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 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
132 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
133 * Since cpu_power is a 'constant', we can use a reciprocal divide.
135 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
137 return reciprocal_divide(load, sg->reciprocal_cpu_power);
141 * Each time a sched group cpu_power is changed,
142 * we must compute its reciprocal value
144 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
146 sg->__cpu_power += val;
147 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
151 static inline int rt_policy(int policy)
153 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
158 static inline int task_has_rt_policy(struct task_struct *p)
160 return rt_policy(p->policy);
164 * This is the priority-queue data structure of the RT scheduling class:
166 struct rt_prio_array {
167 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
168 struct list_head queue[MAX_RT_PRIO];
171 struct rt_bandwidth {
172 /* nests inside the rq lock: */
173 spinlock_t rt_runtime_lock;
176 struct hrtimer rt_period_timer;
179 static struct rt_bandwidth def_rt_bandwidth;
181 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
183 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
185 struct rt_bandwidth *rt_b =
186 container_of(timer, struct rt_bandwidth, rt_period_timer);
192 now = hrtimer_cb_get_time(timer);
193 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
198 idle = do_sched_rt_period_timer(rt_b, overrun);
201 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
205 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
207 rt_b->rt_period = ns_to_ktime(period);
208 rt_b->rt_runtime = runtime;
210 spin_lock_init(&rt_b->rt_runtime_lock);
212 hrtimer_init(&rt_b->rt_period_timer,
213 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
214 rt_b->rt_period_timer.function = sched_rt_period_timer;
217 static inline int rt_bandwidth_enabled(void)
219 return sysctl_sched_rt_runtime >= 0;
222 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
226 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
229 if (hrtimer_active(&rt_b->rt_period_timer))
232 spin_lock(&rt_b->rt_runtime_lock);
234 if (hrtimer_active(&rt_b->rt_period_timer))
237 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
238 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
239 hrtimer_start_expires(&rt_b->rt_period_timer,
242 spin_unlock(&rt_b->rt_runtime_lock);
245 #ifdef CONFIG_RT_GROUP_SCHED
246 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
248 hrtimer_cancel(&rt_b->rt_period_timer);
253 * sched_domains_mutex serializes calls to arch_init_sched_domains,
254 * detach_destroy_domains and partition_sched_domains.
256 static DEFINE_MUTEX(sched_domains_mutex);
258 #ifdef CONFIG_GROUP_SCHED
260 #include <linux/cgroup.h>
264 static LIST_HEAD(task_groups);
266 /* task group related information */
268 #ifdef CONFIG_CGROUP_SCHED
269 struct cgroup_subsys_state css;
272 #ifdef CONFIG_USER_SCHED
276 #ifdef CONFIG_FAIR_GROUP_SCHED
277 /* schedulable entities of this group on each cpu */
278 struct sched_entity **se;
279 /* runqueue "owned" by this group on each cpu */
280 struct cfs_rq **cfs_rq;
281 unsigned long shares;
284 #ifdef CONFIG_RT_GROUP_SCHED
285 struct sched_rt_entity **rt_se;
286 struct rt_rq **rt_rq;
288 struct rt_bandwidth rt_bandwidth;
292 struct list_head list;
294 struct task_group *parent;
295 struct list_head siblings;
296 struct list_head children;
299 #ifdef CONFIG_USER_SCHED
301 /* Helper function to pass uid information to create_sched_user() */
302 void set_tg_uid(struct user_struct *user)
304 user->tg->uid = user->uid;
309 * Every UID task group (including init_task_group aka UID-0) will
310 * be a child to this group.
312 struct task_group root_task_group;
314 #ifdef CONFIG_FAIR_GROUP_SCHED
315 /* Default task group's sched entity on each cpu */
316 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
317 /* Default task group's cfs_rq on each cpu */
318 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
319 #endif /* CONFIG_FAIR_GROUP_SCHED */
321 #ifdef CONFIG_RT_GROUP_SCHED
322 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
323 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
324 #endif /* CONFIG_RT_GROUP_SCHED */
325 #else /* !CONFIG_USER_SCHED */
326 #define root_task_group init_task_group
327 #endif /* CONFIG_USER_SCHED */
329 /* task_group_lock serializes add/remove of task groups and also changes to
330 * a task group's cpu shares.
332 static DEFINE_SPINLOCK(task_group_lock);
335 static int root_task_group_empty(void)
337 return list_empty(&root_task_group.children);
341 #ifdef CONFIG_FAIR_GROUP_SCHED
342 #ifdef CONFIG_USER_SCHED
343 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
344 #else /* !CONFIG_USER_SCHED */
345 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
346 #endif /* CONFIG_USER_SCHED */
349 * A weight of 0 or 1 can cause arithmetics problems.
350 * A weight of a cfs_rq is the sum of weights of which entities
351 * are queued on this cfs_rq, so a weight of a entity should not be
352 * too large, so as the shares value of a task group.
353 * (The default weight is 1024 - so there's no practical
354 * limitation from this.)
357 #define MAX_SHARES (1UL << 18)
359 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
362 /* Default task group.
363 * Every task in system belong to this group at bootup.
365 struct task_group init_task_group;
367 /* return group to which a task belongs */
368 static inline struct task_group *task_group(struct task_struct *p)
370 struct task_group *tg;
372 #ifdef CONFIG_USER_SCHED
374 tg = __task_cred(p)->user->tg;
376 #elif defined(CONFIG_CGROUP_SCHED)
377 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
378 struct task_group, css);
380 tg = &init_task_group;
385 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
386 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
388 #ifdef CONFIG_FAIR_GROUP_SCHED
389 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
390 p->se.parent = task_group(p)->se[cpu];
393 #ifdef CONFIG_RT_GROUP_SCHED
394 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
395 p->rt.parent = task_group(p)->rt_se[cpu];
402 static int root_task_group_empty(void)
408 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
409 static inline struct task_group *task_group(struct task_struct *p)
414 #endif /* CONFIG_GROUP_SCHED */
416 /* CFS-related fields in a runqueue */
418 struct load_weight load;
419 unsigned long nr_running;
424 struct rb_root tasks_timeline;
425 struct rb_node *rb_leftmost;
427 struct list_head tasks;
428 struct list_head *balance_iterator;
431 * 'curr' points to currently running entity on this cfs_rq.
432 * It is set to NULL otherwise (i.e when none are currently running).
434 struct sched_entity *curr, *next, *last;
436 unsigned int nr_spread_over;
438 #ifdef CONFIG_FAIR_GROUP_SCHED
439 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
442 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
443 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
444 * (like users, containers etc.)
446 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
447 * list is used during load balance.
449 struct list_head leaf_cfs_rq_list;
450 struct task_group *tg; /* group that "owns" this runqueue */
454 * the part of load.weight contributed by tasks
456 unsigned long task_weight;
459 * h_load = weight * f(tg)
461 * Where f(tg) is the recursive weight fraction assigned to
464 unsigned long h_load;
467 * this cpu's part of tg->shares
469 unsigned long shares;
472 * load.weight at the time we set shares
474 unsigned long rq_weight;
479 /* Real-Time classes' related field in a runqueue: */
481 struct rt_prio_array active;
482 unsigned long rt_nr_running;
483 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
485 int curr; /* highest queued rt task prio */
487 int next; /* next highest */
492 unsigned long rt_nr_migratory;
494 struct plist_head pushable_tasks;
499 /* Nests inside the rq lock: */
500 spinlock_t rt_runtime_lock;
502 #ifdef CONFIG_RT_GROUP_SCHED
503 unsigned long rt_nr_boosted;
506 struct list_head leaf_rt_rq_list;
507 struct task_group *tg;
508 struct sched_rt_entity *rt_se;
515 * We add the notion of a root-domain which will be used to define per-domain
516 * variables. Each exclusive cpuset essentially defines an island domain by
517 * fully partitioning the member cpus from any other cpuset. Whenever a new
518 * exclusive cpuset is created, we also create and attach a new root-domain
525 cpumask_var_t online;
528 * The "RT overload" flag: it gets set if a CPU has more than
529 * one runnable RT task.
531 cpumask_var_t rto_mask;
534 struct cpupri cpupri;
536 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
538 * Preferred wake up cpu nominated by sched_mc balance that will be
539 * used when most cpus are idle in the system indicating overall very
540 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
542 unsigned int sched_mc_preferred_wakeup_cpu;
547 * By default the system creates a single root-domain with all cpus as
548 * members (mimicking the global state we have today).
550 static struct root_domain def_root_domain;
555 * This is the main, per-CPU runqueue data structure.
557 * Locking rule: those places that want to lock multiple runqueues
558 * (such as the load balancing or the thread migration code), lock
559 * acquire operations must be ordered by ascending &runqueue.
566 * nr_running and cpu_load should be in the same cacheline because
567 * remote CPUs use both these fields when doing load calculation.
569 unsigned long nr_running;
570 #define CPU_LOAD_IDX_MAX 5
571 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
573 unsigned long last_tick_seen;
574 unsigned char in_nohz_recently;
576 /* capture load from *all* tasks on this cpu: */
577 struct load_weight load;
578 unsigned long nr_load_updates;
584 #ifdef CONFIG_FAIR_GROUP_SCHED
585 /* list of leaf cfs_rq on this cpu: */
586 struct list_head leaf_cfs_rq_list;
588 #ifdef CONFIG_RT_GROUP_SCHED
589 struct list_head leaf_rt_rq_list;
593 * This is part of a global counter where only the total sum
594 * over all CPUs matters. A task can increase this counter on
595 * one CPU and if it got migrated afterwards it may decrease
596 * it on another CPU. Always updated under the runqueue lock:
598 unsigned long nr_uninterruptible;
600 struct task_struct *curr, *idle;
601 unsigned long next_balance;
602 struct mm_struct *prev_mm;
609 struct root_domain *rd;
610 struct sched_domain *sd;
612 unsigned char idle_at_tick;
613 /* For active balancing */
616 /* cpu of this runqueue: */
620 unsigned long avg_load_per_task;
622 struct task_struct *migration_thread;
623 struct list_head migration_queue;
626 #ifdef CONFIG_SCHED_HRTICK
628 int hrtick_csd_pending;
629 struct call_single_data hrtick_csd;
631 struct hrtimer hrtick_timer;
634 #ifdef CONFIG_SCHEDSTATS
636 struct sched_info rq_sched_info;
637 unsigned long long rq_cpu_time;
638 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
640 /* sys_sched_yield() stats */
641 unsigned int yld_count;
643 /* schedule() stats */
644 unsigned int sched_switch;
645 unsigned int sched_count;
646 unsigned int sched_goidle;
648 /* try_to_wake_up() stats */
649 unsigned int ttwu_count;
650 unsigned int ttwu_local;
653 unsigned int bkl_count;
657 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
659 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
661 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
664 static inline int cpu_of(struct rq *rq)
674 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
675 * See detach_destroy_domains: synchronize_sched for details.
677 * The domain tree of any CPU may only be accessed from within
678 * preempt-disabled sections.
680 #define for_each_domain(cpu, __sd) \
681 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
683 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
684 #define this_rq() (&__get_cpu_var(runqueues))
685 #define task_rq(p) cpu_rq(task_cpu(p))
686 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
688 static inline void update_rq_clock(struct rq *rq)
690 rq->clock = sched_clock_cpu(cpu_of(rq));
694 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
696 #ifdef CONFIG_SCHED_DEBUG
697 # define const_debug __read_mostly
699 # define const_debug static const
705 * Returns true if the current cpu runqueue is locked.
706 * This interface allows printk to be called with the runqueue lock
707 * held and know whether or not it is OK to wake up the klogd.
709 int runqueue_is_locked(void)
712 struct rq *rq = cpu_rq(cpu);
715 ret = spin_is_locked(&rq->lock);
721 * Debugging: various feature bits
724 #define SCHED_FEAT(name, enabled) \
725 __SCHED_FEAT_##name ,
728 #include "sched_features.h"
733 #define SCHED_FEAT(name, enabled) \
734 (1UL << __SCHED_FEAT_##name) * enabled |
736 const_debug unsigned int sysctl_sched_features =
737 #include "sched_features.h"
742 #ifdef CONFIG_SCHED_DEBUG
743 #define SCHED_FEAT(name, enabled) \
746 static __read_mostly char *sched_feat_names[] = {
747 #include "sched_features.h"
753 static int sched_feat_show(struct seq_file *m, void *v)
757 for (i = 0; sched_feat_names[i]; i++) {
758 if (!(sysctl_sched_features & (1UL << i)))
760 seq_printf(m, "%s ", sched_feat_names[i]);
768 sched_feat_write(struct file *filp, const char __user *ubuf,
769 size_t cnt, loff_t *ppos)
779 if (copy_from_user(&buf, ubuf, cnt))
784 if (strncmp(buf, "NO_", 3) == 0) {
789 for (i = 0; sched_feat_names[i]; i++) {
790 int len = strlen(sched_feat_names[i]);
792 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
794 sysctl_sched_features &= ~(1UL << i);
796 sysctl_sched_features |= (1UL << i);
801 if (!sched_feat_names[i])
809 static int sched_feat_open(struct inode *inode, struct file *filp)
811 return single_open(filp, sched_feat_show, NULL);
814 static struct file_operations sched_feat_fops = {
815 .open = sched_feat_open,
816 .write = sched_feat_write,
819 .release = single_release,
822 static __init int sched_init_debug(void)
824 debugfs_create_file("sched_features", 0644, NULL, NULL,
829 late_initcall(sched_init_debug);
833 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
836 * Number of tasks to iterate in a single balance run.
837 * Limited because this is done with IRQs disabled.
839 const_debug unsigned int sysctl_sched_nr_migrate = 32;
842 * ratelimit for updating the group shares.
845 unsigned int sysctl_sched_shares_ratelimit = 250000;
848 * Inject some fuzzyness into changing the per-cpu group shares
849 * this avoids remote rq-locks at the expense of fairness.
852 unsigned int sysctl_sched_shares_thresh = 4;
855 * period over which we measure -rt task cpu usage in us.
858 unsigned int sysctl_sched_rt_period = 1000000;
860 static __read_mostly int scheduler_running;
863 * part of the period that we allow rt tasks to run in us.
866 int sysctl_sched_rt_runtime = 950000;
868 static inline u64 global_rt_period(void)
870 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
873 static inline u64 global_rt_runtime(void)
875 if (sysctl_sched_rt_runtime < 0)
878 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
881 #ifndef prepare_arch_switch
882 # define prepare_arch_switch(next) do { } while (0)
884 #ifndef finish_arch_switch
885 # define finish_arch_switch(prev) do { } while (0)
888 static inline int task_current(struct rq *rq, struct task_struct *p)
890 return rq->curr == p;
893 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
894 static inline int task_running(struct rq *rq, struct task_struct *p)
896 return task_current(rq, p);
899 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
903 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
905 #ifdef CONFIG_DEBUG_SPINLOCK
906 /* this is a valid case when another task releases the spinlock */
907 rq->lock.owner = current;
910 * If we are tracking spinlock dependencies then we have to
911 * fix up the runqueue lock - which gets 'carried over' from
914 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
916 spin_unlock_irq(&rq->lock);
919 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
920 static inline int task_running(struct rq *rq, struct task_struct *p)
925 return task_current(rq, p);
929 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
933 * We can optimise this out completely for !SMP, because the
934 * SMP rebalancing from interrupt is the only thing that cares
939 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
940 spin_unlock_irq(&rq->lock);
942 spin_unlock(&rq->lock);
946 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
950 * After ->oncpu is cleared, the task can be moved to a different CPU.
951 * We must ensure this doesn't happen until the switch is completely
957 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
961 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
964 * __task_rq_lock - lock the runqueue a given task resides on.
965 * Must be called interrupts disabled.
967 static inline struct rq *__task_rq_lock(struct task_struct *p)
971 struct rq *rq = task_rq(p);
972 spin_lock(&rq->lock);
973 if (likely(rq == task_rq(p)))
975 spin_unlock(&rq->lock);
980 * task_rq_lock - lock the runqueue a given task resides on and disable
981 * interrupts. Note the ordering: we can safely lookup the task_rq without
982 * explicitly disabling preemption.
984 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
990 local_irq_save(*flags);
992 spin_lock(&rq->lock);
993 if (likely(rq == task_rq(p)))
995 spin_unlock_irqrestore(&rq->lock, *flags);
999 void task_rq_unlock_wait(struct task_struct *p)
1001 struct rq *rq = task_rq(p);
1003 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1004 spin_unlock_wait(&rq->lock);
1007 static void __task_rq_unlock(struct rq *rq)
1008 __releases(rq->lock)
1010 spin_unlock(&rq->lock);
1013 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1014 __releases(rq->lock)
1016 spin_unlock_irqrestore(&rq->lock, *flags);
1020 * this_rq_lock - lock this runqueue and disable interrupts.
1022 static struct rq *this_rq_lock(void)
1023 __acquires(rq->lock)
1027 local_irq_disable();
1029 spin_lock(&rq->lock);
1034 #ifdef CONFIG_SCHED_HRTICK
1036 * Use HR-timers to deliver accurate preemption points.
1038 * Its all a bit involved since we cannot program an hrt while holding the
1039 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1042 * When we get rescheduled we reprogram the hrtick_timer outside of the
1048 * - enabled by features
1049 * - hrtimer is actually high res
1051 static inline int hrtick_enabled(struct rq *rq)
1053 if (!sched_feat(HRTICK))
1055 if (!cpu_active(cpu_of(rq)))
1057 return hrtimer_is_hres_active(&rq->hrtick_timer);
1060 static void hrtick_clear(struct rq *rq)
1062 if (hrtimer_active(&rq->hrtick_timer))
1063 hrtimer_cancel(&rq->hrtick_timer);
1067 * High-resolution timer tick.
1068 * Runs from hardirq context with interrupts disabled.
1070 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1072 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1074 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1076 spin_lock(&rq->lock);
1077 update_rq_clock(rq);
1078 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1079 spin_unlock(&rq->lock);
1081 return HRTIMER_NORESTART;
1086 * called from hardirq (IPI) context
1088 static void __hrtick_start(void *arg)
1090 struct rq *rq = arg;
1092 spin_lock(&rq->lock);
1093 hrtimer_restart(&rq->hrtick_timer);
1094 rq->hrtick_csd_pending = 0;
1095 spin_unlock(&rq->lock);
1099 * Called to set the hrtick timer state.
1101 * called with rq->lock held and irqs disabled
1103 static void hrtick_start(struct rq *rq, u64 delay)
1105 struct hrtimer *timer = &rq->hrtick_timer;
1106 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1108 hrtimer_set_expires(timer, time);
1110 if (rq == this_rq()) {
1111 hrtimer_restart(timer);
1112 } else if (!rq->hrtick_csd_pending) {
1113 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd);
1114 rq->hrtick_csd_pending = 1;
1119 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1121 int cpu = (int)(long)hcpu;
1124 case CPU_UP_CANCELED:
1125 case CPU_UP_CANCELED_FROZEN:
1126 case CPU_DOWN_PREPARE:
1127 case CPU_DOWN_PREPARE_FROZEN:
1129 case CPU_DEAD_FROZEN:
1130 hrtick_clear(cpu_rq(cpu));
1137 static __init void init_hrtick(void)
1139 hotcpu_notifier(hotplug_hrtick, 0);
1143 * Called to set the hrtick timer state.
1145 * called with rq->lock held and irqs disabled
1147 static void hrtick_start(struct rq *rq, u64 delay)
1149 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay), HRTIMER_MODE_REL);
1152 static inline void init_hrtick(void)
1155 #endif /* CONFIG_SMP */
1157 static void init_rq_hrtick(struct rq *rq)
1160 rq->hrtick_csd_pending = 0;
1162 rq->hrtick_csd.flags = 0;
1163 rq->hrtick_csd.func = __hrtick_start;
1164 rq->hrtick_csd.info = rq;
1167 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1168 rq->hrtick_timer.function = hrtick;
1170 #else /* CONFIG_SCHED_HRTICK */
1171 static inline void hrtick_clear(struct rq *rq)
1175 static inline void init_rq_hrtick(struct rq *rq)
1179 static inline void init_hrtick(void)
1182 #endif /* CONFIG_SCHED_HRTICK */
1185 * resched_task - mark a task 'to be rescheduled now'.
1187 * On UP this means the setting of the need_resched flag, on SMP it
1188 * might also involve a cross-CPU call to trigger the scheduler on
1193 #ifndef tsk_is_polling
1194 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1197 static void resched_task(struct task_struct *p)
1201 assert_spin_locked(&task_rq(p)->lock);
1203 if (test_tsk_need_resched(p))
1206 set_tsk_need_resched(p);
1209 if (cpu == smp_processor_id())
1212 /* NEED_RESCHED must be visible before we test polling */
1214 if (!tsk_is_polling(p))
1215 smp_send_reschedule(cpu);
1218 static void resched_cpu(int cpu)
1220 struct rq *rq = cpu_rq(cpu);
1221 unsigned long flags;
1223 if (!spin_trylock_irqsave(&rq->lock, flags))
1225 resched_task(cpu_curr(cpu));
1226 spin_unlock_irqrestore(&rq->lock, flags);
1231 * When add_timer_on() enqueues a timer into the timer wheel of an
1232 * idle CPU then this timer might expire before the next timer event
1233 * which is scheduled to wake up that CPU. In case of a completely
1234 * idle system the next event might even be infinite time into the
1235 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1236 * leaves the inner idle loop so the newly added timer is taken into
1237 * account when the CPU goes back to idle and evaluates the timer
1238 * wheel for the next timer event.
1240 void wake_up_idle_cpu(int cpu)
1242 struct rq *rq = cpu_rq(cpu);
1244 if (cpu == smp_processor_id())
1248 * This is safe, as this function is called with the timer
1249 * wheel base lock of (cpu) held. When the CPU is on the way
1250 * to idle and has not yet set rq->curr to idle then it will
1251 * be serialized on the timer wheel base lock and take the new
1252 * timer into account automatically.
1254 if (rq->curr != rq->idle)
1258 * We can set TIF_RESCHED on the idle task of the other CPU
1259 * lockless. The worst case is that the other CPU runs the
1260 * idle task through an additional NOOP schedule()
1262 set_tsk_need_resched(rq->idle);
1264 /* NEED_RESCHED must be visible before we test polling */
1266 if (!tsk_is_polling(rq->idle))
1267 smp_send_reschedule(cpu);
1269 #endif /* CONFIG_NO_HZ */
1271 #else /* !CONFIG_SMP */
1272 static void resched_task(struct task_struct *p)
1274 assert_spin_locked(&task_rq(p)->lock);
1275 set_tsk_need_resched(p);
1277 #endif /* CONFIG_SMP */
1279 #if BITS_PER_LONG == 32
1280 # define WMULT_CONST (~0UL)
1282 # define WMULT_CONST (1UL << 32)
1285 #define WMULT_SHIFT 32
1288 * Shift right and round:
1290 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1293 * delta *= weight / lw
1295 static unsigned long
1296 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1297 struct load_weight *lw)
1301 if (!lw->inv_weight) {
1302 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1305 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1309 tmp = (u64)delta_exec * weight;
1311 * Check whether we'd overflow the 64-bit multiplication:
1313 if (unlikely(tmp > WMULT_CONST))
1314 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1317 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1319 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1322 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1328 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1335 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1336 * of tasks with abnormal "nice" values across CPUs the contribution that
1337 * each task makes to its run queue's load is weighted according to its
1338 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1339 * scaled version of the new time slice allocation that they receive on time
1343 #define WEIGHT_IDLEPRIO 3
1344 #define WMULT_IDLEPRIO 1431655765
1347 * Nice levels are multiplicative, with a gentle 10% change for every
1348 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1349 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1350 * that remained on nice 0.
1352 * The "10% effect" is relative and cumulative: from _any_ nice level,
1353 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1354 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1355 * If a task goes up by ~10% and another task goes down by ~10% then
1356 * the relative distance between them is ~25%.)
1358 static const int prio_to_weight[40] = {
1359 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1360 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1361 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1362 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1363 /* 0 */ 1024, 820, 655, 526, 423,
1364 /* 5 */ 335, 272, 215, 172, 137,
1365 /* 10 */ 110, 87, 70, 56, 45,
1366 /* 15 */ 36, 29, 23, 18, 15,
1370 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1372 * In cases where the weight does not change often, we can use the
1373 * precalculated inverse to speed up arithmetics by turning divisions
1374 * into multiplications:
1376 static const u32 prio_to_wmult[40] = {
1377 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1378 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1379 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1380 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1381 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1382 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1383 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1384 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1387 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1390 * runqueue iterator, to support SMP load-balancing between different
1391 * scheduling classes, without having to expose their internal data
1392 * structures to the load-balancing proper:
1394 struct rq_iterator {
1396 struct task_struct *(*start)(void *);
1397 struct task_struct *(*next)(void *);
1401 static unsigned long
1402 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1403 unsigned long max_load_move, struct sched_domain *sd,
1404 enum cpu_idle_type idle, int *all_pinned,
1405 int *this_best_prio, struct rq_iterator *iterator);
1408 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1409 struct sched_domain *sd, enum cpu_idle_type idle,
1410 struct rq_iterator *iterator);
1413 #ifdef CONFIG_CGROUP_CPUACCT
1414 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1416 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1419 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1421 update_load_add(&rq->load, load);
1424 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1426 update_load_sub(&rq->load, load);
1429 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1430 typedef int (*tg_visitor)(struct task_group *, void *);
1433 * Iterate the full tree, calling @down when first entering a node and @up when
1434 * leaving it for the final time.
1436 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1438 struct task_group *parent, *child;
1442 parent = &root_task_group;
1444 ret = (*down)(parent, data);
1447 list_for_each_entry_rcu(child, &parent->children, siblings) {
1454 ret = (*up)(parent, data);
1459 parent = parent->parent;
1468 static int tg_nop(struct task_group *tg, void *data)
1475 static unsigned long source_load(int cpu, int type);
1476 static unsigned long target_load(int cpu, int type);
1477 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1479 static unsigned long cpu_avg_load_per_task(int cpu)
1481 struct rq *rq = cpu_rq(cpu);
1482 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1485 rq->avg_load_per_task = rq->load.weight / nr_running;
1487 rq->avg_load_per_task = 0;
1489 return rq->avg_load_per_task;
1492 #ifdef CONFIG_FAIR_GROUP_SCHED
1494 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1497 * Calculate and set the cpu's group shares.
1500 update_group_shares_cpu(struct task_group *tg, int cpu,
1501 unsigned long sd_shares, unsigned long sd_rq_weight)
1503 unsigned long shares;
1504 unsigned long rq_weight;
1509 rq_weight = tg->cfs_rq[cpu]->rq_weight;
1512 * \Sum shares * rq_weight
1513 * shares = -----------------------
1517 shares = (sd_shares * rq_weight) / sd_rq_weight;
1518 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1520 if (abs(shares - tg->se[cpu]->load.weight) >
1521 sysctl_sched_shares_thresh) {
1522 struct rq *rq = cpu_rq(cpu);
1523 unsigned long flags;
1525 spin_lock_irqsave(&rq->lock, flags);
1526 tg->cfs_rq[cpu]->shares = shares;
1528 __set_se_shares(tg->se[cpu], shares);
1529 spin_unlock_irqrestore(&rq->lock, flags);
1534 * Re-compute the task group their per cpu shares over the given domain.
1535 * This needs to be done in a bottom-up fashion because the rq weight of a
1536 * parent group depends on the shares of its child groups.
1538 static int tg_shares_up(struct task_group *tg, void *data)
1540 unsigned long weight, rq_weight = 0;
1541 unsigned long shares = 0;
1542 struct sched_domain *sd = data;
1545 for_each_cpu(i, sched_domain_span(sd)) {
1547 * If there are currently no tasks on the cpu pretend there
1548 * is one of average load so that when a new task gets to
1549 * run here it will not get delayed by group starvation.
1551 weight = tg->cfs_rq[i]->load.weight;
1553 weight = NICE_0_LOAD;
1555 tg->cfs_rq[i]->rq_weight = weight;
1556 rq_weight += weight;
1557 shares += tg->cfs_rq[i]->shares;
1560 if ((!shares && rq_weight) || shares > tg->shares)
1561 shares = tg->shares;
1563 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1564 shares = tg->shares;
1566 for_each_cpu(i, sched_domain_span(sd))
1567 update_group_shares_cpu(tg, i, shares, rq_weight);
1573 * Compute the cpu's hierarchical load factor for each task group.
1574 * This needs to be done in a top-down fashion because the load of a child
1575 * group is a fraction of its parents load.
1577 static int tg_load_down(struct task_group *tg, void *data)
1580 long cpu = (long)data;
1583 load = cpu_rq(cpu)->load.weight;
1585 load = tg->parent->cfs_rq[cpu]->h_load;
1586 load *= tg->cfs_rq[cpu]->shares;
1587 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1590 tg->cfs_rq[cpu]->h_load = load;
1595 static void update_shares(struct sched_domain *sd)
1597 u64 now = cpu_clock(raw_smp_processor_id());
1598 s64 elapsed = now - sd->last_update;
1600 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1601 sd->last_update = now;
1602 walk_tg_tree(tg_nop, tg_shares_up, sd);
1606 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1608 spin_unlock(&rq->lock);
1610 spin_lock(&rq->lock);
1613 static void update_h_load(long cpu)
1615 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1620 static inline void update_shares(struct sched_domain *sd)
1624 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1630 #ifdef CONFIG_PREEMPT
1633 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1634 * way at the expense of forcing extra atomic operations in all
1635 * invocations. This assures that the double_lock is acquired using the
1636 * same underlying policy as the spinlock_t on this architecture, which
1637 * reduces latency compared to the unfair variant below. However, it
1638 * also adds more overhead and therefore may reduce throughput.
1640 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1641 __releases(this_rq->lock)
1642 __acquires(busiest->lock)
1643 __acquires(this_rq->lock)
1645 spin_unlock(&this_rq->lock);
1646 double_rq_lock(this_rq, busiest);
1653 * Unfair double_lock_balance: Optimizes throughput at the expense of
1654 * latency by eliminating extra atomic operations when the locks are
1655 * already in proper order on entry. This favors lower cpu-ids and will
1656 * grant the double lock to lower cpus over higher ids under contention,
1657 * regardless of entry order into the function.
1659 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1660 __releases(this_rq->lock)
1661 __acquires(busiest->lock)
1662 __acquires(this_rq->lock)
1666 if (unlikely(!spin_trylock(&busiest->lock))) {
1667 if (busiest < this_rq) {
1668 spin_unlock(&this_rq->lock);
1669 spin_lock(&busiest->lock);
1670 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1673 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1678 #endif /* CONFIG_PREEMPT */
1681 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1683 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1685 if (unlikely(!irqs_disabled())) {
1686 /* printk() doesn't work good under rq->lock */
1687 spin_unlock(&this_rq->lock);
1691 return _double_lock_balance(this_rq, busiest);
1694 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1695 __releases(busiest->lock)
1697 spin_unlock(&busiest->lock);
1698 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1702 #ifdef CONFIG_FAIR_GROUP_SCHED
1703 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1706 cfs_rq->shares = shares;
1711 #include "sched_stats.h"
1712 #include "sched_idletask.c"
1713 #include "sched_fair.c"
1714 #include "sched_rt.c"
1715 #ifdef CONFIG_SCHED_DEBUG
1716 # include "sched_debug.c"
1719 #define sched_class_highest (&rt_sched_class)
1720 #define for_each_class(class) \
1721 for (class = sched_class_highest; class; class = class->next)
1723 static void inc_nr_running(struct rq *rq)
1728 static void dec_nr_running(struct rq *rq)
1733 static void set_load_weight(struct task_struct *p)
1735 if (task_has_rt_policy(p)) {
1736 p->se.load.weight = prio_to_weight[0] * 2;
1737 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1742 * SCHED_IDLE tasks get minimal weight:
1744 if (p->policy == SCHED_IDLE) {
1745 p->se.load.weight = WEIGHT_IDLEPRIO;
1746 p->se.load.inv_weight = WMULT_IDLEPRIO;
1750 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1751 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1754 static void update_avg(u64 *avg, u64 sample)
1756 s64 diff = sample - *avg;
1760 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1763 p->se.start_runtime = p->se.sum_exec_runtime;
1765 sched_info_queued(p);
1766 p->sched_class->enqueue_task(rq, p, wakeup);
1770 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1773 if (p->se.last_wakeup) {
1774 update_avg(&p->se.avg_overlap,
1775 p->se.sum_exec_runtime - p->se.last_wakeup);
1776 p->se.last_wakeup = 0;
1778 update_avg(&p->se.avg_wakeup,
1779 sysctl_sched_wakeup_granularity);
1783 sched_info_dequeued(p);
1784 p->sched_class->dequeue_task(rq, p, sleep);
1789 * __normal_prio - return the priority that is based on the static prio
1791 static inline int __normal_prio(struct task_struct *p)
1793 return p->static_prio;
1797 * Calculate the expected normal priority: i.e. priority
1798 * without taking RT-inheritance into account. Might be
1799 * boosted by interactivity modifiers. Changes upon fork,
1800 * setprio syscalls, and whenever the interactivity
1801 * estimator recalculates.
1803 static inline int normal_prio(struct task_struct *p)
1807 if (task_has_rt_policy(p))
1808 prio = MAX_RT_PRIO-1 - p->rt_priority;
1810 prio = __normal_prio(p);
1815 * Calculate the current priority, i.e. the priority
1816 * taken into account by the scheduler. This value might
1817 * be boosted by RT tasks, or might be boosted by
1818 * interactivity modifiers. Will be RT if the task got
1819 * RT-boosted. If not then it returns p->normal_prio.
1821 static int effective_prio(struct task_struct *p)
1823 p->normal_prio = normal_prio(p);
1825 * If we are RT tasks or we were boosted to RT priority,
1826 * keep the priority unchanged. Otherwise, update priority
1827 * to the normal priority:
1829 if (!rt_prio(p->prio))
1830 return p->normal_prio;
1835 * activate_task - move a task to the runqueue.
1837 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1839 if (task_contributes_to_load(p))
1840 rq->nr_uninterruptible--;
1842 enqueue_task(rq, p, wakeup);
1847 * deactivate_task - remove a task from the runqueue.
1849 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1851 if (task_contributes_to_load(p))
1852 rq->nr_uninterruptible++;
1854 dequeue_task(rq, p, sleep);
1859 * task_curr - is this task currently executing on a CPU?
1860 * @p: the task in question.
1862 inline int task_curr(const struct task_struct *p)
1864 return cpu_curr(task_cpu(p)) == p;
1867 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1869 set_task_rq(p, cpu);
1872 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1873 * successfuly executed on another CPU. We must ensure that updates of
1874 * per-task data have been completed by this moment.
1877 task_thread_info(p)->cpu = cpu;
1881 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1882 const struct sched_class *prev_class,
1883 int oldprio, int running)
1885 if (prev_class != p->sched_class) {
1886 if (prev_class->switched_from)
1887 prev_class->switched_from(rq, p, running);
1888 p->sched_class->switched_to(rq, p, running);
1890 p->sched_class->prio_changed(rq, p, oldprio, running);
1895 /* Used instead of source_load when we know the type == 0 */
1896 static unsigned long weighted_cpuload(const int cpu)
1898 return cpu_rq(cpu)->load.weight;
1902 * Is this task likely cache-hot:
1905 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1910 * Buddy candidates are cache hot:
1912 if (sched_feat(CACHE_HOT_BUDDY) &&
1913 (&p->se == cfs_rq_of(&p->se)->next ||
1914 &p->se == cfs_rq_of(&p->se)->last))
1917 if (p->sched_class != &fair_sched_class)
1920 if (sysctl_sched_migration_cost == -1)
1922 if (sysctl_sched_migration_cost == 0)
1925 delta = now - p->se.exec_start;
1927 return delta < (s64)sysctl_sched_migration_cost;
1931 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1933 int old_cpu = task_cpu(p);
1934 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1935 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1936 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1939 clock_offset = old_rq->clock - new_rq->clock;
1941 trace_sched_migrate_task(p, task_cpu(p), new_cpu);
1943 #ifdef CONFIG_SCHEDSTATS
1944 if (p->se.wait_start)
1945 p->se.wait_start -= clock_offset;
1946 if (p->se.sleep_start)
1947 p->se.sleep_start -= clock_offset;
1948 if (p->se.block_start)
1949 p->se.block_start -= clock_offset;
1950 if (old_cpu != new_cpu) {
1951 schedstat_inc(p, se.nr_migrations);
1952 if (task_hot(p, old_rq->clock, NULL))
1953 schedstat_inc(p, se.nr_forced2_migrations);
1956 p->se.vruntime -= old_cfsrq->min_vruntime -
1957 new_cfsrq->min_vruntime;
1959 __set_task_cpu(p, new_cpu);
1962 struct migration_req {
1963 struct list_head list;
1965 struct task_struct *task;
1968 struct completion done;
1972 * The task's runqueue lock must be held.
1973 * Returns true if you have to wait for migration thread.
1976 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1978 struct rq *rq = task_rq(p);
1981 * If the task is not on a runqueue (and not running), then
1982 * it is sufficient to simply update the task's cpu field.
1984 if (!p->se.on_rq && !task_running(rq, p)) {
1985 set_task_cpu(p, dest_cpu);
1989 init_completion(&req->done);
1991 req->dest_cpu = dest_cpu;
1992 list_add(&req->list, &rq->migration_queue);
1998 * wait_task_inactive - wait for a thread to unschedule.
2000 * If @match_state is nonzero, it's the @p->state value just checked and
2001 * not expected to change. If it changes, i.e. @p might have woken up,
2002 * then return zero. When we succeed in waiting for @p to be off its CPU,
2003 * we return a positive number (its total switch count). If a second call
2004 * a short while later returns the same number, the caller can be sure that
2005 * @p has remained unscheduled the whole time.
2007 * The caller must ensure that the task *will* unschedule sometime soon,
2008 * else this function might spin for a *long* time. This function can't
2009 * be called with interrupts off, or it may introduce deadlock with
2010 * smp_call_function() if an IPI is sent by the same process we are
2011 * waiting to become inactive.
2013 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2015 unsigned long flags;
2022 * We do the initial early heuristics without holding
2023 * any task-queue locks at all. We'll only try to get
2024 * the runqueue lock when things look like they will
2030 * If the task is actively running on another CPU
2031 * still, just relax and busy-wait without holding
2034 * NOTE! Since we don't hold any locks, it's not
2035 * even sure that "rq" stays as the right runqueue!
2036 * But we don't care, since "task_running()" will
2037 * return false if the runqueue has changed and p
2038 * is actually now running somewhere else!
2040 while (task_running(rq, p)) {
2041 if (match_state && unlikely(p->state != match_state))
2047 * Ok, time to look more closely! We need the rq
2048 * lock now, to be *sure*. If we're wrong, we'll
2049 * just go back and repeat.
2051 rq = task_rq_lock(p, &flags);
2052 trace_sched_wait_task(rq, p);
2053 running = task_running(rq, p);
2054 on_rq = p->se.on_rq;
2056 if (!match_state || p->state == match_state)
2057 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2058 task_rq_unlock(rq, &flags);
2061 * If it changed from the expected state, bail out now.
2063 if (unlikely(!ncsw))
2067 * Was it really running after all now that we
2068 * checked with the proper locks actually held?
2070 * Oops. Go back and try again..
2072 if (unlikely(running)) {
2078 * It's not enough that it's not actively running,
2079 * it must be off the runqueue _entirely_, and not
2082 * So if it was still runnable (but just not actively
2083 * running right now), it's preempted, and we should
2084 * yield - it could be a while.
2086 if (unlikely(on_rq)) {
2087 schedule_timeout_uninterruptible(1);
2092 * Ahh, all good. It wasn't running, and it wasn't
2093 * runnable, which means that it will never become
2094 * running in the future either. We're all done!
2103 * kick_process - kick a running thread to enter/exit the kernel
2104 * @p: the to-be-kicked thread
2106 * Cause a process which is running on another CPU to enter
2107 * kernel-mode, without any delay. (to get signals handled.)
2109 * NOTE: this function doesnt have to take the runqueue lock,
2110 * because all it wants to ensure is that the remote task enters
2111 * the kernel. If the IPI races and the task has been migrated
2112 * to another CPU then no harm is done and the purpose has been
2115 void kick_process(struct task_struct *p)
2121 if ((cpu != smp_processor_id()) && task_curr(p))
2122 smp_send_reschedule(cpu);
2127 * Return a low guess at the load of a migration-source cpu weighted
2128 * according to the scheduling class and "nice" value.
2130 * We want to under-estimate the load of migration sources, to
2131 * balance conservatively.
2133 static unsigned long source_load(int cpu, int type)
2135 struct rq *rq = cpu_rq(cpu);
2136 unsigned long total = weighted_cpuload(cpu);
2138 if (type == 0 || !sched_feat(LB_BIAS))
2141 return min(rq->cpu_load[type-1], total);
2145 * Return a high guess at the load of a migration-target cpu weighted
2146 * according to the scheduling class and "nice" value.
2148 static unsigned long target_load(int cpu, int type)
2150 struct rq *rq = cpu_rq(cpu);
2151 unsigned long total = weighted_cpuload(cpu);
2153 if (type == 0 || !sched_feat(LB_BIAS))
2156 return max(rq->cpu_load[type-1], total);
2160 * find_idlest_group finds and returns the least busy CPU group within the
2163 static struct sched_group *
2164 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2166 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2167 unsigned long min_load = ULONG_MAX, this_load = 0;
2168 int load_idx = sd->forkexec_idx;
2169 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2172 unsigned long load, avg_load;
2176 /* Skip over this group if it has no CPUs allowed */
2177 if (!cpumask_intersects(sched_group_cpus(group),
2181 local_group = cpumask_test_cpu(this_cpu,
2182 sched_group_cpus(group));
2184 /* Tally up the load of all CPUs in the group */
2187 for_each_cpu(i, sched_group_cpus(group)) {
2188 /* Bias balancing toward cpus of our domain */
2190 load = source_load(i, load_idx);
2192 load = target_load(i, load_idx);
2197 /* Adjust by relative CPU power of the group */
2198 avg_load = sg_div_cpu_power(group,
2199 avg_load * SCHED_LOAD_SCALE);
2202 this_load = avg_load;
2204 } else if (avg_load < min_load) {
2205 min_load = avg_load;
2208 } while (group = group->next, group != sd->groups);
2210 if (!idlest || 100*this_load < imbalance*min_load)
2216 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2219 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2221 unsigned long load, min_load = ULONG_MAX;
2225 /* Traverse only the allowed CPUs */
2226 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
2227 load = weighted_cpuload(i);
2229 if (load < min_load || (load == min_load && i == this_cpu)) {
2239 * sched_balance_self: balance the current task (running on cpu) in domains
2240 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2243 * Balance, ie. select the least loaded group.
2245 * Returns the target CPU number, or the same CPU if no balancing is needed.
2247 * preempt must be disabled.
2249 static int sched_balance_self(int cpu, int flag)
2251 struct task_struct *t = current;
2252 struct sched_domain *tmp, *sd = NULL;
2254 for_each_domain(cpu, tmp) {
2256 * If power savings logic is enabled for a domain, stop there.
2258 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2260 if (tmp->flags & flag)
2268 struct sched_group *group;
2269 int new_cpu, weight;
2271 if (!(sd->flags & flag)) {
2276 group = find_idlest_group(sd, t, cpu);
2282 new_cpu = find_idlest_cpu(group, t, cpu);
2283 if (new_cpu == -1 || new_cpu == cpu) {
2284 /* Now try balancing at a lower domain level of cpu */
2289 /* Now try balancing at a lower domain level of new_cpu */
2291 weight = cpumask_weight(sched_domain_span(sd));
2293 for_each_domain(cpu, tmp) {
2294 if (weight <= cpumask_weight(sched_domain_span(tmp)))
2296 if (tmp->flags & flag)
2299 /* while loop will break here if sd == NULL */
2305 #endif /* CONFIG_SMP */
2308 * try_to_wake_up - wake up a thread
2309 * @p: the to-be-woken-up thread
2310 * @state: the mask of task states that can be woken
2311 * @sync: do a synchronous wakeup?
2313 * Put it on the run-queue if it's not already there. The "current"
2314 * thread is always on the run-queue (except when the actual
2315 * re-schedule is in progress), and as such you're allowed to do
2316 * the simpler "current->state = TASK_RUNNING" to mark yourself
2317 * runnable without the overhead of this.
2319 * returns failure only if the task is already active.
2321 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2323 int cpu, orig_cpu, this_cpu, success = 0;
2324 unsigned long flags;
2328 if (!sched_feat(SYNC_WAKEUPS))
2332 if (sched_feat(LB_WAKEUP_UPDATE) && !root_task_group_empty()) {
2333 struct sched_domain *sd;
2335 this_cpu = raw_smp_processor_id();
2338 for_each_domain(this_cpu, sd) {
2339 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2348 rq = task_rq_lock(p, &flags);
2349 update_rq_clock(rq);
2350 old_state = p->state;
2351 if (!(old_state & state))
2359 this_cpu = smp_processor_id();
2362 if (unlikely(task_running(rq, p)))
2365 cpu = p->sched_class->select_task_rq(p, sync);
2366 if (cpu != orig_cpu) {
2367 set_task_cpu(p, cpu);
2368 task_rq_unlock(rq, &flags);
2369 /* might preempt at this point */
2370 rq = task_rq_lock(p, &flags);
2371 old_state = p->state;
2372 if (!(old_state & state))
2377 this_cpu = smp_processor_id();
2381 #ifdef CONFIG_SCHEDSTATS
2382 schedstat_inc(rq, ttwu_count);
2383 if (cpu == this_cpu)
2384 schedstat_inc(rq, ttwu_local);
2386 struct sched_domain *sd;
2387 for_each_domain(this_cpu, sd) {
2388 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2389 schedstat_inc(sd, ttwu_wake_remote);
2394 #endif /* CONFIG_SCHEDSTATS */
2397 #endif /* CONFIG_SMP */
2398 schedstat_inc(p, se.nr_wakeups);
2400 schedstat_inc(p, se.nr_wakeups_sync);
2401 if (orig_cpu != cpu)
2402 schedstat_inc(p, se.nr_wakeups_migrate);
2403 if (cpu == this_cpu)
2404 schedstat_inc(p, se.nr_wakeups_local);
2406 schedstat_inc(p, se.nr_wakeups_remote);
2407 activate_task(rq, p, 1);
2411 * Only attribute actual wakeups done by this task.
2413 if (!in_interrupt()) {
2414 struct sched_entity *se = ¤t->se;
2415 u64 sample = se->sum_exec_runtime;
2417 if (se->last_wakeup)
2418 sample -= se->last_wakeup;
2420 sample -= se->start_runtime;
2421 update_avg(&se->avg_wakeup, sample);
2423 se->last_wakeup = se->sum_exec_runtime;
2427 trace_sched_wakeup(rq, p, success);
2428 check_preempt_curr(rq, p, sync);
2430 p->state = TASK_RUNNING;
2432 if (p->sched_class->task_wake_up)
2433 p->sched_class->task_wake_up(rq, p);
2436 task_rq_unlock(rq, &flags);
2441 int wake_up_process(struct task_struct *p)
2443 return try_to_wake_up(p, TASK_ALL, 0);
2445 EXPORT_SYMBOL(wake_up_process);
2447 int wake_up_state(struct task_struct *p, unsigned int state)
2449 return try_to_wake_up(p, state, 0);
2453 * Perform scheduler related setup for a newly forked process p.
2454 * p is forked by current.
2456 * __sched_fork() is basic setup used by init_idle() too:
2458 static void __sched_fork(struct task_struct *p)
2460 p->se.exec_start = 0;
2461 p->se.sum_exec_runtime = 0;
2462 p->se.prev_sum_exec_runtime = 0;
2463 p->se.last_wakeup = 0;
2464 p->se.avg_overlap = 0;
2465 p->se.start_runtime = 0;
2466 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2468 #ifdef CONFIG_SCHEDSTATS
2469 p->se.wait_start = 0;
2470 p->se.sum_sleep_runtime = 0;
2471 p->se.sleep_start = 0;
2472 p->se.block_start = 0;
2473 p->se.sleep_max = 0;
2474 p->se.block_max = 0;
2476 p->se.slice_max = 0;
2480 INIT_LIST_HEAD(&p->rt.run_list);
2482 INIT_LIST_HEAD(&p->se.group_node);
2484 #ifdef CONFIG_PREEMPT_NOTIFIERS
2485 INIT_HLIST_HEAD(&p->preempt_notifiers);
2489 * We mark the process as running here, but have not actually
2490 * inserted it onto the runqueue yet. This guarantees that
2491 * nobody will actually run it, and a signal or other external
2492 * event cannot wake it up and insert it on the runqueue either.
2494 p->state = TASK_RUNNING;
2498 * fork()/clone()-time setup:
2500 void sched_fork(struct task_struct *p, int clone_flags)
2502 int cpu = get_cpu();
2507 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2509 set_task_cpu(p, cpu);
2512 * Make sure we do not leak PI boosting priority to the child:
2514 p->prio = current->normal_prio;
2515 if (!rt_prio(p->prio))
2516 p->sched_class = &fair_sched_class;
2518 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2519 if (likely(sched_info_on()))
2520 memset(&p->sched_info, 0, sizeof(p->sched_info));
2522 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2525 #ifdef CONFIG_PREEMPT
2526 /* Want to start with kernel preemption disabled. */
2527 task_thread_info(p)->preempt_count = 1;
2529 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2535 * wake_up_new_task - wake up a newly created task for the first time.
2537 * This function will do some initial scheduler statistics housekeeping
2538 * that must be done for every newly created context, then puts the task
2539 * on the runqueue and wakes it.
2541 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2543 unsigned long flags;
2546 rq = task_rq_lock(p, &flags);
2547 BUG_ON(p->state != TASK_RUNNING);
2548 update_rq_clock(rq);
2550 p->prio = effective_prio(p);
2552 if (!p->sched_class->task_new || !current->se.on_rq) {
2553 activate_task(rq, p, 0);
2556 * Let the scheduling class do new task startup
2557 * management (if any):
2559 p->sched_class->task_new(rq, p);
2562 trace_sched_wakeup_new(rq, p, 1);
2563 check_preempt_curr(rq, p, 0);
2565 if (p->sched_class->task_wake_up)
2566 p->sched_class->task_wake_up(rq, p);
2568 task_rq_unlock(rq, &flags);
2571 #ifdef CONFIG_PREEMPT_NOTIFIERS
2574 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2575 * @notifier: notifier struct to register
2577 void preempt_notifier_register(struct preempt_notifier *notifier)
2579 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2581 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2584 * preempt_notifier_unregister - no longer interested in preemption notifications
2585 * @notifier: notifier struct to unregister
2587 * This is safe to call from within a preemption notifier.
2589 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2591 hlist_del(¬ifier->link);
2593 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2595 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2597 struct preempt_notifier *notifier;
2598 struct hlist_node *node;
2600 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2601 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2605 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2606 struct task_struct *next)
2608 struct preempt_notifier *notifier;
2609 struct hlist_node *node;
2611 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2612 notifier->ops->sched_out(notifier, next);
2615 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2617 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2622 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2623 struct task_struct *next)
2627 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2630 * prepare_task_switch - prepare to switch tasks
2631 * @rq: the runqueue preparing to switch
2632 * @prev: the current task that is being switched out
2633 * @next: the task we are going to switch to.
2635 * This is called with the rq lock held and interrupts off. It must
2636 * be paired with a subsequent finish_task_switch after the context
2639 * prepare_task_switch sets up locking and calls architecture specific
2643 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2644 struct task_struct *next)
2646 fire_sched_out_preempt_notifiers(prev, next);
2647 prepare_lock_switch(rq, next);
2648 prepare_arch_switch(next);
2652 * finish_task_switch - clean up after a task-switch
2653 * @rq: runqueue associated with task-switch
2654 * @prev: the thread we just switched away from.
2656 * finish_task_switch must be called after the context switch, paired
2657 * with a prepare_task_switch call before the context switch.
2658 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2659 * and do any other architecture-specific cleanup actions.
2661 * Note that we may have delayed dropping an mm in context_switch(). If
2662 * so, we finish that here outside of the runqueue lock. (Doing it
2663 * with the lock held can cause deadlocks; see schedule() for
2666 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2667 __releases(rq->lock)
2669 struct mm_struct *mm = rq->prev_mm;
2672 int post_schedule = 0;
2674 if (current->sched_class->needs_post_schedule)
2675 post_schedule = current->sched_class->needs_post_schedule(rq);
2681 * A task struct has one reference for the use as "current".
2682 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2683 * schedule one last time. The schedule call will never return, and
2684 * the scheduled task must drop that reference.
2685 * The test for TASK_DEAD must occur while the runqueue locks are
2686 * still held, otherwise prev could be scheduled on another cpu, die
2687 * there before we look at prev->state, and then the reference would
2689 * Manfred Spraul <manfred@colorfullife.com>
2691 prev_state = prev->state;
2692 finish_arch_switch(prev);
2693 finish_lock_switch(rq, prev);
2696 current->sched_class->post_schedule(rq);
2699 fire_sched_in_preempt_notifiers(current);
2702 if (unlikely(prev_state == TASK_DEAD)) {
2704 * Remove function-return probe instances associated with this
2705 * task and put them back on the free list.
2707 kprobe_flush_task(prev);
2708 put_task_struct(prev);
2713 * schedule_tail - first thing a freshly forked thread must call.
2714 * @prev: the thread we just switched away from.
2716 asmlinkage void schedule_tail(struct task_struct *prev)
2717 __releases(rq->lock)
2719 struct rq *rq = this_rq();
2721 finish_task_switch(rq, prev);
2722 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2723 /* In this case, finish_task_switch does not reenable preemption */
2726 if (current->set_child_tid)
2727 put_user(task_pid_vnr(current), current->set_child_tid);
2731 * context_switch - switch to the new MM and the new
2732 * thread's register state.
2735 context_switch(struct rq *rq, struct task_struct *prev,
2736 struct task_struct *next)
2738 struct mm_struct *mm, *oldmm;
2740 prepare_task_switch(rq, prev, next);
2741 trace_sched_switch(rq, prev, next);
2743 oldmm = prev->active_mm;
2745 * For paravirt, this is coupled with an exit in switch_to to
2746 * combine the page table reload and the switch backend into
2749 arch_enter_lazy_cpu_mode();
2751 if (unlikely(!mm)) {
2752 next->active_mm = oldmm;
2753 atomic_inc(&oldmm->mm_count);
2754 enter_lazy_tlb(oldmm, next);
2756 switch_mm(oldmm, mm, next);
2758 if (unlikely(!prev->mm)) {
2759 prev->active_mm = NULL;
2760 rq->prev_mm = oldmm;
2763 * Since the runqueue lock will be released by the next
2764 * task (which is an invalid locking op but in the case
2765 * of the scheduler it's an obvious special-case), so we
2766 * do an early lockdep release here:
2768 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2769 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2772 /* Here we just switch the register state and the stack. */
2773 switch_to(prev, next, prev);
2777 * this_rq must be evaluated again because prev may have moved
2778 * CPUs since it called schedule(), thus the 'rq' on its stack
2779 * frame will be invalid.
2781 finish_task_switch(this_rq(), prev);
2785 * nr_running, nr_uninterruptible and nr_context_switches:
2787 * externally visible scheduler statistics: current number of runnable
2788 * threads, current number of uninterruptible-sleeping threads, total
2789 * number of context switches performed since bootup.
2791 unsigned long nr_running(void)
2793 unsigned long i, sum = 0;
2795 for_each_online_cpu(i)
2796 sum += cpu_rq(i)->nr_running;
2801 unsigned long nr_uninterruptible(void)
2803 unsigned long i, sum = 0;
2805 for_each_possible_cpu(i)
2806 sum += cpu_rq(i)->nr_uninterruptible;
2809 * Since we read the counters lockless, it might be slightly
2810 * inaccurate. Do not allow it to go below zero though:
2812 if (unlikely((long)sum < 0))
2818 unsigned long long nr_context_switches(void)
2821 unsigned long long sum = 0;
2823 for_each_possible_cpu(i)
2824 sum += cpu_rq(i)->nr_switches;
2829 unsigned long nr_iowait(void)
2831 unsigned long i, sum = 0;
2833 for_each_possible_cpu(i)
2834 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2839 unsigned long nr_active(void)
2841 unsigned long i, running = 0, uninterruptible = 0;
2843 for_each_online_cpu(i) {
2844 running += cpu_rq(i)->nr_running;
2845 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2848 if (unlikely((long)uninterruptible < 0))
2849 uninterruptible = 0;
2851 return running + uninterruptible;
2855 * Update rq->cpu_load[] statistics. This function is usually called every
2856 * scheduler tick (TICK_NSEC).
2858 static void update_cpu_load(struct rq *this_rq)
2860 unsigned long this_load = this_rq->load.weight;
2863 this_rq->nr_load_updates++;
2865 /* Update our load: */
2866 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2867 unsigned long old_load, new_load;
2869 /* scale is effectively 1 << i now, and >> i divides by scale */
2871 old_load = this_rq->cpu_load[i];
2872 new_load = this_load;
2874 * Round up the averaging division if load is increasing. This
2875 * prevents us from getting stuck on 9 if the load is 10, for
2878 if (new_load > old_load)
2879 new_load += scale-1;
2880 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2887 * double_rq_lock - safely lock two runqueues
2889 * Note this does not disable interrupts like task_rq_lock,
2890 * you need to do so manually before calling.
2892 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2893 __acquires(rq1->lock)
2894 __acquires(rq2->lock)
2896 BUG_ON(!irqs_disabled());
2898 spin_lock(&rq1->lock);
2899 __acquire(rq2->lock); /* Fake it out ;) */
2902 spin_lock(&rq1->lock);
2903 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2905 spin_lock(&rq2->lock);
2906 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2909 update_rq_clock(rq1);
2910 update_rq_clock(rq2);
2914 * double_rq_unlock - safely unlock two runqueues
2916 * Note this does not restore interrupts like task_rq_unlock,
2917 * you need to do so manually after calling.
2919 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2920 __releases(rq1->lock)
2921 __releases(rq2->lock)
2923 spin_unlock(&rq1->lock);
2925 spin_unlock(&rq2->lock);
2927 __release(rq2->lock);
2931 * If dest_cpu is allowed for this process, migrate the task to it.
2932 * This is accomplished by forcing the cpu_allowed mask to only
2933 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2934 * the cpu_allowed mask is restored.
2936 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2938 struct migration_req req;
2939 unsigned long flags;
2942 rq = task_rq_lock(p, &flags);
2943 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
2944 || unlikely(!cpu_active(dest_cpu)))
2947 /* force the process onto the specified CPU */
2948 if (migrate_task(p, dest_cpu, &req)) {
2949 /* Need to wait for migration thread (might exit: take ref). */
2950 struct task_struct *mt = rq->migration_thread;
2952 get_task_struct(mt);
2953 task_rq_unlock(rq, &flags);
2954 wake_up_process(mt);
2955 put_task_struct(mt);
2956 wait_for_completion(&req.done);
2961 task_rq_unlock(rq, &flags);
2965 * sched_exec - execve() is a valuable balancing opportunity, because at
2966 * this point the task has the smallest effective memory and cache footprint.
2968 void sched_exec(void)
2970 int new_cpu, this_cpu = get_cpu();
2971 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2973 if (new_cpu != this_cpu)
2974 sched_migrate_task(current, new_cpu);
2978 * pull_task - move a task from a remote runqueue to the local runqueue.
2979 * Both runqueues must be locked.
2981 static void pull_task(struct rq *src_rq, struct task_struct *p,
2982 struct rq *this_rq, int this_cpu)
2984 deactivate_task(src_rq, p, 0);
2985 set_task_cpu(p, this_cpu);
2986 activate_task(this_rq, p, 0);
2988 * Note that idle threads have a prio of MAX_PRIO, for this test
2989 * to be always true for them.
2991 check_preempt_curr(this_rq, p, 0);
2995 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2998 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2999 struct sched_domain *sd, enum cpu_idle_type idle,
3002 int tsk_cache_hot = 0;
3004 * We do not migrate tasks that are:
3005 * 1) running (obviously), or
3006 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3007 * 3) are cache-hot on their current CPU.
3009 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3010 schedstat_inc(p, se.nr_failed_migrations_affine);
3015 if (task_running(rq, p)) {
3016 schedstat_inc(p, se.nr_failed_migrations_running);
3021 * Aggressive migration if:
3022 * 1) task is cache cold, or
3023 * 2) too many balance attempts have failed.
3026 tsk_cache_hot = task_hot(p, rq->clock, sd);
3027 if (!tsk_cache_hot ||
3028 sd->nr_balance_failed > sd->cache_nice_tries) {
3029 #ifdef CONFIG_SCHEDSTATS
3030 if (tsk_cache_hot) {
3031 schedstat_inc(sd, lb_hot_gained[idle]);
3032 schedstat_inc(p, se.nr_forced_migrations);
3038 if (tsk_cache_hot) {
3039 schedstat_inc(p, se.nr_failed_migrations_hot);
3045 static unsigned long
3046 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3047 unsigned long max_load_move, struct sched_domain *sd,
3048 enum cpu_idle_type idle, int *all_pinned,
3049 int *this_best_prio, struct rq_iterator *iterator)
3051 int loops = 0, pulled = 0, pinned = 0;
3052 struct task_struct *p;
3053 long rem_load_move = max_load_move;
3055 if (max_load_move == 0)
3061 * Start the load-balancing iterator:
3063 p = iterator->start(iterator->arg);
3065 if (!p || loops++ > sysctl_sched_nr_migrate)
3068 if ((p->se.load.weight >> 1) > rem_load_move ||
3069 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3070 p = iterator->next(iterator->arg);
3074 pull_task(busiest, p, this_rq, this_cpu);
3076 rem_load_move -= p->se.load.weight;
3078 #ifdef CONFIG_PREEMPT
3080 * NEWIDLE balancing is a source of latency, so preemptible kernels
3081 * will stop after the first task is pulled to minimize the critical
3084 if (idle == CPU_NEWLY_IDLE)
3089 * We only want to steal up to the prescribed amount of weighted load.
3091 if (rem_load_move > 0) {
3092 if (p->prio < *this_best_prio)
3093 *this_best_prio = p->prio;
3094 p = iterator->next(iterator->arg);
3099 * Right now, this is one of only two places pull_task() is called,
3100 * so we can safely collect pull_task() stats here rather than
3101 * inside pull_task().
3103 schedstat_add(sd, lb_gained[idle], pulled);
3106 *all_pinned = pinned;
3108 return max_load_move - rem_load_move;
3112 * move_tasks tries to move up to max_load_move weighted load from busiest to
3113 * this_rq, as part of a balancing operation within domain "sd".
3114 * Returns 1 if successful and 0 otherwise.
3116 * Called with both runqueues locked.
3118 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3119 unsigned long max_load_move,
3120 struct sched_domain *sd, enum cpu_idle_type idle,
3123 const struct sched_class *class = sched_class_highest;
3124 unsigned long total_load_moved = 0;
3125 int this_best_prio = this_rq->curr->prio;
3129 class->load_balance(this_rq, this_cpu, busiest,
3130 max_load_move - total_load_moved,
3131 sd, idle, all_pinned, &this_best_prio);
3132 class = class->next;
3134 #ifdef CONFIG_PREEMPT
3136 * NEWIDLE balancing is a source of latency, so preemptible
3137 * kernels will stop after the first task is pulled to minimize
3138 * the critical section.
3140 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3143 } while (class && max_load_move > total_load_moved);
3145 return total_load_moved > 0;
3149 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3150 struct sched_domain *sd, enum cpu_idle_type idle,
3151 struct rq_iterator *iterator)
3153 struct task_struct *p = iterator->start(iterator->arg);
3157 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3158 pull_task(busiest, p, this_rq, this_cpu);
3160 * Right now, this is only the second place pull_task()
3161 * is called, so we can safely collect pull_task()
3162 * stats here rather than inside pull_task().
3164 schedstat_inc(sd, lb_gained[idle]);
3168 p = iterator->next(iterator->arg);
3175 * move_one_task tries to move exactly one task from busiest to this_rq, as
3176 * part of active balancing operations within "domain".
3177 * Returns 1 if successful and 0 otherwise.
3179 * Called with both runqueues locked.
3181 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3182 struct sched_domain *sd, enum cpu_idle_type idle)
3184 const struct sched_class *class;
3186 for (class = sched_class_highest; class; class = class->next)
3187 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3192 /********** Helpers for find_busiest_group ************************/
3194 * sd_lb_stats - Structure to store the statistics of a sched_domain
3195 * during load balancing.
3197 struct sd_lb_stats {
3198 struct sched_group *busiest; /* Busiest group in this sd */
3199 struct sched_group *this; /* Local group in this sd */
3200 unsigned long total_load; /* Total load of all groups in sd */
3201 unsigned long total_pwr; /* Total power of all groups in sd */
3202 unsigned long avg_load; /* Average load across all groups in sd */
3204 /** Statistics of this group */
3205 unsigned long this_load;
3206 unsigned long this_load_per_task;
3207 unsigned long this_nr_running;
3209 /* Statistics of the busiest group */
3210 unsigned long max_load;
3211 unsigned long busiest_load_per_task;
3212 unsigned long busiest_nr_running;
3214 int group_imb; /* Is there imbalance in this sd */
3215 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3216 int power_savings_balance; /* Is powersave balance needed for this sd */
3217 struct sched_group *group_min; /* Least loaded group in sd */
3218 struct sched_group *group_leader; /* Group which relieves group_min */
3219 unsigned long min_load_per_task; /* load_per_task in group_min */
3220 unsigned long leader_nr_running; /* Nr running of group_leader */
3221 unsigned long min_nr_running; /* Nr running of group_min */
3226 * sg_lb_stats - stats of a sched_group required for load_balancing
3228 struct sg_lb_stats {
3229 unsigned long avg_load; /*Avg load across the CPUs of the group */
3230 unsigned long group_load; /* Total load over the CPUs of the group */
3231 unsigned long sum_nr_running; /* Nr tasks running in the group */
3232 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3233 unsigned long group_capacity;
3234 int group_imb; /* Is there an imbalance in the group ? */
3238 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3239 * @group: The group whose first cpu is to be returned.
3241 static inline unsigned int group_first_cpu(struct sched_group *group)
3243 return cpumask_first(sched_group_cpus(group));
3247 * get_sd_load_idx - Obtain the load index for a given sched domain.
3248 * @sd: The sched_domain whose load_idx is to be obtained.
3249 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3251 static inline int get_sd_load_idx(struct sched_domain *sd,
3252 enum cpu_idle_type idle)
3258 load_idx = sd->busy_idx;
3261 case CPU_NEWLY_IDLE:
3262 load_idx = sd->newidle_idx;
3265 load_idx = sd->idle_idx;
3273 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3275 * init_sd_power_savings_stats - Initialize power savings statistics for
3276 * the given sched_domain, during load balancing.
3278 * @sd: Sched domain whose power-savings statistics are to be initialized.
3279 * @sds: Variable containing the statistics for sd.
3280 * @idle: Idle status of the CPU at which we're performing load-balancing.
3282 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3283 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3286 * Busy processors will not participate in power savings
3289 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3290 sds->power_savings_balance = 0;
3292 sds->power_savings_balance = 1;
3293 sds->min_nr_running = ULONG_MAX;
3294 sds->leader_nr_running = 0;
3299 * update_sd_power_savings_stats - Update the power saving stats for a
3300 * sched_domain while performing load balancing.
3302 * @group: sched_group belonging to the sched_domain under consideration.
3303 * @sds: Variable containing the statistics of the sched_domain
3304 * @local_group: Does group contain the CPU for which we're performing
3306 * @sgs: Variable containing the statistics of the group.
3308 static inline void update_sd_power_savings_stats(struct sched_group *group,
3309 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3312 if (!sds->power_savings_balance)
3316 * If the local group is idle or completely loaded
3317 * no need to do power savings balance at this domain
3319 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3320 !sds->this_nr_running))
3321 sds->power_savings_balance = 0;
3324 * If a group is already running at full capacity or idle,
3325 * don't include that group in power savings calculations
3327 if (!sds->power_savings_balance ||
3328 sgs->sum_nr_running >= sgs->group_capacity ||
3329 !sgs->sum_nr_running)
3333 * Calculate the group which has the least non-idle load.
3334 * This is the group from where we need to pick up the load
3337 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3338 (sgs->sum_nr_running == sds->min_nr_running &&
3339 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3340 sds->group_min = group;
3341 sds->min_nr_running = sgs->sum_nr_running;
3342 sds->min_load_per_task = sgs->sum_weighted_load /
3343 sgs->sum_nr_running;
3347 * Calculate the group which is almost near its
3348 * capacity but still has some space to pick up some load
3349 * from other group and save more power
3351 if (sgs->sum_nr_running > sgs->group_capacity - 1)
3354 if (sgs->sum_nr_running > sds->leader_nr_running ||
3355 (sgs->sum_nr_running == sds->leader_nr_running &&
3356 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3357 sds->group_leader = group;
3358 sds->leader_nr_running = sgs->sum_nr_running;
3363 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3364 * @sds: Variable containing the statistics of the sched_domain
3365 * under consideration.
3366 * @this_cpu: Cpu at which we're currently performing load-balancing.
3367 * @imbalance: Variable to store the imbalance.
3370 * Check if we have potential to perform some power-savings balance.
3371 * If yes, set the busiest group to be the least loaded group in the
3372 * sched_domain, so that it's CPUs can be put to idle.
3374 * Returns 1 if there is potential to perform power-savings balance.
3377 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3378 int this_cpu, unsigned long *imbalance)
3380 if (!sds->power_savings_balance)
3383 if (sds->this != sds->group_leader ||
3384 sds->group_leader == sds->group_min)
3387 *imbalance = sds->min_load_per_task;
3388 sds->busiest = sds->group_min;
3390 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP) {
3391 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu =
3392 group_first_cpu(sds->group_leader);
3398 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3399 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3400 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3405 static inline void update_sd_power_savings_stats(struct sched_group *group,
3406 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3411 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3412 int this_cpu, unsigned long *imbalance)
3416 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3420 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3421 * @group: sched_group whose statistics are to be updated.
3422 * @this_cpu: Cpu for which load balance is currently performed.
3423 * @idle: Idle status of this_cpu
3424 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3425 * @sd_idle: Idle status of the sched_domain containing group.
3426 * @local_group: Does group contain this_cpu.
3427 * @cpus: Set of cpus considered for load balancing.
3428 * @balance: Should we balance.
3429 * @sgs: variable to hold the statistics for this group.
3431 static inline void update_sg_lb_stats(struct sched_group *group, int this_cpu,
3432 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3433 int local_group, const struct cpumask *cpus,
3434 int *balance, struct sg_lb_stats *sgs)
3436 unsigned long load, max_cpu_load, min_cpu_load;
3438 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3439 unsigned long sum_avg_load_per_task;
3440 unsigned long avg_load_per_task;
3443 balance_cpu = group_first_cpu(group);
3445 /* Tally up the load of all CPUs in the group */
3446 sum_avg_load_per_task = avg_load_per_task = 0;
3448 min_cpu_load = ~0UL;
3450 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3451 struct rq *rq = cpu_rq(i);
3453 if (*sd_idle && rq->nr_running)
3456 /* Bias balancing toward cpus of our domain */
3458 if (idle_cpu(i) && !first_idle_cpu) {
3463 load = target_load(i, load_idx);
3465 load = source_load(i, load_idx);
3466 if (load > max_cpu_load)
3467 max_cpu_load = load;
3468 if (min_cpu_load > load)
3469 min_cpu_load = load;
3472 sgs->group_load += load;
3473 sgs->sum_nr_running += rq->nr_running;
3474 sgs->sum_weighted_load += weighted_cpuload(i);
3476 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3480 * First idle cpu or the first cpu(busiest) in this sched group
3481 * is eligible for doing load balancing at this and above
3482 * domains. In the newly idle case, we will allow all the cpu's
3483 * to do the newly idle load balance.
3485 if (idle != CPU_NEWLY_IDLE && local_group &&
3486 balance_cpu != this_cpu && balance) {
3491 /* Adjust by relative CPU power of the group */
3492 sgs->avg_load = sg_div_cpu_power(group,
3493 sgs->group_load * SCHED_LOAD_SCALE);
3497 * Consider the group unbalanced when the imbalance is larger
3498 * than the average weight of two tasks.
3500 * APZ: with cgroup the avg task weight can vary wildly and
3501 * might not be a suitable number - should we keep a
3502 * normalized nr_running number somewhere that negates
3505 avg_load_per_task = sg_div_cpu_power(group,
3506 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3508 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3511 sgs->group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3516 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3517 * @sd: sched_domain whose statistics are to be updated.
3518 * @this_cpu: Cpu for which load balance is currently performed.
3519 * @idle: Idle status of this_cpu
3520 * @sd_idle: Idle status of the sched_domain containing group.
3521 * @cpus: Set of cpus considered for load balancing.
3522 * @balance: Should we balance.
3523 * @sds: variable to hold the statistics for this sched_domain.
3525 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3526 enum cpu_idle_type idle, int *sd_idle,
3527 const struct cpumask *cpus, int *balance,
3528 struct sd_lb_stats *sds)
3530 struct sched_group *group = sd->groups;
3531 struct sg_lb_stats sgs;
3534 init_sd_power_savings_stats(sd, sds, idle);
3535 load_idx = get_sd_load_idx(sd, idle);
3540 local_group = cpumask_test_cpu(this_cpu,
3541 sched_group_cpus(group));
3542 memset(&sgs, 0, sizeof(sgs));
3543 update_sg_lb_stats(group, this_cpu, idle, load_idx, sd_idle,
3544 local_group, cpus, balance, &sgs);
3546 if (local_group && balance && !(*balance))
3549 sds->total_load += sgs.group_load;
3550 sds->total_pwr += group->__cpu_power;
3553 sds->this_load = sgs.avg_load;
3555 sds->this_nr_running = sgs.sum_nr_running;
3556 sds->this_load_per_task = sgs.sum_weighted_load;
3557 } else if (sgs.avg_load > sds->max_load &&
3558 (sgs.sum_nr_running > sgs.group_capacity ||
3560 sds->max_load = sgs.avg_load;
3561 sds->busiest = group;
3562 sds->busiest_nr_running = sgs.sum_nr_running;
3563 sds->busiest_load_per_task = sgs.sum_weighted_load;
3564 sds->group_imb = sgs.group_imb;
3567 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3568 group = group->next;
3569 } while (group != sd->groups);
3574 * fix_small_imbalance - Calculate the minor imbalance that exists
3575 * amongst the groups of a sched_domain, during
3577 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3578 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3579 * @imbalance: Variable to store the imbalance.
3581 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3582 int this_cpu, unsigned long *imbalance)
3584 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3585 unsigned int imbn = 2;
3587 if (sds->this_nr_running) {
3588 sds->this_load_per_task /= sds->this_nr_running;
3589 if (sds->busiest_load_per_task >
3590 sds->this_load_per_task)
3593 sds->this_load_per_task =
3594 cpu_avg_load_per_task(this_cpu);
3596 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3597 sds->busiest_load_per_task * imbn) {
3598 *imbalance = sds->busiest_load_per_task;
3603 * OK, we don't have enough imbalance to justify moving tasks,
3604 * however we may be able to increase total CPU power used by
3608 pwr_now += sds->busiest->__cpu_power *
3609 min(sds->busiest_load_per_task, sds->max_load);
3610 pwr_now += sds->this->__cpu_power *
3611 min(sds->this_load_per_task, sds->this_load);
3612 pwr_now /= SCHED_LOAD_SCALE;
3614 /* Amount of load we'd subtract */
3615 tmp = sg_div_cpu_power(sds->busiest,
3616 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3617 if (sds->max_load > tmp)
3618 pwr_move += sds->busiest->__cpu_power *
3619 min(sds->busiest_load_per_task, sds->max_load - tmp);
3621 /* Amount of load we'd add */
3622 if (sds->max_load * sds->busiest->__cpu_power <
3623 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3624 tmp = sg_div_cpu_power(sds->this,
3625 sds->max_load * sds->busiest->__cpu_power);
3627 tmp = sg_div_cpu_power(sds->this,
3628 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3629 pwr_move += sds->this->__cpu_power *
3630 min(sds->this_load_per_task, sds->this_load + tmp);
3631 pwr_move /= SCHED_LOAD_SCALE;
3633 /* Move if we gain throughput */
3634 if (pwr_move > pwr_now)
3635 *imbalance = sds->busiest_load_per_task;
3639 * calculate_imbalance - Calculate the amount of imbalance present within the
3640 * groups of a given sched_domain during load balance.
3641 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3642 * @this_cpu: Cpu for which currently load balance is being performed.
3643 * @imbalance: The variable to store the imbalance.
3645 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3646 unsigned long *imbalance)
3648 unsigned long max_pull;
3650 * In the presence of smp nice balancing, certain scenarios can have
3651 * max load less than avg load(as we skip the groups at or below
3652 * its cpu_power, while calculating max_load..)
3654 if (sds->max_load < sds->avg_load) {
3656 return fix_small_imbalance(sds, this_cpu, imbalance);
3659 /* Don't want to pull so many tasks that a group would go idle */
3660 max_pull = min(sds->max_load - sds->avg_load,
3661 sds->max_load - sds->busiest_load_per_task);
3663 /* How much load to actually move to equalise the imbalance */
3664 *imbalance = min(max_pull * sds->busiest->__cpu_power,
3665 (sds->avg_load - sds->this_load) * sds->this->__cpu_power)
3669 * if *imbalance is less than the average load per runnable task
3670 * there is no gaurantee that any tasks will be moved so we'll have
3671 * a think about bumping its value to force at least one task to be
3674 if (*imbalance < sds->busiest_load_per_task)
3675 return fix_small_imbalance(sds, this_cpu, imbalance);
3678 /******* find_busiest_group() helpers end here *********************/
3681 * find_busiest_group - Returns the busiest group within the sched_domain
3682 * if there is an imbalance. If there isn't an imbalance, and
3683 * the user has opted for power-savings, it returns a group whose
3684 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3685 * such a group exists.
3687 * Also calculates the amount of weighted load which should be moved
3688 * to restore balance.
3690 * @sd: The sched_domain whose busiest group is to be returned.
3691 * @this_cpu: The cpu for which load balancing is currently being performed.
3692 * @imbalance: Variable which stores amount of weighted load which should
3693 * be moved to restore balance/put a group to idle.
3694 * @idle: The idle status of this_cpu.
3695 * @sd_idle: The idleness of sd
3696 * @cpus: The set of CPUs under consideration for load-balancing.
3697 * @balance: Pointer to a variable indicating if this_cpu
3698 * is the appropriate cpu to perform load balancing at this_level.
3700 * Returns: - the busiest group if imbalance exists.
3701 * - If no imbalance and user has opted for power-savings balance,
3702 * return the least loaded group whose CPUs can be
3703 * put to idle by rebalancing its tasks onto our group.
3705 static struct sched_group *
3706 find_busiest_group(struct sched_domain *sd, int this_cpu,
3707 unsigned long *imbalance, enum cpu_idle_type idle,
3708 int *sd_idle, const struct cpumask *cpus, int *balance)
3710 struct sd_lb_stats sds;
3712 memset(&sds, 0, sizeof(sds));
3715 * Compute the various statistics relavent for load balancing at
3718 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
3721 /* Cases where imbalance does not exist from POV of this_cpu */
3722 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3724 * 2) There is no busy sibling group to pull from.
3725 * 3) This group is the busiest group.
3726 * 4) This group is more busy than the avg busieness at this
3728 * 5) The imbalance is within the specified limit.
3729 * 6) Any rebalance would lead to ping-pong
3731 if (balance && !(*balance))
3734 if (!sds.busiest || sds.busiest_nr_running == 0)
3737 if (sds.this_load >= sds.max_load)
3740 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
3742 if (sds.this_load >= sds.avg_load)
3745 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
3748 sds.busiest_load_per_task /= sds.busiest_nr_running;
3750 sds.busiest_load_per_task =
3751 min(sds.busiest_load_per_task, sds.avg_load);
3754 * We're trying to get all the cpus to the average_load, so we don't
3755 * want to push ourselves above the average load, nor do we wish to
3756 * reduce the max loaded cpu below the average load, as either of these
3757 * actions would just result in more rebalancing later, and ping-pong
3758 * tasks around. Thus we look for the minimum possible imbalance.
3759 * Negative imbalances (*we* are more loaded than anyone else) will
3760 * be counted as no imbalance for these purposes -- we can't fix that
3761 * by pulling tasks to us. Be careful of negative numbers as they'll
3762 * appear as very large values with unsigned longs.
3764 if (sds.max_load <= sds.busiest_load_per_task)
3767 /* Looks like there is an imbalance. Compute it */
3768 calculate_imbalance(&sds, this_cpu, imbalance);
3773 * There is no obvious imbalance. But check if we can do some balancing
3776 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
3784 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3787 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3788 unsigned long imbalance, const struct cpumask *cpus)
3790 struct rq *busiest = NULL, *rq;
3791 unsigned long max_load = 0;
3794 for_each_cpu(i, sched_group_cpus(group)) {
3797 if (!cpumask_test_cpu(i, cpus))
3801 wl = weighted_cpuload(i);
3803 if (rq->nr_running == 1 && wl > imbalance)
3806 if (wl > max_load) {
3816 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3817 * so long as it is large enough.
3819 #define MAX_PINNED_INTERVAL 512
3821 /* Working cpumask for load_balance and load_balance_newidle. */
3822 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
3825 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3826 * tasks if there is an imbalance.
3828 static int load_balance(int this_cpu, struct rq *this_rq,
3829 struct sched_domain *sd, enum cpu_idle_type idle,
3832 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3833 struct sched_group *group;
3834 unsigned long imbalance;
3836 unsigned long flags;
3837 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
3839 cpumask_setall(cpus);
3842 * When power savings policy is enabled for the parent domain, idle
3843 * sibling can pick up load irrespective of busy siblings. In this case,
3844 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3845 * portraying it as CPU_NOT_IDLE.
3847 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3848 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3851 schedstat_inc(sd, lb_count[idle]);
3855 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3862 schedstat_inc(sd, lb_nobusyg[idle]);
3866 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3868 schedstat_inc(sd, lb_nobusyq[idle]);
3872 BUG_ON(busiest == this_rq);
3874 schedstat_add(sd, lb_imbalance[idle], imbalance);
3877 if (busiest->nr_running > 1) {
3879 * Attempt to move tasks. If find_busiest_group has found
3880 * an imbalance but busiest->nr_running <= 1, the group is
3881 * still unbalanced. ld_moved simply stays zero, so it is
3882 * correctly treated as an imbalance.
3884 local_irq_save(flags);
3885 double_rq_lock(this_rq, busiest);
3886 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3887 imbalance, sd, idle, &all_pinned);
3888 double_rq_unlock(this_rq, busiest);
3889 local_irq_restore(flags);
3892 * some other cpu did the load balance for us.
3894 if (ld_moved && this_cpu != smp_processor_id())
3895 resched_cpu(this_cpu);
3897 /* All tasks on this runqueue were pinned by CPU affinity */
3898 if (unlikely(all_pinned)) {
3899 cpumask_clear_cpu(cpu_of(busiest), cpus);
3900 if (!cpumask_empty(cpus))
3907 schedstat_inc(sd, lb_failed[idle]);
3908 sd->nr_balance_failed++;
3910 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3912 spin_lock_irqsave(&busiest->lock, flags);
3914 /* don't kick the migration_thread, if the curr
3915 * task on busiest cpu can't be moved to this_cpu
3917 if (!cpumask_test_cpu(this_cpu,
3918 &busiest->curr->cpus_allowed)) {
3919 spin_unlock_irqrestore(&busiest->lock, flags);
3921 goto out_one_pinned;
3924 if (!busiest->active_balance) {
3925 busiest->active_balance = 1;
3926 busiest->push_cpu = this_cpu;
3929 spin_unlock_irqrestore(&busiest->lock, flags);
3931 wake_up_process(busiest->migration_thread);
3934 * We've kicked active balancing, reset the failure
3937 sd->nr_balance_failed = sd->cache_nice_tries+1;
3940 sd->nr_balance_failed = 0;
3942 if (likely(!active_balance)) {
3943 /* We were unbalanced, so reset the balancing interval */
3944 sd->balance_interval = sd->min_interval;
3947 * If we've begun active balancing, start to back off. This
3948 * case may not be covered by the all_pinned logic if there
3949 * is only 1 task on the busy runqueue (because we don't call
3952 if (sd->balance_interval < sd->max_interval)
3953 sd->balance_interval *= 2;
3956 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3957 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3963 schedstat_inc(sd, lb_balanced[idle]);
3965 sd->nr_balance_failed = 0;
3968 /* tune up the balancing interval */
3969 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3970 (sd->balance_interval < sd->max_interval))
3971 sd->balance_interval *= 2;
3973 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3974 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3985 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3986 * tasks if there is an imbalance.
3988 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3989 * this_rq is locked.
3992 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
3994 struct sched_group *group;
3995 struct rq *busiest = NULL;
3996 unsigned long imbalance;
4000 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4002 cpumask_setall(cpus);
4005 * When power savings policy is enabled for the parent domain, idle
4006 * sibling can pick up load irrespective of busy siblings. In this case,
4007 * let the state of idle sibling percolate up as IDLE, instead of
4008 * portraying it as CPU_NOT_IDLE.
4010 if (sd->flags & SD_SHARE_CPUPOWER &&
4011 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4014 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4016 update_shares_locked(this_rq, sd);
4017 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4018 &sd_idle, cpus, NULL);
4020 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4024 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4026 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4030 BUG_ON(busiest == this_rq);
4032 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4035 if (busiest->nr_running > 1) {
4036 /* Attempt to move tasks */
4037 double_lock_balance(this_rq, busiest);
4038 /* this_rq->clock is already updated */
4039 update_rq_clock(busiest);
4040 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4041 imbalance, sd, CPU_NEWLY_IDLE,
4043 double_unlock_balance(this_rq, busiest);
4045 if (unlikely(all_pinned)) {
4046 cpumask_clear_cpu(cpu_of(busiest), cpus);
4047 if (!cpumask_empty(cpus))
4053 int active_balance = 0;
4055 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4056 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4057 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4060 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4063 if (sd->nr_balance_failed++ < 2)
4067 * The only task running in a non-idle cpu can be moved to this
4068 * cpu in an attempt to completely freeup the other CPU
4069 * package. The same method used to move task in load_balance()
4070 * have been extended for load_balance_newidle() to speedup
4071 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4073 * The package power saving logic comes from
4074 * find_busiest_group(). If there are no imbalance, then
4075 * f_b_g() will return NULL. However when sched_mc={1,2} then
4076 * f_b_g() will select a group from which a running task may be
4077 * pulled to this cpu in order to make the other package idle.
4078 * If there is no opportunity to make a package idle and if
4079 * there are no imbalance, then f_b_g() will return NULL and no
4080 * action will be taken in load_balance_newidle().
4082 * Under normal task pull operation due to imbalance, there
4083 * will be more than one task in the source run queue and
4084 * move_tasks() will succeed. ld_moved will be true and this
4085 * active balance code will not be triggered.
4088 /* Lock busiest in correct order while this_rq is held */
4089 double_lock_balance(this_rq, busiest);
4092 * don't kick the migration_thread, if the curr
4093 * task on busiest cpu can't be moved to this_cpu
4095 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4096 double_unlock_balance(this_rq, busiest);
4101 if (!busiest->active_balance) {
4102 busiest->active_balance = 1;
4103 busiest->push_cpu = this_cpu;
4107 double_unlock_balance(this_rq, busiest);
4109 * Should not call ttwu while holding a rq->lock
4111 spin_unlock(&this_rq->lock);
4113 wake_up_process(busiest->migration_thread);
4114 spin_lock(&this_rq->lock);
4117 sd->nr_balance_failed = 0;
4119 update_shares_locked(this_rq, sd);
4123 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4124 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4125 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4127 sd->nr_balance_failed = 0;
4133 * idle_balance is called by schedule() if this_cpu is about to become
4134 * idle. Attempts to pull tasks from other CPUs.
4136 static void idle_balance(int this_cpu, struct rq *this_rq)
4138 struct sched_domain *sd;
4139 int pulled_task = 0;
4140 unsigned long next_balance = jiffies + HZ;
4142 for_each_domain(this_cpu, sd) {
4143 unsigned long interval;
4145 if (!(sd->flags & SD_LOAD_BALANCE))
4148 if (sd->flags & SD_BALANCE_NEWIDLE)
4149 /* If we've pulled tasks over stop searching: */
4150 pulled_task = load_balance_newidle(this_cpu, this_rq,
4153 interval = msecs_to_jiffies(sd->balance_interval);
4154 if (time_after(next_balance, sd->last_balance + interval))
4155 next_balance = sd->last_balance + interval;
4159 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4161 * We are going idle. next_balance may be set based on
4162 * a busy processor. So reset next_balance.
4164 this_rq->next_balance = next_balance;
4169 * active_load_balance is run by migration threads. It pushes running tasks
4170 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4171 * running on each physical CPU where possible, and avoids physical /
4172 * logical imbalances.
4174 * Called with busiest_rq locked.
4176 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4178 int target_cpu = busiest_rq->push_cpu;
4179 struct sched_domain *sd;
4180 struct rq *target_rq;
4182 /* Is there any task to move? */
4183 if (busiest_rq->nr_running <= 1)
4186 target_rq = cpu_rq(target_cpu);
4189 * This condition is "impossible", if it occurs
4190 * we need to fix it. Originally reported by
4191 * Bjorn Helgaas on a 128-cpu setup.
4193 BUG_ON(busiest_rq == target_rq);
4195 /* move a task from busiest_rq to target_rq */
4196 double_lock_balance(busiest_rq, target_rq);
4197 update_rq_clock(busiest_rq);
4198 update_rq_clock(target_rq);
4200 /* Search for an sd spanning us and the target CPU. */
4201 for_each_domain(target_cpu, sd) {
4202 if ((sd->flags & SD_LOAD_BALANCE) &&
4203 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4208 schedstat_inc(sd, alb_count);
4210 if (move_one_task(target_rq, target_cpu, busiest_rq,
4212 schedstat_inc(sd, alb_pushed);
4214 schedstat_inc(sd, alb_failed);
4216 double_unlock_balance(busiest_rq, target_rq);
4221 atomic_t load_balancer;
4222 cpumask_var_t cpu_mask;
4223 } nohz ____cacheline_aligned = {
4224 .load_balancer = ATOMIC_INIT(-1),
4228 * This routine will try to nominate the ilb (idle load balancing)
4229 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4230 * load balancing on behalf of all those cpus. If all the cpus in the system
4231 * go into this tickless mode, then there will be no ilb owner (as there is
4232 * no need for one) and all the cpus will sleep till the next wakeup event
4235 * For the ilb owner, tick is not stopped. And this tick will be used
4236 * for idle load balancing. ilb owner will still be part of
4239 * While stopping the tick, this cpu will become the ilb owner if there
4240 * is no other owner. And will be the owner till that cpu becomes busy
4241 * or if all cpus in the system stop their ticks at which point
4242 * there is no need for ilb owner.
4244 * When the ilb owner becomes busy, it nominates another owner, during the
4245 * next busy scheduler_tick()
4247 int select_nohz_load_balancer(int stop_tick)
4249 int cpu = smp_processor_id();
4252 cpu_rq(cpu)->in_nohz_recently = 1;
4254 if (!cpu_active(cpu)) {
4255 if (atomic_read(&nohz.load_balancer) != cpu)
4259 * If we are going offline and still the leader,
4262 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4268 cpumask_set_cpu(cpu, nohz.cpu_mask);
4270 /* time for ilb owner also to sleep */
4271 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4272 if (atomic_read(&nohz.load_balancer) == cpu)
4273 atomic_set(&nohz.load_balancer, -1);
4277 if (atomic_read(&nohz.load_balancer) == -1) {
4278 /* make me the ilb owner */
4279 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4281 } else if (atomic_read(&nohz.load_balancer) == cpu)
4284 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4287 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4289 if (atomic_read(&nohz.load_balancer) == cpu)
4290 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4297 static DEFINE_SPINLOCK(balancing);
4300 * It checks each scheduling domain to see if it is due to be balanced,
4301 * and initiates a balancing operation if so.
4303 * Balancing parameters are set up in arch_init_sched_domains.
4305 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4308 struct rq *rq = cpu_rq(cpu);
4309 unsigned long interval;
4310 struct sched_domain *sd;
4311 /* Earliest time when we have to do rebalance again */
4312 unsigned long next_balance = jiffies + 60*HZ;
4313 int update_next_balance = 0;
4316 for_each_domain(cpu, sd) {
4317 if (!(sd->flags & SD_LOAD_BALANCE))
4320 interval = sd->balance_interval;
4321 if (idle != CPU_IDLE)
4322 interval *= sd->busy_factor;
4324 /* scale ms to jiffies */
4325 interval = msecs_to_jiffies(interval);
4326 if (unlikely(!interval))
4328 if (interval > HZ*NR_CPUS/10)
4329 interval = HZ*NR_CPUS/10;
4331 need_serialize = sd->flags & SD_SERIALIZE;
4333 if (need_serialize) {
4334 if (!spin_trylock(&balancing))
4338 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4339 if (load_balance(cpu, rq, sd, idle, &balance)) {
4341 * We've pulled tasks over so either we're no
4342 * longer idle, or one of our SMT siblings is
4345 idle = CPU_NOT_IDLE;
4347 sd->last_balance = jiffies;
4350 spin_unlock(&balancing);
4352 if (time_after(next_balance, sd->last_balance + interval)) {
4353 next_balance = sd->last_balance + interval;
4354 update_next_balance = 1;
4358 * Stop the load balance at this level. There is another
4359 * CPU in our sched group which is doing load balancing more
4367 * next_balance will be updated only when there is a need.
4368 * When the cpu is attached to null domain for ex, it will not be
4371 if (likely(update_next_balance))
4372 rq->next_balance = next_balance;
4376 * run_rebalance_domains is triggered when needed from the scheduler tick.
4377 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4378 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4380 static void run_rebalance_domains(struct softirq_action *h)
4382 int this_cpu = smp_processor_id();
4383 struct rq *this_rq = cpu_rq(this_cpu);
4384 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4385 CPU_IDLE : CPU_NOT_IDLE;
4387 rebalance_domains(this_cpu, idle);
4391 * If this cpu is the owner for idle load balancing, then do the
4392 * balancing on behalf of the other idle cpus whose ticks are
4395 if (this_rq->idle_at_tick &&
4396 atomic_read(&nohz.load_balancer) == this_cpu) {
4400 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4401 if (balance_cpu == this_cpu)
4405 * If this cpu gets work to do, stop the load balancing
4406 * work being done for other cpus. Next load
4407 * balancing owner will pick it up.
4412 rebalance_domains(balance_cpu, CPU_IDLE);
4414 rq = cpu_rq(balance_cpu);
4415 if (time_after(this_rq->next_balance, rq->next_balance))
4416 this_rq->next_balance = rq->next_balance;
4422 static inline int on_null_domain(int cpu)
4424 return !rcu_dereference(cpu_rq(cpu)->sd);
4428 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4430 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4431 * idle load balancing owner or decide to stop the periodic load balancing,
4432 * if the whole system is idle.
4434 static inline void trigger_load_balance(struct rq *rq, int cpu)
4438 * If we were in the nohz mode recently and busy at the current
4439 * scheduler tick, then check if we need to nominate new idle
4442 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4443 rq->in_nohz_recently = 0;
4445 if (atomic_read(&nohz.load_balancer) == cpu) {
4446 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4447 atomic_set(&nohz.load_balancer, -1);
4450 if (atomic_read(&nohz.load_balancer) == -1) {
4452 * simple selection for now: Nominate the
4453 * first cpu in the nohz list to be the next
4456 * TBD: Traverse the sched domains and nominate
4457 * the nearest cpu in the nohz.cpu_mask.
4459 int ilb = cpumask_first(nohz.cpu_mask);
4461 if (ilb < nr_cpu_ids)
4467 * If this cpu is idle and doing idle load balancing for all the
4468 * cpus with ticks stopped, is it time for that to stop?
4470 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4471 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4477 * If this cpu is idle and the idle load balancing is done by
4478 * someone else, then no need raise the SCHED_SOFTIRQ
4480 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4481 cpumask_test_cpu(cpu, nohz.cpu_mask))
4484 /* Don't need to rebalance while attached to NULL domain */
4485 if (time_after_eq(jiffies, rq->next_balance) &&
4486 likely(!on_null_domain(cpu)))
4487 raise_softirq(SCHED_SOFTIRQ);
4490 #else /* CONFIG_SMP */
4493 * on UP we do not need to balance between CPUs:
4495 static inline void idle_balance(int cpu, struct rq *rq)
4501 DEFINE_PER_CPU(struct kernel_stat, kstat);
4503 EXPORT_PER_CPU_SYMBOL(kstat);
4506 * Return any ns on the sched_clock that have not yet been banked in
4507 * @p in case that task is currently running.
4509 unsigned long long task_delta_exec(struct task_struct *p)
4511 unsigned long flags;
4515 rq = task_rq_lock(p, &flags);
4517 if (task_current(rq, p)) {
4520 update_rq_clock(rq);
4521 delta_exec = rq->clock - p->se.exec_start;
4522 if ((s64)delta_exec > 0)
4526 task_rq_unlock(rq, &flags);
4532 * Account user cpu time to a process.
4533 * @p: the process that the cpu time gets accounted to
4534 * @cputime: the cpu time spent in user space since the last update
4535 * @cputime_scaled: cputime scaled by cpu frequency
4537 void account_user_time(struct task_struct *p, cputime_t cputime,
4538 cputime_t cputime_scaled)
4540 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4543 /* Add user time to process. */
4544 p->utime = cputime_add(p->utime, cputime);
4545 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4546 account_group_user_time(p, cputime);
4548 /* Add user time to cpustat. */
4549 tmp = cputime_to_cputime64(cputime);
4550 if (TASK_NICE(p) > 0)
4551 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4553 cpustat->user = cputime64_add(cpustat->user, tmp);
4554 /* Account for user time used */
4555 acct_update_integrals(p);
4559 * Account guest cpu time to a process.
4560 * @p: the process that the cpu time gets accounted to
4561 * @cputime: the cpu time spent in virtual machine since the last update
4562 * @cputime_scaled: cputime scaled by cpu frequency
4564 static void account_guest_time(struct task_struct *p, cputime_t cputime,
4565 cputime_t cputime_scaled)
4568 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4570 tmp = cputime_to_cputime64(cputime);
4572 /* Add guest time to process. */
4573 p->utime = cputime_add(p->utime, cputime);
4574 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4575 account_group_user_time(p, cputime);
4576 p->gtime = cputime_add(p->gtime, cputime);
4578 /* Add guest time to cpustat. */
4579 cpustat->user = cputime64_add(cpustat->user, tmp);
4580 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4584 * Account system cpu time to a process.
4585 * @p: the process that the cpu time gets accounted to
4586 * @hardirq_offset: the offset to subtract from hardirq_count()
4587 * @cputime: the cpu time spent in kernel space since the last update
4588 * @cputime_scaled: cputime scaled by cpu frequency
4590 void account_system_time(struct task_struct *p, int hardirq_offset,
4591 cputime_t cputime, cputime_t cputime_scaled)
4593 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4596 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4597 account_guest_time(p, cputime, cputime_scaled);
4601 /* Add system time to process. */
4602 p->stime = cputime_add(p->stime, cputime);
4603 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
4604 account_group_system_time(p, cputime);
4606 /* Add system time to cpustat. */
4607 tmp = cputime_to_cputime64(cputime);
4608 if (hardirq_count() - hardirq_offset)
4609 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4610 else if (softirq_count())
4611 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4613 cpustat->system = cputime64_add(cpustat->system, tmp);
4615 /* Account for system time used */
4616 acct_update_integrals(p);
4620 * Account for involuntary wait time.
4621 * @steal: the cpu time spent in involuntary wait
4623 void account_steal_time(cputime_t cputime)
4625 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4626 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4628 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
4632 * Account for idle time.
4633 * @cputime: the cpu time spent in idle wait
4635 void account_idle_time(cputime_t cputime)
4637 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4638 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4639 struct rq *rq = this_rq();
4641 if (atomic_read(&rq->nr_iowait) > 0)
4642 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
4644 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
4647 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4650 * Account a single tick of cpu time.
4651 * @p: the process that the cpu time gets accounted to
4652 * @user_tick: indicates if the tick is a user or a system tick
4654 void account_process_tick(struct task_struct *p, int user_tick)
4656 cputime_t one_jiffy = jiffies_to_cputime(1);
4657 cputime_t one_jiffy_scaled = cputime_to_scaled(one_jiffy);
4658 struct rq *rq = this_rq();
4661 account_user_time(p, one_jiffy, one_jiffy_scaled);
4662 else if (p != rq->idle)
4663 account_system_time(p, HARDIRQ_OFFSET, one_jiffy,
4666 account_idle_time(one_jiffy);
4670 * Account multiple ticks of steal time.
4671 * @p: the process from which the cpu time has been stolen
4672 * @ticks: number of stolen ticks
4674 void account_steal_ticks(unsigned long ticks)
4676 account_steal_time(jiffies_to_cputime(ticks));
4680 * Account multiple ticks of idle time.
4681 * @ticks: number of stolen ticks
4683 void account_idle_ticks(unsigned long ticks)
4685 account_idle_time(jiffies_to_cputime(ticks));
4691 * Use precise platform statistics if available:
4693 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4694 cputime_t task_utime(struct task_struct *p)
4699 cputime_t task_stime(struct task_struct *p)
4704 cputime_t task_utime(struct task_struct *p)
4706 clock_t utime = cputime_to_clock_t(p->utime),
4707 total = utime + cputime_to_clock_t(p->stime);
4711 * Use CFS's precise accounting:
4713 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
4717 do_div(temp, total);
4719 utime = (clock_t)temp;
4721 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
4722 return p->prev_utime;
4725 cputime_t task_stime(struct task_struct *p)
4730 * Use CFS's precise accounting. (we subtract utime from
4731 * the total, to make sure the total observed by userspace
4732 * grows monotonically - apps rely on that):
4734 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
4735 cputime_to_clock_t(task_utime(p));
4738 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
4740 return p->prev_stime;
4744 inline cputime_t task_gtime(struct task_struct *p)
4750 * This function gets called by the timer code, with HZ frequency.
4751 * We call it with interrupts disabled.
4753 * It also gets called by the fork code, when changing the parent's
4756 void scheduler_tick(void)
4758 int cpu = smp_processor_id();
4759 struct rq *rq = cpu_rq(cpu);
4760 struct task_struct *curr = rq->curr;
4764 spin_lock(&rq->lock);
4765 update_rq_clock(rq);
4766 update_cpu_load(rq);
4767 curr->sched_class->task_tick(rq, curr, 0);
4768 spin_unlock(&rq->lock);
4771 rq->idle_at_tick = idle_cpu(cpu);
4772 trigger_load_balance(rq, cpu);
4776 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4777 defined(CONFIG_PREEMPT_TRACER))
4779 static inline unsigned long get_parent_ip(unsigned long addr)
4781 if (in_lock_functions(addr)) {
4782 addr = CALLER_ADDR2;
4783 if (in_lock_functions(addr))
4784 addr = CALLER_ADDR3;
4789 void __kprobes add_preempt_count(int val)
4791 #ifdef CONFIG_DEBUG_PREEMPT
4795 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4798 preempt_count() += val;
4799 #ifdef CONFIG_DEBUG_PREEMPT
4801 * Spinlock count overflowing soon?
4803 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4806 if (preempt_count() == val)
4807 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4809 EXPORT_SYMBOL(add_preempt_count);
4811 void __kprobes sub_preempt_count(int val)
4813 #ifdef CONFIG_DEBUG_PREEMPT
4817 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4820 * Is the spinlock portion underflowing?
4822 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4823 !(preempt_count() & PREEMPT_MASK)))
4827 if (preempt_count() == val)
4828 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4829 preempt_count() -= val;
4831 EXPORT_SYMBOL(sub_preempt_count);
4836 * Print scheduling while atomic bug:
4838 static noinline void __schedule_bug(struct task_struct *prev)
4840 struct pt_regs *regs = get_irq_regs();
4842 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4843 prev->comm, prev->pid, preempt_count());
4845 debug_show_held_locks(prev);
4847 if (irqs_disabled())
4848 print_irqtrace_events(prev);
4857 * Various schedule()-time debugging checks and statistics:
4859 static inline void schedule_debug(struct task_struct *prev)
4862 * Test if we are atomic. Since do_exit() needs to call into
4863 * schedule() atomically, we ignore that path for now.
4864 * Otherwise, whine if we are scheduling when we should not be.
4866 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4867 __schedule_bug(prev);
4869 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4871 schedstat_inc(this_rq(), sched_count);
4872 #ifdef CONFIG_SCHEDSTATS
4873 if (unlikely(prev->lock_depth >= 0)) {
4874 schedstat_inc(this_rq(), bkl_count);
4875 schedstat_inc(prev, sched_info.bkl_count);
4880 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4882 if (prev->state == TASK_RUNNING) {
4883 u64 runtime = prev->se.sum_exec_runtime;
4885 runtime -= prev->se.prev_sum_exec_runtime;
4886 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
4889 * In order to avoid avg_overlap growing stale when we are
4890 * indeed overlapping and hence not getting put to sleep, grow
4891 * the avg_overlap on preemption.
4893 * We use the average preemption runtime because that
4894 * correlates to the amount of cache footprint a task can
4897 update_avg(&prev->se.avg_overlap, runtime);
4899 prev->sched_class->put_prev_task(rq, prev);
4903 * Pick up the highest-prio task:
4905 static inline struct task_struct *
4906 pick_next_task(struct rq *rq)
4908 const struct sched_class *class;
4909 struct task_struct *p;
4912 * Optimization: we know that if all tasks are in
4913 * the fair class we can call that function directly:
4915 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4916 p = fair_sched_class.pick_next_task(rq);
4921 class = sched_class_highest;
4923 p = class->pick_next_task(rq);
4927 * Will never be NULL as the idle class always
4928 * returns a non-NULL p:
4930 class = class->next;
4935 * schedule() is the main scheduler function.
4937 asmlinkage void __sched __schedule(void)
4939 struct task_struct *prev, *next;
4940 unsigned long *switch_count;
4944 cpu = smp_processor_id();
4948 switch_count = &prev->nivcsw;
4950 release_kernel_lock(prev);
4951 need_resched_nonpreemptible:
4953 schedule_debug(prev);
4955 if (sched_feat(HRTICK))
4958 spin_lock_irq(&rq->lock);
4959 update_rq_clock(rq);
4960 clear_tsk_need_resched(prev);
4962 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4963 if (unlikely(signal_pending_state(prev->state, prev)))
4964 prev->state = TASK_RUNNING;
4966 deactivate_task(rq, prev, 1);
4967 switch_count = &prev->nvcsw;
4971 if (prev->sched_class->pre_schedule)
4972 prev->sched_class->pre_schedule(rq, prev);
4975 if (unlikely(!rq->nr_running))
4976 idle_balance(cpu, rq);
4978 put_prev_task(rq, prev);
4979 next = pick_next_task(rq);
4981 if (likely(prev != next)) {
4982 sched_info_switch(prev, next);
4988 context_switch(rq, prev, next); /* unlocks the rq */
4990 * the context switch might have flipped the stack from under
4991 * us, hence refresh the local variables.
4993 cpu = smp_processor_id();
4996 spin_unlock_irq(&rq->lock);
4998 if (unlikely(reacquire_kernel_lock(current) < 0))
4999 goto need_resched_nonpreemptible;
5002 asmlinkage void __sched schedule(void)
5007 preempt_enable_no_resched();
5008 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
5011 EXPORT_SYMBOL(schedule);
5015 * Look out! "owner" is an entirely speculative pointer
5016 * access and not reliable.
5018 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5023 if (!sched_feat(OWNER_SPIN))
5026 #ifdef CONFIG_DEBUG_PAGEALLOC
5028 * Need to access the cpu field knowing that
5029 * DEBUG_PAGEALLOC could have unmapped it if
5030 * the mutex owner just released it and exited.
5032 if (probe_kernel_address(&owner->cpu, cpu))
5039 * Even if the access succeeded (likely case),
5040 * the cpu field may no longer be valid.
5042 if (cpu >= nr_cpumask_bits)
5046 * We need to validate that we can do a
5047 * get_cpu() and that we have the percpu area.
5049 if (!cpu_online(cpu))
5056 * Owner changed, break to re-assess state.
5058 if (lock->owner != owner)
5062 * Is that owner really running on that cpu?
5064 if (task_thread_info(rq->curr) != owner || need_resched())
5074 #ifdef CONFIG_PREEMPT
5076 * this is the entry point to schedule() from in-kernel preemption
5077 * off of preempt_enable. Kernel preemptions off return from interrupt
5078 * occur there and call schedule directly.
5080 asmlinkage void __sched preempt_schedule(void)
5082 struct thread_info *ti = current_thread_info();
5085 * If there is a non-zero preempt_count or interrupts are disabled,
5086 * we do not want to preempt the current task. Just return..
5088 if (likely(ti->preempt_count || irqs_disabled()))
5092 add_preempt_count(PREEMPT_ACTIVE);
5094 sub_preempt_count(PREEMPT_ACTIVE);
5097 * Check again in case we missed a preemption opportunity
5098 * between schedule and now.
5101 } while (need_resched());
5103 EXPORT_SYMBOL(preempt_schedule);
5106 * this is the entry point to schedule() from kernel preemption
5107 * off of irq context.
5108 * Note, that this is called and return with irqs disabled. This will
5109 * protect us against recursive calling from irq.
5111 asmlinkage void __sched preempt_schedule_irq(void)
5113 struct thread_info *ti = current_thread_info();
5115 /* Catch callers which need to be fixed */
5116 BUG_ON(ti->preempt_count || !irqs_disabled());
5119 add_preempt_count(PREEMPT_ACTIVE);
5122 local_irq_disable();
5123 sub_preempt_count(PREEMPT_ACTIVE);
5126 * Check again in case we missed a preemption opportunity
5127 * between schedule and now.
5130 } while (need_resched());
5133 #endif /* CONFIG_PREEMPT */
5135 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
5138 return try_to_wake_up(curr->private, mode, sync);
5140 EXPORT_SYMBOL(default_wake_function);
5143 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5144 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5145 * number) then we wake all the non-exclusive tasks and one exclusive task.
5147 * There are circumstances in which we can try to wake a task which has already
5148 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5149 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5151 void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5152 int nr_exclusive, int sync, void *key)
5154 wait_queue_t *curr, *next;
5156 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5157 unsigned flags = curr->flags;
5159 if (curr->func(curr, mode, sync, key) &&
5160 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5166 * __wake_up - wake up threads blocked on a waitqueue.
5168 * @mode: which threads
5169 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5170 * @key: is directly passed to the wakeup function
5172 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5173 int nr_exclusive, void *key)
5175 unsigned long flags;
5177 spin_lock_irqsave(&q->lock, flags);
5178 __wake_up_common(q, mode, nr_exclusive, 0, key);
5179 spin_unlock_irqrestore(&q->lock, flags);
5181 EXPORT_SYMBOL(__wake_up);
5184 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5186 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5188 __wake_up_common(q, mode, 1, 0, NULL);
5192 * __wake_up_sync - wake up threads blocked on a waitqueue.
5194 * @mode: which threads
5195 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5197 * The sync wakeup differs that the waker knows that it will schedule
5198 * away soon, so while the target thread will be woken up, it will not
5199 * be migrated to another CPU - ie. the two threads are 'synchronized'
5200 * with each other. This can prevent needless bouncing between CPUs.
5202 * On UP it can prevent extra preemption.
5205 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5207 unsigned long flags;
5213 if (unlikely(!nr_exclusive))
5216 spin_lock_irqsave(&q->lock, flags);
5217 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
5218 spin_unlock_irqrestore(&q->lock, flags);
5220 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5223 * complete: - signals a single thread waiting on this completion
5224 * @x: holds the state of this particular completion
5226 * This will wake up a single thread waiting on this completion. Threads will be
5227 * awakened in the same order in which they were queued.
5229 * See also complete_all(), wait_for_completion() and related routines.
5231 void complete(struct completion *x)
5233 unsigned long flags;
5235 spin_lock_irqsave(&x->wait.lock, flags);
5237 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5238 spin_unlock_irqrestore(&x->wait.lock, flags);
5240 EXPORT_SYMBOL(complete);
5243 * complete_all: - signals all threads waiting on this completion
5244 * @x: holds the state of this particular completion
5246 * This will wake up all threads waiting on this particular completion event.
5248 void complete_all(struct completion *x)
5250 unsigned long flags;
5252 spin_lock_irqsave(&x->wait.lock, flags);
5253 x->done += UINT_MAX/2;
5254 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5255 spin_unlock_irqrestore(&x->wait.lock, flags);
5257 EXPORT_SYMBOL(complete_all);
5259 static inline long __sched
5260 do_wait_for_common(struct completion *x, long timeout, int state)
5263 DECLARE_WAITQUEUE(wait, current);
5265 wait.flags |= WQ_FLAG_EXCLUSIVE;
5266 __add_wait_queue_tail(&x->wait, &wait);
5268 if (signal_pending_state(state, current)) {
5269 timeout = -ERESTARTSYS;
5272 __set_current_state(state);
5273 spin_unlock_irq(&x->wait.lock);
5274 timeout = schedule_timeout(timeout);
5275 spin_lock_irq(&x->wait.lock);
5276 } while (!x->done && timeout);
5277 __remove_wait_queue(&x->wait, &wait);
5282 return timeout ?: 1;
5286 wait_for_common(struct completion *x, long timeout, int state)
5290 spin_lock_irq(&x->wait.lock);
5291 timeout = do_wait_for_common(x, timeout, state);
5292 spin_unlock_irq(&x->wait.lock);
5297 * wait_for_completion: - waits for completion of a task
5298 * @x: holds the state of this particular completion
5300 * This waits to be signaled for completion of a specific task. It is NOT
5301 * interruptible and there is no timeout.
5303 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5304 * and interrupt capability. Also see complete().
5306 void __sched wait_for_completion(struct completion *x)
5308 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5310 EXPORT_SYMBOL(wait_for_completion);
5313 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5314 * @x: holds the state of this particular completion
5315 * @timeout: timeout value in jiffies
5317 * This waits for either a completion of a specific task to be signaled or for a
5318 * specified timeout to expire. The timeout is in jiffies. It is not
5321 unsigned long __sched
5322 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5324 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5326 EXPORT_SYMBOL(wait_for_completion_timeout);
5329 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5330 * @x: holds the state of this particular completion
5332 * This waits for completion of a specific task to be signaled. It is
5335 int __sched wait_for_completion_interruptible(struct completion *x)
5337 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5338 if (t == -ERESTARTSYS)
5342 EXPORT_SYMBOL(wait_for_completion_interruptible);
5345 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5346 * @x: holds the state of this particular completion
5347 * @timeout: timeout value in jiffies
5349 * This waits for either a completion of a specific task to be signaled or for a
5350 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5352 unsigned long __sched
5353 wait_for_completion_interruptible_timeout(struct completion *x,
5354 unsigned long timeout)
5356 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5358 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5361 * wait_for_completion_killable: - waits for completion of a task (killable)
5362 * @x: holds the state of this particular completion
5364 * This waits to be signaled for completion of a specific task. It can be
5365 * interrupted by a kill signal.
5367 int __sched wait_for_completion_killable(struct completion *x)
5369 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5370 if (t == -ERESTARTSYS)
5374 EXPORT_SYMBOL(wait_for_completion_killable);
5377 * try_wait_for_completion - try to decrement a completion without blocking
5378 * @x: completion structure
5380 * Returns: 0 if a decrement cannot be done without blocking
5381 * 1 if a decrement succeeded.
5383 * If a completion is being used as a counting completion,
5384 * attempt to decrement the counter without blocking. This
5385 * enables us to avoid waiting if the resource the completion
5386 * is protecting is not available.
5388 bool try_wait_for_completion(struct completion *x)
5392 spin_lock_irq(&x->wait.lock);
5397 spin_unlock_irq(&x->wait.lock);
5400 EXPORT_SYMBOL(try_wait_for_completion);
5403 * completion_done - Test to see if a completion has any waiters
5404 * @x: completion structure
5406 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5407 * 1 if there are no waiters.
5410 bool completion_done(struct completion *x)
5414 spin_lock_irq(&x->wait.lock);
5417 spin_unlock_irq(&x->wait.lock);
5420 EXPORT_SYMBOL(completion_done);
5423 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5425 unsigned long flags;
5428 init_waitqueue_entry(&wait, current);
5430 __set_current_state(state);
5432 spin_lock_irqsave(&q->lock, flags);
5433 __add_wait_queue(q, &wait);
5434 spin_unlock(&q->lock);
5435 timeout = schedule_timeout(timeout);
5436 spin_lock_irq(&q->lock);
5437 __remove_wait_queue(q, &wait);
5438 spin_unlock_irqrestore(&q->lock, flags);
5443 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5445 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5447 EXPORT_SYMBOL(interruptible_sleep_on);
5450 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5452 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5454 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5456 void __sched sleep_on(wait_queue_head_t *q)
5458 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5460 EXPORT_SYMBOL(sleep_on);
5462 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5464 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5466 EXPORT_SYMBOL(sleep_on_timeout);
5468 #ifdef CONFIG_RT_MUTEXES
5471 * rt_mutex_setprio - set the current priority of a task
5473 * @prio: prio value (kernel-internal form)
5475 * This function changes the 'effective' priority of a task. It does
5476 * not touch ->normal_prio like __setscheduler().
5478 * Used by the rt_mutex code to implement priority inheritance logic.
5480 void rt_mutex_setprio(struct task_struct *p, int prio)
5482 unsigned long flags;
5483 int oldprio, on_rq, running;
5485 const struct sched_class *prev_class = p->sched_class;
5487 BUG_ON(prio < 0 || prio > MAX_PRIO);
5489 rq = task_rq_lock(p, &flags);
5490 update_rq_clock(rq);
5493 on_rq = p->se.on_rq;
5494 running = task_current(rq, p);
5496 dequeue_task(rq, p, 0);
5498 p->sched_class->put_prev_task(rq, p);
5501 p->sched_class = &rt_sched_class;
5503 p->sched_class = &fair_sched_class;
5508 p->sched_class->set_curr_task(rq);
5510 enqueue_task(rq, p, 0);
5512 check_class_changed(rq, p, prev_class, oldprio, running);
5514 task_rq_unlock(rq, &flags);
5519 void set_user_nice(struct task_struct *p, long nice)
5521 int old_prio, delta, on_rq;
5522 unsigned long flags;
5525 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5528 * We have to be careful, if called from sys_setpriority(),
5529 * the task might be in the middle of scheduling on another CPU.
5531 rq = task_rq_lock(p, &flags);
5532 update_rq_clock(rq);
5534 * The RT priorities are set via sched_setscheduler(), but we still
5535 * allow the 'normal' nice value to be set - but as expected
5536 * it wont have any effect on scheduling until the task is
5537 * SCHED_FIFO/SCHED_RR:
5539 if (task_has_rt_policy(p)) {
5540 p->static_prio = NICE_TO_PRIO(nice);
5543 on_rq = p->se.on_rq;
5545 dequeue_task(rq, p, 0);
5547 p->static_prio = NICE_TO_PRIO(nice);
5550 p->prio = effective_prio(p);
5551 delta = p->prio - old_prio;
5554 enqueue_task(rq, p, 0);
5556 * If the task increased its priority or is running and
5557 * lowered its priority, then reschedule its CPU:
5559 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5560 resched_task(rq->curr);
5563 task_rq_unlock(rq, &flags);
5565 EXPORT_SYMBOL(set_user_nice);
5568 * can_nice - check if a task can reduce its nice value
5572 int can_nice(const struct task_struct *p, const int nice)
5574 /* convert nice value [19,-20] to rlimit style value [1,40] */
5575 int nice_rlim = 20 - nice;
5577 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
5578 capable(CAP_SYS_NICE));
5581 #ifdef __ARCH_WANT_SYS_NICE
5584 * sys_nice - change the priority of the current process.
5585 * @increment: priority increment
5587 * sys_setpriority is a more generic, but much slower function that
5588 * does similar things.
5590 SYSCALL_DEFINE1(nice, int, increment)
5595 * Setpriority might change our priority at the same moment.
5596 * We don't have to worry. Conceptually one call occurs first
5597 * and we have a single winner.
5599 if (increment < -40)
5604 nice = TASK_NICE(current) + increment;
5610 if (increment < 0 && !can_nice(current, nice))
5613 retval = security_task_setnice(current, nice);
5617 set_user_nice(current, nice);
5624 * task_prio - return the priority value of a given task.
5625 * @p: the task in question.
5627 * This is the priority value as seen by users in /proc.
5628 * RT tasks are offset by -200. Normal tasks are centered
5629 * around 0, value goes from -16 to +15.
5631 int task_prio(const struct task_struct *p)
5633 return p->prio - MAX_RT_PRIO;
5637 * task_nice - return the nice value of a given task.
5638 * @p: the task in question.
5640 int task_nice(const struct task_struct *p)
5642 return TASK_NICE(p);
5644 EXPORT_SYMBOL(task_nice);
5647 * idle_cpu - is a given cpu idle currently?
5648 * @cpu: the processor in question.
5650 int idle_cpu(int cpu)
5652 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5656 * idle_task - return the idle task for a given cpu.
5657 * @cpu: the processor in question.
5659 struct task_struct *idle_task(int cpu)
5661 return cpu_rq(cpu)->idle;
5665 * find_process_by_pid - find a process with a matching PID value.
5666 * @pid: the pid in question.
5668 static struct task_struct *find_process_by_pid(pid_t pid)
5670 return pid ? find_task_by_vpid(pid) : current;
5673 /* Actually do priority change: must hold rq lock. */
5675 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5677 BUG_ON(p->se.on_rq);
5680 switch (p->policy) {
5684 p->sched_class = &fair_sched_class;
5688 p->sched_class = &rt_sched_class;
5692 p->rt_priority = prio;
5693 p->normal_prio = normal_prio(p);
5694 /* we are holding p->pi_lock already */
5695 p->prio = rt_mutex_getprio(p);
5700 * check the target process has a UID that matches the current process's
5702 static bool check_same_owner(struct task_struct *p)
5704 const struct cred *cred = current_cred(), *pcred;
5708 pcred = __task_cred(p);
5709 match = (cred->euid == pcred->euid ||
5710 cred->euid == pcred->uid);
5715 static int __sched_setscheduler(struct task_struct *p, int policy,
5716 struct sched_param *param, bool user)
5718 int retval, oldprio, oldpolicy = -1, on_rq, running;
5719 unsigned long flags;
5720 const struct sched_class *prev_class = p->sched_class;
5723 /* may grab non-irq protected spin_locks */
5724 BUG_ON(in_interrupt());
5726 /* double check policy once rq lock held */
5728 policy = oldpolicy = p->policy;
5729 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5730 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5731 policy != SCHED_IDLE)
5734 * Valid priorities for SCHED_FIFO and SCHED_RR are
5735 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5736 * SCHED_BATCH and SCHED_IDLE is 0.
5738 if (param->sched_priority < 0 ||
5739 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5740 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5742 if (rt_policy(policy) != (param->sched_priority != 0))
5746 * Allow unprivileged RT tasks to decrease priority:
5748 if (user && !capable(CAP_SYS_NICE)) {
5749 if (rt_policy(policy)) {
5750 unsigned long rlim_rtprio;
5752 if (!lock_task_sighand(p, &flags))
5754 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5755 unlock_task_sighand(p, &flags);
5757 /* can't set/change the rt policy */
5758 if (policy != p->policy && !rlim_rtprio)
5761 /* can't increase priority */
5762 if (param->sched_priority > p->rt_priority &&
5763 param->sched_priority > rlim_rtprio)
5767 * Like positive nice levels, dont allow tasks to
5768 * move out of SCHED_IDLE either:
5770 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5773 /* can't change other user's priorities */
5774 if (!check_same_owner(p))
5779 #ifdef CONFIG_RT_GROUP_SCHED
5781 * Do not allow realtime tasks into groups that have no runtime
5784 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5785 task_group(p)->rt_bandwidth.rt_runtime == 0)
5789 retval = security_task_setscheduler(p, policy, param);
5795 * make sure no PI-waiters arrive (or leave) while we are
5796 * changing the priority of the task:
5798 spin_lock_irqsave(&p->pi_lock, flags);
5800 * To be able to change p->policy safely, the apropriate
5801 * runqueue lock must be held.
5803 rq = __task_rq_lock(p);
5804 /* recheck policy now with rq lock held */
5805 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5806 policy = oldpolicy = -1;
5807 __task_rq_unlock(rq);
5808 spin_unlock_irqrestore(&p->pi_lock, flags);
5811 update_rq_clock(rq);
5812 on_rq = p->se.on_rq;
5813 running = task_current(rq, p);
5815 deactivate_task(rq, p, 0);
5817 p->sched_class->put_prev_task(rq, p);
5820 __setscheduler(rq, p, policy, param->sched_priority);
5823 p->sched_class->set_curr_task(rq);
5825 activate_task(rq, p, 0);
5827 check_class_changed(rq, p, prev_class, oldprio, running);
5829 __task_rq_unlock(rq);
5830 spin_unlock_irqrestore(&p->pi_lock, flags);
5832 rt_mutex_adjust_pi(p);
5838 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5839 * @p: the task in question.
5840 * @policy: new policy.
5841 * @param: structure containing the new RT priority.
5843 * NOTE that the task may be already dead.
5845 int sched_setscheduler(struct task_struct *p, int policy,
5846 struct sched_param *param)
5848 return __sched_setscheduler(p, policy, param, true);
5850 EXPORT_SYMBOL_GPL(sched_setscheduler);
5853 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5854 * @p: the task in question.
5855 * @policy: new policy.
5856 * @param: structure containing the new RT priority.
5858 * Just like sched_setscheduler, only don't bother checking if the
5859 * current context has permission. For example, this is needed in
5860 * stop_machine(): we create temporary high priority worker threads,
5861 * but our caller might not have that capability.
5863 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5864 struct sched_param *param)
5866 return __sched_setscheduler(p, policy, param, false);
5870 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5872 struct sched_param lparam;
5873 struct task_struct *p;
5876 if (!param || pid < 0)
5878 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5883 p = find_process_by_pid(pid);
5885 retval = sched_setscheduler(p, policy, &lparam);
5892 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5893 * @pid: the pid in question.
5894 * @policy: new policy.
5895 * @param: structure containing the new RT priority.
5897 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5898 struct sched_param __user *, param)
5900 /* negative values for policy are not valid */
5904 return do_sched_setscheduler(pid, policy, param);
5908 * sys_sched_setparam - set/change the RT priority of a thread
5909 * @pid: the pid in question.
5910 * @param: structure containing the new RT priority.
5912 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5914 return do_sched_setscheduler(pid, -1, param);
5918 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5919 * @pid: the pid in question.
5921 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5923 struct task_struct *p;
5930 read_lock(&tasklist_lock);
5931 p = find_process_by_pid(pid);
5933 retval = security_task_getscheduler(p);
5937 read_unlock(&tasklist_lock);
5942 * sys_sched_getscheduler - get the RT priority of a thread
5943 * @pid: the pid in question.
5944 * @param: structure containing the RT priority.
5946 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5948 struct sched_param lp;
5949 struct task_struct *p;
5952 if (!param || pid < 0)
5955 read_lock(&tasklist_lock);
5956 p = find_process_by_pid(pid);
5961 retval = security_task_getscheduler(p);
5965 lp.sched_priority = p->rt_priority;
5966 read_unlock(&tasklist_lock);
5969 * This one might sleep, we cannot do it with a spinlock held ...
5971 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5976 read_unlock(&tasklist_lock);
5980 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5982 cpumask_var_t cpus_allowed, new_mask;
5983 struct task_struct *p;
5987 read_lock(&tasklist_lock);
5989 p = find_process_by_pid(pid);
5991 read_unlock(&tasklist_lock);
5997 * It is not safe to call set_cpus_allowed with the
5998 * tasklist_lock held. We will bump the task_struct's
5999 * usage count and then drop tasklist_lock.
6002 read_unlock(&tasklist_lock);
6004 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6008 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6010 goto out_free_cpus_allowed;
6013 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6016 retval = security_task_setscheduler(p, 0, NULL);
6020 cpuset_cpus_allowed(p, cpus_allowed);
6021 cpumask_and(new_mask, in_mask, cpus_allowed);
6023 retval = set_cpus_allowed_ptr(p, new_mask);
6026 cpuset_cpus_allowed(p, cpus_allowed);
6027 if (!cpumask_subset(new_mask, cpus_allowed)) {
6029 * We must have raced with a concurrent cpuset
6030 * update. Just reset the cpus_allowed to the
6031 * cpuset's cpus_allowed
6033 cpumask_copy(new_mask, cpus_allowed);
6038 free_cpumask_var(new_mask);
6039 out_free_cpus_allowed:
6040 free_cpumask_var(cpus_allowed);
6047 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6048 struct cpumask *new_mask)
6050 if (len < cpumask_size())
6051 cpumask_clear(new_mask);
6052 else if (len > cpumask_size())
6053 len = cpumask_size();
6055 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6059 * sys_sched_setaffinity - set the cpu affinity of a process
6060 * @pid: pid of the process
6061 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6062 * @user_mask_ptr: user-space pointer to the new cpu mask
6064 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6065 unsigned long __user *, user_mask_ptr)
6067 cpumask_var_t new_mask;
6070 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6073 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6075 retval = sched_setaffinity(pid, new_mask);
6076 free_cpumask_var(new_mask);
6080 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6082 struct task_struct *p;
6086 read_lock(&tasklist_lock);
6089 p = find_process_by_pid(pid);
6093 retval = security_task_getscheduler(p);
6097 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6100 read_unlock(&tasklist_lock);
6107 * sys_sched_getaffinity - get the cpu affinity of a process
6108 * @pid: pid of the process
6109 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6110 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6112 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6113 unsigned long __user *, user_mask_ptr)
6118 if (len < cpumask_size())
6121 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6124 ret = sched_getaffinity(pid, mask);
6126 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6129 ret = cpumask_size();
6131 free_cpumask_var(mask);
6137 * sys_sched_yield - yield the current processor to other threads.
6139 * This function yields the current CPU to other tasks. If there are no
6140 * other threads running on this CPU then this function will return.
6142 SYSCALL_DEFINE0(sched_yield)
6144 struct rq *rq = this_rq_lock();
6146 schedstat_inc(rq, yld_count);
6147 current->sched_class->yield_task(rq);
6150 * Since we are going to call schedule() anyway, there's
6151 * no need to preempt or enable interrupts:
6153 __release(rq->lock);
6154 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6155 _raw_spin_unlock(&rq->lock);
6156 preempt_enable_no_resched();
6163 static void __cond_resched(void)
6165 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6166 __might_sleep(__FILE__, __LINE__);
6169 * The BKS might be reacquired before we have dropped
6170 * PREEMPT_ACTIVE, which could trigger a second
6171 * cond_resched() call.
6174 add_preempt_count(PREEMPT_ACTIVE);
6176 sub_preempt_count(PREEMPT_ACTIVE);
6177 } while (need_resched());
6180 int __sched _cond_resched(void)
6182 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
6183 system_state == SYSTEM_RUNNING) {
6189 EXPORT_SYMBOL(_cond_resched);
6192 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
6193 * call schedule, and on return reacquire the lock.
6195 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6196 * operations here to prevent schedule() from being called twice (once via
6197 * spin_unlock(), once by hand).
6199 int cond_resched_lock(spinlock_t *lock)
6201 int resched = need_resched() && system_state == SYSTEM_RUNNING;
6204 if (spin_needbreak(lock) || resched) {
6206 if (resched && need_resched())
6215 EXPORT_SYMBOL(cond_resched_lock);
6217 int __sched cond_resched_softirq(void)
6219 BUG_ON(!in_softirq());
6221 if (need_resched() && system_state == SYSTEM_RUNNING) {
6229 EXPORT_SYMBOL(cond_resched_softirq);
6232 * yield - yield the current processor to other threads.
6234 * This is a shortcut for kernel-space yielding - it marks the
6235 * thread runnable and calls sys_sched_yield().
6237 void __sched yield(void)
6239 set_current_state(TASK_RUNNING);
6242 EXPORT_SYMBOL(yield);
6245 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6246 * that process accounting knows that this is a task in IO wait state.
6248 * But don't do that if it is a deliberate, throttling IO wait (this task
6249 * has set its backing_dev_info: the queue against which it should throttle)
6251 void __sched io_schedule(void)
6253 struct rq *rq = &__raw_get_cpu_var(runqueues);
6255 delayacct_blkio_start();
6256 atomic_inc(&rq->nr_iowait);
6258 atomic_dec(&rq->nr_iowait);
6259 delayacct_blkio_end();
6261 EXPORT_SYMBOL(io_schedule);
6263 long __sched io_schedule_timeout(long timeout)
6265 struct rq *rq = &__raw_get_cpu_var(runqueues);
6268 delayacct_blkio_start();
6269 atomic_inc(&rq->nr_iowait);
6270 ret = schedule_timeout(timeout);
6271 atomic_dec(&rq->nr_iowait);
6272 delayacct_blkio_end();
6277 * sys_sched_get_priority_max - return maximum RT priority.
6278 * @policy: scheduling class.
6280 * this syscall returns the maximum rt_priority that can be used
6281 * by a given scheduling class.
6283 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6290 ret = MAX_USER_RT_PRIO-1;
6302 * sys_sched_get_priority_min - return minimum RT priority.
6303 * @policy: scheduling class.
6305 * this syscall returns the minimum rt_priority that can be used
6306 * by a given scheduling class.
6308 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6326 * sys_sched_rr_get_interval - return the default timeslice of a process.
6327 * @pid: pid of the process.
6328 * @interval: userspace pointer to the timeslice value.
6330 * this syscall writes the default timeslice value of a given process
6331 * into the user-space timespec buffer. A value of '0' means infinity.
6333 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6334 struct timespec __user *, interval)
6336 struct task_struct *p;
6337 unsigned int time_slice;
6345 read_lock(&tasklist_lock);
6346 p = find_process_by_pid(pid);
6350 retval = security_task_getscheduler(p);
6355 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6356 * tasks that are on an otherwise idle runqueue:
6359 if (p->policy == SCHED_RR) {
6360 time_slice = DEF_TIMESLICE;
6361 } else if (p->policy != SCHED_FIFO) {
6362 struct sched_entity *se = &p->se;
6363 unsigned long flags;
6366 rq = task_rq_lock(p, &flags);
6367 if (rq->cfs.load.weight)
6368 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
6369 task_rq_unlock(rq, &flags);
6371 read_unlock(&tasklist_lock);
6372 jiffies_to_timespec(time_slice, &t);
6373 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6377 read_unlock(&tasklist_lock);
6381 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6383 void sched_show_task(struct task_struct *p)
6385 unsigned long free = 0;
6388 state = p->state ? __ffs(p->state) + 1 : 0;
6389 printk(KERN_INFO "%-13.13s %c", p->comm,
6390 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6391 #if BITS_PER_LONG == 32
6392 if (state == TASK_RUNNING)
6393 printk(KERN_CONT " running ");
6395 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6397 if (state == TASK_RUNNING)
6398 printk(KERN_CONT " running task ");
6400 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6402 #ifdef CONFIG_DEBUG_STACK_USAGE
6403 free = stack_not_used(p);
6405 printk(KERN_CONT "%5lu %5d %6d\n", free,
6406 task_pid_nr(p), task_pid_nr(p->real_parent));
6408 show_stack(p, NULL);
6411 void show_state_filter(unsigned long state_filter)
6413 struct task_struct *g, *p;
6415 #if BITS_PER_LONG == 32
6417 " task PC stack pid father\n");
6420 " task PC stack pid father\n");
6422 read_lock(&tasklist_lock);
6423 do_each_thread(g, p) {
6425 * reset the NMI-timeout, listing all files on a slow
6426 * console might take alot of time:
6428 touch_nmi_watchdog();
6429 if (!state_filter || (p->state & state_filter))
6431 } while_each_thread(g, p);
6433 touch_all_softlockup_watchdogs();
6435 #ifdef CONFIG_SCHED_DEBUG
6436 sysrq_sched_debug_show();
6438 read_unlock(&tasklist_lock);
6440 * Only show locks if all tasks are dumped:
6442 if (state_filter == -1)
6443 debug_show_all_locks();
6446 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6448 idle->sched_class = &idle_sched_class;
6452 * init_idle - set up an idle thread for a given CPU
6453 * @idle: task in question
6454 * @cpu: cpu the idle task belongs to
6456 * NOTE: this function does not set the idle thread's NEED_RESCHED
6457 * flag, to make booting more robust.
6459 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6461 struct rq *rq = cpu_rq(cpu);
6462 unsigned long flags;
6464 spin_lock_irqsave(&rq->lock, flags);
6467 idle->se.exec_start = sched_clock();
6469 idle->prio = idle->normal_prio = MAX_PRIO;
6470 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6471 __set_task_cpu(idle, cpu);
6473 rq->curr = rq->idle = idle;
6474 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6477 spin_unlock_irqrestore(&rq->lock, flags);
6479 /* Set the preempt count _outside_ the spinlocks! */
6480 #if defined(CONFIG_PREEMPT)
6481 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6483 task_thread_info(idle)->preempt_count = 0;
6486 * The idle tasks have their own, simple scheduling class:
6488 idle->sched_class = &idle_sched_class;
6489 ftrace_graph_init_task(idle);
6493 * In a system that switches off the HZ timer nohz_cpu_mask
6494 * indicates which cpus entered this state. This is used
6495 * in the rcu update to wait only for active cpus. For system
6496 * which do not switch off the HZ timer nohz_cpu_mask should
6497 * always be CPU_BITS_NONE.
6499 cpumask_var_t nohz_cpu_mask;
6502 * Increase the granularity value when there are more CPUs,
6503 * because with more CPUs the 'effective latency' as visible
6504 * to users decreases. But the relationship is not linear,
6505 * so pick a second-best guess by going with the log2 of the
6508 * This idea comes from the SD scheduler of Con Kolivas:
6510 static inline void sched_init_granularity(void)
6512 unsigned int factor = 1 + ilog2(num_online_cpus());
6513 const unsigned long limit = 200000000;
6515 sysctl_sched_min_granularity *= factor;
6516 if (sysctl_sched_min_granularity > limit)
6517 sysctl_sched_min_granularity = limit;
6519 sysctl_sched_latency *= factor;
6520 if (sysctl_sched_latency > limit)
6521 sysctl_sched_latency = limit;
6523 sysctl_sched_wakeup_granularity *= factor;
6525 sysctl_sched_shares_ratelimit *= factor;
6530 * This is how migration works:
6532 * 1) we queue a struct migration_req structure in the source CPU's
6533 * runqueue and wake up that CPU's migration thread.
6534 * 2) we down() the locked semaphore => thread blocks.
6535 * 3) migration thread wakes up (implicitly it forces the migrated
6536 * thread off the CPU)
6537 * 4) it gets the migration request and checks whether the migrated
6538 * task is still in the wrong runqueue.
6539 * 5) if it's in the wrong runqueue then the migration thread removes
6540 * it and puts it into the right queue.
6541 * 6) migration thread up()s the semaphore.
6542 * 7) we wake up and the migration is done.
6546 * Change a given task's CPU affinity. Migrate the thread to a
6547 * proper CPU and schedule it away if the CPU it's executing on
6548 * is removed from the allowed bitmask.
6550 * NOTE: the caller must have a valid reference to the task, the
6551 * task must not exit() & deallocate itself prematurely. The
6552 * call is not atomic; no spinlocks may be held.
6554 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6556 struct migration_req req;
6557 unsigned long flags;
6561 rq = task_rq_lock(p, &flags);
6562 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
6567 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
6568 !cpumask_equal(&p->cpus_allowed, new_mask))) {
6573 if (p->sched_class->set_cpus_allowed)
6574 p->sched_class->set_cpus_allowed(p, new_mask);
6576 cpumask_copy(&p->cpus_allowed, new_mask);
6577 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6580 /* Can the task run on the task's current CPU? If so, we're done */
6581 if (cpumask_test_cpu(task_cpu(p), new_mask))
6584 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
6585 /* Need help from migration thread: drop lock and wait. */
6586 task_rq_unlock(rq, &flags);
6587 wake_up_process(rq->migration_thread);
6588 wait_for_completion(&req.done);
6589 tlb_migrate_finish(p->mm);
6593 task_rq_unlock(rq, &flags);
6597 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6600 * Move (not current) task off this cpu, onto dest cpu. We're doing
6601 * this because either it can't run here any more (set_cpus_allowed()
6602 * away from this CPU, or CPU going down), or because we're
6603 * attempting to rebalance this task on exec (sched_exec).
6605 * So we race with normal scheduler movements, but that's OK, as long
6606 * as the task is no longer on this CPU.
6608 * Returns non-zero if task was successfully migrated.
6610 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6612 struct rq *rq_dest, *rq_src;
6615 if (unlikely(!cpu_active(dest_cpu)))
6618 rq_src = cpu_rq(src_cpu);
6619 rq_dest = cpu_rq(dest_cpu);
6621 double_rq_lock(rq_src, rq_dest);
6622 /* Already moved. */
6623 if (task_cpu(p) != src_cpu)
6625 /* Affinity changed (again). */
6626 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6629 on_rq = p->se.on_rq;
6631 deactivate_task(rq_src, p, 0);
6633 set_task_cpu(p, dest_cpu);
6635 activate_task(rq_dest, p, 0);
6636 check_preempt_curr(rq_dest, p, 0);
6641 double_rq_unlock(rq_src, rq_dest);
6646 * migration_thread - this is a highprio system thread that performs
6647 * thread migration by bumping thread off CPU then 'pushing' onto
6650 static int migration_thread(void *data)
6652 int cpu = (long)data;
6656 BUG_ON(rq->migration_thread != current);
6658 set_current_state(TASK_INTERRUPTIBLE);
6659 while (!kthread_should_stop()) {
6660 struct migration_req *req;
6661 struct list_head *head;
6663 spin_lock_irq(&rq->lock);
6665 if (cpu_is_offline(cpu)) {
6666 spin_unlock_irq(&rq->lock);
6670 if (rq->active_balance) {
6671 active_load_balance(rq, cpu);
6672 rq->active_balance = 0;
6675 head = &rq->migration_queue;
6677 if (list_empty(head)) {
6678 spin_unlock_irq(&rq->lock);
6680 set_current_state(TASK_INTERRUPTIBLE);
6683 req = list_entry(head->next, struct migration_req, list);
6684 list_del_init(head->next);
6686 spin_unlock(&rq->lock);
6687 __migrate_task(req->task, cpu, req->dest_cpu);
6690 complete(&req->done);
6692 __set_current_state(TASK_RUNNING);
6696 /* Wait for kthread_stop */
6697 set_current_state(TASK_INTERRUPTIBLE);
6698 while (!kthread_should_stop()) {
6700 set_current_state(TASK_INTERRUPTIBLE);
6702 __set_current_state(TASK_RUNNING);
6706 #ifdef CONFIG_HOTPLUG_CPU
6708 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6712 local_irq_disable();
6713 ret = __migrate_task(p, src_cpu, dest_cpu);
6719 * Figure out where task on dead CPU should go, use force if necessary.
6721 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6724 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
6727 /* Look for allowed, online CPU in same node. */
6728 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
6729 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6732 /* Any allowed, online CPU? */
6733 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
6734 if (dest_cpu < nr_cpu_ids)
6737 /* No more Mr. Nice Guy. */
6738 if (dest_cpu >= nr_cpu_ids) {
6739 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
6740 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
6743 * Don't tell them about moving exiting tasks or
6744 * kernel threads (both mm NULL), since they never
6747 if (p->mm && printk_ratelimit()) {
6748 printk(KERN_INFO "process %d (%s) no "
6749 "longer affine to cpu%d\n",
6750 task_pid_nr(p), p->comm, dead_cpu);
6755 /* It can have affinity changed while we were choosing. */
6756 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
6761 * While a dead CPU has no uninterruptible tasks queued at this point,
6762 * it might still have a nonzero ->nr_uninterruptible counter, because
6763 * for performance reasons the counter is not stricly tracking tasks to
6764 * their home CPUs. So we just add the counter to another CPU's counter,
6765 * to keep the global sum constant after CPU-down:
6767 static void migrate_nr_uninterruptible(struct rq *rq_src)
6769 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
6770 unsigned long flags;
6772 local_irq_save(flags);
6773 double_rq_lock(rq_src, rq_dest);
6774 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6775 rq_src->nr_uninterruptible = 0;
6776 double_rq_unlock(rq_src, rq_dest);
6777 local_irq_restore(flags);
6780 /* Run through task list and migrate tasks from the dead cpu. */
6781 static void migrate_live_tasks(int src_cpu)
6783 struct task_struct *p, *t;
6785 read_lock(&tasklist_lock);
6787 do_each_thread(t, p) {
6791 if (task_cpu(p) == src_cpu)
6792 move_task_off_dead_cpu(src_cpu, p);
6793 } while_each_thread(t, p);
6795 read_unlock(&tasklist_lock);
6799 * Schedules idle task to be the next runnable task on current CPU.
6800 * It does so by boosting its priority to highest possible.
6801 * Used by CPU offline code.
6803 void sched_idle_next(void)
6805 int this_cpu = smp_processor_id();
6806 struct rq *rq = cpu_rq(this_cpu);
6807 struct task_struct *p = rq->idle;
6808 unsigned long flags;
6810 /* cpu has to be offline */
6811 BUG_ON(cpu_online(this_cpu));
6814 * Strictly not necessary since rest of the CPUs are stopped by now
6815 * and interrupts disabled on the current cpu.
6817 spin_lock_irqsave(&rq->lock, flags);
6819 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6821 update_rq_clock(rq);
6822 activate_task(rq, p, 0);
6824 spin_unlock_irqrestore(&rq->lock, flags);
6828 * Ensures that the idle task is using init_mm right before its cpu goes
6831 void idle_task_exit(void)
6833 struct mm_struct *mm = current->active_mm;
6835 BUG_ON(cpu_online(smp_processor_id()));
6838 switch_mm(mm, &init_mm, current);
6842 /* called under rq->lock with disabled interrupts */
6843 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6845 struct rq *rq = cpu_rq(dead_cpu);
6847 /* Must be exiting, otherwise would be on tasklist. */
6848 BUG_ON(!p->exit_state);
6850 /* Cannot have done final schedule yet: would have vanished. */
6851 BUG_ON(p->state == TASK_DEAD);
6856 * Drop lock around migration; if someone else moves it,
6857 * that's OK. No task can be added to this CPU, so iteration is
6860 spin_unlock_irq(&rq->lock);
6861 move_task_off_dead_cpu(dead_cpu, p);
6862 spin_lock_irq(&rq->lock);
6867 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6868 static void migrate_dead_tasks(unsigned int dead_cpu)
6870 struct rq *rq = cpu_rq(dead_cpu);
6871 struct task_struct *next;
6874 if (!rq->nr_running)
6876 update_rq_clock(rq);
6877 next = pick_next_task(rq);
6880 next->sched_class->put_prev_task(rq, next);
6881 migrate_dead(dead_cpu, next);
6885 #endif /* CONFIG_HOTPLUG_CPU */
6887 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6889 static struct ctl_table sd_ctl_dir[] = {
6891 .procname = "sched_domain",
6897 static struct ctl_table sd_ctl_root[] = {
6899 .ctl_name = CTL_KERN,
6900 .procname = "kernel",
6902 .child = sd_ctl_dir,
6907 static struct ctl_table *sd_alloc_ctl_entry(int n)
6909 struct ctl_table *entry =
6910 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6915 static void sd_free_ctl_entry(struct ctl_table **tablep)
6917 struct ctl_table *entry;
6920 * In the intermediate directories, both the child directory and
6921 * procname are dynamically allocated and could fail but the mode
6922 * will always be set. In the lowest directory the names are
6923 * static strings and all have proc handlers.
6925 for (entry = *tablep; entry->mode; entry++) {
6927 sd_free_ctl_entry(&entry->child);
6928 if (entry->proc_handler == NULL)
6929 kfree(entry->procname);
6937 set_table_entry(struct ctl_table *entry,
6938 const char *procname, void *data, int maxlen,
6939 mode_t mode, proc_handler *proc_handler)
6941 entry->procname = procname;
6943 entry->maxlen = maxlen;
6945 entry->proc_handler = proc_handler;
6948 static struct ctl_table *
6949 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6951 struct ctl_table *table = sd_alloc_ctl_entry(13);
6956 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6957 sizeof(long), 0644, proc_doulongvec_minmax);
6958 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6959 sizeof(long), 0644, proc_doulongvec_minmax);
6960 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6961 sizeof(int), 0644, proc_dointvec_minmax);
6962 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6963 sizeof(int), 0644, proc_dointvec_minmax);
6964 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6965 sizeof(int), 0644, proc_dointvec_minmax);
6966 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6967 sizeof(int), 0644, proc_dointvec_minmax);
6968 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6969 sizeof(int), 0644, proc_dointvec_minmax);
6970 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6971 sizeof(int), 0644, proc_dointvec_minmax);
6972 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6973 sizeof(int), 0644, proc_dointvec_minmax);
6974 set_table_entry(&table[9], "cache_nice_tries",
6975 &sd->cache_nice_tries,
6976 sizeof(int), 0644, proc_dointvec_minmax);
6977 set_table_entry(&table[10], "flags", &sd->flags,
6978 sizeof(int), 0644, proc_dointvec_minmax);
6979 set_table_entry(&table[11], "name", sd->name,
6980 CORENAME_MAX_SIZE, 0444, proc_dostring);
6981 /* &table[12] is terminator */
6986 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6988 struct ctl_table *entry, *table;
6989 struct sched_domain *sd;
6990 int domain_num = 0, i;
6993 for_each_domain(cpu, sd)
6995 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7000 for_each_domain(cpu, sd) {
7001 snprintf(buf, 32, "domain%d", i);
7002 entry->procname = kstrdup(buf, GFP_KERNEL);
7004 entry->child = sd_alloc_ctl_domain_table(sd);
7011 static struct ctl_table_header *sd_sysctl_header;
7012 static void register_sched_domain_sysctl(void)
7014 int i, cpu_num = num_online_cpus();
7015 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7018 WARN_ON(sd_ctl_dir[0].child);
7019 sd_ctl_dir[0].child = entry;
7024 for_each_online_cpu(i) {
7025 snprintf(buf, 32, "cpu%d", i);
7026 entry->procname = kstrdup(buf, GFP_KERNEL);
7028 entry->child = sd_alloc_ctl_cpu_table(i);
7032 WARN_ON(sd_sysctl_header);
7033 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7036 /* may be called multiple times per register */
7037 static void unregister_sched_domain_sysctl(void)
7039 if (sd_sysctl_header)
7040 unregister_sysctl_table(sd_sysctl_header);
7041 sd_sysctl_header = NULL;
7042 if (sd_ctl_dir[0].child)
7043 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7046 static void register_sched_domain_sysctl(void)
7049 static void unregister_sched_domain_sysctl(void)
7054 static void set_rq_online(struct rq *rq)
7057 const struct sched_class *class;
7059 cpumask_set_cpu(rq->cpu, rq->rd->online);
7062 for_each_class(class) {
7063 if (class->rq_online)
7064 class->rq_online(rq);
7069 static void set_rq_offline(struct rq *rq)
7072 const struct sched_class *class;
7074 for_each_class(class) {
7075 if (class->rq_offline)
7076 class->rq_offline(rq);
7079 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7085 * migration_call - callback that gets triggered when a CPU is added.
7086 * Here we can start up the necessary migration thread for the new CPU.
7088 static int __cpuinit
7089 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7091 struct task_struct *p;
7092 int cpu = (long)hcpu;
7093 unsigned long flags;
7098 case CPU_UP_PREPARE:
7099 case CPU_UP_PREPARE_FROZEN:
7100 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7103 kthread_bind(p, cpu);
7104 /* Must be high prio: stop_machine expects to yield to it. */
7105 rq = task_rq_lock(p, &flags);
7106 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7107 task_rq_unlock(rq, &flags);
7108 cpu_rq(cpu)->migration_thread = p;
7112 case CPU_ONLINE_FROZEN:
7113 /* Strictly unnecessary, as first user will wake it. */
7114 wake_up_process(cpu_rq(cpu)->migration_thread);
7116 /* Update our root-domain */
7118 spin_lock_irqsave(&rq->lock, flags);
7120 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7124 spin_unlock_irqrestore(&rq->lock, flags);
7127 #ifdef CONFIG_HOTPLUG_CPU
7128 case CPU_UP_CANCELED:
7129 case CPU_UP_CANCELED_FROZEN:
7130 if (!cpu_rq(cpu)->migration_thread)
7132 /* Unbind it from offline cpu so it can run. Fall thru. */
7133 kthread_bind(cpu_rq(cpu)->migration_thread,
7134 cpumask_any(cpu_online_mask));
7135 kthread_stop(cpu_rq(cpu)->migration_thread);
7136 cpu_rq(cpu)->migration_thread = NULL;
7140 case CPU_DEAD_FROZEN:
7141 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7142 migrate_live_tasks(cpu);
7144 kthread_stop(rq->migration_thread);
7145 rq->migration_thread = NULL;
7146 /* Idle task back to normal (off runqueue, low prio) */
7147 spin_lock_irq(&rq->lock);
7148 update_rq_clock(rq);
7149 deactivate_task(rq, rq->idle, 0);
7150 rq->idle->static_prio = MAX_PRIO;
7151 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7152 rq->idle->sched_class = &idle_sched_class;
7153 migrate_dead_tasks(cpu);
7154 spin_unlock_irq(&rq->lock);
7156 migrate_nr_uninterruptible(rq);
7157 BUG_ON(rq->nr_running != 0);
7160 * No need to migrate the tasks: it was best-effort if
7161 * they didn't take sched_hotcpu_mutex. Just wake up
7164 spin_lock_irq(&rq->lock);
7165 while (!list_empty(&rq->migration_queue)) {
7166 struct migration_req *req;
7168 req = list_entry(rq->migration_queue.next,
7169 struct migration_req, list);
7170 list_del_init(&req->list);
7171 spin_unlock_irq(&rq->lock);
7172 complete(&req->done);
7173 spin_lock_irq(&rq->lock);
7175 spin_unlock_irq(&rq->lock);
7179 case CPU_DYING_FROZEN:
7180 /* Update our root-domain */
7182 spin_lock_irqsave(&rq->lock, flags);
7184 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7187 spin_unlock_irqrestore(&rq->lock, flags);
7194 /* Register at highest priority so that task migration (migrate_all_tasks)
7195 * happens before everything else.
7197 static struct notifier_block __cpuinitdata migration_notifier = {
7198 .notifier_call = migration_call,
7202 static int __init migration_init(void)
7204 void *cpu = (void *)(long)smp_processor_id();
7207 /* Start one for the boot CPU: */
7208 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7209 BUG_ON(err == NOTIFY_BAD);
7210 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7211 register_cpu_notifier(&migration_notifier);
7215 early_initcall(migration_init);
7220 #ifdef CONFIG_SCHED_DEBUG
7222 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7223 struct cpumask *groupmask)
7225 struct sched_group *group = sd->groups;
7228 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7229 cpumask_clear(groupmask);
7231 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7233 if (!(sd->flags & SD_LOAD_BALANCE)) {
7234 printk("does not load-balance\n");
7236 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7241 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7243 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7244 printk(KERN_ERR "ERROR: domain->span does not contain "
7247 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7248 printk(KERN_ERR "ERROR: domain->groups does not contain"
7252 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7256 printk(KERN_ERR "ERROR: group is NULL\n");
7260 if (!group->__cpu_power) {
7261 printk(KERN_CONT "\n");
7262 printk(KERN_ERR "ERROR: domain->cpu_power not "
7267 if (!cpumask_weight(sched_group_cpus(group))) {
7268 printk(KERN_CONT "\n");
7269 printk(KERN_ERR "ERROR: empty group\n");
7273 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7274 printk(KERN_CONT "\n");
7275 printk(KERN_ERR "ERROR: repeated CPUs\n");
7279 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7281 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7282 printk(KERN_CONT " %s", str);
7284 group = group->next;
7285 } while (group != sd->groups);
7286 printk(KERN_CONT "\n");
7288 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7289 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7292 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7293 printk(KERN_ERR "ERROR: parent span is not a superset "
7294 "of domain->span\n");
7298 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7300 cpumask_var_t groupmask;
7304 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7308 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7310 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7311 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7316 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7323 free_cpumask_var(groupmask);
7325 #else /* !CONFIG_SCHED_DEBUG */
7326 # define sched_domain_debug(sd, cpu) do { } while (0)
7327 #endif /* CONFIG_SCHED_DEBUG */
7329 static int sd_degenerate(struct sched_domain *sd)
7331 if (cpumask_weight(sched_domain_span(sd)) == 1)
7334 /* Following flags need at least 2 groups */
7335 if (sd->flags & (SD_LOAD_BALANCE |
7336 SD_BALANCE_NEWIDLE |
7340 SD_SHARE_PKG_RESOURCES)) {
7341 if (sd->groups != sd->groups->next)
7345 /* Following flags don't use groups */
7346 if (sd->flags & (SD_WAKE_IDLE |
7355 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7357 unsigned long cflags = sd->flags, pflags = parent->flags;
7359 if (sd_degenerate(parent))
7362 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7365 /* Does parent contain flags not in child? */
7366 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
7367 if (cflags & SD_WAKE_AFFINE)
7368 pflags &= ~SD_WAKE_BALANCE;
7369 /* Flags needing groups don't count if only 1 group in parent */
7370 if (parent->groups == parent->groups->next) {
7371 pflags &= ~(SD_LOAD_BALANCE |
7372 SD_BALANCE_NEWIDLE |
7376 SD_SHARE_PKG_RESOURCES);
7377 if (nr_node_ids == 1)
7378 pflags &= ~SD_SERIALIZE;
7380 if (~cflags & pflags)
7386 static void free_rootdomain(struct root_domain *rd)
7388 cpupri_cleanup(&rd->cpupri);
7390 free_cpumask_var(rd->rto_mask);
7391 free_cpumask_var(rd->online);
7392 free_cpumask_var(rd->span);
7396 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7398 struct root_domain *old_rd = NULL;
7399 unsigned long flags;
7401 spin_lock_irqsave(&rq->lock, flags);
7406 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7409 cpumask_clear_cpu(rq->cpu, old_rd->span);
7412 * If we dont want to free the old_rt yet then
7413 * set old_rd to NULL to skip the freeing later
7416 if (!atomic_dec_and_test(&old_rd->refcount))
7420 atomic_inc(&rd->refcount);
7423 cpumask_set_cpu(rq->cpu, rd->span);
7424 if (cpumask_test_cpu(rq->cpu, cpu_online_mask))
7427 spin_unlock_irqrestore(&rq->lock, flags);
7430 free_rootdomain(old_rd);
7433 static int __init_refok init_rootdomain(struct root_domain *rd, bool bootmem)
7435 memset(rd, 0, sizeof(*rd));
7438 alloc_bootmem_cpumask_var(&def_root_domain.span);
7439 alloc_bootmem_cpumask_var(&def_root_domain.online);
7440 alloc_bootmem_cpumask_var(&def_root_domain.rto_mask);
7441 cpupri_init(&rd->cpupri, true);
7445 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
7447 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
7449 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
7452 if (cpupri_init(&rd->cpupri, false) != 0)
7457 free_cpumask_var(rd->rto_mask);
7459 free_cpumask_var(rd->online);
7461 free_cpumask_var(rd->span);
7466 static void init_defrootdomain(void)
7468 init_rootdomain(&def_root_domain, true);
7470 atomic_set(&def_root_domain.refcount, 1);
7473 static struct root_domain *alloc_rootdomain(void)
7475 struct root_domain *rd;
7477 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7481 if (init_rootdomain(rd, false) != 0) {
7490 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7491 * hold the hotplug lock.
7494 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7496 struct rq *rq = cpu_rq(cpu);
7497 struct sched_domain *tmp;
7499 /* Remove the sched domains which do not contribute to scheduling. */
7500 for (tmp = sd; tmp; ) {
7501 struct sched_domain *parent = tmp->parent;
7505 if (sd_parent_degenerate(tmp, parent)) {
7506 tmp->parent = parent->parent;
7508 parent->parent->child = tmp;
7513 if (sd && sd_degenerate(sd)) {
7519 sched_domain_debug(sd, cpu);
7521 rq_attach_root(rq, rd);
7522 rcu_assign_pointer(rq->sd, sd);
7525 /* cpus with isolated domains */
7526 static cpumask_var_t cpu_isolated_map;
7528 /* Setup the mask of cpus configured for isolated domains */
7529 static int __init isolated_cpu_setup(char *str)
7531 cpulist_parse(str, cpu_isolated_map);
7535 __setup("isolcpus=", isolated_cpu_setup);
7538 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7539 * to a function which identifies what group(along with sched group) a CPU
7540 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7541 * (due to the fact that we keep track of groups covered with a struct cpumask).
7543 * init_sched_build_groups will build a circular linked list of the groups
7544 * covered by the given span, and will set each group's ->cpumask correctly,
7545 * and ->cpu_power to 0.
7548 init_sched_build_groups(const struct cpumask *span,
7549 const struct cpumask *cpu_map,
7550 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
7551 struct sched_group **sg,
7552 struct cpumask *tmpmask),
7553 struct cpumask *covered, struct cpumask *tmpmask)
7555 struct sched_group *first = NULL, *last = NULL;
7558 cpumask_clear(covered);
7560 for_each_cpu(i, span) {
7561 struct sched_group *sg;
7562 int group = group_fn(i, cpu_map, &sg, tmpmask);
7565 if (cpumask_test_cpu(i, covered))
7568 cpumask_clear(sched_group_cpus(sg));
7569 sg->__cpu_power = 0;
7571 for_each_cpu(j, span) {
7572 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
7575 cpumask_set_cpu(j, covered);
7576 cpumask_set_cpu(j, sched_group_cpus(sg));
7587 #define SD_NODES_PER_DOMAIN 16
7592 * find_next_best_node - find the next node to include in a sched_domain
7593 * @node: node whose sched_domain we're building
7594 * @used_nodes: nodes already in the sched_domain
7596 * Find the next node to include in a given scheduling domain. Simply
7597 * finds the closest node not already in the @used_nodes map.
7599 * Should use nodemask_t.
7601 static int find_next_best_node(int node, nodemask_t *used_nodes)
7603 int i, n, val, min_val, best_node = 0;
7607 for (i = 0; i < nr_node_ids; i++) {
7608 /* Start at @node */
7609 n = (node + i) % nr_node_ids;
7611 if (!nr_cpus_node(n))
7614 /* Skip already used nodes */
7615 if (node_isset(n, *used_nodes))
7618 /* Simple min distance search */
7619 val = node_distance(node, n);
7621 if (val < min_val) {
7627 node_set(best_node, *used_nodes);
7632 * sched_domain_node_span - get a cpumask for a node's sched_domain
7633 * @node: node whose cpumask we're constructing
7634 * @span: resulting cpumask
7636 * Given a node, construct a good cpumask for its sched_domain to span. It
7637 * should be one that prevents unnecessary balancing, but also spreads tasks
7640 static void sched_domain_node_span(int node, struct cpumask *span)
7642 nodemask_t used_nodes;
7645 cpumask_clear(span);
7646 nodes_clear(used_nodes);
7648 cpumask_or(span, span, cpumask_of_node(node));
7649 node_set(node, used_nodes);
7651 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7652 int next_node = find_next_best_node(node, &used_nodes);
7654 cpumask_or(span, span, cpumask_of_node(next_node));
7657 #endif /* CONFIG_NUMA */
7659 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7662 * The cpus mask in sched_group and sched_domain hangs off the end.
7663 * FIXME: use cpumask_var_t or dynamic percpu alloc to avoid wasting space
7664 * for nr_cpu_ids < CONFIG_NR_CPUS.
7666 struct static_sched_group {
7667 struct sched_group sg;
7668 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
7671 struct static_sched_domain {
7672 struct sched_domain sd;
7673 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
7677 * SMT sched-domains:
7679 #ifdef CONFIG_SCHED_SMT
7680 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
7681 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
7684 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
7685 struct sched_group **sg, struct cpumask *unused)
7688 *sg = &per_cpu(sched_group_cpus, cpu).sg;
7691 #endif /* CONFIG_SCHED_SMT */
7694 * multi-core sched-domains:
7696 #ifdef CONFIG_SCHED_MC
7697 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
7698 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
7699 #endif /* CONFIG_SCHED_MC */
7701 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7703 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
7704 struct sched_group **sg, struct cpumask *mask)
7708 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
7709 group = cpumask_first(mask);
7711 *sg = &per_cpu(sched_group_core, group).sg;
7714 #elif defined(CONFIG_SCHED_MC)
7716 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
7717 struct sched_group **sg, struct cpumask *unused)
7720 *sg = &per_cpu(sched_group_core, cpu).sg;
7725 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
7726 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
7729 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
7730 struct sched_group **sg, struct cpumask *mask)
7733 #ifdef CONFIG_SCHED_MC
7734 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
7735 group = cpumask_first(mask);
7736 #elif defined(CONFIG_SCHED_SMT)
7737 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
7738 group = cpumask_first(mask);
7743 *sg = &per_cpu(sched_group_phys, group).sg;
7749 * The init_sched_build_groups can't handle what we want to do with node
7750 * groups, so roll our own. Now each node has its own list of groups which
7751 * gets dynamically allocated.
7753 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
7754 static struct sched_group ***sched_group_nodes_bycpu;
7756 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
7757 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
7759 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
7760 struct sched_group **sg,
7761 struct cpumask *nodemask)
7765 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
7766 group = cpumask_first(nodemask);
7769 *sg = &per_cpu(sched_group_allnodes, group).sg;
7773 static void init_numa_sched_groups_power(struct sched_group *group_head)
7775 struct sched_group *sg = group_head;
7781 for_each_cpu(j, sched_group_cpus(sg)) {
7782 struct sched_domain *sd;
7784 sd = &per_cpu(phys_domains, j).sd;
7785 if (j != cpumask_first(sched_group_cpus(sd->groups))) {
7787 * Only add "power" once for each
7793 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7796 } while (sg != group_head);
7798 #endif /* CONFIG_NUMA */
7801 /* Free memory allocated for various sched_group structures */
7802 static void free_sched_groups(const struct cpumask *cpu_map,
7803 struct cpumask *nodemask)
7807 for_each_cpu(cpu, cpu_map) {
7808 struct sched_group **sched_group_nodes
7809 = sched_group_nodes_bycpu[cpu];
7811 if (!sched_group_nodes)
7814 for (i = 0; i < nr_node_ids; i++) {
7815 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7817 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7818 if (cpumask_empty(nodemask))
7828 if (oldsg != sched_group_nodes[i])
7831 kfree(sched_group_nodes);
7832 sched_group_nodes_bycpu[cpu] = NULL;
7835 #else /* !CONFIG_NUMA */
7836 static void free_sched_groups(const struct cpumask *cpu_map,
7837 struct cpumask *nodemask)
7840 #endif /* CONFIG_NUMA */
7843 * Initialize sched groups cpu_power.
7845 * cpu_power indicates the capacity of sched group, which is used while
7846 * distributing the load between different sched groups in a sched domain.
7847 * Typically cpu_power for all the groups in a sched domain will be same unless
7848 * there are asymmetries in the topology. If there are asymmetries, group
7849 * having more cpu_power will pickup more load compared to the group having
7852 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7853 * the maximum number of tasks a group can handle in the presence of other idle
7854 * or lightly loaded groups in the same sched domain.
7856 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7858 struct sched_domain *child;
7859 struct sched_group *group;
7861 WARN_ON(!sd || !sd->groups);
7863 if (cpu != cpumask_first(sched_group_cpus(sd->groups)))
7868 sd->groups->__cpu_power = 0;
7871 * For perf policy, if the groups in child domain share resources
7872 * (for example cores sharing some portions of the cache hierarchy
7873 * or SMT), then set this domain groups cpu_power such that each group
7874 * can handle only one task, when there are other idle groups in the
7875 * same sched domain.
7877 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7879 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7880 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7885 * add cpu_power of each child group to this groups cpu_power
7887 group = child->groups;
7889 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7890 group = group->next;
7891 } while (group != child->groups);
7895 * Initializers for schedule domains
7896 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7899 #ifdef CONFIG_SCHED_DEBUG
7900 # define SD_INIT_NAME(sd, type) sd->name = #type
7902 # define SD_INIT_NAME(sd, type) do { } while (0)
7905 #define SD_INIT(sd, type) sd_init_##type(sd)
7907 #define SD_INIT_FUNC(type) \
7908 static noinline void sd_init_##type(struct sched_domain *sd) \
7910 memset(sd, 0, sizeof(*sd)); \
7911 *sd = SD_##type##_INIT; \
7912 sd->level = SD_LV_##type; \
7913 SD_INIT_NAME(sd, type); \
7918 SD_INIT_FUNC(ALLNODES)
7921 #ifdef CONFIG_SCHED_SMT
7922 SD_INIT_FUNC(SIBLING)
7924 #ifdef CONFIG_SCHED_MC
7928 static int default_relax_domain_level = -1;
7930 static int __init setup_relax_domain_level(char *str)
7934 val = simple_strtoul(str, NULL, 0);
7935 if (val < SD_LV_MAX)
7936 default_relax_domain_level = val;
7940 __setup("relax_domain_level=", setup_relax_domain_level);
7942 static void set_domain_attribute(struct sched_domain *sd,
7943 struct sched_domain_attr *attr)
7947 if (!attr || attr->relax_domain_level < 0) {
7948 if (default_relax_domain_level < 0)
7951 request = default_relax_domain_level;
7953 request = attr->relax_domain_level;
7954 if (request < sd->level) {
7955 /* turn off idle balance on this domain */
7956 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7958 /* turn on idle balance on this domain */
7959 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7964 * Build sched domains for a given set of cpus and attach the sched domains
7965 * to the individual cpus
7967 static int __build_sched_domains(const struct cpumask *cpu_map,
7968 struct sched_domain_attr *attr)
7970 int i, err = -ENOMEM;
7971 struct root_domain *rd;
7972 cpumask_var_t nodemask, this_sibling_map, this_core_map, send_covered,
7975 cpumask_var_t domainspan, covered, notcovered;
7976 struct sched_group **sched_group_nodes = NULL;
7977 int sd_allnodes = 0;
7979 if (!alloc_cpumask_var(&domainspan, GFP_KERNEL))
7981 if (!alloc_cpumask_var(&covered, GFP_KERNEL))
7982 goto free_domainspan;
7983 if (!alloc_cpumask_var(¬covered, GFP_KERNEL))
7987 if (!alloc_cpumask_var(&nodemask, GFP_KERNEL))
7988 goto free_notcovered;
7989 if (!alloc_cpumask_var(&this_sibling_map, GFP_KERNEL))
7991 if (!alloc_cpumask_var(&this_core_map, GFP_KERNEL))
7992 goto free_this_sibling_map;
7993 if (!alloc_cpumask_var(&send_covered, GFP_KERNEL))
7994 goto free_this_core_map;
7995 if (!alloc_cpumask_var(&tmpmask, GFP_KERNEL))
7996 goto free_send_covered;
8000 * Allocate the per-node list of sched groups
8002 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
8004 if (!sched_group_nodes) {
8005 printk(KERN_WARNING "Can not alloc sched group node list\n");
8010 rd = alloc_rootdomain();
8012 printk(KERN_WARNING "Cannot alloc root domain\n");
8013 goto free_sched_groups;
8017 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = sched_group_nodes;
8021 * Set up domains for cpus specified by the cpu_map.
8023 for_each_cpu(i, cpu_map) {
8024 struct sched_domain *sd = NULL, *p;
8026 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(i)), cpu_map);
8029 if (cpumask_weight(cpu_map) >
8030 SD_NODES_PER_DOMAIN*cpumask_weight(nodemask)) {
8031 sd = &per_cpu(allnodes_domains, i).sd;
8032 SD_INIT(sd, ALLNODES);
8033 set_domain_attribute(sd, attr);
8034 cpumask_copy(sched_domain_span(sd), cpu_map);
8035 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
8041 sd = &per_cpu(node_domains, i).sd;
8043 set_domain_attribute(sd, attr);
8044 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8048 cpumask_and(sched_domain_span(sd),
8049 sched_domain_span(sd), cpu_map);
8053 sd = &per_cpu(phys_domains, i).sd;
8055 set_domain_attribute(sd, attr);
8056 cpumask_copy(sched_domain_span(sd), nodemask);
8060 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
8062 #ifdef CONFIG_SCHED_MC
8064 sd = &per_cpu(core_domains, i).sd;
8066 set_domain_attribute(sd, attr);
8067 cpumask_and(sched_domain_span(sd), cpu_map,
8068 cpu_coregroup_mask(i));
8071 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
8074 #ifdef CONFIG_SCHED_SMT
8076 sd = &per_cpu(cpu_domains, i).sd;
8077 SD_INIT(sd, SIBLING);
8078 set_domain_attribute(sd, attr);
8079 cpumask_and(sched_domain_span(sd),
8080 topology_thread_cpumask(i), cpu_map);
8083 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
8087 #ifdef CONFIG_SCHED_SMT
8088 /* Set up CPU (sibling) groups */
8089 for_each_cpu(i, cpu_map) {
8090 cpumask_and(this_sibling_map,
8091 topology_thread_cpumask(i), cpu_map);
8092 if (i != cpumask_first(this_sibling_map))
8095 init_sched_build_groups(this_sibling_map, cpu_map,
8097 send_covered, tmpmask);
8101 #ifdef CONFIG_SCHED_MC
8102 /* Set up multi-core groups */
8103 for_each_cpu(i, cpu_map) {
8104 cpumask_and(this_core_map, cpu_coregroup_mask(i), cpu_map);
8105 if (i != cpumask_first(this_core_map))
8108 init_sched_build_groups(this_core_map, cpu_map,
8110 send_covered, tmpmask);
8114 /* Set up physical groups */
8115 for (i = 0; i < nr_node_ids; i++) {
8116 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8117 if (cpumask_empty(nodemask))
8120 init_sched_build_groups(nodemask, cpu_map,
8122 send_covered, tmpmask);
8126 /* Set up node groups */
8128 init_sched_build_groups(cpu_map, cpu_map,
8129 &cpu_to_allnodes_group,
8130 send_covered, tmpmask);
8133 for (i = 0; i < nr_node_ids; i++) {
8134 /* Set up node groups */
8135 struct sched_group *sg, *prev;
8138 cpumask_clear(covered);
8139 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8140 if (cpumask_empty(nodemask)) {
8141 sched_group_nodes[i] = NULL;
8145 sched_domain_node_span(i, domainspan);
8146 cpumask_and(domainspan, domainspan, cpu_map);
8148 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8151 printk(KERN_WARNING "Can not alloc domain group for "
8155 sched_group_nodes[i] = sg;
8156 for_each_cpu(j, nodemask) {
8157 struct sched_domain *sd;
8159 sd = &per_cpu(node_domains, j).sd;
8162 sg->__cpu_power = 0;
8163 cpumask_copy(sched_group_cpus(sg), nodemask);
8165 cpumask_or(covered, covered, nodemask);
8168 for (j = 0; j < nr_node_ids; j++) {
8169 int n = (i + j) % nr_node_ids;
8171 cpumask_complement(notcovered, covered);
8172 cpumask_and(tmpmask, notcovered, cpu_map);
8173 cpumask_and(tmpmask, tmpmask, domainspan);
8174 if (cpumask_empty(tmpmask))
8177 cpumask_and(tmpmask, tmpmask, cpumask_of_node(n));
8178 if (cpumask_empty(tmpmask))
8181 sg = kmalloc_node(sizeof(struct sched_group) +
8186 "Can not alloc domain group for node %d\n", j);
8189 sg->__cpu_power = 0;
8190 cpumask_copy(sched_group_cpus(sg), tmpmask);
8191 sg->next = prev->next;
8192 cpumask_or(covered, covered, tmpmask);
8199 /* Calculate CPU power for physical packages and nodes */
8200 #ifdef CONFIG_SCHED_SMT
8201 for_each_cpu(i, cpu_map) {
8202 struct sched_domain *sd = &per_cpu(cpu_domains, i).sd;
8204 init_sched_groups_power(i, sd);
8207 #ifdef CONFIG_SCHED_MC
8208 for_each_cpu(i, cpu_map) {
8209 struct sched_domain *sd = &per_cpu(core_domains, i).sd;
8211 init_sched_groups_power(i, sd);
8215 for_each_cpu(i, cpu_map) {
8216 struct sched_domain *sd = &per_cpu(phys_domains, i).sd;
8218 init_sched_groups_power(i, sd);
8222 for (i = 0; i < nr_node_ids; i++)
8223 init_numa_sched_groups_power(sched_group_nodes[i]);
8226 struct sched_group *sg;
8228 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8230 init_numa_sched_groups_power(sg);
8234 /* Attach the domains */
8235 for_each_cpu(i, cpu_map) {
8236 struct sched_domain *sd;
8237 #ifdef CONFIG_SCHED_SMT
8238 sd = &per_cpu(cpu_domains, i).sd;
8239 #elif defined(CONFIG_SCHED_MC)
8240 sd = &per_cpu(core_domains, i).sd;
8242 sd = &per_cpu(phys_domains, i).sd;
8244 cpu_attach_domain(sd, rd, i);
8250 free_cpumask_var(tmpmask);
8252 free_cpumask_var(send_covered);
8254 free_cpumask_var(this_core_map);
8255 free_this_sibling_map:
8256 free_cpumask_var(this_sibling_map);
8258 free_cpumask_var(nodemask);
8261 free_cpumask_var(notcovered);
8263 free_cpumask_var(covered);
8265 free_cpumask_var(domainspan);
8272 kfree(sched_group_nodes);
8278 free_sched_groups(cpu_map, tmpmask);
8279 free_rootdomain(rd);
8284 static int build_sched_domains(const struct cpumask *cpu_map)
8286 return __build_sched_domains(cpu_map, NULL);
8289 static struct cpumask *doms_cur; /* current sched domains */
8290 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8291 static struct sched_domain_attr *dattr_cur;
8292 /* attribues of custom domains in 'doms_cur' */
8295 * Special case: If a kmalloc of a doms_cur partition (array of
8296 * cpumask) fails, then fallback to a single sched domain,
8297 * as determined by the single cpumask fallback_doms.
8299 static cpumask_var_t fallback_doms;
8302 * arch_update_cpu_topology lets virtualized architectures update the
8303 * cpu core maps. It is supposed to return 1 if the topology changed
8304 * or 0 if it stayed the same.
8306 int __attribute__((weak)) arch_update_cpu_topology(void)
8312 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8313 * For now this just excludes isolated cpus, but could be used to
8314 * exclude other special cases in the future.
8316 static int arch_init_sched_domains(const struct cpumask *cpu_map)
8320 arch_update_cpu_topology();
8322 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
8324 doms_cur = fallback_doms;
8325 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
8327 err = build_sched_domains(doms_cur);
8328 register_sched_domain_sysctl();
8333 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
8334 struct cpumask *tmpmask)
8336 free_sched_groups(cpu_map, tmpmask);
8340 * Detach sched domains from a group of cpus specified in cpu_map
8341 * These cpus will now be attached to the NULL domain
8343 static void detach_destroy_domains(const struct cpumask *cpu_map)
8345 /* Save because hotplug lock held. */
8346 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
8349 for_each_cpu(i, cpu_map)
8350 cpu_attach_domain(NULL, &def_root_domain, i);
8351 synchronize_sched();
8352 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
8355 /* handle null as "default" */
8356 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
8357 struct sched_domain_attr *new, int idx_new)
8359 struct sched_domain_attr tmp;
8366 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8367 new ? (new + idx_new) : &tmp,
8368 sizeof(struct sched_domain_attr));
8372 * Partition sched domains as specified by the 'ndoms_new'
8373 * cpumasks in the array doms_new[] of cpumasks. This compares
8374 * doms_new[] to the current sched domain partitioning, doms_cur[].
8375 * It destroys each deleted domain and builds each new domain.
8377 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8378 * The masks don't intersect (don't overlap.) We should setup one
8379 * sched domain for each mask. CPUs not in any of the cpumasks will
8380 * not be load balanced. If the same cpumask appears both in the
8381 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8384 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8385 * ownership of it and will kfree it when done with it. If the caller
8386 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8387 * ndoms_new == 1, and partition_sched_domains() will fallback to
8388 * the single partition 'fallback_doms', it also forces the domains
8391 * If doms_new == NULL it will be replaced with cpu_online_mask.
8392 * ndoms_new == 0 is a special case for destroying existing domains,
8393 * and it will not create the default domain.
8395 * Call with hotplug lock held
8397 /* FIXME: Change to struct cpumask *doms_new[] */
8398 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
8399 struct sched_domain_attr *dattr_new)
8404 mutex_lock(&sched_domains_mutex);
8406 /* always unregister in case we don't destroy any domains */
8407 unregister_sched_domain_sysctl();
8409 /* Let architecture update cpu core mappings. */
8410 new_topology = arch_update_cpu_topology();
8412 n = doms_new ? ndoms_new : 0;
8414 /* Destroy deleted domains */
8415 for (i = 0; i < ndoms_cur; i++) {
8416 for (j = 0; j < n && !new_topology; j++) {
8417 if (cpumask_equal(&doms_cur[i], &doms_new[j])
8418 && dattrs_equal(dattr_cur, i, dattr_new, j))
8421 /* no match - a current sched domain not in new doms_new[] */
8422 detach_destroy_domains(doms_cur + i);
8427 if (doms_new == NULL) {
8429 doms_new = fallback_doms;
8430 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
8431 WARN_ON_ONCE(dattr_new);
8434 /* Build new domains */
8435 for (i = 0; i < ndoms_new; i++) {
8436 for (j = 0; j < ndoms_cur && !new_topology; j++) {
8437 if (cpumask_equal(&doms_new[i], &doms_cur[j])
8438 && dattrs_equal(dattr_new, i, dattr_cur, j))
8441 /* no match - add a new doms_new */
8442 __build_sched_domains(doms_new + i,
8443 dattr_new ? dattr_new + i : NULL);
8448 /* Remember the new sched domains */
8449 if (doms_cur != fallback_doms)
8451 kfree(dattr_cur); /* kfree(NULL) is safe */
8452 doms_cur = doms_new;
8453 dattr_cur = dattr_new;
8454 ndoms_cur = ndoms_new;
8456 register_sched_domain_sysctl();
8458 mutex_unlock(&sched_domains_mutex);
8461 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8462 static void arch_reinit_sched_domains(void)
8466 /* Destroy domains first to force the rebuild */
8467 partition_sched_domains(0, NULL, NULL);
8469 rebuild_sched_domains();
8473 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
8475 unsigned int level = 0;
8477 if (sscanf(buf, "%u", &level) != 1)
8481 * level is always be positive so don't check for
8482 * level < POWERSAVINGS_BALANCE_NONE which is 0
8483 * What happens on 0 or 1 byte write,
8484 * need to check for count as well?
8487 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
8491 sched_smt_power_savings = level;
8493 sched_mc_power_savings = level;
8495 arch_reinit_sched_domains();
8500 #ifdef CONFIG_SCHED_MC
8501 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
8504 return sprintf(page, "%u\n", sched_mc_power_savings);
8506 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
8507 const char *buf, size_t count)
8509 return sched_power_savings_store(buf, count, 0);
8511 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
8512 sched_mc_power_savings_show,
8513 sched_mc_power_savings_store);
8516 #ifdef CONFIG_SCHED_SMT
8517 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
8520 return sprintf(page, "%u\n", sched_smt_power_savings);
8522 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
8523 const char *buf, size_t count)
8525 return sched_power_savings_store(buf, count, 1);
8527 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
8528 sched_smt_power_savings_show,
8529 sched_smt_power_savings_store);
8532 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
8536 #ifdef CONFIG_SCHED_SMT
8538 err = sysfs_create_file(&cls->kset.kobj,
8539 &attr_sched_smt_power_savings.attr);
8541 #ifdef CONFIG_SCHED_MC
8542 if (!err && mc_capable())
8543 err = sysfs_create_file(&cls->kset.kobj,
8544 &attr_sched_mc_power_savings.attr);
8548 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8550 #ifndef CONFIG_CPUSETS
8552 * Add online and remove offline CPUs from the scheduler domains.
8553 * When cpusets are enabled they take over this function.
8555 static int update_sched_domains(struct notifier_block *nfb,
8556 unsigned long action, void *hcpu)
8560 case CPU_ONLINE_FROZEN:
8562 case CPU_DEAD_FROZEN:
8563 partition_sched_domains(1, NULL, NULL);
8572 static int update_runtime(struct notifier_block *nfb,
8573 unsigned long action, void *hcpu)
8575 int cpu = (int)(long)hcpu;
8578 case CPU_DOWN_PREPARE:
8579 case CPU_DOWN_PREPARE_FROZEN:
8580 disable_runtime(cpu_rq(cpu));
8583 case CPU_DOWN_FAILED:
8584 case CPU_DOWN_FAILED_FROZEN:
8586 case CPU_ONLINE_FROZEN:
8587 enable_runtime(cpu_rq(cpu));
8595 void __init sched_init_smp(void)
8597 cpumask_var_t non_isolated_cpus;
8599 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
8601 #if defined(CONFIG_NUMA)
8602 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
8604 BUG_ON(sched_group_nodes_bycpu == NULL);
8607 mutex_lock(&sched_domains_mutex);
8608 arch_init_sched_domains(cpu_online_mask);
8609 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
8610 if (cpumask_empty(non_isolated_cpus))
8611 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
8612 mutex_unlock(&sched_domains_mutex);
8615 #ifndef CONFIG_CPUSETS
8616 /* XXX: Theoretical race here - CPU may be hotplugged now */
8617 hotcpu_notifier(update_sched_domains, 0);
8620 /* RT runtime code needs to handle some hotplug events */
8621 hotcpu_notifier(update_runtime, 0);
8625 /* Move init over to a non-isolated CPU */
8626 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
8628 sched_init_granularity();
8629 free_cpumask_var(non_isolated_cpus);
8631 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
8632 init_sched_rt_class();
8635 void __init sched_init_smp(void)
8637 sched_init_granularity();
8639 #endif /* CONFIG_SMP */
8641 int in_sched_functions(unsigned long addr)
8643 return in_lock_functions(addr) ||
8644 (addr >= (unsigned long)__sched_text_start
8645 && addr < (unsigned long)__sched_text_end);
8648 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
8650 cfs_rq->tasks_timeline = RB_ROOT;
8651 INIT_LIST_HEAD(&cfs_rq->tasks);
8652 #ifdef CONFIG_FAIR_GROUP_SCHED
8655 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8658 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8660 struct rt_prio_array *array;
8663 array = &rt_rq->active;
8664 for (i = 0; i < MAX_RT_PRIO; i++) {
8665 INIT_LIST_HEAD(array->queue + i);
8666 __clear_bit(i, array->bitmap);
8668 /* delimiter for bitsearch: */
8669 __set_bit(MAX_RT_PRIO, array->bitmap);
8671 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8672 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8674 rt_rq->highest_prio.next = MAX_RT_PRIO;
8678 rt_rq->rt_nr_migratory = 0;
8679 rt_rq->overloaded = 0;
8680 plist_head_init(&rq->rt.pushable_tasks, &rq->lock);
8684 rt_rq->rt_throttled = 0;
8685 rt_rq->rt_runtime = 0;
8686 spin_lock_init(&rt_rq->rt_runtime_lock);
8688 #ifdef CONFIG_RT_GROUP_SCHED
8689 rt_rq->rt_nr_boosted = 0;
8694 #ifdef CONFIG_FAIR_GROUP_SCHED
8695 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8696 struct sched_entity *se, int cpu, int add,
8697 struct sched_entity *parent)
8699 struct rq *rq = cpu_rq(cpu);
8700 tg->cfs_rq[cpu] = cfs_rq;
8701 init_cfs_rq(cfs_rq, rq);
8704 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8707 /* se could be NULL for init_task_group */
8712 se->cfs_rq = &rq->cfs;
8714 se->cfs_rq = parent->my_q;
8717 se->load.weight = tg->shares;
8718 se->load.inv_weight = 0;
8719 se->parent = parent;
8723 #ifdef CONFIG_RT_GROUP_SCHED
8724 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8725 struct sched_rt_entity *rt_se, int cpu, int add,
8726 struct sched_rt_entity *parent)
8728 struct rq *rq = cpu_rq(cpu);
8730 tg->rt_rq[cpu] = rt_rq;
8731 init_rt_rq(rt_rq, rq);
8733 rt_rq->rt_se = rt_se;
8734 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8736 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8738 tg->rt_se[cpu] = rt_se;
8743 rt_se->rt_rq = &rq->rt;
8745 rt_se->rt_rq = parent->my_q;
8747 rt_se->my_q = rt_rq;
8748 rt_se->parent = parent;
8749 INIT_LIST_HEAD(&rt_se->run_list);
8753 void __init sched_init(void)
8756 unsigned long alloc_size = 0, ptr;
8758 #ifdef CONFIG_FAIR_GROUP_SCHED
8759 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8761 #ifdef CONFIG_RT_GROUP_SCHED
8762 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8764 #ifdef CONFIG_USER_SCHED
8767 #ifdef CONFIG_CPUMASK_OFFSTACK
8768 alloc_size += num_possible_cpus() * cpumask_size();
8771 * As sched_init() is called before page_alloc is setup,
8772 * we use alloc_bootmem().
8775 ptr = (unsigned long)alloc_bootmem(alloc_size);
8777 #ifdef CONFIG_FAIR_GROUP_SCHED
8778 init_task_group.se = (struct sched_entity **)ptr;
8779 ptr += nr_cpu_ids * sizeof(void **);
8781 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8782 ptr += nr_cpu_ids * sizeof(void **);
8784 #ifdef CONFIG_USER_SCHED
8785 root_task_group.se = (struct sched_entity **)ptr;
8786 ptr += nr_cpu_ids * sizeof(void **);
8788 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8789 ptr += nr_cpu_ids * sizeof(void **);
8790 #endif /* CONFIG_USER_SCHED */
8791 #endif /* CONFIG_FAIR_GROUP_SCHED */
8792 #ifdef CONFIG_RT_GROUP_SCHED
8793 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8794 ptr += nr_cpu_ids * sizeof(void **);
8796 init_task_group.rt_rq = (struct rt_rq **)ptr;
8797 ptr += nr_cpu_ids * sizeof(void **);
8799 #ifdef CONFIG_USER_SCHED
8800 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8801 ptr += nr_cpu_ids * sizeof(void **);
8803 root_task_group.rt_rq = (struct rt_rq **)ptr;
8804 ptr += nr_cpu_ids * sizeof(void **);
8805 #endif /* CONFIG_USER_SCHED */
8806 #endif /* CONFIG_RT_GROUP_SCHED */
8807 #ifdef CONFIG_CPUMASK_OFFSTACK
8808 for_each_possible_cpu(i) {
8809 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
8810 ptr += cpumask_size();
8812 #endif /* CONFIG_CPUMASK_OFFSTACK */
8816 init_defrootdomain();
8819 init_rt_bandwidth(&def_rt_bandwidth,
8820 global_rt_period(), global_rt_runtime());
8822 #ifdef CONFIG_RT_GROUP_SCHED
8823 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8824 global_rt_period(), global_rt_runtime());
8825 #ifdef CONFIG_USER_SCHED
8826 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8827 global_rt_period(), RUNTIME_INF);
8828 #endif /* CONFIG_USER_SCHED */
8829 #endif /* CONFIG_RT_GROUP_SCHED */
8831 #ifdef CONFIG_GROUP_SCHED
8832 list_add(&init_task_group.list, &task_groups);
8833 INIT_LIST_HEAD(&init_task_group.children);
8835 #ifdef CONFIG_USER_SCHED
8836 INIT_LIST_HEAD(&root_task_group.children);
8837 init_task_group.parent = &root_task_group;
8838 list_add(&init_task_group.siblings, &root_task_group.children);
8839 #endif /* CONFIG_USER_SCHED */
8840 #endif /* CONFIG_GROUP_SCHED */
8842 for_each_possible_cpu(i) {
8846 spin_lock_init(&rq->lock);
8848 init_cfs_rq(&rq->cfs, rq);
8849 init_rt_rq(&rq->rt, rq);
8850 #ifdef CONFIG_FAIR_GROUP_SCHED
8851 init_task_group.shares = init_task_group_load;
8852 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8853 #ifdef CONFIG_CGROUP_SCHED
8855 * How much cpu bandwidth does init_task_group get?
8857 * In case of task-groups formed thr' the cgroup filesystem, it
8858 * gets 100% of the cpu resources in the system. This overall
8859 * system cpu resource is divided among the tasks of
8860 * init_task_group and its child task-groups in a fair manner,
8861 * based on each entity's (task or task-group's) weight
8862 * (se->load.weight).
8864 * In other words, if init_task_group has 10 tasks of weight
8865 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8866 * then A0's share of the cpu resource is:
8868 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8870 * We achieve this by letting init_task_group's tasks sit
8871 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8873 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8874 #elif defined CONFIG_USER_SCHED
8875 root_task_group.shares = NICE_0_LOAD;
8876 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8878 * In case of task-groups formed thr' the user id of tasks,
8879 * init_task_group represents tasks belonging to root user.
8880 * Hence it forms a sibling of all subsequent groups formed.
8881 * In this case, init_task_group gets only a fraction of overall
8882 * system cpu resource, based on the weight assigned to root
8883 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8884 * by letting tasks of init_task_group sit in a separate cfs_rq
8885 * (init_cfs_rq) and having one entity represent this group of
8886 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8888 init_tg_cfs_entry(&init_task_group,
8889 &per_cpu(init_cfs_rq, i),
8890 &per_cpu(init_sched_entity, i), i, 1,
8891 root_task_group.se[i]);
8894 #endif /* CONFIG_FAIR_GROUP_SCHED */
8896 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8897 #ifdef CONFIG_RT_GROUP_SCHED
8898 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8899 #ifdef CONFIG_CGROUP_SCHED
8900 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8901 #elif defined CONFIG_USER_SCHED
8902 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8903 init_tg_rt_entry(&init_task_group,
8904 &per_cpu(init_rt_rq, i),
8905 &per_cpu(init_sched_rt_entity, i), i, 1,
8906 root_task_group.rt_se[i]);
8910 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8911 rq->cpu_load[j] = 0;
8915 rq->active_balance = 0;
8916 rq->next_balance = jiffies;
8920 rq->migration_thread = NULL;
8921 INIT_LIST_HEAD(&rq->migration_queue);
8922 rq_attach_root(rq, &def_root_domain);
8925 atomic_set(&rq->nr_iowait, 0);
8928 set_load_weight(&init_task);
8930 #ifdef CONFIG_PREEMPT_NOTIFIERS
8931 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8935 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8938 #ifdef CONFIG_RT_MUTEXES
8939 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8943 * The boot idle thread does lazy MMU switching as well:
8945 atomic_inc(&init_mm.mm_count);
8946 enter_lazy_tlb(&init_mm, current);
8949 * Make us the idle thread. Technically, schedule() should not be
8950 * called from this thread, however somewhere below it might be,
8951 * but because we are the idle thread, we just pick up running again
8952 * when this runqueue becomes "idle".
8954 init_idle(current, smp_processor_id());
8956 * During early bootup we pretend to be a normal task:
8958 current->sched_class = &fair_sched_class;
8960 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8961 alloc_bootmem_cpumask_var(&nohz_cpu_mask);
8964 alloc_bootmem_cpumask_var(&nohz.cpu_mask);
8966 alloc_bootmem_cpumask_var(&cpu_isolated_map);
8969 scheduler_running = 1;
8972 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8973 void __might_sleep(char *file, int line)
8976 static unsigned long prev_jiffy; /* ratelimiting */
8978 if ((!in_atomic() && !irqs_disabled()) ||
8979 system_state != SYSTEM_RUNNING || oops_in_progress)
8981 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8983 prev_jiffy = jiffies;
8986 "BUG: sleeping function called from invalid context at %s:%d\n",
8989 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8990 in_atomic(), irqs_disabled(),
8991 current->pid, current->comm);
8993 debug_show_held_locks(current);
8994 if (irqs_disabled())
8995 print_irqtrace_events(current);
8999 EXPORT_SYMBOL(__might_sleep);
9002 #ifdef CONFIG_MAGIC_SYSRQ
9003 static void normalize_task(struct rq *rq, struct task_struct *p)
9007 update_rq_clock(rq);
9008 on_rq = p->se.on_rq;
9010 deactivate_task(rq, p, 0);
9011 __setscheduler(rq, p, SCHED_NORMAL, 0);
9013 activate_task(rq, p, 0);
9014 resched_task(rq->curr);
9018 void normalize_rt_tasks(void)
9020 struct task_struct *g, *p;
9021 unsigned long flags;
9024 read_lock_irqsave(&tasklist_lock, flags);
9025 do_each_thread(g, p) {
9027 * Only normalize user tasks:
9032 p->se.exec_start = 0;
9033 #ifdef CONFIG_SCHEDSTATS
9034 p->se.wait_start = 0;
9035 p->se.sleep_start = 0;
9036 p->se.block_start = 0;
9041 * Renice negative nice level userspace
9044 if (TASK_NICE(p) < 0 && p->mm)
9045 set_user_nice(p, 0);
9049 spin_lock(&p->pi_lock);
9050 rq = __task_rq_lock(p);
9052 normalize_task(rq, p);
9054 __task_rq_unlock(rq);
9055 spin_unlock(&p->pi_lock);
9056 } while_each_thread(g, p);
9058 read_unlock_irqrestore(&tasklist_lock, flags);
9061 #endif /* CONFIG_MAGIC_SYSRQ */
9065 * These functions are only useful for the IA64 MCA handling.
9067 * They can only be called when the whole system has been
9068 * stopped - every CPU needs to be quiescent, and no scheduling
9069 * activity can take place. Using them for anything else would
9070 * be a serious bug, and as a result, they aren't even visible
9071 * under any other configuration.
9075 * curr_task - return the current task for a given cpu.
9076 * @cpu: the processor in question.
9078 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9080 struct task_struct *curr_task(int cpu)
9082 return cpu_curr(cpu);
9086 * set_curr_task - set the current task for a given cpu.
9087 * @cpu: the processor in question.
9088 * @p: the task pointer to set.
9090 * Description: This function must only be used when non-maskable interrupts
9091 * are serviced on a separate stack. It allows the architecture to switch the
9092 * notion of the current task on a cpu in a non-blocking manner. This function
9093 * must be called with all CPU's synchronized, and interrupts disabled, the
9094 * and caller must save the original value of the current task (see
9095 * curr_task() above) and restore that value before reenabling interrupts and
9096 * re-starting the system.
9098 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9100 void set_curr_task(int cpu, struct task_struct *p)
9107 #ifdef CONFIG_FAIR_GROUP_SCHED
9108 static void free_fair_sched_group(struct task_group *tg)
9112 for_each_possible_cpu(i) {
9114 kfree(tg->cfs_rq[i]);
9124 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9126 struct cfs_rq *cfs_rq;
9127 struct sched_entity *se;
9131 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9134 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9138 tg->shares = NICE_0_LOAD;
9140 for_each_possible_cpu(i) {
9143 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9144 GFP_KERNEL, cpu_to_node(i));
9148 se = kzalloc_node(sizeof(struct sched_entity),
9149 GFP_KERNEL, cpu_to_node(i));
9153 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9162 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9164 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9165 &cpu_rq(cpu)->leaf_cfs_rq_list);
9168 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9170 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9172 #else /* !CONFG_FAIR_GROUP_SCHED */
9173 static inline void free_fair_sched_group(struct task_group *tg)
9178 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9183 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9187 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9190 #endif /* CONFIG_FAIR_GROUP_SCHED */
9192 #ifdef CONFIG_RT_GROUP_SCHED
9193 static void free_rt_sched_group(struct task_group *tg)
9197 destroy_rt_bandwidth(&tg->rt_bandwidth);
9199 for_each_possible_cpu(i) {
9201 kfree(tg->rt_rq[i]);
9203 kfree(tg->rt_se[i]);
9211 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9213 struct rt_rq *rt_rq;
9214 struct sched_rt_entity *rt_se;
9218 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9221 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9225 init_rt_bandwidth(&tg->rt_bandwidth,
9226 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9228 for_each_possible_cpu(i) {
9231 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9232 GFP_KERNEL, cpu_to_node(i));
9236 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9237 GFP_KERNEL, cpu_to_node(i));
9241 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9250 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9252 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9253 &cpu_rq(cpu)->leaf_rt_rq_list);
9256 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9258 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9260 #else /* !CONFIG_RT_GROUP_SCHED */
9261 static inline void free_rt_sched_group(struct task_group *tg)
9266 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9271 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9275 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9278 #endif /* CONFIG_RT_GROUP_SCHED */
9280 #ifdef CONFIG_GROUP_SCHED
9281 static void free_sched_group(struct task_group *tg)
9283 free_fair_sched_group(tg);
9284 free_rt_sched_group(tg);
9288 /* allocate runqueue etc for a new task group */
9289 struct task_group *sched_create_group(struct task_group *parent)
9291 struct task_group *tg;
9292 unsigned long flags;
9295 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
9297 return ERR_PTR(-ENOMEM);
9299 if (!alloc_fair_sched_group(tg, parent))
9302 if (!alloc_rt_sched_group(tg, parent))
9305 spin_lock_irqsave(&task_group_lock, flags);
9306 for_each_possible_cpu(i) {
9307 register_fair_sched_group(tg, i);
9308 register_rt_sched_group(tg, i);
9310 list_add_rcu(&tg->list, &task_groups);
9312 WARN_ON(!parent); /* root should already exist */
9314 tg->parent = parent;
9315 INIT_LIST_HEAD(&tg->children);
9316 list_add_rcu(&tg->siblings, &parent->children);
9317 spin_unlock_irqrestore(&task_group_lock, flags);
9322 free_sched_group(tg);
9323 return ERR_PTR(-ENOMEM);
9326 /* rcu callback to free various structures associated with a task group */
9327 static void free_sched_group_rcu(struct rcu_head *rhp)
9329 /* now it should be safe to free those cfs_rqs */
9330 free_sched_group(container_of(rhp, struct task_group, rcu));
9333 /* Destroy runqueue etc associated with a task group */
9334 void sched_destroy_group(struct task_group *tg)
9336 unsigned long flags;
9339 spin_lock_irqsave(&task_group_lock, flags);
9340 for_each_possible_cpu(i) {
9341 unregister_fair_sched_group(tg, i);
9342 unregister_rt_sched_group(tg, i);
9344 list_del_rcu(&tg->list);
9345 list_del_rcu(&tg->siblings);
9346 spin_unlock_irqrestore(&task_group_lock, flags);
9348 /* wait for possible concurrent references to cfs_rqs complete */
9349 call_rcu(&tg->rcu, free_sched_group_rcu);
9352 /* change task's runqueue when it moves between groups.
9353 * The caller of this function should have put the task in its new group
9354 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9355 * reflect its new group.
9357 void sched_move_task(struct task_struct *tsk)
9360 unsigned long flags;
9363 rq = task_rq_lock(tsk, &flags);
9365 update_rq_clock(rq);
9367 running = task_current(rq, tsk);
9368 on_rq = tsk->se.on_rq;
9371 dequeue_task(rq, tsk, 0);
9372 if (unlikely(running))
9373 tsk->sched_class->put_prev_task(rq, tsk);
9375 set_task_rq(tsk, task_cpu(tsk));
9377 #ifdef CONFIG_FAIR_GROUP_SCHED
9378 if (tsk->sched_class->moved_group)
9379 tsk->sched_class->moved_group(tsk);
9382 if (unlikely(running))
9383 tsk->sched_class->set_curr_task(rq);
9385 enqueue_task(rq, tsk, 0);
9387 task_rq_unlock(rq, &flags);
9389 #endif /* CONFIG_GROUP_SCHED */
9391 #ifdef CONFIG_FAIR_GROUP_SCHED
9392 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
9394 struct cfs_rq *cfs_rq = se->cfs_rq;
9399 dequeue_entity(cfs_rq, se, 0);
9401 se->load.weight = shares;
9402 se->load.inv_weight = 0;
9405 enqueue_entity(cfs_rq, se, 0);
9408 static void set_se_shares(struct sched_entity *se, unsigned long shares)
9410 struct cfs_rq *cfs_rq = se->cfs_rq;
9411 struct rq *rq = cfs_rq->rq;
9412 unsigned long flags;
9414 spin_lock_irqsave(&rq->lock, flags);
9415 __set_se_shares(se, shares);
9416 spin_unlock_irqrestore(&rq->lock, flags);
9419 static DEFINE_MUTEX(shares_mutex);
9421 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9424 unsigned long flags;
9427 * We can't change the weight of the root cgroup.
9432 if (shares < MIN_SHARES)
9433 shares = MIN_SHARES;
9434 else if (shares > MAX_SHARES)
9435 shares = MAX_SHARES;
9437 mutex_lock(&shares_mutex);
9438 if (tg->shares == shares)
9441 spin_lock_irqsave(&task_group_lock, flags);
9442 for_each_possible_cpu(i)
9443 unregister_fair_sched_group(tg, i);
9444 list_del_rcu(&tg->siblings);
9445 spin_unlock_irqrestore(&task_group_lock, flags);
9447 /* wait for any ongoing reference to this group to finish */
9448 synchronize_sched();
9451 * Now we are free to modify the group's share on each cpu
9452 * w/o tripping rebalance_share or load_balance_fair.
9454 tg->shares = shares;
9455 for_each_possible_cpu(i) {
9459 cfs_rq_set_shares(tg->cfs_rq[i], 0);
9460 set_se_shares(tg->se[i], shares);
9464 * Enable load balance activity on this group, by inserting it back on
9465 * each cpu's rq->leaf_cfs_rq_list.
9467 spin_lock_irqsave(&task_group_lock, flags);
9468 for_each_possible_cpu(i)
9469 register_fair_sched_group(tg, i);
9470 list_add_rcu(&tg->siblings, &tg->parent->children);
9471 spin_unlock_irqrestore(&task_group_lock, flags);
9473 mutex_unlock(&shares_mutex);
9477 unsigned long sched_group_shares(struct task_group *tg)
9483 #ifdef CONFIG_RT_GROUP_SCHED
9485 * Ensure that the real time constraints are schedulable.
9487 static DEFINE_MUTEX(rt_constraints_mutex);
9489 static unsigned long to_ratio(u64 period, u64 runtime)
9491 if (runtime == RUNTIME_INF)
9494 return div64_u64(runtime << 20, period);
9497 /* Must be called with tasklist_lock held */
9498 static inline int tg_has_rt_tasks(struct task_group *tg)
9500 struct task_struct *g, *p;
9502 do_each_thread(g, p) {
9503 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
9505 } while_each_thread(g, p);
9510 struct rt_schedulable_data {
9511 struct task_group *tg;
9516 static int tg_schedulable(struct task_group *tg, void *data)
9518 struct rt_schedulable_data *d = data;
9519 struct task_group *child;
9520 unsigned long total, sum = 0;
9521 u64 period, runtime;
9523 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9524 runtime = tg->rt_bandwidth.rt_runtime;
9527 period = d->rt_period;
9528 runtime = d->rt_runtime;
9531 #ifdef CONFIG_USER_SCHED
9532 if (tg == &root_task_group) {
9533 period = global_rt_period();
9534 runtime = global_rt_runtime();
9539 * Cannot have more runtime than the period.
9541 if (runtime > period && runtime != RUNTIME_INF)
9545 * Ensure we don't starve existing RT tasks.
9547 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
9550 total = to_ratio(period, runtime);
9553 * Nobody can have more than the global setting allows.
9555 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
9559 * The sum of our children's runtime should not exceed our own.
9561 list_for_each_entry_rcu(child, &tg->children, siblings) {
9562 period = ktime_to_ns(child->rt_bandwidth.rt_period);
9563 runtime = child->rt_bandwidth.rt_runtime;
9565 if (child == d->tg) {
9566 period = d->rt_period;
9567 runtime = d->rt_runtime;
9570 sum += to_ratio(period, runtime);
9579 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
9581 struct rt_schedulable_data data = {
9583 .rt_period = period,
9584 .rt_runtime = runtime,
9587 return walk_tg_tree(tg_schedulable, tg_nop, &data);
9590 static int tg_set_bandwidth(struct task_group *tg,
9591 u64 rt_period, u64 rt_runtime)
9595 mutex_lock(&rt_constraints_mutex);
9596 read_lock(&tasklist_lock);
9597 err = __rt_schedulable(tg, rt_period, rt_runtime);
9601 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9602 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
9603 tg->rt_bandwidth.rt_runtime = rt_runtime;
9605 for_each_possible_cpu(i) {
9606 struct rt_rq *rt_rq = tg->rt_rq[i];
9608 spin_lock(&rt_rq->rt_runtime_lock);
9609 rt_rq->rt_runtime = rt_runtime;
9610 spin_unlock(&rt_rq->rt_runtime_lock);
9612 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9614 read_unlock(&tasklist_lock);
9615 mutex_unlock(&rt_constraints_mutex);
9620 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
9622 u64 rt_runtime, rt_period;
9624 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9625 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
9626 if (rt_runtime_us < 0)
9627 rt_runtime = RUNTIME_INF;
9629 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9632 long sched_group_rt_runtime(struct task_group *tg)
9636 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
9639 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
9640 do_div(rt_runtime_us, NSEC_PER_USEC);
9641 return rt_runtime_us;
9644 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
9646 u64 rt_runtime, rt_period;
9648 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
9649 rt_runtime = tg->rt_bandwidth.rt_runtime;
9654 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9657 long sched_group_rt_period(struct task_group *tg)
9661 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9662 do_div(rt_period_us, NSEC_PER_USEC);
9663 return rt_period_us;
9666 static int sched_rt_global_constraints(void)
9668 u64 runtime, period;
9671 if (sysctl_sched_rt_period <= 0)
9674 runtime = global_rt_runtime();
9675 period = global_rt_period();
9678 * Sanity check on the sysctl variables.
9680 if (runtime > period && runtime != RUNTIME_INF)
9683 mutex_lock(&rt_constraints_mutex);
9684 read_lock(&tasklist_lock);
9685 ret = __rt_schedulable(NULL, 0, 0);
9686 read_unlock(&tasklist_lock);
9687 mutex_unlock(&rt_constraints_mutex);
9692 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
9694 /* Don't accept realtime tasks when there is no way for them to run */
9695 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
9701 #else /* !CONFIG_RT_GROUP_SCHED */
9702 static int sched_rt_global_constraints(void)
9704 unsigned long flags;
9707 if (sysctl_sched_rt_period <= 0)
9710 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9711 for_each_possible_cpu(i) {
9712 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9714 spin_lock(&rt_rq->rt_runtime_lock);
9715 rt_rq->rt_runtime = global_rt_runtime();
9716 spin_unlock(&rt_rq->rt_runtime_lock);
9718 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9722 #endif /* CONFIG_RT_GROUP_SCHED */
9724 int sched_rt_handler(struct ctl_table *table, int write,
9725 struct file *filp, void __user *buffer, size_t *lenp,
9729 int old_period, old_runtime;
9730 static DEFINE_MUTEX(mutex);
9733 old_period = sysctl_sched_rt_period;
9734 old_runtime = sysctl_sched_rt_runtime;
9736 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
9738 if (!ret && write) {
9739 ret = sched_rt_global_constraints();
9741 sysctl_sched_rt_period = old_period;
9742 sysctl_sched_rt_runtime = old_runtime;
9744 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9745 def_rt_bandwidth.rt_period =
9746 ns_to_ktime(global_rt_period());
9749 mutex_unlock(&mutex);
9754 #ifdef CONFIG_CGROUP_SCHED
9756 /* return corresponding task_group object of a cgroup */
9757 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9759 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9760 struct task_group, css);
9763 static struct cgroup_subsys_state *
9764 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9766 struct task_group *tg, *parent;
9768 if (!cgrp->parent) {
9769 /* This is early initialization for the top cgroup */
9770 return &init_task_group.css;
9773 parent = cgroup_tg(cgrp->parent);
9774 tg = sched_create_group(parent);
9776 return ERR_PTR(-ENOMEM);
9782 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9784 struct task_group *tg = cgroup_tg(cgrp);
9786 sched_destroy_group(tg);
9790 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9791 struct task_struct *tsk)
9793 #ifdef CONFIG_RT_GROUP_SCHED
9794 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
9797 /* We don't support RT-tasks being in separate groups */
9798 if (tsk->sched_class != &fair_sched_class)
9806 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9807 struct cgroup *old_cont, struct task_struct *tsk)
9809 sched_move_task(tsk);
9812 #ifdef CONFIG_FAIR_GROUP_SCHED
9813 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9816 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9819 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9821 struct task_group *tg = cgroup_tg(cgrp);
9823 return (u64) tg->shares;
9825 #endif /* CONFIG_FAIR_GROUP_SCHED */
9827 #ifdef CONFIG_RT_GROUP_SCHED
9828 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9831 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9834 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9836 return sched_group_rt_runtime(cgroup_tg(cgrp));
9839 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9842 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9845 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9847 return sched_group_rt_period(cgroup_tg(cgrp));
9849 #endif /* CONFIG_RT_GROUP_SCHED */
9851 static struct cftype cpu_files[] = {
9852 #ifdef CONFIG_FAIR_GROUP_SCHED
9855 .read_u64 = cpu_shares_read_u64,
9856 .write_u64 = cpu_shares_write_u64,
9859 #ifdef CONFIG_RT_GROUP_SCHED
9861 .name = "rt_runtime_us",
9862 .read_s64 = cpu_rt_runtime_read,
9863 .write_s64 = cpu_rt_runtime_write,
9866 .name = "rt_period_us",
9867 .read_u64 = cpu_rt_period_read_uint,
9868 .write_u64 = cpu_rt_period_write_uint,
9873 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9875 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9878 struct cgroup_subsys cpu_cgroup_subsys = {
9880 .create = cpu_cgroup_create,
9881 .destroy = cpu_cgroup_destroy,
9882 .can_attach = cpu_cgroup_can_attach,
9883 .attach = cpu_cgroup_attach,
9884 .populate = cpu_cgroup_populate,
9885 .subsys_id = cpu_cgroup_subsys_id,
9889 #endif /* CONFIG_CGROUP_SCHED */
9891 #ifdef CONFIG_CGROUP_CPUACCT
9894 * CPU accounting code for task groups.
9896 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9897 * (balbir@in.ibm.com).
9900 /* track cpu usage of a group of tasks and its child groups */
9902 struct cgroup_subsys_state css;
9903 /* cpuusage holds pointer to a u64-type object on every cpu */
9905 struct cpuacct *parent;
9908 struct cgroup_subsys cpuacct_subsys;
9910 /* return cpu accounting group corresponding to this container */
9911 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9913 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9914 struct cpuacct, css);
9917 /* return cpu accounting group to which this task belongs */
9918 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9920 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9921 struct cpuacct, css);
9924 /* create a new cpu accounting group */
9925 static struct cgroup_subsys_state *cpuacct_create(
9926 struct cgroup_subsys *ss, struct cgroup *cgrp)
9928 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9931 return ERR_PTR(-ENOMEM);
9933 ca->cpuusage = alloc_percpu(u64);
9934 if (!ca->cpuusage) {
9936 return ERR_PTR(-ENOMEM);
9940 ca->parent = cgroup_ca(cgrp->parent);
9945 /* destroy an existing cpu accounting group */
9947 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9949 struct cpuacct *ca = cgroup_ca(cgrp);
9951 free_percpu(ca->cpuusage);
9955 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9957 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9960 #ifndef CONFIG_64BIT
9962 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9964 spin_lock_irq(&cpu_rq(cpu)->lock);
9966 spin_unlock_irq(&cpu_rq(cpu)->lock);
9974 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9976 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9978 #ifndef CONFIG_64BIT
9980 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9982 spin_lock_irq(&cpu_rq(cpu)->lock);
9984 spin_unlock_irq(&cpu_rq(cpu)->lock);
9990 /* return total cpu usage (in nanoseconds) of a group */
9991 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9993 struct cpuacct *ca = cgroup_ca(cgrp);
9994 u64 totalcpuusage = 0;
9997 for_each_present_cpu(i)
9998 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10000 return totalcpuusage;
10003 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10006 struct cpuacct *ca = cgroup_ca(cgrp);
10015 for_each_present_cpu(i)
10016 cpuacct_cpuusage_write(ca, i, 0);
10022 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10023 struct seq_file *m)
10025 struct cpuacct *ca = cgroup_ca(cgroup);
10029 for_each_present_cpu(i) {
10030 percpu = cpuacct_cpuusage_read(ca, i);
10031 seq_printf(m, "%llu ", (unsigned long long) percpu);
10033 seq_printf(m, "\n");
10037 static struct cftype files[] = {
10040 .read_u64 = cpuusage_read,
10041 .write_u64 = cpuusage_write,
10044 .name = "usage_percpu",
10045 .read_seq_string = cpuacct_percpu_seq_read,
10050 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10052 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10056 * charge this task's execution time to its accounting group.
10058 * called with rq->lock held.
10060 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10062 struct cpuacct *ca;
10065 if (unlikely(!cpuacct_subsys.active))
10068 cpu = task_cpu(tsk);
10071 for (; ca; ca = ca->parent) {
10072 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10073 *cpuusage += cputime;
10077 struct cgroup_subsys cpuacct_subsys = {
10079 .create = cpuacct_create,
10080 .destroy = cpuacct_destroy,
10081 .populate = cpuacct_populate,
10082 .subsys_id = cpuacct_subsys_id,
10084 #endif /* CONFIG_CGROUP_CPUACCT */