4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/reciprocal_div.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/bootmem.h>
72 #include <linux/debugfs.h>
73 #include <linux/ctype.h>
74 #include <linux/ftrace.h>
75 #include <trace/sched.h>
78 #include <asm/irq_regs.h>
80 #include "sched_cpupri.h"
83 * Convert user-nice values [ -20 ... 0 ... 19 ]
84 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
87 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
88 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
89 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
92 * 'User priority' is the nice value converted to something we
93 * can work with better when scaling various scheduler parameters,
94 * it's a [ 0 ... 39 ] range.
96 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
97 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
98 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
101 * Helpers for converting nanosecond timing to jiffy resolution
103 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
105 #define NICE_0_LOAD SCHED_LOAD_SCALE
106 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
109 * These are the 'tuning knobs' of the scheduler:
111 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
112 * Timeslices get refilled after they expire.
114 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * single value that denotes runtime == period, ie unlimited time.
119 #define RUNTIME_INF ((u64)~0ULL)
121 DEFINE_TRACE(sched_wait_task);
122 DEFINE_TRACE(sched_wakeup);
123 DEFINE_TRACE(sched_wakeup_new);
124 DEFINE_TRACE(sched_switch);
125 DEFINE_TRACE(sched_migrate_task);
129 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
130 * Since cpu_power is a 'constant', we can use a reciprocal divide.
132 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
134 return reciprocal_divide(load, sg->reciprocal_cpu_power);
138 * Each time a sched group cpu_power is changed,
139 * we must compute its reciprocal value
141 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
143 sg->__cpu_power += val;
144 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
148 static inline int rt_policy(int policy)
150 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
155 static inline int task_has_rt_policy(struct task_struct *p)
157 return rt_policy(p->policy);
161 * This is the priority-queue data structure of the RT scheduling class:
163 struct rt_prio_array {
164 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
165 struct list_head queue[MAX_RT_PRIO];
168 struct rt_bandwidth {
169 /* nests inside the rq lock: */
170 spinlock_t rt_runtime_lock;
173 struct hrtimer rt_period_timer;
176 static struct rt_bandwidth def_rt_bandwidth;
178 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
180 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
182 struct rt_bandwidth *rt_b =
183 container_of(timer, struct rt_bandwidth, rt_period_timer);
189 now = hrtimer_cb_get_time(timer);
190 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
195 idle = do_sched_rt_period_timer(rt_b, overrun);
198 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
202 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
204 rt_b->rt_period = ns_to_ktime(period);
205 rt_b->rt_runtime = runtime;
207 spin_lock_init(&rt_b->rt_runtime_lock);
209 hrtimer_init(&rt_b->rt_period_timer,
210 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
211 rt_b->rt_period_timer.function = sched_rt_period_timer;
212 rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_UNLOCKED;
215 static inline int rt_bandwidth_enabled(void)
217 return sysctl_sched_rt_runtime >= 0;
220 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
224 if (rt_bandwidth_enabled() && rt_b->rt_runtime == RUNTIME_INF)
227 if (hrtimer_active(&rt_b->rt_period_timer))
230 spin_lock(&rt_b->rt_runtime_lock);
232 if (hrtimer_active(&rt_b->rt_period_timer))
235 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
236 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
237 hrtimer_start_expires(&rt_b->rt_period_timer,
240 spin_unlock(&rt_b->rt_runtime_lock);
243 #ifdef CONFIG_RT_GROUP_SCHED
244 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
246 hrtimer_cancel(&rt_b->rt_period_timer);
251 * sched_domains_mutex serializes calls to arch_init_sched_domains,
252 * detach_destroy_domains and partition_sched_domains.
254 static DEFINE_MUTEX(sched_domains_mutex);
256 #ifdef CONFIG_GROUP_SCHED
258 #include <linux/cgroup.h>
262 static LIST_HEAD(task_groups);
264 /* task group related information */
266 #ifdef CONFIG_CGROUP_SCHED
267 struct cgroup_subsys_state css;
270 #ifdef CONFIG_FAIR_GROUP_SCHED
271 /* schedulable entities of this group on each cpu */
272 struct sched_entity **se;
273 /* runqueue "owned" by this group on each cpu */
274 struct cfs_rq **cfs_rq;
275 unsigned long shares;
278 #ifdef CONFIG_RT_GROUP_SCHED
279 struct sched_rt_entity **rt_se;
280 struct rt_rq **rt_rq;
282 struct rt_bandwidth rt_bandwidth;
286 struct list_head list;
288 struct task_group *parent;
289 struct list_head siblings;
290 struct list_head children;
293 #ifdef CONFIG_USER_SCHED
297 * Every UID task group (including init_task_group aka UID-0) will
298 * be a child to this group.
300 struct task_group root_task_group;
302 #ifdef CONFIG_FAIR_GROUP_SCHED
303 /* Default task group's sched entity on each cpu */
304 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
305 /* Default task group's cfs_rq on each cpu */
306 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
307 #endif /* CONFIG_FAIR_GROUP_SCHED */
309 #ifdef CONFIG_RT_GROUP_SCHED
310 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
311 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
312 #endif /* CONFIG_RT_GROUP_SCHED */
313 #else /* !CONFIG_USER_SCHED */
314 #define root_task_group init_task_group
315 #endif /* CONFIG_USER_SCHED */
317 /* task_group_lock serializes add/remove of task groups and also changes to
318 * a task group's cpu shares.
320 static DEFINE_SPINLOCK(task_group_lock);
322 #ifdef CONFIG_FAIR_GROUP_SCHED
323 #ifdef CONFIG_USER_SCHED
324 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
325 #else /* !CONFIG_USER_SCHED */
326 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
327 #endif /* CONFIG_USER_SCHED */
330 * A weight of 0 or 1 can cause arithmetics problems.
331 * A weight of a cfs_rq is the sum of weights of which entities
332 * are queued on this cfs_rq, so a weight of a entity should not be
333 * too large, so as the shares value of a task group.
334 * (The default weight is 1024 - so there's no practical
335 * limitation from this.)
338 #define MAX_SHARES (1UL << 18)
340 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
343 /* Default task group.
344 * Every task in system belong to this group at bootup.
346 struct task_group init_task_group;
348 /* return group to which a task belongs */
349 static inline struct task_group *task_group(struct task_struct *p)
351 struct task_group *tg;
353 #ifdef CONFIG_USER_SCHED
355 #elif defined(CONFIG_CGROUP_SCHED)
356 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
357 struct task_group, css);
359 tg = &init_task_group;
364 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
365 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
367 #ifdef CONFIG_FAIR_GROUP_SCHED
368 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
369 p->se.parent = task_group(p)->se[cpu];
372 #ifdef CONFIG_RT_GROUP_SCHED
373 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
374 p->rt.parent = task_group(p)->rt_se[cpu];
380 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
381 static inline struct task_group *task_group(struct task_struct *p)
386 #endif /* CONFIG_GROUP_SCHED */
388 /* CFS-related fields in a runqueue */
390 struct load_weight load;
391 unsigned long nr_running;
396 struct rb_root tasks_timeline;
397 struct rb_node *rb_leftmost;
399 struct list_head tasks;
400 struct list_head *balance_iterator;
403 * 'curr' points to currently running entity on this cfs_rq.
404 * It is set to NULL otherwise (i.e when none are currently running).
406 struct sched_entity *curr, *next, *last;
408 unsigned int nr_spread_over;
410 #ifdef CONFIG_FAIR_GROUP_SCHED
411 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
414 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
415 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
416 * (like users, containers etc.)
418 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
419 * list is used during load balance.
421 struct list_head leaf_cfs_rq_list;
422 struct task_group *tg; /* group that "owns" this runqueue */
426 * the part of load.weight contributed by tasks
428 unsigned long task_weight;
431 * h_load = weight * f(tg)
433 * Where f(tg) is the recursive weight fraction assigned to
436 unsigned long h_load;
439 * this cpu's part of tg->shares
441 unsigned long shares;
444 * load.weight at the time we set shares
446 unsigned long rq_weight;
451 /* Real-Time classes' related field in a runqueue: */
453 struct rt_prio_array active;
454 unsigned long rt_nr_running;
455 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
456 int highest_prio; /* highest queued rt task prio */
459 unsigned long rt_nr_migratory;
465 /* Nests inside the rq lock: */
466 spinlock_t rt_runtime_lock;
468 #ifdef CONFIG_RT_GROUP_SCHED
469 unsigned long rt_nr_boosted;
472 struct list_head leaf_rt_rq_list;
473 struct task_group *tg;
474 struct sched_rt_entity *rt_se;
481 * We add the notion of a root-domain which will be used to define per-domain
482 * variables. Each exclusive cpuset essentially defines an island domain by
483 * fully partitioning the member cpus from any other cpuset. Whenever a new
484 * exclusive cpuset is created, we also create and attach a new root-domain
494 * The "RT overload" flag: it gets set if a CPU has more than
495 * one runnable RT task.
500 struct cpupri cpupri;
505 * By default the system creates a single root-domain with all cpus as
506 * members (mimicking the global state we have today).
508 static struct root_domain def_root_domain;
513 * This is the main, per-CPU runqueue data structure.
515 * Locking rule: those places that want to lock multiple runqueues
516 * (such as the load balancing or the thread migration code), lock
517 * acquire operations must be ordered by ascending &runqueue.
524 * nr_running and cpu_load should be in the same cacheline because
525 * remote CPUs use both these fields when doing load calculation.
527 unsigned long nr_running;
528 #define CPU_LOAD_IDX_MAX 5
529 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
530 unsigned char idle_at_tick;
532 unsigned long last_tick_seen;
533 unsigned char in_nohz_recently;
535 /* capture load from *all* tasks on this cpu: */
536 struct load_weight load;
537 unsigned long nr_load_updates;
543 #ifdef CONFIG_FAIR_GROUP_SCHED
544 /* list of leaf cfs_rq on this cpu: */
545 struct list_head leaf_cfs_rq_list;
547 #ifdef CONFIG_RT_GROUP_SCHED
548 struct list_head leaf_rt_rq_list;
552 * This is part of a global counter where only the total sum
553 * over all CPUs matters. A task can increase this counter on
554 * one CPU and if it got migrated afterwards it may decrease
555 * it on another CPU. Always updated under the runqueue lock:
557 unsigned long nr_uninterruptible;
559 struct task_struct *curr, *idle;
560 unsigned long next_balance;
561 struct mm_struct *prev_mm;
568 struct root_domain *rd;
569 struct sched_domain *sd;
571 /* For active balancing */
574 /* cpu of this runqueue: */
578 unsigned long avg_load_per_task;
580 struct task_struct *migration_thread;
581 struct list_head migration_queue;
584 #ifdef CONFIG_SCHED_HRTICK
586 int hrtick_csd_pending;
587 struct call_single_data hrtick_csd;
589 struct hrtimer hrtick_timer;
592 #ifdef CONFIG_SCHEDSTATS
594 struct sched_info rq_sched_info;
596 /* sys_sched_yield() stats */
597 unsigned int yld_exp_empty;
598 unsigned int yld_act_empty;
599 unsigned int yld_both_empty;
600 unsigned int yld_count;
602 /* schedule() stats */
603 unsigned int sched_switch;
604 unsigned int sched_count;
605 unsigned int sched_goidle;
607 /* try_to_wake_up() stats */
608 unsigned int ttwu_count;
609 unsigned int ttwu_local;
612 unsigned int bkl_count;
616 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
618 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
620 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
623 static inline int cpu_of(struct rq *rq)
633 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
634 * See detach_destroy_domains: synchronize_sched for details.
636 * The domain tree of any CPU may only be accessed from within
637 * preempt-disabled sections.
639 #define for_each_domain(cpu, __sd) \
640 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
642 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
643 #define this_rq() (&__get_cpu_var(runqueues))
644 #define task_rq(p) cpu_rq(task_cpu(p))
645 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
647 static inline void update_rq_clock(struct rq *rq)
649 rq->clock = sched_clock_cpu(cpu_of(rq));
653 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
655 #ifdef CONFIG_SCHED_DEBUG
656 # define const_debug __read_mostly
658 # define const_debug static const
664 * Returns true if the current cpu runqueue is locked.
665 * This interface allows printk to be called with the runqueue lock
666 * held and know whether or not it is OK to wake up the klogd.
668 int runqueue_is_locked(void)
671 struct rq *rq = cpu_rq(cpu);
674 ret = spin_is_locked(&rq->lock);
680 * Debugging: various feature bits
683 #define SCHED_FEAT(name, enabled) \
684 __SCHED_FEAT_##name ,
687 #include "sched_features.h"
692 #define SCHED_FEAT(name, enabled) \
693 (1UL << __SCHED_FEAT_##name) * enabled |
695 const_debug unsigned int sysctl_sched_features =
696 #include "sched_features.h"
701 #ifdef CONFIG_SCHED_DEBUG
702 #define SCHED_FEAT(name, enabled) \
705 static __read_mostly char *sched_feat_names[] = {
706 #include "sched_features.h"
712 static int sched_feat_open(struct inode *inode, struct file *filp)
714 filp->private_data = inode->i_private;
719 sched_feat_read(struct file *filp, char __user *ubuf,
720 size_t cnt, loff_t *ppos)
727 for (i = 0; sched_feat_names[i]; i++) {
728 len += strlen(sched_feat_names[i]);
732 buf = kmalloc(len + 2, GFP_KERNEL);
736 for (i = 0; sched_feat_names[i]; i++) {
737 if (sysctl_sched_features & (1UL << i))
738 r += sprintf(buf + r, "%s ", sched_feat_names[i]);
740 r += sprintf(buf + r, "NO_%s ", sched_feat_names[i]);
743 r += sprintf(buf + r, "\n");
744 WARN_ON(r >= len + 2);
746 r = simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
754 sched_feat_write(struct file *filp, const char __user *ubuf,
755 size_t cnt, loff_t *ppos)
765 if (copy_from_user(&buf, ubuf, cnt))
770 if (strncmp(buf, "NO_", 3) == 0) {
775 for (i = 0; sched_feat_names[i]; i++) {
776 int len = strlen(sched_feat_names[i]);
778 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
780 sysctl_sched_features &= ~(1UL << i);
782 sysctl_sched_features |= (1UL << i);
787 if (!sched_feat_names[i])
795 static struct file_operations sched_feat_fops = {
796 .open = sched_feat_open,
797 .read = sched_feat_read,
798 .write = sched_feat_write,
801 static __init int sched_init_debug(void)
803 debugfs_create_file("sched_features", 0644, NULL, NULL,
808 late_initcall(sched_init_debug);
812 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
815 * Number of tasks to iterate in a single balance run.
816 * Limited because this is done with IRQs disabled.
818 const_debug unsigned int sysctl_sched_nr_migrate = 32;
821 * ratelimit for updating the group shares.
824 unsigned int sysctl_sched_shares_ratelimit = 250000;
827 * Inject some fuzzyness into changing the per-cpu group shares
828 * this avoids remote rq-locks at the expense of fairness.
831 unsigned int sysctl_sched_shares_thresh = 4;
834 * period over which we measure -rt task cpu usage in us.
837 unsigned int sysctl_sched_rt_period = 1000000;
839 static __read_mostly int scheduler_running;
842 * part of the period that we allow rt tasks to run in us.
845 int sysctl_sched_rt_runtime = 950000;
847 static inline u64 global_rt_period(void)
849 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
852 static inline u64 global_rt_runtime(void)
854 if (sysctl_sched_rt_runtime < 0)
857 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
860 #ifndef prepare_arch_switch
861 # define prepare_arch_switch(next) do { } while (0)
863 #ifndef finish_arch_switch
864 # define finish_arch_switch(prev) do { } while (0)
867 static inline int task_current(struct rq *rq, struct task_struct *p)
869 return rq->curr == p;
872 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
873 static inline int task_running(struct rq *rq, struct task_struct *p)
875 return task_current(rq, p);
878 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
882 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
884 #ifdef CONFIG_DEBUG_SPINLOCK
885 /* this is a valid case when another task releases the spinlock */
886 rq->lock.owner = current;
889 * If we are tracking spinlock dependencies then we have to
890 * fix up the runqueue lock - which gets 'carried over' from
893 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
895 spin_unlock_irq(&rq->lock);
898 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
899 static inline int task_running(struct rq *rq, struct task_struct *p)
904 return task_current(rq, p);
908 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
912 * We can optimise this out completely for !SMP, because the
913 * SMP rebalancing from interrupt is the only thing that cares
918 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
919 spin_unlock_irq(&rq->lock);
921 spin_unlock(&rq->lock);
925 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
929 * After ->oncpu is cleared, the task can be moved to a different CPU.
930 * We must ensure this doesn't happen until the switch is completely
936 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
940 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
943 * __task_rq_lock - lock the runqueue a given task resides on.
944 * Must be called interrupts disabled.
946 static inline struct rq *__task_rq_lock(struct task_struct *p)
950 struct rq *rq = task_rq(p);
951 spin_lock(&rq->lock);
952 if (likely(rq == task_rq(p)))
954 spin_unlock(&rq->lock);
959 * task_rq_lock - lock the runqueue a given task resides on and disable
960 * interrupts. Note the ordering: we can safely lookup the task_rq without
961 * explicitly disabling preemption.
963 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
969 local_irq_save(*flags);
971 spin_lock(&rq->lock);
972 if (likely(rq == task_rq(p)))
974 spin_unlock_irqrestore(&rq->lock, *flags);
978 void task_rq_unlock_wait(struct task_struct *p)
980 struct rq *rq = task_rq(p);
982 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
983 spin_unlock_wait(&rq->lock);
986 static void __task_rq_unlock(struct rq *rq)
989 spin_unlock(&rq->lock);
992 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
995 spin_unlock_irqrestore(&rq->lock, *flags);
999 * this_rq_lock - lock this runqueue and disable interrupts.
1001 static struct rq *this_rq_lock(void)
1002 __acquires(rq->lock)
1006 local_irq_disable();
1008 spin_lock(&rq->lock);
1013 #ifdef CONFIG_SCHED_HRTICK
1015 * Use HR-timers to deliver accurate preemption points.
1017 * Its all a bit involved since we cannot program an hrt while holding the
1018 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1021 * When we get rescheduled we reprogram the hrtick_timer outside of the
1027 * - enabled by features
1028 * - hrtimer is actually high res
1030 static inline int hrtick_enabled(struct rq *rq)
1032 if (!sched_feat(HRTICK))
1034 if (!cpu_active(cpu_of(rq)))
1036 return hrtimer_is_hres_active(&rq->hrtick_timer);
1039 static void hrtick_clear(struct rq *rq)
1041 if (hrtimer_active(&rq->hrtick_timer))
1042 hrtimer_cancel(&rq->hrtick_timer);
1046 * High-resolution timer tick.
1047 * Runs from hardirq context with interrupts disabled.
1049 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1051 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1053 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1055 spin_lock(&rq->lock);
1056 update_rq_clock(rq);
1057 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1058 spin_unlock(&rq->lock);
1060 return HRTIMER_NORESTART;
1065 * called from hardirq (IPI) context
1067 static void __hrtick_start(void *arg)
1069 struct rq *rq = arg;
1071 spin_lock(&rq->lock);
1072 hrtimer_restart(&rq->hrtick_timer);
1073 rq->hrtick_csd_pending = 0;
1074 spin_unlock(&rq->lock);
1078 * Called to set the hrtick timer state.
1080 * called with rq->lock held and irqs disabled
1082 static void hrtick_start(struct rq *rq, u64 delay)
1084 struct hrtimer *timer = &rq->hrtick_timer;
1085 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1087 hrtimer_set_expires(timer, time);
1089 if (rq == this_rq()) {
1090 hrtimer_restart(timer);
1091 } else if (!rq->hrtick_csd_pending) {
1092 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd);
1093 rq->hrtick_csd_pending = 1;
1098 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1100 int cpu = (int)(long)hcpu;
1103 case CPU_UP_CANCELED:
1104 case CPU_UP_CANCELED_FROZEN:
1105 case CPU_DOWN_PREPARE:
1106 case CPU_DOWN_PREPARE_FROZEN:
1108 case CPU_DEAD_FROZEN:
1109 hrtick_clear(cpu_rq(cpu));
1116 static __init void init_hrtick(void)
1118 hotcpu_notifier(hotplug_hrtick, 0);
1122 * Called to set the hrtick timer state.
1124 * called with rq->lock held and irqs disabled
1126 static void hrtick_start(struct rq *rq, u64 delay)
1128 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay), HRTIMER_MODE_REL);
1131 static inline void init_hrtick(void)
1134 #endif /* CONFIG_SMP */
1136 static void init_rq_hrtick(struct rq *rq)
1139 rq->hrtick_csd_pending = 0;
1141 rq->hrtick_csd.flags = 0;
1142 rq->hrtick_csd.func = __hrtick_start;
1143 rq->hrtick_csd.info = rq;
1146 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1147 rq->hrtick_timer.function = hrtick;
1148 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_PERCPU;
1150 #else /* CONFIG_SCHED_HRTICK */
1151 static inline void hrtick_clear(struct rq *rq)
1155 static inline void init_rq_hrtick(struct rq *rq)
1159 static inline void init_hrtick(void)
1162 #endif /* CONFIG_SCHED_HRTICK */
1165 * resched_task - mark a task 'to be rescheduled now'.
1167 * On UP this means the setting of the need_resched flag, on SMP it
1168 * might also involve a cross-CPU call to trigger the scheduler on
1173 #ifndef tsk_is_polling
1174 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1177 static void resched_task(struct task_struct *p)
1181 assert_spin_locked(&task_rq(p)->lock);
1183 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1186 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1189 if (cpu == smp_processor_id())
1192 /* NEED_RESCHED must be visible before we test polling */
1194 if (!tsk_is_polling(p))
1195 smp_send_reschedule(cpu);
1198 static void resched_cpu(int cpu)
1200 struct rq *rq = cpu_rq(cpu);
1201 unsigned long flags;
1203 if (!spin_trylock_irqsave(&rq->lock, flags))
1205 resched_task(cpu_curr(cpu));
1206 spin_unlock_irqrestore(&rq->lock, flags);
1211 * When add_timer_on() enqueues a timer into the timer wheel of an
1212 * idle CPU then this timer might expire before the next timer event
1213 * which is scheduled to wake up that CPU. In case of a completely
1214 * idle system the next event might even be infinite time into the
1215 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1216 * leaves the inner idle loop so the newly added timer is taken into
1217 * account when the CPU goes back to idle and evaluates the timer
1218 * wheel for the next timer event.
1220 void wake_up_idle_cpu(int cpu)
1222 struct rq *rq = cpu_rq(cpu);
1224 if (cpu == smp_processor_id())
1228 * This is safe, as this function is called with the timer
1229 * wheel base lock of (cpu) held. When the CPU is on the way
1230 * to idle and has not yet set rq->curr to idle then it will
1231 * be serialized on the timer wheel base lock and take the new
1232 * timer into account automatically.
1234 if (rq->curr != rq->idle)
1238 * We can set TIF_RESCHED on the idle task of the other CPU
1239 * lockless. The worst case is that the other CPU runs the
1240 * idle task through an additional NOOP schedule()
1242 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1244 /* NEED_RESCHED must be visible before we test polling */
1246 if (!tsk_is_polling(rq->idle))
1247 smp_send_reschedule(cpu);
1249 #endif /* CONFIG_NO_HZ */
1251 #else /* !CONFIG_SMP */
1252 static void resched_task(struct task_struct *p)
1254 assert_spin_locked(&task_rq(p)->lock);
1255 set_tsk_need_resched(p);
1257 #endif /* CONFIG_SMP */
1259 #if BITS_PER_LONG == 32
1260 # define WMULT_CONST (~0UL)
1262 # define WMULT_CONST (1UL << 32)
1265 #define WMULT_SHIFT 32
1268 * Shift right and round:
1270 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1273 * delta *= weight / lw
1275 static unsigned long
1276 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1277 struct load_weight *lw)
1281 if (!lw->inv_weight) {
1282 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1285 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1289 tmp = (u64)delta_exec * weight;
1291 * Check whether we'd overflow the 64-bit multiplication:
1293 if (unlikely(tmp > WMULT_CONST))
1294 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1297 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1299 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1302 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1308 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1315 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1316 * of tasks with abnormal "nice" values across CPUs the contribution that
1317 * each task makes to its run queue's load is weighted according to its
1318 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1319 * scaled version of the new time slice allocation that they receive on time
1323 #define WEIGHT_IDLEPRIO 2
1324 #define WMULT_IDLEPRIO (1 << 31)
1327 * Nice levels are multiplicative, with a gentle 10% change for every
1328 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1329 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1330 * that remained on nice 0.
1332 * The "10% effect" is relative and cumulative: from _any_ nice level,
1333 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1334 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1335 * If a task goes up by ~10% and another task goes down by ~10% then
1336 * the relative distance between them is ~25%.)
1338 static const int prio_to_weight[40] = {
1339 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1340 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1341 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1342 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1343 /* 0 */ 1024, 820, 655, 526, 423,
1344 /* 5 */ 335, 272, 215, 172, 137,
1345 /* 10 */ 110, 87, 70, 56, 45,
1346 /* 15 */ 36, 29, 23, 18, 15,
1350 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1352 * In cases where the weight does not change often, we can use the
1353 * precalculated inverse to speed up arithmetics by turning divisions
1354 * into multiplications:
1356 static const u32 prio_to_wmult[40] = {
1357 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1358 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1359 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1360 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1361 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1362 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1363 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1364 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1367 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1370 * runqueue iterator, to support SMP load-balancing between different
1371 * scheduling classes, without having to expose their internal data
1372 * structures to the load-balancing proper:
1374 struct rq_iterator {
1376 struct task_struct *(*start)(void *);
1377 struct task_struct *(*next)(void *);
1381 static unsigned long
1382 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1383 unsigned long max_load_move, struct sched_domain *sd,
1384 enum cpu_idle_type idle, int *all_pinned,
1385 int *this_best_prio, struct rq_iterator *iterator);
1388 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1389 struct sched_domain *sd, enum cpu_idle_type idle,
1390 struct rq_iterator *iterator);
1393 #ifdef CONFIG_CGROUP_CPUACCT
1394 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1396 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1399 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1401 update_load_add(&rq->load, load);
1404 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1406 update_load_sub(&rq->load, load);
1409 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1410 typedef int (*tg_visitor)(struct task_group *, void *);
1413 * Iterate the full tree, calling @down when first entering a node and @up when
1414 * leaving it for the final time.
1416 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1418 struct task_group *parent, *child;
1422 parent = &root_task_group;
1424 ret = (*down)(parent, data);
1427 list_for_each_entry_rcu(child, &parent->children, siblings) {
1434 ret = (*up)(parent, data);
1439 parent = parent->parent;
1448 static int tg_nop(struct task_group *tg, void *data)
1455 static unsigned long source_load(int cpu, int type);
1456 static unsigned long target_load(int cpu, int type);
1457 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1459 static unsigned long cpu_avg_load_per_task(int cpu)
1461 struct rq *rq = cpu_rq(cpu);
1462 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1465 rq->avg_load_per_task = rq->load.weight / nr_running;
1467 rq->avg_load_per_task = 0;
1469 return rq->avg_load_per_task;
1472 #ifdef CONFIG_FAIR_GROUP_SCHED
1474 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1477 * Calculate and set the cpu's group shares.
1480 update_group_shares_cpu(struct task_group *tg, int cpu,
1481 unsigned long sd_shares, unsigned long sd_rq_weight)
1484 unsigned long shares;
1485 unsigned long rq_weight;
1490 rq_weight = tg->cfs_rq[cpu]->load.weight;
1493 * If there are currently no tasks on the cpu pretend there is one of
1494 * average load so that when a new task gets to run here it will not
1495 * get delayed by group starvation.
1499 rq_weight = NICE_0_LOAD;
1502 if (unlikely(rq_weight > sd_rq_weight))
1503 rq_weight = sd_rq_weight;
1506 * \Sum shares * rq_weight
1507 * shares = -----------------------
1511 shares = (sd_shares * rq_weight) / (sd_rq_weight + 1);
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);
1521 * record the actual number of shares, not the boosted amount.
1523 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1524 tg->cfs_rq[cpu]->rq_weight = rq_weight;
1526 __set_se_shares(tg->se[cpu], shares);
1527 spin_unlock_irqrestore(&rq->lock, flags);
1532 * Re-compute the task group their per cpu shares over the given domain.
1533 * This needs to be done in a bottom-up fashion because the rq weight of a
1534 * parent group depends on the shares of its child groups.
1536 static int tg_shares_up(struct task_group *tg, void *data)
1538 unsigned long rq_weight = 0;
1539 unsigned long shares = 0;
1540 struct sched_domain *sd = data;
1543 for_each_cpu_mask(i, sd->span) {
1544 rq_weight += tg->cfs_rq[i]->load.weight;
1545 shares += tg->cfs_rq[i]->shares;
1548 if ((!shares && rq_weight) || shares > tg->shares)
1549 shares = tg->shares;
1551 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1552 shares = tg->shares;
1555 rq_weight = cpus_weight(sd->span) * NICE_0_LOAD;
1557 for_each_cpu_mask(i, sd->span)
1558 update_group_shares_cpu(tg, i, shares, rq_weight);
1564 * Compute the cpu's hierarchical load factor for each task group.
1565 * This needs to be done in a top-down fashion because the load of a child
1566 * group is a fraction of its parents load.
1568 static int tg_load_down(struct task_group *tg, void *data)
1571 long cpu = (long)data;
1574 load = cpu_rq(cpu)->load.weight;
1576 load = tg->parent->cfs_rq[cpu]->h_load;
1577 load *= tg->cfs_rq[cpu]->shares;
1578 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1581 tg->cfs_rq[cpu]->h_load = load;
1586 static void update_shares(struct sched_domain *sd)
1588 u64 now = cpu_clock(raw_smp_processor_id());
1589 s64 elapsed = now - sd->last_update;
1591 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1592 sd->last_update = now;
1593 walk_tg_tree(tg_nop, tg_shares_up, sd);
1597 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1599 spin_unlock(&rq->lock);
1601 spin_lock(&rq->lock);
1604 static void update_h_load(long cpu)
1606 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1611 static inline void update_shares(struct sched_domain *sd)
1615 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1623 #ifdef CONFIG_FAIR_GROUP_SCHED
1624 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1627 cfs_rq->shares = shares;
1632 #include "sched_stats.h"
1633 #include "sched_idletask.c"
1634 #include "sched_fair.c"
1635 #include "sched_rt.c"
1636 #ifdef CONFIG_SCHED_DEBUG
1637 # include "sched_debug.c"
1640 #define sched_class_highest (&rt_sched_class)
1641 #define for_each_class(class) \
1642 for (class = sched_class_highest; class; class = class->next)
1644 static void inc_nr_running(struct rq *rq)
1649 static void dec_nr_running(struct rq *rq)
1654 static void set_load_weight(struct task_struct *p)
1656 if (task_has_rt_policy(p)) {
1657 p->se.load.weight = prio_to_weight[0] * 2;
1658 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1663 * SCHED_IDLE tasks get minimal weight:
1665 if (p->policy == SCHED_IDLE) {
1666 p->se.load.weight = WEIGHT_IDLEPRIO;
1667 p->se.load.inv_weight = WMULT_IDLEPRIO;
1671 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1672 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1675 static void update_avg(u64 *avg, u64 sample)
1677 s64 diff = sample - *avg;
1681 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1683 sched_info_queued(p);
1684 p->sched_class->enqueue_task(rq, p, wakeup);
1688 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1690 if (sleep && p->se.last_wakeup) {
1691 update_avg(&p->se.avg_overlap,
1692 p->se.sum_exec_runtime - p->se.last_wakeup);
1693 p->se.last_wakeup = 0;
1696 sched_info_dequeued(p);
1697 p->sched_class->dequeue_task(rq, p, sleep);
1702 * __normal_prio - return the priority that is based on the static prio
1704 static inline int __normal_prio(struct task_struct *p)
1706 return p->static_prio;
1710 * Calculate the expected normal priority: i.e. priority
1711 * without taking RT-inheritance into account. Might be
1712 * boosted by interactivity modifiers. Changes upon fork,
1713 * setprio syscalls, and whenever the interactivity
1714 * estimator recalculates.
1716 static inline int normal_prio(struct task_struct *p)
1720 if (task_has_rt_policy(p))
1721 prio = MAX_RT_PRIO-1 - p->rt_priority;
1723 prio = __normal_prio(p);
1728 * Calculate the current priority, i.e. the priority
1729 * taken into account by the scheduler. This value might
1730 * be boosted by RT tasks, or might be boosted by
1731 * interactivity modifiers. Will be RT if the task got
1732 * RT-boosted. If not then it returns p->normal_prio.
1734 static int effective_prio(struct task_struct *p)
1736 p->normal_prio = normal_prio(p);
1738 * If we are RT tasks or we were boosted to RT priority,
1739 * keep the priority unchanged. Otherwise, update priority
1740 * to the normal priority:
1742 if (!rt_prio(p->prio))
1743 return p->normal_prio;
1748 * activate_task - move a task to the runqueue.
1750 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1752 if (task_contributes_to_load(p))
1753 rq->nr_uninterruptible--;
1755 enqueue_task(rq, p, wakeup);
1760 * deactivate_task - remove a task from the runqueue.
1762 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1764 if (task_contributes_to_load(p))
1765 rq->nr_uninterruptible++;
1767 dequeue_task(rq, p, sleep);
1772 * task_curr - is this task currently executing on a CPU?
1773 * @p: the task in question.
1775 inline int task_curr(const struct task_struct *p)
1777 return cpu_curr(task_cpu(p)) == p;
1780 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1782 set_task_rq(p, cpu);
1785 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1786 * successfuly executed on another CPU. We must ensure that updates of
1787 * per-task data have been completed by this moment.
1790 task_thread_info(p)->cpu = cpu;
1794 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1795 const struct sched_class *prev_class,
1796 int oldprio, int running)
1798 if (prev_class != p->sched_class) {
1799 if (prev_class->switched_from)
1800 prev_class->switched_from(rq, p, running);
1801 p->sched_class->switched_to(rq, p, running);
1803 p->sched_class->prio_changed(rq, p, oldprio, running);
1808 /* Used instead of source_load when we know the type == 0 */
1809 static unsigned long weighted_cpuload(const int cpu)
1811 return cpu_rq(cpu)->load.weight;
1815 * Is this task likely cache-hot:
1818 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1823 * Buddy candidates are cache hot:
1825 if (sched_feat(CACHE_HOT_BUDDY) &&
1826 (&p->se == cfs_rq_of(&p->se)->next ||
1827 &p->se == cfs_rq_of(&p->se)->last))
1830 if (p->sched_class != &fair_sched_class)
1833 if (sysctl_sched_migration_cost == -1)
1835 if (sysctl_sched_migration_cost == 0)
1838 delta = now - p->se.exec_start;
1840 return delta < (s64)sysctl_sched_migration_cost;
1844 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1846 int old_cpu = task_cpu(p);
1847 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1848 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1849 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1852 clock_offset = old_rq->clock - new_rq->clock;
1854 #ifdef CONFIG_SCHEDSTATS
1855 if (p->se.wait_start)
1856 p->se.wait_start -= clock_offset;
1857 if (p->se.sleep_start)
1858 p->se.sleep_start -= clock_offset;
1859 if (p->se.block_start)
1860 p->se.block_start -= clock_offset;
1861 if (old_cpu != new_cpu) {
1862 schedstat_inc(p, se.nr_migrations);
1863 if (task_hot(p, old_rq->clock, NULL))
1864 schedstat_inc(p, se.nr_forced2_migrations);
1867 p->se.vruntime -= old_cfsrq->min_vruntime -
1868 new_cfsrq->min_vruntime;
1870 __set_task_cpu(p, new_cpu);
1873 struct migration_req {
1874 struct list_head list;
1876 struct task_struct *task;
1879 struct completion done;
1883 * The task's runqueue lock must be held.
1884 * Returns true if you have to wait for migration thread.
1887 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1889 struct rq *rq = task_rq(p);
1892 * If the task is not on a runqueue (and not running), then
1893 * it is sufficient to simply update the task's cpu field.
1895 if (!p->se.on_rq && !task_running(rq, p)) {
1896 set_task_cpu(p, dest_cpu);
1900 init_completion(&req->done);
1902 req->dest_cpu = dest_cpu;
1903 list_add(&req->list, &rq->migration_queue);
1909 * wait_task_inactive - wait for a thread to unschedule.
1911 * If @match_state is nonzero, it's the @p->state value just checked and
1912 * not expected to change. If it changes, i.e. @p might have woken up,
1913 * then return zero. When we succeed in waiting for @p to be off its CPU,
1914 * we return a positive number (its total switch count). If a second call
1915 * a short while later returns the same number, the caller can be sure that
1916 * @p has remained unscheduled the whole time.
1918 * The caller must ensure that the task *will* unschedule sometime soon,
1919 * else this function might spin for a *long* time. This function can't
1920 * be called with interrupts off, or it may introduce deadlock with
1921 * smp_call_function() if an IPI is sent by the same process we are
1922 * waiting to become inactive.
1924 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1926 unsigned long flags;
1933 * We do the initial early heuristics without holding
1934 * any task-queue locks at all. We'll only try to get
1935 * the runqueue lock when things look like they will
1941 * If the task is actively running on another CPU
1942 * still, just relax and busy-wait without holding
1945 * NOTE! Since we don't hold any locks, it's not
1946 * even sure that "rq" stays as the right runqueue!
1947 * But we don't care, since "task_running()" will
1948 * return false if the runqueue has changed and p
1949 * is actually now running somewhere else!
1951 while (task_running(rq, p)) {
1952 if (match_state && unlikely(p->state != match_state))
1958 * Ok, time to look more closely! We need the rq
1959 * lock now, to be *sure*. If we're wrong, we'll
1960 * just go back and repeat.
1962 rq = task_rq_lock(p, &flags);
1963 trace_sched_wait_task(rq, p);
1964 running = task_running(rq, p);
1965 on_rq = p->se.on_rq;
1967 if (!match_state || p->state == match_state)
1968 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1969 task_rq_unlock(rq, &flags);
1972 * If it changed from the expected state, bail out now.
1974 if (unlikely(!ncsw))
1978 * Was it really running after all now that we
1979 * checked with the proper locks actually held?
1981 * Oops. Go back and try again..
1983 if (unlikely(running)) {
1989 * It's not enough that it's not actively running,
1990 * it must be off the runqueue _entirely_, and not
1993 * So if it wa still runnable (but just not actively
1994 * running right now), it's preempted, and we should
1995 * yield - it could be a while.
1997 if (unlikely(on_rq)) {
1998 schedule_timeout_uninterruptible(1);
2003 * Ahh, all good. It wasn't running, and it wasn't
2004 * runnable, which means that it will never become
2005 * running in the future either. We're all done!
2014 * kick_process - kick a running thread to enter/exit the kernel
2015 * @p: the to-be-kicked thread
2017 * Cause a process which is running on another CPU to enter
2018 * kernel-mode, without any delay. (to get signals handled.)
2020 * NOTE: this function doesnt have to take the runqueue lock,
2021 * because all it wants to ensure is that the remote task enters
2022 * the kernel. If the IPI races and the task has been migrated
2023 * to another CPU then no harm is done and the purpose has been
2026 void kick_process(struct task_struct *p)
2032 if ((cpu != smp_processor_id()) && task_curr(p))
2033 smp_send_reschedule(cpu);
2038 * Return a low guess at the load of a migration-source cpu weighted
2039 * according to the scheduling class and "nice" value.
2041 * We want to under-estimate the load of migration sources, to
2042 * balance conservatively.
2044 static unsigned long source_load(int cpu, int type)
2046 struct rq *rq = cpu_rq(cpu);
2047 unsigned long total = weighted_cpuload(cpu);
2049 if (type == 0 || !sched_feat(LB_BIAS))
2052 return min(rq->cpu_load[type-1], total);
2056 * Return a high guess at the load of a migration-target cpu weighted
2057 * according to the scheduling class and "nice" value.
2059 static unsigned long target_load(int cpu, int type)
2061 struct rq *rq = cpu_rq(cpu);
2062 unsigned long total = weighted_cpuload(cpu);
2064 if (type == 0 || !sched_feat(LB_BIAS))
2067 return max(rq->cpu_load[type-1], total);
2071 * find_idlest_group finds and returns the least busy CPU group within the
2074 static struct sched_group *
2075 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2077 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2078 unsigned long min_load = ULONG_MAX, this_load = 0;
2079 int load_idx = sd->forkexec_idx;
2080 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2083 unsigned long load, avg_load;
2087 /* Skip over this group if it has no CPUs allowed */
2088 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
2091 local_group = cpu_isset(this_cpu, group->cpumask);
2093 /* Tally up the load of all CPUs in the group */
2096 for_each_cpu_mask_nr(i, group->cpumask) {
2097 /* Bias balancing toward cpus of our domain */
2099 load = source_load(i, load_idx);
2101 load = target_load(i, load_idx);
2106 /* Adjust by relative CPU power of the group */
2107 avg_load = sg_div_cpu_power(group,
2108 avg_load * SCHED_LOAD_SCALE);
2111 this_load = avg_load;
2113 } else if (avg_load < min_load) {
2114 min_load = avg_load;
2117 } while (group = group->next, group != sd->groups);
2119 if (!idlest || 100*this_load < imbalance*min_load)
2125 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2128 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2131 unsigned long load, min_load = ULONG_MAX;
2135 /* Traverse only the allowed CPUs */
2136 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
2138 for_each_cpu_mask_nr(i, *tmp) {
2139 load = weighted_cpuload(i);
2141 if (load < min_load || (load == min_load && i == this_cpu)) {
2151 * sched_balance_self: balance the current task (running on cpu) in domains
2152 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2155 * Balance, ie. select the least loaded group.
2157 * Returns the target CPU number, or the same CPU if no balancing is needed.
2159 * preempt must be disabled.
2161 static int sched_balance_self(int cpu, int flag)
2163 struct task_struct *t = current;
2164 struct sched_domain *tmp, *sd = NULL;
2166 for_each_domain(cpu, tmp) {
2168 * If power savings logic is enabled for a domain, stop there.
2170 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2172 if (tmp->flags & flag)
2180 cpumask_t span, tmpmask;
2181 struct sched_group *group;
2182 int new_cpu, weight;
2184 if (!(sd->flags & flag)) {
2190 group = find_idlest_group(sd, t, cpu);
2196 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2197 if (new_cpu == -1 || new_cpu == cpu) {
2198 /* Now try balancing at a lower domain level of cpu */
2203 /* Now try balancing at a lower domain level of new_cpu */
2206 weight = cpus_weight(span);
2207 for_each_domain(cpu, tmp) {
2208 if (weight <= cpus_weight(tmp->span))
2210 if (tmp->flags & flag)
2213 /* while loop will break here if sd == NULL */
2219 #endif /* CONFIG_SMP */
2222 * try_to_wake_up - wake up a thread
2223 * @p: the to-be-woken-up thread
2224 * @state: the mask of task states that can be woken
2225 * @sync: do a synchronous wakeup?
2227 * Put it on the run-queue if it's not already there. The "current"
2228 * thread is always on the run-queue (except when the actual
2229 * re-schedule is in progress), and as such you're allowed to do
2230 * the simpler "current->state = TASK_RUNNING" to mark yourself
2231 * runnable without the overhead of this.
2233 * returns failure only if the task is already active.
2235 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2237 int cpu, orig_cpu, this_cpu, success = 0;
2238 unsigned long flags;
2242 if (!sched_feat(SYNC_WAKEUPS))
2246 if (sched_feat(LB_WAKEUP_UPDATE)) {
2247 struct sched_domain *sd;
2249 this_cpu = raw_smp_processor_id();
2252 for_each_domain(this_cpu, sd) {
2253 if (cpu_isset(cpu, sd->span)) {
2262 rq = task_rq_lock(p, &flags);
2263 old_state = p->state;
2264 if (!(old_state & state))
2272 this_cpu = smp_processor_id();
2275 if (unlikely(task_running(rq, p)))
2278 cpu = p->sched_class->select_task_rq(p, sync);
2279 if (cpu != orig_cpu) {
2280 set_task_cpu(p, cpu);
2281 task_rq_unlock(rq, &flags);
2282 /* might preempt at this point */
2283 rq = task_rq_lock(p, &flags);
2284 old_state = p->state;
2285 if (!(old_state & state))
2290 this_cpu = smp_processor_id();
2294 #ifdef CONFIG_SCHEDSTATS
2295 schedstat_inc(rq, ttwu_count);
2296 if (cpu == this_cpu)
2297 schedstat_inc(rq, ttwu_local);
2299 struct sched_domain *sd;
2300 for_each_domain(this_cpu, sd) {
2301 if (cpu_isset(cpu, sd->span)) {
2302 schedstat_inc(sd, ttwu_wake_remote);
2307 #endif /* CONFIG_SCHEDSTATS */
2310 #endif /* CONFIG_SMP */
2311 schedstat_inc(p, se.nr_wakeups);
2313 schedstat_inc(p, se.nr_wakeups_sync);
2314 if (orig_cpu != cpu)
2315 schedstat_inc(p, se.nr_wakeups_migrate);
2316 if (cpu == this_cpu)
2317 schedstat_inc(p, se.nr_wakeups_local);
2319 schedstat_inc(p, se.nr_wakeups_remote);
2320 update_rq_clock(rq);
2321 activate_task(rq, p, 1);
2325 trace_sched_wakeup(rq, p);
2326 check_preempt_curr(rq, p, sync);
2328 p->state = TASK_RUNNING;
2330 if (p->sched_class->task_wake_up)
2331 p->sched_class->task_wake_up(rq, p);
2334 current->se.last_wakeup = current->se.sum_exec_runtime;
2336 task_rq_unlock(rq, &flags);
2341 int wake_up_process(struct task_struct *p)
2343 return try_to_wake_up(p, TASK_ALL, 0);
2345 EXPORT_SYMBOL(wake_up_process);
2347 int wake_up_state(struct task_struct *p, unsigned int state)
2349 return try_to_wake_up(p, state, 0);
2353 * Perform scheduler related setup for a newly forked process p.
2354 * p is forked by current.
2356 * __sched_fork() is basic setup used by init_idle() too:
2358 static void __sched_fork(struct task_struct *p)
2360 p->se.exec_start = 0;
2361 p->se.sum_exec_runtime = 0;
2362 p->se.prev_sum_exec_runtime = 0;
2363 p->se.last_wakeup = 0;
2364 p->se.avg_overlap = 0;
2366 #ifdef CONFIG_SCHEDSTATS
2367 p->se.wait_start = 0;
2368 p->se.sum_sleep_runtime = 0;
2369 p->se.sleep_start = 0;
2370 p->se.block_start = 0;
2371 p->se.sleep_max = 0;
2372 p->se.block_max = 0;
2374 p->se.slice_max = 0;
2378 INIT_LIST_HEAD(&p->rt.run_list);
2380 INIT_LIST_HEAD(&p->se.group_node);
2382 #ifdef CONFIG_PREEMPT_NOTIFIERS
2383 INIT_HLIST_HEAD(&p->preempt_notifiers);
2387 * We mark the process as running here, but have not actually
2388 * inserted it onto the runqueue yet. This guarantees that
2389 * nobody will actually run it, and a signal or other external
2390 * event cannot wake it up and insert it on the runqueue either.
2392 p->state = TASK_RUNNING;
2396 * fork()/clone()-time setup:
2398 void sched_fork(struct task_struct *p, int clone_flags)
2400 int cpu = get_cpu();
2405 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2407 set_task_cpu(p, cpu);
2410 * Make sure we do not leak PI boosting priority to the child:
2412 p->prio = current->normal_prio;
2413 if (!rt_prio(p->prio))
2414 p->sched_class = &fair_sched_class;
2416 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2417 if (likely(sched_info_on()))
2418 memset(&p->sched_info, 0, sizeof(p->sched_info));
2420 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2423 #ifdef CONFIG_PREEMPT
2424 /* Want to start with kernel preemption disabled. */
2425 task_thread_info(p)->preempt_count = 1;
2431 * wake_up_new_task - wake up a newly created task for the first time.
2433 * This function will do some initial scheduler statistics housekeeping
2434 * that must be done for every newly created context, then puts the task
2435 * on the runqueue and wakes it.
2437 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2439 unsigned long flags;
2442 rq = task_rq_lock(p, &flags);
2443 BUG_ON(p->state != TASK_RUNNING);
2444 update_rq_clock(rq);
2446 p->prio = effective_prio(p);
2448 if (!p->sched_class->task_new || !current->se.on_rq) {
2449 activate_task(rq, p, 0);
2452 * Let the scheduling class do new task startup
2453 * management (if any):
2455 p->sched_class->task_new(rq, p);
2458 trace_sched_wakeup_new(rq, p);
2459 check_preempt_curr(rq, p, 0);
2461 if (p->sched_class->task_wake_up)
2462 p->sched_class->task_wake_up(rq, p);
2464 task_rq_unlock(rq, &flags);
2467 #ifdef CONFIG_PREEMPT_NOTIFIERS
2470 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2471 * @notifier: notifier struct to register
2473 void preempt_notifier_register(struct preempt_notifier *notifier)
2475 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2477 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2480 * preempt_notifier_unregister - no longer interested in preemption notifications
2481 * @notifier: notifier struct to unregister
2483 * This is safe to call from within a preemption notifier.
2485 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2487 hlist_del(¬ifier->link);
2489 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2491 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2493 struct preempt_notifier *notifier;
2494 struct hlist_node *node;
2496 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2497 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2501 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2502 struct task_struct *next)
2504 struct preempt_notifier *notifier;
2505 struct hlist_node *node;
2507 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2508 notifier->ops->sched_out(notifier, next);
2511 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2513 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2518 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2519 struct task_struct *next)
2523 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2526 * prepare_task_switch - prepare to switch tasks
2527 * @rq: the runqueue preparing to switch
2528 * @prev: the current task that is being switched out
2529 * @next: the task we are going to switch to.
2531 * This is called with the rq lock held and interrupts off. It must
2532 * be paired with a subsequent finish_task_switch after the context
2535 * prepare_task_switch sets up locking and calls architecture specific
2539 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2540 struct task_struct *next)
2542 fire_sched_out_preempt_notifiers(prev, next);
2543 prepare_lock_switch(rq, next);
2544 prepare_arch_switch(next);
2548 * finish_task_switch - clean up after a task-switch
2549 * @rq: runqueue associated with task-switch
2550 * @prev: the thread we just switched away from.
2552 * finish_task_switch must be called after the context switch, paired
2553 * with a prepare_task_switch call before the context switch.
2554 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2555 * and do any other architecture-specific cleanup actions.
2557 * Note that we may have delayed dropping an mm in context_switch(). If
2558 * so, we finish that here outside of the runqueue lock. (Doing it
2559 * with the lock held can cause deadlocks; see schedule() for
2562 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2563 __releases(rq->lock)
2565 struct mm_struct *mm = rq->prev_mm;
2571 * A task struct has one reference for the use as "current".
2572 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2573 * schedule one last time. The schedule call will never return, and
2574 * the scheduled task must drop that reference.
2575 * The test for TASK_DEAD must occur while the runqueue locks are
2576 * still held, otherwise prev could be scheduled on another cpu, die
2577 * there before we look at prev->state, and then the reference would
2579 * Manfred Spraul <manfred@colorfullife.com>
2581 prev_state = prev->state;
2582 finish_arch_switch(prev);
2583 finish_lock_switch(rq, prev);
2585 if (current->sched_class->post_schedule)
2586 current->sched_class->post_schedule(rq);
2589 fire_sched_in_preempt_notifiers(current);
2592 if (unlikely(prev_state == TASK_DEAD)) {
2594 * Remove function-return probe instances associated with this
2595 * task and put them back on the free list.
2597 kprobe_flush_task(prev);
2598 put_task_struct(prev);
2603 * schedule_tail - first thing a freshly forked thread must call.
2604 * @prev: the thread we just switched away from.
2606 asmlinkage void schedule_tail(struct task_struct *prev)
2607 __releases(rq->lock)
2609 struct rq *rq = this_rq();
2611 finish_task_switch(rq, prev);
2612 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2613 /* In this case, finish_task_switch does not reenable preemption */
2616 if (current->set_child_tid)
2617 put_user(task_pid_vnr(current), current->set_child_tid);
2621 * context_switch - switch to the new MM and the new
2622 * thread's register state.
2625 context_switch(struct rq *rq, struct task_struct *prev,
2626 struct task_struct *next)
2628 struct mm_struct *mm, *oldmm;
2630 prepare_task_switch(rq, prev, next);
2631 trace_sched_switch(rq, prev, next);
2633 oldmm = prev->active_mm;
2635 * For paravirt, this is coupled with an exit in switch_to to
2636 * combine the page table reload and the switch backend into
2639 arch_enter_lazy_cpu_mode();
2641 if (unlikely(!mm)) {
2642 next->active_mm = oldmm;
2643 atomic_inc(&oldmm->mm_count);
2644 enter_lazy_tlb(oldmm, next);
2646 switch_mm(oldmm, mm, next);
2648 if (unlikely(!prev->mm)) {
2649 prev->active_mm = NULL;
2650 rq->prev_mm = oldmm;
2653 * Since the runqueue lock will be released by the next
2654 * task (which is an invalid locking op but in the case
2655 * of the scheduler it's an obvious special-case), so we
2656 * do an early lockdep release here:
2658 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2659 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2662 /* Here we just switch the register state and the stack. */
2663 switch_to(prev, next, prev);
2667 * this_rq must be evaluated again because prev may have moved
2668 * CPUs since it called schedule(), thus the 'rq' on its stack
2669 * frame will be invalid.
2671 finish_task_switch(this_rq(), prev);
2675 * nr_running, nr_uninterruptible and nr_context_switches:
2677 * externally visible scheduler statistics: current number of runnable
2678 * threads, current number of uninterruptible-sleeping threads, total
2679 * number of context switches performed since bootup.
2681 unsigned long nr_running(void)
2683 unsigned long i, sum = 0;
2685 for_each_online_cpu(i)
2686 sum += cpu_rq(i)->nr_running;
2691 unsigned long nr_uninterruptible(void)
2693 unsigned long i, sum = 0;
2695 for_each_possible_cpu(i)
2696 sum += cpu_rq(i)->nr_uninterruptible;
2699 * Since we read the counters lockless, it might be slightly
2700 * inaccurate. Do not allow it to go below zero though:
2702 if (unlikely((long)sum < 0))
2708 unsigned long long nr_context_switches(void)
2711 unsigned long long sum = 0;
2713 for_each_possible_cpu(i)
2714 sum += cpu_rq(i)->nr_switches;
2719 unsigned long nr_iowait(void)
2721 unsigned long i, sum = 0;
2723 for_each_possible_cpu(i)
2724 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2729 unsigned long nr_active(void)
2731 unsigned long i, running = 0, uninterruptible = 0;
2733 for_each_online_cpu(i) {
2734 running += cpu_rq(i)->nr_running;
2735 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2738 if (unlikely((long)uninterruptible < 0))
2739 uninterruptible = 0;
2741 return running + uninterruptible;
2745 * Update rq->cpu_load[] statistics. This function is usually called every
2746 * scheduler tick (TICK_NSEC).
2748 static void update_cpu_load(struct rq *this_rq)
2750 unsigned long this_load = this_rq->load.weight;
2753 this_rq->nr_load_updates++;
2755 /* Update our load: */
2756 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2757 unsigned long old_load, new_load;
2759 /* scale is effectively 1 << i now, and >> i divides by scale */
2761 old_load = this_rq->cpu_load[i];
2762 new_load = this_load;
2764 * Round up the averaging division if load is increasing. This
2765 * prevents us from getting stuck on 9 if the load is 10, for
2768 if (new_load > old_load)
2769 new_load += scale-1;
2770 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2777 * double_rq_lock - safely lock two runqueues
2779 * Note this does not disable interrupts like task_rq_lock,
2780 * you need to do so manually before calling.
2782 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2783 __acquires(rq1->lock)
2784 __acquires(rq2->lock)
2786 BUG_ON(!irqs_disabled());
2788 spin_lock(&rq1->lock);
2789 __acquire(rq2->lock); /* Fake it out ;) */
2792 spin_lock(&rq1->lock);
2793 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2795 spin_lock(&rq2->lock);
2796 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2799 update_rq_clock(rq1);
2800 update_rq_clock(rq2);
2804 * double_rq_unlock - safely unlock two runqueues
2806 * Note this does not restore interrupts like task_rq_unlock,
2807 * you need to do so manually after calling.
2809 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2810 __releases(rq1->lock)
2811 __releases(rq2->lock)
2813 spin_unlock(&rq1->lock);
2815 spin_unlock(&rq2->lock);
2817 __release(rq2->lock);
2821 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2823 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2824 __releases(this_rq->lock)
2825 __acquires(busiest->lock)
2826 __acquires(this_rq->lock)
2830 if (unlikely(!irqs_disabled())) {
2831 /* printk() doesn't work good under rq->lock */
2832 spin_unlock(&this_rq->lock);
2835 if (unlikely(!spin_trylock(&busiest->lock))) {
2836 if (busiest < this_rq) {
2837 spin_unlock(&this_rq->lock);
2838 spin_lock(&busiest->lock);
2839 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
2842 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
2847 static void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
2848 __releases(busiest->lock)
2850 spin_unlock(&busiest->lock);
2851 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
2855 * If dest_cpu is allowed for this process, migrate the task to it.
2856 * This is accomplished by forcing the cpu_allowed mask to only
2857 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2858 * the cpu_allowed mask is restored.
2860 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2862 struct migration_req req;
2863 unsigned long flags;
2866 rq = task_rq_lock(p, &flags);
2867 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2868 || unlikely(!cpu_active(dest_cpu)))
2871 trace_sched_migrate_task(rq, p, dest_cpu);
2872 /* force the process onto the specified CPU */
2873 if (migrate_task(p, dest_cpu, &req)) {
2874 /* Need to wait for migration thread (might exit: take ref). */
2875 struct task_struct *mt = rq->migration_thread;
2877 get_task_struct(mt);
2878 task_rq_unlock(rq, &flags);
2879 wake_up_process(mt);
2880 put_task_struct(mt);
2881 wait_for_completion(&req.done);
2886 task_rq_unlock(rq, &flags);
2890 * sched_exec - execve() is a valuable balancing opportunity, because at
2891 * this point the task has the smallest effective memory and cache footprint.
2893 void sched_exec(void)
2895 int new_cpu, this_cpu = get_cpu();
2896 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2898 if (new_cpu != this_cpu)
2899 sched_migrate_task(current, new_cpu);
2903 * pull_task - move a task from a remote runqueue to the local runqueue.
2904 * Both runqueues must be locked.
2906 static void pull_task(struct rq *src_rq, struct task_struct *p,
2907 struct rq *this_rq, int this_cpu)
2909 deactivate_task(src_rq, p, 0);
2910 set_task_cpu(p, this_cpu);
2911 activate_task(this_rq, p, 0);
2913 * Note that idle threads have a prio of MAX_PRIO, for this test
2914 * to be always true for them.
2916 check_preempt_curr(this_rq, p, 0);
2920 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2923 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2924 struct sched_domain *sd, enum cpu_idle_type idle,
2928 * We do not migrate tasks that are:
2929 * 1) running (obviously), or
2930 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2931 * 3) are cache-hot on their current CPU.
2933 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2934 schedstat_inc(p, se.nr_failed_migrations_affine);
2939 if (task_running(rq, p)) {
2940 schedstat_inc(p, se.nr_failed_migrations_running);
2945 * Aggressive migration if:
2946 * 1) task is cache cold, or
2947 * 2) too many balance attempts have failed.
2950 if (!task_hot(p, rq->clock, sd) ||
2951 sd->nr_balance_failed > sd->cache_nice_tries) {
2952 #ifdef CONFIG_SCHEDSTATS
2953 if (task_hot(p, rq->clock, sd)) {
2954 schedstat_inc(sd, lb_hot_gained[idle]);
2955 schedstat_inc(p, se.nr_forced_migrations);
2961 if (task_hot(p, rq->clock, sd)) {
2962 schedstat_inc(p, se.nr_failed_migrations_hot);
2968 static unsigned long
2969 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2970 unsigned long max_load_move, struct sched_domain *sd,
2971 enum cpu_idle_type idle, int *all_pinned,
2972 int *this_best_prio, struct rq_iterator *iterator)
2974 int loops = 0, pulled = 0, pinned = 0;
2975 struct task_struct *p;
2976 long rem_load_move = max_load_move;
2978 if (max_load_move == 0)
2984 * Start the load-balancing iterator:
2986 p = iterator->start(iterator->arg);
2988 if (!p || loops++ > sysctl_sched_nr_migrate)
2991 if ((p->se.load.weight >> 1) > rem_load_move ||
2992 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2993 p = iterator->next(iterator->arg);
2997 pull_task(busiest, p, this_rq, this_cpu);
2999 rem_load_move -= p->se.load.weight;
3002 * We only want to steal up to the prescribed amount of weighted load.
3004 if (rem_load_move > 0) {
3005 if (p->prio < *this_best_prio)
3006 *this_best_prio = p->prio;
3007 p = iterator->next(iterator->arg);
3012 * Right now, this is one of only two places pull_task() is called,
3013 * so we can safely collect pull_task() stats here rather than
3014 * inside pull_task().
3016 schedstat_add(sd, lb_gained[idle], pulled);
3019 *all_pinned = pinned;
3021 return max_load_move - rem_load_move;
3025 * move_tasks tries to move up to max_load_move weighted load from busiest to
3026 * this_rq, as part of a balancing operation within domain "sd".
3027 * Returns 1 if successful and 0 otherwise.
3029 * Called with both runqueues locked.
3031 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3032 unsigned long max_load_move,
3033 struct sched_domain *sd, enum cpu_idle_type idle,
3036 const struct sched_class *class = sched_class_highest;
3037 unsigned long total_load_moved = 0;
3038 int this_best_prio = this_rq->curr->prio;
3042 class->load_balance(this_rq, this_cpu, busiest,
3043 max_load_move - total_load_moved,
3044 sd, idle, all_pinned, &this_best_prio);
3045 class = class->next;
3047 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3050 } while (class && max_load_move > total_load_moved);
3052 return total_load_moved > 0;
3056 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3057 struct sched_domain *sd, enum cpu_idle_type idle,
3058 struct rq_iterator *iterator)
3060 struct task_struct *p = iterator->start(iterator->arg);
3064 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3065 pull_task(busiest, p, this_rq, this_cpu);
3067 * Right now, this is only the second place pull_task()
3068 * is called, so we can safely collect pull_task()
3069 * stats here rather than inside pull_task().
3071 schedstat_inc(sd, lb_gained[idle]);
3075 p = iterator->next(iterator->arg);
3082 * move_one_task tries to move exactly one task from busiest to this_rq, as
3083 * part of active balancing operations within "domain".
3084 * Returns 1 if successful and 0 otherwise.
3086 * Called with both runqueues locked.
3088 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3089 struct sched_domain *sd, enum cpu_idle_type idle)
3091 const struct sched_class *class;
3093 for (class = sched_class_highest; class; class = class->next)
3094 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3101 * find_busiest_group finds and returns the busiest CPU group within the
3102 * domain. It calculates and returns the amount of weighted load which
3103 * should be moved to restore balance via the imbalance parameter.
3105 static struct sched_group *
3106 find_busiest_group(struct sched_domain *sd, int this_cpu,
3107 unsigned long *imbalance, enum cpu_idle_type idle,
3108 int *sd_idle, const cpumask_t *cpus, int *balance)
3110 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3111 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3112 unsigned long max_pull;
3113 unsigned long busiest_load_per_task, busiest_nr_running;
3114 unsigned long this_load_per_task, this_nr_running;
3115 int load_idx, group_imb = 0;
3116 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3117 int power_savings_balance = 1;
3118 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3119 unsigned long min_nr_running = ULONG_MAX;
3120 struct sched_group *group_min = NULL, *group_leader = NULL;
3123 max_load = this_load = total_load = total_pwr = 0;
3124 busiest_load_per_task = busiest_nr_running = 0;
3125 this_load_per_task = this_nr_running = 0;
3127 if (idle == CPU_NOT_IDLE)
3128 load_idx = sd->busy_idx;
3129 else if (idle == CPU_NEWLY_IDLE)
3130 load_idx = sd->newidle_idx;
3132 load_idx = sd->idle_idx;
3135 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3138 int __group_imb = 0;
3139 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3140 unsigned long sum_nr_running, sum_weighted_load;
3141 unsigned long sum_avg_load_per_task;
3142 unsigned long avg_load_per_task;
3144 local_group = cpu_isset(this_cpu, group->cpumask);
3147 balance_cpu = first_cpu(group->cpumask);
3149 /* Tally up the load of all CPUs in the group */
3150 sum_weighted_load = sum_nr_running = avg_load = 0;
3151 sum_avg_load_per_task = avg_load_per_task = 0;
3154 min_cpu_load = ~0UL;
3156 for_each_cpu_mask_nr(i, group->cpumask) {
3159 if (!cpu_isset(i, *cpus))
3164 if (*sd_idle && rq->nr_running)
3167 /* Bias balancing toward cpus of our domain */
3169 if (idle_cpu(i) && !first_idle_cpu) {
3174 load = target_load(i, load_idx);
3176 load = source_load(i, load_idx);
3177 if (load > max_cpu_load)
3178 max_cpu_load = load;
3179 if (min_cpu_load > load)
3180 min_cpu_load = load;
3184 sum_nr_running += rq->nr_running;
3185 sum_weighted_load += weighted_cpuload(i);
3187 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3191 * First idle cpu or the first cpu(busiest) in this sched group
3192 * is eligible for doing load balancing at this and above
3193 * domains. In the newly idle case, we will allow all the cpu's
3194 * to do the newly idle load balance.
3196 if (idle != CPU_NEWLY_IDLE && local_group &&
3197 balance_cpu != this_cpu && balance) {
3202 total_load += avg_load;
3203 total_pwr += group->__cpu_power;
3205 /* Adjust by relative CPU power of the group */
3206 avg_load = sg_div_cpu_power(group,
3207 avg_load * SCHED_LOAD_SCALE);
3211 * Consider the group unbalanced when the imbalance is larger
3212 * than the average weight of two tasks.
3214 * APZ: with cgroup the avg task weight can vary wildly and
3215 * might not be a suitable number - should we keep a
3216 * normalized nr_running number somewhere that negates
3219 avg_load_per_task = sg_div_cpu_power(group,
3220 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3222 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3225 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3228 this_load = avg_load;
3230 this_nr_running = sum_nr_running;
3231 this_load_per_task = sum_weighted_load;
3232 } else if (avg_load > max_load &&
3233 (sum_nr_running > group_capacity || __group_imb)) {
3234 max_load = avg_load;
3236 busiest_nr_running = sum_nr_running;
3237 busiest_load_per_task = sum_weighted_load;
3238 group_imb = __group_imb;
3241 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3243 * Busy processors will not participate in power savings
3246 if (idle == CPU_NOT_IDLE ||
3247 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3251 * If the local group is idle or completely loaded
3252 * no need to do power savings balance at this domain
3254 if (local_group && (this_nr_running >= group_capacity ||
3256 power_savings_balance = 0;
3259 * If a group is already running at full capacity or idle,
3260 * don't include that group in power savings calculations
3262 if (!power_savings_balance || sum_nr_running >= group_capacity
3267 * Calculate the group which has the least non-idle load.
3268 * This is the group from where we need to pick up the load
3271 if ((sum_nr_running < min_nr_running) ||
3272 (sum_nr_running == min_nr_running &&
3273 first_cpu(group->cpumask) <
3274 first_cpu(group_min->cpumask))) {
3276 min_nr_running = sum_nr_running;
3277 min_load_per_task = sum_weighted_load /
3282 * Calculate the group which is almost near its
3283 * capacity but still has some space to pick up some load
3284 * from other group and save more power
3286 if (sum_nr_running <= group_capacity - 1) {
3287 if (sum_nr_running > leader_nr_running ||
3288 (sum_nr_running == leader_nr_running &&
3289 first_cpu(group->cpumask) >
3290 first_cpu(group_leader->cpumask))) {
3291 group_leader = group;
3292 leader_nr_running = sum_nr_running;
3297 group = group->next;
3298 } while (group != sd->groups);
3300 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3303 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3305 if (this_load >= avg_load ||
3306 100*max_load <= sd->imbalance_pct*this_load)
3309 busiest_load_per_task /= busiest_nr_running;
3311 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3314 * We're trying to get all the cpus to the average_load, so we don't
3315 * want to push ourselves above the average load, nor do we wish to
3316 * reduce the max loaded cpu below the average load, as either of these
3317 * actions would just result in more rebalancing later, and ping-pong
3318 * tasks around. Thus we look for the minimum possible imbalance.
3319 * Negative imbalances (*we* are more loaded than anyone else) will
3320 * be counted as no imbalance for these purposes -- we can't fix that
3321 * by pulling tasks to us. Be careful of negative numbers as they'll
3322 * appear as very large values with unsigned longs.
3324 if (max_load <= busiest_load_per_task)
3328 * In the presence of smp nice balancing, certain scenarios can have
3329 * max load less than avg load(as we skip the groups at or below
3330 * its cpu_power, while calculating max_load..)
3332 if (max_load < avg_load) {
3334 goto small_imbalance;
3337 /* Don't want to pull so many tasks that a group would go idle */
3338 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3340 /* How much load to actually move to equalise the imbalance */
3341 *imbalance = min(max_pull * busiest->__cpu_power,
3342 (avg_load - this_load) * this->__cpu_power)
3346 * if *imbalance is less than the average load per runnable task
3347 * there is no gaurantee that any tasks will be moved so we'll have
3348 * a think about bumping its value to force at least one task to be
3351 if (*imbalance < busiest_load_per_task) {
3352 unsigned long tmp, pwr_now, pwr_move;
3356 pwr_move = pwr_now = 0;
3358 if (this_nr_running) {
3359 this_load_per_task /= this_nr_running;
3360 if (busiest_load_per_task > this_load_per_task)
3363 this_load_per_task = cpu_avg_load_per_task(this_cpu);
3365 if (max_load - this_load + busiest_load_per_task >=
3366 busiest_load_per_task * imbn) {
3367 *imbalance = busiest_load_per_task;
3372 * OK, we don't have enough imbalance to justify moving tasks,
3373 * however we may be able to increase total CPU power used by
3377 pwr_now += busiest->__cpu_power *
3378 min(busiest_load_per_task, max_load);
3379 pwr_now += this->__cpu_power *
3380 min(this_load_per_task, this_load);
3381 pwr_now /= SCHED_LOAD_SCALE;
3383 /* Amount of load we'd subtract */
3384 tmp = sg_div_cpu_power(busiest,
3385 busiest_load_per_task * SCHED_LOAD_SCALE);
3387 pwr_move += busiest->__cpu_power *
3388 min(busiest_load_per_task, max_load - tmp);
3390 /* Amount of load we'd add */
3391 if (max_load * busiest->__cpu_power <
3392 busiest_load_per_task * SCHED_LOAD_SCALE)
3393 tmp = sg_div_cpu_power(this,
3394 max_load * busiest->__cpu_power);
3396 tmp = sg_div_cpu_power(this,
3397 busiest_load_per_task * SCHED_LOAD_SCALE);
3398 pwr_move += this->__cpu_power *
3399 min(this_load_per_task, this_load + tmp);
3400 pwr_move /= SCHED_LOAD_SCALE;
3402 /* Move if we gain throughput */
3403 if (pwr_move > pwr_now)
3404 *imbalance = busiest_load_per_task;
3410 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3411 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3414 if (this == group_leader && group_leader != group_min) {
3415 *imbalance = min_load_per_task;
3425 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3428 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3429 unsigned long imbalance, const cpumask_t *cpus)
3431 struct rq *busiest = NULL, *rq;
3432 unsigned long max_load = 0;
3435 for_each_cpu_mask_nr(i, group->cpumask) {
3438 if (!cpu_isset(i, *cpus))
3442 wl = weighted_cpuload(i);
3444 if (rq->nr_running == 1 && wl > imbalance)
3447 if (wl > max_load) {
3457 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3458 * so long as it is large enough.
3460 #define MAX_PINNED_INTERVAL 512
3463 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3464 * tasks if there is an imbalance.
3466 static int load_balance(int this_cpu, struct rq *this_rq,
3467 struct sched_domain *sd, enum cpu_idle_type idle,
3468 int *balance, cpumask_t *cpus)
3470 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3471 struct sched_group *group;
3472 unsigned long imbalance;
3474 unsigned long flags;
3479 * When power savings policy is enabled for the parent domain, idle
3480 * sibling can pick up load irrespective of busy siblings. In this case,
3481 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3482 * portraying it as CPU_NOT_IDLE.
3484 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3485 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3488 schedstat_inc(sd, lb_count[idle]);
3492 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3499 schedstat_inc(sd, lb_nobusyg[idle]);
3503 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3505 schedstat_inc(sd, lb_nobusyq[idle]);
3509 BUG_ON(busiest == this_rq);
3511 schedstat_add(sd, lb_imbalance[idle], imbalance);
3514 if (busiest->nr_running > 1) {
3516 * Attempt to move tasks. If find_busiest_group has found
3517 * an imbalance but busiest->nr_running <= 1, the group is
3518 * still unbalanced. ld_moved simply stays zero, so it is
3519 * correctly treated as an imbalance.
3521 local_irq_save(flags);
3522 double_rq_lock(this_rq, busiest);
3523 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3524 imbalance, sd, idle, &all_pinned);
3525 double_rq_unlock(this_rq, busiest);
3526 local_irq_restore(flags);
3529 * some other cpu did the load balance for us.
3531 if (ld_moved && this_cpu != smp_processor_id())
3532 resched_cpu(this_cpu);
3534 /* All tasks on this runqueue were pinned by CPU affinity */
3535 if (unlikely(all_pinned)) {
3536 cpu_clear(cpu_of(busiest), *cpus);
3537 if (!cpus_empty(*cpus))
3544 schedstat_inc(sd, lb_failed[idle]);
3545 sd->nr_balance_failed++;
3547 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3549 spin_lock_irqsave(&busiest->lock, flags);
3551 /* don't kick the migration_thread, if the curr
3552 * task on busiest cpu can't be moved to this_cpu
3554 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3555 spin_unlock_irqrestore(&busiest->lock, flags);
3557 goto out_one_pinned;
3560 if (!busiest->active_balance) {
3561 busiest->active_balance = 1;
3562 busiest->push_cpu = this_cpu;
3565 spin_unlock_irqrestore(&busiest->lock, flags);
3567 wake_up_process(busiest->migration_thread);
3570 * We've kicked active balancing, reset the failure
3573 sd->nr_balance_failed = sd->cache_nice_tries+1;
3576 sd->nr_balance_failed = 0;
3578 if (likely(!active_balance)) {
3579 /* We were unbalanced, so reset the balancing interval */
3580 sd->balance_interval = sd->min_interval;
3583 * If we've begun active balancing, start to back off. This
3584 * case may not be covered by the all_pinned logic if there
3585 * is only 1 task on the busy runqueue (because we don't call
3588 if (sd->balance_interval < sd->max_interval)
3589 sd->balance_interval *= 2;
3592 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3593 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3599 schedstat_inc(sd, lb_balanced[idle]);
3601 sd->nr_balance_failed = 0;
3604 /* tune up the balancing interval */
3605 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3606 (sd->balance_interval < sd->max_interval))
3607 sd->balance_interval *= 2;
3609 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3610 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3621 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3622 * tasks if there is an imbalance.
3624 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3625 * this_rq is locked.
3628 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3631 struct sched_group *group;
3632 struct rq *busiest = NULL;
3633 unsigned long imbalance;
3641 * When power savings policy is enabled for the parent domain, idle
3642 * sibling can pick up load irrespective of busy siblings. In this case,
3643 * let the state of idle sibling percolate up as IDLE, instead of
3644 * portraying it as CPU_NOT_IDLE.
3646 if (sd->flags & SD_SHARE_CPUPOWER &&
3647 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3650 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3652 update_shares_locked(this_rq, sd);
3653 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3654 &sd_idle, cpus, NULL);
3656 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3660 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3662 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3666 BUG_ON(busiest == this_rq);
3668 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3671 if (busiest->nr_running > 1) {
3672 /* Attempt to move tasks */
3673 double_lock_balance(this_rq, busiest);
3674 /* this_rq->clock is already updated */
3675 update_rq_clock(busiest);
3676 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3677 imbalance, sd, CPU_NEWLY_IDLE,
3679 double_unlock_balance(this_rq, busiest);
3681 if (unlikely(all_pinned)) {
3682 cpu_clear(cpu_of(busiest), *cpus);
3683 if (!cpus_empty(*cpus))
3689 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3690 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3691 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3694 sd->nr_balance_failed = 0;
3696 update_shares_locked(this_rq, sd);
3700 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3701 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3702 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3704 sd->nr_balance_failed = 0;
3710 * idle_balance is called by schedule() if this_cpu is about to become
3711 * idle. Attempts to pull tasks from other CPUs.
3713 static void idle_balance(int this_cpu, struct rq *this_rq)
3715 struct sched_domain *sd;
3716 int pulled_task = -1;
3717 unsigned long next_balance = jiffies + HZ;
3720 for_each_domain(this_cpu, sd) {
3721 unsigned long interval;
3723 if (!(sd->flags & SD_LOAD_BALANCE))
3726 if (sd->flags & SD_BALANCE_NEWIDLE)
3727 /* If we've pulled tasks over stop searching: */
3728 pulled_task = load_balance_newidle(this_cpu, this_rq,
3731 interval = msecs_to_jiffies(sd->balance_interval);
3732 if (time_after(next_balance, sd->last_balance + interval))
3733 next_balance = sd->last_balance + interval;
3737 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3739 * We are going idle. next_balance may be set based on
3740 * a busy processor. So reset next_balance.
3742 this_rq->next_balance = next_balance;
3747 * active_load_balance is run by migration threads. It pushes running tasks
3748 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3749 * running on each physical CPU where possible, and avoids physical /
3750 * logical imbalances.
3752 * Called with busiest_rq locked.
3754 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3756 int target_cpu = busiest_rq->push_cpu;
3757 struct sched_domain *sd;
3758 struct rq *target_rq;
3760 /* Is there any task to move? */
3761 if (busiest_rq->nr_running <= 1)
3764 target_rq = cpu_rq(target_cpu);
3767 * This condition is "impossible", if it occurs
3768 * we need to fix it. Originally reported by
3769 * Bjorn Helgaas on a 128-cpu setup.
3771 BUG_ON(busiest_rq == target_rq);
3773 /* move a task from busiest_rq to target_rq */
3774 double_lock_balance(busiest_rq, target_rq);
3775 update_rq_clock(busiest_rq);
3776 update_rq_clock(target_rq);
3778 /* Search for an sd spanning us and the target CPU. */
3779 for_each_domain(target_cpu, sd) {
3780 if ((sd->flags & SD_LOAD_BALANCE) &&
3781 cpu_isset(busiest_cpu, sd->span))
3786 schedstat_inc(sd, alb_count);
3788 if (move_one_task(target_rq, target_cpu, busiest_rq,
3790 schedstat_inc(sd, alb_pushed);
3792 schedstat_inc(sd, alb_failed);
3794 double_unlock_balance(busiest_rq, target_rq);
3799 atomic_t load_balancer;
3801 } nohz ____cacheline_aligned = {
3802 .load_balancer = ATOMIC_INIT(-1),
3803 .cpu_mask = CPU_MASK_NONE,
3807 * This routine will try to nominate the ilb (idle load balancing)
3808 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3809 * load balancing on behalf of all those cpus. If all the cpus in the system
3810 * go into this tickless mode, then there will be no ilb owner (as there is
3811 * no need for one) and all the cpus will sleep till the next wakeup event
3814 * For the ilb owner, tick is not stopped. And this tick will be used
3815 * for idle load balancing. ilb owner will still be part of
3818 * While stopping the tick, this cpu will become the ilb owner if there
3819 * is no other owner. And will be the owner till that cpu becomes busy
3820 * or if all cpus in the system stop their ticks at which point
3821 * there is no need for ilb owner.
3823 * When the ilb owner becomes busy, it nominates another owner, during the
3824 * next busy scheduler_tick()
3826 int select_nohz_load_balancer(int stop_tick)
3828 int cpu = smp_processor_id();
3831 cpu_set(cpu, nohz.cpu_mask);
3832 cpu_rq(cpu)->in_nohz_recently = 1;
3835 * If we are going offline and still the leader, give up!
3837 if (!cpu_active(cpu) &&
3838 atomic_read(&nohz.load_balancer) == cpu) {
3839 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3844 /* time for ilb owner also to sleep */
3845 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3846 if (atomic_read(&nohz.load_balancer) == cpu)
3847 atomic_set(&nohz.load_balancer, -1);
3851 if (atomic_read(&nohz.load_balancer) == -1) {
3852 /* make me the ilb owner */
3853 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3855 } else if (atomic_read(&nohz.load_balancer) == cpu)
3858 if (!cpu_isset(cpu, nohz.cpu_mask))
3861 cpu_clear(cpu, nohz.cpu_mask);
3863 if (atomic_read(&nohz.load_balancer) == cpu)
3864 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3871 static DEFINE_SPINLOCK(balancing);
3874 * It checks each scheduling domain to see if it is due to be balanced,
3875 * and initiates a balancing operation if so.
3877 * Balancing parameters are set up in arch_init_sched_domains.
3879 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3882 struct rq *rq = cpu_rq(cpu);
3883 unsigned long interval;
3884 struct sched_domain *sd;
3885 /* Earliest time when we have to do rebalance again */
3886 unsigned long next_balance = jiffies + 60*HZ;
3887 int update_next_balance = 0;
3891 for_each_domain(cpu, sd) {
3892 if (!(sd->flags & SD_LOAD_BALANCE))
3895 interval = sd->balance_interval;
3896 if (idle != CPU_IDLE)
3897 interval *= sd->busy_factor;
3899 /* scale ms to jiffies */
3900 interval = msecs_to_jiffies(interval);
3901 if (unlikely(!interval))
3903 if (interval > HZ*NR_CPUS/10)
3904 interval = HZ*NR_CPUS/10;
3906 need_serialize = sd->flags & SD_SERIALIZE;
3908 if (need_serialize) {
3909 if (!spin_trylock(&balancing))
3913 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3914 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3916 * We've pulled tasks over so either we're no
3917 * longer idle, or one of our SMT siblings is
3920 idle = CPU_NOT_IDLE;
3922 sd->last_balance = jiffies;
3925 spin_unlock(&balancing);
3927 if (time_after(next_balance, sd->last_balance + interval)) {
3928 next_balance = sd->last_balance + interval;
3929 update_next_balance = 1;
3933 * Stop the load balance at this level. There is another
3934 * CPU in our sched group which is doing load balancing more
3942 * next_balance will be updated only when there is a need.
3943 * When the cpu is attached to null domain for ex, it will not be
3946 if (likely(update_next_balance))
3947 rq->next_balance = next_balance;
3951 * run_rebalance_domains is triggered when needed from the scheduler tick.
3952 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3953 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3955 static void run_rebalance_domains(struct softirq_action *h)
3957 int this_cpu = smp_processor_id();
3958 struct rq *this_rq = cpu_rq(this_cpu);
3959 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3960 CPU_IDLE : CPU_NOT_IDLE;
3962 rebalance_domains(this_cpu, idle);
3966 * If this cpu is the owner for idle load balancing, then do the
3967 * balancing on behalf of the other idle cpus whose ticks are
3970 if (this_rq->idle_at_tick &&
3971 atomic_read(&nohz.load_balancer) == this_cpu) {
3972 cpumask_t cpus = nohz.cpu_mask;
3976 cpu_clear(this_cpu, cpus);
3977 for_each_cpu_mask_nr(balance_cpu, cpus) {
3979 * If this cpu gets work to do, stop the load balancing
3980 * work being done for other cpus. Next load
3981 * balancing owner will pick it up.
3986 rebalance_domains(balance_cpu, CPU_IDLE);
3988 rq = cpu_rq(balance_cpu);
3989 if (time_after(this_rq->next_balance, rq->next_balance))
3990 this_rq->next_balance = rq->next_balance;
3997 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3999 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4000 * idle load balancing owner or decide to stop the periodic load balancing,
4001 * if the whole system is idle.
4003 static inline void trigger_load_balance(struct rq *rq, int cpu)
4007 * If we were in the nohz mode recently and busy at the current
4008 * scheduler tick, then check if we need to nominate new idle
4011 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4012 rq->in_nohz_recently = 0;
4014 if (atomic_read(&nohz.load_balancer) == cpu) {
4015 cpu_clear(cpu, nohz.cpu_mask);
4016 atomic_set(&nohz.load_balancer, -1);
4019 if (atomic_read(&nohz.load_balancer) == -1) {
4021 * simple selection for now: Nominate the
4022 * first cpu in the nohz list to be the next
4025 * TBD: Traverse the sched domains and nominate
4026 * the nearest cpu in the nohz.cpu_mask.
4028 int ilb = first_cpu(nohz.cpu_mask);
4030 if (ilb < nr_cpu_ids)
4036 * If this cpu is idle and doing idle load balancing for all the
4037 * cpus with ticks stopped, is it time for that to stop?
4039 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4040 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4046 * If this cpu is idle and the idle load balancing is done by
4047 * someone else, then no need raise the SCHED_SOFTIRQ
4049 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4050 cpu_isset(cpu, nohz.cpu_mask))
4053 if (time_after_eq(jiffies, rq->next_balance))
4054 raise_softirq(SCHED_SOFTIRQ);
4057 #else /* CONFIG_SMP */
4060 * on UP we do not need to balance between CPUs:
4062 static inline void idle_balance(int cpu, struct rq *rq)
4068 DEFINE_PER_CPU(struct kernel_stat, kstat);
4070 EXPORT_PER_CPU_SYMBOL(kstat);
4073 * Return any ns on the sched_clock that have not yet been banked in
4074 * @p in case that task is currently running.
4076 unsigned long long task_delta_exec(struct task_struct *p)
4078 unsigned long flags;
4082 rq = task_rq_lock(p, &flags);
4084 if (task_current(rq, p)) {
4087 update_rq_clock(rq);
4088 delta_exec = rq->clock - p->se.exec_start;
4089 if ((s64)delta_exec > 0)
4093 task_rq_unlock(rq, &flags);
4099 * Account user cpu time to a process.
4100 * @p: the process that the cpu time gets accounted to
4101 * @cputime: the cpu time spent in user space since the last update
4103 void account_user_time(struct task_struct *p, cputime_t cputime)
4105 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4108 p->utime = cputime_add(p->utime, cputime);
4109 account_group_user_time(p, cputime);
4111 /* Add user time to cpustat. */
4112 tmp = cputime_to_cputime64(cputime);
4113 if (TASK_NICE(p) > 0)
4114 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4116 cpustat->user = cputime64_add(cpustat->user, tmp);
4117 /* Account for user time used */
4118 acct_update_integrals(p);
4122 * Account guest cpu time to a process.
4123 * @p: the process that the cpu time gets accounted to
4124 * @cputime: the cpu time spent in virtual machine since the last update
4126 static void account_guest_time(struct task_struct *p, cputime_t cputime)
4129 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4131 tmp = cputime_to_cputime64(cputime);
4133 p->utime = cputime_add(p->utime, cputime);
4134 account_group_user_time(p, cputime);
4135 p->gtime = cputime_add(p->gtime, cputime);
4137 cpustat->user = cputime64_add(cpustat->user, tmp);
4138 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4142 * Account scaled user cpu time to a process.
4143 * @p: the process that the cpu time gets accounted to
4144 * @cputime: the cpu time spent in user space since the last update
4146 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4148 p->utimescaled = cputime_add(p->utimescaled, cputime);
4152 * Account system cpu time to a process.
4153 * @p: the process that the cpu time gets accounted to
4154 * @hardirq_offset: the offset to subtract from hardirq_count()
4155 * @cputime: the cpu time spent in kernel space since the last update
4157 void account_system_time(struct task_struct *p, int hardirq_offset,
4160 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4161 struct rq *rq = this_rq();
4164 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4165 account_guest_time(p, cputime);
4169 p->stime = cputime_add(p->stime, cputime);
4170 account_group_system_time(p, cputime);
4172 /* Add system time to cpustat. */
4173 tmp = cputime_to_cputime64(cputime);
4174 if (hardirq_count() - hardirq_offset)
4175 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4176 else if (softirq_count())
4177 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4178 else if (p != rq->idle)
4179 cpustat->system = cputime64_add(cpustat->system, tmp);
4180 else if (atomic_read(&rq->nr_iowait) > 0)
4181 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4183 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4184 /* Account for system time used */
4185 acct_update_integrals(p);
4189 * Account scaled system cpu time to a process.
4190 * @p: the process that the cpu time gets accounted to
4191 * @hardirq_offset: the offset to subtract from hardirq_count()
4192 * @cputime: the cpu time spent in kernel space since the last update
4194 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4196 p->stimescaled = cputime_add(p->stimescaled, cputime);
4200 * Account for involuntary wait time.
4201 * @p: the process from which the cpu time has been stolen
4202 * @steal: the cpu time spent in involuntary wait
4204 void account_steal_time(struct task_struct *p, cputime_t steal)
4206 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4207 cputime64_t tmp = cputime_to_cputime64(steal);
4208 struct rq *rq = this_rq();
4210 if (p == rq->idle) {
4211 p->stime = cputime_add(p->stime, steal);
4212 account_group_system_time(p, steal);
4213 if (atomic_read(&rq->nr_iowait) > 0)
4214 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4216 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4218 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4222 * Use precise platform statistics if available:
4224 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4225 cputime_t task_utime(struct task_struct *p)
4230 cputime_t task_stime(struct task_struct *p)
4235 cputime_t task_utime(struct task_struct *p)
4237 clock_t utime = cputime_to_clock_t(p->utime),
4238 total = utime + cputime_to_clock_t(p->stime);
4242 * Use CFS's precise accounting:
4244 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
4248 do_div(temp, total);
4250 utime = (clock_t)temp;
4252 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
4253 return p->prev_utime;
4256 cputime_t task_stime(struct task_struct *p)
4261 * Use CFS's precise accounting. (we subtract utime from
4262 * the total, to make sure the total observed by userspace
4263 * grows monotonically - apps rely on that):
4265 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
4266 cputime_to_clock_t(task_utime(p));
4269 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
4271 return p->prev_stime;
4275 inline cputime_t task_gtime(struct task_struct *p)
4281 * This function gets called by the timer code, with HZ frequency.
4282 * We call it with interrupts disabled.
4284 * It also gets called by the fork code, when changing the parent's
4287 void scheduler_tick(void)
4289 int cpu = smp_processor_id();
4290 struct rq *rq = cpu_rq(cpu);
4291 struct task_struct *curr = rq->curr;
4295 spin_lock(&rq->lock);
4296 update_rq_clock(rq);
4297 update_cpu_load(rq);
4298 curr->sched_class->task_tick(rq, curr, 0);
4299 spin_unlock(&rq->lock);
4302 rq->idle_at_tick = idle_cpu(cpu);
4303 trigger_load_balance(rq, cpu);
4307 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4308 defined(CONFIG_PREEMPT_TRACER))
4310 static inline unsigned long get_parent_ip(unsigned long addr)
4312 if (in_lock_functions(addr)) {
4313 addr = CALLER_ADDR2;
4314 if (in_lock_functions(addr))
4315 addr = CALLER_ADDR3;
4320 void __kprobes add_preempt_count(int val)
4322 #ifdef CONFIG_DEBUG_PREEMPT
4326 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4329 preempt_count() += val;
4330 #ifdef CONFIG_DEBUG_PREEMPT
4332 * Spinlock count overflowing soon?
4334 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4337 if (preempt_count() == val)
4338 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4340 EXPORT_SYMBOL(add_preempt_count);
4342 void __kprobes sub_preempt_count(int val)
4344 #ifdef CONFIG_DEBUG_PREEMPT
4348 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4351 * Is the spinlock portion underflowing?
4353 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4354 !(preempt_count() & PREEMPT_MASK)))
4358 if (preempt_count() == val)
4359 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4360 preempt_count() -= val;
4362 EXPORT_SYMBOL(sub_preempt_count);
4367 * Print scheduling while atomic bug:
4369 static noinline void __schedule_bug(struct task_struct *prev)
4371 struct pt_regs *regs = get_irq_regs();
4373 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4374 prev->comm, prev->pid, preempt_count());
4376 debug_show_held_locks(prev);
4378 if (irqs_disabled())
4379 print_irqtrace_events(prev);
4388 * Various schedule()-time debugging checks and statistics:
4390 static inline void schedule_debug(struct task_struct *prev)
4393 * Test if we are atomic. Since do_exit() needs to call into
4394 * schedule() atomically, we ignore that path for now.
4395 * Otherwise, whine if we are scheduling when we should not be.
4397 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4398 __schedule_bug(prev);
4400 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4402 schedstat_inc(this_rq(), sched_count);
4403 #ifdef CONFIG_SCHEDSTATS
4404 if (unlikely(prev->lock_depth >= 0)) {
4405 schedstat_inc(this_rq(), bkl_count);
4406 schedstat_inc(prev, sched_info.bkl_count);
4412 * Pick up the highest-prio task:
4414 static inline struct task_struct *
4415 pick_next_task(struct rq *rq, struct task_struct *prev)
4417 const struct sched_class *class;
4418 struct task_struct *p;
4421 * Optimization: we know that if all tasks are in
4422 * the fair class we can call that function directly:
4424 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4425 p = fair_sched_class.pick_next_task(rq);
4430 class = sched_class_highest;
4432 p = class->pick_next_task(rq);
4436 * Will never be NULL as the idle class always
4437 * returns a non-NULL p:
4439 class = class->next;
4444 * schedule() is the main scheduler function.
4446 asmlinkage void __sched schedule(void)
4448 struct task_struct *prev, *next;
4449 unsigned long *switch_count;
4455 cpu = smp_processor_id();
4459 switch_count = &prev->nivcsw;
4461 release_kernel_lock(prev);
4462 need_resched_nonpreemptible:
4464 schedule_debug(prev);
4466 if (sched_feat(HRTICK))
4469 spin_lock_irq(&rq->lock);
4470 update_rq_clock(rq);
4471 clear_tsk_need_resched(prev);
4473 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4474 if (unlikely(signal_pending_state(prev->state, prev)))
4475 prev->state = TASK_RUNNING;
4477 deactivate_task(rq, prev, 1);
4478 switch_count = &prev->nvcsw;
4482 if (prev->sched_class->pre_schedule)
4483 prev->sched_class->pre_schedule(rq, prev);
4486 if (unlikely(!rq->nr_running))
4487 idle_balance(cpu, rq);
4489 prev->sched_class->put_prev_task(rq, prev);
4490 next = pick_next_task(rq, prev);
4492 if (likely(prev != next)) {
4493 sched_info_switch(prev, next);
4499 context_switch(rq, prev, next); /* unlocks the rq */
4501 * the context switch might have flipped the stack from under
4502 * us, hence refresh the local variables.
4504 cpu = smp_processor_id();
4507 spin_unlock_irq(&rq->lock);
4509 if (unlikely(reacquire_kernel_lock(current) < 0))
4510 goto need_resched_nonpreemptible;
4512 preempt_enable_no_resched();
4513 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4516 EXPORT_SYMBOL(schedule);
4518 #ifdef CONFIG_PREEMPT
4520 * this is the entry point to schedule() from in-kernel preemption
4521 * off of preempt_enable. Kernel preemptions off return from interrupt
4522 * occur there and call schedule directly.
4524 asmlinkage void __sched preempt_schedule(void)
4526 struct thread_info *ti = current_thread_info();
4529 * If there is a non-zero preempt_count or interrupts are disabled,
4530 * we do not want to preempt the current task. Just return..
4532 if (likely(ti->preempt_count || irqs_disabled()))
4536 add_preempt_count(PREEMPT_ACTIVE);
4538 sub_preempt_count(PREEMPT_ACTIVE);
4541 * Check again in case we missed a preemption opportunity
4542 * between schedule and now.
4545 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4547 EXPORT_SYMBOL(preempt_schedule);
4550 * this is the entry point to schedule() from kernel preemption
4551 * off of irq context.
4552 * Note, that this is called and return with irqs disabled. This will
4553 * protect us against recursive calling from irq.
4555 asmlinkage void __sched preempt_schedule_irq(void)
4557 struct thread_info *ti = current_thread_info();
4559 /* Catch callers which need to be fixed */
4560 BUG_ON(ti->preempt_count || !irqs_disabled());
4563 add_preempt_count(PREEMPT_ACTIVE);
4566 local_irq_disable();
4567 sub_preempt_count(PREEMPT_ACTIVE);
4570 * Check again in case we missed a preemption opportunity
4571 * between schedule and now.
4574 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4577 #endif /* CONFIG_PREEMPT */
4579 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4582 return try_to_wake_up(curr->private, mode, sync);
4584 EXPORT_SYMBOL(default_wake_function);
4587 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4588 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4589 * number) then we wake all the non-exclusive tasks and one exclusive task.
4591 * There are circumstances in which we can try to wake a task which has already
4592 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4593 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4595 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4596 int nr_exclusive, int sync, void *key)
4598 wait_queue_t *curr, *next;
4600 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4601 unsigned flags = curr->flags;
4603 if (curr->func(curr, mode, sync, key) &&
4604 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4610 * __wake_up - wake up threads blocked on a waitqueue.
4612 * @mode: which threads
4613 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4614 * @key: is directly passed to the wakeup function
4616 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4617 int nr_exclusive, void *key)
4619 unsigned long flags;
4621 spin_lock_irqsave(&q->lock, flags);
4622 __wake_up_common(q, mode, nr_exclusive, 0, key);
4623 spin_unlock_irqrestore(&q->lock, flags);
4625 EXPORT_SYMBOL(__wake_up);
4628 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4630 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4632 __wake_up_common(q, mode, 1, 0, NULL);
4636 * __wake_up_sync - wake up threads blocked on a waitqueue.
4638 * @mode: which threads
4639 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4641 * The sync wakeup differs that the waker knows that it will schedule
4642 * away soon, so while the target thread will be woken up, it will not
4643 * be migrated to another CPU - ie. the two threads are 'synchronized'
4644 * with each other. This can prevent needless bouncing between CPUs.
4646 * On UP it can prevent extra preemption.
4649 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4651 unsigned long flags;
4657 if (unlikely(!nr_exclusive))
4660 spin_lock_irqsave(&q->lock, flags);
4661 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4662 spin_unlock_irqrestore(&q->lock, flags);
4664 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4667 * complete: - signals a single thread waiting on this completion
4668 * @x: holds the state of this particular completion
4670 * This will wake up a single thread waiting on this completion. Threads will be
4671 * awakened in the same order in which they were queued.
4673 * See also complete_all(), wait_for_completion() and related routines.
4675 void complete(struct completion *x)
4677 unsigned long flags;
4679 spin_lock_irqsave(&x->wait.lock, flags);
4681 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4682 spin_unlock_irqrestore(&x->wait.lock, flags);
4684 EXPORT_SYMBOL(complete);
4687 * complete_all: - signals all threads waiting on this completion
4688 * @x: holds the state of this particular completion
4690 * This will wake up all threads waiting on this particular completion event.
4692 void complete_all(struct completion *x)
4694 unsigned long flags;
4696 spin_lock_irqsave(&x->wait.lock, flags);
4697 x->done += UINT_MAX/2;
4698 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4699 spin_unlock_irqrestore(&x->wait.lock, flags);
4701 EXPORT_SYMBOL(complete_all);
4703 static inline long __sched
4704 do_wait_for_common(struct completion *x, long timeout, int state)
4707 DECLARE_WAITQUEUE(wait, current);
4709 wait.flags |= WQ_FLAG_EXCLUSIVE;
4710 __add_wait_queue_tail(&x->wait, &wait);
4712 if (signal_pending_state(state, current)) {
4713 timeout = -ERESTARTSYS;
4716 __set_current_state(state);
4717 spin_unlock_irq(&x->wait.lock);
4718 timeout = schedule_timeout(timeout);
4719 spin_lock_irq(&x->wait.lock);
4720 } while (!x->done && timeout);
4721 __remove_wait_queue(&x->wait, &wait);
4726 return timeout ?: 1;
4730 wait_for_common(struct completion *x, long timeout, int state)
4734 spin_lock_irq(&x->wait.lock);
4735 timeout = do_wait_for_common(x, timeout, state);
4736 spin_unlock_irq(&x->wait.lock);
4741 * wait_for_completion: - waits for completion of a task
4742 * @x: holds the state of this particular completion
4744 * This waits to be signaled for completion of a specific task. It is NOT
4745 * interruptible and there is no timeout.
4747 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4748 * and interrupt capability. Also see complete().
4750 void __sched wait_for_completion(struct completion *x)
4752 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4754 EXPORT_SYMBOL(wait_for_completion);
4757 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4758 * @x: holds the state of this particular completion
4759 * @timeout: timeout value in jiffies
4761 * This waits for either a completion of a specific task to be signaled or for a
4762 * specified timeout to expire. The timeout is in jiffies. It is not
4765 unsigned long __sched
4766 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4768 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4770 EXPORT_SYMBOL(wait_for_completion_timeout);
4773 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4774 * @x: holds the state of this particular completion
4776 * This waits for completion of a specific task to be signaled. It is
4779 int __sched wait_for_completion_interruptible(struct completion *x)
4781 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4782 if (t == -ERESTARTSYS)
4786 EXPORT_SYMBOL(wait_for_completion_interruptible);
4789 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4790 * @x: holds the state of this particular completion
4791 * @timeout: timeout value in jiffies
4793 * This waits for either a completion of a specific task to be signaled or for a
4794 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4796 unsigned long __sched
4797 wait_for_completion_interruptible_timeout(struct completion *x,
4798 unsigned long timeout)
4800 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4802 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4805 * wait_for_completion_killable: - waits for completion of a task (killable)
4806 * @x: holds the state of this particular completion
4808 * This waits to be signaled for completion of a specific task. It can be
4809 * interrupted by a kill signal.
4811 int __sched wait_for_completion_killable(struct completion *x)
4813 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4814 if (t == -ERESTARTSYS)
4818 EXPORT_SYMBOL(wait_for_completion_killable);
4821 * try_wait_for_completion - try to decrement a completion without blocking
4822 * @x: completion structure
4824 * Returns: 0 if a decrement cannot be done without blocking
4825 * 1 if a decrement succeeded.
4827 * If a completion is being used as a counting completion,
4828 * attempt to decrement the counter without blocking. This
4829 * enables us to avoid waiting if the resource the completion
4830 * is protecting is not available.
4832 bool try_wait_for_completion(struct completion *x)
4836 spin_lock_irq(&x->wait.lock);
4841 spin_unlock_irq(&x->wait.lock);
4844 EXPORT_SYMBOL(try_wait_for_completion);
4847 * completion_done - Test to see if a completion has any waiters
4848 * @x: completion structure
4850 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4851 * 1 if there are no waiters.
4854 bool completion_done(struct completion *x)
4858 spin_lock_irq(&x->wait.lock);
4861 spin_unlock_irq(&x->wait.lock);
4864 EXPORT_SYMBOL(completion_done);
4867 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4869 unsigned long flags;
4872 init_waitqueue_entry(&wait, current);
4874 __set_current_state(state);
4876 spin_lock_irqsave(&q->lock, flags);
4877 __add_wait_queue(q, &wait);
4878 spin_unlock(&q->lock);
4879 timeout = schedule_timeout(timeout);
4880 spin_lock_irq(&q->lock);
4881 __remove_wait_queue(q, &wait);
4882 spin_unlock_irqrestore(&q->lock, flags);
4887 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4889 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4891 EXPORT_SYMBOL(interruptible_sleep_on);
4894 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4896 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4898 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4900 void __sched sleep_on(wait_queue_head_t *q)
4902 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4904 EXPORT_SYMBOL(sleep_on);
4906 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4908 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4910 EXPORT_SYMBOL(sleep_on_timeout);
4912 #ifdef CONFIG_RT_MUTEXES
4915 * rt_mutex_setprio - set the current priority of a task
4917 * @prio: prio value (kernel-internal form)
4919 * This function changes the 'effective' priority of a task. It does
4920 * not touch ->normal_prio like __setscheduler().
4922 * Used by the rt_mutex code to implement priority inheritance logic.
4924 void rt_mutex_setprio(struct task_struct *p, int prio)
4926 unsigned long flags;
4927 int oldprio, on_rq, running;
4929 const struct sched_class *prev_class = p->sched_class;
4931 BUG_ON(prio < 0 || prio > MAX_PRIO);
4933 rq = task_rq_lock(p, &flags);
4934 update_rq_clock(rq);
4937 on_rq = p->se.on_rq;
4938 running = task_current(rq, p);
4940 dequeue_task(rq, p, 0);
4942 p->sched_class->put_prev_task(rq, p);
4945 p->sched_class = &rt_sched_class;
4947 p->sched_class = &fair_sched_class;
4952 p->sched_class->set_curr_task(rq);
4954 enqueue_task(rq, p, 0);
4956 check_class_changed(rq, p, prev_class, oldprio, running);
4958 task_rq_unlock(rq, &flags);
4963 void set_user_nice(struct task_struct *p, long nice)
4965 int old_prio, delta, on_rq;
4966 unsigned long flags;
4969 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4972 * We have to be careful, if called from sys_setpriority(),
4973 * the task might be in the middle of scheduling on another CPU.
4975 rq = task_rq_lock(p, &flags);
4976 update_rq_clock(rq);
4978 * The RT priorities are set via sched_setscheduler(), but we still
4979 * allow the 'normal' nice value to be set - but as expected
4980 * it wont have any effect on scheduling until the task is
4981 * SCHED_FIFO/SCHED_RR:
4983 if (task_has_rt_policy(p)) {
4984 p->static_prio = NICE_TO_PRIO(nice);
4987 on_rq = p->se.on_rq;
4989 dequeue_task(rq, p, 0);
4991 p->static_prio = NICE_TO_PRIO(nice);
4994 p->prio = effective_prio(p);
4995 delta = p->prio - old_prio;
4998 enqueue_task(rq, p, 0);
5000 * If the task increased its priority or is running and
5001 * lowered its priority, then reschedule its CPU:
5003 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5004 resched_task(rq->curr);
5007 task_rq_unlock(rq, &flags);
5009 EXPORT_SYMBOL(set_user_nice);
5012 * can_nice - check if a task can reduce its nice value
5016 int can_nice(const struct task_struct *p, const int nice)
5018 /* convert nice value [19,-20] to rlimit style value [1,40] */
5019 int nice_rlim = 20 - nice;
5021 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
5022 capable(CAP_SYS_NICE));
5025 #ifdef __ARCH_WANT_SYS_NICE
5028 * sys_nice - change the priority of the current process.
5029 * @increment: priority increment
5031 * sys_setpriority is a more generic, but much slower function that
5032 * does similar things.
5034 asmlinkage long sys_nice(int increment)
5039 * Setpriority might change our priority at the same moment.
5040 * We don't have to worry. Conceptually one call occurs first
5041 * and we have a single winner.
5043 if (increment < -40)
5048 nice = PRIO_TO_NICE(current->static_prio) + increment;
5054 if (increment < 0 && !can_nice(current, nice))
5057 retval = security_task_setnice(current, nice);
5061 set_user_nice(current, nice);
5068 * task_prio - return the priority value of a given task.
5069 * @p: the task in question.
5071 * This is the priority value as seen by users in /proc.
5072 * RT tasks are offset by -200. Normal tasks are centered
5073 * around 0, value goes from -16 to +15.
5075 int task_prio(const struct task_struct *p)
5077 return p->prio - MAX_RT_PRIO;
5081 * task_nice - return the nice value of a given task.
5082 * @p: the task in question.
5084 int task_nice(const struct task_struct *p)
5086 return TASK_NICE(p);
5088 EXPORT_SYMBOL(task_nice);
5091 * idle_cpu - is a given cpu idle currently?
5092 * @cpu: the processor in question.
5094 int idle_cpu(int cpu)
5096 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5100 * idle_task - return the idle task for a given cpu.
5101 * @cpu: the processor in question.
5103 struct task_struct *idle_task(int cpu)
5105 return cpu_rq(cpu)->idle;
5109 * find_process_by_pid - find a process with a matching PID value.
5110 * @pid: the pid in question.
5112 static struct task_struct *find_process_by_pid(pid_t pid)
5114 return pid ? find_task_by_vpid(pid) : current;
5117 /* Actually do priority change: must hold rq lock. */
5119 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5121 BUG_ON(p->se.on_rq);
5124 switch (p->policy) {
5128 p->sched_class = &fair_sched_class;
5132 p->sched_class = &rt_sched_class;
5136 p->rt_priority = prio;
5137 p->normal_prio = normal_prio(p);
5138 /* we are holding p->pi_lock already */
5139 p->prio = rt_mutex_getprio(p);
5143 static int __sched_setscheduler(struct task_struct *p, int policy,
5144 struct sched_param *param, bool user)
5146 int retval, oldprio, oldpolicy = -1, on_rq, running;
5147 unsigned long flags;
5148 const struct sched_class *prev_class = p->sched_class;
5151 /* may grab non-irq protected spin_locks */
5152 BUG_ON(in_interrupt());
5154 /* double check policy once rq lock held */
5156 policy = oldpolicy = p->policy;
5157 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5158 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5159 policy != SCHED_IDLE)
5162 * Valid priorities for SCHED_FIFO and SCHED_RR are
5163 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5164 * SCHED_BATCH and SCHED_IDLE is 0.
5166 if (param->sched_priority < 0 ||
5167 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5168 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5170 if (rt_policy(policy) != (param->sched_priority != 0))
5174 * Allow unprivileged RT tasks to decrease priority:
5176 if (user && !capable(CAP_SYS_NICE)) {
5177 if (rt_policy(policy)) {
5178 unsigned long rlim_rtprio;
5180 if (!lock_task_sighand(p, &flags))
5182 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5183 unlock_task_sighand(p, &flags);
5185 /* can't set/change the rt policy */
5186 if (policy != p->policy && !rlim_rtprio)
5189 /* can't increase priority */
5190 if (param->sched_priority > p->rt_priority &&
5191 param->sched_priority > rlim_rtprio)
5195 * Like positive nice levels, dont allow tasks to
5196 * move out of SCHED_IDLE either:
5198 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5201 /* can't change other user's priorities */
5202 if ((current->euid != p->euid) &&
5203 (current->euid != p->uid))
5208 #ifdef CONFIG_RT_GROUP_SCHED
5210 * Do not allow realtime tasks into groups that have no runtime
5213 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5214 task_group(p)->rt_bandwidth.rt_runtime == 0)
5218 retval = security_task_setscheduler(p, policy, param);
5224 * make sure no PI-waiters arrive (or leave) while we are
5225 * changing the priority of the task:
5227 spin_lock_irqsave(&p->pi_lock, flags);
5229 * To be able to change p->policy safely, the apropriate
5230 * runqueue lock must be held.
5232 rq = __task_rq_lock(p);
5233 /* recheck policy now with rq lock held */
5234 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5235 policy = oldpolicy = -1;
5236 __task_rq_unlock(rq);
5237 spin_unlock_irqrestore(&p->pi_lock, flags);
5240 update_rq_clock(rq);
5241 on_rq = p->se.on_rq;
5242 running = task_current(rq, p);
5244 deactivate_task(rq, p, 0);
5246 p->sched_class->put_prev_task(rq, p);
5249 __setscheduler(rq, p, policy, param->sched_priority);
5252 p->sched_class->set_curr_task(rq);
5254 activate_task(rq, p, 0);
5256 check_class_changed(rq, p, prev_class, oldprio, running);
5258 __task_rq_unlock(rq);
5259 spin_unlock_irqrestore(&p->pi_lock, flags);
5261 rt_mutex_adjust_pi(p);
5267 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5268 * @p: the task in question.
5269 * @policy: new policy.
5270 * @param: structure containing the new RT priority.
5272 * NOTE that the task may be already dead.
5274 int sched_setscheduler(struct task_struct *p, int policy,
5275 struct sched_param *param)
5277 return __sched_setscheduler(p, policy, param, true);
5279 EXPORT_SYMBOL_GPL(sched_setscheduler);
5282 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5283 * @p: the task in question.
5284 * @policy: new policy.
5285 * @param: structure containing the new RT priority.
5287 * Just like sched_setscheduler, only don't bother checking if the
5288 * current context has permission. For example, this is needed in
5289 * stop_machine(): we create temporary high priority worker threads,
5290 * but our caller might not have that capability.
5292 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5293 struct sched_param *param)
5295 return __sched_setscheduler(p, policy, param, false);
5299 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5301 struct sched_param lparam;
5302 struct task_struct *p;
5305 if (!param || pid < 0)
5307 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5312 p = find_process_by_pid(pid);
5314 retval = sched_setscheduler(p, policy, &lparam);
5321 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5322 * @pid: the pid in question.
5323 * @policy: new policy.
5324 * @param: structure containing the new RT priority.
5327 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5329 /* negative values for policy are not valid */
5333 return do_sched_setscheduler(pid, policy, param);
5337 * sys_sched_setparam - set/change the RT priority of a thread
5338 * @pid: the pid in question.
5339 * @param: structure containing the new RT priority.
5341 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5343 return do_sched_setscheduler(pid, -1, param);
5347 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5348 * @pid: the pid in question.
5350 asmlinkage long sys_sched_getscheduler(pid_t pid)
5352 struct task_struct *p;
5359 read_lock(&tasklist_lock);
5360 p = find_process_by_pid(pid);
5362 retval = security_task_getscheduler(p);
5366 read_unlock(&tasklist_lock);
5371 * sys_sched_getscheduler - get the RT priority of a thread
5372 * @pid: the pid in question.
5373 * @param: structure containing the RT priority.
5375 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5377 struct sched_param lp;
5378 struct task_struct *p;
5381 if (!param || pid < 0)
5384 read_lock(&tasklist_lock);
5385 p = find_process_by_pid(pid);
5390 retval = security_task_getscheduler(p);
5394 lp.sched_priority = p->rt_priority;
5395 read_unlock(&tasklist_lock);
5398 * This one might sleep, we cannot do it with a spinlock held ...
5400 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5405 read_unlock(&tasklist_lock);
5409 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5411 cpumask_t cpus_allowed;
5412 cpumask_t new_mask = *in_mask;
5413 struct task_struct *p;
5417 read_lock(&tasklist_lock);
5419 p = find_process_by_pid(pid);
5421 read_unlock(&tasklist_lock);
5427 * It is not safe to call set_cpus_allowed with the
5428 * tasklist_lock held. We will bump the task_struct's
5429 * usage count and then drop tasklist_lock.
5432 read_unlock(&tasklist_lock);
5435 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5436 !capable(CAP_SYS_NICE))
5439 retval = security_task_setscheduler(p, 0, NULL);
5443 cpuset_cpus_allowed(p, &cpus_allowed);
5444 cpus_and(new_mask, new_mask, cpus_allowed);
5446 retval = set_cpus_allowed_ptr(p, &new_mask);
5449 cpuset_cpus_allowed(p, &cpus_allowed);
5450 if (!cpus_subset(new_mask, cpus_allowed)) {
5452 * We must have raced with a concurrent cpuset
5453 * update. Just reset the cpus_allowed to the
5454 * cpuset's cpus_allowed
5456 new_mask = cpus_allowed;
5466 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5467 cpumask_t *new_mask)
5469 if (len < sizeof(cpumask_t)) {
5470 memset(new_mask, 0, sizeof(cpumask_t));
5471 } else if (len > sizeof(cpumask_t)) {
5472 len = sizeof(cpumask_t);
5474 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5478 * sys_sched_setaffinity - set the cpu affinity of a process
5479 * @pid: pid of the process
5480 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5481 * @user_mask_ptr: user-space pointer to the new cpu mask
5483 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5484 unsigned long __user *user_mask_ptr)
5489 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5493 return sched_setaffinity(pid, &new_mask);
5496 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5498 struct task_struct *p;
5502 read_lock(&tasklist_lock);
5505 p = find_process_by_pid(pid);
5509 retval = security_task_getscheduler(p);
5513 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5516 read_unlock(&tasklist_lock);
5523 * sys_sched_getaffinity - get the cpu affinity of a process
5524 * @pid: pid of the process
5525 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5526 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5528 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5529 unsigned long __user *user_mask_ptr)
5534 if (len < sizeof(cpumask_t))
5537 ret = sched_getaffinity(pid, &mask);
5541 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5544 return sizeof(cpumask_t);
5548 * sys_sched_yield - yield the current processor to other threads.
5550 * This function yields the current CPU to other tasks. If there are no
5551 * other threads running on this CPU then this function will return.
5553 asmlinkage long sys_sched_yield(void)
5555 struct rq *rq = this_rq_lock();
5557 schedstat_inc(rq, yld_count);
5558 current->sched_class->yield_task(rq);
5561 * Since we are going to call schedule() anyway, there's
5562 * no need to preempt or enable interrupts:
5564 __release(rq->lock);
5565 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5566 _raw_spin_unlock(&rq->lock);
5567 preempt_enable_no_resched();
5574 static void __cond_resched(void)
5576 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5577 __might_sleep(__FILE__, __LINE__);
5580 * The BKS might be reacquired before we have dropped
5581 * PREEMPT_ACTIVE, which could trigger a second
5582 * cond_resched() call.
5585 add_preempt_count(PREEMPT_ACTIVE);
5587 sub_preempt_count(PREEMPT_ACTIVE);
5588 } while (need_resched());
5591 int __sched _cond_resched(void)
5593 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5594 system_state == SYSTEM_RUNNING) {
5600 EXPORT_SYMBOL(_cond_resched);
5603 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5604 * call schedule, and on return reacquire the lock.
5606 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5607 * operations here to prevent schedule() from being called twice (once via
5608 * spin_unlock(), once by hand).
5610 int cond_resched_lock(spinlock_t *lock)
5612 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5615 if (spin_needbreak(lock) || resched) {
5617 if (resched && need_resched())
5626 EXPORT_SYMBOL(cond_resched_lock);
5628 int __sched cond_resched_softirq(void)
5630 BUG_ON(!in_softirq());
5632 if (need_resched() && system_state == SYSTEM_RUNNING) {
5640 EXPORT_SYMBOL(cond_resched_softirq);
5643 * yield - yield the current processor to other threads.
5645 * This is a shortcut for kernel-space yielding - it marks the
5646 * thread runnable and calls sys_sched_yield().
5648 void __sched yield(void)
5650 set_current_state(TASK_RUNNING);
5653 EXPORT_SYMBOL(yield);
5656 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5657 * that process accounting knows that this is a task in IO wait state.
5659 * But don't do that if it is a deliberate, throttling IO wait (this task
5660 * has set its backing_dev_info: the queue against which it should throttle)
5662 void __sched io_schedule(void)
5664 struct rq *rq = &__raw_get_cpu_var(runqueues);
5666 delayacct_blkio_start();
5667 atomic_inc(&rq->nr_iowait);
5669 atomic_dec(&rq->nr_iowait);
5670 delayacct_blkio_end();
5672 EXPORT_SYMBOL(io_schedule);
5674 long __sched io_schedule_timeout(long timeout)
5676 struct rq *rq = &__raw_get_cpu_var(runqueues);
5679 delayacct_blkio_start();
5680 atomic_inc(&rq->nr_iowait);
5681 ret = schedule_timeout(timeout);
5682 atomic_dec(&rq->nr_iowait);
5683 delayacct_blkio_end();
5688 * sys_sched_get_priority_max - return maximum RT priority.
5689 * @policy: scheduling class.
5691 * this syscall returns the maximum rt_priority that can be used
5692 * by a given scheduling class.
5694 asmlinkage long sys_sched_get_priority_max(int policy)
5701 ret = MAX_USER_RT_PRIO-1;
5713 * sys_sched_get_priority_min - return minimum RT priority.
5714 * @policy: scheduling class.
5716 * this syscall returns the minimum rt_priority that can be used
5717 * by a given scheduling class.
5719 asmlinkage long sys_sched_get_priority_min(int policy)
5737 * sys_sched_rr_get_interval - return the default timeslice of a process.
5738 * @pid: pid of the process.
5739 * @interval: userspace pointer to the timeslice value.
5741 * this syscall writes the default timeslice value of a given process
5742 * into the user-space timespec buffer. A value of '0' means infinity.
5745 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5747 struct task_struct *p;
5748 unsigned int time_slice;
5756 read_lock(&tasklist_lock);
5757 p = find_process_by_pid(pid);
5761 retval = security_task_getscheduler(p);
5766 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5767 * tasks that are on an otherwise idle runqueue:
5770 if (p->policy == SCHED_RR) {
5771 time_slice = DEF_TIMESLICE;
5772 } else if (p->policy != SCHED_FIFO) {
5773 struct sched_entity *se = &p->se;
5774 unsigned long flags;
5777 rq = task_rq_lock(p, &flags);
5778 if (rq->cfs.load.weight)
5779 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5780 task_rq_unlock(rq, &flags);
5782 read_unlock(&tasklist_lock);
5783 jiffies_to_timespec(time_slice, &t);
5784 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5788 read_unlock(&tasklist_lock);
5792 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5794 void sched_show_task(struct task_struct *p)
5796 unsigned long free = 0;
5799 state = p->state ? __ffs(p->state) + 1 : 0;
5800 printk(KERN_INFO "%-13.13s %c", p->comm,
5801 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5802 #if BITS_PER_LONG == 32
5803 if (state == TASK_RUNNING)
5804 printk(KERN_CONT " running ");
5806 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5808 if (state == TASK_RUNNING)
5809 printk(KERN_CONT " running task ");
5811 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5813 #ifdef CONFIG_DEBUG_STACK_USAGE
5815 unsigned long *n = end_of_stack(p);
5818 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5821 printk(KERN_CONT "%5lu %5d %6d\n", free,
5822 task_pid_nr(p), task_pid_nr(p->real_parent));
5824 show_stack(p, NULL);
5827 void show_state_filter(unsigned long state_filter)
5829 struct task_struct *g, *p;
5831 #if BITS_PER_LONG == 32
5833 " task PC stack pid father\n");
5836 " task PC stack pid father\n");
5838 read_lock(&tasklist_lock);
5839 do_each_thread(g, p) {
5841 * reset the NMI-timeout, listing all files on a slow
5842 * console might take alot of time:
5844 touch_nmi_watchdog();
5845 if (!state_filter || (p->state & state_filter))
5847 } while_each_thread(g, p);
5849 touch_all_softlockup_watchdogs();
5851 #ifdef CONFIG_SCHED_DEBUG
5852 sysrq_sched_debug_show();
5854 read_unlock(&tasklist_lock);
5856 * Only show locks if all tasks are dumped:
5858 if (state_filter == -1)
5859 debug_show_all_locks();
5862 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5864 idle->sched_class = &idle_sched_class;
5868 * init_idle - set up an idle thread for a given CPU
5869 * @idle: task in question
5870 * @cpu: cpu the idle task belongs to
5872 * NOTE: this function does not set the idle thread's NEED_RESCHED
5873 * flag, to make booting more robust.
5875 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5877 struct rq *rq = cpu_rq(cpu);
5878 unsigned long flags;
5880 spin_lock_irqsave(&rq->lock, flags);
5883 idle->se.exec_start = sched_clock();
5885 idle->prio = idle->normal_prio = MAX_PRIO;
5886 idle->cpus_allowed = cpumask_of_cpu(cpu);
5887 __set_task_cpu(idle, cpu);
5889 rq->curr = rq->idle = idle;
5890 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5893 spin_unlock_irqrestore(&rq->lock, flags);
5895 /* Set the preempt count _outside_ the spinlocks! */
5896 #if defined(CONFIG_PREEMPT)
5897 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5899 task_thread_info(idle)->preempt_count = 0;
5902 * The idle tasks have their own, simple scheduling class:
5904 idle->sched_class = &idle_sched_class;
5905 ftrace_graph_init_task(idle);
5909 * In a system that switches off the HZ timer nohz_cpu_mask
5910 * indicates which cpus entered this state. This is used
5911 * in the rcu update to wait only for active cpus. For system
5912 * which do not switch off the HZ timer nohz_cpu_mask should
5913 * always be CPU_MASK_NONE.
5915 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5918 * Increase the granularity value when there are more CPUs,
5919 * because with more CPUs the 'effective latency' as visible
5920 * to users decreases. But the relationship is not linear,
5921 * so pick a second-best guess by going with the log2 of the
5924 * This idea comes from the SD scheduler of Con Kolivas:
5926 static inline void sched_init_granularity(void)
5928 unsigned int factor = 1 + ilog2(num_online_cpus());
5929 const unsigned long limit = 200000000;
5931 sysctl_sched_min_granularity *= factor;
5932 if (sysctl_sched_min_granularity > limit)
5933 sysctl_sched_min_granularity = limit;
5935 sysctl_sched_latency *= factor;
5936 if (sysctl_sched_latency > limit)
5937 sysctl_sched_latency = limit;
5939 sysctl_sched_wakeup_granularity *= factor;
5941 sysctl_sched_shares_ratelimit *= factor;
5946 * This is how migration works:
5948 * 1) we queue a struct migration_req structure in the source CPU's
5949 * runqueue and wake up that CPU's migration thread.
5950 * 2) we down() the locked semaphore => thread blocks.
5951 * 3) migration thread wakes up (implicitly it forces the migrated
5952 * thread off the CPU)
5953 * 4) it gets the migration request and checks whether the migrated
5954 * task is still in the wrong runqueue.
5955 * 5) if it's in the wrong runqueue then the migration thread removes
5956 * it and puts it into the right queue.
5957 * 6) migration thread up()s the semaphore.
5958 * 7) we wake up and the migration is done.
5962 * Change a given task's CPU affinity. Migrate the thread to a
5963 * proper CPU and schedule it away if the CPU it's executing on
5964 * is removed from the allowed bitmask.
5966 * NOTE: the caller must have a valid reference to the task, the
5967 * task must not exit() & deallocate itself prematurely. The
5968 * call is not atomic; no spinlocks may be held.
5970 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5972 struct migration_req req;
5973 unsigned long flags;
5977 rq = task_rq_lock(p, &flags);
5978 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5983 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5984 !cpus_equal(p->cpus_allowed, *new_mask))) {
5989 if (p->sched_class->set_cpus_allowed)
5990 p->sched_class->set_cpus_allowed(p, new_mask);
5992 p->cpus_allowed = *new_mask;
5993 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5996 /* Can the task run on the task's current CPU? If so, we're done */
5997 if (cpu_isset(task_cpu(p), *new_mask))
6000 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
6001 /* Need help from migration thread: drop lock and wait. */
6002 task_rq_unlock(rq, &flags);
6003 wake_up_process(rq->migration_thread);
6004 wait_for_completion(&req.done);
6005 tlb_migrate_finish(p->mm);
6009 task_rq_unlock(rq, &flags);
6013 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6016 * Move (not current) task off this cpu, onto dest cpu. We're doing
6017 * this because either it can't run here any more (set_cpus_allowed()
6018 * away from this CPU, or CPU going down), or because we're
6019 * attempting to rebalance this task on exec (sched_exec).
6021 * So we race with normal scheduler movements, but that's OK, as long
6022 * as the task is no longer on this CPU.
6024 * Returns non-zero if task was successfully migrated.
6026 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6028 struct rq *rq_dest, *rq_src;
6031 if (unlikely(!cpu_active(dest_cpu)))
6034 rq_src = cpu_rq(src_cpu);
6035 rq_dest = cpu_rq(dest_cpu);
6037 double_rq_lock(rq_src, rq_dest);
6038 /* Already moved. */
6039 if (task_cpu(p) != src_cpu)
6041 /* Affinity changed (again). */
6042 if (!cpu_isset(dest_cpu, p->cpus_allowed))
6045 on_rq = p->se.on_rq;
6047 deactivate_task(rq_src, p, 0);
6049 set_task_cpu(p, dest_cpu);
6051 activate_task(rq_dest, p, 0);
6052 check_preempt_curr(rq_dest, p, 0);
6057 double_rq_unlock(rq_src, rq_dest);
6062 * migration_thread - this is a highprio system thread that performs
6063 * thread migration by bumping thread off CPU then 'pushing' onto
6066 static int migration_thread(void *data)
6068 int cpu = (long)data;
6072 BUG_ON(rq->migration_thread != current);
6074 set_current_state(TASK_INTERRUPTIBLE);
6075 while (!kthread_should_stop()) {
6076 struct migration_req *req;
6077 struct list_head *head;
6079 spin_lock_irq(&rq->lock);
6081 if (cpu_is_offline(cpu)) {
6082 spin_unlock_irq(&rq->lock);
6086 if (rq->active_balance) {
6087 active_load_balance(rq, cpu);
6088 rq->active_balance = 0;
6091 head = &rq->migration_queue;
6093 if (list_empty(head)) {
6094 spin_unlock_irq(&rq->lock);
6096 set_current_state(TASK_INTERRUPTIBLE);
6099 req = list_entry(head->next, struct migration_req, list);
6100 list_del_init(head->next);
6102 spin_unlock(&rq->lock);
6103 __migrate_task(req->task, cpu, req->dest_cpu);
6106 complete(&req->done);
6108 __set_current_state(TASK_RUNNING);
6112 /* Wait for kthread_stop */
6113 set_current_state(TASK_INTERRUPTIBLE);
6114 while (!kthread_should_stop()) {
6116 set_current_state(TASK_INTERRUPTIBLE);
6118 __set_current_state(TASK_RUNNING);
6122 #ifdef CONFIG_HOTPLUG_CPU
6124 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6128 local_irq_disable();
6129 ret = __migrate_task(p, src_cpu, dest_cpu);
6135 * Figure out where task on dead CPU should go, use force if necessary.
6136 * NOTE: interrupts should be disabled by the caller
6138 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6140 unsigned long flags;
6147 mask = node_to_cpumask(cpu_to_node(dead_cpu));
6148 cpus_and(mask, mask, p->cpus_allowed);
6149 dest_cpu = any_online_cpu(mask);
6151 /* On any allowed CPU? */
6152 if (dest_cpu >= nr_cpu_ids)
6153 dest_cpu = any_online_cpu(p->cpus_allowed);
6155 /* No more Mr. Nice Guy. */
6156 if (dest_cpu >= nr_cpu_ids) {
6157 cpumask_t cpus_allowed;
6159 cpuset_cpus_allowed_locked(p, &cpus_allowed);
6161 * Try to stay on the same cpuset, where the
6162 * current cpuset may be a subset of all cpus.
6163 * The cpuset_cpus_allowed_locked() variant of
6164 * cpuset_cpus_allowed() will not block. It must be
6165 * called within calls to cpuset_lock/cpuset_unlock.
6167 rq = task_rq_lock(p, &flags);
6168 p->cpus_allowed = cpus_allowed;
6169 dest_cpu = any_online_cpu(p->cpus_allowed);
6170 task_rq_unlock(rq, &flags);
6173 * Don't tell them about moving exiting tasks or
6174 * kernel threads (both mm NULL), since they never
6177 if (p->mm && printk_ratelimit()) {
6178 printk(KERN_INFO "process %d (%s) no "
6179 "longer affine to cpu%d\n",
6180 task_pid_nr(p), p->comm, dead_cpu);
6183 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
6187 * While a dead CPU has no uninterruptible tasks queued at this point,
6188 * it might still have a nonzero ->nr_uninterruptible counter, because
6189 * for performance reasons the counter is not stricly tracking tasks to
6190 * their home CPUs. So we just add the counter to another CPU's counter,
6191 * to keep the global sum constant after CPU-down:
6193 static void migrate_nr_uninterruptible(struct rq *rq_src)
6195 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
6196 unsigned long flags;
6198 local_irq_save(flags);
6199 double_rq_lock(rq_src, rq_dest);
6200 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6201 rq_src->nr_uninterruptible = 0;
6202 double_rq_unlock(rq_src, rq_dest);
6203 local_irq_restore(flags);
6206 /* Run through task list and migrate tasks from the dead cpu. */
6207 static void migrate_live_tasks(int src_cpu)
6209 struct task_struct *p, *t;
6211 read_lock(&tasklist_lock);
6213 do_each_thread(t, p) {
6217 if (task_cpu(p) == src_cpu)
6218 move_task_off_dead_cpu(src_cpu, p);
6219 } while_each_thread(t, p);
6221 read_unlock(&tasklist_lock);
6225 * Schedules idle task to be the next runnable task on current CPU.
6226 * It does so by boosting its priority to highest possible.
6227 * Used by CPU offline code.
6229 void sched_idle_next(void)
6231 int this_cpu = smp_processor_id();
6232 struct rq *rq = cpu_rq(this_cpu);
6233 struct task_struct *p = rq->idle;
6234 unsigned long flags;
6236 /* cpu has to be offline */
6237 BUG_ON(cpu_online(this_cpu));
6240 * Strictly not necessary since rest of the CPUs are stopped by now
6241 * and interrupts disabled on the current cpu.
6243 spin_lock_irqsave(&rq->lock, flags);
6245 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6247 update_rq_clock(rq);
6248 activate_task(rq, p, 0);
6250 spin_unlock_irqrestore(&rq->lock, flags);
6254 * Ensures that the idle task is using init_mm right before its cpu goes
6257 void idle_task_exit(void)
6259 struct mm_struct *mm = current->active_mm;
6261 BUG_ON(cpu_online(smp_processor_id()));
6264 switch_mm(mm, &init_mm, current);
6268 /* called under rq->lock with disabled interrupts */
6269 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6271 struct rq *rq = cpu_rq(dead_cpu);
6273 /* Must be exiting, otherwise would be on tasklist. */
6274 BUG_ON(!p->exit_state);
6276 /* Cannot have done final schedule yet: would have vanished. */
6277 BUG_ON(p->state == TASK_DEAD);
6282 * Drop lock around migration; if someone else moves it,
6283 * that's OK. No task can be added to this CPU, so iteration is
6286 spin_unlock_irq(&rq->lock);
6287 move_task_off_dead_cpu(dead_cpu, p);
6288 spin_lock_irq(&rq->lock);
6293 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6294 static void migrate_dead_tasks(unsigned int dead_cpu)
6296 struct rq *rq = cpu_rq(dead_cpu);
6297 struct task_struct *next;
6300 if (!rq->nr_running)
6302 update_rq_clock(rq);
6303 next = pick_next_task(rq, rq->curr);
6306 next->sched_class->put_prev_task(rq, next);
6307 migrate_dead(dead_cpu, next);
6311 #endif /* CONFIG_HOTPLUG_CPU */
6313 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6315 static struct ctl_table sd_ctl_dir[] = {
6317 .procname = "sched_domain",
6323 static struct ctl_table sd_ctl_root[] = {
6325 .ctl_name = CTL_KERN,
6326 .procname = "kernel",
6328 .child = sd_ctl_dir,
6333 static struct ctl_table *sd_alloc_ctl_entry(int n)
6335 struct ctl_table *entry =
6336 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6341 static void sd_free_ctl_entry(struct ctl_table **tablep)
6343 struct ctl_table *entry;
6346 * In the intermediate directories, both the child directory and
6347 * procname are dynamically allocated and could fail but the mode
6348 * will always be set. In the lowest directory the names are
6349 * static strings and all have proc handlers.
6351 for (entry = *tablep; entry->mode; entry++) {
6353 sd_free_ctl_entry(&entry->child);
6354 if (entry->proc_handler == NULL)
6355 kfree(entry->procname);
6363 set_table_entry(struct ctl_table *entry,
6364 const char *procname, void *data, int maxlen,
6365 mode_t mode, proc_handler *proc_handler)
6367 entry->procname = procname;
6369 entry->maxlen = maxlen;
6371 entry->proc_handler = proc_handler;
6374 static struct ctl_table *
6375 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6377 struct ctl_table *table = sd_alloc_ctl_entry(13);
6382 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6383 sizeof(long), 0644, proc_doulongvec_minmax);
6384 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6385 sizeof(long), 0644, proc_doulongvec_minmax);
6386 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6387 sizeof(int), 0644, proc_dointvec_minmax);
6388 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6389 sizeof(int), 0644, proc_dointvec_minmax);
6390 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6391 sizeof(int), 0644, proc_dointvec_minmax);
6392 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6393 sizeof(int), 0644, proc_dointvec_minmax);
6394 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6395 sizeof(int), 0644, proc_dointvec_minmax);
6396 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6397 sizeof(int), 0644, proc_dointvec_minmax);
6398 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6399 sizeof(int), 0644, proc_dointvec_minmax);
6400 set_table_entry(&table[9], "cache_nice_tries",
6401 &sd->cache_nice_tries,
6402 sizeof(int), 0644, proc_dointvec_minmax);
6403 set_table_entry(&table[10], "flags", &sd->flags,
6404 sizeof(int), 0644, proc_dointvec_minmax);
6405 set_table_entry(&table[11], "name", sd->name,
6406 CORENAME_MAX_SIZE, 0444, proc_dostring);
6407 /* &table[12] is terminator */
6412 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6414 struct ctl_table *entry, *table;
6415 struct sched_domain *sd;
6416 int domain_num = 0, i;
6419 for_each_domain(cpu, sd)
6421 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6426 for_each_domain(cpu, sd) {
6427 snprintf(buf, 32, "domain%d", i);
6428 entry->procname = kstrdup(buf, GFP_KERNEL);
6430 entry->child = sd_alloc_ctl_domain_table(sd);
6437 static struct ctl_table_header *sd_sysctl_header;
6438 static void register_sched_domain_sysctl(void)
6440 int i, cpu_num = num_online_cpus();
6441 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6444 WARN_ON(sd_ctl_dir[0].child);
6445 sd_ctl_dir[0].child = entry;
6450 for_each_online_cpu(i) {
6451 snprintf(buf, 32, "cpu%d", i);
6452 entry->procname = kstrdup(buf, GFP_KERNEL);
6454 entry->child = sd_alloc_ctl_cpu_table(i);
6458 WARN_ON(sd_sysctl_header);
6459 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6462 /* may be called multiple times per register */
6463 static void unregister_sched_domain_sysctl(void)
6465 if (sd_sysctl_header)
6466 unregister_sysctl_table(sd_sysctl_header);
6467 sd_sysctl_header = NULL;
6468 if (sd_ctl_dir[0].child)
6469 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6472 static void register_sched_domain_sysctl(void)
6475 static void unregister_sched_domain_sysctl(void)
6480 static void set_rq_online(struct rq *rq)
6483 const struct sched_class *class;
6485 cpu_set(rq->cpu, rq->rd->online);
6488 for_each_class(class) {
6489 if (class->rq_online)
6490 class->rq_online(rq);
6495 static void set_rq_offline(struct rq *rq)
6498 const struct sched_class *class;
6500 for_each_class(class) {
6501 if (class->rq_offline)
6502 class->rq_offline(rq);
6505 cpu_clear(rq->cpu, rq->rd->online);
6511 * migration_call - callback that gets triggered when a CPU is added.
6512 * Here we can start up the necessary migration thread for the new CPU.
6514 static int __cpuinit
6515 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6517 struct task_struct *p;
6518 int cpu = (long)hcpu;
6519 unsigned long flags;
6524 case CPU_UP_PREPARE:
6525 case CPU_UP_PREPARE_FROZEN:
6526 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6529 kthread_bind(p, cpu);
6530 /* Must be high prio: stop_machine expects to yield to it. */
6531 rq = task_rq_lock(p, &flags);
6532 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6533 task_rq_unlock(rq, &flags);
6534 cpu_rq(cpu)->migration_thread = p;
6538 case CPU_ONLINE_FROZEN:
6539 /* Strictly unnecessary, as first user will wake it. */
6540 wake_up_process(cpu_rq(cpu)->migration_thread);
6542 /* Update our root-domain */
6544 spin_lock_irqsave(&rq->lock, flags);
6546 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6550 spin_unlock_irqrestore(&rq->lock, flags);
6553 #ifdef CONFIG_HOTPLUG_CPU
6554 case CPU_UP_CANCELED:
6555 case CPU_UP_CANCELED_FROZEN:
6556 if (!cpu_rq(cpu)->migration_thread)
6558 /* Unbind it from offline cpu so it can run. Fall thru. */
6559 kthread_bind(cpu_rq(cpu)->migration_thread,
6560 any_online_cpu(cpu_online_map));
6561 kthread_stop(cpu_rq(cpu)->migration_thread);
6562 cpu_rq(cpu)->migration_thread = NULL;
6566 case CPU_DEAD_FROZEN:
6567 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6568 migrate_live_tasks(cpu);
6570 kthread_stop(rq->migration_thread);
6571 rq->migration_thread = NULL;
6572 /* Idle task back to normal (off runqueue, low prio) */
6573 spin_lock_irq(&rq->lock);
6574 update_rq_clock(rq);
6575 deactivate_task(rq, rq->idle, 0);
6576 rq->idle->static_prio = MAX_PRIO;
6577 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6578 rq->idle->sched_class = &idle_sched_class;
6579 migrate_dead_tasks(cpu);
6580 spin_unlock_irq(&rq->lock);
6582 migrate_nr_uninterruptible(rq);
6583 BUG_ON(rq->nr_running != 0);
6586 * No need to migrate the tasks: it was best-effort if
6587 * they didn't take sched_hotcpu_mutex. Just wake up
6590 spin_lock_irq(&rq->lock);
6591 while (!list_empty(&rq->migration_queue)) {
6592 struct migration_req *req;
6594 req = list_entry(rq->migration_queue.next,
6595 struct migration_req, list);
6596 list_del_init(&req->list);
6597 complete(&req->done);
6599 spin_unlock_irq(&rq->lock);
6603 case CPU_DYING_FROZEN:
6604 /* Update our root-domain */
6606 spin_lock_irqsave(&rq->lock, flags);
6608 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6611 spin_unlock_irqrestore(&rq->lock, flags);
6618 /* Register at highest priority so that task migration (migrate_all_tasks)
6619 * happens before everything else.
6621 static struct notifier_block __cpuinitdata migration_notifier = {
6622 .notifier_call = migration_call,
6626 static int __init migration_init(void)
6628 void *cpu = (void *)(long)smp_processor_id();
6631 /* Start one for the boot CPU: */
6632 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6633 BUG_ON(err == NOTIFY_BAD);
6634 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6635 register_cpu_notifier(&migration_notifier);
6639 early_initcall(migration_init);
6644 #ifdef CONFIG_SCHED_DEBUG
6646 static inline const char *sd_level_to_string(enum sched_domain_level lvl)
6659 case SD_LV_ALLNODES:
6668 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6669 cpumask_t *groupmask)
6671 struct sched_group *group = sd->groups;
6674 cpulist_scnprintf(str, sizeof(str), sd->span);
6675 cpus_clear(*groupmask);
6677 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6679 if (!(sd->flags & SD_LOAD_BALANCE)) {
6680 printk("does not load-balance\n");
6682 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6687 printk(KERN_CONT "span %s level %s\n",
6688 str, sd_level_to_string(sd->level));
6690 if (!cpu_isset(cpu, sd->span)) {
6691 printk(KERN_ERR "ERROR: domain->span does not contain "
6694 if (!cpu_isset(cpu, group->cpumask)) {
6695 printk(KERN_ERR "ERROR: domain->groups does not contain"
6699 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6703 printk(KERN_ERR "ERROR: group is NULL\n");
6707 if (!group->__cpu_power) {
6708 printk(KERN_CONT "\n");
6709 printk(KERN_ERR "ERROR: domain->cpu_power not "
6714 if (!cpus_weight(group->cpumask)) {
6715 printk(KERN_CONT "\n");
6716 printk(KERN_ERR "ERROR: empty group\n");
6720 if (cpus_intersects(*groupmask, group->cpumask)) {
6721 printk(KERN_CONT "\n");
6722 printk(KERN_ERR "ERROR: repeated CPUs\n");
6726 cpus_or(*groupmask, *groupmask, group->cpumask);
6728 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6729 printk(KERN_CONT " %s", str);
6731 group = group->next;
6732 } while (group != sd->groups);
6733 printk(KERN_CONT "\n");
6735 if (!cpus_equal(sd->span, *groupmask))
6736 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6738 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6739 printk(KERN_ERR "ERROR: parent span is not a superset "
6740 "of domain->span\n");
6744 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6746 cpumask_t *groupmask;
6750 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6754 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6756 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6758 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6763 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6772 #else /* !CONFIG_SCHED_DEBUG */
6773 # define sched_domain_debug(sd, cpu) do { } while (0)
6774 #endif /* CONFIG_SCHED_DEBUG */
6776 static int sd_degenerate(struct sched_domain *sd)
6778 if (cpus_weight(sd->span) == 1)
6781 /* Following flags need at least 2 groups */
6782 if (sd->flags & (SD_LOAD_BALANCE |
6783 SD_BALANCE_NEWIDLE |
6787 SD_SHARE_PKG_RESOURCES)) {
6788 if (sd->groups != sd->groups->next)
6792 /* Following flags don't use groups */
6793 if (sd->flags & (SD_WAKE_IDLE |
6802 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6804 unsigned long cflags = sd->flags, pflags = parent->flags;
6806 if (sd_degenerate(parent))
6809 if (!cpus_equal(sd->span, parent->span))
6812 /* Does parent contain flags not in child? */
6813 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6814 if (cflags & SD_WAKE_AFFINE)
6815 pflags &= ~SD_WAKE_BALANCE;
6816 /* Flags needing groups don't count if only 1 group in parent */
6817 if (parent->groups == parent->groups->next) {
6818 pflags &= ~(SD_LOAD_BALANCE |
6819 SD_BALANCE_NEWIDLE |
6823 SD_SHARE_PKG_RESOURCES);
6825 if (~cflags & pflags)
6831 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6833 unsigned long flags;
6835 spin_lock_irqsave(&rq->lock, flags);
6838 struct root_domain *old_rd = rq->rd;
6840 if (cpu_isset(rq->cpu, old_rd->online))
6843 cpu_clear(rq->cpu, old_rd->span);
6845 if (atomic_dec_and_test(&old_rd->refcount))
6849 atomic_inc(&rd->refcount);
6852 cpu_set(rq->cpu, rd->span);
6853 if (cpu_isset(rq->cpu, cpu_online_map))
6856 spin_unlock_irqrestore(&rq->lock, flags);
6859 static void init_rootdomain(struct root_domain *rd)
6861 memset(rd, 0, sizeof(*rd));
6863 cpus_clear(rd->span);
6864 cpus_clear(rd->online);
6866 cpupri_init(&rd->cpupri);
6869 static void init_defrootdomain(void)
6871 init_rootdomain(&def_root_domain);
6872 atomic_set(&def_root_domain.refcount, 1);
6875 static struct root_domain *alloc_rootdomain(void)
6877 struct root_domain *rd;
6879 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6883 init_rootdomain(rd);
6889 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6890 * hold the hotplug lock.
6893 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6895 struct rq *rq = cpu_rq(cpu);
6896 struct sched_domain *tmp;
6898 /* Remove the sched domains which do not contribute to scheduling. */
6899 for (tmp = sd; tmp; ) {
6900 struct sched_domain *parent = tmp->parent;
6904 if (sd_parent_degenerate(tmp, parent)) {
6905 tmp->parent = parent->parent;
6907 parent->parent->child = tmp;
6912 if (sd && sd_degenerate(sd)) {
6918 sched_domain_debug(sd, cpu);
6920 rq_attach_root(rq, rd);
6921 rcu_assign_pointer(rq->sd, sd);
6924 /* cpus with isolated domains */
6925 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6927 /* Setup the mask of cpus configured for isolated domains */
6928 static int __init isolated_cpu_setup(char *str)
6930 static int __initdata ints[NR_CPUS];
6933 str = get_options(str, ARRAY_SIZE(ints), ints);
6934 cpus_clear(cpu_isolated_map);
6935 for (i = 1; i <= ints[0]; i++)
6936 if (ints[i] < NR_CPUS)
6937 cpu_set(ints[i], cpu_isolated_map);
6941 __setup("isolcpus=", isolated_cpu_setup);
6944 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6945 * to a function which identifies what group(along with sched group) a CPU
6946 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6947 * (due to the fact that we keep track of groups covered with a cpumask_t).
6949 * init_sched_build_groups will build a circular linked list of the groups
6950 * covered by the given span, and will set each group's ->cpumask correctly,
6951 * and ->cpu_power to 0.
6954 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6955 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6956 struct sched_group **sg,
6957 cpumask_t *tmpmask),
6958 cpumask_t *covered, cpumask_t *tmpmask)
6960 struct sched_group *first = NULL, *last = NULL;
6963 cpus_clear(*covered);
6965 for_each_cpu_mask_nr(i, *span) {
6966 struct sched_group *sg;
6967 int group = group_fn(i, cpu_map, &sg, tmpmask);
6970 if (cpu_isset(i, *covered))
6973 cpus_clear(sg->cpumask);
6974 sg->__cpu_power = 0;
6976 for_each_cpu_mask_nr(j, *span) {
6977 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6980 cpu_set(j, *covered);
6981 cpu_set(j, sg->cpumask);
6992 #define SD_NODES_PER_DOMAIN 16
6997 * find_next_best_node - find the next node to include in a sched_domain
6998 * @node: node whose sched_domain we're building
6999 * @used_nodes: nodes already in the sched_domain
7001 * Find the next node to include in a given scheduling domain. Simply
7002 * finds the closest node not already in the @used_nodes map.
7004 * Should use nodemask_t.
7006 static int find_next_best_node(int node, nodemask_t *used_nodes)
7008 int i, n, val, min_val, best_node = 0;
7012 for (i = 0; i < nr_node_ids; i++) {
7013 /* Start at @node */
7014 n = (node + i) % nr_node_ids;
7016 if (!nr_cpus_node(n))
7019 /* Skip already used nodes */
7020 if (node_isset(n, *used_nodes))
7023 /* Simple min distance search */
7024 val = node_distance(node, n);
7026 if (val < min_val) {
7032 node_set(best_node, *used_nodes);
7037 * sched_domain_node_span - get a cpumask for a node's sched_domain
7038 * @node: node whose cpumask we're constructing
7039 * @span: resulting cpumask
7041 * Given a node, construct a good cpumask for its sched_domain to span. It
7042 * should be one that prevents unnecessary balancing, but also spreads tasks
7045 static void sched_domain_node_span(int node, cpumask_t *span)
7047 nodemask_t used_nodes;
7048 node_to_cpumask_ptr(nodemask, node);
7052 nodes_clear(used_nodes);
7054 cpus_or(*span, *span, *nodemask);
7055 node_set(node, used_nodes);
7057 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7058 int next_node = find_next_best_node(node, &used_nodes);
7060 node_to_cpumask_ptr_next(nodemask, next_node);
7061 cpus_or(*span, *span, *nodemask);
7064 #endif /* CONFIG_NUMA */
7066 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7069 * SMT sched-domains:
7071 #ifdef CONFIG_SCHED_SMT
7072 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
7073 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
7076 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7080 *sg = &per_cpu(sched_group_cpus, cpu);
7083 #endif /* CONFIG_SCHED_SMT */
7086 * multi-core sched-domains:
7088 #ifdef CONFIG_SCHED_MC
7089 static DEFINE_PER_CPU(struct sched_domain, core_domains);
7090 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
7091 #endif /* CONFIG_SCHED_MC */
7093 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7095 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7100 *mask = per_cpu(cpu_sibling_map, cpu);
7101 cpus_and(*mask, *mask, *cpu_map);
7102 group = first_cpu(*mask);
7104 *sg = &per_cpu(sched_group_core, group);
7107 #elif defined(CONFIG_SCHED_MC)
7109 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7113 *sg = &per_cpu(sched_group_core, cpu);
7118 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
7119 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
7122 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7126 #ifdef CONFIG_SCHED_MC
7127 *mask = cpu_coregroup_map(cpu);
7128 cpus_and(*mask, *mask, *cpu_map);
7129 group = first_cpu(*mask);
7130 #elif defined(CONFIG_SCHED_SMT)
7131 *mask = per_cpu(cpu_sibling_map, cpu);
7132 cpus_and(*mask, *mask, *cpu_map);
7133 group = first_cpu(*mask);
7138 *sg = &per_cpu(sched_group_phys, group);
7144 * The init_sched_build_groups can't handle what we want to do with node
7145 * groups, so roll our own. Now each node has its own list of groups which
7146 * gets dynamically allocated.
7148 static DEFINE_PER_CPU(struct sched_domain, node_domains);
7149 static struct sched_group ***sched_group_nodes_bycpu;
7151 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
7152 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
7154 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
7155 struct sched_group **sg, cpumask_t *nodemask)
7159 *nodemask = node_to_cpumask(cpu_to_node(cpu));
7160 cpus_and(*nodemask, *nodemask, *cpu_map);
7161 group = first_cpu(*nodemask);
7164 *sg = &per_cpu(sched_group_allnodes, group);
7168 static void init_numa_sched_groups_power(struct sched_group *group_head)
7170 struct sched_group *sg = group_head;
7176 for_each_cpu_mask_nr(j, sg->cpumask) {
7177 struct sched_domain *sd;
7179 sd = &per_cpu(phys_domains, j);
7180 if (j != first_cpu(sd->groups->cpumask)) {
7182 * Only add "power" once for each
7188 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7191 } while (sg != group_head);
7193 #endif /* CONFIG_NUMA */
7196 /* Free memory allocated for various sched_group structures */
7197 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7201 for_each_cpu_mask_nr(cpu, *cpu_map) {
7202 struct sched_group **sched_group_nodes
7203 = sched_group_nodes_bycpu[cpu];
7205 if (!sched_group_nodes)
7208 for (i = 0; i < nr_node_ids; i++) {
7209 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7211 *nodemask = node_to_cpumask(i);
7212 cpus_and(*nodemask, *nodemask, *cpu_map);
7213 if (cpus_empty(*nodemask))
7223 if (oldsg != sched_group_nodes[i])
7226 kfree(sched_group_nodes);
7227 sched_group_nodes_bycpu[cpu] = NULL;
7230 #else /* !CONFIG_NUMA */
7231 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7234 #endif /* CONFIG_NUMA */
7237 * Initialize sched groups cpu_power.
7239 * cpu_power indicates the capacity of sched group, which is used while
7240 * distributing the load between different sched groups in a sched domain.
7241 * Typically cpu_power for all the groups in a sched domain will be same unless
7242 * there are asymmetries in the topology. If there are asymmetries, group
7243 * having more cpu_power will pickup more load compared to the group having
7246 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7247 * the maximum number of tasks a group can handle in the presence of other idle
7248 * or lightly loaded groups in the same sched domain.
7250 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7252 struct sched_domain *child;
7253 struct sched_group *group;
7255 WARN_ON(!sd || !sd->groups);
7257 if (cpu != first_cpu(sd->groups->cpumask))
7262 sd->groups->__cpu_power = 0;
7265 * For perf policy, if the groups in child domain share resources
7266 * (for example cores sharing some portions of the cache hierarchy
7267 * or SMT), then set this domain groups cpu_power such that each group
7268 * can handle only one task, when there are other idle groups in the
7269 * same sched domain.
7271 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7273 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7274 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7279 * add cpu_power of each child group to this groups cpu_power
7281 group = child->groups;
7283 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7284 group = group->next;
7285 } while (group != child->groups);
7289 * Initializers for schedule domains
7290 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7293 #ifdef CONFIG_SCHED_DEBUG
7294 # define SD_INIT_NAME(sd, type) sd->name = #type
7296 # define SD_INIT_NAME(sd, type) do { } while (0)
7299 #define SD_INIT(sd, type) sd_init_##type(sd)
7301 #define SD_INIT_FUNC(type) \
7302 static noinline void sd_init_##type(struct sched_domain *sd) \
7304 memset(sd, 0, sizeof(*sd)); \
7305 *sd = SD_##type##_INIT; \
7306 sd->level = SD_LV_##type; \
7307 SD_INIT_NAME(sd, type); \
7312 SD_INIT_FUNC(ALLNODES)
7315 #ifdef CONFIG_SCHED_SMT
7316 SD_INIT_FUNC(SIBLING)
7318 #ifdef CONFIG_SCHED_MC
7323 * To minimize stack usage kmalloc room for cpumasks and share the
7324 * space as the usage in build_sched_domains() dictates. Used only
7325 * if the amount of space is significant.
7328 cpumask_t tmpmask; /* make this one first */
7331 cpumask_t this_sibling_map;
7332 cpumask_t this_core_map;
7334 cpumask_t send_covered;
7337 cpumask_t domainspan;
7339 cpumask_t notcovered;
7344 #define SCHED_CPUMASK_ALLOC 1
7345 #define SCHED_CPUMASK_FREE(v) kfree(v)
7346 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7348 #define SCHED_CPUMASK_ALLOC 0
7349 #define SCHED_CPUMASK_FREE(v)
7350 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7353 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7354 ((unsigned long)(a) + offsetof(struct allmasks, v))
7356 static int default_relax_domain_level = -1;
7358 static int __init setup_relax_domain_level(char *str)
7362 val = simple_strtoul(str, NULL, 0);
7363 if (val < SD_LV_MAX)
7364 default_relax_domain_level = val;
7368 __setup("relax_domain_level=", setup_relax_domain_level);
7370 static void set_domain_attribute(struct sched_domain *sd,
7371 struct sched_domain_attr *attr)
7375 if (!attr || attr->relax_domain_level < 0) {
7376 if (default_relax_domain_level < 0)
7379 request = default_relax_domain_level;
7381 request = attr->relax_domain_level;
7382 if (request < sd->level) {
7383 /* turn off idle balance on this domain */
7384 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7386 /* turn on idle balance on this domain */
7387 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7392 * Build sched domains for a given set of cpus and attach the sched domains
7393 * to the individual cpus
7395 static int __build_sched_domains(const cpumask_t *cpu_map,
7396 struct sched_domain_attr *attr)
7399 struct root_domain *rd;
7400 SCHED_CPUMASK_DECLARE(allmasks);
7403 struct sched_group **sched_group_nodes = NULL;
7404 int sd_allnodes = 0;
7407 * Allocate the per-node list of sched groups
7409 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
7411 if (!sched_group_nodes) {
7412 printk(KERN_WARNING "Can not alloc sched group node list\n");
7417 rd = alloc_rootdomain();
7419 printk(KERN_WARNING "Cannot alloc root domain\n");
7421 kfree(sched_group_nodes);
7426 #if SCHED_CPUMASK_ALLOC
7427 /* get space for all scratch cpumask variables */
7428 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
7430 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7433 kfree(sched_group_nodes);
7438 tmpmask = (cpumask_t *)allmasks;
7442 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7446 * Set up domains for cpus specified by the cpu_map.
7448 for_each_cpu_mask_nr(i, *cpu_map) {
7449 struct sched_domain *sd = NULL, *p;
7450 SCHED_CPUMASK_VAR(nodemask, allmasks);
7452 *nodemask = node_to_cpumask(cpu_to_node(i));
7453 cpus_and(*nodemask, *nodemask, *cpu_map);
7456 if (cpus_weight(*cpu_map) >
7457 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7458 sd = &per_cpu(allnodes_domains, i);
7459 SD_INIT(sd, ALLNODES);
7460 set_domain_attribute(sd, attr);
7461 sd->span = *cpu_map;
7462 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7468 sd = &per_cpu(node_domains, i);
7470 set_domain_attribute(sd, attr);
7471 sched_domain_node_span(cpu_to_node(i), &sd->span);
7475 cpus_and(sd->span, sd->span, *cpu_map);
7479 sd = &per_cpu(phys_domains, i);
7481 set_domain_attribute(sd, attr);
7482 sd->span = *nodemask;
7486 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7488 #ifdef CONFIG_SCHED_MC
7490 sd = &per_cpu(core_domains, i);
7492 set_domain_attribute(sd, attr);
7493 sd->span = cpu_coregroup_map(i);
7494 cpus_and(sd->span, sd->span, *cpu_map);
7497 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7500 #ifdef CONFIG_SCHED_SMT
7502 sd = &per_cpu(cpu_domains, i);
7503 SD_INIT(sd, SIBLING);
7504 set_domain_attribute(sd, attr);
7505 sd->span = per_cpu(cpu_sibling_map, i);
7506 cpus_and(sd->span, sd->span, *cpu_map);
7509 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7513 #ifdef CONFIG_SCHED_SMT
7514 /* Set up CPU (sibling) groups */
7515 for_each_cpu_mask_nr(i, *cpu_map) {
7516 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7517 SCHED_CPUMASK_VAR(send_covered, allmasks);
7519 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7520 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7521 if (i != first_cpu(*this_sibling_map))
7524 init_sched_build_groups(this_sibling_map, cpu_map,
7526 send_covered, tmpmask);
7530 #ifdef CONFIG_SCHED_MC
7531 /* Set up multi-core groups */
7532 for_each_cpu_mask_nr(i, *cpu_map) {
7533 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7534 SCHED_CPUMASK_VAR(send_covered, allmasks);
7536 *this_core_map = cpu_coregroup_map(i);
7537 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7538 if (i != first_cpu(*this_core_map))
7541 init_sched_build_groups(this_core_map, cpu_map,
7543 send_covered, tmpmask);
7547 /* Set up physical groups */
7548 for (i = 0; i < nr_node_ids; i++) {
7549 SCHED_CPUMASK_VAR(nodemask, allmasks);
7550 SCHED_CPUMASK_VAR(send_covered, allmasks);
7552 *nodemask = node_to_cpumask(i);
7553 cpus_and(*nodemask, *nodemask, *cpu_map);
7554 if (cpus_empty(*nodemask))
7557 init_sched_build_groups(nodemask, cpu_map,
7559 send_covered, tmpmask);
7563 /* Set up node groups */
7565 SCHED_CPUMASK_VAR(send_covered, allmasks);
7567 init_sched_build_groups(cpu_map, cpu_map,
7568 &cpu_to_allnodes_group,
7569 send_covered, tmpmask);
7572 for (i = 0; i < nr_node_ids; i++) {
7573 /* Set up node groups */
7574 struct sched_group *sg, *prev;
7575 SCHED_CPUMASK_VAR(nodemask, allmasks);
7576 SCHED_CPUMASK_VAR(domainspan, allmasks);
7577 SCHED_CPUMASK_VAR(covered, allmasks);
7580 *nodemask = node_to_cpumask(i);
7581 cpus_clear(*covered);
7583 cpus_and(*nodemask, *nodemask, *cpu_map);
7584 if (cpus_empty(*nodemask)) {
7585 sched_group_nodes[i] = NULL;
7589 sched_domain_node_span(i, domainspan);
7590 cpus_and(*domainspan, *domainspan, *cpu_map);
7592 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7594 printk(KERN_WARNING "Can not alloc domain group for "
7598 sched_group_nodes[i] = sg;
7599 for_each_cpu_mask_nr(j, *nodemask) {
7600 struct sched_domain *sd;
7602 sd = &per_cpu(node_domains, j);
7605 sg->__cpu_power = 0;
7606 sg->cpumask = *nodemask;
7608 cpus_or(*covered, *covered, *nodemask);
7611 for (j = 0; j < nr_node_ids; j++) {
7612 SCHED_CPUMASK_VAR(notcovered, allmasks);
7613 int n = (i + j) % nr_node_ids;
7614 node_to_cpumask_ptr(pnodemask, n);
7616 cpus_complement(*notcovered, *covered);
7617 cpus_and(*tmpmask, *notcovered, *cpu_map);
7618 cpus_and(*tmpmask, *tmpmask, *domainspan);
7619 if (cpus_empty(*tmpmask))
7622 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7623 if (cpus_empty(*tmpmask))
7626 sg = kmalloc_node(sizeof(struct sched_group),
7630 "Can not alloc domain group for node %d\n", j);
7633 sg->__cpu_power = 0;
7634 sg->cpumask = *tmpmask;
7635 sg->next = prev->next;
7636 cpus_or(*covered, *covered, *tmpmask);
7643 /* Calculate CPU power for physical packages and nodes */
7644 #ifdef CONFIG_SCHED_SMT
7645 for_each_cpu_mask_nr(i, *cpu_map) {
7646 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7648 init_sched_groups_power(i, sd);
7651 #ifdef CONFIG_SCHED_MC
7652 for_each_cpu_mask_nr(i, *cpu_map) {
7653 struct sched_domain *sd = &per_cpu(core_domains, i);
7655 init_sched_groups_power(i, sd);
7659 for_each_cpu_mask_nr(i, *cpu_map) {
7660 struct sched_domain *sd = &per_cpu(phys_domains, i);
7662 init_sched_groups_power(i, sd);
7666 for (i = 0; i < nr_node_ids; i++)
7667 init_numa_sched_groups_power(sched_group_nodes[i]);
7670 struct sched_group *sg;
7672 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7674 init_numa_sched_groups_power(sg);
7678 /* Attach the domains */
7679 for_each_cpu_mask_nr(i, *cpu_map) {
7680 struct sched_domain *sd;
7681 #ifdef CONFIG_SCHED_SMT
7682 sd = &per_cpu(cpu_domains, i);
7683 #elif defined(CONFIG_SCHED_MC)
7684 sd = &per_cpu(core_domains, i);
7686 sd = &per_cpu(phys_domains, i);
7688 cpu_attach_domain(sd, rd, i);
7691 SCHED_CPUMASK_FREE((void *)allmasks);
7696 free_sched_groups(cpu_map, tmpmask);
7697 SCHED_CPUMASK_FREE((void *)allmasks);
7703 static int build_sched_domains(const cpumask_t *cpu_map)
7705 return __build_sched_domains(cpu_map, NULL);
7708 static cpumask_t *doms_cur; /* current sched domains */
7709 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7710 static struct sched_domain_attr *dattr_cur;
7711 /* attribues of custom domains in 'doms_cur' */
7714 * Special case: If a kmalloc of a doms_cur partition (array of
7715 * cpumask_t) fails, then fallback to a single sched domain,
7716 * as determined by the single cpumask_t fallback_doms.
7718 static cpumask_t fallback_doms;
7720 void __attribute__((weak)) arch_update_cpu_topology(void)
7725 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7726 * For now this just excludes isolated cpus, but could be used to
7727 * exclude other special cases in the future.
7729 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7733 arch_update_cpu_topology();
7735 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7737 doms_cur = &fallback_doms;
7738 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7740 err = build_sched_domains(doms_cur);
7741 register_sched_domain_sysctl();
7746 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7749 free_sched_groups(cpu_map, tmpmask);
7753 * Detach sched domains from a group of cpus specified in cpu_map
7754 * These cpus will now be attached to the NULL domain
7756 static void detach_destroy_domains(const cpumask_t *cpu_map)
7761 unregister_sched_domain_sysctl();
7763 for_each_cpu_mask_nr(i, *cpu_map)
7764 cpu_attach_domain(NULL, &def_root_domain, i);
7765 synchronize_sched();
7766 arch_destroy_sched_domains(cpu_map, &tmpmask);
7769 /* handle null as "default" */
7770 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7771 struct sched_domain_attr *new, int idx_new)
7773 struct sched_domain_attr tmp;
7780 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7781 new ? (new + idx_new) : &tmp,
7782 sizeof(struct sched_domain_attr));
7786 * Partition sched domains as specified by the 'ndoms_new'
7787 * cpumasks in the array doms_new[] of cpumasks. This compares
7788 * doms_new[] to the current sched domain partitioning, doms_cur[].
7789 * It destroys each deleted domain and builds each new domain.
7791 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7792 * The masks don't intersect (don't overlap.) We should setup one
7793 * sched domain for each mask. CPUs not in any of the cpumasks will
7794 * not be load balanced. If the same cpumask appears both in the
7795 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7798 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7799 * ownership of it and will kfree it when done with it. If the caller
7800 * failed the kmalloc call, then it can pass in doms_new == NULL &&
7801 * ndoms_new == 1, and partition_sched_domains() will fallback to
7802 * the single partition 'fallback_doms', it also forces the domains
7805 * If doms_new == NULL it will be replaced with cpu_online_map.
7806 * ndoms_new == 0 is a special case for destroying existing domains,
7807 * and it will not create the default domain.
7809 * Call with hotplug lock held
7811 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7812 struct sched_domain_attr *dattr_new)
7816 mutex_lock(&sched_domains_mutex);
7818 /* always unregister in case we don't destroy any domains */
7819 unregister_sched_domain_sysctl();
7821 n = doms_new ? ndoms_new : 0;
7823 /* Destroy deleted domains */
7824 for (i = 0; i < ndoms_cur; i++) {
7825 for (j = 0; j < n; j++) {
7826 if (cpus_equal(doms_cur[i], doms_new[j])
7827 && dattrs_equal(dattr_cur, i, dattr_new, j))
7830 /* no match - a current sched domain not in new doms_new[] */
7831 detach_destroy_domains(doms_cur + i);
7836 if (doms_new == NULL) {
7838 doms_new = &fallback_doms;
7839 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7843 /* Build new domains */
7844 for (i = 0; i < ndoms_new; i++) {
7845 for (j = 0; j < ndoms_cur; j++) {
7846 if (cpus_equal(doms_new[i], doms_cur[j])
7847 && dattrs_equal(dattr_new, i, dattr_cur, j))
7850 /* no match - add a new doms_new */
7851 __build_sched_domains(doms_new + i,
7852 dattr_new ? dattr_new + i : NULL);
7857 /* Remember the new sched domains */
7858 if (doms_cur != &fallback_doms)
7860 kfree(dattr_cur); /* kfree(NULL) is safe */
7861 doms_cur = doms_new;
7862 dattr_cur = dattr_new;
7863 ndoms_cur = ndoms_new;
7865 register_sched_domain_sysctl();
7867 mutex_unlock(&sched_domains_mutex);
7870 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7871 int arch_reinit_sched_domains(void)
7875 /* Destroy domains first to force the rebuild */
7876 partition_sched_domains(0, NULL, NULL);
7878 rebuild_sched_domains();
7884 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7888 if (buf[0] != '0' && buf[0] != '1')
7892 sched_smt_power_savings = (buf[0] == '1');
7894 sched_mc_power_savings = (buf[0] == '1');
7896 ret = arch_reinit_sched_domains();
7898 return ret ? ret : count;
7901 #ifdef CONFIG_SCHED_MC
7902 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7905 return sprintf(page, "%u\n", sched_mc_power_savings);
7907 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7908 const char *buf, size_t count)
7910 return sched_power_savings_store(buf, count, 0);
7912 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7913 sched_mc_power_savings_show,
7914 sched_mc_power_savings_store);
7917 #ifdef CONFIG_SCHED_SMT
7918 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7921 return sprintf(page, "%u\n", sched_smt_power_savings);
7923 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7924 const char *buf, size_t count)
7926 return sched_power_savings_store(buf, count, 1);
7928 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7929 sched_smt_power_savings_show,
7930 sched_smt_power_savings_store);
7933 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7937 #ifdef CONFIG_SCHED_SMT
7939 err = sysfs_create_file(&cls->kset.kobj,
7940 &attr_sched_smt_power_savings.attr);
7942 #ifdef CONFIG_SCHED_MC
7943 if (!err && mc_capable())
7944 err = sysfs_create_file(&cls->kset.kobj,
7945 &attr_sched_mc_power_savings.attr);
7949 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7951 #ifndef CONFIG_CPUSETS
7953 * Add online and remove offline CPUs from the scheduler domains.
7954 * When cpusets are enabled they take over this function.
7956 static int update_sched_domains(struct notifier_block *nfb,
7957 unsigned long action, void *hcpu)
7961 case CPU_ONLINE_FROZEN:
7963 case CPU_DEAD_FROZEN:
7964 partition_sched_domains(1, NULL, NULL);
7973 static int update_runtime(struct notifier_block *nfb,
7974 unsigned long action, void *hcpu)
7976 int cpu = (int)(long)hcpu;
7979 case CPU_DOWN_PREPARE:
7980 case CPU_DOWN_PREPARE_FROZEN:
7981 disable_runtime(cpu_rq(cpu));
7984 case CPU_DOWN_FAILED:
7985 case CPU_DOWN_FAILED_FROZEN:
7987 case CPU_ONLINE_FROZEN:
7988 enable_runtime(cpu_rq(cpu));
7996 void __init sched_init_smp(void)
7998 cpumask_t non_isolated_cpus;
8000 #if defined(CONFIG_NUMA)
8001 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
8003 BUG_ON(sched_group_nodes_bycpu == NULL);
8006 mutex_lock(&sched_domains_mutex);
8007 arch_init_sched_domains(&cpu_online_map);
8008 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
8009 if (cpus_empty(non_isolated_cpus))
8010 cpu_set(smp_processor_id(), non_isolated_cpus);
8011 mutex_unlock(&sched_domains_mutex);
8014 #ifndef CONFIG_CPUSETS
8015 /* XXX: Theoretical race here - CPU may be hotplugged now */
8016 hotcpu_notifier(update_sched_domains, 0);
8019 /* RT runtime code needs to handle some hotplug events */
8020 hotcpu_notifier(update_runtime, 0);
8024 /* Move init over to a non-isolated CPU */
8025 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
8027 sched_init_granularity();
8030 void __init sched_init_smp(void)
8032 sched_init_granularity();
8034 #endif /* CONFIG_SMP */
8036 int in_sched_functions(unsigned long addr)
8038 return in_lock_functions(addr) ||
8039 (addr >= (unsigned long)__sched_text_start
8040 && addr < (unsigned long)__sched_text_end);
8043 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
8045 cfs_rq->tasks_timeline = RB_ROOT;
8046 INIT_LIST_HEAD(&cfs_rq->tasks);
8047 #ifdef CONFIG_FAIR_GROUP_SCHED
8050 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8053 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8055 struct rt_prio_array *array;
8058 array = &rt_rq->active;
8059 for (i = 0; i < MAX_RT_PRIO; i++) {
8060 INIT_LIST_HEAD(array->queue + i);
8061 __clear_bit(i, array->bitmap);
8063 /* delimiter for bitsearch: */
8064 __set_bit(MAX_RT_PRIO, array->bitmap);
8066 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8067 rt_rq->highest_prio = MAX_RT_PRIO;
8070 rt_rq->rt_nr_migratory = 0;
8071 rt_rq->overloaded = 0;
8075 rt_rq->rt_throttled = 0;
8076 rt_rq->rt_runtime = 0;
8077 spin_lock_init(&rt_rq->rt_runtime_lock);
8079 #ifdef CONFIG_RT_GROUP_SCHED
8080 rt_rq->rt_nr_boosted = 0;
8085 #ifdef CONFIG_FAIR_GROUP_SCHED
8086 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8087 struct sched_entity *se, int cpu, int add,
8088 struct sched_entity *parent)
8090 struct rq *rq = cpu_rq(cpu);
8091 tg->cfs_rq[cpu] = cfs_rq;
8092 init_cfs_rq(cfs_rq, rq);
8095 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8098 /* se could be NULL for init_task_group */
8103 se->cfs_rq = &rq->cfs;
8105 se->cfs_rq = parent->my_q;
8108 se->load.weight = tg->shares;
8109 se->load.inv_weight = 0;
8110 se->parent = parent;
8114 #ifdef CONFIG_RT_GROUP_SCHED
8115 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8116 struct sched_rt_entity *rt_se, int cpu, int add,
8117 struct sched_rt_entity *parent)
8119 struct rq *rq = cpu_rq(cpu);
8121 tg->rt_rq[cpu] = rt_rq;
8122 init_rt_rq(rt_rq, rq);
8124 rt_rq->rt_se = rt_se;
8125 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8127 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8129 tg->rt_se[cpu] = rt_se;
8134 rt_se->rt_rq = &rq->rt;
8136 rt_se->rt_rq = parent->my_q;
8138 rt_se->my_q = rt_rq;
8139 rt_se->parent = parent;
8140 INIT_LIST_HEAD(&rt_se->run_list);
8144 void __init sched_init(void)
8147 unsigned long alloc_size = 0, ptr;
8149 #ifdef CONFIG_FAIR_GROUP_SCHED
8150 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8152 #ifdef CONFIG_RT_GROUP_SCHED
8153 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8155 #ifdef CONFIG_USER_SCHED
8159 * As sched_init() is called before page_alloc is setup,
8160 * we use alloc_bootmem().
8163 ptr = (unsigned long)alloc_bootmem(alloc_size);
8165 #ifdef CONFIG_FAIR_GROUP_SCHED
8166 init_task_group.se = (struct sched_entity **)ptr;
8167 ptr += nr_cpu_ids * sizeof(void **);
8169 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8170 ptr += nr_cpu_ids * sizeof(void **);
8172 #ifdef CONFIG_USER_SCHED
8173 root_task_group.se = (struct sched_entity **)ptr;
8174 ptr += nr_cpu_ids * sizeof(void **);
8176 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8177 ptr += nr_cpu_ids * sizeof(void **);
8178 #endif /* CONFIG_USER_SCHED */
8179 #endif /* CONFIG_FAIR_GROUP_SCHED */
8180 #ifdef CONFIG_RT_GROUP_SCHED
8181 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8182 ptr += nr_cpu_ids * sizeof(void **);
8184 init_task_group.rt_rq = (struct rt_rq **)ptr;
8185 ptr += nr_cpu_ids * sizeof(void **);
8187 #ifdef CONFIG_USER_SCHED
8188 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8189 ptr += nr_cpu_ids * sizeof(void **);
8191 root_task_group.rt_rq = (struct rt_rq **)ptr;
8192 ptr += nr_cpu_ids * sizeof(void **);
8193 #endif /* CONFIG_USER_SCHED */
8194 #endif /* CONFIG_RT_GROUP_SCHED */
8198 init_defrootdomain();
8201 init_rt_bandwidth(&def_rt_bandwidth,
8202 global_rt_period(), global_rt_runtime());
8204 #ifdef CONFIG_RT_GROUP_SCHED
8205 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8206 global_rt_period(), global_rt_runtime());
8207 #ifdef CONFIG_USER_SCHED
8208 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8209 global_rt_period(), RUNTIME_INF);
8210 #endif /* CONFIG_USER_SCHED */
8211 #endif /* CONFIG_RT_GROUP_SCHED */
8213 #ifdef CONFIG_GROUP_SCHED
8214 list_add(&init_task_group.list, &task_groups);
8215 INIT_LIST_HEAD(&init_task_group.children);
8217 #ifdef CONFIG_USER_SCHED
8218 INIT_LIST_HEAD(&root_task_group.children);
8219 init_task_group.parent = &root_task_group;
8220 list_add(&init_task_group.siblings, &root_task_group.children);
8221 #endif /* CONFIG_USER_SCHED */
8222 #endif /* CONFIG_GROUP_SCHED */
8224 for_each_possible_cpu(i) {
8228 spin_lock_init(&rq->lock);
8230 init_cfs_rq(&rq->cfs, rq);
8231 init_rt_rq(&rq->rt, rq);
8232 #ifdef CONFIG_FAIR_GROUP_SCHED
8233 init_task_group.shares = init_task_group_load;
8234 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8235 #ifdef CONFIG_CGROUP_SCHED
8237 * How much cpu bandwidth does init_task_group get?
8239 * In case of task-groups formed thr' the cgroup filesystem, it
8240 * gets 100% of the cpu resources in the system. This overall
8241 * system cpu resource is divided among the tasks of
8242 * init_task_group and its child task-groups in a fair manner,
8243 * based on each entity's (task or task-group's) weight
8244 * (se->load.weight).
8246 * In other words, if init_task_group has 10 tasks of weight
8247 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8248 * then A0's share of the cpu resource is:
8250 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8252 * We achieve this by letting init_task_group's tasks sit
8253 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8255 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8256 #elif defined CONFIG_USER_SCHED
8257 root_task_group.shares = NICE_0_LOAD;
8258 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8260 * In case of task-groups formed thr' the user id of tasks,
8261 * init_task_group represents tasks belonging to root user.
8262 * Hence it forms a sibling of all subsequent groups formed.
8263 * In this case, init_task_group gets only a fraction of overall
8264 * system cpu resource, based on the weight assigned to root
8265 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8266 * by letting tasks of init_task_group sit in a separate cfs_rq
8267 * (init_cfs_rq) and having one entity represent this group of
8268 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8270 init_tg_cfs_entry(&init_task_group,
8271 &per_cpu(init_cfs_rq, i),
8272 &per_cpu(init_sched_entity, i), i, 1,
8273 root_task_group.se[i]);
8276 #endif /* CONFIG_FAIR_GROUP_SCHED */
8278 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8279 #ifdef CONFIG_RT_GROUP_SCHED
8280 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8281 #ifdef CONFIG_CGROUP_SCHED
8282 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8283 #elif defined CONFIG_USER_SCHED
8284 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8285 init_tg_rt_entry(&init_task_group,
8286 &per_cpu(init_rt_rq, i),
8287 &per_cpu(init_sched_rt_entity, i), i, 1,
8288 root_task_group.rt_se[i]);
8292 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8293 rq->cpu_load[j] = 0;
8297 rq->active_balance = 0;
8298 rq->next_balance = jiffies;
8302 rq->migration_thread = NULL;
8303 INIT_LIST_HEAD(&rq->migration_queue);
8304 rq_attach_root(rq, &def_root_domain);
8307 atomic_set(&rq->nr_iowait, 0);
8310 set_load_weight(&init_task);
8312 #ifdef CONFIG_PREEMPT_NOTIFIERS
8313 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8317 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8320 #ifdef CONFIG_RT_MUTEXES
8321 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8325 * The boot idle thread does lazy MMU switching as well:
8327 atomic_inc(&init_mm.mm_count);
8328 enter_lazy_tlb(&init_mm, current);
8331 * Make us the idle thread. Technically, schedule() should not be
8332 * called from this thread, however somewhere below it might be,
8333 * but because we are the idle thread, we just pick up running again
8334 * when this runqueue becomes "idle".
8336 init_idle(current, smp_processor_id());
8338 * During early bootup we pretend to be a normal task:
8340 current->sched_class = &fair_sched_class;
8342 scheduler_running = 1;
8345 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8346 void __might_sleep(char *file, int line)
8349 static unsigned long prev_jiffy; /* ratelimiting */
8351 if ((!in_atomic() && !irqs_disabled()) ||
8352 system_state != SYSTEM_RUNNING || oops_in_progress)
8354 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8356 prev_jiffy = jiffies;
8359 "BUG: sleeping function called from invalid context at %s:%d\n",
8362 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8363 in_atomic(), irqs_disabled(),
8364 current->pid, current->comm);
8366 debug_show_held_locks(current);
8367 if (irqs_disabled())
8368 print_irqtrace_events(current);
8372 EXPORT_SYMBOL(__might_sleep);
8375 #ifdef CONFIG_MAGIC_SYSRQ
8376 static void normalize_task(struct rq *rq, struct task_struct *p)
8380 update_rq_clock(rq);
8381 on_rq = p->se.on_rq;
8383 deactivate_task(rq, p, 0);
8384 __setscheduler(rq, p, SCHED_NORMAL, 0);
8386 activate_task(rq, p, 0);
8387 resched_task(rq->curr);
8391 void normalize_rt_tasks(void)
8393 struct task_struct *g, *p;
8394 unsigned long flags;
8397 read_lock_irqsave(&tasklist_lock, flags);
8398 do_each_thread(g, p) {
8400 * Only normalize user tasks:
8405 p->se.exec_start = 0;
8406 #ifdef CONFIG_SCHEDSTATS
8407 p->se.wait_start = 0;
8408 p->se.sleep_start = 0;
8409 p->se.block_start = 0;
8414 * Renice negative nice level userspace
8417 if (TASK_NICE(p) < 0 && p->mm)
8418 set_user_nice(p, 0);
8422 spin_lock(&p->pi_lock);
8423 rq = __task_rq_lock(p);
8425 normalize_task(rq, p);
8427 __task_rq_unlock(rq);
8428 spin_unlock(&p->pi_lock);
8429 } while_each_thread(g, p);
8431 read_unlock_irqrestore(&tasklist_lock, flags);
8434 #endif /* CONFIG_MAGIC_SYSRQ */
8438 * These functions are only useful for the IA64 MCA handling.
8440 * They can only be called when the whole system has been
8441 * stopped - every CPU needs to be quiescent, and no scheduling
8442 * activity can take place. Using them for anything else would
8443 * be a serious bug, and as a result, they aren't even visible
8444 * under any other configuration.
8448 * curr_task - return the current task for a given cpu.
8449 * @cpu: the processor in question.
8451 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8453 struct task_struct *curr_task(int cpu)
8455 return cpu_curr(cpu);
8459 * set_curr_task - set the current task for a given cpu.
8460 * @cpu: the processor in question.
8461 * @p: the task pointer to set.
8463 * Description: This function must only be used when non-maskable interrupts
8464 * are serviced on a separate stack. It allows the architecture to switch the
8465 * notion of the current task on a cpu in a non-blocking manner. This function
8466 * must be called with all CPU's synchronized, and interrupts disabled, the
8467 * and caller must save the original value of the current task (see
8468 * curr_task() above) and restore that value before reenabling interrupts and
8469 * re-starting the system.
8471 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8473 void set_curr_task(int cpu, struct task_struct *p)
8480 #ifdef CONFIG_FAIR_GROUP_SCHED
8481 static void free_fair_sched_group(struct task_group *tg)
8485 for_each_possible_cpu(i) {
8487 kfree(tg->cfs_rq[i]);
8497 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8499 struct cfs_rq *cfs_rq;
8500 struct sched_entity *se, *parent_se;
8504 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8507 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8511 tg->shares = NICE_0_LOAD;
8513 for_each_possible_cpu(i) {
8516 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8517 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8521 se = kmalloc_node(sizeof(struct sched_entity),
8522 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8526 parent_se = parent ? parent->se[i] : NULL;
8527 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
8536 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8538 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8539 &cpu_rq(cpu)->leaf_cfs_rq_list);
8542 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8544 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8546 #else /* !CONFG_FAIR_GROUP_SCHED */
8547 static inline void free_fair_sched_group(struct task_group *tg)
8552 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8557 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8561 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8564 #endif /* CONFIG_FAIR_GROUP_SCHED */
8566 #ifdef CONFIG_RT_GROUP_SCHED
8567 static void free_rt_sched_group(struct task_group *tg)
8571 destroy_rt_bandwidth(&tg->rt_bandwidth);
8573 for_each_possible_cpu(i) {
8575 kfree(tg->rt_rq[i]);
8577 kfree(tg->rt_se[i]);
8585 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8587 struct rt_rq *rt_rq;
8588 struct sched_rt_entity *rt_se, *parent_se;
8592 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8595 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8599 init_rt_bandwidth(&tg->rt_bandwidth,
8600 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8602 for_each_possible_cpu(i) {
8605 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8606 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8610 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8611 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8615 parent_se = parent ? parent->rt_se[i] : NULL;
8616 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
8625 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8627 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8628 &cpu_rq(cpu)->leaf_rt_rq_list);
8631 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8633 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8635 #else /* !CONFIG_RT_GROUP_SCHED */
8636 static inline void free_rt_sched_group(struct task_group *tg)
8641 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8646 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8650 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8653 #endif /* CONFIG_RT_GROUP_SCHED */
8655 #ifdef CONFIG_GROUP_SCHED
8656 static void free_sched_group(struct task_group *tg)
8658 free_fair_sched_group(tg);
8659 free_rt_sched_group(tg);
8663 /* allocate runqueue etc for a new task group */
8664 struct task_group *sched_create_group(struct task_group *parent)
8666 struct task_group *tg;
8667 unsigned long flags;
8670 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8672 return ERR_PTR(-ENOMEM);
8674 if (!alloc_fair_sched_group(tg, parent))
8677 if (!alloc_rt_sched_group(tg, parent))
8680 spin_lock_irqsave(&task_group_lock, flags);
8681 for_each_possible_cpu(i) {
8682 register_fair_sched_group(tg, i);
8683 register_rt_sched_group(tg, i);
8685 list_add_rcu(&tg->list, &task_groups);
8687 WARN_ON(!parent); /* root should already exist */
8689 tg->parent = parent;
8690 INIT_LIST_HEAD(&tg->children);
8691 list_add_rcu(&tg->siblings, &parent->children);
8692 spin_unlock_irqrestore(&task_group_lock, flags);
8697 free_sched_group(tg);
8698 return ERR_PTR(-ENOMEM);
8701 /* rcu callback to free various structures associated with a task group */
8702 static void free_sched_group_rcu(struct rcu_head *rhp)
8704 /* now it should be safe to free those cfs_rqs */
8705 free_sched_group(container_of(rhp, struct task_group, rcu));
8708 /* Destroy runqueue etc associated with a task group */
8709 void sched_destroy_group(struct task_group *tg)
8711 unsigned long flags;
8714 spin_lock_irqsave(&task_group_lock, flags);
8715 for_each_possible_cpu(i) {
8716 unregister_fair_sched_group(tg, i);
8717 unregister_rt_sched_group(tg, i);
8719 list_del_rcu(&tg->list);
8720 list_del_rcu(&tg->siblings);
8721 spin_unlock_irqrestore(&task_group_lock, flags);
8723 /* wait for possible concurrent references to cfs_rqs complete */
8724 call_rcu(&tg->rcu, free_sched_group_rcu);
8727 /* change task's runqueue when it moves between groups.
8728 * The caller of this function should have put the task in its new group
8729 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8730 * reflect its new group.
8732 void sched_move_task(struct task_struct *tsk)
8735 unsigned long flags;
8738 rq = task_rq_lock(tsk, &flags);
8740 update_rq_clock(rq);
8742 running = task_current(rq, tsk);
8743 on_rq = tsk->se.on_rq;
8746 dequeue_task(rq, tsk, 0);
8747 if (unlikely(running))
8748 tsk->sched_class->put_prev_task(rq, tsk);
8750 set_task_rq(tsk, task_cpu(tsk));
8752 #ifdef CONFIG_FAIR_GROUP_SCHED
8753 if (tsk->sched_class->moved_group)
8754 tsk->sched_class->moved_group(tsk);
8757 if (unlikely(running))
8758 tsk->sched_class->set_curr_task(rq);
8760 enqueue_task(rq, tsk, 0);
8762 task_rq_unlock(rq, &flags);
8764 #endif /* CONFIG_GROUP_SCHED */
8766 #ifdef CONFIG_FAIR_GROUP_SCHED
8767 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8769 struct cfs_rq *cfs_rq = se->cfs_rq;
8774 dequeue_entity(cfs_rq, se, 0);
8776 se->load.weight = shares;
8777 se->load.inv_weight = 0;
8780 enqueue_entity(cfs_rq, se, 0);
8783 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8785 struct cfs_rq *cfs_rq = se->cfs_rq;
8786 struct rq *rq = cfs_rq->rq;
8787 unsigned long flags;
8789 spin_lock_irqsave(&rq->lock, flags);
8790 __set_se_shares(se, shares);
8791 spin_unlock_irqrestore(&rq->lock, flags);
8794 static DEFINE_MUTEX(shares_mutex);
8796 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8799 unsigned long flags;
8802 * We can't change the weight of the root cgroup.
8807 if (shares < MIN_SHARES)
8808 shares = MIN_SHARES;
8809 else if (shares > MAX_SHARES)
8810 shares = MAX_SHARES;
8812 mutex_lock(&shares_mutex);
8813 if (tg->shares == shares)
8816 spin_lock_irqsave(&task_group_lock, flags);
8817 for_each_possible_cpu(i)
8818 unregister_fair_sched_group(tg, i);
8819 list_del_rcu(&tg->siblings);
8820 spin_unlock_irqrestore(&task_group_lock, flags);
8822 /* wait for any ongoing reference to this group to finish */
8823 synchronize_sched();
8826 * Now we are free to modify the group's share on each cpu
8827 * w/o tripping rebalance_share or load_balance_fair.
8829 tg->shares = shares;
8830 for_each_possible_cpu(i) {
8834 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8835 set_se_shares(tg->se[i], shares);
8839 * Enable load balance activity on this group, by inserting it back on
8840 * each cpu's rq->leaf_cfs_rq_list.
8842 spin_lock_irqsave(&task_group_lock, flags);
8843 for_each_possible_cpu(i)
8844 register_fair_sched_group(tg, i);
8845 list_add_rcu(&tg->siblings, &tg->parent->children);
8846 spin_unlock_irqrestore(&task_group_lock, flags);
8848 mutex_unlock(&shares_mutex);
8852 unsigned long sched_group_shares(struct task_group *tg)
8858 #ifdef CONFIG_RT_GROUP_SCHED
8860 * Ensure that the real time constraints are schedulable.
8862 static DEFINE_MUTEX(rt_constraints_mutex);
8864 static unsigned long to_ratio(u64 period, u64 runtime)
8866 if (runtime == RUNTIME_INF)
8869 return div64_u64(runtime << 20, period);
8872 /* Must be called with tasklist_lock held */
8873 static inline int tg_has_rt_tasks(struct task_group *tg)
8875 struct task_struct *g, *p;
8877 do_each_thread(g, p) {
8878 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8880 } while_each_thread(g, p);
8885 struct rt_schedulable_data {
8886 struct task_group *tg;
8891 static int tg_schedulable(struct task_group *tg, void *data)
8893 struct rt_schedulable_data *d = data;
8894 struct task_group *child;
8895 unsigned long total, sum = 0;
8896 u64 period, runtime;
8898 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8899 runtime = tg->rt_bandwidth.rt_runtime;
8902 period = d->rt_period;
8903 runtime = d->rt_runtime;
8907 * Cannot have more runtime than the period.
8909 if (runtime > period && runtime != RUNTIME_INF)
8913 * Ensure we don't starve existing RT tasks.
8915 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8918 total = to_ratio(period, runtime);
8921 * Nobody can have more than the global setting allows.
8923 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8927 * The sum of our children's runtime should not exceed our own.
8929 list_for_each_entry_rcu(child, &tg->children, siblings) {
8930 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8931 runtime = child->rt_bandwidth.rt_runtime;
8933 if (child == d->tg) {
8934 period = d->rt_period;
8935 runtime = d->rt_runtime;
8938 sum += to_ratio(period, runtime);
8947 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8949 struct rt_schedulable_data data = {
8951 .rt_period = period,
8952 .rt_runtime = runtime,
8955 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8958 static int tg_set_bandwidth(struct task_group *tg,
8959 u64 rt_period, u64 rt_runtime)
8963 mutex_lock(&rt_constraints_mutex);
8964 read_lock(&tasklist_lock);
8965 err = __rt_schedulable(tg, rt_period, rt_runtime);
8969 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8970 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8971 tg->rt_bandwidth.rt_runtime = rt_runtime;
8973 for_each_possible_cpu(i) {
8974 struct rt_rq *rt_rq = tg->rt_rq[i];
8976 spin_lock(&rt_rq->rt_runtime_lock);
8977 rt_rq->rt_runtime = rt_runtime;
8978 spin_unlock(&rt_rq->rt_runtime_lock);
8980 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8982 read_unlock(&tasklist_lock);
8983 mutex_unlock(&rt_constraints_mutex);
8988 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8990 u64 rt_runtime, rt_period;
8992 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8993 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8994 if (rt_runtime_us < 0)
8995 rt_runtime = RUNTIME_INF;
8997 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9000 long sched_group_rt_runtime(struct task_group *tg)
9004 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
9007 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
9008 do_div(rt_runtime_us, NSEC_PER_USEC);
9009 return rt_runtime_us;
9012 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
9014 u64 rt_runtime, rt_period;
9016 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
9017 rt_runtime = tg->rt_bandwidth.rt_runtime;
9022 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9025 long sched_group_rt_period(struct task_group *tg)
9029 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9030 do_div(rt_period_us, NSEC_PER_USEC);
9031 return rt_period_us;
9034 static int sched_rt_global_constraints(void)
9036 u64 runtime, period;
9039 if (sysctl_sched_rt_period <= 0)
9042 runtime = global_rt_runtime();
9043 period = global_rt_period();
9046 * Sanity check on the sysctl variables.
9048 if (runtime > period && runtime != RUNTIME_INF)
9051 mutex_lock(&rt_constraints_mutex);
9052 read_lock(&tasklist_lock);
9053 ret = __rt_schedulable(NULL, 0, 0);
9054 read_unlock(&tasklist_lock);
9055 mutex_unlock(&rt_constraints_mutex);
9059 #else /* !CONFIG_RT_GROUP_SCHED */
9060 static int sched_rt_global_constraints(void)
9062 unsigned long flags;
9065 if (sysctl_sched_rt_period <= 0)
9068 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9069 for_each_possible_cpu(i) {
9070 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9072 spin_lock(&rt_rq->rt_runtime_lock);
9073 rt_rq->rt_runtime = global_rt_runtime();
9074 spin_unlock(&rt_rq->rt_runtime_lock);
9076 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9080 #endif /* CONFIG_RT_GROUP_SCHED */
9082 int sched_rt_handler(struct ctl_table *table, int write,
9083 struct file *filp, void __user *buffer, size_t *lenp,
9087 int old_period, old_runtime;
9088 static DEFINE_MUTEX(mutex);
9091 old_period = sysctl_sched_rt_period;
9092 old_runtime = sysctl_sched_rt_runtime;
9094 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
9096 if (!ret && write) {
9097 ret = sched_rt_global_constraints();
9099 sysctl_sched_rt_period = old_period;
9100 sysctl_sched_rt_runtime = old_runtime;
9102 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9103 def_rt_bandwidth.rt_period =
9104 ns_to_ktime(global_rt_period());
9107 mutex_unlock(&mutex);
9112 #ifdef CONFIG_CGROUP_SCHED
9114 /* return corresponding task_group object of a cgroup */
9115 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9117 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9118 struct task_group, css);
9121 static struct cgroup_subsys_state *
9122 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9124 struct task_group *tg, *parent;
9126 if (!cgrp->parent) {
9127 /* This is early initialization for the top cgroup */
9128 return &init_task_group.css;
9131 parent = cgroup_tg(cgrp->parent);
9132 tg = sched_create_group(parent);
9134 return ERR_PTR(-ENOMEM);
9140 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9142 struct task_group *tg = cgroup_tg(cgrp);
9144 sched_destroy_group(tg);
9148 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9149 struct task_struct *tsk)
9151 #ifdef CONFIG_RT_GROUP_SCHED
9152 /* Don't accept realtime tasks when there is no way for them to run */
9153 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
9156 /* We don't support RT-tasks being in separate groups */
9157 if (tsk->sched_class != &fair_sched_class)
9165 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9166 struct cgroup *old_cont, struct task_struct *tsk)
9168 sched_move_task(tsk);
9171 #ifdef CONFIG_FAIR_GROUP_SCHED
9172 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9175 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9178 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9180 struct task_group *tg = cgroup_tg(cgrp);
9182 return (u64) tg->shares;
9184 #endif /* CONFIG_FAIR_GROUP_SCHED */
9186 #ifdef CONFIG_RT_GROUP_SCHED
9187 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9190 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9193 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9195 return sched_group_rt_runtime(cgroup_tg(cgrp));
9198 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9201 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9204 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9206 return sched_group_rt_period(cgroup_tg(cgrp));
9208 #endif /* CONFIG_RT_GROUP_SCHED */
9210 static struct cftype cpu_files[] = {
9211 #ifdef CONFIG_FAIR_GROUP_SCHED
9214 .read_u64 = cpu_shares_read_u64,
9215 .write_u64 = cpu_shares_write_u64,
9218 #ifdef CONFIG_RT_GROUP_SCHED
9220 .name = "rt_runtime_us",
9221 .read_s64 = cpu_rt_runtime_read,
9222 .write_s64 = cpu_rt_runtime_write,
9225 .name = "rt_period_us",
9226 .read_u64 = cpu_rt_period_read_uint,
9227 .write_u64 = cpu_rt_period_write_uint,
9232 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9234 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9237 struct cgroup_subsys cpu_cgroup_subsys = {
9239 .create = cpu_cgroup_create,
9240 .destroy = cpu_cgroup_destroy,
9241 .can_attach = cpu_cgroup_can_attach,
9242 .attach = cpu_cgroup_attach,
9243 .populate = cpu_cgroup_populate,
9244 .subsys_id = cpu_cgroup_subsys_id,
9248 #endif /* CONFIG_CGROUP_SCHED */
9250 #ifdef CONFIG_CGROUP_CPUACCT
9253 * CPU accounting code for task groups.
9255 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9256 * (balbir@in.ibm.com).
9259 /* track cpu usage of a group of tasks */
9261 struct cgroup_subsys_state css;
9262 /* cpuusage holds pointer to a u64-type object on every cpu */
9266 struct cgroup_subsys cpuacct_subsys;
9268 /* return cpu accounting group corresponding to this container */
9269 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9271 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9272 struct cpuacct, css);
9275 /* return cpu accounting group to which this task belongs */
9276 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9278 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9279 struct cpuacct, css);
9282 /* create a new cpu accounting group */
9283 static struct cgroup_subsys_state *cpuacct_create(
9284 struct cgroup_subsys *ss, struct cgroup *cgrp)
9286 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9289 return ERR_PTR(-ENOMEM);
9291 ca->cpuusage = alloc_percpu(u64);
9292 if (!ca->cpuusage) {
9294 return ERR_PTR(-ENOMEM);
9300 /* destroy an existing cpu accounting group */
9302 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9304 struct cpuacct *ca = cgroup_ca(cgrp);
9306 free_percpu(ca->cpuusage);
9310 /* return total cpu usage (in nanoseconds) of a group */
9311 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9313 struct cpuacct *ca = cgroup_ca(cgrp);
9314 u64 totalcpuusage = 0;
9317 for_each_possible_cpu(i) {
9318 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9321 * Take rq->lock to make 64-bit addition safe on 32-bit
9324 spin_lock_irq(&cpu_rq(i)->lock);
9325 totalcpuusage += *cpuusage;
9326 spin_unlock_irq(&cpu_rq(i)->lock);
9329 return totalcpuusage;
9332 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9335 struct cpuacct *ca = cgroup_ca(cgrp);
9344 for_each_possible_cpu(i) {
9345 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9347 spin_lock_irq(&cpu_rq(i)->lock);
9349 spin_unlock_irq(&cpu_rq(i)->lock);
9355 static struct cftype files[] = {
9358 .read_u64 = cpuusage_read,
9359 .write_u64 = cpuusage_write,
9363 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9365 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9369 * charge this task's execution time to its accounting group.
9371 * called with rq->lock held.
9373 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9377 if (!cpuacct_subsys.active)
9382 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
9384 *cpuusage += cputime;
9388 struct cgroup_subsys cpuacct_subsys = {
9390 .create = cpuacct_create,
9391 .destroy = cpuacct_destroy,
9392 .populate = cpuacct_populate,
9393 .subsys_id = cpuacct_subsys_id,
9395 #endif /* CONFIG_CGROUP_CPUACCT */