4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/reciprocal_div.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/bootmem.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
76 #include <asm/irq_regs.h>
78 #include "sched_cpupri.h"
81 * Convert user-nice values [ -20 ... 0 ... 19 ]
82 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
85 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
86 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
87 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
90 * 'User priority' is the nice value converted to something we
91 * can work with better when scaling various scheduler parameters,
92 * it's a [ 0 ... 39 ] range.
94 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
95 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
96 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
99 * Helpers for converting nanosecond timing to jiffy resolution
101 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
103 #define NICE_0_LOAD SCHED_LOAD_SCALE
104 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
107 * These are the 'tuning knobs' of the scheduler:
109 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
110 * Timeslices get refilled after they expire.
112 #define DEF_TIMESLICE (100 * HZ / 1000)
115 * single value that denotes runtime == period, ie unlimited time.
117 #define RUNTIME_INF ((u64)~0ULL)
121 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
122 * Since cpu_power is a 'constant', we can use a reciprocal divide.
124 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
126 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
130 * Each time a sched group cpu_power is changed,
131 * we must compute its reciprocal value
133 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
135 sg
->__cpu_power
+= val
;
136 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
140 static inline int rt_policy(int policy
)
142 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
147 static inline int task_has_rt_policy(struct task_struct
*p
)
149 return rt_policy(p
->policy
);
153 * This is the priority-queue data structure of the RT scheduling class:
155 struct rt_prio_array
{
156 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
157 struct list_head queue
[MAX_RT_PRIO
];
160 struct rt_bandwidth
{
161 /* nests inside the rq lock: */
162 spinlock_t rt_runtime_lock
;
165 struct hrtimer rt_period_timer
;
168 static struct rt_bandwidth def_rt_bandwidth
;
170 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
172 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
174 struct rt_bandwidth
*rt_b
=
175 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
181 now
= hrtimer_cb_get_time(timer
);
182 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
187 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
190 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
194 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
196 rt_b
->rt_period
= ns_to_ktime(period
);
197 rt_b
->rt_runtime
= runtime
;
199 spin_lock_init(&rt_b
->rt_runtime_lock
);
201 hrtimer_init(&rt_b
->rt_period_timer
,
202 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
203 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
204 rt_b
->rt_period_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_UNLOCKED
;
207 static inline int rt_bandwidth_enabled(void)
209 return sysctl_sched_rt_runtime
>= 0;
212 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
216 if (rt_bandwidth_enabled() && rt_b
->rt_runtime
== RUNTIME_INF
)
219 if (hrtimer_active(&rt_b
->rt_period_timer
))
222 spin_lock(&rt_b
->rt_runtime_lock
);
224 if (hrtimer_active(&rt_b
->rt_period_timer
))
227 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
228 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
229 hrtimer_start_expires(&rt_b
->rt_period_timer
,
232 spin_unlock(&rt_b
->rt_runtime_lock
);
235 #ifdef CONFIG_RT_GROUP_SCHED
236 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
238 hrtimer_cancel(&rt_b
->rt_period_timer
);
243 * sched_domains_mutex serializes calls to arch_init_sched_domains,
244 * detach_destroy_domains and partition_sched_domains.
246 static DEFINE_MUTEX(sched_domains_mutex
);
248 #ifdef CONFIG_GROUP_SCHED
250 #include <linux/cgroup.h>
254 static LIST_HEAD(task_groups
);
256 /* task group related information */
258 #ifdef CONFIG_CGROUP_SCHED
259 struct cgroup_subsys_state css
;
262 #ifdef CONFIG_FAIR_GROUP_SCHED
263 /* schedulable entities of this group on each cpu */
264 struct sched_entity
**se
;
265 /* runqueue "owned" by this group on each cpu */
266 struct cfs_rq
**cfs_rq
;
267 unsigned long shares
;
270 #ifdef CONFIG_RT_GROUP_SCHED
271 struct sched_rt_entity
**rt_se
;
272 struct rt_rq
**rt_rq
;
274 struct rt_bandwidth rt_bandwidth
;
278 struct list_head list
;
280 struct task_group
*parent
;
281 struct list_head siblings
;
282 struct list_head children
;
285 #ifdef CONFIG_USER_SCHED
289 * Every UID task group (including init_task_group aka UID-0) will
290 * be a child to this group.
292 struct task_group root_task_group
;
294 #ifdef CONFIG_FAIR_GROUP_SCHED
295 /* Default task group's sched entity on each cpu */
296 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
297 /* Default task group's cfs_rq on each cpu */
298 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
299 #endif /* CONFIG_FAIR_GROUP_SCHED */
301 #ifdef CONFIG_RT_GROUP_SCHED
302 static DEFINE_PER_CPU(struct sched_rt_entity
, init_sched_rt_entity
);
303 static DEFINE_PER_CPU(struct rt_rq
, init_rt_rq
) ____cacheline_aligned_in_smp
;
304 #endif /* CONFIG_RT_GROUP_SCHED */
305 #else /* !CONFIG_USER_SCHED */
306 #define root_task_group init_task_group
307 #endif /* CONFIG_USER_SCHED */
309 /* task_group_lock serializes add/remove of task groups and also changes to
310 * a task group's cpu shares.
312 static DEFINE_SPINLOCK(task_group_lock
);
314 #ifdef CONFIG_FAIR_GROUP_SCHED
315 #ifdef CONFIG_USER_SCHED
316 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
317 #else /* !CONFIG_USER_SCHED */
318 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
319 #endif /* CONFIG_USER_SCHED */
322 * A weight of 0 or 1 can cause arithmetics problems.
323 * A weight of a cfs_rq is the sum of weights of which entities
324 * are queued on this cfs_rq, so a weight of a entity should not be
325 * too large, so as the shares value of a task group.
326 * (The default weight is 1024 - so there's no practical
327 * limitation from this.)
330 #define MAX_SHARES (1UL << 18)
332 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
335 /* Default task group.
336 * Every task in system belong to this group at bootup.
338 struct task_group init_task_group
;
340 /* return group to which a task belongs */
341 static inline struct task_group
*task_group(struct task_struct
*p
)
343 struct task_group
*tg
;
345 #ifdef CONFIG_USER_SCHED
347 #elif defined(CONFIG_CGROUP_SCHED)
348 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
349 struct task_group
, css
);
351 tg
= &init_task_group
;
356 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
357 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
359 #ifdef CONFIG_FAIR_GROUP_SCHED
360 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
361 p
->se
.parent
= task_group(p
)->se
[cpu
];
364 #ifdef CONFIG_RT_GROUP_SCHED
365 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
366 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
372 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
373 static inline struct task_group
*task_group(struct task_struct
*p
)
378 #endif /* CONFIG_GROUP_SCHED */
380 /* CFS-related fields in a runqueue */
382 struct load_weight load
;
383 unsigned long nr_running
;
389 struct rb_root tasks_timeline
;
390 struct rb_node
*rb_leftmost
;
392 struct list_head tasks
;
393 struct list_head
*balance_iterator
;
396 * 'curr' points to currently running entity on this cfs_rq.
397 * It is set to NULL otherwise (i.e when none are currently running).
399 struct sched_entity
*curr
, *next
;
401 unsigned long nr_spread_over
;
403 #ifdef CONFIG_FAIR_GROUP_SCHED
404 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
407 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
408 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
409 * (like users, containers etc.)
411 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
412 * list is used during load balance.
414 struct list_head leaf_cfs_rq_list
;
415 struct task_group
*tg
; /* group that "owns" this runqueue */
419 * the part of load.weight contributed by tasks
421 unsigned long task_weight
;
424 * h_load = weight * f(tg)
426 * Where f(tg) is the recursive weight fraction assigned to
429 unsigned long h_load
;
432 * this cpu's part of tg->shares
434 unsigned long shares
;
437 * load.weight at the time we set shares
439 unsigned long rq_weight
;
444 /* Real-Time classes' related field in a runqueue: */
446 struct rt_prio_array active
;
447 unsigned long rt_nr_running
;
448 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
449 int highest_prio
; /* highest queued rt task prio */
452 unsigned long rt_nr_migratory
;
458 /* Nests inside the rq lock: */
459 spinlock_t rt_runtime_lock
;
461 #ifdef CONFIG_RT_GROUP_SCHED
462 unsigned long rt_nr_boosted
;
465 struct list_head leaf_rt_rq_list
;
466 struct task_group
*tg
;
467 struct sched_rt_entity
*rt_se
;
474 * We add the notion of a root-domain which will be used to define per-domain
475 * variables. Each exclusive cpuset essentially defines an island domain by
476 * fully partitioning the member cpus from any other cpuset. Whenever a new
477 * exclusive cpuset is created, we also create and attach a new root-domain
487 * The "RT overload" flag: it gets set if a CPU has more than
488 * one runnable RT task.
493 struct cpupri cpupri
;
498 * By default the system creates a single root-domain with all cpus as
499 * members (mimicking the global state we have today).
501 static struct root_domain def_root_domain
;
506 * This is the main, per-CPU runqueue data structure.
508 * Locking rule: those places that want to lock multiple runqueues
509 * (such as the load balancing or the thread migration code), lock
510 * acquire operations must be ordered by ascending &runqueue.
517 * nr_running and cpu_load should be in the same cacheline because
518 * remote CPUs use both these fields when doing load calculation.
520 unsigned long nr_running
;
521 #define CPU_LOAD_IDX_MAX 5
522 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
523 unsigned char idle_at_tick
;
525 unsigned long last_tick_seen
;
526 unsigned char in_nohz_recently
;
528 /* capture load from *all* tasks on this cpu: */
529 struct load_weight load
;
530 unsigned long nr_load_updates
;
536 #ifdef CONFIG_FAIR_GROUP_SCHED
537 /* list of leaf cfs_rq on this cpu: */
538 struct list_head leaf_cfs_rq_list
;
540 #ifdef CONFIG_RT_GROUP_SCHED
541 struct list_head leaf_rt_rq_list
;
545 * This is part of a global counter where only the total sum
546 * over all CPUs matters. A task can increase this counter on
547 * one CPU and if it got migrated afterwards it may decrease
548 * it on another CPU. Always updated under the runqueue lock:
550 unsigned long nr_uninterruptible
;
552 struct task_struct
*curr
, *idle
;
553 unsigned long next_balance
;
554 struct mm_struct
*prev_mm
;
561 struct root_domain
*rd
;
562 struct sched_domain
*sd
;
564 /* For active balancing */
567 /* cpu of this runqueue: */
571 unsigned long avg_load_per_task
;
573 struct task_struct
*migration_thread
;
574 struct list_head migration_queue
;
577 #ifdef CONFIG_SCHED_HRTICK
579 int hrtick_csd_pending
;
580 struct call_single_data hrtick_csd
;
582 struct hrtimer hrtick_timer
;
585 #ifdef CONFIG_SCHEDSTATS
587 struct sched_info rq_sched_info
;
589 /* sys_sched_yield() stats */
590 unsigned int yld_exp_empty
;
591 unsigned int yld_act_empty
;
592 unsigned int yld_both_empty
;
593 unsigned int yld_count
;
595 /* schedule() stats */
596 unsigned int sched_switch
;
597 unsigned int sched_count
;
598 unsigned int sched_goidle
;
600 /* try_to_wake_up() stats */
601 unsigned int ttwu_count
;
602 unsigned int ttwu_local
;
605 unsigned int bkl_count
;
609 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
611 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int sync
)
613 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, sync
);
616 static inline int cpu_of(struct rq
*rq
)
626 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
627 * See detach_destroy_domains: synchronize_sched for details.
629 * The domain tree of any CPU may only be accessed from within
630 * preempt-disabled sections.
632 #define for_each_domain(cpu, __sd) \
633 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
635 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
636 #define this_rq() (&__get_cpu_var(runqueues))
637 #define task_rq(p) cpu_rq(task_cpu(p))
638 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
640 static inline void update_rq_clock(struct rq
*rq
)
642 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
646 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
648 #ifdef CONFIG_SCHED_DEBUG
649 # define const_debug __read_mostly
651 # define const_debug static const
657 * Returns true if the current cpu runqueue is locked.
658 * This interface allows printk to be called with the runqueue lock
659 * held and know whether or not it is OK to wake up the klogd.
661 int runqueue_is_locked(void)
664 struct rq
*rq
= cpu_rq(cpu
);
667 ret
= spin_is_locked(&rq
->lock
);
673 * Debugging: various feature bits
676 #define SCHED_FEAT(name, enabled) \
677 __SCHED_FEAT_##name ,
680 #include "sched_features.h"
685 #define SCHED_FEAT(name, enabled) \
686 (1UL << __SCHED_FEAT_##name) * enabled |
688 const_debug
unsigned int sysctl_sched_features
=
689 #include "sched_features.h"
694 #ifdef CONFIG_SCHED_DEBUG
695 #define SCHED_FEAT(name, enabled) \
698 static __read_mostly
char *sched_feat_names
[] = {
699 #include "sched_features.h"
705 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
707 filp
->private_data
= inode
->i_private
;
712 sched_feat_read(struct file
*filp
, char __user
*ubuf
,
713 size_t cnt
, loff_t
*ppos
)
720 for (i
= 0; sched_feat_names
[i
]; i
++) {
721 len
+= strlen(sched_feat_names
[i
]);
725 buf
= kmalloc(len
+ 2, GFP_KERNEL
);
729 for (i
= 0; sched_feat_names
[i
]; i
++) {
730 if (sysctl_sched_features
& (1UL << i
))
731 r
+= sprintf(buf
+ r
, "%s ", sched_feat_names
[i
]);
733 r
+= sprintf(buf
+ r
, "NO_%s ", sched_feat_names
[i
]);
736 r
+= sprintf(buf
+ r
, "\n");
737 WARN_ON(r
>= len
+ 2);
739 r
= simple_read_from_buffer(ubuf
, cnt
, ppos
, buf
, r
);
747 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
748 size_t cnt
, loff_t
*ppos
)
758 if (copy_from_user(&buf
, ubuf
, cnt
))
763 if (strncmp(buf
, "NO_", 3) == 0) {
768 for (i
= 0; sched_feat_names
[i
]; i
++) {
769 int len
= strlen(sched_feat_names
[i
]);
771 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
773 sysctl_sched_features
&= ~(1UL << i
);
775 sysctl_sched_features
|= (1UL << i
);
780 if (!sched_feat_names
[i
])
788 static struct file_operations sched_feat_fops
= {
789 .open
= sched_feat_open
,
790 .read
= sched_feat_read
,
791 .write
= sched_feat_write
,
794 static __init
int sched_init_debug(void)
796 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
801 late_initcall(sched_init_debug
);
805 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
808 * Number of tasks to iterate in a single balance run.
809 * Limited because this is done with IRQs disabled.
811 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
814 * ratelimit for updating the group shares.
817 unsigned int sysctl_sched_shares_ratelimit
= 250000;
820 * period over which we measure -rt task cpu usage in us.
823 unsigned int sysctl_sched_rt_period
= 1000000;
825 static __read_mostly
int scheduler_running
;
828 * part of the period that we allow rt tasks to run in us.
831 int sysctl_sched_rt_runtime
= 950000;
833 static inline u64
global_rt_period(void)
835 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
838 static inline u64
global_rt_runtime(void)
840 if (sysctl_sched_rt_runtime
< 0)
843 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
846 #ifndef prepare_arch_switch
847 # define prepare_arch_switch(next) do { } while (0)
849 #ifndef finish_arch_switch
850 # define finish_arch_switch(prev) do { } while (0)
853 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
855 return rq
->curr
== p
;
858 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
859 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
861 return task_current(rq
, p
);
864 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
868 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
870 #ifdef CONFIG_DEBUG_SPINLOCK
871 /* this is a valid case when another task releases the spinlock */
872 rq
->lock
.owner
= current
;
875 * If we are tracking spinlock dependencies then we have to
876 * fix up the runqueue lock - which gets 'carried over' from
879 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
881 spin_unlock_irq(&rq
->lock
);
884 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
885 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
890 return task_current(rq
, p
);
894 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
898 * We can optimise this out completely for !SMP, because the
899 * SMP rebalancing from interrupt is the only thing that cares
904 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
905 spin_unlock_irq(&rq
->lock
);
907 spin_unlock(&rq
->lock
);
911 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
915 * After ->oncpu is cleared, the task can be moved to a different CPU.
916 * We must ensure this doesn't happen until the switch is completely
922 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
926 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
929 * __task_rq_lock - lock the runqueue a given task resides on.
930 * Must be called interrupts disabled.
932 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
936 struct rq
*rq
= task_rq(p
);
937 spin_lock(&rq
->lock
);
938 if (likely(rq
== task_rq(p
)))
940 spin_unlock(&rq
->lock
);
945 * task_rq_lock - lock the runqueue a given task resides on and disable
946 * interrupts. Note the ordering: we can safely lookup the task_rq without
947 * explicitly disabling preemption.
949 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
955 local_irq_save(*flags
);
957 spin_lock(&rq
->lock
);
958 if (likely(rq
== task_rq(p
)))
960 spin_unlock_irqrestore(&rq
->lock
, *flags
);
964 static void __task_rq_unlock(struct rq
*rq
)
967 spin_unlock(&rq
->lock
);
970 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
973 spin_unlock_irqrestore(&rq
->lock
, *flags
);
977 * this_rq_lock - lock this runqueue and disable interrupts.
979 static struct rq
*this_rq_lock(void)
986 spin_lock(&rq
->lock
);
991 #ifdef CONFIG_SCHED_HRTICK
993 * Use HR-timers to deliver accurate preemption points.
995 * Its all a bit involved since we cannot program an hrt while holding the
996 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
999 * When we get rescheduled we reprogram the hrtick_timer outside of the
1005 * - enabled by features
1006 * - hrtimer is actually high res
1008 static inline int hrtick_enabled(struct rq
*rq
)
1010 if (!sched_feat(HRTICK
))
1012 if (!cpu_active(cpu_of(rq
)))
1014 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1017 static void hrtick_clear(struct rq
*rq
)
1019 if (hrtimer_active(&rq
->hrtick_timer
))
1020 hrtimer_cancel(&rq
->hrtick_timer
);
1024 * High-resolution timer tick.
1025 * Runs from hardirq context with interrupts disabled.
1027 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1029 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1031 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1033 spin_lock(&rq
->lock
);
1034 update_rq_clock(rq
);
1035 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1036 spin_unlock(&rq
->lock
);
1038 return HRTIMER_NORESTART
;
1043 * called from hardirq (IPI) context
1045 static void __hrtick_start(void *arg
)
1047 struct rq
*rq
= arg
;
1049 spin_lock(&rq
->lock
);
1050 hrtimer_restart(&rq
->hrtick_timer
);
1051 rq
->hrtick_csd_pending
= 0;
1052 spin_unlock(&rq
->lock
);
1056 * Called to set the hrtick timer state.
1058 * called with rq->lock held and irqs disabled
1060 static void hrtick_start(struct rq
*rq
, u64 delay
)
1062 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1063 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1065 hrtimer_set_expires(timer
, time
);
1067 if (rq
== this_rq()) {
1068 hrtimer_restart(timer
);
1069 } else if (!rq
->hrtick_csd_pending
) {
1070 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
);
1071 rq
->hrtick_csd_pending
= 1;
1076 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1078 int cpu
= (int)(long)hcpu
;
1081 case CPU_UP_CANCELED
:
1082 case CPU_UP_CANCELED_FROZEN
:
1083 case CPU_DOWN_PREPARE
:
1084 case CPU_DOWN_PREPARE_FROZEN
:
1086 case CPU_DEAD_FROZEN
:
1087 hrtick_clear(cpu_rq(cpu
));
1094 static __init
void init_hrtick(void)
1096 hotcpu_notifier(hotplug_hrtick
, 0);
1100 * Called to set the hrtick timer state.
1102 * called with rq->lock held and irqs disabled
1104 static void hrtick_start(struct rq
*rq
, u64 delay
)
1106 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
), HRTIMER_MODE_REL
);
1109 static inline void init_hrtick(void)
1112 #endif /* CONFIG_SMP */
1114 static void init_rq_hrtick(struct rq
*rq
)
1117 rq
->hrtick_csd_pending
= 0;
1119 rq
->hrtick_csd
.flags
= 0;
1120 rq
->hrtick_csd
.func
= __hrtick_start
;
1121 rq
->hrtick_csd
.info
= rq
;
1124 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1125 rq
->hrtick_timer
.function
= hrtick
;
1126 rq
->hrtick_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_PERCPU
;
1128 #else /* CONFIG_SCHED_HRTICK */
1129 static inline void hrtick_clear(struct rq
*rq
)
1133 static inline void init_rq_hrtick(struct rq
*rq
)
1137 static inline void init_hrtick(void)
1140 #endif /* CONFIG_SCHED_HRTICK */
1143 * resched_task - mark a task 'to be rescheduled now'.
1145 * On UP this means the setting of the need_resched flag, on SMP it
1146 * might also involve a cross-CPU call to trigger the scheduler on
1151 #ifndef tsk_is_polling
1152 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1155 static void resched_task(struct task_struct
*p
)
1159 assert_spin_locked(&task_rq(p
)->lock
);
1161 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
1164 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
1167 if (cpu
== smp_processor_id())
1170 /* NEED_RESCHED must be visible before we test polling */
1172 if (!tsk_is_polling(p
))
1173 smp_send_reschedule(cpu
);
1176 static void resched_cpu(int cpu
)
1178 struct rq
*rq
= cpu_rq(cpu
);
1179 unsigned long flags
;
1181 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1183 resched_task(cpu_curr(cpu
));
1184 spin_unlock_irqrestore(&rq
->lock
, flags
);
1189 * When add_timer_on() enqueues a timer into the timer wheel of an
1190 * idle CPU then this timer might expire before the next timer event
1191 * which is scheduled to wake up that CPU. In case of a completely
1192 * idle system the next event might even be infinite time into the
1193 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1194 * leaves the inner idle loop so the newly added timer is taken into
1195 * account when the CPU goes back to idle and evaluates the timer
1196 * wheel for the next timer event.
1198 void wake_up_idle_cpu(int cpu
)
1200 struct rq
*rq
= cpu_rq(cpu
);
1202 if (cpu
== smp_processor_id())
1206 * This is safe, as this function is called with the timer
1207 * wheel base lock of (cpu) held. When the CPU is on the way
1208 * to idle and has not yet set rq->curr to idle then it will
1209 * be serialized on the timer wheel base lock and take the new
1210 * timer into account automatically.
1212 if (rq
->curr
!= rq
->idle
)
1216 * We can set TIF_RESCHED on the idle task of the other CPU
1217 * lockless. The worst case is that the other CPU runs the
1218 * idle task through an additional NOOP schedule()
1220 set_tsk_thread_flag(rq
->idle
, TIF_NEED_RESCHED
);
1222 /* NEED_RESCHED must be visible before we test polling */
1224 if (!tsk_is_polling(rq
->idle
))
1225 smp_send_reschedule(cpu
);
1227 #endif /* CONFIG_NO_HZ */
1229 #else /* !CONFIG_SMP */
1230 static void resched_task(struct task_struct
*p
)
1232 assert_spin_locked(&task_rq(p
)->lock
);
1233 set_tsk_need_resched(p
);
1235 #endif /* CONFIG_SMP */
1237 #if BITS_PER_LONG == 32
1238 # define WMULT_CONST (~0UL)
1240 # define WMULT_CONST (1UL << 32)
1243 #define WMULT_SHIFT 32
1246 * Shift right and round:
1248 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1251 * delta *= weight / lw
1253 static unsigned long
1254 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1255 struct load_weight
*lw
)
1259 if (!lw
->inv_weight
) {
1260 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1263 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1267 tmp
= (u64
)delta_exec
* weight
;
1269 * Check whether we'd overflow the 64-bit multiplication:
1271 if (unlikely(tmp
> WMULT_CONST
))
1272 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1275 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1277 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1280 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1286 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1293 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1294 * of tasks with abnormal "nice" values across CPUs the contribution that
1295 * each task makes to its run queue's load is weighted according to its
1296 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1297 * scaled version of the new time slice allocation that they receive on time
1301 #define WEIGHT_IDLEPRIO 2
1302 #define WMULT_IDLEPRIO (1 << 31)
1305 * Nice levels are multiplicative, with a gentle 10% change for every
1306 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1307 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1308 * that remained on nice 0.
1310 * The "10% effect" is relative and cumulative: from _any_ nice level,
1311 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1312 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1313 * If a task goes up by ~10% and another task goes down by ~10% then
1314 * the relative distance between them is ~25%.)
1316 static const int prio_to_weight
[40] = {
1317 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1318 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1319 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1320 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1321 /* 0 */ 1024, 820, 655, 526, 423,
1322 /* 5 */ 335, 272, 215, 172, 137,
1323 /* 10 */ 110, 87, 70, 56, 45,
1324 /* 15 */ 36, 29, 23, 18, 15,
1328 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1330 * In cases where the weight does not change often, we can use the
1331 * precalculated inverse to speed up arithmetics by turning divisions
1332 * into multiplications:
1334 static const u32 prio_to_wmult
[40] = {
1335 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1336 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1337 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1338 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1339 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1340 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1341 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1342 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1345 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1348 * runqueue iterator, to support SMP load-balancing between different
1349 * scheduling classes, without having to expose their internal data
1350 * structures to the load-balancing proper:
1352 struct rq_iterator
{
1354 struct task_struct
*(*start
)(void *);
1355 struct task_struct
*(*next
)(void *);
1359 static unsigned long
1360 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1361 unsigned long max_load_move
, struct sched_domain
*sd
,
1362 enum cpu_idle_type idle
, int *all_pinned
,
1363 int *this_best_prio
, struct rq_iterator
*iterator
);
1366 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1367 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1368 struct rq_iterator
*iterator
);
1371 #ifdef CONFIG_CGROUP_CPUACCT
1372 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1374 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1377 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1379 update_load_add(&rq
->load
, load
);
1382 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1384 update_load_sub(&rq
->load
, load
);
1387 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1388 typedef int (*tg_visitor
)(struct task_group
*, void *);
1391 * Iterate the full tree, calling @down when first entering a node and @up when
1392 * leaving it for the final time.
1394 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1396 struct task_group
*parent
, *child
;
1400 parent
= &root_task_group
;
1402 ret
= (*down
)(parent
, data
);
1405 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1412 ret
= (*up
)(parent
, data
);
1417 parent
= parent
->parent
;
1426 static int tg_nop(struct task_group
*tg
, void *data
)
1433 static unsigned long source_load(int cpu
, int type
);
1434 static unsigned long target_load(int cpu
, int type
);
1435 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1437 static unsigned long cpu_avg_load_per_task(int cpu
)
1439 struct rq
*rq
= cpu_rq(cpu
);
1442 rq
->avg_load_per_task
= rq
->load
.weight
/ rq
->nr_running
;
1444 return rq
->avg_load_per_task
;
1447 #ifdef CONFIG_FAIR_GROUP_SCHED
1449 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1452 * Calculate and set the cpu's group shares.
1455 __update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1456 unsigned long sd_shares
, unsigned long sd_rq_weight
)
1459 unsigned long shares
;
1460 unsigned long rq_weight
;
1465 rq_weight
= tg
->cfs_rq
[cpu
]->load
.weight
;
1468 * If there are currently no tasks on the cpu pretend there is one of
1469 * average load so that when a new task gets to run here it will not
1470 * get delayed by group starvation.
1474 rq_weight
= NICE_0_LOAD
;
1477 if (unlikely(rq_weight
> sd_rq_weight
))
1478 rq_weight
= sd_rq_weight
;
1481 * \Sum shares * rq_weight
1482 * shares = -----------------------
1486 shares
= (sd_shares
* rq_weight
) / (sd_rq_weight
+ 1);
1489 * record the actual number of shares, not the boosted amount.
1491 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1492 tg
->cfs_rq
[cpu
]->rq_weight
= rq_weight
;
1494 if (shares
< MIN_SHARES
)
1495 shares
= MIN_SHARES
;
1496 else if (shares
> MAX_SHARES
)
1497 shares
= MAX_SHARES
;
1499 __set_se_shares(tg
->se
[cpu
], shares
);
1503 * Re-compute the task group their per cpu shares over the given domain.
1504 * This needs to be done in a bottom-up fashion because the rq weight of a
1505 * parent group depends on the shares of its child groups.
1507 static int tg_shares_up(struct task_group
*tg
, void *data
)
1509 unsigned long rq_weight
= 0;
1510 unsigned long shares
= 0;
1511 struct sched_domain
*sd
= data
;
1514 for_each_cpu_mask(i
, sd
->span
) {
1515 rq_weight
+= tg
->cfs_rq
[i
]->load
.weight
;
1516 shares
+= tg
->cfs_rq
[i
]->shares
;
1519 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1520 shares
= tg
->shares
;
1522 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1523 shares
= tg
->shares
;
1526 rq_weight
= cpus_weight(sd
->span
) * NICE_0_LOAD
;
1528 for_each_cpu_mask(i
, sd
->span
) {
1529 struct rq
*rq
= cpu_rq(i
);
1530 unsigned long flags
;
1532 spin_lock_irqsave(&rq
->lock
, flags
);
1533 __update_group_shares_cpu(tg
, i
, shares
, rq_weight
);
1534 spin_unlock_irqrestore(&rq
->lock
, flags
);
1541 * Compute the cpu's hierarchical load factor for each task group.
1542 * This needs to be done in a top-down fashion because the load of a child
1543 * group is a fraction of its parents load.
1545 static int tg_load_down(struct task_group
*tg
, void *data
)
1548 long cpu
= (long)data
;
1551 load
= cpu_rq(cpu
)->load
.weight
;
1553 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1554 load
*= tg
->cfs_rq
[cpu
]->shares
;
1555 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1558 tg
->cfs_rq
[cpu
]->h_load
= load
;
1563 static void update_shares(struct sched_domain
*sd
)
1565 u64 now
= cpu_clock(raw_smp_processor_id());
1566 s64 elapsed
= now
- sd
->last_update
;
1568 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1569 sd
->last_update
= now
;
1570 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1574 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1576 spin_unlock(&rq
->lock
);
1578 spin_lock(&rq
->lock
);
1581 static void update_h_load(long cpu
)
1583 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1588 static inline void update_shares(struct sched_domain
*sd
)
1592 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1600 #ifdef CONFIG_FAIR_GROUP_SCHED
1601 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1604 cfs_rq
->shares
= shares
;
1609 #include "sched_stats.h"
1610 #include "sched_idletask.c"
1611 #include "sched_fair.c"
1612 #include "sched_rt.c"
1613 #ifdef CONFIG_SCHED_DEBUG
1614 # include "sched_debug.c"
1617 #define sched_class_highest (&rt_sched_class)
1618 #define for_each_class(class) \
1619 for (class = sched_class_highest; class; class = class->next)
1621 static void inc_nr_running(struct rq
*rq
)
1626 static void dec_nr_running(struct rq
*rq
)
1631 static void set_load_weight(struct task_struct
*p
)
1633 if (task_has_rt_policy(p
)) {
1634 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1635 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1640 * SCHED_IDLE tasks get minimal weight:
1642 if (p
->policy
== SCHED_IDLE
) {
1643 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1644 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1648 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1649 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1652 static void update_avg(u64
*avg
, u64 sample
)
1654 s64 diff
= sample
- *avg
;
1658 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1660 sched_info_queued(p
);
1661 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1665 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1667 if (sleep
&& p
->se
.last_wakeup
) {
1668 update_avg(&p
->se
.avg_overlap
,
1669 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1670 p
->se
.last_wakeup
= 0;
1673 sched_info_dequeued(p
);
1674 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1679 * __normal_prio - return the priority that is based on the static prio
1681 static inline int __normal_prio(struct task_struct
*p
)
1683 return p
->static_prio
;
1687 * Calculate the expected normal priority: i.e. priority
1688 * without taking RT-inheritance into account. Might be
1689 * boosted by interactivity modifiers. Changes upon fork,
1690 * setprio syscalls, and whenever the interactivity
1691 * estimator recalculates.
1693 static inline int normal_prio(struct task_struct
*p
)
1697 if (task_has_rt_policy(p
))
1698 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1700 prio
= __normal_prio(p
);
1705 * Calculate the current priority, i.e. the priority
1706 * taken into account by the scheduler. This value might
1707 * be boosted by RT tasks, or might be boosted by
1708 * interactivity modifiers. Will be RT if the task got
1709 * RT-boosted. If not then it returns p->normal_prio.
1711 static int effective_prio(struct task_struct
*p
)
1713 p
->normal_prio
= normal_prio(p
);
1715 * If we are RT tasks or we were boosted to RT priority,
1716 * keep the priority unchanged. Otherwise, update priority
1717 * to the normal priority:
1719 if (!rt_prio(p
->prio
))
1720 return p
->normal_prio
;
1725 * activate_task - move a task to the runqueue.
1727 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1729 if (task_contributes_to_load(p
))
1730 rq
->nr_uninterruptible
--;
1732 enqueue_task(rq
, p
, wakeup
);
1737 * deactivate_task - remove a task from the runqueue.
1739 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1741 if (task_contributes_to_load(p
))
1742 rq
->nr_uninterruptible
++;
1744 dequeue_task(rq
, p
, sleep
);
1749 * task_curr - is this task currently executing on a CPU?
1750 * @p: the task in question.
1752 inline int task_curr(const struct task_struct
*p
)
1754 return cpu_curr(task_cpu(p
)) == p
;
1757 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1759 set_task_rq(p
, cpu
);
1762 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1763 * successfuly executed on another CPU. We must ensure that updates of
1764 * per-task data have been completed by this moment.
1767 task_thread_info(p
)->cpu
= cpu
;
1771 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1772 const struct sched_class
*prev_class
,
1773 int oldprio
, int running
)
1775 if (prev_class
!= p
->sched_class
) {
1776 if (prev_class
->switched_from
)
1777 prev_class
->switched_from(rq
, p
, running
);
1778 p
->sched_class
->switched_to(rq
, p
, running
);
1780 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1785 /* Used instead of source_load when we know the type == 0 */
1786 static unsigned long weighted_cpuload(const int cpu
)
1788 return cpu_rq(cpu
)->load
.weight
;
1792 * Is this task likely cache-hot:
1795 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1800 * Buddy candidates are cache hot:
1802 if (sched_feat(CACHE_HOT_BUDDY
) && (&p
->se
== cfs_rq_of(&p
->se
)->next
))
1805 if (p
->sched_class
!= &fair_sched_class
)
1808 if (sysctl_sched_migration_cost
== -1)
1810 if (sysctl_sched_migration_cost
== 0)
1813 delta
= now
- p
->se
.exec_start
;
1815 return delta
< (s64
)sysctl_sched_migration_cost
;
1819 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1821 int old_cpu
= task_cpu(p
);
1822 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1823 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1824 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1827 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1829 #ifdef CONFIG_SCHEDSTATS
1830 if (p
->se
.wait_start
)
1831 p
->se
.wait_start
-= clock_offset
;
1832 if (p
->se
.sleep_start
)
1833 p
->se
.sleep_start
-= clock_offset
;
1834 if (p
->se
.block_start
)
1835 p
->se
.block_start
-= clock_offset
;
1836 if (old_cpu
!= new_cpu
) {
1837 schedstat_inc(p
, se
.nr_migrations
);
1838 if (task_hot(p
, old_rq
->clock
, NULL
))
1839 schedstat_inc(p
, se
.nr_forced2_migrations
);
1842 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1843 new_cfsrq
->min_vruntime
;
1845 __set_task_cpu(p
, new_cpu
);
1848 struct migration_req
{
1849 struct list_head list
;
1851 struct task_struct
*task
;
1854 struct completion done
;
1858 * The task's runqueue lock must be held.
1859 * Returns true if you have to wait for migration thread.
1862 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1864 struct rq
*rq
= task_rq(p
);
1867 * If the task is not on a runqueue (and not running), then
1868 * it is sufficient to simply update the task's cpu field.
1870 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1871 set_task_cpu(p
, dest_cpu
);
1875 init_completion(&req
->done
);
1877 req
->dest_cpu
= dest_cpu
;
1878 list_add(&req
->list
, &rq
->migration_queue
);
1884 * wait_task_inactive - wait for a thread to unschedule.
1886 * If @match_state is nonzero, it's the @p->state value just checked and
1887 * not expected to change. If it changes, i.e. @p might have woken up,
1888 * then return zero. When we succeed in waiting for @p to be off its CPU,
1889 * we return a positive number (its total switch count). If a second call
1890 * a short while later returns the same number, the caller can be sure that
1891 * @p has remained unscheduled the whole time.
1893 * The caller must ensure that the task *will* unschedule sometime soon,
1894 * else this function might spin for a *long* time. This function can't
1895 * be called with interrupts off, or it may introduce deadlock with
1896 * smp_call_function() if an IPI is sent by the same process we are
1897 * waiting to become inactive.
1899 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1901 unsigned long flags
;
1908 * We do the initial early heuristics without holding
1909 * any task-queue locks at all. We'll only try to get
1910 * the runqueue lock when things look like they will
1916 * If the task is actively running on another CPU
1917 * still, just relax and busy-wait without holding
1920 * NOTE! Since we don't hold any locks, it's not
1921 * even sure that "rq" stays as the right runqueue!
1922 * But we don't care, since "task_running()" will
1923 * return false if the runqueue has changed and p
1924 * is actually now running somewhere else!
1926 while (task_running(rq
, p
)) {
1927 if (match_state
&& unlikely(p
->state
!= match_state
))
1933 * Ok, time to look more closely! We need the rq
1934 * lock now, to be *sure*. If we're wrong, we'll
1935 * just go back and repeat.
1937 rq
= task_rq_lock(p
, &flags
);
1938 running
= task_running(rq
, p
);
1939 on_rq
= p
->se
.on_rq
;
1941 if (!match_state
|| p
->state
== match_state
)
1942 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1943 task_rq_unlock(rq
, &flags
);
1946 * If it changed from the expected state, bail out now.
1948 if (unlikely(!ncsw
))
1952 * Was it really running after all now that we
1953 * checked with the proper locks actually held?
1955 * Oops. Go back and try again..
1957 if (unlikely(running
)) {
1963 * It's not enough that it's not actively running,
1964 * it must be off the runqueue _entirely_, and not
1967 * So if it wa still runnable (but just not actively
1968 * running right now), it's preempted, and we should
1969 * yield - it could be a while.
1971 if (unlikely(on_rq
)) {
1972 schedule_timeout_uninterruptible(1);
1977 * Ahh, all good. It wasn't running, and it wasn't
1978 * runnable, which means that it will never become
1979 * running in the future either. We're all done!
1988 * kick_process - kick a running thread to enter/exit the kernel
1989 * @p: the to-be-kicked thread
1991 * Cause a process which is running on another CPU to enter
1992 * kernel-mode, without any delay. (to get signals handled.)
1994 * NOTE: this function doesnt have to take the runqueue lock,
1995 * because all it wants to ensure is that the remote task enters
1996 * the kernel. If the IPI races and the task has been migrated
1997 * to another CPU then no harm is done and the purpose has been
2000 void kick_process(struct task_struct
*p
)
2006 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2007 smp_send_reschedule(cpu
);
2012 * Return a low guess at the load of a migration-source cpu weighted
2013 * according to the scheduling class and "nice" value.
2015 * We want to under-estimate the load of migration sources, to
2016 * balance conservatively.
2018 static unsigned long source_load(int cpu
, int type
)
2020 struct rq
*rq
= cpu_rq(cpu
);
2021 unsigned long total
= weighted_cpuload(cpu
);
2023 if (type
== 0 || !sched_feat(LB_BIAS
))
2026 return min(rq
->cpu_load
[type
-1], total
);
2030 * Return a high guess at the load of a migration-target cpu weighted
2031 * according to the scheduling class and "nice" value.
2033 static unsigned long target_load(int cpu
, int type
)
2035 struct rq
*rq
= cpu_rq(cpu
);
2036 unsigned long total
= weighted_cpuload(cpu
);
2038 if (type
== 0 || !sched_feat(LB_BIAS
))
2041 return max(rq
->cpu_load
[type
-1], total
);
2045 * find_idlest_group finds and returns the least busy CPU group within the
2048 static struct sched_group
*
2049 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
2051 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2052 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
2053 int load_idx
= sd
->forkexec_idx
;
2054 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
2057 unsigned long load
, avg_load
;
2061 /* Skip over this group if it has no CPUs allowed */
2062 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
2065 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2067 /* Tally up the load of all CPUs in the group */
2070 for_each_cpu_mask_nr(i
, group
->cpumask
) {
2071 /* Bias balancing toward cpus of our domain */
2073 load
= source_load(i
, load_idx
);
2075 load
= target_load(i
, load_idx
);
2080 /* Adjust by relative CPU power of the group */
2081 avg_load
= sg_div_cpu_power(group
,
2082 avg_load
* SCHED_LOAD_SCALE
);
2085 this_load
= avg_load
;
2087 } else if (avg_load
< min_load
) {
2088 min_load
= avg_load
;
2091 } while (group
= group
->next
, group
!= sd
->groups
);
2093 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
2099 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2102 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
,
2105 unsigned long load
, min_load
= ULONG_MAX
;
2109 /* Traverse only the allowed CPUs */
2110 cpus_and(*tmp
, group
->cpumask
, p
->cpus_allowed
);
2112 for_each_cpu_mask_nr(i
, *tmp
) {
2113 load
= weighted_cpuload(i
);
2115 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2125 * sched_balance_self: balance the current task (running on cpu) in domains
2126 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2129 * Balance, ie. select the least loaded group.
2131 * Returns the target CPU number, or the same CPU if no balancing is needed.
2133 * preempt must be disabled.
2135 static int sched_balance_self(int cpu
, int flag
)
2137 struct task_struct
*t
= current
;
2138 struct sched_domain
*tmp
, *sd
= NULL
;
2140 for_each_domain(cpu
, tmp
) {
2142 * If power savings logic is enabled for a domain, stop there.
2144 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2146 if (tmp
->flags
& flag
)
2154 cpumask_t span
, tmpmask
;
2155 struct sched_group
*group
;
2156 int new_cpu
, weight
;
2158 if (!(sd
->flags
& flag
)) {
2164 group
= find_idlest_group(sd
, t
, cpu
);
2170 new_cpu
= find_idlest_cpu(group
, t
, cpu
, &tmpmask
);
2171 if (new_cpu
== -1 || new_cpu
== cpu
) {
2172 /* Now try balancing at a lower domain level of cpu */
2177 /* Now try balancing at a lower domain level of new_cpu */
2180 weight
= cpus_weight(span
);
2181 for_each_domain(cpu
, tmp
) {
2182 if (weight
<= cpus_weight(tmp
->span
))
2184 if (tmp
->flags
& flag
)
2187 /* while loop will break here if sd == NULL */
2193 #endif /* CONFIG_SMP */
2196 * try_to_wake_up - wake up a thread
2197 * @p: the to-be-woken-up thread
2198 * @state: the mask of task states that can be woken
2199 * @sync: do a synchronous wakeup?
2201 * Put it on the run-queue if it's not already there. The "current"
2202 * thread is always on the run-queue (except when the actual
2203 * re-schedule is in progress), and as such you're allowed to do
2204 * the simpler "current->state = TASK_RUNNING" to mark yourself
2205 * runnable without the overhead of this.
2207 * returns failure only if the task is already active.
2209 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
2211 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2212 unsigned long flags
;
2216 if (!sched_feat(SYNC_WAKEUPS
))
2220 if (sched_feat(LB_WAKEUP_UPDATE
)) {
2221 struct sched_domain
*sd
;
2223 this_cpu
= raw_smp_processor_id();
2226 for_each_domain(this_cpu
, sd
) {
2227 if (cpu_isset(cpu
, sd
->span
)) {
2236 rq
= task_rq_lock(p
, &flags
);
2237 old_state
= p
->state
;
2238 if (!(old_state
& state
))
2246 this_cpu
= smp_processor_id();
2249 if (unlikely(task_running(rq
, p
)))
2252 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2253 if (cpu
!= orig_cpu
) {
2254 set_task_cpu(p
, cpu
);
2255 task_rq_unlock(rq
, &flags
);
2256 /* might preempt at this point */
2257 rq
= task_rq_lock(p
, &flags
);
2258 old_state
= p
->state
;
2259 if (!(old_state
& state
))
2264 this_cpu
= smp_processor_id();
2268 #ifdef CONFIG_SCHEDSTATS
2269 schedstat_inc(rq
, ttwu_count
);
2270 if (cpu
== this_cpu
)
2271 schedstat_inc(rq
, ttwu_local
);
2273 struct sched_domain
*sd
;
2274 for_each_domain(this_cpu
, sd
) {
2275 if (cpu_isset(cpu
, sd
->span
)) {
2276 schedstat_inc(sd
, ttwu_wake_remote
);
2281 #endif /* CONFIG_SCHEDSTATS */
2284 #endif /* CONFIG_SMP */
2285 schedstat_inc(p
, se
.nr_wakeups
);
2287 schedstat_inc(p
, se
.nr_wakeups_sync
);
2288 if (orig_cpu
!= cpu
)
2289 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2290 if (cpu
== this_cpu
)
2291 schedstat_inc(p
, se
.nr_wakeups_local
);
2293 schedstat_inc(p
, se
.nr_wakeups_remote
);
2294 update_rq_clock(rq
);
2295 activate_task(rq
, p
, 1);
2299 trace_mark(kernel_sched_wakeup
,
2300 "pid %d state %ld ## rq %p task %p rq->curr %p",
2301 p
->pid
, p
->state
, rq
, p
, rq
->curr
);
2302 check_preempt_curr(rq
, p
, sync
);
2304 p
->state
= TASK_RUNNING
;
2306 if (p
->sched_class
->task_wake_up
)
2307 p
->sched_class
->task_wake_up(rq
, p
);
2310 current
->se
.last_wakeup
= current
->se
.sum_exec_runtime
;
2312 task_rq_unlock(rq
, &flags
);
2317 int wake_up_process(struct task_struct
*p
)
2319 return try_to_wake_up(p
, TASK_ALL
, 0);
2321 EXPORT_SYMBOL(wake_up_process
);
2323 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2325 return try_to_wake_up(p
, state
, 0);
2329 * Perform scheduler related setup for a newly forked process p.
2330 * p is forked by current.
2332 * __sched_fork() is basic setup used by init_idle() too:
2334 static void __sched_fork(struct task_struct
*p
)
2336 p
->se
.exec_start
= 0;
2337 p
->se
.sum_exec_runtime
= 0;
2338 p
->se
.prev_sum_exec_runtime
= 0;
2339 p
->se
.last_wakeup
= 0;
2340 p
->se
.avg_overlap
= 0;
2342 #ifdef CONFIG_SCHEDSTATS
2343 p
->se
.wait_start
= 0;
2344 p
->se
.sum_sleep_runtime
= 0;
2345 p
->se
.sleep_start
= 0;
2346 p
->se
.block_start
= 0;
2347 p
->se
.sleep_max
= 0;
2348 p
->se
.block_max
= 0;
2350 p
->se
.slice_max
= 0;
2354 INIT_LIST_HEAD(&p
->rt
.run_list
);
2356 INIT_LIST_HEAD(&p
->se
.group_node
);
2358 #ifdef CONFIG_PREEMPT_NOTIFIERS
2359 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2363 * We mark the process as running here, but have not actually
2364 * inserted it onto the runqueue yet. This guarantees that
2365 * nobody will actually run it, and a signal or other external
2366 * event cannot wake it up and insert it on the runqueue either.
2368 p
->state
= TASK_RUNNING
;
2372 * fork()/clone()-time setup:
2374 void sched_fork(struct task_struct
*p
, int clone_flags
)
2376 int cpu
= get_cpu();
2381 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2383 set_task_cpu(p
, cpu
);
2386 * Make sure we do not leak PI boosting priority to the child:
2388 p
->prio
= current
->normal_prio
;
2389 if (!rt_prio(p
->prio
))
2390 p
->sched_class
= &fair_sched_class
;
2392 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2393 if (likely(sched_info_on()))
2394 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2396 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2399 #ifdef CONFIG_PREEMPT
2400 /* Want to start with kernel preemption disabled. */
2401 task_thread_info(p
)->preempt_count
= 1;
2407 * wake_up_new_task - wake up a newly created task for the first time.
2409 * This function will do some initial scheduler statistics housekeeping
2410 * that must be done for every newly created context, then puts the task
2411 * on the runqueue and wakes it.
2413 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2415 unsigned long flags
;
2418 rq
= task_rq_lock(p
, &flags
);
2419 BUG_ON(p
->state
!= TASK_RUNNING
);
2420 update_rq_clock(rq
);
2422 p
->prio
= effective_prio(p
);
2424 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2425 activate_task(rq
, p
, 0);
2428 * Let the scheduling class do new task startup
2429 * management (if any):
2431 p
->sched_class
->task_new(rq
, p
);
2434 trace_mark(kernel_sched_wakeup_new
,
2435 "pid %d state %ld ## rq %p task %p rq->curr %p",
2436 p
->pid
, p
->state
, rq
, p
, rq
->curr
);
2437 check_preempt_curr(rq
, p
, 0);
2439 if (p
->sched_class
->task_wake_up
)
2440 p
->sched_class
->task_wake_up(rq
, p
);
2442 task_rq_unlock(rq
, &flags
);
2445 #ifdef CONFIG_PREEMPT_NOTIFIERS
2448 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2449 * @notifier: notifier struct to register
2451 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2453 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2455 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2458 * preempt_notifier_unregister - no longer interested in preemption notifications
2459 * @notifier: notifier struct to unregister
2461 * This is safe to call from within a preemption notifier.
2463 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2465 hlist_del(¬ifier
->link
);
2467 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2469 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2471 struct preempt_notifier
*notifier
;
2472 struct hlist_node
*node
;
2474 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2475 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2479 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2480 struct task_struct
*next
)
2482 struct preempt_notifier
*notifier
;
2483 struct hlist_node
*node
;
2485 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2486 notifier
->ops
->sched_out(notifier
, next
);
2489 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2491 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2496 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2497 struct task_struct
*next
)
2501 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2504 * prepare_task_switch - prepare to switch tasks
2505 * @rq: the runqueue preparing to switch
2506 * @prev: the current task that is being switched out
2507 * @next: the task we are going to switch to.
2509 * This is called with the rq lock held and interrupts off. It must
2510 * be paired with a subsequent finish_task_switch after the context
2513 * prepare_task_switch sets up locking and calls architecture specific
2517 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2518 struct task_struct
*next
)
2520 fire_sched_out_preempt_notifiers(prev
, next
);
2521 prepare_lock_switch(rq
, next
);
2522 prepare_arch_switch(next
);
2526 * finish_task_switch - clean up after a task-switch
2527 * @rq: runqueue associated with task-switch
2528 * @prev: the thread we just switched away from.
2530 * finish_task_switch must be called after the context switch, paired
2531 * with a prepare_task_switch call before the context switch.
2532 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2533 * and do any other architecture-specific cleanup actions.
2535 * Note that we may have delayed dropping an mm in context_switch(). If
2536 * so, we finish that here outside of the runqueue lock. (Doing it
2537 * with the lock held can cause deadlocks; see schedule() for
2540 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2541 __releases(rq
->lock
)
2543 struct mm_struct
*mm
= rq
->prev_mm
;
2549 * A task struct has one reference for the use as "current".
2550 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2551 * schedule one last time. The schedule call will never return, and
2552 * the scheduled task must drop that reference.
2553 * The test for TASK_DEAD must occur while the runqueue locks are
2554 * still held, otherwise prev could be scheduled on another cpu, die
2555 * there before we look at prev->state, and then the reference would
2557 * Manfred Spraul <manfred@colorfullife.com>
2559 prev_state
= prev
->state
;
2560 finish_arch_switch(prev
);
2561 finish_lock_switch(rq
, prev
);
2563 if (current
->sched_class
->post_schedule
)
2564 current
->sched_class
->post_schedule(rq
);
2567 fire_sched_in_preempt_notifiers(current
);
2570 if (unlikely(prev_state
== TASK_DEAD
)) {
2572 * Remove function-return probe instances associated with this
2573 * task and put them back on the free list.
2575 kprobe_flush_task(prev
);
2576 put_task_struct(prev
);
2581 * schedule_tail - first thing a freshly forked thread must call.
2582 * @prev: the thread we just switched away from.
2584 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2585 __releases(rq
->lock
)
2587 struct rq
*rq
= this_rq();
2589 finish_task_switch(rq
, prev
);
2590 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2591 /* In this case, finish_task_switch does not reenable preemption */
2594 if (current
->set_child_tid
)
2595 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2599 * context_switch - switch to the new MM and the new
2600 * thread's register state.
2603 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2604 struct task_struct
*next
)
2606 struct mm_struct
*mm
, *oldmm
;
2608 prepare_task_switch(rq
, prev
, next
);
2609 trace_mark(kernel_sched_schedule
,
2610 "prev_pid %d next_pid %d prev_state %ld "
2611 "## rq %p prev %p next %p",
2612 prev
->pid
, next
->pid
, prev
->state
,
2615 oldmm
= prev
->active_mm
;
2617 * For paravirt, this is coupled with an exit in switch_to to
2618 * combine the page table reload and the switch backend into
2621 arch_enter_lazy_cpu_mode();
2623 if (unlikely(!mm
)) {
2624 next
->active_mm
= oldmm
;
2625 atomic_inc(&oldmm
->mm_count
);
2626 enter_lazy_tlb(oldmm
, next
);
2628 switch_mm(oldmm
, mm
, next
);
2630 if (unlikely(!prev
->mm
)) {
2631 prev
->active_mm
= NULL
;
2632 rq
->prev_mm
= oldmm
;
2635 * Since the runqueue lock will be released by the next
2636 * task (which is an invalid locking op but in the case
2637 * of the scheduler it's an obvious special-case), so we
2638 * do an early lockdep release here:
2640 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2641 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2644 /* Here we just switch the register state and the stack. */
2645 switch_to(prev
, next
, prev
);
2649 * this_rq must be evaluated again because prev may have moved
2650 * CPUs since it called schedule(), thus the 'rq' on its stack
2651 * frame will be invalid.
2653 finish_task_switch(this_rq(), prev
);
2657 * nr_running, nr_uninterruptible and nr_context_switches:
2659 * externally visible scheduler statistics: current number of runnable
2660 * threads, current number of uninterruptible-sleeping threads, total
2661 * number of context switches performed since bootup.
2663 unsigned long nr_running(void)
2665 unsigned long i
, sum
= 0;
2667 for_each_online_cpu(i
)
2668 sum
+= cpu_rq(i
)->nr_running
;
2673 unsigned long nr_uninterruptible(void)
2675 unsigned long i
, sum
= 0;
2677 for_each_possible_cpu(i
)
2678 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2681 * Since we read the counters lockless, it might be slightly
2682 * inaccurate. Do not allow it to go below zero though:
2684 if (unlikely((long)sum
< 0))
2690 unsigned long long nr_context_switches(void)
2693 unsigned long long sum
= 0;
2695 for_each_possible_cpu(i
)
2696 sum
+= cpu_rq(i
)->nr_switches
;
2701 unsigned long nr_iowait(void)
2703 unsigned long i
, sum
= 0;
2705 for_each_possible_cpu(i
)
2706 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2711 unsigned long nr_active(void)
2713 unsigned long i
, running
= 0, uninterruptible
= 0;
2715 for_each_online_cpu(i
) {
2716 running
+= cpu_rq(i
)->nr_running
;
2717 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2720 if (unlikely((long)uninterruptible
< 0))
2721 uninterruptible
= 0;
2723 return running
+ uninterruptible
;
2727 * Update rq->cpu_load[] statistics. This function is usually called every
2728 * scheduler tick (TICK_NSEC).
2730 static void update_cpu_load(struct rq
*this_rq
)
2732 unsigned long this_load
= this_rq
->load
.weight
;
2735 this_rq
->nr_load_updates
++;
2737 /* Update our load: */
2738 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2739 unsigned long old_load
, new_load
;
2741 /* scale is effectively 1 << i now, and >> i divides by scale */
2743 old_load
= this_rq
->cpu_load
[i
];
2744 new_load
= this_load
;
2746 * Round up the averaging division if load is increasing. This
2747 * prevents us from getting stuck on 9 if the load is 10, for
2750 if (new_load
> old_load
)
2751 new_load
+= scale
-1;
2752 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2759 * double_rq_lock - safely lock two runqueues
2761 * Note this does not disable interrupts like task_rq_lock,
2762 * you need to do so manually before calling.
2764 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2765 __acquires(rq1
->lock
)
2766 __acquires(rq2
->lock
)
2768 BUG_ON(!irqs_disabled());
2770 spin_lock(&rq1
->lock
);
2771 __acquire(rq2
->lock
); /* Fake it out ;) */
2774 spin_lock(&rq1
->lock
);
2775 spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
2777 spin_lock(&rq2
->lock
);
2778 spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
2781 update_rq_clock(rq1
);
2782 update_rq_clock(rq2
);
2786 * double_rq_unlock - safely unlock two runqueues
2788 * Note this does not restore interrupts like task_rq_unlock,
2789 * you need to do so manually after calling.
2791 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2792 __releases(rq1
->lock
)
2793 __releases(rq2
->lock
)
2795 spin_unlock(&rq1
->lock
);
2797 spin_unlock(&rq2
->lock
);
2799 __release(rq2
->lock
);
2803 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2805 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2806 __releases(this_rq
->lock
)
2807 __acquires(busiest
->lock
)
2808 __acquires(this_rq
->lock
)
2812 if (unlikely(!irqs_disabled())) {
2813 /* printk() doesn't work good under rq->lock */
2814 spin_unlock(&this_rq
->lock
);
2817 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2818 if (busiest
< this_rq
) {
2819 spin_unlock(&this_rq
->lock
);
2820 spin_lock(&busiest
->lock
);
2821 spin_lock_nested(&this_rq
->lock
, SINGLE_DEPTH_NESTING
);
2824 spin_lock_nested(&busiest
->lock
, SINGLE_DEPTH_NESTING
);
2829 static void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2830 __releases(busiest
->lock
)
2832 spin_unlock(&busiest
->lock
);
2833 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
2837 * If dest_cpu is allowed for this process, migrate the task to it.
2838 * This is accomplished by forcing the cpu_allowed mask to only
2839 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2840 * the cpu_allowed mask is restored.
2842 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2844 struct migration_req req
;
2845 unsigned long flags
;
2848 rq
= task_rq_lock(p
, &flags
);
2849 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2850 || unlikely(!cpu_active(dest_cpu
)))
2853 /* force the process onto the specified CPU */
2854 if (migrate_task(p
, dest_cpu
, &req
)) {
2855 /* Need to wait for migration thread (might exit: take ref). */
2856 struct task_struct
*mt
= rq
->migration_thread
;
2858 get_task_struct(mt
);
2859 task_rq_unlock(rq
, &flags
);
2860 wake_up_process(mt
);
2861 put_task_struct(mt
);
2862 wait_for_completion(&req
.done
);
2867 task_rq_unlock(rq
, &flags
);
2871 * sched_exec - execve() is a valuable balancing opportunity, because at
2872 * this point the task has the smallest effective memory and cache footprint.
2874 void sched_exec(void)
2876 int new_cpu
, this_cpu
= get_cpu();
2877 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2879 if (new_cpu
!= this_cpu
)
2880 sched_migrate_task(current
, new_cpu
);
2884 * pull_task - move a task from a remote runqueue to the local runqueue.
2885 * Both runqueues must be locked.
2887 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2888 struct rq
*this_rq
, int this_cpu
)
2890 deactivate_task(src_rq
, p
, 0);
2891 set_task_cpu(p
, this_cpu
);
2892 activate_task(this_rq
, p
, 0);
2894 * Note that idle threads have a prio of MAX_PRIO, for this test
2895 * to be always true for them.
2897 check_preempt_curr(this_rq
, p
, 0);
2901 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2904 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2905 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2909 * We do not migrate tasks that are:
2910 * 1) running (obviously), or
2911 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2912 * 3) are cache-hot on their current CPU.
2914 if (!cpu_isset(this_cpu
, p
->cpus_allowed
)) {
2915 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2920 if (task_running(rq
, p
)) {
2921 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2926 * Aggressive migration if:
2927 * 1) task is cache cold, or
2928 * 2) too many balance attempts have failed.
2931 if (!task_hot(p
, rq
->clock
, sd
) ||
2932 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2933 #ifdef CONFIG_SCHEDSTATS
2934 if (task_hot(p
, rq
->clock
, sd
)) {
2935 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2936 schedstat_inc(p
, se
.nr_forced_migrations
);
2942 if (task_hot(p
, rq
->clock
, sd
)) {
2943 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
2949 static unsigned long
2950 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2951 unsigned long max_load_move
, struct sched_domain
*sd
,
2952 enum cpu_idle_type idle
, int *all_pinned
,
2953 int *this_best_prio
, struct rq_iterator
*iterator
)
2955 int loops
= 0, pulled
= 0, pinned
= 0;
2956 struct task_struct
*p
;
2957 long rem_load_move
= max_load_move
;
2959 if (max_load_move
== 0)
2965 * Start the load-balancing iterator:
2967 p
= iterator
->start(iterator
->arg
);
2969 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
2972 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
2973 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2974 p
= iterator
->next(iterator
->arg
);
2978 pull_task(busiest
, p
, this_rq
, this_cpu
);
2980 rem_load_move
-= p
->se
.load
.weight
;
2983 * We only want to steal up to the prescribed amount of weighted load.
2985 if (rem_load_move
> 0) {
2986 if (p
->prio
< *this_best_prio
)
2987 *this_best_prio
= p
->prio
;
2988 p
= iterator
->next(iterator
->arg
);
2993 * Right now, this is one of only two places pull_task() is called,
2994 * so we can safely collect pull_task() stats here rather than
2995 * inside pull_task().
2997 schedstat_add(sd
, lb_gained
[idle
], pulled
);
3000 *all_pinned
= pinned
;
3002 return max_load_move
- rem_load_move
;
3006 * move_tasks tries to move up to max_load_move weighted load from busiest to
3007 * this_rq, as part of a balancing operation within domain "sd".
3008 * Returns 1 if successful and 0 otherwise.
3010 * Called with both runqueues locked.
3012 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3013 unsigned long max_load_move
,
3014 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3017 const struct sched_class
*class = sched_class_highest
;
3018 unsigned long total_load_moved
= 0;
3019 int this_best_prio
= this_rq
->curr
->prio
;
3023 class->load_balance(this_rq
, this_cpu
, busiest
,
3024 max_load_move
- total_load_moved
,
3025 sd
, idle
, all_pinned
, &this_best_prio
);
3026 class = class->next
;
3028 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3031 } while (class && max_load_move
> total_load_moved
);
3033 return total_load_moved
> 0;
3037 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3038 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3039 struct rq_iterator
*iterator
)
3041 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3045 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3046 pull_task(busiest
, p
, this_rq
, this_cpu
);
3048 * Right now, this is only the second place pull_task()
3049 * is called, so we can safely collect pull_task()
3050 * stats here rather than inside pull_task().
3052 schedstat_inc(sd
, lb_gained
[idle
]);
3056 p
= iterator
->next(iterator
->arg
);
3063 * move_one_task tries to move exactly one task from busiest to this_rq, as
3064 * part of active balancing operations within "domain".
3065 * Returns 1 if successful and 0 otherwise.
3067 * Called with both runqueues locked.
3069 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3070 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3072 const struct sched_class
*class;
3074 for (class = sched_class_highest
; class; class = class->next
)
3075 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3082 * find_busiest_group finds and returns the busiest CPU group within the
3083 * domain. It calculates and returns the amount of weighted load which
3084 * should be moved to restore balance via the imbalance parameter.
3086 static struct sched_group
*
3087 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
3088 unsigned long *imbalance
, enum cpu_idle_type idle
,
3089 int *sd_idle
, const cpumask_t
*cpus
, int *balance
)
3091 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
3092 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
3093 unsigned long max_pull
;
3094 unsigned long busiest_load_per_task
, busiest_nr_running
;
3095 unsigned long this_load_per_task
, this_nr_running
;
3096 int load_idx
, group_imb
= 0;
3097 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3098 int power_savings_balance
= 1;
3099 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
3100 unsigned long min_nr_running
= ULONG_MAX
;
3101 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
3104 max_load
= this_load
= total_load
= total_pwr
= 0;
3105 busiest_load_per_task
= busiest_nr_running
= 0;
3106 this_load_per_task
= this_nr_running
= 0;
3108 if (idle
== CPU_NOT_IDLE
)
3109 load_idx
= sd
->busy_idx
;
3110 else if (idle
== CPU_NEWLY_IDLE
)
3111 load_idx
= sd
->newidle_idx
;
3113 load_idx
= sd
->idle_idx
;
3116 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
3119 int __group_imb
= 0;
3120 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3121 unsigned long sum_nr_running
, sum_weighted_load
;
3122 unsigned long sum_avg_load_per_task
;
3123 unsigned long avg_load_per_task
;
3125 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
3128 balance_cpu
= first_cpu(group
->cpumask
);
3130 /* Tally up the load of all CPUs in the group */
3131 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
3132 sum_avg_load_per_task
= avg_load_per_task
= 0;
3135 min_cpu_load
= ~0UL;
3137 for_each_cpu_mask_nr(i
, group
->cpumask
) {
3140 if (!cpu_isset(i
, *cpus
))
3145 if (*sd_idle
&& rq
->nr_running
)
3148 /* Bias balancing toward cpus of our domain */
3150 if (idle_cpu(i
) && !first_idle_cpu
) {
3155 load
= target_load(i
, load_idx
);
3157 load
= source_load(i
, load_idx
);
3158 if (load
> max_cpu_load
)
3159 max_cpu_load
= load
;
3160 if (min_cpu_load
> load
)
3161 min_cpu_load
= load
;
3165 sum_nr_running
+= rq
->nr_running
;
3166 sum_weighted_load
+= weighted_cpuload(i
);
3168 sum_avg_load_per_task
+= cpu_avg_load_per_task(i
);
3172 * First idle cpu or the first cpu(busiest) in this sched group
3173 * is eligible for doing load balancing at this and above
3174 * domains. In the newly idle case, we will allow all the cpu's
3175 * to do the newly idle load balance.
3177 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3178 balance_cpu
!= this_cpu
&& balance
) {
3183 total_load
+= avg_load
;
3184 total_pwr
+= group
->__cpu_power
;
3186 /* Adjust by relative CPU power of the group */
3187 avg_load
= sg_div_cpu_power(group
,
3188 avg_load
* SCHED_LOAD_SCALE
);
3192 * Consider the group unbalanced when the imbalance is larger
3193 * than the average weight of two tasks.
3195 * APZ: with cgroup the avg task weight can vary wildly and
3196 * might not be a suitable number - should we keep a
3197 * normalized nr_running number somewhere that negates
3200 avg_load_per_task
= sg_div_cpu_power(group
,
3201 sum_avg_load_per_task
* SCHED_LOAD_SCALE
);
3203 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3206 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
3209 this_load
= avg_load
;
3211 this_nr_running
= sum_nr_running
;
3212 this_load_per_task
= sum_weighted_load
;
3213 } else if (avg_load
> max_load
&&
3214 (sum_nr_running
> group_capacity
|| __group_imb
)) {
3215 max_load
= avg_load
;
3217 busiest_nr_running
= sum_nr_running
;
3218 busiest_load_per_task
= sum_weighted_load
;
3219 group_imb
= __group_imb
;
3222 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3224 * Busy processors will not participate in power savings
3227 if (idle
== CPU_NOT_IDLE
||
3228 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3232 * If the local group is idle or completely loaded
3233 * no need to do power savings balance at this domain
3235 if (local_group
&& (this_nr_running
>= group_capacity
||
3237 power_savings_balance
= 0;
3240 * If a group is already running at full capacity or idle,
3241 * don't include that group in power savings calculations
3243 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
3248 * Calculate the group which has the least non-idle load.
3249 * This is the group from where we need to pick up the load
3252 if ((sum_nr_running
< min_nr_running
) ||
3253 (sum_nr_running
== min_nr_running
&&
3254 first_cpu(group
->cpumask
) <
3255 first_cpu(group_min
->cpumask
))) {
3257 min_nr_running
= sum_nr_running
;
3258 min_load_per_task
= sum_weighted_load
/
3263 * Calculate the group which is almost near its
3264 * capacity but still has some space to pick up some load
3265 * from other group and save more power
3267 if (sum_nr_running
<= group_capacity
- 1) {
3268 if (sum_nr_running
> leader_nr_running
||
3269 (sum_nr_running
== leader_nr_running
&&
3270 first_cpu(group
->cpumask
) >
3271 first_cpu(group_leader
->cpumask
))) {
3272 group_leader
= group
;
3273 leader_nr_running
= sum_nr_running
;
3278 group
= group
->next
;
3279 } while (group
!= sd
->groups
);
3281 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
3284 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
3286 if (this_load
>= avg_load
||
3287 100*max_load
<= sd
->imbalance_pct
*this_load
)
3290 busiest_load_per_task
/= busiest_nr_running
;
3292 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
3295 * We're trying to get all the cpus to the average_load, so we don't
3296 * want to push ourselves above the average load, nor do we wish to
3297 * reduce the max loaded cpu below the average load, as either of these
3298 * actions would just result in more rebalancing later, and ping-pong
3299 * tasks around. Thus we look for the minimum possible imbalance.
3300 * Negative imbalances (*we* are more loaded than anyone else) will
3301 * be counted as no imbalance for these purposes -- we can't fix that
3302 * by pulling tasks to us. Be careful of negative numbers as they'll
3303 * appear as very large values with unsigned longs.
3305 if (max_load
<= busiest_load_per_task
)
3309 * In the presence of smp nice balancing, certain scenarios can have
3310 * max load less than avg load(as we skip the groups at or below
3311 * its cpu_power, while calculating max_load..)
3313 if (max_load
< avg_load
) {
3315 goto small_imbalance
;
3318 /* Don't want to pull so many tasks that a group would go idle */
3319 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
3321 /* How much load to actually move to equalise the imbalance */
3322 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
3323 (avg_load
- this_load
) * this->__cpu_power
)
3327 * if *imbalance is less than the average load per runnable task
3328 * there is no gaurantee that any tasks will be moved so we'll have
3329 * a think about bumping its value to force at least one task to be
3332 if (*imbalance
< busiest_load_per_task
) {
3333 unsigned long tmp
, pwr_now
, pwr_move
;
3337 pwr_move
= pwr_now
= 0;
3339 if (this_nr_running
) {
3340 this_load_per_task
/= this_nr_running
;
3341 if (busiest_load_per_task
> this_load_per_task
)
3344 this_load_per_task
= cpu_avg_load_per_task(this_cpu
);
3346 if (max_load
- this_load
+ 2*busiest_load_per_task
>=
3347 busiest_load_per_task
* imbn
) {
3348 *imbalance
= busiest_load_per_task
;
3353 * OK, we don't have enough imbalance to justify moving tasks,
3354 * however we may be able to increase total CPU power used by
3358 pwr_now
+= busiest
->__cpu_power
*
3359 min(busiest_load_per_task
, max_load
);
3360 pwr_now
+= this->__cpu_power
*
3361 min(this_load_per_task
, this_load
);
3362 pwr_now
/= SCHED_LOAD_SCALE
;
3364 /* Amount of load we'd subtract */
3365 tmp
= sg_div_cpu_power(busiest
,
3366 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3368 pwr_move
+= busiest
->__cpu_power
*
3369 min(busiest_load_per_task
, max_load
- tmp
);
3371 /* Amount of load we'd add */
3372 if (max_load
* busiest
->__cpu_power
<
3373 busiest_load_per_task
* SCHED_LOAD_SCALE
)
3374 tmp
= sg_div_cpu_power(this,
3375 max_load
* busiest
->__cpu_power
);
3377 tmp
= sg_div_cpu_power(this,
3378 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3379 pwr_move
+= this->__cpu_power
*
3380 min(this_load_per_task
, this_load
+ tmp
);
3381 pwr_move
/= SCHED_LOAD_SCALE
;
3383 /* Move if we gain throughput */
3384 if (pwr_move
> pwr_now
)
3385 *imbalance
= busiest_load_per_task
;
3391 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3392 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3395 if (this == group_leader
&& group_leader
!= group_min
) {
3396 *imbalance
= min_load_per_task
;
3406 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3409 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
3410 unsigned long imbalance
, const cpumask_t
*cpus
)
3412 struct rq
*busiest
= NULL
, *rq
;
3413 unsigned long max_load
= 0;
3416 for_each_cpu_mask_nr(i
, group
->cpumask
) {
3419 if (!cpu_isset(i
, *cpus
))
3423 wl
= weighted_cpuload(i
);
3425 if (rq
->nr_running
== 1 && wl
> imbalance
)
3428 if (wl
> max_load
) {
3438 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3439 * so long as it is large enough.
3441 #define MAX_PINNED_INTERVAL 512
3444 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3445 * tasks if there is an imbalance.
3447 static int load_balance(int this_cpu
, struct rq
*this_rq
,
3448 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3449 int *balance
, cpumask_t
*cpus
)
3451 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
3452 struct sched_group
*group
;
3453 unsigned long imbalance
;
3455 unsigned long flags
;
3460 * When power savings policy is enabled for the parent domain, idle
3461 * sibling can pick up load irrespective of busy siblings. In this case,
3462 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3463 * portraying it as CPU_NOT_IDLE.
3465 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3466 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3469 schedstat_inc(sd
, lb_count
[idle
]);
3473 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3480 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3484 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
3486 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3490 BUG_ON(busiest
== this_rq
);
3492 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3495 if (busiest
->nr_running
> 1) {
3497 * Attempt to move tasks. If find_busiest_group has found
3498 * an imbalance but busiest->nr_running <= 1, the group is
3499 * still unbalanced. ld_moved simply stays zero, so it is
3500 * correctly treated as an imbalance.
3502 local_irq_save(flags
);
3503 double_rq_lock(this_rq
, busiest
);
3504 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3505 imbalance
, sd
, idle
, &all_pinned
);
3506 double_rq_unlock(this_rq
, busiest
);
3507 local_irq_restore(flags
);
3510 * some other cpu did the load balance for us.
3512 if (ld_moved
&& this_cpu
!= smp_processor_id())
3513 resched_cpu(this_cpu
);
3515 /* All tasks on this runqueue were pinned by CPU affinity */
3516 if (unlikely(all_pinned
)) {
3517 cpu_clear(cpu_of(busiest
), *cpus
);
3518 if (!cpus_empty(*cpus
))
3525 schedstat_inc(sd
, lb_failed
[idle
]);
3526 sd
->nr_balance_failed
++;
3528 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3530 spin_lock_irqsave(&busiest
->lock
, flags
);
3532 /* don't kick the migration_thread, if the curr
3533 * task on busiest cpu can't be moved to this_cpu
3535 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
3536 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3538 goto out_one_pinned
;
3541 if (!busiest
->active_balance
) {
3542 busiest
->active_balance
= 1;
3543 busiest
->push_cpu
= this_cpu
;
3546 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3548 wake_up_process(busiest
->migration_thread
);
3551 * We've kicked active balancing, reset the failure
3554 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3557 sd
->nr_balance_failed
= 0;
3559 if (likely(!active_balance
)) {
3560 /* We were unbalanced, so reset the balancing interval */
3561 sd
->balance_interval
= sd
->min_interval
;
3564 * If we've begun active balancing, start to back off. This
3565 * case may not be covered by the all_pinned logic if there
3566 * is only 1 task on the busy runqueue (because we don't call
3569 if (sd
->balance_interval
< sd
->max_interval
)
3570 sd
->balance_interval
*= 2;
3573 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3574 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3580 schedstat_inc(sd
, lb_balanced
[idle
]);
3582 sd
->nr_balance_failed
= 0;
3585 /* tune up the balancing interval */
3586 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3587 (sd
->balance_interval
< sd
->max_interval
))
3588 sd
->balance_interval
*= 2;
3590 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3591 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3602 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3603 * tasks if there is an imbalance.
3605 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3606 * this_rq is locked.
3609 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
,
3612 struct sched_group
*group
;
3613 struct rq
*busiest
= NULL
;
3614 unsigned long imbalance
;
3622 * When power savings policy is enabled for the parent domain, idle
3623 * sibling can pick up load irrespective of busy siblings. In this case,
3624 * let the state of idle sibling percolate up as IDLE, instead of
3625 * portraying it as CPU_NOT_IDLE.
3627 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
3628 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3631 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
3633 update_shares_locked(this_rq
, sd
);
3634 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
3635 &sd_idle
, cpus
, NULL
);
3637 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
3641 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
3643 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
3647 BUG_ON(busiest
== this_rq
);
3649 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
3652 if (busiest
->nr_running
> 1) {
3653 /* Attempt to move tasks */
3654 double_lock_balance(this_rq
, busiest
);
3655 /* this_rq->clock is already updated */
3656 update_rq_clock(busiest
);
3657 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3658 imbalance
, sd
, CPU_NEWLY_IDLE
,
3660 double_unlock_balance(this_rq
, busiest
);
3662 if (unlikely(all_pinned
)) {
3663 cpu_clear(cpu_of(busiest
), *cpus
);
3664 if (!cpus_empty(*cpus
))
3670 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
3671 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3672 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3675 sd
->nr_balance_failed
= 0;
3677 update_shares_locked(this_rq
, sd
);
3681 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
3682 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3683 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3685 sd
->nr_balance_failed
= 0;
3691 * idle_balance is called by schedule() if this_cpu is about to become
3692 * idle. Attempts to pull tasks from other CPUs.
3694 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
3696 struct sched_domain
*sd
;
3697 int pulled_task
= -1;
3698 unsigned long next_balance
= jiffies
+ HZ
;
3701 for_each_domain(this_cpu
, sd
) {
3702 unsigned long interval
;
3704 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3707 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
3708 /* If we've pulled tasks over stop searching: */
3709 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
3712 interval
= msecs_to_jiffies(sd
->balance_interval
);
3713 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3714 next_balance
= sd
->last_balance
+ interval
;
3718 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3720 * We are going idle. next_balance may be set based on
3721 * a busy processor. So reset next_balance.
3723 this_rq
->next_balance
= next_balance
;
3728 * active_load_balance is run by migration threads. It pushes running tasks
3729 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3730 * running on each physical CPU where possible, and avoids physical /
3731 * logical imbalances.
3733 * Called with busiest_rq locked.
3735 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
3737 int target_cpu
= busiest_rq
->push_cpu
;
3738 struct sched_domain
*sd
;
3739 struct rq
*target_rq
;
3741 /* Is there any task to move? */
3742 if (busiest_rq
->nr_running
<= 1)
3745 target_rq
= cpu_rq(target_cpu
);
3748 * This condition is "impossible", if it occurs
3749 * we need to fix it. Originally reported by
3750 * Bjorn Helgaas on a 128-cpu setup.
3752 BUG_ON(busiest_rq
== target_rq
);
3754 /* move a task from busiest_rq to target_rq */
3755 double_lock_balance(busiest_rq
, target_rq
);
3756 update_rq_clock(busiest_rq
);
3757 update_rq_clock(target_rq
);
3759 /* Search for an sd spanning us and the target CPU. */
3760 for_each_domain(target_cpu
, sd
) {
3761 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3762 cpu_isset(busiest_cpu
, sd
->span
))
3767 schedstat_inc(sd
, alb_count
);
3769 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3771 schedstat_inc(sd
, alb_pushed
);
3773 schedstat_inc(sd
, alb_failed
);
3775 double_unlock_balance(busiest_rq
, target_rq
);
3780 atomic_t load_balancer
;
3782 } nohz ____cacheline_aligned
= {
3783 .load_balancer
= ATOMIC_INIT(-1),
3784 .cpu_mask
= CPU_MASK_NONE
,
3788 * This routine will try to nominate the ilb (idle load balancing)
3789 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3790 * load balancing on behalf of all those cpus. If all the cpus in the system
3791 * go into this tickless mode, then there will be no ilb owner (as there is
3792 * no need for one) and all the cpus will sleep till the next wakeup event
3795 * For the ilb owner, tick is not stopped. And this tick will be used
3796 * for idle load balancing. ilb owner will still be part of
3799 * While stopping the tick, this cpu will become the ilb owner if there
3800 * is no other owner. And will be the owner till that cpu becomes busy
3801 * or if all cpus in the system stop their ticks at which point
3802 * there is no need for ilb owner.
3804 * When the ilb owner becomes busy, it nominates another owner, during the
3805 * next busy scheduler_tick()
3807 int select_nohz_load_balancer(int stop_tick
)
3809 int cpu
= smp_processor_id();
3812 cpu_set(cpu
, nohz
.cpu_mask
);
3813 cpu_rq(cpu
)->in_nohz_recently
= 1;
3816 * If we are going offline and still the leader, give up!
3818 if (!cpu_active(cpu
) &&
3819 atomic_read(&nohz
.load_balancer
) == cpu
) {
3820 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3825 /* time for ilb owner also to sleep */
3826 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3827 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3828 atomic_set(&nohz
.load_balancer
, -1);
3832 if (atomic_read(&nohz
.load_balancer
) == -1) {
3833 /* make me the ilb owner */
3834 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3836 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3839 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3842 cpu_clear(cpu
, nohz
.cpu_mask
);
3844 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3845 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3852 static DEFINE_SPINLOCK(balancing
);
3855 * It checks each scheduling domain to see if it is due to be balanced,
3856 * and initiates a balancing operation if so.
3858 * Balancing parameters are set up in arch_init_sched_domains.
3860 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3863 struct rq
*rq
= cpu_rq(cpu
);
3864 unsigned long interval
;
3865 struct sched_domain
*sd
;
3866 /* Earliest time when we have to do rebalance again */
3867 unsigned long next_balance
= jiffies
+ 60*HZ
;
3868 int update_next_balance
= 0;
3872 for_each_domain(cpu
, sd
) {
3873 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3876 interval
= sd
->balance_interval
;
3877 if (idle
!= CPU_IDLE
)
3878 interval
*= sd
->busy_factor
;
3880 /* scale ms to jiffies */
3881 interval
= msecs_to_jiffies(interval
);
3882 if (unlikely(!interval
))
3884 if (interval
> HZ
*NR_CPUS
/10)
3885 interval
= HZ
*NR_CPUS
/10;
3887 need_serialize
= sd
->flags
& SD_SERIALIZE
;
3889 if (need_serialize
) {
3890 if (!spin_trylock(&balancing
))
3894 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3895 if (load_balance(cpu
, rq
, sd
, idle
, &balance
, &tmp
)) {
3897 * We've pulled tasks over so either we're no
3898 * longer idle, or one of our SMT siblings is
3901 idle
= CPU_NOT_IDLE
;
3903 sd
->last_balance
= jiffies
;
3906 spin_unlock(&balancing
);
3908 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3909 next_balance
= sd
->last_balance
+ interval
;
3910 update_next_balance
= 1;
3914 * Stop the load balance at this level. There is another
3915 * CPU in our sched group which is doing load balancing more
3923 * next_balance will be updated only when there is a need.
3924 * When the cpu is attached to null domain for ex, it will not be
3927 if (likely(update_next_balance
))
3928 rq
->next_balance
= next_balance
;
3932 * run_rebalance_domains is triggered when needed from the scheduler tick.
3933 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3934 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3936 static void run_rebalance_domains(struct softirq_action
*h
)
3938 int this_cpu
= smp_processor_id();
3939 struct rq
*this_rq
= cpu_rq(this_cpu
);
3940 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3941 CPU_IDLE
: CPU_NOT_IDLE
;
3943 rebalance_domains(this_cpu
, idle
);
3947 * If this cpu is the owner for idle load balancing, then do the
3948 * balancing on behalf of the other idle cpus whose ticks are
3951 if (this_rq
->idle_at_tick
&&
3952 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3953 cpumask_t cpus
= nohz
.cpu_mask
;
3957 cpu_clear(this_cpu
, cpus
);
3958 for_each_cpu_mask_nr(balance_cpu
, cpus
) {
3960 * If this cpu gets work to do, stop the load balancing
3961 * work being done for other cpus. Next load
3962 * balancing owner will pick it up.
3967 rebalance_domains(balance_cpu
, CPU_IDLE
);
3969 rq
= cpu_rq(balance_cpu
);
3970 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3971 this_rq
->next_balance
= rq
->next_balance
;
3978 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3980 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3981 * idle load balancing owner or decide to stop the periodic load balancing,
3982 * if the whole system is idle.
3984 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3988 * If we were in the nohz mode recently and busy at the current
3989 * scheduler tick, then check if we need to nominate new idle
3992 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3993 rq
->in_nohz_recently
= 0;
3995 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3996 cpu_clear(cpu
, nohz
.cpu_mask
);
3997 atomic_set(&nohz
.load_balancer
, -1);
4000 if (atomic_read(&nohz
.load_balancer
) == -1) {
4002 * simple selection for now: Nominate the
4003 * first cpu in the nohz list to be the next
4006 * TBD: Traverse the sched domains and nominate
4007 * the nearest cpu in the nohz.cpu_mask.
4009 int ilb
= first_cpu(nohz
.cpu_mask
);
4011 if (ilb
< nr_cpu_ids
)
4017 * If this cpu is idle and doing idle load balancing for all the
4018 * cpus with ticks stopped, is it time for that to stop?
4020 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4021 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4027 * If this cpu is idle and the idle load balancing is done by
4028 * someone else, then no need raise the SCHED_SOFTIRQ
4030 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4031 cpu_isset(cpu
, nohz
.cpu_mask
))
4034 if (time_after_eq(jiffies
, rq
->next_balance
))
4035 raise_softirq(SCHED_SOFTIRQ
);
4038 #else /* CONFIG_SMP */
4041 * on UP we do not need to balance between CPUs:
4043 static inline void idle_balance(int cpu
, struct rq
*rq
)
4049 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4051 EXPORT_PER_CPU_SYMBOL(kstat
);
4054 * Return p->sum_exec_runtime plus any more ns on the sched_clock
4055 * that have not yet been banked in case the task is currently running.
4057 unsigned long long task_sched_runtime(struct task_struct
*p
)
4059 unsigned long flags
;
4063 rq
= task_rq_lock(p
, &flags
);
4064 ns
= p
->se
.sum_exec_runtime
;
4065 if (task_current(rq
, p
)) {
4066 update_rq_clock(rq
);
4067 delta_exec
= rq
->clock
- p
->se
.exec_start
;
4068 if ((s64
)delta_exec
> 0)
4071 task_rq_unlock(rq
, &flags
);
4077 * Account user cpu time to a process.
4078 * @p: the process that the cpu time gets accounted to
4079 * @cputime: the cpu time spent in user space since the last update
4081 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
4083 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4086 p
->utime
= cputime_add(p
->utime
, cputime
);
4088 /* Add user time to cpustat. */
4089 tmp
= cputime_to_cputime64(cputime
);
4090 if (TASK_NICE(p
) > 0)
4091 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
4093 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4094 /* Account for user time used */
4095 acct_update_integrals(p
);
4099 * Account guest cpu time to a process.
4100 * @p: the process that the cpu time gets accounted to
4101 * @cputime: the cpu time spent in virtual machine since the last update
4103 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
)
4106 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4108 tmp
= cputime_to_cputime64(cputime
);
4110 p
->utime
= cputime_add(p
->utime
, cputime
);
4111 p
->gtime
= cputime_add(p
->gtime
, cputime
);
4113 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4114 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
4118 * Account scaled user cpu time to a process.
4119 * @p: the process that the cpu time gets accounted to
4120 * @cputime: the cpu time spent in user space since the last update
4122 void account_user_time_scaled(struct task_struct
*p
, cputime_t cputime
)
4124 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime
);
4128 * Account system cpu time to a process.
4129 * @p: the process that the cpu time gets accounted to
4130 * @hardirq_offset: the offset to subtract from hardirq_count()
4131 * @cputime: the cpu time spent in kernel space since the last update
4133 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
4136 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4137 struct rq
*rq
= this_rq();
4140 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
4141 account_guest_time(p
, cputime
);
4145 p
->stime
= cputime_add(p
->stime
, cputime
);
4147 /* Add system time to cpustat. */
4148 tmp
= cputime_to_cputime64(cputime
);
4149 if (hardirq_count() - hardirq_offset
)
4150 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
4151 else if (softirq_count())
4152 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
4153 else if (p
!= rq
->idle
)
4154 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
4155 else if (atomic_read(&rq
->nr_iowait
) > 0)
4156 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
4158 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
4159 /* Account for system time used */
4160 acct_update_integrals(p
);
4164 * Account scaled system cpu time to a process.
4165 * @p: the process that the cpu time gets accounted to
4166 * @hardirq_offset: the offset to subtract from hardirq_count()
4167 * @cputime: the cpu time spent in kernel space since the last update
4169 void account_system_time_scaled(struct task_struct
*p
, cputime_t cputime
)
4171 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime
);
4175 * Account for involuntary wait time.
4176 * @p: the process from which the cpu time has been stolen
4177 * @steal: the cpu time spent in involuntary wait
4179 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
4181 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4182 cputime64_t tmp
= cputime_to_cputime64(steal
);
4183 struct rq
*rq
= this_rq();
4185 if (p
== rq
->idle
) {
4186 p
->stime
= cputime_add(p
->stime
, steal
);
4187 if (atomic_read(&rq
->nr_iowait
) > 0)
4188 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
4190 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
4192 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
4196 * Use precise platform statistics if available:
4198 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4199 cputime_t
task_utime(struct task_struct
*p
)
4204 cputime_t
task_stime(struct task_struct
*p
)
4209 cputime_t
task_utime(struct task_struct
*p
)
4211 clock_t utime
= cputime_to_clock_t(p
->utime
),
4212 total
= utime
+ cputime_to_clock_t(p
->stime
);
4216 * Use CFS's precise accounting:
4218 temp
= (u64
)nsec_to_clock_t(p
->se
.sum_exec_runtime
);
4222 do_div(temp
, total
);
4224 utime
= (clock_t)temp
;
4226 p
->prev_utime
= max(p
->prev_utime
, clock_t_to_cputime(utime
));
4227 return p
->prev_utime
;
4230 cputime_t
task_stime(struct task_struct
*p
)
4235 * Use CFS's precise accounting. (we subtract utime from
4236 * the total, to make sure the total observed by userspace
4237 * grows monotonically - apps rely on that):
4239 stime
= nsec_to_clock_t(p
->se
.sum_exec_runtime
) -
4240 cputime_to_clock_t(task_utime(p
));
4243 p
->prev_stime
= max(p
->prev_stime
, clock_t_to_cputime(stime
));
4245 return p
->prev_stime
;
4249 inline cputime_t
task_gtime(struct task_struct
*p
)
4255 * This function gets called by the timer code, with HZ frequency.
4256 * We call it with interrupts disabled.
4258 * It also gets called by the fork code, when changing the parent's
4261 void scheduler_tick(void)
4263 int cpu
= smp_processor_id();
4264 struct rq
*rq
= cpu_rq(cpu
);
4265 struct task_struct
*curr
= rq
->curr
;
4269 spin_lock(&rq
->lock
);
4270 update_rq_clock(rq
);
4271 update_cpu_load(rq
);
4272 curr
->sched_class
->task_tick(rq
, curr
, 0);
4273 spin_unlock(&rq
->lock
);
4276 rq
->idle_at_tick
= idle_cpu(cpu
);
4277 trigger_load_balance(rq
, cpu
);
4281 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4282 defined(CONFIG_PREEMPT_TRACER))
4284 static inline unsigned long get_parent_ip(unsigned long addr
)
4286 if (in_lock_functions(addr
)) {
4287 addr
= CALLER_ADDR2
;
4288 if (in_lock_functions(addr
))
4289 addr
= CALLER_ADDR3
;
4294 void __kprobes
add_preempt_count(int val
)
4296 #ifdef CONFIG_DEBUG_PREEMPT
4300 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4303 preempt_count() += val
;
4304 #ifdef CONFIG_DEBUG_PREEMPT
4306 * Spinlock count overflowing soon?
4308 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4311 if (preempt_count() == val
)
4312 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4314 EXPORT_SYMBOL(add_preempt_count
);
4316 void __kprobes
sub_preempt_count(int val
)
4318 #ifdef CONFIG_DEBUG_PREEMPT
4322 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
4325 * Is the spinlock portion underflowing?
4327 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4328 !(preempt_count() & PREEMPT_MASK
)))
4332 if (preempt_count() == val
)
4333 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4334 preempt_count() -= val
;
4336 EXPORT_SYMBOL(sub_preempt_count
);
4341 * Print scheduling while atomic bug:
4343 static noinline
void __schedule_bug(struct task_struct
*prev
)
4345 struct pt_regs
*regs
= get_irq_regs();
4347 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4348 prev
->comm
, prev
->pid
, preempt_count());
4350 debug_show_held_locks(prev
);
4352 if (irqs_disabled())
4353 print_irqtrace_events(prev
);
4362 * Various schedule()-time debugging checks and statistics:
4364 static inline void schedule_debug(struct task_struct
*prev
)
4367 * Test if we are atomic. Since do_exit() needs to call into
4368 * schedule() atomically, we ignore that path for now.
4369 * Otherwise, whine if we are scheduling when we should not be.
4371 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
4372 __schedule_bug(prev
);
4374 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4376 schedstat_inc(this_rq(), sched_count
);
4377 #ifdef CONFIG_SCHEDSTATS
4378 if (unlikely(prev
->lock_depth
>= 0)) {
4379 schedstat_inc(this_rq(), bkl_count
);
4380 schedstat_inc(prev
, sched_info
.bkl_count
);
4386 * Pick up the highest-prio task:
4388 static inline struct task_struct
*
4389 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
4391 const struct sched_class
*class;
4392 struct task_struct
*p
;
4395 * Optimization: we know that if all tasks are in
4396 * the fair class we can call that function directly:
4398 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
4399 p
= fair_sched_class
.pick_next_task(rq
);
4404 class = sched_class_highest
;
4406 p
= class->pick_next_task(rq
);
4410 * Will never be NULL as the idle class always
4411 * returns a non-NULL p:
4413 class = class->next
;
4418 * schedule() is the main scheduler function.
4420 asmlinkage
void __sched
schedule(void)
4422 struct task_struct
*prev
, *next
;
4423 unsigned long *switch_count
;
4429 cpu
= smp_processor_id();
4433 switch_count
= &prev
->nivcsw
;
4435 release_kernel_lock(prev
);
4436 need_resched_nonpreemptible
:
4438 schedule_debug(prev
);
4440 if (sched_feat(HRTICK
))
4444 * Do the rq-clock update outside the rq lock:
4446 local_irq_disable();
4447 update_rq_clock(rq
);
4448 spin_lock(&rq
->lock
);
4449 clear_tsk_need_resched(prev
);
4451 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
4452 if (unlikely(signal_pending_state(prev
->state
, prev
)))
4453 prev
->state
= TASK_RUNNING
;
4455 deactivate_task(rq
, prev
, 1);
4456 switch_count
= &prev
->nvcsw
;
4460 if (prev
->sched_class
->pre_schedule
)
4461 prev
->sched_class
->pre_schedule(rq
, prev
);
4464 if (unlikely(!rq
->nr_running
))
4465 idle_balance(cpu
, rq
);
4467 prev
->sched_class
->put_prev_task(rq
, prev
);
4468 next
= pick_next_task(rq
, prev
);
4470 if (likely(prev
!= next
)) {
4471 sched_info_switch(prev
, next
);
4477 context_switch(rq
, prev
, next
); /* unlocks the rq */
4479 * the context switch might have flipped the stack from under
4480 * us, hence refresh the local variables.
4482 cpu
= smp_processor_id();
4485 spin_unlock_irq(&rq
->lock
);
4487 if (unlikely(reacquire_kernel_lock(current
) < 0))
4488 goto need_resched_nonpreemptible
;
4490 preempt_enable_no_resched();
4491 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
4494 EXPORT_SYMBOL(schedule
);
4496 #ifdef CONFIG_PREEMPT
4498 * this is the entry point to schedule() from in-kernel preemption
4499 * off of preempt_enable. Kernel preemptions off return from interrupt
4500 * occur there and call schedule directly.
4502 asmlinkage
void __sched
preempt_schedule(void)
4504 struct thread_info
*ti
= current_thread_info();
4507 * If there is a non-zero preempt_count or interrupts are disabled,
4508 * we do not want to preempt the current task. Just return..
4510 if (likely(ti
->preempt_count
|| irqs_disabled()))
4514 add_preempt_count(PREEMPT_ACTIVE
);
4516 sub_preempt_count(PREEMPT_ACTIVE
);
4519 * Check again in case we missed a preemption opportunity
4520 * between schedule and now.
4523 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4525 EXPORT_SYMBOL(preempt_schedule
);
4528 * this is the entry point to schedule() from kernel preemption
4529 * off of irq context.
4530 * Note, that this is called and return with irqs disabled. This will
4531 * protect us against recursive calling from irq.
4533 asmlinkage
void __sched
preempt_schedule_irq(void)
4535 struct thread_info
*ti
= current_thread_info();
4537 /* Catch callers which need to be fixed */
4538 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
4541 add_preempt_count(PREEMPT_ACTIVE
);
4544 local_irq_disable();
4545 sub_preempt_count(PREEMPT_ACTIVE
);
4548 * Check again in case we missed a preemption opportunity
4549 * between schedule and now.
4552 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4555 #endif /* CONFIG_PREEMPT */
4557 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
4560 return try_to_wake_up(curr
->private, mode
, sync
);
4562 EXPORT_SYMBOL(default_wake_function
);
4565 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4566 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4567 * number) then we wake all the non-exclusive tasks and one exclusive task.
4569 * There are circumstances in which we can try to wake a task which has already
4570 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4571 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4573 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4574 int nr_exclusive
, int sync
, void *key
)
4576 wait_queue_t
*curr
, *next
;
4578 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4579 unsigned flags
= curr
->flags
;
4581 if (curr
->func(curr
, mode
, sync
, key
) &&
4582 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4588 * __wake_up - wake up threads blocked on a waitqueue.
4590 * @mode: which threads
4591 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4592 * @key: is directly passed to the wakeup function
4594 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4595 int nr_exclusive
, void *key
)
4597 unsigned long flags
;
4599 spin_lock_irqsave(&q
->lock
, flags
);
4600 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4601 spin_unlock_irqrestore(&q
->lock
, flags
);
4603 EXPORT_SYMBOL(__wake_up
);
4606 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4608 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4610 __wake_up_common(q
, mode
, 1, 0, NULL
);
4614 * __wake_up_sync - wake up threads blocked on a waitqueue.
4616 * @mode: which threads
4617 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4619 * The sync wakeup differs that the waker knows that it will schedule
4620 * away soon, so while the target thread will be woken up, it will not
4621 * be migrated to another CPU - ie. the two threads are 'synchronized'
4622 * with each other. This can prevent needless bouncing between CPUs.
4624 * On UP it can prevent extra preemption.
4627 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4629 unsigned long flags
;
4635 if (unlikely(!nr_exclusive
))
4638 spin_lock_irqsave(&q
->lock
, flags
);
4639 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
4640 spin_unlock_irqrestore(&q
->lock
, flags
);
4642 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4645 * complete: - signals a single thread waiting on this completion
4646 * @x: holds the state of this particular completion
4648 * This will wake up a single thread waiting on this completion. Threads will be
4649 * awakened in the same order in which they were queued.
4651 * See also complete_all(), wait_for_completion() and related routines.
4653 void complete(struct completion
*x
)
4655 unsigned long flags
;
4657 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4659 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4660 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4662 EXPORT_SYMBOL(complete
);
4665 * complete_all: - signals all threads waiting on this completion
4666 * @x: holds the state of this particular completion
4668 * This will wake up all threads waiting on this particular completion event.
4670 void complete_all(struct completion
*x
)
4672 unsigned long flags
;
4674 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4675 x
->done
+= UINT_MAX
/2;
4676 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4677 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4679 EXPORT_SYMBOL(complete_all
);
4681 static inline long __sched
4682 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4685 DECLARE_WAITQUEUE(wait
, current
);
4687 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4688 __add_wait_queue_tail(&x
->wait
, &wait
);
4690 if (signal_pending_state(state
, current
)) {
4691 timeout
= -ERESTARTSYS
;
4694 __set_current_state(state
);
4695 spin_unlock_irq(&x
->wait
.lock
);
4696 timeout
= schedule_timeout(timeout
);
4697 spin_lock_irq(&x
->wait
.lock
);
4698 } while (!x
->done
&& timeout
);
4699 __remove_wait_queue(&x
->wait
, &wait
);
4704 return timeout
?: 1;
4708 wait_for_common(struct completion
*x
, long timeout
, int state
)
4712 spin_lock_irq(&x
->wait
.lock
);
4713 timeout
= do_wait_for_common(x
, timeout
, state
);
4714 spin_unlock_irq(&x
->wait
.lock
);
4719 * wait_for_completion: - waits for completion of a task
4720 * @x: holds the state of this particular completion
4722 * This waits to be signaled for completion of a specific task. It is NOT
4723 * interruptible and there is no timeout.
4725 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4726 * and interrupt capability. Also see complete().
4728 void __sched
wait_for_completion(struct completion
*x
)
4730 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4732 EXPORT_SYMBOL(wait_for_completion
);
4735 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4736 * @x: holds the state of this particular completion
4737 * @timeout: timeout value in jiffies
4739 * This waits for either a completion of a specific task to be signaled or for a
4740 * specified timeout to expire. The timeout is in jiffies. It is not
4743 unsigned long __sched
4744 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4746 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4748 EXPORT_SYMBOL(wait_for_completion_timeout
);
4751 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4752 * @x: holds the state of this particular completion
4754 * This waits for completion of a specific task to be signaled. It is
4757 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4759 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4760 if (t
== -ERESTARTSYS
)
4764 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4767 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4768 * @x: holds the state of this particular completion
4769 * @timeout: timeout value in jiffies
4771 * This waits for either a completion of a specific task to be signaled or for a
4772 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4774 unsigned long __sched
4775 wait_for_completion_interruptible_timeout(struct completion
*x
,
4776 unsigned long timeout
)
4778 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4780 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4783 * wait_for_completion_killable: - waits for completion of a task (killable)
4784 * @x: holds the state of this particular completion
4786 * This waits to be signaled for completion of a specific task. It can be
4787 * interrupted by a kill signal.
4789 int __sched
wait_for_completion_killable(struct completion
*x
)
4791 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4792 if (t
== -ERESTARTSYS
)
4796 EXPORT_SYMBOL(wait_for_completion_killable
);
4799 * try_wait_for_completion - try to decrement a completion without blocking
4800 * @x: completion structure
4802 * Returns: 0 if a decrement cannot be done without blocking
4803 * 1 if a decrement succeeded.
4805 * If a completion is being used as a counting completion,
4806 * attempt to decrement the counter without blocking. This
4807 * enables us to avoid waiting if the resource the completion
4808 * is protecting is not available.
4810 bool try_wait_for_completion(struct completion
*x
)
4814 spin_lock_irq(&x
->wait
.lock
);
4819 spin_unlock_irq(&x
->wait
.lock
);
4822 EXPORT_SYMBOL(try_wait_for_completion
);
4825 * completion_done - Test to see if a completion has any waiters
4826 * @x: completion structure
4828 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4829 * 1 if there are no waiters.
4832 bool completion_done(struct completion
*x
)
4836 spin_lock_irq(&x
->wait
.lock
);
4839 spin_unlock_irq(&x
->wait
.lock
);
4842 EXPORT_SYMBOL(completion_done
);
4845 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4847 unsigned long flags
;
4850 init_waitqueue_entry(&wait
, current
);
4852 __set_current_state(state
);
4854 spin_lock_irqsave(&q
->lock
, flags
);
4855 __add_wait_queue(q
, &wait
);
4856 spin_unlock(&q
->lock
);
4857 timeout
= schedule_timeout(timeout
);
4858 spin_lock_irq(&q
->lock
);
4859 __remove_wait_queue(q
, &wait
);
4860 spin_unlock_irqrestore(&q
->lock
, flags
);
4865 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4867 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4869 EXPORT_SYMBOL(interruptible_sleep_on
);
4872 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4874 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4876 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4878 void __sched
sleep_on(wait_queue_head_t
*q
)
4880 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4882 EXPORT_SYMBOL(sleep_on
);
4884 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4886 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4888 EXPORT_SYMBOL(sleep_on_timeout
);
4890 #ifdef CONFIG_RT_MUTEXES
4893 * rt_mutex_setprio - set the current priority of a task
4895 * @prio: prio value (kernel-internal form)
4897 * This function changes the 'effective' priority of a task. It does
4898 * not touch ->normal_prio like __setscheduler().
4900 * Used by the rt_mutex code to implement priority inheritance logic.
4902 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4904 unsigned long flags
;
4905 int oldprio
, on_rq
, running
;
4907 const struct sched_class
*prev_class
= p
->sched_class
;
4909 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4911 rq
= task_rq_lock(p
, &flags
);
4912 update_rq_clock(rq
);
4915 on_rq
= p
->se
.on_rq
;
4916 running
= task_current(rq
, p
);
4918 dequeue_task(rq
, p
, 0);
4920 p
->sched_class
->put_prev_task(rq
, p
);
4923 p
->sched_class
= &rt_sched_class
;
4925 p
->sched_class
= &fair_sched_class
;
4930 p
->sched_class
->set_curr_task(rq
);
4932 enqueue_task(rq
, p
, 0);
4934 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4936 task_rq_unlock(rq
, &flags
);
4941 void set_user_nice(struct task_struct
*p
, long nice
)
4943 int old_prio
, delta
, on_rq
;
4944 unsigned long flags
;
4947 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4950 * We have to be careful, if called from sys_setpriority(),
4951 * the task might be in the middle of scheduling on another CPU.
4953 rq
= task_rq_lock(p
, &flags
);
4954 update_rq_clock(rq
);
4956 * The RT priorities are set via sched_setscheduler(), but we still
4957 * allow the 'normal' nice value to be set - but as expected
4958 * it wont have any effect on scheduling until the task is
4959 * SCHED_FIFO/SCHED_RR:
4961 if (task_has_rt_policy(p
)) {
4962 p
->static_prio
= NICE_TO_PRIO(nice
);
4965 on_rq
= p
->se
.on_rq
;
4967 dequeue_task(rq
, p
, 0);
4969 p
->static_prio
= NICE_TO_PRIO(nice
);
4972 p
->prio
= effective_prio(p
);
4973 delta
= p
->prio
- old_prio
;
4976 enqueue_task(rq
, p
, 0);
4978 * If the task increased its priority or is running and
4979 * lowered its priority, then reschedule its CPU:
4981 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4982 resched_task(rq
->curr
);
4985 task_rq_unlock(rq
, &flags
);
4987 EXPORT_SYMBOL(set_user_nice
);
4990 * can_nice - check if a task can reduce its nice value
4994 int can_nice(const struct task_struct
*p
, const int nice
)
4996 /* convert nice value [19,-20] to rlimit style value [1,40] */
4997 int nice_rlim
= 20 - nice
;
4999 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
5000 capable(CAP_SYS_NICE
));
5003 #ifdef __ARCH_WANT_SYS_NICE
5006 * sys_nice - change the priority of the current process.
5007 * @increment: priority increment
5009 * sys_setpriority is a more generic, but much slower function that
5010 * does similar things.
5012 asmlinkage
long sys_nice(int increment
)
5017 * Setpriority might change our priority at the same moment.
5018 * We don't have to worry. Conceptually one call occurs first
5019 * and we have a single winner.
5021 if (increment
< -40)
5026 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
5032 if (increment
< 0 && !can_nice(current
, nice
))
5035 retval
= security_task_setnice(current
, nice
);
5039 set_user_nice(current
, nice
);
5046 * task_prio - return the priority value of a given task.
5047 * @p: the task in question.
5049 * This is the priority value as seen by users in /proc.
5050 * RT tasks are offset by -200. Normal tasks are centered
5051 * around 0, value goes from -16 to +15.
5053 int task_prio(const struct task_struct
*p
)
5055 return p
->prio
- MAX_RT_PRIO
;
5059 * task_nice - return the nice value of a given task.
5060 * @p: the task in question.
5062 int task_nice(const struct task_struct
*p
)
5064 return TASK_NICE(p
);
5066 EXPORT_SYMBOL(task_nice
);
5069 * idle_cpu - is a given cpu idle currently?
5070 * @cpu: the processor in question.
5072 int idle_cpu(int cpu
)
5074 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
5078 * idle_task - return the idle task for a given cpu.
5079 * @cpu: the processor in question.
5081 struct task_struct
*idle_task(int cpu
)
5083 return cpu_rq(cpu
)->idle
;
5087 * find_process_by_pid - find a process with a matching PID value.
5088 * @pid: the pid in question.
5090 static struct task_struct
*find_process_by_pid(pid_t pid
)
5092 return pid
? find_task_by_vpid(pid
) : current
;
5095 /* Actually do priority change: must hold rq lock. */
5097 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
5099 BUG_ON(p
->se
.on_rq
);
5102 switch (p
->policy
) {
5106 p
->sched_class
= &fair_sched_class
;
5110 p
->sched_class
= &rt_sched_class
;
5114 p
->rt_priority
= prio
;
5115 p
->normal_prio
= normal_prio(p
);
5116 /* we are holding p->pi_lock already */
5117 p
->prio
= rt_mutex_getprio(p
);
5121 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
5122 struct sched_param
*param
, bool user
)
5124 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
5125 unsigned long flags
;
5126 const struct sched_class
*prev_class
= p
->sched_class
;
5129 /* may grab non-irq protected spin_locks */
5130 BUG_ON(in_interrupt());
5132 /* double check policy once rq lock held */
5134 policy
= oldpolicy
= p
->policy
;
5135 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
5136 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
5137 policy
!= SCHED_IDLE
)
5140 * Valid priorities for SCHED_FIFO and SCHED_RR are
5141 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5142 * SCHED_BATCH and SCHED_IDLE is 0.
5144 if (param
->sched_priority
< 0 ||
5145 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
5146 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
5148 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
5152 * Allow unprivileged RT tasks to decrease priority:
5154 if (user
&& !capable(CAP_SYS_NICE
)) {
5155 if (rt_policy(policy
)) {
5156 unsigned long rlim_rtprio
;
5158 if (!lock_task_sighand(p
, &flags
))
5160 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
5161 unlock_task_sighand(p
, &flags
);
5163 /* can't set/change the rt policy */
5164 if (policy
!= p
->policy
&& !rlim_rtprio
)
5167 /* can't increase priority */
5168 if (param
->sched_priority
> p
->rt_priority
&&
5169 param
->sched_priority
> rlim_rtprio
)
5173 * Like positive nice levels, dont allow tasks to
5174 * move out of SCHED_IDLE either:
5176 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
5179 /* can't change other user's priorities */
5180 if ((current
->euid
!= p
->euid
) &&
5181 (current
->euid
!= p
->uid
))
5186 #ifdef CONFIG_RT_GROUP_SCHED
5188 * Do not allow realtime tasks into groups that have no runtime
5191 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
5192 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
5196 retval
= security_task_setscheduler(p
, policy
, param
);
5202 * make sure no PI-waiters arrive (or leave) while we are
5203 * changing the priority of the task:
5205 spin_lock_irqsave(&p
->pi_lock
, flags
);
5207 * To be able to change p->policy safely, the apropriate
5208 * runqueue lock must be held.
5210 rq
= __task_rq_lock(p
);
5211 /* recheck policy now with rq lock held */
5212 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
5213 policy
= oldpolicy
= -1;
5214 __task_rq_unlock(rq
);
5215 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5218 update_rq_clock(rq
);
5219 on_rq
= p
->se
.on_rq
;
5220 running
= task_current(rq
, p
);
5222 deactivate_task(rq
, p
, 0);
5224 p
->sched_class
->put_prev_task(rq
, p
);
5227 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
5230 p
->sched_class
->set_curr_task(rq
);
5232 activate_task(rq
, p
, 0);
5234 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5236 __task_rq_unlock(rq
);
5237 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5239 rt_mutex_adjust_pi(p
);
5245 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5246 * @p: the task in question.
5247 * @policy: new policy.
5248 * @param: structure containing the new RT priority.
5250 * NOTE that the task may be already dead.
5252 int sched_setscheduler(struct task_struct
*p
, int policy
,
5253 struct sched_param
*param
)
5255 return __sched_setscheduler(p
, policy
, param
, true);
5257 EXPORT_SYMBOL_GPL(sched_setscheduler
);
5260 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5261 * @p: the task in question.
5262 * @policy: new policy.
5263 * @param: structure containing the new RT priority.
5265 * Just like sched_setscheduler, only don't bother checking if the
5266 * current context has permission. For example, this is needed in
5267 * stop_machine(): we create temporary high priority worker threads,
5268 * but our caller might not have that capability.
5270 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
5271 struct sched_param
*param
)
5273 return __sched_setscheduler(p
, policy
, param
, false);
5277 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5279 struct sched_param lparam
;
5280 struct task_struct
*p
;
5283 if (!param
|| pid
< 0)
5285 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
5290 p
= find_process_by_pid(pid
);
5292 retval
= sched_setscheduler(p
, policy
, &lparam
);
5299 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5300 * @pid: the pid in question.
5301 * @policy: new policy.
5302 * @param: structure containing the new RT priority.
5305 sys_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5307 /* negative values for policy are not valid */
5311 return do_sched_setscheduler(pid
, policy
, param
);
5315 * sys_sched_setparam - set/change the RT priority of a thread
5316 * @pid: the pid in question.
5317 * @param: structure containing the new RT priority.
5319 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
5321 return do_sched_setscheduler(pid
, -1, param
);
5325 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5326 * @pid: the pid in question.
5328 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
5330 struct task_struct
*p
;
5337 read_lock(&tasklist_lock
);
5338 p
= find_process_by_pid(pid
);
5340 retval
= security_task_getscheduler(p
);
5344 read_unlock(&tasklist_lock
);
5349 * sys_sched_getscheduler - get the RT priority of a thread
5350 * @pid: the pid in question.
5351 * @param: structure containing the RT priority.
5353 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
5355 struct sched_param lp
;
5356 struct task_struct
*p
;
5359 if (!param
|| pid
< 0)
5362 read_lock(&tasklist_lock
);
5363 p
= find_process_by_pid(pid
);
5368 retval
= security_task_getscheduler(p
);
5372 lp
.sched_priority
= p
->rt_priority
;
5373 read_unlock(&tasklist_lock
);
5376 * This one might sleep, we cannot do it with a spinlock held ...
5378 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
5383 read_unlock(&tasklist_lock
);
5387 long sched_setaffinity(pid_t pid
, const cpumask_t
*in_mask
)
5389 cpumask_t cpus_allowed
;
5390 cpumask_t new_mask
= *in_mask
;
5391 struct task_struct
*p
;
5395 read_lock(&tasklist_lock
);
5397 p
= find_process_by_pid(pid
);
5399 read_unlock(&tasklist_lock
);
5405 * It is not safe to call set_cpus_allowed with the
5406 * tasklist_lock held. We will bump the task_struct's
5407 * usage count and then drop tasklist_lock.
5410 read_unlock(&tasklist_lock
);
5413 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
5414 !capable(CAP_SYS_NICE
))
5417 retval
= security_task_setscheduler(p
, 0, NULL
);
5421 cpuset_cpus_allowed(p
, &cpus_allowed
);
5422 cpus_and(new_mask
, new_mask
, cpus_allowed
);
5424 retval
= set_cpus_allowed_ptr(p
, &new_mask
);
5427 cpuset_cpus_allowed(p
, &cpus_allowed
);
5428 if (!cpus_subset(new_mask
, cpus_allowed
)) {
5430 * We must have raced with a concurrent cpuset
5431 * update. Just reset the cpus_allowed to the
5432 * cpuset's cpus_allowed
5434 new_mask
= cpus_allowed
;
5444 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5445 cpumask_t
*new_mask
)
5447 if (len
< sizeof(cpumask_t
)) {
5448 memset(new_mask
, 0, sizeof(cpumask_t
));
5449 } else if (len
> sizeof(cpumask_t
)) {
5450 len
= sizeof(cpumask_t
);
5452 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
5456 * sys_sched_setaffinity - set the cpu affinity of a process
5457 * @pid: pid of the process
5458 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5459 * @user_mask_ptr: user-space pointer to the new cpu mask
5461 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
5462 unsigned long __user
*user_mask_ptr
)
5467 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
5471 return sched_setaffinity(pid
, &new_mask
);
5474 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
5476 struct task_struct
*p
;
5480 read_lock(&tasklist_lock
);
5483 p
= find_process_by_pid(pid
);
5487 retval
= security_task_getscheduler(p
);
5491 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
5494 read_unlock(&tasklist_lock
);
5501 * sys_sched_getaffinity - get the cpu affinity of a process
5502 * @pid: pid of the process
5503 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5504 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5506 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
5507 unsigned long __user
*user_mask_ptr
)
5512 if (len
< sizeof(cpumask_t
))
5515 ret
= sched_getaffinity(pid
, &mask
);
5519 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
5522 return sizeof(cpumask_t
);
5526 * sys_sched_yield - yield the current processor to other threads.
5528 * This function yields the current CPU to other tasks. If there are no
5529 * other threads running on this CPU then this function will return.
5531 asmlinkage
long sys_sched_yield(void)
5533 struct rq
*rq
= this_rq_lock();
5535 schedstat_inc(rq
, yld_count
);
5536 current
->sched_class
->yield_task(rq
);
5539 * Since we are going to call schedule() anyway, there's
5540 * no need to preempt or enable interrupts:
5542 __release(rq
->lock
);
5543 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5544 _raw_spin_unlock(&rq
->lock
);
5545 preempt_enable_no_resched();
5552 static void __cond_resched(void)
5554 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5555 __might_sleep(__FILE__
, __LINE__
);
5558 * The BKS might be reacquired before we have dropped
5559 * PREEMPT_ACTIVE, which could trigger a second
5560 * cond_resched() call.
5563 add_preempt_count(PREEMPT_ACTIVE
);
5565 sub_preempt_count(PREEMPT_ACTIVE
);
5566 } while (need_resched());
5569 int __sched
_cond_resched(void)
5571 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
5572 system_state
== SYSTEM_RUNNING
) {
5578 EXPORT_SYMBOL(_cond_resched
);
5581 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5582 * call schedule, and on return reacquire the lock.
5584 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5585 * operations here to prevent schedule() from being called twice (once via
5586 * spin_unlock(), once by hand).
5588 int cond_resched_lock(spinlock_t
*lock
)
5590 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
5593 if (spin_needbreak(lock
) || resched
) {
5595 if (resched
&& need_resched())
5604 EXPORT_SYMBOL(cond_resched_lock
);
5606 int __sched
cond_resched_softirq(void)
5608 BUG_ON(!in_softirq());
5610 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
5618 EXPORT_SYMBOL(cond_resched_softirq
);
5621 * yield - yield the current processor to other threads.
5623 * This is a shortcut for kernel-space yielding - it marks the
5624 * thread runnable and calls sys_sched_yield().
5626 void __sched
yield(void)
5628 set_current_state(TASK_RUNNING
);
5631 EXPORT_SYMBOL(yield
);
5634 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5635 * that process accounting knows that this is a task in IO wait state.
5637 * But don't do that if it is a deliberate, throttling IO wait (this task
5638 * has set its backing_dev_info: the queue against which it should throttle)
5640 void __sched
io_schedule(void)
5642 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5644 delayacct_blkio_start();
5645 atomic_inc(&rq
->nr_iowait
);
5647 atomic_dec(&rq
->nr_iowait
);
5648 delayacct_blkio_end();
5650 EXPORT_SYMBOL(io_schedule
);
5652 long __sched
io_schedule_timeout(long timeout
)
5654 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5657 delayacct_blkio_start();
5658 atomic_inc(&rq
->nr_iowait
);
5659 ret
= schedule_timeout(timeout
);
5660 atomic_dec(&rq
->nr_iowait
);
5661 delayacct_blkio_end();
5666 * sys_sched_get_priority_max - return maximum RT priority.
5667 * @policy: scheduling class.
5669 * this syscall returns the maximum rt_priority that can be used
5670 * by a given scheduling class.
5672 asmlinkage
long sys_sched_get_priority_max(int policy
)
5679 ret
= MAX_USER_RT_PRIO
-1;
5691 * sys_sched_get_priority_min - return minimum RT priority.
5692 * @policy: scheduling class.
5694 * this syscall returns the minimum rt_priority that can be used
5695 * by a given scheduling class.
5697 asmlinkage
long sys_sched_get_priority_min(int policy
)
5715 * sys_sched_rr_get_interval - return the default timeslice of a process.
5716 * @pid: pid of the process.
5717 * @interval: userspace pointer to the timeslice value.
5719 * this syscall writes the default timeslice value of a given process
5720 * into the user-space timespec buffer. A value of '0' means infinity.
5723 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
5725 struct task_struct
*p
;
5726 unsigned int time_slice
;
5734 read_lock(&tasklist_lock
);
5735 p
= find_process_by_pid(pid
);
5739 retval
= security_task_getscheduler(p
);
5744 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5745 * tasks that are on an otherwise idle runqueue:
5748 if (p
->policy
== SCHED_RR
) {
5749 time_slice
= DEF_TIMESLICE
;
5750 } else if (p
->policy
!= SCHED_FIFO
) {
5751 struct sched_entity
*se
= &p
->se
;
5752 unsigned long flags
;
5755 rq
= task_rq_lock(p
, &flags
);
5756 if (rq
->cfs
.load
.weight
)
5757 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
5758 task_rq_unlock(rq
, &flags
);
5760 read_unlock(&tasklist_lock
);
5761 jiffies_to_timespec(time_slice
, &t
);
5762 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5766 read_unlock(&tasklist_lock
);
5770 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5772 void sched_show_task(struct task_struct
*p
)
5774 unsigned long free
= 0;
5777 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5778 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5779 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5780 #if BITS_PER_LONG == 32
5781 if (state
== TASK_RUNNING
)
5782 printk(KERN_CONT
" running ");
5784 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5786 if (state
== TASK_RUNNING
)
5787 printk(KERN_CONT
" running task ");
5789 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5791 #ifdef CONFIG_DEBUG_STACK_USAGE
5793 unsigned long *n
= end_of_stack(p
);
5796 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
5799 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
5800 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
5802 show_stack(p
, NULL
);
5805 void show_state_filter(unsigned long state_filter
)
5807 struct task_struct
*g
, *p
;
5809 #if BITS_PER_LONG == 32
5811 " task PC stack pid father\n");
5814 " task PC stack pid father\n");
5816 read_lock(&tasklist_lock
);
5817 do_each_thread(g
, p
) {
5819 * reset the NMI-timeout, listing all files on a slow
5820 * console might take alot of time:
5822 touch_nmi_watchdog();
5823 if (!state_filter
|| (p
->state
& state_filter
))
5825 } while_each_thread(g
, p
);
5827 touch_all_softlockup_watchdogs();
5829 #ifdef CONFIG_SCHED_DEBUG
5830 sysrq_sched_debug_show();
5832 read_unlock(&tasklist_lock
);
5834 * Only show locks if all tasks are dumped:
5836 if (state_filter
== -1)
5837 debug_show_all_locks();
5840 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5842 idle
->sched_class
= &idle_sched_class
;
5846 * init_idle - set up an idle thread for a given CPU
5847 * @idle: task in question
5848 * @cpu: cpu the idle task belongs to
5850 * NOTE: this function does not set the idle thread's NEED_RESCHED
5851 * flag, to make booting more robust.
5853 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5855 struct rq
*rq
= cpu_rq(cpu
);
5856 unsigned long flags
;
5859 idle
->se
.exec_start
= sched_clock();
5861 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
5862 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
5863 __set_task_cpu(idle
, cpu
);
5865 spin_lock_irqsave(&rq
->lock
, flags
);
5866 rq
->curr
= rq
->idle
= idle
;
5867 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5870 spin_unlock_irqrestore(&rq
->lock
, flags
);
5872 /* Set the preempt count _outside_ the spinlocks! */
5873 #if defined(CONFIG_PREEMPT)
5874 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5876 task_thread_info(idle
)->preempt_count
= 0;
5879 * The idle tasks have their own, simple scheduling class:
5881 idle
->sched_class
= &idle_sched_class
;
5885 * In a system that switches off the HZ timer nohz_cpu_mask
5886 * indicates which cpus entered this state. This is used
5887 * in the rcu update to wait only for active cpus. For system
5888 * which do not switch off the HZ timer nohz_cpu_mask should
5889 * always be CPU_MASK_NONE.
5891 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
5894 * Increase the granularity value when there are more CPUs,
5895 * because with more CPUs the 'effective latency' as visible
5896 * to users decreases. But the relationship is not linear,
5897 * so pick a second-best guess by going with the log2 of the
5900 * This idea comes from the SD scheduler of Con Kolivas:
5902 static inline void sched_init_granularity(void)
5904 unsigned int factor
= 1 + ilog2(num_online_cpus());
5905 const unsigned long limit
= 200000000;
5907 sysctl_sched_min_granularity
*= factor
;
5908 if (sysctl_sched_min_granularity
> limit
)
5909 sysctl_sched_min_granularity
= limit
;
5911 sysctl_sched_latency
*= factor
;
5912 if (sysctl_sched_latency
> limit
)
5913 sysctl_sched_latency
= limit
;
5915 sysctl_sched_wakeup_granularity
*= factor
;
5917 sysctl_sched_shares_ratelimit
*= factor
;
5922 * This is how migration works:
5924 * 1) we queue a struct migration_req structure in the source CPU's
5925 * runqueue and wake up that CPU's migration thread.
5926 * 2) we down() the locked semaphore => thread blocks.
5927 * 3) migration thread wakes up (implicitly it forces the migrated
5928 * thread off the CPU)
5929 * 4) it gets the migration request and checks whether the migrated
5930 * task is still in the wrong runqueue.
5931 * 5) if it's in the wrong runqueue then the migration thread removes
5932 * it and puts it into the right queue.
5933 * 6) migration thread up()s the semaphore.
5934 * 7) we wake up and the migration is done.
5938 * Change a given task's CPU affinity. Migrate the thread to a
5939 * proper CPU and schedule it away if the CPU it's executing on
5940 * is removed from the allowed bitmask.
5942 * NOTE: the caller must have a valid reference to the task, the
5943 * task must not exit() & deallocate itself prematurely. The
5944 * call is not atomic; no spinlocks may be held.
5946 int set_cpus_allowed_ptr(struct task_struct
*p
, const cpumask_t
*new_mask
)
5948 struct migration_req req
;
5949 unsigned long flags
;
5953 rq
= task_rq_lock(p
, &flags
);
5954 if (!cpus_intersects(*new_mask
, cpu_online_map
)) {
5959 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
5960 !cpus_equal(p
->cpus_allowed
, *new_mask
))) {
5965 if (p
->sched_class
->set_cpus_allowed
)
5966 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
5968 p
->cpus_allowed
= *new_mask
;
5969 p
->rt
.nr_cpus_allowed
= cpus_weight(*new_mask
);
5972 /* Can the task run on the task's current CPU? If so, we're done */
5973 if (cpu_isset(task_cpu(p
), *new_mask
))
5976 if (migrate_task(p
, any_online_cpu(*new_mask
), &req
)) {
5977 /* Need help from migration thread: drop lock and wait. */
5978 task_rq_unlock(rq
, &flags
);
5979 wake_up_process(rq
->migration_thread
);
5980 wait_for_completion(&req
.done
);
5981 tlb_migrate_finish(p
->mm
);
5985 task_rq_unlock(rq
, &flags
);
5989 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5992 * Move (not current) task off this cpu, onto dest cpu. We're doing
5993 * this because either it can't run here any more (set_cpus_allowed()
5994 * away from this CPU, or CPU going down), or because we're
5995 * attempting to rebalance this task on exec (sched_exec).
5997 * So we race with normal scheduler movements, but that's OK, as long
5998 * as the task is no longer on this CPU.
6000 * Returns non-zero if task was successfully migrated.
6002 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6004 struct rq
*rq_dest
, *rq_src
;
6007 if (unlikely(!cpu_active(dest_cpu
)))
6010 rq_src
= cpu_rq(src_cpu
);
6011 rq_dest
= cpu_rq(dest_cpu
);
6013 double_rq_lock(rq_src
, rq_dest
);
6014 /* Already moved. */
6015 if (task_cpu(p
) != src_cpu
)
6017 /* Affinity changed (again). */
6018 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
6021 on_rq
= p
->se
.on_rq
;
6023 deactivate_task(rq_src
, p
, 0);
6025 set_task_cpu(p
, dest_cpu
);
6027 activate_task(rq_dest
, p
, 0);
6028 check_preempt_curr(rq_dest
, p
, 0);
6033 double_rq_unlock(rq_src
, rq_dest
);
6038 * migration_thread - this is a highprio system thread that performs
6039 * thread migration by bumping thread off CPU then 'pushing' onto
6042 static int migration_thread(void *data
)
6044 int cpu
= (long)data
;
6048 BUG_ON(rq
->migration_thread
!= current
);
6050 set_current_state(TASK_INTERRUPTIBLE
);
6051 while (!kthread_should_stop()) {
6052 struct migration_req
*req
;
6053 struct list_head
*head
;
6055 spin_lock_irq(&rq
->lock
);
6057 if (cpu_is_offline(cpu
)) {
6058 spin_unlock_irq(&rq
->lock
);
6062 if (rq
->active_balance
) {
6063 active_load_balance(rq
, cpu
);
6064 rq
->active_balance
= 0;
6067 head
= &rq
->migration_queue
;
6069 if (list_empty(head
)) {
6070 spin_unlock_irq(&rq
->lock
);
6072 set_current_state(TASK_INTERRUPTIBLE
);
6075 req
= list_entry(head
->next
, struct migration_req
, list
);
6076 list_del_init(head
->next
);
6078 spin_unlock(&rq
->lock
);
6079 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
6082 complete(&req
->done
);
6084 __set_current_state(TASK_RUNNING
);
6088 /* Wait for kthread_stop */
6089 set_current_state(TASK_INTERRUPTIBLE
);
6090 while (!kthread_should_stop()) {
6092 set_current_state(TASK_INTERRUPTIBLE
);
6094 __set_current_state(TASK_RUNNING
);
6098 #ifdef CONFIG_HOTPLUG_CPU
6100 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6104 local_irq_disable();
6105 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
6111 * Figure out where task on dead CPU should go, use force if necessary.
6112 * NOTE: interrupts should be disabled by the caller
6114 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
6116 unsigned long flags
;
6123 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
6124 cpus_and(mask
, mask
, p
->cpus_allowed
);
6125 dest_cpu
= any_online_cpu(mask
);
6127 /* On any allowed CPU? */
6128 if (dest_cpu
>= nr_cpu_ids
)
6129 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
6131 /* No more Mr. Nice Guy. */
6132 if (dest_cpu
>= nr_cpu_ids
) {
6133 cpumask_t cpus_allowed
;
6135 cpuset_cpus_allowed_locked(p
, &cpus_allowed
);
6137 * Try to stay on the same cpuset, where the
6138 * current cpuset may be a subset of all cpus.
6139 * The cpuset_cpus_allowed_locked() variant of
6140 * cpuset_cpus_allowed() will not block. It must be
6141 * called within calls to cpuset_lock/cpuset_unlock.
6143 rq
= task_rq_lock(p
, &flags
);
6144 p
->cpus_allowed
= cpus_allowed
;
6145 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
6146 task_rq_unlock(rq
, &flags
);
6149 * Don't tell them about moving exiting tasks or
6150 * kernel threads (both mm NULL), since they never
6153 if (p
->mm
&& printk_ratelimit()) {
6154 printk(KERN_INFO
"process %d (%s) no "
6155 "longer affine to cpu%d\n",
6156 task_pid_nr(p
), p
->comm
, dead_cpu
);
6159 } while (!__migrate_task_irq(p
, dead_cpu
, dest_cpu
));
6163 * While a dead CPU has no uninterruptible tasks queued at this point,
6164 * it might still have a nonzero ->nr_uninterruptible counter, because
6165 * for performance reasons the counter is not stricly tracking tasks to
6166 * their home CPUs. So we just add the counter to another CPU's counter,
6167 * to keep the global sum constant after CPU-down:
6169 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
6171 struct rq
*rq_dest
= cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR
));
6172 unsigned long flags
;
6174 local_irq_save(flags
);
6175 double_rq_lock(rq_src
, rq_dest
);
6176 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
6177 rq_src
->nr_uninterruptible
= 0;
6178 double_rq_unlock(rq_src
, rq_dest
);
6179 local_irq_restore(flags
);
6182 /* Run through task list and migrate tasks from the dead cpu. */
6183 static void migrate_live_tasks(int src_cpu
)
6185 struct task_struct
*p
, *t
;
6187 read_lock(&tasklist_lock
);
6189 do_each_thread(t
, p
) {
6193 if (task_cpu(p
) == src_cpu
)
6194 move_task_off_dead_cpu(src_cpu
, p
);
6195 } while_each_thread(t
, p
);
6197 read_unlock(&tasklist_lock
);
6201 * Schedules idle task to be the next runnable task on current CPU.
6202 * It does so by boosting its priority to highest possible.
6203 * Used by CPU offline code.
6205 void sched_idle_next(void)
6207 int this_cpu
= smp_processor_id();
6208 struct rq
*rq
= cpu_rq(this_cpu
);
6209 struct task_struct
*p
= rq
->idle
;
6210 unsigned long flags
;
6212 /* cpu has to be offline */
6213 BUG_ON(cpu_online(this_cpu
));
6216 * Strictly not necessary since rest of the CPUs are stopped by now
6217 * and interrupts disabled on the current cpu.
6219 spin_lock_irqsave(&rq
->lock
, flags
);
6221 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6223 update_rq_clock(rq
);
6224 activate_task(rq
, p
, 0);
6226 spin_unlock_irqrestore(&rq
->lock
, flags
);
6230 * Ensures that the idle task is using init_mm right before its cpu goes
6233 void idle_task_exit(void)
6235 struct mm_struct
*mm
= current
->active_mm
;
6237 BUG_ON(cpu_online(smp_processor_id()));
6240 switch_mm(mm
, &init_mm
, current
);
6244 /* called under rq->lock with disabled interrupts */
6245 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
6247 struct rq
*rq
= cpu_rq(dead_cpu
);
6249 /* Must be exiting, otherwise would be on tasklist. */
6250 BUG_ON(!p
->exit_state
);
6252 /* Cannot have done final schedule yet: would have vanished. */
6253 BUG_ON(p
->state
== TASK_DEAD
);
6258 * Drop lock around migration; if someone else moves it,
6259 * that's OK. No task can be added to this CPU, so iteration is
6262 spin_unlock_irq(&rq
->lock
);
6263 move_task_off_dead_cpu(dead_cpu
, p
);
6264 spin_lock_irq(&rq
->lock
);
6269 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6270 static void migrate_dead_tasks(unsigned int dead_cpu
)
6272 struct rq
*rq
= cpu_rq(dead_cpu
);
6273 struct task_struct
*next
;
6276 if (!rq
->nr_running
)
6278 update_rq_clock(rq
);
6279 next
= pick_next_task(rq
, rq
->curr
);
6282 next
->sched_class
->put_prev_task(rq
, next
);
6283 migrate_dead(dead_cpu
, next
);
6287 #endif /* CONFIG_HOTPLUG_CPU */
6289 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6291 static struct ctl_table sd_ctl_dir
[] = {
6293 .procname
= "sched_domain",
6299 static struct ctl_table sd_ctl_root
[] = {
6301 .ctl_name
= CTL_KERN
,
6302 .procname
= "kernel",
6304 .child
= sd_ctl_dir
,
6309 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
6311 struct ctl_table
*entry
=
6312 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
6317 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
6319 struct ctl_table
*entry
;
6322 * In the intermediate directories, both the child directory and
6323 * procname are dynamically allocated and could fail but the mode
6324 * will always be set. In the lowest directory the names are
6325 * static strings and all have proc handlers.
6327 for (entry
= *tablep
; entry
->mode
; entry
++) {
6329 sd_free_ctl_entry(&entry
->child
);
6330 if (entry
->proc_handler
== NULL
)
6331 kfree(entry
->procname
);
6339 set_table_entry(struct ctl_table
*entry
,
6340 const char *procname
, void *data
, int maxlen
,
6341 mode_t mode
, proc_handler
*proc_handler
)
6343 entry
->procname
= procname
;
6345 entry
->maxlen
= maxlen
;
6347 entry
->proc_handler
= proc_handler
;
6350 static struct ctl_table
*
6351 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
6353 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
6358 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
6359 sizeof(long), 0644, proc_doulongvec_minmax
);
6360 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
6361 sizeof(long), 0644, proc_doulongvec_minmax
);
6362 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
6363 sizeof(int), 0644, proc_dointvec_minmax
);
6364 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
6365 sizeof(int), 0644, proc_dointvec_minmax
);
6366 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
6367 sizeof(int), 0644, proc_dointvec_minmax
);
6368 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
6369 sizeof(int), 0644, proc_dointvec_minmax
);
6370 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
6371 sizeof(int), 0644, proc_dointvec_minmax
);
6372 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
6373 sizeof(int), 0644, proc_dointvec_minmax
);
6374 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
6375 sizeof(int), 0644, proc_dointvec_minmax
);
6376 set_table_entry(&table
[9], "cache_nice_tries",
6377 &sd
->cache_nice_tries
,
6378 sizeof(int), 0644, proc_dointvec_minmax
);
6379 set_table_entry(&table
[10], "flags", &sd
->flags
,
6380 sizeof(int), 0644, proc_dointvec_minmax
);
6381 set_table_entry(&table
[11], "name", sd
->name
,
6382 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
6383 /* &table[12] is terminator */
6388 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
6390 struct ctl_table
*entry
, *table
;
6391 struct sched_domain
*sd
;
6392 int domain_num
= 0, i
;
6395 for_each_domain(cpu
, sd
)
6397 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
6402 for_each_domain(cpu
, sd
) {
6403 snprintf(buf
, 32, "domain%d", i
);
6404 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6406 entry
->child
= sd_alloc_ctl_domain_table(sd
);
6413 static struct ctl_table_header
*sd_sysctl_header
;
6414 static void register_sched_domain_sysctl(void)
6416 int i
, cpu_num
= num_online_cpus();
6417 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
6420 WARN_ON(sd_ctl_dir
[0].child
);
6421 sd_ctl_dir
[0].child
= entry
;
6426 for_each_online_cpu(i
) {
6427 snprintf(buf
, 32, "cpu%d", i
);
6428 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6430 entry
->child
= sd_alloc_ctl_cpu_table(i
);
6434 WARN_ON(sd_sysctl_header
);
6435 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
6438 /* may be called multiple times per register */
6439 static void unregister_sched_domain_sysctl(void)
6441 if (sd_sysctl_header
)
6442 unregister_sysctl_table(sd_sysctl_header
);
6443 sd_sysctl_header
= NULL
;
6444 if (sd_ctl_dir
[0].child
)
6445 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
6448 static void register_sched_domain_sysctl(void)
6451 static void unregister_sched_domain_sysctl(void)
6456 static void set_rq_online(struct rq
*rq
)
6459 const struct sched_class
*class;
6461 cpu_set(rq
->cpu
, rq
->rd
->online
);
6464 for_each_class(class) {
6465 if (class->rq_online
)
6466 class->rq_online(rq
);
6471 static void set_rq_offline(struct rq
*rq
)
6474 const struct sched_class
*class;
6476 for_each_class(class) {
6477 if (class->rq_offline
)
6478 class->rq_offline(rq
);
6481 cpu_clear(rq
->cpu
, rq
->rd
->online
);
6487 * migration_call - callback that gets triggered when a CPU is added.
6488 * Here we can start up the necessary migration thread for the new CPU.
6490 static int __cpuinit
6491 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
6493 struct task_struct
*p
;
6494 int cpu
= (long)hcpu
;
6495 unsigned long flags
;
6500 case CPU_UP_PREPARE
:
6501 case CPU_UP_PREPARE_FROZEN
:
6502 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
6505 kthread_bind(p
, cpu
);
6506 /* Must be high prio: stop_machine expects to yield to it. */
6507 rq
= task_rq_lock(p
, &flags
);
6508 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6509 task_rq_unlock(rq
, &flags
);
6510 cpu_rq(cpu
)->migration_thread
= p
;
6514 case CPU_ONLINE_FROZEN
:
6515 /* Strictly unnecessary, as first user will wake it. */
6516 wake_up_process(cpu_rq(cpu
)->migration_thread
);
6518 /* Update our root-domain */
6520 spin_lock_irqsave(&rq
->lock
, flags
);
6522 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6526 spin_unlock_irqrestore(&rq
->lock
, flags
);
6529 #ifdef CONFIG_HOTPLUG_CPU
6530 case CPU_UP_CANCELED
:
6531 case CPU_UP_CANCELED_FROZEN
:
6532 if (!cpu_rq(cpu
)->migration_thread
)
6534 /* Unbind it from offline cpu so it can run. Fall thru. */
6535 kthread_bind(cpu_rq(cpu
)->migration_thread
,
6536 any_online_cpu(cpu_online_map
));
6537 kthread_stop(cpu_rq(cpu
)->migration_thread
);
6538 cpu_rq(cpu
)->migration_thread
= NULL
;
6542 case CPU_DEAD_FROZEN
:
6543 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6544 migrate_live_tasks(cpu
);
6546 kthread_stop(rq
->migration_thread
);
6547 rq
->migration_thread
= NULL
;
6548 /* Idle task back to normal (off runqueue, low prio) */
6549 spin_lock_irq(&rq
->lock
);
6550 update_rq_clock(rq
);
6551 deactivate_task(rq
, rq
->idle
, 0);
6552 rq
->idle
->static_prio
= MAX_PRIO
;
6553 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
6554 rq
->idle
->sched_class
= &idle_sched_class
;
6555 migrate_dead_tasks(cpu
);
6556 spin_unlock_irq(&rq
->lock
);
6558 migrate_nr_uninterruptible(rq
);
6559 BUG_ON(rq
->nr_running
!= 0);
6562 * No need to migrate the tasks: it was best-effort if
6563 * they didn't take sched_hotcpu_mutex. Just wake up
6566 spin_lock_irq(&rq
->lock
);
6567 while (!list_empty(&rq
->migration_queue
)) {
6568 struct migration_req
*req
;
6570 req
= list_entry(rq
->migration_queue
.next
,
6571 struct migration_req
, list
);
6572 list_del_init(&req
->list
);
6573 complete(&req
->done
);
6575 spin_unlock_irq(&rq
->lock
);
6579 case CPU_DYING_FROZEN
:
6580 /* Update our root-domain */
6582 spin_lock_irqsave(&rq
->lock
, flags
);
6584 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6587 spin_unlock_irqrestore(&rq
->lock
, flags
);
6594 /* Register at highest priority so that task migration (migrate_all_tasks)
6595 * happens before everything else.
6597 static struct notifier_block __cpuinitdata migration_notifier
= {
6598 .notifier_call
= migration_call
,
6602 static int __init
migration_init(void)
6604 void *cpu
= (void *)(long)smp_processor_id();
6607 /* Start one for the boot CPU: */
6608 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6609 BUG_ON(err
== NOTIFY_BAD
);
6610 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6611 register_cpu_notifier(&migration_notifier
);
6615 early_initcall(migration_init
);
6620 #ifdef CONFIG_SCHED_DEBUG
6622 static inline const char *sd_level_to_string(enum sched_domain_level lvl
)
6635 case SD_LV_ALLNODES
:
6644 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6645 cpumask_t
*groupmask
)
6647 struct sched_group
*group
= sd
->groups
;
6650 cpulist_scnprintf(str
, sizeof(str
), sd
->span
);
6651 cpus_clear(*groupmask
);
6653 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6655 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6656 printk("does not load-balance\n");
6658 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6663 printk(KERN_CONT
"span %s level %s\n",
6664 str
, sd_level_to_string(sd
->level
));
6666 if (!cpu_isset(cpu
, sd
->span
)) {
6667 printk(KERN_ERR
"ERROR: domain->span does not contain "
6670 if (!cpu_isset(cpu
, group
->cpumask
)) {
6671 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6675 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6679 printk(KERN_ERR
"ERROR: group is NULL\n");
6683 if (!group
->__cpu_power
) {
6684 printk(KERN_CONT
"\n");
6685 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6690 if (!cpus_weight(group
->cpumask
)) {
6691 printk(KERN_CONT
"\n");
6692 printk(KERN_ERR
"ERROR: empty group\n");
6696 if (cpus_intersects(*groupmask
, group
->cpumask
)) {
6697 printk(KERN_CONT
"\n");
6698 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6702 cpus_or(*groupmask
, *groupmask
, group
->cpumask
);
6704 cpulist_scnprintf(str
, sizeof(str
), group
->cpumask
);
6705 printk(KERN_CONT
" %s", str
);
6707 group
= group
->next
;
6708 } while (group
!= sd
->groups
);
6709 printk(KERN_CONT
"\n");
6711 if (!cpus_equal(sd
->span
, *groupmask
))
6712 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6714 if (sd
->parent
&& !cpus_subset(*groupmask
, sd
->parent
->span
))
6715 printk(KERN_ERR
"ERROR: parent span is not a superset "
6716 "of domain->span\n");
6720 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6722 cpumask_t
*groupmask
;
6726 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6730 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6732 groupmask
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
6734 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6739 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6748 #else /* !CONFIG_SCHED_DEBUG */
6749 # define sched_domain_debug(sd, cpu) do { } while (0)
6750 #endif /* CONFIG_SCHED_DEBUG */
6752 static int sd_degenerate(struct sched_domain
*sd
)
6754 if (cpus_weight(sd
->span
) == 1)
6757 /* Following flags need at least 2 groups */
6758 if (sd
->flags
& (SD_LOAD_BALANCE
|
6759 SD_BALANCE_NEWIDLE
|
6763 SD_SHARE_PKG_RESOURCES
)) {
6764 if (sd
->groups
!= sd
->groups
->next
)
6768 /* Following flags don't use groups */
6769 if (sd
->flags
& (SD_WAKE_IDLE
|
6778 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6780 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6782 if (sd_degenerate(parent
))
6785 if (!cpus_equal(sd
->span
, parent
->span
))
6788 /* Does parent contain flags not in child? */
6789 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6790 if (cflags
& SD_WAKE_AFFINE
)
6791 pflags
&= ~SD_WAKE_BALANCE
;
6792 /* Flags needing groups don't count if only 1 group in parent */
6793 if (parent
->groups
== parent
->groups
->next
) {
6794 pflags
&= ~(SD_LOAD_BALANCE
|
6795 SD_BALANCE_NEWIDLE
|
6799 SD_SHARE_PKG_RESOURCES
);
6801 if (~cflags
& pflags
)
6807 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6809 unsigned long flags
;
6811 spin_lock_irqsave(&rq
->lock
, flags
);
6814 struct root_domain
*old_rd
= rq
->rd
;
6816 if (cpu_isset(rq
->cpu
, old_rd
->online
))
6819 cpu_clear(rq
->cpu
, old_rd
->span
);
6821 if (atomic_dec_and_test(&old_rd
->refcount
))
6825 atomic_inc(&rd
->refcount
);
6828 cpu_set(rq
->cpu
, rd
->span
);
6829 if (cpu_isset(rq
->cpu
, cpu_online_map
))
6832 spin_unlock_irqrestore(&rq
->lock
, flags
);
6835 static void init_rootdomain(struct root_domain
*rd
)
6837 memset(rd
, 0, sizeof(*rd
));
6839 cpus_clear(rd
->span
);
6840 cpus_clear(rd
->online
);
6842 cpupri_init(&rd
->cpupri
);
6845 static void init_defrootdomain(void)
6847 init_rootdomain(&def_root_domain
);
6848 atomic_set(&def_root_domain
.refcount
, 1);
6851 static struct root_domain
*alloc_rootdomain(void)
6853 struct root_domain
*rd
;
6855 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6859 init_rootdomain(rd
);
6865 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6866 * hold the hotplug lock.
6869 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6871 struct rq
*rq
= cpu_rq(cpu
);
6872 struct sched_domain
*tmp
;
6874 /* Remove the sched domains which do not contribute to scheduling. */
6875 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
6876 struct sched_domain
*parent
= tmp
->parent
;
6879 if (sd_parent_degenerate(tmp
, parent
)) {
6880 tmp
->parent
= parent
->parent
;
6882 parent
->parent
->child
= tmp
;
6886 if (sd
&& sd_degenerate(sd
)) {
6892 sched_domain_debug(sd
, cpu
);
6894 rq_attach_root(rq
, rd
);
6895 rcu_assign_pointer(rq
->sd
, sd
);
6898 /* cpus with isolated domains */
6899 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
6901 /* Setup the mask of cpus configured for isolated domains */
6902 static int __init
isolated_cpu_setup(char *str
)
6904 static int __initdata ints
[NR_CPUS
];
6907 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
6908 cpus_clear(cpu_isolated_map
);
6909 for (i
= 1; i
<= ints
[0]; i
++)
6910 if (ints
[i
] < NR_CPUS
)
6911 cpu_set(ints
[i
], cpu_isolated_map
);
6915 __setup("isolcpus=", isolated_cpu_setup
);
6918 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6919 * to a function which identifies what group(along with sched group) a CPU
6920 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6921 * (due to the fact that we keep track of groups covered with a cpumask_t).
6923 * init_sched_build_groups will build a circular linked list of the groups
6924 * covered by the given span, and will set each group's ->cpumask correctly,
6925 * and ->cpu_power to 0.
6928 init_sched_build_groups(const cpumask_t
*span
, const cpumask_t
*cpu_map
,
6929 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
6930 struct sched_group
**sg
,
6931 cpumask_t
*tmpmask
),
6932 cpumask_t
*covered
, cpumask_t
*tmpmask
)
6934 struct sched_group
*first
= NULL
, *last
= NULL
;
6937 cpus_clear(*covered
);
6939 for_each_cpu_mask_nr(i
, *span
) {
6940 struct sched_group
*sg
;
6941 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6944 if (cpu_isset(i
, *covered
))
6947 cpus_clear(sg
->cpumask
);
6948 sg
->__cpu_power
= 0;
6950 for_each_cpu_mask_nr(j
, *span
) {
6951 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6954 cpu_set(j
, *covered
);
6955 cpu_set(j
, sg
->cpumask
);
6966 #define SD_NODES_PER_DOMAIN 16
6971 * find_next_best_node - find the next node to include in a sched_domain
6972 * @node: node whose sched_domain we're building
6973 * @used_nodes: nodes already in the sched_domain
6975 * Find the next node to include in a given scheduling domain. Simply
6976 * finds the closest node not already in the @used_nodes map.
6978 * Should use nodemask_t.
6980 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
6982 int i
, n
, val
, min_val
, best_node
= 0;
6986 for (i
= 0; i
< nr_node_ids
; i
++) {
6987 /* Start at @node */
6988 n
= (node
+ i
) % nr_node_ids
;
6990 if (!nr_cpus_node(n
))
6993 /* Skip already used nodes */
6994 if (node_isset(n
, *used_nodes
))
6997 /* Simple min distance search */
6998 val
= node_distance(node
, n
);
7000 if (val
< min_val
) {
7006 node_set(best_node
, *used_nodes
);
7011 * sched_domain_node_span - get a cpumask for a node's sched_domain
7012 * @node: node whose cpumask we're constructing
7013 * @span: resulting cpumask
7015 * Given a node, construct a good cpumask for its sched_domain to span. It
7016 * should be one that prevents unnecessary balancing, but also spreads tasks
7019 static void sched_domain_node_span(int node
, cpumask_t
*span
)
7021 nodemask_t used_nodes
;
7022 node_to_cpumask_ptr(nodemask
, node
);
7026 nodes_clear(used_nodes
);
7028 cpus_or(*span
, *span
, *nodemask
);
7029 node_set(node
, used_nodes
);
7031 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
7032 int next_node
= find_next_best_node(node
, &used_nodes
);
7034 node_to_cpumask_ptr_next(nodemask
, next_node
);
7035 cpus_or(*span
, *span
, *nodemask
);
7038 #endif /* CONFIG_NUMA */
7040 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
7043 * SMT sched-domains:
7045 #ifdef CONFIG_SCHED_SMT
7046 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
7047 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
7050 cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7054 *sg
= &per_cpu(sched_group_cpus
, cpu
);
7057 #endif /* CONFIG_SCHED_SMT */
7060 * multi-core sched-domains:
7062 #ifdef CONFIG_SCHED_MC
7063 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
7064 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
7065 #endif /* CONFIG_SCHED_MC */
7067 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7069 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7074 *mask
= per_cpu(cpu_sibling_map
, cpu
);
7075 cpus_and(*mask
, *mask
, *cpu_map
);
7076 group
= first_cpu(*mask
);
7078 *sg
= &per_cpu(sched_group_core
, group
);
7081 #elif defined(CONFIG_SCHED_MC)
7083 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7087 *sg
= &per_cpu(sched_group_core
, cpu
);
7092 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
7093 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
7096 cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7100 #ifdef CONFIG_SCHED_MC
7101 *mask
= cpu_coregroup_map(cpu
);
7102 cpus_and(*mask
, *mask
, *cpu_map
);
7103 group
= first_cpu(*mask
);
7104 #elif defined(CONFIG_SCHED_SMT)
7105 *mask
= per_cpu(cpu_sibling_map
, cpu
);
7106 cpus_and(*mask
, *mask
, *cpu_map
);
7107 group
= first_cpu(*mask
);
7112 *sg
= &per_cpu(sched_group_phys
, group
);
7118 * The init_sched_build_groups can't handle what we want to do with node
7119 * groups, so roll our own. Now each node has its own list of groups which
7120 * gets dynamically allocated.
7122 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
7123 static struct sched_group
***sched_group_nodes_bycpu
;
7125 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
7126 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
7128 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
7129 struct sched_group
**sg
, cpumask_t
*nodemask
)
7133 *nodemask
= node_to_cpumask(cpu_to_node(cpu
));
7134 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7135 group
= first_cpu(*nodemask
);
7138 *sg
= &per_cpu(sched_group_allnodes
, group
);
7142 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
7144 struct sched_group
*sg
= group_head
;
7150 for_each_cpu_mask_nr(j
, sg
->cpumask
) {
7151 struct sched_domain
*sd
;
7153 sd
= &per_cpu(phys_domains
, j
);
7154 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
7156 * Only add "power" once for each
7162 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
7165 } while (sg
!= group_head
);
7167 #endif /* CONFIG_NUMA */
7170 /* Free memory allocated for various sched_group structures */
7171 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
7175 for_each_cpu_mask_nr(cpu
, *cpu_map
) {
7176 struct sched_group
**sched_group_nodes
7177 = sched_group_nodes_bycpu
[cpu
];
7179 if (!sched_group_nodes
)
7182 for (i
= 0; i
< nr_node_ids
; i
++) {
7183 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
7185 *nodemask
= node_to_cpumask(i
);
7186 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7187 if (cpus_empty(*nodemask
))
7197 if (oldsg
!= sched_group_nodes
[i
])
7200 kfree(sched_group_nodes
);
7201 sched_group_nodes_bycpu
[cpu
] = NULL
;
7204 #else /* !CONFIG_NUMA */
7205 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
7208 #endif /* CONFIG_NUMA */
7211 * Initialize sched groups cpu_power.
7213 * cpu_power indicates the capacity of sched group, which is used while
7214 * distributing the load between different sched groups in a sched domain.
7215 * Typically cpu_power for all the groups in a sched domain will be same unless
7216 * there are asymmetries in the topology. If there are asymmetries, group
7217 * having more cpu_power will pickup more load compared to the group having
7220 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7221 * the maximum number of tasks a group can handle in the presence of other idle
7222 * or lightly loaded groups in the same sched domain.
7224 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
7226 struct sched_domain
*child
;
7227 struct sched_group
*group
;
7229 WARN_ON(!sd
|| !sd
->groups
);
7231 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
7236 sd
->groups
->__cpu_power
= 0;
7239 * For perf policy, if the groups in child domain share resources
7240 * (for example cores sharing some portions of the cache hierarchy
7241 * or SMT), then set this domain groups cpu_power such that each group
7242 * can handle only one task, when there are other idle groups in the
7243 * same sched domain.
7245 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
7247 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
7248 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
7253 * add cpu_power of each child group to this groups cpu_power
7255 group
= child
->groups
;
7257 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
7258 group
= group
->next
;
7259 } while (group
!= child
->groups
);
7263 * Initializers for schedule domains
7264 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7267 #ifdef CONFIG_SCHED_DEBUG
7268 # define SD_INIT_NAME(sd, type) sd->name = #type
7270 # define SD_INIT_NAME(sd, type) do { } while (0)
7273 #define SD_INIT(sd, type) sd_init_##type(sd)
7275 #define SD_INIT_FUNC(type) \
7276 static noinline void sd_init_##type(struct sched_domain *sd) \
7278 memset(sd, 0, sizeof(*sd)); \
7279 *sd = SD_##type##_INIT; \
7280 sd->level = SD_LV_##type; \
7281 SD_INIT_NAME(sd, type); \
7286 SD_INIT_FUNC(ALLNODES
)
7289 #ifdef CONFIG_SCHED_SMT
7290 SD_INIT_FUNC(SIBLING
)
7292 #ifdef CONFIG_SCHED_MC
7297 * To minimize stack usage kmalloc room for cpumasks and share the
7298 * space as the usage in build_sched_domains() dictates. Used only
7299 * if the amount of space is significant.
7302 cpumask_t tmpmask
; /* make this one first */
7305 cpumask_t this_sibling_map
;
7306 cpumask_t this_core_map
;
7308 cpumask_t send_covered
;
7311 cpumask_t domainspan
;
7313 cpumask_t notcovered
;
7318 #define SCHED_CPUMASK_ALLOC 1
7319 #define SCHED_CPUMASK_FREE(v) kfree(v)
7320 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7322 #define SCHED_CPUMASK_ALLOC 0
7323 #define SCHED_CPUMASK_FREE(v)
7324 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7327 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7328 ((unsigned long)(a) + offsetof(struct allmasks, v))
7330 static int default_relax_domain_level
= -1;
7332 static int __init
setup_relax_domain_level(char *str
)
7336 val
= simple_strtoul(str
, NULL
, 0);
7337 if (val
< SD_LV_MAX
)
7338 default_relax_domain_level
= val
;
7342 __setup("relax_domain_level=", setup_relax_domain_level
);
7344 static void set_domain_attribute(struct sched_domain
*sd
,
7345 struct sched_domain_attr
*attr
)
7349 if (!attr
|| attr
->relax_domain_level
< 0) {
7350 if (default_relax_domain_level
< 0)
7353 request
= default_relax_domain_level
;
7355 request
= attr
->relax_domain_level
;
7356 if (request
< sd
->level
) {
7357 /* turn off idle balance on this domain */
7358 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
7360 /* turn on idle balance on this domain */
7361 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
7366 * Build sched domains for a given set of cpus and attach the sched domains
7367 * to the individual cpus
7369 static int __build_sched_domains(const cpumask_t
*cpu_map
,
7370 struct sched_domain_attr
*attr
)
7373 struct root_domain
*rd
;
7374 SCHED_CPUMASK_DECLARE(allmasks
);
7377 struct sched_group
**sched_group_nodes
= NULL
;
7378 int sd_allnodes
= 0;
7381 * Allocate the per-node list of sched groups
7383 sched_group_nodes
= kcalloc(nr_node_ids
, sizeof(struct sched_group
*),
7385 if (!sched_group_nodes
) {
7386 printk(KERN_WARNING
"Can not alloc sched group node list\n");
7391 rd
= alloc_rootdomain();
7393 printk(KERN_WARNING
"Cannot alloc root domain\n");
7395 kfree(sched_group_nodes
);
7400 #if SCHED_CPUMASK_ALLOC
7401 /* get space for all scratch cpumask variables */
7402 allmasks
= kmalloc(sizeof(*allmasks
), GFP_KERNEL
);
7404 printk(KERN_WARNING
"Cannot alloc cpumask array\n");
7407 kfree(sched_group_nodes
);
7412 tmpmask
= (cpumask_t
*)allmasks
;
7416 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
7420 * Set up domains for cpus specified by the cpu_map.
7422 for_each_cpu_mask_nr(i
, *cpu_map
) {
7423 struct sched_domain
*sd
= NULL
, *p
;
7424 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7426 *nodemask
= node_to_cpumask(cpu_to_node(i
));
7427 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7430 if (cpus_weight(*cpu_map
) >
7431 SD_NODES_PER_DOMAIN
*cpus_weight(*nodemask
)) {
7432 sd
= &per_cpu(allnodes_domains
, i
);
7433 SD_INIT(sd
, ALLNODES
);
7434 set_domain_attribute(sd
, attr
);
7435 sd
->span
= *cpu_map
;
7436 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7442 sd
= &per_cpu(node_domains
, i
);
7444 set_domain_attribute(sd
, attr
);
7445 sched_domain_node_span(cpu_to_node(i
), &sd
->span
);
7449 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7453 sd
= &per_cpu(phys_domains
, i
);
7455 set_domain_attribute(sd
, attr
);
7456 sd
->span
= *nodemask
;
7460 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7462 #ifdef CONFIG_SCHED_MC
7464 sd
= &per_cpu(core_domains
, i
);
7466 set_domain_attribute(sd
, attr
);
7467 sd
->span
= cpu_coregroup_map(i
);
7468 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7471 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7474 #ifdef CONFIG_SCHED_SMT
7476 sd
= &per_cpu(cpu_domains
, i
);
7477 SD_INIT(sd
, SIBLING
);
7478 set_domain_attribute(sd
, attr
);
7479 sd
->span
= per_cpu(cpu_sibling_map
, i
);
7480 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7483 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7487 #ifdef CONFIG_SCHED_SMT
7488 /* Set up CPU (sibling) groups */
7489 for_each_cpu_mask_nr(i
, *cpu_map
) {
7490 SCHED_CPUMASK_VAR(this_sibling_map
, allmasks
);
7491 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7493 *this_sibling_map
= per_cpu(cpu_sibling_map
, i
);
7494 cpus_and(*this_sibling_map
, *this_sibling_map
, *cpu_map
);
7495 if (i
!= first_cpu(*this_sibling_map
))
7498 init_sched_build_groups(this_sibling_map
, cpu_map
,
7500 send_covered
, tmpmask
);
7504 #ifdef CONFIG_SCHED_MC
7505 /* Set up multi-core groups */
7506 for_each_cpu_mask_nr(i
, *cpu_map
) {
7507 SCHED_CPUMASK_VAR(this_core_map
, allmasks
);
7508 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7510 *this_core_map
= cpu_coregroup_map(i
);
7511 cpus_and(*this_core_map
, *this_core_map
, *cpu_map
);
7512 if (i
!= first_cpu(*this_core_map
))
7515 init_sched_build_groups(this_core_map
, cpu_map
,
7517 send_covered
, tmpmask
);
7521 /* Set up physical groups */
7522 for (i
= 0; i
< nr_node_ids
; i
++) {
7523 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7524 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7526 *nodemask
= node_to_cpumask(i
);
7527 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7528 if (cpus_empty(*nodemask
))
7531 init_sched_build_groups(nodemask
, cpu_map
,
7533 send_covered
, tmpmask
);
7537 /* Set up node groups */
7539 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7541 init_sched_build_groups(cpu_map
, cpu_map
,
7542 &cpu_to_allnodes_group
,
7543 send_covered
, tmpmask
);
7546 for (i
= 0; i
< nr_node_ids
; i
++) {
7547 /* Set up node groups */
7548 struct sched_group
*sg
, *prev
;
7549 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7550 SCHED_CPUMASK_VAR(domainspan
, allmasks
);
7551 SCHED_CPUMASK_VAR(covered
, allmasks
);
7554 *nodemask
= node_to_cpumask(i
);
7555 cpus_clear(*covered
);
7557 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7558 if (cpus_empty(*nodemask
)) {
7559 sched_group_nodes
[i
] = NULL
;
7563 sched_domain_node_span(i
, domainspan
);
7564 cpus_and(*domainspan
, *domainspan
, *cpu_map
);
7566 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
7568 printk(KERN_WARNING
"Can not alloc domain group for "
7572 sched_group_nodes
[i
] = sg
;
7573 for_each_cpu_mask_nr(j
, *nodemask
) {
7574 struct sched_domain
*sd
;
7576 sd
= &per_cpu(node_domains
, j
);
7579 sg
->__cpu_power
= 0;
7580 sg
->cpumask
= *nodemask
;
7582 cpus_or(*covered
, *covered
, *nodemask
);
7585 for (j
= 0; j
< nr_node_ids
; j
++) {
7586 SCHED_CPUMASK_VAR(notcovered
, allmasks
);
7587 int n
= (i
+ j
) % nr_node_ids
;
7588 node_to_cpumask_ptr(pnodemask
, n
);
7590 cpus_complement(*notcovered
, *covered
);
7591 cpus_and(*tmpmask
, *notcovered
, *cpu_map
);
7592 cpus_and(*tmpmask
, *tmpmask
, *domainspan
);
7593 if (cpus_empty(*tmpmask
))
7596 cpus_and(*tmpmask
, *tmpmask
, *pnodemask
);
7597 if (cpus_empty(*tmpmask
))
7600 sg
= kmalloc_node(sizeof(struct sched_group
),
7604 "Can not alloc domain group for node %d\n", j
);
7607 sg
->__cpu_power
= 0;
7608 sg
->cpumask
= *tmpmask
;
7609 sg
->next
= prev
->next
;
7610 cpus_or(*covered
, *covered
, *tmpmask
);
7617 /* Calculate CPU power for physical packages and nodes */
7618 #ifdef CONFIG_SCHED_SMT
7619 for_each_cpu_mask_nr(i
, *cpu_map
) {
7620 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
7622 init_sched_groups_power(i
, sd
);
7625 #ifdef CONFIG_SCHED_MC
7626 for_each_cpu_mask_nr(i
, *cpu_map
) {
7627 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
7629 init_sched_groups_power(i
, sd
);
7633 for_each_cpu_mask_nr(i
, *cpu_map
) {
7634 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
7636 init_sched_groups_power(i
, sd
);
7640 for (i
= 0; i
< nr_node_ids
; i
++)
7641 init_numa_sched_groups_power(sched_group_nodes
[i
]);
7644 struct sched_group
*sg
;
7646 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
,
7648 init_numa_sched_groups_power(sg
);
7652 /* Attach the domains */
7653 for_each_cpu_mask_nr(i
, *cpu_map
) {
7654 struct sched_domain
*sd
;
7655 #ifdef CONFIG_SCHED_SMT
7656 sd
= &per_cpu(cpu_domains
, i
);
7657 #elif defined(CONFIG_SCHED_MC)
7658 sd
= &per_cpu(core_domains
, i
);
7660 sd
= &per_cpu(phys_domains
, i
);
7662 cpu_attach_domain(sd
, rd
, i
);
7665 SCHED_CPUMASK_FREE((void *)allmasks
);
7670 free_sched_groups(cpu_map
, tmpmask
);
7671 SCHED_CPUMASK_FREE((void *)allmasks
);
7676 static int build_sched_domains(const cpumask_t
*cpu_map
)
7678 return __build_sched_domains(cpu_map
, NULL
);
7681 static cpumask_t
*doms_cur
; /* current sched domains */
7682 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7683 static struct sched_domain_attr
*dattr_cur
;
7684 /* attribues of custom domains in 'doms_cur' */
7687 * Special case: If a kmalloc of a doms_cur partition (array of
7688 * cpumask_t) fails, then fallback to a single sched domain,
7689 * as determined by the single cpumask_t fallback_doms.
7691 static cpumask_t fallback_doms
;
7693 void __attribute__((weak
)) arch_update_cpu_topology(void)
7698 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7699 * For now this just excludes isolated cpus, but could be used to
7700 * exclude other special cases in the future.
7702 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
7706 arch_update_cpu_topology();
7708 doms_cur
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
7710 doms_cur
= &fallback_doms
;
7711 cpus_andnot(*doms_cur
, *cpu_map
, cpu_isolated_map
);
7713 err
= build_sched_domains(doms_cur
);
7714 register_sched_domain_sysctl();
7719 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
,
7722 free_sched_groups(cpu_map
, tmpmask
);
7726 * Detach sched domains from a group of cpus specified in cpu_map
7727 * These cpus will now be attached to the NULL domain
7729 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
7734 unregister_sched_domain_sysctl();
7736 for_each_cpu_mask_nr(i
, *cpu_map
)
7737 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7738 synchronize_sched();
7739 arch_destroy_sched_domains(cpu_map
, &tmpmask
);
7742 /* handle null as "default" */
7743 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7744 struct sched_domain_attr
*new, int idx_new
)
7746 struct sched_domain_attr tmp
;
7753 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7754 new ? (new + idx_new
) : &tmp
,
7755 sizeof(struct sched_domain_attr
));
7759 * Partition sched domains as specified by the 'ndoms_new'
7760 * cpumasks in the array doms_new[] of cpumasks. This compares
7761 * doms_new[] to the current sched domain partitioning, doms_cur[].
7762 * It destroys each deleted domain and builds each new domain.
7764 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7765 * The masks don't intersect (don't overlap.) We should setup one
7766 * sched domain for each mask. CPUs not in any of the cpumasks will
7767 * not be load balanced. If the same cpumask appears both in the
7768 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7771 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7772 * ownership of it and will kfree it when done with it. If the caller
7773 * failed the kmalloc call, then it can pass in doms_new == NULL,
7774 * and partition_sched_domains() will fallback to the single partition
7775 * 'fallback_doms', it also forces the domains to be rebuilt.
7777 * If doms_new==NULL it will be replaced with cpu_online_map.
7778 * ndoms_new==0 is a special case for destroying existing domains.
7779 * It will not create the default domain.
7781 * Call with hotplug lock held
7783 void partition_sched_domains(int ndoms_new
, cpumask_t
*doms_new
,
7784 struct sched_domain_attr
*dattr_new
)
7788 mutex_lock(&sched_domains_mutex
);
7790 /* always unregister in case we don't destroy any domains */
7791 unregister_sched_domain_sysctl();
7793 n
= doms_new
? ndoms_new
: 0;
7795 /* Destroy deleted domains */
7796 for (i
= 0; i
< ndoms_cur
; i
++) {
7797 for (j
= 0; j
< n
; j
++) {
7798 if (cpus_equal(doms_cur
[i
], doms_new
[j
])
7799 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7802 /* no match - a current sched domain not in new doms_new[] */
7803 detach_destroy_domains(doms_cur
+ i
);
7808 if (doms_new
== NULL
) {
7810 doms_new
= &fallback_doms
;
7811 cpus_andnot(doms_new
[0], cpu_online_map
, cpu_isolated_map
);
7815 /* Build new domains */
7816 for (i
= 0; i
< ndoms_new
; i
++) {
7817 for (j
= 0; j
< ndoms_cur
; j
++) {
7818 if (cpus_equal(doms_new
[i
], doms_cur
[j
])
7819 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7822 /* no match - add a new doms_new */
7823 __build_sched_domains(doms_new
+ i
,
7824 dattr_new
? dattr_new
+ i
: NULL
);
7829 /* Remember the new sched domains */
7830 if (doms_cur
!= &fallback_doms
)
7832 kfree(dattr_cur
); /* kfree(NULL) is safe */
7833 doms_cur
= doms_new
;
7834 dattr_cur
= dattr_new
;
7835 ndoms_cur
= ndoms_new
;
7837 register_sched_domain_sysctl();
7839 mutex_unlock(&sched_domains_mutex
);
7842 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7843 int arch_reinit_sched_domains(void)
7847 /* Destroy domains first to force the rebuild */
7848 partition_sched_domains(0, NULL
, NULL
);
7850 rebuild_sched_domains();
7856 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7860 if (buf
[0] != '0' && buf
[0] != '1')
7864 sched_smt_power_savings
= (buf
[0] == '1');
7866 sched_mc_power_savings
= (buf
[0] == '1');
7868 ret
= arch_reinit_sched_domains();
7870 return ret
? ret
: count
;
7873 #ifdef CONFIG_SCHED_MC
7874 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
7877 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7879 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
7880 const char *buf
, size_t count
)
7882 return sched_power_savings_store(buf
, count
, 0);
7884 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
7885 sched_mc_power_savings_show
,
7886 sched_mc_power_savings_store
);
7889 #ifdef CONFIG_SCHED_SMT
7890 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
7893 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7895 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
7896 const char *buf
, size_t count
)
7898 return sched_power_savings_store(buf
, count
, 1);
7900 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
7901 sched_smt_power_savings_show
,
7902 sched_smt_power_savings_store
);
7905 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7909 #ifdef CONFIG_SCHED_SMT
7911 err
= sysfs_create_file(&cls
->kset
.kobj
,
7912 &attr_sched_smt_power_savings
.attr
);
7914 #ifdef CONFIG_SCHED_MC
7915 if (!err
&& mc_capable())
7916 err
= sysfs_create_file(&cls
->kset
.kobj
,
7917 &attr_sched_mc_power_savings
.attr
);
7921 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7923 #ifndef CONFIG_CPUSETS
7925 * Add online and remove offline CPUs from the scheduler domains.
7926 * When cpusets are enabled they take over this function.
7928 static int update_sched_domains(struct notifier_block
*nfb
,
7929 unsigned long action
, void *hcpu
)
7933 case CPU_ONLINE_FROZEN
:
7935 case CPU_DEAD_FROZEN
:
7936 partition_sched_domains(1, NULL
, NULL
);
7945 static int update_runtime(struct notifier_block
*nfb
,
7946 unsigned long action
, void *hcpu
)
7948 int cpu
= (int)(long)hcpu
;
7951 case CPU_DOWN_PREPARE
:
7952 case CPU_DOWN_PREPARE_FROZEN
:
7953 disable_runtime(cpu_rq(cpu
));
7956 case CPU_DOWN_FAILED
:
7957 case CPU_DOWN_FAILED_FROZEN
:
7959 case CPU_ONLINE_FROZEN
:
7960 enable_runtime(cpu_rq(cpu
));
7968 void __init
sched_init_smp(void)
7970 cpumask_t non_isolated_cpus
;
7972 #if defined(CONFIG_NUMA)
7973 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
7975 BUG_ON(sched_group_nodes_bycpu
== NULL
);
7978 mutex_lock(&sched_domains_mutex
);
7979 arch_init_sched_domains(&cpu_online_map
);
7980 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
7981 if (cpus_empty(non_isolated_cpus
))
7982 cpu_set(smp_processor_id(), non_isolated_cpus
);
7983 mutex_unlock(&sched_domains_mutex
);
7986 #ifndef CONFIG_CPUSETS
7987 /* XXX: Theoretical race here - CPU may be hotplugged now */
7988 hotcpu_notifier(update_sched_domains
, 0);
7991 /* RT runtime code needs to handle some hotplug events */
7992 hotcpu_notifier(update_runtime
, 0);
7996 /* Move init over to a non-isolated CPU */
7997 if (set_cpus_allowed_ptr(current
, &non_isolated_cpus
) < 0)
7999 sched_init_granularity();
8002 void __init
sched_init_smp(void)
8004 sched_init_granularity();
8006 #endif /* CONFIG_SMP */
8008 int in_sched_functions(unsigned long addr
)
8010 return in_lock_functions(addr
) ||
8011 (addr
>= (unsigned long)__sched_text_start
8012 && addr
< (unsigned long)__sched_text_end
);
8015 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
8017 cfs_rq
->tasks_timeline
= RB_ROOT
;
8018 INIT_LIST_HEAD(&cfs_rq
->tasks
);
8019 #ifdef CONFIG_FAIR_GROUP_SCHED
8022 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
8025 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
8027 struct rt_prio_array
*array
;
8030 array
= &rt_rq
->active
;
8031 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
8032 INIT_LIST_HEAD(array
->queue
+ i
);
8033 __clear_bit(i
, array
->bitmap
);
8035 /* delimiter for bitsearch: */
8036 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
8038 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8039 rt_rq
->highest_prio
= MAX_RT_PRIO
;
8042 rt_rq
->rt_nr_migratory
= 0;
8043 rt_rq
->overloaded
= 0;
8047 rt_rq
->rt_throttled
= 0;
8048 rt_rq
->rt_runtime
= 0;
8049 spin_lock_init(&rt_rq
->rt_runtime_lock
);
8051 #ifdef CONFIG_RT_GROUP_SCHED
8052 rt_rq
->rt_nr_boosted
= 0;
8057 #ifdef CONFIG_FAIR_GROUP_SCHED
8058 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
8059 struct sched_entity
*se
, int cpu
, int add
,
8060 struct sched_entity
*parent
)
8062 struct rq
*rq
= cpu_rq(cpu
);
8063 tg
->cfs_rq
[cpu
] = cfs_rq
;
8064 init_cfs_rq(cfs_rq
, rq
);
8067 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
8070 /* se could be NULL for init_task_group */
8075 se
->cfs_rq
= &rq
->cfs
;
8077 se
->cfs_rq
= parent
->my_q
;
8080 se
->load
.weight
= tg
->shares
;
8081 se
->load
.inv_weight
= 0;
8082 se
->parent
= parent
;
8086 #ifdef CONFIG_RT_GROUP_SCHED
8087 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
8088 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
8089 struct sched_rt_entity
*parent
)
8091 struct rq
*rq
= cpu_rq(cpu
);
8093 tg
->rt_rq
[cpu
] = rt_rq
;
8094 init_rt_rq(rt_rq
, rq
);
8096 rt_rq
->rt_se
= rt_se
;
8097 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8099 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
8101 tg
->rt_se
[cpu
] = rt_se
;
8106 rt_se
->rt_rq
= &rq
->rt
;
8108 rt_se
->rt_rq
= parent
->my_q
;
8110 rt_se
->my_q
= rt_rq
;
8111 rt_se
->parent
= parent
;
8112 INIT_LIST_HEAD(&rt_se
->run_list
);
8116 void __init
sched_init(void)
8119 unsigned long alloc_size
= 0, ptr
;
8121 #ifdef CONFIG_FAIR_GROUP_SCHED
8122 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8124 #ifdef CONFIG_RT_GROUP_SCHED
8125 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8127 #ifdef CONFIG_USER_SCHED
8131 * As sched_init() is called before page_alloc is setup,
8132 * we use alloc_bootmem().
8135 ptr
= (unsigned long)alloc_bootmem(alloc_size
);
8137 #ifdef CONFIG_FAIR_GROUP_SCHED
8138 init_task_group
.se
= (struct sched_entity
**)ptr
;
8139 ptr
+= nr_cpu_ids
* sizeof(void **);
8141 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8142 ptr
+= nr_cpu_ids
* sizeof(void **);
8144 #ifdef CONFIG_USER_SCHED
8145 root_task_group
.se
= (struct sched_entity
**)ptr
;
8146 ptr
+= nr_cpu_ids
* sizeof(void **);
8148 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8149 ptr
+= nr_cpu_ids
* sizeof(void **);
8150 #endif /* CONFIG_USER_SCHED */
8151 #endif /* CONFIG_FAIR_GROUP_SCHED */
8152 #ifdef CONFIG_RT_GROUP_SCHED
8153 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8154 ptr
+= nr_cpu_ids
* sizeof(void **);
8156 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8157 ptr
+= nr_cpu_ids
* sizeof(void **);
8159 #ifdef CONFIG_USER_SCHED
8160 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8161 ptr
+= nr_cpu_ids
* sizeof(void **);
8163 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8164 ptr
+= nr_cpu_ids
* sizeof(void **);
8165 #endif /* CONFIG_USER_SCHED */
8166 #endif /* CONFIG_RT_GROUP_SCHED */
8170 init_defrootdomain();
8173 init_rt_bandwidth(&def_rt_bandwidth
,
8174 global_rt_period(), global_rt_runtime());
8176 #ifdef CONFIG_RT_GROUP_SCHED
8177 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
8178 global_rt_period(), global_rt_runtime());
8179 #ifdef CONFIG_USER_SCHED
8180 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
8181 global_rt_period(), RUNTIME_INF
);
8182 #endif /* CONFIG_USER_SCHED */
8183 #endif /* CONFIG_RT_GROUP_SCHED */
8185 #ifdef CONFIG_GROUP_SCHED
8186 list_add(&init_task_group
.list
, &task_groups
);
8187 INIT_LIST_HEAD(&init_task_group
.children
);
8189 #ifdef CONFIG_USER_SCHED
8190 INIT_LIST_HEAD(&root_task_group
.children
);
8191 init_task_group
.parent
= &root_task_group
;
8192 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
8193 #endif /* CONFIG_USER_SCHED */
8194 #endif /* CONFIG_GROUP_SCHED */
8196 for_each_possible_cpu(i
) {
8200 spin_lock_init(&rq
->lock
);
8202 init_cfs_rq(&rq
->cfs
, rq
);
8203 init_rt_rq(&rq
->rt
, rq
);
8204 #ifdef CONFIG_FAIR_GROUP_SCHED
8205 init_task_group
.shares
= init_task_group_load
;
8206 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
8207 #ifdef CONFIG_CGROUP_SCHED
8209 * How much cpu bandwidth does init_task_group get?
8211 * In case of task-groups formed thr' the cgroup filesystem, it
8212 * gets 100% of the cpu resources in the system. This overall
8213 * system cpu resource is divided among the tasks of
8214 * init_task_group and its child task-groups in a fair manner,
8215 * based on each entity's (task or task-group's) weight
8216 * (se->load.weight).
8218 * In other words, if init_task_group has 10 tasks of weight
8219 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8220 * then A0's share of the cpu resource is:
8222 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8224 * We achieve this by letting init_task_group's tasks sit
8225 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8227 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
8228 #elif defined CONFIG_USER_SCHED
8229 root_task_group
.shares
= NICE_0_LOAD
;
8230 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
8232 * In case of task-groups formed thr' the user id of tasks,
8233 * init_task_group represents tasks belonging to root user.
8234 * Hence it forms a sibling of all subsequent groups formed.
8235 * In this case, init_task_group gets only a fraction of overall
8236 * system cpu resource, based on the weight assigned to root
8237 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8238 * by letting tasks of init_task_group sit in a separate cfs_rq
8239 * (init_cfs_rq) and having one entity represent this group of
8240 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8242 init_tg_cfs_entry(&init_task_group
,
8243 &per_cpu(init_cfs_rq
, i
),
8244 &per_cpu(init_sched_entity
, i
), i
, 1,
8245 root_task_group
.se
[i
]);
8248 #endif /* CONFIG_FAIR_GROUP_SCHED */
8250 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
8251 #ifdef CONFIG_RT_GROUP_SCHED
8252 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
8253 #ifdef CONFIG_CGROUP_SCHED
8254 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
8255 #elif defined CONFIG_USER_SCHED
8256 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
8257 init_tg_rt_entry(&init_task_group
,
8258 &per_cpu(init_rt_rq
, i
),
8259 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
8260 root_task_group
.rt_se
[i
]);
8264 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
8265 rq
->cpu_load
[j
] = 0;
8269 rq
->active_balance
= 0;
8270 rq
->next_balance
= jiffies
;
8274 rq
->migration_thread
= NULL
;
8275 INIT_LIST_HEAD(&rq
->migration_queue
);
8276 rq_attach_root(rq
, &def_root_domain
);
8279 atomic_set(&rq
->nr_iowait
, 0);
8282 set_load_weight(&init_task
);
8284 #ifdef CONFIG_PREEMPT_NOTIFIERS
8285 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
8289 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
8292 #ifdef CONFIG_RT_MUTEXES
8293 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
8297 * The boot idle thread does lazy MMU switching as well:
8299 atomic_inc(&init_mm
.mm_count
);
8300 enter_lazy_tlb(&init_mm
, current
);
8303 * Make us the idle thread. Technically, schedule() should not be
8304 * called from this thread, however somewhere below it might be,
8305 * but because we are the idle thread, we just pick up running again
8306 * when this runqueue becomes "idle".
8308 init_idle(current
, smp_processor_id());
8310 * During early bootup we pretend to be a normal task:
8312 current
->sched_class
= &fair_sched_class
;
8314 scheduler_running
= 1;
8317 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8318 void __might_sleep(char *file
, int line
)
8321 static unsigned long prev_jiffy
; /* ratelimiting */
8323 if ((!in_atomic() && !irqs_disabled()) ||
8324 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
8326 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
8328 prev_jiffy
= jiffies
;
8331 "BUG: sleeping function called from invalid context at %s:%d\n",
8334 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8335 in_atomic(), irqs_disabled(),
8336 current
->pid
, current
->comm
);
8338 debug_show_held_locks(current
);
8339 if (irqs_disabled())
8340 print_irqtrace_events(current
);
8344 EXPORT_SYMBOL(__might_sleep
);
8347 #ifdef CONFIG_MAGIC_SYSRQ
8348 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
8352 update_rq_clock(rq
);
8353 on_rq
= p
->se
.on_rq
;
8355 deactivate_task(rq
, p
, 0);
8356 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
8358 activate_task(rq
, p
, 0);
8359 resched_task(rq
->curr
);
8363 void normalize_rt_tasks(void)
8365 struct task_struct
*g
, *p
;
8366 unsigned long flags
;
8369 read_lock_irqsave(&tasklist_lock
, flags
);
8370 do_each_thread(g
, p
) {
8372 * Only normalize user tasks:
8377 p
->se
.exec_start
= 0;
8378 #ifdef CONFIG_SCHEDSTATS
8379 p
->se
.wait_start
= 0;
8380 p
->se
.sleep_start
= 0;
8381 p
->se
.block_start
= 0;
8386 * Renice negative nice level userspace
8389 if (TASK_NICE(p
) < 0 && p
->mm
)
8390 set_user_nice(p
, 0);
8394 spin_lock(&p
->pi_lock
);
8395 rq
= __task_rq_lock(p
);
8397 normalize_task(rq
, p
);
8399 __task_rq_unlock(rq
);
8400 spin_unlock(&p
->pi_lock
);
8401 } while_each_thread(g
, p
);
8403 read_unlock_irqrestore(&tasklist_lock
, flags
);
8406 #endif /* CONFIG_MAGIC_SYSRQ */
8410 * These functions are only useful for the IA64 MCA handling.
8412 * They can only be called when the whole system has been
8413 * stopped - every CPU needs to be quiescent, and no scheduling
8414 * activity can take place. Using them for anything else would
8415 * be a serious bug, and as a result, they aren't even visible
8416 * under any other configuration.
8420 * curr_task - return the current task for a given cpu.
8421 * @cpu: the processor in question.
8423 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8425 struct task_struct
*curr_task(int cpu
)
8427 return cpu_curr(cpu
);
8431 * set_curr_task - set the current task for a given cpu.
8432 * @cpu: the processor in question.
8433 * @p: the task pointer to set.
8435 * Description: This function must only be used when non-maskable interrupts
8436 * are serviced on a separate stack. It allows the architecture to switch the
8437 * notion of the current task on a cpu in a non-blocking manner. This function
8438 * must be called with all CPU's synchronized, and interrupts disabled, the
8439 * and caller must save the original value of the current task (see
8440 * curr_task() above) and restore that value before reenabling interrupts and
8441 * re-starting the system.
8443 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8445 void set_curr_task(int cpu
, struct task_struct
*p
)
8452 #ifdef CONFIG_FAIR_GROUP_SCHED
8453 static void free_fair_sched_group(struct task_group
*tg
)
8457 for_each_possible_cpu(i
) {
8459 kfree(tg
->cfs_rq
[i
]);
8469 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8471 struct cfs_rq
*cfs_rq
;
8472 struct sched_entity
*se
, *parent_se
;
8476 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8479 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8483 tg
->shares
= NICE_0_LOAD
;
8485 for_each_possible_cpu(i
) {
8488 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
),
8489 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8493 se
= kmalloc_node(sizeof(struct sched_entity
),
8494 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8498 parent_se
= parent
? parent
->se
[i
] : NULL
;
8499 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent_se
);
8508 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8510 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
8511 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
8514 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8516 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
8518 #else /* !CONFG_FAIR_GROUP_SCHED */
8519 static inline void free_fair_sched_group(struct task_group
*tg
)
8524 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8529 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8533 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8536 #endif /* CONFIG_FAIR_GROUP_SCHED */
8538 #ifdef CONFIG_RT_GROUP_SCHED
8539 static void free_rt_sched_group(struct task_group
*tg
)
8543 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8545 for_each_possible_cpu(i
) {
8547 kfree(tg
->rt_rq
[i
]);
8549 kfree(tg
->rt_se
[i
]);
8557 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8559 struct rt_rq
*rt_rq
;
8560 struct sched_rt_entity
*rt_se
, *parent_se
;
8564 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8567 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8571 init_rt_bandwidth(&tg
->rt_bandwidth
,
8572 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8574 for_each_possible_cpu(i
) {
8577 rt_rq
= kmalloc_node(sizeof(struct rt_rq
),
8578 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8582 rt_se
= kmalloc_node(sizeof(struct sched_rt_entity
),
8583 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8587 parent_se
= parent
? parent
->rt_se
[i
] : NULL
;
8588 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent_se
);
8597 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8599 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8600 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8603 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8605 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8607 #else /* !CONFIG_RT_GROUP_SCHED */
8608 static inline void free_rt_sched_group(struct task_group
*tg
)
8613 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8618 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8622 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8625 #endif /* CONFIG_RT_GROUP_SCHED */
8627 #ifdef CONFIG_GROUP_SCHED
8628 static void free_sched_group(struct task_group
*tg
)
8630 free_fair_sched_group(tg
);
8631 free_rt_sched_group(tg
);
8635 /* allocate runqueue etc for a new task group */
8636 struct task_group
*sched_create_group(struct task_group
*parent
)
8638 struct task_group
*tg
;
8639 unsigned long flags
;
8642 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8644 return ERR_PTR(-ENOMEM
);
8646 if (!alloc_fair_sched_group(tg
, parent
))
8649 if (!alloc_rt_sched_group(tg
, parent
))
8652 spin_lock_irqsave(&task_group_lock
, flags
);
8653 for_each_possible_cpu(i
) {
8654 register_fair_sched_group(tg
, i
);
8655 register_rt_sched_group(tg
, i
);
8657 list_add_rcu(&tg
->list
, &task_groups
);
8659 WARN_ON(!parent
); /* root should already exist */
8661 tg
->parent
= parent
;
8662 INIT_LIST_HEAD(&tg
->children
);
8663 list_add_rcu(&tg
->siblings
, &parent
->children
);
8664 spin_unlock_irqrestore(&task_group_lock
, flags
);
8669 free_sched_group(tg
);
8670 return ERR_PTR(-ENOMEM
);
8673 /* rcu callback to free various structures associated with a task group */
8674 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8676 /* now it should be safe to free those cfs_rqs */
8677 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8680 /* Destroy runqueue etc associated with a task group */
8681 void sched_destroy_group(struct task_group
*tg
)
8683 unsigned long flags
;
8686 spin_lock_irqsave(&task_group_lock
, flags
);
8687 for_each_possible_cpu(i
) {
8688 unregister_fair_sched_group(tg
, i
);
8689 unregister_rt_sched_group(tg
, i
);
8691 list_del_rcu(&tg
->list
);
8692 list_del_rcu(&tg
->siblings
);
8693 spin_unlock_irqrestore(&task_group_lock
, flags
);
8695 /* wait for possible concurrent references to cfs_rqs complete */
8696 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8699 /* change task's runqueue when it moves between groups.
8700 * The caller of this function should have put the task in its new group
8701 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8702 * reflect its new group.
8704 void sched_move_task(struct task_struct
*tsk
)
8707 unsigned long flags
;
8710 rq
= task_rq_lock(tsk
, &flags
);
8712 update_rq_clock(rq
);
8714 running
= task_current(rq
, tsk
);
8715 on_rq
= tsk
->se
.on_rq
;
8718 dequeue_task(rq
, tsk
, 0);
8719 if (unlikely(running
))
8720 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8722 set_task_rq(tsk
, task_cpu(tsk
));
8724 #ifdef CONFIG_FAIR_GROUP_SCHED
8725 if (tsk
->sched_class
->moved_group
)
8726 tsk
->sched_class
->moved_group(tsk
);
8729 if (unlikely(running
))
8730 tsk
->sched_class
->set_curr_task(rq
);
8732 enqueue_task(rq
, tsk
, 0);
8734 task_rq_unlock(rq
, &flags
);
8736 #endif /* CONFIG_GROUP_SCHED */
8738 #ifdef CONFIG_FAIR_GROUP_SCHED
8739 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8741 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8746 dequeue_entity(cfs_rq
, se
, 0);
8748 se
->load
.weight
= shares
;
8749 se
->load
.inv_weight
= 0;
8752 enqueue_entity(cfs_rq
, se
, 0);
8755 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8757 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8758 struct rq
*rq
= cfs_rq
->rq
;
8759 unsigned long flags
;
8761 spin_lock_irqsave(&rq
->lock
, flags
);
8762 __set_se_shares(se
, shares
);
8763 spin_unlock_irqrestore(&rq
->lock
, flags
);
8766 static DEFINE_MUTEX(shares_mutex
);
8768 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8771 unsigned long flags
;
8774 * We can't change the weight of the root cgroup.
8779 if (shares
< MIN_SHARES
)
8780 shares
= MIN_SHARES
;
8781 else if (shares
> MAX_SHARES
)
8782 shares
= MAX_SHARES
;
8784 mutex_lock(&shares_mutex
);
8785 if (tg
->shares
== shares
)
8788 spin_lock_irqsave(&task_group_lock
, flags
);
8789 for_each_possible_cpu(i
)
8790 unregister_fair_sched_group(tg
, i
);
8791 list_del_rcu(&tg
->siblings
);
8792 spin_unlock_irqrestore(&task_group_lock
, flags
);
8794 /* wait for any ongoing reference to this group to finish */
8795 synchronize_sched();
8798 * Now we are free to modify the group's share on each cpu
8799 * w/o tripping rebalance_share or load_balance_fair.
8801 tg
->shares
= shares
;
8802 for_each_possible_cpu(i
) {
8806 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
8807 set_se_shares(tg
->se
[i
], shares
);
8811 * Enable load balance activity on this group, by inserting it back on
8812 * each cpu's rq->leaf_cfs_rq_list.
8814 spin_lock_irqsave(&task_group_lock
, flags
);
8815 for_each_possible_cpu(i
)
8816 register_fair_sched_group(tg
, i
);
8817 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
8818 spin_unlock_irqrestore(&task_group_lock
, flags
);
8820 mutex_unlock(&shares_mutex
);
8824 unsigned long sched_group_shares(struct task_group
*tg
)
8830 #ifdef CONFIG_RT_GROUP_SCHED
8832 * Ensure that the real time constraints are schedulable.
8834 static DEFINE_MUTEX(rt_constraints_mutex
);
8836 static unsigned long to_ratio(u64 period
, u64 runtime
)
8838 if (runtime
== RUNTIME_INF
)
8841 return div64_u64(runtime
<< 20, period
);
8844 /* Must be called with tasklist_lock held */
8845 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8847 struct task_struct
*g
, *p
;
8849 do_each_thread(g
, p
) {
8850 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8852 } while_each_thread(g
, p
);
8857 struct rt_schedulable_data
{
8858 struct task_group
*tg
;
8863 static int tg_schedulable(struct task_group
*tg
, void *data
)
8865 struct rt_schedulable_data
*d
= data
;
8866 struct task_group
*child
;
8867 unsigned long total
, sum
= 0;
8868 u64 period
, runtime
;
8870 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8871 runtime
= tg
->rt_bandwidth
.rt_runtime
;
8874 period
= d
->rt_period
;
8875 runtime
= d
->rt_runtime
;
8879 * Cannot have more runtime than the period.
8881 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8885 * Ensure we don't starve existing RT tasks.
8887 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
8890 total
= to_ratio(period
, runtime
);
8893 * Nobody can have more than the global setting allows.
8895 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
8899 * The sum of our children's runtime should not exceed our own.
8901 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
8902 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
8903 runtime
= child
->rt_bandwidth
.rt_runtime
;
8905 if (child
== d
->tg
) {
8906 period
= d
->rt_period
;
8907 runtime
= d
->rt_runtime
;
8910 sum
+= to_ratio(period
, runtime
);
8919 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8921 struct rt_schedulable_data data
= {
8923 .rt_period
= period
,
8924 .rt_runtime
= runtime
,
8927 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
8930 static int tg_set_bandwidth(struct task_group
*tg
,
8931 u64 rt_period
, u64 rt_runtime
)
8935 mutex_lock(&rt_constraints_mutex
);
8936 read_lock(&tasklist_lock
);
8937 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
8941 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8942 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8943 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8945 for_each_possible_cpu(i
) {
8946 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8948 spin_lock(&rt_rq
->rt_runtime_lock
);
8949 rt_rq
->rt_runtime
= rt_runtime
;
8950 spin_unlock(&rt_rq
->rt_runtime_lock
);
8952 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8954 read_unlock(&tasklist_lock
);
8955 mutex_unlock(&rt_constraints_mutex
);
8960 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8962 u64 rt_runtime
, rt_period
;
8964 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8965 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8966 if (rt_runtime_us
< 0)
8967 rt_runtime
= RUNTIME_INF
;
8969 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8972 long sched_group_rt_runtime(struct task_group
*tg
)
8976 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8979 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8980 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8981 return rt_runtime_us
;
8984 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8986 u64 rt_runtime
, rt_period
;
8988 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8989 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8994 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8997 long sched_group_rt_period(struct task_group
*tg
)
9001 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9002 do_div(rt_period_us
, NSEC_PER_USEC
);
9003 return rt_period_us
;
9006 static int sched_rt_global_constraints(void)
9008 u64 runtime
, period
;
9011 if (sysctl_sched_rt_period
<= 0)
9014 runtime
= global_rt_runtime();
9015 period
= global_rt_period();
9018 * Sanity check on the sysctl variables.
9020 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
9023 mutex_lock(&rt_constraints_mutex
);
9024 read_lock(&tasklist_lock
);
9025 ret
= __rt_schedulable(NULL
, 0, 0);
9026 read_unlock(&tasklist_lock
);
9027 mutex_unlock(&rt_constraints_mutex
);
9031 #else /* !CONFIG_RT_GROUP_SCHED */
9032 static int sched_rt_global_constraints(void)
9034 unsigned long flags
;
9037 if (sysctl_sched_rt_period
<= 0)
9040 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9041 for_each_possible_cpu(i
) {
9042 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
9044 spin_lock(&rt_rq
->rt_runtime_lock
);
9045 rt_rq
->rt_runtime
= global_rt_runtime();
9046 spin_unlock(&rt_rq
->rt_runtime_lock
);
9048 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9052 #endif /* CONFIG_RT_GROUP_SCHED */
9054 int sched_rt_handler(struct ctl_table
*table
, int write
,
9055 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
9059 int old_period
, old_runtime
;
9060 static DEFINE_MUTEX(mutex
);
9063 old_period
= sysctl_sched_rt_period
;
9064 old_runtime
= sysctl_sched_rt_runtime
;
9066 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
9068 if (!ret
&& write
) {
9069 ret
= sched_rt_global_constraints();
9071 sysctl_sched_rt_period
= old_period
;
9072 sysctl_sched_rt_runtime
= old_runtime
;
9074 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
9075 def_rt_bandwidth
.rt_period
=
9076 ns_to_ktime(global_rt_period());
9079 mutex_unlock(&mutex
);
9084 #ifdef CONFIG_CGROUP_SCHED
9086 /* return corresponding task_group object of a cgroup */
9087 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
9089 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
9090 struct task_group
, css
);
9093 static struct cgroup_subsys_state
*
9094 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9096 struct task_group
*tg
, *parent
;
9098 if (!cgrp
->parent
) {
9099 /* This is early initialization for the top cgroup */
9100 return &init_task_group
.css
;
9103 parent
= cgroup_tg(cgrp
->parent
);
9104 tg
= sched_create_group(parent
);
9106 return ERR_PTR(-ENOMEM
);
9112 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9114 struct task_group
*tg
= cgroup_tg(cgrp
);
9116 sched_destroy_group(tg
);
9120 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9121 struct task_struct
*tsk
)
9123 #ifdef CONFIG_RT_GROUP_SCHED
9124 /* Don't accept realtime tasks when there is no way for them to run */
9125 if (rt_task(tsk
) && cgroup_tg(cgrp
)->rt_bandwidth
.rt_runtime
== 0)
9128 /* We don't support RT-tasks being in separate groups */
9129 if (tsk
->sched_class
!= &fair_sched_class
)
9137 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9138 struct cgroup
*old_cont
, struct task_struct
*tsk
)
9140 sched_move_task(tsk
);
9143 #ifdef CONFIG_FAIR_GROUP_SCHED
9144 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
9147 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
9150 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
9152 struct task_group
*tg
= cgroup_tg(cgrp
);
9154 return (u64
) tg
->shares
;
9156 #endif /* CONFIG_FAIR_GROUP_SCHED */
9158 #ifdef CONFIG_RT_GROUP_SCHED
9159 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
9162 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
9165 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9167 return sched_group_rt_runtime(cgroup_tg(cgrp
));
9170 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
9173 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
9176 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
9178 return sched_group_rt_period(cgroup_tg(cgrp
));
9180 #endif /* CONFIG_RT_GROUP_SCHED */
9182 static struct cftype cpu_files
[] = {
9183 #ifdef CONFIG_FAIR_GROUP_SCHED
9186 .read_u64
= cpu_shares_read_u64
,
9187 .write_u64
= cpu_shares_write_u64
,
9190 #ifdef CONFIG_RT_GROUP_SCHED
9192 .name
= "rt_runtime_us",
9193 .read_s64
= cpu_rt_runtime_read
,
9194 .write_s64
= cpu_rt_runtime_write
,
9197 .name
= "rt_period_us",
9198 .read_u64
= cpu_rt_period_read_uint
,
9199 .write_u64
= cpu_rt_period_write_uint
,
9204 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
9206 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
9209 struct cgroup_subsys cpu_cgroup_subsys
= {
9211 .create
= cpu_cgroup_create
,
9212 .destroy
= cpu_cgroup_destroy
,
9213 .can_attach
= cpu_cgroup_can_attach
,
9214 .attach
= cpu_cgroup_attach
,
9215 .populate
= cpu_cgroup_populate
,
9216 .subsys_id
= cpu_cgroup_subsys_id
,
9220 #endif /* CONFIG_CGROUP_SCHED */
9222 #ifdef CONFIG_CGROUP_CPUACCT
9225 * CPU accounting code for task groups.
9227 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9228 * (balbir@in.ibm.com).
9231 /* track cpu usage of a group of tasks */
9233 struct cgroup_subsys_state css
;
9234 /* cpuusage holds pointer to a u64-type object on every cpu */
9238 struct cgroup_subsys cpuacct_subsys
;
9240 /* return cpu accounting group corresponding to this container */
9241 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
9243 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
9244 struct cpuacct
, css
);
9247 /* return cpu accounting group to which this task belongs */
9248 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
9250 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
9251 struct cpuacct
, css
);
9254 /* create a new cpu accounting group */
9255 static struct cgroup_subsys_state
*cpuacct_create(
9256 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9258 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
9261 return ERR_PTR(-ENOMEM
);
9263 ca
->cpuusage
= alloc_percpu(u64
);
9264 if (!ca
->cpuusage
) {
9266 return ERR_PTR(-ENOMEM
);
9272 /* destroy an existing cpu accounting group */
9274 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9276 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9278 free_percpu(ca
->cpuusage
);
9282 /* return total cpu usage (in nanoseconds) of a group */
9283 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9285 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9286 u64 totalcpuusage
= 0;
9289 for_each_possible_cpu(i
) {
9290 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
9293 * Take rq->lock to make 64-bit addition safe on 32-bit
9296 spin_lock_irq(&cpu_rq(i
)->lock
);
9297 totalcpuusage
+= *cpuusage
;
9298 spin_unlock_irq(&cpu_rq(i
)->lock
);
9301 return totalcpuusage
;
9304 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
9307 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9316 for_each_possible_cpu(i
) {
9317 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
9319 spin_lock_irq(&cpu_rq(i
)->lock
);
9321 spin_unlock_irq(&cpu_rq(i
)->lock
);
9327 static struct cftype files
[] = {
9330 .read_u64
= cpuusage_read
,
9331 .write_u64
= cpuusage_write
,
9335 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9337 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
9341 * charge this task's execution time to its accounting group.
9343 * called with rq->lock held.
9345 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
9349 if (!cpuacct_subsys
.active
)
9354 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, task_cpu(tsk
));
9356 *cpuusage
+= cputime
;
9360 struct cgroup_subsys cpuacct_subsys
= {
9362 .create
= cpuacct_create
,
9363 .destroy
= cpuacct_destroy
,
9364 .populate
= cpuacct_populate
,
9365 .subsys_id
= cpuacct_subsys_id
,
9367 #endif /* CONFIG_CGROUP_CPUACCT */