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_USER_SCHED
268 #ifdef CONFIG_FAIR_GROUP_SCHED
269 /* schedulable entities of this group on each cpu */
270 struct sched_entity
**se
;
271 /* runqueue "owned" by this group on each cpu */
272 struct cfs_rq
**cfs_rq
;
273 unsigned long shares
;
276 #ifdef CONFIG_RT_GROUP_SCHED
277 struct sched_rt_entity
**rt_se
;
278 struct rt_rq
**rt_rq
;
280 struct rt_bandwidth rt_bandwidth
;
284 struct list_head list
;
286 struct task_group
*parent
;
287 struct list_head siblings
;
288 struct list_head children
;
291 #ifdef CONFIG_USER_SCHED
293 /* Helper function to pass uid information to create_sched_user() */
294 void set_tg_uid(struct user_struct
*user
)
296 user
->tg
->uid
= user
->uid
;
301 * Every UID task group (including init_task_group aka UID-0) will
302 * be a child to this group.
304 struct task_group root_task_group
;
306 #ifdef CONFIG_FAIR_GROUP_SCHED
307 /* Default task group's sched entity on each cpu */
308 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
309 /* Default task group's cfs_rq on each cpu */
310 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
311 #endif /* CONFIG_FAIR_GROUP_SCHED */
313 #ifdef CONFIG_RT_GROUP_SCHED
314 static DEFINE_PER_CPU(struct sched_rt_entity
, init_sched_rt_entity
);
315 static DEFINE_PER_CPU(struct rt_rq
, init_rt_rq
) ____cacheline_aligned_in_smp
;
316 #endif /* CONFIG_RT_GROUP_SCHED */
317 #else /* !CONFIG_USER_SCHED */
318 #define root_task_group init_task_group
319 #endif /* CONFIG_USER_SCHED */
321 /* task_group_lock serializes add/remove of task groups and also changes to
322 * a task group's cpu shares.
324 static DEFINE_SPINLOCK(task_group_lock
);
326 #ifdef CONFIG_FAIR_GROUP_SCHED
327 #ifdef CONFIG_USER_SCHED
328 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
329 #else /* !CONFIG_USER_SCHED */
330 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
331 #endif /* CONFIG_USER_SCHED */
334 * A weight of 0 or 1 can cause arithmetics problems.
335 * A weight of a cfs_rq is the sum of weights of which entities
336 * are queued on this cfs_rq, so a weight of a entity should not be
337 * too large, so as the shares value of a task group.
338 * (The default weight is 1024 - so there's no practical
339 * limitation from this.)
342 #define MAX_SHARES (1UL << 18)
344 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
347 /* Default task group.
348 * Every task in system belong to this group at bootup.
350 struct task_group init_task_group
;
352 /* return group to which a task belongs */
353 static inline struct task_group
*task_group(struct task_struct
*p
)
355 struct task_group
*tg
;
357 #ifdef CONFIG_USER_SCHED
359 #elif defined(CONFIG_CGROUP_SCHED)
360 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
361 struct task_group
, css
);
363 tg
= &init_task_group
;
368 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
369 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
371 #ifdef CONFIG_FAIR_GROUP_SCHED
372 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
373 p
->se
.parent
= task_group(p
)->se
[cpu
];
376 #ifdef CONFIG_RT_GROUP_SCHED
377 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
378 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
384 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
385 static inline struct task_group
*task_group(struct task_struct
*p
)
390 #endif /* CONFIG_GROUP_SCHED */
392 /* CFS-related fields in a runqueue */
394 struct load_weight load
;
395 unsigned long nr_running
;
400 struct rb_root tasks_timeline
;
401 struct rb_node
*rb_leftmost
;
403 struct list_head tasks
;
404 struct list_head
*balance_iterator
;
407 * 'curr' points to currently running entity on this cfs_rq.
408 * It is set to NULL otherwise (i.e when none are currently running).
410 struct sched_entity
*curr
, *next
, *last
;
412 unsigned int nr_spread_over
;
414 #ifdef CONFIG_FAIR_GROUP_SCHED
415 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
418 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
419 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
420 * (like users, containers etc.)
422 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
423 * list is used during load balance.
425 struct list_head leaf_cfs_rq_list
;
426 struct task_group
*tg
; /* group that "owns" this runqueue */
430 * the part of load.weight contributed by tasks
432 unsigned long task_weight
;
435 * h_load = weight * f(tg)
437 * Where f(tg) is the recursive weight fraction assigned to
440 unsigned long h_load
;
443 * this cpu's part of tg->shares
445 unsigned long shares
;
448 * load.weight at the time we set shares
450 unsigned long rq_weight
;
455 /* Real-Time classes' related field in a runqueue: */
457 struct rt_prio_array active
;
458 unsigned long rt_nr_running
;
459 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
460 int highest_prio
; /* highest queued rt task prio */
463 unsigned long rt_nr_migratory
;
469 /* Nests inside the rq lock: */
470 spinlock_t rt_runtime_lock
;
472 #ifdef CONFIG_RT_GROUP_SCHED
473 unsigned long rt_nr_boosted
;
476 struct list_head leaf_rt_rq_list
;
477 struct task_group
*tg
;
478 struct sched_rt_entity
*rt_se
;
485 * We add the notion of a root-domain which will be used to define per-domain
486 * variables. Each exclusive cpuset essentially defines an island domain by
487 * fully partitioning the member cpus from any other cpuset. Whenever a new
488 * exclusive cpuset is created, we also create and attach a new root-domain
498 * The "RT overload" flag: it gets set if a CPU has more than
499 * one runnable RT task.
504 struct cpupri cpupri
;
509 * By default the system creates a single root-domain with all cpus as
510 * members (mimicking the global state we have today).
512 static struct root_domain def_root_domain
;
517 * This is the main, per-CPU runqueue data structure.
519 * Locking rule: those places that want to lock multiple runqueues
520 * (such as the load balancing or the thread migration code), lock
521 * acquire operations must be ordered by ascending &runqueue.
528 * nr_running and cpu_load should be in the same cacheline because
529 * remote CPUs use both these fields when doing load calculation.
531 unsigned long nr_running
;
532 #define CPU_LOAD_IDX_MAX 5
533 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
534 unsigned char idle_at_tick
;
536 unsigned long last_tick_seen
;
537 unsigned char in_nohz_recently
;
539 /* capture load from *all* tasks on this cpu: */
540 struct load_weight load
;
541 unsigned long nr_load_updates
;
547 #ifdef CONFIG_FAIR_GROUP_SCHED
548 /* list of leaf cfs_rq on this cpu: */
549 struct list_head leaf_cfs_rq_list
;
551 #ifdef CONFIG_RT_GROUP_SCHED
552 struct list_head leaf_rt_rq_list
;
556 * This is part of a global counter where only the total sum
557 * over all CPUs matters. A task can increase this counter on
558 * one CPU and if it got migrated afterwards it may decrease
559 * it on another CPU. Always updated under the runqueue lock:
561 unsigned long nr_uninterruptible
;
563 struct task_struct
*curr
, *idle
;
564 unsigned long next_balance
;
565 struct mm_struct
*prev_mm
;
572 struct root_domain
*rd
;
573 struct sched_domain
*sd
;
575 /* For active balancing */
578 /* cpu of this runqueue: */
582 unsigned long avg_load_per_task
;
584 struct task_struct
*migration_thread
;
585 struct list_head migration_queue
;
588 #ifdef CONFIG_SCHED_HRTICK
590 int hrtick_csd_pending
;
591 struct call_single_data hrtick_csd
;
593 struct hrtimer hrtick_timer
;
596 #ifdef CONFIG_SCHEDSTATS
598 struct sched_info rq_sched_info
;
600 /* sys_sched_yield() stats */
601 unsigned int yld_exp_empty
;
602 unsigned int yld_act_empty
;
603 unsigned int yld_both_empty
;
604 unsigned int yld_count
;
606 /* schedule() stats */
607 unsigned int sched_switch
;
608 unsigned int sched_count
;
609 unsigned int sched_goidle
;
611 /* try_to_wake_up() stats */
612 unsigned int ttwu_count
;
613 unsigned int ttwu_local
;
616 unsigned int bkl_count
;
620 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
622 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int sync
)
624 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, sync
);
627 static inline int cpu_of(struct rq
*rq
)
637 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
638 * See detach_destroy_domains: synchronize_sched for details.
640 * The domain tree of any CPU may only be accessed from within
641 * preempt-disabled sections.
643 #define for_each_domain(cpu, __sd) \
644 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
646 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
647 #define this_rq() (&__get_cpu_var(runqueues))
648 #define task_rq(p) cpu_rq(task_cpu(p))
649 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
651 static inline void update_rq_clock(struct rq
*rq
)
653 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
657 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
659 #ifdef CONFIG_SCHED_DEBUG
660 # define const_debug __read_mostly
662 # define const_debug static const
668 * Returns true if the current cpu runqueue is locked.
669 * This interface allows printk to be called with the runqueue lock
670 * held and know whether or not it is OK to wake up the klogd.
672 int runqueue_is_locked(void)
675 struct rq
*rq
= cpu_rq(cpu
);
678 ret
= spin_is_locked(&rq
->lock
);
684 * Debugging: various feature bits
687 #define SCHED_FEAT(name, enabled) \
688 __SCHED_FEAT_##name ,
691 #include "sched_features.h"
696 #define SCHED_FEAT(name, enabled) \
697 (1UL << __SCHED_FEAT_##name) * enabled |
699 const_debug
unsigned int sysctl_sched_features
=
700 #include "sched_features.h"
705 #ifdef CONFIG_SCHED_DEBUG
706 #define SCHED_FEAT(name, enabled) \
709 static __read_mostly
char *sched_feat_names
[] = {
710 #include "sched_features.h"
716 static int sched_feat_show(struct seq_file
*m
, void *v
)
720 for (i
= 0; sched_feat_names
[i
]; i
++) {
721 if (!(sysctl_sched_features
& (1UL << i
)))
723 seq_printf(m
, "%s ", sched_feat_names
[i
]);
731 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
732 size_t cnt
, loff_t
*ppos
)
742 if (copy_from_user(&buf
, ubuf
, cnt
))
747 if (strncmp(buf
, "NO_", 3) == 0) {
752 for (i
= 0; sched_feat_names
[i
]; i
++) {
753 int len
= strlen(sched_feat_names
[i
]);
755 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
757 sysctl_sched_features
&= ~(1UL << i
);
759 sysctl_sched_features
|= (1UL << i
);
764 if (!sched_feat_names
[i
])
772 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
774 return single_open(filp
, sched_feat_show
, NULL
);
777 static struct file_operations sched_feat_fops
= {
778 .open
= sched_feat_open
,
779 .write
= sched_feat_write
,
782 .release
= single_release
,
785 static __init
int sched_init_debug(void)
787 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
792 late_initcall(sched_init_debug
);
796 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
799 * Number of tasks to iterate in a single balance run.
800 * Limited because this is done with IRQs disabled.
802 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
805 * ratelimit for updating the group shares.
808 unsigned int sysctl_sched_shares_ratelimit
= 250000;
811 * Inject some fuzzyness into changing the per-cpu group shares
812 * this avoids remote rq-locks at the expense of fairness.
815 unsigned int sysctl_sched_shares_thresh
= 4;
818 * period over which we measure -rt task cpu usage in us.
821 unsigned int sysctl_sched_rt_period
= 1000000;
823 static __read_mostly
int scheduler_running
;
826 * part of the period that we allow rt tasks to run in us.
829 int sysctl_sched_rt_runtime
= 950000;
831 static inline u64
global_rt_period(void)
833 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
836 static inline u64
global_rt_runtime(void)
838 if (sysctl_sched_rt_runtime
< 0)
841 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
844 #ifndef prepare_arch_switch
845 # define prepare_arch_switch(next) do { } while (0)
847 #ifndef finish_arch_switch
848 # define finish_arch_switch(prev) do { } while (0)
851 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
853 return rq
->curr
== p
;
856 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
857 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
859 return task_current(rq
, p
);
862 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
866 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
868 #ifdef CONFIG_DEBUG_SPINLOCK
869 /* this is a valid case when another task releases the spinlock */
870 rq
->lock
.owner
= current
;
873 * If we are tracking spinlock dependencies then we have to
874 * fix up the runqueue lock - which gets 'carried over' from
877 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
879 spin_unlock_irq(&rq
->lock
);
882 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
883 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
888 return task_current(rq
, p
);
892 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
896 * We can optimise this out completely for !SMP, because the
897 * SMP rebalancing from interrupt is the only thing that cares
902 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
903 spin_unlock_irq(&rq
->lock
);
905 spin_unlock(&rq
->lock
);
909 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
913 * After ->oncpu is cleared, the task can be moved to a different CPU.
914 * We must ensure this doesn't happen until the switch is completely
920 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
924 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
927 * __task_rq_lock - lock the runqueue a given task resides on.
928 * Must be called interrupts disabled.
930 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
934 struct rq
*rq
= task_rq(p
);
935 spin_lock(&rq
->lock
);
936 if (likely(rq
== task_rq(p
)))
938 spin_unlock(&rq
->lock
);
943 * task_rq_lock - lock the runqueue a given task resides on and disable
944 * interrupts. Note the ordering: we can safely lookup the task_rq without
945 * explicitly disabling preemption.
947 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
953 local_irq_save(*flags
);
955 spin_lock(&rq
->lock
);
956 if (likely(rq
== task_rq(p
)))
958 spin_unlock_irqrestore(&rq
->lock
, *flags
);
962 void task_rq_unlock_wait(struct task_struct
*p
)
964 struct rq
*rq
= task_rq(p
);
966 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
967 spin_unlock_wait(&rq
->lock
);
970 static void __task_rq_unlock(struct rq
*rq
)
973 spin_unlock(&rq
->lock
);
976 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
979 spin_unlock_irqrestore(&rq
->lock
, *flags
);
983 * this_rq_lock - lock this runqueue and disable interrupts.
985 static struct rq
*this_rq_lock(void)
992 spin_lock(&rq
->lock
);
997 #ifdef CONFIG_SCHED_HRTICK
999 * Use HR-timers to deliver accurate preemption points.
1001 * Its all a bit involved since we cannot program an hrt while holding the
1002 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1005 * When we get rescheduled we reprogram the hrtick_timer outside of the
1011 * - enabled by features
1012 * - hrtimer is actually high res
1014 static inline int hrtick_enabled(struct rq
*rq
)
1016 if (!sched_feat(HRTICK
))
1018 if (!cpu_active(cpu_of(rq
)))
1020 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1023 static void hrtick_clear(struct rq
*rq
)
1025 if (hrtimer_active(&rq
->hrtick_timer
))
1026 hrtimer_cancel(&rq
->hrtick_timer
);
1030 * High-resolution timer tick.
1031 * Runs from hardirq context with interrupts disabled.
1033 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1035 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1037 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1039 spin_lock(&rq
->lock
);
1040 update_rq_clock(rq
);
1041 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1042 spin_unlock(&rq
->lock
);
1044 return HRTIMER_NORESTART
;
1049 * called from hardirq (IPI) context
1051 static void __hrtick_start(void *arg
)
1053 struct rq
*rq
= arg
;
1055 spin_lock(&rq
->lock
);
1056 hrtimer_restart(&rq
->hrtick_timer
);
1057 rq
->hrtick_csd_pending
= 0;
1058 spin_unlock(&rq
->lock
);
1062 * Called to set the hrtick timer state.
1064 * called with rq->lock held and irqs disabled
1066 static void hrtick_start(struct rq
*rq
, u64 delay
)
1068 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1069 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1071 hrtimer_set_expires(timer
, time
);
1073 if (rq
== this_rq()) {
1074 hrtimer_restart(timer
);
1075 } else if (!rq
->hrtick_csd_pending
) {
1076 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
);
1077 rq
->hrtick_csd_pending
= 1;
1082 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1084 int cpu
= (int)(long)hcpu
;
1087 case CPU_UP_CANCELED
:
1088 case CPU_UP_CANCELED_FROZEN
:
1089 case CPU_DOWN_PREPARE
:
1090 case CPU_DOWN_PREPARE_FROZEN
:
1092 case CPU_DEAD_FROZEN
:
1093 hrtick_clear(cpu_rq(cpu
));
1100 static __init
void init_hrtick(void)
1102 hotcpu_notifier(hotplug_hrtick
, 0);
1106 * Called to set the hrtick timer state.
1108 * called with rq->lock held and irqs disabled
1110 static void hrtick_start(struct rq
*rq
, u64 delay
)
1112 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
), HRTIMER_MODE_REL
);
1115 static inline void init_hrtick(void)
1118 #endif /* CONFIG_SMP */
1120 static void init_rq_hrtick(struct rq
*rq
)
1123 rq
->hrtick_csd_pending
= 0;
1125 rq
->hrtick_csd
.flags
= 0;
1126 rq
->hrtick_csd
.func
= __hrtick_start
;
1127 rq
->hrtick_csd
.info
= rq
;
1130 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1131 rq
->hrtick_timer
.function
= hrtick
;
1132 rq
->hrtick_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_PERCPU
;
1134 #else /* CONFIG_SCHED_HRTICK */
1135 static inline void hrtick_clear(struct rq
*rq
)
1139 static inline void init_rq_hrtick(struct rq
*rq
)
1143 static inline void init_hrtick(void)
1146 #endif /* CONFIG_SCHED_HRTICK */
1149 * resched_task - mark a task 'to be rescheduled now'.
1151 * On UP this means the setting of the need_resched flag, on SMP it
1152 * might also involve a cross-CPU call to trigger the scheduler on
1157 #ifndef tsk_is_polling
1158 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1161 static void resched_task(struct task_struct
*p
)
1165 assert_spin_locked(&task_rq(p
)->lock
);
1167 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
1170 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
1173 if (cpu
== smp_processor_id())
1176 /* NEED_RESCHED must be visible before we test polling */
1178 if (!tsk_is_polling(p
))
1179 smp_send_reschedule(cpu
);
1182 static void resched_cpu(int cpu
)
1184 struct rq
*rq
= cpu_rq(cpu
);
1185 unsigned long flags
;
1187 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1189 resched_task(cpu_curr(cpu
));
1190 spin_unlock_irqrestore(&rq
->lock
, flags
);
1195 * When add_timer_on() enqueues a timer into the timer wheel of an
1196 * idle CPU then this timer might expire before the next timer event
1197 * which is scheduled to wake up that CPU. In case of a completely
1198 * idle system the next event might even be infinite time into the
1199 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1200 * leaves the inner idle loop so the newly added timer is taken into
1201 * account when the CPU goes back to idle and evaluates the timer
1202 * wheel for the next timer event.
1204 void wake_up_idle_cpu(int cpu
)
1206 struct rq
*rq
= cpu_rq(cpu
);
1208 if (cpu
== smp_processor_id())
1212 * This is safe, as this function is called with the timer
1213 * wheel base lock of (cpu) held. When the CPU is on the way
1214 * to idle and has not yet set rq->curr to idle then it will
1215 * be serialized on the timer wheel base lock and take the new
1216 * timer into account automatically.
1218 if (rq
->curr
!= rq
->idle
)
1222 * We can set TIF_RESCHED on the idle task of the other CPU
1223 * lockless. The worst case is that the other CPU runs the
1224 * idle task through an additional NOOP schedule()
1226 set_tsk_thread_flag(rq
->idle
, TIF_NEED_RESCHED
);
1228 /* NEED_RESCHED must be visible before we test polling */
1230 if (!tsk_is_polling(rq
->idle
))
1231 smp_send_reschedule(cpu
);
1233 #endif /* CONFIG_NO_HZ */
1235 #else /* !CONFIG_SMP */
1236 static void resched_task(struct task_struct
*p
)
1238 assert_spin_locked(&task_rq(p
)->lock
);
1239 set_tsk_need_resched(p
);
1241 #endif /* CONFIG_SMP */
1243 #if BITS_PER_LONG == 32
1244 # define WMULT_CONST (~0UL)
1246 # define WMULT_CONST (1UL << 32)
1249 #define WMULT_SHIFT 32
1252 * Shift right and round:
1254 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1257 * delta *= weight / lw
1259 static unsigned long
1260 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1261 struct load_weight
*lw
)
1265 if (!lw
->inv_weight
) {
1266 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1269 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1273 tmp
= (u64
)delta_exec
* weight
;
1275 * Check whether we'd overflow the 64-bit multiplication:
1277 if (unlikely(tmp
> WMULT_CONST
))
1278 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1281 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1283 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1286 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1292 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1299 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1300 * of tasks with abnormal "nice" values across CPUs the contribution that
1301 * each task makes to its run queue's load is weighted according to its
1302 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1303 * scaled version of the new time slice allocation that they receive on time
1307 #define WEIGHT_IDLEPRIO 2
1308 #define WMULT_IDLEPRIO (1 << 31)
1311 * Nice levels are multiplicative, with a gentle 10% change for every
1312 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1313 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1314 * that remained on nice 0.
1316 * The "10% effect" is relative and cumulative: from _any_ nice level,
1317 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1318 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1319 * If a task goes up by ~10% and another task goes down by ~10% then
1320 * the relative distance between them is ~25%.)
1322 static const int prio_to_weight
[40] = {
1323 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1324 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1325 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1326 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1327 /* 0 */ 1024, 820, 655, 526, 423,
1328 /* 5 */ 335, 272, 215, 172, 137,
1329 /* 10 */ 110, 87, 70, 56, 45,
1330 /* 15 */ 36, 29, 23, 18, 15,
1334 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1336 * In cases where the weight does not change often, we can use the
1337 * precalculated inverse to speed up arithmetics by turning divisions
1338 * into multiplications:
1340 static const u32 prio_to_wmult
[40] = {
1341 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1342 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1343 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1344 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1345 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1346 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1347 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1348 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1351 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1354 * runqueue iterator, to support SMP load-balancing between different
1355 * scheduling classes, without having to expose their internal data
1356 * structures to the load-balancing proper:
1358 struct rq_iterator
{
1360 struct task_struct
*(*start
)(void *);
1361 struct task_struct
*(*next
)(void *);
1365 static unsigned long
1366 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1367 unsigned long max_load_move
, struct sched_domain
*sd
,
1368 enum cpu_idle_type idle
, int *all_pinned
,
1369 int *this_best_prio
, struct rq_iterator
*iterator
);
1372 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1373 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1374 struct rq_iterator
*iterator
);
1377 #ifdef CONFIG_CGROUP_CPUACCT
1378 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1380 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1383 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1385 update_load_add(&rq
->load
, load
);
1388 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1390 update_load_sub(&rq
->load
, load
);
1393 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1394 typedef int (*tg_visitor
)(struct task_group
*, void *);
1397 * Iterate the full tree, calling @down when first entering a node and @up when
1398 * leaving it for the final time.
1400 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1402 struct task_group
*parent
, *child
;
1406 parent
= &root_task_group
;
1408 ret
= (*down
)(parent
, data
);
1411 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1418 ret
= (*up
)(parent
, data
);
1423 parent
= parent
->parent
;
1432 static int tg_nop(struct task_group
*tg
, void *data
)
1439 static unsigned long source_load(int cpu
, int type
);
1440 static unsigned long target_load(int cpu
, int type
);
1441 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1443 static unsigned long cpu_avg_load_per_task(int cpu
)
1445 struct rq
*rq
= cpu_rq(cpu
);
1446 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1449 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1451 rq
->avg_load_per_task
= 0;
1453 return rq
->avg_load_per_task
;
1456 #ifdef CONFIG_FAIR_GROUP_SCHED
1458 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1461 * Calculate and set the cpu's group shares.
1464 update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1465 unsigned long sd_shares
, unsigned long sd_rq_weight
)
1467 unsigned long shares
;
1468 unsigned long rq_weight
;
1473 rq_weight
= tg
->cfs_rq
[cpu
]->rq_weight
;
1476 * \Sum shares * rq_weight
1477 * shares = -----------------------
1481 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1482 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1484 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1485 sysctl_sched_shares_thresh
) {
1486 struct rq
*rq
= cpu_rq(cpu
);
1487 unsigned long flags
;
1489 spin_lock_irqsave(&rq
->lock
, flags
);
1490 tg
->cfs_rq
[cpu
]->shares
= shares
;
1492 __set_se_shares(tg
->se
[cpu
], shares
);
1493 spin_unlock_irqrestore(&rq
->lock
, flags
);
1498 * Re-compute the task group their per cpu shares over the given domain.
1499 * This needs to be done in a bottom-up fashion because the rq weight of a
1500 * parent group depends on the shares of its child groups.
1502 static int tg_shares_up(struct task_group
*tg
, void *data
)
1504 unsigned long weight
, rq_weight
= 0;
1505 unsigned long shares
= 0;
1506 struct sched_domain
*sd
= data
;
1509 for_each_cpu_mask(i
, sd
->span
) {
1511 * If there are currently no tasks on the cpu pretend there
1512 * is one of average load so that when a new task gets to
1513 * run here it will not get delayed by group starvation.
1515 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1517 weight
= NICE_0_LOAD
;
1519 tg
->cfs_rq
[i
]->rq_weight
= weight
;
1520 rq_weight
+= weight
;
1521 shares
+= tg
->cfs_rq
[i
]->shares
;
1524 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1525 shares
= tg
->shares
;
1527 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1528 shares
= tg
->shares
;
1530 for_each_cpu_mask(i
, sd
->span
)
1531 update_group_shares_cpu(tg
, i
, shares
, rq_weight
);
1537 * Compute the cpu's hierarchical load factor for each task group.
1538 * This needs to be done in a top-down fashion because the load of a child
1539 * group is a fraction of its parents load.
1541 static int tg_load_down(struct task_group
*tg
, void *data
)
1544 long cpu
= (long)data
;
1547 load
= cpu_rq(cpu
)->load
.weight
;
1549 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1550 load
*= tg
->cfs_rq
[cpu
]->shares
;
1551 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1554 tg
->cfs_rq
[cpu
]->h_load
= load
;
1559 static void update_shares(struct sched_domain
*sd
)
1561 u64 now
= cpu_clock(raw_smp_processor_id());
1562 s64 elapsed
= now
- sd
->last_update
;
1564 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1565 sd
->last_update
= now
;
1566 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1570 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1572 spin_unlock(&rq
->lock
);
1574 spin_lock(&rq
->lock
);
1577 static void update_h_load(long cpu
)
1579 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1584 static inline void update_shares(struct sched_domain
*sd
)
1588 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1595 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1597 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1598 __releases(this_rq
->lock
)
1599 __acquires(busiest
->lock
)
1600 __acquires(this_rq
->lock
)
1604 if (unlikely(!irqs_disabled())) {
1605 /* printk() doesn't work good under rq->lock */
1606 spin_unlock(&this_rq
->lock
);
1609 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1610 if (busiest
< this_rq
) {
1611 spin_unlock(&this_rq
->lock
);
1612 spin_lock(&busiest
->lock
);
1613 spin_lock_nested(&this_rq
->lock
, SINGLE_DEPTH_NESTING
);
1616 spin_lock_nested(&busiest
->lock
, SINGLE_DEPTH_NESTING
);
1621 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1622 __releases(busiest
->lock
)
1624 spin_unlock(&busiest
->lock
);
1625 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1629 #ifdef CONFIG_FAIR_GROUP_SCHED
1630 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1633 cfs_rq
->shares
= shares
;
1638 #include "sched_stats.h"
1639 #include "sched_idletask.c"
1640 #include "sched_fair.c"
1641 #include "sched_rt.c"
1642 #ifdef CONFIG_SCHED_DEBUG
1643 # include "sched_debug.c"
1646 #define sched_class_highest (&rt_sched_class)
1647 #define for_each_class(class) \
1648 for (class = sched_class_highest; class; class = class->next)
1650 static void inc_nr_running(struct rq
*rq
)
1655 static void dec_nr_running(struct rq
*rq
)
1660 static void set_load_weight(struct task_struct
*p
)
1662 if (task_has_rt_policy(p
)) {
1663 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1664 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1669 * SCHED_IDLE tasks get minimal weight:
1671 if (p
->policy
== SCHED_IDLE
) {
1672 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1673 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1677 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1678 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1681 static void update_avg(u64
*avg
, u64 sample
)
1683 s64 diff
= sample
- *avg
;
1687 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1689 sched_info_queued(p
);
1690 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1694 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1696 if (sleep
&& p
->se
.last_wakeup
) {
1697 update_avg(&p
->se
.avg_overlap
,
1698 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1699 p
->se
.last_wakeup
= 0;
1702 sched_info_dequeued(p
);
1703 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1708 * __normal_prio - return the priority that is based on the static prio
1710 static inline int __normal_prio(struct task_struct
*p
)
1712 return p
->static_prio
;
1716 * Calculate the expected normal priority: i.e. priority
1717 * without taking RT-inheritance into account. Might be
1718 * boosted by interactivity modifiers. Changes upon fork,
1719 * setprio syscalls, and whenever the interactivity
1720 * estimator recalculates.
1722 static inline int normal_prio(struct task_struct
*p
)
1726 if (task_has_rt_policy(p
))
1727 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1729 prio
= __normal_prio(p
);
1734 * Calculate the current priority, i.e. the priority
1735 * taken into account by the scheduler. This value might
1736 * be boosted by RT tasks, or might be boosted by
1737 * interactivity modifiers. Will be RT if the task got
1738 * RT-boosted. If not then it returns p->normal_prio.
1740 static int effective_prio(struct task_struct
*p
)
1742 p
->normal_prio
= normal_prio(p
);
1744 * If we are RT tasks or we were boosted to RT priority,
1745 * keep the priority unchanged. Otherwise, update priority
1746 * to the normal priority:
1748 if (!rt_prio(p
->prio
))
1749 return p
->normal_prio
;
1754 * activate_task - move a task to the runqueue.
1756 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1758 if (task_contributes_to_load(p
))
1759 rq
->nr_uninterruptible
--;
1761 enqueue_task(rq
, p
, wakeup
);
1766 * deactivate_task - remove a task from the runqueue.
1768 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1770 if (task_contributes_to_load(p
))
1771 rq
->nr_uninterruptible
++;
1773 dequeue_task(rq
, p
, sleep
);
1778 * task_curr - is this task currently executing on a CPU?
1779 * @p: the task in question.
1781 inline int task_curr(const struct task_struct
*p
)
1783 return cpu_curr(task_cpu(p
)) == p
;
1786 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1788 set_task_rq(p
, cpu
);
1791 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1792 * successfuly executed on another CPU. We must ensure that updates of
1793 * per-task data have been completed by this moment.
1796 task_thread_info(p
)->cpu
= cpu
;
1800 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1801 const struct sched_class
*prev_class
,
1802 int oldprio
, int running
)
1804 if (prev_class
!= p
->sched_class
) {
1805 if (prev_class
->switched_from
)
1806 prev_class
->switched_from(rq
, p
, running
);
1807 p
->sched_class
->switched_to(rq
, p
, running
);
1809 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1814 /* Used instead of source_load when we know the type == 0 */
1815 static unsigned long weighted_cpuload(const int cpu
)
1817 return cpu_rq(cpu
)->load
.weight
;
1821 * Is this task likely cache-hot:
1824 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1829 * Buddy candidates are cache hot:
1831 if (sched_feat(CACHE_HOT_BUDDY
) &&
1832 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
1833 &p
->se
== cfs_rq_of(&p
->se
)->last
))
1836 if (p
->sched_class
!= &fair_sched_class
)
1839 if (sysctl_sched_migration_cost
== -1)
1841 if (sysctl_sched_migration_cost
== 0)
1844 delta
= now
- p
->se
.exec_start
;
1846 return delta
< (s64
)sysctl_sched_migration_cost
;
1850 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1852 int old_cpu
= task_cpu(p
);
1853 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1854 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1855 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1858 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1860 #ifdef CONFIG_SCHEDSTATS
1861 if (p
->se
.wait_start
)
1862 p
->se
.wait_start
-= clock_offset
;
1863 if (p
->se
.sleep_start
)
1864 p
->se
.sleep_start
-= clock_offset
;
1865 if (p
->se
.block_start
)
1866 p
->se
.block_start
-= clock_offset
;
1867 if (old_cpu
!= new_cpu
) {
1868 schedstat_inc(p
, se
.nr_migrations
);
1869 if (task_hot(p
, old_rq
->clock
, NULL
))
1870 schedstat_inc(p
, se
.nr_forced2_migrations
);
1873 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1874 new_cfsrq
->min_vruntime
;
1876 __set_task_cpu(p
, new_cpu
);
1879 struct migration_req
{
1880 struct list_head list
;
1882 struct task_struct
*task
;
1885 struct completion done
;
1889 * The task's runqueue lock must be held.
1890 * Returns true if you have to wait for migration thread.
1893 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1895 struct rq
*rq
= task_rq(p
);
1898 * If the task is not on a runqueue (and not running), then
1899 * it is sufficient to simply update the task's cpu field.
1901 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1902 set_task_cpu(p
, dest_cpu
);
1906 init_completion(&req
->done
);
1908 req
->dest_cpu
= dest_cpu
;
1909 list_add(&req
->list
, &rq
->migration_queue
);
1915 * wait_task_inactive - wait for a thread to unschedule.
1917 * If @match_state is nonzero, it's the @p->state value just checked and
1918 * not expected to change. If it changes, i.e. @p might have woken up,
1919 * then return zero. When we succeed in waiting for @p to be off its CPU,
1920 * we return a positive number (its total switch count). If a second call
1921 * a short while later returns the same number, the caller can be sure that
1922 * @p has remained unscheduled the whole time.
1924 * The caller must ensure that the task *will* unschedule sometime soon,
1925 * else this function might spin for a *long* time. This function can't
1926 * be called with interrupts off, or it may introduce deadlock with
1927 * smp_call_function() if an IPI is sent by the same process we are
1928 * waiting to become inactive.
1930 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1932 unsigned long flags
;
1939 * We do the initial early heuristics without holding
1940 * any task-queue locks at all. We'll only try to get
1941 * the runqueue lock when things look like they will
1947 * If the task is actively running on another CPU
1948 * still, just relax and busy-wait without holding
1951 * NOTE! Since we don't hold any locks, it's not
1952 * even sure that "rq" stays as the right runqueue!
1953 * But we don't care, since "task_running()" will
1954 * return false if the runqueue has changed and p
1955 * is actually now running somewhere else!
1957 while (task_running(rq
, p
)) {
1958 if (match_state
&& unlikely(p
->state
!= match_state
))
1964 * Ok, time to look more closely! We need the rq
1965 * lock now, to be *sure*. If we're wrong, we'll
1966 * just go back and repeat.
1968 rq
= task_rq_lock(p
, &flags
);
1969 trace_sched_wait_task(rq
, p
);
1970 running
= task_running(rq
, p
);
1971 on_rq
= p
->se
.on_rq
;
1973 if (!match_state
|| p
->state
== match_state
)
1974 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1975 task_rq_unlock(rq
, &flags
);
1978 * If it changed from the expected state, bail out now.
1980 if (unlikely(!ncsw
))
1984 * Was it really running after all now that we
1985 * checked with the proper locks actually held?
1987 * Oops. Go back and try again..
1989 if (unlikely(running
)) {
1995 * It's not enough that it's not actively running,
1996 * it must be off the runqueue _entirely_, and not
1999 * So if it wa still runnable (but just not actively
2000 * running right now), it's preempted, and we should
2001 * yield - it could be a while.
2003 if (unlikely(on_rq
)) {
2004 schedule_timeout_uninterruptible(1);
2009 * Ahh, all good. It wasn't running, and it wasn't
2010 * runnable, which means that it will never become
2011 * running in the future either. We're all done!
2020 * kick_process - kick a running thread to enter/exit the kernel
2021 * @p: the to-be-kicked thread
2023 * Cause a process which is running on another CPU to enter
2024 * kernel-mode, without any delay. (to get signals handled.)
2026 * NOTE: this function doesnt have to take the runqueue lock,
2027 * because all it wants to ensure is that the remote task enters
2028 * the kernel. If the IPI races and the task has been migrated
2029 * to another CPU then no harm is done and the purpose has been
2032 void kick_process(struct task_struct
*p
)
2038 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2039 smp_send_reschedule(cpu
);
2044 * Return a low guess at the load of a migration-source cpu weighted
2045 * according to the scheduling class and "nice" value.
2047 * We want to under-estimate the load of migration sources, to
2048 * balance conservatively.
2050 static unsigned long source_load(int cpu
, int type
)
2052 struct rq
*rq
= cpu_rq(cpu
);
2053 unsigned long total
= weighted_cpuload(cpu
);
2055 if (type
== 0 || !sched_feat(LB_BIAS
))
2058 return min(rq
->cpu_load
[type
-1], total
);
2062 * Return a high guess at the load of a migration-target cpu weighted
2063 * according to the scheduling class and "nice" value.
2065 static unsigned long target_load(int cpu
, int type
)
2067 struct rq
*rq
= cpu_rq(cpu
);
2068 unsigned long total
= weighted_cpuload(cpu
);
2070 if (type
== 0 || !sched_feat(LB_BIAS
))
2073 return max(rq
->cpu_load
[type
-1], total
);
2077 * find_idlest_group finds and returns the least busy CPU group within the
2080 static struct sched_group
*
2081 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
2083 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2084 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
2085 int load_idx
= sd
->forkexec_idx
;
2086 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
2089 unsigned long load
, avg_load
;
2093 /* Skip over this group if it has no CPUs allowed */
2094 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
2097 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2099 /* Tally up the load of all CPUs in the group */
2102 for_each_cpu_mask_nr(i
, group
->cpumask
) {
2103 /* Bias balancing toward cpus of our domain */
2105 load
= source_load(i
, load_idx
);
2107 load
= target_load(i
, load_idx
);
2112 /* Adjust by relative CPU power of the group */
2113 avg_load
= sg_div_cpu_power(group
,
2114 avg_load
* SCHED_LOAD_SCALE
);
2117 this_load
= avg_load
;
2119 } else if (avg_load
< min_load
) {
2120 min_load
= avg_load
;
2123 } while (group
= group
->next
, group
!= sd
->groups
);
2125 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
2131 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2134 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
,
2137 unsigned long load
, min_load
= ULONG_MAX
;
2141 /* Traverse only the allowed CPUs */
2142 cpus_and(*tmp
, group
->cpumask
, p
->cpus_allowed
);
2144 for_each_cpu_mask_nr(i
, *tmp
) {
2145 load
= weighted_cpuload(i
);
2147 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2157 * sched_balance_self: balance the current task (running on cpu) in domains
2158 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2161 * Balance, ie. select the least loaded group.
2163 * Returns the target CPU number, or the same CPU if no balancing is needed.
2165 * preempt must be disabled.
2167 static int sched_balance_self(int cpu
, int flag
)
2169 struct task_struct
*t
= current
;
2170 struct sched_domain
*tmp
, *sd
= NULL
;
2172 for_each_domain(cpu
, tmp
) {
2174 * If power savings logic is enabled for a domain, stop there.
2176 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2178 if (tmp
->flags
& flag
)
2186 cpumask_t span
, tmpmask
;
2187 struct sched_group
*group
;
2188 int new_cpu
, weight
;
2190 if (!(sd
->flags
& flag
)) {
2196 group
= find_idlest_group(sd
, t
, cpu
);
2202 new_cpu
= find_idlest_cpu(group
, t
, cpu
, &tmpmask
);
2203 if (new_cpu
== -1 || new_cpu
== cpu
) {
2204 /* Now try balancing at a lower domain level of cpu */
2209 /* Now try balancing at a lower domain level of new_cpu */
2212 weight
= cpus_weight(span
);
2213 for_each_domain(cpu
, tmp
) {
2214 if (weight
<= cpus_weight(tmp
->span
))
2216 if (tmp
->flags
& flag
)
2219 /* while loop will break here if sd == NULL */
2225 #endif /* CONFIG_SMP */
2228 * try_to_wake_up - wake up a thread
2229 * @p: the to-be-woken-up thread
2230 * @state: the mask of task states that can be woken
2231 * @sync: do a synchronous wakeup?
2233 * Put it on the run-queue if it's not already there. The "current"
2234 * thread is always on the run-queue (except when the actual
2235 * re-schedule is in progress), and as such you're allowed to do
2236 * the simpler "current->state = TASK_RUNNING" to mark yourself
2237 * runnable without the overhead of this.
2239 * returns failure only if the task is already active.
2241 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
2243 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2244 unsigned long flags
;
2248 if (!sched_feat(SYNC_WAKEUPS
))
2252 if (sched_feat(LB_WAKEUP_UPDATE
)) {
2253 struct sched_domain
*sd
;
2255 this_cpu
= raw_smp_processor_id();
2258 for_each_domain(this_cpu
, sd
) {
2259 if (cpu_isset(cpu
, sd
->span
)) {
2268 rq
= task_rq_lock(p
, &flags
);
2269 old_state
= p
->state
;
2270 if (!(old_state
& state
))
2278 this_cpu
= smp_processor_id();
2281 if (unlikely(task_running(rq
, p
)))
2284 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2285 if (cpu
!= orig_cpu
) {
2286 set_task_cpu(p
, cpu
);
2287 task_rq_unlock(rq
, &flags
);
2288 /* might preempt at this point */
2289 rq
= task_rq_lock(p
, &flags
);
2290 old_state
= p
->state
;
2291 if (!(old_state
& state
))
2296 this_cpu
= smp_processor_id();
2300 #ifdef CONFIG_SCHEDSTATS
2301 schedstat_inc(rq
, ttwu_count
);
2302 if (cpu
== this_cpu
)
2303 schedstat_inc(rq
, ttwu_local
);
2305 struct sched_domain
*sd
;
2306 for_each_domain(this_cpu
, sd
) {
2307 if (cpu_isset(cpu
, sd
->span
)) {
2308 schedstat_inc(sd
, ttwu_wake_remote
);
2313 #endif /* CONFIG_SCHEDSTATS */
2316 #endif /* CONFIG_SMP */
2317 schedstat_inc(p
, se
.nr_wakeups
);
2319 schedstat_inc(p
, se
.nr_wakeups_sync
);
2320 if (orig_cpu
!= cpu
)
2321 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2322 if (cpu
== this_cpu
)
2323 schedstat_inc(p
, se
.nr_wakeups_local
);
2325 schedstat_inc(p
, se
.nr_wakeups_remote
);
2326 update_rq_clock(rq
);
2327 activate_task(rq
, p
, 1);
2331 trace_sched_wakeup(rq
, p
);
2332 check_preempt_curr(rq
, p
, sync
);
2334 p
->state
= TASK_RUNNING
;
2336 if (p
->sched_class
->task_wake_up
)
2337 p
->sched_class
->task_wake_up(rq
, p
);
2340 current
->se
.last_wakeup
= current
->se
.sum_exec_runtime
;
2342 task_rq_unlock(rq
, &flags
);
2347 int wake_up_process(struct task_struct
*p
)
2349 return try_to_wake_up(p
, TASK_ALL
, 0);
2351 EXPORT_SYMBOL(wake_up_process
);
2353 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2355 return try_to_wake_up(p
, state
, 0);
2359 * Perform scheduler related setup for a newly forked process p.
2360 * p is forked by current.
2362 * __sched_fork() is basic setup used by init_idle() too:
2364 static void __sched_fork(struct task_struct
*p
)
2366 p
->se
.exec_start
= 0;
2367 p
->se
.sum_exec_runtime
= 0;
2368 p
->se
.prev_sum_exec_runtime
= 0;
2369 p
->se
.last_wakeup
= 0;
2370 p
->se
.avg_overlap
= 0;
2372 #ifdef CONFIG_SCHEDSTATS
2373 p
->se
.wait_start
= 0;
2374 p
->se
.sum_sleep_runtime
= 0;
2375 p
->se
.sleep_start
= 0;
2376 p
->se
.block_start
= 0;
2377 p
->se
.sleep_max
= 0;
2378 p
->se
.block_max
= 0;
2380 p
->se
.slice_max
= 0;
2384 INIT_LIST_HEAD(&p
->rt
.run_list
);
2386 INIT_LIST_HEAD(&p
->se
.group_node
);
2388 #ifdef CONFIG_PREEMPT_NOTIFIERS
2389 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2393 * We mark the process as running here, but have not actually
2394 * inserted it onto the runqueue yet. This guarantees that
2395 * nobody will actually run it, and a signal or other external
2396 * event cannot wake it up and insert it on the runqueue either.
2398 p
->state
= TASK_RUNNING
;
2402 * fork()/clone()-time setup:
2404 void sched_fork(struct task_struct
*p
, int clone_flags
)
2406 int cpu
= get_cpu();
2411 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2413 set_task_cpu(p
, cpu
);
2416 * Make sure we do not leak PI boosting priority to the child:
2418 p
->prio
= current
->normal_prio
;
2419 if (!rt_prio(p
->prio
))
2420 p
->sched_class
= &fair_sched_class
;
2422 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2423 if (likely(sched_info_on()))
2424 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2426 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2429 #ifdef CONFIG_PREEMPT
2430 /* Want to start with kernel preemption disabled. */
2431 task_thread_info(p
)->preempt_count
= 1;
2437 * wake_up_new_task - wake up a newly created task for the first time.
2439 * This function will do some initial scheduler statistics housekeeping
2440 * that must be done for every newly created context, then puts the task
2441 * on the runqueue and wakes it.
2443 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2445 unsigned long flags
;
2448 rq
= task_rq_lock(p
, &flags
);
2449 BUG_ON(p
->state
!= TASK_RUNNING
);
2450 update_rq_clock(rq
);
2452 p
->prio
= effective_prio(p
);
2454 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2455 activate_task(rq
, p
, 0);
2458 * Let the scheduling class do new task startup
2459 * management (if any):
2461 p
->sched_class
->task_new(rq
, p
);
2464 trace_sched_wakeup_new(rq
, p
);
2465 check_preempt_curr(rq
, p
, 0);
2467 if (p
->sched_class
->task_wake_up
)
2468 p
->sched_class
->task_wake_up(rq
, p
);
2470 task_rq_unlock(rq
, &flags
);
2473 #ifdef CONFIG_PREEMPT_NOTIFIERS
2476 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2477 * @notifier: notifier struct to register
2479 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2481 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2483 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2486 * preempt_notifier_unregister - no longer interested in preemption notifications
2487 * @notifier: notifier struct to unregister
2489 * This is safe to call from within a preemption notifier.
2491 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2493 hlist_del(¬ifier
->link
);
2495 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2497 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2499 struct preempt_notifier
*notifier
;
2500 struct hlist_node
*node
;
2502 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2503 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2507 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2508 struct task_struct
*next
)
2510 struct preempt_notifier
*notifier
;
2511 struct hlist_node
*node
;
2513 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2514 notifier
->ops
->sched_out(notifier
, next
);
2517 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2519 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2524 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2525 struct task_struct
*next
)
2529 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2532 * prepare_task_switch - prepare to switch tasks
2533 * @rq: the runqueue preparing to switch
2534 * @prev: the current task that is being switched out
2535 * @next: the task we are going to switch to.
2537 * This is called with the rq lock held and interrupts off. It must
2538 * be paired with a subsequent finish_task_switch after the context
2541 * prepare_task_switch sets up locking and calls architecture specific
2545 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2546 struct task_struct
*next
)
2548 fire_sched_out_preempt_notifiers(prev
, next
);
2549 prepare_lock_switch(rq
, next
);
2550 prepare_arch_switch(next
);
2554 * finish_task_switch - clean up after a task-switch
2555 * @rq: runqueue associated with task-switch
2556 * @prev: the thread we just switched away from.
2558 * finish_task_switch must be called after the context switch, paired
2559 * with a prepare_task_switch call before the context switch.
2560 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2561 * and do any other architecture-specific cleanup actions.
2563 * Note that we may have delayed dropping an mm in context_switch(). If
2564 * so, we finish that here outside of the runqueue lock. (Doing it
2565 * with the lock held can cause deadlocks; see schedule() for
2568 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2569 __releases(rq
->lock
)
2571 struct mm_struct
*mm
= rq
->prev_mm
;
2577 * A task struct has one reference for the use as "current".
2578 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2579 * schedule one last time. The schedule call will never return, and
2580 * the scheduled task must drop that reference.
2581 * The test for TASK_DEAD must occur while the runqueue locks are
2582 * still held, otherwise prev could be scheduled on another cpu, die
2583 * there before we look at prev->state, and then the reference would
2585 * Manfred Spraul <manfred@colorfullife.com>
2587 prev_state
= prev
->state
;
2588 finish_arch_switch(prev
);
2589 finish_lock_switch(rq
, prev
);
2591 if (current
->sched_class
->post_schedule
)
2592 current
->sched_class
->post_schedule(rq
);
2595 fire_sched_in_preempt_notifiers(current
);
2598 if (unlikely(prev_state
== TASK_DEAD
)) {
2600 * Remove function-return probe instances associated with this
2601 * task and put them back on the free list.
2603 kprobe_flush_task(prev
);
2604 put_task_struct(prev
);
2609 * schedule_tail - first thing a freshly forked thread must call.
2610 * @prev: the thread we just switched away from.
2612 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2613 __releases(rq
->lock
)
2615 struct rq
*rq
= this_rq();
2617 finish_task_switch(rq
, prev
);
2618 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2619 /* In this case, finish_task_switch does not reenable preemption */
2622 if (current
->set_child_tid
)
2623 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2627 * context_switch - switch to the new MM and the new
2628 * thread's register state.
2631 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2632 struct task_struct
*next
)
2634 struct mm_struct
*mm
, *oldmm
;
2636 prepare_task_switch(rq
, prev
, next
);
2637 trace_sched_switch(rq
, prev
, next
);
2639 oldmm
= prev
->active_mm
;
2641 * For paravirt, this is coupled with an exit in switch_to to
2642 * combine the page table reload and the switch backend into
2645 arch_enter_lazy_cpu_mode();
2647 if (unlikely(!mm
)) {
2648 next
->active_mm
= oldmm
;
2649 atomic_inc(&oldmm
->mm_count
);
2650 enter_lazy_tlb(oldmm
, next
);
2652 switch_mm(oldmm
, mm
, next
);
2654 if (unlikely(!prev
->mm
)) {
2655 prev
->active_mm
= NULL
;
2656 rq
->prev_mm
= oldmm
;
2659 * Since the runqueue lock will be released by the next
2660 * task (which is an invalid locking op but in the case
2661 * of the scheduler it's an obvious special-case), so we
2662 * do an early lockdep release here:
2664 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2665 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2668 /* Here we just switch the register state and the stack. */
2669 switch_to(prev
, next
, prev
);
2673 * this_rq must be evaluated again because prev may have moved
2674 * CPUs since it called schedule(), thus the 'rq' on its stack
2675 * frame will be invalid.
2677 finish_task_switch(this_rq(), prev
);
2681 * nr_running, nr_uninterruptible and nr_context_switches:
2683 * externally visible scheduler statistics: current number of runnable
2684 * threads, current number of uninterruptible-sleeping threads, total
2685 * number of context switches performed since bootup.
2687 unsigned long nr_running(void)
2689 unsigned long i
, sum
= 0;
2691 for_each_online_cpu(i
)
2692 sum
+= cpu_rq(i
)->nr_running
;
2697 unsigned long nr_uninterruptible(void)
2699 unsigned long i
, sum
= 0;
2701 for_each_possible_cpu(i
)
2702 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2705 * Since we read the counters lockless, it might be slightly
2706 * inaccurate. Do not allow it to go below zero though:
2708 if (unlikely((long)sum
< 0))
2714 unsigned long long nr_context_switches(void)
2717 unsigned long long sum
= 0;
2719 for_each_possible_cpu(i
)
2720 sum
+= cpu_rq(i
)->nr_switches
;
2725 unsigned long nr_iowait(void)
2727 unsigned long i
, sum
= 0;
2729 for_each_possible_cpu(i
)
2730 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2735 unsigned long nr_active(void)
2737 unsigned long i
, running
= 0, uninterruptible
= 0;
2739 for_each_online_cpu(i
) {
2740 running
+= cpu_rq(i
)->nr_running
;
2741 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2744 if (unlikely((long)uninterruptible
< 0))
2745 uninterruptible
= 0;
2747 return running
+ uninterruptible
;
2751 * Update rq->cpu_load[] statistics. This function is usually called every
2752 * scheduler tick (TICK_NSEC).
2754 static void update_cpu_load(struct rq
*this_rq
)
2756 unsigned long this_load
= this_rq
->load
.weight
;
2759 this_rq
->nr_load_updates
++;
2761 /* Update our load: */
2762 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2763 unsigned long old_load
, new_load
;
2765 /* scale is effectively 1 << i now, and >> i divides by scale */
2767 old_load
= this_rq
->cpu_load
[i
];
2768 new_load
= this_load
;
2770 * Round up the averaging division if load is increasing. This
2771 * prevents us from getting stuck on 9 if the load is 10, for
2774 if (new_load
> old_load
)
2775 new_load
+= scale
-1;
2776 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2783 * double_rq_lock - safely lock two runqueues
2785 * Note this does not disable interrupts like task_rq_lock,
2786 * you need to do so manually before calling.
2788 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2789 __acquires(rq1
->lock
)
2790 __acquires(rq2
->lock
)
2792 BUG_ON(!irqs_disabled());
2794 spin_lock(&rq1
->lock
);
2795 __acquire(rq2
->lock
); /* Fake it out ;) */
2798 spin_lock(&rq1
->lock
);
2799 spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
2801 spin_lock(&rq2
->lock
);
2802 spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
2805 update_rq_clock(rq1
);
2806 update_rq_clock(rq2
);
2810 * double_rq_unlock - safely unlock two runqueues
2812 * Note this does not restore interrupts like task_rq_unlock,
2813 * you need to do so manually after calling.
2815 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2816 __releases(rq1
->lock
)
2817 __releases(rq2
->lock
)
2819 spin_unlock(&rq1
->lock
);
2821 spin_unlock(&rq2
->lock
);
2823 __release(rq2
->lock
);
2827 * If dest_cpu is allowed for this process, migrate the task to it.
2828 * This is accomplished by forcing the cpu_allowed mask to only
2829 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2830 * the cpu_allowed mask is restored.
2832 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2834 struct migration_req req
;
2835 unsigned long flags
;
2838 rq
= task_rq_lock(p
, &flags
);
2839 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2840 || unlikely(!cpu_active(dest_cpu
)))
2843 trace_sched_migrate_task(rq
, p
, dest_cpu
);
2844 /* force the process onto the specified CPU */
2845 if (migrate_task(p
, dest_cpu
, &req
)) {
2846 /* Need to wait for migration thread (might exit: take ref). */
2847 struct task_struct
*mt
= rq
->migration_thread
;
2849 get_task_struct(mt
);
2850 task_rq_unlock(rq
, &flags
);
2851 wake_up_process(mt
);
2852 put_task_struct(mt
);
2853 wait_for_completion(&req
.done
);
2858 task_rq_unlock(rq
, &flags
);
2862 * sched_exec - execve() is a valuable balancing opportunity, because at
2863 * this point the task has the smallest effective memory and cache footprint.
2865 void sched_exec(void)
2867 int new_cpu
, this_cpu
= get_cpu();
2868 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2870 if (new_cpu
!= this_cpu
)
2871 sched_migrate_task(current
, new_cpu
);
2875 * pull_task - move a task from a remote runqueue to the local runqueue.
2876 * Both runqueues must be locked.
2878 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2879 struct rq
*this_rq
, int this_cpu
)
2881 deactivate_task(src_rq
, p
, 0);
2882 set_task_cpu(p
, this_cpu
);
2883 activate_task(this_rq
, p
, 0);
2885 * Note that idle threads have a prio of MAX_PRIO, for this test
2886 * to be always true for them.
2888 check_preempt_curr(this_rq
, p
, 0);
2892 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2895 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2896 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2900 * We do not migrate tasks that are:
2901 * 1) running (obviously), or
2902 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2903 * 3) are cache-hot on their current CPU.
2905 if (!cpu_isset(this_cpu
, p
->cpus_allowed
)) {
2906 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2911 if (task_running(rq
, p
)) {
2912 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2917 * Aggressive migration if:
2918 * 1) task is cache cold, or
2919 * 2) too many balance attempts have failed.
2922 if (!task_hot(p
, rq
->clock
, sd
) ||
2923 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2924 #ifdef CONFIG_SCHEDSTATS
2925 if (task_hot(p
, rq
->clock
, sd
)) {
2926 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2927 schedstat_inc(p
, se
.nr_forced_migrations
);
2933 if (task_hot(p
, rq
->clock
, sd
)) {
2934 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
2940 static unsigned long
2941 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2942 unsigned long max_load_move
, struct sched_domain
*sd
,
2943 enum cpu_idle_type idle
, int *all_pinned
,
2944 int *this_best_prio
, struct rq_iterator
*iterator
)
2946 int loops
= 0, pulled
= 0, pinned
= 0;
2947 struct task_struct
*p
;
2948 long rem_load_move
= max_load_move
;
2950 if (max_load_move
== 0)
2956 * Start the load-balancing iterator:
2958 p
= iterator
->start(iterator
->arg
);
2960 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
2963 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
2964 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2965 p
= iterator
->next(iterator
->arg
);
2969 pull_task(busiest
, p
, this_rq
, this_cpu
);
2971 rem_load_move
-= p
->se
.load
.weight
;
2974 * We only want to steal up to the prescribed amount of weighted load.
2976 if (rem_load_move
> 0) {
2977 if (p
->prio
< *this_best_prio
)
2978 *this_best_prio
= p
->prio
;
2979 p
= iterator
->next(iterator
->arg
);
2984 * Right now, this is one of only two places pull_task() is called,
2985 * so we can safely collect pull_task() stats here rather than
2986 * inside pull_task().
2988 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2991 *all_pinned
= pinned
;
2993 return max_load_move
- rem_load_move
;
2997 * move_tasks tries to move up to max_load_move weighted load from busiest to
2998 * this_rq, as part of a balancing operation within domain "sd".
2999 * Returns 1 if successful and 0 otherwise.
3001 * Called with both runqueues locked.
3003 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3004 unsigned long max_load_move
,
3005 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3008 const struct sched_class
*class = sched_class_highest
;
3009 unsigned long total_load_moved
= 0;
3010 int this_best_prio
= this_rq
->curr
->prio
;
3014 class->load_balance(this_rq
, this_cpu
, busiest
,
3015 max_load_move
- total_load_moved
,
3016 sd
, idle
, all_pinned
, &this_best_prio
);
3017 class = class->next
;
3019 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3022 } while (class && max_load_move
> total_load_moved
);
3024 return total_load_moved
> 0;
3028 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3029 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3030 struct rq_iterator
*iterator
)
3032 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3036 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3037 pull_task(busiest
, p
, this_rq
, this_cpu
);
3039 * Right now, this is only the second place pull_task()
3040 * is called, so we can safely collect pull_task()
3041 * stats here rather than inside pull_task().
3043 schedstat_inc(sd
, lb_gained
[idle
]);
3047 p
= iterator
->next(iterator
->arg
);
3054 * move_one_task tries to move exactly one task from busiest to this_rq, as
3055 * part of active balancing operations within "domain".
3056 * Returns 1 if successful and 0 otherwise.
3058 * Called with both runqueues locked.
3060 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3061 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3063 const struct sched_class
*class;
3065 for (class = sched_class_highest
; class; class = class->next
)
3066 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3073 * find_busiest_group finds and returns the busiest CPU group within the
3074 * domain. It calculates and returns the amount of weighted load which
3075 * should be moved to restore balance via the imbalance parameter.
3077 static struct sched_group
*
3078 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
3079 unsigned long *imbalance
, enum cpu_idle_type idle
,
3080 int *sd_idle
, const cpumask_t
*cpus
, int *balance
)
3082 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
3083 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
3084 unsigned long max_pull
;
3085 unsigned long busiest_load_per_task
, busiest_nr_running
;
3086 unsigned long this_load_per_task
, this_nr_running
;
3087 int load_idx
, group_imb
= 0;
3088 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3089 int power_savings_balance
= 1;
3090 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
3091 unsigned long min_nr_running
= ULONG_MAX
;
3092 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
3095 max_load
= this_load
= total_load
= total_pwr
= 0;
3096 busiest_load_per_task
= busiest_nr_running
= 0;
3097 this_load_per_task
= this_nr_running
= 0;
3099 if (idle
== CPU_NOT_IDLE
)
3100 load_idx
= sd
->busy_idx
;
3101 else if (idle
== CPU_NEWLY_IDLE
)
3102 load_idx
= sd
->newidle_idx
;
3104 load_idx
= sd
->idle_idx
;
3107 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
3110 int __group_imb
= 0;
3111 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3112 unsigned long sum_nr_running
, sum_weighted_load
;
3113 unsigned long sum_avg_load_per_task
;
3114 unsigned long avg_load_per_task
;
3116 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
3119 balance_cpu
= first_cpu(group
->cpumask
);
3121 /* Tally up the load of all CPUs in the group */
3122 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
3123 sum_avg_load_per_task
= avg_load_per_task
= 0;
3126 min_cpu_load
= ~0UL;
3128 for_each_cpu_mask_nr(i
, group
->cpumask
) {
3131 if (!cpu_isset(i
, *cpus
))
3136 if (*sd_idle
&& rq
->nr_running
)
3139 /* Bias balancing toward cpus of our domain */
3141 if (idle_cpu(i
) && !first_idle_cpu
) {
3146 load
= target_load(i
, load_idx
);
3148 load
= source_load(i
, load_idx
);
3149 if (load
> max_cpu_load
)
3150 max_cpu_load
= load
;
3151 if (min_cpu_load
> load
)
3152 min_cpu_load
= load
;
3156 sum_nr_running
+= rq
->nr_running
;
3157 sum_weighted_load
+= weighted_cpuload(i
);
3159 sum_avg_load_per_task
+= cpu_avg_load_per_task(i
);
3163 * First idle cpu or the first cpu(busiest) in this sched group
3164 * is eligible for doing load balancing at this and above
3165 * domains. In the newly idle case, we will allow all the cpu's
3166 * to do the newly idle load balance.
3168 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3169 balance_cpu
!= this_cpu
&& balance
) {
3174 total_load
+= avg_load
;
3175 total_pwr
+= group
->__cpu_power
;
3177 /* Adjust by relative CPU power of the group */
3178 avg_load
= sg_div_cpu_power(group
,
3179 avg_load
* SCHED_LOAD_SCALE
);
3183 * Consider the group unbalanced when the imbalance is larger
3184 * than the average weight of two tasks.
3186 * APZ: with cgroup the avg task weight can vary wildly and
3187 * might not be a suitable number - should we keep a
3188 * normalized nr_running number somewhere that negates
3191 avg_load_per_task
= sg_div_cpu_power(group
,
3192 sum_avg_load_per_task
* SCHED_LOAD_SCALE
);
3194 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3197 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
3200 this_load
= avg_load
;
3202 this_nr_running
= sum_nr_running
;
3203 this_load_per_task
= sum_weighted_load
;
3204 } else if (avg_load
> max_load
&&
3205 (sum_nr_running
> group_capacity
|| __group_imb
)) {
3206 max_load
= avg_load
;
3208 busiest_nr_running
= sum_nr_running
;
3209 busiest_load_per_task
= sum_weighted_load
;
3210 group_imb
= __group_imb
;
3213 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3215 * Busy processors will not participate in power savings
3218 if (idle
== CPU_NOT_IDLE
||
3219 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3223 * If the local group is idle or completely loaded
3224 * no need to do power savings balance at this domain
3226 if (local_group
&& (this_nr_running
>= group_capacity
||
3228 power_savings_balance
= 0;
3231 * If a group is already running at full capacity or idle,
3232 * don't include that group in power savings calculations
3234 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
3239 * Calculate the group which has the least non-idle load.
3240 * This is the group from where we need to pick up the load
3243 if ((sum_nr_running
< min_nr_running
) ||
3244 (sum_nr_running
== min_nr_running
&&
3245 first_cpu(group
->cpumask
) <
3246 first_cpu(group_min
->cpumask
))) {
3248 min_nr_running
= sum_nr_running
;
3249 min_load_per_task
= sum_weighted_load
/
3254 * Calculate the group which is almost near its
3255 * capacity but still has some space to pick up some load
3256 * from other group and save more power
3258 if (sum_nr_running
<= group_capacity
- 1) {
3259 if (sum_nr_running
> leader_nr_running
||
3260 (sum_nr_running
== leader_nr_running
&&
3261 first_cpu(group
->cpumask
) >
3262 first_cpu(group_leader
->cpumask
))) {
3263 group_leader
= group
;
3264 leader_nr_running
= sum_nr_running
;
3269 group
= group
->next
;
3270 } while (group
!= sd
->groups
);
3272 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
3275 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
3277 if (this_load
>= avg_load
||
3278 100*max_load
<= sd
->imbalance_pct
*this_load
)
3281 busiest_load_per_task
/= busiest_nr_running
;
3283 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
3286 * We're trying to get all the cpus to the average_load, so we don't
3287 * want to push ourselves above the average load, nor do we wish to
3288 * reduce the max loaded cpu below the average load, as either of these
3289 * actions would just result in more rebalancing later, and ping-pong
3290 * tasks around. Thus we look for the minimum possible imbalance.
3291 * Negative imbalances (*we* are more loaded than anyone else) will
3292 * be counted as no imbalance for these purposes -- we can't fix that
3293 * by pulling tasks to us. Be careful of negative numbers as they'll
3294 * appear as very large values with unsigned longs.
3296 if (max_load
<= busiest_load_per_task
)
3300 * In the presence of smp nice balancing, certain scenarios can have
3301 * max load less than avg load(as we skip the groups at or below
3302 * its cpu_power, while calculating max_load..)
3304 if (max_load
< avg_load
) {
3306 goto small_imbalance
;
3309 /* Don't want to pull so many tasks that a group would go idle */
3310 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
3312 /* How much load to actually move to equalise the imbalance */
3313 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
3314 (avg_load
- this_load
) * this->__cpu_power
)
3318 * if *imbalance is less than the average load per runnable task
3319 * there is no gaurantee that any tasks will be moved so we'll have
3320 * a think about bumping its value to force at least one task to be
3323 if (*imbalance
< busiest_load_per_task
) {
3324 unsigned long tmp
, pwr_now
, pwr_move
;
3328 pwr_move
= pwr_now
= 0;
3330 if (this_nr_running
) {
3331 this_load_per_task
/= this_nr_running
;
3332 if (busiest_load_per_task
> this_load_per_task
)
3335 this_load_per_task
= cpu_avg_load_per_task(this_cpu
);
3337 if (max_load
- this_load
+ busiest_load_per_task
>=
3338 busiest_load_per_task
* imbn
) {
3339 *imbalance
= busiest_load_per_task
;
3344 * OK, we don't have enough imbalance to justify moving tasks,
3345 * however we may be able to increase total CPU power used by
3349 pwr_now
+= busiest
->__cpu_power
*
3350 min(busiest_load_per_task
, max_load
);
3351 pwr_now
+= this->__cpu_power
*
3352 min(this_load_per_task
, this_load
);
3353 pwr_now
/= SCHED_LOAD_SCALE
;
3355 /* Amount of load we'd subtract */
3356 tmp
= sg_div_cpu_power(busiest
,
3357 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3359 pwr_move
+= busiest
->__cpu_power
*
3360 min(busiest_load_per_task
, max_load
- tmp
);
3362 /* Amount of load we'd add */
3363 if (max_load
* busiest
->__cpu_power
<
3364 busiest_load_per_task
* SCHED_LOAD_SCALE
)
3365 tmp
= sg_div_cpu_power(this,
3366 max_load
* busiest
->__cpu_power
);
3368 tmp
= sg_div_cpu_power(this,
3369 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3370 pwr_move
+= this->__cpu_power
*
3371 min(this_load_per_task
, this_load
+ tmp
);
3372 pwr_move
/= SCHED_LOAD_SCALE
;
3374 /* Move if we gain throughput */
3375 if (pwr_move
> pwr_now
)
3376 *imbalance
= busiest_load_per_task
;
3382 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3383 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3386 if (this == group_leader
&& group_leader
!= group_min
) {
3387 *imbalance
= min_load_per_task
;
3397 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3400 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
3401 unsigned long imbalance
, const cpumask_t
*cpus
)
3403 struct rq
*busiest
= NULL
, *rq
;
3404 unsigned long max_load
= 0;
3407 for_each_cpu_mask_nr(i
, group
->cpumask
) {
3410 if (!cpu_isset(i
, *cpus
))
3414 wl
= weighted_cpuload(i
);
3416 if (rq
->nr_running
== 1 && wl
> imbalance
)
3419 if (wl
> max_load
) {
3429 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3430 * so long as it is large enough.
3432 #define MAX_PINNED_INTERVAL 512
3435 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3436 * tasks if there is an imbalance.
3438 static int load_balance(int this_cpu
, struct rq
*this_rq
,
3439 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3440 int *balance
, cpumask_t
*cpus
)
3442 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
3443 struct sched_group
*group
;
3444 unsigned long imbalance
;
3446 unsigned long flags
;
3451 * When power savings policy is enabled for the parent domain, idle
3452 * sibling can pick up load irrespective of busy siblings. In this case,
3453 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3454 * portraying it as CPU_NOT_IDLE.
3456 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3457 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3460 schedstat_inc(sd
, lb_count
[idle
]);
3464 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3471 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3475 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
3477 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3481 BUG_ON(busiest
== this_rq
);
3483 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3486 if (busiest
->nr_running
> 1) {
3488 * Attempt to move tasks. If find_busiest_group has found
3489 * an imbalance but busiest->nr_running <= 1, the group is
3490 * still unbalanced. ld_moved simply stays zero, so it is
3491 * correctly treated as an imbalance.
3493 local_irq_save(flags
);
3494 double_rq_lock(this_rq
, busiest
);
3495 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3496 imbalance
, sd
, idle
, &all_pinned
);
3497 double_rq_unlock(this_rq
, busiest
);
3498 local_irq_restore(flags
);
3501 * some other cpu did the load balance for us.
3503 if (ld_moved
&& this_cpu
!= smp_processor_id())
3504 resched_cpu(this_cpu
);
3506 /* All tasks on this runqueue were pinned by CPU affinity */
3507 if (unlikely(all_pinned
)) {
3508 cpu_clear(cpu_of(busiest
), *cpus
);
3509 if (!cpus_empty(*cpus
))
3516 schedstat_inc(sd
, lb_failed
[idle
]);
3517 sd
->nr_balance_failed
++;
3519 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3521 spin_lock_irqsave(&busiest
->lock
, flags
);
3523 /* don't kick the migration_thread, if the curr
3524 * task on busiest cpu can't be moved to this_cpu
3526 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
3527 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3529 goto out_one_pinned
;
3532 if (!busiest
->active_balance
) {
3533 busiest
->active_balance
= 1;
3534 busiest
->push_cpu
= this_cpu
;
3537 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3539 wake_up_process(busiest
->migration_thread
);
3542 * We've kicked active balancing, reset the failure
3545 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3548 sd
->nr_balance_failed
= 0;
3550 if (likely(!active_balance
)) {
3551 /* We were unbalanced, so reset the balancing interval */
3552 sd
->balance_interval
= sd
->min_interval
;
3555 * If we've begun active balancing, start to back off. This
3556 * case may not be covered by the all_pinned logic if there
3557 * is only 1 task on the busy runqueue (because we don't call
3560 if (sd
->balance_interval
< sd
->max_interval
)
3561 sd
->balance_interval
*= 2;
3564 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3565 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3571 schedstat_inc(sd
, lb_balanced
[idle
]);
3573 sd
->nr_balance_failed
= 0;
3576 /* tune up the balancing interval */
3577 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3578 (sd
->balance_interval
< sd
->max_interval
))
3579 sd
->balance_interval
*= 2;
3581 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3582 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3593 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3594 * tasks if there is an imbalance.
3596 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3597 * this_rq is locked.
3600 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
,
3603 struct sched_group
*group
;
3604 struct rq
*busiest
= NULL
;
3605 unsigned long imbalance
;
3613 * When power savings policy is enabled for the parent domain, idle
3614 * sibling can pick up load irrespective of busy siblings. In this case,
3615 * let the state of idle sibling percolate up as IDLE, instead of
3616 * portraying it as CPU_NOT_IDLE.
3618 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
3619 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3622 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
3624 update_shares_locked(this_rq
, sd
);
3625 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
3626 &sd_idle
, cpus
, NULL
);
3628 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
3632 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
3634 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
3638 BUG_ON(busiest
== this_rq
);
3640 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
3643 if (busiest
->nr_running
> 1) {
3644 /* Attempt to move tasks */
3645 double_lock_balance(this_rq
, busiest
);
3646 /* this_rq->clock is already updated */
3647 update_rq_clock(busiest
);
3648 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3649 imbalance
, sd
, CPU_NEWLY_IDLE
,
3651 double_unlock_balance(this_rq
, busiest
);
3653 if (unlikely(all_pinned
)) {
3654 cpu_clear(cpu_of(busiest
), *cpus
);
3655 if (!cpus_empty(*cpus
))
3661 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
3662 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3663 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3666 sd
->nr_balance_failed
= 0;
3668 update_shares_locked(this_rq
, sd
);
3672 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
3673 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3674 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3676 sd
->nr_balance_failed
= 0;
3682 * idle_balance is called by schedule() if this_cpu is about to become
3683 * idle. Attempts to pull tasks from other CPUs.
3685 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
3687 struct sched_domain
*sd
;
3688 int pulled_task
= 0;
3689 unsigned long next_balance
= jiffies
+ HZ
;
3692 for_each_domain(this_cpu
, sd
) {
3693 unsigned long interval
;
3695 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3698 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
3699 /* If we've pulled tasks over stop searching: */
3700 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
3703 interval
= msecs_to_jiffies(sd
->balance_interval
);
3704 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3705 next_balance
= sd
->last_balance
+ interval
;
3709 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3711 * We are going idle. next_balance may be set based on
3712 * a busy processor. So reset next_balance.
3714 this_rq
->next_balance
= next_balance
;
3719 * active_load_balance is run by migration threads. It pushes running tasks
3720 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3721 * running on each physical CPU where possible, and avoids physical /
3722 * logical imbalances.
3724 * Called with busiest_rq locked.
3726 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
3728 int target_cpu
= busiest_rq
->push_cpu
;
3729 struct sched_domain
*sd
;
3730 struct rq
*target_rq
;
3732 /* Is there any task to move? */
3733 if (busiest_rq
->nr_running
<= 1)
3736 target_rq
= cpu_rq(target_cpu
);
3739 * This condition is "impossible", if it occurs
3740 * we need to fix it. Originally reported by
3741 * Bjorn Helgaas on a 128-cpu setup.
3743 BUG_ON(busiest_rq
== target_rq
);
3745 /* move a task from busiest_rq to target_rq */
3746 double_lock_balance(busiest_rq
, target_rq
);
3747 update_rq_clock(busiest_rq
);
3748 update_rq_clock(target_rq
);
3750 /* Search for an sd spanning us and the target CPU. */
3751 for_each_domain(target_cpu
, sd
) {
3752 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3753 cpu_isset(busiest_cpu
, sd
->span
))
3758 schedstat_inc(sd
, alb_count
);
3760 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3762 schedstat_inc(sd
, alb_pushed
);
3764 schedstat_inc(sd
, alb_failed
);
3766 double_unlock_balance(busiest_rq
, target_rq
);
3771 atomic_t load_balancer
;
3773 } nohz ____cacheline_aligned
= {
3774 .load_balancer
= ATOMIC_INIT(-1),
3775 .cpu_mask
= CPU_MASK_NONE
,
3779 * This routine will try to nominate the ilb (idle load balancing)
3780 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3781 * load balancing on behalf of all those cpus. If all the cpus in the system
3782 * go into this tickless mode, then there will be no ilb owner (as there is
3783 * no need for one) and all the cpus will sleep till the next wakeup event
3786 * For the ilb owner, tick is not stopped. And this tick will be used
3787 * for idle load balancing. ilb owner will still be part of
3790 * While stopping the tick, this cpu will become the ilb owner if there
3791 * is no other owner. And will be the owner till that cpu becomes busy
3792 * or if all cpus in the system stop their ticks at which point
3793 * there is no need for ilb owner.
3795 * When the ilb owner becomes busy, it nominates another owner, during the
3796 * next busy scheduler_tick()
3798 int select_nohz_load_balancer(int stop_tick
)
3800 int cpu
= smp_processor_id();
3803 cpu_set(cpu
, nohz
.cpu_mask
);
3804 cpu_rq(cpu
)->in_nohz_recently
= 1;
3807 * If we are going offline and still the leader, give up!
3809 if (!cpu_active(cpu
) &&
3810 atomic_read(&nohz
.load_balancer
) == cpu
) {
3811 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3816 /* time for ilb owner also to sleep */
3817 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3818 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3819 atomic_set(&nohz
.load_balancer
, -1);
3823 if (atomic_read(&nohz
.load_balancer
) == -1) {
3824 /* make me the ilb owner */
3825 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3827 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3830 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3833 cpu_clear(cpu
, nohz
.cpu_mask
);
3835 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3836 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3843 static DEFINE_SPINLOCK(balancing
);
3846 * It checks each scheduling domain to see if it is due to be balanced,
3847 * and initiates a balancing operation if so.
3849 * Balancing parameters are set up in arch_init_sched_domains.
3851 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3854 struct rq
*rq
= cpu_rq(cpu
);
3855 unsigned long interval
;
3856 struct sched_domain
*sd
;
3857 /* Earliest time when we have to do rebalance again */
3858 unsigned long next_balance
= jiffies
+ 60*HZ
;
3859 int update_next_balance
= 0;
3863 for_each_domain(cpu
, sd
) {
3864 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3867 interval
= sd
->balance_interval
;
3868 if (idle
!= CPU_IDLE
)
3869 interval
*= sd
->busy_factor
;
3871 /* scale ms to jiffies */
3872 interval
= msecs_to_jiffies(interval
);
3873 if (unlikely(!interval
))
3875 if (interval
> HZ
*NR_CPUS
/10)
3876 interval
= HZ
*NR_CPUS
/10;
3878 need_serialize
= sd
->flags
& SD_SERIALIZE
;
3880 if (need_serialize
) {
3881 if (!spin_trylock(&balancing
))
3885 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3886 if (load_balance(cpu
, rq
, sd
, idle
, &balance
, &tmp
)) {
3888 * We've pulled tasks over so either we're no
3889 * longer idle, or one of our SMT siblings is
3892 idle
= CPU_NOT_IDLE
;
3894 sd
->last_balance
= jiffies
;
3897 spin_unlock(&balancing
);
3899 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3900 next_balance
= sd
->last_balance
+ interval
;
3901 update_next_balance
= 1;
3905 * Stop the load balance at this level. There is another
3906 * CPU in our sched group which is doing load balancing more
3914 * next_balance will be updated only when there is a need.
3915 * When the cpu is attached to null domain for ex, it will not be
3918 if (likely(update_next_balance
))
3919 rq
->next_balance
= next_balance
;
3923 * run_rebalance_domains is triggered when needed from the scheduler tick.
3924 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3925 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3927 static void run_rebalance_domains(struct softirq_action
*h
)
3929 int this_cpu
= smp_processor_id();
3930 struct rq
*this_rq
= cpu_rq(this_cpu
);
3931 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3932 CPU_IDLE
: CPU_NOT_IDLE
;
3934 rebalance_domains(this_cpu
, idle
);
3938 * If this cpu is the owner for idle load balancing, then do the
3939 * balancing on behalf of the other idle cpus whose ticks are
3942 if (this_rq
->idle_at_tick
&&
3943 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3944 cpumask_t cpus
= nohz
.cpu_mask
;
3948 cpu_clear(this_cpu
, cpus
);
3949 for_each_cpu_mask_nr(balance_cpu
, cpus
) {
3951 * If this cpu gets work to do, stop the load balancing
3952 * work being done for other cpus. Next load
3953 * balancing owner will pick it up.
3958 rebalance_domains(balance_cpu
, CPU_IDLE
);
3960 rq
= cpu_rq(balance_cpu
);
3961 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3962 this_rq
->next_balance
= rq
->next_balance
;
3969 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3971 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3972 * idle load balancing owner or decide to stop the periodic load balancing,
3973 * if the whole system is idle.
3975 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3979 * If we were in the nohz mode recently and busy at the current
3980 * scheduler tick, then check if we need to nominate new idle
3983 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3984 rq
->in_nohz_recently
= 0;
3986 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3987 cpu_clear(cpu
, nohz
.cpu_mask
);
3988 atomic_set(&nohz
.load_balancer
, -1);
3991 if (atomic_read(&nohz
.load_balancer
) == -1) {
3993 * simple selection for now: Nominate the
3994 * first cpu in the nohz list to be the next
3997 * TBD: Traverse the sched domains and nominate
3998 * the nearest cpu in the nohz.cpu_mask.
4000 int ilb
= first_cpu(nohz
.cpu_mask
);
4002 if (ilb
< nr_cpu_ids
)
4008 * If this cpu is idle and doing idle load balancing for all the
4009 * cpus with ticks stopped, is it time for that to stop?
4011 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4012 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4018 * If this cpu is idle and the idle load balancing is done by
4019 * someone else, then no need raise the SCHED_SOFTIRQ
4021 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4022 cpu_isset(cpu
, nohz
.cpu_mask
))
4025 if (time_after_eq(jiffies
, rq
->next_balance
))
4026 raise_softirq(SCHED_SOFTIRQ
);
4029 #else /* CONFIG_SMP */
4032 * on UP we do not need to balance between CPUs:
4034 static inline void idle_balance(int cpu
, struct rq
*rq
)
4040 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4042 EXPORT_PER_CPU_SYMBOL(kstat
);
4045 * Return any ns on the sched_clock that have not yet been banked in
4046 * @p in case that task is currently running.
4048 unsigned long long task_delta_exec(struct task_struct
*p
)
4050 unsigned long flags
;
4054 rq
= task_rq_lock(p
, &flags
);
4056 if (task_current(rq
, p
)) {
4059 update_rq_clock(rq
);
4060 delta_exec
= rq
->clock
- p
->se
.exec_start
;
4061 if ((s64
)delta_exec
> 0)
4065 task_rq_unlock(rq
, &flags
);
4071 * Account user cpu time to a process.
4072 * @p: the process that the cpu time gets accounted to
4073 * @cputime: the cpu time spent in user space since the last update
4075 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
4077 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4080 p
->utime
= cputime_add(p
->utime
, cputime
);
4081 account_group_user_time(p
, cputime
);
4083 /* Add user time to cpustat. */
4084 tmp
= cputime_to_cputime64(cputime
);
4085 if (TASK_NICE(p
) > 0)
4086 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
4088 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4089 /* Account for user time used */
4090 acct_update_integrals(p
);
4094 * Account guest cpu time to a process.
4095 * @p: the process that the cpu time gets accounted to
4096 * @cputime: the cpu time spent in virtual machine since the last update
4098 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
)
4101 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4103 tmp
= cputime_to_cputime64(cputime
);
4105 p
->utime
= cputime_add(p
->utime
, cputime
);
4106 account_group_user_time(p
, cputime
);
4107 p
->gtime
= cputime_add(p
->gtime
, cputime
);
4109 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4110 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
4114 * Account scaled user cpu time to a process.
4115 * @p: the process that the cpu time gets accounted to
4116 * @cputime: the cpu time spent in user space since the last update
4118 void account_user_time_scaled(struct task_struct
*p
, cputime_t cputime
)
4120 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime
);
4124 * Account system cpu time to a process.
4125 * @p: the process that the cpu time gets accounted to
4126 * @hardirq_offset: the offset to subtract from hardirq_count()
4127 * @cputime: the cpu time spent in kernel space since the last update
4129 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
4132 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4133 struct rq
*rq
= this_rq();
4136 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
4137 account_guest_time(p
, cputime
);
4141 p
->stime
= cputime_add(p
->stime
, cputime
);
4142 account_group_system_time(p
, cputime
);
4144 /* Add system time to cpustat. */
4145 tmp
= cputime_to_cputime64(cputime
);
4146 if (hardirq_count() - hardirq_offset
)
4147 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
4148 else if (softirq_count())
4149 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
4150 else if (p
!= rq
->idle
)
4151 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
4152 else if (atomic_read(&rq
->nr_iowait
) > 0)
4153 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
4155 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
4156 /* Account for system time used */
4157 acct_update_integrals(p
);
4161 * Account scaled system cpu time to a process.
4162 * @p: the process that the cpu time gets accounted to
4163 * @hardirq_offset: the offset to subtract from hardirq_count()
4164 * @cputime: the cpu time spent in kernel space since the last update
4166 void account_system_time_scaled(struct task_struct
*p
, cputime_t cputime
)
4168 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime
);
4172 * Account for involuntary wait time.
4173 * @p: the process from which the cpu time has been stolen
4174 * @steal: the cpu time spent in involuntary wait
4176 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
4178 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4179 cputime64_t tmp
= cputime_to_cputime64(steal
);
4180 struct rq
*rq
= this_rq();
4182 if (p
== rq
->idle
) {
4183 p
->stime
= cputime_add(p
->stime
, steal
);
4184 account_group_system_time(p
, steal
);
4185 if (atomic_read(&rq
->nr_iowait
) > 0)
4186 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
4188 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
4190 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
4194 * Use precise platform statistics if available:
4196 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4197 cputime_t
task_utime(struct task_struct
*p
)
4202 cputime_t
task_stime(struct task_struct
*p
)
4207 cputime_t
task_utime(struct task_struct
*p
)
4209 clock_t utime
= cputime_to_clock_t(p
->utime
),
4210 total
= utime
+ cputime_to_clock_t(p
->stime
);
4214 * Use CFS's precise accounting:
4216 temp
= (u64
)nsec_to_clock_t(p
->se
.sum_exec_runtime
);
4220 do_div(temp
, total
);
4222 utime
= (clock_t)temp
;
4224 p
->prev_utime
= max(p
->prev_utime
, clock_t_to_cputime(utime
));
4225 return p
->prev_utime
;
4228 cputime_t
task_stime(struct task_struct
*p
)
4233 * Use CFS's precise accounting. (we subtract utime from
4234 * the total, to make sure the total observed by userspace
4235 * grows monotonically - apps rely on that):
4237 stime
= nsec_to_clock_t(p
->se
.sum_exec_runtime
) -
4238 cputime_to_clock_t(task_utime(p
));
4241 p
->prev_stime
= max(p
->prev_stime
, clock_t_to_cputime(stime
));
4243 return p
->prev_stime
;
4247 inline cputime_t
task_gtime(struct task_struct
*p
)
4253 * This function gets called by the timer code, with HZ frequency.
4254 * We call it with interrupts disabled.
4256 * It also gets called by the fork code, when changing the parent's
4259 void scheduler_tick(void)
4261 int cpu
= smp_processor_id();
4262 struct rq
*rq
= cpu_rq(cpu
);
4263 struct task_struct
*curr
= rq
->curr
;
4267 spin_lock(&rq
->lock
);
4268 update_rq_clock(rq
);
4269 update_cpu_load(rq
);
4270 curr
->sched_class
->task_tick(rq
, curr
, 0);
4271 spin_unlock(&rq
->lock
);
4274 rq
->idle_at_tick
= idle_cpu(cpu
);
4275 trigger_load_balance(rq
, cpu
);
4279 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4280 defined(CONFIG_PREEMPT_TRACER))
4282 static inline unsigned long get_parent_ip(unsigned long addr
)
4284 if (in_lock_functions(addr
)) {
4285 addr
= CALLER_ADDR2
;
4286 if (in_lock_functions(addr
))
4287 addr
= CALLER_ADDR3
;
4292 void __kprobes
add_preempt_count(int val
)
4294 #ifdef CONFIG_DEBUG_PREEMPT
4298 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4301 preempt_count() += val
;
4302 #ifdef CONFIG_DEBUG_PREEMPT
4304 * Spinlock count overflowing soon?
4306 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4309 if (preempt_count() == val
)
4310 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4312 EXPORT_SYMBOL(add_preempt_count
);
4314 void __kprobes
sub_preempt_count(int val
)
4316 #ifdef CONFIG_DEBUG_PREEMPT
4320 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
4323 * Is the spinlock portion underflowing?
4325 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4326 !(preempt_count() & PREEMPT_MASK
)))
4330 if (preempt_count() == val
)
4331 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4332 preempt_count() -= val
;
4334 EXPORT_SYMBOL(sub_preempt_count
);
4339 * Print scheduling while atomic bug:
4341 static noinline
void __schedule_bug(struct task_struct
*prev
)
4343 struct pt_regs
*regs
= get_irq_regs();
4345 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4346 prev
->comm
, prev
->pid
, preempt_count());
4348 debug_show_held_locks(prev
);
4350 if (irqs_disabled())
4351 print_irqtrace_events(prev
);
4360 * Various schedule()-time debugging checks and statistics:
4362 static inline void schedule_debug(struct task_struct
*prev
)
4365 * Test if we are atomic. Since do_exit() needs to call into
4366 * schedule() atomically, we ignore that path for now.
4367 * Otherwise, whine if we are scheduling when we should not be.
4369 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
4370 __schedule_bug(prev
);
4372 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4374 schedstat_inc(this_rq(), sched_count
);
4375 #ifdef CONFIG_SCHEDSTATS
4376 if (unlikely(prev
->lock_depth
>= 0)) {
4377 schedstat_inc(this_rq(), bkl_count
);
4378 schedstat_inc(prev
, sched_info
.bkl_count
);
4384 * Pick up the highest-prio task:
4386 static inline struct task_struct
*
4387 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
4389 const struct sched_class
*class;
4390 struct task_struct
*p
;
4393 * Optimization: we know that if all tasks are in
4394 * the fair class we can call that function directly:
4396 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
4397 p
= fair_sched_class
.pick_next_task(rq
);
4402 class = sched_class_highest
;
4404 p
= class->pick_next_task(rq
);
4408 * Will never be NULL as the idle class always
4409 * returns a non-NULL p:
4411 class = class->next
;
4416 * schedule() is the main scheduler function.
4418 asmlinkage
void __sched
schedule(void)
4420 struct task_struct
*prev
, *next
;
4421 unsigned long *switch_count
;
4427 cpu
= smp_processor_id();
4431 switch_count
= &prev
->nivcsw
;
4433 release_kernel_lock(prev
);
4434 need_resched_nonpreemptible
:
4436 schedule_debug(prev
);
4438 if (sched_feat(HRTICK
))
4441 spin_lock_irq(&rq
->lock
);
4442 update_rq_clock(rq
);
4443 clear_tsk_need_resched(prev
);
4445 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
4446 if (unlikely(signal_pending_state(prev
->state
, prev
)))
4447 prev
->state
= TASK_RUNNING
;
4449 deactivate_task(rq
, prev
, 1);
4450 switch_count
= &prev
->nvcsw
;
4454 if (prev
->sched_class
->pre_schedule
)
4455 prev
->sched_class
->pre_schedule(rq
, prev
);
4458 if (unlikely(!rq
->nr_running
))
4459 idle_balance(cpu
, rq
);
4461 prev
->sched_class
->put_prev_task(rq
, prev
);
4462 next
= pick_next_task(rq
, prev
);
4464 if (likely(prev
!= next
)) {
4465 sched_info_switch(prev
, next
);
4471 context_switch(rq
, prev
, next
); /* unlocks the rq */
4473 * the context switch might have flipped the stack from under
4474 * us, hence refresh the local variables.
4476 cpu
= smp_processor_id();
4479 spin_unlock_irq(&rq
->lock
);
4481 if (unlikely(reacquire_kernel_lock(current
) < 0))
4482 goto need_resched_nonpreemptible
;
4484 preempt_enable_no_resched();
4485 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
4488 EXPORT_SYMBOL(schedule
);
4490 #ifdef CONFIG_PREEMPT
4492 * this is the entry point to schedule() from in-kernel preemption
4493 * off of preempt_enable. Kernel preemptions off return from interrupt
4494 * occur there and call schedule directly.
4496 asmlinkage
void __sched
preempt_schedule(void)
4498 struct thread_info
*ti
= current_thread_info();
4501 * If there is a non-zero preempt_count or interrupts are disabled,
4502 * we do not want to preempt the current task. Just return..
4504 if (likely(ti
->preempt_count
|| irqs_disabled()))
4508 add_preempt_count(PREEMPT_ACTIVE
);
4510 sub_preempt_count(PREEMPT_ACTIVE
);
4513 * Check again in case we missed a preemption opportunity
4514 * between schedule and now.
4517 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4519 EXPORT_SYMBOL(preempt_schedule
);
4522 * this is the entry point to schedule() from kernel preemption
4523 * off of irq context.
4524 * Note, that this is called and return with irqs disabled. This will
4525 * protect us against recursive calling from irq.
4527 asmlinkage
void __sched
preempt_schedule_irq(void)
4529 struct thread_info
*ti
= current_thread_info();
4531 /* Catch callers which need to be fixed */
4532 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
4535 add_preempt_count(PREEMPT_ACTIVE
);
4538 local_irq_disable();
4539 sub_preempt_count(PREEMPT_ACTIVE
);
4542 * Check again in case we missed a preemption opportunity
4543 * between schedule and now.
4546 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4549 #endif /* CONFIG_PREEMPT */
4551 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
4554 return try_to_wake_up(curr
->private, mode
, sync
);
4556 EXPORT_SYMBOL(default_wake_function
);
4559 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4560 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4561 * number) then we wake all the non-exclusive tasks and one exclusive task.
4563 * There are circumstances in which we can try to wake a task which has already
4564 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4565 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4567 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4568 int nr_exclusive
, int sync
, void *key
)
4570 wait_queue_t
*curr
, *next
;
4572 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4573 unsigned flags
= curr
->flags
;
4575 if (curr
->func(curr
, mode
, sync
, key
) &&
4576 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4582 * __wake_up - wake up threads blocked on a waitqueue.
4584 * @mode: which threads
4585 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4586 * @key: is directly passed to the wakeup function
4588 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4589 int nr_exclusive
, void *key
)
4591 unsigned long flags
;
4593 spin_lock_irqsave(&q
->lock
, flags
);
4594 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4595 spin_unlock_irqrestore(&q
->lock
, flags
);
4597 EXPORT_SYMBOL(__wake_up
);
4600 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4602 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4604 __wake_up_common(q
, mode
, 1, 0, NULL
);
4608 * __wake_up_sync - wake up threads blocked on a waitqueue.
4610 * @mode: which threads
4611 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4613 * The sync wakeup differs that the waker knows that it will schedule
4614 * away soon, so while the target thread will be woken up, it will not
4615 * be migrated to another CPU - ie. the two threads are 'synchronized'
4616 * with each other. This can prevent needless bouncing between CPUs.
4618 * On UP it can prevent extra preemption.
4621 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4623 unsigned long flags
;
4629 if (unlikely(!nr_exclusive
))
4632 spin_lock_irqsave(&q
->lock
, flags
);
4633 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
4634 spin_unlock_irqrestore(&q
->lock
, flags
);
4636 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4639 * complete: - signals a single thread waiting on this completion
4640 * @x: holds the state of this particular completion
4642 * This will wake up a single thread waiting on this completion. Threads will be
4643 * awakened in the same order in which they were queued.
4645 * See also complete_all(), wait_for_completion() and related routines.
4647 void complete(struct completion
*x
)
4649 unsigned long flags
;
4651 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4653 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4654 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4656 EXPORT_SYMBOL(complete
);
4659 * complete_all: - signals all threads waiting on this completion
4660 * @x: holds the state of this particular completion
4662 * This will wake up all threads waiting on this particular completion event.
4664 void complete_all(struct completion
*x
)
4666 unsigned long flags
;
4668 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4669 x
->done
+= UINT_MAX
/2;
4670 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4671 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4673 EXPORT_SYMBOL(complete_all
);
4675 static inline long __sched
4676 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4679 DECLARE_WAITQUEUE(wait
, current
);
4681 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4682 __add_wait_queue_tail(&x
->wait
, &wait
);
4684 if (signal_pending_state(state
, current
)) {
4685 timeout
= -ERESTARTSYS
;
4688 __set_current_state(state
);
4689 spin_unlock_irq(&x
->wait
.lock
);
4690 timeout
= schedule_timeout(timeout
);
4691 spin_lock_irq(&x
->wait
.lock
);
4692 } while (!x
->done
&& timeout
);
4693 __remove_wait_queue(&x
->wait
, &wait
);
4698 return timeout
?: 1;
4702 wait_for_common(struct completion
*x
, long timeout
, int state
)
4706 spin_lock_irq(&x
->wait
.lock
);
4707 timeout
= do_wait_for_common(x
, timeout
, state
);
4708 spin_unlock_irq(&x
->wait
.lock
);
4713 * wait_for_completion: - waits for completion of a task
4714 * @x: holds the state of this particular completion
4716 * This waits to be signaled for completion of a specific task. It is NOT
4717 * interruptible and there is no timeout.
4719 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4720 * and interrupt capability. Also see complete().
4722 void __sched
wait_for_completion(struct completion
*x
)
4724 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4726 EXPORT_SYMBOL(wait_for_completion
);
4729 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4730 * @x: holds the state of this particular completion
4731 * @timeout: timeout value in jiffies
4733 * This waits for either a completion of a specific task to be signaled or for a
4734 * specified timeout to expire. The timeout is in jiffies. It is not
4737 unsigned long __sched
4738 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4740 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4742 EXPORT_SYMBOL(wait_for_completion_timeout
);
4745 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4746 * @x: holds the state of this particular completion
4748 * This waits for completion of a specific task to be signaled. It is
4751 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4753 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4754 if (t
== -ERESTARTSYS
)
4758 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4761 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4762 * @x: holds the state of this particular completion
4763 * @timeout: timeout value in jiffies
4765 * This waits for either a completion of a specific task to be signaled or for a
4766 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4768 unsigned long __sched
4769 wait_for_completion_interruptible_timeout(struct completion
*x
,
4770 unsigned long timeout
)
4772 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4774 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4777 * wait_for_completion_killable: - waits for completion of a task (killable)
4778 * @x: holds the state of this particular completion
4780 * This waits to be signaled for completion of a specific task. It can be
4781 * interrupted by a kill signal.
4783 int __sched
wait_for_completion_killable(struct completion
*x
)
4785 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4786 if (t
== -ERESTARTSYS
)
4790 EXPORT_SYMBOL(wait_for_completion_killable
);
4793 * try_wait_for_completion - try to decrement a completion without blocking
4794 * @x: completion structure
4796 * Returns: 0 if a decrement cannot be done without blocking
4797 * 1 if a decrement succeeded.
4799 * If a completion is being used as a counting completion,
4800 * attempt to decrement the counter without blocking. This
4801 * enables us to avoid waiting if the resource the completion
4802 * is protecting is not available.
4804 bool try_wait_for_completion(struct completion
*x
)
4808 spin_lock_irq(&x
->wait
.lock
);
4813 spin_unlock_irq(&x
->wait
.lock
);
4816 EXPORT_SYMBOL(try_wait_for_completion
);
4819 * completion_done - Test to see if a completion has any waiters
4820 * @x: completion structure
4822 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4823 * 1 if there are no waiters.
4826 bool completion_done(struct completion
*x
)
4830 spin_lock_irq(&x
->wait
.lock
);
4833 spin_unlock_irq(&x
->wait
.lock
);
4836 EXPORT_SYMBOL(completion_done
);
4839 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4841 unsigned long flags
;
4844 init_waitqueue_entry(&wait
, current
);
4846 __set_current_state(state
);
4848 spin_lock_irqsave(&q
->lock
, flags
);
4849 __add_wait_queue(q
, &wait
);
4850 spin_unlock(&q
->lock
);
4851 timeout
= schedule_timeout(timeout
);
4852 spin_lock_irq(&q
->lock
);
4853 __remove_wait_queue(q
, &wait
);
4854 spin_unlock_irqrestore(&q
->lock
, flags
);
4859 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4861 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4863 EXPORT_SYMBOL(interruptible_sleep_on
);
4866 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4868 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4870 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4872 void __sched
sleep_on(wait_queue_head_t
*q
)
4874 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4876 EXPORT_SYMBOL(sleep_on
);
4878 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4880 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4882 EXPORT_SYMBOL(sleep_on_timeout
);
4884 #ifdef CONFIG_RT_MUTEXES
4887 * rt_mutex_setprio - set the current priority of a task
4889 * @prio: prio value (kernel-internal form)
4891 * This function changes the 'effective' priority of a task. It does
4892 * not touch ->normal_prio like __setscheduler().
4894 * Used by the rt_mutex code to implement priority inheritance logic.
4896 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4898 unsigned long flags
;
4899 int oldprio
, on_rq
, running
;
4901 const struct sched_class
*prev_class
= p
->sched_class
;
4903 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4905 rq
= task_rq_lock(p
, &flags
);
4906 update_rq_clock(rq
);
4909 on_rq
= p
->se
.on_rq
;
4910 running
= task_current(rq
, p
);
4912 dequeue_task(rq
, p
, 0);
4914 p
->sched_class
->put_prev_task(rq
, p
);
4917 p
->sched_class
= &rt_sched_class
;
4919 p
->sched_class
= &fair_sched_class
;
4924 p
->sched_class
->set_curr_task(rq
);
4926 enqueue_task(rq
, p
, 0);
4928 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4930 task_rq_unlock(rq
, &flags
);
4935 void set_user_nice(struct task_struct
*p
, long nice
)
4937 int old_prio
, delta
, on_rq
;
4938 unsigned long flags
;
4941 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4944 * We have to be careful, if called from sys_setpriority(),
4945 * the task might be in the middle of scheduling on another CPU.
4947 rq
= task_rq_lock(p
, &flags
);
4948 update_rq_clock(rq
);
4950 * The RT priorities are set via sched_setscheduler(), but we still
4951 * allow the 'normal' nice value to be set - but as expected
4952 * it wont have any effect on scheduling until the task is
4953 * SCHED_FIFO/SCHED_RR:
4955 if (task_has_rt_policy(p
)) {
4956 p
->static_prio
= NICE_TO_PRIO(nice
);
4959 on_rq
= p
->se
.on_rq
;
4961 dequeue_task(rq
, p
, 0);
4963 p
->static_prio
= NICE_TO_PRIO(nice
);
4966 p
->prio
= effective_prio(p
);
4967 delta
= p
->prio
- old_prio
;
4970 enqueue_task(rq
, p
, 0);
4972 * If the task increased its priority or is running and
4973 * lowered its priority, then reschedule its CPU:
4975 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4976 resched_task(rq
->curr
);
4979 task_rq_unlock(rq
, &flags
);
4981 EXPORT_SYMBOL(set_user_nice
);
4984 * can_nice - check if a task can reduce its nice value
4988 int can_nice(const struct task_struct
*p
, const int nice
)
4990 /* convert nice value [19,-20] to rlimit style value [1,40] */
4991 int nice_rlim
= 20 - nice
;
4993 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4994 capable(CAP_SYS_NICE
));
4997 #ifdef __ARCH_WANT_SYS_NICE
5000 * sys_nice - change the priority of the current process.
5001 * @increment: priority increment
5003 * sys_setpriority is a more generic, but much slower function that
5004 * does similar things.
5006 asmlinkage
long sys_nice(int increment
)
5011 * Setpriority might change our priority at the same moment.
5012 * We don't have to worry. Conceptually one call occurs first
5013 * and we have a single winner.
5015 if (increment
< -40)
5020 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
5026 if (increment
< 0 && !can_nice(current
, nice
))
5029 retval
= security_task_setnice(current
, nice
);
5033 set_user_nice(current
, nice
);
5040 * task_prio - return the priority value of a given task.
5041 * @p: the task in question.
5043 * This is the priority value as seen by users in /proc.
5044 * RT tasks are offset by -200. Normal tasks are centered
5045 * around 0, value goes from -16 to +15.
5047 int task_prio(const struct task_struct
*p
)
5049 return p
->prio
- MAX_RT_PRIO
;
5053 * task_nice - return the nice value of a given task.
5054 * @p: the task in question.
5056 int task_nice(const struct task_struct
*p
)
5058 return TASK_NICE(p
);
5060 EXPORT_SYMBOL(task_nice
);
5063 * idle_cpu - is a given cpu idle currently?
5064 * @cpu: the processor in question.
5066 int idle_cpu(int cpu
)
5068 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
5072 * idle_task - return the idle task for a given cpu.
5073 * @cpu: the processor in question.
5075 struct task_struct
*idle_task(int cpu
)
5077 return cpu_rq(cpu
)->idle
;
5081 * find_process_by_pid - find a process with a matching PID value.
5082 * @pid: the pid in question.
5084 static struct task_struct
*find_process_by_pid(pid_t pid
)
5086 return pid
? find_task_by_vpid(pid
) : current
;
5089 /* Actually do priority change: must hold rq lock. */
5091 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
5093 BUG_ON(p
->se
.on_rq
);
5096 switch (p
->policy
) {
5100 p
->sched_class
= &fair_sched_class
;
5104 p
->sched_class
= &rt_sched_class
;
5108 p
->rt_priority
= prio
;
5109 p
->normal_prio
= normal_prio(p
);
5110 /* we are holding p->pi_lock already */
5111 p
->prio
= rt_mutex_getprio(p
);
5115 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
5116 struct sched_param
*param
, bool user
)
5118 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
5119 unsigned long flags
;
5120 const struct sched_class
*prev_class
= p
->sched_class
;
5123 /* may grab non-irq protected spin_locks */
5124 BUG_ON(in_interrupt());
5126 /* double check policy once rq lock held */
5128 policy
= oldpolicy
= p
->policy
;
5129 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
5130 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
5131 policy
!= SCHED_IDLE
)
5134 * Valid priorities for SCHED_FIFO and SCHED_RR are
5135 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5136 * SCHED_BATCH and SCHED_IDLE is 0.
5138 if (param
->sched_priority
< 0 ||
5139 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
5140 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
5142 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
5146 * Allow unprivileged RT tasks to decrease priority:
5148 if (user
&& !capable(CAP_SYS_NICE
)) {
5149 if (rt_policy(policy
)) {
5150 unsigned long rlim_rtprio
;
5152 if (!lock_task_sighand(p
, &flags
))
5154 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
5155 unlock_task_sighand(p
, &flags
);
5157 /* can't set/change the rt policy */
5158 if (policy
!= p
->policy
&& !rlim_rtprio
)
5161 /* can't increase priority */
5162 if (param
->sched_priority
> p
->rt_priority
&&
5163 param
->sched_priority
> rlim_rtprio
)
5167 * Like positive nice levels, dont allow tasks to
5168 * move out of SCHED_IDLE either:
5170 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
5173 /* can't change other user's priorities */
5174 if ((current
->euid
!= p
->euid
) &&
5175 (current
->euid
!= p
->uid
))
5180 #ifdef CONFIG_RT_GROUP_SCHED
5182 * Do not allow realtime tasks into groups that have no runtime
5185 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
5186 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
5190 retval
= security_task_setscheduler(p
, policy
, param
);
5196 * make sure no PI-waiters arrive (or leave) while we are
5197 * changing the priority of the task:
5199 spin_lock_irqsave(&p
->pi_lock
, flags
);
5201 * To be able to change p->policy safely, the apropriate
5202 * runqueue lock must be held.
5204 rq
= __task_rq_lock(p
);
5205 /* recheck policy now with rq lock held */
5206 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
5207 policy
= oldpolicy
= -1;
5208 __task_rq_unlock(rq
);
5209 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5212 update_rq_clock(rq
);
5213 on_rq
= p
->se
.on_rq
;
5214 running
= task_current(rq
, p
);
5216 deactivate_task(rq
, p
, 0);
5218 p
->sched_class
->put_prev_task(rq
, p
);
5221 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
5224 p
->sched_class
->set_curr_task(rq
);
5226 activate_task(rq
, p
, 0);
5228 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5230 __task_rq_unlock(rq
);
5231 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5233 rt_mutex_adjust_pi(p
);
5239 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5240 * @p: the task in question.
5241 * @policy: new policy.
5242 * @param: structure containing the new RT priority.
5244 * NOTE that the task may be already dead.
5246 int sched_setscheduler(struct task_struct
*p
, int policy
,
5247 struct sched_param
*param
)
5249 return __sched_setscheduler(p
, policy
, param
, true);
5251 EXPORT_SYMBOL_GPL(sched_setscheduler
);
5254 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5255 * @p: the task in question.
5256 * @policy: new policy.
5257 * @param: structure containing the new RT priority.
5259 * Just like sched_setscheduler, only don't bother checking if the
5260 * current context has permission. For example, this is needed in
5261 * stop_machine(): we create temporary high priority worker threads,
5262 * but our caller might not have that capability.
5264 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
5265 struct sched_param
*param
)
5267 return __sched_setscheduler(p
, policy
, param
, false);
5271 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5273 struct sched_param lparam
;
5274 struct task_struct
*p
;
5277 if (!param
|| pid
< 0)
5279 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
5284 p
= find_process_by_pid(pid
);
5286 retval
= sched_setscheduler(p
, policy
, &lparam
);
5293 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5294 * @pid: the pid in question.
5295 * @policy: new policy.
5296 * @param: structure containing the new RT priority.
5299 sys_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5301 /* negative values for policy are not valid */
5305 return do_sched_setscheduler(pid
, policy
, param
);
5309 * sys_sched_setparam - set/change the RT priority of a thread
5310 * @pid: the pid in question.
5311 * @param: structure containing the new RT priority.
5313 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
5315 return do_sched_setscheduler(pid
, -1, param
);
5319 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5320 * @pid: the pid in question.
5322 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
5324 struct task_struct
*p
;
5331 read_lock(&tasklist_lock
);
5332 p
= find_process_by_pid(pid
);
5334 retval
= security_task_getscheduler(p
);
5338 read_unlock(&tasklist_lock
);
5343 * sys_sched_getscheduler - get the RT priority of a thread
5344 * @pid: the pid in question.
5345 * @param: structure containing the RT priority.
5347 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
5349 struct sched_param lp
;
5350 struct task_struct
*p
;
5353 if (!param
|| pid
< 0)
5356 read_lock(&tasklist_lock
);
5357 p
= find_process_by_pid(pid
);
5362 retval
= security_task_getscheduler(p
);
5366 lp
.sched_priority
= p
->rt_priority
;
5367 read_unlock(&tasklist_lock
);
5370 * This one might sleep, we cannot do it with a spinlock held ...
5372 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
5377 read_unlock(&tasklist_lock
);
5381 long sched_setaffinity(pid_t pid
, const cpumask_t
*in_mask
)
5383 cpumask_t cpus_allowed
;
5384 cpumask_t new_mask
= *in_mask
;
5385 struct task_struct
*p
;
5389 read_lock(&tasklist_lock
);
5391 p
= find_process_by_pid(pid
);
5393 read_unlock(&tasklist_lock
);
5399 * It is not safe to call set_cpus_allowed with the
5400 * tasklist_lock held. We will bump the task_struct's
5401 * usage count and then drop tasklist_lock.
5404 read_unlock(&tasklist_lock
);
5407 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
5408 !capable(CAP_SYS_NICE
))
5411 retval
= security_task_setscheduler(p
, 0, NULL
);
5415 cpuset_cpus_allowed(p
, &cpus_allowed
);
5416 cpus_and(new_mask
, new_mask
, cpus_allowed
);
5418 retval
= set_cpus_allowed_ptr(p
, &new_mask
);
5421 cpuset_cpus_allowed(p
, &cpus_allowed
);
5422 if (!cpus_subset(new_mask
, cpus_allowed
)) {
5424 * We must have raced with a concurrent cpuset
5425 * update. Just reset the cpus_allowed to the
5426 * cpuset's cpus_allowed
5428 new_mask
= cpus_allowed
;
5438 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5439 cpumask_t
*new_mask
)
5441 if (len
< sizeof(cpumask_t
)) {
5442 memset(new_mask
, 0, sizeof(cpumask_t
));
5443 } else if (len
> sizeof(cpumask_t
)) {
5444 len
= sizeof(cpumask_t
);
5446 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
5450 * sys_sched_setaffinity - set the cpu affinity of a process
5451 * @pid: pid of the process
5452 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5453 * @user_mask_ptr: user-space pointer to the new cpu mask
5455 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
5456 unsigned long __user
*user_mask_ptr
)
5461 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
5465 return sched_setaffinity(pid
, &new_mask
);
5468 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
5470 struct task_struct
*p
;
5474 read_lock(&tasklist_lock
);
5477 p
= find_process_by_pid(pid
);
5481 retval
= security_task_getscheduler(p
);
5485 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
5488 read_unlock(&tasklist_lock
);
5495 * sys_sched_getaffinity - get the cpu affinity of a process
5496 * @pid: pid of the process
5497 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5498 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5500 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
5501 unsigned long __user
*user_mask_ptr
)
5506 if (len
< sizeof(cpumask_t
))
5509 ret
= sched_getaffinity(pid
, &mask
);
5513 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
5516 return sizeof(cpumask_t
);
5520 * sys_sched_yield - yield the current processor to other threads.
5522 * This function yields the current CPU to other tasks. If there are no
5523 * other threads running on this CPU then this function will return.
5525 asmlinkage
long sys_sched_yield(void)
5527 struct rq
*rq
= this_rq_lock();
5529 schedstat_inc(rq
, yld_count
);
5530 current
->sched_class
->yield_task(rq
);
5533 * Since we are going to call schedule() anyway, there's
5534 * no need to preempt or enable interrupts:
5536 __release(rq
->lock
);
5537 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5538 _raw_spin_unlock(&rq
->lock
);
5539 preempt_enable_no_resched();
5546 static void __cond_resched(void)
5548 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5549 __might_sleep(__FILE__
, __LINE__
);
5552 * The BKS might be reacquired before we have dropped
5553 * PREEMPT_ACTIVE, which could trigger a second
5554 * cond_resched() call.
5557 add_preempt_count(PREEMPT_ACTIVE
);
5559 sub_preempt_count(PREEMPT_ACTIVE
);
5560 } while (need_resched());
5563 int __sched
_cond_resched(void)
5565 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
5566 system_state
== SYSTEM_RUNNING
) {
5572 EXPORT_SYMBOL(_cond_resched
);
5575 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5576 * call schedule, and on return reacquire the lock.
5578 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5579 * operations here to prevent schedule() from being called twice (once via
5580 * spin_unlock(), once by hand).
5582 int cond_resched_lock(spinlock_t
*lock
)
5584 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
5587 if (spin_needbreak(lock
) || resched
) {
5589 if (resched
&& need_resched())
5598 EXPORT_SYMBOL(cond_resched_lock
);
5600 int __sched
cond_resched_softirq(void)
5602 BUG_ON(!in_softirq());
5604 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
5612 EXPORT_SYMBOL(cond_resched_softirq
);
5615 * yield - yield the current processor to other threads.
5617 * This is a shortcut for kernel-space yielding - it marks the
5618 * thread runnable and calls sys_sched_yield().
5620 void __sched
yield(void)
5622 set_current_state(TASK_RUNNING
);
5625 EXPORT_SYMBOL(yield
);
5628 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5629 * that process accounting knows that this is a task in IO wait state.
5631 * But don't do that if it is a deliberate, throttling IO wait (this task
5632 * has set its backing_dev_info: the queue against which it should throttle)
5634 void __sched
io_schedule(void)
5636 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5638 delayacct_blkio_start();
5639 atomic_inc(&rq
->nr_iowait
);
5641 atomic_dec(&rq
->nr_iowait
);
5642 delayacct_blkio_end();
5644 EXPORT_SYMBOL(io_schedule
);
5646 long __sched
io_schedule_timeout(long timeout
)
5648 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5651 delayacct_blkio_start();
5652 atomic_inc(&rq
->nr_iowait
);
5653 ret
= schedule_timeout(timeout
);
5654 atomic_dec(&rq
->nr_iowait
);
5655 delayacct_blkio_end();
5660 * sys_sched_get_priority_max - return maximum RT priority.
5661 * @policy: scheduling class.
5663 * this syscall returns the maximum rt_priority that can be used
5664 * by a given scheduling class.
5666 asmlinkage
long sys_sched_get_priority_max(int policy
)
5673 ret
= MAX_USER_RT_PRIO
-1;
5685 * sys_sched_get_priority_min - return minimum RT priority.
5686 * @policy: scheduling class.
5688 * this syscall returns the minimum rt_priority that can be used
5689 * by a given scheduling class.
5691 asmlinkage
long sys_sched_get_priority_min(int policy
)
5709 * sys_sched_rr_get_interval - return the default timeslice of a process.
5710 * @pid: pid of the process.
5711 * @interval: userspace pointer to the timeslice value.
5713 * this syscall writes the default timeslice value of a given process
5714 * into the user-space timespec buffer. A value of '0' means infinity.
5717 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
5719 struct task_struct
*p
;
5720 unsigned int time_slice
;
5728 read_lock(&tasklist_lock
);
5729 p
= find_process_by_pid(pid
);
5733 retval
= security_task_getscheduler(p
);
5738 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5739 * tasks that are on an otherwise idle runqueue:
5742 if (p
->policy
== SCHED_RR
) {
5743 time_slice
= DEF_TIMESLICE
;
5744 } else if (p
->policy
!= SCHED_FIFO
) {
5745 struct sched_entity
*se
= &p
->se
;
5746 unsigned long flags
;
5749 rq
= task_rq_lock(p
, &flags
);
5750 if (rq
->cfs
.load
.weight
)
5751 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
5752 task_rq_unlock(rq
, &flags
);
5754 read_unlock(&tasklist_lock
);
5755 jiffies_to_timespec(time_slice
, &t
);
5756 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5760 read_unlock(&tasklist_lock
);
5764 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5766 void sched_show_task(struct task_struct
*p
)
5768 unsigned long free
= 0;
5771 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5772 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5773 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5774 #if BITS_PER_LONG == 32
5775 if (state
== TASK_RUNNING
)
5776 printk(KERN_CONT
" running ");
5778 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5780 if (state
== TASK_RUNNING
)
5781 printk(KERN_CONT
" running task ");
5783 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5785 #ifdef CONFIG_DEBUG_STACK_USAGE
5787 unsigned long *n
= end_of_stack(p
);
5790 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
5793 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
5794 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
5796 show_stack(p
, NULL
);
5799 void show_state_filter(unsigned long state_filter
)
5801 struct task_struct
*g
, *p
;
5803 #if BITS_PER_LONG == 32
5805 " task PC stack pid father\n");
5808 " task PC stack pid father\n");
5810 read_lock(&tasklist_lock
);
5811 do_each_thread(g
, p
) {
5813 * reset the NMI-timeout, listing all files on a slow
5814 * console might take alot of time:
5816 touch_nmi_watchdog();
5817 if (!state_filter
|| (p
->state
& state_filter
))
5819 } while_each_thread(g
, p
);
5821 touch_all_softlockup_watchdogs();
5823 #ifdef CONFIG_SCHED_DEBUG
5824 sysrq_sched_debug_show();
5826 read_unlock(&tasklist_lock
);
5828 * Only show locks if all tasks are dumped:
5830 if (state_filter
== -1)
5831 debug_show_all_locks();
5834 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5836 idle
->sched_class
= &idle_sched_class
;
5840 * init_idle - set up an idle thread for a given CPU
5841 * @idle: task in question
5842 * @cpu: cpu the idle task belongs to
5844 * NOTE: this function does not set the idle thread's NEED_RESCHED
5845 * flag, to make booting more robust.
5847 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5849 struct rq
*rq
= cpu_rq(cpu
);
5850 unsigned long flags
;
5852 spin_lock_irqsave(&rq
->lock
, flags
);
5855 idle
->se
.exec_start
= sched_clock();
5857 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
5858 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
5859 __set_task_cpu(idle
, cpu
);
5861 rq
->curr
= rq
->idle
= idle
;
5862 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5865 spin_unlock_irqrestore(&rq
->lock
, flags
);
5867 /* Set the preempt count _outside_ the spinlocks! */
5868 #if defined(CONFIG_PREEMPT)
5869 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5871 task_thread_info(idle
)->preempt_count
= 0;
5874 * The idle tasks have their own, simple scheduling class:
5876 idle
->sched_class
= &idle_sched_class
;
5880 * In a system that switches off the HZ timer nohz_cpu_mask
5881 * indicates which cpus entered this state. This is used
5882 * in the rcu update to wait only for active cpus. For system
5883 * which do not switch off the HZ timer nohz_cpu_mask should
5884 * always be CPU_MASK_NONE.
5886 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
5889 * Increase the granularity value when there are more CPUs,
5890 * because with more CPUs the 'effective latency' as visible
5891 * to users decreases. But the relationship is not linear,
5892 * so pick a second-best guess by going with the log2 of the
5895 * This idea comes from the SD scheduler of Con Kolivas:
5897 static inline void sched_init_granularity(void)
5899 unsigned int factor
= 1 + ilog2(num_online_cpus());
5900 const unsigned long limit
= 200000000;
5902 sysctl_sched_min_granularity
*= factor
;
5903 if (sysctl_sched_min_granularity
> limit
)
5904 sysctl_sched_min_granularity
= limit
;
5906 sysctl_sched_latency
*= factor
;
5907 if (sysctl_sched_latency
> limit
)
5908 sysctl_sched_latency
= limit
;
5910 sysctl_sched_wakeup_granularity
*= factor
;
5912 sysctl_sched_shares_ratelimit
*= factor
;
5917 * This is how migration works:
5919 * 1) we queue a struct migration_req structure in the source CPU's
5920 * runqueue and wake up that CPU's migration thread.
5921 * 2) we down() the locked semaphore => thread blocks.
5922 * 3) migration thread wakes up (implicitly it forces the migrated
5923 * thread off the CPU)
5924 * 4) it gets the migration request and checks whether the migrated
5925 * task is still in the wrong runqueue.
5926 * 5) if it's in the wrong runqueue then the migration thread removes
5927 * it and puts it into the right queue.
5928 * 6) migration thread up()s the semaphore.
5929 * 7) we wake up and the migration is done.
5933 * Change a given task's CPU affinity. Migrate the thread to a
5934 * proper CPU and schedule it away if the CPU it's executing on
5935 * is removed from the allowed bitmask.
5937 * NOTE: the caller must have a valid reference to the task, the
5938 * task must not exit() & deallocate itself prematurely. The
5939 * call is not atomic; no spinlocks may be held.
5941 int set_cpus_allowed_ptr(struct task_struct
*p
, const cpumask_t
*new_mask
)
5943 struct migration_req req
;
5944 unsigned long flags
;
5948 rq
= task_rq_lock(p
, &flags
);
5949 if (!cpus_intersects(*new_mask
, cpu_online_map
)) {
5954 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
5955 !cpus_equal(p
->cpus_allowed
, *new_mask
))) {
5960 if (p
->sched_class
->set_cpus_allowed
)
5961 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
5963 p
->cpus_allowed
= *new_mask
;
5964 p
->rt
.nr_cpus_allowed
= cpus_weight(*new_mask
);
5967 /* Can the task run on the task's current CPU? If so, we're done */
5968 if (cpu_isset(task_cpu(p
), *new_mask
))
5971 if (migrate_task(p
, any_online_cpu(*new_mask
), &req
)) {
5972 /* Need help from migration thread: drop lock and wait. */
5973 task_rq_unlock(rq
, &flags
);
5974 wake_up_process(rq
->migration_thread
);
5975 wait_for_completion(&req
.done
);
5976 tlb_migrate_finish(p
->mm
);
5980 task_rq_unlock(rq
, &flags
);
5984 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5987 * Move (not current) task off this cpu, onto dest cpu. We're doing
5988 * this because either it can't run here any more (set_cpus_allowed()
5989 * away from this CPU, or CPU going down), or because we're
5990 * attempting to rebalance this task on exec (sched_exec).
5992 * So we race with normal scheduler movements, but that's OK, as long
5993 * as the task is no longer on this CPU.
5995 * Returns non-zero if task was successfully migrated.
5997 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5999 struct rq
*rq_dest
, *rq_src
;
6002 if (unlikely(!cpu_active(dest_cpu
)))
6005 rq_src
= cpu_rq(src_cpu
);
6006 rq_dest
= cpu_rq(dest_cpu
);
6008 double_rq_lock(rq_src
, rq_dest
);
6009 /* Already moved. */
6010 if (task_cpu(p
) != src_cpu
)
6012 /* Affinity changed (again). */
6013 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
6016 on_rq
= p
->se
.on_rq
;
6018 deactivate_task(rq_src
, p
, 0);
6020 set_task_cpu(p
, dest_cpu
);
6022 activate_task(rq_dest
, p
, 0);
6023 check_preempt_curr(rq_dest
, p
, 0);
6028 double_rq_unlock(rq_src
, rq_dest
);
6033 * migration_thread - this is a highprio system thread that performs
6034 * thread migration by bumping thread off CPU then 'pushing' onto
6037 static int migration_thread(void *data
)
6039 int cpu
= (long)data
;
6043 BUG_ON(rq
->migration_thread
!= current
);
6045 set_current_state(TASK_INTERRUPTIBLE
);
6046 while (!kthread_should_stop()) {
6047 struct migration_req
*req
;
6048 struct list_head
*head
;
6050 spin_lock_irq(&rq
->lock
);
6052 if (cpu_is_offline(cpu
)) {
6053 spin_unlock_irq(&rq
->lock
);
6057 if (rq
->active_balance
) {
6058 active_load_balance(rq
, cpu
);
6059 rq
->active_balance
= 0;
6062 head
= &rq
->migration_queue
;
6064 if (list_empty(head
)) {
6065 spin_unlock_irq(&rq
->lock
);
6067 set_current_state(TASK_INTERRUPTIBLE
);
6070 req
= list_entry(head
->next
, struct migration_req
, list
);
6071 list_del_init(head
->next
);
6073 spin_unlock(&rq
->lock
);
6074 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
6077 complete(&req
->done
);
6079 __set_current_state(TASK_RUNNING
);
6083 /* Wait for kthread_stop */
6084 set_current_state(TASK_INTERRUPTIBLE
);
6085 while (!kthread_should_stop()) {
6087 set_current_state(TASK_INTERRUPTIBLE
);
6089 __set_current_state(TASK_RUNNING
);
6093 #ifdef CONFIG_HOTPLUG_CPU
6095 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6099 local_irq_disable();
6100 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
6106 * Figure out where task on dead CPU should go, use force if necessary.
6108 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
6110 unsigned long flags
;
6117 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
6118 cpus_and(mask
, mask
, p
->cpus_allowed
);
6119 dest_cpu
= any_online_cpu(mask
);
6121 /* On any allowed CPU? */
6122 if (dest_cpu
>= nr_cpu_ids
)
6123 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
6125 /* No more Mr. Nice Guy. */
6126 if (dest_cpu
>= nr_cpu_ids
) {
6127 cpumask_t cpus_allowed
;
6129 cpuset_cpus_allowed_locked(p
, &cpus_allowed
);
6131 * Try to stay on the same cpuset, where the
6132 * current cpuset may be a subset of all cpus.
6133 * The cpuset_cpus_allowed_locked() variant of
6134 * cpuset_cpus_allowed() will not block. It must be
6135 * called within calls to cpuset_lock/cpuset_unlock.
6137 rq
= task_rq_lock(p
, &flags
);
6138 p
->cpus_allowed
= cpus_allowed
;
6139 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
6140 task_rq_unlock(rq
, &flags
);
6143 * Don't tell them about moving exiting tasks or
6144 * kernel threads (both mm NULL), since they never
6147 if (p
->mm
&& printk_ratelimit()) {
6148 printk(KERN_INFO
"process %d (%s) no "
6149 "longer affine to cpu%d\n",
6150 task_pid_nr(p
), p
->comm
, dead_cpu
);
6153 } while (!__migrate_task_irq(p
, dead_cpu
, dest_cpu
));
6157 * While a dead CPU has no uninterruptible tasks queued at this point,
6158 * it might still have a nonzero ->nr_uninterruptible counter, because
6159 * for performance reasons the counter is not stricly tracking tasks to
6160 * their home CPUs. So we just add the counter to another CPU's counter,
6161 * to keep the global sum constant after CPU-down:
6163 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
6165 struct rq
*rq_dest
= cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR
));
6166 unsigned long flags
;
6168 local_irq_save(flags
);
6169 double_rq_lock(rq_src
, rq_dest
);
6170 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
6171 rq_src
->nr_uninterruptible
= 0;
6172 double_rq_unlock(rq_src
, rq_dest
);
6173 local_irq_restore(flags
);
6176 /* Run through task list and migrate tasks from the dead cpu. */
6177 static void migrate_live_tasks(int src_cpu
)
6179 struct task_struct
*p
, *t
;
6181 read_lock(&tasklist_lock
);
6183 do_each_thread(t
, p
) {
6187 if (task_cpu(p
) == src_cpu
)
6188 move_task_off_dead_cpu(src_cpu
, p
);
6189 } while_each_thread(t
, p
);
6191 read_unlock(&tasklist_lock
);
6195 * Schedules idle task to be the next runnable task on current CPU.
6196 * It does so by boosting its priority to highest possible.
6197 * Used by CPU offline code.
6199 void sched_idle_next(void)
6201 int this_cpu
= smp_processor_id();
6202 struct rq
*rq
= cpu_rq(this_cpu
);
6203 struct task_struct
*p
= rq
->idle
;
6204 unsigned long flags
;
6206 /* cpu has to be offline */
6207 BUG_ON(cpu_online(this_cpu
));
6210 * Strictly not necessary since rest of the CPUs are stopped by now
6211 * and interrupts disabled on the current cpu.
6213 spin_lock_irqsave(&rq
->lock
, flags
);
6215 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6217 update_rq_clock(rq
);
6218 activate_task(rq
, p
, 0);
6220 spin_unlock_irqrestore(&rq
->lock
, flags
);
6224 * Ensures that the idle task is using init_mm right before its cpu goes
6227 void idle_task_exit(void)
6229 struct mm_struct
*mm
= current
->active_mm
;
6231 BUG_ON(cpu_online(smp_processor_id()));
6234 switch_mm(mm
, &init_mm
, current
);
6238 /* called under rq->lock with disabled interrupts */
6239 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
6241 struct rq
*rq
= cpu_rq(dead_cpu
);
6243 /* Must be exiting, otherwise would be on tasklist. */
6244 BUG_ON(!p
->exit_state
);
6246 /* Cannot have done final schedule yet: would have vanished. */
6247 BUG_ON(p
->state
== TASK_DEAD
);
6252 * Drop lock around migration; if someone else moves it,
6253 * that's OK. No task can be added to this CPU, so iteration is
6256 spin_unlock_irq(&rq
->lock
);
6257 move_task_off_dead_cpu(dead_cpu
, p
);
6258 spin_lock_irq(&rq
->lock
);
6263 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6264 static void migrate_dead_tasks(unsigned int dead_cpu
)
6266 struct rq
*rq
= cpu_rq(dead_cpu
);
6267 struct task_struct
*next
;
6270 if (!rq
->nr_running
)
6272 update_rq_clock(rq
);
6273 next
= pick_next_task(rq
, rq
->curr
);
6276 next
->sched_class
->put_prev_task(rq
, next
);
6277 migrate_dead(dead_cpu
, next
);
6281 #endif /* CONFIG_HOTPLUG_CPU */
6283 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6285 static struct ctl_table sd_ctl_dir
[] = {
6287 .procname
= "sched_domain",
6293 static struct ctl_table sd_ctl_root
[] = {
6295 .ctl_name
= CTL_KERN
,
6296 .procname
= "kernel",
6298 .child
= sd_ctl_dir
,
6303 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
6305 struct ctl_table
*entry
=
6306 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
6311 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
6313 struct ctl_table
*entry
;
6316 * In the intermediate directories, both the child directory and
6317 * procname are dynamically allocated and could fail but the mode
6318 * will always be set. In the lowest directory the names are
6319 * static strings and all have proc handlers.
6321 for (entry
= *tablep
; entry
->mode
; entry
++) {
6323 sd_free_ctl_entry(&entry
->child
);
6324 if (entry
->proc_handler
== NULL
)
6325 kfree(entry
->procname
);
6333 set_table_entry(struct ctl_table
*entry
,
6334 const char *procname
, void *data
, int maxlen
,
6335 mode_t mode
, proc_handler
*proc_handler
)
6337 entry
->procname
= procname
;
6339 entry
->maxlen
= maxlen
;
6341 entry
->proc_handler
= proc_handler
;
6344 static struct ctl_table
*
6345 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
6347 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
6352 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
6353 sizeof(long), 0644, proc_doulongvec_minmax
);
6354 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
6355 sizeof(long), 0644, proc_doulongvec_minmax
);
6356 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
6357 sizeof(int), 0644, proc_dointvec_minmax
);
6358 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
6359 sizeof(int), 0644, proc_dointvec_minmax
);
6360 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
6361 sizeof(int), 0644, proc_dointvec_minmax
);
6362 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
6363 sizeof(int), 0644, proc_dointvec_minmax
);
6364 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
6365 sizeof(int), 0644, proc_dointvec_minmax
);
6366 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
6367 sizeof(int), 0644, proc_dointvec_minmax
);
6368 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
6369 sizeof(int), 0644, proc_dointvec_minmax
);
6370 set_table_entry(&table
[9], "cache_nice_tries",
6371 &sd
->cache_nice_tries
,
6372 sizeof(int), 0644, proc_dointvec_minmax
);
6373 set_table_entry(&table
[10], "flags", &sd
->flags
,
6374 sizeof(int), 0644, proc_dointvec_minmax
);
6375 set_table_entry(&table
[11], "name", sd
->name
,
6376 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
6377 /* &table[12] is terminator */
6382 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
6384 struct ctl_table
*entry
, *table
;
6385 struct sched_domain
*sd
;
6386 int domain_num
= 0, i
;
6389 for_each_domain(cpu
, sd
)
6391 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
6396 for_each_domain(cpu
, sd
) {
6397 snprintf(buf
, 32, "domain%d", i
);
6398 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6400 entry
->child
= sd_alloc_ctl_domain_table(sd
);
6407 static struct ctl_table_header
*sd_sysctl_header
;
6408 static void register_sched_domain_sysctl(void)
6410 int i
, cpu_num
= num_online_cpus();
6411 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
6414 WARN_ON(sd_ctl_dir
[0].child
);
6415 sd_ctl_dir
[0].child
= entry
;
6420 for_each_online_cpu(i
) {
6421 snprintf(buf
, 32, "cpu%d", i
);
6422 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6424 entry
->child
= sd_alloc_ctl_cpu_table(i
);
6428 WARN_ON(sd_sysctl_header
);
6429 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
6432 /* may be called multiple times per register */
6433 static void unregister_sched_domain_sysctl(void)
6435 if (sd_sysctl_header
)
6436 unregister_sysctl_table(sd_sysctl_header
);
6437 sd_sysctl_header
= NULL
;
6438 if (sd_ctl_dir
[0].child
)
6439 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
6442 static void register_sched_domain_sysctl(void)
6445 static void unregister_sched_domain_sysctl(void)
6450 static void set_rq_online(struct rq
*rq
)
6453 const struct sched_class
*class;
6455 cpu_set(rq
->cpu
, rq
->rd
->online
);
6458 for_each_class(class) {
6459 if (class->rq_online
)
6460 class->rq_online(rq
);
6465 static void set_rq_offline(struct rq
*rq
)
6468 const struct sched_class
*class;
6470 for_each_class(class) {
6471 if (class->rq_offline
)
6472 class->rq_offline(rq
);
6475 cpu_clear(rq
->cpu
, rq
->rd
->online
);
6481 * migration_call - callback that gets triggered when a CPU is added.
6482 * Here we can start up the necessary migration thread for the new CPU.
6484 static int __cpuinit
6485 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
6487 struct task_struct
*p
;
6488 int cpu
= (long)hcpu
;
6489 unsigned long flags
;
6494 case CPU_UP_PREPARE
:
6495 case CPU_UP_PREPARE_FROZEN
:
6496 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
6499 kthread_bind(p
, cpu
);
6500 /* Must be high prio: stop_machine expects to yield to it. */
6501 rq
= task_rq_lock(p
, &flags
);
6502 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6503 task_rq_unlock(rq
, &flags
);
6504 cpu_rq(cpu
)->migration_thread
= p
;
6508 case CPU_ONLINE_FROZEN
:
6509 /* Strictly unnecessary, as first user will wake it. */
6510 wake_up_process(cpu_rq(cpu
)->migration_thread
);
6512 /* Update our root-domain */
6514 spin_lock_irqsave(&rq
->lock
, flags
);
6516 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6520 spin_unlock_irqrestore(&rq
->lock
, flags
);
6523 #ifdef CONFIG_HOTPLUG_CPU
6524 case CPU_UP_CANCELED
:
6525 case CPU_UP_CANCELED_FROZEN
:
6526 if (!cpu_rq(cpu
)->migration_thread
)
6528 /* Unbind it from offline cpu so it can run. Fall thru. */
6529 kthread_bind(cpu_rq(cpu
)->migration_thread
,
6530 any_online_cpu(cpu_online_map
));
6531 kthread_stop(cpu_rq(cpu
)->migration_thread
);
6532 cpu_rq(cpu
)->migration_thread
= NULL
;
6536 case CPU_DEAD_FROZEN
:
6537 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6538 migrate_live_tasks(cpu
);
6540 kthread_stop(rq
->migration_thread
);
6541 rq
->migration_thread
= NULL
;
6542 /* Idle task back to normal (off runqueue, low prio) */
6543 spin_lock_irq(&rq
->lock
);
6544 update_rq_clock(rq
);
6545 deactivate_task(rq
, rq
->idle
, 0);
6546 rq
->idle
->static_prio
= MAX_PRIO
;
6547 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
6548 rq
->idle
->sched_class
= &idle_sched_class
;
6549 migrate_dead_tasks(cpu
);
6550 spin_unlock_irq(&rq
->lock
);
6552 migrate_nr_uninterruptible(rq
);
6553 BUG_ON(rq
->nr_running
!= 0);
6556 * No need to migrate the tasks: it was best-effort if
6557 * they didn't take sched_hotcpu_mutex. Just wake up
6560 spin_lock_irq(&rq
->lock
);
6561 while (!list_empty(&rq
->migration_queue
)) {
6562 struct migration_req
*req
;
6564 req
= list_entry(rq
->migration_queue
.next
,
6565 struct migration_req
, list
);
6566 list_del_init(&req
->list
);
6567 spin_unlock_irq(&rq
->lock
);
6568 complete(&req
->done
);
6569 spin_lock_irq(&rq
->lock
);
6571 spin_unlock_irq(&rq
->lock
);
6575 case CPU_DYING_FROZEN
:
6576 /* Update our root-domain */
6578 spin_lock_irqsave(&rq
->lock
, flags
);
6580 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6583 spin_unlock_irqrestore(&rq
->lock
, flags
);
6590 /* Register at highest priority so that task migration (migrate_all_tasks)
6591 * happens before everything else.
6593 static struct notifier_block __cpuinitdata migration_notifier
= {
6594 .notifier_call
= migration_call
,
6598 static int __init
migration_init(void)
6600 void *cpu
= (void *)(long)smp_processor_id();
6603 /* Start one for the boot CPU: */
6604 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6605 BUG_ON(err
== NOTIFY_BAD
);
6606 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6607 register_cpu_notifier(&migration_notifier
);
6611 early_initcall(migration_init
);
6616 #ifdef CONFIG_SCHED_DEBUG
6618 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6619 cpumask_t
*groupmask
)
6621 struct sched_group
*group
= sd
->groups
;
6624 cpulist_scnprintf(str
, sizeof(str
), sd
->span
);
6625 cpus_clear(*groupmask
);
6627 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6629 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6630 printk("does not load-balance\n");
6632 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6637 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
6639 if (!cpu_isset(cpu
, sd
->span
)) {
6640 printk(KERN_ERR
"ERROR: domain->span does not contain "
6643 if (!cpu_isset(cpu
, group
->cpumask
)) {
6644 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6648 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6652 printk(KERN_ERR
"ERROR: group is NULL\n");
6656 if (!group
->__cpu_power
) {
6657 printk(KERN_CONT
"\n");
6658 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6663 if (!cpus_weight(group
->cpumask
)) {
6664 printk(KERN_CONT
"\n");
6665 printk(KERN_ERR
"ERROR: empty group\n");
6669 if (cpus_intersects(*groupmask
, group
->cpumask
)) {
6670 printk(KERN_CONT
"\n");
6671 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6675 cpus_or(*groupmask
, *groupmask
, group
->cpumask
);
6677 cpulist_scnprintf(str
, sizeof(str
), group
->cpumask
);
6678 printk(KERN_CONT
" %s", str
);
6680 group
= group
->next
;
6681 } while (group
!= sd
->groups
);
6682 printk(KERN_CONT
"\n");
6684 if (!cpus_equal(sd
->span
, *groupmask
))
6685 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6687 if (sd
->parent
&& !cpus_subset(*groupmask
, sd
->parent
->span
))
6688 printk(KERN_ERR
"ERROR: parent span is not a superset "
6689 "of domain->span\n");
6693 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6695 cpumask_t
*groupmask
;
6699 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6703 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6705 groupmask
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
6707 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6712 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6721 #else /* !CONFIG_SCHED_DEBUG */
6722 # define sched_domain_debug(sd, cpu) do { } while (0)
6723 #endif /* CONFIG_SCHED_DEBUG */
6725 static int sd_degenerate(struct sched_domain
*sd
)
6727 if (cpus_weight(sd
->span
) == 1)
6730 /* Following flags need at least 2 groups */
6731 if (sd
->flags
& (SD_LOAD_BALANCE
|
6732 SD_BALANCE_NEWIDLE
|
6736 SD_SHARE_PKG_RESOURCES
)) {
6737 if (sd
->groups
!= sd
->groups
->next
)
6741 /* Following flags don't use groups */
6742 if (sd
->flags
& (SD_WAKE_IDLE
|
6751 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6753 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6755 if (sd_degenerate(parent
))
6758 if (!cpus_equal(sd
->span
, parent
->span
))
6761 /* Does parent contain flags not in child? */
6762 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6763 if (cflags
& SD_WAKE_AFFINE
)
6764 pflags
&= ~SD_WAKE_BALANCE
;
6765 /* Flags needing groups don't count if only 1 group in parent */
6766 if (parent
->groups
== parent
->groups
->next
) {
6767 pflags
&= ~(SD_LOAD_BALANCE
|
6768 SD_BALANCE_NEWIDLE
|
6772 SD_SHARE_PKG_RESOURCES
);
6773 if (nr_node_ids
== 1)
6774 pflags
&= ~SD_SERIALIZE
;
6776 if (~cflags
& pflags
)
6782 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6784 unsigned long flags
;
6786 spin_lock_irqsave(&rq
->lock
, flags
);
6789 struct root_domain
*old_rd
= rq
->rd
;
6791 if (cpu_isset(rq
->cpu
, old_rd
->online
))
6794 cpu_clear(rq
->cpu
, old_rd
->span
);
6796 if (atomic_dec_and_test(&old_rd
->refcount
))
6800 atomic_inc(&rd
->refcount
);
6803 cpu_set(rq
->cpu
, rd
->span
);
6804 if (cpu_isset(rq
->cpu
, cpu_online_map
))
6807 spin_unlock_irqrestore(&rq
->lock
, flags
);
6810 static void init_rootdomain(struct root_domain
*rd
)
6812 memset(rd
, 0, sizeof(*rd
));
6814 cpus_clear(rd
->span
);
6815 cpus_clear(rd
->online
);
6817 cpupri_init(&rd
->cpupri
);
6820 static void init_defrootdomain(void)
6822 init_rootdomain(&def_root_domain
);
6823 atomic_set(&def_root_domain
.refcount
, 1);
6826 static struct root_domain
*alloc_rootdomain(void)
6828 struct root_domain
*rd
;
6830 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6834 init_rootdomain(rd
);
6840 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6841 * hold the hotplug lock.
6844 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6846 struct rq
*rq
= cpu_rq(cpu
);
6847 struct sched_domain
*tmp
;
6849 /* Remove the sched domains which do not contribute to scheduling. */
6850 for (tmp
= sd
; tmp
; ) {
6851 struct sched_domain
*parent
= tmp
->parent
;
6855 if (sd_parent_degenerate(tmp
, parent
)) {
6856 tmp
->parent
= parent
->parent
;
6858 parent
->parent
->child
= tmp
;
6863 if (sd
&& sd_degenerate(sd
)) {
6869 sched_domain_debug(sd
, cpu
);
6871 rq_attach_root(rq
, rd
);
6872 rcu_assign_pointer(rq
->sd
, sd
);
6875 /* cpus with isolated domains */
6876 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
6878 /* Setup the mask of cpus configured for isolated domains */
6879 static int __init
isolated_cpu_setup(char *str
)
6881 static int __initdata ints
[NR_CPUS
];
6884 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
6885 cpus_clear(cpu_isolated_map
);
6886 for (i
= 1; i
<= ints
[0]; i
++)
6887 if (ints
[i
] < NR_CPUS
)
6888 cpu_set(ints
[i
], cpu_isolated_map
);
6892 __setup("isolcpus=", isolated_cpu_setup
);
6895 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6896 * to a function which identifies what group(along with sched group) a CPU
6897 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6898 * (due to the fact that we keep track of groups covered with a cpumask_t).
6900 * init_sched_build_groups will build a circular linked list of the groups
6901 * covered by the given span, and will set each group's ->cpumask correctly,
6902 * and ->cpu_power to 0.
6905 init_sched_build_groups(const cpumask_t
*span
, const cpumask_t
*cpu_map
,
6906 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
6907 struct sched_group
**sg
,
6908 cpumask_t
*tmpmask
),
6909 cpumask_t
*covered
, cpumask_t
*tmpmask
)
6911 struct sched_group
*first
= NULL
, *last
= NULL
;
6914 cpus_clear(*covered
);
6916 for_each_cpu_mask_nr(i
, *span
) {
6917 struct sched_group
*sg
;
6918 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6921 if (cpu_isset(i
, *covered
))
6924 cpus_clear(sg
->cpumask
);
6925 sg
->__cpu_power
= 0;
6927 for_each_cpu_mask_nr(j
, *span
) {
6928 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6931 cpu_set(j
, *covered
);
6932 cpu_set(j
, sg
->cpumask
);
6943 #define SD_NODES_PER_DOMAIN 16
6948 * find_next_best_node - find the next node to include in a sched_domain
6949 * @node: node whose sched_domain we're building
6950 * @used_nodes: nodes already in the sched_domain
6952 * Find the next node to include in a given scheduling domain. Simply
6953 * finds the closest node not already in the @used_nodes map.
6955 * Should use nodemask_t.
6957 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
6959 int i
, n
, val
, min_val
, best_node
= 0;
6963 for (i
= 0; i
< nr_node_ids
; i
++) {
6964 /* Start at @node */
6965 n
= (node
+ i
) % nr_node_ids
;
6967 if (!nr_cpus_node(n
))
6970 /* Skip already used nodes */
6971 if (node_isset(n
, *used_nodes
))
6974 /* Simple min distance search */
6975 val
= node_distance(node
, n
);
6977 if (val
< min_val
) {
6983 node_set(best_node
, *used_nodes
);
6988 * sched_domain_node_span - get a cpumask for a node's sched_domain
6989 * @node: node whose cpumask we're constructing
6990 * @span: resulting cpumask
6992 * Given a node, construct a good cpumask for its sched_domain to span. It
6993 * should be one that prevents unnecessary balancing, but also spreads tasks
6996 static void sched_domain_node_span(int node
, cpumask_t
*span
)
6998 nodemask_t used_nodes
;
6999 node_to_cpumask_ptr(nodemask
, node
);
7003 nodes_clear(used_nodes
);
7005 cpus_or(*span
, *span
, *nodemask
);
7006 node_set(node
, used_nodes
);
7008 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
7009 int next_node
= find_next_best_node(node
, &used_nodes
);
7011 node_to_cpumask_ptr_next(nodemask
, next_node
);
7012 cpus_or(*span
, *span
, *nodemask
);
7015 #endif /* CONFIG_NUMA */
7017 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
7020 * SMT sched-domains:
7022 #ifdef CONFIG_SCHED_SMT
7023 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
7024 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
7027 cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7031 *sg
= &per_cpu(sched_group_cpus
, cpu
);
7034 #endif /* CONFIG_SCHED_SMT */
7037 * multi-core sched-domains:
7039 #ifdef CONFIG_SCHED_MC
7040 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
7041 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
7042 #endif /* CONFIG_SCHED_MC */
7044 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7046 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7051 *mask
= per_cpu(cpu_sibling_map
, cpu
);
7052 cpus_and(*mask
, *mask
, *cpu_map
);
7053 group
= first_cpu(*mask
);
7055 *sg
= &per_cpu(sched_group_core
, group
);
7058 #elif defined(CONFIG_SCHED_MC)
7060 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7064 *sg
= &per_cpu(sched_group_core
, cpu
);
7069 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
7070 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
7073 cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7077 #ifdef CONFIG_SCHED_MC
7078 *mask
= cpu_coregroup_map(cpu
);
7079 cpus_and(*mask
, *mask
, *cpu_map
);
7080 group
= first_cpu(*mask
);
7081 #elif defined(CONFIG_SCHED_SMT)
7082 *mask
= per_cpu(cpu_sibling_map
, cpu
);
7083 cpus_and(*mask
, *mask
, *cpu_map
);
7084 group
= first_cpu(*mask
);
7089 *sg
= &per_cpu(sched_group_phys
, group
);
7095 * The init_sched_build_groups can't handle what we want to do with node
7096 * groups, so roll our own. Now each node has its own list of groups which
7097 * gets dynamically allocated.
7099 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
7100 static struct sched_group
***sched_group_nodes_bycpu
;
7102 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
7103 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
7105 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
7106 struct sched_group
**sg
, cpumask_t
*nodemask
)
7110 *nodemask
= node_to_cpumask(cpu_to_node(cpu
));
7111 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7112 group
= first_cpu(*nodemask
);
7115 *sg
= &per_cpu(sched_group_allnodes
, group
);
7119 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
7121 struct sched_group
*sg
= group_head
;
7127 for_each_cpu_mask_nr(j
, sg
->cpumask
) {
7128 struct sched_domain
*sd
;
7130 sd
= &per_cpu(phys_domains
, j
);
7131 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
7133 * Only add "power" once for each
7139 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
7142 } while (sg
!= group_head
);
7144 #endif /* CONFIG_NUMA */
7147 /* Free memory allocated for various sched_group structures */
7148 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
7152 for_each_cpu_mask_nr(cpu
, *cpu_map
) {
7153 struct sched_group
**sched_group_nodes
7154 = sched_group_nodes_bycpu
[cpu
];
7156 if (!sched_group_nodes
)
7159 for (i
= 0; i
< nr_node_ids
; i
++) {
7160 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
7162 *nodemask
= node_to_cpumask(i
);
7163 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7164 if (cpus_empty(*nodemask
))
7174 if (oldsg
!= sched_group_nodes
[i
])
7177 kfree(sched_group_nodes
);
7178 sched_group_nodes_bycpu
[cpu
] = NULL
;
7181 #else /* !CONFIG_NUMA */
7182 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
7185 #endif /* CONFIG_NUMA */
7188 * Initialize sched groups cpu_power.
7190 * cpu_power indicates the capacity of sched group, which is used while
7191 * distributing the load between different sched groups in a sched domain.
7192 * Typically cpu_power for all the groups in a sched domain will be same unless
7193 * there are asymmetries in the topology. If there are asymmetries, group
7194 * having more cpu_power will pickup more load compared to the group having
7197 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7198 * the maximum number of tasks a group can handle in the presence of other idle
7199 * or lightly loaded groups in the same sched domain.
7201 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
7203 struct sched_domain
*child
;
7204 struct sched_group
*group
;
7206 WARN_ON(!sd
|| !sd
->groups
);
7208 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
7213 sd
->groups
->__cpu_power
= 0;
7216 * For perf policy, if the groups in child domain share resources
7217 * (for example cores sharing some portions of the cache hierarchy
7218 * or SMT), then set this domain groups cpu_power such that each group
7219 * can handle only one task, when there are other idle groups in the
7220 * same sched domain.
7222 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
7224 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
7225 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
7230 * add cpu_power of each child group to this groups cpu_power
7232 group
= child
->groups
;
7234 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
7235 group
= group
->next
;
7236 } while (group
!= child
->groups
);
7240 * Initializers for schedule domains
7241 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7244 #ifdef CONFIG_SCHED_DEBUG
7245 # define SD_INIT_NAME(sd, type) sd->name = #type
7247 # define SD_INIT_NAME(sd, type) do { } while (0)
7250 #define SD_INIT(sd, type) sd_init_##type(sd)
7252 #define SD_INIT_FUNC(type) \
7253 static noinline void sd_init_##type(struct sched_domain *sd) \
7255 memset(sd, 0, sizeof(*sd)); \
7256 *sd = SD_##type##_INIT; \
7257 sd->level = SD_LV_##type; \
7258 SD_INIT_NAME(sd, type); \
7263 SD_INIT_FUNC(ALLNODES
)
7266 #ifdef CONFIG_SCHED_SMT
7267 SD_INIT_FUNC(SIBLING
)
7269 #ifdef CONFIG_SCHED_MC
7274 * To minimize stack usage kmalloc room for cpumasks and share the
7275 * space as the usage in build_sched_domains() dictates. Used only
7276 * if the amount of space is significant.
7279 cpumask_t tmpmask
; /* make this one first */
7282 cpumask_t this_sibling_map
;
7283 cpumask_t this_core_map
;
7285 cpumask_t send_covered
;
7288 cpumask_t domainspan
;
7290 cpumask_t notcovered
;
7295 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7296 static inline void sched_cpumask_alloc(struct allmasks
**masks
)
7298 *masks
= kmalloc(sizeof(**masks
), GFP_KERNEL
);
7300 static inline void sched_cpumask_free(struct allmasks
*masks
)
7305 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7306 static inline void sched_cpumask_alloc(struct allmasks
**masks
)
7308 static inline void sched_cpumask_free(struct allmasks
*masks
)
7312 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7313 ((unsigned long)(a) + offsetof(struct allmasks, v))
7315 static int default_relax_domain_level
= -1;
7317 static int __init
setup_relax_domain_level(char *str
)
7321 val
= simple_strtoul(str
, NULL
, 0);
7322 if (val
< SD_LV_MAX
)
7323 default_relax_domain_level
= val
;
7327 __setup("relax_domain_level=", setup_relax_domain_level
);
7329 static void set_domain_attribute(struct sched_domain
*sd
,
7330 struct sched_domain_attr
*attr
)
7334 if (!attr
|| attr
->relax_domain_level
< 0) {
7335 if (default_relax_domain_level
< 0)
7338 request
= default_relax_domain_level
;
7340 request
= attr
->relax_domain_level
;
7341 if (request
< sd
->level
) {
7342 /* turn off idle balance on this domain */
7343 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
7345 /* turn on idle balance on this domain */
7346 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
7351 * Build sched domains for a given set of cpus and attach the sched domains
7352 * to the individual cpus
7354 static int __build_sched_domains(const cpumask_t
*cpu_map
,
7355 struct sched_domain_attr
*attr
)
7358 struct root_domain
*rd
;
7359 SCHED_CPUMASK_DECLARE(allmasks
);
7362 struct sched_group
**sched_group_nodes
= NULL
;
7363 int sd_allnodes
= 0;
7366 * Allocate the per-node list of sched groups
7368 sched_group_nodes
= kcalloc(nr_node_ids
, sizeof(struct sched_group
*),
7370 if (!sched_group_nodes
) {
7371 printk(KERN_WARNING
"Can not alloc sched group node list\n");
7376 rd
= alloc_rootdomain();
7378 printk(KERN_WARNING
"Cannot alloc root domain\n");
7380 kfree(sched_group_nodes
);
7385 /* get space for all scratch cpumask variables */
7386 sched_cpumask_alloc(&allmasks
);
7388 printk(KERN_WARNING
"Cannot alloc cpumask array\n");
7391 kfree(sched_group_nodes
);
7396 tmpmask
= (cpumask_t
*)allmasks
;
7400 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
7404 * Set up domains for cpus specified by the cpu_map.
7406 for_each_cpu_mask_nr(i
, *cpu_map
) {
7407 struct sched_domain
*sd
= NULL
, *p
;
7408 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7410 *nodemask
= node_to_cpumask(cpu_to_node(i
));
7411 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7414 if (cpus_weight(*cpu_map
) >
7415 SD_NODES_PER_DOMAIN
*cpus_weight(*nodemask
)) {
7416 sd
= &per_cpu(allnodes_domains
, i
);
7417 SD_INIT(sd
, ALLNODES
);
7418 set_domain_attribute(sd
, attr
);
7419 sd
->span
= *cpu_map
;
7420 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7426 sd
= &per_cpu(node_domains
, i
);
7428 set_domain_attribute(sd
, attr
);
7429 sched_domain_node_span(cpu_to_node(i
), &sd
->span
);
7433 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7437 sd
= &per_cpu(phys_domains
, i
);
7439 set_domain_attribute(sd
, attr
);
7440 sd
->span
= *nodemask
;
7444 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7446 #ifdef CONFIG_SCHED_MC
7448 sd
= &per_cpu(core_domains
, i
);
7450 set_domain_attribute(sd
, attr
);
7451 sd
->span
= cpu_coregroup_map(i
);
7452 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7455 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7458 #ifdef CONFIG_SCHED_SMT
7460 sd
= &per_cpu(cpu_domains
, i
);
7461 SD_INIT(sd
, SIBLING
);
7462 set_domain_attribute(sd
, attr
);
7463 sd
->span
= per_cpu(cpu_sibling_map
, i
);
7464 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7467 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7471 #ifdef CONFIG_SCHED_SMT
7472 /* Set up CPU (sibling) groups */
7473 for_each_cpu_mask_nr(i
, *cpu_map
) {
7474 SCHED_CPUMASK_VAR(this_sibling_map
, allmasks
);
7475 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7477 *this_sibling_map
= per_cpu(cpu_sibling_map
, i
);
7478 cpus_and(*this_sibling_map
, *this_sibling_map
, *cpu_map
);
7479 if (i
!= first_cpu(*this_sibling_map
))
7482 init_sched_build_groups(this_sibling_map
, cpu_map
,
7484 send_covered
, tmpmask
);
7488 #ifdef CONFIG_SCHED_MC
7489 /* Set up multi-core groups */
7490 for_each_cpu_mask_nr(i
, *cpu_map
) {
7491 SCHED_CPUMASK_VAR(this_core_map
, allmasks
);
7492 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7494 *this_core_map
= cpu_coregroup_map(i
);
7495 cpus_and(*this_core_map
, *this_core_map
, *cpu_map
);
7496 if (i
!= first_cpu(*this_core_map
))
7499 init_sched_build_groups(this_core_map
, cpu_map
,
7501 send_covered
, tmpmask
);
7505 /* Set up physical groups */
7506 for (i
= 0; i
< nr_node_ids
; i
++) {
7507 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7508 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7510 *nodemask
= node_to_cpumask(i
);
7511 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7512 if (cpus_empty(*nodemask
))
7515 init_sched_build_groups(nodemask
, cpu_map
,
7517 send_covered
, tmpmask
);
7521 /* Set up node groups */
7523 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7525 init_sched_build_groups(cpu_map
, cpu_map
,
7526 &cpu_to_allnodes_group
,
7527 send_covered
, tmpmask
);
7530 for (i
= 0; i
< nr_node_ids
; i
++) {
7531 /* Set up node groups */
7532 struct sched_group
*sg
, *prev
;
7533 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7534 SCHED_CPUMASK_VAR(domainspan
, allmasks
);
7535 SCHED_CPUMASK_VAR(covered
, allmasks
);
7538 *nodemask
= node_to_cpumask(i
);
7539 cpus_clear(*covered
);
7541 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7542 if (cpus_empty(*nodemask
)) {
7543 sched_group_nodes
[i
] = NULL
;
7547 sched_domain_node_span(i
, domainspan
);
7548 cpus_and(*domainspan
, *domainspan
, *cpu_map
);
7550 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
7552 printk(KERN_WARNING
"Can not alloc domain group for "
7556 sched_group_nodes
[i
] = sg
;
7557 for_each_cpu_mask_nr(j
, *nodemask
) {
7558 struct sched_domain
*sd
;
7560 sd
= &per_cpu(node_domains
, j
);
7563 sg
->__cpu_power
= 0;
7564 sg
->cpumask
= *nodemask
;
7566 cpus_or(*covered
, *covered
, *nodemask
);
7569 for (j
= 0; j
< nr_node_ids
; j
++) {
7570 SCHED_CPUMASK_VAR(notcovered
, allmasks
);
7571 int n
= (i
+ j
) % nr_node_ids
;
7572 node_to_cpumask_ptr(pnodemask
, n
);
7574 cpus_complement(*notcovered
, *covered
);
7575 cpus_and(*tmpmask
, *notcovered
, *cpu_map
);
7576 cpus_and(*tmpmask
, *tmpmask
, *domainspan
);
7577 if (cpus_empty(*tmpmask
))
7580 cpus_and(*tmpmask
, *tmpmask
, *pnodemask
);
7581 if (cpus_empty(*tmpmask
))
7584 sg
= kmalloc_node(sizeof(struct sched_group
),
7588 "Can not alloc domain group for node %d\n", j
);
7591 sg
->__cpu_power
= 0;
7592 sg
->cpumask
= *tmpmask
;
7593 sg
->next
= prev
->next
;
7594 cpus_or(*covered
, *covered
, *tmpmask
);
7601 /* Calculate CPU power for physical packages and nodes */
7602 #ifdef CONFIG_SCHED_SMT
7603 for_each_cpu_mask_nr(i
, *cpu_map
) {
7604 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
7606 init_sched_groups_power(i
, sd
);
7609 #ifdef CONFIG_SCHED_MC
7610 for_each_cpu_mask_nr(i
, *cpu_map
) {
7611 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
7613 init_sched_groups_power(i
, sd
);
7617 for_each_cpu_mask_nr(i
, *cpu_map
) {
7618 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
7620 init_sched_groups_power(i
, sd
);
7624 for (i
= 0; i
< nr_node_ids
; i
++)
7625 init_numa_sched_groups_power(sched_group_nodes
[i
]);
7628 struct sched_group
*sg
;
7630 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
,
7632 init_numa_sched_groups_power(sg
);
7636 /* Attach the domains */
7637 for_each_cpu_mask_nr(i
, *cpu_map
) {
7638 struct sched_domain
*sd
;
7639 #ifdef CONFIG_SCHED_SMT
7640 sd
= &per_cpu(cpu_domains
, i
);
7641 #elif defined(CONFIG_SCHED_MC)
7642 sd
= &per_cpu(core_domains
, i
);
7644 sd
= &per_cpu(phys_domains
, i
);
7646 cpu_attach_domain(sd
, rd
, i
);
7649 sched_cpumask_free(allmasks
);
7654 free_sched_groups(cpu_map
, tmpmask
);
7655 sched_cpumask_free(allmasks
);
7661 static int build_sched_domains(const cpumask_t
*cpu_map
)
7663 return __build_sched_domains(cpu_map
, NULL
);
7666 static cpumask_t
*doms_cur
; /* current sched domains */
7667 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7668 static struct sched_domain_attr
*dattr_cur
;
7669 /* attribues of custom domains in 'doms_cur' */
7672 * Special case: If a kmalloc of a doms_cur partition (array of
7673 * cpumask_t) fails, then fallback to a single sched domain,
7674 * as determined by the single cpumask_t fallback_doms.
7676 static cpumask_t fallback_doms
;
7679 * arch_update_cpu_topology lets virtualized architectures update the
7680 * cpu core maps. It is supposed to return 1 if the topology changed
7681 * or 0 if it stayed the same.
7683 int __attribute__((weak
)) arch_update_cpu_topology(void)
7689 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7690 * For now this just excludes isolated cpus, but could be used to
7691 * exclude other special cases in the future.
7693 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
7697 arch_update_cpu_topology();
7699 doms_cur
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
7701 doms_cur
= &fallback_doms
;
7702 cpus_andnot(*doms_cur
, *cpu_map
, cpu_isolated_map
);
7704 err
= build_sched_domains(doms_cur
);
7705 register_sched_domain_sysctl();
7710 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
,
7713 free_sched_groups(cpu_map
, tmpmask
);
7717 * Detach sched domains from a group of cpus specified in cpu_map
7718 * These cpus will now be attached to the NULL domain
7720 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
7725 for_each_cpu_mask_nr(i
, *cpu_map
)
7726 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7727 synchronize_sched();
7728 arch_destroy_sched_domains(cpu_map
, &tmpmask
);
7731 /* handle null as "default" */
7732 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7733 struct sched_domain_attr
*new, int idx_new
)
7735 struct sched_domain_attr tmp
;
7742 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7743 new ? (new + idx_new
) : &tmp
,
7744 sizeof(struct sched_domain_attr
));
7748 * Partition sched domains as specified by the 'ndoms_new'
7749 * cpumasks in the array doms_new[] of cpumasks. This compares
7750 * doms_new[] to the current sched domain partitioning, doms_cur[].
7751 * It destroys each deleted domain and builds each new domain.
7753 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7754 * The masks don't intersect (don't overlap.) We should setup one
7755 * sched domain for each mask. CPUs not in any of the cpumasks will
7756 * not be load balanced. If the same cpumask appears both in the
7757 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7760 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7761 * ownership of it and will kfree it when done with it. If the caller
7762 * failed the kmalloc call, then it can pass in doms_new == NULL &&
7763 * ndoms_new == 1, and partition_sched_domains() will fallback to
7764 * the single partition 'fallback_doms', it also forces the domains
7767 * If doms_new == NULL it will be replaced with cpu_online_map.
7768 * ndoms_new == 0 is a special case for destroying existing domains,
7769 * and it will not create the default domain.
7771 * Call with hotplug lock held
7773 void partition_sched_domains(int ndoms_new
, cpumask_t
*doms_new
,
7774 struct sched_domain_attr
*dattr_new
)
7779 mutex_lock(&sched_domains_mutex
);
7781 /* always unregister in case we don't destroy any domains */
7782 unregister_sched_domain_sysctl();
7784 /* Let architecture update cpu core mappings. */
7785 new_topology
= arch_update_cpu_topology();
7787 n
= doms_new
? ndoms_new
: 0;
7789 /* Destroy deleted domains */
7790 for (i
= 0; i
< ndoms_cur
; i
++) {
7791 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7792 if (cpus_equal(doms_cur
[i
], doms_new
[j
])
7793 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7796 /* no match - a current sched domain not in new doms_new[] */
7797 detach_destroy_domains(doms_cur
+ i
);
7802 if (doms_new
== NULL
) {
7804 doms_new
= &fallback_doms
;
7805 cpus_andnot(doms_new
[0], cpu_online_map
, cpu_isolated_map
);
7806 WARN_ON_ONCE(dattr_new
);
7809 /* Build new domains */
7810 for (i
= 0; i
< ndoms_new
; i
++) {
7811 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
7812 if (cpus_equal(doms_new
[i
], doms_cur
[j
])
7813 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7816 /* no match - add a new doms_new */
7817 __build_sched_domains(doms_new
+ i
,
7818 dattr_new
? dattr_new
+ i
: NULL
);
7823 /* Remember the new sched domains */
7824 if (doms_cur
!= &fallback_doms
)
7826 kfree(dattr_cur
); /* kfree(NULL) is safe */
7827 doms_cur
= doms_new
;
7828 dattr_cur
= dattr_new
;
7829 ndoms_cur
= ndoms_new
;
7831 register_sched_domain_sysctl();
7833 mutex_unlock(&sched_domains_mutex
);
7836 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7837 int arch_reinit_sched_domains(void)
7841 /* Destroy domains first to force the rebuild */
7842 partition_sched_domains(0, NULL
, NULL
);
7844 rebuild_sched_domains();
7850 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7854 if (buf
[0] != '0' && buf
[0] != '1')
7858 sched_smt_power_savings
= (buf
[0] == '1');
7860 sched_mc_power_savings
= (buf
[0] == '1');
7862 ret
= arch_reinit_sched_domains();
7864 return ret
? ret
: count
;
7867 #ifdef CONFIG_SCHED_MC
7868 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
7871 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7873 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
7874 const char *buf
, size_t count
)
7876 return sched_power_savings_store(buf
, count
, 0);
7878 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
7879 sched_mc_power_savings_show
,
7880 sched_mc_power_savings_store
);
7883 #ifdef CONFIG_SCHED_SMT
7884 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
7887 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7889 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
7890 const char *buf
, size_t count
)
7892 return sched_power_savings_store(buf
, count
, 1);
7894 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
7895 sched_smt_power_savings_show
,
7896 sched_smt_power_savings_store
);
7899 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7903 #ifdef CONFIG_SCHED_SMT
7905 err
= sysfs_create_file(&cls
->kset
.kobj
,
7906 &attr_sched_smt_power_savings
.attr
);
7908 #ifdef CONFIG_SCHED_MC
7909 if (!err
&& mc_capable())
7910 err
= sysfs_create_file(&cls
->kset
.kobj
,
7911 &attr_sched_mc_power_savings
.attr
);
7915 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7917 #ifndef CONFIG_CPUSETS
7919 * Add online and remove offline CPUs from the scheduler domains.
7920 * When cpusets are enabled they take over this function.
7922 static int update_sched_domains(struct notifier_block
*nfb
,
7923 unsigned long action
, void *hcpu
)
7927 case CPU_ONLINE_FROZEN
:
7929 case CPU_DEAD_FROZEN
:
7930 partition_sched_domains(1, NULL
, NULL
);
7939 static int update_runtime(struct notifier_block
*nfb
,
7940 unsigned long action
, void *hcpu
)
7942 int cpu
= (int)(long)hcpu
;
7945 case CPU_DOWN_PREPARE
:
7946 case CPU_DOWN_PREPARE_FROZEN
:
7947 disable_runtime(cpu_rq(cpu
));
7950 case CPU_DOWN_FAILED
:
7951 case CPU_DOWN_FAILED_FROZEN
:
7953 case CPU_ONLINE_FROZEN
:
7954 enable_runtime(cpu_rq(cpu
));
7962 void __init
sched_init_smp(void)
7964 cpumask_t non_isolated_cpus
;
7966 #if defined(CONFIG_NUMA)
7967 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
7969 BUG_ON(sched_group_nodes_bycpu
== NULL
);
7972 mutex_lock(&sched_domains_mutex
);
7973 arch_init_sched_domains(&cpu_online_map
);
7974 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
7975 if (cpus_empty(non_isolated_cpus
))
7976 cpu_set(smp_processor_id(), non_isolated_cpus
);
7977 mutex_unlock(&sched_domains_mutex
);
7980 #ifndef CONFIG_CPUSETS
7981 /* XXX: Theoretical race here - CPU may be hotplugged now */
7982 hotcpu_notifier(update_sched_domains
, 0);
7985 /* RT runtime code needs to handle some hotplug events */
7986 hotcpu_notifier(update_runtime
, 0);
7990 /* Move init over to a non-isolated CPU */
7991 if (set_cpus_allowed_ptr(current
, &non_isolated_cpus
) < 0)
7993 sched_init_granularity();
7996 void __init
sched_init_smp(void)
7998 sched_init_granularity();
8000 #endif /* CONFIG_SMP */
8002 int in_sched_functions(unsigned long addr
)
8004 return in_lock_functions(addr
) ||
8005 (addr
>= (unsigned long)__sched_text_start
8006 && addr
< (unsigned long)__sched_text_end
);
8009 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
8011 cfs_rq
->tasks_timeline
= RB_ROOT
;
8012 INIT_LIST_HEAD(&cfs_rq
->tasks
);
8013 #ifdef CONFIG_FAIR_GROUP_SCHED
8016 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
8019 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
8021 struct rt_prio_array
*array
;
8024 array
= &rt_rq
->active
;
8025 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
8026 INIT_LIST_HEAD(array
->queue
+ i
);
8027 __clear_bit(i
, array
->bitmap
);
8029 /* delimiter for bitsearch: */
8030 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
8032 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8033 rt_rq
->highest_prio
= MAX_RT_PRIO
;
8036 rt_rq
->rt_nr_migratory
= 0;
8037 rt_rq
->overloaded
= 0;
8041 rt_rq
->rt_throttled
= 0;
8042 rt_rq
->rt_runtime
= 0;
8043 spin_lock_init(&rt_rq
->rt_runtime_lock
);
8045 #ifdef CONFIG_RT_GROUP_SCHED
8046 rt_rq
->rt_nr_boosted
= 0;
8051 #ifdef CONFIG_FAIR_GROUP_SCHED
8052 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
8053 struct sched_entity
*se
, int cpu
, int add
,
8054 struct sched_entity
*parent
)
8056 struct rq
*rq
= cpu_rq(cpu
);
8057 tg
->cfs_rq
[cpu
] = cfs_rq
;
8058 init_cfs_rq(cfs_rq
, rq
);
8061 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
8064 /* se could be NULL for init_task_group */
8069 se
->cfs_rq
= &rq
->cfs
;
8071 se
->cfs_rq
= parent
->my_q
;
8074 se
->load
.weight
= tg
->shares
;
8075 se
->load
.inv_weight
= 0;
8076 se
->parent
= parent
;
8080 #ifdef CONFIG_RT_GROUP_SCHED
8081 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
8082 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
8083 struct sched_rt_entity
*parent
)
8085 struct rq
*rq
= cpu_rq(cpu
);
8087 tg
->rt_rq
[cpu
] = rt_rq
;
8088 init_rt_rq(rt_rq
, rq
);
8090 rt_rq
->rt_se
= rt_se
;
8091 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8093 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
8095 tg
->rt_se
[cpu
] = rt_se
;
8100 rt_se
->rt_rq
= &rq
->rt
;
8102 rt_se
->rt_rq
= parent
->my_q
;
8104 rt_se
->my_q
= rt_rq
;
8105 rt_se
->parent
= parent
;
8106 INIT_LIST_HEAD(&rt_se
->run_list
);
8110 void __init
sched_init(void)
8113 unsigned long alloc_size
= 0, ptr
;
8115 #ifdef CONFIG_FAIR_GROUP_SCHED
8116 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8118 #ifdef CONFIG_RT_GROUP_SCHED
8119 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8121 #ifdef CONFIG_USER_SCHED
8125 * As sched_init() is called before page_alloc is setup,
8126 * we use alloc_bootmem().
8129 ptr
= (unsigned long)alloc_bootmem(alloc_size
);
8131 #ifdef CONFIG_FAIR_GROUP_SCHED
8132 init_task_group
.se
= (struct sched_entity
**)ptr
;
8133 ptr
+= nr_cpu_ids
* sizeof(void **);
8135 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8136 ptr
+= nr_cpu_ids
* sizeof(void **);
8138 #ifdef CONFIG_USER_SCHED
8139 root_task_group
.se
= (struct sched_entity
**)ptr
;
8140 ptr
+= nr_cpu_ids
* sizeof(void **);
8142 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8143 ptr
+= nr_cpu_ids
* sizeof(void **);
8144 #endif /* CONFIG_USER_SCHED */
8145 #endif /* CONFIG_FAIR_GROUP_SCHED */
8146 #ifdef CONFIG_RT_GROUP_SCHED
8147 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8148 ptr
+= nr_cpu_ids
* sizeof(void **);
8150 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8151 ptr
+= nr_cpu_ids
* sizeof(void **);
8153 #ifdef CONFIG_USER_SCHED
8154 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8155 ptr
+= nr_cpu_ids
* sizeof(void **);
8157 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8158 ptr
+= nr_cpu_ids
* sizeof(void **);
8159 #endif /* CONFIG_USER_SCHED */
8160 #endif /* CONFIG_RT_GROUP_SCHED */
8164 init_defrootdomain();
8167 init_rt_bandwidth(&def_rt_bandwidth
,
8168 global_rt_period(), global_rt_runtime());
8170 #ifdef CONFIG_RT_GROUP_SCHED
8171 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
8172 global_rt_period(), global_rt_runtime());
8173 #ifdef CONFIG_USER_SCHED
8174 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
8175 global_rt_period(), RUNTIME_INF
);
8176 #endif /* CONFIG_USER_SCHED */
8177 #endif /* CONFIG_RT_GROUP_SCHED */
8179 #ifdef CONFIG_GROUP_SCHED
8180 list_add(&init_task_group
.list
, &task_groups
);
8181 INIT_LIST_HEAD(&init_task_group
.children
);
8183 #ifdef CONFIG_USER_SCHED
8184 INIT_LIST_HEAD(&root_task_group
.children
);
8185 init_task_group
.parent
= &root_task_group
;
8186 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
8187 #endif /* CONFIG_USER_SCHED */
8188 #endif /* CONFIG_GROUP_SCHED */
8190 for_each_possible_cpu(i
) {
8194 spin_lock_init(&rq
->lock
);
8196 init_cfs_rq(&rq
->cfs
, rq
);
8197 init_rt_rq(&rq
->rt
, rq
);
8198 #ifdef CONFIG_FAIR_GROUP_SCHED
8199 init_task_group
.shares
= init_task_group_load
;
8200 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
8201 #ifdef CONFIG_CGROUP_SCHED
8203 * How much cpu bandwidth does init_task_group get?
8205 * In case of task-groups formed thr' the cgroup filesystem, it
8206 * gets 100% of the cpu resources in the system. This overall
8207 * system cpu resource is divided among the tasks of
8208 * init_task_group and its child task-groups in a fair manner,
8209 * based on each entity's (task or task-group's) weight
8210 * (se->load.weight).
8212 * In other words, if init_task_group has 10 tasks of weight
8213 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8214 * then A0's share of the cpu resource is:
8216 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8218 * We achieve this by letting init_task_group's tasks sit
8219 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8221 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
8222 #elif defined CONFIG_USER_SCHED
8223 root_task_group
.shares
= NICE_0_LOAD
;
8224 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
8226 * In case of task-groups formed thr' the user id of tasks,
8227 * init_task_group represents tasks belonging to root user.
8228 * Hence it forms a sibling of all subsequent groups formed.
8229 * In this case, init_task_group gets only a fraction of overall
8230 * system cpu resource, based on the weight assigned to root
8231 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8232 * by letting tasks of init_task_group sit in a separate cfs_rq
8233 * (init_cfs_rq) and having one entity represent this group of
8234 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8236 init_tg_cfs_entry(&init_task_group
,
8237 &per_cpu(init_cfs_rq
, i
),
8238 &per_cpu(init_sched_entity
, i
), i
, 1,
8239 root_task_group
.se
[i
]);
8242 #endif /* CONFIG_FAIR_GROUP_SCHED */
8244 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
8245 #ifdef CONFIG_RT_GROUP_SCHED
8246 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
8247 #ifdef CONFIG_CGROUP_SCHED
8248 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
8249 #elif defined CONFIG_USER_SCHED
8250 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
8251 init_tg_rt_entry(&init_task_group
,
8252 &per_cpu(init_rt_rq
, i
),
8253 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
8254 root_task_group
.rt_se
[i
]);
8258 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
8259 rq
->cpu_load
[j
] = 0;
8263 rq
->active_balance
= 0;
8264 rq
->next_balance
= jiffies
;
8268 rq
->migration_thread
= NULL
;
8269 INIT_LIST_HEAD(&rq
->migration_queue
);
8270 rq_attach_root(rq
, &def_root_domain
);
8273 atomic_set(&rq
->nr_iowait
, 0);
8276 set_load_weight(&init_task
);
8278 #ifdef CONFIG_PREEMPT_NOTIFIERS
8279 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
8283 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
8286 #ifdef CONFIG_RT_MUTEXES
8287 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
8291 * The boot idle thread does lazy MMU switching as well:
8293 atomic_inc(&init_mm
.mm_count
);
8294 enter_lazy_tlb(&init_mm
, current
);
8297 * Make us the idle thread. Technically, schedule() should not be
8298 * called from this thread, however somewhere below it might be,
8299 * but because we are the idle thread, we just pick up running again
8300 * when this runqueue becomes "idle".
8302 init_idle(current
, smp_processor_id());
8304 * During early bootup we pretend to be a normal task:
8306 current
->sched_class
= &fair_sched_class
;
8308 scheduler_running
= 1;
8311 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8312 void __might_sleep(char *file
, int line
)
8315 static unsigned long prev_jiffy
; /* ratelimiting */
8317 if ((!in_atomic() && !irqs_disabled()) ||
8318 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
8320 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
8322 prev_jiffy
= jiffies
;
8325 "BUG: sleeping function called from invalid context at %s:%d\n",
8328 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8329 in_atomic(), irqs_disabled(),
8330 current
->pid
, current
->comm
);
8332 debug_show_held_locks(current
);
8333 if (irqs_disabled())
8334 print_irqtrace_events(current
);
8338 EXPORT_SYMBOL(__might_sleep
);
8341 #ifdef CONFIG_MAGIC_SYSRQ
8342 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
8346 update_rq_clock(rq
);
8347 on_rq
= p
->se
.on_rq
;
8349 deactivate_task(rq
, p
, 0);
8350 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
8352 activate_task(rq
, p
, 0);
8353 resched_task(rq
->curr
);
8357 void normalize_rt_tasks(void)
8359 struct task_struct
*g
, *p
;
8360 unsigned long flags
;
8363 read_lock_irqsave(&tasklist_lock
, flags
);
8364 do_each_thread(g
, p
) {
8366 * Only normalize user tasks:
8371 p
->se
.exec_start
= 0;
8372 #ifdef CONFIG_SCHEDSTATS
8373 p
->se
.wait_start
= 0;
8374 p
->se
.sleep_start
= 0;
8375 p
->se
.block_start
= 0;
8380 * Renice negative nice level userspace
8383 if (TASK_NICE(p
) < 0 && p
->mm
)
8384 set_user_nice(p
, 0);
8388 spin_lock(&p
->pi_lock
);
8389 rq
= __task_rq_lock(p
);
8391 normalize_task(rq
, p
);
8393 __task_rq_unlock(rq
);
8394 spin_unlock(&p
->pi_lock
);
8395 } while_each_thread(g
, p
);
8397 read_unlock_irqrestore(&tasklist_lock
, flags
);
8400 #endif /* CONFIG_MAGIC_SYSRQ */
8404 * These functions are only useful for the IA64 MCA handling.
8406 * They can only be called when the whole system has been
8407 * stopped - every CPU needs to be quiescent, and no scheduling
8408 * activity can take place. Using them for anything else would
8409 * be a serious bug, and as a result, they aren't even visible
8410 * under any other configuration.
8414 * curr_task - return the current task for a given cpu.
8415 * @cpu: the processor in question.
8417 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8419 struct task_struct
*curr_task(int cpu
)
8421 return cpu_curr(cpu
);
8425 * set_curr_task - set the current task for a given cpu.
8426 * @cpu: the processor in question.
8427 * @p: the task pointer to set.
8429 * Description: This function must only be used when non-maskable interrupts
8430 * are serviced on a separate stack. It allows the architecture to switch the
8431 * notion of the current task on a cpu in a non-blocking manner. This function
8432 * must be called with all CPU's synchronized, and interrupts disabled, the
8433 * and caller must save the original value of the current task (see
8434 * curr_task() above) and restore that value before reenabling interrupts and
8435 * re-starting the system.
8437 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8439 void set_curr_task(int cpu
, struct task_struct
*p
)
8446 #ifdef CONFIG_FAIR_GROUP_SCHED
8447 static void free_fair_sched_group(struct task_group
*tg
)
8451 for_each_possible_cpu(i
) {
8453 kfree(tg
->cfs_rq
[i
]);
8463 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8465 struct cfs_rq
*cfs_rq
;
8466 struct sched_entity
*se
;
8470 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8473 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8477 tg
->shares
= NICE_0_LOAD
;
8479 for_each_possible_cpu(i
) {
8482 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
8483 GFP_KERNEL
, cpu_to_node(i
));
8487 se
= kzalloc_node(sizeof(struct sched_entity
),
8488 GFP_KERNEL
, cpu_to_node(i
));
8492 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
8501 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8503 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
8504 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
8507 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8509 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
8511 #else /* !CONFG_FAIR_GROUP_SCHED */
8512 static inline void free_fair_sched_group(struct task_group
*tg
)
8517 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8522 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8526 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8529 #endif /* CONFIG_FAIR_GROUP_SCHED */
8531 #ifdef CONFIG_RT_GROUP_SCHED
8532 static void free_rt_sched_group(struct task_group
*tg
)
8536 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8538 for_each_possible_cpu(i
) {
8540 kfree(tg
->rt_rq
[i
]);
8542 kfree(tg
->rt_se
[i
]);
8550 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8552 struct rt_rq
*rt_rq
;
8553 struct sched_rt_entity
*rt_se
;
8557 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8560 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8564 init_rt_bandwidth(&tg
->rt_bandwidth
,
8565 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8567 for_each_possible_cpu(i
) {
8570 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
8571 GFP_KERNEL
, cpu_to_node(i
));
8575 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
8576 GFP_KERNEL
, cpu_to_node(i
));
8580 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
8589 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8591 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8592 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8595 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8597 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8599 #else /* !CONFIG_RT_GROUP_SCHED */
8600 static inline void free_rt_sched_group(struct task_group
*tg
)
8605 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8610 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8614 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8617 #endif /* CONFIG_RT_GROUP_SCHED */
8619 #ifdef CONFIG_GROUP_SCHED
8620 static void free_sched_group(struct task_group
*tg
)
8622 free_fair_sched_group(tg
);
8623 free_rt_sched_group(tg
);
8627 /* allocate runqueue etc for a new task group */
8628 struct task_group
*sched_create_group(struct task_group
*parent
)
8630 struct task_group
*tg
;
8631 unsigned long flags
;
8634 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8636 return ERR_PTR(-ENOMEM
);
8638 if (!alloc_fair_sched_group(tg
, parent
))
8641 if (!alloc_rt_sched_group(tg
, parent
))
8644 spin_lock_irqsave(&task_group_lock
, flags
);
8645 for_each_possible_cpu(i
) {
8646 register_fair_sched_group(tg
, i
);
8647 register_rt_sched_group(tg
, i
);
8649 list_add_rcu(&tg
->list
, &task_groups
);
8651 WARN_ON(!parent
); /* root should already exist */
8653 tg
->parent
= parent
;
8654 INIT_LIST_HEAD(&tg
->children
);
8655 list_add_rcu(&tg
->siblings
, &parent
->children
);
8656 spin_unlock_irqrestore(&task_group_lock
, flags
);
8661 free_sched_group(tg
);
8662 return ERR_PTR(-ENOMEM
);
8665 /* rcu callback to free various structures associated with a task group */
8666 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8668 /* now it should be safe to free those cfs_rqs */
8669 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8672 /* Destroy runqueue etc associated with a task group */
8673 void sched_destroy_group(struct task_group
*tg
)
8675 unsigned long flags
;
8678 spin_lock_irqsave(&task_group_lock
, flags
);
8679 for_each_possible_cpu(i
) {
8680 unregister_fair_sched_group(tg
, i
);
8681 unregister_rt_sched_group(tg
, i
);
8683 list_del_rcu(&tg
->list
);
8684 list_del_rcu(&tg
->siblings
);
8685 spin_unlock_irqrestore(&task_group_lock
, flags
);
8687 /* wait for possible concurrent references to cfs_rqs complete */
8688 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8691 /* change task's runqueue when it moves between groups.
8692 * The caller of this function should have put the task in its new group
8693 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8694 * reflect its new group.
8696 void sched_move_task(struct task_struct
*tsk
)
8699 unsigned long flags
;
8702 rq
= task_rq_lock(tsk
, &flags
);
8704 update_rq_clock(rq
);
8706 running
= task_current(rq
, tsk
);
8707 on_rq
= tsk
->se
.on_rq
;
8710 dequeue_task(rq
, tsk
, 0);
8711 if (unlikely(running
))
8712 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8714 set_task_rq(tsk
, task_cpu(tsk
));
8716 #ifdef CONFIG_FAIR_GROUP_SCHED
8717 if (tsk
->sched_class
->moved_group
)
8718 tsk
->sched_class
->moved_group(tsk
);
8721 if (unlikely(running
))
8722 tsk
->sched_class
->set_curr_task(rq
);
8724 enqueue_task(rq
, tsk
, 0);
8726 task_rq_unlock(rq
, &flags
);
8728 #endif /* CONFIG_GROUP_SCHED */
8730 #ifdef CONFIG_FAIR_GROUP_SCHED
8731 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8733 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8738 dequeue_entity(cfs_rq
, se
, 0);
8740 se
->load
.weight
= shares
;
8741 se
->load
.inv_weight
= 0;
8744 enqueue_entity(cfs_rq
, se
, 0);
8747 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8749 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8750 struct rq
*rq
= cfs_rq
->rq
;
8751 unsigned long flags
;
8753 spin_lock_irqsave(&rq
->lock
, flags
);
8754 __set_se_shares(se
, shares
);
8755 spin_unlock_irqrestore(&rq
->lock
, flags
);
8758 static DEFINE_MUTEX(shares_mutex
);
8760 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8763 unsigned long flags
;
8766 * We can't change the weight of the root cgroup.
8771 if (shares
< MIN_SHARES
)
8772 shares
= MIN_SHARES
;
8773 else if (shares
> MAX_SHARES
)
8774 shares
= MAX_SHARES
;
8776 mutex_lock(&shares_mutex
);
8777 if (tg
->shares
== shares
)
8780 spin_lock_irqsave(&task_group_lock
, flags
);
8781 for_each_possible_cpu(i
)
8782 unregister_fair_sched_group(tg
, i
);
8783 list_del_rcu(&tg
->siblings
);
8784 spin_unlock_irqrestore(&task_group_lock
, flags
);
8786 /* wait for any ongoing reference to this group to finish */
8787 synchronize_sched();
8790 * Now we are free to modify the group's share on each cpu
8791 * w/o tripping rebalance_share or load_balance_fair.
8793 tg
->shares
= shares
;
8794 for_each_possible_cpu(i
) {
8798 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
8799 set_se_shares(tg
->se
[i
], shares
);
8803 * Enable load balance activity on this group, by inserting it back on
8804 * each cpu's rq->leaf_cfs_rq_list.
8806 spin_lock_irqsave(&task_group_lock
, flags
);
8807 for_each_possible_cpu(i
)
8808 register_fair_sched_group(tg
, i
);
8809 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
8810 spin_unlock_irqrestore(&task_group_lock
, flags
);
8812 mutex_unlock(&shares_mutex
);
8816 unsigned long sched_group_shares(struct task_group
*tg
)
8822 #ifdef CONFIG_RT_GROUP_SCHED
8824 * Ensure that the real time constraints are schedulable.
8826 static DEFINE_MUTEX(rt_constraints_mutex
);
8828 static unsigned long to_ratio(u64 period
, u64 runtime
)
8830 if (runtime
== RUNTIME_INF
)
8833 return div64_u64(runtime
<< 20, period
);
8836 /* Must be called with tasklist_lock held */
8837 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8839 struct task_struct
*g
, *p
;
8841 do_each_thread(g
, p
) {
8842 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8844 } while_each_thread(g
, p
);
8849 struct rt_schedulable_data
{
8850 struct task_group
*tg
;
8855 static int tg_schedulable(struct task_group
*tg
, void *data
)
8857 struct rt_schedulable_data
*d
= data
;
8858 struct task_group
*child
;
8859 unsigned long total
, sum
= 0;
8860 u64 period
, runtime
;
8862 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8863 runtime
= tg
->rt_bandwidth
.rt_runtime
;
8866 period
= d
->rt_period
;
8867 runtime
= d
->rt_runtime
;
8871 * Cannot have more runtime than the period.
8873 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8877 * Ensure we don't starve existing RT tasks.
8879 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
8882 total
= to_ratio(period
, runtime
);
8885 * Nobody can have more than the global setting allows.
8887 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
8891 * The sum of our children's runtime should not exceed our own.
8893 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
8894 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
8895 runtime
= child
->rt_bandwidth
.rt_runtime
;
8897 if (child
== d
->tg
) {
8898 period
= d
->rt_period
;
8899 runtime
= d
->rt_runtime
;
8902 sum
+= to_ratio(period
, runtime
);
8911 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8913 struct rt_schedulable_data data
= {
8915 .rt_period
= period
,
8916 .rt_runtime
= runtime
,
8919 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
8922 static int tg_set_bandwidth(struct task_group
*tg
,
8923 u64 rt_period
, u64 rt_runtime
)
8927 mutex_lock(&rt_constraints_mutex
);
8928 read_lock(&tasklist_lock
);
8929 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
8933 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8934 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8935 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8937 for_each_possible_cpu(i
) {
8938 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8940 spin_lock(&rt_rq
->rt_runtime_lock
);
8941 rt_rq
->rt_runtime
= rt_runtime
;
8942 spin_unlock(&rt_rq
->rt_runtime_lock
);
8944 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8946 read_unlock(&tasklist_lock
);
8947 mutex_unlock(&rt_constraints_mutex
);
8952 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8954 u64 rt_runtime
, rt_period
;
8956 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8957 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8958 if (rt_runtime_us
< 0)
8959 rt_runtime
= RUNTIME_INF
;
8961 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8964 long sched_group_rt_runtime(struct task_group
*tg
)
8968 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8971 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8972 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8973 return rt_runtime_us
;
8976 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8978 u64 rt_runtime
, rt_period
;
8980 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8981 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8986 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8989 long sched_group_rt_period(struct task_group
*tg
)
8993 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8994 do_div(rt_period_us
, NSEC_PER_USEC
);
8995 return rt_period_us
;
8998 static int sched_rt_global_constraints(void)
9000 u64 runtime
, period
;
9003 if (sysctl_sched_rt_period
<= 0)
9006 runtime
= global_rt_runtime();
9007 period
= global_rt_period();
9010 * Sanity check on the sysctl variables.
9012 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
9015 mutex_lock(&rt_constraints_mutex
);
9016 read_lock(&tasklist_lock
);
9017 ret
= __rt_schedulable(NULL
, 0, 0);
9018 read_unlock(&tasklist_lock
);
9019 mutex_unlock(&rt_constraints_mutex
);
9023 #else /* !CONFIG_RT_GROUP_SCHED */
9024 static int sched_rt_global_constraints(void)
9026 unsigned long flags
;
9029 if (sysctl_sched_rt_period
<= 0)
9032 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9033 for_each_possible_cpu(i
) {
9034 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
9036 spin_lock(&rt_rq
->rt_runtime_lock
);
9037 rt_rq
->rt_runtime
= global_rt_runtime();
9038 spin_unlock(&rt_rq
->rt_runtime_lock
);
9040 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9044 #endif /* CONFIG_RT_GROUP_SCHED */
9046 int sched_rt_handler(struct ctl_table
*table
, int write
,
9047 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
9051 int old_period
, old_runtime
;
9052 static DEFINE_MUTEX(mutex
);
9055 old_period
= sysctl_sched_rt_period
;
9056 old_runtime
= sysctl_sched_rt_runtime
;
9058 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
9060 if (!ret
&& write
) {
9061 ret
= sched_rt_global_constraints();
9063 sysctl_sched_rt_period
= old_period
;
9064 sysctl_sched_rt_runtime
= old_runtime
;
9066 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
9067 def_rt_bandwidth
.rt_period
=
9068 ns_to_ktime(global_rt_period());
9071 mutex_unlock(&mutex
);
9076 #ifdef CONFIG_CGROUP_SCHED
9078 /* return corresponding task_group object of a cgroup */
9079 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
9081 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
9082 struct task_group
, css
);
9085 static struct cgroup_subsys_state
*
9086 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9088 struct task_group
*tg
, *parent
;
9090 if (!cgrp
->parent
) {
9091 /* This is early initialization for the top cgroup */
9092 return &init_task_group
.css
;
9095 parent
= cgroup_tg(cgrp
->parent
);
9096 tg
= sched_create_group(parent
);
9098 return ERR_PTR(-ENOMEM
);
9104 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9106 struct task_group
*tg
= cgroup_tg(cgrp
);
9108 sched_destroy_group(tg
);
9112 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9113 struct task_struct
*tsk
)
9115 #ifdef CONFIG_RT_GROUP_SCHED
9116 /* Don't accept realtime tasks when there is no way for them to run */
9117 if (rt_task(tsk
) && cgroup_tg(cgrp
)->rt_bandwidth
.rt_runtime
== 0)
9120 /* We don't support RT-tasks being in separate groups */
9121 if (tsk
->sched_class
!= &fair_sched_class
)
9129 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9130 struct cgroup
*old_cont
, struct task_struct
*tsk
)
9132 sched_move_task(tsk
);
9135 #ifdef CONFIG_FAIR_GROUP_SCHED
9136 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
9139 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
9142 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
9144 struct task_group
*tg
= cgroup_tg(cgrp
);
9146 return (u64
) tg
->shares
;
9148 #endif /* CONFIG_FAIR_GROUP_SCHED */
9150 #ifdef CONFIG_RT_GROUP_SCHED
9151 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
9154 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
9157 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9159 return sched_group_rt_runtime(cgroup_tg(cgrp
));
9162 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
9165 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
9168 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
9170 return sched_group_rt_period(cgroup_tg(cgrp
));
9172 #endif /* CONFIG_RT_GROUP_SCHED */
9174 static struct cftype cpu_files
[] = {
9175 #ifdef CONFIG_FAIR_GROUP_SCHED
9178 .read_u64
= cpu_shares_read_u64
,
9179 .write_u64
= cpu_shares_write_u64
,
9182 #ifdef CONFIG_RT_GROUP_SCHED
9184 .name
= "rt_runtime_us",
9185 .read_s64
= cpu_rt_runtime_read
,
9186 .write_s64
= cpu_rt_runtime_write
,
9189 .name
= "rt_period_us",
9190 .read_u64
= cpu_rt_period_read_uint
,
9191 .write_u64
= cpu_rt_period_write_uint
,
9196 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
9198 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
9201 struct cgroup_subsys cpu_cgroup_subsys
= {
9203 .create
= cpu_cgroup_create
,
9204 .destroy
= cpu_cgroup_destroy
,
9205 .can_attach
= cpu_cgroup_can_attach
,
9206 .attach
= cpu_cgroup_attach
,
9207 .populate
= cpu_cgroup_populate
,
9208 .subsys_id
= cpu_cgroup_subsys_id
,
9212 #endif /* CONFIG_CGROUP_SCHED */
9214 #ifdef CONFIG_CGROUP_CPUACCT
9217 * CPU accounting code for task groups.
9219 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9220 * (balbir@in.ibm.com).
9223 /* track cpu usage of a group of tasks and its child groups */
9225 struct cgroup_subsys_state css
;
9226 /* cpuusage holds pointer to a u64-type object on every cpu */
9228 struct cpuacct
*parent
;
9231 struct cgroup_subsys cpuacct_subsys
;
9233 /* return cpu accounting group corresponding to this container */
9234 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
9236 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
9237 struct cpuacct
, css
);
9240 /* return cpu accounting group to which this task belongs */
9241 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
9243 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
9244 struct cpuacct
, css
);
9247 /* create a new cpu accounting group */
9248 static struct cgroup_subsys_state
*cpuacct_create(
9249 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9251 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
9254 return ERR_PTR(-ENOMEM
);
9256 ca
->cpuusage
= alloc_percpu(u64
);
9257 if (!ca
->cpuusage
) {
9259 return ERR_PTR(-ENOMEM
);
9263 ca
->parent
= cgroup_ca(cgrp
->parent
);
9268 /* destroy an existing cpu accounting group */
9270 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9272 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9274 free_percpu(ca
->cpuusage
);
9278 /* return total cpu usage (in nanoseconds) of a group */
9279 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9281 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9282 u64 totalcpuusage
= 0;
9285 for_each_possible_cpu(i
) {
9286 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
9289 * Take rq->lock to make 64-bit addition safe on 32-bit
9292 spin_lock_irq(&cpu_rq(i
)->lock
);
9293 totalcpuusage
+= *cpuusage
;
9294 spin_unlock_irq(&cpu_rq(i
)->lock
);
9297 return totalcpuusage
;
9300 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
9303 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9312 for_each_possible_cpu(i
) {
9313 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
9315 spin_lock_irq(&cpu_rq(i
)->lock
);
9317 spin_unlock_irq(&cpu_rq(i
)->lock
);
9323 static struct cftype files
[] = {
9326 .read_u64
= cpuusage_read
,
9327 .write_u64
= cpuusage_write
,
9331 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9333 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
9337 * charge this task's execution time to its accounting group.
9339 * called with rq->lock held.
9341 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
9346 if (!cpuacct_subsys
.active
)
9349 cpu
= task_cpu(tsk
);
9352 for (; ca
; ca
= ca
->parent
) {
9353 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, cpu
);
9354 *cpuusage
+= cputime
;
9358 struct cgroup_subsys cpuacct_subsys
= {
9360 .create
= cpuacct_create
,
9361 .destroy
= cpuacct_destroy
,
9362 .populate
= cpuacct_populate
,
9363 .subsys_id
= cpuacct_subsys_id
,
9365 #endif /* CONFIG_CGROUP_CPUACCT */