4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/reciprocal_div.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/bootmem.h>
72 #include <linux/debugfs.h>
73 #include <linux/ctype.h>
74 #include <linux/ftrace.h>
75 #include <trace/sched.h>
78 #include <asm/irq_regs.h>
80 #include "sched_cpupri.h"
83 * Convert user-nice values [ -20 ... 0 ... 19 ]
84 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
87 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
88 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
89 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
92 * 'User priority' is the nice value converted to something we
93 * can work with better when scaling various scheduler parameters,
94 * it's a [ 0 ... 39 ] range.
96 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
97 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
98 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
101 * Helpers for converting nanosecond timing to jiffy resolution
103 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
105 #define NICE_0_LOAD SCHED_LOAD_SCALE
106 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
109 * These are the 'tuning knobs' of the scheduler:
111 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
112 * Timeslices get refilled after they expire.
114 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * single value that denotes runtime == period, ie unlimited time.
119 #define RUNTIME_INF ((u64)~0ULL)
121 DEFINE_TRACE(sched_wait_task
);
122 DEFINE_TRACE(sched_wakeup
);
123 DEFINE_TRACE(sched_wakeup_new
);
124 DEFINE_TRACE(sched_switch
);
125 DEFINE_TRACE(sched_migrate_task
);
129 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
130 * Since cpu_power is a 'constant', we can use a reciprocal divide.
132 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
134 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
138 * Each time a sched group cpu_power is changed,
139 * we must compute its reciprocal value
141 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
143 sg
->__cpu_power
+= val
;
144 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
148 static inline int rt_policy(int policy
)
150 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
155 static inline int task_has_rt_policy(struct task_struct
*p
)
157 return rt_policy(p
->policy
);
161 * This is the priority-queue data structure of the RT scheduling class:
163 struct rt_prio_array
{
164 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
165 struct list_head queue
[MAX_RT_PRIO
];
168 struct rt_bandwidth
{
169 /* nests inside the rq lock: */
170 spinlock_t rt_runtime_lock
;
173 struct hrtimer rt_period_timer
;
176 static struct rt_bandwidth def_rt_bandwidth
;
178 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
180 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
182 struct rt_bandwidth
*rt_b
=
183 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
189 now
= hrtimer_cb_get_time(timer
);
190 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
195 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
198 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
202 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
204 rt_b
->rt_period
= ns_to_ktime(period
);
205 rt_b
->rt_runtime
= runtime
;
207 spin_lock_init(&rt_b
->rt_runtime_lock
);
209 hrtimer_init(&rt_b
->rt_period_timer
,
210 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
211 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
212 rt_b
->rt_period_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_UNLOCKED
;
215 static inline int rt_bandwidth_enabled(void)
217 return sysctl_sched_rt_runtime
>= 0;
220 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
224 if (rt_bandwidth_enabled() && rt_b
->rt_runtime
== RUNTIME_INF
)
227 if (hrtimer_active(&rt_b
->rt_period_timer
))
230 spin_lock(&rt_b
->rt_runtime_lock
);
232 if (hrtimer_active(&rt_b
->rt_period_timer
))
235 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
236 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
237 hrtimer_start_expires(&rt_b
->rt_period_timer
,
240 spin_unlock(&rt_b
->rt_runtime_lock
);
243 #ifdef CONFIG_RT_GROUP_SCHED
244 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
246 hrtimer_cancel(&rt_b
->rt_period_timer
);
251 * sched_domains_mutex serializes calls to arch_init_sched_domains,
252 * detach_destroy_domains and partition_sched_domains.
254 static DEFINE_MUTEX(sched_domains_mutex
);
256 #ifdef CONFIG_GROUP_SCHED
258 #include <linux/cgroup.h>
262 static LIST_HEAD(task_groups
);
264 /* task group related information */
266 #ifdef CONFIG_CGROUP_SCHED
267 struct cgroup_subsys_state css
;
270 #ifdef CONFIG_FAIR_GROUP_SCHED
271 /* schedulable entities of this group on each cpu */
272 struct sched_entity
**se
;
273 /* runqueue "owned" by this group on each cpu */
274 struct cfs_rq
**cfs_rq
;
275 unsigned long shares
;
278 #ifdef CONFIG_RT_GROUP_SCHED
279 struct sched_rt_entity
**rt_se
;
280 struct rt_rq
**rt_rq
;
282 struct rt_bandwidth rt_bandwidth
;
286 struct list_head list
;
288 struct task_group
*parent
;
289 struct list_head siblings
;
290 struct list_head children
;
293 #ifdef CONFIG_USER_SCHED
297 * Every UID task group (including init_task_group aka UID-0) will
298 * be a child to this group.
300 struct task_group root_task_group
;
302 #ifdef CONFIG_FAIR_GROUP_SCHED
303 /* Default task group's sched entity on each cpu */
304 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
305 /* Default task group's cfs_rq on each cpu */
306 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
307 #endif /* CONFIG_FAIR_GROUP_SCHED */
309 #ifdef CONFIG_RT_GROUP_SCHED
310 static DEFINE_PER_CPU(struct sched_rt_entity
, init_sched_rt_entity
);
311 static DEFINE_PER_CPU(struct rt_rq
, init_rt_rq
) ____cacheline_aligned_in_smp
;
312 #endif /* CONFIG_RT_GROUP_SCHED */
313 #else /* !CONFIG_USER_SCHED */
314 #define root_task_group init_task_group
315 #endif /* CONFIG_USER_SCHED */
317 /* task_group_lock serializes add/remove of task groups and also changes to
318 * a task group's cpu shares.
320 static DEFINE_SPINLOCK(task_group_lock
);
322 #ifdef CONFIG_FAIR_GROUP_SCHED
323 #ifdef CONFIG_USER_SCHED
324 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
325 #else /* !CONFIG_USER_SCHED */
326 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
327 #endif /* CONFIG_USER_SCHED */
330 * A weight of 0 or 1 can cause arithmetics problems.
331 * A weight of a cfs_rq is the sum of weights of which entities
332 * are queued on this cfs_rq, so a weight of a entity should not be
333 * too large, so as the shares value of a task group.
334 * (The default weight is 1024 - so there's no practical
335 * limitation from this.)
338 #define MAX_SHARES (1UL << 18)
340 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
343 /* Default task group.
344 * Every task in system belong to this group at bootup.
346 struct task_group init_task_group
;
348 /* return group to which a task belongs */
349 static inline struct task_group
*task_group(struct task_struct
*p
)
351 struct task_group
*tg
;
353 #ifdef CONFIG_USER_SCHED
355 #elif defined(CONFIG_CGROUP_SCHED)
356 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
357 struct task_group
, css
);
359 tg
= &init_task_group
;
364 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
365 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
367 #ifdef CONFIG_FAIR_GROUP_SCHED
368 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
369 p
->se
.parent
= task_group(p
)->se
[cpu
];
372 #ifdef CONFIG_RT_GROUP_SCHED
373 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
374 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
380 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
381 static inline struct task_group
*task_group(struct task_struct
*p
)
386 #endif /* CONFIG_GROUP_SCHED */
388 /* CFS-related fields in a runqueue */
390 struct load_weight load
;
391 unsigned long nr_running
;
396 struct rb_root tasks_timeline
;
397 struct rb_node
*rb_leftmost
;
399 struct list_head tasks
;
400 struct list_head
*balance_iterator
;
403 * 'curr' points to currently running entity on this cfs_rq.
404 * It is set to NULL otherwise (i.e when none are currently running).
406 struct sched_entity
*curr
, *next
, *last
;
408 unsigned int nr_spread_over
;
410 #ifdef CONFIG_FAIR_GROUP_SCHED
411 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
414 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
415 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
416 * (like users, containers etc.)
418 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
419 * list is used during load balance.
421 struct list_head leaf_cfs_rq_list
;
422 struct task_group
*tg
; /* group that "owns" this runqueue */
426 * the part of load.weight contributed by tasks
428 unsigned long task_weight
;
431 * h_load = weight * f(tg)
433 * Where f(tg) is the recursive weight fraction assigned to
436 unsigned long h_load
;
439 * this cpu's part of tg->shares
441 unsigned long shares
;
444 * load.weight at the time we set shares
446 unsigned long rq_weight
;
451 /* Real-Time classes' related field in a runqueue: */
453 struct rt_prio_array active
;
454 unsigned long rt_nr_running
;
455 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
456 int highest_prio
; /* highest queued rt task prio */
459 unsigned long rt_nr_migratory
;
465 /* Nests inside the rq lock: */
466 spinlock_t rt_runtime_lock
;
468 #ifdef CONFIG_RT_GROUP_SCHED
469 unsigned long rt_nr_boosted
;
472 struct list_head leaf_rt_rq_list
;
473 struct task_group
*tg
;
474 struct sched_rt_entity
*rt_se
;
481 * We add the notion of a root-domain which will be used to define per-domain
482 * variables. Each exclusive cpuset essentially defines an island domain by
483 * fully partitioning the member cpus from any other cpuset. Whenever a new
484 * exclusive cpuset is created, we also create and attach a new root-domain
494 * The "RT overload" flag: it gets set if a CPU has more than
495 * one runnable RT task.
500 struct cpupri cpupri
;
505 * By default the system creates a single root-domain with all cpus as
506 * members (mimicking the global state we have today).
508 static struct root_domain def_root_domain
;
513 * This is the main, per-CPU runqueue data structure.
515 * Locking rule: those places that want to lock multiple runqueues
516 * (such as the load balancing or the thread migration code), lock
517 * acquire operations must be ordered by ascending &runqueue.
524 * nr_running and cpu_load should be in the same cacheline because
525 * remote CPUs use both these fields when doing load calculation.
527 unsigned long nr_running
;
528 #define CPU_LOAD_IDX_MAX 5
529 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
530 unsigned char idle_at_tick
;
532 unsigned long last_tick_seen
;
533 unsigned char in_nohz_recently
;
535 /* capture load from *all* tasks on this cpu: */
536 struct load_weight load
;
537 unsigned long nr_load_updates
;
543 #ifdef CONFIG_FAIR_GROUP_SCHED
544 /* list of leaf cfs_rq on this cpu: */
545 struct list_head leaf_cfs_rq_list
;
547 #ifdef CONFIG_RT_GROUP_SCHED
548 struct list_head leaf_rt_rq_list
;
552 * This is part of a global counter where only the total sum
553 * over all CPUs matters. A task can increase this counter on
554 * one CPU and if it got migrated afterwards it may decrease
555 * it on another CPU. Always updated under the runqueue lock:
557 unsigned long nr_uninterruptible
;
559 struct task_struct
*curr
, *idle
;
560 unsigned long next_balance
;
561 struct mm_struct
*prev_mm
;
568 struct root_domain
*rd
;
569 struct sched_domain
*sd
;
571 /* For active balancing */
574 /* cpu of this runqueue: */
578 unsigned long avg_load_per_task
;
580 struct task_struct
*migration_thread
;
581 struct list_head migration_queue
;
584 #ifdef CONFIG_SCHED_HRTICK
586 int hrtick_csd_pending
;
587 struct call_single_data hrtick_csd
;
589 struct hrtimer hrtick_timer
;
592 #ifdef CONFIG_SCHEDSTATS
594 struct sched_info rq_sched_info
;
596 /* sys_sched_yield() stats */
597 unsigned int yld_exp_empty
;
598 unsigned int yld_act_empty
;
599 unsigned int yld_both_empty
;
600 unsigned int yld_count
;
602 /* schedule() stats */
603 unsigned int sched_switch
;
604 unsigned int sched_count
;
605 unsigned int sched_goidle
;
607 /* try_to_wake_up() stats */
608 unsigned int ttwu_count
;
609 unsigned int ttwu_local
;
612 unsigned int bkl_count
;
616 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
618 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int sync
)
620 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, sync
);
623 static inline int cpu_of(struct rq
*rq
)
633 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
634 * See detach_destroy_domains: synchronize_sched for details.
636 * The domain tree of any CPU may only be accessed from within
637 * preempt-disabled sections.
639 #define for_each_domain(cpu, __sd) \
640 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
642 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
643 #define this_rq() (&__get_cpu_var(runqueues))
644 #define task_rq(p) cpu_rq(task_cpu(p))
645 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
647 static inline void update_rq_clock(struct rq
*rq
)
649 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
653 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
655 #ifdef CONFIG_SCHED_DEBUG
656 # define const_debug __read_mostly
658 # define const_debug static const
664 * Returns true if the current cpu runqueue is locked.
665 * This interface allows printk to be called with the runqueue lock
666 * held and know whether or not it is OK to wake up the klogd.
668 int runqueue_is_locked(void)
671 struct rq
*rq
= cpu_rq(cpu
);
674 ret
= spin_is_locked(&rq
->lock
);
680 * Debugging: various feature bits
683 #define SCHED_FEAT(name, enabled) \
684 __SCHED_FEAT_##name ,
687 #include "sched_features.h"
692 #define SCHED_FEAT(name, enabled) \
693 (1UL << __SCHED_FEAT_##name) * enabled |
695 const_debug
unsigned int sysctl_sched_features
=
696 #include "sched_features.h"
701 #ifdef CONFIG_SCHED_DEBUG
702 #define SCHED_FEAT(name, enabled) \
705 static __read_mostly
char *sched_feat_names
[] = {
706 #include "sched_features.h"
712 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
714 filp
->private_data
= inode
->i_private
;
719 sched_feat_read(struct file
*filp
, char __user
*ubuf
,
720 size_t cnt
, loff_t
*ppos
)
727 for (i
= 0; sched_feat_names
[i
]; i
++) {
728 len
+= strlen(sched_feat_names
[i
]);
732 buf
= kmalloc(len
+ 2, GFP_KERNEL
);
736 for (i
= 0; sched_feat_names
[i
]; i
++) {
737 if (sysctl_sched_features
& (1UL << i
))
738 r
+= sprintf(buf
+ r
, "%s ", sched_feat_names
[i
]);
740 r
+= sprintf(buf
+ r
, "NO_%s ", sched_feat_names
[i
]);
743 r
+= sprintf(buf
+ r
, "\n");
744 WARN_ON(r
>= len
+ 2);
746 r
= simple_read_from_buffer(ubuf
, cnt
, ppos
, buf
, r
);
754 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
755 size_t cnt
, loff_t
*ppos
)
765 if (copy_from_user(&buf
, ubuf
, cnt
))
770 if (strncmp(buf
, "NO_", 3) == 0) {
775 for (i
= 0; sched_feat_names
[i
]; i
++) {
776 int len
= strlen(sched_feat_names
[i
]);
778 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
780 sysctl_sched_features
&= ~(1UL << i
);
782 sysctl_sched_features
|= (1UL << i
);
787 if (!sched_feat_names
[i
])
795 static struct file_operations sched_feat_fops
= {
796 .open
= sched_feat_open
,
797 .read
= sched_feat_read
,
798 .write
= sched_feat_write
,
801 static __init
int sched_init_debug(void)
803 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
808 late_initcall(sched_init_debug
);
812 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
815 * Number of tasks to iterate in a single balance run.
816 * Limited because this is done with IRQs disabled.
818 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
821 * ratelimit for updating the group shares.
824 unsigned int sysctl_sched_shares_ratelimit
= 250000;
827 * Inject some fuzzyness into changing the per-cpu group shares
828 * this avoids remote rq-locks at the expense of fairness.
831 unsigned int sysctl_sched_shares_thresh
= 4;
834 * period over which we measure -rt task cpu usage in us.
837 unsigned int sysctl_sched_rt_period
= 1000000;
839 static __read_mostly
int scheduler_running
;
842 * part of the period that we allow rt tasks to run in us.
845 int sysctl_sched_rt_runtime
= 950000;
847 static inline u64
global_rt_period(void)
849 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
852 static inline u64
global_rt_runtime(void)
854 if (sysctl_sched_rt_runtime
< 0)
857 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
860 #ifndef prepare_arch_switch
861 # define prepare_arch_switch(next) do { } while (0)
863 #ifndef finish_arch_switch
864 # define finish_arch_switch(prev) do { } while (0)
867 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
869 return rq
->curr
== p
;
872 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
873 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
875 return task_current(rq
, p
);
878 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
882 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
884 #ifdef CONFIG_DEBUG_SPINLOCK
885 /* this is a valid case when another task releases the spinlock */
886 rq
->lock
.owner
= current
;
889 * If we are tracking spinlock dependencies then we have to
890 * fix up the runqueue lock - which gets 'carried over' from
893 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
895 spin_unlock_irq(&rq
->lock
);
898 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
899 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
904 return task_current(rq
, p
);
908 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
912 * We can optimise this out completely for !SMP, because the
913 * SMP rebalancing from interrupt is the only thing that cares
918 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
919 spin_unlock_irq(&rq
->lock
);
921 spin_unlock(&rq
->lock
);
925 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
929 * After ->oncpu is cleared, the task can be moved to a different CPU.
930 * We must ensure this doesn't happen until the switch is completely
936 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
940 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
943 * __task_rq_lock - lock the runqueue a given task resides on.
944 * Must be called interrupts disabled.
946 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
950 struct rq
*rq
= task_rq(p
);
951 spin_lock(&rq
->lock
);
952 if (likely(rq
== task_rq(p
)))
954 spin_unlock(&rq
->lock
);
959 * task_rq_lock - lock the runqueue a given task resides on and disable
960 * interrupts. Note the ordering: we can safely lookup the task_rq without
961 * explicitly disabling preemption.
963 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
969 local_irq_save(*flags
);
971 spin_lock(&rq
->lock
);
972 if (likely(rq
== task_rq(p
)))
974 spin_unlock_irqrestore(&rq
->lock
, *flags
);
978 void task_rq_unlock_wait(struct task_struct
*p
)
980 struct rq
*rq
= task_rq(p
);
982 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
983 spin_unlock_wait(&rq
->lock
);
986 static void __task_rq_unlock(struct rq
*rq
)
989 spin_unlock(&rq
->lock
);
992 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
995 spin_unlock_irqrestore(&rq
->lock
, *flags
);
999 * this_rq_lock - lock this runqueue and disable interrupts.
1001 static struct rq
*this_rq_lock(void)
1002 __acquires(rq
->lock
)
1006 local_irq_disable();
1008 spin_lock(&rq
->lock
);
1013 #ifdef CONFIG_SCHED_HRTICK
1015 * Use HR-timers to deliver accurate preemption points.
1017 * Its all a bit involved since we cannot program an hrt while holding the
1018 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1021 * When we get rescheduled we reprogram the hrtick_timer outside of the
1027 * - enabled by features
1028 * - hrtimer is actually high res
1030 static inline int hrtick_enabled(struct rq
*rq
)
1032 if (!sched_feat(HRTICK
))
1034 if (!cpu_active(cpu_of(rq
)))
1036 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1039 static void hrtick_clear(struct rq
*rq
)
1041 if (hrtimer_active(&rq
->hrtick_timer
))
1042 hrtimer_cancel(&rq
->hrtick_timer
);
1046 * High-resolution timer tick.
1047 * Runs from hardirq context with interrupts disabled.
1049 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1051 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1053 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1055 spin_lock(&rq
->lock
);
1056 update_rq_clock(rq
);
1057 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1058 spin_unlock(&rq
->lock
);
1060 return HRTIMER_NORESTART
;
1065 * called from hardirq (IPI) context
1067 static void __hrtick_start(void *arg
)
1069 struct rq
*rq
= arg
;
1071 spin_lock(&rq
->lock
);
1072 hrtimer_restart(&rq
->hrtick_timer
);
1073 rq
->hrtick_csd_pending
= 0;
1074 spin_unlock(&rq
->lock
);
1078 * Called to set the hrtick timer state.
1080 * called with rq->lock held and irqs disabled
1082 static void hrtick_start(struct rq
*rq
, u64 delay
)
1084 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1085 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1087 hrtimer_set_expires(timer
, time
);
1089 if (rq
== this_rq()) {
1090 hrtimer_restart(timer
);
1091 } else if (!rq
->hrtick_csd_pending
) {
1092 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
);
1093 rq
->hrtick_csd_pending
= 1;
1098 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1100 int cpu
= (int)(long)hcpu
;
1103 case CPU_UP_CANCELED
:
1104 case CPU_UP_CANCELED_FROZEN
:
1105 case CPU_DOWN_PREPARE
:
1106 case CPU_DOWN_PREPARE_FROZEN
:
1108 case CPU_DEAD_FROZEN
:
1109 hrtick_clear(cpu_rq(cpu
));
1116 static __init
void init_hrtick(void)
1118 hotcpu_notifier(hotplug_hrtick
, 0);
1122 * Called to set the hrtick timer state.
1124 * called with rq->lock held and irqs disabled
1126 static void hrtick_start(struct rq
*rq
, u64 delay
)
1128 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
), HRTIMER_MODE_REL
);
1131 static inline void init_hrtick(void)
1134 #endif /* CONFIG_SMP */
1136 static void init_rq_hrtick(struct rq
*rq
)
1139 rq
->hrtick_csd_pending
= 0;
1141 rq
->hrtick_csd
.flags
= 0;
1142 rq
->hrtick_csd
.func
= __hrtick_start
;
1143 rq
->hrtick_csd
.info
= rq
;
1146 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1147 rq
->hrtick_timer
.function
= hrtick
;
1148 rq
->hrtick_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_PERCPU
;
1150 #else /* CONFIG_SCHED_HRTICK */
1151 static inline void hrtick_clear(struct rq
*rq
)
1155 static inline void init_rq_hrtick(struct rq
*rq
)
1159 static inline void init_hrtick(void)
1162 #endif /* CONFIG_SCHED_HRTICK */
1165 * resched_task - mark a task 'to be rescheduled now'.
1167 * On UP this means the setting of the need_resched flag, on SMP it
1168 * might also involve a cross-CPU call to trigger the scheduler on
1173 #ifndef tsk_is_polling
1174 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1177 static void resched_task(struct task_struct
*p
)
1181 assert_spin_locked(&task_rq(p
)->lock
);
1183 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
1186 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
1189 if (cpu
== smp_processor_id())
1192 /* NEED_RESCHED must be visible before we test polling */
1194 if (!tsk_is_polling(p
))
1195 smp_send_reschedule(cpu
);
1198 static void resched_cpu(int cpu
)
1200 struct rq
*rq
= cpu_rq(cpu
);
1201 unsigned long flags
;
1203 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1205 resched_task(cpu_curr(cpu
));
1206 spin_unlock_irqrestore(&rq
->lock
, flags
);
1211 * When add_timer_on() enqueues a timer into the timer wheel of an
1212 * idle CPU then this timer might expire before the next timer event
1213 * which is scheduled to wake up that CPU. In case of a completely
1214 * idle system the next event might even be infinite time into the
1215 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1216 * leaves the inner idle loop so the newly added timer is taken into
1217 * account when the CPU goes back to idle and evaluates the timer
1218 * wheel for the next timer event.
1220 void wake_up_idle_cpu(int cpu
)
1222 struct rq
*rq
= cpu_rq(cpu
);
1224 if (cpu
== smp_processor_id())
1228 * This is safe, as this function is called with the timer
1229 * wheel base lock of (cpu) held. When the CPU is on the way
1230 * to idle and has not yet set rq->curr to idle then it will
1231 * be serialized on the timer wheel base lock and take the new
1232 * timer into account automatically.
1234 if (rq
->curr
!= rq
->idle
)
1238 * We can set TIF_RESCHED on the idle task of the other CPU
1239 * lockless. The worst case is that the other CPU runs the
1240 * idle task through an additional NOOP schedule()
1242 set_tsk_thread_flag(rq
->idle
, TIF_NEED_RESCHED
);
1244 /* NEED_RESCHED must be visible before we test polling */
1246 if (!tsk_is_polling(rq
->idle
))
1247 smp_send_reschedule(cpu
);
1249 #endif /* CONFIG_NO_HZ */
1251 #else /* !CONFIG_SMP */
1252 static void resched_task(struct task_struct
*p
)
1254 assert_spin_locked(&task_rq(p
)->lock
);
1255 set_tsk_need_resched(p
);
1257 #endif /* CONFIG_SMP */
1259 #if BITS_PER_LONG == 32
1260 # define WMULT_CONST (~0UL)
1262 # define WMULT_CONST (1UL << 32)
1265 #define WMULT_SHIFT 32
1268 * Shift right and round:
1270 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1273 * delta *= weight / lw
1275 static unsigned long
1276 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1277 struct load_weight
*lw
)
1281 if (!lw
->inv_weight
) {
1282 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1285 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1289 tmp
= (u64
)delta_exec
* weight
;
1291 * Check whether we'd overflow the 64-bit multiplication:
1293 if (unlikely(tmp
> WMULT_CONST
))
1294 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1297 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1299 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1302 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1308 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1315 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1316 * of tasks with abnormal "nice" values across CPUs the contribution that
1317 * each task makes to its run queue's load is weighted according to its
1318 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1319 * scaled version of the new time slice allocation that they receive on time
1323 #define WEIGHT_IDLEPRIO 2
1324 #define WMULT_IDLEPRIO (1 << 31)
1327 * Nice levels are multiplicative, with a gentle 10% change for every
1328 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1329 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1330 * that remained on nice 0.
1332 * The "10% effect" is relative and cumulative: from _any_ nice level,
1333 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1334 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1335 * If a task goes up by ~10% and another task goes down by ~10% then
1336 * the relative distance between them is ~25%.)
1338 static const int prio_to_weight
[40] = {
1339 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1340 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1341 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1342 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1343 /* 0 */ 1024, 820, 655, 526, 423,
1344 /* 5 */ 335, 272, 215, 172, 137,
1345 /* 10 */ 110, 87, 70, 56, 45,
1346 /* 15 */ 36, 29, 23, 18, 15,
1350 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1352 * In cases where the weight does not change often, we can use the
1353 * precalculated inverse to speed up arithmetics by turning divisions
1354 * into multiplications:
1356 static const u32 prio_to_wmult
[40] = {
1357 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1358 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1359 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1360 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1361 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1362 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1363 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1364 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1367 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1370 * runqueue iterator, to support SMP load-balancing between different
1371 * scheduling classes, without having to expose their internal data
1372 * structures to the load-balancing proper:
1374 struct rq_iterator
{
1376 struct task_struct
*(*start
)(void *);
1377 struct task_struct
*(*next
)(void *);
1381 static unsigned long
1382 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1383 unsigned long max_load_move
, struct sched_domain
*sd
,
1384 enum cpu_idle_type idle
, int *all_pinned
,
1385 int *this_best_prio
, struct rq_iterator
*iterator
);
1388 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1389 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1390 struct rq_iterator
*iterator
);
1393 #ifdef CONFIG_CGROUP_CPUACCT
1394 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1396 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1399 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1401 update_load_add(&rq
->load
, load
);
1404 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1406 update_load_sub(&rq
->load
, load
);
1409 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1410 typedef int (*tg_visitor
)(struct task_group
*, void *);
1413 * Iterate the full tree, calling @down when first entering a node and @up when
1414 * leaving it for the final time.
1416 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1418 struct task_group
*parent
, *child
;
1422 parent
= &root_task_group
;
1424 ret
= (*down
)(parent
, data
);
1427 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1434 ret
= (*up
)(parent
, data
);
1439 parent
= parent
->parent
;
1448 static int tg_nop(struct task_group
*tg
, void *data
)
1455 static unsigned long source_load(int cpu
, int type
);
1456 static unsigned long target_load(int cpu
, int type
);
1457 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1459 static unsigned long cpu_avg_load_per_task(int cpu
)
1461 struct rq
*rq
= cpu_rq(cpu
);
1464 rq
->avg_load_per_task
= rq
->load
.weight
/ rq
->nr_running
;
1466 rq
->avg_load_per_task
= 0;
1468 return rq
->avg_load_per_task
;
1471 #ifdef CONFIG_FAIR_GROUP_SCHED
1473 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1476 * Calculate and set the cpu's group shares.
1479 update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1480 unsigned long sd_shares
, unsigned long sd_rq_weight
)
1483 unsigned long shares
;
1484 unsigned long rq_weight
;
1489 rq_weight
= tg
->cfs_rq
[cpu
]->load
.weight
;
1492 * If there are currently no tasks on the cpu pretend there is one of
1493 * average load so that when a new task gets to run here it will not
1494 * get delayed by group starvation.
1498 rq_weight
= NICE_0_LOAD
;
1501 if (unlikely(rq_weight
> sd_rq_weight
))
1502 rq_weight
= sd_rq_weight
;
1505 * \Sum shares * rq_weight
1506 * shares = -----------------------
1510 shares
= (sd_shares
* rq_weight
) / (sd_rq_weight
+ 1);
1511 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1513 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1514 sysctl_sched_shares_thresh
) {
1515 struct rq
*rq
= cpu_rq(cpu
);
1516 unsigned long flags
;
1518 spin_lock_irqsave(&rq
->lock
, flags
);
1520 * record the actual number of shares, not the boosted amount.
1522 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1523 tg
->cfs_rq
[cpu
]->rq_weight
= rq_weight
;
1525 __set_se_shares(tg
->se
[cpu
], shares
);
1526 spin_unlock_irqrestore(&rq
->lock
, flags
);
1531 * Re-compute the task group their per cpu shares over the given domain.
1532 * This needs to be done in a bottom-up fashion because the rq weight of a
1533 * parent group depends on the shares of its child groups.
1535 static int tg_shares_up(struct task_group
*tg
, void *data
)
1537 unsigned long rq_weight
= 0;
1538 unsigned long shares
= 0;
1539 struct sched_domain
*sd
= data
;
1542 for_each_cpu_mask(i
, sd
->span
) {
1543 rq_weight
+= tg
->cfs_rq
[i
]->load
.weight
;
1544 shares
+= tg
->cfs_rq
[i
]->shares
;
1547 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1548 shares
= tg
->shares
;
1550 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1551 shares
= tg
->shares
;
1554 rq_weight
= cpus_weight(sd
->span
) * NICE_0_LOAD
;
1556 for_each_cpu_mask(i
, sd
->span
)
1557 update_group_shares_cpu(tg
, i
, shares
, rq_weight
);
1563 * Compute the cpu's hierarchical load factor for each task group.
1564 * This needs to be done in a top-down fashion because the load of a child
1565 * group is a fraction of its parents load.
1567 static int tg_load_down(struct task_group
*tg
, void *data
)
1570 long cpu
= (long)data
;
1573 load
= cpu_rq(cpu
)->load
.weight
;
1575 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1576 load
*= tg
->cfs_rq
[cpu
]->shares
;
1577 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1580 tg
->cfs_rq
[cpu
]->h_load
= load
;
1585 static void update_shares(struct sched_domain
*sd
)
1587 u64 now
= cpu_clock(raw_smp_processor_id());
1588 s64 elapsed
= now
- sd
->last_update
;
1590 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1591 sd
->last_update
= now
;
1592 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1596 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1598 spin_unlock(&rq
->lock
);
1600 spin_lock(&rq
->lock
);
1603 static void update_h_load(long cpu
)
1605 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1610 static inline void update_shares(struct sched_domain
*sd
)
1614 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1622 #ifdef CONFIG_FAIR_GROUP_SCHED
1623 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1626 cfs_rq
->shares
= shares
;
1631 #include "sched_stats.h"
1632 #include "sched_idletask.c"
1633 #include "sched_fair.c"
1634 #include "sched_rt.c"
1635 #ifdef CONFIG_SCHED_DEBUG
1636 # include "sched_debug.c"
1639 #define sched_class_highest (&rt_sched_class)
1640 #define for_each_class(class) \
1641 for (class = sched_class_highest; class; class = class->next)
1643 static void inc_nr_running(struct rq
*rq
)
1648 static void dec_nr_running(struct rq
*rq
)
1653 static void set_load_weight(struct task_struct
*p
)
1655 if (task_has_rt_policy(p
)) {
1656 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1657 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1662 * SCHED_IDLE tasks get minimal weight:
1664 if (p
->policy
== SCHED_IDLE
) {
1665 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1666 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1670 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1671 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1674 static void update_avg(u64
*avg
, u64 sample
)
1676 s64 diff
= sample
- *avg
;
1680 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1682 sched_info_queued(p
);
1683 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1687 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1689 if (sleep
&& p
->se
.last_wakeup
) {
1690 update_avg(&p
->se
.avg_overlap
,
1691 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1692 p
->se
.last_wakeup
= 0;
1695 sched_info_dequeued(p
);
1696 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1701 * __normal_prio - return the priority that is based on the static prio
1703 static inline int __normal_prio(struct task_struct
*p
)
1705 return p
->static_prio
;
1709 * Calculate the expected normal priority: i.e. priority
1710 * without taking RT-inheritance into account. Might be
1711 * boosted by interactivity modifiers. Changes upon fork,
1712 * setprio syscalls, and whenever the interactivity
1713 * estimator recalculates.
1715 static inline int normal_prio(struct task_struct
*p
)
1719 if (task_has_rt_policy(p
))
1720 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1722 prio
= __normal_prio(p
);
1727 * Calculate the current priority, i.e. the priority
1728 * taken into account by the scheduler. This value might
1729 * be boosted by RT tasks, or might be boosted by
1730 * interactivity modifiers. Will be RT if the task got
1731 * RT-boosted. If not then it returns p->normal_prio.
1733 static int effective_prio(struct task_struct
*p
)
1735 p
->normal_prio
= normal_prio(p
);
1737 * If we are RT tasks or we were boosted to RT priority,
1738 * keep the priority unchanged. Otherwise, update priority
1739 * to the normal priority:
1741 if (!rt_prio(p
->prio
))
1742 return p
->normal_prio
;
1747 * activate_task - move a task to the runqueue.
1749 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1751 if (task_contributes_to_load(p
))
1752 rq
->nr_uninterruptible
--;
1754 enqueue_task(rq
, p
, wakeup
);
1759 * deactivate_task - remove a task from the runqueue.
1761 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1763 if (task_contributes_to_load(p
))
1764 rq
->nr_uninterruptible
++;
1766 dequeue_task(rq
, p
, sleep
);
1771 * task_curr - is this task currently executing on a CPU?
1772 * @p: the task in question.
1774 inline int task_curr(const struct task_struct
*p
)
1776 return cpu_curr(task_cpu(p
)) == p
;
1779 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1781 set_task_rq(p
, cpu
);
1784 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1785 * successfuly executed on another CPU. We must ensure that updates of
1786 * per-task data have been completed by this moment.
1789 task_thread_info(p
)->cpu
= cpu
;
1793 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1794 const struct sched_class
*prev_class
,
1795 int oldprio
, int running
)
1797 if (prev_class
!= p
->sched_class
) {
1798 if (prev_class
->switched_from
)
1799 prev_class
->switched_from(rq
, p
, running
);
1800 p
->sched_class
->switched_to(rq
, p
, running
);
1802 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1807 /* Used instead of source_load when we know the type == 0 */
1808 static unsigned long weighted_cpuload(const int cpu
)
1810 return cpu_rq(cpu
)->load
.weight
;
1814 * Is this task likely cache-hot:
1817 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1822 * Buddy candidates are cache hot:
1824 if (sched_feat(CACHE_HOT_BUDDY
) &&
1825 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
1826 &p
->se
== cfs_rq_of(&p
->se
)->last
))
1829 if (p
->sched_class
!= &fair_sched_class
)
1832 if (sysctl_sched_migration_cost
== -1)
1834 if (sysctl_sched_migration_cost
== 0)
1837 delta
= now
- p
->se
.exec_start
;
1839 return delta
< (s64
)sysctl_sched_migration_cost
;
1843 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1845 int old_cpu
= task_cpu(p
);
1846 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1847 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1848 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1851 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1853 #ifdef CONFIG_SCHEDSTATS
1854 if (p
->se
.wait_start
)
1855 p
->se
.wait_start
-= clock_offset
;
1856 if (p
->se
.sleep_start
)
1857 p
->se
.sleep_start
-= clock_offset
;
1858 if (p
->se
.block_start
)
1859 p
->se
.block_start
-= clock_offset
;
1860 if (old_cpu
!= new_cpu
) {
1861 schedstat_inc(p
, se
.nr_migrations
);
1862 if (task_hot(p
, old_rq
->clock
, NULL
))
1863 schedstat_inc(p
, se
.nr_forced2_migrations
);
1866 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1867 new_cfsrq
->min_vruntime
;
1869 __set_task_cpu(p
, new_cpu
);
1872 struct migration_req
{
1873 struct list_head list
;
1875 struct task_struct
*task
;
1878 struct completion done
;
1882 * The task's runqueue lock must be held.
1883 * Returns true if you have to wait for migration thread.
1886 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1888 struct rq
*rq
= task_rq(p
);
1891 * If the task is not on a runqueue (and not running), then
1892 * it is sufficient to simply update the task's cpu field.
1894 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1895 set_task_cpu(p
, dest_cpu
);
1899 init_completion(&req
->done
);
1901 req
->dest_cpu
= dest_cpu
;
1902 list_add(&req
->list
, &rq
->migration_queue
);
1908 * wait_task_inactive - wait for a thread to unschedule.
1910 * If @match_state is nonzero, it's the @p->state value just checked and
1911 * not expected to change. If it changes, i.e. @p might have woken up,
1912 * then return zero. When we succeed in waiting for @p to be off its CPU,
1913 * we return a positive number (its total switch count). If a second call
1914 * a short while later returns the same number, the caller can be sure that
1915 * @p has remained unscheduled the whole time.
1917 * The caller must ensure that the task *will* unschedule sometime soon,
1918 * else this function might spin for a *long* time. This function can't
1919 * be called with interrupts off, or it may introduce deadlock with
1920 * smp_call_function() if an IPI is sent by the same process we are
1921 * waiting to become inactive.
1923 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1925 unsigned long flags
;
1932 * We do the initial early heuristics without holding
1933 * any task-queue locks at all. We'll only try to get
1934 * the runqueue lock when things look like they will
1940 * If the task is actively running on another CPU
1941 * still, just relax and busy-wait without holding
1944 * NOTE! Since we don't hold any locks, it's not
1945 * even sure that "rq" stays as the right runqueue!
1946 * But we don't care, since "task_running()" will
1947 * return false if the runqueue has changed and p
1948 * is actually now running somewhere else!
1950 while (task_running(rq
, p
)) {
1951 if (match_state
&& unlikely(p
->state
!= match_state
))
1957 * Ok, time to look more closely! We need the rq
1958 * lock now, to be *sure*. If we're wrong, we'll
1959 * just go back and repeat.
1961 rq
= task_rq_lock(p
, &flags
);
1962 trace_sched_wait_task(rq
, p
);
1963 running
= task_running(rq
, p
);
1964 on_rq
= p
->se
.on_rq
;
1966 if (!match_state
|| p
->state
== match_state
)
1967 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1968 task_rq_unlock(rq
, &flags
);
1971 * If it changed from the expected state, bail out now.
1973 if (unlikely(!ncsw
))
1977 * Was it really running after all now that we
1978 * checked with the proper locks actually held?
1980 * Oops. Go back and try again..
1982 if (unlikely(running
)) {
1988 * It's not enough that it's not actively running,
1989 * it must be off the runqueue _entirely_, and not
1992 * So if it wa still runnable (but just not actively
1993 * running right now), it's preempted, and we should
1994 * yield - it could be a while.
1996 if (unlikely(on_rq
)) {
1997 schedule_timeout_uninterruptible(1);
2002 * Ahh, all good. It wasn't running, and it wasn't
2003 * runnable, which means that it will never become
2004 * running in the future either. We're all done!
2013 * kick_process - kick a running thread to enter/exit the kernel
2014 * @p: the to-be-kicked thread
2016 * Cause a process which is running on another CPU to enter
2017 * kernel-mode, without any delay. (to get signals handled.)
2019 * NOTE: this function doesnt have to take the runqueue lock,
2020 * because all it wants to ensure is that the remote task enters
2021 * the kernel. If the IPI races and the task has been migrated
2022 * to another CPU then no harm is done and the purpose has been
2025 void kick_process(struct task_struct
*p
)
2031 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2032 smp_send_reschedule(cpu
);
2037 * Return a low guess at the load of a migration-source cpu weighted
2038 * according to the scheduling class and "nice" value.
2040 * We want to under-estimate the load of migration sources, to
2041 * balance conservatively.
2043 static unsigned long source_load(int cpu
, int type
)
2045 struct rq
*rq
= cpu_rq(cpu
);
2046 unsigned long total
= weighted_cpuload(cpu
);
2048 if (type
== 0 || !sched_feat(LB_BIAS
))
2051 return min(rq
->cpu_load
[type
-1], total
);
2055 * Return a high guess at the load of a migration-target cpu weighted
2056 * according to the scheduling class and "nice" value.
2058 static unsigned long target_load(int cpu
, int type
)
2060 struct rq
*rq
= cpu_rq(cpu
);
2061 unsigned long total
= weighted_cpuload(cpu
);
2063 if (type
== 0 || !sched_feat(LB_BIAS
))
2066 return max(rq
->cpu_load
[type
-1], total
);
2070 * find_idlest_group finds and returns the least busy CPU group within the
2073 static struct sched_group
*
2074 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
2076 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2077 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
2078 int load_idx
= sd
->forkexec_idx
;
2079 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
2082 unsigned long load
, avg_load
;
2086 /* Skip over this group if it has no CPUs allowed */
2087 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
2090 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2092 /* Tally up the load of all CPUs in the group */
2095 for_each_cpu_mask_nr(i
, group
->cpumask
) {
2096 /* Bias balancing toward cpus of our domain */
2098 load
= source_load(i
, load_idx
);
2100 load
= target_load(i
, load_idx
);
2105 /* Adjust by relative CPU power of the group */
2106 avg_load
= sg_div_cpu_power(group
,
2107 avg_load
* SCHED_LOAD_SCALE
);
2110 this_load
= avg_load
;
2112 } else if (avg_load
< min_load
) {
2113 min_load
= avg_load
;
2116 } while (group
= group
->next
, group
!= sd
->groups
);
2118 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
2124 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2127 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
,
2130 unsigned long load
, min_load
= ULONG_MAX
;
2134 /* Traverse only the allowed CPUs */
2135 cpus_and(*tmp
, group
->cpumask
, p
->cpus_allowed
);
2137 for_each_cpu_mask_nr(i
, *tmp
) {
2138 load
= weighted_cpuload(i
);
2140 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2150 * sched_balance_self: balance the current task (running on cpu) in domains
2151 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2154 * Balance, ie. select the least loaded group.
2156 * Returns the target CPU number, or the same CPU if no balancing is needed.
2158 * preempt must be disabled.
2160 static int sched_balance_self(int cpu
, int flag
)
2162 struct task_struct
*t
= current
;
2163 struct sched_domain
*tmp
, *sd
= NULL
;
2165 for_each_domain(cpu
, tmp
) {
2167 * If power savings logic is enabled for a domain, stop there.
2169 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2171 if (tmp
->flags
& flag
)
2179 cpumask_t span
, tmpmask
;
2180 struct sched_group
*group
;
2181 int new_cpu
, weight
;
2183 if (!(sd
->flags
& flag
)) {
2189 group
= find_idlest_group(sd
, t
, cpu
);
2195 new_cpu
= find_idlest_cpu(group
, t
, cpu
, &tmpmask
);
2196 if (new_cpu
== -1 || new_cpu
== cpu
) {
2197 /* Now try balancing at a lower domain level of cpu */
2202 /* Now try balancing at a lower domain level of new_cpu */
2205 weight
= cpus_weight(span
);
2206 for_each_domain(cpu
, tmp
) {
2207 if (weight
<= cpus_weight(tmp
->span
))
2209 if (tmp
->flags
& flag
)
2212 /* while loop will break here if sd == NULL */
2218 #endif /* CONFIG_SMP */
2221 * try_to_wake_up - wake up a thread
2222 * @p: the to-be-woken-up thread
2223 * @state: the mask of task states that can be woken
2224 * @sync: do a synchronous wakeup?
2226 * Put it on the run-queue if it's not already there. The "current"
2227 * thread is always on the run-queue (except when the actual
2228 * re-schedule is in progress), and as such you're allowed to do
2229 * the simpler "current->state = TASK_RUNNING" to mark yourself
2230 * runnable without the overhead of this.
2232 * returns failure only if the task is already active.
2234 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
2236 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2237 unsigned long flags
;
2241 if (!sched_feat(SYNC_WAKEUPS
))
2245 if (sched_feat(LB_WAKEUP_UPDATE
)) {
2246 struct sched_domain
*sd
;
2248 this_cpu
= raw_smp_processor_id();
2251 for_each_domain(this_cpu
, sd
) {
2252 if (cpu_isset(cpu
, sd
->span
)) {
2261 rq
= task_rq_lock(p
, &flags
);
2262 old_state
= p
->state
;
2263 if (!(old_state
& state
))
2271 this_cpu
= smp_processor_id();
2274 if (unlikely(task_running(rq
, p
)))
2277 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2278 if (cpu
!= orig_cpu
) {
2279 set_task_cpu(p
, cpu
);
2280 task_rq_unlock(rq
, &flags
);
2281 /* might preempt at this point */
2282 rq
= task_rq_lock(p
, &flags
);
2283 old_state
= p
->state
;
2284 if (!(old_state
& state
))
2289 this_cpu
= smp_processor_id();
2293 #ifdef CONFIG_SCHEDSTATS
2294 schedstat_inc(rq
, ttwu_count
);
2295 if (cpu
== this_cpu
)
2296 schedstat_inc(rq
, ttwu_local
);
2298 struct sched_domain
*sd
;
2299 for_each_domain(this_cpu
, sd
) {
2300 if (cpu_isset(cpu
, sd
->span
)) {
2301 schedstat_inc(sd
, ttwu_wake_remote
);
2306 #endif /* CONFIG_SCHEDSTATS */
2309 #endif /* CONFIG_SMP */
2310 schedstat_inc(p
, se
.nr_wakeups
);
2312 schedstat_inc(p
, se
.nr_wakeups_sync
);
2313 if (orig_cpu
!= cpu
)
2314 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2315 if (cpu
== this_cpu
)
2316 schedstat_inc(p
, se
.nr_wakeups_local
);
2318 schedstat_inc(p
, se
.nr_wakeups_remote
);
2319 update_rq_clock(rq
);
2320 activate_task(rq
, p
, 1);
2324 trace_sched_wakeup(rq
, p
);
2325 check_preempt_curr(rq
, p
, sync
);
2327 p
->state
= TASK_RUNNING
;
2329 if (p
->sched_class
->task_wake_up
)
2330 p
->sched_class
->task_wake_up(rq
, p
);
2333 current
->se
.last_wakeup
= current
->se
.sum_exec_runtime
;
2335 task_rq_unlock(rq
, &flags
);
2340 int wake_up_process(struct task_struct
*p
)
2342 return try_to_wake_up(p
, TASK_ALL
, 0);
2344 EXPORT_SYMBOL(wake_up_process
);
2346 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2348 return try_to_wake_up(p
, state
, 0);
2352 * Perform scheduler related setup for a newly forked process p.
2353 * p is forked by current.
2355 * __sched_fork() is basic setup used by init_idle() too:
2357 static void __sched_fork(struct task_struct
*p
)
2359 p
->se
.exec_start
= 0;
2360 p
->se
.sum_exec_runtime
= 0;
2361 p
->se
.prev_sum_exec_runtime
= 0;
2362 p
->se
.last_wakeup
= 0;
2363 p
->se
.avg_overlap
= 0;
2365 #ifdef CONFIG_SCHEDSTATS
2366 p
->se
.wait_start
= 0;
2367 p
->se
.sum_sleep_runtime
= 0;
2368 p
->se
.sleep_start
= 0;
2369 p
->se
.block_start
= 0;
2370 p
->se
.sleep_max
= 0;
2371 p
->se
.block_max
= 0;
2373 p
->se
.slice_max
= 0;
2377 INIT_LIST_HEAD(&p
->rt
.run_list
);
2379 INIT_LIST_HEAD(&p
->se
.group_node
);
2381 #ifdef CONFIG_PREEMPT_NOTIFIERS
2382 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2386 * We mark the process as running here, but have not actually
2387 * inserted it onto the runqueue yet. This guarantees that
2388 * nobody will actually run it, and a signal or other external
2389 * event cannot wake it up and insert it on the runqueue either.
2391 p
->state
= TASK_RUNNING
;
2395 * fork()/clone()-time setup:
2397 void sched_fork(struct task_struct
*p
, int clone_flags
)
2399 int cpu
= get_cpu();
2404 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2406 set_task_cpu(p
, cpu
);
2409 * Make sure we do not leak PI boosting priority to the child:
2411 p
->prio
= current
->normal_prio
;
2412 if (!rt_prio(p
->prio
))
2413 p
->sched_class
= &fair_sched_class
;
2415 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2416 if (likely(sched_info_on()))
2417 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2419 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2422 #ifdef CONFIG_PREEMPT
2423 /* Want to start with kernel preemption disabled. */
2424 task_thread_info(p
)->preempt_count
= 1;
2430 * wake_up_new_task - wake up a newly created task for the first time.
2432 * This function will do some initial scheduler statistics housekeeping
2433 * that must be done for every newly created context, then puts the task
2434 * on the runqueue and wakes it.
2436 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2438 unsigned long flags
;
2441 rq
= task_rq_lock(p
, &flags
);
2442 BUG_ON(p
->state
!= TASK_RUNNING
);
2443 update_rq_clock(rq
);
2445 p
->prio
= effective_prio(p
);
2447 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2448 activate_task(rq
, p
, 0);
2451 * Let the scheduling class do new task startup
2452 * management (if any):
2454 p
->sched_class
->task_new(rq
, p
);
2457 trace_sched_wakeup_new(rq
, p
);
2458 check_preempt_curr(rq
, p
, 0);
2460 if (p
->sched_class
->task_wake_up
)
2461 p
->sched_class
->task_wake_up(rq
, p
);
2463 task_rq_unlock(rq
, &flags
);
2466 #ifdef CONFIG_PREEMPT_NOTIFIERS
2469 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2470 * @notifier: notifier struct to register
2472 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2474 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2476 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2479 * preempt_notifier_unregister - no longer interested in preemption notifications
2480 * @notifier: notifier struct to unregister
2482 * This is safe to call from within a preemption notifier.
2484 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2486 hlist_del(¬ifier
->link
);
2488 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2490 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2492 struct preempt_notifier
*notifier
;
2493 struct hlist_node
*node
;
2495 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2496 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2500 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2501 struct task_struct
*next
)
2503 struct preempt_notifier
*notifier
;
2504 struct hlist_node
*node
;
2506 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2507 notifier
->ops
->sched_out(notifier
, next
);
2510 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2512 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2517 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2518 struct task_struct
*next
)
2522 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2525 * prepare_task_switch - prepare to switch tasks
2526 * @rq: the runqueue preparing to switch
2527 * @prev: the current task that is being switched out
2528 * @next: the task we are going to switch to.
2530 * This is called with the rq lock held and interrupts off. It must
2531 * be paired with a subsequent finish_task_switch after the context
2534 * prepare_task_switch sets up locking and calls architecture specific
2538 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2539 struct task_struct
*next
)
2541 fire_sched_out_preempt_notifiers(prev
, next
);
2542 prepare_lock_switch(rq
, next
);
2543 prepare_arch_switch(next
);
2547 * finish_task_switch - clean up after a task-switch
2548 * @rq: runqueue associated with task-switch
2549 * @prev: the thread we just switched away from.
2551 * finish_task_switch must be called after the context switch, paired
2552 * with a prepare_task_switch call before the context switch.
2553 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2554 * and do any other architecture-specific cleanup actions.
2556 * Note that we may have delayed dropping an mm in context_switch(). If
2557 * so, we finish that here outside of the runqueue lock. (Doing it
2558 * with the lock held can cause deadlocks; see schedule() for
2561 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2562 __releases(rq
->lock
)
2564 struct mm_struct
*mm
= rq
->prev_mm
;
2570 * A task struct has one reference for the use as "current".
2571 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2572 * schedule one last time. The schedule call will never return, and
2573 * the scheduled task must drop that reference.
2574 * The test for TASK_DEAD must occur while the runqueue locks are
2575 * still held, otherwise prev could be scheduled on another cpu, die
2576 * there before we look at prev->state, and then the reference would
2578 * Manfred Spraul <manfred@colorfullife.com>
2580 prev_state
= prev
->state
;
2581 finish_arch_switch(prev
);
2582 finish_lock_switch(rq
, prev
);
2584 if (current
->sched_class
->post_schedule
)
2585 current
->sched_class
->post_schedule(rq
);
2588 fire_sched_in_preempt_notifiers(current
);
2591 if (unlikely(prev_state
== TASK_DEAD
)) {
2593 * Remove function-return probe instances associated with this
2594 * task and put them back on the free list.
2596 kprobe_flush_task(prev
);
2597 put_task_struct(prev
);
2602 * schedule_tail - first thing a freshly forked thread must call.
2603 * @prev: the thread we just switched away from.
2605 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2606 __releases(rq
->lock
)
2608 struct rq
*rq
= this_rq();
2610 finish_task_switch(rq
, prev
);
2611 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2612 /* In this case, finish_task_switch does not reenable preemption */
2615 if (current
->set_child_tid
)
2616 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2620 * context_switch - switch to the new MM and the new
2621 * thread's register state.
2624 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2625 struct task_struct
*next
)
2627 struct mm_struct
*mm
, *oldmm
;
2629 prepare_task_switch(rq
, prev
, next
);
2630 trace_sched_switch(rq
, prev
, next
);
2632 oldmm
= prev
->active_mm
;
2634 * For paravirt, this is coupled with an exit in switch_to to
2635 * combine the page table reload and the switch backend into
2638 arch_enter_lazy_cpu_mode();
2640 if (unlikely(!mm
)) {
2641 next
->active_mm
= oldmm
;
2642 atomic_inc(&oldmm
->mm_count
);
2643 enter_lazy_tlb(oldmm
, next
);
2645 switch_mm(oldmm
, mm
, next
);
2647 if (unlikely(!prev
->mm
)) {
2648 prev
->active_mm
= NULL
;
2649 rq
->prev_mm
= oldmm
;
2652 * Since the runqueue lock will be released by the next
2653 * task (which is an invalid locking op but in the case
2654 * of the scheduler it's an obvious special-case), so we
2655 * do an early lockdep release here:
2657 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2658 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2661 /* Here we just switch the register state and the stack. */
2662 switch_to(prev
, next
, prev
);
2666 * this_rq must be evaluated again because prev may have moved
2667 * CPUs since it called schedule(), thus the 'rq' on its stack
2668 * frame will be invalid.
2670 finish_task_switch(this_rq(), prev
);
2674 * nr_running, nr_uninterruptible and nr_context_switches:
2676 * externally visible scheduler statistics: current number of runnable
2677 * threads, current number of uninterruptible-sleeping threads, total
2678 * number of context switches performed since bootup.
2680 unsigned long nr_running(void)
2682 unsigned long i
, sum
= 0;
2684 for_each_online_cpu(i
)
2685 sum
+= cpu_rq(i
)->nr_running
;
2690 unsigned long nr_uninterruptible(void)
2692 unsigned long i
, sum
= 0;
2694 for_each_possible_cpu(i
)
2695 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2698 * Since we read the counters lockless, it might be slightly
2699 * inaccurate. Do not allow it to go below zero though:
2701 if (unlikely((long)sum
< 0))
2707 unsigned long long nr_context_switches(void)
2710 unsigned long long sum
= 0;
2712 for_each_possible_cpu(i
)
2713 sum
+= cpu_rq(i
)->nr_switches
;
2718 unsigned long nr_iowait(void)
2720 unsigned long i
, sum
= 0;
2722 for_each_possible_cpu(i
)
2723 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2728 unsigned long nr_active(void)
2730 unsigned long i
, running
= 0, uninterruptible
= 0;
2732 for_each_online_cpu(i
) {
2733 running
+= cpu_rq(i
)->nr_running
;
2734 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2737 if (unlikely((long)uninterruptible
< 0))
2738 uninterruptible
= 0;
2740 return running
+ uninterruptible
;
2744 * Update rq->cpu_load[] statistics. This function is usually called every
2745 * scheduler tick (TICK_NSEC).
2747 static void update_cpu_load(struct rq
*this_rq
)
2749 unsigned long this_load
= this_rq
->load
.weight
;
2752 this_rq
->nr_load_updates
++;
2754 /* Update our load: */
2755 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2756 unsigned long old_load
, new_load
;
2758 /* scale is effectively 1 << i now, and >> i divides by scale */
2760 old_load
= this_rq
->cpu_load
[i
];
2761 new_load
= this_load
;
2763 * Round up the averaging division if load is increasing. This
2764 * prevents us from getting stuck on 9 if the load is 10, for
2767 if (new_load
> old_load
)
2768 new_load
+= scale
-1;
2769 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2776 * double_rq_lock - safely lock two runqueues
2778 * Note this does not disable interrupts like task_rq_lock,
2779 * you need to do so manually before calling.
2781 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2782 __acquires(rq1
->lock
)
2783 __acquires(rq2
->lock
)
2785 BUG_ON(!irqs_disabled());
2787 spin_lock(&rq1
->lock
);
2788 __acquire(rq2
->lock
); /* Fake it out ;) */
2791 spin_lock(&rq1
->lock
);
2792 spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
2794 spin_lock(&rq2
->lock
);
2795 spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
2798 update_rq_clock(rq1
);
2799 update_rq_clock(rq2
);
2803 * double_rq_unlock - safely unlock two runqueues
2805 * Note this does not restore interrupts like task_rq_unlock,
2806 * you need to do so manually after calling.
2808 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2809 __releases(rq1
->lock
)
2810 __releases(rq2
->lock
)
2812 spin_unlock(&rq1
->lock
);
2814 spin_unlock(&rq2
->lock
);
2816 __release(rq2
->lock
);
2820 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2822 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2823 __releases(this_rq
->lock
)
2824 __acquires(busiest
->lock
)
2825 __acquires(this_rq
->lock
)
2829 if (unlikely(!irqs_disabled())) {
2830 /* printk() doesn't work good under rq->lock */
2831 spin_unlock(&this_rq
->lock
);
2834 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2835 if (busiest
< this_rq
) {
2836 spin_unlock(&this_rq
->lock
);
2837 spin_lock(&busiest
->lock
);
2838 spin_lock_nested(&this_rq
->lock
, SINGLE_DEPTH_NESTING
);
2841 spin_lock_nested(&busiest
->lock
, SINGLE_DEPTH_NESTING
);
2846 static void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2847 __releases(busiest
->lock
)
2849 spin_unlock(&busiest
->lock
);
2850 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
2854 * If dest_cpu is allowed for this process, migrate the task to it.
2855 * This is accomplished by forcing the cpu_allowed mask to only
2856 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2857 * the cpu_allowed mask is restored.
2859 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2861 struct migration_req req
;
2862 unsigned long flags
;
2865 rq
= task_rq_lock(p
, &flags
);
2866 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2867 || unlikely(!cpu_active(dest_cpu
)))
2870 trace_sched_migrate_task(rq
, p
, dest_cpu
);
2871 /* force the process onto the specified CPU */
2872 if (migrate_task(p
, dest_cpu
, &req
)) {
2873 /* Need to wait for migration thread (might exit: take ref). */
2874 struct task_struct
*mt
= rq
->migration_thread
;
2876 get_task_struct(mt
);
2877 task_rq_unlock(rq
, &flags
);
2878 wake_up_process(mt
);
2879 put_task_struct(mt
);
2880 wait_for_completion(&req
.done
);
2885 task_rq_unlock(rq
, &flags
);
2889 * sched_exec - execve() is a valuable balancing opportunity, because at
2890 * this point the task has the smallest effective memory and cache footprint.
2892 void sched_exec(void)
2894 int new_cpu
, this_cpu
= get_cpu();
2895 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2897 if (new_cpu
!= this_cpu
)
2898 sched_migrate_task(current
, new_cpu
);
2902 * pull_task - move a task from a remote runqueue to the local runqueue.
2903 * Both runqueues must be locked.
2905 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2906 struct rq
*this_rq
, int this_cpu
)
2908 deactivate_task(src_rq
, p
, 0);
2909 set_task_cpu(p
, this_cpu
);
2910 activate_task(this_rq
, p
, 0);
2912 * Note that idle threads have a prio of MAX_PRIO, for this test
2913 * to be always true for them.
2915 check_preempt_curr(this_rq
, p
, 0);
2919 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2922 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2923 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2927 * We do not migrate tasks that are:
2928 * 1) running (obviously), or
2929 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2930 * 3) are cache-hot on their current CPU.
2932 if (!cpu_isset(this_cpu
, p
->cpus_allowed
)) {
2933 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2938 if (task_running(rq
, p
)) {
2939 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2944 * Aggressive migration if:
2945 * 1) task is cache cold, or
2946 * 2) too many balance attempts have failed.
2949 if (!task_hot(p
, rq
->clock
, sd
) ||
2950 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2951 #ifdef CONFIG_SCHEDSTATS
2952 if (task_hot(p
, rq
->clock
, sd
)) {
2953 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2954 schedstat_inc(p
, se
.nr_forced_migrations
);
2960 if (task_hot(p
, rq
->clock
, sd
)) {
2961 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
2967 static unsigned long
2968 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2969 unsigned long max_load_move
, struct sched_domain
*sd
,
2970 enum cpu_idle_type idle
, int *all_pinned
,
2971 int *this_best_prio
, struct rq_iterator
*iterator
)
2973 int loops
= 0, pulled
= 0, pinned
= 0;
2974 struct task_struct
*p
;
2975 long rem_load_move
= max_load_move
;
2977 if (max_load_move
== 0)
2983 * Start the load-balancing iterator:
2985 p
= iterator
->start(iterator
->arg
);
2987 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
2990 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
2991 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2992 p
= iterator
->next(iterator
->arg
);
2996 pull_task(busiest
, p
, this_rq
, this_cpu
);
2998 rem_load_move
-= p
->se
.load
.weight
;
3001 * We only want to steal up to the prescribed amount of weighted load.
3003 if (rem_load_move
> 0) {
3004 if (p
->prio
< *this_best_prio
)
3005 *this_best_prio
= p
->prio
;
3006 p
= iterator
->next(iterator
->arg
);
3011 * Right now, this is one of only two places pull_task() is called,
3012 * so we can safely collect pull_task() stats here rather than
3013 * inside pull_task().
3015 schedstat_add(sd
, lb_gained
[idle
], pulled
);
3018 *all_pinned
= pinned
;
3020 return max_load_move
- rem_load_move
;
3024 * move_tasks tries to move up to max_load_move weighted load from busiest to
3025 * this_rq, as part of a balancing operation within domain "sd".
3026 * Returns 1 if successful and 0 otherwise.
3028 * Called with both runqueues locked.
3030 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3031 unsigned long max_load_move
,
3032 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3035 const struct sched_class
*class = sched_class_highest
;
3036 unsigned long total_load_moved
= 0;
3037 int this_best_prio
= this_rq
->curr
->prio
;
3041 class->load_balance(this_rq
, this_cpu
, busiest
,
3042 max_load_move
- total_load_moved
,
3043 sd
, idle
, all_pinned
, &this_best_prio
);
3044 class = class->next
;
3046 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3049 } while (class && max_load_move
> total_load_moved
);
3051 return total_load_moved
> 0;
3055 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3056 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3057 struct rq_iterator
*iterator
)
3059 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3063 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3064 pull_task(busiest
, p
, this_rq
, this_cpu
);
3066 * Right now, this is only the second place pull_task()
3067 * is called, so we can safely collect pull_task()
3068 * stats here rather than inside pull_task().
3070 schedstat_inc(sd
, lb_gained
[idle
]);
3074 p
= iterator
->next(iterator
->arg
);
3081 * move_one_task tries to move exactly one task from busiest to this_rq, as
3082 * part of active balancing operations within "domain".
3083 * Returns 1 if successful and 0 otherwise.
3085 * Called with both runqueues locked.
3087 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3088 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3090 const struct sched_class
*class;
3092 for (class = sched_class_highest
; class; class = class->next
)
3093 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3100 * find_busiest_group finds and returns the busiest CPU group within the
3101 * domain. It calculates and returns the amount of weighted load which
3102 * should be moved to restore balance via the imbalance parameter.
3104 static struct sched_group
*
3105 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
3106 unsigned long *imbalance
, enum cpu_idle_type idle
,
3107 int *sd_idle
, const cpumask_t
*cpus
, int *balance
)
3109 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
3110 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
3111 unsigned long max_pull
;
3112 unsigned long busiest_load_per_task
, busiest_nr_running
;
3113 unsigned long this_load_per_task
, this_nr_running
;
3114 int load_idx
, group_imb
= 0;
3115 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3116 int power_savings_balance
= 1;
3117 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
3118 unsigned long min_nr_running
= ULONG_MAX
;
3119 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
3122 max_load
= this_load
= total_load
= total_pwr
= 0;
3123 busiest_load_per_task
= busiest_nr_running
= 0;
3124 this_load_per_task
= this_nr_running
= 0;
3126 if (idle
== CPU_NOT_IDLE
)
3127 load_idx
= sd
->busy_idx
;
3128 else if (idle
== CPU_NEWLY_IDLE
)
3129 load_idx
= sd
->newidle_idx
;
3131 load_idx
= sd
->idle_idx
;
3134 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
3137 int __group_imb
= 0;
3138 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3139 unsigned long sum_nr_running
, sum_weighted_load
;
3140 unsigned long sum_avg_load_per_task
;
3141 unsigned long avg_load_per_task
;
3143 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
3146 balance_cpu
= first_cpu(group
->cpumask
);
3148 /* Tally up the load of all CPUs in the group */
3149 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
3150 sum_avg_load_per_task
= avg_load_per_task
= 0;
3153 min_cpu_load
= ~0UL;
3155 for_each_cpu_mask_nr(i
, group
->cpumask
) {
3158 if (!cpu_isset(i
, *cpus
))
3163 if (*sd_idle
&& rq
->nr_running
)
3166 /* Bias balancing toward cpus of our domain */
3168 if (idle_cpu(i
) && !first_idle_cpu
) {
3173 load
= target_load(i
, load_idx
);
3175 load
= source_load(i
, load_idx
);
3176 if (load
> max_cpu_load
)
3177 max_cpu_load
= load
;
3178 if (min_cpu_load
> load
)
3179 min_cpu_load
= load
;
3183 sum_nr_running
+= rq
->nr_running
;
3184 sum_weighted_load
+= weighted_cpuload(i
);
3186 sum_avg_load_per_task
+= cpu_avg_load_per_task(i
);
3190 * First idle cpu or the first cpu(busiest) in this sched group
3191 * is eligible for doing load balancing at this and above
3192 * domains. In the newly idle case, we will allow all the cpu's
3193 * to do the newly idle load balance.
3195 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3196 balance_cpu
!= this_cpu
&& balance
) {
3201 total_load
+= avg_load
;
3202 total_pwr
+= group
->__cpu_power
;
3204 /* Adjust by relative CPU power of the group */
3205 avg_load
= sg_div_cpu_power(group
,
3206 avg_load
* SCHED_LOAD_SCALE
);
3210 * Consider the group unbalanced when the imbalance is larger
3211 * than the average weight of two tasks.
3213 * APZ: with cgroup the avg task weight can vary wildly and
3214 * might not be a suitable number - should we keep a
3215 * normalized nr_running number somewhere that negates
3218 avg_load_per_task
= sg_div_cpu_power(group
,
3219 sum_avg_load_per_task
* SCHED_LOAD_SCALE
);
3221 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3224 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
3227 this_load
= avg_load
;
3229 this_nr_running
= sum_nr_running
;
3230 this_load_per_task
= sum_weighted_load
;
3231 } else if (avg_load
> max_load
&&
3232 (sum_nr_running
> group_capacity
|| __group_imb
)) {
3233 max_load
= avg_load
;
3235 busiest_nr_running
= sum_nr_running
;
3236 busiest_load_per_task
= sum_weighted_load
;
3237 group_imb
= __group_imb
;
3240 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3242 * Busy processors will not participate in power savings
3245 if (idle
== CPU_NOT_IDLE
||
3246 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3250 * If the local group is idle or completely loaded
3251 * no need to do power savings balance at this domain
3253 if (local_group
&& (this_nr_running
>= group_capacity
||
3255 power_savings_balance
= 0;
3258 * If a group is already running at full capacity or idle,
3259 * don't include that group in power savings calculations
3261 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
3266 * Calculate the group which has the least non-idle load.
3267 * This is the group from where we need to pick up the load
3270 if ((sum_nr_running
< min_nr_running
) ||
3271 (sum_nr_running
== min_nr_running
&&
3272 first_cpu(group
->cpumask
) <
3273 first_cpu(group_min
->cpumask
))) {
3275 min_nr_running
= sum_nr_running
;
3276 min_load_per_task
= sum_weighted_load
/
3281 * Calculate the group which is almost near its
3282 * capacity but still has some space to pick up some load
3283 * from other group and save more power
3285 if (sum_nr_running
<= group_capacity
- 1) {
3286 if (sum_nr_running
> leader_nr_running
||
3287 (sum_nr_running
== leader_nr_running
&&
3288 first_cpu(group
->cpumask
) >
3289 first_cpu(group_leader
->cpumask
))) {
3290 group_leader
= group
;
3291 leader_nr_running
= sum_nr_running
;
3296 group
= group
->next
;
3297 } while (group
!= sd
->groups
);
3299 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
3302 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
3304 if (this_load
>= avg_load
||
3305 100*max_load
<= sd
->imbalance_pct
*this_load
)
3308 busiest_load_per_task
/= busiest_nr_running
;
3310 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
3313 * We're trying to get all the cpus to the average_load, so we don't
3314 * want to push ourselves above the average load, nor do we wish to
3315 * reduce the max loaded cpu below the average load, as either of these
3316 * actions would just result in more rebalancing later, and ping-pong
3317 * tasks around. Thus we look for the minimum possible imbalance.
3318 * Negative imbalances (*we* are more loaded than anyone else) will
3319 * be counted as no imbalance for these purposes -- we can't fix that
3320 * by pulling tasks to us. Be careful of negative numbers as they'll
3321 * appear as very large values with unsigned longs.
3323 if (max_load
<= busiest_load_per_task
)
3327 * In the presence of smp nice balancing, certain scenarios can have
3328 * max load less than avg load(as we skip the groups at or below
3329 * its cpu_power, while calculating max_load..)
3331 if (max_load
< avg_load
) {
3333 goto small_imbalance
;
3336 /* Don't want to pull so many tasks that a group would go idle */
3337 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
3339 /* How much load to actually move to equalise the imbalance */
3340 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
3341 (avg_load
- this_load
) * this->__cpu_power
)
3345 * if *imbalance is less than the average load per runnable task
3346 * there is no gaurantee that any tasks will be moved so we'll have
3347 * a think about bumping its value to force at least one task to be
3350 if (*imbalance
< busiest_load_per_task
) {
3351 unsigned long tmp
, pwr_now
, pwr_move
;
3355 pwr_move
= pwr_now
= 0;
3357 if (this_nr_running
) {
3358 this_load_per_task
/= this_nr_running
;
3359 if (busiest_load_per_task
> this_load_per_task
)
3362 this_load_per_task
= cpu_avg_load_per_task(this_cpu
);
3364 if (max_load
- this_load
+ busiest_load_per_task
>=
3365 busiest_load_per_task
* imbn
) {
3366 *imbalance
= busiest_load_per_task
;
3371 * OK, we don't have enough imbalance to justify moving tasks,
3372 * however we may be able to increase total CPU power used by
3376 pwr_now
+= busiest
->__cpu_power
*
3377 min(busiest_load_per_task
, max_load
);
3378 pwr_now
+= this->__cpu_power
*
3379 min(this_load_per_task
, this_load
);
3380 pwr_now
/= SCHED_LOAD_SCALE
;
3382 /* Amount of load we'd subtract */
3383 tmp
= sg_div_cpu_power(busiest
,
3384 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3386 pwr_move
+= busiest
->__cpu_power
*
3387 min(busiest_load_per_task
, max_load
- tmp
);
3389 /* Amount of load we'd add */
3390 if (max_load
* busiest
->__cpu_power
<
3391 busiest_load_per_task
* SCHED_LOAD_SCALE
)
3392 tmp
= sg_div_cpu_power(this,
3393 max_load
* busiest
->__cpu_power
);
3395 tmp
= sg_div_cpu_power(this,
3396 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3397 pwr_move
+= this->__cpu_power
*
3398 min(this_load_per_task
, this_load
+ tmp
);
3399 pwr_move
/= SCHED_LOAD_SCALE
;
3401 /* Move if we gain throughput */
3402 if (pwr_move
> pwr_now
)
3403 *imbalance
= busiest_load_per_task
;
3409 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3410 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3413 if (this == group_leader
&& group_leader
!= group_min
) {
3414 *imbalance
= min_load_per_task
;
3424 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3427 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
3428 unsigned long imbalance
, const cpumask_t
*cpus
)
3430 struct rq
*busiest
= NULL
, *rq
;
3431 unsigned long max_load
= 0;
3434 for_each_cpu_mask_nr(i
, group
->cpumask
) {
3437 if (!cpu_isset(i
, *cpus
))
3441 wl
= weighted_cpuload(i
);
3443 if (rq
->nr_running
== 1 && wl
> imbalance
)
3446 if (wl
> max_load
) {
3456 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3457 * so long as it is large enough.
3459 #define MAX_PINNED_INTERVAL 512
3462 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3463 * tasks if there is an imbalance.
3465 static int load_balance(int this_cpu
, struct rq
*this_rq
,
3466 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3467 int *balance
, cpumask_t
*cpus
)
3469 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
3470 struct sched_group
*group
;
3471 unsigned long imbalance
;
3473 unsigned long flags
;
3478 * When power savings policy is enabled for the parent domain, idle
3479 * sibling can pick up load irrespective of busy siblings. In this case,
3480 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3481 * portraying it as CPU_NOT_IDLE.
3483 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3484 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3487 schedstat_inc(sd
, lb_count
[idle
]);
3491 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3498 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3502 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
3504 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3508 BUG_ON(busiest
== this_rq
);
3510 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3513 if (busiest
->nr_running
> 1) {
3515 * Attempt to move tasks. If find_busiest_group has found
3516 * an imbalance but busiest->nr_running <= 1, the group is
3517 * still unbalanced. ld_moved simply stays zero, so it is
3518 * correctly treated as an imbalance.
3520 local_irq_save(flags
);
3521 double_rq_lock(this_rq
, busiest
);
3522 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3523 imbalance
, sd
, idle
, &all_pinned
);
3524 double_rq_unlock(this_rq
, busiest
);
3525 local_irq_restore(flags
);
3528 * some other cpu did the load balance for us.
3530 if (ld_moved
&& this_cpu
!= smp_processor_id())
3531 resched_cpu(this_cpu
);
3533 /* All tasks on this runqueue were pinned by CPU affinity */
3534 if (unlikely(all_pinned
)) {
3535 cpu_clear(cpu_of(busiest
), *cpus
);
3536 if (!cpus_empty(*cpus
))
3543 schedstat_inc(sd
, lb_failed
[idle
]);
3544 sd
->nr_balance_failed
++;
3546 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3548 spin_lock_irqsave(&busiest
->lock
, flags
);
3550 /* don't kick the migration_thread, if the curr
3551 * task on busiest cpu can't be moved to this_cpu
3553 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
3554 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3556 goto out_one_pinned
;
3559 if (!busiest
->active_balance
) {
3560 busiest
->active_balance
= 1;
3561 busiest
->push_cpu
= this_cpu
;
3564 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3566 wake_up_process(busiest
->migration_thread
);
3569 * We've kicked active balancing, reset the failure
3572 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3575 sd
->nr_balance_failed
= 0;
3577 if (likely(!active_balance
)) {
3578 /* We were unbalanced, so reset the balancing interval */
3579 sd
->balance_interval
= sd
->min_interval
;
3582 * If we've begun active balancing, start to back off. This
3583 * case may not be covered by the all_pinned logic if there
3584 * is only 1 task on the busy runqueue (because we don't call
3587 if (sd
->balance_interval
< sd
->max_interval
)
3588 sd
->balance_interval
*= 2;
3591 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3592 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3598 schedstat_inc(sd
, lb_balanced
[idle
]);
3600 sd
->nr_balance_failed
= 0;
3603 /* tune up the balancing interval */
3604 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3605 (sd
->balance_interval
< sd
->max_interval
))
3606 sd
->balance_interval
*= 2;
3608 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3609 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3620 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3621 * tasks if there is an imbalance.
3623 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3624 * this_rq is locked.
3627 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
,
3630 struct sched_group
*group
;
3631 struct rq
*busiest
= NULL
;
3632 unsigned long imbalance
;
3640 * When power savings policy is enabled for the parent domain, idle
3641 * sibling can pick up load irrespective of busy siblings. In this case,
3642 * let the state of idle sibling percolate up as IDLE, instead of
3643 * portraying it as CPU_NOT_IDLE.
3645 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
3646 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3649 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
3651 update_shares_locked(this_rq
, sd
);
3652 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
3653 &sd_idle
, cpus
, NULL
);
3655 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
3659 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
3661 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
3665 BUG_ON(busiest
== this_rq
);
3667 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
3670 if (busiest
->nr_running
> 1) {
3671 /* Attempt to move tasks */
3672 double_lock_balance(this_rq
, busiest
);
3673 /* this_rq->clock is already updated */
3674 update_rq_clock(busiest
);
3675 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3676 imbalance
, sd
, CPU_NEWLY_IDLE
,
3678 double_unlock_balance(this_rq
, busiest
);
3680 if (unlikely(all_pinned
)) {
3681 cpu_clear(cpu_of(busiest
), *cpus
);
3682 if (!cpus_empty(*cpus
))
3688 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
3689 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3690 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3693 sd
->nr_balance_failed
= 0;
3695 update_shares_locked(this_rq
, sd
);
3699 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
3700 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3701 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3703 sd
->nr_balance_failed
= 0;
3709 * idle_balance is called by schedule() if this_cpu is about to become
3710 * idle. Attempts to pull tasks from other CPUs.
3712 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
3714 struct sched_domain
*sd
;
3715 int pulled_task
= -1;
3716 unsigned long next_balance
= jiffies
+ HZ
;
3719 for_each_domain(this_cpu
, sd
) {
3720 unsigned long interval
;
3722 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3725 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
3726 /* If we've pulled tasks over stop searching: */
3727 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
3730 interval
= msecs_to_jiffies(sd
->balance_interval
);
3731 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3732 next_balance
= sd
->last_balance
+ interval
;
3736 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3738 * We are going idle. next_balance may be set based on
3739 * a busy processor. So reset next_balance.
3741 this_rq
->next_balance
= next_balance
;
3746 * active_load_balance is run by migration threads. It pushes running tasks
3747 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3748 * running on each physical CPU where possible, and avoids physical /
3749 * logical imbalances.
3751 * Called with busiest_rq locked.
3753 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
3755 int target_cpu
= busiest_rq
->push_cpu
;
3756 struct sched_domain
*sd
;
3757 struct rq
*target_rq
;
3759 /* Is there any task to move? */
3760 if (busiest_rq
->nr_running
<= 1)
3763 target_rq
= cpu_rq(target_cpu
);
3766 * This condition is "impossible", if it occurs
3767 * we need to fix it. Originally reported by
3768 * Bjorn Helgaas on a 128-cpu setup.
3770 BUG_ON(busiest_rq
== target_rq
);
3772 /* move a task from busiest_rq to target_rq */
3773 double_lock_balance(busiest_rq
, target_rq
);
3774 update_rq_clock(busiest_rq
);
3775 update_rq_clock(target_rq
);
3777 /* Search for an sd spanning us and the target CPU. */
3778 for_each_domain(target_cpu
, sd
) {
3779 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3780 cpu_isset(busiest_cpu
, sd
->span
))
3785 schedstat_inc(sd
, alb_count
);
3787 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3789 schedstat_inc(sd
, alb_pushed
);
3791 schedstat_inc(sd
, alb_failed
);
3793 double_unlock_balance(busiest_rq
, target_rq
);
3798 atomic_t load_balancer
;
3800 } nohz ____cacheline_aligned
= {
3801 .load_balancer
= ATOMIC_INIT(-1),
3802 .cpu_mask
= CPU_MASK_NONE
,
3806 * This routine will try to nominate the ilb (idle load balancing)
3807 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3808 * load balancing on behalf of all those cpus. If all the cpus in the system
3809 * go into this tickless mode, then there will be no ilb owner (as there is
3810 * no need for one) and all the cpus will sleep till the next wakeup event
3813 * For the ilb owner, tick is not stopped. And this tick will be used
3814 * for idle load balancing. ilb owner will still be part of
3817 * While stopping the tick, this cpu will become the ilb owner if there
3818 * is no other owner. And will be the owner till that cpu becomes busy
3819 * or if all cpus in the system stop their ticks at which point
3820 * there is no need for ilb owner.
3822 * When the ilb owner becomes busy, it nominates another owner, during the
3823 * next busy scheduler_tick()
3825 int select_nohz_load_balancer(int stop_tick
)
3827 int cpu
= smp_processor_id();
3830 cpu_set(cpu
, nohz
.cpu_mask
);
3831 cpu_rq(cpu
)->in_nohz_recently
= 1;
3834 * If we are going offline and still the leader, give up!
3836 if (!cpu_active(cpu
) &&
3837 atomic_read(&nohz
.load_balancer
) == cpu
) {
3838 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3843 /* time for ilb owner also to sleep */
3844 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3845 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3846 atomic_set(&nohz
.load_balancer
, -1);
3850 if (atomic_read(&nohz
.load_balancer
) == -1) {
3851 /* make me the ilb owner */
3852 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3854 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3857 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3860 cpu_clear(cpu
, nohz
.cpu_mask
);
3862 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3863 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3870 static DEFINE_SPINLOCK(balancing
);
3873 * It checks each scheduling domain to see if it is due to be balanced,
3874 * and initiates a balancing operation if so.
3876 * Balancing parameters are set up in arch_init_sched_domains.
3878 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3881 struct rq
*rq
= cpu_rq(cpu
);
3882 unsigned long interval
;
3883 struct sched_domain
*sd
;
3884 /* Earliest time when we have to do rebalance again */
3885 unsigned long next_balance
= jiffies
+ 60*HZ
;
3886 int update_next_balance
= 0;
3890 for_each_domain(cpu
, sd
) {
3891 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3894 interval
= sd
->balance_interval
;
3895 if (idle
!= CPU_IDLE
)
3896 interval
*= sd
->busy_factor
;
3898 /* scale ms to jiffies */
3899 interval
= msecs_to_jiffies(interval
);
3900 if (unlikely(!interval
))
3902 if (interval
> HZ
*NR_CPUS
/10)
3903 interval
= HZ
*NR_CPUS
/10;
3905 need_serialize
= sd
->flags
& SD_SERIALIZE
;
3907 if (need_serialize
) {
3908 if (!spin_trylock(&balancing
))
3912 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3913 if (load_balance(cpu
, rq
, sd
, idle
, &balance
, &tmp
)) {
3915 * We've pulled tasks over so either we're no
3916 * longer idle, or one of our SMT siblings is
3919 idle
= CPU_NOT_IDLE
;
3921 sd
->last_balance
= jiffies
;
3924 spin_unlock(&balancing
);
3926 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3927 next_balance
= sd
->last_balance
+ interval
;
3928 update_next_balance
= 1;
3932 * Stop the load balance at this level. There is another
3933 * CPU in our sched group which is doing load balancing more
3941 * next_balance will be updated only when there is a need.
3942 * When the cpu is attached to null domain for ex, it will not be
3945 if (likely(update_next_balance
))
3946 rq
->next_balance
= next_balance
;
3950 * run_rebalance_domains is triggered when needed from the scheduler tick.
3951 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3952 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3954 static void run_rebalance_domains(struct softirq_action
*h
)
3956 int this_cpu
= smp_processor_id();
3957 struct rq
*this_rq
= cpu_rq(this_cpu
);
3958 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3959 CPU_IDLE
: CPU_NOT_IDLE
;
3961 rebalance_domains(this_cpu
, idle
);
3965 * If this cpu is the owner for idle load balancing, then do the
3966 * balancing on behalf of the other idle cpus whose ticks are
3969 if (this_rq
->idle_at_tick
&&
3970 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3971 cpumask_t cpus
= nohz
.cpu_mask
;
3975 cpu_clear(this_cpu
, cpus
);
3976 for_each_cpu_mask_nr(balance_cpu
, cpus
) {
3978 * If this cpu gets work to do, stop the load balancing
3979 * work being done for other cpus. Next load
3980 * balancing owner will pick it up.
3985 rebalance_domains(balance_cpu
, CPU_IDLE
);
3987 rq
= cpu_rq(balance_cpu
);
3988 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3989 this_rq
->next_balance
= rq
->next_balance
;
3996 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3998 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3999 * idle load balancing owner or decide to stop the periodic load balancing,
4000 * if the whole system is idle.
4002 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
4006 * If we were in the nohz mode recently and busy at the current
4007 * scheduler tick, then check if we need to nominate new idle
4010 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
4011 rq
->in_nohz_recently
= 0;
4013 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4014 cpu_clear(cpu
, nohz
.cpu_mask
);
4015 atomic_set(&nohz
.load_balancer
, -1);
4018 if (atomic_read(&nohz
.load_balancer
) == -1) {
4020 * simple selection for now: Nominate the
4021 * first cpu in the nohz list to be the next
4024 * TBD: Traverse the sched domains and nominate
4025 * the nearest cpu in the nohz.cpu_mask.
4027 int ilb
= first_cpu(nohz
.cpu_mask
);
4029 if (ilb
< nr_cpu_ids
)
4035 * If this cpu is idle and doing idle load balancing for all the
4036 * cpus with ticks stopped, is it time for that to stop?
4038 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4039 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4045 * If this cpu is idle and the idle load balancing is done by
4046 * someone else, then no need raise the SCHED_SOFTIRQ
4048 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4049 cpu_isset(cpu
, nohz
.cpu_mask
))
4052 if (time_after_eq(jiffies
, rq
->next_balance
))
4053 raise_softirq(SCHED_SOFTIRQ
);
4056 #else /* CONFIG_SMP */
4059 * on UP we do not need to balance between CPUs:
4061 static inline void idle_balance(int cpu
, struct rq
*rq
)
4067 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4069 EXPORT_PER_CPU_SYMBOL(kstat
);
4072 * Return any ns on the sched_clock that have not yet been banked in
4073 * @p in case that task is currently running.
4075 unsigned long long task_delta_exec(struct task_struct
*p
)
4077 unsigned long flags
;
4081 rq
= task_rq_lock(p
, &flags
);
4083 if (task_current(rq
, p
)) {
4086 update_rq_clock(rq
);
4087 delta_exec
= rq
->clock
- p
->se
.exec_start
;
4088 if ((s64
)delta_exec
> 0)
4092 task_rq_unlock(rq
, &flags
);
4098 * Account user cpu time to a process.
4099 * @p: the process that the cpu time gets accounted to
4100 * @cputime: the cpu time spent in user space since the last update
4102 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
4104 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4107 p
->utime
= cputime_add(p
->utime
, cputime
);
4108 account_group_user_time(p
, cputime
);
4110 /* Add user time to cpustat. */
4111 tmp
= cputime_to_cputime64(cputime
);
4112 if (TASK_NICE(p
) > 0)
4113 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
4115 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4116 /* Account for user time used */
4117 acct_update_integrals(p
);
4121 * Account guest cpu time to a process.
4122 * @p: the process that the cpu time gets accounted to
4123 * @cputime: the cpu time spent in virtual machine since the last update
4125 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
)
4128 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4130 tmp
= cputime_to_cputime64(cputime
);
4132 p
->utime
= cputime_add(p
->utime
, cputime
);
4133 account_group_user_time(p
, cputime
);
4134 p
->gtime
= cputime_add(p
->gtime
, cputime
);
4136 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4137 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
4141 * Account scaled user cpu time to a process.
4142 * @p: the process that the cpu time gets accounted to
4143 * @cputime: the cpu time spent in user space since the last update
4145 void account_user_time_scaled(struct task_struct
*p
, cputime_t cputime
)
4147 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime
);
4151 * Account system cpu time to a process.
4152 * @p: the process that the cpu time gets accounted to
4153 * @hardirq_offset: the offset to subtract from hardirq_count()
4154 * @cputime: the cpu time spent in kernel space since the last update
4156 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
4159 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4160 struct rq
*rq
= this_rq();
4163 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
4164 account_guest_time(p
, cputime
);
4168 p
->stime
= cputime_add(p
->stime
, cputime
);
4169 account_group_system_time(p
, cputime
);
4171 /* Add system time to cpustat. */
4172 tmp
= cputime_to_cputime64(cputime
);
4173 if (hardirq_count() - hardirq_offset
)
4174 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
4175 else if (softirq_count())
4176 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
4177 else if (p
!= rq
->idle
)
4178 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
4179 else if (atomic_read(&rq
->nr_iowait
) > 0)
4180 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
4182 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
4183 /* Account for system time used */
4184 acct_update_integrals(p
);
4188 * Account scaled system cpu time to a process.
4189 * @p: the process that the cpu time gets accounted to
4190 * @hardirq_offset: the offset to subtract from hardirq_count()
4191 * @cputime: the cpu time spent in kernel space since the last update
4193 void account_system_time_scaled(struct task_struct
*p
, cputime_t cputime
)
4195 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime
);
4199 * Account for involuntary wait time.
4200 * @p: the process from which the cpu time has been stolen
4201 * @steal: the cpu time spent in involuntary wait
4203 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
4205 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4206 cputime64_t tmp
= cputime_to_cputime64(steal
);
4207 struct rq
*rq
= this_rq();
4209 if (p
== rq
->idle
) {
4210 p
->stime
= cputime_add(p
->stime
, steal
);
4211 account_group_system_time(p
, steal
);
4212 if (atomic_read(&rq
->nr_iowait
) > 0)
4213 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
4215 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
4217 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
4221 * Use precise platform statistics if available:
4223 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4224 cputime_t
task_utime(struct task_struct
*p
)
4229 cputime_t
task_stime(struct task_struct
*p
)
4234 cputime_t
task_utime(struct task_struct
*p
)
4236 clock_t utime
= cputime_to_clock_t(p
->utime
),
4237 total
= utime
+ cputime_to_clock_t(p
->stime
);
4241 * Use CFS's precise accounting:
4243 temp
= (u64
)nsec_to_clock_t(p
->se
.sum_exec_runtime
);
4247 do_div(temp
, total
);
4249 utime
= (clock_t)temp
;
4251 p
->prev_utime
= max(p
->prev_utime
, clock_t_to_cputime(utime
));
4252 return p
->prev_utime
;
4255 cputime_t
task_stime(struct task_struct
*p
)
4260 * Use CFS's precise accounting. (we subtract utime from
4261 * the total, to make sure the total observed by userspace
4262 * grows monotonically - apps rely on that):
4264 stime
= nsec_to_clock_t(p
->se
.sum_exec_runtime
) -
4265 cputime_to_clock_t(task_utime(p
));
4268 p
->prev_stime
= max(p
->prev_stime
, clock_t_to_cputime(stime
));
4270 return p
->prev_stime
;
4274 inline cputime_t
task_gtime(struct task_struct
*p
)
4280 * This function gets called by the timer code, with HZ frequency.
4281 * We call it with interrupts disabled.
4283 * It also gets called by the fork code, when changing the parent's
4286 void scheduler_tick(void)
4288 int cpu
= smp_processor_id();
4289 struct rq
*rq
= cpu_rq(cpu
);
4290 struct task_struct
*curr
= rq
->curr
;
4294 spin_lock(&rq
->lock
);
4295 update_rq_clock(rq
);
4296 update_cpu_load(rq
);
4297 curr
->sched_class
->task_tick(rq
, curr
, 0);
4298 spin_unlock(&rq
->lock
);
4301 rq
->idle_at_tick
= idle_cpu(cpu
);
4302 trigger_load_balance(rq
, cpu
);
4306 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4307 defined(CONFIG_PREEMPT_TRACER))
4309 static inline unsigned long get_parent_ip(unsigned long addr
)
4311 if (in_lock_functions(addr
)) {
4312 addr
= CALLER_ADDR2
;
4313 if (in_lock_functions(addr
))
4314 addr
= CALLER_ADDR3
;
4319 void __kprobes
add_preempt_count(int val
)
4321 #ifdef CONFIG_DEBUG_PREEMPT
4325 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4328 preempt_count() += val
;
4329 #ifdef CONFIG_DEBUG_PREEMPT
4331 * Spinlock count overflowing soon?
4333 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4336 if (preempt_count() == val
)
4337 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4339 EXPORT_SYMBOL(add_preempt_count
);
4341 void __kprobes
sub_preempt_count(int val
)
4343 #ifdef CONFIG_DEBUG_PREEMPT
4347 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
4350 * Is the spinlock portion underflowing?
4352 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4353 !(preempt_count() & PREEMPT_MASK
)))
4357 if (preempt_count() == val
)
4358 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4359 preempt_count() -= val
;
4361 EXPORT_SYMBOL(sub_preempt_count
);
4366 * Print scheduling while atomic bug:
4368 static noinline
void __schedule_bug(struct task_struct
*prev
)
4370 struct pt_regs
*regs
= get_irq_regs();
4372 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4373 prev
->comm
, prev
->pid
, preempt_count());
4375 debug_show_held_locks(prev
);
4377 if (irqs_disabled())
4378 print_irqtrace_events(prev
);
4387 * Various schedule()-time debugging checks and statistics:
4389 static inline void schedule_debug(struct task_struct
*prev
)
4392 * Test if we are atomic. Since do_exit() needs to call into
4393 * schedule() atomically, we ignore that path for now.
4394 * Otherwise, whine if we are scheduling when we should not be.
4396 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
4397 __schedule_bug(prev
);
4399 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4401 schedstat_inc(this_rq(), sched_count
);
4402 #ifdef CONFIG_SCHEDSTATS
4403 if (unlikely(prev
->lock_depth
>= 0)) {
4404 schedstat_inc(this_rq(), bkl_count
);
4405 schedstat_inc(prev
, sched_info
.bkl_count
);
4411 * Pick up the highest-prio task:
4413 static inline struct task_struct
*
4414 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
4416 const struct sched_class
*class;
4417 struct task_struct
*p
;
4420 * Optimization: we know that if all tasks are in
4421 * the fair class we can call that function directly:
4423 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
4424 p
= fair_sched_class
.pick_next_task(rq
);
4429 class = sched_class_highest
;
4431 p
= class->pick_next_task(rq
);
4435 * Will never be NULL as the idle class always
4436 * returns a non-NULL p:
4438 class = class->next
;
4443 * schedule() is the main scheduler function.
4445 asmlinkage
void __sched
schedule(void)
4447 struct task_struct
*prev
, *next
;
4448 unsigned long *switch_count
;
4454 cpu
= smp_processor_id();
4458 switch_count
= &prev
->nivcsw
;
4460 release_kernel_lock(prev
);
4461 need_resched_nonpreemptible
:
4463 schedule_debug(prev
);
4465 if (sched_feat(HRTICK
))
4468 spin_lock_irq(&rq
->lock
);
4469 update_rq_clock(rq
);
4470 clear_tsk_need_resched(prev
);
4472 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
4473 if (unlikely(signal_pending_state(prev
->state
, prev
)))
4474 prev
->state
= TASK_RUNNING
;
4476 deactivate_task(rq
, prev
, 1);
4477 switch_count
= &prev
->nvcsw
;
4481 if (prev
->sched_class
->pre_schedule
)
4482 prev
->sched_class
->pre_schedule(rq
, prev
);
4485 if (unlikely(!rq
->nr_running
))
4486 idle_balance(cpu
, rq
);
4488 prev
->sched_class
->put_prev_task(rq
, prev
);
4489 next
= pick_next_task(rq
, prev
);
4491 if (likely(prev
!= next
)) {
4492 sched_info_switch(prev
, next
);
4498 context_switch(rq
, prev
, next
); /* unlocks the rq */
4500 * the context switch might have flipped the stack from under
4501 * us, hence refresh the local variables.
4503 cpu
= smp_processor_id();
4506 spin_unlock_irq(&rq
->lock
);
4508 if (unlikely(reacquire_kernel_lock(current
) < 0))
4509 goto need_resched_nonpreemptible
;
4511 preempt_enable_no_resched();
4512 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
4515 EXPORT_SYMBOL(schedule
);
4517 #ifdef CONFIG_PREEMPT
4519 * this is the entry point to schedule() from in-kernel preemption
4520 * off of preempt_enable. Kernel preemptions off return from interrupt
4521 * occur there and call schedule directly.
4523 asmlinkage
void __sched
preempt_schedule(void)
4525 struct thread_info
*ti
= current_thread_info();
4528 * If there is a non-zero preempt_count or interrupts are disabled,
4529 * we do not want to preempt the current task. Just return..
4531 if (likely(ti
->preempt_count
|| irqs_disabled()))
4535 add_preempt_count(PREEMPT_ACTIVE
);
4537 sub_preempt_count(PREEMPT_ACTIVE
);
4540 * Check again in case we missed a preemption opportunity
4541 * between schedule and now.
4544 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4546 EXPORT_SYMBOL(preempt_schedule
);
4549 * this is the entry point to schedule() from kernel preemption
4550 * off of irq context.
4551 * Note, that this is called and return with irqs disabled. This will
4552 * protect us against recursive calling from irq.
4554 asmlinkage
void __sched
preempt_schedule_irq(void)
4556 struct thread_info
*ti
= current_thread_info();
4558 /* Catch callers which need to be fixed */
4559 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
4562 add_preempt_count(PREEMPT_ACTIVE
);
4565 local_irq_disable();
4566 sub_preempt_count(PREEMPT_ACTIVE
);
4569 * Check again in case we missed a preemption opportunity
4570 * between schedule and now.
4573 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4576 #endif /* CONFIG_PREEMPT */
4578 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
4581 return try_to_wake_up(curr
->private, mode
, sync
);
4583 EXPORT_SYMBOL(default_wake_function
);
4586 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4587 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4588 * number) then we wake all the non-exclusive tasks and one exclusive task.
4590 * There are circumstances in which we can try to wake a task which has already
4591 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4592 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4594 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4595 int nr_exclusive
, int sync
, void *key
)
4597 wait_queue_t
*curr
, *next
;
4599 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4600 unsigned flags
= curr
->flags
;
4602 if (curr
->func(curr
, mode
, sync
, key
) &&
4603 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4609 * __wake_up - wake up threads blocked on a waitqueue.
4611 * @mode: which threads
4612 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4613 * @key: is directly passed to the wakeup function
4615 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4616 int nr_exclusive
, void *key
)
4618 unsigned long flags
;
4620 spin_lock_irqsave(&q
->lock
, flags
);
4621 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4622 spin_unlock_irqrestore(&q
->lock
, flags
);
4624 EXPORT_SYMBOL(__wake_up
);
4627 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4629 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4631 __wake_up_common(q
, mode
, 1, 0, NULL
);
4635 * __wake_up_sync - wake up threads blocked on a waitqueue.
4637 * @mode: which threads
4638 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4640 * The sync wakeup differs that the waker knows that it will schedule
4641 * away soon, so while the target thread will be woken up, it will not
4642 * be migrated to another CPU - ie. the two threads are 'synchronized'
4643 * with each other. This can prevent needless bouncing between CPUs.
4645 * On UP it can prevent extra preemption.
4648 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4650 unsigned long flags
;
4656 if (unlikely(!nr_exclusive
))
4659 spin_lock_irqsave(&q
->lock
, flags
);
4660 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
4661 spin_unlock_irqrestore(&q
->lock
, flags
);
4663 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4666 * complete: - signals a single thread waiting on this completion
4667 * @x: holds the state of this particular completion
4669 * This will wake up a single thread waiting on this completion. Threads will be
4670 * awakened in the same order in which they were queued.
4672 * See also complete_all(), wait_for_completion() and related routines.
4674 void complete(struct completion
*x
)
4676 unsigned long flags
;
4678 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4680 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4681 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4683 EXPORT_SYMBOL(complete
);
4686 * complete_all: - signals all threads waiting on this completion
4687 * @x: holds the state of this particular completion
4689 * This will wake up all threads waiting on this particular completion event.
4691 void complete_all(struct completion
*x
)
4693 unsigned long flags
;
4695 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4696 x
->done
+= UINT_MAX
/2;
4697 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4698 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4700 EXPORT_SYMBOL(complete_all
);
4702 static inline long __sched
4703 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4706 DECLARE_WAITQUEUE(wait
, current
);
4708 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4709 __add_wait_queue_tail(&x
->wait
, &wait
);
4711 if (signal_pending_state(state
, current
)) {
4712 timeout
= -ERESTARTSYS
;
4715 __set_current_state(state
);
4716 spin_unlock_irq(&x
->wait
.lock
);
4717 timeout
= schedule_timeout(timeout
);
4718 spin_lock_irq(&x
->wait
.lock
);
4719 } while (!x
->done
&& timeout
);
4720 __remove_wait_queue(&x
->wait
, &wait
);
4725 return timeout
?: 1;
4729 wait_for_common(struct completion
*x
, long timeout
, int state
)
4733 spin_lock_irq(&x
->wait
.lock
);
4734 timeout
= do_wait_for_common(x
, timeout
, state
);
4735 spin_unlock_irq(&x
->wait
.lock
);
4740 * wait_for_completion: - waits for completion of a task
4741 * @x: holds the state of this particular completion
4743 * This waits to be signaled for completion of a specific task. It is NOT
4744 * interruptible and there is no timeout.
4746 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4747 * and interrupt capability. Also see complete().
4749 void __sched
wait_for_completion(struct completion
*x
)
4751 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4753 EXPORT_SYMBOL(wait_for_completion
);
4756 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4757 * @x: holds the state of this particular completion
4758 * @timeout: timeout value in jiffies
4760 * This waits for either a completion of a specific task to be signaled or for a
4761 * specified timeout to expire. The timeout is in jiffies. It is not
4764 unsigned long __sched
4765 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4767 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4769 EXPORT_SYMBOL(wait_for_completion_timeout
);
4772 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4773 * @x: holds the state of this particular completion
4775 * This waits for completion of a specific task to be signaled. It is
4778 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4780 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4781 if (t
== -ERESTARTSYS
)
4785 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4788 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4789 * @x: holds the state of this particular completion
4790 * @timeout: timeout value in jiffies
4792 * This waits for either a completion of a specific task to be signaled or for a
4793 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4795 unsigned long __sched
4796 wait_for_completion_interruptible_timeout(struct completion
*x
,
4797 unsigned long timeout
)
4799 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4801 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4804 * wait_for_completion_killable: - waits for completion of a task (killable)
4805 * @x: holds the state of this particular completion
4807 * This waits to be signaled for completion of a specific task. It can be
4808 * interrupted by a kill signal.
4810 int __sched
wait_for_completion_killable(struct completion
*x
)
4812 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4813 if (t
== -ERESTARTSYS
)
4817 EXPORT_SYMBOL(wait_for_completion_killable
);
4820 * try_wait_for_completion - try to decrement a completion without blocking
4821 * @x: completion structure
4823 * Returns: 0 if a decrement cannot be done without blocking
4824 * 1 if a decrement succeeded.
4826 * If a completion is being used as a counting completion,
4827 * attempt to decrement the counter without blocking. This
4828 * enables us to avoid waiting if the resource the completion
4829 * is protecting is not available.
4831 bool try_wait_for_completion(struct completion
*x
)
4835 spin_lock_irq(&x
->wait
.lock
);
4840 spin_unlock_irq(&x
->wait
.lock
);
4843 EXPORT_SYMBOL(try_wait_for_completion
);
4846 * completion_done - Test to see if a completion has any waiters
4847 * @x: completion structure
4849 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4850 * 1 if there are no waiters.
4853 bool completion_done(struct completion
*x
)
4857 spin_lock_irq(&x
->wait
.lock
);
4860 spin_unlock_irq(&x
->wait
.lock
);
4863 EXPORT_SYMBOL(completion_done
);
4866 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4868 unsigned long flags
;
4871 init_waitqueue_entry(&wait
, current
);
4873 __set_current_state(state
);
4875 spin_lock_irqsave(&q
->lock
, flags
);
4876 __add_wait_queue(q
, &wait
);
4877 spin_unlock(&q
->lock
);
4878 timeout
= schedule_timeout(timeout
);
4879 spin_lock_irq(&q
->lock
);
4880 __remove_wait_queue(q
, &wait
);
4881 spin_unlock_irqrestore(&q
->lock
, flags
);
4886 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4888 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4890 EXPORT_SYMBOL(interruptible_sleep_on
);
4893 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4895 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4897 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4899 void __sched
sleep_on(wait_queue_head_t
*q
)
4901 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4903 EXPORT_SYMBOL(sleep_on
);
4905 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4907 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4909 EXPORT_SYMBOL(sleep_on_timeout
);
4911 #ifdef CONFIG_RT_MUTEXES
4914 * rt_mutex_setprio - set the current priority of a task
4916 * @prio: prio value (kernel-internal form)
4918 * This function changes the 'effective' priority of a task. It does
4919 * not touch ->normal_prio like __setscheduler().
4921 * Used by the rt_mutex code to implement priority inheritance logic.
4923 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4925 unsigned long flags
;
4926 int oldprio
, on_rq
, running
;
4928 const struct sched_class
*prev_class
= p
->sched_class
;
4930 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4932 rq
= task_rq_lock(p
, &flags
);
4933 update_rq_clock(rq
);
4936 on_rq
= p
->se
.on_rq
;
4937 running
= task_current(rq
, p
);
4939 dequeue_task(rq
, p
, 0);
4941 p
->sched_class
->put_prev_task(rq
, p
);
4944 p
->sched_class
= &rt_sched_class
;
4946 p
->sched_class
= &fair_sched_class
;
4951 p
->sched_class
->set_curr_task(rq
);
4953 enqueue_task(rq
, p
, 0);
4955 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4957 task_rq_unlock(rq
, &flags
);
4962 void set_user_nice(struct task_struct
*p
, long nice
)
4964 int old_prio
, delta
, on_rq
;
4965 unsigned long flags
;
4968 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4971 * We have to be careful, if called from sys_setpriority(),
4972 * the task might be in the middle of scheduling on another CPU.
4974 rq
= task_rq_lock(p
, &flags
);
4975 update_rq_clock(rq
);
4977 * The RT priorities are set via sched_setscheduler(), but we still
4978 * allow the 'normal' nice value to be set - but as expected
4979 * it wont have any effect on scheduling until the task is
4980 * SCHED_FIFO/SCHED_RR:
4982 if (task_has_rt_policy(p
)) {
4983 p
->static_prio
= NICE_TO_PRIO(nice
);
4986 on_rq
= p
->se
.on_rq
;
4988 dequeue_task(rq
, p
, 0);
4990 p
->static_prio
= NICE_TO_PRIO(nice
);
4993 p
->prio
= effective_prio(p
);
4994 delta
= p
->prio
- old_prio
;
4997 enqueue_task(rq
, p
, 0);
4999 * If the task increased its priority or is running and
5000 * lowered its priority, then reschedule its CPU:
5002 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
5003 resched_task(rq
->curr
);
5006 task_rq_unlock(rq
, &flags
);
5008 EXPORT_SYMBOL(set_user_nice
);
5011 * can_nice - check if a task can reduce its nice value
5015 int can_nice(const struct task_struct
*p
, const int nice
)
5017 /* convert nice value [19,-20] to rlimit style value [1,40] */
5018 int nice_rlim
= 20 - nice
;
5020 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
5021 capable(CAP_SYS_NICE
));
5024 #ifdef __ARCH_WANT_SYS_NICE
5027 * sys_nice - change the priority of the current process.
5028 * @increment: priority increment
5030 * sys_setpriority is a more generic, but much slower function that
5031 * does similar things.
5033 asmlinkage
long sys_nice(int increment
)
5038 * Setpriority might change our priority at the same moment.
5039 * We don't have to worry. Conceptually one call occurs first
5040 * and we have a single winner.
5042 if (increment
< -40)
5047 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
5053 if (increment
< 0 && !can_nice(current
, nice
))
5056 retval
= security_task_setnice(current
, nice
);
5060 set_user_nice(current
, nice
);
5067 * task_prio - return the priority value of a given task.
5068 * @p: the task in question.
5070 * This is the priority value as seen by users in /proc.
5071 * RT tasks are offset by -200. Normal tasks are centered
5072 * around 0, value goes from -16 to +15.
5074 int task_prio(const struct task_struct
*p
)
5076 return p
->prio
- MAX_RT_PRIO
;
5080 * task_nice - return the nice value of a given task.
5081 * @p: the task in question.
5083 int task_nice(const struct task_struct
*p
)
5085 return TASK_NICE(p
);
5087 EXPORT_SYMBOL(task_nice
);
5090 * idle_cpu - is a given cpu idle currently?
5091 * @cpu: the processor in question.
5093 int idle_cpu(int cpu
)
5095 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
5099 * idle_task - return the idle task for a given cpu.
5100 * @cpu: the processor in question.
5102 struct task_struct
*idle_task(int cpu
)
5104 return cpu_rq(cpu
)->idle
;
5108 * find_process_by_pid - find a process with a matching PID value.
5109 * @pid: the pid in question.
5111 static struct task_struct
*find_process_by_pid(pid_t pid
)
5113 return pid
? find_task_by_vpid(pid
) : current
;
5116 /* Actually do priority change: must hold rq lock. */
5118 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
5120 BUG_ON(p
->se
.on_rq
);
5123 switch (p
->policy
) {
5127 p
->sched_class
= &fair_sched_class
;
5131 p
->sched_class
= &rt_sched_class
;
5135 p
->rt_priority
= prio
;
5136 p
->normal_prio
= normal_prio(p
);
5137 /* we are holding p->pi_lock already */
5138 p
->prio
= rt_mutex_getprio(p
);
5142 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
5143 struct sched_param
*param
, bool user
)
5145 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
5146 unsigned long flags
;
5147 const struct sched_class
*prev_class
= p
->sched_class
;
5150 /* may grab non-irq protected spin_locks */
5151 BUG_ON(in_interrupt());
5153 /* double check policy once rq lock held */
5155 policy
= oldpolicy
= p
->policy
;
5156 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
5157 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
5158 policy
!= SCHED_IDLE
)
5161 * Valid priorities for SCHED_FIFO and SCHED_RR are
5162 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5163 * SCHED_BATCH and SCHED_IDLE is 0.
5165 if (param
->sched_priority
< 0 ||
5166 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
5167 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
5169 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
5173 * Allow unprivileged RT tasks to decrease priority:
5175 if (user
&& !capable(CAP_SYS_NICE
)) {
5176 if (rt_policy(policy
)) {
5177 unsigned long rlim_rtprio
;
5179 if (!lock_task_sighand(p
, &flags
))
5181 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
5182 unlock_task_sighand(p
, &flags
);
5184 /* can't set/change the rt policy */
5185 if (policy
!= p
->policy
&& !rlim_rtprio
)
5188 /* can't increase priority */
5189 if (param
->sched_priority
> p
->rt_priority
&&
5190 param
->sched_priority
> rlim_rtprio
)
5194 * Like positive nice levels, dont allow tasks to
5195 * move out of SCHED_IDLE either:
5197 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
5200 /* can't change other user's priorities */
5201 if ((current
->euid
!= p
->euid
) &&
5202 (current
->euid
!= p
->uid
))
5207 #ifdef CONFIG_RT_GROUP_SCHED
5209 * Do not allow realtime tasks into groups that have no runtime
5212 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
5213 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
5217 retval
= security_task_setscheduler(p
, policy
, param
);
5223 * make sure no PI-waiters arrive (or leave) while we are
5224 * changing the priority of the task:
5226 spin_lock_irqsave(&p
->pi_lock
, flags
);
5228 * To be able to change p->policy safely, the apropriate
5229 * runqueue lock must be held.
5231 rq
= __task_rq_lock(p
);
5232 /* recheck policy now with rq lock held */
5233 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
5234 policy
= oldpolicy
= -1;
5235 __task_rq_unlock(rq
);
5236 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5239 update_rq_clock(rq
);
5240 on_rq
= p
->se
.on_rq
;
5241 running
= task_current(rq
, p
);
5243 deactivate_task(rq
, p
, 0);
5245 p
->sched_class
->put_prev_task(rq
, p
);
5248 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
5251 p
->sched_class
->set_curr_task(rq
);
5253 activate_task(rq
, p
, 0);
5255 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5257 __task_rq_unlock(rq
);
5258 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5260 rt_mutex_adjust_pi(p
);
5266 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5267 * @p: the task in question.
5268 * @policy: new policy.
5269 * @param: structure containing the new RT priority.
5271 * NOTE that the task may be already dead.
5273 int sched_setscheduler(struct task_struct
*p
, int policy
,
5274 struct sched_param
*param
)
5276 return __sched_setscheduler(p
, policy
, param
, true);
5278 EXPORT_SYMBOL_GPL(sched_setscheduler
);
5281 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5282 * @p: the task in question.
5283 * @policy: new policy.
5284 * @param: structure containing the new RT priority.
5286 * Just like sched_setscheduler, only don't bother checking if the
5287 * current context has permission. For example, this is needed in
5288 * stop_machine(): we create temporary high priority worker threads,
5289 * but our caller might not have that capability.
5291 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
5292 struct sched_param
*param
)
5294 return __sched_setscheduler(p
, policy
, param
, false);
5298 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5300 struct sched_param lparam
;
5301 struct task_struct
*p
;
5304 if (!param
|| pid
< 0)
5306 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
5311 p
= find_process_by_pid(pid
);
5313 retval
= sched_setscheduler(p
, policy
, &lparam
);
5320 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5321 * @pid: the pid in question.
5322 * @policy: new policy.
5323 * @param: structure containing the new RT priority.
5326 sys_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5328 /* negative values for policy are not valid */
5332 return do_sched_setscheduler(pid
, policy
, param
);
5336 * sys_sched_setparam - set/change the RT priority of a thread
5337 * @pid: the pid in question.
5338 * @param: structure containing the new RT priority.
5340 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
5342 return do_sched_setscheduler(pid
, -1, param
);
5346 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5347 * @pid: the pid in question.
5349 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
5351 struct task_struct
*p
;
5358 read_lock(&tasklist_lock
);
5359 p
= find_process_by_pid(pid
);
5361 retval
= security_task_getscheduler(p
);
5365 read_unlock(&tasklist_lock
);
5370 * sys_sched_getscheduler - get the RT priority of a thread
5371 * @pid: the pid in question.
5372 * @param: structure containing the RT priority.
5374 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
5376 struct sched_param lp
;
5377 struct task_struct
*p
;
5380 if (!param
|| pid
< 0)
5383 read_lock(&tasklist_lock
);
5384 p
= find_process_by_pid(pid
);
5389 retval
= security_task_getscheduler(p
);
5393 lp
.sched_priority
= p
->rt_priority
;
5394 read_unlock(&tasklist_lock
);
5397 * This one might sleep, we cannot do it with a spinlock held ...
5399 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
5404 read_unlock(&tasklist_lock
);
5408 long sched_setaffinity(pid_t pid
, const cpumask_t
*in_mask
)
5410 cpumask_t cpus_allowed
;
5411 cpumask_t new_mask
= *in_mask
;
5412 struct task_struct
*p
;
5416 read_lock(&tasklist_lock
);
5418 p
= find_process_by_pid(pid
);
5420 read_unlock(&tasklist_lock
);
5426 * It is not safe to call set_cpus_allowed with the
5427 * tasklist_lock held. We will bump the task_struct's
5428 * usage count and then drop tasklist_lock.
5431 read_unlock(&tasklist_lock
);
5434 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
5435 !capable(CAP_SYS_NICE
))
5438 retval
= security_task_setscheduler(p
, 0, NULL
);
5442 cpuset_cpus_allowed(p
, &cpus_allowed
);
5443 cpus_and(new_mask
, new_mask
, cpus_allowed
);
5445 retval
= set_cpus_allowed_ptr(p
, &new_mask
);
5448 cpuset_cpus_allowed(p
, &cpus_allowed
);
5449 if (!cpus_subset(new_mask
, cpus_allowed
)) {
5451 * We must have raced with a concurrent cpuset
5452 * update. Just reset the cpus_allowed to the
5453 * cpuset's cpus_allowed
5455 new_mask
= cpus_allowed
;
5465 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5466 cpumask_t
*new_mask
)
5468 if (len
< sizeof(cpumask_t
)) {
5469 memset(new_mask
, 0, sizeof(cpumask_t
));
5470 } else if (len
> sizeof(cpumask_t
)) {
5471 len
= sizeof(cpumask_t
);
5473 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
5477 * sys_sched_setaffinity - set the cpu affinity of a process
5478 * @pid: pid of the process
5479 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5480 * @user_mask_ptr: user-space pointer to the new cpu mask
5482 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
5483 unsigned long __user
*user_mask_ptr
)
5488 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
5492 return sched_setaffinity(pid
, &new_mask
);
5495 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
5497 struct task_struct
*p
;
5501 read_lock(&tasklist_lock
);
5504 p
= find_process_by_pid(pid
);
5508 retval
= security_task_getscheduler(p
);
5512 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
5515 read_unlock(&tasklist_lock
);
5522 * sys_sched_getaffinity - get the cpu affinity of a process
5523 * @pid: pid of the process
5524 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5525 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5527 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
5528 unsigned long __user
*user_mask_ptr
)
5533 if (len
< sizeof(cpumask_t
))
5536 ret
= sched_getaffinity(pid
, &mask
);
5540 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
5543 return sizeof(cpumask_t
);
5547 * sys_sched_yield - yield the current processor to other threads.
5549 * This function yields the current CPU to other tasks. If there are no
5550 * other threads running on this CPU then this function will return.
5552 asmlinkage
long sys_sched_yield(void)
5554 struct rq
*rq
= this_rq_lock();
5556 schedstat_inc(rq
, yld_count
);
5557 current
->sched_class
->yield_task(rq
);
5560 * Since we are going to call schedule() anyway, there's
5561 * no need to preempt or enable interrupts:
5563 __release(rq
->lock
);
5564 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5565 _raw_spin_unlock(&rq
->lock
);
5566 preempt_enable_no_resched();
5573 static void __cond_resched(void)
5575 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5576 __might_sleep(__FILE__
, __LINE__
);
5579 * The BKS might be reacquired before we have dropped
5580 * PREEMPT_ACTIVE, which could trigger a second
5581 * cond_resched() call.
5584 add_preempt_count(PREEMPT_ACTIVE
);
5586 sub_preempt_count(PREEMPT_ACTIVE
);
5587 } while (need_resched());
5590 int __sched
_cond_resched(void)
5592 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
5593 system_state
== SYSTEM_RUNNING
) {
5599 EXPORT_SYMBOL(_cond_resched
);
5602 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5603 * call schedule, and on return reacquire the lock.
5605 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5606 * operations here to prevent schedule() from being called twice (once via
5607 * spin_unlock(), once by hand).
5609 int cond_resched_lock(spinlock_t
*lock
)
5611 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
5614 if (spin_needbreak(lock
) || resched
) {
5616 if (resched
&& need_resched())
5625 EXPORT_SYMBOL(cond_resched_lock
);
5627 int __sched
cond_resched_softirq(void)
5629 BUG_ON(!in_softirq());
5631 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
5639 EXPORT_SYMBOL(cond_resched_softirq
);
5642 * yield - yield the current processor to other threads.
5644 * This is a shortcut for kernel-space yielding - it marks the
5645 * thread runnable and calls sys_sched_yield().
5647 void __sched
yield(void)
5649 set_current_state(TASK_RUNNING
);
5652 EXPORT_SYMBOL(yield
);
5655 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5656 * that process accounting knows that this is a task in IO wait state.
5658 * But don't do that if it is a deliberate, throttling IO wait (this task
5659 * has set its backing_dev_info: the queue against which it should throttle)
5661 void __sched
io_schedule(void)
5663 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5665 delayacct_blkio_start();
5666 atomic_inc(&rq
->nr_iowait
);
5668 atomic_dec(&rq
->nr_iowait
);
5669 delayacct_blkio_end();
5671 EXPORT_SYMBOL(io_schedule
);
5673 long __sched
io_schedule_timeout(long timeout
)
5675 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5678 delayacct_blkio_start();
5679 atomic_inc(&rq
->nr_iowait
);
5680 ret
= schedule_timeout(timeout
);
5681 atomic_dec(&rq
->nr_iowait
);
5682 delayacct_blkio_end();
5687 * sys_sched_get_priority_max - return maximum RT priority.
5688 * @policy: scheduling class.
5690 * this syscall returns the maximum rt_priority that can be used
5691 * by a given scheduling class.
5693 asmlinkage
long sys_sched_get_priority_max(int policy
)
5700 ret
= MAX_USER_RT_PRIO
-1;
5712 * sys_sched_get_priority_min - return minimum RT priority.
5713 * @policy: scheduling class.
5715 * this syscall returns the minimum rt_priority that can be used
5716 * by a given scheduling class.
5718 asmlinkage
long sys_sched_get_priority_min(int policy
)
5736 * sys_sched_rr_get_interval - return the default timeslice of a process.
5737 * @pid: pid of the process.
5738 * @interval: userspace pointer to the timeslice value.
5740 * this syscall writes the default timeslice value of a given process
5741 * into the user-space timespec buffer. A value of '0' means infinity.
5744 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
5746 struct task_struct
*p
;
5747 unsigned int time_slice
;
5755 read_lock(&tasklist_lock
);
5756 p
= find_process_by_pid(pid
);
5760 retval
= security_task_getscheduler(p
);
5765 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5766 * tasks that are on an otherwise idle runqueue:
5769 if (p
->policy
== SCHED_RR
) {
5770 time_slice
= DEF_TIMESLICE
;
5771 } else if (p
->policy
!= SCHED_FIFO
) {
5772 struct sched_entity
*se
= &p
->se
;
5773 unsigned long flags
;
5776 rq
= task_rq_lock(p
, &flags
);
5777 if (rq
->cfs
.load
.weight
)
5778 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
5779 task_rq_unlock(rq
, &flags
);
5781 read_unlock(&tasklist_lock
);
5782 jiffies_to_timespec(time_slice
, &t
);
5783 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5787 read_unlock(&tasklist_lock
);
5791 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5793 void sched_show_task(struct task_struct
*p
)
5795 unsigned long free
= 0;
5798 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5799 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5800 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5801 #if BITS_PER_LONG == 32
5802 if (state
== TASK_RUNNING
)
5803 printk(KERN_CONT
" running ");
5805 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5807 if (state
== TASK_RUNNING
)
5808 printk(KERN_CONT
" running task ");
5810 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5812 #ifdef CONFIG_DEBUG_STACK_USAGE
5814 unsigned long *n
= end_of_stack(p
);
5817 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
5820 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
5821 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
5823 show_stack(p
, NULL
);
5826 void show_state_filter(unsigned long state_filter
)
5828 struct task_struct
*g
, *p
;
5830 #if BITS_PER_LONG == 32
5832 " task PC stack pid father\n");
5835 " task PC stack pid father\n");
5837 read_lock(&tasklist_lock
);
5838 do_each_thread(g
, p
) {
5840 * reset the NMI-timeout, listing all files on a slow
5841 * console might take alot of time:
5843 touch_nmi_watchdog();
5844 if (!state_filter
|| (p
->state
& state_filter
))
5846 } while_each_thread(g
, p
);
5848 touch_all_softlockup_watchdogs();
5850 #ifdef CONFIG_SCHED_DEBUG
5851 sysrq_sched_debug_show();
5853 read_unlock(&tasklist_lock
);
5855 * Only show locks if all tasks are dumped:
5857 if (state_filter
== -1)
5858 debug_show_all_locks();
5861 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5863 idle
->sched_class
= &idle_sched_class
;
5867 * init_idle - set up an idle thread for a given CPU
5868 * @idle: task in question
5869 * @cpu: cpu the idle task belongs to
5871 * NOTE: this function does not set the idle thread's NEED_RESCHED
5872 * flag, to make booting more robust.
5874 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5876 struct rq
*rq
= cpu_rq(cpu
);
5877 unsigned long flags
;
5879 spin_lock_irqsave(&rq
->lock
, flags
);
5882 idle
->se
.exec_start
= sched_clock();
5884 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
5885 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
5886 __set_task_cpu(idle
, cpu
);
5888 rq
->curr
= rq
->idle
= idle
;
5889 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5892 spin_unlock_irqrestore(&rq
->lock
, flags
);
5894 /* Set the preempt count _outside_ the spinlocks! */
5895 #if defined(CONFIG_PREEMPT)
5896 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5898 task_thread_info(idle
)->preempt_count
= 0;
5901 * The idle tasks have their own, simple scheduling class:
5903 idle
->sched_class
= &idle_sched_class
;
5904 #ifdef CONFIG_FUNCTION_RET_TRACER
5905 ftrace_retfunc_init_task(idle
);
5910 * In a system that switches off the HZ timer nohz_cpu_mask
5911 * indicates which cpus entered this state. This is used
5912 * in the rcu update to wait only for active cpus. For system
5913 * which do not switch off the HZ timer nohz_cpu_mask should
5914 * always be CPU_MASK_NONE.
5916 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
5919 * Increase the granularity value when there are more CPUs,
5920 * because with more CPUs the 'effective latency' as visible
5921 * to users decreases. But the relationship is not linear,
5922 * so pick a second-best guess by going with the log2 of the
5925 * This idea comes from the SD scheduler of Con Kolivas:
5927 static inline void sched_init_granularity(void)
5929 unsigned int factor
= 1 + ilog2(num_online_cpus());
5930 const unsigned long limit
= 200000000;
5932 sysctl_sched_min_granularity
*= factor
;
5933 if (sysctl_sched_min_granularity
> limit
)
5934 sysctl_sched_min_granularity
= limit
;
5936 sysctl_sched_latency
*= factor
;
5937 if (sysctl_sched_latency
> limit
)
5938 sysctl_sched_latency
= limit
;
5940 sysctl_sched_wakeup_granularity
*= factor
;
5942 sysctl_sched_shares_ratelimit
*= factor
;
5947 * This is how migration works:
5949 * 1) we queue a struct migration_req structure in the source CPU's
5950 * runqueue and wake up that CPU's migration thread.
5951 * 2) we down() the locked semaphore => thread blocks.
5952 * 3) migration thread wakes up (implicitly it forces the migrated
5953 * thread off the CPU)
5954 * 4) it gets the migration request and checks whether the migrated
5955 * task is still in the wrong runqueue.
5956 * 5) if it's in the wrong runqueue then the migration thread removes
5957 * it and puts it into the right queue.
5958 * 6) migration thread up()s the semaphore.
5959 * 7) we wake up and the migration is done.
5963 * Change a given task's CPU affinity. Migrate the thread to a
5964 * proper CPU and schedule it away if the CPU it's executing on
5965 * is removed from the allowed bitmask.
5967 * NOTE: the caller must have a valid reference to the task, the
5968 * task must not exit() & deallocate itself prematurely. The
5969 * call is not atomic; no spinlocks may be held.
5971 int set_cpus_allowed_ptr(struct task_struct
*p
, const cpumask_t
*new_mask
)
5973 struct migration_req req
;
5974 unsigned long flags
;
5978 rq
= task_rq_lock(p
, &flags
);
5979 if (!cpus_intersects(*new_mask
, cpu_online_map
)) {
5984 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
5985 !cpus_equal(p
->cpus_allowed
, *new_mask
))) {
5990 if (p
->sched_class
->set_cpus_allowed
)
5991 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
5993 p
->cpus_allowed
= *new_mask
;
5994 p
->rt
.nr_cpus_allowed
= cpus_weight(*new_mask
);
5997 /* Can the task run on the task's current CPU? If so, we're done */
5998 if (cpu_isset(task_cpu(p
), *new_mask
))
6001 if (migrate_task(p
, any_online_cpu(*new_mask
), &req
)) {
6002 /* Need help from migration thread: drop lock and wait. */
6003 task_rq_unlock(rq
, &flags
);
6004 wake_up_process(rq
->migration_thread
);
6005 wait_for_completion(&req
.done
);
6006 tlb_migrate_finish(p
->mm
);
6010 task_rq_unlock(rq
, &flags
);
6014 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
6017 * Move (not current) task off this cpu, onto dest cpu. We're doing
6018 * this because either it can't run here any more (set_cpus_allowed()
6019 * away from this CPU, or CPU going down), or because we're
6020 * attempting to rebalance this task on exec (sched_exec).
6022 * So we race with normal scheduler movements, but that's OK, as long
6023 * as the task is no longer on this CPU.
6025 * Returns non-zero if task was successfully migrated.
6027 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6029 struct rq
*rq_dest
, *rq_src
;
6032 if (unlikely(!cpu_active(dest_cpu
)))
6035 rq_src
= cpu_rq(src_cpu
);
6036 rq_dest
= cpu_rq(dest_cpu
);
6038 double_rq_lock(rq_src
, rq_dest
);
6039 /* Already moved. */
6040 if (task_cpu(p
) != src_cpu
)
6042 /* Affinity changed (again). */
6043 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
6046 on_rq
= p
->se
.on_rq
;
6048 deactivate_task(rq_src
, p
, 0);
6050 set_task_cpu(p
, dest_cpu
);
6052 activate_task(rq_dest
, p
, 0);
6053 check_preempt_curr(rq_dest
, p
, 0);
6058 double_rq_unlock(rq_src
, rq_dest
);
6063 * migration_thread - this is a highprio system thread that performs
6064 * thread migration by bumping thread off CPU then 'pushing' onto
6067 static int migration_thread(void *data
)
6069 int cpu
= (long)data
;
6073 BUG_ON(rq
->migration_thread
!= current
);
6075 set_current_state(TASK_INTERRUPTIBLE
);
6076 while (!kthread_should_stop()) {
6077 struct migration_req
*req
;
6078 struct list_head
*head
;
6080 spin_lock_irq(&rq
->lock
);
6082 if (cpu_is_offline(cpu
)) {
6083 spin_unlock_irq(&rq
->lock
);
6087 if (rq
->active_balance
) {
6088 active_load_balance(rq
, cpu
);
6089 rq
->active_balance
= 0;
6092 head
= &rq
->migration_queue
;
6094 if (list_empty(head
)) {
6095 spin_unlock_irq(&rq
->lock
);
6097 set_current_state(TASK_INTERRUPTIBLE
);
6100 req
= list_entry(head
->next
, struct migration_req
, list
);
6101 list_del_init(head
->next
);
6103 spin_unlock(&rq
->lock
);
6104 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
6107 complete(&req
->done
);
6109 __set_current_state(TASK_RUNNING
);
6113 /* Wait for kthread_stop */
6114 set_current_state(TASK_INTERRUPTIBLE
);
6115 while (!kthread_should_stop()) {
6117 set_current_state(TASK_INTERRUPTIBLE
);
6119 __set_current_state(TASK_RUNNING
);
6123 #ifdef CONFIG_HOTPLUG_CPU
6125 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6129 local_irq_disable();
6130 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
6136 * Figure out where task on dead CPU should go, use force if necessary.
6137 * NOTE: interrupts should be disabled by the caller
6139 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
6141 unsigned long flags
;
6148 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
6149 cpus_and(mask
, mask
, p
->cpus_allowed
);
6150 dest_cpu
= any_online_cpu(mask
);
6152 /* On any allowed CPU? */
6153 if (dest_cpu
>= nr_cpu_ids
)
6154 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
6156 /* No more Mr. Nice Guy. */
6157 if (dest_cpu
>= nr_cpu_ids
) {
6158 cpumask_t cpus_allowed
;
6160 cpuset_cpus_allowed_locked(p
, &cpus_allowed
);
6162 * Try to stay on the same cpuset, where the
6163 * current cpuset may be a subset of all cpus.
6164 * The cpuset_cpus_allowed_locked() variant of
6165 * cpuset_cpus_allowed() will not block. It must be
6166 * called within calls to cpuset_lock/cpuset_unlock.
6168 rq
= task_rq_lock(p
, &flags
);
6169 p
->cpus_allowed
= cpus_allowed
;
6170 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
6171 task_rq_unlock(rq
, &flags
);
6174 * Don't tell them about moving exiting tasks or
6175 * kernel threads (both mm NULL), since they never
6178 if (p
->mm
&& printk_ratelimit()) {
6179 printk(KERN_INFO
"process %d (%s) no "
6180 "longer affine to cpu%d\n",
6181 task_pid_nr(p
), p
->comm
, dead_cpu
);
6184 } while (!__migrate_task_irq(p
, dead_cpu
, dest_cpu
));
6188 * While a dead CPU has no uninterruptible tasks queued at this point,
6189 * it might still have a nonzero ->nr_uninterruptible counter, because
6190 * for performance reasons the counter is not stricly tracking tasks to
6191 * their home CPUs. So we just add the counter to another CPU's counter,
6192 * to keep the global sum constant after CPU-down:
6194 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
6196 struct rq
*rq_dest
= cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR
));
6197 unsigned long flags
;
6199 local_irq_save(flags
);
6200 double_rq_lock(rq_src
, rq_dest
);
6201 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
6202 rq_src
->nr_uninterruptible
= 0;
6203 double_rq_unlock(rq_src
, rq_dest
);
6204 local_irq_restore(flags
);
6207 /* Run through task list and migrate tasks from the dead cpu. */
6208 static void migrate_live_tasks(int src_cpu
)
6210 struct task_struct
*p
, *t
;
6212 read_lock(&tasklist_lock
);
6214 do_each_thread(t
, p
) {
6218 if (task_cpu(p
) == src_cpu
)
6219 move_task_off_dead_cpu(src_cpu
, p
);
6220 } while_each_thread(t
, p
);
6222 read_unlock(&tasklist_lock
);
6226 * Schedules idle task to be the next runnable task on current CPU.
6227 * It does so by boosting its priority to highest possible.
6228 * Used by CPU offline code.
6230 void sched_idle_next(void)
6232 int this_cpu
= smp_processor_id();
6233 struct rq
*rq
= cpu_rq(this_cpu
);
6234 struct task_struct
*p
= rq
->idle
;
6235 unsigned long flags
;
6237 /* cpu has to be offline */
6238 BUG_ON(cpu_online(this_cpu
));
6241 * Strictly not necessary since rest of the CPUs are stopped by now
6242 * and interrupts disabled on the current cpu.
6244 spin_lock_irqsave(&rq
->lock
, flags
);
6246 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6248 update_rq_clock(rq
);
6249 activate_task(rq
, p
, 0);
6251 spin_unlock_irqrestore(&rq
->lock
, flags
);
6255 * Ensures that the idle task is using init_mm right before its cpu goes
6258 void idle_task_exit(void)
6260 struct mm_struct
*mm
= current
->active_mm
;
6262 BUG_ON(cpu_online(smp_processor_id()));
6265 switch_mm(mm
, &init_mm
, current
);
6269 /* called under rq->lock with disabled interrupts */
6270 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
6272 struct rq
*rq
= cpu_rq(dead_cpu
);
6274 /* Must be exiting, otherwise would be on tasklist. */
6275 BUG_ON(!p
->exit_state
);
6277 /* Cannot have done final schedule yet: would have vanished. */
6278 BUG_ON(p
->state
== TASK_DEAD
);
6283 * Drop lock around migration; if someone else moves it,
6284 * that's OK. No task can be added to this CPU, so iteration is
6287 spin_unlock_irq(&rq
->lock
);
6288 move_task_off_dead_cpu(dead_cpu
, p
);
6289 spin_lock_irq(&rq
->lock
);
6294 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6295 static void migrate_dead_tasks(unsigned int dead_cpu
)
6297 struct rq
*rq
= cpu_rq(dead_cpu
);
6298 struct task_struct
*next
;
6301 if (!rq
->nr_running
)
6303 update_rq_clock(rq
);
6304 next
= pick_next_task(rq
, rq
->curr
);
6307 next
->sched_class
->put_prev_task(rq
, next
);
6308 migrate_dead(dead_cpu
, next
);
6312 #endif /* CONFIG_HOTPLUG_CPU */
6314 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6316 static struct ctl_table sd_ctl_dir
[] = {
6318 .procname
= "sched_domain",
6324 static struct ctl_table sd_ctl_root
[] = {
6326 .ctl_name
= CTL_KERN
,
6327 .procname
= "kernel",
6329 .child
= sd_ctl_dir
,
6334 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
6336 struct ctl_table
*entry
=
6337 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
6342 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
6344 struct ctl_table
*entry
;
6347 * In the intermediate directories, both the child directory and
6348 * procname are dynamically allocated and could fail but the mode
6349 * will always be set. In the lowest directory the names are
6350 * static strings and all have proc handlers.
6352 for (entry
= *tablep
; entry
->mode
; entry
++) {
6354 sd_free_ctl_entry(&entry
->child
);
6355 if (entry
->proc_handler
== NULL
)
6356 kfree(entry
->procname
);
6364 set_table_entry(struct ctl_table
*entry
,
6365 const char *procname
, void *data
, int maxlen
,
6366 mode_t mode
, proc_handler
*proc_handler
)
6368 entry
->procname
= procname
;
6370 entry
->maxlen
= maxlen
;
6372 entry
->proc_handler
= proc_handler
;
6375 static struct ctl_table
*
6376 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
6378 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
6383 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
6384 sizeof(long), 0644, proc_doulongvec_minmax
);
6385 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
6386 sizeof(long), 0644, proc_doulongvec_minmax
);
6387 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
6388 sizeof(int), 0644, proc_dointvec_minmax
);
6389 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
6390 sizeof(int), 0644, proc_dointvec_minmax
);
6391 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
6392 sizeof(int), 0644, proc_dointvec_minmax
);
6393 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
6394 sizeof(int), 0644, proc_dointvec_minmax
);
6395 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
6396 sizeof(int), 0644, proc_dointvec_minmax
);
6397 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
6398 sizeof(int), 0644, proc_dointvec_minmax
);
6399 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
6400 sizeof(int), 0644, proc_dointvec_minmax
);
6401 set_table_entry(&table
[9], "cache_nice_tries",
6402 &sd
->cache_nice_tries
,
6403 sizeof(int), 0644, proc_dointvec_minmax
);
6404 set_table_entry(&table
[10], "flags", &sd
->flags
,
6405 sizeof(int), 0644, proc_dointvec_minmax
);
6406 set_table_entry(&table
[11], "name", sd
->name
,
6407 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
6408 /* &table[12] is terminator */
6413 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
6415 struct ctl_table
*entry
, *table
;
6416 struct sched_domain
*sd
;
6417 int domain_num
= 0, i
;
6420 for_each_domain(cpu
, sd
)
6422 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
6427 for_each_domain(cpu
, sd
) {
6428 snprintf(buf
, 32, "domain%d", i
);
6429 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6431 entry
->child
= sd_alloc_ctl_domain_table(sd
);
6438 static struct ctl_table_header
*sd_sysctl_header
;
6439 static void register_sched_domain_sysctl(void)
6441 int i
, cpu_num
= num_online_cpus();
6442 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
6445 WARN_ON(sd_ctl_dir
[0].child
);
6446 sd_ctl_dir
[0].child
= entry
;
6451 for_each_online_cpu(i
) {
6452 snprintf(buf
, 32, "cpu%d", i
);
6453 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6455 entry
->child
= sd_alloc_ctl_cpu_table(i
);
6459 WARN_ON(sd_sysctl_header
);
6460 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
6463 /* may be called multiple times per register */
6464 static void unregister_sched_domain_sysctl(void)
6466 if (sd_sysctl_header
)
6467 unregister_sysctl_table(sd_sysctl_header
);
6468 sd_sysctl_header
= NULL
;
6469 if (sd_ctl_dir
[0].child
)
6470 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
6473 static void register_sched_domain_sysctl(void)
6476 static void unregister_sched_domain_sysctl(void)
6481 static void set_rq_online(struct rq
*rq
)
6484 const struct sched_class
*class;
6486 cpu_set(rq
->cpu
, rq
->rd
->online
);
6489 for_each_class(class) {
6490 if (class->rq_online
)
6491 class->rq_online(rq
);
6496 static void set_rq_offline(struct rq
*rq
)
6499 const struct sched_class
*class;
6501 for_each_class(class) {
6502 if (class->rq_offline
)
6503 class->rq_offline(rq
);
6506 cpu_clear(rq
->cpu
, rq
->rd
->online
);
6512 * migration_call - callback that gets triggered when a CPU is added.
6513 * Here we can start up the necessary migration thread for the new CPU.
6515 static int __cpuinit
6516 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
6518 struct task_struct
*p
;
6519 int cpu
= (long)hcpu
;
6520 unsigned long flags
;
6525 case CPU_UP_PREPARE
:
6526 case CPU_UP_PREPARE_FROZEN
:
6527 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
6530 kthread_bind(p
, cpu
);
6531 /* Must be high prio: stop_machine expects to yield to it. */
6532 rq
= task_rq_lock(p
, &flags
);
6533 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6534 task_rq_unlock(rq
, &flags
);
6535 cpu_rq(cpu
)->migration_thread
= p
;
6539 case CPU_ONLINE_FROZEN
:
6540 /* Strictly unnecessary, as first user will wake it. */
6541 wake_up_process(cpu_rq(cpu
)->migration_thread
);
6543 /* Update our root-domain */
6545 spin_lock_irqsave(&rq
->lock
, flags
);
6547 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6551 spin_unlock_irqrestore(&rq
->lock
, flags
);
6554 #ifdef CONFIG_HOTPLUG_CPU
6555 case CPU_UP_CANCELED
:
6556 case CPU_UP_CANCELED_FROZEN
:
6557 if (!cpu_rq(cpu
)->migration_thread
)
6559 /* Unbind it from offline cpu so it can run. Fall thru. */
6560 kthread_bind(cpu_rq(cpu
)->migration_thread
,
6561 any_online_cpu(cpu_online_map
));
6562 kthread_stop(cpu_rq(cpu
)->migration_thread
);
6563 cpu_rq(cpu
)->migration_thread
= NULL
;
6567 case CPU_DEAD_FROZEN
:
6568 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6569 migrate_live_tasks(cpu
);
6571 kthread_stop(rq
->migration_thread
);
6572 rq
->migration_thread
= NULL
;
6573 /* Idle task back to normal (off runqueue, low prio) */
6574 spin_lock_irq(&rq
->lock
);
6575 update_rq_clock(rq
);
6576 deactivate_task(rq
, rq
->idle
, 0);
6577 rq
->idle
->static_prio
= MAX_PRIO
;
6578 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
6579 rq
->idle
->sched_class
= &idle_sched_class
;
6580 migrate_dead_tasks(cpu
);
6581 spin_unlock_irq(&rq
->lock
);
6583 migrate_nr_uninterruptible(rq
);
6584 BUG_ON(rq
->nr_running
!= 0);
6587 * No need to migrate the tasks: it was best-effort if
6588 * they didn't take sched_hotcpu_mutex. Just wake up
6591 spin_lock_irq(&rq
->lock
);
6592 while (!list_empty(&rq
->migration_queue
)) {
6593 struct migration_req
*req
;
6595 req
= list_entry(rq
->migration_queue
.next
,
6596 struct migration_req
, list
);
6597 list_del_init(&req
->list
);
6598 complete(&req
->done
);
6600 spin_unlock_irq(&rq
->lock
);
6604 case CPU_DYING_FROZEN
:
6605 /* Update our root-domain */
6607 spin_lock_irqsave(&rq
->lock
, flags
);
6609 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6612 spin_unlock_irqrestore(&rq
->lock
, flags
);
6619 /* Register at highest priority so that task migration (migrate_all_tasks)
6620 * happens before everything else.
6622 static struct notifier_block __cpuinitdata migration_notifier
= {
6623 .notifier_call
= migration_call
,
6627 static int __init
migration_init(void)
6629 void *cpu
= (void *)(long)smp_processor_id();
6632 /* Start one for the boot CPU: */
6633 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6634 BUG_ON(err
== NOTIFY_BAD
);
6635 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6636 register_cpu_notifier(&migration_notifier
);
6640 early_initcall(migration_init
);
6645 #ifdef CONFIG_SCHED_DEBUG
6647 static inline const char *sd_level_to_string(enum sched_domain_level lvl
)
6660 case SD_LV_ALLNODES
:
6669 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6670 cpumask_t
*groupmask
)
6672 struct sched_group
*group
= sd
->groups
;
6675 cpulist_scnprintf(str
, sizeof(str
), sd
->span
);
6676 cpus_clear(*groupmask
);
6678 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6680 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6681 printk("does not load-balance\n");
6683 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6688 printk(KERN_CONT
"span %s level %s\n",
6689 str
, sd_level_to_string(sd
->level
));
6691 if (!cpu_isset(cpu
, sd
->span
)) {
6692 printk(KERN_ERR
"ERROR: domain->span does not contain "
6695 if (!cpu_isset(cpu
, group
->cpumask
)) {
6696 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6700 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6704 printk(KERN_ERR
"ERROR: group is NULL\n");
6708 if (!group
->__cpu_power
) {
6709 printk(KERN_CONT
"\n");
6710 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6715 if (!cpus_weight(group
->cpumask
)) {
6716 printk(KERN_CONT
"\n");
6717 printk(KERN_ERR
"ERROR: empty group\n");
6721 if (cpus_intersects(*groupmask
, group
->cpumask
)) {
6722 printk(KERN_CONT
"\n");
6723 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6727 cpus_or(*groupmask
, *groupmask
, group
->cpumask
);
6729 cpulist_scnprintf(str
, sizeof(str
), group
->cpumask
);
6730 printk(KERN_CONT
" %s", str
);
6732 group
= group
->next
;
6733 } while (group
!= sd
->groups
);
6734 printk(KERN_CONT
"\n");
6736 if (!cpus_equal(sd
->span
, *groupmask
))
6737 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6739 if (sd
->parent
&& !cpus_subset(*groupmask
, sd
->parent
->span
))
6740 printk(KERN_ERR
"ERROR: parent span is not a superset "
6741 "of domain->span\n");
6745 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6747 cpumask_t
*groupmask
;
6751 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6755 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6757 groupmask
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
6759 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6764 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6773 #else /* !CONFIG_SCHED_DEBUG */
6774 # define sched_domain_debug(sd, cpu) do { } while (0)
6775 #endif /* CONFIG_SCHED_DEBUG */
6777 static int sd_degenerate(struct sched_domain
*sd
)
6779 if (cpus_weight(sd
->span
) == 1)
6782 /* Following flags need at least 2 groups */
6783 if (sd
->flags
& (SD_LOAD_BALANCE
|
6784 SD_BALANCE_NEWIDLE
|
6788 SD_SHARE_PKG_RESOURCES
)) {
6789 if (sd
->groups
!= sd
->groups
->next
)
6793 /* Following flags don't use groups */
6794 if (sd
->flags
& (SD_WAKE_IDLE
|
6803 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6805 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6807 if (sd_degenerate(parent
))
6810 if (!cpus_equal(sd
->span
, parent
->span
))
6813 /* Does parent contain flags not in child? */
6814 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6815 if (cflags
& SD_WAKE_AFFINE
)
6816 pflags
&= ~SD_WAKE_BALANCE
;
6817 /* Flags needing groups don't count if only 1 group in parent */
6818 if (parent
->groups
== parent
->groups
->next
) {
6819 pflags
&= ~(SD_LOAD_BALANCE
|
6820 SD_BALANCE_NEWIDLE
|
6824 SD_SHARE_PKG_RESOURCES
);
6826 if (~cflags
& pflags
)
6832 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6834 unsigned long flags
;
6836 spin_lock_irqsave(&rq
->lock
, flags
);
6839 struct root_domain
*old_rd
= rq
->rd
;
6841 if (cpu_isset(rq
->cpu
, old_rd
->online
))
6844 cpu_clear(rq
->cpu
, old_rd
->span
);
6846 if (atomic_dec_and_test(&old_rd
->refcount
))
6850 atomic_inc(&rd
->refcount
);
6853 cpu_set(rq
->cpu
, rd
->span
);
6854 if (cpu_isset(rq
->cpu
, cpu_online_map
))
6857 spin_unlock_irqrestore(&rq
->lock
, flags
);
6860 static void init_rootdomain(struct root_domain
*rd
)
6862 memset(rd
, 0, sizeof(*rd
));
6864 cpus_clear(rd
->span
);
6865 cpus_clear(rd
->online
);
6867 cpupri_init(&rd
->cpupri
);
6870 static void init_defrootdomain(void)
6872 init_rootdomain(&def_root_domain
);
6873 atomic_set(&def_root_domain
.refcount
, 1);
6876 static struct root_domain
*alloc_rootdomain(void)
6878 struct root_domain
*rd
;
6880 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6884 init_rootdomain(rd
);
6890 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6891 * hold the hotplug lock.
6894 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6896 struct rq
*rq
= cpu_rq(cpu
);
6897 struct sched_domain
*tmp
;
6899 /* Remove the sched domains which do not contribute to scheduling. */
6900 for (tmp
= sd
; tmp
; ) {
6901 struct sched_domain
*parent
= tmp
->parent
;
6905 if (sd_parent_degenerate(tmp
, parent
)) {
6906 tmp
->parent
= parent
->parent
;
6908 parent
->parent
->child
= tmp
;
6913 if (sd
&& sd_degenerate(sd
)) {
6919 sched_domain_debug(sd
, cpu
);
6921 rq_attach_root(rq
, rd
);
6922 rcu_assign_pointer(rq
->sd
, sd
);
6925 /* cpus with isolated domains */
6926 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
6928 /* Setup the mask of cpus configured for isolated domains */
6929 static int __init
isolated_cpu_setup(char *str
)
6931 static int __initdata ints
[NR_CPUS
];
6934 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
6935 cpus_clear(cpu_isolated_map
);
6936 for (i
= 1; i
<= ints
[0]; i
++)
6937 if (ints
[i
] < NR_CPUS
)
6938 cpu_set(ints
[i
], cpu_isolated_map
);
6942 __setup("isolcpus=", isolated_cpu_setup
);
6945 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6946 * to a function which identifies what group(along with sched group) a CPU
6947 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6948 * (due to the fact that we keep track of groups covered with a cpumask_t).
6950 * init_sched_build_groups will build a circular linked list of the groups
6951 * covered by the given span, and will set each group's ->cpumask correctly,
6952 * and ->cpu_power to 0.
6955 init_sched_build_groups(const cpumask_t
*span
, const cpumask_t
*cpu_map
,
6956 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
6957 struct sched_group
**sg
,
6958 cpumask_t
*tmpmask
),
6959 cpumask_t
*covered
, cpumask_t
*tmpmask
)
6961 struct sched_group
*first
= NULL
, *last
= NULL
;
6964 cpus_clear(*covered
);
6966 for_each_cpu_mask_nr(i
, *span
) {
6967 struct sched_group
*sg
;
6968 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6971 if (cpu_isset(i
, *covered
))
6974 cpus_clear(sg
->cpumask
);
6975 sg
->__cpu_power
= 0;
6977 for_each_cpu_mask_nr(j
, *span
) {
6978 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6981 cpu_set(j
, *covered
);
6982 cpu_set(j
, sg
->cpumask
);
6993 #define SD_NODES_PER_DOMAIN 16
6998 * find_next_best_node - find the next node to include in a sched_domain
6999 * @node: node whose sched_domain we're building
7000 * @used_nodes: nodes already in the sched_domain
7002 * Find the next node to include in a given scheduling domain. Simply
7003 * finds the closest node not already in the @used_nodes map.
7005 * Should use nodemask_t.
7007 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
7009 int i
, n
, val
, min_val
, best_node
= 0;
7013 for (i
= 0; i
< nr_node_ids
; i
++) {
7014 /* Start at @node */
7015 n
= (node
+ i
) % nr_node_ids
;
7017 if (!nr_cpus_node(n
))
7020 /* Skip already used nodes */
7021 if (node_isset(n
, *used_nodes
))
7024 /* Simple min distance search */
7025 val
= node_distance(node
, n
);
7027 if (val
< min_val
) {
7033 node_set(best_node
, *used_nodes
);
7038 * sched_domain_node_span - get a cpumask for a node's sched_domain
7039 * @node: node whose cpumask we're constructing
7040 * @span: resulting cpumask
7042 * Given a node, construct a good cpumask for its sched_domain to span. It
7043 * should be one that prevents unnecessary balancing, but also spreads tasks
7046 static void sched_domain_node_span(int node
, cpumask_t
*span
)
7048 nodemask_t used_nodes
;
7049 node_to_cpumask_ptr(nodemask
, node
);
7053 nodes_clear(used_nodes
);
7055 cpus_or(*span
, *span
, *nodemask
);
7056 node_set(node
, used_nodes
);
7058 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
7059 int next_node
= find_next_best_node(node
, &used_nodes
);
7061 node_to_cpumask_ptr_next(nodemask
, next_node
);
7062 cpus_or(*span
, *span
, *nodemask
);
7065 #endif /* CONFIG_NUMA */
7067 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
7070 * SMT sched-domains:
7072 #ifdef CONFIG_SCHED_SMT
7073 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
7074 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
7077 cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7081 *sg
= &per_cpu(sched_group_cpus
, cpu
);
7084 #endif /* CONFIG_SCHED_SMT */
7087 * multi-core sched-domains:
7089 #ifdef CONFIG_SCHED_MC
7090 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
7091 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
7092 #endif /* CONFIG_SCHED_MC */
7094 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7096 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7101 *mask
= per_cpu(cpu_sibling_map
, cpu
);
7102 cpus_and(*mask
, *mask
, *cpu_map
);
7103 group
= first_cpu(*mask
);
7105 *sg
= &per_cpu(sched_group_core
, group
);
7108 #elif defined(CONFIG_SCHED_MC)
7110 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7114 *sg
= &per_cpu(sched_group_core
, cpu
);
7119 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
7120 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
7123 cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7127 #ifdef CONFIG_SCHED_MC
7128 *mask
= cpu_coregroup_map(cpu
);
7129 cpus_and(*mask
, *mask
, *cpu_map
);
7130 group
= first_cpu(*mask
);
7131 #elif defined(CONFIG_SCHED_SMT)
7132 *mask
= per_cpu(cpu_sibling_map
, cpu
);
7133 cpus_and(*mask
, *mask
, *cpu_map
);
7134 group
= first_cpu(*mask
);
7139 *sg
= &per_cpu(sched_group_phys
, group
);
7145 * The init_sched_build_groups can't handle what we want to do with node
7146 * groups, so roll our own. Now each node has its own list of groups which
7147 * gets dynamically allocated.
7149 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
7150 static struct sched_group
***sched_group_nodes_bycpu
;
7152 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
7153 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
7155 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
7156 struct sched_group
**sg
, cpumask_t
*nodemask
)
7160 *nodemask
= node_to_cpumask(cpu_to_node(cpu
));
7161 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7162 group
= first_cpu(*nodemask
);
7165 *sg
= &per_cpu(sched_group_allnodes
, group
);
7169 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
7171 struct sched_group
*sg
= group_head
;
7177 for_each_cpu_mask_nr(j
, sg
->cpumask
) {
7178 struct sched_domain
*sd
;
7180 sd
= &per_cpu(phys_domains
, j
);
7181 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
7183 * Only add "power" once for each
7189 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
7192 } while (sg
!= group_head
);
7194 #endif /* CONFIG_NUMA */
7197 /* Free memory allocated for various sched_group structures */
7198 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
7202 for_each_cpu_mask_nr(cpu
, *cpu_map
) {
7203 struct sched_group
**sched_group_nodes
7204 = sched_group_nodes_bycpu
[cpu
];
7206 if (!sched_group_nodes
)
7209 for (i
= 0; i
< nr_node_ids
; i
++) {
7210 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
7212 *nodemask
= node_to_cpumask(i
);
7213 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7214 if (cpus_empty(*nodemask
))
7224 if (oldsg
!= sched_group_nodes
[i
])
7227 kfree(sched_group_nodes
);
7228 sched_group_nodes_bycpu
[cpu
] = NULL
;
7231 #else /* !CONFIG_NUMA */
7232 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
7235 #endif /* CONFIG_NUMA */
7238 * Initialize sched groups cpu_power.
7240 * cpu_power indicates the capacity of sched group, which is used while
7241 * distributing the load between different sched groups in a sched domain.
7242 * Typically cpu_power for all the groups in a sched domain will be same unless
7243 * there are asymmetries in the topology. If there are asymmetries, group
7244 * having more cpu_power will pickup more load compared to the group having
7247 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7248 * the maximum number of tasks a group can handle in the presence of other idle
7249 * or lightly loaded groups in the same sched domain.
7251 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
7253 struct sched_domain
*child
;
7254 struct sched_group
*group
;
7256 WARN_ON(!sd
|| !sd
->groups
);
7258 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
7263 sd
->groups
->__cpu_power
= 0;
7266 * For perf policy, if the groups in child domain share resources
7267 * (for example cores sharing some portions of the cache hierarchy
7268 * or SMT), then set this domain groups cpu_power such that each group
7269 * can handle only one task, when there are other idle groups in the
7270 * same sched domain.
7272 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
7274 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
7275 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
7280 * add cpu_power of each child group to this groups cpu_power
7282 group
= child
->groups
;
7284 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
7285 group
= group
->next
;
7286 } while (group
!= child
->groups
);
7290 * Initializers for schedule domains
7291 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7294 #ifdef CONFIG_SCHED_DEBUG
7295 # define SD_INIT_NAME(sd, type) sd->name = #type
7297 # define SD_INIT_NAME(sd, type) do { } while (0)
7300 #define SD_INIT(sd, type) sd_init_##type(sd)
7302 #define SD_INIT_FUNC(type) \
7303 static noinline void sd_init_##type(struct sched_domain *sd) \
7305 memset(sd, 0, sizeof(*sd)); \
7306 *sd = SD_##type##_INIT; \
7307 sd->level = SD_LV_##type; \
7308 SD_INIT_NAME(sd, type); \
7313 SD_INIT_FUNC(ALLNODES
)
7316 #ifdef CONFIG_SCHED_SMT
7317 SD_INIT_FUNC(SIBLING
)
7319 #ifdef CONFIG_SCHED_MC
7324 * To minimize stack usage kmalloc room for cpumasks and share the
7325 * space as the usage in build_sched_domains() dictates. Used only
7326 * if the amount of space is significant.
7329 cpumask_t tmpmask
; /* make this one first */
7332 cpumask_t this_sibling_map
;
7333 cpumask_t this_core_map
;
7335 cpumask_t send_covered
;
7338 cpumask_t domainspan
;
7340 cpumask_t notcovered
;
7345 #define SCHED_CPUMASK_ALLOC 1
7346 #define SCHED_CPUMASK_FREE(v) kfree(v)
7347 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7349 #define SCHED_CPUMASK_ALLOC 0
7350 #define SCHED_CPUMASK_FREE(v)
7351 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7354 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7355 ((unsigned long)(a) + offsetof(struct allmasks, v))
7357 static int default_relax_domain_level
= -1;
7359 static int __init
setup_relax_domain_level(char *str
)
7363 val
= simple_strtoul(str
, NULL
, 0);
7364 if (val
< SD_LV_MAX
)
7365 default_relax_domain_level
= val
;
7369 __setup("relax_domain_level=", setup_relax_domain_level
);
7371 static void set_domain_attribute(struct sched_domain
*sd
,
7372 struct sched_domain_attr
*attr
)
7376 if (!attr
|| attr
->relax_domain_level
< 0) {
7377 if (default_relax_domain_level
< 0)
7380 request
= default_relax_domain_level
;
7382 request
= attr
->relax_domain_level
;
7383 if (request
< sd
->level
) {
7384 /* turn off idle balance on this domain */
7385 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
7387 /* turn on idle balance on this domain */
7388 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
7393 * Build sched domains for a given set of cpus and attach the sched domains
7394 * to the individual cpus
7396 static int __build_sched_domains(const cpumask_t
*cpu_map
,
7397 struct sched_domain_attr
*attr
)
7400 struct root_domain
*rd
;
7401 SCHED_CPUMASK_DECLARE(allmasks
);
7404 struct sched_group
**sched_group_nodes
= NULL
;
7405 int sd_allnodes
= 0;
7408 * Allocate the per-node list of sched groups
7410 sched_group_nodes
= kcalloc(nr_node_ids
, sizeof(struct sched_group
*),
7412 if (!sched_group_nodes
) {
7413 printk(KERN_WARNING
"Can not alloc sched group node list\n");
7418 rd
= alloc_rootdomain();
7420 printk(KERN_WARNING
"Cannot alloc root domain\n");
7422 kfree(sched_group_nodes
);
7427 #if SCHED_CPUMASK_ALLOC
7428 /* get space for all scratch cpumask variables */
7429 allmasks
= kmalloc(sizeof(*allmasks
), GFP_KERNEL
);
7431 printk(KERN_WARNING
"Cannot alloc cpumask array\n");
7434 kfree(sched_group_nodes
);
7439 tmpmask
= (cpumask_t
*)allmasks
;
7443 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
7447 * Set up domains for cpus specified by the cpu_map.
7449 for_each_cpu_mask_nr(i
, *cpu_map
) {
7450 struct sched_domain
*sd
= NULL
, *p
;
7451 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7453 *nodemask
= node_to_cpumask(cpu_to_node(i
));
7454 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7457 if (cpus_weight(*cpu_map
) >
7458 SD_NODES_PER_DOMAIN
*cpus_weight(*nodemask
)) {
7459 sd
= &per_cpu(allnodes_domains
, i
);
7460 SD_INIT(sd
, ALLNODES
);
7461 set_domain_attribute(sd
, attr
);
7462 sd
->span
= *cpu_map
;
7463 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7469 sd
= &per_cpu(node_domains
, i
);
7471 set_domain_attribute(sd
, attr
);
7472 sched_domain_node_span(cpu_to_node(i
), &sd
->span
);
7476 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7480 sd
= &per_cpu(phys_domains
, i
);
7482 set_domain_attribute(sd
, attr
);
7483 sd
->span
= *nodemask
;
7487 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7489 #ifdef CONFIG_SCHED_MC
7491 sd
= &per_cpu(core_domains
, i
);
7493 set_domain_attribute(sd
, attr
);
7494 sd
->span
= cpu_coregroup_map(i
);
7495 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7498 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7501 #ifdef CONFIG_SCHED_SMT
7503 sd
= &per_cpu(cpu_domains
, i
);
7504 SD_INIT(sd
, SIBLING
);
7505 set_domain_attribute(sd
, attr
);
7506 sd
->span
= per_cpu(cpu_sibling_map
, i
);
7507 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7510 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7514 #ifdef CONFIG_SCHED_SMT
7515 /* Set up CPU (sibling) groups */
7516 for_each_cpu_mask_nr(i
, *cpu_map
) {
7517 SCHED_CPUMASK_VAR(this_sibling_map
, allmasks
);
7518 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7520 *this_sibling_map
= per_cpu(cpu_sibling_map
, i
);
7521 cpus_and(*this_sibling_map
, *this_sibling_map
, *cpu_map
);
7522 if (i
!= first_cpu(*this_sibling_map
))
7525 init_sched_build_groups(this_sibling_map
, cpu_map
,
7527 send_covered
, tmpmask
);
7531 #ifdef CONFIG_SCHED_MC
7532 /* Set up multi-core groups */
7533 for_each_cpu_mask_nr(i
, *cpu_map
) {
7534 SCHED_CPUMASK_VAR(this_core_map
, allmasks
);
7535 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7537 *this_core_map
= cpu_coregroup_map(i
);
7538 cpus_and(*this_core_map
, *this_core_map
, *cpu_map
);
7539 if (i
!= first_cpu(*this_core_map
))
7542 init_sched_build_groups(this_core_map
, cpu_map
,
7544 send_covered
, tmpmask
);
7548 /* Set up physical groups */
7549 for (i
= 0; i
< nr_node_ids
; i
++) {
7550 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7551 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7553 *nodemask
= node_to_cpumask(i
);
7554 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7555 if (cpus_empty(*nodemask
))
7558 init_sched_build_groups(nodemask
, cpu_map
,
7560 send_covered
, tmpmask
);
7564 /* Set up node groups */
7566 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7568 init_sched_build_groups(cpu_map
, cpu_map
,
7569 &cpu_to_allnodes_group
,
7570 send_covered
, tmpmask
);
7573 for (i
= 0; i
< nr_node_ids
; i
++) {
7574 /* Set up node groups */
7575 struct sched_group
*sg
, *prev
;
7576 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7577 SCHED_CPUMASK_VAR(domainspan
, allmasks
);
7578 SCHED_CPUMASK_VAR(covered
, allmasks
);
7581 *nodemask
= node_to_cpumask(i
);
7582 cpus_clear(*covered
);
7584 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7585 if (cpus_empty(*nodemask
)) {
7586 sched_group_nodes
[i
] = NULL
;
7590 sched_domain_node_span(i
, domainspan
);
7591 cpus_and(*domainspan
, *domainspan
, *cpu_map
);
7593 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
7595 printk(KERN_WARNING
"Can not alloc domain group for "
7599 sched_group_nodes
[i
] = sg
;
7600 for_each_cpu_mask_nr(j
, *nodemask
) {
7601 struct sched_domain
*sd
;
7603 sd
= &per_cpu(node_domains
, j
);
7606 sg
->__cpu_power
= 0;
7607 sg
->cpumask
= *nodemask
;
7609 cpus_or(*covered
, *covered
, *nodemask
);
7612 for (j
= 0; j
< nr_node_ids
; j
++) {
7613 SCHED_CPUMASK_VAR(notcovered
, allmasks
);
7614 int n
= (i
+ j
) % nr_node_ids
;
7615 node_to_cpumask_ptr(pnodemask
, n
);
7617 cpus_complement(*notcovered
, *covered
);
7618 cpus_and(*tmpmask
, *notcovered
, *cpu_map
);
7619 cpus_and(*tmpmask
, *tmpmask
, *domainspan
);
7620 if (cpus_empty(*tmpmask
))
7623 cpus_and(*tmpmask
, *tmpmask
, *pnodemask
);
7624 if (cpus_empty(*tmpmask
))
7627 sg
= kmalloc_node(sizeof(struct sched_group
),
7631 "Can not alloc domain group for node %d\n", j
);
7634 sg
->__cpu_power
= 0;
7635 sg
->cpumask
= *tmpmask
;
7636 sg
->next
= prev
->next
;
7637 cpus_or(*covered
, *covered
, *tmpmask
);
7644 /* Calculate CPU power for physical packages and nodes */
7645 #ifdef CONFIG_SCHED_SMT
7646 for_each_cpu_mask_nr(i
, *cpu_map
) {
7647 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
7649 init_sched_groups_power(i
, sd
);
7652 #ifdef CONFIG_SCHED_MC
7653 for_each_cpu_mask_nr(i
, *cpu_map
) {
7654 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
7656 init_sched_groups_power(i
, sd
);
7660 for_each_cpu_mask_nr(i
, *cpu_map
) {
7661 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
7663 init_sched_groups_power(i
, sd
);
7667 for (i
= 0; i
< nr_node_ids
; i
++)
7668 init_numa_sched_groups_power(sched_group_nodes
[i
]);
7671 struct sched_group
*sg
;
7673 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
,
7675 init_numa_sched_groups_power(sg
);
7679 /* Attach the domains */
7680 for_each_cpu_mask_nr(i
, *cpu_map
) {
7681 struct sched_domain
*sd
;
7682 #ifdef CONFIG_SCHED_SMT
7683 sd
= &per_cpu(cpu_domains
, i
);
7684 #elif defined(CONFIG_SCHED_MC)
7685 sd
= &per_cpu(core_domains
, i
);
7687 sd
= &per_cpu(phys_domains
, i
);
7689 cpu_attach_domain(sd
, rd
, i
);
7692 SCHED_CPUMASK_FREE((void *)allmasks
);
7697 free_sched_groups(cpu_map
, tmpmask
);
7698 SCHED_CPUMASK_FREE((void *)allmasks
);
7704 static int build_sched_domains(const cpumask_t
*cpu_map
)
7706 return __build_sched_domains(cpu_map
, NULL
);
7709 static cpumask_t
*doms_cur
; /* current sched domains */
7710 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7711 static struct sched_domain_attr
*dattr_cur
;
7712 /* attribues of custom domains in 'doms_cur' */
7715 * Special case: If a kmalloc of a doms_cur partition (array of
7716 * cpumask_t) fails, then fallback to a single sched domain,
7717 * as determined by the single cpumask_t fallback_doms.
7719 static cpumask_t fallback_doms
;
7721 void __attribute__((weak
)) arch_update_cpu_topology(void)
7726 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7727 * For now this just excludes isolated cpus, but could be used to
7728 * exclude other special cases in the future.
7730 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
7734 arch_update_cpu_topology();
7736 doms_cur
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
7738 doms_cur
= &fallback_doms
;
7739 cpus_andnot(*doms_cur
, *cpu_map
, cpu_isolated_map
);
7741 err
= build_sched_domains(doms_cur
);
7742 register_sched_domain_sysctl();
7747 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
,
7750 free_sched_groups(cpu_map
, tmpmask
);
7754 * Detach sched domains from a group of cpus specified in cpu_map
7755 * These cpus will now be attached to the NULL domain
7757 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
7762 unregister_sched_domain_sysctl();
7764 for_each_cpu_mask_nr(i
, *cpu_map
)
7765 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7766 synchronize_sched();
7767 arch_destroy_sched_domains(cpu_map
, &tmpmask
);
7770 /* handle null as "default" */
7771 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7772 struct sched_domain_attr
*new, int idx_new
)
7774 struct sched_domain_attr tmp
;
7781 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7782 new ? (new + idx_new
) : &tmp
,
7783 sizeof(struct sched_domain_attr
));
7787 * Partition sched domains as specified by the 'ndoms_new'
7788 * cpumasks in the array doms_new[] of cpumasks. This compares
7789 * doms_new[] to the current sched domain partitioning, doms_cur[].
7790 * It destroys each deleted domain and builds each new domain.
7792 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7793 * The masks don't intersect (don't overlap.) We should setup one
7794 * sched domain for each mask. CPUs not in any of the cpumasks will
7795 * not be load balanced. If the same cpumask appears both in the
7796 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7799 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7800 * ownership of it and will kfree it when done with it. If the caller
7801 * failed the kmalloc call, then it can pass in doms_new == NULL &&
7802 * ndoms_new == 1, and partition_sched_domains() will fallback to
7803 * the single partition 'fallback_doms', it also forces the domains
7806 * If doms_new == NULL it will be replaced with cpu_online_map.
7807 * ndoms_new == 0 is a special case for destroying existing domains,
7808 * and it will not create the default domain.
7810 * Call with hotplug lock held
7812 void partition_sched_domains(int ndoms_new
, cpumask_t
*doms_new
,
7813 struct sched_domain_attr
*dattr_new
)
7817 mutex_lock(&sched_domains_mutex
);
7819 /* always unregister in case we don't destroy any domains */
7820 unregister_sched_domain_sysctl();
7822 n
= doms_new
? ndoms_new
: 0;
7824 /* Destroy deleted domains */
7825 for (i
= 0; i
< ndoms_cur
; i
++) {
7826 for (j
= 0; j
< n
; j
++) {
7827 if (cpus_equal(doms_cur
[i
], doms_new
[j
])
7828 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7831 /* no match - a current sched domain not in new doms_new[] */
7832 detach_destroy_domains(doms_cur
+ i
);
7837 if (doms_new
== NULL
) {
7839 doms_new
= &fallback_doms
;
7840 cpus_andnot(doms_new
[0], cpu_online_map
, cpu_isolated_map
);
7844 /* Build new domains */
7845 for (i
= 0; i
< ndoms_new
; i
++) {
7846 for (j
= 0; j
< ndoms_cur
; j
++) {
7847 if (cpus_equal(doms_new
[i
], doms_cur
[j
])
7848 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7851 /* no match - add a new doms_new */
7852 __build_sched_domains(doms_new
+ i
,
7853 dattr_new
? dattr_new
+ i
: NULL
);
7858 /* Remember the new sched domains */
7859 if (doms_cur
!= &fallback_doms
)
7861 kfree(dattr_cur
); /* kfree(NULL) is safe */
7862 doms_cur
= doms_new
;
7863 dattr_cur
= dattr_new
;
7864 ndoms_cur
= ndoms_new
;
7866 register_sched_domain_sysctl();
7868 mutex_unlock(&sched_domains_mutex
);
7871 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7872 int arch_reinit_sched_domains(void)
7876 /* Destroy domains first to force the rebuild */
7877 partition_sched_domains(0, NULL
, NULL
);
7879 rebuild_sched_domains();
7885 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7889 if (buf
[0] != '0' && buf
[0] != '1')
7893 sched_smt_power_savings
= (buf
[0] == '1');
7895 sched_mc_power_savings
= (buf
[0] == '1');
7897 ret
= arch_reinit_sched_domains();
7899 return ret
? ret
: count
;
7902 #ifdef CONFIG_SCHED_MC
7903 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
7906 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7908 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
7909 const char *buf
, size_t count
)
7911 return sched_power_savings_store(buf
, count
, 0);
7913 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
7914 sched_mc_power_savings_show
,
7915 sched_mc_power_savings_store
);
7918 #ifdef CONFIG_SCHED_SMT
7919 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
7922 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7924 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
7925 const char *buf
, size_t count
)
7927 return sched_power_savings_store(buf
, count
, 1);
7929 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
7930 sched_smt_power_savings_show
,
7931 sched_smt_power_savings_store
);
7934 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7938 #ifdef CONFIG_SCHED_SMT
7940 err
= sysfs_create_file(&cls
->kset
.kobj
,
7941 &attr_sched_smt_power_savings
.attr
);
7943 #ifdef CONFIG_SCHED_MC
7944 if (!err
&& mc_capable())
7945 err
= sysfs_create_file(&cls
->kset
.kobj
,
7946 &attr_sched_mc_power_savings
.attr
);
7950 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7952 #ifndef CONFIG_CPUSETS
7954 * Add online and remove offline CPUs from the scheduler domains.
7955 * When cpusets are enabled they take over this function.
7957 static int update_sched_domains(struct notifier_block
*nfb
,
7958 unsigned long action
, void *hcpu
)
7962 case CPU_ONLINE_FROZEN
:
7964 case CPU_DEAD_FROZEN
:
7965 partition_sched_domains(1, NULL
, NULL
);
7974 static int update_runtime(struct notifier_block
*nfb
,
7975 unsigned long action
, void *hcpu
)
7977 int cpu
= (int)(long)hcpu
;
7980 case CPU_DOWN_PREPARE
:
7981 case CPU_DOWN_PREPARE_FROZEN
:
7982 disable_runtime(cpu_rq(cpu
));
7985 case CPU_DOWN_FAILED
:
7986 case CPU_DOWN_FAILED_FROZEN
:
7988 case CPU_ONLINE_FROZEN
:
7989 enable_runtime(cpu_rq(cpu
));
7997 void __init
sched_init_smp(void)
7999 cpumask_t non_isolated_cpus
;
8001 #if defined(CONFIG_NUMA)
8002 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
8004 BUG_ON(sched_group_nodes_bycpu
== NULL
);
8007 mutex_lock(&sched_domains_mutex
);
8008 arch_init_sched_domains(&cpu_online_map
);
8009 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
8010 if (cpus_empty(non_isolated_cpus
))
8011 cpu_set(smp_processor_id(), non_isolated_cpus
);
8012 mutex_unlock(&sched_domains_mutex
);
8015 #ifndef CONFIG_CPUSETS
8016 /* XXX: Theoretical race here - CPU may be hotplugged now */
8017 hotcpu_notifier(update_sched_domains
, 0);
8020 /* RT runtime code needs to handle some hotplug events */
8021 hotcpu_notifier(update_runtime
, 0);
8025 /* Move init over to a non-isolated CPU */
8026 if (set_cpus_allowed_ptr(current
, &non_isolated_cpus
) < 0)
8028 sched_init_granularity();
8031 void __init
sched_init_smp(void)
8033 sched_init_granularity();
8035 #endif /* CONFIG_SMP */
8037 int in_sched_functions(unsigned long addr
)
8039 return in_lock_functions(addr
) ||
8040 (addr
>= (unsigned long)__sched_text_start
8041 && addr
< (unsigned long)__sched_text_end
);
8044 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
8046 cfs_rq
->tasks_timeline
= RB_ROOT
;
8047 INIT_LIST_HEAD(&cfs_rq
->tasks
);
8048 #ifdef CONFIG_FAIR_GROUP_SCHED
8051 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
8054 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
8056 struct rt_prio_array
*array
;
8059 array
= &rt_rq
->active
;
8060 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
8061 INIT_LIST_HEAD(array
->queue
+ i
);
8062 __clear_bit(i
, array
->bitmap
);
8064 /* delimiter for bitsearch: */
8065 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
8067 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8068 rt_rq
->highest_prio
= MAX_RT_PRIO
;
8071 rt_rq
->rt_nr_migratory
= 0;
8072 rt_rq
->overloaded
= 0;
8076 rt_rq
->rt_throttled
= 0;
8077 rt_rq
->rt_runtime
= 0;
8078 spin_lock_init(&rt_rq
->rt_runtime_lock
);
8080 #ifdef CONFIG_RT_GROUP_SCHED
8081 rt_rq
->rt_nr_boosted
= 0;
8086 #ifdef CONFIG_FAIR_GROUP_SCHED
8087 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
8088 struct sched_entity
*se
, int cpu
, int add
,
8089 struct sched_entity
*parent
)
8091 struct rq
*rq
= cpu_rq(cpu
);
8092 tg
->cfs_rq
[cpu
] = cfs_rq
;
8093 init_cfs_rq(cfs_rq
, rq
);
8096 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
8099 /* se could be NULL for init_task_group */
8104 se
->cfs_rq
= &rq
->cfs
;
8106 se
->cfs_rq
= parent
->my_q
;
8109 se
->load
.weight
= tg
->shares
;
8110 se
->load
.inv_weight
= 0;
8111 se
->parent
= parent
;
8115 #ifdef CONFIG_RT_GROUP_SCHED
8116 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
8117 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
8118 struct sched_rt_entity
*parent
)
8120 struct rq
*rq
= cpu_rq(cpu
);
8122 tg
->rt_rq
[cpu
] = rt_rq
;
8123 init_rt_rq(rt_rq
, rq
);
8125 rt_rq
->rt_se
= rt_se
;
8126 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8128 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
8130 tg
->rt_se
[cpu
] = rt_se
;
8135 rt_se
->rt_rq
= &rq
->rt
;
8137 rt_se
->rt_rq
= parent
->my_q
;
8139 rt_se
->my_q
= rt_rq
;
8140 rt_se
->parent
= parent
;
8141 INIT_LIST_HEAD(&rt_se
->run_list
);
8145 void __init
sched_init(void)
8148 unsigned long alloc_size
= 0, ptr
;
8150 #ifdef CONFIG_FAIR_GROUP_SCHED
8151 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8153 #ifdef CONFIG_RT_GROUP_SCHED
8154 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8156 #ifdef CONFIG_USER_SCHED
8160 * As sched_init() is called before page_alloc is setup,
8161 * we use alloc_bootmem().
8164 ptr
= (unsigned long)alloc_bootmem(alloc_size
);
8166 #ifdef CONFIG_FAIR_GROUP_SCHED
8167 init_task_group
.se
= (struct sched_entity
**)ptr
;
8168 ptr
+= nr_cpu_ids
* sizeof(void **);
8170 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8171 ptr
+= nr_cpu_ids
* sizeof(void **);
8173 #ifdef CONFIG_USER_SCHED
8174 root_task_group
.se
= (struct sched_entity
**)ptr
;
8175 ptr
+= nr_cpu_ids
* sizeof(void **);
8177 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8178 ptr
+= nr_cpu_ids
* sizeof(void **);
8179 #endif /* CONFIG_USER_SCHED */
8180 #endif /* CONFIG_FAIR_GROUP_SCHED */
8181 #ifdef CONFIG_RT_GROUP_SCHED
8182 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8183 ptr
+= nr_cpu_ids
* sizeof(void **);
8185 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8186 ptr
+= nr_cpu_ids
* sizeof(void **);
8188 #ifdef CONFIG_USER_SCHED
8189 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8190 ptr
+= nr_cpu_ids
* sizeof(void **);
8192 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8193 ptr
+= nr_cpu_ids
* sizeof(void **);
8194 #endif /* CONFIG_USER_SCHED */
8195 #endif /* CONFIG_RT_GROUP_SCHED */
8199 init_defrootdomain();
8202 init_rt_bandwidth(&def_rt_bandwidth
,
8203 global_rt_period(), global_rt_runtime());
8205 #ifdef CONFIG_RT_GROUP_SCHED
8206 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
8207 global_rt_period(), global_rt_runtime());
8208 #ifdef CONFIG_USER_SCHED
8209 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
8210 global_rt_period(), RUNTIME_INF
);
8211 #endif /* CONFIG_USER_SCHED */
8212 #endif /* CONFIG_RT_GROUP_SCHED */
8214 #ifdef CONFIG_GROUP_SCHED
8215 list_add(&init_task_group
.list
, &task_groups
);
8216 INIT_LIST_HEAD(&init_task_group
.children
);
8218 #ifdef CONFIG_USER_SCHED
8219 INIT_LIST_HEAD(&root_task_group
.children
);
8220 init_task_group
.parent
= &root_task_group
;
8221 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
8222 #endif /* CONFIG_USER_SCHED */
8223 #endif /* CONFIG_GROUP_SCHED */
8225 for_each_possible_cpu(i
) {
8229 spin_lock_init(&rq
->lock
);
8231 init_cfs_rq(&rq
->cfs
, rq
);
8232 init_rt_rq(&rq
->rt
, rq
);
8233 #ifdef CONFIG_FAIR_GROUP_SCHED
8234 init_task_group
.shares
= init_task_group_load
;
8235 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
8236 #ifdef CONFIG_CGROUP_SCHED
8238 * How much cpu bandwidth does init_task_group get?
8240 * In case of task-groups formed thr' the cgroup filesystem, it
8241 * gets 100% of the cpu resources in the system. This overall
8242 * system cpu resource is divided among the tasks of
8243 * init_task_group and its child task-groups in a fair manner,
8244 * based on each entity's (task or task-group's) weight
8245 * (se->load.weight).
8247 * In other words, if init_task_group has 10 tasks of weight
8248 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8249 * then A0's share of the cpu resource is:
8251 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8253 * We achieve this by letting init_task_group's tasks sit
8254 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8256 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
8257 #elif defined CONFIG_USER_SCHED
8258 root_task_group
.shares
= NICE_0_LOAD
;
8259 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
8261 * In case of task-groups formed thr' the user id of tasks,
8262 * init_task_group represents tasks belonging to root user.
8263 * Hence it forms a sibling of all subsequent groups formed.
8264 * In this case, init_task_group gets only a fraction of overall
8265 * system cpu resource, based on the weight assigned to root
8266 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8267 * by letting tasks of init_task_group sit in a separate cfs_rq
8268 * (init_cfs_rq) and having one entity represent this group of
8269 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8271 init_tg_cfs_entry(&init_task_group
,
8272 &per_cpu(init_cfs_rq
, i
),
8273 &per_cpu(init_sched_entity
, i
), i
, 1,
8274 root_task_group
.se
[i
]);
8277 #endif /* CONFIG_FAIR_GROUP_SCHED */
8279 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
8280 #ifdef CONFIG_RT_GROUP_SCHED
8281 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
8282 #ifdef CONFIG_CGROUP_SCHED
8283 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
8284 #elif defined CONFIG_USER_SCHED
8285 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
8286 init_tg_rt_entry(&init_task_group
,
8287 &per_cpu(init_rt_rq
, i
),
8288 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
8289 root_task_group
.rt_se
[i
]);
8293 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
8294 rq
->cpu_load
[j
] = 0;
8298 rq
->active_balance
= 0;
8299 rq
->next_balance
= jiffies
;
8303 rq
->migration_thread
= NULL
;
8304 INIT_LIST_HEAD(&rq
->migration_queue
);
8305 rq_attach_root(rq
, &def_root_domain
);
8308 atomic_set(&rq
->nr_iowait
, 0);
8311 set_load_weight(&init_task
);
8313 #ifdef CONFIG_PREEMPT_NOTIFIERS
8314 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
8318 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
8321 #ifdef CONFIG_RT_MUTEXES
8322 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
8326 * The boot idle thread does lazy MMU switching as well:
8328 atomic_inc(&init_mm
.mm_count
);
8329 enter_lazy_tlb(&init_mm
, current
);
8332 * Make us the idle thread. Technically, schedule() should not be
8333 * called from this thread, however somewhere below it might be,
8334 * but because we are the idle thread, we just pick up running again
8335 * when this runqueue becomes "idle".
8337 init_idle(current
, smp_processor_id());
8339 * During early bootup we pretend to be a normal task:
8341 current
->sched_class
= &fair_sched_class
;
8343 scheduler_running
= 1;
8346 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8347 void __might_sleep(char *file
, int line
)
8350 static unsigned long prev_jiffy
; /* ratelimiting */
8352 if ((!in_atomic() && !irqs_disabled()) ||
8353 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
8355 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
8357 prev_jiffy
= jiffies
;
8360 "BUG: sleeping function called from invalid context at %s:%d\n",
8363 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8364 in_atomic(), irqs_disabled(),
8365 current
->pid
, current
->comm
);
8367 debug_show_held_locks(current
);
8368 if (irqs_disabled())
8369 print_irqtrace_events(current
);
8373 EXPORT_SYMBOL(__might_sleep
);
8376 #ifdef CONFIG_MAGIC_SYSRQ
8377 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
8381 update_rq_clock(rq
);
8382 on_rq
= p
->se
.on_rq
;
8384 deactivate_task(rq
, p
, 0);
8385 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
8387 activate_task(rq
, p
, 0);
8388 resched_task(rq
->curr
);
8392 void normalize_rt_tasks(void)
8394 struct task_struct
*g
, *p
;
8395 unsigned long flags
;
8398 read_lock_irqsave(&tasklist_lock
, flags
);
8399 do_each_thread(g
, p
) {
8401 * Only normalize user tasks:
8406 p
->se
.exec_start
= 0;
8407 #ifdef CONFIG_SCHEDSTATS
8408 p
->se
.wait_start
= 0;
8409 p
->se
.sleep_start
= 0;
8410 p
->se
.block_start
= 0;
8415 * Renice negative nice level userspace
8418 if (TASK_NICE(p
) < 0 && p
->mm
)
8419 set_user_nice(p
, 0);
8423 spin_lock(&p
->pi_lock
);
8424 rq
= __task_rq_lock(p
);
8426 normalize_task(rq
, p
);
8428 __task_rq_unlock(rq
);
8429 spin_unlock(&p
->pi_lock
);
8430 } while_each_thread(g
, p
);
8432 read_unlock_irqrestore(&tasklist_lock
, flags
);
8435 #endif /* CONFIG_MAGIC_SYSRQ */
8439 * These functions are only useful for the IA64 MCA handling.
8441 * They can only be called when the whole system has been
8442 * stopped - every CPU needs to be quiescent, and no scheduling
8443 * activity can take place. Using them for anything else would
8444 * be a serious bug, and as a result, they aren't even visible
8445 * under any other configuration.
8449 * curr_task - return the current task for a given cpu.
8450 * @cpu: the processor in question.
8452 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8454 struct task_struct
*curr_task(int cpu
)
8456 return cpu_curr(cpu
);
8460 * set_curr_task - set the current task for a given cpu.
8461 * @cpu: the processor in question.
8462 * @p: the task pointer to set.
8464 * Description: This function must only be used when non-maskable interrupts
8465 * are serviced on a separate stack. It allows the architecture to switch the
8466 * notion of the current task on a cpu in a non-blocking manner. This function
8467 * must be called with all CPU's synchronized, and interrupts disabled, the
8468 * and caller must save the original value of the current task (see
8469 * curr_task() above) and restore that value before reenabling interrupts and
8470 * re-starting the system.
8472 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8474 void set_curr_task(int cpu
, struct task_struct
*p
)
8481 #ifdef CONFIG_FAIR_GROUP_SCHED
8482 static void free_fair_sched_group(struct task_group
*tg
)
8486 for_each_possible_cpu(i
) {
8488 kfree(tg
->cfs_rq
[i
]);
8498 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8500 struct cfs_rq
*cfs_rq
;
8501 struct sched_entity
*se
, *parent_se
;
8505 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8508 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8512 tg
->shares
= NICE_0_LOAD
;
8514 for_each_possible_cpu(i
) {
8517 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
),
8518 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8522 se
= kmalloc_node(sizeof(struct sched_entity
),
8523 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8527 parent_se
= parent
? parent
->se
[i
] : NULL
;
8528 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent_se
);
8537 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8539 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
8540 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
8543 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8545 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
8547 #else /* !CONFG_FAIR_GROUP_SCHED */
8548 static inline void free_fair_sched_group(struct task_group
*tg
)
8553 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8558 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8562 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8565 #endif /* CONFIG_FAIR_GROUP_SCHED */
8567 #ifdef CONFIG_RT_GROUP_SCHED
8568 static void free_rt_sched_group(struct task_group
*tg
)
8572 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8574 for_each_possible_cpu(i
) {
8576 kfree(tg
->rt_rq
[i
]);
8578 kfree(tg
->rt_se
[i
]);
8586 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8588 struct rt_rq
*rt_rq
;
8589 struct sched_rt_entity
*rt_se
, *parent_se
;
8593 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8596 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8600 init_rt_bandwidth(&tg
->rt_bandwidth
,
8601 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8603 for_each_possible_cpu(i
) {
8606 rt_rq
= kmalloc_node(sizeof(struct rt_rq
),
8607 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8611 rt_se
= kmalloc_node(sizeof(struct sched_rt_entity
),
8612 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8616 parent_se
= parent
? parent
->rt_se
[i
] : NULL
;
8617 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent_se
);
8626 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8628 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8629 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8632 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8634 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8636 #else /* !CONFIG_RT_GROUP_SCHED */
8637 static inline void free_rt_sched_group(struct task_group
*tg
)
8642 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8647 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8651 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8654 #endif /* CONFIG_RT_GROUP_SCHED */
8656 #ifdef CONFIG_GROUP_SCHED
8657 static void free_sched_group(struct task_group
*tg
)
8659 free_fair_sched_group(tg
);
8660 free_rt_sched_group(tg
);
8664 /* allocate runqueue etc for a new task group */
8665 struct task_group
*sched_create_group(struct task_group
*parent
)
8667 struct task_group
*tg
;
8668 unsigned long flags
;
8671 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8673 return ERR_PTR(-ENOMEM
);
8675 if (!alloc_fair_sched_group(tg
, parent
))
8678 if (!alloc_rt_sched_group(tg
, parent
))
8681 spin_lock_irqsave(&task_group_lock
, flags
);
8682 for_each_possible_cpu(i
) {
8683 register_fair_sched_group(tg
, i
);
8684 register_rt_sched_group(tg
, i
);
8686 list_add_rcu(&tg
->list
, &task_groups
);
8688 WARN_ON(!parent
); /* root should already exist */
8690 tg
->parent
= parent
;
8691 INIT_LIST_HEAD(&tg
->children
);
8692 list_add_rcu(&tg
->siblings
, &parent
->children
);
8693 spin_unlock_irqrestore(&task_group_lock
, flags
);
8698 free_sched_group(tg
);
8699 return ERR_PTR(-ENOMEM
);
8702 /* rcu callback to free various structures associated with a task group */
8703 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8705 /* now it should be safe to free those cfs_rqs */
8706 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8709 /* Destroy runqueue etc associated with a task group */
8710 void sched_destroy_group(struct task_group
*tg
)
8712 unsigned long flags
;
8715 spin_lock_irqsave(&task_group_lock
, flags
);
8716 for_each_possible_cpu(i
) {
8717 unregister_fair_sched_group(tg
, i
);
8718 unregister_rt_sched_group(tg
, i
);
8720 list_del_rcu(&tg
->list
);
8721 list_del_rcu(&tg
->siblings
);
8722 spin_unlock_irqrestore(&task_group_lock
, flags
);
8724 /* wait for possible concurrent references to cfs_rqs complete */
8725 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8728 /* change task's runqueue when it moves between groups.
8729 * The caller of this function should have put the task in its new group
8730 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8731 * reflect its new group.
8733 void sched_move_task(struct task_struct
*tsk
)
8736 unsigned long flags
;
8739 rq
= task_rq_lock(tsk
, &flags
);
8741 update_rq_clock(rq
);
8743 running
= task_current(rq
, tsk
);
8744 on_rq
= tsk
->se
.on_rq
;
8747 dequeue_task(rq
, tsk
, 0);
8748 if (unlikely(running
))
8749 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8751 set_task_rq(tsk
, task_cpu(tsk
));
8753 #ifdef CONFIG_FAIR_GROUP_SCHED
8754 if (tsk
->sched_class
->moved_group
)
8755 tsk
->sched_class
->moved_group(tsk
);
8758 if (unlikely(running
))
8759 tsk
->sched_class
->set_curr_task(rq
);
8761 enqueue_task(rq
, tsk
, 0);
8763 task_rq_unlock(rq
, &flags
);
8765 #endif /* CONFIG_GROUP_SCHED */
8767 #ifdef CONFIG_FAIR_GROUP_SCHED
8768 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8770 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8775 dequeue_entity(cfs_rq
, se
, 0);
8777 se
->load
.weight
= shares
;
8778 se
->load
.inv_weight
= 0;
8781 enqueue_entity(cfs_rq
, se
, 0);
8784 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8786 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8787 struct rq
*rq
= cfs_rq
->rq
;
8788 unsigned long flags
;
8790 spin_lock_irqsave(&rq
->lock
, flags
);
8791 __set_se_shares(se
, shares
);
8792 spin_unlock_irqrestore(&rq
->lock
, flags
);
8795 static DEFINE_MUTEX(shares_mutex
);
8797 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8800 unsigned long flags
;
8803 * We can't change the weight of the root cgroup.
8808 if (shares
< MIN_SHARES
)
8809 shares
= MIN_SHARES
;
8810 else if (shares
> MAX_SHARES
)
8811 shares
= MAX_SHARES
;
8813 mutex_lock(&shares_mutex
);
8814 if (tg
->shares
== shares
)
8817 spin_lock_irqsave(&task_group_lock
, flags
);
8818 for_each_possible_cpu(i
)
8819 unregister_fair_sched_group(tg
, i
);
8820 list_del_rcu(&tg
->siblings
);
8821 spin_unlock_irqrestore(&task_group_lock
, flags
);
8823 /* wait for any ongoing reference to this group to finish */
8824 synchronize_sched();
8827 * Now we are free to modify the group's share on each cpu
8828 * w/o tripping rebalance_share or load_balance_fair.
8830 tg
->shares
= shares
;
8831 for_each_possible_cpu(i
) {
8835 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
8836 set_se_shares(tg
->se
[i
], shares
);
8840 * Enable load balance activity on this group, by inserting it back on
8841 * each cpu's rq->leaf_cfs_rq_list.
8843 spin_lock_irqsave(&task_group_lock
, flags
);
8844 for_each_possible_cpu(i
)
8845 register_fair_sched_group(tg
, i
);
8846 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
8847 spin_unlock_irqrestore(&task_group_lock
, flags
);
8849 mutex_unlock(&shares_mutex
);
8853 unsigned long sched_group_shares(struct task_group
*tg
)
8859 #ifdef CONFIG_RT_GROUP_SCHED
8861 * Ensure that the real time constraints are schedulable.
8863 static DEFINE_MUTEX(rt_constraints_mutex
);
8865 static unsigned long to_ratio(u64 period
, u64 runtime
)
8867 if (runtime
== RUNTIME_INF
)
8870 return div64_u64(runtime
<< 20, period
);
8873 /* Must be called with tasklist_lock held */
8874 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8876 struct task_struct
*g
, *p
;
8878 do_each_thread(g
, p
) {
8879 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8881 } while_each_thread(g
, p
);
8886 struct rt_schedulable_data
{
8887 struct task_group
*tg
;
8892 static int tg_schedulable(struct task_group
*tg
, void *data
)
8894 struct rt_schedulable_data
*d
= data
;
8895 struct task_group
*child
;
8896 unsigned long total
, sum
= 0;
8897 u64 period
, runtime
;
8899 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8900 runtime
= tg
->rt_bandwidth
.rt_runtime
;
8903 period
= d
->rt_period
;
8904 runtime
= d
->rt_runtime
;
8908 * Cannot have more runtime than the period.
8910 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8914 * Ensure we don't starve existing RT tasks.
8916 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
8919 total
= to_ratio(period
, runtime
);
8922 * Nobody can have more than the global setting allows.
8924 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
8928 * The sum of our children's runtime should not exceed our own.
8930 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
8931 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
8932 runtime
= child
->rt_bandwidth
.rt_runtime
;
8934 if (child
== d
->tg
) {
8935 period
= d
->rt_period
;
8936 runtime
= d
->rt_runtime
;
8939 sum
+= to_ratio(period
, runtime
);
8948 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8950 struct rt_schedulable_data data
= {
8952 .rt_period
= period
,
8953 .rt_runtime
= runtime
,
8956 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
8959 static int tg_set_bandwidth(struct task_group
*tg
,
8960 u64 rt_period
, u64 rt_runtime
)
8964 mutex_lock(&rt_constraints_mutex
);
8965 read_lock(&tasklist_lock
);
8966 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
8970 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8971 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8972 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8974 for_each_possible_cpu(i
) {
8975 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8977 spin_lock(&rt_rq
->rt_runtime_lock
);
8978 rt_rq
->rt_runtime
= rt_runtime
;
8979 spin_unlock(&rt_rq
->rt_runtime_lock
);
8981 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8983 read_unlock(&tasklist_lock
);
8984 mutex_unlock(&rt_constraints_mutex
);
8989 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8991 u64 rt_runtime
, rt_period
;
8993 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8994 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8995 if (rt_runtime_us
< 0)
8996 rt_runtime
= RUNTIME_INF
;
8998 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
9001 long sched_group_rt_runtime(struct task_group
*tg
)
9005 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
9008 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
9009 do_div(rt_runtime_us
, NSEC_PER_USEC
);
9010 return rt_runtime_us
;
9013 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
9015 u64 rt_runtime
, rt_period
;
9017 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
9018 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
9023 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
9026 long sched_group_rt_period(struct task_group
*tg
)
9030 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9031 do_div(rt_period_us
, NSEC_PER_USEC
);
9032 return rt_period_us
;
9035 static int sched_rt_global_constraints(void)
9037 u64 runtime
, period
;
9040 if (sysctl_sched_rt_period
<= 0)
9043 runtime
= global_rt_runtime();
9044 period
= global_rt_period();
9047 * Sanity check on the sysctl variables.
9049 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
9052 mutex_lock(&rt_constraints_mutex
);
9053 read_lock(&tasklist_lock
);
9054 ret
= __rt_schedulable(NULL
, 0, 0);
9055 read_unlock(&tasklist_lock
);
9056 mutex_unlock(&rt_constraints_mutex
);
9060 #else /* !CONFIG_RT_GROUP_SCHED */
9061 static int sched_rt_global_constraints(void)
9063 unsigned long flags
;
9066 if (sysctl_sched_rt_period
<= 0)
9069 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9070 for_each_possible_cpu(i
) {
9071 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
9073 spin_lock(&rt_rq
->rt_runtime_lock
);
9074 rt_rq
->rt_runtime
= global_rt_runtime();
9075 spin_unlock(&rt_rq
->rt_runtime_lock
);
9077 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9081 #endif /* CONFIG_RT_GROUP_SCHED */
9083 int sched_rt_handler(struct ctl_table
*table
, int write
,
9084 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
9088 int old_period
, old_runtime
;
9089 static DEFINE_MUTEX(mutex
);
9092 old_period
= sysctl_sched_rt_period
;
9093 old_runtime
= sysctl_sched_rt_runtime
;
9095 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
9097 if (!ret
&& write
) {
9098 ret
= sched_rt_global_constraints();
9100 sysctl_sched_rt_period
= old_period
;
9101 sysctl_sched_rt_runtime
= old_runtime
;
9103 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
9104 def_rt_bandwidth
.rt_period
=
9105 ns_to_ktime(global_rt_period());
9108 mutex_unlock(&mutex
);
9113 #ifdef CONFIG_CGROUP_SCHED
9115 /* return corresponding task_group object of a cgroup */
9116 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
9118 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
9119 struct task_group
, css
);
9122 static struct cgroup_subsys_state
*
9123 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9125 struct task_group
*tg
, *parent
;
9127 if (!cgrp
->parent
) {
9128 /* This is early initialization for the top cgroup */
9129 return &init_task_group
.css
;
9132 parent
= cgroup_tg(cgrp
->parent
);
9133 tg
= sched_create_group(parent
);
9135 return ERR_PTR(-ENOMEM
);
9141 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9143 struct task_group
*tg
= cgroup_tg(cgrp
);
9145 sched_destroy_group(tg
);
9149 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9150 struct task_struct
*tsk
)
9152 #ifdef CONFIG_RT_GROUP_SCHED
9153 /* Don't accept realtime tasks when there is no way for them to run */
9154 if (rt_task(tsk
) && cgroup_tg(cgrp
)->rt_bandwidth
.rt_runtime
== 0)
9157 /* We don't support RT-tasks being in separate groups */
9158 if (tsk
->sched_class
!= &fair_sched_class
)
9166 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9167 struct cgroup
*old_cont
, struct task_struct
*tsk
)
9169 sched_move_task(tsk
);
9172 #ifdef CONFIG_FAIR_GROUP_SCHED
9173 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
9176 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
9179 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
9181 struct task_group
*tg
= cgroup_tg(cgrp
);
9183 return (u64
) tg
->shares
;
9185 #endif /* CONFIG_FAIR_GROUP_SCHED */
9187 #ifdef CONFIG_RT_GROUP_SCHED
9188 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
9191 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
9194 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9196 return sched_group_rt_runtime(cgroup_tg(cgrp
));
9199 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
9202 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
9205 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
9207 return sched_group_rt_period(cgroup_tg(cgrp
));
9209 #endif /* CONFIG_RT_GROUP_SCHED */
9211 static struct cftype cpu_files
[] = {
9212 #ifdef CONFIG_FAIR_GROUP_SCHED
9215 .read_u64
= cpu_shares_read_u64
,
9216 .write_u64
= cpu_shares_write_u64
,
9219 #ifdef CONFIG_RT_GROUP_SCHED
9221 .name
= "rt_runtime_us",
9222 .read_s64
= cpu_rt_runtime_read
,
9223 .write_s64
= cpu_rt_runtime_write
,
9226 .name
= "rt_period_us",
9227 .read_u64
= cpu_rt_period_read_uint
,
9228 .write_u64
= cpu_rt_period_write_uint
,
9233 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
9235 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
9238 struct cgroup_subsys cpu_cgroup_subsys
= {
9240 .create
= cpu_cgroup_create
,
9241 .destroy
= cpu_cgroup_destroy
,
9242 .can_attach
= cpu_cgroup_can_attach
,
9243 .attach
= cpu_cgroup_attach
,
9244 .populate
= cpu_cgroup_populate
,
9245 .subsys_id
= cpu_cgroup_subsys_id
,
9249 #endif /* CONFIG_CGROUP_SCHED */
9251 #ifdef CONFIG_CGROUP_CPUACCT
9254 * CPU accounting code for task groups.
9256 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9257 * (balbir@in.ibm.com).
9260 /* track cpu usage of a group of tasks */
9262 struct cgroup_subsys_state css
;
9263 /* cpuusage holds pointer to a u64-type object on every cpu */
9267 struct cgroup_subsys cpuacct_subsys
;
9269 /* return cpu accounting group corresponding to this container */
9270 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
9272 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
9273 struct cpuacct
, css
);
9276 /* return cpu accounting group to which this task belongs */
9277 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
9279 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
9280 struct cpuacct
, css
);
9283 /* create a new cpu accounting group */
9284 static struct cgroup_subsys_state
*cpuacct_create(
9285 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9287 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
9290 return ERR_PTR(-ENOMEM
);
9292 ca
->cpuusage
= alloc_percpu(u64
);
9293 if (!ca
->cpuusage
) {
9295 return ERR_PTR(-ENOMEM
);
9301 /* destroy an existing cpu accounting group */
9303 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9305 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9307 free_percpu(ca
->cpuusage
);
9311 /* return total cpu usage (in nanoseconds) of a group */
9312 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9314 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9315 u64 totalcpuusage
= 0;
9318 for_each_possible_cpu(i
) {
9319 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
9322 * Take rq->lock to make 64-bit addition safe on 32-bit
9325 spin_lock_irq(&cpu_rq(i
)->lock
);
9326 totalcpuusage
+= *cpuusage
;
9327 spin_unlock_irq(&cpu_rq(i
)->lock
);
9330 return totalcpuusage
;
9333 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
9336 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9345 for_each_possible_cpu(i
) {
9346 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
9348 spin_lock_irq(&cpu_rq(i
)->lock
);
9350 spin_unlock_irq(&cpu_rq(i
)->lock
);
9356 static struct cftype files
[] = {
9359 .read_u64
= cpuusage_read
,
9360 .write_u64
= cpuusage_write
,
9364 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9366 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
9370 * charge this task's execution time to its accounting group.
9372 * called with rq->lock held.
9374 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
9378 if (!cpuacct_subsys
.active
)
9383 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, task_cpu(tsk
));
9385 *cpuusage
+= cputime
;
9389 struct cgroup_subsys cpuacct_subsys
= {
9391 .create
= cpuacct_create
,
9392 .destroy
= cpuacct_destroy
,
9393 .populate
= cpuacct_populate
,
9394 .subsys_id
= cpuacct_subsys_id
,
9396 #endif /* CONFIG_CGROUP_CPUACCT */