4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/reciprocal_div.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/bootmem.h>
72 #include <linux/debugfs.h>
73 #include <linux/ctype.h>
74 #include <linux/ftrace.h>
75 #include <trace/sched.h>
78 #include <asm/irq_regs.h>
80 #include "sched_cpupri.h"
83 * Convert user-nice values [ -20 ... 0 ... 19 ]
84 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
87 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
88 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
89 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
92 * 'User priority' is the nice value converted to something we
93 * can work with better when scaling various scheduler parameters,
94 * it's a [ 0 ... 39 ] range.
96 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
97 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
98 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
101 * Helpers for converting nanosecond timing to jiffy resolution
103 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
105 #define NICE_0_LOAD SCHED_LOAD_SCALE
106 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
109 * These are the 'tuning knobs' of the scheduler:
111 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
112 * Timeslices get refilled after they expire.
114 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * single value that denotes runtime == period, ie unlimited time.
119 #define RUNTIME_INF ((u64)~0ULL)
121 DEFINE_TRACE(sched_wait_task);
122 DEFINE_TRACE(sched_wakeup);
123 DEFINE_TRACE(sched_wakeup_new);
124 DEFINE_TRACE(sched_switch);
125 DEFINE_TRACE(sched_migrate_task);
129 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
130 * Since cpu_power is a 'constant', we can use a reciprocal divide.
132 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
134 return reciprocal_divide(load, sg->reciprocal_cpu_power);
138 * Each time a sched group cpu_power is changed,
139 * we must compute its reciprocal value
141 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
143 sg->__cpu_power += val;
144 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
148 static inline int rt_policy(int policy)
150 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
155 static inline int task_has_rt_policy(struct task_struct *p)
157 return rt_policy(p->policy);
161 * This is the priority-queue data structure of the RT scheduling class:
163 struct rt_prio_array {
164 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
165 struct list_head queue[MAX_RT_PRIO];
168 struct rt_bandwidth {
169 /* nests inside the rq lock: */
170 spinlock_t rt_runtime_lock;
173 struct hrtimer rt_period_timer;
176 static struct rt_bandwidth def_rt_bandwidth;
178 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
180 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
182 struct rt_bandwidth *rt_b =
183 container_of(timer, struct rt_bandwidth, rt_period_timer);
189 now = hrtimer_cb_get_time(timer);
190 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
195 idle = do_sched_rt_period_timer(rt_b, overrun);
198 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
202 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
204 rt_b->rt_period = ns_to_ktime(period);
205 rt_b->rt_runtime = runtime;
207 spin_lock_init(&rt_b->rt_runtime_lock);
209 hrtimer_init(&rt_b->rt_period_timer,
210 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
211 rt_b->rt_period_timer.function = sched_rt_period_timer;
212 rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_UNLOCKED;
215 static inline int rt_bandwidth_enabled(void)
217 return sysctl_sched_rt_runtime >= 0;
220 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
224 if (rt_bandwidth_enabled() && rt_b->rt_runtime == RUNTIME_INF)
227 if (hrtimer_active(&rt_b->rt_period_timer))
230 spin_lock(&rt_b->rt_runtime_lock);
232 if (hrtimer_active(&rt_b->rt_period_timer))
235 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
236 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
237 hrtimer_start_expires(&rt_b->rt_period_timer,
240 spin_unlock(&rt_b->rt_runtime_lock);
243 #ifdef CONFIG_RT_GROUP_SCHED
244 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
246 hrtimer_cancel(&rt_b->rt_period_timer);
251 * sched_domains_mutex serializes calls to arch_init_sched_domains,
252 * detach_destroy_domains and partition_sched_domains.
254 static DEFINE_MUTEX(sched_domains_mutex);
256 #ifdef CONFIG_GROUP_SCHED
258 #include <linux/cgroup.h>
262 static LIST_HEAD(task_groups);
264 /* task group related information */
266 #ifdef CONFIG_CGROUP_SCHED
267 struct cgroup_subsys_state css;
270 #ifdef CONFIG_USER_SCHED
274 #ifdef CONFIG_FAIR_GROUP_SCHED
275 /* schedulable entities of this group on each cpu */
276 struct sched_entity **se;
277 /* runqueue "owned" by this group on each cpu */
278 struct cfs_rq **cfs_rq;
279 unsigned long shares;
282 #ifdef CONFIG_RT_GROUP_SCHED
283 struct sched_rt_entity **rt_se;
284 struct rt_rq **rt_rq;
286 struct rt_bandwidth rt_bandwidth;
290 struct list_head list;
292 struct task_group *parent;
293 struct list_head siblings;
294 struct list_head children;
297 #ifdef CONFIG_USER_SCHED
299 /* Helper function to pass uid information to create_sched_user() */
300 void set_tg_uid(struct user_struct *user)
302 user->tg->uid = user->uid;
307 * Every UID task group (including init_task_group aka UID-0) will
308 * be a child to this group.
310 struct task_group root_task_group;
312 #ifdef CONFIG_FAIR_GROUP_SCHED
313 /* Default task group's sched entity on each cpu */
314 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
315 /* Default task group's cfs_rq on each cpu */
316 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
317 #endif /* CONFIG_FAIR_GROUP_SCHED */
319 #ifdef CONFIG_RT_GROUP_SCHED
320 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
321 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
322 #endif /* CONFIG_RT_GROUP_SCHED */
323 #else /* !CONFIG_USER_SCHED */
324 #define root_task_group init_task_group
325 #endif /* CONFIG_USER_SCHED */
327 /* task_group_lock serializes add/remove of task groups and also changes to
328 * a task group's cpu shares.
330 static DEFINE_SPINLOCK(task_group_lock);
332 #ifdef CONFIG_FAIR_GROUP_SCHED
333 #ifdef CONFIG_USER_SCHED
334 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
335 #else /* !CONFIG_USER_SCHED */
336 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
337 #endif /* CONFIG_USER_SCHED */
340 * A weight of 0 or 1 can cause arithmetics problems.
341 * A weight of a cfs_rq is the sum of weights of which entities
342 * are queued on this cfs_rq, so a weight of a entity should not be
343 * too large, so as the shares value of a task group.
344 * (The default weight is 1024 - so there's no practical
345 * limitation from this.)
348 #define MAX_SHARES (1UL << 18)
350 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
353 /* Default task group.
354 * Every task in system belong to this group at bootup.
356 struct task_group init_task_group;
358 /* return group to which a task belongs */
359 static inline struct task_group *task_group(struct task_struct *p)
361 struct task_group *tg;
363 #ifdef CONFIG_USER_SCHED
365 tg = __task_cred(p)->user->tg;
367 #elif defined(CONFIG_CGROUP_SCHED)
368 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
369 struct task_group, css);
371 tg = &init_task_group;
376 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
377 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
379 #ifdef CONFIG_FAIR_GROUP_SCHED
380 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
381 p->se.parent = task_group(p)->se[cpu];
384 #ifdef CONFIG_RT_GROUP_SCHED
385 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
386 p->rt.parent = task_group(p)->rt_se[cpu];
392 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
393 static inline struct task_group *task_group(struct task_struct *p)
398 #endif /* CONFIG_GROUP_SCHED */
400 /* CFS-related fields in a runqueue */
402 struct load_weight load;
403 unsigned long nr_running;
408 struct rb_root tasks_timeline;
409 struct rb_node *rb_leftmost;
411 struct list_head tasks;
412 struct list_head *balance_iterator;
415 * 'curr' points to currently running entity on this cfs_rq.
416 * It is set to NULL otherwise (i.e when none are currently running).
418 struct sched_entity *curr, *next, *last;
420 unsigned int nr_spread_over;
422 #ifdef CONFIG_FAIR_GROUP_SCHED
423 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
426 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
427 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
428 * (like users, containers etc.)
430 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
431 * list is used during load balance.
433 struct list_head leaf_cfs_rq_list;
434 struct task_group *tg; /* group that "owns" this runqueue */
438 * the part of load.weight contributed by tasks
440 unsigned long task_weight;
443 * h_load = weight * f(tg)
445 * Where f(tg) is the recursive weight fraction assigned to
448 unsigned long h_load;
451 * this cpu's part of tg->shares
453 unsigned long shares;
456 * load.weight at the time we set shares
458 unsigned long rq_weight;
463 /* Real-Time classes' related field in a runqueue: */
465 struct rt_prio_array active;
466 unsigned long rt_nr_running;
467 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
468 int highest_prio; /* highest queued rt task prio */
471 unsigned long rt_nr_migratory;
477 /* Nests inside the rq lock: */
478 spinlock_t rt_runtime_lock;
480 #ifdef CONFIG_RT_GROUP_SCHED
481 unsigned long rt_nr_boosted;
484 struct list_head leaf_rt_rq_list;
485 struct task_group *tg;
486 struct sched_rt_entity *rt_se;
493 * We add the notion of a root-domain which will be used to define per-domain
494 * variables. Each exclusive cpuset essentially defines an island domain by
495 * fully partitioning the member cpus from any other cpuset. Whenever a new
496 * exclusive cpuset is created, we also create and attach a new root-domain
506 * The "RT overload" flag: it gets set if a CPU has more than
507 * one runnable RT task.
512 struct cpupri cpupri;
517 * By default the system creates a single root-domain with all cpus as
518 * members (mimicking the global state we have today).
520 static struct root_domain def_root_domain;
525 * This is the main, per-CPU runqueue data structure.
527 * Locking rule: those places that want to lock multiple runqueues
528 * (such as the load balancing or the thread migration code), lock
529 * acquire operations must be ordered by ascending &runqueue.
536 * nr_running and cpu_load should be in the same cacheline because
537 * remote CPUs use both these fields when doing load calculation.
539 unsigned long nr_running;
540 #define CPU_LOAD_IDX_MAX 5
541 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
542 unsigned char idle_at_tick;
544 unsigned long last_tick_seen;
545 unsigned char in_nohz_recently;
547 /* capture load from *all* tasks on this cpu: */
548 struct load_weight load;
549 unsigned long nr_load_updates;
555 #ifdef CONFIG_FAIR_GROUP_SCHED
556 /* list of leaf cfs_rq on this cpu: */
557 struct list_head leaf_cfs_rq_list;
559 #ifdef CONFIG_RT_GROUP_SCHED
560 struct list_head leaf_rt_rq_list;
564 * This is part of a global counter where only the total sum
565 * over all CPUs matters. A task can increase this counter on
566 * one CPU and if it got migrated afterwards it may decrease
567 * it on another CPU. Always updated under the runqueue lock:
569 unsigned long nr_uninterruptible;
571 struct task_struct *curr, *idle;
572 unsigned long next_balance;
573 struct mm_struct *prev_mm;
580 struct root_domain *rd;
581 struct sched_domain *sd;
583 /* For active balancing */
586 /* cpu of this runqueue: */
590 unsigned long avg_load_per_task;
592 struct task_struct *migration_thread;
593 struct list_head migration_queue;
596 #ifdef CONFIG_SCHED_HRTICK
598 int hrtick_csd_pending;
599 struct call_single_data hrtick_csd;
601 struct hrtimer hrtick_timer;
604 #ifdef CONFIG_SCHEDSTATS
606 struct sched_info rq_sched_info;
607 unsigned long long rq_cpu_time;
608 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
610 /* sys_sched_yield() stats */
611 unsigned int yld_exp_empty;
612 unsigned int yld_act_empty;
613 unsigned int yld_both_empty;
614 unsigned int yld_count;
616 /* schedule() stats */
617 unsigned int sched_switch;
618 unsigned int sched_count;
619 unsigned int sched_goidle;
621 /* try_to_wake_up() stats */
622 unsigned int ttwu_count;
623 unsigned int ttwu_local;
626 unsigned int bkl_count;
630 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
632 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
634 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
637 static inline int cpu_of(struct rq *rq)
647 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
648 * See detach_destroy_domains: synchronize_sched for details.
650 * The domain tree of any CPU may only be accessed from within
651 * preempt-disabled sections.
653 #define for_each_domain(cpu, __sd) \
654 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
656 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
657 #define this_rq() (&__get_cpu_var(runqueues))
658 #define task_rq(p) cpu_rq(task_cpu(p))
659 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
661 static inline void update_rq_clock(struct rq *rq)
663 rq->clock = sched_clock_cpu(cpu_of(rq));
667 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
669 #ifdef CONFIG_SCHED_DEBUG
670 # define const_debug __read_mostly
672 # define const_debug static const
678 * Returns true if the current cpu runqueue is locked.
679 * This interface allows printk to be called with the runqueue lock
680 * held and know whether or not it is OK to wake up the klogd.
682 int runqueue_is_locked(void)
685 struct rq *rq = cpu_rq(cpu);
688 ret = spin_is_locked(&rq->lock);
694 * Debugging: various feature bits
697 #define SCHED_FEAT(name, enabled) \
698 __SCHED_FEAT_##name ,
701 #include "sched_features.h"
706 #define SCHED_FEAT(name, enabled) \
707 (1UL << __SCHED_FEAT_##name) * enabled |
709 const_debug unsigned int sysctl_sched_features =
710 #include "sched_features.h"
715 #ifdef CONFIG_SCHED_DEBUG
716 #define SCHED_FEAT(name, enabled) \
719 static __read_mostly char *sched_feat_names[] = {
720 #include "sched_features.h"
726 static int sched_feat_show(struct seq_file *m, void *v)
730 for (i = 0; sched_feat_names[i]; i++) {
731 if (!(sysctl_sched_features & (1UL << i)))
733 seq_printf(m, "%s ", sched_feat_names[i]);
741 sched_feat_write(struct file *filp, const char __user *ubuf,
742 size_t cnt, loff_t *ppos)
752 if (copy_from_user(&buf, ubuf, cnt))
757 if (strncmp(buf, "NO_", 3) == 0) {
762 for (i = 0; sched_feat_names[i]; i++) {
763 int len = strlen(sched_feat_names[i]);
765 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
767 sysctl_sched_features &= ~(1UL << i);
769 sysctl_sched_features |= (1UL << i);
774 if (!sched_feat_names[i])
782 static int sched_feat_open(struct inode *inode, struct file *filp)
784 return single_open(filp, sched_feat_show, NULL);
787 static struct file_operations sched_feat_fops = {
788 .open = sched_feat_open,
789 .write = sched_feat_write,
792 .release = single_release,
795 static __init int sched_init_debug(void)
797 debugfs_create_file("sched_features", 0644, NULL, NULL,
802 late_initcall(sched_init_debug);
806 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
809 * Number of tasks to iterate in a single balance run.
810 * Limited because this is done with IRQs disabled.
812 const_debug unsigned int sysctl_sched_nr_migrate = 32;
815 * ratelimit for updating the group shares.
818 unsigned int sysctl_sched_shares_ratelimit = 250000;
821 * Inject some fuzzyness into changing the per-cpu group shares
822 * this avoids remote rq-locks at the expense of fairness.
825 unsigned int sysctl_sched_shares_thresh = 4;
828 * period over which we measure -rt task cpu usage in us.
831 unsigned int sysctl_sched_rt_period = 1000000;
833 static __read_mostly int scheduler_running;
836 * part of the period that we allow rt tasks to run in us.
839 int sysctl_sched_rt_runtime = 950000;
841 static inline u64 global_rt_period(void)
843 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
846 static inline u64 global_rt_runtime(void)
848 if (sysctl_sched_rt_runtime < 0)
851 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
854 #ifndef prepare_arch_switch
855 # define prepare_arch_switch(next) do { } while (0)
857 #ifndef finish_arch_switch
858 # define finish_arch_switch(prev) do { } while (0)
861 static inline int task_current(struct rq *rq, struct task_struct *p)
863 return rq->curr == p;
866 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
867 static inline int task_running(struct rq *rq, struct task_struct *p)
869 return task_current(rq, p);
872 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
876 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
878 #ifdef CONFIG_DEBUG_SPINLOCK
879 /* this is a valid case when another task releases the spinlock */
880 rq->lock.owner = current;
883 * If we are tracking spinlock dependencies then we have to
884 * fix up the runqueue lock - which gets 'carried over' from
887 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
889 spin_unlock_irq(&rq->lock);
892 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
893 static inline int task_running(struct rq *rq, struct task_struct *p)
898 return task_current(rq, p);
902 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
906 * We can optimise this out completely for !SMP, because the
907 * SMP rebalancing from interrupt is the only thing that cares
912 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
913 spin_unlock_irq(&rq->lock);
915 spin_unlock(&rq->lock);
919 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
923 * After ->oncpu is cleared, the task can be moved to a different CPU.
924 * We must ensure this doesn't happen until the switch is completely
930 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
934 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
937 * __task_rq_lock - lock the runqueue a given task resides on.
938 * Must be called interrupts disabled.
940 static inline struct rq *__task_rq_lock(struct task_struct *p)
944 struct rq *rq = task_rq(p);
945 spin_lock(&rq->lock);
946 if (likely(rq == task_rq(p)))
948 spin_unlock(&rq->lock);
953 * task_rq_lock - lock the runqueue a given task resides on and disable
954 * interrupts. Note the ordering: we can safely lookup the task_rq without
955 * explicitly disabling preemption.
957 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
963 local_irq_save(*flags);
965 spin_lock(&rq->lock);
966 if (likely(rq == task_rq(p)))
968 spin_unlock_irqrestore(&rq->lock, *flags);
972 void task_rq_unlock_wait(struct task_struct *p)
974 struct rq *rq = task_rq(p);
976 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
977 spin_unlock_wait(&rq->lock);
980 static void __task_rq_unlock(struct rq *rq)
983 spin_unlock(&rq->lock);
986 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
989 spin_unlock_irqrestore(&rq->lock, *flags);
993 * this_rq_lock - lock this runqueue and disable interrupts.
995 static struct rq *this_rq_lock(void)
1000 local_irq_disable();
1002 spin_lock(&rq->lock);
1007 #ifdef CONFIG_SCHED_HRTICK
1009 * Use HR-timers to deliver accurate preemption points.
1011 * Its all a bit involved since we cannot program an hrt while holding the
1012 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1015 * When we get rescheduled we reprogram the hrtick_timer outside of the
1021 * - enabled by features
1022 * - hrtimer is actually high res
1024 static inline int hrtick_enabled(struct rq *rq)
1026 if (!sched_feat(HRTICK))
1028 if (!cpu_active(cpu_of(rq)))
1030 return hrtimer_is_hres_active(&rq->hrtick_timer);
1033 static void hrtick_clear(struct rq *rq)
1035 if (hrtimer_active(&rq->hrtick_timer))
1036 hrtimer_cancel(&rq->hrtick_timer);
1040 * High-resolution timer tick.
1041 * Runs from hardirq context with interrupts disabled.
1043 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1045 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1047 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1049 spin_lock(&rq->lock);
1050 update_rq_clock(rq);
1051 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1052 spin_unlock(&rq->lock);
1054 return HRTIMER_NORESTART;
1059 * called from hardirq (IPI) context
1061 static void __hrtick_start(void *arg)
1063 struct rq *rq = arg;
1065 spin_lock(&rq->lock);
1066 hrtimer_restart(&rq->hrtick_timer);
1067 rq->hrtick_csd_pending = 0;
1068 spin_unlock(&rq->lock);
1072 * Called to set the hrtick timer state.
1074 * called with rq->lock held and irqs disabled
1076 static void hrtick_start(struct rq *rq, u64 delay)
1078 struct hrtimer *timer = &rq->hrtick_timer;
1079 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1081 hrtimer_set_expires(timer, time);
1083 if (rq == this_rq()) {
1084 hrtimer_restart(timer);
1085 } else if (!rq->hrtick_csd_pending) {
1086 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd);
1087 rq->hrtick_csd_pending = 1;
1092 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1094 int cpu = (int)(long)hcpu;
1097 case CPU_UP_CANCELED:
1098 case CPU_UP_CANCELED_FROZEN:
1099 case CPU_DOWN_PREPARE:
1100 case CPU_DOWN_PREPARE_FROZEN:
1102 case CPU_DEAD_FROZEN:
1103 hrtick_clear(cpu_rq(cpu));
1110 static __init void init_hrtick(void)
1112 hotcpu_notifier(hotplug_hrtick, 0);
1116 * Called to set the hrtick timer state.
1118 * called with rq->lock held and irqs disabled
1120 static void hrtick_start(struct rq *rq, u64 delay)
1122 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay), HRTIMER_MODE_REL);
1125 static inline void init_hrtick(void)
1128 #endif /* CONFIG_SMP */
1130 static void init_rq_hrtick(struct rq *rq)
1133 rq->hrtick_csd_pending = 0;
1135 rq->hrtick_csd.flags = 0;
1136 rq->hrtick_csd.func = __hrtick_start;
1137 rq->hrtick_csd.info = rq;
1140 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1141 rq->hrtick_timer.function = hrtick;
1142 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_PERCPU;
1144 #else /* CONFIG_SCHED_HRTICK */
1145 static inline void hrtick_clear(struct rq *rq)
1149 static inline void init_rq_hrtick(struct rq *rq)
1153 static inline void init_hrtick(void)
1156 #endif /* CONFIG_SCHED_HRTICK */
1159 * resched_task - mark a task 'to be rescheduled now'.
1161 * On UP this means the setting of the need_resched flag, on SMP it
1162 * might also involve a cross-CPU call to trigger the scheduler on
1167 #ifndef tsk_is_polling
1168 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1171 static void resched_task(struct task_struct *p)
1175 assert_spin_locked(&task_rq(p)->lock);
1177 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1180 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1183 if (cpu == smp_processor_id())
1186 /* NEED_RESCHED must be visible before we test polling */
1188 if (!tsk_is_polling(p))
1189 smp_send_reschedule(cpu);
1192 static void resched_cpu(int cpu)
1194 struct rq *rq = cpu_rq(cpu);
1195 unsigned long flags;
1197 if (!spin_trylock_irqsave(&rq->lock, flags))
1199 resched_task(cpu_curr(cpu));
1200 spin_unlock_irqrestore(&rq->lock, flags);
1205 * When add_timer_on() enqueues a timer into the timer wheel of an
1206 * idle CPU then this timer might expire before the next timer event
1207 * which is scheduled to wake up that CPU. In case of a completely
1208 * idle system the next event might even be infinite time into the
1209 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1210 * leaves the inner idle loop so the newly added timer is taken into
1211 * account when the CPU goes back to idle and evaluates the timer
1212 * wheel for the next timer event.
1214 void wake_up_idle_cpu(int cpu)
1216 struct rq *rq = cpu_rq(cpu);
1218 if (cpu == smp_processor_id())
1222 * This is safe, as this function is called with the timer
1223 * wheel base lock of (cpu) held. When the CPU is on the way
1224 * to idle and has not yet set rq->curr to idle then it will
1225 * be serialized on the timer wheel base lock and take the new
1226 * timer into account automatically.
1228 if (rq->curr != rq->idle)
1232 * We can set TIF_RESCHED on the idle task of the other CPU
1233 * lockless. The worst case is that the other CPU runs the
1234 * idle task through an additional NOOP schedule()
1236 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1238 /* NEED_RESCHED must be visible before we test polling */
1240 if (!tsk_is_polling(rq->idle))
1241 smp_send_reschedule(cpu);
1243 #endif /* CONFIG_NO_HZ */
1245 #else /* !CONFIG_SMP */
1246 static void resched_task(struct task_struct *p)
1248 assert_spin_locked(&task_rq(p)->lock);
1249 set_tsk_need_resched(p);
1251 #endif /* CONFIG_SMP */
1253 #if BITS_PER_LONG == 32
1254 # define WMULT_CONST (~0UL)
1256 # define WMULT_CONST (1UL << 32)
1259 #define WMULT_SHIFT 32
1262 * Shift right and round:
1264 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1267 * delta *= weight / lw
1269 static unsigned long
1270 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1271 struct load_weight *lw)
1275 if (!lw->inv_weight) {
1276 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1279 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1283 tmp = (u64)delta_exec * weight;
1285 * Check whether we'd overflow the 64-bit multiplication:
1287 if (unlikely(tmp > WMULT_CONST))
1288 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1291 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1293 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1296 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1302 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1309 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1310 * of tasks with abnormal "nice" values across CPUs the contribution that
1311 * each task makes to its run queue's load is weighted according to its
1312 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1313 * scaled version of the new time slice allocation that they receive on time
1317 #define WEIGHT_IDLEPRIO 2
1318 #define WMULT_IDLEPRIO (1 << 31)
1321 * Nice levels are multiplicative, with a gentle 10% change for every
1322 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1323 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1324 * that remained on nice 0.
1326 * The "10% effect" is relative and cumulative: from _any_ nice level,
1327 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1328 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1329 * If a task goes up by ~10% and another task goes down by ~10% then
1330 * the relative distance between them is ~25%.)
1332 static const int prio_to_weight[40] = {
1333 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1334 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1335 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1336 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1337 /* 0 */ 1024, 820, 655, 526, 423,
1338 /* 5 */ 335, 272, 215, 172, 137,
1339 /* 10 */ 110, 87, 70, 56, 45,
1340 /* 15 */ 36, 29, 23, 18, 15,
1344 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1346 * In cases where the weight does not change often, we can use the
1347 * precalculated inverse to speed up arithmetics by turning divisions
1348 * into multiplications:
1350 static const u32 prio_to_wmult[40] = {
1351 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1352 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1353 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1354 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1355 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1356 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1357 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1358 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1361 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1364 * runqueue iterator, to support SMP load-balancing between different
1365 * scheduling classes, without having to expose their internal data
1366 * structures to the load-balancing proper:
1368 struct rq_iterator {
1370 struct task_struct *(*start)(void *);
1371 struct task_struct *(*next)(void *);
1375 static unsigned long
1376 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1377 unsigned long max_load_move, struct sched_domain *sd,
1378 enum cpu_idle_type idle, int *all_pinned,
1379 int *this_best_prio, struct rq_iterator *iterator);
1382 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1383 struct sched_domain *sd, enum cpu_idle_type idle,
1384 struct rq_iterator *iterator);
1387 #ifdef CONFIG_CGROUP_CPUACCT
1388 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1390 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1393 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1395 update_load_add(&rq->load, load);
1398 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1400 update_load_sub(&rq->load, load);
1403 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1404 typedef int (*tg_visitor)(struct task_group *, void *);
1407 * Iterate the full tree, calling @down when first entering a node and @up when
1408 * leaving it for the final time.
1410 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1412 struct task_group *parent, *child;
1416 parent = &root_task_group;
1418 ret = (*down)(parent, data);
1421 list_for_each_entry_rcu(child, &parent->children, siblings) {
1428 ret = (*up)(parent, data);
1433 parent = parent->parent;
1442 static int tg_nop(struct task_group *tg, void *data)
1449 static unsigned long source_load(int cpu, int type);
1450 static unsigned long target_load(int cpu, int type);
1451 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1453 static unsigned long cpu_avg_load_per_task(int cpu)
1455 struct rq *rq = cpu_rq(cpu);
1456 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1459 rq->avg_load_per_task = rq->load.weight / nr_running;
1461 rq->avg_load_per_task = 0;
1463 return rq->avg_load_per_task;
1466 #ifdef CONFIG_FAIR_GROUP_SCHED
1468 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1471 * Calculate and set the cpu's group shares.
1474 update_group_shares_cpu(struct task_group *tg, int cpu,
1475 unsigned long sd_shares, unsigned long sd_rq_weight)
1477 unsigned long shares;
1478 unsigned long rq_weight;
1483 rq_weight = tg->cfs_rq[cpu]->rq_weight;
1486 * \Sum shares * rq_weight
1487 * shares = -----------------------
1491 shares = (sd_shares * rq_weight) / sd_rq_weight;
1492 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1494 if (abs(shares - tg->se[cpu]->load.weight) >
1495 sysctl_sched_shares_thresh) {
1496 struct rq *rq = cpu_rq(cpu);
1497 unsigned long flags;
1499 spin_lock_irqsave(&rq->lock, flags);
1500 tg->cfs_rq[cpu]->shares = shares;
1502 __set_se_shares(tg->se[cpu], shares);
1503 spin_unlock_irqrestore(&rq->lock, flags);
1508 * Re-compute the task group their per cpu shares over the given domain.
1509 * This needs to be done in a bottom-up fashion because the rq weight of a
1510 * parent group depends on the shares of its child groups.
1512 static int tg_shares_up(struct task_group *tg, void *data)
1514 unsigned long weight, rq_weight = 0;
1515 unsigned long shares = 0;
1516 struct sched_domain *sd = data;
1519 for_each_cpu_mask(i, sd->span) {
1521 * If there are currently no tasks on the cpu pretend there
1522 * is one of average load so that when a new task gets to
1523 * run here it will not get delayed by group starvation.
1525 weight = tg->cfs_rq[i]->load.weight;
1527 weight = NICE_0_LOAD;
1529 tg->cfs_rq[i]->rq_weight = weight;
1530 rq_weight += weight;
1531 shares += tg->cfs_rq[i]->shares;
1534 if ((!shares && rq_weight) || shares > tg->shares)
1535 shares = tg->shares;
1537 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1538 shares = tg->shares;
1540 for_each_cpu_mask(i, sd->span)
1541 update_group_shares_cpu(tg, i, shares, rq_weight);
1547 * Compute the cpu's hierarchical load factor for each task group.
1548 * This needs to be done in a top-down fashion because the load of a child
1549 * group is a fraction of its parents load.
1551 static int tg_load_down(struct task_group *tg, void *data)
1554 long cpu = (long)data;
1557 load = cpu_rq(cpu)->load.weight;
1559 load = tg->parent->cfs_rq[cpu]->h_load;
1560 load *= tg->cfs_rq[cpu]->shares;
1561 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1564 tg->cfs_rq[cpu]->h_load = load;
1569 static void update_shares(struct sched_domain *sd)
1571 u64 now = cpu_clock(raw_smp_processor_id());
1572 s64 elapsed = now - sd->last_update;
1574 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1575 sd->last_update = now;
1576 walk_tg_tree(tg_nop, tg_shares_up, sd);
1580 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1582 spin_unlock(&rq->lock);
1584 spin_lock(&rq->lock);
1587 static void update_h_load(long cpu)
1589 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1594 static inline void update_shares(struct sched_domain *sd)
1598 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1605 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1607 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1608 __releases(this_rq->lock)
1609 __acquires(busiest->lock)
1610 __acquires(this_rq->lock)
1614 if (unlikely(!irqs_disabled())) {
1615 /* printk() doesn't work good under rq->lock */
1616 spin_unlock(&this_rq->lock);
1619 if (unlikely(!spin_trylock(&busiest->lock))) {
1620 if (busiest < this_rq) {
1621 spin_unlock(&this_rq->lock);
1622 spin_lock(&busiest->lock);
1623 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1626 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1631 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1632 __releases(busiest->lock)
1634 spin_unlock(&busiest->lock);
1635 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1639 #ifdef CONFIG_FAIR_GROUP_SCHED
1640 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1643 cfs_rq->shares = shares;
1648 #include "sched_stats.h"
1649 #include "sched_idletask.c"
1650 #include "sched_fair.c"
1651 #include "sched_rt.c"
1652 #ifdef CONFIG_SCHED_DEBUG
1653 # include "sched_debug.c"
1656 #define sched_class_highest (&rt_sched_class)
1657 #define for_each_class(class) \
1658 for (class = sched_class_highest; class; class = class->next)
1660 static void inc_nr_running(struct rq *rq)
1665 static void dec_nr_running(struct rq *rq)
1670 static void set_load_weight(struct task_struct *p)
1672 if (task_has_rt_policy(p)) {
1673 p->se.load.weight = prio_to_weight[0] * 2;
1674 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1679 * SCHED_IDLE tasks get minimal weight:
1681 if (p->policy == SCHED_IDLE) {
1682 p->se.load.weight = WEIGHT_IDLEPRIO;
1683 p->se.load.inv_weight = WMULT_IDLEPRIO;
1687 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1688 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1691 static void update_avg(u64 *avg, u64 sample)
1693 s64 diff = sample - *avg;
1697 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1699 sched_info_queued(p);
1700 p->sched_class->enqueue_task(rq, p, wakeup);
1704 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1706 if (sleep && p->se.last_wakeup) {
1707 update_avg(&p->se.avg_overlap,
1708 p->se.sum_exec_runtime - p->se.last_wakeup);
1709 p->se.last_wakeup = 0;
1712 sched_info_dequeued(p);
1713 p->sched_class->dequeue_task(rq, p, sleep);
1718 * __normal_prio - return the priority that is based on the static prio
1720 static inline int __normal_prio(struct task_struct *p)
1722 return p->static_prio;
1726 * Calculate the expected normal priority: i.e. priority
1727 * without taking RT-inheritance into account. Might be
1728 * boosted by interactivity modifiers. Changes upon fork,
1729 * setprio syscalls, and whenever the interactivity
1730 * estimator recalculates.
1732 static inline int normal_prio(struct task_struct *p)
1736 if (task_has_rt_policy(p))
1737 prio = MAX_RT_PRIO-1 - p->rt_priority;
1739 prio = __normal_prio(p);
1744 * Calculate the current priority, i.e. the priority
1745 * taken into account by the scheduler. This value might
1746 * be boosted by RT tasks, or might be boosted by
1747 * interactivity modifiers. Will be RT if the task got
1748 * RT-boosted. If not then it returns p->normal_prio.
1750 static int effective_prio(struct task_struct *p)
1752 p->normal_prio = normal_prio(p);
1754 * If we are RT tasks or we were boosted to RT priority,
1755 * keep the priority unchanged. Otherwise, update priority
1756 * to the normal priority:
1758 if (!rt_prio(p->prio))
1759 return p->normal_prio;
1764 * activate_task - move a task to the runqueue.
1766 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1768 if (task_contributes_to_load(p))
1769 rq->nr_uninterruptible--;
1771 enqueue_task(rq, p, wakeup);
1776 * deactivate_task - remove a task from the runqueue.
1778 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1780 if (task_contributes_to_load(p))
1781 rq->nr_uninterruptible++;
1783 dequeue_task(rq, p, sleep);
1788 * task_curr - is this task currently executing on a CPU?
1789 * @p: the task in question.
1791 inline int task_curr(const struct task_struct *p)
1793 return cpu_curr(task_cpu(p)) == p;
1796 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1798 set_task_rq(p, cpu);
1801 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1802 * successfuly executed on another CPU. We must ensure that updates of
1803 * per-task data have been completed by this moment.
1806 task_thread_info(p)->cpu = cpu;
1810 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1811 const struct sched_class *prev_class,
1812 int oldprio, int running)
1814 if (prev_class != p->sched_class) {
1815 if (prev_class->switched_from)
1816 prev_class->switched_from(rq, p, running);
1817 p->sched_class->switched_to(rq, p, running);
1819 p->sched_class->prio_changed(rq, p, oldprio, running);
1824 /* Used instead of source_load when we know the type == 0 */
1825 static unsigned long weighted_cpuload(const int cpu)
1827 return cpu_rq(cpu)->load.weight;
1831 * Is this task likely cache-hot:
1834 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1839 * Buddy candidates are cache hot:
1841 if (sched_feat(CACHE_HOT_BUDDY) &&
1842 (&p->se == cfs_rq_of(&p->se)->next ||
1843 &p->se == cfs_rq_of(&p->se)->last))
1846 if (p->sched_class != &fair_sched_class)
1849 if (sysctl_sched_migration_cost == -1)
1851 if (sysctl_sched_migration_cost == 0)
1854 delta = now - p->se.exec_start;
1856 return delta < (s64)sysctl_sched_migration_cost;
1860 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1862 int old_cpu = task_cpu(p);
1863 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1864 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1865 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1868 clock_offset = old_rq->clock - new_rq->clock;
1870 trace_sched_migrate_task(p, task_cpu(p), new_cpu);
1872 #ifdef CONFIG_SCHEDSTATS
1873 if (p->se.wait_start)
1874 p->se.wait_start -= clock_offset;
1875 if (p->se.sleep_start)
1876 p->se.sleep_start -= clock_offset;
1877 if (p->se.block_start)
1878 p->se.block_start -= clock_offset;
1879 if (old_cpu != new_cpu) {
1880 schedstat_inc(p, se.nr_migrations);
1881 if (task_hot(p, old_rq->clock, NULL))
1882 schedstat_inc(p, se.nr_forced2_migrations);
1885 p->se.vruntime -= old_cfsrq->min_vruntime -
1886 new_cfsrq->min_vruntime;
1888 __set_task_cpu(p, new_cpu);
1891 struct migration_req {
1892 struct list_head list;
1894 struct task_struct *task;
1897 struct completion done;
1901 * The task's runqueue lock must be held.
1902 * Returns true if you have to wait for migration thread.
1905 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1907 struct rq *rq = task_rq(p);
1910 * If the task is not on a runqueue (and not running), then
1911 * it is sufficient to simply update the task's cpu field.
1913 if (!p->se.on_rq && !task_running(rq, p)) {
1914 set_task_cpu(p, dest_cpu);
1918 init_completion(&req->done);
1920 req->dest_cpu = dest_cpu;
1921 list_add(&req->list, &rq->migration_queue);
1927 * wait_task_inactive - wait for a thread to unschedule.
1929 * If @match_state is nonzero, it's the @p->state value just checked and
1930 * not expected to change. If it changes, i.e. @p might have woken up,
1931 * then return zero. When we succeed in waiting for @p to be off its CPU,
1932 * we return a positive number (its total switch count). If a second call
1933 * a short while later returns the same number, the caller can be sure that
1934 * @p has remained unscheduled the whole time.
1936 * The caller must ensure that the task *will* unschedule sometime soon,
1937 * else this function might spin for a *long* time. This function can't
1938 * be called with interrupts off, or it may introduce deadlock with
1939 * smp_call_function() if an IPI is sent by the same process we are
1940 * waiting to become inactive.
1942 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1944 unsigned long flags;
1951 * We do the initial early heuristics without holding
1952 * any task-queue locks at all. We'll only try to get
1953 * the runqueue lock when things look like they will
1959 * If the task is actively running on another CPU
1960 * still, just relax and busy-wait without holding
1963 * NOTE! Since we don't hold any locks, it's not
1964 * even sure that "rq" stays as the right runqueue!
1965 * But we don't care, since "task_running()" will
1966 * return false if the runqueue has changed and p
1967 * is actually now running somewhere else!
1969 while (task_running(rq, p)) {
1970 if (match_state && unlikely(p->state != match_state))
1976 * Ok, time to look more closely! We need the rq
1977 * lock now, to be *sure*. If we're wrong, we'll
1978 * just go back and repeat.
1980 rq = task_rq_lock(p, &flags);
1981 trace_sched_wait_task(rq, p);
1982 running = task_running(rq, p);
1983 on_rq = p->se.on_rq;
1985 if (!match_state || p->state == match_state)
1986 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1987 task_rq_unlock(rq, &flags);
1990 * If it changed from the expected state, bail out now.
1992 if (unlikely(!ncsw))
1996 * Was it really running after all now that we
1997 * checked with the proper locks actually held?
1999 * Oops. Go back and try again..
2001 if (unlikely(running)) {
2007 * It's not enough that it's not actively running,
2008 * it must be off the runqueue _entirely_, and not
2011 * So if it wa still runnable (but just not actively
2012 * running right now), it's preempted, and we should
2013 * yield - it could be a while.
2015 if (unlikely(on_rq)) {
2016 schedule_timeout_uninterruptible(1);
2021 * Ahh, all good. It wasn't running, and it wasn't
2022 * runnable, which means that it will never become
2023 * running in the future either. We're all done!
2032 * kick_process - kick a running thread to enter/exit the kernel
2033 * @p: the to-be-kicked thread
2035 * Cause a process which is running on another CPU to enter
2036 * kernel-mode, without any delay. (to get signals handled.)
2038 * NOTE: this function doesnt have to take the runqueue lock,
2039 * because all it wants to ensure is that the remote task enters
2040 * the kernel. If the IPI races and the task has been migrated
2041 * to another CPU then no harm is done and the purpose has been
2044 void kick_process(struct task_struct *p)
2050 if ((cpu != smp_processor_id()) && task_curr(p))
2051 smp_send_reschedule(cpu);
2056 * Return a low guess at the load of a migration-source cpu weighted
2057 * according to the scheduling class and "nice" value.
2059 * We want to under-estimate the load of migration sources, to
2060 * balance conservatively.
2062 static unsigned long source_load(int cpu, int type)
2064 struct rq *rq = cpu_rq(cpu);
2065 unsigned long total = weighted_cpuload(cpu);
2067 if (type == 0 || !sched_feat(LB_BIAS))
2070 return min(rq->cpu_load[type-1], total);
2074 * Return a high guess at the load of a migration-target cpu weighted
2075 * according to the scheduling class and "nice" value.
2077 static unsigned long target_load(int cpu, int type)
2079 struct rq *rq = cpu_rq(cpu);
2080 unsigned long total = weighted_cpuload(cpu);
2082 if (type == 0 || !sched_feat(LB_BIAS))
2085 return max(rq->cpu_load[type-1], total);
2089 * find_idlest_group finds and returns the least busy CPU group within the
2092 static struct sched_group *
2093 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2095 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2096 unsigned long min_load = ULONG_MAX, this_load = 0;
2097 int load_idx = sd->forkexec_idx;
2098 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2101 unsigned long load, avg_load;
2105 /* Skip over this group if it has no CPUs allowed */
2106 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
2109 local_group = cpu_isset(this_cpu, group->cpumask);
2111 /* Tally up the load of all CPUs in the group */
2114 for_each_cpu_mask_nr(i, group->cpumask) {
2115 /* Bias balancing toward cpus of our domain */
2117 load = source_load(i, load_idx);
2119 load = target_load(i, load_idx);
2124 /* Adjust by relative CPU power of the group */
2125 avg_load = sg_div_cpu_power(group,
2126 avg_load * SCHED_LOAD_SCALE);
2129 this_load = avg_load;
2131 } else if (avg_load < min_load) {
2132 min_load = avg_load;
2135 } while (group = group->next, group != sd->groups);
2137 if (!idlest || 100*this_load < imbalance*min_load)
2143 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2146 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2149 unsigned long load, min_load = ULONG_MAX;
2153 /* Traverse only the allowed CPUs */
2154 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
2156 for_each_cpu_mask_nr(i, *tmp) {
2157 load = weighted_cpuload(i);
2159 if (load < min_load || (load == min_load && i == this_cpu)) {
2169 * sched_balance_self: balance the current task (running on cpu) in domains
2170 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2173 * Balance, ie. select the least loaded group.
2175 * Returns the target CPU number, or the same CPU if no balancing is needed.
2177 * preempt must be disabled.
2179 static int sched_balance_self(int cpu, int flag)
2181 struct task_struct *t = current;
2182 struct sched_domain *tmp, *sd = NULL;
2184 for_each_domain(cpu, tmp) {
2186 * If power savings logic is enabled for a domain, stop there.
2188 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2190 if (tmp->flags & flag)
2198 cpumask_t span, tmpmask;
2199 struct sched_group *group;
2200 int new_cpu, weight;
2202 if (!(sd->flags & flag)) {
2208 group = find_idlest_group(sd, t, cpu);
2214 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2215 if (new_cpu == -1 || new_cpu == cpu) {
2216 /* Now try balancing at a lower domain level of cpu */
2221 /* Now try balancing at a lower domain level of new_cpu */
2224 weight = cpus_weight(span);
2225 for_each_domain(cpu, tmp) {
2226 if (weight <= cpus_weight(tmp->span))
2228 if (tmp->flags & flag)
2231 /* while loop will break here if sd == NULL */
2237 #endif /* CONFIG_SMP */
2240 * try_to_wake_up - wake up a thread
2241 * @p: the to-be-woken-up thread
2242 * @state: the mask of task states that can be woken
2243 * @sync: do a synchronous wakeup?
2245 * Put it on the run-queue if it's not already there. The "current"
2246 * thread is always on the run-queue (except when the actual
2247 * re-schedule is in progress), and as such you're allowed to do
2248 * the simpler "current->state = TASK_RUNNING" to mark yourself
2249 * runnable without the overhead of this.
2251 * returns failure only if the task is already active.
2253 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2255 int cpu, orig_cpu, this_cpu, success = 0;
2256 unsigned long flags;
2260 if (!sched_feat(SYNC_WAKEUPS))
2264 if (sched_feat(LB_WAKEUP_UPDATE)) {
2265 struct sched_domain *sd;
2267 this_cpu = raw_smp_processor_id();
2270 for_each_domain(this_cpu, sd) {
2271 if (cpu_isset(cpu, sd->span)) {
2280 rq = task_rq_lock(p, &flags);
2281 update_rq_clock(rq);
2282 old_state = p->state;
2283 if (!(old_state & state))
2291 this_cpu = smp_processor_id();
2294 if (unlikely(task_running(rq, p)))
2297 cpu = p->sched_class->select_task_rq(p, sync);
2298 if (cpu != orig_cpu) {
2299 set_task_cpu(p, cpu);
2300 task_rq_unlock(rq, &flags);
2301 /* might preempt at this point */
2302 rq = task_rq_lock(p, &flags);
2303 old_state = p->state;
2304 if (!(old_state & state))
2309 this_cpu = smp_processor_id();
2313 #ifdef CONFIG_SCHEDSTATS
2314 schedstat_inc(rq, ttwu_count);
2315 if (cpu == this_cpu)
2316 schedstat_inc(rq, ttwu_local);
2318 struct sched_domain *sd;
2319 for_each_domain(this_cpu, sd) {
2320 if (cpu_isset(cpu, sd->span)) {
2321 schedstat_inc(sd, ttwu_wake_remote);
2326 #endif /* CONFIG_SCHEDSTATS */
2329 #endif /* CONFIG_SMP */
2330 schedstat_inc(p, se.nr_wakeups);
2332 schedstat_inc(p, se.nr_wakeups_sync);
2333 if (orig_cpu != cpu)
2334 schedstat_inc(p, se.nr_wakeups_migrate);
2335 if (cpu == this_cpu)
2336 schedstat_inc(p, se.nr_wakeups_local);
2338 schedstat_inc(p, se.nr_wakeups_remote);
2339 activate_task(rq, p, 1);
2343 trace_sched_wakeup(rq, p, success);
2344 check_preempt_curr(rq, p, sync);
2346 p->state = TASK_RUNNING;
2348 if (p->sched_class->task_wake_up)
2349 p->sched_class->task_wake_up(rq, p);
2352 current->se.last_wakeup = current->se.sum_exec_runtime;
2354 task_rq_unlock(rq, &flags);
2359 int wake_up_process(struct task_struct *p)
2361 return try_to_wake_up(p, TASK_ALL, 0);
2363 EXPORT_SYMBOL(wake_up_process);
2365 int wake_up_state(struct task_struct *p, unsigned int state)
2367 return try_to_wake_up(p, state, 0);
2371 * Perform scheduler related setup for a newly forked process p.
2372 * p is forked by current.
2374 * __sched_fork() is basic setup used by init_idle() too:
2376 static void __sched_fork(struct task_struct *p)
2378 p->se.exec_start = 0;
2379 p->se.sum_exec_runtime = 0;
2380 p->se.prev_sum_exec_runtime = 0;
2381 p->se.last_wakeup = 0;
2382 p->se.avg_overlap = 0;
2384 #ifdef CONFIG_SCHEDSTATS
2385 p->se.wait_start = 0;
2386 p->se.sum_sleep_runtime = 0;
2387 p->se.sleep_start = 0;
2388 p->se.block_start = 0;
2389 p->se.sleep_max = 0;
2390 p->se.block_max = 0;
2392 p->se.slice_max = 0;
2396 INIT_LIST_HEAD(&p->rt.run_list);
2398 INIT_LIST_HEAD(&p->se.group_node);
2400 #ifdef CONFIG_PREEMPT_NOTIFIERS
2401 INIT_HLIST_HEAD(&p->preempt_notifiers);
2405 * We mark the process as running here, but have not actually
2406 * inserted it onto the runqueue yet. This guarantees that
2407 * nobody will actually run it, and a signal or other external
2408 * event cannot wake it up and insert it on the runqueue either.
2410 p->state = TASK_RUNNING;
2414 * fork()/clone()-time setup:
2416 void sched_fork(struct task_struct *p, int clone_flags)
2418 int cpu = get_cpu();
2423 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2425 set_task_cpu(p, cpu);
2428 * Make sure we do not leak PI boosting priority to the child:
2430 p->prio = current->normal_prio;
2431 if (!rt_prio(p->prio))
2432 p->sched_class = &fair_sched_class;
2434 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2435 if (likely(sched_info_on()))
2436 memset(&p->sched_info, 0, sizeof(p->sched_info));
2438 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2441 #ifdef CONFIG_PREEMPT
2442 /* Want to start with kernel preemption disabled. */
2443 task_thread_info(p)->preempt_count = 1;
2449 * wake_up_new_task - wake up a newly created task for the first time.
2451 * This function will do some initial scheduler statistics housekeeping
2452 * that must be done for every newly created context, then puts the task
2453 * on the runqueue and wakes it.
2455 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2457 unsigned long flags;
2460 rq = task_rq_lock(p, &flags);
2461 BUG_ON(p->state != TASK_RUNNING);
2462 update_rq_clock(rq);
2464 p->prio = effective_prio(p);
2466 if (!p->sched_class->task_new || !current->se.on_rq) {
2467 activate_task(rq, p, 0);
2470 * Let the scheduling class do new task startup
2471 * management (if any):
2473 p->sched_class->task_new(rq, p);
2476 trace_sched_wakeup_new(rq, p, 1);
2477 check_preempt_curr(rq, p, 0);
2479 if (p->sched_class->task_wake_up)
2480 p->sched_class->task_wake_up(rq, p);
2482 task_rq_unlock(rq, &flags);
2485 #ifdef CONFIG_PREEMPT_NOTIFIERS
2488 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2489 * @notifier: notifier struct to register
2491 void preempt_notifier_register(struct preempt_notifier *notifier)
2493 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2495 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2498 * preempt_notifier_unregister - no longer interested in preemption notifications
2499 * @notifier: notifier struct to unregister
2501 * This is safe to call from within a preemption notifier.
2503 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2505 hlist_del(¬ifier->link);
2507 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2509 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2511 struct preempt_notifier *notifier;
2512 struct hlist_node *node;
2514 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2515 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2519 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2520 struct task_struct *next)
2522 struct preempt_notifier *notifier;
2523 struct hlist_node *node;
2525 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2526 notifier->ops->sched_out(notifier, next);
2529 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2531 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2536 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2537 struct task_struct *next)
2541 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2544 * prepare_task_switch - prepare to switch tasks
2545 * @rq: the runqueue preparing to switch
2546 * @prev: the current task that is being switched out
2547 * @next: the task we are going to switch to.
2549 * This is called with the rq lock held and interrupts off. It must
2550 * be paired with a subsequent finish_task_switch after the context
2553 * prepare_task_switch sets up locking and calls architecture specific
2557 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2558 struct task_struct *next)
2560 fire_sched_out_preempt_notifiers(prev, next);
2561 prepare_lock_switch(rq, next);
2562 prepare_arch_switch(next);
2566 * finish_task_switch - clean up after a task-switch
2567 * @rq: runqueue associated with task-switch
2568 * @prev: the thread we just switched away from.
2570 * finish_task_switch must be called after the context switch, paired
2571 * with a prepare_task_switch call before the context switch.
2572 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2573 * and do any other architecture-specific cleanup actions.
2575 * Note that we may have delayed dropping an mm in context_switch(). If
2576 * so, we finish that here outside of the runqueue lock. (Doing it
2577 * with the lock held can cause deadlocks; see schedule() for
2580 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2581 __releases(rq->lock)
2583 struct mm_struct *mm = rq->prev_mm;
2589 * A task struct has one reference for the use as "current".
2590 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2591 * schedule one last time. The schedule call will never return, and
2592 * the scheduled task must drop that reference.
2593 * The test for TASK_DEAD must occur while the runqueue locks are
2594 * still held, otherwise prev could be scheduled on another cpu, die
2595 * there before we look at prev->state, and then the reference would
2597 * Manfred Spraul <manfred@colorfullife.com>
2599 prev_state = prev->state;
2600 finish_arch_switch(prev);
2601 finish_lock_switch(rq, prev);
2603 if (current->sched_class->post_schedule)
2604 current->sched_class->post_schedule(rq);
2607 fire_sched_in_preempt_notifiers(current);
2610 if (unlikely(prev_state == TASK_DEAD)) {
2612 * Remove function-return probe instances associated with this
2613 * task and put them back on the free list.
2615 kprobe_flush_task(prev);
2616 put_task_struct(prev);
2621 * schedule_tail - first thing a freshly forked thread must call.
2622 * @prev: the thread we just switched away from.
2624 asmlinkage void schedule_tail(struct task_struct *prev)
2625 __releases(rq->lock)
2627 struct rq *rq = this_rq();
2629 finish_task_switch(rq, prev);
2630 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2631 /* In this case, finish_task_switch does not reenable preemption */
2634 if (current->set_child_tid)
2635 put_user(task_pid_vnr(current), current->set_child_tid);
2639 * context_switch - switch to the new MM and the new
2640 * thread's register state.
2643 context_switch(struct rq *rq, struct task_struct *prev,
2644 struct task_struct *next)
2646 struct mm_struct *mm, *oldmm;
2648 prepare_task_switch(rq, prev, next);
2649 trace_sched_switch(rq, prev, next);
2651 oldmm = prev->active_mm;
2653 * For paravirt, this is coupled with an exit in switch_to to
2654 * combine the page table reload and the switch backend into
2657 arch_enter_lazy_cpu_mode();
2659 if (unlikely(!mm)) {
2660 next->active_mm = oldmm;
2661 atomic_inc(&oldmm->mm_count);
2662 enter_lazy_tlb(oldmm, next);
2664 switch_mm(oldmm, mm, next);
2666 if (unlikely(!prev->mm)) {
2667 prev->active_mm = NULL;
2668 rq->prev_mm = oldmm;
2671 * Since the runqueue lock will be released by the next
2672 * task (which is an invalid locking op but in the case
2673 * of the scheduler it's an obvious special-case), so we
2674 * do an early lockdep release here:
2676 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2677 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2680 /* Here we just switch the register state and the stack. */
2681 switch_to(prev, next, prev);
2685 * this_rq must be evaluated again because prev may have moved
2686 * CPUs since it called schedule(), thus the 'rq' on its stack
2687 * frame will be invalid.
2689 finish_task_switch(this_rq(), prev);
2693 * nr_running, nr_uninterruptible and nr_context_switches:
2695 * externally visible scheduler statistics: current number of runnable
2696 * threads, current number of uninterruptible-sleeping threads, total
2697 * number of context switches performed since bootup.
2699 unsigned long nr_running(void)
2701 unsigned long i, sum = 0;
2703 for_each_online_cpu(i)
2704 sum += cpu_rq(i)->nr_running;
2709 unsigned long nr_uninterruptible(void)
2711 unsigned long i, sum = 0;
2713 for_each_possible_cpu(i)
2714 sum += cpu_rq(i)->nr_uninterruptible;
2717 * Since we read the counters lockless, it might be slightly
2718 * inaccurate. Do not allow it to go below zero though:
2720 if (unlikely((long)sum < 0))
2726 unsigned long long nr_context_switches(void)
2729 unsigned long long sum = 0;
2731 for_each_possible_cpu(i)
2732 sum += cpu_rq(i)->nr_switches;
2737 unsigned long nr_iowait(void)
2739 unsigned long i, sum = 0;
2741 for_each_possible_cpu(i)
2742 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2747 unsigned long nr_active(void)
2749 unsigned long i, running = 0, uninterruptible = 0;
2751 for_each_online_cpu(i) {
2752 running += cpu_rq(i)->nr_running;
2753 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2756 if (unlikely((long)uninterruptible < 0))
2757 uninterruptible = 0;
2759 return running + uninterruptible;
2763 * Update rq->cpu_load[] statistics. This function is usually called every
2764 * scheduler tick (TICK_NSEC).
2766 static void update_cpu_load(struct rq *this_rq)
2768 unsigned long this_load = this_rq->load.weight;
2771 this_rq->nr_load_updates++;
2773 /* Update our load: */
2774 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2775 unsigned long old_load, new_load;
2777 /* scale is effectively 1 << i now, and >> i divides by scale */
2779 old_load = this_rq->cpu_load[i];
2780 new_load = this_load;
2782 * Round up the averaging division if load is increasing. This
2783 * prevents us from getting stuck on 9 if the load is 10, for
2786 if (new_load > old_load)
2787 new_load += scale-1;
2788 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2795 * double_rq_lock - safely lock two runqueues
2797 * Note this does not disable interrupts like task_rq_lock,
2798 * you need to do so manually before calling.
2800 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2801 __acquires(rq1->lock)
2802 __acquires(rq2->lock)
2804 BUG_ON(!irqs_disabled());
2806 spin_lock(&rq1->lock);
2807 __acquire(rq2->lock); /* Fake it out ;) */
2810 spin_lock(&rq1->lock);
2811 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2813 spin_lock(&rq2->lock);
2814 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2817 update_rq_clock(rq1);
2818 update_rq_clock(rq2);
2822 * double_rq_unlock - safely unlock two runqueues
2824 * Note this does not restore interrupts like task_rq_unlock,
2825 * you need to do so manually after calling.
2827 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2828 __releases(rq1->lock)
2829 __releases(rq2->lock)
2831 spin_unlock(&rq1->lock);
2833 spin_unlock(&rq2->lock);
2835 __release(rq2->lock);
2839 * If dest_cpu is allowed for this process, migrate the task to it.
2840 * This is accomplished by forcing the cpu_allowed mask to only
2841 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2842 * the cpu_allowed mask is restored.
2844 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2846 struct migration_req req;
2847 unsigned long flags;
2850 rq = task_rq_lock(p, &flags);
2851 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2852 || unlikely(!cpu_active(dest_cpu)))
2855 /* force the process onto the specified CPU */
2856 if (migrate_task(p, dest_cpu, &req)) {
2857 /* Need to wait for migration thread (might exit: take ref). */
2858 struct task_struct *mt = rq->migration_thread;
2860 get_task_struct(mt);
2861 task_rq_unlock(rq, &flags);
2862 wake_up_process(mt);
2863 put_task_struct(mt);
2864 wait_for_completion(&req.done);
2869 task_rq_unlock(rq, &flags);
2873 * sched_exec - execve() is a valuable balancing opportunity, because at
2874 * this point the task has the smallest effective memory and cache footprint.
2876 void sched_exec(void)
2878 int new_cpu, this_cpu = get_cpu();
2879 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2881 if (new_cpu != this_cpu)
2882 sched_migrate_task(current, new_cpu);
2886 * pull_task - move a task from a remote runqueue to the local runqueue.
2887 * Both runqueues must be locked.
2889 static void pull_task(struct rq *src_rq, struct task_struct *p,
2890 struct rq *this_rq, int this_cpu)
2892 deactivate_task(src_rq, p, 0);
2893 set_task_cpu(p, this_cpu);
2894 activate_task(this_rq, p, 0);
2896 * Note that idle threads have a prio of MAX_PRIO, for this test
2897 * to be always true for them.
2899 check_preempt_curr(this_rq, p, 0);
2903 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2906 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2907 struct sched_domain *sd, enum cpu_idle_type idle,
2911 * We do not migrate tasks that are:
2912 * 1) running (obviously), or
2913 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2914 * 3) are cache-hot on their current CPU.
2916 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2917 schedstat_inc(p, se.nr_failed_migrations_affine);
2922 if (task_running(rq, p)) {
2923 schedstat_inc(p, se.nr_failed_migrations_running);
2928 * Aggressive migration if:
2929 * 1) task is cache cold, or
2930 * 2) too many balance attempts have failed.
2933 if (!task_hot(p, rq->clock, sd) ||
2934 sd->nr_balance_failed > sd->cache_nice_tries) {
2935 #ifdef CONFIG_SCHEDSTATS
2936 if (task_hot(p, rq->clock, sd)) {
2937 schedstat_inc(sd, lb_hot_gained[idle]);
2938 schedstat_inc(p, se.nr_forced_migrations);
2944 if (task_hot(p, rq->clock, sd)) {
2945 schedstat_inc(p, se.nr_failed_migrations_hot);
2951 static unsigned long
2952 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2953 unsigned long max_load_move, struct sched_domain *sd,
2954 enum cpu_idle_type idle, int *all_pinned,
2955 int *this_best_prio, struct rq_iterator *iterator)
2957 int loops = 0, pulled = 0, pinned = 0;
2958 struct task_struct *p;
2959 long rem_load_move = max_load_move;
2961 if (max_load_move == 0)
2967 * Start the load-balancing iterator:
2969 p = iterator->start(iterator->arg);
2971 if (!p || loops++ > sysctl_sched_nr_migrate)
2974 if ((p->se.load.weight >> 1) > rem_load_move ||
2975 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2976 p = iterator->next(iterator->arg);
2980 pull_task(busiest, p, this_rq, this_cpu);
2982 rem_load_move -= p->se.load.weight;
2985 * We only want to steal up to the prescribed amount of weighted load.
2987 if (rem_load_move > 0) {
2988 if (p->prio < *this_best_prio)
2989 *this_best_prio = p->prio;
2990 p = iterator->next(iterator->arg);
2995 * Right now, this is one of only two places pull_task() is called,
2996 * so we can safely collect pull_task() stats here rather than
2997 * inside pull_task().
2999 schedstat_add(sd, lb_gained[idle], pulled);
3002 *all_pinned = pinned;
3004 return max_load_move - rem_load_move;
3008 * move_tasks tries to move up to max_load_move weighted load from busiest to
3009 * this_rq, as part of a balancing operation within domain "sd".
3010 * Returns 1 if successful and 0 otherwise.
3012 * Called with both runqueues locked.
3014 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3015 unsigned long max_load_move,
3016 struct sched_domain *sd, enum cpu_idle_type idle,
3019 const struct sched_class *class = sched_class_highest;
3020 unsigned long total_load_moved = 0;
3021 int this_best_prio = this_rq->curr->prio;
3025 class->load_balance(this_rq, this_cpu, busiest,
3026 max_load_move - total_load_moved,
3027 sd, idle, all_pinned, &this_best_prio);
3028 class = class->next;
3030 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3033 } while (class && max_load_move > total_load_moved);
3035 return total_load_moved > 0;
3039 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3040 struct sched_domain *sd, enum cpu_idle_type idle,
3041 struct rq_iterator *iterator)
3043 struct task_struct *p = iterator->start(iterator->arg);
3047 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3048 pull_task(busiest, p, this_rq, this_cpu);
3050 * Right now, this is only the second place pull_task()
3051 * is called, so we can safely collect pull_task()
3052 * stats here rather than inside pull_task().
3054 schedstat_inc(sd, lb_gained[idle]);
3058 p = iterator->next(iterator->arg);
3065 * move_one_task tries to move exactly one task from busiest to this_rq, as
3066 * part of active balancing operations within "domain".
3067 * Returns 1 if successful and 0 otherwise.
3069 * Called with both runqueues locked.
3071 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3072 struct sched_domain *sd, enum cpu_idle_type idle)
3074 const struct sched_class *class;
3076 for (class = sched_class_highest; class; class = class->next)
3077 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3084 * find_busiest_group finds and returns the busiest CPU group within the
3085 * domain. It calculates and returns the amount of weighted load which
3086 * should be moved to restore balance via the imbalance parameter.
3088 static struct sched_group *
3089 find_busiest_group(struct sched_domain *sd, int this_cpu,
3090 unsigned long *imbalance, enum cpu_idle_type idle,
3091 int *sd_idle, const cpumask_t *cpus, int *balance)
3093 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3094 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3095 unsigned long max_pull;
3096 unsigned long busiest_load_per_task, busiest_nr_running;
3097 unsigned long this_load_per_task, this_nr_running;
3098 int load_idx, group_imb = 0;
3099 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3100 int power_savings_balance = 1;
3101 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3102 unsigned long min_nr_running = ULONG_MAX;
3103 struct sched_group *group_min = NULL, *group_leader = NULL;
3106 max_load = this_load = total_load = total_pwr = 0;
3107 busiest_load_per_task = busiest_nr_running = 0;
3108 this_load_per_task = this_nr_running = 0;
3110 if (idle == CPU_NOT_IDLE)
3111 load_idx = sd->busy_idx;
3112 else if (idle == CPU_NEWLY_IDLE)
3113 load_idx = sd->newidle_idx;
3115 load_idx = sd->idle_idx;
3118 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3121 int __group_imb = 0;
3122 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3123 unsigned long sum_nr_running, sum_weighted_load;
3124 unsigned long sum_avg_load_per_task;
3125 unsigned long avg_load_per_task;
3127 local_group = cpu_isset(this_cpu, group->cpumask);
3130 balance_cpu = first_cpu(group->cpumask);
3132 /* Tally up the load of all CPUs in the group */
3133 sum_weighted_load = sum_nr_running = avg_load = 0;
3134 sum_avg_load_per_task = avg_load_per_task = 0;
3137 min_cpu_load = ~0UL;
3139 for_each_cpu_mask_nr(i, group->cpumask) {
3142 if (!cpu_isset(i, *cpus))
3147 if (*sd_idle && rq->nr_running)
3150 /* Bias balancing toward cpus of our domain */
3152 if (idle_cpu(i) && !first_idle_cpu) {
3157 load = target_load(i, load_idx);
3159 load = source_load(i, load_idx);
3160 if (load > max_cpu_load)
3161 max_cpu_load = load;
3162 if (min_cpu_load > load)
3163 min_cpu_load = load;
3167 sum_nr_running += rq->nr_running;
3168 sum_weighted_load += weighted_cpuload(i);
3170 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3174 * First idle cpu or the first cpu(busiest) in this sched group
3175 * is eligible for doing load balancing at this and above
3176 * domains. In the newly idle case, we will allow all the cpu's
3177 * to do the newly idle load balance.
3179 if (idle != CPU_NEWLY_IDLE && local_group &&
3180 balance_cpu != this_cpu && balance) {
3185 total_load += avg_load;
3186 total_pwr += group->__cpu_power;
3188 /* Adjust by relative CPU power of the group */
3189 avg_load = sg_div_cpu_power(group,
3190 avg_load * SCHED_LOAD_SCALE);
3194 * Consider the group unbalanced when the imbalance is larger
3195 * than the average weight of two tasks.
3197 * APZ: with cgroup the avg task weight can vary wildly and
3198 * might not be a suitable number - should we keep a
3199 * normalized nr_running number somewhere that negates
3202 avg_load_per_task = sg_div_cpu_power(group,
3203 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3205 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3208 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3211 this_load = avg_load;
3213 this_nr_running = sum_nr_running;
3214 this_load_per_task = sum_weighted_load;
3215 } else if (avg_load > max_load &&
3216 (sum_nr_running > group_capacity || __group_imb)) {
3217 max_load = avg_load;
3219 busiest_nr_running = sum_nr_running;
3220 busiest_load_per_task = sum_weighted_load;
3221 group_imb = __group_imb;
3224 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3226 * Busy processors will not participate in power savings
3229 if (idle == CPU_NOT_IDLE ||
3230 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3234 * If the local group is idle or completely loaded
3235 * no need to do power savings balance at this domain
3237 if (local_group && (this_nr_running >= group_capacity ||
3239 power_savings_balance = 0;
3242 * If a group is already running at full capacity or idle,
3243 * don't include that group in power savings calculations
3245 if (!power_savings_balance || sum_nr_running >= group_capacity
3250 * Calculate the group which has the least non-idle load.
3251 * This is the group from where we need to pick up the load
3254 if ((sum_nr_running < min_nr_running) ||
3255 (sum_nr_running == min_nr_running &&
3256 first_cpu(group->cpumask) <
3257 first_cpu(group_min->cpumask))) {
3259 min_nr_running = sum_nr_running;
3260 min_load_per_task = sum_weighted_load /
3265 * Calculate the group which is almost near its
3266 * capacity but still has some space to pick up some load
3267 * from other group and save more power
3269 if (sum_nr_running <= group_capacity - 1) {
3270 if (sum_nr_running > leader_nr_running ||
3271 (sum_nr_running == leader_nr_running &&
3272 first_cpu(group->cpumask) >
3273 first_cpu(group_leader->cpumask))) {
3274 group_leader = group;
3275 leader_nr_running = sum_nr_running;
3280 group = group->next;
3281 } while (group != sd->groups);
3283 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3286 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3288 if (this_load >= avg_load ||
3289 100*max_load <= sd->imbalance_pct*this_load)
3292 busiest_load_per_task /= busiest_nr_running;
3294 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3297 * We're trying to get all the cpus to the average_load, so we don't
3298 * want to push ourselves above the average load, nor do we wish to
3299 * reduce the max loaded cpu below the average load, as either of these
3300 * actions would just result in more rebalancing later, and ping-pong
3301 * tasks around. Thus we look for the minimum possible imbalance.
3302 * Negative imbalances (*we* are more loaded than anyone else) will
3303 * be counted as no imbalance for these purposes -- we can't fix that
3304 * by pulling tasks to us. Be careful of negative numbers as they'll
3305 * appear as very large values with unsigned longs.
3307 if (max_load <= busiest_load_per_task)
3311 * In the presence of smp nice balancing, certain scenarios can have
3312 * max load less than avg load(as we skip the groups at or below
3313 * its cpu_power, while calculating max_load..)
3315 if (max_load < avg_load) {
3317 goto small_imbalance;
3320 /* Don't want to pull so many tasks that a group would go idle */
3321 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3323 /* How much load to actually move to equalise the imbalance */
3324 *imbalance = min(max_pull * busiest->__cpu_power,
3325 (avg_load - this_load) * this->__cpu_power)
3329 * if *imbalance is less than the average load per runnable task
3330 * there is no gaurantee that any tasks will be moved so we'll have
3331 * a think about bumping its value to force at least one task to be
3334 if (*imbalance < busiest_load_per_task) {
3335 unsigned long tmp, pwr_now, pwr_move;
3339 pwr_move = pwr_now = 0;
3341 if (this_nr_running) {
3342 this_load_per_task /= this_nr_running;
3343 if (busiest_load_per_task > this_load_per_task)
3346 this_load_per_task = cpu_avg_load_per_task(this_cpu);
3348 if (max_load - this_load + busiest_load_per_task >=
3349 busiest_load_per_task * imbn) {
3350 *imbalance = busiest_load_per_task;
3355 * OK, we don't have enough imbalance to justify moving tasks,
3356 * however we may be able to increase total CPU power used by
3360 pwr_now += busiest->__cpu_power *
3361 min(busiest_load_per_task, max_load);
3362 pwr_now += this->__cpu_power *
3363 min(this_load_per_task, this_load);
3364 pwr_now /= SCHED_LOAD_SCALE;
3366 /* Amount of load we'd subtract */
3367 tmp = sg_div_cpu_power(busiest,
3368 busiest_load_per_task * SCHED_LOAD_SCALE);
3370 pwr_move += busiest->__cpu_power *
3371 min(busiest_load_per_task, max_load - tmp);
3373 /* Amount of load we'd add */
3374 if (max_load * busiest->__cpu_power <
3375 busiest_load_per_task * SCHED_LOAD_SCALE)
3376 tmp = sg_div_cpu_power(this,
3377 max_load * busiest->__cpu_power);
3379 tmp = sg_div_cpu_power(this,
3380 busiest_load_per_task * SCHED_LOAD_SCALE);
3381 pwr_move += this->__cpu_power *
3382 min(this_load_per_task, this_load + tmp);
3383 pwr_move /= SCHED_LOAD_SCALE;
3385 /* Move if we gain throughput */
3386 if (pwr_move > pwr_now)
3387 *imbalance = busiest_load_per_task;
3393 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3394 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3397 if (this == group_leader && group_leader != group_min) {
3398 *imbalance = min_load_per_task;
3408 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3411 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3412 unsigned long imbalance, const cpumask_t *cpus)
3414 struct rq *busiest = NULL, *rq;
3415 unsigned long max_load = 0;
3418 for_each_cpu_mask_nr(i, group->cpumask) {
3421 if (!cpu_isset(i, *cpus))
3425 wl = weighted_cpuload(i);
3427 if (rq->nr_running == 1 && wl > imbalance)
3430 if (wl > max_load) {
3440 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3441 * so long as it is large enough.
3443 #define MAX_PINNED_INTERVAL 512
3446 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3447 * tasks if there is an imbalance.
3449 static int load_balance(int this_cpu, struct rq *this_rq,
3450 struct sched_domain *sd, enum cpu_idle_type idle,
3451 int *balance, cpumask_t *cpus)
3453 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3454 struct sched_group *group;
3455 unsigned long imbalance;
3457 unsigned long flags;
3462 * When power savings policy is enabled for the parent domain, idle
3463 * sibling can pick up load irrespective of busy siblings. In this case,
3464 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3465 * portraying it as CPU_NOT_IDLE.
3467 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3468 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3471 schedstat_inc(sd, lb_count[idle]);
3475 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3482 schedstat_inc(sd, lb_nobusyg[idle]);
3486 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3488 schedstat_inc(sd, lb_nobusyq[idle]);
3492 BUG_ON(busiest == this_rq);
3494 schedstat_add(sd, lb_imbalance[idle], imbalance);
3497 if (busiest->nr_running > 1) {
3499 * Attempt to move tasks. If find_busiest_group has found
3500 * an imbalance but busiest->nr_running <= 1, the group is
3501 * still unbalanced. ld_moved simply stays zero, so it is
3502 * correctly treated as an imbalance.
3504 local_irq_save(flags);
3505 double_rq_lock(this_rq, busiest);
3506 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3507 imbalance, sd, idle, &all_pinned);
3508 double_rq_unlock(this_rq, busiest);
3509 local_irq_restore(flags);
3512 * some other cpu did the load balance for us.
3514 if (ld_moved && this_cpu != smp_processor_id())
3515 resched_cpu(this_cpu);
3517 /* All tasks on this runqueue were pinned by CPU affinity */
3518 if (unlikely(all_pinned)) {
3519 cpu_clear(cpu_of(busiest), *cpus);
3520 if (!cpus_empty(*cpus))
3527 schedstat_inc(sd, lb_failed[idle]);
3528 sd->nr_balance_failed++;
3530 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3532 spin_lock_irqsave(&busiest->lock, flags);
3534 /* don't kick the migration_thread, if the curr
3535 * task on busiest cpu can't be moved to this_cpu
3537 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3538 spin_unlock_irqrestore(&busiest->lock, flags);
3540 goto out_one_pinned;
3543 if (!busiest->active_balance) {
3544 busiest->active_balance = 1;
3545 busiest->push_cpu = this_cpu;
3548 spin_unlock_irqrestore(&busiest->lock, flags);
3550 wake_up_process(busiest->migration_thread);
3553 * We've kicked active balancing, reset the failure
3556 sd->nr_balance_failed = sd->cache_nice_tries+1;
3559 sd->nr_balance_failed = 0;
3561 if (likely(!active_balance)) {
3562 /* We were unbalanced, so reset the balancing interval */
3563 sd->balance_interval = sd->min_interval;
3566 * If we've begun active balancing, start to back off. This
3567 * case may not be covered by the all_pinned logic if there
3568 * is only 1 task on the busy runqueue (because we don't call
3571 if (sd->balance_interval < sd->max_interval)
3572 sd->balance_interval *= 2;
3575 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3576 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3582 schedstat_inc(sd, lb_balanced[idle]);
3584 sd->nr_balance_failed = 0;
3587 /* tune up the balancing interval */
3588 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3589 (sd->balance_interval < sd->max_interval))
3590 sd->balance_interval *= 2;
3592 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3593 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3604 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3605 * tasks if there is an imbalance.
3607 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3608 * this_rq is locked.
3611 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3614 struct sched_group *group;
3615 struct rq *busiest = NULL;
3616 unsigned long imbalance;
3624 * When power savings policy is enabled for the parent domain, idle
3625 * sibling can pick up load irrespective of busy siblings. In this case,
3626 * let the state of idle sibling percolate up as IDLE, instead of
3627 * portraying it as CPU_NOT_IDLE.
3629 if (sd->flags & SD_SHARE_CPUPOWER &&
3630 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3633 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3635 update_shares_locked(this_rq, sd);
3636 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3637 &sd_idle, cpus, NULL);
3639 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3643 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3645 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3649 BUG_ON(busiest == this_rq);
3651 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3654 if (busiest->nr_running > 1) {
3655 /* Attempt to move tasks */
3656 double_lock_balance(this_rq, busiest);
3657 /* this_rq->clock is already updated */
3658 update_rq_clock(busiest);
3659 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3660 imbalance, sd, CPU_NEWLY_IDLE,
3662 double_unlock_balance(this_rq, busiest);
3664 if (unlikely(all_pinned)) {
3665 cpu_clear(cpu_of(busiest), *cpus);
3666 if (!cpus_empty(*cpus))
3672 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3673 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3674 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3677 sd->nr_balance_failed = 0;
3679 update_shares_locked(this_rq, sd);
3683 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3684 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3685 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3687 sd->nr_balance_failed = 0;
3693 * idle_balance is called by schedule() if this_cpu is about to become
3694 * idle. Attempts to pull tasks from other CPUs.
3696 static void idle_balance(int this_cpu, struct rq *this_rq)
3698 struct sched_domain *sd;
3699 int pulled_task = 0;
3700 unsigned long next_balance = jiffies + HZ;
3703 for_each_domain(this_cpu, sd) {
3704 unsigned long interval;
3706 if (!(sd->flags & SD_LOAD_BALANCE))
3709 if (sd->flags & SD_BALANCE_NEWIDLE)
3710 /* If we've pulled tasks over stop searching: */
3711 pulled_task = load_balance_newidle(this_cpu, this_rq,
3714 interval = msecs_to_jiffies(sd->balance_interval);
3715 if (time_after(next_balance, sd->last_balance + interval))
3716 next_balance = sd->last_balance + interval;
3720 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3722 * We are going idle. next_balance may be set based on
3723 * a busy processor. So reset next_balance.
3725 this_rq->next_balance = next_balance;
3730 * active_load_balance is run by migration threads. It pushes running tasks
3731 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3732 * running on each physical CPU where possible, and avoids physical /
3733 * logical imbalances.
3735 * Called with busiest_rq locked.
3737 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3739 int target_cpu = busiest_rq->push_cpu;
3740 struct sched_domain *sd;
3741 struct rq *target_rq;
3743 /* Is there any task to move? */
3744 if (busiest_rq->nr_running <= 1)
3747 target_rq = cpu_rq(target_cpu);
3750 * This condition is "impossible", if it occurs
3751 * we need to fix it. Originally reported by
3752 * Bjorn Helgaas on a 128-cpu setup.
3754 BUG_ON(busiest_rq == target_rq);
3756 /* move a task from busiest_rq to target_rq */
3757 double_lock_balance(busiest_rq, target_rq);
3758 update_rq_clock(busiest_rq);
3759 update_rq_clock(target_rq);
3761 /* Search for an sd spanning us and the target CPU. */
3762 for_each_domain(target_cpu, sd) {
3763 if ((sd->flags & SD_LOAD_BALANCE) &&
3764 cpu_isset(busiest_cpu, sd->span))
3769 schedstat_inc(sd, alb_count);
3771 if (move_one_task(target_rq, target_cpu, busiest_rq,
3773 schedstat_inc(sd, alb_pushed);
3775 schedstat_inc(sd, alb_failed);
3777 double_unlock_balance(busiest_rq, target_rq);
3782 atomic_t load_balancer;
3784 } nohz ____cacheline_aligned = {
3785 .load_balancer = ATOMIC_INIT(-1),
3786 .cpu_mask = CPU_MASK_NONE,
3790 * This routine will try to nominate the ilb (idle load balancing)
3791 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3792 * load balancing on behalf of all those cpus. If all the cpus in the system
3793 * go into this tickless mode, then there will be no ilb owner (as there is
3794 * no need for one) and all the cpus will sleep till the next wakeup event
3797 * For the ilb owner, tick is not stopped. And this tick will be used
3798 * for idle load balancing. ilb owner will still be part of
3801 * While stopping the tick, this cpu will become the ilb owner if there
3802 * is no other owner. And will be the owner till that cpu becomes busy
3803 * or if all cpus in the system stop their ticks at which point
3804 * there is no need for ilb owner.
3806 * When the ilb owner becomes busy, it nominates another owner, during the
3807 * next busy scheduler_tick()
3809 int select_nohz_load_balancer(int stop_tick)
3811 int cpu = smp_processor_id();
3814 cpu_set(cpu, nohz.cpu_mask);
3815 cpu_rq(cpu)->in_nohz_recently = 1;
3818 * If we are going offline and still the leader, give up!
3820 if (!cpu_active(cpu) &&
3821 atomic_read(&nohz.load_balancer) == cpu) {
3822 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3827 /* time for ilb owner also to sleep */
3828 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3829 if (atomic_read(&nohz.load_balancer) == cpu)
3830 atomic_set(&nohz.load_balancer, -1);
3834 if (atomic_read(&nohz.load_balancer) == -1) {
3835 /* make me the ilb owner */
3836 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3838 } else if (atomic_read(&nohz.load_balancer) == cpu)
3841 if (!cpu_isset(cpu, nohz.cpu_mask))
3844 cpu_clear(cpu, nohz.cpu_mask);
3846 if (atomic_read(&nohz.load_balancer) == cpu)
3847 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3854 static DEFINE_SPINLOCK(balancing);
3857 * It checks each scheduling domain to see if it is due to be balanced,
3858 * and initiates a balancing operation if so.
3860 * Balancing parameters are set up in arch_init_sched_domains.
3862 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3865 struct rq *rq = cpu_rq(cpu);
3866 unsigned long interval;
3867 struct sched_domain *sd;
3868 /* Earliest time when we have to do rebalance again */
3869 unsigned long next_balance = jiffies + 60*HZ;
3870 int update_next_balance = 0;
3874 for_each_domain(cpu, sd) {
3875 if (!(sd->flags & SD_LOAD_BALANCE))
3878 interval = sd->balance_interval;
3879 if (idle != CPU_IDLE)
3880 interval *= sd->busy_factor;
3882 /* scale ms to jiffies */
3883 interval = msecs_to_jiffies(interval);
3884 if (unlikely(!interval))
3886 if (interval > HZ*NR_CPUS/10)
3887 interval = HZ*NR_CPUS/10;
3889 need_serialize = sd->flags & SD_SERIALIZE;
3891 if (need_serialize) {
3892 if (!spin_trylock(&balancing))
3896 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3897 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3899 * We've pulled tasks over so either we're no
3900 * longer idle, or one of our SMT siblings is
3903 idle = CPU_NOT_IDLE;
3905 sd->last_balance = jiffies;
3908 spin_unlock(&balancing);
3910 if (time_after(next_balance, sd->last_balance + interval)) {
3911 next_balance = sd->last_balance + interval;
3912 update_next_balance = 1;
3916 * Stop the load balance at this level. There is another
3917 * CPU in our sched group which is doing load balancing more
3925 * next_balance will be updated only when there is a need.
3926 * When the cpu is attached to null domain for ex, it will not be
3929 if (likely(update_next_balance))
3930 rq->next_balance = next_balance;
3934 * run_rebalance_domains is triggered when needed from the scheduler tick.
3935 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3936 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3938 static void run_rebalance_domains(struct softirq_action *h)
3940 int this_cpu = smp_processor_id();
3941 struct rq *this_rq = cpu_rq(this_cpu);
3942 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3943 CPU_IDLE : CPU_NOT_IDLE;
3945 rebalance_domains(this_cpu, idle);
3949 * If this cpu is the owner for idle load balancing, then do the
3950 * balancing on behalf of the other idle cpus whose ticks are
3953 if (this_rq->idle_at_tick &&
3954 atomic_read(&nohz.load_balancer) == this_cpu) {
3955 cpumask_t cpus = nohz.cpu_mask;
3959 cpu_clear(this_cpu, cpus);
3960 for_each_cpu_mask_nr(balance_cpu, cpus) {
3962 * If this cpu gets work to do, stop the load balancing
3963 * work being done for other cpus. Next load
3964 * balancing owner will pick it up.
3969 rebalance_domains(balance_cpu, CPU_IDLE);
3971 rq = cpu_rq(balance_cpu);
3972 if (time_after(this_rq->next_balance, rq->next_balance))
3973 this_rq->next_balance = rq->next_balance;
3980 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3982 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3983 * idle load balancing owner or decide to stop the periodic load balancing,
3984 * if the whole system is idle.
3986 static inline void trigger_load_balance(struct rq *rq, int cpu)
3990 * If we were in the nohz mode recently and busy at the current
3991 * scheduler tick, then check if we need to nominate new idle
3994 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3995 rq->in_nohz_recently = 0;
3997 if (atomic_read(&nohz.load_balancer) == cpu) {
3998 cpu_clear(cpu, nohz.cpu_mask);
3999 atomic_set(&nohz.load_balancer, -1);
4002 if (atomic_read(&nohz.load_balancer) == -1) {
4004 * simple selection for now: Nominate the
4005 * first cpu in the nohz list to be the next
4008 * TBD: Traverse the sched domains and nominate
4009 * the nearest cpu in the nohz.cpu_mask.
4011 int ilb = first_cpu(nohz.cpu_mask);
4013 if (ilb < nr_cpu_ids)
4019 * If this cpu is idle and doing idle load balancing for all the
4020 * cpus with ticks stopped, is it time for that to stop?
4022 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4023 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4029 * If this cpu is idle and the idle load balancing is done by
4030 * someone else, then no need raise the SCHED_SOFTIRQ
4032 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4033 cpu_isset(cpu, nohz.cpu_mask))
4036 if (time_after_eq(jiffies, rq->next_balance))
4037 raise_softirq(SCHED_SOFTIRQ);
4040 #else /* CONFIG_SMP */
4043 * on UP we do not need to balance between CPUs:
4045 static inline void idle_balance(int cpu, struct rq *rq)
4051 DEFINE_PER_CPU(struct kernel_stat, kstat);
4053 EXPORT_PER_CPU_SYMBOL(kstat);
4056 * Return any ns on the sched_clock that have not yet been banked in
4057 * @p in case that task is currently running.
4059 unsigned long long task_delta_exec(struct task_struct *p)
4061 unsigned long flags;
4065 rq = task_rq_lock(p, &flags);
4067 if (task_current(rq, p)) {
4070 update_rq_clock(rq);
4071 delta_exec = rq->clock - p->se.exec_start;
4072 if ((s64)delta_exec > 0)
4076 task_rq_unlock(rq, &flags);
4082 * Account user cpu time to a process.
4083 * @p: the process that the cpu time gets accounted to
4084 * @cputime: the cpu time spent in user space since the last update
4086 void account_user_time(struct task_struct *p, cputime_t cputime)
4088 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4091 p->utime = cputime_add(p->utime, cputime);
4092 account_group_user_time(p, cputime);
4094 /* Add user time to cpustat. */
4095 tmp = cputime_to_cputime64(cputime);
4096 if (TASK_NICE(p) > 0)
4097 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4099 cpustat->user = cputime64_add(cpustat->user, tmp);
4100 /* Account for user time used */
4101 acct_update_integrals(p);
4105 * Account guest cpu time to a process.
4106 * @p: the process that the cpu time gets accounted to
4107 * @cputime: the cpu time spent in virtual machine since the last update
4109 static void account_guest_time(struct task_struct *p, cputime_t cputime)
4112 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4114 tmp = cputime_to_cputime64(cputime);
4116 p->utime = cputime_add(p->utime, cputime);
4117 account_group_user_time(p, cputime);
4118 p->gtime = cputime_add(p->gtime, cputime);
4120 cpustat->user = cputime64_add(cpustat->user, tmp);
4121 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4125 * Account scaled user cpu time to a process.
4126 * @p: the process that the cpu time gets accounted to
4127 * @cputime: the cpu time spent in user space since the last update
4129 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4131 p->utimescaled = cputime_add(p->utimescaled, cputime);
4135 * Account system cpu time to a process.
4136 * @p: the process that the cpu time gets accounted to
4137 * @hardirq_offset: the offset to subtract from hardirq_count()
4138 * @cputime: the cpu time spent in kernel space since the last update
4140 void account_system_time(struct task_struct *p, int hardirq_offset,
4143 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4144 struct rq *rq = this_rq();
4147 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4148 account_guest_time(p, cputime);
4152 p->stime = cputime_add(p->stime, cputime);
4153 account_group_system_time(p, cputime);
4155 /* Add system time to cpustat. */
4156 tmp = cputime_to_cputime64(cputime);
4157 if (hardirq_count() - hardirq_offset)
4158 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4159 else if (softirq_count())
4160 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4161 else if (p != rq->idle)
4162 cpustat->system = cputime64_add(cpustat->system, tmp);
4163 else if (atomic_read(&rq->nr_iowait) > 0)
4164 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4166 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4167 /* Account for system time used */
4168 acct_update_integrals(p);
4172 * Account scaled system cpu time to a process.
4173 * @p: the process that the cpu time gets accounted to
4174 * @hardirq_offset: the offset to subtract from hardirq_count()
4175 * @cputime: the cpu time spent in kernel space since the last update
4177 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4179 p->stimescaled = cputime_add(p->stimescaled, cputime);
4183 * Account for involuntary wait time.
4184 * @p: the process from which the cpu time has been stolen
4185 * @steal: the cpu time spent in involuntary wait
4187 void account_steal_time(struct task_struct *p, cputime_t steal)
4189 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4190 cputime64_t tmp = cputime_to_cputime64(steal);
4191 struct rq *rq = this_rq();
4193 if (p == rq->idle) {
4194 p->stime = cputime_add(p->stime, steal);
4195 if (atomic_read(&rq->nr_iowait) > 0)
4196 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4198 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4200 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4204 * Use precise platform statistics if available:
4206 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4207 cputime_t task_utime(struct task_struct *p)
4212 cputime_t task_stime(struct task_struct *p)
4217 cputime_t task_utime(struct task_struct *p)
4219 clock_t utime = cputime_to_clock_t(p->utime),
4220 total = utime + cputime_to_clock_t(p->stime);
4224 * Use CFS's precise accounting:
4226 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
4230 do_div(temp, total);
4232 utime = (clock_t)temp;
4234 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
4235 return p->prev_utime;
4238 cputime_t task_stime(struct task_struct *p)
4243 * Use CFS's precise accounting. (we subtract utime from
4244 * the total, to make sure the total observed by userspace
4245 * grows monotonically - apps rely on that):
4247 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
4248 cputime_to_clock_t(task_utime(p));
4251 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
4253 return p->prev_stime;
4257 inline cputime_t task_gtime(struct task_struct *p)
4263 * This function gets called by the timer code, with HZ frequency.
4264 * We call it with interrupts disabled.
4266 * It also gets called by the fork code, when changing the parent's
4269 void scheduler_tick(void)
4271 int cpu = smp_processor_id();
4272 struct rq *rq = cpu_rq(cpu);
4273 struct task_struct *curr = rq->curr;
4277 spin_lock(&rq->lock);
4278 update_rq_clock(rq);
4279 update_cpu_load(rq);
4280 curr->sched_class->task_tick(rq, curr, 0);
4281 spin_unlock(&rq->lock);
4284 rq->idle_at_tick = idle_cpu(cpu);
4285 trigger_load_balance(rq, cpu);
4289 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4290 defined(CONFIG_PREEMPT_TRACER))
4292 static inline unsigned long get_parent_ip(unsigned long addr)
4294 if (in_lock_functions(addr)) {
4295 addr = CALLER_ADDR2;
4296 if (in_lock_functions(addr))
4297 addr = CALLER_ADDR3;
4302 void __kprobes add_preempt_count(int val)
4304 #ifdef CONFIG_DEBUG_PREEMPT
4308 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4311 preempt_count() += val;
4312 #ifdef CONFIG_DEBUG_PREEMPT
4314 * Spinlock count overflowing soon?
4316 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4319 if (preempt_count() == val)
4320 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4322 EXPORT_SYMBOL(add_preempt_count);
4324 void __kprobes sub_preempt_count(int val)
4326 #ifdef CONFIG_DEBUG_PREEMPT
4330 if (DEBUG_LOCKS_WARN_ON(val > preempt_count() - (!!kernel_locked())))
4333 * Is the spinlock portion underflowing?
4335 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4336 !(preempt_count() & PREEMPT_MASK)))
4340 if (preempt_count() == val)
4341 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4342 preempt_count() -= val;
4344 EXPORT_SYMBOL(sub_preempt_count);
4349 * Print scheduling while atomic bug:
4351 static noinline void __schedule_bug(struct task_struct *prev)
4353 struct pt_regs *regs = get_irq_regs();
4355 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4356 prev->comm, prev->pid, preempt_count());
4358 debug_show_held_locks(prev);
4360 if (irqs_disabled())
4361 print_irqtrace_events(prev);
4370 * Various schedule()-time debugging checks and statistics:
4372 static inline void schedule_debug(struct task_struct *prev)
4375 * Test if we are atomic. Since do_exit() needs to call into
4376 * schedule() atomically, we ignore that path for now.
4377 * Otherwise, whine if we are scheduling when we should not be.
4379 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4380 __schedule_bug(prev);
4382 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4384 schedstat_inc(this_rq(), sched_count);
4385 #ifdef CONFIG_SCHEDSTATS
4386 if (unlikely(prev->lock_depth >= 0)) {
4387 schedstat_inc(this_rq(), bkl_count);
4388 schedstat_inc(prev, sched_info.bkl_count);
4394 * Pick up the highest-prio task:
4396 static inline struct task_struct *
4397 pick_next_task(struct rq *rq, struct task_struct *prev)
4399 const struct sched_class *class;
4400 struct task_struct *p;
4403 * Optimization: we know that if all tasks are in
4404 * the fair class we can call that function directly:
4406 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4407 p = fair_sched_class.pick_next_task(rq);
4412 class = sched_class_highest;
4414 p = class->pick_next_task(rq);
4418 * Will never be NULL as the idle class always
4419 * returns a non-NULL p:
4421 class = class->next;
4426 * schedule() is the main scheduler function.
4428 asmlinkage void __sched schedule(void)
4430 struct task_struct *prev, *next;
4431 unsigned long *switch_count;
4437 cpu = smp_processor_id();
4441 switch_count = &prev->nivcsw;
4443 release_kernel_lock(prev);
4444 need_resched_nonpreemptible:
4446 schedule_debug(prev);
4448 if (sched_feat(HRTICK))
4451 spin_lock_irq(&rq->lock);
4452 update_rq_clock(rq);
4453 clear_tsk_need_resched(prev);
4455 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4456 if (unlikely(signal_pending_state(prev->state, prev)))
4457 prev->state = TASK_RUNNING;
4459 deactivate_task(rq, prev, 1);
4460 switch_count = &prev->nvcsw;
4464 if (prev->sched_class->pre_schedule)
4465 prev->sched_class->pre_schedule(rq, prev);
4468 if (unlikely(!rq->nr_running))
4469 idle_balance(cpu, rq);
4471 prev->sched_class->put_prev_task(rq, prev);
4472 next = pick_next_task(rq, prev);
4474 if (likely(prev != next)) {
4475 sched_info_switch(prev, next);
4481 context_switch(rq, prev, next); /* unlocks the rq */
4483 * the context switch might have flipped the stack from under
4484 * us, hence refresh the local variables.
4486 cpu = smp_processor_id();
4489 spin_unlock_irq(&rq->lock);
4491 if (unlikely(reacquire_kernel_lock(current) < 0))
4492 goto need_resched_nonpreemptible;
4494 preempt_enable_no_resched();
4495 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4498 EXPORT_SYMBOL(schedule);
4500 #ifdef CONFIG_PREEMPT
4502 * this is the entry point to schedule() from in-kernel preemption
4503 * off of preempt_enable. Kernel preemptions off return from interrupt
4504 * occur there and call schedule directly.
4506 asmlinkage void __sched preempt_schedule(void)
4508 struct thread_info *ti = current_thread_info();
4511 * If there is a non-zero preempt_count or interrupts are disabled,
4512 * we do not want to preempt the current task. Just return..
4514 if (likely(ti->preempt_count || irqs_disabled()))
4518 add_preempt_count(PREEMPT_ACTIVE);
4520 sub_preempt_count(PREEMPT_ACTIVE);
4523 * Check again in case we missed a preemption opportunity
4524 * between schedule and now.
4527 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4529 EXPORT_SYMBOL(preempt_schedule);
4532 * this is the entry point to schedule() from kernel preemption
4533 * off of irq context.
4534 * Note, that this is called and return with irqs disabled. This will
4535 * protect us against recursive calling from irq.
4537 asmlinkage void __sched preempt_schedule_irq(void)
4539 struct thread_info *ti = current_thread_info();
4541 /* Catch callers which need to be fixed */
4542 BUG_ON(ti->preempt_count || !irqs_disabled());
4545 add_preempt_count(PREEMPT_ACTIVE);
4548 local_irq_disable();
4549 sub_preempt_count(PREEMPT_ACTIVE);
4552 * Check again in case we missed a preemption opportunity
4553 * between schedule and now.
4556 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4559 #endif /* CONFIG_PREEMPT */
4561 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4564 return try_to_wake_up(curr->private, mode, sync);
4566 EXPORT_SYMBOL(default_wake_function);
4569 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4570 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4571 * number) then we wake all the non-exclusive tasks and one exclusive task.
4573 * There are circumstances in which we can try to wake a task which has already
4574 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4575 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4577 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4578 int nr_exclusive, int sync, void *key)
4580 wait_queue_t *curr, *next;
4582 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4583 unsigned flags = curr->flags;
4585 if (curr->func(curr, mode, sync, key) &&
4586 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4592 * __wake_up - wake up threads blocked on a waitqueue.
4594 * @mode: which threads
4595 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4596 * @key: is directly passed to the wakeup function
4598 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4599 int nr_exclusive, void *key)
4601 unsigned long flags;
4603 spin_lock_irqsave(&q->lock, flags);
4604 __wake_up_common(q, mode, nr_exclusive, 0, key);
4605 spin_unlock_irqrestore(&q->lock, flags);
4607 EXPORT_SYMBOL(__wake_up);
4610 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4612 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4614 __wake_up_common(q, mode, 1, 0, NULL);
4618 * __wake_up_sync - wake up threads blocked on a waitqueue.
4620 * @mode: which threads
4621 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4623 * The sync wakeup differs that the waker knows that it will schedule
4624 * away soon, so while the target thread will be woken up, it will not
4625 * be migrated to another CPU - ie. the two threads are 'synchronized'
4626 * with each other. This can prevent needless bouncing between CPUs.
4628 * On UP it can prevent extra preemption.
4631 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4633 unsigned long flags;
4639 if (unlikely(!nr_exclusive))
4642 spin_lock_irqsave(&q->lock, flags);
4643 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4644 spin_unlock_irqrestore(&q->lock, flags);
4646 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4649 * complete: - signals a single thread waiting on this completion
4650 * @x: holds the state of this particular completion
4652 * This will wake up a single thread waiting on this completion. Threads will be
4653 * awakened in the same order in which they were queued.
4655 * See also complete_all(), wait_for_completion() and related routines.
4657 void complete(struct completion *x)
4659 unsigned long flags;
4661 spin_lock_irqsave(&x->wait.lock, flags);
4663 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4664 spin_unlock_irqrestore(&x->wait.lock, flags);
4666 EXPORT_SYMBOL(complete);
4669 * complete_all: - signals all threads waiting on this completion
4670 * @x: holds the state of this particular completion
4672 * This will wake up all threads waiting on this particular completion event.
4674 void complete_all(struct completion *x)
4676 unsigned long flags;
4678 spin_lock_irqsave(&x->wait.lock, flags);
4679 x->done += UINT_MAX/2;
4680 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4681 spin_unlock_irqrestore(&x->wait.lock, flags);
4683 EXPORT_SYMBOL(complete_all);
4685 static inline long __sched
4686 do_wait_for_common(struct completion *x, long timeout, int state)
4689 DECLARE_WAITQUEUE(wait, current);
4691 wait.flags |= WQ_FLAG_EXCLUSIVE;
4692 __add_wait_queue_tail(&x->wait, &wait);
4694 if (signal_pending_state(state, current)) {
4695 timeout = -ERESTARTSYS;
4698 __set_current_state(state);
4699 spin_unlock_irq(&x->wait.lock);
4700 timeout = schedule_timeout(timeout);
4701 spin_lock_irq(&x->wait.lock);
4702 } while (!x->done && timeout);
4703 __remove_wait_queue(&x->wait, &wait);
4708 return timeout ?: 1;
4712 wait_for_common(struct completion *x, long timeout, int state)
4716 spin_lock_irq(&x->wait.lock);
4717 timeout = do_wait_for_common(x, timeout, state);
4718 spin_unlock_irq(&x->wait.lock);
4723 * wait_for_completion: - waits for completion of a task
4724 * @x: holds the state of this particular completion
4726 * This waits to be signaled for completion of a specific task. It is NOT
4727 * interruptible and there is no timeout.
4729 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4730 * and interrupt capability. Also see complete().
4732 void __sched wait_for_completion(struct completion *x)
4734 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4736 EXPORT_SYMBOL(wait_for_completion);
4739 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4740 * @x: holds the state of this particular completion
4741 * @timeout: timeout value in jiffies
4743 * This waits for either a completion of a specific task to be signaled or for a
4744 * specified timeout to expire. The timeout is in jiffies. It is not
4747 unsigned long __sched
4748 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4750 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4752 EXPORT_SYMBOL(wait_for_completion_timeout);
4755 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4756 * @x: holds the state of this particular completion
4758 * This waits for completion of a specific task to be signaled. It is
4761 int __sched wait_for_completion_interruptible(struct completion *x)
4763 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4764 if (t == -ERESTARTSYS)
4768 EXPORT_SYMBOL(wait_for_completion_interruptible);
4771 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4772 * @x: holds the state of this particular completion
4773 * @timeout: timeout value in jiffies
4775 * This waits for either a completion of a specific task to be signaled or for a
4776 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4778 unsigned long __sched
4779 wait_for_completion_interruptible_timeout(struct completion *x,
4780 unsigned long timeout)
4782 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4784 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4787 * wait_for_completion_killable: - waits for completion of a task (killable)
4788 * @x: holds the state of this particular completion
4790 * This waits to be signaled for completion of a specific task. It can be
4791 * interrupted by a kill signal.
4793 int __sched wait_for_completion_killable(struct completion *x)
4795 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4796 if (t == -ERESTARTSYS)
4800 EXPORT_SYMBOL(wait_for_completion_killable);
4803 * try_wait_for_completion - try to decrement a completion without blocking
4804 * @x: completion structure
4806 * Returns: 0 if a decrement cannot be done without blocking
4807 * 1 if a decrement succeeded.
4809 * If a completion is being used as a counting completion,
4810 * attempt to decrement the counter without blocking. This
4811 * enables us to avoid waiting if the resource the completion
4812 * is protecting is not available.
4814 bool try_wait_for_completion(struct completion *x)
4818 spin_lock_irq(&x->wait.lock);
4823 spin_unlock_irq(&x->wait.lock);
4826 EXPORT_SYMBOL(try_wait_for_completion);
4829 * completion_done - Test to see if a completion has any waiters
4830 * @x: completion structure
4832 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4833 * 1 if there are no waiters.
4836 bool completion_done(struct completion *x)
4840 spin_lock_irq(&x->wait.lock);
4843 spin_unlock_irq(&x->wait.lock);
4846 EXPORT_SYMBOL(completion_done);
4849 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4851 unsigned long flags;
4854 init_waitqueue_entry(&wait, current);
4856 __set_current_state(state);
4858 spin_lock_irqsave(&q->lock, flags);
4859 __add_wait_queue(q, &wait);
4860 spin_unlock(&q->lock);
4861 timeout = schedule_timeout(timeout);
4862 spin_lock_irq(&q->lock);
4863 __remove_wait_queue(q, &wait);
4864 spin_unlock_irqrestore(&q->lock, flags);
4869 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4871 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4873 EXPORT_SYMBOL(interruptible_sleep_on);
4876 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4878 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4880 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4882 void __sched sleep_on(wait_queue_head_t *q)
4884 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4886 EXPORT_SYMBOL(sleep_on);
4888 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4890 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4892 EXPORT_SYMBOL(sleep_on_timeout);
4894 #ifdef CONFIG_RT_MUTEXES
4897 * rt_mutex_setprio - set the current priority of a task
4899 * @prio: prio value (kernel-internal form)
4901 * This function changes the 'effective' priority of a task. It does
4902 * not touch ->normal_prio like __setscheduler().
4904 * Used by the rt_mutex code to implement priority inheritance logic.
4906 void rt_mutex_setprio(struct task_struct *p, int prio)
4908 unsigned long flags;
4909 int oldprio, on_rq, running;
4911 const struct sched_class *prev_class = p->sched_class;
4913 BUG_ON(prio < 0 || prio > MAX_PRIO);
4915 rq = task_rq_lock(p, &flags);
4916 update_rq_clock(rq);
4919 on_rq = p->se.on_rq;
4920 running = task_current(rq, p);
4922 dequeue_task(rq, p, 0);
4924 p->sched_class->put_prev_task(rq, p);
4927 p->sched_class = &rt_sched_class;
4929 p->sched_class = &fair_sched_class;
4934 p->sched_class->set_curr_task(rq);
4936 enqueue_task(rq, p, 0);
4938 check_class_changed(rq, p, prev_class, oldprio, running);
4940 task_rq_unlock(rq, &flags);
4945 void set_user_nice(struct task_struct *p, long nice)
4947 int old_prio, delta, on_rq;
4948 unsigned long flags;
4951 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4954 * We have to be careful, if called from sys_setpriority(),
4955 * the task might be in the middle of scheduling on another CPU.
4957 rq = task_rq_lock(p, &flags);
4958 update_rq_clock(rq);
4960 * The RT priorities are set via sched_setscheduler(), but we still
4961 * allow the 'normal' nice value to be set - but as expected
4962 * it wont have any effect on scheduling until the task is
4963 * SCHED_FIFO/SCHED_RR:
4965 if (task_has_rt_policy(p)) {
4966 p->static_prio = NICE_TO_PRIO(nice);
4969 on_rq = p->se.on_rq;
4971 dequeue_task(rq, p, 0);
4973 p->static_prio = NICE_TO_PRIO(nice);
4976 p->prio = effective_prio(p);
4977 delta = p->prio - old_prio;
4980 enqueue_task(rq, p, 0);
4982 * If the task increased its priority or is running and
4983 * lowered its priority, then reschedule its CPU:
4985 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4986 resched_task(rq->curr);
4989 task_rq_unlock(rq, &flags);
4991 EXPORT_SYMBOL(set_user_nice);
4994 * can_nice - check if a task can reduce its nice value
4998 int can_nice(const struct task_struct *p, const int nice)
5000 /* convert nice value [19,-20] to rlimit style value [1,40] */
5001 int nice_rlim = 20 - nice;
5003 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
5004 capable(CAP_SYS_NICE));
5007 #ifdef __ARCH_WANT_SYS_NICE
5010 * sys_nice - change the priority of the current process.
5011 * @increment: priority increment
5013 * sys_setpriority is a more generic, but much slower function that
5014 * does similar things.
5016 asmlinkage long sys_nice(int increment)
5021 * Setpriority might change our priority at the same moment.
5022 * We don't have to worry. Conceptually one call occurs first
5023 * and we have a single winner.
5025 if (increment < -40)
5030 nice = PRIO_TO_NICE(current->static_prio) + increment;
5036 if (increment < 0 && !can_nice(current, nice))
5039 retval = security_task_setnice(current, nice);
5043 set_user_nice(current, nice);
5050 * task_prio - return the priority value of a given task.
5051 * @p: the task in question.
5053 * This is the priority value as seen by users in /proc.
5054 * RT tasks are offset by -200. Normal tasks are centered
5055 * around 0, value goes from -16 to +15.
5057 int task_prio(const struct task_struct *p)
5059 return p->prio - MAX_RT_PRIO;
5063 * task_nice - return the nice value of a given task.
5064 * @p: the task in question.
5066 int task_nice(const struct task_struct *p)
5068 return TASK_NICE(p);
5070 EXPORT_SYMBOL(task_nice);
5073 * idle_cpu - is a given cpu idle currently?
5074 * @cpu: the processor in question.
5076 int idle_cpu(int cpu)
5078 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5082 * idle_task - return the idle task for a given cpu.
5083 * @cpu: the processor in question.
5085 struct task_struct *idle_task(int cpu)
5087 return cpu_rq(cpu)->idle;
5091 * find_process_by_pid - find a process with a matching PID value.
5092 * @pid: the pid in question.
5094 static struct task_struct *find_process_by_pid(pid_t pid)
5096 return pid ? find_task_by_vpid(pid) : current;
5099 /* Actually do priority change: must hold rq lock. */
5101 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5103 BUG_ON(p->se.on_rq);
5106 switch (p->policy) {
5110 p->sched_class = &fair_sched_class;
5114 p->sched_class = &rt_sched_class;
5118 p->rt_priority = prio;
5119 p->normal_prio = normal_prio(p);
5120 /* we are holding p->pi_lock already */
5121 p->prio = rt_mutex_getprio(p);
5126 * check the target process has a UID that matches the current process's
5128 static bool check_same_owner(struct task_struct *p)
5130 const struct cred *cred = current_cred(), *pcred;
5134 pcred = __task_cred(p);
5135 match = (cred->euid == pcred->euid ||
5136 cred->euid == pcred->uid);
5141 static int __sched_setscheduler(struct task_struct *p, int policy,
5142 struct sched_param *param, bool user)
5144 int retval, oldprio, oldpolicy = -1, on_rq, running;
5145 unsigned long flags;
5146 const struct sched_class *prev_class = p->sched_class;
5149 /* may grab non-irq protected spin_locks */
5150 BUG_ON(in_interrupt());
5152 /* double check policy once rq lock held */
5154 policy = oldpolicy = p->policy;
5155 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5156 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5157 policy != SCHED_IDLE)
5160 * Valid priorities for SCHED_FIFO and SCHED_RR are
5161 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5162 * SCHED_BATCH and SCHED_IDLE is 0.
5164 if (param->sched_priority < 0 ||
5165 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5166 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5168 if (rt_policy(policy) != (param->sched_priority != 0))
5172 * Allow unprivileged RT tasks to decrease priority:
5174 if (user && !capable(CAP_SYS_NICE)) {
5175 if (rt_policy(policy)) {
5176 unsigned long rlim_rtprio;
5178 if (!lock_task_sighand(p, &flags))
5180 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5181 unlock_task_sighand(p, &flags);
5183 /* can't set/change the rt policy */
5184 if (policy != p->policy && !rlim_rtprio)
5187 /* can't increase priority */
5188 if (param->sched_priority > p->rt_priority &&
5189 param->sched_priority > rlim_rtprio)
5193 * Like positive nice levels, dont allow tasks to
5194 * move out of SCHED_IDLE either:
5196 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5199 /* can't change other user's priorities */
5200 if (!check_same_owner(p))
5205 #ifdef CONFIG_RT_GROUP_SCHED
5207 * Do not allow realtime tasks into groups that have no runtime
5210 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5211 task_group(p)->rt_bandwidth.rt_runtime == 0)
5215 retval = security_task_setscheduler(p, policy, param);
5221 * make sure no PI-waiters arrive (or leave) while we are
5222 * changing the priority of the task:
5224 spin_lock_irqsave(&p->pi_lock, flags);
5226 * To be able to change p->policy safely, the apropriate
5227 * runqueue lock must be held.
5229 rq = __task_rq_lock(p);
5230 /* recheck policy now with rq lock held */
5231 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5232 policy = oldpolicy = -1;
5233 __task_rq_unlock(rq);
5234 spin_unlock_irqrestore(&p->pi_lock, flags);
5237 update_rq_clock(rq);
5238 on_rq = p->se.on_rq;
5239 running = task_current(rq, p);
5241 deactivate_task(rq, p, 0);
5243 p->sched_class->put_prev_task(rq, p);
5246 __setscheduler(rq, p, policy, param->sched_priority);
5249 p->sched_class->set_curr_task(rq);
5251 activate_task(rq, p, 0);
5253 check_class_changed(rq, p, prev_class, oldprio, running);
5255 __task_rq_unlock(rq);
5256 spin_unlock_irqrestore(&p->pi_lock, flags);
5258 rt_mutex_adjust_pi(p);
5264 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5265 * @p: the task in question.
5266 * @policy: new policy.
5267 * @param: structure containing the new RT priority.
5269 * NOTE that the task may be already dead.
5271 int sched_setscheduler(struct task_struct *p, int policy,
5272 struct sched_param *param)
5274 return __sched_setscheduler(p, policy, param, true);
5276 EXPORT_SYMBOL_GPL(sched_setscheduler);
5279 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5280 * @p: the task in question.
5281 * @policy: new policy.
5282 * @param: structure containing the new RT priority.
5284 * Just like sched_setscheduler, only don't bother checking if the
5285 * current context has permission. For example, this is needed in
5286 * stop_machine(): we create temporary high priority worker threads,
5287 * but our caller might not have that capability.
5289 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5290 struct sched_param *param)
5292 return __sched_setscheduler(p, policy, param, false);
5296 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5298 struct sched_param lparam;
5299 struct task_struct *p;
5302 if (!param || pid < 0)
5304 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5309 p = find_process_by_pid(pid);
5311 retval = sched_setscheduler(p, policy, &lparam);
5318 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5319 * @pid: the pid in question.
5320 * @policy: new policy.
5321 * @param: structure containing the new RT priority.
5324 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5326 /* negative values for policy are not valid */
5330 return do_sched_setscheduler(pid, policy, param);
5334 * sys_sched_setparam - set/change the RT priority of a thread
5335 * @pid: the pid in question.
5336 * @param: structure containing the new RT priority.
5338 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5340 return do_sched_setscheduler(pid, -1, param);
5344 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5345 * @pid: the pid in question.
5347 asmlinkage long sys_sched_getscheduler(pid_t pid)
5349 struct task_struct *p;
5356 read_lock(&tasklist_lock);
5357 p = find_process_by_pid(pid);
5359 retval = security_task_getscheduler(p);
5363 read_unlock(&tasklist_lock);
5368 * sys_sched_getscheduler - get the RT priority of a thread
5369 * @pid: the pid in question.
5370 * @param: structure containing the RT priority.
5372 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5374 struct sched_param lp;
5375 struct task_struct *p;
5378 if (!param || pid < 0)
5381 read_lock(&tasklist_lock);
5382 p = find_process_by_pid(pid);
5387 retval = security_task_getscheduler(p);
5391 lp.sched_priority = p->rt_priority;
5392 read_unlock(&tasklist_lock);
5395 * This one might sleep, we cannot do it with a spinlock held ...
5397 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5402 read_unlock(&tasklist_lock);
5406 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5408 cpumask_t cpus_allowed;
5409 cpumask_t new_mask = *in_mask;
5410 struct task_struct *p;
5414 read_lock(&tasklist_lock);
5416 p = find_process_by_pid(pid);
5418 read_unlock(&tasklist_lock);
5424 * It is not safe to call set_cpus_allowed with the
5425 * tasklist_lock held. We will bump the task_struct's
5426 * usage count and then drop tasklist_lock.
5429 read_unlock(&tasklist_lock);
5432 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
5435 retval = security_task_setscheduler(p, 0, NULL);
5439 cpuset_cpus_allowed(p, &cpus_allowed);
5440 cpus_and(new_mask, new_mask, cpus_allowed);
5442 retval = set_cpus_allowed_ptr(p, &new_mask);
5445 cpuset_cpus_allowed(p, &cpus_allowed);
5446 if (!cpus_subset(new_mask, cpus_allowed)) {
5448 * We must have raced with a concurrent cpuset
5449 * update. Just reset the cpus_allowed to the
5450 * cpuset's cpus_allowed
5452 new_mask = cpus_allowed;
5462 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5463 cpumask_t *new_mask)
5465 if (len < sizeof(cpumask_t)) {
5466 memset(new_mask, 0, sizeof(cpumask_t));
5467 } else if (len > sizeof(cpumask_t)) {
5468 len = sizeof(cpumask_t);
5470 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5474 * sys_sched_setaffinity - set the cpu affinity of a process
5475 * @pid: pid of the process
5476 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5477 * @user_mask_ptr: user-space pointer to the new cpu mask
5479 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5480 unsigned long __user *user_mask_ptr)
5485 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5489 return sched_setaffinity(pid, &new_mask);
5492 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5494 struct task_struct *p;
5498 read_lock(&tasklist_lock);
5501 p = find_process_by_pid(pid);
5505 retval = security_task_getscheduler(p);
5509 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5512 read_unlock(&tasklist_lock);
5519 * sys_sched_getaffinity - get the cpu affinity of a process
5520 * @pid: pid of the process
5521 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5522 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5524 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5525 unsigned long __user *user_mask_ptr)
5530 if (len < sizeof(cpumask_t))
5533 ret = sched_getaffinity(pid, &mask);
5537 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5540 return sizeof(cpumask_t);
5544 * sys_sched_yield - yield the current processor to other threads.
5546 * This function yields the current CPU to other tasks. If there are no
5547 * other threads running on this CPU then this function will return.
5549 asmlinkage long sys_sched_yield(void)
5551 struct rq *rq = this_rq_lock();
5553 schedstat_inc(rq, yld_count);
5554 current->sched_class->yield_task(rq);
5557 * Since we are going to call schedule() anyway, there's
5558 * no need to preempt or enable interrupts:
5560 __release(rq->lock);
5561 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5562 _raw_spin_unlock(&rq->lock);
5563 preempt_enable_no_resched();
5570 static void __cond_resched(void)
5572 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5573 __might_sleep(__FILE__, __LINE__);
5576 * The BKS might be reacquired before we have dropped
5577 * PREEMPT_ACTIVE, which could trigger a second
5578 * cond_resched() call.
5581 add_preempt_count(PREEMPT_ACTIVE);
5583 sub_preempt_count(PREEMPT_ACTIVE);
5584 } while (need_resched());
5587 int __sched _cond_resched(void)
5589 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5590 system_state == SYSTEM_RUNNING) {
5596 EXPORT_SYMBOL(_cond_resched);
5599 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5600 * call schedule, and on return reacquire the lock.
5602 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5603 * operations here to prevent schedule() from being called twice (once via
5604 * spin_unlock(), once by hand).
5606 int cond_resched_lock(spinlock_t *lock)
5608 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5611 if (spin_needbreak(lock) || resched) {
5613 if (resched && need_resched())
5622 EXPORT_SYMBOL(cond_resched_lock);
5624 int __sched cond_resched_softirq(void)
5626 BUG_ON(!in_softirq());
5628 if (need_resched() && system_state == SYSTEM_RUNNING) {
5636 EXPORT_SYMBOL(cond_resched_softirq);
5639 * yield - yield the current processor to other threads.
5641 * This is a shortcut for kernel-space yielding - it marks the
5642 * thread runnable and calls sys_sched_yield().
5644 void __sched yield(void)
5646 set_current_state(TASK_RUNNING);
5649 EXPORT_SYMBOL(yield);
5652 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5653 * that process accounting knows that this is a task in IO wait state.
5655 * But don't do that if it is a deliberate, throttling IO wait (this task
5656 * has set its backing_dev_info: the queue against which it should throttle)
5658 void __sched io_schedule(void)
5660 struct rq *rq = &__raw_get_cpu_var(runqueues);
5662 delayacct_blkio_start();
5663 atomic_inc(&rq->nr_iowait);
5665 atomic_dec(&rq->nr_iowait);
5666 delayacct_blkio_end();
5668 EXPORT_SYMBOL(io_schedule);
5670 long __sched io_schedule_timeout(long timeout)
5672 struct rq *rq = &__raw_get_cpu_var(runqueues);
5675 delayacct_blkio_start();
5676 atomic_inc(&rq->nr_iowait);
5677 ret = schedule_timeout(timeout);
5678 atomic_dec(&rq->nr_iowait);
5679 delayacct_blkio_end();
5684 * sys_sched_get_priority_max - return maximum RT priority.
5685 * @policy: scheduling class.
5687 * this syscall returns the maximum rt_priority that can be used
5688 * by a given scheduling class.
5690 asmlinkage long sys_sched_get_priority_max(int policy)
5697 ret = MAX_USER_RT_PRIO-1;
5709 * sys_sched_get_priority_min - return minimum RT priority.
5710 * @policy: scheduling class.
5712 * this syscall returns the minimum rt_priority that can be used
5713 * by a given scheduling class.
5715 asmlinkage long sys_sched_get_priority_min(int policy)
5733 * sys_sched_rr_get_interval - return the default timeslice of a process.
5734 * @pid: pid of the process.
5735 * @interval: userspace pointer to the timeslice value.
5737 * this syscall writes the default timeslice value of a given process
5738 * into the user-space timespec buffer. A value of '0' means infinity.
5741 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5743 struct task_struct *p;
5744 unsigned int time_slice;
5752 read_lock(&tasklist_lock);
5753 p = find_process_by_pid(pid);
5757 retval = security_task_getscheduler(p);
5762 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5763 * tasks that are on an otherwise idle runqueue:
5766 if (p->policy == SCHED_RR) {
5767 time_slice = DEF_TIMESLICE;
5768 } else if (p->policy != SCHED_FIFO) {
5769 struct sched_entity *se = &p->se;
5770 unsigned long flags;
5773 rq = task_rq_lock(p, &flags);
5774 if (rq->cfs.load.weight)
5775 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5776 task_rq_unlock(rq, &flags);
5778 read_unlock(&tasklist_lock);
5779 jiffies_to_timespec(time_slice, &t);
5780 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5784 read_unlock(&tasklist_lock);
5788 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5790 void sched_show_task(struct task_struct *p)
5792 unsigned long free = 0;
5795 state = p->state ? __ffs(p->state) + 1 : 0;
5796 printk(KERN_INFO "%-13.13s %c", p->comm,
5797 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5798 #if BITS_PER_LONG == 32
5799 if (state == TASK_RUNNING)
5800 printk(KERN_CONT " running ");
5802 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5804 if (state == TASK_RUNNING)
5805 printk(KERN_CONT " running task ");
5807 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5809 #ifdef CONFIG_DEBUG_STACK_USAGE
5811 unsigned long *n = end_of_stack(p);
5814 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5817 printk(KERN_CONT "%5lu %5d %6d\n", free,
5818 task_pid_nr(p), task_pid_nr(p->real_parent));
5820 show_stack(p, NULL);
5823 void show_state_filter(unsigned long state_filter)
5825 struct task_struct *g, *p;
5827 #if BITS_PER_LONG == 32
5829 " task PC stack pid father\n");
5832 " task PC stack pid father\n");
5834 read_lock(&tasklist_lock);
5835 do_each_thread(g, p) {
5837 * reset the NMI-timeout, listing all files on a slow
5838 * console might take alot of time:
5840 touch_nmi_watchdog();
5841 if (!state_filter || (p->state & state_filter))
5843 } while_each_thread(g, p);
5845 touch_all_softlockup_watchdogs();
5847 #ifdef CONFIG_SCHED_DEBUG
5848 sysrq_sched_debug_show();
5850 read_unlock(&tasklist_lock);
5852 * Only show locks if all tasks are dumped:
5854 if (state_filter == -1)
5855 debug_show_all_locks();
5858 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5860 idle->sched_class = &idle_sched_class;
5864 * init_idle - set up an idle thread for a given CPU
5865 * @idle: task in question
5866 * @cpu: cpu the idle task belongs to
5868 * NOTE: this function does not set the idle thread's NEED_RESCHED
5869 * flag, to make booting more robust.
5871 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5873 struct rq *rq = cpu_rq(cpu);
5874 unsigned long flags;
5876 spin_lock_irqsave(&rq->lock, flags);
5879 idle->se.exec_start = sched_clock();
5881 idle->prio = idle->normal_prio = MAX_PRIO;
5882 idle->cpus_allowed = cpumask_of_cpu(cpu);
5883 __set_task_cpu(idle, cpu);
5885 rq->curr = rq->idle = idle;
5886 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5889 spin_unlock_irqrestore(&rq->lock, flags);
5891 /* Set the preempt count _outside_ the spinlocks! */
5892 #if defined(CONFIG_PREEMPT)
5893 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5895 task_thread_info(idle)->preempt_count = 0;
5898 * The idle tasks have their own, simple scheduling class:
5900 idle->sched_class = &idle_sched_class;
5901 ftrace_graph_init_task(idle);
5905 * In a system that switches off the HZ timer nohz_cpu_mask
5906 * indicates which cpus entered this state. This is used
5907 * in the rcu update to wait only for active cpus. For system
5908 * which do not switch off the HZ timer nohz_cpu_mask should
5909 * always be CPU_MASK_NONE.
5911 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5914 * Increase the granularity value when there are more CPUs,
5915 * because with more CPUs the 'effective latency' as visible
5916 * to users decreases. But the relationship is not linear,
5917 * so pick a second-best guess by going with the log2 of the
5920 * This idea comes from the SD scheduler of Con Kolivas:
5922 static inline void sched_init_granularity(void)
5924 unsigned int factor = 1 + ilog2(num_online_cpus());
5925 const unsigned long limit = 200000000;
5927 sysctl_sched_min_granularity *= factor;
5928 if (sysctl_sched_min_granularity > limit)
5929 sysctl_sched_min_granularity = limit;
5931 sysctl_sched_latency *= factor;
5932 if (sysctl_sched_latency > limit)
5933 sysctl_sched_latency = limit;
5935 sysctl_sched_wakeup_granularity *= factor;
5937 sysctl_sched_shares_ratelimit *= factor;
5942 * This is how migration works:
5944 * 1) we queue a struct migration_req structure in the source CPU's
5945 * runqueue and wake up that CPU's migration thread.
5946 * 2) we down() the locked semaphore => thread blocks.
5947 * 3) migration thread wakes up (implicitly it forces the migrated
5948 * thread off the CPU)
5949 * 4) it gets the migration request and checks whether the migrated
5950 * task is still in the wrong runqueue.
5951 * 5) if it's in the wrong runqueue then the migration thread removes
5952 * it and puts it into the right queue.
5953 * 6) migration thread up()s the semaphore.
5954 * 7) we wake up and the migration is done.
5958 * Change a given task's CPU affinity. Migrate the thread to a
5959 * proper CPU and schedule it away if the CPU it's executing on
5960 * is removed from the allowed bitmask.
5962 * NOTE: the caller must have a valid reference to the task, the
5963 * task must not exit() & deallocate itself prematurely. The
5964 * call is not atomic; no spinlocks may be held.
5966 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5968 struct migration_req req;
5969 unsigned long flags;
5973 rq = task_rq_lock(p, &flags);
5974 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5979 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5980 !cpus_equal(p->cpus_allowed, *new_mask))) {
5985 if (p->sched_class->set_cpus_allowed)
5986 p->sched_class->set_cpus_allowed(p, new_mask);
5988 p->cpus_allowed = *new_mask;
5989 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5992 /* Can the task run on the task's current CPU? If so, we're done */
5993 if (cpu_isset(task_cpu(p), *new_mask))
5996 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5997 /* Need help from migration thread: drop lock and wait. */
5998 task_rq_unlock(rq, &flags);
5999 wake_up_process(rq->migration_thread);
6000 wait_for_completion(&req.done);
6001 tlb_migrate_finish(p->mm);
6005 task_rq_unlock(rq, &flags);
6009 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6012 * Move (not current) task off this cpu, onto dest cpu. We're doing
6013 * this because either it can't run here any more (set_cpus_allowed()
6014 * away from this CPU, or CPU going down), or because we're
6015 * attempting to rebalance this task on exec (sched_exec).
6017 * So we race with normal scheduler movements, but that's OK, as long
6018 * as the task is no longer on this CPU.
6020 * Returns non-zero if task was successfully migrated.
6022 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6024 struct rq *rq_dest, *rq_src;
6027 if (unlikely(!cpu_active(dest_cpu)))
6030 rq_src = cpu_rq(src_cpu);
6031 rq_dest = cpu_rq(dest_cpu);
6033 double_rq_lock(rq_src, rq_dest);
6034 /* Already moved. */
6035 if (task_cpu(p) != src_cpu)
6037 /* Affinity changed (again). */
6038 if (!cpu_isset(dest_cpu, p->cpus_allowed))
6041 on_rq = p->se.on_rq;
6043 deactivate_task(rq_src, p, 0);
6045 set_task_cpu(p, dest_cpu);
6047 activate_task(rq_dest, p, 0);
6048 check_preempt_curr(rq_dest, p, 0);
6053 double_rq_unlock(rq_src, rq_dest);
6058 * migration_thread - this is a highprio system thread that performs
6059 * thread migration by bumping thread off CPU then 'pushing' onto
6062 static int migration_thread(void *data)
6064 int cpu = (long)data;
6068 BUG_ON(rq->migration_thread != current);
6070 set_current_state(TASK_INTERRUPTIBLE);
6071 while (!kthread_should_stop()) {
6072 struct migration_req *req;
6073 struct list_head *head;
6075 spin_lock_irq(&rq->lock);
6077 if (cpu_is_offline(cpu)) {
6078 spin_unlock_irq(&rq->lock);
6082 if (rq->active_balance) {
6083 active_load_balance(rq, cpu);
6084 rq->active_balance = 0;
6087 head = &rq->migration_queue;
6089 if (list_empty(head)) {
6090 spin_unlock_irq(&rq->lock);
6092 set_current_state(TASK_INTERRUPTIBLE);
6095 req = list_entry(head->next, struct migration_req, list);
6096 list_del_init(head->next);
6098 spin_unlock(&rq->lock);
6099 __migrate_task(req->task, cpu, req->dest_cpu);
6102 complete(&req->done);
6104 __set_current_state(TASK_RUNNING);
6108 /* Wait for kthread_stop */
6109 set_current_state(TASK_INTERRUPTIBLE);
6110 while (!kthread_should_stop()) {
6112 set_current_state(TASK_INTERRUPTIBLE);
6114 __set_current_state(TASK_RUNNING);
6118 #ifdef CONFIG_HOTPLUG_CPU
6120 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6124 local_irq_disable();
6125 ret = __migrate_task(p, src_cpu, dest_cpu);
6131 * Figure out where task on dead CPU should go, use force if necessary.
6133 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6135 unsigned long flags;
6142 mask = node_to_cpumask(cpu_to_node(dead_cpu));
6143 cpus_and(mask, mask, p->cpus_allowed);
6144 dest_cpu = any_online_cpu(mask);
6146 /* On any allowed CPU? */
6147 if (dest_cpu >= nr_cpu_ids)
6148 dest_cpu = any_online_cpu(p->cpus_allowed);
6150 /* No more Mr. Nice Guy. */
6151 if (dest_cpu >= nr_cpu_ids) {
6152 cpumask_t cpus_allowed;
6154 cpuset_cpus_allowed_locked(p, &cpus_allowed);
6156 * Try to stay on the same cpuset, where the
6157 * current cpuset may be a subset of all cpus.
6158 * The cpuset_cpus_allowed_locked() variant of
6159 * cpuset_cpus_allowed() will not block. It must be
6160 * called within calls to cpuset_lock/cpuset_unlock.
6162 rq = task_rq_lock(p, &flags);
6163 p->cpus_allowed = cpus_allowed;
6164 dest_cpu = any_online_cpu(p->cpus_allowed);
6165 task_rq_unlock(rq, &flags);
6168 * Don't tell them about moving exiting tasks or
6169 * kernel threads (both mm NULL), since they never
6172 if (p->mm && printk_ratelimit()) {
6173 printk(KERN_INFO "process %d (%s) no "
6174 "longer affine to cpu%d\n",
6175 task_pid_nr(p), p->comm, dead_cpu);
6178 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
6182 * While a dead CPU has no uninterruptible tasks queued at this point,
6183 * it might still have a nonzero ->nr_uninterruptible counter, because
6184 * for performance reasons the counter is not stricly tracking tasks to
6185 * their home CPUs. So we just add the counter to another CPU's counter,
6186 * to keep the global sum constant after CPU-down:
6188 static void migrate_nr_uninterruptible(struct rq *rq_src)
6190 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
6191 unsigned long flags;
6193 local_irq_save(flags);
6194 double_rq_lock(rq_src, rq_dest);
6195 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6196 rq_src->nr_uninterruptible = 0;
6197 double_rq_unlock(rq_src, rq_dest);
6198 local_irq_restore(flags);
6201 /* Run through task list and migrate tasks from the dead cpu. */
6202 static void migrate_live_tasks(int src_cpu)
6204 struct task_struct *p, *t;
6206 read_lock(&tasklist_lock);
6208 do_each_thread(t, p) {
6212 if (task_cpu(p) == src_cpu)
6213 move_task_off_dead_cpu(src_cpu, p);
6214 } while_each_thread(t, p);
6216 read_unlock(&tasklist_lock);
6220 * Schedules idle task to be the next runnable task on current CPU.
6221 * It does so by boosting its priority to highest possible.
6222 * Used by CPU offline code.
6224 void sched_idle_next(void)
6226 int this_cpu = smp_processor_id();
6227 struct rq *rq = cpu_rq(this_cpu);
6228 struct task_struct *p = rq->idle;
6229 unsigned long flags;
6231 /* cpu has to be offline */
6232 BUG_ON(cpu_online(this_cpu));
6235 * Strictly not necessary since rest of the CPUs are stopped by now
6236 * and interrupts disabled on the current cpu.
6238 spin_lock_irqsave(&rq->lock, flags);
6240 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6242 update_rq_clock(rq);
6243 activate_task(rq, p, 0);
6245 spin_unlock_irqrestore(&rq->lock, flags);
6249 * Ensures that the idle task is using init_mm right before its cpu goes
6252 void idle_task_exit(void)
6254 struct mm_struct *mm = current->active_mm;
6256 BUG_ON(cpu_online(smp_processor_id()));
6259 switch_mm(mm, &init_mm, current);
6263 /* called under rq->lock with disabled interrupts */
6264 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6266 struct rq *rq = cpu_rq(dead_cpu);
6268 /* Must be exiting, otherwise would be on tasklist. */
6269 BUG_ON(!p->exit_state);
6271 /* Cannot have done final schedule yet: would have vanished. */
6272 BUG_ON(p->state == TASK_DEAD);
6277 * Drop lock around migration; if someone else moves it,
6278 * that's OK. No task can be added to this CPU, so iteration is
6281 spin_unlock_irq(&rq->lock);
6282 move_task_off_dead_cpu(dead_cpu, p);
6283 spin_lock_irq(&rq->lock);
6288 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6289 static void migrate_dead_tasks(unsigned int dead_cpu)
6291 struct rq *rq = cpu_rq(dead_cpu);
6292 struct task_struct *next;
6295 if (!rq->nr_running)
6297 update_rq_clock(rq);
6298 next = pick_next_task(rq, rq->curr);
6301 next->sched_class->put_prev_task(rq, next);
6302 migrate_dead(dead_cpu, next);
6306 #endif /* CONFIG_HOTPLUG_CPU */
6308 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6310 static struct ctl_table sd_ctl_dir[] = {
6312 .procname = "sched_domain",
6318 static struct ctl_table sd_ctl_root[] = {
6320 .ctl_name = CTL_KERN,
6321 .procname = "kernel",
6323 .child = sd_ctl_dir,
6328 static struct ctl_table *sd_alloc_ctl_entry(int n)
6330 struct ctl_table *entry =
6331 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6336 static void sd_free_ctl_entry(struct ctl_table **tablep)
6338 struct ctl_table *entry;
6341 * In the intermediate directories, both the child directory and
6342 * procname are dynamically allocated and could fail but the mode
6343 * will always be set. In the lowest directory the names are
6344 * static strings and all have proc handlers.
6346 for (entry = *tablep; entry->mode; entry++) {
6348 sd_free_ctl_entry(&entry->child);
6349 if (entry->proc_handler == NULL)
6350 kfree(entry->procname);
6358 set_table_entry(struct ctl_table *entry,
6359 const char *procname, void *data, int maxlen,
6360 mode_t mode, proc_handler *proc_handler)
6362 entry->procname = procname;
6364 entry->maxlen = maxlen;
6366 entry->proc_handler = proc_handler;
6369 static struct ctl_table *
6370 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6372 struct ctl_table *table = sd_alloc_ctl_entry(13);
6377 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6378 sizeof(long), 0644, proc_doulongvec_minmax);
6379 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6380 sizeof(long), 0644, proc_doulongvec_minmax);
6381 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6382 sizeof(int), 0644, proc_dointvec_minmax);
6383 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6384 sizeof(int), 0644, proc_dointvec_minmax);
6385 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6386 sizeof(int), 0644, proc_dointvec_minmax);
6387 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6388 sizeof(int), 0644, proc_dointvec_minmax);
6389 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6390 sizeof(int), 0644, proc_dointvec_minmax);
6391 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6392 sizeof(int), 0644, proc_dointvec_minmax);
6393 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6394 sizeof(int), 0644, proc_dointvec_minmax);
6395 set_table_entry(&table[9], "cache_nice_tries",
6396 &sd->cache_nice_tries,
6397 sizeof(int), 0644, proc_dointvec_minmax);
6398 set_table_entry(&table[10], "flags", &sd->flags,
6399 sizeof(int), 0644, proc_dointvec_minmax);
6400 set_table_entry(&table[11], "name", sd->name,
6401 CORENAME_MAX_SIZE, 0444, proc_dostring);
6402 /* &table[12] is terminator */
6407 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6409 struct ctl_table *entry, *table;
6410 struct sched_domain *sd;
6411 int domain_num = 0, i;
6414 for_each_domain(cpu, sd)
6416 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6421 for_each_domain(cpu, sd) {
6422 snprintf(buf, 32, "domain%d", i);
6423 entry->procname = kstrdup(buf, GFP_KERNEL);
6425 entry->child = sd_alloc_ctl_domain_table(sd);
6432 static struct ctl_table_header *sd_sysctl_header;
6433 static void register_sched_domain_sysctl(void)
6435 int i, cpu_num = num_online_cpus();
6436 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6439 WARN_ON(sd_ctl_dir[0].child);
6440 sd_ctl_dir[0].child = entry;
6445 for_each_online_cpu(i) {
6446 snprintf(buf, 32, "cpu%d", i);
6447 entry->procname = kstrdup(buf, GFP_KERNEL);
6449 entry->child = sd_alloc_ctl_cpu_table(i);
6453 WARN_ON(sd_sysctl_header);
6454 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6457 /* may be called multiple times per register */
6458 static void unregister_sched_domain_sysctl(void)
6460 if (sd_sysctl_header)
6461 unregister_sysctl_table(sd_sysctl_header);
6462 sd_sysctl_header = NULL;
6463 if (sd_ctl_dir[0].child)
6464 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6467 static void register_sched_domain_sysctl(void)
6470 static void unregister_sched_domain_sysctl(void)
6475 static void set_rq_online(struct rq *rq)
6478 const struct sched_class *class;
6480 cpu_set(rq->cpu, rq->rd->online);
6483 for_each_class(class) {
6484 if (class->rq_online)
6485 class->rq_online(rq);
6490 static void set_rq_offline(struct rq *rq)
6493 const struct sched_class *class;
6495 for_each_class(class) {
6496 if (class->rq_offline)
6497 class->rq_offline(rq);
6500 cpu_clear(rq->cpu, rq->rd->online);
6506 * migration_call - callback that gets triggered when a CPU is added.
6507 * Here we can start up the necessary migration thread for the new CPU.
6509 static int __cpuinit
6510 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6512 struct task_struct *p;
6513 int cpu = (long)hcpu;
6514 unsigned long flags;
6519 case CPU_UP_PREPARE:
6520 case CPU_UP_PREPARE_FROZEN:
6521 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6524 kthread_bind(p, cpu);
6525 /* Must be high prio: stop_machine expects to yield to it. */
6526 rq = task_rq_lock(p, &flags);
6527 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6528 task_rq_unlock(rq, &flags);
6529 cpu_rq(cpu)->migration_thread = p;
6533 case CPU_ONLINE_FROZEN:
6534 /* Strictly unnecessary, as first user will wake it. */
6535 wake_up_process(cpu_rq(cpu)->migration_thread);
6537 /* Update our root-domain */
6539 spin_lock_irqsave(&rq->lock, flags);
6541 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6545 spin_unlock_irqrestore(&rq->lock, flags);
6548 #ifdef CONFIG_HOTPLUG_CPU
6549 case CPU_UP_CANCELED:
6550 case CPU_UP_CANCELED_FROZEN:
6551 if (!cpu_rq(cpu)->migration_thread)
6553 /* Unbind it from offline cpu so it can run. Fall thru. */
6554 kthread_bind(cpu_rq(cpu)->migration_thread,
6555 any_online_cpu(cpu_online_map));
6556 kthread_stop(cpu_rq(cpu)->migration_thread);
6557 cpu_rq(cpu)->migration_thread = NULL;
6561 case CPU_DEAD_FROZEN:
6562 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6563 migrate_live_tasks(cpu);
6565 kthread_stop(rq->migration_thread);
6566 rq->migration_thread = NULL;
6567 /* Idle task back to normal (off runqueue, low prio) */
6568 spin_lock_irq(&rq->lock);
6569 update_rq_clock(rq);
6570 deactivate_task(rq, rq->idle, 0);
6571 rq->idle->static_prio = MAX_PRIO;
6572 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6573 rq->idle->sched_class = &idle_sched_class;
6574 migrate_dead_tasks(cpu);
6575 spin_unlock_irq(&rq->lock);
6577 migrate_nr_uninterruptible(rq);
6578 BUG_ON(rq->nr_running != 0);
6581 * No need to migrate the tasks: it was best-effort if
6582 * they didn't take sched_hotcpu_mutex. Just wake up
6585 spin_lock_irq(&rq->lock);
6586 while (!list_empty(&rq->migration_queue)) {
6587 struct migration_req *req;
6589 req = list_entry(rq->migration_queue.next,
6590 struct migration_req, list);
6591 list_del_init(&req->list);
6592 spin_unlock_irq(&rq->lock);
6593 complete(&req->done);
6594 spin_lock_irq(&rq->lock);
6596 spin_unlock_irq(&rq->lock);
6600 case CPU_DYING_FROZEN:
6601 /* Update our root-domain */
6603 spin_lock_irqsave(&rq->lock, flags);
6605 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6608 spin_unlock_irqrestore(&rq->lock, flags);
6615 /* Register at highest priority so that task migration (migrate_all_tasks)
6616 * happens before everything else.
6618 static struct notifier_block __cpuinitdata migration_notifier = {
6619 .notifier_call = migration_call,
6623 static int __init migration_init(void)
6625 void *cpu = (void *)(long)smp_processor_id();
6628 /* Start one for the boot CPU: */
6629 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6630 BUG_ON(err == NOTIFY_BAD);
6631 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6632 register_cpu_notifier(&migration_notifier);
6636 early_initcall(migration_init);
6641 #ifdef CONFIG_SCHED_DEBUG
6643 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6644 cpumask_t *groupmask)
6646 struct sched_group *group = sd->groups;
6649 cpulist_scnprintf(str, sizeof(str), sd->span);
6650 cpus_clear(*groupmask);
6652 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6654 if (!(sd->flags & SD_LOAD_BALANCE)) {
6655 printk("does not load-balance\n");
6657 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6662 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6664 if (!cpu_isset(cpu, sd->span)) {
6665 printk(KERN_ERR "ERROR: domain->span does not contain "
6668 if (!cpu_isset(cpu, group->cpumask)) {
6669 printk(KERN_ERR "ERROR: domain->groups does not contain"
6673 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6677 printk(KERN_ERR "ERROR: group is NULL\n");
6681 if (!group->__cpu_power) {
6682 printk(KERN_CONT "\n");
6683 printk(KERN_ERR "ERROR: domain->cpu_power not "
6688 if (!cpus_weight(group->cpumask)) {
6689 printk(KERN_CONT "\n");
6690 printk(KERN_ERR "ERROR: empty group\n");
6694 if (cpus_intersects(*groupmask, group->cpumask)) {
6695 printk(KERN_CONT "\n");
6696 printk(KERN_ERR "ERROR: repeated CPUs\n");
6700 cpus_or(*groupmask, *groupmask, group->cpumask);
6702 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6703 printk(KERN_CONT " %s", str);
6705 group = group->next;
6706 } while (group != sd->groups);
6707 printk(KERN_CONT "\n");
6709 if (!cpus_equal(sd->span, *groupmask))
6710 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6712 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6713 printk(KERN_ERR "ERROR: parent span is not a superset "
6714 "of domain->span\n");
6718 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6720 cpumask_t *groupmask;
6724 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6728 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6730 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6732 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6737 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6746 #else /* !CONFIG_SCHED_DEBUG */
6747 # define sched_domain_debug(sd, cpu) do { } while (0)
6748 #endif /* CONFIG_SCHED_DEBUG */
6750 static int sd_degenerate(struct sched_domain *sd)
6752 if (cpus_weight(sd->span) == 1)
6755 /* Following flags need at least 2 groups */
6756 if (sd->flags & (SD_LOAD_BALANCE |
6757 SD_BALANCE_NEWIDLE |
6761 SD_SHARE_PKG_RESOURCES)) {
6762 if (sd->groups != sd->groups->next)
6766 /* Following flags don't use groups */
6767 if (sd->flags & (SD_WAKE_IDLE |
6776 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6778 unsigned long cflags = sd->flags, pflags = parent->flags;
6780 if (sd_degenerate(parent))
6783 if (!cpus_equal(sd->span, parent->span))
6786 /* Does parent contain flags not in child? */
6787 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6788 if (cflags & SD_WAKE_AFFINE)
6789 pflags &= ~SD_WAKE_BALANCE;
6790 /* Flags needing groups don't count if only 1 group in parent */
6791 if (parent->groups == parent->groups->next) {
6792 pflags &= ~(SD_LOAD_BALANCE |
6793 SD_BALANCE_NEWIDLE |
6797 SD_SHARE_PKG_RESOURCES);
6798 if (nr_node_ids == 1)
6799 pflags &= ~SD_SERIALIZE;
6801 if (~cflags & pflags)
6807 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6809 unsigned long flags;
6811 spin_lock_irqsave(&rq->lock, flags);
6814 struct root_domain *old_rd = rq->rd;
6816 if (cpu_isset(rq->cpu, old_rd->online))
6819 cpu_clear(rq->cpu, old_rd->span);
6821 if (atomic_dec_and_test(&old_rd->refcount))
6825 atomic_inc(&rd->refcount);
6828 cpu_set(rq->cpu, rd->span);
6829 if (cpu_isset(rq->cpu, cpu_online_map))
6832 spin_unlock_irqrestore(&rq->lock, flags);
6835 static void init_rootdomain(struct root_domain *rd)
6837 memset(rd, 0, sizeof(*rd));
6839 cpus_clear(rd->span);
6840 cpus_clear(rd->online);
6842 cpupri_init(&rd->cpupri);
6845 static void init_defrootdomain(void)
6847 init_rootdomain(&def_root_domain);
6848 atomic_set(&def_root_domain.refcount, 1);
6851 static struct root_domain *alloc_rootdomain(void)
6853 struct root_domain *rd;
6855 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6859 init_rootdomain(rd);
6865 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6866 * hold the hotplug lock.
6869 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6871 struct rq *rq = cpu_rq(cpu);
6872 struct sched_domain *tmp;
6874 /* Remove the sched domains which do not contribute to scheduling. */
6875 for (tmp = sd; tmp; ) {
6876 struct sched_domain *parent = tmp->parent;
6880 if (sd_parent_degenerate(tmp, parent)) {
6881 tmp->parent = parent->parent;
6883 parent->parent->child = tmp;
6888 if (sd && sd_degenerate(sd)) {
6894 sched_domain_debug(sd, cpu);
6896 rq_attach_root(rq, rd);
6897 rcu_assign_pointer(rq->sd, sd);
6900 /* cpus with isolated domains */
6901 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6903 /* Setup the mask of cpus configured for isolated domains */
6904 static int __init isolated_cpu_setup(char *str)
6906 static int __initdata ints[NR_CPUS];
6909 str = get_options(str, ARRAY_SIZE(ints), ints);
6910 cpus_clear(cpu_isolated_map);
6911 for (i = 1; i <= ints[0]; i++)
6912 if (ints[i] < NR_CPUS)
6913 cpu_set(ints[i], cpu_isolated_map);
6917 __setup("isolcpus=", isolated_cpu_setup);
6920 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6921 * to a function which identifies what group(along with sched group) a CPU
6922 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6923 * (due to the fact that we keep track of groups covered with a cpumask_t).
6925 * init_sched_build_groups will build a circular linked list of the groups
6926 * covered by the given span, and will set each group's ->cpumask correctly,
6927 * and ->cpu_power to 0.
6930 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6931 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6932 struct sched_group **sg,
6933 cpumask_t *tmpmask),
6934 cpumask_t *covered, cpumask_t *tmpmask)
6936 struct sched_group *first = NULL, *last = NULL;
6939 cpus_clear(*covered);
6941 for_each_cpu_mask_nr(i, *span) {
6942 struct sched_group *sg;
6943 int group = group_fn(i, cpu_map, &sg, tmpmask);
6946 if (cpu_isset(i, *covered))
6949 cpus_clear(sg->cpumask);
6950 sg->__cpu_power = 0;
6952 for_each_cpu_mask_nr(j, *span) {
6953 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6956 cpu_set(j, *covered);
6957 cpu_set(j, sg->cpumask);
6968 #define SD_NODES_PER_DOMAIN 16
6973 * find_next_best_node - find the next node to include in a sched_domain
6974 * @node: node whose sched_domain we're building
6975 * @used_nodes: nodes already in the sched_domain
6977 * Find the next node to include in a given scheduling domain. Simply
6978 * finds the closest node not already in the @used_nodes map.
6980 * Should use nodemask_t.
6982 static int find_next_best_node(int node, nodemask_t *used_nodes)
6984 int i, n, val, min_val, best_node = 0;
6988 for (i = 0; i < nr_node_ids; i++) {
6989 /* Start at @node */
6990 n = (node + i) % nr_node_ids;
6992 if (!nr_cpus_node(n))
6995 /* Skip already used nodes */
6996 if (node_isset(n, *used_nodes))
6999 /* Simple min distance search */
7000 val = node_distance(node, n);
7002 if (val < min_val) {
7008 node_set(best_node, *used_nodes);
7013 * sched_domain_node_span - get a cpumask for a node's sched_domain
7014 * @node: node whose cpumask we're constructing
7015 * @span: resulting cpumask
7017 * Given a node, construct a good cpumask for its sched_domain to span. It
7018 * should be one that prevents unnecessary balancing, but also spreads tasks
7021 static void sched_domain_node_span(int node, cpumask_t *span)
7023 nodemask_t used_nodes;
7024 node_to_cpumask_ptr(nodemask, node);
7028 nodes_clear(used_nodes);
7030 cpus_or(*span, *span, *nodemask);
7031 node_set(node, used_nodes);
7033 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7034 int next_node = find_next_best_node(node, &used_nodes);
7036 node_to_cpumask_ptr_next(nodemask, next_node);
7037 cpus_or(*span, *span, *nodemask);
7040 #endif /* CONFIG_NUMA */
7042 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7045 * SMT sched-domains:
7047 #ifdef CONFIG_SCHED_SMT
7048 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
7049 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
7052 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7056 *sg = &per_cpu(sched_group_cpus, cpu);
7059 #endif /* CONFIG_SCHED_SMT */
7062 * multi-core sched-domains:
7064 #ifdef CONFIG_SCHED_MC
7065 static DEFINE_PER_CPU(struct sched_domain, core_domains);
7066 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
7067 #endif /* CONFIG_SCHED_MC */
7069 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7071 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7076 *mask = per_cpu(cpu_sibling_map, cpu);
7077 cpus_and(*mask, *mask, *cpu_map);
7078 group = first_cpu(*mask);
7080 *sg = &per_cpu(sched_group_core, group);
7083 #elif defined(CONFIG_SCHED_MC)
7085 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7089 *sg = &per_cpu(sched_group_core, cpu);
7094 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
7095 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
7098 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7102 #ifdef CONFIG_SCHED_MC
7103 *mask = cpu_coregroup_map(cpu);
7104 cpus_and(*mask, *mask, *cpu_map);
7105 group = first_cpu(*mask);
7106 #elif defined(CONFIG_SCHED_SMT)
7107 *mask = per_cpu(cpu_sibling_map, cpu);
7108 cpus_and(*mask, *mask, *cpu_map);
7109 group = first_cpu(*mask);
7114 *sg = &per_cpu(sched_group_phys, group);
7120 * The init_sched_build_groups can't handle what we want to do with node
7121 * groups, so roll our own. Now each node has its own list of groups which
7122 * gets dynamically allocated.
7124 static DEFINE_PER_CPU(struct sched_domain, node_domains);
7125 static struct sched_group ***sched_group_nodes_bycpu;
7127 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
7128 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
7130 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
7131 struct sched_group **sg, cpumask_t *nodemask)
7135 *nodemask = node_to_cpumask(cpu_to_node(cpu));
7136 cpus_and(*nodemask, *nodemask, *cpu_map);
7137 group = first_cpu(*nodemask);
7140 *sg = &per_cpu(sched_group_allnodes, group);
7144 static void init_numa_sched_groups_power(struct sched_group *group_head)
7146 struct sched_group *sg = group_head;
7152 for_each_cpu_mask_nr(j, sg->cpumask) {
7153 struct sched_domain *sd;
7155 sd = &per_cpu(phys_domains, j);
7156 if (j != first_cpu(sd->groups->cpumask)) {
7158 * Only add "power" once for each
7164 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7167 } while (sg != group_head);
7169 #endif /* CONFIG_NUMA */
7172 /* Free memory allocated for various sched_group structures */
7173 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7177 for_each_cpu_mask_nr(cpu, *cpu_map) {
7178 struct sched_group **sched_group_nodes
7179 = sched_group_nodes_bycpu[cpu];
7181 if (!sched_group_nodes)
7184 for (i = 0; i < nr_node_ids; i++) {
7185 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7187 *nodemask = node_to_cpumask(i);
7188 cpus_and(*nodemask, *nodemask, *cpu_map);
7189 if (cpus_empty(*nodemask))
7199 if (oldsg != sched_group_nodes[i])
7202 kfree(sched_group_nodes);
7203 sched_group_nodes_bycpu[cpu] = NULL;
7206 #else /* !CONFIG_NUMA */
7207 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7210 #endif /* CONFIG_NUMA */
7213 * Initialize sched groups cpu_power.
7215 * cpu_power indicates the capacity of sched group, which is used while
7216 * distributing the load between different sched groups in a sched domain.
7217 * Typically cpu_power for all the groups in a sched domain will be same unless
7218 * there are asymmetries in the topology. If there are asymmetries, group
7219 * having more cpu_power will pickup more load compared to the group having
7222 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7223 * the maximum number of tasks a group can handle in the presence of other idle
7224 * or lightly loaded groups in the same sched domain.
7226 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7228 struct sched_domain *child;
7229 struct sched_group *group;
7231 WARN_ON(!sd || !sd->groups);
7233 if (cpu != first_cpu(sd->groups->cpumask))
7238 sd->groups->__cpu_power = 0;
7241 * For perf policy, if the groups in child domain share resources
7242 * (for example cores sharing some portions of the cache hierarchy
7243 * or SMT), then set this domain groups cpu_power such that each group
7244 * can handle only one task, when there are other idle groups in the
7245 * same sched domain.
7247 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7249 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7250 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7255 * add cpu_power of each child group to this groups cpu_power
7257 group = child->groups;
7259 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7260 group = group->next;
7261 } while (group != child->groups);
7265 * Initializers for schedule domains
7266 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7269 #ifdef CONFIG_SCHED_DEBUG
7270 # define SD_INIT_NAME(sd, type) sd->name = #type
7272 # define SD_INIT_NAME(sd, type) do { } while (0)
7275 #define SD_INIT(sd, type) sd_init_##type(sd)
7277 #define SD_INIT_FUNC(type) \
7278 static noinline void sd_init_##type(struct sched_domain *sd) \
7280 memset(sd, 0, sizeof(*sd)); \
7281 *sd = SD_##type##_INIT; \
7282 sd->level = SD_LV_##type; \
7283 SD_INIT_NAME(sd, type); \
7288 SD_INIT_FUNC(ALLNODES)
7291 #ifdef CONFIG_SCHED_SMT
7292 SD_INIT_FUNC(SIBLING)
7294 #ifdef CONFIG_SCHED_MC
7299 * To minimize stack usage kmalloc room for cpumasks and share the
7300 * space as the usage in build_sched_domains() dictates. Used only
7301 * if the amount of space is significant.
7304 cpumask_t tmpmask; /* make this one first */
7307 cpumask_t this_sibling_map;
7308 cpumask_t this_core_map;
7310 cpumask_t send_covered;
7313 cpumask_t domainspan;
7315 cpumask_t notcovered;
7320 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7321 static inline void sched_cpumask_alloc(struct allmasks **masks)
7323 *masks = kmalloc(sizeof(**masks), GFP_KERNEL);
7325 static inline void sched_cpumask_free(struct allmasks *masks)
7330 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7331 static inline void sched_cpumask_alloc(struct allmasks **masks)
7333 static inline void sched_cpumask_free(struct allmasks *masks)
7337 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7338 ((unsigned long)(a) + offsetof(struct allmasks, v))
7340 static int default_relax_domain_level = -1;
7342 static int __init setup_relax_domain_level(char *str)
7346 val = simple_strtoul(str, NULL, 0);
7347 if (val < SD_LV_MAX)
7348 default_relax_domain_level = val;
7352 __setup("relax_domain_level=", setup_relax_domain_level);
7354 static void set_domain_attribute(struct sched_domain *sd,
7355 struct sched_domain_attr *attr)
7359 if (!attr || attr->relax_domain_level < 0) {
7360 if (default_relax_domain_level < 0)
7363 request = default_relax_domain_level;
7365 request = attr->relax_domain_level;
7366 if (request < sd->level) {
7367 /* turn off idle balance on this domain */
7368 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7370 /* turn on idle balance on this domain */
7371 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7376 * Build sched domains for a given set of cpus and attach the sched domains
7377 * to the individual cpus
7379 static int __build_sched_domains(const cpumask_t *cpu_map,
7380 struct sched_domain_attr *attr)
7383 struct root_domain *rd;
7384 SCHED_CPUMASK_DECLARE(allmasks);
7387 struct sched_group **sched_group_nodes = NULL;
7388 int sd_allnodes = 0;
7391 * Allocate the per-node list of sched groups
7393 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
7395 if (!sched_group_nodes) {
7396 printk(KERN_WARNING "Can not alloc sched group node list\n");
7401 rd = alloc_rootdomain();
7403 printk(KERN_WARNING "Cannot alloc root domain\n");
7405 kfree(sched_group_nodes);
7410 /* get space for all scratch cpumask variables */
7411 sched_cpumask_alloc(&allmasks);
7413 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7416 kfree(sched_group_nodes);
7421 tmpmask = (cpumask_t *)allmasks;
7425 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7429 * Set up domains for cpus specified by the cpu_map.
7431 for_each_cpu_mask_nr(i, *cpu_map) {
7432 struct sched_domain *sd = NULL, *p;
7433 SCHED_CPUMASK_VAR(nodemask, allmasks);
7435 *nodemask = node_to_cpumask(cpu_to_node(i));
7436 cpus_and(*nodemask, *nodemask, *cpu_map);
7439 if (cpus_weight(*cpu_map) >
7440 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7441 sd = &per_cpu(allnodes_domains, i);
7442 SD_INIT(sd, ALLNODES);
7443 set_domain_attribute(sd, attr);
7444 sd->span = *cpu_map;
7445 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7451 sd = &per_cpu(node_domains, i);
7453 set_domain_attribute(sd, attr);
7454 sched_domain_node_span(cpu_to_node(i), &sd->span);
7458 cpus_and(sd->span, sd->span, *cpu_map);
7462 sd = &per_cpu(phys_domains, i);
7464 set_domain_attribute(sd, attr);
7465 sd->span = *nodemask;
7469 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7471 #ifdef CONFIG_SCHED_MC
7473 sd = &per_cpu(core_domains, i);
7475 set_domain_attribute(sd, attr);
7476 sd->span = cpu_coregroup_map(i);
7477 cpus_and(sd->span, sd->span, *cpu_map);
7480 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7483 #ifdef CONFIG_SCHED_SMT
7485 sd = &per_cpu(cpu_domains, i);
7486 SD_INIT(sd, SIBLING);
7487 set_domain_attribute(sd, attr);
7488 sd->span = per_cpu(cpu_sibling_map, i);
7489 cpus_and(sd->span, sd->span, *cpu_map);
7492 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7496 #ifdef CONFIG_SCHED_SMT
7497 /* Set up CPU (sibling) groups */
7498 for_each_cpu_mask_nr(i, *cpu_map) {
7499 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7500 SCHED_CPUMASK_VAR(send_covered, allmasks);
7502 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7503 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7504 if (i != first_cpu(*this_sibling_map))
7507 init_sched_build_groups(this_sibling_map, cpu_map,
7509 send_covered, tmpmask);
7513 #ifdef CONFIG_SCHED_MC
7514 /* Set up multi-core groups */
7515 for_each_cpu_mask_nr(i, *cpu_map) {
7516 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7517 SCHED_CPUMASK_VAR(send_covered, allmasks);
7519 *this_core_map = cpu_coregroup_map(i);
7520 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7521 if (i != first_cpu(*this_core_map))
7524 init_sched_build_groups(this_core_map, cpu_map,
7526 send_covered, tmpmask);
7530 /* Set up physical groups */
7531 for (i = 0; i < nr_node_ids; i++) {
7532 SCHED_CPUMASK_VAR(nodemask, allmasks);
7533 SCHED_CPUMASK_VAR(send_covered, allmasks);
7535 *nodemask = node_to_cpumask(i);
7536 cpus_and(*nodemask, *nodemask, *cpu_map);
7537 if (cpus_empty(*nodemask))
7540 init_sched_build_groups(nodemask, cpu_map,
7542 send_covered, tmpmask);
7546 /* Set up node groups */
7548 SCHED_CPUMASK_VAR(send_covered, allmasks);
7550 init_sched_build_groups(cpu_map, cpu_map,
7551 &cpu_to_allnodes_group,
7552 send_covered, tmpmask);
7555 for (i = 0; i < nr_node_ids; i++) {
7556 /* Set up node groups */
7557 struct sched_group *sg, *prev;
7558 SCHED_CPUMASK_VAR(nodemask, allmasks);
7559 SCHED_CPUMASK_VAR(domainspan, allmasks);
7560 SCHED_CPUMASK_VAR(covered, allmasks);
7563 *nodemask = node_to_cpumask(i);
7564 cpus_clear(*covered);
7566 cpus_and(*nodemask, *nodemask, *cpu_map);
7567 if (cpus_empty(*nodemask)) {
7568 sched_group_nodes[i] = NULL;
7572 sched_domain_node_span(i, domainspan);
7573 cpus_and(*domainspan, *domainspan, *cpu_map);
7575 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7577 printk(KERN_WARNING "Can not alloc domain group for "
7581 sched_group_nodes[i] = sg;
7582 for_each_cpu_mask_nr(j, *nodemask) {
7583 struct sched_domain *sd;
7585 sd = &per_cpu(node_domains, j);
7588 sg->__cpu_power = 0;
7589 sg->cpumask = *nodemask;
7591 cpus_or(*covered, *covered, *nodemask);
7594 for (j = 0; j < nr_node_ids; j++) {
7595 SCHED_CPUMASK_VAR(notcovered, allmasks);
7596 int n = (i + j) % nr_node_ids;
7597 node_to_cpumask_ptr(pnodemask, n);
7599 cpus_complement(*notcovered, *covered);
7600 cpus_and(*tmpmask, *notcovered, *cpu_map);
7601 cpus_and(*tmpmask, *tmpmask, *domainspan);
7602 if (cpus_empty(*tmpmask))
7605 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7606 if (cpus_empty(*tmpmask))
7609 sg = kmalloc_node(sizeof(struct sched_group),
7613 "Can not alloc domain group for node %d\n", j);
7616 sg->__cpu_power = 0;
7617 sg->cpumask = *tmpmask;
7618 sg->next = prev->next;
7619 cpus_or(*covered, *covered, *tmpmask);
7626 /* Calculate CPU power for physical packages and nodes */
7627 #ifdef CONFIG_SCHED_SMT
7628 for_each_cpu_mask_nr(i, *cpu_map) {
7629 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7631 init_sched_groups_power(i, sd);
7634 #ifdef CONFIG_SCHED_MC
7635 for_each_cpu_mask_nr(i, *cpu_map) {
7636 struct sched_domain *sd = &per_cpu(core_domains, i);
7638 init_sched_groups_power(i, sd);
7642 for_each_cpu_mask_nr(i, *cpu_map) {
7643 struct sched_domain *sd = &per_cpu(phys_domains, i);
7645 init_sched_groups_power(i, sd);
7649 for (i = 0; i < nr_node_ids; i++)
7650 init_numa_sched_groups_power(sched_group_nodes[i]);
7653 struct sched_group *sg;
7655 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7657 init_numa_sched_groups_power(sg);
7661 /* Attach the domains */
7662 for_each_cpu_mask_nr(i, *cpu_map) {
7663 struct sched_domain *sd;
7664 #ifdef CONFIG_SCHED_SMT
7665 sd = &per_cpu(cpu_domains, i);
7666 #elif defined(CONFIG_SCHED_MC)
7667 sd = &per_cpu(core_domains, i);
7669 sd = &per_cpu(phys_domains, i);
7671 cpu_attach_domain(sd, rd, i);
7674 sched_cpumask_free(allmasks);
7679 free_sched_groups(cpu_map, tmpmask);
7680 sched_cpumask_free(allmasks);
7686 static int build_sched_domains(const cpumask_t *cpu_map)
7688 return __build_sched_domains(cpu_map, NULL);
7691 static cpumask_t *doms_cur; /* current sched domains */
7692 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7693 static struct sched_domain_attr *dattr_cur;
7694 /* attribues of custom domains in 'doms_cur' */
7697 * Special case: If a kmalloc of a doms_cur partition (array of
7698 * cpumask_t) fails, then fallback to a single sched domain,
7699 * as determined by the single cpumask_t fallback_doms.
7701 static cpumask_t fallback_doms;
7704 * arch_update_cpu_topology lets virtualized architectures update the
7705 * cpu core maps. It is supposed to return 1 if the topology changed
7706 * or 0 if it stayed the same.
7708 int __attribute__((weak)) arch_update_cpu_topology(void)
7714 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7715 * For now this just excludes isolated cpus, but could be used to
7716 * exclude other special cases in the future.
7718 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7722 arch_update_cpu_topology();
7724 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7726 doms_cur = &fallback_doms;
7727 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7729 err = build_sched_domains(doms_cur);
7730 register_sched_domain_sysctl();
7735 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7738 free_sched_groups(cpu_map, tmpmask);
7742 * Detach sched domains from a group of cpus specified in cpu_map
7743 * These cpus will now be attached to the NULL domain
7745 static void detach_destroy_domains(const cpumask_t *cpu_map)
7750 for_each_cpu_mask_nr(i, *cpu_map)
7751 cpu_attach_domain(NULL, &def_root_domain, i);
7752 synchronize_sched();
7753 arch_destroy_sched_domains(cpu_map, &tmpmask);
7756 /* handle null as "default" */
7757 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7758 struct sched_domain_attr *new, int idx_new)
7760 struct sched_domain_attr tmp;
7767 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7768 new ? (new + idx_new) : &tmp,
7769 sizeof(struct sched_domain_attr));
7773 * Partition sched domains as specified by the 'ndoms_new'
7774 * cpumasks in the array doms_new[] of cpumasks. This compares
7775 * doms_new[] to the current sched domain partitioning, doms_cur[].
7776 * It destroys each deleted domain and builds each new domain.
7778 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7779 * The masks don't intersect (don't overlap.) We should setup one
7780 * sched domain for each mask. CPUs not in any of the cpumasks will
7781 * not be load balanced. If the same cpumask appears both in the
7782 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7785 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7786 * ownership of it and will kfree it when done with it. If the caller
7787 * failed the kmalloc call, then it can pass in doms_new == NULL &&
7788 * ndoms_new == 1, and partition_sched_domains() will fallback to
7789 * the single partition 'fallback_doms', it also forces the domains
7792 * If doms_new == NULL it will be replaced with cpu_online_map.
7793 * ndoms_new == 0 is a special case for destroying existing domains,
7794 * and it will not create the default domain.
7796 * Call with hotplug lock held
7798 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7799 struct sched_domain_attr *dattr_new)
7804 mutex_lock(&sched_domains_mutex);
7806 /* always unregister in case we don't destroy any domains */
7807 unregister_sched_domain_sysctl();
7809 /* Let architecture update cpu core mappings. */
7810 new_topology = arch_update_cpu_topology();
7812 n = doms_new ? ndoms_new : 0;
7814 /* Destroy deleted domains */
7815 for (i = 0; i < ndoms_cur; i++) {
7816 for (j = 0; j < n && !new_topology; j++) {
7817 if (cpus_equal(doms_cur[i], doms_new[j])
7818 && dattrs_equal(dattr_cur, i, dattr_new, j))
7821 /* no match - a current sched domain not in new doms_new[] */
7822 detach_destroy_domains(doms_cur + i);
7827 if (doms_new == NULL) {
7829 doms_new = &fallback_doms;
7830 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7831 WARN_ON_ONCE(dattr_new);
7834 /* Build new domains */
7835 for (i = 0; i < ndoms_new; i++) {
7836 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7837 if (cpus_equal(doms_new[i], doms_cur[j])
7838 && dattrs_equal(dattr_new, i, dattr_cur, j))
7841 /* no match - add a new doms_new */
7842 __build_sched_domains(doms_new + i,
7843 dattr_new ? dattr_new + i : NULL);
7848 /* Remember the new sched domains */
7849 if (doms_cur != &fallback_doms)
7851 kfree(dattr_cur); /* kfree(NULL) is safe */
7852 doms_cur = doms_new;
7853 dattr_cur = dattr_new;
7854 ndoms_cur = ndoms_new;
7856 register_sched_domain_sysctl();
7858 mutex_unlock(&sched_domains_mutex);
7861 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7862 int arch_reinit_sched_domains(void)
7866 /* Destroy domains first to force the rebuild */
7867 partition_sched_domains(0, NULL, NULL);
7869 rebuild_sched_domains();
7875 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7879 if (buf[0] != '0' && buf[0] != '1')
7883 sched_smt_power_savings = (buf[0] == '1');
7885 sched_mc_power_savings = (buf[0] == '1');
7887 ret = arch_reinit_sched_domains();
7889 return ret ? ret : count;
7892 #ifdef CONFIG_SCHED_MC
7893 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7896 return sprintf(page, "%u\n", sched_mc_power_savings);
7898 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7899 const char *buf, size_t count)
7901 return sched_power_savings_store(buf, count, 0);
7903 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7904 sched_mc_power_savings_show,
7905 sched_mc_power_savings_store);
7908 #ifdef CONFIG_SCHED_SMT
7909 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7912 return sprintf(page, "%u\n", sched_smt_power_savings);
7914 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7915 const char *buf, size_t count)
7917 return sched_power_savings_store(buf, count, 1);
7919 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7920 sched_smt_power_savings_show,
7921 sched_smt_power_savings_store);
7924 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7928 #ifdef CONFIG_SCHED_SMT
7930 err = sysfs_create_file(&cls->kset.kobj,
7931 &attr_sched_smt_power_savings.attr);
7933 #ifdef CONFIG_SCHED_MC
7934 if (!err && mc_capable())
7935 err = sysfs_create_file(&cls->kset.kobj,
7936 &attr_sched_mc_power_savings.attr);
7940 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7942 #ifndef CONFIG_CPUSETS
7944 * Add online and remove offline CPUs from the scheduler domains.
7945 * When cpusets are enabled they take over this function.
7947 static int update_sched_domains(struct notifier_block *nfb,
7948 unsigned long action, void *hcpu)
7952 case CPU_ONLINE_FROZEN:
7954 case CPU_DEAD_FROZEN:
7955 partition_sched_domains(1, NULL, NULL);
7964 static int update_runtime(struct notifier_block *nfb,
7965 unsigned long action, void *hcpu)
7967 int cpu = (int)(long)hcpu;
7970 case CPU_DOWN_PREPARE:
7971 case CPU_DOWN_PREPARE_FROZEN:
7972 disable_runtime(cpu_rq(cpu));
7975 case CPU_DOWN_FAILED:
7976 case CPU_DOWN_FAILED_FROZEN:
7978 case CPU_ONLINE_FROZEN:
7979 enable_runtime(cpu_rq(cpu));
7987 void __init sched_init_smp(void)
7989 cpumask_t non_isolated_cpus;
7991 #if defined(CONFIG_NUMA)
7992 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7994 BUG_ON(sched_group_nodes_bycpu == NULL);
7997 mutex_lock(&sched_domains_mutex);
7998 arch_init_sched_domains(&cpu_online_map);
7999 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
8000 if (cpus_empty(non_isolated_cpus))
8001 cpu_set(smp_processor_id(), non_isolated_cpus);
8002 mutex_unlock(&sched_domains_mutex);
8005 #ifndef CONFIG_CPUSETS
8006 /* XXX: Theoretical race here - CPU may be hotplugged now */
8007 hotcpu_notifier(update_sched_domains, 0);
8010 /* RT runtime code needs to handle some hotplug events */
8011 hotcpu_notifier(update_runtime, 0);
8015 /* Move init over to a non-isolated CPU */
8016 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
8018 sched_init_granularity();
8021 void __init sched_init_smp(void)
8023 sched_init_granularity();
8025 #endif /* CONFIG_SMP */
8027 int in_sched_functions(unsigned long addr)
8029 return in_lock_functions(addr) ||
8030 (addr >= (unsigned long)__sched_text_start
8031 && addr < (unsigned long)__sched_text_end);
8034 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
8036 cfs_rq->tasks_timeline = RB_ROOT;
8037 INIT_LIST_HEAD(&cfs_rq->tasks);
8038 #ifdef CONFIG_FAIR_GROUP_SCHED
8041 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8044 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8046 struct rt_prio_array *array;
8049 array = &rt_rq->active;
8050 for (i = 0; i < MAX_RT_PRIO; i++) {
8051 INIT_LIST_HEAD(array->queue + i);
8052 __clear_bit(i, array->bitmap);
8054 /* delimiter for bitsearch: */
8055 __set_bit(MAX_RT_PRIO, array->bitmap);
8057 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8058 rt_rq->highest_prio = MAX_RT_PRIO;
8061 rt_rq->rt_nr_migratory = 0;
8062 rt_rq->overloaded = 0;
8066 rt_rq->rt_throttled = 0;
8067 rt_rq->rt_runtime = 0;
8068 spin_lock_init(&rt_rq->rt_runtime_lock);
8070 #ifdef CONFIG_RT_GROUP_SCHED
8071 rt_rq->rt_nr_boosted = 0;
8076 #ifdef CONFIG_FAIR_GROUP_SCHED
8077 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8078 struct sched_entity *se, int cpu, int add,
8079 struct sched_entity *parent)
8081 struct rq *rq = cpu_rq(cpu);
8082 tg->cfs_rq[cpu] = cfs_rq;
8083 init_cfs_rq(cfs_rq, rq);
8086 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8089 /* se could be NULL for init_task_group */
8094 se->cfs_rq = &rq->cfs;
8096 se->cfs_rq = parent->my_q;
8099 se->load.weight = tg->shares;
8100 se->load.inv_weight = 0;
8101 se->parent = parent;
8105 #ifdef CONFIG_RT_GROUP_SCHED
8106 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8107 struct sched_rt_entity *rt_se, int cpu, int add,
8108 struct sched_rt_entity *parent)
8110 struct rq *rq = cpu_rq(cpu);
8112 tg->rt_rq[cpu] = rt_rq;
8113 init_rt_rq(rt_rq, rq);
8115 rt_rq->rt_se = rt_se;
8116 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8118 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8120 tg->rt_se[cpu] = rt_se;
8125 rt_se->rt_rq = &rq->rt;
8127 rt_se->rt_rq = parent->my_q;
8129 rt_se->my_q = rt_rq;
8130 rt_se->parent = parent;
8131 INIT_LIST_HEAD(&rt_se->run_list);
8135 void __init sched_init(void)
8138 unsigned long alloc_size = 0, ptr;
8140 #ifdef CONFIG_FAIR_GROUP_SCHED
8141 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8143 #ifdef CONFIG_RT_GROUP_SCHED
8144 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8146 #ifdef CONFIG_USER_SCHED
8150 * As sched_init() is called before page_alloc is setup,
8151 * we use alloc_bootmem().
8154 ptr = (unsigned long)alloc_bootmem(alloc_size);
8156 #ifdef CONFIG_FAIR_GROUP_SCHED
8157 init_task_group.se = (struct sched_entity **)ptr;
8158 ptr += nr_cpu_ids * sizeof(void **);
8160 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8161 ptr += nr_cpu_ids * sizeof(void **);
8163 #ifdef CONFIG_USER_SCHED
8164 root_task_group.se = (struct sched_entity **)ptr;
8165 ptr += nr_cpu_ids * sizeof(void **);
8167 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8168 ptr += nr_cpu_ids * sizeof(void **);
8169 #endif /* CONFIG_USER_SCHED */
8170 #endif /* CONFIG_FAIR_GROUP_SCHED */
8171 #ifdef CONFIG_RT_GROUP_SCHED
8172 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8173 ptr += nr_cpu_ids * sizeof(void **);
8175 init_task_group.rt_rq = (struct rt_rq **)ptr;
8176 ptr += nr_cpu_ids * sizeof(void **);
8178 #ifdef CONFIG_USER_SCHED
8179 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8180 ptr += nr_cpu_ids * sizeof(void **);
8182 root_task_group.rt_rq = (struct rt_rq **)ptr;
8183 ptr += nr_cpu_ids * sizeof(void **);
8184 #endif /* CONFIG_USER_SCHED */
8185 #endif /* CONFIG_RT_GROUP_SCHED */
8189 init_defrootdomain();
8192 init_rt_bandwidth(&def_rt_bandwidth,
8193 global_rt_period(), global_rt_runtime());
8195 #ifdef CONFIG_RT_GROUP_SCHED
8196 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8197 global_rt_period(), global_rt_runtime());
8198 #ifdef CONFIG_USER_SCHED
8199 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8200 global_rt_period(), RUNTIME_INF);
8201 #endif /* CONFIG_USER_SCHED */
8202 #endif /* CONFIG_RT_GROUP_SCHED */
8204 #ifdef CONFIG_GROUP_SCHED
8205 list_add(&init_task_group.list, &task_groups);
8206 INIT_LIST_HEAD(&init_task_group.children);
8208 #ifdef CONFIG_USER_SCHED
8209 INIT_LIST_HEAD(&root_task_group.children);
8210 init_task_group.parent = &root_task_group;
8211 list_add(&init_task_group.siblings, &root_task_group.children);
8212 #endif /* CONFIG_USER_SCHED */
8213 #endif /* CONFIG_GROUP_SCHED */
8215 for_each_possible_cpu(i) {
8219 spin_lock_init(&rq->lock);
8221 init_cfs_rq(&rq->cfs, rq);
8222 init_rt_rq(&rq->rt, rq);
8223 #ifdef CONFIG_FAIR_GROUP_SCHED
8224 init_task_group.shares = init_task_group_load;
8225 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8226 #ifdef CONFIG_CGROUP_SCHED
8228 * How much cpu bandwidth does init_task_group get?
8230 * In case of task-groups formed thr' the cgroup filesystem, it
8231 * gets 100% of the cpu resources in the system. This overall
8232 * system cpu resource is divided among the tasks of
8233 * init_task_group and its child task-groups in a fair manner,
8234 * based on each entity's (task or task-group's) weight
8235 * (se->load.weight).
8237 * In other words, if init_task_group has 10 tasks of weight
8238 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8239 * then A0's share of the cpu resource is:
8241 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8243 * We achieve this by letting init_task_group's tasks sit
8244 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8246 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8247 #elif defined CONFIG_USER_SCHED
8248 root_task_group.shares = NICE_0_LOAD;
8249 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8251 * In case of task-groups formed thr' the user id of tasks,
8252 * init_task_group represents tasks belonging to root user.
8253 * Hence it forms a sibling of all subsequent groups formed.
8254 * In this case, init_task_group gets only a fraction of overall
8255 * system cpu resource, based on the weight assigned to root
8256 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8257 * by letting tasks of init_task_group sit in a separate cfs_rq
8258 * (init_cfs_rq) and having one entity represent this group of
8259 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8261 init_tg_cfs_entry(&init_task_group,
8262 &per_cpu(init_cfs_rq, i),
8263 &per_cpu(init_sched_entity, i), i, 1,
8264 root_task_group.se[i]);
8267 #endif /* CONFIG_FAIR_GROUP_SCHED */
8269 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8270 #ifdef CONFIG_RT_GROUP_SCHED
8271 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8272 #ifdef CONFIG_CGROUP_SCHED
8273 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8274 #elif defined CONFIG_USER_SCHED
8275 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8276 init_tg_rt_entry(&init_task_group,
8277 &per_cpu(init_rt_rq, i),
8278 &per_cpu(init_sched_rt_entity, i), i, 1,
8279 root_task_group.rt_se[i]);
8283 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8284 rq->cpu_load[j] = 0;
8288 rq->active_balance = 0;
8289 rq->next_balance = jiffies;
8293 rq->migration_thread = NULL;
8294 INIT_LIST_HEAD(&rq->migration_queue);
8295 rq_attach_root(rq, &def_root_domain);
8298 atomic_set(&rq->nr_iowait, 0);
8301 set_load_weight(&init_task);
8303 #ifdef CONFIG_PREEMPT_NOTIFIERS
8304 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8308 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8311 #ifdef CONFIG_RT_MUTEXES
8312 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8316 * The boot idle thread does lazy MMU switching as well:
8318 atomic_inc(&init_mm.mm_count);
8319 enter_lazy_tlb(&init_mm, current);
8322 * Make us the idle thread. Technically, schedule() should not be
8323 * called from this thread, however somewhere below it might be,
8324 * but because we are the idle thread, we just pick up running again
8325 * when this runqueue becomes "idle".
8327 init_idle(current, smp_processor_id());
8329 * During early bootup we pretend to be a normal task:
8331 current->sched_class = &fair_sched_class;
8333 scheduler_running = 1;
8336 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8337 void __might_sleep(char *file, int line)
8340 static unsigned long prev_jiffy; /* ratelimiting */
8342 if ((!in_atomic() && !irqs_disabled()) ||
8343 system_state != SYSTEM_RUNNING || oops_in_progress)
8345 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8347 prev_jiffy = jiffies;
8350 "BUG: sleeping function called from invalid context at %s:%d\n",
8353 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8354 in_atomic(), irqs_disabled(),
8355 current->pid, current->comm);
8357 debug_show_held_locks(current);
8358 if (irqs_disabled())
8359 print_irqtrace_events(current);
8363 EXPORT_SYMBOL(__might_sleep);
8366 #ifdef CONFIG_MAGIC_SYSRQ
8367 static void normalize_task(struct rq *rq, struct task_struct *p)
8371 update_rq_clock(rq);
8372 on_rq = p->se.on_rq;
8374 deactivate_task(rq, p, 0);
8375 __setscheduler(rq, p, SCHED_NORMAL, 0);
8377 activate_task(rq, p, 0);
8378 resched_task(rq->curr);
8382 void normalize_rt_tasks(void)
8384 struct task_struct *g, *p;
8385 unsigned long flags;
8388 read_lock_irqsave(&tasklist_lock, flags);
8389 do_each_thread(g, p) {
8391 * Only normalize user tasks:
8396 p->se.exec_start = 0;
8397 #ifdef CONFIG_SCHEDSTATS
8398 p->se.wait_start = 0;
8399 p->se.sleep_start = 0;
8400 p->se.block_start = 0;
8405 * Renice negative nice level userspace
8408 if (TASK_NICE(p) < 0 && p->mm)
8409 set_user_nice(p, 0);
8413 spin_lock(&p->pi_lock);
8414 rq = __task_rq_lock(p);
8416 normalize_task(rq, p);
8418 __task_rq_unlock(rq);
8419 spin_unlock(&p->pi_lock);
8420 } while_each_thread(g, p);
8422 read_unlock_irqrestore(&tasklist_lock, flags);
8425 #endif /* CONFIG_MAGIC_SYSRQ */
8429 * These functions are only useful for the IA64 MCA handling.
8431 * They can only be called when the whole system has been
8432 * stopped - every CPU needs to be quiescent, and no scheduling
8433 * activity can take place. Using them for anything else would
8434 * be a serious bug, and as a result, they aren't even visible
8435 * under any other configuration.
8439 * curr_task - return the current task for a given cpu.
8440 * @cpu: the processor in question.
8442 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8444 struct task_struct *curr_task(int cpu)
8446 return cpu_curr(cpu);
8450 * set_curr_task - set the current task for a given cpu.
8451 * @cpu: the processor in question.
8452 * @p: the task pointer to set.
8454 * Description: This function must only be used when non-maskable interrupts
8455 * are serviced on a separate stack. It allows the architecture to switch the
8456 * notion of the current task on a cpu in a non-blocking manner. This function
8457 * must be called with all CPU's synchronized, and interrupts disabled, the
8458 * and caller must save the original value of the current task (see
8459 * curr_task() above) and restore that value before reenabling interrupts and
8460 * re-starting the system.
8462 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8464 void set_curr_task(int cpu, struct task_struct *p)
8471 #ifdef CONFIG_FAIR_GROUP_SCHED
8472 static void free_fair_sched_group(struct task_group *tg)
8476 for_each_possible_cpu(i) {
8478 kfree(tg->cfs_rq[i]);
8488 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8490 struct cfs_rq *cfs_rq;
8491 struct sched_entity *se;
8495 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8498 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8502 tg->shares = NICE_0_LOAD;
8504 for_each_possible_cpu(i) {
8507 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8508 GFP_KERNEL, cpu_to_node(i));
8512 se = kzalloc_node(sizeof(struct sched_entity),
8513 GFP_KERNEL, cpu_to_node(i));
8517 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8526 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8528 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8529 &cpu_rq(cpu)->leaf_cfs_rq_list);
8532 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8534 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8536 #else /* !CONFG_FAIR_GROUP_SCHED */
8537 static inline void free_fair_sched_group(struct task_group *tg)
8542 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8547 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8551 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8554 #endif /* CONFIG_FAIR_GROUP_SCHED */
8556 #ifdef CONFIG_RT_GROUP_SCHED
8557 static void free_rt_sched_group(struct task_group *tg)
8561 destroy_rt_bandwidth(&tg->rt_bandwidth);
8563 for_each_possible_cpu(i) {
8565 kfree(tg->rt_rq[i]);
8567 kfree(tg->rt_se[i]);
8575 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8577 struct rt_rq *rt_rq;
8578 struct sched_rt_entity *rt_se;
8582 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8585 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8589 init_rt_bandwidth(&tg->rt_bandwidth,
8590 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8592 for_each_possible_cpu(i) {
8595 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8596 GFP_KERNEL, cpu_to_node(i));
8600 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8601 GFP_KERNEL, cpu_to_node(i));
8605 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8614 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8616 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8617 &cpu_rq(cpu)->leaf_rt_rq_list);
8620 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8622 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8624 #else /* !CONFIG_RT_GROUP_SCHED */
8625 static inline void free_rt_sched_group(struct task_group *tg)
8630 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8635 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8639 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8642 #endif /* CONFIG_RT_GROUP_SCHED */
8644 #ifdef CONFIG_GROUP_SCHED
8645 static void free_sched_group(struct task_group *tg)
8647 free_fair_sched_group(tg);
8648 free_rt_sched_group(tg);
8652 /* allocate runqueue etc for a new task group */
8653 struct task_group *sched_create_group(struct task_group *parent)
8655 struct task_group *tg;
8656 unsigned long flags;
8659 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8661 return ERR_PTR(-ENOMEM);
8663 if (!alloc_fair_sched_group(tg, parent))
8666 if (!alloc_rt_sched_group(tg, parent))
8669 spin_lock_irqsave(&task_group_lock, flags);
8670 for_each_possible_cpu(i) {
8671 register_fair_sched_group(tg, i);
8672 register_rt_sched_group(tg, i);
8674 list_add_rcu(&tg->list, &task_groups);
8676 WARN_ON(!parent); /* root should already exist */
8678 tg->parent = parent;
8679 INIT_LIST_HEAD(&tg->children);
8680 list_add_rcu(&tg->siblings, &parent->children);
8681 spin_unlock_irqrestore(&task_group_lock, flags);
8686 free_sched_group(tg);
8687 return ERR_PTR(-ENOMEM);
8690 /* rcu callback to free various structures associated with a task group */
8691 static void free_sched_group_rcu(struct rcu_head *rhp)
8693 /* now it should be safe to free those cfs_rqs */
8694 free_sched_group(container_of(rhp, struct task_group, rcu));
8697 /* Destroy runqueue etc associated with a task group */
8698 void sched_destroy_group(struct task_group *tg)
8700 unsigned long flags;
8703 spin_lock_irqsave(&task_group_lock, flags);
8704 for_each_possible_cpu(i) {
8705 unregister_fair_sched_group(tg, i);
8706 unregister_rt_sched_group(tg, i);
8708 list_del_rcu(&tg->list);
8709 list_del_rcu(&tg->siblings);
8710 spin_unlock_irqrestore(&task_group_lock, flags);
8712 /* wait for possible concurrent references to cfs_rqs complete */
8713 call_rcu(&tg->rcu, free_sched_group_rcu);
8716 /* change task's runqueue when it moves between groups.
8717 * The caller of this function should have put the task in its new group
8718 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8719 * reflect its new group.
8721 void sched_move_task(struct task_struct *tsk)
8724 unsigned long flags;
8727 rq = task_rq_lock(tsk, &flags);
8729 update_rq_clock(rq);
8731 running = task_current(rq, tsk);
8732 on_rq = tsk->se.on_rq;
8735 dequeue_task(rq, tsk, 0);
8736 if (unlikely(running))
8737 tsk->sched_class->put_prev_task(rq, tsk);
8739 set_task_rq(tsk, task_cpu(tsk));
8741 #ifdef CONFIG_FAIR_GROUP_SCHED
8742 if (tsk->sched_class->moved_group)
8743 tsk->sched_class->moved_group(tsk);
8746 if (unlikely(running))
8747 tsk->sched_class->set_curr_task(rq);
8749 enqueue_task(rq, tsk, 0);
8751 task_rq_unlock(rq, &flags);
8753 #endif /* CONFIG_GROUP_SCHED */
8755 #ifdef CONFIG_FAIR_GROUP_SCHED
8756 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8758 struct cfs_rq *cfs_rq = se->cfs_rq;
8763 dequeue_entity(cfs_rq, se, 0);
8765 se->load.weight = shares;
8766 se->load.inv_weight = 0;
8769 enqueue_entity(cfs_rq, se, 0);
8772 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8774 struct cfs_rq *cfs_rq = se->cfs_rq;
8775 struct rq *rq = cfs_rq->rq;
8776 unsigned long flags;
8778 spin_lock_irqsave(&rq->lock, flags);
8779 __set_se_shares(se, shares);
8780 spin_unlock_irqrestore(&rq->lock, flags);
8783 static DEFINE_MUTEX(shares_mutex);
8785 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8788 unsigned long flags;
8791 * We can't change the weight of the root cgroup.
8796 if (shares < MIN_SHARES)
8797 shares = MIN_SHARES;
8798 else if (shares > MAX_SHARES)
8799 shares = MAX_SHARES;
8801 mutex_lock(&shares_mutex);
8802 if (tg->shares == shares)
8805 spin_lock_irqsave(&task_group_lock, flags);
8806 for_each_possible_cpu(i)
8807 unregister_fair_sched_group(tg, i);
8808 list_del_rcu(&tg->siblings);
8809 spin_unlock_irqrestore(&task_group_lock, flags);
8811 /* wait for any ongoing reference to this group to finish */
8812 synchronize_sched();
8815 * Now we are free to modify the group's share on each cpu
8816 * w/o tripping rebalance_share or load_balance_fair.
8818 tg->shares = shares;
8819 for_each_possible_cpu(i) {
8823 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8824 set_se_shares(tg->se[i], shares);
8828 * Enable load balance activity on this group, by inserting it back on
8829 * each cpu's rq->leaf_cfs_rq_list.
8831 spin_lock_irqsave(&task_group_lock, flags);
8832 for_each_possible_cpu(i)
8833 register_fair_sched_group(tg, i);
8834 list_add_rcu(&tg->siblings, &tg->parent->children);
8835 spin_unlock_irqrestore(&task_group_lock, flags);
8837 mutex_unlock(&shares_mutex);
8841 unsigned long sched_group_shares(struct task_group *tg)
8847 #ifdef CONFIG_RT_GROUP_SCHED
8849 * Ensure that the real time constraints are schedulable.
8851 static DEFINE_MUTEX(rt_constraints_mutex);
8853 static unsigned long to_ratio(u64 period, u64 runtime)
8855 if (runtime == RUNTIME_INF)
8858 return div64_u64(runtime << 20, period);
8861 /* Must be called with tasklist_lock held */
8862 static inline int tg_has_rt_tasks(struct task_group *tg)
8864 struct task_struct *g, *p;
8866 do_each_thread(g, p) {
8867 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8869 } while_each_thread(g, p);
8874 struct rt_schedulable_data {
8875 struct task_group *tg;
8880 static int tg_schedulable(struct task_group *tg, void *data)
8882 struct rt_schedulable_data *d = data;
8883 struct task_group *child;
8884 unsigned long total, sum = 0;
8885 u64 period, runtime;
8887 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8888 runtime = tg->rt_bandwidth.rt_runtime;
8891 period = d->rt_period;
8892 runtime = d->rt_runtime;
8896 * Cannot have more runtime than the period.
8898 if (runtime > period && runtime != RUNTIME_INF)
8902 * Ensure we don't starve existing RT tasks.
8904 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8907 total = to_ratio(period, runtime);
8910 * Nobody can have more than the global setting allows.
8912 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8916 * The sum of our children's runtime should not exceed our own.
8918 list_for_each_entry_rcu(child, &tg->children, siblings) {
8919 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8920 runtime = child->rt_bandwidth.rt_runtime;
8922 if (child == d->tg) {
8923 period = d->rt_period;
8924 runtime = d->rt_runtime;
8927 sum += to_ratio(period, runtime);
8936 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8938 struct rt_schedulable_data data = {
8940 .rt_period = period,
8941 .rt_runtime = runtime,
8944 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8947 static int tg_set_bandwidth(struct task_group *tg,
8948 u64 rt_period, u64 rt_runtime)
8952 mutex_lock(&rt_constraints_mutex);
8953 read_lock(&tasklist_lock);
8954 err = __rt_schedulable(tg, rt_period, rt_runtime);
8958 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8959 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8960 tg->rt_bandwidth.rt_runtime = rt_runtime;
8962 for_each_possible_cpu(i) {
8963 struct rt_rq *rt_rq = tg->rt_rq[i];
8965 spin_lock(&rt_rq->rt_runtime_lock);
8966 rt_rq->rt_runtime = rt_runtime;
8967 spin_unlock(&rt_rq->rt_runtime_lock);
8969 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8971 read_unlock(&tasklist_lock);
8972 mutex_unlock(&rt_constraints_mutex);
8977 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8979 u64 rt_runtime, rt_period;
8981 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8982 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8983 if (rt_runtime_us < 0)
8984 rt_runtime = RUNTIME_INF;
8986 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8989 long sched_group_rt_runtime(struct task_group *tg)
8993 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8996 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8997 do_div(rt_runtime_us, NSEC_PER_USEC);
8998 return rt_runtime_us;
9001 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
9003 u64 rt_runtime, rt_period;
9005 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
9006 rt_runtime = tg->rt_bandwidth.rt_runtime;
9011 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9014 long sched_group_rt_period(struct task_group *tg)
9018 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9019 do_div(rt_period_us, NSEC_PER_USEC);
9020 return rt_period_us;
9023 static int sched_rt_global_constraints(void)
9025 u64 runtime, period;
9028 if (sysctl_sched_rt_period <= 0)
9031 runtime = global_rt_runtime();
9032 period = global_rt_period();
9035 * Sanity check on the sysctl variables.
9037 if (runtime > period && runtime != RUNTIME_INF)
9040 mutex_lock(&rt_constraints_mutex);
9041 read_lock(&tasklist_lock);
9042 ret = __rt_schedulable(NULL, 0, 0);
9043 read_unlock(&tasklist_lock);
9044 mutex_unlock(&rt_constraints_mutex);
9048 #else /* !CONFIG_RT_GROUP_SCHED */
9049 static int sched_rt_global_constraints(void)
9051 unsigned long flags;
9054 if (sysctl_sched_rt_period <= 0)
9057 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9058 for_each_possible_cpu(i) {
9059 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9061 spin_lock(&rt_rq->rt_runtime_lock);
9062 rt_rq->rt_runtime = global_rt_runtime();
9063 spin_unlock(&rt_rq->rt_runtime_lock);
9065 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9069 #endif /* CONFIG_RT_GROUP_SCHED */
9071 int sched_rt_handler(struct ctl_table *table, int write,
9072 struct file *filp, void __user *buffer, size_t *lenp,
9076 int old_period, old_runtime;
9077 static DEFINE_MUTEX(mutex);
9080 old_period = sysctl_sched_rt_period;
9081 old_runtime = sysctl_sched_rt_runtime;
9083 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
9085 if (!ret && write) {
9086 ret = sched_rt_global_constraints();
9088 sysctl_sched_rt_period = old_period;
9089 sysctl_sched_rt_runtime = old_runtime;
9091 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9092 def_rt_bandwidth.rt_period =
9093 ns_to_ktime(global_rt_period());
9096 mutex_unlock(&mutex);
9101 #ifdef CONFIG_CGROUP_SCHED
9103 /* return corresponding task_group object of a cgroup */
9104 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9106 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9107 struct task_group, css);
9110 static struct cgroup_subsys_state *
9111 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9113 struct task_group *tg, *parent;
9115 if (!cgrp->parent) {
9116 /* This is early initialization for the top cgroup */
9117 return &init_task_group.css;
9120 parent = cgroup_tg(cgrp->parent);
9121 tg = sched_create_group(parent);
9123 return ERR_PTR(-ENOMEM);
9129 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9131 struct task_group *tg = cgroup_tg(cgrp);
9133 sched_destroy_group(tg);
9137 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9138 struct task_struct *tsk)
9140 #ifdef CONFIG_RT_GROUP_SCHED
9141 /* Don't accept realtime tasks when there is no way for them to run */
9142 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
9145 /* We don't support RT-tasks being in separate groups */
9146 if (tsk->sched_class != &fair_sched_class)
9154 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9155 struct cgroup *old_cont, struct task_struct *tsk)
9157 sched_move_task(tsk);
9160 #ifdef CONFIG_FAIR_GROUP_SCHED
9161 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9164 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9167 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9169 struct task_group *tg = cgroup_tg(cgrp);
9171 return (u64) tg->shares;
9173 #endif /* CONFIG_FAIR_GROUP_SCHED */
9175 #ifdef CONFIG_RT_GROUP_SCHED
9176 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9179 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9182 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9184 return sched_group_rt_runtime(cgroup_tg(cgrp));
9187 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9190 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9193 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9195 return sched_group_rt_period(cgroup_tg(cgrp));
9197 #endif /* CONFIG_RT_GROUP_SCHED */
9199 static struct cftype cpu_files[] = {
9200 #ifdef CONFIG_FAIR_GROUP_SCHED
9203 .read_u64 = cpu_shares_read_u64,
9204 .write_u64 = cpu_shares_write_u64,
9207 #ifdef CONFIG_RT_GROUP_SCHED
9209 .name = "rt_runtime_us",
9210 .read_s64 = cpu_rt_runtime_read,
9211 .write_s64 = cpu_rt_runtime_write,
9214 .name = "rt_period_us",
9215 .read_u64 = cpu_rt_period_read_uint,
9216 .write_u64 = cpu_rt_period_write_uint,
9221 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9223 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9226 struct cgroup_subsys cpu_cgroup_subsys = {
9228 .create = cpu_cgroup_create,
9229 .destroy = cpu_cgroup_destroy,
9230 .can_attach = cpu_cgroup_can_attach,
9231 .attach = cpu_cgroup_attach,
9232 .populate = cpu_cgroup_populate,
9233 .subsys_id = cpu_cgroup_subsys_id,
9237 #endif /* CONFIG_CGROUP_SCHED */
9239 #ifdef CONFIG_CGROUP_CPUACCT
9242 * CPU accounting code for task groups.
9244 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9245 * (balbir@in.ibm.com).
9248 /* track cpu usage of a group of tasks and its child groups */
9250 struct cgroup_subsys_state css;
9251 /* cpuusage holds pointer to a u64-type object on every cpu */
9253 struct cpuacct *parent;
9256 struct cgroup_subsys cpuacct_subsys;
9258 /* return cpu accounting group corresponding to this container */
9259 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9261 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9262 struct cpuacct, css);
9265 /* return cpu accounting group to which this task belongs */
9266 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9268 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9269 struct cpuacct, css);
9272 /* create a new cpu accounting group */
9273 static struct cgroup_subsys_state *cpuacct_create(
9274 struct cgroup_subsys *ss, struct cgroup *cgrp)
9276 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9279 return ERR_PTR(-ENOMEM);
9281 ca->cpuusage = alloc_percpu(u64);
9282 if (!ca->cpuusage) {
9284 return ERR_PTR(-ENOMEM);
9288 ca->parent = cgroup_ca(cgrp->parent);
9293 /* destroy an existing cpu accounting group */
9295 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9297 struct cpuacct *ca = cgroup_ca(cgrp);
9299 free_percpu(ca->cpuusage);
9303 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9305 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9308 #ifndef CONFIG_64BIT
9310 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9312 spin_lock_irq(&cpu_rq(cpu)->lock);
9314 spin_unlock_irq(&cpu_rq(cpu)->lock);
9322 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9324 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9326 #ifndef CONFIG_64BIT
9328 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9330 spin_lock_irq(&cpu_rq(cpu)->lock);
9332 spin_unlock_irq(&cpu_rq(cpu)->lock);
9338 /* return total cpu usage (in nanoseconds) of a group */
9339 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9341 struct cpuacct *ca = cgroup_ca(cgrp);
9342 u64 totalcpuusage = 0;
9345 for_each_present_cpu(i)
9346 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9348 return totalcpuusage;
9351 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9354 struct cpuacct *ca = cgroup_ca(cgrp);
9363 for_each_present_cpu(i)
9364 cpuacct_cpuusage_write(ca, i, 0);
9370 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9373 struct cpuacct *ca = cgroup_ca(cgroup);
9377 for_each_present_cpu(i) {
9378 percpu = cpuacct_cpuusage_read(ca, i);
9379 seq_printf(m, "%llu ", (unsigned long long) percpu);
9381 seq_printf(m, "\n");
9385 static struct cftype files[] = {
9388 .read_u64 = cpuusage_read,
9389 .write_u64 = cpuusage_write,
9392 .name = "usage_percpu",
9393 .read_seq_string = cpuacct_percpu_seq_read,
9398 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9400 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9404 * charge this task's execution time to its accounting group.
9406 * called with rq->lock held.
9408 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9413 if (!cpuacct_subsys.active)
9416 cpu = task_cpu(tsk);
9419 for (; ca; ca = ca->parent) {
9420 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9421 *cpuusage += cputime;
9425 struct cgroup_subsys cpuacct_subsys = {
9427 .create = cpuacct_create,
9428 .destroy = cpuacct_destroy,
9429 .populate = cpuacct_populate,
9430 .subsys_id = cpuacct_subsys_id,
9432 #endif /* CONFIG_CGROUP_CPUACCT */