4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/reciprocal_div.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/bootmem.h>
72 #include <linux/debugfs.h>
73 #include <linux/ctype.h>
74 #include <linux/ftrace.h>
75 #include <trace/sched.h>
78 #include <asm/irq_regs.h>
80 #include "sched_cpupri.h"
83 * Convert user-nice values [ -20 ... 0 ... 19 ]
84 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
87 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
88 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
89 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
92 * 'User priority' is the nice value converted to something we
93 * can work with better when scaling various scheduler parameters,
94 * it's a [ 0 ... 39 ] range.
96 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
97 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
98 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
101 * Helpers for converting nanosecond timing to jiffy resolution
103 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
105 #define NICE_0_LOAD SCHED_LOAD_SCALE
106 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
109 * These are the 'tuning knobs' of the scheduler:
111 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
112 * Timeslices get refilled after they expire.
114 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * single value that denotes runtime == period, ie unlimited time.
119 #define RUNTIME_INF ((u64)~0ULL)
121 DEFINE_TRACE(sched_wait_task);
122 DEFINE_TRACE(sched_wakeup);
123 DEFINE_TRACE(sched_wakeup_new);
124 DEFINE_TRACE(sched_switch);
125 DEFINE_TRACE(sched_migrate_task);
129 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
130 * Since cpu_power is a 'constant', we can use a reciprocal divide.
132 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
134 return reciprocal_divide(load, sg->reciprocal_cpu_power);
138 * Each time a sched group cpu_power is changed,
139 * we must compute its reciprocal value
141 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
143 sg->__cpu_power += val;
144 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
148 static inline int rt_policy(int policy)
150 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
155 static inline int task_has_rt_policy(struct task_struct *p)
157 return rt_policy(p->policy);
161 * This is the priority-queue data structure of the RT scheduling class:
163 struct rt_prio_array {
164 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
165 struct list_head queue[MAX_RT_PRIO];
168 struct rt_bandwidth {
169 /* nests inside the rq lock: */
170 spinlock_t rt_runtime_lock;
173 struct hrtimer rt_period_timer;
176 static struct rt_bandwidth def_rt_bandwidth;
178 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
180 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
182 struct rt_bandwidth *rt_b =
183 container_of(timer, struct rt_bandwidth, rt_period_timer);
189 now = hrtimer_cb_get_time(timer);
190 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
195 idle = do_sched_rt_period_timer(rt_b, overrun);
198 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
202 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
204 rt_b->rt_period = ns_to_ktime(period);
205 rt_b->rt_runtime = runtime;
207 spin_lock_init(&rt_b->rt_runtime_lock);
209 hrtimer_init(&rt_b->rt_period_timer,
210 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
211 rt_b->rt_period_timer.function = sched_rt_period_timer;
212 rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_UNLOCKED;
215 static inline int rt_bandwidth_enabled(void)
217 return sysctl_sched_rt_runtime >= 0;
220 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
224 if (rt_bandwidth_enabled() && rt_b->rt_runtime == RUNTIME_INF)
227 if (hrtimer_active(&rt_b->rt_period_timer))
230 spin_lock(&rt_b->rt_runtime_lock);
232 if (hrtimer_active(&rt_b->rt_period_timer))
235 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
236 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
237 hrtimer_start_expires(&rt_b->rt_period_timer,
240 spin_unlock(&rt_b->rt_runtime_lock);
243 #ifdef CONFIG_RT_GROUP_SCHED
244 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
246 hrtimer_cancel(&rt_b->rt_period_timer);
251 * sched_domains_mutex serializes calls to arch_init_sched_domains,
252 * detach_destroy_domains and partition_sched_domains.
254 static DEFINE_MUTEX(sched_domains_mutex);
256 #ifdef CONFIG_GROUP_SCHED
258 #include <linux/cgroup.h>
262 static LIST_HEAD(task_groups);
264 /* task group related information */
266 #ifdef CONFIG_CGROUP_SCHED
267 struct cgroup_subsys_state css;
270 #ifdef CONFIG_FAIR_GROUP_SCHED
271 /* schedulable entities of this group on each cpu */
272 struct sched_entity **se;
273 /* runqueue "owned" by this group on each cpu */
274 struct cfs_rq **cfs_rq;
275 unsigned long shares;
278 #ifdef CONFIG_RT_GROUP_SCHED
279 struct sched_rt_entity **rt_se;
280 struct rt_rq **rt_rq;
282 struct rt_bandwidth rt_bandwidth;
286 struct list_head list;
288 struct task_group *parent;
289 struct list_head siblings;
290 struct list_head children;
293 #ifdef CONFIG_USER_SCHED
297 * Every UID task group (including init_task_group aka UID-0) will
298 * be a child to this group.
300 struct task_group root_task_group;
302 #ifdef CONFIG_FAIR_GROUP_SCHED
303 /* Default task group's sched entity on each cpu */
304 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
305 /* Default task group's cfs_rq on each cpu */
306 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
307 #endif /* CONFIG_FAIR_GROUP_SCHED */
309 #ifdef CONFIG_RT_GROUP_SCHED
310 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
311 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
312 #endif /* CONFIG_RT_GROUP_SCHED */
313 #else /* !CONFIG_USER_SCHED */
314 #define root_task_group init_task_group
315 #endif /* CONFIG_USER_SCHED */
317 /* task_group_lock serializes add/remove of task groups and also changes to
318 * a task group's cpu shares.
320 static DEFINE_SPINLOCK(task_group_lock);
322 #ifdef CONFIG_FAIR_GROUP_SCHED
323 #ifdef CONFIG_USER_SCHED
324 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
325 #else /* !CONFIG_USER_SCHED */
326 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
327 #endif /* CONFIG_USER_SCHED */
330 * A weight of 0 or 1 can cause arithmetics problems.
331 * A weight of a cfs_rq is the sum of weights of which entities
332 * are queued on this cfs_rq, so a weight of a entity should not be
333 * too large, so as the shares value of a task group.
334 * (The default weight is 1024 - so there's no practical
335 * limitation from this.)
338 #define MAX_SHARES (1UL << 18)
340 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
343 /* Default task group.
344 * Every task in system belong to this group at bootup.
346 struct task_group init_task_group;
348 /* return group to which a task belongs */
349 static inline struct task_group *task_group(struct task_struct *p)
351 struct task_group *tg;
353 #ifdef CONFIG_USER_SCHED
355 #elif defined(CONFIG_CGROUP_SCHED)
356 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
357 struct task_group, css);
359 tg = &init_task_group;
364 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
365 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
367 #ifdef CONFIG_FAIR_GROUP_SCHED
368 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
369 p->se.parent = task_group(p)->se[cpu];
372 #ifdef CONFIG_RT_GROUP_SCHED
373 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
374 p->rt.parent = task_group(p)->rt_se[cpu];
380 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
381 static inline struct task_group *task_group(struct task_struct *p)
386 #endif /* CONFIG_GROUP_SCHED */
388 /* CFS-related fields in a runqueue */
390 struct load_weight load;
391 unsigned long nr_running;
396 struct rb_root tasks_timeline;
397 struct rb_node *rb_leftmost;
399 struct list_head tasks;
400 struct list_head *balance_iterator;
403 * 'curr' points to currently running entity on this cfs_rq.
404 * It is set to NULL otherwise (i.e when none are currently running).
406 struct sched_entity *curr, *next, *last;
408 unsigned int nr_spread_over;
410 #ifdef CONFIG_FAIR_GROUP_SCHED
411 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
414 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
415 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
416 * (like users, containers etc.)
418 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
419 * list is used during load balance.
421 struct list_head leaf_cfs_rq_list;
422 struct task_group *tg; /* group that "owns" this runqueue */
426 * the part of load.weight contributed by tasks
428 unsigned long task_weight;
431 * h_load = weight * f(tg)
433 * Where f(tg) is the recursive weight fraction assigned to
436 unsigned long h_load;
439 * this cpu's part of tg->shares
441 unsigned long shares;
444 * load.weight at the time we set shares
446 unsigned long rq_weight;
451 /* Real-Time classes' related field in a runqueue: */
453 struct rt_prio_array active;
454 unsigned long rt_nr_running;
455 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
456 int highest_prio; /* highest queued rt task prio */
459 unsigned long rt_nr_migratory;
465 /* Nests inside the rq lock: */
466 spinlock_t rt_runtime_lock;
468 #ifdef CONFIG_RT_GROUP_SCHED
469 unsigned long rt_nr_boosted;
472 struct list_head leaf_rt_rq_list;
473 struct task_group *tg;
474 struct sched_rt_entity *rt_se;
481 * We add the notion of a root-domain which will be used to define per-domain
482 * variables. Each exclusive cpuset essentially defines an island domain by
483 * fully partitioning the member cpus from any other cpuset. Whenever a new
484 * exclusive cpuset is created, we also create and attach a new root-domain
491 cpumask_var_t online;
494 * The "RT overload" flag: it gets set if a CPU has more than
495 * one runnable RT task.
497 cpumask_var_t rto_mask;
500 struct cpupri cpupri;
505 * By default the system creates a single root-domain with all cpus as
506 * members (mimicking the global state we have today).
508 static struct root_domain def_root_domain;
513 * This is the main, per-CPU runqueue data structure.
515 * Locking rule: those places that want to lock multiple runqueues
516 * (such as the load balancing or the thread migration code), lock
517 * acquire operations must be ordered by ascending &runqueue.
524 * nr_running and cpu_load should be in the same cacheline because
525 * remote CPUs use both these fields when doing load calculation.
527 unsigned long nr_running;
528 #define CPU_LOAD_IDX_MAX 5
529 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
530 unsigned char idle_at_tick;
532 unsigned long last_tick_seen;
533 unsigned char in_nohz_recently;
535 /* capture load from *all* tasks on this cpu: */
536 struct load_weight load;
537 unsigned long nr_load_updates;
543 #ifdef CONFIG_FAIR_GROUP_SCHED
544 /* list of leaf cfs_rq on this cpu: */
545 struct list_head leaf_cfs_rq_list;
547 #ifdef CONFIG_RT_GROUP_SCHED
548 struct list_head leaf_rt_rq_list;
552 * This is part of a global counter where only the total sum
553 * over all CPUs matters. A task can increase this counter on
554 * one CPU and if it got migrated afterwards it may decrease
555 * it on another CPU. Always updated under the runqueue lock:
557 unsigned long nr_uninterruptible;
559 struct task_struct *curr, *idle;
560 unsigned long next_balance;
561 struct mm_struct *prev_mm;
568 struct root_domain *rd;
569 struct sched_domain *sd;
571 /* For active balancing */
574 /* cpu of this runqueue: */
578 unsigned long avg_load_per_task;
580 struct task_struct *migration_thread;
581 struct list_head migration_queue;
584 #ifdef CONFIG_SCHED_HRTICK
586 int hrtick_csd_pending;
587 struct call_single_data hrtick_csd;
589 struct hrtimer hrtick_timer;
592 #ifdef CONFIG_SCHEDSTATS
594 struct sched_info rq_sched_info;
596 /* sys_sched_yield() stats */
597 unsigned int yld_exp_empty;
598 unsigned int yld_act_empty;
599 unsigned int yld_both_empty;
600 unsigned int yld_count;
602 /* schedule() stats */
603 unsigned int sched_switch;
604 unsigned int sched_count;
605 unsigned int sched_goidle;
607 /* try_to_wake_up() stats */
608 unsigned int ttwu_count;
609 unsigned int ttwu_local;
612 unsigned int bkl_count;
616 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
618 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
620 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
623 static inline int cpu_of(struct rq *rq)
633 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
634 * See detach_destroy_domains: synchronize_sched for details.
636 * The domain tree of any CPU may only be accessed from within
637 * preempt-disabled sections.
639 #define for_each_domain(cpu, __sd) \
640 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
642 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
643 #define this_rq() (&__get_cpu_var(runqueues))
644 #define task_rq(p) cpu_rq(task_cpu(p))
645 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
647 static inline void update_rq_clock(struct rq *rq)
649 rq->clock = sched_clock_cpu(cpu_of(rq));
653 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
655 #ifdef CONFIG_SCHED_DEBUG
656 # define const_debug __read_mostly
658 # define const_debug static const
664 * Returns true if the current cpu runqueue is locked.
665 * This interface allows printk to be called with the runqueue lock
666 * held and know whether or not it is OK to wake up the klogd.
668 int runqueue_is_locked(void)
671 struct rq *rq = cpu_rq(cpu);
674 ret = spin_is_locked(&rq->lock);
680 * Debugging: various feature bits
683 #define SCHED_FEAT(name, enabled) \
684 __SCHED_FEAT_##name ,
687 #include "sched_features.h"
692 #define SCHED_FEAT(name, enabled) \
693 (1UL << __SCHED_FEAT_##name) * enabled |
695 const_debug unsigned int sysctl_sched_features =
696 #include "sched_features.h"
701 #ifdef CONFIG_SCHED_DEBUG
702 #define SCHED_FEAT(name, enabled) \
705 static __read_mostly char *sched_feat_names[] = {
706 #include "sched_features.h"
712 static int sched_feat_show(struct seq_file *m, void *v)
716 for (i = 0; sched_feat_names[i]; i++) {
717 if (!(sysctl_sched_features & (1UL << i)))
719 seq_printf(m, "%s ", sched_feat_names[i]);
727 sched_feat_write(struct file *filp, const char __user *ubuf,
728 size_t cnt, loff_t *ppos)
738 if (copy_from_user(&buf, ubuf, cnt))
743 if (strncmp(buf, "NO_", 3) == 0) {
748 for (i = 0; sched_feat_names[i]; i++) {
749 int len = strlen(sched_feat_names[i]);
751 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
753 sysctl_sched_features &= ~(1UL << i);
755 sysctl_sched_features |= (1UL << i);
760 if (!sched_feat_names[i])
768 static int sched_feat_open(struct inode *inode, struct file *filp)
770 return single_open(filp, sched_feat_show, NULL);
773 static struct file_operations sched_feat_fops = {
774 .open = sched_feat_open,
775 .write = sched_feat_write,
778 .release = single_release,
781 static __init int sched_init_debug(void)
783 debugfs_create_file("sched_features", 0644, NULL, NULL,
788 late_initcall(sched_init_debug);
792 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
795 * Number of tasks to iterate in a single balance run.
796 * Limited because this is done with IRQs disabled.
798 const_debug unsigned int sysctl_sched_nr_migrate = 32;
801 * ratelimit for updating the group shares.
804 unsigned int sysctl_sched_shares_ratelimit = 250000;
807 * Inject some fuzzyness into changing the per-cpu group shares
808 * this avoids remote rq-locks at the expense of fairness.
811 unsigned int sysctl_sched_shares_thresh = 4;
814 * period over which we measure -rt task cpu usage in us.
817 unsigned int sysctl_sched_rt_period = 1000000;
819 static __read_mostly int scheduler_running;
822 * part of the period that we allow rt tasks to run in us.
825 int sysctl_sched_rt_runtime = 950000;
827 static inline u64 global_rt_period(void)
829 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
832 static inline u64 global_rt_runtime(void)
834 if (sysctl_sched_rt_runtime < 0)
837 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
840 #ifndef prepare_arch_switch
841 # define prepare_arch_switch(next) do { } while (0)
843 #ifndef finish_arch_switch
844 # define finish_arch_switch(prev) do { } while (0)
847 static inline int task_current(struct rq *rq, struct task_struct *p)
849 return rq->curr == p;
852 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
853 static inline int task_running(struct rq *rq, struct task_struct *p)
855 return task_current(rq, p);
858 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
862 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
864 #ifdef CONFIG_DEBUG_SPINLOCK
865 /* this is a valid case when another task releases the spinlock */
866 rq->lock.owner = current;
869 * If we are tracking spinlock dependencies then we have to
870 * fix up the runqueue lock - which gets 'carried over' from
873 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
875 spin_unlock_irq(&rq->lock);
878 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
879 static inline int task_running(struct rq *rq, struct task_struct *p)
884 return task_current(rq, p);
888 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
892 * We can optimise this out completely for !SMP, because the
893 * SMP rebalancing from interrupt is the only thing that cares
898 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
899 spin_unlock_irq(&rq->lock);
901 spin_unlock(&rq->lock);
905 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
909 * After ->oncpu is cleared, the task can be moved to a different CPU.
910 * We must ensure this doesn't happen until the switch is completely
916 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
920 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
923 * __task_rq_lock - lock the runqueue a given task resides on.
924 * Must be called interrupts disabled.
926 static inline struct rq *__task_rq_lock(struct task_struct *p)
930 struct rq *rq = task_rq(p);
931 spin_lock(&rq->lock);
932 if (likely(rq == task_rq(p)))
934 spin_unlock(&rq->lock);
939 * task_rq_lock - lock the runqueue a given task resides on and disable
940 * interrupts. Note the ordering: we can safely lookup the task_rq without
941 * explicitly disabling preemption.
943 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
949 local_irq_save(*flags);
951 spin_lock(&rq->lock);
952 if (likely(rq == task_rq(p)))
954 spin_unlock_irqrestore(&rq->lock, *flags);
958 void task_rq_unlock_wait(struct task_struct *p)
960 struct rq *rq = task_rq(p);
962 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
963 spin_unlock_wait(&rq->lock);
966 static void __task_rq_unlock(struct rq *rq)
969 spin_unlock(&rq->lock);
972 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
975 spin_unlock_irqrestore(&rq->lock, *flags);
979 * this_rq_lock - lock this runqueue and disable interrupts.
981 static struct rq *this_rq_lock(void)
988 spin_lock(&rq->lock);
993 #ifdef CONFIG_SCHED_HRTICK
995 * Use HR-timers to deliver accurate preemption points.
997 * Its all a bit involved since we cannot program an hrt while holding the
998 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1001 * When we get rescheduled we reprogram the hrtick_timer outside of the
1007 * - enabled by features
1008 * - hrtimer is actually high res
1010 static inline int hrtick_enabled(struct rq *rq)
1012 if (!sched_feat(HRTICK))
1014 if (!cpu_active(cpu_of(rq)))
1016 return hrtimer_is_hres_active(&rq->hrtick_timer);
1019 static void hrtick_clear(struct rq *rq)
1021 if (hrtimer_active(&rq->hrtick_timer))
1022 hrtimer_cancel(&rq->hrtick_timer);
1026 * High-resolution timer tick.
1027 * Runs from hardirq context with interrupts disabled.
1029 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1031 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1033 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1035 spin_lock(&rq->lock);
1036 update_rq_clock(rq);
1037 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1038 spin_unlock(&rq->lock);
1040 return HRTIMER_NORESTART;
1045 * called from hardirq (IPI) context
1047 static void __hrtick_start(void *arg)
1049 struct rq *rq = arg;
1051 spin_lock(&rq->lock);
1052 hrtimer_restart(&rq->hrtick_timer);
1053 rq->hrtick_csd_pending = 0;
1054 spin_unlock(&rq->lock);
1058 * Called to set the hrtick timer state.
1060 * called with rq->lock held and irqs disabled
1062 static void hrtick_start(struct rq *rq, u64 delay)
1064 struct hrtimer *timer = &rq->hrtick_timer;
1065 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1067 hrtimer_set_expires(timer, time);
1069 if (rq == this_rq()) {
1070 hrtimer_restart(timer);
1071 } else if (!rq->hrtick_csd_pending) {
1072 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd);
1073 rq->hrtick_csd_pending = 1;
1078 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1080 int cpu = (int)(long)hcpu;
1083 case CPU_UP_CANCELED:
1084 case CPU_UP_CANCELED_FROZEN:
1085 case CPU_DOWN_PREPARE:
1086 case CPU_DOWN_PREPARE_FROZEN:
1088 case CPU_DEAD_FROZEN:
1089 hrtick_clear(cpu_rq(cpu));
1096 static __init void init_hrtick(void)
1098 hotcpu_notifier(hotplug_hrtick, 0);
1102 * Called to set the hrtick timer state.
1104 * called with rq->lock held and irqs disabled
1106 static void hrtick_start(struct rq *rq, u64 delay)
1108 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay), HRTIMER_MODE_REL);
1111 static inline void init_hrtick(void)
1114 #endif /* CONFIG_SMP */
1116 static void init_rq_hrtick(struct rq *rq)
1119 rq->hrtick_csd_pending = 0;
1121 rq->hrtick_csd.flags = 0;
1122 rq->hrtick_csd.func = __hrtick_start;
1123 rq->hrtick_csd.info = rq;
1126 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1127 rq->hrtick_timer.function = hrtick;
1128 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_PERCPU;
1130 #else /* CONFIG_SCHED_HRTICK */
1131 static inline void hrtick_clear(struct rq *rq)
1135 static inline void init_rq_hrtick(struct rq *rq)
1139 static inline void init_hrtick(void)
1142 #endif /* CONFIG_SCHED_HRTICK */
1145 * resched_task - mark a task 'to be rescheduled now'.
1147 * On UP this means the setting of the need_resched flag, on SMP it
1148 * might also involve a cross-CPU call to trigger the scheduler on
1153 #ifndef tsk_is_polling
1154 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1157 static void resched_task(struct task_struct *p)
1161 assert_spin_locked(&task_rq(p)->lock);
1163 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1166 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1169 if (cpu == smp_processor_id())
1172 /* NEED_RESCHED must be visible before we test polling */
1174 if (!tsk_is_polling(p))
1175 smp_send_reschedule(cpu);
1178 static void resched_cpu(int cpu)
1180 struct rq *rq = cpu_rq(cpu);
1181 unsigned long flags;
1183 if (!spin_trylock_irqsave(&rq->lock, flags))
1185 resched_task(cpu_curr(cpu));
1186 spin_unlock_irqrestore(&rq->lock, flags);
1191 * When add_timer_on() enqueues a timer into the timer wheel of an
1192 * idle CPU then this timer might expire before the next timer event
1193 * which is scheduled to wake up that CPU. In case of a completely
1194 * idle system the next event might even be infinite time into the
1195 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1196 * leaves the inner idle loop so the newly added timer is taken into
1197 * account when the CPU goes back to idle and evaluates the timer
1198 * wheel for the next timer event.
1200 void wake_up_idle_cpu(int cpu)
1202 struct rq *rq = cpu_rq(cpu);
1204 if (cpu == smp_processor_id())
1208 * This is safe, as this function is called with the timer
1209 * wheel base lock of (cpu) held. When the CPU is on the way
1210 * to idle and has not yet set rq->curr to idle then it will
1211 * be serialized on the timer wheel base lock and take the new
1212 * timer into account automatically.
1214 if (rq->curr != rq->idle)
1218 * We can set TIF_RESCHED on the idle task of the other CPU
1219 * lockless. The worst case is that the other CPU runs the
1220 * idle task through an additional NOOP schedule()
1222 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1224 /* NEED_RESCHED must be visible before we test polling */
1226 if (!tsk_is_polling(rq->idle))
1227 smp_send_reschedule(cpu);
1229 #endif /* CONFIG_NO_HZ */
1231 #else /* !CONFIG_SMP */
1232 static void resched_task(struct task_struct *p)
1234 assert_spin_locked(&task_rq(p)->lock);
1235 set_tsk_need_resched(p);
1237 #endif /* CONFIG_SMP */
1239 #if BITS_PER_LONG == 32
1240 # define WMULT_CONST (~0UL)
1242 # define WMULT_CONST (1UL << 32)
1245 #define WMULT_SHIFT 32
1248 * Shift right and round:
1250 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1253 * delta *= weight / lw
1255 static unsigned long
1256 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1257 struct load_weight *lw)
1261 if (!lw->inv_weight) {
1262 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1265 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1269 tmp = (u64)delta_exec * weight;
1271 * Check whether we'd overflow the 64-bit multiplication:
1273 if (unlikely(tmp > WMULT_CONST))
1274 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1277 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1279 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1282 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1288 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1295 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1296 * of tasks with abnormal "nice" values across CPUs the contribution that
1297 * each task makes to its run queue's load is weighted according to its
1298 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1299 * scaled version of the new time slice allocation that they receive on time
1303 #define WEIGHT_IDLEPRIO 2
1304 #define WMULT_IDLEPRIO (1 << 31)
1307 * Nice levels are multiplicative, with a gentle 10% change for every
1308 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1309 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1310 * that remained on nice 0.
1312 * The "10% effect" is relative and cumulative: from _any_ nice level,
1313 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1314 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1315 * If a task goes up by ~10% and another task goes down by ~10% then
1316 * the relative distance between them is ~25%.)
1318 static const int prio_to_weight[40] = {
1319 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1320 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1321 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1322 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1323 /* 0 */ 1024, 820, 655, 526, 423,
1324 /* 5 */ 335, 272, 215, 172, 137,
1325 /* 10 */ 110, 87, 70, 56, 45,
1326 /* 15 */ 36, 29, 23, 18, 15,
1330 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1332 * In cases where the weight does not change often, we can use the
1333 * precalculated inverse to speed up arithmetics by turning divisions
1334 * into multiplications:
1336 static const u32 prio_to_wmult[40] = {
1337 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1338 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1339 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1340 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1341 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1342 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1343 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1344 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1347 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1350 * runqueue iterator, to support SMP load-balancing between different
1351 * scheduling classes, without having to expose their internal data
1352 * structures to the load-balancing proper:
1354 struct rq_iterator {
1356 struct task_struct *(*start)(void *);
1357 struct task_struct *(*next)(void *);
1361 static unsigned long
1362 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1363 unsigned long max_load_move, struct sched_domain *sd,
1364 enum cpu_idle_type idle, int *all_pinned,
1365 int *this_best_prio, struct rq_iterator *iterator);
1368 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1369 struct sched_domain *sd, enum cpu_idle_type idle,
1370 struct rq_iterator *iterator);
1373 #ifdef CONFIG_CGROUP_CPUACCT
1374 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1376 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1379 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1381 update_load_add(&rq->load, load);
1384 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1386 update_load_sub(&rq->load, load);
1389 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1390 typedef int (*tg_visitor)(struct task_group *, void *);
1393 * Iterate the full tree, calling @down when first entering a node and @up when
1394 * leaving it for the final time.
1396 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1398 struct task_group *parent, *child;
1402 parent = &root_task_group;
1404 ret = (*down)(parent, data);
1407 list_for_each_entry_rcu(child, &parent->children, siblings) {
1414 ret = (*up)(parent, data);
1419 parent = parent->parent;
1428 static int tg_nop(struct task_group *tg, void *data)
1435 static unsigned long source_load(int cpu, int type);
1436 static unsigned long target_load(int cpu, int type);
1437 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1439 static unsigned long cpu_avg_load_per_task(int cpu)
1441 struct rq *rq = cpu_rq(cpu);
1444 rq->avg_load_per_task = rq->load.weight / rq->nr_running;
1446 rq->avg_load_per_task = 0;
1448 return rq->avg_load_per_task;
1451 #ifdef CONFIG_FAIR_GROUP_SCHED
1453 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1456 * Calculate and set the cpu's group shares.
1459 update_group_shares_cpu(struct task_group *tg, int cpu,
1460 unsigned long sd_shares, unsigned long sd_rq_weight)
1462 unsigned long shares;
1463 unsigned long rq_weight;
1468 rq_weight = tg->cfs_rq[cpu]->rq_weight;
1471 * \Sum shares * rq_weight
1472 * shares = -----------------------
1476 shares = (sd_shares * rq_weight) / sd_rq_weight;
1477 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1479 if (abs(shares - tg->se[cpu]->load.weight) >
1480 sysctl_sched_shares_thresh) {
1481 struct rq *rq = cpu_rq(cpu);
1482 unsigned long flags;
1484 spin_lock_irqsave(&rq->lock, flags);
1485 tg->cfs_rq[cpu]->shares = shares;
1487 __set_se_shares(tg->se[cpu], shares);
1488 spin_unlock_irqrestore(&rq->lock, flags);
1493 * Re-compute the task group their per cpu shares over the given domain.
1494 * This needs to be done in a bottom-up fashion because the rq weight of a
1495 * parent group depends on the shares of its child groups.
1497 static int tg_shares_up(struct task_group *tg, void *data)
1499 unsigned long weight, rq_weight = 0;
1500 unsigned long shares = 0;
1501 struct sched_domain *sd = data;
1504 for_each_cpu(i, sched_domain_span(sd)) {
1506 * If there are currently no tasks on the cpu pretend there
1507 * is one of average load so that when a new task gets to
1508 * run here it will not get delayed by group starvation.
1510 weight = tg->cfs_rq[i]->load.weight;
1512 weight = NICE_0_LOAD;
1514 tg->cfs_rq[i]->rq_weight = weight;
1515 rq_weight += weight;
1516 shares += tg->cfs_rq[i]->shares;
1519 if ((!shares && rq_weight) || shares > tg->shares)
1520 shares = tg->shares;
1522 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1523 shares = tg->shares;
1525 for_each_cpu(i, sched_domain_span(sd))
1526 update_group_shares_cpu(tg, i, shares, rq_weight);
1532 * Compute the cpu's hierarchical load factor for each task group.
1533 * This needs to be done in a top-down fashion because the load of a child
1534 * group is a fraction of its parents load.
1536 static int tg_load_down(struct task_group *tg, void *data)
1539 long cpu = (long)data;
1542 load = cpu_rq(cpu)->load.weight;
1544 load = tg->parent->cfs_rq[cpu]->h_load;
1545 load *= tg->cfs_rq[cpu]->shares;
1546 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1549 tg->cfs_rq[cpu]->h_load = load;
1554 static void update_shares(struct sched_domain *sd)
1556 u64 now = cpu_clock(raw_smp_processor_id());
1557 s64 elapsed = now - sd->last_update;
1559 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1560 sd->last_update = now;
1561 walk_tg_tree(tg_nop, tg_shares_up, sd);
1565 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1567 spin_unlock(&rq->lock);
1569 spin_lock(&rq->lock);
1572 static void update_h_load(long cpu)
1574 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1579 static inline void update_shares(struct sched_domain *sd)
1583 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1591 #ifdef CONFIG_FAIR_GROUP_SCHED
1592 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1595 cfs_rq->shares = shares;
1600 #include "sched_stats.h"
1601 #include "sched_idletask.c"
1602 #include "sched_fair.c"
1603 #include "sched_rt.c"
1604 #ifdef CONFIG_SCHED_DEBUG
1605 # include "sched_debug.c"
1608 #define sched_class_highest (&rt_sched_class)
1609 #define for_each_class(class) \
1610 for (class = sched_class_highest; class; class = class->next)
1612 static void inc_nr_running(struct rq *rq)
1617 static void dec_nr_running(struct rq *rq)
1622 static void set_load_weight(struct task_struct *p)
1624 if (task_has_rt_policy(p)) {
1625 p->se.load.weight = prio_to_weight[0] * 2;
1626 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1631 * SCHED_IDLE tasks get minimal weight:
1633 if (p->policy == SCHED_IDLE) {
1634 p->se.load.weight = WEIGHT_IDLEPRIO;
1635 p->se.load.inv_weight = WMULT_IDLEPRIO;
1639 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1640 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1643 static void update_avg(u64 *avg, u64 sample)
1645 s64 diff = sample - *avg;
1649 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1651 sched_info_queued(p);
1652 p->sched_class->enqueue_task(rq, p, wakeup);
1656 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1658 if (sleep && p->se.last_wakeup) {
1659 update_avg(&p->se.avg_overlap,
1660 p->se.sum_exec_runtime - p->se.last_wakeup);
1661 p->se.last_wakeup = 0;
1664 sched_info_dequeued(p);
1665 p->sched_class->dequeue_task(rq, p, sleep);
1670 * __normal_prio - return the priority that is based on the static prio
1672 static inline int __normal_prio(struct task_struct *p)
1674 return p->static_prio;
1678 * Calculate the expected normal priority: i.e. priority
1679 * without taking RT-inheritance into account. Might be
1680 * boosted by interactivity modifiers. Changes upon fork,
1681 * setprio syscalls, and whenever the interactivity
1682 * estimator recalculates.
1684 static inline int normal_prio(struct task_struct *p)
1688 if (task_has_rt_policy(p))
1689 prio = MAX_RT_PRIO-1 - p->rt_priority;
1691 prio = __normal_prio(p);
1696 * Calculate the current priority, i.e. the priority
1697 * taken into account by the scheduler. This value might
1698 * be boosted by RT tasks, or might be boosted by
1699 * interactivity modifiers. Will be RT if the task got
1700 * RT-boosted. If not then it returns p->normal_prio.
1702 static int effective_prio(struct task_struct *p)
1704 p->normal_prio = normal_prio(p);
1706 * If we are RT tasks or we were boosted to RT priority,
1707 * keep the priority unchanged. Otherwise, update priority
1708 * to the normal priority:
1710 if (!rt_prio(p->prio))
1711 return p->normal_prio;
1716 * activate_task - move a task to the runqueue.
1718 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1720 if (task_contributes_to_load(p))
1721 rq->nr_uninterruptible--;
1723 enqueue_task(rq, p, wakeup);
1728 * deactivate_task - remove a task from the runqueue.
1730 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1732 if (task_contributes_to_load(p))
1733 rq->nr_uninterruptible++;
1735 dequeue_task(rq, p, sleep);
1740 * task_curr - is this task currently executing on a CPU?
1741 * @p: the task in question.
1743 inline int task_curr(const struct task_struct *p)
1745 return cpu_curr(task_cpu(p)) == p;
1748 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1750 set_task_rq(p, cpu);
1753 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1754 * successfuly executed on another CPU. We must ensure that updates of
1755 * per-task data have been completed by this moment.
1758 task_thread_info(p)->cpu = cpu;
1762 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1763 const struct sched_class *prev_class,
1764 int oldprio, int running)
1766 if (prev_class != p->sched_class) {
1767 if (prev_class->switched_from)
1768 prev_class->switched_from(rq, p, running);
1769 p->sched_class->switched_to(rq, p, running);
1771 p->sched_class->prio_changed(rq, p, oldprio, running);
1776 /* Used instead of source_load when we know the type == 0 */
1777 static unsigned long weighted_cpuload(const int cpu)
1779 return cpu_rq(cpu)->load.weight;
1783 * Is this task likely cache-hot:
1786 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1791 * Buddy candidates are cache hot:
1793 if (sched_feat(CACHE_HOT_BUDDY) &&
1794 (&p->se == cfs_rq_of(&p->se)->next ||
1795 &p->se == cfs_rq_of(&p->se)->last))
1798 if (p->sched_class != &fair_sched_class)
1801 if (sysctl_sched_migration_cost == -1)
1803 if (sysctl_sched_migration_cost == 0)
1806 delta = now - p->se.exec_start;
1808 return delta < (s64)sysctl_sched_migration_cost;
1812 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1814 int old_cpu = task_cpu(p);
1815 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1816 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1817 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1820 clock_offset = old_rq->clock - new_rq->clock;
1822 #ifdef CONFIG_SCHEDSTATS
1823 if (p->se.wait_start)
1824 p->se.wait_start -= clock_offset;
1825 if (p->se.sleep_start)
1826 p->se.sleep_start -= clock_offset;
1827 if (p->se.block_start)
1828 p->se.block_start -= clock_offset;
1829 if (old_cpu != new_cpu) {
1830 schedstat_inc(p, se.nr_migrations);
1831 if (task_hot(p, old_rq->clock, NULL))
1832 schedstat_inc(p, se.nr_forced2_migrations);
1835 p->se.vruntime -= old_cfsrq->min_vruntime -
1836 new_cfsrq->min_vruntime;
1838 __set_task_cpu(p, new_cpu);
1841 struct migration_req {
1842 struct list_head list;
1844 struct task_struct *task;
1847 struct completion done;
1851 * The task's runqueue lock must be held.
1852 * Returns true if you have to wait for migration thread.
1855 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1857 struct rq *rq = task_rq(p);
1860 * If the task is not on a runqueue (and not running), then
1861 * it is sufficient to simply update the task's cpu field.
1863 if (!p->se.on_rq && !task_running(rq, p)) {
1864 set_task_cpu(p, dest_cpu);
1868 init_completion(&req->done);
1870 req->dest_cpu = dest_cpu;
1871 list_add(&req->list, &rq->migration_queue);
1877 * wait_task_inactive - wait for a thread to unschedule.
1879 * If @match_state is nonzero, it's the @p->state value just checked and
1880 * not expected to change. If it changes, i.e. @p might have woken up,
1881 * then return zero. When we succeed in waiting for @p to be off its CPU,
1882 * we return a positive number (its total switch count). If a second call
1883 * a short while later returns the same number, the caller can be sure that
1884 * @p has remained unscheduled the whole time.
1886 * The caller must ensure that the task *will* unschedule sometime soon,
1887 * else this function might spin for a *long* time. This function can't
1888 * be called with interrupts off, or it may introduce deadlock with
1889 * smp_call_function() if an IPI is sent by the same process we are
1890 * waiting to become inactive.
1892 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1894 unsigned long flags;
1901 * We do the initial early heuristics without holding
1902 * any task-queue locks at all. We'll only try to get
1903 * the runqueue lock when things look like they will
1909 * If the task is actively running on another CPU
1910 * still, just relax and busy-wait without holding
1913 * NOTE! Since we don't hold any locks, it's not
1914 * even sure that "rq" stays as the right runqueue!
1915 * But we don't care, since "task_running()" will
1916 * return false if the runqueue has changed and p
1917 * is actually now running somewhere else!
1919 while (task_running(rq, p)) {
1920 if (match_state && unlikely(p->state != match_state))
1926 * Ok, time to look more closely! We need the rq
1927 * lock now, to be *sure*. If we're wrong, we'll
1928 * just go back and repeat.
1930 rq = task_rq_lock(p, &flags);
1931 trace_sched_wait_task(rq, p);
1932 running = task_running(rq, p);
1933 on_rq = p->se.on_rq;
1935 if (!match_state || p->state == match_state)
1936 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1937 task_rq_unlock(rq, &flags);
1940 * If it changed from the expected state, bail out now.
1942 if (unlikely(!ncsw))
1946 * Was it really running after all now that we
1947 * checked with the proper locks actually held?
1949 * Oops. Go back and try again..
1951 if (unlikely(running)) {
1957 * It's not enough that it's not actively running,
1958 * it must be off the runqueue _entirely_, and not
1961 * So if it wa still runnable (but just not actively
1962 * running right now), it's preempted, and we should
1963 * yield - it could be a while.
1965 if (unlikely(on_rq)) {
1966 schedule_timeout_uninterruptible(1);
1971 * Ahh, all good. It wasn't running, and it wasn't
1972 * runnable, which means that it will never become
1973 * running in the future either. We're all done!
1982 * kick_process - kick a running thread to enter/exit the kernel
1983 * @p: the to-be-kicked thread
1985 * Cause a process which is running on another CPU to enter
1986 * kernel-mode, without any delay. (to get signals handled.)
1988 * NOTE: this function doesnt have to take the runqueue lock,
1989 * because all it wants to ensure is that the remote task enters
1990 * the kernel. If the IPI races and the task has been migrated
1991 * to another CPU then no harm is done and the purpose has been
1994 void kick_process(struct task_struct *p)
2000 if ((cpu != smp_processor_id()) && task_curr(p))
2001 smp_send_reschedule(cpu);
2006 * Return a low guess at the load of a migration-source cpu weighted
2007 * according to the scheduling class and "nice" value.
2009 * We want to under-estimate the load of migration sources, to
2010 * balance conservatively.
2012 static unsigned long source_load(int cpu, int type)
2014 struct rq *rq = cpu_rq(cpu);
2015 unsigned long total = weighted_cpuload(cpu);
2017 if (type == 0 || !sched_feat(LB_BIAS))
2020 return min(rq->cpu_load[type-1], total);
2024 * Return a high guess at the load of a migration-target cpu weighted
2025 * according to the scheduling class and "nice" value.
2027 static unsigned long target_load(int cpu, int type)
2029 struct rq *rq = cpu_rq(cpu);
2030 unsigned long total = weighted_cpuload(cpu);
2032 if (type == 0 || !sched_feat(LB_BIAS))
2035 return max(rq->cpu_load[type-1], total);
2039 * find_idlest_group finds and returns the least busy CPU group within the
2042 static struct sched_group *
2043 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2045 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2046 unsigned long min_load = ULONG_MAX, this_load = 0;
2047 int load_idx = sd->forkexec_idx;
2048 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2051 unsigned long load, avg_load;
2055 /* Skip over this group if it has no CPUs allowed */
2056 if (!cpumask_intersects(sched_group_cpus(group),
2060 local_group = cpumask_test_cpu(this_cpu,
2061 sched_group_cpus(group));
2063 /* Tally up the load of all CPUs in the group */
2066 for_each_cpu(i, sched_group_cpus(group)) {
2067 /* Bias balancing toward cpus of our domain */
2069 load = source_load(i, load_idx);
2071 load = target_load(i, load_idx);
2076 /* Adjust by relative CPU power of the group */
2077 avg_load = sg_div_cpu_power(group,
2078 avg_load * SCHED_LOAD_SCALE);
2081 this_load = avg_load;
2083 } else if (avg_load < min_load) {
2084 min_load = avg_load;
2087 } while (group = group->next, group != sd->groups);
2089 if (!idlest || 100*this_load < imbalance*min_load)
2095 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2098 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2100 unsigned long load, min_load = ULONG_MAX;
2104 /* Traverse only the allowed CPUs */
2105 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
2106 load = weighted_cpuload(i);
2108 if (load < min_load || (load == min_load && i == this_cpu)) {
2118 * sched_balance_self: balance the current task (running on cpu) in domains
2119 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2122 * Balance, ie. select the least loaded group.
2124 * Returns the target CPU number, or the same CPU if no balancing is needed.
2126 * preempt must be disabled.
2128 static int sched_balance_self(int cpu, int flag)
2130 struct task_struct *t = current;
2131 struct sched_domain *tmp, *sd = NULL;
2133 for_each_domain(cpu, tmp) {
2135 * If power savings logic is enabled for a domain, stop there.
2137 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2139 if (tmp->flags & flag)
2147 struct sched_group *group;
2148 int new_cpu, weight;
2150 if (!(sd->flags & flag)) {
2155 group = find_idlest_group(sd, t, cpu);
2161 new_cpu = find_idlest_cpu(group, t, cpu);
2162 if (new_cpu == -1 || new_cpu == cpu) {
2163 /* Now try balancing at a lower domain level of cpu */
2168 /* Now try balancing at a lower domain level of new_cpu */
2170 weight = cpumask_weight(sched_domain_span(sd));
2172 for_each_domain(cpu, tmp) {
2173 if (weight <= cpumask_weight(sched_domain_span(tmp)))
2175 if (tmp->flags & flag)
2178 /* while loop will break here if sd == NULL */
2184 #endif /* CONFIG_SMP */
2187 * try_to_wake_up - wake up a thread
2188 * @p: the to-be-woken-up thread
2189 * @state: the mask of task states that can be woken
2190 * @sync: do a synchronous wakeup?
2192 * Put it on the run-queue if it's not already there. The "current"
2193 * thread is always on the run-queue (except when the actual
2194 * re-schedule is in progress), and as such you're allowed to do
2195 * the simpler "current->state = TASK_RUNNING" to mark yourself
2196 * runnable without the overhead of this.
2198 * returns failure only if the task is already active.
2200 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2202 int cpu, orig_cpu, this_cpu, success = 0;
2203 unsigned long flags;
2207 if (!sched_feat(SYNC_WAKEUPS))
2211 if (sched_feat(LB_WAKEUP_UPDATE)) {
2212 struct sched_domain *sd;
2214 this_cpu = raw_smp_processor_id();
2217 for_each_domain(this_cpu, sd) {
2218 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2227 rq = task_rq_lock(p, &flags);
2228 old_state = p->state;
2229 if (!(old_state & state))
2237 this_cpu = smp_processor_id();
2240 if (unlikely(task_running(rq, p)))
2243 cpu = p->sched_class->select_task_rq(p, sync);
2244 if (cpu != orig_cpu) {
2245 set_task_cpu(p, cpu);
2246 task_rq_unlock(rq, &flags);
2247 /* might preempt at this point */
2248 rq = task_rq_lock(p, &flags);
2249 old_state = p->state;
2250 if (!(old_state & state))
2255 this_cpu = smp_processor_id();
2259 #ifdef CONFIG_SCHEDSTATS
2260 schedstat_inc(rq, ttwu_count);
2261 if (cpu == this_cpu)
2262 schedstat_inc(rq, ttwu_local);
2264 struct sched_domain *sd;
2265 for_each_domain(this_cpu, sd) {
2266 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2267 schedstat_inc(sd, ttwu_wake_remote);
2272 #endif /* CONFIG_SCHEDSTATS */
2275 #endif /* CONFIG_SMP */
2276 schedstat_inc(p, se.nr_wakeups);
2278 schedstat_inc(p, se.nr_wakeups_sync);
2279 if (orig_cpu != cpu)
2280 schedstat_inc(p, se.nr_wakeups_migrate);
2281 if (cpu == this_cpu)
2282 schedstat_inc(p, se.nr_wakeups_local);
2284 schedstat_inc(p, se.nr_wakeups_remote);
2285 update_rq_clock(rq);
2286 activate_task(rq, p, 1);
2290 trace_sched_wakeup(rq, p);
2291 check_preempt_curr(rq, p, sync);
2293 p->state = TASK_RUNNING;
2295 if (p->sched_class->task_wake_up)
2296 p->sched_class->task_wake_up(rq, p);
2299 current->se.last_wakeup = current->se.sum_exec_runtime;
2301 task_rq_unlock(rq, &flags);
2306 int wake_up_process(struct task_struct *p)
2308 return try_to_wake_up(p, TASK_ALL, 0);
2310 EXPORT_SYMBOL(wake_up_process);
2312 int wake_up_state(struct task_struct *p, unsigned int state)
2314 return try_to_wake_up(p, state, 0);
2318 * Perform scheduler related setup for a newly forked process p.
2319 * p is forked by current.
2321 * __sched_fork() is basic setup used by init_idle() too:
2323 static void __sched_fork(struct task_struct *p)
2325 p->se.exec_start = 0;
2326 p->se.sum_exec_runtime = 0;
2327 p->se.prev_sum_exec_runtime = 0;
2328 p->se.last_wakeup = 0;
2329 p->se.avg_overlap = 0;
2331 #ifdef CONFIG_SCHEDSTATS
2332 p->se.wait_start = 0;
2333 p->se.sum_sleep_runtime = 0;
2334 p->se.sleep_start = 0;
2335 p->se.block_start = 0;
2336 p->se.sleep_max = 0;
2337 p->se.block_max = 0;
2339 p->se.slice_max = 0;
2343 INIT_LIST_HEAD(&p->rt.run_list);
2345 INIT_LIST_HEAD(&p->se.group_node);
2347 #ifdef CONFIG_PREEMPT_NOTIFIERS
2348 INIT_HLIST_HEAD(&p->preempt_notifiers);
2352 * We mark the process as running here, but have not actually
2353 * inserted it onto the runqueue yet. This guarantees that
2354 * nobody will actually run it, and a signal or other external
2355 * event cannot wake it up and insert it on the runqueue either.
2357 p->state = TASK_RUNNING;
2361 * fork()/clone()-time setup:
2363 void sched_fork(struct task_struct *p, int clone_flags)
2365 int cpu = get_cpu();
2370 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2372 set_task_cpu(p, cpu);
2375 * Make sure we do not leak PI boosting priority to the child:
2377 p->prio = current->normal_prio;
2378 if (!rt_prio(p->prio))
2379 p->sched_class = &fair_sched_class;
2381 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2382 if (likely(sched_info_on()))
2383 memset(&p->sched_info, 0, sizeof(p->sched_info));
2385 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2388 #ifdef CONFIG_PREEMPT
2389 /* Want to start with kernel preemption disabled. */
2390 task_thread_info(p)->preempt_count = 1;
2396 * wake_up_new_task - wake up a newly created task for the first time.
2398 * This function will do some initial scheduler statistics housekeeping
2399 * that must be done for every newly created context, then puts the task
2400 * on the runqueue and wakes it.
2402 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2404 unsigned long flags;
2407 rq = task_rq_lock(p, &flags);
2408 BUG_ON(p->state != TASK_RUNNING);
2409 update_rq_clock(rq);
2411 p->prio = effective_prio(p);
2413 if (!p->sched_class->task_new || !current->se.on_rq) {
2414 activate_task(rq, p, 0);
2417 * Let the scheduling class do new task startup
2418 * management (if any):
2420 p->sched_class->task_new(rq, p);
2423 trace_sched_wakeup_new(rq, p);
2424 check_preempt_curr(rq, p, 0);
2426 if (p->sched_class->task_wake_up)
2427 p->sched_class->task_wake_up(rq, p);
2429 task_rq_unlock(rq, &flags);
2432 #ifdef CONFIG_PREEMPT_NOTIFIERS
2435 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2436 * @notifier: notifier struct to register
2438 void preempt_notifier_register(struct preempt_notifier *notifier)
2440 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2442 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2445 * preempt_notifier_unregister - no longer interested in preemption notifications
2446 * @notifier: notifier struct to unregister
2448 * This is safe to call from within a preemption notifier.
2450 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2452 hlist_del(¬ifier->link);
2454 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2456 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2458 struct preempt_notifier *notifier;
2459 struct hlist_node *node;
2461 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2462 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2466 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2467 struct task_struct *next)
2469 struct preempt_notifier *notifier;
2470 struct hlist_node *node;
2472 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2473 notifier->ops->sched_out(notifier, next);
2476 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2478 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2483 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2484 struct task_struct *next)
2488 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2491 * prepare_task_switch - prepare to switch tasks
2492 * @rq: the runqueue preparing to switch
2493 * @prev: the current task that is being switched out
2494 * @next: the task we are going to switch to.
2496 * This is called with the rq lock held and interrupts off. It must
2497 * be paired with a subsequent finish_task_switch after the context
2500 * prepare_task_switch sets up locking and calls architecture specific
2504 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2505 struct task_struct *next)
2507 fire_sched_out_preempt_notifiers(prev, next);
2508 prepare_lock_switch(rq, next);
2509 prepare_arch_switch(next);
2513 * finish_task_switch - clean up after a task-switch
2514 * @rq: runqueue associated with task-switch
2515 * @prev: the thread we just switched away from.
2517 * finish_task_switch must be called after the context switch, paired
2518 * with a prepare_task_switch call before the context switch.
2519 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2520 * and do any other architecture-specific cleanup actions.
2522 * Note that we may have delayed dropping an mm in context_switch(). If
2523 * so, we finish that here outside of the runqueue lock. (Doing it
2524 * with the lock held can cause deadlocks; see schedule() for
2527 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2528 __releases(rq->lock)
2530 struct mm_struct *mm = rq->prev_mm;
2536 * A task struct has one reference for the use as "current".
2537 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2538 * schedule one last time. The schedule call will never return, and
2539 * the scheduled task must drop that reference.
2540 * The test for TASK_DEAD must occur while the runqueue locks are
2541 * still held, otherwise prev could be scheduled on another cpu, die
2542 * there before we look at prev->state, and then the reference would
2544 * Manfred Spraul <manfred@colorfullife.com>
2546 prev_state = prev->state;
2547 finish_arch_switch(prev);
2548 finish_lock_switch(rq, prev);
2550 if (current->sched_class->post_schedule)
2551 current->sched_class->post_schedule(rq);
2554 fire_sched_in_preempt_notifiers(current);
2557 if (unlikely(prev_state == TASK_DEAD)) {
2559 * Remove function-return probe instances associated with this
2560 * task and put them back on the free list.
2562 kprobe_flush_task(prev);
2563 put_task_struct(prev);
2568 * schedule_tail - first thing a freshly forked thread must call.
2569 * @prev: the thread we just switched away from.
2571 asmlinkage void schedule_tail(struct task_struct *prev)
2572 __releases(rq->lock)
2574 struct rq *rq = this_rq();
2576 finish_task_switch(rq, prev);
2577 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2578 /* In this case, finish_task_switch does not reenable preemption */
2581 if (current->set_child_tid)
2582 put_user(task_pid_vnr(current), current->set_child_tid);
2586 * context_switch - switch to the new MM and the new
2587 * thread's register state.
2590 context_switch(struct rq *rq, struct task_struct *prev,
2591 struct task_struct *next)
2593 struct mm_struct *mm, *oldmm;
2595 prepare_task_switch(rq, prev, next);
2596 trace_sched_switch(rq, prev, next);
2598 oldmm = prev->active_mm;
2600 * For paravirt, this is coupled with an exit in switch_to to
2601 * combine the page table reload and the switch backend into
2604 arch_enter_lazy_cpu_mode();
2606 if (unlikely(!mm)) {
2607 next->active_mm = oldmm;
2608 atomic_inc(&oldmm->mm_count);
2609 enter_lazy_tlb(oldmm, next);
2611 switch_mm(oldmm, mm, next);
2613 if (unlikely(!prev->mm)) {
2614 prev->active_mm = NULL;
2615 rq->prev_mm = oldmm;
2618 * Since the runqueue lock will be released by the next
2619 * task (which is an invalid locking op but in the case
2620 * of the scheduler it's an obvious special-case), so we
2621 * do an early lockdep release here:
2623 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2624 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2627 /* Here we just switch the register state and the stack. */
2628 switch_to(prev, next, prev);
2632 * this_rq must be evaluated again because prev may have moved
2633 * CPUs since it called schedule(), thus the 'rq' on its stack
2634 * frame will be invalid.
2636 finish_task_switch(this_rq(), prev);
2640 * nr_running, nr_uninterruptible and nr_context_switches:
2642 * externally visible scheduler statistics: current number of runnable
2643 * threads, current number of uninterruptible-sleeping threads, total
2644 * number of context switches performed since bootup.
2646 unsigned long nr_running(void)
2648 unsigned long i, sum = 0;
2650 for_each_online_cpu(i)
2651 sum += cpu_rq(i)->nr_running;
2656 unsigned long nr_uninterruptible(void)
2658 unsigned long i, sum = 0;
2660 for_each_possible_cpu(i)
2661 sum += cpu_rq(i)->nr_uninterruptible;
2664 * Since we read the counters lockless, it might be slightly
2665 * inaccurate. Do not allow it to go below zero though:
2667 if (unlikely((long)sum < 0))
2673 unsigned long long nr_context_switches(void)
2676 unsigned long long sum = 0;
2678 for_each_possible_cpu(i)
2679 sum += cpu_rq(i)->nr_switches;
2684 unsigned long nr_iowait(void)
2686 unsigned long i, sum = 0;
2688 for_each_possible_cpu(i)
2689 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2694 unsigned long nr_active(void)
2696 unsigned long i, running = 0, uninterruptible = 0;
2698 for_each_online_cpu(i) {
2699 running += cpu_rq(i)->nr_running;
2700 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2703 if (unlikely((long)uninterruptible < 0))
2704 uninterruptible = 0;
2706 return running + uninterruptible;
2710 * Update rq->cpu_load[] statistics. This function is usually called every
2711 * scheduler tick (TICK_NSEC).
2713 static void update_cpu_load(struct rq *this_rq)
2715 unsigned long this_load = this_rq->load.weight;
2718 this_rq->nr_load_updates++;
2720 /* Update our load: */
2721 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2722 unsigned long old_load, new_load;
2724 /* scale is effectively 1 << i now, and >> i divides by scale */
2726 old_load = this_rq->cpu_load[i];
2727 new_load = this_load;
2729 * Round up the averaging division if load is increasing. This
2730 * prevents us from getting stuck on 9 if the load is 10, for
2733 if (new_load > old_load)
2734 new_load += scale-1;
2735 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2742 * double_rq_lock - safely lock two runqueues
2744 * Note this does not disable interrupts like task_rq_lock,
2745 * you need to do so manually before calling.
2747 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2748 __acquires(rq1->lock)
2749 __acquires(rq2->lock)
2751 BUG_ON(!irqs_disabled());
2753 spin_lock(&rq1->lock);
2754 __acquire(rq2->lock); /* Fake it out ;) */
2757 spin_lock(&rq1->lock);
2758 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2760 spin_lock(&rq2->lock);
2761 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2764 update_rq_clock(rq1);
2765 update_rq_clock(rq2);
2769 * double_rq_unlock - safely unlock two runqueues
2771 * Note this does not restore interrupts like task_rq_unlock,
2772 * you need to do so manually after calling.
2774 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2775 __releases(rq1->lock)
2776 __releases(rq2->lock)
2778 spin_unlock(&rq1->lock);
2780 spin_unlock(&rq2->lock);
2782 __release(rq2->lock);
2786 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2788 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2789 __releases(this_rq->lock)
2790 __acquires(busiest->lock)
2791 __acquires(this_rq->lock)
2795 if (unlikely(!irqs_disabled())) {
2796 /* printk() doesn't work good under rq->lock */
2797 spin_unlock(&this_rq->lock);
2800 if (unlikely(!spin_trylock(&busiest->lock))) {
2801 if (busiest < this_rq) {
2802 spin_unlock(&this_rq->lock);
2803 spin_lock(&busiest->lock);
2804 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
2807 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
2812 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
2813 __releases(busiest->lock)
2815 spin_unlock(&busiest->lock);
2816 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
2820 * If dest_cpu is allowed for this process, migrate the task to it.
2821 * This is accomplished by forcing the cpu_allowed mask to only
2822 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2823 * the cpu_allowed mask is restored.
2825 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2827 struct migration_req req;
2828 unsigned long flags;
2831 rq = task_rq_lock(p, &flags);
2832 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2833 || unlikely(!cpu_active(dest_cpu)))
2836 trace_sched_migrate_task(rq, p, dest_cpu);
2837 /* force the process onto the specified CPU */
2838 if (migrate_task(p, dest_cpu, &req)) {
2839 /* Need to wait for migration thread (might exit: take ref). */
2840 struct task_struct *mt = rq->migration_thread;
2842 get_task_struct(mt);
2843 task_rq_unlock(rq, &flags);
2844 wake_up_process(mt);
2845 put_task_struct(mt);
2846 wait_for_completion(&req.done);
2851 task_rq_unlock(rq, &flags);
2855 * sched_exec - execve() is a valuable balancing opportunity, because at
2856 * this point the task has the smallest effective memory and cache footprint.
2858 void sched_exec(void)
2860 int new_cpu, this_cpu = get_cpu();
2861 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2863 if (new_cpu != this_cpu)
2864 sched_migrate_task(current, new_cpu);
2868 * pull_task - move a task from a remote runqueue to the local runqueue.
2869 * Both runqueues must be locked.
2871 static void pull_task(struct rq *src_rq, struct task_struct *p,
2872 struct rq *this_rq, int this_cpu)
2874 deactivate_task(src_rq, p, 0);
2875 set_task_cpu(p, this_cpu);
2876 activate_task(this_rq, p, 0);
2878 * Note that idle threads have a prio of MAX_PRIO, for this test
2879 * to be always true for them.
2881 check_preempt_curr(this_rq, p, 0);
2885 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2888 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2889 struct sched_domain *sd, enum cpu_idle_type idle,
2893 * We do not migrate tasks that are:
2894 * 1) running (obviously), or
2895 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2896 * 3) are cache-hot on their current CPU.
2898 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2899 schedstat_inc(p, se.nr_failed_migrations_affine);
2904 if (task_running(rq, p)) {
2905 schedstat_inc(p, se.nr_failed_migrations_running);
2910 * Aggressive migration if:
2911 * 1) task is cache cold, or
2912 * 2) too many balance attempts have failed.
2915 if (!task_hot(p, rq->clock, sd) ||
2916 sd->nr_balance_failed > sd->cache_nice_tries) {
2917 #ifdef CONFIG_SCHEDSTATS
2918 if (task_hot(p, rq->clock, sd)) {
2919 schedstat_inc(sd, lb_hot_gained[idle]);
2920 schedstat_inc(p, se.nr_forced_migrations);
2926 if (task_hot(p, rq->clock, sd)) {
2927 schedstat_inc(p, se.nr_failed_migrations_hot);
2933 static unsigned long
2934 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2935 unsigned long max_load_move, struct sched_domain *sd,
2936 enum cpu_idle_type idle, int *all_pinned,
2937 int *this_best_prio, struct rq_iterator *iterator)
2939 int loops = 0, pulled = 0, pinned = 0;
2940 struct task_struct *p;
2941 long rem_load_move = max_load_move;
2943 if (max_load_move == 0)
2949 * Start the load-balancing iterator:
2951 p = iterator->start(iterator->arg);
2953 if (!p || loops++ > sysctl_sched_nr_migrate)
2956 if ((p->se.load.weight >> 1) > rem_load_move ||
2957 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2958 p = iterator->next(iterator->arg);
2962 pull_task(busiest, p, this_rq, this_cpu);
2964 rem_load_move -= p->se.load.weight;
2967 * We only want to steal up to the prescribed amount of weighted load.
2969 if (rem_load_move > 0) {
2970 if (p->prio < *this_best_prio)
2971 *this_best_prio = p->prio;
2972 p = iterator->next(iterator->arg);
2977 * Right now, this is one of only two places pull_task() is called,
2978 * so we can safely collect pull_task() stats here rather than
2979 * inside pull_task().
2981 schedstat_add(sd, lb_gained[idle], pulled);
2984 *all_pinned = pinned;
2986 return max_load_move - rem_load_move;
2990 * move_tasks tries to move up to max_load_move weighted load from busiest to
2991 * this_rq, as part of a balancing operation within domain "sd".
2992 * Returns 1 if successful and 0 otherwise.
2994 * Called with both runqueues locked.
2996 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2997 unsigned long max_load_move,
2998 struct sched_domain *sd, enum cpu_idle_type idle,
3001 const struct sched_class *class = sched_class_highest;
3002 unsigned long total_load_moved = 0;
3003 int this_best_prio = this_rq->curr->prio;
3007 class->load_balance(this_rq, this_cpu, busiest,
3008 max_load_move - total_load_moved,
3009 sd, idle, all_pinned, &this_best_prio);
3010 class = class->next;
3012 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3015 } while (class && max_load_move > total_load_moved);
3017 return total_load_moved > 0;
3021 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3022 struct sched_domain *sd, enum cpu_idle_type idle,
3023 struct rq_iterator *iterator)
3025 struct task_struct *p = iterator->start(iterator->arg);
3029 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3030 pull_task(busiest, p, this_rq, this_cpu);
3032 * Right now, this is only the second place pull_task()
3033 * is called, so we can safely collect pull_task()
3034 * stats here rather than inside pull_task().
3036 schedstat_inc(sd, lb_gained[idle]);
3040 p = iterator->next(iterator->arg);
3047 * move_one_task tries to move exactly one task from busiest to this_rq, as
3048 * part of active balancing operations within "domain".
3049 * Returns 1 if successful and 0 otherwise.
3051 * Called with both runqueues locked.
3053 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3054 struct sched_domain *sd, enum cpu_idle_type idle)
3056 const struct sched_class *class;
3058 for (class = sched_class_highest; class; class = class->next)
3059 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3066 * find_busiest_group finds and returns the busiest CPU group within the
3067 * domain. It calculates and returns the amount of weighted load which
3068 * should be moved to restore balance via the imbalance parameter.
3070 static struct sched_group *
3071 find_busiest_group(struct sched_domain *sd, int this_cpu,
3072 unsigned long *imbalance, enum cpu_idle_type idle,
3073 int *sd_idle, const cpumask_t *cpus, int *balance)
3075 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3076 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3077 unsigned long max_pull;
3078 unsigned long busiest_load_per_task, busiest_nr_running;
3079 unsigned long this_load_per_task, this_nr_running;
3080 int load_idx, group_imb = 0;
3081 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3082 int power_savings_balance = 1;
3083 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3084 unsigned long min_nr_running = ULONG_MAX;
3085 struct sched_group *group_min = NULL, *group_leader = NULL;
3088 max_load = this_load = total_load = total_pwr = 0;
3089 busiest_load_per_task = busiest_nr_running = 0;
3090 this_load_per_task = this_nr_running = 0;
3092 if (idle == CPU_NOT_IDLE)
3093 load_idx = sd->busy_idx;
3094 else if (idle == CPU_NEWLY_IDLE)
3095 load_idx = sd->newidle_idx;
3097 load_idx = sd->idle_idx;
3100 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3103 int __group_imb = 0;
3104 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3105 unsigned long sum_nr_running, sum_weighted_load;
3106 unsigned long sum_avg_load_per_task;
3107 unsigned long avg_load_per_task;
3109 local_group = cpumask_test_cpu(this_cpu,
3110 sched_group_cpus(group));
3113 balance_cpu = cpumask_first(sched_group_cpus(group));
3115 /* Tally up the load of all CPUs in the group */
3116 sum_weighted_load = sum_nr_running = avg_load = 0;
3117 sum_avg_load_per_task = avg_load_per_task = 0;
3120 min_cpu_load = ~0UL;
3122 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3123 struct rq *rq = cpu_rq(i);
3125 if (*sd_idle && rq->nr_running)
3128 /* Bias balancing toward cpus of our domain */
3130 if (idle_cpu(i) && !first_idle_cpu) {
3135 load = target_load(i, load_idx);
3137 load = source_load(i, load_idx);
3138 if (load > max_cpu_load)
3139 max_cpu_load = load;
3140 if (min_cpu_load > load)
3141 min_cpu_load = load;
3145 sum_nr_running += rq->nr_running;
3146 sum_weighted_load += weighted_cpuload(i);
3148 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3152 * First idle cpu or the first cpu(busiest) in this sched group
3153 * is eligible for doing load balancing at this and above
3154 * domains. In the newly idle case, we will allow all the cpu's
3155 * to do the newly idle load balance.
3157 if (idle != CPU_NEWLY_IDLE && local_group &&
3158 balance_cpu != this_cpu && balance) {
3163 total_load += avg_load;
3164 total_pwr += group->__cpu_power;
3166 /* Adjust by relative CPU power of the group */
3167 avg_load = sg_div_cpu_power(group,
3168 avg_load * SCHED_LOAD_SCALE);
3172 * Consider the group unbalanced when the imbalance is larger
3173 * than the average weight of two tasks.
3175 * APZ: with cgroup the avg task weight can vary wildly and
3176 * might not be a suitable number - should we keep a
3177 * normalized nr_running number somewhere that negates
3180 avg_load_per_task = sg_div_cpu_power(group,
3181 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3183 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3186 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3189 this_load = avg_load;
3191 this_nr_running = sum_nr_running;
3192 this_load_per_task = sum_weighted_load;
3193 } else if (avg_load > max_load &&
3194 (sum_nr_running > group_capacity || __group_imb)) {
3195 max_load = avg_load;
3197 busiest_nr_running = sum_nr_running;
3198 busiest_load_per_task = sum_weighted_load;
3199 group_imb = __group_imb;
3202 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3204 * Busy processors will not participate in power savings
3207 if (idle == CPU_NOT_IDLE ||
3208 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3212 * If the local group is idle or completely loaded
3213 * no need to do power savings balance at this domain
3215 if (local_group && (this_nr_running >= group_capacity ||
3217 power_savings_balance = 0;
3220 * If a group is already running at full capacity or idle,
3221 * don't include that group in power savings calculations
3223 if (!power_savings_balance || sum_nr_running >= group_capacity
3228 * Calculate the group which has the least non-idle load.
3229 * This is the group from where we need to pick up the load
3232 if ((sum_nr_running < min_nr_running) ||
3233 (sum_nr_running == min_nr_running &&
3234 cpumask_first(sched_group_cpus(group)) <
3235 cpumask_first(sched_group_cpus(group_min)))) {
3237 min_nr_running = sum_nr_running;
3238 min_load_per_task = sum_weighted_load /
3243 * Calculate the group which is almost near its
3244 * capacity but still has some space to pick up some load
3245 * from other group and save more power
3247 if (sum_nr_running <= group_capacity - 1) {
3248 if (sum_nr_running > leader_nr_running ||
3249 (sum_nr_running == leader_nr_running &&
3250 cpumask_first(sched_group_cpus(group)) >
3251 cpumask_first(sched_group_cpus(group_leader)))) {
3252 group_leader = group;
3253 leader_nr_running = sum_nr_running;
3258 group = group->next;
3259 } while (group != sd->groups);
3261 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3264 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3266 if (this_load >= avg_load ||
3267 100*max_load <= sd->imbalance_pct*this_load)
3270 busiest_load_per_task /= busiest_nr_running;
3272 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3275 * We're trying to get all the cpus to the average_load, so we don't
3276 * want to push ourselves above the average load, nor do we wish to
3277 * reduce the max loaded cpu below the average load, as either of these
3278 * actions would just result in more rebalancing later, and ping-pong
3279 * tasks around. Thus we look for the minimum possible imbalance.
3280 * Negative imbalances (*we* are more loaded than anyone else) will
3281 * be counted as no imbalance for these purposes -- we can't fix that
3282 * by pulling tasks to us. Be careful of negative numbers as they'll
3283 * appear as very large values with unsigned longs.
3285 if (max_load <= busiest_load_per_task)
3289 * In the presence of smp nice balancing, certain scenarios can have
3290 * max load less than avg load(as we skip the groups at or below
3291 * its cpu_power, while calculating max_load..)
3293 if (max_load < avg_load) {
3295 goto small_imbalance;
3298 /* Don't want to pull so many tasks that a group would go idle */
3299 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3301 /* How much load to actually move to equalise the imbalance */
3302 *imbalance = min(max_pull * busiest->__cpu_power,
3303 (avg_load - this_load) * this->__cpu_power)
3307 * if *imbalance is less than the average load per runnable task
3308 * there is no gaurantee that any tasks will be moved so we'll have
3309 * a think about bumping its value to force at least one task to be
3312 if (*imbalance < busiest_load_per_task) {
3313 unsigned long tmp, pwr_now, pwr_move;
3317 pwr_move = pwr_now = 0;
3319 if (this_nr_running) {
3320 this_load_per_task /= this_nr_running;
3321 if (busiest_load_per_task > this_load_per_task)
3324 this_load_per_task = cpu_avg_load_per_task(this_cpu);
3326 if (max_load - this_load + busiest_load_per_task >=
3327 busiest_load_per_task * imbn) {
3328 *imbalance = busiest_load_per_task;
3333 * OK, we don't have enough imbalance to justify moving tasks,
3334 * however we may be able to increase total CPU power used by
3338 pwr_now += busiest->__cpu_power *
3339 min(busiest_load_per_task, max_load);
3340 pwr_now += this->__cpu_power *
3341 min(this_load_per_task, this_load);
3342 pwr_now /= SCHED_LOAD_SCALE;
3344 /* Amount of load we'd subtract */
3345 tmp = sg_div_cpu_power(busiest,
3346 busiest_load_per_task * SCHED_LOAD_SCALE);
3348 pwr_move += busiest->__cpu_power *
3349 min(busiest_load_per_task, max_load - tmp);
3351 /* Amount of load we'd add */
3352 if (max_load * busiest->__cpu_power <
3353 busiest_load_per_task * SCHED_LOAD_SCALE)
3354 tmp = sg_div_cpu_power(this,
3355 max_load * busiest->__cpu_power);
3357 tmp = sg_div_cpu_power(this,
3358 busiest_load_per_task * SCHED_LOAD_SCALE);
3359 pwr_move += this->__cpu_power *
3360 min(this_load_per_task, this_load + tmp);
3361 pwr_move /= SCHED_LOAD_SCALE;
3363 /* Move if we gain throughput */
3364 if (pwr_move > pwr_now)
3365 *imbalance = busiest_load_per_task;
3371 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3372 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3375 if (this == group_leader && group_leader != group_min) {
3376 *imbalance = min_load_per_task;
3386 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3389 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3390 unsigned long imbalance, const cpumask_t *cpus)
3392 struct rq *busiest = NULL, *rq;
3393 unsigned long max_load = 0;
3396 for_each_cpu(i, sched_group_cpus(group)) {
3399 if (!cpu_isset(i, *cpus))
3403 wl = weighted_cpuload(i);
3405 if (rq->nr_running == 1 && wl > imbalance)
3408 if (wl > max_load) {
3418 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3419 * so long as it is large enough.
3421 #define MAX_PINNED_INTERVAL 512
3424 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3425 * tasks if there is an imbalance.
3427 static int load_balance(int this_cpu, struct rq *this_rq,
3428 struct sched_domain *sd, enum cpu_idle_type idle,
3429 int *balance, cpumask_t *cpus)
3431 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3432 struct sched_group *group;
3433 unsigned long imbalance;
3435 unsigned long flags;
3440 * When power savings policy is enabled for the parent domain, idle
3441 * sibling can pick up load irrespective of busy siblings. In this case,
3442 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3443 * portraying it as CPU_NOT_IDLE.
3445 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3446 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3449 schedstat_inc(sd, lb_count[idle]);
3453 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3460 schedstat_inc(sd, lb_nobusyg[idle]);
3464 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3466 schedstat_inc(sd, lb_nobusyq[idle]);
3470 BUG_ON(busiest == this_rq);
3472 schedstat_add(sd, lb_imbalance[idle], imbalance);
3475 if (busiest->nr_running > 1) {
3477 * Attempt to move tasks. If find_busiest_group has found
3478 * an imbalance but busiest->nr_running <= 1, the group is
3479 * still unbalanced. ld_moved simply stays zero, so it is
3480 * correctly treated as an imbalance.
3482 local_irq_save(flags);
3483 double_rq_lock(this_rq, busiest);
3484 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3485 imbalance, sd, idle, &all_pinned);
3486 double_rq_unlock(this_rq, busiest);
3487 local_irq_restore(flags);
3490 * some other cpu did the load balance for us.
3492 if (ld_moved && this_cpu != smp_processor_id())
3493 resched_cpu(this_cpu);
3495 /* All tasks on this runqueue were pinned by CPU affinity */
3496 if (unlikely(all_pinned)) {
3497 cpu_clear(cpu_of(busiest), *cpus);
3498 if (!cpus_empty(*cpus))
3505 schedstat_inc(sd, lb_failed[idle]);
3506 sd->nr_balance_failed++;
3508 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3510 spin_lock_irqsave(&busiest->lock, flags);
3512 /* don't kick the migration_thread, if the curr
3513 * task on busiest cpu can't be moved to this_cpu
3515 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3516 spin_unlock_irqrestore(&busiest->lock, flags);
3518 goto out_one_pinned;
3521 if (!busiest->active_balance) {
3522 busiest->active_balance = 1;
3523 busiest->push_cpu = this_cpu;
3526 spin_unlock_irqrestore(&busiest->lock, flags);
3528 wake_up_process(busiest->migration_thread);
3531 * We've kicked active balancing, reset the failure
3534 sd->nr_balance_failed = sd->cache_nice_tries+1;
3537 sd->nr_balance_failed = 0;
3539 if (likely(!active_balance)) {
3540 /* We were unbalanced, so reset the balancing interval */
3541 sd->balance_interval = sd->min_interval;
3544 * If we've begun active balancing, start to back off. This
3545 * case may not be covered by the all_pinned logic if there
3546 * is only 1 task on the busy runqueue (because we don't call
3549 if (sd->balance_interval < sd->max_interval)
3550 sd->balance_interval *= 2;
3553 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3554 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3560 schedstat_inc(sd, lb_balanced[idle]);
3562 sd->nr_balance_failed = 0;
3565 /* tune up the balancing interval */
3566 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3567 (sd->balance_interval < sd->max_interval))
3568 sd->balance_interval *= 2;
3570 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3571 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3582 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3583 * tasks if there is an imbalance.
3585 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3586 * this_rq is locked.
3589 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3592 struct sched_group *group;
3593 struct rq *busiest = NULL;
3594 unsigned long imbalance;
3602 * When power savings policy is enabled for the parent domain, idle
3603 * sibling can pick up load irrespective of busy siblings. In this case,
3604 * let the state of idle sibling percolate up as IDLE, instead of
3605 * portraying it as CPU_NOT_IDLE.
3607 if (sd->flags & SD_SHARE_CPUPOWER &&
3608 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3611 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3613 update_shares_locked(this_rq, sd);
3614 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3615 &sd_idle, cpus, NULL);
3617 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3621 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3623 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3627 BUG_ON(busiest == this_rq);
3629 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3632 if (busiest->nr_running > 1) {
3633 /* Attempt to move tasks */
3634 double_lock_balance(this_rq, busiest);
3635 /* this_rq->clock is already updated */
3636 update_rq_clock(busiest);
3637 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3638 imbalance, sd, CPU_NEWLY_IDLE,
3640 double_unlock_balance(this_rq, busiest);
3642 if (unlikely(all_pinned)) {
3643 cpu_clear(cpu_of(busiest), *cpus);
3644 if (!cpus_empty(*cpus))
3650 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3651 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3652 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3655 sd->nr_balance_failed = 0;
3657 update_shares_locked(this_rq, sd);
3661 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3662 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3663 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3665 sd->nr_balance_failed = 0;
3671 * idle_balance is called by schedule() if this_cpu is about to become
3672 * idle. Attempts to pull tasks from other CPUs.
3674 static void idle_balance(int this_cpu, struct rq *this_rq)
3676 struct sched_domain *sd;
3677 int pulled_task = -1;
3678 unsigned long next_balance = jiffies + HZ;
3679 cpumask_var_t tmpmask;
3681 if (!alloc_cpumask_var(&tmpmask, GFP_ATOMIC))
3684 for_each_domain(this_cpu, sd) {
3685 unsigned long interval;
3687 if (!(sd->flags & SD_LOAD_BALANCE))
3690 if (sd->flags & SD_BALANCE_NEWIDLE)
3691 /* If we've pulled tasks over stop searching: */
3692 pulled_task = load_balance_newidle(this_cpu, this_rq,
3695 interval = msecs_to_jiffies(sd->balance_interval);
3696 if (time_after(next_balance, sd->last_balance + interval))
3697 next_balance = sd->last_balance + interval;
3701 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3703 * We are going idle. next_balance may be set based on
3704 * a busy processor. So reset next_balance.
3706 this_rq->next_balance = next_balance;
3708 free_cpumask_var(tmpmask);
3712 * active_load_balance is run by migration threads. It pushes running tasks
3713 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3714 * running on each physical CPU where possible, and avoids physical /
3715 * logical imbalances.
3717 * Called with busiest_rq locked.
3719 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3721 int target_cpu = busiest_rq->push_cpu;
3722 struct sched_domain *sd;
3723 struct rq *target_rq;
3725 /* Is there any task to move? */
3726 if (busiest_rq->nr_running <= 1)
3729 target_rq = cpu_rq(target_cpu);
3732 * This condition is "impossible", if it occurs
3733 * we need to fix it. Originally reported by
3734 * Bjorn Helgaas on a 128-cpu setup.
3736 BUG_ON(busiest_rq == target_rq);
3738 /* move a task from busiest_rq to target_rq */
3739 double_lock_balance(busiest_rq, target_rq);
3740 update_rq_clock(busiest_rq);
3741 update_rq_clock(target_rq);
3743 /* Search for an sd spanning us and the target CPU. */
3744 for_each_domain(target_cpu, sd) {
3745 if ((sd->flags & SD_LOAD_BALANCE) &&
3746 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
3751 schedstat_inc(sd, alb_count);
3753 if (move_one_task(target_rq, target_cpu, busiest_rq,
3755 schedstat_inc(sd, alb_pushed);
3757 schedstat_inc(sd, alb_failed);
3759 double_unlock_balance(busiest_rq, target_rq);
3764 atomic_t load_balancer;
3765 cpumask_var_t cpu_mask;
3766 } nohz ____cacheline_aligned = {
3767 .load_balancer = ATOMIC_INIT(-1),
3771 * This routine will try to nominate the ilb (idle load balancing)
3772 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3773 * load balancing on behalf of all those cpus. If all the cpus in the system
3774 * go into this tickless mode, then there will be no ilb owner (as there is
3775 * no need for one) and all the cpus will sleep till the next wakeup event
3778 * For the ilb owner, tick is not stopped. And this tick will be used
3779 * for idle load balancing. ilb owner will still be part of
3782 * While stopping the tick, this cpu will become the ilb owner if there
3783 * is no other owner. And will be the owner till that cpu becomes busy
3784 * or if all cpus in the system stop their ticks at which point
3785 * there is no need for ilb owner.
3787 * When the ilb owner becomes busy, it nominates another owner, during the
3788 * next busy scheduler_tick()
3790 int select_nohz_load_balancer(int stop_tick)
3792 int cpu = smp_processor_id();
3795 cpumask_set_cpu(cpu, nohz.cpu_mask);
3796 cpu_rq(cpu)->in_nohz_recently = 1;
3799 * If we are going offline and still the leader, give up!
3801 if (!cpu_active(cpu) &&
3802 atomic_read(&nohz.load_balancer) == cpu) {
3803 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3808 /* time for ilb owner also to sleep */
3809 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
3810 if (atomic_read(&nohz.load_balancer) == cpu)
3811 atomic_set(&nohz.load_balancer, -1);
3815 if (atomic_read(&nohz.load_balancer) == -1) {
3816 /* make me the ilb owner */
3817 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3819 } else if (atomic_read(&nohz.load_balancer) == cpu)
3822 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
3825 cpumask_clear_cpu(cpu, nohz.cpu_mask);
3827 if (atomic_read(&nohz.load_balancer) == cpu)
3828 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3835 static DEFINE_SPINLOCK(balancing);
3838 * It checks each scheduling domain to see if it is due to be balanced,
3839 * and initiates a balancing operation if so.
3841 * Balancing parameters are set up in arch_init_sched_domains.
3843 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3846 struct rq *rq = cpu_rq(cpu);
3847 unsigned long interval;
3848 struct sched_domain *sd;
3849 /* Earliest time when we have to do rebalance again */
3850 unsigned long next_balance = jiffies + 60*HZ;
3851 int update_next_balance = 0;
3855 /* Fails alloc? Rebalancing probably not a priority right now. */
3856 if (!alloc_cpumask_var(&tmp, GFP_ATOMIC))
3859 for_each_domain(cpu, sd) {
3860 if (!(sd->flags & SD_LOAD_BALANCE))
3863 interval = sd->balance_interval;
3864 if (idle != CPU_IDLE)
3865 interval *= sd->busy_factor;
3867 /* scale ms to jiffies */
3868 interval = msecs_to_jiffies(interval);
3869 if (unlikely(!interval))
3871 if (interval > HZ*NR_CPUS/10)
3872 interval = HZ*NR_CPUS/10;
3874 need_serialize = sd->flags & SD_SERIALIZE;
3876 if (need_serialize) {
3877 if (!spin_trylock(&balancing))
3881 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3882 if (load_balance(cpu, rq, sd, idle, &balance, tmp)) {
3884 * We've pulled tasks over so either we're no
3885 * longer idle, or one of our SMT siblings is
3888 idle = CPU_NOT_IDLE;
3890 sd->last_balance = jiffies;
3893 spin_unlock(&balancing);
3895 if (time_after(next_balance, sd->last_balance + interval)) {
3896 next_balance = sd->last_balance + interval;
3897 update_next_balance = 1;
3901 * Stop the load balance at this level. There is another
3902 * CPU in our sched group which is doing load balancing more
3910 * next_balance will be updated only when there is a need.
3911 * When the cpu is attached to null domain for ex, it will not be
3914 if (likely(update_next_balance))
3915 rq->next_balance = next_balance;
3917 free_cpumask_var(tmp);
3921 * run_rebalance_domains is triggered when needed from the scheduler tick.
3922 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3923 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3925 static void run_rebalance_domains(struct softirq_action *h)
3927 int this_cpu = smp_processor_id();
3928 struct rq *this_rq = cpu_rq(this_cpu);
3929 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3930 CPU_IDLE : CPU_NOT_IDLE;
3932 rebalance_domains(this_cpu, idle);
3936 * If this cpu is the owner for idle load balancing, then do the
3937 * balancing on behalf of the other idle cpus whose ticks are
3940 if (this_rq->idle_at_tick &&
3941 atomic_read(&nohz.load_balancer) == this_cpu) {
3945 for_each_cpu(balance_cpu, nohz.cpu_mask) {
3946 if (balance_cpu == this_cpu)
3950 * If this cpu gets work to do, stop the load balancing
3951 * work being done for other cpus. Next load
3952 * balancing owner will pick it up.
3957 rebalance_domains(balance_cpu, CPU_IDLE);
3959 rq = cpu_rq(balance_cpu);
3960 if (time_after(this_rq->next_balance, rq->next_balance))
3961 this_rq->next_balance = rq->next_balance;
3968 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3970 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3971 * idle load balancing owner or decide to stop the periodic load balancing,
3972 * if the whole system is idle.
3974 static inline void trigger_load_balance(struct rq *rq, int cpu)
3978 * If we were in the nohz mode recently and busy at the current
3979 * scheduler tick, then check if we need to nominate new idle
3982 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3983 rq->in_nohz_recently = 0;
3985 if (atomic_read(&nohz.load_balancer) == cpu) {
3986 cpumask_clear_cpu(cpu, nohz.cpu_mask);
3987 atomic_set(&nohz.load_balancer, -1);
3990 if (atomic_read(&nohz.load_balancer) == -1) {
3992 * simple selection for now: Nominate the
3993 * first cpu in the nohz list to be the next
3996 * TBD: Traverse the sched domains and nominate
3997 * the nearest cpu in the nohz.cpu_mask.
3999 int ilb = cpumask_first(nohz.cpu_mask);
4001 if (ilb < nr_cpu_ids)
4007 * If this cpu is idle and doing idle load balancing for all the
4008 * cpus with ticks stopped, is it time for that to stop?
4010 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4011 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4017 * If this cpu is idle and the idle load balancing is done by
4018 * someone else, then no need raise the SCHED_SOFTIRQ
4020 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4021 cpumask_test_cpu(cpu, nohz.cpu_mask))
4024 if (time_after_eq(jiffies, rq->next_balance))
4025 raise_softirq(SCHED_SOFTIRQ);
4028 #else /* CONFIG_SMP */
4031 * on UP we do not need to balance between CPUs:
4033 static inline void idle_balance(int cpu, struct rq *rq)
4039 DEFINE_PER_CPU(struct kernel_stat, kstat);
4041 EXPORT_PER_CPU_SYMBOL(kstat);
4044 * Return any ns on the sched_clock that have not yet been banked in
4045 * @p in case that task is currently running.
4047 unsigned long long task_delta_exec(struct task_struct *p)
4049 unsigned long flags;
4053 rq = task_rq_lock(p, &flags);
4055 if (task_current(rq, p)) {
4058 update_rq_clock(rq);
4059 delta_exec = rq->clock - p->se.exec_start;
4060 if ((s64)delta_exec > 0)
4064 task_rq_unlock(rq, &flags);
4070 * Account user cpu time to a process.
4071 * @p: the process that the cpu time gets accounted to
4072 * @cputime: the cpu time spent in user space since the last update
4074 void account_user_time(struct task_struct *p, cputime_t cputime)
4076 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4079 p->utime = cputime_add(p->utime, cputime);
4080 account_group_user_time(p, cputime);
4082 /* Add user time to cpustat. */
4083 tmp = cputime_to_cputime64(cputime);
4084 if (TASK_NICE(p) > 0)
4085 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4087 cpustat->user = cputime64_add(cpustat->user, tmp);
4088 /* Account for user time used */
4089 acct_update_integrals(p);
4093 * Account guest cpu time to a process.
4094 * @p: the process that the cpu time gets accounted to
4095 * @cputime: the cpu time spent in virtual machine since the last update
4097 static void account_guest_time(struct task_struct *p, cputime_t cputime)
4100 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4102 tmp = cputime_to_cputime64(cputime);
4104 p->utime = cputime_add(p->utime, cputime);
4105 account_group_user_time(p, cputime);
4106 p->gtime = cputime_add(p->gtime, cputime);
4108 cpustat->user = cputime64_add(cpustat->user, tmp);
4109 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4113 * Account scaled user cpu time to a process.
4114 * @p: the process that the cpu time gets accounted to
4115 * @cputime: the cpu time spent in user space since the last update
4117 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4119 p->utimescaled = cputime_add(p->utimescaled, cputime);
4123 * Account system cpu time to a process.
4124 * @p: the process that the cpu time gets accounted to
4125 * @hardirq_offset: the offset to subtract from hardirq_count()
4126 * @cputime: the cpu time spent in kernel space since the last update
4128 void account_system_time(struct task_struct *p, int hardirq_offset,
4131 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4132 struct rq *rq = this_rq();
4135 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4136 account_guest_time(p, cputime);
4140 p->stime = cputime_add(p->stime, cputime);
4141 account_group_system_time(p, cputime);
4143 /* Add system time to cpustat. */
4144 tmp = cputime_to_cputime64(cputime);
4145 if (hardirq_count() - hardirq_offset)
4146 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4147 else if (softirq_count())
4148 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4149 else if (p != rq->idle)
4150 cpustat->system = cputime64_add(cpustat->system, tmp);
4151 else if (atomic_read(&rq->nr_iowait) > 0)
4152 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4154 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4155 /* Account for system time used */
4156 acct_update_integrals(p);
4160 * Account scaled system cpu time to a process.
4161 * @p: the process that the cpu time gets accounted to
4162 * @hardirq_offset: the offset to subtract from hardirq_count()
4163 * @cputime: the cpu time spent in kernel space since the last update
4165 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4167 p->stimescaled = cputime_add(p->stimescaled, cputime);
4171 * Account for involuntary wait time.
4172 * @p: the process from which the cpu time has been stolen
4173 * @steal: the cpu time spent in involuntary wait
4175 void account_steal_time(struct task_struct *p, cputime_t steal)
4177 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4178 cputime64_t tmp = cputime_to_cputime64(steal);
4179 struct rq *rq = this_rq();
4181 if (p == rq->idle) {
4182 p->stime = cputime_add(p->stime, steal);
4183 if (atomic_read(&rq->nr_iowait) > 0)
4184 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4186 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4188 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4192 * Use precise platform statistics if available:
4194 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4195 cputime_t task_utime(struct task_struct *p)
4200 cputime_t task_stime(struct task_struct *p)
4205 cputime_t task_utime(struct task_struct *p)
4207 clock_t utime = cputime_to_clock_t(p->utime),
4208 total = utime + cputime_to_clock_t(p->stime);
4212 * Use CFS's precise accounting:
4214 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
4218 do_div(temp, total);
4220 utime = (clock_t)temp;
4222 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
4223 return p->prev_utime;
4226 cputime_t task_stime(struct task_struct *p)
4231 * Use CFS's precise accounting. (we subtract utime from
4232 * the total, to make sure the total observed by userspace
4233 * grows monotonically - apps rely on that):
4235 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
4236 cputime_to_clock_t(task_utime(p));
4239 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
4241 return p->prev_stime;
4245 inline cputime_t task_gtime(struct task_struct *p)
4251 * This function gets called by the timer code, with HZ frequency.
4252 * We call it with interrupts disabled.
4254 * It also gets called by the fork code, when changing the parent's
4257 void scheduler_tick(void)
4259 int cpu = smp_processor_id();
4260 struct rq *rq = cpu_rq(cpu);
4261 struct task_struct *curr = rq->curr;
4265 spin_lock(&rq->lock);
4266 update_rq_clock(rq);
4267 update_cpu_load(rq);
4268 curr->sched_class->task_tick(rq, curr, 0);
4269 spin_unlock(&rq->lock);
4272 rq->idle_at_tick = idle_cpu(cpu);
4273 trigger_load_balance(rq, cpu);
4277 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4278 defined(CONFIG_PREEMPT_TRACER))
4280 static inline unsigned long get_parent_ip(unsigned long addr)
4282 if (in_lock_functions(addr)) {
4283 addr = CALLER_ADDR2;
4284 if (in_lock_functions(addr))
4285 addr = CALLER_ADDR3;
4290 void __kprobes add_preempt_count(int val)
4292 #ifdef CONFIG_DEBUG_PREEMPT
4296 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4299 preempt_count() += val;
4300 #ifdef CONFIG_DEBUG_PREEMPT
4302 * Spinlock count overflowing soon?
4304 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4307 if (preempt_count() == val)
4308 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4310 EXPORT_SYMBOL(add_preempt_count);
4312 void __kprobes sub_preempt_count(int val)
4314 #ifdef CONFIG_DEBUG_PREEMPT
4318 if (DEBUG_LOCKS_WARN_ON(val > preempt_count() - (!!kernel_locked())))
4321 * Is the spinlock portion underflowing?
4323 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4324 !(preempt_count() & PREEMPT_MASK)))
4328 if (preempt_count() == val)
4329 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4330 preempt_count() -= val;
4332 EXPORT_SYMBOL(sub_preempt_count);
4337 * Print scheduling while atomic bug:
4339 static noinline void __schedule_bug(struct task_struct *prev)
4341 struct pt_regs *regs = get_irq_regs();
4343 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4344 prev->comm, prev->pid, preempt_count());
4346 debug_show_held_locks(prev);
4348 if (irqs_disabled())
4349 print_irqtrace_events(prev);
4358 * Various schedule()-time debugging checks and statistics:
4360 static inline void schedule_debug(struct task_struct *prev)
4363 * Test if we are atomic. Since do_exit() needs to call into
4364 * schedule() atomically, we ignore that path for now.
4365 * Otherwise, whine if we are scheduling when we should not be.
4367 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4368 __schedule_bug(prev);
4370 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4372 schedstat_inc(this_rq(), sched_count);
4373 #ifdef CONFIG_SCHEDSTATS
4374 if (unlikely(prev->lock_depth >= 0)) {
4375 schedstat_inc(this_rq(), bkl_count);
4376 schedstat_inc(prev, sched_info.bkl_count);
4382 * Pick up the highest-prio task:
4384 static inline struct task_struct *
4385 pick_next_task(struct rq *rq, struct task_struct *prev)
4387 const struct sched_class *class;
4388 struct task_struct *p;
4391 * Optimization: we know that if all tasks are in
4392 * the fair class we can call that function directly:
4394 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4395 p = fair_sched_class.pick_next_task(rq);
4400 class = sched_class_highest;
4402 p = class->pick_next_task(rq);
4406 * Will never be NULL as the idle class always
4407 * returns a non-NULL p:
4409 class = class->next;
4414 * schedule() is the main scheduler function.
4416 asmlinkage void __sched schedule(void)
4418 struct task_struct *prev, *next;
4419 unsigned long *switch_count;
4425 cpu = smp_processor_id();
4429 switch_count = &prev->nivcsw;
4431 release_kernel_lock(prev);
4432 need_resched_nonpreemptible:
4434 schedule_debug(prev);
4436 if (sched_feat(HRTICK))
4439 spin_lock_irq(&rq->lock);
4440 update_rq_clock(rq);
4441 clear_tsk_need_resched(prev);
4443 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4444 if (unlikely(signal_pending_state(prev->state, prev)))
4445 prev->state = TASK_RUNNING;
4447 deactivate_task(rq, prev, 1);
4448 switch_count = &prev->nvcsw;
4452 if (prev->sched_class->pre_schedule)
4453 prev->sched_class->pre_schedule(rq, prev);
4456 if (unlikely(!rq->nr_running))
4457 idle_balance(cpu, rq);
4459 prev->sched_class->put_prev_task(rq, prev);
4460 next = pick_next_task(rq, prev);
4462 if (likely(prev != next)) {
4463 sched_info_switch(prev, next);
4469 context_switch(rq, prev, next); /* unlocks the rq */
4471 * the context switch might have flipped the stack from under
4472 * us, hence refresh the local variables.
4474 cpu = smp_processor_id();
4477 spin_unlock_irq(&rq->lock);
4479 if (unlikely(reacquire_kernel_lock(current) < 0))
4480 goto need_resched_nonpreemptible;
4482 preempt_enable_no_resched();
4483 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4486 EXPORT_SYMBOL(schedule);
4488 #ifdef CONFIG_PREEMPT
4490 * this is the entry point to schedule() from in-kernel preemption
4491 * off of preempt_enable. Kernel preemptions off return from interrupt
4492 * occur there and call schedule directly.
4494 asmlinkage void __sched preempt_schedule(void)
4496 struct thread_info *ti = current_thread_info();
4499 * If there is a non-zero preempt_count or interrupts are disabled,
4500 * we do not want to preempt the current task. Just return..
4502 if (likely(ti->preempt_count || irqs_disabled()))
4506 add_preempt_count(PREEMPT_ACTIVE);
4508 sub_preempt_count(PREEMPT_ACTIVE);
4511 * Check again in case we missed a preemption opportunity
4512 * between schedule and now.
4515 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4517 EXPORT_SYMBOL(preempt_schedule);
4520 * this is the entry point to schedule() from kernel preemption
4521 * off of irq context.
4522 * Note, that this is called and return with irqs disabled. This will
4523 * protect us against recursive calling from irq.
4525 asmlinkage void __sched preempt_schedule_irq(void)
4527 struct thread_info *ti = current_thread_info();
4529 /* Catch callers which need to be fixed */
4530 BUG_ON(ti->preempt_count || !irqs_disabled());
4533 add_preempt_count(PREEMPT_ACTIVE);
4536 local_irq_disable();
4537 sub_preempt_count(PREEMPT_ACTIVE);
4540 * Check again in case we missed a preemption opportunity
4541 * between schedule and now.
4544 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4547 #endif /* CONFIG_PREEMPT */
4549 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4552 return try_to_wake_up(curr->private, mode, sync);
4554 EXPORT_SYMBOL(default_wake_function);
4557 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4558 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4559 * number) then we wake all the non-exclusive tasks and one exclusive task.
4561 * There are circumstances in which we can try to wake a task which has already
4562 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4563 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4565 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4566 int nr_exclusive, int sync, void *key)
4568 wait_queue_t *curr, *next;
4570 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4571 unsigned flags = curr->flags;
4573 if (curr->func(curr, mode, sync, key) &&
4574 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4580 * __wake_up - wake up threads blocked on a waitqueue.
4582 * @mode: which threads
4583 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4584 * @key: is directly passed to the wakeup function
4586 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4587 int nr_exclusive, void *key)
4589 unsigned long flags;
4591 spin_lock_irqsave(&q->lock, flags);
4592 __wake_up_common(q, mode, nr_exclusive, 0, key);
4593 spin_unlock_irqrestore(&q->lock, flags);
4595 EXPORT_SYMBOL(__wake_up);
4598 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4600 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4602 __wake_up_common(q, mode, 1, 0, NULL);
4606 * __wake_up_sync - wake up threads blocked on a waitqueue.
4608 * @mode: which threads
4609 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4611 * The sync wakeup differs that the waker knows that it will schedule
4612 * away soon, so while the target thread will be woken up, it will not
4613 * be migrated to another CPU - ie. the two threads are 'synchronized'
4614 * with each other. This can prevent needless bouncing between CPUs.
4616 * On UP it can prevent extra preemption.
4619 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4621 unsigned long flags;
4627 if (unlikely(!nr_exclusive))
4630 spin_lock_irqsave(&q->lock, flags);
4631 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4632 spin_unlock_irqrestore(&q->lock, flags);
4634 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4637 * complete: - signals a single thread waiting on this completion
4638 * @x: holds the state of this particular completion
4640 * This will wake up a single thread waiting on this completion. Threads will be
4641 * awakened in the same order in which they were queued.
4643 * See also complete_all(), wait_for_completion() and related routines.
4645 void complete(struct completion *x)
4647 unsigned long flags;
4649 spin_lock_irqsave(&x->wait.lock, flags);
4651 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4652 spin_unlock_irqrestore(&x->wait.lock, flags);
4654 EXPORT_SYMBOL(complete);
4657 * complete_all: - signals all threads waiting on this completion
4658 * @x: holds the state of this particular completion
4660 * This will wake up all threads waiting on this particular completion event.
4662 void complete_all(struct completion *x)
4664 unsigned long flags;
4666 spin_lock_irqsave(&x->wait.lock, flags);
4667 x->done += UINT_MAX/2;
4668 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4669 spin_unlock_irqrestore(&x->wait.lock, flags);
4671 EXPORT_SYMBOL(complete_all);
4673 static inline long __sched
4674 do_wait_for_common(struct completion *x, long timeout, int state)
4677 DECLARE_WAITQUEUE(wait, current);
4679 wait.flags |= WQ_FLAG_EXCLUSIVE;
4680 __add_wait_queue_tail(&x->wait, &wait);
4682 if (signal_pending_state(state, current)) {
4683 timeout = -ERESTARTSYS;
4686 __set_current_state(state);
4687 spin_unlock_irq(&x->wait.lock);
4688 timeout = schedule_timeout(timeout);
4689 spin_lock_irq(&x->wait.lock);
4690 } while (!x->done && timeout);
4691 __remove_wait_queue(&x->wait, &wait);
4696 return timeout ?: 1;
4700 wait_for_common(struct completion *x, long timeout, int state)
4704 spin_lock_irq(&x->wait.lock);
4705 timeout = do_wait_for_common(x, timeout, state);
4706 spin_unlock_irq(&x->wait.lock);
4711 * wait_for_completion: - waits for completion of a task
4712 * @x: holds the state of this particular completion
4714 * This waits to be signaled for completion of a specific task. It is NOT
4715 * interruptible and there is no timeout.
4717 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4718 * and interrupt capability. Also see complete().
4720 void __sched wait_for_completion(struct completion *x)
4722 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4724 EXPORT_SYMBOL(wait_for_completion);
4727 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4728 * @x: holds the state of this particular completion
4729 * @timeout: timeout value in jiffies
4731 * This waits for either a completion of a specific task to be signaled or for a
4732 * specified timeout to expire. The timeout is in jiffies. It is not
4735 unsigned long __sched
4736 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4738 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4740 EXPORT_SYMBOL(wait_for_completion_timeout);
4743 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4744 * @x: holds the state of this particular completion
4746 * This waits for completion of a specific task to be signaled. It is
4749 int __sched wait_for_completion_interruptible(struct completion *x)
4751 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4752 if (t == -ERESTARTSYS)
4756 EXPORT_SYMBOL(wait_for_completion_interruptible);
4759 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4760 * @x: holds the state of this particular completion
4761 * @timeout: timeout value in jiffies
4763 * This waits for either a completion of a specific task to be signaled or for a
4764 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4766 unsigned long __sched
4767 wait_for_completion_interruptible_timeout(struct completion *x,
4768 unsigned long timeout)
4770 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4772 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4775 * wait_for_completion_killable: - waits for completion of a task (killable)
4776 * @x: holds the state of this particular completion
4778 * This waits to be signaled for completion of a specific task. It can be
4779 * interrupted by a kill signal.
4781 int __sched wait_for_completion_killable(struct completion *x)
4783 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4784 if (t == -ERESTARTSYS)
4788 EXPORT_SYMBOL(wait_for_completion_killable);
4791 * try_wait_for_completion - try to decrement a completion without blocking
4792 * @x: completion structure
4794 * Returns: 0 if a decrement cannot be done without blocking
4795 * 1 if a decrement succeeded.
4797 * If a completion is being used as a counting completion,
4798 * attempt to decrement the counter without blocking. This
4799 * enables us to avoid waiting if the resource the completion
4800 * is protecting is not available.
4802 bool try_wait_for_completion(struct completion *x)
4806 spin_lock_irq(&x->wait.lock);
4811 spin_unlock_irq(&x->wait.lock);
4814 EXPORT_SYMBOL(try_wait_for_completion);
4817 * completion_done - Test to see if a completion has any waiters
4818 * @x: completion structure
4820 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4821 * 1 if there are no waiters.
4824 bool completion_done(struct completion *x)
4828 spin_lock_irq(&x->wait.lock);
4831 spin_unlock_irq(&x->wait.lock);
4834 EXPORT_SYMBOL(completion_done);
4837 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4839 unsigned long flags;
4842 init_waitqueue_entry(&wait, current);
4844 __set_current_state(state);
4846 spin_lock_irqsave(&q->lock, flags);
4847 __add_wait_queue(q, &wait);
4848 spin_unlock(&q->lock);
4849 timeout = schedule_timeout(timeout);
4850 spin_lock_irq(&q->lock);
4851 __remove_wait_queue(q, &wait);
4852 spin_unlock_irqrestore(&q->lock, flags);
4857 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4859 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4861 EXPORT_SYMBOL(interruptible_sleep_on);
4864 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4866 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4868 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4870 void __sched sleep_on(wait_queue_head_t *q)
4872 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4874 EXPORT_SYMBOL(sleep_on);
4876 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4878 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4880 EXPORT_SYMBOL(sleep_on_timeout);
4882 #ifdef CONFIG_RT_MUTEXES
4885 * rt_mutex_setprio - set the current priority of a task
4887 * @prio: prio value (kernel-internal form)
4889 * This function changes the 'effective' priority of a task. It does
4890 * not touch ->normal_prio like __setscheduler().
4892 * Used by the rt_mutex code to implement priority inheritance logic.
4894 void rt_mutex_setprio(struct task_struct *p, int prio)
4896 unsigned long flags;
4897 int oldprio, on_rq, running;
4899 const struct sched_class *prev_class = p->sched_class;
4901 BUG_ON(prio < 0 || prio > MAX_PRIO);
4903 rq = task_rq_lock(p, &flags);
4904 update_rq_clock(rq);
4907 on_rq = p->se.on_rq;
4908 running = task_current(rq, p);
4910 dequeue_task(rq, p, 0);
4912 p->sched_class->put_prev_task(rq, p);
4915 p->sched_class = &rt_sched_class;
4917 p->sched_class = &fair_sched_class;
4922 p->sched_class->set_curr_task(rq);
4924 enqueue_task(rq, p, 0);
4926 check_class_changed(rq, p, prev_class, oldprio, running);
4928 task_rq_unlock(rq, &flags);
4933 void set_user_nice(struct task_struct *p, long nice)
4935 int old_prio, delta, on_rq;
4936 unsigned long flags;
4939 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4942 * We have to be careful, if called from sys_setpriority(),
4943 * the task might be in the middle of scheduling on another CPU.
4945 rq = task_rq_lock(p, &flags);
4946 update_rq_clock(rq);
4948 * The RT priorities are set via sched_setscheduler(), but we still
4949 * allow the 'normal' nice value to be set - but as expected
4950 * it wont have any effect on scheduling until the task is
4951 * SCHED_FIFO/SCHED_RR:
4953 if (task_has_rt_policy(p)) {
4954 p->static_prio = NICE_TO_PRIO(nice);
4957 on_rq = p->se.on_rq;
4959 dequeue_task(rq, p, 0);
4961 p->static_prio = NICE_TO_PRIO(nice);
4964 p->prio = effective_prio(p);
4965 delta = p->prio - old_prio;
4968 enqueue_task(rq, p, 0);
4970 * If the task increased its priority or is running and
4971 * lowered its priority, then reschedule its CPU:
4973 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4974 resched_task(rq->curr);
4977 task_rq_unlock(rq, &flags);
4979 EXPORT_SYMBOL(set_user_nice);
4982 * can_nice - check if a task can reduce its nice value
4986 int can_nice(const struct task_struct *p, const int nice)
4988 /* convert nice value [19,-20] to rlimit style value [1,40] */
4989 int nice_rlim = 20 - nice;
4991 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4992 capable(CAP_SYS_NICE));
4995 #ifdef __ARCH_WANT_SYS_NICE
4998 * sys_nice - change the priority of the current process.
4999 * @increment: priority increment
5001 * sys_setpriority is a more generic, but much slower function that
5002 * does similar things.
5004 asmlinkage long sys_nice(int increment)
5009 * Setpriority might change our priority at the same moment.
5010 * We don't have to worry. Conceptually one call occurs first
5011 * and we have a single winner.
5013 if (increment < -40)
5018 nice = PRIO_TO_NICE(current->static_prio) + increment;
5024 if (increment < 0 && !can_nice(current, nice))
5027 retval = security_task_setnice(current, nice);
5031 set_user_nice(current, nice);
5038 * task_prio - return the priority value of a given task.
5039 * @p: the task in question.
5041 * This is the priority value as seen by users in /proc.
5042 * RT tasks are offset by -200. Normal tasks are centered
5043 * around 0, value goes from -16 to +15.
5045 int task_prio(const struct task_struct *p)
5047 return p->prio - MAX_RT_PRIO;
5051 * task_nice - return the nice value of a given task.
5052 * @p: the task in question.
5054 int task_nice(const struct task_struct *p)
5056 return TASK_NICE(p);
5058 EXPORT_SYMBOL(task_nice);
5061 * idle_cpu - is a given cpu idle currently?
5062 * @cpu: the processor in question.
5064 int idle_cpu(int cpu)
5066 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5070 * idle_task - return the idle task for a given cpu.
5071 * @cpu: the processor in question.
5073 struct task_struct *idle_task(int cpu)
5075 return cpu_rq(cpu)->idle;
5079 * find_process_by_pid - find a process with a matching PID value.
5080 * @pid: the pid in question.
5082 static struct task_struct *find_process_by_pid(pid_t pid)
5084 return pid ? find_task_by_vpid(pid) : current;
5087 /* Actually do priority change: must hold rq lock. */
5089 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5091 BUG_ON(p->se.on_rq);
5094 switch (p->policy) {
5098 p->sched_class = &fair_sched_class;
5102 p->sched_class = &rt_sched_class;
5106 p->rt_priority = prio;
5107 p->normal_prio = normal_prio(p);
5108 /* we are holding p->pi_lock already */
5109 p->prio = rt_mutex_getprio(p);
5113 static int __sched_setscheduler(struct task_struct *p, int policy,
5114 struct sched_param *param, bool user)
5116 int retval, oldprio, oldpolicy = -1, on_rq, running;
5117 unsigned long flags;
5118 const struct sched_class *prev_class = p->sched_class;
5121 /* may grab non-irq protected spin_locks */
5122 BUG_ON(in_interrupt());
5124 /* double check policy once rq lock held */
5126 policy = oldpolicy = p->policy;
5127 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5128 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5129 policy != SCHED_IDLE)
5132 * Valid priorities for SCHED_FIFO and SCHED_RR are
5133 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5134 * SCHED_BATCH and SCHED_IDLE is 0.
5136 if (param->sched_priority < 0 ||
5137 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5138 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5140 if (rt_policy(policy) != (param->sched_priority != 0))
5144 * Allow unprivileged RT tasks to decrease priority:
5146 if (user && !capable(CAP_SYS_NICE)) {
5147 if (rt_policy(policy)) {
5148 unsigned long rlim_rtprio;
5150 if (!lock_task_sighand(p, &flags))
5152 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5153 unlock_task_sighand(p, &flags);
5155 /* can't set/change the rt policy */
5156 if (policy != p->policy && !rlim_rtprio)
5159 /* can't increase priority */
5160 if (param->sched_priority > p->rt_priority &&
5161 param->sched_priority > rlim_rtprio)
5165 * Like positive nice levels, dont allow tasks to
5166 * move out of SCHED_IDLE either:
5168 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5171 /* can't change other user's priorities */
5172 if ((current->euid != p->euid) &&
5173 (current->euid != p->uid))
5178 #ifdef CONFIG_RT_GROUP_SCHED
5180 * Do not allow realtime tasks into groups that have no runtime
5183 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5184 task_group(p)->rt_bandwidth.rt_runtime == 0)
5188 retval = security_task_setscheduler(p, policy, param);
5194 * make sure no PI-waiters arrive (or leave) while we are
5195 * changing the priority of the task:
5197 spin_lock_irqsave(&p->pi_lock, flags);
5199 * To be able to change p->policy safely, the apropriate
5200 * runqueue lock must be held.
5202 rq = __task_rq_lock(p);
5203 /* recheck policy now with rq lock held */
5204 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5205 policy = oldpolicy = -1;
5206 __task_rq_unlock(rq);
5207 spin_unlock_irqrestore(&p->pi_lock, flags);
5210 update_rq_clock(rq);
5211 on_rq = p->se.on_rq;
5212 running = task_current(rq, p);
5214 deactivate_task(rq, p, 0);
5216 p->sched_class->put_prev_task(rq, p);
5219 __setscheduler(rq, p, policy, param->sched_priority);
5222 p->sched_class->set_curr_task(rq);
5224 activate_task(rq, p, 0);
5226 check_class_changed(rq, p, prev_class, oldprio, running);
5228 __task_rq_unlock(rq);
5229 spin_unlock_irqrestore(&p->pi_lock, flags);
5231 rt_mutex_adjust_pi(p);
5237 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5238 * @p: the task in question.
5239 * @policy: new policy.
5240 * @param: structure containing the new RT priority.
5242 * NOTE that the task may be already dead.
5244 int sched_setscheduler(struct task_struct *p, int policy,
5245 struct sched_param *param)
5247 return __sched_setscheduler(p, policy, param, true);
5249 EXPORT_SYMBOL_GPL(sched_setscheduler);
5252 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5253 * @p: the task in question.
5254 * @policy: new policy.
5255 * @param: structure containing the new RT priority.
5257 * Just like sched_setscheduler, only don't bother checking if the
5258 * current context has permission. For example, this is needed in
5259 * stop_machine(): we create temporary high priority worker threads,
5260 * but our caller might not have that capability.
5262 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5263 struct sched_param *param)
5265 return __sched_setscheduler(p, policy, param, false);
5269 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5271 struct sched_param lparam;
5272 struct task_struct *p;
5275 if (!param || pid < 0)
5277 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5282 p = find_process_by_pid(pid);
5284 retval = sched_setscheduler(p, policy, &lparam);
5291 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5292 * @pid: the pid in question.
5293 * @policy: new policy.
5294 * @param: structure containing the new RT priority.
5297 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5299 /* negative values for policy are not valid */
5303 return do_sched_setscheduler(pid, policy, param);
5307 * sys_sched_setparam - set/change the RT priority of a thread
5308 * @pid: the pid in question.
5309 * @param: structure containing the new RT priority.
5311 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5313 return do_sched_setscheduler(pid, -1, param);
5317 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5318 * @pid: the pid in question.
5320 asmlinkage long sys_sched_getscheduler(pid_t pid)
5322 struct task_struct *p;
5329 read_lock(&tasklist_lock);
5330 p = find_process_by_pid(pid);
5332 retval = security_task_getscheduler(p);
5336 read_unlock(&tasklist_lock);
5341 * sys_sched_getscheduler - get the RT priority of a thread
5342 * @pid: the pid in question.
5343 * @param: structure containing the RT priority.
5345 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5347 struct sched_param lp;
5348 struct task_struct *p;
5351 if (!param || pid < 0)
5354 read_lock(&tasklist_lock);
5355 p = find_process_by_pid(pid);
5360 retval = security_task_getscheduler(p);
5364 lp.sched_priority = p->rt_priority;
5365 read_unlock(&tasklist_lock);
5368 * This one might sleep, we cannot do it with a spinlock held ...
5370 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5375 read_unlock(&tasklist_lock);
5379 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5381 cpumask_var_t cpus_allowed, new_mask;
5382 struct task_struct *p;
5386 read_lock(&tasklist_lock);
5388 p = find_process_by_pid(pid);
5390 read_unlock(&tasklist_lock);
5396 * It is not safe to call set_cpus_allowed with the
5397 * tasklist_lock held. We will bump the task_struct's
5398 * usage count and then drop tasklist_lock.
5401 read_unlock(&tasklist_lock);
5403 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5407 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5409 goto out_free_cpus_allowed;
5412 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5413 !capable(CAP_SYS_NICE))
5416 retval = security_task_setscheduler(p, 0, NULL);
5420 cpuset_cpus_allowed(p, cpus_allowed);
5421 cpumask_and(new_mask, in_mask, cpus_allowed);
5423 retval = set_cpus_allowed_ptr(p, new_mask);
5426 cpuset_cpus_allowed(p, cpus_allowed);
5427 if (!cpumask_subset(new_mask, cpus_allowed)) {
5429 * We must have raced with a concurrent cpuset
5430 * update. Just reset the cpus_allowed to the
5431 * cpuset's cpus_allowed
5433 cpumask_copy(new_mask, cpus_allowed);
5438 free_cpumask_var(new_mask);
5439 out_free_cpus_allowed:
5440 free_cpumask_var(cpus_allowed);
5447 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5448 cpumask_t *new_mask)
5450 if (len < sizeof(cpumask_t)) {
5451 memset(new_mask, 0, sizeof(cpumask_t));
5452 } else if (len > sizeof(cpumask_t)) {
5453 len = sizeof(cpumask_t);
5455 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5459 * sys_sched_setaffinity - set the cpu affinity of a process
5460 * @pid: pid of the process
5461 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5462 * @user_mask_ptr: user-space pointer to the new cpu mask
5464 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5465 unsigned long __user *user_mask_ptr)
5467 cpumask_var_t new_mask;
5470 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5473 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5475 retval = sched_setaffinity(pid, new_mask);
5476 free_cpumask_var(new_mask);
5480 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5482 struct task_struct *p;
5486 read_lock(&tasklist_lock);
5489 p = find_process_by_pid(pid);
5493 retval = security_task_getscheduler(p);
5497 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5500 read_unlock(&tasklist_lock);
5507 * sys_sched_getaffinity - get the cpu affinity of a process
5508 * @pid: pid of the process
5509 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5510 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5512 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5513 unsigned long __user *user_mask_ptr)
5518 if (len < cpumask_size())
5521 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5524 ret = sched_getaffinity(pid, mask);
5526 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
5529 ret = cpumask_size();
5531 free_cpumask_var(mask);
5537 * sys_sched_yield - yield the current processor to other threads.
5539 * This function yields the current CPU to other tasks. If there are no
5540 * other threads running on this CPU then this function will return.
5542 asmlinkage long sys_sched_yield(void)
5544 struct rq *rq = this_rq_lock();
5546 schedstat_inc(rq, yld_count);
5547 current->sched_class->yield_task(rq);
5550 * Since we are going to call schedule() anyway, there's
5551 * no need to preempt or enable interrupts:
5553 __release(rq->lock);
5554 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5555 _raw_spin_unlock(&rq->lock);
5556 preempt_enable_no_resched();
5563 static void __cond_resched(void)
5565 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5566 __might_sleep(__FILE__, __LINE__);
5569 * The BKS might be reacquired before we have dropped
5570 * PREEMPT_ACTIVE, which could trigger a second
5571 * cond_resched() call.
5574 add_preempt_count(PREEMPT_ACTIVE);
5576 sub_preempt_count(PREEMPT_ACTIVE);
5577 } while (need_resched());
5580 int __sched _cond_resched(void)
5582 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5583 system_state == SYSTEM_RUNNING) {
5589 EXPORT_SYMBOL(_cond_resched);
5592 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5593 * call schedule, and on return reacquire the lock.
5595 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5596 * operations here to prevent schedule() from being called twice (once via
5597 * spin_unlock(), once by hand).
5599 int cond_resched_lock(spinlock_t *lock)
5601 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5604 if (spin_needbreak(lock) || resched) {
5606 if (resched && need_resched())
5615 EXPORT_SYMBOL(cond_resched_lock);
5617 int __sched cond_resched_softirq(void)
5619 BUG_ON(!in_softirq());
5621 if (need_resched() && system_state == SYSTEM_RUNNING) {
5629 EXPORT_SYMBOL(cond_resched_softirq);
5632 * yield - yield the current processor to other threads.
5634 * This is a shortcut for kernel-space yielding - it marks the
5635 * thread runnable and calls sys_sched_yield().
5637 void __sched yield(void)
5639 set_current_state(TASK_RUNNING);
5642 EXPORT_SYMBOL(yield);
5645 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5646 * that process accounting knows that this is a task in IO wait state.
5648 * But don't do that if it is a deliberate, throttling IO wait (this task
5649 * has set its backing_dev_info: the queue against which it should throttle)
5651 void __sched io_schedule(void)
5653 struct rq *rq = &__raw_get_cpu_var(runqueues);
5655 delayacct_blkio_start();
5656 atomic_inc(&rq->nr_iowait);
5658 atomic_dec(&rq->nr_iowait);
5659 delayacct_blkio_end();
5661 EXPORT_SYMBOL(io_schedule);
5663 long __sched io_schedule_timeout(long timeout)
5665 struct rq *rq = &__raw_get_cpu_var(runqueues);
5668 delayacct_blkio_start();
5669 atomic_inc(&rq->nr_iowait);
5670 ret = schedule_timeout(timeout);
5671 atomic_dec(&rq->nr_iowait);
5672 delayacct_blkio_end();
5677 * sys_sched_get_priority_max - return maximum RT priority.
5678 * @policy: scheduling class.
5680 * this syscall returns the maximum rt_priority that can be used
5681 * by a given scheduling class.
5683 asmlinkage long sys_sched_get_priority_max(int policy)
5690 ret = MAX_USER_RT_PRIO-1;
5702 * sys_sched_get_priority_min - return minimum RT priority.
5703 * @policy: scheduling class.
5705 * this syscall returns the minimum rt_priority that can be used
5706 * by a given scheduling class.
5708 asmlinkage long sys_sched_get_priority_min(int policy)
5726 * sys_sched_rr_get_interval - return the default timeslice of a process.
5727 * @pid: pid of the process.
5728 * @interval: userspace pointer to the timeslice value.
5730 * this syscall writes the default timeslice value of a given process
5731 * into the user-space timespec buffer. A value of '0' means infinity.
5734 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5736 struct task_struct *p;
5737 unsigned int time_slice;
5745 read_lock(&tasklist_lock);
5746 p = find_process_by_pid(pid);
5750 retval = security_task_getscheduler(p);
5755 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5756 * tasks that are on an otherwise idle runqueue:
5759 if (p->policy == SCHED_RR) {
5760 time_slice = DEF_TIMESLICE;
5761 } else if (p->policy != SCHED_FIFO) {
5762 struct sched_entity *se = &p->se;
5763 unsigned long flags;
5766 rq = task_rq_lock(p, &flags);
5767 if (rq->cfs.load.weight)
5768 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5769 task_rq_unlock(rq, &flags);
5771 read_unlock(&tasklist_lock);
5772 jiffies_to_timespec(time_slice, &t);
5773 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5777 read_unlock(&tasklist_lock);
5781 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5783 void sched_show_task(struct task_struct *p)
5785 unsigned long free = 0;
5788 state = p->state ? __ffs(p->state) + 1 : 0;
5789 printk(KERN_INFO "%-13.13s %c", p->comm,
5790 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5791 #if BITS_PER_LONG == 32
5792 if (state == TASK_RUNNING)
5793 printk(KERN_CONT " running ");
5795 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5797 if (state == TASK_RUNNING)
5798 printk(KERN_CONT " running task ");
5800 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5802 #ifdef CONFIG_DEBUG_STACK_USAGE
5804 unsigned long *n = end_of_stack(p);
5807 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5810 printk(KERN_CONT "%5lu %5d %6d\n", free,
5811 task_pid_nr(p), task_pid_nr(p->real_parent));
5813 show_stack(p, NULL);
5816 void show_state_filter(unsigned long state_filter)
5818 struct task_struct *g, *p;
5820 #if BITS_PER_LONG == 32
5822 " task PC stack pid father\n");
5825 " task PC stack pid father\n");
5827 read_lock(&tasklist_lock);
5828 do_each_thread(g, p) {
5830 * reset the NMI-timeout, listing all files on a slow
5831 * console might take alot of time:
5833 touch_nmi_watchdog();
5834 if (!state_filter || (p->state & state_filter))
5836 } while_each_thread(g, p);
5838 touch_all_softlockup_watchdogs();
5840 #ifdef CONFIG_SCHED_DEBUG
5841 sysrq_sched_debug_show();
5843 read_unlock(&tasklist_lock);
5845 * Only show locks if all tasks are dumped:
5847 if (state_filter == -1)
5848 debug_show_all_locks();
5851 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5853 idle->sched_class = &idle_sched_class;
5857 * init_idle - set up an idle thread for a given CPU
5858 * @idle: task in question
5859 * @cpu: cpu the idle task belongs to
5861 * NOTE: this function does not set the idle thread's NEED_RESCHED
5862 * flag, to make booting more robust.
5864 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5866 struct rq *rq = cpu_rq(cpu);
5867 unsigned long flags;
5869 spin_lock_irqsave(&rq->lock, flags);
5872 idle->se.exec_start = sched_clock();
5874 idle->prio = idle->normal_prio = MAX_PRIO;
5875 idle->cpus_allowed = cpumask_of_cpu(cpu);
5876 __set_task_cpu(idle, cpu);
5878 rq->curr = rq->idle = idle;
5879 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5882 spin_unlock_irqrestore(&rq->lock, flags);
5884 /* Set the preempt count _outside_ the spinlocks! */
5885 #if defined(CONFIG_PREEMPT)
5886 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5888 task_thread_info(idle)->preempt_count = 0;
5891 * The idle tasks have their own, simple scheduling class:
5893 idle->sched_class = &idle_sched_class;
5894 ftrace_retfunc_init_task(idle);
5898 * In a system that switches off the HZ timer nohz_cpu_mask
5899 * indicates which cpus entered this state. This is used
5900 * in the rcu update to wait only for active cpus. For system
5901 * which do not switch off the HZ timer nohz_cpu_mask should
5902 * always be CPU_BITS_NONE.
5904 cpumask_var_t nohz_cpu_mask;
5907 * Increase the granularity value when there are more CPUs,
5908 * because with more CPUs the 'effective latency' as visible
5909 * to users decreases. But the relationship is not linear,
5910 * so pick a second-best guess by going with the log2 of the
5913 * This idea comes from the SD scheduler of Con Kolivas:
5915 static inline void sched_init_granularity(void)
5917 unsigned int factor = 1 + ilog2(num_online_cpus());
5918 const unsigned long limit = 200000000;
5920 sysctl_sched_min_granularity *= factor;
5921 if (sysctl_sched_min_granularity > limit)
5922 sysctl_sched_min_granularity = limit;
5924 sysctl_sched_latency *= factor;
5925 if (sysctl_sched_latency > limit)
5926 sysctl_sched_latency = limit;
5928 sysctl_sched_wakeup_granularity *= factor;
5930 sysctl_sched_shares_ratelimit *= factor;
5935 * This is how migration works:
5937 * 1) we queue a struct migration_req structure in the source CPU's
5938 * runqueue and wake up that CPU's migration thread.
5939 * 2) we down() the locked semaphore => thread blocks.
5940 * 3) migration thread wakes up (implicitly it forces the migrated
5941 * thread off the CPU)
5942 * 4) it gets the migration request and checks whether the migrated
5943 * task is still in the wrong runqueue.
5944 * 5) if it's in the wrong runqueue then the migration thread removes
5945 * it and puts it into the right queue.
5946 * 6) migration thread up()s the semaphore.
5947 * 7) we wake up and the migration is done.
5951 * Change a given task's CPU affinity. Migrate the thread to a
5952 * proper CPU and schedule it away if the CPU it's executing on
5953 * is removed from the allowed bitmask.
5955 * NOTE: the caller must have a valid reference to the task, the
5956 * task must not exit() & deallocate itself prematurely. The
5957 * call is not atomic; no spinlocks may be held.
5959 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5961 struct migration_req req;
5962 unsigned long flags;
5966 rq = task_rq_lock(p, &flags);
5967 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5972 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5973 !cpus_equal(p->cpus_allowed, *new_mask))) {
5978 if (p->sched_class->set_cpus_allowed)
5979 p->sched_class->set_cpus_allowed(p, new_mask);
5981 p->cpus_allowed = *new_mask;
5982 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5985 /* Can the task run on the task's current CPU? If so, we're done */
5986 if (cpu_isset(task_cpu(p), *new_mask))
5989 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
5990 /* Need help from migration thread: drop lock and wait. */
5991 task_rq_unlock(rq, &flags);
5992 wake_up_process(rq->migration_thread);
5993 wait_for_completion(&req.done);
5994 tlb_migrate_finish(p->mm);
5998 task_rq_unlock(rq, &flags);
6002 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6005 * Move (not current) task off this cpu, onto dest cpu. We're doing
6006 * this because either it can't run here any more (set_cpus_allowed()
6007 * away from this CPU, or CPU going down), or because we're
6008 * attempting to rebalance this task on exec (sched_exec).
6010 * So we race with normal scheduler movements, but that's OK, as long
6011 * as the task is no longer on this CPU.
6013 * Returns non-zero if task was successfully migrated.
6015 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6017 struct rq *rq_dest, *rq_src;
6020 if (unlikely(!cpu_active(dest_cpu)))
6023 rq_src = cpu_rq(src_cpu);
6024 rq_dest = cpu_rq(dest_cpu);
6026 double_rq_lock(rq_src, rq_dest);
6027 /* Already moved. */
6028 if (task_cpu(p) != src_cpu)
6030 /* Affinity changed (again). */
6031 if (!cpu_isset(dest_cpu, p->cpus_allowed))
6034 on_rq = p->se.on_rq;
6036 deactivate_task(rq_src, p, 0);
6038 set_task_cpu(p, dest_cpu);
6040 activate_task(rq_dest, p, 0);
6041 check_preempt_curr(rq_dest, p, 0);
6046 double_rq_unlock(rq_src, rq_dest);
6051 * migration_thread - this is a highprio system thread that performs
6052 * thread migration by bumping thread off CPU then 'pushing' onto
6055 static int migration_thread(void *data)
6057 int cpu = (long)data;
6061 BUG_ON(rq->migration_thread != current);
6063 set_current_state(TASK_INTERRUPTIBLE);
6064 while (!kthread_should_stop()) {
6065 struct migration_req *req;
6066 struct list_head *head;
6068 spin_lock_irq(&rq->lock);
6070 if (cpu_is_offline(cpu)) {
6071 spin_unlock_irq(&rq->lock);
6075 if (rq->active_balance) {
6076 active_load_balance(rq, cpu);
6077 rq->active_balance = 0;
6080 head = &rq->migration_queue;
6082 if (list_empty(head)) {
6083 spin_unlock_irq(&rq->lock);
6085 set_current_state(TASK_INTERRUPTIBLE);
6088 req = list_entry(head->next, struct migration_req, list);
6089 list_del_init(head->next);
6091 spin_unlock(&rq->lock);
6092 __migrate_task(req->task, cpu, req->dest_cpu);
6095 complete(&req->done);
6097 __set_current_state(TASK_RUNNING);
6101 /* Wait for kthread_stop */
6102 set_current_state(TASK_INTERRUPTIBLE);
6103 while (!kthread_should_stop()) {
6105 set_current_state(TASK_INTERRUPTIBLE);
6107 __set_current_state(TASK_RUNNING);
6111 #ifdef CONFIG_HOTPLUG_CPU
6113 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6117 local_irq_disable();
6118 ret = __migrate_task(p, src_cpu, dest_cpu);
6124 * Figure out where task on dead CPU should go, use force if necessary.
6126 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6128 unsigned long flags;
6131 /* FIXME: Use cpumask_of_node here. */
6132 cpumask_t _nodemask = node_to_cpumask(cpu_to_node(dead_cpu));
6133 const struct cpumask *nodemask = &_nodemask;
6136 /* Look for allowed, online CPU in same node. */
6137 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
6138 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6141 /* Any allowed, online CPU? */
6142 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
6143 if (dest_cpu < nr_cpu_ids)
6146 /* No more Mr. Nice Guy. */
6147 if (dest_cpu >= nr_cpu_ids) {
6148 rq = task_rq_lock(p, &flags);
6149 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
6150 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
6151 task_rq_unlock(rq, &flags);
6154 * Don't tell them about moving exiting tasks or
6155 * kernel threads (both mm NULL), since they never
6158 if (p->mm && printk_ratelimit()) {
6159 printk(KERN_INFO "process %d (%s) no "
6160 "longer affine to cpu%d\n",
6161 task_pid_nr(p), p->comm, dead_cpu);
6166 /* It can have affinity changed while we were choosing. */
6167 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
6172 * While a dead CPU has no uninterruptible tasks queued at this point,
6173 * it might still have a nonzero ->nr_uninterruptible counter, because
6174 * for performance reasons the counter is not stricly tracking tasks to
6175 * their home CPUs. So we just add the counter to another CPU's counter,
6176 * to keep the global sum constant after CPU-down:
6178 static void migrate_nr_uninterruptible(struct rq *rq_src)
6180 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
6181 unsigned long flags;
6183 local_irq_save(flags);
6184 double_rq_lock(rq_src, rq_dest);
6185 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6186 rq_src->nr_uninterruptible = 0;
6187 double_rq_unlock(rq_src, rq_dest);
6188 local_irq_restore(flags);
6191 /* Run through task list and migrate tasks from the dead cpu. */
6192 static void migrate_live_tasks(int src_cpu)
6194 struct task_struct *p, *t;
6196 read_lock(&tasklist_lock);
6198 do_each_thread(t, p) {
6202 if (task_cpu(p) == src_cpu)
6203 move_task_off_dead_cpu(src_cpu, p);
6204 } while_each_thread(t, p);
6206 read_unlock(&tasklist_lock);
6210 * Schedules idle task to be the next runnable task on current CPU.
6211 * It does so by boosting its priority to highest possible.
6212 * Used by CPU offline code.
6214 void sched_idle_next(void)
6216 int this_cpu = smp_processor_id();
6217 struct rq *rq = cpu_rq(this_cpu);
6218 struct task_struct *p = rq->idle;
6219 unsigned long flags;
6221 /* cpu has to be offline */
6222 BUG_ON(cpu_online(this_cpu));
6225 * Strictly not necessary since rest of the CPUs are stopped by now
6226 * and interrupts disabled on the current cpu.
6228 spin_lock_irqsave(&rq->lock, flags);
6230 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6232 update_rq_clock(rq);
6233 activate_task(rq, p, 0);
6235 spin_unlock_irqrestore(&rq->lock, flags);
6239 * Ensures that the idle task is using init_mm right before its cpu goes
6242 void idle_task_exit(void)
6244 struct mm_struct *mm = current->active_mm;
6246 BUG_ON(cpu_online(smp_processor_id()));
6249 switch_mm(mm, &init_mm, current);
6253 /* called under rq->lock with disabled interrupts */
6254 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6256 struct rq *rq = cpu_rq(dead_cpu);
6258 /* Must be exiting, otherwise would be on tasklist. */
6259 BUG_ON(!p->exit_state);
6261 /* Cannot have done final schedule yet: would have vanished. */
6262 BUG_ON(p->state == TASK_DEAD);
6267 * Drop lock around migration; if someone else moves it,
6268 * that's OK. No task can be added to this CPU, so iteration is
6271 spin_unlock_irq(&rq->lock);
6272 move_task_off_dead_cpu(dead_cpu, p);
6273 spin_lock_irq(&rq->lock);
6278 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6279 static void migrate_dead_tasks(unsigned int dead_cpu)
6281 struct rq *rq = cpu_rq(dead_cpu);
6282 struct task_struct *next;
6285 if (!rq->nr_running)
6287 update_rq_clock(rq);
6288 next = pick_next_task(rq, rq->curr);
6291 next->sched_class->put_prev_task(rq, next);
6292 migrate_dead(dead_cpu, next);
6296 #endif /* CONFIG_HOTPLUG_CPU */
6298 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6300 static struct ctl_table sd_ctl_dir[] = {
6302 .procname = "sched_domain",
6308 static struct ctl_table sd_ctl_root[] = {
6310 .ctl_name = CTL_KERN,
6311 .procname = "kernel",
6313 .child = sd_ctl_dir,
6318 static struct ctl_table *sd_alloc_ctl_entry(int n)
6320 struct ctl_table *entry =
6321 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6326 static void sd_free_ctl_entry(struct ctl_table **tablep)
6328 struct ctl_table *entry;
6331 * In the intermediate directories, both the child directory and
6332 * procname are dynamically allocated and could fail but the mode
6333 * will always be set. In the lowest directory the names are
6334 * static strings and all have proc handlers.
6336 for (entry = *tablep; entry->mode; entry++) {
6338 sd_free_ctl_entry(&entry->child);
6339 if (entry->proc_handler == NULL)
6340 kfree(entry->procname);
6348 set_table_entry(struct ctl_table *entry,
6349 const char *procname, void *data, int maxlen,
6350 mode_t mode, proc_handler *proc_handler)
6352 entry->procname = procname;
6354 entry->maxlen = maxlen;
6356 entry->proc_handler = proc_handler;
6359 static struct ctl_table *
6360 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6362 struct ctl_table *table = sd_alloc_ctl_entry(13);
6367 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6368 sizeof(long), 0644, proc_doulongvec_minmax);
6369 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6370 sizeof(long), 0644, proc_doulongvec_minmax);
6371 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6372 sizeof(int), 0644, proc_dointvec_minmax);
6373 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6374 sizeof(int), 0644, proc_dointvec_minmax);
6375 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6376 sizeof(int), 0644, proc_dointvec_minmax);
6377 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6378 sizeof(int), 0644, proc_dointvec_minmax);
6379 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6380 sizeof(int), 0644, proc_dointvec_minmax);
6381 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6382 sizeof(int), 0644, proc_dointvec_minmax);
6383 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6384 sizeof(int), 0644, proc_dointvec_minmax);
6385 set_table_entry(&table[9], "cache_nice_tries",
6386 &sd->cache_nice_tries,
6387 sizeof(int), 0644, proc_dointvec_minmax);
6388 set_table_entry(&table[10], "flags", &sd->flags,
6389 sizeof(int), 0644, proc_dointvec_minmax);
6390 set_table_entry(&table[11], "name", sd->name,
6391 CORENAME_MAX_SIZE, 0444, proc_dostring);
6392 /* &table[12] is terminator */
6397 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6399 struct ctl_table *entry, *table;
6400 struct sched_domain *sd;
6401 int domain_num = 0, i;
6404 for_each_domain(cpu, sd)
6406 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6411 for_each_domain(cpu, sd) {
6412 snprintf(buf, 32, "domain%d", i);
6413 entry->procname = kstrdup(buf, GFP_KERNEL);
6415 entry->child = sd_alloc_ctl_domain_table(sd);
6422 static struct ctl_table_header *sd_sysctl_header;
6423 static void register_sched_domain_sysctl(void)
6425 int i, cpu_num = num_online_cpus();
6426 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6429 WARN_ON(sd_ctl_dir[0].child);
6430 sd_ctl_dir[0].child = entry;
6435 for_each_online_cpu(i) {
6436 snprintf(buf, 32, "cpu%d", i);
6437 entry->procname = kstrdup(buf, GFP_KERNEL);
6439 entry->child = sd_alloc_ctl_cpu_table(i);
6443 WARN_ON(sd_sysctl_header);
6444 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6447 /* may be called multiple times per register */
6448 static void unregister_sched_domain_sysctl(void)
6450 if (sd_sysctl_header)
6451 unregister_sysctl_table(sd_sysctl_header);
6452 sd_sysctl_header = NULL;
6453 if (sd_ctl_dir[0].child)
6454 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6457 static void register_sched_domain_sysctl(void)
6460 static void unregister_sched_domain_sysctl(void)
6465 static void set_rq_online(struct rq *rq)
6468 const struct sched_class *class;
6470 cpumask_set_cpu(rq->cpu, rq->rd->online);
6473 for_each_class(class) {
6474 if (class->rq_online)
6475 class->rq_online(rq);
6480 static void set_rq_offline(struct rq *rq)
6483 const struct sched_class *class;
6485 for_each_class(class) {
6486 if (class->rq_offline)
6487 class->rq_offline(rq);
6490 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6496 * migration_call - callback that gets triggered when a CPU is added.
6497 * Here we can start up the necessary migration thread for the new CPU.
6499 static int __cpuinit
6500 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6502 struct task_struct *p;
6503 int cpu = (long)hcpu;
6504 unsigned long flags;
6509 case CPU_UP_PREPARE:
6510 case CPU_UP_PREPARE_FROZEN:
6511 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6514 kthread_bind(p, cpu);
6515 /* Must be high prio: stop_machine expects to yield to it. */
6516 rq = task_rq_lock(p, &flags);
6517 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6518 task_rq_unlock(rq, &flags);
6519 cpu_rq(cpu)->migration_thread = p;
6523 case CPU_ONLINE_FROZEN:
6524 /* Strictly unnecessary, as first user will wake it. */
6525 wake_up_process(cpu_rq(cpu)->migration_thread);
6527 /* Update our root-domain */
6529 spin_lock_irqsave(&rq->lock, flags);
6531 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6535 spin_unlock_irqrestore(&rq->lock, flags);
6538 #ifdef CONFIG_HOTPLUG_CPU
6539 case CPU_UP_CANCELED:
6540 case CPU_UP_CANCELED_FROZEN:
6541 if (!cpu_rq(cpu)->migration_thread)
6543 /* Unbind it from offline cpu so it can run. Fall thru. */
6544 kthread_bind(cpu_rq(cpu)->migration_thread,
6545 cpumask_any(cpu_online_mask));
6546 kthread_stop(cpu_rq(cpu)->migration_thread);
6547 cpu_rq(cpu)->migration_thread = NULL;
6551 case CPU_DEAD_FROZEN:
6552 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6553 migrate_live_tasks(cpu);
6555 kthread_stop(rq->migration_thread);
6556 rq->migration_thread = NULL;
6557 /* Idle task back to normal (off runqueue, low prio) */
6558 spin_lock_irq(&rq->lock);
6559 update_rq_clock(rq);
6560 deactivate_task(rq, rq->idle, 0);
6561 rq->idle->static_prio = MAX_PRIO;
6562 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6563 rq->idle->sched_class = &idle_sched_class;
6564 migrate_dead_tasks(cpu);
6565 spin_unlock_irq(&rq->lock);
6567 migrate_nr_uninterruptible(rq);
6568 BUG_ON(rq->nr_running != 0);
6571 * No need to migrate the tasks: it was best-effort if
6572 * they didn't take sched_hotcpu_mutex. Just wake up
6575 spin_lock_irq(&rq->lock);
6576 while (!list_empty(&rq->migration_queue)) {
6577 struct migration_req *req;
6579 req = list_entry(rq->migration_queue.next,
6580 struct migration_req, list);
6581 list_del_init(&req->list);
6582 complete(&req->done);
6584 spin_unlock_irq(&rq->lock);
6588 case CPU_DYING_FROZEN:
6589 /* Update our root-domain */
6591 spin_lock_irqsave(&rq->lock, flags);
6593 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6596 spin_unlock_irqrestore(&rq->lock, flags);
6603 /* Register at highest priority so that task migration (migrate_all_tasks)
6604 * happens before everything else.
6606 static struct notifier_block __cpuinitdata migration_notifier = {
6607 .notifier_call = migration_call,
6611 static int __init migration_init(void)
6613 void *cpu = (void *)(long)smp_processor_id();
6616 /* Start one for the boot CPU: */
6617 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6618 BUG_ON(err == NOTIFY_BAD);
6619 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6620 register_cpu_notifier(&migration_notifier);
6624 early_initcall(migration_init);
6629 #ifdef CONFIG_SCHED_DEBUG
6631 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6632 cpumask_t *groupmask)
6634 struct sched_group *group = sd->groups;
6637 cpulist_scnprintf(str, sizeof(str), *sched_domain_span(sd));
6638 cpus_clear(*groupmask);
6640 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6642 if (!(sd->flags & SD_LOAD_BALANCE)) {
6643 printk("does not load-balance\n");
6645 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6650 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6652 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6653 printk(KERN_ERR "ERROR: domain->span does not contain "
6656 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6657 printk(KERN_ERR "ERROR: domain->groups does not contain"
6661 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6665 printk(KERN_ERR "ERROR: group is NULL\n");
6669 if (!group->__cpu_power) {
6670 printk(KERN_CONT "\n");
6671 printk(KERN_ERR "ERROR: domain->cpu_power not "
6676 if (!cpumask_weight(sched_group_cpus(group))) {
6677 printk(KERN_CONT "\n");
6678 printk(KERN_ERR "ERROR: empty group\n");
6682 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6683 printk(KERN_CONT "\n");
6684 printk(KERN_ERR "ERROR: repeated CPUs\n");
6688 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6690 cpulist_scnprintf(str, sizeof(str), *sched_group_cpus(group));
6691 printk(KERN_CONT " %s", str);
6693 group = group->next;
6694 } while (group != sd->groups);
6695 printk(KERN_CONT "\n");
6697 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6698 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6701 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6702 printk(KERN_ERR "ERROR: parent span is not a superset "
6703 "of domain->span\n");
6707 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6709 cpumask_var_t groupmask;
6713 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6717 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6719 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6720 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6725 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6732 free_cpumask_var(groupmask);
6734 #else /* !CONFIG_SCHED_DEBUG */
6735 # define sched_domain_debug(sd, cpu) do { } while (0)
6736 #endif /* CONFIG_SCHED_DEBUG */
6738 static int sd_degenerate(struct sched_domain *sd)
6740 if (cpumask_weight(sched_domain_span(sd)) == 1)
6743 /* Following flags need at least 2 groups */
6744 if (sd->flags & (SD_LOAD_BALANCE |
6745 SD_BALANCE_NEWIDLE |
6749 SD_SHARE_PKG_RESOURCES)) {
6750 if (sd->groups != sd->groups->next)
6754 /* Following flags don't use groups */
6755 if (sd->flags & (SD_WAKE_IDLE |
6764 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6766 unsigned long cflags = sd->flags, pflags = parent->flags;
6768 if (sd_degenerate(parent))
6771 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6774 /* Does parent contain flags not in child? */
6775 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6776 if (cflags & SD_WAKE_AFFINE)
6777 pflags &= ~SD_WAKE_BALANCE;
6778 /* Flags needing groups don't count if only 1 group in parent */
6779 if (parent->groups == parent->groups->next) {
6780 pflags &= ~(SD_LOAD_BALANCE |
6781 SD_BALANCE_NEWIDLE |
6785 SD_SHARE_PKG_RESOURCES);
6787 if (~cflags & pflags)
6793 static void free_rootdomain(struct root_domain *rd)
6795 cpupri_cleanup(&rd->cpupri);
6797 free_cpumask_var(rd->rto_mask);
6798 free_cpumask_var(rd->online);
6799 free_cpumask_var(rd->span);
6803 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6805 unsigned long flags;
6807 spin_lock_irqsave(&rq->lock, flags);
6810 struct root_domain *old_rd = rq->rd;
6812 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6815 cpumask_clear_cpu(rq->cpu, old_rd->span);
6817 if (atomic_dec_and_test(&old_rd->refcount))
6818 free_rootdomain(old_rd);
6821 atomic_inc(&rd->refcount);
6824 cpumask_set_cpu(rq->cpu, rd->span);
6825 if (cpumask_test_cpu(rq->cpu, cpu_online_mask))
6828 spin_unlock_irqrestore(&rq->lock, flags);
6831 static int init_rootdomain(struct root_domain *rd, bool bootmem)
6833 memset(rd, 0, sizeof(*rd));
6836 alloc_bootmem_cpumask_var(&def_root_domain.span);
6837 alloc_bootmem_cpumask_var(&def_root_domain.online);
6838 alloc_bootmem_cpumask_var(&def_root_domain.rto_mask);
6839 cpupri_init(&rd->cpupri, true);
6843 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6845 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6847 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6850 if (cpupri_init(&rd->cpupri, false) != 0)
6855 free_cpumask_var(rd->rto_mask);
6857 free_cpumask_var(rd->online);
6859 free_cpumask_var(rd->span);
6865 static void init_defrootdomain(void)
6867 init_rootdomain(&def_root_domain, true);
6869 atomic_set(&def_root_domain.refcount, 1);
6872 static struct root_domain *alloc_rootdomain(void)
6874 struct root_domain *rd;
6876 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6880 if (init_rootdomain(rd, false) != 0) {
6889 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6890 * hold the hotplug lock.
6893 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6895 struct rq *rq = cpu_rq(cpu);
6896 struct sched_domain *tmp;
6898 /* Remove the sched domains which do not contribute to scheduling. */
6899 for (tmp = sd; tmp; ) {
6900 struct sched_domain *parent = tmp->parent;
6904 if (sd_parent_degenerate(tmp, parent)) {
6905 tmp->parent = parent->parent;
6907 parent->parent->child = tmp;
6912 if (sd && sd_degenerate(sd)) {
6918 sched_domain_debug(sd, cpu);
6920 rq_attach_root(rq, rd);
6921 rcu_assign_pointer(rq->sd, sd);
6924 /* cpus with isolated domains */
6925 static cpumask_var_t cpu_isolated_map;
6927 /* Setup the mask of cpus configured for isolated domains */
6928 static int __init isolated_cpu_setup(char *str)
6930 cpulist_parse(str, *cpu_isolated_map);
6934 __setup("isolcpus=", isolated_cpu_setup);
6937 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6938 * to a function which identifies what group(along with sched group) a CPU
6939 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6940 * (due to the fact that we keep track of groups covered with a cpumask_t).
6942 * init_sched_build_groups will build a circular linked list of the groups
6943 * covered by the given span, and will set each group's ->cpumask correctly,
6944 * and ->cpu_power to 0.
6947 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6948 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6949 struct sched_group **sg,
6950 cpumask_t *tmpmask),
6951 cpumask_t *covered, cpumask_t *tmpmask)
6953 struct sched_group *first = NULL, *last = NULL;
6956 cpus_clear(*covered);
6958 for_each_cpu(i, span) {
6959 struct sched_group *sg;
6960 int group = group_fn(i, cpu_map, &sg, tmpmask);
6963 if (cpumask_test_cpu(i, covered))
6966 cpumask_clear(sched_group_cpus(sg));
6967 sg->__cpu_power = 0;
6969 for_each_cpu(j, span) {
6970 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6973 cpu_set(j, *covered);
6974 cpumask_set_cpu(j, sched_group_cpus(sg));
6985 #define SD_NODES_PER_DOMAIN 16
6990 * find_next_best_node - find the next node to include in a sched_domain
6991 * @node: node whose sched_domain we're building
6992 * @used_nodes: nodes already in the sched_domain
6994 * Find the next node to include in a given scheduling domain. Simply
6995 * finds the closest node not already in the @used_nodes map.
6997 * Should use nodemask_t.
6999 static int find_next_best_node(int node, nodemask_t *used_nodes)
7001 int i, n, val, min_val, best_node = 0;
7005 for (i = 0; i < nr_node_ids; i++) {
7006 /* Start at @node */
7007 n = (node + i) % nr_node_ids;
7009 if (!nr_cpus_node(n))
7012 /* Skip already used nodes */
7013 if (node_isset(n, *used_nodes))
7016 /* Simple min distance search */
7017 val = node_distance(node, n);
7019 if (val < min_val) {
7025 node_set(best_node, *used_nodes);
7030 * sched_domain_node_span - get a cpumask for a node's sched_domain
7031 * @node: node whose cpumask we're constructing
7032 * @span: resulting cpumask
7034 * Given a node, construct a good cpumask for its sched_domain to span. It
7035 * should be one that prevents unnecessary balancing, but also spreads tasks
7038 static void sched_domain_node_span(int node, cpumask_t *span)
7040 nodemask_t used_nodes;
7041 node_to_cpumask_ptr(nodemask, node);
7045 nodes_clear(used_nodes);
7047 cpus_or(*span, *span, *nodemask);
7048 node_set(node, used_nodes);
7050 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7051 int next_node = find_next_best_node(node, &used_nodes);
7053 node_to_cpumask_ptr_next(nodemask, next_node);
7054 cpus_or(*span, *span, *nodemask);
7057 #endif /* CONFIG_NUMA */
7059 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7062 * The cpus mask in sched_group and sched_domain hangs off the end.
7063 * FIXME: use cpumask_var_t or dynamic percpu alloc to avoid wasting space
7064 * for nr_cpu_ids < CONFIG_NR_CPUS.
7066 struct static_sched_group {
7067 struct sched_group sg;
7068 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
7071 struct static_sched_domain {
7072 struct sched_domain sd;
7073 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
7077 * SMT sched-domains:
7079 #ifdef CONFIG_SCHED_SMT
7080 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
7081 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
7084 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7088 *sg = &per_cpu(sched_group_cpus, cpu).sg;
7091 #endif /* CONFIG_SCHED_SMT */
7094 * multi-core sched-domains:
7096 #ifdef CONFIG_SCHED_MC
7097 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
7098 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
7099 #endif /* CONFIG_SCHED_MC */
7101 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7103 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7108 *mask = per_cpu(cpu_sibling_map, cpu);
7109 cpus_and(*mask, *mask, *cpu_map);
7110 group = first_cpu(*mask);
7112 *sg = &per_cpu(sched_group_core, group).sg;
7115 #elif defined(CONFIG_SCHED_MC)
7117 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7121 *sg = &per_cpu(sched_group_core, cpu).sg;
7126 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
7127 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
7130 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7134 #ifdef CONFIG_SCHED_MC
7135 *mask = cpu_coregroup_map(cpu);
7136 cpus_and(*mask, *mask, *cpu_map);
7137 group = first_cpu(*mask);
7138 #elif defined(CONFIG_SCHED_SMT)
7139 *mask = per_cpu(cpu_sibling_map, cpu);
7140 cpus_and(*mask, *mask, *cpu_map);
7141 group = first_cpu(*mask);
7146 *sg = &per_cpu(sched_group_phys, group).sg;
7152 * The init_sched_build_groups can't handle what we want to do with node
7153 * groups, so roll our own. Now each node has its own list of groups which
7154 * gets dynamically allocated.
7156 static DEFINE_PER_CPU(struct sched_domain, node_domains);
7157 static struct sched_group ***sched_group_nodes_bycpu;
7159 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
7160 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
7162 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
7163 struct sched_group **sg, cpumask_t *nodemask)
7166 node_to_cpumask_ptr(pnodemask, cpu_to_node(cpu));
7168 cpus_and(*nodemask, *pnodemask, *cpu_map);
7169 group = first_cpu(*nodemask);
7172 *sg = &per_cpu(sched_group_allnodes, group).sg;
7176 static void init_numa_sched_groups_power(struct sched_group *group_head)
7178 struct sched_group *sg = group_head;
7184 for_each_cpu(j, sched_group_cpus(sg)) {
7185 struct sched_domain *sd;
7187 sd = &per_cpu(phys_domains, j).sd;
7188 if (j != cpumask_first(sched_group_cpus(sd->groups))) {
7190 * Only add "power" once for each
7196 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7199 } while (sg != group_head);
7201 #endif /* CONFIG_NUMA */
7204 /* Free memory allocated for various sched_group structures */
7205 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7209 for_each_cpu(cpu, cpu_map) {
7210 struct sched_group **sched_group_nodes
7211 = sched_group_nodes_bycpu[cpu];
7213 if (!sched_group_nodes)
7216 for (i = 0; i < nr_node_ids; i++) {
7217 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7218 node_to_cpumask_ptr(pnodemask, i);
7220 cpus_and(*nodemask, *pnodemask, *cpu_map);
7221 if (cpus_empty(*nodemask))
7231 if (oldsg != sched_group_nodes[i])
7234 kfree(sched_group_nodes);
7235 sched_group_nodes_bycpu[cpu] = NULL;
7238 #else /* !CONFIG_NUMA */
7239 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7242 #endif /* CONFIG_NUMA */
7245 * Initialize sched groups cpu_power.
7247 * cpu_power indicates the capacity of sched group, which is used while
7248 * distributing the load between different sched groups in a sched domain.
7249 * Typically cpu_power for all the groups in a sched domain will be same unless
7250 * there are asymmetries in the topology. If there are asymmetries, group
7251 * having more cpu_power will pickup more load compared to the group having
7254 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7255 * the maximum number of tasks a group can handle in the presence of other idle
7256 * or lightly loaded groups in the same sched domain.
7258 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7260 struct sched_domain *child;
7261 struct sched_group *group;
7263 WARN_ON(!sd || !sd->groups);
7265 if (cpu != cpumask_first(sched_group_cpus(sd->groups)))
7270 sd->groups->__cpu_power = 0;
7273 * For perf policy, if the groups in child domain share resources
7274 * (for example cores sharing some portions of the cache hierarchy
7275 * or SMT), then set this domain groups cpu_power such that each group
7276 * can handle only one task, when there are other idle groups in the
7277 * same sched domain.
7279 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7281 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7282 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7287 * add cpu_power of each child group to this groups cpu_power
7289 group = child->groups;
7291 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7292 group = group->next;
7293 } while (group != child->groups);
7297 * Initializers for schedule domains
7298 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7301 #ifdef CONFIG_SCHED_DEBUG
7302 # define SD_INIT_NAME(sd, type) sd->name = #type
7304 # define SD_INIT_NAME(sd, type) do { } while (0)
7307 #define SD_INIT(sd, type) sd_init_##type(sd)
7309 #define SD_INIT_FUNC(type) \
7310 static noinline void sd_init_##type(struct sched_domain *sd) \
7312 memset(sd, 0, sizeof(*sd)); \
7313 *sd = SD_##type##_INIT; \
7314 sd->level = SD_LV_##type; \
7315 SD_INIT_NAME(sd, type); \
7320 SD_INIT_FUNC(ALLNODES)
7323 #ifdef CONFIG_SCHED_SMT
7324 SD_INIT_FUNC(SIBLING)
7326 #ifdef CONFIG_SCHED_MC
7330 static int default_relax_domain_level = -1;
7332 static int __init setup_relax_domain_level(char *str)
7336 val = simple_strtoul(str, NULL, 0);
7337 if (val < SD_LV_MAX)
7338 default_relax_domain_level = val;
7342 __setup("relax_domain_level=", setup_relax_domain_level);
7344 static void set_domain_attribute(struct sched_domain *sd,
7345 struct sched_domain_attr *attr)
7349 if (!attr || attr->relax_domain_level < 0) {
7350 if (default_relax_domain_level < 0)
7353 request = default_relax_domain_level;
7355 request = attr->relax_domain_level;
7356 if (request < sd->level) {
7357 /* turn off idle balance on this domain */
7358 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7360 /* turn on idle balance on this domain */
7361 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7366 * Build sched domains for a given set of cpus and attach the sched domains
7367 * to the individual cpus
7369 static int __build_sched_domains(const cpumask_t *cpu_map,
7370 struct sched_domain_attr *attr)
7372 int i, err = -ENOMEM;
7373 struct root_domain *rd;
7374 cpumask_var_t nodemask, this_sibling_map, this_core_map, send_covered,
7377 cpumask_var_t domainspan, covered, notcovered;
7378 struct sched_group **sched_group_nodes = NULL;
7379 int sd_allnodes = 0;
7381 if (!alloc_cpumask_var(&domainspan, GFP_KERNEL))
7383 if (!alloc_cpumask_var(&covered, GFP_KERNEL))
7384 goto free_domainspan;
7385 if (!alloc_cpumask_var(¬covered, GFP_KERNEL))
7389 if (!alloc_cpumask_var(&nodemask, GFP_KERNEL))
7390 goto free_notcovered;
7391 if (!alloc_cpumask_var(&this_sibling_map, GFP_KERNEL))
7393 if (!alloc_cpumask_var(&this_core_map, GFP_KERNEL))
7394 goto free_this_sibling_map;
7395 if (!alloc_cpumask_var(&send_covered, GFP_KERNEL))
7396 goto free_this_core_map;
7397 if (!alloc_cpumask_var(&tmpmask, GFP_KERNEL))
7398 goto free_send_covered;
7402 * Allocate the per-node list of sched groups
7404 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
7406 if (!sched_group_nodes) {
7407 printk(KERN_WARNING "Can not alloc sched group node list\n");
7412 rd = alloc_rootdomain();
7414 printk(KERN_WARNING "Cannot alloc root domain\n");
7415 goto free_sched_groups;
7419 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7423 * Set up domains for cpus specified by the cpu_map.
7425 for_each_cpu(i, cpu_map) {
7426 struct sched_domain *sd = NULL, *p;
7428 *nodemask = node_to_cpumask(cpu_to_node(i));
7429 cpus_and(*nodemask, *nodemask, *cpu_map);
7432 if (cpus_weight(*cpu_map) >
7433 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7434 sd = &per_cpu(allnodes_domains, i);
7435 SD_INIT(sd, ALLNODES);
7436 set_domain_attribute(sd, attr);
7437 cpumask_copy(sched_domain_span(sd), cpu_map);
7438 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7444 sd = &per_cpu(node_domains, i);
7446 set_domain_attribute(sd, attr);
7447 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
7451 cpumask_and(sched_domain_span(sd),
7452 sched_domain_span(sd), cpu_map);
7456 sd = &per_cpu(phys_domains, i).sd;
7458 set_domain_attribute(sd, attr);
7459 cpumask_copy(sched_domain_span(sd), nodemask);
7463 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7465 #ifdef CONFIG_SCHED_MC
7467 sd = &per_cpu(core_domains, i).sd;
7469 set_domain_attribute(sd, attr);
7470 *sched_domain_span(sd) = cpu_coregroup_map(i);
7471 cpumask_and(sched_domain_span(sd),
7472 sched_domain_span(sd), cpu_map);
7475 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7478 #ifdef CONFIG_SCHED_SMT
7480 sd = &per_cpu(cpu_domains, i).sd;
7481 SD_INIT(sd, SIBLING);
7482 set_domain_attribute(sd, attr);
7483 cpumask_and(sched_domain_span(sd),
7484 &per_cpu(cpu_sibling_map, i), cpu_map);
7487 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7491 #ifdef CONFIG_SCHED_SMT
7492 /* Set up CPU (sibling) groups */
7493 for_each_cpu(i, cpu_map) {
7494 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7495 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7496 if (i != first_cpu(*this_sibling_map))
7499 init_sched_build_groups(this_sibling_map, cpu_map,
7501 send_covered, tmpmask);
7505 #ifdef CONFIG_SCHED_MC
7506 /* Set up multi-core groups */
7507 for_each_cpu(i, cpu_map) {
7508 *this_core_map = cpu_coregroup_map(i);
7509 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7510 if (i != first_cpu(*this_core_map))
7513 init_sched_build_groups(this_core_map, cpu_map,
7515 send_covered, tmpmask);
7519 /* Set up physical groups */
7520 for (i = 0; i < nr_node_ids; i++) {
7521 *nodemask = node_to_cpumask(i);
7522 cpus_and(*nodemask, *nodemask, *cpu_map);
7523 if (cpus_empty(*nodemask))
7526 init_sched_build_groups(nodemask, cpu_map,
7528 send_covered, tmpmask);
7532 /* Set up node groups */
7534 init_sched_build_groups(cpu_map, cpu_map,
7535 &cpu_to_allnodes_group,
7536 send_covered, tmpmask);
7539 for (i = 0; i < nr_node_ids; i++) {
7540 /* Set up node groups */
7541 struct sched_group *sg, *prev;
7544 *nodemask = node_to_cpumask(i);
7545 cpus_clear(*covered);
7547 cpus_and(*nodemask, *nodemask, *cpu_map);
7548 if (cpus_empty(*nodemask)) {
7549 sched_group_nodes[i] = NULL;
7553 sched_domain_node_span(i, domainspan);
7554 cpus_and(*domainspan, *domainspan, *cpu_map);
7556 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
7559 printk(KERN_WARNING "Can not alloc domain group for "
7563 sched_group_nodes[i] = sg;
7564 for_each_cpu(j, nodemask) {
7565 struct sched_domain *sd;
7567 sd = &per_cpu(node_domains, j);
7570 sg->__cpu_power = 0;
7571 cpumask_copy(sched_group_cpus(sg), nodemask);
7573 cpus_or(*covered, *covered, *nodemask);
7576 for (j = 0; j < nr_node_ids; j++) {
7577 int n = (i + j) % nr_node_ids;
7578 node_to_cpumask_ptr(pnodemask, n);
7580 cpus_complement(*notcovered, *covered);
7581 cpus_and(*tmpmask, *notcovered, *cpu_map);
7582 cpus_and(*tmpmask, *tmpmask, *domainspan);
7583 if (cpus_empty(*tmpmask))
7586 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7587 if (cpus_empty(*tmpmask))
7590 sg = kmalloc_node(sizeof(struct sched_group) +
7595 "Can not alloc domain group for node %d\n", j);
7598 sg->__cpu_power = 0;
7599 cpumask_copy(sched_group_cpus(sg), tmpmask);
7600 sg->next = prev->next;
7601 cpus_or(*covered, *covered, *tmpmask);
7608 /* Calculate CPU power for physical packages and nodes */
7609 #ifdef CONFIG_SCHED_SMT
7610 for_each_cpu(i, cpu_map) {
7611 struct sched_domain *sd = &per_cpu(cpu_domains, i).sd;
7613 init_sched_groups_power(i, sd);
7616 #ifdef CONFIG_SCHED_MC
7617 for_each_cpu(i, cpu_map) {
7618 struct sched_domain *sd = &per_cpu(core_domains, i).sd;
7620 init_sched_groups_power(i, sd);
7624 for_each_cpu(i, cpu_map) {
7625 struct sched_domain *sd = &per_cpu(phys_domains, i).sd;
7627 init_sched_groups_power(i, sd);
7631 for (i = 0; i < nr_node_ids; i++)
7632 init_numa_sched_groups_power(sched_group_nodes[i]);
7635 struct sched_group *sg;
7637 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7639 init_numa_sched_groups_power(sg);
7643 /* Attach the domains */
7644 for_each_cpu(i, cpu_map) {
7645 struct sched_domain *sd;
7646 #ifdef CONFIG_SCHED_SMT
7647 sd = &per_cpu(cpu_domains, i).sd;
7648 #elif defined(CONFIG_SCHED_MC)
7649 sd = &per_cpu(core_domains, i).sd;
7651 sd = &per_cpu(phys_domains, i).sd;
7653 cpu_attach_domain(sd, rd, i);
7659 free_cpumask_var(tmpmask);
7661 free_cpumask_var(send_covered);
7663 free_cpumask_var(this_core_map);
7664 free_this_sibling_map:
7665 free_cpumask_var(this_sibling_map);
7667 free_cpumask_var(nodemask);
7670 free_cpumask_var(notcovered);
7672 free_cpumask_var(covered);
7674 free_cpumask_var(domainspan);
7681 kfree(sched_group_nodes);
7687 free_sched_groups(cpu_map, tmpmask);
7688 free_rootdomain(rd);
7693 static int build_sched_domains(const cpumask_t *cpu_map)
7695 return __build_sched_domains(cpu_map, NULL);
7698 static cpumask_t *doms_cur; /* current sched domains */
7699 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7700 static struct sched_domain_attr *dattr_cur;
7701 /* attribues of custom domains in 'doms_cur' */
7704 * Special case: If a kmalloc of a doms_cur partition (array of
7705 * cpumask) fails, then fallback to a single sched domain,
7706 * as determined by the single cpumask fallback_doms.
7708 static cpumask_var_t fallback_doms;
7710 void __attribute__((weak)) arch_update_cpu_topology(void)
7715 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7716 * For now this just excludes isolated cpus, but could be used to
7717 * exclude other special cases in the future.
7719 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7723 arch_update_cpu_topology();
7725 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7727 doms_cur = fallback_doms;
7728 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
7730 err = build_sched_domains(doms_cur);
7731 register_sched_domain_sysctl();
7736 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7739 free_sched_groups(cpu_map, tmpmask);
7743 * Detach sched domains from a group of cpus specified in cpu_map
7744 * These cpus will now be attached to the NULL domain
7746 static void detach_destroy_domains(const cpumask_t *cpu_map)
7751 for_each_cpu(i, cpu_map)
7752 cpu_attach_domain(NULL, &def_root_domain, i);
7753 synchronize_sched();
7754 arch_destroy_sched_domains(cpu_map, &tmpmask);
7757 /* handle null as "default" */
7758 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7759 struct sched_domain_attr *new, int idx_new)
7761 struct sched_domain_attr tmp;
7768 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7769 new ? (new + idx_new) : &tmp,
7770 sizeof(struct sched_domain_attr));
7774 * Partition sched domains as specified by the 'ndoms_new'
7775 * cpumasks in the array doms_new[] of cpumasks. This compares
7776 * doms_new[] to the current sched domain partitioning, doms_cur[].
7777 * It destroys each deleted domain and builds each new domain.
7779 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7780 * The masks don't intersect (don't overlap.) We should setup one
7781 * sched domain for each mask. CPUs not in any of the cpumasks will
7782 * not be load balanced. If the same cpumask appears both in the
7783 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7786 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7787 * ownership of it and will kfree it when done with it. If the caller
7788 * failed the kmalloc call, then it can pass in doms_new == NULL &&
7789 * ndoms_new == 1, and partition_sched_domains() will fallback to
7790 * the single partition 'fallback_doms', it also forces the domains
7793 * If doms_new == NULL it will be replaced with cpu_online_map.
7794 * ndoms_new == 0 is a special case for destroying existing domains,
7795 * and it will not create the default domain.
7797 * Call with hotplug lock held
7799 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7800 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 n = doms_new ? ndoms_new : 0;
7811 /* Destroy deleted domains */
7812 for (i = 0; i < ndoms_cur; i++) {
7813 for (j = 0; j < n; j++) {
7814 if (cpus_equal(doms_cur[i], doms_new[j])
7815 && dattrs_equal(dattr_cur, i, dattr_new, j))
7818 /* no match - a current sched domain not in new doms_new[] */
7819 detach_destroy_domains(doms_cur + i);
7824 if (doms_new == NULL) {
7826 doms_new = fallback_doms;
7827 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
7828 WARN_ON_ONCE(dattr_new);
7831 /* Build new domains */
7832 for (i = 0; i < ndoms_new; i++) {
7833 for (j = 0; j < ndoms_cur; j++) {
7834 if (cpus_equal(doms_new[i], doms_cur[j])
7835 && dattrs_equal(dattr_new, i, dattr_cur, j))
7838 /* no match - add a new doms_new */
7839 __build_sched_domains(doms_new + i,
7840 dattr_new ? dattr_new + i : NULL);
7845 /* Remember the new sched domains */
7846 if (doms_cur != fallback_doms)
7848 kfree(dattr_cur); /* kfree(NULL) is safe */
7849 doms_cur = doms_new;
7850 dattr_cur = dattr_new;
7851 ndoms_cur = ndoms_new;
7853 register_sched_domain_sysctl();
7855 mutex_unlock(&sched_domains_mutex);
7858 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7859 int arch_reinit_sched_domains(void)
7863 /* Destroy domains first to force the rebuild */
7864 partition_sched_domains(0, NULL, NULL);
7866 rebuild_sched_domains();
7872 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7876 if (buf[0] != '0' && buf[0] != '1')
7880 sched_smt_power_savings = (buf[0] == '1');
7882 sched_mc_power_savings = (buf[0] == '1');
7884 ret = arch_reinit_sched_domains();
7886 return ret ? ret : count;
7889 #ifdef CONFIG_SCHED_MC
7890 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7893 return sprintf(page, "%u\n", sched_mc_power_savings);
7895 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7896 const char *buf, size_t count)
7898 return sched_power_savings_store(buf, count, 0);
7900 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7901 sched_mc_power_savings_show,
7902 sched_mc_power_savings_store);
7905 #ifdef CONFIG_SCHED_SMT
7906 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7909 return sprintf(page, "%u\n", sched_smt_power_savings);
7911 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7912 const char *buf, size_t count)
7914 return sched_power_savings_store(buf, count, 1);
7916 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7917 sched_smt_power_savings_show,
7918 sched_smt_power_savings_store);
7921 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7925 #ifdef CONFIG_SCHED_SMT
7927 err = sysfs_create_file(&cls->kset.kobj,
7928 &attr_sched_smt_power_savings.attr);
7930 #ifdef CONFIG_SCHED_MC
7931 if (!err && mc_capable())
7932 err = sysfs_create_file(&cls->kset.kobj,
7933 &attr_sched_mc_power_savings.attr);
7937 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7939 #ifndef CONFIG_CPUSETS
7941 * Add online and remove offline CPUs from the scheduler domains.
7942 * When cpusets are enabled they take over this function.
7944 static int update_sched_domains(struct notifier_block *nfb,
7945 unsigned long action, void *hcpu)
7949 case CPU_ONLINE_FROZEN:
7951 case CPU_DEAD_FROZEN:
7952 partition_sched_domains(1, NULL, NULL);
7961 static int update_runtime(struct notifier_block *nfb,
7962 unsigned long action, void *hcpu)
7964 int cpu = (int)(long)hcpu;
7967 case CPU_DOWN_PREPARE:
7968 case CPU_DOWN_PREPARE_FROZEN:
7969 disable_runtime(cpu_rq(cpu));
7972 case CPU_DOWN_FAILED:
7973 case CPU_DOWN_FAILED_FROZEN:
7975 case CPU_ONLINE_FROZEN:
7976 enable_runtime(cpu_rq(cpu));
7984 void __init sched_init_smp(void)
7986 cpumask_var_t non_isolated_cpus;
7988 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7990 #if defined(CONFIG_NUMA)
7991 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7993 BUG_ON(sched_group_nodes_bycpu == NULL);
7996 mutex_lock(&sched_domains_mutex);
7997 arch_init_sched_domains(cpu_online_mask);
7998 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7999 if (cpumask_empty(non_isolated_cpus))
8000 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
8001 mutex_unlock(&sched_domains_mutex);
8004 #ifndef CONFIG_CPUSETS
8005 /* XXX: Theoretical race here - CPU may be hotplugged now */
8006 hotcpu_notifier(update_sched_domains, 0);
8009 /* RT runtime code needs to handle some hotplug events */
8010 hotcpu_notifier(update_runtime, 0);
8014 /* Move init over to a non-isolated CPU */
8015 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
8017 sched_init_granularity();
8018 free_cpumask_var(non_isolated_cpus);
8020 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
8023 void __init sched_init_smp(void)
8025 sched_init_granularity();
8027 #endif /* CONFIG_SMP */
8029 int in_sched_functions(unsigned long addr)
8031 return in_lock_functions(addr) ||
8032 (addr >= (unsigned long)__sched_text_start
8033 && addr < (unsigned long)__sched_text_end);
8036 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
8038 cfs_rq->tasks_timeline = RB_ROOT;
8039 INIT_LIST_HEAD(&cfs_rq->tasks);
8040 #ifdef CONFIG_FAIR_GROUP_SCHED
8043 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8046 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8048 struct rt_prio_array *array;
8051 array = &rt_rq->active;
8052 for (i = 0; i < MAX_RT_PRIO; i++) {
8053 INIT_LIST_HEAD(array->queue + i);
8054 __clear_bit(i, array->bitmap);
8056 /* delimiter for bitsearch: */
8057 __set_bit(MAX_RT_PRIO, array->bitmap);
8059 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8060 rt_rq->highest_prio = MAX_RT_PRIO;
8063 rt_rq->rt_nr_migratory = 0;
8064 rt_rq->overloaded = 0;
8068 rt_rq->rt_throttled = 0;
8069 rt_rq->rt_runtime = 0;
8070 spin_lock_init(&rt_rq->rt_runtime_lock);
8072 #ifdef CONFIG_RT_GROUP_SCHED
8073 rt_rq->rt_nr_boosted = 0;
8078 #ifdef CONFIG_FAIR_GROUP_SCHED
8079 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8080 struct sched_entity *se, int cpu, int add,
8081 struct sched_entity *parent)
8083 struct rq *rq = cpu_rq(cpu);
8084 tg->cfs_rq[cpu] = cfs_rq;
8085 init_cfs_rq(cfs_rq, rq);
8088 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8091 /* se could be NULL for init_task_group */
8096 se->cfs_rq = &rq->cfs;
8098 se->cfs_rq = parent->my_q;
8101 se->load.weight = tg->shares;
8102 se->load.inv_weight = 0;
8103 se->parent = parent;
8107 #ifdef CONFIG_RT_GROUP_SCHED
8108 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8109 struct sched_rt_entity *rt_se, int cpu, int add,
8110 struct sched_rt_entity *parent)
8112 struct rq *rq = cpu_rq(cpu);
8114 tg->rt_rq[cpu] = rt_rq;
8115 init_rt_rq(rt_rq, rq);
8117 rt_rq->rt_se = rt_se;
8118 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8120 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8122 tg->rt_se[cpu] = rt_se;
8127 rt_se->rt_rq = &rq->rt;
8129 rt_se->rt_rq = parent->my_q;
8131 rt_se->my_q = rt_rq;
8132 rt_se->parent = parent;
8133 INIT_LIST_HEAD(&rt_se->run_list);
8137 void __init sched_init(void)
8140 unsigned long alloc_size = 0, ptr;
8142 #ifdef CONFIG_FAIR_GROUP_SCHED
8143 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8145 #ifdef CONFIG_RT_GROUP_SCHED
8146 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8148 #ifdef CONFIG_USER_SCHED
8152 * As sched_init() is called before page_alloc is setup,
8153 * we use alloc_bootmem().
8156 ptr = (unsigned long)alloc_bootmem(alloc_size);
8158 #ifdef CONFIG_FAIR_GROUP_SCHED
8159 init_task_group.se = (struct sched_entity **)ptr;
8160 ptr += nr_cpu_ids * sizeof(void **);
8162 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8163 ptr += nr_cpu_ids * sizeof(void **);
8165 #ifdef CONFIG_USER_SCHED
8166 root_task_group.se = (struct sched_entity **)ptr;
8167 ptr += nr_cpu_ids * sizeof(void **);
8169 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8170 ptr += nr_cpu_ids * sizeof(void **);
8171 #endif /* CONFIG_USER_SCHED */
8172 #endif /* CONFIG_FAIR_GROUP_SCHED */
8173 #ifdef CONFIG_RT_GROUP_SCHED
8174 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8175 ptr += nr_cpu_ids * sizeof(void **);
8177 init_task_group.rt_rq = (struct rt_rq **)ptr;
8178 ptr += nr_cpu_ids * sizeof(void **);
8180 #ifdef CONFIG_USER_SCHED
8181 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8182 ptr += nr_cpu_ids * sizeof(void **);
8184 root_task_group.rt_rq = (struct rt_rq **)ptr;
8185 ptr += nr_cpu_ids * sizeof(void **);
8186 #endif /* CONFIG_USER_SCHED */
8187 #endif /* CONFIG_RT_GROUP_SCHED */
8191 init_defrootdomain();
8194 init_rt_bandwidth(&def_rt_bandwidth,
8195 global_rt_period(), global_rt_runtime());
8197 #ifdef CONFIG_RT_GROUP_SCHED
8198 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8199 global_rt_period(), global_rt_runtime());
8200 #ifdef CONFIG_USER_SCHED
8201 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8202 global_rt_period(), RUNTIME_INF);
8203 #endif /* CONFIG_USER_SCHED */
8204 #endif /* CONFIG_RT_GROUP_SCHED */
8206 #ifdef CONFIG_GROUP_SCHED
8207 list_add(&init_task_group.list, &task_groups);
8208 INIT_LIST_HEAD(&init_task_group.children);
8210 #ifdef CONFIG_USER_SCHED
8211 INIT_LIST_HEAD(&root_task_group.children);
8212 init_task_group.parent = &root_task_group;
8213 list_add(&init_task_group.siblings, &root_task_group.children);
8214 #endif /* CONFIG_USER_SCHED */
8215 #endif /* CONFIG_GROUP_SCHED */
8217 for_each_possible_cpu(i) {
8221 spin_lock_init(&rq->lock);
8223 init_cfs_rq(&rq->cfs, rq);
8224 init_rt_rq(&rq->rt, rq);
8225 #ifdef CONFIG_FAIR_GROUP_SCHED
8226 init_task_group.shares = init_task_group_load;
8227 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8228 #ifdef CONFIG_CGROUP_SCHED
8230 * How much cpu bandwidth does init_task_group get?
8232 * In case of task-groups formed thr' the cgroup filesystem, it
8233 * gets 100% of the cpu resources in the system. This overall
8234 * system cpu resource is divided among the tasks of
8235 * init_task_group and its child task-groups in a fair manner,
8236 * based on each entity's (task or task-group's) weight
8237 * (se->load.weight).
8239 * In other words, if init_task_group has 10 tasks of weight
8240 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8241 * then A0's share of the cpu resource is:
8243 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8245 * We achieve this by letting init_task_group's tasks sit
8246 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8248 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8249 #elif defined CONFIG_USER_SCHED
8250 root_task_group.shares = NICE_0_LOAD;
8251 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8253 * In case of task-groups formed thr' the user id of tasks,
8254 * init_task_group represents tasks belonging to root user.
8255 * Hence it forms a sibling of all subsequent groups formed.
8256 * In this case, init_task_group gets only a fraction of overall
8257 * system cpu resource, based on the weight assigned to root
8258 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8259 * by letting tasks of init_task_group sit in a separate cfs_rq
8260 * (init_cfs_rq) and having one entity represent this group of
8261 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8263 init_tg_cfs_entry(&init_task_group,
8264 &per_cpu(init_cfs_rq, i),
8265 &per_cpu(init_sched_entity, i), i, 1,
8266 root_task_group.se[i]);
8269 #endif /* CONFIG_FAIR_GROUP_SCHED */
8271 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8272 #ifdef CONFIG_RT_GROUP_SCHED
8273 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8274 #ifdef CONFIG_CGROUP_SCHED
8275 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8276 #elif defined CONFIG_USER_SCHED
8277 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8278 init_tg_rt_entry(&init_task_group,
8279 &per_cpu(init_rt_rq, i),
8280 &per_cpu(init_sched_rt_entity, i), i, 1,
8281 root_task_group.rt_se[i]);
8285 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8286 rq->cpu_load[j] = 0;
8290 rq->active_balance = 0;
8291 rq->next_balance = jiffies;
8295 rq->migration_thread = NULL;
8296 INIT_LIST_HEAD(&rq->migration_queue);
8297 rq_attach_root(rq, &def_root_domain);
8300 atomic_set(&rq->nr_iowait, 0);
8303 set_load_weight(&init_task);
8305 #ifdef CONFIG_PREEMPT_NOTIFIERS
8306 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8310 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8313 #ifdef CONFIG_RT_MUTEXES
8314 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8318 * The boot idle thread does lazy MMU switching as well:
8320 atomic_inc(&init_mm.mm_count);
8321 enter_lazy_tlb(&init_mm, current);
8324 * Make us the idle thread. Technically, schedule() should not be
8325 * called from this thread, however somewhere below it might be,
8326 * but because we are the idle thread, we just pick up running again
8327 * when this runqueue becomes "idle".
8329 init_idle(current, smp_processor_id());
8331 * During early bootup we pretend to be a normal task:
8333 current->sched_class = &fair_sched_class;
8335 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8336 alloc_bootmem_cpumask_var(&nohz_cpu_mask);
8338 alloc_bootmem_cpumask_var(&nohz.cpu_mask);
8340 alloc_bootmem_cpumask_var(&cpu_isolated_map);
8342 scheduler_running = 1;
8345 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8346 void __might_sleep(char *file, int line)
8349 static unsigned long prev_jiffy; /* ratelimiting */
8351 if ((!in_atomic() && !irqs_disabled()) ||
8352 system_state != SYSTEM_RUNNING || oops_in_progress)
8354 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8356 prev_jiffy = jiffies;
8359 "BUG: sleeping function called from invalid context at %s:%d\n",
8362 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8363 in_atomic(), irqs_disabled(),
8364 current->pid, current->comm);
8366 debug_show_held_locks(current);
8367 if (irqs_disabled())
8368 print_irqtrace_events(current);
8372 EXPORT_SYMBOL(__might_sleep);
8375 #ifdef CONFIG_MAGIC_SYSRQ
8376 static void normalize_task(struct rq *rq, struct task_struct *p)
8380 update_rq_clock(rq);
8381 on_rq = p->se.on_rq;
8383 deactivate_task(rq, p, 0);
8384 __setscheduler(rq, p, SCHED_NORMAL, 0);
8386 activate_task(rq, p, 0);
8387 resched_task(rq->curr);
8391 void normalize_rt_tasks(void)
8393 struct task_struct *g, *p;
8394 unsigned long flags;
8397 read_lock_irqsave(&tasklist_lock, flags);
8398 do_each_thread(g, p) {
8400 * Only normalize user tasks:
8405 p->se.exec_start = 0;
8406 #ifdef CONFIG_SCHEDSTATS
8407 p->se.wait_start = 0;
8408 p->se.sleep_start = 0;
8409 p->se.block_start = 0;
8414 * Renice negative nice level userspace
8417 if (TASK_NICE(p) < 0 && p->mm)
8418 set_user_nice(p, 0);
8422 spin_lock(&p->pi_lock);
8423 rq = __task_rq_lock(p);
8425 normalize_task(rq, p);
8427 __task_rq_unlock(rq);
8428 spin_unlock(&p->pi_lock);
8429 } while_each_thread(g, p);
8431 read_unlock_irqrestore(&tasklist_lock, flags);
8434 #endif /* CONFIG_MAGIC_SYSRQ */
8438 * These functions are only useful for the IA64 MCA handling.
8440 * They can only be called when the whole system has been
8441 * stopped - every CPU needs to be quiescent, and no scheduling
8442 * activity can take place. Using them for anything else would
8443 * be a serious bug, and as a result, they aren't even visible
8444 * under any other configuration.
8448 * curr_task - return the current task for a given cpu.
8449 * @cpu: the processor in question.
8451 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8453 struct task_struct *curr_task(int cpu)
8455 return cpu_curr(cpu);
8459 * set_curr_task - set the current task for a given cpu.
8460 * @cpu: the processor in question.
8461 * @p: the task pointer to set.
8463 * Description: This function must only be used when non-maskable interrupts
8464 * are serviced on a separate stack. It allows the architecture to switch the
8465 * notion of the current task on a cpu in a non-blocking manner. This function
8466 * must be called with all CPU's synchronized, and interrupts disabled, the
8467 * and caller must save the original value of the current task (see
8468 * curr_task() above) and restore that value before reenabling interrupts and
8469 * re-starting the system.
8471 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8473 void set_curr_task(int cpu, struct task_struct *p)
8480 #ifdef CONFIG_FAIR_GROUP_SCHED
8481 static void free_fair_sched_group(struct task_group *tg)
8485 for_each_possible_cpu(i) {
8487 kfree(tg->cfs_rq[i]);
8497 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8499 struct cfs_rq *cfs_rq;
8500 struct sched_entity *se;
8504 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8507 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8511 tg->shares = NICE_0_LOAD;
8513 for_each_possible_cpu(i) {
8516 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8517 GFP_KERNEL, cpu_to_node(i));
8521 se = kzalloc_node(sizeof(struct sched_entity),
8522 GFP_KERNEL, cpu_to_node(i));
8526 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8535 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8537 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8538 &cpu_rq(cpu)->leaf_cfs_rq_list);
8541 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8543 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8545 #else /* !CONFG_FAIR_GROUP_SCHED */
8546 static inline void free_fair_sched_group(struct task_group *tg)
8551 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8556 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8560 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8563 #endif /* CONFIG_FAIR_GROUP_SCHED */
8565 #ifdef CONFIG_RT_GROUP_SCHED
8566 static void free_rt_sched_group(struct task_group *tg)
8570 destroy_rt_bandwidth(&tg->rt_bandwidth);
8572 for_each_possible_cpu(i) {
8574 kfree(tg->rt_rq[i]);
8576 kfree(tg->rt_se[i]);
8584 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8586 struct rt_rq *rt_rq;
8587 struct sched_rt_entity *rt_se;
8591 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8594 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8598 init_rt_bandwidth(&tg->rt_bandwidth,
8599 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8601 for_each_possible_cpu(i) {
8604 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8605 GFP_KERNEL, cpu_to_node(i));
8609 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8610 GFP_KERNEL, cpu_to_node(i));
8614 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8623 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8625 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8626 &cpu_rq(cpu)->leaf_rt_rq_list);
8629 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8631 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8633 #else /* !CONFIG_RT_GROUP_SCHED */
8634 static inline void free_rt_sched_group(struct task_group *tg)
8639 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8644 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8648 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8651 #endif /* CONFIG_RT_GROUP_SCHED */
8653 #ifdef CONFIG_GROUP_SCHED
8654 static void free_sched_group(struct task_group *tg)
8656 free_fair_sched_group(tg);
8657 free_rt_sched_group(tg);
8661 /* allocate runqueue etc for a new task group */
8662 struct task_group *sched_create_group(struct task_group *parent)
8664 struct task_group *tg;
8665 unsigned long flags;
8668 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8670 return ERR_PTR(-ENOMEM);
8672 if (!alloc_fair_sched_group(tg, parent))
8675 if (!alloc_rt_sched_group(tg, parent))
8678 spin_lock_irqsave(&task_group_lock, flags);
8679 for_each_possible_cpu(i) {
8680 register_fair_sched_group(tg, i);
8681 register_rt_sched_group(tg, i);
8683 list_add_rcu(&tg->list, &task_groups);
8685 WARN_ON(!parent); /* root should already exist */
8687 tg->parent = parent;
8688 INIT_LIST_HEAD(&tg->children);
8689 list_add_rcu(&tg->siblings, &parent->children);
8690 spin_unlock_irqrestore(&task_group_lock, flags);
8695 free_sched_group(tg);
8696 return ERR_PTR(-ENOMEM);
8699 /* rcu callback to free various structures associated with a task group */
8700 static void free_sched_group_rcu(struct rcu_head *rhp)
8702 /* now it should be safe to free those cfs_rqs */
8703 free_sched_group(container_of(rhp, struct task_group, rcu));
8706 /* Destroy runqueue etc associated with a task group */
8707 void sched_destroy_group(struct task_group *tg)
8709 unsigned long flags;
8712 spin_lock_irqsave(&task_group_lock, flags);
8713 for_each_possible_cpu(i) {
8714 unregister_fair_sched_group(tg, i);
8715 unregister_rt_sched_group(tg, i);
8717 list_del_rcu(&tg->list);
8718 list_del_rcu(&tg->siblings);
8719 spin_unlock_irqrestore(&task_group_lock, flags);
8721 /* wait for possible concurrent references to cfs_rqs complete */
8722 call_rcu(&tg->rcu, free_sched_group_rcu);
8725 /* change task's runqueue when it moves between groups.
8726 * The caller of this function should have put the task in its new group
8727 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8728 * reflect its new group.
8730 void sched_move_task(struct task_struct *tsk)
8733 unsigned long flags;
8736 rq = task_rq_lock(tsk, &flags);
8738 update_rq_clock(rq);
8740 running = task_current(rq, tsk);
8741 on_rq = tsk->se.on_rq;
8744 dequeue_task(rq, tsk, 0);
8745 if (unlikely(running))
8746 tsk->sched_class->put_prev_task(rq, tsk);
8748 set_task_rq(tsk, task_cpu(tsk));
8750 #ifdef CONFIG_FAIR_GROUP_SCHED
8751 if (tsk->sched_class->moved_group)
8752 tsk->sched_class->moved_group(tsk);
8755 if (unlikely(running))
8756 tsk->sched_class->set_curr_task(rq);
8758 enqueue_task(rq, tsk, 0);
8760 task_rq_unlock(rq, &flags);
8762 #endif /* CONFIG_GROUP_SCHED */
8764 #ifdef CONFIG_FAIR_GROUP_SCHED
8765 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8767 struct cfs_rq *cfs_rq = se->cfs_rq;
8772 dequeue_entity(cfs_rq, se, 0);
8774 se->load.weight = shares;
8775 se->load.inv_weight = 0;
8778 enqueue_entity(cfs_rq, se, 0);
8781 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8783 struct cfs_rq *cfs_rq = se->cfs_rq;
8784 struct rq *rq = cfs_rq->rq;
8785 unsigned long flags;
8787 spin_lock_irqsave(&rq->lock, flags);
8788 __set_se_shares(se, shares);
8789 spin_unlock_irqrestore(&rq->lock, flags);
8792 static DEFINE_MUTEX(shares_mutex);
8794 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8797 unsigned long flags;
8800 * We can't change the weight of the root cgroup.
8805 if (shares < MIN_SHARES)
8806 shares = MIN_SHARES;
8807 else if (shares > MAX_SHARES)
8808 shares = MAX_SHARES;
8810 mutex_lock(&shares_mutex);
8811 if (tg->shares == shares)
8814 spin_lock_irqsave(&task_group_lock, flags);
8815 for_each_possible_cpu(i)
8816 unregister_fair_sched_group(tg, i);
8817 list_del_rcu(&tg->siblings);
8818 spin_unlock_irqrestore(&task_group_lock, flags);
8820 /* wait for any ongoing reference to this group to finish */
8821 synchronize_sched();
8824 * Now we are free to modify the group's share on each cpu
8825 * w/o tripping rebalance_share or load_balance_fair.
8827 tg->shares = shares;
8828 for_each_possible_cpu(i) {
8832 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8833 set_se_shares(tg->se[i], shares);
8837 * Enable load balance activity on this group, by inserting it back on
8838 * each cpu's rq->leaf_cfs_rq_list.
8840 spin_lock_irqsave(&task_group_lock, flags);
8841 for_each_possible_cpu(i)
8842 register_fair_sched_group(tg, i);
8843 list_add_rcu(&tg->siblings, &tg->parent->children);
8844 spin_unlock_irqrestore(&task_group_lock, flags);
8846 mutex_unlock(&shares_mutex);
8850 unsigned long sched_group_shares(struct task_group *tg)
8856 #ifdef CONFIG_RT_GROUP_SCHED
8858 * Ensure that the real time constraints are schedulable.
8860 static DEFINE_MUTEX(rt_constraints_mutex);
8862 static unsigned long to_ratio(u64 period, u64 runtime)
8864 if (runtime == RUNTIME_INF)
8867 return div64_u64(runtime << 20, period);
8870 /* Must be called with tasklist_lock held */
8871 static inline int tg_has_rt_tasks(struct task_group *tg)
8873 struct task_struct *g, *p;
8875 do_each_thread(g, p) {
8876 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8878 } while_each_thread(g, p);
8883 struct rt_schedulable_data {
8884 struct task_group *tg;
8889 static int tg_schedulable(struct task_group *tg, void *data)
8891 struct rt_schedulable_data *d = data;
8892 struct task_group *child;
8893 unsigned long total, sum = 0;
8894 u64 period, runtime;
8896 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8897 runtime = tg->rt_bandwidth.rt_runtime;
8900 period = d->rt_period;
8901 runtime = d->rt_runtime;
8905 * Cannot have more runtime than the period.
8907 if (runtime > period && runtime != RUNTIME_INF)
8911 * Ensure we don't starve existing RT tasks.
8913 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8916 total = to_ratio(period, runtime);
8919 * Nobody can have more than the global setting allows.
8921 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8925 * The sum of our children's runtime should not exceed our own.
8927 list_for_each_entry_rcu(child, &tg->children, siblings) {
8928 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8929 runtime = child->rt_bandwidth.rt_runtime;
8931 if (child == d->tg) {
8932 period = d->rt_period;
8933 runtime = d->rt_runtime;
8936 sum += to_ratio(period, runtime);
8945 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8947 struct rt_schedulable_data data = {
8949 .rt_period = period,
8950 .rt_runtime = runtime,
8953 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8956 static int tg_set_bandwidth(struct task_group *tg,
8957 u64 rt_period, u64 rt_runtime)
8961 mutex_lock(&rt_constraints_mutex);
8962 read_lock(&tasklist_lock);
8963 err = __rt_schedulable(tg, rt_period, rt_runtime);
8967 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8968 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8969 tg->rt_bandwidth.rt_runtime = rt_runtime;
8971 for_each_possible_cpu(i) {
8972 struct rt_rq *rt_rq = tg->rt_rq[i];
8974 spin_lock(&rt_rq->rt_runtime_lock);
8975 rt_rq->rt_runtime = rt_runtime;
8976 spin_unlock(&rt_rq->rt_runtime_lock);
8978 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8980 read_unlock(&tasklist_lock);
8981 mutex_unlock(&rt_constraints_mutex);
8986 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8988 u64 rt_runtime, rt_period;
8990 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8991 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8992 if (rt_runtime_us < 0)
8993 rt_runtime = RUNTIME_INF;
8995 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8998 long sched_group_rt_runtime(struct task_group *tg)
9002 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
9005 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
9006 do_div(rt_runtime_us, NSEC_PER_USEC);
9007 return rt_runtime_us;
9010 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
9012 u64 rt_runtime, rt_period;
9014 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
9015 rt_runtime = tg->rt_bandwidth.rt_runtime;
9020 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9023 long sched_group_rt_period(struct task_group *tg)
9027 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9028 do_div(rt_period_us, NSEC_PER_USEC);
9029 return rt_period_us;
9032 static int sched_rt_global_constraints(void)
9034 u64 runtime, period;
9037 if (sysctl_sched_rt_period <= 0)
9040 runtime = global_rt_runtime();
9041 period = global_rt_period();
9044 * Sanity check on the sysctl variables.
9046 if (runtime > period && runtime != RUNTIME_INF)
9049 mutex_lock(&rt_constraints_mutex);
9050 read_lock(&tasklist_lock);
9051 ret = __rt_schedulable(NULL, 0, 0);
9052 read_unlock(&tasklist_lock);
9053 mutex_unlock(&rt_constraints_mutex);
9057 #else /* !CONFIG_RT_GROUP_SCHED */
9058 static int sched_rt_global_constraints(void)
9060 unsigned long flags;
9063 if (sysctl_sched_rt_period <= 0)
9066 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9067 for_each_possible_cpu(i) {
9068 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9070 spin_lock(&rt_rq->rt_runtime_lock);
9071 rt_rq->rt_runtime = global_rt_runtime();
9072 spin_unlock(&rt_rq->rt_runtime_lock);
9074 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9078 #endif /* CONFIG_RT_GROUP_SCHED */
9080 int sched_rt_handler(struct ctl_table *table, int write,
9081 struct file *filp, void __user *buffer, size_t *lenp,
9085 int old_period, old_runtime;
9086 static DEFINE_MUTEX(mutex);
9089 old_period = sysctl_sched_rt_period;
9090 old_runtime = sysctl_sched_rt_runtime;
9092 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
9094 if (!ret && write) {
9095 ret = sched_rt_global_constraints();
9097 sysctl_sched_rt_period = old_period;
9098 sysctl_sched_rt_runtime = old_runtime;
9100 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9101 def_rt_bandwidth.rt_period =
9102 ns_to_ktime(global_rt_period());
9105 mutex_unlock(&mutex);
9110 #ifdef CONFIG_CGROUP_SCHED
9112 /* return corresponding task_group object of a cgroup */
9113 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9115 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9116 struct task_group, css);
9119 static struct cgroup_subsys_state *
9120 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9122 struct task_group *tg, *parent;
9124 if (!cgrp->parent) {
9125 /* This is early initialization for the top cgroup */
9126 return &init_task_group.css;
9129 parent = cgroup_tg(cgrp->parent);
9130 tg = sched_create_group(parent);
9132 return ERR_PTR(-ENOMEM);
9138 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9140 struct task_group *tg = cgroup_tg(cgrp);
9142 sched_destroy_group(tg);
9146 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9147 struct task_struct *tsk)
9149 #ifdef CONFIG_RT_GROUP_SCHED
9150 /* Don't accept realtime tasks when there is no way for them to run */
9151 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
9154 /* We don't support RT-tasks being in separate groups */
9155 if (tsk->sched_class != &fair_sched_class)
9163 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9164 struct cgroup *old_cont, struct task_struct *tsk)
9166 sched_move_task(tsk);
9169 #ifdef CONFIG_FAIR_GROUP_SCHED
9170 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9173 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9176 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9178 struct task_group *tg = cgroup_tg(cgrp);
9180 return (u64) tg->shares;
9182 #endif /* CONFIG_FAIR_GROUP_SCHED */
9184 #ifdef CONFIG_RT_GROUP_SCHED
9185 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9188 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9191 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9193 return sched_group_rt_runtime(cgroup_tg(cgrp));
9196 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9199 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9202 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9204 return sched_group_rt_period(cgroup_tg(cgrp));
9206 #endif /* CONFIG_RT_GROUP_SCHED */
9208 static struct cftype cpu_files[] = {
9209 #ifdef CONFIG_FAIR_GROUP_SCHED
9212 .read_u64 = cpu_shares_read_u64,
9213 .write_u64 = cpu_shares_write_u64,
9216 #ifdef CONFIG_RT_GROUP_SCHED
9218 .name = "rt_runtime_us",
9219 .read_s64 = cpu_rt_runtime_read,
9220 .write_s64 = cpu_rt_runtime_write,
9223 .name = "rt_period_us",
9224 .read_u64 = cpu_rt_period_read_uint,
9225 .write_u64 = cpu_rt_period_write_uint,
9230 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9232 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9235 struct cgroup_subsys cpu_cgroup_subsys = {
9237 .create = cpu_cgroup_create,
9238 .destroy = cpu_cgroup_destroy,
9239 .can_attach = cpu_cgroup_can_attach,
9240 .attach = cpu_cgroup_attach,
9241 .populate = cpu_cgroup_populate,
9242 .subsys_id = cpu_cgroup_subsys_id,
9246 #endif /* CONFIG_CGROUP_SCHED */
9248 #ifdef CONFIG_CGROUP_CPUACCT
9251 * CPU accounting code for task groups.
9253 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9254 * (balbir@in.ibm.com).
9257 /* track cpu usage of a group of tasks and its child groups */
9259 struct cgroup_subsys_state css;
9260 /* cpuusage holds pointer to a u64-type object on every cpu */
9262 struct cpuacct *parent;
9265 struct cgroup_subsys cpuacct_subsys;
9267 /* return cpu accounting group corresponding to this container */
9268 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9270 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9271 struct cpuacct, css);
9274 /* return cpu accounting group to which this task belongs */
9275 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9277 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9278 struct cpuacct, css);
9281 /* create a new cpu accounting group */
9282 static struct cgroup_subsys_state *cpuacct_create(
9283 struct cgroup_subsys *ss, struct cgroup *cgrp)
9285 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9288 return ERR_PTR(-ENOMEM);
9290 ca->cpuusage = alloc_percpu(u64);
9291 if (!ca->cpuusage) {
9293 return ERR_PTR(-ENOMEM);
9297 ca->parent = cgroup_ca(cgrp->parent);
9302 /* destroy an existing cpu accounting group */
9304 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9306 struct cpuacct *ca = cgroup_ca(cgrp);
9308 free_percpu(ca->cpuusage);
9312 /* return total cpu usage (in nanoseconds) of a group */
9313 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9315 struct cpuacct *ca = cgroup_ca(cgrp);
9316 u64 totalcpuusage = 0;
9319 for_each_possible_cpu(i) {
9320 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9323 * Take rq->lock to make 64-bit addition safe on 32-bit
9326 spin_lock_irq(&cpu_rq(i)->lock);
9327 totalcpuusage += *cpuusage;
9328 spin_unlock_irq(&cpu_rq(i)->lock);
9331 return totalcpuusage;
9334 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9337 struct cpuacct *ca = cgroup_ca(cgrp);
9346 for_each_possible_cpu(i) {
9347 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9349 spin_lock_irq(&cpu_rq(i)->lock);
9351 spin_unlock_irq(&cpu_rq(i)->lock);
9357 static struct cftype files[] = {
9360 .read_u64 = cpuusage_read,
9361 .write_u64 = cpuusage_write,
9365 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9367 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9371 * charge this task's execution time to its accounting group.
9373 * called with rq->lock held.
9375 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9380 if (!cpuacct_subsys.active)
9383 cpu = task_cpu(tsk);
9386 for (; ca; ca = ca->parent) {
9387 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9388 *cpuusage += cputime;
9392 struct cgroup_subsys cpuacct_subsys = {
9394 .create = cpuacct_create,
9395 .destroy = cpuacct_destroy,
9396 .populate = cpuacct_populate,
9397 .subsys_id = cpuacct_subsys_id,
9399 #endif /* CONFIG_CGROUP_CPUACCT */