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 int 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 void task_rq_unlock_wait(struct task_struct *p)
954 struct rq *rq = task_rq(p);
956 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
957 spin_unlock_wait(&rq->lock);
960 static void __task_rq_unlock(struct rq *rq)
963 spin_unlock(&rq->lock);
966 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
969 spin_unlock_irqrestore(&rq->lock, *flags);
973 * this_rq_lock - lock this runqueue and disable interrupts.
975 static struct rq *this_rq_lock(void)
982 spin_lock(&rq->lock);
987 #ifdef CONFIG_SCHED_HRTICK
989 * Use HR-timers to deliver accurate preemption points.
991 * Its all a bit involved since we cannot program an hrt while holding the
992 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
995 * When we get rescheduled we reprogram the hrtick_timer outside of the
1001 * - enabled by features
1002 * - hrtimer is actually high res
1004 static inline int hrtick_enabled(struct rq *rq)
1006 if (!sched_feat(HRTICK))
1008 if (!cpu_active(cpu_of(rq)))
1010 return hrtimer_is_hres_active(&rq->hrtick_timer);
1013 static void hrtick_clear(struct rq *rq)
1015 if (hrtimer_active(&rq->hrtick_timer))
1016 hrtimer_cancel(&rq->hrtick_timer);
1020 * High-resolution timer tick.
1021 * Runs from hardirq context with interrupts disabled.
1023 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1025 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1027 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1029 spin_lock(&rq->lock);
1030 update_rq_clock(rq);
1031 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1032 spin_unlock(&rq->lock);
1034 return HRTIMER_NORESTART;
1039 * called from hardirq (IPI) context
1041 static void __hrtick_start(void *arg)
1043 struct rq *rq = arg;
1045 spin_lock(&rq->lock);
1046 hrtimer_restart(&rq->hrtick_timer);
1047 rq->hrtick_csd_pending = 0;
1048 spin_unlock(&rq->lock);
1052 * Called to set the hrtick timer state.
1054 * called with rq->lock held and irqs disabled
1056 static void hrtick_start(struct rq *rq, u64 delay)
1058 struct hrtimer *timer = &rq->hrtick_timer;
1059 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1061 hrtimer_set_expires(timer, time);
1063 if (rq == this_rq()) {
1064 hrtimer_restart(timer);
1065 } else if (!rq->hrtick_csd_pending) {
1066 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd);
1067 rq->hrtick_csd_pending = 1;
1072 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1074 int cpu = (int)(long)hcpu;
1077 case CPU_UP_CANCELED:
1078 case CPU_UP_CANCELED_FROZEN:
1079 case CPU_DOWN_PREPARE:
1080 case CPU_DOWN_PREPARE_FROZEN:
1082 case CPU_DEAD_FROZEN:
1083 hrtick_clear(cpu_rq(cpu));
1090 static __init void init_hrtick(void)
1092 hotcpu_notifier(hotplug_hrtick, 0);
1096 * Called to set the hrtick timer state.
1098 * called with rq->lock held and irqs disabled
1100 static void hrtick_start(struct rq *rq, u64 delay)
1102 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay), HRTIMER_MODE_REL);
1105 static inline void init_hrtick(void)
1108 #endif /* CONFIG_SMP */
1110 static void init_rq_hrtick(struct rq *rq)
1113 rq->hrtick_csd_pending = 0;
1115 rq->hrtick_csd.flags = 0;
1116 rq->hrtick_csd.func = __hrtick_start;
1117 rq->hrtick_csd.info = rq;
1120 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1121 rq->hrtick_timer.function = hrtick;
1122 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_PERCPU;
1124 #else /* CONFIG_SCHED_HRTICK */
1125 static inline void hrtick_clear(struct rq *rq)
1129 static inline void init_rq_hrtick(struct rq *rq)
1133 static inline void init_hrtick(void)
1136 #endif /* CONFIG_SCHED_HRTICK */
1139 * resched_task - mark a task 'to be rescheduled now'.
1141 * On UP this means the setting of the need_resched flag, on SMP it
1142 * might also involve a cross-CPU call to trigger the scheduler on
1147 #ifndef tsk_is_polling
1148 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1151 static void resched_task(struct task_struct *p)
1155 assert_spin_locked(&task_rq(p)->lock);
1157 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1160 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1163 if (cpu == smp_processor_id())
1166 /* NEED_RESCHED must be visible before we test polling */
1168 if (!tsk_is_polling(p))
1169 smp_send_reschedule(cpu);
1172 static void resched_cpu(int cpu)
1174 struct rq *rq = cpu_rq(cpu);
1175 unsigned long flags;
1177 if (!spin_trylock_irqsave(&rq->lock, flags))
1179 resched_task(cpu_curr(cpu));
1180 spin_unlock_irqrestore(&rq->lock, flags);
1185 * When add_timer_on() enqueues a timer into the timer wheel of an
1186 * idle CPU then this timer might expire before the next timer event
1187 * which is scheduled to wake up that CPU. In case of a completely
1188 * idle system the next event might even be infinite time into the
1189 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1190 * leaves the inner idle loop so the newly added timer is taken into
1191 * account when the CPU goes back to idle and evaluates the timer
1192 * wheel for the next timer event.
1194 void wake_up_idle_cpu(int cpu)
1196 struct rq *rq = cpu_rq(cpu);
1198 if (cpu == smp_processor_id())
1202 * This is safe, as this function is called with the timer
1203 * wheel base lock of (cpu) held. When the CPU is on the way
1204 * to idle and has not yet set rq->curr to idle then it will
1205 * be serialized on the timer wheel base lock and take the new
1206 * timer into account automatically.
1208 if (rq->curr != rq->idle)
1212 * We can set TIF_RESCHED on the idle task of the other CPU
1213 * lockless. The worst case is that the other CPU runs the
1214 * idle task through an additional NOOP schedule()
1216 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1218 /* NEED_RESCHED must be visible before we test polling */
1220 if (!tsk_is_polling(rq->idle))
1221 smp_send_reschedule(cpu);
1223 #endif /* CONFIG_NO_HZ */
1225 #else /* !CONFIG_SMP */
1226 static void resched_task(struct task_struct *p)
1228 assert_spin_locked(&task_rq(p)->lock);
1229 set_tsk_need_resched(p);
1231 #endif /* CONFIG_SMP */
1233 #if BITS_PER_LONG == 32
1234 # define WMULT_CONST (~0UL)
1236 # define WMULT_CONST (1UL << 32)
1239 #define WMULT_SHIFT 32
1242 * Shift right and round:
1244 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1247 * delta *= weight / lw
1249 static unsigned long
1250 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1251 struct load_weight *lw)
1255 if (!lw->inv_weight) {
1256 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1259 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1263 tmp = (u64)delta_exec * weight;
1265 * Check whether we'd overflow the 64-bit multiplication:
1267 if (unlikely(tmp > WMULT_CONST))
1268 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1271 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1273 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1276 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1282 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1289 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1290 * of tasks with abnormal "nice" values across CPUs the contribution that
1291 * each task makes to its run queue's load is weighted according to its
1292 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1293 * scaled version of the new time slice allocation that they receive on time
1297 #define WEIGHT_IDLEPRIO 2
1298 #define WMULT_IDLEPRIO (1 << 31)
1301 * Nice levels are multiplicative, with a gentle 10% change for every
1302 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1303 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1304 * that remained on nice 0.
1306 * The "10% effect" is relative and cumulative: from _any_ nice level,
1307 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1308 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1309 * If a task goes up by ~10% and another task goes down by ~10% then
1310 * the relative distance between them is ~25%.)
1312 static const int prio_to_weight[40] = {
1313 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1314 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1315 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1316 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1317 /* 0 */ 1024, 820, 655, 526, 423,
1318 /* 5 */ 335, 272, 215, 172, 137,
1319 /* 10 */ 110, 87, 70, 56, 45,
1320 /* 15 */ 36, 29, 23, 18, 15,
1324 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1326 * In cases where the weight does not change often, we can use the
1327 * precalculated inverse to speed up arithmetics by turning divisions
1328 * into multiplications:
1330 static const u32 prio_to_wmult[40] = {
1331 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1332 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1333 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1334 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1335 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1336 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1337 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1338 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1341 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1344 * runqueue iterator, to support SMP load-balancing between different
1345 * scheduling classes, without having to expose their internal data
1346 * structures to the load-balancing proper:
1348 struct rq_iterator {
1350 struct task_struct *(*start)(void *);
1351 struct task_struct *(*next)(void *);
1355 static unsigned long
1356 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1357 unsigned long max_load_move, struct sched_domain *sd,
1358 enum cpu_idle_type idle, int *all_pinned,
1359 int *this_best_prio, struct rq_iterator *iterator);
1362 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1363 struct sched_domain *sd, enum cpu_idle_type idle,
1364 struct rq_iterator *iterator);
1367 #ifdef CONFIG_CGROUP_CPUACCT
1368 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1370 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1373 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1375 update_load_add(&rq->load, load);
1378 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1380 update_load_sub(&rq->load, load);
1383 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1384 typedef int (*tg_visitor)(struct task_group *, void *);
1387 * Iterate the full tree, calling @down when first entering a node and @up when
1388 * leaving it for the final time.
1390 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1392 struct task_group *parent, *child;
1396 parent = &root_task_group;
1398 ret = (*down)(parent, data);
1401 list_for_each_entry_rcu(child, &parent->children, siblings) {
1408 ret = (*up)(parent, data);
1413 parent = parent->parent;
1422 static int tg_nop(struct task_group *tg, void *data)
1429 static unsigned long source_load(int cpu, int type);
1430 static unsigned long target_load(int cpu, int type);
1431 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1433 static unsigned long cpu_avg_load_per_task(int cpu)
1435 struct rq *rq = cpu_rq(cpu);
1438 rq->avg_load_per_task = rq->load.weight / rq->nr_running;
1440 rq->avg_load_per_task = 0;
1442 return rq->avg_load_per_task;
1445 #ifdef CONFIG_FAIR_GROUP_SCHED
1447 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1450 * Calculate and set the cpu's group shares.
1453 update_group_shares_cpu(struct task_group *tg, int cpu,
1454 unsigned long sd_shares, unsigned long sd_rq_weight)
1457 unsigned long shares;
1458 unsigned long rq_weight;
1463 rq_weight = tg->cfs_rq[cpu]->load.weight;
1466 * If there are currently no tasks on the cpu pretend there is one of
1467 * average load so that when a new task gets to run here it will not
1468 * get delayed by group starvation.
1472 rq_weight = NICE_0_LOAD;
1475 if (unlikely(rq_weight > sd_rq_weight))
1476 rq_weight = sd_rq_weight;
1479 * \Sum shares * rq_weight
1480 * shares = -----------------------
1484 shares = (sd_shares * rq_weight) / (sd_rq_weight + 1);
1485 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1487 if (abs(shares - tg->se[cpu]->load.weight) >
1488 sysctl_sched_shares_thresh) {
1489 struct rq *rq = cpu_rq(cpu);
1490 unsigned long flags;
1492 spin_lock_irqsave(&rq->lock, flags);
1494 * record the actual number of shares, not the boosted amount.
1496 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1497 tg->cfs_rq[cpu]->rq_weight = rq_weight;
1499 __set_se_shares(tg->se[cpu], shares);
1500 spin_unlock_irqrestore(&rq->lock, flags);
1505 * Re-compute the task group their per cpu shares over the given domain.
1506 * This needs to be done in a bottom-up fashion because the rq weight of a
1507 * parent group depends on the shares of its child groups.
1509 static int tg_shares_up(struct task_group *tg, void *data)
1511 unsigned long rq_weight = 0;
1512 unsigned long shares = 0;
1513 struct sched_domain *sd = data;
1516 for_each_cpu_mask(i, sd->span) {
1517 rq_weight += tg->cfs_rq[i]->load.weight;
1518 shares += tg->cfs_rq[i]->shares;
1521 if ((!shares && rq_weight) || shares > tg->shares)
1522 shares = tg->shares;
1524 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1525 shares = tg->shares;
1528 rq_weight = cpus_weight(sd->span) * NICE_0_LOAD;
1530 for_each_cpu_mask(i, sd->span)
1531 update_group_shares_cpu(tg, i, shares, rq_weight);
1537 * Compute the cpu's hierarchical load factor for each task group.
1538 * This needs to be done in a top-down fashion because the load of a child
1539 * group is a fraction of its parents load.
1541 static int tg_load_down(struct task_group *tg, void *data)
1544 long cpu = (long)data;
1547 load = cpu_rq(cpu)->load.weight;
1549 load = tg->parent->cfs_rq[cpu]->h_load;
1550 load *= tg->cfs_rq[cpu]->shares;
1551 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1554 tg->cfs_rq[cpu]->h_load = load;
1559 static void update_shares(struct sched_domain *sd)
1561 u64 now = cpu_clock(raw_smp_processor_id());
1562 s64 elapsed = now - sd->last_update;
1564 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1565 sd->last_update = now;
1566 walk_tg_tree(tg_nop, tg_shares_up, sd);
1570 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1572 spin_unlock(&rq->lock);
1574 spin_lock(&rq->lock);
1577 static void update_h_load(long cpu)
1579 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1584 static inline void update_shares(struct sched_domain *sd)
1588 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1596 #ifdef CONFIG_FAIR_GROUP_SCHED
1597 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1600 cfs_rq->shares = shares;
1605 #include "sched_stats.h"
1606 #include "sched_idletask.c"
1607 #include "sched_fair.c"
1608 #include "sched_rt.c"
1609 #ifdef CONFIG_SCHED_DEBUG
1610 # include "sched_debug.c"
1613 #define sched_class_highest (&rt_sched_class)
1614 #define for_each_class(class) \
1615 for (class = sched_class_highest; class; class = class->next)
1617 static void inc_nr_running(struct rq *rq)
1622 static void dec_nr_running(struct rq *rq)
1627 static void set_load_weight(struct task_struct *p)
1629 if (task_has_rt_policy(p)) {
1630 p->se.load.weight = prio_to_weight[0] * 2;
1631 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1636 * SCHED_IDLE tasks get minimal weight:
1638 if (p->policy == SCHED_IDLE) {
1639 p->se.load.weight = WEIGHT_IDLEPRIO;
1640 p->se.load.inv_weight = WMULT_IDLEPRIO;
1644 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1645 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1648 static void update_avg(u64 *avg, u64 sample)
1650 s64 diff = sample - *avg;
1654 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1656 sched_info_queued(p);
1657 p->sched_class->enqueue_task(rq, p, wakeup);
1661 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1663 if (sleep && p->se.last_wakeup) {
1664 update_avg(&p->se.avg_overlap,
1665 p->se.sum_exec_runtime - p->se.last_wakeup);
1666 p->se.last_wakeup = 0;
1669 sched_info_dequeued(p);
1670 p->sched_class->dequeue_task(rq, p, sleep);
1675 * __normal_prio - return the priority that is based on the static prio
1677 static inline int __normal_prio(struct task_struct *p)
1679 return p->static_prio;
1683 * Calculate the expected normal priority: i.e. priority
1684 * without taking RT-inheritance into account. Might be
1685 * boosted by interactivity modifiers. Changes upon fork,
1686 * setprio syscalls, and whenever the interactivity
1687 * estimator recalculates.
1689 static inline int normal_prio(struct task_struct *p)
1693 if (task_has_rt_policy(p))
1694 prio = MAX_RT_PRIO-1 - p->rt_priority;
1696 prio = __normal_prio(p);
1701 * Calculate the current priority, i.e. the priority
1702 * taken into account by the scheduler. This value might
1703 * be boosted by RT tasks, or might be boosted by
1704 * interactivity modifiers. Will be RT if the task got
1705 * RT-boosted. If not then it returns p->normal_prio.
1707 static int effective_prio(struct task_struct *p)
1709 p->normal_prio = normal_prio(p);
1711 * If we are RT tasks or we were boosted to RT priority,
1712 * keep the priority unchanged. Otherwise, update priority
1713 * to the normal priority:
1715 if (!rt_prio(p->prio))
1716 return p->normal_prio;
1721 * activate_task - move a task to the runqueue.
1723 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1725 if (task_contributes_to_load(p))
1726 rq->nr_uninterruptible--;
1728 enqueue_task(rq, p, wakeup);
1733 * deactivate_task - remove a task from the runqueue.
1735 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1737 if (task_contributes_to_load(p))
1738 rq->nr_uninterruptible++;
1740 dequeue_task(rq, p, sleep);
1745 * task_curr - is this task currently executing on a CPU?
1746 * @p: the task in question.
1748 inline int task_curr(const struct task_struct *p)
1750 return cpu_curr(task_cpu(p)) == p;
1753 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1755 set_task_rq(p, cpu);
1758 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1759 * successfuly executed on another CPU. We must ensure that updates of
1760 * per-task data have been completed by this moment.
1763 task_thread_info(p)->cpu = cpu;
1767 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1768 const struct sched_class *prev_class,
1769 int oldprio, int running)
1771 if (prev_class != p->sched_class) {
1772 if (prev_class->switched_from)
1773 prev_class->switched_from(rq, p, running);
1774 p->sched_class->switched_to(rq, p, running);
1776 p->sched_class->prio_changed(rq, p, oldprio, running);
1781 /* Used instead of source_load when we know the type == 0 */
1782 static unsigned long weighted_cpuload(const int cpu)
1784 return cpu_rq(cpu)->load.weight;
1788 * Is this task likely cache-hot:
1791 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1796 * Buddy candidates are cache hot:
1798 if (sched_feat(CACHE_HOT_BUDDY) &&
1799 (&p->se == cfs_rq_of(&p->se)->next ||
1800 &p->se == cfs_rq_of(&p->se)->last))
1803 if (p->sched_class != &fair_sched_class)
1806 if (sysctl_sched_migration_cost == -1)
1808 if (sysctl_sched_migration_cost == 0)
1811 delta = now - p->se.exec_start;
1813 return delta < (s64)sysctl_sched_migration_cost;
1817 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1819 int old_cpu = task_cpu(p);
1820 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1821 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1822 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1825 clock_offset = old_rq->clock - new_rq->clock;
1827 #ifdef CONFIG_SCHEDSTATS
1828 if (p->se.wait_start)
1829 p->se.wait_start -= clock_offset;
1830 if (p->se.sleep_start)
1831 p->se.sleep_start -= clock_offset;
1832 if (p->se.block_start)
1833 p->se.block_start -= clock_offset;
1834 if (old_cpu != new_cpu) {
1835 schedstat_inc(p, se.nr_migrations);
1836 if (task_hot(p, old_rq->clock, NULL))
1837 schedstat_inc(p, se.nr_forced2_migrations);
1840 p->se.vruntime -= old_cfsrq->min_vruntime -
1841 new_cfsrq->min_vruntime;
1843 __set_task_cpu(p, new_cpu);
1846 struct migration_req {
1847 struct list_head list;
1849 struct task_struct *task;
1852 struct completion done;
1856 * The task's runqueue lock must be held.
1857 * Returns true if you have to wait for migration thread.
1860 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1862 struct rq *rq = task_rq(p);
1865 * If the task is not on a runqueue (and not running), then
1866 * it is sufficient to simply update the task's cpu field.
1868 if (!p->se.on_rq && !task_running(rq, p)) {
1869 set_task_cpu(p, dest_cpu);
1873 init_completion(&req->done);
1875 req->dest_cpu = dest_cpu;
1876 list_add(&req->list, &rq->migration_queue);
1882 * wait_task_inactive - wait for a thread to unschedule.
1884 * If @match_state is nonzero, it's the @p->state value just checked and
1885 * not expected to change. If it changes, i.e. @p might have woken up,
1886 * then return zero. When we succeed in waiting for @p to be off its CPU,
1887 * we return a positive number (its total switch count). If a second call
1888 * a short while later returns the same number, the caller can be sure that
1889 * @p has remained unscheduled the whole time.
1891 * The caller must ensure that the task *will* unschedule sometime soon,
1892 * else this function might spin for a *long* time. This function can't
1893 * be called with interrupts off, or it may introduce deadlock with
1894 * smp_call_function() if an IPI is sent by the same process we are
1895 * waiting to become inactive.
1897 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1899 unsigned long flags;
1906 * We do the initial early heuristics without holding
1907 * any task-queue locks at all. We'll only try to get
1908 * the runqueue lock when things look like they will
1914 * If the task is actively running on another CPU
1915 * still, just relax and busy-wait without holding
1918 * NOTE! Since we don't hold any locks, it's not
1919 * even sure that "rq" stays as the right runqueue!
1920 * But we don't care, since "task_running()" will
1921 * return false if the runqueue has changed and p
1922 * is actually now running somewhere else!
1924 while (task_running(rq, p)) {
1925 if (match_state && unlikely(p->state != match_state))
1931 * Ok, time to look more closely! We need the rq
1932 * lock now, to be *sure*. If we're wrong, we'll
1933 * just go back and repeat.
1935 rq = task_rq_lock(p, &flags);
1936 trace_sched_wait_task(rq, p);
1937 running = task_running(rq, p);
1938 on_rq = p->se.on_rq;
1940 if (!match_state || p->state == match_state)
1941 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1942 task_rq_unlock(rq, &flags);
1945 * If it changed from the expected state, bail out now.
1947 if (unlikely(!ncsw))
1951 * Was it really running after all now that we
1952 * checked with the proper locks actually held?
1954 * Oops. Go back and try again..
1956 if (unlikely(running)) {
1962 * It's not enough that it's not actively running,
1963 * it must be off the runqueue _entirely_, and not
1966 * So if it wa still runnable (but just not actively
1967 * running right now), it's preempted, and we should
1968 * yield - it could be a while.
1970 if (unlikely(on_rq)) {
1971 schedule_timeout_uninterruptible(1);
1976 * Ahh, all good. It wasn't running, and it wasn't
1977 * runnable, which means that it will never become
1978 * running in the future either. We're all done!
1987 * kick_process - kick a running thread to enter/exit the kernel
1988 * @p: the to-be-kicked thread
1990 * Cause a process which is running on another CPU to enter
1991 * kernel-mode, without any delay. (to get signals handled.)
1993 * NOTE: this function doesnt have to take the runqueue lock,
1994 * because all it wants to ensure is that the remote task enters
1995 * the kernel. If the IPI races and the task has been migrated
1996 * to another CPU then no harm is done and the purpose has been
1999 void kick_process(struct task_struct *p)
2005 if ((cpu != smp_processor_id()) && task_curr(p))
2006 smp_send_reschedule(cpu);
2011 * Return a low guess at the load of a migration-source cpu weighted
2012 * according to the scheduling class and "nice" value.
2014 * We want to under-estimate the load of migration sources, to
2015 * balance conservatively.
2017 static unsigned long source_load(int cpu, int type)
2019 struct rq *rq = cpu_rq(cpu);
2020 unsigned long total = weighted_cpuload(cpu);
2022 if (type == 0 || !sched_feat(LB_BIAS))
2025 return min(rq->cpu_load[type-1], total);
2029 * Return a high guess at the load of a migration-target cpu weighted
2030 * according to the scheduling class and "nice" value.
2032 static unsigned long target_load(int cpu, int type)
2034 struct rq *rq = cpu_rq(cpu);
2035 unsigned long total = weighted_cpuload(cpu);
2037 if (type == 0 || !sched_feat(LB_BIAS))
2040 return max(rq->cpu_load[type-1], total);
2044 * find_idlest_group finds and returns the least busy CPU group within the
2047 static struct sched_group *
2048 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2050 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2051 unsigned long min_load = ULONG_MAX, this_load = 0;
2052 int load_idx = sd->forkexec_idx;
2053 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2056 unsigned long load, avg_load;
2060 /* Skip over this group if it has no CPUs allowed */
2061 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
2064 local_group = cpu_isset(this_cpu, group->cpumask);
2066 /* Tally up the load of all CPUs in the group */
2069 for_each_cpu_mask_nr(i, group->cpumask) {
2070 /* Bias balancing toward cpus of our domain */
2072 load = source_load(i, load_idx);
2074 load = target_load(i, load_idx);
2079 /* Adjust by relative CPU power of the group */
2080 avg_load = sg_div_cpu_power(group,
2081 avg_load * SCHED_LOAD_SCALE);
2084 this_load = avg_load;
2086 } else if (avg_load < min_load) {
2087 min_load = avg_load;
2090 } while (group = group->next, group != sd->groups);
2092 if (!idlest || 100*this_load < imbalance*min_load)
2098 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2101 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2104 unsigned long load, min_load = ULONG_MAX;
2108 /* Traverse only the allowed CPUs */
2109 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
2111 for_each_cpu_mask_nr(i, *tmp) {
2112 load = weighted_cpuload(i);
2114 if (load < min_load || (load == min_load && i == this_cpu)) {
2124 * sched_balance_self: balance the current task (running on cpu) in domains
2125 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2128 * Balance, ie. select the least loaded group.
2130 * Returns the target CPU number, or the same CPU if no balancing is needed.
2132 * preempt must be disabled.
2134 static int sched_balance_self(int cpu, int flag)
2136 struct task_struct *t = current;
2137 struct sched_domain *tmp, *sd = NULL;
2139 for_each_domain(cpu, tmp) {
2141 * If power savings logic is enabled for a domain, stop there.
2143 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2145 if (tmp->flags & flag)
2153 cpumask_t span, tmpmask;
2154 struct sched_group *group;
2155 int new_cpu, weight;
2157 if (!(sd->flags & flag)) {
2163 group = find_idlest_group(sd, t, cpu);
2169 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2170 if (new_cpu == -1 || new_cpu == cpu) {
2171 /* Now try balancing at a lower domain level of cpu */
2176 /* Now try balancing at a lower domain level of new_cpu */
2179 weight = cpus_weight(span);
2180 for_each_domain(cpu, tmp) {
2181 if (weight <= cpus_weight(tmp->span))
2183 if (tmp->flags & flag)
2186 /* while loop will break here if sd == NULL */
2192 #endif /* CONFIG_SMP */
2195 * try_to_wake_up - wake up a thread
2196 * @p: the to-be-woken-up thread
2197 * @state: the mask of task states that can be woken
2198 * @sync: do a synchronous wakeup?
2200 * Put it on the run-queue if it's not already there. The "current"
2201 * thread is always on the run-queue (except when the actual
2202 * re-schedule is in progress), and as such you're allowed to do
2203 * the simpler "current->state = TASK_RUNNING" to mark yourself
2204 * runnable without the overhead of this.
2206 * returns failure only if the task is already active.
2208 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2210 int cpu, orig_cpu, this_cpu, success = 0;
2211 unsigned long flags;
2215 if (!sched_feat(SYNC_WAKEUPS))
2219 if (sched_feat(LB_WAKEUP_UPDATE)) {
2220 struct sched_domain *sd;
2222 this_cpu = raw_smp_processor_id();
2225 for_each_domain(this_cpu, sd) {
2226 if (cpu_isset(cpu, sd->span)) {
2235 rq = task_rq_lock(p, &flags);
2236 old_state = p->state;
2237 if (!(old_state & state))
2245 this_cpu = smp_processor_id();
2248 if (unlikely(task_running(rq, p)))
2251 cpu = p->sched_class->select_task_rq(p, sync);
2252 if (cpu != orig_cpu) {
2253 set_task_cpu(p, cpu);
2254 task_rq_unlock(rq, &flags);
2255 /* might preempt at this point */
2256 rq = task_rq_lock(p, &flags);
2257 old_state = p->state;
2258 if (!(old_state & state))
2263 this_cpu = smp_processor_id();
2267 #ifdef CONFIG_SCHEDSTATS
2268 schedstat_inc(rq, ttwu_count);
2269 if (cpu == this_cpu)
2270 schedstat_inc(rq, ttwu_local);
2272 struct sched_domain *sd;
2273 for_each_domain(this_cpu, sd) {
2274 if (cpu_isset(cpu, sd->span)) {
2275 schedstat_inc(sd, ttwu_wake_remote);
2280 #endif /* CONFIG_SCHEDSTATS */
2283 #endif /* CONFIG_SMP */
2284 schedstat_inc(p, se.nr_wakeups);
2286 schedstat_inc(p, se.nr_wakeups_sync);
2287 if (orig_cpu != cpu)
2288 schedstat_inc(p, se.nr_wakeups_migrate);
2289 if (cpu == this_cpu)
2290 schedstat_inc(p, se.nr_wakeups_local);
2292 schedstat_inc(p, se.nr_wakeups_remote);
2293 update_rq_clock(rq);
2294 activate_task(rq, p, 1);
2298 trace_sched_wakeup(rq, p);
2299 check_preempt_curr(rq, p, sync);
2301 p->state = TASK_RUNNING;
2303 if (p->sched_class->task_wake_up)
2304 p->sched_class->task_wake_up(rq, p);
2307 current->se.last_wakeup = current->se.sum_exec_runtime;
2309 task_rq_unlock(rq, &flags);
2314 int wake_up_process(struct task_struct *p)
2316 return try_to_wake_up(p, TASK_ALL, 0);
2318 EXPORT_SYMBOL(wake_up_process);
2320 int wake_up_state(struct task_struct *p, unsigned int state)
2322 return try_to_wake_up(p, state, 0);
2326 * Perform scheduler related setup for a newly forked process p.
2327 * p is forked by current.
2329 * __sched_fork() is basic setup used by init_idle() too:
2331 static void __sched_fork(struct task_struct *p)
2333 p->se.exec_start = 0;
2334 p->se.sum_exec_runtime = 0;
2335 p->se.prev_sum_exec_runtime = 0;
2336 p->se.last_wakeup = 0;
2337 p->se.avg_overlap = 0;
2339 #ifdef CONFIG_SCHEDSTATS
2340 p->se.wait_start = 0;
2341 p->se.sum_sleep_runtime = 0;
2342 p->se.sleep_start = 0;
2343 p->se.block_start = 0;
2344 p->se.sleep_max = 0;
2345 p->se.block_max = 0;
2347 p->se.slice_max = 0;
2351 INIT_LIST_HEAD(&p->rt.run_list);
2353 INIT_LIST_HEAD(&p->se.group_node);
2355 #ifdef CONFIG_PREEMPT_NOTIFIERS
2356 INIT_HLIST_HEAD(&p->preempt_notifiers);
2360 * We mark the process as running here, but have not actually
2361 * inserted it onto the runqueue yet. This guarantees that
2362 * nobody will actually run it, and a signal or other external
2363 * event cannot wake it up and insert it on the runqueue either.
2365 p->state = TASK_RUNNING;
2369 * fork()/clone()-time setup:
2371 void sched_fork(struct task_struct *p, int clone_flags)
2373 int cpu = get_cpu();
2378 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2380 set_task_cpu(p, cpu);
2383 * Make sure we do not leak PI boosting priority to the child:
2385 p->prio = current->normal_prio;
2386 if (!rt_prio(p->prio))
2387 p->sched_class = &fair_sched_class;
2389 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2390 if (likely(sched_info_on()))
2391 memset(&p->sched_info, 0, sizeof(p->sched_info));
2393 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2396 #ifdef CONFIG_PREEMPT
2397 /* Want to start with kernel preemption disabled. */
2398 task_thread_info(p)->preempt_count = 1;
2404 * wake_up_new_task - wake up a newly created task for the first time.
2406 * This function will do some initial scheduler statistics housekeeping
2407 * that must be done for every newly created context, then puts the task
2408 * on the runqueue and wakes it.
2410 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2412 unsigned long flags;
2415 rq = task_rq_lock(p, &flags);
2416 BUG_ON(p->state != TASK_RUNNING);
2417 update_rq_clock(rq);
2419 p->prio = effective_prio(p);
2421 if (!p->sched_class->task_new || !current->se.on_rq) {
2422 activate_task(rq, p, 0);
2425 * Let the scheduling class do new task startup
2426 * management (if any):
2428 p->sched_class->task_new(rq, p);
2431 trace_sched_wakeup_new(rq, p);
2432 check_preempt_curr(rq, p, 0);
2434 if (p->sched_class->task_wake_up)
2435 p->sched_class->task_wake_up(rq, p);
2437 task_rq_unlock(rq, &flags);
2440 #ifdef CONFIG_PREEMPT_NOTIFIERS
2443 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2444 * @notifier: notifier struct to register
2446 void preempt_notifier_register(struct preempt_notifier *notifier)
2448 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2450 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2453 * preempt_notifier_unregister - no longer interested in preemption notifications
2454 * @notifier: notifier struct to unregister
2456 * This is safe to call from within a preemption notifier.
2458 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2460 hlist_del(¬ifier->link);
2462 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2464 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2466 struct preempt_notifier *notifier;
2467 struct hlist_node *node;
2469 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2470 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2474 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2475 struct task_struct *next)
2477 struct preempt_notifier *notifier;
2478 struct hlist_node *node;
2480 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2481 notifier->ops->sched_out(notifier, next);
2484 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2486 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2491 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2492 struct task_struct *next)
2496 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2499 * prepare_task_switch - prepare to switch tasks
2500 * @rq: the runqueue preparing to switch
2501 * @prev: the current task that is being switched out
2502 * @next: the task we are going to switch to.
2504 * This is called with the rq lock held and interrupts off. It must
2505 * be paired with a subsequent finish_task_switch after the context
2508 * prepare_task_switch sets up locking and calls architecture specific
2512 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2513 struct task_struct *next)
2515 fire_sched_out_preempt_notifiers(prev, next);
2516 prepare_lock_switch(rq, next);
2517 prepare_arch_switch(next);
2521 * finish_task_switch - clean up after a task-switch
2522 * @rq: runqueue associated with task-switch
2523 * @prev: the thread we just switched away from.
2525 * finish_task_switch must be called after the context switch, paired
2526 * with a prepare_task_switch call before the context switch.
2527 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2528 * and do any other architecture-specific cleanup actions.
2530 * Note that we may have delayed dropping an mm in context_switch(). If
2531 * so, we finish that here outside of the runqueue lock. (Doing it
2532 * with the lock held can cause deadlocks; see schedule() for
2535 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2536 __releases(rq->lock)
2538 struct mm_struct *mm = rq->prev_mm;
2544 * A task struct has one reference for the use as "current".
2545 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2546 * schedule one last time. The schedule call will never return, and
2547 * the scheduled task must drop that reference.
2548 * The test for TASK_DEAD must occur while the runqueue locks are
2549 * still held, otherwise prev could be scheduled on another cpu, die
2550 * there before we look at prev->state, and then the reference would
2552 * Manfred Spraul <manfred@colorfullife.com>
2554 prev_state = prev->state;
2555 finish_arch_switch(prev);
2556 finish_lock_switch(rq, prev);
2558 if (current->sched_class->post_schedule)
2559 current->sched_class->post_schedule(rq);
2562 fire_sched_in_preempt_notifiers(current);
2565 if (unlikely(prev_state == TASK_DEAD)) {
2567 * Remove function-return probe instances associated with this
2568 * task and put them back on the free list.
2570 kprobe_flush_task(prev);
2571 put_task_struct(prev);
2576 * schedule_tail - first thing a freshly forked thread must call.
2577 * @prev: the thread we just switched away from.
2579 asmlinkage void schedule_tail(struct task_struct *prev)
2580 __releases(rq->lock)
2582 struct rq *rq = this_rq();
2584 finish_task_switch(rq, prev);
2585 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2586 /* In this case, finish_task_switch does not reenable preemption */
2589 if (current->set_child_tid)
2590 put_user(task_pid_vnr(current), current->set_child_tid);
2594 * context_switch - switch to the new MM and the new
2595 * thread's register state.
2598 context_switch(struct rq *rq, struct task_struct *prev,
2599 struct task_struct *next)
2601 struct mm_struct *mm, *oldmm;
2603 prepare_task_switch(rq, prev, next);
2604 trace_sched_switch(rq, prev, next);
2606 oldmm = prev->active_mm;
2608 * For paravirt, this is coupled with an exit in switch_to to
2609 * combine the page table reload and the switch backend into
2612 arch_enter_lazy_cpu_mode();
2614 if (unlikely(!mm)) {
2615 next->active_mm = oldmm;
2616 atomic_inc(&oldmm->mm_count);
2617 enter_lazy_tlb(oldmm, next);
2619 switch_mm(oldmm, mm, next);
2621 if (unlikely(!prev->mm)) {
2622 prev->active_mm = NULL;
2623 rq->prev_mm = oldmm;
2626 * Since the runqueue lock will be released by the next
2627 * task (which is an invalid locking op but in the case
2628 * of the scheduler it's an obvious special-case), so we
2629 * do an early lockdep release here:
2631 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2632 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2635 /* Here we just switch the register state and the stack. */
2636 switch_to(prev, next, prev);
2640 * this_rq must be evaluated again because prev may have moved
2641 * CPUs since it called schedule(), thus the 'rq' on its stack
2642 * frame will be invalid.
2644 finish_task_switch(this_rq(), prev);
2648 * nr_running, nr_uninterruptible and nr_context_switches:
2650 * externally visible scheduler statistics: current number of runnable
2651 * threads, current number of uninterruptible-sleeping threads, total
2652 * number of context switches performed since bootup.
2654 unsigned long nr_running(void)
2656 unsigned long i, sum = 0;
2658 for_each_online_cpu(i)
2659 sum += cpu_rq(i)->nr_running;
2664 unsigned long nr_uninterruptible(void)
2666 unsigned long i, sum = 0;
2668 for_each_possible_cpu(i)
2669 sum += cpu_rq(i)->nr_uninterruptible;
2672 * Since we read the counters lockless, it might be slightly
2673 * inaccurate. Do not allow it to go below zero though:
2675 if (unlikely((long)sum < 0))
2681 unsigned long long nr_context_switches(void)
2684 unsigned long long sum = 0;
2686 for_each_possible_cpu(i)
2687 sum += cpu_rq(i)->nr_switches;
2692 unsigned long nr_iowait(void)
2694 unsigned long i, sum = 0;
2696 for_each_possible_cpu(i)
2697 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2702 unsigned long nr_active(void)
2704 unsigned long i, running = 0, uninterruptible = 0;
2706 for_each_online_cpu(i) {
2707 running += cpu_rq(i)->nr_running;
2708 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2711 if (unlikely((long)uninterruptible < 0))
2712 uninterruptible = 0;
2714 return running + uninterruptible;
2718 * Update rq->cpu_load[] statistics. This function is usually called every
2719 * scheduler tick (TICK_NSEC).
2721 static void update_cpu_load(struct rq *this_rq)
2723 unsigned long this_load = this_rq->load.weight;
2726 this_rq->nr_load_updates++;
2728 /* Update our load: */
2729 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2730 unsigned long old_load, new_load;
2732 /* scale is effectively 1 << i now, and >> i divides by scale */
2734 old_load = this_rq->cpu_load[i];
2735 new_load = this_load;
2737 * Round up the averaging division if load is increasing. This
2738 * prevents us from getting stuck on 9 if the load is 10, for
2741 if (new_load > old_load)
2742 new_load += scale-1;
2743 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2750 * double_rq_lock - safely lock two runqueues
2752 * Note this does not disable interrupts like task_rq_lock,
2753 * you need to do so manually before calling.
2755 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2756 __acquires(rq1->lock)
2757 __acquires(rq2->lock)
2759 BUG_ON(!irqs_disabled());
2761 spin_lock(&rq1->lock);
2762 __acquire(rq2->lock); /* Fake it out ;) */
2765 spin_lock(&rq1->lock);
2766 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2768 spin_lock(&rq2->lock);
2769 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2772 update_rq_clock(rq1);
2773 update_rq_clock(rq2);
2777 * double_rq_unlock - safely unlock two runqueues
2779 * Note this does not restore interrupts like task_rq_unlock,
2780 * you need to do so manually after calling.
2782 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2783 __releases(rq1->lock)
2784 __releases(rq2->lock)
2786 spin_unlock(&rq1->lock);
2788 spin_unlock(&rq2->lock);
2790 __release(rq2->lock);
2794 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2796 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2797 __releases(this_rq->lock)
2798 __acquires(busiest->lock)
2799 __acquires(this_rq->lock)
2803 if (unlikely(!irqs_disabled())) {
2804 /* printk() doesn't work good under rq->lock */
2805 spin_unlock(&this_rq->lock);
2808 if (unlikely(!spin_trylock(&busiest->lock))) {
2809 if (busiest < this_rq) {
2810 spin_unlock(&this_rq->lock);
2811 spin_lock(&busiest->lock);
2812 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
2815 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
2820 static void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
2821 __releases(busiest->lock)
2823 spin_unlock(&busiest->lock);
2824 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
2828 * If dest_cpu is allowed for this process, migrate the task to it.
2829 * This is accomplished by forcing the cpu_allowed mask to only
2830 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2831 * the cpu_allowed mask is restored.
2833 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2835 struct migration_req req;
2836 unsigned long flags;
2839 rq = task_rq_lock(p, &flags);
2840 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2841 || unlikely(!cpu_active(dest_cpu)))
2844 trace_sched_migrate_task(rq, p, dest_cpu);
2845 /* force the process onto the specified CPU */
2846 if (migrate_task(p, dest_cpu, &req)) {
2847 /* Need to wait for migration thread (might exit: take ref). */
2848 struct task_struct *mt = rq->migration_thread;
2850 get_task_struct(mt);
2851 task_rq_unlock(rq, &flags);
2852 wake_up_process(mt);
2853 put_task_struct(mt);
2854 wait_for_completion(&req.done);
2859 task_rq_unlock(rq, &flags);
2863 * sched_exec - execve() is a valuable balancing opportunity, because at
2864 * this point the task has the smallest effective memory and cache footprint.
2866 void sched_exec(void)
2868 int new_cpu, this_cpu = get_cpu();
2869 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2871 if (new_cpu != this_cpu)
2872 sched_migrate_task(current, new_cpu);
2876 * pull_task - move a task from a remote runqueue to the local runqueue.
2877 * Both runqueues must be locked.
2879 static void pull_task(struct rq *src_rq, struct task_struct *p,
2880 struct rq *this_rq, int this_cpu)
2882 deactivate_task(src_rq, p, 0);
2883 set_task_cpu(p, this_cpu);
2884 activate_task(this_rq, p, 0);
2886 * Note that idle threads have a prio of MAX_PRIO, for this test
2887 * to be always true for them.
2889 check_preempt_curr(this_rq, p, 0);
2893 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2896 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2897 struct sched_domain *sd, enum cpu_idle_type idle,
2901 * We do not migrate tasks that are:
2902 * 1) running (obviously), or
2903 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2904 * 3) are cache-hot on their current CPU.
2906 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2907 schedstat_inc(p, se.nr_failed_migrations_affine);
2912 if (task_running(rq, p)) {
2913 schedstat_inc(p, se.nr_failed_migrations_running);
2918 * Aggressive migration if:
2919 * 1) task is cache cold, or
2920 * 2) too many balance attempts have failed.
2923 if (!task_hot(p, rq->clock, sd) ||
2924 sd->nr_balance_failed > sd->cache_nice_tries) {
2925 #ifdef CONFIG_SCHEDSTATS
2926 if (task_hot(p, rq->clock, sd)) {
2927 schedstat_inc(sd, lb_hot_gained[idle]);
2928 schedstat_inc(p, se.nr_forced_migrations);
2934 if (task_hot(p, rq->clock, sd)) {
2935 schedstat_inc(p, se.nr_failed_migrations_hot);
2941 static unsigned long
2942 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2943 unsigned long max_load_move, struct sched_domain *sd,
2944 enum cpu_idle_type idle, int *all_pinned,
2945 int *this_best_prio, struct rq_iterator *iterator)
2947 int loops = 0, pulled = 0, pinned = 0;
2948 struct task_struct *p;
2949 long rem_load_move = max_load_move;
2951 if (max_load_move == 0)
2957 * Start the load-balancing iterator:
2959 p = iterator->start(iterator->arg);
2961 if (!p || loops++ > sysctl_sched_nr_migrate)
2964 if ((p->se.load.weight >> 1) > rem_load_move ||
2965 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2966 p = iterator->next(iterator->arg);
2970 pull_task(busiest, p, this_rq, this_cpu);
2972 rem_load_move -= p->se.load.weight;
2975 * We only want to steal up to the prescribed amount of weighted load.
2977 if (rem_load_move > 0) {
2978 if (p->prio < *this_best_prio)
2979 *this_best_prio = p->prio;
2980 p = iterator->next(iterator->arg);
2985 * Right now, this is one of only two places pull_task() is called,
2986 * so we can safely collect pull_task() stats here rather than
2987 * inside pull_task().
2989 schedstat_add(sd, lb_gained[idle], pulled);
2992 *all_pinned = pinned;
2994 return max_load_move - rem_load_move;
2998 * move_tasks tries to move up to max_load_move weighted load from busiest to
2999 * this_rq, as part of a balancing operation within domain "sd".
3000 * Returns 1 if successful and 0 otherwise.
3002 * Called with both runqueues locked.
3004 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3005 unsigned long max_load_move,
3006 struct sched_domain *sd, enum cpu_idle_type idle,
3009 const struct sched_class *class = sched_class_highest;
3010 unsigned long total_load_moved = 0;
3011 int this_best_prio = this_rq->curr->prio;
3015 class->load_balance(this_rq, this_cpu, busiest,
3016 max_load_move - total_load_moved,
3017 sd, idle, all_pinned, &this_best_prio);
3018 class = class->next;
3020 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3023 } while (class && max_load_move > total_load_moved);
3025 return total_load_moved > 0;
3029 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3030 struct sched_domain *sd, enum cpu_idle_type idle,
3031 struct rq_iterator *iterator)
3033 struct task_struct *p = iterator->start(iterator->arg);
3037 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3038 pull_task(busiest, p, this_rq, this_cpu);
3040 * Right now, this is only the second place pull_task()
3041 * is called, so we can safely collect pull_task()
3042 * stats here rather than inside pull_task().
3044 schedstat_inc(sd, lb_gained[idle]);
3048 p = iterator->next(iterator->arg);
3055 * move_one_task tries to move exactly one task from busiest to this_rq, as
3056 * part of active balancing operations within "domain".
3057 * Returns 1 if successful and 0 otherwise.
3059 * Called with both runqueues locked.
3061 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3062 struct sched_domain *sd, enum cpu_idle_type idle)
3064 const struct sched_class *class;
3066 for (class = sched_class_highest; class; class = class->next)
3067 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3074 * find_busiest_group finds and returns the busiest CPU group within the
3075 * domain. It calculates and returns the amount of weighted load which
3076 * should be moved to restore balance via the imbalance parameter.
3078 static struct sched_group *
3079 find_busiest_group(struct sched_domain *sd, int this_cpu,
3080 unsigned long *imbalance, enum cpu_idle_type idle,
3081 int *sd_idle, const cpumask_t *cpus, int *balance)
3083 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3084 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3085 unsigned long max_pull;
3086 unsigned long busiest_load_per_task, busiest_nr_running;
3087 unsigned long this_load_per_task, this_nr_running;
3088 int load_idx, group_imb = 0;
3089 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3090 int power_savings_balance = 1;
3091 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3092 unsigned long min_nr_running = ULONG_MAX;
3093 struct sched_group *group_min = NULL, *group_leader = NULL;
3096 max_load = this_load = total_load = total_pwr = 0;
3097 busiest_load_per_task = busiest_nr_running = 0;
3098 this_load_per_task = this_nr_running = 0;
3100 if (idle == CPU_NOT_IDLE)
3101 load_idx = sd->busy_idx;
3102 else if (idle == CPU_NEWLY_IDLE)
3103 load_idx = sd->newidle_idx;
3105 load_idx = sd->idle_idx;
3108 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3111 int __group_imb = 0;
3112 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3113 unsigned long sum_nr_running, sum_weighted_load;
3114 unsigned long sum_avg_load_per_task;
3115 unsigned long avg_load_per_task;
3117 local_group = cpu_isset(this_cpu, group->cpumask);
3120 balance_cpu = first_cpu(group->cpumask);
3122 /* Tally up the load of all CPUs in the group */
3123 sum_weighted_load = sum_nr_running = avg_load = 0;
3124 sum_avg_load_per_task = avg_load_per_task = 0;
3127 min_cpu_load = ~0UL;
3129 for_each_cpu_mask_nr(i, group->cpumask) {
3132 if (!cpu_isset(i, *cpus))
3137 if (*sd_idle && rq->nr_running)
3140 /* Bias balancing toward cpus of our domain */
3142 if (idle_cpu(i) && !first_idle_cpu) {
3147 load = target_load(i, load_idx);
3149 load = source_load(i, load_idx);
3150 if (load > max_cpu_load)
3151 max_cpu_load = load;
3152 if (min_cpu_load > load)
3153 min_cpu_load = load;
3157 sum_nr_running += rq->nr_running;
3158 sum_weighted_load += weighted_cpuload(i);
3160 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3164 * First idle cpu or the first cpu(busiest) in this sched group
3165 * is eligible for doing load balancing at this and above
3166 * domains. In the newly idle case, we will allow all the cpu's
3167 * to do the newly idle load balance.
3169 if (idle != CPU_NEWLY_IDLE && local_group &&
3170 balance_cpu != this_cpu && balance) {
3175 total_load += avg_load;
3176 total_pwr += group->__cpu_power;
3178 /* Adjust by relative CPU power of the group */
3179 avg_load = sg_div_cpu_power(group,
3180 avg_load * SCHED_LOAD_SCALE);
3184 * Consider the group unbalanced when the imbalance is larger
3185 * than the average weight of two tasks.
3187 * APZ: with cgroup the avg task weight can vary wildly and
3188 * might not be a suitable number - should we keep a
3189 * normalized nr_running number somewhere that negates
3192 avg_load_per_task = sg_div_cpu_power(group,
3193 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3195 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3198 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3201 this_load = avg_load;
3203 this_nr_running = sum_nr_running;
3204 this_load_per_task = sum_weighted_load;
3205 } else if (avg_load > max_load &&
3206 (sum_nr_running > group_capacity || __group_imb)) {
3207 max_load = avg_load;
3209 busiest_nr_running = sum_nr_running;
3210 busiest_load_per_task = sum_weighted_load;
3211 group_imb = __group_imb;
3214 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3216 * Busy processors will not participate in power savings
3219 if (idle == CPU_NOT_IDLE ||
3220 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3224 * If the local group is idle or completely loaded
3225 * no need to do power savings balance at this domain
3227 if (local_group && (this_nr_running >= group_capacity ||
3229 power_savings_balance = 0;
3232 * If a group is already running at full capacity or idle,
3233 * don't include that group in power savings calculations
3235 if (!power_savings_balance || sum_nr_running >= group_capacity
3240 * Calculate the group which has the least non-idle load.
3241 * This is the group from where we need to pick up the load
3244 if ((sum_nr_running < min_nr_running) ||
3245 (sum_nr_running == min_nr_running &&
3246 first_cpu(group->cpumask) <
3247 first_cpu(group_min->cpumask))) {
3249 min_nr_running = sum_nr_running;
3250 min_load_per_task = sum_weighted_load /
3255 * Calculate the group which is almost near its
3256 * capacity but still has some space to pick up some load
3257 * from other group and save more power
3259 if (sum_nr_running <= group_capacity - 1) {
3260 if (sum_nr_running > leader_nr_running ||
3261 (sum_nr_running == leader_nr_running &&
3262 first_cpu(group->cpumask) >
3263 first_cpu(group_leader->cpumask))) {
3264 group_leader = group;
3265 leader_nr_running = sum_nr_running;
3270 group = group->next;
3271 } while (group != sd->groups);
3273 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3276 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3278 if (this_load >= avg_load ||
3279 100*max_load <= sd->imbalance_pct*this_load)
3282 busiest_load_per_task /= busiest_nr_running;
3284 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3287 * We're trying to get all the cpus to the average_load, so we don't
3288 * want to push ourselves above the average load, nor do we wish to
3289 * reduce the max loaded cpu below the average load, as either of these
3290 * actions would just result in more rebalancing later, and ping-pong
3291 * tasks around. Thus we look for the minimum possible imbalance.
3292 * Negative imbalances (*we* are more loaded than anyone else) will
3293 * be counted as no imbalance for these purposes -- we can't fix that
3294 * by pulling tasks to us. Be careful of negative numbers as they'll
3295 * appear as very large values with unsigned longs.
3297 if (max_load <= busiest_load_per_task)
3301 * In the presence of smp nice balancing, certain scenarios can have
3302 * max load less than avg load(as we skip the groups at or below
3303 * its cpu_power, while calculating max_load..)
3305 if (max_load < avg_load) {
3307 goto small_imbalance;
3310 /* Don't want to pull so many tasks that a group would go idle */
3311 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3313 /* How much load to actually move to equalise the imbalance */
3314 *imbalance = min(max_pull * busiest->__cpu_power,
3315 (avg_load - this_load) * this->__cpu_power)
3319 * if *imbalance is less than the average load per runnable task
3320 * there is no gaurantee that any tasks will be moved so we'll have
3321 * a think about bumping its value to force at least one task to be
3324 if (*imbalance < busiest_load_per_task) {
3325 unsigned long tmp, pwr_now, pwr_move;
3329 pwr_move = pwr_now = 0;
3331 if (this_nr_running) {
3332 this_load_per_task /= this_nr_running;
3333 if (busiest_load_per_task > this_load_per_task)
3336 this_load_per_task = cpu_avg_load_per_task(this_cpu);
3338 if (max_load - this_load + busiest_load_per_task >=
3339 busiest_load_per_task * imbn) {
3340 *imbalance = busiest_load_per_task;
3345 * OK, we don't have enough imbalance to justify moving tasks,
3346 * however we may be able to increase total CPU power used by
3350 pwr_now += busiest->__cpu_power *
3351 min(busiest_load_per_task, max_load);
3352 pwr_now += this->__cpu_power *
3353 min(this_load_per_task, this_load);
3354 pwr_now /= SCHED_LOAD_SCALE;
3356 /* Amount of load we'd subtract */
3357 tmp = sg_div_cpu_power(busiest,
3358 busiest_load_per_task * SCHED_LOAD_SCALE);
3360 pwr_move += busiest->__cpu_power *
3361 min(busiest_load_per_task, max_load - tmp);
3363 /* Amount of load we'd add */
3364 if (max_load * busiest->__cpu_power <
3365 busiest_load_per_task * SCHED_LOAD_SCALE)
3366 tmp = sg_div_cpu_power(this,
3367 max_load * busiest->__cpu_power);
3369 tmp = sg_div_cpu_power(this,
3370 busiest_load_per_task * SCHED_LOAD_SCALE);
3371 pwr_move += this->__cpu_power *
3372 min(this_load_per_task, this_load + tmp);
3373 pwr_move /= SCHED_LOAD_SCALE;
3375 /* Move if we gain throughput */
3376 if (pwr_move > pwr_now)
3377 *imbalance = busiest_load_per_task;
3383 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3384 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3387 if (this == group_leader && group_leader != group_min) {
3388 *imbalance = min_load_per_task;
3398 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3401 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3402 unsigned long imbalance, const cpumask_t *cpus)
3404 struct rq *busiest = NULL, *rq;
3405 unsigned long max_load = 0;
3408 for_each_cpu_mask_nr(i, group->cpumask) {
3411 if (!cpu_isset(i, *cpus))
3415 wl = weighted_cpuload(i);
3417 if (rq->nr_running == 1 && wl > imbalance)
3420 if (wl > max_load) {
3430 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3431 * so long as it is large enough.
3433 #define MAX_PINNED_INTERVAL 512
3436 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3437 * tasks if there is an imbalance.
3439 static int load_balance(int this_cpu, struct rq *this_rq,
3440 struct sched_domain *sd, enum cpu_idle_type idle,
3441 int *balance, cpumask_t *cpus)
3443 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3444 struct sched_group *group;
3445 unsigned long imbalance;
3447 unsigned long flags;
3452 * When power savings policy is enabled for the parent domain, idle
3453 * sibling can pick up load irrespective of busy siblings. In this case,
3454 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3455 * portraying it as CPU_NOT_IDLE.
3457 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3458 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3461 schedstat_inc(sd, lb_count[idle]);
3465 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3472 schedstat_inc(sd, lb_nobusyg[idle]);
3476 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3478 schedstat_inc(sd, lb_nobusyq[idle]);
3482 BUG_ON(busiest == this_rq);
3484 schedstat_add(sd, lb_imbalance[idle], imbalance);
3487 if (busiest->nr_running > 1) {
3489 * Attempt to move tasks. If find_busiest_group has found
3490 * an imbalance but busiest->nr_running <= 1, the group is
3491 * still unbalanced. ld_moved simply stays zero, so it is
3492 * correctly treated as an imbalance.
3494 local_irq_save(flags);
3495 double_rq_lock(this_rq, busiest);
3496 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3497 imbalance, sd, idle, &all_pinned);
3498 double_rq_unlock(this_rq, busiest);
3499 local_irq_restore(flags);
3502 * some other cpu did the load balance for us.
3504 if (ld_moved && this_cpu != smp_processor_id())
3505 resched_cpu(this_cpu);
3507 /* All tasks on this runqueue were pinned by CPU affinity */
3508 if (unlikely(all_pinned)) {
3509 cpu_clear(cpu_of(busiest), *cpus);
3510 if (!cpus_empty(*cpus))
3517 schedstat_inc(sd, lb_failed[idle]);
3518 sd->nr_balance_failed++;
3520 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3522 spin_lock_irqsave(&busiest->lock, flags);
3524 /* don't kick the migration_thread, if the curr
3525 * task on busiest cpu can't be moved to this_cpu
3527 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3528 spin_unlock_irqrestore(&busiest->lock, flags);
3530 goto out_one_pinned;
3533 if (!busiest->active_balance) {
3534 busiest->active_balance = 1;
3535 busiest->push_cpu = this_cpu;
3538 spin_unlock_irqrestore(&busiest->lock, flags);
3540 wake_up_process(busiest->migration_thread);
3543 * We've kicked active balancing, reset the failure
3546 sd->nr_balance_failed = sd->cache_nice_tries+1;
3549 sd->nr_balance_failed = 0;
3551 if (likely(!active_balance)) {
3552 /* We were unbalanced, so reset the balancing interval */
3553 sd->balance_interval = sd->min_interval;
3556 * If we've begun active balancing, start to back off. This
3557 * case may not be covered by the all_pinned logic if there
3558 * is only 1 task on the busy runqueue (because we don't call
3561 if (sd->balance_interval < sd->max_interval)
3562 sd->balance_interval *= 2;
3565 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3566 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3572 schedstat_inc(sd, lb_balanced[idle]);
3574 sd->nr_balance_failed = 0;
3577 /* tune up the balancing interval */
3578 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3579 (sd->balance_interval < sd->max_interval))
3580 sd->balance_interval *= 2;
3582 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3583 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3594 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3595 * tasks if there is an imbalance.
3597 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3598 * this_rq is locked.
3601 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3604 struct sched_group *group;
3605 struct rq *busiest = NULL;
3606 unsigned long imbalance;
3614 * When power savings policy is enabled for the parent domain, idle
3615 * sibling can pick up load irrespective of busy siblings. In this case,
3616 * let the state of idle sibling percolate up as IDLE, instead of
3617 * portraying it as CPU_NOT_IDLE.
3619 if (sd->flags & SD_SHARE_CPUPOWER &&
3620 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3623 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3625 update_shares_locked(this_rq, sd);
3626 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3627 &sd_idle, cpus, NULL);
3629 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3633 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3635 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3639 BUG_ON(busiest == this_rq);
3641 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3644 if (busiest->nr_running > 1) {
3645 /* Attempt to move tasks */
3646 double_lock_balance(this_rq, busiest);
3647 /* this_rq->clock is already updated */
3648 update_rq_clock(busiest);
3649 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3650 imbalance, sd, CPU_NEWLY_IDLE,
3652 double_unlock_balance(this_rq, busiest);
3654 if (unlikely(all_pinned)) {
3655 cpu_clear(cpu_of(busiest), *cpus);
3656 if (!cpus_empty(*cpus))
3662 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3663 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3664 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3667 sd->nr_balance_failed = 0;
3669 update_shares_locked(this_rq, sd);
3673 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3674 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3675 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3677 sd->nr_balance_failed = 0;
3683 * idle_balance is called by schedule() if this_cpu is about to become
3684 * idle. Attempts to pull tasks from other CPUs.
3686 static void idle_balance(int this_cpu, struct rq *this_rq)
3688 struct sched_domain *sd;
3689 int pulled_task = -1;
3690 unsigned long next_balance = jiffies + HZ;
3693 for_each_domain(this_cpu, sd) {
3694 unsigned long interval;
3696 if (!(sd->flags & SD_LOAD_BALANCE))
3699 if (sd->flags & SD_BALANCE_NEWIDLE)
3700 /* If we've pulled tasks over stop searching: */
3701 pulled_task = load_balance_newidle(this_cpu, this_rq,
3704 interval = msecs_to_jiffies(sd->balance_interval);
3705 if (time_after(next_balance, sd->last_balance + interval))
3706 next_balance = sd->last_balance + interval;
3710 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3712 * We are going idle. next_balance may be set based on
3713 * a busy processor. So reset next_balance.
3715 this_rq->next_balance = next_balance;
3720 * active_load_balance is run by migration threads. It pushes running tasks
3721 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3722 * running on each physical CPU where possible, and avoids physical /
3723 * logical imbalances.
3725 * Called with busiest_rq locked.
3727 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3729 int target_cpu = busiest_rq->push_cpu;
3730 struct sched_domain *sd;
3731 struct rq *target_rq;
3733 /* Is there any task to move? */
3734 if (busiest_rq->nr_running <= 1)
3737 target_rq = cpu_rq(target_cpu);
3740 * This condition is "impossible", if it occurs
3741 * we need to fix it. Originally reported by
3742 * Bjorn Helgaas on a 128-cpu setup.
3744 BUG_ON(busiest_rq == target_rq);
3746 /* move a task from busiest_rq to target_rq */
3747 double_lock_balance(busiest_rq, target_rq);
3748 update_rq_clock(busiest_rq);
3749 update_rq_clock(target_rq);
3751 /* Search for an sd spanning us and the target CPU. */
3752 for_each_domain(target_cpu, sd) {
3753 if ((sd->flags & SD_LOAD_BALANCE) &&
3754 cpu_isset(busiest_cpu, sd->span))
3759 schedstat_inc(sd, alb_count);
3761 if (move_one_task(target_rq, target_cpu, busiest_rq,
3763 schedstat_inc(sd, alb_pushed);
3765 schedstat_inc(sd, alb_failed);
3767 double_unlock_balance(busiest_rq, target_rq);
3772 atomic_t load_balancer;
3774 } nohz ____cacheline_aligned = {
3775 .load_balancer = ATOMIC_INIT(-1),
3776 .cpu_mask = CPU_MASK_NONE,
3780 * This routine will try to nominate the ilb (idle load balancing)
3781 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3782 * load balancing on behalf of all those cpus. If all the cpus in the system
3783 * go into this tickless mode, then there will be no ilb owner (as there is
3784 * no need for one) and all the cpus will sleep till the next wakeup event
3787 * For the ilb owner, tick is not stopped. And this tick will be used
3788 * for idle load balancing. ilb owner will still be part of
3791 * While stopping the tick, this cpu will become the ilb owner if there
3792 * is no other owner. And will be the owner till that cpu becomes busy
3793 * or if all cpus in the system stop their ticks at which point
3794 * there is no need for ilb owner.
3796 * When the ilb owner becomes busy, it nominates another owner, during the
3797 * next busy scheduler_tick()
3799 int select_nohz_load_balancer(int stop_tick)
3801 int cpu = smp_processor_id();
3804 cpu_set(cpu, nohz.cpu_mask);
3805 cpu_rq(cpu)->in_nohz_recently = 1;
3808 * If we are going offline and still the leader, give up!
3810 if (!cpu_active(cpu) &&
3811 atomic_read(&nohz.load_balancer) == cpu) {
3812 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3817 /* time for ilb owner also to sleep */
3818 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3819 if (atomic_read(&nohz.load_balancer) == cpu)
3820 atomic_set(&nohz.load_balancer, -1);
3824 if (atomic_read(&nohz.load_balancer) == -1) {
3825 /* make me the ilb owner */
3826 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3828 } else if (atomic_read(&nohz.load_balancer) == cpu)
3831 if (!cpu_isset(cpu, nohz.cpu_mask))
3834 cpu_clear(cpu, nohz.cpu_mask);
3836 if (atomic_read(&nohz.load_balancer) == cpu)
3837 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3844 static DEFINE_SPINLOCK(balancing);
3847 * It checks each scheduling domain to see if it is due to be balanced,
3848 * and initiates a balancing operation if so.
3850 * Balancing parameters are set up in arch_init_sched_domains.
3852 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3855 struct rq *rq = cpu_rq(cpu);
3856 unsigned long interval;
3857 struct sched_domain *sd;
3858 /* Earliest time when we have to do rebalance again */
3859 unsigned long next_balance = jiffies + 60*HZ;
3860 int update_next_balance = 0;
3864 for_each_domain(cpu, sd) {
3865 if (!(sd->flags & SD_LOAD_BALANCE))
3868 interval = sd->balance_interval;
3869 if (idle != CPU_IDLE)
3870 interval *= sd->busy_factor;
3872 /* scale ms to jiffies */
3873 interval = msecs_to_jiffies(interval);
3874 if (unlikely(!interval))
3876 if (interval > HZ*NR_CPUS/10)
3877 interval = HZ*NR_CPUS/10;
3879 need_serialize = sd->flags & SD_SERIALIZE;
3881 if (need_serialize) {
3882 if (!spin_trylock(&balancing))
3886 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3887 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3889 * We've pulled tasks over so either we're no
3890 * longer idle, or one of our SMT siblings is
3893 idle = CPU_NOT_IDLE;
3895 sd->last_balance = jiffies;
3898 spin_unlock(&balancing);
3900 if (time_after(next_balance, sd->last_balance + interval)) {
3901 next_balance = sd->last_balance + interval;
3902 update_next_balance = 1;
3906 * Stop the load balance at this level. There is another
3907 * CPU in our sched group which is doing load balancing more
3915 * next_balance will be updated only when there is a need.
3916 * When the cpu is attached to null domain for ex, it will not be
3919 if (likely(update_next_balance))
3920 rq->next_balance = next_balance;
3924 * run_rebalance_domains is triggered when needed from the scheduler tick.
3925 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3926 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3928 static void run_rebalance_domains(struct softirq_action *h)
3930 int this_cpu = smp_processor_id();
3931 struct rq *this_rq = cpu_rq(this_cpu);
3932 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3933 CPU_IDLE : CPU_NOT_IDLE;
3935 rebalance_domains(this_cpu, idle);
3939 * If this cpu is the owner for idle load balancing, then do the
3940 * balancing on behalf of the other idle cpus whose ticks are
3943 if (this_rq->idle_at_tick &&
3944 atomic_read(&nohz.load_balancer) == this_cpu) {
3945 cpumask_t cpus = nohz.cpu_mask;
3949 cpu_clear(this_cpu, cpus);
3950 for_each_cpu_mask_nr(balance_cpu, cpus) {
3952 * If this cpu gets work to do, stop the load balancing
3953 * work being done for other cpus. Next load
3954 * balancing owner will pick it up.
3959 rebalance_domains(balance_cpu, CPU_IDLE);
3961 rq = cpu_rq(balance_cpu);
3962 if (time_after(this_rq->next_balance, rq->next_balance))
3963 this_rq->next_balance = rq->next_balance;
3970 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3972 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3973 * idle load balancing owner or decide to stop the periodic load balancing,
3974 * if the whole system is idle.
3976 static inline void trigger_load_balance(struct rq *rq, int cpu)
3980 * If we were in the nohz mode recently and busy at the current
3981 * scheduler tick, then check if we need to nominate new idle
3984 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3985 rq->in_nohz_recently = 0;
3987 if (atomic_read(&nohz.load_balancer) == cpu) {
3988 cpu_clear(cpu, nohz.cpu_mask);
3989 atomic_set(&nohz.load_balancer, -1);
3992 if (atomic_read(&nohz.load_balancer) == -1) {
3994 * simple selection for now: Nominate the
3995 * first cpu in the nohz list to be the next
3998 * TBD: Traverse the sched domains and nominate
3999 * the nearest cpu in the nohz.cpu_mask.
4001 int ilb = first_cpu(nohz.cpu_mask);
4003 if (ilb < nr_cpu_ids)
4009 * If this cpu is idle and doing idle load balancing for all the
4010 * cpus with ticks stopped, is it time for that to stop?
4012 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4013 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4019 * If this cpu is idle and the idle load balancing is done by
4020 * someone else, then no need raise the SCHED_SOFTIRQ
4022 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4023 cpu_isset(cpu, nohz.cpu_mask))
4026 if (time_after_eq(jiffies, rq->next_balance))
4027 raise_softirq(SCHED_SOFTIRQ);
4030 #else /* CONFIG_SMP */
4033 * on UP we do not need to balance between CPUs:
4035 static inline void idle_balance(int cpu, struct rq *rq)
4041 DEFINE_PER_CPU(struct kernel_stat, kstat);
4043 EXPORT_PER_CPU_SYMBOL(kstat);
4046 * Return any ns on the sched_clock that have not yet been banked in
4047 * @p in case that task is currently running.
4049 unsigned long long task_delta_exec(struct task_struct *p)
4051 unsigned long flags;
4055 rq = task_rq_lock(p, &flags);
4057 if (task_current(rq, p)) {
4060 update_rq_clock(rq);
4061 delta_exec = rq->clock - p->se.exec_start;
4062 if ((s64)delta_exec > 0)
4066 task_rq_unlock(rq, &flags);
4072 * Account user cpu time to a process.
4073 * @p: the process that the cpu time gets accounted to
4074 * @cputime: the cpu time spent in user space since the last update
4076 void account_user_time(struct task_struct *p, cputime_t cputime)
4078 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4081 p->utime = cputime_add(p->utime, cputime);
4082 account_group_user_time(p, cputime);
4084 /* Add user time to cpustat. */
4085 tmp = cputime_to_cputime64(cputime);
4086 if (TASK_NICE(p) > 0)
4087 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4089 cpustat->user = cputime64_add(cpustat->user, tmp);
4090 /* Account for user time used */
4091 acct_update_integrals(p);
4095 * Account guest cpu time to a process.
4096 * @p: the process that the cpu time gets accounted to
4097 * @cputime: the cpu time spent in virtual machine since the last update
4099 static void account_guest_time(struct task_struct *p, cputime_t cputime)
4102 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4104 tmp = cputime_to_cputime64(cputime);
4106 p->utime = cputime_add(p->utime, cputime);
4107 account_group_user_time(p, cputime);
4108 p->gtime = cputime_add(p->gtime, cputime);
4110 cpustat->user = cputime64_add(cpustat->user, tmp);
4111 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4115 * Account scaled user cpu time to a process.
4116 * @p: the process that the cpu time gets accounted to
4117 * @cputime: the cpu time spent in user space since the last update
4119 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4121 p->utimescaled = cputime_add(p->utimescaled, cputime);
4125 * Account system cpu time to a process.
4126 * @p: the process that the cpu time gets accounted to
4127 * @hardirq_offset: the offset to subtract from hardirq_count()
4128 * @cputime: the cpu time spent in kernel space since the last update
4130 void account_system_time(struct task_struct *p, int hardirq_offset,
4133 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4134 struct rq *rq = this_rq();
4137 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4138 account_guest_time(p, cputime);
4142 p->stime = cputime_add(p->stime, cputime);
4143 account_group_system_time(p, cputime);
4145 /* Add system time to cpustat. */
4146 tmp = cputime_to_cputime64(cputime);
4147 if (hardirq_count() - hardirq_offset)
4148 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4149 else if (softirq_count())
4150 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4151 else if (p != rq->idle)
4152 cpustat->system = cputime64_add(cpustat->system, tmp);
4153 else if (atomic_read(&rq->nr_iowait) > 0)
4154 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4156 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4157 /* Account for system time used */
4158 acct_update_integrals(p);
4162 * Account scaled system cpu time to a process.
4163 * @p: the process that the cpu time gets accounted to
4164 * @hardirq_offset: the offset to subtract from hardirq_count()
4165 * @cputime: the cpu time spent in kernel space since the last update
4167 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4169 p->stimescaled = cputime_add(p->stimescaled, cputime);
4173 * Account for involuntary wait time.
4174 * @p: the process from which the cpu time has been stolen
4175 * @steal: the cpu time spent in involuntary wait
4177 void account_steal_time(struct task_struct *p, cputime_t steal)
4179 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4180 cputime64_t tmp = cputime_to_cputime64(steal);
4181 struct rq *rq = this_rq();
4183 if (p == rq->idle) {
4184 p->stime = cputime_add(p->stime, steal);
4185 account_group_system_time(p, steal);
4186 if (atomic_read(&rq->nr_iowait) > 0)
4187 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4189 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4191 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4195 * Use precise platform statistics if available:
4197 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4198 cputime_t task_utime(struct task_struct *p)
4203 cputime_t task_stime(struct task_struct *p)
4208 cputime_t task_utime(struct task_struct *p)
4210 clock_t utime = cputime_to_clock_t(p->utime),
4211 total = utime + cputime_to_clock_t(p->stime);
4215 * Use CFS's precise accounting:
4217 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
4221 do_div(temp, total);
4223 utime = (clock_t)temp;
4225 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
4226 return p->prev_utime;
4229 cputime_t task_stime(struct task_struct *p)
4234 * Use CFS's precise accounting. (we subtract utime from
4235 * the total, to make sure the total observed by userspace
4236 * grows monotonically - apps rely on that):
4238 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
4239 cputime_to_clock_t(task_utime(p));
4242 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
4244 return p->prev_stime;
4248 inline cputime_t task_gtime(struct task_struct *p)
4254 * This function gets called by the timer code, with HZ frequency.
4255 * We call it with interrupts disabled.
4257 * It also gets called by the fork code, when changing the parent's
4260 void scheduler_tick(void)
4262 int cpu = smp_processor_id();
4263 struct rq *rq = cpu_rq(cpu);
4264 struct task_struct *curr = rq->curr;
4268 spin_lock(&rq->lock);
4269 update_rq_clock(rq);
4270 update_cpu_load(rq);
4271 curr->sched_class->task_tick(rq, curr, 0);
4272 spin_unlock(&rq->lock);
4275 rq->idle_at_tick = idle_cpu(cpu);
4276 trigger_load_balance(rq, cpu);
4280 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4281 defined(CONFIG_PREEMPT_TRACER))
4283 static inline unsigned long get_parent_ip(unsigned long addr)
4285 if (in_lock_functions(addr)) {
4286 addr = CALLER_ADDR2;
4287 if (in_lock_functions(addr))
4288 addr = CALLER_ADDR3;
4293 void __kprobes add_preempt_count(int val)
4295 #ifdef CONFIG_DEBUG_PREEMPT
4299 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4302 preempt_count() += val;
4303 #ifdef CONFIG_DEBUG_PREEMPT
4305 * Spinlock count overflowing soon?
4307 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4310 if (preempt_count() == val)
4311 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4313 EXPORT_SYMBOL(add_preempt_count);
4315 void __kprobes sub_preempt_count(int val)
4317 #ifdef CONFIG_DEBUG_PREEMPT
4321 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4324 * Is the spinlock portion underflowing?
4326 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4327 !(preempt_count() & PREEMPT_MASK)))
4331 if (preempt_count() == val)
4332 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4333 preempt_count() -= val;
4335 EXPORT_SYMBOL(sub_preempt_count);
4340 * Print scheduling while atomic bug:
4342 static noinline void __schedule_bug(struct task_struct *prev)
4344 struct pt_regs *regs = get_irq_regs();
4346 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4347 prev->comm, prev->pid, preempt_count());
4349 debug_show_held_locks(prev);
4351 if (irqs_disabled())
4352 print_irqtrace_events(prev);
4361 * Various schedule()-time debugging checks and statistics:
4363 static inline void schedule_debug(struct task_struct *prev)
4366 * Test if we are atomic. Since do_exit() needs to call into
4367 * schedule() atomically, we ignore that path for now.
4368 * Otherwise, whine if we are scheduling when we should not be.
4370 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4371 __schedule_bug(prev);
4373 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4375 schedstat_inc(this_rq(), sched_count);
4376 #ifdef CONFIG_SCHEDSTATS
4377 if (unlikely(prev->lock_depth >= 0)) {
4378 schedstat_inc(this_rq(), bkl_count);
4379 schedstat_inc(prev, sched_info.bkl_count);
4385 * Pick up the highest-prio task:
4387 static inline struct task_struct *
4388 pick_next_task(struct rq *rq, struct task_struct *prev)
4390 const struct sched_class *class;
4391 struct task_struct *p;
4394 * Optimization: we know that if all tasks are in
4395 * the fair class we can call that function directly:
4397 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4398 p = fair_sched_class.pick_next_task(rq);
4403 class = sched_class_highest;
4405 p = class->pick_next_task(rq);
4409 * Will never be NULL as the idle class always
4410 * returns a non-NULL p:
4412 class = class->next;
4417 * schedule() is the main scheduler function.
4419 asmlinkage void __sched schedule(void)
4421 struct task_struct *prev, *next;
4422 unsigned long *switch_count;
4428 cpu = smp_processor_id();
4432 switch_count = &prev->nivcsw;
4434 release_kernel_lock(prev);
4435 need_resched_nonpreemptible:
4437 schedule_debug(prev);
4439 if (sched_feat(HRTICK))
4442 spin_lock_irq(&rq->lock);
4443 update_rq_clock(rq);
4444 clear_tsk_need_resched(prev);
4446 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4447 if (unlikely(signal_pending_state(prev->state, prev)))
4448 prev->state = TASK_RUNNING;
4450 deactivate_task(rq, prev, 1);
4451 switch_count = &prev->nvcsw;
4455 if (prev->sched_class->pre_schedule)
4456 prev->sched_class->pre_schedule(rq, prev);
4459 if (unlikely(!rq->nr_running))
4460 idle_balance(cpu, rq);
4462 prev->sched_class->put_prev_task(rq, prev);
4463 next = pick_next_task(rq, prev);
4465 if (likely(prev != next)) {
4466 sched_info_switch(prev, next);
4472 context_switch(rq, prev, next); /* unlocks the rq */
4474 * the context switch might have flipped the stack from under
4475 * us, hence refresh the local variables.
4477 cpu = smp_processor_id();
4480 spin_unlock_irq(&rq->lock);
4482 if (unlikely(reacquire_kernel_lock(current) < 0))
4483 goto need_resched_nonpreemptible;
4485 preempt_enable_no_resched();
4486 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4489 EXPORT_SYMBOL(schedule);
4491 #ifdef CONFIG_PREEMPT
4493 * this is the entry point to schedule() from in-kernel preemption
4494 * off of preempt_enable. Kernel preemptions off return from interrupt
4495 * occur there and call schedule directly.
4497 asmlinkage void __sched preempt_schedule(void)
4499 struct thread_info *ti = current_thread_info();
4502 * If there is a non-zero preempt_count or interrupts are disabled,
4503 * we do not want to preempt the current task. Just return..
4505 if (likely(ti->preempt_count || irqs_disabled()))
4509 add_preempt_count(PREEMPT_ACTIVE);
4511 sub_preempt_count(PREEMPT_ACTIVE);
4514 * Check again in case we missed a preemption opportunity
4515 * between schedule and now.
4518 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4520 EXPORT_SYMBOL(preempt_schedule);
4523 * this is the entry point to schedule() from kernel preemption
4524 * off of irq context.
4525 * Note, that this is called and return with irqs disabled. This will
4526 * protect us against recursive calling from irq.
4528 asmlinkage void __sched preempt_schedule_irq(void)
4530 struct thread_info *ti = current_thread_info();
4532 /* Catch callers which need to be fixed */
4533 BUG_ON(ti->preempt_count || !irqs_disabled());
4536 add_preempt_count(PREEMPT_ACTIVE);
4539 local_irq_disable();
4540 sub_preempt_count(PREEMPT_ACTIVE);
4543 * Check again in case we missed a preemption opportunity
4544 * between schedule and now.
4547 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4550 #endif /* CONFIG_PREEMPT */
4552 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4555 return try_to_wake_up(curr->private, mode, sync);
4557 EXPORT_SYMBOL(default_wake_function);
4560 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4561 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4562 * number) then we wake all the non-exclusive tasks and one exclusive task.
4564 * There are circumstances in which we can try to wake a task which has already
4565 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4566 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4568 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4569 int nr_exclusive, int sync, void *key)
4571 wait_queue_t *curr, *next;
4573 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4574 unsigned flags = curr->flags;
4576 if (curr->func(curr, mode, sync, key) &&
4577 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4583 * __wake_up - wake up threads blocked on a waitqueue.
4585 * @mode: which threads
4586 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4587 * @key: is directly passed to the wakeup function
4589 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4590 int nr_exclusive, void *key)
4592 unsigned long flags;
4594 spin_lock_irqsave(&q->lock, flags);
4595 __wake_up_common(q, mode, nr_exclusive, 0, key);
4596 spin_unlock_irqrestore(&q->lock, flags);
4598 EXPORT_SYMBOL(__wake_up);
4601 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4603 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4605 __wake_up_common(q, mode, 1, 0, NULL);
4609 * __wake_up_sync - wake up threads blocked on a waitqueue.
4611 * @mode: which threads
4612 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4614 * The sync wakeup differs that the waker knows that it will schedule
4615 * away soon, so while the target thread will be woken up, it will not
4616 * be migrated to another CPU - ie. the two threads are 'synchronized'
4617 * with each other. This can prevent needless bouncing between CPUs.
4619 * On UP it can prevent extra preemption.
4622 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4624 unsigned long flags;
4630 if (unlikely(!nr_exclusive))
4633 spin_lock_irqsave(&q->lock, flags);
4634 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4635 spin_unlock_irqrestore(&q->lock, flags);
4637 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4640 * complete: - signals a single thread waiting on this completion
4641 * @x: holds the state of this particular completion
4643 * This will wake up a single thread waiting on this completion. Threads will be
4644 * awakened in the same order in which they were queued.
4646 * See also complete_all(), wait_for_completion() and related routines.
4648 void complete(struct completion *x)
4650 unsigned long flags;
4652 spin_lock_irqsave(&x->wait.lock, flags);
4654 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4655 spin_unlock_irqrestore(&x->wait.lock, flags);
4657 EXPORT_SYMBOL(complete);
4660 * complete_all: - signals all threads waiting on this completion
4661 * @x: holds the state of this particular completion
4663 * This will wake up all threads waiting on this particular completion event.
4665 void complete_all(struct completion *x)
4667 unsigned long flags;
4669 spin_lock_irqsave(&x->wait.lock, flags);
4670 x->done += UINT_MAX/2;
4671 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4672 spin_unlock_irqrestore(&x->wait.lock, flags);
4674 EXPORT_SYMBOL(complete_all);
4676 static inline long __sched
4677 do_wait_for_common(struct completion *x, long timeout, int state)
4680 DECLARE_WAITQUEUE(wait, current);
4682 wait.flags |= WQ_FLAG_EXCLUSIVE;
4683 __add_wait_queue_tail(&x->wait, &wait);
4685 if (signal_pending_state(state, current)) {
4686 timeout = -ERESTARTSYS;
4689 __set_current_state(state);
4690 spin_unlock_irq(&x->wait.lock);
4691 timeout = schedule_timeout(timeout);
4692 spin_lock_irq(&x->wait.lock);
4693 } while (!x->done && timeout);
4694 __remove_wait_queue(&x->wait, &wait);
4699 return timeout ?: 1;
4703 wait_for_common(struct completion *x, long timeout, int state)
4707 spin_lock_irq(&x->wait.lock);
4708 timeout = do_wait_for_common(x, timeout, state);
4709 spin_unlock_irq(&x->wait.lock);
4714 * wait_for_completion: - waits for completion of a task
4715 * @x: holds the state of this particular completion
4717 * This waits to be signaled for completion of a specific task. It is NOT
4718 * interruptible and there is no timeout.
4720 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4721 * and interrupt capability. Also see complete().
4723 void __sched wait_for_completion(struct completion *x)
4725 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4727 EXPORT_SYMBOL(wait_for_completion);
4730 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4731 * @x: holds the state of this particular completion
4732 * @timeout: timeout value in jiffies
4734 * This waits for either a completion of a specific task to be signaled or for a
4735 * specified timeout to expire. The timeout is in jiffies. It is not
4738 unsigned long __sched
4739 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4741 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4743 EXPORT_SYMBOL(wait_for_completion_timeout);
4746 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4747 * @x: holds the state of this particular completion
4749 * This waits for completion of a specific task to be signaled. It is
4752 int __sched wait_for_completion_interruptible(struct completion *x)
4754 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4755 if (t == -ERESTARTSYS)
4759 EXPORT_SYMBOL(wait_for_completion_interruptible);
4762 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4763 * @x: holds the state of this particular completion
4764 * @timeout: timeout value in jiffies
4766 * This waits for either a completion of a specific task to be signaled or for a
4767 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4769 unsigned long __sched
4770 wait_for_completion_interruptible_timeout(struct completion *x,
4771 unsigned long timeout)
4773 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4775 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4778 * wait_for_completion_killable: - waits for completion of a task (killable)
4779 * @x: holds the state of this particular completion
4781 * This waits to be signaled for completion of a specific task. It can be
4782 * interrupted by a kill signal.
4784 int __sched wait_for_completion_killable(struct completion *x)
4786 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4787 if (t == -ERESTARTSYS)
4791 EXPORT_SYMBOL(wait_for_completion_killable);
4794 * try_wait_for_completion - try to decrement a completion without blocking
4795 * @x: completion structure
4797 * Returns: 0 if a decrement cannot be done without blocking
4798 * 1 if a decrement succeeded.
4800 * If a completion is being used as a counting completion,
4801 * attempt to decrement the counter without blocking. This
4802 * enables us to avoid waiting if the resource the completion
4803 * is protecting is not available.
4805 bool try_wait_for_completion(struct completion *x)
4809 spin_lock_irq(&x->wait.lock);
4814 spin_unlock_irq(&x->wait.lock);
4817 EXPORT_SYMBOL(try_wait_for_completion);
4820 * completion_done - Test to see if a completion has any waiters
4821 * @x: completion structure
4823 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4824 * 1 if there are no waiters.
4827 bool completion_done(struct completion *x)
4831 spin_lock_irq(&x->wait.lock);
4834 spin_unlock_irq(&x->wait.lock);
4837 EXPORT_SYMBOL(completion_done);
4840 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4842 unsigned long flags;
4845 init_waitqueue_entry(&wait, current);
4847 __set_current_state(state);
4849 spin_lock_irqsave(&q->lock, flags);
4850 __add_wait_queue(q, &wait);
4851 spin_unlock(&q->lock);
4852 timeout = schedule_timeout(timeout);
4853 spin_lock_irq(&q->lock);
4854 __remove_wait_queue(q, &wait);
4855 spin_unlock_irqrestore(&q->lock, flags);
4860 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4862 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4864 EXPORT_SYMBOL(interruptible_sleep_on);
4867 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4869 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4871 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4873 void __sched sleep_on(wait_queue_head_t *q)
4875 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4877 EXPORT_SYMBOL(sleep_on);
4879 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4881 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4883 EXPORT_SYMBOL(sleep_on_timeout);
4885 #ifdef CONFIG_RT_MUTEXES
4888 * rt_mutex_setprio - set the current priority of a task
4890 * @prio: prio value (kernel-internal form)
4892 * This function changes the 'effective' priority of a task. It does
4893 * not touch ->normal_prio like __setscheduler().
4895 * Used by the rt_mutex code to implement priority inheritance logic.
4897 void rt_mutex_setprio(struct task_struct *p, int prio)
4899 unsigned long flags;
4900 int oldprio, on_rq, running;
4902 const struct sched_class *prev_class = p->sched_class;
4904 BUG_ON(prio < 0 || prio > MAX_PRIO);
4906 rq = task_rq_lock(p, &flags);
4907 update_rq_clock(rq);
4910 on_rq = p->se.on_rq;
4911 running = task_current(rq, p);
4913 dequeue_task(rq, p, 0);
4915 p->sched_class->put_prev_task(rq, p);
4918 p->sched_class = &rt_sched_class;
4920 p->sched_class = &fair_sched_class;
4925 p->sched_class->set_curr_task(rq);
4927 enqueue_task(rq, p, 0);
4929 check_class_changed(rq, p, prev_class, oldprio, running);
4931 task_rq_unlock(rq, &flags);
4936 void set_user_nice(struct task_struct *p, long nice)
4938 int old_prio, delta, on_rq;
4939 unsigned long flags;
4942 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4945 * We have to be careful, if called from sys_setpriority(),
4946 * the task might be in the middle of scheduling on another CPU.
4948 rq = task_rq_lock(p, &flags);
4949 update_rq_clock(rq);
4951 * The RT priorities are set via sched_setscheduler(), but we still
4952 * allow the 'normal' nice value to be set - but as expected
4953 * it wont have any effect on scheduling until the task is
4954 * SCHED_FIFO/SCHED_RR:
4956 if (task_has_rt_policy(p)) {
4957 p->static_prio = NICE_TO_PRIO(nice);
4960 on_rq = p->se.on_rq;
4962 dequeue_task(rq, p, 0);
4964 p->static_prio = NICE_TO_PRIO(nice);
4967 p->prio = effective_prio(p);
4968 delta = p->prio - old_prio;
4971 enqueue_task(rq, p, 0);
4973 * If the task increased its priority or is running and
4974 * lowered its priority, then reschedule its CPU:
4976 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4977 resched_task(rq->curr);
4980 task_rq_unlock(rq, &flags);
4982 EXPORT_SYMBOL(set_user_nice);
4985 * can_nice - check if a task can reduce its nice value
4989 int can_nice(const struct task_struct *p, const int nice)
4991 /* convert nice value [19,-20] to rlimit style value [1,40] */
4992 int nice_rlim = 20 - nice;
4994 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4995 capable(CAP_SYS_NICE));
4998 #ifdef __ARCH_WANT_SYS_NICE
5001 * sys_nice - change the priority of the current process.
5002 * @increment: priority increment
5004 * sys_setpriority is a more generic, but much slower function that
5005 * does similar things.
5007 asmlinkage long sys_nice(int increment)
5012 * Setpriority might change our priority at the same moment.
5013 * We don't have to worry. Conceptually one call occurs first
5014 * and we have a single winner.
5016 if (increment < -40)
5021 nice = PRIO_TO_NICE(current->static_prio) + increment;
5027 if (increment < 0 && !can_nice(current, nice))
5030 retval = security_task_setnice(current, nice);
5034 set_user_nice(current, nice);
5041 * task_prio - return the priority value of a given task.
5042 * @p: the task in question.
5044 * This is the priority value as seen by users in /proc.
5045 * RT tasks are offset by -200. Normal tasks are centered
5046 * around 0, value goes from -16 to +15.
5048 int task_prio(const struct task_struct *p)
5050 return p->prio - MAX_RT_PRIO;
5054 * task_nice - return the nice value of a given task.
5055 * @p: the task in question.
5057 int task_nice(const struct task_struct *p)
5059 return TASK_NICE(p);
5061 EXPORT_SYMBOL(task_nice);
5064 * idle_cpu - is a given cpu idle currently?
5065 * @cpu: the processor in question.
5067 int idle_cpu(int cpu)
5069 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5073 * idle_task - return the idle task for a given cpu.
5074 * @cpu: the processor in question.
5076 struct task_struct *idle_task(int cpu)
5078 return cpu_rq(cpu)->idle;
5082 * find_process_by_pid - find a process with a matching PID value.
5083 * @pid: the pid in question.
5085 static struct task_struct *find_process_by_pid(pid_t pid)
5087 return pid ? find_task_by_vpid(pid) : current;
5090 /* Actually do priority change: must hold rq lock. */
5092 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5094 BUG_ON(p->se.on_rq);
5097 switch (p->policy) {
5101 p->sched_class = &fair_sched_class;
5105 p->sched_class = &rt_sched_class;
5109 p->rt_priority = prio;
5110 p->normal_prio = normal_prio(p);
5111 /* we are holding p->pi_lock already */
5112 p->prio = rt_mutex_getprio(p);
5116 static int __sched_setscheduler(struct task_struct *p, int policy,
5117 struct sched_param *param, bool user)
5119 int retval, oldprio, oldpolicy = -1, on_rq, running;
5120 unsigned long flags;
5121 const struct sched_class *prev_class = p->sched_class;
5124 /* may grab non-irq protected spin_locks */
5125 BUG_ON(in_interrupt());
5127 /* double check policy once rq lock held */
5129 policy = oldpolicy = p->policy;
5130 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5131 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5132 policy != SCHED_IDLE)
5135 * Valid priorities for SCHED_FIFO and SCHED_RR are
5136 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5137 * SCHED_BATCH and SCHED_IDLE is 0.
5139 if (param->sched_priority < 0 ||
5140 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5141 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5143 if (rt_policy(policy) != (param->sched_priority != 0))
5147 * Allow unprivileged RT tasks to decrease priority:
5149 if (user && !capable(CAP_SYS_NICE)) {
5150 if (rt_policy(policy)) {
5151 unsigned long rlim_rtprio;
5153 if (!lock_task_sighand(p, &flags))
5155 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5156 unlock_task_sighand(p, &flags);
5158 /* can't set/change the rt policy */
5159 if (policy != p->policy && !rlim_rtprio)
5162 /* can't increase priority */
5163 if (param->sched_priority > p->rt_priority &&
5164 param->sched_priority > rlim_rtprio)
5168 * Like positive nice levels, dont allow tasks to
5169 * move out of SCHED_IDLE either:
5171 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5174 /* can't change other user's priorities */
5175 if ((current->euid != p->euid) &&
5176 (current->euid != p->uid))
5181 #ifdef CONFIG_RT_GROUP_SCHED
5183 * Do not allow realtime tasks into groups that have no runtime
5186 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5187 task_group(p)->rt_bandwidth.rt_runtime == 0)
5191 retval = security_task_setscheduler(p, policy, param);
5197 * make sure no PI-waiters arrive (or leave) while we are
5198 * changing the priority of the task:
5200 spin_lock_irqsave(&p->pi_lock, flags);
5202 * To be able to change p->policy safely, the apropriate
5203 * runqueue lock must be held.
5205 rq = __task_rq_lock(p);
5206 /* recheck policy now with rq lock held */
5207 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5208 policy = oldpolicy = -1;
5209 __task_rq_unlock(rq);
5210 spin_unlock_irqrestore(&p->pi_lock, flags);
5213 update_rq_clock(rq);
5214 on_rq = p->se.on_rq;
5215 running = task_current(rq, p);
5217 deactivate_task(rq, p, 0);
5219 p->sched_class->put_prev_task(rq, p);
5222 __setscheduler(rq, p, policy, param->sched_priority);
5225 p->sched_class->set_curr_task(rq);
5227 activate_task(rq, p, 0);
5229 check_class_changed(rq, p, prev_class, oldprio, running);
5231 __task_rq_unlock(rq);
5232 spin_unlock_irqrestore(&p->pi_lock, flags);
5234 rt_mutex_adjust_pi(p);
5240 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5241 * @p: the task in question.
5242 * @policy: new policy.
5243 * @param: structure containing the new RT priority.
5245 * NOTE that the task may be already dead.
5247 int sched_setscheduler(struct task_struct *p, int policy,
5248 struct sched_param *param)
5250 return __sched_setscheduler(p, policy, param, true);
5252 EXPORT_SYMBOL_GPL(sched_setscheduler);
5255 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5256 * @p: the task in question.
5257 * @policy: new policy.
5258 * @param: structure containing the new RT priority.
5260 * Just like sched_setscheduler, only don't bother checking if the
5261 * current context has permission. For example, this is needed in
5262 * stop_machine(): we create temporary high priority worker threads,
5263 * but our caller might not have that capability.
5265 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5266 struct sched_param *param)
5268 return __sched_setscheduler(p, policy, param, false);
5272 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5274 struct sched_param lparam;
5275 struct task_struct *p;
5278 if (!param || pid < 0)
5280 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5285 p = find_process_by_pid(pid);
5287 retval = sched_setscheduler(p, policy, &lparam);
5294 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5295 * @pid: the pid in question.
5296 * @policy: new policy.
5297 * @param: structure containing the new RT priority.
5300 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5302 /* negative values for policy are not valid */
5306 return do_sched_setscheduler(pid, policy, param);
5310 * sys_sched_setparam - set/change the RT priority of a thread
5311 * @pid: the pid in question.
5312 * @param: structure containing the new RT priority.
5314 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5316 return do_sched_setscheduler(pid, -1, param);
5320 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5321 * @pid: the pid in question.
5323 asmlinkage long sys_sched_getscheduler(pid_t pid)
5325 struct task_struct *p;
5332 read_lock(&tasklist_lock);
5333 p = find_process_by_pid(pid);
5335 retval = security_task_getscheduler(p);
5339 read_unlock(&tasklist_lock);
5344 * sys_sched_getscheduler - get the RT priority of a thread
5345 * @pid: the pid in question.
5346 * @param: structure containing the RT priority.
5348 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5350 struct sched_param lp;
5351 struct task_struct *p;
5354 if (!param || pid < 0)
5357 read_lock(&tasklist_lock);
5358 p = find_process_by_pid(pid);
5363 retval = security_task_getscheduler(p);
5367 lp.sched_priority = p->rt_priority;
5368 read_unlock(&tasklist_lock);
5371 * This one might sleep, we cannot do it with a spinlock held ...
5373 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5378 read_unlock(&tasklist_lock);
5382 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5384 cpumask_t cpus_allowed;
5385 cpumask_t new_mask = *in_mask;
5386 struct task_struct *p;
5390 read_lock(&tasklist_lock);
5392 p = find_process_by_pid(pid);
5394 read_unlock(&tasklist_lock);
5400 * It is not safe to call set_cpus_allowed with the
5401 * tasklist_lock held. We will bump the task_struct's
5402 * usage count and then drop tasklist_lock.
5405 read_unlock(&tasklist_lock);
5408 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5409 !capable(CAP_SYS_NICE))
5412 retval = security_task_setscheduler(p, 0, NULL);
5416 cpuset_cpus_allowed(p, &cpus_allowed);
5417 cpus_and(new_mask, new_mask, cpus_allowed);
5419 retval = set_cpus_allowed_ptr(p, &new_mask);
5422 cpuset_cpus_allowed(p, &cpus_allowed);
5423 if (!cpus_subset(new_mask, cpus_allowed)) {
5425 * We must have raced with a concurrent cpuset
5426 * update. Just reset the cpus_allowed to the
5427 * cpuset's cpus_allowed
5429 new_mask = cpus_allowed;
5439 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5440 cpumask_t *new_mask)
5442 if (len < sizeof(cpumask_t)) {
5443 memset(new_mask, 0, sizeof(cpumask_t));
5444 } else if (len > sizeof(cpumask_t)) {
5445 len = sizeof(cpumask_t);
5447 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5451 * sys_sched_setaffinity - set the cpu affinity of a process
5452 * @pid: pid of the process
5453 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5454 * @user_mask_ptr: user-space pointer to the new cpu mask
5456 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5457 unsigned long __user *user_mask_ptr)
5462 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5466 return sched_setaffinity(pid, &new_mask);
5469 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5471 struct task_struct *p;
5475 read_lock(&tasklist_lock);
5478 p = find_process_by_pid(pid);
5482 retval = security_task_getscheduler(p);
5486 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5489 read_unlock(&tasklist_lock);
5496 * sys_sched_getaffinity - get the cpu affinity of a process
5497 * @pid: pid of the process
5498 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5499 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5501 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5502 unsigned long __user *user_mask_ptr)
5507 if (len < sizeof(cpumask_t))
5510 ret = sched_getaffinity(pid, &mask);
5514 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5517 return sizeof(cpumask_t);
5521 * sys_sched_yield - yield the current processor to other threads.
5523 * This function yields the current CPU to other tasks. If there are no
5524 * other threads running on this CPU then this function will return.
5526 asmlinkage long sys_sched_yield(void)
5528 struct rq *rq = this_rq_lock();
5530 schedstat_inc(rq, yld_count);
5531 current->sched_class->yield_task(rq);
5534 * Since we are going to call schedule() anyway, there's
5535 * no need to preempt or enable interrupts:
5537 __release(rq->lock);
5538 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5539 _raw_spin_unlock(&rq->lock);
5540 preempt_enable_no_resched();
5547 static void __cond_resched(void)
5549 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5550 __might_sleep(__FILE__, __LINE__);
5553 * The BKS might be reacquired before we have dropped
5554 * PREEMPT_ACTIVE, which could trigger a second
5555 * cond_resched() call.
5558 add_preempt_count(PREEMPT_ACTIVE);
5560 sub_preempt_count(PREEMPT_ACTIVE);
5561 } while (need_resched());
5564 int __sched _cond_resched(void)
5566 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5567 system_state == SYSTEM_RUNNING) {
5573 EXPORT_SYMBOL(_cond_resched);
5576 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5577 * call schedule, and on return reacquire the lock.
5579 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5580 * operations here to prevent schedule() from being called twice (once via
5581 * spin_unlock(), once by hand).
5583 int cond_resched_lock(spinlock_t *lock)
5585 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5588 if (spin_needbreak(lock) || resched) {
5590 if (resched && need_resched())
5599 EXPORT_SYMBOL(cond_resched_lock);
5601 int __sched cond_resched_softirq(void)
5603 BUG_ON(!in_softirq());
5605 if (need_resched() && system_state == SYSTEM_RUNNING) {
5613 EXPORT_SYMBOL(cond_resched_softirq);
5616 * yield - yield the current processor to other threads.
5618 * This is a shortcut for kernel-space yielding - it marks the
5619 * thread runnable and calls sys_sched_yield().
5621 void __sched yield(void)
5623 set_current_state(TASK_RUNNING);
5626 EXPORT_SYMBOL(yield);
5629 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5630 * that process accounting knows that this is a task in IO wait state.
5632 * But don't do that if it is a deliberate, throttling IO wait (this task
5633 * has set its backing_dev_info: the queue against which it should throttle)
5635 void __sched io_schedule(void)
5637 struct rq *rq = &__raw_get_cpu_var(runqueues);
5639 delayacct_blkio_start();
5640 atomic_inc(&rq->nr_iowait);
5642 atomic_dec(&rq->nr_iowait);
5643 delayacct_blkio_end();
5645 EXPORT_SYMBOL(io_schedule);
5647 long __sched io_schedule_timeout(long timeout)
5649 struct rq *rq = &__raw_get_cpu_var(runqueues);
5652 delayacct_blkio_start();
5653 atomic_inc(&rq->nr_iowait);
5654 ret = schedule_timeout(timeout);
5655 atomic_dec(&rq->nr_iowait);
5656 delayacct_blkio_end();
5661 * sys_sched_get_priority_max - return maximum RT priority.
5662 * @policy: scheduling class.
5664 * this syscall returns the maximum rt_priority that can be used
5665 * by a given scheduling class.
5667 asmlinkage long sys_sched_get_priority_max(int policy)
5674 ret = MAX_USER_RT_PRIO-1;
5686 * sys_sched_get_priority_min - return minimum RT priority.
5687 * @policy: scheduling class.
5689 * this syscall returns the minimum rt_priority that can be used
5690 * by a given scheduling class.
5692 asmlinkage long sys_sched_get_priority_min(int policy)
5710 * sys_sched_rr_get_interval - return the default timeslice of a process.
5711 * @pid: pid of the process.
5712 * @interval: userspace pointer to the timeslice value.
5714 * this syscall writes the default timeslice value of a given process
5715 * into the user-space timespec buffer. A value of '0' means infinity.
5718 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5720 struct task_struct *p;
5721 unsigned int time_slice;
5729 read_lock(&tasklist_lock);
5730 p = find_process_by_pid(pid);
5734 retval = security_task_getscheduler(p);
5739 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5740 * tasks that are on an otherwise idle runqueue:
5743 if (p->policy == SCHED_RR) {
5744 time_slice = DEF_TIMESLICE;
5745 } else if (p->policy != SCHED_FIFO) {
5746 struct sched_entity *se = &p->se;
5747 unsigned long flags;
5750 rq = task_rq_lock(p, &flags);
5751 if (rq->cfs.load.weight)
5752 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5753 task_rq_unlock(rq, &flags);
5755 read_unlock(&tasklist_lock);
5756 jiffies_to_timespec(time_slice, &t);
5757 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5761 read_unlock(&tasklist_lock);
5765 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5767 void sched_show_task(struct task_struct *p)
5769 unsigned long free = 0;
5772 state = p->state ? __ffs(p->state) + 1 : 0;
5773 printk(KERN_INFO "%-13.13s %c", p->comm,
5774 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5775 #if BITS_PER_LONG == 32
5776 if (state == TASK_RUNNING)
5777 printk(KERN_CONT " running ");
5779 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5781 if (state == TASK_RUNNING)
5782 printk(KERN_CONT " running task ");
5784 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5786 #ifdef CONFIG_DEBUG_STACK_USAGE
5788 unsigned long *n = end_of_stack(p);
5791 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5794 printk(KERN_CONT "%5lu %5d %6d\n", free,
5795 task_pid_nr(p), task_pid_nr(p->real_parent));
5797 show_stack(p, NULL);
5800 void show_state_filter(unsigned long state_filter)
5802 struct task_struct *g, *p;
5804 #if BITS_PER_LONG == 32
5806 " task PC stack pid father\n");
5809 " task PC stack pid father\n");
5811 read_lock(&tasklist_lock);
5812 do_each_thread(g, p) {
5814 * reset the NMI-timeout, listing all files on a slow
5815 * console might take alot of time:
5817 touch_nmi_watchdog();
5818 if (!state_filter || (p->state & state_filter))
5820 } while_each_thread(g, p);
5822 touch_all_softlockup_watchdogs();
5824 #ifdef CONFIG_SCHED_DEBUG
5825 sysrq_sched_debug_show();
5827 read_unlock(&tasklist_lock);
5829 * Only show locks if all tasks are dumped:
5831 if (state_filter == -1)
5832 debug_show_all_locks();
5835 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5837 idle->sched_class = &idle_sched_class;
5841 * init_idle - set up an idle thread for a given CPU
5842 * @idle: task in question
5843 * @cpu: cpu the idle task belongs to
5845 * NOTE: this function does not set the idle thread's NEED_RESCHED
5846 * flag, to make booting more robust.
5848 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5850 struct rq *rq = cpu_rq(cpu);
5851 unsigned long flags;
5853 spin_lock_irqsave(&rq->lock, flags);
5856 idle->se.exec_start = sched_clock();
5858 idle->prio = idle->normal_prio = MAX_PRIO;
5859 idle->cpus_allowed = cpumask_of_cpu(cpu);
5860 __set_task_cpu(idle, cpu);
5862 rq->curr = rq->idle = idle;
5863 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5866 spin_unlock_irqrestore(&rq->lock, flags);
5868 /* Set the preempt count _outside_ the spinlocks! */
5869 #if defined(CONFIG_PREEMPT)
5870 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5872 task_thread_info(idle)->preempt_count = 0;
5875 * The idle tasks have their own, simple scheduling class:
5877 idle->sched_class = &idle_sched_class;
5881 * In a system that switches off the HZ timer nohz_cpu_mask
5882 * indicates which cpus entered this state. This is used
5883 * in the rcu update to wait only for active cpus. For system
5884 * which do not switch off the HZ timer nohz_cpu_mask should
5885 * always be CPU_MASK_NONE.
5887 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5890 * Increase the granularity value when there are more CPUs,
5891 * because with more CPUs the 'effective latency' as visible
5892 * to users decreases. But the relationship is not linear,
5893 * so pick a second-best guess by going with the log2 of the
5896 * This idea comes from the SD scheduler of Con Kolivas:
5898 static inline void sched_init_granularity(void)
5900 unsigned int factor = 1 + ilog2(num_online_cpus());
5901 const unsigned long limit = 200000000;
5903 sysctl_sched_min_granularity *= factor;
5904 if (sysctl_sched_min_granularity > limit)
5905 sysctl_sched_min_granularity = limit;
5907 sysctl_sched_latency *= factor;
5908 if (sysctl_sched_latency > limit)
5909 sysctl_sched_latency = limit;
5911 sysctl_sched_wakeup_granularity *= factor;
5913 sysctl_sched_shares_ratelimit *= factor;
5918 * This is how migration works:
5920 * 1) we queue a struct migration_req structure in the source CPU's
5921 * runqueue and wake up that CPU's migration thread.
5922 * 2) we down() the locked semaphore => thread blocks.
5923 * 3) migration thread wakes up (implicitly it forces the migrated
5924 * thread off the CPU)
5925 * 4) it gets the migration request and checks whether the migrated
5926 * task is still in the wrong runqueue.
5927 * 5) if it's in the wrong runqueue then the migration thread removes
5928 * it and puts it into the right queue.
5929 * 6) migration thread up()s the semaphore.
5930 * 7) we wake up and the migration is done.
5934 * Change a given task's CPU affinity. Migrate the thread to a
5935 * proper CPU and schedule it away if the CPU it's executing on
5936 * is removed from the allowed bitmask.
5938 * NOTE: the caller must have a valid reference to the task, the
5939 * task must not exit() & deallocate itself prematurely. The
5940 * call is not atomic; no spinlocks may be held.
5942 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5944 struct migration_req req;
5945 unsigned long flags;
5949 rq = task_rq_lock(p, &flags);
5950 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5955 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5956 !cpus_equal(p->cpus_allowed, *new_mask))) {
5961 if (p->sched_class->set_cpus_allowed)
5962 p->sched_class->set_cpus_allowed(p, new_mask);
5964 p->cpus_allowed = *new_mask;
5965 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5968 /* Can the task run on the task's current CPU? If so, we're done */
5969 if (cpu_isset(task_cpu(p), *new_mask))
5972 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5973 /* Need help from migration thread: drop lock and wait. */
5974 task_rq_unlock(rq, &flags);
5975 wake_up_process(rq->migration_thread);
5976 wait_for_completion(&req.done);
5977 tlb_migrate_finish(p->mm);
5981 task_rq_unlock(rq, &flags);
5985 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5988 * Move (not current) task off this cpu, onto dest cpu. We're doing
5989 * this because either it can't run here any more (set_cpus_allowed()
5990 * away from this CPU, or CPU going down), or because we're
5991 * attempting to rebalance this task on exec (sched_exec).
5993 * So we race with normal scheduler movements, but that's OK, as long
5994 * as the task is no longer on this CPU.
5996 * Returns non-zero if task was successfully migrated.
5998 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6000 struct rq *rq_dest, *rq_src;
6003 if (unlikely(!cpu_active(dest_cpu)))
6006 rq_src = cpu_rq(src_cpu);
6007 rq_dest = cpu_rq(dest_cpu);
6009 double_rq_lock(rq_src, rq_dest);
6010 /* Already moved. */
6011 if (task_cpu(p) != src_cpu)
6013 /* Affinity changed (again). */
6014 if (!cpu_isset(dest_cpu, p->cpus_allowed))
6017 on_rq = p->se.on_rq;
6019 deactivate_task(rq_src, p, 0);
6021 set_task_cpu(p, dest_cpu);
6023 activate_task(rq_dest, p, 0);
6024 check_preempt_curr(rq_dest, p, 0);
6029 double_rq_unlock(rq_src, rq_dest);
6034 * migration_thread - this is a highprio system thread that performs
6035 * thread migration by bumping thread off CPU then 'pushing' onto
6038 static int migration_thread(void *data)
6040 int cpu = (long)data;
6044 BUG_ON(rq->migration_thread != current);
6046 set_current_state(TASK_INTERRUPTIBLE);
6047 while (!kthread_should_stop()) {
6048 struct migration_req *req;
6049 struct list_head *head;
6051 spin_lock_irq(&rq->lock);
6053 if (cpu_is_offline(cpu)) {
6054 spin_unlock_irq(&rq->lock);
6058 if (rq->active_balance) {
6059 active_load_balance(rq, cpu);
6060 rq->active_balance = 0;
6063 head = &rq->migration_queue;
6065 if (list_empty(head)) {
6066 spin_unlock_irq(&rq->lock);
6068 set_current_state(TASK_INTERRUPTIBLE);
6071 req = list_entry(head->next, struct migration_req, list);
6072 list_del_init(head->next);
6074 spin_unlock(&rq->lock);
6075 __migrate_task(req->task, cpu, req->dest_cpu);
6078 complete(&req->done);
6080 __set_current_state(TASK_RUNNING);
6084 /* Wait for kthread_stop */
6085 set_current_state(TASK_INTERRUPTIBLE);
6086 while (!kthread_should_stop()) {
6088 set_current_state(TASK_INTERRUPTIBLE);
6090 __set_current_state(TASK_RUNNING);
6094 #ifdef CONFIG_HOTPLUG_CPU
6096 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6100 local_irq_disable();
6101 ret = __migrate_task(p, src_cpu, dest_cpu);
6107 * Figure out where task on dead CPU should go, use force if necessary.
6108 * NOTE: interrupts should be disabled by the caller
6110 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6112 unsigned long flags;
6119 mask = node_to_cpumask(cpu_to_node(dead_cpu));
6120 cpus_and(mask, mask, p->cpus_allowed);
6121 dest_cpu = any_online_cpu(mask);
6123 /* On any allowed CPU? */
6124 if (dest_cpu >= nr_cpu_ids)
6125 dest_cpu = any_online_cpu(p->cpus_allowed);
6127 /* No more Mr. Nice Guy. */
6128 if (dest_cpu >= nr_cpu_ids) {
6129 cpumask_t cpus_allowed;
6131 cpuset_cpus_allowed_locked(p, &cpus_allowed);
6133 * Try to stay on the same cpuset, where the
6134 * current cpuset may be a subset of all cpus.
6135 * The cpuset_cpus_allowed_locked() variant of
6136 * cpuset_cpus_allowed() will not block. It must be
6137 * called within calls to cpuset_lock/cpuset_unlock.
6139 rq = task_rq_lock(p, &flags);
6140 p->cpus_allowed = cpus_allowed;
6141 dest_cpu = any_online_cpu(p->cpus_allowed);
6142 task_rq_unlock(rq, &flags);
6145 * Don't tell them about moving exiting tasks or
6146 * kernel threads (both mm NULL), since they never
6149 if (p->mm && printk_ratelimit()) {
6150 printk(KERN_INFO "process %d (%s) no "
6151 "longer affine to cpu%d\n",
6152 task_pid_nr(p), p->comm, dead_cpu);
6155 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
6159 * While a dead CPU has no uninterruptible tasks queued at this point,
6160 * it might still have a nonzero ->nr_uninterruptible counter, because
6161 * for performance reasons the counter is not stricly tracking tasks to
6162 * their home CPUs. So we just add the counter to another CPU's counter,
6163 * to keep the global sum constant after CPU-down:
6165 static void migrate_nr_uninterruptible(struct rq *rq_src)
6167 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
6168 unsigned long flags;
6170 local_irq_save(flags);
6171 double_rq_lock(rq_src, rq_dest);
6172 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6173 rq_src->nr_uninterruptible = 0;
6174 double_rq_unlock(rq_src, rq_dest);
6175 local_irq_restore(flags);
6178 /* Run through task list and migrate tasks from the dead cpu. */
6179 static void migrate_live_tasks(int src_cpu)
6181 struct task_struct *p, *t;
6183 read_lock(&tasklist_lock);
6185 do_each_thread(t, p) {
6189 if (task_cpu(p) == src_cpu)
6190 move_task_off_dead_cpu(src_cpu, p);
6191 } while_each_thread(t, p);
6193 read_unlock(&tasklist_lock);
6197 * Schedules idle task to be the next runnable task on current CPU.
6198 * It does so by boosting its priority to highest possible.
6199 * Used by CPU offline code.
6201 void sched_idle_next(void)
6203 int this_cpu = smp_processor_id();
6204 struct rq *rq = cpu_rq(this_cpu);
6205 struct task_struct *p = rq->idle;
6206 unsigned long flags;
6208 /* cpu has to be offline */
6209 BUG_ON(cpu_online(this_cpu));
6212 * Strictly not necessary since rest of the CPUs are stopped by now
6213 * and interrupts disabled on the current cpu.
6215 spin_lock_irqsave(&rq->lock, flags);
6217 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6219 update_rq_clock(rq);
6220 activate_task(rq, p, 0);
6222 spin_unlock_irqrestore(&rq->lock, flags);
6226 * Ensures that the idle task is using init_mm right before its cpu goes
6229 void idle_task_exit(void)
6231 struct mm_struct *mm = current->active_mm;
6233 BUG_ON(cpu_online(smp_processor_id()));
6236 switch_mm(mm, &init_mm, current);
6240 /* called under rq->lock with disabled interrupts */
6241 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6243 struct rq *rq = cpu_rq(dead_cpu);
6245 /* Must be exiting, otherwise would be on tasklist. */
6246 BUG_ON(!p->exit_state);
6248 /* Cannot have done final schedule yet: would have vanished. */
6249 BUG_ON(p->state == TASK_DEAD);
6254 * Drop lock around migration; if someone else moves it,
6255 * that's OK. No task can be added to this CPU, so iteration is
6258 spin_unlock_irq(&rq->lock);
6259 move_task_off_dead_cpu(dead_cpu, p);
6260 spin_lock_irq(&rq->lock);
6265 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6266 static void migrate_dead_tasks(unsigned int dead_cpu)
6268 struct rq *rq = cpu_rq(dead_cpu);
6269 struct task_struct *next;
6272 if (!rq->nr_running)
6274 update_rq_clock(rq);
6275 next = pick_next_task(rq, rq->curr);
6278 next->sched_class->put_prev_task(rq, next);
6279 migrate_dead(dead_cpu, next);
6283 #endif /* CONFIG_HOTPLUG_CPU */
6285 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6287 static struct ctl_table sd_ctl_dir[] = {
6289 .procname = "sched_domain",
6295 static struct ctl_table sd_ctl_root[] = {
6297 .ctl_name = CTL_KERN,
6298 .procname = "kernel",
6300 .child = sd_ctl_dir,
6305 static struct ctl_table *sd_alloc_ctl_entry(int n)
6307 struct ctl_table *entry =
6308 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6313 static void sd_free_ctl_entry(struct ctl_table **tablep)
6315 struct ctl_table *entry;
6318 * In the intermediate directories, both the child directory and
6319 * procname are dynamically allocated and could fail but the mode
6320 * will always be set. In the lowest directory the names are
6321 * static strings and all have proc handlers.
6323 for (entry = *tablep; entry->mode; entry++) {
6325 sd_free_ctl_entry(&entry->child);
6326 if (entry->proc_handler == NULL)
6327 kfree(entry->procname);
6335 set_table_entry(struct ctl_table *entry,
6336 const char *procname, void *data, int maxlen,
6337 mode_t mode, proc_handler *proc_handler)
6339 entry->procname = procname;
6341 entry->maxlen = maxlen;
6343 entry->proc_handler = proc_handler;
6346 static struct ctl_table *
6347 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6349 struct ctl_table *table = sd_alloc_ctl_entry(13);
6354 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6355 sizeof(long), 0644, proc_doulongvec_minmax);
6356 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6357 sizeof(long), 0644, proc_doulongvec_minmax);
6358 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6359 sizeof(int), 0644, proc_dointvec_minmax);
6360 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6361 sizeof(int), 0644, proc_dointvec_minmax);
6362 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6363 sizeof(int), 0644, proc_dointvec_minmax);
6364 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6365 sizeof(int), 0644, proc_dointvec_minmax);
6366 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6367 sizeof(int), 0644, proc_dointvec_minmax);
6368 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6369 sizeof(int), 0644, proc_dointvec_minmax);
6370 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6371 sizeof(int), 0644, proc_dointvec_minmax);
6372 set_table_entry(&table[9], "cache_nice_tries",
6373 &sd->cache_nice_tries,
6374 sizeof(int), 0644, proc_dointvec_minmax);
6375 set_table_entry(&table[10], "flags", &sd->flags,
6376 sizeof(int), 0644, proc_dointvec_minmax);
6377 set_table_entry(&table[11], "name", sd->name,
6378 CORENAME_MAX_SIZE, 0444, proc_dostring);
6379 /* &table[12] is terminator */
6384 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6386 struct ctl_table *entry, *table;
6387 struct sched_domain *sd;
6388 int domain_num = 0, i;
6391 for_each_domain(cpu, sd)
6393 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6398 for_each_domain(cpu, sd) {
6399 snprintf(buf, 32, "domain%d", i);
6400 entry->procname = kstrdup(buf, GFP_KERNEL);
6402 entry->child = sd_alloc_ctl_domain_table(sd);
6409 static struct ctl_table_header *sd_sysctl_header;
6410 static void register_sched_domain_sysctl(void)
6412 int i, cpu_num = num_online_cpus();
6413 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6416 WARN_ON(sd_ctl_dir[0].child);
6417 sd_ctl_dir[0].child = entry;
6422 for_each_online_cpu(i) {
6423 snprintf(buf, 32, "cpu%d", i);
6424 entry->procname = kstrdup(buf, GFP_KERNEL);
6426 entry->child = sd_alloc_ctl_cpu_table(i);
6430 WARN_ON(sd_sysctl_header);
6431 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6434 /* may be called multiple times per register */
6435 static void unregister_sched_domain_sysctl(void)
6437 if (sd_sysctl_header)
6438 unregister_sysctl_table(sd_sysctl_header);
6439 sd_sysctl_header = NULL;
6440 if (sd_ctl_dir[0].child)
6441 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6444 static void register_sched_domain_sysctl(void)
6447 static void unregister_sched_domain_sysctl(void)
6452 static void set_rq_online(struct rq *rq)
6455 const struct sched_class *class;
6457 cpu_set(rq->cpu, rq->rd->online);
6460 for_each_class(class) {
6461 if (class->rq_online)
6462 class->rq_online(rq);
6467 static void set_rq_offline(struct rq *rq)
6470 const struct sched_class *class;
6472 for_each_class(class) {
6473 if (class->rq_offline)
6474 class->rq_offline(rq);
6477 cpu_clear(rq->cpu, rq->rd->online);
6483 * migration_call - callback that gets triggered when a CPU is added.
6484 * Here we can start up the necessary migration thread for the new CPU.
6486 static int __cpuinit
6487 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6489 struct task_struct *p;
6490 int cpu = (long)hcpu;
6491 unsigned long flags;
6496 case CPU_UP_PREPARE:
6497 case CPU_UP_PREPARE_FROZEN:
6498 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6501 kthread_bind(p, cpu);
6502 /* Must be high prio: stop_machine expects to yield to it. */
6503 rq = task_rq_lock(p, &flags);
6504 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6505 task_rq_unlock(rq, &flags);
6506 cpu_rq(cpu)->migration_thread = p;
6510 case CPU_ONLINE_FROZEN:
6511 /* Strictly unnecessary, as first user will wake it. */
6512 wake_up_process(cpu_rq(cpu)->migration_thread);
6514 /* Update our root-domain */
6516 spin_lock_irqsave(&rq->lock, flags);
6518 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6522 spin_unlock_irqrestore(&rq->lock, flags);
6525 #ifdef CONFIG_HOTPLUG_CPU
6526 case CPU_UP_CANCELED:
6527 case CPU_UP_CANCELED_FROZEN:
6528 if (!cpu_rq(cpu)->migration_thread)
6530 /* Unbind it from offline cpu so it can run. Fall thru. */
6531 kthread_bind(cpu_rq(cpu)->migration_thread,
6532 any_online_cpu(cpu_online_map));
6533 kthread_stop(cpu_rq(cpu)->migration_thread);
6534 cpu_rq(cpu)->migration_thread = NULL;
6538 case CPU_DEAD_FROZEN:
6539 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6540 migrate_live_tasks(cpu);
6542 kthread_stop(rq->migration_thread);
6543 rq->migration_thread = NULL;
6544 /* Idle task back to normal (off runqueue, low prio) */
6545 spin_lock_irq(&rq->lock);
6546 update_rq_clock(rq);
6547 deactivate_task(rq, rq->idle, 0);
6548 rq->idle->static_prio = MAX_PRIO;
6549 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6550 rq->idle->sched_class = &idle_sched_class;
6551 migrate_dead_tasks(cpu);
6552 spin_unlock_irq(&rq->lock);
6554 migrate_nr_uninterruptible(rq);
6555 BUG_ON(rq->nr_running != 0);
6558 * No need to migrate the tasks: it was best-effort if
6559 * they didn't take sched_hotcpu_mutex. Just wake up
6562 spin_lock_irq(&rq->lock);
6563 while (!list_empty(&rq->migration_queue)) {
6564 struct migration_req *req;
6566 req = list_entry(rq->migration_queue.next,
6567 struct migration_req, list);
6568 list_del_init(&req->list);
6569 complete(&req->done);
6571 spin_unlock_irq(&rq->lock);
6575 case CPU_DYING_FROZEN:
6576 /* Update our root-domain */
6578 spin_lock_irqsave(&rq->lock, flags);
6580 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6583 spin_unlock_irqrestore(&rq->lock, flags);
6590 /* Register at highest priority so that task migration (migrate_all_tasks)
6591 * happens before everything else.
6593 static struct notifier_block __cpuinitdata migration_notifier = {
6594 .notifier_call = migration_call,
6598 static int __init migration_init(void)
6600 void *cpu = (void *)(long)smp_processor_id();
6603 /* Start one for the boot CPU: */
6604 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6605 BUG_ON(err == NOTIFY_BAD);
6606 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6607 register_cpu_notifier(&migration_notifier);
6611 early_initcall(migration_init);
6616 #ifdef CONFIG_SCHED_DEBUG
6618 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6619 cpumask_t *groupmask)
6621 struct sched_group *group = sd->groups;
6624 cpulist_scnprintf(str, sizeof(str), sd->span);
6625 cpus_clear(*groupmask);
6627 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6629 if (!(sd->flags & SD_LOAD_BALANCE)) {
6630 printk("does not load-balance\n");
6632 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6637 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6639 if (!cpu_isset(cpu, sd->span)) {
6640 printk(KERN_ERR "ERROR: domain->span does not contain "
6643 if (!cpu_isset(cpu, group->cpumask)) {
6644 printk(KERN_ERR "ERROR: domain->groups does not contain"
6648 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6652 printk(KERN_ERR "ERROR: group is NULL\n");
6656 if (!group->__cpu_power) {
6657 printk(KERN_CONT "\n");
6658 printk(KERN_ERR "ERROR: domain->cpu_power not "
6663 if (!cpus_weight(group->cpumask)) {
6664 printk(KERN_CONT "\n");
6665 printk(KERN_ERR "ERROR: empty group\n");
6669 if (cpus_intersects(*groupmask, group->cpumask)) {
6670 printk(KERN_CONT "\n");
6671 printk(KERN_ERR "ERROR: repeated CPUs\n");
6675 cpus_or(*groupmask, *groupmask, group->cpumask);
6677 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6678 printk(KERN_CONT " %s", str);
6680 group = group->next;
6681 } while (group != sd->groups);
6682 printk(KERN_CONT "\n");
6684 if (!cpus_equal(sd->span, *groupmask))
6685 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6687 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6688 printk(KERN_ERR "ERROR: parent span is not a superset "
6689 "of domain->span\n");
6693 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6695 cpumask_t *groupmask;
6699 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6703 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6705 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6707 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6712 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6721 #else /* !CONFIG_SCHED_DEBUG */
6722 # define sched_domain_debug(sd, cpu) do { } while (0)
6723 #endif /* CONFIG_SCHED_DEBUG */
6725 static int sd_degenerate(struct sched_domain *sd)
6727 if (cpus_weight(sd->span) == 1)
6730 /* Following flags need at least 2 groups */
6731 if (sd->flags & (SD_LOAD_BALANCE |
6732 SD_BALANCE_NEWIDLE |
6736 SD_SHARE_PKG_RESOURCES)) {
6737 if (sd->groups != sd->groups->next)
6741 /* Following flags don't use groups */
6742 if (sd->flags & (SD_WAKE_IDLE |
6751 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6753 unsigned long cflags = sd->flags, pflags = parent->flags;
6755 if (sd_degenerate(parent))
6758 if (!cpus_equal(sd->span, parent->span))
6761 /* Does parent contain flags not in child? */
6762 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6763 if (cflags & SD_WAKE_AFFINE)
6764 pflags &= ~SD_WAKE_BALANCE;
6765 /* Flags needing groups don't count if only 1 group in parent */
6766 if (parent->groups == parent->groups->next) {
6767 pflags &= ~(SD_LOAD_BALANCE |
6768 SD_BALANCE_NEWIDLE |
6772 SD_SHARE_PKG_RESOURCES);
6774 if (~cflags & pflags)
6780 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6782 unsigned long flags;
6784 spin_lock_irqsave(&rq->lock, flags);
6787 struct root_domain *old_rd = rq->rd;
6789 if (cpu_isset(rq->cpu, old_rd->online))
6792 cpu_clear(rq->cpu, old_rd->span);
6794 if (atomic_dec_and_test(&old_rd->refcount))
6798 atomic_inc(&rd->refcount);
6801 cpu_set(rq->cpu, rd->span);
6802 if (cpu_isset(rq->cpu, cpu_online_map))
6805 spin_unlock_irqrestore(&rq->lock, flags);
6808 static void init_rootdomain(struct root_domain *rd)
6810 memset(rd, 0, sizeof(*rd));
6812 cpus_clear(rd->span);
6813 cpus_clear(rd->online);
6815 cpupri_init(&rd->cpupri);
6818 static void init_defrootdomain(void)
6820 init_rootdomain(&def_root_domain);
6821 atomic_set(&def_root_domain.refcount, 1);
6824 static struct root_domain *alloc_rootdomain(void)
6826 struct root_domain *rd;
6828 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6832 init_rootdomain(rd);
6838 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6839 * hold the hotplug lock.
6842 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6844 struct rq *rq = cpu_rq(cpu);
6845 struct sched_domain *tmp;
6847 /* Remove the sched domains which do not contribute to scheduling. */
6848 for (tmp = sd; tmp; ) {
6849 struct sched_domain *parent = tmp->parent;
6853 if (sd_parent_degenerate(tmp, parent)) {
6854 tmp->parent = parent->parent;
6856 parent->parent->child = tmp;
6861 if (sd && sd_degenerate(sd)) {
6867 sched_domain_debug(sd, cpu);
6869 rq_attach_root(rq, rd);
6870 rcu_assign_pointer(rq->sd, sd);
6873 /* cpus with isolated domains */
6874 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6876 /* Setup the mask of cpus configured for isolated domains */
6877 static int __init isolated_cpu_setup(char *str)
6879 static int __initdata ints[NR_CPUS];
6882 str = get_options(str, ARRAY_SIZE(ints), ints);
6883 cpus_clear(cpu_isolated_map);
6884 for (i = 1; i <= ints[0]; i++)
6885 if (ints[i] < NR_CPUS)
6886 cpu_set(ints[i], cpu_isolated_map);
6890 __setup("isolcpus=", isolated_cpu_setup);
6893 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6894 * to a function which identifies what group(along with sched group) a CPU
6895 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6896 * (due to the fact that we keep track of groups covered with a cpumask_t).
6898 * init_sched_build_groups will build a circular linked list of the groups
6899 * covered by the given span, and will set each group's ->cpumask correctly,
6900 * and ->cpu_power to 0.
6903 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6904 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6905 struct sched_group **sg,
6906 cpumask_t *tmpmask),
6907 cpumask_t *covered, cpumask_t *tmpmask)
6909 struct sched_group *first = NULL, *last = NULL;
6912 cpus_clear(*covered);
6914 for_each_cpu_mask_nr(i, *span) {
6915 struct sched_group *sg;
6916 int group = group_fn(i, cpu_map, &sg, tmpmask);
6919 if (cpu_isset(i, *covered))
6922 cpus_clear(sg->cpumask);
6923 sg->__cpu_power = 0;
6925 for_each_cpu_mask_nr(j, *span) {
6926 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6929 cpu_set(j, *covered);
6930 cpu_set(j, sg->cpumask);
6941 #define SD_NODES_PER_DOMAIN 16
6946 * find_next_best_node - find the next node to include in a sched_domain
6947 * @node: node whose sched_domain we're building
6948 * @used_nodes: nodes already in the sched_domain
6950 * Find the next node to include in a given scheduling domain. Simply
6951 * finds the closest node not already in the @used_nodes map.
6953 * Should use nodemask_t.
6955 static int find_next_best_node(int node, nodemask_t *used_nodes)
6957 int i, n, val, min_val, best_node = 0;
6961 for (i = 0; i < nr_node_ids; i++) {
6962 /* Start at @node */
6963 n = (node + i) % nr_node_ids;
6965 if (!nr_cpus_node(n))
6968 /* Skip already used nodes */
6969 if (node_isset(n, *used_nodes))
6972 /* Simple min distance search */
6973 val = node_distance(node, n);
6975 if (val < min_val) {
6981 node_set(best_node, *used_nodes);
6986 * sched_domain_node_span - get a cpumask for a node's sched_domain
6987 * @node: node whose cpumask we're constructing
6988 * @span: resulting cpumask
6990 * Given a node, construct a good cpumask for its sched_domain to span. It
6991 * should be one that prevents unnecessary balancing, but also spreads tasks
6994 static void sched_domain_node_span(int node, cpumask_t *span)
6996 nodemask_t used_nodes;
6997 node_to_cpumask_ptr(nodemask, node);
7001 nodes_clear(used_nodes);
7003 cpus_or(*span, *span, *nodemask);
7004 node_set(node, used_nodes);
7006 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7007 int next_node = find_next_best_node(node, &used_nodes);
7009 node_to_cpumask_ptr_next(nodemask, next_node);
7010 cpus_or(*span, *span, *nodemask);
7013 #endif /* CONFIG_NUMA */
7015 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7018 * SMT sched-domains:
7020 #ifdef CONFIG_SCHED_SMT
7021 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
7022 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
7025 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7029 *sg = &per_cpu(sched_group_cpus, cpu);
7032 #endif /* CONFIG_SCHED_SMT */
7035 * multi-core sched-domains:
7037 #ifdef CONFIG_SCHED_MC
7038 static DEFINE_PER_CPU(struct sched_domain, core_domains);
7039 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
7040 #endif /* CONFIG_SCHED_MC */
7042 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7044 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7049 *mask = per_cpu(cpu_sibling_map, cpu);
7050 cpus_and(*mask, *mask, *cpu_map);
7051 group = first_cpu(*mask);
7053 *sg = &per_cpu(sched_group_core, group);
7056 #elif defined(CONFIG_SCHED_MC)
7058 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7062 *sg = &per_cpu(sched_group_core, cpu);
7067 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
7068 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
7071 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7075 #ifdef CONFIG_SCHED_MC
7076 *mask = cpu_coregroup_map(cpu);
7077 cpus_and(*mask, *mask, *cpu_map);
7078 group = first_cpu(*mask);
7079 #elif defined(CONFIG_SCHED_SMT)
7080 *mask = per_cpu(cpu_sibling_map, cpu);
7081 cpus_and(*mask, *mask, *cpu_map);
7082 group = first_cpu(*mask);
7087 *sg = &per_cpu(sched_group_phys, group);
7093 * The init_sched_build_groups can't handle what we want to do with node
7094 * groups, so roll our own. Now each node has its own list of groups which
7095 * gets dynamically allocated.
7097 static DEFINE_PER_CPU(struct sched_domain, node_domains);
7098 static struct sched_group ***sched_group_nodes_bycpu;
7100 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
7101 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
7103 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
7104 struct sched_group **sg, cpumask_t *nodemask)
7108 *nodemask = node_to_cpumask(cpu_to_node(cpu));
7109 cpus_and(*nodemask, *nodemask, *cpu_map);
7110 group = first_cpu(*nodemask);
7113 *sg = &per_cpu(sched_group_allnodes, group);
7117 static void init_numa_sched_groups_power(struct sched_group *group_head)
7119 struct sched_group *sg = group_head;
7125 for_each_cpu_mask_nr(j, sg->cpumask) {
7126 struct sched_domain *sd;
7128 sd = &per_cpu(phys_domains, j);
7129 if (j != first_cpu(sd->groups->cpumask)) {
7131 * Only add "power" once for each
7137 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7140 } while (sg != group_head);
7142 #endif /* CONFIG_NUMA */
7145 /* Free memory allocated for various sched_group structures */
7146 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7150 for_each_cpu_mask_nr(cpu, *cpu_map) {
7151 struct sched_group **sched_group_nodes
7152 = sched_group_nodes_bycpu[cpu];
7154 if (!sched_group_nodes)
7157 for (i = 0; i < nr_node_ids; i++) {
7158 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7160 *nodemask = node_to_cpumask(i);
7161 cpus_and(*nodemask, *nodemask, *cpu_map);
7162 if (cpus_empty(*nodemask))
7172 if (oldsg != sched_group_nodes[i])
7175 kfree(sched_group_nodes);
7176 sched_group_nodes_bycpu[cpu] = NULL;
7179 #else /* !CONFIG_NUMA */
7180 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7183 #endif /* CONFIG_NUMA */
7186 * Initialize sched groups cpu_power.
7188 * cpu_power indicates the capacity of sched group, which is used while
7189 * distributing the load between different sched groups in a sched domain.
7190 * Typically cpu_power for all the groups in a sched domain will be same unless
7191 * there are asymmetries in the topology. If there are asymmetries, group
7192 * having more cpu_power will pickup more load compared to the group having
7195 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7196 * the maximum number of tasks a group can handle in the presence of other idle
7197 * or lightly loaded groups in the same sched domain.
7199 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7201 struct sched_domain *child;
7202 struct sched_group *group;
7204 WARN_ON(!sd || !sd->groups);
7206 if (cpu != first_cpu(sd->groups->cpumask))
7211 sd->groups->__cpu_power = 0;
7214 * For perf policy, if the groups in child domain share resources
7215 * (for example cores sharing some portions of the cache hierarchy
7216 * or SMT), then set this domain groups cpu_power such that each group
7217 * can handle only one task, when there are other idle groups in the
7218 * same sched domain.
7220 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7222 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7223 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7228 * add cpu_power of each child group to this groups cpu_power
7230 group = child->groups;
7232 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7233 group = group->next;
7234 } while (group != child->groups);
7238 * Initializers for schedule domains
7239 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7242 #ifdef CONFIG_SCHED_DEBUG
7243 # define SD_INIT_NAME(sd, type) sd->name = #type
7245 # define SD_INIT_NAME(sd, type) do { } while (0)
7248 #define SD_INIT(sd, type) sd_init_##type(sd)
7250 #define SD_INIT_FUNC(type) \
7251 static noinline void sd_init_##type(struct sched_domain *sd) \
7253 memset(sd, 0, sizeof(*sd)); \
7254 *sd = SD_##type##_INIT; \
7255 sd->level = SD_LV_##type; \
7256 SD_INIT_NAME(sd, type); \
7261 SD_INIT_FUNC(ALLNODES)
7264 #ifdef CONFIG_SCHED_SMT
7265 SD_INIT_FUNC(SIBLING)
7267 #ifdef CONFIG_SCHED_MC
7272 * To minimize stack usage kmalloc room for cpumasks and share the
7273 * space as the usage in build_sched_domains() dictates. Used only
7274 * if the amount of space is significant.
7277 cpumask_t tmpmask; /* make this one first */
7280 cpumask_t this_sibling_map;
7281 cpumask_t this_core_map;
7283 cpumask_t send_covered;
7286 cpumask_t domainspan;
7288 cpumask_t notcovered;
7293 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7294 static inline void sched_cpumask_alloc(struct allmasks **masks)
7296 *masks = kmalloc(sizeof(**masks), GFP_KERNEL);
7298 static inline void sched_cpumask_free(struct allmasks *masks)
7303 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7304 static inline void sched_cpumask_alloc(struct allmasks **masks)
7306 static inline void sched_cpumask_free(struct allmasks *masks)
7310 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7311 ((unsigned long)(a) + offsetof(struct allmasks, v))
7313 static int default_relax_domain_level = -1;
7315 static int __init setup_relax_domain_level(char *str)
7319 val = simple_strtoul(str, NULL, 0);
7320 if (val < SD_LV_MAX)
7321 default_relax_domain_level = val;
7325 __setup("relax_domain_level=", setup_relax_domain_level);
7327 static void set_domain_attribute(struct sched_domain *sd,
7328 struct sched_domain_attr *attr)
7332 if (!attr || attr->relax_domain_level < 0) {
7333 if (default_relax_domain_level < 0)
7336 request = default_relax_domain_level;
7338 request = attr->relax_domain_level;
7339 if (request < sd->level) {
7340 /* turn off idle balance on this domain */
7341 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7343 /* turn on idle balance on this domain */
7344 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7349 * Build sched domains for a given set of cpus and attach the sched domains
7350 * to the individual cpus
7352 static int __build_sched_domains(const cpumask_t *cpu_map,
7353 struct sched_domain_attr *attr)
7356 struct root_domain *rd;
7357 SCHED_CPUMASK_DECLARE(allmasks);
7360 struct sched_group **sched_group_nodes = NULL;
7361 int sd_allnodes = 0;
7364 * Allocate the per-node list of sched groups
7366 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
7368 if (!sched_group_nodes) {
7369 printk(KERN_WARNING "Can not alloc sched group node list\n");
7374 rd = alloc_rootdomain();
7376 printk(KERN_WARNING "Cannot alloc root domain\n");
7378 kfree(sched_group_nodes);
7383 /* get space for all scratch cpumask variables */
7384 sched_cpumask_alloc(&allmasks);
7386 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7389 kfree(sched_group_nodes);
7394 tmpmask = (cpumask_t *)allmasks;
7398 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7402 * Set up domains for cpus specified by the cpu_map.
7404 for_each_cpu_mask_nr(i, *cpu_map) {
7405 struct sched_domain *sd = NULL, *p;
7406 SCHED_CPUMASK_VAR(nodemask, allmasks);
7408 *nodemask = node_to_cpumask(cpu_to_node(i));
7409 cpus_and(*nodemask, *nodemask, *cpu_map);
7412 if (cpus_weight(*cpu_map) >
7413 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7414 sd = &per_cpu(allnodes_domains, i);
7415 SD_INIT(sd, ALLNODES);
7416 set_domain_attribute(sd, attr);
7417 sd->span = *cpu_map;
7418 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7424 sd = &per_cpu(node_domains, i);
7426 set_domain_attribute(sd, attr);
7427 sched_domain_node_span(cpu_to_node(i), &sd->span);
7431 cpus_and(sd->span, sd->span, *cpu_map);
7435 sd = &per_cpu(phys_domains, i);
7437 set_domain_attribute(sd, attr);
7438 sd->span = *nodemask;
7442 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7444 #ifdef CONFIG_SCHED_MC
7446 sd = &per_cpu(core_domains, i);
7448 set_domain_attribute(sd, attr);
7449 sd->span = cpu_coregroup_map(i);
7450 cpus_and(sd->span, sd->span, *cpu_map);
7453 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7456 #ifdef CONFIG_SCHED_SMT
7458 sd = &per_cpu(cpu_domains, i);
7459 SD_INIT(sd, SIBLING);
7460 set_domain_attribute(sd, attr);
7461 sd->span = per_cpu(cpu_sibling_map, i);
7462 cpus_and(sd->span, sd->span, *cpu_map);
7465 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7469 #ifdef CONFIG_SCHED_SMT
7470 /* Set up CPU (sibling) groups */
7471 for_each_cpu_mask_nr(i, *cpu_map) {
7472 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7473 SCHED_CPUMASK_VAR(send_covered, allmasks);
7475 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7476 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7477 if (i != first_cpu(*this_sibling_map))
7480 init_sched_build_groups(this_sibling_map, cpu_map,
7482 send_covered, tmpmask);
7486 #ifdef CONFIG_SCHED_MC
7487 /* Set up multi-core groups */
7488 for_each_cpu_mask_nr(i, *cpu_map) {
7489 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7490 SCHED_CPUMASK_VAR(send_covered, allmasks);
7492 *this_core_map = cpu_coregroup_map(i);
7493 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7494 if (i != first_cpu(*this_core_map))
7497 init_sched_build_groups(this_core_map, cpu_map,
7499 send_covered, tmpmask);
7503 /* Set up physical groups */
7504 for (i = 0; i < nr_node_ids; i++) {
7505 SCHED_CPUMASK_VAR(nodemask, allmasks);
7506 SCHED_CPUMASK_VAR(send_covered, allmasks);
7508 *nodemask = node_to_cpumask(i);
7509 cpus_and(*nodemask, *nodemask, *cpu_map);
7510 if (cpus_empty(*nodemask))
7513 init_sched_build_groups(nodemask, cpu_map,
7515 send_covered, tmpmask);
7519 /* Set up node groups */
7521 SCHED_CPUMASK_VAR(send_covered, allmasks);
7523 init_sched_build_groups(cpu_map, cpu_map,
7524 &cpu_to_allnodes_group,
7525 send_covered, tmpmask);
7528 for (i = 0; i < nr_node_ids; i++) {
7529 /* Set up node groups */
7530 struct sched_group *sg, *prev;
7531 SCHED_CPUMASK_VAR(nodemask, allmasks);
7532 SCHED_CPUMASK_VAR(domainspan, allmasks);
7533 SCHED_CPUMASK_VAR(covered, allmasks);
7536 *nodemask = node_to_cpumask(i);
7537 cpus_clear(*covered);
7539 cpus_and(*nodemask, *nodemask, *cpu_map);
7540 if (cpus_empty(*nodemask)) {
7541 sched_group_nodes[i] = NULL;
7545 sched_domain_node_span(i, domainspan);
7546 cpus_and(*domainspan, *domainspan, *cpu_map);
7548 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7550 printk(KERN_WARNING "Can not alloc domain group for "
7554 sched_group_nodes[i] = sg;
7555 for_each_cpu_mask_nr(j, *nodemask) {
7556 struct sched_domain *sd;
7558 sd = &per_cpu(node_domains, j);
7561 sg->__cpu_power = 0;
7562 sg->cpumask = *nodemask;
7564 cpus_or(*covered, *covered, *nodemask);
7567 for (j = 0; j < nr_node_ids; j++) {
7568 SCHED_CPUMASK_VAR(notcovered, allmasks);
7569 int n = (i + j) % nr_node_ids;
7570 node_to_cpumask_ptr(pnodemask, n);
7572 cpus_complement(*notcovered, *covered);
7573 cpus_and(*tmpmask, *notcovered, *cpu_map);
7574 cpus_and(*tmpmask, *tmpmask, *domainspan);
7575 if (cpus_empty(*tmpmask))
7578 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7579 if (cpus_empty(*tmpmask))
7582 sg = kmalloc_node(sizeof(struct sched_group),
7586 "Can not alloc domain group for node %d\n", j);
7589 sg->__cpu_power = 0;
7590 sg->cpumask = *tmpmask;
7591 sg->next = prev->next;
7592 cpus_or(*covered, *covered, *tmpmask);
7599 /* Calculate CPU power for physical packages and nodes */
7600 #ifdef CONFIG_SCHED_SMT
7601 for_each_cpu_mask_nr(i, *cpu_map) {
7602 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7604 init_sched_groups_power(i, sd);
7607 #ifdef CONFIG_SCHED_MC
7608 for_each_cpu_mask_nr(i, *cpu_map) {
7609 struct sched_domain *sd = &per_cpu(core_domains, i);
7611 init_sched_groups_power(i, sd);
7615 for_each_cpu_mask_nr(i, *cpu_map) {
7616 struct sched_domain *sd = &per_cpu(phys_domains, i);
7618 init_sched_groups_power(i, sd);
7622 for (i = 0; i < nr_node_ids; i++)
7623 init_numa_sched_groups_power(sched_group_nodes[i]);
7626 struct sched_group *sg;
7628 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7630 init_numa_sched_groups_power(sg);
7634 /* Attach the domains */
7635 for_each_cpu_mask_nr(i, *cpu_map) {
7636 struct sched_domain *sd;
7637 #ifdef CONFIG_SCHED_SMT
7638 sd = &per_cpu(cpu_domains, i);
7639 #elif defined(CONFIG_SCHED_MC)
7640 sd = &per_cpu(core_domains, i);
7642 sd = &per_cpu(phys_domains, i);
7644 cpu_attach_domain(sd, rd, i);
7647 sched_cpumask_free(allmasks);
7652 free_sched_groups(cpu_map, tmpmask);
7653 sched_cpumask_free(allmasks);
7659 static int build_sched_domains(const cpumask_t *cpu_map)
7661 return __build_sched_domains(cpu_map, NULL);
7664 static cpumask_t *doms_cur; /* current sched domains */
7665 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7666 static struct sched_domain_attr *dattr_cur;
7667 /* attribues of custom domains in 'doms_cur' */
7670 * Special case: If a kmalloc of a doms_cur partition (array of
7671 * cpumask_t) fails, then fallback to a single sched domain,
7672 * as determined by the single cpumask_t fallback_doms.
7674 static cpumask_t fallback_doms;
7676 void __attribute__((weak)) arch_update_cpu_topology(void)
7681 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7682 * For now this just excludes isolated cpus, but could be used to
7683 * exclude other special cases in the future.
7685 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7689 arch_update_cpu_topology();
7691 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7693 doms_cur = &fallback_doms;
7694 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7696 err = build_sched_domains(doms_cur);
7697 register_sched_domain_sysctl();
7702 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7705 free_sched_groups(cpu_map, tmpmask);
7709 * Detach sched domains from a group of cpus specified in cpu_map
7710 * These cpus will now be attached to the NULL domain
7712 static void detach_destroy_domains(const cpumask_t *cpu_map)
7717 for_each_cpu_mask_nr(i, *cpu_map)
7718 cpu_attach_domain(NULL, &def_root_domain, i);
7719 synchronize_sched();
7720 arch_destroy_sched_domains(cpu_map, &tmpmask);
7723 /* handle null as "default" */
7724 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7725 struct sched_domain_attr *new, int idx_new)
7727 struct sched_domain_attr tmp;
7734 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7735 new ? (new + idx_new) : &tmp,
7736 sizeof(struct sched_domain_attr));
7740 * Partition sched domains as specified by the 'ndoms_new'
7741 * cpumasks in the array doms_new[] of cpumasks. This compares
7742 * doms_new[] to the current sched domain partitioning, doms_cur[].
7743 * It destroys each deleted domain and builds each new domain.
7745 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7746 * The masks don't intersect (don't overlap.) We should setup one
7747 * sched domain for each mask. CPUs not in any of the cpumasks will
7748 * not be load balanced. If the same cpumask appears both in the
7749 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7752 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7753 * ownership of it and will kfree it when done with it. If the caller
7754 * failed the kmalloc call, then it can pass in doms_new == NULL &&
7755 * ndoms_new == 1, and partition_sched_domains() will fallback to
7756 * the single partition 'fallback_doms', it also forces the domains
7759 * If doms_new == NULL it will be replaced with cpu_online_map.
7760 * ndoms_new == 0 is a special case for destroying existing domains,
7761 * and it will not create the default domain.
7763 * Call with hotplug lock held
7765 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7766 struct sched_domain_attr *dattr_new)
7770 mutex_lock(&sched_domains_mutex);
7772 /* always unregister in case we don't destroy any domains */
7773 unregister_sched_domain_sysctl();
7775 n = doms_new ? ndoms_new : 0;
7777 /* Destroy deleted domains */
7778 for (i = 0; i < ndoms_cur; i++) {
7779 for (j = 0; j < n; j++) {
7780 if (cpus_equal(doms_cur[i], doms_new[j])
7781 && dattrs_equal(dattr_cur, i, dattr_new, j))
7784 /* no match - a current sched domain not in new doms_new[] */
7785 detach_destroy_domains(doms_cur + i);
7790 if (doms_new == NULL) {
7792 doms_new = &fallback_doms;
7793 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7794 WARN_ON_ONCE(dattr_new);
7797 /* Build new domains */
7798 for (i = 0; i < ndoms_new; i++) {
7799 for (j = 0; j < ndoms_cur; j++) {
7800 if (cpus_equal(doms_new[i], doms_cur[j])
7801 && dattrs_equal(dattr_new, i, dattr_cur, j))
7804 /* no match - add a new doms_new */
7805 __build_sched_domains(doms_new + i,
7806 dattr_new ? dattr_new + i : NULL);
7811 /* Remember the new sched domains */
7812 if (doms_cur != &fallback_doms)
7814 kfree(dattr_cur); /* kfree(NULL) is safe */
7815 doms_cur = doms_new;
7816 dattr_cur = dattr_new;
7817 ndoms_cur = ndoms_new;
7819 register_sched_domain_sysctl();
7821 mutex_unlock(&sched_domains_mutex);
7824 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7825 int arch_reinit_sched_domains(void)
7829 /* Destroy domains first to force the rebuild */
7830 partition_sched_domains(0, NULL, NULL);
7832 rebuild_sched_domains();
7838 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7842 if (buf[0] != '0' && buf[0] != '1')
7846 sched_smt_power_savings = (buf[0] == '1');
7848 sched_mc_power_savings = (buf[0] == '1');
7850 ret = arch_reinit_sched_domains();
7852 return ret ? ret : count;
7855 #ifdef CONFIG_SCHED_MC
7856 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7859 return sprintf(page, "%u\n", sched_mc_power_savings);
7861 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7862 const char *buf, size_t count)
7864 return sched_power_savings_store(buf, count, 0);
7866 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7867 sched_mc_power_savings_show,
7868 sched_mc_power_savings_store);
7871 #ifdef CONFIG_SCHED_SMT
7872 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7875 return sprintf(page, "%u\n", sched_smt_power_savings);
7877 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7878 const char *buf, size_t count)
7880 return sched_power_savings_store(buf, count, 1);
7882 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7883 sched_smt_power_savings_show,
7884 sched_smt_power_savings_store);
7887 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7891 #ifdef CONFIG_SCHED_SMT
7893 err = sysfs_create_file(&cls->kset.kobj,
7894 &attr_sched_smt_power_savings.attr);
7896 #ifdef CONFIG_SCHED_MC
7897 if (!err && mc_capable())
7898 err = sysfs_create_file(&cls->kset.kobj,
7899 &attr_sched_mc_power_savings.attr);
7903 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7905 #ifndef CONFIG_CPUSETS
7907 * Add online and remove offline CPUs from the scheduler domains.
7908 * When cpusets are enabled they take over this function.
7910 static int update_sched_domains(struct notifier_block *nfb,
7911 unsigned long action, void *hcpu)
7915 case CPU_ONLINE_FROZEN:
7917 case CPU_DEAD_FROZEN:
7918 partition_sched_domains(1, NULL, NULL);
7927 static int update_runtime(struct notifier_block *nfb,
7928 unsigned long action, void *hcpu)
7930 int cpu = (int)(long)hcpu;
7933 case CPU_DOWN_PREPARE:
7934 case CPU_DOWN_PREPARE_FROZEN:
7935 disable_runtime(cpu_rq(cpu));
7938 case CPU_DOWN_FAILED:
7939 case CPU_DOWN_FAILED_FROZEN:
7941 case CPU_ONLINE_FROZEN:
7942 enable_runtime(cpu_rq(cpu));
7950 void __init sched_init_smp(void)
7952 cpumask_t non_isolated_cpus;
7954 #if defined(CONFIG_NUMA)
7955 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7957 BUG_ON(sched_group_nodes_bycpu == NULL);
7960 mutex_lock(&sched_domains_mutex);
7961 arch_init_sched_domains(&cpu_online_map);
7962 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7963 if (cpus_empty(non_isolated_cpus))
7964 cpu_set(smp_processor_id(), non_isolated_cpus);
7965 mutex_unlock(&sched_domains_mutex);
7968 #ifndef CONFIG_CPUSETS
7969 /* XXX: Theoretical race here - CPU may be hotplugged now */
7970 hotcpu_notifier(update_sched_domains, 0);
7973 /* RT runtime code needs to handle some hotplug events */
7974 hotcpu_notifier(update_runtime, 0);
7978 /* Move init over to a non-isolated CPU */
7979 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7981 sched_init_granularity();
7984 void __init sched_init_smp(void)
7986 sched_init_granularity();
7988 #endif /* CONFIG_SMP */
7990 int in_sched_functions(unsigned long addr)
7992 return in_lock_functions(addr) ||
7993 (addr >= (unsigned long)__sched_text_start
7994 && addr < (unsigned long)__sched_text_end);
7997 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7999 cfs_rq->tasks_timeline = RB_ROOT;
8000 INIT_LIST_HEAD(&cfs_rq->tasks);
8001 #ifdef CONFIG_FAIR_GROUP_SCHED
8004 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8007 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8009 struct rt_prio_array *array;
8012 array = &rt_rq->active;
8013 for (i = 0; i < MAX_RT_PRIO; i++) {
8014 INIT_LIST_HEAD(array->queue + i);
8015 __clear_bit(i, array->bitmap);
8017 /* delimiter for bitsearch: */
8018 __set_bit(MAX_RT_PRIO, array->bitmap);
8020 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8021 rt_rq->highest_prio = MAX_RT_PRIO;
8024 rt_rq->rt_nr_migratory = 0;
8025 rt_rq->overloaded = 0;
8029 rt_rq->rt_throttled = 0;
8030 rt_rq->rt_runtime = 0;
8031 spin_lock_init(&rt_rq->rt_runtime_lock);
8033 #ifdef CONFIG_RT_GROUP_SCHED
8034 rt_rq->rt_nr_boosted = 0;
8039 #ifdef CONFIG_FAIR_GROUP_SCHED
8040 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8041 struct sched_entity *se, int cpu, int add,
8042 struct sched_entity *parent)
8044 struct rq *rq = cpu_rq(cpu);
8045 tg->cfs_rq[cpu] = cfs_rq;
8046 init_cfs_rq(cfs_rq, rq);
8049 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8052 /* se could be NULL for init_task_group */
8057 se->cfs_rq = &rq->cfs;
8059 se->cfs_rq = parent->my_q;
8062 se->load.weight = tg->shares;
8063 se->load.inv_weight = 0;
8064 se->parent = parent;
8068 #ifdef CONFIG_RT_GROUP_SCHED
8069 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8070 struct sched_rt_entity *rt_se, int cpu, int add,
8071 struct sched_rt_entity *parent)
8073 struct rq *rq = cpu_rq(cpu);
8075 tg->rt_rq[cpu] = rt_rq;
8076 init_rt_rq(rt_rq, rq);
8078 rt_rq->rt_se = rt_se;
8079 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8081 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8083 tg->rt_se[cpu] = rt_se;
8088 rt_se->rt_rq = &rq->rt;
8090 rt_se->rt_rq = parent->my_q;
8092 rt_se->my_q = rt_rq;
8093 rt_se->parent = parent;
8094 INIT_LIST_HEAD(&rt_se->run_list);
8098 void __init sched_init(void)
8101 unsigned long alloc_size = 0, ptr;
8103 #ifdef CONFIG_FAIR_GROUP_SCHED
8104 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8106 #ifdef CONFIG_RT_GROUP_SCHED
8107 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8109 #ifdef CONFIG_USER_SCHED
8113 * As sched_init() is called before page_alloc is setup,
8114 * we use alloc_bootmem().
8117 ptr = (unsigned long)alloc_bootmem(alloc_size);
8119 #ifdef CONFIG_FAIR_GROUP_SCHED
8120 init_task_group.se = (struct sched_entity **)ptr;
8121 ptr += nr_cpu_ids * sizeof(void **);
8123 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8124 ptr += nr_cpu_ids * sizeof(void **);
8126 #ifdef CONFIG_USER_SCHED
8127 root_task_group.se = (struct sched_entity **)ptr;
8128 ptr += nr_cpu_ids * sizeof(void **);
8130 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8131 ptr += nr_cpu_ids * sizeof(void **);
8132 #endif /* CONFIG_USER_SCHED */
8133 #endif /* CONFIG_FAIR_GROUP_SCHED */
8134 #ifdef CONFIG_RT_GROUP_SCHED
8135 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8136 ptr += nr_cpu_ids * sizeof(void **);
8138 init_task_group.rt_rq = (struct rt_rq **)ptr;
8139 ptr += nr_cpu_ids * sizeof(void **);
8141 #ifdef CONFIG_USER_SCHED
8142 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8143 ptr += nr_cpu_ids * sizeof(void **);
8145 root_task_group.rt_rq = (struct rt_rq **)ptr;
8146 ptr += nr_cpu_ids * sizeof(void **);
8147 #endif /* CONFIG_USER_SCHED */
8148 #endif /* CONFIG_RT_GROUP_SCHED */
8152 init_defrootdomain();
8155 init_rt_bandwidth(&def_rt_bandwidth,
8156 global_rt_period(), global_rt_runtime());
8158 #ifdef CONFIG_RT_GROUP_SCHED
8159 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8160 global_rt_period(), global_rt_runtime());
8161 #ifdef CONFIG_USER_SCHED
8162 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8163 global_rt_period(), RUNTIME_INF);
8164 #endif /* CONFIG_USER_SCHED */
8165 #endif /* CONFIG_RT_GROUP_SCHED */
8167 #ifdef CONFIG_GROUP_SCHED
8168 list_add(&init_task_group.list, &task_groups);
8169 INIT_LIST_HEAD(&init_task_group.children);
8171 #ifdef CONFIG_USER_SCHED
8172 INIT_LIST_HEAD(&root_task_group.children);
8173 init_task_group.parent = &root_task_group;
8174 list_add(&init_task_group.siblings, &root_task_group.children);
8175 #endif /* CONFIG_USER_SCHED */
8176 #endif /* CONFIG_GROUP_SCHED */
8178 for_each_possible_cpu(i) {
8182 spin_lock_init(&rq->lock);
8184 init_cfs_rq(&rq->cfs, rq);
8185 init_rt_rq(&rq->rt, rq);
8186 #ifdef CONFIG_FAIR_GROUP_SCHED
8187 init_task_group.shares = init_task_group_load;
8188 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8189 #ifdef CONFIG_CGROUP_SCHED
8191 * How much cpu bandwidth does init_task_group get?
8193 * In case of task-groups formed thr' the cgroup filesystem, it
8194 * gets 100% of the cpu resources in the system. This overall
8195 * system cpu resource is divided among the tasks of
8196 * init_task_group and its child task-groups in a fair manner,
8197 * based on each entity's (task or task-group's) weight
8198 * (se->load.weight).
8200 * In other words, if init_task_group has 10 tasks of weight
8201 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8202 * then A0's share of the cpu resource is:
8204 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8206 * We achieve this by letting init_task_group's tasks sit
8207 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8209 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8210 #elif defined CONFIG_USER_SCHED
8211 root_task_group.shares = NICE_0_LOAD;
8212 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8214 * In case of task-groups formed thr' the user id of tasks,
8215 * init_task_group represents tasks belonging to root user.
8216 * Hence it forms a sibling of all subsequent groups formed.
8217 * In this case, init_task_group gets only a fraction of overall
8218 * system cpu resource, based on the weight assigned to root
8219 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8220 * by letting tasks of init_task_group sit in a separate cfs_rq
8221 * (init_cfs_rq) and having one entity represent this group of
8222 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8224 init_tg_cfs_entry(&init_task_group,
8225 &per_cpu(init_cfs_rq, i),
8226 &per_cpu(init_sched_entity, i), i, 1,
8227 root_task_group.se[i]);
8230 #endif /* CONFIG_FAIR_GROUP_SCHED */
8232 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8233 #ifdef CONFIG_RT_GROUP_SCHED
8234 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8235 #ifdef CONFIG_CGROUP_SCHED
8236 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8237 #elif defined CONFIG_USER_SCHED
8238 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8239 init_tg_rt_entry(&init_task_group,
8240 &per_cpu(init_rt_rq, i),
8241 &per_cpu(init_sched_rt_entity, i), i, 1,
8242 root_task_group.rt_se[i]);
8246 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8247 rq->cpu_load[j] = 0;
8251 rq->active_balance = 0;
8252 rq->next_balance = jiffies;
8256 rq->migration_thread = NULL;
8257 INIT_LIST_HEAD(&rq->migration_queue);
8258 rq_attach_root(rq, &def_root_domain);
8261 atomic_set(&rq->nr_iowait, 0);
8264 set_load_weight(&init_task);
8266 #ifdef CONFIG_PREEMPT_NOTIFIERS
8267 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8271 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8274 #ifdef CONFIG_RT_MUTEXES
8275 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8279 * The boot idle thread does lazy MMU switching as well:
8281 atomic_inc(&init_mm.mm_count);
8282 enter_lazy_tlb(&init_mm, current);
8285 * Make us the idle thread. Technically, schedule() should not be
8286 * called from this thread, however somewhere below it might be,
8287 * but because we are the idle thread, we just pick up running again
8288 * when this runqueue becomes "idle".
8290 init_idle(current, smp_processor_id());
8292 * During early bootup we pretend to be a normal task:
8294 current->sched_class = &fair_sched_class;
8296 scheduler_running = 1;
8299 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8300 void __might_sleep(char *file, int line)
8303 static unsigned long prev_jiffy; /* ratelimiting */
8305 if ((!in_atomic() && !irqs_disabled()) ||
8306 system_state != SYSTEM_RUNNING || oops_in_progress)
8308 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8310 prev_jiffy = jiffies;
8313 "BUG: sleeping function called from invalid context at %s:%d\n",
8316 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8317 in_atomic(), irqs_disabled(),
8318 current->pid, current->comm);
8320 debug_show_held_locks(current);
8321 if (irqs_disabled())
8322 print_irqtrace_events(current);
8326 EXPORT_SYMBOL(__might_sleep);
8329 #ifdef CONFIG_MAGIC_SYSRQ
8330 static void normalize_task(struct rq *rq, struct task_struct *p)
8334 update_rq_clock(rq);
8335 on_rq = p->se.on_rq;
8337 deactivate_task(rq, p, 0);
8338 __setscheduler(rq, p, SCHED_NORMAL, 0);
8340 activate_task(rq, p, 0);
8341 resched_task(rq->curr);
8345 void normalize_rt_tasks(void)
8347 struct task_struct *g, *p;
8348 unsigned long flags;
8351 read_lock_irqsave(&tasklist_lock, flags);
8352 do_each_thread(g, p) {
8354 * Only normalize user tasks:
8359 p->se.exec_start = 0;
8360 #ifdef CONFIG_SCHEDSTATS
8361 p->se.wait_start = 0;
8362 p->se.sleep_start = 0;
8363 p->se.block_start = 0;
8368 * Renice negative nice level userspace
8371 if (TASK_NICE(p) < 0 && p->mm)
8372 set_user_nice(p, 0);
8376 spin_lock(&p->pi_lock);
8377 rq = __task_rq_lock(p);
8379 normalize_task(rq, p);
8381 __task_rq_unlock(rq);
8382 spin_unlock(&p->pi_lock);
8383 } while_each_thread(g, p);
8385 read_unlock_irqrestore(&tasklist_lock, flags);
8388 #endif /* CONFIG_MAGIC_SYSRQ */
8392 * These functions are only useful for the IA64 MCA handling.
8394 * They can only be called when the whole system has been
8395 * stopped - every CPU needs to be quiescent, and no scheduling
8396 * activity can take place. Using them for anything else would
8397 * be a serious bug, and as a result, they aren't even visible
8398 * under any other configuration.
8402 * curr_task - return the current task for a given cpu.
8403 * @cpu: the processor in question.
8405 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8407 struct task_struct *curr_task(int cpu)
8409 return cpu_curr(cpu);
8413 * set_curr_task - set the current task for a given cpu.
8414 * @cpu: the processor in question.
8415 * @p: the task pointer to set.
8417 * Description: This function must only be used when non-maskable interrupts
8418 * are serviced on a separate stack. It allows the architecture to switch the
8419 * notion of the current task on a cpu in a non-blocking manner. This function
8420 * must be called with all CPU's synchronized, and interrupts disabled, the
8421 * and caller must save the original value of the current task (see
8422 * curr_task() above) and restore that value before reenabling interrupts and
8423 * re-starting the system.
8425 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8427 void set_curr_task(int cpu, struct task_struct *p)
8434 #ifdef CONFIG_FAIR_GROUP_SCHED
8435 static void free_fair_sched_group(struct task_group *tg)
8439 for_each_possible_cpu(i) {
8441 kfree(tg->cfs_rq[i]);
8451 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8453 struct cfs_rq *cfs_rq;
8454 struct sched_entity *se;
8458 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8461 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8465 tg->shares = NICE_0_LOAD;
8467 for_each_possible_cpu(i) {
8470 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8471 GFP_KERNEL, cpu_to_node(i));
8475 se = kzalloc_node(sizeof(struct sched_entity),
8476 GFP_KERNEL, cpu_to_node(i));
8480 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8489 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8491 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8492 &cpu_rq(cpu)->leaf_cfs_rq_list);
8495 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8497 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8499 #else /* !CONFG_FAIR_GROUP_SCHED */
8500 static inline void free_fair_sched_group(struct task_group *tg)
8505 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8510 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8514 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8517 #endif /* CONFIG_FAIR_GROUP_SCHED */
8519 #ifdef CONFIG_RT_GROUP_SCHED
8520 static void free_rt_sched_group(struct task_group *tg)
8524 destroy_rt_bandwidth(&tg->rt_bandwidth);
8526 for_each_possible_cpu(i) {
8528 kfree(tg->rt_rq[i]);
8530 kfree(tg->rt_se[i]);
8538 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8540 struct rt_rq *rt_rq;
8541 struct sched_rt_entity *rt_se;
8545 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8548 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8552 init_rt_bandwidth(&tg->rt_bandwidth,
8553 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8555 for_each_possible_cpu(i) {
8558 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8559 GFP_KERNEL, cpu_to_node(i));
8563 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8564 GFP_KERNEL, cpu_to_node(i));
8568 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8577 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8579 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8580 &cpu_rq(cpu)->leaf_rt_rq_list);
8583 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8585 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8587 #else /* !CONFIG_RT_GROUP_SCHED */
8588 static inline void free_rt_sched_group(struct task_group *tg)
8593 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8598 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8602 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8605 #endif /* CONFIG_RT_GROUP_SCHED */
8607 #ifdef CONFIG_GROUP_SCHED
8608 static void free_sched_group(struct task_group *tg)
8610 free_fair_sched_group(tg);
8611 free_rt_sched_group(tg);
8615 /* allocate runqueue etc for a new task group */
8616 struct task_group *sched_create_group(struct task_group *parent)
8618 struct task_group *tg;
8619 unsigned long flags;
8622 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8624 return ERR_PTR(-ENOMEM);
8626 if (!alloc_fair_sched_group(tg, parent))
8629 if (!alloc_rt_sched_group(tg, parent))
8632 spin_lock_irqsave(&task_group_lock, flags);
8633 for_each_possible_cpu(i) {
8634 register_fair_sched_group(tg, i);
8635 register_rt_sched_group(tg, i);
8637 list_add_rcu(&tg->list, &task_groups);
8639 WARN_ON(!parent); /* root should already exist */
8641 tg->parent = parent;
8642 INIT_LIST_HEAD(&tg->children);
8643 list_add_rcu(&tg->siblings, &parent->children);
8644 spin_unlock_irqrestore(&task_group_lock, flags);
8649 free_sched_group(tg);
8650 return ERR_PTR(-ENOMEM);
8653 /* rcu callback to free various structures associated with a task group */
8654 static void free_sched_group_rcu(struct rcu_head *rhp)
8656 /* now it should be safe to free those cfs_rqs */
8657 free_sched_group(container_of(rhp, struct task_group, rcu));
8660 /* Destroy runqueue etc associated with a task group */
8661 void sched_destroy_group(struct task_group *tg)
8663 unsigned long flags;
8666 spin_lock_irqsave(&task_group_lock, flags);
8667 for_each_possible_cpu(i) {
8668 unregister_fair_sched_group(tg, i);
8669 unregister_rt_sched_group(tg, i);
8671 list_del_rcu(&tg->list);
8672 list_del_rcu(&tg->siblings);
8673 spin_unlock_irqrestore(&task_group_lock, flags);
8675 /* wait for possible concurrent references to cfs_rqs complete */
8676 call_rcu(&tg->rcu, free_sched_group_rcu);
8679 /* change task's runqueue when it moves between groups.
8680 * The caller of this function should have put the task in its new group
8681 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8682 * reflect its new group.
8684 void sched_move_task(struct task_struct *tsk)
8687 unsigned long flags;
8690 rq = task_rq_lock(tsk, &flags);
8692 update_rq_clock(rq);
8694 running = task_current(rq, tsk);
8695 on_rq = tsk->se.on_rq;
8698 dequeue_task(rq, tsk, 0);
8699 if (unlikely(running))
8700 tsk->sched_class->put_prev_task(rq, tsk);
8702 set_task_rq(tsk, task_cpu(tsk));
8704 #ifdef CONFIG_FAIR_GROUP_SCHED
8705 if (tsk->sched_class->moved_group)
8706 tsk->sched_class->moved_group(tsk);
8709 if (unlikely(running))
8710 tsk->sched_class->set_curr_task(rq);
8712 enqueue_task(rq, tsk, 0);
8714 task_rq_unlock(rq, &flags);
8716 #endif /* CONFIG_GROUP_SCHED */
8718 #ifdef CONFIG_FAIR_GROUP_SCHED
8719 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8721 struct cfs_rq *cfs_rq = se->cfs_rq;
8726 dequeue_entity(cfs_rq, se, 0);
8728 se->load.weight = shares;
8729 se->load.inv_weight = 0;
8732 enqueue_entity(cfs_rq, se, 0);
8735 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8737 struct cfs_rq *cfs_rq = se->cfs_rq;
8738 struct rq *rq = cfs_rq->rq;
8739 unsigned long flags;
8741 spin_lock_irqsave(&rq->lock, flags);
8742 __set_se_shares(se, shares);
8743 spin_unlock_irqrestore(&rq->lock, flags);
8746 static DEFINE_MUTEX(shares_mutex);
8748 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8751 unsigned long flags;
8754 * We can't change the weight of the root cgroup.
8759 if (shares < MIN_SHARES)
8760 shares = MIN_SHARES;
8761 else if (shares > MAX_SHARES)
8762 shares = MAX_SHARES;
8764 mutex_lock(&shares_mutex);
8765 if (tg->shares == shares)
8768 spin_lock_irqsave(&task_group_lock, flags);
8769 for_each_possible_cpu(i)
8770 unregister_fair_sched_group(tg, i);
8771 list_del_rcu(&tg->siblings);
8772 spin_unlock_irqrestore(&task_group_lock, flags);
8774 /* wait for any ongoing reference to this group to finish */
8775 synchronize_sched();
8778 * Now we are free to modify the group's share on each cpu
8779 * w/o tripping rebalance_share or load_balance_fair.
8781 tg->shares = shares;
8782 for_each_possible_cpu(i) {
8786 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8787 set_se_shares(tg->se[i], shares);
8791 * Enable load balance activity on this group, by inserting it back on
8792 * each cpu's rq->leaf_cfs_rq_list.
8794 spin_lock_irqsave(&task_group_lock, flags);
8795 for_each_possible_cpu(i)
8796 register_fair_sched_group(tg, i);
8797 list_add_rcu(&tg->siblings, &tg->parent->children);
8798 spin_unlock_irqrestore(&task_group_lock, flags);
8800 mutex_unlock(&shares_mutex);
8804 unsigned long sched_group_shares(struct task_group *tg)
8810 #ifdef CONFIG_RT_GROUP_SCHED
8812 * Ensure that the real time constraints are schedulable.
8814 static DEFINE_MUTEX(rt_constraints_mutex);
8816 static unsigned long to_ratio(u64 period, u64 runtime)
8818 if (runtime == RUNTIME_INF)
8821 return div64_u64(runtime << 20, period);
8824 /* Must be called with tasklist_lock held */
8825 static inline int tg_has_rt_tasks(struct task_group *tg)
8827 struct task_struct *g, *p;
8829 do_each_thread(g, p) {
8830 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8832 } while_each_thread(g, p);
8837 struct rt_schedulable_data {
8838 struct task_group *tg;
8843 static int tg_schedulable(struct task_group *tg, void *data)
8845 struct rt_schedulable_data *d = data;
8846 struct task_group *child;
8847 unsigned long total, sum = 0;
8848 u64 period, runtime;
8850 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8851 runtime = tg->rt_bandwidth.rt_runtime;
8854 period = d->rt_period;
8855 runtime = d->rt_runtime;
8859 * Cannot have more runtime than the period.
8861 if (runtime > period && runtime != RUNTIME_INF)
8865 * Ensure we don't starve existing RT tasks.
8867 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8870 total = to_ratio(period, runtime);
8873 * Nobody can have more than the global setting allows.
8875 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8879 * The sum of our children's runtime should not exceed our own.
8881 list_for_each_entry_rcu(child, &tg->children, siblings) {
8882 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8883 runtime = child->rt_bandwidth.rt_runtime;
8885 if (child == d->tg) {
8886 period = d->rt_period;
8887 runtime = d->rt_runtime;
8890 sum += to_ratio(period, runtime);
8899 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8901 struct rt_schedulable_data data = {
8903 .rt_period = period,
8904 .rt_runtime = runtime,
8907 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8910 static int tg_set_bandwidth(struct task_group *tg,
8911 u64 rt_period, u64 rt_runtime)
8915 mutex_lock(&rt_constraints_mutex);
8916 read_lock(&tasklist_lock);
8917 err = __rt_schedulable(tg, rt_period, rt_runtime);
8921 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8922 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8923 tg->rt_bandwidth.rt_runtime = rt_runtime;
8925 for_each_possible_cpu(i) {
8926 struct rt_rq *rt_rq = tg->rt_rq[i];
8928 spin_lock(&rt_rq->rt_runtime_lock);
8929 rt_rq->rt_runtime = rt_runtime;
8930 spin_unlock(&rt_rq->rt_runtime_lock);
8932 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8934 read_unlock(&tasklist_lock);
8935 mutex_unlock(&rt_constraints_mutex);
8940 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8942 u64 rt_runtime, rt_period;
8944 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8945 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8946 if (rt_runtime_us < 0)
8947 rt_runtime = RUNTIME_INF;
8949 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8952 long sched_group_rt_runtime(struct task_group *tg)
8956 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8959 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8960 do_div(rt_runtime_us, NSEC_PER_USEC);
8961 return rt_runtime_us;
8964 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8966 u64 rt_runtime, rt_period;
8968 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8969 rt_runtime = tg->rt_bandwidth.rt_runtime;
8974 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8977 long sched_group_rt_period(struct task_group *tg)
8981 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8982 do_div(rt_period_us, NSEC_PER_USEC);
8983 return rt_period_us;
8986 static int sched_rt_global_constraints(void)
8988 u64 runtime, period;
8991 if (sysctl_sched_rt_period <= 0)
8994 runtime = global_rt_runtime();
8995 period = global_rt_period();
8998 * Sanity check on the sysctl variables.
9000 if (runtime > period && runtime != RUNTIME_INF)
9003 mutex_lock(&rt_constraints_mutex);
9004 read_lock(&tasklist_lock);
9005 ret = __rt_schedulable(NULL, 0, 0);
9006 read_unlock(&tasklist_lock);
9007 mutex_unlock(&rt_constraints_mutex);
9011 #else /* !CONFIG_RT_GROUP_SCHED */
9012 static int sched_rt_global_constraints(void)
9014 unsigned long flags;
9017 if (sysctl_sched_rt_period <= 0)
9020 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9021 for_each_possible_cpu(i) {
9022 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9024 spin_lock(&rt_rq->rt_runtime_lock);
9025 rt_rq->rt_runtime = global_rt_runtime();
9026 spin_unlock(&rt_rq->rt_runtime_lock);
9028 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9032 #endif /* CONFIG_RT_GROUP_SCHED */
9034 int sched_rt_handler(struct ctl_table *table, int write,
9035 struct file *filp, void __user *buffer, size_t *lenp,
9039 int old_period, old_runtime;
9040 static DEFINE_MUTEX(mutex);
9043 old_period = sysctl_sched_rt_period;
9044 old_runtime = sysctl_sched_rt_runtime;
9046 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
9048 if (!ret && write) {
9049 ret = sched_rt_global_constraints();
9051 sysctl_sched_rt_period = old_period;
9052 sysctl_sched_rt_runtime = old_runtime;
9054 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9055 def_rt_bandwidth.rt_period =
9056 ns_to_ktime(global_rt_period());
9059 mutex_unlock(&mutex);
9064 #ifdef CONFIG_CGROUP_SCHED
9066 /* return corresponding task_group object of a cgroup */
9067 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9069 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9070 struct task_group, css);
9073 static struct cgroup_subsys_state *
9074 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9076 struct task_group *tg, *parent;
9078 if (!cgrp->parent) {
9079 /* This is early initialization for the top cgroup */
9080 return &init_task_group.css;
9083 parent = cgroup_tg(cgrp->parent);
9084 tg = sched_create_group(parent);
9086 return ERR_PTR(-ENOMEM);
9092 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9094 struct task_group *tg = cgroup_tg(cgrp);
9096 sched_destroy_group(tg);
9100 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9101 struct task_struct *tsk)
9103 #ifdef CONFIG_RT_GROUP_SCHED
9104 /* Don't accept realtime tasks when there is no way for them to run */
9105 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
9108 /* We don't support RT-tasks being in separate groups */
9109 if (tsk->sched_class != &fair_sched_class)
9117 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9118 struct cgroup *old_cont, struct task_struct *tsk)
9120 sched_move_task(tsk);
9123 #ifdef CONFIG_FAIR_GROUP_SCHED
9124 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9127 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9130 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9132 struct task_group *tg = cgroup_tg(cgrp);
9134 return (u64) tg->shares;
9136 #endif /* CONFIG_FAIR_GROUP_SCHED */
9138 #ifdef CONFIG_RT_GROUP_SCHED
9139 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9142 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9145 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9147 return sched_group_rt_runtime(cgroup_tg(cgrp));
9150 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9153 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9156 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9158 return sched_group_rt_period(cgroup_tg(cgrp));
9160 #endif /* CONFIG_RT_GROUP_SCHED */
9162 static struct cftype cpu_files[] = {
9163 #ifdef CONFIG_FAIR_GROUP_SCHED
9166 .read_u64 = cpu_shares_read_u64,
9167 .write_u64 = cpu_shares_write_u64,
9170 #ifdef CONFIG_RT_GROUP_SCHED
9172 .name = "rt_runtime_us",
9173 .read_s64 = cpu_rt_runtime_read,
9174 .write_s64 = cpu_rt_runtime_write,
9177 .name = "rt_period_us",
9178 .read_u64 = cpu_rt_period_read_uint,
9179 .write_u64 = cpu_rt_period_write_uint,
9184 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9186 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9189 struct cgroup_subsys cpu_cgroup_subsys = {
9191 .create = cpu_cgroup_create,
9192 .destroy = cpu_cgroup_destroy,
9193 .can_attach = cpu_cgroup_can_attach,
9194 .attach = cpu_cgroup_attach,
9195 .populate = cpu_cgroup_populate,
9196 .subsys_id = cpu_cgroup_subsys_id,
9200 #endif /* CONFIG_CGROUP_SCHED */
9202 #ifdef CONFIG_CGROUP_CPUACCT
9205 * CPU accounting code for task groups.
9207 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9208 * (balbir@in.ibm.com).
9211 /* track cpu usage of a group of tasks and its child groups */
9213 struct cgroup_subsys_state css;
9214 /* cpuusage holds pointer to a u64-type object on every cpu */
9216 struct cpuacct *parent;
9219 struct cgroup_subsys cpuacct_subsys;
9221 /* return cpu accounting group corresponding to this container */
9222 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9224 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9225 struct cpuacct, css);
9228 /* return cpu accounting group to which this task belongs */
9229 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9231 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9232 struct cpuacct, css);
9235 /* create a new cpu accounting group */
9236 static struct cgroup_subsys_state *cpuacct_create(
9237 struct cgroup_subsys *ss, struct cgroup *cgrp)
9239 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9242 return ERR_PTR(-ENOMEM);
9244 ca->cpuusage = alloc_percpu(u64);
9245 if (!ca->cpuusage) {
9247 return ERR_PTR(-ENOMEM);
9251 ca->parent = cgroup_ca(cgrp->parent);
9256 /* destroy an existing cpu accounting group */
9258 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9260 struct cpuacct *ca = cgroup_ca(cgrp);
9262 free_percpu(ca->cpuusage);
9266 /* return total cpu usage (in nanoseconds) of a group */
9267 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9269 struct cpuacct *ca = cgroup_ca(cgrp);
9270 u64 totalcpuusage = 0;
9273 for_each_possible_cpu(i) {
9274 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9277 * Take rq->lock to make 64-bit addition safe on 32-bit
9280 spin_lock_irq(&cpu_rq(i)->lock);
9281 totalcpuusage += *cpuusage;
9282 spin_unlock_irq(&cpu_rq(i)->lock);
9285 return totalcpuusage;
9288 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9291 struct cpuacct *ca = cgroup_ca(cgrp);
9300 for_each_possible_cpu(i) {
9301 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9303 spin_lock_irq(&cpu_rq(i)->lock);
9305 spin_unlock_irq(&cpu_rq(i)->lock);
9311 static struct cftype files[] = {
9314 .read_u64 = cpuusage_read,
9315 .write_u64 = cpuusage_write,
9319 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9321 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9325 * charge this task's execution time to its accounting group.
9327 * called with rq->lock held.
9329 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9334 if (!cpuacct_subsys.active)
9337 cpu = task_cpu(tsk);
9340 for (; ca; ca = ca->parent) {
9341 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9342 *cpuusage += cputime;
9346 struct cgroup_subsys cpuacct_subsys = {
9348 .create = cpuacct_create,
9349 .destroy = cpuacct_destroy,
9350 .populate = cpuacct_populate,
9351 .subsys_id = cpuacct_subsys_id,
9353 #endif /* CONFIG_CGROUP_CPUACCT */