for system time used */
4146 acct_update_integrals(p
);
4150 * Account scaled system cpu time to a process.
4151 * @p: the process that the cpu time gets accounted to
4152 * @hardirq_offset: the offset to subtract from hardirq_count()
4153 * @cputime: the cpu time spent in kernel space since the last update
4155 void account_system_time_scaled(struct task_struct
*p
, cputime_t cputime
)
4157 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime
);
4161 * Account for involuntary wait time.
4162 * @p: the process from which the cpu time has been stolen
4163 * @steal: the cpu time spent in involuntary wait
4165 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
4167 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4168 cputime64_t tmp
= cputime_to_cputime64(steal
);
4169 struct rq
*rq
= this_rq();
4171 if (p
== rq
->idle
) {
4172 p
->stime
= cputime_add(p
->stime
, steal
);
4173 account_group_system_time(p
, steal
);
4174 if (atomic_read(&rq
->nr_iowait
) > 0)
4175 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
4177 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
4179 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
4183 * Use precise platform statistics if available:
4185 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4186 cputime_t
task_utime(struct task_struct
*p
)
4191 cputime_t
task_stime(struct task_struct
*p
)
4196 cputime_t
task_utime(struct task_struct
*p
)
4198 clock_t utime
= cputime_to_clock_t(p
->utime
),
4199 total
= utime
+ cputime_to_clock_t(p
->stime
);
4203 * Use CFS's precise accounting:
4205 temp
= (u64
)nsec_to_clock_t(p
->se
.sum_exec_runtime
);
4209 do_div(temp
, total
);
4211 utime
= (clock_t)temp
;
4213 p
->prev_utime
= max(p
->prev_utime
, clock_t_to_cputime(utime
));
4214 return p
->prev_utime
;
4217 cputime_t
task_stime(struct task_struct
*p
)
4222 * Use CFS's precise accounting. (we subtract utime from
4223 * the total, to make sure the total observed by userspace
4224 * grows monotonically - apps rely on that):
4226 stime
= nsec_to_clock_t(p
->se
.sum_exec_runtime
) -
4227 cputime_to_clock_t(task_utime(p
));
4230 p
->prev_stime
= max(p
->prev_stime
, clock_t_to_cputime(stime
));
4232 return p
->prev_stime
;
4236 inline cputime_t
task_gtime(struct task_struct
*p
)
4242 * This function gets called by the timer code, with HZ frequency.
4243 * We call it with interrupts disabled.
4245 * It also gets called by the fork code, when changing the parent's
4248 void scheduler_tick(void)
4250 int cpu
= smp_processor_id();
4251 struct rq
*rq
= cpu_rq(cpu
);
4252 struct task_struct
*curr
= rq
->curr
;
4256 spin_lock(&rq
->lock
);
4257 update_rq_clock(rq
);
4258 update_cpu_load(rq
);
4259 curr
->sched_class
->task_tick(rq
, curr
, 0);
4260 spin_unlock(&rq
->lock
);
4263 rq
->idle_at_tick
= idle_cpu(cpu
);
4264 trigger_load_balance(rq
, cpu
);
4268 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4269 defined(CONFIG_PREEMPT_TRACER))
4271 static inline unsigned long get_parent_ip(unsigned long addr
)
4273 if (in_lock_functions(addr
)) {
4274 addr
= CALLER_ADDR2
;
4275 if (in_lock_functions(addr
))
4276 addr
= CALLER_ADDR3
;
4281 void __kprobes
add_preempt_count(int val
)
4283 #ifdef CONFIG_DEBUG_PREEMPT
4287 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4290 preempt_count() += val
;
4291 #ifdef CONFIG_DEBUG_PREEMPT
4293 * Spinlock count overflowing soon?
4295 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4298 if (preempt_count() == val
)
4299 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4301 EXPORT_SYMBOL(add_preempt_count
);
4303 void __kprobes
sub_preempt_count(int val
)
4305 #ifdef CONFIG_DEBUG_PREEMPT
4309 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
4312 * Is the spinlock portion underflowing?
4314 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4315 !(preempt_count() & PREEMPT_MASK
)))
4319 if (preempt_count() == val
)
4320 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4321 preempt_count() -= val
;
4323 EXPORT_SYMBOL(sub_preempt_count
);
4328 * Print scheduling while atomic bug:
4330 static noinline
void __schedule_bug(struct task_struct
*prev
)
4332 struct pt_regs
*regs
= get_irq_regs();
4334 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4335 prev
->comm
, prev
->pid
, preempt_count());
4337 debug_show_held_locks(prev
);
4339 if (irqs_disabled())
4340 print_irqtrace_events(prev
);
4349 * Various schedule()-time debugging checks and statistics:
4351 static inline void schedule_debug(struct task_struct
*prev
)
4354 * Test if we are atomic. Since do_exit() needs to call into
4355 * schedule() atomically, we ignore that path for now.
4356 * Otherwise, whine if we are scheduling when we should not be.
4358 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
4359 __schedule_bug(prev
);
4361 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4363 schedstat_inc(this_rq(), sched_count
);
4364 #ifdef CONFIG_SCHEDSTATS
4365 if (unlikely(prev
->lock_depth
>= 0)) {
4366 schedstat_inc(this_rq(), bkl_count
);
4367 schedstat_inc(prev
, sched_info
.bkl_count
);
4373 * Pick up the highest-prio task:
4375 static inline struct task_struct
*
4376 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
4378 const struct sched_class
*class;
4379 struct task_struct
*p
;
4382 * Optimization: we know that if all tasks are in
4383 * the fair class we can call that function directly:
4385 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
4386 p
= fair_sched_class
.pick_next_task(rq
);
4391 class = sched_class_highest
;
4393 p
= class->pick_next_task(rq
);
4397 * Will never be NULL as the idle class always
4398 * returns a non-NULL p:
4400 class = class->next
;
4405 * schedule() is the main scheduler function.
4407 asmlinkage
void __sched
schedule(void)
4409 struct task_struct
*prev
, *next
;
4410 unsigned long *switch_count
;
4416 cpu
= smp_processor_id();
4420 switch_count
= &prev
->nivcsw
;
4422 release_kernel_lock(prev
);
4423 need_resched_nonpreemptible
:
4425 schedule_debug(prev
);
4427 if (sched_feat(HRTICK
))
4430 spin_lock_irq(&rq
->lock
);
4431 update_rq_clock(rq
);
4432 clear_tsk_need_resched(prev
);
4434 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
4435 if (unlikely(signal_pending_state(prev
->state
, prev
)))
4436 prev
->state
= TASK_RUNNING
;
4438 deactivate_task(rq
, prev
, 1);
4439 switch_count
= &prev
->nvcsw
;
4443 if (prev
->sched_class
->pre_schedule
)
4444 prev
->sched_class
->pre_schedule(rq
, prev
);
4447 if (unlikely(!rq
->nr_running
))
4448 idle_balance(cpu
, rq
);
4450 prev
->sched_class
->put_prev_task(rq
, prev
);
4451 next
= pick_next_task(rq
, prev
);
4453 if (likely(prev
!= next
)) {
4454 sched_info_switch(prev
, next
);
4460 context_switch(rq
, prev
, next
); /* unlocks the rq */
4462 * the context switch might have flipped the stack from under
4463 * us, hence refresh the local variables.
4465 cpu
= smp_processor_id();
4468 spin_unlock_irq(&rq
->lock
);
4470 if (unlikely(reacquire_kernel_lock(current
) < 0))
4471 goto need_resched_nonpreemptible
;
4473 preempt_enable_no_resched();
4474 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
4477 EXPORT_SYMBOL(schedule
);
4479 #ifdef CONFIG_PREEMPT
4481 * this is the entry point to schedule() from in-kernel preemption
4482 * off of preempt_enable. Kernel preemptions off return from interrupt
4483 * occur there and call schedule directly.
4485 asmlinkage
void __sched
preempt_schedule(void)
4487 struct thread_info
*ti
= current_thread_info();
4490 * If there is a non-zero preempt_count or interrupts are disabled,
4491 * we do not want to preempt the current task. Just return..
4493 if (likely(ti
->preempt_count
|| irqs_disabled()))
4497 add_preempt_count(PREEMPT_ACTIVE
);
4499 sub_preempt_count(PREEMPT_ACTIVE
);
4502 * Check again in case we missed a preemption opportunity
4503 * between schedule and now.
4506 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4508 EXPORT_SYMBOL(preempt_schedule
);
4511 * this is the entry point to schedule() from kernel preemption
4512 * off of irq context.
4513 * Note, that this is called and return with irqs disabled. This will
4514 * protect us against recursive calling from irq.
4516 asmlinkage
void __sched
preempt_schedule_irq(void)
4518 struct thread_info
*ti
= current_thread_info();
4520 /* Catch callers which need to be fixed */
4521 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
4524 add_preempt_count(PREEMPT_ACTIVE
);
4527 local_irq_disable();
4528 sub_preempt_count(PREEMPT_ACTIVE
);
4531 * Check again in case we missed a preemption opportunity
4532 * between schedule and now.
4535 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4538 #endif /* CONFIG_PREEMPT */
4540 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
4543 return try_to_wake_up(curr
->private, mode
, sync
);
4545 EXPORT_SYMBOL(default_wake_function
);
4548 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4549 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4550 * number) then we wake all the non-exclusive tasks and one exclusive task.
4552 * There are circumstances in which we can try to wake a task which has already
4553 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4554 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4556 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4557 int nr_exclusive
, int sync
, void *key
)
4559 wait_queue_t
*curr
, *next
;
4561 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4562 unsigned flags
= curr
->flags
;
4564 if (curr
->func(curr
, mode
, sync
, key
) &&
4565 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4571 * __wake_up - wake up threads blocked on a waitqueue.
4573 * @mode: which threads
4574 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4575 * @key: is directly passed to the wakeup function
4577 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4578 int nr_exclusive
, void *key
)
4580 unsigned long flags
;
4582 spin_lock_irqsave(&q
->lock
, flags
);
4583 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4584 spin_unlock_irqrestore(&q
->lock
, flags
);
4586 EXPORT_SYMBOL(__wake_up
);
4589 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4591 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4593 __wake_up_common(q
, mode
, 1, 0, NULL
);
4597 * __wake_up_sync - wake up threads blocked on a waitqueue.
4599 * @mode: which threads
4600 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4602 * The sync wakeup differs that the waker knows that it will schedule
4603 * away soon, so while the target thread will be woken up, it will not
4604 * be migrated to another CPU - ie. the two threads are 'synchronized'
4605 * with each other. This can prevent needless bouncing between CPUs.
4607 * On UP it can prevent extra preemption.
4610 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4612 unsigned long flags
;
4618 if (unlikely(!nr_exclusive
))
4621 spin_lock_irqsave(&q
->lock
, flags
);
4622 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
4623 spin_unlock_irqrestore(&q
->lock
, flags
);
4625 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4628 * complete: - signals a single thread waiting on this completion
4629 * @x: holds the state of this particular completion
4631 * This will wake up a single thread waiting on this completion. Threads will be
4632 * awakened in the same order in which they were queued.
4634 * See also complete_all(), wait_for_completion() and related routines.
4636 void complete(struct completion
*x
)
4638 unsigned long flags
;
4640 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4642 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4643 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4645 EXPORT_SYMBOL(complete
);
4648 * complete_all: - signals all threads waiting on this completion
4649 * @x: holds the state of this particular completion
4651 * This will wake up all threads waiting on this particular completion event.
4653 void complete_all(struct completion
*x
)
4655 unsigned long flags
;
4657 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4658 x
->done
+= UINT_MAX
/2;
4659 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4660 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4662 EXPORT_SYMBOL(complete_all
);
4664 static inline long __sched
4665 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4668 DECLARE_WAITQUEUE(wait
, current
);
4670 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4671 __add_wait_queue_tail(&x
->wait
, &wait
);
4673 if (signal_pending_state(state
, current
)) {
4674 timeout
= -ERESTARTSYS
;
4677 __set_current_state(state
);
4678 spin_unlock_irq(&x
->wait
.lock
);
4679 timeout
= schedule_timeout(timeout
);
4680 spin_lock_irq(&x
->wait
.lock
);
4681 } while (!x
->done
&& timeout
);
4682 __remove_wait_queue(&x
->wait
, &wait
);
4687 return timeout
?: 1;
4691 wait_for_common(struct completion
*x
, long timeout
, int state
)
4695 spin_lock_irq(&x
->wait
.lock
);
4696 timeout
= do_wait_for_common(x
, timeout
, state
);
4697 spin_unlock_irq(&x
->wait
.lock
);
4702 * wait_for_completion: - waits for completion of a task
4703 * @x: holds the state of this particular completion
4705 * This waits to be signaled for completion of a specific task. It is NOT
4706 * interruptible and there is no timeout.
4708 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4709 * and interrupt capability. Also see complete().
4711 void __sched
wait_for_completion(struct completion
*x
)
4713 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4715 EXPORT_SYMBOL(wait_for_completion
);
4718 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4719 * @x: holds the state of this particular completion
4720 * @timeout: timeout value in jiffies
4722 * This waits for either a completion of a specific task to be signaled or for a
4723 * specified timeout to expire. The timeout is in jiffies. It is not
4726 unsigned long __sched
4727 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4729 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4731 EXPORT_SYMBOL(wait_for_completion_timeout
);
4734 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4735 * @x: holds the state of this particular completion
4737 * This waits for completion of a specific task to be signaled. It is
4740 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4742 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4743 if (t
== -ERESTARTSYS
)
4747 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4750 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4751 * @x: holds the state of this particular completion
4752 * @timeout: timeout value in jiffies
4754 * This waits for either a completion of a specific task to be signaled or for a
4755 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4757 unsigned long __sched
4758 wait_for_completion_interruptible_timeout(struct completion
*x
,
4759 unsigned long timeout
)
4761 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4763 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4766 * wait_for_completion_killable: - waits for completion of a task (killable)
4767 * @x: holds the state of this particular completion
4769 * This waits to be signaled for completion of a specific task. It can be
4770 * interrupted by a kill signal.
4772 int __sched
wait_for_completion_killable(struct completion
*x
)
4774 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4775 if (t
== -ERESTARTSYS
)
4779 EXPORT_SYMBOL(wait_for_completion_killable
);
4782 * try_wait_for_completion - try to decrement a completion without blocking
4783 * @x: completion structure
4785 * Returns: 0 if a decrement cannot be done without blocking
4786 * 1 if a decrement succeeded.
4788 * If a completion is being used as a counting completion,
4789 * attempt to decrement the counter without blocking. This
4790 * enables us to avoid waiting if the resource the completion
4791 * is protecting is not available.
4793 bool try_wait_for_completion(struct completion
*x
)
4797 spin_lock_irq(&x
->wait
.lock
);
4802 spin_unlock_irq(&x
->wait
.lock
);
4805 EXPORT_SYMBOL(try_wait_for_completion
);
4808 * completion_done - Test to see if a completion has any waiters
4809 * @x: completion structure
4811 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4812 * 1 if there are no waiters.
4815 bool completion_done(struct completion
*x
)
4819 spin_lock_irq(&x
->wait
.lock
);
4822 spin_unlock_irq(&x
->wait
.lock
);
4825 EXPORT_SYMBOL(completion_done
);
4828 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4830 unsigned long flags
;
4833 init_waitqueue_entry(&wait
, current
);
4835 __set_current_state(state
);
4837 spin_lock_irqsave(&q
->lock
, flags
);
4838 __add_wait_queue(q
, &wait
);
4839 spin_unlock(&q
->lock
);
4840 timeout
= schedule_timeout(timeout
);
4841 spin_lock_irq(&q
->lock
);
4842 __remove_wait_queue(q
, &wait
);
4843 spin_unlock_irqrestore(&q
->lock
, flags
);
4848 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4850 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4852 EXPORT_SYMBOL(interruptible_sleep_on
);
4855 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4857 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4859 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4861 void __sched
sleep_on(wait_queue_head_t
*q
)
4863 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4865 EXPORT_SYMBOL(sleep_on
);
4867 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4869 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4871 EXPORT_SYMBOL(sleep_on_timeout
);
4873 #ifdef CONFIG_RT_MUTEXES
4876 * rt_mutex_setprio - set the current priority of a task
4878 * @prio: prio value (kernel-internal form)
4880 * This function changes the 'effective' priority of a task. It does
4881 * not touch ->normal_prio like __setscheduler().
4883 * Used by the rt_mutex code to implement priority inheritance logic.
4885 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4887 unsigned long flags
;
4888 int oldprio
, on_rq
, running
;
4890 const struct sched_class
*prev_class
= p
->sched_class
;
4892 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4894 rq
= task_rq_lock(p
, &flags
);
4895 update_rq_clock(rq
);
4898 on_rq
= p
->se
.on_rq
;
4899 running
= task_current(rq
, p
);
4901 dequeue_task(rq
, p
, 0);
4903 p
->sched_class
->put_prev_task(rq
, p
);
4906 p
->sched_class
= &rt_sched_class
;
4908 p
->sched_class
= &fair_sched_class
;
4913 p
->sched_class
->set_curr_task(rq
);
4915 enqueue_task(rq
, p
, 0);
4917 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4919 task_rq_unlock(rq
, &flags
);
4924 void set_user_nice(struct task_struct
*p
, long nice
)
4926 int old_prio
, delta
, on_rq
;
4927 unsigned long flags
;
4930 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4933 * We have to be careful, if called from sys_setpriority(),
4934 * the task might be in the middle of scheduling on another CPU.
4936 rq
= task_rq_lock(p
, &flags
);
4937 update_rq_clock(rq
);
4939 * The RT priorities are set via sched_setscheduler(), but we still
4940 * allow the 'normal' nice value to be set - but as expected
4941 * it wont have any effect on scheduling until the task is
4942 * SCHED_FIFO/SCHED_RR:
4944 if (task_has_rt_policy(p
)) {
4945 p
->static_prio
= NICE_TO_PRIO(nice
);
4948 on_rq
= p
->se
.on_rq
;
4950 dequeue_task(rq
, p
, 0);
4952 p
->static_prio
= NICE_TO_PRIO(nice
);
4955 p
->prio
= effective_prio(p
);
4956 delta
= p
->prio
- old_prio
;
4959 enqueue_task(rq
, p
, 0);
4961 * If the task increased its priority or is running and
4962 * lowered its priority, then reschedule its CPU:
4964 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4965 resched_task(rq
->curr
);
4968 task_rq_unlock(rq
, &flags
);
4970 EXPORT_SYMBOL(set_user_nice
);
4973 * can_nice - check if a task can reduce its nice value
4977 int can_nice(const struct task_struct
*p
, const int nice
)
4979 /* convert nice value [19,-20] to rlimit style value [1,40] */
4980 int nice_rlim
= 20 - nice
;
4982 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4983 capable(CAP_SYS_NICE
));
4986 #ifdef __ARCH_WANT_SYS_NICE
4989 * sys_nice - change the priority of the current process.
4990 * @increment: priority increment
4992 * sys_setpriority is a more generic, but much slower function that
4993 * does similar things.
4995 asmlinkage
long sys_nice(int increment
)
5000 * Setpriority might change our priority at the same moment.
5001 * We don't have to worry. Conceptually one call occurs first
5002 * and we have a single winner.
5004 if (increment
< -40)
5009 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
5015 if (increment
< 0 && !can_nice(current
, nice
))
5018 retval
= security_task_setnice(current
, nice
);
5022 set_user_nice(current
, nice
);
5029 * task_prio - return the priority value of a given task.
5030 * @p: the task in question.
5032 * This is the priority value as seen by users in /proc.
5033 * RT tasks are offset by -200. Normal tasks are centered
5034 * around 0, value goes from -16 to +15.
5036 int task_prio(const struct task_struct
*p
)
5038 return p
->prio
- MAX_RT_PRIO
;
5042 * task_nice - return the nice value of a given task.
5043 * @p: the task in question.
5045 int task_nice(const struct task_struct
*p
)
5047 return TASK_NICE(p
);
5049 EXPORT_SYMBOL(task_nice
);
5052 * idle_cpu - is a given cpu idle currently?
5053 * @cpu: the processor in question.
5055 int idle_cpu(int cpu
)
5057 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
5061 * idle_task - return the idle task for a given cpu.
5062 * @cpu: the processor in question.
5064 struct task_struct
*idle_task(int cpu
)
5066 return cpu_rq(cpu
)->idle
;
5070 * find_process_by_pid - find a process with a matching PID value.
5071 * @pid: the pid in question.
5073 static struct task_struct
*find_process_by_pid(pid_t pid
)
5075 return pid
? find_task_by_vpid(pid
) : current
;
5078 /* Actually do priority change: must hold rq lock. */
5080 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
5082 BUG_ON(p
->se
.on_rq
);
5085 switch (p
->policy
) {
5089 p
->sched_class
= &fair_sched_class
;
5093 p
->sched_class
= &rt_sched_class
;
5097 p
->rt_priority
= prio
;
5098 p
->normal_prio
= normal_prio(p
);
5099 /* we are holding p->pi_lock already */
5100 p
->prio
= rt_mutex_getprio(p
);
5104 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
5105 struct sched_param
*param
, bool user
)
5107 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
5108 unsigned long flags
;
5109 const struct sched_class
*prev_class
= p
->sched_class
;
5112 /* may grab non-irq protected spin_locks */
5113 BUG_ON(in_interrupt());
5115 /* double check policy once rq lock held */
5117 policy
= oldpolicy
= p
->policy
;
5118 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
5119 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
5120 policy
!= SCHED_IDLE
)
5123 * Valid priorities for SCHED_FIFO and SCHED_RR are
5124 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5125 * SCHED_BATCH and SCHED_IDLE is 0.
5127 if (param
->sched_priority
< 0 ||
5128 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
5129 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
5131 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
5135 * Allow unprivileged RT tasks to decrease priority:
5137 if (user
&& !capable(CAP_SYS_NICE
)) {
5138 if (rt_policy(policy
)) {
5139 unsigned long rlim_rtprio
;
5141 if (!lock_task_sighand(p
, &flags
))
5143 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
5144 unlock_task_sighand(p
, &flags
);
5146 /* can't set/change the rt policy */
5147 if (policy
!= p
->policy
&& !rlim_rtprio
)
5150 /* can't increase priority */
5151 if (param
->sched_priority
> p
->rt_priority
&&
5152 param
->sched_priority
> rlim_rtprio
)
5156 * Like positive nice levels, dont allow tasks to
5157 * move out of SCHED_IDLE either:
5159 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
5162 /* can't change other user's priorities */
5163 if ((current
->euid
!= p
->euid
) &&
5164 (current
->euid
!= p
->uid
))
5169 #ifdef CONFIG_RT_GROUP_SCHED
5171 * Do not allow realtime tasks into groups that have no runtime
5174 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
5175 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
5179 retval
= security_task_setscheduler(p
, policy
, param
);
5185 * make sure no PI-waiters arrive (or leave) while we are
5186 * changing the priority of the task:
5188 spin_lock_irqsave(&p
->pi_lock
, flags
);
5190 * To be able to change p->policy safely, the apropriate
5191 * runqueue lock must be held.
5193 rq
= __task_rq_lock(p
);
5194 /* recheck policy now with rq lock held */
5195 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
5196 policy
= oldpolicy
= -1;
5197 __task_rq_unlock(rq
);
5198 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5201 update_rq_clock(rq
);
5202 on_rq
= p
->se
.on_rq
;
5203 running
= task_current(rq
, p
);
5205 deactivate_task(rq
, p
, 0);
5207 p
->sched_class
->put_prev_task(rq
, p
);
5210 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
5213 p
->sched_class
->set_curr_task(rq
);
5215 activate_task(rq
, p
, 0);
5217 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5219 __task_rq_unlock(rq
);
5220 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5222 rt_mutex_adjust_pi(p
);
5228 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5229 * @p: the task in question.
5230 * @policy: new policy.
5231 * @param: structure containing the new RT priority.
5233 * NOTE that the task may be already dead.
5235 int sched_setscheduler(struct task_struct
*p
, int policy
,
5236 struct sched_param
*param
)
5238 return __sched_setscheduler(p
, policy
, param
, true);
5240 EXPORT_SYMBOL_GPL(sched_setscheduler
);
5243 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5244 * @p: the task in question.
5245 * @policy: new policy.
5246 * @param: structure containing the new RT priority.
5248 * Just like sched_setscheduler, only don't bother checking if the
5249 * current context has permission. For example, this is needed in
5250 * stop_machine(): we create temporary high priority worker threads,
5251 * but our caller might not have that capability.
5253 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
5254 struct sched_param
*param
)
5256 return __sched_setscheduler(p
, policy
, param
, false);
5260 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5262 struct sched_param lparam
;
5263 struct task_struct
*p
;
5266 if (!param
|| pid
< 0)
5268 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
5273 p
= find_process_by_pid(pid
);
5275 retval
= sched_setscheduler(p
, policy
, &lparam
);
5282 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5283 * @pid: the pid in question.
5284 * @policy: new policy.
5285 * @param: structure containing the new RT priority.
5288 sys_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5290 /* negative values for policy are not valid */
5294 return do_sched_setscheduler(pid
, policy
, param
);
5298 * sys_sched_setparam - set/change the RT priority of a thread
5299 * @pid: the pid in question.
5300 * @param: structure containing the new RT priority.
5302 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
5304 return do_sched_setscheduler(pid
, -1, param
);
5308 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5309 * @pid: the pid in question.
5311 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
5313 struct task_struct
*p
;
5320 read_lock(&tasklist_lock
);
5321 p
= find_process_by_pid(pid
);
5323 retval
= security_task_getscheduler(p
);
5327 read_unlock(&tasklist_lock
);
5332 * sys_sched_getscheduler - get the RT priority of a thread
5333 * @pid: the pid in question.
5334 * @param: structure containing the RT priority.
5336 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
5338 struct sched_param lp
;
5339 struct task_struct
*p
;
5342 if (!param
|| pid
< 0)
5345 read_lock(&tasklist_lock
);
5346 p
= find_process_by_pid(pid
);
5351 retval
= security_task_getscheduler(p
);
5355 lp
.sched_priority
= p
->rt_priority
;
5356 read_unlock(&tasklist_lock
);
5359 * This one might sleep, we cannot do it with a spinlock held ...
5361 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
5366 read_unlock(&tasklist_lock
);
5370 long sched_setaffinity(pid_t pid
, const cpumask_t
*in_mask
)
5372 cpumask_t cpus_allowed
;
5373 cpumask_t new_mask
= *in_mask
;
5374 struct task_struct
*p
;
5378 read_lock(&tasklist_lock
);
5380 p
= find_process_by_pid(pid
);
5382 read_unlock(&tasklist_lock
);
5388 * It is not safe to call set_cpus_allowed with the
5389 * tasklist_lock held. We will bump the task_struct's
5390 * usage count and then drop tasklist_lock.
5393 read_unlock(&tasklist_lock
);
5396 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
5397 !capable(CAP_SYS_NICE
))
5400 retval
= security_task_setscheduler(p
, 0, NULL
);
5404 cpuset_cpus_allowed(p
, &cpus_allowed
);
5405 cpus_and(new_mask
, new_mask
, cpus_allowed
);
5407 retval
= set_cpus_allowed_ptr(p
, &new_mask
);
5410 cpuset_cpus_allowed(p
, &cpus_allowed
);
5411 if (!cpus_subset(new_mask
, cpus_allowed
)) {
5413 * We must have raced with a concurrent cpuset
5414 * update. Just reset the cpus_allowed to the
5415 * cpuset's cpus_allowed
5417 new_mask
= cpus_allowed
;
5427 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5428 cpumask_t
*new_mask
)
5430 if (len
< sizeof(cpumask_t
)) {
5431 memset(new_mask
, 0, sizeof(cpumask_t
));
5432 } else if (len
> sizeof(cpumask_t
)) {
5433 len
= sizeof(cpumask_t
);
5435 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
5439 * sys_sched_setaffinity - set the cpu affinity of a process
5440 * @pid: pid of the process
5441 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5442 * @user_mask_ptr: user-space pointer to the new cpu mask
5444 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
5445 unsigned long __user
*user_mask_ptr
)
5450 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
5454 return sched_setaffinity(pid
, &new_mask
);
5457 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
5459 struct task_struct
*p
;
5463 read_lock(&tasklist_lock
);
5466 p
= find_process_by_pid(pid
);
5470 retval
= security_task_getscheduler(p
);
5474 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
5477 read_unlock(&tasklist_lock
);
5484 * sys_sched_getaffinity - get the cpu affinity of a process
5485 * @pid: pid of the process
5486 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5487 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5489 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
5490 unsigned long __user
*user_mask_ptr
)
5495 if (len
< sizeof(cpumask_t
))
5498 ret
= sched_getaffinity(pid
, &mask
);
5502 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
5505 return sizeof(cpumask_t
);
5509 * sys_sched_yield - yield the current processor to other threads.
5511 * This function yields the current CPU to other tasks. If there are no
5512 * other threads running on this CPU then this function will return.
5514 asmlinkage
long sys_sched_yield(void)
5516 struct rq
*rq
= this_rq_lock();
5518 schedstat_inc(rq
, yld_count
);
5519 current
->sched_class
->yield_task(rq
);
5522 * Since we are going to call schedule() anyway, there's
5523 * no need to preempt or enable interrupts:
5525 __release(rq
->lock
);
5526 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5527 _raw_spin_unlock(&rq
->lock
);
5528 preempt_enable_no_resched();
5535 static void __cond_resched(void)
5537 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5538 __might_sleep(__FILE__
, __LINE__
);
5541 * The BKS might be reacquired before we have dropped
5542 * PREEMPT_ACTIVE, which could trigger a second
5543 * cond_resched() call.
5546 add_preempt_count(PREEMPT_ACTIVE
);
5548 sub_preempt_count(PREEMPT_ACTIVE
);
5549 } while (need_resched());
5552 int __sched
_cond_resched(void)
5554 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
5555 system_state
== SYSTEM_RUNNING
) {
5561 EXPORT_SYMBOL(_cond_resched
);
5564 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5565 * call schedule, and on return reacquire the lock.
5567 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5568 * operations here to prevent schedule() from being called twice (once via
5569 * spin_unlock(), once by hand).
5571 int cond_resched_lock(spinlock_t
*lock
)
5573 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
5576 if (spin_needbreak(lock
) || resched
) {
5578 if (resched
&& need_resched())
5587 EXPORT_SYMBOL(cond_resched_lock
);
5589 int __sched
cond_resched_softirq(void)
5591 BUG_ON(!in_softirq());
5593 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
5601 EXPORT_SYMBOL(cond_resched_softirq
);
5604 * yield - yield the current processor to other threads.
5606 * This is a shortcut for kernel-space yielding - it marks the
5607 * thread runnable and calls sys_sched_yield().
5609 void __sched
yield(void)
5611 set_current_state(TASK_RUNNING
);
5614 EXPORT_SYMBOL(yield
);
5617 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5618 * that process accounting knows that this is a task in IO wait state.
5620 * But don't do that if it is a deliberate, throttling IO wait (this task
5621 * has set its backing_dev_info: the queue against which it should throttle)
5623 void __sched
io_schedule(void)
5625 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5627 delayacct_blkio_start();
5628 atomic_inc(&rq
->nr_iowait
);
5630 atomic_dec(&rq
->nr_iowait
);
5631 delayacct_blkio_end();
5633 EXPORT_SYMBOL(io_schedule
);
5635 long __sched
io_schedule_timeout(long timeout
)
5637 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5640 delayacct_blkio_start();
5641 atomic_inc(&rq
->nr_iowait
);
5642 ret
= schedule_timeout(timeout
);
5643 atomic_dec(&rq
->nr_iowait
);
5644 delayacct_blkio_end();
5649 * sys_sched_get_priority_max - return maximum RT priority.
5650 * @policy: scheduling class.
5652 * this syscall returns the maximum rt_priority that can be used
5653 * by a given scheduling class.
5655 asmlinkage
long sys_sched_get_priority_max(int policy
)
5662 ret
= MAX_USER_RT_PRIO
-1;
5674 * sys_sched_get_priority_min - return minimum RT priority.
5675 * @policy: scheduling class.
5677 * this syscall returns the minimum rt_priority that can be used
5678 * by a given scheduling class.
5680 asmlinkage
long sys_sched_get_priority_min(int policy
)
5698 * sys_sched_rr_get_interval - return the default timeslice of a process.
5699 * @pid: pid of the process.
5700 * @interval: userspace pointer to the timeslice value.
5702 * this syscall writes the default timeslice value of a given process
5703 * into the user-space timespec buffer. A value of '0' means infinity.
5706 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
5708 struct task_struct
*p
;
5709 unsigned int time_slice
;
5717 read_lock(&tasklist_lock
);
5718 p
= find_process_by_pid(pid
);
5722 retval
= security_task_getscheduler(p
);
5727 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5728 * tasks that are on an otherwise idle runqueue:
5731 if (p
->policy
== SCHED_RR
) {
5732 time_slice
= DEF_TIMESLICE
;
5733 } else if (p
->policy
!= SCHED_FIFO
) {
5734 struct sched_entity
*se
= &p
->se
;
5735 unsigned long flags
;
5738 rq
= task_rq_lock(p
, &flags
);
5739 if (rq
->cfs
.load
.weight
)
5740 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
5741 task_rq_unlock(rq
, &flags
);
5743 read_unlock(&tasklist_lock
);
5744 jiffies_to_timespec(time_slice
, &t
);
5745 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5749 read_unlock(&tasklist_lock
);
5753 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5755 void sched_show_task(struct task_struct
*p
)
5757 unsigned long free
= 0;
5760 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5761 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5762 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5763 #if BITS_PER_LONG == 32
5764 if (state
== TASK_RUNNING
)
5765 printk(KERN_CONT
" running ");
5767 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5769 if (state
== TASK_RUNNING
)
5770 printk(KERN_CONT
" running task ");
5772 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5774 #ifdef CONFIG_DEBUG_STACK_USAGE
5776 unsigned long *n
= end_of_stack(p
);
5779 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
5782 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
5783 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
5785 show_stack(p
, NULL
);
5788 void show_state_filter(unsigned long state_filter
)
5790 struct task_struct
*g
, *p
;
5792 #if BITS_PER_LONG == 32
5794 " task PC stack pid father\n");
5797 " task PC stack pid father\n");
5799 read_lock(&tasklist_lock
);
5800 do_each_thread(g
, p
) {
5802 * reset the NMI-timeout, listing all files on a slow
5803 * console might take alot of time:
5805 touch_nmi_watchdog();
5806 if (!state_filter
|| (p
->state
& state_filter
))
5808 } while_each_thread(g
, p
);
5810 touch_all_softlockup_watchdogs();
5812 #ifdef CONFIG_SCHED_DEBUG
5813 sysrq_sched_debug_show();
5815 read_unlock(&tasklist_lock
);
5817 * Only show locks if all tasks are dumped:
5819 if (state_filter
== -1)
5820 debug_show_all_locks();
5823 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5825 idle
->sched_class
= &idle_sched_class
;
5829 * init_idle - set up an idle thread for a given CPU
5830 * @idle: task in question
5831 * @cpu: cpu the idle task belongs to
5833 * NOTE: this function does not set the idle thread's NEED_RESCHED
5834 * flag, to make booting more robust.
5836 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5838 struct rq
*rq
= cpu_rq(cpu
);
5839 unsigned long flags
;
5842 idle
->se
.exec_start
= sched_clock();
5844 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
5845 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
5846 __set_task_cpu(idle
, cpu
);
5848 spin_lock_irqsave(&rq
->lock
, flags
);
5849 rq
->curr
= rq
->idle
= idle
;
5850 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5853 spin_unlock_irqrestore(&rq
->lock
, flags
);
5855 /* Set the preempt count _outside_ the spinlocks! */
5856 #if defined(CONFIG_PREEMPT)
5857 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5859 task_thread_info(idle
)->preempt_count
= 0;
5862 * The idle tasks have their own, simple scheduling class:
5864 idle
->sched_class
= &idle_sched_class
;
5868 * In a system that switches off the HZ timer nohz_cpu_mask
5869 * indicates which cpus entered this state. This is used
5870 * in the rcu update to wait only for active cpus. For system
5871 * which do not switch off the HZ timer nohz_cpu_mask should
5872 * always be CPU_MASK_NONE.
5874 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
5877 * Increase the granularity value when there are more CPUs,
5878 * because with more CPUs the 'effective latency' as visible
5879 * to users decreases. But the relationship is not linear,
5880 * so pick a second-best guess by going with the log2 of the
5883 * This idea comes from the SD scheduler of Con Kolivas:
5885 static inline void sched_init_granularity(void)
5887 unsigned int factor
= 1 + ilog2(num_online_cpus());
5888 const unsigned long limit
= 200000000;
5890 sysctl_sched_min_granularity
*= factor
;
5891 if (sysctl_sched_min_granularity
> limit
)
5892 sysctl_sched_min_granularity
= limit
;
5894 sysctl_sched_latency
*= factor
;
5895 if (sysctl_sched_latency
> limit
)
5896 sysctl_sched_latency
= limit
;
5898 sysctl_sched_wakeup_granularity
*= factor
;
5900 sysctl_sched_shares_ratelimit
*= factor
;
5905 * This is how migration works:
5907 * 1) we queue a struct migration_req structure in the source CPU's
5908 * runqueue and wake up that CPU's migration thread.
5909 * 2) we down() the locked semaphore => thread blocks.
5910 * 3) migration thread wakes up (implicitly it forces the migrated
5911 * thread off the CPU)
5912 * 4) it gets the migration request and checks whether the migrated
5913 * task is still in the wrong runqueue.
5914 * 5) if it's in the wrong runqueue then the migration thread removes
5915 * it and puts it into the right queue.
5916 * 6) migration thread up()s the semaphore.
5917 * 7) we wake up and the migration is done.
5921 * Change a given task's CPU affinity. Migrate the thread to a
5922 * proper CPU and schedule it away if the CPU it's executing on
5923 * is removed from the allowed bitmask.
5925 * NOTE: the caller must have a valid reference to the task, the
5926 * task must not exit() & deallocate itself prematurely. The
5927 * call is not atomic; no spinlocks may be held.
5929 int set_cpus_allowed_ptr(struct task_struct
*p
, const cpumask_t
*new_mask
)
5931 struct migration_req req
;
5932 unsigned long flags
;
5936 rq
= task_rq_lock(p
, &flags
);
5937 if (!cpus_intersects(*new_mask
, cpu_online_map
)) {
5942 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
5943 !cpus_equal(p
->cpus_allowed
, *new_mask
))) {
5948 if (p
->sched_class
->set_cpus_allowed
)
5949 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
5951 p
->cpus_allowed
= *new_mask
;
5952 p
->rt
.nr_cpus_allowed
= cpus_weight(*new_mask
);
5955 /* Can the task run on the task's current CPU? If so, we're done */
5956 if (cpu_isset(task_cpu(p
), *new_mask
))
5959 if (migrate_task(p
, any_online_cpu(*new_mask
), &req
)) {
5960 /* Need help from migration thread: drop lock and wait. */
5961 task_rq_unlock(rq
, &flags
);
5962 wake_up_process(rq
->migration_thread
);
5963 wait_for_completion(&req
.done
);
5964 tlb_migrate_finish(p
->mm
);
5968 task_rq_unlock(rq
, &flags
);
5972 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5975 * Move (not current) task off this cpu, onto dest cpu. We're doing
5976 * this because either it can't run here any more (set_cpus_allowed()
5977 * away from this CPU, or CPU going down), or because we're
5978 * attempting to rebalance this task on exec (sched_exec).
5980 * So we race with normal scheduler movements, but that's OK, as long
5981 * as the task is no longer on this CPU.
5983 * Returns non-zero if task was successfully migrated.
5985 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5987 struct rq
*rq_dest
, *rq_src
;
5990 if (unlikely(!cpu_active(dest_cpu
)))
5993 rq_src
= cpu_rq(src_cpu
);
5994 rq_dest
= cpu_rq(dest_cpu
);
5996 double_rq_lock(rq_src
, rq_dest
);
5997 /* Already moved. */
5998 if (task_cpu(p
) != src_cpu
)
6000 /* Affinity changed (again). */
6001 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
6004 on_rq
= p
->se
.on_rq
;
6006 deactivate_task(rq_src
, p
, 0);
6008 set_task_cpu(p
, dest_cpu
);
6010 activate_task(rq_dest
, p
, 0);
6011 check_preempt_curr(rq_dest
, p
, 0);
6016 double_rq_unlock(rq_src
, rq_dest
);
6021 * migration_thread - this is a highprio system thread that performs
6022 * thread migration by bumping thread off CPU then 'pushing' onto
6025 static int migration_thread(void *data
)
6027 int cpu
= (long)data
;
6031 BUG_ON(rq
->migration_thread
!= current
);
6033 set_current_state(TASK_INTERRUPTIBLE
);
6034 while (!kthread_should_stop()) {
6035 struct migration_req
*req
;
6036 struct list_head
*head
;
6038 spin_lock_irq(&rq
->lock
);
6040 if (cpu_is_offline(cpu
)) {
6041 spin_unlock_irq(&rq
->lock
);
6045 if (rq
->active_balance
) {
6046 active_load_balance(rq
, cpu
);
6047 rq
->active_balance
= 0;
6050 head
= &rq
->migration_queue
;
6052 if (list_empty(head
)) {
6053 spin_unlock_irq(&rq
->lock
);
6055 set_current_state(TASK_INTERRUPTIBLE
);
6058 req
= list_entry(head
->next
, struct migration_req
, list
);
6059 list_del_init(head
->next
);
6061 spin_unlock(&rq
->lock
);
6062 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
6065 complete(&req
->done
);
6067 __set_current_state(TASK_RUNNING
);
6071 /* Wait for kthread_stop */
6072 set_current_state(TASK_INTERRUPTIBLE
);
6073 while (!kthread_should_stop()) {
6075 set_current_state(TASK_INTERRUPTIBLE
);
6077 __set_current_state(TASK_RUNNING
);
6081 #ifdef CONFIG_HOTPLUG_CPU
6083 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6087 local_irq_disable();
6088 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
6094 * Figure out where task on dead CPU should go, use force if necessary.
6095 * NOTE: interrupts should be disabled by the caller
6097 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
6099 unsigned long flags
;
6106 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
6107 cpus_and(mask
, mask
, p
->cpus_allowed
);
6108 dest_cpu
= any_online_cpu(mask
);
6110 /* On any allowed CPU? */
6111 if (dest_cpu
>= nr_cpu_ids
)
6112 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
6114 /* No more Mr. Nice Guy. */
6115 if (dest_cpu
>= nr_cpu_ids
) {
6116 cpumask_t cpus_allowed
;
6118 cpuset_cpus_allowed_locked(p
, &cpus_allowed
);
6120 * Try to stay on the same cpuset, where the
6121 * current cpuset may be a subset of all cpus.
6122 * The cpuset_cpus_allowed_locked() variant of
6123 * cpuset_cpus_allowed() will not block. It must be
6124 * called within calls to cpuset_lock/cpuset_unlock.
6126 rq
= task_rq_lock(p
, &flags
);
6127 p
->cpus_allowed
= cpus_allowed
;
6128 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
6129 task_rq_unlock(rq
, &flags
);
6132 * Don't tell them about moving exiting tasks or
6133 * kernel threads (both mm NULL), since they never
6136 if (p
->mm
&& printk_ratelimit()) {
6137 printk(KERN_INFO
"process %d (%s) no "
6138 "longer affine to cpu%d\n",
6139 task_pid_nr(p
), p
->comm
, dead_cpu
);
6142 } while (!__migrate_task_irq(p
, dead_cpu
, dest_cpu
));
6146 * While a dead CPU has no uninterruptible tasks queued at this point,
6147 * it might still have a nonzero ->nr_uninterruptible counter, because
6148 * for performance reasons the counter is not stricly tracking tasks to
6149 * their home CPUs. So we just add the counter to another CPU's counter,
6150 * to keep the global sum constant after CPU-down:
6152 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
6154 struct rq
*rq_dest
= cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR
));
6155 unsigned long flags
;
6157 local_irq_save(flags
);
6158 double_rq_lock(rq_src
, rq_dest
);
6159 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
6160 rq_src
->nr_uninterruptible
= 0;
6161 double_rq_unlock(rq_src
, rq_dest
);
6162 local_irq_restore(flags
);
6165 /* Run through task list and migrate tasks from the dead cpu. */
6166 static void migrate_live_tasks(int src_cpu
)
6168 struct task_struct
*p
, *t
;
6170 read_lock(&tasklist_lock
);
6172 do_each_thread(t
, p
) {
6176 if (task_cpu(p
) == src_cpu
)
6177 move_task_off_dead_cpu(src_cpu
, p
);
6178 } while_each_thread(t
, p
);
6180 read_unlock(&tasklist_lock
);
6184 * Schedules idle task to be the next runnable task on current CPU.
6185 * It does so by boosting its priority to highest possible.
6186 * Used by CPU offline code.
6188 void sched_idle_next(void)
6190 int this_cpu
= smp_processor_id();
6191 struct rq
*rq
= cpu_rq(this_cpu
);
6192 struct task_struct
*p
= rq
->idle
;
6193 unsigned long flags
;
6195 /* cpu has to be offline */
6196 BUG_ON(cpu_online(this_cpu
));
6199 * Strictly not necessary since rest of the CPUs are stopped by now
6200 * and interrupts disabled on the current cpu.
6202 spin_lock_irqsave(&rq
->lock
, flags
);
6204 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6206 update_rq_clock(rq
);
6207 activate_task(rq
, p
, 0);
6209 spin_unlock_irqrestore(&rq
->lock
, flags
);
6213 * Ensures that the idle task is using init_mm right before its cpu goes
6216 void idle_task_exit(void)
6218 struct mm_struct
*mm
= current
->active_mm
;
6220 BUG_ON(cpu_online(smp_processor_id()));
6223 switch_mm(mm
, &init_mm
, current
);
6227 /* called under rq->lock with disabled interrupts */
6228 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
6230 struct rq
*rq
= cpu_rq(dead_cpu
);
6232 /* Must be exiting, otherwise would be on tasklist. */
6233 BUG_ON(!p
->exit_state
);
6235 /* Cannot have done final schedule yet: would have vanished. */
6236 BUG_ON(p
->state
== TASK_DEAD
);
6241 * Drop lock around migration; if someone else moves it,
6242 * that's OK. No task can be added to this CPU, so iteration is
6245 spin_unlock_irq(&rq
->lock
);
6246 move_task_off_dead_cpu(dead_cpu
, p
);
6247 spin_lock_irq(&rq
->lock
);
6252 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6253 static void migrate_dead_tasks(unsigned int dead_cpu
)
6255 struct rq
*rq
= cpu_rq(dead_cpu
);
6256 struct task_struct
*next
;
6259 if (!rq
->nr_running
)
6261 update_rq_clock(rq
);
6262 next
= pick_next_task(rq
, rq
->curr
);
6265 next
->sched_class
->put_prev_task(rq
, next
);
6266 migrate_dead(dead_cpu
, next
);
6270 #endif /* CONFIG_HOTPLUG_CPU */
6272 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6274 static struct ctl_table sd_ctl_dir
[] = {
6276 .procname
= "sched_domain",
6282 static struct ctl_table sd_ctl_root
[] = {
6284 .ctl_name
= CTL_KERN
,
6285 .procname
= "kernel",
6287 .child
= sd_ctl_dir
,
6292 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
6294 struct ctl_table
*entry
=
6295 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
6300 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
6302 struct ctl_table
*entry
;
6305 * In the intermediate directories, both the child directory and
6306 * procname are dynamically allocated and could fail but the mode
6307 * will always be set. In the lowest directory the names are
6308 * static strings and all have proc handlers.
6310 for (entry
= *tablep
; entry
->mode
; entry
++) {
6312 sd_free_ctl_entry(&entry
->child
);
6313 if (entry
->proc_handler
== NULL
)
6314 kfree(entry
->procname
);
6322 set_table_entry(struct ctl_table
*entry
,
6323 const char *procname
, void *data
, int maxlen
,
6324 mode_t mode
, proc_handler
*proc_handler
)
6326 entry
->procname
= procname
;
6328 entry
->maxlen
= maxlen
;
6330 entry
->proc_handler
= proc_handler
;
6333 static struct ctl_table
*
6334 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
6336 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
6341 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
6342 sizeof(long), 0644, proc_doulongvec_minmax
);
6343 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
6344 sizeof(long), 0644, proc_doulongvec_minmax
);
6345 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
6346 sizeof(int), 0644, proc_dointvec_minmax
);
6347 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
6348 sizeof(int), 0644, proc_dointvec_minmax
);
6349 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
6350 sizeof(int), 0644, proc_dointvec_minmax
);
6351 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
6352 sizeof(int), 0644, proc_dointvec_minmax
);
6353 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
6354 sizeof(int), 0644, proc_dointvec_minmax
);
6355 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
6356 sizeof(int), 0644, proc_dointvec_minmax
);
6357 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
6358 sizeof(int), 0644, proc_dointvec_minmax
);
6359 set_table_entry(&table
[9], "cache_nice_tries",
6360 &sd
->cache_nice_tries
,
6361 sizeof(int), 0644, proc_dointvec_minmax
);
6362 set_table_entry(&table
[10], "flags", &sd
->flags
,
6363 sizeof(int), 0644, proc_dointvec_minmax
);
6364 set_table_entry(&table
[11], "name", sd
->name
,
6365 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
6366 /* &table[12] is terminator */
6371 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
6373 struct ctl_table
*entry
, *table
;
6374 struct sched_domain
*sd
;
6375 int domain_num
= 0, i
;
6378 for_each_domain(cpu
, sd
)
6380 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
6385 for_each_domain(cpu
, sd
) {
6386 snprintf(buf
, 32, "domain%d", i
);
6387 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6389 entry
->child
= sd_alloc_ctl_domain_table(sd
);
6396 static struct ctl_table_header
*sd_sysctl_header
;
6397 static void register_sched_domain_sysctl(void)
6399 int i
, cpu_num
= num_online_cpus();
6400 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
6403 WARN_ON(sd_ctl_dir
[0].child
);
6404 sd_ctl_dir
[0].child
= entry
;
6409 for_each_online_cpu(i
) {
6410 snprintf(buf
, 32, "cpu%d", i
);
6411 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6413 entry
->child
= sd_alloc_ctl_cpu_table(i
);
6417 WARN_ON(sd_sysctl_header
);
6418 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
6421 /* may be called multiple times per register */
6422 static void unregister_sched_domain_sysctl(void)
6424 if (sd_sysctl_header
)
6425 unregister_sysctl_table(sd_sysctl_header
);
6426 sd_sysctl_header
= NULL
;
6427 if (sd_ctl_dir
[0].child
)
6428 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
6431 static void register_sched_domain_sysctl(void)
6434 static void unregister_sched_domain_sysctl(void)
6439 static void set_rq_online(struct rq
*rq
)
6442 const struct sched_class
*class;
6444 cpu_set(rq
->cpu
, rq
->rd
->online
);
6447 for_each_class(class) {
6448 if (class->rq_online
)
6449 class->rq_online(rq
);
6454 static void set_rq_offline(struct rq
*rq
)
6457 const struct sched_class
*class;
6459 for_each_class(class) {
6460 if (class->rq_offline
)
6461 class->rq_offline(rq
);
6464 cpu_clear(rq
->cpu
, rq
->rd
->online
);
6470 * migration_call - callback that gets triggered when a CPU is added.
6471 * Here we can start up the necessary migration thread for the new CPU.
6473 static int __cpuinit
6474 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
6476 struct task_struct
*p
;
6477 int cpu
= (long)hcpu
;
6478 unsigned long flags
;
6483 case CPU_UP_PREPARE
:
6484 case CPU_UP_PREPARE_FROZEN
:
6485 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
6488 kthread_bind(p
, cpu
);
6489 /* Must be high prio: stop_machine expects to yield to it. */
6490 rq
= task_rq_lock(p
, &flags
);
6491 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6492 task_rq_unlock(rq
, &flags
);
6493 cpu_rq(cpu
)->migration_thread
= p
;
6497 case CPU_ONLINE_FROZEN
:
6498 /* Strictly unnecessary, as first user will wake it. */
6499 wake_up_process(cpu_rq(cpu
)->migration_thread
);
6501 /* Update our root-domain */
6503 spin_lock_irqsave(&rq
->lock
, flags
);
6505 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6509 spin_unlock_irqrestore(&rq
->lock
, flags
);
6512 #ifdef CONFIG_HOTPLUG_CPU
6513 case CPU_UP_CANCELED
:
6514 case CPU_UP_CANCELED_FROZEN
:
6515 if (!cpu_rq(cpu
)->migration_thread
)
6517 /* Unbind it from offline cpu so it can run. Fall thru. */
6518 kthread_bind(cpu_rq(cpu
)->migration_thread
,
6519 any_online_cpu(cpu_online_map
));
6520 kthread_stop(cpu_rq(cpu
)->migration_thread
);
6521 cpu_rq(cpu
)->migration_thread
= NULL
;
6525 case CPU_DEAD_FROZEN
:
6526 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6527 migrate_live_tasks(cpu
);
6529 kthread_stop(rq
->migration_thread
);
6530 rq
->migration_thread
= NULL
;
6531 /* Idle task back to normal (off runqueue, low prio) */
6532 spin_lock_irq(&rq
->lock
);
6533 update_rq_clock(rq
);
6534 deactivate_task(rq
, rq
->idle
, 0);
6535 rq
->idle
->static_prio
= MAX_PRIO
;
6536 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
6537 rq
->idle
->sched_class
= &idle_sched_class
;
6538 migrate_dead_tasks(cpu
);
6539 spin_unlock_irq(&rq
->lock
);
6541 migrate_nr_uninterruptible(rq
);
6542 BUG_ON(rq
->nr_running
!= 0);
6545 * No need to migrate the tasks: it was best-effort if
6546 * they didn't take sched_hotcpu_mutex. Just wake up
6549 spin_lock_irq(&rq
->lock
);
6550 while (!list_empty(&rq
->migration_queue
)) {
6551 struct migration_req
*req
;
6553 req
= list_entry(rq
->migration_queue
.next
,
6554 struct migration_req
, list
);
6555 list_del_init(&req
->list
);
6556 complete(&req
->done
);
6558 spin_unlock_irq(&rq
->lock
);
6562 case CPU_DYING_FROZEN
:
6563 /* Update our root-domain */
6565 spin_lock_irqsave(&rq
->lock
, flags
);
6567 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6570 spin_unlock_irqrestore(&rq
->lock
, flags
);
6577 /* Register at highest priority so that task migration (migrate_all_tasks)
6578 * happens before everything else.
6580 static struct notifier_block __cpuinitdata migration_notifier
= {
6581 .notifier_call
= migration_call
,
6585 static int __init
migration_init(void)
6587 void *cpu
= (void *)(long)smp_processor_id();
6590 /* Start one for the boot CPU: */
6591 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6592 BUG_ON(err
== NOTIFY_BAD
);
6593 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6594 register_cpu_notifier(&migration_notifier
);
6598 early_initcall(migration_init
);
6603 #ifdef CONFIG_SCHED_DEBUG
6605 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6606 cpumask_t
*groupmask
)
6608 struct sched_group
*group
= sd
->groups
;
6611 cpulist_scnprintf(str
, sizeof(str
), sd
->span
);
6612 cpus_clear(*groupmask
);
6614 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6616 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6617 printk("does not load-balance\n");
6619 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6624 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
6626 if (!cpu_isset(cpu
, sd
->span
)) {
6627 printk(KERN_ERR
"ERROR: domain->span does not contain "
6630 if (!cpu_isset(cpu
, group
->cpumask
)) {
6631 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6635 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6639 printk(KERN_ERR
"ERROR: group is NULL\n");
6643 if (!group
->__cpu_power
) {
6644 printk(KERN_CONT
"\n");
6645 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6650 if (!cpus_weight(group
->cpumask
)) {
6651 printk(KERN_CONT
"\n");
6652 printk(KERN_ERR
"ERROR: empty group\n");
6656 if (cpus_intersects(*groupmask
, group
->cpumask
)) {
6657 printk(KERN_CONT
"\n");
6658 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6662 cpus_or(*groupmask
, *groupmask
, group
->cpumask
);
6664 cpulist_scnprintf(str
, sizeof(str
), group
->cpumask
);
6665 printk(KERN_CONT
" %s", str
);
6667 group
= group
->next
;
6668 } while (group
!= sd
->groups
);
6669 printk(KERN_CONT
"\n");
6671 if (!cpus_equal(sd
->span
, *groupmask
))
6672 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6674 if (sd
->parent
&& !cpus_subset(*groupmask
, sd
->parent
->span
))
6675 printk(KERN_ERR
"ERROR: parent span is not a superset "
6676 "of domain->span\n");
6680 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6682 cpumask_t
*groupmask
;
6686 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6690 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6692 groupmask
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
6694 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6699 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6708 #else /* !CONFIG_SCHED_DEBUG */
6709 # define sched_domain_debug(sd, cpu) do { } while (0)
6710 #endif /* CONFIG_SCHED_DEBUG */
6712 static int sd_degenerate(struct sched_domain
*sd
)
6714 if (cpus_weight(sd
->span
) == 1)
6717 /* Following flags need at least 2 groups */
6718 if (sd
->flags
& (SD_LOAD_BALANCE
|
6719 SD_BALANCE_NEWIDLE
|
6723 SD_SHARE_PKG_RESOURCES
)) {
6724 if (sd
->groups
!= sd
->groups
->next
)
6728 /* Following flags don't use groups */
6729 if (sd
->flags
& (SD_WAKE_IDLE
|
6738 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6740 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6742 if (sd_degenerate(parent
))
6745 if (!cpus_equal(sd
->span
, parent
->span
))
6748 /* Does parent contain flags not in child? */
6749 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6750 if (cflags
& SD_WAKE_AFFINE
)
6751 pflags
&= ~SD_WAKE_BALANCE
;
6752 /* Flags needing groups don't count if only 1 group in parent */
6753 if (parent
->groups
== parent
->groups
->next
) {
6754 pflags
&= ~(SD_LOAD_BALANCE
|
6755 SD_BALANCE_NEWIDLE
|
6759 SD_SHARE_PKG_RESOURCES
);
6761 if (~cflags
& pflags
)
6767 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6769 unsigned long flags
;
6771 spin_lock_irqsave(&rq
->lock
, flags
);
6774 struct root_domain
*old_rd
= rq
->rd
;
6776 if (cpu_isset(rq
->cpu
, old_rd
->online
))
6779 cpu_clear(rq
->cpu
, old_rd
->span
);
6781 if (atomic_dec_and_test(&old_rd
->refcount
))
6785 atomic_inc(&rd
->refcount
);
6788 cpu_set(rq
->cpu
, rd
->span
);
6789 if (cpu_isset(rq
->cpu
, cpu_online_map
))
6792 spin_unlock_irqrestore(&rq
->lock
, flags
);
6795 static void init_rootdomain(struct root_domain
*rd
)
6797 memset(rd
, 0, sizeof(*rd
));
6799 cpus_clear(rd
->span
);
6800 cpus_clear(rd
->online
);
6802 cpupri_init(&rd
->cpupri
);
6805 static void init_defrootdomain(void)
6807 init_rootdomain(&def_root_domain
);
6808 atomic_set(&def_root_domain
.refcount
, 1);
6811 static struct root_domain
*alloc_rootdomain(void)
6813 struct root_domain
*rd
;
6815 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6819 init_rootdomain(rd
);
6825 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6826 * hold the hotplug lock.
6829 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6831 struct rq
*rq
= cpu_rq(cpu
);
6832 struct sched_domain
*tmp
;
6834 /* Remove the sched domains which do not contribute to scheduling. */
6835 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
6836 struct sched_domain
*parent
= tmp
->parent
;
6839 if (sd_parent_degenerate(tmp
, parent
)) {
6840 tmp
->parent
= parent
->parent
;
6842 parent
->parent
->child
= tmp
;
6846 if (sd
&& sd_degenerate(sd
)) {
6852 sched_domain_debug(sd
, cpu
);
6854 rq_attach_root(rq
, rd
);
6855 rcu_assign_pointer(rq
->sd
, sd
);
6858 /* cpus with isolated domains */
6859 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
6861 /* Setup the mask of cpus configured for isolated domains */
6862 static int __init
isolated_cpu_setup(char *str
)
6864 static int __initdata ints
[NR_CPUS
];
6867 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
6868 cpus_clear(cpu_isolated_map
);
6869 for (i
= 1; i
<= ints
[0]; i
++)
6870 if (ints
[i
] < NR_CPUS
)
6871 cpu_set(ints
[i
], cpu_isolated_map
);
6875 __setup("isolcpus=", isolated_cpu_setup
);
6878 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6879 * to a function which identifies what group(along with sched group) a CPU
6880 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6881 * (due to the fact that we keep track of groups covered with a cpumask_t).
6883 * init_sched_build_groups will build a circular linked list of the groups
6884 * covered by the given span, and will set each group's ->cpumask correctly,
6885 * and ->cpu_power to 0.
6888 init_sched_build_groups(const cpumask_t
*span
, const cpumask_t
*cpu_map
,
6889 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
6890 struct sched_group
**sg
,
6891 cpumask_t
*tmpmask
),
6892 cpumask_t
*covered
, cpumask_t
*tmpmask
)
6894 struct sched_group
*first
= NULL
, *last
= NULL
;
6897 cpus_clear(*covered
);
6899 for_each_cpu_mask_nr(i
, *span
) {
6900 struct sched_group
*sg
;
6901 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6904 if (cpu_isset(i
, *covered
))
6907 cpus_clear(sg
->cpumask
);
6908 sg
->__cpu_power
= 0;
6910 for_each_cpu_mask_nr(j
, *span
) {
6911 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6914 cpu_set(j
, *covered
);
6915 cpu_set(j
, sg
->cpumask
);
6926 #define SD_NODES_PER_DOMAIN 16
6931 * find_next_best_node - find the next node to include in a sched_domain
6932 * @node: node whose sched_domain we're building
6933 * @used_nodes: nodes already in the sched_domain
6935 * Find the next node to include in a given scheduling domain. Simply
6936 * finds the closest node not already in the @used_nodes map.
6938 * Should use nodemask_t.
6940 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
6942 int i
, n
, val
, min_val
, best_node
= 0;
6946 for (i
= 0; i
< nr_node_ids
; i
++) {
6947 /* Start at @node */
6948 n
= (node
+ i
) % nr_node_ids
;
6950 if (!nr_cpus_node(n
))
6953 /* Skip already used nodes */
6954 if (node_isset(n
, *used_nodes
))
6957 /* Simple min distance search */
6958 val
= node_distance(node
, n
);
6960 if (val
< min_val
) {
6966 node_set(best_node
, *used_nodes
);
6971 * sched_domain_node_span - get a cpumask for a node's sched_domain
6972 * @node: node whose cpumask we're constructing
6973 * @span: resulting cpumask
6975 * Given a node, construct a good cpumask for its sched_domain to span. It
6976 * should be one that prevents unnecessary balancing, but also spreads tasks
6979 static void sched_domain_node_span(int node
, cpumask_t
*span
)
6981 nodemask_t used_nodes
;
6982 node_to_cpumask_ptr(nodemask
, node
);
6986 nodes_clear(used_nodes
);
6988 cpus_or(*span
, *span
, *nodemask
);
6989 node_set(node
, used_nodes
);
6991 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6992 int next_node
= find_next_best_node(node
, &used_nodes
);
6994 node_to_cpumask_ptr_next(nodemask
, next_node
);
6995 cpus_or(*span
, *span
, *nodemask
);
6998 #endif /* CONFIG_NUMA */
7000 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
7003 * SMT sched-domains:
7005 #ifdef CONFIG_SCHED_SMT
7006 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
7007 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
7010 cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7014 *sg
= &per_cpu(sched_group_cpus
, cpu
);
7017 #endif /* CONFIG_SCHED_SMT */
7020 * multi-core sched-domains:
7022 #ifdef CONFIG_SCHED_MC
7023 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
7024 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
7025 #endif /* CONFIG_SCHED_MC */
7027 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7029 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7034 *mask
= per_cpu(cpu_sibling_map
, cpu
);
7035 cpus_and(*mask
, *mask
, *cpu_map
);
7036 group
= first_cpu(*mask
);
7038 *sg
= &per_cpu(sched_group_core
, group
);
7041 #elif defined(CONFIG_SCHED_MC)
7043 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7047 *sg
= &per_cpu(sched_group_core
, cpu
);
7052 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
7053 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
7056 cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7060 #ifdef CONFIG_SCHED_MC
7061 *mask
= cpu_coregroup_map(cpu
);
7062 cpus_and(*mask
, *mask
, *cpu_map
);
7063 group
= first_cpu(*mask
);
7064 #elif defined(CONFIG_SCHED_SMT)
7065 *mask
= per_cpu(cpu_sibling_map
, cpu
);
7066 cpus_and(*mask
, *mask
, *cpu_map
);
7067 group
= first_cpu(*mask
);
7072 *sg
= &per_cpu(sched_group_phys
, group
);
7078 * The init_sched_build_groups can't handle what we want to do with node
7079 * groups, so roll our own. Now each node has its own list of groups which
7080 * gets dynamically allocated.
7082 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
7083 static struct sched_group
***sched_group_nodes_bycpu
;
7085 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
7086 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
7088 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
7089 struct sched_group
**sg
, cpumask_t
*nodemask
)
7093 *nodemask
= node_to_cpumask(cpu_to_node(cpu
));
7094 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7095 group
= first_cpu(*nodemask
);
7098 *sg
= &per_cpu(sched_group_allnodes
, group
);
7102 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
7104 struct sched_group
*sg
= group_head
;
7110 for_each_cpu_mask_nr(j
, sg
->cpumask
) {
7111 struct sched_domain
*sd
;
7113 sd
= &per_cpu(phys_domains
, j
);
7114 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
7116 * Only add "power" once for each
7122 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
7125 } while (sg
!= group_head
);
7127 #endif /* CONFIG_NUMA */
7130 /* Free memory allocated for various sched_group structures */
7131 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
7135 for_each_cpu_mask_nr(cpu
, *cpu_map
) {
7136 struct sched_group
**sched_group_nodes
7137 = sched_group_nodes_bycpu
[cpu
];
7139 if (!sched_group_nodes
)
7142 for (i
= 0; i
< nr_node_ids
; i
++) {
7143 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
7145 *nodemask
= node_to_cpumask(i
);
7146 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7147 if (cpus_empty(*nodemask
))
7157 if (oldsg
!= sched_group_nodes
[i
])
7160 kfree(sched_group_nodes
);
7161 sched_group_nodes_bycpu
[cpu
] = NULL
;
7164 #else /* !CONFIG_NUMA */
7165 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
7168 #endif /* CONFIG_NUMA */
7171 * Initialize sched groups cpu_power.
7173 * cpu_power indicates the capacity of sched group, which is used while
7174 * distributing the load between different sched groups in a sched domain.
7175 * Typically cpu_power for all the groups in a sched domain will be same unless
7176 * there are asymmetries in the topology. If there are asymmetries, group
7177 * having more cpu_power will pickup more load compared to the group having
7180 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7181 * the maximum number of tasks a group can handle in the presence of other idle
7182 * or lightly loaded groups in the same sched domain.
7184 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
7186 struct sched_domain
*child
;
7187 struct sched_group
*group
;
7189 WARN_ON(!sd
|| !sd
->groups
);
7191 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
7196 sd
->groups
->__cpu_power
= 0;
7199 * For perf policy, if the groups in child domain share resources
7200 * (for example cores sharing some portions of the cache hierarchy
7201 * or SMT), then set this domain groups cpu_power such that each group
7202 * can handle only one task, when there are other idle groups in the
7203 * same sched domain.
7205 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
7207 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
7208 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
7213 * add cpu_power of each child group to this groups cpu_power
7215 group
= child
->groups
;
7217 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
7218 group
= group
->next
;
7219 } while (group
!= child
->groups
);
7223 * Initializers for schedule domains
7224 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7227 #ifdef CONFIG_SCHED_DEBUG
7228 # define SD_INIT_NAME(sd, type) sd->name = #type
7230 # define SD_INIT_NAME(sd, type) do { } while (0)
7233 #define SD_INIT(sd, type) sd_init_##type(sd)
7235 #define SD_INIT_FUNC(type) \
7236 static noinline void sd_init_##type(struct sched_domain *sd) \
7238 memset(sd, 0, sizeof(*sd)); \
7239 *sd = SD_##type##_INIT; \
7240 sd->level = SD_LV_##type; \
7241 SD_INIT_NAME(sd, type); \
7246 SD_INIT_FUNC(ALLNODES
)
7249 #ifdef CONFIG_SCHED_SMT
7250 SD_INIT_FUNC(SIBLING
)
7252 #ifdef CONFIG_SCHED_MC
7257 * To minimize stack usage kmalloc room for cpumasks and share the
7258 * space as the usage in build_sched_domains() dictates. Used only
7259 * if the amount of space is significant.
7262 cpumask_t tmpmask
; /* make this one first */
7265 cpumask_t this_sibling_map
;
7266 cpumask_t this_core_map
;
7268 cpumask_t send_covered
;
7271 cpumask_t domainspan
;
7273 cpumask_t notcovered
;
7278 #define SCHED_CPUMASK_ALLOC 1
7279 #define SCHED_CPUMASK_FREE(v) kfree(v)
7280 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7282 #define SCHED_CPUMASK_ALLOC 0
7283 #define SCHED_CPUMASK_FREE(v)
7284 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7287 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7288 ((unsigned long)(a) + offsetof(struct allmasks, v))
7290 static int default_relax_domain_level
= -1;
7292 static int __init
setup_relax_domain_level(char *str
)
7296 val
= simple_strtoul(str
, NULL
, 0);
7297 if (val
< SD_LV_MAX
)
7298 default_relax_domain_level
= val
;
7302 __setup("relax_domain_level=", setup_relax_domain_level
);
7304 static void set_domain_attribute(struct sched_domain
*sd
,
7305 struct sched_domain_attr
*attr
)
7309 if (!attr
|| attr
->relax_domain_level
< 0) {
7310 if (default_relax_domain_level
< 0)
7313 request
= default_relax_domain_level
;
7315 request
= attr
->relax_domain_level
;
7316 if (request
< sd
->level
) {
7317 /* turn off idle balance on this domain */
7318 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
7320 /* turn on idle balance on this domain */
7321 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
7326 * Build sched domains for a given set of cpus and attach the sched domains
7327 * to the individual cpus
7329 static int __build_sched_domains(const cpumask_t
*cpu_map
,
7330 struct sched_domain_attr
*attr
)
7333 struct root_domain
*rd
;
7334 SCHED_CPUMASK_DECLARE(allmasks
);
7337 struct sched_group
**sched_group_nodes
= NULL
;
7338 int sd_allnodes
= 0;
7341 * Allocate the per-node list of sched groups
7343 sched_group_nodes
= kcalloc(nr_node_ids
, sizeof(struct sched_group
*),
7345 if (!sched_group_nodes
) {
7346 printk(KERN_WARNING
"Can not alloc sched group node list\n");
7351 rd
= alloc_rootdomain();
7353 printk(KERN_WARNING
"Cannot alloc root domain\n");
7355 kfree(sched_group_nodes
);
7360 #if SCHED_CPUMASK_ALLOC
7361 /* get space for all scratch cpumask variables */
7362 allmasks
= kmalloc(sizeof(*allmasks
), GFP_KERNEL
);
7364 printk(KERN_WARNING
"Cannot alloc cpumask array\n");
7367 kfree(sched_group_nodes
);
7372 tmpmask
= (cpumask_t
*)allmasks
;
7376 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
7380 * Set up domains for cpus specified by the cpu_map.
7382 for_each_cpu_mask_nr(i
, *cpu_map
) {
7383 struct sched_domain
*sd
= NULL
, *p
;
7384 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7386 *nodemask
= node_to_cpumask(cpu_to_node(i
));
7387 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7390 if (cpus_weight(*cpu_map
) >
7391 SD_NODES_PER_DOMAIN
*cpus_weight(*nodemask
)) {
7392 sd
= &per_cpu(allnodes_domains
, i
);
7393 SD_INIT(sd
, ALLNODES
);
7394 set_domain_attribute(sd
, attr
);
7395 sd
->span
= *cpu_map
;
7396 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7402 sd
= &per_cpu(node_domains
, i
);
7404 set_domain_attribute(sd
, attr
);
7405 sched_domain_node_span(cpu_to_node(i
), &sd
->span
);
7409 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7413 sd
= &per_cpu(phys_domains
, i
);
7415 set_domain_attribute(sd
, attr
);
7416 sd
->span
= *nodemask
;
7420 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7422 #ifdef CONFIG_SCHED_MC
7424 sd
= &per_cpu(core_domains
, i
);
7426 set_domain_attribute(sd
, attr
);
7427 sd
->span
= cpu_coregroup_map(i
);
7428 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7431 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7434 #ifdef CONFIG_SCHED_SMT
7436 sd
= &per_cpu(cpu_domains
, i
);
7437 SD_INIT(sd
, SIBLING
);
7438 set_domain_attribute(sd
, attr
);
7439 sd
->span
= per_cpu(cpu_sibling_map
, i
);
7440 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7443 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7447 #ifdef CONFIG_SCHED_SMT
7448 /* Set up CPU (sibling) groups */
7449 for_each_cpu_mask_nr(i
, *cpu_map
) {
7450 SCHED_CPUMASK_VAR(this_sibling_map
, allmasks
);
7451 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7453 *this_sibling_map
= per_cpu(cpu_sibling_map
, i
);
7454 cpus_and(*this_sibling_map
, *this_sibling_map
, *cpu_map
);
7455 if (i
!= first_cpu(*this_sibling_map
))
7458 init_sched_build_groups(this_sibling_map
, cpu_map
,
7460 send_covered
, tmpmask
);
7464 #ifdef CONFIG_SCHED_MC
7465 /* Set up multi-core groups */
7466 for_each_cpu_mask_nr(i
, *cpu_map
) {
7467 SCHED_CPUMASK_VAR(this_core_map
, allmasks
);
7468 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7470 *this_core_map
= cpu_coregroup_map(i
);
7471 cpus_and(*this_core_map
, *this_core_map
, *cpu_map
);
7472 if (i
!= first_cpu(*this_core_map
))
7475 init_sched_build_groups(this_core_map
, cpu_map
,
7477 send_covered
, tmpmask
);
7481 /* Set up physical groups */
7482 for (i
= 0; i
< nr_node_ids
; i
++) {
7483 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7484 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7486 *nodemask
= node_to_cpumask(i
);
7487 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7488 if (cpus_empty(*nodemask
))
7491 init_sched_build_groups(nodemask
, cpu_map
,
7493 send_covered
, tmpmask
);
7497 /* Set up node groups */
7499 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7501 init_sched_build_groups(cpu_map
, cpu_map
,
7502 &cpu_to_allnodes_group
,
7503 send_covered
, tmpmask
);
7506 for (i
= 0; i
< nr_node_ids
; i
++) {
7507 /* Set up node groups */
7508 struct sched_group
*sg
, *prev
;
7509 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7510 SCHED_CPUMASK_VAR(domainspan
, allmasks
);
7511 SCHED_CPUMASK_VAR(covered
, allmasks
);
7514 *nodemask
= node_to_cpumask(i
);
7515 cpus_clear(*covered
);
7517 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7518 if (cpus_empty(*nodemask
)) {
7519 sched_group_nodes
[i
] = NULL
;
7523 sched_domain_node_span(i
, domainspan
);
7524 cpus_and(*domainspan
, *domainspan
, *cpu_map
);
7526 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
7528 printk(KERN_WARNING
"Can not alloc domain group for "
7532 sched_group_nodes
[i
] = sg
;
7533 for_each_cpu_mask_nr(j
, *nodemask
) {
7534 struct sched_domain
*sd
;
7536 sd
= &per_cpu(node_domains
, j
);
7539 sg
->__cpu_power
= 0;
7540 sg
->cpumask
= *nodemask
;
7542 cpus_or(*covered
, *covered
, *nodemask
);
7545 for (j
= 0; j
< nr_node_ids
; j
++) {
7546 SCHED_CPUMASK_VAR(notcovered
, allmasks
);
7547 int n
= (i
+ j
) % nr_node_ids
;
7548 node_to_cpumask_ptr(pnodemask
, n
);
7550 cpus_complement(*notcovered
, *covered
);
7551 cpus_and(*tmpmask
, *notcovered
, *cpu_map
);
7552 cpus_and(*tmpmask
, *tmpmask
, *domainspan
);
7553 if (cpus_empty(*tmpmask
))
7556 cpus_and(*tmpmask
, *tmpmask
, *pnodemask
);
7557 if (cpus_empty(*tmpmask
))
7560 sg
= kmalloc_node(sizeof(struct sched_group
),
7564 "Can not alloc domain group for node %d\n", j
);
7567 sg
->__cpu_power
= 0;
7568 sg
->cpumask
= *tmpmask
;
7569 sg
->next
= prev
->next
;
7570 cpus_or(*covered
, *covered
, *tmpmask
);
7577 /* Calculate CPU power for physical packages and nodes */
7578 #ifdef CONFIG_SCHED_SMT
7579 for_each_cpu_mask_nr(i
, *cpu_map
) {
7580 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
7582 init_sched_groups_power(i
, sd
);
7585 #ifdef CONFIG_SCHED_MC
7586 for_each_cpu_mask_nr(i
, *cpu_map
) {
7587 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
7589 init_sched_groups_power(i
, sd
);
7593 for_each_cpu_mask_nr(i
, *cpu_map
) {
7594 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
7596 init_sched_groups_power(i
, sd
);
7600 for (i
= 0; i
< nr_node_ids
; i
++)
7601 init_numa_sched_groups_power(sched_group_nodes
[i
]);
7604 struct sched_group
*sg
;
7606 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
,
7608 init_numa_sched_groups_power(sg
);
7612 /* Attach the domains */
7613 for_each_cpu_mask_nr(i
, *cpu_map
) {
7614 struct sched_domain
*sd
;
7615 #ifdef CONFIG_SCHED_SMT
7616 sd
= &per_cpu(cpu_domains
, i
);
7617 #elif defined(CONFIG_SCHED_MC)
7618 sd
= &per_cpu(core_domains
, i
);
7620 sd
= &per_cpu(phys_domains
, i
);
7622 cpu_attach_domain(sd
, rd
, i
);
7625 SCHED_CPUMASK_FREE((void *)allmasks
);
7630 free_sched_groups(cpu_map
, tmpmask
);
7631 SCHED_CPUMASK_FREE((void *)allmasks
);
7636 static int build_sched_domains(const cpumask_t
*cpu_map
)
7638 return __build_sched_domains(cpu_map
, NULL
);
7641 static cpumask_t
*doms_cur
; /* current sched domains */
7642 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7643 static struct sched_domain_attr
*dattr_cur
;
7644 /* attribues of custom domains in 'doms_cur' */
7647 * Special case: If a kmalloc of a doms_cur partition (array of
7648 * cpumask_t) fails, then fallback to a single sched domain,
7649 * as determined by the single cpumask_t fallback_doms.
7651 static cpumask_t fallback_doms
;
7653 void __attribute__((weak
)) arch_update_cpu_topology(void)
7658 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7659 * For now this just excludes isolated cpus, but could be used to
7660 * exclude other special cases in the future.
7662 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
7666 arch_update_cpu_topology();
7668 doms_cur
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
7670 doms_cur
= &fallback_doms
;
7671 cpus_andnot(*doms_cur
, *cpu_map
, cpu_isolated_map
);
7673 err
= build_sched_domains(doms_cur
);
7674 register_sched_domain_sysctl();
7679 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
,
7682 free_sched_groups(cpu_map
, tmpmask
);
7686 * Detach sched domains from a group of cpus specified in cpu_map
7687 * These cpus will now be attached to the NULL domain
7689 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
7694 for_each_cpu_mask_nr(i
, *cpu_map
)
7695 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7696 synchronize_sched();
7697 arch_destroy_sched_domains(cpu_map
, &tmpmask
);
7700 /* handle null as "default" */
7701 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7702 struct sched_domain_attr
*new, int idx_new
)
7704 struct sched_domain_attr tmp
;
7711 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7712 new ? (new + idx_new
) : &tmp
,
7713 sizeof(struct sched_domain_attr
));
7717 * Partition sched domains as specified by the 'ndoms_new'
7718 * cpumasks in the array doms_new[] of cpumasks. This compares
7719 * doms_new[] to the current sched domain partitioning, doms_cur[].
7720 * It destroys each deleted domain and builds each new domain.
7722 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7723 * The masks don't intersect (don't overlap.) We should setup one
7724 * sched domain for each mask. CPUs not in any of the cpumasks will
7725 * not be load balanced. If the same cpumask appears both in the
7726 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7729 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7730 * ownership of it and will kfree it when done with it. If the caller
7731 * failed the kmalloc call, then it can pass in doms_new == NULL,
7732 * and partition_sched_domains() will fallback to the single partition
7733 * 'fallback_doms', it also forces the domains to be rebuilt.
7735 * If doms_new==NULL it will be replaced with cpu_online_map.
7736 * ndoms_new==0 is a special case for destroying existing domains.
7737 * It will not create the default domain.
7739 * Call with hotplug lock held
7741 void partition_sched_domains(int ndoms_new
, cpumask_t
*doms_new
,
7742 struct sched_domain_attr
*dattr_new
)
7746 mutex_lock(&sched_domains_mutex
);
7748 /* always unregister in case we don't destroy any domains */
7749 unregister_sched_domain_sysctl();
7751 n
= doms_new
? ndoms_new
: 0;
7753 /* Destroy deleted domains */
7754 for (i
= 0; i
< ndoms_cur
; i
++) {
7755 for (j
= 0; j
< n
; j
++) {
7756 if (cpus_equal(doms_cur
[i
], doms_new
[j
])
7757 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7760 /* no match - a current sched domain not in new doms_new[] */
7761 detach_destroy_domains(doms_cur
+ i
);
7766 if (doms_new
== NULL
) {
7768 doms_new
= &fallback_doms
;
7769 cpus_andnot(doms_new
[0], cpu_online_map
, cpu_isolated_map
);
7773 /* Build new domains */
7774 for (i
= 0; i
< ndoms_new
; i
++) {
7775 for (j
= 0; j
< ndoms_cur
; j
++) {
7776 if (cpus_equal(doms_new
[i
], doms_cur
[j
])
7777 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7780 /* no match - add a new doms_new */
7781 __build_sched_domains(doms_new
+ i
,
7782 dattr_new
? dattr_new
+ i
: NULL
);
7787 /* Remember the new sched domains */
7788 if (doms_cur
!= &fallback_doms
)
7790 kfree(dattr_cur
); /* kfree(NULL) is safe */
7791 doms_cur
= doms_new
;
7792 dattr_cur
= dattr_new
;
7793 ndoms_cur
= ndoms_new
;
7795 register_sched_domain_sysctl();
7797 mutex_unlock(&sched_domains_mutex
);
7800 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7801 int arch_reinit_sched_domains(void)
7805 /* Destroy domains first to force the rebuild */
7806 partition_sched_domains(0, NULL
, NULL
);
7808 rebuild_sched_domains();
7814 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7818 if (buf
[0] != '0' && buf
[0] != '1')
7822 sched_smt_power_savings
= (buf
[0] == '1');
7824 sched_mc_power_savings
= (buf
[0] == '1');
7826 ret
= arch_reinit_sched_domains();
7828 return ret
? ret
: count
;
7831 #ifdef CONFIG_SCHED_MC
7832 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
7835 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7837 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
7838 const char *buf
, size_t count
)
7840 return sched_power_savings_store(buf
, count
, 0);
7842 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
7843 sched_mc_power_savings_show
,
7844 sched_mc_power_savings_store
);
7847 #ifdef CONFIG_SCHED_SMT
7848 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
7851 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7853 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
7854 const char *buf
, size_t count
)
7856 return sched_power_savings_store(buf
, count
, 1);
7858 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
7859 sched_smt_power_savings_show
,
7860 sched_smt_power_savings_store
);
7863 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7867 #ifdef CONFIG_SCHED_SMT
7869 err
= sysfs_create_file(&cls
->kset
.kobj
,
7870 &attr_sched_smt_power_savings
.attr
);
7872 #ifdef CONFIG_SCHED_MC
7873 if (!err
&& mc_capable())
7874 err
= sysfs_create_file(&cls
->kset
.kobj
,
7875 &attr_sched_mc_power_savings
.attr
);
7879 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7881 #ifndef CONFIG_CPUSETS
7883 * Add online and remove offline CPUs from the scheduler domains.
7884 * When cpusets are enabled they take over this function.
7886 static int update_sched_domains(struct notifier_block
*nfb
,
7887 unsigned long action
, void *hcpu
)
7891 case CPU_ONLINE_FROZEN
:
7893 case CPU_DEAD_FROZEN
:
7894 partition_sched_domains(1, NULL
, NULL
);
7903 static int update_runtime(struct notifier_block
*nfb
,
7904 unsigned long action
, void *hcpu
)
7906 int cpu
= (int)(long)hcpu
;
7909 case CPU_DOWN_PREPARE
:
7910 case CPU_DOWN_PREPARE_FROZEN
:
7911 disable_runtime(cpu_rq(cpu
));
7914 case CPU_DOWN_FAILED
:
7915 case CPU_DOWN_FAILED_FROZEN
:
7917 case CPU_ONLINE_FROZEN
:
7918 enable_runtime(cpu_rq(cpu
));
7926 void __init
sched_init_smp(void)
7928 cpumask_t non_isolated_cpus
;
7930 #if defined(CONFIG_NUMA)
7931 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
7933 BUG_ON(sched_group_nodes_bycpu
== NULL
);
7936 mutex_lock(&sched_domains_mutex
);
7937 arch_init_sched_domains(&cpu_online_map
);
7938 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
7939 if (cpus_empty(non_isolated_cpus
))
7940 cpu_set(smp_processor_id(), non_isolated_cpus
);
7941 mutex_unlock(&sched_domains_mutex
);
7944 #ifndef CONFIG_CPUSETS
7945 /* XXX: Theoretical race here - CPU may be hotplugged now */
7946 hotcpu_notifier(update_sched_domains
, 0);
7949 /* RT runtime code needs to handle some hotplug events */
7950 hotcpu_notifier(update_runtime
, 0);
7954 /* Move init over to a non-isolated CPU */
7955 if (set_cpus_allowed_ptr(current
, &non_isolated_cpus
) < 0)
7957 sched_init_granularity();
7960 void __init
sched_init_smp(void)
7962 sched_init_granularity();
7964 #endif /* CONFIG_SMP */
7966 int in_sched_functions(unsigned long addr
)
7968 return in_lock_functions(addr
) ||
7969 (addr
>= (unsigned long)__sched_text_start
7970 && addr
< (unsigned long)__sched_text_end
);
7973 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7975 cfs_rq
->tasks_timeline
= RB_ROOT
;
7976 INIT_LIST_HEAD(&cfs_rq
->tasks
);
7977 #ifdef CONFIG_FAIR_GROUP_SCHED
7980 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7983 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7985 struct rt_prio_array
*array
;
7988 array
= &rt_rq
->active
;
7989 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
7990 INIT_LIST_HEAD(array
->queue
+ i
);
7991 __clear_bit(i
, array
->bitmap
);
7993 /* delimiter for bitsearch: */
7994 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
7996 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7997 rt_rq
->highest_prio
= MAX_RT_PRIO
;
8000 rt_rq
->rt_nr_migratory
= 0;
8001 rt_rq
->overloaded
= 0;
8005 rt_rq
->rt_throttled
= 0;
8006 rt_rq
->rt_runtime
= 0;
8007 spin_lock_init(&rt_rq
->rt_runtime_lock
);
8009 #ifdef CONFIG_RT_GROUP_SCHED
8010 rt_rq
->rt_nr_boosted
= 0;
8015 #ifdef CONFIG_FAIR_GROUP_SCHED
8016 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
8017 struct sched_entity
*se
, int cpu
, int add
,
8018 struct sched_entity
*parent
)
8020 struct rq
*rq
= cpu_rq(cpu
);
8021 tg
->cfs_rq
[cpu
] = cfs_rq
;
8022 init_cfs_rq(cfs_rq
, rq
);
8025 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
8028 /* se could be NULL for init_task_group */
8033 se
->cfs_rq
= &rq
->cfs
;
8035 se
->cfs_rq
= parent
->my_q
;
8038 se
->load
.weight
= tg
->shares
;
8039 se
->load
.inv_weight
= 0;
8040 se
->parent
= parent
;
8044 #ifdef CONFIG_RT_GROUP_SCHED
8045 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
8046 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
8047 struct sched_rt_entity
*parent
)
8049 struct rq
*rq
= cpu_rq(cpu
);
8051 tg
->rt_rq
[cpu
] = rt_rq
;
8052 init_rt_rq(rt_rq
, rq
);
8054 rt_rq
->rt_se
= rt_se
;
8055 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8057 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
8059 tg
->rt_se
[cpu
] = rt_se
;
8064 rt_se
->rt_rq
= &rq
->rt
;
8066 rt_se
->rt_rq
= parent
->my_q
;
8068 rt_se
->my_q
= rt_rq
;
8069 rt_se
->parent
= parent
;
8070 INIT_LIST_HEAD(&rt_se
->run_list
);
8074 void __init
sched_init(void)
8077 unsigned long alloc_size
= 0, ptr
;
8079 #ifdef CONFIG_FAIR_GROUP_SCHED
8080 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8082 #ifdef CONFIG_RT_GROUP_SCHED
8083 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8085 #ifdef CONFIG_USER_SCHED
8089 * As sched_init() is called before page_alloc is setup,
8090 * we use alloc_bootmem().
8093 ptr
= (unsigned long)alloc_bootmem(alloc_size
);
8095 #ifdef CONFIG_FAIR_GROUP_SCHED
8096 init_task_group
.se
= (struct sched_entity
**)ptr
;
8097 ptr
+= nr_cpu_ids
* sizeof(void **);
8099 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8100 ptr
+= nr_cpu_ids
* sizeof(void **);
8102 #ifdef CONFIG_USER_SCHED
8103 root_task_group
.se
= (struct sched_entity
**)ptr
;
8104 ptr
+= nr_cpu_ids
* sizeof(void **);
8106 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8107 ptr
+= nr_cpu_ids
* sizeof(void **);
8108 #endif /* CONFIG_USER_SCHED */
8109 #endif /* CONFIG_FAIR_GROUP_SCHED */
8110 #ifdef CONFIG_RT_GROUP_SCHED
8111 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8112 ptr
+= nr_cpu_ids
* sizeof(void **);
8114 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8115 ptr
+= nr_cpu_ids
* sizeof(void **);
8117 #ifdef CONFIG_USER_SCHED
8118 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8119 ptr
+= nr_cpu_ids
* sizeof(void **);
8121 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8122 ptr
+= nr_cpu_ids
* sizeof(void **);
8123 #endif /* CONFIG_USER_SCHED */
8124 #endif /* CONFIG_RT_GROUP_SCHED */
8128 init_defrootdomain();
8131 init_rt_bandwidth(&def_rt_bandwidth
,
8132 global_rt_period(), global_rt_runtime());
8134 #ifdef CONFIG_RT_GROUP_SCHED
8135 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
8136 global_rt_period(), global_rt_runtime());
8137 #ifdef CONFIG_USER_SCHED
8138 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
8139 global_rt_period(), RUNTIME_INF
);
8140 #endif /* CONFIG_USER_SCHED */
8141 #endif /* CONFIG_RT_GROUP_SCHED */
8143 #ifdef CONFIG_GROUP_SCHED
8144 list_add(&init_task_group
.list
, &task_groups
);
8145 INIT_LIST_HEAD(&init_task_group
.children
);
8147 #ifdef CONFIG_USER_SCHED
8148 INIT_LIST_HEAD(&root_task_group
.children
);
8149 init_task_group
.parent
= &root_task_group
;
8150 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
8151 #endif /* CONFIG_USER_SCHED */
8152 #endif /* CONFIG_GROUP_SCHED */
8154 for_each_possible_cpu(i
) {
8158 spin_lock_init(&rq
->lock
);
8160 init_cfs_rq(&rq
->cfs
, rq
);
8161 init_rt_rq(&rq
->rt
, rq
);
8162 #ifdef CONFIG_FAIR_GROUP_SCHED
8163 init_task_group
.shares
= init_task_group_load
;
8164 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
8165 #ifdef CONFIG_CGROUP_SCHED
8167 * How much cpu bandwidth does init_task_group get?
8169 * In case of task-groups formed thr' the cgroup filesystem, it
8170 * gets 100% of the cpu resources in the system. This overall
8171 * system cpu resource is divided among the tasks of
8172 * init_task_group and its child task-groups in a fair manner,
8173 * based on each entity's (task or task-group's) weight
8174 * (se->load.weight).
8176 * In other words, if init_task_group has 10 tasks of weight
8177 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8178 * then A0's share of the cpu resource is:
8180 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8182 * We achieve this by letting init_task_group's tasks sit
8183 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8185 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
8186 #elif defined CONFIG_USER_SCHED
8187 root_task_group
.shares
= NICE_0_LOAD
;
8188 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
8190 * In case of task-groups formed thr' the user id of tasks,
8191 * init_task_group represents tasks belonging to root user.
8192 * Hence it forms a sibling of all subsequent groups formed.
8193 * In this case, init_task_group gets only a fraction of overall
8194 * system cpu resource, based on the weight assigned to root
8195 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8196 * by letting tasks of init_task_group sit in a separate cfs_rq
8197 * (init_cfs_rq) and having one entity represent this group of
8198 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8200 init_tg_cfs_entry(&init_task_group
,
8201 &per_cpu(init_cfs_rq
, i
),
8202 &per_cpu(init_sched_entity
, i
), i
, 1,
8203 root_task_group
.se
[i
]);
8206 #endif /* CONFIG_FAIR_GROUP_SCHED */
8208 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
8209 #ifdef CONFIG_RT_GROUP_SCHED
8210 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
8211 #ifdef CONFIG_CGROUP_SCHED
8212 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
8213 #elif defined CONFIG_USER_SCHED
8214 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
8215 init_tg_rt_entry(&init_task_group
,
8216 &per_cpu(init_rt_rq
, i
),
8217 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
8218 root_task_group
.rt_se
[i
]);
8222 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
8223 rq
->cpu_load
[j
] = 0;
8227 rq
->active_balance
= 0;
8228 rq
->next_balance
= jiffies
;
8232 rq
->migration_thread
= NULL
;
8233 INIT_LIST_HEAD(&rq
->migration_queue
);
8234 rq_attach_root(rq
, &def_root_domain
);
8237 atomic_set(&rq
->nr_iowait
, 0);
8240 set_load_weight(&init_task
);
8242 #ifdef CONFIG_PREEMPT_NOTIFIERS
8243 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
8247 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
8250 #ifdef CONFIG_RT_MUTEXES
8251 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
8255 * The boot idle thread does lazy MMU switching as well:
8257 atomic_inc(&init_mm
.mm_count
);
8258 enter_lazy_tlb(&init_mm
, current
);
8261 * Make us the idle thread. Technically, schedule() should not be
8262 * called from this thread, however somewhere below it might be,
8263 * but because we are the idle thread, we just pick up running again
8264 * when this runqueue becomes "idle".
8266 init_idle(current
, smp_processor_id());
8268 * During early bootup we pretend to be a normal task:
8270 current
->sched_class
= &fair_sched_class
;
8272 scheduler_running
= 1;
8275 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8276 void __might_sleep(char *file
, int line
)
8279 static unsigned long prev_jiffy
; /* ratelimiting */
8281 if ((!in_atomic() && !irqs_disabled()) ||
8282 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
8284 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
8286 prev_jiffy
= jiffies
;
8289 "BUG: sleeping function called from invalid context at %s:%d\n",
8292 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8293 in_atomic(), irqs_disabled(),
8294 current
->pid
, current
->comm
);
8296 debug_show_held_locks(current
);
8297 if (irqs_disabled())
8298 print_irqtrace_events(current
);
8302 EXPORT_SYMBOL(__might_sleep
);
8305 #ifdef CONFIG_MAGIC_SYSRQ
8306 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
8310 update_rq_clock(rq
);
8311 on_rq
= p
->se
.on_rq
;
8313 deactivate_task(rq
, p
, 0);
8314 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
8316 activate_task(rq
, p
, 0);
8317 resched_task(rq
->curr
);
8321 void normalize_rt_tasks(void)
8323 struct task_struct
*g
, *p
;
8324 unsigned long flags
;
8327 read_lock_irqsave(&tasklist_lock
, flags
);
8328 do_each_thread(g
, p
) {
8330 * Only normalize user tasks:
8335 p
->se
.exec_start
= 0;
8336 #ifdef CONFIG_SCHEDSTATS
8337 p
->se
.wait_start
= 0;
8338 p
->se
.sleep_start
= 0;
8339 p
->se
.block_start
= 0;
8344 * Renice negative nice level userspace
8347 if (TASK_NICE(p
) < 0 && p
->mm
)
8348 set_user_nice(p
, 0);
8352 spin_lock(&p
->pi_lock
);
8353 rq
= __task_rq_lock(p
);
8355 normalize_task(rq
, p
);
8357 __task_rq_unlock(rq
);
8358 spin_unlock(&p
->pi_lock
);
8359 } while_each_thread(g
, p
);
8361 read_unlock_irqrestore(&tasklist_lock
, flags
);
8364 #endif /* CONFIG_MAGIC_SYSRQ */
8368 * These functions are only useful for the IA64 MCA handling.
8370 * They can only be called when the whole system has been
8371 * stopped - every CPU needs to be quiescent, and no scheduling
8372 * activity can take place. Using them for anything else would
8373 * be a serious bug, and as a result, they aren't even visible
8374 * under any other configuration.
8378 * curr_task - return the current task for a given cpu.
8379 * @cpu: the processor in question.
8381 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8383 struct task_struct
*curr_task(int cpu
)
8385 return cpu_curr(cpu
);
8389 * set_curr_task - set the current task for a given cpu.
8390 * @cpu: the processor in question.
8391 * @p: the task pointer to set.
8393 * Description: This function must only be used when non-maskable interrupts
8394 * are serviced on a separate stack. It allows the architecture to switch the
8395 * notion of the current task on a cpu in a non-blocking manner. This function
8396 * must be called with all CPU's synchronized, and interrupts disabled, the
8397 * and caller must save the original value of the current task (see
8398 * curr_task() above) and restore that value before reenabling interrupts and
8399 * re-starting the system.
8401 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8403 void set_curr_task(int cpu
, struct task_struct
*p
)
8410 #ifdef CONFIG_FAIR_GROUP_SCHED
8411 static void free_fair_sched_group(struct task_group
*tg
)
8415 for_each_possible_cpu(i
) {
8417 kfree(tg
->cfs_rq
[i
]);
8427 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8429 struct cfs_rq
*cfs_rq
;
8430 struct sched_entity
*se
;
8434 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8437 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8441 tg
->shares
= NICE_0_LOAD
;
8443 for_each_possible_cpu(i
) {
8446 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
8447 GFP_KERNEL
, cpu_to_node(i
));
8451 se
= kzalloc_node(sizeof(struct sched_entity
),
8452 GFP_KERNEL
, cpu_to_node(i
));
8456 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
8465 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8467 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
8468 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
8471 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8473 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
8475 #else /* !CONFG_FAIR_GROUP_SCHED */
8476 static inline void free_fair_sched_group(struct task_group
*tg
)
8481 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8486 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8490 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8493 #endif /* CONFIG_FAIR_GROUP_SCHED */
8495 #ifdef CONFIG_RT_GROUP_SCHED
8496 static void free_rt_sched_group(struct task_group
*tg
)
8500 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8502 for_each_possible_cpu(i
) {
8504 kfree(tg
->rt_rq
[i
]);
8506 kfree(tg
->rt_se
[i
]);
8514 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8516 struct rt_rq
*rt_rq
;
8517 struct sched_rt_entity
*rt_se
;
8521 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8524 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8528 init_rt_bandwidth(&tg
->rt_bandwidth
,
8529 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8531 for_each_possible_cpu(i
) {
8534 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
8535 GFP_KERNEL
, cpu_to_node(i
));
8539 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
8540 GFP_KERNEL
, cpu_to_node(i
));
8544 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
8553 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8555 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8556 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8559 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8561 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8563 #else /* !CONFIG_RT_GROUP_SCHED */
8564 static inline void free_rt_sched_group(struct task_group
*tg
)
8569 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8574 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8578 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8581 #endif /* CONFIG_RT_GROUP_SCHED */
8583 #ifdef CONFIG_GROUP_SCHED
8584 static void free_sched_group(struct task_group
*tg
)
8586 free_fair_sched_group(tg
);
8587 free_rt_sched_group(tg
);
8591 /* allocate runqueue etc for a new task group */
8592 struct task_group
*sched_create_group(struct task_group
*parent
)
8594 struct task_group
*tg
;
8595 unsigned long flags
;
8598 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8600 return ERR_PTR(-ENOMEM
);
8602 if (!alloc_fair_sched_group(tg
, parent
))
8605 if (!alloc_rt_sched_group(tg
, parent
))
8608 spin_lock_irqsave(&task_group_lock
, flags
);
8609 for_each_possible_cpu(i
) {
8610 register_fair_sched_group(tg
, i
);
8611 register_rt_sched_group(tg
, i
);
8613 list_add_rcu(&tg
->list
, &task_groups
);
8615 WARN_ON(!parent
); /* root should already exist */
8617 tg
->parent
= parent
;
8618 INIT_LIST_HEAD(&tg
->children
);
8619 list_add_rcu(&tg
->siblings
, &parent
->children
);
8620 spin_unlock_irqrestore(&task_group_lock
, flags
);
8625 free_sched_group(tg
);
8626 return ERR_PTR(-ENOMEM
);
8629 /* rcu callback to free various structures associated with a task group */
8630 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8632 /* now it should be safe to free those cfs_rqs */
8633 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8636 /* Destroy runqueue etc associated with a task group */
8637 void sched_destroy_group(struct task_group
*tg
)
8639 unsigned long flags
;
8642 spin_lock_irqsave(&task_group_lock
, flags
);
8643 for_each_possible_cpu(i
) {
8644 unregister_fair_sched_group(tg
, i
);
8645 unregister_rt_sched_group(tg
, i
);
8647 list_del_rcu(&tg
->list
);
8648 list_del_rcu(&tg
->siblings
);
8649 spin_unlock_irqrestore(&task_group_lock
, flags
);
8651 /* wait for possible concurrent references to cfs_rqs complete */
8652 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8655 /* change task's runqueue when it moves between groups.
8656 * The caller of this function should have put the task in its new group
8657 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8658 * reflect its new group.
8660 void sched_move_task(struct task_struct
*tsk
)
8663 unsigned long flags
;
8666 rq
= task_rq_lock(tsk
, &flags
);
8668 update_rq_clock(rq
);
8670 running
= task_current(rq
, tsk
);
8671 on_rq
= tsk
->se
.on_rq
;
8674 dequeue_task(rq
, tsk
, 0);
8675 if (unlikely(running
))
8676 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8678 set_task_rq(tsk
, task_cpu(tsk
));
8680 #ifdef CONFIG_FAIR_GROUP_SCHED
8681 if (tsk
->sched_class
->moved_group
)
8682 tsk
->sched_class
->moved_group(tsk
);
8685 if (unlikely(running
))
8686 tsk
->sched_class
->set_curr_task(rq
);
8688 enqueue_task(rq
, tsk
, 0);
8690 task_rq_unlock(rq
, &flags
);
8692 #endif /* CONFIG_GROUP_SCHED */
8694 #ifdef CONFIG_FAIR_GROUP_SCHED
8695 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8697 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8702 dequeue_entity(cfs_rq
, se
, 0);
8704 se
->load
.weight
= shares
;
8705 se
->load
.inv_weight
= 0;
8708 enqueue_entity(cfs_rq
, se
, 0);
8711 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8713 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8714 struct rq
*rq
= cfs_rq
->rq
;
8715 unsigned long flags
;
8717 spin_lock_irqsave(&rq
->lock
, flags
);
8718 __set_se_shares(se
, shares
);
8719 spin_unlock_irqrestore(&rq
->lock
, flags
);
8722 static DEFINE_MUTEX(shares_mutex
);
8724 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8727 unsigned long flags
;
8730 * We can't change the weight of the root cgroup.
8735 if (shares
< MIN_SHARES
)
8736 shares
= MIN_SHARES
;
8737 else if (shares
> MAX_SHARES
)
8738 shares
= MAX_SHARES
;
8740 mutex_lock(&shares_mutex
);
8741 if (tg
->shares
== shares
)
8744 spin_lock_irqsave(&task_group_lock
, flags
);
8745 for_each_possible_cpu(i
)
8746 unregister_fair_sched_group(tg
, i
);
8747 list_del_rcu(&tg
->siblings
);
8748 spin_unlock_irqrestore(&task_group_lock
, flags
);
8750 /* wait for any ongoing reference to this group to finish */
8751 synchronize_sched();
8754 * Now we are free to modify the group's share on each cpu
8755 * w/o tripping rebalance_share or load_balance_fair.
8757 tg
->shares
= shares
;
8758 for_each_possible_cpu(i
) {
8762 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
8763 set_se_shares(tg
->se
[i
], shares
);
8767 * Enable load balance activity on this group, by inserting it back on
8768 * each cpu's rq->leaf_cfs_rq_list.
8770 spin_lock_irqsave(&task_group_lock
, flags
);
8771 for_each_possible_cpu(i
)
8772 register_fair_sched_group(tg
, i
);
8773 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
8774 spin_unlock_irqrestore(&task_group_lock
, flags
);
8776 mutex_unlock(&shares_mutex
);
8780 unsigned long sched_group_shares(struct task_group
*tg
)
8786 #ifdef CONFIG_RT_GROUP_SCHED
8788 * Ensure that the real time constraints are schedulable.
8790 static DEFINE_MUTEX(rt_constraints_mutex
);
8792 static unsigned long to_ratio(u64 period
, u64 runtime
)
8794 if (runtime
== RUNTIME_INF
)
8797 return div64_u64(runtime
<< 20, period
);
8800 /* Must be called with tasklist_lock held */
8801 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8803 struct task_struct
*g
, *p
;
8805 do_each_thread(g
, p
) {
8806 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8808 } while_each_thread(g
, p
);
8813 struct rt_schedulable_data
{
8814 struct task_group
*tg
;
8819 static int tg_schedulable(struct task_group
*tg
, void *data
)
8821 struct rt_schedulable_data
*d
= data
;
8822 struct task_group
*child
;
8823 unsigned long total
, sum
= 0;
8824 u64 period
, runtime
;
8826 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8827 runtime
= tg
->rt_bandwidth
.rt_runtime
;
8830 period
= d
->rt_period
;
8831 runtime
= d
->rt_runtime
;
8835 * Cannot have more runtime than the period.
8837 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8841 * Ensure we don't starve existing RT tasks.
8843 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
8846 total
= to_ratio(period
, runtime
);
8849 * Nobody can have more than the global setting allows.
8851 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
8855 * The sum of our children's runtime should not exceed our own.
8857 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
8858 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
8859 runtime
= child
->rt_bandwidth
.rt_runtime
;
8861 if (child
== d
->tg
) {
8862 period
= d
->rt_period
;
8863 runtime
= d
->rt_runtime
;
8866 sum
+= to_ratio(period
, runtime
);
8875 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8877 struct rt_schedulable_data data
= {
8879 .rt_period
= period
,
8880 .rt_runtime
= runtime
,
8883 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
8886 static int tg_set_bandwidth(struct task_group
*tg
,
8887 u64 rt_period
, u64 rt_runtime
)
8891 mutex_lock(&rt_constraints_mutex
);
8892 read_lock(&tasklist_lock
);
8893 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
8897 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8898 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8899 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8901 for_each_possible_cpu(i
) {
8902 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8904 spin_lock(&rt_rq
->rt_runtime_lock
);
8905 rt_rq
->rt_runtime
= rt_runtime
;
8906 spin_unlock(&rt_rq
->rt_runtime_lock
);
8908 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8910 read_unlock(&tasklist_lock
);
8911 mutex_unlock(&rt_constraints_mutex
);
8916 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8918 u64 rt_runtime
, rt_period
;
8920 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8921 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8922 if (rt_runtime_us
< 0)
8923 rt_runtime
= RUNTIME_INF
;
8925 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8928 long sched_group_rt_runtime(struct task_group
*tg
)
8932 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8935 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8936 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8937 return rt_runtime_us
;
8940 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8942 u64 rt_runtime
, rt_period
;
8944 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8945 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8950 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8953 long sched_group_rt_period(struct task_group
*tg
)
8957 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8958 do_div(rt_period_us
, NSEC_PER_USEC
);
8959 return rt_period_us
;
8962 static int sched_rt_global_constraints(void)
8964 u64 runtime
, period
;
8967 if (sysctl_sched_rt_period
<= 0)
8970 runtime
= global_rt_runtime();
8971 period
= global_rt_period();
8974 * Sanity check on the sysctl variables.
8976 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8979 mutex_lock(&rt_constraints_mutex
);
8980 read_lock(&tasklist_lock
);
8981 ret
= __rt_schedulable(NULL
, 0, 0);
8982 read_unlock(&tasklist_lock
);
8983 mutex_unlock(&rt_constraints_mutex
);
8987 #else /* !CONFIG_RT_GROUP_SCHED */
8988 static int sched_rt_global_constraints(void)
8990 unsigned long flags
;
8993 if (sysctl_sched_rt_period
<= 0)
8996 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8997 for_each_possible_cpu(i
) {
8998 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
9000 spin_lock(&rt_rq
->rt_runtime_lock
);
9001 rt_rq
->rt_runtime
= global_rt_runtime();
9002 spin_unlock(&rt_rq
->rt_runtime_lock
);
9004 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9008 #endif /* CONFIG_RT_GROUP_SCHED */
9010 int sched_rt_handler(struct ctl_table
*table
, int write
,
9011 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
9015 int old_period
, old_runtime
;
9016 static DEFINE_MUTEX(mutex
);
9019 old_period
= sysctl_sched_rt_period
;
9020 old_runtime
= sysctl_sched_rt_runtime
;
9022 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
9024 if (!ret
&& write
) {
9025 ret
= sched_rt_global_constraints();
9027 sysctl_sched_rt_period
= old_period
;
9028 sysctl_sched_rt_runtime
= old_runtime
;
9030 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
9031 def_rt_bandwidth
.rt_period
=
9032 ns_to_ktime(global_rt_period());
9035 mutex_unlock(&mutex
);
9040 #ifdef CONFIG_CGROUP_SCHED
9042 /* return corresponding task_group object of a cgroup */
9043 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
9045 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
9046 struct task_group
, css
);
9049 static struct cgroup_subsys_state
*
9050 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9052 struct task_group
*tg
, *parent
;
9054 if (!cgrp
->parent
) {
9055 /* This is early initialization for the top cgroup */
9056 return &init_task_group
.css
;
9059 parent
= cgroup_tg(cgrp
->parent
);
9060 tg
= sched_create_group(parent
);
9062 return ERR_PTR(-ENOMEM
);
9068 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9070 struct task_group
*tg
= cgroup_tg(cgrp
);
9072 sched_destroy_group(tg
);
9076 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9077 struct task_struct
*tsk
)
9079 #ifdef CONFIG_RT_GROUP_SCHED
9080 /* Don't accept realtime tasks when there is no way for them to run */
9081 if (rt_task(tsk
) && cgroup_tg(cgrp
)->rt_bandwidth
.rt_runtime
== 0)
9084 /* We don't support RT-tasks being in separate groups */
9085 if (tsk
->sched_class
!= &fair_sched_class
)
9093 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9094 struct cgroup
*old_cont
, struct task_struct
*tsk
)
9096 sched_move_task(tsk
);
9099 #ifdef CONFIG_FAIR_GROUP_SCHED
9100 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
9103 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
9106 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
9108 struct task_group
*tg
= cgroup_tg(cgrp
);
9110 return (u64
) tg
->shares
;
9112 #endif /* CONFIG_FAIR_GROUP_SCHED */
9114 #ifdef CONFIG_RT_GROUP_SCHED
9115 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
9118 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
9121 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9123 return sched_group_rt_runtime(cgroup_tg(cgrp
));
9126 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
9129 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
9132 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
9134 return sched_group_rt_period(cgroup_tg(cgrp
));
9136 #endif /* CONFIG_RT_GROUP_SCHED */
9138 static struct cftype cpu_files
[] = {
9139 #ifdef CONFIG_FAIR_GROUP_SCHED
9142 .read_u64
= cpu_shares_read_u64
,
9143 .write_u64
= cpu_shares_write_u64
,
9146 #ifdef CONFIG_RT_GROUP_SCHED
9148 .name
= "rt_runtime_us",
9149 .read_s64
= cpu_rt_runtime_read
,
9150 .write_s64
= cpu_rt_runtime_write
,
9153 .name
= "rt_period_us",
9154 .read_u64
= cpu_rt_period_read_uint
,
9155 .write_u64
= cpu_rt_period_write_uint
,
9160 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
9162 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
9165 struct cgroup_subsys cpu_cgroup_subsys
= {
9167 .create
= cpu_cgroup_create
,
9168 .destroy
= cpu_cgroup_destroy
,
9169 .can_attach
= cpu_cgroup_can_attach
,
9170 .attach
= cpu_cgroup_attach
,
9171 .populate
= cpu_cgroup_populate
,
9172 .subsys_id
= cpu_cgroup_subsys_id
,
9176 #endif /* CONFIG_CGROUP_SCHED */
9178 #ifdef CONFIG_CGROUP_CPUACCT
9181 * CPU accounting code for task groups.
9183 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9184 * (balbir@in.ibm.com).
9187 /* track cpu usage of a group of tasks */
9189 struct cgroup_subsys_state css
;
9190 /* cpuusage holds pointer to a u64-type object on every cpu */
9194 struct cgroup_subsys cpuacct_subsys
;
9196 /* return cpu accounting group corresponding to this container */
9197 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
9199 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
9200 struct cpuacct
, css
);
9203 /* return cpu accounting group to which this task belongs */
9204 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
9206 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
9207 struct cpuacct
, css
);
9210 /* create a new cpu accounting group */
9211 static struct cgroup_subsys_state
*cpuacct_create(
9212 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9214 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
9217 return ERR_PTR(-ENOMEM
);
9219 ca
->cpuusage
= alloc_percpu(u64
);
9220 if (!ca
->cpuusage
) {
9222 return ERR_PTR(-ENOMEM
);
9228 /* destroy an existing cpu accounting group */
9230 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9232 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9234 free_percpu(ca
->cpuusage
);
9238 /* return total cpu usage (in nanoseconds) of a group */
9239 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9241 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9242 u64 totalcpuusage
= 0;
9245 for_each_possible_cpu(i
) {
9246 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
9249 * Take rq->lock to make 64-bit addition safe on 32-bit
9252 spin_lock_irq(&cpu_rq(i
)->lock
);
9253 totalcpuusage
+= *cpuusage
;
9254 spin_unlock_irq(&cpu_rq(i
)->lock
);
9257 return totalcpuusage
;
9260 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
9263 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9272 for_each_possible_cpu(i
) {
9273 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
9275 spin_lock_irq(&cpu_rq(i
)->lock
);
9277 spin_unlock_irq(&cpu_rq(i
)->lock
);
9283 static struct cftype files
[] = {
9286 .read_u64
= cpuusage_read
,
9287 .write_u64
= cpuusage_write
,
9291 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9293 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
9297 * charge this task's execution time to its accounting group.
9299 * called with rq->lock held.
9301 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
9305 if (!cpuacct_subsys
.active
)
9310 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, task_cpu(tsk
));
9312 *cpuusage
+= cputime
;
9316 struct cgroup_subsys cpuacct_subsys
= {
9318 .create
= cpuacct_create
,
9319 .destroy
= cpuacct_destroy
,
9320 .populate
= cpuacct_populate
,
9321 .subsys_id
= cpuacct_subsys_id
,
9323 #endif /* CONFIG_CGROUP_CPUACCT */