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/perf_counter.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/kthread.h>
59 #include <linux/proc_fs.h>
60 #include <linux/seq_file.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/reciprocal_div.h>
68 #include <linux/unistd.h>
69 #include <linux/pagemap.h>
70 #include <linux/hrtimer.h>
71 #include <linux/tick.h>
72 #include <linux/debugfs.h>
73 #include <linux/ctype.h>
74 #include <linux/ftrace.h>
77 #include <asm/irq_regs.h>
79 #include "sched_cpupri.h"
81 #define CREATE_TRACE_POINTS
82 #include <trace/events/sched.h>
85 * Convert user-nice values [ -20 ... 0 ... 19 ]
86 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
89 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
90 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
91 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
94 * 'User priority' is the nice value converted to something we
95 * can work with better when scaling various scheduler parameters,
96 * it's a [ 0 ... 39 ] range.
98 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
99 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
100 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
103 * Helpers for converting nanosecond timing to jiffy resolution
105 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
107 #define NICE_0_LOAD SCHED_LOAD_SCALE
108 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
111 * These are the 'tuning knobs' of the scheduler:
113 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
114 * Timeslices get refilled after they expire.
116 #define DEF_TIMESLICE (100 * HZ / 1000)
119 * single value that denotes runtime == period, ie unlimited time.
121 #define RUNTIME_INF ((u64)~0ULL)
125 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
128 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
129 * Since cpu_power is a 'constant', we can use a reciprocal divide.
131 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
133 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
137 * Each time a sched group cpu_power is changed,
138 * we must compute its reciprocal value
140 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
142 sg
->__cpu_power
+= val
;
143 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
147 static inline int rt_policy(int policy
)
149 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
154 static inline int task_has_rt_policy(struct task_struct
*p
)
156 return rt_policy(p
->policy
);
160 * This is the priority-queue data structure of the RT scheduling class:
162 struct rt_prio_array
{
163 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
164 struct list_head queue
[MAX_RT_PRIO
];
167 struct rt_bandwidth
{
168 /* nests inside the rq lock: */
169 spinlock_t rt_runtime_lock
;
172 struct hrtimer rt_period_timer
;
175 static struct rt_bandwidth def_rt_bandwidth
;
177 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
179 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
181 struct rt_bandwidth
*rt_b
=
182 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
188 now
= hrtimer_cb_get_time(timer
);
189 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
194 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
197 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
201 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
203 rt_b
->rt_period
= ns_to_ktime(period
);
204 rt_b
->rt_runtime
= runtime
;
206 spin_lock_init(&rt_b
->rt_runtime_lock
);
208 hrtimer_init(&rt_b
->rt_period_timer
,
209 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
210 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
213 static inline int rt_bandwidth_enabled(void)
215 return sysctl_sched_rt_runtime
>= 0;
218 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
222 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
225 if (hrtimer_active(&rt_b
->rt_period_timer
))
228 spin_lock(&rt_b
->rt_runtime_lock
);
233 if (hrtimer_active(&rt_b
->rt_period_timer
))
236 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
237 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
239 soft
= hrtimer_get_softexpires(&rt_b
->rt_period_timer
);
240 hard
= hrtimer_get_expires(&rt_b
->rt_period_timer
);
241 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
242 __hrtimer_start_range_ns(&rt_b
->rt_period_timer
, soft
, delta
,
243 HRTIMER_MODE_ABS
, 0);
245 spin_unlock(&rt_b
->rt_runtime_lock
);
248 #ifdef CONFIG_RT_GROUP_SCHED
249 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
251 hrtimer_cancel(&rt_b
->rt_period_timer
);
256 * sched_domains_mutex serializes calls to arch_init_sched_domains,
257 * detach_destroy_domains and partition_sched_domains.
259 static DEFINE_MUTEX(sched_domains_mutex
);
261 #ifdef CONFIG_GROUP_SCHED
263 #include <linux/cgroup.h>
267 static LIST_HEAD(task_groups
);
269 /* task group related information */
271 #ifdef CONFIG_CGROUP_SCHED
272 struct cgroup_subsys_state css
;
275 #ifdef CONFIG_USER_SCHED
279 #ifdef CONFIG_FAIR_GROUP_SCHED
280 /* schedulable entities of this group on each cpu */
281 struct sched_entity
**se
;
282 /* runqueue "owned" by this group on each cpu */
283 struct cfs_rq
**cfs_rq
;
284 unsigned long shares
;
287 #ifdef CONFIG_RT_GROUP_SCHED
288 struct sched_rt_entity
**rt_se
;
289 struct rt_rq
**rt_rq
;
291 struct rt_bandwidth rt_bandwidth
;
295 struct list_head list
;
297 struct task_group
*parent
;
298 struct list_head siblings
;
299 struct list_head children
;
302 #ifdef CONFIG_USER_SCHED
304 /* Helper function to pass uid information to create_sched_user() */
305 void set_tg_uid(struct user_struct
*user
)
307 user
->tg
->uid
= user
->uid
;
312 * Every UID task group (including init_task_group aka UID-0) will
313 * be a child to this group.
315 struct task_group root_task_group
;
317 #ifdef CONFIG_FAIR_GROUP_SCHED
318 /* Default task group's sched entity on each cpu */
319 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
320 /* Default task group's cfs_rq on each cpu */
321 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
322 #endif /* CONFIG_FAIR_GROUP_SCHED */
324 #ifdef CONFIG_RT_GROUP_SCHED
325 static DEFINE_PER_CPU(struct sched_rt_entity
, init_sched_rt_entity
);
326 static DEFINE_PER_CPU(struct rt_rq
, init_rt_rq
) ____cacheline_aligned_in_smp
;
327 #endif /* CONFIG_RT_GROUP_SCHED */
328 #else /* !CONFIG_USER_SCHED */
329 #define root_task_group init_task_group
330 #endif /* CONFIG_USER_SCHED */
332 /* task_group_lock serializes add/remove of task groups and also changes to
333 * a task group's cpu shares.
335 static DEFINE_SPINLOCK(task_group_lock
);
338 static int root_task_group_empty(void)
340 return list_empty(&root_task_group
.children
);
344 #ifdef CONFIG_FAIR_GROUP_SCHED
345 #ifdef CONFIG_USER_SCHED
346 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
347 #else /* !CONFIG_USER_SCHED */
348 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
349 #endif /* CONFIG_USER_SCHED */
352 * A weight of 0 or 1 can cause arithmetics problems.
353 * A weight of a cfs_rq is the sum of weights of which entities
354 * are queued on this cfs_rq, so a weight of a entity should not be
355 * too large, so as the shares value of a task group.
356 * (The default weight is 1024 - so there's no practical
357 * limitation from this.)
360 #define MAX_SHARES (1UL << 18)
362 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
365 /* Default task group.
366 * Every task in system belong to this group at bootup.
368 struct task_group init_task_group
;
370 /* return group to which a task belongs */
371 static inline struct task_group
*task_group(struct task_struct
*p
)
373 struct task_group
*tg
;
375 #ifdef CONFIG_USER_SCHED
377 tg
= __task_cred(p
)->user
->tg
;
379 #elif defined(CONFIG_CGROUP_SCHED)
380 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
381 struct task_group
, css
);
383 tg
= &init_task_group
;
388 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
389 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
391 #ifdef CONFIG_FAIR_GROUP_SCHED
392 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
393 p
->se
.parent
= task_group(p
)->se
[cpu
];
396 #ifdef CONFIG_RT_GROUP_SCHED
397 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
398 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
405 static int root_task_group_empty(void)
411 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
412 static inline struct task_group
*task_group(struct task_struct
*p
)
417 #endif /* CONFIG_GROUP_SCHED */
419 /* CFS-related fields in a runqueue */
421 struct load_weight load
;
422 unsigned long nr_running
;
427 struct rb_root tasks_timeline
;
428 struct rb_node
*rb_leftmost
;
430 struct list_head tasks
;
431 struct list_head
*balance_iterator
;
434 * 'curr' points to currently running entity on this cfs_rq.
435 * It is set to NULL otherwise (i.e when none are currently running).
437 struct sched_entity
*curr
, *next
, *last
;
439 unsigned int nr_spread_over
;
441 #ifdef CONFIG_FAIR_GROUP_SCHED
442 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
445 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
446 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
447 * (like users, containers etc.)
449 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
450 * list is used during load balance.
452 struct list_head leaf_cfs_rq_list
;
453 struct task_group
*tg
; /* group that "owns" this runqueue */
457 * the part of load.weight contributed by tasks
459 unsigned long task_weight
;
462 * h_load = weight * f(tg)
464 * Where f(tg) is the recursive weight fraction assigned to
467 unsigned long h_load
;
470 * this cpu's part of tg->shares
472 unsigned long shares
;
475 * load.weight at the time we set shares
477 unsigned long rq_weight
;
482 /* Real-Time classes' related field in a runqueue: */
484 struct rt_prio_array active
;
485 unsigned long rt_nr_running
;
486 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
488 int curr
; /* highest queued rt task prio */
490 int next
; /* next highest */
495 unsigned long rt_nr_migratory
;
497 struct plist_head pushable_tasks
;
502 /* Nests inside the rq lock: */
503 spinlock_t rt_runtime_lock
;
505 #ifdef CONFIG_RT_GROUP_SCHED
506 unsigned long rt_nr_boosted
;
509 struct list_head leaf_rt_rq_list
;
510 struct task_group
*tg
;
511 struct sched_rt_entity
*rt_se
;
518 * We add the notion of a root-domain which will be used to define per-domain
519 * variables. Each exclusive cpuset essentially defines an island domain by
520 * fully partitioning the member cpus from any other cpuset. Whenever a new
521 * exclusive cpuset is created, we also create and attach a new root-domain
528 cpumask_var_t online
;
531 * The "RT overload" flag: it gets set if a CPU has more than
532 * one runnable RT task.
534 cpumask_var_t rto_mask
;
537 struct cpupri cpupri
;
539 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
541 * Preferred wake up cpu nominated by sched_mc balance that will be
542 * used when most cpus are idle in the system indicating overall very
543 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
545 unsigned int sched_mc_preferred_wakeup_cpu
;
550 * By default the system creates a single root-domain with all cpus as
551 * members (mimicking the global state we have today).
553 static struct root_domain def_root_domain
;
558 * This is the main, per-CPU runqueue data structure.
560 * Locking rule: those places that want to lock multiple runqueues
561 * (such as the load balancing or the thread migration code), lock
562 * acquire operations must be ordered by ascending &runqueue.
569 * nr_running and cpu_load should be in the same cacheline because
570 * remote CPUs use both these fields when doing load calculation.
572 unsigned long nr_running
;
573 #define CPU_LOAD_IDX_MAX 5
574 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
576 unsigned long last_tick_seen
;
577 unsigned char in_nohz_recently
;
579 /* capture load from *all* tasks on this cpu: */
580 struct load_weight load
;
581 unsigned long nr_load_updates
;
583 u64 nr_migrations_in
;
588 #ifdef CONFIG_FAIR_GROUP_SCHED
589 /* list of leaf cfs_rq on this cpu: */
590 struct list_head leaf_cfs_rq_list
;
592 #ifdef CONFIG_RT_GROUP_SCHED
593 struct list_head leaf_rt_rq_list
;
597 * This is part of a global counter where only the total sum
598 * over all CPUs matters. A task can increase this counter on
599 * one CPU and if it got migrated afterwards it may decrease
600 * it on another CPU. Always updated under the runqueue lock:
602 unsigned long nr_uninterruptible
;
604 struct task_struct
*curr
, *idle
;
605 unsigned long next_balance
;
606 struct mm_struct
*prev_mm
;
613 struct root_domain
*rd
;
614 struct sched_domain
*sd
;
616 unsigned char idle_at_tick
;
617 /* For active balancing */
620 /* cpu of this runqueue: */
624 unsigned long avg_load_per_task
;
626 struct task_struct
*migration_thread
;
627 struct list_head migration_queue
;
630 /* calc_load related fields */
631 unsigned long calc_load_update
;
632 long calc_load_active
;
634 #ifdef CONFIG_SCHED_HRTICK
636 int hrtick_csd_pending
;
637 struct call_single_data hrtick_csd
;
639 struct hrtimer hrtick_timer
;
642 #ifdef CONFIG_SCHEDSTATS
644 struct sched_info rq_sched_info
;
645 unsigned long long rq_cpu_time
;
646 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
648 /* sys_sched_yield() stats */
649 unsigned int yld_count
;
651 /* schedule() stats */
652 unsigned int sched_switch
;
653 unsigned int sched_count
;
654 unsigned int sched_goidle
;
656 /* try_to_wake_up() stats */
657 unsigned int ttwu_count
;
658 unsigned int ttwu_local
;
661 unsigned int bkl_count
;
665 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
667 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int sync
)
669 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, sync
);
672 static inline int cpu_of(struct rq
*rq
)
682 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
683 * See detach_destroy_domains: synchronize_sched for details.
685 * The domain tree of any CPU may only be accessed from within
686 * preempt-disabled sections.
688 #define for_each_domain(cpu, __sd) \
689 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
691 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
692 #define this_rq() (&__get_cpu_var(runqueues))
693 #define task_rq(p) cpu_rq(task_cpu(p))
694 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
696 inline void update_rq_clock(struct rq
*rq
)
698 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
702 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
704 #ifdef CONFIG_SCHED_DEBUG
705 # define const_debug __read_mostly
707 # define const_debug static const
713 * Returns true if the current cpu runqueue is locked.
714 * This interface allows printk to be called with the runqueue lock
715 * held and know whether or not it is OK to wake up the klogd.
717 int runqueue_is_locked(void)
720 struct rq
*rq
= cpu_rq(cpu
);
723 ret
= spin_is_locked(&rq
->lock
);
729 * Debugging: various feature bits
732 #define SCHED_FEAT(name, enabled) \
733 __SCHED_FEAT_##name ,
736 #include "sched_features.h"
741 #define SCHED_FEAT(name, enabled) \
742 (1UL << __SCHED_FEAT_##name) * enabled |
744 const_debug
unsigned int sysctl_sched_features
=
745 #include "sched_features.h"
750 #ifdef CONFIG_SCHED_DEBUG
751 #define SCHED_FEAT(name, enabled) \
754 static __read_mostly
char *sched_feat_names
[] = {
755 #include "sched_features.h"
761 static int sched_feat_show(struct seq_file
*m
, void *v
)
765 for (i
= 0; sched_feat_names
[i
]; i
++) {
766 if (!(sysctl_sched_features
& (1UL << i
)))
768 seq_printf(m
, "%s ", sched_feat_names
[i
]);
776 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
777 size_t cnt
, loff_t
*ppos
)
787 if (copy_from_user(&buf
, ubuf
, cnt
))
792 if (strncmp(buf
, "NO_", 3) == 0) {
797 for (i
= 0; sched_feat_names
[i
]; i
++) {
798 int len
= strlen(sched_feat_names
[i
]);
800 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
802 sysctl_sched_features
&= ~(1UL << i
);
804 sysctl_sched_features
|= (1UL << i
);
809 if (!sched_feat_names
[i
])
817 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
819 return single_open(filp
, sched_feat_show
, NULL
);
822 static struct file_operations sched_feat_fops
= {
823 .open
= sched_feat_open
,
824 .write
= sched_feat_write
,
827 .release
= single_release
,
830 static __init
int sched_init_debug(void)
832 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
837 late_initcall(sched_init_debug
);
841 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
844 * Number of tasks to iterate in a single balance run.
845 * Limited because this is done with IRQs disabled.
847 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
850 * ratelimit for updating the group shares.
853 unsigned int sysctl_sched_shares_ratelimit
= 250000;
856 * Inject some fuzzyness into changing the per-cpu group shares
857 * this avoids remote rq-locks at the expense of fairness.
860 unsigned int sysctl_sched_shares_thresh
= 4;
863 * period over which we measure -rt task cpu usage in us.
866 unsigned int sysctl_sched_rt_period
= 1000000;
868 static __read_mostly
int scheduler_running
;
871 * part of the period that we allow rt tasks to run in us.
874 int sysctl_sched_rt_runtime
= 950000;
876 static inline u64
global_rt_period(void)
878 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
881 static inline u64
global_rt_runtime(void)
883 if (sysctl_sched_rt_runtime
< 0)
886 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
889 #ifndef prepare_arch_switch
890 # define prepare_arch_switch(next) do { } while (0)
892 #ifndef finish_arch_switch
893 # define finish_arch_switch(prev) do { } while (0)
896 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
898 return rq
->curr
== p
;
901 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
902 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
904 return task_current(rq
, p
);
907 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
911 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
913 #ifdef CONFIG_DEBUG_SPINLOCK
914 /* this is a valid case when another task releases the spinlock */
915 rq
->lock
.owner
= current
;
918 * If we are tracking spinlock dependencies then we have to
919 * fix up the runqueue lock - which gets 'carried over' from
922 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
924 spin_unlock_irq(&rq
->lock
);
927 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
928 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
933 return task_current(rq
, p
);
937 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
941 * We can optimise this out completely for !SMP, because the
942 * SMP rebalancing from interrupt is the only thing that cares
947 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
948 spin_unlock_irq(&rq
->lock
);
950 spin_unlock(&rq
->lock
);
954 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
958 * After ->oncpu is cleared, the task can be moved to a different CPU.
959 * We must ensure this doesn't happen until the switch is completely
965 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
969 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
972 * __task_rq_lock - lock the runqueue a given task resides on.
973 * Must be called interrupts disabled.
975 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
979 struct rq
*rq
= task_rq(p
);
980 spin_lock(&rq
->lock
);
981 if (likely(rq
== task_rq(p
)))
983 spin_unlock(&rq
->lock
);
988 * task_rq_lock - lock the runqueue a given task resides on and disable
989 * interrupts. Note the ordering: we can safely lookup the task_rq without
990 * explicitly disabling preemption.
992 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
998 local_irq_save(*flags
);
1000 spin_lock(&rq
->lock
);
1001 if (likely(rq
== task_rq(p
)))
1003 spin_unlock_irqrestore(&rq
->lock
, *flags
);
1007 void task_rq_unlock_wait(struct task_struct
*p
)
1009 struct rq
*rq
= task_rq(p
);
1011 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1012 spin_unlock_wait(&rq
->lock
);
1015 static void __task_rq_unlock(struct rq
*rq
)
1016 __releases(rq
->lock
)
1018 spin_unlock(&rq
->lock
);
1021 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
1022 __releases(rq
->lock
)
1024 spin_unlock_irqrestore(&rq
->lock
, *flags
);
1028 * this_rq_lock - lock this runqueue and disable interrupts.
1030 static struct rq
*this_rq_lock(void)
1031 __acquires(rq
->lock
)
1035 local_irq_disable();
1037 spin_lock(&rq
->lock
);
1042 #ifdef CONFIG_SCHED_HRTICK
1044 * Use HR-timers to deliver accurate preemption points.
1046 * Its all a bit involved since we cannot program an hrt while holding the
1047 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1050 * When we get rescheduled we reprogram the hrtick_timer outside of the
1056 * - enabled by features
1057 * - hrtimer is actually high res
1059 static inline int hrtick_enabled(struct rq
*rq
)
1061 if (!sched_feat(HRTICK
))
1063 if (!cpu_active(cpu_of(rq
)))
1065 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1068 static void hrtick_clear(struct rq
*rq
)
1070 if (hrtimer_active(&rq
->hrtick_timer
))
1071 hrtimer_cancel(&rq
->hrtick_timer
);
1075 * High-resolution timer tick.
1076 * Runs from hardirq context with interrupts disabled.
1078 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1080 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1082 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1084 spin_lock(&rq
->lock
);
1085 update_rq_clock(rq
);
1086 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1087 spin_unlock(&rq
->lock
);
1089 return HRTIMER_NORESTART
;
1094 * called from hardirq (IPI) context
1096 static void __hrtick_start(void *arg
)
1098 struct rq
*rq
= arg
;
1100 spin_lock(&rq
->lock
);
1101 hrtimer_restart(&rq
->hrtick_timer
);
1102 rq
->hrtick_csd_pending
= 0;
1103 spin_unlock(&rq
->lock
);
1107 * Called to set the hrtick timer state.
1109 * called with rq->lock held and irqs disabled
1111 static void hrtick_start(struct rq
*rq
, u64 delay
)
1113 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1114 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1116 hrtimer_set_expires(timer
, time
);
1118 if (rq
== this_rq()) {
1119 hrtimer_restart(timer
);
1120 } else if (!rq
->hrtick_csd_pending
) {
1121 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1122 rq
->hrtick_csd_pending
= 1;
1127 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1129 int cpu
= (int)(long)hcpu
;
1132 case CPU_UP_CANCELED
:
1133 case CPU_UP_CANCELED_FROZEN
:
1134 case CPU_DOWN_PREPARE
:
1135 case CPU_DOWN_PREPARE_FROZEN
:
1137 case CPU_DEAD_FROZEN
:
1138 hrtick_clear(cpu_rq(cpu
));
1145 static __init
void init_hrtick(void)
1147 hotcpu_notifier(hotplug_hrtick
, 0);
1151 * Called to set the hrtick timer state.
1153 * called with rq->lock held and irqs disabled
1155 static void hrtick_start(struct rq
*rq
, u64 delay
)
1157 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1158 HRTIMER_MODE_REL
, 0);
1161 static inline void init_hrtick(void)
1164 #endif /* CONFIG_SMP */
1166 static void init_rq_hrtick(struct rq
*rq
)
1169 rq
->hrtick_csd_pending
= 0;
1171 rq
->hrtick_csd
.flags
= 0;
1172 rq
->hrtick_csd
.func
= __hrtick_start
;
1173 rq
->hrtick_csd
.info
= rq
;
1176 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1177 rq
->hrtick_timer
.function
= hrtick
;
1179 #else /* CONFIG_SCHED_HRTICK */
1180 static inline void hrtick_clear(struct rq
*rq
)
1184 static inline void init_rq_hrtick(struct rq
*rq
)
1188 static inline void init_hrtick(void)
1191 #endif /* CONFIG_SCHED_HRTICK */
1194 * resched_task - mark a task 'to be rescheduled now'.
1196 * On UP this means the setting of the need_resched flag, on SMP it
1197 * might also involve a cross-CPU call to trigger the scheduler on
1202 #ifndef tsk_is_polling
1203 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1206 static void resched_task(struct task_struct
*p
)
1210 assert_spin_locked(&task_rq(p
)->lock
);
1212 if (test_tsk_need_resched(p
))
1215 set_tsk_need_resched(p
);
1218 if (cpu
== smp_processor_id())
1221 /* NEED_RESCHED must be visible before we test polling */
1223 if (!tsk_is_polling(p
))
1224 smp_send_reschedule(cpu
);
1227 static void resched_cpu(int cpu
)
1229 struct rq
*rq
= cpu_rq(cpu
);
1230 unsigned long flags
;
1232 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1234 resched_task(cpu_curr(cpu
));
1235 spin_unlock_irqrestore(&rq
->lock
, flags
);
1240 * When add_timer_on() enqueues a timer into the timer wheel of an
1241 * idle CPU then this timer might expire before the next timer event
1242 * which is scheduled to wake up that CPU. In case of a completely
1243 * idle system the next event might even be infinite time into the
1244 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1245 * leaves the inner idle loop so the newly added timer is taken into
1246 * account when the CPU goes back to idle and evaluates the timer
1247 * wheel for the next timer event.
1249 void wake_up_idle_cpu(int cpu
)
1251 struct rq
*rq
= cpu_rq(cpu
);
1253 if (cpu
== smp_processor_id())
1257 * This is safe, as this function is called with the timer
1258 * wheel base lock of (cpu) held. When the CPU is on the way
1259 * to idle and has not yet set rq->curr to idle then it will
1260 * be serialized on the timer wheel base lock and take the new
1261 * timer into account automatically.
1263 if (rq
->curr
!= rq
->idle
)
1267 * We can set TIF_RESCHED on the idle task of the other CPU
1268 * lockless. The worst case is that the other CPU runs the
1269 * idle task through an additional NOOP schedule()
1271 set_tsk_need_resched(rq
->idle
);
1273 /* NEED_RESCHED must be visible before we test polling */
1275 if (!tsk_is_polling(rq
->idle
))
1276 smp_send_reschedule(cpu
);
1278 #endif /* CONFIG_NO_HZ */
1280 #else /* !CONFIG_SMP */
1281 static void resched_task(struct task_struct
*p
)
1283 assert_spin_locked(&task_rq(p
)->lock
);
1284 set_tsk_need_resched(p
);
1286 #endif /* CONFIG_SMP */
1288 #if BITS_PER_LONG == 32
1289 # define WMULT_CONST (~0UL)
1291 # define WMULT_CONST (1UL << 32)
1294 #define WMULT_SHIFT 32
1297 * Shift right and round:
1299 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1302 * delta *= weight / lw
1304 static unsigned long
1305 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1306 struct load_weight
*lw
)
1310 if (!lw
->inv_weight
) {
1311 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1314 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1318 tmp
= (u64
)delta_exec
* weight
;
1320 * Check whether we'd overflow the 64-bit multiplication:
1322 if (unlikely(tmp
> WMULT_CONST
))
1323 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1326 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1328 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1331 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1337 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1344 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1345 * of tasks with abnormal "nice" values across CPUs the contribution that
1346 * each task makes to its run queue's load is weighted according to its
1347 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1348 * scaled version of the new time slice allocation that they receive on time
1352 #define WEIGHT_IDLEPRIO 3
1353 #define WMULT_IDLEPRIO 1431655765
1356 * Nice levels are multiplicative, with a gentle 10% change for every
1357 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1358 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1359 * that remained on nice 0.
1361 * The "10% effect" is relative and cumulative: from _any_ nice level,
1362 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1363 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1364 * If a task goes up by ~10% and another task goes down by ~10% then
1365 * the relative distance between them is ~25%.)
1367 static const int prio_to_weight
[40] = {
1368 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1369 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1370 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1371 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1372 /* 0 */ 1024, 820, 655, 526, 423,
1373 /* 5 */ 335, 272, 215, 172, 137,
1374 /* 10 */ 110, 87, 70, 56, 45,
1375 /* 15 */ 36, 29, 23, 18, 15,
1379 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1381 * In cases where the weight does not change often, we can use the
1382 * precalculated inverse to speed up arithmetics by turning divisions
1383 * into multiplications:
1385 static const u32 prio_to_wmult
[40] = {
1386 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1387 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1388 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1389 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1390 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1391 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1392 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1393 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1396 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1399 * runqueue iterator, to support SMP load-balancing between different
1400 * scheduling classes, without having to expose their internal data
1401 * structures to the load-balancing proper:
1403 struct rq_iterator
{
1405 struct task_struct
*(*start
)(void *);
1406 struct task_struct
*(*next
)(void *);
1410 static unsigned long
1411 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1412 unsigned long max_load_move
, struct sched_domain
*sd
,
1413 enum cpu_idle_type idle
, int *all_pinned
,
1414 int *this_best_prio
, struct rq_iterator
*iterator
);
1417 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1418 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1419 struct rq_iterator
*iterator
);
1422 /* Time spent by the tasks of the cpu accounting group executing in ... */
1423 enum cpuacct_stat_index
{
1424 CPUACCT_STAT_USER
, /* ... user mode */
1425 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1427 CPUACCT_STAT_NSTATS
,
1430 #ifdef CONFIG_CGROUP_CPUACCT
1431 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1432 static void cpuacct_update_stats(struct task_struct
*tsk
,
1433 enum cpuacct_stat_index idx
, cputime_t val
);
1435 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1436 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1437 enum cpuacct_stat_index idx
, cputime_t val
) {}
1440 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1442 update_load_add(&rq
->load
, load
);
1445 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1447 update_load_sub(&rq
->load
, load
);
1450 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1451 typedef int (*tg_visitor
)(struct task_group
*, void *);
1454 * Iterate the full tree, calling @down when first entering a node and @up when
1455 * leaving it for the final time.
1457 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1459 struct task_group
*parent
, *child
;
1463 parent
= &root_task_group
;
1465 ret
= (*down
)(parent
, data
);
1468 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1475 ret
= (*up
)(parent
, data
);
1480 parent
= parent
->parent
;
1489 static int tg_nop(struct task_group
*tg
, void *data
)
1496 static unsigned long source_load(int cpu
, int type
);
1497 static unsigned long target_load(int cpu
, int type
);
1498 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1500 static unsigned long cpu_avg_load_per_task(int cpu
)
1502 struct rq
*rq
= cpu_rq(cpu
);
1503 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1506 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1508 rq
->avg_load_per_task
= 0;
1510 return rq
->avg_load_per_task
;
1513 #ifdef CONFIG_FAIR_GROUP_SCHED
1515 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1518 * Calculate and set the cpu's group shares.
1521 update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1522 unsigned long sd_shares
, unsigned long sd_rq_weight
)
1524 unsigned long shares
;
1525 unsigned long rq_weight
;
1530 rq_weight
= tg
->cfs_rq
[cpu
]->rq_weight
;
1533 * \Sum shares * rq_weight
1534 * shares = -----------------------
1538 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1539 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1541 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1542 sysctl_sched_shares_thresh
) {
1543 struct rq
*rq
= cpu_rq(cpu
);
1544 unsigned long flags
;
1546 spin_lock_irqsave(&rq
->lock
, flags
);
1547 tg
->cfs_rq
[cpu
]->shares
= shares
;
1549 __set_se_shares(tg
->se
[cpu
], shares
);
1550 spin_unlock_irqrestore(&rq
->lock
, flags
);
1555 * Re-compute the task group their per cpu shares over the given domain.
1556 * This needs to be done in a bottom-up fashion because the rq weight of a
1557 * parent group depends on the shares of its child groups.
1559 static int tg_shares_up(struct task_group
*tg
, void *data
)
1561 unsigned long weight
, rq_weight
= 0;
1562 unsigned long shares
= 0;
1563 struct sched_domain
*sd
= data
;
1566 for_each_cpu(i
, sched_domain_span(sd
)) {
1568 * If there are currently no tasks on the cpu pretend there
1569 * is one of average load so that when a new task gets to
1570 * run here it will not get delayed by group starvation.
1572 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1574 weight
= NICE_0_LOAD
;
1576 tg
->cfs_rq
[i
]->rq_weight
= weight
;
1577 rq_weight
+= weight
;
1578 shares
+= tg
->cfs_rq
[i
]->shares
;
1581 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1582 shares
= tg
->shares
;
1584 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1585 shares
= tg
->shares
;
1587 for_each_cpu(i
, sched_domain_span(sd
))
1588 update_group_shares_cpu(tg
, i
, shares
, rq_weight
);
1594 * Compute the cpu's hierarchical load factor for each task group.
1595 * This needs to be done in a top-down fashion because the load of a child
1596 * group is a fraction of its parents load.
1598 static int tg_load_down(struct task_group
*tg
, void *data
)
1601 long cpu
= (long)data
;
1604 load
= cpu_rq(cpu
)->load
.weight
;
1606 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1607 load
*= tg
->cfs_rq
[cpu
]->shares
;
1608 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1611 tg
->cfs_rq
[cpu
]->h_load
= load
;
1616 static void update_shares(struct sched_domain
*sd
)
1618 u64 now
= cpu_clock(raw_smp_processor_id());
1619 s64 elapsed
= now
- sd
->last_update
;
1621 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1622 sd
->last_update
= now
;
1623 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1627 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1629 spin_unlock(&rq
->lock
);
1631 spin_lock(&rq
->lock
);
1634 static void update_h_load(long cpu
)
1636 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1641 static inline void update_shares(struct sched_domain
*sd
)
1645 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1651 #ifdef CONFIG_PREEMPT
1654 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1655 * way at the expense of forcing extra atomic operations in all
1656 * invocations. This assures that the double_lock is acquired using the
1657 * same underlying policy as the spinlock_t on this architecture, which
1658 * reduces latency compared to the unfair variant below. However, it
1659 * also adds more overhead and therefore may reduce throughput.
1661 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1662 __releases(this_rq
->lock
)
1663 __acquires(busiest
->lock
)
1664 __acquires(this_rq
->lock
)
1666 spin_unlock(&this_rq
->lock
);
1667 double_rq_lock(this_rq
, busiest
);
1674 * Unfair double_lock_balance: Optimizes throughput at the expense of
1675 * latency by eliminating extra atomic operations when the locks are
1676 * already in proper order on entry. This favors lower cpu-ids and will
1677 * grant the double lock to lower cpus over higher ids under contention,
1678 * regardless of entry order into the function.
1680 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1681 __releases(this_rq
->lock
)
1682 __acquires(busiest
->lock
)
1683 __acquires(this_rq
->lock
)
1687 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1688 if (busiest
< this_rq
) {
1689 spin_unlock(&this_rq
->lock
);
1690 spin_lock(&busiest
->lock
);
1691 spin_lock_nested(&this_rq
->lock
, SINGLE_DEPTH_NESTING
);
1694 spin_lock_nested(&busiest
->lock
, SINGLE_DEPTH_NESTING
);
1699 #endif /* CONFIG_PREEMPT */
1702 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1704 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1706 if (unlikely(!irqs_disabled())) {
1707 /* printk() doesn't work good under rq->lock */
1708 spin_unlock(&this_rq
->lock
);
1712 return _double_lock_balance(this_rq
, busiest
);
1715 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1716 __releases(busiest
->lock
)
1718 spin_unlock(&busiest
->lock
);
1719 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1723 #ifdef CONFIG_FAIR_GROUP_SCHED
1724 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1727 cfs_rq
->shares
= shares
;
1732 static void calc_load_account_active(struct rq
*this_rq
);
1734 #include "sched_stats.h"
1735 #include "sched_idletask.c"
1736 #include "sched_fair.c"
1737 #include "sched_rt.c"
1738 #ifdef CONFIG_SCHED_DEBUG
1739 # include "sched_debug.c"
1742 #define sched_class_highest (&rt_sched_class)
1743 #define for_each_class(class) \
1744 for (class = sched_class_highest; class; class = class->next)
1746 static void inc_nr_running(struct rq
*rq
)
1751 static void dec_nr_running(struct rq
*rq
)
1756 static void set_load_weight(struct task_struct
*p
)
1758 if (task_has_rt_policy(p
)) {
1759 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1760 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1765 * SCHED_IDLE tasks get minimal weight:
1767 if (p
->policy
== SCHED_IDLE
) {
1768 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1769 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1773 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1774 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1777 static void update_avg(u64
*avg
, u64 sample
)
1779 s64 diff
= sample
- *avg
;
1783 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1786 p
->se
.start_runtime
= p
->se
.sum_exec_runtime
;
1788 sched_info_queued(p
);
1789 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1793 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1796 if (p
->se
.last_wakeup
) {
1797 update_avg(&p
->se
.avg_overlap
,
1798 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1799 p
->se
.last_wakeup
= 0;
1801 update_avg(&p
->se
.avg_wakeup
,
1802 sysctl_sched_wakeup_granularity
);
1806 sched_info_dequeued(p
);
1807 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1812 * __normal_prio - return the priority that is based on the static prio
1814 static inline int __normal_prio(struct task_struct
*p
)
1816 return p
->static_prio
;
1820 * Calculate the expected normal priority: i.e. priority
1821 * without taking RT-inheritance into account. Might be
1822 * boosted by interactivity modifiers. Changes upon fork,
1823 * setprio syscalls, and whenever the interactivity
1824 * estimator recalculates.
1826 static inline int normal_prio(struct task_struct
*p
)
1830 if (task_has_rt_policy(p
))
1831 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1833 prio
= __normal_prio(p
);
1838 * Calculate the current priority, i.e. the priority
1839 * taken into account by the scheduler. This value might
1840 * be boosted by RT tasks, or might be boosted by
1841 * interactivity modifiers. Will be RT if the task got
1842 * RT-boosted. If not then it returns p->normal_prio.
1844 static int effective_prio(struct task_struct
*p
)
1846 p
->normal_prio
= normal_prio(p
);
1848 * If we are RT tasks or we were boosted to RT priority,
1849 * keep the priority unchanged. Otherwise, update priority
1850 * to the normal priority:
1852 if (!rt_prio(p
->prio
))
1853 return p
->normal_prio
;
1858 * activate_task - move a task to the runqueue.
1860 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1862 if (task_contributes_to_load(p
))
1863 rq
->nr_uninterruptible
--;
1865 enqueue_task(rq
, p
, wakeup
);
1870 * deactivate_task - remove a task from the runqueue.
1872 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1874 if (task_contributes_to_load(p
))
1875 rq
->nr_uninterruptible
++;
1877 dequeue_task(rq
, p
, sleep
);
1882 * task_curr - is this task currently executing on a CPU?
1883 * @p: the task in question.
1885 inline int task_curr(const struct task_struct
*p
)
1887 return cpu_curr(task_cpu(p
)) == p
;
1890 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1892 set_task_rq(p
, cpu
);
1895 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1896 * successfuly executed on another CPU. We must ensure that updates of
1897 * per-task data have been completed by this moment.
1900 task_thread_info(p
)->cpu
= cpu
;
1904 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1905 const struct sched_class
*prev_class
,
1906 int oldprio
, int running
)
1908 if (prev_class
!= p
->sched_class
) {
1909 if (prev_class
->switched_from
)
1910 prev_class
->switched_from(rq
, p
, running
);
1911 p
->sched_class
->switched_to(rq
, p
, running
);
1913 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1918 /* Used instead of source_load when we know the type == 0 */
1919 static unsigned long weighted_cpuload(const int cpu
)
1921 return cpu_rq(cpu
)->load
.weight
;
1925 * Is this task likely cache-hot:
1928 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1933 * Buddy candidates are cache hot:
1935 if (sched_feat(CACHE_HOT_BUDDY
) &&
1936 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
1937 &p
->se
== cfs_rq_of(&p
->se
)->last
))
1940 if (p
->sched_class
!= &fair_sched_class
)
1943 if (sysctl_sched_migration_cost
== -1)
1945 if (sysctl_sched_migration_cost
== 0)
1948 delta
= now
- p
->se
.exec_start
;
1950 return delta
< (s64
)sysctl_sched_migration_cost
;
1954 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1956 int old_cpu
= task_cpu(p
);
1957 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1958 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1959 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1962 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1964 trace_sched_migrate_task(p
, new_cpu
);
1966 #ifdef CONFIG_SCHEDSTATS
1967 if (p
->se
.wait_start
)
1968 p
->se
.wait_start
-= clock_offset
;
1969 if (p
->se
.sleep_start
)
1970 p
->se
.sleep_start
-= clock_offset
;
1971 if (p
->se
.block_start
)
1972 p
->se
.block_start
-= clock_offset
;
1974 if (old_cpu
!= new_cpu
) {
1975 p
->se
.nr_migrations
++;
1976 new_rq
->nr_migrations_in
++;
1977 #ifdef CONFIG_SCHEDSTATS
1978 if (task_hot(p
, old_rq
->clock
, NULL
))
1979 schedstat_inc(p
, se
.nr_forced2_migrations
);
1981 perf_counter_task_migration(p
, new_cpu
);
1983 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1984 new_cfsrq
->min_vruntime
;
1986 __set_task_cpu(p
, new_cpu
);
1989 struct migration_req
{
1990 struct list_head list
;
1992 struct task_struct
*task
;
1995 struct completion done
;
1999 * The task's runqueue lock must be held.
2000 * Returns true if you have to wait for migration thread.
2003 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
2005 struct rq
*rq
= task_rq(p
);
2008 * If the task is not on a runqueue (and not running), then
2009 * it is sufficient to simply update the task's cpu field.
2011 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
2012 set_task_cpu(p
, dest_cpu
);
2016 init_completion(&req
->done
);
2018 req
->dest_cpu
= dest_cpu
;
2019 list_add(&req
->list
, &rq
->migration_queue
);
2025 * wait_task_context_switch - wait for a thread to complete at least one
2028 * @p must not be current.
2030 void wait_task_context_switch(struct task_struct
*p
)
2032 unsigned long nvcsw
, nivcsw
, flags
;
2040 * The runqueue is assigned before the actual context
2041 * switch. We need to take the runqueue lock.
2043 * We could check initially without the lock but it is
2044 * very likely that we need to take the lock in every
2047 rq
= task_rq_lock(p
, &flags
);
2048 running
= task_running(rq
, p
);
2049 task_rq_unlock(rq
, &flags
);
2051 if (likely(!running
))
2054 * The switch count is incremented before the actual
2055 * context switch. We thus wait for two switches to be
2056 * sure at least one completed.
2058 if ((p
->nvcsw
- nvcsw
) > 1)
2060 if ((p
->nivcsw
- nivcsw
) > 1)
2068 * wait_task_inactive - wait for a thread to unschedule.
2070 * If @match_state is nonzero, it's the @p->state value just checked and
2071 * not expected to change. If it changes, i.e. @p might have woken up,
2072 * then return zero. When we succeed in waiting for @p to be off its CPU,
2073 * we return a positive number (its total switch count). If a second call
2074 * a short while later returns the same number, the caller can be sure that
2075 * @p has remained unscheduled the whole time.
2077 * The caller must ensure that the task *will* unschedule sometime soon,
2078 * else this function might spin for a *long* time. This function can't
2079 * be called with interrupts off, or it may introduce deadlock with
2080 * smp_call_function() if an IPI is sent by the same process we are
2081 * waiting to become inactive.
2083 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2085 unsigned long flags
;
2092 * We do the initial early heuristics without holding
2093 * any task-queue locks at all. We'll only try to get
2094 * the runqueue lock when things look like they will
2100 * If the task is actively running on another CPU
2101 * still, just relax and busy-wait without holding
2104 * NOTE! Since we don't hold any locks, it's not
2105 * even sure that "rq" stays as the right runqueue!
2106 * But we don't care, since "task_running()" will
2107 * return false if the runqueue has changed and p
2108 * is actually now running somewhere else!
2110 while (task_running(rq
, p
)) {
2111 if (match_state
&& unlikely(p
->state
!= match_state
))
2117 * Ok, time to look more closely! We need the rq
2118 * lock now, to be *sure*. If we're wrong, we'll
2119 * just go back and repeat.
2121 rq
= task_rq_lock(p
, &flags
);
2122 trace_sched_wait_task(rq
, p
);
2123 running
= task_running(rq
, p
);
2124 on_rq
= p
->se
.on_rq
;
2126 if (!match_state
|| p
->state
== match_state
)
2127 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2128 task_rq_unlock(rq
, &flags
);
2131 * If it changed from the expected state, bail out now.
2133 if (unlikely(!ncsw
))
2137 * Was it really running after all now that we
2138 * checked with the proper locks actually held?
2140 * Oops. Go back and try again..
2142 if (unlikely(running
)) {
2148 * It's not enough that it's not actively running,
2149 * it must be off the runqueue _entirely_, and not
2152 * So if it was still runnable (but just not actively
2153 * running right now), it's preempted, and we should
2154 * yield - it could be a while.
2156 if (unlikely(on_rq
)) {
2157 schedule_timeout_uninterruptible(1);
2162 * Ahh, all good. It wasn't running, and it wasn't
2163 * runnable, which means that it will never become
2164 * running in the future either. We're all done!
2173 * kick_process - kick a running thread to enter/exit the kernel
2174 * @p: the to-be-kicked thread
2176 * Cause a process which is running on another CPU to enter
2177 * kernel-mode, without any delay. (to get signals handled.)
2179 * NOTE: this function doesnt have to take the runqueue lock,
2180 * because all it wants to ensure is that the remote task enters
2181 * the kernel. If the IPI races and the task has been migrated
2182 * to another CPU then no harm is done and the purpose has been
2185 void kick_process(struct task_struct
*p
)
2191 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2192 smp_send_reschedule(cpu
);
2195 EXPORT_SYMBOL_GPL(kick_process
);
2198 * Return a low guess at the load of a migration-source cpu weighted
2199 * according to the scheduling class and "nice" value.
2201 * We want to under-estimate the load of migration sources, to
2202 * balance conservatively.
2204 static unsigned long source_load(int cpu
, int type
)
2206 struct rq
*rq
= cpu_rq(cpu
);
2207 unsigned long total
= weighted_cpuload(cpu
);
2209 if (type
== 0 || !sched_feat(LB_BIAS
))
2212 return min(rq
->cpu_load
[type
-1], total
);
2216 * Return a high guess at the load of a migration-target cpu weighted
2217 * according to the scheduling class and "nice" value.
2219 static unsigned long target_load(int cpu
, int type
)
2221 struct rq
*rq
= cpu_rq(cpu
);
2222 unsigned long total
= weighted_cpuload(cpu
);
2224 if (type
== 0 || !sched_feat(LB_BIAS
))
2227 return max(rq
->cpu_load
[type
-1], total
);
2231 * find_idlest_group finds and returns the least busy CPU group within the
2234 static struct sched_group
*
2235 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
2237 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2238 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
2239 int load_idx
= sd
->forkexec_idx
;
2240 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
2243 unsigned long load
, avg_load
;
2247 /* Skip over this group if it has no CPUs allowed */
2248 if (!cpumask_intersects(sched_group_cpus(group
),
2252 local_group
= cpumask_test_cpu(this_cpu
,
2253 sched_group_cpus(group
));
2255 /* Tally up the load of all CPUs in the group */
2258 for_each_cpu(i
, sched_group_cpus(group
)) {
2259 /* Bias balancing toward cpus of our domain */
2261 load
= source_load(i
, load_idx
);
2263 load
= target_load(i
, load_idx
);
2268 /* Adjust by relative CPU power of the group */
2269 avg_load
= sg_div_cpu_power(group
,
2270 avg_load
* SCHED_LOAD_SCALE
);
2273 this_load
= avg_load
;
2275 } else if (avg_load
< min_load
) {
2276 min_load
= avg_load
;
2279 } while (group
= group
->next
, group
!= sd
->groups
);
2281 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
2287 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2290 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
2292 unsigned long load
, min_load
= ULONG_MAX
;
2296 /* Traverse only the allowed CPUs */
2297 for_each_cpu_and(i
, sched_group_cpus(group
), &p
->cpus_allowed
) {
2298 load
= weighted_cpuload(i
);
2300 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2310 * sched_balance_self: balance the current task (running on cpu) in domains
2311 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2314 * Balance, ie. select the least loaded group.
2316 * Returns the target CPU number, or the same CPU if no balancing is needed.
2318 * preempt must be disabled.
2320 static int sched_balance_self(int cpu
, int flag
)
2322 struct task_struct
*t
= current
;
2323 struct sched_domain
*tmp
, *sd
= NULL
;
2325 for_each_domain(cpu
, tmp
) {
2327 * If power savings logic is enabled for a domain, stop there.
2329 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2331 if (tmp
->flags
& flag
)
2339 struct sched_group
*group
;
2340 int new_cpu
, weight
;
2342 if (!(sd
->flags
& flag
)) {
2347 group
= find_idlest_group(sd
, t
, cpu
);
2353 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
2354 if (new_cpu
== -1 || new_cpu
== cpu
) {
2355 /* Now try balancing at a lower domain level of cpu */
2360 /* Now try balancing at a lower domain level of new_cpu */
2362 weight
= cpumask_weight(sched_domain_span(sd
));
2364 for_each_domain(cpu
, tmp
) {
2365 if (weight
<= cpumask_weight(sched_domain_span(tmp
)))
2367 if (tmp
->flags
& flag
)
2370 /* while loop will break here if sd == NULL */
2376 #endif /* CONFIG_SMP */
2379 * task_oncpu_function_call - call a function on the cpu on which a task runs
2380 * @p: the task to evaluate
2381 * @func: the function to be called
2382 * @info: the function call argument
2384 * Calls the function @func when the task is currently running. This might
2385 * be on the current CPU, which just calls the function directly
2387 void task_oncpu_function_call(struct task_struct
*p
,
2388 void (*func
) (void *info
), void *info
)
2395 smp_call_function_single(cpu
, func
, info
, 1);
2400 * try_to_wake_up - wake up a thread
2401 * @p: the to-be-woken-up thread
2402 * @state: the mask of task states that can be woken
2403 * @sync: do a synchronous wakeup?
2405 * Put it on the run-queue if it's not already there. The "current"
2406 * thread is always on the run-queue (except when the actual
2407 * re-schedule is in progress), and as such you're allowed to do
2408 * the simpler "current->state = TASK_RUNNING" to mark yourself
2409 * runnable without the overhead of this.
2411 * returns failure only if the task is already active.
2413 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
2415 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2416 unsigned long flags
;
2420 if (!sched_feat(SYNC_WAKEUPS
))
2424 if (sched_feat(LB_WAKEUP_UPDATE
) && !root_task_group_empty()) {
2425 struct sched_domain
*sd
;
2427 this_cpu
= raw_smp_processor_id();
2430 for_each_domain(this_cpu
, sd
) {
2431 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2440 rq
= task_rq_lock(p
, &flags
);
2441 update_rq_clock(rq
);
2442 old_state
= p
->state
;
2443 if (!(old_state
& state
))
2451 this_cpu
= smp_processor_id();
2454 if (unlikely(task_running(rq
, p
)))
2457 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2458 if (cpu
!= orig_cpu
) {
2459 set_task_cpu(p
, cpu
);
2460 task_rq_unlock(rq
, &flags
);
2461 /* might preempt at this point */
2462 rq
= task_rq_lock(p
, &flags
);
2463 old_state
= p
->state
;
2464 if (!(old_state
& state
))
2469 this_cpu
= smp_processor_id();
2473 #ifdef CONFIG_SCHEDSTATS
2474 schedstat_inc(rq
, ttwu_count
);
2475 if (cpu
== this_cpu
)
2476 schedstat_inc(rq
, ttwu_local
);
2478 struct sched_domain
*sd
;
2479 for_each_domain(this_cpu
, sd
) {
2480 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2481 schedstat_inc(sd
, ttwu_wake_remote
);
2486 #endif /* CONFIG_SCHEDSTATS */
2489 #endif /* CONFIG_SMP */
2490 schedstat_inc(p
, se
.nr_wakeups
);
2492 schedstat_inc(p
, se
.nr_wakeups_sync
);
2493 if (orig_cpu
!= cpu
)
2494 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2495 if (cpu
== this_cpu
)
2496 schedstat_inc(p
, se
.nr_wakeups_local
);
2498 schedstat_inc(p
, se
.nr_wakeups_remote
);
2499 activate_task(rq
, p
, 1);
2503 * Only attribute actual wakeups done by this task.
2505 if (!in_interrupt()) {
2506 struct sched_entity
*se
= ¤t
->se
;
2507 u64 sample
= se
->sum_exec_runtime
;
2509 if (se
->last_wakeup
)
2510 sample
-= se
->last_wakeup
;
2512 sample
-= se
->start_runtime
;
2513 update_avg(&se
->avg_wakeup
, sample
);
2515 se
->last_wakeup
= se
->sum_exec_runtime
;
2519 trace_sched_wakeup(rq
, p
, success
);
2520 check_preempt_curr(rq
, p
, sync
);
2522 p
->state
= TASK_RUNNING
;
2524 if (p
->sched_class
->task_wake_up
)
2525 p
->sched_class
->task_wake_up(rq
, p
);
2528 task_rq_unlock(rq
, &flags
);
2534 * wake_up_process - Wake up a specific process
2535 * @p: The process to be woken up.
2537 * Attempt to wake up the nominated process and move it to the set of runnable
2538 * processes. Returns 1 if the process was woken up, 0 if it was already
2541 * It may be assumed that this function implies a write memory barrier before
2542 * changing the task state if and only if any tasks are woken up.
2544 int wake_up_process(struct task_struct
*p
)
2546 return try_to_wake_up(p
, TASK_ALL
, 0);
2548 EXPORT_SYMBOL(wake_up_process
);
2550 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2552 return try_to_wake_up(p
, state
, 0);
2556 * Perform scheduler related setup for a newly forked process p.
2557 * p is forked by current.
2559 * __sched_fork() is basic setup used by init_idle() too:
2561 static void __sched_fork(struct task_struct
*p
)
2563 p
->se
.exec_start
= 0;
2564 p
->se
.sum_exec_runtime
= 0;
2565 p
->se
.prev_sum_exec_runtime
= 0;
2566 p
->se
.nr_migrations
= 0;
2567 p
->se
.last_wakeup
= 0;
2568 p
->se
.avg_overlap
= 0;
2569 p
->se
.start_runtime
= 0;
2570 p
->se
.avg_wakeup
= sysctl_sched_wakeup_granularity
;
2572 #ifdef CONFIG_SCHEDSTATS
2573 p
->se
.wait_start
= 0;
2574 p
->se
.sum_sleep_runtime
= 0;
2575 p
->se
.sleep_start
= 0;
2576 p
->se
.block_start
= 0;
2577 p
->se
.sleep_max
= 0;
2578 p
->se
.block_max
= 0;
2580 p
->se
.slice_max
= 0;
2584 INIT_LIST_HEAD(&p
->rt
.run_list
);
2586 INIT_LIST_HEAD(&p
->se
.group_node
);
2588 #ifdef CONFIG_PREEMPT_NOTIFIERS
2589 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2593 * We mark the process as running here, but have not actually
2594 * inserted it onto the runqueue yet. This guarantees that
2595 * nobody will actually run it, and a signal or other external
2596 * event cannot wake it up and insert it on the runqueue either.
2598 p
->state
= TASK_RUNNING
;
2602 * fork()/clone()-time setup:
2604 void sched_fork(struct task_struct
*p
, int clone_flags
)
2606 int cpu
= get_cpu();
2611 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2613 set_task_cpu(p
, cpu
);
2616 * Make sure we do not leak PI boosting priority to the child.
2618 p
->prio
= current
->normal_prio
;
2621 * Revert to default priority/policy on fork if requested.
2623 if (unlikely(p
->sched_reset_on_fork
)) {
2624 if (p
->policy
== SCHED_FIFO
|| p
->policy
== SCHED_RR
)
2625 p
->policy
= SCHED_NORMAL
;
2627 if (p
->normal_prio
< DEFAULT_PRIO
)
2628 p
->prio
= DEFAULT_PRIO
;
2631 * We don't need the reset flag anymore after the fork. It has
2632 * fulfilled its duty:
2634 p
->sched_reset_on_fork
= 0;
2637 if (!rt_prio(p
->prio
))
2638 p
->sched_class
= &fair_sched_class
;
2640 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2641 if (likely(sched_info_on()))
2642 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2644 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2647 #ifdef CONFIG_PREEMPT
2648 /* Want to start with kernel preemption disabled. */
2649 task_thread_info(p
)->preempt_count
= 1;
2651 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2657 * wake_up_new_task - wake up a newly created task for the first time.
2659 * This function will do some initial scheduler statistics housekeeping
2660 * that must be done for every newly created context, then puts the task
2661 * on the runqueue and wakes it.
2663 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2665 unsigned long flags
;
2668 rq
= task_rq_lock(p
, &flags
);
2669 BUG_ON(p
->state
!= TASK_RUNNING
);
2670 update_rq_clock(rq
);
2672 p
->prio
= effective_prio(p
);
2674 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2675 activate_task(rq
, p
, 0);
2678 * Let the scheduling class do new task startup
2679 * management (if any):
2681 p
->sched_class
->task_new(rq
, p
);
2684 trace_sched_wakeup_new(rq
, p
, 1);
2685 check_preempt_curr(rq
, p
, 0);
2687 if (p
->sched_class
->task_wake_up
)
2688 p
->sched_class
->task_wake_up(rq
, p
);
2690 task_rq_unlock(rq
, &flags
);
2693 #ifdef CONFIG_PREEMPT_NOTIFIERS
2696 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2697 * @notifier: notifier struct to register
2699 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2701 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2703 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2706 * preempt_notifier_unregister - no longer interested in preemption notifications
2707 * @notifier: notifier struct to unregister
2709 * This is safe to call from within a preemption notifier.
2711 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2713 hlist_del(¬ifier
->link
);
2715 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2717 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2719 struct preempt_notifier
*notifier
;
2720 struct hlist_node
*node
;
2722 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2723 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2727 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2728 struct task_struct
*next
)
2730 struct preempt_notifier
*notifier
;
2731 struct hlist_node
*node
;
2733 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2734 notifier
->ops
->sched_out(notifier
, next
);
2737 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2739 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2744 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2745 struct task_struct
*next
)
2749 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2752 * prepare_task_switch - prepare to switch tasks
2753 * @rq: the runqueue preparing to switch
2754 * @prev: the current task that is being switched out
2755 * @next: the task we are going to switch to.
2757 * This is called with the rq lock held and interrupts off. It must
2758 * be paired with a subsequent finish_task_switch after the context
2761 * prepare_task_switch sets up locking and calls architecture specific
2765 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2766 struct task_struct
*next
)
2768 fire_sched_out_preempt_notifiers(prev
, next
);
2769 prepare_lock_switch(rq
, next
);
2770 prepare_arch_switch(next
);
2774 * finish_task_switch - clean up after a task-switch
2775 * @rq: runqueue associated with task-switch
2776 * @prev: the thread we just switched away from.
2778 * finish_task_switch must be called after the context switch, paired
2779 * with a prepare_task_switch call before the context switch.
2780 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2781 * and do any other architecture-specific cleanup actions.
2783 * Note that we may have delayed dropping an mm in context_switch(). If
2784 * so, we finish that here outside of the runqueue lock. (Doing it
2785 * with the lock held can cause deadlocks; see schedule() for
2788 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2789 __releases(rq
->lock
)
2791 struct mm_struct
*mm
= rq
->prev_mm
;
2794 int post_schedule
= 0;
2796 if (current
->sched_class
->needs_post_schedule
)
2797 post_schedule
= current
->sched_class
->needs_post_schedule(rq
);
2803 * A task struct has one reference for the use as "current".
2804 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2805 * schedule one last time. The schedule call will never return, and
2806 * the scheduled task must drop that reference.
2807 * The test for TASK_DEAD must occur while the runqueue locks are
2808 * still held, otherwise prev could be scheduled on another cpu, die
2809 * there before we look at prev->state, and then the reference would
2811 * Manfred Spraul <manfred@colorfullife.com>
2813 prev_state
= prev
->state
;
2814 finish_arch_switch(prev
);
2815 perf_counter_task_sched_in(current
, cpu_of(rq
));
2816 finish_lock_switch(rq
, prev
);
2819 current
->sched_class
->post_schedule(rq
);
2822 fire_sched_in_preempt_notifiers(current
);
2825 if (unlikely(prev_state
== TASK_DEAD
)) {
2827 * Remove function-return probe instances associated with this
2828 * task and put them back on the free list.
2830 kprobe_flush_task(prev
);
2831 put_task_struct(prev
);
2836 * schedule_tail - first thing a freshly forked thread must call.
2837 * @prev: the thread we just switched away from.
2839 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2840 __releases(rq
->lock
)
2842 struct rq
*rq
= this_rq();
2844 finish_task_switch(rq
, prev
);
2845 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2846 /* In this case, finish_task_switch does not reenable preemption */
2849 if (current
->set_child_tid
)
2850 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2854 * context_switch - switch to the new MM and the new
2855 * thread's register state.
2858 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2859 struct task_struct
*next
)
2861 struct mm_struct
*mm
, *oldmm
;
2863 prepare_task_switch(rq
, prev
, next
);
2864 trace_sched_switch(rq
, prev
, next
);
2866 oldmm
= prev
->active_mm
;
2868 * For paravirt, this is coupled with an exit in switch_to to
2869 * combine the page table reload and the switch backend into
2872 arch_start_context_switch(prev
);
2874 if (unlikely(!mm
)) {
2875 next
->active_mm
= oldmm
;
2876 atomic_inc(&oldmm
->mm_count
);
2877 enter_lazy_tlb(oldmm
, next
);
2879 switch_mm(oldmm
, mm
, next
);
2881 if (unlikely(!prev
->mm
)) {
2882 prev
->active_mm
= NULL
;
2883 rq
->prev_mm
= oldmm
;
2886 * Since the runqueue lock will be released by the next
2887 * task (which is an invalid locking op but in the case
2888 * of the scheduler it's an obvious special-case), so we
2889 * do an early lockdep release here:
2891 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2892 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2895 /* Here we just switch the register state and the stack. */
2896 switch_to(prev
, next
, prev
);
2900 * this_rq must be evaluated again because prev may have moved
2901 * CPUs since it called schedule(), thus the 'rq' on its stack
2902 * frame will be invalid.
2904 finish_task_switch(this_rq(), prev
);
2908 * nr_running, nr_uninterruptible and nr_context_switches:
2910 * externally visible scheduler statistics: current number of runnable
2911 * threads, current number of uninterruptible-sleeping threads, total
2912 * number of context switches performed since bootup.
2914 unsigned long nr_running(void)
2916 unsigned long i
, sum
= 0;
2918 for_each_online_cpu(i
)
2919 sum
+= cpu_rq(i
)->nr_running
;
2924 unsigned long nr_uninterruptible(void)
2926 unsigned long i
, sum
= 0;
2928 for_each_possible_cpu(i
)
2929 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2932 * Since we read the counters lockless, it might be slightly
2933 * inaccurate. Do not allow it to go below zero though:
2935 if (unlikely((long)sum
< 0))
2941 unsigned long long nr_context_switches(void)
2944 unsigned long long sum
= 0;
2946 for_each_possible_cpu(i
)
2947 sum
+= cpu_rq(i
)->nr_switches
;
2952 unsigned long nr_iowait(void)
2954 unsigned long i
, sum
= 0;
2956 for_each_possible_cpu(i
)
2957 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2962 /* Variables and functions for calc_load */
2963 static atomic_long_t calc_load_tasks
;
2964 static unsigned long calc_load_update
;
2965 unsigned long avenrun
[3];
2966 EXPORT_SYMBOL(avenrun
);
2969 * get_avenrun - get the load average array
2970 * @loads: pointer to dest load array
2971 * @offset: offset to add
2972 * @shift: shift count to shift the result left
2974 * These values are estimates at best, so no need for locking.
2976 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
2978 loads
[0] = (avenrun
[0] + offset
) << shift
;
2979 loads
[1] = (avenrun
[1] + offset
) << shift
;
2980 loads
[2] = (avenrun
[2] + offset
) << shift
;
2983 static unsigned long
2984 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
2987 load
+= active
* (FIXED_1
- exp
);
2988 return load
>> FSHIFT
;
2992 * calc_load - update the avenrun load estimates 10 ticks after the
2993 * CPUs have updated calc_load_tasks.
2995 void calc_global_load(void)
2997 unsigned long upd
= calc_load_update
+ 10;
3000 if (time_before(jiffies
, upd
))
3003 active
= atomic_long_read(&calc_load_tasks
);
3004 active
= active
> 0 ? active
* FIXED_1
: 0;
3006 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
3007 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
3008 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
3010 calc_load_update
+= LOAD_FREQ
;
3014 * Either called from update_cpu_load() or from a cpu going idle
3016 static void calc_load_account_active(struct rq
*this_rq
)
3018 long nr_active
, delta
;
3020 nr_active
= this_rq
->nr_running
;
3021 nr_active
+= (long) this_rq
->nr_uninterruptible
;
3023 if (nr_active
!= this_rq
->calc_load_active
) {
3024 delta
= nr_active
- this_rq
->calc_load_active
;
3025 this_rq
->calc_load_active
= nr_active
;
3026 atomic_long_add(delta
, &calc_load_tasks
);
3031 * Externally visible per-cpu scheduler statistics:
3032 * cpu_nr_migrations(cpu) - number of migrations into that cpu
3034 u64
cpu_nr_migrations(int cpu
)
3036 return cpu_rq(cpu
)->nr_migrations_in
;
3040 * Update rq->cpu_load[] statistics. This function is usually called every
3041 * scheduler tick (TICK_NSEC).
3043 static void update_cpu_load(struct rq
*this_rq
)
3045 unsigned long this_load
= this_rq
->load
.weight
;
3048 this_rq
->nr_load_updates
++;
3050 /* Update our load: */
3051 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3052 unsigned long old_load
, new_load
;
3054 /* scale is effectively 1 << i now, and >> i divides by scale */
3056 old_load
= this_rq
->cpu_load
[i
];
3057 new_load
= this_load
;
3059 * Round up the averaging division if load is increasing. This
3060 * prevents us from getting stuck on 9 if the load is 10, for
3063 if (new_load
> old_load
)
3064 new_load
+= scale
-1;
3065 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
3068 if (time_after_eq(jiffies
, this_rq
->calc_load_update
)) {
3069 this_rq
->calc_load_update
+= LOAD_FREQ
;
3070 calc_load_account_active(this_rq
);
3077 * double_rq_lock - safely lock two runqueues
3079 * Note this does not disable interrupts like task_rq_lock,
3080 * you need to do so manually before calling.
3082 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
3083 __acquires(rq1
->lock
)
3084 __acquires(rq2
->lock
)
3086 BUG_ON(!irqs_disabled());
3088 spin_lock(&rq1
->lock
);
3089 __acquire(rq2
->lock
); /* Fake it out ;) */
3092 spin_lock(&rq1
->lock
);
3093 spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
3095 spin_lock(&rq2
->lock
);
3096 spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
3099 update_rq_clock(rq1
);
3100 update_rq_clock(rq2
);
3104 * double_rq_unlock - safely unlock two runqueues
3106 * Note this does not restore interrupts like task_rq_unlock,
3107 * you need to do so manually after calling.
3109 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
3110 __releases(rq1
->lock
)
3111 __releases(rq2
->lock
)
3113 spin_unlock(&rq1
->lock
);
3115 spin_unlock(&rq2
->lock
);
3117 __release(rq2
->lock
);
3121 * If dest_cpu is allowed for this process, migrate the task to it.
3122 * This is accomplished by forcing the cpu_allowed mask to only
3123 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3124 * the cpu_allowed mask is restored.
3126 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
3128 struct migration_req req
;
3129 unsigned long flags
;
3132 rq
= task_rq_lock(p
, &flags
);
3133 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
)
3134 || unlikely(!cpu_active(dest_cpu
)))
3137 /* force the process onto the specified CPU */
3138 if (migrate_task(p
, dest_cpu
, &req
)) {
3139 /* Need to wait for migration thread (might exit: take ref). */
3140 struct task_struct
*mt
= rq
->migration_thread
;
3142 get_task_struct(mt
);
3143 task_rq_unlock(rq
, &flags
);
3144 wake_up_process(mt
);
3145 put_task_struct(mt
);
3146 wait_for_completion(&req
.done
);
3151 task_rq_unlock(rq
, &flags
);
3155 * sched_exec - execve() is a valuable balancing opportunity, because at
3156 * this point the task has the smallest effective memory and cache footprint.
3158 void sched_exec(void)
3160 int new_cpu
, this_cpu
= get_cpu();
3161 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
3163 if (new_cpu
!= this_cpu
)
3164 sched_migrate_task(current
, new_cpu
);
3168 * pull_task - move a task from a remote runqueue to the local runqueue.
3169 * Both runqueues must be locked.
3171 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
3172 struct rq
*this_rq
, int this_cpu
)
3174 deactivate_task(src_rq
, p
, 0);
3175 set_task_cpu(p
, this_cpu
);
3176 activate_task(this_rq
, p
, 0);
3178 * Note that idle threads have a prio of MAX_PRIO, for this test
3179 * to be always true for them.
3181 check_preempt_curr(this_rq
, p
, 0);
3185 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3188 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
3189 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3192 int tsk_cache_hot
= 0;
3194 * We do not migrate tasks that are:
3195 * 1) running (obviously), or
3196 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3197 * 3) are cache-hot on their current CPU.
3199 if (!cpumask_test_cpu(this_cpu
, &p
->cpus_allowed
)) {
3200 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
3205 if (task_running(rq
, p
)) {
3206 schedstat_inc(p
, se
.nr_failed_migrations_running
);
3211 * Aggressive migration if:
3212 * 1) task is cache cold, or
3213 * 2) too many balance attempts have failed.
3216 tsk_cache_hot
= task_hot(p
, rq
->clock
, sd
);
3217 if (!tsk_cache_hot
||
3218 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
3219 #ifdef CONFIG_SCHEDSTATS
3220 if (tsk_cache_hot
) {
3221 schedstat_inc(sd
, lb_hot_gained
[idle
]);
3222 schedstat_inc(p
, se
.nr_forced_migrations
);
3228 if (tsk_cache_hot
) {
3229 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
3235 static unsigned long
3236 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3237 unsigned long max_load_move
, struct sched_domain
*sd
,
3238 enum cpu_idle_type idle
, int *all_pinned
,
3239 int *this_best_prio
, struct rq_iterator
*iterator
)
3241 int loops
= 0, pulled
= 0, pinned
= 0;
3242 struct task_struct
*p
;
3243 long rem_load_move
= max_load_move
;
3245 if (max_load_move
== 0)
3251 * Start the load-balancing iterator:
3253 p
= iterator
->start(iterator
->arg
);
3255 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
3258 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
3259 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3260 p
= iterator
->next(iterator
->arg
);
3264 pull_task(busiest
, p
, this_rq
, this_cpu
);
3266 rem_load_move
-= p
->se
.load
.weight
;
3268 #ifdef CONFIG_PREEMPT
3270 * NEWIDLE balancing is a source of latency, so preemptible kernels
3271 * will stop after the first task is pulled to minimize the critical
3274 if (idle
== CPU_NEWLY_IDLE
)
3279 * We only want to steal up to the prescribed amount of weighted load.
3281 if (rem_load_move
> 0) {
3282 if (p
->prio
< *this_best_prio
)
3283 *this_best_prio
= p
->prio
;
3284 p
= iterator
->next(iterator
->arg
);
3289 * Right now, this is one of only two places pull_task() is called,
3290 * so we can safely collect pull_task() stats here rather than
3291 * inside pull_task().
3293 schedstat_add(sd
, lb_gained
[idle
], pulled
);
3296 *all_pinned
= pinned
;
3298 return max_load_move
- rem_load_move
;
3302 * move_tasks tries to move up to max_load_move weighted load from busiest to
3303 * this_rq, as part of a balancing operation within domain "sd".
3304 * Returns 1 if successful and 0 otherwise.
3306 * Called with both runqueues locked.
3308 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3309 unsigned long max_load_move
,
3310 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3313 const struct sched_class
*class = sched_class_highest
;
3314 unsigned long total_load_moved
= 0;
3315 int this_best_prio
= this_rq
->curr
->prio
;
3319 class->load_balance(this_rq
, this_cpu
, busiest
,
3320 max_load_move
- total_load_moved
,
3321 sd
, idle
, all_pinned
, &this_best_prio
);
3322 class = class->next
;
3324 #ifdef CONFIG_PREEMPT
3326 * NEWIDLE balancing is a source of latency, so preemptible
3327 * kernels will stop after the first task is pulled to minimize
3328 * the critical section.
3330 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3333 } while (class && max_load_move
> total_load_moved
);
3335 return total_load_moved
> 0;
3339 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3340 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3341 struct rq_iterator
*iterator
)
3343 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3347 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3348 pull_task(busiest
, p
, this_rq
, this_cpu
);
3350 * Right now, this is only the second place pull_task()
3351 * is called, so we can safely collect pull_task()
3352 * stats here rather than inside pull_task().
3354 schedstat_inc(sd
, lb_gained
[idle
]);
3358 p
= iterator
->next(iterator
->arg
);
3365 * move_one_task tries to move exactly one task from busiest to this_rq, as
3366 * part of active balancing operations within "domain".
3367 * Returns 1 if successful and 0 otherwise.
3369 * Called with both runqueues locked.
3371 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3372 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3374 const struct sched_class
*class;
3376 for (class = sched_class_highest
; class; class = class->next
)
3377 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3382 /********** Helpers for find_busiest_group ************************/
3384 * sd_lb_stats - Structure to store the statistics of a sched_domain
3385 * during load balancing.
3387 struct sd_lb_stats
{
3388 struct sched_group
*busiest
; /* Busiest group in this sd */
3389 struct sched_group
*this; /* Local group in this sd */
3390 unsigned long total_load
; /* Total load of all groups in sd */
3391 unsigned long total_pwr
; /* Total power of all groups in sd */
3392 unsigned long avg_load
; /* Average load across all groups in sd */
3394 /** Statistics of this group */
3395 unsigned long this_load
;
3396 unsigned long this_load_per_task
;
3397 unsigned long this_nr_running
;
3399 /* Statistics of the busiest group */
3400 unsigned long max_load
;
3401 unsigned long busiest_load_per_task
;
3402 unsigned long busiest_nr_running
;
3404 int group_imb
; /* Is there imbalance in this sd */
3405 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3406 int power_savings_balance
; /* Is powersave balance needed for this sd */
3407 struct sched_group
*group_min
; /* Least loaded group in sd */
3408 struct sched_group
*group_leader
; /* Group which relieves group_min */
3409 unsigned long min_load_per_task
; /* load_per_task in group_min */
3410 unsigned long leader_nr_running
; /* Nr running of group_leader */
3411 unsigned long min_nr_running
; /* Nr running of group_min */
3416 * sg_lb_stats - stats of a sched_group required for load_balancing
3418 struct sg_lb_stats
{
3419 unsigned long avg_load
; /*Avg load across the CPUs of the group */
3420 unsigned long group_load
; /* Total load over the CPUs of the group */
3421 unsigned long sum_nr_running
; /* Nr tasks running in the group */
3422 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
3423 unsigned long group_capacity
;
3424 int group_imb
; /* Is there an imbalance in the group ? */
3428 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3429 * @group: The group whose first cpu is to be returned.
3431 static inline unsigned int group_first_cpu(struct sched_group
*group
)
3433 return cpumask_first(sched_group_cpus(group
));
3437 * get_sd_load_idx - Obtain the load index for a given sched domain.
3438 * @sd: The sched_domain whose load_idx is to be obtained.
3439 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3441 static inline int get_sd_load_idx(struct sched_domain
*sd
,
3442 enum cpu_idle_type idle
)
3448 load_idx
= sd
->busy_idx
;
3451 case CPU_NEWLY_IDLE
:
3452 load_idx
= sd
->newidle_idx
;
3455 load_idx
= sd
->idle_idx
;
3463 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3465 * init_sd_power_savings_stats - Initialize power savings statistics for
3466 * the given sched_domain, during load balancing.
3468 * @sd: Sched domain whose power-savings statistics are to be initialized.
3469 * @sds: Variable containing the statistics for sd.
3470 * @idle: Idle status of the CPU at which we're performing load-balancing.
3472 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3473 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3476 * Busy processors will not participate in power savings
3479 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3480 sds
->power_savings_balance
= 0;
3482 sds
->power_savings_balance
= 1;
3483 sds
->min_nr_running
= ULONG_MAX
;
3484 sds
->leader_nr_running
= 0;
3489 * update_sd_power_savings_stats - Update the power saving stats for a
3490 * sched_domain while performing load balancing.
3492 * @group: sched_group belonging to the sched_domain under consideration.
3493 * @sds: Variable containing the statistics of the sched_domain
3494 * @local_group: Does group contain the CPU for which we're performing
3496 * @sgs: Variable containing the statistics of the group.
3498 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3499 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3502 if (!sds
->power_savings_balance
)
3506 * If the local group is idle or completely loaded
3507 * no need to do power savings balance at this domain
3509 if (local_group
&& (sds
->this_nr_running
>= sgs
->group_capacity
||
3510 !sds
->this_nr_running
))
3511 sds
->power_savings_balance
= 0;
3514 * If a group is already running at full capacity or idle,
3515 * don't include that group in power savings calculations
3517 if (!sds
->power_savings_balance
||
3518 sgs
->sum_nr_running
>= sgs
->group_capacity
||
3519 !sgs
->sum_nr_running
)
3523 * Calculate the group which has the least non-idle load.
3524 * This is the group from where we need to pick up the load
3527 if ((sgs
->sum_nr_running
< sds
->min_nr_running
) ||
3528 (sgs
->sum_nr_running
== sds
->min_nr_running
&&
3529 group_first_cpu(group
) > group_first_cpu(sds
->group_min
))) {
3530 sds
->group_min
= group
;
3531 sds
->min_nr_running
= sgs
->sum_nr_running
;
3532 sds
->min_load_per_task
= sgs
->sum_weighted_load
/
3533 sgs
->sum_nr_running
;
3537 * Calculate the group which is almost near its
3538 * capacity but still has some space to pick up some load
3539 * from other group and save more power
3541 if (sgs
->sum_nr_running
> sgs
->group_capacity
- 1)
3544 if (sgs
->sum_nr_running
> sds
->leader_nr_running
||
3545 (sgs
->sum_nr_running
== sds
->leader_nr_running
&&
3546 group_first_cpu(group
) < group_first_cpu(sds
->group_leader
))) {
3547 sds
->group_leader
= group
;
3548 sds
->leader_nr_running
= sgs
->sum_nr_running
;
3553 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3554 * @sds: Variable containing the statistics of the sched_domain
3555 * under consideration.
3556 * @this_cpu: Cpu at which we're currently performing load-balancing.
3557 * @imbalance: Variable to store the imbalance.
3560 * Check if we have potential to perform some power-savings balance.
3561 * If yes, set the busiest group to be the least loaded group in the
3562 * sched_domain, so that it's CPUs can be put to idle.
3564 * Returns 1 if there is potential to perform power-savings balance.
3567 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3568 int this_cpu
, unsigned long *imbalance
)
3570 if (!sds
->power_savings_balance
)
3573 if (sds
->this != sds
->group_leader
||
3574 sds
->group_leader
== sds
->group_min
)
3577 *imbalance
= sds
->min_load_per_task
;
3578 sds
->busiest
= sds
->group_min
;
3580 if (sched_mc_power_savings
>= POWERSAVINGS_BALANCE_WAKEUP
) {
3581 cpu_rq(this_cpu
)->rd
->sched_mc_preferred_wakeup_cpu
=
3582 group_first_cpu(sds
->group_leader
);
3588 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3589 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3590 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3595 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3596 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3601 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3602 int this_cpu
, unsigned long *imbalance
)
3606 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3610 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3611 * @group: sched_group whose statistics are to be updated.
3612 * @this_cpu: Cpu for which load balance is currently performed.
3613 * @idle: Idle status of this_cpu
3614 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3615 * @sd_idle: Idle status of the sched_domain containing group.
3616 * @local_group: Does group contain this_cpu.
3617 * @cpus: Set of cpus considered for load balancing.
3618 * @balance: Should we balance.
3619 * @sgs: variable to hold the statistics for this group.
3621 static inline void update_sg_lb_stats(struct sched_group
*group
, int this_cpu
,
3622 enum cpu_idle_type idle
, int load_idx
, int *sd_idle
,
3623 int local_group
, const struct cpumask
*cpus
,
3624 int *balance
, struct sg_lb_stats
*sgs
)
3626 unsigned long load
, max_cpu_load
, min_cpu_load
;
3628 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3629 unsigned long sum_avg_load_per_task
;
3630 unsigned long avg_load_per_task
;
3633 balance_cpu
= group_first_cpu(group
);
3635 /* Tally up the load of all CPUs in the group */
3636 sum_avg_load_per_task
= avg_load_per_task
= 0;
3638 min_cpu_load
= ~0UL;
3640 for_each_cpu_and(i
, sched_group_cpus(group
), cpus
) {
3641 struct rq
*rq
= cpu_rq(i
);
3643 if (*sd_idle
&& rq
->nr_running
)
3646 /* Bias balancing toward cpus of our domain */
3648 if (idle_cpu(i
) && !first_idle_cpu
) {
3653 load
= target_load(i
, load_idx
);
3655 load
= source_load(i
, load_idx
);
3656 if (load
> max_cpu_load
)
3657 max_cpu_load
= load
;
3658 if (min_cpu_load
> load
)
3659 min_cpu_load
= load
;
3662 sgs
->group_load
+= load
;
3663 sgs
->sum_nr_running
+= rq
->nr_running
;
3664 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
3666 sum_avg_load_per_task
+= cpu_avg_load_per_task(i
);
3670 * First idle cpu or the first cpu(busiest) in this sched group
3671 * is eligible for doing load balancing at this and above
3672 * domains. In the newly idle case, we will allow all the cpu's
3673 * to do the newly idle load balance.
3675 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3676 balance_cpu
!= this_cpu
&& balance
) {
3681 /* Adjust by relative CPU power of the group */
3682 sgs
->avg_load
= sg_div_cpu_power(group
,
3683 sgs
->group_load
* SCHED_LOAD_SCALE
);
3687 * Consider the group unbalanced when the imbalance is larger
3688 * than the average weight of two tasks.
3690 * APZ: with cgroup the avg task weight can vary wildly and
3691 * might not be a suitable number - should we keep a
3692 * normalized nr_running number somewhere that negates
3695 avg_load_per_task
= sg_div_cpu_power(group
,
3696 sum_avg_load_per_task
* SCHED_LOAD_SCALE
);
3698 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3701 sgs
->group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
3706 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3707 * @sd: sched_domain whose statistics are to be updated.
3708 * @this_cpu: Cpu for which load balance is currently performed.
3709 * @idle: Idle status of this_cpu
3710 * @sd_idle: Idle status of the sched_domain containing group.
3711 * @cpus: Set of cpus considered for load balancing.
3712 * @balance: Should we balance.
3713 * @sds: variable to hold the statistics for this sched_domain.
3715 static inline void update_sd_lb_stats(struct sched_domain
*sd
, int this_cpu
,
3716 enum cpu_idle_type idle
, int *sd_idle
,
3717 const struct cpumask
*cpus
, int *balance
,
3718 struct sd_lb_stats
*sds
)
3720 struct sched_group
*group
= sd
->groups
;
3721 struct sg_lb_stats sgs
;
3724 init_sd_power_savings_stats(sd
, sds
, idle
);
3725 load_idx
= get_sd_load_idx(sd
, idle
);
3730 local_group
= cpumask_test_cpu(this_cpu
,
3731 sched_group_cpus(group
));
3732 memset(&sgs
, 0, sizeof(sgs
));
3733 update_sg_lb_stats(group
, this_cpu
, idle
, load_idx
, sd_idle
,
3734 local_group
, cpus
, balance
, &sgs
);
3736 if (local_group
&& balance
&& !(*balance
))
3739 sds
->total_load
+= sgs
.group_load
;
3740 sds
->total_pwr
+= group
->__cpu_power
;
3743 sds
->this_load
= sgs
.avg_load
;
3745 sds
->this_nr_running
= sgs
.sum_nr_running
;
3746 sds
->this_load_per_task
= sgs
.sum_weighted_load
;
3747 } else if (sgs
.avg_load
> sds
->max_load
&&
3748 (sgs
.sum_nr_running
> sgs
.group_capacity
||
3750 sds
->max_load
= sgs
.avg_load
;
3751 sds
->busiest
= group
;
3752 sds
->busiest_nr_running
= sgs
.sum_nr_running
;
3753 sds
->busiest_load_per_task
= sgs
.sum_weighted_load
;
3754 sds
->group_imb
= sgs
.group_imb
;
3757 update_sd_power_savings_stats(group
, sds
, local_group
, &sgs
);
3758 group
= group
->next
;
3759 } while (group
!= sd
->groups
);
3764 * fix_small_imbalance - Calculate the minor imbalance that exists
3765 * amongst the groups of a sched_domain, during
3767 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3768 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3769 * @imbalance: Variable to store the imbalance.
3771 static inline void fix_small_imbalance(struct sd_lb_stats
*sds
,
3772 int this_cpu
, unsigned long *imbalance
)
3774 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
3775 unsigned int imbn
= 2;
3777 if (sds
->this_nr_running
) {
3778 sds
->this_load_per_task
/= sds
->this_nr_running
;
3779 if (sds
->busiest_load_per_task
>
3780 sds
->this_load_per_task
)
3783 sds
->this_load_per_task
=
3784 cpu_avg_load_per_task(this_cpu
);
3786 if (sds
->max_load
- sds
->this_load
+ sds
->busiest_load_per_task
>=
3787 sds
->busiest_load_per_task
* imbn
) {
3788 *imbalance
= sds
->busiest_load_per_task
;
3793 * OK, we don't have enough imbalance to justify moving tasks,
3794 * however we may be able to increase total CPU power used by
3798 pwr_now
+= sds
->busiest
->__cpu_power
*
3799 min(sds
->busiest_load_per_task
, sds
->max_load
);
3800 pwr_now
+= sds
->this->__cpu_power
*
3801 min(sds
->this_load_per_task
, sds
->this_load
);
3802 pwr_now
/= SCHED_LOAD_SCALE
;
3804 /* Amount of load we'd subtract */
3805 tmp
= sg_div_cpu_power(sds
->busiest
,
3806 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
);
3807 if (sds
->max_load
> tmp
)
3808 pwr_move
+= sds
->busiest
->__cpu_power
*
3809 min(sds
->busiest_load_per_task
, sds
->max_load
- tmp
);
3811 /* Amount of load we'd add */
3812 if (sds
->max_load
* sds
->busiest
->__cpu_power
<
3813 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
)
3814 tmp
= sg_div_cpu_power(sds
->this,
3815 sds
->max_load
* sds
->busiest
->__cpu_power
);
3817 tmp
= sg_div_cpu_power(sds
->this,
3818 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
);
3819 pwr_move
+= sds
->this->__cpu_power
*
3820 min(sds
->this_load_per_task
, sds
->this_load
+ tmp
);
3821 pwr_move
/= SCHED_LOAD_SCALE
;
3823 /* Move if we gain throughput */
3824 if (pwr_move
> pwr_now
)
3825 *imbalance
= sds
->busiest_load_per_task
;
3829 * calculate_imbalance - Calculate the amount of imbalance present within the
3830 * groups of a given sched_domain during load balance.
3831 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3832 * @this_cpu: Cpu for which currently load balance is being performed.
3833 * @imbalance: The variable to store the imbalance.
3835 static inline void calculate_imbalance(struct sd_lb_stats
*sds
, int this_cpu
,
3836 unsigned long *imbalance
)
3838 unsigned long max_pull
;
3840 * In the presence of smp nice balancing, certain scenarios can have
3841 * max load less than avg load(as we skip the groups at or below
3842 * its cpu_power, while calculating max_load..)
3844 if (sds
->max_load
< sds
->avg_load
) {
3846 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3849 /* Don't want to pull so many tasks that a group would go idle */
3850 max_pull
= min(sds
->max_load
- sds
->avg_load
,
3851 sds
->max_load
- sds
->busiest_load_per_task
);
3853 /* How much load to actually move to equalise the imbalance */
3854 *imbalance
= min(max_pull
* sds
->busiest
->__cpu_power
,
3855 (sds
->avg_load
- sds
->this_load
) * sds
->this->__cpu_power
)
3859 * if *imbalance is less than the average load per runnable task
3860 * there is no gaurantee that any tasks will be moved so we'll have
3861 * a think about bumping its value to force at least one task to be
3864 if (*imbalance
< sds
->busiest_load_per_task
)
3865 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3868 /******* find_busiest_group() helpers end here *********************/
3871 * find_busiest_group - Returns the busiest group within the sched_domain
3872 * if there is an imbalance. If there isn't an imbalance, and
3873 * the user has opted for power-savings, it returns a group whose
3874 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3875 * such a group exists.
3877 * Also calculates the amount of weighted load which should be moved
3878 * to restore balance.
3880 * @sd: The sched_domain whose busiest group is to be returned.
3881 * @this_cpu: The cpu for which load balancing is currently being performed.
3882 * @imbalance: Variable which stores amount of weighted load which should
3883 * be moved to restore balance/put a group to idle.
3884 * @idle: The idle status of this_cpu.
3885 * @sd_idle: The idleness of sd
3886 * @cpus: The set of CPUs under consideration for load-balancing.
3887 * @balance: Pointer to a variable indicating if this_cpu
3888 * is the appropriate cpu to perform load balancing at this_level.
3890 * Returns: - the busiest group if imbalance exists.
3891 * - If no imbalance and user has opted for power-savings balance,
3892 * return the least loaded group whose CPUs can be
3893 * put to idle by rebalancing its tasks onto our group.
3895 static struct sched_group
*
3896 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
3897 unsigned long *imbalance
, enum cpu_idle_type idle
,
3898 int *sd_idle
, const struct cpumask
*cpus
, int *balance
)
3900 struct sd_lb_stats sds
;
3902 memset(&sds
, 0, sizeof(sds
));
3905 * Compute the various statistics relavent for load balancing at
3908 update_sd_lb_stats(sd
, this_cpu
, idle
, sd_idle
, cpus
,
3911 /* Cases where imbalance does not exist from POV of this_cpu */
3912 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3914 * 2) There is no busy sibling group to pull from.
3915 * 3) This group is the busiest group.
3916 * 4) This group is more busy than the avg busieness at this
3918 * 5) The imbalance is within the specified limit.
3919 * 6) Any rebalance would lead to ping-pong
3921 if (balance
&& !(*balance
))
3924 if (!sds
.busiest
|| sds
.busiest_nr_running
== 0)
3927 if (sds
.this_load
>= sds
.max_load
)
3930 sds
.avg_load
= (SCHED_LOAD_SCALE
* sds
.total_load
) / sds
.total_pwr
;
3932 if (sds
.this_load
>= sds
.avg_load
)
3935 if (100 * sds
.max_load
<= sd
->imbalance_pct
* sds
.this_load
)
3938 sds
.busiest_load_per_task
/= sds
.busiest_nr_running
;
3940 sds
.busiest_load_per_task
=
3941 min(sds
.busiest_load_per_task
, sds
.avg_load
);
3944 * We're trying to get all the cpus to the average_load, so we don't
3945 * want to push ourselves above the average load, nor do we wish to
3946 * reduce the max loaded cpu below the average load, as either of these
3947 * actions would just result in more rebalancing later, and ping-pong
3948 * tasks around. Thus we look for the minimum possible imbalance.
3949 * Negative imbalances (*we* are more loaded than anyone else) will
3950 * be counted as no imbalance for these purposes -- we can't fix that
3951 * by pulling tasks to us. Be careful of negative numbers as they'll
3952 * appear as very large values with unsigned longs.
3954 if (sds
.max_load
<= sds
.busiest_load_per_task
)
3957 /* Looks like there is an imbalance. Compute it */
3958 calculate_imbalance(&sds
, this_cpu
, imbalance
);
3963 * There is no obvious imbalance. But check if we can do some balancing
3966 if (check_power_save_busiest_group(&sds
, this_cpu
, imbalance
))
3974 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3977 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
3978 unsigned long imbalance
, const struct cpumask
*cpus
)
3980 struct rq
*busiest
= NULL
, *rq
;
3981 unsigned long max_load
= 0;
3984 for_each_cpu(i
, sched_group_cpus(group
)) {
3987 if (!cpumask_test_cpu(i
, cpus
))
3991 wl
= weighted_cpuload(i
);
3993 if (rq
->nr_running
== 1 && wl
> imbalance
)
3996 if (wl
> max_load
) {
4006 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4007 * so long as it is large enough.
4009 #define MAX_PINNED_INTERVAL 512
4011 /* Working cpumask for load_balance and load_balance_newidle. */
4012 static DEFINE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
4015 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4016 * tasks if there is an imbalance.
4018 static int load_balance(int this_cpu
, struct rq
*this_rq
,
4019 struct sched_domain
*sd
, enum cpu_idle_type idle
,
4022 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
4023 struct sched_group
*group
;
4024 unsigned long imbalance
;
4026 unsigned long flags
;
4027 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4029 cpumask_setall(cpus
);
4032 * When power savings policy is enabled for the parent domain, idle
4033 * sibling can pick up load irrespective of busy siblings. In this case,
4034 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4035 * portraying it as CPU_NOT_IDLE.
4037 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4038 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4041 schedstat_inc(sd
, lb_count
[idle
]);
4045 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
4052 schedstat_inc(sd
, lb_nobusyg
[idle
]);
4056 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
4058 schedstat_inc(sd
, lb_nobusyq
[idle
]);
4062 BUG_ON(busiest
== this_rq
);
4064 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
4067 if (busiest
->nr_running
> 1) {
4069 * Attempt to move tasks. If find_busiest_group has found
4070 * an imbalance but busiest->nr_running <= 1, the group is
4071 * still unbalanced. ld_moved simply stays zero, so it is
4072 * correctly treated as an imbalance.
4074 local_irq_save(flags
);
4075 double_rq_lock(this_rq
, busiest
);
4076 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4077 imbalance
, sd
, idle
, &all_pinned
);
4078 double_rq_unlock(this_rq
, busiest
);
4079 local_irq_restore(flags
);
4082 * some other cpu did the load balance for us.
4084 if (ld_moved
&& this_cpu
!= smp_processor_id())
4085 resched_cpu(this_cpu
);
4087 /* All tasks on this runqueue were pinned by CPU affinity */
4088 if (unlikely(all_pinned
)) {
4089 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4090 if (!cpumask_empty(cpus
))
4097 schedstat_inc(sd
, lb_failed
[idle
]);
4098 sd
->nr_balance_failed
++;
4100 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
4102 spin_lock_irqsave(&busiest
->lock
, flags
);
4104 /* don't kick the migration_thread, if the curr
4105 * task on busiest cpu can't be moved to this_cpu
4107 if (!cpumask_test_cpu(this_cpu
,
4108 &busiest
->curr
->cpus_allowed
)) {
4109 spin_unlock_irqrestore(&busiest
->lock
, flags
);
4111 goto out_one_pinned
;
4114 if (!busiest
->active_balance
) {
4115 busiest
->active_balance
= 1;
4116 busiest
->push_cpu
= this_cpu
;
4119 spin_unlock_irqrestore(&busiest
->lock
, flags
);
4121 wake_up_process(busiest
->migration_thread
);
4124 * We've kicked active balancing, reset the failure
4127 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
4130 sd
->nr_balance_failed
= 0;
4132 if (likely(!active_balance
)) {
4133 /* We were unbalanced, so reset the balancing interval */
4134 sd
->balance_interval
= sd
->min_interval
;
4137 * If we've begun active balancing, start to back off. This
4138 * case may not be covered by the all_pinned logic if there
4139 * is only 1 task on the busy runqueue (because we don't call
4142 if (sd
->balance_interval
< sd
->max_interval
)
4143 sd
->balance_interval
*= 2;
4146 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4147 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4153 schedstat_inc(sd
, lb_balanced
[idle
]);
4155 sd
->nr_balance_failed
= 0;
4158 /* tune up the balancing interval */
4159 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
4160 (sd
->balance_interval
< sd
->max_interval
))
4161 sd
->balance_interval
*= 2;
4163 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4164 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4175 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4176 * tasks if there is an imbalance.
4178 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4179 * this_rq is locked.
4182 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
4184 struct sched_group
*group
;
4185 struct rq
*busiest
= NULL
;
4186 unsigned long imbalance
;
4190 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4192 cpumask_setall(cpus
);
4195 * When power savings policy is enabled for the parent domain, idle
4196 * sibling can pick up load irrespective of busy siblings. In this case,
4197 * let the state of idle sibling percolate up as IDLE, instead of
4198 * portraying it as CPU_NOT_IDLE.
4200 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
4201 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4204 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
4206 update_shares_locked(this_rq
, sd
);
4207 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
4208 &sd_idle
, cpus
, NULL
);
4210 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
4214 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
4216 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
4220 BUG_ON(busiest
== this_rq
);
4222 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
4225 if (busiest
->nr_running
> 1) {
4226 /* Attempt to move tasks */
4227 double_lock_balance(this_rq
, busiest
);
4228 /* this_rq->clock is already updated */
4229 update_rq_clock(busiest
);
4230 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4231 imbalance
, sd
, CPU_NEWLY_IDLE
,
4233 double_unlock_balance(this_rq
, busiest
);
4235 if (unlikely(all_pinned
)) {
4236 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4237 if (!cpumask_empty(cpus
))
4243 int active_balance
= 0;
4245 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
4246 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4247 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4250 if (sched_mc_power_savings
< POWERSAVINGS_BALANCE_WAKEUP
)
4253 if (sd
->nr_balance_failed
++ < 2)
4257 * The only task running in a non-idle cpu can be moved to this
4258 * cpu in an attempt to completely freeup the other CPU
4259 * package. The same method used to move task in load_balance()
4260 * have been extended for load_balance_newidle() to speedup
4261 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4263 * The package power saving logic comes from
4264 * find_busiest_group(). If there are no imbalance, then
4265 * f_b_g() will return NULL. However when sched_mc={1,2} then
4266 * f_b_g() will select a group from which a running task may be
4267 * pulled to this cpu in order to make the other package idle.
4268 * If there is no opportunity to make a package idle and if
4269 * there are no imbalance, then f_b_g() will return NULL and no
4270 * action will be taken in load_balance_newidle().
4272 * Under normal task pull operation due to imbalance, there
4273 * will be more than one task in the source run queue and
4274 * move_tasks() will succeed. ld_moved will be true and this
4275 * active balance code will not be triggered.
4278 /* Lock busiest in correct order while this_rq is held */
4279 double_lock_balance(this_rq
, busiest
);
4282 * don't kick the migration_thread, if the curr
4283 * task on busiest cpu can't be moved to this_cpu
4285 if (!cpumask_test_cpu(this_cpu
, &busiest
->curr
->cpus_allowed
)) {
4286 double_unlock_balance(this_rq
, busiest
);
4291 if (!busiest
->active_balance
) {
4292 busiest
->active_balance
= 1;
4293 busiest
->push_cpu
= this_cpu
;
4297 double_unlock_balance(this_rq
, busiest
);
4299 * Should not call ttwu while holding a rq->lock
4301 spin_unlock(&this_rq
->lock
);
4303 wake_up_process(busiest
->migration_thread
);
4304 spin_lock(&this_rq
->lock
);
4307 sd
->nr_balance_failed
= 0;
4309 update_shares_locked(this_rq
, sd
);
4313 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
4314 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4315 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4317 sd
->nr_balance_failed
= 0;
4323 * idle_balance is called by schedule() if this_cpu is about to become
4324 * idle. Attempts to pull tasks from other CPUs.
4326 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
4328 struct sched_domain
*sd
;
4329 int pulled_task
= 0;
4330 unsigned long next_balance
= jiffies
+ HZ
;
4332 for_each_domain(this_cpu
, sd
) {
4333 unsigned long interval
;
4335 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4338 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
4339 /* If we've pulled tasks over stop searching: */
4340 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
4343 interval
= msecs_to_jiffies(sd
->balance_interval
);
4344 if (time_after(next_balance
, sd
->last_balance
+ interval
))
4345 next_balance
= sd
->last_balance
+ interval
;
4349 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
4351 * We are going idle. next_balance may be set based on
4352 * a busy processor. So reset next_balance.
4354 this_rq
->next_balance
= next_balance
;
4359 * active_load_balance is run by migration threads. It pushes running tasks
4360 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4361 * running on each physical CPU where possible, and avoids physical /
4362 * logical imbalances.
4364 * Called with busiest_rq locked.
4366 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
4368 int target_cpu
= busiest_rq
->push_cpu
;
4369 struct sched_domain
*sd
;
4370 struct rq
*target_rq
;
4372 /* Is there any task to move? */
4373 if (busiest_rq
->nr_running
<= 1)
4376 target_rq
= cpu_rq(target_cpu
);
4379 * This condition is "impossible", if it occurs
4380 * we need to fix it. Originally reported by
4381 * Bjorn Helgaas on a 128-cpu setup.
4383 BUG_ON(busiest_rq
== target_rq
);
4385 /* move a task from busiest_rq to target_rq */
4386 double_lock_balance(busiest_rq
, target_rq
);
4387 update_rq_clock(busiest_rq
);
4388 update_rq_clock(target_rq
);
4390 /* Search for an sd spanning us and the target CPU. */
4391 for_each_domain(target_cpu
, sd
) {
4392 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
4393 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
4398 schedstat_inc(sd
, alb_count
);
4400 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
4402 schedstat_inc(sd
, alb_pushed
);
4404 schedstat_inc(sd
, alb_failed
);
4406 double_unlock_balance(busiest_rq
, target_rq
);
4411 atomic_t load_balancer
;
4412 cpumask_var_t cpu_mask
;
4413 cpumask_var_t ilb_grp_nohz_mask
;
4414 } nohz ____cacheline_aligned
= {
4415 .load_balancer
= ATOMIC_INIT(-1),
4418 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4420 * lowest_flag_domain - Return lowest sched_domain containing flag.
4421 * @cpu: The cpu whose lowest level of sched domain is to
4423 * @flag: The flag to check for the lowest sched_domain
4424 * for the given cpu.
4426 * Returns the lowest sched_domain of a cpu which contains the given flag.
4428 static inline struct sched_domain
*lowest_flag_domain(int cpu
, int flag
)
4430 struct sched_domain
*sd
;
4432 for_each_domain(cpu
, sd
)
4433 if (sd
&& (sd
->flags
& flag
))
4440 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4441 * @cpu: The cpu whose domains we're iterating over.
4442 * @sd: variable holding the value of the power_savings_sd
4444 * @flag: The flag to filter the sched_domains to be iterated.
4446 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4447 * set, starting from the lowest sched_domain to the highest.
4449 #define for_each_flag_domain(cpu, sd, flag) \
4450 for (sd = lowest_flag_domain(cpu, flag); \
4451 (sd && (sd->flags & flag)); sd = sd->parent)
4454 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4455 * @ilb_group: group to be checked for semi-idleness
4457 * Returns: 1 if the group is semi-idle. 0 otherwise.
4459 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4460 * and atleast one non-idle CPU. This helper function checks if the given
4461 * sched_group is semi-idle or not.
4463 static inline int is_semi_idle_group(struct sched_group
*ilb_group
)
4465 cpumask_and(nohz
.ilb_grp_nohz_mask
, nohz
.cpu_mask
,
4466 sched_group_cpus(ilb_group
));
4469 * A sched_group is semi-idle when it has atleast one busy cpu
4470 * and atleast one idle cpu.
4472 if (cpumask_empty(nohz
.ilb_grp_nohz_mask
))
4475 if (cpumask_equal(nohz
.ilb_grp_nohz_mask
, sched_group_cpus(ilb_group
)))
4481 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4482 * @cpu: The cpu which is nominating a new idle_load_balancer.
4484 * Returns: Returns the id of the idle load balancer if it exists,
4485 * Else, returns >= nr_cpu_ids.
4487 * This algorithm picks the idle load balancer such that it belongs to a
4488 * semi-idle powersavings sched_domain. The idea is to try and avoid
4489 * completely idle packages/cores just for the purpose of idle load balancing
4490 * when there are other idle cpu's which are better suited for that job.
4492 static int find_new_ilb(int cpu
)
4494 struct sched_domain
*sd
;
4495 struct sched_group
*ilb_group
;
4498 * Have idle load balancer selection from semi-idle packages only
4499 * when power-aware load balancing is enabled
4501 if (!(sched_smt_power_savings
|| sched_mc_power_savings
))
4505 * Optimize for the case when we have no idle CPUs or only one
4506 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4508 if (cpumask_weight(nohz
.cpu_mask
) < 2)
4511 for_each_flag_domain(cpu
, sd
, SD_POWERSAVINGS_BALANCE
) {
4512 ilb_group
= sd
->groups
;
4515 if (is_semi_idle_group(ilb_group
))
4516 return cpumask_first(nohz
.ilb_grp_nohz_mask
);
4518 ilb_group
= ilb_group
->next
;
4520 } while (ilb_group
!= sd
->groups
);
4524 return cpumask_first(nohz
.cpu_mask
);
4526 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4527 static inline int find_new_ilb(int call_cpu
)
4529 return cpumask_first(nohz
.cpu_mask
);
4534 * This routine will try to nominate the ilb (idle load balancing)
4535 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4536 * load balancing on behalf of all those cpus. If all the cpus in the system
4537 * go into this tickless mode, then there will be no ilb owner (as there is
4538 * no need for one) and all the cpus will sleep till the next wakeup event
4541 * For the ilb owner, tick is not stopped. And this tick will be used
4542 * for idle load balancing. ilb owner will still be part of
4545 * While stopping the tick, this cpu will become the ilb owner if there
4546 * is no other owner. And will be the owner till that cpu becomes busy
4547 * or if all cpus in the system stop their ticks at which point
4548 * there is no need for ilb owner.
4550 * When the ilb owner becomes busy, it nominates another owner, during the
4551 * next busy scheduler_tick()
4553 int select_nohz_load_balancer(int stop_tick
)
4555 int cpu
= smp_processor_id();
4558 cpu_rq(cpu
)->in_nohz_recently
= 1;
4560 if (!cpu_active(cpu
)) {
4561 if (atomic_read(&nohz
.load_balancer
) != cpu
)
4565 * If we are going offline and still the leader,
4568 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4574 cpumask_set_cpu(cpu
, nohz
.cpu_mask
);
4576 /* time for ilb owner also to sleep */
4577 if (cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4578 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4579 atomic_set(&nohz
.load_balancer
, -1);
4583 if (atomic_read(&nohz
.load_balancer
) == -1) {
4584 /* make me the ilb owner */
4585 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
4587 } else if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4590 if (!(sched_smt_power_savings
||
4591 sched_mc_power_savings
))
4594 * Check to see if there is a more power-efficient
4597 new_ilb
= find_new_ilb(cpu
);
4598 if (new_ilb
< nr_cpu_ids
&& new_ilb
!= cpu
) {
4599 atomic_set(&nohz
.load_balancer
, -1);
4600 resched_cpu(new_ilb
);
4606 if (!cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4609 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4611 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4612 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4619 static DEFINE_SPINLOCK(balancing
);
4622 * It checks each scheduling domain to see if it is due to be balanced,
4623 * and initiates a balancing operation if so.
4625 * Balancing parameters are set up in arch_init_sched_domains.
4627 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
4630 struct rq
*rq
= cpu_rq(cpu
);
4631 unsigned long interval
;
4632 struct sched_domain
*sd
;
4633 /* Earliest time when we have to do rebalance again */
4634 unsigned long next_balance
= jiffies
+ 60*HZ
;
4635 int update_next_balance
= 0;
4638 for_each_domain(cpu
, sd
) {
4639 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4642 interval
= sd
->balance_interval
;
4643 if (idle
!= CPU_IDLE
)
4644 interval
*= sd
->busy_factor
;
4646 /* scale ms to jiffies */
4647 interval
= msecs_to_jiffies(interval
);
4648 if (unlikely(!interval
))
4650 if (interval
> HZ
*NR_CPUS
/10)
4651 interval
= HZ
*NR_CPUS
/10;
4653 need_serialize
= sd
->flags
& SD_SERIALIZE
;
4655 if (need_serialize
) {
4656 if (!spin_trylock(&balancing
))
4660 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
4661 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
4663 * We've pulled tasks over so either we're no
4664 * longer idle, or one of our SMT siblings is
4667 idle
= CPU_NOT_IDLE
;
4669 sd
->last_balance
= jiffies
;
4672 spin_unlock(&balancing
);
4674 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
4675 next_balance
= sd
->last_balance
+ interval
;
4676 update_next_balance
= 1;
4680 * Stop the load balance at this level. There is another
4681 * CPU in our sched group which is doing load balancing more
4689 * next_balance will be updated only when there is a need.
4690 * When the cpu is attached to null domain for ex, it will not be
4693 if (likely(update_next_balance
))
4694 rq
->next_balance
= next_balance
;
4698 * run_rebalance_domains is triggered when needed from the scheduler tick.
4699 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4700 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4702 static void run_rebalance_domains(struct softirq_action
*h
)
4704 int this_cpu
= smp_processor_id();
4705 struct rq
*this_rq
= cpu_rq(this_cpu
);
4706 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
4707 CPU_IDLE
: CPU_NOT_IDLE
;
4709 rebalance_domains(this_cpu
, idle
);
4713 * If this cpu is the owner for idle load balancing, then do the
4714 * balancing on behalf of the other idle cpus whose ticks are
4717 if (this_rq
->idle_at_tick
&&
4718 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
4722 for_each_cpu(balance_cpu
, nohz
.cpu_mask
) {
4723 if (balance_cpu
== this_cpu
)
4727 * If this cpu gets work to do, stop the load balancing
4728 * work being done for other cpus. Next load
4729 * balancing owner will pick it up.
4734 rebalance_domains(balance_cpu
, CPU_IDLE
);
4736 rq
= cpu_rq(balance_cpu
);
4737 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
4738 this_rq
->next_balance
= rq
->next_balance
;
4744 static inline int on_null_domain(int cpu
)
4746 return !rcu_dereference(cpu_rq(cpu
)->sd
);
4750 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4752 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4753 * idle load balancing owner or decide to stop the periodic load balancing,
4754 * if the whole system is idle.
4756 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
4760 * If we were in the nohz mode recently and busy at the current
4761 * scheduler tick, then check if we need to nominate new idle
4764 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
4765 rq
->in_nohz_recently
= 0;
4767 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4768 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4769 atomic_set(&nohz
.load_balancer
, -1);
4772 if (atomic_read(&nohz
.load_balancer
) == -1) {
4773 int ilb
= find_new_ilb(cpu
);
4775 if (ilb
< nr_cpu_ids
)
4781 * If this cpu is idle and doing idle load balancing for all the
4782 * cpus with ticks stopped, is it time for that to stop?
4784 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4785 cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4791 * If this cpu is idle and the idle load balancing is done by
4792 * someone else, then no need raise the SCHED_SOFTIRQ
4794 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4795 cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4798 /* Don't need to rebalance while attached to NULL domain */
4799 if (time_after_eq(jiffies
, rq
->next_balance
) &&
4800 likely(!on_null_domain(cpu
)))
4801 raise_softirq(SCHED_SOFTIRQ
);
4804 #else /* CONFIG_SMP */
4807 * on UP we do not need to balance between CPUs:
4809 static inline void idle_balance(int cpu
, struct rq
*rq
)
4815 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4817 EXPORT_PER_CPU_SYMBOL(kstat
);
4820 * Return any ns on the sched_clock that have not yet been accounted in
4821 * @p in case that task is currently running.
4823 * Called with task_rq_lock() held on @rq.
4825 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
4829 if (task_current(rq
, p
)) {
4830 update_rq_clock(rq
);
4831 ns
= rq
->clock
- p
->se
.exec_start
;
4839 unsigned long long task_delta_exec(struct task_struct
*p
)
4841 unsigned long flags
;
4845 rq
= task_rq_lock(p
, &flags
);
4846 ns
= do_task_delta_exec(p
, rq
);
4847 task_rq_unlock(rq
, &flags
);
4853 * Return accounted runtime for the task.
4854 * In case the task is currently running, return the runtime plus current's
4855 * pending runtime that have not been accounted yet.
4857 unsigned long long task_sched_runtime(struct task_struct
*p
)
4859 unsigned long flags
;
4863 rq
= task_rq_lock(p
, &flags
);
4864 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
4865 task_rq_unlock(rq
, &flags
);
4871 * Return sum_exec_runtime for the thread group.
4872 * In case the task is currently running, return the sum plus current's
4873 * pending runtime that have not been accounted yet.
4875 * Note that the thread group might have other running tasks as well,
4876 * so the return value not includes other pending runtime that other
4877 * running tasks might have.
4879 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
4881 struct task_cputime totals
;
4882 unsigned long flags
;
4886 rq
= task_rq_lock(p
, &flags
);
4887 thread_group_cputime(p
, &totals
);
4888 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
4889 task_rq_unlock(rq
, &flags
);
4895 * Account user cpu time to a process.
4896 * @p: the process that the cpu time gets accounted to
4897 * @cputime: the cpu time spent in user space since the last update
4898 * @cputime_scaled: cputime scaled by cpu frequency
4900 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
4901 cputime_t cputime_scaled
)
4903 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4906 /* Add user time to process. */
4907 p
->utime
= cputime_add(p
->utime
, cputime
);
4908 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
4909 account_group_user_time(p
, cputime
);
4911 /* Add user time to cpustat. */
4912 tmp
= cputime_to_cputime64(cputime
);
4913 if (TASK_NICE(p
) > 0)
4914 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
4916 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4918 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
4919 /* Account for user time used */
4920 acct_update_integrals(p
);
4924 * Account guest cpu time to a process.
4925 * @p: the process that the cpu time gets accounted to
4926 * @cputime: the cpu time spent in virtual machine since the last update
4927 * @cputime_scaled: cputime scaled by cpu frequency
4929 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
4930 cputime_t cputime_scaled
)
4933 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4935 tmp
= cputime_to_cputime64(cputime
);
4937 /* Add guest time to process. */
4938 p
->utime
= cputime_add(p
->utime
, cputime
);
4939 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
4940 account_group_user_time(p
, cputime
);
4941 p
->gtime
= cputime_add(p
->gtime
, cputime
);
4943 /* Add guest time to cpustat. */
4944 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4945 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
4949 * Account system cpu time to a process.
4950 * @p: the process that the cpu time gets accounted to
4951 * @hardirq_offset: the offset to subtract from hardirq_count()
4952 * @cputime: the cpu time spent in kernel space since the last update
4953 * @cputime_scaled: cputime scaled by cpu frequency
4955 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
4956 cputime_t cputime
, cputime_t cputime_scaled
)
4958 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4961 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
4962 account_guest_time(p
, cputime
, cputime_scaled
);
4966 /* Add system time to process. */
4967 p
->stime
= cputime_add(p
->stime
, cputime
);
4968 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
4969 account_group_system_time(p
, cputime
);
4971 /* Add system time to cpustat. */
4972 tmp
= cputime_to_cputime64(cputime
);
4973 if (hardirq_count() - hardirq_offset
)
4974 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
4975 else if (softirq_count())
4976 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
4978 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
4980 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
4982 /* Account for system time used */
4983 acct_update_integrals(p
);
4987 * Account for involuntary wait time.
4988 * @steal: the cpu time spent in involuntary wait
4990 void account_steal_time(cputime_t cputime
)
4992 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4993 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
4995 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
4999 * Account for idle time.
5000 * @cputime: the cpu time spent in idle wait
5002 void account_idle_time(cputime_t cputime
)
5004 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5005 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
5006 struct rq
*rq
= this_rq();
5008 if (atomic_read(&rq
->nr_iowait
) > 0)
5009 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
5011 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
5014 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5017 * Account a single tick of cpu time.
5018 * @p: the process that the cpu time gets accounted to
5019 * @user_tick: indicates if the tick is a user or a system tick
5021 void account_process_tick(struct task_struct
*p
, int user_tick
)
5023 cputime_t one_jiffy
= jiffies_to_cputime(1);
5024 cputime_t one_jiffy_scaled
= cputime_to_scaled(one_jiffy
);
5025 struct rq
*rq
= this_rq();
5028 account_user_time(p
, one_jiffy
, one_jiffy_scaled
);
5029 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
5030 account_system_time(p
, HARDIRQ_OFFSET
, one_jiffy
,
5033 account_idle_time(one_jiffy
);
5037 * Account multiple ticks of steal time.
5038 * @p: the process from which the cpu time has been stolen
5039 * @ticks: number of stolen ticks
5041 void account_steal_ticks(unsigned long ticks
)
5043 account_steal_time(jiffies_to_cputime(ticks
));
5047 * Account multiple ticks of idle time.
5048 * @ticks: number of stolen ticks
5050 void account_idle_ticks(unsigned long ticks
)
5052 account_idle_time(jiffies_to_cputime(ticks
));
5058 * Use precise platform statistics if available:
5060 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5061 cputime_t
task_utime(struct task_struct
*p
)
5066 cputime_t
task_stime(struct task_struct
*p
)
5071 cputime_t
task_utime(struct task_struct
*p
)
5073 clock_t utime
= cputime_to_clock_t(p
->utime
),
5074 total
= utime
+ cputime_to_clock_t(p
->stime
);
5078 * Use CFS's precise accounting:
5080 temp
= (u64
)nsec_to_clock_t(p
->se
.sum_exec_runtime
);
5084 do_div(temp
, total
);
5086 utime
= (clock_t)temp
;
5088 p
->prev_utime
= max(p
->prev_utime
, clock_t_to_cputime(utime
));
5089 return p
->prev_utime
;
5092 cputime_t
task_stime(struct task_struct
*p
)
5097 * Use CFS's precise accounting. (we subtract utime from
5098 * the total, to make sure the total observed by userspace
5099 * grows monotonically - apps rely on that):
5101 stime
= nsec_to_clock_t(p
->se
.sum_exec_runtime
) -
5102 cputime_to_clock_t(task_utime(p
));
5105 p
->prev_stime
= max(p
->prev_stime
, clock_t_to_cputime(stime
));
5107 return p
->prev_stime
;
5111 inline cputime_t
task_gtime(struct task_struct
*p
)
5117 * This function gets called by the timer code, with HZ frequency.
5118 * We call it with interrupts disabled.
5120 * It also gets called by the fork code, when changing the parent's
5123 void scheduler_tick(void)
5125 int cpu
= smp_processor_id();
5126 struct rq
*rq
= cpu_rq(cpu
);
5127 struct task_struct
*curr
= rq
->curr
;
5131 spin_lock(&rq
->lock
);
5132 update_rq_clock(rq
);
5133 update_cpu_load(rq
);
5134 curr
->sched_class
->task_tick(rq
, curr
, 0);
5135 spin_unlock(&rq
->lock
);
5137 perf_counter_task_tick(curr
, cpu
);
5140 rq
->idle_at_tick
= idle_cpu(cpu
);
5141 trigger_load_balance(rq
, cpu
);
5145 notrace
unsigned long get_parent_ip(unsigned long addr
)
5147 if (in_lock_functions(addr
)) {
5148 addr
= CALLER_ADDR2
;
5149 if (in_lock_functions(addr
))
5150 addr
= CALLER_ADDR3
;
5155 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5156 defined(CONFIG_PREEMPT_TRACER))
5158 void __kprobes
add_preempt_count(int val
)
5160 #ifdef CONFIG_DEBUG_PREEMPT
5164 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5167 preempt_count() += val
;
5168 #ifdef CONFIG_DEBUG_PREEMPT
5170 * Spinlock count overflowing soon?
5172 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
5175 if (preempt_count() == val
)
5176 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
5178 EXPORT_SYMBOL(add_preempt_count
);
5180 void __kprobes
sub_preempt_count(int val
)
5182 #ifdef CONFIG_DEBUG_PREEMPT
5186 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
5189 * Is the spinlock portion underflowing?
5191 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
5192 !(preempt_count() & PREEMPT_MASK
)))
5196 if (preempt_count() == val
)
5197 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
5198 preempt_count() -= val
;
5200 EXPORT_SYMBOL(sub_preempt_count
);
5205 * Print scheduling while atomic bug:
5207 static noinline
void __schedule_bug(struct task_struct
*prev
)
5209 struct pt_regs
*regs
= get_irq_regs();
5211 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
5212 prev
->comm
, prev
->pid
, preempt_count());
5214 debug_show_held_locks(prev
);
5216 if (irqs_disabled())
5217 print_irqtrace_events(prev
);
5226 * Various schedule()-time debugging checks and statistics:
5228 static inline void schedule_debug(struct task_struct
*prev
)
5231 * Test if we are atomic. Since do_exit() needs to call into
5232 * schedule() atomically, we ignore that path for now.
5233 * Otherwise, whine if we are scheduling when we should not be.
5235 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
5236 __schedule_bug(prev
);
5238 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
5240 schedstat_inc(this_rq(), sched_count
);
5241 #ifdef CONFIG_SCHEDSTATS
5242 if (unlikely(prev
->lock_depth
>= 0)) {
5243 schedstat_inc(this_rq(), bkl_count
);
5244 schedstat_inc(prev
, sched_info
.bkl_count
);
5249 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
5251 if (prev
->state
== TASK_RUNNING
) {
5252 u64 runtime
= prev
->se
.sum_exec_runtime
;
5254 runtime
-= prev
->se
.prev_sum_exec_runtime
;
5255 runtime
= min_t(u64
, runtime
, 2*sysctl_sched_migration_cost
);
5258 * In order to avoid avg_overlap growing stale when we are
5259 * indeed overlapping and hence not getting put to sleep, grow
5260 * the avg_overlap on preemption.
5262 * We use the average preemption runtime because that
5263 * correlates to the amount of cache footprint a task can
5266 update_avg(&prev
->se
.avg_overlap
, runtime
);
5268 prev
->sched_class
->put_prev_task(rq
, prev
);
5272 * Pick up the highest-prio task:
5274 static inline struct task_struct
*
5275 pick_next_task(struct rq
*rq
)
5277 const struct sched_class
*class;
5278 struct task_struct
*p
;
5281 * Optimization: we know that if all tasks are in
5282 * the fair class we can call that function directly:
5284 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
5285 p
= fair_sched_class
.pick_next_task(rq
);
5290 class = sched_class_highest
;
5292 p
= class->pick_next_task(rq
);
5296 * Will never be NULL as the idle class always
5297 * returns a non-NULL p:
5299 class = class->next
;
5304 * schedule() is the main scheduler function.
5306 asmlinkage
void __sched
schedule(void)
5308 struct task_struct
*prev
, *next
;
5309 unsigned long *switch_count
;
5315 cpu
= smp_processor_id();
5319 switch_count
= &prev
->nivcsw
;
5321 release_kernel_lock(prev
);
5322 need_resched_nonpreemptible
:
5324 schedule_debug(prev
);
5326 if (sched_feat(HRTICK
))
5329 spin_lock_irq(&rq
->lock
);
5330 update_rq_clock(rq
);
5331 clear_tsk_need_resched(prev
);
5333 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
5334 if (unlikely(signal_pending_state(prev
->state
, prev
)))
5335 prev
->state
= TASK_RUNNING
;
5337 deactivate_task(rq
, prev
, 1);
5338 switch_count
= &prev
->nvcsw
;
5342 if (prev
->sched_class
->pre_schedule
)
5343 prev
->sched_class
->pre_schedule(rq
, prev
);
5346 if (unlikely(!rq
->nr_running
))
5347 idle_balance(cpu
, rq
);
5349 put_prev_task(rq
, prev
);
5350 next
= pick_next_task(rq
);
5352 if (likely(prev
!= next
)) {
5353 sched_info_switch(prev
, next
);
5354 perf_counter_task_sched_out(prev
, next
, cpu
);
5360 context_switch(rq
, prev
, next
); /* unlocks the rq */
5362 * the context switch might have flipped the stack from under
5363 * us, hence refresh the local variables.
5365 cpu
= smp_processor_id();
5368 spin_unlock_irq(&rq
->lock
);
5370 if (unlikely(reacquire_kernel_lock(current
) < 0))
5371 goto need_resched_nonpreemptible
;
5373 preempt_enable_no_resched();
5377 EXPORT_SYMBOL(schedule
);
5381 * Look out! "owner" is an entirely speculative pointer
5382 * access and not reliable.
5384 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
5389 if (!sched_feat(OWNER_SPIN
))
5392 #ifdef CONFIG_DEBUG_PAGEALLOC
5394 * Need to access the cpu field knowing that
5395 * DEBUG_PAGEALLOC could have unmapped it if
5396 * the mutex owner just released it and exited.
5398 if (probe_kernel_address(&owner
->cpu
, cpu
))
5405 * Even if the access succeeded (likely case),
5406 * the cpu field may no longer be valid.
5408 if (cpu
>= nr_cpumask_bits
)
5412 * We need to validate that we can do a
5413 * get_cpu() and that we have the percpu area.
5415 if (!cpu_online(cpu
))
5422 * Owner changed, break to re-assess state.
5424 if (lock
->owner
!= owner
)
5428 * Is that owner really running on that cpu?
5430 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
5440 #ifdef CONFIG_PREEMPT
5442 * this is the entry point to schedule() from in-kernel preemption
5443 * off of preempt_enable. Kernel preemptions off return from interrupt
5444 * occur there and call schedule directly.
5446 asmlinkage
void __sched
preempt_schedule(void)
5448 struct thread_info
*ti
= current_thread_info();
5451 * If there is a non-zero preempt_count or interrupts are disabled,
5452 * we do not want to preempt the current task. Just return..
5454 if (likely(ti
->preempt_count
|| irqs_disabled()))
5458 add_preempt_count(PREEMPT_ACTIVE
);
5460 sub_preempt_count(PREEMPT_ACTIVE
);
5463 * Check again in case we missed a preemption opportunity
5464 * between schedule and now.
5467 } while (need_resched());
5469 EXPORT_SYMBOL(preempt_schedule
);
5472 * this is the entry point to schedule() from kernel preemption
5473 * off of irq context.
5474 * Note, that this is called and return with irqs disabled. This will
5475 * protect us against recursive calling from irq.
5477 asmlinkage
void __sched
preempt_schedule_irq(void)
5479 struct thread_info
*ti
= current_thread_info();
5481 /* Catch callers which need to be fixed */
5482 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
5485 add_preempt_count(PREEMPT_ACTIVE
);
5488 local_irq_disable();
5489 sub_preempt_count(PREEMPT_ACTIVE
);
5492 * Check again in case we missed a preemption opportunity
5493 * between schedule and now.
5496 } while (need_resched());
5499 #endif /* CONFIG_PREEMPT */
5501 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
5504 return try_to_wake_up(curr
->private, mode
, sync
);
5506 EXPORT_SYMBOL(default_wake_function
);
5509 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5510 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5511 * number) then we wake all the non-exclusive tasks and one exclusive task.
5513 * There are circumstances in which we can try to wake a task which has already
5514 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5515 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5517 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
5518 int nr_exclusive
, int sync
, void *key
)
5520 wait_queue_t
*curr
, *next
;
5522 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
5523 unsigned flags
= curr
->flags
;
5525 if (curr
->func(curr
, mode
, sync
, key
) &&
5526 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
5532 * __wake_up - wake up threads blocked on a waitqueue.
5534 * @mode: which threads
5535 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5536 * @key: is directly passed to the wakeup function
5538 * It may be assumed that this function implies a write memory barrier before
5539 * changing the task state if and only if any tasks are woken up.
5541 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
5542 int nr_exclusive
, void *key
)
5544 unsigned long flags
;
5546 spin_lock_irqsave(&q
->lock
, flags
);
5547 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
5548 spin_unlock_irqrestore(&q
->lock
, flags
);
5550 EXPORT_SYMBOL(__wake_up
);
5553 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5555 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
5557 __wake_up_common(q
, mode
, 1, 0, NULL
);
5560 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
5562 __wake_up_common(q
, mode
, 1, 0, key
);
5566 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5568 * @mode: which threads
5569 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5570 * @key: opaque value to be passed to wakeup targets
5572 * The sync wakeup differs that the waker knows that it will schedule
5573 * away soon, so while the target thread will be woken up, it will not
5574 * be migrated to another CPU - ie. the two threads are 'synchronized'
5575 * with each other. This can prevent needless bouncing between CPUs.
5577 * On UP it can prevent extra preemption.
5579 * It may be assumed that this function implies a write memory barrier before
5580 * changing the task state if and only if any tasks are woken up.
5582 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
5583 int nr_exclusive
, void *key
)
5585 unsigned long flags
;
5591 if (unlikely(!nr_exclusive
))
5594 spin_lock_irqsave(&q
->lock
, flags
);
5595 __wake_up_common(q
, mode
, nr_exclusive
, sync
, key
);
5596 spin_unlock_irqrestore(&q
->lock
, flags
);
5598 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
5601 * __wake_up_sync - see __wake_up_sync_key()
5603 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
5605 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
5607 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
5610 * complete: - signals a single thread waiting on this completion
5611 * @x: holds the state of this particular completion
5613 * This will wake up a single thread waiting on this completion. Threads will be
5614 * awakened in the same order in which they were queued.
5616 * See also complete_all(), wait_for_completion() and related routines.
5618 * It may be assumed that this function implies a write memory barrier before
5619 * changing the task state if and only if any tasks are woken up.
5621 void complete(struct completion
*x
)
5623 unsigned long flags
;
5625 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5627 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
5628 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5630 EXPORT_SYMBOL(complete
);
5633 * complete_all: - signals all threads waiting on this completion
5634 * @x: holds the state of this particular completion
5636 * This will wake up all threads waiting on this particular completion event.
5638 * It may be assumed that this function implies a write memory barrier before
5639 * changing the task state if and only if any tasks are woken up.
5641 void complete_all(struct completion
*x
)
5643 unsigned long flags
;
5645 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5646 x
->done
+= UINT_MAX
/2;
5647 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
5648 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5650 EXPORT_SYMBOL(complete_all
);
5652 static inline long __sched
5653 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
5656 DECLARE_WAITQUEUE(wait
, current
);
5658 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
5659 __add_wait_queue_tail(&x
->wait
, &wait
);
5661 if (signal_pending_state(state
, current
)) {
5662 timeout
= -ERESTARTSYS
;
5665 __set_current_state(state
);
5666 spin_unlock_irq(&x
->wait
.lock
);
5667 timeout
= schedule_timeout(timeout
);
5668 spin_lock_irq(&x
->wait
.lock
);
5669 } while (!x
->done
&& timeout
);
5670 __remove_wait_queue(&x
->wait
, &wait
);
5675 return timeout
?: 1;
5679 wait_for_common(struct completion
*x
, long timeout
, int state
)
5683 spin_lock_irq(&x
->wait
.lock
);
5684 timeout
= do_wait_for_common(x
, timeout
, state
);
5685 spin_unlock_irq(&x
->wait
.lock
);
5690 * wait_for_completion: - waits for completion of a task
5691 * @x: holds the state of this particular completion
5693 * This waits to be signaled for completion of a specific task. It is NOT
5694 * interruptible and there is no timeout.
5696 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5697 * and interrupt capability. Also see complete().
5699 void __sched
wait_for_completion(struct completion
*x
)
5701 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
5703 EXPORT_SYMBOL(wait_for_completion
);
5706 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5707 * @x: holds the state of this particular completion
5708 * @timeout: timeout value in jiffies
5710 * This waits for either a completion of a specific task to be signaled or for a
5711 * specified timeout to expire. The timeout is in jiffies. It is not
5714 unsigned long __sched
5715 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
5717 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
5719 EXPORT_SYMBOL(wait_for_completion_timeout
);
5722 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5723 * @x: holds the state of this particular completion
5725 * This waits for completion of a specific task to be signaled. It is
5728 int __sched
wait_for_completion_interruptible(struct completion
*x
)
5730 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
5731 if (t
== -ERESTARTSYS
)
5735 EXPORT_SYMBOL(wait_for_completion_interruptible
);
5738 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5739 * @x: holds the state of this particular completion
5740 * @timeout: timeout value in jiffies
5742 * This waits for either a completion of a specific task to be signaled or for a
5743 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5745 unsigned long __sched
5746 wait_for_completion_interruptible_timeout(struct completion
*x
,
5747 unsigned long timeout
)
5749 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
5751 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
5754 * wait_for_completion_killable: - waits for completion of a task (killable)
5755 * @x: holds the state of this particular completion
5757 * This waits to be signaled for completion of a specific task. It can be
5758 * interrupted by a kill signal.
5760 int __sched
wait_for_completion_killable(struct completion
*x
)
5762 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
5763 if (t
== -ERESTARTSYS
)
5767 EXPORT_SYMBOL(wait_for_completion_killable
);
5770 * try_wait_for_completion - try to decrement a completion without blocking
5771 * @x: completion structure
5773 * Returns: 0 if a decrement cannot be done without blocking
5774 * 1 if a decrement succeeded.
5776 * If a completion is being used as a counting completion,
5777 * attempt to decrement the counter without blocking. This
5778 * enables us to avoid waiting if the resource the completion
5779 * is protecting is not available.
5781 bool try_wait_for_completion(struct completion
*x
)
5785 spin_lock_irq(&x
->wait
.lock
);
5790 spin_unlock_irq(&x
->wait
.lock
);
5793 EXPORT_SYMBOL(try_wait_for_completion
);
5796 * completion_done - Test to see if a completion has any waiters
5797 * @x: completion structure
5799 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5800 * 1 if there are no waiters.
5803 bool completion_done(struct completion
*x
)
5807 spin_lock_irq(&x
->wait
.lock
);
5810 spin_unlock_irq(&x
->wait
.lock
);
5813 EXPORT_SYMBOL(completion_done
);
5816 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
5818 unsigned long flags
;
5821 init_waitqueue_entry(&wait
, current
);
5823 __set_current_state(state
);
5825 spin_lock_irqsave(&q
->lock
, flags
);
5826 __add_wait_queue(q
, &wait
);
5827 spin_unlock(&q
->lock
);
5828 timeout
= schedule_timeout(timeout
);
5829 spin_lock_irq(&q
->lock
);
5830 __remove_wait_queue(q
, &wait
);
5831 spin_unlock_irqrestore(&q
->lock
, flags
);
5836 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
5838 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5840 EXPORT_SYMBOL(interruptible_sleep_on
);
5843 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5845 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
5847 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
5849 void __sched
sleep_on(wait_queue_head_t
*q
)
5851 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5853 EXPORT_SYMBOL(sleep_on
);
5855 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5857 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
5859 EXPORT_SYMBOL(sleep_on_timeout
);
5861 #ifdef CONFIG_RT_MUTEXES
5864 * rt_mutex_setprio - set the current priority of a task
5866 * @prio: prio value (kernel-internal form)
5868 * This function changes the 'effective' priority of a task. It does
5869 * not touch ->normal_prio like __setscheduler().
5871 * Used by the rt_mutex code to implement priority inheritance logic.
5873 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
5875 unsigned long flags
;
5876 int oldprio
, on_rq
, running
;
5878 const struct sched_class
*prev_class
= p
->sched_class
;
5880 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
5882 rq
= task_rq_lock(p
, &flags
);
5883 update_rq_clock(rq
);
5886 on_rq
= p
->se
.on_rq
;
5887 running
= task_current(rq
, p
);
5889 dequeue_task(rq
, p
, 0);
5891 p
->sched_class
->put_prev_task(rq
, p
);
5894 p
->sched_class
= &rt_sched_class
;
5896 p
->sched_class
= &fair_sched_class
;
5901 p
->sched_class
->set_curr_task(rq
);
5903 enqueue_task(rq
, p
, 0);
5905 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5907 task_rq_unlock(rq
, &flags
);
5912 void set_user_nice(struct task_struct
*p
, long nice
)
5914 int old_prio
, delta
, on_rq
;
5915 unsigned long flags
;
5918 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
5921 * We have to be careful, if called from sys_setpriority(),
5922 * the task might be in the middle of scheduling on another CPU.
5924 rq
= task_rq_lock(p
, &flags
);
5925 update_rq_clock(rq
);
5927 * The RT priorities are set via sched_setscheduler(), but we still
5928 * allow the 'normal' nice value to be set - but as expected
5929 * it wont have any effect on scheduling until the task is
5930 * SCHED_FIFO/SCHED_RR:
5932 if (task_has_rt_policy(p
)) {
5933 p
->static_prio
= NICE_TO_PRIO(nice
);
5936 on_rq
= p
->se
.on_rq
;
5938 dequeue_task(rq
, p
, 0);
5940 p
->static_prio
= NICE_TO_PRIO(nice
);
5943 p
->prio
= effective_prio(p
);
5944 delta
= p
->prio
- old_prio
;
5947 enqueue_task(rq
, p
, 0);
5949 * If the task increased its priority or is running and
5950 * lowered its priority, then reschedule its CPU:
5952 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
5953 resched_task(rq
->curr
);
5956 task_rq_unlock(rq
, &flags
);
5958 EXPORT_SYMBOL(set_user_nice
);
5961 * can_nice - check if a task can reduce its nice value
5965 int can_nice(const struct task_struct
*p
, const int nice
)
5967 /* convert nice value [19,-20] to rlimit style value [1,40] */
5968 int nice_rlim
= 20 - nice
;
5970 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
5971 capable(CAP_SYS_NICE
));
5974 #ifdef __ARCH_WANT_SYS_NICE
5977 * sys_nice - change the priority of the current process.
5978 * @increment: priority increment
5980 * sys_setpriority is a more generic, but much slower function that
5981 * does similar things.
5983 SYSCALL_DEFINE1(nice
, int, increment
)
5988 * Setpriority might change our priority at the same moment.
5989 * We don't have to worry. Conceptually one call occurs first
5990 * and we have a single winner.
5992 if (increment
< -40)
5997 nice
= TASK_NICE(current
) + increment
;
6003 if (increment
< 0 && !can_nice(current
, nice
))
6006 retval
= security_task_setnice(current
, nice
);
6010 set_user_nice(current
, nice
);
6017 * task_prio - return the priority value of a given task.
6018 * @p: the task in question.
6020 * This is the priority value as seen by users in /proc.
6021 * RT tasks are offset by -200. Normal tasks are centered
6022 * around 0, value goes from -16 to +15.
6024 int task_prio(const struct task_struct
*p
)
6026 return p
->prio
- MAX_RT_PRIO
;
6030 * task_nice - return the nice value of a given task.
6031 * @p: the task in question.
6033 int task_nice(const struct task_struct
*p
)
6035 return TASK_NICE(p
);
6037 EXPORT_SYMBOL(task_nice
);
6040 * idle_cpu - is a given cpu idle currently?
6041 * @cpu: the processor in question.
6043 int idle_cpu(int cpu
)
6045 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
6049 * idle_task - return the idle task for a given cpu.
6050 * @cpu: the processor in question.
6052 struct task_struct
*idle_task(int cpu
)
6054 return cpu_rq(cpu
)->idle
;
6058 * find_process_by_pid - find a process with a matching PID value.
6059 * @pid: the pid in question.
6061 static struct task_struct
*find_process_by_pid(pid_t pid
)
6063 return pid
? find_task_by_vpid(pid
) : current
;
6066 /* Actually do priority change: must hold rq lock. */
6068 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
6070 BUG_ON(p
->se
.on_rq
);
6073 switch (p
->policy
) {
6077 p
->sched_class
= &fair_sched_class
;
6081 p
->sched_class
= &rt_sched_class
;
6085 p
->rt_priority
= prio
;
6086 p
->normal_prio
= normal_prio(p
);
6087 /* we are holding p->pi_lock already */
6088 p
->prio
= rt_mutex_getprio(p
);
6093 * check the target process has a UID that matches the current process's
6095 static bool check_same_owner(struct task_struct
*p
)
6097 const struct cred
*cred
= current_cred(), *pcred
;
6101 pcred
= __task_cred(p
);
6102 match
= (cred
->euid
== pcred
->euid
||
6103 cred
->euid
== pcred
->uid
);
6108 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
6109 struct sched_param
*param
, bool user
)
6111 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
6112 unsigned long flags
;
6113 const struct sched_class
*prev_class
= p
->sched_class
;
6117 /* may grab non-irq protected spin_locks */
6118 BUG_ON(in_interrupt());
6120 /* double check policy once rq lock held */
6122 reset_on_fork
= p
->sched_reset_on_fork
;
6123 policy
= oldpolicy
= p
->policy
;
6125 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
6126 policy
&= ~SCHED_RESET_ON_FORK
;
6128 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
6129 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
6130 policy
!= SCHED_IDLE
)
6135 * Valid priorities for SCHED_FIFO and SCHED_RR are
6136 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6137 * SCHED_BATCH and SCHED_IDLE is 0.
6139 if (param
->sched_priority
< 0 ||
6140 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
6141 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
6143 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
6147 * Allow unprivileged RT tasks to decrease priority:
6149 if (user
&& !capable(CAP_SYS_NICE
)) {
6150 if (rt_policy(policy
)) {
6151 unsigned long rlim_rtprio
;
6153 if (!lock_task_sighand(p
, &flags
))
6155 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
6156 unlock_task_sighand(p
, &flags
);
6158 /* can't set/change the rt policy */
6159 if (policy
!= p
->policy
&& !rlim_rtprio
)
6162 /* can't increase priority */
6163 if (param
->sched_priority
> p
->rt_priority
&&
6164 param
->sched_priority
> rlim_rtprio
)
6168 * Like positive nice levels, dont allow tasks to
6169 * move out of SCHED_IDLE either:
6171 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
6174 /* can't change other user's priorities */
6175 if (!check_same_owner(p
))
6178 /* Normal users shall not reset the sched_reset_on_fork flag */
6179 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
6184 #ifdef CONFIG_RT_GROUP_SCHED
6186 * Do not allow realtime tasks into groups that have no runtime
6189 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
6190 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
6194 retval
= security_task_setscheduler(p
, policy
, param
);
6200 * make sure no PI-waiters arrive (or leave) while we are
6201 * changing the priority of the task:
6203 spin_lock_irqsave(&p
->pi_lock
, flags
);
6205 * To be able to change p->policy safely, the apropriate
6206 * runqueue lock must be held.
6208 rq
= __task_rq_lock(p
);
6209 /* recheck policy now with rq lock held */
6210 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
6211 policy
= oldpolicy
= -1;
6212 __task_rq_unlock(rq
);
6213 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6216 update_rq_clock(rq
);
6217 on_rq
= p
->se
.on_rq
;
6218 running
= task_current(rq
, p
);
6220 deactivate_task(rq
, p
, 0);
6222 p
->sched_class
->put_prev_task(rq
, p
);
6224 p
->sched_reset_on_fork
= reset_on_fork
;
6227 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
6230 p
->sched_class
->set_curr_task(rq
);
6232 activate_task(rq
, p
, 0);
6234 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
6236 __task_rq_unlock(rq
);
6237 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6239 rt_mutex_adjust_pi(p
);
6245 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6246 * @p: the task in question.
6247 * @policy: new policy.
6248 * @param: structure containing the new RT priority.
6250 * NOTE that the task may be already dead.
6252 int sched_setscheduler(struct task_struct
*p
, int policy
,
6253 struct sched_param
*param
)
6255 return __sched_setscheduler(p
, policy
, param
, true);
6257 EXPORT_SYMBOL_GPL(sched_setscheduler
);
6260 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6261 * @p: the task in question.
6262 * @policy: new policy.
6263 * @param: structure containing the new RT priority.
6265 * Just like sched_setscheduler, only don't bother checking if the
6266 * current context has permission. For example, this is needed in
6267 * stop_machine(): we create temporary high priority worker threads,
6268 * but our caller might not have that capability.
6270 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
6271 struct sched_param
*param
)
6273 return __sched_setscheduler(p
, policy
, param
, false);
6277 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
6279 struct sched_param lparam
;
6280 struct task_struct
*p
;
6283 if (!param
|| pid
< 0)
6285 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
6290 p
= find_process_by_pid(pid
);
6292 retval
= sched_setscheduler(p
, policy
, &lparam
);
6299 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6300 * @pid: the pid in question.
6301 * @policy: new policy.
6302 * @param: structure containing the new RT priority.
6304 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
6305 struct sched_param __user
*, param
)
6307 /* negative values for policy are not valid */
6311 return do_sched_setscheduler(pid
, policy
, param
);
6315 * sys_sched_setparam - set/change the RT priority of a thread
6316 * @pid: the pid in question.
6317 * @param: structure containing the new RT priority.
6319 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6321 return do_sched_setscheduler(pid
, -1, param
);
6325 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6326 * @pid: the pid in question.
6328 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
6330 struct task_struct
*p
;
6337 read_lock(&tasklist_lock
);
6338 p
= find_process_by_pid(pid
);
6340 retval
= security_task_getscheduler(p
);
6343 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
6345 read_unlock(&tasklist_lock
);
6350 * sys_sched_getparam - get the RT priority of a thread
6351 * @pid: the pid in question.
6352 * @param: structure containing the RT priority.
6354 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6356 struct sched_param lp
;
6357 struct task_struct
*p
;
6360 if (!param
|| pid
< 0)
6363 read_lock(&tasklist_lock
);
6364 p
= find_process_by_pid(pid
);
6369 retval
= security_task_getscheduler(p
);
6373 lp
.sched_priority
= p
->rt_priority
;
6374 read_unlock(&tasklist_lock
);
6377 * This one might sleep, we cannot do it with a spinlock held ...
6379 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
6384 read_unlock(&tasklist_lock
);
6388 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
6390 cpumask_var_t cpus_allowed
, new_mask
;
6391 struct task_struct
*p
;
6395 read_lock(&tasklist_lock
);
6397 p
= find_process_by_pid(pid
);
6399 read_unlock(&tasklist_lock
);
6405 * It is not safe to call set_cpus_allowed with the
6406 * tasklist_lock held. We will bump the task_struct's
6407 * usage count and then drop tasklist_lock.
6410 read_unlock(&tasklist_lock
);
6412 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
6416 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
6418 goto out_free_cpus_allowed
;
6421 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
6424 retval
= security_task_setscheduler(p
, 0, NULL
);
6428 cpuset_cpus_allowed(p
, cpus_allowed
);
6429 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
6431 retval
= set_cpus_allowed_ptr(p
, new_mask
);
6434 cpuset_cpus_allowed(p
, cpus_allowed
);
6435 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
6437 * We must have raced with a concurrent cpuset
6438 * update. Just reset the cpus_allowed to the
6439 * cpuset's cpus_allowed
6441 cpumask_copy(new_mask
, cpus_allowed
);
6446 free_cpumask_var(new_mask
);
6447 out_free_cpus_allowed
:
6448 free_cpumask_var(cpus_allowed
);
6455 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
6456 struct cpumask
*new_mask
)
6458 if (len
< cpumask_size())
6459 cpumask_clear(new_mask
);
6460 else if (len
> cpumask_size())
6461 len
= cpumask_size();
6463 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
6467 * sys_sched_setaffinity - set the cpu affinity of a process
6468 * @pid: pid of the process
6469 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6470 * @user_mask_ptr: user-space pointer to the new cpu mask
6472 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
6473 unsigned long __user
*, user_mask_ptr
)
6475 cpumask_var_t new_mask
;
6478 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
6481 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
6483 retval
= sched_setaffinity(pid
, new_mask
);
6484 free_cpumask_var(new_mask
);
6488 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
6490 struct task_struct
*p
;
6494 read_lock(&tasklist_lock
);
6497 p
= find_process_by_pid(pid
);
6501 retval
= security_task_getscheduler(p
);
6505 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
6508 read_unlock(&tasklist_lock
);
6515 * sys_sched_getaffinity - get the cpu affinity of a process
6516 * @pid: pid of the process
6517 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6518 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6520 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
6521 unsigned long __user
*, user_mask_ptr
)
6526 if (len
< cpumask_size())
6529 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
6532 ret
= sched_getaffinity(pid
, mask
);
6534 if (copy_to_user(user_mask_ptr
, mask
, cpumask_size()))
6537 ret
= cpumask_size();
6539 free_cpumask_var(mask
);
6545 * sys_sched_yield - yield the current processor to other threads.
6547 * This function yields the current CPU to other tasks. If there are no
6548 * other threads running on this CPU then this function will return.
6550 SYSCALL_DEFINE0(sched_yield
)
6552 struct rq
*rq
= this_rq_lock();
6554 schedstat_inc(rq
, yld_count
);
6555 current
->sched_class
->yield_task(rq
);
6558 * Since we are going to call schedule() anyway, there's
6559 * no need to preempt or enable interrupts:
6561 __release(rq
->lock
);
6562 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
6563 _raw_spin_unlock(&rq
->lock
);
6564 preempt_enable_no_resched();
6571 static void __cond_resched(void)
6573 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6574 __might_sleep(__FILE__
, __LINE__
);
6577 * The BKS might be reacquired before we have dropped
6578 * PREEMPT_ACTIVE, which could trigger a second
6579 * cond_resched() call.
6582 add_preempt_count(PREEMPT_ACTIVE
);
6584 sub_preempt_count(PREEMPT_ACTIVE
);
6585 } while (need_resched());
6588 int __sched
_cond_resched(void)
6590 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
6591 system_state
== SYSTEM_RUNNING
) {
6597 EXPORT_SYMBOL(_cond_resched
);
6600 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
6601 * call schedule, and on return reacquire the lock.
6603 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6604 * operations here to prevent schedule() from being called twice (once via
6605 * spin_unlock(), once by hand).
6607 int cond_resched_lock(spinlock_t
*lock
)
6609 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
6612 if (spin_needbreak(lock
) || resched
) {
6614 if (resched
&& need_resched())
6623 EXPORT_SYMBOL(cond_resched_lock
);
6625 int __sched
cond_resched_softirq(void)
6627 BUG_ON(!in_softirq());
6629 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
6637 EXPORT_SYMBOL(cond_resched_softirq
);
6640 * yield - yield the current processor to other threads.
6642 * This is a shortcut for kernel-space yielding - it marks the
6643 * thread runnable and calls sys_sched_yield().
6645 void __sched
yield(void)
6647 set_current_state(TASK_RUNNING
);
6650 EXPORT_SYMBOL(yield
);
6653 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6654 * that process accounting knows that this is a task in IO wait state.
6656 * But don't do that if it is a deliberate, throttling IO wait (this task
6657 * has set its backing_dev_info: the queue against which it should throttle)
6659 void __sched
io_schedule(void)
6661 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
6663 delayacct_blkio_start();
6664 atomic_inc(&rq
->nr_iowait
);
6666 atomic_dec(&rq
->nr_iowait
);
6667 delayacct_blkio_end();
6669 EXPORT_SYMBOL(io_schedule
);
6671 long __sched
io_schedule_timeout(long timeout
)
6673 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
6676 delayacct_blkio_start();
6677 atomic_inc(&rq
->nr_iowait
);
6678 ret
= schedule_timeout(timeout
);
6679 atomic_dec(&rq
->nr_iowait
);
6680 delayacct_blkio_end();
6685 * sys_sched_get_priority_max - return maximum RT priority.
6686 * @policy: scheduling class.
6688 * this syscall returns the maximum rt_priority that can be used
6689 * by a given scheduling class.
6691 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
6698 ret
= MAX_USER_RT_PRIO
-1;
6710 * sys_sched_get_priority_min - return minimum RT priority.
6711 * @policy: scheduling class.
6713 * this syscall returns the minimum rt_priority that can be used
6714 * by a given scheduling class.
6716 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
6734 * sys_sched_rr_get_interval - return the default timeslice of a process.
6735 * @pid: pid of the process.
6736 * @interval: userspace pointer to the timeslice value.
6738 * this syscall writes the default timeslice value of a given process
6739 * into the user-space timespec buffer. A value of '0' means infinity.
6741 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
6742 struct timespec __user
*, interval
)
6744 struct task_struct
*p
;
6745 unsigned int time_slice
;
6753 read_lock(&tasklist_lock
);
6754 p
= find_process_by_pid(pid
);
6758 retval
= security_task_getscheduler(p
);
6763 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6764 * tasks that are on an otherwise idle runqueue:
6767 if (p
->policy
== SCHED_RR
) {
6768 time_slice
= DEF_TIMESLICE
;
6769 } else if (p
->policy
!= SCHED_FIFO
) {
6770 struct sched_entity
*se
= &p
->se
;
6771 unsigned long flags
;
6774 rq
= task_rq_lock(p
, &flags
);
6775 if (rq
->cfs
.load
.weight
)
6776 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
6777 task_rq_unlock(rq
, &flags
);
6779 read_unlock(&tasklist_lock
);
6780 jiffies_to_timespec(time_slice
, &t
);
6781 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
6785 read_unlock(&tasklist_lock
);
6789 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
6791 void sched_show_task(struct task_struct
*p
)
6793 unsigned long free
= 0;
6796 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
6797 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
6798 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
6799 #if BITS_PER_LONG == 32
6800 if (state
== TASK_RUNNING
)
6801 printk(KERN_CONT
" running ");
6803 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
6805 if (state
== TASK_RUNNING
)
6806 printk(KERN_CONT
" running task ");
6808 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
6810 #ifdef CONFIG_DEBUG_STACK_USAGE
6811 free
= stack_not_used(p
);
6813 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
6814 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
6815 (unsigned long)task_thread_info(p
)->flags
);
6817 show_stack(p
, NULL
);
6820 void show_state_filter(unsigned long state_filter
)
6822 struct task_struct
*g
, *p
;
6824 #if BITS_PER_LONG == 32
6826 " task PC stack pid father\n");
6829 " task PC stack pid father\n");
6831 read_lock(&tasklist_lock
);
6832 do_each_thread(g
, p
) {
6834 * reset the NMI-timeout, listing all files on a slow
6835 * console might take alot of time:
6837 touch_nmi_watchdog();
6838 if (!state_filter
|| (p
->state
& state_filter
))
6840 } while_each_thread(g
, p
);
6842 touch_all_softlockup_watchdogs();
6844 #ifdef CONFIG_SCHED_DEBUG
6845 sysrq_sched_debug_show();
6847 read_unlock(&tasklist_lock
);
6849 * Only show locks if all tasks are dumped:
6851 if (state_filter
== -1)
6852 debug_show_all_locks();
6855 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
6857 idle
->sched_class
= &idle_sched_class
;
6861 * init_idle - set up an idle thread for a given CPU
6862 * @idle: task in question
6863 * @cpu: cpu the idle task belongs to
6865 * NOTE: this function does not set the idle thread's NEED_RESCHED
6866 * flag, to make booting more robust.
6868 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
6870 struct rq
*rq
= cpu_rq(cpu
);
6871 unsigned long flags
;
6873 spin_lock_irqsave(&rq
->lock
, flags
);
6876 idle
->se
.exec_start
= sched_clock();
6878 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
6879 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
6880 __set_task_cpu(idle
, cpu
);
6882 rq
->curr
= rq
->idle
= idle
;
6883 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6886 spin_unlock_irqrestore(&rq
->lock
, flags
);
6888 /* Set the preempt count _outside_ the spinlocks! */
6889 #if defined(CONFIG_PREEMPT)
6890 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
6892 task_thread_info(idle
)->preempt_count
= 0;
6895 * The idle tasks have their own, simple scheduling class:
6897 idle
->sched_class
= &idle_sched_class
;
6898 ftrace_graph_init_task(idle
);
6902 * In a system that switches off the HZ timer nohz_cpu_mask
6903 * indicates which cpus entered this state. This is used
6904 * in the rcu update to wait only for active cpus. For system
6905 * which do not switch off the HZ timer nohz_cpu_mask should
6906 * always be CPU_BITS_NONE.
6908 cpumask_var_t nohz_cpu_mask
;
6911 * Increase the granularity value when there are more CPUs,
6912 * because with more CPUs the 'effective latency' as visible
6913 * to users decreases. But the relationship is not linear,
6914 * so pick a second-best guess by going with the log2 of the
6917 * This idea comes from the SD scheduler of Con Kolivas:
6919 static inline void sched_init_granularity(void)
6921 unsigned int factor
= 1 + ilog2(num_online_cpus());
6922 const unsigned long limit
= 200000000;
6924 sysctl_sched_min_granularity
*= factor
;
6925 if (sysctl_sched_min_granularity
> limit
)
6926 sysctl_sched_min_granularity
= limit
;
6928 sysctl_sched_latency
*= factor
;
6929 if (sysctl_sched_latency
> limit
)
6930 sysctl_sched_latency
= limit
;
6932 sysctl_sched_wakeup_granularity
*= factor
;
6934 sysctl_sched_shares_ratelimit
*= factor
;
6939 * This is how migration works:
6941 * 1) we queue a struct migration_req structure in the source CPU's
6942 * runqueue and wake up that CPU's migration thread.
6943 * 2) we down() the locked semaphore => thread blocks.
6944 * 3) migration thread wakes up (implicitly it forces the migrated
6945 * thread off the CPU)
6946 * 4) it gets the migration request and checks whether the migrated
6947 * task is still in the wrong runqueue.
6948 * 5) if it's in the wrong runqueue then the migration thread removes
6949 * it and puts it into the right queue.
6950 * 6) migration thread up()s the semaphore.
6951 * 7) we wake up and the migration is done.
6955 * Change a given task's CPU affinity. Migrate the thread to a
6956 * proper CPU and schedule it away if the CPU it's executing on
6957 * is removed from the allowed bitmask.
6959 * NOTE: the caller must have a valid reference to the task, the
6960 * task must not exit() & deallocate itself prematurely. The
6961 * call is not atomic; no spinlocks may be held.
6963 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
6965 struct migration_req req
;
6966 unsigned long flags
;
6970 rq
= task_rq_lock(p
, &flags
);
6971 if (!cpumask_intersects(new_mask
, cpu_online_mask
)) {
6976 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
6977 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
6982 if (p
->sched_class
->set_cpus_allowed
)
6983 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
6985 cpumask_copy(&p
->cpus_allowed
, new_mask
);
6986 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
6989 /* Can the task run on the task's current CPU? If so, we're done */
6990 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
6993 if (migrate_task(p
, cpumask_any_and(cpu_online_mask
, new_mask
), &req
)) {
6994 /* Need help from migration thread: drop lock and wait. */
6995 task_rq_unlock(rq
, &flags
);
6996 wake_up_process(rq
->migration_thread
);
6997 wait_for_completion(&req
.done
);
6998 tlb_migrate_finish(p
->mm
);
7002 task_rq_unlock(rq
, &flags
);
7006 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
7009 * Move (not current) task off this cpu, onto dest cpu. We're doing
7010 * this because either it can't run here any more (set_cpus_allowed()
7011 * away from this CPU, or CPU going down), or because we're
7012 * attempting to rebalance this task on exec (sched_exec).
7014 * So we race with normal scheduler movements, but that's OK, as long
7015 * as the task is no longer on this CPU.
7017 * Returns non-zero if task was successfully migrated.
7019 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
7021 struct rq
*rq_dest
, *rq_src
;
7024 if (unlikely(!cpu_active(dest_cpu
)))
7027 rq_src
= cpu_rq(src_cpu
);
7028 rq_dest
= cpu_rq(dest_cpu
);
7030 double_rq_lock(rq_src
, rq_dest
);
7031 /* Already moved. */
7032 if (task_cpu(p
) != src_cpu
)
7034 /* Affinity changed (again). */
7035 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
7038 on_rq
= p
->se
.on_rq
;
7040 deactivate_task(rq_src
, p
, 0);
7042 set_task_cpu(p
, dest_cpu
);
7044 activate_task(rq_dest
, p
, 0);
7045 check_preempt_curr(rq_dest
, p
, 0);
7050 double_rq_unlock(rq_src
, rq_dest
);
7055 * migration_thread - this is a highprio system thread that performs
7056 * thread migration by bumping thread off CPU then 'pushing' onto
7059 static int migration_thread(void *data
)
7061 int cpu
= (long)data
;
7065 BUG_ON(rq
->migration_thread
!= current
);
7067 set_current_state(TASK_INTERRUPTIBLE
);
7068 while (!kthread_should_stop()) {
7069 struct migration_req
*req
;
7070 struct list_head
*head
;
7072 spin_lock_irq(&rq
->lock
);
7074 if (cpu_is_offline(cpu
)) {
7075 spin_unlock_irq(&rq
->lock
);
7079 if (rq
->active_balance
) {
7080 active_load_balance(rq
, cpu
);
7081 rq
->active_balance
= 0;
7084 head
= &rq
->migration_queue
;
7086 if (list_empty(head
)) {
7087 spin_unlock_irq(&rq
->lock
);
7089 set_current_state(TASK_INTERRUPTIBLE
);
7092 req
= list_entry(head
->next
, struct migration_req
, list
);
7093 list_del_init(head
->next
);
7095 spin_unlock(&rq
->lock
);
7096 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
7099 complete(&req
->done
);
7101 __set_current_state(TASK_RUNNING
);
7105 /* Wait for kthread_stop */
7106 set_current_state(TASK_INTERRUPTIBLE
);
7107 while (!kthread_should_stop()) {
7109 set_current_state(TASK_INTERRUPTIBLE
);
7111 __set_current_state(TASK_RUNNING
);
7115 #ifdef CONFIG_HOTPLUG_CPU
7117 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
7121 local_irq_disable();
7122 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
7128 * Figure out where task on dead CPU should go, use force if necessary.
7130 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
7133 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(dead_cpu
));
7136 /* Look for allowed, online CPU in same node. */
7137 for_each_cpu_and(dest_cpu
, nodemask
, cpu_online_mask
)
7138 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
7141 /* Any allowed, online CPU? */
7142 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_online_mask
);
7143 if (dest_cpu
< nr_cpu_ids
)
7146 /* No more Mr. Nice Guy. */
7147 if (dest_cpu
>= nr_cpu_ids
) {
7148 cpuset_cpus_allowed_locked(p
, &p
->cpus_allowed
);
7149 dest_cpu
= cpumask_any_and(cpu_online_mask
, &p
->cpus_allowed
);
7152 * Don't tell them about moving exiting tasks or
7153 * kernel threads (both mm NULL), since they never
7156 if (p
->mm
&& printk_ratelimit()) {
7157 printk(KERN_INFO
"process %d (%s) no "
7158 "longer affine to cpu%d\n",
7159 task_pid_nr(p
), p
->comm
, dead_cpu
);
7164 /* It can have affinity changed while we were choosing. */
7165 if (unlikely(!__migrate_task_irq(p
, dead_cpu
, dest_cpu
)))
7170 * While a dead CPU has no uninterruptible tasks queued at this point,
7171 * it might still have a nonzero ->nr_uninterruptible counter, because
7172 * for performance reasons the counter is not stricly tracking tasks to
7173 * their home CPUs. So we just add the counter to another CPU's counter,
7174 * to keep the global sum constant after CPU-down:
7176 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
7178 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_online_mask
));
7179 unsigned long flags
;
7181 local_irq_save(flags
);
7182 double_rq_lock(rq_src
, rq_dest
);
7183 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
7184 rq_src
->nr_uninterruptible
= 0;
7185 double_rq_unlock(rq_src
, rq_dest
);
7186 local_irq_restore(flags
);
7189 /* Run through task list and migrate tasks from the dead cpu. */
7190 static void migrate_live_tasks(int src_cpu
)
7192 struct task_struct
*p
, *t
;
7194 read_lock(&tasklist_lock
);
7196 do_each_thread(t
, p
) {
7200 if (task_cpu(p
) == src_cpu
)
7201 move_task_off_dead_cpu(src_cpu
, p
);
7202 } while_each_thread(t
, p
);
7204 read_unlock(&tasklist_lock
);
7208 * Schedules idle task to be the next runnable task on current CPU.
7209 * It does so by boosting its priority to highest possible.
7210 * Used by CPU offline code.
7212 void sched_idle_next(void)
7214 int this_cpu
= smp_processor_id();
7215 struct rq
*rq
= cpu_rq(this_cpu
);
7216 struct task_struct
*p
= rq
->idle
;
7217 unsigned long flags
;
7219 /* cpu has to be offline */
7220 BUG_ON(cpu_online(this_cpu
));
7223 * Strictly not necessary since rest of the CPUs are stopped by now
7224 * and interrupts disabled on the current cpu.
7226 spin_lock_irqsave(&rq
->lock
, flags
);
7228 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7230 update_rq_clock(rq
);
7231 activate_task(rq
, p
, 0);
7233 spin_unlock_irqrestore(&rq
->lock
, flags
);
7237 * Ensures that the idle task is using init_mm right before its cpu goes
7240 void idle_task_exit(void)
7242 struct mm_struct
*mm
= current
->active_mm
;
7244 BUG_ON(cpu_online(smp_processor_id()));
7247 switch_mm(mm
, &init_mm
, current
);
7251 /* called under rq->lock with disabled interrupts */
7252 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
7254 struct rq
*rq
= cpu_rq(dead_cpu
);
7256 /* Must be exiting, otherwise would be on tasklist. */
7257 BUG_ON(!p
->exit_state
);
7259 /* Cannot have done final schedule yet: would have vanished. */
7260 BUG_ON(p
->state
== TASK_DEAD
);
7265 * Drop lock around migration; if someone else moves it,
7266 * that's OK. No task can be added to this CPU, so iteration is
7269 spin_unlock_irq(&rq
->lock
);
7270 move_task_off_dead_cpu(dead_cpu
, p
);
7271 spin_lock_irq(&rq
->lock
);
7276 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7277 static void migrate_dead_tasks(unsigned int dead_cpu
)
7279 struct rq
*rq
= cpu_rq(dead_cpu
);
7280 struct task_struct
*next
;
7283 if (!rq
->nr_running
)
7285 update_rq_clock(rq
);
7286 next
= pick_next_task(rq
);
7289 next
->sched_class
->put_prev_task(rq
, next
);
7290 migrate_dead(dead_cpu
, next
);
7296 * remove the tasks which were accounted by rq from calc_load_tasks.
7298 static void calc_global_load_remove(struct rq
*rq
)
7300 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
7302 #endif /* CONFIG_HOTPLUG_CPU */
7304 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7306 static struct ctl_table sd_ctl_dir
[] = {
7308 .procname
= "sched_domain",
7314 static struct ctl_table sd_ctl_root
[] = {
7316 .ctl_name
= CTL_KERN
,
7317 .procname
= "kernel",
7319 .child
= sd_ctl_dir
,
7324 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
7326 struct ctl_table
*entry
=
7327 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
7332 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
7334 struct ctl_table
*entry
;
7337 * In the intermediate directories, both the child directory and
7338 * procname are dynamically allocated and could fail but the mode
7339 * will always be set. In the lowest directory the names are
7340 * static strings and all have proc handlers.
7342 for (entry
= *tablep
; entry
->mode
; entry
++) {
7344 sd_free_ctl_entry(&entry
->child
);
7345 if (entry
->proc_handler
== NULL
)
7346 kfree(entry
->procname
);
7354 set_table_entry(struct ctl_table
*entry
,
7355 const char *procname
, void *data
, int maxlen
,
7356 mode_t mode
, proc_handler
*proc_handler
)
7358 entry
->procname
= procname
;
7360 entry
->maxlen
= maxlen
;
7362 entry
->proc_handler
= proc_handler
;
7365 static struct ctl_table
*
7366 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
7368 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
7373 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
7374 sizeof(long), 0644, proc_doulongvec_minmax
);
7375 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
7376 sizeof(long), 0644, proc_doulongvec_minmax
);
7377 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
7378 sizeof(int), 0644, proc_dointvec_minmax
);
7379 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
7380 sizeof(int), 0644, proc_dointvec_minmax
);
7381 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
7382 sizeof(int), 0644, proc_dointvec_minmax
);
7383 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
7384 sizeof(int), 0644, proc_dointvec_minmax
);
7385 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
7386 sizeof(int), 0644, proc_dointvec_minmax
);
7387 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
7388 sizeof(int), 0644, proc_dointvec_minmax
);
7389 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
7390 sizeof(int), 0644, proc_dointvec_minmax
);
7391 set_table_entry(&table
[9], "cache_nice_tries",
7392 &sd
->cache_nice_tries
,
7393 sizeof(int), 0644, proc_dointvec_minmax
);
7394 set_table_entry(&table
[10], "flags", &sd
->flags
,
7395 sizeof(int), 0644, proc_dointvec_minmax
);
7396 set_table_entry(&table
[11], "name", sd
->name
,
7397 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
7398 /* &table[12] is terminator */
7403 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
7405 struct ctl_table
*entry
, *table
;
7406 struct sched_domain
*sd
;
7407 int domain_num
= 0, i
;
7410 for_each_domain(cpu
, sd
)
7412 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
7417 for_each_domain(cpu
, sd
) {
7418 snprintf(buf
, 32, "domain%d", i
);
7419 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7421 entry
->child
= sd_alloc_ctl_domain_table(sd
);
7428 static struct ctl_table_header
*sd_sysctl_header
;
7429 static void register_sched_domain_sysctl(void)
7431 int i
, cpu_num
= num_online_cpus();
7432 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
7435 WARN_ON(sd_ctl_dir
[0].child
);
7436 sd_ctl_dir
[0].child
= entry
;
7441 for_each_online_cpu(i
) {
7442 snprintf(buf
, 32, "cpu%d", i
);
7443 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7445 entry
->child
= sd_alloc_ctl_cpu_table(i
);
7449 WARN_ON(sd_sysctl_header
);
7450 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
7453 /* may be called multiple times per register */
7454 static void unregister_sched_domain_sysctl(void)
7456 if (sd_sysctl_header
)
7457 unregister_sysctl_table(sd_sysctl_header
);
7458 sd_sysctl_header
= NULL
;
7459 if (sd_ctl_dir
[0].child
)
7460 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
7463 static void register_sched_domain_sysctl(void)
7466 static void unregister_sched_domain_sysctl(void)
7471 static void set_rq_online(struct rq
*rq
)
7474 const struct sched_class
*class;
7476 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
7479 for_each_class(class) {
7480 if (class->rq_online
)
7481 class->rq_online(rq
);
7486 static void set_rq_offline(struct rq
*rq
)
7489 const struct sched_class
*class;
7491 for_each_class(class) {
7492 if (class->rq_offline
)
7493 class->rq_offline(rq
);
7496 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
7502 * migration_call - callback that gets triggered when a CPU is added.
7503 * Here we can start up the necessary migration thread for the new CPU.
7505 static int __cpuinit
7506 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
7508 struct task_struct
*p
;
7509 int cpu
= (long)hcpu
;
7510 unsigned long flags
;
7515 case CPU_UP_PREPARE
:
7516 case CPU_UP_PREPARE_FROZEN
:
7517 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
7520 kthread_bind(p
, cpu
);
7521 /* Must be high prio: stop_machine expects to yield to it. */
7522 rq
= task_rq_lock(p
, &flags
);
7523 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7524 task_rq_unlock(rq
, &flags
);
7525 cpu_rq(cpu
)->migration_thread
= p
;
7529 case CPU_ONLINE_FROZEN
:
7530 /* Strictly unnecessary, as first user will wake it. */
7531 wake_up_process(cpu_rq(cpu
)->migration_thread
);
7533 /* Update our root-domain */
7535 spin_lock_irqsave(&rq
->lock
, flags
);
7536 rq
->calc_load_update
= calc_load_update
;
7537 rq
->calc_load_active
= 0;
7539 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7543 spin_unlock_irqrestore(&rq
->lock
, flags
);
7546 #ifdef CONFIG_HOTPLUG_CPU
7547 case CPU_UP_CANCELED
:
7548 case CPU_UP_CANCELED_FROZEN
:
7549 if (!cpu_rq(cpu
)->migration_thread
)
7551 /* Unbind it from offline cpu so it can run. Fall thru. */
7552 kthread_bind(cpu_rq(cpu
)->migration_thread
,
7553 cpumask_any(cpu_online_mask
));
7554 kthread_stop(cpu_rq(cpu
)->migration_thread
);
7555 cpu_rq(cpu
)->migration_thread
= NULL
;
7559 case CPU_DEAD_FROZEN
:
7560 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7561 migrate_live_tasks(cpu
);
7563 kthread_stop(rq
->migration_thread
);
7564 rq
->migration_thread
= NULL
;
7565 /* Idle task back to normal (off runqueue, low prio) */
7566 spin_lock_irq(&rq
->lock
);
7567 update_rq_clock(rq
);
7568 deactivate_task(rq
, rq
->idle
, 0);
7569 rq
->idle
->static_prio
= MAX_PRIO
;
7570 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
7571 rq
->idle
->sched_class
= &idle_sched_class
;
7572 migrate_dead_tasks(cpu
);
7573 spin_unlock_irq(&rq
->lock
);
7575 migrate_nr_uninterruptible(rq
);
7576 BUG_ON(rq
->nr_running
!= 0);
7577 calc_global_load_remove(rq
);
7579 * No need to migrate the tasks: it was best-effort if
7580 * they didn't take sched_hotcpu_mutex. Just wake up
7583 spin_lock_irq(&rq
->lock
);
7584 while (!list_empty(&rq
->migration_queue
)) {
7585 struct migration_req
*req
;
7587 req
= list_entry(rq
->migration_queue
.next
,
7588 struct migration_req
, list
);
7589 list_del_init(&req
->list
);
7590 spin_unlock_irq(&rq
->lock
);
7591 complete(&req
->done
);
7592 spin_lock_irq(&rq
->lock
);
7594 spin_unlock_irq(&rq
->lock
);
7598 case CPU_DYING_FROZEN
:
7599 /* Update our root-domain */
7601 spin_lock_irqsave(&rq
->lock
, flags
);
7603 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7606 spin_unlock_irqrestore(&rq
->lock
, flags
);
7614 * Register at high priority so that task migration (migrate_all_tasks)
7615 * happens before everything else. This has to be lower priority than
7616 * the notifier in the perf_counter subsystem, though.
7618 static struct notifier_block __cpuinitdata migration_notifier
= {
7619 .notifier_call
= migration_call
,
7623 static int __init
migration_init(void)
7625 void *cpu
= (void *)(long)smp_processor_id();
7628 /* Start one for the boot CPU: */
7629 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
7630 BUG_ON(err
== NOTIFY_BAD
);
7631 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
7632 register_cpu_notifier(&migration_notifier
);
7636 early_initcall(migration_init
);
7641 #ifdef CONFIG_SCHED_DEBUG
7643 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
7644 struct cpumask
*groupmask
)
7646 struct sched_group
*group
= sd
->groups
;
7649 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
7650 cpumask_clear(groupmask
);
7652 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
7654 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
7655 printk("does not load-balance\n");
7657 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
7662 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
7664 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
7665 printk(KERN_ERR
"ERROR: domain->span does not contain "
7668 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
7669 printk(KERN_ERR
"ERROR: domain->groups does not contain"
7673 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
7677 printk(KERN_ERR
"ERROR: group is NULL\n");
7681 if (!group
->__cpu_power
) {
7682 printk(KERN_CONT
"\n");
7683 printk(KERN_ERR
"ERROR: domain->cpu_power not "
7688 if (!cpumask_weight(sched_group_cpus(group
))) {
7689 printk(KERN_CONT
"\n");
7690 printk(KERN_ERR
"ERROR: empty group\n");
7694 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
7695 printk(KERN_CONT
"\n");
7696 printk(KERN_ERR
"ERROR: repeated CPUs\n");
7700 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
7702 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
7704 printk(KERN_CONT
" %s", str
);
7705 if (group
->__cpu_power
!= SCHED_LOAD_SCALE
) {
7706 printk(KERN_CONT
" (__cpu_power = %d)",
7707 group
->__cpu_power
);
7710 group
= group
->next
;
7711 } while (group
!= sd
->groups
);
7712 printk(KERN_CONT
"\n");
7714 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
7715 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
7718 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
7719 printk(KERN_ERR
"ERROR: parent span is not a superset "
7720 "of domain->span\n");
7724 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
7726 cpumask_var_t groupmask
;
7730 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
7734 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
7736 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
7737 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
7742 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
7749 free_cpumask_var(groupmask
);
7751 #else /* !CONFIG_SCHED_DEBUG */
7752 # define sched_domain_debug(sd, cpu) do { } while (0)
7753 #endif /* CONFIG_SCHED_DEBUG */
7755 static int sd_degenerate(struct sched_domain
*sd
)
7757 if (cpumask_weight(sched_domain_span(sd
)) == 1)
7760 /* Following flags need at least 2 groups */
7761 if (sd
->flags
& (SD_LOAD_BALANCE
|
7762 SD_BALANCE_NEWIDLE
|
7766 SD_SHARE_PKG_RESOURCES
)) {
7767 if (sd
->groups
!= sd
->groups
->next
)
7771 /* Following flags don't use groups */
7772 if (sd
->flags
& (SD_WAKE_IDLE
|
7781 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
7783 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
7785 if (sd_degenerate(parent
))
7788 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
7791 /* Does parent contain flags not in child? */
7792 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
7793 if (cflags
& SD_WAKE_AFFINE
)
7794 pflags
&= ~SD_WAKE_BALANCE
;
7795 /* Flags needing groups don't count if only 1 group in parent */
7796 if (parent
->groups
== parent
->groups
->next
) {
7797 pflags
&= ~(SD_LOAD_BALANCE
|
7798 SD_BALANCE_NEWIDLE
|
7802 SD_SHARE_PKG_RESOURCES
);
7803 if (nr_node_ids
== 1)
7804 pflags
&= ~SD_SERIALIZE
;
7806 if (~cflags
& pflags
)
7812 static void free_rootdomain(struct root_domain
*rd
)
7814 cpupri_cleanup(&rd
->cpupri
);
7816 free_cpumask_var(rd
->rto_mask
);
7817 free_cpumask_var(rd
->online
);
7818 free_cpumask_var(rd
->span
);
7822 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
7824 struct root_domain
*old_rd
= NULL
;
7825 unsigned long flags
;
7827 spin_lock_irqsave(&rq
->lock
, flags
);
7832 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
7835 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
7838 * If we dont want to free the old_rt yet then
7839 * set old_rd to NULL to skip the freeing later
7842 if (!atomic_dec_and_test(&old_rd
->refcount
))
7846 atomic_inc(&rd
->refcount
);
7849 cpumask_set_cpu(rq
->cpu
, rd
->span
);
7850 if (cpumask_test_cpu(rq
->cpu
, cpu_online_mask
))
7853 spin_unlock_irqrestore(&rq
->lock
, flags
);
7856 free_rootdomain(old_rd
);
7859 static int __init_refok
init_rootdomain(struct root_domain
*rd
, bool bootmem
)
7861 gfp_t gfp
= GFP_KERNEL
;
7863 memset(rd
, 0, sizeof(*rd
));
7868 if (!alloc_cpumask_var(&rd
->span
, gfp
))
7870 if (!alloc_cpumask_var(&rd
->online
, gfp
))
7872 if (!alloc_cpumask_var(&rd
->rto_mask
, gfp
))
7875 if (cpupri_init(&rd
->cpupri
, bootmem
) != 0)
7880 free_cpumask_var(rd
->rto_mask
);
7882 free_cpumask_var(rd
->online
);
7884 free_cpumask_var(rd
->span
);
7889 static void init_defrootdomain(void)
7891 init_rootdomain(&def_root_domain
, true);
7893 atomic_set(&def_root_domain
.refcount
, 1);
7896 static struct root_domain
*alloc_rootdomain(void)
7898 struct root_domain
*rd
;
7900 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
7904 if (init_rootdomain(rd
, false) != 0) {
7913 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7914 * hold the hotplug lock.
7917 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
7919 struct rq
*rq
= cpu_rq(cpu
);
7920 struct sched_domain
*tmp
;
7922 /* Remove the sched domains which do not contribute to scheduling. */
7923 for (tmp
= sd
; tmp
; ) {
7924 struct sched_domain
*parent
= tmp
->parent
;
7928 if (sd_parent_degenerate(tmp
, parent
)) {
7929 tmp
->parent
= parent
->parent
;
7931 parent
->parent
->child
= tmp
;
7936 if (sd
&& sd_degenerate(sd
)) {
7942 sched_domain_debug(sd
, cpu
);
7944 rq_attach_root(rq
, rd
);
7945 rcu_assign_pointer(rq
->sd
, sd
);
7948 /* cpus with isolated domains */
7949 static cpumask_var_t cpu_isolated_map
;
7951 /* Setup the mask of cpus configured for isolated domains */
7952 static int __init
isolated_cpu_setup(char *str
)
7954 cpulist_parse(str
, cpu_isolated_map
);
7958 __setup("isolcpus=", isolated_cpu_setup
);
7961 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7962 * to a function which identifies what group(along with sched group) a CPU
7963 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7964 * (due to the fact that we keep track of groups covered with a struct cpumask).
7966 * init_sched_build_groups will build a circular linked list of the groups
7967 * covered by the given span, and will set each group's ->cpumask correctly,
7968 * and ->cpu_power to 0.
7971 init_sched_build_groups(const struct cpumask
*span
,
7972 const struct cpumask
*cpu_map
,
7973 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
7974 struct sched_group
**sg
,
7975 struct cpumask
*tmpmask
),
7976 struct cpumask
*covered
, struct cpumask
*tmpmask
)
7978 struct sched_group
*first
= NULL
, *last
= NULL
;
7981 cpumask_clear(covered
);
7983 for_each_cpu(i
, span
) {
7984 struct sched_group
*sg
;
7985 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
7988 if (cpumask_test_cpu(i
, covered
))
7991 cpumask_clear(sched_group_cpus(sg
));
7992 sg
->__cpu_power
= 0;
7994 for_each_cpu(j
, span
) {
7995 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
7998 cpumask_set_cpu(j
, covered
);
7999 cpumask_set_cpu(j
, sched_group_cpus(sg
));
8010 #define SD_NODES_PER_DOMAIN 16
8015 * find_next_best_node - find the next node to include in a sched_domain
8016 * @node: node whose sched_domain we're building
8017 * @used_nodes: nodes already in the sched_domain
8019 * Find the next node to include in a given scheduling domain. Simply
8020 * finds the closest node not already in the @used_nodes map.
8022 * Should use nodemask_t.
8024 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
8026 int i
, n
, val
, min_val
, best_node
= 0;
8030 for (i
= 0; i
< nr_node_ids
; i
++) {
8031 /* Start at @node */
8032 n
= (node
+ i
) % nr_node_ids
;
8034 if (!nr_cpus_node(n
))
8037 /* Skip already used nodes */
8038 if (node_isset(n
, *used_nodes
))
8041 /* Simple min distance search */
8042 val
= node_distance(node
, n
);
8044 if (val
< min_val
) {
8050 node_set(best_node
, *used_nodes
);
8055 * sched_domain_node_span - get a cpumask for a node's sched_domain
8056 * @node: node whose cpumask we're constructing
8057 * @span: resulting cpumask
8059 * Given a node, construct a good cpumask for its sched_domain to span. It
8060 * should be one that prevents unnecessary balancing, but also spreads tasks
8063 static void sched_domain_node_span(int node
, struct cpumask
*span
)
8065 nodemask_t used_nodes
;
8068 cpumask_clear(span
);
8069 nodes_clear(used_nodes
);
8071 cpumask_or(span
, span
, cpumask_of_node(node
));
8072 node_set(node
, used_nodes
);
8074 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
8075 int next_node
= find_next_best_node(node
, &used_nodes
);
8077 cpumask_or(span
, span
, cpumask_of_node(next_node
));
8080 #endif /* CONFIG_NUMA */
8082 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
8085 * The cpus mask in sched_group and sched_domain hangs off the end.
8087 * ( See the the comments in include/linux/sched.h:struct sched_group
8088 * and struct sched_domain. )
8090 struct static_sched_group
{
8091 struct sched_group sg
;
8092 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
8095 struct static_sched_domain
{
8096 struct sched_domain sd
;
8097 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
8101 * SMT sched-domains:
8103 #ifdef CONFIG_SCHED_SMT
8104 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
8105 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_cpus
);
8108 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
8109 struct sched_group
**sg
, struct cpumask
*unused
)
8112 *sg
= &per_cpu(sched_group_cpus
, cpu
).sg
;
8115 #endif /* CONFIG_SCHED_SMT */
8118 * multi-core sched-domains:
8120 #ifdef CONFIG_SCHED_MC
8121 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
8122 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
8123 #endif /* CONFIG_SCHED_MC */
8125 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8127 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
8128 struct sched_group
**sg
, struct cpumask
*mask
)
8132 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
8133 group
= cpumask_first(mask
);
8135 *sg
= &per_cpu(sched_group_core
, group
).sg
;
8138 #elif defined(CONFIG_SCHED_MC)
8140 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
8141 struct sched_group
**sg
, struct cpumask
*unused
)
8144 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
8149 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
8150 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
8153 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
8154 struct sched_group
**sg
, struct cpumask
*mask
)
8157 #ifdef CONFIG_SCHED_MC
8158 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
8159 group
= cpumask_first(mask
);
8160 #elif defined(CONFIG_SCHED_SMT)
8161 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
8162 group
= cpumask_first(mask
);
8167 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
8173 * The init_sched_build_groups can't handle what we want to do with node
8174 * groups, so roll our own. Now each node has its own list of groups which
8175 * gets dynamically allocated.
8177 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
8178 static struct sched_group
***sched_group_nodes_bycpu
;
8180 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
8181 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
8183 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
8184 struct sched_group
**sg
,
8185 struct cpumask
*nodemask
)
8189 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
8190 group
= cpumask_first(nodemask
);
8193 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
8197 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
8199 struct sched_group
*sg
= group_head
;
8205 for_each_cpu(j
, sched_group_cpus(sg
)) {
8206 struct sched_domain
*sd
;
8208 sd
= &per_cpu(phys_domains
, j
).sd
;
8209 if (j
!= group_first_cpu(sd
->groups
)) {
8211 * Only add "power" once for each
8217 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
8220 } while (sg
!= group_head
);
8222 #endif /* CONFIG_NUMA */
8225 /* Free memory allocated for various sched_group structures */
8226 static void free_sched_groups(const struct cpumask
*cpu_map
,
8227 struct cpumask
*nodemask
)
8231 for_each_cpu(cpu
, cpu_map
) {
8232 struct sched_group
**sched_group_nodes
8233 = sched_group_nodes_bycpu
[cpu
];
8235 if (!sched_group_nodes
)
8238 for (i
= 0; i
< nr_node_ids
; i
++) {
8239 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
8241 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8242 if (cpumask_empty(nodemask
))
8252 if (oldsg
!= sched_group_nodes
[i
])
8255 kfree(sched_group_nodes
);
8256 sched_group_nodes_bycpu
[cpu
] = NULL
;
8259 #else /* !CONFIG_NUMA */
8260 static void free_sched_groups(const struct cpumask
*cpu_map
,
8261 struct cpumask
*nodemask
)
8264 #endif /* CONFIG_NUMA */
8267 * Initialize sched groups cpu_power.
8269 * cpu_power indicates the capacity of sched group, which is used while
8270 * distributing the load between different sched groups in a sched domain.
8271 * Typically cpu_power for all the groups in a sched domain will be same unless
8272 * there are asymmetries in the topology. If there are asymmetries, group
8273 * having more cpu_power will pickup more load compared to the group having
8276 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
8277 * the maximum number of tasks a group can handle in the presence of other idle
8278 * or lightly loaded groups in the same sched domain.
8280 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
8282 struct sched_domain
*child
;
8283 struct sched_group
*group
;
8285 WARN_ON(!sd
|| !sd
->groups
);
8287 if (cpu
!= group_first_cpu(sd
->groups
))
8292 sd
->groups
->__cpu_power
= 0;
8295 * For perf policy, if the groups in child domain share resources
8296 * (for example cores sharing some portions of the cache hierarchy
8297 * or SMT), then set this domain groups cpu_power such that each group
8298 * can handle only one task, when there are other idle groups in the
8299 * same sched domain.
8301 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
8303 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
8304 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
8309 * add cpu_power of each child group to this groups cpu_power
8311 group
= child
->groups
;
8313 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
8314 group
= group
->next
;
8315 } while (group
!= child
->groups
);
8319 * Initializers for schedule domains
8320 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8323 #ifdef CONFIG_SCHED_DEBUG
8324 # define SD_INIT_NAME(sd, type) sd->name = #type
8326 # define SD_INIT_NAME(sd, type) do { } while (0)
8329 #define SD_INIT(sd, type) sd_init_##type(sd)
8331 #define SD_INIT_FUNC(type) \
8332 static noinline void sd_init_##type(struct sched_domain *sd) \
8334 memset(sd, 0, sizeof(*sd)); \
8335 *sd = SD_##type##_INIT; \
8336 sd->level = SD_LV_##type; \
8337 SD_INIT_NAME(sd, type); \
8342 SD_INIT_FUNC(ALLNODES
)
8345 #ifdef CONFIG_SCHED_SMT
8346 SD_INIT_FUNC(SIBLING
)
8348 #ifdef CONFIG_SCHED_MC
8352 static int default_relax_domain_level
= -1;
8354 static int __init
setup_relax_domain_level(char *str
)
8358 val
= simple_strtoul(str
, NULL
, 0);
8359 if (val
< SD_LV_MAX
)
8360 default_relax_domain_level
= val
;
8364 __setup("relax_domain_level=", setup_relax_domain_level
);
8366 static void set_domain_attribute(struct sched_domain
*sd
,
8367 struct sched_domain_attr
*attr
)
8371 if (!attr
|| attr
->relax_domain_level
< 0) {
8372 if (default_relax_domain_level
< 0)
8375 request
= default_relax_domain_level
;
8377 request
= attr
->relax_domain_level
;
8378 if (request
< sd
->level
) {
8379 /* turn off idle balance on this domain */
8380 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
8382 /* turn on idle balance on this domain */
8383 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
8388 * Build sched domains for a given set of cpus and attach the sched domains
8389 * to the individual cpus
8391 static int __build_sched_domains(const struct cpumask
*cpu_map
,
8392 struct sched_domain_attr
*attr
)
8394 int i
, err
= -ENOMEM
;
8395 struct root_domain
*rd
;
8396 cpumask_var_t nodemask
, this_sibling_map
, this_core_map
, send_covered
,
8399 cpumask_var_t domainspan
, covered
, notcovered
;
8400 struct sched_group
**sched_group_nodes
= NULL
;
8401 int sd_allnodes
= 0;
8403 if (!alloc_cpumask_var(&domainspan
, GFP_KERNEL
))
8405 if (!alloc_cpumask_var(&covered
, GFP_KERNEL
))
8406 goto free_domainspan
;
8407 if (!alloc_cpumask_var(¬covered
, GFP_KERNEL
))
8411 if (!alloc_cpumask_var(&nodemask
, GFP_KERNEL
))
8412 goto free_notcovered
;
8413 if (!alloc_cpumask_var(&this_sibling_map
, GFP_KERNEL
))
8415 if (!alloc_cpumask_var(&this_core_map
, GFP_KERNEL
))
8416 goto free_this_sibling_map
;
8417 if (!alloc_cpumask_var(&send_covered
, GFP_KERNEL
))
8418 goto free_this_core_map
;
8419 if (!alloc_cpumask_var(&tmpmask
, GFP_KERNEL
))
8420 goto free_send_covered
;
8424 * Allocate the per-node list of sched groups
8426 sched_group_nodes
= kcalloc(nr_node_ids
, sizeof(struct sched_group
*),
8428 if (!sched_group_nodes
) {
8429 printk(KERN_WARNING
"Can not alloc sched group node list\n");
8434 rd
= alloc_rootdomain();
8436 printk(KERN_WARNING
"Cannot alloc root domain\n");
8437 goto free_sched_groups
;
8441 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = sched_group_nodes
;
8445 * Set up domains for cpus specified by the cpu_map.
8447 for_each_cpu(i
, cpu_map
) {
8448 struct sched_domain
*sd
= NULL
, *p
;
8450 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(i
)), cpu_map
);
8453 if (cpumask_weight(cpu_map
) >
8454 SD_NODES_PER_DOMAIN
*cpumask_weight(nodemask
)) {
8455 sd
= &per_cpu(allnodes_domains
, i
).sd
;
8456 SD_INIT(sd
, ALLNODES
);
8457 set_domain_attribute(sd
, attr
);
8458 cpumask_copy(sched_domain_span(sd
), cpu_map
);
8459 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8465 sd
= &per_cpu(node_domains
, i
).sd
;
8467 set_domain_attribute(sd
, attr
);
8468 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
8472 cpumask_and(sched_domain_span(sd
),
8473 sched_domain_span(sd
), cpu_map
);
8477 sd
= &per_cpu(phys_domains
, i
).sd
;
8479 set_domain_attribute(sd
, attr
);
8480 cpumask_copy(sched_domain_span(sd
), nodemask
);
8484 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8486 #ifdef CONFIG_SCHED_MC
8488 sd
= &per_cpu(core_domains
, i
).sd
;
8490 set_domain_attribute(sd
, attr
);
8491 cpumask_and(sched_domain_span(sd
), cpu_map
,
8492 cpu_coregroup_mask(i
));
8495 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8498 #ifdef CONFIG_SCHED_SMT
8500 sd
= &per_cpu(cpu_domains
, i
).sd
;
8501 SD_INIT(sd
, SIBLING
);
8502 set_domain_attribute(sd
, attr
);
8503 cpumask_and(sched_domain_span(sd
),
8504 topology_thread_cpumask(i
), cpu_map
);
8507 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8511 #ifdef CONFIG_SCHED_SMT
8512 /* Set up CPU (sibling) groups */
8513 for_each_cpu(i
, cpu_map
) {
8514 cpumask_and(this_sibling_map
,
8515 topology_thread_cpumask(i
), cpu_map
);
8516 if (i
!= cpumask_first(this_sibling_map
))
8519 init_sched_build_groups(this_sibling_map
, cpu_map
,
8521 send_covered
, tmpmask
);
8525 #ifdef CONFIG_SCHED_MC
8526 /* Set up multi-core groups */
8527 for_each_cpu(i
, cpu_map
) {
8528 cpumask_and(this_core_map
, cpu_coregroup_mask(i
), cpu_map
);
8529 if (i
!= cpumask_first(this_core_map
))
8532 init_sched_build_groups(this_core_map
, cpu_map
,
8534 send_covered
, tmpmask
);
8538 /* Set up physical groups */
8539 for (i
= 0; i
< nr_node_ids
; i
++) {
8540 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8541 if (cpumask_empty(nodemask
))
8544 init_sched_build_groups(nodemask
, cpu_map
,
8546 send_covered
, tmpmask
);
8550 /* Set up node groups */
8552 init_sched_build_groups(cpu_map
, cpu_map
,
8553 &cpu_to_allnodes_group
,
8554 send_covered
, tmpmask
);
8557 for (i
= 0; i
< nr_node_ids
; i
++) {
8558 /* Set up node groups */
8559 struct sched_group
*sg
, *prev
;
8562 cpumask_clear(covered
);
8563 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8564 if (cpumask_empty(nodemask
)) {
8565 sched_group_nodes
[i
] = NULL
;
8569 sched_domain_node_span(i
, domainspan
);
8570 cpumask_and(domainspan
, domainspan
, cpu_map
);
8572 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
8575 printk(KERN_WARNING
"Can not alloc domain group for "
8579 sched_group_nodes
[i
] = sg
;
8580 for_each_cpu(j
, nodemask
) {
8581 struct sched_domain
*sd
;
8583 sd
= &per_cpu(node_domains
, j
).sd
;
8586 sg
->__cpu_power
= 0;
8587 cpumask_copy(sched_group_cpus(sg
), nodemask
);
8589 cpumask_or(covered
, covered
, nodemask
);
8592 for (j
= 0; j
< nr_node_ids
; j
++) {
8593 int n
= (i
+ j
) % nr_node_ids
;
8595 cpumask_complement(notcovered
, covered
);
8596 cpumask_and(tmpmask
, notcovered
, cpu_map
);
8597 cpumask_and(tmpmask
, tmpmask
, domainspan
);
8598 if (cpumask_empty(tmpmask
))
8601 cpumask_and(tmpmask
, tmpmask
, cpumask_of_node(n
));
8602 if (cpumask_empty(tmpmask
))
8605 sg
= kmalloc_node(sizeof(struct sched_group
) +
8610 "Can not alloc domain group for node %d\n", j
);
8613 sg
->__cpu_power
= 0;
8614 cpumask_copy(sched_group_cpus(sg
), tmpmask
);
8615 sg
->next
= prev
->next
;
8616 cpumask_or(covered
, covered
, tmpmask
);
8623 /* Calculate CPU power for physical packages and nodes */
8624 #ifdef CONFIG_SCHED_SMT
8625 for_each_cpu(i
, cpu_map
) {
8626 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
).sd
;
8628 init_sched_groups_power(i
, sd
);
8631 #ifdef CONFIG_SCHED_MC
8632 for_each_cpu(i
, cpu_map
) {
8633 struct sched_domain
*sd
= &per_cpu(core_domains
, i
).sd
;
8635 init_sched_groups_power(i
, sd
);
8639 for_each_cpu(i
, cpu_map
) {
8640 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
).sd
;
8642 init_sched_groups_power(i
, sd
);
8646 for (i
= 0; i
< nr_node_ids
; i
++)
8647 init_numa_sched_groups_power(sched_group_nodes
[i
]);
8650 struct sched_group
*sg
;
8652 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
8654 init_numa_sched_groups_power(sg
);
8658 /* Attach the domains */
8659 for_each_cpu(i
, cpu_map
) {
8660 struct sched_domain
*sd
;
8661 #ifdef CONFIG_SCHED_SMT
8662 sd
= &per_cpu(cpu_domains
, i
).sd
;
8663 #elif defined(CONFIG_SCHED_MC)
8664 sd
= &per_cpu(core_domains
, i
).sd
;
8666 sd
= &per_cpu(phys_domains
, i
).sd
;
8668 cpu_attach_domain(sd
, rd
, i
);
8674 free_cpumask_var(tmpmask
);
8676 free_cpumask_var(send_covered
);
8678 free_cpumask_var(this_core_map
);
8679 free_this_sibling_map
:
8680 free_cpumask_var(this_sibling_map
);
8682 free_cpumask_var(nodemask
);
8685 free_cpumask_var(notcovered
);
8687 free_cpumask_var(covered
);
8689 free_cpumask_var(domainspan
);
8696 kfree(sched_group_nodes
);
8702 free_sched_groups(cpu_map
, tmpmask
);
8703 free_rootdomain(rd
);
8708 static int build_sched_domains(const struct cpumask
*cpu_map
)
8710 return __build_sched_domains(cpu_map
, NULL
);
8713 static struct cpumask
*doms_cur
; /* current sched domains */
8714 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
8715 static struct sched_domain_attr
*dattr_cur
;
8716 /* attribues of custom domains in 'doms_cur' */
8719 * Special case: If a kmalloc of a doms_cur partition (array of
8720 * cpumask) fails, then fallback to a single sched domain,
8721 * as determined by the single cpumask fallback_doms.
8723 static cpumask_var_t fallback_doms
;
8726 * arch_update_cpu_topology lets virtualized architectures update the
8727 * cpu core maps. It is supposed to return 1 if the topology changed
8728 * or 0 if it stayed the same.
8730 int __attribute__((weak
)) arch_update_cpu_topology(void)
8736 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8737 * For now this just excludes isolated cpus, but could be used to
8738 * exclude other special cases in the future.
8740 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
8744 arch_update_cpu_topology();
8746 doms_cur
= kmalloc(cpumask_size(), GFP_KERNEL
);
8748 doms_cur
= fallback_doms
;
8749 cpumask_andnot(doms_cur
, cpu_map
, cpu_isolated_map
);
8751 err
= build_sched_domains(doms_cur
);
8752 register_sched_domain_sysctl();
8757 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
8758 struct cpumask
*tmpmask
)
8760 free_sched_groups(cpu_map
, tmpmask
);
8764 * Detach sched domains from a group of cpus specified in cpu_map
8765 * These cpus will now be attached to the NULL domain
8767 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
8769 /* Save because hotplug lock held. */
8770 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
8773 for_each_cpu(i
, cpu_map
)
8774 cpu_attach_domain(NULL
, &def_root_domain
, i
);
8775 synchronize_sched();
8776 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
8779 /* handle null as "default" */
8780 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
8781 struct sched_domain_attr
*new, int idx_new
)
8783 struct sched_domain_attr tmp
;
8790 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
8791 new ? (new + idx_new
) : &tmp
,
8792 sizeof(struct sched_domain_attr
));
8796 * Partition sched domains as specified by the 'ndoms_new'
8797 * cpumasks in the array doms_new[] of cpumasks. This compares
8798 * doms_new[] to the current sched domain partitioning, doms_cur[].
8799 * It destroys each deleted domain and builds each new domain.
8801 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8802 * The masks don't intersect (don't overlap.) We should setup one
8803 * sched domain for each mask. CPUs not in any of the cpumasks will
8804 * not be load balanced. If the same cpumask appears both in the
8805 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8808 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8809 * ownership of it and will kfree it when done with it. If the caller
8810 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8811 * ndoms_new == 1, and partition_sched_domains() will fallback to
8812 * the single partition 'fallback_doms', it also forces the domains
8815 * If doms_new == NULL it will be replaced with cpu_online_mask.
8816 * ndoms_new == 0 is a special case for destroying existing domains,
8817 * and it will not create the default domain.
8819 * Call with hotplug lock held
8821 /* FIXME: Change to struct cpumask *doms_new[] */
8822 void partition_sched_domains(int ndoms_new
, struct cpumask
*doms_new
,
8823 struct sched_domain_attr
*dattr_new
)
8828 mutex_lock(&sched_domains_mutex
);
8830 /* always unregister in case we don't destroy any domains */
8831 unregister_sched_domain_sysctl();
8833 /* Let architecture update cpu core mappings. */
8834 new_topology
= arch_update_cpu_topology();
8836 n
= doms_new
? ndoms_new
: 0;
8838 /* Destroy deleted domains */
8839 for (i
= 0; i
< ndoms_cur
; i
++) {
8840 for (j
= 0; j
< n
&& !new_topology
; j
++) {
8841 if (cpumask_equal(&doms_cur
[i
], &doms_new
[j
])
8842 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
8845 /* no match - a current sched domain not in new doms_new[] */
8846 detach_destroy_domains(doms_cur
+ i
);
8851 if (doms_new
== NULL
) {
8853 doms_new
= fallback_doms
;
8854 cpumask_andnot(&doms_new
[0], cpu_online_mask
, cpu_isolated_map
);
8855 WARN_ON_ONCE(dattr_new
);
8858 /* Build new domains */
8859 for (i
= 0; i
< ndoms_new
; i
++) {
8860 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
8861 if (cpumask_equal(&doms_new
[i
], &doms_cur
[j
])
8862 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
8865 /* no match - add a new doms_new */
8866 __build_sched_domains(doms_new
+ i
,
8867 dattr_new
? dattr_new
+ i
: NULL
);
8872 /* Remember the new sched domains */
8873 if (doms_cur
!= fallback_doms
)
8875 kfree(dattr_cur
); /* kfree(NULL) is safe */
8876 doms_cur
= doms_new
;
8877 dattr_cur
= dattr_new
;
8878 ndoms_cur
= ndoms_new
;
8880 register_sched_domain_sysctl();
8882 mutex_unlock(&sched_domains_mutex
);
8885 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8886 static void arch_reinit_sched_domains(void)
8890 /* Destroy domains first to force the rebuild */
8891 partition_sched_domains(0, NULL
, NULL
);
8893 rebuild_sched_domains();
8897 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
8899 unsigned int level
= 0;
8901 if (sscanf(buf
, "%u", &level
) != 1)
8905 * level is always be positive so don't check for
8906 * level < POWERSAVINGS_BALANCE_NONE which is 0
8907 * What happens on 0 or 1 byte write,
8908 * need to check for count as well?
8911 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
8915 sched_smt_power_savings
= level
;
8917 sched_mc_power_savings
= level
;
8919 arch_reinit_sched_domains();
8924 #ifdef CONFIG_SCHED_MC
8925 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
8928 return sprintf(page
, "%u\n", sched_mc_power_savings
);
8930 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
8931 const char *buf
, size_t count
)
8933 return sched_power_savings_store(buf
, count
, 0);
8935 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
8936 sched_mc_power_savings_show
,
8937 sched_mc_power_savings_store
);
8940 #ifdef CONFIG_SCHED_SMT
8941 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
8944 return sprintf(page
, "%u\n", sched_smt_power_savings
);
8946 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
8947 const char *buf
, size_t count
)
8949 return sched_power_savings_store(buf
, count
, 1);
8951 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
8952 sched_smt_power_savings_show
,
8953 sched_smt_power_savings_store
);
8956 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
8960 #ifdef CONFIG_SCHED_SMT
8962 err
= sysfs_create_file(&cls
->kset
.kobj
,
8963 &attr_sched_smt_power_savings
.attr
);
8965 #ifdef CONFIG_SCHED_MC
8966 if (!err
&& mc_capable())
8967 err
= sysfs_create_file(&cls
->kset
.kobj
,
8968 &attr_sched_mc_power_savings
.attr
);
8972 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8974 #ifndef CONFIG_CPUSETS
8976 * Add online and remove offline CPUs from the scheduler domains.
8977 * When cpusets are enabled they take over this function.
8979 static int update_sched_domains(struct notifier_block
*nfb
,
8980 unsigned long action
, void *hcpu
)
8984 case CPU_ONLINE_FROZEN
:
8986 case CPU_DEAD_FROZEN
:
8987 partition_sched_domains(1, NULL
, NULL
);
8996 static int update_runtime(struct notifier_block
*nfb
,
8997 unsigned long action
, void *hcpu
)
8999 int cpu
= (int)(long)hcpu
;
9002 case CPU_DOWN_PREPARE
:
9003 case CPU_DOWN_PREPARE_FROZEN
:
9004 disable_runtime(cpu_rq(cpu
));
9007 case CPU_DOWN_FAILED
:
9008 case CPU_DOWN_FAILED_FROZEN
:
9010 case CPU_ONLINE_FROZEN
:
9011 enable_runtime(cpu_rq(cpu
));
9019 void __init
sched_init_smp(void)
9021 cpumask_var_t non_isolated_cpus
;
9023 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
9025 #if defined(CONFIG_NUMA)
9026 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
9028 BUG_ON(sched_group_nodes_bycpu
== NULL
);
9031 mutex_lock(&sched_domains_mutex
);
9032 arch_init_sched_domains(cpu_online_mask
);
9033 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
9034 if (cpumask_empty(non_isolated_cpus
))
9035 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
9036 mutex_unlock(&sched_domains_mutex
);
9039 #ifndef CONFIG_CPUSETS
9040 /* XXX: Theoretical race here - CPU may be hotplugged now */
9041 hotcpu_notifier(update_sched_domains
, 0);
9044 /* RT runtime code needs to handle some hotplug events */
9045 hotcpu_notifier(update_runtime
, 0);
9049 /* Move init over to a non-isolated CPU */
9050 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
9052 sched_init_granularity();
9053 free_cpumask_var(non_isolated_cpus
);
9055 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
9056 init_sched_rt_class();
9059 void __init
sched_init_smp(void)
9061 sched_init_granularity();
9063 #endif /* CONFIG_SMP */
9065 int in_sched_functions(unsigned long addr
)
9067 return in_lock_functions(addr
) ||
9068 (addr
>= (unsigned long)__sched_text_start
9069 && addr
< (unsigned long)__sched_text_end
);
9072 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
9074 cfs_rq
->tasks_timeline
= RB_ROOT
;
9075 INIT_LIST_HEAD(&cfs_rq
->tasks
);
9076 #ifdef CONFIG_FAIR_GROUP_SCHED
9079 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
9082 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
9084 struct rt_prio_array
*array
;
9087 array
= &rt_rq
->active
;
9088 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
9089 INIT_LIST_HEAD(array
->queue
+ i
);
9090 __clear_bit(i
, array
->bitmap
);
9092 /* delimiter for bitsearch: */
9093 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
9095 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9096 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
9098 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
9102 rt_rq
->rt_nr_migratory
= 0;
9103 rt_rq
->overloaded
= 0;
9104 plist_head_init(&rq
->rt
.pushable_tasks
, &rq
->lock
);
9108 rt_rq
->rt_throttled
= 0;
9109 rt_rq
->rt_runtime
= 0;
9110 spin_lock_init(&rt_rq
->rt_runtime_lock
);
9112 #ifdef CONFIG_RT_GROUP_SCHED
9113 rt_rq
->rt_nr_boosted
= 0;
9118 #ifdef CONFIG_FAIR_GROUP_SCHED
9119 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
9120 struct sched_entity
*se
, int cpu
, int add
,
9121 struct sched_entity
*parent
)
9123 struct rq
*rq
= cpu_rq(cpu
);
9124 tg
->cfs_rq
[cpu
] = cfs_rq
;
9125 init_cfs_rq(cfs_rq
, rq
);
9128 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
9131 /* se could be NULL for init_task_group */
9136 se
->cfs_rq
= &rq
->cfs
;
9138 se
->cfs_rq
= parent
->my_q
;
9141 se
->load
.weight
= tg
->shares
;
9142 se
->load
.inv_weight
= 0;
9143 se
->parent
= parent
;
9147 #ifdef CONFIG_RT_GROUP_SCHED
9148 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
9149 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
9150 struct sched_rt_entity
*parent
)
9152 struct rq
*rq
= cpu_rq(cpu
);
9154 tg
->rt_rq
[cpu
] = rt_rq
;
9155 init_rt_rq(rt_rq
, rq
);
9157 rt_rq
->rt_se
= rt_se
;
9158 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
9160 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
9162 tg
->rt_se
[cpu
] = rt_se
;
9167 rt_se
->rt_rq
= &rq
->rt
;
9169 rt_se
->rt_rq
= parent
->my_q
;
9171 rt_se
->my_q
= rt_rq
;
9172 rt_se
->parent
= parent
;
9173 INIT_LIST_HEAD(&rt_se
->run_list
);
9177 void __init
sched_init(void)
9180 unsigned long alloc_size
= 0, ptr
;
9182 #ifdef CONFIG_FAIR_GROUP_SCHED
9183 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
9185 #ifdef CONFIG_RT_GROUP_SCHED
9186 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
9188 #ifdef CONFIG_USER_SCHED
9191 #ifdef CONFIG_CPUMASK_OFFSTACK
9192 alloc_size
+= num_possible_cpus() * cpumask_size();
9195 * As sched_init() is called before page_alloc is setup,
9196 * we use alloc_bootmem().
9199 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
9201 #ifdef CONFIG_FAIR_GROUP_SCHED
9202 init_task_group
.se
= (struct sched_entity
**)ptr
;
9203 ptr
+= nr_cpu_ids
* sizeof(void **);
9205 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
9206 ptr
+= nr_cpu_ids
* sizeof(void **);
9208 #ifdef CONFIG_USER_SCHED
9209 root_task_group
.se
= (struct sched_entity
**)ptr
;
9210 ptr
+= nr_cpu_ids
* sizeof(void **);
9212 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
9213 ptr
+= nr_cpu_ids
* sizeof(void **);
9214 #endif /* CONFIG_USER_SCHED */
9215 #endif /* CONFIG_FAIR_GROUP_SCHED */
9216 #ifdef CONFIG_RT_GROUP_SCHED
9217 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
9218 ptr
+= nr_cpu_ids
* sizeof(void **);
9220 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
9221 ptr
+= nr_cpu_ids
* sizeof(void **);
9223 #ifdef CONFIG_USER_SCHED
9224 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
9225 ptr
+= nr_cpu_ids
* sizeof(void **);
9227 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
9228 ptr
+= nr_cpu_ids
* sizeof(void **);
9229 #endif /* CONFIG_USER_SCHED */
9230 #endif /* CONFIG_RT_GROUP_SCHED */
9231 #ifdef CONFIG_CPUMASK_OFFSTACK
9232 for_each_possible_cpu(i
) {
9233 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
9234 ptr
+= cpumask_size();
9236 #endif /* CONFIG_CPUMASK_OFFSTACK */
9240 init_defrootdomain();
9243 init_rt_bandwidth(&def_rt_bandwidth
,
9244 global_rt_period(), global_rt_runtime());
9246 #ifdef CONFIG_RT_GROUP_SCHED
9247 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
9248 global_rt_period(), global_rt_runtime());
9249 #ifdef CONFIG_USER_SCHED
9250 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
9251 global_rt_period(), RUNTIME_INF
);
9252 #endif /* CONFIG_USER_SCHED */
9253 #endif /* CONFIG_RT_GROUP_SCHED */
9255 #ifdef CONFIG_GROUP_SCHED
9256 list_add(&init_task_group
.list
, &task_groups
);
9257 INIT_LIST_HEAD(&init_task_group
.children
);
9259 #ifdef CONFIG_USER_SCHED
9260 INIT_LIST_HEAD(&root_task_group
.children
);
9261 init_task_group
.parent
= &root_task_group
;
9262 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
9263 #endif /* CONFIG_USER_SCHED */
9264 #endif /* CONFIG_GROUP_SCHED */
9266 for_each_possible_cpu(i
) {
9270 spin_lock_init(&rq
->lock
);
9272 rq
->calc_load_active
= 0;
9273 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
9274 init_cfs_rq(&rq
->cfs
, rq
);
9275 init_rt_rq(&rq
->rt
, rq
);
9276 #ifdef CONFIG_FAIR_GROUP_SCHED
9277 init_task_group
.shares
= init_task_group_load
;
9278 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
9279 #ifdef CONFIG_CGROUP_SCHED
9281 * How much cpu bandwidth does init_task_group get?
9283 * In case of task-groups formed thr' the cgroup filesystem, it
9284 * gets 100% of the cpu resources in the system. This overall
9285 * system cpu resource is divided among the tasks of
9286 * init_task_group and its child task-groups in a fair manner,
9287 * based on each entity's (task or task-group's) weight
9288 * (se->load.weight).
9290 * In other words, if init_task_group has 10 tasks of weight
9291 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9292 * then A0's share of the cpu resource is:
9294 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9296 * We achieve this by letting init_task_group's tasks sit
9297 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9299 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
9300 #elif defined CONFIG_USER_SCHED
9301 root_task_group
.shares
= NICE_0_LOAD
;
9302 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
9304 * In case of task-groups formed thr' the user id of tasks,
9305 * init_task_group represents tasks belonging to root user.
9306 * Hence it forms a sibling of all subsequent groups formed.
9307 * In this case, init_task_group gets only a fraction of overall
9308 * system cpu resource, based on the weight assigned to root
9309 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9310 * by letting tasks of init_task_group sit in a separate cfs_rq
9311 * (init_cfs_rq) and having one entity represent this group of
9312 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9314 init_tg_cfs_entry(&init_task_group
,
9315 &per_cpu(init_cfs_rq
, i
),
9316 &per_cpu(init_sched_entity
, i
), i
, 1,
9317 root_task_group
.se
[i
]);
9320 #endif /* CONFIG_FAIR_GROUP_SCHED */
9322 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
9323 #ifdef CONFIG_RT_GROUP_SCHED
9324 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
9325 #ifdef CONFIG_CGROUP_SCHED
9326 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
9327 #elif defined CONFIG_USER_SCHED
9328 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
9329 init_tg_rt_entry(&init_task_group
,
9330 &per_cpu(init_rt_rq
, i
),
9331 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
9332 root_task_group
.rt_se
[i
]);
9336 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
9337 rq
->cpu_load
[j
] = 0;
9341 rq
->active_balance
= 0;
9342 rq
->next_balance
= jiffies
;
9346 rq
->migration_thread
= NULL
;
9347 INIT_LIST_HEAD(&rq
->migration_queue
);
9348 rq_attach_root(rq
, &def_root_domain
);
9351 atomic_set(&rq
->nr_iowait
, 0);
9354 set_load_weight(&init_task
);
9356 #ifdef CONFIG_PREEMPT_NOTIFIERS
9357 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
9361 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
9364 #ifdef CONFIG_RT_MUTEXES
9365 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
9369 * The boot idle thread does lazy MMU switching as well:
9371 atomic_inc(&init_mm
.mm_count
);
9372 enter_lazy_tlb(&init_mm
, current
);
9375 * Make us the idle thread. Technically, schedule() should not be
9376 * called from this thread, however somewhere below it might be,
9377 * but because we are the idle thread, we just pick up running again
9378 * when this runqueue becomes "idle".
9380 init_idle(current
, smp_processor_id());
9382 calc_load_update
= jiffies
+ LOAD_FREQ
;
9385 * During early bootup we pretend to be a normal task:
9387 current
->sched_class
= &fair_sched_class
;
9389 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9390 alloc_cpumask_var(&nohz_cpu_mask
, GFP_NOWAIT
);
9393 alloc_cpumask_var(&nohz
.cpu_mask
, GFP_NOWAIT
);
9394 alloc_cpumask_var(&nohz
.ilb_grp_nohz_mask
, GFP_NOWAIT
);
9396 alloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
9399 perf_counter_init();
9401 scheduler_running
= 1;
9404 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9405 void __might_sleep(char *file
, int line
)
9408 static unsigned long prev_jiffy
; /* ratelimiting */
9410 if ((!in_atomic() && !irqs_disabled()) ||
9411 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
9413 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
9415 prev_jiffy
= jiffies
;
9418 "BUG: sleeping function called from invalid context at %s:%d\n",
9421 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9422 in_atomic(), irqs_disabled(),
9423 current
->pid
, current
->comm
);
9425 debug_show_held_locks(current
);
9426 if (irqs_disabled())
9427 print_irqtrace_events(current
);
9431 EXPORT_SYMBOL(__might_sleep
);
9434 #ifdef CONFIG_MAGIC_SYSRQ
9435 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
9439 update_rq_clock(rq
);
9440 on_rq
= p
->se
.on_rq
;
9442 deactivate_task(rq
, p
, 0);
9443 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
9445 activate_task(rq
, p
, 0);
9446 resched_task(rq
->curr
);
9450 void normalize_rt_tasks(void)
9452 struct task_struct
*g
, *p
;
9453 unsigned long flags
;
9456 read_lock_irqsave(&tasklist_lock
, flags
);
9457 do_each_thread(g
, p
) {
9459 * Only normalize user tasks:
9464 p
->se
.exec_start
= 0;
9465 #ifdef CONFIG_SCHEDSTATS
9466 p
->se
.wait_start
= 0;
9467 p
->se
.sleep_start
= 0;
9468 p
->se
.block_start
= 0;
9473 * Renice negative nice level userspace
9476 if (TASK_NICE(p
) < 0 && p
->mm
)
9477 set_user_nice(p
, 0);
9481 spin_lock(&p
->pi_lock
);
9482 rq
= __task_rq_lock(p
);
9484 normalize_task(rq
, p
);
9486 __task_rq_unlock(rq
);
9487 spin_unlock(&p
->pi_lock
);
9488 } while_each_thread(g
, p
);
9490 read_unlock_irqrestore(&tasklist_lock
, flags
);
9493 #endif /* CONFIG_MAGIC_SYSRQ */
9497 * These functions are only useful for the IA64 MCA handling.
9499 * They can only be called when the whole system has been
9500 * stopped - every CPU needs to be quiescent, and no scheduling
9501 * activity can take place. Using them for anything else would
9502 * be a serious bug, and as a result, they aren't even visible
9503 * under any other configuration.
9507 * curr_task - return the current task for a given cpu.
9508 * @cpu: the processor in question.
9510 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9512 struct task_struct
*curr_task(int cpu
)
9514 return cpu_curr(cpu
);
9518 * set_curr_task - set the current task for a given cpu.
9519 * @cpu: the processor in question.
9520 * @p: the task pointer to set.
9522 * Description: This function must only be used when non-maskable interrupts
9523 * are serviced on a separate stack. It allows the architecture to switch the
9524 * notion of the current task on a cpu in a non-blocking manner. This function
9525 * must be called with all CPU's synchronized, and interrupts disabled, the
9526 * and caller must save the original value of the current task (see
9527 * curr_task() above) and restore that value before reenabling interrupts and
9528 * re-starting the system.
9530 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9532 void set_curr_task(int cpu
, struct task_struct
*p
)
9539 #ifdef CONFIG_FAIR_GROUP_SCHED
9540 static void free_fair_sched_group(struct task_group
*tg
)
9544 for_each_possible_cpu(i
) {
9546 kfree(tg
->cfs_rq
[i
]);
9556 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9558 struct cfs_rq
*cfs_rq
;
9559 struct sched_entity
*se
;
9563 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9566 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
9570 tg
->shares
= NICE_0_LOAD
;
9572 for_each_possible_cpu(i
) {
9575 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
9576 GFP_KERNEL
, cpu_to_node(i
));
9580 se
= kzalloc_node(sizeof(struct sched_entity
),
9581 GFP_KERNEL
, cpu_to_node(i
));
9585 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
9594 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9596 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
9597 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
9600 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9602 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
9604 #else /* !CONFG_FAIR_GROUP_SCHED */
9605 static inline void free_fair_sched_group(struct task_group
*tg
)
9610 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9615 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9619 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9622 #endif /* CONFIG_FAIR_GROUP_SCHED */
9624 #ifdef CONFIG_RT_GROUP_SCHED
9625 static void free_rt_sched_group(struct task_group
*tg
)
9629 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
9631 for_each_possible_cpu(i
) {
9633 kfree(tg
->rt_rq
[i
]);
9635 kfree(tg
->rt_se
[i
]);
9643 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9645 struct rt_rq
*rt_rq
;
9646 struct sched_rt_entity
*rt_se
;
9650 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9653 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
9657 init_rt_bandwidth(&tg
->rt_bandwidth
,
9658 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
9660 for_each_possible_cpu(i
) {
9663 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
9664 GFP_KERNEL
, cpu_to_node(i
));
9668 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
9669 GFP_KERNEL
, cpu_to_node(i
));
9673 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
9682 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9684 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
9685 &cpu_rq(cpu
)->leaf_rt_rq_list
);
9688 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9690 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
9692 #else /* !CONFIG_RT_GROUP_SCHED */
9693 static inline void free_rt_sched_group(struct task_group
*tg
)
9698 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9703 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9707 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9710 #endif /* CONFIG_RT_GROUP_SCHED */
9712 #ifdef CONFIG_GROUP_SCHED
9713 static void free_sched_group(struct task_group
*tg
)
9715 free_fair_sched_group(tg
);
9716 free_rt_sched_group(tg
);
9720 /* allocate runqueue etc for a new task group */
9721 struct task_group
*sched_create_group(struct task_group
*parent
)
9723 struct task_group
*tg
;
9724 unsigned long flags
;
9727 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
9729 return ERR_PTR(-ENOMEM
);
9731 if (!alloc_fair_sched_group(tg
, parent
))
9734 if (!alloc_rt_sched_group(tg
, parent
))
9737 spin_lock_irqsave(&task_group_lock
, flags
);
9738 for_each_possible_cpu(i
) {
9739 register_fair_sched_group(tg
, i
);
9740 register_rt_sched_group(tg
, i
);
9742 list_add_rcu(&tg
->list
, &task_groups
);
9744 WARN_ON(!parent
); /* root should already exist */
9746 tg
->parent
= parent
;
9747 INIT_LIST_HEAD(&tg
->children
);
9748 list_add_rcu(&tg
->siblings
, &parent
->children
);
9749 spin_unlock_irqrestore(&task_group_lock
, flags
);
9754 free_sched_group(tg
);
9755 return ERR_PTR(-ENOMEM
);
9758 /* rcu callback to free various structures associated with a task group */
9759 static void free_sched_group_rcu(struct rcu_head
*rhp
)
9761 /* now it should be safe to free those cfs_rqs */
9762 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
9765 /* Destroy runqueue etc associated with a task group */
9766 void sched_destroy_group(struct task_group
*tg
)
9768 unsigned long flags
;
9771 spin_lock_irqsave(&task_group_lock
, flags
);
9772 for_each_possible_cpu(i
) {
9773 unregister_fair_sched_group(tg
, i
);
9774 unregister_rt_sched_group(tg
, i
);
9776 list_del_rcu(&tg
->list
);
9777 list_del_rcu(&tg
->siblings
);
9778 spin_unlock_irqrestore(&task_group_lock
, flags
);
9780 /* wait for possible concurrent references to cfs_rqs complete */
9781 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
9784 /* change task's runqueue when it moves between groups.
9785 * The caller of this function should have put the task in its new group
9786 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9787 * reflect its new group.
9789 void sched_move_task(struct task_struct
*tsk
)
9792 unsigned long flags
;
9795 rq
= task_rq_lock(tsk
, &flags
);
9797 update_rq_clock(rq
);
9799 running
= task_current(rq
, tsk
);
9800 on_rq
= tsk
->se
.on_rq
;
9803 dequeue_task(rq
, tsk
, 0);
9804 if (unlikely(running
))
9805 tsk
->sched_class
->put_prev_task(rq
, tsk
);
9807 set_task_rq(tsk
, task_cpu(tsk
));
9809 #ifdef CONFIG_FAIR_GROUP_SCHED
9810 if (tsk
->sched_class
->moved_group
)
9811 tsk
->sched_class
->moved_group(tsk
);
9814 if (unlikely(running
))
9815 tsk
->sched_class
->set_curr_task(rq
);
9817 enqueue_task(rq
, tsk
, 0);
9819 task_rq_unlock(rq
, &flags
);
9821 #endif /* CONFIG_GROUP_SCHED */
9823 #ifdef CONFIG_FAIR_GROUP_SCHED
9824 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
9826 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
9831 dequeue_entity(cfs_rq
, se
, 0);
9833 se
->load
.weight
= shares
;
9834 se
->load
.inv_weight
= 0;
9837 enqueue_entity(cfs_rq
, se
, 0);
9840 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
9842 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
9843 struct rq
*rq
= cfs_rq
->rq
;
9844 unsigned long flags
;
9846 spin_lock_irqsave(&rq
->lock
, flags
);
9847 __set_se_shares(se
, shares
);
9848 spin_unlock_irqrestore(&rq
->lock
, flags
);
9851 static DEFINE_MUTEX(shares_mutex
);
9853 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
9856 unsigned long flags
;
9859 * We can't change the weight of the root cgroup.
9864 if (shares
< MIN_SHARES
)
9865 shares
= MIN_SHARES
;
9866 else if (shares
> MAX_SHARES
)
9867 shares
= MAX_SHARES
;
9869 mutex_lock(&shares_mutex
);
9870 if (tg
->shares
== shares
)
9873 spin_lock_irqsave(&task_group_lock
, flags
);
9874 for_each_possible_cpu(i
)
9875 unregister_fair_sched_group(tg
, i
);
9876 list_del_rcu(&tg
->siblings
);
9877 spin_unlock_irqrestore(&task_group_lock
, flags
);
9879 /* wait for any ongoing reference to this group to finish */
9880 synchronize_sched();
9883 * Now we are free to modify the group's share on each cpu
9884 * w/o tripping rebalance_share or load_balance_fair.
9886 tg
->shares
= shares
;
9887 for_each_possible_cpu(i
) {
9891 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
9892 set_se_shares(tg
->se
[i
], shares
);
9896 * Enable load balance activity on this group, by inserting it back on
9897 * each cpu's rq->leaf_cfs_rq_list.
9899 spin_lock_irqsave(&task_group_lock
, flags
);
9900 for_each_possible_cpu(i
)
9901 register_fair_sched_group(tg
, i
);
9902 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
9903 spin_unlock_irqrestore(&task_group_lock
, flags
);
9905 mutex_unlock(&shares_mutex
);
9909 unsigned long sched_group_shares(struct task_group
*tg
)
9915 #ifdef CONFIG_RT_GROUP_SCHED
9917 * Ensure that the real time constraints are schedulable.
9919 static DEFINE_MUTEX(rt_constraints_mutex
);
9921 static unsigned long to_ratio(u64 period
, u64 runtime
)
9923 if (runtime
== RUNTIME_INF
)
9926 return div64_u64(runtime
<< 20, period
);
9929 /* Must be called with tasklist_lock held */
9930 static inline int tg_has_rt_tasks(struct task_group
*tg
)
9932 struct task_struct
*g
, *p
;
9934 do_each_thread(g
, p
) {
9935 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
9937 } while_each_thread(g
, p
);
9942 struct rt_schedulable_data
{
9943 struct task_group
*tg
;
9948 static int tg_schedulable(struct task_group
*tg
, void *data
)
9950 struct rt_schedulable_data
*d
= data
;
9951 struct task_group
*child
;
9952 unsigned long total
, sum
= 0;
9953 u64 period
, runtime
;
9955 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9956 runtime
= tg
->rt_bandwidth
.rt_runtime
;
9959 period
= d
->rt_period
;
9960 runtime
= d
->rt_runtime
;
9963 #ifdef CONFIG_USER_SCHED
9964 if (tg
== &root_task_group
) {
9965 period
= global_rt_period();
9966 runtime
= global_rt_runtime();
9971 * Cannot have more runtime than the period.
9973 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
9977 * Ensure we don't starve existing RT tasks.
9979 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
9982 total
= to_ratio(period
, runtime
);
9985 * Nobody can have more than the global setting allows.
9987 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
9991 * The sum of our children's runtime should not exceed our own.
9993 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
9994 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
9995 runtime
= child
->rt_bandwidth
.rt_runtime
;
9997 if (child
== d
->tg
) {
9998 period
= d
->rt_period
;
9999 runtime
= d
->rt_runtime
;
10002 sum
+= to_ratio(period
, runtime
);
10011 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
10013 struct rt_schedulable_data data
= {
10015 .rt_period
= period
,
10016 .rt_runtime
= runtime
,
10019 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
10022 static int tg_set_bandwidth(struct task_group
*tg
,
10023 u64 rt_period
, u64 rt_runtime
)
10027 mutex_lock(&rt_constraints_mutex
);
10028 read_lock(&tasklist_lock
);
10029 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
10033 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
10034 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
10035 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
10037 for_each_possible_cpu(i
) {
10038 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
10040 spin_lock(&rt_rq
->rt_runtime_lock
);
10041 rt_rq
->rt_runtime
= rt_runtime
;
10042 spin_unlock(&rt_rq
->rt_runtime_lock
);
10044 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
10046 read_unlock(&tasklist_lock
);
10047 mutex_unlock(&rt_constraints_mutex
);
10052 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
10054 u64 rt_runtime
, rt_period
;
10056 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10057 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
10058 if (rt_runtime_us
< 0)
10059 rt_runtime
= RUNTIME_INF
;
10061 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
10064 long sched_group_rt_runtime(struct task_group
*tg
)
10068 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
10071 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
10072 do_div(rt_runtime_us
, NSEC_PER_USEC
);
10073 return rt_runtime_us
;
10076 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
10078 u64 rt_runtime
, rt_period
;
10080 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
10081 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
10083 if (rt_period
== 0)
10086 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
10089 long sched_group_rt_period(struct task_group
*tg
)
10093 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10094 do_div(rt_period_us
, NSEC_PER_USEC
);
10095 return rt_period_us
;
10098 static int sched_rt_global_constraints(void)
10100 u64 runtime
, period
;
10103 if (sysctl_sched_rt_period
<= 0)
10106 runtime
= global_rt_runtime();
10107 period
= global_rt_period();
10110 * Sanity check on the sysctl variables.
10112 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
10115 mutex_lock(&rt_constraints_mutex
);
10116 read_lock(&tasklist_lock
);
10117 ret
= __rt_schedulable(NULL
, 0, 0);
10118 read_unlock(&tasklist_lock
);
10119 mutex_unlock(&rt_constraints_mutex
);
10124 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
10126 /* Don't accept realtime tasks when there is no way for them to run */
10127 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
10133 #else /* !CONFIG_RT_GROUP_SCHED */
10134 static int sched_rt_global_constraints(void)
10136 unsigned long flags
;
10139 if (sysctl_sched_rt_period
<= 0)
10143 * There's always some RT tasks in the root group
10144 * -- migration, kstopmachine etc..
10146 if (sysctl_sched_rt_runtime
== 0)
10149 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
10150 for_each_possible_cpu(i
) {
10151 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
10153 spin_lock(&rt_rq
->rt_runtime_lock
);
10154 rt_rq
->rt_runtime
= global_rt_runtime();
10155 spin_unlock(&rt_rq
->rt_runtime_lock
);
10157 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
10161 #endif /* CONFIG_RT_GROUP_SCHED */
10163 int sched_rt_handler(struct ctl_table
*table
, int write
,
10164 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
10168 int old_period
, old_runtime
;
10169 static DEFINE_MUTEX(mutex
);
10171 mutex_lock(&mutex
);
10172 old_period
= sysctl_sched_rt_period
;
10173 old_runtime
= sysctl_sched_rt_runtime
;
10175 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
10177 if (!ret
&& write
) {
10178 ret
= sched_rt_global_constraints();
10180 sysctl_sched_rt_period
= old_period
;
10181 sysctl_sched_rt_runtime
= old_runtime
;
10183 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
10184 def_rt_bandwidth
.rt_period
=
10185 ns_to_ktime(global_rt_period());
10188 mutex_unlock(&mutex
);
10193 #ifdef CONFIG_CGROUP_SCHED
10195 /* return corresponding task_group object of a cgroup */
10196 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
10198 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
10199 struct task_group
, css
);
10202 static struct cgroup_subsys_state
*
10203 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10205 struct task_group
*tg
, *parent
;
10207 if (!cgrp
->parent
) {
10208 /* This is early initialization for the top cgroup */
10209 return &init_task_group
.css
;
10212 parent
= cgroup_tg(cgrp
->parent
);
10213 tg
= sched_create_group(parent
);
10215 return ERR_PTR(-ENOMEM
);
10221 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10223 struct task_group
*tg
= cgroup_tg(cgrp
);
10225 sched_destroy_group(tg
);
10229 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
10230 struct task_struct
*tsk
)
10232 #ifdef CONFIG_RT_GROUP_SCHED
10233 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
10236 /* We don't support RT-tasks being in separate groups */
10237 if (tsk
->sched_class
!= &fair_sched_class
)
10245 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
10246 struct cgroup
*old_cont
, struct task_struct
*tsk
)
10248 sched_move_task(tsk
);
10251 #ifdef CONFIG_FAIR_GROUP_SCHED
10252 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
10255 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
10258 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
10260 struct task_group
*tg
= cgroup_tg(cgrp
);
10262 return (u64
) tg
->shares
;
10264 #endif /* CONFIG_FAIR_GROUP_SCHED */
10266 #ifdef CONFIG_RT_GROUP_SCHED
10267 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
10270 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
10273 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10275 return sched_group_rt_runtime(cgroup_tg(cgrp
));
10278 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
10281 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
10284 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
10286 return sched_group_rt_period(cgroup_tg(cgrp
));
10288 #endif /* CONFIG_RT_GROUP_SCHED */
10290 static struct cftype cpu_files
[] = {
10291 #ifdef CONFIG_FAIR_GROUP_SCHED
10294 .read_u64
= cpu_shares_read_u64
,
10295 .write_u64
= cpu_shares_write_u64
,
10298 #ifdef CONFIG_RT_GROUP_SCHED
10300 .name
= "rt_runtime_us",
10301 .read_s64
= cpu_rt_runtime_read
,
10302 .write_s64
= cpu_rt_runtime_write
,
10305 .name
= "rt_period_us",
10306 .read_u64
= cpu_rt_period_read_uint
,
10307 .write_u64
= cpu_rt_period_write_uint
,
10312 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
10314 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
10317 struct cgroup_subsys cpu_cgroup_subsys
= {
10319 .create
= cpu_cgroup_create
,
10320 .destroy
= cpu_cgroup_destroy
,
10321 .can_attach
= cpu_cgroup_can_attach
,
10322 .attach
= cpu_cgroup_attach
,
10323 .populate
= cpu_cgroup_populate
,
10324 .subsys_id
= cpu_cgroup_subsys_id
,
10328 #endif /* CONFIG_CGROUP_SCHED */
10330 #ifdef CONFIG_CGROUP_CPUACCT
10333 * CPU accounting code for task groups.
10335 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10336 * (balbir@in.ibm.com).
10339 /* track cpu usage of a group of tasks and its child groups */
10341 struct cgroup_subsys_state css
;
10342 /* cpuusage holds pointer to a u64-type object on every cpu */
10344 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
10345 struct cpuacct
*parent
;
10348 struct cgroup_subsys cpuacct_subsys
;
10350 /* return cpu accounting group corresponding to this container */
10351 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
10353 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
10354 struct cpuacct
, css
);
10357 /* return cpu accounting group to which this task belongs */
10358 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
10360 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
10361 struct cpuacct
, css
);
10364 /* create a new cpu accounting group */
10365 static struct cgroup_subsys_state
*cpuacct_create(
10366 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10368 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
10374 ca
->cpuusage
= alloc_percpu(u64
);
10378 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10379 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
10380 goto out_free_counters
;
10383 ca
->parent
= cgroup_ca(cgrp
->parent
);
10389 percpu_counter_destroy(&ca
->cpustat
[i
]);
10390 free_percpu(ca
->cpuusage
);
10394 return ERR_PTR(-ENOMEM
);
10397 /* destroy an existing cpu accounting group */
10399 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10401 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10404 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10405 percpu_counter_destroy(&ca
->cpustat
[i
]);
10406 free_percpu(ca
->cpuusage
);
10410 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
10412 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10415 #ifndef CONFIG_64BIT
10417 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10419 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10421 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10429 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
10431 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10433 #ifndef CONFIG_64BIT
10435 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10437 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10439 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10445 /* return total cpu usage (in nanoseconds) of a group */
10446 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10448 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10449 u64 totalcpuusage
= 0;
10452 for_each_present_cpu(i
)
10453 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
10455 return totalcpuusage
;
10458 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
10461 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10470 for_each_present_cpu(i
)
10471 cpuacct_cpuusage_write(ca
, i
, 0);
10477 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
10478 struct seq_file
*m
)
10480 struct cpuacct
*ca
= cgroup_ca(cgroup
);
10484 for_each_present_cpu(i
) {
10485 percpu
= cpuacct_cpuusage_read(ca
, i
);
10486 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
10488 seq_printf(m
, "\n");
10492 static const char *cpuacct_stat_desc
[] = {
10493 [CPUACCT_STAT_USER
] = "user",
10494 [CPUACCT_STAT_SYSTEM
] = "system",
10497 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
10498 struct cgroup_map_cb
*cb
)
10500 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10503 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
10504 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
10505 val
= cputime64_to_clock_t(val
);
10506 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
10511 static struct cftype files
[] = {
10514 .read_u64
= cpuusage_read
,
10515 .write_u64
= cpuusage_write
,
10518 .name
= "usage_percpu",
10519 .read_seq_string
= cpuacct_percpu_seq_read
,
10523 .read_map
= cpuacct_stats_show
,
10527 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10529 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
10533 * charge this task's execution time to its accounting group.
10535 * called with rq->lock held.
10537 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
10539 struct cpuacct
*ca
;
10542 if (unlikely(!cpuacct_subsys
.active
))
10545 cpu
= task_cpu(tsk
);
10551 for (; ca
; ca
= ca
->parent
) {
10552 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10553 *cpuusage
+= cputime
;
10560 * Charge the system/user time to the task's accounting group.
10562 static void cpuacct_update_stats(struct task_struct
*tsk
,
10563 enum cpuacct_stat_index idx
, cputime_t val
)
10565 struct cpuacct
*ca
;
10567 if (unlikely(!cpuacct_subsys
.active
))
10574 percpu_counter_add(&ca
->cpustat
[idx
], val
);
10580 struct cgroup_subsys cpuacct_subsys
= {
10582 .create
= cpuacct_create
,
10583 .destroy
= cpuacct_destroy
,
10584 .populate
= cpuacct_populate
,
10585 .subsys_id
= cpuacct_subsys_id
,
10587 #endif /* CONFIG_CGROUP_CPUACCT */