4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/reciprocal_div.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/bootmem.h>
72 #include <linux/debugfs.h>
73 #include <linux/ctype.h>
74 #include <linux/ftrace.h>
75 #include <trace/sched.h>
78 #include <asm/irq_regs.h>
80 #include "sched_cpupri.h"
83 * Convert user-nice values [ -20 ... 0 ... 19 ]
84 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
87 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
88 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
89 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
92 * 'User priority' is the nice value converted to something we
93 * can work with better when scaling various scheduler parameters,
94 * it's a [ 0 ... 39 ] range.
96 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
97 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
98 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
101 * Helpers for converting nanosecond timing to jiffy resolution
103 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
105 #define NICE_0_LOAD SCHED_LOAD_SCALE
106 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
109 * These are the 'tuning knobs' of the scheduler:
111 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
112 * Timeslices get refilled after they expire.
114 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * single value that denotes runtime == period, ie unlimited time.
119 #define RUNTIME_INF ((u64)~0ULL)
121 DEFINE_TRACE(sched_wait_task);
122 DEFINE_TRACE(sched_wakeup);
123 DEFINE_TRACE(sched_wakeup_new);
124 DEFINE_TRACE(sched_switch);
125 DEFINE_TRACE(sched_migrate_task);
129 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
130 * Since cpu_power is a 'constant', we can use a reciprocal divide.
132 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
134 return reciprocal_divide(load, sg->reciprocal_cpu_power);
138 * Each time a sched group cpu_power is changed,
139 * we must compute its reciprocal value
141 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
143 sg->__cpu_power += val;
144 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
148 static inline int rt_policy(int policy)
150 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
155 static inline int task_has_rt_policy(struct task_struct *p)
157 return rt_policy(p->policy);
161 * This is the priority-queue data structure of the RT scheduling class:
163 struct rt_prio_array {
164 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
165 struct list_head queue[MAX_RT_PRIO];
168 struct rt_bandwidth {
169 /* nests inside the rq lock: */
170 spinlock_t rt_runtime_lock;
173 struct hrtimer rt_period_timer;
176 static struct rt_bandwidth def_rt_bandwidth;
178 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
180 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
182 struct rt_bandwidth *rt_b =
183 container_of(timer, struct rt_bandwidth, rt_period_timer);
189 now = hrtimer_cb_get_time(timer);
190 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
195 idle = do_sched_rt_period_timer(rt_b, overrun);
198 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
202 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
204 rt_b->rt_period = ns_to_ktime(period);
205 rt_b->rt_runtime = runtime;
207 spin_lock_init(&rt_b->rt_runtime_lock);
209 hrtimer_init(&rt_b->rt_period_timer,
210 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
211 rt_b->rt_period_timer.function = sched_rt_period_timer;
212 rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_UNLOCKED;
215 static inline int rt_bandwidth_enabled(void)
217 return sysctl_sched_rt_runtime >= 0;
220 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
224 if (rt_bandwidth_enabled() && rt_b->rt_runtime == RUNTIME_INF)
227 if (hrtimer_active(&rt_b->rt_period_timer))
230 spin_lock(&rt_b->rt_runtime_lock);
232 if (hrtimer_active(&rt_b->rt_period_timer))
235 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
236 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
237 hrtimer_start_expires(&rt_b->rt_period_timer,
240 spin_unlock(&rt_b->rt_runtime_lock);
243 #ifdef CONFIG_RT_GROUP_SCHED
244 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
246 hrtimer_cancel(&rt_b->rt_period_timer);
251 * sched_domains_mutex serializes calls to arch_init_sched_domains,
252 * detach_destroy_domains and partition_sched_domains.
254 static DEFINE_MUTEX(sched_domains_mutex);
256 #ifdef CONFIG_GROUP_SCHED
258 #include <linux/cgroup.h>
262 static LIST_HEAD(task_groups);
264 /* task group related information */
266 #ifdef CONFIG_CGROUP_SCHED
267 struct cgroup_subsys_state css;
270 #ifdef CONFIG_USER_SCHED
274 #ifdef CONFIG_FAIR_GROUP_SCHED
275 /* schedulable entities of this group on each cpu */
276 struct sched_entity **se;
277 /* runqueue "owned" by this group on each cpu */
278 struct cfs_rq **cfs_rq;
279 unsigned long shares;
282 #ifdef CONFIG_RT_GROUP_SCHED
283 struct sched_rt_entity **rt_se;
284 struct rt_rq **rt_rq;
286 struct rt_bandwidth rt_bandwidth;
290 struct list_head list;
292 struct task_group *parent;
293 struct list_head siblings;
294 struct list_head children;
297 #ifdef CONFIG_USER_SCHED
299 /* Helper function to pass uid information to create_sched_user() */
300 void set_tg_uid(struct user_struct *user)
302 user->tg->uid = user->uid;
307 * Every UID task group (including init_task_group aka UID-0) will
308 * be a child to this group.
310 struct task_group root_task_group;
312 #ifdef CONFIG_FAIR_GROUP_SCHED
313 /* Default task group's sched entity on each cpu */
314 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
315 /* Default task group's cfs_rq on each cpu */
316 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
317 #endif /* CONFIG_FAIR_GROUP_SCHED */
319 #ifdef CONFIG_RT_GROUP_SCHED
320 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
321 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
322 #endif /* CONFIG_RT_GROUP_SCHED */
323 #else /* !CONFIG_USER_SCHED */
324 #define root_task_group init_task_group
325 #endif /* CONFIG_USER_SCHED */
327 /* task_group_lock serializes add/remove of task groups and also changes to
328 * a task group's cpu shares.
330 static DEFINE_SPINLOCK(task_group_lock);
332 #ifdef CONFIG_FAIR_GROUP_SCHED
333 #ifdef CONFIG_USER_SCHED
334 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
335 #else /* !CONFIG_USER_SCHED */
336 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
337 #endif /* CONFIG_USER_SCHED */
340 * A weight of 0 or 1 can cause arithmetics problems.
341 * A weight of a cfs_rq is the sum of weights of which entities
342 * are queued on this cfs_rq, so a weight of a entity should not be
343 * too large, so as the shares value of a task group.
344 * (The default weight is 1024 - so there's no practical
345 * limitation from this.)
348 #define MAX_SHARES (1UL << 18)
350 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
353 /* Default task group.
354 * Every task in system belong to this group at bootup.
356 struct task_group init_task_group;
358 /* return group to which a task belongs */
359 static inline struct task_group *task_group(struct task_struct *p)
361 struct task_group *tg;
363 #ifdef CONFIG_USER_SCHED
365 tg = __task_cred(p)->user->tg;
367 #elif defined(CONFIG_CGROUP_SCHED)
368 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
369 struct task_group, css);
371 tg = &init_task_group;
376 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
377 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
379 #ifdef CONFIG_FAIR_GROUP_SCHED
380 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
381 p->se.parent = task_group(p)->se[cpu];
384 #ifdef CONFIG_RT_GROUP_SCHED
385 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
386 p->rt.parent = task_group(p)->rt_se[cpu];
392 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
393 static inline struct task_group *task_group(struct task_struct *p)
398 #endif /* CONFIG_GROUP_SCHED */
400 /* CFS-related fields in a runqueue */
402 struct load_weight load;
403 unsigned long nr_running;
408 struct rb_root tasks_timeline;
409 struct rb_node *rb_leftmost;
411 struct list_head tasks;
412 struct list_head *balance_iterator;
415 * 'curr' points to currently running entity on this cfs_rq.
416 * It is set to NULL otherwise (i.e when none are currently running).
418 struct sched_entity *curr, *next, *last;
420 unsigned int nr_spread_over;
422 #ifdef CONFIG_FAIR_GROUP_SCHED
423 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
426 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
427 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
428 * (like users, containers etc.)
430 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
431 * list is used during load balance.
433 struct list_head leaf_cfs_rq_list;
434 struct task_group *tg; /* group that "owns" this runqueue */
438 * the part of load.weight contributed by tasks
440 unsigned long task_weight;
443 * h_load = weight * f(tg)
445 * Where f(tg) is the recursive weight fraction assigned to
448 unsigned long h_load;
451 * this cpu's part of tg->shares
453 unsigned long shares;
456 * load.weight at the time we set shares
458 unsigned long rq_weight;
463 /* Real-Time classes' related field in a runqueue: */
465 struct rt_prio_array active;
466 unsigned long rt_nr_running;
467 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
468 int highest_prio; /* highest queued rt task prio */
471 unsigned long rt_nr_migratory;
477 /* Nests inside the rq lock: */
478 spinlock_t rt_runtime_lock;
480 #ifdef CONFIG_RT_GROUP_SCHED
481 unsigned long rt_nr_boosted;
484 struct list_head leaf_rt_rq_list;
485 struct task_group *tg;
486 struct sched_rt_entity *rt_se;
493 * We add the notion of a root-domain which will be used to define per-domain
494 * variables. Each exclusive cpuset essentially defines an island domain by
495 * fully partitioning the member cpus from any other cpuset. Whenever a new
496 * exclusive cpuset is created, we also create and attach a new root-domain
506 * The "RT overload" flag: it gets set if a CPU has more than
507 * one runnable RT task.
512 struct cpupri cpupri;
517 * By default the system creates a single root-domain with all cpus as
518 * members (mimicking the global state we have today).
520 static struct root_domain def_root_domain;
525 * This is the main, per-CPU runqueue data structure.
527 * Locking rule: those places that want to lock multiple runqueues
528 * (such as the load balancing or the thread migration code), lock
529 * acquire operations must be ordered by ascending &runqueue.
536 * nr_running and cpu_load should be in the same cacheline because
537 * remote CPUs use both these fields when doing load calculation.
539 unsigned long nr_running;
540 #define CPU_LOAD_IDX_MAX 5
541 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
542 unsigned char idle_at_tick;
544 unsigned long last_tick_seen;
545 unsigned char in_nohz_recently;
547 /* capture load from *all* tasks on this cpu: */
548 struct load_weight load;
549 unsigned long nr_load_updates;
555 #ifdef CONFIG_FAIR_GROUP_SCHED
556 /* list of leaf cfs_rq on this cpu: */
557 struct list_head leaf_cfs_rq_list;
559 #ifdef CONFIG_RT_GROUP_SCHED
560 struct list_head leaf_rt_rq_list;
564 * This is part of a global counter where only the total sum
565 * over all CPUs matters. A task can increase this counter on
566 * one CPU and if it got migrated afterwards it may decrease
567 * it on another CPU. Always updated under the runqueue lock:
569 unsigned long nr_uninterruptible;
571 struct task_struct *curr, *idle;
572 unsigned long next_balance;
573 struct mm_struct *prev_mm;
580 struct root_domain *rd;
581 struct sched_domain *sd;
583 /* For active balancing */
586 /* cpu of this runqueue: */
590 unsigned long avg_load_per_task;
592 struct task_struct *migration_thread;
593 struct list_head migration_queue;
596 #ifdef CONFIG_SCHED_HRTICK
598 int hrtick_csd_pending;
599 struct call_single_data hrtick_csd;
601 struct hrtimer hrtick_timer;
604 #ifdef CONFIG_SCHEDSTATS
606 struct sched_info rq_sched_info;
607 unsigned long long rq_cpu_time;
608 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
610 /* sys_sched_yield() stats */
611 unsigned int yld_exp_empty;
612 unsigned int yld_act_empty;
613 unsigned int yld_both_empty;
614 unsigned int yld_count;
616 /* schedule() stats */
617 unsigned int sched_switch;
618 unsigned int sched_count;
619 unsigned int sched_goidle;
621 /* try_to_wake_up() stats */
622 unsigned int ttwu_count;
623 unsigned int ttwu_local;
626 unsigned int bkl_count;
630 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
632 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
634 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
637 static inline int cpu_of(struct rq *rq)
647 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
648 * See detach_destroy_domains: synchronize_sched for details.
650 * The domain tree of any CPU may only be accessed from within
651 * preempt-disabled sections.
653 #define for_each_domain(cpu, __sd) \
654 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
656 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
657 #define this_rq() (&__get_cpu_var(runqueues))
658 #define task_rq(p) cpu_rq(task_cpu(p))
659 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
661 inline void update_rq_clock(struct rq *rq)
663 rq->clock = sched_clock_cpu(cpu_of(rq));
667 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
669 #ifdef CONFIG_SCHED_DEBUG
670 # define const_debug __read_mostly
672 # define const_debug static const
678 * Returns true if the current cpu runqueue is locked.
679 * This interface allows printk to be called with the runqueue lock
680 * held and know whether or not it is OK to wake up the klogd.
682 int runqueue_is_locked(void)
685 struct rq *rq = cpu_rq(cpu);
688 ret = spin_is_locked(&rq->lock);
694 * Debugging: various feature bits
697 #define SCHED_FEAT(name, enabled) \
698 __SCHED_FEAT_##name ,
701 #include "sched_features.h"
706 #define SCHED_FEAT(name, enabled) \
707 (1UL << __SCHED_FEAT_##name) * enabled |
709 const_debug unsigned int sysctl_sched_features =
710 #include "sched_features.h"
715 #ifdef CONFIG_SCHED_DEBUG
716 #define SCHED_FEAT(name, enabled) \
719 static __read_mostly char *sched_feat_names[] = {
720 #include "sched_features.h"
726 static int sched_feat_show(struct seq_file *m, void *v)
730 for (i = 0; sched_feat_names[i]; i++) {
731 if (!(sysctl_sched_features & (1UL << i)))
733 seq_printf(m, "%s ", sched_feat_names[i]);
741 sched_feat_write(struct file *filp, const char __user *ubuf,
742 size_t cnt, loff_t *ppos)
752 if (copy_from_user(&buf, ubuf, cnt))
757 if (strncmp(buf, "NO_", 3) == 0) {
762 for (i = 0; sched_feat_names[i]; i++) {
763 int len = strlen(sched_feat_names[i]);
765 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
767 sysctl_sched_features &= ~(1UL << i);
769 sysctl_sched_features |= (1UL << i);
774 if (!sched_feat_names[i])
782 static int sched_feat_open(struct inode *inode, struct file *filp)
784 return single_open(filp, sched_feat_show, NULL);
787 static struct file_operations sched_feat_fops = {
788 .open = sched_feat_open,
789 .write = sched_feat_write,
792 .release = single_release,
795 static __init int sched_init_debug(void)
797 debugfs_create_file("sched_features", 0644, NULL, NULL,
802 late_initcall(sched_init_debug);
806 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
809 * Number of tasks to iterate in a single balance run.
810 * Limited because this is done with IRQs disabled.
812 const_debug unsigned int sysctl_sched_nr_migrate = 32;
815 * ratelimit for updating the group shares.
818 unsigned int sysctl_sched_shares_ratelimit = 250000;
821 * Inject some fuzzyness into changing the per-cpu group shares
822 * this avoids remote rq-locks at the expense of fairness.
825 unsigned int sysctl_sched_shares_thresh = 4;
828 * period over which we measure -rt task cpu usage in us.
831 unsigned int sysctl_sched_rt_period = 1000000;
833 static __read_mostly int scheduler_running;
836 * part of the period that we allow rt tasks to run in us.
839 int sysctl_sched_rt_runtime = 950000;
841 static inline u64 global_rt_period(void)
843 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
846 static inline u64 global_rt_runtime(void)
848 if (sysctl_sched_rt_runtime < 0)
851 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
854 #ifndef prepare_arch_switch
855 # define prepare_arch_switch(next) do { } while (0)
857 #ifndef finish_arch_switch
858 # define finish_arch_switch(prev) do { } while (0)
861 static inline int task_current(struct rq *rq, struct task_struct *p)
863 return rq->curr == p;
866 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
867 static inline int task_running(struct rq *rq, struct task_struct *p)
869 return task_current(rq, p);
872 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
876 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
878 #ifdef CONFIG_DEBUG_SPINLOCK
879 /* this is a valid case when another task releases the spinlock */
880 rq->lock.owner = current;
883 * If we are tracking spinlock dependencies then we have to
884 * fix up the runqueue lock - which gets 'carried over' from
887 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
889 spin_unlock_irq(&rq->lock);
892 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
893 static inline int task_running(struct rq *rq, struct task_struct *p)
898 return task_current(rq, p);
902 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
906 * We can optimise this out completely for !SMP, because the
907 * SMP rebalancing from interrupt is the only thing that cares
912 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
913 spin_unlock_irq(&rq->lock);
915 spin_unlock(&rq->lock);
919 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
923 * After ->oncpu is cleared, the task can be moved to a different CPU.
924 * We must ensure this doesn't happen until the switch is completely
930 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
934 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
937 * __task_rq_lock - lock the runqueue a given task resides on.
938 * Must be called interrupts disabled.
940 static inline struct rq *__task_rq_lock(struct task_struct *p)
944 struct rq *rq = task_rq(p);
945 spin_lock(&rq->lock);
946 if (likely(rq == task_rq(p)))
948 spin_unlock(&rq->lock);
953 * task_rq_lock - lock the runqueue a given task resides on and disable
954 * interrupts. Note the ordering: we can safely lookup the task_rq without
955 * explicitly disabling preemption.
957 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
963 local_irq_save(*flags);
965 spin_lock(&rq->lock);
966 if (likely(rq == task_rq(p)))
968 spin_unlock_irqrestore(&rq->lock, *flags);
972 void curr_rq_lock_irq_save(unsigned long *flags)
977 local_irq_save(*flags);
978 rq = cpu_rq(smp_processor_id());
979 spin_lock(&rq->lock);
982 void curr_rq_unlock_irq_restore(unsigned long *flags)
987 rq = cpu_rq(smp_processor_id());
988 spin_unlock(&rq->lock);
989 local_irq_restore(*flags);
992 void task_rq_unlock_wait(struct task_struct *p)
994 struct rq *rq = task_rq(p);
996 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
997 spin_unlock_wait(&rq->lock);
1000 static void __task_rq_unlock(struct rq *rq)
1001 __releases(rq->lock)
1003 spin_unlock(&rq->lock);
1006 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1007 __releases(rq->lock)
1009 spin_unlock_irqrestore(&rq->lock, *flags);
1013 * this_rq_lock - lock this runqueue and disable interrupts.
1015 static struct rq *this_rq_lock(void)
1016 __acquires(rq->lock)
1020 local_irq_disable();
1022 spin_lock(&rq->lock);
1027 #ifdef CONFIG_SCHED_HRTICK
1029 * Use HR-timers to deliver accurate preemption points.
1031 * Its all a bit involved since we cannot program an hrt while holding the
1032 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1035 * When we get rescheduled we reprogram the hrtick_timer outside of the
1041 * - enabled by features
1042 * - hrtimer is actually high res
1044 static inline int hrtick_enabled(struct rq *rq)
1046 if (!sched_feat(HRTICK))
1048 if (!cpu_active(cpu_of(rq)))
1050 return hrtimer_is_hres_active(&rq->hrtick_timer);
1053 static void hrtick_clear(struct rq *rq)
1055 if (hrtimer_active(&rq->hrtick_timer))
1056 hrtimer_cancel(&rq->hrtick_timer);
1060 * High-resolution timer tick.
1061 * Runs from hardirq context with interrupts disabled.
1063 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1065 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1067 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1069 spin_lock(&rq->lock);
1070 update_rq_clock(rq);
1071 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1072 spin_unlock(&rq->lock);
1074 return HRTIMER_NORESTART;
1079 * called from hardirq (IPI) context
1081 static void __hrtick_start(void *arg)
1083 struct rq *rq = arg;
1085 spin_lock(&rq->lock);
1086 hrtimer_restart(&rq->hrtick_timer);
1087 rq->hrtick_csd_pending = 0;
1088 spin_unlock(&rq->lock);
1092 * Called to set the hrtick timer state.
1094 * called with rq->lock held and irqs disabled
1096 static void hrtick_start(struct rq *rq, u64 delay)
1098 struct hrtimer *timer = &rq->hrtick_timer;
1099 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1101 hrtimer_set_expires(timer, time);
1103 if (rq == this_rq()) {
1104 hrtimer_restart(timer);
1105 } else if (!rq->hrtick_csd_pending) {
1106 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd);
1107 rq->hrtick_csd_pending = 1;
1112 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1114 int cpu = (int)(long)hcpu;
1117 case CPU_UP_CANCELED:
1118 case CPU_UP_CANCELED_FROZEN:
1119 case CPU_DOWN_PREPARE:
1120 case CPU_DOWN_PREPARE_FROZEN:
1122 case CPU_DEAD_FROZEN:
1123 hrtick_clear(cpu_rq(cpu));
1130 static __init void init_hrtick(void)
1132 hotcpu_notifier(hotplug_hrtick, 0);
1136 * Called to set the hrtick timer state.
1138 * called with rq->lock held and irqs disabled
1140 static void hrtick_start(struct rq *rq, u64 delay)
1142 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay), HRTIMER_MODE_REL);
1145 static inline void init_hrtick(void)
1148 #endif /* CONFIG_SMP */
1150 static void init_rq_hrtick(struct rq *rq)
1153 rq->hrtick_csd_pending = 0;
1155 rq->hrtick_csd.flags = 0;
1156 rq->hrtick_csd.func = __hrtick_start;
1157 rq->hrtick_csd.info = rq;
1160 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1161 rq->hrtick_timer.function = hrtick;
1162 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_PERCPU;
1164 #else /* CONFIG_SCHED_HRTICK */
1165 static inline void hrtick_clear(struct rq *rq)
1169 static inline void init_rq_hrtick(struct rq *rq)
1173 static inline void init_hrtick(void)
1176 #endif /* CONFIG_SCHED_HRTICK */
1179 * resched_task - mark a task 'to be rescheduled now'.
1181 * On UP this means the setting of the need_resched flag, on SMP it
1182 * might also involve a cross-CPU call to trigger the scheduler on
1187 #ifndef tsk_is_polling
1188 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1191 static void resched_task(struct task_struct *p)
1195 assert_spin_locked(&task_rq(p)->lock);
1197 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1200 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1203 if (cpu == smp_processor_id())
1206 /* NEED_RESCHED must be visible before we test polling */
1208 if (!tsk_is_polling(p))
1209 smp_send_reschedule(cpu);
1212 static void resched_cpu(int cpu)
1214 struct rq *rq = cpu_rq(cpu);
1215 unsigned long flags;
1217 if (!spin_trylock_irqsave(&rq->lock, flags))
1219 resched_task(cpu_curr(cpu));
1220 spin_unlock_irqrestore(&rq->lock, flags);
1225 * When add_timer_on() enqueues a timer into the timer wheel of an
1226 * idle CPU then this timer might expire before the next timer event
1227 * which is scheduled to wake up that CPU. In case of a completely
1228 * idle system the next event might even be infinite time into the
1229 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1230 * leaves the inner idle loop so the newly added timer is taken into
1231 * account when the CPU goes back to idle and evaluates the timer
1232 * wheel for the next timer event.
1234 void wake_up_idle_cpu(int cpu)
1236 struct rq *rq = cpu_rq(cpu);
1238 if (cpu == smp_processor_id())
1242 * This is safe, as this function is called with the timer
1243 * wheel base lock of (cpu) held. When the CPU is on the way
1244 * to idle and has not yet set rq->curr to idle then it will
1245 * be serialized on the timer wheel base lock and take the new
1246 * timer into account automatically.
1248 if (rq->curr != rq->idle)
1252 * We can set TIF_RESCHED on the idle task of the other CPU
1253 * lockless. The worst case is that the other CPU runs the
1254 * idle task through an additional NOOP schedule()
1256 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1258 /* NEED_RESCHED must be visible before we test polling */
1260 if (!tsk_is_polling(rq->idle))
1261 smp_send_reschedule(cpu);
1263 #endif /* CONFIG_NO_HZ */
1265 #else /* !CONFIG_SMP */
1266 static void resched_task(struct task_struct *p)
1268 assert_spin_locked(&task_rq(p)->lock);
1269 set_tsk_need_resched(p);
1271 #endif /* CONFIG_SMP */
1273 #if BITS_PER_LONG == 32
1274 # define WMULT_CONST (~0UL)
1276 # define WMULT_CONST (1UL << 32)
1279 #define WMULT_SHIFT 32
1282 * Shift right and round:
1284 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1287 * delta *= weight / lw
1289 static unsigned long
1290 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1291 struct load_weight *lw)
1295 if (!lw->inv_weight) {
1296 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1299 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1303 tmp = (u64)delta_exec * weight;
1305 * Check whether we'd overflow the 64-bit multiplication:
1307 if (unlikely(tmp > WMULT_CONST))
1308 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1311 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1313 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1316 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1322 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1329 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1330 * of tasks with abnormal "nice" values across CPUs the contribution that
1331 * each task makes to its run queue's load is weighted according to its
1332 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1333 * scaled version of the new time slice allocation that they receive on time
1337 #define WEIGHT_IDLEPRIO 2
1338 #define WMULT_IDLEPRIO (1 << 31)
1341 * Nice levels are multiplicative, with a gentle 10% change for every
1342 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1343 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1344 * that remained on nice 0.
1346 * The "10% effect" is relative and cumulative: from _any_ nice level,
1347 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1348 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1349 * If a task goes up by ~10% and another task goes down by ~10% then
1350 * the relative distance between them is ~25%.)
1352 static const int prio_to_weight[40] = {
1353 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1354 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1355 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1356 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1357 /* 0 */ 1024, 820, 655, 526, 423,
1358 /* 5 */ 335, 272, 215, 172, 137,
1359 /* 10 */ 110, 87, 70, 56, 45,
1360 /* 15 */ 36, 29, 23, 18, 15,
1364 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1366 * In cases where the weight does not change often, we can use the
1367 * precalculated inverse to speed up arithmetics by turning divisions
1368 * into multiplications:
1370 static const u32 prio_to_wmult[40] = {
1371 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1372 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1373 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1374 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1375 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1376 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1377 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1378 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1381 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1384 * runqueue iterator, to support SMP load-balancing between different
1385 * scheduling classes, without having to expose their internal data
1386 * structures to the load-balancing proper:
1388 struct rq_iterator {
1390 struct task_struct *(*start)(void *);
1391 struct task_struct *(*next)(void *);
1395 static unsigned long
1396 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1397 unsigned long max_load_move, struct sched_domain *sd,
1398 enum cpu_idle_type idle, int *all_pinned,
1399 int *this_best_prio, struct rq_iterator *iterator);
1402 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1403 struct sched_domain *sd, enum cpu_idle_type idle,
1404 struct rq_iterator *iterator);
1407 #ifdef CONFIG_CGROUP_CPUACCT
1408 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1410 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1413 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1415 update_load_add(&rq->load, load);
1418 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1420 update_load_sub(&rq->load, load);
1423 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1424 typedef int (*tg_visitor)(struct task_group *, void *);
1427 * Iterate the full tree, calling @down when first entering a node and @up when
1428 * leaving it for the final time.
1430 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1432 struct task_group *parent, *child;
1436 parent = &root_task_group;
1438 ret = (*down)(parent, data);
1441 list_for_each_entry_rcu(child, &parent->children, siblings) {
1448 ret = (*up)(parent, data);
1453 parent = parent->parent;
1462 static int tg_nop(struct task_group *tg, void *data)
1469 static unsigned long source_load(int cpu, int type);
1470 static unsigned long target_load(int cpu, int type);
1471 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1473 static unsigned long cpu_avg_load_per_task(int cpu)
1475 struct rq *rq = cpu_rq(cpu);
1476 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1479 rq->avg_load_per_task = rq->load.weight / nr_running;
1481 rq->avg_load_per_task = 0;
1483 return rq->avg_load_per_task;
1486 #ifdef CONFIG_FAIR_GROUP_SCHED
1488 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1491 * Calculate and set the cpu's group shares.
1494 update_group_shares_cpu(struct task_group *tg, int cpu,
1495 unsigned long sd_shares, unsigned long sd_rq_weight)
1497 unsigned long shares;
1498 unsigned long rq_weight;
1503 rq_weight = tg->cfs_rq[cpu]->rq_weight;
1506 * \Sum shares * rq_weight
1507 * shares = -----------------------
1511 shares = (sd_shares * rq_weight) / sd_rq_weight;
1512 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1514 if (abs(shares - tg->se[cpu]->load.weight) >
1515 sysctl_sched_shares_thresh) {
1516 struct rq *rq = cpu_rq(cpu);
1517 unsigned long flags;
1519 spin_lock_irqsave(&rq->lock, flags);
1520 tg->cfs_rq[cpu]->shares = shares;
1522 __set_se_shares(tg->se[cpu], shares);
1523 spin_unlock_irqrestore(&rq->lock, flags);
1528 * Re-compute the task group their per cpu shares over the given domain.
1529 * This needs to be done in a bottom-up fashion because the rq weight of a
1530 * parent group depends on the shares of its child groups.
1532 static int tg_shares_up(struct task_group *tg, void *data)
1534 unsigned long weight, rq_weight = 0;
1535 unsigned long shares = 0;
1536 struct sched_domain *sd = data;
1539 for_each_cpu_mask(i, sd->span) {
1541 * If there are currently no tasks on the cpu pretend there
1542 * is one of average load so that when a new task gets to
1543 * run here it will not get delayed by group starvation.
1545 weight = tg->cfs_rq[i]->load.weight;
1547 weight = NICE_0_LOAD;
1549 tg->cfs_rq[i]->rq_weight = weight;
1550 rq_weight += weight;
1551 shares += tg->cfs_rq[i]->shares;
1554 if ((!shares && rq_weight) || shares > tg->shares)
1555 shares = tg->shares;
1557 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1558 shares = tg->shares;
1560 for_each_cpu_mask(i, sd->span)
1561 update_group_shares_cpu(tg, i, shares, rq_weight);
1567 * Compute the cpu's hierarchical load factor for each task group.
1568 * This needs to be done in a top-down fashion because the load of a child
1569 * group is a fraction of its parents load.
1571 static int tg_load_down(struct task_group *tg, void *data)
1574 long cpu = (long)data;
1577 load = cpu_rq(cpu)->load.weight;
1579 load = tg->parent->cfs_rq[cpu]->h_load;
1580 load *= tg->cfs_rq[cpu]->shares;
1581 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1584 tg->cfs_rq[cpu]->h_load = load;
1589 static void update_shares(struct sched_domain *sd)
1591 u64 now = cpu_clock(raw_smp_processor_id());
1592 s64 elapsed = now - sd->last_update;
1594 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1595 sd->last_update = now;
1596 walk_tg_tree(tg_nop, tg_shares_up, sd);
1600 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1602 spin_unlock(&rq->lock);
1604 spin_lock(&rq->lock);
1607 static void update_h_load(long cpu)
1609 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1614 static inline void update_shares(struct sched_domain *sd)
1618 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1625 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1627 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1628 __releases(this_rq->lock)
1629 __acquires(busiest->lock)
1630 __acquires(this_rq->lock)
1634 if (unlikely(!irqs_disabled())) {
1635 /* printk() doesn't work good under rq->lock */
1636 spin_unlock(&this_rq->lock);
1639 if (unlikely(!spin_trylock(&busiest->lock))) {
1640 if (busiest < this_rq) {
1641 spin_unlock(&this_rq->lock);
1642 spin_lock(&busiest->lock);
1643 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1646 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1651 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1652 __releases(busiest->lock)
1654 spin_unlock(&busiest->lock);
1655 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1659 #ifdef CONFIG_FAIR_GROUP_SCHED
1660 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1663 cfs_rq->shares = shares;
1668 #include "sched_stats.h"
1669 #include "sched_idletask.c"
1670 #include "sched_fair.c"
1671 #include "sched_rt.c"
1672 #ifdef CONFIG_SCHED_DEBUG
1673 # include "sched_debug.c"
1676 #define sched_class_highest (&rt_sched_class)
1677 #define for_each_class(class) \
1678 for (class = sched_class_highest; class; class = class->next)
1680 static void inc_nr_running(struct rq *rq)
1685 static void dec_nr_running(struct rq *rq)
1690 static void set_load_weight(struct task_struct *p)
1692 if (task_has_rt_policy(p)) {
1693 p->se.load.weight = prio_to_weight[0] * 2;
1694 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1699 * SCHED_IDLE tasks get minimal weight:
1701 if (p->policy == SCHED_IDLE) {
1702 p->se.load.weight = WEIGHT_IDLEPRIO;
1703 p->se.load.inv_weight = WMULT_IDLEPRIO;
1707 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1708 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1711 static void update_avg(u64 *avg, u64 sample)
1713 s64 diff = sample - *avg;
1717 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1719 sched_info_queued(p);
1720 p->sched_class->enqueue_task(rq, p, wakeup);
1724 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1726 if (sleep && p->se.last_wakeup) {
1727 update_avg(&p->se.avg_overlap,
1728 p->se.sum_exec_runtime - p->se.last_wakeup);
1729 p->se.last_wakeup = 0;
1732 sched_info_dequeued(p);
1733 p->sched_class->dequeue_task(rq, p, sleep);
1738 * __normal_prio - return the priority that is based on the static prio
1740 static inline int __normal_prio(struct task_struct *p)
1742 return p->static_prio;
1746 * Calculate the expected normal priority: i.e. priority
1747 * without taking RT-inheritance into account. Might be
1748 * boosted by interactivity modifiers. Changes upon fork,
1749 * setprio syscalls, and whenever the interactivity
1750 * estimator recalculates.
1752 static inline int normal_prio(struct task_struct *p)
1756 if (task_has_rt_policy(p))
1757 prio = MAX_RT_PRIO-1 - p->rt_priority;
1759 prio = __normal_prio(p);
1764 * Calculate the current priority, i.e. the priority
1765 * taken into account by the scheduler. This value might
1766 * be boosted by RT tasks, or might be boosted by
1767 * interactivity modifiers. Will be RT if the task got
1768 * RT-boosted. If not then it returns p->normal_prio.
1770 static int effective_prio(struct task_struct *p)
1772 p->normal_prio = normal_prio(p);
1774 * If we are RT tasks or we were boosted to RT priority,
1775 * keep the priority unchanged. Otherwise, update priority
1776 * to the normal priority:
1778 if (!rt_prio(p->prio))
1779 return p->normal_prio;
1784 * activate_task - move a task to the runqueue.
1786 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1788 if (task_contributes_to_load(p))
1789 rq->nr_uninterruptible--;
1791 enqueue_task(rq, p, wakeup);
1796 * deactivate_task - remove a task from the runqueue.
1798 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1800 if (task_contributes_to_load(p))
1801 rq->nr_uninterruptible++;
1803 dequeue_task(rq, p, sleep);
1808 * task_curr - is this task currently executing on a CPU?
1809 * @p: the task in question.
1811 inline int task_curr(const struct task_struct *p)
1813 return cpu_curr(task_cpu(p)) == p;
1816 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1818 set_task_rq(p, cpu);
1821 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1822 * successfuly executed on another CPU. We must ensure that updates of
1823 * per-task data have been completed by this moment.
1826 task_thread_info(p)->cpu = cpu;
1830 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1831 const struct sched_class *prev_class,
1832 int oldprio, int running)
1834 if (prev_class != p->sched_class) {
1835 if (prev_class->switched_from)
1836 prev_class->switched_from(rq, p, running);
1837 p->sched_class->switched_to(rq, p, running);
1839 p->sched_class->prio_changed(rq, p, oldprio, running);
1844 /* Used instead of source_load when we know the type == 0 */
1845 static unsigned long weighted_cpuload(const int cpu)
1847 return cpu_rq(cpu)->load.weight;
1851 * Is this task likely cache-hot:
1854 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1859 * Buddy candidates are cache hot:
1861 if (sched_feat(CACHE_HOT_BUDDY) &&
1862 (&p->se == cfs_rq_of(&p->se)->next ||
1863 &p->se == cfs_rq_of(&p->se)->last))
1866 if (p->sched_class != &fair_sched_class)
1869 if (sysctl_sched_migration_cost == -1)
1871 if (sysctl_sched_migration_cost == 0)
1874 delta = now - p->se.exec_start;
1876 return delta < (s64)sysctl_sched_migration_cost;
1880 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1882 int old_cpu = task_cpu(p);
1883 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1884 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1885 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1888 clock_offset = old_rq->clock - new_rq->clock;
1890 trace_sched_migrate_task(p, task_cpu(p), new_cpu);
1892 #ifdef CONFIG_SCHEDSTATS
1893 if (p->se.wait_start)
1894 p->se.wait_start -= clock_offset;
1895 if (p->se.sleep_start)
1896 p->se.sleep_start -= clock_offset;
1897 if (p->se.block_start)
1898 p->se.block_start -= clock_offset;
1900 if (old_cpu != new_cpu) {
1901 p->se.nr_migrations++;
1902 #ifdef CONFIG_SCHEDSTATS
1903 if (task_hot(p, old_rq->clock, NULL))
1904 schedstat_inc(p, se.nr_forced2_migrations);
1907 p->se.vruntime -= old_cfsrq->min_vruntime -
1908 new_cfsrq->min_vruntime;
1910 __set_task_cpu(p, new_cpu);
1913 struct migration_req {
1914 struct list_head list;
1916 struct task_struct *task;
1919 struct completion done;
1923 * The task's runqueue lock must be held.
1924 * Returns true if you have to wait for migration thread.
1927 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1929 struct rq *rq = task_rq(p);
1932 * If the task is not on a runqueue (and not running), then
1933 * it is sufficient to simply update the task's cpu field.
1935 if (!p->se.on_rq && !task_running(rq, p)) {
1936 set_task_cpu(p, dest_cpu);
1940 init_completion(&req->done);
1942 req->dest_cpu = dest_cpu;
1943 list_add(&req->list, &rq->migration_queue);
1949 * wait_task_inactive - wait for a thread to unschedule.
1951 * If @match_state is nonzero, it's the @p->state value just checked and
1952 * not expected to change. If it changes, i.e. @p might have woken up,
1953 * then return zero. When we succeed in waiting for @p to be off its CPU,
1954 * we return a positive number (its total switch count). If a second call
1955 * a short while later returns the same number, the caller can be sure that
1956 * @p has remained unscheduled the whole time.
1958 * The caller must ensure that the task *will* unschedule sometime soon,
1959 * else this function might spin for a *long* time. This function can't
1960 * be called with interrupts off, or it may introduce deadlock with
1961 * smp_call_function() if an IPI is sent by the same process we are
1962 * waiting to become inactive.
1964 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1966 unsigned long flags;
1973 * We do the initial early heuristics without holding
1974 * any task-queue locks at all. We'll only try to get
1975 * the runqueue lock when things look like they will
1981 * If the task is actively running on another CPU
1982 * still, just relax and busy-wait without holding
1985 * NOTE! Since we don't hold any locks, it's not
1986 * even sure that "rq" stays as the right runqueue!
1987 * But we don't care, since "task_running()" will
1988 * return false if the runqueue has changed and p
1989 * is actually now running somewhere else!
1991 while (task_running(rq, p)) {
1992 if (match_state && unlikely(p->state != match_state))
1998 * Ok, time to look more closely! We need the rq
1999 * lock now, to be *sure*. If we're wrong, we'll
2000 * just go back and repeat.
2002 rq = task_rq_lock(p, &flags);
2003 trace_sched_wait_task(rq, p);
2004 running = task_running(rq, p);
2005 on_rq = p->se.on_rq;
2007 if (!match_state || p->state == match_state)
2008 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2009 task_rq_unlock(rq, &flags);
2012 * If it changed from the expected state, bail out now.
2014 if (unlikely(!ncsw))
2018 * Was it really running after all now that we
2019 * checked with the proper locks actually held?
2021 * Oops. Go back and try again..
2023 if (unlikely(running)) {
2029 * It's not enough that it's not actively running,
2030 * it must be off the runqueue _entirely_, and not
2033 * So if it wa still runnable (but just not actively
2034 * running right now), it's preempted, and we should
2035 * yield - it could be a while.
2037 if (unlikely(on_rq)) {
2038 schedule_timeout_uninterruptible(1);
2043 * Ahh, all good. It wasn't running, and it wasn't
2044 * runnable, which means that it will never become
2045 * running in the future either. We're all done!
2054 * kick_process - kick a running thread to enter/exit the kernel
2055 * @p: the to-be-kicked thread
2057 * Cause a process which is running on another CPU to enter
2058 * kernel-mode, without any delay. (to get signals handled.)
2060 * NOTE: this function doesnt have to take the runqueue lock,
2061 * because all it wants to ensure is that the remote task enters
2062 * the kernel. If the IPI races and the task has been migrated
2063 * to another CPU then no harm is done and the purpose has been
2066 void kick_process(struct task_struct *p)
2072 if ((cpu != smp_processor_id()) && task_curr(p))
2073 smp_send_reschedule(cpu);
2078 * Return a low guess at the load of a migration-source cpu weighted
2079 * according to the scheduling class and "nice" value.
2081 * We want to under-estimate the load of migration sources, to
2082 * balance conservatively.
2084 static unsigned long source_load(int cpu, int type)
2086 struct rq *rq = cpu_rq(cpu);
2087 unsigned long total = weighted_cpuload(cpu);
2089 if (type == 0 || !sched_feat(LB_BIAS))
2092 return min(rq->cpu_load[type-1], total);
2096 * Return a high guess at the load of a migration-target cpu weighted
2097 * according to the scheduling class and "nice" value.
2099 static unsigned long target_load(int cpu, int type)
2101 struct rq *rq = cpu_rq(cpu);
2102 unsigned long total = weighted_cpuload(cpu);
2104 if (type == 0 || !sched_feat(LB_BIAS))
2107 return max(rq->cpu_load[type-1], total);
2111 * find_idlest_group finds and returns the least busy CPU group within the
2114 static struct sched_group *
2115 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2117 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2118 unsigned long min_load = ULONG_MAX, this_load = 0;
2119 int load_idx = sd->forkexec_idx;
2120 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2123 unsigned long load, avg_load;
2127 /* Skip over this group if it has no CPUs allowed */
2128 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
2131 local_group = cpu_isset(this_cpu, group->cpumask);
2133 /* Tally up the load of all CPUs in the group */
2136 for_each_cpu_mask_nr(i, group->cpumask) {
2137 /* Bias balancing toward cpus of our domain */
2139 load = source_load(i, load_idx);
2141 load = target_load(i, load_idx);
2146 /* Adjust by relative CPU power of the group */
2147 avg_load = sg_div_cpu_power(group,
2148 avg_load * SCHED_LOAD_SCALE);
2151 this_load = avg_load;
2153 } else if (avg_load < min_load) {
2154 min_load = avg_load;
2157 } while (group = group->next, group != sd->groups);
2159 if (!idlest || 100*this_load < imbalance*min_load)
2165 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2168 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2171 unsigned long load, min_load = ULONG_MAX;
2175 /* Traverse only the allowed CPUs */
2176 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
2178 for_each_cpu_mask_nr(i, *tmp) {
2179 load = weighted_cpuload(i);
2181 if (load < min_load || (load == min_load && i == this_cpu)) {
2191 * sched_balance_self: balance the current task (running on cpu) in domains
2192 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2195 * Balance, ie. select the least loaded group.
2197 * Returns the target CPU number, or the same CPU if no balancing is needed.
2199 * preempt must be disabled.
2201 static int sched_balance_self(int cpu, int flag)
2203 struct task_struct *t = current;
2204 struct sched_domain *tmp, *sd = NULL;
2206 for_each_domain(cpu, tmp) {
2208 * If power savings logic is enabled for a domain, stop there.
2210 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2212 if (tmp->flags & flag)
2220 cpumask_t span, tmpmask;
2221 struct sched_group *group;
2222 int new_cpu, weight;
2224 if (!(sd->flags & flag)) {
2230 group = find_idlest_group(sd, t, cpu);
2236 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2237 if (new_cpu == -1 || new_cpu == cpu) {
2238 /* Now try balancing at a lower domain level of cpu */
2243 /* Now try balancing at a lower domain level of new_cpu */
2246 weight = cpus_weight(span);
2247 for_each_domain(cpu, tmp) {
2248 if (weight <= cpus_weight(tmp->span))
2250 if (tmp->flags & flag)
2253 /* while loop will break here if sd == NULL */
2259 #endif /* CONFIG_SMP */
2262 * task_oncpu_function_call - call a function on the cpu on which a task runs
2263 * @p: the task to evaluate
2264 * @func: the function to be called
2265 * @info: the function call argument
2267 * Calls the function @func when the task is currently running. This might
2268 * be on the current CPU, which just calls the function directly
2270 void task_oncpu_function_call(struct task_struct *p,
2271 void (*func) (void *info), void *info)
2278 smp_call_function_single(cpu, func, info, 1);
2283 * try_to_wake_up - wake up a thread
2284 * @p: the to-be-woken-up thread
2285 * @state: the mask of task states that can be woken
2286 * @sync: do a synchronous wakeup?
2288 * Put it on the run-queue if it's not already there. The "current"
2289 * thread is always on the run-queue (except when the actual
2290 * re-schedule is in progress), and as such you're allowed to do
2291 * the simpler "current->state = TASK_RUNNING" to mark yourself
2292 * runnable without the overhead of this.
2294 * returns failure only if the task is already active.
2296 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2298 int cpu, orig_cpu, this_cpu, success = 0;
2299 unsigned long flags;
2303 if (!sched_feat(SYNC_WAKEUPS))
2307 if (sched_feat(LB_WAKEUP_UPDATE)) {
2308 struct sched_domain *sd;
2310 this_cpu = raw_smp_processor_id();
2313 for_each_domain(this_cpu, sd) {
2314 if (cpu_isset(cpu, sd->span)) {
2323 rq = task_rq_lock(p, &flags);
2324 update_rq_clock(rq);
2325 old_state = p->state;
2326 if (!(old_state & state))
2334 this_cpu = smp_processor_id();
2337 if (unlikely(task_running(rq, p)))
2340 cpu = p->sched_class->select_task_rq(p, sync);
2341 if (cpu != orig_cpu) {
2342 set_task_cpu(p, cpu);
2343 task_rq_unlock(rq, &flags);
2344 /* might preempt at this point */
2345 rq = task_rq_lock(p, &flags);
2346 old_state = p->state;
2347 if (!(old_state & state))
2352 this_cpu = smp_processor_id();
2356 #ifdef CONFIG_SCHEDSTATS
2357 schedstat_inc(rq, ttwu_count);
2358 if (cpu == this_cpu)
2359 schedstat_inc(rq, ttwu_local);
2361 struct sched_domain *sd;
2362 for_each_domain(this_cpu, sd) {
2363 if (cpu_isset(cpu, sd->span)) {
2364 schedstat_inc(sd, ttwu_wake_remote);
2369 #endif /* CONFIG_SCHEDSTATS */
2372 #endif /* CONFIG_SMP */
2373 schedstat_inc(p, se.nr_wakeups);
2375 schedstat_inc(p, se.nr_wakeups_sync);
2376 if (orig_cpu != cpu)
2377 schedstat_inc(p, se.nr_wakeups_migrate);
2378 if (cpu == this_cpu)
2379 schedstat_inc(p, se.nr_wakeups_local);
2381 schedstat_inc(p, se.nr_wakeups_remote);
2382 activate_task(rq, p, 1);
2386 trace_sched_wakeup(rq, p, success);
2387 check_preempt_curr(rq, p, sync);
2389 p->state = TASK_RUNNING;
2391 if (p->sched_class->task_wake_up)
2392 p->sched_class->task_wake_up(rq, p);
2395 current->se.last_wakeup = current->se.sum_exec_runtime;
2397 task_rq_unlock(rq, &flags);
2402 int wake_up_process(struct task_struct *p)
2404 return try_to_wake_up(p, TASK_ALL, 0);
2406 EXPORT_SYMBOL(wake_up_process);
2408 int wake_up_state(struct task_struct *p, unsigned int state)
2410 return try_to_wake_up(p, state, 0);
2414 * Perform scheduler related setup for a newly forked process p.
2415 * p is forked by current.
2417 * __sched_fork() is basic setup used by init_idle() too:
2419 static void __sched_fork(struct task_struct *p)
2421 p->se.exec_start = 0;
2422 p->se.sum_exec_runtime = 0;
2423 p->se.prev_sum_exec_runtime = 0;
2424 p->se.nr_migrations = 0;
2425 p->se.last_wakeup = 0;
2426 p->se.avg_overlap = 0;
2428 #ifdef CONFIG_SCHEDSTATS
2429 p->se.wait_start = 0;
2430 p->se.sum_sleep_runtime = 0;
2431 p->se.sleep_start = 0;
2432 p->se.block_start = 0;
2433 p->se.sleep_max = 0;
2434 p->se.block_max = 0;
2436 p->se.slice_max = 0;
2440 INIT_LIST_HEAD(&p->rt.run_list);
2442 INIT_LIST_HEAD(&p->se.group_node);
2444 #ifdef CONFIG_PREEMPT_NOTIFIERS
2445 INIT_HLIST_HEAD(&p->preempt_notifiers);
2449 * We mark the process as running here, but have not actually
2450 * inserted it onto the runqueue yet. This guarantees that
2451 * nobody will actually run it, and a signal or other external
2452 * event cannot wake it up and insert it on the runqueue either.
2454 p->state = TASK_RUNNING;
2458 * fork()/clone()-time setup:
2460 void sched_fork(struct task_struct *p, int clone_flags)
2462 int cpu = get_cpu();
2467 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2469 set_task_cpu(p, cpu);
2472 * Make sure we do not leak PI boosting priority to the child:
2474 p->prio = current->normal_prio;
2475 if (!rt_prio(p->prio))
2476 p->sched_class = &fair_sched_class;
2478 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2479 if (likely(sched_info_on()))
2480 memset(&p->sched_info, 0, sizeof(p->sched_info));
2482 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2485 #ifdef CONFIG_PREEMPT
2486 /* Want to start with kernel preemption disabled. */
2487 task_thread_info(p)->preempt_count = 1;
2493 * wake_up_new_task - wake up a newly created task for the first time.
2495 * This function will do some initial scheduler statistics housekeeping
2496 * that must be done for every newly created context, then puts the task
2497 * on the runqueue and wakes it.
2499 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2501 unsigned long flags;
2504 rq = task_rq_lock(p, &flags);
2505 BUG_ON(p->state != TASK_RUNNING);
2506 update_rq_clock(rq);
2508 p->prio = effective_prio(p);
2510 if (!p->sched_class->task_new || !current->se.on_rq) {
2511 activate_task(rq, p, 0);
2514 * Let the scheduling class do new task startup
2515 * management (if any):
2517 p->sched_class->task_new(rq, p);
2520 trace_sched_wakeup_new(rq, p, 1);
2521 check_preempt_curr(rq, p, 0);
2523 if (p->sched_class->task_wake_up)
2524 p->sched_class->task_wake_up(rq, p);
2526 task_rq_unlock(rq, &flags);
2529 #ifdef CONFIG_PREEMPT_NOTIFIERS
2532 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2533 * @notifier: notifier struct to register
2535 void preempt_notifier_register(struct preempt_notifier *notifier)
2537 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2539 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2542 * preempt_notifier_unregister - no longer interested in preemption notifications
2543 * @notifier: notifier struct to unregister
2545 * This is safe to call from within a preemption notifier.
2547 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2549 hlist_del(¬ifier->link);
2551 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2553 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2555 struct preempt_notifier *notifier;
2556 struct hlist_node *node;
2558 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2559 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2563 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2564 struct task_struct *next)
2566 struct preempt_notifier *notifier;
2567 struct hlist_node *node;
2569 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2570 notifier->ops->sched_out(notifier, next);
2573 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2575 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2580 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2581 struct task_struct *next)
2585 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2588 * prepare_task_switch - prepare to switch tasks
2589 * @rq: the runqueue preparing to switch
2590 * @prev: the current task that is being switched out
2591 * @next: the task we are going to switch to.
2593 * This is called with the rq lock held and interrupts off. It must
2594 * be paired with a subsequent finish_task_switch after the context
2597 * prepare_task_switch sets up locking and calls architecture specific
2601 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2602 struct task_struct *next)
2604 fire_sched_out_preempt_notifiers(prev, next);
2605 prepare_lock_switch(rq, next);
2606 prepare_arch_switch(next);
2610 * finish_task_switch - clean up after a task-switch
2611 * @rq: runqueue associated with task-switch
2612 * @prev: the thread we just switched away from.
2614 * finish_task_switch must be called after the context switch, paired
2615 * with a prepare_task_switch call before the context switch.
2616 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2617 * and do any other architecture-specific cleanup actions.
2619 * Note that we may have delayed dropping an mm in context_switch(). If
2620 * so, we finish that here outside of the runqueue lock. (Doing it
2621 * with the lock held can cause deadlocks; see schedule() for
2624 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2625 __releases(rq->lock)
2627 struct mm_struct *mm = rq->prev_mm;
2633 * A task struct has one reference for the use as "current".
2634 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2635 * schedule one last time. The schedule call will never return, and
2636 * the scheduled task must drop that reference.
2637 * The test for TASK_DEAD must occur while the runqueue locks are
2638 * still held, otherwise prev could be scheduled on another cpu, die
2639 * there before we look at prev->state, and then the reference would
2641 * Manfred Spraul <manfred@colorfullife.com>
2643 prev_state = prev->state;
2644 finish_arch_switch(prev);
2645 perf_counter_task_sched_in(current, cpu_of(rq));
2646 finish_lock_switch(rq, prev);
2648 if (current->sched_class->post_schedule)
2649 current->sched_class->post_schedule(rq);
2652 fire_sched_in_preempt_notifiers(current);
2655 if (unlikely(prev_state == TASK_DEAD)) {
2657 * Remove function-return probe instances associated with this
2658 * task and put them back on the free list.
2660 kprobe_flush_task(prev);
2661 put_task_struct(prev);
2666 * schedule_tail - first thing a freshly forked thread must call.
2667 * @prev: the thread we just switched away from.
2669 asmlinkage void schedule_tail(struct task_struct *prev)
2670 __releases(rq->lock)
2672 struct rq *rq = this_rq();
2674 finish_task_switch(rq, prev);
2675 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2676 /* In this case, finish_task_switch does not reenable preemption */
2679 if (current->set_child_tid)
2680 put_user(task_pid_vnr(current), current->set_child_tid);
2684 * context_switch - switch to the new MM and the new
2685 * thread's register state.
2688 context_switch(struct rq *rq, struct task_struct *prev,
2689 struct task_struct *next)
2691 struct mm_struct *mm, *oldmm;
2693 prepare_task_switch(rq, prev, next);
2694 trace_sched_switch(rq, prev, next);
2696 oldmm = prev->active_mm;
2698 * For paravirt, this is coupled with an exit in switch_to to
2699 * combine the page table reload and the switch backend into
2702 arch_enter_lazy_cpu_mode();
2704 if (unlikely(!mm)) {
2705 next->active_mm = oldmm;
2706 atomic_inc(&oldmm->mm_count);
2707 enter_lazy_tlb(oldmm, next);
2709 switch_mm(oldmm, mm, next);
2711 if (unlikely(!prev->mm)) {
2712 prev->active_mm = NULL;
2713 rq->prev_mm = oldmm;
2716 * Since the runqueue lock will be released by the next
2717 * task (which is an invalid locking op but in the case
2718 * of the scheduler it's an obvious special-case), so we
2719 * do an early lockdep release here:
2721 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2722 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2725 /* Here we just switch the register state and the stack. */
2726 switch_to(prev, next, prev);
2730 * this_rq must be evaluated again because prev may have moved
2731 * CPUs since it called schedule(), thus the 'rq' on its stack
2732 * frame will be invalid.
2734 finish_task_switch(this_rq(), prev);
2738 * nr_running, nr_uninterruptible and nr_context_switches:
2740 * externally visible scheduler statistics: current number of runnable
2741 * threads, current number of uninterruptible-sleeping threads, total
2742 * number of context switches performed since bootup.
2744 unsigned long nr_running(void)
2746 unsigned long i, sum = 0;
2748 for_each_online_cpu(i)
2749 sum += cpu_rq(i)->nr_running;
2754 unsigned long nr_uninterruptible(void)
2756 unsigned long i, sum = 0;
2758 for_each_possible_cpu(i)
2759 sum += cpu_rq(i)->nr_uninterruptible;
2762 * Since we read the counters lockless, it might be slightly
2763 * inaccurate. Do not allow it to go below zero though:
2765 if (unlikely((long)sum < 0))
2771 unsigned long long nr_context_switches(void)
2774 unsigned long long sum = 0;
2776 for_each_possible_cpu(i)
2777 sum += cpu_rq(i)->nr_switches;
2782 unsigned long nr_iowait(void)
2784 unsigned long i, sum = 0;
2786 for_each_possible_cpu(i)
2787 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2792 unsigned long nr_active(void)
2794 unsigned long i, running = 0, uninterruptible = 0;
2796 for_each_online_cpu(i) {
2797 running += cpu_rq(i)->nr_running;
2798 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2801 if (unlikely((long)uninterruptible < 0))
2802 uninterruptible = 0;
2804 return running + uninterruptible;
2808 * Update rq->cpu_load[] statistics. This function is usually called every
2809 * scheduler tick (TICK_NSEC).
2811 static void update_cpu_load(struct rq *this_rq)
2813 unsigned long this_load = this_rq->load.weight;
2816 this_rq->nr_load_updates++;
2818 /* Update our load: */
2819 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2820 unsigned long old_load, new_load;
2822 /* scale is effectively 1 << i now, and >> i divides by scale */
2824 old_load = this_rq->cpu_load[i];
2825 new_load = this_load;
2827 * Round up the averaging division if load is increasing. This
2828 * prevents us from getting stuck on 9 if the load is 10, for
2831 if (new_load > old_load)
2832 new_load += scale-1;
2833 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2840 * double_rq_lock - safely lock two runqueues
2842 * Note this does not disable interrupts like task_rq_lock,
2843 * you need to do so manually before calling.
2845 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2846 __acquires(rq1->lock)
2847 __acquires(rq2->lock)
2849 BUG_ON(!irqs_disabled());
2851 spin_lock(&rq1->lock);
2852 __acquire(rq2->lock); /* Fake it out ;) */
2855 spin_lock(&rq1->lock);
2856 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2858 spin_lock(&rq2->lock);
2859 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2862 update_rq_clock(rq1);
2863 update_rq_clock(rq2);
2867 * double_rq_unlock - safely unlock two runqueues
2869 * Note this does not restore interrupts like task_rq_unlock,
2870 * you need to do so manually after calling.
2872 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2873 __releases(rq1->lock)
2874 __releases(rq2->lock)
2876 spin_unlock(&rq1->lock);
2878 spin_unlock(&rq2->lock);
2880 __release(rq2->lock);
2884 * If dest_cpu is allowed for this process, migrate the task to it.
2885 * This is accomplished by forcing the cpu_allowed mask to only
2886 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2887 * the cpu_allowed mask is restored.
2889 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2891 struct migration_req req;
2892 unsigned long flags;
2895 rq = task_rq_lock(p, &flags);
2896 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2897 || unlikely(!cpu_active(dest_cpu)))
2900 /* force the process onto the specified CPU */
2901 if (migrate_task(p, dest_cpu, &req)) {
2902 /* Need to wait for migration thread (might exit: take ref). */
2903 struct task_struct *mt = rq->migration_thread;
2905 get_task_struct(mt);
2906 task_rq_unlock(rq, &flags);
2907 wake_up_process(mt);
2908 put_task_struct(mt);
2909 wait_for_completion(&req.done);
2914 task_rq_unlock(rq, &flags);
2918 * sched_exec - execve() is a valuable balancing opportunity, because at
2919 * this point the task has the smallest effective memory and cache footprint.
2921 void sched_exec(void)
2923 int new_cpu, this_cpu = get_cpu();
2924 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2926 if (new_cpu != this_cpu)
2927 sched_migrate_task(current, new_cpu);
2931 * pull_task - move a task from a remote runqueue to the local runqueue.
2932 * Both runqueues must be locked.
2934 static void pull_task(struct rq *src_rq, struct task_struct *p,
2935 struct rq *this_rq, int this_cpu)
2937 deactivate_task(src_rq, p, 0);
2938 set_task_cpu(p, this_cpu);
2939 activate_task(this_rq, p, 0);
2941 * Note that idle threads have a prio of MAX_PRIO, for this test
2942 * to be always true for them.
2944 check_preempt_curr(this_rq, p, 0);
2948 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2951 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2952 struct sched_domain *sd, enum cpu_idle_type idle,
2956 * We do not migrate tasks that are:
2957 * 1) running (obviously), or
2958 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2959 * 3) are cache-hot on their current CPU.
2961 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2962 schedstat_inc(p, se.nr_failed_migrations_affine);
2967 if (task_running(rq, p)) {
2968 schedstat_inc(p, se.nr_failed_migrations_running);
2973 * Aggressive migration if:
2974 * 1) task is cache cold, or
2975 * 2) too many balance attempts have failed.
2978 if (!task_hot(p, rq->clock, sd) ||
2979 sd->nr_balance_failed > sd->cache_nice_tries) {
2980 #ifdef CONFIG_SCHEDSTATS
2981 if (task_hot(p, rq->clock, sd)) {
2982 schedstat_inc(sd, lb_hot_gained[idle]);
2983 schedstat_inc(p, se.nr_forced_migrations);
2989 if (task_hot(p, rq->clock, sd)) {
2990 schedstat_inc(p, se.nr_failed_migrations_hot);
2996 static unsigned long
2997 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2998 unsigned long max_load_move, struct sched_domain *sd,
2999 enum cpu_idle_type idle, int *all_pinned,
3000 int *this_best_prio, struct rq_iterator *iterator)
3002 int loops = 0, pulled = 0, pinned = 0;
3003 struct task_struct *p;
3004 long rem_load_move = max_load_move;
3006 if (max_load_move == 0)
3012 * Start the load-balancing iterator:
3014 p = iterator->start(iterator->arg);
3016 if (!p || loops++ > sysctl_sched_nr_migrate)
3019 if ((p->se.load.weight >> 1) > rem_load_move ||
3020 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3021 p = iterator->next(iterator->arg);
3025 pull_task(busiest, p, this_rq, this_cpu);
3027 rem_load_move -= p->se.load.weight;
3030 * We only want to steal up to the prescribed amount of weighted load.
3032 if (rem_load_move > 0) {
3033 if (p->prio < *this_best_prio)
3034 *this_best_prio = p->prio;
3035 p = iterator->next(iterator->arg);
3040 * Right now, this is one of only two places pull_task() is called,
3041 * so we can safely collect pull_task() stats here rather than
3042 * inside pull_task().
3044 schedstat_add(sd, lb_gained[idle], pulled);
3047 *all_pinned = pinned;
3049 return max_load_move - rem_load_move;
3053 * move_tasks tries to move up to max_load_move weighted load from busiest to
3054 * this_rq, as part of a balancing operation within domain "sd".
3055 * Returns 1 if successful and 0 otherwise.
3057 * Called with both runqueues locked.
3059 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3060 unsigned long max_load_move,
3061 struct sched_domain *sd, enum cpu_idle_type idle,
3064 const struct sched_class *class = sched_class_highest;
3065 unsigned long total_load_moved = 0;
3066 int this_best_prio = this_rq->curr->prio;
3070 class->load_balance(this_rq, this_cpu, busiest,
3071 max_load_move - total_load_moved,
3072 sd, idle, all_pinned, &this_best_prio);
3073 class = class->next;
3075 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3078 } while (class && max_load_move > total_load_moved);
3080 return total_load_moved > 0;
3084 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3085 struct sched_domain *sd, enum cpu_idle_type idle,
3086 struct rq_iterator *iterator)
3088 struct task_struct *p = iterator->start(iterator->arg);
3092 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3093 pull_task(busiest, p, this_rq, this_cpu);
3095 * Right now, this is only the second place pull_task()
3096 * is called, so we can safely collect pull_task()
3097 * stats here rather than inside pull_task().
3099 schedstat_inc(sd, lb_gained[idle]);
3103 p = iterator->next(iterator->arg);
3110 * move_one_task tries to move exactly one task from busiest to this_rq, as
3111 * part of active balancing operations within "domain".
3112 * Returns 1 if successful and 0 otherwise.
3114 * Called with both runqueues locked.
3116 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3117 struct sched_domain *sd, enum cpu_idle_type idle)
3119 const struct sched_class *class;
3121 for (class = sched_class_highest; class; class = class->next)
3122 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3129 * find_busiest_group finds and returns the busiest CPU group within the
3130 * domain. It calculates and returns the amount of weighted load which
3131 * should be moved to restore balance via the imbalance parameter.
3133 static struct sched_group *
3134 find_busiest_group(struct sched_domain *sd, int this_cpu,
3135 unsigned long *imbalance, enum cpu_idle_type idle,
3136 int *sd_idle, const cpumask_t *cpus, int *balance)
3138 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3139 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3140 unsigned long max_pull;
3141 unsigned long busiest_load_per_task, busiest_nr_running;
3142 unsigned long this_load_per_task, this_nr_running;
3143 int load_idx, group_imb = 0;
3144 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3145 int power_savings_balance = 1;
3146 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3147 unsigned long min_nr_running = ULONG_MAX;
3148 struct sched_group *group_min = NULL, *group_leader = NULL;
3151 max_load = this_load = total_load = total_pwr = 0;
3152 busiest_load_per_task = busiest_nr_running = 0;
3153 this_load_per_task = this_nr_running = 0;
3155 if (idle == CPU_NOT_IDLE)
3156 load_idx = sd->busy_idx;
3157 else if (idle == CPU_NEWLY_IDLE)
3158 load_idx = sd->newidle_idx;
3160 load_idx = sd->idle_idx;
3163 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3166 int __group_imb = 0;
3167 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3168 unsigned long sum_nr_running, sum_weighted_load;
3169 unsigned long sum_avg_load_per_task;
3170 unsigned long avg_load_per_task;
3172 local_group = cpu_isset(this_cpu, group->cpumask);
3175 balance_cpu = first_cpu(group->cpumask);
3177 /* Tally up the load of all CPUs in the group */
3178 sum_weighted_load = sum_nr_running = avg_load = 0;
3179 sum_avg_load_per_task = avg_load_per_task = 0;
3182 min_cpu_load = ~0UL;
3184 for_each_cpu_mask_nr(i, group->cpumask) {
3187 if (!cpu_isset(i, *cpus))
3192 if (*sd_idle && rq->nr_running)
3195 /* Bias balancing toward cpus of our domain */
3197 if (idle_cpu(i) && !first_idle_cpu) {
3202 load = target_load(i, load_idx);
3204 load = source_load(i, load_idx);
3205 if (load > max_cpu_load)
3206 max_cpu_load = load;
3207 if (min_cpu_load > load)
3208 min_cpu_load = load;
3212 sum_nr_running += rq->nr_running;
3213 sum_weighted_load += weighted_cpuload(i);
3215 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3219 * First idle cpu or the first cpu(busiest) in this sched group
3220 * is eligible for doing load balancing at this and above
3221 * domains. In the newly idle case, we will allow all the cpu's
3222 * to do the newly idle load balance.
3224 if (idle != CPU_NEWLY_IDLE && local_group &&
3225 balance_cpu != this_cpu && balance) {
3230 total_load += avg_load;
3231 total_pwr += group->__cpu_power;
3233 /* Adjust by relative CPU power of the group */
3234 avg_load = sg_div_cpu_power(group,
3235 avg_load * SCHED_LOAD_SCALE);
3239 * Consider the group unbalanced when the imbalance is larger
3240 * than the average weight of two tasks.
3242 * APZ: with cgroup the avg task weight can vary wildly and
3243 * might not be a suitable number - should we keep a
3244 * normalized nr_running number somewhere that negates
3247 avg_load_per_task = sg_div_cpu_power(group,
3248 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3250 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3253 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3256 this_load = avg_load;
3258 this_nr_running = sum_nr_running;
3259 this_load_per_task = sum_weighted_load;
3260 } else if (avg_load > max_load &&
3261 (sum_nr_running > group_capacity || __group_imb)) {
3262 max_load = avg_load;
3264 busiest_nr_running = sum_nr_running;
3265 busiest_load_per_task = sum_weighted_load;
3266 group_imb = __group_imb;
3269 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3271 * Busy processors will not participate in power savings
3274 if (idle == CPU_NOT_IDLE ||
3275 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3279 * If the local group is idle or completely loaded
3280 * no need to do power savings balance at this domain
3282 if (local_group && (this_nr_running >= group_capacity ||
3284 power_savings_balance = 0;
3287 * If a group is already running at full capacity or idle,
3288 * don't include that group in power savings calculations
3290 if (!power_savings_balance || sum_nr_running >= group_capacity
3295 * Calculate the group which has the least non-idle load.
3296 * This is the group from where we need to pick up the load
3299 if ((sum_nr_running < min_nr_running) ||
3300 (sum_nr_running == min_nr_running &&
3301 first_cpu(group->cpumask) <
3302 first_cpu(group_min->cpumask))) {
3304 min_nr_running = sum_nr_running;
3305 min_load_per_task = sum_weighted_load /
3310 * Calculate the group which is almost near its
3311 * capacity but still has some space to pick up some load
3312 * from other group and save more power
3314 if (sum_nr_running <= group_capacity - 1) {
3315 if (sum_nr_running > leader_nr_running ||
3316 (sum_nr_running == leader_nr_running &&
3317 first_cpu(group->cpumask) >
3318 first_cpu(group_leader->cpumask))) {
3319 group_leader = group;
3320 leader_nr_running = sum_nr_running;
3325 group = group->next;
3326 } while (group != sd->groups);
3328 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3331 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3333 if (this_load >= avg_load ||
3334 100*max_load <= sd->imbalance_pct*this_load)
3337 busiest_load_per_task /= busiest_nr_running;
3339 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3342 * We're trying to get all the cpus to the average_load, so we don't
3343 * want to push ourselves above the average load, nor do we wish to
3344 * reduce the max loaded cpu below the average load, as either of these
3345 * actions would just result in more rebalancing later, and ping-pong
3346 * tasks around. Thus we look for the minimum possible imbalance.
3347 * Negative imbalances (*we* are more loaded than anyone else) will
3348 * be counted as no imbalance for these purposes -- we can't fix that
3349 * by pulling tasks to us. Be careful of negative numbers as they'll
3350 * appear as very large values with unsigned longs.
3352 if (max_load <= busiest_load_per_task)
3356 * In the presence of smp nice balancing, certain scenarios can have
3357 * max load less than avg load(as we skip the groups at or below
3358 * its cpu_power, while calculating max_load..)
3360 if (max_load < avg_load) {
3362 goto small_imbalance;
3365 /* Don't want to pull so many tasks that a group would go idle */
3366 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3368 /* How much load to actually move to equalise the imbalance */
3369 *imbalance = min(max_pull * busiest->__cpu_power,
3370 (avg_load - this_load) * this->__cpu_power)
3374 * if *imbalance is less than the average load per runnable task
3375 * there is no gaurantee that any tasks will be moved so we'll have
3376 * a think about bumping its value to force at least one task to be
3379 if (*imbalance < busiest_load_per_task) {
3380 unsigned long tmp, pwr_now, pwr_move;
3384 pwr_move = pwr_now = 0;
3386 if (this_nr_running) {
3387 this_load_per_task /= this_nr_running;
3388 if (busiest_load_per_task > this_load_per_task)
3391 this_load_per_task = cpu_avg_load_per_task(this_cpu);
3393 if (max_load - this_load + busiest_load_per_task >=
3394 busiest_load_per_task * imbn) {
3395 *imbalance = busiest_load_per_task;
3400 * OK, we don't have enough imbalance to justify moving tasks,
3401 * however we may be able to increase total CPU power used by
3405 pwr_now += busiest->__cpu_power *
3406 min(busiest_load_per_task, max_load);
3407 pwr_now += this->__cpu_power *
3408 min(this_load_per_task, this_load);
3409 pwr_now /= SCHED_LOAD_SCALE;
3411 /* Amount of load we'd subtract */
3412 tmp = sg_div_cpu_power(busiest,
3413 busiest_load_per_task * SCHED_LOAD_SCALE);
3415 pwr_move += busiest->__cpu_power *
3416 min(busiest_load_per_task, max_load - tmp);
3418 /* Amount of load we'd add */
3419 if (max_load * busiest->__cpu_power <
3420 busiest_load_per_task * SCHED_LOAD_SCALE)
3421 tmp = sg_div_cpu_power(this,
3422 max_load * busiest->__cpu_power);
3424 tmp = sg_div_cpu_power(this,
3425 busiest_load_per_task * SCHED_LOAD_SCALE);
3426 pwr_move += this->__cpu_power *
3427 min(this_load_per_task, this_load + tmp);
3428 pwr_move /= SCHED_LOAD_SCALE;
3430 /* Move if we gain throughput */
3431 if (pwr_move > pwr_now)
3432 *imbalance = busiest_load_per_task;
3438 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3439 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3442 if (this == group_leader && group_leader != group_min) {
3443 *imbalance = min_load_per_task;
3453 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3456 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3457 unsigned long imbalance, const cpumask_t *cpus)
3459 struct rq *busiest = NULL, *rq;
3460 unsigned long max_load = 0;
3463 for_each_cpu_mask_nr(i, group->cpumask) {
3466 if (!cpu_isset(i, *cpus))
3470 wl = weighted_cpuload(i);
3472 if (rq->nr_running == 1 && wl > imbalance)
3475 if (wl > max_load) {
3485 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3486 * so long as it is large enough.
3488 #define MAX_PINNED_INTERVAL 512
3491 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3492 * tasks if there is an imbalance.
3494 static int load_balance(int this_cpu, struct rq *this_rq,
3495 struct sched_domain *sd, enum cpu_idle_type idle,
3496 int *balance, cpumask_t *cpus)
3498 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3499 struct sched_group *group;
3500 unsigned long imbalance;
3502 unsigned long flags;
3507 * When power savings policy is enabled for the parent domain, idle
3508 * sibling can pick up load irrespective of busy siblings. In this case,
3509 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3510 * portraying it as CPU_NOT_IDLE.
3512 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3513 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3516 schedstat_inc(sd, lb_count[idle]);
3520 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3527 schedstat_inc(sd, lb_nobusyg[idle]);
3531 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3533 schedstat_inc(sd, lb_nobusyq[idle]);
3537 BUG_ON(busiest == this_rq);
3539 schedstat_add(sd, lb_imbalance[idle], imbalance);
3542 if (busiest->nr_running > 1) {
3544 * Attempt to move tasks. If find_busiest_group has found
3545 * an imbalance but busiest->nr_running <= 1, the group is
3546 * still unbalanced. ld_moved simply stays zero, so it is
3547 * correctly treated as an imbalance.
3549 local_irq_save(flags);
3550 double_rq_lock(this_rq, busiest);
3551 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3552 imbalance, sd, idle, &all_pinned);
3553 double_rq_unlock(this_rq, busiest);
3554 local_irq_restore(flags);
3557 * some other cpu did the load balance for us.
3559 if (ld_moved && this_cpu != smp_processor_id())
3560 resched_cpu(this_cpu);
3562 /* All tasks on this runqueue were pinned by CPU affinity */
3563 if (unlikely(all_pinned)) {
3564 cpu_clear(cpu_of(busiest), *cpus);
3565 if (!cpus_empty(*cpus))
3572 schedstat_inc(sd, lb_failed[idle]);
3573 sd->nr_balance_failed++;
3575 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3577 spin_lock_irqsave(&busiest->lock, flags);
3579 /* don't kick the migration_thread, if the curr
3580 * task on busiest cpu can't be moved to this_cpu
3582 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3583 spin_unlock_irqrestore(&busiest->lock, flags);
3585 goto out_one_pinned;
3588 if (!busiest->active_balance) {
3589 busiest->active_balance = 1;
3590 busiest->push_cpu = this_cpu;
3593 spin_unlock_irqrestore(&busiest->lock, flags);
3595 wake_up_process(busiest->migration_thread);
3598 * We've kicked active balancing, reset the failure
3601 sd->nr_balance_failed = sd->cache_nice_tries+1;
3604 sd->nr_balance_failed = 0;
3606 if (likely(!active_balance)) {
3607 /* We were unbalanced, so reset the balancing interval */
3608 sd->balance_interval = sd->min_interval;
3611 * If we've begun active balancing, start to back off. This
3612 * case may not be covered by the all_pinned logic if there
3613 * is only 1 task on the busy runqueue (because we don't call
3616 if (sd->balance_interval < sd->max_interval)
3617 sd->balance_interval *= 2;
3620 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3621 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3627 schedstat_inc(sd, lb_balanced[idle]);
3629 sd->nr_balance_failed = 0;
3632 /* tune up the balancing interval */
3633 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3634 (sd->balance_interval < sd->max_interval))
3635 sd->balance_interval *= 2;
3637 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3638 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3649 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3650 * tasks if there is an imbalance.
3652 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3653 * this_rq is locked.
3656 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3659 struct sched_group *group;
3660 struct rq *busiest = NULL;
3661 unsigned long imbalance;
3669 * When power savings policy is enabled for the parent domain, idle
3670 * sibling can pick up load irrespective of busy siblings. In this case,
3671 * let the state of idle sibling percolate up as IDLE, instead of
3672 * portraying it as CPU_NOT_IDLE.
3674 if (sd->flags & SD_SHARE_CPUPOWER &&
3675 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3678 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3680 update_shares_locked(this_rq, sd);
3681 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3682 &sd_idle, cpus, NULL);
3684 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3688 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3690 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3694 BUG_ON(busiest == this_rq);
3696 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3699 if (busiest->nr_running > 1) {
3700 /* Attempt to move tasks */
3701 double_lock_balance(this_rq, busiest);
3702 /* this_rq->clock is already updated */
3703 update_rq_clock(busiest);
3704 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3705 imbalance, sd, CPU_NEWLY_IDLE,
3707 double_unlock_balance(this_rq, busiest);
3709 if (unlikely(all_pinned)) {
3710 cpu_clear(cpu_of(busiest), *cpus);
3711 if (!cpus_empty(*cpus))
3717 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3718 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3719 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3722 sd->nr_balance_failed = 0;
3724 update_shares_locked(this_rq, sd);
3728 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3729 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3730 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3732 sd->nr_balance_failed = 0;
3738 * idle_balance is called by schedule() if this_cpu is about to become
3739 * idle. Attempts to pull tasks from other CPUs.
3741 static void idle_balance(int this_cpu, struct rq *this_rq)
3743 struct sched_domain *sd;
3744 int pulled_task = 0;
3745 unsigned long next_balance = jiffies + HZ;
3748 for_each_domain(this_cpu, sd) {
3749 unsigned long interval;
3751 if (!(sd->flags & SD_LOAD_BALANCE))
3754 if (sd->flags & SD_BALANCE_NEWIDLE)
3755 /* If we've pulled tasks over stop searching: */
3756 pulled_task = load_balance_newidle(this_cpu, this_rq,
3759 interval = msecs_to_jiffies(sd->balance_interval);
3760 if (time_after(next_balance, sd->last_balance + interval))
3761 next_balance = sd->last_balance + interval;
3765 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3767 * We are going idle. next_balance may be set based on
3768 * a busy processor. So reset next_balance.
3770 this_rq->next_balance = next_balance;
3775 * active_load_balance is run by migration threads. It pushes running tasks
3776 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3777 * running on each physical CPU where possible, and avoids physical /
3778 * logical imbalances.
3780 * Called with busiest_rq locked.
3782 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3784 int target_cpu = busiest_rq->push_cpu;
3785 struct sched_domain *sd;
3786 struct rq *target_rq;
3788 /* Is there any task to move? */
3789 if (busiest_rq->nr_running <= 1)
3792 target_rq = cpu_rq(target_cpu);
3795 * This condition is "impossible", if it occurs
3796 * we need to fix it. Originally reported by
3797 * Bjorn Helgaas on a 128-cpu setup.
3799 BUG_ON(busiest_rq == target_rq);
3801 /* move a task from busiest_rq to target_rq */
3802 double_lock_balance(busiest_rq, target_rq);
3803 update_rq_clock(busiest_rq);
3804 update_rq_clock(target_rq);
3806 /* Search for an sd spanning us and the target CPU. */
3807 for_each_domain(target_cpu, sd) {
3808 if ((sd->flags & SD_LOAD_BALANCE) &&
3809 cpu_isset(busiest_cpu, sd->span))
3814 schedstat_inc(sd, alb_count);
3816 if (move_one_task(target_rq, target_cpu, busiest_rq,
3818 schedstat_inc(sd, alb_pushed);
3820 schedstat_inc(sd, alb_failed);
3822 double_unlock_balance(busiest_rq, target_rq);
3827 atomic_t load_balancer;
3829 } nohz ____cacheline_aligned = {
3830 .load_balancer = ATOMIC_INIT(-1),
3831 .cpu_mask = CPU_MASK_NONE,
3835 * This routine will try to nominate the ilb (idle load balancing)
3836 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3837 * load balancing on behalf of all those cpus. If all the cpus in the system
3838 * go into this tickless mode, then there will be no ilb owner (as there is
3839 * no need for one) and all the cpus will sleep till the next wakeup event
3842 * For the ilb owner, tick is not stopped. And this tick will be used
3843 * for idle load balancing. ilb owner will still be part of
3846 * While stopping the tick, this cpu will become the ilb owner if there
3847 * is no other owner. And will be the owner till that cpu becomes busy
3848 * or if all cpus in the system stop their ticks at which point
3849 * there is no need for ilb owner.
3851 * When the ilb owner becomes busy, it nominates another owner, during the
3852 * next busy scheduler_tick()
3854 int select_nohz_load_balancer(int stop_tick)
3856 int cpu = smp_processor_id();
3859 cpu_set(cpu, nohz.cpu_mask);
3860 cpu_rq(cpu)->in_nohz_recently = 1;
3863 * If we are going offline and still the leader, give up!
3865 if (!cpu_active(cpu) &&
3866 atomic_read(&nohz.load_balancer) == cpu) {
3867 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3872 /* time for ilb owner also to sleep */
3873 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3874 if (atomic_read(&nohz.load_balancer) == cpu)
3875 atomic_set(&nohz.load_balancer, -1);
3879 if (atomic_read(&nohz.load_balancer) == -1) {
3880 /* make me the ilb owner */
3881 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3883 } else if (atomic_read(&nohz.load_balancer) == cpu)
3886 if (!cpu_isset(cpu, nohz.cpu_mask))
3889 cpu_clear(cpu, nohz.cpu_mask);
3891 if (atomic_read(&nohz.load_balancer) == cpu)
3892 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3899 static DEFINE_SPINLOCK(balancing);
3902 * It checks each scheduling domain to see if it is due to be balanced,
3903 * and initiates a balancing operation if so.
3905 * Balancing parameters are set up in arch_init_sched_domains.
3907 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3910 struct rq *rq = cpu_rq(cpu);
3911 unsigned long interval;
3912 struct sched_domain *sd;
3913 /* Earliest time when we have to do rebalance again */
3914 unsigned long next_balance = jiffies + 60*HZ;
3915 int update_next_balance = 0;
3919 for_each_domain(cpu, sd) {
3920 if (!(sd->flags & SD_LOAD_BALANCE))
3923 interval = sd->balance_interval;
3924 if (idle != CPU_IDLE)
3925 interval *= sd->busy_factor;
3927 /* scale ms to jiffies */
3928 interval = msecs_to_jiffies(interval);
3929 if (unlikely(!interval))
3931 if (interval > HZ*NR_CPUS/10)
3932 interval = HZ*NR_CPUS/10;
3934 need_serialize = sd->flags & SD_SERIALIZE;
3936 if (need_serialize) {
3937 if (!spin_trylock(&balancing))
3941 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3942 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3944 * We've pulled tasks over so either we're no
3945 * longer idle, or one of our SMT siblings is
3948 idle = CPU_NOT_IDLE;
3950 sd->last_balance = jiffies;
3953 spin_unlock(&balancing);
3955 if (time_after(next_balance, sd->last_balance + interval)) {
3956 next_balance = sd->last_balance + interval;
3957 update_next_balance = 1;
3961 * Stop the load balance at this level. There is another
3962 * CPU in our sched group which is doing load balancing more
3970 * next_balance will be updated only when there is a need.
3971 * When the cpu is attached to null domain for ex, it will not be
3974 if (likely(update_next_balance))
3975 rq->next_balance = next_balance;
3979 * run_rebalance_domains is triggered when needed from the scheduler tick.
3980 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3981 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3983 static void run_rebalance_domains(struct softirq_action *h)
3985 int this_cpu = smp_processor_id();
3986 struct rq *this_rq = cpu_rq(this_cpu);
3987 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3988 CPU_IDLE : CPU_NOT_IDLE;
3990 rebalance_domains(this_cpu, idle);
3994 * If this cpu is the owner for idle load balancing, then do the
3995 * balancing on behalf of the other idle cpus whose ticks are
3998 if (this_rq->idle_at_tick &&
3999 atomic_read(&nohz.load_balancer) == this_cpu) {
4000 cpumask_t cpus = nohz.cpu_mask;
4004 cpu_clear(this_cpu, cpus);
4005 for_each_cpu_mask_nr(balance_cpu, cpus) {
4007 * If this cpu gets work to do, stop the load balancing
4008 * work being done for other cpus. Next load
4009 * balancing owner will pick it up.
4014 rebalance_domains(balance_cpu, CPU_IDLE);
4016 rq = cpu_rq(balance_cpu);
4017 if (time_after(this_rq->next_balance, rq->next_balance))
4018 this_rq->next_balance = rq->next_balance;
4025 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4027 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4028 * idle load balancing owner or decide to stop the periodic load balancing,
4029 * if the whole system is idle.
4031 static inline void trigger_load_balance(struct rq *rq, int cpu)
4035 * If we were in the nohz mode recently and busy at the current
4036 * scheduler tick, then check if we need to nominate new idle
4039 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4040 rq->in_nohz_recently = 0;
4042 if (atomic_read(&nohz.load_balancer) == cpu) {
4043 cpu_clear(cpu, nohz.cpu_mask);
4044 atomic_set(&nohz.load_balancer, -1);
4047 if (atomic_read(&nohz.load_balancer) == -1) {
4049 * simple selection for now: Nominate the
4050 * first cpu in the nohz list to be the next
4053 * TBD: Traverse the sched domains and nominate
4054 * the nearest cpu in the nohz.cpu_mask.
4056 int ilb = first_cpu(nohz.cpu_mask);
4058 if (ilb < nr_cpu_ids)
4064 * If this cpu is idle and doing idle load balancing for all the
4065 * cpus with ticks stopped, is it time for that to stop?
4067 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4068 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4074 * If this cpu is idle and the idle load balancing is done by
4075 * someone else, then no need raise the SCHED_SOFTIRQ
4077 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4078 cpu_isset(cpu, nohz.cpu_mask))
4081 if (time_after_eq(jiffies, rq->next_balance))
4082 raise_softirq(SCHED_SOFTIRQ);
4085 #else /* CONFIG_SMP */
4088 * on UP we do not need to balance between CPUs:
4090 static inline void idle_balance(int cpu, struct rq *rq)
4096 DEFINE_PER_CPU(struct kernel_stat, kstat);
4098 EXPORT_PER_CPU_SYMBOL(kstat);
4101 * Return any ns on the sched_clock that have not yet been banked in
4102 * @p in case that task is currently running.
4104 unsigned long long __task_delta_exec(struct task_struct *p, int update)
4110 WARN_ON_ONCE(!runqueue_is_locked());
4111 WARN_ON_ONCE(!task_current(rq, p));
4114 update_rq_clock(rq);
4116 delta_exec = rq->clock - p->se.exec_start;
4118 WARN_ON_ONCE(delta_exec < 0);
4124 * Return any ns on the sched_clock that have not yet been banked in
4125 * @p in case that task is currently running.
4127 unsigned long long task_delta_exec(struct task_struct *p)
4129 unsigned long flags;
4133 rq = task_rq_lock(p, &flags);
4135 if (task_current(rq, p)) {
4138 update_rq_clock(rq);
4139 delta_exec = rq->clock - p->se.exec_start;
4140 if ((s64)delta_exec > 0)
4144 task_rq_unlock(rq, &flags);
4150 * Account user cpu time to a process.
4151 * @p: the process that the cpu time gets accounted to
4152 * @cputime: the cpu time spent in user space since the last update
4154 void account_user_time(struct task_struct *p, cputime_t cputime)
4156 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4159 p->utime = cputime_add(p->utime, cputime);
4160 account_group_user_time(p, cputime);
4162 /* Add user time to cpustat. */
4163 tmp = cputime_to_cputime64(cputime);
4164 if (TASK_NICE(p) > 0)
4165 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4167 cpustat->user = cputime64_add(cpustat->user, tmp);
4168 /* Account for user time used */
4169 acct_update_integrals(p);
4173 * Account guest cpu time to a process.
4174 * @p: the process that the cpu time gets accounted to
4175 * @cputime: the cpu time spent in virtual machine since the last update
4177 static void account_guest_time(struct task_struct *p, cputime_t cputime)
4180 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4182 tmp = cputime_to_cputime64(cputime);
4184 p->utime = cputime_add(p->utime, cputime);
4185 account_group_user_time(p, cputime);
4186 p->gtime = cputime_add(p->gtime, cputime);
4188 cpustat->user = cputime64_add(cpustat->user, tmp);
4189 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4193 * Account scaled user cpu time to a process.
4194 * @p: the process that the cpu time gets accounted to
4195 * @cputime: the cpu time spent in user space since the last update
4197 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4199 p->utimescaled = cputime_add(p->utimescaled, cputime);
4203 * Account system cpu time to a process.
4204 * @p: the process that the cpu time gets accounted to
4205 * @hardirq_offset: the offset to subtract from hardirq_count()
4206 * @cputime: the cpu time spent in kernel space since the last update
4208 void account_system_time(struct task_struct *p, int hardirq_offset,
4211 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4212 struct rq *rq = this_rq();
4215 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4216 account_guest_time(p, cputime);
4220 p->stime = cputime_add(p->stime, cputime);
4221 account_group_system_time(p, cputime);
4223 /* Add system time to cpustat. */
4224 tmp = cputime_to_cputime64(cputime);
4225 if (hardirq_count() - hardirq_offset)
4226 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4227 else if (softirq_count())
4228 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4229 else if (p != rq->idle)
4230 cpustat->system = cputime64_add(cpustat->system, tmp);
4231 else if (atomic_read(&rq->nr_iowait) > 0)
4232 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4234 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4235 /* Account for system time used */
4236 acct_update_integrals(p);
4240 * Account scaled system cpu time to a process.
4241 * @p: the process that the cpu time gets accounted to
4242 * @hardirq_offset: the offset to subtract from hardirq_count()
4243 * @cputime: the cpu time spent in kernel space since the last update
4245 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4247 p->stimescaled = cputime_add(p->stimescaled, cputime);
4251 * Account for involuntary wait time.
4252 * @p: the process from which the cpu time has been stolen
4253 * @steal: the cpu time spent in involuntary wait
4255 void account_steal_time(struct task_struct *p, cputime_t steal)
4257 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4258 cputime64_t tmp = cputime_to_cputime64(steal);
4259 struct rq *rq = this_rq();
4261 if (p == rq->idle) {
4262 p->stime = cputime_add(p->stime, steal);
4263 account_group_system_time(p, steal);
4264 if (atomic_read(&rq->nr_iowait) > 0)
4265 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4267 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4269 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4273 * Use precise platform statistics if available:
4275 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4276 cputime_t task_utime(struct task_struct *p)
4281 cputime_t task_stime(struct task_struct *p)
4286 cputime_t task_utime(struct task_struct *p)
4288 clock_t utime = cputime_to_clock_t(p->utime),
4289 total = utime + cputime_to_clock_t(p->stime);
4293 * Use CFS's precise accounting:
4295 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
4299 do_div(temp, total);
4301 utime = (clock_t)temp;
4303 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
4304 return p->prev_utime;
4307 cputime_t task_stime(struct task_struct *p)
4312 * Use CFS's precise accounting. (we subtract utime from
4313 * the total, to make sure the total observed by userspace
4314 * grows monotonically - apps rely on that):
4316 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
4317 cputime_to_clock_t(task_utime(p));
4320 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
4322 return p->prev_stime;
4326 inline cputime_t task_gtime(struct task_struct *p)
4332 * This function gets called by the timer code, with HZ frequency.
4333 * We call it with interrupts disabled.
4335 * It also gets called by the fork code, when changing the parent's
4338 void scheduler_tick(void)
4340 int cpu = smp_processor_id();
4341 struct rq *rq = cpu_rq(cpu);
4342 struct task_struct *curr = rq->curr;
4346 spin_lock(&rq->lock);
4347 update_rq_clock(rq);
4348 update_cpu_load(rq);
4349 curr->sched_class->task_tick(rq, curr, 0);
4350 perf_counter_task_tick(curr, cpu);
4351 spin_unlock(&rq->lock);
4354 rq->idle_at_tick = idle_cpu(cpu);
4355 trigger_load_balance(rq, cpu);
4359 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4360 defined(CONFIG_PREEMPT_TRACER))
4362 static inline unsigned long get_parent_ip(unsigned long addr)
4364 if (in_lock_functions(addr)) {
4365 addr = CALLER_ADDR2;
4366 if (in_lock_functions(addr))
4367 addr = CALLER_ADDR3;
4372 void __kprobes add_preempt_count(int val)
4374 #ifdef CONFIG_DEBUG_PREEMPT
4378 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4381 preempt_count() += val;
4382 #ifdef CONFIG_DEBUG_PREEMPT
4384 * Spinlock count overflowing soon?
4386 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4389 if (preempt_count() == val)
4390 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4392 EXPORT_SYMBOL(add_preempt_count);
4394 void __kprobes sub_preempt_count(int val)
4396 #ifdef CONFIG_DEBUG_PREEMPT
4400 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4403 * Is the spinlock portion underflowing?
4405 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4406 !(preempt_count() & PREEMPT_MASK)))
4410 if (preempt_count() == val)
4411 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4412 preempt_count() -= val;
4414 EXPORT_SYMBOL(sub_preempt_count);
4419 * Print scheduling while atomic bug:
4421 static noinline void __schedule_bug(struct task_struct *prev)
4423 struct pt_regs *regs = get_irq_regs();
4425 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4426 prev->comm, prev->pid, preempt_count());
4428 debug_show_held_locks(prev);
4430 if (irqs_disabled())
4431 print_irqtrace_events(prev);
4440 * Various schedule()-time debugging checks and statistics:
4442 static inline void schedule_debug(struct task_struct *prev)
4445 * Test if we are atomic. Since do_exit() needs to call into
4446 * schedule() atomically, we ignore that path for now.
4447 * Otherwise, whine if we are scheduling when we should not be.
4449 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4450 __schedule_bug(prev);
4452 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4454 schedstat_inc(this_rq(), sched_count);
4455 #ifdef CONFIG_SCHEDSTATS
4456 if (unlikely(prev->lock_depth >= 0)) {
4457 schedstat_inc(this_rq(), bkl_count);
4458 schedstat_inc(prev, sched_info.bkl_count);
4464 * Pick up the highest-prio task:
4466 static inline struct task_struct *
4467 pick_next_task(struct rq *rq, struct task_struct *prev)
4469 const struct sched_class *class;
4470 struct task_struct *p;
4473 * Optimization: we know that if all tasks are in
4474 * the fair class we can call that function directly:
4476 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4477 p = fair_sched_class.pick_next_task(rq);
4482 class = sched_class_highest;
4484 p = class->pick_next_task(rq);
4488 * Will never be NULL as the idle class always
4489 * returns a non-NULL p:
4491 class = class->next;
4496 * schedule() is the main scheduler function.
4498 asmlinkage void __sched schedule(void)
4500 struct task_struct *prev, *next;
4501 unsigned long *switch_count;
4507 cpu = smp_processor_id();
4511 switch_count = &prev->nivcsw;
4513 release_kernel_lock(prev);
4514 need_resched_nonpreemptible:
4516 schedule_debug(prev);
4518 if (sched_feat(HRTICK))
4521 spin_lock_irq(&rq->lock);
4522 update_rq_clock(rq);
4523 clear_tsk_need_resched(prev);
4525 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4526 if (unlikely(signal_pending_state(prev->state, prev)))
4527 prev->state = TASK_RUNNING;
4529 deactivate_task(rq, prev, 1);
4530 switch_count = &prev->nvcsw;
4534 if (prev->sched_class->pre_schedule)
4535 prev->sched_class->pre_schedule(rq, prev);
4538 if (unlikely(!rq->nr_running))
4539 idle_balance(cpu, rq);
4541 prev->sched_class->put_prev_task(rq, prev);
4542 next = pick_next_task(rq, prev);
4544 if (likely(prev != next)) {
4545 sched_info_switch(prev, next);
4546 perf_counter_task_sched_out(prev, cpu);
4552 context_switch(rq, prev, next); /* unlocks the rq */
4554 * the context switch might have flipped the stack from under
4555 * us, hence refresh the local variables.
4557 cpu = smp_processor_id();
4560 spin_unlock_irq(&rq->lock);
4562 if (unlikely(reacquire_kernel_lock(current) < 0))
4563 goto need_resched_nonpreemptible;
4565 preempt_enable_no_resched();
4566 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4569 EXPORT_SYMBOL(schedule);
4571 #ifdef CONFIG_PREEMPT
4573 * this is the entry point to schedule() from in-kernel preemption
4574 * off of preempt_enable. Kernel preemptions off return from interrupt
4575 * occur there and call schedule directly.
4577 asmlinkage void __sched preempt_schedule(void)
4579 struct thread_info *ti = current_thread_info();
4582 * If there is a non-zero preempt_count or interrupts are disabled,
4583 * we do not want to preempt the current task. Just return..
4585 if (likely(ti->preempt_count || irqs_disabled()))
4589 add_preempt_count(PREEMPT_ACTIVE);
4591 sub_preempt_count(PREEMPT_ACTIVE);
4594 * Check again in case we missed a preemption opportunity
4595 * between schedule and now.
4598 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4600 EXPORT_SYMBOL(preempt_schedule);
4603 * this is the entry point to schedule() from kernel preemption
4604 * off of irq context.
4605 * Note, that this is called and return with irqs disabled. This will
4606 * protect us against recursive calling from irq.
4608 asmlinkage void __sched preempt_schedule_irq(void)
4610 struct thread_info *ti = current_thread_info();
4612 /* Catch callers which need to be fixed */
4613 BUG_ON(ti->preempt_count || !irqs_disabled());
4616 add_preempt_count(PREEMPT_ACTIVE);
4619 local_irq_disable();
4620 sub_preempt_count(PREEMPT_ACTIVE);
4623 * Check again in case we missed a preemption opportunity
4624 * between schedule and now.
4627 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4630 #endif /* CONFIG_PREEMPT */
4632 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4635 return try_to_wake_up(curr->private, mode, sync);
4637 EXPORT_SYMBOL(default_wake_function);
4640 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4641 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4642 * number) then we wake all the non-exclusive tasks and one exclusive task.
4644 * There are circumstances in which we can try to wake a task which has already
4645 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4646 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4648 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4649 int nr_exclusive, int sync, void *key)
4651 wait_queue_t *curr, *next;
4653 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4654 unsigned flags = curr->flags;
4656 if (curr->func(curr, mode, sync, key) &&
4657 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4663 * __wake_up - wake up threads blocked on a waitqueue.
4665 * @mode: which threads
4666 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4667 * @key: is directly passed to the wakeup function
4669 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4670 int nr_exclusive, void *key)
4672 unsigned long flags;
4674 spin_lock_irqsave(&q->lock, flags);
4675 __wake_up_common(q, mode, nr_exclusive, 0, key);
4676 spin_unlock_irqrestore(&q->lock, flags);
4678 EXPORT_SYMBOL(__wake_up);
4681 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4683 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4685 __wake_up_common(q, mode, 1, 0, NULL);
4689 * __wake_up_sync - wake up threads blocked on a waitqueue.
4691 * @mode: which threads
4692 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4694 * The sync wakeup differs that the waker knows that it will schedule
4695 * away soon, so while the target thread will be woken up, it will not
4696 * be migrated to another CPU - ie. the two threads are 'synchronized'
4697 * with each other. This can prevent needless bouncing between CPUs.
4699 * On UP it can prevent extra preemption.
4702 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4704 unsigned long flags;
4710 if (unlikely(!nr_exclusive))
4713 spin_lock_irqsave(&q->lock, flags);
4714 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4715 spin_unlock_irqrestore(&q->lock, flags);
4717 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4720 * complete: - signals a single thread waiting on this completion
4721 * @x: holds the state of this particular completion
4723 * This will wake up a single thread waiting on this completion. Threads will be
4724 * awakened in the same order in which they were queued.
4726 * See also complete_all(), wait_for_completion() and related routines.
4728 void complete(struct completion *x)
4730 unsigned long flags;
4732 spin_lock_irqsave(&x->wait.lock, flags);
4734 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4735 spin_unlock_irqrestore(&x->wait.lock, flags);
4737 EXPORT_SYMBOL(complete);
4740 * complete_all: - signals all threads waiting on this completion
4741 * @x: holds the state of this particular completion
4743 * This will wake up all threads waiting on this particular completion event.
4745 void complete_all(struct completion *x)
4747 unsigned long flags;
4749 spin_lock_irqsave(&x->wait.lock, flags);
4750 x->done += UINT_MAX/2;
4751 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4752 spin_unlock_irqrestore(&x->wait.lock, flags);
4754 EXPORT_SYMBOL(complete_all);
4756 static inline long __sched
4757 do_wait_for_common(struct completion *x, long timeout, int state)
4760 DECLARE_WAITQUEUE(wait, current);
4762 wait.flags |= WQ_FLAG_EXCLUSIVE;
4763 __add_wait_queue_tail(&x->wait, &wait);
4765 if (signal_pending_state(state, current)) {
4766 timeout = -ERESTARTSYS;
4769 __set_current_state(state);
4770 spin_unlock_irq(&x->wait.lock);
4771 timeout = schedule_timeout(timeout);
4772 spin_lock_irq(&x->wait.lock);
4773 } while (!x->done && timeout);
4774 __remove_wait_queue(&x->wait, &wait);
4779 return timeout ?: 1;
4783 wait_for_common(struct completion *x, long timeout, int state)
4787 spin_lock_irq(&x->wait.lock);
4788 timeout = do_wait_for_common(x, timeout, state);
4789 spin_unlock_irq(&x->wait.lock);
4794 * wait_for_completion: - waits for completion of a task
4795 * @x: holds the state of this particular completion
4797 * This waits to be signaled for completion of a specific task. It is NOT
4798 * interruptible and there is no timeout.
4800 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4801 * and interrupt capability. Also see complete().
4803 void __sched wait_for_completion(struct completion *x)
4805 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4807 EXPORT_SYMBOL(wait_for_completion);
4810 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4811 * @x: holds the state of this particular completion
4812 * @timeout: timeout value in jiffies
4814 * This waits for either a completion of a specific task to be signaled or for a
4815 * specified timeout to expire. The timeout is in jiffies. It is not
4818 unsigned long __sched
4819 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4821 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4823 EXPORT_SYMBOL(wait_for_completion_timeout);
4826 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4827 * @x: holds the state of this particular completion
4829 * This waits for completion of a specific task to be signaled. It is
4832 int __sched wait_for_completion_interruptible(struct completion *x)
4834 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4835 if (t == -ERESTARTSYS)
4839 EXPORT_SYMBOL(wait_for_completion_interruptible);
4842 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4843 * @x: holds the state of this particular completion
4844 * @timeout: timeout value in jiffies
4846 * This waits for either a completion of a specific task to be signaled or for a
4847 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4849 unsigned long __sched
4850 wait_for_completion_interruptible_timeout(struct completion *x,
4851 unsigned long timeout)
4853 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4855 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4858 * wait_for_completion_killable: - waits for completion of a task (killable)
4859 * @x: holds the state of this particular completion
4861 * This waits to be signaled for completion of a specific task. It can be
4862 * interrupted by a kill signal.
4864 int __sched wait_for_completion_killable(struct completion *x)
4866 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4867 if (t == -ERESTARTSYS)
4871 EXPORT_SYMBOL(wait_for_completion_killable);
4874 * try_wait_for_completion - try to decrement a completion without blocking
4875 * @x: completion structure
4877 * Returns: 0 if a decrement cannot be done without blocking
4878 * 1 if a decrement succeeded.
4880 * If a completion is being used as a counting completion,
4881 * attempt to decrement the counter without blocking. This
4882 * enables us to avoid waiting if the resource the completion
4883 * is protecting is not available.
4885 bool try_wait_for_completion(struct completion *x)
4889 spin_lock_irq(&x->wait.lock);
4894 spin_unlock_irq(&x->wait.lock);
4897 EXPORT_SYMBOL(try_wait_for_completion);
4900 * completion_done - Test to see if a completion has any waiters
4901 * @x: completion structure
4903 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4904 * 1 if there are no waiters.
4907 bool completion_done(struct completion *x)
4911 spin_lock_irq(&x->wait.lock);
4914 spin_unlock_irq(&x->wait.lock);
4917 EXPORT_SYMBOL(completion_done);
4920 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4922 unsigned long flags;
4925 init_waitqueue_entry(&wait, current);
4927 __set_current_state(state);
4929 spin_lock_irqsave(&q->lock, flags);
4930 __add_wait_queue(q, &wait);
4931 spin_unlock(&q->lock);
4932 timeout = schedule_timeout(timeout);
4933 spin_lock_irq(&q->lock);
4934 __remove_wait_queue(q, &wait);
4935 spin_unlock_irqrestore(&q->lock, flags);
4940 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4942 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4944 EXPORT_SYMBOL(interruptible_sleep_on);
4947 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4949 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4951 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4953 void __sched sleep_on(wait_queue_head_t *q)
4955 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4957 EXPORT_SYMBOL(sleep_on);
4959 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4961 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4963 EXPORT_SYMBOL(sleep_on_timeout);
4965 #ifdef CONFIG_RT_MUTEXES
4968 * rt_mutex_setprio - set the current priority of a task
4970 * @prio: prio value (kernel-internal form)
4972 * This function changes the 'effective' priority of a task. It does
4973 * not touch ->normal_prio like __setscheduler().
4975 * Used by the rt_mutex code to implement priority inheritance logic.
4977 void rt_mutex_setprio(struct task_struct *p, int prio)
4979 unsigned long flags;
4980 int oldprio, on_rq, running;
4982 const struct sched_class *prev_class = p->sched_class;
4984 BUG_ON(prio < 0 || prio > MAX_PRIO);
4986 rq = task_rq_lock(p, &flags);
4987 update_rq_clock(rq);
4990 on_rq = p->se.on_rq;
4991 running = task_current(rq, p);
4993 dequeue_task(rq, p, 0);
4995 p->sched_class->put_prev_task(rq, p);
4998 p->sched_class = &rt_sched_class;
5000 p->sched_class = &fair_sched_class;
5005 p->sched_class->set_curr_task(rq);
5007 enqueue_task(rq, p, 0);
5009 check_class_changed(rq, p, prev_class, oldprio, running);
5011 task_rq_unlock(rq, &flags);
5016 void set_user_nice(struct task_struct *p, long nice)
5018 int old_prio, delta, on_rq;
5019 unsigned long flags;
5022 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5025 * We have to be careful, if called from sys_setpriority(),
5026 * the task might be in the middle of scheduling on another CPU.
5028 rq = task_rq_lock(p, &flags);
5029 update_rq_clock(rq);
5031 * The RT priorities are set via sched_setscheduler(), but we still
5032 * allow the 'normal' nice value to be set - but as expected
5033 * it wont have any effect on scheduling until the task is
5034 * SCHED_FIFO/SCHED_RR:
5036 if (task_has_rt_policy(p)) {
5037 p->static_prio = NICE_TO_PRIO(nice);
5040 on_rq = p->se.on_rq;
5042 dequeue_task(rq, p, 0);
5044 p->static_prio = NICE_TO_PRIO(nice);
5047 p->prio = effective_prio(p);
5048 delta = p->prio - old_prio;
5051 enqueue_task(rq, p, 0);
5053 * If the task increased its priority or is running and
5054 * lowered its priority, then reschedule its CPU:
5056 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5057 resched_task(rq->curr);
5060 task_rq_unlock(rq, &flags);
5062 EXPORT_SYMBOL(set_user_nice);
5065 * can_nice - check if a task can reduce its nice value
5069 int can_nice(const struct task_struct *p, const int nice)
5071 /* convert nice value [19,-20] to rlimit style value [1,40] */
5072 int nice_rlim = 20 - nice;
5074 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
5075 capable(CAP_SYS_NICE));
5078 #ifdef __ARCH_WANT_SYS_NICE
5081 * sys_nice - change the priority of the current process.
5082 * @increment: priority increment
5084 * sys_setpriority is a more generic, but much slower function that
5085 * does similar things.
5087 asmlinkage long sys_nice(int increment)
5092 * Setpriority might change our priority at the same moment.
5093 * We don't have to worry. Conceptually one call occurs first
5094 * and we have a single winner.
5096 if (increment < -40)
5101 nice = PRIO_TO_NICE(current->static_prio) + increment;
5107 if (increment < 0 && !can_nice(current, nice))
5110 retval = security_task_setnice(current, nice);
5114 set_user_nice(current, nice);
5121 * task_prio - return the priority value of a given task.
5122 * @p: the task in question.
5124 * This is the priority value as seen by users in /proc.
5125 * RT tasks are offset by -200. Normal tasks are centered
5126 * around 0, value goes from -16 to +15.
5128 int task_prio(const struct task_struct *p)
5130 return p->prio - MAX_RT_PRIO;
5134 * task_nice - return the nice value of a given task.
5135 * @p: the task in question.
5137 int task_nice(const struct task_struct *p)
5139 return TASK_NICE(p);
5141 EXPORT_SYMBOL(task_nice);
5144 * idle_cpu - is a given cpu idle currently?
5145 * @cpu: the processor in question.
5147 int idle_cpu(int cpu)
5149 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5153 * idle_task - return the idle task for a given cpu.
5154 * @cpu: the processor in question.
5156 struct task_struct *idle_task(int cpu)
5158 return cpu_rq(cpu)->idle;
5162 * find_process_by_pid - find a process with a matching PID value.
5163 * @pid: the pid in question.
5165 static struct task_struct *find_process_by_pid(pid_t pid)
5167 return pid ? find_task_by_vpid(pid) : current;
5170 /* Actually do priority change: must hold rq lock. */
5172 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5174 BUG_ON(p->se.on_rq);
5177 switch (p->policy) {
5181 p->sched_class = &fair_sched_class;
5185 p->sched_class = &rt_sched_class;
5189 p->rt_priority = prio;
5190 p->normal_prio = normal_prio(p);
5191 /* we are holding p->pi_lock already */
5192 p->prio = rt_mutex_getprio(p);
5197 * check the target process has a UID that matches the current process's
5199 static bool check_same_owner(struct task_struct *p)
5201 const struct cred *cred = current_cred(), *pcred;
5205 pcred = __task_cred(p);
5206 match = (cred->euid == pcred->euid ||
5207 cred->euid == pcred->uid);
5212 static int __sched_setscheduler(struct task_struct *p, int policy,
5213 struct sched_param *param, bool user)
5215 int retval, oldprio, oldpolicy = -1, on_rq, running;
5216 unsigned long flags;
5217 const struct sched_class *prev_class = p->sched_class;
5220 /* may grab non-irq protected spin_locks */
5221 BUG_ON(in_interrupt());
5223 /* double check policy once rq lock held */
5225 policy = oldpolicy = p->policy;
5226 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5227 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5228 policy != SCHED_IDLE)
5231 * Valid priorities for SCHED_FIFO and SCHED_RR are
5232 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5233 * SCHED_BATCH and SCHED_IDLE is 0.
5235 if (param->sched_priority < 0 ||
5236 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5237 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5239 if (rt_policy(policy) != (param->sched_priority != 0))
5243 * Allow unprivileged RT tasks to decrease priority:
5245 if (user && !capable(CAP_SYS_NICE)) {
5246 if (rt_policy(policy)) {
5247 unsigned long rlim_rtprio;
5249 if (!lock_task_sighand(p, &flags))
5251 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5252 unlock_task_sighand(p, &flags);
5254 /* can't set/change the rt policy */
5255 if (policy != p->policy && !rlim_rtprio)
5258 /* can't increase priority */
5259 if (param->sched_priority > p->rt_priority &&
5260 param->sched_priority > rlim_rtprio)
5264 * Like positive nice levels, dont allow tasks to
5265 * move out of SCHED_IDLE either:
5267 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5270 /* can't change other user's priorities */
5271 if (!check_same_owner(p))
5276 #ifdef CONFIG_RT_GROUP_SCHED
5278 * Do not allow realtime tasks into groups that have no runtime
5281 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5282 task_group(p)->rt_bandwidth.rt_runtime == 0)
5286 retval = security_task_setscheduler(p, policy, param);
5292 * make sure no PI-waiters arrive (or leave) while we are
5293 * changing the priority of the task:
5295 spin_lock_irqsave(&p->pi_lock, flags);
5297 * To be able to change p->policy safely, the apropriate
5298 * runqueue lock must be held.
5300 rq = __task_rq_lock(p);
5301 /* recheck policy now with rq lock held */
5302 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5303 policy = oldpolicy = -1;
5304 __task_rq_unlock(rq);
5305 spin_unlock_irqrestore(&p->pi_lock, flags);
5308 update_rq_clock(rq);
5309 on_rq = p->se.on_rq;
5310 running = task_current(rq, p);
5312 deactivate_task(rq, p, 0);
5314 p->sched_class->put_prev_task(rq, p);
5317 __setscheduler(rq, p, policy, param->sched_priority);
5320 p->sched_class->set_curr_task(rq);
5322 activate_task(rq, p, 0);
5324 check_class_changed(rq, p, prev_class, oldprio, running);
5326 __task_rq_unlock(rq);
5327 spin_unlock_irqrestore(&p->pi_lock, flags);
5329 rt_mutex_adjust_pi(p);
5335 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5336 * @p: the task in question.
5337 * @policy: new policy.
5338 * @param: structure containing the new RT priority.
5340 * NOTE that the task may be already dead.
5342 int sched_setscheduler(struct task_struct *p, int policy,
5343 struct sched_param *param)
5345 return __sched_setscheduler(p, policy, param, true);
5347 EXPORT_SYMBOL_GPL(sched_setscheduler);
5350 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5351 * @p: the task in question.
5352 * @policy: new policy.
5353 * @param: structure containing the new RT priority.
5355 * Just like sched_setscheduler, only don't bother checking if the
5356 * current context has permission. For example, this is needed in
5357 * stop_machine(): we create temporary high priority worker threads,
5358 * but our caller might not have that capability.
5360 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5361 struct sched_param *param)
5363 return __sched_setscheduler(p, policy, param, false);
5367 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5369 struct sched_param lparam;
5370 struct task_struct *p;
5373 if (!param || pid < 0)
5375 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5380 p = find_process_by_pid(pid);
5382 retval = sched_setscheduler(p, policy, &lparam);
5389 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5390 * @pid: the pid in question.
5391 * @policy: new policy.
5392 * @param: structure containing the new RT priority.
5395 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5397 /* negative values for policy are not valid */
5401 return do_sched_setscheduler(pid, policy, param);
5405 * sys_sched_setparam - set/change the RT priority of a thread
5406 * @pid: the pid in question.
5407 * @param: structure containing the new RT priority.
5409 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5411 return do_sched_setscheduler(pid, -1, param);
5415 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5416 * @pid: the pid in question.
5418 asmlinkage long sys_sched_getscheduler(pid_t pid)
5420 struct task_struct *p;
5427 read_lock(&tasklist_lock);
5428 p = find_process_by_pid(pid);
5430 retval = security_task_getscheduler(p);
5434 read_unlock(&tasklist_lock);
5439 * sys_sched_getscheduler - get the RT priority of a thread
5440 * @pid: the pid in question.
5441 * @param: structure containing the RT priority.
5443 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5445 struct sched_param lp;
5446 struct task_struct *p;
5449 if (!param || pid < 0)
5452 read_lock(&tasklist_lock);
5453 p = find_process_by_pid(pid);
5458 retval = security_task_getscheduler(p);
5462 lp.sched_priority = p->rt_priority;
5463 read_unlock(&tasklist_lock);
5466 * This one might sleep, we cannot do it with a spinlock held ...
5468 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5473 read_unlock(&tasklist_lock);
5477 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5479 cpumask_t cpus_allowed;
5480 cpumask_t new_mask = *in_mask;
5481 struct task_struct *p;
5485 read_lock(&tasklist_lock);
5487 p = find_process_by_pid(pid);
5489 read_unlock(&tasklist_lock);
5495 * It is not safe to call set_cpus_allowed with the
5496 * tasklist_lock held. We will bump the task_struct's
5497 * usage count and then drop tasklist_lock.
5500 read_unlock(&tasklist_lock);
5503 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
5506 retval = security_task_setscheduler(p, 0, NULL);
5510 cpuset_cpus_allowed(p, &cpus_allowed);
5511 cpus_and(new_mask, new_mask, cpus_allowed);
5513 retval = set_cpus_allowed_ptr(p, &new_mask);
5516 cpuset_cpus_allowed(p, &cpus_allowed);
5517 if (!cpus_subset(new_mask, cpus_allowed)) {
5519 * We must have raced with a concurrent cpuset
5520 * update. Just reset the cpus_allowed to the
5521 * cpuset's cpus_allowed
5523 new_mask = cpus_allowed;
5533 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5534 cpumask_t *new_mask)
5536 if (len < sizeof(cpumask_t)) {
5537 memset(new_mask, 0, sizeof(cpumask_t));
5538 } else if (len > sizeof(cpumask_t)) {
5539 len = sizeof(cpumask_t);
5541 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5545 * sys_sched_setaffinity - set the cpu affinity of a process
5546 * @pid: pid of the process
5547 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5548 * @user_mask_ptr: user-space pointer to the new cpu mask
5550 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5551 unsigned long __user *user_mask_ptr)
5556 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5560 return sched_setaffinity(pid, &new_mask);
5563 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5565 struct task_struct *p;
5569 read_lock(&tasklist_lock);
5572 p = find_process_by_pid(pid);
5576 retval = security_task_getscheduler(p);
5580 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5583 read_unlock(&tasklist_lock);
5590 * sys_sched_getaffinity - get the cpu affinity of a process
5591 * @pid: pid of the process
5592 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5593 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5595 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5596 unsigned long __user *user_mask_ptr)
5601 if (len < sizeof(cpumask_t))
5604 ret = sched_getaffinity(pid, &mask);
5608 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5611 return sizeof(cpumask_t);
5615 * sys_sched_yield - yield the current processor to other threads.
5617 * This function yields the current CPU to other tasks. If there are no
5618 * other threads running on this CPU then this function will return.
5620 asmlinkage long sys_sched_yield(void)
5622 struct rq *rq = this_rq_lock();
5624 schedstat_inc(rq, yld_count);
5625 current->sched_class->yield_task(rq);
5628 * Since we are going to call schedule() anyway, there's
5629 * no need to preempt or enable interrupts:
5631 __release(rq->lock);
5632 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5633 _raw_spin_unlock(&rq->lock);
5634 preempt_enable_no_resched();
5641 static void __cond_resched(void)
5643 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5644 __might_sleep(__FILE__, __LINE__);
5647 * The BKS might be reacquired before we have dropped
5648 * PREEMPT_ACTIVE, which could trigger a second
5649 * cond_resched() call.
5652 add_preempt_count(PREEMPT_ACTIVE);
5654 sub_preempt_count(PREEMPT_ACTIVE);
5655 } while (need_resched());
5658 int __sched _cond_resched(void)
5660 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5661 system_state == SYSTEM_RUNNING) {
5667 EXPORT_SYMBOL(_cond_resched);
5670 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5671 * call schedule, and on return reacquire the lock.
5673 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5674 * operations here to prevent schedule() from being called twice (once via
5675 * spin_unlock(), once by hand).
5677 int cond_resched_lock(spinlock_t *lock)
5679 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5682 if (spin_needbreak(lock) || resched) {
5684 if (resched && need_resched())
5693 EXPORT_SYMBOL(cond_resched_lock);
5695 int __sched cond_resched_softirq(void)
5697 BUG_ON(!in_softirq());
5699 if (need_resched() && system_state == SYSTEM_RUNNING) {
5707 EXPORT_SYMBOL(cond_resched_softirq);
5710 * yield - yield the current processor to other threads.
5712 * This is a shortcut for kernel-space yielding - it marks the
5713 * thread runnable and calls sys_sched_yield().
5715 void __sched yield(void)
5717 set_current_state(TASK_RUNNING);
5720 EXPORT_SYMBOL(yield);
5723 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5724 * that process accounting knows that this is a task in IO wait state.
5726 * But don't do that if it is a deliberate, throttling IO wait (this task
5727 * has set its backing_dev_info: the queue against which it should throttle)
5729 void __sched io_schedule(void)
5731 struct rq *rq = &__raw_get_cpu_var(runqueues);
5733 delayacct_blkio_start();
5734 atomic_inc(&rq->nr_iowait);
5736 atomic_dec(&rq->nr_iowait);
5737 delayacct_blkio_end();
5739 EXPORT_SYMBOL(io_schedule);
5741 long __sched io_schedule_timeout(long timeout)
5743 struct rq *rq = &__raw_get_cpu_var(runqueues);
5746 delayacct_blkio_start();
5747 atomic_inc(&rq->nr_iowait);
5748 ret = schedule_timeout(timeout);
5749 atomic_dec(&rq->nr_iowait);
5750 delayacct_blkio_end();
5755 * sys_sched_get_priority_max - return maximum RT priority.
5756 * @policy: scheduling class.
5758 * this syscall returns the maximum rt_priority that can be used
5759 * by a given scheduling class.
5761 asmlinkage long sys_sched_get_priority_max(int policy)
5768 ret = MAX_USER_RT_PRIO-1;
5780 * sys_sched_get_priority_min - return minimum RT priority.
5781 * @policy: scheduling class.
5783 * this syscall returns the minimum rt_priority that can be used
5784 * by a given scheduling class.
5786 asmlinkage long sys_sched_get_priority_min(int policy)
5804 * sys_sched_rr_get_interval - return the default timeslice of a process.
5805 * @pid: pid of the process.
5806 * @interval: userspace pointer to the timeslice value.
5808 * this syscall writes the default timeslice value of a given process
5809 * into the user-space timespec buffer. A value of '0' means infinity.
5812 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5814 struct task_struct *p;
5815 unsigned int time_slice;
5823 read_lock(&tasklist_lock);
5824 p = find_process_by_pid(pid);
5828 retval = security_task_getscheduler(p);
5833 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5834 * tasks that are on an otherwise idle runqueue:
5837 if (p->policy == SCHED_RR) {
5838 time_slice = DEF_TIMESLICE;
5839 } else if (p->policy != SCHED_FIFO) {
5840 struct sched_entity *se = &p->se;
5841 unsigned long flags;
5844 rq = task_rq_lock(p, &flags);
5845 if (rq->cfs.load.weight)
5846 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5847 task_rq_unlock(rq, &flags);
5849 read_unlock(&tasklist_lock);
5850 jiffies_to_timespec(time_slice, &t);
5851 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5855 read_unlock(&tasklist_lock);
5859 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5861 void sched_show_task(struct task_struct *p)
5863 unsigned long free = 0;
5866 state = p->state ? __ffs(p->state) + 1 : 0;
5867 printk(KERN_INFO "%-13.13s %c", p->comm,
5868 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5869 #if BITS_PER_LONG == 32
5870 if (state == TASK_RUNNING)
5871 printk(KERN_CONT " running ");
5873 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5875 if (state == TASK_RUNNING)
5876 printk(KERN_CONT " running task ");
5878 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5880 #ifdef CONFIG_DEBUG_STACK_USAGE
5882 unsigned long *n = end_of_stack(p);
5885 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5888 printk(KERN_CONT "%5lu %5d %6d\n", free,
5889 task_pid_nr(p), task_pid_nr(p->real_parent));
5891 show_stack(p, NULL);
5894 void show_state_filter(unsigned long state_filter)
5896 struct task_struct *g, *p;
5898 #if BITS_PER_LONG == 32
5900 " task PC stack pid father\n");
5903 " task PC stack pid father\n");
5905 read_lock(&tasklist_lock);
5906 do_each_thread(g, p) {
5908 * reset the NMI-timeout, listing all files on a slow
5909 * console might take alot of time:
5911 touch_nmi_watchdog();
5912 if (!state_filter || (p->state & state_filter))
5914 } while_each_thread(g, p);
5916 touch_all_softlockup_watchdogs();
5918 #ifdef CONFIG_SCHED_DEBUG
5919 sysrq_sched_debug_show();
5921 read_unlock(&tasklist_lock);
5923 * Only show locks if all tasks are dumped:
5925 if (state_filter == -1)
5926 debug_show_all_locks();
5929 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5931 idle->sched_class = &idle_sched_class;
5935 * init_idle - set up an idle thread for a given CPU
5936 * @idle: task in question
5937 * @cpu: cpu the idle task belongs to
5939 * NOTE: this function does not set the idle thread's NEED_RESCHED
5940 * flag, to make booting more robust.
5942 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5944 struct rq *rq = cpu_rq(cpu);
5945 unsigned long flags;
5947 spin_lock_irqsave(&rq->lock, flags);
5950 idle->se.exec_start = sched_clock();
5952 idle->prio = idle->normal_prio = MAX_PRIO;
5953 idle->cpus_allowed = cpumask_of_cpu(cpu);
5954 __set_task_cpu(idle, cpu);
5956 rq->curr = rq->idle = idle;
5957 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5960 spin_unlock_irqrestore(&rq->lock, flags);
5962 /* Set the preempt count _outside_ the spinlocks! */
5963 #if defined(CONFIG_PREEMPT)
5964 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5966 task_thread_info(idle)->preempt_count = 0;
5969 * The idle tasks have their own, simple scheduling class:
5971 idle->sched_class = &idle_sched_class;
5972 ftrace_graph_init_task(idle);
5976 * In a system that switches off the HZ timer nohz_cpu_mask
5977 * indicates which cpus entered this state. This is used
5978 * in the rcu update to wait only for active cpus. For system
5979 * which do not switch off the HZ timer nohz_cpu_mask should
5980 * always be CPU_MASK_NONE.
5982 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5985 * Increase the granularity value when there are more CPUs,
5986 * because with more CPUs the 'effective latency' as visible
5987 * to users decreases. But the relationship is not linear,
5988 * so pick a second-best guess by going with the log2 of the
5991 * This idea comes from the SD scheduler of Con Kolivas:
5993 static inline void sched_init_granularity(void)
5995 unsigned int factor = 1 + ilog2(num_online_cpus());
5996 const unsigned long limit = 200000000;
5998 sysctl_sched_min_granularity *= factor;
5999 if (sysctl_sched_min_granularity > limit)
6000 sysctl_sched_min_granularity = limit;
6002 sysctl_sched_latency *= factor;
6003 if (sysctl_sched_latency > limit)
6004 sysctl_sched_latency = limit;
6006 sysctl_sched_wakeup_granularity *= factor;
6008 sysctl_sched_shares_ratelimit *= factor;
6013 * This is how migration works:
6015 * 1) we queue a struct migration_req structure in the source CPU's
6016 * runqueue and wake up that CPU's migration thread.
6017 * 2) we down() the locked semaphore => thread blocks.
6018 * 3) migration thread wakes up (implicitly it forces the migrated
6019 * thread off the CPU)
6020 * 4) it gets the migration request and checks whether the migrated
6021 * task is still in the wrong runqueue.
6022 * 5) if it's in the wrong runqueue then the migration thread removes
6023 * it and puts it into the right queue.
6024 * 6) migration thread up()s the semaphore.
6025 * 7) we wake up and the migration is done.
6029 * Change a given task's CPU affinity. Migrate the thread to a
6030 * proper CPU and schedule it away if the CPU it's executing on
6031 * is removed from the allowed bitmask.
6033 * NOTE: the caller must have a valid reference to the task, the
6034 * task must not exit() & deallocate itself prematurely. The
6035 * call is not atomic; no spinlocks may be held.
6037 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
6039 struct migration_req req;
6040 unsigned long flags;
6044 rq = task_rq_lock(p, &flags);
6045 if (!cpus_intersects(*new_mask, cpu_online_map)) {
6050 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
6051 !cpus_equal(p->cpus_allowed, *new_mask))) {
6056 if (p->sched_class->set_cpus_allowed)
6057 p->sched_class->set_cpus_allowed(p, new_mask);
6059 p->cpus_allowed = *new_mask;
6060 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
6063 /* Can the task run on the task's current CPU? If so, we're done */
6064 if (cpu_isset(task_cpu(p), *new_mask))
6067 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
6068 /* Need help from migration thread: drop lock and wait. */
6069 task_rq_unlock(rq, &flags);
6070 wake_up_process(rq->migration_thread);
6071 wait_for_completion(&req.done);
6072 tlb_migrate_finish(p->mm);
6076 task_rq_unlock(rq, &flags);
6080 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6083 * Move (not current) task off this cpu, onto dest cpu. We're doing
6084 * this because either it can't run here any more (set_cpus_allowed()
6085 * away from this CPU, or CPU going down), or because we're
6086 * attempting to rebalance this task on exec (sched_exec).
6088 * So we race with normal scheduler movements, but that's OK, as long
6089 * as the task is no longer on this CPU.
6091 * Returns non-zero if task was successfully migrated.
6093 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6095 struct rq *rq_dest, *rq_src;
6098 if (unlikely(!cpu_active(dest_cpu)))
6101 rq_src = cpu_rq(src_cpu);
6102 rq_dest = cpu_rq(dest_cpu);
6104 double_rq_lock(rq_src, rq_dest);
6105 /* Already moved. */
6106 if (task_cpu(p) != src_cpu)
6108 /* Affinity changed (again). */
6109 if (!cpu_isset(dest_cpu, p->cpus_allowed))
6112 on_rq = p->se.on_rq;
6114 deactivate_task(rq_src, p, 0);
6116 set_task_cpu(p, dest_cpu);
6118 activate_task(rq_dest, p, 0);
6119 check_preempt_curr(rq_dest, p, 0);
6124 double_rq_unlock(rq_src, rq_dest);
6129 * migration_thread - this is a highprio system thread that performs
6130 * thread migration by bumping thread off CPU then 'pushing' onto
6133 static int migration_thread(void *data)
6135 int cpu = (long)data;
6139 BUG_ON(rq->migration_thread != current);
6141 set_current_state(TASK_INTERRUPTIBLE);
6142 while (!kthread_should_stop()) {
6143 struct migration_req *req;
6144 struct list_head *head;
6146 spin_lock_irq(&rq->lock);
6148 if (cpu_is_offline(cpu)) {
6149 spin_unlock_irq(&rq->lock);
6153 if (rq->active_balance) {
6154 active_load_balance(rq, cpu);
6155 rq->active_balance = 0;
6158 head = &rq->migration_queue;
6160 if (list_empty(head)) {
6161 spin_unlock_irq(&rq->lock);
6163 set_current_state(TASK_INTERRUPTIBLE);
6166 req = list_entry(head->next, struct migration_req, list);
6167 list_del_init(head->next);
6169 spin_unlock(&rq->lock);
6170 __migrate_task(req->task, cpu, req->dest_cpu);
6173 complete(&req->done);
6175 __set_current_state(TASK_RUNNING);
6179 /* Wait for kthread_stop */
6180 set_current_state(TASK_INTERRUPTIBLE);
6181 while (!kthread_should_stop()) {
6183 set_current_state(TASK_INTERRUPTIBLE);
6185 __set_current_state(TASK_RUNNING);
6189 #ifdef CONFIG_HOTPLUG_CPU
6191 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6195 local_irq_disable();
6196 ret = __migrate_task(p, src_cpu, dest_cpu);
6202 * Figure out where task on dead CPU should go, use force if necessary.
6204 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6206 unsigned long flags;
6213 mask = node_to_cpumask(cpu_to_node(dead_cpu));
6214 cpus_and(mask, mask, p->cpus_allowed);
6215 dest_cpu = any_online_cpu(mask);
6217 /* On any allowed CPU? */
6218 if (dest_cpu >= nr_cpu_ids)
6219 dest_cpu = any_online_cpu(p->cpus_allowed);
6221 /* No more Mr. Nice Guy. */
6222 if (dest_cpu >= nr_cpu_ids) {
6223 cpumask_t cpus_allowed;
6225 cpuset_cpus_allowed_locked(p, &cpus_allowed);
6227 * Try to stay on the same cpuset, where the
6228 * current cpuset may be a subset of all cpus.
6229 * The cpuset_cpus_allowed_locked() variant of
6230 * cpuset_cpus_allowed() will not block. It must be
6231 * called within calls to cpuset_lock/cpuset_unlock.
6233 rq = task_rq_lock(p, &flags);
6234 p->cpus_allowed = cpus_allowed;
6235 dest_cpu = any_online_cpu(p->cpus_allowed);
6236 task_rq_unlock(rq, &flags);
6239 * Don't tell them about moving exiting tasks or
6240 * kernel threads (both mm NULL), since they never
6243 if (p->mm && printk_ratelimit()) {
6244 printk(KERN_INFO "process %d (%s) no "
6245 "longer affine to cpu%d\n",
6246 task_pid_nr(p), p->comm, dead_cpu);
6249 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
6253 * While a dead CPU has no uninterruptible tasks queued at this point,
6254 * it might still have a nonzero ->nr_uninterruptible counter, because
6255 * for performance reasons the counter is not stricly tracking tasks to
6256 * their home CPUs. So we just add the counter to another CPU's counter,
6257 * to keep the global sum constant after CPU-down:
6259 static void migrate_nr_uninterruptible(struct rq *rq_src)
6261 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
6262 unsigned long flags;
6264 local_irq_save(flags);
6265 double_rq_lock(rq_src, rq_dest);
6266 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6267 rq_src->nr_uninterruptible = 0;
6268 double_rq_unlock(rq_src, rq_dest);
6269 local_irq_restore(flags);
6272 /* Run through task list and migrate tasks from the dead cpu. */
6273 static void migrate_live_tasks(int src_cpu)
6275 struct task_struct *p, *t;
6277 read_lock(&tasklist_lock);
6279 do_each_thread(t, p) {
6283 if (task_cpu(p) == src_cpu)
6284 move_task_off_dead_cpu(src_cpu, p);
6285 } while_each_thread(t, p);
6287 read_unlock(&tasklist_lock);
6291 * Schedules idle task to be the next runnable task on current CPU.
6292 * It does so by boosting its priority to highest possible.
6293 * Used by CPU offline code.
6295 void sched_idle_next(void)
6297 int this_cpu = smp_processor_id();
6298 struct rq *rq = cpu_rq(this_cpu);
6299 struct task_struct *p = rq->idle;
6300 unsigned long flags;
6302 /* cpu has to be offline */
6303 BUG_ON(cpu_online(this_cpu));
6306 * Strictly not necessary since rest of the CPUs are stopped by now
6307 * and interrupts disabled on the current cpu.
6309 spin_lock_irqsave(&rq->lock, flags);
6311 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6313 update_rq_clock(rq);
6314 activate_task(rq, p, 0);
6316 spin_unlock_irqrestore(&rq->lock, flags);
6320 * Ensures that the idle task is using init_mm right before its cpu goes
6323 void idle_task_exit(void)
6325 struct mm_struct *mm = current->active_mm;
6327 BUG_ON(cpu_online(smp_processor_id()));
6330 switch_mm(mm, &init_mm, current);
6334 /* called under rq->lock with disabled interrupts */
6335 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6337 struct rq *rq = cpu_rq(dead_cpu);
6339 /* Must be exiting, otherwise would be on tasklist. */
6340 BUG_ON(!p->exit_state);
6342 /* Cannot have done final schedule yet: would have vanished. */
6343 BUG_ON(p->state == TASK_DEAD);
6348 * Drop lock around migration; if someone else moves it,
6349 * that's OK. No task can be added to this CPU, so iteration is
6352 spin_unlock_irq(&rq->lock);
6353 move_task_off_dead_cpu(dead_cpu, p);
6354 spin_lock_irq(&rq->lock);
6359 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6360 static void migrate_dead_tasks(unsigned int dead_cpu)
6362 struct rq *rq = cpu_rq(dead_cpu);
6363 struct task_struct *next;
6366 if (!rq->nr_running)
6368 update_rq_clock(rq);
6369 next = pick_next_task(rq, rq->curr);
6372 next->sched_class->put_prev_task(rq, next);
6373 migrate_dead(dead_cpu, next);
6377 #endif /* CONFIG_HOTPLUG_CPU */
6379 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6381 static struct ctl_table sd_ctl_dir[] = {
6383 .procname = "sched_domain",
6389 static struct ctl_table sd_ctl_root[] = {
6391 .ctl_name = CTL_KERN,
6392 .procname = "kernel",
6394 .child = sd_ctl_dir,
6399 static struct ctl_table *sd_alloc_ctl_entry(int n)
6401 struct ctl_table *entry =
6402 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6407 static void sd_free_ctl_entry(struct ctl_table **tablep)
6409 struct ctl_table *entry;
6412 * In the intermediate directories, both the child directory and
6413 * procname are dynamically allocated and could fail but the mode
6414 * will always be set. In the lowest directory the names are
6415 * static strings and all have proc handlers.
6417 for (entry = *tablep; entry->mode; entry++) {
6419 sd_free_ctl_entry(&entry->child);
6420 if (entry->proc_handler == NULL)
6421 kfree(entry->procname);
6429 set_table_entry(struct ctl_table *entry,
6430 const char *procname, void *data, int maxlen,
6431 mode_t mode, proc_handler *proc_handler)
6433 entry->procname = procname;
6435 entry->maxlen = maxlen;
6437 entry->proc_handler = proc_handler;
6440 static struct ctl_table *
6441 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6443 struct ctl_table *table = sd_alloc_ctl_entry(13);
6448 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6449 sizeof(long), 0644, proc_doulongvec_minmax);
6450 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6451 sizeof(long), 0644, proc_doulongvec_minmax);
6452 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6453 sizeof(int), 0644, proc_dointvec_minmax);
6454 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6455 sizeof(int), 0644, proc_dointvec_minmax);
6456 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6457 sizeof(int), 0644, proc_dointvec_minmax);
6458 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6459 sizeof(int), 0644, proc_dointvec_minmax);
6460 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6461 sizeof(int), 0644, proc_dointvec_minmax);
6462 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6463 sizeof(int), 0644, proc_dointvec_minmax);
6464 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6465 sizeof(int), 0644, proc_dointvec_minmax);
6466 set_table_entry(&table[9], "cache_nice_tries",
6467 &sd->cache_nice_tries,
6468 sizeof(int), 0644, proc_dointvec_minmax);
6469 set_table_entry(&table[10], "flags", &sd->flags,
6470 sizeof(int), 0644, proc_dointvec_minmax);
6471 set_table_entry(&table[11], "name", sd->name,
6472 CORENAME_MAX_SIZE, 0444, proc_dostring);
6473 /* &table[12] is terminator */
6478 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6480 struct ctl_table *entry, *table;
6481 struct sched_domain *sd;
6482 int domain_num = 0, i;
6485 for_each_domain(cpu, sd)
6487 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6492 for_each_domain(cpu, sd) {
6493 snprintf(buf, 32, "domain%d", i);
6494 entry->procname = kstrdup(buf, GFP_KERNEL);
6496 entry->child = sd_alloc_ctl_domain_table(sd);
6503 static struct ctl_table_header *sd_sysctl_header;
6504 static void register_sched_domain_sysctl(void)
6506 int i, cpu_num = num_online_cpus();
6507 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6510 WARN_ON(sd_ctl_dir[0].child);
6511 sd_ctl_dir[0].child = entry;
6516 for_each_online_cpu(i) {
6517 snprintf(buf, 32, "cpu%d", i);
6518 entry->procname = kstrdup(buf, GFP_KERNEL);
6520 entry->child = sd_alloc_ctl_cpu_table(i);
6524 WARN_ON(sd_sysctl_header);
6525 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6528 /* may be called multiple times per register */
6529 static void unregister_sched_domain_sysctl(void)
6531 if (sd_sysctl_header)
6532 unregister_sysctl_table(sd_sysctl_header);
6533 sd_sysctl_header = NULL;
6534 if (sd_ctl_dir[0].child)
6535 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6538 static void register_sched_domain_sysctl(void)
6541 static void unregister_sched_domain_sysctl(void)
6546 static void set_rq_online(struct rq *rq)
6549 const struct sched_class *class;
6551 cpu_set(rq->cpu, rq->rd->online);
6554 for_each_class(class) {
6555 if (class->rq_online)
6556 class->rq_online(rq);
6561 static void set_rq_offline(struct rq *rq)
6564 const struct sched_class *class;
6566 for_each_class(class) {
6567 if (class->rq_offline)
6568 class->rq_offline(rq);
6571 cpu_clear(rq->cpu, rq->rd->online);
6577 * migration_call - callback that gets triggered when a CPU is added.
6578 * Here we can start up the necessary migration thread for the new CPU.
6580 static int __cpuinit
6581 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6583 struct task_struct *p;
6584 int cpu = (long)hcpu;
6585 unsigned long flags;
6590 case CPU_UP_PREPARE:
6591 case CPU_UP_PREPARE_FROZEN:
6592 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6595 kthread_bind(p, cpu);
6596 /* Must be high prio: stop_machine expects to yield to it. */
6597 rq = task_rq_lock(p, &flags);
6598 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6599 task_rq_unlock(rq, &flags);
6600 cpu_rq(cpu)->migration_thread = p;
6604 case CPU_ONLINE_FROZEN:
6605 /* Strictly unnecessary, as first user will wake it. */
6606 wake_up_process(cpu_rq(cpu)->migration_thread);
6608 /* Update our root-domain */
6610 spin_lock_irqsave(&rq->lock, flags);
6612 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6616 spin_unlock_irqrestore(&rq->lock, flags);
6619 #ifdef CONFIG_HOTPLUG_CPU
6620 case CPU_UP_CANCELED:
6621 case CPU_UP_CANCELED_FROZEN:
6622 if (!cpu_rq(cpu)->migration_thread)
6624 /* Unbind it from offline cpu so it can run. Fall thru. */
6625 kthread_bind(cpu_rq(cpu)->migration_thread,
6626 any_online_cpu(cpu_online_map));
6627 kthread_stop(cpu_rq(cpu)->migration_thread);
6628 cpu_rq(cpu)->migration_thread = NULL;
6632 case CPU_DEAD_FROZEN:
6633 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6634 migrate_live_tasks(cpu);
6636 kthread_stop(rq->migration_thread);
6637 rq->migration_thread = NULL;
6638 /* Idle task back to normal (off runqueue, low prio) */
6639 spin_lock_irq(&rq->lock);
6640 update_rq_clock(rq);
6641 deactivate_task(rq, rq->idle, 0);
6642 rq->idle->static_prio = MAX_PRIO;
6643 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6644 rq->idle->sched_class = &idle_sched_class;
6645 migrate_dead_tasks(cpu);
6646 spin_unlock_irq(&rq->lock);
6648 migrate_nr_uninterruptible(rq);
6649 BUG_ON(rq->nr_running != 0);
6652 * No need to migrate the tasks: it was best-effort if
6653 * they didn't take sched_hotcpu_mutex. Just wake up
6656 spin_lock_irq(&rq->lock);
6657 while (!list_empty(&rq->migration_queue)) {
6658 struct migration_req *req;
6660 req = list_entry(rq->migration_queue.next,
6661 struct migration_req, list);
6662 list_del_init(&req->list);
6663 spin_unlock_irq(&rq->lock);
6664 complete(&req->done);
6665 spin_lock_irq(&rq->lock);
6667 spin_unlock_irq(&rq->lock);
6671 case CPU_DYING_FROZEN:
6672 /* Update our root-domain */
6674 spin_lock_irqsave(&rq->lock, flags);
6676 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6679 spin_unlock_irqrestore(&rq->lock, flags);
6686 /* Register at highest priority so that task migration (migrate_all_tasks)
6687 * happens before everything else.
6689 static struct notifier_block __cpuinitdata migration_notifier = {
6690 .notifier_call = migration_call,
6694 static int __init migration_init(void)
6696 void *cpu = (void *)(long)smp_processor_id();
6699 /* Start one for the boot CPU: */
6700 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6701 BUG_ON(err == NOTIFY_BAD);
6702 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6703 register_cpu_notifier(&migration_notifier);
6707 early_initcall(migration_init);
6712 #ifdef CONFIG_SCHED_DEBUG
6714 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6715 cpumask_t *groupmask)
6717 struct sched_group *group = sd->groups;
6720 cpulist_scnprintf(str, sizeof(str), sd->span);
6721 cpus_clear(*groupmask);
6723 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6725 if (!(sd->flags & SD_LOAD_BALANCE)) {
6726 printk("does not load-balance\n");
6728 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6733 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6735 if (!cpu_isset(cpu, sd->span)) {
6736 printk(KERN_ERR "ERROR: domain->span does not contain "
6739 if (!cpu_isset(cpu, group->cpumask)) {
6740 printk(KERN_ERR "ERROR: domain->groups does not contain"
6744 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6748 printk(KERN_ERR "ERROR: group is NULL\n");
6752 if (!group->__cpu_power) {
6753 printk(KERN_CONT "\n");
6754 printk(KERN_ERR "ERROR: domain->cpu_power not "
6759 if (!cpus_weight(group->cpumask)) {
6760 printk(KERN_CONT "\n");
6761 printk(KERN_ERR "ERROR: empty group\n");
6765 if (cpus_intersects(*groupmask, group->cpumask)) {
6766 printk(KERN_CONT "\n");
6767 printk(KERN_ERR "ERROR: repeated CPUs\n");
6771 cpus_or(*groupmask, *groupmask, group->cpumask);
6773 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6774 printk(KERN_CONT " %s", str);
6776 group = group->next;
6777 } while (group != sd->groups);
6778 printk(KERN_CONT "\n");
6780 if (!cpus_equal(sd->span, *groupmask))
6781 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6783 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6784 printk(KERN_ERR "ERROR: parent span is not a superset "
6785 "of domain->span\n");
6789 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6791 cpumask_t *groupmask;
6795 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6799 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6801 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6803 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6808 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6817 #else /* !CONFIG_SCHED_DEBUG */
6818 # define sched_domain_debug(sd, cpu) do { } while (0)
6819 #endif /* CONFIG_SCHED_DEBUG */
6821 static int sd_degenerate(struct sched_domain *sd)
6823 if (cpus_weight(sd->span) == 1)
6826 /* Following flags need at least 2 groups */
6827 if (sd->flags & (SD_LOAD_BALANCE |
6828 SD_BALANCE_NEWIDLE |
6832 SD_SHARE_PKG_RESOURCES)) {
6833 if (sd->groups != sd->groups->next)
6837 /* Following flags don't use groups */
6838 if (sd->flags & (SD_WAKE_IDLE |
6847 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6849 unsigned long cflags = sd->flags, pflags = parent->flags;
6851 if (sd_degenerate(parent))
6854 if (!cpus_equal(sd->span, parent->span))
6857 /* Does parent contain flags not in child? */
6858 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6859 if (cflags & SD_WAKE_AFFINE)
6860 pflags &= ~SD_WAKE_BALANCE;
6861 /* Flags needing groups don't count if only 1 group in parent */
6862 if (parent->groups == parent->groups->next) {
6863 pflags &= ~(SD_LOAD_BALANCE |
6864 SD_BALANCE_NEWIDLE |
6868 SD_SHARE_PKG_RESOURCES);
6869 if (nr_node_ids == 1)
6870 pflags &= ~SD_SERIALIZE;
6872 if (~cflags & pflags)
6878 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6880 unsigned long flags;
6882 spin_lock_irqsave(&rq->lock, flags);
6885 struct root_domain *old_rd = rq->rd;
6887 if (cpu_isset(rq->cpu, old_rd->online))
6890 cpu_clear(rq->cpu, old_rd->span);
6892 if (atomic_dec_and_test(&old_rd->refcount))
6896 atomic_inc(&rd->refcount);
6899 cpu_set(rq->cpu, rd->span);
6900 if (cpu_isset(rq->cpu, cpu_online_map))
6903 spin_unlock_irqrestore(&rq->lock, flags);
6906 static void init_rootdomain(struct root_domain *rd)
6908 memset(rd, 0, sizeof(*rd));
6910 cpus_clear(rd->span);
6911 cpus_clear(rd->online);
6913 cpupri_init(&rd->cpupri);
6916 static void init_defrootdomain(void)
6918 init_rootdomain(&def_root_domain);
6919 atomic_set(&def_root_domain.refcount, 1);
6922 static struct root_domain *alloc_rootdomain(void)
6924 struct root_domain *rd;
6926 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6930 init_rootdomain(rd);
6936 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6937 * hold the hotplug lock.
6940 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6942 struct rq *rq = cpu_rq(cpu);
6943 struct sched_domain *tmp;
6945 /* Remove the sched domains which do not contribute to scheduling. */
6946 for (tmp = sd; tmp; ) {
6947 struct sched_domain *parent = tmp->parent;
6951 if (sd_parent_degenerate(tmp, parent)) {
6952 tmp->parent = parent->parent;
6954 parent->parent->child = tmp;
6959 if (sd && sd_degenerate(sd)) {
6965 sched_domain_debug(sd, cpu);
6967 rq_attach_root(rq, rd);
6968 rcu_assign_pointer(rq->sd, sd);
6971 /* cpus with isolated domains */
6972 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6974 /* Setup the mask of cpus configured for isolated domains */
6975 static int __init isolated_cpu_setup(char *str)
6977 static int __initdata ints[NR_CPUS];
6980 str = get_options(str, ARRAY_SIZE(ints), ints);
6981 cpus_clear(cpu_isolated_map);
6982 for (i = 1; i <= ints[0]; i++)
6983 if (ints[i] < NR_CPUS)
6984 cpu_set(ints[i], cpu_isolated_map);
6988 __setup("isolcpus=", isolated_cpu_setup);
6991 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6992 * to a function which identifies what group(along with sched group) a CPU
6993 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6994 * (due to the fact that we keep track of groups covered with a cpumask_t).
6996 * init_sched_build_groups will build a circular linked list of the groups
6997 * covered by the given span, and will set each group's ->cpumask correctly,
6998 * and ->cpu_power to 0.
7001 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
7002 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
7003 struct sched_group **sg,
7004 cpumask_t *tmpmask),
7005 cpumask_t *covered, cpumask_t *tmpmask)
7007 struct sched_group *first = NULL, *last = NULL;
7010 cpus_clear(*covered);
7012 for_each_cpu_mask_nr(i, *span) {
7013 struct sched_group *sg;
7014 int group = group_fn(i, cpu_map, &sg, tmpmask);
7017 if (cpu_isset(i, *covered))
7020 cpus_clear(sg->cpumask);
7021 sg->__cpu_power = 0;
7023 for_each_cpu_mask_nr(j, *span) {
7024 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
7027 cpu_set(j, *covered);
7028 cpu_set(j, sg->cpumask);
7039 #define SD_NODES_PER_DOMAIN 16
7044 * find_next_best_node - find the next node to include in a sched_domain
7045 * @node: node whose sched_domain we're building
7046 * @used_nodes: nodes already in the sched_domain
7048 * Find the next node to include in a given scheduling domain. Simply
7049 * finds the closest node not already in the @used_nodes map.
7051 * Should use nodemask_t.
7053 static int find_next_best_node(int node, nodemask_t *used_nodes)
7055 int i, n, val, min_val, best_node = 0;
7059 for (i = 0; i < nr_node_ids; i++) {
7060 /* Start at @node */
7061 n = (node + i) % nr_node_ids;
7063 if (!nr_cpus_node(n))
7066 /* Skip already used nodes */
7067 if (node_isset(n, *used_nodes))
7070 /* Simple min distance search */
7071 val = node_distance(node, n);
7073 if (val < min_val) {
7079 node_set(best_node, *used_nodes);
7084 * sched_domain_node_span - get a cpumask for a node's sched_domain
7085 * @node: node whose cpumask we're constructing
7086 * @span: resulting cpumask
7088 * Given a node, construct a good cpumask for its sched_domain to span. It
7089 * should be one that prevents unnecessary balancing, but also spreads tasks
7092 static void sched_domain_node_span(int node, cpumask_t *span)
7094 nodemask_t used_nodes;
7095 node_to_cpumask_ptr(nodemask, node);
7099 nodes_clear(used_nodes);
7101 cpus_or(*span, *span, *nodemask);
7102 node_set(node, used_nodes);
7104 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7105 int next_node = find_next_best_node(node, &used_nodes);
7107 node_to_cpumask_ptr_next(nodemask, next_node);
7108 cpus_or(*span, *span, *nodemask);
7111 #endif /* CONFIG_NUMA */
7113 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7116 * SMT sched-domains:
7118 #ifdef CONFIG_SCHED_SMT
7119 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
7120 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
7123 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7127 *sg = &per_cpu(sched_group_cpus, cpu);
7130 #endif /* CONFIG_SCHED_SMT */
7133 * multi-core sched-domains:
7135 #ifdef CONFIG_SCHED_MC
7136 static DEFINE_PER_CPU(struct sched_domain, core_domains);
7137 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
7138 #endif /* CONFIG_SCHED_MC */
7140 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7142 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7147 *mask = per_cpu(cpu_sibling_map, cpu);
7148 cpus_and(*mask, *mask, *cpu_map);
7149 group = first_cpu(*mask);
7151 *sg = &per_cpu(sched_group_core, group);
7154 #elif defined(CONFIG_SCHED_MC)
7156 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7160 *sg = &per_cpu(sched_group_core, cpu);
7165 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
7166 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
7169 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7173 #ifdef CONFIG_SCHED_MC
7174 *mask = cpu_coregroup_map(cpu);
7175 cpus_and(*mask, *mask, *cpu_map);
7176 group = first_cpu(*mask);
7177 #elif defined(CONFIG_SCHED_SMT)
7178 *mask = per_cpu(cpu_sibling_map, cpu);
7179 cpus_and(*mask, *mask, *cpu_map);
7180 group = first_cpu(*mask);
7185 *sg = &per_cpu(sched_group_phys, group);
7191 * The init_sched_build_groups can't handle what we want to do with node
7192 * groups, so roll our own. Now each node has its own list of groups which
7193 * gets dynamically allocated.
7195 static DEFINE_PER_CPU(struct sched_domain, node_domains);
7196 static struct sched_group ***sched_group_nodes_bycpu;
7198 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
7199 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
7201 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
7202 struct sched_group **sg, cpumask_t *nodemask)
7206 *nodemask = node_to_cpumask(cpu_to_node(cpu));
7207 cpus_and(*nodemask, *nodemask, *cpu_map);
7208 group = first_cpu(*nodemask);
7211 *sg = &per_cpu(sched_group_allnodes, group);
7215 static void init_numa_sched_groups_power(struct sched_group *group_head)
7217 struct sched_group *sg = group_head;
7223 for_each_cpu_mask_nr(j, sg->cpumask) {
7224 struct sched_domain *sd;
7226 sd = &per_cpu(phys_domains, j);
7227 if (j != first_cpu(sd->groups->cpumask)) {
7229 * Only add "power" once for each
7235 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7238 } while (sg != group_head);
7240 #endif /* CONFIG_NUMA */
7243 /* Free memory allocated for various sched_group structures */
7244 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7248 for_each_cpu_mask_nr(cpu, *cpu_map) {
7249 struct sched_group **sched_group_nodes
7250 = sched_group_nodes_bycpu[cpu];
7252 if (!sched_group_nodes)
7255 for (i = 0; i < nr_node_ids; i++) {
7256 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7258 *nodemask = node_to_cpumask(i);
7259 cpus_and(*nodemask, *nodemask, *cpu_map);
7260 if (cpus_empty(*nodemask))
7270 if (oldsg != sched_group_nodes[i])
7273 kfree(sched_group_nodes);
7274 sched_group_nodes_bycpu[cpu] = NULL;
7277 #else /* !CONFIG_NUMA */
7278 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7281 #endif /* CONFIG_NUMA */
7284 * Initialize sched groups cpu_power.
7286 * cpu_power indicates the capacity of sched group, which is used while
7287 * distributing the load between different sched groups in a sched domain.
7288 * Typically cpu_power for all the groups in a sched domain will be same unless
7289 * there are asymmetries in the topology. If there are asymmetries, group
7290 * having more cpu_power will pickup more load compared to the group having
7293 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7294 * the maximum number of tasks a group can handle in the presence of other idle
7295 * or lightly loaded groups in the same sched domain.
7297 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7299 struct sched_domain *child;
7300 struct sched_group *group;
7302 WARN_ON(!sd || !sd->groups);
7304 if (cpu != first_cpu(sd->groups->cpumask))
7309 sd->groups->__cpu_power = 0;
7312 * For perf policy, if the groups in child domain share resources
7313 * (for example cores sharing some portions of the cache hierarchy
7314 * or SMT), then set this domain groups cpu_power such that each group
7315 * can handle only one task, when there are other idle groups in the
7316 * same sched domain.
7318 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7320 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7321 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7326 * add cpu_power of each child group to this groups cpu_power
7328 group = child->groups;
7330 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7331 group = group->next;
7332 } while (group != child->groups);
7336 * Initializers for schedule domains
7337 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7340 #ifdef CONFIG_SCHED_DEBUG
7341 # define SD_INIT_NAME(sd, type) sd->name = #type
7343 # define SD_INIT_NAME(sd, type) do { } while (0)
7346 #define SD_INIT(sd, type) sd_init_##type(sd)
7348 #define SD_INIT_FUNC(type) \
7349 static noinline void sd_init_##type(struct sched_domain *sd) \
7351 memset(sd, 0, sizeof(*sd)); \
7352 *sd = SD_##type##_INIT; \
7353 sd->level = SD_LV_##type; \
7354 SD_INIT_NAME(sd, type); \
7359 SD_INIT_FUNC(ALLNODES)
7362 #ifdef CONFIG_SCHED_SMT
7363 SD_INIT_FUNC(SIBLING)
7365 #ifdef CONFIG_SCHED_MC
7370 * To minimize stack usage kmalloc room for cpumasks and share the
7371 * space as the usage in build_sched_domains() dictates. Used only
7372 * if the amount of space is significant.
7375 cpumask_t tmpmask; /* make this one first */
7378 cpumask_t this_sibling_map;
7379 cpumask_t this_core_map;
7381 cpumask_t send_covered;
7384 cpumask_t domainspan;
7386 cpumask_t notcovered;
7391 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7392 static inline void sched_cpumask_alloc(struct allmasks **masks)
7394 *masks = kmalloc(sizeof(**masks), GFP_KERNEL);
7396 static inline void sched_cpumask_free(struct allmasks *masks)
7401 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7402 static inline void sched_cpumask_alloc(struct allmasks **masks)
7404 static inline void sched_cpumask_free(struct allmasks *masks)
7408 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7409 ((unsigned long)(a) + offsetof(struct allmasks, v))
7411 static int default_relax_domain_level = -1;
7413 static int __init setup_relax_domain_level(char *str)
7417 val = simple_strtoul(str, NULL, 0);
7418 if (val < SD_LV_MAX)
7419 default_relax_domain_level = val;
7423 __setup("relax_domain_level=", setup_relax_domain_level);
7425 static void set_domain_attribute(struct sched_domain *sd,
7426 struct sched_domain_attr *attr)
7430 if (!attr || attr->relax_domain_level < 0) {
7431 if (default_relax_domain_level < 0)
7434 request = default_relax_domain_level;
7436 request = attr->relax_domain_level;
7437 if (request < sd->level) {
7438 /* turn off idle balance on this domain */
7439 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7441 /* turn on idle balance on this domain */
7442 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7447 * Build sched domains for a given set of cpus and attach the sched domains
7448 * to the individual cpus
7450 static int __build_sched_domains(const cpumask_t *cpu_map,
7451 struct sched_domain_attr *attr)
7454 struct root_domain *rd;
7455 SCHED_CPUMASK_DECLARE(allmasks);
7458 struct sched_group **sched_group_nodes = NULL;
7459 int sd_allnodes = 0;
7462 * Allocate the per-node list of sched groups
7464 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
7466 if (!sched_group_nodes) {
7467 printk(KERN_WARNING "Can not alloc sched group node list\n");
7472 rd = alloc_rootdomain();
7474 printk(KERN_WARNING "Cannot alloc root domain\n");
7476 kfree(sched_group_nodes);
7481 /* get space for all scratch cpumask variables */
7482 sched_cpumask_alloc(&allmasks);
7484 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7487 kfree(sched_group_nodes);
7492 tmpmask = (cpumask_t *)allmasks;
7496 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7500 * Set up domains for cpus specified by the cpu_map.
7502 for_each_cpu_mask_nr(i, *cpu_map) {
7503 struct sched_domain *sd = NULL, *p;
7504 SCHED_CPUMASK_VAR(nodemask, allmasks);
7506 *nodemask = node_to_cpumask(cpu_to_node(i));
7507 cpus_and(*nodemask, *nodemask, *cpu_map);
7510 if (cpus_weight(*cpu_map) >
7511 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7512 sd = &per_cpu(allnodes_domains, i);
7513 SD_INIT(sd, ALLNODES);
7514 set_domain_attribute(sd, attr);
7515 sd->span = *cpu_map;
7516 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7522 sd = &per_cpu(node_domains, i);
7524 set_domain_attribute(sd, attr);
7525 sched_domain_node_span(cpu_to_node(i), &sd->span);
7529 cpus_and(sd->span, sd->span, *cpu_map);
7533 sd = &per_cpu(phys_domains, i);
7535 set_domain_attribute(sd, attr);
7536 sd->span = *nodemask;
7540 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7542 #ifdef CONFIG_SCHED_MC
7544 sd = &per_cpu(core_domains, i);
7546 set_domain_attribute(sd, attr);
7547 sd->span = cpu_coregroup_map(i);
7548 cpus_and(sd->span, sd->span, *cpu_map);
7551 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7554 #ifdef CONFIG_SCHED_SMT
7556 sd = &per_cpu(cpu_domains, i);
7557 SD_INIT(sd, SIBLING);
7558 set_domain_attribute(sd, attr);
7559 sd->span = per_cpu(cpu_sibling_map, i);
7560 cpus_and(sd->span, sd->span, *cpu_map);
7563 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7567 #ifdef CONFIG_SCHED_SMT
7568 /* Set up CPU (sibling) groups */
7569 for_each_cpu_mask_nr(i, *cpu_map) {
7570 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7571 SCHED_CPUMASK_VAR(send_covered, allmasks);
7573 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7574 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7575 if (i != first_cpu(*this_sibling_map))
7578 init_sched_build_groups(this_sibling_map, cpu_map,
7580 send_covered, tmpmask);
7584 #ifdef CONFIG_SCHED_MC
7585 /* Set up multi-core groups */
7586 for_each_cpu_mask_nr(i, *cpu_map) {
7587 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7588 SCHED_CPUMASK_VAR(send_covered, allmasks);
7590 *this_core_map = cpu_coregroup_map(i);
7591 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7592 if (i != first_cpu(*this_core_map))
7595 init_sched_build_groups(this_core_map, cpu_map,
7597 send_covered, tmpmask);
7601 /* Set up physical groups */
7602 for (i = 0; i < nr_node_ids; i++) {
7603 SCHED_CPUMASK_VAR(nodemask, allmasks);
7604 SCHED_CPUMASK_VAR(send_covered, allmasks);
7606 *nodemask = node_to_cpumask(i);
7607 cpus_and(*nodemask, *nodemask, *cpu_map);
7608 if (cpus_empty(*nodemask))
7611 init_sched_build_groups(nodemask, cpu_map,
7613 send_covered, tmpmask);
7617 /* Set up node groups */
7619 SCHED_CPUMASK_VAR(send_covered, allmasks);
7621 init_sched_build_groups(cpu_map, cpu_map,
7622 &cpu_to_allnodes_group,
7623 send_covered, tmpmask);
7626 for (i = 0; i < nr_node_ids; i++) {
7627 /* Set up node groups */
7628 struct sched_group *sg, *prev;
7629 SCHED_CPUMASK_VAR(nodemask, allmasks);
7630 SCHED_CPUMASK_VAR(domainspan, allmasks);
7631 SCHED_CPUMASK_VAR(covered, allmasks);
7634 *nodemask = node_to_cpumask(i);
7635 cpus_clear(*covered);
7637 cpus_and(*nodemask, *nodemask, *cpu_map);
7638 if (cpus_empty(*nodemask)) {
7639 sched_group_nodes[i] = NULL;
7643 sched_domain_node_span(i, domainspan);
7644 cpus_and(*domainspan, *domainspan, *cpu_map);
7646 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7648 printk(KERN_WARNING "Can not alloc domain group for "
7652 sched_group_nodes[i] = sg;
7653 for_each_cpu_mask_nr(j, *nodemask) {
7654 struct sched_domain *sd;
7656 sd = &per_cpu(node_domains, j);
7659 sg->__cpu_power = 0;
7660 sg->cpumask = *nodemask;
7662 cpus_or(*covered, *covered, *nodemask);
7665 for (j = 0; j < nr_node_ids; j++) {
7666 SCHED_CPUMASK_VAR(notcovered, allmasks);
7667 int n = (i + j) % nr_node_ids;
7668 node_to_cpumask_ptr(pnodemask, n);
7670 cpus_complement(*notcovered, *covered);
7671 cpus_and(*tmpmask, *notcovered, *cpu_map);
7672 cpus_and(*tmpmask, *tmpmask, *domainspan);
7673 if (cpus_empty(*tmpmask))
7676 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7677 if (cpus_empty(*tmpmask))
7680 sg = kmalloc_node(sizeof(struct sched_group),
7684 "Can not alloc domain group for node %d\n", j);
7687 sg->__cpu_power = 0;
7688 sg->cpumask = *tmpmask;
7689 sg->next = prev->next;
7690 cpus_or(*covered, *covered, *tmpmask);
7697 /* Calculate CPU power for physical packages and nodes */
7698 #ifdef CONFIG_SCHED_SMT
7699 for_each_cpu_mask_nr(i, *cpu_map) {
7700 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7702 init_sched_groups_power(i, sd);
7705 #ifdef CONFIG_SCHED_MC
7706 for_each_cpu_mask_nr(i, *cpu_map) {
7707 struct sched_domain *sd = &per_cpu(core_domains, i);
7709 init_sched_groups_power(i, sd);
7713 for_each_cpu_mask_nr(i, *cpu_map) {
7714 struct sched_domain *sd = &per_cpu(phys_domains, i);
7716 init_sched_groups_power(i, sd);
7720 for (i = 0; i < nr_node_ids; i++)
7721 init_numa_sched_groups_power(sched_group_nodes[i]);
7724 struct sched_group *sg;
7726 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7728 init_numa_sched_groups_power(sg);
7732 /* Attach the domains */
7733 for_each_cpu_mask_nr(i, *cpu_map) {
7734 struct sched_domain *sd;
7735 #ifdef CONFIG_SCHED_SMT
7736 sd = &per_cpu(cpu_domains, i);
7737 #elif defined(CONFIG_SCHED_MC)
7738 sd = &per_cpu(core_domains, i);
7740 sd = &per_cpu(phys_domains, i);
7742 cpu_attach_domain(sd, rd, i);
7745 sched_cpumask_free(allmasks);
7750 free_sched_groups(cpu_map, tmpmask);
7751 sched_cpumask_free(allmasks);
7757 static int build_sched_domains(const cpumask_t *cpu_map)
7759 return __build_sched_domains(cpu_map, NULL);
7762 static cpumask_t *doms_cur; /* current sched domains */
7763 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7764 static struct sched_domain_attr *dattr_cur;
7765 /* attribues of custom domains in 'doms_cur' */
7768 * Special case: If a kmalloc of a doms_cur partition (array of
7769 * cpumask_t) fails, then fallback to a single sched domain,
7770 * as determined by the single cpumask_t fallback_doms.
7772 static cpumask_t fallback_doms;
7775 * arch_update_cpu_topology lets virtualized architectures update the
7776 * cpu core maps. It is supposed to return 1 if the topology changed
7777 * or 0 if it stayed the same.
7779 int __attribute__((weak)) arch_update_cpu_topology(void)
7785 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7786 * For now this just excludes isolated cpus, but could be used to
7787 * exclude other special cases in the future.
7789 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7793 arch_update_cpu_topology();
7795 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7797 doms_cur = &fallback_doms;
7798 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7800 err = build_sched_domains(doms_cur);
7801 register_sched_domain_sysctl();
7806 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7809 free_sched_groups(cpu_map, tmpmask);
7813 * Detach sched domains from a group of cpus specified in cpu_map
7814 * These cpus will now be attached to the NULL domain
7816 static void detach_destroy_domains(const cpumask_t *cpu_map)
7821 for_each_cpu_mask_nr(i, *cpu_map)
7822 cpu_attach_domain(NULL, &def_root_domain, i);
7823 synchronize_sched();
7824 arch_destroy_sched_domains(cpu_map, &tmpmask);
7827 /* handle null as "default" */
7828 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7829 struct sched_domain_attr *new, int idx_new)
7831 struct sched_domain_attr tmp;
7838 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7839 new ? (new + idx_new) : &tmp,
7840 sizeof(struct sched_domain_attr));
7844 * Partition sched domains as specified by the 'ndoms_new'
7845 * cpumasks in the array doms_new[] of cpumasks. This compares
7846 * doms_new[] to the current sched domain partitioning, doms_cur[].
7847 * It destroys each deleted domain and builds each new domain.
7849 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7850 * The masks don't intersect (don't overlap.) We should setup one
7851 * sched domain for each mask. CPUs not in any of the cpumasks will
7852 * not be load balanced. If the same cpumask appears both in the
7853 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7856 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7857 * ownership of it and will kfree it when done with it. If the caller
7858 * failed the kmalloc call, then it can pass in doms_new == NULL &&
7859 * ndoms_new == 1, and partition_sched_domains() will fallback to
7860 * the single partition 'fallback_doms', it also forces the domains
7863 * If doms_new == NULL it will be replaced with cpu_online_map.
7864 * ndoms_new == 0 is a special case for destroying existing domains,
7865 * and it will not create the default domain.
7867 * Call with hotplug lock held
7869 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7870 struct sched_domain_attr *dattr_new)
7875 mutex_lock(&sched_domains_mutex);
7877 /* always unregister in case we don't destroy any domains */
7878 unregister_sched_domain_sysctl();
7880 /* Let architecture update cpu core mappings. */
7881 new_topology = arch_update_cpu_topology();
7883 n = doms_new ? ndoms_new : 0;
7885 /* Destroy deleted domains */
7886 for (i = 0; i < ndoms_cur; i++) {
7887 for (j = 0; j < n && !new_topology; j++) {
7888 if (cpus_equal(doms_cur[i], doms_new[j])
7889 && dattrs_equal(dattr_cur, i, dattr_new, j))
7892 /* no match - a current sched domain not in new doms_new[] */
7893 detach_destroy_domains(doms_cur + i);
7898 if (doms_new == NULL) {
7900 doms_new = &fallback_doms;
7901 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7902 WARN_ON_ONCE(dattr_new);
7905 /* Build new domains */
7906 for (i = 0; i < ndoms_new; i++) {
7907 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7908 if (cpus_equal(doms_new[i], doms_cur[j])
7909 && dattrs_equal(dattr_new, i, dattr_cur, j))
7912 /* no match - add a new doms_new */
7913 __build_sched_domains(doms_new + i,
7914 dattr_new ? dattr_new + i : NULL);
7919 /* Remember the new sched domains */
7920 if (doms_cur != &fallback_doms)
7922 kfree(dattr_cur); /* kfree(NULL) is safe */
7923 doms_cur = doms_new;
7924 dattr_cur = dattr_new;
7925 ndoms_cur = ndoms_new;
7927 register_sched_domain_sysctl();
7929 mutex_unlock(&sched_domains_mutex);
7932 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7933 int arch_reinit_sched_domains(void)
7937 /* Destroy domains first to force the rebuild */
7938 partition_sched_domains(0, NULL, NULL);
7940 rebuild_sched_domains();
7946 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7950 if (buf[0] != '0' && buf[0] != '1')
7954 sched_smt_power_savings = (buf[0] == '1');
7956 sched_mc_power_savings = (buf[0] == '1');
7958 ret = arch_reinit_sched_domains();
7960 return ret ? ret : count;
7963 #ifdef CONFIG_SCHED_MC
7964 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7967 return sprintf(page, "%u\n", sched_mc_power_savings);
7969 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7970 const char *buf, size_t count)
7972 return sched_power_savings_store(buf, count, 0);
7974 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7975 sched_mc_power_savings_show,
7976 sched_mc_power_savings_store);
7979 #ifdef CONFIG_SCHED_SMT
7980 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7983 return sprintf(page, "%u\n", sched_smt_power_savings);
7985 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7986 const char *buf, size_t count)
7988 return sched_power_savings_store(buf, count, 1);
7990 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7991 sched_smt_power_savings_show,
7992 sched_smt_power_savings_store);
7995 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7999 #ifdef CONFIG_SCHED_SMT
8001 err = sysfs_create_file(&cls->kset.kobj,
8002 &attr_sched_smt_power_savings.attr);
8004 #ifdef CONFIG_SCHED_MC
8005 if (!err && mc_capable())
8006 err = sysfs_create_file(&cls->kset.kobj,
8007 &attr_sched_mc_power_savings.attr);
8011 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8013 #ifndef CONFIG_CPUSETS
8015 * Add online and remove offline CPUs from the scheduler domains.
8016 * When cpusets are enabled they take over this function.
8018 static int update_sched_domains(struct notifier_block *nfb,
8019 unsigned long action, void *hcpu)
8023 case CPU_ONLINE_FROZEN:
8025 case CPU_DEAD_FROZEN:
8026 partition_sched_domains(1, NULL, NULL);
8035 static int update_runtime(struct notifier_block *nfb,
8036 unsigned long action, void *hcpu)
8038 int cpu = (int)(long)hcpu;
8041 case CPU_DOWN_PREPARE:
8042 case CPU_DOWN_PREPARE_FROZEN:
8043 disable_runtime(cpu_rq(cpu));
8046 case CPU_DOWN_FAILED:
8047 case CPU_DOWN_FAILED_FROZEN:
8049 case CPU_ONLINE_FROZEN:
8050 enable_runtime(cpu_rq(cpu));
8058 void __init sched_init_smp(void)
8060 cpumask_t non_isolated_cpus;
8062 #if defined(CONFIG_NUMA)
8063 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
8065 BUG_ON(sched_group_nodes_bycpu == NULL);
8068 mutex_lock(&sched_domains_mutex);
8069 arch_init_sched_domains(&cpu_online_map);
8070 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
8071 if (cpus_empty(non_isolated_cpus))
8072 cpu_set(smp_processor_id(), non_isolated_cpus);
8073 mutex_unlock(&sched_domains_mutex);
8076 #ifndef CONFIG_CPUSETS
8077 /* XXX: Theoretical race here - CPU may be hotplugged now */
8078 hotcpu_notifier(update_sched_domains, 0);
8081 /* RT runtime code needs to handle some hotplug events */
8082 hotcpu_notifier(update_runtime, 0);
8086 /* Move init over to a non-isolated CPU */
8087 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
8089 sched_init_granularity();
8092 void __init sched_init_smp(void)
8094 sched_init_granularity();
8096 #endif /* CONFIG_SMP */
8098 int in_sched_functions(unsigned long addr)
8100 return in_lock_functions(addr) ||
8101 (addr >= (unsigned long)__sched_text_start
8102 && addr < (unsigned long)__sched_text_end);
8105 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
8107 cfs_rq->tasks_timeline = RB_ROOT;
8108 INIT_LIST_HEAD(&cfs_rq->tasks);
8109 #ifdef CONFIG_FAIR_GROUP_SCHED
8112 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8115 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8117 struct rt_prio_array *array;
8120 array = &rt_rq->active;
8121 for (i = 0; i < MAX_RT_PRIO; i++) {
8122 INIT_LIST_HEAD(array->queue + i);
8123 __clear_bit(i, array->bitmap);
8125 /* delimiter for bitsearch: */
8126 __set_bit(MAX_RT_PRIO, array->bitmap);
8128 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8129 rt_rq->highest_prio = MAX_RT_PRIO;
8132 rt_rq->rt_nr_migratory = 0;
8133 rt_rq->overloaded = 0;
8137 rt_rq->rt_throttled = 0;
8138 rt_rq->rt_runtime = 0;
8139 spin_lock_init(&rt_rq->rt_runtime_lock);
8141 #ifdef CONFIG_RT_GROUP_SCHED
8142 rt_rq->rt_nr_boosted = 0;
8147 #ifdef CONFIG_FAIR_GROUP_SCHED
8148 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8149 struct sched_entity *se, int cpu, int add,
8150 struct sched_entity *parent)
8152 struct rq *rq = cpu_rq(cpu);
8153 tg->cfs_rq[cpu] = cfs_rq;
8154 init_cfs_rq(cfs_rq, rq);
8157 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8160 /* se could be NULL for init_task_group */
8165 se->cfs_rq = &rq->cfs;
8167 se->cfs_rq = parent->my_q;
8170 se->load.weight = tg->shares;
8171 se->load.inv_weight = 0;
8172 se->parent = parent;
8176 #ifdef CONFIG_RT_GROUP_SCHED
8177 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8178 struct sched_rt_entity *rt_se, int cpu, int add,
8179 struct sched_rt_entity *parent)
8181 struct rq *rq = cpu_rq(cpu);
8183 tg->rt_rq[cpu] = rt_rq;
8184 init_rt_rq(rt_rq, rq);
8186 rt_rq->rt_se = rt_se;
8187 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8189 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8191 tg->rt_se[cpu] = rt_se;
8196 rt_se->rt_rq = &rq->rt;
8198 rt_se->rt_rq = parent->my_q;
8200 rt_se->my_q = rt_rq;
8201 rt_se->parent = parent;
8202 INIT_LIST_HEAD(&rt_se->run_list);
8206 void __init sched_init(void)
8209 unsigned long alloc_size = 0, ptr;
8211 #ifdef CONFIG_FAIR_GROUP_SCHED
8212 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8214 #ifdef CONFIG_RT_GROUP_SCHED
8215 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8217 #ifdef CONFIG_USER_SCHED
8221 * As sched_init() is called before page_alloc is setup,
8222 * we use alloc_bootmem().
8225 ptr = (unsigned long)alloc_bootmem(alloc_size);
8227 #ifdef CONFIG_FAIR_GROUP_SCHED
8228 init_task_group.se = (struct sched_entity **)ptr;
8229 ptr += nr_cpu_ids * sizeof(void **);
8231 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8232 ptr += nr_cpu_ids * sizeof(void **);
8234 #ifdef CONFIG_USER_SCHED
8235 root_task_group.se = (struct sched_entity **)ptr;
8236 ptr += nr_cpu_ids * sizeof(void **);
8238 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8239 ptr += nr_cpu_ids * sizeof(void **);
8240 #endif /* CONFIG_USER_SCHED */
8241 #endif /* CONFIG_FAIR_GROUP_SCHED */
8242 #ifdef CONFIG_RT_GROUP_SCHED
8243 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8244 ptr += nr_cpu_ids * sizeof(void **);
8246 init_task_group.rt_rq = (struct rt_rq **)ptr;
8247 ptr += nr_cpu_ids * sizeof(void **);
8249 #ifdef CONFIG_USER_SCHED
8250 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8251 ptr += nr_cpu_ids * sizeof(void **);
8253 root_task_group.rt_rq = (struct rt_rq **)ptr;
8254 ptr += nr_cpu_ids * sizeof(void **);
8255 #endif /* CONFIG_USER_SCHED */
8256 #endif /* CONFIG_RT_GROUP_SCHED */
8260 init_defrootdomain();
8263 init_rt_bandwidth(&def_rt_bandwidth,
8264 global_rt_period(), global_rt_runtime());
8266 #ifdef CONFIG_RT_GROUP_SCHED
8267 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8268 global_rt_period(), global_rt_runtime());
8269 #ifdef CONFIG_USER_SCHED
8270 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8271 global_rt_period(), RUNTIME_INF);
8272 #endif /* CONFIG_USER_SCHED */
8273 #endif /* CONFIG_RT_GROUP_SCHED */
8275 #ifdef CONFIG_GROUP_SCHED
8276 list_add(&init_task_group.list, &task_groups);
8277 INIT_LIST_HEAD(&init_task_group.children);
8279 #ifdef CONFIG_USER_SCHED
8280 INIT_LIST_HEAD(&root_task_group.children);
8281 init_task_group.parent = &root_task_group;
8282 list_add(&init_task_group.siblings, &root_task_group.children);
8283 #endif /* CONFIG_USER_SCHED */
8284 #endif /* CONFIG_GROUP_SCHED */
8286 for_each_possible_cpu(i) {
8290 spin_lock_init(&rq->lock);
8292 init_cfs_rq(&rq->cfs, rq);
8293 init_rt_rq(&rq->rt, rq);
8294 #ifdef CONFIG_FAIR_GROUP_SCHED
8295 init_task_group.shares = init_task_group_load;
8296 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8297 #ifdef CONFIG_CGROUP_SCHED
8299 * How much cpu bandwidth does init_task_group get?
8301 * In case of task-groups formed thr' the cgroup filesystem, it
8302 * gets 100% of the cpu resources in the system. This overall
8303 * system cpu resource is divided among the tasks of
8304 * init_task_group and its child task-groups in a fair manner,
8305 * based on each entity's (task or task-group's) weight
8306 * (se->load.weight).
8308 * In other words, if init_task_group has 10 tasks of weight
8309 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8310 * then A0's share of the cpu resource is:
8312 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8314 * We achieve this by letting init_task_group's tasks sit
8315 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8317 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8318 #elif defined CONFIG_USER_SCHED
8319 root_task_group.shares = NICE_0_LOAD;
8320 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8322 * In case of task-groups formed thr' the user id of tasks,
8323 * init_task_group represents tasks belonging to root user.
8324 * Hence it forms a sibling of all subsequent groups formed.
8325 * In this case, init_task_group gets only a fraction of overall
8326 * system cpu resource, based on the weight assigned to root
8327 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8328 * by letting tasks of init_task_group sit in a separate cfs_rq
8329 * (init_cfs_rq) and having one entity represent this group of
8330 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8332 init_tg_cfs_entry(&init_task_group,
8333 &per_cpu(init_cfs_rq, i),
8334 &per_cpu(init_sched_entity, i), i, 1,
8335 root_task_group.se[i]);
8338 #endif /* CONFIG_FAIR_GROUP_SCHED */
8340 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8341 #ifdef CONFIG_RT_GROUP_SCHED
8342 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8343 #ifdef CONFIG_CGROUP_SCHED
8344 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8345 #elif defined CONFIG_USER_SCHED
8346 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8347 init_tg_rt_entry(&init_task_group,
8348 &per_cpu(init_rt_rq, i),
8349 &per_cpu(init_sched_rt_entity, i), i, 1,
8350 root_task_group.rt_se[i]);
8354 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8355 rq->cpu_load[j] = 0;
8359 rq->active_balance = 0;
8360 rq->next_balance = jiffies;
8364 rq->migration_thread = NULL;
8365 INIT_LIST_HEAD(&rq->migration_queue);
8366 rq_attach_root(rq, &def_root_domain);
8369 atomic_set(&rq->nr_iowait, 0);
8372 set_load_weight(&init_task);
8374 #ifdef CONFIG_PREEMPT_NOTIFIERS
8375 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8379 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8382 #ifdef CONFIG_RT_MUTEXES
8383 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8387 * The boot idle thread does lazy MMU switching as well:
8389 atomic_inc(&init_mm.mm_count);
8390 enter_lazy_tlb(&init_mm, current);
8393 * Make us the idle thread. Technically, schedule() should not be
8394 * called from this thread, however somewhere below it might be,
8395 * but because we are the idle thread, we just pick up running again
8396 * when this runqueue becomes "idle".
8398 init_idle(current, smp_processor_id());
8400 * During early bootup we pretend to be a normal task:
8402 current->sched_class = &fair_sched_class;
8404 scheduler_running = 1;
8407 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8408 void __might_sleep(char *file, int line)
8411 static unsigned long prev_jiffy; /* ratelimiting */
8413 if ((!in_atomic() && !irqs_disabled()) ||
8414 system_state != SYSTEM_RUNNING || oops_in_progress)
8416 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8418 prev_jiffy = jiffies;
8421 "BUG: sleeping function called from invalid context at %s:%d\n",
8424 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8425 in_atomic(), irqs_disabled(),
8426 current->pid, current->comm);
8428 debug_show_held_locks(current);
8429 if (irqs_disabled())
8430 print_irqtrace_events(current);
8434 EXPORT_SYMBOL(__might_sleep);
8437 #ifdef CONFIG_MAGIC_SYSRQ
8438 static void normalize_task(struct rq *rq, struct task_struct *p)
8442 update_rq_clock(rq);
8443 on_rq = p->se.on_rq;
8445 deactivate_task(rq, p, 0);
8446 __setscheduler(rq, p, SCHED_NORMAL, 0);
8448 activate_task(rq, p, 0);
8449 resched_task(rq->curr);
8453 void normalize_rt_tasks(void)
8455 struct task_struct *g, *p;
8456 unsigned long flags;
8459 read_lock_irqsave(&tasklist_lock, flags);
8460 do_each_thread(g, p) {
8462 * Only normalize user tasks:
8467 p->se.exec_start = 0;
8468 #ifdef CONFIG_SCHEDSTATS
8469 p->se.wait_start = 0;
8470 p->se.sleep_start = 0;
8471 p->se.block_start = 0;
8476 * Renice negative nice level userspace
8479 if (TASK_NICE(p) < 0 && p->mm)
8480 set_user_nice(p, 0);
8484 spin_lock(&p->pi_lock);
8485 rq = __task_rq_lock(p);
8487 normalize_task(rq, p);
8489 __task_rq_unlock(rq);
8490 spin_unlock(&p->pi_lock);
8491 } while_each_thread(g, p);
8493 read_unlock_irqrestore(&tasklist_lock, flags);
8496 #endif /* CONFIG_MAGIC_SYSRQ */
8500 * These functions are only useful for the IA64 MCA handling.
8502 * They can only be called when the whole system has been
8503 * stopped - every CPU needs to be quiescent, and no scheduling
8504 * activity can take place. Using them for anything else would
8505 * be a serious bug, and as a result, they aren't even visible
8506 * under any other configuration.
8510 * curr_task - return the current task for a given cpu.
8511 * @cpu: the processor in question.
8513 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8515 struct task_struct *curr_task(int cpu)
8517 return cpu_curr(cpu);
8521 * set_curr_task - set the current task for a given cpu.
8522 * @cpu: the processor in question.
8523 * @p: the task pointer to set.
8525 * Description: This function must only be used when non-maskable interrupts
8526 * are serviced on a separate stack. It allows the architecture to switch the
8527 * notion of the current task on a cpu in a non-blocking manner. This function
8528 * must be called with all CPU's synchronized, and interrupts disabled, the
8529 * and caller must save the original value of the current task (see
8530 * curr_task() above) and restore that value before reenabling interrupts and
8531 * re-starting the system.
8533 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8535 void set_curr_task(int cpu, struct task_struct *p)
8542 #ifdef CONFIG_FAIR_GROUP_SCHED
8543 static void free_fair_sched_group(struct task_group *tg)
8547 for_each_possible_cpu(i) {
8549 kfree(tg->cfs_rq[i]);
8559 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8561 struct cfs_rq *cfs_rq;
8562 struct sched_entity *se;
8566 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8569 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8573 tg->shares = NICE_0_LOAD;
8575 for_each_possible_cpu(i) {
8578 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8579 GFP_KERNEL, cpu_to_node(i));
8583 se = kzalloc_node(sizeof(struct sched_entity),
8584 GFP_KERNEL, cpu_to_node(i));
8588 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8597 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8599 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8600 &cpu_rq(cpu)->leaf_cfs_rq_list);
8603 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8605 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8607 #else /* !CONFG_FAIR_GROUP_SCHED */
8608 static inline void free_fair_sched_group(struct task_group *tg)
8613 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8618 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8622 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8625 #endif /* CONFIG_FAIR_GROUP_SCHED */
8627 #ifdef CONFIG_RT_GROUP_SCHED
8628 static void free_rt_sched_group(struct task_group *tg)
8632 destroy_rt_bandwidth(&tg->rt_bandwidth);
8634 for_each_possible_cpu(i) {
8636 kfree(tg->rt_rq[i]);
8638 kfree(tg->rt_se[i]);
8646 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8648 struct rt_rq *rt_rq;
8649 struct sched_rt_entity *rt_se;
8653 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8656 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8660 init_rt_bandwidth(&tg->rt_bandwidth,
8661 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8663 for_each_possible_cpu(i) {
8666 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8667 GFP_KERNEL, cpu_to_node(i));
8671 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8672 GFP_KERNEL, cpu_to_node(i));
8676 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8685 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8687 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8688 &cpu_rq(cpu)->leaf_rt_rq_list);
8691 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8693 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8695 #else /* !CONFIG_RT_GROUP_SCHED */
8696 static inline void free_rt_sched_group(struct task_group *tg)
8701 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8706 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8710 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8713 #endif /* CONFIG_RT_GROUP_SCHED */
8715 #ifdef CONFIG_GROUP_SCHED
8716 static void free_sched_group(struct task_group *tg)
8718 free_fair_sched_group(tg);
8719 free_rt_sched_group(tg);
8723 /* allocate runqueue etc for a new task group */
8724 struct task_group *sched_create_group(struct task_group *parent)
8726 struct task_group *tg;
8727 unsigned long flags;
8730 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8732 return ERR_PTR(-ENOMEM);
8734 if (!alloc_fair_sched_group(tg, parent))
8737 if (!alloc_rt_sched_group(tg, parent))
8740 spin_lock_irqsave(&task_group_lock, flags);
8741 for_each_possible_cpu(i) {
8742 register_fair_sched_group(tg, i);
8743 register_rt_sched_group(tg, i);
8745 list_add_rcu(&tg->list, &task_groups);
8747 WARN_ON(!parent); /* root should already exist */
8749 tg->parent = parent;
8750 INIT_LIST_HEAD(&tg->children);
8751 list_add_rcu(&tg->siblings, &parent->children);
8752 spin_unlock_irqrestore(&task_group_lock, flags);
8757 free_sched_group(tg);
8758 return ERR_PTR(-ENOMEM);
8761 /* rcu callback to free various structures associated with a task group */
8762 static void free_sched_group_rcu(struct rcu_head *rhp)
8764 /* now it should be safe to free those cfs_rqs */
8765 free_sched_group(container_of(rhp, struct task_group, rcu));
8768 /* Destroy runqueue etc associated with a task group */
8769 void sched_destroy_group(struct task_group *tg)
8771 unsigned long flags;
8774 spin_lock_irqsave(&task_group_lock, flags);
8775 for_each_possible_cpu(i) {
8776 unregister_fair_sched_group(tg, i);
8777 unregister_rt_sched_group(tg, i);
8779 list_del_rcu(&tg->list);
8780 list_del_rcu(&tg->siblings);
8781 spin_unlock_irqrestore(&task_group_lock, flags);
8783 /* wait for possible concurrent references to cfs_rqs complete */
8784 call_rcu(&tg->rcu, free_sched_group_rcu);
8787 /* change task's runqueue when it moves between groups.
8788 * The caller of this function should have put the task in its new group
8789 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8790 * reflect its new group.
8792 void sched_move_task(struct task_struct *tsk)
8795 unsigned long flags;
8798 rq = task_rq_lock(tsk, &flags);
8800 update_rq_clock(rq);
8802 running = task_current(rq, tsk);
8803 on_rq = tsk->se.on_rq;
8806 dequeue_task(rq, tsk, 0);
8807 if (unlikely(running))
8808 tsk->sched_class->put_prev_task(rq, tsk);
8810 set_task_rq(tsk, task_cpu(tsk));
8812 #ifdef CONFIG_FAIR_GROUP_SCHED
8813 if (tsk->sched_class->moved_group)
8814 tsk->sched_class->moved_group(tsk);
8817 if (unlikely(running))
8818 tsk->sched_class->set_curr_task(rq);
8820 enqueue_task(rq, tsk, 0);
8822 task_rq_unlock(rq, &flags);
8824 #endif /* CONFIG_GROUP_SCHED */
8826 #ifdef CONFIG_FAIR_GROUP_SCHED
8827 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8829 struct cfs_rq *cfs_rq = se->cfs_rq;
8834 dequeue_entity(cfs_rq, se, 0);
8836 se->load.weight = shares;
8837 se->load.inv_weight = 0;
8840 enqueue_entity(cfs_rq, se, 0);
8843 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8845 struct cfs_rq *cfs_rq = se->cfs_rq;
8846 struct rq *rq = cfs_rq->rq;
8847 unsigned long flags;
8849 spin_lock_irqsave(&rq->lock, flags);
8850 __set_se_shares(se, shares);
8851 spin_unlock_irqrestore(&rq->lock, flags);
8854 static DEFINE_MUTEX(shares_mutex);
8856 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8859 unsigned long flags;
8862 * We can't change the weight of the root cgroup.
8867 if (shares < MIN_SHARES)
8868 shares = MIN_SHARES;
8869 else if (shares > MAX_SHARES)
8870 shares = MAX_SHARES;
8872 mutex_lock(&shares_mutex);
8873 if (tg->shares == shares)
8876 spin_lock_irqsave(&task_group_lock, flags);
8877 for_each_possible_cpu(i)
8878 unregister_fair_sched_group(tg, i);
8879 list_del_rcu(&tg->siblings);
8880 spin_unlock_irqrestore(&task_group_lock, flags);
8882 /* wait for any ongoing reference to this group to finish */
8883 synchronize_sched();
8886 * Now we are free to modify the group's share on each cpu
8887 * w/o tripping rebalance_share or load_balance_fair.
8889 tg->shares = shares;
8890 for_each_possible_cpu(i) {
8894 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8895 set_se_shares(tg->se[i], shares);
8899 * Enable load balance activity on this group, by inserting it back on
8900 * each cpu's rq->leaf_cfs_rq_list.
8902 spin_lock_irqsave(&task_group_lock, flags);
8903 for_each_possible_cpu(i)
8904 register_fair_sched_group(tg, i);
8905 list_add_rcu(&tg->siblings, &tg->parent->children);
8906 spin_unlock_irqrestore(&task_group_lock, flags);
8908 mutex_unlock(&shares_mutex);
8912 unsigned long sched_group_shares(struct task_group *tg)
8918 #ifdef CONFIG_RT_GROUP_SCHED
8920 * Ensure that the real time constraints are schedulable.
8922 static DEFINE_MUTEX(rt_constraints_mutex);
8924 static unsigned long to_ratio(u64 period, u64 runtime)
8926 if (runtime == RUNTIME_INF)
8929 return div64_u64(runtime << 20, period);
8932 /* Must be called with tasklist_lock held */
8933 static inline int tg_has_rt_tasks(struct task_group *tg)
8935 struct task_struct *g, *p;
8937 do_each_thread(g, p) {
8938 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8940 } while_each_thread(g, p);
8945 struct rt_schedulable_data {
8946 struct task_group *tg;
8951 static int tg_schedulable(struct task_group *tg, void *data)
8953 struct rt_schedulable_data *d = data;
8954 struct task_group *child;
8955 unsigned long total, sum = 0;
8956 u64 period, runtime;
8958 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8959 runtime = tg->rt_bandwidth.rt_runtime;
8962 period = d->rt_period;
8963 runtime = d->rt_runtime;
8967 * Cannot have more runtime than the period.
8969 if (runtime > period && runtime != RUNTIME_INF)
8973 * Ensure we don't starve existing RT tasks.
8975 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8978 total = to_ratio(period, runtime);
8981 * Nobody can have more than the global setting allows.
8983 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8987 * The sum of our children's runtime should not exceed our own.
8989 list_for_each_entry_rcu(child, &tg->children, siblings) {
8990 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8991 runtime = child->rt_bandwidth.rt_runtime;
8993 if (child == d->tg) {
8994 period = d->rt_period;
8995 runtime = d->rt_runtime;
8998 sum += to_ratio(period, runtime);
9007 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
9009 struct rt_schedulable_data data = {
9011 .rt_period = period,
9012 .rt_runtime = runtime,
9015 return walk_tg_tree(tg_schedulable, tg_nop, &data);
9018 static int tg_set_bandwidth(struct task_group *tg,
9019 u64 rt_period, u64 rt_runtime)
9023 mutex_lock(&rt_constraints_mutex);
9024 read_lock(&tasklist_lock);
9025 err = __rt_schedulable(tg, rt_period, rt_runtime);
9029 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9030 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
9031 tg->rt_bandwidth.rt_runtime = rt_runtime;
9033 for_each_possible_cpu(i) {
9034 struct rt_rq *rt_rq = tg->rt_rq[i];
9036 spin_lock(&rt_rq->rt_runtime_lock);
9037 rt_rq->rt_runtime = rt_runtime;
9038 spin_unlock(&rt_rq->rt_runtime_lock);
9040 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9042 read_unlock(&tasklist_lock);
9043 mutex_unlock(&rt_constraints_mutex);
9048 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
9050 u64 rt_runtime, rt_period;
9052 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9053 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
9054 if (rt_runtime_us < 0)
9055 rt_runtime = RUNTIME_INF;
9057 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9060 long sched_group_rt_runtime(struct task_group *tg)
9064 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
9067 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
9068 do_div(rt_runtime_us, NSEC_PER_USEC);
9069 return rt_runtime_us;
9072 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
9074 u64 rt_runtime, rt_period;
9076 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
9077 rt_runtime = tg->rt_bandwidth.rt_runtime;
9082 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9085 long sched_group_rt_period(struct task_group *tg)
9089 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9090 do_div(rt_period_us, NSEC_PER_USEC);
9091 return rt_period_us;
9094 static int sched_rt_global_constraints(void)
9096 u64 runtime, period;
9099 if (sysctl_sched_rt_period <= 0)
9102 runtime = global_rt_runtime();
9103 period = global_rt_period();
9106 * Sanity check on the sysctl variables.
9108 if (runtime > period && runtime != RUNTIME_INF)
9111 mutex_lock(&rt_constraints_mutex);
9112 read_lock(&tasklist_lock);
9113 ret = __rt_schedulable(NULL, 0, 0);
9114 read_unlock(&tasklist_lock);
9115 mutex_unlock(&rt_constraints_mutex);
9119 #else /* !CONFIG_RT_GROUP_SCHED */
9120 static int sched_rt_global_constraints(void)
9122 unsigned long flags;
9125 if (sysctl_sched_rt_period <= 0)
9128 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9129 for_each_possible_cpu(i) {
9130 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9132 spin_lock(&rt_rq->rt_runtime_lock);
9133 rt_rq->rt_runtime = global_rt_runtime();
9134 spin_unlock(&rt_rq->rt_runtime_lock);
9136 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9140 #endif /* CONFIG_RT_GROUP_SCHED */
9142 int sched_rt_handler(struct ctl_table *table, int write,
9143 struct file *filp, void __user *buffer, size_t *lenp,
9147 int old_period, old_runtime;
9148 static DEFINE_MUTEX(mutex);
9151 old_period = sysctl_sched_rt_period;
9152 old_runtime = sysctl_sched_rt_runtime;
9154 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
9156 if (!ret && write) {
9157 ret = sched_rt_global_constraints();
9159 sysctl_sched_rt_period = old_period;
9160 sysctl_sched_rt_runtime = old_runtime;
9162 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9163 def_rt_bandwidth.rt_period =
9164 ns_to_ktime(global_rt_period());
9167 mutex_unlock(&mutex);
9172 #ifdef CONFIG_CGROUP_SCHED
9174 /* return corresponding task_group object of a cgroup */
9175 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9177 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9178 struct task_group, css);
9181 static struct cgroup_subsys_state *
9182 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9184 struct task_group *tg, *parent;
9186 if (!cgrp->parent) {
9187 /* This is early initialization for the top cgroup */
9188 return &init_task_group.css;
9191 parent = cgroup_tg(cgrp->parent);
9192 tg = sched_create_group(parent);
9194 return ERR_PTR(-ENOMEM);
9200 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9202 struct task_group *tg = cgroup_tg(cgrp);
9204 sched_destroy_group(tg);
9208 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9209 struct task_struct *tsk)
9211 #ifdef CONFIG_RT_GROUP_SCHED
9212 /* Don't accept realtime tasks when there is no way for them to run */
9213 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
9216 /* We don't support RT-tasks being in separate groups */
9217 if (tsk->sched_class != &fair_sched_class)
9225 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9226 struct cgroup *old_cont, struct task_struct *tsk)
9228 sched_move_task(tsk);
9231 #ifdef CONFIG_FAIR_GROUP_SCHED
9232 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9235 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9238 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9240 struct task_group *tg = cgroup_tg(cgrp);
9242 return (u64) tg->shares;
9244 #endif /* CONFIG_FAIR_GROUP_SCHED */
9246 #ifdef CONFIG_RT_GROUP_SCHED
9247 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9250 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9253 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9255 return sched_group_rt_runtime(cgroup_tg(cgrp));
9258 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9261 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9264 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9266 return sched_group_rt_period(cgroup_tg(cgrp));
9268 #endif /* CONFIG_RT_GROUP_SCHED */
9270 static struct cftype cpu_files[] = {
9271 #ifdef CONFIG_FAIR_GROUP_SCHED
9274 .read_u64 = cpu_shares_read_u64,
9275 .write_u64 = cpu_shares_write_u64,
9278 #ifdef CONFIG_RT_GROUP_SCHED
9280 .name = "rt_runtime_us",
9281 .read_s64 = cpu_rt_runtime_read,
9282 .write_s64 = cpu_rt_runtime_write,
9285 .name = "rt_period_us",
9286 .read_u64 = cpu_rt_period_read_uint,
9287 .write_u64 = cpu_rt_period_write_uint,
9292 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9294 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9297 struct cgroup_subsys cpu_cgroup_subsys = {
9299 .create = cpu_cgroup_create,
9300 .destroy = cpu_cgroup_destroy,
9301 .can_attach = cpu_cgroup_can_attach,
9302 .attach = cpu_cgroup_attach,
9303 .populate = cpu_cgroup_populate,
9304 .subsys_id = cpu_cgroup_subsys_id,
9308 #endif /* CONFIG_CGROUP_SCHED */
9310 #ifdef CONFIG_CGROUP_CPUACCT
9313 * CPU accounting code for task groups.
9315 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9316 * (balbir@in.ibm.com).
9319 /* track cpu usage of a group of tasks and its child groups */
9321 struct cgroup_subsys_state css;
9322 /* cpuusage holds pointer to a u64-type object on every cpu */
9324 struct cpuacct *parent;
9327 struct cgroup_subsys cpuacct_subsys;
9329 /* return cpu accounting group corresponding to this container */
9330 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9332 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9333 struct cpuacct, css);
9336 /* return cpu accounting group to which this task belongs */
9337 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9339 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9340 struct cpuacct, css);
9343 /* create a new cpu accounting group */
9344 static struct cgroup_subsys_state *cpuacct_create(
9345 struct cgroup_subsys *ss, struct cgroup *cgrp)
9347 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9350 return ERR_PTR(-ENOMEM);
9352 ca->cpuusage = alloc_percpu(u64);
9353 if (!ca->cpuusage) {
9355 return ERR_PTR(-ENOMEM);
9359 ca->parent = cgroup_ca(cgrp->parent);
9364 /* destroy an existing cpu accounting group */
9366 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9368 struct cpuacct *ca = cgroup_ca(cgrp);
9370 free_percpu(ca->cpuusage);
9374 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9376 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9379 #ifndef CONFIG_64BIT
9381 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9383 spin_lock_irq(&cpu_rq(cpu)->lock);
9385 spin_unlock_irq(&cpu_rq(cpu)->lock);
9393 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9395 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9397 #ifndef CONFIG_64BIT
9399 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9401 spin_lock_irq(&cpu_rq(cpu)->lock);
9403 spin_unlock_irq(&cpu_rq(cpu)->lock);
9409 /* return total cpu usage (in nanoseconds) of a group */
9410 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9412 struct cpuacct *ca = cgroup_ca(cgrp);
9413 u64 totalcpuusage = 0;
9416 for_each_present_cpu(i)
9417 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9419 return totalcpuusage;
9422 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9425 struct cpuacct *ca = cgroup_ca(cgrp);
9434 for_each_present_cpu(i)
9435 cpuacct_cpuusage_write(ca, i, 0);
9441 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9444 struct cpuacct *ca = cgroup_ca(cgroup);
9448 for_each_present_cpu(i) {
9449 percpu = cpuacct_cpuusage_read(ca, i);
9450 seq_printf(m, "%llu ", (unsigned long long) percpu);
9452 seq_printf(m, "\n");
9456 static struct cftype files[] = {
9459 .read_u64 = cpuusage_read,
9460 .write_u64 = cpuusage_write,
9463 .name = "usage_percpu",
9464 .read_seq_string = cpuacct_percpu_seq_read,
9469 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9471 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9475 * charge this task's execution time to its accounting group.
9477 * called with rq->lock held.
9479 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9484 if (!cpuacct_subsys.active)
9487 cpu = task_cpu(tsk);
9490 for (; ca; ca = ca->parent) {
9491 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9492 *cpuusage += cputime;
9496 struct cgroup_subsys cpuacct_subsys = {
9498 .create = cpuacct_create,
9499 .destroy = cpuacct_destroy,
9500 .populate = cpuacct_populate,
9501 .subsys_id = cpuacct_subsys_id,
9503 #endif /* CONFIG_CGROUP_CPUACCT */