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)
123 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
124 * Since cpu_power is a 'constant', we can use a reciprocal divide.
126 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
128 return reciprocal_divide(load, sg->reciprocal_cpu_power);
132 * Each time a sched group cpu_power is changed,
133 * we must compute its reciprocal value
135 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
137 sg->__cpu_power += val;
138 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
142 static inline int rt_policy(int policy)
144 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
149 static inline int task_has_rt_policy(struct task_struct *p)
151 return rt_policy(p->policy);
155 * This is the priority-queue data structure of the RT scheduling class:
157 struct rt_prio_array {
158 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
159 struct list_head queue[MAX_RT_PRIO];
162 struct rt_bandwidth {
163 /* nests inside the rq lock: */
164 spinlock_t rt_runtime_lock;
167 struct hrtimer rt_period_timer;
170 static struct rt_bandwidth def_rt_bandwidth;
172 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
174 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
176 struct rt_bandwidth *rt_b =
177 container_of(timer, struct rt_bandwidth, rt_period_timer);
183 now = hrtimer_cb_get_time(timer);
184 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
189 idle = do_sched_rt_period_timer(rt_b, overrun);
192 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
196 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
198 rt_b->rt_period = ns_to_ktime(period);
199 rt_b->rt_runtime = runtime;
201 spin_lock_init(&rt_b->rt_runtime_lock);
203 hrtimer_init(&rt_b->rt_period_timer,
204 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
205 rt_b->rt_period_timer.function = sched_rt_period_timer;
206 rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_UNLOCKED;
209 static inline int rt_bandwidth_enabled(void)
211 return sysctl_sched_rt_runtime >= 0;
214 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
218 if (rt_bandwidth_enabled() && rt_b->rt_runtime == RUNTIME_INF)
221 if (hrtimer_active(&rt_b->rt_period_timer))
224 spin_lock(&rt_b->rt_runtime_lock);
226 if (hrtimer_active(&rt_b->rt_period_timer))
229 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
230 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
231 hrtimer_start_expires(&rt_b->rt_period_timer,
234 spin_unlock(&rt_b->rt_runtime_lock);
237 #ifdef CONFIG_RT_GROUP_SCHED
238 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
240 hrtimer_cancel(&rt_b->rt_period_timer);
245 * sched_domains_mutex serializes calls to arch_init_sched_domains,
246 * detach_destroy_domains and partition_sched_domains.
248 static DEFINE_MUTEX(sched_domains_mutex);
250 #ifdef CONFIG_GROUP_SCHED
252 #include <linux/cgroup.h>
256 static LIST_HEAD(task_groups);
258 /* task group related information */
260 #ifdef CONFIG_CGROUP_SCHED
261 struct cgroup_subsys_state css;
264 #ifdef CONFIG_FAIR_GROUP_SCHED
265 /* schedulable entities of this group on each cpu */
266 struct sched_entity **se;
267 /* runqueue "owned" by this group on each cpu */
268 struct cfs_rq **cfs_rq;
269 unsigned long shares;
272 #ifdef CONFIG_RT_GROUP_SCHED
273 struct sched_rt_entity **rt_se;
274 struct rt_rq **rt_rq;
276 struct rt_bandwidth rt_bandwidth;
280 struct list_head list;
282 struct task_group *parent;
283 struct list_head siblings;
284 struct list_head children;
287 #ifdef CONFIG_USER_SCHED
291 * Every UID task group (including init_task_group aka UID-0) will
292 * be a child to this group.
294 struct task_group root_task_group;
296 #ifdef CONFIG_FAIR_GROUP_SCHED
297 /* Default task group's sched entity on each cpu */
298 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
299 /* Default task group's cfs_rq on each cpu */
300 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
301 #endif /* CONFIG_FAIR_GROUP_SCHED */
303 #ifdef CONFIG_RT_GROUP_SCHED
304 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
305 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
306 #endif /* CONFIG_RT_GROUP_SCHED */
307 #else /* !CONFIG_USER_SCHED */
308 #define root_task_group init_task_group
309 #endif /* CONFIG_USER_SCHED */
311 /* task_group_lock serializes add/remove of task groups and also changes to
312 * a task group's cpu shares.
314 static DEFINE_SPINLOCK(task_group_lock);
316 #ifdef CONFIG_FAIR_GROUP_SCHED
317 #ifdef CONFIG_USER_SCHED
318 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
319 #else /* !CONFIG_USER_SCHED */
320 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
321 #endif /* CONFIG_USER_SCHED */
324 * A weight of 0 or 1 can cause arithmetics problems.
325 * A weight of a cfs_rq is the sum of weights of which entities
326 * are queued on this cfs_rq, so a weight of a entity should not be
327 * too large, so as the shares value of a task group.
328 * (The default weight is 1024 - so there's no practical
329 * limitation from this.)
332 #define MAX_SHARES (1UL << 18)
334 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
337 /* Default task group.
338 * Every task in system belong to this group at bootup.
340 struct task_group init_task_group;
342 /* return group to which a task belongs */
343 static inline struct task_group *task_group(struct task_struct *p)
345 struct task_group *tg;
347 #ifdef CONFIG_USER_SCHED
349 #elif defined(CONFIG_CGROUP_SCHED)
350 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
351 struct task_group, css);
353 tg = &init_task_group;
358 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
359 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
361 #ifdef CONFIG_FAIR_GROUP_SCHED
362 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
363 p->se.parent = task_group(p)->se[cpu];
366 #ifdef CONFIG_RT_GROUP_SCHED
367 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
368 p->rt.parent = task_group(p)->rt_se[cpu];
374 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
375 static inline struct task_group *task_group(struct task_struct *p)
380 #endif /* CONFIG_GROUP_SCHED */
382 /* CFS-related fields in a runqueue */
384 struct load_weight load;
385 unsigned long nr_running;
390 struct rb_root tasks_timeline;
391 struct rb_node *rb_leftmost;
393 struct list_head tasks;
394 struct list_head *balance_iterator;
397 * 'curr' points to currently running entity on this cfs_rq.
398 * It is set to NULL otherwise (i.e when none are currently running).
400 struct sched_entity *curr, *next, *last;
402 unsigned long nr_spread_over;
404 #ifdef CONFIG_FAIR_GROUP_SCHED
405 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
408 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
409 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
410 * (like users, containers etc.)
412 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
413 * list is used during load balance.
415 struct list_head leaf_cfs_rq_list;
416 struct task_group *tg; /* group that "owns" this runqueue */
420 * the part of load.weight contributed by tasks
422 unsigned long task_weight;
425 * h_load = weight * f(tg)
427 * Where f(tg) is the recursive weight fraction assigned to
430 unsigned long h_load;
433 * this cpu's part of tg->shares
435 unsigned long shares;
438 * load.weight at the time we set shares
440 unsigned long rq_weight;
445 /* Real-Time classes' related field in a runqueue: */
447 struct rt_prio_array active;
448 unsigned long rt_nr_running;
449 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
450 int highest_prio; /* highest queued rt task prio */
453 unsigned long rt_nr_migratory;
459 /* Nests inside the rq lock: */
460 spinlock_t rt_runtime_lock;
462 #ifdef CONFIG_RT_GROUP_SCHED
463 unsigned long rt_nr_boosted;
466 struct list_head leaf_rt_rq_list;
467 struct task_group *tg;
468 struct sched_rt_entity *rt_se;
475 * We add the notion of a root-domain which will be used to define per-domain
476 * variables. Each exclusive cpuset essentially defines an island domain by
477 * fully partitioning the member cpus from any other cpuset. Whenever a new
478 * exclusive cpuset is created, we also create and attach a new root-domain
488 * The "RT overload" flag: it gets set if a CPU has more than
489 * one runnable RT task.
494 struct cpupri cpupri;
499 * By default the system creates a single root-domain with all cpus as
500 * members (mimicking the global state we have today).
502 static struct root_domain def_root_domain;
507 * This is the main, per-CPU runqueue data structure.
509 * Locking rule: those places that want to lock multiple runqueues
510 * (such as the load balancing or the thread migration code), lock
511 * acquire operations must be ordered by ascending &runqueue.
518 * nr_running and cpu_load should be in the same cacheline because
519 * remote CPUs use both these fields when doing load calculation.
521 unsigned long nr_running;
522 #define CPU_LOAD_IDX_MAX 5
523 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
524 unsigned char idle_at_tick;
526 unsigned long last_tick_seen;
527 unsigned char in_nohz_recently;
529 /* capture load from *all* tasks on this cpu: */
530 struct load_weight load;
531 unsigned long nr_load_updates;
537 #ifdef CONFIG_FAIR_GROUP_SCHED
538 /* list of leaf cfs_rq on this cpu: */
539 struct list_head leaf_cfs_rq_list;
541 #ifdef CONFIG_RT_GROUP_SCHED
542 struct list_head leaf_rt_rq_list;
546 * This is part of a global counter where only the total sum
547 * over all CPUs matters. A task can increase this counter on
548 * one CPU and if it got migrated afterwards it may decrease
549 * it on another CPU. Always updated under the runqueue lock:
551 unsigned long nr_uninterruptible;
553 struct task_struct *curr, *idle;
554 unsigned long next_balance;
555 struct mm_struct *prev_mm;
562 struct root_domain *rd;
563 struct sched_domain *sd;
565 /* For active balancing */
568 /* cpu of this runqueue: */
572 unsigned long avg_load_per_task;
574 struct task_struct *migration_thread;
575 struct list_head migration_queue;
578 #ifdef CONFIG_SCHED_HRTICK
580 int hrtick_csd_pending;
581 struct call_single_data hrtick_csd;
583 struct hrtimer hrtick_timer;
586 #ifdef CONFIG_SCHEDSTATS
588 struct sched_info rq_sched_info;
590 /* sys_sched_yield() stats */
591 unsigned int yld_exp_empty;
592 unsigned int yld_act_empty;
593 unsigned int yld_both_empty;
594 unsigned int yld_count;
596 /* schedule() stats */
597 unsigned int sched_switch;
598 unsigned int sched_count;
599 unsigned int sched_goidle;
601 /* try_to_wake_up() stats */
602 unsigned int ttwu_count;
603 unsigned int ttwu_local;
606 unsigned int bkl_count;
610 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
612 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
614 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
617 static inline int cpu_of(struct rq *rq)
627 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
628 * See detach_destroy_domains: synchronize_sched for details.
630 * The domain tree of any CPU may only be accessed from within
631 * preempt-disabled sections.
633 #define for_each_domain(cpu, __sd) \
634 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
636 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
637 #define this_rq() (&__get_cpu_var(runqueues))
638 #define task_rq(p) cpu_rq(task_cpu(p))
639 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
641 static inline void update_rq_clock(struct rq *rq)
643 rq->clock = sched_clock_cpu(cpu_of(rq));
647 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
649 #ifdef CONFIG_SCHED_DEBUG
650 # define const_debug __read_mostly
652 # define const_debug static const
658 * Returns true if the current cpu runqueue is locked.
659 * This interface allows printk to be called with the runqueue lock
660 * held and know whether or not it is OK to wake up the klogd.
662 int runqueue_is_locked(void)
665 struct rq *rq = cpu_rq(cpu);
668 ret = spin_is_locked(&rq->lock);
674 * Debugging: various feature bits
677 #define SCHED_FEAT(name, enabled) \
678 __SCHED_FEAT_##name ,
681 #include "sched_features.h"
686 #define SCHED_FEAT(name, enabled) \
687 (1UL << __SCHED_FEAT_##name) * enabled |
689 const_debug unsigned int sysctl_sched_features =
690 #include "sched_features.h"
695 #ifdef CONFIG_SCHED_DEBUG
696 #define SCHED_FEAT(name, enabled) \
699 static __read_mostly char *sched_feat_names[] = {
700 #include "sched_features.h"
706 static int sched_feat_show(struct seq_file *m, void *v)
710 for (i = 0; sched_feat_names[i]; i++) {
711 if (!(sysctl_sched_features & (1UL << i)))
713 seq_printf(m, "%s ", sched_feat_names[i]);
721 sched_feat_write(struct file *filp, const char __user *ubuf,
722 size_t cnt, loff_t *ppos)
732 if (copy_from_user(&buf, ubuf, cnt))
737 if (strncmp(buf, "NO_", 3) == 0) {
742 for (i = 0; sched_feat_names[i]; i++) {
743 int len = strlen(sched_feat_names[i]);
745 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
747 sysctl_sched_features &= ~(1UL << i);
749 sysctl_sched_features |= (1UL << i);
754 if (!sched_feat_names[i])
762 static int sched_feat_open(struct inode *inode, struct file *filp)
764 return single_open(filp, sched_feat_show, NULL);
767 static struct file_operations sched_feat_fops = {
768 .open = sched_feat_open,
769 .write = sched_feat_write,
772 .release = single_release,
775 static __init int sched_init_debug(void)
777 debugfs_create_file("sched_features", 0644, NULL, NULL,
782 late_initcall(sched_init_debug);
786 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
789 * Number of tasks to iterate in a single balance run.
790 * Limited because this is done with IRQs disabled.
792 const_debug unsigned int sysctl_sched_nr_migrate = 32;
795 * ratelimit for updating the group shares.
798 unsigned int sysctl_sched_shares_ratelimit = 250000;
801 * Inject some fuzzyness into changing the per-cpu group shares
802 * this avoids remote rq-locks at the expense of fairness.
805 unsigned int sysctl_sched_shares_thresh = 4;
808 * period over which we measure -rt task cpu usage in us.
811 unsigned int sysctl_sched_rt_period = 1000000;
813 static __read_mostly int scheduler_running;
816 * part of the period that we allow rt tasks to run in us.
819 int sysctl_sched_rt_runtime = 950000;
821 static inline u64 global_rt_period(void)
823 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
826 static inline u64 global_rt_runtime(void)
828 if (sysctl_sched_rt_runtime < 0)
831 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
834 #ifndef prepare_arch_switch
835 # define prepare_arch_switch(next) do { } while (0)
837 #ifndef finish_arch_switch
838 # define finish_arch_switch(prev) do { } while (0)
841 static inline int task_current(struct rq *rq, struct task_struct *p)
843 return rq->curr == p;
846 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
847 static inline int task_running(struct rq *rq, struct task_struct *p)
849 return task_current(rq, p);
852 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
856 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
858 #ifdef CONFIG_DEBUG_SPINLOCK
859 /* this is a valid case when another task releases the spinlock */
860 rq->lock.owner = current;
863 * If we are tracking spinlock dependencies then we have to
864 * fix up the runqueue lock - which gets 'carried over' from
867 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
869 spin_unlock_irq(&rq->lock);
872 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
873 static inline int task_running(struct rq *rq, struct task_struct *p)
878 return task_current(rq, p);
882 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
886 * We can optimise this out completely for !SMP, because the
887 * SMP rebalancing from interrupt is the only thing that cares
892 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
893 spin_unlock_irq(&rq->lock);
895 spin_unlock(&rq->lock);
899 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
903 * After ->oncpu is cleared, the task can be moved to a different CPU.
904 * We must ensure this doesn't happen until the switch is completely
910 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
914 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
917 * __task_rq_lock - lock the runqueue a given task resides on.
918 * Must be called interrupts disabled.
920 static inline struct rq *__task_rq_lock(struct task_struct *p)
924 struct rq *rq = task_rq(p);
925 spin_lock(&rq->lock);
926 if (likely(rq == task_rq(p)))
928 spin_unlock(&rq->lock);
933 * task_rq_lock - lock the runqueue a given task resides on and disable
934 * interrupts. Note the ordering: we can safely lookup the task_rq without
935 * explicitly disabling preemption.
937 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
943 local_irq_save(*flags);
945 spin_lock(&rq->lock);
946 if (likely(rq == task_rq(p)))
948 spin_unlock_irqrestore(&rq->lock, *flags);
952 static void __task_rq_unlock(struct rq *rq)
955 spin_unlock(&rq->lock);
958 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
961 spin_unlock_irqrestore(&rq->lock, *flags);
965 * this_rq_lock - lock this runqueue and disable interrupts.
967 static struct rq *this_rq_lock(void)
974 spin_lock(&rq->lock);
979 #ifdef CONFIG_SCHED_HRTICK
981 * Use HR-timers to deliver accurate preemption points.
983 * Its all a bit involved since we cannot program an hrt while holding the
984 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
987 * When we get rescheduled we reprogram the hrtick_timer outside of the
993 * - enabled by features
994 * - hrtimer is actually high res
996 static inline int hrtick_enabled(struct rq *rq)
998 if (!sched_feat(HRTICK))
1000 if (!cpu_active(cpu_of(rq)))
1002 return hrtimer_is_hres_active(&rq->hrtick_timer);
1005 static void hrtick_clear(struct rq *rq)
1007 if (hrtimer_active(&rq->hrtick_timer))
1008 hrtimer_cancel(&rq->hrtick_timer);
1012 * High-resolution timer tick.
1013 * Runs from hardirq context with interrupts disabled.
1015 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1017 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1019 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1021 spin_lock(&rq->lock);
1022 update_rq_clock(rq);
1023 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1024 spin_unlock(&rq->lock);
1026 return HRTIMER_NORESTART;
1031 * called from hardirq (IPI) context
1033 static void __hrtick_start(void *arg)
1035 struct rq *rq = arg;
1037 spin_lock(&rq->lock);
1038 hrtimer_restart(&rq->hrtick_timer);
1039 rq->hrtick_csd_pending = 0;
1040 spin_unlock(&rq->lock);
1044 * Called to set the hrtick timer state.
1046 * called with rq->lock held and irqs disabled
1048 static void hrtick_start(struct rq *rq, u64 delay)
1050 struct hrtimer *timer = &rq->hrtick_timer;
1051 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1053 hrtimer_set_expires(timer, time);
1055 if (rq == this_rq()) {
1056 hrtimer_restart(timer);
1057 } else if (!rq->hrtick_csd_pending) {
1058 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd);
1059 rq->hrtick_csd_pending = 1;
1064 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1066 int cpu = (int)(long)hcpu;
1069 case CPU_UP_CANCELED:
1070 case CPU_UP_CANCELED_FROZEN:
1071 case CPU_DOWN_PREPARE:
1072 case CPU_DOWN_PREPARE_FROZEN:
1074 case CPU_DEAD_FROZEN:
1075 hrtick_clear(cpu_rq(cpu));
1082 static __init void init_hrtick(void)
1084 hotcpu_notifier(hotplug_hrtick, 0);
1088 * Called to set the hrtick timer state.
1090 * called with rq->lock held and irqs disabled
1092 static void hrtick_start(struct rq *rq, u64 delay)
1094 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay), HRTIMER_MODE_REL);
1097 static inline void init_hrtick(void)
1100 #endif /* CONFIG_SMP */
1102 static void init_rq_hrtick(struct rq *rq)
1105 rq->hrtick_csd_pending = 0;
1107 rq->hrtick_csd.flags = 0;
1108 rq->hrtick_csd.func = __hrtick_start;
1109 rq->hrtick_csd.info = rq;
1112 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1113 rq->hrtick_timer.function = hrtick;
1114 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_PERCPU;
1116 #else /* CONFIG_SCHED_HRTICK */
1117 static inline void hrtick_clear(struct rq *rq)
1121 static inline void init_rq_hrtick(struct rq *rq)
1125 static inline void init_hrtick(void)
1128 #endif /* CONFIG_SCHED_HRTICK */
1131 * resched_task - mark a task 'to be rescheduled now'.
1133 * On UP this means the setting of the need_resched flag, on SMP it
1134 * might also involve a cross-CPU call to trigger the scheduler on
1139 #ifndef tsk_is_polling
1140 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1143 static void resched_task(struct task_struct *p)
1147 assert_spin_locked(&task_rq(p)->lock);
1149 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1152 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1155 if (cpu == smp_processor_id())
1158 /* NEED_RESCHED must be visible before we test polling */
1160 if (!tsk_is_polling(p))
1161 smp_send_reschedule(cpu);
1164 static void resched_cpu(int cpu)
1166 struct rq *rq = cpu_rq(cpu);
1167 unsigned long flags;
1169 if (!spin_trylock_irqsave(&rq->lock, flags))
1171 resched_task(cpu_curr(cpu));
1172 spin_unlock_irqrestore(&rq->lock, flags);
1177 * When add_timer_on() enqueues a timer into the timer wheel of an
1178 * idle CPU then this timer might expire before the next timer event
1179 * which is scheduled to wake up that CPU. In case of a completely
1180 * idle system the next event might even be infinite time into the
1181 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1182 * leaves the inner idle loop so the newly added timer is taken into
1183 * account when the CPU goes back to idle and evaluates the timer
1184 * wheel for the next timer event.
1186 void wake_up_idle_cpu(int cpu)
1188 struct rq *rq = cpu_rq(cpu);
1190 if (cpu == smp_processor_id())
1194 * This is safe, as this function is called with the timer
1195 * wheel base lock of (cpu) held. When the CPU is on the way
1196 * to idle and has not yet set rq->curr to idle then it will
1197 * be serialized on the timer wheel base lock and take the new
1198 * timer into account automatically.
1200 if (rq->curr != rq->idle)
1204 * We can set TIF_RESCHED on the idle task of the other CPU
1205 * lockless. The worst case is that the other CPU runs the
1206 * idle task through an additional NOOP schedule()
1208 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1210 /* NEED_RESCHED must be visible before we test polling */
1212 if (!tsk_is_polling(rq->idle))
1213 smp_send_reschedule(cpu);
1215 #endif /* CONFIG_NO_HZ */
1217 #else /* !CONFIG_SMP */
1218 static void resched_task(struct task_struct *p)
1220 assert_spin_locked(&task_rq(p)->lock);
1221 set_tsk_need_resched(p);
1223 #endif /* CONFIG_SMP */
1225 #if BITS_PER_LONG == 32
1226 # define WMULT_CONST (~0UL)
1228 # define WMULT_CONST (1UL << 32)
1231 #define WMULT_SHIFT 32
1234 * Shift right and round:
1236 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1239 * delta *= weight / lw
1241 static unsigned long
1242 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1243 struct load_weight *lw)
1247 if (!lw->inv_weight) {
1248 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1251 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1255 tmp = (u64)delta_exec * weight;
1257 * Check whether we'd overflow the 64-bit multiplication:
1259 if (unlikely(tmp > WMULT_CONST))
1260 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1263 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1265 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1268 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1274 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1281 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1282 * of tasks with abnormal "nice" values across CPUs the contribution that
1283 * each task makes to its run queue's load is weighted according to its
1284 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1285 * scaled version of the new time slice allocation that they receive on time
1289 #define WEIGHT_IDLEPRIO 2
1290 #define WMULT_IDLEPRIO (1 << 31)
1293 * Nice levels are multiplicative, with a gentle 10% change for every
1294 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1295 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1296 * that remained on nice 0.
1298 * The "10% effect" is relative and cumulative: from _any_ nice level,
1299 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1300 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1301 * If a task goes up by ~10% and another task goes down by ~10% then
1302 * the relative distance between them is ~25%.)
1304 static const int prio_to_weight[40] = {
1305 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1306 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1307 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1308 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1309 /* 0 */ 1024, 820, 655, 526, 423,
1310 /* 5 */ 335, 272, 215, 172, 137,
1311 /* 10 */ 110, 87, 70, 56, 45,
1312 /* 15 */ 36, 29, 23, 18, 15,
1316 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1318 * In cases where the weight does not change often, we can use the
1319 * precalculated inverse to speed up arithmetics by turning divisions
1320 * into multiplications:
1322 static const u32 prio_to_wmult[40] = {
1323 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1324 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1325 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1326 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1327 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1328 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1329 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1330 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1333 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1336 * runqueue iterator, to support SMP load-balancing between different
1337 * scheduling classes, without having to expose their internal data
1338 * structures to the load-balancing proper:
1340 struct rq_iterator {
1342 struct task_struct *(*start)(void *);
1343 struct task_struct *(*next)(void *);
1347 static unsigned long
1348 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1349 unsigned long max_load_move, struct sched_domain *sd,
1350 enum cpu_idle_type idle, int *all_pinned,
1351 int *this_best_prio, struct rq_iterator *iterator);
1354 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1355 struct sched_domain *sd, enum cpu_idle_type idle,
1356 struct rq_iterator *iterator);
1359 #ifdef CONFIG_CGROUP_CPUACCT
1360 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1362 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1365 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1367 update_load_add(&rq->load, load);
1370 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1372 update_load_sub(&rq->load, load);
1375 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1376 typedef int (*tg_visitor)(struct task_group *, void *);
1379 * Iterate the full tree, calling @down when first entering a node and @up when
1380 * leaving it for the final time.
1382 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1384 struct task_group *parent, *child;
1388 parent = &root_task_group;
1390 ret = (*down)(parent, data);
1393 list_for_each_entry_rcu(child, &parent->children, siblings) {
1400 ret = (*up)(parent, data);
1405 parent = parent->parent;
1414 static int tg_nop(struct task_group *tg, void *data)
1421 static unsigned long source_load(int cpu, int type);
1422 static unsigned long target_load(int cpu, int type);
1423 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1425 static unsigned long cpu_avg_load_per_task(int cpu)
1427 struct rq *rq = cpu_rq(cpu);
1430 rq->avg_load_per_task = rq->load.weight / rq->nr_running;
1432 return rq->avg_load_per_task;
1435 #ifdef CONFIG_FAIR_GROUP_SCHED
1437 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1440 * Calculate and set the cpu's group shares.
1443 update_group_shares_cpu(struct task_group *tg, int cpu,
1444 unsigned long sd_shares, unsigned long sd_rq_weight)
1447 unsigned long shares;
1448 unsigned long rq_weight;
1453 rq_weight = tg->cfs_rq[cpu]->load.weight;
1456 * If there are currently no tasks on the cpu pretend there is one of
1457 * average load so that when a new task gets to run here it will not
1458 * get delayed by group starvation.
1462 rq_weight = NICE_0_LOAD;
1465 if (unlikely(rq_weight > sd_rq_weight))
1466 rq_weight = sd_rq_weight;
1469 * \Sum shares * rq_weight
1470 * shares = -----------------------
1474 shares = (sd_shares * rq_weight) / (sd_rq_weight + 1);
1475 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1477 if (abs(shares - tg->se[cpu]->load.weight) >
1478 sysctl_sched_shares_thresh) {
1479 struct rq *rq = cpu_rq(cpu);
1480 unsigned long flags;
1482 spin_lock_irqsave(&rq->lock, flags);
1484 * record the actual number of shares, not the boosted amount.
1486 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1487 tg->cfs_rq[cpu]->rq_weight = rq_weight;
1489 __set_se_shares(tg->se[cpu], shares);
1490 spin_unlock_irqrestore(&rq->lock, flags);
1495 * Re-compute the task group their per cpu shares over the given domain.
1496 * This needs to be done in a bottom-up fashion because the rq weight of a
1497 * parent group depends on the shares of its child groups.
1499 static int tg_shares_up(struct task_group *tg, void *data)
1501 unsigned long rq_weight = 0;
1502 unsigned long shares = 0;
1503 struct sched_domain *sd = data;
1506 for_each_cpu_mask(i, sd->span) {
1507 rq_weight += tg->cfs_rq[i]->load.weight;
1508 shares += tg->cfs_rq[i]->shares;
1511 if ((!shares && rq_weight) || shares > tg->shares)
1512 shares = tg->shares;
1514 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1515 shares = tg->shares;
1518 rq_weight = cpus_weight(sd->span) * NICE_0_LOAD;
1520 for_each_cpu_mask(i, sd->span)
1521 update_group_shares_cpu(tg, i, shares, rq_weight);
1527 * Compute the cpu's hierarchical load factor for each task group.
1528 * This needs to be done in a top-down fashion because the load of a child
1529 * group is a fraction of its parents load.
1531 static int tg_load_down(struct task_group *tg, void *data)
1534 long cpu = (long)data;
1537 load = cpu_rq(cpu)->load.weight;
1539 load = tg->parent->cfs_rq[cpu]->h_load;
1540 load *= tg->cfs_rq[cpu]->shares;
1541 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1544 tg->cfs_rq[cpu]->h_load = load;
1549 static void update_shares(struct sched_domain *sd)
1551 u64 now = cpu_clock(raw_smp_processor_id());
1552 s64 elapsed = now - sd->last_update;
1554 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1555 sd->last_update = now;
1556 walk_tg_tree(tg_nop, tg_shares_up, sd);
1560 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1562 spin_unlock(&rq->lock);
1564 spin_lock(&rq->lock);
1567 static void update_h_load(long cpu)
1569 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1574 static inline void update_shares(struct sched_domain *sd)
1578 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1586 #ifdef CONFIG_FAIR_GROUP_SCHED
1587 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1590 cfs_rq->shares = shares;
1595 #include "sched_stats.h"
1596 #include "sched_idletask.c"
1597 #include "sched_fair.c"
1598 #include "sched_rt.c"
1599 #ifdef CONFIG_SCHED_DEBUG
1600 # include "sched_debug.c"
1603 #define sched_class_highest (&rt_sched_class)
1604 #define for_each_class(class) \
1605 for (class = sched_class_highest; class; class = class->next)
1607 static void inc_nr_running(struct rq *rq)
1612 static void dec_nr_running(struct rq *rq)
1617 static void set_load_weight(struct task_struct *p)
1619 if (task_has_rt_policy(p)) {
1620 p->se.load.weight = prio_to_weight[0] * 2;
1621 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1626 * SCHED_IDLE tasks get minimal weight:
1628 if (p->policy == SCHED_IDLE) {
1629 p->se.load.weight = WEIGHT_IDLEPRIO;
1630 p->se.load.inv_weight = WMULT_IDLEPRIO;
1634 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1635 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1638 static void update_avg(u64 *avg, u64 sample)
1640 s64 diff = sample - *avg;
1644 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1646 sched_info_queued(p);
1647 p->sched_class->enqueue_task(rq, p, wakeup);
1651 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1653 if (sleep && p->se.last_wakeup) {
1654 update_avg(&p->se.avg_overlap,
1655 p->se.sum_exec_runtime - p->se.last_wakeup);
1656 p->se.last_wakeup = 0;
1659 sched_info_dequeued(p);
1660 p->sched_class->dequeue_task(rq, p, sleep);
1665 * __normal_prio - return the priority that is based on the static prio
1667 static inline int __normal_prio(struct task_struct *p)
1669 return p->static_prio;
1673 * Calculate the expected normal priority: i.e. priority
1674 * without taking RT-inheritance into account. Might be
1675 * boosted by interactivity modifiers. Changes upon fork,
1676 * setprio syscalls, and whenever the interactivity
1677 * estimator recalculates.
1679 static inline int normal_prio(struct task_struct *p)
1683 if (task_has_rt_policy(p))
1684 prio = MAX_RT_PRIO-1 - p->rt_priority;
1686 prio = __normal_prio(p);
1691 * Calculate the current priority, i.e. the priority
1692 * taken into account by the scheduler. This value might
1693 * be boosted by RT tasks, or might be boosted by
1694 * interactivity modifiers. Will be RT if the task got
1695 * RT-boosted. If not then it returns p->normal_prio.
1697 static int effective_prio(struct task_struct *p)
1699 p->normal_prio = normal_prio(p);
1701 * If we are RT tasks or we were boosted to RT priority,
1702 * keep the priority unchanged. Otherwise, update priority
1703 * to the normal priority:
1705 if (!rt_prio(p->prio))
1706 return p->normal_prio;
1711 * activate_task - move a task to the runqueue.
1713 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1715 if (task_contributes_to_load(p))
1716 rq->nr_uninterruptible--;
1718 enqueue_task(rq, p, wakeup);
1723 * deactivate_task - remove a task from the runqueue.
1725 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1727 if (task_contributes_to_load(p))
1728 rq->nr_uninterruptible++;
1730 dequeue_task(rq, p, sleep);
1735 * task_curr - is this task currently executing on a CPU?
1736 * @p: the task in question.
1738 inline int task_curr(const struct task_struct *p)
1740 return cpu_curr(task_cpu(p)) == p;
1743 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1745 set_task_rq(p, cpu);
1748 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1749 * successfuly executed on another CPU. We must ensure that updates of
1750 * per-task data have been completed by this moment.
1753 task_thread_info(p)->cpu = cpu;
1757 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1758 const struct sched_class *prev_class,
1759 int oldprio, int running)
1761 if (prev_class != p->sched_class) {
1762 if (prev_class->switched_from)
1763 prev_class->switched_from(rq, p, running);
1764 p->sched_class->switched_to(rq, p, running);
1766 p->sched_class->prio_changed(rq, p, oldprio, running);
1771 /* Used instead of source_load when we know the type == 0 */
1772 static unsigned long weighted_cpuload(const int cpu)
1774 return cpu_rq(cpu)->load.weight;
1778 * Is this task likely cache-hot:
1781 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1786 * Buddy candidates are cache hot:
1788 if (sched_feat(CACHE_HOT_BUDDY) &&
1789 (&p->se == cfs_rq_of(&p->se)->next ||
1790 &p->se == cfs_rq_of(&p->se)->last))
1793 if (p->sched_class != &fair_sched_class)
1796 if (sysctl_sched_migration_cost == -1)
1798 if (sysctl_sched_migration_cost == 0)
1801 delta = now - p->se.exec_start;
1803 return delta < (s64)sysctl_sched_migration_cost;
1807 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1809 int old_cpu = task_cpu(p);
1810 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1811 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1812 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1815 clock_offset = old_rq->clock - new_rq->clock;
1817 #ifdef CONFIG_SCHEDSTATS
1818 if (p->se.wait_start)
1819 p->se.wait_start -= clock_offset;
1820 if (p->se.sleep_start)
1821 p->se.sleep_start -= clock_offset;
1822 if (p->se.block_start)
1823 p->se.block_start -= clock_offset;
1824 if (old_cpu != new_cpu) {
1825 schedstat_inc(p, se.nr_migrations);
1826 if (task_hot(p, old_rq->clock, NULL))
1827 schedstat_inc(p, se.nr_forced2_migrations);
1830 p->se.vruntime -= old_cfsrq->min_vruntime -
1831 new_cfsrq->min_vruntime;
1833 __set_task_cpu(p, new_cpu);
1836 struct migration_req {
1837 struct list_head list;
1839 struct task_struct *task;
1842 struct completion done;
1846 * The task's runqueue lock must be held.
1847 * Returns true if you have to wait for migration thread.
1850 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1852 struct rq *rq = task_rq(p);
1855 * If the task is not on a runqueue (and not running), then
1856 * it is sufficient to simply update the task's cpu field.
1858 if (!p->se.on_rq && !task_running(rq, p)) {
1859 set_task_cpu(p, dest_cpu);
1863 init_completion(&req->done);
1865 req->dest_cpu = dest_cpu;
1866 list_add(&req->list, &rq->migration_queue);
1872 * wait_task_inactive - wait for a thread to unschedule.
1874 * If @match_state is nonzero, it's the @p->state value just checked and
1875 * not expected to change. If it changes, i.e. @p might have woken up,
1876 * then return zero. When we succeed in waiting for @p to be off its CPU,
1877 * we return a positive number (its total switch count). If a second call
1878 * a short while later returns the same number, the caller can be sure that
1879 * @p has remained unscheduled the whole time.
1881 * The caller must ensure that the task *will* unschedule sometime soon,
1882 * else this function might spin for a *long* time. This function can't
1883 * be called with interrupts off, or it may introduce deadlock with
1884 * smp_call_function() if an IPI is sent by the same process we are
1885 * waiting to become inactive.
1887 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1889 unsigned long flags;
1896 * We do the initial early heuristics without holding
1897 * any task-queue locks at all. We'll only try to get
1898 * the runqueue lock when things look like they will
1904 * If the task is actively running on another CPU
1905 * still, just relax and busy-wait without holding
1908 * NOTE! Since we don't hold any locks, it's not
1909 * even sure that "rq" stays as the right runqueue!
1910 * But we don't care, since "task_running()" will
1911 * return false if the runqueue has changed and p
1912 * is actually now running somewhere else!
1914 while (task_running(rq, p)) {
1915 if (match_state && unlikely(p->state != match_state))
1921 * Ok, time to look more closely! We need the rq
1922 * lock now, to be *sure*. If we're wrong, we'll
1923 * just go back and repeat.
1925 rq = task_rq_lock(p, &flags);
1926 trace_sched_wait_task(rq, p);
1927 running = task_running(rq, p);
1928 on_rq = p->se.on_rq;
1930 if (!match_state || p->state == match_state)
1931 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1932 task_rq_unlock(rq, &flags);
1935 * If it changed from the expected state, bail out now.
1937 if (unlikely(!ncsw))
1941 * Was it really running after all now that we
1942 * checked with the proper locks actually held?
1944 * Oops. Go back and try again..
1946 if (unlikely(running)) {
1952 * It's not enough that it's not actively running,
1953 * it must be off the runqueue _entirely_, and not
1956 * So if it wa still runnable (but just not actively
1957 * running right now), it's preempted, and we should
1958 * yield - it could be a while.
1960 if (unlikely(on_rq)) {
1961 schedule_timeout_uninterruptible(1);
1966 * Ahh, all good. It wasn't running, and it wasn't
1967 * runnable, which means that it will never become
1968 * running in the future either. We're all done!
1977 * kick_process - kick a running thread to enter/exit the kernel
1978 * @p: the to-be-kicked thread
1980 * Cause a process which is running on another CPU to enter
1981 * kernel-mode, without any delay. (to get signals handled.)
1983 * NOTE: this function doesnt have to take the runqueue lock,
1984 * because all it wants to ensure is that the remote task enters
1985 * the kernel. If the IPI races and the task has been migrated
1986 * to another CPU then no harm is done and the purpose has been
1989 void kick_process(struct task_struct *p)
1995 if ((cpu != smp_processor_id()) && task_curr(p))
1996 smp_send_reschedule(cpu);
2001 * Return a low guess at the load of a migration-source cpu weighted
2002 * according to the scheduling class and "nice" value.
2004 * We want to under-estimate the load of migration sources, to
2005 * balance conservatively.
2007 static unsigned long source_load(int cpu, int type)
2009 struct rq *rq = cpu_rq(cpu);
2010 unsigned long total = weighted_cpuload(cpu);
2012 if (type == 0 || !sched_feat(LB_BIAS))
2015 return min(rq->cpu_load[type-1], total);
2019 * Return a high guess at the load of a migration-target cpu weighted
2020 * according to the scheduling class and "nice" value.
2022 static unsigned long target_load(int cpu, int type)
2024 struct rq *rq = cpu_rq(cpu);
2025 unsigned long total = weighted_cpuload(cpu);
2027 if (type == 0 || !sched_feat(LB_BIAS))
2030 return max(rq->cpu_load[type-1], total);
2034 * find_idlest_group finds and returns the least busy CPU group within the
2037 static struct sched_group *
2038 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2040 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2041 unsigned long min_load = ULONG_MAX, this_load = 0;
2042 int load_idx = sd->forkexec_idx;
2043 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2046 unsigned long load, avg_load;
2050 /* Skip over this group if it has no CPUs allowed */
2051 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
2054 local_group = cpu_isset(this_cpu, group->cpumask);
2056 /* Tally up the load of all CPUs in the group */
2059 for_each_cpu_mask_nr(i, group->cpumask) {
2060 /* Bias balancing toward cpus of our domain */
2062 load = source_load(i, load_idx);
2064 load = target_load(i, load_idx);
2069 /* Adjust by relative CPU power of the group */
2070 avg_load = sg_div_cpu_power(group,
2071 avg_load * SCHED_LOAD_SCALE);
2074 this_load = avg_load;
2076 } else if (avg_load < min_load) {
2077 min_load = avg_load;
2080 } while (group = group->next, group != sd->groups);
2082 if (!idlest || 100*this_load < imbalance*min_load)
2088 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2091 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2094 unsigned long load, min_load = ULONG_MAX;
2098 /* Traverse only the allowed CPUs */
2099 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
2101 for_each_cpu_mask_nr(i, *tmp) {
2102 load = weighted_cpuload(i);
2104 if (load < min_load || (load == min_load && i == this_cpu)) {
2114 * sched_balance_self: balance the current task (running on cpu) in domains
2115 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2118 * Balance, ie. select the least loaded group.
2120 * Returns the target CPU number, or the same CPU if no balancing is needed.
2122 * preempt must be disabled.
2124 static int sched_balance_self(int cpu, int flag)
2126 struct task_struct *t = current;
2127 struct sched_domain *tmp, *sd = NULL;
2129 for_each_domain(cpu, tmp) {
2131 * If power savings logic is enabled for a domain, stop there.
2133 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2135 if (tmp->flags & flag)
2143 cpumask_t span, tmpmask;
2144 struct sched_group *group;
2145 int new_cpu, weight;
2147 if (!(sd->flags & flag)) {
2153 group = find_idlest_group(sd, t, cpu);
2159 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2160 if (new_cpu == -1 || new_cpu == cpu) {
2161 /* Now try balancing at a lower domain level of cpu */
2166 /* Now try balancing at a lower domain level of new_cpu */
2169 weight = cpus_weight(span);
2170 for_each_domain(cpu, tmp) {
2171 if (weight <= cpus_weight(tmp->span))
2173 if (tmp->flags & flag)
2176 /* while loop will break here if sd == NULL */
2182 #endif /* CONFIG_SMP */
2185 * try_to_wake_up - wake up a thread
2186 * @p: the to-be-woken-up thread
2187 * @state: the mask of task states that can be woken
2188 * @sync: do a synchronous wakeup?
2190 * Put it on the run-queue if it's not already there. The "current"
2191 * thread is always on the run-queue (except when the actual
2192 * re-schedule is in progress), and as such you're allowed to do
2193 * the simpler "current->state = TASK_RUNNING" to mark yourself
2194 * runnable without the overhead of this.
2196 * returns failure only if the task is already active.
2198 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2200 int cpu, orig_cpu, this_cpu, success = 0;
2201 unsigned long flags;
2205 if (!sched_feat(SYNC_WAKEUPS))
2209 if (sched_feat(LB_WAKEUP_UPDATE)) {
2210 struct sched_domain *sd;
2212 this_cpu = raw_smp_processor_id();
2215 for_each_domain(this_cpu, sd) {
2216 if (cpu_isset(cpu, sd->span)) {
2225 rq = task_rq_lock(p, &flags);
2226 old_state = p->state;
2227 if (!(old_state & state))
2235 this_cpu = smp_processor_id();
2238 if (unlikely(task_running(rq, p)))
2241 cpu = p->sched_class->select_task_rq(p, sync);
2242 if (cpu != orig_cpu) {
2243 set_task_cpu(p, cpu);
2244 task_rq_unlock(rq, &flags);
2245 /* might preempt at this point */
2246 rq = task_rq_lock(p, &flags);
2247 old_state = p->state;
2248 if (!(old_state & state))
2253 this_cpu = smp_processor_id();
2257 #ifdef CONFIG_SCHEDSTATS
2258 schedstat_inc(rq, ttwu_count);
2259 if (cpu == this_cpu)
2260 schedstat_inc(rq, ttwu_local);
2262 struct sched_domain *sd;
2263 for_each_domain(this_cpu, sd) {
2264 if (cpu_isset(cpu, sd->span)) {
2265 schedstat_inc(sd, ttwu_wake_remote);
2270 #endif /* CONFIG_SCHEDSTATS */
2273 #endif /* CONFIG_SMP */
2274 schedstat_inc(p, se.nr_wakeups);
2276 schedstat_inc(p, se.nr_wakeups_sync);
2277 if (orig_cpu != cpu)
2278 schedstat_inc(p, se.nr_wakeups_migrate);
2279 if (cpu == this_cpu)
2280 schedstat_inc(p, se.nr_wakeups_local);
2282 schedstat_inc(p, se.nr_wakeups_remote);
2283 update_rq_clock(rq);
2284 activate_task(rq, p, 1);
2288 trace_sched_wakeup(rq, p);
2289 check_preempt_curr(rq, p, sync);
2291 p->state = TASK_RUNNING;
2293 if (p->sched_class->task_wake_up)
2294 p->sched_class->task_wake_up(rq, p);
2297 current->se.last_wakeup = current->se.sum_exec_runtime;
2299 task_rq_unlock(rq, &flags);
2304 int wake_up_process(struct task_struct *p)
2306 return try_to_wake_up(p, TASK_ALL, 0);
2308 EXPORT_SYMBOL(wake_up_process);
2310 int wake_up_state(struct task_struct *p, unsigned int state)
2312 return try_to_wake_up(p, state, 0);
2316 * Perform scheduler related setup for a newly forked process p.
2317 * p is forked by current.
2319 * __sched_fork() is basic setup used by init_idle() too:
2321 static void __sched_fork(struct task_struct *p)
2323 p->se.exec_start = 0;
2324 p->se.sum_exec_runtime = 0;
2325 p->se.prev_sum_exec_runtime = 0;
2326 p->se.last_wakeup = 0;
2327 p->se.avg_overlap = 0;
2329 #ifdef CONFIG_SCHEDSTATS
2330 p->se.wait_start = 0;
2331 p->se.sum_sleep_runtime = 0;
2332 p->se.sleep_start = 0;
2333 p->se.block_start = 0;
2334 p->se.sleep_max = 0;
2335 p->se.block_max = 0;
2337 p->se.slice_max = 0;
2341 INIT_LIST_HEAD(&p->rt.run_list);
2343 INIT_LIST_HEAD(&p->se.group_node);
2345 #ifdef CONFIG_PREEMPT_NOTIFIERS
2346 INIT_HLIST_HEAD(&p->preempt_notifiers);
2350 * We mark the process as running here, but have not actually
2351 * inserted it onto the runqueue yet. This guarantees that
2352 * nobody will actually run it, and a signal or other external
2353 * event cannot wake it up and insert it on the runqueue either.
2355 p->state = TASK_RUNNING;
2359 * fork()/clone()-time setup:
2361 void sched_fork(struct task_struct *p, int clone_flags)
2363 int cpu = get_cpu();
2368 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2370 set_task_cpu(p, cpu);
2373 * Make sure we do not leak PI boosting priority to the child:
2375 p->prio = current->normal_prio;
2376 if (!rt_prio(p->prio))
2377 p->sched_class = &fair_sched_class;
2379 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2380 if (likely(sched_info_on()))
2381 memset(&p->sched_info, 0, sizeof(p->sched_info));
2383 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2386 #ifdef CONFIG_PREEMPT
2387 /* Want to start with kernel preemption disabled. */
2388 task_thread_info(p)->preempt_count = 1;
2394 * wake_up_new_task - wake up a newly created task for the first time.
2396 * This function will do some initial scheduler statistics housekeeping
2397 * that must be done for every newly created context, then puts the task
2398 * on the runqueue and wakes it.
2400 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2402 unsigned long flags;
2405 rq = task_rq_lock(p, &flags);
2406 BUG_ON(p->state != TASK_RUNNING);
2407 update_rq_clock(rq);
2409 p->prio = effective_prio(p);
2411 if (!p->sched_class->task_new || !current->se.on_rq) {
2412 activate_task(rq, p, 0);
2415 * Let the scheduling class do new task startup
2416 * management (if any):
2418 p->sched_class->task_new(rq, p);
2421 trace_sched_wakeup_new(rq, p);
2422 check_preempt_curr(rq, p, 0);
2424 if (p->sched_class->task_wake_up)
2425 p->sched_class->task_wake_up(rq, p);
2427 task_rq_unlock(rq, &flags);
2430 #ifdef CONFIG_PREEMPT_NOTIFIERS
2433 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2434 * @notifier: notifier struct to register
2436 void preempt_notifier_register(struct preempt_notifier *notifier)
2438 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2440 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2443 * preempt_notifier_unregister - no longer interested in preemption notifications
2444 * @notifier: notifier struct to unregister
2446 * This is safe to call from within a preemption notifier.
2448 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2450 hlist_del(¬ifier->link);
2452 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2454 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2456 struct preempt_notifier *notifier;
2457 struct hlist_node *node;
2459 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2460 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2464 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2465 struct task_struct *next)
2467 struct preempt_notifier *notifier;
2468 struct hlist_node *node;
2470 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2471 notifier->ops->sched_out(notifier, next);
2474 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2476 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2481 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2482 struct task_struct *next)
2486 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2489 * prepare_task_switch - prepare to switch tasks
2490 * @rq: the runqueue preparing to switch
2491 * @prev: the current task that is being switched out
2492 * @next: the task we are going to switch to.
2494 * This is called with the rq lock held and interrupts off. It must
2495 * be paired with a subsequent finish_task_switch after the context
2498 * prepare_task_switch sets up locking and calls architecture specific
2502 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2503 struct task_struct *next)
2505 fire_sched_out_preempt_notifiers(prev, next);
2506 prepare_lock_switch(rq, next);
2507 prepare_arch_switch(next);
2511 * finish_task_switch - clean up after a task-switch
2512 * @rq: runqueue associated with task-switch
2513 * @prev: the thread we just switched away from.
2515 * finish_task_switch must be called after the context switch, paired
2516 * with a prepare_task_switch call before the context switch.
2517 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2518 * and do any other architecture-specific cleanup actions.
2520 * Note that we may have delayed dropping an mm in context_switch(). If
2521 * so, we finish that here outside of the runqueue lock. (Doing it
2522 * with the lock held can cause deadlocks; see schedule() for
2525 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2526 __releases(rq->lock)
2528 struct mm_struct *mm = rq->prev_mm;
2534 * A task struct has one reference for the use as "current".
2535 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2536 * schedule one last time. The schedule call will never return, and
2537 * the scheduled task must drop that reference.
2538 * The test for TASK_DEAD must occur while the runqueue locks are
2539 * still held, otherwise prev could be scheduled on another cpu, die
2540 * there before we look at prev->state, and then the reference would
2542 * Manfred Spraul <manfred@colorfullife.com>
2544 prev_state = prev->state;
2545 finish_arch_switch(prev);
2546 finish_lock_switch(rq, prev);
2548 if (current->sched_class->post_schedule)
2549 current->sched_class->post_schedule(rq);
2552 fire_sched_in_preempt_notifiers(current);
2555 if (unlikely(prev_state == TASK_DEAD)) {
2557 * Remove function-return probe instances associated with this
2558 * task and put them back on the free list.
2560 kprobe_flush_task(prev);
2561 put_task_struct(prev);
2566 * schedule_tail - first thing a freshly forked thread must call.
2567 * @prev: the thread we just switched away from.
2569 asmlinkage void schedule_tail(struct task_struct *prev)
2570 __releases(rq->lock)
2572 struct rq *rq = this_rq();
2574 finish_task_switch(rq, prev);
2575 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2576 /* In this case, finish_task_switch does not reenable preemption */
2579 if (current->set_child_tid)
2580 put_user(task_pid_vnr(current), current->set_child_tid);
2584 * context_switch - switch to the new MM and the new
2585 * thread's register state.
2588 context_switch(struct rq *rq, struct task_struct *prev,
2589 struct task_struct *next)
2591 struct mm_struct *mm, *oldmm;
2593 prepare_task_switch(rq, prev, next);
2594 trace_sched_switch(rq, prev, next);
2596 oldmm = prev->active_mm;
2598 * For paravirt, this is coupled with an exit in switch_to to
2599 * combine the page table reload and the switch backend into
2602 arch_enter_lazy_cpu_mode();
2604 if (unlikely(!mm)) {
2605 next->active_mm = oldmm;
2606 atomic_inc(&oldmm->mm_count);
2607 enter_lazy_tlb(oldmm, next);
2609 switch_mm(oldmm, mm, next);
2611 if (unlikely(!prev->mm)) {
2612 prev->active_mm = NULL;
2613 rq->prev_mm = oldmm;
2616 * Since the runqueue lock will be released by the next
2617 * task (which is an invalid locking op but in the case
2618 * of the scheduler it's an obvious special-case), so we
2619 * do an early lockdep release here:
2621 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2622 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2625 /* Here we just switch the register state and the stack. */
2626 switch_to(prev, next, prev);
2630 * this_rq must be evaluated again because prev may have moved
2631 * CPUs since it called schedule(), thus the 'rq' on its stack
2632 * frame will be invalid.
2634 finish_task_switch(this_rq(), prev);
2638 * nr_running, nr_uninterruptible and nr_context_switches:
2640 * externally visible scheduler statistics: current number of runnable
2641 * threads, current number of uninterruptible-sleeping threads, total
2642 * number of context switches performed since bootup.
2644 unsigned long nr_running(void)
2646 unsigned long i, sum = 0;
2648 for_each_online_cpu(i)
2649 sum += cpu_rq(i)->nr_running;
2654 unsigned long nr_uninterruptible(void)
2656 unsigned long i, sum = 0;
2658 for_each_possible_cpu(i)
2659 sum += cpu_rq(i)->nr_uninterruptible;
2662 * Since we read the counters lockless, it might be slightly
2663 * inaccurate. Do not allow it to go below zero though:
2665 if (unlikely((long)sum < 0))
2671 unsigned long long nr_context_switches(void)
2674 unsigned long long sum = 0;
2676 for_each_possible_cpu(i)
2677 sum += cpu_rq(i)->nr_switches;
2682 unsigned long nr_iowait(void)
2684 unsigned long i, sum = 0;
2686 for_each_possible_cpu(i)
2687 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2692 unsigned long nr_active(void)
2694 unsigned long i, running = 0, uninterruptible = 0;
2696 for_each_online_cpu(i) {
2697 running += cpu_rq(i)->nr_running;
2698 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2701 if (unlikely((long)uninterruptible < 0))
2702 uninterruptible = 0;
2704 return running + uninterruptible;
2708 * Update rq->cpu_load[] statistics. This function is usually called every
2709 * scheduler tick (TICK_NSEC).
2711 static void update_cpu_load(struct rq *this_rq)
2713 unsigned long this_load = this_rq->load.weight;
2716 this_rq->nr_load_updates++;
2718 /* Update our load: */
2719 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2720 unsigned long old_load, new_load;
2722 /* scale is effectively 1 << i now, and >> i divides by scale */
2724 old_load = this_rq->cpu_load[i];
2725 new_load = this_load;
2727 * Round up the averaging division if load is increasing. This
2728 * prevents us from getting stuck on 9 if the load is 10, for
2731 if (new_load > old_load)
2732 new_load += scale-1;
2733 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2740 * double_rq_lock - safely lock two runqueues
2742 * Note this does not disable interrupts like task_rq_lock,
2743 * you need to do so manually before calling.
2745 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2746 __acquires(rq1->lock)
2747 __acquires(rq2->lock)
2749 BUG_ON(!irqs_disabled());
2751 spin_lock(&rq1->lock);
2752 __acquire(rq2->lock); /* Fake it out ;) */
2755 spin_lock(&rq1->lock);
2756 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2758 spin_lock(&rq2->lock);
2759 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2762 update_rq_clock(rq1);
2763 update_rq_clock(rq2);
2767 * double_rq_unlock - safely unlock two runqueues
2769 * Note this does not restore interrupts like task_rq_unlock,
2770 * you need to do so manually after calling.
2772 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2773 __releases(rq1->lock)
2774 __releases(rq2->lock)
2776 spin_unlock(&rq1->lock);
2778 spin_unlock(&rq2->lock);
2780 __release(rq2->lock);
2784 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2786 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2787 __releases(this_rq->lock)
2788 __acquires(busiest->lock)
2789 __acquires(this_rq->lock)
2793 if (unlikely(!irqs_disabled())) {
2794 /* printk() doesn't work good under rq->lock */
2795 spin_unlock(&this_rq->lock);
2798 if (unlikely(!spin_trylock(&busiest->lock))) {
2799 if (busiest < this_rq) {
2800 spin_unlock(&this_rq->lock);
2801 spin_lock(&busiest->lock);
2802 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
2805 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
2810 static void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
2811 __releases(busiest->lock)
2813 spin_unlock(&busiest->lock);
2814 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
2818 * If dest_cpu is allowed for this process, migrate the task to it.
2819 * This is accomplished by forcing the cpu_allowed mask to only
2820 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2821 * the cpu_allowed mask is restored.
2823 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2825 struct migration_req req;
2826 unsigned long flags;
2829 rq = task_rq_lock(p, &flags);
2830 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2831 || unlikely(!cpu_active(dest_cpu)))
2834 trace_sched_migrate_task(rq, p, dest_cpu);
2835 /* force the process onto the specified CPU */
2836 if (migrate_task(p, dest_cpu, &req)) {
2837 /* Need to wait for migration thread (might exit: take ref). */
2838 struct task_struct *mt = rq->migration_thread;
2840 get_task_struct(mt);
2841 task_rq_unlock(rq, &flags);
2842 wake_up_process(mt);
2843 put_task_struct(mt);
2844 wait_for_completion(&req.done);
2849 task_rq_unlock(rq, &flags);
2853 * sched_exec - execve() is a valuable balancing opportunity, because at
2854 * this point the task has the smallest effective memory and cache footprint.
2856 void sched_exec(void)
2858 int new_cpu, this_cpu = get_cpu();
2859 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2861 if (new_cpu != this_cpu)
2862 sched_migrate_task(current, new_cpu);
2866 * pull_task - move a task from a remote runqueue to the local runqueue.
2867 * Both runqueues must be locked.
2869 static void pull_task(struct rq *src_rq, struct task_struct *p,
2870 struct rq *this_rq, int this_cpu)
2872 deactivate_task(src_rq, p, 0);
2873 set_task_cpu(p, this_cpu);
2874 activate_task(this_rq, p, 0);
2876 * Note that idle threads have a prio of MAX_PRIO, for this test
2877 * to be always true for them.
2879 check_preempt_curr(this_rq, p, 0);
2883 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2886 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2887 struct sched_domain *sd, enum cpu_idle_type idle,
2891 * We do not migrate tasks that are:
2892 * 1) running (obviously), or
2893 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2894 * 3) are cache-hot on their current CPU.
2896 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2897 schedstat_inc(p, se.nr_failed_migrations_affine);
2902 if (task_running(rq, p)) {
2903 schedstat_inc(p, se.nr_failed_migrations_running);
2908 * Aggressive migration if:
2909 * 1) task is cache cold, or
2910 * 2) too many balance attempts have failed.
2913 if (!task_hot(p, rq->clock, sd) ||
2914 sd->nr_balance_failed > sd->cache_nice_tries) {
2915 #ifdef CONFIG_SCHEDSTATS
2916 if (task_hot(p, rq->clock, sd)) {
2917 schedstat_inc(sd, lb_hot_gained[idle]);
2918 schedstat_inc(p, se.nr_forced_migrations);
2924 if (task_hot(p, rq->clock, sd)) {
2925 schedstat_inc(p, se.nr_failed_migrations_hot);
2931 static unsigned long
2932 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2933 unsigned long max_load_move, struct sched_domain *sd,
2934 enum cpu_idle_type idle, int *all_pinned,
2935 int *this_best_prio, struct rq_iterator *iterator)
2937 int loops = 0, pulled = 0, pinned = 0;
2938 struct task_struct *p;
2939 long rem_load_move = max_load_move;
2941 if (max_load_move == 0)
2947 * Start the load-balancing iterator:
2949 p = iterator->start(iterator->arg);
2951 if (!p || loops++ > sysctl_sched_nr_migrate)
2954 if ((p->se.load.weight >> 1) > rem_load_move ||
2955 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2956 p = iterator->next(iterator->arg);
2960 pull_task(busiest, p, this_rq, this_cpu);
2962 rem_load_move -= p->se.load.weight;
2965 * We only want to steal up to the prescribed amount of weighted load.
2967 if (rem_load_move > 0) {
2968 if (p->prio < *this_best_prio)
2969 *this_best_prio = p->prio;
2970 p = iterator->next(iterator->arg);
2975 * Right now, this is one of only two places pull_task() is called,
2976 * so we can safely collect pull_task() stats here rather than
2977 * inside pull_task().
2979 schedstat_add(sd, lb_gained[idle], pulled);
2982 *all_pinned = pinned;
2984 return max_load_move - rem_load_move;
2988 * move_tasks tries to move up to max_load_move weighted load from busiest to
2989 * this_rq, as part of a balancing operation within domain "sd".
2990 * Returns 1 if successful and 0 otherwise.
2992 * Called with both runqueues locked.
2994 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2995 unsigned long max_load_move,
2996 struct sched_domain *sd, enum cpu_idle_type idle,
2999 const struct sched_class *class = sched_class_highest;
3000 unsigned long total_load_moved = 0;
3001 int this_best_prio = this_rq->curr->prio;
3005 class->load_balance(this_rq, this_cpu, busiest,
3006 max_load_move - total_load_moved,
3007 sd, idle, all_pinned, &this_best_prio);
3008 class = class->next;
3010 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3013 } while (class && max_load_move > total_load_moved);
3015 return total_load_moved > 0;
3019 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3020 struct sched_domain *sd, enum cpu_idle_type idle,
3021 struct rq_iterator *iterator)
3023 struct task_struct *p = iterator->start(iterator->arg);
3027 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3028 pull_task(busiest, p, this_rq, this_cpu);
3030 * Right now, this is only the second place pull_task()
3031 * is called, so we can safely collect pull_task()
3032 * stats here rather than inside pull_task().
3034 schedstat_inc(sd, lb_gained[idle]);
3038 p = iterator->next(iterator->arg);
3045 * move_one_task tries to move exactly one task from busiest to this_rq, as
3046 * part of active balancing operations within "domain".
3047 * Returns 1 if successful and 0 otherwise.
3049 * Called with both runqueues locked.
3051 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3052 struct sched_domain *sd, enum cpu_idle_type idle)
3054 const struct sched_class *class;
3056 for (class = sched_class_highest; class; class = class->next)
3057 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3064 * find_busiest_group finds and returns the busiest CPU group within the
3065 * domain. It calculates and returns the amount of weighted load which
3066 * should be moved to restore balance via the imbalance parameter.
3068 static struct sched_group *
3069 find_busiest_group(struct sched_domain *sd, int this_cpu,
3070 unsigned long *imbalance, enum cpu_idle_type idle,
3071 int *sd_idle, const cpumask_t *cpus, int *balance)
3073 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3074 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3075 unsigned long max_pull;
3076 unsigned long busiest_load_per_task, busiest_nr_running;
3077 unsigned long this_load_per_task, this_nr_running;
3078 int load_idx, group_imb = 0;
3079 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3080 int power_savings_balance = 1;
3081 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3082 unsigned long min_nr_running = ULONG_MAX;
3083 struct sched_group *group_min = NULL, *group_leader = NULL;
3086 max_load = this_load = total_load = total_pwr = 0;
3087 busiest_load_per_task = busiest_nr_running = 0;
3088 this_load_per_task = this_nr_running = 0;
3090 if (idle == CPU_NOT_IDLE)
3091 load_idx = sd->busy_idx;
3092 else if (idle == CPU_NEWLY_IDLE)
3093 load_idx = sd->newidle_idx;
3095 load_idx = sd->idle_idx;
3098 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3101 int __group_imb = 0;
3102 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3103 unsigned long sum_nr_running, sum_weighted_load;
3104 unsigned long sum_avg_load_per_task;
3105 unsigned long avg_load_per_task;
3107 local_group = cpu_isset(this_cpu, group->cpumask);
3110 balance_cpu = first_cpu(group->cpumask);
3112 /* Tally up the load of all CPUs in the group */
3113 sum_weighted_load = sum_nr_running = avg_load = 0;
3114 sum_avg_load_per_task = avg_load_per_task = 0;
3117 min_cpu_load = ~0UL;
3119 for_each_cpu_mask_nr(i, group->cpumask) {
3122 if (!cpu_isset(i, *cpus))
3127 if (*sd_idle && rq->nr_running)
3130 /* Bias balancing toward cpus of our domain */
3132 if (idle_cpu(i) && !first_idle_cpu) {
3137 load = target_load(i, load_idx);
3139 load = source_load(i, load_idx);
3140 if (load > max_cpu_load)
3141 max_cpu_load = load;
3142 if (min_cpu_load > load)
3143 min_cpu_load = load;
3147 sum_nr_running += rq->nr_running;
3148 sum_weighted_load += weighted_cpuload(i);
3150 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3154 * First idle cpu or the first cpu(busiest) in this sched group
3155 * is eligible for doing load balancing at this and above
3156 * domains. In the newly idle case, we will allow all the cpu's
3157 * to do the newly idle load balance.
3159 if (idle != CPU_NEWLY_IDLE && local_group &&
3160 balance_cpu != this_cpu && balance) {
3165 total_load += avg_load;
3166 total_pwr += group->__cpu_power;
3168 /* Adjust by relative CPU power of the group */
3169 avg_load = sg_div_cpu_power(group,
3170 avg_load * SCHED_LOAD_SCALE);
3174 * Consider the group unbalanced when the imbalance is larger
3175 * than the average weight of two tasks.
3177 * APZ: with cgroup the avg task weight can vary wildly and
3178 * might not be a suitable number - should we keep a
3179 * normalized nr_running number somewhere that negates
3182 avg_load_per_task = sg_div_cpu_power(group,
3183 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3185 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3188 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3191 this_load = avg_load;
3193 this_nr_running = sum_nr_running;
3194 this_load_per_task = sum_weighted_load;
3195 } else if (avg_load > max_load &&
3196 (sum_nr_running > group_capacity || __group_imb)) {
3197 max_load = avg_load;
3199 busiest_nr_running = sum_nr_running;
3200 busiest_load_per_task = sum_weighted_load;
3201 group_imb = __group_imb;
3204 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3206 * Busy processors will not participate in power savings
3209 if (idle == CPU_NOT_IDLE ||
3210 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3214 * If the local group is idle or completely loaded
3215 * no need to do power savings balance at this domain
3217 if (local_group && (this_nr_running >= group_capacity ||
3219 power_savings_balance = 0;
3222 * If a group is already running at full capacity or idle,
3223 * don't include that group in power savings calculations
3225 if (!power_savings_balance || sum_nr_running >= group_capacity
3230 * Calculate the group which has the least non-idle load.
3231 * This is the group from where we need to pick up the load
3234 if ((sum_nr_running < min_nr_running) ||
3235 (sum_nr_running == min_nr_running &&
3236 first_cpu(group->cpumask) <
3237 first_cpu(group_min->cpumask))) {
3239 min_nr_running = sum_nr_running;
3240 min_load_per_task = sum_weighted_load /
3245 * Calculate the group which is almost near its
3246 * capacity but still has some space to pick up some load
3247 * from other group and save more power
3249 if (sum_nr_running <= group_capacity - 1) {
3250 if (sum_nr_running > leader_nr_running ||
3251 (sum_nr_running == leader_nr_running &&
3252 first_cpu(group->cpumask) >
3253 first_cpu(group_leader->cpumask))) {
3254 group_leader = group;
3255 leader_nr_running = sum_nr_running;
3260 group = group->next;
3261 } while (group != sd->groups);
3263 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3266 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3268 if (this_load >= avg_load ||
3269 100*max_load <= sd->imbalance_pct*this_load)
3272 busiest_load_per_task /= busiest_nr_running;
3274 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3277 * We're trying to get all the cpus to the average_load, so we don't
3278 * want to push ourselves above the average load, nor do we wish to
3279 * reduce the max loaded cpu below the average load, as either of these
3280 * actions would just result in more rebalancing later, and ping-pong
3281 * tasks around. Thus we look for the minimum possible imbalance.
3282 * Negative imbalances (*we* are more loaded than anyone else) will
3283 * be counted as no imbalance for these purposes -- we can't fix that
3284 * by pulling tasks to us. Be careful of negative numbers as they'll
3285 * appear as very large values with unsigned longs.
3287 if (max_load <= busiest_load_per_task)
3291 * In the presence of smp nice balancing, certain scenarios can have
3292 * max load less than avg load(as we skip the groups at or below
3293 * its cpu_power, while calculating max_load..)
3295 if (max_load < avg_load) {
3297 goto small_imbalance;
3300 /* Don't want to pull so many tasks that a group would go idle */
3301 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3303 /* How much load to actually move to equalise the imbalance */
3304 *imbalance = min(max_pull * busiest->__cpu_power,
3305 (avg_load - this_load) * this->__cpu_power)
3309 * if *imbalance is less than the average load per runnable task
3310 * there is no gaurantee that any tasks will be moved so we'll have
3311 * a think about bumping its value to force at least one task to be
3314 if (*imbalance < busiest_load_per_task) {
3315 unsigned long tmp, pwr_now, pwr_move;
3319 pwr_move = pwr_now = 0;
3321 if (this_nr_running) {
3322 this_load_per_task /= this_nr_running;
3323 if (busiest_load_per_task > this_load_per_task)
3326 this_load_per_task = cpu_avg_load_per_task(this_cpu);
3328 if (max_load - this_load + busiest_load_per_task >=
3329 busiest_load_per_task * imbn) {
3330 *imbalance = busiest_load_per_task;
3335 * OK, we don't have enough imbalance to justify moving tasks,
3336 * however we may be able to increase total CPU power used by
3340 pwr_now += busiest->__cpu_power *
3341 min(busiest_load_per_task, max_load);
3342 pwr_now += this->__cpu_power *
3343 min(this_load_per_task, this_load);
3344 pwr_now /= SCHED_LOAD_SCALE;
3346 /* Amount of load we'd subtract */
3347 tmp = sg_div_cpu_power(busiest,
3348 busiest_load_per_task * SCHED_LOAD_SCALE);
3350 pwr_move += busiest->__cpu_power *
3351 min(busiest_load_per_task, max_load - tmp);
3353 /* Amount of load we'd add */
3354 if (max_load * busiest->__cpu_power <
3355 busiest_load_per_task * SCHED_LOAD_SCALE)
3356 tmp = sg_div_cpu_power(this,
3357 max_load * busiest->__cpu_power);
3359 tmp = sg_div_cpu_power(this,
3360 busiest_load_per_task * SCHED_LOAD_SCALE);
3361 pwr_move += this->__cpu_power *
3362 min(this_load_per_task, this_load + tmp);
3363 pwr_move /= SCHED_LOAD_SCALE;
3365 /* Move if we gain throughput */
3366 if (pwr_move > pwr_now)
3367 *imbalance = busiest_load_per_task;
3373 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3374 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3377 if (this == group_leader && group_leader != group_min) {
3378 *imbalance = min_load_per_task;
3388 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3391 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3392 unsigned long imbalance, const cpumask_t *cpus)
3394 struct rq *busiest = NULL, *rq;
3395 unsigned long max_load = 0;
3398 for_each_cpu_mask_nr(i, group->cpumask) {
3401 if (!cpu_isset(i, *cpus))
3405 wl = weighted_cpuload(i);
3407 if (rq->nr_running == 1 && wl > imbalance)
3410 if (wl > max_load) {
3420 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3421 * so long as it is large enough.
3423 #define MAX_PINNED_INTERVAL 512
3426 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3427 * tasks if there is an imbalance.
3429 static int load_balance(int this_cpu, struct rq *this_rq,
3430 struct sched_domain *sd, enum cpu_idle_type idle,
3431 int *balance, cpumask_t *cpus)
3433 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3434 struct sched_group *group;
3435 unsigned long imbalance;
3437 unsigned long flags;
3442 * When power savings policy is enabled for the parent domain, idle
3443 * sibling can pick up load irrespective of busy siblings. In this case,
3444 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3445 * portraying it as CPU_NOT_IDLE.
3447 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3448 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3451 schedstat_inc(sd, lb_count[idle]);
3455 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3462 schedstat_inc(sd, lb_nobusyg[idle]);
3466 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3468 schedstat_inc(sd, lb_nobusyq[idle]);
3472 BUG_ON(busiest == this_rq);
3474 schedstat_add(sd, lb_imbalance[idle], imbalance);
3477 if (busiest->nr_running > 1) {
3479 * Attempt to move tasks. If find_busiest_group has found
3480 * an imbalance but busiest->nr_running <= 1, the group is
3481 * still unbalanced. ld_moved simply stays zero, so it is
3482 * correctly treated as an imbalance.
3484 local_irq_save(flags);
3485 double_rq_lock(this_rq, busiest);
3486 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3487 imbalance, sd, idle, &all_pinned);
3488 double_rq_unlock(this_rq, busiest);
3489 local_irq_restore(flags);
3492 * some other cpu did the load balance for us.
3494 if (ld_moved && this_cpu != smp_processor_id())
3495 resched_cpu(this_cpu);
3497 /* All tasks on this runqueue were pinned by CPU affinity */
3498 if (unlikely(all_pinned)) {
3499 cpu_clear(cpu_of(busiest), *cpus);
3500 if (!cpus_empty(*cpus))
3507 schedstat_inc(sd, lb_failed[idle]);
3508 sd->nr_balance_failed++;
3510 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3512 spin_lock_irqsave(&busiest->lock, flags);
3514 /* don't kick the migration_thread, if the curr
3515 * task on busiest cpu can't be moved to this_cpu
3517 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3518 spin_unlock_irqrestore(&busiest->lock, flags);
3520 goto out_one_pinned;
3523 if (!busiest->active_balance) {
3524 busiest->active_balance = 1;
3525 busiest->push_cpu = this_cpu;
3528 spin_unlock_irqrestore(&busiest->lock, flags);
3530 wake_up_process(busiest->migration_thread);
3533 * We've kicked active balancing, reset the failure
3536 sd->nr_balance_failed = sd->cache_nice_tries+1;
3539 sd->nr_balance_failed = 0;
3541 if (likely(!active_balance)) {
3542 /* We were unbalanced, so reset the balancing interval */
3543 sd->balance_interval = sd->min_interval;
3546 * If we've begun active balancing, start to back off. This
3547 * case may not be covered by the all_pinned logic if there
3548 * is only 1 task on the busy runqueue (because we don't call
3551 if (sd->balance_interval < sd->max_interval)
3552 sd->balance_interval *= 2;
3555 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3556 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3562 schedstat_inc(sd, lb_balanced[idle]);
3564 sd->nr_balance_failed = 0;
3567 /* tune up the balancing interval */
3568 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3569 (sd->balance_interval < sd->max_interval))
3570 sd->balance_interval *= 2;
3572 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3573 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3584 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3585 * tasks if there is an imbalance.
3587 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3588 * this_rq is locked.
3591 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3594 struct sched_group *group;
3595 struct rq *busiest = NULL;
3596 unsigned long imbalance;
3604 * When power savings policy is enabled for the parent domain, idle
3605 * sibling can pick up load irrespective of busy siblings. In this case,
3606 * let the state of idle sibling percolate up as IDLE, instead of
3607 * portraying it as CPU_NOT_IDLE.
3609 if (sd->flags & SD_SHARE_CPUPOWER &&
3610 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3613 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3615 update_shares_locked(this_rq, sd);
3616 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3617 &sd_idle, cpus, NULL);
3619 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3623 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3625 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3629 BUG_ON(busiest == this_rq);
3631 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3634 if (busiest->nr_running > 1) {
3635 /* Attempt to move tasks */
3636 double_lock_balance(this_rq, busiest);
3637 /* this_rq->clock is already updated */
3638 update_rq_clock(busiest);
3639 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3640 imbalance, sd, CPU_NEWLY_IDLE,
3642 double_unlock_balance(this_rq, busiest);
3644 if (unlikely(all_pinned)) {
3645 cpu_clear(cpu_of(busiest), *cpus);
3646 if (!cpus_empty(*cpus))
3652 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3653 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3654 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3657 sd->nr_balance_failed = 0;
3659 update_shares_locked(this_rq, sd);
3663 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3664 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3665 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3667 sd->nr_balance_failed = 0;
3673 * idle_balance is called by schedule() if this_cpu is about to become
3674 * idle. Attempts to pull tasks from other CPUs.
3676 static void idle_balance(int this_cpu, struct rq *this_rq)
3678 struct sched_domain *sd;
3679 int pulled_task = -1;
3680 unsigned long next_balance = jiffies + HZ;
3683 for_each_domain(this_cpu, sd) {
3684 unsigned long interval;
3686 if (!(sd->flags & SD_LOAD_BALANCE))
3689 if (sd->flags & SD_BALANCE_NEWIDLE)
3690 /* If we've pulled tasks over stop searching: */
3691 pulled_task = load_balance_newidle(this_cpu, this_rq,
3694 interval = msecs_to_jiffies(sd->balance_interval);
3695 if (time_after(next_balance, sd->last_balance + interval))
3696 next_balance = sd->last_balance + interval;
3700 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3702 * We are going idle. next_balance may be set based on
3703 * a busy processor. So reset next_balance.
3705 this_rq->next_balance = next_balance;
3710 * active_load_balance is run by migration threads. It pushes running tasks
3711 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3712 * running on each physical CPU where possible, and avoids physical /
3713 * logical imbalances.
3715 * Called with busiest_rq locked.
3717 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3719 int target_cpu = busiest_rq->push_cpu;
3720 struct sched_domain *sd;
3721 struct rq *target_rq;
3723 /* Is there any task to move? */
3724 if (busiest_rq->nr_running <= 1)
3727 target_rq = cpu_rq(target_cpu);
3730 * This condition is "impossible", if it occurs
3731 * we need to fix it. Originally reported by
3732 * Bjorn Helgaas on a 128-cpu setup.
3734 BUG_ON(busiest_rq == target_rq);
3736 /* move a task from busiest_rq to target_rq */
3737 double_lock_balance(busiest_rq, target_rq);
3738 update_rq_clock(busiest_rq);
3739 update_rq_clock(target_rq);
3741 /* Search for an sd spanning us and the target CPU. */
3742 for_each_domain(target_cpu, sd) {
3743 if ((sd->flags & SD_LOAD_BALANCE) &&
3744 cpu_isset(busiest_cpu, sd->span))
3749 schedstat_inc(sd, alb_count);
3751 if (move_one_task(target_rq, target_cpu, busiest_rq,
3753 schedstat_inc(sd, alb_pushed);
3755 schedstat_inc(sd, alb_failed);
3757 double_unlock_balance(busiest_rq, target_rq);
3762 atomic_t load_balancer;
3764 } nohz ____cacheline_aligned = {
3765 .load_balancer = ATOMIC_INIT(-1),
3766 .cpu_mask = CPU_MASK_NONE,
3770 * This routine will try to nominate the ilb (idle load balancing)
3771 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3772 * load balancing on behalf of all those cpus. If all the cpus in the system
3773 * go into this tickless mode, then there will be no ilb owner (as there is
3774 * no need for one) and all the cpus will sleep till the next wakeup event
3777 * For the ilb owner, tick is not stopped. And this tick will be used
3778 * for idle load balancing. ilb owner will still be part of
3781 * While stopping the tick, this cpu will become the ilb owner if there
3782 * is no other owner. And will be the owner till that cpu becomes busy
3783 * or if all cpus in the system stop their ticks at which point
3784 * there is no need for ilb owner.
3786 * When the ilb owner becomes busy, it nominates another owner, during the
3787 * next busy scheduler_tick()
3789 int select_nohz_load_balancer(int stop_tick)
3791 int cpu = smp_processor_id();
3794 cpu_set(cpu, nohz.cpu_mask);
3795 cpu_rq(cpu)->in_nohz_recently = 1;
3798 * If we are going offline and still the leader, give up!
3800 if (!cpu_active(cpu) &&
3801 atomic_read(&nohz.load_balancer) == cpu) {
3802 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3807 /* time for ilb owner also to sleep */
3808 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3809 if (atomic_read(&nohz.load_balancer) == cpu)
3810 atomic_set(&nohz.load_balancer, -1);
3814 if (atomic_read(&nohz.load_balancer) == -1) {
3815 /* make me the ilb owner */
3816 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3818 } else if (atomic_read(&nohz.load_balancer) == cpu)
3821 if (!cpu_isset(cpu, nohz.cpu_mask))
3824 cpu_clear(cpu, nohz.cpu_mask);
3826 if (atomic_read(&nohz.load_balancer) == cpu)
3827 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3834 static DEFINE_SPINLOCK(balancing);
3837 * It checks each scheduling domain to see if it is due to be balanced,
3838 * and initiates a balancing operation if so.
3840 * Balancing parameters are set up in arch_init_sched_domains.
3842 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3845 struct rq *rq = cpu_rq(cpu);
3846 unsigned long interval;
3847 struct sched_domain *sd;
3848 /* Earliest time when we have to do rebalance again */
3849 unsigned long next_balance = jiffies + 60*HZ;
3850 int update_next_balance = 0;
3854 for_each_domain(cpu, sd) {
3855 if (!(sd->flags & SD_LOAD_BALANCE))
3858 interval = sd->balance_interval;
3859 if (idle != CPU_IDLE)
3860 interval *= sd->busy_factor;
3862 /* scale ms to jiffies */
3863 interval = msecs_to_jiffies(interval);
3864 if (unlikely(!interval))
3866 if (interval > HZ*NR_CPUS/10)
3867 interval = HZ*NR_CPUS/10;
3869 need_serialize = sd->flags & SD_SERIALIZE;
3871 if (need_serialize) {
3872 if (!spin_trylock(&balancing))
3876 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3877 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3879 * We've pulled tasks over so either we're no
3880 * longer idle, or one of our SMT siblings is
3883 idle = CPU_NOT_IDLE;
3885 sd->last_balance = jiffies;
3888 spin_unlock(&balancing);
3890 if (time_after(next_balance, sd->last_balance + interval)) {
3891 next_balance = sd->last_balance + interval;
3892 update_next_balance = 1;
3896 * Stop the load balance at this level. There is another
3897 * CPU in our sched group which is doing load balancing more
3905 * next_balance will be updated only when there is a need.
3906 * When the cpu is attached to null domain for ex, it will not be
3909 if (likely(update_next_balance))
3910 rq->next_balance = next_balance;
3914 * run_rebalance_domains is triggered when needed from the scheduler tick.
3915 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3916 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3918 static void run_rebalance_domains(struct softirq_action *h)
3920 int this_cpu = smp_processor_id();
3921 struct rq *this_rq = cpu_rq(this_cpu);
3922 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3923 CPU_IDLE : CPU_NOT_IDLE;
3925 rebalance_domains(this_cpu, idle);
3929 * If this cpu is the owner for idle load balancing, then do the
3930 * balancing on behalf of the other idle cpus whose ticks are
3933 if (this_rq->idle_at_tick &&
3934 atomic_read(&nohz.load_balancer) == this_cpu) {
3935 cpumask_t cpus = nohz.cpu_mask;
3939 cpu_clear(this_cpu, cpus);
3940 for_each_cpu_mask_nr(balance_cpu, cpus) {
3942 * If this cpu gets work to do, stop the load balancing
3943 * work being done for other cpus. Next load
3944 * balancing owner will pick it up.
3949 rebalance_domains(balance_cpu, CPU_IDLE);
3951 rq = cpu_rq(balance_cpu);
3952 if (time_after(this_rq->next_balance, rq->next_balance))
3953 this_rq->next_balance = rq->next_balance;
3960 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3962 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3963 * idle load balancing owner or decide to stop the periodic load balancing,
3964 * if the whole system is idle.
3966 static inline void trigger_load_balance(struct rq *rq, int cpu)
3970 * If we were in the nohz mode recently and busy at the current
3971 * scheduler tick, then check if we need to nominate new idle
3974 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3975 rq->in_nohz_recently = 0;
3977 if (atomic_read(&nohz.load_balancer) == cpu) {
3978 cpu_clear(cpu, nohz.cpu_mask);
3979 atomic_set(&nohz.load_balancer, -1);
3982 if (atomic_read(&nohz.load_balancer) == -1) {
3984 * simple selection for now: Nominate the
3985 * first cpu in the nohz list to be the next
3988 * TBD: Traverse the sched domains and nominate
3989 * the nearest cpu in the nohz.cpu_mask.
3991 int ilb = first_cpu(nohz.cpu_mask);
3993 if (ilb < nr_cpu_ids)
3999 * If this cpu is idle and doing idle load balancing for all the
4000 * cpus with ticks stopped, is it time for that to stop?
4002 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4003 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4009 * If this cpu is idle and the idle load balancing is done by
4010 * someone else, then no need raise the SCHED_SOFTIRQ
4012 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4013 cpu_isset(cpu, nohz.cpu_mask))
4016 if (time_after_eq(jiffies, rq->next_balance))
4017 raise_softirq(SCHED_SOFTIRQ);
4020 #else /* CONFIG_SMP */
4023 * on UP we do not need to balance between CPUs:
4025 static inline void idle_balance(int cpu, struct rq *rq)
4031 DEFINE_PER_CPU(struct kernel_stat, kstat);
4033 EXPORT_PER_CPU_SYMBOL(kstat);
4036 * Return any ns on the sched_clock that have not yet been banked in
4037 * @p in case that task is currently running.
4039 unsigned long long task_delta_exec(struct task_struct *p)
4041 unsigned long flags;
4045 rq = task_rq_lock(p, &flags);
4047 if (task_current(rq, p)) {
4050 update_rq_clock(rq);
4051 delta_exec = rq->clock - p->se.exec_start;
4052 if ((s64)delta_exec > 0)
4056 task_rq_unlock(rq, &flags);
4062 * Account user cpu time to a process.
4063 * @p: the process that the cpu time gets accounted to
4064 * @cputime: the cpu time spent in user space since the last update
4066 void account_user_time(struct task_struct *p, cputime_t cputime)
4068 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4071 p->utime = cputime_add(p->utime, cputime);
4072 account_group_user_time(p, cputime);
4074 /* Add user time to cpustat. */
4075 tmp = cputime_to_cputime64(cputime);
4076 if (TASK_NICE(p) > 0)
4077 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4079 cpustat->user = cputime64_add(cpustat->user, tmp);
4080 /* Account for user time used */
4081 acct_update_integrals(p);
4085 * Account guest cpu time to a process.
4086 * @p: the process that the cpu time gets accounted to
4087 * @cputime: the cpu time spent in virtual machine since the last update
4089 static void account_guest_time(struct task_struct *p, cputime_t cputime)
4092 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4094 tmp = cputime_to_cputime64(cputime);
4096 p->utime = cputime_add(p->utime, cputime);
4097 account_group_user_time(p, cputime);
4098 p->gtime = cputime_add(p->gtime, cputime);
4100 cpustat->user = cputime64_add(cpustat->user, tmp);
4101 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4105 * Account scaled user cpu time to a process.
4106 * @p: the process that the cpu time gets accounted to
4107 * @cputime: the cpu time spent in user space since the last update
4109 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4111 p->utimescaled = cputime_add(p->utimescaled, cputime);
4115 * Account system cpu time to a process.
4116 * @p: the process that the cpu time gets accounted to
4117 * @hardirq_offset: the offset to subtract from hardirq_count()
4118 * @cputime: the cpu time spent in kernel space since the last update
4120 void account_system_time(struct task_struct *p, int hardirq_offset,
4123 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4124 struct rq *rq = this_rq();
4127 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4128 account_guest_time(p, cputime);
4132 p->stime = cputime_add(p->stime, cputime);
4133 account_group_system_time(p, cputime);
4135 /* Add system time to cpustat. */
4136 tmp = cputime_to_cputime64(cputime);
4137 if (hardirq_count() - hardirq_offset)
4138 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4139 else if (softirq_count())
4140 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4141 else if (p != rq->idle)
4142 cpustat->system = cputime64_add(cpustat->system, tmp);
4143 else if (atomic_read(&rq->nr_iowait) > 0)
4144 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4146 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4147 /* Account for system time used */
4148 acct_update_integrals(p);
4152 * Account scaled system cpu time to a process.
4153 * @p: the process that the cpu time gets accounted to
4154 * @hardirq_offset: the offset to subtract from hardirq_count()
4155 * @cputime: the cpu time spent in kernel space since the last update
4157 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4159 p->stimescaled = cputime_add(p->stimescaled, cputime);
4163 * Account for involuntary wait time.
4164 * @p: the process from which the cpu time has been stolen
4165 * @steal: the cpu time spent in involuntary wait
4167 void account_steal_time(struct task_struct *p, cputime_t steal)
4169 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4170 cputime64_t tmp = cputime_to_cputime64(steal);
4171 struct rq *rq = this_rq();
4173 if (p == rq->idle) {
4174 p->stime = cputime_add(p->stime, steal);
4175 account_group_system_time(p, steal);
4176 if (atomic_read(&rq->nr_iowait) > 0)
4177 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4179 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4181 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4185 * Use precise platform statistics if available:
4187 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4188 cputime_t task_utime(struct task_struct *p)
4193 cputime_t task_stime(struct task_struct *p)
4198 cputime_t task_utime(struct task_struct *p)
4200 clock_t utime = cputime_to_clock_t(p->utime),
4201 total = utime + cputime_to_clock_t(p->stime);
4205 * Use CFS's precise accounting:
4207 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
4211 do_div(temp, total);
4213 utime = (clock_t)temp;
4215 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
4216 return p->prev_utime;
4219 cputime_t task_stime(struct task_struct *p)
4224 * Use CFS's precise accounting. (we subtract utime from
4225 * the total, to make sure the total observed by userspace
4226 * grows monotonically - apps rely on that):
4228 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
4229 cputime_to_clock_t(task_utime(p));
4232 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
4234 return p->prev_stime;
4238 inline cputime_t task_gtime(struct task_struct *p)
4244 * This function gets called by the timer code, with HZ frequency.
4245 * We call it with interrupts disabled.
4247 * It also gets called by the fork code, when changing the parent's
4250 void scheduler_tick(void)
4252 int cpu = smp_processor_id();
4253 struct rq *rq = cpu_rq(cpu);
4254 struct task_struct *curr = rq->curr;
4258 spin_lock(&rq->lock);
4259 update_rq_clock(rq);
4260 update_cpu_load(rq);
4261 curr->sched_class->task_tick(rq, curr, 0);
4262 spin_unlock(&rq->lock);
4265 rq->idle_at_tick = idle_cpu(cpu);
4266 trigger_load_balance(rq, cpu);
4270 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4271 defined(CONFIG_PREEMPT_TRACER))
4273 static inline unsigned long get_parent_ip(unsigned long addr)
4275 if (in_lock_functions(addr)) {
4276 addr = CALLER_ADDR2;
4277 if (in_lock_functions(addr))
4278 addr = CALLER_ADDR3;
4283 void __kprobes add_preempt_count(int val)
4285 #ifdef CONFIG_DEBUG_PREEMPT
4289 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4292 preempt_count() += val;
4293 #ifdef CONFIG_DEBUG_PREEMPT
4295 * Spinlock count overflowing soon?
4297 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4300 if (preempt_count() == val)
4301 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4303 EXPORT_SYMBOL(add_preempt_count);
4305 void __kprobes sub_preempt_count(int val)
4307 #ifdef CONFIG_DEBUG_PREEMPT
4311 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4314 * Is the spinlock portion underflowing?
4316 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4317 !(preempt_count() & PREEMPT_MASK)))
4321 if (preempt_count() == val)
4322 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4323 preempt_count() -= val;
4325 EXPORT_SYMBOL(sub_preempt_count);
4330 * Print scheduling while atomic bug:
4332 static noinline void __schedule_bug(struct task_struct *prev)
4334 struct pt_regs *regs = get_irq_regs();
4336 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4337 prev->comm, prev->pid, preempt_count());
4339 debug_show_held_locks(prev);
4341 if (irqs_disabled())
4342 print_irqtrace_events(prev);
4351 * Various schedule()-time debugging checks and statistics:
4353 static inline void schedule_debug(struct task_struct *prev)
4356 * Test if we are atomic. Since do_exit() needs to call into
4357 * schedule() atomically, we ignore that path for now.
4358 * Otherwise, whine if we are scheduling when we should not be.
4360 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4361 __schedule_bug(prev);
4363 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4365 schedstat_inc(this_rq(), sched_count);
4366 #ifdef CONFIG_SCHEDSTATS
4367 if (unlikely(prev->lock_depth >= 0)) {
4368 schedstat_inc(this_rq(), bkl_count);
4369 schedstat_inc(prev, sched_info.bkl_count);
4375 * Pick up the highest-prio task:
4377 static inline struct task_struct *
4378 pick_next_task(struct rq *rq, struct task_struct *prev)
4380 const struct sched_class *class;
4381 struct task_struct *p;
4384 * Optimization: we know that if all tasks are in
4385 * the fair class we can call that function directly:
4387 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4388 p = fair_sched_class.pick_next_task(rq);
4393 class = sched_class_highest;
4395 p = class->pick_next_task(rq);
4399 * Will never be NULL as the idle class always
4400 * returns a non-NULL p:
4402 class = class->next;
4407 * schedule() is the main scheduler function.
4409 asmlinkage void __sched schedule(void)
4411 struct task_struct *prev, *next;
4412 unsigned long *switch_count;
4418 cpu = smp_processor_id();
4422 switch_count = &prev->nivcsw;
4424 release_kernel_lock(prev);
4425 need_resched_nonpreemptible:
4427 schedule_debug(prev);
4429 if (sched_feat(HRTICK))
4432 spin_lock_irq(&rq->lock);
4433 update_rq_clock(rq);
4434 clear_tsk_need_resched(prev);
4436 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4437 if (unlikely(signal_pending_state(prev->state, prev)))
4438 prev->state = TASK_RUNNING;
4440 deactivate_task(rq, prev, 1);
4441 switch_count = &prev->nvcsw;
4445 if (prev->sched_class->pre_schedule)
4446 prev->sched_class->pre_schedule(rq, prev);
4449 if (unlikely(!rq->nr_running))
4450 idle_balance(cpu, rq);
4452 prev->sched_class->put_prev_task(rq, prev);
4453 next = pick_next_task(rq, prev);
4455 if (likely(prev != next)) {
4456 sched_info_switch(prev, next);
4462 context_switch(rq, prev, next); /* unlocks the rq */
4464 * the context switch might have flipped the stack from under
4465 * us, hence refresh the local variables.
4467 cpu = smp_processor_id();
4470 spin_unlock_irq(&rq->lock);
4472 if (unlikely(reacquire_kernel_lock(current) < 0))
4473 goto need_resched_nonpreemptible;
4475 preempt_enable_no_resched();
4476 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4479 EXPORT_SYMBOL(schedule);
4481 #ifdef CONFIG_PREEMPT
4483 * this is the entry point to schedule() from in-kernel preemption
4484 * off of preempt_enable. Kernel preemptions off return from interrupt
4485 * occur there and call schedule directly.
4487 asmlinkage void __sched preempt_schedule(void)
4489 struct thread_info *ti = current_thread_info();
4492 * If there is a non-zero preempt_count or interrupts are disabled,
4493 * we do not want to preempt the current task. Just return..
4495 if (likely(ti->preempt_count || irqs_disabled()))
4499 add_preempt_count(PREEMPT_ACTIVE);
4501 sub_preempt_count(PREEMPT_ACTIVE);
4504 * Check again in case we missed a preemption opportunity
4505 * between schedule and now.
4508 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4510 EXPORT_SYMBOL(preempt_schedule);
4513 * this is the entry point to schedule() from kernel preemption
4514 * off of irq context.
4515 * Note, that this is called and return with irqs disabled. This will
4516 * protect us against recursive calling from irq.
4518 asmlinkage void __sched preempt_schedule_irq(void)
4520 struct thread_info *ti = current_thread_info();
4522 /* Catch callers which need to be fixed */
4523 BUG_ON(ti->preempt_count || !irqs_disabled());
4526 add_preempt_count(PREEMPT_ACTIVE);
4529 local_irq_disable();
4530 sub_preempt_count(PREEMPT_ACTIVE);
4533 * Check again in case we missed a preemption opportunity
4534 * between schedule and now.
4537 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4540 #endif /* CONFIG_PREEMPT */
4542 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4545 return try_to_wake_up(curr->private, mode, sync);
4547 EXPORT_SYMBOL(default_wake_function);
4550 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4551 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4552 * number) then we wake all the non-exclusive tasks and one exclusive task.
4554 * There are circumstances in which we can try to wake a task which has already
4555 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4556 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4558 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4559 int nr_exclusive, int sync, void *key)
4561 wait_queue_t *curr, *next;
4563 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4564 unsigned flags = curr->flags;
4566 if (curr->func(curr, mode, sync, key) &&
4567 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4573 * __wake_up - wake up threads blocked on a waitqueue.
4575 * @mode: which threads
4576 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4577 * @key: is directly passed to the wakeup function
4579 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4580 int nr_exclusive, void *key)
4582 unsigned long flags;
4584 spin_lock_irqsave(&q->lock, flags);
4585 __wake_up_common(q, mode, nr_exclusive, 0, key);
4586 spin_unlock_irqrestore(&q->lock, flags);
4588 EXPORT_SYMBOL(__wake_up);
4591 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4593 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4595 __wake_up_common(q, mode, 1, 0, NULL);
4599 * __wake_up_sync - wake up threads blocked on a waitqueue.
4601 * @mode: which threads
4602 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4604 * The sync wakeup differs that the waker knows that it will schedule
4605 * away soon, so while the target thread will be woken up, it will not
4606 * be migrated to another CPU - ie. the two threads are 'synchronized'
4607 * with each other. This can prevent needless bouncing between CPUs.
4609 * On UP it can prevent extra preemption.
4612 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4614 unsigned long flags;
4620 if (unlikely(!nr_exclusive))
4623 spin_lock_irqsave(&q->lock, flags);
4624 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4625 spin_unlock_irqrestore(&q->lock, flags);
4627 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4630 * complete: - signals a single thread waiting on this completion
4631 * @x: holds the state of this particular completion
4633 * This will wake up a single thread waiting on this completion. Threads will be
4634 * awakened in the same order in which they were queued.
4636 * See also complete_all(), wait_for_completion() and related routines.
4638 void complete(struct completion *x)
4640 unsigned long flags;
4642 spin_lock_irqsave(&x->wait.lock, flags);
4644 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4645 spin_unlock_irqrestore(&x->wait.lock, flags);
4647 EXPORT_SYMBOL(complete);
4650 * complete_all: - signals all threads waiting on this completion
4651 * @x: holds the state of this particular completion
4653 * This will wake up all threads waiting on this particular completion event.
4655 void complete_all(struct completion *x)
4657 unsigned long flags;
4659 spin_lock_irqsave(&x->wait.lock, flags);
4660 x->done += UINT_MAX/2;
4661 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4662 spin_unlock_irqrestore(&x->wait.lock, flags);
4664 EXPORT_SYMBOL(complete_all);
4666 static inline long __sched
4667 do_wait_for_common(struct completion *x, long timeout, int state)
4670 DECLARE_WAITQUEUE(wait, current);
4672 wait.flags |= WQ_FLAG_EXCLUSIVE;
4673 __add_wait_queue_tail(&x->wait, &wait);
4675 if (signal_pending_state(state, current)) {
4676 timeout = -ERESTARTSYS;
4679 __set_current_state(state);
4680 spin_unlock_irq(&x->wait.lock);
4681 timeout = schedule_timeout(timeout);
4682 spin_lock_irq(&x->wait.lock);
4683 } while (!x->done && timeout);
4684 __remove_wait_queue(&x->wait, &wait);
4689 return timeout ?: 1;
4693 wait_for_common(struct completion *x, long timeout, int state)
4697 spin_lock_irq(&x->wait.lock);
4698 timeout = do_wait_for_common(x, timeout, state);
4699 spin_unlock_irq(&x->wait.lock);
4704 * wait_for_completion: - waits for completion of a task
4705 * @x: holds the state of this particular completion
4707 * This waits to be signaled for completion of a specific task. It is NOT
4708 * interruptible and there is no timeout.
4710 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4711 * and interrupt capability. Also see complete().
4713 void __sched wait_for_completion(struct completion *x)
4715 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4717 EXPORT_SYMBOL(wait_for_completion);
4720 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4721 * @x: holds the state of this particular completion
4722 * @timeout: timeout value in jiffies
4724 * This waits for either a completion of a specific task to be signaled or for a
4725 * specified timeout to expire. The timeout is in jiffies. It is not
4728 unsigned long __sched
4729 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4731 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4733 EXPORT_SYMBOL(wait_for_completion_timeout);
4736 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4737 * @x: holds the state of this particular completion
4739 * This waits for completion of a specific task to be signaled. It is
4742 int __sched wait_for_completion_interruptible(struct completion *x)
4744 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4745 if (t == -ERESTARTSYS)
4749 EXPORT_SYMBOL(wait_for_completion_interruptible);
4752 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4753 * @x: holds the state of this particular completion
4754 * @timeout: timeout value in jiffies
4756 * This waits for either a completion of a specific task to be signaled or for a
4757 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4759 unsigned long __sched
4760 wait_for_completion_interruptible_timeout(struct completion *x,
4761 unsigned long timeout)
4763 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4765 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4768 * wait_for_completion_killable: - waits for completion of a task (killable)
4769 * @x: holds the state of this particular completion
4771 * This waits to be signaled for completion of a specific task. It can be
4772 * interrupted by a kill signal.
4774 int __sched wait_for_completion_killable(struct completion *x)
4776 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4777 if (t == -ERESTARTSYS)
4781 EXPORT_SYMBOL(wait_for_completion_killable);
4784 * try_wait_for_completion - try to decrement a completion without blocking
4785 * @x: completion structure
4787 * Returns: 0 if a decrement cannot be done without blocking
4788 * 1 if a decrement succeeded.
4790 * If a completion is being used as a counting completion,
4791 * attempt to decrement the counter without blocking. This
4792 * enables us to avoid waiting if the resource the completion
4793 * is protecting is not available.
4795 bool try_wait_for_completion(struct completion *x)
4799 spin_lock_irq(&x->wait.lock);
4804 spin_unlock_irq(&x->wait.lock);
4807 EXPORT_SYMBOL(try_wait_for_completion);
4810 * completion_done - Test to see if a completion has any waiters
4811 * @x: completion structure
4813 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4814 * 1 if there are no waiters.
4817 bool completion_done(struct completion *x)
4821 spin_lock_irq(&x->wait.lock);
4824 spin_unlock_irq(&x->wait.lock);
4827 EXPORT_SYMBOL(completion_done);
4830 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4832 unsigned long flags;
4835 init_waitqueue_entry(&wait, current);
4837 __set_current_state(state);
4839 spin_lock_irqsave(&q->lock, flags);
4840 __add_wait_queue(q, &wait);
4841 spin_unlock(&q->lock);
4842 timeout = schedule_timeout(timeout);
4843 spin_lock_irq(&q->lock);
4844 __remove_wait_queue(q, &wait);
4845 spin_unlock_irqrestore(&q->lock, flags);
4850 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4852 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4854 EXPORT_SYMBOL(interruptible_sleep_on);
4857 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4859 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4861 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4863 void __sched sleep_on(wait_queue_head_t *q)
4865 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4867 EXPORT_SYMBOL(sleep_on);
4869 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4871 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4873 EXPORT_SYMBOL(sleep_on_timeout);
4875 #ifdef CONFIG_RT_MUTEXES
4878 * rt_mutex_setprio - set the current priority of a task
4880 * @prio: prio value (kernel-internal form)
4882 * This function changes the 'effective' priority of a task. It does
4883 * not touch ->normal_prio like __setscheduler().
4885 * Used by the rt_mutex code to implement priority inheritance logic.
4887 void rt_mutex_setprio(struct task_struct *p, int prio)
4889 unsigned long flags;
4890 int oldprio, on_rq, running;
4892 const struct sched_class *prev_class = p->sched_class;
4894 BUG_ON(prio < 0 || prio > MAX_PRIO);
4896 rq = task_rq_lock(p, &flags);
4897 update_rq_clock(rq);
4900 on_rq = p->se.on_rq;
4901 running = task_current(rq, p);
4903 dequeue_task(rq, p, 0);
4905 p->sched_class->put_prev_task(rq, p);
4908 p->sched_class = &rt_sched_class;
4910 p->sched_class = &fair_sched_class;
4915 p->sched_class->set_curr_task(rq);
4917 enqueue_task(rq, p, 0);
4919 check_class_changed(rq, p, prev_class, oldprio, running);
4921 task_rq_unlock(rq, &flags);
4926 void set_user_nice(struct task_struct *p, long nice)
4928 int old_prio, delta, on_rq;
4929 unsigned long flags;
4932 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4935 * We have to be careful, if called from sys_setpriority(),
4936 * the task might be in the middle of scheduling on another CPU.
4938 rq = task_rq_lock(p, &flags);
4939 update_rq_clock(rq);
4941 * The RT priorities are set via sched_setscheduler(), but we still
4942 * allow the 'normal' nice value to be set - but as expected
4943 * it wont have any effect on scheduling until the task is
4944 * SCHED_FIFO/SCHED_RR:
4946 if (task_has_rt_policy(p)) {
4947 p->static_prio = NICE_TO_PRIO(nice);
4950 on_rq = p->se.on_rq;
4952 dequeue_task(rq, p, 0);
4954 p->static_prio = NICE_TO_PRIO(nice);
4957 p->prio = effective_prio(p);
4958 delta = p->prio - old_prio;
4961 enqueue_task(rq, p, 0);
4963 * If the task increased its priority or is running and
4964 * lowered its priority, then reschedule its CPU:
4966 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4967 resched_task(rq->curr);
4970 task_rq_unlock(rq, &flags);
4972 EXPORT_SYMBOL(set_user_nice);
4975 * can_nice - check if a task can reduce its nice value
4979 int can_nice(const struct task_struct *p, const int nice)
4981 /* convert nice value [19,-20] to rlimit style value [1,40] */
4982 int nice_rlim = 20 - nice;
4984 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4985 capable(CAP_SYS_NICE));
4988 #ifdef __ARCH_WANT_SYS_NICE
4991 * sys_nice - change the priority of the current process.
4992 * @increment: priority increment
4994 * sys_setpriority is a more generic, but much slower function that
4995 * does similar things.
4997 asmlinkage long sys_nice(int increment)
5002 * Setpriority might change our priority at the same moment.
5003 * We don't have to worry. Conceptually one call occurs first
5004 * and we have a single winner.
5006 if (increment < -40)
5011 nice = PRIO_TO_NICE(current->static_prio) + increment;
5017 if (increment < 0 && !can_nice(current, nice))
5020 retval = security_task_setnice(current, nice);
5024 set_user_nice(current, nice);
5031 * task_prio - return the priority value of a given task.
5032 * @p: the task in question.
5034 * This is the priority value as seen by users in /proc.
5035 * RT tasks are offset by -200. Normal tasks are centered
5036 * around 0, value goes from -16 to +15.
5038 int task_prio(const struct task_struct *p)
5040 return p->prio - MAX_RT_PRIO;
5044 * task_nice - return the nice value of a given task.
5045 * @p: the task in question.
5047 int task_nice(const struct task_struct *p)
5049 return TASK_NICE(p);
5051 EXPORT_SYMBOL(task_nice);
5054 * idle_cpu - is a given cpu idle currently?
5055 * @cpu: the processor in question.
5057 int idle_cpu(int cpu)
5059 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5063 * idle_task - return the idle task for a given cpu.
5064 * @cpu: the processor in question.
5066 struct task_struct *idle_task(int cpu)
5068 return cpu_rq(cpu)->idle;
5072 * find_process_by_pid - find a process with a matching PID value.
5073 * @pid: the pid in question.
5075 static struct task_struct *find_process_by_pid(pid_t pid)
5077 return pid ? find_task_by_vpid(pid) : current;
5080 /* Actually do priority change: must hold rq lock. */
5082 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5084 BUG_ON(p->se.on_rq);
5087 switch (p->policy) {
5091 p->sched_class = &fair_sched_class;
5095 p->sched_class = &rt_sched_class;
5099 p->rt_priority = prio;
5100 p->normal_prio = normal_prio(p);
5101 /* we are holding p->pi_lock already */
5102 p->prio = rt_mutex_getprio(p);
5106 static int __sched_setscheduler(struct task_struct *p, int policy,
5107 struct sched_param *param, bool user)
5109 int retval, oldprio, oldpolicy = -1, on_rq, running;
5110 unsigned long flags;
5111 const struct sched_class *prev_class = p->sched_class;
5114 /* may grab non-irq protected spin_locks */
5115 BUG_ON(in_interrupt());
5117 /* double check policy once rq lock held */
5119 policy = oldpolicy = p->policy;
5120 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5121 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5122 policy != SCHED_IDLE)
5125 * Valid priorities for SCHED_FIFO and SCHED_RR are
5126 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5127 * SCHED_BATCH and SCHED_IDLE is 0.
5129 if (param->sched_priority < 0 ||
5130 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5131 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5133 if (rt_policy(policy) != (param->sched_priority != 0))
5137 * Allow unprivileged RT tasks to decrease priority:
5139 if (user && !capable(CAP_SYS_NICE)) {
5140 if (rt_policy(policy)) {
5141 unsigned long rlim_rtprio;
5143 if (!lock_task_sighand(p, &flags))
5145 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5146 unlock_task_sighand(p, &flags);
5148 /* can't set/change the rt policy */
5149 if (policy != p->policy && !rlim_rtprio)
5152 /* can't increase priority */
5153 if (param->sched_priority > p->rt_priority &&
5154 param->sched_priority > rlim_rtprio)
5158 * Like positive nice levels, dont allow tasks to
5159 * move out of SCHED_IDLE either:
5161 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5164 /* can't change other user's priorities */
5165 if ((current->euid != p->euid) &&
5166 (current->euid != p->uid))
5171 #ifdef CONFIG_RT_GROUP_SCHED
5173 * Do not allow realtime tasks into groups that have no runtime
5176 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5177 task_group(p)->rt_bandwidth.rt_runtime == 0)
5181 retval = security_task_setscheduler(p, policy, param);
5187 * make sure no PI-waiters arrive (or leave) while we are
5188 * changing the priority of the task:
5190 spin_lock_irqsave(&p->pi_lock, flags);
5192 * To be able to change p->policy safely, the apropriate
5193 * runqueue lock must be held.
5195 rq = __task_rq_lock(p);
5196 /* recheck policy now with rq lock held */
5197 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5198 policy = oldpolicy = -1;
5199 __task_rq_unlock(rq);
5200 spin_unlock_irqrestore(&p->pi_lock, flags);
5203 update_rq_clock(rq);
5204 on_rq = p->se.on_rq;
5205 running = task_current(rq, p);
5207 deactivate_task(rq, p, 0);
5209 p->sched_class->put_prev_task(rq, p);
5212 __setscheduler(rq, p, policy, param->sched_priority);
5215 p->sched_class->set_curr_task(rq);
5217 activate_task(rq, p, 0);
5219 check_class_changed(rq, p, prev_class, oldprio, running);
5221 __task_rq_unlock(rq);
5222 spin_unlock_irqrestore(&p->pi_lock, flags);
5224 rt_mutex_adjust_pi(p);
5230 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5231 * @p: the task in question.
5232 * @policy: new policy.
5233 * @param: structure containing the new RT priority.
5235 * NOTE that the task may be already dead.
5237 int sched_setscheduler(struct task_struct *p, int policy,
5238 struct sched_param *param)
5240 return __sched_setscheduler(p, policy, param, true);
5242 EXPORT_SYMBOL_GPL(sched_setscheduler);
5245 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5246 * @p: the task in question.
5247 * @policy: new policy.
5248 * @param: structure containing the new RT priority.
5250 * Just like sched_setscheduler, only don't bother checking if the
5251 * current context has permission. For example, this is needed in
5252 * stop_machine(): we create temporary high priority worker threads,
5253 * but our caller might not have that capability.
5255 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5256 struct sched_param *param)
5258 return __sched_setscheduler(p, policy, param, false);
5262 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5264 struct sched_param lparam;
5265 struct task_struct *p;
5268 if (!param || pid < 0)
5270 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5275 p = find_process_by_pid(pid);
5277 retval = sched_setscheduler(p, policy, &lparam);
5284 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5285 * @pid: the pid in question.
5286 * @policy: new policy.
5287 * @param: structure containing the new RT priority.
5290 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5292 /* negative values for policy are not valid */
5296 return do_sched_setscheduler(pid, policy, param);
5300 * sys_sched_setparam - set/change the RT priority of a thread
5301 * @pid: the pid in question.
5302 * @param: structure containing the new RT priority.
5304 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5306 return do_sched_setscheduler(pid, -1, param);
5310 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5311 * @pid: the pid in question.
5313 asmlinkage long sys_sched_getscheduler(pid_t pid)
5315 struct task_struct *p;
5322 read_lock(&tasklist_lock);
5323 p = find_process_by_pid(pid);
5325 retval = security_task_getscheduler(p);
5329 read_unlock(&tasklist_lock);
5334 * sys_sched_getscheduler - get the RT priority of a thread
5335 * @pid: the pid in question.
5336 * @param: structure containing the RT priority.
5338 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5340 struct sched_param lp;
5341 struct task_struct *p;
5344 if (!param || pid < 0)
5347 read_lock(&tasklist_lock);
5348 p = find_process_by_pid(pid);
5353 retval = security_task_getscheduler(p);
5357 lp.sched_priority = p->rt_priority;
5358 read_unlock(&tasklist_lock);
5361 * This one might sleep, we cannot do it with a spinlock held ...
5363 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5368 read_unlock(&tasklist_lock);
5372 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5374 cpumask_t cpus_allowed;
5375 cpumask_t new_mask = *in_mask;
5376 struct task_struct *p;
5380 read_lock(&tasklist_lock);
5382 p = find_process_by_pid(pid);
5384 read_unlock(&tasklist_lock);
5390 * It is not safe to call set_cpus_allowed with the
5391 * tasklist_lock held. We will bump the task_struct's
5392 * usage count and then drop tasklist_lock.
5395 read_unlock(&tasklist_lock);
5398 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5399 !capable(CAP_SYS_NICE))
5402 retval = security_task_setscheduler(p, 0, NULL);
5406 cpuset_cpus_allowed(p, &cpus_allowed);
5407 cpus_and(new_mask, new_mask, cpus_allowed);
5409 retval = set_cpus_allowed_ptr(p, &new_mask);
5412 cpuset_cpus_allowed(p, &cpus_allowed);
5413 if (!cpus_subset(new_mask, cpus_allowed)) {
5415 * We must have raced with a concurrent cpuset
5416 * update. Just reset the cpus_allowed to the
5417 * cpuset's cpus_allowed
5419 new_mask = cpus_allowed;
5429 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5430 cpumask_t *new_mask)
5432 if (len < sizeof(cpumask_t)) {
5433 memset(new_mask, 0, sizeof(cpumask_t));
5434 } else if (len > sizeof(cpumask_t)) {
5435 len = sizeof(cpumask_t);
5437 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5441 * sys_sched_setaffinity - set the cpu affinity of a process
5442 * @pid: pid of the process
5443 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5444 * @user_mask_ptr: user-space pointer to the new cpu mask
5446 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5447 unsigned long __user *user_mask_ptr)
5452 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5456 return sched_setaffinity(pid, &new_mask);
5459 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5461 struct task_struct *p;
5465 read_lock(&tasklist_lock);
5468 p = find_process_by_pid(pid);
5472 retval = security_task_getscheduler(p);
5476 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5479 read_unlock(&tasklist_lock);
5486 * sys_sched_getaffinity - get the cpu affinity of a process
5487 * @pid: pid of the process
5488 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5489 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5491 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5492 unsigned long __user *user_mask_ptr)
5497 if (len < sizeof(cpumask_t))
5500 ret = sched_getaffinity(pid, &mask);
5504 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5507 return sizeof(cpumask_t);
5511 * sys_sched_yield - yield the current processor to other threads.
5513 * This function yields the current CPU to other tasks. If there are no
5514 * other threads running on this CPU then this function will return.
5516 asmlinkage long sys_sched_yield(void)
5518 struct rq *rq = this_rq_lock();
5520 schedstat_inc(rq, yld_count);
5521 current->sched_class->yield_task(rq);
5524 * Since we are going to call schedule() anyway, there's
5525 * no need to preempt or enable interrupts:
5527 __release(rq->lock);
5528 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5529 _raw_spin_unlock(&rq->lock);
5530 preempt_enable_no_resched();
5537 static void __cond_resched(void)
5539 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5540 __might_sleep(__FILE__, __LINE__);
5543 * The BKS might be reacquired before we have dropped
5544 * PREEMPT_ACTIVE, which could trigger a second
5545 * cond_resched() call.
5548 add_preempt_count(PREEMPT_ACTIVE);
5550 sub_preempt_count(PREEMPT_ACTIVE);
5551 } while (need_resched());
5554 int __sched _cond_resched(void)
5556 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5557 system_state == SYSTEM_RUNNING) {
5563 EXPORT_SYMBOL(_cond_resched);
5566 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5567 * call schedule, and on return reacquire the lock.
5569 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5570 * operations here to prevent schedule() from being called twice (once via
5571 * spin_unlock(), once by hand).
5573 int cond_resched_lock(spinlock_t *lock)
5575 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5578 if (spin_needbreak(lock) || resched) {
5580 if (resched && need_resched())
5589 EXPORT_SYMBOL(cond_resched_lock);
5591 int __sched cond_resched_softirq(void)
5593 BUG_ON(!in_softirq());
5595 if (need_resched() && system_state == SYSTEM_RUNNING) {
5603 EXPORT_SYMBOL(cond_resched_softirq);
5606 * yield - yield the current processor to other threads.
5608 * This is a shortcut for kernel-space yielding - it marks the
5609 * thread runnable and calls sys_sched_yield().
5611 void __sched yield(void)
5613 set_current_state(TASK_RUNNING);
5616 EXPORT_SYMBOL(yield);
5619 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5620 * that process accounting knows that this is a task in IO wait state.
5622 * But don't do that if it is a deliberate, throttling IO wait (this task
5623 * has set its backing_dev_info: the queue against which it should throttle)
5625 void __sched io_schedule(void)
5627 struct rq *rq = &__raw_get_cpu_var(runqueues);
5629 delayacct_blkio_start();
5630 atomic_inc(&rq->nr_iowait);
5632 atomic_dec(&rq->nr_iowait);
5633 delayacct_blkio_end();
5635 EXPORT_SYMBOL(io_schedule);
5637 long __sched io_schedule_timeout(long timeout)
5639 struct rq *rq = &__raw_get_cpu_var(runqueues);
5642 delayacct_blkio_start();
5643 atomic_inc(&rq->nr_iowait);
5644 ret = schedule_timeout(timeout);
5645 atomic_dec(&rq->nr_iowait);
5646 delayacct_blkio_end();
5651 * sys_sched_get_priority_max - return maximum RT priority.
5652 * @policy: scheduling class.
5654 * this syscall returns the maximum rt_priority that can be used
5655 * by a given scheduling class.
5657 asmlinkage long sys_sched_get_priority_max(int policy)
5664 ret = MAX_USER_RT_PRIO-1;
5676 * sys_sched_get_priority_min - return minimum RT priority.
5677 * @policy: scheduling class.
5679 * this syscall returns the minimum rt_priority that can be used
5680 * by a given scheduling class.
5682 asmlinkage long sys_sched_get_priority_min(int policy)
5700 * sys_sched_rr_get_interval - return the default timeslice of a process.
5701 * @pid: pid of the process.
5702 * @interval: userspace pointer to the timeslice value.
5704 * this syscall writes the default timeslice value of a given process
5705 * into the user-space timespec buffer. A value of '0' means infinity.
5708 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5710 struct task_struct *p;
5711 unsigned int time_slice;
5719 read_lock(&tasklist_lock);
5720 p = find_process_by_pid(pid);
5724 retval = security_task_getscheduler(p);
5729 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5730 * tasks that are on an otherwise idle runqueue:
5733 if (p->policy == SCHED_RR) {
5734 time_slice = DEF_TIMESLICE;
5735 } else if (p->policy != SCHED_FIFO) {
5736 struct sched_entity *se = &p->se;
5737 unsigned long flags;
5740 rq = task_rq_lock(p, &flags);
5741 if (rq->cfs.load.weight)
5742 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5743 task_rq_unlock(rq, &flags);
5745 read_unlock(&tasklist_lock);
5746 jiffies_to_timespec(time_slice, &t);
5747 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5751 read_unlock(&tasklist_lock);
5755 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5757 void sched_show_task(struct task_struct *p)
5759 unsigned long free = 0;
5762 state = p->state ? __ffs(p->state) + 1 : 0;
5763 printk(KERN_INFO "%-13.13s %c", p->comm,
5764 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5765 #if BITS_PER_LONG == 32
5766 if (state == TASK_RUNNING)
5767 printk(KERN_CONT " running ");
5769 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5771 if (state == TASK_RUNNING)
5772 printk(KERN_CONT " running task ");
5774 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5776 #ifdef CONFIG_DEBUG_STACK_USAGE
5778 unsigned long *n = end_of_stack(p);
5781 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5784 printk(KERN_CONT "%5lu %5d %6d\n", free,
5785 task_pid_nr(p), task_pid_nr(p->real_parent));
5787 show_stack(p, NULL);
5790 void show_state_filter(unsigned long state_filter)
5792 struct task_struct *g, *p;
5794 #if BITS_PER_LONG == 32
5796 " task PC stack pid father\n");
5799 " task PC stack pid father\n");
5801 read_lock(&tasklist_lock);
5802 do_each_thread(g, p) {
5804 * reset the NMI-timeout, listing all files on a slow
5805 * console might take alot of time:
5807 touch_nmi_watchdog();
5808 if (!state_filter || (p->state & state_filter))
5810 } while_each_thread(g, p);
5812 touch_all_softlockup_watchdogs();
5814 #ifdef CONFIG_SCHED_DEBUG
5815 sysrq_sched_debug_show();
5817 read_unlock(&tasklist_lock);
5819 * Only show locks if all tasks are dumped:
5821 if (state_filter == -1)
5822 debug_show_all_locks();
5825 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5827 idle->sched_class = &idle_sched_class;
5831 * init_idle - set up an idle thread for a given CPU
5832 * @idle: task in question
5833 * @cpu: cpu the idle task belongs to
5835 * NOTE: this function does not set the idle thread's NEED_RESCHED
5836 * flag, to make booting more robust.
5838 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5840 struct rq *rq = cpu_rq(cpu);
5841 unsigned long flags;
5844 idle->se.exec_start = sched_clock();
5846 idle->prio = idle->normal_prio = MAX_PRIO;
5847 idle->cpus_allowed = cpumask_of_cpu(cpu);
5848 __set_task_cpu(idle, cpu);
5850 spin_lock_irqsave(&rq->lock, flags);
5851 rq->curr = rq->idle = idle;
5852 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5855 spin_unlock_irqrestore(&rq->lock, flags);
5857 /* Set the preempt count _outside_ the spinlocks! */
5858 #if defined(CONFIG_PREEMPT)
5859 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5861 task_thread_info(idle)->preempt_count = 0;
5864 * The idle tasks have their own, simple scheduling class:
5866 idle->sched_class = &idle_sched_class;
5870 * In a system that switches off the HZ timer nohz_cpu_mask
5871 * indicates which cpus entered this state. This is used
5872 * in the rcu update to wait only for active cpus. For system
5873 * which do not switch off the HZ timer nohz_cpu_mask should
5874 * always be CPU_MASK_NONE.
5876 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5879 * Increase the granularity value when there are more CPUs,
5880 * because with more CPUs the 'effective latency' as visible
5881 * to users decreases. But the relationship is not linear,
5882 * so pick a second-best guess by going with the log2 of the
5885 * This idea comes from the SD scheduler of Con Kolivas:
5887 static inline void sched_init_granularity(void)
5889 unsigned int factor = 1 + ilog2(num_online_cpus());
5890 const unsigned long limit = 200000000;
5892 sysctl_sched_min_granularity *= factor;
5893 if (sysctl_sched_min_granularity > limit)
5894 sysctl_sched_min_granularity = limit;
5896 sysctl_sched_latency *= factor;
5897 if (sysctl_sched_latency > limit)
5898 sysctl_sched_latency = limit;
5900 sysctl_sched_wakeup_granularity *= factor;
5902 sysctl_sched_shares_ratelimit *= factor;
5907 * This is how migration works:
5909 * 1) we queue a struct migration_req structure in the source CPU's
5910 * runqueue and wake up that CPU's migration thread.
5911 * 2) we down() the locked semaphore => thread blocks.
5912 * 3) migration thread wakes up (implicitly it forces the migrated
5913 * thread off the CPU)
5914 * 4) it gets the migration request and checks whether the migrated
5915 * task is still in the wrong runqueue.
5916 * 5) if it's in the wrong runqueue then the migration thread removes
5917 * it and puts it into the right queue.
5918 * 6) migration thread up()s the semaphore.
5919 * 7) we wake up and the migration is done.
5923 * Change a given task's CPU affinity. Migrate the thread to a
5924 * proper CPU and schedule it away if the CPU it's executing on
5925 * is removed from the allowed bitmask.
5927 * NOTE: the caller must have a valid reference to the task, the
5928 * task must not exit() & deallocate itself prematurely. The
5929 * call is not atomic; no spinlocks may be held.
5931 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5933 struct migration_req req;
5934 unsigned long flags;
5938 rq = task_rq_lock(p, &flags);
5939 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5944 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5945 !cpus_equal(p->cpus_allowed, *new_mask))) {
5950 if (p->sched_class->set_cpus_allowed)
5951 p->sched_class->set_cpus_allowed(p, new_mask);
5953 p->cpus_allowed = *new_mask;
5954 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5957 /* Can the task run on the task's current CPU? If so, we're done */
5958 if (cpu_isset(task_cpu(p), *new_mask))
5961 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5962 /* Need help from migration thread: drop lock and wait. */
5963 task_rq_unlock(rq, &flags);
5964 wake_up_process(rq->migration_thread);
5965 wait_for_completion(&req.done);
5966 tlb_migrate_finish(p->mm);
5970 task_rq_unlock(rq, &flags);
5974 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5977 * Move (not current) task off this cpu, onto dest cpu. We're doing
5978 * this because either it can't run here any more (set_cpus_allowed()
5979 * away from this CPU, or CPU going down), or because we're
5980 * attempting to rebalance this task on exec (sched_exec).
5982 * So we race with normal scheduler movements, but that's OK, as long
5983 * as the task is no longer on this CPU.
5985 * Returns non-zero if task was successfully migrated.
5987 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5989 struct rq *rq_dest, *rq_src;
5992 if (unlikely(!cpu_active(dest_cpu)))
5995 rq_src = cpu_rq(src_cpu);
5996 rq_dest = cpu_rq(dest_cpu);
5998 double_rq_lock(rq_src, rq_dest);
5999 /* Already moved. */
6000 if (task_cpu(p) != src_cpu)
6002 /* Affinity changed (again). */
6003 if (!cpu_isset(dest_cpu, p->cpus_allowed))
6006 on_rq = p->se.on_rq;
6008 deactivate_task(rq_src, p, 0);
6010 set_task_cpu(p, dest_cpu);
6012 activate_task(rq_dest, p, 0);
6013 check_preempt_curr(rq_dest, p, 0);
6018 double_rq_unlock(rq_src, rq_dest);
6023 * migration_thread - this is a highprio system thread that performs
6024 * thread migration by bumping thread off CPU then 'pushing' onto
6027 static int migration_thread(void *data)
6029 int cpu = (long)data;
6033 BUG_ON(rq->migration_thread != current);
6035 set_current_state(TASK_INTERRUPTIBLE);
6036 while (!kthread_should_stop()) {
6037 struct migration_req *req;
6038 struct list_head *head;
6040 spin_lock_irq(&rq->lock);
6042 if (cpu_is_offline(cpu)) {
6043 spin_unlock_irq(&rq->lock);
6047 if (rq->active_balance) {
6048 active_load_balance(rq, cpu);
6049 rq->active_balance = 0;
6052 head = &rq->migration_queue;
6054 if (list_empty(head)) {
6055 spin_unlock_irq(&rq->lock);
6057 set_current_state(TASK_INTERRUPTIBLE);
6060 req = list_entry(head->next, struct migration_req, list);
6061 list_del_init(head->next);
6063 spin_unlock(&rq->lock);
6064 __migrate_task(req->task, cpu, req->dest_cpu);
6067 complete(&req->done);
6069 __set_current_state(TASK_RUNNING);
6073 /* Wait for kthread_stop */
6074 set_current_state(TASK_INTERRUPTIBLE);
6075 while (!kthread_should_stop()) {
6077 set_current_state(TASK_INTERRUPTIBLE);
6079 __set_current_state(TASK_RUNNING);
6083 #ifdef CONFIG_HOTPLUG_CPU
6085 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6089 local_irq_disable();
6090 ret = __migrate_task(p, src_cpu, dest_cpu);
6096 * Figure out where task on dead CPU should go, use force if necessary.
6097 * NOTE: interrupts should be disabled by the caller
6099 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6101 unsigned long flags;
6108 mask = node_to_cpumask(cpu_to_node(dead_cpu));
6109 cpus_and(mask, mask, p->cpus_allowed);
6110 dest_cpu = any_online_cpu(mask);
6112 /* On any allowed CPU? */
6113 if (dest_cpu >= nr_cpu_ids)
6114 dest_cpu = any_online_cpu(p->cpus_allowed);
6116 /* No more Mr. Nice Guy. */
6117 if (dest_cpu >= nr_cpu_ids) {
6118 cpumask_t cpus_allowed;
6120 cpuset_cpus_allowed_locked(p, &cpus_allowed);
6122 * Try to stay on the same cpuset, where the
6123 * current cpuset may be a subset of all cpus.
6124 * The cpuset_cpus_allowed_locked() variant of
6125 * cpuset_cpus_allowed() will not block. It must be
6126 * called within calls to cpuset_lock/cpuset_unlock.
6128 rq = task_rq_lock(p, &flags);
6129 p->cpus_allowed = cpus_allowed;
6130 dest_cpu = any_online_cpu(p->cpus_allowed);
6131 task_rq_unlock(rq, &flags);
6134 * Don't tell them about moving exiting tasks or
6135 * kernel threads (both mm NULL), since they never
6138 if (p->mm && printk_ratelimit()) {
6139 printk(KERN_INFO "process %d (%s) no "
6140 "longer affine to cpu%d\n",
6141 task_pid_nr(p), p->comm, dead_cpu);
6144 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
6148 * While a dead CPU has no uninterruptible tasks queued at this point,
6149 * it might still have a nonzero ->nr_uninterruptible counter, because
6150 * for performance reasons the counter is not stricly tracking tasks to
6151 * their home CPUs. So we just add the counter to another CPU's counter,
6152 * to keep the global sum constant after CPU-down:
6154 static void migrate_nr_uninterruptible(struct rq *rq_src)
6156 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
6157 unsigned long flags;
6159 local_irq_save(flags);
6160 double_rq_lock(rq_src, rq_dest);
6161 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6162 rq_src->nr_uninterruptible = 0;
6163 double_rq_unlock(rq_src, rq_dest);
6164 local_irq_restore(flags);
6167 /* Run through task list and migrate tasks from the dead cpu. */
6168 static void migrate_live_tasks(int src_cpu)
6170 struct task_struct *p, *t;
6172 read_lock(&tasklist_lock);
6174 do_each_thread(t, p) {
6178 if (task_cpu(p) == src_cpu)
6179 move_task_off_dead_cpu(src_cpu, p);
6180 } while_each_thread(t, p);
6182 read_unlock(&tasklist_lock);
6186 * Schedules idle task to be the next runnable task on current CPU.
6187 * It does so by boosting its priority to highest possible.
6188 * Used by CPU offline code.
6190 void sched_idle_next(void)
6192 int this_cpu = smp_processor_id();
6193 struct rq *rq = cpu_rq(this_cpu);
6194 struct task_struct *p = rq->idle;
6195 unsigned long flags;
6197 /* cpu has to be offline */
6198 BUG_ON(cpu_online(this_cpu));
6201 * Strictly not necessary since rest of the CPUs are stopped by now
6202 * and interrupts disabled on the current cpu.
6204 spin_lock_irqsave(&rq->lock, flags);
6206 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6208 update_rq_clock(rq);
6209 activate_task(rq, p, 0);
6211 spin_unlock_irqrestore(&rq->lock, flags);
6215 * Ensures that the idle task is using init_mm right before its cpu goes
6218 void idle_task_exit(void)
6220 struct mm_struct *mm = current->active_mm;
6222 BUG_ON(cpu_online(smp_processor_id()));
6225 switch_mm(mm, &init_mm, current);
6229 /* called under rq->lock with disabled interrupts */
6230 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6232 struct rq *rq = cpu_rq(dead_cpu);
6234 /* Must be exiting, otherwise would be on tasklist. */
6235 BUG_ON(!p->exit_state);
6237 /* Cannot have done final schedule yet: would have vanished. */
6238 BUG_ON(p->state == TASK_DEAD);
6243 * Drop lock around migration; if someone else moves it,
6244 * that's OK. No task can be added to this CPU, so iteration is
6247 spin_unlock_irq(&rq->lock);
6248 move_task_off_dead_cpu(dead_cpu, p);
6249 spin_lock_irq(&rq->lock);
6254 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6255 static void migrate_dead_tasks(unsigned int dead_cpu)
6257 struct rq *rq = cpu_rq(dead_cpu);
6258 struct task_struct *next;
6261 if (!rq->nr_running)
6263 update_rq_clock(rq);
6264 next = pick_next_task(rq, rq->curr);
6267 next->sched_class->put_prev_task(rq, next);
6268 migrate_dead(dead_cpu, next);
6272 #endif /* CONFIG_HOTPLUG_CPU */
6274 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6276 static struct ctl_table sd_ctl_dir[] = {
6278 .procname = "sched_domain",
6284 static struct ctl_table sd_ctl_root[] = {
6286 .ctl_name = CTL_KERN,
6287 .procname = "kernel",
6289 .child = sd_ctl_dir,
6294 static struct ctl_table *sd_alloc_ctl_entry(int n)
6296 struct ctl_table *entry =
6297 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6302 static void sd_free_ctl_entry(struct ctl_table **tablep)
6304 struct ctl_table *entry;
6307 * In the intermediate directories, both the child directory and
6308 * procname are dynamically allocated and could fail but the mode
6309 * will always be set. In the lowest directory the names are
6310 * static strings and all have proc handlers.
6312 for (entry = *tablep; entry->mode; entry++) {
6314 sd_free_ctl_entry(&entry->child);
6315 if (entry->proc_handler == NULL)
6316 kfree(entry->procname);
6324 set_table_entry(struct ctl_table *entry,
6325 const char *procname, void *data, int maxlen,
6326 mode_t mode, proc_handler *proc_handler)
6328 entry->procname = procname;
6330 entry->maxlen = maxlen;
6332 entry->proc_handler = proc_handler;
6335 static struct ctl_table *
6336 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6338 struct ctl_table *table = sd_alloc_ctl_entry(13);
6343 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6344 sizeof(long), 0644, proc_doulongvec_minmax);
6345 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6346 sizeof(long), 0644, proc_doulongvec_minmax);
6347 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6348 sizeof(int), 0644, proc_dointvec_minmax);
6349 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6350 sizeof(int), 0644, proc_dointvec_minmax);
6351 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6352 sizeof(int), 0644, proc_dointvec_minmax);
6353 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6354 sizeof(int), 0644, proc_dointvec_minmax);
6355 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6356 sizeof(int), 0644, proc_dointvec_minmax);
6357 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6358 sizeof(int), 0644, proc_dointvec_minmax);
6359 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6360 sizeof(int), 0644, proc_dointvec_minmax);
6361 set_table_entry(&table[9], "cache_nice_tries",
6362 &sd->cache_nice_tries,
6363 sizeof(int), 0644, proc_dointvec_minmax);
6364 set_table_entry(&table[10], "flags", &sd->flags,
6365 sizeof(int), 0644, proc_dointvec_minmax);
6366 set_table_entry(&table[11], "name", sd->name,
6367 CORENAME_MAX_SIZE, 0444, proc_dostring);
6368 /* &table[12] is terminator */
6373 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6375 struct ctl_table *entry, *table;
6376 struct sched_domain *sd;
6377 int domain_num = 0, i;
6380 for_each_domain(cpu, sd)
6382 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6387 for_each_domain(cpu, sd) {
6388 snprintf(buf, 32, "domain%d", i);
6389 entry->procname = kstrdup(buf, GFP_KERNEL);
6391 entry->child = sd_alloc_ctl_domain_table(sd);
6398 static struct ctl_table_header *sd_sysctl_header;
6399 static void register_sched_domain_sysctl(void)
6401 int i, cpu_num = num_online_cpus();
6402 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6405 WARN_ON(sd_ctl_dir[0].child);
6406 sd_ctl_dir[0].child = entry;
6411 for_each_online_cpu(i) {
6412 snprintf(buf, 32, "cpu%d", i);
6413 entry->procname = kstrdup(buf, GFP_KERNEL);
6415 entry->child = sd_alloc_ctl_cpu_table(i);
6419 WARN_ON(sd_sysctl_header);
6420 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6423 /* may be called multiple times per register */
6424 static void unregister_sched_domain_sysctl(void)
6426 if (sd_sysctl_header)
6427 unregister_sysctl_table(sd_sysctl_header);
6428 sd_sysctl_header = NULL;
6429 if (sd_ctl_dir[0].child)
6430 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6433 static void register_sched_domain_sysctl(void)
6436 static void unregister_sched_domain_sysctl(void)
6441 static void set_rq_online(struct rq *rq)
6444 const struct sched_class *class;
6446 cpu_set(rq->cpu, rq->rd->online);
6449 for_each_class(class) {
6450 if (class->rq_online)
6451 class->rq_online(rq);
6456 static void set_rq_offline(struct rq *rq)
6459 const struct sched_class *class;
6461 for_each_class(class) {
6462 if (class->rq_offline)
6463 class->rq_offline(rq);
6466 cpu_clear(rq->cpu, rq->rd->online);
6472 * migration_call - callback that gets triggered when a CPU is added.
6473 * Here we can start up the necessary migration thread for the new CPU.
6475 static int __cpuinit
6476 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6478 struct task_struct *p;
6479 int cpu = (long)hcpu;
6480 unsigned long flags;
6485 case CPU_UP_PREPARE:
6486 case CPU_UP_PREPARE_FROZEN:
6487 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6490 kthread_bind(p, cpu);
6491 /* Must be high prio: stop_machine expects to yield to it. */
6492 rq = task_rq_lock(p, &flags);
6493 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6494 task_rq_unlock(rq, &flags);
6495 cpu_rq(cpu)->migration_thread = p;
6499 case CPU_ONLINE_FROZEN:
6500 /* Strictly unnecessary, as first user will wake it. */
6501 wake_up_process(cpu_rq(cpu)->migration_thread);
6503 /* Update our root-domain */
6505 spin_lock_irqsave(&rq->lock, flags);
6507 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6511 spin_unlock_irqrestore(&rq->lock, flags);
6514 #ifdef CONFIG_HOTPLUG_CPU
6515 case CPU_UP_CANCELED:
6516 case CPU_UP_CANCELED_FROZEN:
6517 if (!cpu_rq(cpu)->migration_thread)
6519 /* Unbind it from offline cpu so it can run. Fall thru. */
6520 kthread_bind(cpu_rq(cpu)->migration_thread,
6521 any_online_cpu(cpu_online_map));
6522 kthread_stop(cpu_rq(cpu)->migration_thread);
6523 cpu_rq(cpu)->migration_thread = NULL;
6527 case CPU_DEAD_FROZEN:
6528 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6529 migrate_live_tasks(cpu);
6531 kthread_stop(rq->migration_thread);
6532 rq->migration_thread = NULL;
6533 /* Idle task back to normal (off runqueue, low prio) */
6534 spin_lock_irq(&rq->lock);
6535 update_rq_clock(rq);
6536 deactivate_task(rq, rq->idle, 0);
6537 rq->idle->static_prio = MAX_PRIO;
6538 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6539 rq->idle->sched_class = &idle_sched_class;
6540 migrate_dead_tasks(cpu);
6541 spin_unlock_irq(&rq->lock);
6543 migrate_nr_uninterruptible(rq);
6544 BUG_ON(rq->nr_running != 0);
6547 * No need to migrate the tasks: it was best-effort if
6548 * they didn't take sched_hotcpu_mutex. Just wake up
6551 spin_lock_irq(&rq->lock);
6552 while (!list_empty(&rq->migration_queue)) {
6553 struct migration_req *req;
6555 req = list_entry(rq->migration_queue.next,
6556 struct migration_req, list);
6557 list_del_init(&req->list);
6558 complete(&req->done);
6560 spin_unlock_irq(&rq->lock);
6564 case CPU_DYING_FROZEN:
6565 /* Update our root-domain */
6567 spin_lock_irqsave(&rq->lock, flags);
6569 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6572 spin_unlock_irqrestore(&rq->lock, flags);
6579 /* Register at highest priority so that task migration (migrate_all_tasks)
6580 * happens before everything else.
6582 static struct notifier_block __cpuinitdata migration_notifier = {
6583 .notifier_call = migration_call,
6587 static int __init migration_init(void)
6589 void *cpu = (void *)(long)smp_processor_id();
6592 /* Start one for the boot CPU: */
6593 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6594 BUG_ON(err == NOTIFY_BAD);
6595 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6596 register_cpu_notifier(&migration_notifier);
6600 early_initcall(migration_init);
6605 #ifdef CONFIG_SCHED_DEBUG
6607 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6608 cpumask_t *groupmask)
6610 struct sched_group *group = sd->groups;
6613 cpulist_scnprintf(str, sizeof(str), sd->span);
6614 cpus_clear(*groupmask);
6616 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6618 if (!(sd->flags & SD_LOAD_BALANCE)) {
6619 printk("does not load-balance\n");
6621 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6626 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6628 if (!cpu_isset(cpu, sd->span)) {
6629 printk(KERN_ERR "ERROR: domain->span does not contain "
6632 if (!cpu_isset(cpu, group->cpumask)) {
6633 printk(KERN_ERR "ERROR: domain->groups does not contain"
6637 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6641 printk(KERN_ERR "ERROR: group is NULL\n");
6645 if (!group->__cpu_power) {
6646 printk(KERN_CONT "\n");
6647 printk(KERN_ERR "ERROR: domain->cpu_power not "
6652 if (!cpus_weight(group->cpumask)) {
6653 printk(KERN_CONT "\n");
6654 printk(KERN_ERR "ERROR: empty group\n");
6658 if (cpus_intersects(*groupmask, group->cpumask)) {
6659 printk(KERN_CONT "\n");
6660 printk(KERN_ERR "ERROR: repeated CPUs\n");
6664 cpus_or(*groupmask, *groupmask, group->cpumask);
6666 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6667 printk(KERN_CONT " %s", str);
6669 group = group->next;
6670 } while (group != sd->groups);
6671 printk(KERN_CONT "\n");
6673 if (!cpus_equal(sd->span, *groupmask))
6674 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6676 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6677 printk(KERN_ERR "ERROR: parent span is not a superset "
6678 "of domain->span\n");
6682 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6684 cpumask_t *groupmask;
6688 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6692 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6694 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6696 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6701 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6710 #else /* !CONFIG_SCHED_DEBUG */
6711 # define sched_domain_debug(sd, cpu) do { } while (0)
6712 #endif /* CONFIG_SCHED_DEBUG */
6714 static int sd_degenerate(struct sched_domain *sd)
6716 if (cpus_weight(sd->span) == 1)
6719 /* Following flags need at least 2 groups */
6720 if (sd->flags & (SD_LOAD_BALANCE |
6721 SD_BALANCE_NEWIDLE |
6725 SD_SHARE_PKG_RESOURCES)) {
6726 if (sd->groups != sd->groups->next)
6730 /* Following flags don't use groups */
6731 if (sd->flags & (SD_WAKE_IDLE |
6740 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6742 unsigned long cflags = sd->flags, pflags = parent->flags;
6744 if (sd_degenerate(parent))
6747 if (!cpus_equal(sd->span, parent->span))
6750 /* Does parent contain flags not in child? */
6751 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6752 if (cflags & SD_WAKE_AFFINE)
6753 pflags &= ~SD_WAKE_BALANCE;
6754 /* Flags needing groups don't count if only 1 group in parent */
6755 if (parent->groups == parent->groups->next) {
6756 pflags &= ~(SD_LOAD_BALANCE |
6757 SD_BALANCE_NEWIDLE |
6761 SD_SHARE_PKG_RESOURCES);
6763 if (~cflags & pflags)
6769 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6771 unsigned long flags;
6773 spin_lock_irqsave(&rq->lock, flags);
6776 struct root_domain *old_rd = rq->rd;
6778 if (cpu_isset(rq->cpu, old_rd->online))
6781 cpu_clear(rq->cpu, old_rd->span);
6783 if (atomic_dec_and_test(&old_rd->refcount))
6787 atomic_inc(&rd->refcount);
6790 cpu_set(rq->cpu, rd->span);
6791 if (cpu_isset(rq->cpu, cpu_online_map))
6794 spin_unlock_irqrestore(&rq->lock, flags);
6797 static void init_rootdomain(struct root_domain *rd)
6799 memset(rd, 0, sizeof(*rd));
6801 cpus_clear(rd->span);
6802 cpus_clear(rd->online);
6804 cpupri_init(&rd->cpupri);
6807 static void init_defrootdomain(void)
6809 init_rootdomain(&def_root_domain);
6810 atomic_set(&def_root_domain.refcount, 1);
6813 static struct root_domain *alloc_rootdomain(void)
6815 struct root_domain *rd;
6817 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6821 init_rootdomain(rd);
6827 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6828 * hold the hotplug lock.
6831 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6833 struct rq *rq = cpu_rq(cpu);
6834 struct sched_domain *tmp;
6836 /* Remove the sched domains which do not contribute to scheduling. */
6837 for (tmp = sd; tmp; ) {
6838 struct sched_domain *parent = tmp->parent;
6842 if (sd_parent_degenerate(tmp, parent)) {
6843 tmp->parent = parent->parent;
6845 parent->parent->child = tmp;
6850 if (sd && sd_degenerate(sd)) {
6856 sched_domain_debug(sd, cpu);
6858 rq_attach_root(rq, rd);
6859 rcu_assign_pointer(rq->sd, sd);
6862 /* cpus with isolated domains */
6863 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6865 /* Setup the mask of cpus configured for isolated domains */
6866 static int __init isolated_cpu_setup(char *str)
6868 static int __initdata ints[NR_CPUS];
6871 str = get_options(str, ARRAY_SIZE(ints), ints);
6872 cpus_clear(cpu_isolated_map);
6873 for (i = 1; i <= ints[0]; i++)
6874 if (ints[i] < NR_CPUS)
6875 cpu_set(ints[i], cpu_isolated_map);
6879 __setup("isolcpus=", isolated_cpu_setup);
6882 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6883 * to a function which identifies what group(along with sched group) a CPU
6884 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6885 * (due to the fact that we keep track of groups covered with a cpumask_t).
6887 * init_sched_build_groups will build a circular linked list of the groups
6888 * covered by the given span, and will set each group's ->cpumask correctly,
6889 * and ->cpu_power to 0.
6892 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6893 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6894 struct sched_group **sg,
6895 cpumask_t *tmpmask),
6896 cpumask_t *covered, cpumask_t *tmpmask)
6898 struct sched_group *first = NULL, *last = NULL;
6901 cpus_clear(*covered);
6903 for_each_cpu_mask_nr(i, *span) {
6904 struct sched_group *sg;
6905 int group = group_fn(i, cpu_map, &sg, tmpmask);
6908 if (cpu_isset(i, *covered))
6911 cpus_clear(sg->cpumask);
6912 sg->__cpu_power = 0;
6914 for_each_cpu_mask_nr(j, *span) {
6915 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6918 cpu_set(j, *covered);
6919 cpu_set(j, sg->cpumask);
6930 #define SD_NODES_PER_DOMAIN 16
6935 * find_next_best_node - find the next node to include in a sched_domain
6936 * @node: node whose sched_domain we're building
6937 * @used_nodes: nodes already in the sched_domain
6939 * Find the next node to include in a given scheduling domain. Simply
6940 * finds the closest node not already in the @used_nodes map.
6942 * Should use nodemask_t.
6944 static int find_next_best_node(int node, nodemask_t *used_nodes)
6946 int i, n, val, min_val, best_node = 0;
6950 for (i = 0; i < nr_node_ids; i++) {
6951 /* Start at @node */
6952 n = (node + i) % nr_node_ids;
6954 if (!nr_cpus_node(n))
6957 /* Skip already used nodes */
6958 if (node_isset(n, *used_nodes))
6961 /* Simple min distance search */
6962 val = node_distance(node, n);
6964 if (val < min_val) {
6970 node_set(best_node, *used_nodes);
6975 * sched_domain_node_span - get a cpumask for a node's sched_domain
6976 * @node: node whose cpumask we're constructing
6977 * @span: resulting cpumask
6979 * Given a node, construct a good cpumask for its sched_domain to span. It
6980 * should be one that prevents unnecessary balancing, but also spreads tasks
6983 static void sched_domain_node_span(int node, cpumask_t *span)
6985 nodemask_t used_nodes;
6986 node_to_cpumask_ptr(nodemask, node);
6990 nodes_clear(used_nodes);
6992 cpus_or(*span, *span, *nodemask);
6993 node_set(node, used_nodes);
6995 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6996 int next_node = find_next_best_node(node, &used_nodes);
6998 node_to_cpumask_ptr_next(nodemask, next_node);
6999 cpus_or(*span, *span, *nodemask);
7002 #endif /* CONFIG_NUMA */
7004 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7007 * SMT sched-domains:
7009 #ifdef CONFIG_SCHED_SMT
7010 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
7011 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
7014 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7018 *sg = &per_cpu(sched_group_cpus, cpu);
7021 #endif /* CONFIG_SCHED_SMT */
7024 * multi-core sched-domains:
7026 #ifdef CONFIG_SCHED_MC
7027 static DEFINE_PER_CPU(struct sched_domain, core_domains);
7028 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
7029 #endif /* CONFIG_SCHED_MC */
7031 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7033 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7038 *mask = per_cpu(cpu_sibling_map, cpu);
7039 cpus_and(*mask, *mask, *cpu_map);
7040 group = first_cpu(*mask);
7042 *sg = &per_cpu(sched_group_core, group);
7045 #elif defined(CONFIG_SCHED_MC)
7047 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7051 *sg = &per_cpu(sched_group_core, cpu);
7056 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
7057 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
7060 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7064 #ifdef CONFIG_SCHED_MC
7065 *mask = cpu_coregroup_map(cpu);
7066 cpus_and(*mask, *mask, *cpu_map);
7067 group = first_cpu(*mask);
7068 #elif defined(CONFIG_SCHED_SMT)
7069 *mask = per_cpu(cpu_sibling_map, cpu);
7070 cpus_and(*mask, *mask, *cpu_map);
7071 group = first_cpu(*mask);
7076 *sg = &per_cpu(sched_group_phys, group);
7082 * The init_sched_build_groups can't handle what we want to do with node
7083 * groups, so roll our own. Now each node has its own list of groups which
7084 * gets dynamically allocated.
7086 static DEFINE_PER_CPU(struct sched_domain, node_domains);
7087 static struct sched_group ***sched_group_nodes_bycpu;
7089 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
7090 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
7092 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
7093 struct sched_group **sg, cpumask_t *nodemask)
7097 *nodemask = node_to_cpumask(cpu_to_node(cpu));
7098 cpus_and(*nodemask, *nodemask, *cpu_map);
7099 group = first_cpu(*nodemask);
7102 *sg = &per_cpu(sched_group_allnodes, group);
7106 static void init_numa_sched_groups_power(struct sched_group *group_head)
7108 struct sched_group *sg = group_head;
7114 for_each_cpu_mask_nr(j, sg->cpumask) {
7115 struct sched_domain *sd;
7117 sd = &per_cpu(phys_domains, j);
7118 if (j != first_cpu(sd->groups->cpumask)) {
7120 * Only add "power" once for each
7126 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7129 } while (sg != group_head);
7131 #endif /* CONFIG_NUMA */
7134 /* Free memory allocated for various sched_group structures */
7135 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7139 for_each_cpu_mask_nr(cpu, *cpu_map) {
7140 struct sched_group **sched_group_nodes
7141 = sched_group_nodes_bycpu[cpu];
7143 if (!sched_group_nodes)
7146 for (i = 0; i < nr_node_ids; i++) {
7147 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7149 *nodemask = node_to_cpumask(i);
7150 cpus_and(*nodemask, *nodemask, *cpu_map);
7151 if (cpus_empty(*nodemask))
7161 if (oldsg != sched_group_nodes[i])
7164 kfree(sched_group_nodes);
7165 sched_group_nodes_bycpu[cpu] = NULL;
7168 #else /* !CONFIG_NUMA */
7169 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7172 #endif /* CONFIG_NUMA */
7175 * Initialize sched groups cpu_power.
7177 * cpu_power indicates the capacity of sched group, which is used while
7178 * distributing the load between different sched groups in a sched domain.
7179 * Typically cpu_power for all the groups in a sched domain will be same unless
7180 * there are asymmetries in the topology. If there are asymmetries, group
7181 * having more cpu_power will pickup more load compared to the group having
7184 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7185 * the maximum number of tasks a group can handle in the presence of other idle
7186 * or lightly loaded groups in the same sched domain.
7188 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7190 struct sched_domain *child;
7191 struct sched_group *group;
7193 WARN_ON(!sd || !sd->groups);
7195 if (cpu != first_cpu(sd->groups->cpumask))
7200 sd->groups->__cpu_power = 0;
7203 * For perf policy, if the groups in child domain share resources
7204 * (for example cores sharing some portions of the cache hierarchy
7205 * or SMT), then set this domain groups cpu_power such that each group
7206 * can handle only one task, when there are other idle groups in the
7207 * same sched domain.
7209 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7211 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7212 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7217 * add cpu_power of each child group to this groups cpu_power
7219 group = child->groups;
7221 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7222 group = group->next;
7223 } while (group != child->groups);
7227 * Initializers for schedule domains
7228 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7231 #ifdef CONFIG_SCHED_DEBUG
7232 # define SD_INIT_NAME(sd, type) sd->name = #type
7234 # define SD_INIT_NAME(sd, type) do { } while (0)
7237 #define SD_INIT(sd, type) sd_init_##type(sd)
7239 #define SD_INIT_FUNC(type) \
7240 static noinline void sd_init_##type(struct sched_domain *sd) \
7242 memset(sd, 0, sizeof(*sd)); \
7243 *sd = SD_##type##_INIT; \
7244 sd->level = SD_LV_##type; \
7245 SD_INIT_NAME(sd, type); \
7250 SD_INIT_FUNC(ALLNODES)
7253 #ifdef CONFIG_SCHED_SMT
7254 SD_INIT_FUNC(SIBLING)
7256 #ifdef CONFIG_SCHED_MC
7261 * To minimize stack usage kmalloc room for cpumasks and share the
7262 * space as the usage in build_sched_domains() dictates. Used only
7263 * if the amount of space is significant.
7266 cpumask_t tmpmask; /* make this one first */
7269 cpumask_t this_sibling_map;
7270 cpumask_t this_core_map;
7272 cpumask_t send_covered;
7275 cpumask_t domainspan;
7277 cpumask_t notcovered;
7282 #define SCHED_CPUMASK_ALLOC 1
7283 #define SCHED_CPUMASK_FREE(v) kfree(v)
7284 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7286 #define SCHED_CPUMASK_ALLOC 0
7287 #define SCHED_CPUMASK_FREE(v)
7288 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7291 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7292 ((unsigned long)(a) + offsetof(struct allmasks, v))
7294 static int default_relax_domain_level = -1;
7296 static int __init setup_relax_domain_level(char *str)
7300 val = simple_strtoul(str, NULL, 0);
7301 if (val < SD_LV_MAX)
7302 default_relax_domain_level = val;
7306 __setup("relax_domain_level=", setup_relax_domain_level);
7308 static void set_domain_attribute(struct sched_domain *sd,
7309 struct sched_domain_attr *attr)
7313 if (!attr || attr->relax_domain_level < 0) {
7314 if (default_relax_domain_level < 0)
7317 request = default_relax_domain_level;
7319 request = attr->relax_domain_level;
7320 if (request < sd->level) {
7321 /* turn off idle balance on this domain */
7322 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7324 /* turn on idle balance on this domain */
7325 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7330 * Build sched domains for a given set of cpus and attach the sched domains
7331 * to the individual cpus
7333 static int __build_sched_domains(const cpumask_t *cpu_map,
7334 struct sched_domain_attr *attr)
7337 struct root_domain *rd;
7338 SCHED_CPUMASK_DECLARE(allmasks);
7341 struct sched_group **sched_group_nodes = NULL;
7342 int sd_allnodes = 0;
7345 * Allocate the per-node list of sched groups
7347 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
7349 if (!sched_group_nodes) {
7350 printk(KERN_WARNING "Can not alloc sched group node list\n");
7355 rd = alloc_rootdomain();
7357 printk(KERN_WARNING "Cannot alloc root domain\n");
7359 kfree(sched_group_nodes);
7364 #if SCHED_CPUMASK_ALLOC
7365 /* get space for all scratch cpumask variables */
7366 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
7368 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7371 kfree(sched_group_nodes);
7376 tmpmask = (cpumask_t *)allmasks;
7380 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7384 * Set up domains for cpus specified by the cpu_map.
7386 for_each_cpu_mask_nr(i, *cpu_map) {
7387 struct sched_domain *sd = NULL, *p;
7388 SCHED_CPUMASK_VAR(nodemask, allmasks);
7390 *nodemask = node_to_cpumask(cpu_to_node(i));
7391 cpus_and(*nodemask, *nodemask, *cpu_map);
7394 if (cpus_weight(*cpu_map) >
7395 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7396 sd = &per_cpu(allnodes_domains, i);
7397 SD_INIT(sd, ALLNODES);
7398 set_domain_attribute(sd, attr);
7399 sd->span = *cpu_map;
7400 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7406 sd = &per_cpu(node_domains, i);
7408 set_domain_attribute(sd, attr);
7409 sched_domain_node_span(cpu_to_node(i), &sd->span);
7413 cpus_and(sd->span, sd->span, *cpu_map);
7417 sd = &per_cpu(phys_domains, i);
7419 set_domain_attribute(sd, attr);
7420 sd->span = *nodemask;
7424 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7426 #ifdef CONFIG_SCHED_MC
7428 sd = &per_cpu(core_domains, i);
7430 set_domain_attribute(sd, attr);
7431 sd->span = cpu_coregroup_map(i);
7432 cpus_and(sd->span, sd->span, *cpu_map);
7435 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7438 #ifdef CONFIG_SCHED_SMT
7440 sd = &per_cpu(cpu_domains, i);
7441 SD_INIT(sd, SIBLING);
7442 set_domain_attribute(sd, attr);
7443 sd->span = per_cpu(cpu_sibling_map, i);
7444 cpus_and(sd->span, sd->span, *cpu_map);
7447 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7451 #ifdef CONFIG_SCHED_SMT
7452 /* Set up CPU (sibling) groups */
7453 for_each_cpu_mask_nr(i, *cpu_map) {
7454 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7455 SCHED_CPUMASK_VAR(send_covered, allmasks);
7457 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7458 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7459 if (i != first_cpu(*this_sibling_map))
7462 init_sched_build_groups(this_sibling_map, cpu_map,
7464 send_covered, tmpmask);
7468 #ifdef CONFIG_SCHED_MC
7469 /* Set up multi-core groups */
7470 for_each_cpu_mask_nr(i, *cpu_map) {
7471 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7472 SCHED_CPUMASK_VAR(send_covered, allmasks);
7474 *this_core_map = cpu_coregroup_map(i);
7475 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7476 if (i != first_cpu(*this_core_map))
7479 init_sched_build_groups(this_core_map, cpu_map,
7481 send_covered, tmpmask);
7485 /* Set up physical groups */
7486 for (i = 0; i < nr_node_ids; i++) {
7487 SCHED_CPUMASK_VAR(nodemask, allmasks);
7488 SCHED_CPUMASK_VAR(send_covered, allmasks);
7490 *nodemask = node_to_cpumask(i);
7491 cpus_and(*nodemask, *nodemask, *cpu_map);
7492 if (cpus_empty(*nodemask))
7495 init_sched_build_groups(nodemask, cpu_map,
7497 send_covered, tmpmask);
7501 /* Set up node groups */
7503 SCHED_CPUMASK_VAR(send_covered, allmasks);
7505 init_sched_build_groups(cpu_map, cpu_map,
7506 &cpu_to_allnodes_group,
7507 send_covered, tmpmask);
7510 for (i = 0; i < nr_node_ids; i++) {
7511 /* Set up node groups */
7512 struct sched_group *sg, *prev;
7513 SCHED_CPUMASK_VAR(nodemask, allmasks);
7514 SCHED_CPUMASK_VAR(domainspan, allmasks);
7515 SCHED_CPUMASK_VAR(covered, allmasks);
7518 *nodemask = node_to_cpumask(i);
7519 cpus_clear(*covered);
7521 cpus_and(*nodemask, *nodemask, *cpu_map);
7522 if (cpus_empty(*nodemask)) {
7523 sched_group_nodes[i] = NULL;
7527 sched_domain_node_span(i, domainspan);
7528 cpus_and(*domainspan, *domainspan, *cpu_map);
7530 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7532 printk(KERN_WARNING "Can not alloc domain group for "
7536 sched_group_nodes[i] = sg;
7537 for_each_cpu_mask_nr(j, *nodemask) {
7538 struct sched_domain *sd;
7540 sd = &per_cpu(node_domains, j);
7543 sg->__cpu_power = 0;
7544 sg->cpumask = *nodemask;
7546 cpus_or(*covered, *covered, *nodemask);
7549 for (j = 0; j < nr_node_ids; j++) {
7550 SCHED_CPUMASK_VAR(notcovered, allmasks);
7551 int n = (i + j) % nr_node_ids;
7552 node_to_cpumask_ptr(pnodemask, n);
7554 cpus_complement(*notcovered, *covered);
7555 cpus_and(*tmpmask, *notcovered, *cpu_map);
7556 cpus_and(*tmpmask, *tmpmask, *domainspan);
7557 if (cpus_empty(*tmpmask))
7560 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7561 if (cpus_empty(*tmpmask))
7564 sg = kmalloc_node(sizeof(struct sched_group),
7568 "Can not alloc domain group for node %d\n", j);
7571 sg->__cpu_power = 0;
7572 sg->cpumask = *tmpmask;
7573 sg->next = prev->next;
7574 cpus_or(*covered, *covered, *tmpmask);
7581 /* Calculate CPU power for physical packages and nodes */
7582 #ifdef CONFIG_SCHED_SMT
7583 for_each_cpu_mask_nr(i, *cpu_map) {
7584 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7586 init_sched_groups_power(i, sd);
7589 #ifdef CONFIG_SCHED_MC
7590 for_each_cpu_mask_nr(i, *cpu_map) {
7591 struct sched_domain *sd = &per_cpu(core_domains, i);
7593 init_sched_groups_power(i, sd);
7597 for_each_cpu_mask_nr(i, *cpu_map) {
7598 struct sched_domain *sd = &per_cpu(phys_domains, i);
7600 init_sched_groups_power(i, sd);
7604 for (i = 0; i < nr_node_ids; i++)
7605 init_numa_sched_groups_power(sched_group_nodes[i]);
7608 struct sched_group *sg;
7610 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7612 init_numa_sched_groups_power(sg);
7616 /* Attach the domains */
7617 for_each_cpu_mask_nr(i, *cpu_map) {
7618 struct sched_domain *sd;
7619 #ifdef CONFIG_SCHED_SMT
7620 sd = &per_cpu(cpu_domains, i);
7621 #elif defined(CONFIG_SCHED_MC)
7622 sd = &per_cpu(core_domains, i);
7624 sd = &per_cpu(phys_domains, i);
7626 cpu_attach_domain(sd, rd, i);
7629 SCHED_CPUMASK_FREE((void *)allmasks);
7634 free_sched_groups(cpu_map, tmpmask);
7635 SCHED_CPUMASK_FREE((void *)allmasks);
7641 static int build_sched_domains(const cpumask_t *cpu_map)
7643 return __build_sched_domains(cpu_map, NULL);
7646 static cpumask_t *doms_cur; /* current sched domains */
7647 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7648 static struct sched_domain_attr *dattr_cur;
7649 /* attribues of custom domains in 'doms_cur' */
7652 * Special case: If a kmalloc of a doms_cur partition (array of
7653 * cpumask_t) fails, then fallback to a single sched domain,
7654 * as determined by the single cpumask_t fallback_doms.
7656 static cpumask_t fallback_doms;
7658 void __attribute__((weak)) arch_update_cpu_topology(void)
7663 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7664 * For now this just excludes isolated cpus, but could be used to
7665 * exclude other special cases in the future.
7667 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7671 arch_update_cpu_topology();
7673 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7675 doms_cur = &fallback_doms;
7676 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7678 err = build_sched_domains(doms_cur);
7679 register_sched_domain_sysctl();
7684 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7687 free_sched_groups(cpu_map, tmpmask);
7691 * Detach sched domains from a group of cpus specified in cpu_map
7692 * These cpus will now be attached to the NULL domain
7694 static void detach_destroy_domains(const cpumask_t *cpu_map)
7699 for_each_cpu_mask_nr(i, *cpu_map)
7700 cpu_attach_domain(NULL, &def_root_domain, i);
7701 synchronize_sched();
7702 arch_destroy_sched_domains(cpu_map, &tmpmask);
7705 /* handle null as "default" */
7706 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7707 struct sched_domain_attr *new, int idx_new)
7709 struct sched_domain_attr tmp;
7716 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7717 new ? (new + idx_new) : &tmp,
7718 sizeof(struct sched_domain_attr));
7722 * Partition sched domains as specified by the 'ndoms_new'
7723 * cpumasks in the array doms_new[] of cpumasks. This compares
7724 * doms_new[] to the current sched domain partitioning, doms_cur[].
7725 * It destroys each deleted domain and builds each new domain.
7727 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7728 * The masks don't intersect (don't overlap.) We should setup one
7729 * sched domain for each mask. CPUs not in any of the cpumasks will
7730 * not be load balanced. If the same cpumask appears both in the
7731 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7734 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7735 * ownership of it and will kfree it when done with it. If the caller
7736 * failed the kmalloc call, then it can pass in doms_new == NULL,
7737 * and partition_sched_domains() will fallback to the single partition
7738 * 'fallback_doms', it also forces the domains to be rebuilt.
7740 * If doms_new==NULL it will be replaced with cpu_online_map.
7741 * ndoms_new==0 is a special case for destroying existing domains.
7742 * It will not create the default domain.
7744 * Call with hotplug lock held
7746 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7747 struct sched_domain_attr *dattr_new)
7751 mutex_lock(&sched_domains_mutex);
7753 /* always unregister in case we don't destroy any domains */
7754 unregister_sched_domain_sysctl();
7756 n = doms_new ? ndoms_new : 0;
7758 /* Destroy deleted domains */
7759 for (i = 0; i < ndoms_cur; i++) {
7760 for (j = 0; j < n; j++) {
7761 if (cpus_equal(doms_cur[i], doms_new[j])
7762 && dattrs_equal(dattr_cur, i, dattr_new, j))
7765 /* no match - a current sched domain not in new doms_new[] */
7766 detach_destroy_domains(doms_cur + i);
7771 if (doms_new == NULL) {
7773 doms_new = &fallback_doms;
7774 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7775 WARN_ON_ONCE(dattr_new);
7778 /* Build new domains */
7779 for (i = 0; i < ndoms_new; i++) {
7780 for (j = 0; j < ndoms_cur; j++) {
7781 if (cpus_equal(doms_new[i], doms_cur[j])
7782 && dattrs_equal(dattr_new, i, dattr_cur, j))
7785 /* no match - add a new doms_new */
7786 __build_sched_domains(doms_new + i,
7787 dattr_new ? dattr_new + i : NULL);
7792 /* Remember the new sched domains */
7793 if (doms_cur != &fallback_doms)
7795 kfree(dattr_cur); /* kfree(NULL) is safe */
7796 doms_cur = doms_new;
7797 dattr_cur = dattr_new;
7798 ndoms_cur = ndoms_new;
7800 register_sched_domain_sysctl();
7802 mutex_unlock(&sched_domains_mutex);
7805 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7806 int arch_reinit_sched_domains(void)
7810 /* Destroy domains first to force the rebuild */
7811 partition_sched_domains(0, NULL, NULL);
7813 rebuild_sched_domains();
7819 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7823 if (buf[0] != '0' && buf[0] != '1')
7827 sched_smt_power_savings = (buf[0] == '1');
7829 sched_mc_power_savings = (buf[0] == '1');
7831 ret = arch_reinit_sched_domains();
7833 return ret ? ret : count;
7836 #ifdef CONFIG_SCHED_MC
7837 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7840 return sprintf(page, "%u\n", sched_mc_power_savings);
7842 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7843 const char *buf, size_t count)
7845 return sched_power_savings_store(buf, count, 0);
7847 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7848 sched_mc_power_savings_show,
7849 sched_mc_power_savings_store);
7852 #ifdef CONFIG_SCHED_SMT
7853 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7856 return sprintf(page, "%u\n", sched_smt_power_savings);
7858 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7859 const char *buf, size_t count)
7861 return sched_power_savings_store(buf, count, 1);
7863 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7864 sched_smt_power_savings_show,
7865 sched_smt_power_savings_store);
7868 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7872 #ifdef CONFIG_SCHED_SMT
7874 err = sysfs_create_file(&cls->kset.kobj,
7875 &attr_sched_smt_power_savings.attr);
7877 #ifdef CONFIG_SCHED_MC
7878 if (!err && mc_capable())
7879 err = sysfs_create_file(&cls->kset.kobj,
7880 &attr_sched_mc_power_savings.attr);
7884 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7886 #ifndef CONFIG_CPUSETS
7888 * Add online and remove offline CPUs from the scheduler domains.
7889 * When cpusets are enabled they take over this function.
7891 static int update_sched_domains(struct notifier_block *nfb,
7892 unsigned long action, void *hcpu)
7896 case CPU_ONLINE_FROZEN:
7898 case CPU_DEAD_FROZEN:
7899 partition_sched_domains(1, NULL, NULL);
7908 static int update_runtime(struct notifier_block *nfb,
7909 unsigned long action, void *hcpu)
7911 int cpu = (int)(long)hcpu;
7914 case CPU_DOWN_PREPARE:
7915 case CPU_DOWN_PREPARE_FROZEN:
7916 disable_runtime(cpu_rq(cpu));
7919 case CPU_DOWN_FAILED:
7920 case CPU_DOWN_FAILED_FROZEN:
7922 case CPU_ONLINE_FROZEN:
7923 enable_runtime(cpu_rq(cpu));
7931 void __init sched_init_smp(void)
7933 cpumask_t non_isolated_cpus;
7935 #if defined(CONFIG_NUMA)
7936 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7938 BUG_ON(sched_group_nodes_bycpu == NULL);
7941 mutex_lock(&sched_domains_mutex);
7942 arch_init_sched_domains(&cpu_online_map);
7943 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7944 if (cpus_empty(non_isolated_cpus))
7945 cpu_set(smp_processor_id(), non_isolated_cpus);
7946 mutex_unlock(&sched_domains_mutex);
7949 #ifndef CONFIG_CPUSETS
7950 /* XXX: Theoretical race here - CPU may be hotplugged now */
7951 hotcpu_notifier(update_sched_domains, 0);
7954 /* RT runtime code needs to handle some hotplug events */
7955 hotcpu_notifier(update_runtime, 0);
7959 /* Move init over to a non-isolated CPU */
7960 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7962 sched_init_granularity();
7965 void __init sched_init_smp(void)
7967 sched_init_granularity();
7969 #endif /* CONFIG_SMP */
7971 int in_sched_functions(unsigned long addr)
7973 return in_lock_functions(addr) ||
7974 (addr >= (unsigned long)__sched_text_start
7975 && addr < (unsigned long)__sched_text_end);
7978 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7980 cfs_rq->tasks_timeline = RB_ROOT;
7981 INIT_LIST_HEAD(&cfs_rq->tasks);
7982 #ifdef CONFIG_FAIR_GROUP_SCHED
7985 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7988 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7990 struct rt_prio_array *array;
7993 array = &rt_rq->active;
7994 for (i = 0; i < MAX_RT_PRIO; i++) {
7995 INIT_LIST_HEAD(array->queue + i);
7996 __clear_bit(i, array->bitmap);
7998 /* delimiter for bitsearch: */
7999 __set_bit(MAX_RT_PRIO, array->bitmap);
8001 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8002 rt_rq->highest_prio = MAX_RT_PRIO;
8005 rt_rq->rt_nr_migratory = 0;
8006 rt_rq->overloaded = 0;
8010 rt_rq->rt_throttled = 0;
8011 rt_rq->rt_runtime = 0;
8012 spin_lock_init(&rt_rq->rt_runtime_lock);
8014 #ifdef CONFIG_RT_GROUP_SCHED
8015 rt_rq->rt_nr_boosted = 0;
8020 #ifdef CONFIG_FAIR_GROUP_SCHED
8021 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8022 struct sched_entity *se, int cpu, int add,
8023 struct sched_entity *parent)
8025 struct rq *rq = cpu_rq(cpu);
8026 tg->cfs_rq[cpu] = cfs_rq;
8027 init_cfs_rq(cfs_rq, rq);
8030 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8033 /* se could be NULL for init_task_group */
8038 se->cfs_rq = &rq->cfs;
8040 se->cfs_rq = parent->my_q;
8043 se->load.weight = tg->shares;
8044 se->load.inv_weight = 0;
8045 se->parent = parent;
8049 #ifdef CONFIG_RT_GROUP_SCHED
8050 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8051 struct sched_rt_entity *rt_se, int cpu, int add,
8052 struct sched_rt_entity *parent)
8054 struct rq *rq = cpu_rq(cpu);
8056 tg->rt_rq[cpu] = rt_rq;
8057 init_rt_rq(rt_rq, rq);
8059 rt_rq->rt_se = rt_se;
8060 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8062 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8064 tg->rt_se[cpu] = rt_se;
8069 rt_se->rt_rq = &rq->rt;
8071 rt_se->rt_rq = parent->my_q;
8073 rt_se->my_q = rt_rq;
8074 rt_se->parent = parent;
8075 INIT_LIST_HEAD(&rt_se->run_list);
8079 void __init sched_init(void)
8082 unsigned long alloc_size = 0, ptr;
8084 #ifdef CONFIG_FAIR_GROUP_SCHED
8085 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8087 #ifdef CONFIG_RT_GROUP_SCHED
8088 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8090 #ifdef CONFIG_USER_SCHED
8094 * As sched_init() is called before page_alloc is setup,
8095 * we use alloc_bootmem().
8098 ptr = (unsigned long)alloc_bootmem(alloc_size);
8100 #ifdef CONFIG_FAIR_GROUP_SCHED
8101 init_task_group.se = (struct sched_entity **)ptr;
8102 ptr += nr_cpu_ids * sizeof(void **);
8104 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8105 ptr += nr_cpu_ids * sizeof(void **);
8107 #ifdef CONFIG_USER_SCHED
8108 root_task_group.se = (struct sched_entity **)ptr;
8109 ptr += nr_cpu_ids * sizeof(void **);
8111 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8112 ptr += nr_cpu_ids * sizeof(void **);
8113 #endif /* CONFIG_USER_SCHED */
8114 #endif /* CONFIG_FAIR_GROUP_SCHED */
8115 #ifdef CONFIG_RT_GROUP_SCHED
8116 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8117 ptr += nr_cpu_ids * sizeof(void **);
8119 init_task_group.rt_rq = (struct rt_rq **)ptr;
8120 ptr += nr_cpu_ids * sizeof(void **);
8122 #ifdef CONFIG_USER_SCHED
8123 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8124 ptr += nr_cpu_ids * sizeof(void **);
8126 root_task_group.rt_rq = (struct rt_rq **)ptr;
8127 ptr += nr_cpu_ids * sizeof(void **);
8128 #endif /* CONFIG_USER_SCHED */
8129 #endif /* CONFIG_RT_GROUP_SCHED */
8133 init_defrootdomain();
8136 init_rt_bandwidth(&def_rt_bandwidth,
8137 global_rt_period(), global_rt_runtime());
8139 #ifdef CONFIG_RT_GROUP_SCHED
8140 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8141 global_rt_period(), global_rt_runtime());
8142 #ifdef CONFIG_USER_SCHED
8143 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8144 global_rt_period(), RUNTIME_INF);
8145 #endif /* CONFIG_USER_SCHED */
8146 #endif /* CONFIG_RT_GROUP_SCHED */
8148 #ifdef CONFIG_GROUP_SCHED
8149 list_add(&init_task_group.list, &task_groups);
8150 INIT_LIST_HEAD(&init_task_group.children);
8152 #ifdef CONFIG_USER_SCHED
8153 INIT_LIST_HEAD(&root_task_group.children);
8154 init_task_group.parent = &root_task_group;
8155 list_add(&init_task_group.siblings, &root_task_group.children);
8156 #endif /* CONFIG_USER_SCHED */
8157 #endif /* CONFIG_GROUP_SCHED */
8159 for_each_possible_cpu(i) {
8163 spin_lock_init(&rq->lock);
8165 init_cfs_rq(&rq->cfs, rq);
8166 init_rt_rq(&rq->rt, rq);
8167 #ifdef CONFIG_FAIR_GROUP_SCHED
8168 init_task_group.shares = init_task_group_load;
8169 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8170 #ifdef CONFIG_CGROUP_SCHED
8172 * How much cpu bandwidth does init_task_group get?
8174 * In case of task-groups formed thr' the cgroup filesystem, it
8175 * gets 100% of the cpu resources in the system. This overall
8176 * system cpu resource is divided among the tasks of
8177 * init_task_group and its child task-groups in a fair manner,
8178 * based on each entity's (task or task-group's) weight
8179 * (se->load.weight).
8181 * In other words, if init_task_group has 10 tasks of weight
8182 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8183 * then A0's share of the cpu resource is:
8185 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8187 * We achieve this by letting init_task_group's tasks sit
8188 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8190 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8191 #elif defined CONFIG_USER_SCHED
8192 root_task_group.shares = NICE_0_LOAD;
8193 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8195 * In case of task-groups formed thr' the user id of tasks,
8196 * init_task_group represents tasks belonging to root user.
8197 * Hence it forms a sibling of all subsequent groups formed.
8198 * In this case, init_task_group gets only a fraction of overall
8199 * system cpu resource, based on the weight assigned to root
8200 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8201 * by letting tasks of init_task_group sit in a separate cfs_rq
8202 * (init_cfs_rq) and having one entity represent this group of
8203 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8205 init_tg_cfs_entry(&init_task_group,
8206 &per_cpu(init_cfs_rq, i),
8207 &per_cpu(init_sched_entity, i), i, 1,
8208 root_task_group.se[i]);
8211 #endif /* CONFIG_FAIR_GROUP_SCHED */
8213 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8214 #ifdef CONFIG_RT_GROUP_SCHED
8215 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8216 #ifdef CONFIG_CGROUP_SCHED
8217 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8218 #elif defined CONFIG_USER_SCHED
8219 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8220 init_tg_rt_entry(&init_task_group,
8221 &per_cpu(init_rt_rq, i),
8222 &per_cpu(init_sched_rt_entity, i), i, 1,
8223 root_task_group.rt_se[i]);
8227 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8228 rq->cpu_load[j] = 0;
8232 rq->active_balance = 0;
8233 rq->next_balance = jiffies;
8237 rq->migration_thread = NULL;
8238 INIT_LIST_HEAD(&rq->migration_queue);
8239 rq_attach_root(rq, &def_root_domain);
8242 atomic_set(&rq->nr_iowait, 0);
8245 set_load_weight(&init_task);
8247 #ifdef CONFIG_PREEMPT_NOTIFIERS
8248 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8252 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8255 #ifdef CONFIG_RT_MUTEXES
8256 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8260 * The boot idle thread does lazy MMU switching as well:
8262 atomic_inc(&init_mm.mm_count);
8263 enter_lazy_tlb(&init_mm, current);
8266 * Make us the idle thread. Technically, schedule() should not be
8267 * called from this thread, however somewhere below it might be,
8268 * but because we are the idle thread, we just pick up running again
8269 * when this runqueue becomes "idle".
8271 init_idle(current, smp_processor_id());
8273 * During early bootup we pretend to be a normal task:
8275 current->sched_class = &fair_sched_class;
8277 scheduler_running = 1;
8280 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8281 void __might_sleep(char *file, int line)
8284 static unsigned long prev_jiffy; /* ratelimiting */
8286 if ((!in_atomic() && !irqs_disabled()) ||
8287 system_state != SYSTEM_RUNNING || oops_in_progress)
8289 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8291 prev_jiffy = jiffies;
8294 "BUG: sleeping function called from invalid context at %s:%d\n",
8297 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8298 in_atomic(), irqs_disabled(),
8299 current->pid, current->comm);
8301 debug_show_held_locks(current);
8302 if (irqs_disabled())
8303 print_irqtrace_events(current);
8307 EXPORT_SYMBOL(__might_sleep);
8310 #ifdef CONFIG_MAGIC_SYSRQ
8311 static void normalize_task(struct rq *rq, struct task_struct *p)
8315 update_rq_clock(rq);
8316 on_rq = p->se.on_rq;
8318 deactivate_task(rq, p, 0);
8319 __setscheduler(rq, p, SCHED_NORMAL, 0);
8321 activate_task(rq, p, 0);
8322 resched_task(rq->curr);
8326 void normalize_rt_tasks(void)
8328 struct task_struct *g, *p;
8329 unsigned long flags;
8332 read_lock_irqsave(&tasklist_lock, flags);
8333 do_each_thread(g, p) {
8335 * Only normalize user tasks:
8340 p->se.exec_start = 0;
8341 #ifdef CONFIG_SCHEDSTATS
8342 p->se.wait_start = 0;
8343 p->se.sleep_start = 0;
8344 p->se.block_start = 0;
8349 * Renice negative nice level userspace
8352 if (TASK_NICE(p) < 0 && p->mm)
8353 set_user_nice(p, 0);
8357 spin_lock(&p->pi_lock);
8358 rq = __task_rq_lock(p);
8360 normalize_task(rq, p);
8362 __task_rq_unlock(rq);
8363 spin_unlock(&p->pi_lock);
8364 } while_each_thread(g, p);
8366 read_unlock_irqrestore(&tasklist_lock, flags);
8369 #endif /* CONFIG_MAGIC_SYSRQ */
8373 * These functions are only useful for the IA64 MCA handling.
8375 * They can only be called when the whole system has been
8376 * stopped - every CPU needs to be quiescent, and no scheduling
8377 * activity can take place. Using them for anything else would
8378 * be a serious bug, and as a result, they aren't even visible
8379 * under any other configuration.
8383 * curr_task - return the current task for a given cpu.
8384 * @cpu: the processor in question.
8386 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8388 struct task_struct *curr_task(int cpu)
8390 return cpu_curr(cpu);
8394 * set_curr_task - set the current task for a given cpu.
8395 * @cpu: the processor in question.
8396 * @p: the task pointer to set.
8398 * Description: This function must only be used when non-maskable interrupts
8399 * are serviced on a separate stack. It allows the architecture to switch the
8400 * notion of the current task on a cpu in a non-blocking manner. This function
8401 * must be called with all CPU's synchronized, and interrupts disabled, the
8402 * and caller must save the original value of the current task (see
8403 * curr_task() above) and restore that value before reenabling interrupts and
8404 * re-starting the system.
8406 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8408 void set_curr_task(int cpu, struct task_struct *p)
8415 #ifdef CONFIG_FAIR_GROUP_SCHED
8416 static void free_fair_sched_group(struct task_group *tg)
8420 for_each_possible_cpu(i) {
8422 kfree(tg->cfs_rq[i]);
8432 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8434 struct cfs_rq *cfs_rq;
8435 struct sched_entity *se;
8439 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8442 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8446 tg->shares = NICE_0_LOAD;
8448 for_each_possible_cpu(i) {
8451 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8452 GFP_KERNEL, cpu_to_node(i));
8456 se = kzalloc_node(sizeof(struct sched_entity),
8457 GFP_KERNEL, cpu_to_node(i));
8461 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8470 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8472 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8473 &cpu_rq(cpu)->leaf_cfs_rq_list);
8476 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8478 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8480 #else /* !CONFG_FAIR_GROUP_SCHED */
8481 static inline void free_fair_sched_group(struct task_group *tg)
8486 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8491 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8495 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8498 #endif /* CONFIG_FAIR_GROUP_SCHED */
8500 #ifdef CONFIG_RT_GROUP_SCHED
8501 static void free_rt_sched_group(struct task_group *tg)
8505 destroy_rt_bandwidth(&tg->rt_bandwidth);
8507 for_each_possible_cpu(i) {
8509 kfree(tg->rt_rq[i]);
8511 kfree(tg->rt_se[i]);
8519 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8521 struct rt_rq *rt_rq;
8522 struct sched_rt_entity *rt_se;
8526 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8529 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8533 init_rt_bandwidth(&tg->rt_bandwidth,
8534 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8536 for_each_possible_cpu(i) {
8539 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8540 GFP_KERNEL, cpu_to_node(i));
8544 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8545 GFP_KERNEL, cpu_to_node(i));
8549 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8558 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8560 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8561 &cpu_rq(cpu)->leaf_rt_rq_list);
8564 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8566 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8568 #else /* !CONFIG_RT_GROUP_SCHED */
8569 static inline void free_rt_sched_group(struct task_group *tg)
8574 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8579 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8583 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8586 #endif /* CONFIG_RT_GROUP_SCHED */
8588 #ifdef CONFIG_GROUP_SCHED
8589 static void free_sched_group(struct task_group *tg)
8591 free_fair_sched_group(tg);
8592 free_rt_sched_group(tg);
8596 /* allocate runqueue etc for a new task group */
8597 struct task_group *sched_create_group(struct task_group *parent)
8599 struct task_group *tg;
8600 unsigned long flags;
8603 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8605 return ERR_PTR(-ENOMEM);
8607 if (!alloc_fair_sched_group(tg, parent))
8610 if (!alloc_rt_sched_group(tg, parent))
8613 spin_lock_irqsave(&task_group_lock, flags);
8614 for_each_possible_cpu(i) {
8615 register_fair_sched_group(tg, i);
8616 register_rt_sched_group(tg, i);
8618 list_add_rcu(&tg->list, &task_groups);
8620 WARN_ON(!parent); /* root should already exist */
8622 tg->parent = parent;
8623 INIT_LIST_HEAD(&tg->children);
8624 list_add_rcu(&tg->siblings, &parent->children);
8625 spin_unlock_irqrestore(&task_group_lock, flags);
8630 free_sched_group(tg);
8631 return ERR_PTR(-ENOMEM);
8634 /* rcu callback to free various structures associated with a task group */
8635 static void free_sched_group_rcu(struct rcu_head *rhp)
8637 /* now it should be safe to free those cfs_rqs */
8638 free_sched_group(container_of(rhp, struct task_group, rcu));
8641 /* Destroy runqueue etc associated with a task group */
8642 void sched_destroy_group(struct task_group *tg)
8644 unsigned long flags;
8647 spin_lock_irqsave(&task_group_lock, flags);
8648 for_each_possible_cpu(i) {
8649 unregister_fair_sched_group(tg, i);
8650 unregister_rt_sched_group(tg, i);
8652 list_del_rcu(&tg->list);
8653 list_del_rcu(&tg->siblings);
8654 spin_unlock_irqrestore(&task_group_lock, flags);
8656 /* wait for possible concurrent references to cfs_rqs complete */
8657 call_rcu(&tg->rcu, free_sched_group_rcu);
8660 /* change task's runqueue when it moves between groups.
8661 * The caller of this function should have put the task in its new group
8662 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8663 * reflect its new group.
8665 void sched_move_task(struct task_struct *tsk)
8668 unsigned long flags;
8671 rq = task_rq_lock(tsk, &flags);
8673 update_rq_clock(rq);
8675 running = task_current(rq, tsk);
8676 on_rq = tsk->se.on_rq;
8679 dequeue_task(rq, tsk, 0);
8680 if (unlikely(running))
8681 tsk->sched_class->put_prev_task(rq, tsk);
8683 set_task_rq(tsk, task_cpu(tsk));
8685 #ifdef CONFIG_FAIR_GROUP_SCHED
8686 if (tsk->sched_class->moved_group)
8687 tsk->sched_class->moved_group(tsk);
8690 if (unlikely(running))
8691 tsk->sched_class->set_curr_task(rq);
8693 enqueue_task(rq, tsk, 0);
8695 task_rq_unlock(rq, &flags);
8697 #endif /* CONFIG_GROUP_SCHED */
8699 #ifdef CONFIG_FAIR_GROUP_SCHED
8700 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8702 struct cfs_rq *cfs_rq = se->cfs_rq;
8707 dequeue_entity(cfs_rq, se, 0);
8709 se->load.weight = shares;
8710 se->load.inv_weight = 0;
8713 enqueue_entity(cfs_rq, se, 0);
8716 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8718 struct cfs_rq *cfs_rq = se->cfs_rq;
8719 struct rq *rq = cfs_rq->rq;
8720 unsigned long flags;
8722 spin_lock_irqsave(&rq->lock, flags);
8723 __set_se_shares(se, shares);
8724 spin_unlock_irqrestore(&rq->lock, flags);
8727 static DEFINE_MUTEX(shares_mutex);
8729 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8732 unsigned long flags;
8735 * We can't change the weight of the root cgroup.
8740 if (shares < MIN_SHARES)
8741 shares = MIN_SHARES;
8742 else if (shares > MAX_SHARES)
8743 shares = MAX_SHARES;
8745 mutex_lock(&shares_mutex);
8746 if (tg->shares == shares)
8749 spin_lock_irqsave(&task_group_lock, flags);
8750 for_each_possible_cpu(i)
8751 unregister_fair_sched_group(tg, i);
8752 list_del_rcu(&tg->siblings);
8753 spin_unlock_irqrestore(&task_group_lock, flags);
8755 /* wait for any ongoing reference to this group to finish */
8756 synchronize_sched();
8759 * Now we are free to modify the group's share on each cpu
8760 * w/o tripping rebalance_share or load_balance_fair.
8762 tg->shares = shares;
8763 for_each_possible_cpu(i) {
8767 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8768 set_se_shares(tg->se[i], shares);
8772 * Enable load balance activity on this group, by inserting it back on
8773 * each cpu's rq->leaf_cfs_rq_list.
8775 spin_lock_irqsave(&task_group_lock, flags);
8776 for_each_possible_cpu(i)
8777 register_fair_sched_group(tg, i);
8778 list_add_rcu(&tg->siblings, &tg->parent->children);
8779 spin_unlock_irqrestore(&task_group_lock, flags);
8781 mutex_unlock(&shares_mutex);
8785 unsigned long sched_group_shares(struct task_group *tg)
8791 #ifdef CONFIG_RT_GROUP_SCHED
8793 * Ensure that the real time constraints are schedulable.
8795 static DEFINE_MUTEX(rt_constraints_mutex);
8797 static unsigned long to_ratio(u64 period, u64 runtime)
8799 if (runtime == RUNTIME_INF)
8802 return div64_u64(runtime << 20, period);
8805 /* Must be called with tasklist_lock held */
8806 static inline int tg_has_rt_tasks(struct task_group *tg)
8808 struct task_struct *g, *p;
8810 do_each_thread(g, p) {
8811 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8813 } while_each_thread(g, p);
8818 struct rt_schedulable_data {
8819 struct task_group *tg;
8824 static int tg_schedulable(struct task_group *tg, void *data)
8826 struct rt_schedulable_data *d = data;
8827 struct task_group *child;
8828 unsigned long total, sum = 0;
8829 u64 period, runtime;
8831 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8832 runtime = tg->rt_bandwidth.rt_runtime;
8835 period = d->rt_period;
8836 runtime = d->rt_runtime;
8840 * Cannot have more runtime than the period.
8842 if (runtime > period && runtime != RUNTIME_INF)
8846 * Ensure we don't starve existing RT tasks.
8848 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8851 total = to_ratio(period, runtime);
8854 * Nobody can have more than the global setting allows.
8856 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8860 * The sum of our children's runtime should not exceed our own.
8862 list_for_each_entry_rcu(child, &tg->children, siblings) {
8863 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8864 runtime = child->rt_bandwidth.rt_runtime;
8866 if (child == d->tg) {
8867 period = d->rt_period;
8868 runtime = d->rt_runtime;
8871 sum += to_ratio(period, runtime);
8880 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8882 struct rt_schedulable_data data = {
8884 .rt_period = period,
8885 .rt_runtime = runtime,
8888 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8891 static int tg_set_bandwidth(struct task_group *tg,
8892 u64 rt_period, u64 rt_runtime)
8896 mutex_lock(&rt_constraints_mutex);
8897 read_lock(&tasklist_lock);
8898 err = __rt_schedulable(tg, rt_period, rt_runtime);
8902 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8903 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8904 tg->rt_bandwidth.rt_runtime = rt_runtime;
8906 for_each_possible_cpu(i) {
8907 struct rt_rq *rt_rq = tg->rt_rq[i];
8909 spin_lock(&rt_rq->rt_runtime_lock);
8910 rt_rq->rt_runtime = rt_runtime;
8911 spin_unlock(&rt_rq->rt_runtime_lock);
8913 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8915 read_unlock(&tasklist_lock);
8916 mutex_unlock(&rt_constraints_mutex);
8921 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8923 u64 rt_runtime, rt_period;
8925 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8926 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8927 if (rt_runtime_us < 0)
8928 rt_runtime = RUNTIME_INF;
8930 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8933 long sched_group_rt_runtime(struct task_group *tg)
8937 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8940 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8941 do_div(rt_runtime_us, NSEC_PER_USEC);
8942 return rt_runtime_us;
8945 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8947 u64 rt_runtime, rt_period;
8949 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8950 rt_runtime = tg->rt_bandwidth.rt_runtime;
8955 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8958 long sched_group_rt_period(struct task_group *tg)
8962 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8963 do_div(rt_period_us, NSEC_PER_USEC);
8964 return rt_period_us;
8967 static int sched_rt_global_constraints(void)
8969 u64 runtime, period;
8972 if (sysctl_sched_rt_period <= 0)
8975 runtime = global_rt_runtime();
8976 period = global_rt_period();
8979 * Sanity check on the sysctl variables.
8981 if (runtime > period && runtime != RUNTIME_INF)
8984 mutex_lock(&rt_constraints_mutex);
8985 read_lock(&tasklist_lock);
8986 ret = __rt_schedulable(NULL, 0, 0);
8987 read_unlock(&tasklist_lock);
8988 mutex_unlock(&rt_constraints_mutex);
8992 #else /* !CONFIG_RT_GROUP_SCHED */
8993 static int sched_rt_global_constraints(void)
8995 unsigned long flags;
8998 if (sysctl_sched_rt_period <= 0)
9001 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9002 for_each_possible_cpu(i) {
9003 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9005 spin_lock(&rt_rq->rt_runtime_lock);
9006 rt_rq->rt_runtime = global_rt_runtime();
9007 spin_unlock(&rt_rq->rt_runtime_lock);
9009 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9013 #endif /* CONFIG_RT_GROUP_SCHED */
9015 int sched_rt_handler(struct ctl_table *table, int write,
9016 struct file *filp, void __user *buffer, size_t *lenp,
9020 int old_period, old_runtime;
9021 static DEFINE_MUTEX(mutex);
9024 old_period = sysctl_sched_rt_period;
9025 old_runtime = sysctl_sched_rt_runtime;
9027 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
9029 if (!ret && write) {
9030 ret = sched_rt_global_constraints();
9032 sysctl_sched_rt_period = old_period;
9033 sysctl_sched_rt_runtime = old_runtime;
9035 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9036 def_rt_bandwidth.rt_period =
9037 ns_to_ktime(global_rt_period());
9040 mutex_unlock(&mutex);
9045 #ifdef CONFIG_CGROUP_SCHED
9047 /* return corresponding task_group object of a cgroup */
9048 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9050 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9051 struct task_group, css);
9054 static struct cgroup_subsys_state *
9055 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9057 struct task_group *tg, *parent;
9059 if (!cgrp->parent) {
9060 /* This is early initialization for the top cgroup */
9061 return &init_task_group.css;
9064 parent = cgroup_tg(cgrp->parent);
9065 tg = sched_create_group(parent);
9067 return ERR_PTR(-ENOMEM);
9073 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9075 struct task_group *tg = cgroup_tg(cgrp);
9077 sched_destroy_group(tg);
9081 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9082 struct task_struct *tsk)
9084 #ifdef CONFIG_RT_GROUP_SCHED
9085 /* Don't accept realtime tasks when there is no way for them to run */
9086 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
9089 /* We don't support RT-tasks being in separate groups */
9090 if (tsk->sched_class != &fair_sched_class)
9098 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9099 struct cgroup *old_cont, struct task_struct *tsk)
9101 sched_move_task(tsk);
9104 #ifdef CONFIG_FAIR_GROUP_SCHED
9105 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9108 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9111 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9113 struct task_group *tg = cgroup_tg(cgrp);
9115 return (u64) tg->shares;
9117 #endif /* CONFIG_FAIR_GROUP_SCHED */
9119 #ifdef CONFIG_RT_GROUP_SCHED
9120 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9123 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9126 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9128 return sched_group_rt_runtime(cgroup_tg(cgrp));
9131 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9134 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9137 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9139 return sched_group_rt_period(cgroup_tg(cgrp));
9141 #endif /* CONFIG_RT_GROUP_SCHED */
9143 static struct cftype cpu_files[] = {
9144 #ifdef CONFIG_FAIR_GROUP_SCHED
9147 .read_u64 = cpu_shares_read_u64,
9148 .write_u64 = cpu_shares_write_u64,
9151 #ifdef CONFIG_RT_GROUP_SCHED
9153 .name = "rt_runtime_us",
9154 .read_s64 = cpu_rt_runtime_read,
9155 .write_s64 = cpu_rt_runtime_write,
9158 .name = "rt_period_us",
9159 .read_u64 = cpu_rt_period_read_uint,
9160 .write_u64 = cpu_rt_period_write_uint,
9165 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9167 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9170 struct cgroup_subsys cpu_cgroup_subsys = {
9172 .create = cpu_cgroup_create,
9173 .destroy = cpu_cgroup_destroy,
9174 .can_attach = cpu_cgroup_can_attach,
9175 .attach = cpu_cgroup_attach,
9176 .populate = cpu_cgroup_populate,
9177 .subsys_id = cpu_cgroup_subsys_id,
9181 #endif /* CONFIG_CGROUP_SCHED */
9183 #ifdef CONFIG_CGROUP_CPUACCT
9186 * CPU accounting code for task groups.
9188 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9189 * (balbir@in.ibm.com).
9192 /* track cpu usage of a group of tasks */
9194 struct cgroup_subsys_state css;
9195 /* cpuusage holds pointer to a u64-type object on every cpu */
9199 struct cgroup_subsys cpuacct_subsys;
9201 /* return cpu accounting group corresponding to this container */
9202 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9204 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9205 struct cpuacct, css);
9208 /* return cpu accounting group to which this task belongs */
9209 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9211 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9212 struct cpuacct, css);
9215 /* create a new cpu accounting group */
9216 static struct cgroup_subsys_state *cpuacct_create(
9217 struct cgroup_subsys *ss, struct cgroup *cgrp)
9219 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9222 return ERR_PTR(-ENOMEM);
9224 ca->cpuusage = alloc_percpu(u64);
9225 if (!ca->cpuusage) {
9227 return ERR_PTR(-ENOMEM);
9233 /* destroy an existing cpu accounting group */
9235 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9237 struct cpuacct *ca = cgroup_ca(cgrp);
9239 free_percpu(ca->cpuusage);
9243 /* return total cpu usage (in nanoseconds) of a group */
9244 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9246 struct cpuacct *ca = cgroup_ca(cgrp);
9247 u64 totalcpuusage = 0;
9250 for_each_possible_cpu(i) {
9251 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9254 * Take rq->lock to make 64-bit addition safe on 32-bit
9257 spin_lock_irq(&cpu_rq(i)->lock);
9258 totalcpuusage += *cpuusage;
9259 spin_unlock_irq(&cpu_rq(i)->lock);
9262 return totalcpuusage;
9265 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9268 struct cpuacct *ca = cgroup_ca(cgrp);
9277 for_each_possible_cpu(i) {
9278 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9280 spin_lock_irq(&cpu_rq(i)->lock);
9282 spin_unlock_irq(&cpu_rq(i)->lock);
9288 static struct cftype files[] = {
9291 .read_u64 = cpuusage_read,
9292 .write_u64 = cpuusage_write,
9296 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9298 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9302 * charge this task's execution time to its accounting group.
9304 * called with rq->lock held.
9306 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9310 if (!cpuacct_subsys.active)
9315 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
9317 *cpuusage += cputime;
9321 struct cgroup_subsys cpuacct_subsys = {
9323 .create = cpuacct_create,
9324 .destroy = cpuacct_destroy,
9325 .populate = cpuacct_populate,
9326 .subsys_id = cpuacct_subsys_id,
9328 #endif /* CONFIG_CGROUP_CPUACCT */