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 tg = __task_cred(p)->user->tg;
351 #elif defined(CONFIG_CGROUP_SCHED)
352 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
353 struct task_group, css);
355 tg = &init_task_group;
360 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
361 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
363 #ifdef CONFIG_FAIR_GROUP_SCHED
364 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
365 p->se.parent = task_group(p)->se[cpu];
368 #ifdef CONFIG_RT_GROUP_SCHED
369 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
370 p->rt.parent = task_group(p)->rt_se[cpu];
376 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
377 static inline struct task_group *task_group(struct task_struct *p)
382 #endif /* CONFIG_GROUP_SCHED */
384 /* CFS-related fields in a runqueue */
386 struct load_weight load;
387 unsigned long nr_running;
392 struct rb_root tasks_timeline;
393 struct rb_node *rb_leftmost;
395 struct list_head tasks;
396 struct list_head *balance_iterator;
399 * 'curr' points to currently running entity on this cfs_rq.
400 * It is set to NULL otherwise (i.e when none are currently running).
402 struct sched_entity *curr, *next, *last;
404 unsigned int nr_spread_over;
406 #ifdef CONFIG_FAIR_GROUP_SCHED
407 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
410 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
411 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
412 * (like users, containers etc.)
414 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
415 * list is used during load balance.
417 struct list_head leaf_cfs_rq_list;
418 struct task_group *tg; /* group that "owns" this runqueue */
422 * the part of load.weight contributed by tasks
424 unsigned long task_weight;
427 * h_load = weight * f(tg)
429 * Where f(tg) is the recursive weight fraction assigned to
432 unsigned long h_load;
435 * this cpu's part of tg->shares
437 unsigned long shares;
440 * load.weight at the time we set shares
442 unsigned long rq_weight;
447 /* Real-Time classes' related field in a runqueue: */
449 struct rt_prio_array active;
450 unsigned long rt_nr_running;
451 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
452 int highest_prio; /* highest queued rt task prio */
455 unsigned long rt_nr_migratory;
461 /* Nests inside the rq lock: */
462 spinlock_t rt_runtime_lock;
464 #ifdef CONFIG_RT_GROUP_SCHED
465 unsigned long rt_nr_boosted;
468 struct list_head leaf_rt_rq_list;
469 struct task_group *tg;
470 struct sched_rt_entity *rt_se;
477 * We add the notion of a root-domain which will be used to define per-domain
478 * variables. Each exclusive cpuset essentially defines an island domain by
479 * fully partitioning the member cpus from any other cpuset. Whenever a new
480 * exclusive cpuset is created, we also create and attach a new root-domain
490 * The "RT overload" flag: it gets set if a CPU has more than
491 * one runnable RT task.
496 struct cpupri cpupri;
501 * By default the system creates a single root-domain with all cpus as
502 * members (mimicking the global state we have today).
504 static struct root_domain def_root_domain;
509 * This is the main, per-CPU runqueue data structure.
511 * Locking rule: those places that want to lock multiple runqueues
512 * (such as the load balancing or the thread migration code), lock
513 * acquire operations must be ordered by ascending &runqueue.
520 * nr_running and cpu_load should be in the same cacheline because
521 * remote CPUs use both these fields when doing load calculation.
523 unsigned long nr_running;
524 #define CPU_LOAD_IDX_MAX 5
525 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
526 unsigned char idle_at_tick;
528 unsigned long last_tick_seen;
529 unsigned char in_nohz_recently;
531 /* capture load from *all* tasks on this cpu: */
532 struct load_weight load;
533 unsigned long nr_load_updates;
539 #ifdef CONFIG_FAIR_GROUP_SCHED
540 /* list of leaf cfs_rq on this cpu: */
541 struct list_head leaf_cfs_rq_list;
543 #ifdef CONFIG_RT_GROUP_SCHED
544 struct list_head leaf_rt_rq_list;
548 * This is part of a global counter where only the total sum
549 * over all CPUs matters. A task can increase this counter on
550 * one CPU and if it got migrated afterwards it may decrease
551 * it on another CPU. Always updated under the runqueue lock:
553 unsigned long nr_uninterruptible;
555 struct task_struct *curr, *idle;
556 unsigned long next_balance;
557 struct mm_struct *prev_mm;
564 struct root_domain *rd;
565 struct sched_domain *sd;
567 /* For active balancing */
570 /* cpu of this runqueue: */
574 unsigned long avg_load_per_task;
576 struct task_struct *migration_thread;
577 struct list_head migration_queue;
580 #ifdef CONFIG_SCHED_HRTICK
582 int hrtick_csd_pending;
583 struct call_single_data hrtick_csd;
585 struct hrtimer hrtick_timer;
588 #ifdef CONFIG_SCHEDSTATS
590 struct sched_info rq_sched_info;
592 /* sys_sched_yield() stats */
593 unsigned int yld_exp_empty;
594 unsigned int yld_act_empty;
595 unsigned int yld_both_empty;
596 unsigned int yld_count;
598 /* schedule() stats */
599 unsigned int sched_switch;
600 unsigned int sched_count;
601 unsigned int sched_goidle;
603 /* try_to_wake_up() stats */
604 unsigned int ttwu_count;
605 unsigned int ttwu_local;
608 unsigned int bkl_count;
612 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
614 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
616 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
619 static inline int cpu_of(struct rq *rq)
629 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
630 * See detach_destroy_domains: synchronize_sched for details.
632 * The domain tree of any CPU may only be accessed from within
633 * preempt-disabled sections.
635 #define for_each_domain(cpu, __sd) \
636 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
638 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
639 #define this_rq() (&__get_cpu_var(runqueues))
640 #define task_rq(p) cpu_rq(task_cpu(p))
641 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
643 static inline void update_rq_clock(struct rq *rq)
645 rq->clock = sched_clock_cpu(cpu_of(rq));
649 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
651 #ifdef CONFIG_SCHED_DEBUG
652 # define const_debug __read_mostly
654 # define const_debug static const
660 * Returns true if the current cpu runqueue is locked.
661 * This interface allows printk to be called with the runqueue lock
662 * held and know whether or not it is OK to wake up the klogd.
664 int runqueue_is_locked(void)
667 struct rq *rq = cpu_rq(cpu);
670 ret = spin_is_locked(&rq->lock);
676 * Debugging: various feature bits
679 #define SCHED_FEAT(name, enabled) \
680 __SCHED_FEAT_##name ,
683 #include "sched_features.h"
688 #define SCHED_FEAT(name, enabled) \
689 (1UL << __SCHED_FEAT_##name) * enabled |
691 const_debug unsigned int sysctl_sched_features =
692 #include "sched_features.h"
697 #ifdef CONFIG_SCHED_DEBUG
698 #define SCHED_FEAT(name, enabled) \
701 static __read_mostly char *sched_feat_names[] = {
702 #include "sched_features.h"
708 static int sched_feat_open(struct inode *inode, struct file *filp)
710 filp->private_data = inode->i_private;
715 sched_feat_read(struct file *filp, char __user *ubuf,
716 size_t cnt, loff_t *ppos)
723 for (i = 0; sched_feat_names[i]; i++) {
724 len += strlen(sched_feat_names[i]);
728 buf = kmalloc(len + 2, GFP_KERNEL);
732 for (i = 0; sched_feat_names[i]; i++) {
733 if (sysctl_sched_features & (1UL << i))
734 r += sprintf(buf + r, "%s ", sched_feat_names[i]);
736 r += sprintf(buf + r, "NO_%s ", sched_feat_names[i]);
739 r += sprintf(buf + r, "\n");
740 WARN_ON(r >= len + 2);
742 r = simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
750 sched_feat_write(struct file *filp, const char __user *ubuf,
751 size_t cnt, loff_t *ppos)
761 if (copy_from_user(&buf, ubuf, cnt))
766 if (strncmp(buf, "NO_", 3) == 0) {
771 for (i = 0; sched_feat_names[i]; i++) {
772 int len = strlen(sched_feat_names[i]);
774 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
776 sysctl_sched_features &= ~(1UL << i);
778 sysctl_sched_features |= (1UL << i);
783 if (!sched_feat_names[i])
791 static struct file_operations sched_feat_fops = {
792 .open = sched_feat_open,
793 .read = sched_feat_read,
794 .write = sched_feat_write,
797 static __init int sched_init_debug(void)
799 debugfs_create_file("sched_features", 0644, NULL, NULL,
804 late_initcall(sched_init_debug);
808 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
811 * Number of tasks to iterate in a single balance run.
812 * Limited because this is done with IRQs disabled.
814 const_debug unsigned int sysctl_sched_nr_migrate = 32;
817 * ratelimit for updating the group shares.
820 unsigned int sysctl_sched_shares_ratelimit = 250000;
823 * Inject some fuzzyness into changing the per-cpu group shares
824 * this avoids remote rq-locks at the expense of fairness.
827 unsigned int sysctl_sched_shares_thresh = 4;
830 * period over which we measure -rt task cpu usage in us.
833 unsigned int sysctl_sched_rt_period = 1000000;
835 static __read_mostly int scheduler_running;
838 * part of the period that we allow rt tasks to run in us.
841 int sysctl_sched_rt_runtime = 950000;
843 static inline u64 global_rt_period(void)
845 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
848 static inline u64 global_rt_runtime(void)
850 if (sysctl_sched_rt_runtime < 0)
853 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
856 #ifndef prepare_arch_switch
857 # define prepare_arch_switch(next) do { } while (0)
859 #ifndef finish_arch_switch
860 # define finish_arch_switch(prev) do { } while (0)
863 static inline int task_current(struct rq *rq, struct task_struct *p)
865 return rq->curr == p;
868 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
869 static inline int task_running(struct rq *rq, struct task_struct *p)
871 return task_current(rq, p);
874 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
878 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
880 #ifdef CONFIG_DEBUG_SPINLOCK
881 /* this is a valid case when another task releases the spinlock */
882 rq->lock.owner = current;
885 * If we are tracking spinlock dependencies then we have to
886 * fix up the runqueue lock - which gets 'carried over' from
889 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
891 spin_unlock_irq(&rq->lock);
894 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
895 static inline int task_running(struct rq *rq, struct task_struct *p)
900 return task_current(rq, p);
904 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
908 * We can optimise this out completely for !SMP, because the
909 * SMP rebalancing from interrupt is the only thing that cares
914 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
915 spin_unlock_irq(&rq->lock);
917 spin_unlock(&rq->lock);
921 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
925 * After ->oncpu is cleared, the task can be moved to a different CPU.
926 * We must ensure this doesn't happen until the switch is completely
932 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
936 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
939 * __task_rq_lock - lock the runqueue a given task resides on.
940 * Must be called interrupts disabled.
942 static inline struct rq *__task_rq_lock(struct task_struct *p)
946 struct rq *rq = task_rq(p);
947 spin_lock(&rq->lock);
948 if (likely(rq == task_rq(p)))
950 spin_unlock(&rq->lock);
955 * task_rq_lock - lock the runqueue a given task resides on and disable
956 * interrupts. Note the ordering: we can safely lookup the task_rq without
957 * explicitly disabling preemption.
959 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
965 local_irq_save(*flags);
967 spin_lock(&rq->lock);
968 if (likely(rq == task_rq(p)))
970 spin_unlock_irqrestore(&rq->lock, *flags);
974 void task_rq_unlock_wait(struct task_struct *p)
976 struct rq *rq = task_rq(p);
978 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
979 spin_unlock_wait(&rq->lock);
982 static void __task_rq_unlock(struct rq *rq)
985 spin_unlock(&rq->lock);
988 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
991 spin_unlock_irqrestore(&rq->lock, *flags);
995 * this_rq_lock - lock this runqueue and disable interrupts.
997 static struct rq *this_rq_lock(void)
1002 local_irq_disable();
1004 spin_lock(&rq->lock);
1009 #ifdef CONFIG_SCHED_HRTICK
1011 * Use HR-timers to deliver accurate preemption points.
1013 * Its all a bit involved since we cannot program an hrt while holding the
1014 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1017 * When we get rescheduled we reprogram the hrtick_timer outside of the
1023 * - enabled by features
1024 * - hrtimer is actually high res
1026 static inline int hrtick_enabled(struct rq *rq)
1028 if (!sched_feat(HRTICK))
1030 if (!cpu_active(cpu_of(rq)))
1032 return hrtimer_is_hres_active(&rq->hrtick_timer);
1035 static void hrtick_clear(struct rq *rq)
1037 if (hrtimer_active(&rq->hrtick_timer))
1038 hrtimer_cancel(&rq->hrtick_timer);
1042 * High-resolution timer tick.
1043 * Runs from hardirq context with interrupts disabled.
1045 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1047 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1049 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1051 spin_lock(&rq->lock);
1052 update_rq_clock(rq);
1053 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1054 spin_unlock(&rq->lock);
1056 return HRTIMER_NORESTART;
1061 * called from hardirq (IPI) context
1063 static void __hrtick_start(void *arg)
1065 struct rq *rq = arg;
1067 spin_lock(&rq->lock);
1068 hrtimer_restart(&rq->hrtick_timer);
1069 rq->hrtick_csd_pending = 0;
1070 spin_unlock(&rq->lock);
1074 * Called to set the hrtick timer state.
1076 * called with rq->lock held and irqs disabled
1078 static void hrtick_start(struct rq *rq, u64 delay)
1080 struct hrtimer *timer = &rq->hrtick_timer;
1081 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1083 hrtimer_set_expires(timer, time);
1085 if (rq == this_rq()) {
1086 hrtimer_restart(timer);
1087 } else if (!rq->hrtick_csd_pending) {
1088 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd);
1089 rq->hrtick_csd_pending = 1;
1094 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1096 int cpu = (int)(long)hcpu;
1099 case CPU_UP_CANCELED:
1100 case CPU_UP_CANCELED_FROZEN:
1101 case CPU_DOWN_PREPARE:
1102 case CPU_DOWN_PREPARE_FROZEN:
1104 case CPU_DEAD_FROZEN:
1105 hrtick_clear(cpu_rq(cpu));
1112 static __init void init_hrtick(void)
1114 hotcpu_notifier(hotplug_hrtick, 0);
1118 * Called to set the hrtick timer state.
1120 * called with rq->lock held and irqs disabled
1122 static void hrtick_start(struct rq *rq, u64 delay)
1124 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay), HRTIMER_MODE_REL);
1127 static inline void init_hrtick(void)
1130 #endif /* CONFIG_SMP */
1132 static void init_rq_hrtick(struct rq *rq)
1135 rq->hrtick_csd_pending = 0;
1137 rq->hrtick_csd.flags = 0;
1138 rq->hrtick_csd.func = __hrtick_start;
1139 rq->hrtick_csd.info = rq;
1142 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1143 rq->hrtick_timer.function = hrtick;
1144 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_PERCPU;
1146 #else /* CONFIG_SCHED_HRTICK */
1147 static inline void hrtick_clear(struct rq *rq)
1151 static inline void init_rq_hrtick(struct rq *rq)
1155 static inline void init_hrtick(void)
1158 #endif /* CONFIG_SCHED_HRTICK */
1161 * resched_task - mark a task 'to be rescheduled now'.
1163 * On UP this means the setting of the need_resched flag, on SMP it
1164 * might also involve a cross-CPU call to trigger the scheduler on
1169 #ifndef tsk_is_polling
1170 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1173 static void resched_task(struct task_struct *p)
1177 assert_spin_locked(&task_rq(p)->lock);
1179 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1182 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1185 if (cpu == smp_processor_id())
1188 /* NEED_RESCHED must be visible before we test polling */
1190 if (!tsk_is_polling(p))
1191 smp_send_reschedule(cpu);
1194 static void resched_cpu(int cpu)
1196 struct rq *rq = cpu_rq(cpu);
1197 unsigned long flags;
1199 if (!spin_trylock_irqsave(&rq->lock, flags))
1201 resched_task(cpu_curr(cpu));
1202 spin_unlock_irqrestore(&rq->lock, flags);
1207 * When add_timer_on() enqueues a timer into the timer wheel of an
1208 * idle CPU then this timer might expire before the next timer event
1209 * which is scheduled to wake up that CPU. In case of a completely
1210 * idle system the next event might even be infinite time into the
1211 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1212 * leaves the inner idle loop so the newly added timer is taken into
1213 * account when the CPU goes back to idle and evaluates the timer
1214 * wheel for the next timer event.
1216 void wake_up_idle_cpu(int cpu)
1218 struct rq *rq = cpu_rq(cpu);
1220 if (cpu == smp_processor_id())
1224 * This is safe, as this function is called with the timer
1225 * wheel base lock of (cpu) held. When the CPU is on the way
1226 * to idle and has not yet set rq->curr to idle then it will
1227 * be serialized on the timer wheel base lock and take the new
1228 * timer into account automatically.
1230 if (rq->curr != rq->idle)
1234 * We can set TIF_RESCHED on the idle task of the other CPU
1235 * lockless. The worst case is that the other CPU runs the
1236 * idle task through an additional NOOP schedule()
1238 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1240 /* NEED_RESCHED must be visible before we test polling */
1242 if (!tsk_is_polling(rq->idle))
1243 smp_send_reschedule(cpu);
1245 #endif /* CONFIG_NO_HZ */
1247 #else /* !CONFIG_SMP */
1248 static void resched_task(struct task_struct *p)
1250 assert_spin_locked(&task_rq(p)->lock);
1251 set_tsk_need_resched(p);
1253 #endif /* CONFIG_SMP */
1255 #if BITS_PER_LONG == 32
1256 # define WMULT_CONST (~0UL)
1258 # define WMULT_CONST (1UL << 32)
1261 #define WMULT_SHIFT 32
1264 * Shift right and round:
1266 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1269 * delta *= weight / lw
1271 static unsigned long
1272 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1273 struct load_weight *lw)
1277 if (!lw->inv_weight) {
1278 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1281 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1285 tmp = (u64)delta_exec * weight;
1287 * Check whether we'd overflow the 64-bit multiplication:
1289 if (unlikely(tmp > WMULT_CONST))
1290 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1293 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1295 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1298 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1304 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1311 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1312 * of tasks with abnormal "nice" values across CPUs the contribution that
1313 * each task makes to its run queue's load is weighted according to its
1314 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1315 * scaled version of the new time slice allocation that they receive on time
1319 #define WEIGHT_IDLEPRIO 2
1320 #define WMULT_IDLEPRIO (1 << 31)
1323 * Nice levels are multiplicative, with a gentle 10% change for every
1324 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1325 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1326 * that remained on nice 0.
1328 * The "10% effect" is relative and cumulative: from _any_ nice level,
1329 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1330 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1331 * If a task goes up by ~10% and another task goes down by ~10% then
1332 * the relative distance between them is ~25%.)
1334 static const int prio_to_weight[40] = {
1335 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1336 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1337 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1338 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1339 /* 0 */ 1024, 820, 655, 526, 423,
1340 /* 5 */ 335, 272, 215, 172, 137,
1341 /* 10 */ 110, 87, 70, 56, 45,
1342 /* 15 */ 36, 29, 23, 18, 15,
1346 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1348 * In cases where the weight does not change often, we can use the
1349 * precalculated inverse to speed up arithmetics by turning divisions
1350 * into multiplications:
1352 static const u32 prio_to_wmult[40] = {
1353 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1354 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1355 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1356 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1357 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1358 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1359 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1360 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1363 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1366 * runqueue iterator, to support SMP load-balancing between different
1367 * scheduling classes, without having to expose their internal data
1368 * structures to the load-balancing proper:
1370 struct rq_iterator {
1372 struct task_struct *(*start)(void *);
1373 struct task_struct *(*next)(void *);
1377 static unsigned long
1378 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1379 unsigned long max_load_move, struct sched_domain *sd,
1380 enum cpu_idle_type idle, int *all_pinned,
1381 int *this_best_prio, struct rq_iterator *iterator);
1384 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1385 struct sched_domain *sd, enum cpu_idle_type idle,
1386 struct rq_iterator *iterator);
1389 #ifdef CONFIG_CGROUP_CPUACCT
1390 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1392 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1395 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1397 update_load_add(&rq->load, load);
1400 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1402 update_load_sub(&rq->load, load);
1405 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1406 typedef int (*tg_visitor)(struct task_group *, void *);
1409 * Iterate the full tree, calling @down when first entering a node and @up when
1410 * leaving it for the final time.
1412 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1414 struct task_group *parent, *child;
1418 parent = &root_task_group;
1420 ret = (*down)(parent, data);
1423 list_for_each_entry_rcu(child, &parent->children, siblings) {
1430 ret = (*up)(parent, data);
1435 parent = parent->parent;
1444 static int tg_nop(struct task_group *tg, void *data)
1451 static unsigned long source_load(int cpu, int type);
1452 static unsigned long target_load(int cpu, int type);
1453 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1455 static unsigned long cpu_avg_load_per_task(int cpu)
1457 struct rq *rq = cpu_rq(cpu);
1460 rq->avg_load_per_task = rq->load.weight / rq->nr_running;
1462 rq->avg_load_per_task = 0;
1464 return rq->avg_load_per_task;
1467 #ifdef CONFIG_FAIR_GROUP_SCHED
1469 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1472 * Calculate and set the cpu's group shares.
1475 update_group_shares_cpu(struct task_group *tg, int cpu,
1476 unsigned long sd_shares, unsigned long sd_rq_weight)
1479 unsigned long shares;
1480 unsigned long rq_weight;
1485 rq_weight = tg->cfs_rq[cpu]->load.weight;
1488 * If there are currently no tasks on the cpu pretend there is one of
1489 * average load so that when a new task gets to run here it will not
1490 * get delayed by group starvation.
1494 rq_weight = NICE_0_LOAD;
1497 if (unlikely(rq_weight > sd_rq_weight))
1498 rq_weight = sd_rq_weight;
1501 * \Sum shares * rq_weight
1502 * shares = -----------------------
1506 shares = (sd_shares * rq_weight) / (sd_rq_weight + 1);
1507 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1509 if (abs(shares - tg->se[cpu]->load.weight) >
1510 sysctl_sched_shares_thresh) {
1511 struct rq *rq = cpu_rq(cpu);
1512 unsigned long flags;
1514 spin_lock_irqsave(&rq->lock, flags);
1516 * record the actual number of shares, not the boosted amount.
1518 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1519 tg->cfs_rq[cpu]->rq_weight = rq_weight;
1521 __set_se_shares(tg->se[cpu], shares);
1522 spin_unlock_irqrestore(&rq->lock, flags);
1527 * Re-compute the task group their per cpu shares over the given domain.
1528 * This needs to be done in a bottom-up fashion because the rq weight of a
1529 * parent group depends on the shares of its child groups.
1531 static int tg_shares_up(struct task_group *tg, void *data)
1533 unsigned long rq_weight = 0;
1534 unsigned long shares = 0;
1535 struct sched_domain *sd = data;
1538 for_each_cpu_mask(i, sd->span) {
1539 rq_weight += tg->cfs_rq[i]->load.weight;
1540 shares += tg->cfs_rq[i]->shares;
1543 if ((!shares && rq_weight) || shares > tg->shares)
1544 shares = tg->shares;
1546 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1547 shares = tg->shares;
1550 rq_weight = cpus_weight(sd->span) * NICE_0_LOAD;
1552 for_each_cpu_mask(i, sd->span)
1553 update_group_shares_cpu(tg, i, shares, rq_weight);
1559 * Compute the cpu's hierarchical load factor for each task group.
1560 * This needs to be done in a top-down fashion because the load of a child
1561 * group is a fraction of its parents load.
1563 static int tg_load_down(struct task_group *tg, void *data)
1566 long cpu = (long)data;
1569 load = cpu_rq(cpu)->load.weight;
1571 load = tg->parent->cfs_rq[cpu]->h_load;
1572 load *= tg->cfs_rq[cpu]->shares;
1573 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1576 tg->cfs_rq[cpu]->h_load = load;
1581 static void update_shares(struct sched_domain *sd)
1583 u64 now = cpu_clock(raw_smp_processor_id());
1584 s64 elapsed = now - sd->last_update;
1586 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1587 sd->last_update = now;
1588 walk_tg_tree(tg_nop, tg_shares_up, sd);
1592 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1594 spin_unlock(&rq->lock);
1596 spin_lock(&rq->lock);
1599 static void update_h_load(long cpu)
1601 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1606 static inline void update_shares(struct sched_domain *sd)
1610 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1618 #ifdef CONFIG_FAIR_GROUP_SCHED
1619 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1622 cfs_rq->shares = shares;
1627 #include "sched_stats.h"
1628 #include "sched_idletask.c"
1629 #include "sched_fair.c"
1630 #include "sched_rt.c"
1631 #ifdef CONFIG_SCHED_DEBUG
1632 # include "sched_debug.c"
1635 #define sched_class_highest (&rt_sched_class)
1636 #define for_each_class(class) \
1637 for (class = sched_class_highest; class; class = class->next)
1639 static void inc_nr_running(struct rq *rq)
1644 static void dec_nr_running(struct rq *rq)
1649 static void set_load_weight(struct task_struct *p)
1651 if (task_has_rt_policy(p)) {
1652 p->se.load.weight = prio_to_weight[0] * 2;
1653 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1658 * SCHED_IDLE tasks get minimal weight:
1660 if (p->policy == SCHED_IDLE) {
1661 p->se.load.weight = WEIGHT_IDLEPRIO;
1662 p->se.load.inv_weight = WMULT_IDLEPRIO;
1666 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1667 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1670 static void update_avg(u64 *avg, u64 sample)
1672 s64 diff = sample - *avg;
1676 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1678 sched_info_queued(p);
1679 p->sched_class->enqueue_task(rq, p, wakeup);
1683 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1685 if (sleep && p->se.last_wakeup) {
1686 update_avg(&p->se.avg_overlap,
1687 p->se.sum_exec_runtime - p->se.last_wakeup);
1688 p->se.last_wakeup = 0;
1691 sched_info_dequeued(p);
1692 p->sched_class->dequeue_task(rq, p, sleep);
1697 * __normal_prio - return the priority that is based on the static prio
1699 static inline int __normal_prio(struct task_struct *p)
1701 return p->static_prio;
1705 * Calculate the expected normal priority: i.e. priority
1706 * without taking RT-inheritance into account. Might be
1707 * boosted by interactivity modifiers. Changes upon fork,
1708 * setprio syscalls, and whenever the interactivity
1709 * estimator recalculates.
1711 static inline int normal_prio(struct task_struct *p)
1715 if (task_has_rt_policy(p))
1716 prio = MAX_RT_PRIO-1 - p->rt_priority;
1718 prio = __normal_prio(p);
1723 * Calculate the current priority, i.e. the priority
1724 * taken into account by the scheduler. This value might
1725 * be boosted by RT tasks, or might be boosted by
1726 * interactivity modifiers. Will be RT if the task got
1727 * RT-boosted. If not then it returns p->normal_prio.
1729 static int effective_prio(struct task_struct *p)
1731 p->normal_prio = normal_prio(p);
1733 * If we are RT tasks or we were boosted to RT priority,
1734 * keep the priority unchanged. Otherwise, update priority
1735 * to the normal priority:
1737 if (!rt_prio(p->prio))
1738 return p->normal_prio;
1743 * activate_task - move a task to the runqueue.
1745 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1747 if (task_contributes_to_load(p))
1748 rq->nr_uninterruptible--;
1750 enqueue_task(rq, p, wakeup);
1755 * deactivate_task - remove a task from the runqueue.
1757 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1759 if (task_contributes_to_load(p))
1760 rq->nr_uninterruptible++;
1762 dequeue_task(rq, p, sleep);
1767 * task_curr - is this task currently executing on a CPU?
1768 * @p: the task in question.
1770 inline int task_curr(const struct task_struct *p)
1772 return cpu_curr(task_cpu(p)) == p;
1775 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1777 set_task_rq(p, cpu);
1780 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1781 * successfuly executed on another CPU. We must ensure that updates of
1782 * per-task data have been completed by this moment.
1785 task_thread_info(p)->cpu = cpu;
1789 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1790 const struct sched_class *prev_class,
1791 int oldprio, int running)
1793 if (prev_class != p->sched_class) {
1794 if (prev_class->switched_from)
1795 prev_class->switched_from(rq, p, running);
1796 p->sched_class->switched_to(rq, p, running);
1798 p->sched_class->prio_changed(rq, p, oldprio, running);
1803 /* Used instead of source_load when we know the type == 0 */
1804 static unsigned long weighted_cpuload(const int cpu)
1806 return cpu_rq(cpu)->load.weight;
1810 * Is this task likely cache-hot:
1813 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1818 * Buddy candidates are cache hot:
1820 if (sched_feat(CACHE_HOT_BUDDY) &&
1821 (&p->se == cfs_rq_of(&p->se)->next ||
1822 &p->se == cfs_rq_of(&p->se)->last))
1825 if (p->sched_class != &fair_sched_class)
1828 if (sysctl_sched_migration_cost == -1)
1830 if (sysctl_sched_migration_cost == 0)
1833 delta = now - p->se.exec_start;
1835 return delta < (s64)sysctl_sched_migration_cost;
1839 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1841 int old_cpu = task_cpu(p);
1842 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1843 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1844 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1847 clock_offset = old_rq->clock - new_rq->clock;
1849 #ifdef CONFIG_SCHEDSTATS
1850 if (p->se.wait_start)
1851 p->se.wait_start -= clock_offset;
1852 if (p->se.sleep_start)
1853 p->se.sleep_start -= clock_offset;
1854 if (p->se.block_start)
1855 p->se.block_start -= clock_offset;
1856 if (old_cpu != new_cpu) {
1857 schedstat_inc(p, se.nr_migrations);
1858 if (task_hot(p, old_rq->clock, NULL))
1859 schedstat_inc(p, se.nr_forced2_migrations);
1862 p->se.vruntime -= old_cfsrq->min_vruntime -
1863 new_cfsrq->min_vruntime;
1865 __set_task_cpu(p, new_cpu);
1868 struct migration_req {
1869 struct list_head list;
1871 struct task_struct *task;
1874 struct completion done;
1878 * The task's runqueue lock must be held.
1879 * Returns true if you have to wait for migration thread.
1882 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1884 struct rq *rq = task_rq(p);
1887 * If the task is not on a runqueue (and not running), then
1888 * it is sufficient to simply update the task's cpu field.
1890 if (!p->se.on_rq && !task_running(rq, p)) {
1891 set_task_cpu(p, dest_cpu);
1895 init_completion(&req->done);
1897 req->dest_cpu = dest_cpu;
1898 list_add(&req->list, &rq->migration_queue);
1904 * wait_task_inactive - wait for a thread to unschedule.
1906 * If @match_state is nonzero, it's the @p->state value just checked and
1907 * not expected to change. If it changes, i.e. @p might have woken up,
1908 * then return zero. When we succeed in waiting for @p to be off its CPU,
1909 * we return a positive number (its total switch count). If a second call
1910 * a short while later returns the same number, the caller can be sure that
1911 * @p has remained unscheduled the whole time.
1913 * The caller must ensure that the task *will* unschedule sometime soon,
1914 * else this function might spin for a *long* time. This function can't
1915 * be called with interrupts off, or it may introduce deadlock with
1916 * smp_call_function() if an IPI is sent by the same process we are
1917 * waiting to become inactive.
1919 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1921 unsigned long flags;
1928 * We do the initial early heuristics without holding
1929 * any task-queue locks at all. We'll only try to get
1930 * the runqueue lock when things look like they will
1936 * If the task is actively running on another CPU
1937 * still, just relax and busy-wait without holding
1940 * NOTE! Since we don't hold any locks, it's not
1941 * even sure that "rq" stays as the right runqueue!
1942 * But we don't care, since "task_running()" will
1943 * return false if the runqueue has changed and p
1944 * is actually now running somewhere else!
1946 while (task_running(rq, p)) {
1947 if (match_state && unlikely(p->state != match_state))
1953 * Ok, time to look more closely! We need the rq
1954 * lock now, to be *sure*. If we're wrong, we'll
1955 * just go back and repeat.
1957 rq = task_rq_lock(p, &flags);
1958 trace_sched_wait_task(rq, p);
1959 running = task_running(rq, p);
1960 on_rq = p->se.on_rq;
1962 if (!match_state || p->state == match_state)
1963 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1964 task_rq_unlock(rq, &flags);
1967 * If it changed from the expected state, bail out now.
1969 if (unlikely(!ncsw))
1973 * Was it really running after all now that we
1974 * checked with the proper locks actually held?
1976 * Oops. Go back and try again..
1978 if (unlikely(running)) {
1984 * It's not enough that it's not actively running,
1985 * it must be off the runqueue _entirely_, and not
1988 * So if it wa still runnable (but just not actively
1989 * running right now), it's preempted, and we should
1990 * yield - it could be a while.
1992 if (unlikely(on_rq)) {
1993 schedule_timeout_uninterruptible(1);
1998 * Ahh, all good. It wasn't running, and it wasn't
1999 * runnable, which means that it will never become
2000 * running in the future either. We're all done!
2009 * kick_process - kick a running thread to enter/exit the kernel
2010 * @p: the to-be-kicked thread
2012 * Cause a process which is running on another CPU to enter
2013 * kernel-mode, without any delay. (to get signals handled.)
2015 * NOTE: this function doesnt have to take the runqueue lock,
2016 * because all it wants to ensure is that the remote task enters
2017 * the kernel. If the IPI races and the task has been migrated
2018 * to another CPU then no harm is done and the purpose has been
2021 void kick_process(struct task_struct *p)
2027 if ((cpu != smp_processor_id()) && task_curr(p))
2028 smp_send_reschedule(cpu);
2033 * Return a low guess at the load of a migration-source cpu weighted
2034 * according to the scheduling class and "nice" value.
2036 * We want to under-estimate the load of migration sources, to
2037 * balance conservatively.
2039 static unsigned long source_load(int cpu, int type)
2041 struct rq *rq = cpu_rq(cpu);
2042 unsigned long total = weighted_cpuload(cpu);
2044 if (type == 0 || !sched_feat(LB_BIAS))
2047 return min(rq->cpu_load[type-1], total);
2051 * Return a high guess at the load of a migration-target cpu weighted
2052 * according to the scheduling class and "nice" value.
2054 static unsigned long target_load(int cpu, int type)
2056 struct rq *rq = cpu_rq(cpu);
2057 unsigned long total = weighted_cpuload(cpu);
2059 if (type == 0 || !sched_feat(LB_BIAS))
2062 return max(rq->cpu_load[type-1], total);
2066 * find_idlest_group finds and returns the least busy CPU group within the
2069 static struct sched_group *
2070 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2072 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2073 unsigned long min_load = ULONG_MAX, this_load = 0;
2074 int load_idx = sd->forkexec_idx;
2075 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2078 unsigned long load, avg_load;
2082 /* Skip over this group if it has no CPUs allowed */
2083 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
2086 local_group = cpu_isset(this_cpu, group->cpumask);
2088 /* Tally up the load of all CPUs in the group */
2091 for_each_cpu_mask_nr(i, group->cpumask) {
2092 /* Bias balancing toward cpus of our domain */
2094 load = source_load(i, load_idx);
2096 load = target_load(i, load_idx);
2101 /* Adjust by relative CPU power of the group */
2102 avg_load = sg_div_cpu_power(group,
2103 avg_load * SCHED_LOAD_SCALE);
2106 this_load = avg_load;
2108 } else if (avg_load < min_load) {
2109 min_load = avg_load;
2112 } while (group = group->next, group != sd->groups);
2114 if (!idlest || 100*this_load < imbalance*min_load)
2120 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2123 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2126 unsigned long load, min_load = ULONG_MAX;
2130 /* Traverse only the allowed CPUs */
2131 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
2133 for_each_cpu_mask_nr(i, *tmp) {
2134 load = weighted_cpuload(i);
2136 if (load < min_load || (load == min_load && i == this_cpu)) {
2146 * sched_balance_self: balance the current task (running on cpu) in domains
2147 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2150 * Balance, ie. select the least loaded group.
2152 * Returns the target CPU number, or the same CPU if no balancing is needed.
2154 * preempt must be disabled.
2156 static int sched_balance_self(int cpu, int flag)
2158 struct task_struct *t = current;
2159 struct sched_domain *tmp, *sd = NULL;
2161 for_each_domain(cpu, tmp) {
2163 * If power savings logic is enabled for a domain, stop there.
2165 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2167 if (tmp->flags & flag)
2175 cpumask_t span, tmpmask;
2176 struct sched_group *group;
2177 int new_cpu, weight;
2179 if (!(sd->flags & flag)) {
2185 group = find_idlest_group(sd, t, cpu);
2191 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2192 if (new_cpu == -1 || new_cpu == cpu) {
2193 /* Now try balancing at a lower domain level of cpu */
2198 /* Now try balancing at a lower domain level of new_cpu */
2201 weight = cpus_weight(span);
2202 for_each_domain(cpu, tmp) {
2203 if (weight <= cpus_weight(tmp->span))
2205 if (tmp->flags & flag)
2208 /* while loop will break here if sd == NULL */
2214 #endif /* CONFIG_SMP */
2217 * try_to_wake_up - wake up a thread
2218 * @p: the to-be-woken-up thread
2219 * @state: the mask of task states that can be woken
2220 * @sync: do a synchronous wakeup?
2222 * Put it on the run-queue if it's not already there. The "current"
2223 * thread is always on the run-queue (except when the actual
2224 * re-schedule is in progress), and as such you're allowed to do
2225 * the simpler "current->state = TASK_RUNNING" to mark yourself
2226 * runnable without the overhead of this.
2228 * returns failure only if the task is already active.
2230 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2232 int cpu, orig_cpu, this_cpu, success = 0;
2233 unsigned long flags;
2237 if (!sched_feat(SYNC_WAKEUPS))
2241 if (sched_feat(LB_WAKEUP_UPDATE)) {
2242 struct sched_domain *sd;
2244 this_cpu = raw_smp_processor_id();
2247 for_each_domain(this_cpu, sd) {
2248 if (cpu_isset(cpu, sd->span)) {
2257 rq = task_rq_lock(p, &flags);
2258 old_state = p->state;
2259 if (!(old_state & state))
2267 this_cpu = smp_processor_id();
2270 if (unlikely(task_running(rq, p)))
2273 cpu = p->sched_class->select_task_rq(p, sync);
2274 if (cpu != orig_cpu) {
2275 set_task_cpu(p, cpu);
2276 task_rq_unlock(rq, &flags);
2277 /* might preempt at this point */
2278 rq = task_rq_lock(p, &flags);
2279 old_state = p->state;
2280 if (!(old_state & state))
2285 this_cpu = smp_processor_id();
2289 #ifdef CONFIG_SCHEDSTATS
2290 schedstat_inc(rq, ttwu_count);
2291 if (cpu == this_cpu)
2292 schedstat_inc(rq, ttwu_local);
2294 struct sched_domain *sd;
2295 for_each_domain(this_cpu, sd) {
2296 if (cpu_isset(cpu, sd->span)) {
2297 schedstat_inc(sd, ttwu_wake_remote);
2302 #endif /* CONFIG_SCHEDSTATS */
2305 #endif /* CONFIG_SMP */
2306 schedstat_inc(p, se.nr_wakeups);
2308 schedstat_inc(p, se.nr_wakeups_sync);
2309 if (orig_cpu != cpu)
2310 schedstat_inc(p, se.nr_wakeups_migrate);
2311 if (cpu == this_cpu)
2312 schedstat_inc(p, se.nr_wakeups_local);
2314 schedstat_inc(p, se.nr_wakeups_remote);
2315 update_rq_clock(rq);
2316 activate_task(rq, p, 1);
2320 trace_sched_wakeup(rq, p);
2321 check_preempt_curr(rq, p, sync);
2323 p->state = TASK_RUNNING;
2325 if (p->sched_class->task_wake_up)
2326 p->sched_class->task_wake_up(rq, p);
2329 current->se.last_wakeup = current->se.sum_exec_runtime;
2331 task_rq_unlock(rq, &flags);
2336 int wake_up_process(struct task_struct *p)
2338 return try_to_wake_up(p, TASK_ALL, 0);
2340 EXPORT_SYMBOL(wake_up_process);
2342 int wake_up_state(struct task_struct *p, unsigned int state)
2344 return try_to_wake_up(p, state, 0);
2348 * Perform scheduler related setup for a newly forked process p.
2349 * p is forked by current.
2351 * __sched_fork() is basic setup used by init_idle() too:
2353 static void __sched_fork(struct task_struct *p)
2355 p->se.exec_start = 0;
2356 p->se.sum_exec_runtime = 0;
2357 p->se.prev_sum_exec_runtime = 0;
2358 p->se.last_wakeup = 0;
2359 p->se.avg_overlap = 0;
2361 #ifdef CONFIG_SCHEDSTATS
2362 p->se.wait_start = 0;
2363 p->se.sum_sleep_runtime = 0;
2364 p->se.sleep_start = 0;
2365 p->se.block_start = 0;
2366 p->se.sleep_max = 0;
2367 p->se.block_max = 0;
2369 p->se.slice_max = 0;
2373 INIT_LIST_HEAD(&p->rt.run_list);
2375 INIT_LIST_HEAD(&p->se.group_node);
2377 #ifdef CONFIG_PREEMPT_NOTIFIERS
2378 INIT_HLIST_HEAD(&p->preempt_notifiers);
2382 * We mark the process as running here, but have not actually
2383 * inserted it onto the runqueue yet. This guarantees that
2384 * nobody will actually run it, and a signal or other external
2385 * event cannot wake it up and insert it on the runqueue either.
2387 p->state = TASK_RUNNING;
2391 * fork()/clone()-time setup:
2393 void sched_fork(struct task_struct *p, int clone_flags)
2395 int cpu = get_cpu();
2400 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2402 set_task_cpu(p, cpu);
2405 * Make sure we do not leak PI boosting priority to the child:
2407 p->prio = current->normal_prio;
2408 if (!rt_prio(p->prio))
2409 p->sched_class = &fair_sched_class;
2411 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2412 if (likely(sched_info_on()))
2413 memset(&p->sched_info, 0, sizeof(p->sched_info));
2415 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2418 #ifdef CONFIG_PREEMPT
2419 /* Want to start with kernel preemption disabled. */
2420 task_thread_info(p)->preempt_count = 1;
2426 * wake_up_new_task - wake up a newly created task for the first time.
2428 * This function will do some initial scheduler statistics housekeeping
2429 * that must be done for every newly created context, then puts the task
2430 * on the runqueue and wakes it.
2432 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2434 unsigned long flags;
2437 rq = task_rq_lock(p, &flags);
2438 BUG_ON(p->state != TASK_RUNNING);
2439 update_rq_clock(rq);
2441 p->prio = effective_prio(p);
2443 if (!p->sched_class->task_new || !current->se.on_rq) {
2444 activate_task(rq, p, 0);
2447 * Let the scheduling class do new task startup
2448 * management (if any):
2450 p->sched_class->task_new(rq, p);
2453 trace_sched_wakeup_new(rq, p);
2454 check_preempt_curr(rq, p, 0);
2456 if (p->sched_class->task_wake_up)
2457 p->sched_class->task_wake_up(rq, p);
2459 task_rq_unlock(rq, &flags);
2462 #ifdef CONFIG_PREEMPT_NOTIFIERS
2465 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2466 * @notifier: notifier struct to register
2468 void preempt_notifier_register(struct preempt_notifier *notifier)
2470 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2472 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2475 * preempt_notifier_unregister - no longer interested in preemption notifications
2476 * @notifier: notifier struct to unregister
2478 * This is safe to call from within a preemption notifier.
2480 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2482 hlist_del(¬ifier->link);
2484 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2486 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2488 struct preempt_notifier *notifier;
2489 struct hlist_node *node;
2491 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2492 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2496 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2497 struct task_struct *next)
2499 struct preempt_notifier *notifier;
2500 struct hlist_node *node;
2502 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2503 notifier->ops->sched_out(notifier, next);
2506 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2508 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2513 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2514 struct task_struct *next)
2518 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2521 * prepare_task_switch - prepare to switch tasks
2522 * @rq: the runqueue preparing to switch
2523 * @prev: the current task that is being switched out
2524 * @next: the task we are going to switch to.
2526 * This is called with the rq lock held and interrupts off. It must
2527 * be paired with a subsequent finish_task_switch after the context
2530 * prepare_task_switch sets up locking and calls architecture specific
2534 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2535 struct task_struct *next)
2537 fire_sched_out_preempt_notifiers(prev, next);
2538 prepare_lock_switch(rq, next);
2539 prepare_arch_switch(next);
2543 * finish_task_switch - clean up after a task-switch
2544 * @rq: runqueue associated with task-switch
2545 * @prev: the thread we just switched away from.
2547 * finish_task_switch must be called after the context switch, paired
2548 * with a prepare_task_switch call before the context switch.
2549 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2550 * and do any other architecture-specific cleanup actions.
2552 * Note that we may have delayed dropping an mm in context_switch(). If
2553 * so, we finish that here outside of the runqueue lock. (Doing it
2554 * with the lock held can cause deadlocks; see schedule() for
2557 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2558 __releases(rq->lock)
2560 struct mm_struct *mm = rq->prev_mm;
2566 * A task struct has one reference for the use as "current".
2567 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2568 * schedule one last time. The schedule call will never return, and
2569 * the scheduled task must drop that reference.
2570 * The test for TASK_DEAD must occur while the runqueue locks are
2571 * still held, otherwise prev could be scheduled on another cpu, die
2572 * there before we look at prev->state, and then the reference would
2574 * Manfred Spraul <manfred@colorfullife.com>
2576 prev_state = prev->state;
2577 finish_arch_switch(prev);
2578 finish_lock_switch(rq, prev);
2580 if (current->sched_class->post_schedule)
2581 current->sched_class->post_schedule(rq);
2584 fire_sched_in_preempt_notifiers(current);
2587 if (unlikely(prev_state == TASK_DEAD)) {
2589 * Remove function-return probe instances associated with this
2590 * task and put them back on the free list.
2592 kprobe_flush_task(prev);
2593 put_task_struct(prev);
2598 * schedule_tail - first thing a freshly forked thread must call.
2599 * @prev: the thread we just switched away from.
2601 asmlinkage void schedule_tail(struct task_struct *prev)
2602 __releases(rq->lock)
2604 struct rq *rq = this_rq();
2606 finish_task_switch(rq, prev);
2607 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2608 /* In this case, finish_task_switch does not reenable preemption */
2611 if (current->set_child_tid)
2612 put_user(task_pid_vnr(current), current->set_child_tid);
2616 * context_switch - switch to the new MM and the new
2617 * thread's register state.
2620 context_switch(struct rq *rq, struct task_struct *prev,
2621 struct task_struct *next)
2623 struct mm_struct *mm, *oldmm;
2625 prepare_task_switch(rq, prev, next);
2626 trace_sched_switch(rq, prev, next);
2628 oldmm = prev->active_mm;
2630 * For paravirt, this is coupled with an exit in switch_to to
2631 * combine the page table reload and the switch backend into
2634 arch_enter_lazy_cpu_mode();
2636 if (unlikely(!mm)) {
2637 next->active_mm = oldmm;
2638 atomic_inc(&oldmm->mm_count);
2639 enter_lazy_tlb(oldmm, next);
2641 switch_mm(oldmm, mm, next);
2643 if (unlikely(!prev->mm)) {
2644 prev->active_mm = NULL;
2645 rq->prev_mm = oldmm;
2648 * Since the runqueue lock will be released by the next
2649 * task (which is an invalid locking op but in the case
2650 * of the scheduler it's an obvious special-case), so we
2651 * do an early lockdep release here:
2653 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2654 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2657 /* Here we just switch the register state and the stack. */
2658 switch_to(prev, next, prev);
2662 * this_rq must be evaluated again because prev may have moved
2663 * CPUs since it called schedule(), thus the 'rq' on its stack
2664 * frame will be invalid.
2666 finish_task_switch(this_rq(), prev);
2670 * nr_running, nr_uninterruptible and nr_context_switches:
2672 * externally visible scheduler statistics: current number of runnable
2673 * threads, current number of uninterruptible-sleeping threads, total
2674 * number of context switches performed since bootup.
2676 unsigned long nr_running(void)
2678 unsigned long i, sum = 0;
2680 for_each_online_cpu(i)
2681 sum += cpu_rq(i)->nr_running;
2686 unsigned long nr_uninterruptible(void)
2688 unsigned long i, sum = 0;
2690 for_each_possible_cpu(i)
2691 sum += cpu_rq(i)->nr_uninterruptible;
2694 * Since we read the counters lockless, it might be slightly
2695 * inaccurate. Do not allow it to go below zero though:
2697 if (unlikely((long)sum < 0))
2703 unsigned long long nr_context_switches(void)
2706 unsigned long long sum = 0;
2708 for_each_possible_cpu(i)
2709 sum += cpu_rq(i)->nr_switches;
2714 unsigned long nr_iowait(void)
2716 unsigned long i, sum = 0;
2718 for_each_possible_cpu(i)
2719 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2724 unsigned long nr_active(void)
2726 unsigned long i, running = 0, uninterruptible = 0;
2728 for_each_online_cpu(i) {
2729 running += cpu_rq(i)->nr_running;
2730 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2733 if (unlikely((long)uninterruptible < 0))
2734 uninterruptible = 0;
2736 return running + uninterruptible;
2740 * Update rq->cpu_load[] statistics. This function is usually called every
2741 * scheduler tick (TICK_NSEC).
2743 static void update_cpu_load(struct rq *this_rq)
2745 unsigned long this_load = this_rq->load.weight;
2748 this_rq->nr_load_updates++;
2750 /* Update our load: */
2751 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2752 unsigned long old_load, new_load;
2754 /* scale is effectively 1 << i now, and >> i divides by scale */
2756 old_load = this_rq->cpu_load[i];
2757 new_load = this_load;
2759 * Round up the averaging division if load is increasing. This
2760 * prevents us from getting stuck on 9 if the load is 10, for
2763 if (new_load > old_load)
2764 new_load += scale-1;
2765 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2772 * double_rq_lock - safely lock two runqueues
2774 * Note this does not disable interrupts like task_rq_lock,
2775 * you need to do so manually before calling.
2777 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2778 __acquires(rq1->lock)
2779 __acquires(rq2->lock)
2781 BUG_ON(!irqs_disabled());
2783 spin_lock(&rq1->lock);
2784 __acquire(rq2->lock); /* Fake it out ;) */
2787 spin_lock(&rq1->lock);
2788 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2790 spin_lock(&rq2->lock);
2791 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2794 update_rq_clock(rq1);
2795 update_rq_clock(rq2);
2799 * double_rq_unlock - safely unlock two runqueues
2801 * Note this does not restore interrupts like task_rq_unlock,
2802 * you need to do so manually after calling.
2804 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2805 __releases(rq1->lock)
2806 __releases(rq2->lock)
2808 spin_unlock(&rq1->lock);
2810 spin_unlock(&rq2->lock);
2812 __release(rq2->lock);
2816 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2818 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2819 __releases(this_rq->lock)
2820 __acquires(busiest->lock)
2821 __acquires(this_rq->lock)
2825 if (unlikely(!irqs_disabled())) {
2826 /* printk() doesn't work good under rq->lock */
2827 spin_unlock(&this_rq->lock);
2830 if (unlikely(!spin_trylock(&busiest->lock))) {
2831 if (busiest < this_rq) {
2832 spin_unlock(&this_rq->lock);
2833 spin_lock(&busiest->lock);
2834 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
2837 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
2842 static void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
2843 __releases(busiest->lock)
2845 spin_unlock(&busiest->lock);
2846 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
2850 * If dest_cpu is allowed for this process, migrate the task to it.
2851 * This is accomplished by forcing the cpu_allowed mask to only
2852 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2853 * the cpu_allowed mask is restored.
2855 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2857 struct migration_req req;
2858 unsigned long flags;
2861 rq = task_rq_lock(p, &flags);
2862 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2863 || unlikely(!cpu_active(dest_cpu)))
2866 trace_sched_migrate_task(rq, p, dest_cpu);
2867 /* force the process onto the specified CPU */
2868 if (migrate_task(p, dest_cpu, &req)) {
2869 /* Need to wait for migration thread (might exit: take ref). */
2870 struct task_struct *mt = rq->migration_thread;
2872 get_task_struct(mt);
2873 task_rq_unlock(rq, &flags);
2874 wake_up_process(mt);
2875 put_task_struct(mt);
2876 wait_for_completion(&req.done);
2881 task_rq_unlock(rq, &flags);
2885 * sched_exec - execve() is a valuable balancing opportunity, because at
2886 * this point the task has the smallest effective memory and cache footprint.
2888 void sched_exec(void)
2890 int new_cpu, this_cpu = get_cpu();
2891 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2893 if (new_cpu != this_cpu)
2894 sched_migrate_task(current, new_cpu);
2898 * pull_task - move a task from a remote runqueue to the local runqueue.
2899 * Both runqueues must be locked.
2901 static void pull_task(struct rq *src_rq, struct task_struct *p,
2902 struct rq *this_rq, int this_cpu)
2904 deactivate_task(src_rq, p, 0);
2905 set_task_cpu(p, this_cpu);
2906 activate_task(this_rq, p, 0);
2908 * Note that idle threads have a prio of MAX_PRIO, for this test
2909 * to be always true for them.
2911 check_preempt_curr(this_rq, p, 0);
2915 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2918 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2919 struct sched_domain *sd, enum cpu_idle_type idle,
2923 * We do not migrate tasks that are:
2924 * 1) running (obviously), or
2925 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2926 * 3) are cache-hot on their current CPU.
2928 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2929 schedstat_inc(p, se.nr_failed_migrations_affine);
2934 if (task_running(rq, p)) {
2935 schedstat_inc(p, se.nr_failed_migrations_running);
2940 * Aggressive migration if:
2941 * 1) task is cache cold, or
2942 * 2) too many balance attempts have failed.
2945 if (!task_hot(p, rq->clock, sd) ||
2946 sd->nr_balance_failed > sd->cache_nice_tries) {
2947 #ifdef CONFIG_SCHEDSTATS
2948 if (task_hot(p, rq->clock, sd)) {
2949 schedstat_inc(sd, lb_hot_gained[idle]);
2950 schedstat_inc(p, se.nr_forced_migrations);
2956 if (task_hot(p, rq->clock, sd)) {
2957 schedstat_inc(p, se.nr_failed_migrations_hot);
2963 static unsigned long
2964 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2965 unsigned long max_load_move, struct sched_domain *sd,
2966 enum cpu_idle_type idle, int *all_pinned,
2967 int *this_best_prio, struct rq_iterator *iterator)
2969 int loops = 0, pulled = 0, pinned = 0;
2970 struct task_struct *p;
2971 long rem_load_move = max_load_move;
2973 if (max_load_move == 0)
2979 * Start the load-balancing iterator:
2981 p = iterator->start(iterator->arg);
2983 if (!p || loops++ > sysctl_sched_nr_migrate)
2986 if ((p->se.load.weight >> 1) > rem_load_move ||
2987 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2988 p = iterator->next(iterator->arg);
2992 pull_task(busiest, p, this_rq, this_cpu);
2994 rem_load_move -= p->se.load.weight;
2997 * We only want to steal up to the prescribed amount of weighted load.
2999 if (rem_load_move > 0) {
3000 if (p->prio < *this_best_prio)
3001 *this_best_prio = p->prio;
3002 p = iterator->next(iterator->arg);
3007 * Right now, this is one of only two places pull_task() is called,
3008 * so we can safely collect pull_task() stats here rather than
3009 * inside pull_task().
3011 schedstat_add(sd, lb_gained[idle], pulled);
3014 *all_pinned = pinned;
3016 return max_load_move - rem_load_move;
3020 * move_tasks tries to move up to max_load_move weighted load from busiest to
3021 * this_rq, as part of a balancing operation within domain "sd".
3022 * Returns 1 if successful and 0 otherwise.
3024 * Called with both runqueues locked.
3026 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3027 unsigned long max_load_move,
3028 struct sched_domain *sd, enum cpu_idle_type idle,
3031 const struct sched_class *class = sched_class_highest;
3032 unsigned long total_load_moved = 0;
3033 int this_best_prio = this_rq->curr->prio;
3037 class->load_balance(this_rq, this_cpu, busiest,
3038 max_load_move - total_load_moved,
3039 sd, idle, all_pinned, &this_best_prio);
3040 class = class->next;
3042 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3045 } while (class && max_load_move > total_load_moved);
3047 return total_load_moved > 0;
3051 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3052 struct sched_domain *sd, enum cpu_idle_type idle,
3053 struct rq_iterator *iterator)
3055 struct task_struct *p = iterator->start(iterator->arg);
3059 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3060 pull_task(busiest, p, this_rq, this_cpu);
3062 * Right now, this is only the second place pull_task()
3063 * is called, so we can safely collect pull_task()
3064 * stats here rather than inside pull_task().
3066 schedstat_inc(sd, lb_gained[idle]);
3070 p = iterator->next(iterator->arg);
3077 * move_one_task tries to move exactly one task from busiest to this_rq, as
3078 * part of active balancing operations within "domain".
3079 * Returns 1 if successful and 0 otherwise.
3081 * Called with both runqueues locked.
3083 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3084 struct sched_domain *sd, enum cpu_idle_type idle)
3086 const struct sched_class *class;
3088 for (class = sched_class_highest; class; class = class->next)
3089 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3096 * find_busiest_group finds and returns the busiest CPU group within the
3097 * domain. It calculates and returns the amount of weighted load which
3098 * should be moved to restore balance via the imbalance parameter.
3100 static struct sched_group *
3101 find_busiest_group(struct sched_domain *sd, int this_cpu,
3102 unsigned long *imbalance, enum cpu_idle_type idle,
3103 int *sd_idle, const cpumask_t *cpus, int *balance)
3105 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3106 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3107 unsigned long max_pull;
3108 unsigned long busiest_load_per_task, busiest_nr_running;
3109 unsigned long this_load_per_task, this_nr_running;
3110 int load_idx, group_imb = 0;
3111 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3112 int power_savings_balance = 1;
3113 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3114 unsigned long min_nr_running = ULONG_MAX;
3115 struct sched_group *group_min = NULL, *group_leader = NULL;
3118 max_load = this_load = total_load = total_pwr = 0;
3119 busiest_load_per_task = busiest_nr_running = 0;
3120 this_load_per_task = this_nr_running = 0;
3122 if (idle == CPU_NOT_IDLE)
3123 load_idx = sd->busy_idx;
3124 else if (idle == CPU_NEWLY_IDLE)
3125 load_idx = sd->newidle_idx;
3127 load_idx = sd->idle_idx;
3130 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3133 int __group_imb = 0;
3134 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3135 unsigned long sum_nr_running, sum_weighted_load;
3136 unsigned long sum_avg_load_per_task;
3137 unsigned long avg_load_per_task;
3139 local_group = cpu_isset(this_cpu, group->cpumask);
3142 balance_cpu = first_cpu(group->cpumask);
3144 /* Tally up the load of all CPUs in the group */
3145 sum_weighted_load = sum_nr_running = avg_load = 0;
3146 sum_avg_load_per_task = avg_load_per_task = 0;
3149 min_cpu_load = ~0UL;
3151 for_each_cpu_mask_nr(i, group->cpumask) {
3154 if (!cpu_isset(i, *cpus))
3159 if (*sd_idle && rq->nr_running)
3162 /* Bias balancing toward cpus of our domain */
3164 if (idle_cpu(i) && !first_idle_cpu) {
3169 load = target_load(i, load_idx);
3171 load = source_load(i, load_idx);
3172 if (load > max_cpu_load)
3173 max_cpu_load = load;
3174 if (min_cpu_load > load)
3175 min_cpu_load = load;
3179 sum_nr_running += rq->nr_running;
3180 sum_weighted_load += weighted_cpuload(i);
3182 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3186 * First idle cpu or the first cpu(busiest) in this sched group
3187 * is eligible for doing load balancing at this and above
3188 * domains. In the newly idle case, we will allow all the cpu's
3189 * to do the newly idle load balance.
3191 if (idle != CPU_NEWLY_IDLE && local_group &&
3192 balance_cpu != this_cpu && balance) {
3197 total_load += avg_load;
3198 total_pwr += group->__cpu_power;
3200 /* Adjust by relative CPU power of the group */
3201 avg_load = sg_div_cpu_power(group,
3202 avg_load * SCHED_LOAD_SCALE);
3206 * Consider the group unbalanced when the imbalance is larger
3207 * than the average weight of two tasks.
3209 * APZ: with cgroup the avg task weight can vary wildly and
3210 * might not be a suitable number - should we keep a
3211 * normalized nr_running number somewhere that negates
3214 avg_load_per_task = sg_div_cpu_power(group,
3215 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3217 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3220 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3223 this_load = avg_load;
3225 this_nr_running = sum_nr_running;
3226 this_load_per_task = sum_weighted_load;
3227 } else if (avg_load > max_load &&
3228 (sum_nr_running > group_capacity || __group_imb)) {
3229 max_load = avg_load;
3231 busiest_nr_running = sum_nr_running;
3232 busiest_load_per_task = sum_weighted_load;
3233 group_imb = __group_imb;
3236 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3238 * Busy processors will not participate in power savings
3241 if (idle == CPU_NOT_IDLE ||
3242 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3246 * If the local group is idle or completely loaded
3247 * no need to do power savings balance at this domain
3249 if (local_group && (this_nr_running >= group_capacity ||
3251 power_savings_balance = 0;
3254 * If a group is already running at full capacity or idle,
3255 * don't include that group in power savings calculations
3257 if (!power_savings_balance || sum_nr_running >= group_capacity
3262 * Calculate the group which has the least non-idle load.
3263 * This is the group from where we need to pick up the load
3266 if ((sum_nr_running < min_nr_running) ||
3267 (sum_nr_running == min_nr_running &&
3268 first_cpu(group->cpumask) <
3269 first_cpu(group_min->cpumask))) {
3271 min_nr_running = sum_nr_running;
3272 min_load_per_task = sum_weighted_load /
3277 * Calculate the group which is almost near its
3278 * capacity but still has some space to pick up some load
3279 * from other group and save more power
3281 if (sum_nr_running <= group_capacity - 1) {
3282 if (sum_nr_running > leader_nr_running ||
3283 (sum_nr_running == leader_nr_running &&
3284 first_cpu(group->cpumask) >
3285 first_cpu(group_leader->cpumask))) {
3286 group_leader = group;
3287 leader_nr_running = sum_nr_running;
3292 group = group->next;
3293 } while (group != sd->groups);
3295 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3298 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3300 if (this_load >= avg_load ||
3301 100*max_load <= sd->imbalance_pct*this_load)
3304 busiest_load_per_task /= busiest_nr_running;
3306 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3309 * We're trying to get all the cpus to the average_load, so we don't
3310 * want to push ourselves above the average load, nor do we wish to
3311 * reduce the max loaded cpu below the average load, as either of these
3312 * actions would just result in more rebalancing later, and ping-pong
3313 * tasks around. Thus we look for the minimum possible imbalance.
3314 * Negative imbalances (*we* are more loaded than anyone else) will
3315 * be counted as no imbalance for these purposes -- we can't fix that
3316 * by pulling tasks to us. Be careful of negative numbers as they'll
3317 * appear as very large values with unsigned longs.
3319 if (max_load <= busiest_load_per_task)
3323 * In the presence of smp nice balancing, certain scenarios can have
3324 * max load less than avg load(as we skip the groups at or below
3325 * its cpu_power, while calculating max_load..)
3327 if (max_load < avg_load) {
3329 goto small_imbalance;
3332 /* Don't want to pull so many tasks that a group would go idle */
3333 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3335 /* How much load to actually move to equalise the imbalance */
3336 *imbalance = min(max_pull * busiest->__cpu_power,
3337 (avg_load - this_load) * this->__cpu_power)
3341 * if *imbalance is less than the average load per runnable task
3342 * there is no gaurantee that any tasks will be moved so we'll have
3343 * a think about bumping its value to force at least one task to be
3346 if (*imbalance < busiest_load_per_task) {
3347 unsigned long tmp, pwr_now, pwr_move;
3351 pwr_move = pwr_now = 0;
3353 if (this_nr_running) {
3354 this_load_per_task /= this_nr_running;
3355 if (busiest_load_per_task > this_load_per_task)
3358 this_load_per_task = cpu_avg_load_per_task(this_cpu);
3360 if (max_load - this_load + busiest_load_per_task >=
3361 busiest_load_per_task * imbn) {
3362 *imbalance = busiest_load_per_task;
3367 * OK, we don't have enough imbalance to justify moving tasks,
3368 * however we may be able to increase total CPU power used by
3372 pwr_now += busiest->__cpu_power *
3373 min(busiest_load_per_task, max_load);
3374 pwr_now += this->__cpu_power *
3375 min(this_load_per_task, this_load);
3376 pwr_now /= SCHED_LOAD_SCALE;
3378 /* Amount of load we'd subtract */
3379 tmp = sg_div_cpu_power(busiest,
3380 busiest_load_per_task * SCHED_LOAD_SCALE);
3382 pwr_move += busiest->__cpu_power *
3383 min(busiest_load_per_task, max_load - tmp);
3385 /* Amount of load we'd add */
3386 if (max_load * busiest->__cpu_power <
3387 busiest_load_per_task * SCHED_LOAD_SCALE)
3388 tmp = sg_div_cpu_power(this,
3389 max_load * busiest->__cpu_power);
3391 tmp = sg_div_cpu_power(this,
3392 busiest_load_per_task * SCHED_LOAD_SCALE);
3393 pwr_move += this->__cpu_power *
3394 min(this_load_per_task, this_load + tmp);
3395 pwr_move /= SCHED_LOAD_SCALE;
3397 /* Move if we gain throughput */
3398 if (pwr_move > pwr_now)
3399 *imbalance = busiest_load_per_task;
3405 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3406 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3409 if (this == group_leader && group_leader != group_min) {
3410 *imbalance = min_load_per_task;
3420 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3423 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3424 unsigned long imbalance, const cpumask_t *cpus)
3426 struct rq *busiest = NULL, *rq;
3427 unsigned long max_load = 0;
3430 for_each_cpu_mask_nr(i, group->cpumask) {
3433 if (!cpu_isset(i, *cpus))
3437 wl = weighted_cpuload(i);
3439 if (rq->nr_running == 1 && wl > imbalance)
3442 if (wl > max_load) {
3452 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3453 * so long as it is large enough.
3455 #define MAX_PINNED_INTERVAL 512
3458 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3459 * tasks if there is an imbalance.
3461 static int load_balance(int this_cpu, struct rq *this_rq,
3462 struct sched_domain *sd, enum cpu_idle_type idle,
3463 int *balance, cpumask_t *cpus)
3465 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3466 struct sched_group *group;
3467 unsigned long imbalance;
3469 unsigned long flags;
3474 * When power savings policy is enabled for the parent domain, idle
3475 * sibling can pick up load irrespective of busy siblings. In this case,
3476 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3477 * portraying it as CPU_NOT_IDLE.
3479 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3480 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3483 schedstat_inc(sd, lb_count[idle]);
3487 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3494 schedstat_inc(sd, lb_nobusyg[idle]);
3498 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3500 schedstat_inc(sd, lb_nobusyq[idle]);
3504 BUG_ON(busiest == this_rq);
3506 schedstat_add(sd, lb_imbalance[idle], imbalance);
3509 if (busiest->nr_running > 1) {
3511 * Attempt to move tasks. If find_busiest_group has found
3512 * an imbalance but busiest->nr_running <= 1, the group is
3513 * still unbalanced. ld_moved simply stays zero, so it is
3514 * correctly treated as an imbalance.
3516 local_irq_save(flags);
3517 double_rq_lock(this_rq, busiest);
3518 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3519 imbalance, sd, idle, &all_pinned);
3520 double_rq_unlock(this_rq, busiest);
3521 local_irq_restore(flags);
3524 * some other cpu did the load balance for us.
3526 if (ld_moved && this_cpu != smp_processor_id())
3527 resched_cpu(this_cpu);
3529 /* All tasks on this runqueue were pinned by CPU affinity */
3530 if (unlikely(all_pinned)) {
3531 cpu_clear(cpu_of(busiest), *cpus);
3532 if (!cpus_empty(*cpus))
3539 schedstat_inc(sd, lb_failed[idle]);
3540 sd->nr_balance_failed++;
3542 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3544 spin_lock_irqsave(&busiest->lock, flags);
3546 /* don't kick the migration_thread, if the curr
3547 * task on busiest cpu can't be moved to this_cpu
3549 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3550 spin_unlock_irqrestore(&busiest->lock, flags);
3552 goto out_one_pinned;
3555 if (!busiest->active_balance) {
3556 busiest->active_balance = 1;
3557 busiest->push_cpu = this_cpu;
3560 spin_unlock_irqrestore(&busiest->lock, flags);
3562 wake_up_process(busiest->migration_thread);
3565 * We've kicked active balancing, reset the failure
3568 sd->nr_balance_failed = sd->cache_nice_tries+1;
3571 sd->nr_balance_failed = 0;
3573 if (likely(!active_balance)) {
3574 /* We were unbalanced, so reset the balancing interval */
3575 sd->balance_interval = sd->min_interval;
3578 * If we've begun active balancing, start to back off. This
3579 * case may not be covered by the all_pinned logic if there
3580 * is only 1 task on the busy runqueue (because we don't call
3583 if (sd->balance_interval < sd->max_interval)
3584 sd->balance_interval *= 2;
3587 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3588 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3594 schedstat_inc(sd, lb_balanced[idle]);
3596 sd->nr_balance_failed = 0;
3599 /* tune up the balancing interval */
3600 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3601 (sd->balance_interval < sd->max_interval))
3602 sd->balance_interval *= 2;
3604 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3605 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3616 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3617 * tasks if there is an imbalance.
3619 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3620 * this_rq is locked.
3623 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3626 struct sched_group *group;
3627 struct rq *busiest = NULL;
3628 unsigned long imbalance;
3636 * When power savings policy is enabled for the parent domain, idle
3637 * sibling can pick up load irrespective of busy siblings. In this case,
3638 * let the state of idle sibling percolate up as IDLE, instead of
3639 * portraying it as CPU_NOT_IDLE.
3641 if (sd->flags & SD_SHARE_CPUPOWER &&
3642 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3645 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3647 update_shares_locked(this_rq, sd);
3648 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3649 &sd_idle, cpus, NULL);
3651 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3655 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3657 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3661 BUG_ON(busiest == this_rq);
3663 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3666 if (busiest->nr_running > 1) {
3667 /* Attempt to move tasks */
3668 double_lock_balance(this_rq, busiest);
3669 /* this_rq->clock is already updated */
3670 update_rq_clock(busiest);
3671 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3672 imbalance, sd, CPU_NEWLY_IDLE,
3674 double_unlock_balance(this_rq, busiest);
3676 if (unlikely(all_pinned)) {
3677 cpu_clear(cpu_of(busiest), *cpus);
3678 if (!cpus_empty(*cpus))
3684 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3685 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3686 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3689 sd->nr_balance_failed = 0;
3691 update_shares_locked(this_rq, sd);
3695 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3696 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3697 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3699 sd->nr_balance_failed = 0;
3705 * idle_balance is called by schedule() if this_cpu is about to become
3706 * idle. Attempts to pull tasks from other CPUs.
3708 static void idle_balance(int this_cpu, struct rq *this_rq)
3710 struct sched_domain *sd;
3711 int pulled_task = -1;
3712 unsigned long next_balance = jiffies + HZ;
3715 for_each_domain(this_cpu, sd) {
3716 unsigned long interval;
3718 if (!(sd->flags & SD_LOAD_BALANCE))
3721 if (sd->flags & SD_BALANCE_NEWIDLE)
3722 /* If we've pulled tasks over stop searching: */
3723 pulled_task = load_balance_newidle(this_cpu, this_rq,
3726 interval = msecs_to_jiffies(sd->balance_interval);
3727 if (time_after(next_balance, sd->last_balance + interval))
3728 next_balance = sd->last_balance + interval;
3732 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3734 * We are going idle. next_balance may be set based on
3735 * a busy processor. So reset next_balance.
3737 this_rq->next_balance = next_balance;
3742 * active_load_balance is run by migration threads. It pushes running tasks
3743 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3744 * running on each physical CPU where possible, and avoids physical /
3745 * logical imbalances.
3747 * Called with busiest_rq locked.
3749 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3751 int target_cpu = busiest_rq->push_cpu;
3752 struct sched_domain *sd;
3753 struct rq *target_rq;
3755 /* Is there any task to move? */
3756 if (busiest_rq->nr_running <= 1)
3759 target_rq = cpu_rq(target_cpu);
3762 * This condition is "impossible", if it occurs
3763 * we need to fix it. Originally reported by
3764 * Bjorn Helgaas on a 128-cpu setup.
3766 BUG_ON(busiest_rq == target_rq);
3768 /* move a task from busiest_rq to target_rq */
3769 double_lock_balance(busiest_rq, target_rq);
3770 update_rq_clock(busiest_rq);
3771 update_rq_clock(target_rq);
3773 /* Search for an sd spanning us and the target CPU. */
3774 for_each_domain(target_cpu, sd) {
3775 if ((sd->flags & SD_LOAD_BALANCE) &&
3776 cpu_isset(busiest_cpu, sd->span))
3781 schedstat_inc(sd, alb_count);
3783 if (move_one_task(target_rq, target_cpu, busiest_rq,
3785 schedstat_inc(sd, alb_pushed);
3787 schedstat_inc(sd, alb_failed);
3789 double_unlock_balance(busiest_rq, target_rq);
3794 atomic_t load_balancer;
3796 } nohz ____cacheline_aligned = {
3797 .load_balancer = ATOMIC_INIT(-1),
3798 .cpu_mask = CPU_MASK_NONE,
3802 * This routine will try to nominate the ilb (idle load balancing)
3803 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3804 * load balancing on behalf of all those cpus. If all the cpus in the system
3805 * go into this tickless mode, then there will be no ilb owner (as there is
3806 * no need for one) and all the cpus will sleep till the next wakeup event
3809 * For the ilb owner, tick is not stopped. And this tick will be used
3810 * for idle load balancing. ilb owner will still be part of
3813 * While stopping the tick, this cpu will become the ilb owner if there
3814 * is no other owner. And will be the owner till that cpu becomes busy
3815 * or if all cpus in the system stop their ticks at which point
3816 * there is no need for ilb owner.
3818 * When the ilb owner becomes busy, it nominates another owner, during the
3819 * next busy scheduler_tick()
3821 int select_nohz_load_balancer(int stop_tick)
3823 int cpu = smp_processor_id();
3826 cpu_set(cpu, nohz.cpu_mask);
3827 cpu_rq(cpu)->in_nohz_recently = 1;
3830 * If we are going offline and still the leader, give up!
3832 if (!cpu_active(cpu) &&
3833 atomic_read(&nohz.load_balancer) == cpu) {
3834 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3839 /* time for ilb owner also to sleep */
3840 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3841 if (atomic_read(&nohz.load_balancer) == cpu)
3842 atomic_set(&nohz.load_balancer, -1);
3846 if (atomic_read(&nohz.load_balancer) == -1) {
3847 /* make me the ilb owner */
3848 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3850 } else if (atomic_read(&nohz.load_balancer) == cpu)
3853 if (!cpu_isset(cpu, nohz.cpu_mask))
3856 cpu_clear(cpu, nohz.cpu_mask);
3858 if (atomic_read(&nohz.load_balancer) == cpu)
3859 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3866 static DEFINE_SPINLOCK(balancing);
3869 * It checks each scheduling domain to see if it is due to be balanced,
3870 * and initiates a balancing operation if so.
3872 * Balancing parameters are set up in arch_init_sched_domains.
3874 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3877 struct rq *rq = cpu_rq(cpu);
3878 unsigned long interval;
3879 struct sched_domain *sd;
3880 /* Earliest time when we have to do rebalance again */
3881 unsigned long next_balance = jiffies + 60*HZ;
3882 int update_next_balance = 0;
3886 for_each_domain(cpu, sd) {
3887 if (!(sd->flags & SD_LOAD_BALANCE))
3890 interval = sd->balance_interval;
3891 if (idle != CPU_IDLE)
3892 interval *= sd->busy_factor;
3894 /* scale ms to jiffies */
3895 interval = msecs_to_jiffies(interval);
3896 if (unlikely(!interval))
3898 if (interval > HZ*NR_CPUS/10)
3899 interval = HZ*NR_CPUS/10;
3901 need_serialize = sd->flags & SD_SERIALIZE;
3903 if (need_serialize) {
3904 if (!spin_trylock(&balancing))
3908 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3909 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3911 * We've pulled tasks over so either we're no
3912 * longer idle, or one of our SMT siblings is
3915 idle = CPU_NOT_IDLE;
3917 sd->last_balance = jiffies;
3920 spin_unlock(&balancing);
3922 if (time_after(next_balance, sd->last_balance + interval)) {
3923 next_balance = sd->last_balance + interval;
3924 update_next_balance = 1;
3928 * Stop the load balance at this level. There is another
3929 * CPU in our sched group which is doing load balancing more
3937 * next_balance will be updated only when there is a need.
3938 * When the cpu is attached to null domain for ex, it will not be
3941 if (likely(update_next_balance))
3942 rq->next_balance = next_balance;
3946 * run_rebalance_domains is triggered when needed from the scheduler tick.
3947 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3948 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3950 static void run_rebalance_domains(struct softirq_action *h)
3952 int this_cpu = smp_processor_id();
3953 struct rq *this_rq = cpu_rq(this_cpu);
3954 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3955 CPU_IDLE : CPU_NOT_IDLE;
3957 rebalance_domains(this_cpu, idle);
3961 * If this cpu is the owner for idle load balancing, then do the
3962 * balancing on behalf of the other idle cpus whose ticks are
3965 if (this_rq->idle_at_tick &&
3966 atomic_read(&nohz.load_balancer) == this_cpu) {
3967 cpumask_t cpus = nohz.cpu_mask;
3971 cpu_clear(this_cpu, cpus);
3972 for_each_cpu_mask_nr(balance_cpu, cpus) {
3974 * If this cpu gets work to do, stop the load balancing
3975 * work being done for other cpus. Next load
3976 * balancing owner will pick it up.
3981 rebalance_domains(balance_cpu, CPU_IDLE);
3983 rq = cpu_rq(balance_cpu);
3984 if (time_after(this_rq->next_balance, rq->next_balance))
3985 this_rq->next_balance = rq->next_balance;
3992 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3994 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3995 * idle load balancing owner or decide to stop the periodic load balancing,
3996 * if the whole system is idle.
3998 static inline void trigger_load_balance(struct rq *rq, int cpu)
4002 * If we were in the nohz mode recently and busy at the current
4003 * scheduler tick, then check if we need to nominate new idle
4006 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4007 rq->in_nohz_recently = 0;
4009 if (atomic_read(&nohz.load_balancer) == cpu) {
4010 cpu_clear(cpu, nohz.cpu_mask);
4011 atomic_set(&nohz.load_balancer, -1);
4014 if (atomic_read(&nohz.load_balancer) == -1) {
4016 * simple selection for now: Nominate the
4017 * first cpu in the nohz list to be the next
4020 * TBD: Traverse the sched domains and nominate
4021 * the nearest cpu in the nohz.cpu_mask.
4023 int ilb = first_cpu(nohz.cpu_mask);
4025 if (ilb < nr_cpu_ids)
4031 * If this cpu is idle and doing idle load balancing for all the
4032 * cpus with ticks stopped, is it time for that to stop?
4034 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4035 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4041 * If this cpu is idle and the idle load balancing is done by
4042 * someone else, then no need raise the SCHED_SOFTIRQ
4044 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4045 cpu_isset(cpu, nohz.cpu_mask))
4048 if (time_after_eq(jiffies, rq->next_balance))
4049 raise_softirq(SCHED_SOFTIRQ);
4052 #else /* CONFIG_SMP */
4055 * on UP we do not need to balance between CPUs:
4057 static inline void idle_balance(int cpu, struct rq *rq)
4063 DEFINE_PER_CPU(struct kernel_stat, kstat);
4065 EXPORT_PER_CPU_SYMBOL(kstat);
4068 * Return any ns on the sched_clock that have not yet been banked in
4069 * @p in case that task is currently running.
4071 unsigned long long task_delta_exec(struct task_struct *p)
4073 unsigned long flags;
4077 rq = task_rq_lock(p, &flags);
4079 if (task_current(rq, p)) {
4082 update_rq_clock(rq);
4083 delta_exec = rq->clock - p->se.exec_start;
4084 if ((s64)delta_exec > 0)
4088 task_rq_unlock(rq, &flags);
4094 * Account user cpu time to a process.
4095 * @p: the process that the cpu time gets accounted to
4096 * @cputime: the cpu time spent in user space since the last update
4098 void account_user_time(struct task_struct *p, cputime_t cputime)
4100 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4103 p->utime = cputime_add(p->utime, cputime);
4104 account_group_user_time(p, cputime);
4106 /* Add user time to cpustat. */
4107 tmp = cputime_to_cputime64(cputime);
4108 if (TASK_NICE(p) > 0)
4109 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4111 cpustat->user = cputime64_add(cpustat->user, tmp);
4112 /* Account for user time used */
4113 acct_update_integrals(p);
4117 * Account guest cpu time to a process.
4118 * @p: the process that the cpu time gets accounted to
4119 * @cputime: the cpu time spent in virtual machine since the last update
4121 static void account_guest_time(struct task_struct *p, cputime_t cputime)
4124 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4126 tmp = cputime_to_cputime64(cputime);
4128 p->utime = cputime_add(p->utime, cputime);
4129 account_group_user_time(p, cputime);
4130 p->gtime = cputime_add(p->gtime, cputime);
4132 cpustat->user = cputime64_add(cpustat->user, tmp);
4133 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4137 * Account scaled user cpu time to a process.
4138 * @p: the process that the cpu time gets accounted to
4139 * @cputime: the cpu time spent in user space since the last update
4141 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4143 p->utimescaled = cputime_add(p->utimescaled, cputime);
4147 * Account system cpu time to a process.
4148 * @p: the process that the cpu time gets accounted to
4149 * @hardirq_offset: the offset to subtract from hardirq_count()
4150 * @cputime: the cpu time spent in kernel space since the last update
4152 void account_system_time(struct task_struct *p, int hardirq_offset,
4155 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4156 struct rq *rq = this_rq();
4159 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4160 account_guest_time(p, cputime);
4164 p->stime = cputime_add(p->stime, cputime);
4165 account_group_system_time(p, cputime);
4167 /* Add system time to cpustat. */
4168 tmp = cputime_to_cputime64(cputime);
4169 if (hardirq_count() - hardirq_offset)
4170 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4171 else if (softirq_count())
4172 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4173 else if (p != rq->idle)
4174 cpustat->system = cputime64_add(cpustat->system, tmp);
4175 else if (atomic_read(&rq->nr_iowait) > 0)
4176 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4178 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4179 /* Account for system time used */
4180 acct_update_integrals(p);
4184 * Account scaled system cpu time to a process.
4185 * @p: the process that the cpu time gets accounted to
4186 * @hardirq_offset: the offset to subtract from hardirq_count()
4187 * @cputime: the cpu time spent in kernel space since the last update
4189 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4191 p->stimescaled = cputime_add(p->stimescaled, cputime);
4195 * Account for involuntary wait time.
4196 * @p: the process from which the cpu time has been stolen
4197 * @steal: the cpu time spent in involuntary wait
4199 void account_steal_time(struct task_struct *p, cputime_t steal)
4201 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4202 cputime64_t tmp = cputime_to_cputime64(steal);
4203 struct rq *rq = this_rq();
4205 if (p == rq->idle) {
4206 p->stime = cputime_add(p->stime, steal);
4207 account_group_system_time(p, steal);
4208 if (atomic_read(&rq->nr_iowait) > 0)
4209 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4211 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4213 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4217 * Use precise platform statistics if available:
4219 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4220 cputime_t task_utime(struct task_struct *p)
4225 cputime_t task_stime(struct task_struct *p)
4230 cputime_t task_utime(struct task_struct *p)
4232 clock_t utime = cputime_to_clock_t(p->utime),
4233 total = utime + cputime_to_clock_t(p->stime);
4237 * Use CFS's precise accounting:
4239 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
4243 do_div(temp, total);
4245 utime = (clock_t)temp;
4247 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
4248 return p->prev_utime;
4251 cputime_t task_stime(struct task_struct *p)
4256 * Use CFS's precise accounting. (we subtract utime from
4257 * the total, to make sure the total observed by userspace
4258 * grows monotonically - apps rely on that):
4260 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
4261 cputime_to_clock_t(task_utime(p));
4264 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
4266 return p->prev_stime;
4270 inline cputime_t task_gtime(struct task_struct *p)
4276 * This function gets called by the timer code, with HZ frequency.
4277 * We call it with interrupts disabled.
4279 * It also gets called by the fork code, when changing the parent's
4282 void scheduler_tick(void)
4284 int cpu = smp_processor_id();
4285 struct rq *rq = cpu_rq(cpu);
4286 struct task_struct *curr = rq->curr;
4290 spin_lock(&rq->lock);
4291 update_rq_clock(rq);
4292 update_cpu_load(rq);
4293 curr->sched_class->task_tick(rq, curr, 0);
4294 spin_unlock(&rq->lock);
4297 rq->idle_at_tick = idle_cpu(cpu);
4298 trigger_load_balance(rq, cpu);
4302 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4303 defined(CONFIG_PREEMPT_TRACER))
4305 static inline unsigned long get_parent_ip(unsigned long addr)
4307 if (in_lock_functions(addr)) {
4308 addr = CALLER_ADDR2;
4309 if (in_lock_functions(addr))
4310 addr = CALLER_ADDR3;
4315 void __kprobes add_preempt_count(int val)
4317 #ifdef CONFIG_DEBUG_PREEMPT
4321 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4324 preempt_count() += val;
4325 #ifdef CONFIG_DEBUG_PREEMPT
4327 * Spinlock count overflowing soon?
4329 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4332 if (preempt_count() == val)
4333 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4335 EXPORT_SYMBOL(add_preempt_count);
4337 void __kprobes sub_preempt_count(int val)
4339 #ifdef CONFIG_DEBUG_PREEMPT
4343 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4346 * Is the spinlock portion underflowing?
4348 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4349 !(preempt_count() & PREEMPT_MASK)))
4353 if (preempt_count() == val)
4354 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4355 preempt_count() -= val;
4357 EXPORT_SYMBOL(sub_preempt_count);
4362 * Print scheduling while atomic bug:
4364 static noinline void __schedule_bug(struct task_struct *prev)
4366 struct pt_regs *regs = get_irq_regs();
4368 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4369 prev->comm, prev->pid, preempt_count());
4371 debug_show_held_locks(prev);
4373 if (irqs_disabled())
4374 print_irqtrace_events(prev);
4383 * Various schedule()-time debugging checks and statistics:
4385 static inline void schedule_debug(struct task_struct *prev)
4388 * Test if we are atomic. Since do_exit() needs to call into
4389 * schedule() atomically, we ignore that path for now.
4390 * Otherwise, whine if we are scheduling when we should not be.
4392 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4393 __schedule_bug(prev);
4395 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4397 schedstat_inc(this_rq(), sched_count);
4398 #ifdef CONFIG_SCHEDSTATS
4399 if (unlikely(prev->lock_depth >= 0)) {
4400 schedstat_inc(this_rq(), bkl_count);
4401 schedstat_inc(prev, sched_info.bkl_count);
4407 * Pick up the highest-prio task:
4409 static inline struct task_struct *
4410 pick_next_task(struct rq *rq, struct task_struct *prev)
4412 const struct sched_class *class;
4413 struct task_struct *p;
4416 * Optimization: we know that if all tasks are in
4417 * the fair class we can call that function directly:
4419 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4420 p = fair_sched_class.pick_next_task(rq);
4425 class = sched_class_highest;
4427 p = class->pick_next_task(rq);
4431 * Will never be NULL as the idle class always
4432 * returns a non-NULL p:
4434 class = class->next;
4439 * schedule() is the main scheduler function.
4441 asmlinkage void __sched schedule(void)
4443 struct task_struct *prev, *next;
4444 unsigned long *switch_count;
4450 cpu = smp_processor_id();
4454 switch_count = &prev->nivcsw;
4456 release_kernel_lock(prev);
4457 need_resched_nonpreemptible:
4459 schedule_debug(prev);
4461 if (sched_feat(HRTICK))
4464 spin_lock_irq(&rq->lock);
4465 update_rq_clock(rq);
4466 clear_tsk_need_resched(prev);
4468 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4469 if (unlikely(signal_pending_state(prev->state, prev)))
4470 prev->state = TASK_RUNNING;
4472 deactivate_task(rq, prev, 1);
4473 switch_count = &prev->nvcsw;
4477 if (prev->sched_class->pre_schedule)
4478 prev->sched_class->pre_schedule(rq, prev);
4481 if (unlikely(!rq->nr_running))
4482 idle_balance(cpu, rq);
4484 prev->sched_class->put_prev_task(rq, prev);
4485 next = pick_next_task(rq, prev);
4487 if (likely(prev != next)) {
4488 sched_info_switch(prev, next);
4494 context_switch(rq, prev, next); /* unlocks the rq */
4496 * the context switch might have flipped the stack from under
4497 * us, hence refresh the local variables.
4499 cpu = smp_processor_id();
4502 spin_unlock_irq(&rq->lock);
4504 if (unlikely(reacquire_kernel_lock(current) < 0))
4505 goto need_resched_nonpreemptible;
4507 preempt_enable_no_resched();
4508 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4511 EXPORT_SYMBOL(schedule);
4513 #ifdef CONFIG_PREEMPT
4515 * this is the entry point to schedule() from in-kernel preemption
4516 * off of preempt_enable. Kernel preemptions off return from interrupt
4517 * occur there and call schedule directly.
4519 asmlinkage void __sched preempt_schedule(void)
4521 struct thread_info *ti = current_thread_info();
4524 * If there is a non-zero preempt_count or interrupts are disabled,
4525 * we do not want to preempt the current task. Just return..
4527 if (likely(ti->preempt_count || irqs_disabled()))
4531 add_preempt_count(PREEMPT_ACTIVE);
4533 sub_preempt_count(PREEMPT_ACTIVE);
4536 * Check again in case we missed a preemption opportunity
4537 * between schedule and now.
4540 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4542 EXPORT_SYMBOL(preempt_schedule);
4545 * this is the entry point to schedule() from kernel preemption
4546 * off of irq context.
4547 * Note, that this is called and return with irqs disabled. This will
4548 * protect us against recursive calling from irq.
4550 asmlinkage void __sched preempt_schedule_irq(void)
4552 struct thread_info *ti = current_thread_info();
4554 /* Catch callers which need to be fixed */
4555 BUG_ON(ti->preempt_count || !irqs_disabled());
4558 add_preempt_count(PREEMPT_ACTIVE);
4561 local_irq_disable();
4562 sub_preempt_count(PREEMPT_ACTIVE);
4565 * Check again in case we missed a preemption opportunity
4566 * between schedule and now.
4569 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4572 #endif /* CONFIG_PREEMPT */
4574 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4577 return try_to_wake_up(curr->private, mode, sync);
4579 EXPORT_SYMBOL(default_wake_function);
4582 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4583 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4584 * number) then we wake all the non-exclusive tasks and one exclusive task.
4586 * There are circumstances in which we can try to wake a task which has already
4587 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4588 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4590 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4591 int nr_exclusive, int sync, void *key)
4593 wait_queue_t *curr, *next;
4595 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4596 unsigned flags = curr->flags;
4598 if (curr->func(curr, mode, sync, key) &&
4599 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4605 * __wake_up - wake up threads blocked on a waitqueue.
4607 * @mode: which threads
4608 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4609 * @key: is directly passed to the wakeup function
4611 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4612 int nr_exclusive, void *key)
4614 unsigned long flags;
4616 spin_lock_irqsave(&q->lock, flags);
4617 __wake_up_common(q, mode, nr_exclusive, 0, key);
4618 spin_unlock_irqrestore(&q->lock, flags);
4620 EXPORT_SYMBOL(__wake_up);
4623 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4625 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4627 __wake_up_common(q, mode, 1, 0, NULL);
4631 * __wake_up_sync - wake up threads blocked on a waitqueue.
4633 * @mode: which threads
4634 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4636 * The sync wakeup differs that the waker knows that it will schedule
4637 * away soon, so while the target thread will be woken up, it will not
4638 * be migrated to another CPU - ie. the two threads are 'synchronized'
4639 * with each other. This can prevent needless bouncing between CPUs.
4641 * On UP it can prevent extra preemption.
4644 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4646 unsigned long flags;
4652 if (unlikely(!nr_exclusive))
4655 spin_lock_irqsave(&q->lock, flags);
4656 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4657 spin_unlock_irqrestore(&q->lock, flags);
4659 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4662 * complete: - signals a single thread waiting on this completion
4663 * @x: holds the state of this particular completion
4665 * This will wake up a single thread waiting on this completion. Threads will be
4666 * awakened in the same order in which they were queued.
4668 * See also complete_all(), wait_for_completion() and related routines.
4670 void complete(struct completion *x)
4672 unsigned long flags;
4674 spin_lock_irqsave(&x->wait.lock, flags);
4676 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4677 spin_unlock_irqrestore(&x->wait.lock, flags);
4679 EXPORT_SYMBOL(complete);
4682 * complete_all: - signals all threads waiting on this completion
4683 * @x: holds the state of this particular completion
4685 * This will wake up all threads waiting on this particular completion event.
4687 void complete_all(struct completion *x)
4689 unsigned long flags;
4691 spin_lock_irqsave(&x->wait.lock, flags);
4692 x->done += UINT_MAX/2;
4693 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4694 spin_unlock_irqrestore(&x->wait.lock, flags);
4696 EXPORT_SYMBOL(complete_all);
4698 static inline long __sched
4699 do_wait_for_common(struct completion *x, long timeout, int state)
4702 DECLARE_WAITQUEUE(wait, current);
4704 wait.flags |= WQ_FLAG_EXCLUSIVE;
4705 __add_wait_queue_tail(&x->wait, &wait);
4707 if (signal_pending_state(state, current)) {
4708 timeout = -ERESTARTSYS;
4711 __set_current_state(state);
4712 spin_unlock_irq(&x->wait.lock);
4713 timeout = schedule_timeout(timeout);
4714 spin_lock_irq(&x->wait.lock);
4715 } while (!x->done && timeout);
4716 __remove_wait_queue(&x->wait, &wait);
4721 return timeout ?: 1;
4725 wait_for_common(struct completion *x, long timeout, int state)
4729 spin_lock_irq(&x->wait.lock);
4730 timeout = do_wait_for_common(x, timeout, state);
4731 spin_unlock_irq(&x->wait.lock);
4736 * wait_for_completion: - waits for completion of a task
4737 * @x: holds the state of this particular completion
4739 * This waits to be signaled for completion of a specific task. It is NOT
4740 * interruptible and there is no timeout.
4742 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4743 * and interrupt capability. Also see complete().
4745 void __sched wait_for_completion(struct completion *x)
4747 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4749 EXPORT_SYMBOL(wait_for_completion);
4752 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4753 * @x: holds the state of this particular completion
4754 * @timeout: timeout value in jiffies
4756 * This waits for either a completion of a specific task to be signaled or for a
4757 * specified timeout to expire. The timeout is in jiffies. It is not
4760 unsigned long __sched
4761 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4763 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4765 EXPORT_SYMBOL(wait_for_completion_timeout);
4768 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4769 * @x: holds the state of this particular completion
4771 * This waits for completion of a specific task to be signaled. It is
4774 int __sched wait_for_completion_interruptible(struct completion *x)
4776 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4777 if (t == -ERESTARTSYS)
4781 EXPORT_SYMBOL(wait_for_completion_interruptible);
4784 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4785 * @x: holds the state of this particular completion
4786 * @timeout: timeout value in jiffies
4788 * This waits for either a completion of a specific task to be signaled or for a
4789 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4791 unsigned long __sched
4792 wait_for_completion_interruptible_timeout(struct completion *x,
4793 unsigned long timeout)
4795 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4797 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4800 * wait_for_completion_killable: - waits for completion of a task (killable)
4801 * @x: holds the state of this particular completion
4803 * This waits to be signaled for completion of a specific task. It can be
4804 * interrupted by a kill signal.
4806 int __sched wait_for_completion_killable(struct completion *x)
4808 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4809 if (t == -ERESTARTSYS)
4813 EXPORT_SYMBOL(wait_for_completion_killable);
4816 * try_wait_for_completion - try to decrement a completion without blocking
4817 * @x: completion structure
4819 * Returns: 0 if a decrement cannot be done without blocking
4820 * 1 if a decrement succeeded.
4822 * If a completion is being used as a counting completion,
4823 * attempt to decrement the counter without blocking. This
4824 * enables us to avoid waiting if the resource the completion
4825 * is protecting is not available.
4827 bool try_wait_for_completion(struct completion *x)
4831 spin_lock_irq(&x->wait.lock);
4836 spin_unlock_irq(&x->wait.lock);
4839 EXPORT_SYMBOL(try_wait_for_completion);
4842 * completion_done - Test to see if a completion has any waiters
4843 * @x: completion structure
4845 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4846 * 1 if there are no waiters.
4849 bool completion_done(struct completion *x)
4853 spin_lock_irq(&x->wait.lock);
4856 spin_unlock_irq(&x->wait.lock);
4859 EXPORT_SYMBOL(completion_done);
4862 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4864 unsigned long flags;
4867 init_waitqueue_entry(&wait, current);
4869 __set_current_state(state);
4871 spin_lock_irqsave(&q->lock, flags);
4872 __add_wait_queue(q, &wait);
4873 spin_unlock(&q->lock);
4874 timeout = schedule_timeout(timeout);
4875 spin_lock_irq(&q->lock);
4876 __remove_wait_queue(q, &wait);
4877 spin_unlock_irqrestore(&q->lock, flags);
4882 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4884 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4886 EXPORT_SYMBOL(interruptible_sleep_on);
4889 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4891 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4893 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4895 void __sched sleep_on(wait_queue_head_t *q)
4897 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4899 EXPORT_SYMBOL(sleep_on);
4901 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4903 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4905 EXPORT_SYMBOL(sleep_on_timeout);
4907 #ifdef CONFIG_RT_MUTEXES
4910 * rt_mutex_setprio - set the current priority of a task
4912 * @prio: prio value (kernel-internal form)
4914 * This function changes the 'effective' priority of a task. It does
4915 * not touch ->normal_prio like __setscheduler().
4917 * Used by the rt_mutex code to implement priority inheritance logic.
4919 void rt_mutex_setprio(struct task_struct *p, int prio)
4921 unsigned long flags;
4922 int oldprio, on_rq, running;
4924 const struct sched_class *prev_class = p->sched_class;
4926 BUG_ON(prio < 0 || prio > MAX_PRIO);
4928 rq = task_rq_lock(p, &flags);
4929 update_rq_clock(rq);
4932 on_rq = p->se.on_rq;
4933 running = task_current(rq, p);
4935 dequeue_task(rq, p, 0);
4937 p->sched_class->put_prev_task(rq, p);
4940 p->sched_class = &rt_sched_class;
4942 p->sched_class = &fair_sched_class;
4947 p->sched_class->set_curr_task(rq);
4949 enqueue_task(rq, p, 0);
4951 check_class_changed(rq, p, prev_class, oldprio, running);
4953 task_rq_unlock(rq, &flags);
4958 void set_user_nice(struct task_struct *p, long nice)
4960 int old_prio, delta, on_rq;
4961 unsigned long flags;
4964 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4967 * We have to be careful, if called from sys_setpriority(),
4968 * the task might be in the middle of scheduling on another CPU.
4970 rq = task_rq_lock(p, &flags);
4971 update_rq_clock(rq);
4973 * The RT priorities are set via sched_setscheduler(), but we still
4974 * allow the 'normal' nice value to be set - but as expected
4975 * it wont have any effect on scheduling until the task is
4976 * SCHED_FIFO/SCHED_RR:
4978 if (task_has_rt_policy(p)) {
4979 p->static_prio = NICE_TO_PRIO(nice);
4982 on_rq = p->se.on_rq;
4984 dequeue_task(rq, p, 0);
4986 p->static_prio = NICE_TO_PRIO(nice);
4989 p->prio = effective_prio(p);
4990 delta = p->prio - old_prio;
4993 enqueue_task(rq, p, 0);
4995 * If the task increased its priority or is running and
4996 * lowered its priority, then reschedule its CPU:
4998 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4999 resched_task(rq->curr);
5002 task_rq_unlock(rq, &flags);
5004 EXPORT_SYMBOL(set_user_nice);
5007 * can_nice - check if a task can reduce its nice value
5011 int can_nice(const struct task_struct *p, const int nice)
5013 /* convert nice value [19,-20] to rlimit style value [1,40] */
5014 int nice_rlim = 20 - nice;
5016 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
5017 capable(CAP_SYS_NICE));
5020 #ifdef __ARCH_WANT_SYS_NICE
5023 * sys_nice - change the priority of the current process.
5024 * @increment: priority increment
5026 * sys_setpriority is a more generic, but much slower function that
5027 * does similar things.
5029 asmlinkage long sys_nice(int increment)
5034 * Setpriority might change our priority at the same moment.
5035 * We don't have to worry. Conceptually one call occurs first
5036 * and we have a single winner.
5038 if (increment < -40)
5043 nice = PRIO_TO_NICE(current->static_prio) + increment;
5049 if (increment < 0 && !can_nice(current, nice))
5052 retval = security_task_setnice(current, nice);
5056 set_user_nice(current, nice);
5063 * task_prio - return the priority value of a given task.
5064 * @p: the task in question.
5066 * This is the priority value as seen by users in /proc.
5067 * RT tasks are offset by -200. Normal tasks are centered
5068 * around 0, value goes from -16 to +15.
5070 int task_prio(const struct task_struct *p)
5072 return p->prio - MAX_RT_PRIO;
5076 * task_nice - return the nice value of a given task.
5077 * @p: the task in question.
5079 int task_nice(const struct task_struct *p)
5081 return TASK_NICE(p);
5083 EXPORT_SYMBOL(task_nice);
5086 * idle_cpu - is a given cpu idle currently?
5087 * @cpu: the processor in question.
5089 int idle_cpu(int cpu)
5091 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5095 * idle_task - return the idle task for a given cpu.
5096 * @cpu: the processor in question.
5098 struct task_struct *idle_task(int cpu)
5100 return cpu_rq(cpu)->idle;
5104 * find_process_by_pid - find a process with a matching PID value.
5105 * @pid: the pid in question.
5107 static struct task_struct *find_process_by_pid(pid_t pid)
5109 return pid ? find_task_by_vpid(pid) : current;
5112 /* Actually do priority change: must hold rq lock. */
5114 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5116 BUG_ON(p->se.on_rq);
5119 switch (p->policy) {
5123 p->sched_class = &fair_sched_class;
5127 p->sched_class = &rt_sched_class;
5131 p->rt_priority = prio;
5132 p->normal_prio = normal_prio(p);
5133 /* we are holding p->pi_lock already */
5134 p->prio = rt_mutex_getprio(p);
5139 * check the target process has a UID that matches the current process's
5141 static bool check_same_owner(struct task_struct *p)
5143 const struct cred *cred = current_cred(), *pcred;
5147 pcred = __task_cred(p);
5148 match = (cred->euid == pcred->euid ||
5149 cred->euid == pcred->uid);
5154 static int __sched_setscheduler(struct task_struct *p, int policy,
5155 struct sched_param *param, bool user)
5157 int retval, oldprio, oldpolicy = -1, on_rq, running;
5158 unsigned long flags;
5159 const struct sched_class *prev_class = p->sched_class;
5162 /* may grab non-irq protected spin_locks */
5163 BUG_ON(in_interrupt());
5165 /* double check policy once rq lock held */
5167 policy = oldpolicy = p->policy;
5168 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5169 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5170 policy != SCHED_IDLE)
5173 * Valid priorities for SCHED_FIFO and SCHED_RR are
5174 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5175 * SCHED_BATCH and SCHED_IDLE is 0.
5177 if (param->sched_priority < 0 ||
5178 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5179 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5181 if (rt_policy(policy) != (param->sched_priority != 0))
5185 * Allow unprivileged RT tasks to decrease priority:
5187 if (user && !capable(CAP_SYS_NICE)) {
5188 if (rt_policy(policy)) {
5189 unsigned long rlim_rtprio;
5191 if (!lock_task_sighand(p, &flags))
5193 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5194 unlock_task_sighand(p, &flags);
5196 /* can't set/change the rt policy */
5197 if (policy != p->policy && !rlim_rtprio)
5200 /* can't increase priority */
5201 if (param->sched_priority > p->rt_priority &&
5202 param->sched_priority > rlim_rtprio)
5206 * Like positive nice levels, dont allow tasks to
5207 * move out of SCHED_IDLE either:
5209 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5212 /* can't change other user's priorities */
5213 if (!check_same_owner(p))
5218 #ifdef CONFIG_RT_GROUP_SCHED
5220 * Do not allow realtime tasks into groups that have no runtime
5223 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5224 task_group(p)->rt_bandwidth.rt_runtime == 0)
5228 retval = security_task_setscheduler(p, policy, param);
5234 * make sure no PI-waiters arrive (or leave) while we are
5235 * changing the priority of the task:
5237 spin_lock_irqsave(&p->pi_lock, flags);
5239 * To be able to change p->policy safely, the apropriate
5240 * runqueue lock must be held.
5242 rq = __task_rq_lock(p);
5243 /* recheck policy now with rq lock held */
5244 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5245 policy = oldpolicy = -1;
5246 __task_rq_unlock(rq);
5247 spin_unlock_irqrestore(&p->pi_lock, flags);
5250 update_rq_clock(rq);
5251 on_rq = p->se.on_rq;
5252 running = task_current(rq, p);
5254 deactivate_task(rq, p, 0);
5256 p->sched_class->put_prev_task(rq, p);
5259 __setscheduler(rq, p, policy, param->sched_priority);
5262 p->sched_class->set_curr_task(rq);
5264 activate_task(rq, p, 0);
5266 check_class_changed(rq, p, prev_class, oldprio, running);
5268 __task_rq_unlock(rq);
5269 spin_unlock_irqrestore(&p->pi_lock, flags);
5271 rt_mutex_adjust_pi(p);
5277 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5278 * @p: the task in question.
5279 * @policy: new policy.
5280 * @param: structure containing the new RT priority.
5282 * NOTE that the task may be already dead.
5284 int sched_setscheduler(struct task_struct *p, int policy,
5285 struct sched_param *param)
5287 return __sched_setscheduler(p, policy, param, true);
5289 EXPORT_SYMBOL_GPL(sched_setscheduler);
5292 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5293 * @p: the task in question.
5294 * @policy: new policy.
5295 * @param: structure containing the new RT priority.
5297 * Just like sched_setscheduler, only don't bother checking if the
5298 * current context has permission. For example, this is needed in
5299 * stop_machine(): we create temporary high priority worker threads,
5300 * but our caller might not have that capability.
5302 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5303 struct sched_param *param)
5305 return __sched_setscheduler(p, policy, param, false);
5309 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5311 struct sched_param lparam;
5312 struct task_struct *p;
5315 if (!param || pid < 0)
5317 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5322 p = find_process_by_pid(pid);
5324 retval = sched_setscheduler(p, policy, &lparam);
5331 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5332 * @pid: the pid in question.
5333 * @policy: new policy.
5334 * @param: structure containing the new RT priority.
5337 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5339 /* negative values for policy are not valid */
5343 return do_sched_setscheduler(pid, policy, param);
5347 * sys_sched_setparam - set/change the RT priority of a thread
5348 * @pid: the pid in question.
5349 * @param: structure containing the new RT priority.
5351 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5353 return do_sched_setscheduler(pid, -1, param);
5357 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5358 * @pid: the pid in question.
5360 asmlinkage long sys_sched_getscheduler(pid_t pid)
5362 struct task_struct *p;
5369 read_lock(&tasklist_lock);
5370 p = find_process_by_pid(pid);
5372 retval = security_task_getscheduler(p);
5376 read_unlock(&tasklist_lock);
5381 * sys_sched_getscheduler - get the RT priority of a thread
5382 * @pid: the pid in question.
5383 * @param: structure containing the RT priority.
5385 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5387 struct sched_param lp;
5388 struct task_struct *p;
5391 if (!param || pid < 0)
5394 read_lock(&tasklist_lock);
5395 p = find_process_by_pid(pid);
5400 retval = security_task_getscheduler(p);
5404 lp.sched_priority = p->rt_priority;
5405 read_unlock(&tasklist_lock);
5408 * This one might sleep, we cannot do it with a spinlock held ...
5410 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5415 read_unlock(&tasklist_lock);
5419 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5421 cpumask_t cpus_allowed;
5422 cpumask_t new_mask = *in_mask;
5423 struct task_struct *p;
5427 read_lock(&tasklist_lock);
5429 p = find_process_by_pid(pid);
5431 read_unlock(&tasklist_lock);
5437 * It is not safe to call set_cpus_allowed with the
5438 * tasklist_lock held. We will bump the task_struct's
5439 * usage count and then drop tasklist_lock.
5442 read_unlock(&tasklist_lock);
5445 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
5448 retval = security_task_setscheduler(p, 0, NULL);
5452 cpuset_cpus_allowed(p, &cpus_allowed);
5453 cpus_and(new_mask, new_mask, cpus_allowed);
5455 retval = set_cpus_allowed_ptr(p, &new_mask);
5458 cpuset_cpus_allowed(p, &cpus_allowed);
5459 if (!cpus_subset(new_mask, cpus_allowed)) {
5461 * We must have raced with a concurrent cpuset
5462 * update. Just reset the cpus_allowed to the
5463 * cpuset's cpus_allowed
5465 new_mask = cpus_allowed;
5475 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5476 cpumask_t *new_mask)
5478 if (len < sizeof(cpumask_t)) {
5479 memset(new_mask, 0, sizeof(cpumask_t));
5480 } else if (len > sizeof(cpumask_t)) {
5481 len = sizeof(cpumask_t);
5483 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5487 * sys_sched_setaffinity - set the cpu affinity of a process
5488 * @pid: pid of the process
5489 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5490 * @user_mask_ptr: user-space pointer to the new cpu mask
5492 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5493 unsigned long __user *user_mask_ptr)
5498 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5502 return sched_setaffinity(pid, &new_mask);
5505 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5507 struct task_struct *p;
5511 read_lock(&tasklist_lock);
5514 p = find_process_by_pid(pid);
5518 retval = security_task_getscheduler(p);
5522 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5525 read_unlock(&tasklist_lock);
5532 * sys_sched_getaffinity - get the cpu affinity of a process
5533 * @pid: pid of the process
5534 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5535 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5537 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5538 unsigned long __user *user_mask_ptr)
5543 if (len < sizeof(cpumask_t))
5546 ret = sched_getaffinity(pid, &mask);
5550 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5553 return sizeof(cpumask_t);
5557 * sys_sched_yield - yield the current processor to other threads.
5559 * This function yields the current CPU to other tasks. If there are no
5560 * other threads running on this CPU then this function will return.
5562 asmlinkage long sys_sched_yield(void)
5564 struct rq *rq = this_rq_lock();
5566 schedstat_inc(rq, yld_count);
5567 current->sched_class->yield_task(rq);
5570 * Since we are going to call schedule() anyway, there's
5571 * no need to preempt or enable interrupts:
5573 __release(rq->lock);
5574 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5575 _raw_spin_unlock(&rq->lock);
5576 preempt_enable_no_resched();
5583 static void __cond_resched(void)
5585 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5586 __might_sleep(__FILE__, __LINE__);
5589 * The BKS might be reacquired before we have dropped
5590 * PREEMPT_ACTIVE, which could trigger a second
5591 * cond_resched() call.
5594 add_preempt_count(PREEMPT_ACTIVE);
5596 sub_preempt_count(PREEMPT_ACTIVE);
5597 } while (need_resched());
5600 int __sched _cond_resched(void)
5602 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5603 system_state == SYSTEM_RUNNING) {
5609 EXPORT_SYMBOL(_cond_resched);
5612 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5613 * call schedule, and on return reacquire the lock.
5615 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5616 * operations here to prevent schedule() from being called twice (once via
5617 * spin_unlock(), once by hand).
5619 int cond_resched_lock(spinlock_t *lock)
5621 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5624 if (spin_needbreak(lock) || resched) {
5626 if (resched && need_resched())
5635 EXPORT_SYMBOL(cond_resched_lock);
5637 int __sched cond_resched_softirq(void)
5639 BUG_ON(!in_softirq());
5641 if (need_resched() && system_state == SYSTEM_RUNNING) {
5649 EXPORT_SYMBOL(cond_resched_softirq);
5652 * yield - yield the current processor to other threads.
5654 * This is a shortcut for kernel-space yielding - it marks the
5655 * thread runnable and calls sys_sched_yield().
5657 void __sched yield(void)
5659 set_current_state(TASK_RUNNING);
5662 EXPORT_SYMBOL(yield);
5665 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5666 * that process accounting knows that this is a task in IO wait state.
5668 * But don't do that if it is a deliberate, throttling IO wait (this task
5669 * has set its backing_dev_info: the queue against which it should throttle)
5671 void __sched io_schedule(void)
5673 struct rq *rq = &__raw_get_cpu_var(runqueues);
5675 delayacct_blkio_start();
5676 atomic_inc(&rq->nr_iowait);
5678 atomic_dec(&rq->nr_iowait);
5679 delayacct_blkio_end();
5681 EXPORT_SYMBOL(io_schedule);
5683 long __sched io_schedule_timeout(long timeout)
5685 struct rq *rq = &__raw_get_cpu_var(runqueues);
5688 delayacct_blkio_start();
5689 atomic_inc(&rq->nr_iowait);
5690 ret = schedule_timeout(timeout);
5691 atomic_dec(&rq->nr_iowait);
5692 delayacct_blkio_end();
5697 * sys_sched_get_priority_max - return maximum RT priority.
5698 * @policy: scheduling class.
5700 * this syscall returns the maximum rt_priority that can be used
5701 * by a given scheduling class.
5703 asmlinkage long sys_sched_get_priority_max(int policy)
5710 ret = MAX_USER_RT_PRIO-1;
5722 * sys_sched_get_priority_min - return minimum RT priority.
5723 * @policy: scheduling class.
5725 * this syscall returns the minimum rt_priority that can be used
5726 * by a given scheduling class.
5728 asmlinkage long sys_sched_get_priority_min(int policy)
5746 * sys_sched_rr_get_interval - return the default timeslice of a process.
5747 * @pid: pid of the process.
5748 * @interval: userspace pointer to the timeslice value.
5750 * this syscall writes the default timeslice value of a given process
5751 * into the user-space timespec buffer. A value of '0' means infinity.
5754 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5756 struct task_struct *p;
5757 unsigned int time_slice;
5765 read_lock(&tasklist_lock);
5766 p = find_process_by_pid(pid);
5770 retval = security_task_getscheduler(p);
5775 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5776 * tasks that are on an otherwise idle runqueue:
5779 if (p->policy == SCHED_RR) {
5780 time_slice = DEF_TIMESLICE;
5781 } else if (p->policy != SCHED_FIFO) {
5782 struct sched_entity *se = &p->se;
5783 unsigned long flags;
5786 rq = task_rq_lock(p, &flags);
5787 if (rq->cfs.load.weight)
5788 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5789 task_rq_unlock(rq, &flags);
5791 read_unlock(&tasklist_lock);
5792 jiffies_to_timespec(time_slice, &t);
5793 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5797 read_unlock(&tasklist_lock);
5801 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5803 void sched_show_task(struct task_struct *p)
5805 unsigned long free = 0;
5808 state = p->state ? __ffs(p->state) + 1 : 0;
5809 printk(KERN_INFO "%-13.13s %c", p->comm,
5810 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5811 #if BITS_PER_LONG == 32
5812 if (state == TASK_RUNNING)
5813 printk(KERN_CONT " running ");
5815 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5817 if (state == TASK_RUNNING)
5818 printk(KERN_CONT " running task ");
5820 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5822 #ifdef CONFIG_DEBUG_STACK_USAGE
5824 unsigned long *n = end_of_stack(p);
5827 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5830 printk(KERN_CONT "%5lu %5d %6d\n", free,
5831 task_pid_nr(p), task_pid_nr(p->real_parent));
5833 show_stack(p, NULL);
5836 void show_state_filter(unsigned long state_filter)
5838 struct task_struct *g, *p;
5840 #if BITS_PER_LONG == 32
5842 " task PC stack pid father\n");
5845 " task PC stack pid father\n");
5847 read_lock(&tasklist_lock);
5848 do_each_thread(g, p) {
5850 * reset the NMI-timeout, listing all files on a slow
5851 * console might take alot of time:
5853 touch_nmi_watchdog();
5854 if (!state_filter || (p->state & state_filter))
5856 } while_each_thread(g, p);
5858 touch_all_softlockup_watchdogs();
5860 #ifdef CONFIG_SCHED_DEBUG
5861 sysrq_sched_debug_show();
5863 read_unlock(&tasklist_lock);
5865 * Only show locks if all tasks are dumped:
5867 if (state_filter == -1)
5868 debug_show_all_locks();
5871 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5873 idle->sched_class = &idle_sched_class;
5877 * init_idle - set up an idle thread for a given CPU
5878 * @idle: task in question
5879 * @cpu: cpu the idle task belongs to
5881 * NOTE: this function does not set the idle thread's NEED_RESCHED
5882 * flag, to make booting more robust.
5884 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5886 struct rq *rq = cpu_rq(cpu);
5887 unsigned long flags;
5889 spin_lock_irqsave(&rq->lock, flags);
5892 idle->se.exec_start = sched_clock();
5894 idle->prio = idle->normal_prio = MAX_PRIO;
5895 idle->cpus_allowed = cpumask_of_cpu(cpu);
5896 __set_task_cpu(idle, cpu);
5898 rq->curr = rq->idle = idle;
5899 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5902 spin_unlock_irqrestore(&rq->lock, flags);
5904 /* Set the preempt count _outside_ the spinlocks! */
5905 #if defined(CONFIG_PREEMPT)
5906 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5908 task_thread_info(idle)->preempt_count = 0;
5911 * The idle tasks have their own, simple scheduling class:
5913 idle->sched_class = &idle_sched_class;
5917 * In a system that switches off the HZ timer nohz_cpu_mask
5918 * indicates which cpus entered this state. This is used
5919 * in the rcu update to wait only for active cpus. For system
5920 * which do not switch off the HZ timer nohz_cpu_mask should
5921 * always be CPU_MASK_NONE.
5923 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5926 * Increase the granularity value when there are more CPUs,
5927 * because with more CPUs the 'effective latency' as visible
5928 * to users decreases. But the relationship is not linear,
5929 * so pick a second-best guess by going with the log2 of the
5932 * This idea comes from the SD scheduler of Con Kolivas:
5934 static inline void sched_init_granularity(void)
5936 unsigned int factor = 1 + ilog2(num_online_cpus());
5937 const unsigned long limit = 200000000;
5939 sysctl_sched_min_granularity *= factor;
5940 if (sysctl_sched_min_granularity > limit)
5941 sysctl_sched_min_granularity = limit;
5943 sysctl_sched_latency *= factor;
5944 if (sysctl_sched_latency > limit)
5945 sysctl_sched_latency = limit;
5947 sysctl_sched_wakeup_granularity *= factor;
5949 sysctl_sched_shares_ratelimit *= factor;
5954 * This is how migration works:
5956 * 1) we queue a struct migration_req structure in the source CPU's
5957 * runqueue and wake up that CPU's migration thread.
5958 * 2) we down() the locked semaphore => thread blocks.
5959 * 3) migration thread wakes up (implicitly it forces the migrated
5960 * thread off the CPU)
5961 * 4) it gets the migration request and checks whether the migrated
5962 * task is still in the wrong runqueue.
5963 * 5) if it's in the wrong runqueue then the migration thread removes
5964 * it and puts it into the right queue.
5965 * 6) migration thread up()s the semaphore.
5966 * 7) we wake up and the migration is done.
5970 * Change a given task's CPU affinity. Migrate the thread to a
5971 * proper CPU and schedule it away if the CPU it's executing on
5972 * is removed from the allowed bitmask.
5974 * NOTE: the caller must have a valid reference to the task, the
5975 * task must not exit() & deallocate itself prematurely. The
5976 * call is not atomic; no spinlocks may be held.
5978 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5980 struct migration_req req;
5981 unsigned long flags;
5985 rq = task_rq_lock(p, &flags);
5986 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5991 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5992 !cpus_equal(p->cpus_allowed, *new_mask))) {
5997 if (p->sched_class->set_cpus_allowed)
5998 p->sched_class->set_cpus_allowed(p, new_mask);
6000 p->cpus_allowed = *new_mask;
6001 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
6004 /* Can the task run on the task's current CPU? If so, we're done */
6005 if (cpu_isset(task_cpu(p), *new_mask))
6008 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
6009 /* Need help from migration thread: drop lock and wait. */
6010 task_rq_unlock(rq, &flags);
6011 wake_up_process(rq->migration_thread);
6012 wait_for_completion(&req.done);
6013 tlb_migrate_finish(p->mm);
6017 task_rq_unlock(rq, &flags);
6021 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6024 * Move (not current) task off this cpu, onto dest cpu. We're doing
6025 * this because either it can't run here any more (set_cpus_allowed()
6026 * away from this CPU, or CPU going down), or because we're
6027 * attempting to rebalance this task on exec (sched_exec).
6029 * So we race with normal scheduler movements, but that's OK, as long
6030 * as the task is no longer on this CPU.
6032 * Returns non-zero if task was successfully migrated.
6034 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6036 struct rq *rq_dest, *rq_src;
6039 if (unlikely(!cpu_active(dest_cpu)))
6042 rq_src = cpu_rq(src_cpu);
6043 rq_dest = cpu_rq(dest_cpu);
6045 double_rq_lock(rq_src, rq_dest);
6046 /* Already moved. */
6047 if (task_cpu(p) != src_cpu)
6049 /* Affinity changed (again). */
6050 if (!cpu_isset(dest_cpu, p->cpus_allowed))
6053 on_rq = p->se.on_rq;
6055 deactivate_task(rq_src, p, 0);
6057 set_task_cpu(p, dest_cpu);
6059 activate_task(rq_dest, p, 0);
6060 check_preempt_curr(rq_dest, p, 0);
6065 double_rq_unlock(rq_src, rq_dest);
6070 * migration_thread - this is a highprio system thread that performs
6071 * thread migration by bumping thread off CPU then 'pushing' onto
6074 static int migration_thread(void *data)
6076 int cpu = (long)data;
6080 BUG_ON(rq->migration_thread != current);
6082 set_current_state(TASK_INTERRUPTIBLE);
6083 while (!kthread_should_stop()) {
6084 struct migration_req *req;
6085 struct list_head *head;
6087 spin_lock_irq(&rq->lock);
6089 if (cpu_is_offline(cpu)) {
6090 spin_unlock_irq(&rq->lock);
6094 if (rq->active_balance) {
6095 active_load_balance(rq, cpu);
6096 rq->active_balance = 0;
6099 head = &rq->migration_queue;
6101 if (list_empty(head)) {
6102 spin_unlock_irq(&rq->lock);
6104 set_current_state(TASK_INTERRUPTIBLE);
6107 req = list_entry(head->next, struct migration_req, list);
6108 list_del_init(head->next);
6110 spin_unlock(&rq->lock);
6111 __migrate_task(req->task, cpu, req->dest_cpu);
6114 complete(&req->done);
6116 __set_current_state(TASK_RUNNING);
6120 /* Wait for kthread_stop */
6121 set_current_state(TASK_INTERRUPTIBLE);
6122 while (!kthread_should_stop()) {
6124 set_current_state(TASK_INTERRUPTIBLE);
6126 __set_current_state(TASK_RUNNING);
6130 #ifdef CONFIG_HOTPLUG_CPU
6132 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6136 local_irq_disable();
6137 ret = __migrate_task(p, src_cpu, dest_cpu);
6143 * Figure out where task on dead CPU should go, use force if necessary.
6144 * NOTE: interrupts should be disabled by the caller
6146 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6148 unsigned long flags;
6155 mask = node_to_cpumask(cpu_to_node(dead_cpu));
6156 cpus_and(mask, mask, p->cpus_allowed);
6157 dest_cpu = any_online_cpu(mask);
6159 /* On any allowed CPU? */
6160 if (dest_cpu >= nr_cpu_ids)
6161 dest_cpu = any_online_cpu(p->cpus_allowed);
6163 /* No more Mr. Nice Guy. */
6164 if (dest_cpu >= nr_cpu_ids) {
6165 cpumask_t cpus_allowed;
6167 cpuset_cpus_allowed_locked(p, &cpus_allowed);
6169 * Try to stay on the same cpuset, where the
6170 * current cpuset may be a subset of all cpus.
6171 * The cpuset_cpus_allowed_locked() variant of
6172 * cpuset_cpus_allowed() will not block. It must be
6173 * called within calls to cpuset_lock/cpuset_unlock.
6175 rq = task_rq_lock(p, &flags);
6176 p->cpus_allowed = cpus_allowed;
6177 dest_cpu = any_online_cpu(p->cpus_allowed);
6178 task_rq_unlock(rq, &flags);
6181 * Don't tell them about moving exiting tasks or
6182 * kernel threads (both mm NULL), since they never
6185 if (p->mm && printk_ratelimit()) {
6186 printk(KERN_INFO "process %d (%s) no "
6187 "longer affine to cpu%d\n",
6188 task_pid_nr(p), p->comm, dead_cpu);
6191 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
6195 * While a dead CPU has no uninterruptible tasks queued at this point,
6196 * it might still have a nonzero ->nr_uninterruptible counter, because
6197 * for performance reasons the counter is not stricly tracking tasks to
6198 * their home CPUs. So we just add the counter to another CPU's counter,
6199 * to keep the global sum constant after CPU-down:
6201 static void migrate_nr_uninterruptible(struct rq *rq_src)
6203 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
6204 unsigned long flags;
6206 local_irq_save(flags);
6207 double_rq_lock(rq_src, rq_dest);
6208 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6209 rq_src->nr_uninterruptible = 0;
6210 double_rq_unlock(rq_src, rq_dest);
6211 local_irq_restore(flags);
6214 /* Run through task list and migrate tasks from the dead cpu. */
6215 static void migrate_live_tasks(int src_cpu)
6217 struct task_struct *p, *t;
6219 read_lock(&tasklist_lock);
6221 do_each_thread(t, p) {
6225 if (task_cpu(p) == src_cpu)
6226 move_task_off_dead_cpu(src_cpu, p);
6227 } while_each_thread(t, p);
6229 read_unlock(&tasklist_lock);
6233 * Schedules idle task to be the next runnable task on current CPU.
6234 * It does so by boosting its priority to highest possible.
6235 * Used by CPU offline code.
6237 void sched_idle_next(void)
6239 int this_cpu = smp_processor_id();
6240 struct rq *rq = cpu_rq(this_cpu);
6241 struct task_struct *p = rq->idle;
6242 unsigned long flags;
6244 /* cpu has to be offline */
6245 BUG_ON(cpu_online(this_cpu));
6248 * Strictly not necessary since rest of the CPUs are stopped by now
6249 * and interrupts disabled on the current cpu.
6251 spin_lock_irqsave(&rq->lock, flags);
6253 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6255 update_rq_clock(rq);
6256 activate_task(rq, p, 0);
6258 spin_unlock_irqrestore(&rq->lock, flags);
6262 * Ensures that the idle task is using init_mm right before its cpu goes
6265 void idle_task_exit(void)
6267 struct mm_struct *mm = current->active_mm;
6269 BUG_ON(cpu_online(smp_processor_id()));
6272 switch_mm(mm, &init_mm, current);
6276 /* called under rq->lock with disabled interrupts */
6277 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6279 struct rq *rq = cpu_rq(dead_cpu);
6281 /* Must be exiting, otherwise would be on tasklist. */
6282 BUG_ON(!p->exit_state);
6284 /* Cannot have done final schedule yet: would have vanished. */
6285 BUG_ON(p->state == TASK_DEAD);
6290 * Drop lock around migration; if someone else moves it,
6291 * that's OK. No task can be added to this CPU, so iteration is
6294 spin_unlock_irq(&rq->lock);
6295 move_task_off_dead_cpu(dead_cpu, p);
6296 spin_lock_irq(&rq->lock);
6301 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6302 static void migrate_dead_tasks(unsigned int dead_cpu)
6304 struct rq *rq = cpu_rq(dead_cpu);
6305 struct task_struct *next;
6308 if (!rq->nr_running)
6310 update_rq_clock(rq);
6311 next = pick_next_task(rq, rq->curr);
6314 next->sched_class->put_prev_task(rq, next);
6315 migrate_dead(dead_cpu, next);
6319 #endif /* CONFIG_HOTPLUG_CPU */
6321 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6323 static struct ctl_table sd_ctl_dir[] = {
6325 .procname = "sched_domain",
6331 static struct ctl_table sd_ctl_root[] = {
6333 .ctl_name = CTL_KERN,
6334 .procname = "kernel",
6336 .child = sd_ctl_dir,
6341 static struct ctl_table *sd_alloc_ctl_entry(int n)
6343 struct ctl_table *entry =
6344 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6349 static void sd_free_ctl_entry(struct ctl_table **tablep)
6351 struct ctl_table *entry;
6354 * In the intermediate directories, both the child directory and
6355 * procname are dynamically allocated and could fail but the mode
6356 * will always be set. In the lowest directory the names are
6357 * static strings and all have proc handlers.
6359 for (entry = *tablep; entry->mode; entry++) {
6361 sd_free_ctl_entry(&entry->child);
6362 if (entry->proc_handler == NULL)
6363 kfree(entry->procname);
6371 set_table_entry(struct ctl_table *entry,
6372 const char *procname, void *data, int maxlen,
6373 mode_t mode, proc_handler *proc_handler)
6375 entry->procname = procname;
6377 entry->maxlen = maxlen;
6379 entry->proc_handler = proc_handler;
6382 static struct ctl_table *
6383 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6385 struct ctl_table *table = sd_alloc_ctl_entry(13);
6390 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6391 sizeof(long), 0644, proc_doulongvec_minmax);
6392 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6393 sizeof(long), 0644, proc_doulongvec_minmax);
6394 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6395 sizeof(int), 0644, proc_dointvec_minmax);
6396 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6397 sizeof(int), 0644, proc_dointvec_minmax);
6398 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6399 sizeof(int), 0644, proc_dointvec_minmax);
6400 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6401 sizeof(int), 0644, proc_dointvec_minmax);
6402 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6403 sizeof(int), 0644, proc_dointvec_minmax);
6404 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6405 sizeof(int), 0644, proc_dointvec_minmax);
6406 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6407 sizeof(int), 0644, proc_dointvec_minmax);
6408 set_table_entry(&table[9], "cache_nice_tries",
6409 &sd->cache_nice_tries,
6410 sizeof(int), 0644, proc_dointvec_minmax);
6411 set_table_entry(&table[10], "flags", &sd->flags,
6412 sizeof(int), 0644, proc_dointvec_minmax);
6413 set_table_entry(&table[11], "name", sd->name,
6414 CORENAME_MAX_SIZE, 0444, proc_dostring);
6415 /* &table[12] is terminator */
6420 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6422 struct ctl_table *entry, *table;
6423 struct sched_domain *sd;
6424 int domain_num = 0, i;
6427 for_each_domain(cpu, sd)
6429 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6434 for_each_domain(cpu, sd) {
6435 snprintf(buf, 32, "domain%d", i);
6436 entry->procname = kstrdup(buf, GFP_KERNEL);
6438 entry->child = sd_alloc_ctl_domain_table(sd);
6445 static struct ctl_table_header *sd_sysctl_header;
6446 static void register_sched_domain_sysctl(void)
6448 int i, cpu_num = num_online_cpus();
6449 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6452 WARN_ON(sd_ctl_dir[0].child);
6453 sd_ctl_dir[0].child = entry;
6458 for_each_online_cpu(i) {
6459 snprintf(buf, 32, "cpu%d", i);
6460 entry->procname = kstrdup(buf, GFP_KERNEL);
6462 entry->child = sd_alloc_ctl_cpu_table(i);
6466 WARN_ON(sd_sysctl_header);
6467 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6470 /* may be called multiple times per register */
6471 static void unregister_sched_domain_sysctl(void)
6473 if (sd_sysctl_header)
6474 unregister_sysctl_table(sd_sysctl_header);
6475 sd_sysctl_header = NULL;
6476 if (sd_ctl_dir[0].child)
6477 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6480 static void register_sched_domain_sysctl(void)
6483 static void unregister_sched_domain_sysctl(void)
6488 static void set_rq_online(struct rq *rq)
6491 const struct sched_class *class;
6493 cpu_set(rq->cpu, rq->rd->online);
6496 for_each_class(class) {
6497 if (class->rq_online)
6498 class->rq_online(rq);
6503 static void set_rq_offline(struct rq *rq)
6506 const struct sched_class *class;
6508 for_each_class(class) {
6509 if (class->rq_offline)
6510 class->rq_offline(rq);
6513 cpu_clear(rq->cpu, rq->rd->online);
6519 * migration_call - callback that gets triggered when a CPU is added.
6520 * Here we can start up the necessary migration thread for the new CPU.
6522 static int __cpuinit
6523 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6525 struct task_struct *p;
6526 int cpu = (long)hcpu;
6527 unsigned long flags;
6532 case CPU_UP_PREPARE:
6533 case CPU_UP_PREPARE_FROZEN:
6534 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6537 kthread_bind(p, cpu);
6538 /* Must be high prio: stop_machine expects to yield to it. */
6539 rq = task_rq_lock(p, &flags);
6540 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6541 task_rq_unlock(rq, &flags);
6542 cpu_rq(cpu)->migration_thread = p;
6546 case CPU_ONLINE_FROZEN:
6547 /* Strictly unnecessary, as first user will wake it. */
6548 wake_up_process(cpu_rq(cpu)->migration_thread);
6550 /* Update our root-domain */
6552 spin_lock_irqsave(&rq->lock, flags);
6554 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6558 spin_unlock_irqrestore(&rq->lock, flags);
6561 #ifdef CONFIG_HOTPLUG_CPU
6562 case CPU_UP_CANCELED:
6563 case CPU_UP_CANCELED_FROZEN:
6564 if (!cpu_rq(cpu)->migration_thread)
6566 /* Unbind it from offline cpu so it can run. Fall thru. */
6567 kthread_bind(cpu_rq(cpu)->migration_thread,
6568 any_online_cpu(cpu_online_map));
6569 kthread_stop(cpu_rq(cpu)->migration_thread);
6570 cpu_rq(cpu)->migration_thread = NULL;
6574 case CPU_DEAD_FROZEN:
6575 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6576 migrate_live_tasks(cpu);
6578 kthread_stop(rq->migration_thread);
6579 rq->migration_thread = NULL;
6580 /* Idle task back to normal (off runqueue, low prio) */
6581 spin_lock_irq(&rq->lock);
6582 update_rq_clock(rq);
6583 deactivate_task(rq, rq->idle, 0);
6584 rq->idle->static_prio = MAX_PRIO;
6585 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6586 rq->idle->sched_class = &idle_sched_class;
6587 migrate_dead_tasks(cpu);
6588 spin_unlock_irq(&rq->lock);
6590 migrate_nr_uninterruptible(rq);
6591 BUG_ON(rq->nr_running != 0);
6594 * No need to migrate the tasks: it was best-effort if
6595 * they didn't take sched_hotcpu_mutex. Just wake up
6598 spin_lock_irq(&rq->lock);
6599 while (!list_empty(&rq->migration_queue)) {
6600 struct migration_req *req;
6602 req = list_entry(rq->migration_queue.next,
6603 struct migration_req, list);
6604 list_del_init(&req->list);
6605 complete(&req->done);
6607 spin_unlock_irq(&rq->lock);
6611 case CPU_DYING_FROZEN:
6612 /* Update our root-domain */
6614 spin_lock_irqsave(&rq->lock, flags);
6616 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6619 spin_unlock_irqrestore(&rq->lock, flags);
6626 /* Register at highest priority so that task migration (migrate_all_tasks)
6627 * happens before everything else.
6629 static struct notifier_block __cpuinitdata migration_notifier = {
6630 .notifier_call = migration_call,
6634 static int __init migration_init(void)
6636 void *cpu = (void *)(long)smp_processor_id();
6639 /* Start one for the boot CPU: */
6640 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6641 BUG_ON(err == NOTIFY_BAD);
6642 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6643 register_cpu_notifier(&migration_notifier);
6647 early_initcall(migration_init);
6652 #ifdef CONFIG_SCHED_DEBUG
6654 static inline const char *sd_level_to_string(enum sched_domain_level lvl)
6667 case SD_LV_ALLNODES:
6676 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6677 cpumask_t *groupmask)
6679 struct sched_group *group = sd->groups;
6682 cpulist_scnprintf(str, sizeof(str), sd->span);
6683 cpus_clear(*groupmask);
6685 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6687 if (!(sd->flags & SD_LOAD_BALANCE)) {
6688 printk("does not load-balance\n");
6690 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6695 printk(KERN_CONT "span %s level %s\n",
6696 str, sd_level_to_string(sd->level));
6698 if (!cpu_isset(cpu, sd->span)) {
6699 printk(KERN_ERR "ERROR: domain->span does not contain "
6702 if (!cpu_isset(cpu, group->cpumask)) {
6703 printk(KERN_ERR "ERROR: domain->groups does not contain"
6707 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6711 printk(KERN_ERR "ERROR: group is NULL\n");
6715 if (!group->__cpu_power) {
6716 printk(KERN_CONT "\n");
6717 printk(KERN_ERR "ERROR: domain->cpu_power not "
6722 if (!cpus_weight(group->cpumask)) {
6723 printk(KERN_CONT "\n");
6724 printk(KERN_ERR "ERROR: empty group\n");
6728 if (cpus_intersects(*groupmask, group->cpumask)) {
6729 printk(KERN_CONT "\n");
6730 printk(KERN_ERR "ERROR: repeated CPUs\n");
6734 cpus_or(*groupmask, *groupmask, group->cpumask);
6736 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6737 printk(KERN_CONT " %s", str);
6739 group = group->next;
6740 } while (group != sd->groups);
6741 printk(KERN_CONT "\n");
6743 if (!cpus_equal(sd->span, *groupmask))
6744 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6746 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6747 printk(KERN_ERR "ERROR: parent span is not a superset "
6748 "of domain->span\n");
6752 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6754 cpumask_t *groupmask;
6758 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6762 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6764 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6766 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6771 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6780 #else /* !CONFIG_SCHED_DEBUG */
6781 # define sched_domain_debug(sd, cpu) do { } while (0)
6782 #endif /* CONFIG_SCHED_DEBUG */
6784 static int sd_degenerate(struct sched_domain *sd)
6786 if (cpus_weight(sd->span) == 1)
6789 /* Following flags need at least 2 groups */
6790 if (sd->flags & (SD_LOAD_BALANCE |
6791 SD_BALANCE_NEWIDLE |
6795 SD_SHARE_PKG_RESOURCES)) {
6796 if (sd->groups != sd->groups->next)
6800 /* Following flags don't use groups */
6801 if (sd->flags & (SD_WAKE_IDLE |
6810 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6812 unsigned long cflags = sd->flags, pflags = parent->flags;
6814 if (sd_degenerate(parent))
6817 if (!cpus_equal(sd->span, parent->span))
6820 /* Does parent contain flags not in child? */
6821 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6822 if (cflags & SD_WAKE_AFFINE)
6823 pflags &= ~SD_WAKE_BALANCE;
6824 /* Flags needing groups don't count if only 1 group in parent */
6825 if (parent->groups == parent->groups->next) {
6826 pflags &= ~(SD_LOAD_BALANCE |
6827 SD_BALANCE_NEWIDLE |
6831 SD_SHARE_PKG_RESOURCES);
6833 if (~cflags & pflags)
6839 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6841 unsigned long flags;
6843 spin_lock_irqsave(&rq->lock, flags);
6846 struct root_domain *old_rd = rq->rd;
6848 if (cpu_isset(rq->cpu, old_rd->online))
6851 cpu_clear(rq->cpu, old_rd->span);
6853 if (atomic_dec_and_test(&old_rd->refcount))
6857 atomic_inc(&rd->refcount);
6860 cpu_set(rq->cpu, rd->span);
6861 if (cpu_isset(rq->cpu, cpu_online_map))
6864 spin_unlock_irqrestore(&rq->lock, flags);
6867 static void init_rootdomain(struct root_domain *rd)
6869 memset(rd, 0, sizeof(*rd));
6871 cpus_clear(rd->span);
6872 cpus_clear(rd->online);
6874 cpupri_init(&rd->cpupri);
6877 static void init_defrootdomain(void)
6879 init_rootdomain(&def_root_domain);
6880 atomic_set(&def_root_domain.refcount, 1);
6883 static struct root_domain *alloc_rootdomain(void)
6885 struct root_domain *rd;
6887 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6891 init_rootdomain(rd);
6897 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6898 * hold the hotplug lock.
6901 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6903 struct rq *rq = cpu_rq(cpu);
6904 struct sched_domain *tmp;
6906 /* Remove the sched domains which do not contribute to scheduling. */
6907 for (tmp = sd; tmp; ) {
6908 struct sched_domain *parent = tmp->parent;
6912 if (sd_parent_degenerate(tmp, parent)) {
6913 tmp->parent = parent->parent;
6915 parent->parent->child = tmp;
6920 if (sd && sd_degenerate(sd)) {
6926 sched_domain_debug(sd, cpu);
6928 rq_attach_root(rq, rd);
6929 rcu_assign_pointer(rq->sd, sd);
6932 /* cpus with isolated domains */
6933 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6935 /* Setup the mask of cpus configured for isolated domains */
6936 static int __init isolated_cpu_setup(char *str)
6938 static int __initdata ints[NR_CPUS];
6941 str = get_options(str, ARRAY_SIZE(ints), ints);
6942 cpus_clear(cpu_isolated_map);
6943 for (i = 1; i <= ints[0]; i++)
6944 if (ints[i] < NR_CPUS)
6945 cpu_set(ints[i], cpu_isolated_map);
6949 __setup("isolcpus=", isolated_cpu_setup);
6952 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6953 * to a function which identifies what group(along with sched group) a CPU
6954 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6955 * (due to the fact that we keep track of groups covered with a cpumask_t).
6957 * init_sched_build_groups will build a circular linked list of the groups
6958 * covered by the given span, and will set each group's ->cpumask correctly,
6959 * and ->cpu_power to 0.
6962 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6963 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6964 struct sched_group **sg,
6965 cpumask_t *tmpmask),
6966 cpumask_t *covered, cpumask_t *tmpmask)
6968 struct sched_group *first = NULL, *last = NULL;
6971 cpus_clear(*covered);
6973 for_each_cpu_mask_nr(i, *span) {
6974 struct sched_group *sg;
6975 int group = group_fn(i, cpu_map, &sg, tmpmask);
6978 if (cpu_isset(i, *covered))
6981 cpus_clear(sg->cpumask);
6982 sg->__cpu_power = 0;
6984 for_each_cpu_mask_nr(j, *span) {
6985 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6988 cpu_set(j, *covered);
6989 cpu_set(j, sg->cpumask);
7000 #define SD_NODES_PER_DOMAIN 16
7005 * find_next_best_node - find the next node to include in a sched_domain
7006 * @node: node whose sched_domain we're building
7007 * @used_nodes: nodes already in the sched_domain
7009 * Find the next node to include in a given scheduling domain. Simply
7010 * finds the closest node not already in the @used_nodes map.
7012 * Should use nodemask_t.
7014 static int find_next_best_node(int node, nodemask_t *used_nodes)
7016 int i, n, val, min_val, best_node = 0;
7020 for (i = 0; i < nr_node_ids; i++) {
7021 /* Start at @node */
7022 n = (node + i) % nr_node_ids;
7024 if (!nr_cpus_node(n))
7027 /* Skip already used nodes */
7028 if (node_isset(n, *used_nodes))
7031 /* Simple min distance search */
7032 val = node_distance(node, n);
7034 if (val < min_val) {
7040 node_set(best_node, *used_nodes);
7045 * sched_domain_node_span - get a cpumask for a node's sched_domain
7046 * @node: node whose cpumask we're constructing
7047 * @span: resulting cpumask
7049 * Given a node, construct a good cpumask for its sched_domain to span. It
7050 * should be one that prevents unnecessary balancing, but also spreads tasks
7053 static void sched_domain_node_span(int node, cpumask_t *span)
7055 nodemask_t used_nodes;
7056 node_to_cpumask_ptr(nodemask, node);
7060 nodes_clear(used_nodes);
7062 cpus_or(*span, *span, *nodemask);
7063 node_set(node, used_nodes);
7065 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7066 int next_node = find_next_best_node(node, &used_nodes);
7068 node_to_cpumask_ptr_next(nodemask, next_node);
7069 cpus_or(*span, *span, *nodemask);
7072 #endif /* CONFIG_NUMA */
7074 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7077 * SMT sched-domains:
7079 #ifdef CONFIG_SCHED_SMT
7080 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
7081 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
7084 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7088 *sg = &per_cpu(sched_group_cpus, cpu);
7091 #endif /* CONFIG_SCHED_SMT */
7094 * multi-core sched-domains:
7096 #ifdef CONFIG_SCHED_MC
7097 static DEFINE_PER_CPU(struct sched_domain, core_domains);
7098 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
7099 #endif /* CONFIG_SCHED_MC */
7101 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7103 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7108 *mask = per_cpu(cpu_sibling_map, cpu);
7109 cpus_and(*mask, *mask, *cpu_map);
7110 group = first_cpu(*mask);
7112 *sg = &per_cpu(sched_group_core, group);
7115 #elif defined(CONFIG_SCHED_MC)
7117 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7121 *sg = &per_cpu(sched_group_core, cpu);
7126 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
7127 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
7130 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7134 #ifdef CONFIG_SCHED_MC
7135 *mask = cpu_coregroup_map(cpu);
7136 cpus_and(*mask, *mask, *cpu_map);
7137 group = first_cpu(*mask);
7138 #elif defined(CONFIG_SCHED_SMT)
7139 *mask = per_cpu(cpu_sibling_map, cpu);
7140 cpus_and(*mask, *mask, *cpu_map);
7141 group = first_cpu(*mask);
7146 *sg = &per_cpu(sched_group_phys, group);
7152 * The init_sched_build_groups can't handle what we want to do with node
7153 * groups, so roll our own. Now each node has its own list of groups which
7154 * gets dynamically allocated.
7156 static DEFINE_PER_CPU(struct sched_domain, node_domains);
7157 static struct sched_group ***sched_group_nodes_bycpu;
7159 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
7160 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
7162 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
7163 struct sched_group **sg, cpumask_t *nodemask)
7167 *nodemask = node_to_cpumask(cpu_to_node(cpu));
7168 cpus_and(*nodemask, *nodemask, *cpu_map);
7169 group = first_cpu(*nodemask);
7172 *sg = &per_cpu(sched_group_allnodes, group);
7176 static void init_numa_sched_groups_power(struct sched_group *group_head)
7178 struct sched_group *sg = group_head;
7184 for_each_cpu_mask_nr(j, sg->cpumask) {
7185 struct sched_domain *sd;
7187 sd = &per_cpu(phys_domains, j);
7188 if (j != first_cpu(sd->groups->cpumask)) {
7190 * Only add "power" once for each
7196 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7199 } while (sg != group_head);
7201 #endif /* CONFIG_NUMA */
7204 /* Free memory allocated for various sched_group structures */
7205 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7209 for_each_cpu_mask_nr(cpu, *cpu_map) {
7210 struct sched_group **sched_group_nodes
7211 = sched_group_nodes_bycpu[cpu];
7213 if (!sched_group_nodes)
7216 for (i = 0; i < nr_node_ids; i++) {
7217 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7219 *nodemask = node_to_cpumask(i);
7220 cpus_and(*nodemask, *nodemask, *cpu_map);
7221 if (cpus_empty(*nodemask))
7231 if (oldsg != sched_group_nodes[i])
7234 kfree(sched_group_nodes);
7235 sched_group_nodes_bycpu[cpu] = NULL;
7238 #else /* !CONFIG_NUMA */
7239 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7242 #endif /* CONFIG_NUMA */
7245 * Initialize sched groups cpu_power.
7247 * cpu_power indicates the capacity of sched group, which is used while
7248 * distributing the load between different sched groups in a sched domain.
7249 * Typically cpu_power for all the groups in a sched domain will be same unless
7250 * there are asymmetries in the topology. If there are asymmetries, group
7251 * having more cpu_power will pickup more load compared to the group having
7254 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7255 * the maximum number of tasks a group can handle in the presence of other idle
7256 * or lightly loaded groups in the same sched domain.
7258 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7260 struct sched_domain *child;
7261 struct sched_group *group;
7263 WARN_ON(!sd || !sd->groups);
7265 if (cpu != first_cpu(sd->groups->cpumask))
7270 sd->groups->__cpu_power = 0;
7273 * For perf policy, if the groups in child domain share resources
7274 * (for example cores sharing some portions of the cache hierarchy
7275 * or SMT), then set this domain groups cpu_power such that each group
7276 * can handle only one task, when there are other idle groups in the
7277 * same sched domain.
7279 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7281 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7282 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7287 * add cpu_power of each child group to this groups cpu_power
7289 group = child->groups;
7291 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7292 group = group->next;
7293 } while (group != child->groups);
7297 * Initializers for schedule domains
7298 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7301 #ifdef CONFIG_SCHED_DEBUG
7302 # define SD_INIT_NAME(sd, type) sd->name = #type
7304 # define SD_INIT_NAME(sd, type) do { } while (0)
7307 #define SD_INIT(sd, type) sd_init_##type(sd)
7309 #define SD_INIT_FUNC(type) \
7310 static noinline void sd_init_##type(struct sched_domain *sd) \
7312 memset(sd, 0, sizeof(*sd)); \
7313 *sd = SD_##type##_INIT; \
7314 sd->level = SD_LV_##type; \
7315 SD_INIT_NAME(sd, type); \
7320 SD_INIT_FUNC(ALLNODES)
7323 #ifdef CONFIG_SCHED_SMT
7324 SD_INIT_FUNC(SIBLING)
7326 #ifdef CONFIG_SCHED_MC
7331 * To minimize stack usage kmalloc room for cpumasks and share the
7332 * space as the usage in build_sched_domains() dictates. Used only
7333 * if the amount of space is significant.
7336 cpumask_t tmpmask; /* make this one first */
7339 cpumask_t this_sibling_map;
7340 cpumask_t this_core_map;
7342 cpumask_t send_covered;
7345 cpumask_t domainspan;
7347 cpumask_t notcovered;
7352 #define SCHED_CPUMASK_ALLOC 1
7353 #define SCHED_CPUMASK_FREE(v) kfree(v)
7354 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7356 #define SCHED_CPUMASK_ALLOC 0
7357 #define SCHED_CPUMASK_FREE(v)
7358 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7361 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7362 ((unsigned long)(a) + offsetof(struct allmasks, v))
7364 static int default_relax_domain_level = -1;
7366 static int __init setup_relax_domain_level(char *str)
7370 val = simple_strtoul(str, NULL, 0);
7371 if (val < SD_LV_MAX)
7372 default_relax_domain_level = val;
7376 __setup("relax_domain_level=", setup_relax_domain_level);
7378 static void set_domain_attribute(struct sched_domain *sd,
7379 struct sched_domain_attr *attr)
7383 if (!attr || attr->relax_domain_level < 0) {
7384 if (default_relax_domain_level < 0)
7387 request = default_relax_domain_level;
7389 request = attr->relax_domain_level;
7390 if (request < sd->level) {
7391 /* turn off idle balance on this domain */
7392 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7394 /* turn on idle balance on this domain */
7395 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7400 * Build sched domains for a given set of cpus and attach the sched domains
7401 * to the individual cpus
7403 static int __build_sched_domains(const cpumask_t *cpu_map,
7404 struct sched_domain_attr *attr)
7407 struct root_domain *rd;
7408 SCHED_CPUMASK_DECLARE(allmasks);
7411 struct sched_group **sched_group_nodes = NULL;
7412 int sd_allnodes = 0;
7415 * Allocate the per-node list of sched groups
7417 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
7419 if (!sched_group_nodes) {
7420 printk(KERN_WARNING "Can not alloc sched group node list\n");
7425 rd = alloc_rootdomain();
7427 printk(KERN_WARNING "Cannot alloc root domain\n");
7429 kfree(sched_group_nodes);
7434 #if SCHED_CPUMASK_ALLOC
7435 /* get space for all scratch cpumask variables */
7436 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
7438 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7441 kfree(sched_group_nodes);
7446 tmpmask = (cpumask_t *)allmasks;
7450 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7454 * Set up domains for cpus specified by the cpu_map.
7456 for_each_cpu_mask_nr(i, *cpu_map) {
7457 struct sched_domain *sd = NULL, *p;
7458 SCHED_CPUMASK_VAR(nodemask, allmasks);
7460 *nodemask = node_to_cpumask(cpu_to_node(i));
7461 cpus_and(*nodemask, *nodemask, *cpu_map);
7464 if (cpus_weight(*cpu_map) >
7465 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7466 sd = &per_cpu(allnodes_domains, i);
7467 SD_INIT(sd, ALLNODES);
7468 set_domain_attribute(sd, attr);
7469 sd->span = *cpu_map;
7470 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7476 sd = &per_cpu(node_domains, i);
7478 set_domain_attribute(sd, attr);
7479 sched_domain_node_span(cpu_to_node(i), &sd->span);
7483 cpus_and(sd->span, sd->span, *cpu_map);
7487 sd = &per_cpu(phys_domains, i);
7489 set_domain_attribute(sd, attr);
7490 sd->span = *nodemask;
7494 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7496 #ifdef CONFIG_SCHED_MC
7498 sd = &per_cpu(core_domains, i);
7500 set_domain_attribute(sd, attr);
7501 sd->span = cpu_coregroup_map(i);
7502 cpus_and(sd->span, sd->span, *cpu_map);
7505 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7508 #ifdef CONFIG_SCHED_SMT
7510 sd = &per_cpu(cpu_domains, i);
7511 SD_INIT(sd, SIBLING);
7512 set_domain_attribute(sd, attr);
7513 sd->span = per_cpu(cpu_sibling_map, i);
7514 cpus_and(sd->span, sd->span, *cpu_map);
7517 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7521 #ifdef CONFIG_SCHED_SMT
7522 /* Set up CPU (sibling) groups */
7523 for_each_cpu_mask_nr(i, *cpu_map) {
7524 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7525 SCHED_CPUMASK_VAR(send_covered, allmasks);
7527 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7528 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7529 if (i != first_cpu(*this_sibling_map))
7532 init_sched_build_groups(this_sibling_map, cpu_map,
7534 send_covered, tmpmask);
7538 #ifdef CONFIG_SCHED_MC
7539 /* Set up multi-core groups */
7540 for_each_cpu_mask_nr(i, *cpu_map) {
7541 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7542 SCHED_CPUMASK_VAR(send_covered, allmasks);
7544 *this_core_map = cpu_coregroup_map(i);
7545 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7546 if (i != first_cpu(*this_core_map))
7549 init_sched_build_groups(this_core_map, cpu_map,
7551 send_covered, tmpmask);
7555 /* Set up physical groups */
7556 for (i = 0; i < nr_node_ids; i++) {
7557 SCHED_CPUMASK_VAR(nodemask, allmasks);
7558 SCHED_CPUMASK_VAR(send_covered, allmasks);
7560 *nodemask = node_to_cpumask(i);
7561 cpus_and(*nodemask, *nodemask, *cpu_map);
7562 if (cpus_empty(*nodemask))
7565 init_sched_build_groups(nodemask, cpu_map,
7567 send_covered, tmpmask);
7571 /* Set up node groups */
7573 SCHED_CPUMASK_VAR(send_covered, allmasks);
7575 init_sched_build_groups(cpu_map, cpu_map,
7576 &cpu_to_allnodes_group,
7577 send_covered, tmpmask);
7580 for (i = 0; i < nr_node_ids; i++) {
7581 /* Set up node groups */
7582 struct sched_group *sg, *prev;
7583 SCHED_CPUMASK_VAR(nodemask, allmasks);
7584 SCHED_CPUMASK_VAR(domainspan, allmasks);
7585 SCHED_CPUMASK_VAR(covered, allmasks);
7588 *nodemask = node_to_cpumask(i);
7589 cpus_clear(*covered);
7591 cpus_and(*nodemask, *nodemask, *cpu_map);
7592 if (cpus_empty(*nodemask)) {
7593 sched_group_nodes[i] = NULL;
7597 sched_domain_node_span(i, domainspan);
7598 cpus_and(*domainspan, *domainspan, *cpu_map);
7600 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7602 printk(KERN_WARNING "Can not alloc domain group for "
7606 sched_group_nodes[i] = sg;
7607 for_each_cpu_mask_nr(j, *nodemask) {
7608 struct sched_domain *sd;
7610 sd = &per_cpu(node_domains, j);
7613 sg->__cpu_power = 0;
7614 sg->cpumask = *nodemask;
7616 cpus_or(*covered, *covered, *nodemask);
7619 for (j = 0; j < nr_node_ids; j++) {
7620 SCHED_CPUMASK_VAR(notcovered, allmasks);
7621 int n = (i + j) % nr_node_ids;
7622 node_to_cpumask_ptr(pnodemask, n);
7624 cpus_complement(*notcovered, *covered);
7625 cpus_and(*tmpmask, *notcovered, *cpu_map);
7626 cpus_and(*tmpmask, *tmpmask, *domainspan);
7627 if (cpus_empty(*tmpmask))
7630 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7631 if (cpus_empty(*tmpmask))
7634 sg = kmalloc_node(sizeof(struct sched_group),
7638 "Can not alloc domain group for node %d\n", j);
7641 sg->__cpu_power = 0;
7642 sg->cpumask = *tmpmask;
7643 sg->next = prev->next;
7644 cpus_or(*covered, *covered, *tmpmask);
7651 /* Calculate CPU power for physical packages and nodes */
7652 #ifdef CONFIG_SCHED_SMT
7653 for_each_cpu_mask_nr(i, *cpu_map) {
7654 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7656 init_sched_groups_power(i, sd);
7659 #ifdef CONFIG_SCHED_MC
7660 for_each_cpu_mask_nr(i, *cpu_map) {
7661 struct sched_domain *sd = &per_cpu(core_domains, i);
7663 init_sched_groups_power(i, sd);
7667 for_each_cpu_mask_nr(i, *cpu_map) {
7668 struct sched_domain *sd = &per_cpu(phys_domains, i);
7670 init_sched_groups_power(i, sd);
7674 for (i = 0; i < nr_node_ids; i++)
7675 init_numa_sched_groups_power(sched_group_nodes[i]);
7678 struct sched_group *sg;
7680 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7682 init_numa_sched_groups_power(sg);
7686 /* Attach the domains */
7687 for_each_cpu_mask_nr(i, *cpu_map) {
7688 struct sched_domain *sd;
7689 #ifdef CONFIG_SCHED_SMT
7690 sd = &per_cpu(cpu_domains, i);
7691 #elif defined(CONFIG_SCHED_MC)
7692 sd = &per_cpu(core_domains, i);
7694 sd = &per_cpu(phys_domains, i);
7696 cpu_attach_domain(sd, rd, i);
7699 SCHED_CPUMASK_FREE((void *)allmasks);
7704 free_sched_groups(cpu_map, tmpmask);
7705 SCHED_CPUMASK_FREE((void *)allmasks);
7711 static int build_sched_domains(const cpumask_t *cpu_map)
7713 return __build_sched_domains(cpu_map, NULL);
7716 static cpumask_t *doms_cur; /* current sched domains */
7717 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7718 static struct sched_domain_attr *dattr_cur;
7719 /* attribues of custom domains in 'doms_cur' */
7722 * Special case: If a kmalloc of a doms_cur partition (array of
7723 * cpumask_t) fails, then fallback to a single sched domain,
7724 * as determined by the single cpumask_t fallback_doms.
7726 static cpumask_t fallback_doms;
7728 void __attribute__((weak)) arch_update_cpu_topology(void)
7733 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7734 * For now this just excludes isolated cpus, but could be used to
7735 * exclude other special cases in the future.
7737 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7741 arch_update_cpu_topology();
7743 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7745 doms_cur = &fallback_doms;
7746 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7748 err = build_sched_domains(doms_cur);
7749 register_sched_domain_sysctl();
7754 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7757 free_sched_groups(cpu_map, tmpmask);
7761 * Detach sched domains from a group of cpus specified in cpu_map
7762 * These cpus will now be attached to the NULL domain
7764 static void detach_destroy_domains(const cpumask_t *cpu_map)
7769 unregister_sched_domain_sysctl();
7771 for_each_cpu_mask_nr(i, *cpu_map)
7772 cpu_attach_domain(NULL, &def_root_domain, i);
7773 synchronize_sched();
7774 arch_destroy_sched_domains(cpu_map, &tmpmask);
7777 /* handle null as "default" */
7778 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7779 struct sched_domain_attr *new, int idx_new)
7781 struct sched_domain_attr tmp;
7788 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7789 new ? (new + idx_new) : &tmp,
7790 sizeof(struct sched_domain_attr));
7794 * Partition sched domains as specified by the 'ndoms_new'
7795 * cpumasks in the array doms_new[] of cpumasks. This compares
7796 * doms_new[] to the current sched domain partitioning, doms_cur[].
7797 * It destroys each deleted domain and builds each new domain.
7799 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7800 * The masks don't intersect (don't overlap.) We should setup one
7801 * sched domain for each mask. CPUs not in any of the cpumasks will
7802 * not be load balanced. If the same cpumask appears both in the
7803 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7806 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7807 * ownership of it and will kfree it when done with it. If the caller
7808 * failed the kmalloc call, then it can pass in doms_new == NULL,
7809 * and partition_sched_domains() will fallback to the single partition
7810 * 'fallback_doms', it also forces the domains to be rebuilt.
7812 * If doms_new==NULL it will be replaced with cpu_online_map.
7813 * ndoms_new==0 is a special case for destroying existing domains.
7814 * It will not create the default domain.
7816 * Call with hotplug lock held
7818 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7819 struct sched_domain_attr *dattr_new)
7823 mutex_lock(&sched_domains_mutex);
7825 /* always unregister in case we don't destroy any domains */
7826 unregister_sched_domain_sysctl();
7828 n = doms_new ? ndoms_new : 0;
7830 /* Destroy deleted domains */
7831 for (i = 0; i < ndoms_cur; i++) {
7832 for (j = 0; j < n; j++) {
7833 if (cpus_equal(doms_cur[i], doms_new[j])
7834 && dattrs_equal(dattr_cur, i, dattr_new, j))
7837 /* no match - a current sched domain not in new doms_new[] */
7838 detach_destroy_domains(doms_cur + i);
7843 if (doms_new == NULL) {
7845 doms_new = &fallback_doms;
7846 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7850 /* Build new domains */
7851 for (i = 0; i < ndoms_new; i++) {
7852 for (j = 0; j < ndoms_cur; j++) {
7853 if (cpus_equal(doms_new[i], doms_cur[j])
7854 && dattrs_equal(dattr_new, i, dattr_cur, j))
7857 /* no match - add a new doms_new */
7858 __build_sched_domains(doms_new + i,
7859 dattr_new ? dattr_new + i : NULL);
7864 /* Remember the new sched domains */
7865 if (doms_cur != &fallback_doms)
7867 kfree(dattr_cur); /* kfree(NULL) is safe */
7868 doms_cur = doms_new;
7869 dattr_cur = dattr_new;
7870 ndoms_cur = ndoms_new;
7872 register_sched_domain_sysctl();
7874 mutex_unlock(&sched_domains_mutex);
7877 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7878 int arch_reinit_sched_domains(void)
7882 /* Destroy domains first to force the rebuild */
7883 partition_sched_domains(0, NULL, NULL);
7885 rebuild_sched_domains();
7891 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7895 if (buf[0] != '0' && buf[0] != '1')
7899 sched_smt_power_savings = (buf[0] == '1');
7901 sched_mc_power_savings = (buf[0] == '1');
7903 ret = arch_reinit_sched_domains();
7905 return ret ? ret : count;
7908 #ifdef CONFIG_SCHED_MC
7909 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7912 return sprintf(page, "%u\n", sched_mc_power_savings);
7914 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7915 const char *buf, size_t count)
7917 return sched_power_savings_store(buf, count, 0);
7919 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7920 sched_mc_power_savings_show,
7921 sched_mc_power_savings_store);
7924 #ifdef CONFIG_SCHED_SMT
7925 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7928 return sprintf(page, "%u\n", sched_smt_power_savings);
7930 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7931 const char *buf, size_t count)
7933 return sched_power_savings_store(buf, count, 1);
7935 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7936 sched_smt_power_savings_show,
7937 sched_smt_power_savings_store);
7940 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7944 #ifdef CONFIG_SCHED_SMT
7946 err = sysfs_create_file(&cls->kset.kobj,
7947 &attr_sched_smt_power_savings.attr);
7949 #ifdef CONFIG_SCHED_MC
7950 if (!err && mc_capable())
7951 err = sysfs_create_file(&cls->kset.kobj,
7952 &attr_sched_mc_power_savings.attr);
7956 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7958 #ifndef CONFIG_CPUSETS
7960 * Add online and remove offline CPUs from the scheduler domains.
7961 * When cpusets are enabled they take over this function.
7963 static int update_sched_domains(struct notifier_block *nfb,
7964 unsigned long action, void *hcpu)
7968 case CPU_ONLINE_FROZEN:
7970 case CPU_DEAD_FROZEN:
7971 partition_sched_domains(1, NULL, NULL);
7980 static int update_runtime(struct notifier_block *nfb,
7981 unsigned long action, void *hcpu)
7983 int cpu = (int)(long)hcpu;
7986 case CPU_DOWN_PREPARE:
7987 case CPU_DOWN_PREPARE_FROZEN:
7988 disable_runtime(cpu_rq(cpu));
7991 case CPU_DOWN_FAILED:
7992 case CPU_DOWN_FAILED_FROZEN:
7994 case CPU_ONLINE_FROZEN:
7995 enable_runtime(cpu_rq(cpu));
8003 void __init sched_init_smp(void)
8005 cpumask_t non_isolated_cpus;
8007 #if defined(CONFIG_NUMA)
8008 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
8010 BUG_ON(sched_group_nodes_bycpu == NULL);
8013 mutex_lock(&sched_domains_mutex);
8014 arch_init_sched_domains(&cpu_online_map);
8015 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
8016 if (cpus_empty(non_isolated_cpus))
8017 cpu_set(smp_processor_id(), non_isolated_cpus);
8018 mutex_unlock(&sched_domains_mutex);
8021 #ifndef CONFIG_CPUSETS
8022 /* XXX: Theoretical race here - CPU may be hotplugged now */
8023 hotcpu_notifier(update_sched_domains, 0);
8026 /* RT runtime code needs to handle some hotplug events */
8027 hotcpu_notifier(update_runtime, 0);
8031 /* Move init over to a non-isolated CPU */
8032 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
8034 sched_init_granularity();
8037 void __init sched_init_smp(void)
8039 sched_init_granularity();
8041 #endif /* CONFIG_SMP */
8043 int in_sched_functions(unsigned long addr)
8045 return in_lock_functions(addr) ||
8046 (addr >= (unsigned long)__sched_text_start
8047 && addr < (unsigned long)__sched_text_end);
8050 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
8052 cfs_rq->tasks_timeline = RB_ROOT;
8053 INIT_LIST_HEAD(&cfs_rq->tasks);
8054 #ifdef CONFIG_FAIR_GROUP_SCHED
8057 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8060 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8062 struct rt_prio_array *array;
8065 array = &rt_rq->active;
8066 for (i = 0; i < MAX_RT_PRIO; i++) {
8067 INIT_LIST_HEAD(array->queue + i);
8068 __clear_bit(i, array->bitmap);
8070 /* delimiter for bitsearch: */
8071 __set_bit(MAX_RT_PRIO, array->bitmap);
8073 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8074 rt_rq->highest_prio = MAX_RT_PRIO;
8077 rt_rq->rt_nr_migratory = 0;
8078 rt_rq->overloaded = 0;
8082 rt_rq->rt_throttled = 0;
8083 rt_rq->rt_runtime = 0;
8084 spin_lock_init(&rt_rq->rt_runtime_lock);
8086 #ifdef CONFIG_RT_GROUP_SCHED
8087 rt_rq->rt_nr_boosted = 0;
8092 #ifdef CONFIG_FAIR_GROUP_SCHED
8093 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8094 struct sched_entity *se, int cpu, int add,
8095 struct sched_entity *parent)
8097 struct rq *rq = cpu_rq(cpu);
8098 tg->cfs_rq[cpu] = cfs_rq;
8099 init_cfs_rq(cfs_rq, rq);
8102 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8105 /* se could be NULL for init_task_group */
8110 se->cfs_rq = &rq->cfs;
8112 se->cfs_rq = parent->my_q;
8115 se->load.weight = tg->shares;
8116 se->load.inv_weight = 0;
8117 se->parent = parent;
8121 #ifdef CONFIG_RT_GROUP_SCHED
8122 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8123 struct sched_rt_entity *rt_se, int cpu, int add,
8124 struct sched_rt_entity *parent)
8126 struct rq *rq = cpu_rq(cpu);
8128 tg->rt_rq[cpu] = rt_rq;
8129 init_rt_rq(rt_rq, rq);
8131 rt_rq->rt_se = rt_se;
8132 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8134 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8136 tg->rt_se[cpu] = rt_se;
8141 rt_se->rt_rq = &rq->rt;
8143 rt_se->rt_rq = parent->my_q;
8145 rt_se->my_q = rt_rq;
8146 rt_se->parent = parent;
8147 INIT_LIST_HEAD(&rt_se->run_list);
8151 void __init sched_init(void)
8154 unsigned long alloc_size = 0, ptr;
8156 #ifdef CONFIG_FAIR_GROUP_SCHED
8157 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8159 #ifdef CONFIG_RT_GROUP_SCHED
8160 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8162 #ifdef CONFIG_USER_SCHED
8166 * As sched_init() is called before page_alloc is setup,
8167 * we use alloc_bootmem().
8170 ptr = (unsigned long)alloc_bootmem(alloc_size);
8172 #ifdef CONFIG_FAIR_GROUP_SCHED
8173 init_task_group.se = (struct sched_entity **)ptr;
8174 ptr += nr_cpu_ids * sizeof(void **);
8176 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8177 ptr += nr_cpu_ids * sizeof(void **);
8179 #ifdef CONFIG_USER_SCHED
8180 root_task_group.se = (struct sched_entity **)ptr;
8181 ptr += nr_cpu_ids * sizeof(void **);
8183 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8184 ptr += nr_cpu_ids * sizeof(void **);
8185 #endif /* CONFIG_USER_SCHED */
8186 #endif /* CONFIG_FAIR_GROUP_SCHED */
8187 #ifdef CONFIG_RT_GROUP_SCHED
8188 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8189 ptr += nr_cpu_ids * sizeof(void **);
8191 init_task_group.rt_rq = (struct rt_rq **)ptr;
8192 ptr += nr_cpu_ids * sizeof(void **);
8194 #ifdef CONFIG_USER_SCHED
8195 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8196 ptr += nr_cpu_ids * sizeof(void **);
8198 root_task_group.rt_rq = (struct rt_rq **)ptr;
8199 ptr += nr_cpu_ids * sizeof(void **);
8200 #endif /* CONFIG_USER_SCHED */
8201 #endif /* CONFIG_RT_GROUP_SCHED */
8205 init_defrootdomain();
8208 init_rt_bandwidth(&def_rt_bandwidth,
8209 global_rt_period(), global_rt_runtime());
8211 #ifdef CONFIG_RT_GROUP_SCHED
8212 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8213 global_rt_period(), global_rt_runtime());
8214 #ifdef CONFIG_USER_SCHED
8215 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8216 global_rt_period(), RUNTIME_INF);
8217 #endif /* CONFIG_USER_SCHED */
8218 #endif /* CONFIG_RT_GROUP_SCHED */
8220 #ifdef CONFIG_GROUP_SCHED
8221 list_add(&init_task_group.list, &task_groups);
8222 INIT_LIST_HEAD(&init_task_group.children);
8224 #ifdef CONFIG_USER_SCHED
8225 INIT_LIST_HEAD(&root_task_group.children);
8226 init_task_group.parent = &root_task_group;
8227 list_add(&init_task_group.siblings, &root_task_group.children);
8228 #endif /* CONFIG_USER_SCHED */
8229 #endif /* CONFIG_GROUP_SCHED */
8231 for_each_possible_cpu(i) {
8235 spin_lock_init(&rq->lock);
8237 init_cfs_rq(&rq->cfs, rq);
8238 init_rt_rq(&rq->rt, rq);
8239 #ifdef CONFIG_FAIR_GROUP_SCHED
8240 init_task_group.shares = init_task_group_load;
8241 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8242 #ifdef CONFIG_CGROUP_SCHED
8244 * How much cpu bandwidth does init_task_group get?
8246 * In case of task-groups formed thr' the cgroup filesystem, it
8247 * gets 100% of the cpu resources in the system. This overall
8248 * system cpu resource is divided among the tasks of
8249 * init_task_group and its child task-groups in a fair manner,
8250 * based on each entity's (task or task-group's) weight
8251 * (se->load.weight).
8253 * In other words, if init_task_group has 10 tasks of weight
8254 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8255 * then A0's share of the cpu resource is:
8257 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8259 * We achieve this by letting init_task_group's tasks sit
8260 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8262 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8263 #elif defined CONFIG_USER_SCHED
8264 root_task_group.shares = NICE_0_LOAD;
8265 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8267 * In case of task-groups formed thr' the user id of tasks,
8268 * init_task_group represents tasks belonging to root user.
8269 * Hence it forms a sibling of all subsequent groups formed.
8270 * In this case, init_task_group gets only a fraction of overall
8271 * system cpu resource, based on the weight assigned to root
8272 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8273 * by letting tasks of init_task_group sit in a separate cfs_rq
8274 * (init_cfs_rq) and having one entity represent this group of
8275 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8277 init_tg_cfs_entry(&init_task_group,
8278 &per_cpu(init_cfs_rq, i),
8279 &per_cpu(init_sched_entity, i), i, 1,
8280 root_task_group.se[i]);
8283 #endif /* CONFIG_FAIR_GROUP_SCHED */
8285 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8286 #ifdef CONFIG_RT_GROUP_SCHED
8287 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8288 #ifdef CONFIG_CGROUP_SCHED
8289 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8290 #elif defined CONFIG_USER_SCHED
8291 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8292 init_tg_rt_entry(&init_task_group,
8293 &per_cpu(init_rt_rq, i),
8294 &per_cpu(init_sched_rt_entity, i), i, 1,
8295 root_task_group.rt_se[i]);
8299 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8300 rq->cpu_load[j] = 0;
8304 rq->active_balance = 0;
8305 rq->next_balance = jiffies;
8309 rq->migration_thread = NULL;
8310 INIT_LIST_HEAD(&rq->migration_queue);
8311 rq_attach_root(rq, &def_root_domain);
8314 atomic_set(&rq->nr_iowait, 0);
8317 set_load_weight(&init_task);
8319 #ifdef CONFIG_PREEMPT_NOTIFIERS
8320 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8324 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8327 #ifdef CONFIG_RT_MUTEXES
8328 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8332 * The boot idle thread does lazy MMU switching as well:
8334 atomic_inc(&init_mm.mm_count);
8335 enter_lazy_tlb(&init_mm, current);
8338 * Make us the idle thread. Technically, schedule() should not be
8339 * called from this thread, however somewhere below it might be,
8340 * but because we are the idle thread, we just pick up running again
8341 * when this runqueue becomes "idle".
8343 init_idle(current, smp_processor_id());
8345 * During early bootup we pretend to be a normal task:
8347 current->sched_class = &fair_sched_class;
8349 scheduler_running = 1;
8352 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8353 void __might_sleep(char *file, int line)
8356 static unsigned long prev_jiffy; /* ratelimiting */
8358 if ((!in_atomic() && !irqs_disabled()) ||
8359 system_state != SYSTEM_RUNNING || oops_in_progress)
8361 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8363 prev_jiffy = jiffies;
8366 "BUG: sleeping function called from invalid context at %s:%d\n",
8369 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8370 in_atomic(), irqs_disabled(),
8371 current->pid, current->comm);
8373 debug_show_held_locks(current);
8374 if (irqs_disabled())
8375 print_irqtrace_events(current);
8379 EXPORT_SYMBOL(__might_sleep);
8382 #ifdef CONFIG_MAGIC_SYSRQ
8383 static void normalize_task(struct rq *rq, struct task_struct *p)
8387 update_rq_clock(rq);
8388 on_rq = p->se.on_rq;
8390 deactivate_task(rq, p, 0);
8391 __setscheduler(rq, p, SCHED_NORMAL, 0);
8393 activate_task(rq, p, 0);
8394 resched_task(rq->curr);
8398 void normalize_rt_tasks(void)
8400 struct task_struct *g, *p;
8401 unsigned long flags;
8404 read_lock_irqsave(&tasklist_lock, flags);
8405 do_each_thread(g, p) {
8407 * Only normalize user tasks:
8412 p->se.exec_start = 0;
8413 #ifdef CONFIG_SCHEDSTATS
8414 p->se.wait_start = 0;
8415 p->se.sleep_start = 0;
8416 p->se.block_start = 0;
8421 * Renice negative nice level userspace
8424 if (TASK_NICE(p) < 0 && p->mm)
8425 set_user_nice(p, 0);
8429 spin_lock(&p->pi_lock);
8430 rq = __task_rq_lock(p);
8432 normalize_task(rq, p);
8434 __task_rq_unlock(rq);
8435 spin_unlock(&p->pi_lock);
8436 } while_each_thread(g, p);
8438 read_unlock_irqrestore(&tasklist_lock, flags);
8441 #endif /* CONFIG_MAGIC_SYSRQ */
8445 * These functions are only useful for the IA64 MCA handling.
8447 * They can only be called when the whole system has been
8448 * stopped - every CPU needs to be quiescent, and no scheduling
8449 * activity can take place. Using them for anything else would
8450 * be a serious bug, and as a result, they aren't even visible
8451 * under any other configuration.
8455 * curr_task - return the current task for a given cpu.
8456 * @cpu: the processor in question.
8458 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8460 struct task_struct *curr_task(int cpu)
8462 return cpu_curr(cpu);
8466 * set_curr_task - set the current task for a given cpu.
8467 * @cpu: the processor in question.
8468 * @p: the task pointer to set.
8470 * Description: This function must only be used when non-maskable interrupts
8471 * are serviced on a separate stack. It allows the architecture to switch the
8472 * notion of the current task on a cpu in a non-blocking manner. This function
8473 * must be called with all CPU's synchronized, and interrupts disabled, the
8474 * and caller must save the original value of the current task (see
8475 * curr_task() above) and restore that value before reenabling interrupts and
8476 * re-starting the system.
8478 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8480 void set_curr_task(int cpu, struct task_struct *p)
8487 #ifdef CONFIG_FAIR_GROUP_SCHED
8488 static void free_fair_sched_group(struct task_group *tg)
8492 for_each_possible_cpu(i) {
8494 kfree(tg->cfs_rq[i]);
8504 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8506 struct cfs_rq *cfs_rq;
8507 struct sched_entity *se, *parent_se;
8511 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8514 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8518 tg->shares = NICE_0_LOAD;
8520 for_each_possible_cpu(i) {
8523 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8524 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8528 se = kmalloc_node(sizeof(struct sched_entity),
8529 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8533 parent_se = parent ? parent->se[i] : NULL;
8534 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
8543 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8545 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8546 &cpu_rq(cpu)->leaf_cfs_rq_list);
8549 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8551 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8553 #else /* !CONFG_FAIR_GROUP_SCHED */
8554 static inline void free_fair_sched_group(struct task_group *tg)
8559 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8564 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8568 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8571 #endif /* CONFIG_FAIR_GROUP_SCHED */
8573 #ifdef CONFIG_RT_GROUP_SCHED
8574 static void free_rt_sched_group(struct task_group *tg)
8578 destroy_rt_bandwidth(&tg->rt_bandwidth);
8580 for_each_possible_cpu(i) {
8582 kfree(tg->rt_rq[i]);
8584 kfree(tg->rt_se[i]);
8592 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8594 struct rt_rq *rt_rq;
8595 struct sched_rt_entity *rt_se, *parent_se;
8599 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8602 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8606 init_rt_bandwidth(&tg->rt_bandwidth,
8607 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8609 for_each_possible_cpu(i) {
8612 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8613 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8617 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8618 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8622 parent_se = parent ? parent->rt_se[i] : NULL;
8623 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
8632 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8634 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8635 &cpu_rq(cpu)->leaf_rt_rq_list);
8638 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8640 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8642 #else /* !CONFIG_RT_GROUP_SCHED */
8643 static inline void free_rt_sched_group(struct task_group *tg)
8648 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8653 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8657 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8660 #endif /* CONFIG_RT_GROUP_SCHED */
8662 #ifdef CONFIG_GROUP_SCHED
8663 static void free_sched_group(struct task_group *tg)
8665 free_fair_sched_group(tg);
8666 free_rt_sched_group(tg);
8670 /* allocate runqueue etc for a new task group */
8671 struct task_group *sched_create_group(struct task_group *parent)
8673 struct task_group *tg;
8674 unsigned long flags;
8677 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8679 return ERR_PTR(-ENOMEM);
8681 if (!alloc_fair_sched_group(tg, parent))
8684 if (!alloc_rt_sched_group(tg, parent))
8687 spin_lock_irqsave(&task_group_lock, flags);
8688 for_each_possible_cpu(i) {
8689 register_fair_sched_group(tg, i);
8690 register_rt_sched_group(tg, i);
8692 list_add_rcu(&tg->list, &task_groups);
8694 WARN_ON(!parent); /* root should already exist */
8696 tg->parent = parent;
8697 INIT_LIST_HEAD(&tg->children);
8698 list_add_rcu(&tg->siblings, &parent->children);
8699 spin_unlock_irqrestore(&task_group_lock, flags);
8704 free_sched_group(tg);
8705 return ERR_PTR(-ENOMEM);
8708 /* rcu callback to free various structures associated with a task group */
8709 static void free_sched_group_rcu(struct rcu_head *rhp)
8711 /* now it should be safe to free those cfs_rqs */
8712 free_sched_group(container_of(rhp, struct task_group, rcu));
8715 /* Destroy runqueue etc associated with a task group */
8716 void sched_destroy_group(struct task_group *tg)
8718 unsigned long flags;
8721 spin_lock_irqsave(&task_group_lock, flags);
8722 for_each_possible_cpu(i) {
8723 unregister_fair_sched_group(tg, i);
8724 unregister_rt_sched_group(tg, i);
8726 list_del_rcu(&tg->list);
8727 list_del_rcu(&tg->siblings);
8728 spin_unlock_irqrestore(&task_group_lock, flags);
8730 /* wait for possible concurrent references to cfs_rqs complete */
8731 call_rcu(&tg->rcu, free_sched_group_rcu);
8734 /* change task's runqueue when it moves between groups.
8735 * The caller of this function should have put the task in its new group
8736 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8737 * reflect its new group.
8739 void sched_move_task(struct task_struct *tsk)
8742 unsigned long flags;
8745 rq = task_rq_lock(tsk, &flags);
8747 update_rq_clock(rq);
8749 running = task_current(rq, tsk);
8750 on_rq = tsk->se.on_rq;
8753 dequeue_task(rq, tsk, 0);
8754 if (unlikely(running))
8755 tsk->sched_class->put_prev_task(rq, tsk);
8757 set_task_rq(tsk, task_cpu(tsk));
8759 #ifdef CONFIG_FAIR_GROUP_SCHED
8760 if (tsk->sched_class->moved_group)
8761 tsk->sched_class->moved_group(tsk);
8764 if (unlikely(running))
8765 tsk->sched_class->set_curr_task(rq);
8767 enqueue_task(rq, tsk, 0);
8769 task_rq_unlock(rq, &flags);
8771 #endif /* CONFIG_GROUP_SCHED */
8773 #ifdef CONFIG_FAIR_GROUP_SCHED
8774 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8776 struct cfs_rq *cfs_rq = se->cfs_rq;
8781 dequeue_entity(cfs_rq, se, 0);
8783 se->load.weight = shares;
8784 se->load.inv_weight = 0;
8787 enqueue_entity(cfs_rq, se, 0);
8790 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8792 struct cfs_rq *cfs_rq = se->cfs_rq;
8793 struct rq *rq = cfs_rq->rq;
8794 unsigned long flags;
8796 spin_lock_irqsave(&rq->lock, flags);
8797 __set_se_shares(se, shares);
8798 spin_unlock_irqrestore(&rq->lock, flags);
8801 static DEFINE_MUTEX(shares_mutex);
8803 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8806 unsigned long flags;
8809 * We can't change the weight of the root cgroup.
8814 if (shares < MIN_SHARES)
8815 shares = MIN_SHARES;
8816 else if (shares > MAX_SHARES)
8817 shares = MAX_SHARES;
8819 mutex_lock(&shares_mutex);
8820 if (tg->shares == shares)
8823 spin_lock_irqsave(&task_group_lock, flags);
8824 for_each_possible_cpu(i)
8825 unregister_fair_sched_group(tg, i);
8826 list_del_rcu(&tg->siblings);
8827 spin_unlock_irqrestore(&task_group_lock, flags);
8829 /* wait for any ongoing reference to this group to finish */
8830 synchronize_sched();
8833 * Now we are free to modify the group's share on each cpu
8834 * w/o tripping rebalance_share or load_balance_fair.
8836 tg->shares = shares;
8837 for_each_possible_cpu(i) {
8841 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8842 set_se_shares(tg->se[i], shares);
8846 * Enable load balance activity on this group, by inserting it back on
8847 * each cpu's rq->leaf_cfs_rq_list.
8849 spin_lock_irqsave(&task_group_lock, flags);
8850 for_each_possible_cpu(i)
8851 register_fair_sched_group(tg, i);
8852 list_add_rcu(&tg->siblings, &tg->parent->children);
8853 spin_unlock_irqrestore(&task_group_lock, flags);
8855 mutex_unlock(&shares_mutex);
8859 unsigned long sched_group_shares(struct task_group *tg)
8865 #ifdef CONFIG_RT_GROUP_SCHED
8867 * Ensure that the real time constraints are schedulable.
8869 static DEFINE_MUTEX(rt_constraints_mutex);
8871 static unsigned long to_ratio(u64 period, u64 runtime)
8873 if (runtime == RUNTIME_INF)
8876 return div64_u64(runtime << 20, period);
8879 /* Must be called with tasklist_lock held */
8880 static inline int tg_has_rt_tasks(struct task_group *tg)
8882 struct task_struct *g, *p;
8884 do_each_thread(g, p) {
8885 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8887 } while_each_thread(g, p);
8892 struct rt_schedulable_data {
8893 struct task_group *tg;
8898 static int tg_schedulable(struct task_group *tg, void *data)
8900 struct rt_schedulable_data *d = data;
8901 struct task_group *child;
8902 unsigned long total, sum = 0;
8903 u64 period, runtime;
8905 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8906 runtime = tg->rt_bandwidth.rt_runtime;
8909 period = d->rt_period;
8910 runtime = d->rt_runtime;
8914 * Cannot have more runtime than the period.
8916 if (runtime > period && runtime != RUNTIME_INF)
8920 * Ensure we don't starve existing RT tasks.
8922 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8925 total = to_ratio(period, runtime);
8928 * Nobody can have more than the global setting allows.
8930 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8934 * The sum of our children's runtime should not exceed our own.
8936 list_for_each_entry_rcu(child, &tg->children, siblings) {
8937 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8938 runtime = child->rt_bandwidth.rt_runtime;
8940 if (child == d->tg) {
8941 period = d->rt_period;
8942 runtime = d->rt_runtime;
8945 sum += to_ratio(period, runtime);
8954 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8956 struct rt_schedulable_data data = {
8958 .rt_period = period,
8959 .rt_runtime = runtime,
8962 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8965 static int tg_set_bandwidth(struct task_group *tg,
8966 u64 rt_period, u64 rt_runtime)
8970 mutex_lock(&rt_constraints_mutex);
8971 read_lock(&tasklist_lock);
8972 err = __rt_schedulable(tg, rt_period, rt_runtime);
8976 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8977 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8978 tg->rt_bandwidth.rt_runtime = rt_runtime;
8980 for_each_possible_cpu(i) {
8981 struct rt_rq *rt_rq = tg->rt_rq[i];
8983 spin_lock(&rt_rq->rt_runtime_lock);
8984 rt_rq->rt_runtime = rt_runtime;
8985 spin_unlock(&rt_rq->rt_runtime_lock);
8987 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8989 read_unlock(&tasklist_lock);
8990 mutex_unlock(&rt_constraints_mutex);
8995 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8997 u64 rt_runtime, rt_period;
8999 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9000 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
9001 if (rt_runtime_us < 0)
9002 rt_runtime = RUNTIME_INF;
9004 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9007 long sched_group_rt_runtime(struct task_group *tg)
9011 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
9014 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
9015 do_div(rt_runtime_us, NSEC_PER_USEC);
9016 return rt_runtime_us;
9019 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
9021 u64 rt_runtime, rt_period;
9023 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
9024 rt_runtime = tg->rt_bandwidth.rt_runtime;
9029 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9032 long sched_group_rt_period(struct task_group *tg)
9036 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9037 do_div(rt_period_us, NSEC_PER_USEC);
9038 return rt_period_us;
9041 static int sched_rt_global_constraints(void)
9043 u64 runtime, period;
9046 if (sysctl_sched_rt_period <= 0)
9049 runtime = global_rt_runtime();
9050 period = global_rt_period();
9053 * Sanity check on the sysctl variables.
9055 if (runtime > period && runtime != RUNTIME_INF)
9058 mutex_lock(&rt_constraints_mutex);
9059 read_lock(&tasklist_lock);
9060 ret = __rt_schedulable(NULL, 0, 0);
9061 read_unlock(&tasklist_lock);
9062 mutex_unlock(&rt_constraints_mutex);
9066 #else /* !CONFIG_RT_GROUP_SCHED */
9067 static int sched_rt_global_constraints(void)
9069 unsigned long flags;
9072 if (sysctl_sched_rt_period <= 0)
9075 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9076 for_each_possible_cpu(i) {
9077 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9079 spin_lock(&rt_rq->rt_runtime_lock);
9080 rt_rq->rt_runtime = global_rt_runtime();
9081 spin_unlock(&rt_rq->rt_runtime_lock);
9083 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9087 #endif /* CONFIG_RT_GROUP_SCHED */
9089 int sched_rt_handler(struct ctl_table *table, int write,
9090 struct file *filp, void __user *buffer, size_t *lenp,
9094 int old_period, old_runtime;
9095 static DEFINE_MUTEX(mutex);
9098 old_period = sysctl_sched_rt_period;
9099 old_runtime = sysctl_sched_rt_runtime;
9101 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
9103 if (!ret && write) {
9104 ret = sched_rt_global_constraints();
9106 sysctl_sched_rt_period = old_period;
9107 sysctl_sched_rt_runtime = old_runtime;
9109 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9110 def_rt_bandwidth.rt_period =
9111 ns_to_ktime(global_rt_period());
9114 mutex_unlock(&mutex);
9119 #ifdef CONFIG_CGROUP_SCHED
9121 /* return corresponding task_group object of a cgroup */
9122 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9124 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9125 struct task_group, css);
9128 static struct cgroup_subsys_state *
9129 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9131 struct task_group *tg, *parent;
9133 if (!cgrp->parent) {
9134 /* This is early initialization for the top cgroup */
9135 return &init_task_group.css;
9138 parent = cgroup_tg(cgrp->parent);
9139 tg = sched_create_group(parent);
9141 return ERR_PTR(-ENOMEM);
9147 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9149 struct task_group *tg = cgroup_tg(cgrp);
9151 sched_destroy_group(tg);
9155 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9156 struct task_struct *tsk)
9158 #ifdef CONFIG_RT_GROUP_SCHED
9159 /* Don't accept realtime tasks when there is no way for them to run */
9160 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
9163 /* We don't support RT-tasks being in separate groups */
9164 if (tsk->sched_class != &fair_sched_class)
9172 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9173 struct cgroup *old_cont, struct task_struct *tsk)
9175 sched_move_task(tsk);
9178 #ifdef CONFIG_FAIR_GROUP_SCHED
9179 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9182 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9185 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9187 struct task_group *tg = cgroup_tg(cgrp);
9189 return (u64) tg->shares;
9191 #endif /* CONFIG_FAIR_GROUP_SCHED */
9193 #ifdef CONFIG_RT_GROUP_SCHED
9194 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9197 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9200 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9202 return sched_group_rt_runtime(cgroup_tg(cgrp));
9205 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9208 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9211 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9213 return sched_group_rt_period(cgroup_tg(cgrp));
9215 #endif /* CONFIG_RT_GROUP_SCHED */
9217 static struct cftype cpu_files[] = {
9218 #ifdef CONFIG_FAIR_GROUP_SCHED
9221 .read_u64 = cpu_shares_read_u64,
9222 .write_u64 = cpu_shares_write_u64,
9225 #ifdef CONFIG_RT_GROUP_SCHED
9227 .name = "rt_runtime_us",
9228 .read_s64 = cpu_rt_runtime_read,
9229 .write_s64 = cpu_rt_runtime_write,
9232 .name = "rt_period_us",
9233 .read_u64 = cpu_rt_period_read_uint,
9234 .write_u64 = cpu_rt_period_write_uint,
9239 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9241 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9244 struct cgroup_subsys cpu_cgroup_subsys = {
9246 .create = cpu_cgroup_create,
9247 .destroy = cpu_cgroup_destroy,
9248 .can_attach = cpu_cgroup_can_attach,
9249 .attach = cpu_cgroup_attach,
9250 .populate = cpu_cgroup_populate,
9251 .subsys_id = cpu_cgroup_subsys_id,
9255 #endif /* CONFIG_CGROUP_SCHED */
9257 #ifdef CONFIG_CGROUP_CPUACCT
9260 * CPU accounting code for task groups.
9262 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9263 * (balbir@in.ibm.com).
9266 /* track cpu usage of a group of tasks */
9268 struct cgroup_subsys_state css;
9269 /* cpuusage holds pointer to a u64-type object on every cpu */
9273 struct cgroup_subsys cpuacct_subsys;
9275 /* return cpu accounting group corresponding to this container */
9276 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9278 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9279 struct cpuacct, css);
9282 /* return cpu accounting group to which this task belongs */
9283 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9285 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9286 struct cpuacct, css);
9289 /* create a new cpu accounting group */
9290 static struct cgroup_subsys_state *cpuacct_create(
9291 struct cgroup_subsys *ss, struct cgroup *cgrp)
9293 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9296 return ERR_PTR(-ENOMEM);
9298 ca->cpuusage = alloc_percpu(u64);
9299 if (!ca->cpuusage) {
9301 return ERR_PTR(-ENOMEM);
9307 /* destroy an existing cpu accounting group */
9309 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9311 struct cpuacct *ca = cgroup_ca(cgrp);
9313 free_percpu(ca->cpuusage);
9317 /* return total cpu usage (in nanoseconds) of a group */
9318 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9320 struct cpuacct *ca = cgroup_ca(cgrp);
9321 u64 totalcpuusage = 0;
9324 for_each_possible_cpu(i) {
9325 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9328 * Take rq->lock to make 64-bit addition safe on 32-bit
9331 spin_lock_irq(&cpu_rq(i)->lock);
9332 totalcpuusage += *cpuusage;
9333 spin_unlock_irq(&cpu_rq(i)->lock);
9336 return totalcpuusage;
9339 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9342 struct cpuacct *ca = cgroup_ca(cgrp);
9351 for_each_possible_cpu(i) {
9352 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9354 spin_lock_irq(&cpu_rq(i)->lock);
9356 spin_unlock_irq(&cpu_rq(i)->lock);
9362 static struct cftype files[] = {
9365 .read_u64 = cpuusage_read,
9366 .write_u64 = cpuusage_write,
9370 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9372 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9376 * charge this task's execution time to its accounting group.
9378 * called with rq->lock held.
9380 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9384 if (!cpuacct_subsys.active)
9389 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
9391 *cpuusage += cputime;
9395 struct cgroup_subsys cpuacct_subsys = {
9397 .create = cpuacct_create,
9398 .destroy = cpuacct_destroy,
9399 .populate = cpuacct_populate,
9400 .subsys_id = cpuacct_subsys_id,
9402 #endif /* CONFIG_CGROUP_CPUACCT */