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
, *last
;
402 unsigned int 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 void task_rq_unlock_wait(struct task_struct
*p
)
954 struct rq
*rq
= task_rq(p
);
956 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
957 spin_unlock_wait(&rq
->lock
);
960 static void __task_rq_unlock(struct rq
*rq
)
963 spin_unlock(&rq
->lock
);
966 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
969 spin_unlock_irqrestore(&rq
->lock
, *flags
);
973 * this_rq_lock - lock this runqueue and disable interrupts.
975 static struct rq
*this_rq_lock(void)
982 spin_lock(&rq
->lock
);
987 #ifdef CONFIG_SCHED_HRTICK
989 * Use HR-timers to deliver accurate preemption points.
991 * Its all a bit involved since we cannot program an hrt while holding the
992 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
995 * When we get rescheduled we reprogram the hrtick_timer outside of the
1001 * - enabled by features
1002 * - hrtimer is actually high res
1004 static inline int hrtick_enabled(struct rq
*rq
)
1006 if (!sched_feat(HRTICK
))
1008 if (!cpu_active(cpu_of(rq
)))
1010 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1013 static void hrtick_clear(struct rq
*rq
)
1015 if (hrtimer_active(&rq
->hrtick_timer
))
1016 hrtimer_cancel(&rq
->hrtick_timer
);
1020 * High-resolution timer tick.
1021 * Runs from hardirq context with interrupts disabled.
1023 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1025 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1027 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1029 spin_lock(&rq
->lock
);
1030 update_rq_clock(rq
);
1031 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1032 spin_unlock(&rq
->lock
);
1034 return HRTIMER_NORESTART
;
1039 * called from hardirq (IPI) context
1041 static void __hrtick_start(void *arg
)
1043 struct rq
*rq
= arg
;
1045 spin_lock(&rq
->lock
);
1046 hrtimer_restart(&rq
->hrtick_timer
);
1047 rq
->hrtick_csd_pending
= 0;
1048 spin_unlock(&rq
->lock
);
1052 * Called to set the hrtick timer state.
1054 * called with rq->lock held and irqs disabled
1056 static void hrtick_start(struct rq
*rq
, u64 delay
)
1058 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1059 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1061 hrtimer_set_expires(timer
, time
);
1063 if (rq
== this_rq()) {
1064 hrtimer_restart(timer
);
1065 } else if (!rq
->hrtick_csd_pending
) {
1066 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
);
1067 rq
->hrtick_csd_pending
= 1;
1072 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1074 int cpu
= (int)(long)hcpu
;
1077 case CPU_UP_CANCELED
:
1078 case CPU_UP_CANCELED_FROZEN
:
1079 case CPU_DOWN_PREPARE
:
1080 case CPU_DOWN_PREPARE_FROZEN
:
1082 case CPU_DEAD_FROZEN
:
1083 hrtick_clear(cpu_rq(cpu
));
1090 static __init
void init_hrtick(void)
1092 hotcpu_notifier(hotplug_hrtick
, 0);
1096 * Called to set the hrtick timer state.
1098 * called with rq->lock held and irqs disabled
1100 static void hrtick_start(struct rq
*rq
, u64 delay
)
1102 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
), HRTIMER_MODE_REL
);
1105 static inline void init_hrtick(void)
1108 #endif /* CONFIG_SMP */
1110 static void init_rq_hrtick(struct rq
*rq
)
1113 rq
->hrtick_csd_pending
= 0;
1115 rq
->hrtick_csd
.flags
= 0;
1116 rq
->hrtick_csd
.func
= __hrtick_start
;
1117 rq
->hrtick_csd
.info
= rq
;
1120 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1121 rq
->hrtick_timer
.function
= hrtick
;
1122 rq
->hrtick_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_PERCPU
;
1124 #else /* CONFIG_SCHED_HRTICK */
1125 static inline void hrtick_clear(struct rq
*rq
)
1129 static inline void init_rq_hrtick(struct rq
*rq
)
1133 static inline void init_hrtick(void)
1136 #endif /* CONFIG_SCHED_HRTICK */
1139 * resched_task - mark a task 'to be rescheduled now'.
1141 * On UP this means the setting of the need_resched flag, on SMP it
1142 * might also involve a cross-CPU call to trigger the scheduler on
1147 #ifndef tsk_is_polling
1148 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1151 static void resched_task(struct task_struct
*p
)
1155 assert_spin_locked(&task_rq(p
)->lock
);
1157 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
1160 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
1163 if (cpu
== smp_processor_id())
1166 /* NEED_RESCHED must be visible before we test polling */
1168 if (!tsk_is_polling(p
))
1169 smp_send_reschedule(cpu
);
1172 static void resched_cpu(int cpu
)
1174 struct rq
*rq
= cpu_rq(cpu
);
1175 unsigned long flags
;
1177 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1179 resched_task(cpu_curr(cpu
));
1180 spin_unlock_irqrestore(&rq
->lock
, flags
);
1185 * When add_timer_on() enqueues a timer into the timer wheel of an
1186 * idle CPU then this timer might expire before the next timer event
1187 * which is scheduled to wake up that CPU. In case of a completely
1188 * idle system the next event might even be infinite time into the
1189 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1190 * leaves the inner idle loop so the newly added timer is taken into
1191 * account when the CPU goes back to idle and evaluates the timer
1192 * wheel for the next timer event.
1194 void wake_up_idle_cpu(int cpu
)
1196 struct rq
*rq
= cpu_rq(cpu
);
1198 if (cpu
== smp_processor_id())
1202 * This is safe, as this function is called with the timer
1203 * wheel base lock of (cpu) held. When the CPU is on the way
1204 * to idle and has not yet set rq->curr to idle then it will
1205 * be serialized on the timer wheel base lock and take the new
1206 * timer into account automatically.
1208 if (rq
->curr
!= rq
->idle
)
1212 * We can set TIF_RESCHED on the idle task of the other CPU
1213 * lockless. The worst case is that the other CPU runs the
1214 * idle task through an additional NOOP schedule()
1216 set_tsk_thread_flag(rq
->idle
, TIF_NEED_RESCHED
);
1218 /* NEED_RESCHED must be visible before we test polling */
1220 if (!tsk_is_polling(rq
->idle
))
1221 smp_send_reschedule(cpu
);
1223 #endif /* CONFIG_NO_HZ */
1225 #else /* !CONFIG_SMP */
1226 static void resched_task(struct task_struct
*p
)
1228 assert_spin_locked(&task_rq(p
)->lock
);
1229 set_tsk_need_resched(p
);
1231 #endif /* CONFIG_SMP */
1233 #if BITS_PER_LONG == 32
1234 # define WMULT_CONST (~0UL)
1236 # define WMULT_CONST (1UL << 32)
1239 #define WMULT_SHIFT 32
1242 * Shift right and round:
1244 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1247 * delta *= weight / lw
1249 static unsigned long
1250 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1251 struct load_weight
*lw
)
1255 if (!lw
->inv_weight
) {
1256 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1259 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1263 tmp
= (u64
)delta_exec
* weight
;
1265 * Check whether we'd overflow the 64-bit multiplication:
1267 if (unlikely(tmp
> WMULT_CONST
))
1268 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1271 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1273 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1276 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1282 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1289 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1290 * of tasks with abnormal "nice" values across CPUs the contribution that
1291 * each task makes to its run queue's load is weighted according to its
1292 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1293 * scaled version of the new time slice allocation that they receive on time
1297 #define WEIGHT_IDLEPRIO 2
1298 #define WMULT_IDLEPRIO (1 << 31)
1301 * Nice levels are multiplicative, with a gentle 10% change for every
1302 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1303 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1304 * that remained on nice 0.
1306 * The "10% effect" is relative and cumulative: from _any_ nice level,
1307 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1308 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1309 * If a task goes up by ~10% and another task goes down by ~10% then
1310 * the relative distance between them is ~25%.)
1312 static const int prio_to_weight
[40] = {
1313 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1314 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1315 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1316 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1317 /* 0 */ 1024, 820, 655, 526, 423,
1318 /* 5 */ 335, 272, 215, 172, 137,
1319 /* 10 */ 110, 87, 70, 56, 45,
1320 /* 15 */ 36, 29, 23, 18, 15,
1324 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1326 * In cases where the weight does not change often, we can use the
1327 * precalculated inverse to speed up arithmetics by turning divisions
1328 * into multiplications:
1330 static const u32 prio_to_wmult
[40] = {
1331 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1332 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1333 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1334 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1335 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1336 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1337 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1338 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1341 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1344 * runqueue iterator, to support SMP load-balancing between different
1345 * scheduling classes, without having to expose their internal data
1346 * structures to the load-balancing proper:
1348 struct rq_iterator
{
1350 struct task_struct
*(*start
)(void *);
1351 struct task_struct
*(*next
)(void *);
1355 static unsigned long
1356 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1357 unsigned long max_load_move
, struct sched_domain
*sd
,
1358 enum cpu_idle_type idle
, int *all_pinned
,
1359 int *this_best_prio
, struct rq_iterator
*iterator
);
1362 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1363 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1364 struct rq_iterator
*iterator
);
1367 #ifdef CONFIG_CGROUP_CPUACCT
1368 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1370 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1373 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1375 update_load_add(&rq
->load
, load
);
1378 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1380 update_load_sub(&rq
->load
, load
);
1383 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1384 typedef int (*tg_visitor
)(struct task_group
*, void *);
1387 * Iterate the full tree, calling @down when first entering a node and @up when
1388 * leaving it for the final time.
1390 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1392 struct task_group
*parent
, *child
;
1396 parent
= &root_task_group
;
1398 ret
= (*down
)(parent
, data
);
1401 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1408 ret
= (*up
)(parent
, data
);
1413 parent
= parent
->parent
;
1422 static int tg_nop(struct task_group
*tg
, void *data
)
1429 static unsigned long source_load(int cpu
, int type
);
1430 static unsigned long target_load(int cpu
, int type
);
1431 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1433 static unsigned long cpu_avg_load_per_task(int cpu
)
1435 struct rq
*rq
= cpu_rq(cpu
);
1438 rq
->avg_load_per_task
= rq
->load
.weight
/ rq
->nr_running
;
1440 rq
->avg_load_per_task
= 0;
1442 return rq
->avg_load_per_task
;
1445 #ifdef CONFIG_FAIR_GROUP_SCHED
1447 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1450 * Calculate and set the cpu's group shares.
1453 update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1454 unsigned long sd_shares
, unsigned long sd_rq_weight
)
1456 unsigned long shares
;
1457 unsigned long rq_weight
;
1462 rq_weight
= tg
->cfs_rq
[cpu
]->rq_weight
;
1465 * \Sum shares * rq_weight
1466 * shares = -----------------------
1470 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1471 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1473 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1474 sysctl_sched_shares_thresh
) {
1475 struct rq
*rq
= cpu_rq(cpu
);
1476 unsigned long flags
;
1478 spin_lock_irqsave(&rq
->lock
, flags
);
1479 tg
->cfs_rq
[cpu
]->shares
= shares
;
1481 __set_se_shares(tg
->se
[cpu
], shares
);
1482 spin_unlock_irqrestore(&rq
->lock
, flags
);
1487 * Re-compute the task group their per cpu shares over the given domain.
1488 * This needs to be done in a bottom-up fashion because the rq weight of a
1489 * parent group depends on the shares of its child groups.
1491 static int tg_shares_up(struct task_group
*tg
, void *data
)
1493 unsigned long weight
, rq_weight
= 0;
1494 unsigned long shares
= 0;
1495 struct sched_domain
*sd
= data
;
1498 for_each_cpu_mask(i
, sd
->span
) {
1500 * If there are currently no tasks on the cpu pretend there
1501 * is one of average load so that when a new task gets to
1502 * run here it will not get delayed by group starvation.
1504 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1506 weight
= NICE_0_LOAD
;
1508 tg
->cfs_rq
[i
]->rq_weight
= weight
;
1509 rq_weight
+= weight
;
1510 shares
+= tg
->cfs_rq
[i
]->shares
;
1513 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1514 shares
= tg
->shares
;
1516 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1517 shares
= tg
->shares
;
1519 for_each_cpu_mask(i
, sd
->span
)
1520 update_group_shares_cpu(tg
, i
, shares
, rq_weight
);
1526 * Compute the cpu's hierarchical load factor for each task group.
1527 * This needs to be done in a top-down fashion because the load of a child
1528 * group is a fraction of its parents load.
1530 static int tg_load_down(struct task_group
*tg
, void *data
)
1533 long cpu
= (long)data
;
1536 load
= cpu_rq(cpu
)->load
.weight
;
1538 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1539 load
*= tg
->cfs_rq
[cpu
]->shares
;
1540 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1543 tg
->cfs_rq
[cpu
]->h_load
= load
;
1548 static void update_shares(struct sched_domain
*sd
)
1550 u64 now
= cpu_clock(raw_smp_processor_id());
1551 s64 elapsed
= now
- sd
->last_update
;
1553 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1554 sd
->last_update
= now
;
1555 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1559 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1561 spin_unlock(&rq
->lock
);
1563 spin_lock(&rq
->lock
);
1566 static void update_h_load(long cpu
)
1568 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1573 static inline void update_shares(struct sched_domain
*sd
)
1577 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1585 #ifdef CONFIG_FAIR_GROUP_SCHED
1586 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1589 cfs_rq
->shares
= shares
;
1594 #include "sched_stats.h"
1595 #include "sched_idletask.c"
1596 #include "sched_fair.c"
1597 #include "sched_rt.c"
1598 #ifdef CONFIG_SCHED_DEBUG
1599 # include "sched_debug.c"
1602 #define sched_class_highest (&rt_sched_class)
1603 #define for_each_class(class) \
1604 for (class = sched_class_highest; class; class = class->next)
1606 static void inc_nr_running(struct rq
*rq
)
1611 static void dec_nr_running(struct rq
*rq
)
1616 static void set_load_weight(struct task_struct
*p
)
1618 if (task_has_rt_policy(p
)) {
1619 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1620 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1625 * SCHED_IDLE tasks get minimal weight:
1627 if (p
->policy
== SCHED_IDLE
) {
1628 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1629 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1633 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1634 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1637 static void update_avg(u64
*avg
, u64 sample
)
1639 s64 diff
= sample
- *avg
;
1643 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1645 sched_info_queued(p
);
1646 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1650 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1652 if (sleep
&& p
->se
.last_wakeup
) {
1653 update_avg(&p
->se
.avg_overlap
,
1654 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1655 p
->se
.last_wakeup
= 0;
1658 sched_info_dequeued(p
);
1659 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1664 * __normal_prio - return the priority that is based on the static prio
1666 static inline int __normal_prio(struct task_struct
*p
)
1668 return p
->static_prio
;
1672 * Calculate the expected normal priority: i.e. priority
1673 * without taking RT-inheritance into account. Might be
1674 * boosted by interactivity modifiers. Changes upon fork,
1675 * setprio syscalls, and whenever the interactivity
1676 * estimator recalculates.
1678 static inline int normal_prio(struct task_struct
*p
)
1682 if (task_has_rt_policy(p
))
1683 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1685 prio
= __normal_prio(p
);
1690 * Calculate the current priority, i.e. the priority
1691 * taken into account by the scheduler. This value might
1692 * be boosted by RT tasks, or might be boosted by
1693 * interactivity modifiers. Will be RT if the task got
1694 * RT-boosted. If not then it returns p->normal_prio.
1696 static int effective_prio(struct task_struct
*p
)
1698 p
->normal_prio
= normal_prio(p
);
1700 * If we are RT tasks or we were boosted to RT priority,
1701 * keep the priority unchanged. Otherwise, update priority
1702 * to the normal priority:
1704 if (!rt_prio(p
->prio
))
1705 return p
->normal_prio
;
1710 * activate_task - move a task to the runqueue.
1712 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1714 if (task_contributes_to_load(p
))
1715 rq
->nr_uninterruptible
--;
1717 enqueue_task(rq
, p
, wakeup
);
1722 * deactivate_task - remove a task from the runqueue.
1724 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1726 if (task_contributes_to_load(p
))
1727 rq
->nr_uninterruptible
++;
1729 dequeue_task(rq
, p
, sleep
);
1734 * task_curr - is this task currently executing on a CPU?
1735 * @p: the task in question.
1737 inline int task_curr(const struct task_struct
*p
)
1739 return cpu_curr(task_cpu(p
)) == p
;
1742 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1744 set_task_rq(p
, cpu
);
1747 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1748 * successfuly executed on another CPU. We must ensure that updates of
1749 * per-task data have been completed by this moment.
1752 task_thread_info(p
)->cpu
= cpu
;
1756 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1757 const struct sched_class
*prev_class
,
1758 int oldprio
, int running
)
1760 if (prev_class
!= p
->sched_class
) {
1761 if (prev_class
->switched_from
)
1762 prev_class
->switched_from(rq
, p
, running
);
1763 p
->sched_class
->switched_to(rq
, p
, running
);
1765 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1770 /* Used instead of source_load when we know the type == 0 */
1771 static unsigned long weighted_cpuload(const int cpu
)
1773 return cpu_rq(cpu
)->load
.weight
;
1777 * Is this task likely cache-hot:
1780 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1785 * Buddy candidates are cache hot:
1787 if (sched_feat(CACHE_HOT_BUDDY
) &&
1788 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
1789 &p
->se
== cfs_rq_of(&p
->se
)->last
))
1792 if (p
->sched_class
!= &fair_sched_class
)
1795 if (sysctl_sched_migration_cost
== -1)
1797 if (sysctl_sched_migration_cost
== 0)
1800 delta
= now
- p
->se
.exec_start
;
1802 return delta
< (s64
)sysctl_sched_migration_cost
;
1806 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1808 int old_cpu
= task_cpu(p
);
1809 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1810 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1811 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1814 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1816 #ifdef CONFIG_SCHEDSTATS
1817 if (p
->se
.wait_start
)
1818 p
->se
.wait_start
-= clock_offset
;
1819 if (p
->se
.sleep_start
)
1820 p
->se
.sleep_start
-= clock_offset
;
1821 if (p
->se
.block_start
)
1822 p
->se
.block_start
-= clock_offset
;
1823 if (old_cpu
!= new_cpu
) {
1824 schedstat_inc(p
, se
.nr_migrations
);
1825 if (task_hot(p
, old_rq
->clock
, NULL
))
1826 schedstat_inc(p
, se
.nr_forced2_migrations
);
1829 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1830 new_cfsrq
->min_vruntime
;
1832 __set_task_cpu(p
, new_cpu
);
1835 struct migration_req
{
1836 struct list_head list
;
1838 struct task_struct
*task
;
1841 struct completion done
;
1845 * The task's runqueue lock must be held.
1846 * Returns true if you have to wait for migration thread.
1849 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1851 struct rq
*rq
= task_rq(p
);
1854 * If the task is not on a runqueue (and not running), then
1855 * it is sufficient to simply update the task's cpu field.
1857 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1858 set_task_cpu(p
, dest_cpu
);
1862 init_completion(&req
->done
);
1864 req
->dest_cpu
= dest_cpu
;
1865 list_add(&req
->list
, &rq
->migration_queue
);
1871 * wait_task_inactive - wait for a thread to unschedule.
1873 * If @match_state is nonzero, it's the @p->state value just checked and
1874 * not expected to change. If it changes, i.e. @p might have woken up,
1875 * then return zero. When we succeed in waiting for @p to be off its CPU,
1876 * we return a positive number (its total switch count). If a second call
1877 * a short while later returns the same number, the caller can be sure that
1878 * @p has remained unscheduled the whole time.
1880 * The caller must ensure that the task *will* unschedule sometime soon,
1881 * else this function might spin for a *long* time. This function can't
1882 * be called with interrupts off, or it may introduce deadlock with
1883 * smp_call_function() if an IPI is sent by the same process we are
1884 * waiting to become inactive.
1886 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1888 unsigned long flags
;
1895 * We do the initial early heuristics without holding
1896 * any task-queue locks at all. We'll only try to get
1897 * the runqueue lock when things look like they will
1903 * If the task is actively running on another CPU
1904 * still, just relax and busy-wait without holding
1907 * NOTE! Since we don't hold any locks, it's not
1908 * even sure that "rq" stays as the right runqueue!
1909 * But we don't care, since "task_running()" will
1910 * return false if the runqueue has changed and p
1911 * is actually now running somewhere else!
1913 while (task_running(rq
, p
)) {
1914 if (match_state
&& unlikely(p
->state
!= match_state
))
1920 * Ok, time to look more closely! We need the rq
1921 * lock now, to be *sure*. If we're wrong, we'll
1922 * just go back and repeat.
1924 rq
= task_rq_lock(p
, &flags
);
1925 trace_sched_wait_task(rq
, p
);
1926 running
= task_running(rq
, p
);
1927 on_rq
= p
->se
.on_rq
;
1929 if (!match_state
|| p
->state
== match_state
)
1930 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1931 task_rq_unlock(rq
, &flags
);
1934 * If it changed from the expected state, bail out now.
1936 if (unlikely(!ncsw
))
1940 * Was it really running after all now that we
1941 * checked with the proper locks actually held?
1943 * Oops. Go back and try again..
1945 if (unlikely(running
)) {
1951 * It's not enough that it's not actively running,
1952 * it must be off the runqueue _entirely_, and not
1955 * So if it wa still runnable (but just not actively
1956 * running right now), it's preempted, and we should
1957 * yield - it could be a while.
1959 if (unlikely(on_rq
)) {
1960 schedule_timeout_uninterruptible(1);
1965 * Ahh, all good. It wasn't running, and it wasn't
1966 * runnable, which means that it will never become
1967 * running in the future either. We're all done!
1976 * kick_process - kick a running thread to enter/exit the kernel
1977 * @p: the to-be-kicked thread
1979 * Cause a process which is running on another CPU to enter
1980 * kernel-mode, without any delay. (to get signals handled.)
1982 * NOTE: this function doesnt have to take the runqueue lock,
1983 * because all it wants to ensure is that the remote task enters
1984 * the kernel. If the IPI races and the task has been migrated
1985 * to another CPU then no harm is done and the purpose has been
1988 void kick_process(struct task_struct
*p
)
1994 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1995 smp_send_reschedule(cpu
);
2000 * Return a low guess at the load of a migration-source cpu weighted
2001 * according to the scheduling class and "nice" value.
2003 * We want to under-estimate the load of migration sources, to
2004 * balance conservatively.
2006 static unsigned long source_load(int cpu
, int type
)
2008 struct rq
*rq
= cpu_rq(cpu
);
2009 unsigned long total
= weighted_cpuload(cpu
);
2011 if (type
== 0 || !sched_feat(LB_BIAS
))
2014 return min(rq
->cpu_load
[type
-1], total
);
2018 * Return a high guess at the load of a migration-target cpu weighted
2019 * according to the scheduling class and "nice" value.
2021 static unsigned long target_load(int cpu
, int type
)
2023 struct rq
*rq
= cpu_rq(cpu
);
2024 unsigned long total
= weighted_cpuload(cpu
);
2026 if (type
== 0 || !sched_feat(LB_BIAS
))
2029 return max(rq
->cpu_load
[type
-1], total
);
2033 * find_idlest_group finds and returns the least busy CPU group within the
2036 static struct sched_group
*
2037 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
2039 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2040 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
2041 int load_idx
= sd
->forkexec_idx
;
2042 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
2045 unsigned long load
, avg_load
;
2049 /* Skip over this group if it has no CPUs allowed */
2050 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
2053 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2055 /* Tally up the load of all CPUs in the group */
2058 for_each_cpu_mask_nr(i
, group
->cpumask
) {
2059 /* Bias balancing toward cpus of our domain */
2061 load
= source_load(i
, load_idx
);
2063 load
= target_load(i
, load_idx
);
2068 /* Adjust by relative CPU power of the group */
2069 avg_load
= sg_div_cpu_power(group
,
2070 avg_load
* SCHED_LOAD_SCALE
);
2073 this_load
= avg_load
;
2075 } else if (avg_load
< min_load
) {
2076 min_load
= avg_load
;
2079 } while (group
= group
->next
, group
!= sd
->groups
);
2081 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
2087 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2090 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
,
2093 unsigned long load
, min_load
= ULONG_MAX
;
2097 /* Traverse only the allowed CPUs */
2098 cpus_and(*tmp
, group
->cpumask
, p
->cpus_allowed
);
2100 for_each_cpu_mask_nr(i
, *tmp
) {
2101 load
= weighted_cpuload(i
);
2103 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2113 * sched_balance_self: balance the current task (running on cpu) in domains
2114 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2117 * Balance, ie. select the least loaded group.
2119 * Returns the target CPU number, or the same CPU if no balancing is needed.
2121 * preempt must be disabled.
2123 static int sched_balance_self(int cpu
, int flag
)
2125 struct task_struct
*t
= current
;
2126 struct sched_domain
*tmp
, *sd
= NULL
;
2128 for_each_domain(cpu
, tmp
) {
2130 * If power savings logic is enabled for a domain, stop there.
2132 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2134 if (tmp
->flags
& flag
)
2142 cpumask_t span
, tmpmask
;
2143 struct sched_group
*group
;
2144 int new_cpu
, weight
;
2146 if (!(sd
->flags
& flag
)) {
2152 group
= find_idlest_group(sd
, t
, cpu
);
2158 new_cpu
= find_idlest_cpu(group
, t
, cpu
, &tmpmask
);
2159 if (new_cpu
== -1 || new_cpu
== cpu
) {
2160 /* Now try balancing at a lower domain level of cpu */
2165 /* Now try balancing at a lower domain level of new_cpu */
2168 weight
= cpus_weight(span
);
2169 for_each_domain(cpu
, tmp
) {
2170 if (weight
<= cpus_weight(tmp
->span
))
2172 if (tmp
->flags
& flag
)
2175 /* while loop will break here if sd == NULL */
2181 #endif /* CONFIG_SMP */
2184 * try_to_wake_up - wake up a thread
2185 * @p: the to-be-woken-up thread
2186 * @state: the mask of task states that can be woken
2187 * @sync: do a synchronous wakeup?
2189 * Put it on the run-queue if it's not already there. The "current"
2190 * thread is always on the run-queue (except when the actual
2191 * re-schedule is in progress), and as such you're allowed to do
2192 * the simpler "current->state = TASK_RUNNING" to mark yourself
2193 * runnable without the overhead of this.
2195 * returns failure only if the task is already active.
2197 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
2199 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2200 unsigned long flags
;
2204 if (!sched_feat(SYNC_WAKEUPS
))
2208 if (sched_feat(LB_WAKEUP_UPDATE
)) {
2209 struct sched_domain
*sd
;
2211 this_cpu
= raw_smp_processor_id();
2214 for_each_domain(this_cpu
, sd
) {
2215 if (cpu_isset(cpu
, sd
->span
)) {
2224 rq
= task_rq_lock(p
, &flags
);
2225 old_state
= p
->state
;
2226 if (!(old_state
& state
))
2234 this_cpu
= smp_processor_id();
2237 if (unlikely(task_running(rq
, p
)))
2240 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2241 if (cpu
!= orig_cpu
) {
2242 set_task_cpu(p
, cpu
);
2243 task_rq_unlock(rq
, &flags
);
2244 /* might preempt at this point */
2245 rq
= task_rq_lock(p
, &flags
);
2246 old_state
= p
->state
;
2247 if (!(old_state
& state
))
2252 this_cpu
= smp_processor_id();
2256 #ifdef CONFIG_SCHEDSTATS
2257 schedstat_inc(rq
, ttwu_count
);
2258 if (cpu
== this_cpu
)
2259 schedstat_inc(rq
, ttwu_local
);
2261 struct sched_domain
*sd
;
2262 for_each_domain(this_cpu
, sd
) {
2263 if (cpu_isset(cpu
, sd
->span
)) {
2264 schedstat_inc(sd
, ttwu_wake_remote
);
2269 #endif /* CONFIG_SCHEDSTATS */
2272 #endif /* CONFIG_SMP */
2273 schedstat_inc(p
, se
.nr_wakeups
);
2275 schedstat_inc(p
, se
.nr_wakeups_sync
);
2276 if (orig_cpu
!= cpu
)
2277 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2278 if (cpu
== this_cpu
)
2279 schedstat_inc(p
, se
.nr_wakeups_local
);
2281 schedstat_inc(p
, se
.nr_wakeups_remote
);
2282 update_rq_clock(rq
);
2283 activate_task(rq
, p
, 1);
2287 trace_sched_wakeup(rq
, p
);
2288 check_preempt_curr(rq
, p
, sync
);
2290 p
->state
= TASK_RUNNING
;
2292 if (p
->sched_class
->task_wake_up
)
2293 p
->sched_class
->task_wake_up(rq
, p
);
2296 current
->se
.last_wakeup
= current
->se
.sum_exec_runtime
;
2298 task_rq_unlock(rq
, &flags
);
2303 int wake_up_process(struct task_struct
*p
)
2305 return try_to_wake_up(p
, TASK_ALL
, 0);
2307 EXPORT_SYMBOL(wake_up_process
);
2309 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2311 return try_to_wake_up(p
, state
, 0);
2315 * Perform scheduler related setup for a newly forked process p.
2316 * p is forked by current.
2318 * __sched_fork() is basic setup used by init_idle() too:
2320 static void __sched_fork(struct task_struct
*p
)
2322 p
->se
.exec_start
= 0;
2323 p
->se
.sum_exec_runtime
= 0;
2324 p
->se
.prev_sum_exec_runtime
= 0;
2325 p
->se
.last_wakeup
= 0;
2326 p
->se
.avg_overlap
= 0;
2328 #ifdef CONFIG_SCHEDSTATS
2329 p
->se
.wait_start
= 0;
2330 p
->se
.sum_sleep_runtime
= 0;
2331 p
->se
.sleep_start
= 0;
2332 p
->se
.block_start
= 0;
2333 p
->se
.sleep_max
= 0;
2334 p
->se
.block_max
= 0;
2336 p
->se
.slice_max
= 0;
2340 INIT_LIST_HEAD(&p
->rt
.run_list
);
2342 INIT_LIST_HEAD(&p
->se
.group_node
);
2344 #ifdef CONFIG_PREEMPT_NOTIFIERS
2345 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2349 * We mark the process as running here, but have not actually
2350 * inserted it onto the runqueue yet. This guarantees that
2351 * nobody will actually run it, and a signal or other external
2352 * event cannot wake it up and insert it on the runqueue either.
2354 p
->state
= TASK_RUNNING
;
2358 * fork()/clone()-time setup:
2360 void sched_fork(struct task_struct
*p
, int clone_flags
)
2362 int cpu
= get_cpu();
2367 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2369 set_task_cpu(p
, cpu
);
2372 * Make sure we do not leak PI boosting priority to the child:
2374 p
->prio
= current
->normal_prio
;
2375 if (!rt_prio(p
->prio
))
2376 p
->sched_class
= &fair_sched_class
;
2378 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2379 if (likely(sched_info_on()))
2380 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2382 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2385 #ifdef CONFIG_PREEMPT
2386 /* Want to start with kernel preemption disabled. */
2387 task_thread_info(p
)->preempt_count
= 1;
2393 * wake_up_new_task - wake up a newly created task for the first time.
2395 * This function will do some initial scheduler statistics housekeeping
2396 * that must be done for every newly created context, then puts the task
2397 * on the runqueue and wakes it.
2399 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2401 unsigned long flags
;
2404 rq
= task_rq_lock(p
, &flags
);
2405 BUG_ON(p
->state
!= TASK_RUNNING
);
2406 update_rq_clock(rq
);
2408 p
->prio
= effective_prio(p
);
2410 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2411 activate_task(rq
, p
, 0);
2414 * Let the scheduling class do new task startup
2415 * management (if any):
2417 p
->sched_class
->task_new(rq
, p
);
2420 trace_sched_wakeup_new(rq
, p
);
2421 check_preempt_curr(rq
, p
, 0);
2423 if (p
->sched_class
->task_wake_up
)
2424 p
->sched_class
->task_wake_up(rq
, p
);
2426 task_rq_unlock(rq
, &flags
);
2429 #ifdef CONFIG_PREEMPT_NOTIFIERS
2432 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2433 * @notifier: notifier struct to register
2435 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2437 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2439 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2442 * preempt_notifier_unregister - no longer interested in preemption notifications
2443 * @notifier: notifier struct to unregister
2445 * This is safe to call from within a preemption notifier.
2447 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2449 hlist_del(¬ifier
->link
);
2451 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2453 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2455 struct preempt_notifier
*notifier
;
2456 struct hlist_node
*node
;
2458 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2459 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2463 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2464 struct task_struct
*next
)
2466 struct preempt_notifier
*notifier
;
2467 struct hlist_node
*node
;
2469 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2470 notifier
->ops
->sched_out(notifier
, next
);
2473 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2475 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2480 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2481 struct task_struct
*next
)
2485 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2488 * prepare_task_switch - prepare to switch tasks
2489 * @rq: the runqueue preparing to switch
2490 * @prev: the current task that is being switched out
2491 * @next: the task we are going to switch to.
2493 * This is called with the rq lock held and interrupts off. It must
2494 * be paired with a subsequent finish_task_switch after the context
2497 * prepare_task_switch sets up locking and calls architecture specific
2501 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2502 struct task_struct
*next
)
2504 fire_sched_out_preempt_notifiers(prev
, next
);
2505 prepare_lock_switch(rq
, next
);
2506 prepare_arch_switch(next
);
2510 * finish_task_switch - clean up after a task-switch
2511 * @rq: runqueue associated with task-switch
2512 * @prev: the thread we just switched away from.
2514 * finish_task_switch must be called after the context switch, paired
2515 * with a prepare_task_switch call before the context switch.
2516 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2517 * and do any other architecture-specific cleanup actions.
2519 * Note that we may have delayed dropping an mm in context_switch(). If
2520 * so, we finish that here outside of the runqueue lock. (Doing it
2521 * with the lock held can cause deadlocks; see schedule() for
2524 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2525 __releases(rq
->lock
)
2527 struct mm_struct
*mm
= rq
->prev_mm
;
2533 * A task struct has one reference for the use as "current".
2534 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2535 * schedule one last time. The schedule call will never return, and
2536 * the scheduled task must drop that reference.
2537 * The test for TASK_DEAD must occur while the runqueue locks are
2538 * still held, otherwise prev could be scheduled on another cpu, die
2539 * there before we look at prev->state, and then the reference would
2541 * Manfred Spraul <manfred@colorfullife.com>
2543 prev_state
= prev
->state
;
2544 finish_arch_switch(prev
);
2545 finish_lock_switch(rq
, prev
);
2547 if (current
->sched_class
->post_schedule
)
2548 current
->sched_class
->post_schedule(rq
);
2551 fire_sched_in_preempt_notifiers(current
);
2554 if (unlikely(prev_state
== TASK_DEAD
)) {
2556 * Remove function-return probe instances associated with this
2557 * task and put them back on the free list.
2559 kprobe_flush_task(prev
);
2560 put_task_struct(prev
);
2565 * schedule_tail - first thing a freshly forked thread must call.
2566 * @prev: the thread we just switched away from.
2568 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2569 __releases(rq
->lock
)
2571 struct rq
*rq
= this_rq();
2573 finish_task_switch(rq
, prev
);
2574 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2575 /* In this case, finish_task_switch does not reenable preemption */
2578 if (current
->set_child_tid
)
2579 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2583 * context_switch - switch to the new MM and the new
2584 * thread's register state.
2587 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2588 struct task_struct
*next
)
2590 struct mm_struct
*mm
, *oldmm
;
2592 prepare_task_switch(rq
, prev
, next
);
2593 trace_sched_switch(rq
, prev
, next
);
2595 oldmm
= prev
->active_mm
;
2597 * For paravirt, this is coupled with an exit in switch_to to
2598 * combine the page table reload and the switch backend into
2601 arch_enter_lazy_cpu_mode();
2603 if (unlikely(!mm
)) {
2604 next
->active_mm
= oldmm
;
2605 atomic_inc(&oldmm
->mm_count
);
2606 enter_lazy_tlb(oldmm
, next
);
2608 switch_mm(oldmm
, mm
, next
);
2610 if (unlikely(!prev
->mm
)) {
2611 prev
->active_mm
= NULL
;
2612 rq
->prev_mm
= oldmm
;
2615 * Since the runqueue lock will be released by the next
2616 * task (which is an invalid locking op but in the case
2617 * of the scheduler it's an obvious special-case), so we
2618 * do an early lockdep release here:
2620 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2621 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2624 /* Here we just switch the register state and the stack. */
2625 switch_to(prev
, next
, prev
);
2629 * this_rq must be evaluated again because prev may have moved
2630 * CPUs since it called schedule(), thus the 'rq' on its stack
2631 * frame will be invalid.
2633 finish_task_switch(this_rq(), prev
);
2637 * nr_running, nr_uninterruptible and nr_context_switches:
2639 * externally visible scheduler statistics: current number of runnable
2640 * threads, current number of uninterruptible-sleeping threads, total
2641 * number of context switches performed since bootup.
2643 unsigned long nr_running(void)
2645 unsigned long i
, sum
= 0;
2647 for_each_online_cpu(i
)
2648 sum
+= cpu_rq(i
)->nr_running
;
2653 unsigned long nr_uninterruptible(void)
2655 unsigned long i
, sum
= 0;
2657 for_each_possible_cpu(i
)
2658 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2661 * Since we read the counters lockless, it might be slightly
2662 * inaccurate. Do not allow it to go below zero though:
2664 if (unlikely((long)sum
< 0))
2670 unsigned long long nr_context_switches(void)
2673 unsigned long long sum
= 0;
2675 for_each_possible_cpu(i
)
2676 sum
+= cpu_rq(i
)->nr_switches
;
2681 unsigned long nr_iowait(void)
2683 unsigned long i
, sum
= 0;
2685 for_each_possible_cpu(i
)
2686 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2691 unsigned long nr_active(void)
2693 unsigned long i
, running
= 0, uninterruptible
= 0;
2695 for_each_online_cpu(i
) {
2696 running
+= cpu_rq(i
)->nr_running
;
2697 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2700 if (unlikely((long)uninterruptible
< 0))
2701 uninterruptible
= 0;
2703 return running
+ uninterruptible
;
2707 * Update rq->cpu_load[] statistics. This function is usually called every
2708 * scheduler tick (TICK_NSEC).
2710 static void update_cpu_load(struct rq
*this_rq
)
2712 unsigned long this_load
= this_rq
->load
.weight
;
2715 this_rq
->nr_load_updates
++;
2717 /* Update our load: */
2718 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2719 unsigned long old_load
, new_load
;
2721 /* scale is effectively 1 << i now, and >> i divides by scale */
2723 old_load
= this_rq
->cpu_load
[i
];
2724 new_load
= this_load
;
2726 * Round up the averaging division if load is increasing. This
2727 * prevents us from getting stuck on 9 if the load is 10, for
2730 if (new_load
> old_load
)
2731 new_load
+= scale
-1;
2732 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2739 * double_rq_lock - safely lock two runqueues
2741 * Note this does not disable interrupts like task_rq_lock,
2742 * you need to do so manually before calling.
2744 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2745 __acquires(rq1
->lock
)
2746 __acquires(rq2
->lock
)
2748 BUG_ON(!irqs_disabled());
2750 spin_lock(&rq1
->lock
);
2751 __acquire(rq2
->lock
); /* Fake it out ;) */
2754 spin_lock(&rq1
->lock
);
2755 spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
2757 spin_lock(&rq2
->lock
);
2758 spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
2761 update_rq_clock(rq1
);
2762 update_rq_clock(rq2
);
2766 * double_rq_unlock - safely unlock two runqueues
2768 * Note this does not restore interrupts like task_rq_unlock,
2769 * you need to do so manually after calling.
2771 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2772 __releases(rq1
->lock
)
2773 __releases(rq2
->lock
)
2775 spin_unlock(&rq1
->lock
);
2777 spin_unlock(&rq2
->lock
);
2779 __release(rq2
->lock
);
2783 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2785 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2786 __releases(this_rq
->lock
)
2787 __acquires(busiest
->lock
)
2788 __acquires(this_rq
->lock
)
2792 if (unlikely(!irqs_disabled())) {
2793 /* printk() doesn't work good under rq->lock */
2794 spin_unlock(&this_rq
->lock
);
2797 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2798 if (busiest
< this_rq
) {
2799 spin_unlock(&this_rq
->lock
);
2800 spin_lock(&busiest
->lock
);
2801 spin_lock_nested(&this_rq
->lock
, SINGLE_DEPTH_NESTING
);
2804 spin_lock_nested(&busiest
->lock
, SINGLE_DEPTH_NESTING
);
2809 static void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2810 __releases(busiest
->lock
)
2812 spin_unlock(&busiest
->lock
);
2813 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
2817 * If dest_cpu is allowed for this process, migrate the task to it.
2818 * This is accomplished by forcing the cpu_allowed mask to only
2819 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2820 * the cpu_allowed mask is restored.
2822 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2824 struct migration_req req
;
2825 unsigned long flags
;
2828 rq
= task_rq_lock(p
, &flags
);
2829 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2830 || unlikely(!cpu_active(dest_cpu
)))
2833 trace_sched_migrate_task(rq
, p
, dest_cpu
);
2834 /* force the process onto the specified CPU */
2835 if (migrate_task(p
, dest_cpu
, &req
)) {
2836 /* Need to wait for migration thread (might exit: take ref). */
2837 struct task_struct
*mt
= rq
->migration_thread
;
2839 get_task_struct(mt
);
2840 task_rq_unlock(rq
, &flags
);
2841 wake_up_process(mt
);
2842 put_task_struct(mt
);
2843 wait_for_completion(&req
.done
);
2848 task_rq_unlock(rq
, &flags
);
2852 * sched_exec - execve() is a valuable balancing opportunity, because at
2853 * this point the task has the smallest effective memory and cache footprint.
2855 void sched_exec(void)
2857 int new_cpu
, this_cpu
= get_cpu();
2858 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2860 if (new_cpu
!= this_cpu
)
2861 sched_migrate_task(current
, new_cpu
);
2865 * pull_task - move a task from a remote runqueue to the local runqueue.
2866 * Both runqueues must be locked.
2868 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2869 struct rq
*this_rq
, int this_cpu
)
2871 deactivate_task(src_rq
, p
, 0);
2872 set_task_cpu(p
, this_cpu
);
2873 activate_task(this_rq
, p
, 0);
2875 * Note that idle threads have a prio of MAX_PRIO, for this test
2876 * to be always true for them.
2878 check_preempt_curr(this_rq
, p
, 0);
2882 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2885 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2886 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2890 * We do not migrate tasks that are:
2891 * 1) running (obviously), or
2892 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2893 * 3) are cache-hot on their current CPU.
2895 if (!cpu_isset(this_cpu
, p
->cpus_allowed
)) {
2896 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2901 if (task_running(rq
, p
)) {
2902 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2907 * Aggressive migration if:
2908 * 1) task is cache cold, or
2909 * 2) too many balance attempts have failed.
2912 if (!task_hot(p
, rq
->clock
, sd
) ||
2913 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2914 #ifdef CONFIG_SCHEDSTATS
2915 if (task_hot(p
, rq
->clock
, sd
)) {
2916 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2917 schedstat_inc(p
, se
.nr_forced_migrations
);
2923 if (task_hot(p
, rq
->clock
, sd
)) {
2924 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
2930 static unsigned long
2931 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2932 unsigned long max_load_move
, struct sched_domain
*sd
,
2933 enum cpu_idle_type idle
, int *all_pinned
,
2934 int *this_best_prio
, struct rq_iterator
*iterator
)
2936 int loops
= 0, pulled
= 0, pinned
= 0;
2937 struct task_struct
*p
;
2938 long rem_load_move
= max_load_move
;
2940 if (max_load_move
== 0)
2946 * Start the load-balancing iterator:
2948 p
= iterator
->start(iterator
->arg
);
2950 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
2953 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
2954 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2955 p
= iterator
->next(iterator
->arg
);
2959 pull_task(busiest
, p
, this_rq
, this_cpu
);
2961 rem_load_move
-= p
->se
.load
.weight
;
2964 * We only want to steal up to the prescribed amount of weighted load.
2966 if (rem_load_move
> 0) {
2967 if (p
->prio
< *this_best_prio
)
2968 *this_best_prio
= p
->prio
;
2969 p
= iterator
->next(iterator
->arg
);
2974 * Right now, this is one of only two places pull_task() is called,
2975 * so we can safely collect pull_task() stats here rather than
2976 * inside pull_task().
2978 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2981 *all_pinned
= pinned
;
2983 return max_load_move
- rem_load_move
;
2987 * move_tasks tries to move up to max_load_move weighted load from busiest to
2988 * this_rq, as part of a balancing operation within domain "sd".
2989 * Returns 1 if successful and 0 otherwise.
2991 * Called with both runqueues locked.
2993 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2994 unsigned long max_load_move
,
2995 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2998 const struct sched_class
*class = sched_class_highest
;
2999 unsigned long total_load_moved
= 0;
3000 int this_best_prio
= this_rq
->curr
->prio
;
3004 class->load_balance(this_rq
, this_cpu
, busiest
,
3005 max_load_move
- total_load_moved
,
3006 sd
, idle
, all_pinned
, &this_best_prio
);
3007 class = class->next
;
3009 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3012 } while (class && max_load_move
> total_load_moved
);
3014 return total_load_moved
> 0;
3018 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3019 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3020 struct rq_iterator
*iterator
)
3022 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3026 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3027 pull_task(busiest
, p
, this_rq
, this_cpu
);
3029 * Right now, this is only the second place pull_task()
3030 * is called, so we can safely collect pull_task()
3031 * stats here rather than inside pull_task().
3033 schedstat_inc(sd
, lb_gained
[idle
]);
3037 p
= iterator
->next(iterator
->arg
);
3044 * move_one_task tries to move exactly one task from busiest to this_rq, as
3045 * part of active balancing operations within "domain".
3046 * Returns 1 if successful and 0 otherwise.
3048 * Called with both runqueues locked.
3050 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3051 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3053 const struct sched_class
*class;
3055 for (class = sched_class_highest
; class; class = class->next
)
3056 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3063 * find_busiest_group finds and returns the busiest CPU group within the
3064 * domain. It calculates and returns the amount of weighted load which
3065 * should be moved to restore balance via the imbalance parameter.
3067 static struct sched_group
*
3068 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
3069 unsigned long *imbalance
, enum cpu_idle_type idle
,
3070 int *sd_idle
, const cpumask_t
*cpus
, int *balance
)
3072 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
3073 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
3074 unsigned long max_pull
;
3075 unsigned long busiest_load_per_task
, busiest_nr_running
;
3076 unsigned long this_load_per_task
, this_nr_running
;
3077 int load_idx
, group_imb
= 0;
3078 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3079 int power_savings_balance
= 1;
3080 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
3081 unsigned long min_nr_running
= ULONG_MAX
;
3082 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
3085 max_load
= this_load
= total_load
= total_pwr
= 0;
3086 busiest_load_per_task
= busiest_nr_running
= 0;
3087 this_load_per_task
= this_nr_running
= 0;
3089 if (idle
== CPU_NOT_IDLE
)
3090 load_idx
= sd
->busy_idx
;
3091 else if (idle
== CPU_NEWLY_IDLE
)
3092 load_idx
= sd
->newidle_idx
;
3094 load_idx
= sd
->idle_idx
;
3097 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
3100 int __group_imb
= 0;
3101 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3102 unsigned long sum_nr_running
, sum_weighted_load
;
3103 unsigned long sum_avg_load_per_task
;
3104 unsigned long avg_load_per_task
;
3106 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
3109 balance_cpu
= first_cpu(group
->cpumask
);
3111 /* Tally up the load of all CPUs in the group */
3112 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
3113 sum_avg_load_per_task
= avg_load_per_task
= 0;
3116 min_cpu_load
= ~0UL;
3118 for_each_cpu_mask_nr(i
, group
->cpumask
) {
3121 if (!cpu_isset(i
, *cpus
))
3126 if (*sd_idle
&& rq
->nr_running
)
3129 /* Bias balancing toward cpus of our domain */
3131 if (idle_cpu(i
) && !first_idle_cpu
) {
3136 load
= target_load(i
, load_idx
);
3138 load
= source_load(i
, load_idx
);
3139 if (load
> max_cpu_load
)
3140 max_cpu_load
= load
;
3141 if (min_cpu_load
> load
)
3142 min_cpu_load
= load
;
3146 sum_nr_running
+= rq
->nr_running
;
3147 sum_weighted_load
+= weighted_cpuload(i
);
3149 sum_avg_load_per_task
+= cpu_avg_load_per_task(i
);
3153 * First idle cpu or the first cpu(busiest) in this sched group
3154 * is eligible for doing load balancing at this and above
3155 * domains. In the newly idle case, we will allow all the cpu's
3156 * to do the newly idle load balance.
3158 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3159 balance_cpu
!= this_cpu
&& balance
) {
3164 total_load
+= avg_load
;
3165 total_pwr
+= group
->__cpu_power
;
3167 /* Adjust by relative CPU power of the group */
3168 avg_load
= sg_div_cpu_power(group
,
3169 avg_load
* SCHED_LOAD_SCALE
);
3173 * Consider the group unbalanced when the imbalance is larger
3174 * than the average weight of two tasks.
3176 * APZ: with cgroup the avg task weight can vary wildly and
3177 * might not be a suitable number - should we keep a
3178 * normalized nr_running number somewhere that negates
3181 avg_load_per_task
= sg_div_cpu_power(group
,
3182 sum_avg_load_per_task
* SCHED_LOAD_SCALE
);
3184 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3187 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
3190 this_load
= avg_load
;
3192 this_nr_running
= sum_nr_running
;
3193 this_load_per_task
= sum_weighted_load
;
3194 } else if (avg_load
> max_load
&&
3195 (sum_nr_running
> group_capacity
|| __group_imb
)) {
3196 max_load
= avg_load
;
3198 busiest_nr_running
= sum_nr_running
;
3199 busiest_load_per_task
= sum_weighted_load
;
3200 group_imb
= __group_imb
;
3203 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3205 * Busy processors will not participate in power savings
3208 if (idle
== CPU_NOT_IDLE
||
3209 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3213 * If the local group is idle or completely loaded
3214 * no need to do power savings balance at this domain
3216 if (local_group
&& (this_nr_running
>= group_capacity
||
3218 power_savings_balance
= 0;
3221 * If a group is already running at full capacity or idle,
3222 * don't include that group in power savings calculations
3224 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
3229 * Calculate the group which has the least non-idle load.
3230 * This is the group from where we need to pick up the load
3233 if ((sum_nr_running
< min_nr_running
) ||
3234 (sum_nr_running
== min_nr_running
&&
3235 first_cpu(group
->cpumask
) <
3236 first_cpu(group_min
->cpumask
))) {
3238 min_nr_running
= sum_nr_running
;
3239 min_load_per_task
= sum_weighted_load
/
3244 * Calculate the group which is almost near its
3245 * capacity but still has some space to pick up some load
3246 * from other group and save more power
3248 if (sum_nr_running
<= group_capacity
- 1) {
3249 if (sum_nr_running
> leader_nr_running
||
3250 (sum_nr_running
== leader_nr_running
&&
3251 first_cpu(group
->cpumask
) >
3252 first_cpu(group_leader
->cpumask
))) {
3253 group_leader
= group
;
3254 leader_nr_running
= sum_nr_running
;
3259 group
= group
->next
;
3260 } while (group
!= sd
->groups
);
3262 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
3265 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
3267 if (this_load
>= avg_load
||
3268 100*max_load
<= sd
->imbalance_pct
*this_load
)
3271 busiest_load_per_task
/= busiest_nr_running
;
3273 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
3276 * We're trying to get all the cpus to the average_load, so we don't
3277 * want to push ourselves above the average load, nor do we wish to
3278 * reduce the max loaded cpu below the average load, as either of these
3279 * actions would just result in more rebalancing later, and ping-pong
3280 * tasks around. Thus we look for the minimum possible imbalance.
3281 * Negative imbalances (*we* are more loaded than anyone else) will
3282 * be counted as no imbalance for these purposes -- we can't fix that
3283 * by pulling tasks to us. Be careful of negative numbers as they'll
3284 * appear as very large values with unsigned longs.
3286 if (max_load
<= busiest_load_per_task
)
3290 * In the presence of smp nice balancing, certain scenarios can have
3291 * max load less than avg load(as we skip the groups at or below
3292 * its cpu_power, while calculating max_load..)
3294 if (max_load
< avg_load
) {
3296 goto small_imbalance
;
3299 /* Don't want to pull so many tasks that a group would go idle */
3300 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
3302 /* How much load to actually move to equalise the imbalance */
3303 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
3304 (avg_load
- this_load
) * this->__cpu_power
)
3308 * if *imbalance is less than the average load per runnable task
3309 * there is no gaurantee that any tasks will be moved so we'll have
3310 * a think about bumping its value to force at least one task to be
3313 if (*imbalance
< busiest_load_per_task
) {
3314 unsigned long tmp
, pwr_now
, pwr_move
;
3318 pwr_move
= pwr_now
= 0;
3320 if (this_nr_running
) {
3321 this_load_per_task
/= this_nr_running
;
3322 if (busiest_load_per_task
> this_load_per_task
)
3325 this_load_per_task
= cpu_avg_load_per_task(this_cpu
);
3327 if (max_load
- this_load
+ busiest_load_per_task
>=
3328 busiest_load_per_task
* imbn
) {
3329 *imbalance
= busiest_load_per_task
;
3334 * OK, we don't have enough imbalance to justify moving tasks,
3335 * however we may be able to increase total CPU power used by
3339 pwr_now
+= busiest
->__cpu_power
*
3340 min(busiest_load_per_task
, max_load
);
3341 pwr_now
+= this->__cpu_power
*
3342 min(this_load_per_task
, this_load
);
3343 pwr_now
/= SCHED_LOAD_SCALE
;
3345 /* Amount of load we'd subtract */
3346 tmp
= sg_div_cpu_power(busiest
,
3347 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3349 pwr_move
+= busiest
->__cpu_power
*
3350 min(busiest_load_per_task
, max_load
- tmp
);
3352 /* Amount of load we'd add */
3353 if (max_load
* busiest
->__cpu_power
<
3354 busiest_load_per_task
* SCHED_LOAD_SCALE
)
3355 tmp
= sg_div_cpu_power(this,
3356 max_load
* busiest
->__cpu_power
);
3358 tmp
= sg_div_cpu_power(this,
3359 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3360 pwr_move
+= this->__cpu_power
*
3361 min(this_load_per_task
, this_load
+ tmp
);
3362 pwr_move
/= SCHED_LOAD_SCALE
;
3364 /* Move if we gain throughput */
3365 if (pwr_move
> pwr_now
)
3366 *imbalance
= busiest_load_per_task
;
3372 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3373 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3376 if (this == group_leader
&& group_leader
!= group_min
) {
3377 *imbalance
= min_load_per_task
;
3387 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3390 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
3391 unsigned long imbalance
, const cpumask_t
*cpus
)
3393 struct rq
*busiest
= NULL
, *rq
;
3394 unsigned long max_load
= 0;
3397 for_each_cpu_mask_nr(i
, group
->cpumask
) {
3400 if (!cpu_isset(i
, *cpus
))
3404 wl
= weighted_cpuload(i
);
3406 if (rq
->nr_running
== 1 && wl
> imbalance
)
3409 if (wl
> max_load
) {
3419 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3420 * so long as it is large enough.
3422 #define MAX_PINNED_INTERVAL 512
3425 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3426 * tasks if there is an imbalance.
3428 static int load_balance(int this_cpu
, struct rq
*this_rq
,
3429 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3430 int *balance
, cpumask_t
*cpus
)
3432 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
3433 struct sched_group
*group
;
3434 unsigned long imbalance
;
3436 unsigned long flags
;
3441 * When power savings policy is enabled for the parent domain, idle
3442 * sibling can pick up load irrespective of busy siblings. In this case,
3443 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3444 * portraying it as CPU_NOT_IDLE.
3446 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3447 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3450 schedstat_inc(sd
, lb_count
[idle
]);
3454 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3461 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3465 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
3467 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3471 BUG_ON(busiest
== this_rq
);
3473 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3476 if (busiest
->nr_running
> 1) {
3478 * Attempt to move tasks. If find_busiest_group has found
3479 * an imbalance but busiest->nr_running <= 1, the group is
3480 * still unbalanced. ld_moved simply stays zero, so it is
3481 * correctly treated as an imbalance.
3483 local_irq_save(flags
);
3484 double_rq_lock(this_rq
, busiest
);
3485 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3486 imbalance
, sd
, idle
, &all_pinned
);
3487 double_rq_unlock(this_rq
, busiest
);
3488 local_irq_restore(flags
);
3491 * some other cpu did the load balance for us.
3493 if (ld_moved
&& this_cpu
!= smp_processor_id())
3494 resched_cpu(this_cpu
);
3496 /* All tasks on this runqueue were pinned by CPU affinity */
3497 if (unlikely(all_pinned
)) {
3498 cpu_clear(cpu_of(busiest
), *cpus
);
3499 if (!cpus_empty(*cpus
))
3506 schedstat_inc(sd
, lb_failed
[idle
]);
3507 sd
->nr_balance_failed
++;
3509 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3511 spin_lock_irqsave(&busiest
->lock
, flags
);
3513 /* don't kick the migration_thread, if the curr
3514 * task on busiest cpu can't be moved to this_cpu
3516 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
3517 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3519 goto out_one_pinned
;
3522 if (!busiest
->active_balance
) {
3523 busiest
->active_balance
= 1;
3524 busiest
->push_cpu
= this_cpu
;
3527 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3529 wake_up_process(busiest
->migration_thread
);
3532 * We've kicked active balancing, reset the failure
3535 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3538 sd
->nr_balance_failed
= 0;
3540 if (likely(!active_balance
)) {
3541 /* We were unbalanced, so reset the balancing interval */
3542 sd
->balance_interval
= sd
->min_interval
;
3545 * If we've begun active balancing, start to back off. This
3546 * case may not be covered by the all_pinned logic if there
3547 * is only 1 task on the busy runqueue (because we don't call
3550 if (sd
->balance_interval
< sd
->max_interval
)
3551 sd
->balance_interval
*= 2;
3554 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3555 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3561 schedstat_inc(sd
, lb_balanced
[idle
]);
3563 sd
->nr_balance_failed
= 0;
3566 /* tune up the balancing interval */
3567 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3568 (sd
->balance_interval
< sd
->max_interval
))
3569 sd
->balance_interval
*= 2;
3571 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3572 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3583 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3584 * tasks if there is an imbalance.
3586 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3587 * this_rq is locked.
3590 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
,
3593 struct sched_group
*group
;
3594 struct rq
*busiest
= NULL
;
3595 unsigned long imbalance
;
3603 * When power savings policy is enabled for the parent domain, idle
3604 * sibling can pick up load irrespective of busy siblings. In this case,
3605 * let the state of idle sibling percolate up as IDLE, instead of
3606 * portraying it as CPU_NOT_IDLE.
3608 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
3609 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3612 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
3614 update_shares_locked(this_rq
, sd
);
3615 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
3616 &sd_idle
, cpus
, NULL
);
3618 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
3622 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
3624 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
3628 BUG_ON(busiest
== this_rq
);
3630 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
3633 if (busiest
->nr_running
> 1) {
3634 /* Attempt to move tasks */
3635 double_lock_balance(this_rq
, busiest
);
3636 /* this_rq->clock is already updated */
3637 update_rq_clock(busiest
);
3638 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3639 imbalance
, sd
, CPU_NEWLY_IDLE
,
3641 double_unlock_balance(this_rq
, busiest
);
3643 if (unlikely(all_pinned
)) {
3644 cpu_clear(cpu_of(busiest
), *cpus
);
3645 if (!cpus_empty(*cpus
))
3651 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
3652 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3653 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3656 sd
->nr_balance_failed
= 0;
3658 update_shares_locked(this_rq
, sd
);
3662 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
3663 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3664 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3666 sd
->nr_balance_failed
= 0;
3672 * idle_balance is called by schedule() if this_cpu is about to become
3673 * idle. Attempts to pull tasks from other CPUs.
3675 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
3677 struct sched_domain
*sd
;
3678 int pulled_task
= -1;
3679 unsigned long next_balance
= jiffies
+ HZ
;
3682 for_each_domain(this_cpu
, sd
) {
3683 unsigned long interval
;
3685 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3688 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
3689 /* If we've pulled tasks over stop searching: */
3690 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
3693 interval
= msecs_to_jiffies(sd
->balance_interval
);
3694 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3695 next_balance
= sd
->last_balance
+ interval
;
3699 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3701 * We are going idle. next_balance may be set based on
3702 * a busy processor. So reset next_balance.
3704 this_rq
->next_balance
= next_balance
;
3709 * active_load_balance is run by migration threads. It pushes running tasks
3710 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3711 * running on each physical CPU where possible, and avoids physical /
3712 * logical imbalances.
3714 * Called with busiest_rq locked.
3716 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
3718 int target_cpu
= busiest_rq
->push_cpu
;
3719 struct sched_domain
*sd
;
3720 struct rq
*target_rq
;
3722 /* Is there any task to move? */
3723 if (busiest_rq
->nr_running
<= 1)
3726 target_rq
= cpu_rq(target_cpu
);
3729 * This condition is "impossible", if it occurs
3730 * we need to fix it. Originally reported by
3731 * Bjorn Helgaas on a 128-cpu setup.
3733 BUG_ON(busiest_rq
== target_rq
);
3735 /* move a task from busiest_rq to target_rq */
3736 double_lock_balance(busiest_rq
, target_rq
);
3737 update_rq_clock(busiest_rq
);
3738 update_rq_clock(target_rq
);
3740 /* Search for an sd spanning us and the target CPU. */
3741 for_each_domain(target_cpu
, sd
) {
3742 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3743 cpu_isset(busiest_cpu
, sd
->span
))
3748 schedstat_inc(sd
, alb_count
);
3750 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3752 schedstat_inc(sd
, alb_pushed
);
3754 schedstat_inc(sd
, alb_failed
);
3756 double_unlock_balance(busiest_rq
, target_rq
);
3761 atomic_t load_balancer
;
3763 } nohz ____cacheline_aligned
= {
3764 .load_balancer
= ATOMIC_INIT(-1),
3765 .cpu_mask
= CPU_MASK_NONE
,
3769 * This routine will try to nominate the ilb (idle load balancing)
3770 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3771 * load balancing on behalf of all those cpus. If all the cpus in the system
3772 * go into this tickless mode, then there will be no ilb owner (as there is
3773 * no need for one) and all the cpus will sleep till the next wakeup event
3776 * For the ilb owner, tick is not stopped. And this tick will be used
3777 * for idle load balancing. ilb owner will still be part of
3780 * While stopping the tick, this cpu will become the ilb owner if there
3781 * is no other owner. And will be the owner till that cpu becomes busy
3782 * or if all cpus in the system stop their ticks at which point
3783 * there is no need for ilb owner.
3785 * When the ilb owner becomes busy, it nominates another owner, during the
3786 * next busy scheduler_tick()
3788 int select_nohz_load_balancer(int stop_tick
)
3790 int cpu
= smp_processor_id();
3793 cpu_set(cpu
, nohz
.cpu_mask
);
3794 cpu_rq(cpu
)->in_nohz_recently
= 1;
3797 * If we are going offline and still the leader, give up!
3799 if (!cpu_active(cpu
) &&
3800 atomic_read(&nohz
.load_balancer
) == cpu
) {
3801 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3806 /* time for ilb owner also to sleep */
3807 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3808 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3809 atomic_set(&nohz
.load_balancer
, -1);
3813 if (atomic_read(&nohz
.load_balancer
) == -1) {
3814 /* make me the ilb owner */
3815 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3817 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3820 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3823 cpu_clear(cpu
, nohz
.cpu_mask
);
3825 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3826 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3833 static DEFINE_SPINLOCK(balancing
);
3836 * It checks each scheduling domain to see if it is due to be balanced,
3837 * and initiates a balancing operation if so.
3839 * Balancing parameters are set up in arch_init_sched_domains.
3841 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3844 struct rq
*rq
= cpu_rq(cpu
);
3845 unsigned long interval
;
3846 struct sched_domain
*sd
;
3847 /* Earliest time when we have to do rebalance again */
3848 unsigned long next_balance
= jiffies
+ 60*HZ
;
3849 int update_next_balance
= 0;
3853 for_each_domain(cpu
, sd
) {
3854 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3857 interval
= sd
->balance_interval
;
3858 if (idle
!= CPU_IDLE
)
3859 interval
*= sd
->busy_factor
;
3861 /* scale ms to jiffies */
3862 interval
= msecs_to_jiffies(interval
);
3863 if (unlikely(!interval
))
3865 if (interval
> HZ
*NR_CPUS
/10)
3866 interval
= HZ
*NR_CPUS
/10;
3868 need_serialize
= sd
->flags
& SD_SERIALIZE
;
3870 if (need_serialize
) {
3871 if (!spin_trylock(&balancing
))
3875 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3876 if (load_balance(cpu
, rq
, sd
, idle
, &balance
, &tmp
)) {
3878 * We've pulled tasks over so either we're no
3879 * longer idle, or one of our SMT siblings is
3882 idle
= CPU_NOT_IDLE
;
3884 sd
->last_balance
= jiffies
;
3887 spin_unlock(&balancing
);
3889 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3890 next_balance
= sd
->last_balance
+ interval
;
3891 update_next_balance
= 1;
3895 * Stop the load balance at this level. There is another
3896 * CPU in our sched group which is doing load balancing more
3904 * next_balance will be updated only when there is a need.
3905 * When the cpu is attached to null domain for ex, it will not be
3908 if (likely(update_next_balance
))
3909 rq
->next_balance
= next_balance
;
3913 * run_rebalance_domains is triggered when needed from the scheduler tick.
3914 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3915 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3917 static void run_rebalance_domains(struct softirq_action
*h
)
3919 int this_cpu
= smp_processor_id();
3920 struct rq
*this_rq
= cpu_rq(this_cpu
);
3921 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3922 CPU_IDLE
: CPU_NOT_IDLE
;
3924 rebalance_domains(this_cpu
, idle
);
3928 * If this cpu is the owner for idle load balancing, then do the
3929 * balancing on behalf of the other idle cpus whose ticks are
3932 if (this_rq
->idle_at_tick
&&
3933 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3934 cpumask_t cpus
= nohz
.cpu_mask
;
3938 cpu_clear(this_cpu
, cpus
);
3939 for_each_cpu_mask_nr(balance_cpu
, cpus
) {
3941 * If this cpu gets work to do, stop the load balancing
3942 * work being done for other cpus. Next load
3943 * balancing owner will pick it up.
3948 rebalance_domains(balance_cpu
, CPU_IDLE
);
3950 rq
= cpu_rq(balance_cpu
);
3951 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3952 this_rq
->next_balance
= rq
->next_balance
;
3959 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3961 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3962 * idle load balancing owner or decide to stop the periodic load balancing,
3963 * if the whole system is idle.
3965 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3969 * If we were in the nohz mode recently and busy at the current
3970 * scheduler tick, then check if we need to nominate new idle
3973 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3974 rq
->in_nohz_recently
= 0;
3976 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3977 cpu_clear(cpu
, nohz
.cpu_mask
);
3978 atomic_set(&nohz
.load_balancer
, -1);
3981 if (atomic_read(&nohz
.load_balancer
) == -1) {
3983 * simple selection for now: Nominate the
3984 * first cpu in the nohz list to be the next
3987 * TBD: Traverse the sched domains and nominate
3988 * the nearest cpu in the nohz.cpu_mask.
3990 int ilb
= first_cpu(nohz
.cpu_mask
);
3992 if (ilb
< nr_cpu_ids
)
3998 * If this cpu is idle and doing idle load balancing for all the
3999 * cpus with ticks stopped, is it time for that to stop?
4001 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4002 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4008 * If this cpu is idle and the idle load balancing is done by
4009 * someone else, then no need raise the SCHED_SOFTIRQ
4011 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4012 cpu_isset(cpu
, nohz
.cpu_mask
))
4015 if (time_after_eq(jiffies
, rq
->next_balance
))
4016 raise_softirq(SCHED_SOFTIRQ
);
4019 #else /* CONFIG_SMP */
4022 * on UP we do not need to balance between CPUs:
4024 static inline void idle_balance(int cpu
, struct rq
*rq
)
4030 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4032 EXPORT_PER_CPU_SYMBOL(kstat
);
4035 * Return any ns on the sched_clock that have not yet been banked in
4036 * @p in case that task is currently running.
4038 unsigned long long task_delta_exec(struct task_struct
*p
)
4040 unsigned long flags
;
4044 rq
= task_rq_lock(p
, &flags
);
4046 if (task_current(rq
, p
)) {
4049 update_rq_clock(rq
);
4050 delta_exec
= rq
->clock
- p
->se
.exec_start
;
4051 if ((s64
)delta_exec
> 0)
4055 task_rq_unlock(rq
, &flags
);
4061 * Account user cpu time to a process.
4062 * @p: the process that the cpu time gets accounted to
4063 * @cputime: the cpu time spent in user space since the last update
4065 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
4067 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4070 p
->utime
= cputime_add(p
->utime
, cputime
);
4071 account_group_user_time(p
, cputime
);
4073 /* Add user time to cpustat. */
4074 tmp
= cputime_to_cputime64(cputime
);
4075 if (TASK_NICE(p
) > 0)
4076 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
4078 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4079 /* Account for user time used */
4080 acct_update_integrals(p
);
4084 * Account guest cpu time to a process.
4085 * @p: the process that the cpu time gets accounted to
4086 * @cputime: the cpu time spent in virtual machine since the last update
4088 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
)
4091 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4093 tmp
= cputime_to_cputime64(cputime
);
4095 p
->utime
= cputime_add(p
->utime
, cputime
);
4096 account_group_user_time(p
, cputime
);
4097 p
->gtime
= cputime_add(p
->gtime
, cputime
);
4099 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4100 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
4104 * Account scaled user cpu time to a process.
4105 * @p: the process that the cpu time gets accounted to
4106 * @cputime: the cpu time spent in user space since the last update
4108 void account_user_time_scaled(struct task_struct
*p
, cputime_t cputime
)
4110 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime
);
4114 * Account system cpu time to a process.
4115 * @p: the process that the cpu time gets accounted to
4116 * @hardirq_offset: the offset to subtract from hardirq_count()
4117 * @cputime: the cpu time spent in kernel space since the last update
4119 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
4122 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4123 struct rq
*rq
= this_rq();
4126 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
4127 account_guest_time(p
, cputime
);
4131 p
->stime
= cputime_add(p
->stime
, cputime
);
4132 account_group_system_time(p
, cputime
);
4134 /* Add system time to cpustat. */
4135 tmp
= cputime_to_cputime64(cputime
);
4136 if (hardirq_count() - hardirq_offset
)
4137 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
4138 else if (softirq_count())
4139 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
4140 else if (p
!= rq
->idle
)
4141 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
4142 else if (atomic_read(&rq
->nr_iowait
) > 0)
4143 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
4145 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
4146 /* Account for system time used */
4147 acct_update_integrals(p
);
4151 * Account scaled system cpu time to a process.
4152 * @p: the process that the cpu time gets accounted to
4153 * @hardirq_offset: the offset to subtract from hardirq_count()
4154 * @cputime: the cpu time spent in kernel space since the last update
4156 void account_system_time_scaled(struct task_struct
*p
, cputime_t cputime
)
4158 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime
);
4162 * Account for involuntary wait time.
4163 * @p: the process from which the cpu time has been stolen
4164 * @steal: the cpu time spent in involuntary wait
4166 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
4168 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4169 cputime64_t tmp
= cputime_to_cputime64(steal
);
4170 struct rq
*rq
= this_rq();
4172 if (p
== rq
->idle
) {
4173 p
->stime
= cputime_add(p
->stime
, steal
);
4174 account_group_system_time(p
, steal
);
4175 if (atomic_read(&rq
->nr_iowait
) > 0)
4176 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
4178 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
4180 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
4184 * Use precise platform statistics if available:
4186 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4187 cputime_t
task_utime(struct task_struct
*p
)
4192 cputime_t
task_stime(struct task_struct
*p
)
4197 cputime_t
task_utime(struct task_struct
*p
)
4199 clock_t utime
= cputime_to_clock_t(p
->utime
),
4200 total
= utime
+ cputime_to_clock_t(p
->stime
);
4204 * Use CFS's precise accounting:
4206 temp
= (u64
)nsec_to_clock_t(p
->se
.sum_exec_runtime
);
4210 do_div(temp
, total
);
4212 utime
= (clock_t)temp
;
4214 p
->prev_utime
= max(p
->prev_utime
, clock_t_to_cputime(utime
));
4215 return p
->prev_utime
;
4218 cputime_t
task_stime(struct task_struct
*p
)
4223 * Use CFS's precise accounting. (we subtract utime from
4224 * the total, to make sure the total observed by userspace
4225 * grows monotonically - apps rely on that):
4227 stime
= nsec_to_clock_t(p
->se
.sum_exec_runtime
) -
4228 cputime_to_clock_t(task_utime(p
));
4231 p
->prev_stime
= max(p
->prev_stime
, clock_t_to_cputime(stime
));
4233 return p
->prev_stime
;
4237 inline cputime_t
task_gtime(struct task_struct
*p
)
4243 * This function gets called by the timer code, with HZ frequency.
4244 * We call it with interrupts disabled.
4246 * It also gets called by the fork code, when changing the parent's
4249 void scheduler_tick(void)
4251 int cpu
= smp_processor_id();
4252 struct rq
*rq
= cpu_rq(cpu
);
4253 struct task_struct
*curr
= rq
->curr
;
4257 spin_lock(&rq
->lock
);
4258 update_rq_clock(rq
);
4259 update_cpu_load(rq
);
4260 curr
->sched_class
->task_tick(rq
, curr
, 0);
4261 spin_unlock(&rq
->lock
);
4264 rq
->idle_at_tick
= idle_cpu(cpu
);
4265 trigger_load_balance(rq
, cpu
);
4269 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4270 defined(CONFIG_PREEMPT_TRACER))
4272 static inline unsigned long get_parent_ip(unsigned long addr
)
4274 if (in_lock_functions(addr
)) {
4275 addr
= CALLER_ADDR2
;
4276 if (in_lock_functions(addr
))
4277 addr
= CALLER_ADDR3
;
4282 void __kprobes
add_preempt_count(int val
)
4284 #ifdef CONFIG_DEBUG_PREEMPT
4288 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4291 preempt_count() += val
;
4292 #ifdef CONFIG_DEBUG_PREEMPT
4294 * Spinlock count overflowing soon?
4296 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4299 if (preempt_count() == val
)
4300 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4302 EXPORT_SYMBOL(add_preempt_count
);
4304 void __kprobes
sub_preempt_count(int val
)
4306 #ifdef CONFIG_DEBUG_PREEMPT
4310 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
4313 * Is the spinlock portion underflowing?
4315 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4316 !(preempt_count() & PREEMPT_MASK
)))
4320 if (preempt_count() == val
)
4321 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4322 preempt_count() -= val
;
4324 EXPORT_SYMBOL(sub_preempt_count
);
4329 * Print scheduling while atomic bug:
4331 static noinline
void __schedule_bug(struct task_struct
*prev
)
4333 struct pt_regs
*regs
= get_irq_regs();
4335 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4336 prev
->comm
, prev
->pid
, preempt_count());
4338 debug_show_held_locks(prev
);
4340 if (irqs_disabled())
4341 print_irqtrace_events(prev
);
4350 * Various schedule()-time debugging checks and statistics:
4352 static inline void schedule_debug(struct task_struct
*prev
)
4355 * Test if we are atomic. Since do_exit() needs to call into
4356 * schedule() atomically, we ignore that path for now.
4357 * Otherwise, whine if we are scheduling when we should not be.
4359 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
4360 __schedule_bug(prev
);
4362 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4364 schedstat_inc(this_rq(), sched_count
);
4365 #ifdef CONFIG_SCHEDSTATS
4366 if (unlikely(prev
->lock_depth
>= 0)) {
4367 schedstat_inc(this_rq(), bkl_count
);
4368 schedstat_inc(prev
, sched_info
.bkl_count
);
4374 * Pick up the highest-prio task:
4376 static inline struct task_struct
*
4377 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
4379 const struct sched_class
*class;
4380 struct task_struct
*p
;
4383 * Optimization: we know that if all tasks are in
4384 * the fair class we can call that function directly:
4386 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
4387 p
= fair_sched_class
.pick_next_task(rq
);
4392 class = sched_class_highest
;
4394 p
= class->pick_next_task(rq
);
4398 * Will never be NULL as the idle class always
4399 * returns a non-NULL p:
4401 class = class->next
;
4406 * schedule() is the main scheduler function.
4408 asmlinkage
void __sched
schedule(void)
4410 struct task_struct
*prev
, *next
;
4411 unsigned long *switch_count
;
4417 cpu
= smp_processor_id();
4421 switch_count
= &prev
->nivcsw
;
4423 release_kernel_lock(prev
);
4424 need_resched_nonpreemptible
:
4426 schedule_debug(prev
);
4428 if (sched_feat(HRTICK
))
4431 spin_lock_irq(&rq
->lock
);
4432 update_rq_clock(rq
);
4433 clear_tsk_need_resched(prev
);
4435 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
4436 if (unlikely(signal_pending_state(prev
->state
, prev
)))
4437 prev
->state
= TASK_RUNNING
;
4439 deactivate_task(rq
, prev
, 1);
4440 switch_count
= &prev
->nvcsw
;
4444 if (prev
->sched_class
->pre_schedule
)
4445 prev
->sched_class
->pre_schedule(rq
, prev
);
4448 if (unlikely(!rq
->nr_running
))
4449 idle_balance(cpu
, rq
);
4451 prev
->sched_class
->put_prev_task(rq
, prev
);
4452 next
= pick_next_task(rq
, prev
);
4454 if (likely(prev
!= next
)) {
4455 sched_info_switch(prev
, next
);
4461 context_switch(rq
, prev
, next
); /* unlocks the rq */
4463 * the context switch might have flipped the stack from under
4464 * us, hence refresh the local variables.
4466 cpu
= smp_processor_id();
4469 spin_unlock_irq(&rq
->lock
);
4471 if (unlikely(reacquire_kernel_lock(current
) < 0))
4472 goto need_resched_nonpreemptible
;
4474 preempt_enable_no_resched();
4475 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
4478 EXPORT_SYMBOL(schedule
);
4480 #ifdef CONFIG_PREEMPT
4482 * this is the entry point to schedule() from in-kernel preemption
4483 * off of preempt_enable. Kernel preemptions off return from interrupt
4484 * occur there and call schedule directly.
4486 asmlinkage
void __sched
preempt_schedule(void)
4488 struct thread_info
*ti
= current_thread_info();
4491 * If there is a non-zero preempt_count or interrupts are disabled,
4492 * we do not want to preempt the current task. Just return..
4494 if (likely(ti
->preempt_count
|| irqs_disabled()))
4498 add_preempt_count(PREEMPT_ACTIVE
);
4500 sub_preempt_count(PREEMPT_ACTIVE
);
4503 * Check again in case we missed a preemption opportunity
4504 * between schedule and now.
4507 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4509 EXPORT_SYMBOL(preempt_schedule
);
4512 * this is the entry point to schedule() from kernel preemption
4513 * off of irq context.
4514 * Note, that this is called and return with irqs disabled. This will
4515 * protect us against recursive calling from irq.
4517 asmlinkage
void __sched
preempt_schedule_irq(void)
4519 struct thread_info
*ti
= current_thread_info();
4521 /* Catch callers which need to be fixed */
4522 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
4525 add_preempt_count(PREEMPT_ACTIVE
);
4528 local_irq_disable();
4529 sub_preempt_count(PREEMPT_ACTIVE
);
4532 * Check again in case we missed a preemption opportunity
4533 * between schedule and now.
4536 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4539 #endif /* CONFIG_PREEMPT */
4541 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
4544 return try_to_wake_up(curr
->private, mode
, sync
);
4546 EXPORT_SYMBOL(default_wake_function
);
4549 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4550 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4551 * number) then we wake all the non-exclusive tasks and one exclusive task.
4553 * There are circumstances in which we can try to wake a task which has already
4554 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4555 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4557 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4558 int nr_exclusive
, int sync
, void *key
)
4560 wait_queue_t
*curr
, *next
;
4562 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4563 unsigned flags
= curr
->flags
;
4565 if (curr
->func(curr
, mode
, sync
, key
) &&
4566 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4572 * __wake_up - wake up threads blocked on a waitqueue.
4574 * @mode: which threads
4575 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4576 * @key: is directly passed to the wakeup function
4578 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4579 int nr_exclusive
, void *key
)
4581 unsigned long flags
;
4583 spin_lock_irqsave(&q
->lock
, flags
);
4584 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4585 spin_unlock_irqrestore(&q
->lock
, flags
);
4587 EXPORT_SYMBOL(__wake_up
);
4590 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4592 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4594 __wake_up_common(q
, mode
, 1, 0, NULL
);
4598 * __wake_up_sync - wake up threads blocked on a waitqueue.
4600 * @mode: which threads
4601 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4603 * The sync wakeup differs that the waker knows that it will schedule
4604 * away soon, so while the target thread will be woken up, it will not
4605 * be migrated to another CPU - ie. the two threads are 'synchronized'
4606 * with each other. This can prevent needless bouncing between CPUs.
4608 * On UP it can prevent extra preemption.
4611 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4613 unsigned long flags
;
4619 if (unlikely(!nr_exclusive
))
4622 spin_lock_irqsave(&q
->lock
, flags
);
4623 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
4624 spin_unlock_irqrestore(&q
->lock
, flags
);
4626 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4629 * complete: - signals a single thread waiting on this completion
4630 * @x: holds the state of this particular completion
4632 * This will wake up a single thread waiting on this completion. Threads will be
4633 * awakened in the same order in which they were queued.
4635 * See also complete_all(), wait_for_completion() and related routines.
4637 void complete(struct completion
*x
)
4639 unsigned long flags
;
4641 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4643 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4644 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4646 EXPORT_SYMBOL(complete
);
4649 * complete_all: - signals all threads waiting on this completion
4650 * @x: holds the state of this particular completion
4652 * This will wake up all threads waiting on this particular completion event.
4654 void complete_all(struct completion
*x
)
4656 unsigned long flags
;
4658 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4659 x
->done
+= UINT_MAX
/2;
4660 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4661 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4663 EXPORT_SYMBOL(complete_all
);
4665 static inline long __sched
4666 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4669 DECLARE_WAITQUEUE(wait
, current
);
4671 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4672 __add_wait_queue_tail(&x
->wait
, &wait
);
4674 if (signal_pending_state(state
, current
)) {
4675 timeout
= -ERESTARTSYS
;
4678 __set_current_state(state
);
4679 spin_unlock_irq(&x
->wait
.lock
);
4680 timeout
= schedule_timeout(timeout
);
4681 spin_lock_irq(&x
->wait
.lock
);
4682 } while (!x
->done
&& timeout
);
4683 __remove_wait_queue(&x
->wait
, &wait
);
4688 return timeout
?: 1;
4692 wait_for_common(struct completion
*x
, long timeout
, int state
)
4696 spin_lock_irq(&x
->wait
.lock
);
4697 timeout
= do_wait_for_common(x
, timeout
, state
);
4698 spin_unlock_irq(&x
->wait
.lock
);
4703 * wait_for_completion: - waits for completion of a task
4704 * @x: holds the state of this particular completion
4706 * This waits to be signaled for completion of a specific task. It is NOT
4707 * interruptible and there is no timeout.
4709 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4710 * and interrupt capability. Also see complete().
4712 void __sched
wait_for_completion(struct completion
*x
)
4714 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4716 EXPORT_SYMBOL(wait_for_completion
);
4719 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4720 * @x: holds the state of this particular completion
4721 * @timeout: timeout value in jiffies
4723 * This waits for either a completion of a specific task to be signaled or for a
4724 * specified timeout to expire. The timeout is in jiffies. It is not
4727 unsigned long __sched
4728 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4730 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4732 EXPORT_SYMBOL(wait_for_completion_timeout
);
4735 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4736 * @x: holds the state of this particular completion
4738 * This waits for completion of a specific task to be signaled. It is
4741 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4743 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4744 if (t
== -ERESTARTSYS
)
4748 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4751 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4752 * @x: holds the state of this particular completion
4753 * @timeout: timeout value in jiffies
4755 * This waits for either a completion of a specific task to be signaled or for a
4756 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4758 unsigned long __sched
4759 wait_for_completion_interruptible_timeout(struct completion
*x
,
4760 unsigned long timeout
)
4762 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4764 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4767 * wait_for_completion_killable: - waits for completion of a task (killable)
4768 * @x: holds the state of this particular completion
4770 * This waits to be signaled for completion of a specific task. It can be
4771 * interrupted by a kill signal.
4773 int __sched
wait_for_completion_killable(struct completion
*x
)
4775 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4776 if (t
== -ERESTARTSYS
)
4780 EXPORT_SYMBOL(wait_for_completion_killable
);
4783 * try_wait_for_completion - try to decrement a completion without blocking
4784 * @x: completion structure
4786 * Returns: 0 if a decrement cannot be done without blocking
4787 * 1 if a decrement succeeded.
4789 * If a completion is being used as a counting completion,
4790 * attempt to decrement the counter without blocking. This
4791 * enables us to avoid waiting if the resource the completion
4792 * is protecting is not available.
4794 bool try_wait_for_completion(struct completion
*x
)
4798 spin_lock_irq(&x
->wait
.lock
);
4803 spin_unlock_irq(&x
->wait
.lock
);
4806 EXPORT_SYMBOL(try_wait_for_completion
);
4809 * completion_done - Test to see if a completion has any waiters
4810 * @x: completion structure
4812 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4813 * 1 if there are no waiters.
4816 bool completion_done(struct completion
*x
)
4820 spin_lock_irq(&x
->wait
.lock
);
4823 spin_unlock_irq(&x
->wait
.lock
);
4826 EXPORT_SYMBOL(completion_done
);
4829 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4831 unsigned long flags
;
4834 init_waitqueue_entry(&wait
, current
);
4836 __set_current_state(state
);
4838 spin_lock_irqsave(&q
->lock
, flags
);
4839 __add_wait_queue(q
, &wait
);
4840 spin_unlock(&q
->lock
);
4841 timeout
= schedule_timeout(timeout
);
4842 spin_lock_irq(&q
->lock
);
4843 __remove_wait_queue(q
, &wait
);
4844 spin_unlock_irqrestore(&q
->lock
, flags
);
4849 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4851 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4853 EXPORT_SYMBOL(interruptible_sleep_on
);
4856 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4858 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4860 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4862 void __sched
sleep_on(wait_queue_head_t
*q
)
4864 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4866 EXPORT_SYMBOL(sleep_on
);
4868 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4870 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4872 EXPORT_SYMBOL(sleep_on_timeout
);
4874 #ifdef CONFIG_RT_MUTEXES
4877 * rt_mutex_setprio - set the current priority of a task
4879 * @prio: prio value (kernel-internal form)
4881 * This function changes the 'effective' priority of a task. It does
4882 * not touch ->normal_prio like __setscheduler().
4884 * Used by the rt_mutex code to implement priority inheritance logic.
4886 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4888 unsigned long flags
;
4889 int oldprio
, on_rq
, running
;
4891 const struct sched_class
*prev_class
= p
->sched_class
;
4893 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4895 rq
= task_rq_lock(p
, &flags
);
4896 update_rq_clock(rq
);
4899 on_rq
= p
->se
.on_rq
;
4900 running
= task_current(rq
, p
);
4902 dequeue_task(rq
, p
, 0);
4904 p
->sched_class
->put_prev_task(rq
, p
);
4907 p
->sched_class
= &rt_sched_class
;
4909 p
->sched_class
= &fair_sched_class
;
4914 p
->sched_class
->set_curr_task(rq
);
4916 enqueue_task(rq
, p
, 0);
4918 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4920 task_rq_unlock(rq
, &flags
);
4925 void set_user_nice(struct task_struct
*p
, long nice
)
4927 int old_prio
, delta
, on_rq
;
4928 unsigned long flags
;
4931 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4934 * We have to be careful, if called from sys_setpriority(),
4935 * the task might be in the middle of scheduling on another CPU.
4937 rq
= task_rq_lock(p
, &flags
);
4938 update_rq_clock(rq
);
4940 * The RT priorities are set via sched_setscheduler(), but we still
4941 * allow the 'normal' nice value to be set - but as expected
4942 * it wont have any effect on scheduling until the task is
4943 * SCHED_FIFO/SCHED_RR:
4945 if (task_has_rt_policy(p
)) {
4946 p
->static_prio
= NICE_TO_PRIO(nice
);
4949 on_rq
= p
->se
.on_rq
;
4951 dequeue_task(rq
, p
, 0);
4953 p
->static_prio
= NICE_TO_PRIO(nice
);
4956 p
->prio
= effective_prio(p
);
4957 delta
= p
->prio
- old_prio
;
4960 enqueue_task(rq
, p
, 0);
4962 * If the task increased its priority or is running and
4963 * lowered its priority, then reschedule its CPU:
4965 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4966 resched_task(rq
->curr
);
4969 task_rq_unlock(rq
, &flags
);
4971 EXPORT_SYMBOL(set_user_nice
);
4974 * can_nice - check if a task can reduce its nice value
4978 int can_nice(const struct task_struct
*p
, const int nice
)
4980 /* convert nice value [19,-20] to rlimit style value [1,40] */
4981 int nice_rlim
= 20 - nice
;
4983 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4984 capable(CAP_SYS_NICE
));
4987 #ifdef __ARCH_WANT_SYS_NICE
4990 * sys_nice - change the priority of the current process.
4991 * @increment: priority increment
4993 * sys_setpriority is a more generic, but much slower function that
4994 * does similar things.
4996 asmlinkage
long sys_nice(int increment
)
5001 * Setpriority might change our priority at the same moment.
5002 * We don't have to worry. Conceptually one call occurs first
5003 * and we have a single winner.
5005 if (increment
< -40)
5010 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
5016 if (increment
< 0 && !can_nice(current
, nice
))
5019 retval
= security_task_setnice(current
, nice
);
5023 set_user_nice(current
, nice
);
5030 * task_prio - return the priority value of a given task.
5031 * @p: the task in question.
5033 * This is the priority value as seen by users in /proc.
5034 * RT tasks are offset by -200. Normal tasks are centered
5035 * around 0, value goes from -16 to +15.
5037 int task_prio(const struct task_struct
*p
)
5039 return p
->prio
- MAX_RT_PRIO
;
5043 * task_nice - return the nice value of a given task.
5044 * @p: the task in question.
5046 int task_nice(const struct task_struct
*p
)
5048 return TASK_NICE(p
);
5050 EXPORT_SYMBOL(task_nice
);
5053 * idle_cpu - is a given cpu idle currently?
5054 * @cpu: the processor in question.
5056 int idle_cpu(int cpu
)
5058 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
5062 * idle_task - return the idle task for a given cpu.
5063 * @cpu: the processor in question.
5065 struct task_struct
*idle_task(int cpu
)
5067 return cpu_rq(cpu
)->idle
;
5071 * find_process_by_pid - find a process with a matching PID value.
5072 * @pid: the pid in question.
5074 static struct task_struct
*find_process_by_pid(pid_t pid
)
5076 return pid
? find_task_by_vpid(pid
) : current
;
5079 /* Actually do priority change: must hold rq lock. */
5081 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
5083 BUG_ON(p
->se
.on_rq
);
5086 switch (p
->policy
) {
5090 p
->sched_class
= &fair_sched_class
;
5094 p
->sched_class
= &rt_sched_class
;
5098 p
->rt_priority
= prio
;
5099 p
->normal_prio
= normal_prio(p
);
5100 /* we are holding p->pi_lock already */
5101 p
->prio
= rt_mutex_getprio(p
);
5105 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
5106 struct sched_param
*param
, bool user
)
5108 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
5109 unsigned long flags
;
5110 const struct sched_class
*prev_class
= p
->sched_class
;
5113 /* may grab non-irq protected spin_locks */
5114 BUG_ON(in_interrupt());
5116 /* double check policy once rq lock held */
5118 policy
= oldpolicy
= p
->policy
;
5119 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
5120 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
5121 policy
!= SCHED_IDLE
)
5124 * Valid priorities for SCHED_FIFO and SCHED_RR are
5125 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5126 * SCHED_BATCH and SCHED_IDLE is 0.
5128 if (param
->sched_priority
< 0 ||
5129 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
5130 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
5132 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
5136 * Allow unprivileged RT tasks to decrease priority:
5138 if (user
&& !capable(CAP_SYS_NICE
)) {
5139 if (rt_policy(policy
)) {
5140 unsigned long rlim_rtprio
;
5142 if (!lock_task_sighand(p
, &flags
))
5144 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
5145 unlock_task_sighand(p
, &flags
);
5147 /* can't set/change the rt policy */
5148 if (policy
!= p
->policy
&& !rlim_rtprio
)
5151 /* can't increase priority */
5152 if (param
->sched_priority
> p
->rt_priority
&&
5153 param
->sched_priority
> rlim_rtprio
)
5157 * Like positive nice levels, dont allow tasks to
5158 * move out of SCHED_IDLE either:
5160 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
5163 /* can't change other user's priorities */
5164 if ((current
->euid
!= p
->euid
) &&
5165 (current
->euid
!= p
->uid
))
5170 #ifdef CONFIG_RT_GROUP_SCHED
5172 * Do not allow realtime tasks into groups that have no runtime
5175 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
5176 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
5180 retval
= security_task_setscheduler(p
, policy
, param
);
5186 * make sure no PI-waiters arrive (or leave) while we are
5187 * changing the priority of the task:
5189 spin_lock_irqsave(&p
->pi_lock
, flags
);
5191 * To be able to change p->policy safely, the apropriate
5192 * runqueue lock must be held.
5194 rq
= __task_rq_lock(p
);
5195 /* recheck policy now with rq lock held */
5196 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
5197 policy
= oldpolicy
= -1;
5198 __task_rq_unlock(rq
);
5199 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5202 update_rq_clock(rq
);
5203 on_rq
= p
->se
.on_rq
;
5204 running
= task_current(rq
, p
);
5206 deactivate_task(rq
, p
, 0);
5208 p
->sched_class
->put_prev_task(rq
, p
);
5211 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
5214 p
->sched_class
->set_curr_task(rq
);
5216 activate_task(rq
, p
, 0);
5218 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5220 __task_rq_unlock(rq
);
5221 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5223 rt_mutex_adjust_pi(p
);
5229 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5230 * @p: the task in question.
5231 * @policy: new policy.
5232 * @param: structure containing the new RT priority.
5234 * NOTE that the task may be already dead.
5236 int sched_setscheduler(struct task_struct
*p
, int policy
,
5237 struct sched_param
*param
)
5239 return __sched_setscheduler(p
, policy
, param
, true);
5241 EXPORT_SYMBOL_GPL(sched_setscheduler
);
5244 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5245 * @p: the task in question.
5246 * @policy: new policy.
5247 * @param: structure containing the new RT priority.
5249 * Just like sched_setscheduler, only don't bother checking if the
5250 * current context has permission. For example, this is needed in
5251 * stop_machine(): we create temporary high priority worker threads,
5252 * but our caller might not have that capability.
5254 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
5255 struct sched_param
*param
)
5257 return __sched_setscheduler(p
, policy
, param
, false);
5261 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5263 struct sched_param lparam
;
5264 struct task_struct
*p
;
5267 if (!param
|| pid
< 0)
5269 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
5274 p
= find_process_by_pid(pid
);
5276 retval
= sched_setscheduler(p
, policy
, &lparam
);
5283 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5284 * @pid: the pid in question.
5285 * @policy: new policy.
5286 * @param: structure containing the new RT priority.
5289 sys_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5291 /* negative values for policy are not valid */
5295 return do_sched_setscheduler(pid
, policy
, param
);
5299 * sys_sched_setparam - set/change the RT priority of a thread
5300 * @pid: the pid in question.
5301 * @param: structure containing the new RT priority.
5303 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
5305 return do_sched_setscheduler(pid
, -1, param
);
5309 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5310 * @pid: the pid in question.
5312 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
5314 struct task_struct
*p
;
5321 read_lock(&tasklist_lock
);
5322 p
= find_process_by_pid(pid
);
5324 retval
= security_task_getscheduler(p
);
5328 read_unlock(&tasklist_lock
);
5333 * sys_sched_getscheduler - get the RT priority of a thread
5334 * @pid: the pid in question.
5335 * @param: structure containing the RT priority.
5337 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
5339 struct sched_param lp
;
5340 struct task_struct
*p
;
5343 if (!param
|| pid
< 0)
5346 read_lock(&tasklist_lock
);
5347 p
= find_process_by_pid(pid
);
5352 retval
= security_task_getscheduler(p
);
5356 lp
.sched_priority
= p
->rt_priority
;
5357 read_unlock(&tasklist_lock
);
5360 * This one might sleep, we cannot do it with a spinlock held ...
5362 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
5367 read_unlock(&tasklist_lock
);
5371 long sched_setaffinity(pid_t pid
, const cpumask_t
*in_mask
)
5373 cpumask_t cpus_allowed
;
5374 cpumask_t new_mask
= *in_mask
;
5375 struct task_struct
*p
;
5379 read_lock(&tasklist_lock
);
5381 p
= find_process_by_pid(pid
);
5383 read_unlock(&tasklist_lock
);
5389 * It is not safe to call set_cpus_allowed with the
5390 * tasklist_lock held. We will bump the task_struct's
5391 * usage count and then drop tasklist_lock.
5394 read_unlock(&tasklist_lock
);
5397 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
5398 !capable(CAP_SYS_NICE
))
5401 retval
= security_task_setscheduler(p
, 0, NULL
);
5405 cpuset_cpus_allowed(p
, &cpus_allowed
);
5406 cpus_and(new_mask
, new_mask
, cpus_allowed
);
5408 retval
= set_cpus_allowed_ptr(p
, &new_mask
);
5411 cpuset_cpus_allowed(p
, &cpus_allowed
);
5412 if (!cpus_subset(new_mask
, cpus_allowed
)) {
5414 * We must have raced with a concurrent cpuset
5415 * update. Just reset the cpus_allowed to the
5416 * cpuset's cpus_allowed
5418 new_mask
= cpus_allowed
;
5428 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5429 cpumask_t
*new_mask
)
5431 if (len
< sizeof(cpumask_t
)) {
5432 memset(new_mask
, 0, sizeof(cpumask_t
));
5433 } else if (len
> sizeof(cpumask_t
)) {
5434 len
= sizeof(cpumask_t
);
5436 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
5440 * sys_sched_setaffinity - set the cpu affinity of a process
5441 * @pid: pid of the process
5442 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5443 * @user_mask_ptr: user-space pointer to the new cpu mask
5445 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
5446 unsigned long __user
*user_mask_ptr
)
5451 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
5455 return sched_setaffinity(pid
, &new_mask
);
5458 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
5460 struct task_struct
*p
;
5464 read_lock(&tasklist_lock
);
5467 p
= find_process_by_pid(pid
);
5471 retval
= security_task_getscheduler(p
);
5475 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
5478 read_unlock(&tasklist_lock
);
5485 * sys_sched_getaffinity - get the cpu affinity of a process
5486 * @pid: pid of the process
5487 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5488 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5490 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
5491 unsigned long __user
*user_mask_ptr
)
5496 if (len
< sizeof(cpumask_t
))
5499 ret
= sched_getaffinity(pid
, &mask
);
5503 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
5506 return sizeof(cpumask_t
);
5510 * sys_sched_yield - yield the current processor to other threads.
5512 * This function yields the current CPU to other tasks. If there are no
5513 * other threads running on this CPU then this function will return.
5515 asmlinkage
long sys_sched_yield(void)
5517 struct rq
*rq
= this_rq_lock();
5519 schedstat_inc(rq
, yld_count
);
5520 current
->sched_class
->yield_task(rq
);
5523 * Since we are going to call schedule() anyway, there's
5524 * no need to preempt or enable interrupts:
5526 __release(rq
->lock
);
5527 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5528 _raw_spin_unlock(&rq
->lock
);
5529 preempt_enable_no_resched();
5536 static void __cond_resched(void)
5538 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5539 __might_sleep(__FILE__
, __LINE__
);
5542 * The BKS might be reacquired before we have dropped
5543 * PREEMPT_ACTIVE, which could trigger a second
5544 * cond_resched() call.
5547 add_preempt_count(PREEMPT_ACTIVE
);
5549 sub_preempt_count(PREEMPT_ACTIVE
);
5550 } while (need_resched());
5553 int __sched
_cond_resched(void)
5555 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
5556 system_state
== SYSTEM_RUNNING
) {
5562 EXPORT_SYMBOL(_cond_resched
);
5565 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5566 * call schedule, and on return reacquire the lock.
5568 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5569 * operations here to prevent schedule() from being called twice (once via
5570 * spin_unlock(), once by hand).
5572 int cond_resched_lock(spinlock_t
*lock
)
5574 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
5577 if (spin_needbreak(lock
) || resched
) {
5579 if (resched
&& need_resched())
5588 EXPORT_SYMBOL(cond_resched_lock
);
5590 int __sched
cond_resched_softirq(void)
5592 BUG_ON(!in_softirq());
5594 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
5602 EXPORT_SYMBOL(cond_resched_softirq
);
5605 * yield - yield the current processor to other threads.
5607 * This is a shortcut for kernel-space yielding - it marks the
5608 * thread runnable and calls sys_sched_yield().
5610 void __sched
yield(void)
5612 set_current_state(TASK_RUNNING
);
5615 EXPORT_SYMBOL(yield
);
5618 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5619 * that process accounting knows that this is a task in IO wait state.
5621 * But don't do that if it is a deliberate, throttling IO wait (this task
5622 * has set its backing_dev_info: the queue against which it should throttle)
5624 void __sched
io_schedule(void)
5626 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5628 delayacct_blkio_start();
5629 atomic_inc(&rq
->nr_iowait
);
5631 atomic_dec(&rq
->nr_iowait
);
5632 delayacct_blkio_end();
5634 EXPORT_SYMBOL(io_schedule
);
5636 long __sched
io_schedule_timeout(long timeout
)
5638 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5641 delayacct_blkio_start();
5642 atomic_inc(&rq
->nr_iowait
);
5643 ret
= schedule_timeout(timeout
);
5644 atomic_dec(&rq
->nr_iowait
);
5645 delayacct_blkio_end();
5650 * sys_sched_get_priority_max - return maximum RT priority.
5651 * @policy: scheduling class.
5653 * this syscall returns the maximum rt_priority that can be used
5654 * by a given scheduling class.
5656 asmlinkage
long sys_sched_get_priority_max(int policy
)
5663 ret
= MAX_USER_RT_PRIO
-1;
5675 * sys_sched_get_priority_min - return minimum RT priority.
5676 * @policy: scheduling class.
5678 * this syscall returns the minimum rt_priority that can be used
5679 * by a given scheduling class.
5681 asmlinkage
long sys_sched_get_priority_min(int policy
)
5699 * sys_sched_rr_get_interval - return the default timeslice of a process.
5700 * @pid: pid of the process.
5701 * @interval: userspace pointer to the timeslice value.
5703 * this syscall writes the default timeslice value of a given process
5704 * into the user-space timespec buffer. A value of '0' means infinity.
5707 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
5709 struct task_struct
*p
;
5710 unsigned int time_slice
;
5718 read_lock(&tasklist_lock
);
5719 p
= find_process_by_pid(pid
);
5723 retval
= security_task_getscheduler(p
);
5728 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5729 * tasks that are on an otherwise idle runqueue:
5732 if (p
->policy
== SCHED_RR
) {
5733 time_slice
= DEF_TIMESLICE
;
5734 } else if (p
->policy
!= SCHED_FIFO
) {
5735 struct sched_entity
*se
= &p
->se
;
5736 unsigned long flags
;
5739 rq
= task_rq_lock(p
, &flags
);
5740 if (rq
->cfs
.load
.weight
)
5741 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
5742 task_rq_unlock(rq
, &flags
);
5744 read_unlock(&tasklist_lock
);
5745 jiffies_to_timespec(time_slice
, &t
);
5746 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5750 read_unlock(&tasklist_lock
);
5754 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5756 void sched_show_task(struct task_struct
*p
)
5758 unsigned long free
= 0;
5761 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5762 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5763 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5764 #if BITS_PER_LONG == 32
5765 if (state
== TASK_RUNNING
)
5766 printk(KERN_CONT
" running ");
5768 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5770 if (state
== TASK_RUNNING
)
5771 printk(KERN_CONT
" running task ");
5773 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5775 #ifdef CONFIG_DEBUG_STACK_USAGE
5777 unsigned long *n
= end_of_stack(p
);
5780 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
5783 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
5784 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
5786 show_stack(p
, NULL
);
5789 void show_state_filter(unsigned long state_filter
)
5791 struct task_struct
*g
, *p
;
5793 #if BITS_PER_LONG == 32
5795 " task PC stack pid father\n");
5798 " task PC stack pid father\n");
5800 read_lock(&tasklist_lock
);
5801 do_each_thread(g
, p
) {
5803 * reset the NMI-timeout, listing all files on a slow
5804 * console might take alot of time:
5806 touch_nmi_watchdog();
5807 if (!state_filter
|| (p
->state
& state_filter
))
5809 } while_each_thread(g
, p
);
5811 touch_all_softlockup_watchdogs();
5813 #ifdef CONFIG_SCHED_DEBUG
5814 sysrq_sched_debug_show();
5816 read_unlock(&tasklist_lock
);
5818 * Only show locks if all tasks are dumped:
5820 if (state_filter
== -1)
5821 debug_show_all_locks();
5824 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5826 idle
->sched_class
= &idle_sched_class
;
5830 * init_idle - set up an idle thread for a given CPU
5831 * @idle: task in question
5832 * @cpu: cpu the idle task belongs to
5834 * NOTE: this function does not set the idle thread's NEED_RESCHED
5835 * flag, to make booting more robust.
5837 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5839 struct rq
*rq
= cpu_rq(cpu
);
5840 unsigned long flags
;
5842 spin_lock_irqsave(&rq
->lock
, flags
);
5845 idle
->se
.exec_start
= sched_clock();
5847 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
5848 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
5849 __set_task_cpu(idle
, cpu
);
5851 rq
->curr
= rq
->idle
= idle
;
5852 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5855 spin_unlock_irqrestore(&rq
->lock
, flags
);
5857 /* Set the preempt count _outside_ the spinlocks! */
5858 #if defined(CONFIG_PREEMPT)
5859 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5861 task_thread_info(idle
)->preempt_count
= 0;
5864 * The idle tasks have their own, simple scheduling class:
5866 idle
->sched_class
= &idle_sched_class
;
5870 * In a system that switches off the HZ timer nohz_cpu_mask
5871 * indicates which cpus entered this state. This is used
5872 * in the rcu update to wait only for active cpus. For system
5873 * which do not switch off the HZ timer nohz_cpu_mask should
5874 * always be CPU_MASK_NONE.
5876 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
5879 * Increase the granularity value when there are more CPUs,
5880 * because with more CPUs the 'effective latency' as visible
5881 * to users decreases. But the relationship is not linear,
5882 * so pick a second-best guess by going with the log2 of the
5885 * This idea comes from the SD scheduler of Con Kolivas:
5887 static inline void sched_init_granularity(void)
5889 unsigned int factor
= 1 + ilog2(num_online_cpus());
5890 const unsigned long limit
= 200000000;
5892 sysctl_sched_min_granularity
*= factor
;
5893 if (sysctl_sched_min_granularity
> limit
)
5894 sysctl_sched_min_granularity
= limit
;
5896 sysctl_sched_latency
*= factor
;
5897 if (sysctl_sched_latency
> limit
)
5898 sysctl_sched_latency
= limit
;
5900 sysctl_sched_wakeup_granularity
*= factor
;
5902 sysctl_sched_shares_ratelimit
*= factor
;
5907 * This is how migration works:
5909 * 1) we queue a struct migration_req structure in the source CPU's
5910 * runqueue and wake up that CPU's migration thread.
5911 * 2) we down() the locked semaphore => thread blocks.
5912 * 3) migration thread wakes up (implicitly it forces the migrated
5913 * thread off the CPU)
5914 * 4) it gets the migration request and checks whether the migrated
5915 * task is still in the wrong runqueue.
5916 * 5) if it's in the wrong runqueue then the migration thread removes
5917 * it and puts it into the right queue.
5918 * 6) migration thread up()s the semaphore.
5919 * 7) we wake up and the migration is done.
5923 * Change a given task's CPU affinity. Migrate the thread to a
5924 * proper CPU and schedule it away if the CPU it's executing on
5925 * is removed from the allowed bitmask.
5927 * NOTE: the caller must have a valid reference to the task, the
5928 * task must not exit() & deallocate itself prematurely. The
5929 * call is not atomic; no spinlocks may be held.
5931 int set_cpus_allowed_ptr(struct task_struct
*p
, const cpumask_t
*new_mask
)
5933 struct migration_req req
;
5934 unsigned long flags
;
5938 rq
= task_rq_lock(p
, &flags
);
5939 if (!cpus_intersects(*new_mask
, cpu_online_map
)) {
5944 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
5945 !cpus_equal(p
->cpus_allowed
, *new_mask
))) {
5950 if (p
->sched_class
->set_cpus_allowed
)
5951 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
5953 p
->cpus_allowed
= *new_mask
;
5954 p
->rt
.nr_cpus_allowed
= cpus_weight(*new_mask
);
5957 /* Can the task run on the task's current CPU? If so, we're done */
5958 if (cpu_isset(task_cpu(p
), *new_mask
))
5961 if (migrate_task(p
, any_online_cpu(*new_mask
), &req
)) {
5962 /* Need help from migration thread: drop lock and wait. */
5963 task_rq_unlock(rq
, &flags
);
5964 wake_up_process(rq
->migration_thread
);
5965 wait_for_completion(&req
.done
);
5966 tlb_migrate_finish(p
->mm
);
5970 task_rq_unlock(rq
, &flags
);
5974 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5977 * Move (not current) task off this cpu, onto dest cpu. We're doing
5978 * this because either it can't run here any more (set_cpus_allowed()
5979 * away from this CPU, or CPU going down), or because we're
5980 * attempting to rebalance this task on exec (sched_exec).
5982 * So we race with normal scheduler movements, but that's OK, as long
5983 * as the task is no longer on this CPU.
5985 * Returns non-zero if task was successfully migrated.
5987 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5989 struct rq
*rq_dest
, *rq_src
;
5992 if (unlikely(!cpu_active(dest_cpu
)))
5995 rq_src
= cpu_rq(src_cpu
);
5996 rq_dest
= cpu_rq(dest_cpu
);
5998 double_rq_lock(rq_src
, rq_dest
);
5999 /* Already moved. */
6000 if (task_cpu(p
) != src_cpu
)
6002 /* Affinity changed (again). */
6003 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
6006 on_rq
= p
->se
.on_rq
;
6008 deactivate_task(rq_src
, p
, 0);
6010 set_task_cpu(p
, dest_cpu
);
6012 activate_task(rq_dest
, p
, 0);
6013 check_preempt_curr(rq_dest
, p
, 0);
6018 double_rq_unlock(rq_src
, rq_dest
);
6023 * migration_thread - this is a highprio system thread that performs
6024 * thread migration by bumping thread off CPU then 'pushing' onto
6027 static int migration_thread(void *data
)
6029 int cpu
= (long)data
;
6033 BUG_ON(rq
->migration_thread
!= current
);
6035 set_current_state(TASK_INTERRUPTIBLE
);
6036 while (!kthread_should_stop()) {
6037 struct migration_req
*req
;
6038 struct list_head
*head
;
6040 spin_lock_irq(&rq
->lock
);
6042 if (cpu_is_offline(cpu
)) {
6043 spin_unlock_irq(&rq
->lock
);
6047 if (rq
->active_balance
) {
6048 active_load_balance(rq
, cpu
);
6049 rq
->active_balance
= 0;
6052 head
= &rq
->migration_queue
;
6054 if (list_empty(head
)) {
6055 spin_unlock_irq(&rq
->lock
);
6057 set_current_state(TASK_INTERRUPTIBLE
);
6060 req
= list_entry(head
->next
, struct migration_req
, list
);
6061 list_del_init(head
->next
);
6063 spin_unlock(&rq
->lock
);
6064 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
6067 complete(&req
->done
);
6069 __set_current_state(TASK_RUNNING
);
6073 /* Wait for kthread_stop */
6074 set_current_state(TASK_INTERRUPTIBLE
);
6075 while (!kthread_should_stop()) {
6077 set_current_state(TASK_INTERRUPTIBLE
);
6079 __set_current_state(TASK_RUNNING
);
6083 #ifdef CONFIG_HOTPLUG_CPU
6085 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6089 local_irq_disable();
6090 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
6096 * Figure out where task on dead CPU should go, use force if necessary.
6098 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
6100 unsigned long flags
;
6107 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
6108 cpus_and(mask
, mask
, p
->cpus_allowed
);
6109 dest_cpu
= any_online_cpu(mask
);
6111 /* On any allowed CPU? */
6112 if (dest_cpu
>= nr_cpu_ids
)
6113 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
6115 /* No more Mr. Nice Guy. */
6116 if (dest_cpu
>= nr_cpu_ids
) {
6117 cpumask_t cpus_allowed
;
6119 cpuset_cpus_allowed_locked(p
, &cpus_allowed
);
6121 * Try to stay on the same cpuset, where the
6122 * current cpuset may be a subset of all cpus.
6123 * The cpuset_cpus_allowed_locked() variant of
6124 * cpuset_cpus_allowed() will not block. It must be
6125 * called within calls to cpuset_lock/cpuset_unlock.
6127 rq
= task_rq_lock(p
, &flags
);
6128 p
->cpus_allowed
= cpus_allowed
;
6129 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
6130 task_rq_unlock(rq
, &flags
);
6133 * Don't tell them about moving exiting tasks or
6134 * kernel threads (both mm NULL), since they never
6137 if (p
->mm
&& printk_ratelimit()) {
6138 printk(KERN_INFO
"process %d (%s) no "
6139 "longer affine to cpu%d\n",
6140 task_pid_nr(p
), p
->comm
, dead_cpu
);
6143 } while (!__migrate_task_irq(p
, dead_cpu
, dest_cpu
));
6147 * While a dead CPU has no uninterruptible tasks queued at this point,
6148 * it might still have a nonzero ->nr_uninterruptible counter, because
6149 * for performance reasons the counter is not stricly tracking tasks to
6150 * their home CPUs. So we just add the counter to another CPU's counter,
6151 * to keep the global sum constant after CPU-down:
6153 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
6155 struct rq
*rq_dest
= cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR
));
6156 unsigned long flags
;
6158 local_irq_save(flags
);
6159 double_rq_lock(rq_src
, rq_dest
);
6160 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
6161 rq_src
->nr_uninterruptible
= 0;
6162 double_rq_unlock(rq_src
, rq_dest
);
6163 local_irq_restore(flags
);
6166 /* Run through task list and migrate tasks from the dead cpu. */
6167 static void migrate_live_tasks(int src_cpu
)
6169 struct task_struct
*p
, *t
;
6171 read_lock(&tasklist_lock
);
6173 do_each_thread(t
, p
) {
6177 if (task_cpu(p
) == src_cpu
)
6178 move_task_off_dead_cpu(src_cpu
, p
);
6179 } while_each_thread(t
, p
);
6181 read_unlock(&tasklist_lock
);
6185 * Schedules idle task to be the next runnable task on current CPU.
6186 * It does so by boosting its priority to highest possible.
6187 * Used by CPU offline code.
6189 void sched_idle_next(void)
6191 int this_cpu
= smp_processor_id();
6192 struct rq
*rq
= cpu_rq(this_cpu
);
6193 struct task_struct
*p
= rq
->idle
;
6194 unsigned long flags
;
6196 /* cpu has to be offline */
6197 BUG_ON(cpu_online(this_cpu
));
6200 * Strictly not necessary since rest of the CPUs are stopped by now
6201 * and interrupts disabled on the current cpu.
6203 spin_lock_irqsave(&rq
->lock
, flags
);
6205 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6207 update_rq_clock(rq
);
6208 activate_task(rq
, p
, 0);
6210 spin_unlock_irqrestore(&rq
->lock
, flags
);
6214 * Ensures that the idle task is using init_mm right before its cpu goes
6217 void idle_task_exit(void)
6219 struct mm_struct
*mm
= current
->active_mm
;
6221 BUG_ON(cpu_online(smp_processor_id()));
6224 switch_mm(mm
, &init_mm
, current
);
6228 /* called under rq->lock with disabled interrupts */
6229 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
6231 struct rq
*rq
= cpu_rq(dead_cpu
);
6233 /* Must be exiting, otherwise would be on tasklist. */
6234 BUG_ON(!p
->exit_state
);
6236 /* Cannot have done final schedule yet: would have vanished. */
6237 BUG_ON(p
->state
== TASK_DEAD
);
6242 * Drop lock around migration; if someone else moves it,
6243 * that's OK. No task can be added to this CPU, so iteration is
6246 spin_unlock_irq(&rq
->lock
);
6247 move_task_off_dead_cpu(dead_cpu
, p
);
6248 spin_lock_irq(&rq
->lock
);
6253 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6254 static void migrate_dead_tasks(unsigned int dead_cpu
)
6256 struct rq
*rq
= cpu_rq(dead_cpu
);
6257 struct task_struct
*next
;
6260 if (!rq
->nr_running
)
6262 update_rq_clock(rq
);
6263 next
= pick_next_task(rq
, rq
->curr
);
6266 next
->sched_class
->put_prev_task(rq
, next
);
6267 migrate_dead(dead_cpu
, next
);
6271 #endif /* CONFIG_HOTPLUG_CPU */
6273 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6275 static struct ctl_table sd_ctl_dir
[] = {
6277 .procname
= "sched_domain",
6283 static struct ctl_table sd_ctl_root
[] = {
6285 .ctl_name
= CTL_KERN
,
6286 .procname
= "kernel",
6288 .child
= sd_ctl_dir
,
6293 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
6295 struct ctl_table
*entry
=
6296 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
6301 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
6303 struct ctl_table
*entry
;
6306 * In the intermediate directories, both the child directory and
6307 * procname are dynamically allocated and could fail but the mode
6308 * will always be set. In the lowest directory the names are
6309 * static strings and all have proc handlers.
6311 for (entry
= *tablep
; entry
->mode
; entry
++) {
6313 sd_free_ctl_entry(&entry
->child
);
6314 if (entry
->proc_handler
== NULL
)
6315 kfree(entry
->procname
);
6323 set_table_entry(struct ctl_table
*entry
,
6324 const char *procname
, void *data
, int maxlen
,
6325 mode_t mode
, proc_handler
*proc_handler
)
6327 entry
->procname
= procname
;
6329 entry
->maxlen
= maxlen
;
6331 entry
->proc_handler
= proc_handler
;
6334 static struct ctl_table
*
6335 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
6337 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
6342 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
6343 sizeof(long), 0644, proc_doulongvec_minmax
);
6344 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
6345 sizeof(long), 0644, proc_doulongvec_minmax
);
6346 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
6347 sizeof(int), 0644, proc_dointvec_minmax
);
6348 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
6349 sizeof(int), 0644, proc_dointvec_minmax
);
6350 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
6351 sizeof(int), 0644, proc_dointvec_minmax
);
6352 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
6353 sizeof(int), 0644, proc_dointvec_minmax
);
6354 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
6355 sizeof(int), 0644, proc_dointvec_minmax
);
6356 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
6357 sizeof(int), 0644, proc_dointvec_minmax
);
6358 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
6359 sizeof(int), 0644, proc_dointvec_minmax
);
6360 set_table_entry(&table
[9], "cache_nice_tries",
6361 &sd
->cache_nice_tries
,
6362 sizeof(int), 0644, proc_dointvec_minmax
);
6363 set_table_entry(&table
[10], "flags", &sd
->flags
,
6364 sizeof(int), 0644, proc_dointvec_minmax
);
6365 set_table_entry(&table
[11], "name", sd
->name
,
6366 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
6367 /* &table[12] is terminator */
6372 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
6374 struct ctl_table
*entry
, *table
;
6375 struct sched_domain
*sd
;
6376 int domain_num
= 0, i
;
6379 for_each_domain(cpu
, sd
)
6381 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
6386 for_each_domain(cpu
, sd
) {
6387 snprintf(buf
, 32, "domain%d", i
);
6388 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6390 entry
->child
= sd_alloc_ctl_domain_table(sd
);
6397 static struct ctl_table_header
*sd_sysctl_header
;
6398 static void register_sched_domain_sysctl(void)
6400 int i
, cpu_num
= num_online_cpus();
6401 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
6404 WARN_ON(sd_ctl_dir
[0].child
);
6405 sd_ctl_dir
[0].child
= entry
;
6410 for_each_online_cpu(i
) {
6411 snprintf(buf
, 32, "cpu%d", i
);
6412 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6414 entry
->child
= sd_alloc_ctl_cpu_table(i
);
6418 WARN_ON(sd_sysctl_header
);
6419 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
6422 /* may be called multiple times per register */
6423 static void unregister_sched_domain_sysctl(void)
6425 if (sd_sysctl_header
)
6426 unregister_sysctl_table(sd_sysctl_header
);
6427 sd_sysctl_header
= NULL
;
6428 if (sd_ctl_dir
[0].child
)
6429 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
6432 static void register_sched_domain_sysctl(void)
6435 static void unregister_sched_domain_sysctl(void)
6440 static void set_rq_online(struct rq
*rq
)
6443 const struct sched_class
*class;
6445 cpu_set(rq
->cpu
, rq
->rd
->online
);
6448 for_each_class(class) {
6449 if (class->rq_online
)
6450 class->rq_online(rq
);
6455 static void set_rq_offline(struct rq
*rq
)
6458 const struct sched_class
*class;
6460 for_each_class(class) {
6461 if (class->rq_offline
)
6462 class->rq_offline(rq
);
6465 cpu_clear(rq
->cpu
, rq
->rd
->online
);
6471 * migration_call - callback that gets triggered when a CPU is added.
6472 * Here we can start up the necessary migration thread for the new CPU.
6474 static int __cpuinit
6475 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
6477 struct task_struct
*p
;
6478 int cpu
= (long)hcpu
;
6479 unsigned long flags
;
6484 case CPU_UP_PREPARE
:
6485 case CPU_UP_PREPARE_FROZEN
:
6486 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
6489 kthread_bind(p
, cpu
);
6490 /* Must be high prio: stop_machine expects to yield to it. */
6491 rq
= task_rq_lock(p
, &flags
);
6492 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6493 task_rq_unlock(rq
, &flags
);
6494 cpu_rq(cpu
)->migration_thread
= p
;
6498 case CPU_ONLINE_FROZEN
:
6499 /* Strictly unnecessary, as first user will wake it. */
6500 wake_up_process(cpu_rq(cpu
)->migration_thread
);
6502 /* Update our root-domain */
6504 spin_lock_irqsave(&rq
->lock
, flags
);
6506 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6510 spin_unlock_irqrestore(&rq
->lock
, flags
);
6513 #ifdef CONFIG_HOTPLUG_CPU
6514 case CPU_UP_CANCELED
:
6515 case CPU_UP_CANCELED_FROZEN
:
6516 if (!cpu_rq(cpu
)->migration_thread
)
6518 /* Unbind it from offline cpu so it can run. Fall thru. */
6519 kthread_bind(cpu_rq(cpu
)->migration_thread
,
6520 any_online_cpu(cpu_online_map
));
6521 kthread_stop(cpu_rq(cpu
)->migration_thread
);
6522 cpu_rq(cpu
)->migration_thread
= NULL
;
6526 case CPU_DEAD_FROZEN
:
6527 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6528 migrate_live_tasks(cpu
);
6530 kthread_stop(rq
->migration_thread
);
6531 rq
->migration_thread
= NULL
;
6532 /* Idle task back to normal (off runqueue, low prio) */
6533 spin_lock_irq(&rq
->lock
);
6534 update_rq_clock(rq
);
6535 deactivate_task(rq
, rq
->idle
, 0);
6536 rq
->idle
->static_prio
= MAX_PRIO
;
6537 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
6538 rq
->idle
->sched_class
= &idle_sched_class
;
6539 migrate_dead_tasks(cpu
);
6540 spin_unlock_irq(&rq
->lock
);
6542 migrate_nr_uninterruptible(rq
);
6543 BUG_ON(rq
->nr_running
!= 0);
6546 * No need to migrate the tasks: it was best-effort if
6547 * they didn't take sched_hotcpu_mutex. Just wake up
6550 spin_lock_irq(&rq
->lock
);
6551 while (!list_empty(&rq
->migration_queue
)) {
6552 struct migration_req
*req
;
6554 req
= list_entry(rq
->migration_queue
.next
,
6555 struct migration_req
, list
);
6556 list_del_init(&req
->list
);
6557 complete(&req
->done
);
6559 spin_unlock_irq(&rq
->lock
);
6563 case CPU_DYING_FROZEN
:
6564 /* Update our root-domain */
6566 spin_lock_irqsave(&rq
->lock
, flags
);
6568 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6571 spin_unlock_irqrestore(&rq
->lock
, flags
);
6578 /* Register at highest priority so that task migration (migrate_all_tasks)
6579 * happens before everything else.
6581 static struct notifier_block __cpuinitdata migration_notifier
= {
6582 .notifier_call
= migration_call
,
6586 static int __init
migration_init(void)
6588 void *cpu
= (void *)(long)smp_processor_id();
6591 /* Start one for the boot CPU: */
6592 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6593 BUG_ON(err
== NOTIFY_BAD
);
6594 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6595 register_cpu_notifier(&migration_notifier
);
6599 early_initcall(migration_init
);
6604 #ifdef CONFIG_SCHED_DEBUG
6606 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6607 cpumask_t
*groupmask
)
6609 struct sched_group
*group
= sd
->groups
;
6612 cpulist_scnprintf(str
, sizeof(str
), sd
->span
);
6613 cpus_clear(*groupmask
);
6615 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6617 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6618 printk("does not load-balance\n");
6620 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6625 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
6627 if (!cpu_isset(cpu
, sd
->span
)) {
6628 printk(KERN_ERR
"ERROR: domain->span does not contain "
6631 if (!cpu_isset(cpu
, group
->cpumask
)) {
6632 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6636 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6640 printk(KERN_ERR
"ERROR: group is NULL\n");
6644 if (!group
->__cpu_power
) {
6645 printk(KERN_CONT
"\n");
6646 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6651 if (!cpus_weight(group
->cpumask
)) {
6652 printk(KERN_CONT
"\n");
6653 printk(KERN_ERR
"ERROR: empty group\n");
6657 if (cpus_intersects(*groupmask
, group
->cpumask
)) {
6658 printk(KERN_CONT
"\n");
6659 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6663 cpus_or(*groupmask
, *groupmask
, group
->cpumask
);
6665 cpulist_scnprintf(str
, sizeof(str
), group
->cpumask
);
6666 printk(KERN_CONT
" %s", str
);
6668 group
= group
->next
;
6669 } while (group
!= sd
->groups
);
6670 printk(KERN_CONT
"\n");
6672 if (!cpus_equal(sd
->span
, *groupmask
))
6673 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6675 if (sd
->parent
&& !cpus_subset(*groupmask
, sd
->parent
->span
))
6676 printk(KERN_ERR
"ERROR: parent span is not a superset "
6677 "of domain->span\n");
6681 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6683 cpumask_t
*groupmask
;
6687 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6691 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6693 groupmask
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
6695 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6700 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6709 #else /* !CONFIG_SCHED_DEBUG */
6710 # define sched_domain_debug(sd, cpu) do { } while (0)
6711 #endif /* CONFIG_SCHED_DEBUG */
6713 static int sd_degenerate(struct sched_domain
*sd
)
6715 if (cpus_weight(sd
->span
) == 1)
6718 /* Following flags need at least 2 groups */
6719 if (sd
->flags
& (SD_LOAD_BALANCE
|
6720 SD_BALANCE_NEWIDLE
|
6724 SD_SHARE_PKG_RESOURCES
)) {
6725 if (sd
->groups
!= sd
->groups
->next
)
6729 /* Following flags don't use groups */
6730 if (sd
->flags
& (SD_WAKE_IDLE
|
6739 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6741 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6743 if (sd_degenerate(parent
))
6746 if (!cpus_equal(sd
->span
, parent
->span
))
6749 /* Does parent contain flags not in child? */
6750 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6751 if (cflags
& SD_WAKE_AFFINE
)
6752 pflags
&= ~SD_WAKE_BALANCE
;
6753 /* Flags needing groups don't count if only 1 group in parent */
6754 if (parent
->groups
== parent
->groups
->next
) {
6755 pflags
&= ~(SD_LOAD_BALANCE
|
6756 SD_BALANCE_NEWIDLE
|
6760 SD_SHARE_PKG_RESOURCES
);
6762 if (~cflags
& pflags
)
6768 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6770 unsigned long flags
;
6772 spin_lock_irqsave(&rq
->lock
, flags
);
6775 struct root_domain
*old_rd
= rq
->rd
;
6777 if (cpu_isset(rq
->cpu
, old_rd
->online
))
6780 cpu_clear(rq
->cpu
, old_rd
->span
);
6782 if (atomic_dec_and_test(&old_rd
->refcount
))
6786 atomic_inc(&rd
->refcount
);
6789 cpu_set(rq
->cpu
, rd
->span
);
6790 if (cpu_isset(rq
->cpu
, cpu_online_map
))
6793 spin_unlock_irqrestore(&rq
->lock
, flags
);
6796 static void init_rootdomain(struct root_domain
*rd
)
6798 memset(rd
, 0, sizeof(*rd
));
6800 cpus_clear(rd
->span
);
6801 cpus_clear(rd
->online
);
6803 cpupri_init(&rd
->cpupri
);
6806 static void init_defrootdomain(void)
6808 init_rootdomain(&def_root_domain
);
6809 atomic_set(&def_root_domain
.refcount
, 1);
6812 static struct root_domain
*alloc_rootdomain(void)
6814 struct root_domain
*rd
;
6816 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6820 init_rootdomain(rd
);
6826 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6827 * hold the hotplug lock.
6830 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6832 struct rq
*rq
= cpu_rq(cpu
);
6833 struct sched_domain
*tmp
;
6835 /* Remove the sched domains which do not contribute to scheduling. */
6836 for (tmp
= sd
; tmp
; ) {
6837 struct sched_domain
*parent
= tmp
->parent
;
6841 if (sd_parent_degenerate(tmp
, parent
)) {
6842 tmp
->parent
= parent
->parent
;
6844 parent
->parent
->child
= tmp
;
6849 if (sd
&& sd_degenerate(sd
)) {
6855 sched_domain_debug(sd
, cpu
);
6857 rq_attach_root(rq
, rd
);
6858 rcu_assign_pointer(rq
->sd
, sd
);
6861 /* cpus with isolated domains */
6862 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
6864 /* Setup the mask of cpus configured for isolated domains */
6865 static int __init
isolated_cpu_setup(char *str
)
6867 static int __initdata ints
[NR_CPUS
];
6870 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
6871 cpus_clear(cpu_isolated_map
);
6872 for (i
= 1; i
<= ints
[0]; i
++)
6873 if (ints
[i
] < NR_CPUS
)
6874 cpu_set(ints
[i
], cpu_isolated_map
);
6878 __setup("isolcpus=", isolated_cpu_setup
);
6881 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6882 * to a function which identifies what group(along with sched group) a CPU
6883 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6884 * (due to the fact that we keep track of groups covered with a cpumask_t).
6886 * init_sched_build_groups will build a circular linked list of the groups
6887 * covered by the given span, and will set each group's ->cpumask correctly,
6888 * and ->cpu_power to 0.
6891 init_sched_build_groups(const cpumask_t
*span
, const cpumask_t
*cpu_map
,
6892 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
6893 struct sched_group
**sg
,
6894 cpumask_t
*tmpmask
),
6895 cpumask_t
*covered
, cpumask_t
*tmpmask
)
6897 struct sched_group
*first
= NULL
, *last
= NULL
;
6900 cpus_clear(*covered
);
6902 for_each_cpu_mask_nr(i
, *span
) {
6903 struct sched_group
*sg
;
6904 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6907 if (cpu_isset(i
, *covered
))
6910 cpus_clear(sg
->cpumask
);
6911 sg
->__cpu_power
= 0;
6913 for_each_cpu_mask_nr(j
, *span
) {
6914 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6917 cpu_set(j
, *covered
);
6918 cpu_set(j
, sg
->cpumask
);
6929 #define SD_NODES_PER_DOMAIN 16
6934 * find_next_best_node - find the next node to include in a sched_domain
6935 * @node: node whose sched_domain we're building
6936 * @used_nodes: nodes already in the sched_domain
6938 * Find the next node to include in a given scheduling domain. Simply
6939 * finds the closest node not already in the @used_nodes map.
6941 * Should use nodemask_t.
6943 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
6945 int i
, n
, val
, min_val
, best_node
= 0;
6949 for (i
= 0; i
< nr_node_ids
; i
++) {
6950 /* Start at @node */
6951 n
= (node
+ i
) % nr_node_ids
;
6953 if (!nr_cpus_node(n
))
6956 /* Skip already used nodes */
6957 if (node_isset(n
, *used_nodes
))
6960 /* Simple min distance search */
6961 val
= node_distance(node
, n
);
6963 if (val
< min_val
) {
6969 node_set(best_node
, *used_nodes
);
6974 * sched_domain_node_span - get a cpumask for a node's sched_domain
6975 * @node: node whose cpumask we're constructing
6976 * @span: resulting cpumask
6978 * Given a node, construct a good cpumask for its sched_domain to span. It
6979 * should be one that prevents unnecessary balancing, but also spreads tasks
6982 static void sched_domain_node_span(int node
, cpumask_t
*span
)
6984 nodemask_t used_nodes
;
6985 node_to_cpumask_ptr(nodemask
, node
);
6989 nodes_clear(used_nodes
);
6991 cpus_or(*span
, *span
, *nodemask
);
6992 node_set(node
, used_nodes
);
6994 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6995 int next_node
= find_next_best_node(node
, &used_nodes
);
6997 node_to_cpumask_ptr_next(nodemask
, next_node
);
6998 cpus_or(*span
, *span
, *nodemask
);
7001 #endif /* CONFIG_NUMA */
7003 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
7006 * SMT sched-domains:
7008 #ifdef CONFIG_SCHED_SMT
7009 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
7010 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
7013 cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7017 *sg
= &per_cpu(sched_group_cpus
, cpu
);
7020 #endif /* CONFIG_SCHED_SMT */
7023 * multi-core sched-domains:
7025 #ifdef CONFIG_SCHED_MC
7026 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
7027 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
7028 #endif /* CONFIG_SCHED_MC */
7030 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7032 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7037 *mask
= per_cpu(cpu_sibling_map
, cpu
);
7038 cpus_and(*mask
, *mask
, *cpu_map
);
7039 group
= first_cpu(*mask
);
7041 *sg
= &per_cpu(sched_group_core
, group
);
7044 #elif defined(CONFIG_SCHED_MC)
7046 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7050 *sg
= &per_cpu(sched_group_core
, cpu
);
7055 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
7056 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
7059 cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7063 #ifdef CONFIG_SCHED_MC
7064 *mask
= cpu_coregroup_map(cpu
);
7065 cpus_and(*mask
, *mask
, *cpu_map
);
7066 group
= first_cpu(*mask
);
7067 #elif defined(CONFIG_SCHED_SMT)
7068 *mask
= per_cpu(cpu_sibling_map
, cpu
);
7069 cpus_and(*mask
, *mask
, *cpu_map
);
7070 group
= first_cpu(*mask
);
7075 *sg
= &per_cpu(sched_group_phys
, group
);
7081 * The init_sched_build_groups can't handle what we want to do with node
7082 * groups, so roll our own. Now each node has its own list of groups which
7083 * gets dynamically allocated.
7085 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
7086 static struct sched_group
***sched_group_nodes_bycpu
;
7088 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
7089 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
7091 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
7092 struct sched_group
**sg
, cpumask_t
*nodemask
)
7096 *nodemask
= node_to_cpumask(cpu_to_node(cpu
));
7097 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7098 group
= first_cpu(*nodemask
);
7101 *sg
= &per_cpu(sched_group_allnodes
, group
);
7105 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
7107 struct sched_group
*sg
= group_head
;
7113 for_each_cpu_mask_nr(j
, sg
->cpumask
) {
7114 struct sched_domain
*sd
;
7116 sd
= &per_cpu(phys_domains
, j
);
7117 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
7119 * Only add "power" once for each
7125 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
7128 } while (sg
!= group_head
);
7130 #endif /* CONFIG_NUMA */
7133 /* Free memory allocated for various sched_group structures */
7134 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
7138 for_each_cpu_mask_nr(cpu
, *cpu_map
) {
7139 struct sched_group
**sched_group_nodes
7140 = sched_group_nodes_bycpu
[cpu
];
7142 if (!sched_group_nodes
)
7145 for (i
= 0; i
< nr_node_ids
; i
++) {
7146 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
7148 *nodemask
= node_to_cpumask(i
);
7149 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7150 if (cpus_empty(*nodemask
))
7160 if (oldsg
!= sched_group_nodes
[i
])
7163 kfree(sched_group_nodes
);
7164 sched_group_nodes_bycpu
[cpu
] = NULL
;
7167 #else /* !CONFIG_NUMA */
7168 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
7171 #endif /* CONFIG_NUMA */
7174 * Initialize sched groups cpu_power.
7176 * cpu_power indicates the capacity of sched group, which is used while
7177 * distributing the load between different sched groups in a sched domain.
7178 * Typically cpu_power for all the groups in a sched domain will be same unless
7179 * there are asymmetries in the topology. If there are asymmetries, group
7180 * having more cpu_power will pickup more load compared to the group having
7183 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7184 * the maximum number of tasks a group can handle in the presence of other idle
7185 * or lightly loaded groups in the same sched domain.
7187 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
7189 struct sched_domain
*child
;
7190 struct sched_group
*group
;
7192 WARN_ON(!sd
|| !sd
->groups
);
7194 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
7199 sd
->groups
->__cpu_power
= 0;
7202 * For perf policy, if the groups in child domain share resources
7203 * (for example cores sharing some portions of the cache hierarchy
7204 * or SMT), then set this domain groups cpu_power such that each group
7205 * can handle only one task, when there are other idle groups in the
7206 * same sched domain.
7208 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
7210 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
7211 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
7216 * add cpu_power of each child group to this groups cpu_power
7218 group
= child
->groups
;
7220 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
7221 group
= group
->next
;
7222 } while (group
!= child
->groups
);
7226 * Initializers for schedule domains
7227 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7230 #ifdef CONFIG_SCHED_DEBUG
7231 # define SD_INIT_NAME(sd, type) sd->name = #type
7233 # define SD_INIT_NAME(sd, type) do { } while (0)
7236 #define SD_INIT(sd, type) sd_init_##type(sd)
7238 #define SD_INIT_FUNC(type) \
7239 static noinline void sd_init_##type(struct sched_domain *sd) \
7241 memset(sd, 0, sizeof(*sd)); \
7242 *sd = SD_##type##_INIT; \
7243 sd->level = SD_LV_##type; \
7244 SD_INIT_NAME(sd, type); \
7249 SD_INIT_FUNC(ALLNODES
)
7252 #ifdef CONFIG_SCHED_SMT
7253 SD_INIT_FUNC(SIBLING
)
7255 #ifdef CONFIG_SCHED_MC
7260 * To minimize stack usage kmalloc room for cpumasks and share the
7261 * space as the usage in build_sched_domains() dictates. Used only
7262 * if the amount of space is significant.
7265 cpumask_t tmpmask
; /* make this one first */
7268 cpumask_t this_sibling_map
;
7269 cpumask_t this_core_map
;
7271 cpumask_t send_covered
;
7274 cpumask_t domainspan
;
7276 cpumask_t notcovered
;
7281 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7282 static inline void sched_cpumask_alloc(struct allmasks
**masks
)
7284 *masks
= kmalloc(sizeof(**masks
), GFP_KERNEL
);
7286 static inline void sched_cpumask_free(struct allmasks
*masks
)
7291 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7292 static inline void sched_cpumask_alloc(struct allmasks
**masks
)
7294 static inline void sched_cpumask_free(struct allmasks
*masks
)
7298 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7299 ((unsigned long)(a) + offsetof(struct allmasks, v))
7301 static int default_relax_domain_level
= -1;
7303 static int __init
setup_relax_domain_level(char *str
)
7307 val
= simple_strtoul(str
, NULL
, 0);
7308 if (val
< SD_LV_MAX
)
7309 default_relax_domain_level
= val
;
7313 __setup("relax_domain_level=", setup_relax_domain_level
);
7315 static void set_domain_attribute(struct sched_domain
*sd
,
7316 struct sched_domain_attr
*attr
)
7320 if (!attr
|| attr
->relax_domain_level
< 0) {
7321 if (default_relax_domain_level
< 0)
7324 request
= default_relax_domain_level
;
7326 request
= attr
->relax_domain_level
;
7327 if (request
< sd
->level
) {
7328 /* turn off idle balance on this domain */
7329 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
7331 /* turn on idle balance on this domain */
7332 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
7337 * Build sched domains for a given set of cpus and attach the sched domains
7338 * to the individual cpus
7340 static int __build_sched_domains(const cpumask_t
*cpu_map
,
7341 struct sched_domain_attr
*attr
)
7344 struct root_domain
*rd
;
7345 SCHED_CPUMASK_DECLARE(allmasks
);
7348 struct sched_group
**sched_group_nodes
= NULL
;
7349 int sd_allnodes
= 0;
7352 * Allocate the per-node list of sched groups
7354 sched_group_nodes
= kcalloc(nr_node_ids
, sizeof(struct sched_group
*),
7356 if (!sched_group_nodes
) {
7357 printk(KERN_WARNING
"Can not alloc sched group node list\n");
7362 rd
= alloc_rootdomain();
7364 printk(KERN_WARNING
"Cannot alloc root domain\n");
7366 kfree(sched_group_nodes
);
7371 /* get space for all scratch cpumask variables */
7372 sched_cpumask_alloc(&allmasks
);
7374 printk(KERN_WARNING
"Cannot alloc cpumask array\n");
7377 kfree(sched_group_nodes
);
7382 tmpmask
= (cpumask_t
*)allmasks
;
7386 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
7390 * Set up domains for cpus specified by the cpu_map.
7392 for_each_cpu_mask_nr(i
, *cpu_map
) {
7393 struct sched_domain
*sd
= NULL
, *p
;
7394 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7396 *nodemask
= node_to_cpumask(cpu_to_node(i
));
7397 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7400 if (cpus_weight(*cpu_map
) >
7401 SD_NODES_PER_DOMAIN
*cpus_weight(*nodemask
)) {
7402 sd
= &per_cpu(allnodes_domains
, i
);
7403 SD_INIT(sd
, ALLNODES
);
7404 set_domain_attribute(sd
, attr
);
7405 sd
->span
= *cpu_map
;
7406 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7412 sd
= &per_cpu(node_domains
, i
);
7414 set_domain_attribute(sd
, attr
);
7415 sched_domain_node_span(cpu_to_node(i
), &sd
->span
);
7419 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7423 sd
= &per_cpu(phys_domains
, i
);
7425 set_domain_attribute(sd
, attr
);
7426 sd
->span
= *nodemask
;
7430 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7432 #ifdef CONFIG_SCHED_MC
7434 sd
= &per_cpu(core_domains
, i
);
7436 set_domain_attribute(sd
, attr
);
7437 sd
->span
= cpu_coregroup_map(i
);
7438 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7441 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7444 #ifdef CONFIG_SCHED_SMT
7446 sd
= &per_cpu(cpu_domains
, i
);
7447 SD_INIT(sd
, SIBLING
);
7448 set_domain_attribute(sd
, attr
);
7449 sd
->span
= per_cpu(cpu_sibling_map
, i
);
7450 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7453 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7457 #ifdef CONFIG_SCHED_SMT
7458 /* Set up CPU (sibling) groups */
7459 for_each_cpu_mask_nr(i
, *cpu_map
) {
7460 SCHED_CPUMASK_VAR(this_sibling_map
, allmasks
);
7461 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7463 *this_sibling_map
= per_cpu(cpu_sibling_map
, i
);
7464 cpus_and(*this_sibling_map
, *this_sibling_map
, *cpu_map
);
7465 if (i
!= first_cpu(*this_sibling_map
))
7468 init_sched_build_groups(this_sibling_map
, cpu_map
,
7470 send_covered
, tmpmask
);
7474 #ifdef CONFIG_SCHED_MC
7475 /* Set up multi-core groups */
7476 for_each_cpu_mask_nr(i
, *cpu_map
) {
7477 SCHED_CPUMASK_VAR(this_core_map
, allmasks
);
7478 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7480 *this_core_map
= cpu_coregroup_map(i
);
7481 cpus_and(*this_core_map
, *this_core_map
, *cpu_map
);
7482 if (i
!= first_cpu(*this_core_map
))
7485 init_sched_build_groups(this_core_map
, cpu_map
,
7487 send_covered
, tmpmask
);
7491 /* Set up physical groups */
7492 for (i
= 0; i
< nr_node_ids
; i
++) {
7493 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7494 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7496 *nodemask
= node_to_cpumask(i
);
7497 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7498 if (cpus_empty(*nodemask
))
7501 init_sched_build_groups(nodemask
, cpu_map
,
7503 send_covered
, tmpmask
);
7507 /* Set up node groups */
7509 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7511 init_sched_build_groups(cpu_map
, cpu_map
,
7512 &cpu_to_allnodes_group
,
7513 send_covered
, tmpmask
);
7516 for (i
= 0; i
< nr_node_ids
; i
++) {
7517 /* Set up node groups */
7518 struct sched_group
*sg
, *prev
;
7519 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7520 SCHED_CPUMASK_VAR(domainspan
, allmasks
);
7521 SCHED_CPUMASK_VAR(covered
, allmasks
);
7524 *nodemask
= node_to_cpumask(i
);
7525 cpus_clear(*covered
);
7527 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7528 if (cpus_empty(*nodemask
)) {
7529 sched_group_nodes
[i
] = NULL
;
7533 sched_domain_node_span(i
, domainspan
);
7534 cpus_and(*domainspan
, *domainspan
, *cpu_map
);
7536 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
7538 printk(KERN_WARNING
"Can not alloc domain group for "
7542 sched_group_nodes
[i
] = sg
;
7543 for_each_cpu_mask_nr(j
, *nodemask
) {
7544 struct sched_domain
*sd
;
7546 sd
= &per_cpu(node_domains
, j
);
7549 sg
->__cpu_power
= 0;
7550 sg
->cpumask
= *nodemask
;
7552 cpus_or(*covered
, *covered
, *nodemask
);
7555 for (j
= 0; j
< nr_node_ids
; j
++) {
7556 SCHED_CPUMASK_VAR(notcovered
, allmasks
);
7557 int n
= (i
+ j
) % nr_node_ids
;
7558 node_to_cpumask_ptr(pnodemask
, n
);
7560 cpus_complement(*notcovered
, *covered
);
7561 cpus_and(*tmpmask
, *notcovered
, *cpu_map
);
7562 cpus_and(*tmpmask
, *tmpmask
, *domainspan
);
7563 if (cpus_empty(*tmpmask
))
7566 cpus_and(*tmpmask
, *tmpmask
, *pnodemask
);
7567 if (cpus_empty(*tmpmask
))
7570 sg
= kmalloc_node(sizeof(struct sched_group
),
7574 "Can not alloc domain group for node %d\n", j
);
7577 sg
->__cpu_power
= 0;
7578 sg
->cpumask
= *tmpmask
;
7579 sg
->next
= prev
->next
;
7580 cpus_or(*covered
, *covered
, *tmpmask
);
7587 /* Calculate CPU power for physical packages and nodes */
7588 #ifdef CONFIG_SCHED_SMT
7589 for_each_cpu_mask_nr(i
, *cpu_map
) {
7590 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
7592 init_sched_groups_power(i
, sd
);
7595 #ifdef CONFIG_SCHED_MC
7596 for_each_cpu_mask_nr(i
, *cpu_map
) {
7597 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
7599 init_sched_groups_power(i
, sd
);
7603 for_each_cpu_mask_nr(i
, *cpu_map
) {
7604 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
7606 init_sched_groups_power(i
, sd
);
7610 for (i
= 0; i
< nr_node_ids
; i
++)
7611 init_numa_sched_groups_power(sched_group_nodes
[i
]);
7614 struct sched_group
*sg
;
7616 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
,
7618 init_numa_sched_groups_power(sg
);
7622 /* Attach the domains */
7623 for_each_cpu_mask_nr(i
, *cpu_map
) {
7624 struct sched_domain
*sd
;
7625 #ifdef CONFIG_SCHED_SMT
7626 sd
= &per_cpu(cpu_domains
, i
);
7627 #elif defined(CONFIG_SCHED_MC)
7628 sd
= &per_cpu(core_domains
, i
);
7630 sd
= &per_cpu(phys_domains
, i
);
7632 cpu_attach_domain(sd
, rd
, i
);
7635 sched_cpumask_free(allmasks
);
7640 free_sched_groups(cpu_map
, tmpmask
);
7641 sched_cpumask_free(allmasks
);
7647 static int build_sched_domains(const cpumask_t
*cpu_map
)
7649 return __build_sched_domains(cpu_map
, NULL
);
7652 static cpumask_t
*doms_cur
; /* current sched domains */
7653 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7654 static struct sched_domain_attr
*dattr_cur
;
7655 /* attribues of custom domains in 'doms_cur' */
7658 * Special case: If a kmalloc of a doms_cur partition (array of
7659 * cpumask_t) fails, then fallback to a single sched domain,
7660 * as determined by the single cpumask_t fallback_doms.
7662 static cpumask_t fallback_doms
;
7664 void __attribute__((weak
)) arch_update_cpu_topology(void)
7669 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7670 * For now this just excludes isolated cpus, but could be used to
7671 * exclude other special cases in the future.
7673 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
7677 arch_update_cpu_topology();
7679 doms_cur
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
7681 doms_cur
= &fallback_doms
;
7682 cpus_andnot(*doms_cur
, *cpu_map
, cpu_isolated_map
);
7684 err
= build_sched_domains(doms_cur
);
7685 register_sched_domain_sysctl();
7690 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
,
7693 free_sched_groups(cpu_map
, tmpmask
);
7697 * Detach sched domains from a group of cpus specified in cpu_map
7698 * These cpus will now be attached to the NULL domain
7700 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
7705 for_each_cpu_mask_nr(i
, *cpu_map
)
7706 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7707 synchronize_sched();
7708 arch_destroy_sched_domains(cpu_map
, &tmpmask
);
7711 /* handle null as "default" */
7712 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7713 struct sched_domain_attr
*new, int idx_new
)
7715 struct sched_domain_attr tmp
;
7722 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7723 new ? (new + idx_new
) : &tmp
,
7724 sizeof(struct sched_domain_attr
));
7728 * Partition sched domains as specified by the 'ndoms_new'
7729 * cpumasks in the array doms_new[] of cpumasks. This compares
7730 * doms_new[] to the current sched domain partitioning, doms_cur[].
7731 * It destroys each deleted domain and builds each new domain.
7733 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7734 * The masks don't intersect (don't overlap.) We should setup one
7735 * sched domain for each mask. CPUs not in any of the cpumasks will
7736 * not be load balanced. If the same cpumask appears both in the
7737 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7740 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7741 * ownership of it and will kfree it when done with it. If the caller
7742 * failed the kmalloc call, then it can pass in doms_new == NULL &&
7743 * ndoms_new == 1, and partition_sched_domains() will fallback to
7744 * the single partition 'fallback_doms', it also forces the domains
7747 * If doms_new == NULL it will be replaced with cpu_online_map.
7748 * ndoms_new == 0 is a special case for destroying existing domains,
7749 * and it will not create the default domain.
7751 * Call with hotplug lock held
7753 void partition_sched_domains(int ndoms_new
, cpumask_t
*doms_new
,
7754 struct sched_domain_attr
*dattr_new
)
7758 mutex_lock(&sched_domains_mutex
);
7760 /* always unregister in case we don't destroy any domains */
7761 unregister_sched_domain_sysctl();
7763 n
= doms_new
? ndoms_new
: 0;
7765 /* Destroy deleted domains */
7766 for (i
= 0; i
< ndoms_cur
; i
++) {
7767 for (j
= 0; j
< n
; j
++) {
7768 if (cpus_equal(doms_cur
[i
], doms_new
[j
])
7769 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7772 /* no match - a current sched domain not in new doms_new[] */
7773 detach_destroy_domains(doms_cur
+ i
);
7778 if (doms_new
== NULL
) {
7780 doms_new
= &fallback_doms
;
7781 cpus_andnot(doms_new
[0], cpu_online_map
, cpu_isolated_map
);
7782 WARN_ON_ONCE(dattr_new
);
7785 /* Build new domains */
7786 for (i
= 0; i
< ndoms_new
; i
++) {
7787 for (j
= 0; j
< ndoms_cur
; j
++) {
7788 if (cpus_equal(doms_new
[i
], doms_cur
[j
])
7789 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7792 /* no match - add a new doms_new */
7793 __build_sched_domains(doms_new
+ i
,
7794 dattr_new
? dattr_new
+ i
: NULL
);
7799 /* Remember the new sched domains */
7800 if (doms_cur
!= &fallback_doms
)
7802 kfree(dattr_cur
); /* kfree(NULL) is safe */
7803 doms_cur
= doms_new
;
7804 dattr_cur
= dattr_new
;
7805 ndoms_cur
= ndoms_new
;
7807 register_sched_domain_sysctl();
7809 mutex_unlock(&sched_domains_mutex
);
7812 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7813 int arch_reinit_sched_domains(void)
7817 /* Destroy domains first to force the rebuild */
7818 partition_sched_domains(0, NULL
, NULL
);
7820 rebuild_sched_domains();
7826 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7830 if (buf
[0] != '0' && buf
[0] != '1')
7834 sched_smt_power_savings
= (buf
[0] == '1');
7836 sched_mc_power_savings
= (buf
[0] == '1');
7838 ret
= arch_reinit_sched_domains();
7840 return ret
? ret
: count
;
7843 #ifdef CONFIG_SCHED_MC
7844 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
7847 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7849 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
7850 const char *buf
, size_t count
)
7852 return sched_power_savings_store(buf
, count
, 0);
7854 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
7855 sched_mc_power_savings_show
,
7856 sched_mc_power_savings_store
);
7859 #ifdef CONFIG_SCHED_SMT
7860 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
7863 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7865 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
7866 const char *buf
, size_t count
)
7868 return sched_power_savings_store(buf
, count
, 1);
7870 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
7871 sched_smt_power_savings_show
,
7872 sched_smt_power_savings_store
);
7875 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7879 #ifdef CONFIG_SCHED_SMT
7881 err
= sysfs_create_file(&cls
->kset
.kobj
,
7882 &attr_sched_smt_power_savings
.attr
);
7884 #ifdef CONFIG_SCHED_MC
7885 if (!err
&& mc_capable())
7886 err
= sysfs_create_file(&cls
->kset
.kobj
,
7887 &attr_sched_mc_power_savings
.attr
);
7891 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7893 #ifndef CONFIG_CPUSETS
7895 * Add online and remove offline CPUs from the scheduler domains.
7896 * When cpusets are enabled they take over this function.
7898 static int update_sched_domains(struct notifier_block
*nfb
,
7899 unsigned long action
, void *hcpu
)
7903 case CPU_ONLINE_FROZEN
:
7905 case CPU_DEAD_FROZEN
:
7906 partition_sched_domains(1, NULL
, NULL
);
7915 static int update_runtime(struct notifier_block
*nfb
,
7916 unsigned long action
, void *hcpu
)
7918 int cpu
= (int)(long)hcpu
;
7921 case CPU_DOWN_PREPARE
:
7922 case CPU_DOWN_PREPARE_FROZEN
:
7923 disable_runtime(cpu_rq(cpu
));
7926 case CPU_DOWN_FAILED
:
7927 case CPU_DOWN_FAILED_FROZEN
:
7929 case CPU_ONLINE_FROZEN
:
7930 enable_runtime(cpu_rq(cpu
));
7938 void __init
sched_init_smp(void)
7940 cpumask_t non_isolated_cpus
;
7942 #if defined(CONFIG_NUMA)
7943 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
7945 BUG_ON(sched_group_nodes_bycpu
== NULL
);
7948 mutex_lock(&sched_domains_mutex
);
7949 arch_init_sched_domains(&cpu_online_map
);
7950 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
7951 if (cpus_empty(non_isolated_cpus
))
7952 cpu_set(smp_processor_id(), non_isolated_cpus
);
7953 mutex_unlock(&sched_domains_mutex
);
7956 #ifndef CONFIG_CPUSETS
7957 /* XXX: Theoretical race here - CPU may be hotplugged now */
7958 hotcpu_notifier(update_sched_domains
, 0);
7961 /* RT runtime code needs to handle some hotplug events */
7962 hotcpu_notifier(update_runtime
, 0);
7966 /* Move init over to a non-isolated CPU */
7967 if (set_cpus_allowed_ptr(current
, &non_isolated_cpus
) < 0)
7969 sched_init_granularity();
7972 void __init
sched_init_smp(void)
7974 sched_init_granularity();
7976 #endif /* CONFIG_SMP */
7978 int in_sched_functions(unsigned long addr
)
7980 return in_lock_functions(addr
) ||
7981 (addr
>= (unsigned long)__sched_text_start
7982 && addr
< (unsigned long)__sched_text_end
);
7985 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7987 cfs_rq
->tasks_timeline
= RB_ROOT
;
7988 INIT_LIST_HEAD(&cfs_rq
->tasks
);
7989 #ifdef CONFIG_FAIR_GROUP_SCHED
7992 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7995 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7997 struct rt_prio_array
*array
;
8000 array
= &rt_rq
->active
;
8001 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
8002 INIT_LIST_HEAD(array
->queue
+ i
);
8003 __clear_bit(i
, array
->bitmap
);
8005 /* delimiter for bitsearch: */
8006 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
8008 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8009 rt_rq
->highest_prio
= MAX_RT_PRIO
;
8012 rt_rq
->rt_nr_migratory
= 0;
8013 rt_rq
->overloaded
= 0;
8017 rt_rq
->rt_throttled
= 0;
8018 rt_rq
->rt_runtime
= 0;
8019 spin_lock_init(&rt_rq
->rt_runtime_lock
);
8021 #ifdef CONFIG_RT_GROUP_SCHED
8022 rt_rq
->rt_nr_boosted
= 0;
8027 #ifdef CONFIG_FAIR_GROUP_SCHED
8028 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
8029 struct sched_entity
*se
, int cpu
, int add
,
8030 struct sched_entity
*parent
)
8032 struct rq
*rq
= cpu_rq(cpu
);
8033 tg
->cfs_rq
[cpu
] = cfs_rq
;
8034 init_cfs_rq(cfs_rq
, rq
);
8037 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
8040 /* se could be NULL for init_task_group */
8045 se
->cfs_rq
= &rq
->cfs
;
8047 se
->cfs_rq
= parent
->my_q
;
8050 se
->load
.weight
= tg
->shares
;
8051 se
->load
.inv_weight
= 0;
8052 se
->parent
= parent
;
8056 #ifdef CONFIG_RT_GROUP_SCHED
8057 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
8058 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
8059 struct sched_rt_entity
*parent
)
8061 struct rq
*rq
= cpu_rq(cpu
);
8063 tg
->rt_rq
[cpu
] = rt_rq
;
8064 init_rt_rq(rt_rq
, rq
);
8066 rt_rq
->rt_se
= rt_se
;
8067 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8069 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
8071 tg
->rt_se
[cpu
] = rt_se
;
8076 rt_se
->rt_rq
= &rq
->rt
;
8078 rt_se
->rt_rq
= parent
->my_q
;
8080 rt_se
->my_q
= rt_rq
;
8081 rt_se
->parent
= parent
;
8082 INIT_LIST_HEAD(&rt_se
->run_list
);
8086 void __init
sched_init(void)
8089 unsigned long alloc_size
= 0, ptr
;
8091 #ifdef CONFIG_FAIR_GROUP_SCHED
8092 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8094 #ifdef CONFIG_RT_GROUP_SCHED
8095 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8097 #ifdef CONFIG_USER_SCHED
8101 * As sched_init() is called before page_alloc is setup,
8102 * we use alloc_bootmem().
8105 ptr
= (unsigned long)alloc_bootmem(alloc_size
);
8107 #ifdef CONFIG_FAIR_GROUP_SCHED
8108 init_task_group
.se
= (struct sched_entity
**)ptr
;
8109 ptr
+= nr_cpu_ids
* sizeof(void **);
8111 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8112 ptr
+= nr_cpu_ids
* sizeof(void **);
8114 #ifdef CONFIG_USER_SCHED
8115 root_task_group
.se
= (struct sched_entity
**)ptr
;
8116 ptr
+= nr_cpu_ids
* sizeof(void **);
8118 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8119 ptr
+= nr_cpu_ids
* sizeof(void **);
8120 #endif /* CONFIG_USER_SCHED */
8121 #endif /* CONFIG_FAIR_GROUP_SCHED */
8122 #ifdef CONFIG_RT_GROUP_SCHED
8123 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8124 ptr
+= nr_cpu_ids
* sizeof(void **);
8126 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8127 ptr
+= nr_cpu_ids
* sizeof(void **);
8129 #ifdef CONFIG_USER_SCHED
8130 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8131 ptr
+= nr_cpu_ids
* sizeof(void **);
8133 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8134 ptr
+= nr_cpu_ids
* sizeof(void **);
8135 #endif /* CONFIG_USER_SCHED */
8136 #endif /* CONFIG_RT_GROUP_SCHED */
8140 init_defrootdomain();
8143 init_rt_bandwidth(&def_rt_bandwidth
,
8144 global_rt_period(), global_rt_runtime());
8146 #ifdef CONFIG_RT_GROUP_SCHED
8147 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
8148 global_rt_period(), global_rt_runtime());
8149 #ifdef CONFIG_USER_SCHED
8150 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
8151 global_rt_period(), RUNTIME_INF
);
8152 #endif /* CONFIG_USER_SCHED */
8153 #endif /* CONFIG_RT_GROUP_SCHED */
8155 #ifdef CONFIG_GROUP_SCHED
8156 list_add(&init_task_group
.list
, &task_groups
);
8157 INIT_LIST_HEAD(&init_task_group
.children
);
8159 #ifdef CONFIG_USER_SCHED
8160 INIT_LIST_HEAD(&root_task_group
.children
);
8161 init_task_group
.parent
= &root_task_group
;
8162 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
8163 #endif /* CONFIG_USER_SCHED */
8164 #endif /* CONFIG_GROUP_SCHED */
8166 for_each_possible_cpu(i
) {
8170 spin_lock_init(&rq
->lock
);
8172 init_cfs_rq(&rq
->cfs
, rq
);
8173 init_rt_rq(&rq
->rt
, rq
);
8174 #ifdef CONFIG_FAIR_GROUP_SCHED
8175 init_task_group
.shares
= init_task_group_load
;
8176 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
8177 #ifdef CONFIG_CGROUP_SCHED
8179 * How much cpu bandwidth does init_task_group get?
8181 * In case of task-groups formed thr' the cgroup filesystem, it
8182 * gets 100% of the cpu resources in the system. This overall
8183 * system cpu resource is divided among the tasks of
8184 * init_task_group and its child task-groups in a fair manner,
8185 * based on each entity's (task or task-group's) weight
8186 * (se->load.weight).
8188 * In other words, if init_task_group has 10 tasks of weight
8189 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8190 * then A0's share of the cpu resource is:
8192 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8194 * We achieve this by letting init_task_group's tasks sit
8195 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8197 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
8198 #elif defined CONFIG_USER_SCHED
8199 root_task_group
.shares
= NICE_0_LOAD
;
8200 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
8202 * In case of task-groups formed thr' the user id of tasks,
8203 * init_task_group represents tasks belonging to root user.
8204 * Hence it forms a sibling of all subsequent groups formed.
8205 * In this case, init_task_group gets only a fraction of overall
8206 * system cpu resource, based on the weight assigned to root
8207 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8208 * by letting tasks of init_task_group sit in a separate cfs_rq
8209 * (init_cfs_rq) and having one entity represent this group of
8210 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8212 init_tg_cfs_entry(&init_task_group
,
8213 &per_cpu(init_cfs_rq
, i
),
8214 &per_cpu(init_sched_entity
, i
), i
, 1,
8215 root_task_group
.se
[i
]);
8218 #endif /* CONFIG_FAIR_GROUP_SCHED */
8220 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
8221 #ifdef CONFIG_RT_GROUP_SCHED
8222 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
8223 #ifdef CONFIG_CGROUP_SCHED
8224 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
8225 #elif defined CONFIG_USER_SCHED
8226 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
8227 init_tg_rt_entry(&init_task_group
,
8228 &per_cpu(init_rt_rq
, i
),
8229 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
8230 root_task_group
.rt_se
[i
]);
8234 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
8235 rq
->cpu_load
[j
] = 0;
8239 rq
->active_balance
= 0;
8240 rq
->next_balance
= jiffies
;
8244 rq
->migration_thread
= NULL
;
8245 INIT_LIST_HEAD(&rq
->migration_queue
);
8246 rq_attach_root(rq
, &def_root_domain
);
8249 atomic_set(&rq
->nr_iowait
, 0);
8252 set_load_weight(&init_task
);
8254 #ifdef CONFIG_PREEMPT_NOTIFIERS
8255 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
8259 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
8262 #ifdef CONFIG_RT_MUTEXES
8263 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
8267 * The boot idle thread does lazy MMU switching as well:
8269 atomic_inc(&init_mm
.mm_count
);
8270 enter_lazy_tlb(&init_mm
, current
);
8273 * Make us the idle thread. Technically, schedule() should not be
8274 * called from this thread, however somewhere below it might be,
8275 * but because we are the idle thread, we just pick up running again
8276 * when this runqueue becomes "idle".
8278 init_idle(current
, smp_processor_id());
8280 * During early bootup we pretend to be a normal task:
8282 current
->sched_class
= &fair_sched_class
;
8284 scheduler_running
= 1;
8287 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8288 void __might_sleep(char *file
, int line
)
8291 static unsigned long prev_jiffy
; /* ratelimiting */
8293 if ((!in_atomic() && !irqs_disabled()) ||
8294 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
8296 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
8298 prev_jiffy
= jiffies
;
8301 "BUG: sleeping function called from invalid context at %s:%d\n",
8304 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8305 in_atomic(), irqs_disabled(),
8306 current
->pid
, current
->comm
);
8308 debug_show_held_locks(current
);
8309 if (irqs_disabled())
8310 print_irqtrace_events(current
);
8314 EXPORT_SYMBOL(__might_sleep
);
8317 #ifdef CONFIG_MAGIC_SYSRQ
8318 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
8322 update_rq_clock(rq
);
8323 on_rq
= p
->se
.on_rq
;
8325 deactivate_task(rq
, p
, 0);
8326 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
8328 activate_task(rq
, p
, 0);
8329 resched_task(rq
->curr
);
8333 void normalize_rt_tasks(void)
8335 struct task_struct
*g
, *p
;
8336 unsigned long flags
;
8339 read_lock_irqsave(&tasklist_lock
, flags
);
8340 do_each_thread(g
, p
) {
8342 * Only normalize user tasks:
8347 p
->se
.exec_start
= 0;
8348 #ifdef CONFIG_SCHEDSTATS
8349 p
->se
.wait_start
= 0;
8350 p
->se
.sleep_start
= 0;
8351 p
->se
.block_start
= 0;
8356 * Renice negative nice level userspace
8359 if (TASK_NICE(p
) < 0 && p
->mm
)
8360 set_user_nice(p
, 0);
8364 spin_lock(&p
->pi_lock
);
8365 rq
= __task_rq_lock(p
);
8367 normalize_task(rq
, p
);
8369 __task_rq_unlock(rq
);
8370 spin_unlock(&p
->pi_lock
);
8371 } while_each_thread(g
, p
);
8373 read_unlock_irqrestore(&tasklist_lock
, flags
);
8376 #endif /* CONFIG_MAGIC_SYSRQ */
8380 * These functions are only useful for the IA64 MCA handling.
8382 * They can only be called when the whole system has been
8383 * stopped - every CPU needs to be quiescent, and no scheduling
8384 * activity can take place. Using them for anything else would
8385 * be a serious bug, and as a result, they aren't even visible
8386 * under any other configuration.
8390 * curr_task - return the current task for a given cpu.
8391 * @cpu: the processor in question.
8393 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8395 struct task_struct
*curr_task(int cpu
)
8397 return cpu_curr(cpu
);
8401 * set_curr_task - set the current task for a given cpu.
8402 * @cpu: the processor in question.
8403 * @p: the task pointer to set.
8405 * Description: This function must only be used when non-maskable interrupts
8406 * are serviced on a separate stack. It allows the architecture to switch the
8407 * notion of the current task on a cpu in a non-blocking manner. This function
8408 * must be called with all CPU's synchronized, and interrupts disabled, the
8409 * and caller must save the original value of the current task (see
8410 * curr_task() above) and restore that value before reenabling interrupts and
8411 * re-starting the system.
8413 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8415 void set_curr_task(int cpu
, struct task_struct
*p
)
8422 #ifdef CONFIG_FAIR_GROUP_SCHED
8423 static void free_fair_sched_group(struct task_group
*tg
)
8427 for_each_possible_cpu(i
) {
8429 kfree(tg
->cfs_rq
[i
]);
8439 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8441 struct cfs_rq
*cfs_rq
;
8442 struct sched_entity
*se
;
8446 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8449 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8453 tg
->shares
= NICE_0_LOAD
;
8455 for_each_possible_cpu(i
) {
8458 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
8459 GFP_KERNEL
, cpu_to_node(i
));
8463 se
= kzalloc_node(sizeof(struct sched_entity
),
8464 GFP_KERNEL
, cpu_to_node(i
));
8468 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
8477 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8479 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
8480 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
8483 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8485 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
8487 #else /* !CONFG_FAIR_GROUP_SCHED */
8488 static inline void free_fair_sched_group(struct task_group
*tg
)
8493 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8498 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8502 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8505 #endif /* CONFIG_FAIR_GROUP_SCHED */
8507 #ifdef CONFIG_RT_GROUP_SCHED
8508 static void free_rt_sched_group(struct task_group
*tg
)
8512 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8514 for_each_possible_cpu(i
) {
8516 kfree(tg
->rt_rq
[i
]);
8518 kfree(tg
->rt_se
[i
]);
8526 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8528 struct rt_rq
*rt_rq
;
8529 struct sched_rt_entity
*rt_se
;
8533 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8536 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8540 init_rt_bandwidth(&tg
->rt_bandwidth
,
8541 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8543 for_each_possible_cpu(i
) {
8546 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
8547 GFP_KERNEL
, cpu_to_node(i
));
8551 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
8552 GFP_KERNEL
, cpu_to_node(i
));
8556 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
8565 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8567 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8568 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8571 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8573 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8575 #else /* !CONFIG_RT_GROUP_SCHED */
8576 static inline void free_rt_sched_group(struct task_group
*tg
)
8581 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8586 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8590 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8593 #endif /* CONFIG_RT_GROUP_SCHED */
8595 #ifdef CONFIG_GROUP_SCHED
8596 static void free_sched_group(struct task_group
*tg
)
8598 free_fair_sched_group(tg
);
8599 free_rt_sched_group(tg
);
8603 /* allocate runqueue etc for a new task group */
8604 struct task_group
*sched_create_group(struct task_group
*parent
)
8606 struct task_group
*tg
;
8607 unsigned long flags
;
8610 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8612 return ERR_PTR(-ENOMEM
);
8614 if (!alloc_fair_sched_group(tg
, parent
))
8617 if (!alloc_rt_sched_group(tg
, parent
))
8620 spin_lock_irqsave(&task_group_lock
, flags
);
8621 for_each_possible_cpu(i
) {
8622 register_fair_sched_group(tg
, i
);
8623 register_rt_sched_group(tg
, i
);
8625 list_add_rcu(&tg
->list
, &task_groups
);
8627 WARN_ON(!parent
); /* root should already exist */
8629 tg
->parent
= parent
;
8630 INIT_LIST_HEAD(&tg
->children
);
8631 list_add_rcu(&tg
->siblings
, &parent
->children
);
8632 spin_unlock_irqrestore(&task_group_lock
, flags
);
8637 free_sched_group(tg
);
8638 return ERR_PTR(-ENOMEM
);
8641 /* rcu callback to free various structures associated with a task group */
8642 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8644 /* now it should be safe to free those cfs_rqs */
8645 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8648 /* Destroy runqueue etc associated with a task group */
8649 void sched_destroy_group(struct task_group
*tg
)
8651 unsigned long flags
;
8654 spin_lock_irqsave(&task_group_lock
, flags
);
8655 for_each_possible_cpu(i
) {
8656 unregister_fair_sched_group(tg
, i
);
8657 unregister_rt_sched_group(tg
, i
);
8659 list_del_rcu(&tg
->list
);
8660 list_del_rcu(&tg
->siblings
);
8661 spin_unlock_irqrestore(&task_group_lock
, flags
);
8663 /* wait for possible concurrent references to cfs_rqs complete */
8664 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8667 /* change task's runqueue when it moves between groups.
8668 * The caller of this function should have put the task in its new group
8669 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8670 * reflect its new group.
8672 void sched_move_task(struct task_struct
*tsk
)
8675 unsigned long flags
;
8678 rq
= task_rq_lock(tsk
, &flags
);
8680 update_rq_clock(rq
);
8682 running
= task_current(rq
, tsk
);
8683 on_rq
= tsk
->se
.on_rq
;
8686 dequeue_task(rq
, tsk
, 0);
8687 if (unlikely(running
))
8688 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8690 set_task_rq(tsk
, task_cpu(tsk
));
8692 #ifdef CONFIG_FAIR_GROUP_SCHED
8693 if (tsk
->sched_class
->moved_group
)
8694 tsk
->sched_class
->moved_group(tsk
);
8697 if (unlikely(running
))
8698 tsk
->sched_class
->set_curr_task(rq
);
8700 enqueue_task(rq
, tsk
, 0);
8702 task_rq_unlock(rq
, &flags
);
8704 #endif /* CONFIG_GROUP_SCHED */
8706 #ifdef CONFIG_FAIR_GROUP_SCHED
8707 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8709 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8714 dequeue_entity(cfs_rq
, se
, 0);
8716 se
->load
.weight
= shares
;
8717 se
->load
.inv_weight
= 0;
8720 enqueue_entity(cfs_rq
, se
, 0);
8723 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8725 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8726 struct rq
*rq
= cfs_rq
->rq
;
8727 unsigned long flags
;
8729 spin_lock_irqsave(&rq
->lock
, flags
);
8730 __set_se_shares(se
, shares
);
8731 spin_unlock_irqrestore(&rq
->lock
, flags
);
8734 static DEFINE_MUTEX(shares_mutex
);
8736 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8739 unsigned long flags
;
8742 * We can't change the weight of the root cgroup.
8747 if (shares
< MIN_SHARES
)
8748 shares
= MIN_SHARES
;
8749 else if (shares
> MAX_SHARES
)
8750 shares
= MAX_SHARES
;
8752 mutex_lock(&shares_mutex
);
8753 if (tg
->shares
== shares
)
8756 spin_lock_irqsave(&task_group_lock
, flags
);
8757 for_each_possible_cpu(i
)
8758 unregister_fair_sched_group(tg
, i
);
8759 list_del_rcu(&tg
->siblings
);
8760 spin_unlock_irqrestore(&task_group_lock
, flags
);
8762 /* wait for any ongoing reference to this group to finish */
8763 synchronize_sched();
8766 * Now we are free to modify the group's share on each cpu
8767 * w/o tripping rebalance_share or load_balance_fair.
8769 tg
->shares
= shares
;
8770 for_each_possible_cpu(i
) {
8774 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
8775 set_se_shares(tg
->se
[i
], shares
);
8779 * Enable load balance activity on this group, by inserting it back on
8780 * each cpu's rq->leaf_cfs_rq_list.
8782 spin_lock_irqsave(&task_group_lock
, flags
);
8783 for_each_possible_cpu(i
)
8784 register_fair_sched_group(tg
, i
);
8785 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
8786 spin_unlock_irqrestore(&task_group_lock
, flags
);
8788 mutex_unlock(&shares_mutex
);
8792 unsigned long sched_group_shares(struct task_group
*tg
)
8798 #ifdef CONFIG_RT_GROUP_SCHED
8800 * Ensure that the real time constraints are schedulable.
8802 static DEFINE_MUTEX(rt_constraints_mutex
);
8804 static unsigned long to_ratio(u64 period
, u64 runtime
)
8806 if (runtime
== RUNTIME_INF
)
8809 return div64_u64(runtime
<< 20, period
);
8812 /* Must be called with tasklist_lock held */
8813 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8815 struct task_struct
*g
, *p
;
8817 do_each_thread(g
, p
) {
8818 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8820 } while_each_thread(g
, p
);
8825 struct rt_schedulable_data
{
8826 struct task_group
*tg
;
8831 static int tg_schedulable(struct task_group
*tg
, void *data
)
8833 struct rt_schedulable_data
*d
= data
;
8834 struct task_group
*child
;
8835 unsigned long total
, sum
= 0;
8836 u64 period
, runtime
;
8838 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8839 runtime
= tg
->rt_bandwidth
.rt_runtime
;
8842 period
= d
->rt_period
;
8843 runtime
= d
->rt_runtime
;
8847 * Cannot have more runtime than the period.
8849 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8853 * Ensure we don't starve existing RT tasks.
8855 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
8858 total
= to_ratio(period
, runtime
);
8861 * Nobody can have more than the global setting allows.
8863 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
8867 * The sum of our children's runtime should not exceed our own.
8869 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
8870 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
8871 runtime
= child
->rt_bandwidth
.rt_runtime
;
8873 if (child
== d
->tg
) {
8874 period
= d
->rt_period
;
8875 runtime
= d
->rt_runtime
;
8878 sum
+= to_ratio(period
, runtime
);
8887 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8889 struct rt_schedulable_data data
= {
8891 .rt_period
= period
,
8892 .rt_runtime
= runtime
,
8895 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
8898 static int tg_set_bandwidth(struct task_group
*tg
,
8899 u64 rt_period
, u64 rt_runtime
)
8903 mutex_lock(&rt_constraints_mutex
);
8904 read_lock(&tasklist_lock
);
8905 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
8909 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8910 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8911 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8913 for_each_possible_cpu(i
) {
8914 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8916 spin_lock(&rt_rq
->rt_runtime_lock
);
8917 rt_rq
->rt_runtime
= rt_runtime
;
8918 spin_unlock(&rt_rq
->rt_runtime_lock
);
8920 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8922 read_unlock(&tasklist_lock
);
8923 mutex_unlock(&rt_constraints_mutex
);
8928 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8930 u64 rt_runtime
, rt_period
;
8932 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8933 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8934 if (rt_runtime_us
< 0)
8935 rt_runtime
= RUNTIME_INF
;
8937 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8940 long sched_group_rt_runtime(struct task_group
*tg
)
8944 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8947 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8948 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8949 return rt_runtime_us
;
8952 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8954 u64 rt_runtime
, rt_period
;
8956 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8957 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8962 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8965 long sched_group_rt_period(struct task_group
*tg
)
8969 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8970 do_div(rt_period_us
, NSEC_PER_USEC
);
8971 return rt_period_us
;
8974 static int sched_rt_global_constraints(void)
8976 u64 runtime
, period
;
8979 if (sysctl_sched_rt_period
<= 0)
8982 runtime
= global_rt_runtime();
8983 period
= global_rt_period();
8986 * Sanity check on the sysctl variables.
8988 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8991 mutex_lock(&rt_constraints_mutex
);
8992 read_lock(&tasklist_lock
);
8993 ret
= __rt_schedulable(NULL
, 0, 0);
8994 read_unlock(&tasklist_lock
);
8995 mutex_unlock(&rt_constraints_mutex
);
8999 #else /* !CONFIG_RT_GROUP_SCHED */
9000 static int sched_rt_global_constraints(void)
9002 unsigned long flags
;
9005 if (sysctl_sched_rt_period
<= 0)
9008 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9009 for_each_possible_cpu(i
) {
9010 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
9012 spin_lock(&rt_rq
->rt_runtime_lock
);
9013 rt_rq
->rt_runtime
= global_rt_runtime();
9014 spin_unlock(&rt_rq
->rt_runtime_lock
);
9016 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9020 #endif /* CONFIG_RT_GROUP_SCHED */
9022 int sched_rt_handler(struct ctl_table
*table
, int write
,
9023 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
9027 int old_period
, old_runtime
;
9028 static DEFINE_MUTEX(mutex
);
9031 old_period
= sysctl_sched_rt_period
;
9032 old_runtime
= sysctl_sched_rt_runtime
;
9034 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
9036 if (!ret
&& write
) {
9037 ret
= sched_rt_global_constraints();
9039 sysctl_sched_rt_period
= old_period
;
9040 sysctl_sched_rt_runtime
= old_runtime
;
9042 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
9043 def_rt_bandwidth
.rt_period
=
9044 ns_to_ktime(global_rt_period());
9047 mutex_unlock(&mutex
);
9052 #ifdef CONFIG_CGROUP_SCHED
9054 /* return corresponding task_group object of a cgroup */
9055 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
9057 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
9058 struct task_group
, css
);
9061 static struct cgroup_subsys_state
*
9062 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9064 struct task_group
*tg
, *parent
;
9066 if (!cgrp
->parent
) {
9067 /* This is early initialization for the top cgroup */
9068 return &init_task_group
.css
;
9071 parent
= cgroup_tg(cgrp
->parent
);
9072 tg
= sched_create_group(parent
);
9074 return ERR_PTR(-ENOMEM
);
9080 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9082 struct task_group
*tg
= cgroup_tg(cgrp
);
9084 sched_destroy_group(tg
);
9088 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9089 struct task_struct
*tsk
)
9091 #ifdef CONFIG_RT_GROUP_SCHED
9092 /* Don't accept realtime tasks when there is no way for them to run */
9093 if (rt_task(tsk
) && cgroup_tg(cgrp
)->rt_bandwidth
.rt_runtime
== 0)
9096 /* We don't support RT-tasks being in separate groups */
9097 if (tsk
->sched_class
!= &fair_sched_class
)
9105 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9106 struct cgroup
*old_cont
, struct task_struct
*tsk
)
9108 sched_move_task(tsk
);
9111 #ifdef CONFIG_FAIR_GROUP_SCHED
9112 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
9115 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
9118 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
9120 struct task_group
*tg
= cgroup_tg(cgrp
);
9122 return (u64
) tg
->shares
;
9124 #endif /* CONFIG_FAIR_GROUP_SCHED */
9126 #ifdef CONFIG_RT_GROUP_SCHED
9127 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
9130 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
9133 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9135 return sched_group_rt_runtime(cgroup_tg(cgrp
));
9138 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
9141 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
9144 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
9146 return sched_group_rt_period(cgroup_tg(cgrp
));
9148 #endif /* CONFIG_RT_GROUP_SCHED */
9150 static struct cftype cpu_files
[] = {
9151 #ifdef CONFIG_FAIR_GROUP_SCHED
9154 .read_u64
= cpu_shares_read_u64
,
9155 .write_u64
= cpu_shares_write_u64
,
9158 #ifdef CONFIG_RT_GROUP_SCHED
9160 .name
= "rt_runtime_us",
9161 .read_s64
= cpu_rt_runtime_read
,
9162 .write_s64
= cpu_rt_runtime_write
,
9165 .name
= "rt_period_us",
9166 .read_u64
= cpu_rt_period_read_uint
,
9167 .write_u64
= cpu_rt_period_write_uint
,
9172 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
9174 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
9177 struct cgroup_subsys cpu_cgroup_subsys
= {
9179 .create
= cpu_cgroup_create
,
9180 .destroy
= cpu_cgroup_destroy
,
9181 .can_attach
= cpu_cgroup_can_attach
,
9182 .attach
= cpu_cgroup_attach
,
9183 .populate
= cpu_cgroup_populate
,
9184 .subsys_id
= cpu_cgroup_subsys_id
,
9188 #endif /* CONFIG_CGROUP_SCHED */
9190 #ifdef CONFIG_CGROUP_CPUACCT
9193 * CPU accounting code for task groups.
9195 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9196 * (balbir@in.ibm.com).
9199 /* track cpu usage of a group of tasks and its child groups */
9201 struct cgroup_subsys_state css
;
9202 /* cpuusage holds pointer to a u64-type object on every cpu */
9204 struct cpuacct
*parent
;
9207 struct cgroup_subsys cpuacct_subsys
;
9209 /* return cpu accounting group corresponding to this container */
9210 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
9212 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
9213 struct cpuacct
, css
);
9216 /* return cpu accounting group to which this task belongs */
9217 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
9219 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
9220 struct cpuacct
, css
);
9223 /* create a new cpu accounting group */
9224 static struct cgroup_subsys_state
*cpuacct_create(
9225 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9227 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
9230 return ERR_PTR(-ENOMEM
);
9232 ca
->cpuusage
= alloc_percpu(u64
);
9233 if (!ca
->cpuusage
) {
9235 return ERR_PTR(-ENOMEM
);
9239 ca
->parent
= cgroup_ca(cgrp
->parent
);
9244 /* destroy an existing cpu accounting group */
9246 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9248 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9250 free_percpu(ca
->cpuusage
);
9254 /* return total cpu usage (in nanoseconds) of a group */
9255 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9257 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9258 u64 totalcpuusage
= 0;
9261 for_each_possible_cpu(i
) {
9262 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
9265 * Take rq->lock to make 64-bit addition safe on 32-bit
9268 spin_lock_irq(&cpu_rq(i
)->lock
);
9269 totalcpuusage
+= *cpuusage
;
9270 spin_unlock_irq(&cpu_rq(i
)->lock
);
9273 return totalcpuusage
;
9276 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
9279 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9288 for_each_possible_cpu(i
) {
9289 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
9291 spin_lock_irq(&cpu_rq(i
)->lock
);
9293 spin_unlock_irq(&cpu_rq(i
)->lock
);
9299 static struct cftype files
[] = {
9302 .read_u64
= cpuusage_read
,
9303 .write_u64
= cpuusage_write
,
9307 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9309 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
9313 * charge this task's execution time to its accounting group.
9315 * called with rq->lock held.
9317 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
9322 if (!cpuacct_subsys
.active
)
9325 cpu
= task_cpu(tsk
);
9328 for (; ca
; ca
= ca
->parent
) {
9329 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, cpu
);
9330 *cpuusage
+= cputime
;
9334 struct cgroup_subsys cpuacct_subsys
= {
9336 .create
= cpuacct_create
,
9337 .destroy
= cpuacct_destroy
,
9338 .populate
= cpuacct_populate
,
9339 .subsys_id
= cpuacct_subsys_id
,
9341 #endif /* CONFIG_CGROUP_CPUACCT */