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/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/reciprocal_div.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/bootmem.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
75 #include <asm/irq_regs.h>
77 #include "sched_cpupri.h"
80 * Convert user-nice values [ -20 ... 0 ... 19 ]
81 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
84 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
85 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
86 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
89 * 'User priority' is the nice value converted to something we
90 * can work with better when scaling various scheduler parameters,
91 * it's a [ 0 ... 39 ] range.
93 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
94 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
95 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
98 * Helpers for converting nanosecond timing to jiffy resolution
100 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
102 #define NICE_0_LOAD SCHED_LOAD_SCALE
103 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
106 * These are the 'tuning knobs' of the scheduler:
108 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
109 * Timeslices get refilled after they expire.
111 #define DEF_TIMESLICE (100 * HZ / 1000)
114 * single value that denotes runtime == period, ie unlimited time.
116 #define RUNTIME_INF ((u64)~0ULL)
120 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
121 * Since cpu_power is a 'constant', we can use a reciprocal divide.
123 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
125 return reciprocal_divide(load, sg->reciprocal_cpu_power);
129 * Each time a sched group cpu_power is changed,
130 * we must compute its reciprocal value
132 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
134 sg->__cpu_power += val;
135 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
139 static inline int rt_policy(int policy)
141 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
146 static inline int task_has_rt_policy(struct task_struct *p)
148 return rt_policy(p->policy);
152 * This is the priority-queue data structure of the RT scheduling class:
154 struct rt_prio_array {
155 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
156 struct list_head queue[MAX_RT_PRIO];
159 struct rt_bandwidth {
160 /* nests inside the rq lock: */
161 spinlock_t rt_runtime_lock;
164 struct hrtimer rt_period_timer;
167 static struct rt_bandwidth def_rt_bandwidth;
169 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
171 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
173 struct rt_bandwidth *rt_b =
174 container_of(timer, struct rt_bandwidth, rt_period_timer);
180 now = hrtimer_cb_get_time(timer);
181 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
186 idle = do_sched_rt_period_timer(rt_b, overrun);
189 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
193 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
195 rt_b->rt_period = ns_to_ktime(period);
196 rt_b->rt_runtime = runtime;
198 spin_lock_init(&rt_b->rt_runtime_lock);
200 hrtimer_init(&rt_b->rt_period_timer,
201 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
202 rt_b->rt_period_timer.function = sched_rt_period_timer;
203 rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
206 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
210 if (rt_b->rt_runtime == RUNTIME_INF)
213 if (hrtimer_active(&rt_b->rt_period_timer))
216 spin_lock(&rt_b->rt_runtime_lock);
218 if (hrtimer_active(&rt_b->rt_period_timer))
221 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
222 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
223 hrtimer_start(&rt_b->rt_period_timer,
224 rt_b->rt_period_timer.expires,
227 spin_unlock(&rt_b->rt_runtime_lock);
230 #ifdef CONFIG_RT_GROUP_SCHED
231 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
233 hrtimer_cancel(&rt_b->rt_period_timer);
238 * sched_domains_mutex serializes calls to arch_init_sched_domains,
239 * detach_destroy_domains and partition_sched_domains.
241 static DEFINE_MUTEX(sched_domains_mutex);
243 #ifdef CONFIG_GROUP_SCHED
245 #include <linux/cgroup.h>
249 static LIST_HEAD(task_groups);
251 /* task group related information */
253 #ifdef CONFIG_CGROUP_SCHED
254 struct cgroup_subsys_state css;
257 #ifdef CONFIG_FAIR_GROUP_SCHED
258 /* schedulable entities of this group on each cpu */
259 struct sched_entity **se;
260 /* runqueue "owned" by this group on each cpu */
261 struct cfs_rq **cfs_rq;
262 unsigned long shares;
265 #ifdef CONFIG_RT_GROUP_SCHED
266 struct sched_rt_entity **rt_se;
267 struct rt_rq **rt_rq;
269 struct rt_bandwidth rt_bandwidth;
273 struct list_head list;
275 struct task_group *parent;
276 struct list_head siblings;
277 struct list_head children;
280 #ifdef CONFIG_USER_SCHED
284 * Every UID task group (including init_task_group aka UID-0) will
285 * be a child to this group.
287 struct task_group root_task_group;
289 #ifdef CONFIG_FAIR_GROUP_SCHED
290 /* Default task group's sched entity on each cpu */
291 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
292 /* Default task group's cfs_rq on each cpu */
293 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
294 #endif /* CONFIG_FAIR_GROUP_SCHED */
296 #ifdef CONFIG_RT_GROUP_SCHED
297 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
298 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
299 #endif /* CONFIG_RT_GROUP_SCHED */
300 #else /* !CONFIG_FAIR_GROUP_SCHED */
301 #define root_task_group init_task_group
302 #endif /* CONFIG_FAIR_GROUP_SCHED */
304 /* task_group_lock serializes add/remove of task groups and also changes to
305 * a task group's cpu shares.
307 static DEFINE_SPINLOCK(task_group_lock);
309 #ifdef CONFIG_FAIR_GROUP_SCHED
310 #ifdef CONFIG_USER_SCHED
311 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
312 #else /* !CONFIG_USER_SCHED */
313 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
314 #endif /* CONFIG_USER_SCHED */
317 * A weight of 0 or 1 can cause arithmetics problems.
318 * A weight of a cfs_rq is the sum of weights of which entities
319 * are queued on this cfs_rq, so a weight of a entity should not be
320 * too large, so as the shares value of a task group.
321 * (The default weight is 1024 - so there's no practical
322 * limitation from this.)
325 #define MAX_SHARES (1UL << 18)
327 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
330 /* Default task group.
331 * Every task in system belong to this group at bootup.
333 struct task_group init_task_group;
335 /* return group to which a task belongs */
336 static inline struct task_group *task_group(struct task_struct *p)
338 struct task_group *tg;
340 #ifdef CONFIG_USER_SCHED
342 #elif defined(CONFIG_CGROUP_SCHED)
343 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
344 struct task_group, css);
346 tg = &init_task_group;
351 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
352 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
354 #ifdef CONFIG_FAIR_GROUP_SCHED
355 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
356 p->se.parent = task_group(p)->se[cpu];
359 #ifdef CONFIG_RT_GROUP_SCHED
360 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
361 p->rt.parent = task_group(p)->rt_se[cpu];
367 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
369 #endif /* CONFIG_GROUP_SCHED */
371 /* CFS-related fields in a runqueue */
373 struct load_weight load;
374 unsigned long nr_running;
379 struct rb_root tasks_timeline;
380 struct rb_node *rb_leftmost;
382 struct list_head tasks;
383 struct list_head *balance_iterator;
386 * 'curr' points to currently running entity on this cfs_rq.
387 * It is set to NULL otherwise (i.e when none are currently running).
389 struct sched_entity *curr, *next;
391 unsigned long nr_spread_over;
393 #ifdef CONFIG_FAIR_GROUP_SCHED
394 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
397 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
398 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
399 * (like users, containers etc.)
401 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
402 * list is used during load balance.
404 struct list_head leaf_cfs_rq_list;
405 struct task_group *tg; /* group that "owns" this runqueue */
408 unsigned long task_weight;
409 unsigned long shares;
411 * We need space to build a sched_domain wide view of the full task
412 * group tree, in order to avoid depending on dynamic memory allocation
413 * during the load balancing we place this in the per cpu task group
414 * hierarchy. This limits the load balancing to one instance per cpu,
415 * but more should not be needed anyway.
417 struct aggregate_struct {
419 * load = weight(cpus) * f(tg)
421 * Where f(tg) is the recursive weight fraction assigned to
427 * part of the group weight distributed to this span.
429 unsigned long shares;
432 * The sum of all runqueue weights within this span.
434 unsigned long rq_weight;
437 * Weight contributed by tasks; this is the part we can
438 * influence by moving tasks around.
440 unsigned long task_weight;
446 /* Real-Time classes' related field in a runqueue: */
448 struct rt_prio_array active;
449 unsigned long rt_nr_running;
450 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
451 int highest_prio; /* highest queued rt task prio */
454 unsigned long rt_nr_migratory;
460 /* Nests inside the rq lock: */
461 spinlock_t rt_runtime_lock;
463 #ifdef CONFIG_RT_GROUP_SCHED
464 unsigned long rt_nr_boosted;
467 struct list_head leaf_rt_rq_list;
468 struct task_group *tg;
469 struct sched_rt_entity *rt_se;
476 * We add the notion of a root-domain which will be used to define per-domain
477 * variables. Each exclusive cpuset essentially defines an island domain by
478 * fully partitioning the member cpus from any other cpuset. Whenever a new
479 * exclusive cpuset is created, we also create and attach a new root-domain
489 * The "RT overload" flag: it gets set if a CPU has more than
490 * one runnable RT task.
495 struct cpupri cpupri;
500 * By default the system creates a single root-domain with all cpus as
501 * members (mimicking the global state we have today).
503 static struct root_domain def_root_domain;
508 * This is the main, per-CPU runqueue data structure.
510 * Locking rule: those places that want to lock multiple runqueues
511 * (such as the load balancing or the thread migration code), lock
512 * acquire operations must be ordered by ascending &runqueue.
519 * nr_running and cpu_load should be in the same cacheline because
520 * remote CPUs use both these fields when doing load calculation.
522 unsigned long nr_running;
523 #define CPU_LOAD_IDX_MAX 5
524 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
525 unsigned char idle_at_tick;
527 unsigned long last_tick_seen;
528 unsigned char in_nohz_recently;
530 /* capture load from *all* tasks on this cpu: */
531 struct load_weight load;
532 unsigned long nr_load_updates;
538 #ifdef CONFIG_FAIR_GROUP_SCHED
539 /* list of leaf cfs_rq on this cpu: */
540 struct list_head leaf_cfs_rq_list;
542 #ifdef CONFIG_RT_GROUP_SCHED
543 struct list_head leaf_rt_rq_list;
547 * This is part of a global counter where only the total sum
548 * over all CPUs matters. A task can increase this counter on
549 * one CPU and if it got migrated afterwards it may decrease
550 * it on another CPU. Always updated under the runqueue lock:
552 unsigned long nr_uninterruptible;
554 struct task_struct *curr, *idle;
555 unsigned long next_balance;
556 struct mm_struct *prev_mm;
563 struct root_domain *rd;
564 struct sched_domain *sd;
566 /* For active balancing */
569 /* cpu of this runqueue: */
573 struct task_struct *migration_thread;
574 struct list_head migration_queue;
577 #ifdef CONFIG_SCHED_HRTICK
578 unsigned long hrtick_flags;
579 ktime_t hrtick_expire;
580 struct hrtimer hrtick_timer;
583 #ifdef CONFIG_SCHEDSTATS
585 struct sched_info rq_sched_info;
587 /* sys_sched_yield() stats */
588 unsigned int yld_exp_empty;
589 unsigned int yld_act_empty;
590 unsigned int yld_both_empty;
591 unsigned int yld_count;
593 /* schedule() stats */
594 unsigned int sched_switch;
595 unsigned int sched_count;
596 unsigned int sched_goidle;
598 /* try_to_wake_up() stats */
599 unsigned int ttwu_count;
600 unsigned int ttwu_local;
603 unsigned int bkl_count;
605 struct lock_class_key rq_lock_key;
608 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
610 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
612 rq->curr->sched_class->check_preempt_curr(rq, p);
615 static inline int cpu_of(struct rq *rq)
625 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
626 * See detach_destroy_domains: synchronize_sched for details.
628 * The domain tree of any CPU may only be accessed from within
629 * preempt-disabled sections.
631 #define for_each_domain(cpu, __sd) \
632 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
634 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
635 #define this_rq() (&__get_cpu_var(runqueues))
636 #define task_rq(p) cpu_rq(task_cpu(p))
637 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
639 static inline void update_rq_clock(struct rq *rq)
641 rq->clock = sched_clock_cpu(cpu_of(rq));
645 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
647 #ifdef CONFIG_SCHED_DEBUG
648 # define const_debug __read_mostly
650 # define const_debug static const
654 * Debugging: various feature bits
657 #define SCHED_FEAT(name, enabled) \
658 __SCHED_FEAT_##name ,
661 #include "sched_features.h"
666 #define SCHED_FEAT(name, enabled) \
667 (1UL << __SCHED_FEAT_##name) * enabled |
669 const_debug unsigned int sysctl_sched_features =
670 #include "sched_features.h"
675 #ifdef CONFIG_SCHED_DEBUG
676 #define SCHED_FEAT(name, enabled) \
679 static __read_mostly char *sched_feat_names[] = {
680 #include "sched_features.h"
686 static int sched_feat_open(struct inode *inode, struct file *filp)
688 filp->private_data = inode->i_private;
693 sched_feat_read(struct file *filp, char __user *ubuf,
694 size_t cnt, loff_t *ppos)
701 for (i = 0; sched_feat_names[i]; i++) {
702 len += strlen(sched_feat_names[i]);
706 buf = kmalloc(len + 2, GFP_KERNEL);
710 for (i = 0; sched_feat_names[i]; i++) {
711 if (sysctl_sched_features & (1UL << i))
712 r += sprintf(buf + r, "%s ", sched_feat_names[i]);
714 r += sprintf(buf + r, "NO_%s ", sched_feat_names[i]);
717 r += sprintf(buf + r, "\n");
718 WARN_ON(r >= len + 2);
720 r = simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
728 sched_feat_write(struct file *filp, const char __user *ubuf,
729 size_t cnt, loff_t *ppos)
739 if (copy_from_user(&buf, ubuf, cnt))
744 if (strncmp(buf, "NO_", 3) == 0) {
749 for (i = 0; sched_feat_names[i]; i++) {
750 int len = strlen(sched_feat_names[i]);
752 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
754 sysctl_sched_features &= ~(1UL << i);
756 sysctl_sched_features |= (1UL << i);
761 if (!sched_feat_names[i])
769 static struct file_operations sched_feat_fops = {
770 .open = sched_feat_open,
771 .read = sched_feat_read,
772 .write = sched_feat_write,
775 static __init int sched_init_debug(void)
777 debugfs_create_file("sched_features", 0644, NULL, NULL,
782 late_initcall(sched_init_debug);
786 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
789 * Number of tasks to iterate in a single balance run.
790 * Limited because this is done with IRQs disabled.
792 const_debug unsigned int sysctl_sched_nr_migrate = 32;
795 * period over which we measure -rt task cpu usage in us.
798 unsigned int sysctl_sched_rt_period = 1000000;
800 static __read_mostly int scheduler_running;
803 * part of the period that we allow rt tasks to run in us.
806 int sysctl_sched_rt_runtime = 950000;
808 static inline u64 global_rt_period(void)
810 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
813 static inline u64 global_rt_runtime(void)
815 if (sysctl_sched_rt_period < 0)
818 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
821 #ifndef prepare_arch_switch
822 # define prepare_arch_switch(next) do { } while (0)
824 #ifndef finish_arch_switch
825 # define finish_arch_switch(prev) do { } while (0)
828 static inline int task_current(struct rq *rq, struct task_struct *p)
830 return rq->curr == p;
833 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
834 static inline int task_running(struct rq *rq, struct task_struct *p)
836 return task_current(rq, p);
839 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
843 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
845 #ifdef CONFIG_DEBUG_SPINLOCK
846 /* this is a valid case when another task releases the spinlock */
847 rq->lock.owner = current;
850 * If we are tracking spinlock dependencies then we have to
851 * fix up the runqueue lock - which gets 'carried over' from
854 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
856 spin_unlock_irq(&rq->lock);
859 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
860 static inline int task_running(struct rq *rq, struct task_struct *p)
865 return task_current(rq, p);
869 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
873 * We can optimise this out completely for !SMP, because the
874 * SMP rebalancing from interrupt is the only thing that cares
879 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
880 spin_unlock_irq(&rq->lock);
882 spin_unlock(&rq->lock);
886 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
890 * After ->oncpu is cleared, the task can be moved to a different CPU.
891 * We must ensure this doesn't happen until the switch is completely
897 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
901 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
904 * __task_rq_lock - lock the runqueue a given task resides on.
905 * Must be called interrupts disabled.
907 static inline struct rq *__task_rq_lock(struct task_struct *p)
911 struct rq *rq = task_rq(p);
912 spin_lock(&rq->lock);
913 if (likely(rq == task_rq(p)))
915 spin_unlock(&rq->lock);
920 * task_rq_lock - lock the runqueue a given task resides on and disable
921 * interrupts. Note the ordering: we can safely lookup the task_rq without
922 * explicitly disabling preemption.
924 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
930 local_irq_save(*flags);
932 spin_lock(&rq->lock);
933 if (likely(rq == task_rq(p)))
935 spin_unlock_irqrestore(&rq->lock, *flags);
939 static void __task_rq_unlock(struct rq *rq)
942 spin_unlock(&rq->lock);
945 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
948 spin_unlock_irqrestore(&rq->lock, *flags);
952 * this_rq_lock - lock this runqueue and disable interrupts.
954 static struct rq *this_rq_lock(void)
961 spin_lock(&rq->lock);
966 static void __resched_task(struct task_struct *p, int tif_bit);
968 static inline void resched_task(struct task_struct *p)
970 __resched_task(p, TIF_NEED_RESCHED);
973 #ifdef CONFIG_SCHED_HRTICK
975 * Use HR-timers to deliver accurate preemption points.
977 * Its all a bit involved since we cannot program an hrt while holding the
978 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
981 * When we get rescheduled we reprogram the hrtick_timer outside of the
984 static inline void resched_hrt(struct task_struct *p)
986 __resched_task(p, TIF_HRTICK_RESCHED);
989 static inline void resched_rq(struct rq *rq)
993 spin_lock_irqsave(&rq->lock, flags);
994 resched_task(rq->curr);
995 spin_unlock_irqrestore(&rq->lock, flags);
999 HRTICK_SET, /* re-programm hrtick_timer */
1000 HRTICK_RESET, /* not a new slice */
1001 HRTICK_BLOCK, /* stop hrtick operations */
1006 * - enabled by features
1007 * - hrtimer is actually high res
1009 static inline int hrtick_enabled(struct rq *rq)
1011 if (!sched_feat(HRTICK))
1013 if (unlikely(test_bit(HRTICK_BLOCK, &rq->hrtick_flags)))
1015 return hrtimer_is_hres_active(&rq->hrtick_timer);
1019 * Called to set the hrtick timer state.
1021 * called with rq->lock held and irqs disabled
1023 static void hrtick_start(struct rq *rq, u64 delay, int reset)
1025 assert_spin_locked(&rq->lock);
1028 * preempt at: now + delay
1031 ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
1033 * indicate we need to program the timer
1035 __set_bit(HRTICK_SET, &rq->hrtick_flags);
1037 __set_bit(HRTICK_RESET, &rq->hrtick_flags);
1040 * New slices are called from the schedule path and don't need a
1041 * forced reschedule.
1044 resched_hrt(rq->curr);
1047 static void hrtick_clear(struct rq *rq)
1049 if (hrtimer_active(&rq->hrtick_timer))
1050 hrtimer_cancel(&rq->hrtick_timer);
1054 * Update the timer from the possible pending state.
1056 static void hrtick_set(struct rq *rq)
1060 unsigned long flags;
1062 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1064 spin_lock_irqsave(&rq->lock, flags);
1065 set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
1066 reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
1067 time = rq->hrtick_expire;
1068 clear_thread_flag(TIF_HRTICK_RESCHED);
1069 spin_unlock_irqrestore(&rq->lock, flags);
1072 hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
1073 if (reset && !hrtimer_active(&rq->hrtick_timer))
1080 * High-resolution timer tick.
1081 * Runs from hardirq context with interrupts disabled.
1083 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1085 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1087 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1089 spin_lock(&rq->lock);
1090 update_rq_clock(rq);
1091 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1092 spin_unlock(&rq->lock);
1094 return HRTIMER_NORESTART;
1098 static void hotplug_hrtick_disable(int cpu)
1100 struct rq *rq = cpu_rq(cpu);
1101 unsigned long flags;
1103 spin_lock_irqsave(&rq->lock, flags);
1104 rq->hrtick_flags = 0;
1105 __set_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1106 spin_unlock_irqrestore(&rq->lock, flags);
1111 static void hotplug_hrtick_enable(int cpu)
1113 struct rq *rq = cpu_rq(cpu);
1114 unsigned long flags;
1116 spin_lock_irqsave(&rq->lock, flags);
1117 __clear_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1118 spin_unlock_irqrestore(&rq->lock, flags);
1122 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1124 int cpu = (int)(long)hcpu;
1127 case CPU_UP_CANCELED:
1128 case CPU_UP_CANCELED_FROZEN:
1129 case CPU_DOWN_PREPARE:
1130 case CPU_DOWN_PREPARE_FROZEN:
1132 case CPU_DEAD_FROZEN:
1133 hotplug_hrtick_disable(cpu);
1136 case CPU_UP_PREPARE:
1137 case CPU_UP_PREPARE_FROZEN:
1138 case CPU_DOWN_FAILED:
1139 case CPU_DOWN_FAILED_FROZEN:
1141 case CPU_ONLINE_FROZEN:
1142 hotplug_hrtick_enable(cpu);
1149 static void init_hrtick(void)
1151 hotcpu_notifier(hotplug_hrtick, 0);
1153 #endif /* CONFIG_SMP */
1155 static void init_rq_hrtick(struct rq *rq)
1157 rq->hrtick_flags = 0;
1158 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1159 rq->hrtick_timer.function = hrtick;
1160 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1163 void hrtick_resched(void)
1166 unsigned long flags;
1168 if (!test_thread_flag(TIF_HRTICK_RESCHED))
1171 local_irq_save(flags);
1172 rq = cpu_rq(smp_processor_id());
1174 local_irq_restore(flags);
1177 static inline void hrtick_clear(struct rq *rq)
1181 static inline void hrtick_set(struct rq *rq)
1185 static inline void init_rq_hrtick(struct rq *rq)
1189 void hrtick_resched(void)
1193 static inline void init_hrtick(void)
1199 * resched_task - mark a task 'to be rescheduled now'.
1201 * On UP this means the setting of the need_resched flag, on SMP it
1202 * might also involve a cross-CPU call to trigger the scheduler on
1207 #ifndef tsk_is_polling
1208 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1211 static void __resched_task(struct task_struct *p, int tif_bit)
1215 assert_spin_locked(&task_rq(p)->lock);
1217 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
1220 set_tsk_thread_flag(p, tif_bit);
1223 if (cpu == smp_processor_id())
1226 /* NEED_RESCHED must be visible before we test polling */
1228 if (!tsk_is_polling(p))
1229 smp_send_reschedule(cpu);
1232 static void resched_cpu(int cpu)
1234 struct rq *rq = cpu_rq(cpu);
1235 unsigned long flags;
1237 if (!spin_trylock_irqsave(&rq->lock, flags))
1239 resched_task(cpu_curr(cpu));
1240 spin_unlock_irqrestore(&rq->lock, flags);
1245 * When add_timer_on() enqueues a timer into the timer wheel of an
1246 * idle CPU then this timer might expire before the next timer event
1247 * which is scheduled to wake up that CPU. In case of a completely
1248 * idle system the next event might even be infinite time into the
1249 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1250 * leaves the inner idle loop so the newly added timer is taken into
1251 * account when the CPU goes back to idle and evaluates the timer
1252 * wheel for the next timer event.
1254 void wake_up_idle_cpu(int cpu)
1256 struct rq *rq = cpu_rq(cpu);
1258 if (cpu == smp_processor_id())
1262 * This is safe, as this function is called with the timer
1263 * wheel base lock of (cpu) held. When the CPU is on the way
1264 * to idle and has not yet set rq->curr to idle then it will
1265 * be serialized on the timer wheel base lock and take the new
1266 * timer into account automatically.
1268 if (rq->curr != rq->idle)
1272 * We can set TIF_RESCHED on the idle task of the other CPU
1273 * lockless. The worst case is that the other CPU runs the
1274 * idle task through an additional NOOP schedule()
1276 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1278 /* NEED_RESCHED must be visible before we test polling */
1280 if (!tsk_is_polling(rq->idle))
1281 smp_send_reschedule(cpu);
1283 #endif /* CONFIG_NO_HZ */
1285 #else /* !CONFIG_SMP */
1286 static void __resched_task(struct task_struct *p, int tif_bit)
1288 assert_spin_locked(&task_rq(p)->lock);
1289 set_tsk_thread_flag(p, tif_bit);
1291 #endif /* CONFIG_SMP */
1293 #if BITS_PER_LONG == 32
1294 # define WMULT_CONST (~0UL)
1296 # define WMULT_CONST (1UL << 32)
1299 #define WMULT_SHIFT 32
1302 * Shift right and round:
1304 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1307 * delta *= weight / lw
1309 static unsigned long
1310 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1311 struct load_weight *lw)
1315 if (!lw->inv_weight) {
1316 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1319 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1323 tmp = (u64)delta_exec * weight;
1325 * Check whether we'd overflow the 64-bit multiplication:
1327 if (unlikely(tmp > WMULT_CONST))
1328 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1331 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1333 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1336 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1342 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1349 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1350 * of tasks with abnormal "nice" values across CPUs the contribution that
1351 * each task makes to its run queue's load is weighted according to its
1352 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1353 * scaled version of the new time slice allocation that they receive on time
1357 #define WEIGHT_IDLEPRIO 2
1358 #define WMULT_IDLEPRIO (1 << 31)
1361 * Nice levels are multiplicative, with a gentle 10% change for every
1362 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1363 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1364 * that remained on nice 0.
1366 * The "10% effect" is relative and cumulative: from _any_ nice level,
1367 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1368 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1369 * If a task goes up by ~10% and another task goes down by ~10% then
1370 * the relative distance between them is ~25%.)
1372 static const int prio_to_weight[40] = {
1373 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1374 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1375 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1376 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1377 /* 0 */ 1024, 820, 655, 526, 423,
1378 /* 5 */ 335, 272, 215, 172, 137,
1379 /* 10 */ 110, 87, 70, 56, 45,
1380 /* 15 */ 36, 29, 23, 18, 15,
1384 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1386 * In cases where the weight does not change often, we can use the
1387 * precalculated inverse to speed up arithmetics by turning divisions
1388 * into multiplications:
1390 static const u32 prio_to_wmult[40] = {
1391 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1392 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1393 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1394 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1395 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1396 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1397 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1398 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1401 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1404 * runqueue iterator, to support SMP load-balancing between different
1405 * scheduling classes, without having to expose their internal data
1406 * structures to the load-balancing proper:
1408 struct rq_iterator {
1410 struct task_struct *(*start)(void *);
1411 struct task_struct *(*next)(void *);
1415 static unsigned long
1416 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1417 unsigned long max_load_move, struct sched_domain *sd,
1418 enum cpu_idle_type idle, int *all_pinned,
1419 int *this_best_prio, struct rq_iterator *iterator);
1422 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1423 struct sched_domain *sd, enum cpu_idle_type idle,
1424 struct rq_iterator *iterator);
1427 #ifdef CONFIG_CGROUP_CPUACCT
1428 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1430 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1433 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1435 update_load_add(&rq->load, load);
1438 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1440 update_load_sub(&rq->load, load);
1444 static unsigned long source_load(int cpu, int type);
1445 static unsigned long target_load(int cpu, int type);
1446 static unsigned long cpu_avg_load_per_task(int cpu);
1447 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1449 #ifdef CONFIG_FAIR_GROUP_SCHED
1452 * Group load balancing.
1454 * We calculate a few balance domain wide aggregate numbers; load and weight.
1455 * Given the pictures below, and assuming each item has equal weight:
1466 * A and B get 1/3-rd of the total load. C and D get 1/3-rd of A's 1/3-rd,
1467 * which equals 1/9-th of the total load.
1470 * The weight of this group on the selected cpus.
1473 * Direct sum of all the cpu's their rq weight, e.g. A would get 3 while
1477 * Part of the rq_weight contributed by tasks; all groups except B would
1481 static inline struct aggregate_struct *
1482 aggregate(struct task_group *tg, struct sched_domain *sd)
1484 return &tg->cfs_rq[sd->first_cpu]->aggregate;
1487 typedef void (*aggregate_func)(struct task_group *, struct sched_domain *);
1490 * Iterate the full tree, calling @down when first entering a node and @up when
1491 * leaving it for the final time.
1494 void aggregate_walk_tree(aggregate_func down, aggregate_func up,
1495 struct sched_domain *sd)
1497 struct task_group *parent, *child;
1500 parent = &root_task_group;
1502 (*down)(parent, sd);
1503 list_for_each_entry_rcu(child, &parent->children, siblings) {
1513 parent = parent->parent;
1520 * Calculate the aggregate runqueue weight.
1523 void aggregate_group_weight(struct task_group *tg, struct sched_domain *sd)
1525 unsigned long rq_weight = 0;
1526 unsigned long task_weight = 0;
1529 for_each_cpu_mask(i, sd->span) {
1530 rq_weight += tg->cfs_rq[i]->load.weight;
1531 task_weight += tg->cfs_rq[i]->task_weight;
1534 aggregate(tg, sd)->rq_weight = rq_weight;
1535 aggregate(tg, sd)->task_weight = task_weight;
1539 * Compute the weight of this group on the given cpus.
1542 void aggregate_group_shares(struct task_group *tg, struct sched_domain *sd)
1544 unsigned long shares = 0;
1547 for_each_cpu_mask(i, sd->span)
1548 shares += tg->cfs_rq[i]->shares;
1550 if ((!shares && aggregate(tg, sd)->rq_weight) || shares > tg->shares)
1551 shares = tg->shares;
1553 aggregate(tg, sd)->shares = shares;
1557 * Compute the load fraction assigned to this group, relies on the aggregate
1558 * weight and this group's parent's load, i.e. top-down.
1561 void aggregate_group_load(struct task_group *tg, struct sched_domain *sd)
1569 for_each_cpu_mask(i, sd->span)
1570 load += cpu_rq(i)->load.weight;
1573 load = aggregate(tg->parent, sd)->load;
1576 * shares is our weight in the parent's rq so
1577 * shares/parent->rq_weight gives our fraction of the load
1579 load *= aggregate(tg, sd)->shares;
1580 load /= aggregate(tg->parent, sd)->rq_weight + 1;
1583 aggregate(tg, sd)->load = load;
1586 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1589 * Calculate and set the cpu's group shares.
1592 __update_group_shares_cpu(struct task_group *tg, struct sched_domain *sd,
1596 unsigned long shares;
1597 unsigned long rq_weight;
1602 rq_weight = tg->cfs_rq[tcpu]->load.weight;
1605 * If there are currently no tasks on the cpu pretend there is one of
1606 * average load so that when a new task gets to run here it will not
1607 * get delayed by group starvation.
1611 rq_weight = NICE_0_LOAD;
1615 * \Sum shares * rq_weight
1616 * shares = -----------------------
1620 shares = aggregate(tg, sd)->shares * rq_weight;
1621 shares /= aggregate(tg, sd)->rq_weight + 1;
1624 * record the actual number of shares, not the boosted amount.
1626 tg->cfs_rq[tcpu]->shares = boost ? 0 : shares;
1628 if (shares < MIN_SHARES)
1629 shares = MIN_SHARES;
1630 else if (shares > MAX_SHARES)
1631 shares = MAX_SHARES;
1633 __set_se_shares(tg->se[tcpu], shares);
1637 * Re-adjust the weights on the cpu the task came from and on the cpu the
1641 __move_group_shares(struct task_group *tg, struct sched_domain *sd,
1644 unsigned long shares;
1646 shares = tg->cfs_rq[scpu]->shares + tg->cfs_rq[dcpu]->shares;
1648 __update_group_shares_cpu(tg, sd, scpu);
1649 __update_group_shares_cpu(tg, sd, dcpu);
1652 * ensure we never loose shares due to rounding errors in the
1653 * above redistribution.
1655 shares -= tg->cfs_rq[scpu]->shares + tg->cfs_rq[dcpu]->shares;
1657 tg->cfs_rq[dcpu]->shares += shares;
1661 * Because changing a group's shares changes the weight of the super-group
1662 * we need to walk up the tree and change all shares until we hit the root.
1665 move_group_shares(struct task_group *tg, struct sched_domain *sd,
1669 __move_group_shares(tg, sd, scpu, dcpu);
1675 void aggregate_group_set_shares(struct task_group *tg, struct sched_domain *sd)
1677 unsigned long shares = aggregate(tg, sd)->shares;
1680 for_each_cpu_mask(i, sd->span) {
1681 struct rq *rq = cpu_rq(i);
1682 unsigned long flags;
1684 spin_lock_irqsave(&rq->lock, flags);
1685 __update_group_shares_cpu(tg, sd, i);
1686 spin_unlock_irqrestore(&rq->lock, flags);
1689 aggregate_group_shares(tg, sd);
1692 * ensure we never loose shares due to rounding errors in the
1693 * above redistribution.
1695 shares -= aggregate(tg, sd)->shares;
1697 tg->cfs_rq[sd->first_cpu]->shares += shares;
1698 aggregate(tg, sd)->shares += shares;
1703 * Calculate the accumulative weight and recursive load of each task group
1704 * while walking down the tree.
1707 void aggregate_get_down(struct task_group *tg, struct sched_domain *sd)
1709 aggregate_group_weight(tg, sd);
1710 aggregate_group_shares(tg, sd);
1711 aggregate_group_load(tg, sd);
1715 * Rebalance the cpu shares while walking back up the tree.
1718 void aggregate_get_up(struct task_group *tg, struct sched_domain *sd)
1720 aggregate_group_set_shares(tg, sd);
1723 static DEFINE_PER_CPU(spinlock_t, aggregate_lock);
1725 static void __init init_aggregate(void)
1729 for_each_possible_cpu(i)
1730 spin_lock_init(&per_cpu(aggregate_lock, i));
1733 static int get_aggregate(struct sched_domain *sd)
1735 if (!spin_trylock(&per_cpu(aggregate_lock, sd->first_cpu)))
1738 aggregate_walk_tree(aggregate_get_down, aggregate_get_up, sd);
1742 static void put_aggregate(struct sched_domain *sd)
1744 spin_unlock(&per_cpu(aggregate_lock, sd->first_cpu));
1747 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1749 cfs_rq->shares = shares;
1754 static inline void init_aggregate(void)
1758 static inline int get_aggregate(struct sched_domain *sd)
1763 static inline void put_aggregate(struct sched_domain *sd)
1770 #include "sched_stats.h"
1771 #include "sched_idletask.c"
1772 #include "sched_fair.c"
1773 #include "sched_rt.c"
1774 #ifdef CONFIG_SCHED_DEBUG
1775 # include "sched_debug.c"
1778 #define sched_class_highest (&rt_sched_class)
1779 #define for_each_class(class) \
1780 for (class = sched_class_highest; class; class = class->next)
1782 static void inc_nr_running(struct rq *rq)
1787 static void dec_nr_running(struct rq *rq)
1792 static void set_load_weight(struct task_struct *p)
1794 if (task_has_rt_policy(p)) {
1795 p->se.load.weight = prio_to_weight[0] * 2;
1796 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1801 * SCHED_IDLE tasks get minimal weight:
1803 if (p->policy == SCHED_IDLE) {
1804 p->se.load.weight = WEIGHT_IDLEPRIO;
1805 p->se.load.inv_weight = WMULT_IDLEPRIO;
1809 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1810 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1813 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1815 sched_info_queued(p);
1816 p->sched_class->enqueue_task(rq, p, wakeup);
1820 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1822 p->sched_class->dequeue_task(rq, p, sleep);
1827 * __normal_prio - return the priority that is based on the static prio
1829 static inline int __normal_prio(struct task_struct *p)
1831 return p->static_prio;
1835 * Calculate the expected normal priority: i.e. priority
1836 * without taking RT-inheritance into account. Might be
1837 * boosted by interactivity modifiers. Changes upon fork,
1838 * setprio syscalls, and whenever the interactivity
1839 * estimator recalculates.
1841 static inline int normal_prio(struct task_struct *p)
1845 if (task_has_rt_policy(p))
1846 prio = MAX_RT_PRIO-1 - p->rt_priority;
1848 prio = __normal_prio(p);
1853 * Calculate the current priority, i.e. the priority
1854 * taken into account by the scheduler. This value might
1855 * be boosted by RT tasks, or might be boosted by
1856 * interactivity modifiers. Will be RT if the task got
1857 * RT-boosted. If not then it returns p->normal_prio.
1859 static int effective_prio(struct task_struct *p)
1861 p->normal_prio = normal_prio(p);
1863 * If we are RT tasks or we were boosted to RT priority,
1864 * keep the priority unchanged. Otherwise, update priority
1865 * to the normal priority:
1867 if (!rt_prio(p->prio))
1868 return p->normal_prio;
1873 * activate_task - move a task to the runqueue.
1875 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1877 if (task_contributes_to_load(p))
1878 rq->nr_uninterruptible--;
1880 enqueue_task(rq, p, wakeup);
1885 * deactivate_task - remove a task from the runqueue.
1887 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1889 if (task_contributes_to_load(p))
1890 rq->nr_uninterruptible++;
1892 dequeue_task(rq, p, sleep);
1897 * task_curr - is this task currently executing on a CPU?
1898 * @p: the task in question.
1900 inline int task_curr(const struct task_struct *p)
1902 return cpu_curr(task_cpu(p)) == p;
1905 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1907 set_task_rq(p, cpu);
1910 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1911 * successfuly executed on another CPU. We must ensure that updates of
1912 * per-task data have been completed by this moment.
1915 task_thread_info(p)->cpu = cpu;
1919 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1920 const struct sched_class *prev_class,
1921 int oldprio, int running)
1923 if (prev_class != p->sched_class) {
1924 if (prev_class->switched_from)
1925 prev_class->switched_from(rq, p, running);
1926 p->sched_class->switched_to(rq, p, running);
1928 p->sched_class->prio_changed(rq, p, oldprio, running);
1933 /* Used instead of source_load when we know the type == 0 */
1934 static unsigned long weighted_cpuload(const int cpu)
1936 return cpu_rq(cpu)->load.weight;
1940 * Is this task likely cache-hot:
1943 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1948 * Buddy candidates are cache hot:
1950 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
1953 if (p->sched_class != &fair_sched_class)
1956 if (sysctl_sched_migration_cost == -1)
1958 if (sysctl_sched_migration_cost == 0)
1961 delta = now - p->se.exec_start;
1963 return delta < (s64)sysctl_sched_migration_cost;
1967 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1969 int old_cpu = task_cpu(p);
1970 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1971 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1972 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1975 clock_offset = old_rq->clock - new_rq->clock;
1977 #ifdef CONFIG_SCHEDSTATS
1978 if (p->se.wait_start)
1979 p->se.wait_start -= clock_offset;
1980 if (p->se.sleep_start)
1981 p->se.sleep_start -= clock_offset;
1982 if (p->se.block_start)
1983 p->se.block_start -= clock_offset;
1984 if (old_cpu != new_cpu) {
1985 schedstat_inc(p, se.nr_migrations);
1986 if (task_hot(p, old_rq->clock, NULL))
1987 schedstat_inc(p, se.nr_forced2_migrations);
1990 p->se.vruntime -= old_cfsrq->min_vruntime -
1991 new_cfsrq->min_vruntime;
1993 __set_task_cpu(p, new_cpu);
1996 struct migration_req {
1997 struct list_head list;
1999 struct task_struct *task;
2002 struct completion done;
2006 * The task's runqueue lock must be held.
2007 * Returns true if you have to wait for migration thread.
2010 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2012 struct rq *rq = task_rq(p);
2015 * If the task is not on a runqueue (and not running), then
2016 * it is sufficient to simply update the task's cpu field.
2018 if (!p->se.on_rq && !task_running(rq, p)) {
2019 set_task_cpu(p, dest_cpu);
2023 init_completion(&req->done);
2025 req->dest_cpu = dest_cpu;
2026 list_add(&req->list, &rq->migration_queue);
2032 * wait_task_inactive - wait for a thread to unschedule.
2034 * The caller must ensure that the task *will* unschedule sometime soon,
2035 * else this function might spin for a *long* time. This function can't
2036 * be called with interrupts off, or it may introduce deadlock with
2037 * smp_call_function() if an IPI is sent by the same process we are
2038 * waiting to become inactive.
2040 void wait_task_inactive(struct task_struct *p)
2042 unsigned long flags;
2048 * We do the initial early heuristics without holding
2049 * any task-queue locks at all. We'll only try to get
2050 * the runqueue lock when things look like they will
2056 * If the task is actively running on another CPU
2057 * still, just relax and busy-wait without holding
2060 * NOTE! Since we don't hold any locks, it's not
2061 * even sure that "rq" stays as the right runqueue!
2062 * But we don't care, since "task_running()" will
2063 * return false if the runqueue has changed and p
2064 * is actually now running somewhere else!
2066 while (task_running(rq, p))
2070 * Ok, time to look more closely! We need the rq
2071 * lock now, to be *sure*. If we're wrong, we'll
2072 * just go back and repeat.
2074 rq = task_rq_lock(p, &flags);
2075 running = task_running(rq, p);
2076 on_rq = p->se.on_rq;
2077 task_rq_unlock(rq, &flags);
2080 * Was it really running after all now that we
2081 * checked with the proper locks actually held?
2083 * Oops. Go back and try again..
2085 if (unlikely(running)) {
2091 * It's not enough that it's not actively running,
2092 * it must be off the runqueue _entirely_, and not
2095 * So if it wa still runnable (but just not actively
2096 * running right now), it's preempted, and we should
2097 * yield - it could be a while.
2099 if (unlikely(on_rq)) {
2100 schedule_timeout_uninterruptible(1);
2105 * Ahh, all good. It wasn't running, and it wasn't
2106 * runnable, which means that it will never become
2107 * running in the future either. We're all done!
2114 * kick_process - kick a running thread to enter/exit the kernel
2115 * @p: the to-be-kicked thread
2117 * Cause a process which is running on another CPU to enter
2118 * kernel-mode, without any delay. (to get signals handled.)
2120 * NOTE: this function doesnt have to take the runqueue lock,
2121 * because all it wants to ensure is that the remote task enters
2122 * the kernel. If the IPI races and the task has been migrated
2123 * to another CPU then no harm is done and the purpose has been
2126 void kick_process(struct task_struct *p)
2132 if ((cpu != smp_processor_id()) && task_curr(p))
2133 smp_send_reschedule(cpu);
2138 * Return a low guess at the load of a migration-source cpu weighted
2139 * according to the scheduling class and "nice" value.
2141 * We want to under-estimate the load of migration sources, to
2142 * balance conservatively.
2144 static unsigned long source_load(int cpu, int type)
2146 struct rq *rq = cpu_rq(cpu);
2147 unsigned long total = weighted_cpuload(cpu);
2152 return min(rq->cpu_load[type-1], total);
2156 * Return a high guess at the load of a migration-target cpu weighted
2157 * according to the scheduling class and "nice" value.
2159 static unsigned long target_load(int cpu, int type)
2161 struct rq *rq = cpu_rq(cpu);
2162 unsigned long total = weighted_cpuload(cpu);
2167 return max(rq->cpu_load[type-1], total);
2171 * Return the average load per task on the cpu's run queue
2173 static unsigned long cpu_avg_load_per_task(int cpu)
2175 struct rq *rq = cpu_rq(cpu);
2176 unsigned long total = weighted_cpuload(cpu);
2177 unsigned long n = rq->nr_running;
2179 return n ? total / n : SCHED_LOAD_SCALE;
2183 * find_idlest_group finds and returns the least busy CPU group within the
2186 static struct sched_group *
2187 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2189 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2190 unsigned long min_load = ULONG_MAX, this_load = 0;
2191 int load_idx = sd->forkexec_idx;
2192 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2195 unsigned long load, avg_load;
2199 /* Skip over this group if it has no CPUs allowed */
2200 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
2203 local_group = cpu_isset(this_cpu, group->cpumask);
2205 /* Tally up the load of all CPUs in the group */
2208 for_each_cpu_mask(i, group->cpumask) {
2209 /* Bias balancing toward cpus of our domain */
2211 load = source_load(i, load_idx);
2213 load = target_load(i, load_idx);
2218 /* Adjust by relative CPU power of the group */
2219 avg_load = sg_div_cpu_power(group,
2220 avg_load * SCHED_LOAD_SCALE);
2223 this_load = avg_load;
2225 } else if (avg_load < min_load) {
2226 min_load = avg_load;
2229 } while (group = group->next, group != sd->groups);
2231 if (!idlest || 100*this_load < imbalance*min_load)
2237 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2240 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2243 unsigned long load, min_load = ULONG_MAX;
2247 /* Traverse only the allowed CPUs */
2248 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
2250 for_each_cpu_mask(i, *tmp) {
2251 load = weighted_cpuload(i);
2253 if (load < min_load || (load == min_load && i == this_cpu)) {
2263 * sched_balance_self: balance the current task (running on cpu) in domains
2264 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2267 * Balance, ie. select the least loaded group.
2269 * Returns the target CPU number, or the same CPU if no balancing is needed.
2271 * preempt must be disabled.
2273 static int sched_balance_self(int cpu, int flag)
2275 struct task_struct *t = current;
2276 struct sched_domain *tmp, *sd = NULL;
2278 for_each_domain(cpu, tmp) {
2280 * If power savings logic is enabled for a domain, stop there.
2282 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2284 if (tmp->flags & flag)
2289 cpumask_t span, tmpmask;
2290 struct sched_group *group;
2291 int new_cpu, weight;
2293 if (!(sd->flags & flag)) {
2299 group = find_idlest_group(sd, t, cpu);
2305 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2306 if (new_cpu == -1 || new_cpu == cpu) {
2307 /* Now try balancing at a lower domain level of cpu */
2312 /* Now try balancing at a lower domain level of new_cpu */
2315 weight = cpus_weight(span);
2316 for_each_domain(cpu, tmp) {
2317 if (weight <= cpus_weight(tmp->span))
2319 if (tmp->flags & flag)
2322 /* while loop will break here if sd == NULL */
2328 #endif /* CONFIG_SMP */
2331 * try_to_wake_up - wake up a thread
2332 * @p: the to-be-woken-up thread
2333 * @state: the mask of task states that can be woken
2334 * @sync: do a synchronous wakeup?
2336 * Put it on the run-queue if it's not already there. The "current"
2337 * thread is always on the run-queue (except when the actual
2338 * re-schedule is in progress), and as such you're allowed to do
2339 * the simpler "current->state = TASK_RUNNING" to mark yourself
2340 * runnable without the overhead of this.
2342 * returns failure only if the task is already active.
2344 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2346 int cpu, orig_cpu, this_cpu, success = 0;
2347 unsigned long flags;
2351 if (!sched_feat(SYNC_WAKEUPS))
2355 rq = task_rq_lock(p, &flags);
2356 old_state = p->state;
2357 if (!(old_state & state))
2365 this_cpu = smp_processor_id();
2368 if (unlikely(task_running(rq, p)))
2371 cpu = p->sched_class->select_task_rq(p, sync);
2372 if (cpu != orig_cpu) {
2373 set_task_cpu(p, cpu);
2374 task_rq_unlock(rq, &flags);
2375 /* might preempt at this point */
2376 rq = task_rq_lock(p, &flags);
2377 old_state = p->state;
2378 if (!(old_state & state))
2383 this_cpu = smp_processor_id();
2387 #ifdef CONFIG_SCHEDSTATS
2388 schedstat_inc(rq, ttwu_count);
2389 if (cpu == this_cpu)
2390 schedstat_inc(rq, ttwu_local);
2392 struct sched_domain *sd;
2393 for_each_domain(this_cpu, sd) {
2394 if (cpu_isset(cpu, sd->span)) {
2395 schedstat_inc(sd, ttwu_wake_remote);
2400 #endif /* CONFIG_SCHEDSTATS */
2403 #endif /* CONFIG_SMP */
2404 schedstat_inc(p, se.nr_wakeups);
2406 schedstat_inc(p, se.nr_wakeups_sync);
2407 if (orig_cpu != cpu)
2408 schedstat_inc(p, se.nr_wakeups_migrate);
2409 if (cpu == this_cpu)
2410 schedstat_inc(p, se.nr_wakeups_local);
2412 schedstat_inc(p, se.nr_wakeups_remote);
2413 update_rq_clock(rq);
2414 activate_task(rq, p, 1);
2418 check_preempt_curr(rq, p);
2420 p->state = TASK_RUNNING;
2422 if (p->sched_class->task_wake_up)
2423 p->sched_class->task_wake_up(rq, p);
2426 task_rq_unlock(rq, &flags);
2431 int wake_up_process(struct task_struct *p)
2433 return try_to_wake_up(p, TASK_ALL, 0);
2435 EXPORT_SYMBOL(wake_up_process);
2437 int wake_up_state(struct task_struct *p, unsigned int state)
2439 return try_to_wake_up(p, state, 0);
2443 * Perform scheduler related setup for a newly forked process p.
2444 * p is forked by current.
2446 * __sched_fork() is basic setup used by init_idle() too:
2448 static void __sched_fork(struct task_struct *p)
2450 p->se.exec_start = 0;
2451 p->se.sum_exec_runtime = 0;
2452 p->se.prev_sum_exec_runtime = 0;
2453 p->se.last_wakeup = 0;
2454 p->se.avg_overlap = 0;
2456 #ifdef CONFIG_SCHEDSTATS
2457 p->se.wait_start = 0;
2458 p->se.sum_sleep_runtime = 0;
2459 p->se.sleep_start = 0;
2460 p->se.block_start = 0;
2461 p->se.sleep_max = 0;
2462 p->se.block_max = 0;
2464 p->se.slice_max = 0;
2468 INIT_LIST_HEAD(&p->rt.run_list);
2470 INIT_LIST_HEAD(&p->se.group_node);
2472 #ifdef CONFIG_PREEMPT_NOTIFIERS
2473 INIT_HLIST_HEAD(&p->preempt_notifiers);
2477 * We mark the process as running here, but have not actually
2478 * inserted it onto the runqueue yet. This guarantees that
2479 * nobody will actually run it, and a signal or other external
2480 * event cannot wake it up and insert it on the runqueue either.
2482 p->state = TASK_RUNNING;
2486 * fork()/clone()-time setup:
2488 void sched_fork(struct task_struct *p, int clone_flags)
2490 int cpu = get_cpu();
2495 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2497 set_task_cpu(p, cpu);
2500 * Make sure we do not leak PI boosting priority to the child:
2502 p->prio = current->normal_prio;
2503 if (!rt_prio(p->prio))
2504 p->sched_class = &fair_sched_class;
2506 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2507 if (likely(sched_info_on()))
2508 memset(&p->sched_info, 0, sizeof(p->sched_info));
2510 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2513 #ifdef CONFIG_PREEMPT
2514 /* Want to start with kernel preemption disabled. */
2515 task_thread_info(p)->preempt_count = 1;
2521 * wake_up_new_task - wake up a newly created task for the first time.
2523 * This function will do some initial scheduler statistics housekeeping
2524 * that must be done for every newly created context, then puts the task
2525 * on the runqueue and wakes it.
2527 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2529 unsigned long flags;
2532 rq = task_rq_lock(p, &flags);
2533 BUG_ON(p->state != TASK_RUNNING);
2534 update_rq_clock(rq);
2536 p->prio = effective_prio(p);
2538 if (!p->sched_class->task_new || !current->se.on_rq) {
2539 activate_task(rq, p, 0);
2542 * Let the scheduling class do new task startup
2543 * management (if any):
2545 p->sched_class->task_new(rq, p);
2548 check_preempt_curr(rq, p);
2550 if (p->sched_class->task_wake_up)
2551 p->sched_class->task_wake_up(rq, p);
2553 task_rq_unlock(rq, &flags);
2556 #ifdef CONFIG_PREEMPT_NOTIFIERS
2559 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2560 * @notifier: notifier struct to register
2562 void preempt_notifier_register(struct preempt_notifier *notifier)
2564 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2566 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2569 * preempt_notifier_unregister - no longer interested in preemption notifications
2570 * @notifier: notifier struct to unregister
2572 * This is safe to call from within a preemption notifier.
2574 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2576 hlist_del(¬ifier->link);
2578 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2580 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2582 struct preempt_notifier *notifier;
2583 struct hlist_node *node;
2585 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2586 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2590 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2591 struct task_struct *next)
2593 struct preempt_notifier *notifier;
2594 struct hlist_node *node;
2596 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2597 notifier->ops->sched_out(notifier, next);
2600 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2602 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2607 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2608 struct task_struct *next)
2612 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2615 * prepare_task_switch - prepare to switch tasks
2616 * @rq: the runqueue preparing to switch
2617 * @prev: the current task that is being switched out
2618 * @next: the task we are going to switch to.
2620 * This is called with the rq lock held and interrupts off. It must
2621 * be paired with a subsequent finish_task_switch after the context
2624 * prepare_task_switch sets up locking and calls architecture specific
2628 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2629 struct task_struct *next)
2631 fire_sched_out_preempt_notifiers(prev, next);
2632 prepare_lock_switch(rq, next);
2633 prepare_arch_switch(next);
2637 * finish_task_switch - clean up after a task-switch
2638 * @rq: runqueue associated with task-switch
2639 * @prev: the thread we just switched away from.
2641 * finish_task_switch must be called after the context switch, paired
2642 * with a prepare_task_switch call before the context switch.
2643 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2644 * and do any other architecture-specific cleanup actions.
2646 * Note that we may have delayed dropping an mm in context_switch(). If
2647 * so, we finish that here outside of the runqueue lock. (Doing it
2648 * with the lock held can cause deadlocks; see schedule() for
2651 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2652 __releases(rq->lock)
2654 struct mm_struct *mm = rq->prev_mm;
2660 * A task struct has one reference for the use as "current".
2661 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2662 * schedule one last time. The schedule call will never return, and
2663 * the scheduled task must drop that reference.
2664 * The test for TASK_DEAD must occur while the runqueue locks are
2665 * still held, otherwise prev could be scheduled on another cpu, die
2666 * there before we look at prev->state, and then the reference would
2668 * Manfred Spraul <manfred@colorfullife.com>
2670 prev_state = prev->state;
2671 finish_arch_switch(prev);
2672 finish_lock_switch(rq, prev);
2674 if (current->sched_class->post_schedule)
2675 current->sched_class->post_schedule(rq);
2678 fire_sched_in_preempt_notifiers(current);
2681 if (unlikely(prev_state == TASK_DEAD)) {
2683 * Remove function-return probe instances associated with this
2684 * task and put them back on the free list.
2686 kprobe_flush_task(prev);
2687 put_task_struct(prev);
2692 * schedule_tail - first thing a freshly forked thread must call.
2693 * @prev: the thread we just switched away from.
2695 asmlinkage void schedule_tail(struct task_struct *prev)
2696 __releases(rq->lock)
2698 struct rq *rq = this_rq();
2700 finish_task_switch(rq, prev);
2701 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2702 /* In this case, finish_task_switch does not reenable preemption */
2705 if (current->set_child_tid)
2706 put_user(task_pid_vnr(current), current->set_child_tid);
2710 * context_switch - switch to the new MM and the new
2711 * thread's register state.
2714 context_switch(struct rq *rq, struct task_struct *prev,
2715 struct task_struct *next)
2717 struct mm_struct *mm, *oldmm;
2719 prepare_task_switch(rq, prev, next);
2721 oldmm = prev->active_mm;
2723 * For paravirt, this is coupled with an exit in switch_to to
2724 * combine the page table reload and the switch backend into
2727 arch_enter_lazy_cpu_mode();
2729 if (unlikely(!mm)) {
2730 next->active_mm = oldmm;
2731 atomic_inc(&oldmm->mm_count);
2732 enter_lazy_tlb(oldmm, next);
2734 switch_mm(oldmm, mm, next);
2736 if (unlikely(!prev->mm)) {
2737 prev->active_mm = NULL;
2738 rq->prev_mm = oldmm;
2741 * Since the runqueue lock will be released by the next
2742 * task (which is an invalid locking op but in the case
2743 * of the scheduler it's an obvious special-case), so we
2744 * do an early lockdep release here:
2746 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2747 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2750 /* Here we just switch the register state and the stack. */
2751 switch_to(prev, next, prev);
2755 * this_rq must be evaluated again because prev may have moved
2756 * CPUs since it called schedule(), thus the 'rq' on its stack
2757 * frame will be invalid.
2759 finish_task_switch(this_rq(), prev);
2763 * nr_running, nr_uninterruptible and nr_context_switches:
2765 * externally visible scheduler statistics: current number of runnable
2766 * threads, current number of uninterruptible-sleeping threads, total
2767 * number of context switches performed since bootup.
2769 unsigned long nr_running(void)
2771 unsigned long i, sum = 0;
2773 for_each_online_cpu(i)
2774 sum += cpu_rq(i)->nr_running;
2779 unsigned long nr_uninterruptible(void)
2781 unsigned long i, sum = 0;
2783 for_each_possible_cpu(i)
2784 sum += cpu_rq(i)->nr_uninterruptible;
2787 * Since we read the counters lockless, it might be slightly
2788 * inaccurate. Do not allow it to go below zero though:
2790 if (unlikely((long)sum < 0))
2796 unsigned long long nr_context_switches(void)
2799 unsigned long long sum = 0;
2801 for_each_possible_cpu(i)
2802 sum += cpu_rq(i)->nr_switches;
2807 unsigned long nr_iowait(void)
2809 unsigned long i, sum = 0;
2811 for_each_possible_cpu(i)
2812 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2817 unsigned long nr_active(void)
2819 unsigned long i, running = 0, uninterruptible = 0;
2821 for_each_online_cpu(i) {
2822 running += cpu_rq(i)->nr_running;
2823 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2826 if (unlikely((long)uninterruptible < 0))
2827 uninterruptible = 0;
2829 return running + uninterruptible;
2833 * Update rq->cpu_load[] statistics. This function is usually called every
2834 * scheduler tick (TICK_NSEC).
2836 static void update_cpu_load(struct rq *this_rq)
2838 unsigned long this_load = this_rq->load.weight;
2841 this_rq->nr_load_updates++;
2843 /* Update our load: */
2844 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2845 unsigned long old_load, new_load;
2847 /* scale is effectively 1 << i now, and >> i divides by scale */
2849 old_load = this_rq->cpu_load[i];
2850 new_load = this_load;
2852 * Round up the averaging division if load is increasing. This
2853 * prevents us from getting stuck on 9 if the load is 10, for
2856 if (new_load > old_load)
2857 new_load += scale-1;
2858 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2865 * double_rq_lock - safely lock two runqueues
2867 * Note this does not disable interrupts like task_rq_lock,
2868 * you need to do so manually before calling.
2870 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2871 __acquires(rq1->lock)
2872 __acquires(rq2->lock)
2874 BUG_ON(!irqs_disabled());
2876 spin_lock(&rq1->lock);
2877 __acquire(rq2->lock); /* Fake it out ;) */
2880 spin_lock(&rq1->lock);
2881 spin_lock(&rq2->lock);
2883 spin_lock(&rq2->lock);
2884 spin_lock(&rq1->lock);
2887 update_rq_clock(rq1);
2888 update_rq_clock(rq2);
2892 * double_rq_unlock - safely unlock two runqueues
2894 * Note this does not restore interrupts like task_rq_unlock,
2895 * you need to do so manually after calling.
2897 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2898 __releases(rq1->lock)
2899 __releases(rq2->lock)
2901 spin_unlock(&rq1->lock);
2903 spin_unlock(&rq2->lock);
2905 __release(rq2->lock);
2909 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2911 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2912 __releases(this_rq->lock)
2913 __acquires(busiest->lock)
2914 __acquires(this_rq->lock)
2918 if (unlikely(!irqs_disabled())) {
2919 /* printk() doesn't work good under rq->lock */
2920 spin_unlock(&this_rq->lock);
2923 if (unlikely(!spin_trylock(&busiest->lock))) {
2924 if (busiest < this_rq) {
2925 spin_unlock(&this_rq->lock);
2926 spin_lock(&busiest->lock);
2927 spin_lock(&this_rq->lock);
2930 spin_lock(&busiest->lock);
2936 * If dest_cpu is allowed for this process, migrate the task to it.
2937 * This is accomplished by forcing the cpu_allowed mask to only
2938 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2939 * the cpu_allowed mask is restored.
2941 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2943 struct migration_req req;
2944 unsigned long flags;
2947 rq = task_rq_lock(p, &flags);
2948 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2949 || unlikely(cpu_is_offline(dest_cpu)))
2952 /* force the process onto the specified CPU */
2953 if (migrate_task(p, dest_cpu, &req)) {
2954 /* Need to wait for migration thread (might exit: take ref). */
2955 struct task_struct *mt = rq->migration_thread;
2957 get_task_struct(mt);
2958 task_rq_unlock(rq, &flags);
2959 wake_up_process(mt);
2960 put_task_struct(mt);
2961 wait_for_completion(&req.done);
2966 task_rq_unlock(rq, &flags);
2970 * sched_exec - execve() is a valuable balancing opportunity, because at
2971 * this point the task has the smallest effective memory and cache footprint.
2973 void sched_exec(void)
2975 int new_cpu, this_cpu = get_cpu();
2976 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2978 if (new_cpu != this_cpu)
2979 sched_migrate_task(current, new_cpu);
2983 * pull_task - move a task from a remote runqueue to the local runqueue.
2984 * Both runqueues must be locked.
2986 static void pull_task(struct rq *src_rq, struct task_struct *p,
2987 struct rq *this_rq, int this_cpu)
2989 deactivate_task(src_rq, p, 0);
2990 set_task_cpu(p, this_cpu);
2991 activate_task(this_rq, p, 0);
2993 * Note that idle threads have a prio of MAX_PRIO, for this test
2994 * to be always true for them.
2996 check_preempt_curr(this_rq, p);
3000 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3003 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3004 struct sched_domain *sd, enum cpu_idle_type idle,
3008 * We do not migrate tasks that are:
3009 * 1) running (obviously), or
3010 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3011 * 3) are cache-hot on their current CPU.
3013 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
3014 schedstat_inc(p, se.nr_failed_migrations_affine);
3019 if (task_running(rq, p)) {
3020 schedstat_inc(p, se.nr_failed_migrations_running);
3025 * Aggressive migration if:
3026 * 1) task is cache cold, or
3027 * 2) too many balance attempts have failed.
3030 if (!task_hot(p, rq->clock, sd) ||
3031 sd->nr_balance_failed > sd->cache_nice_tries) {
3032 #ifdef CONFIG_SCHEDSTATS
3033 if (task_hot(p, rq->clock, sd)) {
3034 schedstat_inc(sd, lb_hot_gained[idle]);
3035 schedstat_inc(p, se.nr_forced_migrations);
3041 if (task_hot(p, rq->clock, sd)) {
3042 schedstat_inc(p, se.nr_failed_migrations_hot);
3048 static unsigned long
3049 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3050 unsigned long max_load_move, struct sched_domain *sd,
3051 enum cpu_idle_type idle, int *all_pinned,
3052 int *this_best_prio, struct rq_iterator *iterator)
3054 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
3055 struct task_struct *p;
3056 long rem_load_move = max_load_move;
3058 if (max_load_move == 0)
3064 * Start the load-balancing iterator:
3066 p = iterator->start(iterator->arg);
3068 if (!p || loops++ > sysctl_sched_nr_migrate)
3071 * To help distribute high priority tasks across CPUs we don't
3072 * skip a task if it will be the highest priority task (i.e. smallest
3073 * prio value) on its new queue regardless of its load weight
3075 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
3076 SCHED_LOAD_SCALE_FUZZ;
3077 if ((skip_for_load && p->prio >= *this_best_prio) ||
3078 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3079 p = iterator->next(iterator->arg);
3083 pull_task(busiest, p, this_rq, this_cpu);
3085 rem_load_move -= p->se.load.weight;
3088 * We only want to steal up to the prescribed amount of weighted load.
3090 if (rem_load_move > 0) {
3091 if (p->prio < *this_best_prio)
3092 *this_best_prio = p->prio;
3093 p = iterator->next(iterator->arg);
3098 * Right now, this is one of only two places pull_task() is called,
3099 * so we can safely collect pull_task() stats here rather than
3100 * inside pull_task().
3102 schedstat_add(sd, lb_gained[idle], pulled);
3105 *all_pinned = pinned;
3107 return max_load_move - rem_load_move;
3111 * move_tasks tries to move up to max_load_move weighted load from busiest to
3112 * this_rq, as part of a balancing operation within domain "sd".
3113 * Returns 1 if successful and 0 otherwise.
3115 * Called with both runqueues locked.
3117 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3118 unsigned long max_load_move,
3119 struct sched_domain *sd, enum cpu_idle_type idle,
3122 const struct sched_class *class = sched_class_highest;
3123 unsigned long total_load_moved = 0;
3124 int this_best_prio = this_rq->curr->prio;
3128 class->load_balance(this_rq, this_cpu, busiest,
3129 max_load_move - total_load_moved,
3130 sd, idle, all_pinned, &this_best_prio);
3131 class = class->next;
3132 } while (class && max_load_move > total_load_moved);
3134 return total_load_moved > 0;
3138 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3139 struct sched_domain *sd, enum cpu_idle_type idle,
3140 struct rq_iterator *iterator)
3142 struct task_struct *p = iterator->start(iterator->arg);
3146 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3147 pull_task(busiest, p, this_rq, this_cpu);
3149 * Right now, this is only the second place pull_task()
3150 * is called, so we can safely collect pull_task()
3151 * stats here rather than inside pull_task().
3153 schedstat_inc(sd, lb_gained[idle]);
3157 p = iterator->next(iterator->arg);
3164 * move_one_task tries to move exactly one task from busiest to this_rq, as
3165 * part of active balancing operations within "domain".
3166 * Returns 1 if successful and 0 otherwise.
3168 * Called with both runqueues locked.
3170 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3171 struct sched_domain *sd, enum cpu_idle_type idle)
3173 const struct sched_class *class;
3175 for (class = sched_class_highest; class; class = class->next)
3176 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3183 * find_busiest_group finds and returns the busiest CPU group within the
3184 * domain. It calculates and returns the amount of weighted load which
3185 * should be moved to restore balance via the imbalance parameter.
3187 static struct sched_group *
3188 find_busiest_group(struct sched_domain *sd, int this_cpu,
3189 unsigned long *imbalance, enum cpu_idle_type idle,
3190 int *sd_idle, const cpumask_t *cpus, int *balance)
3192 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3193 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3194 unsigned long max_pull;
3195 unsigned long busiest_load_per_task, busiest_nr_running;
3196 unsigned long this_load_per_task, this_nr_running;
3197 int load_idx, group_imb = 0;
3198 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3199 int power_savings_balance = 1;
3200 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3201 unsigned long min_nr_running = ULONG_MAX;
3202 struct sched_group *group_min = NULL, *group_leader = NULL;
3205 max_load = this_load = total_load = total_pwr = 0;
3206 busiest_load_per_task = busiest_nr_running = 0;
3207 this_load_per_task = this_nr_running = 0;
3208 if (idle == CPU_NOT_IDLE)
3209 load_idx = sd->busy_idx;
3210 else if (idle == CPU_NEWLY_IDLE)
3211 load_idx = sd->newidle_idx;
3213 load_idx = sd->idle_idx;
3216 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3219 int __group_imb = 0;
3220 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3221 unsigned long sum_nr_running, sum_weighted_load;
3223 local_group = cpu_isset(this_cpu, group->cpumask);
3226 balance_cpu = first_cpu(group->cpumask);
3228 /* Tally up the load of all CPUs in the group */
3229 sum_weighted_load = sum_nr_running = avg_load = 0;
3231 min_cpu_load = ~0UL;
3233 for_each_cpu_mask(i, group->cpumask) {
3236 if (!cpu_isset(i, *cpus))
3241 if (*sd_idle && rq->nr_running)
3244 /* Bias balancing toward cpus of our domain */
3246 if (idle_cpu(i) && !first_idle_cpu) {
3251 load = target_load(i, load_idx);
3253 load = source_load(i, load_idx);
3254 if (load > max_cpu_load)
3255 max_cpu_load = load;
3256 if (min_cpu_load > load)
3257 min_cpu_load = load;
3261 sum_nr_running += rq->nr_running;
3262 sum_weighted_load += weighted_cpuload(i);
3266 * First idle cpu or the first cpu(busiest) in this sched group
3267 * is eligible for doing load balancing at this and above
3268 * domains. In the newly idle case, we will allow all the cpu's
3269 * to do the newly idle load balance.
3271 if (idle != CPU_NEWLY_IDLE && local_group &&
3272 balance_cpu != this_cpu && balance) {
3277 total_load += avg_load;
3278 total_pwr += group->__cpu_power;
3280 /* Adjust by relative CPU power of the group */
3281 avg_load = sg_div_cpu_power(group,
3282 avg_load * SCHED_LOAD_SCALE);
3284 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
3287 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3290 this_load = avg_load;
3292 this_nr_running = sum_nr_running;
3293 this_load_per_task = sum_weighted_load;
3294 } else if (avg_load > max_load &&
3295 (sum_nr_running > group_capacity || __group_imb)) {
3296 max_load = avg_load;
3298 busiest_nr_running = sum_nr_running;
3299 busiest_load_per_task = sum_weighted_load;
3300 group_imb = __group_imb;
3303 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3305 * Busy processors will not participate in power savings
3308 if (idle == CPU_NOT_IDLE ||
3309 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3313 * If the local group is idle or completely loaded
3314 * no need to do power savings balance at this domain
3316 if (local_group && (this_nr_running >= group_capacity ||
3318 power_savings_balance = 0;
3321 * If a group is already running at full capacity or idle,
3322 * don't include that group in power savings calculations
3324 if (!power_savings_balance || sum_nr_running >= group_capacity
3329 * Calculate the group which has the least non-idle load.
3330 * This is the group from where we need to pick up the load
3333 if ((sum_nr_running < min_nr_running) ||
3334 (sum_nr_running == min_nr_running &&
3335 first_cpu(group->cpumask) <
3336 first_cpu(group_min->cpumask))) {
3338 min_nr_running = sum_nr_running;
3339 min_load_per_task = sum_weighted_load /
3344 * Calculate the group which is almost near its
3345 * capacity but still has some space to pick up some load
3346 * from other group and save more power
3348 if (sum_nr_running <= group_capacity - 1) {
3349 if (sum_nr_running > leader_nr_running ||
3350 (sum_nr_running == leader_nr_running &&
3351 first_cpu(group->cpumask) >
3352 first_cpu(group_leader->cpumask))) {
3353 group_leader = group;
3354 leader_nr_running = sum_nr_running;
3359 group = group->next;
3360 } while (group != sd->groups);
3362 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3365 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3367 if (this_load >= avg_load ||
3368 100*max_load <= sd->imbalance_pct*this_load)
3371 busiest_load_per_task /= busiest_nr_running;
3373 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3376 * We're trying to get all the cpus to the average_load, so we don't
3377 * want to push ourselves above the average load, nor do we wish to
3378 * reduce the max loaded cpu below the average load, as either of these
3379 * actions would just result in more rebalancing later, and ping-pong
3380 * tasks around. Thus we look for the minimum possible imbalance.
3381 * Negative imbalances (*we* are more loaded than anyone else) will
3382 * be counted as no imbalance for these purposes -- we can't fix that
3383 * by pulling tasks to us. Be careful of negative numbers as they'll
3384 * appear as very large values with unsigned longs.
3386 if (max_load <= busiest_load_per_task)
3390 * In the presence of smp nice balancing, certain scenarios can have
3391 * max load less than avg load(as we skip the groups at or below
3392 * its cpu_power, while calculating max_load..)
3394 if (max_load < avg_load) {
3396 goto small_imbalance;
3399 /* Don't want to pull so many tasks that a group would go idle */
3400 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3402 /* How much load to actually move to equalise the imbalance */
3403 *imbalance = min(max_pull * busiest->__cpu_power,
3404 (avg_load - this_load) * this->__cpu_power)
3408 * if *imbalance is less than the average load per runnable task
3409 * there is no gaurantee that any tasks will be moved so we'll have
3410 * a think about bumping its value to force at least one task to be
3413 if (*imbalance < busiest_load_per_task) {
3414 unsigned long tmp, pwr_now, pwr_move;
3418 pwr_move = pwr_now = 0;
3420 if (this_nr_running) {
3421 this_load_per_task /= this_nr_running;
3422 if (busiest_load_per_task > this_load_per_task)
3425 this_load_per_task = SCHED_LOAD_SCALE;
3427 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
3428 busiest_load_per_task * imbn) {
3429 *imbalance = busiest_load_per_task;
3434 * OK, we don't have enough imbalance to justify moving tasks,
3435 * however we may be able to increase total CPU power used by
3439 pwr_now += busiest->__cpu_power *
3440 min(busiest_load_per_task, max_load);
3441 pwr_now += this->__cpu_power *
3442 min(this_load_per_task, this_load);
3443 pwr_now /= SCHED_LOAD_SCALE;
3445 /* Amount of load we'd subtract */
3446 tmp = sg_div_cpu_power(busiest,
3447 busiest_load_per_task * SCHED_LOAD_SCALE);
3449 pwr_move += busiest->__cpu_power *
3450 min(busiest_load_per_task, max_load - tmp);
3452 /* Amount of load we'd add */
3453 if (max_load * busiest->__cpu_power <
3454 busiest_load_per_task * SCHED_LOAD_SCALE)
3455 tmp = sg_div_cpu_power(this,
3456 max_load * busiest->__cpu_power);
3458 tmp = sg_div_cpu_power(this,
3459 busiest_load_per_task * SCHED_LOAD_SCALE);
3460 pwr_move += this->__cpu_power *
3461 min(this_load_per_task, this_load + tmp);
3462 pwr_move /= SCHED_LOAD_SCALE;
3464 /* Move if we gain throughput */
3465 if (pwr_move > pwr_now)
3466 *imbalance = busiest_load_per_task;
3472 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3473 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3476 if (this == group_leader && group_leader != group_min) {
3477 *imbalance = min_load_per_task;
3487 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3490 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3491 unsigned long imbalance, const cpumask_t *cpus)
3493 struct rq *busiest = NULL, *rq;
3494 unsigned long max_load = 0;
3497 for_each_cpu_mask(i, group->cpumask) {
3500 if (!cpu_isset(i, *cpus))
3504 wl = weighted_cpuload(i);
3506 if (rq->nr_running == 1 && wl > imbalance)
3509 if (wl > max_load) {
3519 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3520 * so long as it is large enough.
3522 #define MAX_PINNED_INTERVAL 512
3525 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3526 * tasks if there is an imbalance.
3528 static int load_balance(int this_cpu, struct rq *this_rq,
3529 struct sched_domain *sd, enum cpu_idle_type idle,
3530 int *balance, cpumask_t *cpus)
3532 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3533 struct sched_group *group;
3534 unsigned long imbalance;
3536 unsigned long flags;
3537 int unlock_aggregate;
3541 unlock_aggregate = get_aggregate(sd);
3544 * When power savings policy is enabled for the parent domain, idle
3545 * sibling can pick up load irrespective of busy siblings. In this case,
3546 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3547 * portraying it as CPU_NOT_IDLE.
3549 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3550 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3553 schedstat_inc(sd, lb_count[idle]);
3556 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3563 schedstat_inc(sd, lb_nobusyg[idle]);
3567 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3569 schedstat_inc(sd, lb_nobusyq[idle]);
3573 BUG_ON(busiest == this_rq);
3575 schedstat_add(sd, lb_imbalance[idle], imbalance);
3578 if (busiest->nr_running > 1) {
3580 * Attempt to move tasks. If find_busiest_group has found
3581 * an imbalance but busiest->nr_running <= 1, the group is
3582 * still unbalanced. ld_moved simply stays zero, so it is
3583 * correctly treated as an imbalance.
3585 local_irq_save(flags);
3586 double_rq_lock(this_rq, busiest);
3587 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3588 imbalance, sd, idle, &all_pinned);
3589 double_rq_unlock(this_rq, busiest);
3590 local_irq_restore(flags);
3593 * some other cpu did the load balance for us.
3595 if (ld_moved && this_cpu != smp_processor_id())
3596 resched_cpu(this_cpu);
3598 /* All tasks on this runqueue were pinned by CPU affinity */
3599 if (unlikely(all_pinned)) {
3600 cpu_clear(cpu_of(busiest), *cpus);
3601 if (!cpus_empty(*cpus))
3608 schedstat_inc(sd, lb_failed[idle]);
3609 sd->nr_balance_failed++;
3611 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3613 spin_lock_irqsave(&busiest->lock, flags);
3615 /* don't kick the migration_thread, if the curr
3616 * task on busiest cpu can't be moved to this_cpu
3618 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3619 spin_unlock_irqrestore(&busiest->lock, flags);
3621 goto out_one_pinned;
3624 if (!busiest->active_balance) {
3625 busiest->active_balance = 1;
3626 busiest->push_cpu = this_cpu;
3629 spin_unlock_irqrestore(&busiest->lock, flags);
3631 wake_up_process(busiest->migration_thread);
3634 * We've kicked active balancing, reset the failure
3637 sd->nr_balance_failed = sd->cache_nice_tries+1;
3640 sd->nr_balance_failed = 0;
3642 if (likely(!active_balance)) {
3643 /* We were unbalanced, so reset the balancing interval */
3644 sd->balance_interval = sd->min_interval;
3647 * If we've begun active balancing, start to back off. This
3648 * case may not be covered by the all_pinned logic if there
3649 * is only 1 task on the busy runqueue (because we don't call
3652 if (sd->balance_interval < sd->max_interval)
3653 sd->balance_interval *= 2;
3656 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3657 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3663 schedstat_inc(sd, lb_balanced[idle]);
3665 sd->nr_balance_failed = 0;
3668 /* tune up the balancing interval */
3669 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3670 (sd->balance_interval < sd->max_interval))
3671 sd->balance_interval *= 2;
3673 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3674 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3679 if (unlock_aggregate)
3685 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3686 * tasks if there is an imbalance.
3688 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3689 * this_rq is locked.
3692 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3695 struct sched_group *group;
3696 struct rq *busiest = NULL;
3697 unsigned long imbalance;
3705 * When power savings policy is enabled for the parent domain, idle
3706 * sibling can pick up load irrespective of busy siblings. In this case,
3707 * let the state of idle sibling percolate up as IDLE, instead of
3708 * portraying it as CPU_NOT_IDLE.
3710 if (sd->flags & SD_SHARE_CPUPOWER &&
3711 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3714 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3716 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3717 &sd_idle, cpus, NULL);
3719 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3723 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3725 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3729 BUG_ON(busiest == this_rq);
3731 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3734 if (busiest->nr_running > 1) {
3735 /* Attempt to move tasks */
3736 double_lock_balance(this_rq, busiest);
3737 /* this_rq->clock is already updated */
3738 update_rq_clock(busiest);
3739 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3740 imbalance, sd, CPU_NEWLY_IDLE,
3742 spin_unlock(&busiest->lock);
3744 if (unlikely(all_pinned)) {
3745 cpu_clear(cpu_of(busiest), *cpus);
3746 if (!cpus_empty(*cpus))
3752 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3753 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3754 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3757 sd->nr_balance_failed = 0;
3762 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3763 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3764 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3766 sd->nr_balance_failed = 0;
3772 * idle_balance is called by schedule() if this_cpu is about to become
3773 * idle. Attempts to pull tasks from other CPUs.
3775 static void idle_balance(int this_cpu, struct rq *this_rq)
3777 struct sched_domain *sd;
3778 int pulled_task = -1;
3779 unsigned long next_balance = jiffies + HZ;
3782 for_each_domain(this_cpu, sd) {
3783 unsigned long interval;
3785 if (!(sd->flags & SD_LOAD_BALANCE))
3788 if (sd->flags & SD_BALANCE_NEWIDLE)
3789 /* If we've pulled tasks over stop searching: */
3790 pulled_task = load_balance_newidle(this_cpu, this_rq,
3793 interval = msecs_to_jiffies(sd->balance_interval);
3794 if (time_after(next_balance, sd->last_balance + interval))
3795 next_balance = sd->last_balance + interval;
3799 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3801 * We are going idle. next_balance may be set based on
3802 * a busy processor. So reset next_balance.
3804 this_rq->next_balance = next_balance;
3809 * active_load_balance is run by migration threads. It pushes running tasks
3810 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3811 * running on each physical CPU where possible, and avoids physical /
3812 * logical imbalances.
3814 * Called with busiest_rq locked.
3816 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3818 int target_cpu = busiest_rq->push_cpu;
3819 struct sched_domain *sd;
3820 struct rq *target_rq;
3822 /* Is there any task to move? */
3823 if (busiest_rq->nr_running <= 1)
3826 target_rq = cpu_rq(target_cpu);
3829 * This condition is "impossible", if it occurs
3830 * we need to fix it. Originally reported by
3831 * Bjorn Helgaas on a 128-cpu setup.
3833 BUG_ON(busiest_rq == target_rq);
3835 /* move a task from busiest_rq to target_rq */
3836 double_lock_balance(busiest_rq, target_rq);
3837 update_rq_clock(busiest_rq);
3838 update_rq_clock(target_rq);
3840 /* Search for an sd spanning us and the target CPU. */
3841 for_each_domain(target_cpu, sd) {
3842 if ((sd->flags & SD_LOAD_BALANCE) &&
3843 cpu_isset(busiest_cpu, sd->span))
3848 schedstat_inc(sd, alb_count);
3850 if (move_one_task(target_rq, target_cpu, busiest_rq,
3852 schedstat_inc(sd, alb_pushed);
3854 schedstat_inc(sd, alb_failed);
3856 spin_unlock(&target_rq->lock);
3861 atomic_t load_balancer;
3863 } nohz ____cacheline_aligned = {
3864 .load_balancer = ATOMIC_INIT(-1),
3865 .cpu_mask = CPU_MASK_NONE,
3869 * This routine will try to nominate the ilb (idle load balancing)
3870 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3871 * load balancing on behalf of all those cpus. If all the cpus in the system
3872 * go into this tickless mode, then there will be no ilb owner (as there is
3873 * no need for one) and all the cpus will sleep till the next wakeup event
3876 * For the ilb owner, tick is not stopped. And this tick will be used
3877 * for idle load balancing. ilb owner will still be part of
3880 * While stopping the tick, this cpu will become the ilb owner if there
3881 * is no other owner. And will be the owner till that cpu becomes busy
3882 * or if all cpus in the system stop their ticks at which point
3883 * there is no need for ilb owner.
3885 * When the ilb owner becomes busy, it nominates another owner, during the
3886 * next busy scheduler_tick()
3888 int select_nohz_load_balancer(int stop_tick)
3890 int cpu = smp_processor_id();
3893 cpu_set(cpu, nohz.cpu_mask);
3894 cpu_rq(cpu)->in_nohz_recently = 1;
3897 * If we are going offline and still the leader, give up!
3899 if (cpu_is_offline(cpu) &&
3900 atomic_read(&nohz.load_balancer) == cpu) {
3901 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3906 /* time for ilb owner also to sleep */
3907 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3908 if (atomic_read(&nohz.load_balancer) == cpu)
3909 atomic_set(&nohz.load_balancer, -1);
3913 if (atomic_read(&nohz.load_balancer) == -1) {
3914 /* make me the ilb owner */
3915 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3917 } else if (atomic_read(&nohz.load_balancer) == cpu)
3920 if (!cpu_isset(cpu, nohz.cpu_mask))
3923 cpu_clear(cpu, nohz.cpu_mask);
3925 if (atomic_read(&nohz.load_balancer) == cpu)
3926 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3933 static DEFINE_SPINLOCK(balancing);
3936 * It checks each scheduling domain to see if it is due to be balanced,
3937 * and initiates a balancing operation if so.
3939 * Balancing parameters are set up in arch_init_sched_domains.
3941 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3944 struct rq *rq = cpu_rq(cpu);
3945 unsigned long interval;
3946 struct sched_domain *sd;
3947 /* Earliest time when we have to do rebalance again */
3948 unsigned long next_balance = jiffies + 60*HZ;
3949 int update_next_balance = 0;
3953 for_each_domain(cpu, sd) {
3954 if (!(sd->flags & SD_LOAD_BALANCE))
3957 interval = sd->balance_interval;
3958 if (idle != CPU_IDLE)
3959 interval *= sd->busy_factor;
3961 /* scale ms to jiffies */
3962 interval = msecs_to_jiffies(interval);
3963 if (unlikely(!interval))
3965 if (interval > HZ*NR_CPUS/10)
3966 interval = HZ*NR_CPUS/10;
3968 need_serialize = sd->flags & SD_SERIALIZE;
3970 if (need_serialize) {
3971 if (!spin_trylock(&balancing))
3975 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3976 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3978 * We've pulled tasks over so either we're no
3979 * longer idle, or one of our SMT siblings is
3982 idle = CPU_NOT_IDLE;
3984 sd->last_balance = jiffies;
3987 spin_unlock(&balancing);
3989 if (time_after(next_balance, sd->last_balance + interval)) {
3990 next_balance = sd->last_balance + interval;
3991 update_next_balance = 1;
3995 * Stop the load balance at this level. There is another
3996 * CPU in our sched group which is doing load balancing more
4004 * next_balance will be updated only when there is a need.
4005 * When the cpu is attached to null domain for ex, it will not be
4008 if (likely(update_next_balance))
4009 rq->next_balance = next_balance;
4013 * run_rebalance_domains is triggered when needed from the scheduler tick.
4014 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4015 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4017 static void run_rebalance_domains(struct softirq_action *h)
4019 int this_cpu = smp_processor_id();
4020 struct rq *this_rq = cpu_rq(this_cpu);
4021 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4022 CPU_IDLE : CPU_NOT_IDLE;
4024 rebalance_domains(this_cpu, idle);
4028 * If this cpu is the owner for idle load balancing, then do the
4029 * balancing on behalf of the other idle cpus whose ticks are
4032 if (this_rq->idle_at_tick &&
4033 atomic_read(&nohz.load_balancer) == this_cpu) {
4034 cpumask_t cpus = nohz.cpu_mask;
4038 cpu_clear(this_cpu, cpus);
4039 for_each_cpu_mask(balance_cpu, cpus) {
4041 * If this cpu gets work to do, stop the load balancing
4042 * work being done for other cpus. Next load
4043 * balancing owner will pick it up.
4048 rebalance_domains(balance_cpu, CPU_IDLE);
4050 rq = cpu_rq(balance_cpu);
4051 if (time_after(this_rq->next_balance, rq->next_balance))
4052 this_rq->next_balance = rq->next_balance;
4059 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4061 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4062 * idle load balancing owner or decide to stop the periodic load balancing,
4063 * if the whole system is idle.
4065 static inline void trigger_load_balance(struct rq *rq, int cpu)
4069 * If we were in the nohz mode recently and busy at the current
4070 * scheduler tick, then check if we need to nominate new idle
4073 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4074 rq->in_nohz_recently = 0;
4076 if (atomic_read(&nohz.load_balancer) == cpu) {
4077 cpu_clear(cpu, nohz.cpu_mask);
4078 atomic_set(&nohz.load_balancer, -1);
4081 if (atomic_read(&nohz.load_balancer) == -1) {
4083 * simple selection for now: Nominate the
4084 * first cpu in the nohz list to be the next
4087 * TBD: Traverse the sched domains and nominate
4088 * the nearest cpu in the nohz.cpu_mask.
4090 int ilb = first_cpu(nohz.cpu_mask);
4092 if (ilb < nr_cpu_ids)
4098 * If this cpu is idle and doing idle load balancing for all the
4099 * cpus with ticks stopped, is it time for that to stop?
4101 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4102 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4108 * If this cpu is idle and the idle load balancing is done by
4109 * someone else, then no need raise the SCHED_SOFTIRQ
4111 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4112 cpu_isset(cpu, nohz.cpu_mask))
4115 if (time_after_eq(jiffies, rq->next_balance))
4116 raise_softirq(SCHED_SOFTIRQ);
4119 #else /* CONFIG_SMP */
4122 * on UP we do not need to balance between CPUs:
4124 static inline void idle_balance(int cpu, struct rq *rq)
4130 DEFINE_PER_CPU(struct kernel_stat, kstat);
4132 EXPORT_PER_CPU_SYMBOL(kstat);
4135 * Return p->sum_exec_runtime plus any more ns on the sched_clock
4136 * that have not yet been banked in case the task is currently running.
4138 unsigned long long task_sched_runtime(struct task_struct *p)
4140 unsigned long flags;
4144 rq = task_rq_lock(p, &flags);
4145 ns = p->se.sum_exec_runtime;
4146 if (task_current(rq, p)) {
4147 update_rq_clock(rq);
4148 delta_exec = rq->clock - p->se.exec_start;
4149 if ((s64)delta_exec > 0)
4152 task_rq_unlock(rq, &flags);
4158 * Account user cpu time to a process.
4159 * @p: the process that the cpu time gets accounted to
4160 * @cputime: the cpu time spent in user space since the last update
4162 void account_user_time(struct task_struct *p, cputime_t cputime)
4164 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4167 p->utime = cputime_add(p->utime, cputime);
4169 /* Add user time to cpustat. */
4170 tmp = cputime_to_cputime64(cputime);
4171 if (TASK_NICE(p) > 0)
4172 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4174 cpustat->user = cputime64_add(cpustat->user, tmp);
4178 * Account guest cpu time to a process.
4179 * @p: the process that the cpu time gets accounted to
4180 * @cputime: the cpu time spent in virtual machine since the last update
4182 static void account_guest_time(struct task_struct *p, cputime_t cputime)
4185 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4187 tmp = cputime_to_cputime64(cputime);
4189 p->utime = cputime_add(p->utime, cputime);
4190 p->gtime = cputime_add(p->gtime, cputime);
4192 cpustat->user = cputime64_add(cpustat->user, tmp);
4193 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4197 * Account scaled user cpu time to a process.
4198 * @p: the process that the cpu time gets accounted to
4199 * @cputime: the cpu time spent in user space since the last update
4201 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4203 p->utimescaled = cputime_add(p->utimescaled, cputime);
4207 * Account system cpu time to a process.
4208 * @p: the process that the cpu time gets accounted to
4209 * @hardirq_offset: the offset to subtract from hardirq_count()
4210 * @cputime: the cpu time spent in kernel space since the last update
4212 void account_system_time(struct task_struct *p, int hardirq_offset,
4215 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4216 struct rq *rq = this_rq();
4219 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4220 account_guest_time(p, cputime);
4224 p->stime = cputime_add(p->stime, cputime);
4226 /* Add system time to cpustat. */
4227 tmp = cputime_to_cputime64(cputime);
4228 if (hardirq_count() - hardirq_offset)
4229 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4230 else if (softirq_count())
4231 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4232 else if (p != rq->idle)
4233 cpustat->system = cputime64_add(cpustat->system, tmp);
4234 else if (atomic_read(&rq->nr_iowait) > 0)
4235 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4237 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4238 /* Account for system time used */
4239 acct_update_integrals(p);
4243 * Account scaled system cpu time to a process.
4244 * @p: the process that the cpu time gets accounted to
4245 * @hardirq_offset: the offset to subtract from hardirq_count()
4246 * @cputime: the cpu time spent in kernel space since the last update
4248 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4250 p->stimescaled = cputime_add(p->stimescaled, cputime);
4254 * Account for involuntary wait time.
4255 * @p: the process from which the cpu time has been stolen
4256 * @steal: the cpu time spent in involuntary wait
4258 void account_steal_time(struct task_struct *p, cputime_t steal)
4260 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4261 cputime64_t tmp = cputime_to_cputime64(steal);
4262 struct rq *rq = this_rq();
4264 if (p == rq->idle) {
4265 p->stime = cputime_add(p->stime, steal);
4266 if (atomic_read(&rq->nr_iowait) > 0)
4267 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4269 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4271 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4275 * This function gets called by the timer code, with HZ frequency.
4276 * We call it with interrupts disabled.
4278 * It also gets called by the fork code, when changing the parent's
4281 void scheduler_tick(void)
4283 int cpu = smp_processor_id();
4284 struct rq *rq = cpu_rq(cpu);
4285 struct task_struct *curr = rq->curr;
4289 spin_lock(&rq->lock);
4290 update_rq_clock(rq);
4291 update_cpu_load(rq);
4292 curr->sched_class->task_tick(rq, curr, 0);
4293 spin_unlock(&rq->lock);
4296 rq->idle_at_tick = idle_cpu(cpu);
4297 trigger_load_balance(rq, cpu);
4301 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
4303 void __kprobes add_preempt_count(int val)
4308 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4310 preempt_count() += val;
4312 * Spinlock count overflowing soon?
4314 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4317 EXPORT_SYMBOL(add_preempt_count);
4319 void __kprobes sub_preempt_count(int val)
4324 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4327 * Is the spinlock portion underflowing?
4329 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4330 !(preempt_count() & PREEMPT_MASK)))
4333 preempt_count() -= val;
4335 EXPORT_SYMBOL(sub_preempt_count);
4340 * Print scheduling while atomic bug:
4342 static noinline void __schedule_bug(struct task_struct *prev)
4344 struct pt_regs *regs = get_irq_regs();
4346 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4347 prev->comm, prev->pid, preempt_count());
4349 debug_show_held_locks(prev);
4351 if (irqs_disabled())
4352 print_irqtrace_events(prev);
4361 * Various schedule()-time debugging checks and statistics:
4363 static inline void schedule_debug(struct task_struct *prev)
4366 * Test if we are atomic. Since do_exit() needs to call into
4367 * schedule() atomically, we ignore that path for now.
4368 * Otherwise, whine if we are scheduling when we should not be.
4370 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4371 __schedule_bug(prev);
4373 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4375 schedstat_inc(this_rq(), sched_count);
4376 #ifdef CONFIG_SCHEDSTATS
4377 if (unlikely(prev->lock_depth >= 0)) {
4378 schedstat_inc(this_rq(), bkl_count);
4379 schedstat_inc(prev, sched_info.bkl_count);
4385 * Pick up the highest-prio task:
4387 static inline struct task_struct *
4388 pick_next_task(struct rq *rq, struct task_struct *prev)
4390 const struct sched_class *class;
4391 struct task_struct *p;
4394 * Optimization: we know that if all tasks are in
4395 * the fair class we can call that function directly:
4397 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4398 p = fair_sched_class.pick_next_task(rq);
4403 class = sched_class_highest;
4405 p = class->pick_next_task(rq);
4409 * Will never be NULL as the idle class always
4410 * returns a non-NULL p:
4412 class = class->next;
4417 * schedule() is the main scheduler function.
4419 asmlinkage void __sched schedule(void)
4421 struct task_struct *prev, *next;
4422 unsigned long *switch_count;
4424 int cpu, hrtick = sched_feat(HRTICK);
4428 cpu = smp_processor_id();
4432 switch_count = &prev->nivcsw;
4434 release_kernel_lock(prev);
4435 need_resched_nonpreemptible:
4437 schedule_debug(prev);
4443 * Do the rq-clock update outside the rq lock:
4445 local_irq_disable();
4446 update_rq_clock(rq);
4447 spin_lock(&rq->lock);
4448 clear_tsk_need_resched(prev);
4450 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4451 if (unlikely(signal_pending_state(prev->state, prev)))
4452 prev->state = TASK_RUNNING;
4454 deactivate_task(rq, prev, 1);
4455 switch_count = &prev->nvcsw;
4459 if (prev->sched_class->pre_schedule)
4460 prev->sched_class->pre_schedule(rq, prev);
4463 if (unlikely(!rq->nr_running))
4464 idle_balance(cpu, rq);
4466 prev->sched_class->put_prev_task(rq, prev);
4467 next = pick_next_task(rq, prev);
4469 if (likely(prev != next)) {
4470 sched_info_switch(prev, next);
4476 context_switch(rq, prev, next); /* unlocks the rq */
4478 * the context switch might have flipped the stack from under
4479 * us, hence refresh the local variables.
4481 cpu = smp_processor_id();
4484 spin_unlock_irq(&rq->lock);
4489 if (unlikely(reacquire_kernel_lock(current) < 0))
4490 goto need_resched_nonpreemptible;
4492 preempt_enable_no_resched();
4493 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4496 EXPORT_SYMBOL(schedule);
4498 #ifdef CONFIG_PREEMPT
4500 * this is the entry point to schedule() from in-kernel preemption
4501 * off of preempt_enable. Kernel preemptions off return from interrupt
4502 * occur there and call schedule directly.
4504 asmlinkage void __sched preempt_schedule(void)
4506 struct thread_info *ti = current_thread_info();
4509 * If there is a non-zero preempt_count or interrupts are disabled,
4510 * we do not want to preempt the current task. Just return..
4512 if (likely(ti->preempt_count || irqs_disabled()))
4516 add_preempt_count(PREEMPT_ACTIVE);
4518 sub_preempt_count(PREEMPT_ACTIVE);
4521 * Check again in case we missed a preemption opportunity
4522 * between schedule and now.
4525 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4527 EXPORT_SYMBOL(preempt_schedule);
4530 * this is the entry point to schedule() from kernel preemption
4531 * off of irq context.
4532 * Note, that this is called and return with irqs disabled. This will
4533 * protect us against recursive calling from irq.
4535 asmlinkage void __sched preempt_schedule_irq(void)
4537 struct thread_info *ti = current_thread_info();
4539 /* Catch callers which need to be fixed */
4540 BUG_ON(ti->preempt_count || !irqs_disabled());
4543 add_preempt_count(PREEMPT_ACTIVE);
4546 local_irq_disable();
4547 sub_preempt_count(PREEMPT_ACTIVE);
4550 * Check again in case we missed a preemption opportunity
4551 * between schedule and now.
4554 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4557 #endif /* CONFIG_PREEMPT */
4559 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4562 return try_to_wake_up(curr->private, mode, sync);
4564 EXPORT_SYMBOL(default_wake_function);
4567 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4568 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4569 * number) then we wake all the non-exclusive tasks and one exclusive task.
4571 * There are circumstances in which we can try to wake a task which has already
4572 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4573 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4575 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4576 int nr_exclusive, int sync, void *key)
4578 wait_queue_t *curr, *next;
4580 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4581 unsigned flags = curr->flags;
4583 if (curr->func(curr, mode, sync, key) &&
4584 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4590 * __wake_up - wake up threads blocked on a waitqueue.
4592 * @mode: which threads
4593 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4594 * @key: is directly passed to the wakeup function
4596 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4597 int nr_exclusive, void *key)
4599 unsigned long flags;
4601 spin_lock_irqsave(&q->lock, flags);
4602 __wake_up_common(q, mode, nr_exclusive, 0, key);
4603 spin_unlock_irqrestore(&q->lock, flags);
4605 EXPORT_SYMBOL(__wake_up);
4608 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4610 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4612 __wake_up_common(q, mode, 1, 0, NULL);
4616 * __wake_up_sync - wake up threads blocked on a waitqueue.
4618 * @mode: which threads
4619 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4621 * The sync wakeup differs that the waker knows that it will schedule
4622 * away soon, so while the target thread will be woken up, it will not
4623 * be migrated to another CPU - ie. the two threads are 'synchronized'
4624 * with each other. This can prevent needless bouncing between CPUs.
4626 * On UP it can prevent extra preemption.
4629 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4631 unsigned long flags;
4637 if (unlikely(!nr_exclusive))
4640 spin_lock_irqsave(&q->lock, flags);
4641 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4642 spin_unlock_irqrestore(&q->lock, flags);
4644 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4646 void complete(struct completion *x)
4648 unsigned long flags;
4650 spin_lock_irqsave(&x->wait.lock, flags);
4652 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4653 spin_unlock_irqrestore(&x->wait.lock, flags);
4655 EXPORT_SYMBOL(complete);
4657 void complete_all(struct completion *x)
4659 unsigned long flags;
4661 spin_lock_irqsave(&x->wait.lock, flags);
4662 x->done += UINT_MAX/2;
4663 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4664 spin_unlock_irqrestore(&x->wait.lock, flags);
4666 EXPORT_SYMBOL(complete_all);
4668 static inline long __sched
4669 do_wait_for_common(struct completion *x, long timeout, int state)
4672 DECLARE_WAITQUEUE(wait, current);
4674 wait.flags |= WQ_FLAG_EXCLUSIVE;
4675 __add_wait_queue_tail(&x->wait, &wait);
4677 if ((state == TASK_INTERRUPTIBLE &&
4678 signal_pending(current)) ||
4679 (state == TASK_KILLABLE &&
4680 fatal_signal_pending(current))) {
4681 timeout = -ERESTARTSYS;
4684 __set_current_state(state);
4685 spin_unlock_irq(&x->wait.lock);
4686 timeout = schedule_timeout(timeout);
4687 spin_lock_irq(&x->wait.lock);
4688 } while (!x->done && timeout);
4689 __remove_wait_queue(&x->wait, &wait);
4694 return timeout ?: 1;
4698 wait_for_common(struct completion *x, long timeout, int state)
4702 spin_lock_irq(&x->wait.lock);
4703 timeout = do_wait_for_common(x, timeout, state);
4704 spin_unlock_irq(&x->wait.lock);
4708 void __sched wait_for_completion(struct completion *x)
4710 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4712 EXPORT_SYMBOL(wait_for_completion);
4714 unsigned long __sched
4715 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4717 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4719 EXPORT_SYMBOL(wait_for_completion_timeout);
4721 int __sched wait_for_completion_interruptible(struct completion *x)
4723 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4724 if (t == -ERESTARTSYS)
4728 EXPORT_SYMBOL(wait_for_completion_interruptible);
4730 unsigned long __sched
4731 wait_for_completion_interruptible_timeout(struct completion *x,
4732 unsigned long timeout)
4734 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4736 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4738 int __sched wait_for_completion_killable(struct completion *x)
4740 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4741 if (t == -ERESTARTSYS)
4745 EXPORT_SYMBOL(wait_for_completion_killable);
4748 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4750 unsigned long flags;
4753 init_waitqueue_entry(&wait, current);
4755 __set_current_state(state);
4757 spin_lock_irqsave(&q->lock, flags);
4758 __add_wait_queue(q, &wait);
4759 spin_unlock(&q->lock);
4760 timeout = schedule_timeout(timeout);
4761 spin_lock_irq(&q->lock);
4762 __remove_wait_queue(q, &wait);
4763 spin_unlock_irqrestore(&q->lock, flags);
4768 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4770 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4772 EXPORT_SYMBOL(interruptible_sleep_on);
4775 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4777 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4779 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4781 void __sched sleep_on(wait_queue_head_t *q)
4783 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4785 EXPORT_SYMBOL(sleep_on);
4787 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4789 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4791 EXPORT_SYMBOL(sleep_on_timeout);
4793 #ifdef CONFIG_RT_MUTEXES
4796 * rt_mutex_setprio - set the current priority of a task
4798 * @prio: prio value (kernel-internal form)
4800 * This function changes the 'effective' priority of a task. It does
4801 * not touch ->normal_prio like __setscheduler().
4803 * Used by the rt_mutex code to implement priority inheritance logic.
4805 void rt_mutex_setprio(struct task_struct *p, int prio)
4807 unsigned long flags;
4808 int oldprio, on_rq, running;
4810 const struct sched_class *prev_class = p->sched_class;
4812 BUG_ON(prio < 0 || prio > MAX_PRIO);
4814 rq = task_rq_lock(p, &flags);
4815 update_rq_clock(rq);
4818 on_rq = p->se.on_rq;
4819 running = task_current(rq, p);
4821 dequeue_task(rq, p, 0);
4823 p->sched_class->put_prev_task(rq, p);
4826 p->sched_class = &rt_sched_class;
4828 p->sched_class = &fair_sched_class;
4833 p->sched_class->set_curr_task(rq);
4835 enqueue_task(rq, p, 0);
4837 check_class_changed(rq, p, prev_class, oldprio, running);
4839 task_rq_unlock(rq, &flags);
4844 void set_user_nice(struct task_struct *p, long nice)
4846 int old_prio, delta, on_rq;
4847 unsigned long flags;
4850 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4853 * We have to be careful, if called from sys_setpriority(),
4854 * the task might be in the middle of scheduling on another CPU.
4856 rq = task_rq_lock(p, &flags);
4857 update_rq_clock(rq);
4859 * The RT priorities are set via sched_setscheduler(), but we still
4860 * allow the 'normal' nice value to be set - but as expected
4861 * it wont have any effect on scheduling until the task is
4862 * SCHED_FIFO/SCHED_RR:
4864 if (task_has_rt_policy(p)) {
4865 p->static_prio = NICE_TO_PRIO(nice);
4868 on_rq = p->se.on_rq;
4870 dequeue_task(rq, p, 0);
4872 p->static_prio = NICE_TO_PRIO(nice);
4875 p->prio = effective_prio(p);
4876 delta = p->prio - old_prio;
4879 enqueue_task(rq, p, 0);
4881 * If the task increased its priority or is running and
4882 * lowered its priority, then reschedule its CPU:
4884 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4885 resched_task(rq->curr);
4888 task_rq_unlock(rq, &flags);
4890 EXPORT_SYMBOL(set_user_nice);
4893 * can_nice - check if a task can reduce its nice value
4897 int can_nice(const struct task_struct *p, const int nice)
4899 /* convert nice value [19,-20] to rlimit style value [1,40] */
4900 int nice_rlim = 20 - nice;
4902 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4903 capable(CAP_SYS_NICE));
4906 #ifdef __ARCH_WANT_SYS_NICE
4909 * sys_nice - change the priority of the current process.
4910 * @increment: priority increment
4912 * sys_setpriority is a more generic, but much slower function that
4913 * does similar things.
4915 asmlinkage long sys_nice(int increment)
4920 * Setpriority might change our priority at the same moment.
4921 * We don't have to worry. Conceptually one call occurs first
4922 * and we have a single winner.
4924 if (increment < -40)
4929 nice = PRIO_TO_NICE(current->static_prio) + increment;
4935 if (increment < 0 && !can_nice(current, nice))
4938 retval = security_task_setnice(current, nice);
4942 set_user_nice(current, nice);
4949 * task_prio - return the priority value of a given task.
4950 * @p: the task in question.
4952 * This is the priority value as seen by users in /proc.
4953 * RT tasks are offset by -200. Normal tasks are centered
4954 * around 0, value goes from -16 to +15.
4956 int task_prio(const struct task_struct *p)
4958 return p->prio - MAX_RT_PRIO;
4962 * task_nice - return the nice value of a given task.
4963 * @p: the task in question.
4965 int task_nice(const struct task_struct *p)
4967 return TASK_NICE(p);
4969 EXPORT_SYMBOL(task_nice);
4972 * idle_cpu - is a given cpu idle currently?
4973 * @cpu: the processor in question.
4975 int idle_cpu(int cpu)
4977 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4981 * idle_task - return the idle task for a given cpu.
4982 * @cpu: the processor in question.
4984 struct task_struct *idle_task(int cpu)
4986 return cpu_rq(cpu)->idle;
4990 * find_process_by_pid - find a process with a matching PID value.
4991 * @pid: the pid in question.
4993 static struct task_struct *find_process_by_pid(pid_t pid)
4995 return pid ? find_task_by_vpid(pid) : current;
4998 /* Actually do priority change: must hold rq lock. */
5000 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5002 BUG_ON(p->se.on_rq);
5005 switch (p->policy) {
5009 p->sched_class = &fair_sched_class;
5013 p->sched_class = &rt_sched_class;
5017 p->rt_priority = prio;
5018 p->normal_prio = normal_prio(p);
5019 /* we are holding p->pi_lock already */
5020 p->prio = rt_mutex_getprio(p);
5025 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5026 * @p: the task in question.
5027 * @policy: new policy.
5028 * @param: structure containing the new RT priority.
5030 * NOTE that the task may be already dead.
5032 int sched_setscheduler(struct task_struct *p, int policy,
5033 struct sched_param *param)
5035 int retval, oldprio, oldpolicy = -1, on_rq, running;
5036 unsigned long flags;
5037 const struct sched_class *prev_class = p->sched_class;
5040 /* may grab non-irq protected spin_locks */
5041 BUG_ON(in_interrupt());
5043 /* double check policy once rq lock held */
5045 policy = oldpolicy = p->policy;
5046 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5047 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5048 policy != SCHED_IDLE)
5051 * Valid priorities for SCHED_FIFO and SCHED_RR are
5052 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5053 * SCHED_BATCH and SCHED_IDLE is 0.
5055 if (param->sched_priority < 0 ||
5056 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5057 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5059 if (rt_policy(policy) != (param->sched_priority != 0))
5063 * Allow unprivileged RT tasks to decrease priority:
5065 if (!capable(CAP_SYS_NICE)) {
5066 if (rt_policy(policy)) {
5067 unsigned long rlim_rtprio;
5069 if (!lock_task_sighand(p, &flags))
5071 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5072 unlock_task_sighand(p, &flags);
5074 /* can't set/change the rt policy */
5075 if (policy != p->policy && !rlim_rtprio)
5078 /* can't increase priority */
5079 if (param->sched_priority > p->rt_priority &&
5080 param->sched_priority > rlim_rtprio)
5084 * Like positive nice levels, dont allow tasks to
5085 * move out of SCHED_IDLE either:
5087 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5090 /* can't change other user's priorities */
5091 if ((current->euid != p->euid) &&
5092 (current->euid != p->uid))
5096 #ifdef CONFIG_RT_GROUP_SCHED
5098 * Do not allow realtime tasks into groups that have no runtime
5101 if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
5105 retval = security_task_setscheduler(p, policy, param);
5109 * make sure no PI-waiters arrive (or leave) while we are
5110 * changing the priority of the task:
5112 spin_lock_irqsave(&p->pi_lock, flags);
5114 * To be able to change p->policy safely, the apropriate
5115 * runqueue lock must be held.
5117 rq = __task_rq_lock(p);
5118 /* recheck policy now with rq lock held */
5119 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5120 policy = oldpolicy = -1;
5121 __task_rq_unlock(rq);
5122 spin_unlock_irqrestore(&p->pi_lock, flags);
5125 update_rq_clock(rq);
5126 on_rq = p->se.on_rq;
5127 running = task_current(rq, p);
5129 deactivate_task(rq, p, 0);
5131 p->sched_class->put_prev_task(rq, p);
5134 __setscheduler(rq, p, policy, param->sched_priority);
5137 p->sched_class->set_curr_task(rq);
5139 activate_task(rq, p, 0);
5141 check_class_changed(rq, p, prev_class, oldprio, running);
5143 __task_rq_unlock(rq);
5144 spin_unlock_irqrestore(&p->pi_lock, flags);
5146 rt_mutex_adjust_pi(p);
5150 EXPORT_SYMBOL_GPL(sched_setscheduler);
5153 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5155 struct sched_param lparam;
5156 struct task_struct *p;
5159 if (!param || pid < 0)
5161 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5166 p = find_process_by_pid(pid);
5168 retval = sched_setscheduler(p, policy, &lparam);
5175 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5176 * @pid: the pid in question.
5177 * @policy: new policy.
5178 * @param: structure containing the new RT priority.
5181 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5183 /* negative values for policy are not valid */
5187 return do_sched_setscheduler(pid, policy, param);
5191 * sys_sched_setparam - set/change the RT priority of a thread
5192 * @pid: the pid in question.
5193 * @param: structure containing the new RT priority.
5195 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5197 return do_sched_setscheduler(pid, -1, param);
5201 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5202 * @pid: the pid in question.
5204 asmlinkage long sys_sched_getscheduler(pid_t pid)
5206 struct task_struct *p;
5213 read_lock(&tasklist_lock);
5214 p = find_process_by_pid(pid);
5216 retval = security_task_getscheduler(p);
5220 read_unlock(&tasklist_lock);
5225 * sys_sched_getscheduler - get the RT priority of a thread
5226 * @pid: the pid in question.
5227 * @param: structure containing the RT priority.
5229 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5231 struct sched_param lp;
5232 struct task_struct *p;
5235 if (!param || pid < 0)
5238 read_lock(&tasklist_lock);
5239 p = find_process_by_pid(pid);
5244 retval = security_task_getscheduler(p);
5248 lp.sched_priority = p->rt_priority;
5249 read_unlock(&tasklist_lock);
5252 * This one might sleep, we cannot do it with a spinlock held ...
5254 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5259 read_unlock(&tasklist_lock);
5263 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5265 cpumask_t cpus_allowed;
5266 cpumask_t new_mask = *in_mask;
5267 struct task_struct *p;
5271 read_lock(&tasklist_lock);
5273 p = find_process_by_pid(pid);
5275 read_unlock(&tasklist_lock);
5281 * It is not safe to call set_cpus_allowed with the
5282 * tasklist_lock held. We will bump the task_struct's
5283 * usage count and then drop tasklist_lock.
5286 read_unlock(&tasklist_lock);
5289 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5290 !capable(CAP_SYS_NICE))
5293 retval = security_task_setscheduler(p, 0, NULL);
5297 cpuset_cpus_allowed(p, &cpus_allowed);
5298 cpus_and(new_mask, new_mask, cpus_allowed);
5300 retval = set_cpus_allowed_ptr(p, &new_mask);
5303 cpuset_cpus_allowed(p, &cpus_allowed);
5304 if (!cpus_subset(new_mask, cpus_allowed)) {
5306 * We must have raced with a concurrent cpuset
5307 * update. Just reset the cpus_allowed to the
5308 * cpuset's cpus_allowed
5310 new_mask = cpus_allowed;
5320 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5321 cpumask_t *new_mask)
5323 if (len < sizeof(cpumask_t)) {
5324 memset(new_mask, 0, sizeof(cpumask_t));
5325 } else if (len > sizeof(cpumask_t)) {
5326 len = sizeof(cpumask_t);
5328 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5332 * sys_sched_setaffinity - set the cpu affinity of a process
5333 * @pid: pid of the process
5334 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5335 * @user_mask_ptr: user-space pointer to the new cpu mask
5337 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5338 unsigned long __user *user_mask_ptr)
5343 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5347 return sched_setaffinity(pid, &new_mask);
5350 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5352 struct task_struct *p;
5356 read_lock(&tasklist_lock);
5359 p = find_process_by_pid(pid);
5363 retval = security_task_getscheduler(p);
5367 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5370 read_unlock(&tasklist_lock);
5377 * sys_sched_getaffinity - get the cpu affinity of a process
5378 * @pid: pid of the process
5379 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5380 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5382 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5383 unsigned long __user *user_mask_ptr)
5388 if (len < sizeof(cpumask_t))
5391 ret = sched_getaffinity(pid, &mask);
5395 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5398 return sizeof(cpumask_t);
5402 * sys_sched_yield - yield the current processor to other threads.
5404 * This function yields the current CPU to other tasks. If there are no
5405 * other threads running on this CPU then this function will return.
5407 asmlinkage long sys_sched_yield(void)
5409 struct rq *rq = this_rq_lock();
5411 schedstat_inc(rq, yld_count);
5412 current->sched_class->yield_task(rq);
5415 * Since we are going to call schedule() anyway, there's
5416 * no need to preempt or enable interrupts:
5418 __release(rq->lock);
5419 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5420 _raw_spin_unlock(&rq->lock);
5421 preempt_enable_no_resched();
5428 static void __cond_resched(void)
5430 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5431 __might_sleep(__FILE__, __LINE__);
5434 * The BKS might be reacquired before we have dropped
5435 * PREEMPT_ACTIVE, which could trigger a second
5436 * cond_resched() call.
5439 add_preempt_count(PREEMPT_ACTIVE);
5441 sub_preempt_count(PREEMPT_ACTIVE);
5442 } while (need_resched());
5445 int __sched _cond_resched(void)
5447 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5448 system_state == SYSTEM_RUNNING) {
5454 EXPORT_SYMBOL(_cond_resched);
5457 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5458 * call schedule, and on return reacquire the lock.
5460 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5461 * operations here to prevent schedule() from being called twice (once via
5462 * spin_unlock(), once by hand).
5464 int cond_resched_lock(spinlock_t *lock)
5466 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5469 if (spin_needbreak(lock) || resched) {
5471 if (resched && need_resched())
5480 EXPORT_SYMBOL(cond_resched_lock);
5482 int __sched cond_resched_softirq(void)
5484 BUG_ON(!in_softirq());
5486 if (need_resched() && system_state == SYSTEM_RUNNING) {
5494 EXPORT_SYMBOL(cond_resched_softirq);
5497 * yield - yield the current processor to other threads.
5499 * This is a shortcut for kernel-space yielding - it marks the
5500 * thread runnable and calls sys_sched_yield().
5502 void __sched yield(void)
5504 set_current_state(TASK_RUNNING);
5507 EXPORT_SYMBOL(yield);
5510 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5511 * that process accounting knows that this is a task in IO wait state.
5513 * But don't do that if it is a deliberate, throttling IO wait (this task
5514 * has set its backing_dev_info: the queue against which it should throttle)
5516 void __sched io_schedule(void)
5518 struct rq *rq = &__raw_get_cpu_var(runqueues);
5520 delayacct_blkio_start();
5521 atomic_inc(&rq->nr_iowait);
5523 atomic_dec(&rq->nr_iowait);
5524 delayacct_blkio_end();
5526 EXPORT_SYMBOL(io_schedule);
5528 long __sched io_schedule_timeout(long timeout)
5530 struct rq *rq = &__raw_get_cpu_var(runqueues);
5533 delayacct_blkio_start();
5534 atomic_inc(&rq->nr_iowait);
5535 ret = schedule_timeout(timeout);
5536 atomic_dec(&rq->nr_iowait);
5537 delayacct_blkio_end();
5542 * sys_sched_get_priority_max - return maximum RT priority.
5543 * @policy: scheduling class.
5545 * this syscall returns the maximum rt_priority that can be used
5546 * by a given scheduling class.
5548 asmlinkage long sys_sched_get_priority_max(int policy)
5555 ret = MAX_USER_RT_PRIO-1;
5567 * sys_sched_get_priority_min - return minimum RT priority.
5568 * @policy: scheduling class.
5570 * this syscall returns the minimum rt_priority that can be used
5571 * by a given scheduling class.
5573 asmlinkage long sys_sched_get_priority_min(int policy)
5591 * sys_sched_rr_get_interval - return the default timeslice of a process.
5592 * @pid: pid of the process.
5593 * @interval: userspace pointer to the timeslice value.
5595 * this syscall writes the default timeslice value of a given process
5596 * into the user-space timespec buffer. A value of '0' means infinity.
5599 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5601 struct task_struct *p;
5602 unsigned int time_slice;
5610 read_lock(&tasklist_lock);
5611 p = find_process_by_pid(pid);
5615 retval = security_task_getscheduler(p);
5620 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5621 * tasks that are on an otherwise idle runqueue:
5624 if (p->policy == SCHED_RR) {
5625 time_slice = DEF_TIMESLICE;
5626 } else if (p->policy != SCHED_FIFO) {
5627 struct sched_entity *se = &p->se;
5628 unsigned long flags;
5631 rq = task_rq_lock(p, &flags);
5632 if (rq->cfs.load.weight)
5633 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5634 task_rq_unlock(rq, &flags);
5636 read_unlock(&tasklist_lock);
5637 jiffies_to_timespec(time_slice, &t);
5638 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5642 read_unlock(&tasklist_lock);
5646 static const char stat_nam[] = "RSDTtZX";
5648 void sched_show_task(struct task_struct *p)
5650 unsigned long free = 0;
5653 state = p->state ? __ffs(p->state) + 1 : 0;
5654 printk(KERN_INFO "%-13.13s %c", p->comm,
5655 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5656 #if BITS_PER_LONG == 32
5657 if (state == TASK_RUNNING)
5658 printk(KERN_CONT " running ");
5660 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5662 if (state == TASK_RUNNING)
5663 printk(KERN_CONT " running task ");
5665 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5667 #ifdef CONFIG_DEBUG_STACK_USAGE
5669 unsigned long *n = end_of_stack(p);
5672 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5675 printk(KERN_CONT "%5lu %5d %6d\n", free,
5676 task_pid_nr(p), task_pid_nr(p->real_parent));
5678 show_stack(p, NULL);
5681 void show_state_filter(unsigned long state_filter)
5683 struct task_struct *g, *p;
5685 #if BITS_PER_LONG == 32
5687 " task PC stack pid father\n");
5690 " task PC stack pid father\n");
5692 read_lock(&tasklist_lock);
5693 do_each_thread(g, p) {
5695 * reset the NMI-timeout, listing all files on a slow
5696 * console might take alot of time:
5698 touch_nmi_watchdog();
5699 if (!state_filter || (p->state & state_filter))
5701 } while_each_thread(g, p);
5703 touch_all_softlockup_watchdogs();
5705 #ifdef CONFIG_SCHED_DEBUG
5706 sysrq_sched_debug_show();
5708 read_unlock(&tasklist_lock);
5710 * Only show locks if all tasks are dumped:
5712 if (state_filter == -1)
5713 debug_show_all_locks();
5716 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5718 idle->sched_class = &idle_sched_class;
5722 * init_idle - set up an idle thread for a given CPU
5723 * @idle: task in question
5724 * @cpu: cpu the idle task belongs to
5726 * NOTE: this function does not set the idle thread's NEED_RESCHED
5727 * flag, to make booting more robust.
5729 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5731 struct rq *rq = cpu_rq(cpu);
5732 unsigned long flags;
5735 idle->se.exec_start = sched_clock();
5737 idle->prio = idle->normal_prio = MAX_PRIO;
5738 idle->cpus_allowed = cpumask_of_cpu(cpu);
5739 __set_task_cpu(idle, cpu);
5741 spin_lock_irqsave(&rq->lock, flags);
5742 rq->curr = rq->idle = idle;
5743 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5746 spin_unlock_irqrestore(&rq->lock, flags);
5748 /* Set the preempt count _outside_ the spinlocks! */
5749 #if defined(CONFIG_PREEMPT)
5750 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5752 task_thread_info(idle)->preempt_count = 0;
5755 * The idle tasks have their own, simple scheduling class:
5757 idle->sched_class = &idle_sched_class;
5761 * In a system that switches off the HZ timer nohz_cpu_mask
5762 * indicates which cpus entered this state. This is used
5763 * in the rcu update to wait only for active cpus. For system
5764 * which do not switch off the HZ timer nohz_cpu_mask should
5765 * always be CPU_MASK_NONE.
5767 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5770 * Increase the granularity value when there are more CPUs,
5771 * because with more CPUs the 'effective latency' as visible
5772 * to users decreases. But the relationship is not linear,
5773 * so pick a second-best guess by going with the log2 of the
5776 * This idea comes from the SD scheduler of Con Kolivas:
5778 static inline void sched_init_granularity(void)
5780 unsigned int factor = 1 + ilog2(num_online_cpus());
5781 const unsigned long limit = 200000000;
5783 sysctl_sched_min_granularity *= factor;
5784 if (sysctl_sched_min_granularity > limit)
5785 sysctl_sched_min_granularity = limit;
5787 sysctl_sched_latency *= factor;
5788 if (sysctl_sched_latency > limit)
5789 sysctl_sched_latency = limit;
5791 sysctl_sched_wakeup_granularity *= factor;
5796 * This is how migration works:
5798 * 1) we queue a struct migration_req structure in the source CPU's
5799 * runqueue and wake up that CPU's migration thread.
5800 * 2) we down() the locked semaphore => thread blocks.
5801 * 3) migration thread wakes up (implicitly it forces the migrated
5802 * thread off the CPU)
5803 * 4) it gets the migration request and checks whether the migrated
5804 * task is still in the wrong runqueue.
5805 * 5) if it's in the wrong runqueue then the migration thread removes
5806 * it and puts it into the right queue.
5807 * 6) migration thread up()s the semaphore.
5808 * 7) we wake up and the migration is done.
5812 * Change a given task's CPU affinity. Migrate the thread to a
5813 * proper CPU and schedule it away if the CPU it's executing on
5814 * is removed from the allowed bitmask.
5816 * NOTE: the caller must have a valid reference to the task, the
5817 * task must not exit() & deallocate itself prematurely. The
5818 * call is not atomic; no spinlocks may be held.
5820 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5822 struct migration_req req;
5823 unsigned long flags;
5827 rq = task_rq_lock(p, &flags);
5828 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5833 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5834 !cpus_equal(p->cpus_allowed, *new_mask))) {
5839 if (p->sched_class->set_cpus_allowed)
5840 p->sched_class->set_cpus_allowed(p, new_mask);
5842 p->cpus_allowed = *new_mask;
5843 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5846 /* Can the task run on the task's current CPU? If so, we're done */
5847 if (cpu_isset(task_cpu(p), *new_mask))
5850 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5851 /* Need help from migration thread: drop lock and wait. */
5852 task_rq_unlock(rq, &flags);
5853 wake_up_process(rq->migration_thread);
5854 wait_for_completion(&req.done);
5855 tlb_migrate_finish(p->mm);
5859 task_rq_unlock(rq, &flags);
5863 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5866 * Move (not current) task off this cpu, onto dest cpu. We're doing
5867 * this because either it can't run here any more (set_cpus_allowed()
5868 * away from this CPU, or CPU going down), or because we're
5869 * attempting to rebalance this task on exec (sched_exec).
5871 * So we race with normal scheduler movements, but that's OK, as long
5872 * as the task is no longer on this CPU.
5874 * Returns non-zero if task was successfully migrated.
5876 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5878 struct rq *rq_dest, *rq_src;
5881 if (unlikely(cpu_is_offline(dest_cpu)))
5884 rq_src = cpu_rq(src_cpu);
5885 rq_dest = cpu_rq(dest_cpu);
5887 double_rq_lock(rq_src, rq_dest);
5888 /* Already moved. */
5889 if (task_cpu(p) != src_cpu)
5891 /* Affinity changed (again). */
5892 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5895 on_rq = p->se.on_rq;
5897 deactivate_task(rq_src, p, 0);
5899 set_task_cpu(p, dest_cpu);
5901 activate_task(rq_dest, p, 0);
5902 check_preempt_curr(rq_dest, p);
5906 double_rq_unlock(rq_src, rq_dest);
5911 * migration_thread - this is a highprio system thread that performs
5912 * thread migration by bumping thread off CPU then 'pushing' onto
5915 static int migration_thread(void *data)
5917 int cpu = (long)data;
5921 BUG_ON(rq->migration_thread != current);
5923 set_current_state(TASK_INTERRUPTIBLE);
5924 while (!kthread_should_stop()) {
5925 struct migration_req *req;
5926 struct list_head *head;
5928 spin_lock_irq(&rq->lock);
5930 if (cpu_is_offline(cpu)) {
5931 spin_unlock_irq(&rq->lock);
5935 if (rq->active_balance) {
5936 active_load_balance(rq, cpu);
5937 rq->active_balance = 0;
5940 head = &rq->migration_queue;
5942 if (list_empty(head)) {
5943 spin_unlock_irq(&rq->lock);
5945 set_current_state(TASK_INTERRUPTIBLE);
5948 req = list_entry(head->next, struct migration_req, list);
5949 list_del_init(head->next);
5951 spin_unlock(&rq->lock);
5952 __migrate_task(req->task, cpu, req->dest_cpu);
5955 complete(&req->done);
5957 __set_current_state(TASK_RUNNING);
5961 /* Wait for kthread_stop */
5962 set_current_state(TASK_INTERRUPTIBLE);
5963 while (!kthread_should_stop()) {
5965 set_current_state(TASK_INTERRUPTIBLE);
5967 __set_current_state(TASK_RUNNING);
5971 #ifdef CONFIG_HOTPLUG_CPU
5973 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5977 local_irq_disable();
5978 ret = __migrate_task(p, src_cpu, dest_cpu);
5984 * Figure out where task on dead CPU should go, use force if necessary.
5985 * NOTE: interrupts should be disabled by the caller
5987 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5989 unsigned long flags;
5996 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5997 cpus_and(mask, mask, p->cpus_allowed);
5998 dest_cpu = any_online_cpu(mask);
6000 /* On any allowed CPU? */
6001 if (dest_cpu >= nr_cpu_ids)
6002 dest_cpu = any_online_cpu(p->cpus_allowed);
6004 /* No more Mr. Nice Guy. */
6005 if (dest_cpu >= nr_cpu_ids) {
6006 cpumask_t cpus_allowed;
6008 cpuset_cpus_allowed_locked(p, &cpus_allowed);
6010 * Try to stay on the same cpuset, where the
6011 * current cpuset may be a subset of all cpus.
6012 * The cpuset_cpus_allowed_locked() variant of
6013 * cpuset_cpus_allowed() will not block. It must be
6014 * called within calls to cpuset_lock/cpuset_unlock.
6016 rq = task_rq_lock(p, &flags);
6017 p->cpus_allowed = cpus_allowed;
6018 dest_cpu = any_online_cpu(p->cpus_allowed);
6019 task_rq_unlock(rq, &flags);
6022 * Don't tell them about moving exiting tasks or
6023 * kernel threads (both mm NULL), since they never
6026 if (p->mm && printk_ratelimit()) {
6027 printk(KERN_INFO "process %d (%s) no "
6028 "longer affine to cpu%d\n",
6029 task_pid_nr(p), p->comm, dead_cpu);
6032 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
6036 * While a dead CPU has no uninterruptible tasks queued at this point,
6037 * it might still have a nonzero ->nr_uninterruptible counter, because
6038 * for performance reasons the counter is not stricly tracking tasks to
6039 * their home CPUs. So we just add the counter to another CPU's counter,
6040 * to keep the global sum constant after CPU-down:
6042 static void migrate_nr_uninterruptible(struct rq *rq_src)
6044 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
6045 unsigned long flags;
6047 local_irq_save(flags);
6048 double_rq_lock(rq_src, rq_dest);
6049 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6050 rq_src->nr_uninterruptible = 0;
6051 double_rq_unlock(rq_src, rq_dest);
6052 local_irq_restore(flags);
6055 /* Run through task list and migrate tasks from the dead cpu. */
6056 static void migrate_live_tasks(int src_cpu)
6058 struct task_struct *p, *t;
6060 read_lock(&tasklist_lock);
6062 do_each_thread(t, p) {
6066 if (task_cpu(p) == src_cpu)
6067 move_task_off_dead_cpu(src_cpu, p);
6068 } while_each_thread(t, p);
6070 read_unlock(&tasklist_lock);
6074 * Schedules idle task to be the next runnable task on current CPU.
6075 * It does so by boosting its priority to highest possible.
6076 * Used by CPU offline code.
6078 void sched_idle_next(void)
6080 int this_cpu = smp_processor_id();
6081 struct rq *rq = cpu_rq(this_cpu);
6082 struct task_struct *p = rq->idle;
6083 unsigned long flags;
6085 /* cpu has to be offline */
6086 BUG_ON(cpu_online(this_cpu));
6089 * Strictly not necessary since rest of the CPUs are stopped by now
6090 * and interrupts disabled on the current cpu.
6092 spin_lock_irqsave(&rq->lock, flags);
6094 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6096 update_rq_clock(rq);
6097 activate_task(rq, p, 0);
6099 spin_unlock_irqrestore(&rq->lock, flags);
6103 * Ensures that the idle task is using init_mm right before its cpu goes
6106 void idle_task_exit(void)
6108 struct mm_struct *mm = current->active_mm;
6110 BUG_ON(cpu_online(smp_processor_id()));
6113 switch_mm(mm, &init_mm, current);
6117 /* called under rq->lock with disabled interrupts */
6118 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6120 struct rq *rq = cpu_rq(dead_cpu);
6122 /* Must be exiting, otherwise would be on tasklist. */
6123 BUG_ON(!p->exit_state);
6125 /* Cannot have done final schedule yet: would have vanished. */
6126 BUG_ON(p->state == TASK_DEAD);
6131 * Drop lock around migration; if someone else moves it,
6132 * that's OK. No task can be added to this CPU, so iteration is
6135 spin_unlock_irq(&rq->lock);
6136 move_task_off_dead_cpu(dead_cpu, p);
6137 spin_lock_irq(&rq->lock);
6142 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6143 static void migrate_dead_tasks(unsigned int dead_cpu)
6145 struct rq *rq = cpu_rq(dead_cpu);
6146 struct task_struct *next;
6149 if (!rq->nr_running)
6151 update_rq_clock(rq);
6152 next = pick_next_task(rq, rq->curr);
6155 migrate_dead(dead_cpu, next);
6159 #endif /* CONFIG_HOTPLUG_CPU */
6161 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6163 static struct ctl_table sd_ctl_dir[] = {
6165 .procname = "sched_domain",
6171 static struct ctl_table sd_ctl_root[] = {
6173 .ctl_name = CTL_KERN,
6174 .procname = "kernel",
6176 .child = sd_ctl_dir,
6181 static struct ctl_table *sd_alloc_ctl_entry(int n)
6183 struct ctl_table *entry =
6184 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6189 static void sd_free_ctl_entry(struct ctl_table **tablep)
6191 struct ctl_table *entry;
6194 * In the intermediate directories, both the child directory and
6195 * procname are dynamically allocated and could fail but the mode
6196 * will always be set. In the lowest directory the names are
6197 * static strings and all have proc handlers.
6199 for (entry = *tablep; entry->mode; entry++) {
6201 sd_free_ctl_entry(&entry->child);
6202 if (entry->proc_handler == NULL)
6203 kfree(entry->procname);
6211 set_table_entry(struct ctl_table *entry,
6212 const char *procname, void *data, int maxlen,
6213 mode_t mode, proc_handler *proc_handler)
6215 entry->procname = procname;
6217 entry->maxlen = maxlen;
6219 entry->proc_handler = proc_handler;
6222 static struct ctl_table *
6223 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6225 struct ctl_table *table = sd_alloc_ctl_entry(12);
6230 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6231 sizeof(long), 0644, proc_doulongvec_minmax);
6232 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6233 sizeof(long), 0644, proc_doulongvec_minmax);
6234 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6235 sizeof(int), 0644, proc_dointvec_minmax);
6236 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6237 sizeof(int), 0644, proc_dointvec_minmax);
6238 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6239 sizeof(int), 0644, proc_dointvec_minmax);
6240 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6241 sizeof(int), 0644, proc_dointvec_minmax);
6242 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6243 sizeof(int), 0644, proc_dointvec_minmax);
6244 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6245 sizeof(int), 0644, proc_dointvec_minmax);
6246 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6247 sizeof(int), 0644, proc_dointvec_minmax);
6248 set_table_entry(&table[9], "cache_nice_tries",
6249 &sd->cache_nice_tries,
6250 sizeof(int), 0644, proc_dointvec_minmax);
6251 set_table_entry(&table[10], "flags", &sd->flags,
6252 sizeof(int), 0644, proc_dointvec_minmax);
6253 /* &table[11] is terminator */
6258 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6260 struct ctl_table *entry, *table;
6261 struct sched_domain *sd;
6262 int domain_num = 0, i;
6265 for_each_domain(cpu, sd)
6267 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6272 for_each_domain(cpu, sd) {
6273 snprintf(buf, 32, "domain%d", i);
6274 entry->procname = kstrdup(buf, GFP_KERNEL);
6276 entry->child = sd_alloc_ctl_domain_table(sd);
6283 static struct ctl_table_header *sd_sysctl_header;
6284 static void register_sched_domain_sysctl(void)
6286 int i, cpu_num = num_online_cpus();
6287 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6290 WARN_ON(sd_ctl_dir[0].child);
6291 sd_ctl_dir[0].child = entry;
6296 for_each_online_cpu(i) {
6297 snprintf(buf, 32, "cpu%d", i);
6298 entry->procname = kstrdup(buf, GFP_KERNEL);
6300 entry->child = sd_alloc_ctl_cpu_table(i);
6304 WARN_ON(sd_sysctl_header);
6305 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6308 /* may be called multiple times per register */
6309 static void unregister_sched_domain_sysctl(void)
6311 if (sd_sysctl_header)
6312 unregister_sysctl_table(sd_sysctl_header);
6313 sd_sysctl_header = NULL;
6314 if (sd_ctl_dir[0].child)
6315 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6318 static void register_sched_domain_sysctl(void)
6321 static void unregister_sched_domain_sysctl(void)
6326 static void set_rq_online(struct rq *rq)
6329 const struct sched_class *class;
6331 cpu_set(rq->cpu, rq->rd->online);
6334 for_each_class(class) {
6335 if (class->rq_online)
6336 class->rq_online(rq);
6341 static void set_rq_offline(struct rq *rq)
6344 const struct sched_class *class;
6346 for_each_class(class) {
6347 if (class->rq_offline)
6348 class->rq_offline(rq);
6351 cpu_clear(rq->cpu, rq->rd->online);
6357 * migration_call - callback that gets triggered when a CPU is added.
6358 * Here we can start up the necessary migration thread for the new CPU.
6360 static int __cpuinit
6361 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6363 struct task_struct *p;
6364 int cpu = (long)hcpu;
6365 unsigned long flags;
6370 case CPU_UP_PREPARE:
6371 case CPU_UP_PREPARE_FROZEN:
6372 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6375 kthread_bind(p, cpu);
6376 /* Must be high prio: stop_machine expects to yield to it. */
6377 rq = task_rq_lock(p, &flags);
6378 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6379 task_rq_unlock(rq, &flags);
6380 cpu_rq(cpu)->migration_thread = p;
6384 case CPU_ONLINE_FROZEN:
6385 /* Strictly unnecessary, as first user will wake it. */
6386 wake_up_process(cpu_rq(cpu)->migration_thread);
6388 /* Update our root-domain */
6390 spin_lock_irqsave(&rq->lock, flags);
6392 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6396 spin_unlock_irqrestore(&rq->lock, flags);
6399 #ifdef CONFIG_HOTPLUG_CPU
6400 case CPU_UP_CANCELED:
6401 case CPU_UP_CANCELED_FROZEN:
6402 if (!cpu_rq(cpu)->migration_thread)
6404 /* Unbind it from offline cpu so it can run. Fall thru. */
6405 kthread_bind(cpu_rq(cpu)->migration_thread,
6406 any_online_cpu(cpu_online_map));
6407 kthread_stop(cpu_rq(cpu)->migration_thread);
6408 cpu_rq(cpu)->migration_thread = NULL;
6412 case CPU_DEAD_FROZEN:
6413 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6414 migrate_live_tasks(cpu);
6416 kthread_stop(rq->migration_thread);
6417 rq->migration_thread = NULL;
6418 /* Idle task back to normal (off runqueue, low prio) */
6419 spin_lock_irq(&rq->lock);
6420 update_rq_clock(rq);
6421 deactivate_task(rq, rq->idle, 0);
6422 rq->idle->static_prio = MAX_PRIO;
6423 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6424 rq->idle->sched_class = &idle_sched_class;
6425 migrate_dead_tasks(cpu);
6426 spin_unlock_irq(&rq->lock);
6428 migrate_nr_uninterruptible(rq);
6429 BUG_ON(rq->nr_running != 0);
6432 * No need to migrate the tasks: it was best-effort if
6433 * they didn't take sched_hotcpu_mutex. Just wake up
6436 spin_lock_irq(&rq->lock);
6437 while (!list_empty(&rq->migration_queue)) {
6438 struct migration_req *req;
6440 req = list_entry(rq->migration_queue.next,
6441 struct migration_req, list);
6442 list_del_init(&req->list);
6443 complete(&req->done);
6445 spin_unlock_irq(&rq->lock);
6449 case CPU_DYING_FROZEN:
6450 /* Update our root-domain */
6452 spin_lock_irqsave(&rq->lock, flags);
6454 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6457 spin_unlock_irqrestore(&rq->lock, flags);
6464 /* Register at highest priority so that task migration (migrate_all_tasks)
6465 * happens before everything else.
6467 static struct notifier_block __cpuinitdata migration_notifier = {
6468 .notifier_call = migration_call,
6472 void __init migration_init(void)
6474 void *cpu = (void *)(long)smp_processor_id();
6477 /* Start one for the boot CPU: */
6478 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6479 BUG_ON(err == NOTIFY_BAD);
6480 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6481 register_cpu_notifier(&migration_notifier);
6487 #ifdef CONFIG_SCHED_DEBUG
6489 static inline const char *sd_level_to_string(enum sched_domain_level lvl)
6502 case SD_LV_ALLNODES:
6511 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6512 cpumask_t *groupmask)
6514 struct sched_group *group = sd->groups;
6517 cpulist_scnprintf(str, sizeof(str), sd->span);
6518 cpus_clear(*groupmask);
6520 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6522 if (!(sd->flags & SD_LOAD_BALANCE)) {
6523 printk("does not load-balance\n");
6525 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6530 printk(KERN_CONT "span %s level %s\n",
6531 str, sd_level_to_string(sd->level));
6533 if (!cpu_isset(cpu, sd->span)) {
6534 printk(KERN_ERR "ERROR: domain->span does not contain "
6537 if (!cpu_isset(cpu, group->cpumask)) {
6538 printk(KERN_ERR "ERROR: domain->groups does not contain"
6542 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6546 printk(KERN_ERR "ERROR: group is NULL\n");
6550 if (!group->__cpu_power) {
6551 printk(KERN_CONT "\n");
6552 printk(KERN_ERR "ERROR: domain->cpu_power not "
6557 if (!cpus_weight(group->cpumask)) {
6558 printk(KERN_CONT "\n");
6559 printk(KERN_ERR "ERROR: empty group\n");
6563 if (cpus_intersects(*groupmask, group->cpumask)) {
6564 printk(KERN_CONT "\n");
6565 printk(KERN_ERR "ERROR: repeated CPUs\n");
6569 cpus_or(*groupmask, *groupmask, group->cpumask);
6571 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6572 printk(KERN_CONT " %s", str);
6574 group = group->next;
6575 } while (group != sd->groups);
6576 printk(KERN_CONT "\n");
6578 if (!cpus_equal(sd->span, *groupmask))
6579 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6581 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6582 printk(KERN_ERR "ERROR: parent span is not a superset "
6583 "of domain->span\n");
6587 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6589 cpumask_t *groupmask;
6593 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6597 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6599 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6601 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6606 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6615 #else /* !CONFIG_SCHED_DEBUG */
6616 # define sched_domain_debug(sd, cpu) do { } while (0)
6617 #endif /* CONFIG_SCHED_DEBUG */
6619 static int sd_degenerate(struct sched_domain *sd)
6621 if (cpus_weight(sd->span) == 1)
6624 /* Following flags need at least 2 groups */
6625 if (sd->flags & (SD_LOAD_BALANCE |
6626 SD_BALANCE_NEWIDLE |
6630 SD_SHARE_PKG_RESOURCES)) {
6631 if (sd->groups != sd->groups->next)
6635 /* Following flags don't use groups */
6636 if (sd->flags & (SD_WAKE_IDLE |
6645 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6647 unsigned long cflags = sd->flags, pflags = parent->flags;
6649 if (sd_degenerate(parent))
6652 if (!cpus_equal(sd->span, parent->span))
6655 /* Does parent contain flags not in child? */
6656 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6657 if (cflags & SD_WAKE_AFFINE)
6658 pflags &= ~SD_WAKE_BALANCE;
6659 /* Flags needing groups don't count if only 1 group in parent */
6660 if (parent->groups == parent->groups->next) {
6661 pflags &= ~(SD_LOAD_BALANCE |
6662 SD_BALANCE_NEWIDLE |
6666 SD_SHARE_PKG_RESOURCES);
6668 if (~cflags & pflags)
6674 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6676 unsigned long flags;
6678 spin_lock_irqsave(&rq->lock, flags);
6681 struct root_domain *old_rd = rq->rd;
6683 if (cpu_isset(rq->cpu, old_rd->online))
6686 cpu_clear(rq->cpu, old_rd->span);
6688 if (atomic_dec_and_test(&old_rd->refcount))
6692 atomic_inc(&rd->refcount);
6695 cpu_set(rq->cpu, rd->span);
6696 if (cpu_isset(rq->cpu, cpu_online_map))
6699 spin_unlock_irqrestore(&rq->lock, flags);
6702 static void init_rootdomain(struct root_domain *rd)
6704 memset(rd, 0, sizeof(*rd));
6706 cpus_clear(rd->span);
6707 cpus_clear(rd->online);
6709 cpupri_init(&rd->cpupri);
6712 static void init_defrootdomain(void)
6714 init_rootdomain(&def_root_domain);
6715 atomic_set(&def_root_domain.refcount, 1);
6718 static struct root_domain *alloc_rootdomain(void)
6720 struct root_domain *rd;
6722 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6726 init_rootdomain(rd);
6732 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6733 * hold the hotplug lock.
6736 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6738 struct rq *rq = cpu_rq(cpu);
6739 struct sched_domain *tmp;
6741 /* Remove the sched domains which do not contribute to scheduling. */
6742 for (tmp = sd; tmp; tmp = tmp->parent) {
6743 struct sched_domain *parent = tmp->parent;
6746 if (sd_parent_degenerate(tmp, parent)) {
6747 tmp->parent = parent->parent;
6749 parent->parent->child = tmp;
6753 if (sd && sd_degenerate(sd)) {
6759 sched_domain_debug(sd, cpu);
6761 rq_attach_root(rq, rd);
6762 rcu_assign_pointer(rq->sd, sd);
6765 /* cpus with isolated domains */
6766 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6768 /* Setup the mask of cpus configured for isolated domains */
6769 static int __init isolated_cpu_setup(char *str)
6771 int ints[NR_CPUS], i;
6773 str = get_options(str, ARRAY_SIZE(ints), ints);
6774 cpus_clear(cpu_isolated_map);
6775 for (i = 1; i <= ints[0]; i++)
6776 if (ints[i] < NR_CPUS)
6777 cpu_set(ints[i], cpu_isolated_map);
6781 __setup("isolcpus=", isolated_cpu_setup);
6784 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6785 * to a function which identifies what group(along with sched group) a CPU
6786 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6787 * (due to the fact that we keep track of groups covered with a cpumask_t).
6789 * init_sched_build_groups will build a circular linked list of the groups
6790 * covered by the given span, and will set each group's ->cpumask correctly,
6791 * and ->cpu_power to 0.
6794 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6795 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6796 struct sched_group **sg,
6797 cpumask_t *tmpmask),
6798 cpumask_t *covered, cpumask_t *tmpmask)
6800 struct sched_group *first = NULL, *last = NULL;
6803 cpus_clear(*covered);
6805 for_each_cpu_mask(i, *span) {
6806 struct sched_group *sg;
6807 int group = group_fn(i, cpu_map, &sg, tmpmask);
6810 if (cpu_isset(i, *covered))
6813 cpus_clear(sg->cpumask);
6814 sg->__cpu_power = 0;
6816 for_each_cpu_mask(j, *span) {
6817 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6820 cpu_set(j, *covered);
6821 cpu_set(j, sg->cpumask);
6832 #define SD_NODES_PER_DOMAIN 16
6837 * find_next_best_node - find the next node to include in a sched_domain
6838 * @node: node whose sched_domain we're building
6839 * @used_nodes: nodes already in the sched_domain
6841 * Find the next node to include in a given scheduling domain. Simply
6842 * finds the closest node not already in the @used_nodes map.
6844 * Should use nodemask_t.
6846 static int find_next_best_node(int node, nodemask_t *used_nodes)
6848 int i, n, val, min_val, best_node = 0;
6852 for (i = 0; i < MAX_NUMNODES; i++) {
6853 /* Start at @node */
6854 n = (node + i) % MAX_NUMNODES;
6856 if (!nr_cpus_node(n))
6859 /* Skip already used nodes */
6860 if (node_isset(n, *used_nodes))
6863 /* Simple min distance search */
6864 val = node_distance(node, n);
6866 if (val < min_val) {
6872 node_set(best_node, *used_nodes);
6877 * sched_domain_node_span - get a cpumask for a node's sched_domain
6878 * @node: node whose cpumask we're constructing
6879 * @span: resulting cpumask
6881 * Given a node, construct a good cpumask for its sched_domain to span. It
6882 * should be one that prevents unnecessary balancing, but also spreads tasks
6885 static void sched_domain_node_span(int node, cpumask_t *span)
6887 nodemask_t used_nodes;
6888 node_to_cpumask_ptr(nodemask, node);
6892 nodes_clear(used_nodes);
6894 cpus_or(*span, *span, *nodemask);
6895 node_set(node, used_nodes);
6897 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6898 int next_node = find_next_best_node(node, &used_nodes);
6900 node_to_cpumask_ptr_next(nodemask, next_node);
6901 cpus_or(*span, *span, *nodemask);
6904 #endif /* CONFIG_NUMA */
6906 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6909 * SMT sched-domains:
6911 #ifdef CONFIG_SCHED_SMT
6912 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6913 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6916 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6920 *sg = &per_cpu(sched_group_cpus, cpu);
6923 #endif /* CONFIG_SCHED_SMT */
6926 * multi-core sched-domains:
6928 #ifdef CONFIG_SCHED_MC
6929 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6930 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6931 #endif /* CONFIG_SCHED_MC */
6933 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6935 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6940 *mask = per_cpu(cpu_sibling_map, cpu);
6941 cpus_and(*mask, *mask, *cpu_map);
6942 group = first_cpu(*mask);
6944 *sg = &per_cpu(sched_group_core, group);
6947 #elif defined(CONFIG_SCHED_MC)
6949 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6953 *sg = &per_cpu(sched_group_core, cpu);
6958 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6959 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6962 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6966 #ifdef CONFIG_SCHED_MC
6967 *mask = cpu_coregroup_map(cpu);
6968 cpus_and(*mask, *mask, *cpu_map);
6969 group = first_cpu(*mask);
6970 #elif defined(CONFIG_SCHED_SMT)
6971 *mask = per_cpu(cpu_sibling_map, cpu);
6972 cpus_and(*mask, *mask, *cpu_map);
6973 group = first_cpu(*mask);
6978 *sg = &per_cpu(sched_group_phys, group);
6984 * The init_sched_build_groups can't handle what we want to do with node
6985 * groups, so roll our own. Now each node has its own list of groups which
6986 * gets dynamically allocated.
6988 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6989 static struct sched_group ***sched_group_nodes_bycpu;
6991 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6992 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6994 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6995 struct sched_group **sg, cpumask_t *nodemask)
6999 *nodemask = node_to_cpumask(cpu_to_node(cpu));
7000 cpus_and(*nodemask, *nodemask, *cpu_map);
7001 group = first_cpu(*nodemask);
7004 *sg = &per_cpu(sched_group_allnodes, group);
7008 static void init_numa_sched_groups_power(struct sched_group *group_head)
7010 struct sched_group *sg = group_head;
7016 for_each_cpu_mask(j, sg->cpumask) {
7017 struct sched_domain *sd;
7019 sd = &per_cpu(phys_domains, j);
7020 if (j != first_cpu(sd->groups->cpumask)) {
7022 * Only add "power" once for each
7028 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7031 } while (sg != group_head);
7033 #endif /* CONFIG_NUMA */
7036 /* Free memory allocated for various sched_group structures */
7037 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7041 for_each_cpu_mask(cpu, *cpu_map) {
7042 struct sched_group **sched_group_nodes
7043 = sched_group_nodes_bycpu[cpu];
7045 if (!sched_group_nodes)
7048 for (i = 0; i < MAX_NUMNODES; i++) {
7049 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7051 *nodemask = node_to_cpumask(i);
7052 cpus_and(*nodemask, *nodemask, *cpu_map);
7053 if (cpus_empty(*nodemask))
7063 if (oldsg != sched_group_nodes[i])
7066 kfree(sched_group_nodes);
7067 sched_group_nodes_bycpu[cpu] = NULL;
7070 #else /* !CONFIG_NUMA */
7071 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7074 #endif /* CONFIG_NUMA */
7077 * Initialize sched groups cpu_power.
7079 * cpu_power indicates the capacity of sched group, which is used while
7080 * distributing the load between different sched groups in a sched domain.
7081 * Typically cpu_power for all the groups in a sched domain will be same unless
7082 * there are asymmetries in the topology. If there are asymmetries, group
7083 * having more cpu_power will pickup more load compared to the group having
7086 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7087 * the maximum number of tasks a group can handle in the presence of other idle
7088 * or lightly loaded groups in the same sched domain.
7090 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7092 struct sched_domain *child;
7093 struct sched_group *group;
7095 WARN_ON(!sd || !sd->groups);
7097 if (cpu != first_cpu(sd->groups->cpumask))
7102 sd->groups->__cpu_power = 0;
7105 * For perf policy, if the groups in child domain share resources
7106 * (for example cores sharing some portions of the cache hierarchy
7107 * or SMT), then set this domain groups cpu_power such that each group
7108 * can handle only one task, when there are other idle groups in the
7109 * same sched domain.
7111 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7113 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7114 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7119 * add cpu_power of each child group to this groups cpu_power
7121 group = child->groups;
7123 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7124 group = group->next;
7125 } while (group != child->groups);
7129 * Initializers for schedule domains
7130 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7133 #define SD_INIT(sd, type) sd_init_##type(sd)
7134 #define SD_INIT_FUNC(type) \
7135 static noinline void sd_init_##type(struct sched_domain *sd) \
7137 memset(sd, 0, sizeof(*sd)); \
7138 *sd = SD_##type##_INIT; \
7139 sd->level = SD_LV_##type; \
7144 SD_INIT_FUNC(ALLNODES)
7147 #ifdef CONFIG_SCHED_SMT
7148 SD_INIT_FUNC(SIBLING)
7150 #ifdef CONFIG_SCHED_MC
7155 * To minimize stack usage kmalloc room for cpumasks and share the
7156 * space as the usage in build_sched_domains() dictates. Used only
7157 * if the amount of space is significant.
7160 cpumask_t tmpmask; /* make this one first */
7163 cpumask_t this_sibling_map;
7164 cpumask_t this_core_map;
7166 cpumask_t send_covered;
7169 cpumask_t domainspan;
7171 cpumask_t notcovered;
7176 #define SCHED_CPUMASK_ALLOC 1
7177 #define SCHED_CPUMASK_FREE(v) kfree(v)
7178 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7180 #define SCHED_CPUMASK_ALLOC 0
7181 #define SCHED_CPUMASK_FREE(v)
7182 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7185 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7186 ((unsigned long)(a) + offsetof(struct allmasks, v))
7188 static int default_relax_domain_level = -1;
7190 static int __init setup_relax_domain_level(char *str)
7194 val = simple_strtoul(str, NULL, 0);
7195 if (val < SD_LV_MAX)
7196 default_relax_domain_level = val;
7200 __setup("relax_domain_level=", setup_relax_domain_level);
7202 static void set_domain_attribute(struct sched_domain *sd,
7203 struct sched_domain_attr *attr)
7207 if (!attr || attr->relax_domain_level < 0) {
7208 if (default_relax_domain_level < 0)
7211 request = default_relax_domain_level;
7213 request = attr->relax_domain_level;
7214 if (request < sd->level) {
7215 /* turn off idle balance on this domain */
7216 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7218 /* turn on idle balance on this domain */
7219 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7224 * Build sched domains for a given set of cpus and attach the sched domains
7225 * to the individual cpus
7227 static int __build_sched_domains(const cpumask_t *cpu_map,
7228 struct sched_domain_attr *attr)
7231 struct root_domain *rd;
7232 SCHED_CPUMASK_DECLARE(allmasks);
7235 struct sched_group **sched_group_nodes = NULL;
7236 int sd_allnodes = 0;
7239 * Allocate the per-node list of sched groups
7241 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
7243 if (!sched_group_nodes) {
7244 printk(KERN_WARNING "Can not alloc sched group node list\n");
7249 rd = alloc_rootdomain();
7251 printk(KERN_WARNING "Cannot alloc root domain\n");
7253 kfree(sched_group_nodes);
7258 #if SCHED_CPUMASK_ALLOC
7259 /* get space for all scratch cpumask variables */
7260 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
7262 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7265 kfree(sched_group_nodes);
7270 tmpmask = (cpumask_t *)allmasks;
7274 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7278 * Set up domains for cpus specified by the cpu_map.
7280 for_each_cpu_mask(i, *cpu_map) {
7281 struct sched_domain *sd = NULL, *p;
7282 SCHED_CPUMASK_VAR(nodemask, allmasks);
7284 *nodemask = node_to_cpumask(cpu_to_node(i));
7285 cpus_and(*nodemask, *nodemask, *cpu_map);
7288 if (cpus_weight(*cpu_map) >
7289 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7290 sd = &per_cpu(allnodes_domains, i);
7291 SD_INIT(sd, ALLNODES);
7292 set_domain_attribute(sd, attr);
7293 sd->span = *cpu_map;
7294 sd->first_cpu = first_cpu(sd->span);
7295 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7301 sd = &per_cpu(node_domains, i);
7303 set_domain_attribute(sd, attr);
7304 sched_domain_node_span(cpu_to_node(i), &sd->span);
7305 sd->first_cpu = first_cpu(sd->span);
7309 cpus_and(sd->span, sd->span, *cpu_map);
7313 sd = &per_cpu(phys_domains, i);
7315 set_domain_attribute(sd, attr);
7316 sd->span = *nodemask;
7317 sd->first_cpu = first_cpu(sd->span);
7321 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7323 #ifdef CONFIG_SCHED_MC
7325 sd = &per_cpu(core_domains, i);
7327 set_domain_attribute(sd, attr);
7328 sd->span = cpu_coregroup_map(i);
7329 sd->first_cpu = first_cpu(sd->span);
7330 cpus_and(sd->span, sd->span, *cpu_map);
7333 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7336 #ifdef CONFIG_SCHED_SMT
7338 sd = &per_cpu(cpu_domains, i);
7339 SD_INIT(sd, SIBLING);
7340 set_domain_attribute(sd, attr);
7341 sd->span = per_cpu(cpu_sibling_map, i);
7342 sd->first_cpu = first_cpu(sd->span);
7343 cpus_and(sd->span, sd->span, *cpu_map);
7346 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7350 #ifdef CONFIG_SCHED_SMT
7351 /* Set up CPU (sibling) groups */
7352 for_each_cpu_mask(i, *cpu_map) {
7353 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7354 SCHED_CPUMASK_VAR(send_covered, allmasks);
7356 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7357 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7358 if (i != first_cpu(*this_sibling_map))
7361 init_sched_build_groups(this_sibling_map, cpu_map,
7363 send_covered, tmpmask);
7367 #ifdef CONFIG_SCHED_MC
7368 /* Set up multi-core groups */
7369 for_each_cpu_mask(i, *cpu_map) {
7370 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7371 SCHED_CPUMASK_VAR(send_covered, allmasks);
7373 *this_core_map = cpu_coregroup_map(i);
7374 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7375 if (i != first_cpu(*this_core_map))
7378 init_sched_build_groups(this_core_map, cpu_map,
7380 send_covered, tmpmask);
7384 /* Set up physical groups */
7385 for (i = 0; i < MAX_NUMNODES; i++) {
7386 SCHED_CPUMASK_VAR(nodemask, allmasks);
7387 SCHED_CPUMASK_VAR(send_covered, allmasks);
7389 *nodemask = node_to_cpumask(i);
7390 cpus_and(*nodemask, *nodemask, *cpu_map);
7391 if (cpus_empty(*nodemask))
7394 init_sched_build_groups(nodemask, cpu_map,
7396 send_covered, tmpmask);
7400 /* Set up node groups */
7402 SCHED_CPUMASK_VAR(send_covered, allmasks);
7404 init_sched_build_groups(cpu_map, cpu_map,
7405 &cpu_to_allnodes_group,
7406 send_covered, tmpmask);
7409 for (i = 0; i < MAX_NUMNODES; i++) {
7410 /* Set up node groups */
7411 struct sched_group *sg, *prev;
7412 SCHED_CPUMASK_VAR(nodemask, allmasks);
7413 SCHED_CPUMASK_VAR(domainspan, allmasks);
7414 SCHED_CPUMASK_VAR(covered, allmasks);
7417 *nodemask = node_to_cpumask(i);
7418 cpus_clear(*covered);
7420 cpus_and(*nodemask, *nodemask, *cpu_map);
7421 if (cpus_empty(*nodemask)) {
7422 sched_group_nodes[i] = NULL;
7426 sched_domain_node_span(i, domainspan);
7427 cpus_and(*domainspan, *domainspan, *cpu_map);
7429 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7431 printk(KERN_WARNING "Can not alloc domain group for "
7435 sched_group_nodes[i] = sg;
7436 for_each_cpu_mask(j, *nodemask) {
7437 struct sched_domain *sd;
7439 sd = &per_cpu(node_domains, j);
7442 sg->__cpu_power = 0;
7443 sg->cpumask = *nodemask;
7445 cpus_or(*covered, *covered, *nodemask);
7448 for (j = 0; j < MAX_NUMNODES; j++) {
7449 SCHED_CPUMASK_VAR(notcovered, allmasks);
7450 int n = (i + j) % MAX_NUMNODES;
7451 node_to_cpumask_ptr(pnodemask, n);
7453 cpus_complement(*notcovered, *covered);
7454 cpus_and(*tmpmask, *notcovered, *cpu_map);
7455 cpus_and(*tmpmask, *tmpmask, *domainspan);
7456 if (cpus_empty(*tmpmask))
7459 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7460 if (cpus_empty(*tmpmask))
7463 sg = kmalloc_node(sizeof(struct sched_group),
7467 "Can not alloc domain group for node %d\n", j);
7470 sg->__cpu_power = 0;
7471 sg->cpumask = *tmpmask;
7472 sg->next = prev->next;
7473 cpus_or(*covered, *covered, *tmpmask);
7480 /* Calculate CPU power for physical packages and nodes */
7481 #ifdef CONFIG_SCHED_SMT
7482 for_each_cpu_mask(i, *cpu_map) {
7483 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7485 init_sched_groups_power(i, sd);
7488 #ifdef CONFIG_SCHED_MC
7489 for_each_cpu_mask(i, *cpu_map) {
7490 struct sched_domain *sd = &per_cpu(core_domains, i);
7492 init_sched_groups_power(i, sd);
7496 for_each_cpu_mask(i, *cpu_map) {
7497 struct sched_domain *sd = &per_cpu(phys_domains, i);
7499 init_sched_groups_power(i, sd);
7503 for (i = 0; i < MAX_NUMNODES; i++)
7504 init_numa_sched_groups_power(sched_group_nodes[i]);
7507 struct sched_group *sg;
7509 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7511 init_numa_sched_groups_power(sg);
7515 /* Attach the domains */
7516 for_each_cpu_mask(i, *cpu_map) {
7517 struct sched_domain *sd;
7518 #ifdef CONFIG_SCHED_SMT
7519 sd = &per_cpu(cpu_domains, i);
7520 #elif defined(CONFIG_SCHED_MC)
7521 sd = &per_cpu(core_domains, i);
7523 sd = &per_cpu(phys_domains, i);
7525 cpu_attach_domain(sd, rd, i);
7528 SCHED_CPUMASK_FREE((void *)allmasks);
7533 free_sched_groups(cpu_map, tmpmask);
7534 SCHED_CPUMASK_FREE((void *)allmasks);
7539 static int build_sched_domains(const cpumask_t *cpu_map)
7541 return __build_sched_domains(cpu_map, NULL);
7544 static cpumask_t *doms_cur; /* current sched domains */
7545 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7546 static struct sched_domain_attr *dattr_cur;
7547 /* attribues of custom domains in 'doms_cur' */
7550 * Special case: If a kmalloc of a doms_cur partition (array of
7551 * cpumask_t) fails, then fallback to a single sched domain,
7552 * as determined by the single cpumask_t fallback_doms.
7554 static cpumask_t fallback_doms;
7556 void __attribute__((weak)) arch_update_cpu_topology(void)
7561 * Free current domain masks.
7562 * Called after all cpus are attached to NULL domain.
7564 static void free_sched_domains(void)
7567 if (doms_cur != &fallback_doms)
7569 doms_cur = &fallback_doms;
7573 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7574 * For now this just excludes isolated cpus, but could be used to
7575 * exclude other special cases in the future.
7577 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7581 arch_update_cpu_topology();
7583 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7585 doms_cur = &fallback_doms;
7586 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7588 err = build_sched_domains(doms_cur);
7589 register_sched_domain_sysctl();
7594 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7597 free_sched_groups(cpu_map, tmpmask);
7601 * Detach sched domains from a group of cpus specified in cpu_map
7602 * These cpus will now be attached to the NULL domain
7604 static void detach_destroy_domains(const cpumask_t *cpu_map)
7609 unregister_sched_domain_sysctl();
7611 for_each_cpu_mask(i, *cpu_map)
7612 cpu_attach_domain(NULL, &def_root_domain, i);
7613 synchronize_sched();
7614 arch_destroy_sched_domains(cpu_map, &tmpmask);
7617 /* handle null as "default" */
7618 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7619 struct sched_domain_attr *new, int idx_new)
7621 struct sched_domain_attr tmp;
7628 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7629 new ? (new + idx_new) : &tmp,
7630 sizeof(struct sched_domain_attr));
7634 * Partition sched domains as specified by the 'ndoms_new'
7635 * cpumasks in the array doms_new[] of cpumasks. This compares
7636 * doms_new[] to the current sched domain partitioning, doms_cur[].
7637 * It destroys each deleted domain and builds each new domain.
7639 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7640 * The masks don't intersect (don't overlap.) We should setup one
7641 * sched domain for each mask. CPUs not in any of the cpumasks will
7642 * not be load balanced. If the same cpumask appears both in the
7643 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7646 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7647 * ownership of it and will kfree it when done with it. If the caller
7648 * failed the kmalloc call, then it can pass in doms_new == NULL,
7649 * and partition_sched_domains() will fallback to the single partition
7652 * Call with hotplug lock held
7654 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7655 struct sched_domain_attr *dattr_new)
7659 mutex_lock(&sched_domains_mutex);
7661 /* always unregister in case we don't destroy any domains */
7662 unregister_sched_domain_sysctl();
7664 if (doms_new == NULL) {
7666 doms_new = &fallback_doms;
7667 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7671 /* Destroy deleted domains */
7672 for (i = 0; i < ndoms_cur; i++) {
7673 for (j = 0; j < ndoms_new; j++) {
7674 if (cpus_equal(doms_cur[i], doms_new[j])
7675 && dattrs_equal(dattr_cur, i, dattr_new, j))
7678 /* no match - a current sched domain not in new doms_new[] */
7679 detach_destroy_domains(doms_cur + i);
7684 /* Build new domains */
7685 for (i = 0; i < ndoms_new; i++) {
7686 for (j = 0; j < ndoms_cur; j++) {
7687 if (cpus_equal(doms_new[i], doms_cur[j])
7688 && dattrs_equal(dattr_new, i, dattr_cur, j))
7691 /* no match - add a new doms_new */
7692 __build_sched_domains(doms_new + i,
7693 dattr_new ? dattr_new + i : NULL);
7698 /* Remember the new sched domains */
7699 if (doms_cur != &fallback_doms)
7701 kfree(dattr_cur); /* kfree(NULL) is safe */
7702 doms_cur = doms_new;
7703 dattr_cur = dattr_new;
7704 ndoms_cur = ndoms_new;
7706 register_sched_domain_sysctl();
7708 mutex_unlock(&sched_domains_mutex);
7711 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7712 int arch_reinit_sched_domains(void)
7717 mutex_lock(&sched_domains_mutex);
7718 detach_destroy_domains(&cpu_online_map);
7719 free_sched_domains();
7720 err = arch_init_sched_domains(&cpu_online_map);
7721 mutex_unlock(&sched_domains_mutex);
7727 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7731 if (buf[0] != '0' && buf[0] != '1')
7735 sched_smt_power_savings = (buf[0] == '1');
7737 sched_mc_power_savings = (buf[0] == '1');
7739 ret = arch_reinit_sched_domains();
7741 return ret ? ret : count;
7744 #ifdef CONFIG_SCHED_MC
7745 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7747 return sprintf(page, "%u\n", sched_mc_power_savings);
7749 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7750 const char *buf, size_t count)
7752 return sched_power_savings_store(buf, count, 0);
7754 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7755 sched_mc_power_savings_store);
7758 #ifdef CONFIG_SCHED_SMT
7759 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7761 return sprintf(page, "%u\n", sched_smt_power_savings);
7763 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7764 const char *buf, size_t count)
7766 return sched_power_savings_store(buf, count, 1);
7768 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7769 sched_smt_power_savings_store);
7772 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7776 #ifdef CONFIG_SCHED_SMT
7778 err = sysfs_create_file(&cls->kset.kobj,
7779 &attr_sched_smt_power_savings.attr);
7781 #ifdef CONFIG_SCHED_MC
7782 if (!err && mc_capable())
7783 err = sysfs_create_file(&cls->kset.kobj,
7784 &attr_sched_mc_power_savings.attr);
7788 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7791 * Force a reinitialization of the sched domains hierarchy. The domains
7792 * and groups cannot be updated in place without racing with the balancing
7793 * code, so we temporarily attach all running cpus to the NULL domain
7794 * which will prevent rebalancing while the sched domains are recalculated.
7796 static int update_sched_domains(struct notifier_block *nfb,
7797 unsigned long action, void *hcpu)
7799 int cpu = (int)(long)hcpu;
7802 case CPU_DOWN_PREPARE:
7803 case CPU_DOWN_PREPARE_FROZEN:
7804 disable_runtime(cpu_rq(cpu));
7806 case CPU_UP_PREPARE:
7807 case CPU_UP_PREPARE_FROZEN:
7808 detach_destroy_domains(&cpu_online_map);
7809 free_sched_domains();
7813 case CPU_DOWN_FAILED:
7814 case CPU_DOWN_FAILED_FROZEN:
7816 case CPU_ONLINE_FROZEN:
7817 enable_runtime(cpu_rq(cpu));
7819 case CPU_UP_CANCELED:
7820 case CPU_UP_CANCELED_FROZEN:
7822 case CPU_DEAD_FROZEN:
7824 * Fall through and re-initialise the domains.
7831 #ifndef CONFIG_CPUSETS
7833 * Create default domain partitioning if cpusets are disabled.
7834 * Otherwise we let cpusets rebuild the domains based on the
7838 /* The hotplug lock is already held by cpu_up/cpu_down */
7839 arch_init_sched_domains(&cpu_online_map);
7845 void __init sched_init_smp(void)
7847 cpumask_t non_isolated_cpus;
7849 #if defined(CONFIG_NUMA)
7850 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7852 BUG_ON(sched_group_nodes_bycpu == NULL);
7855 mutex_lock(&sched_domains_mutex);
7856 arch_init_sched_domains(&cpu_online_map);
7857 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7858 if (cpus_empty(non_isolated_cpus))
7859 cpu_set(smp_processor_id(), non_isolated_cpus);
7860 mutex_unlock(&sched_domains_mutex);
7862 /* XXX: Theoretical race here - CPU may be hotplugged now */
7863 hotcpu_notifier(update_sched_domains, 0);
7866 /* Move init over to a non-isolated CPU */
7867 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7869 sched_init_granularity();
7872 void __init sched_init_smp(void)
7874 sched_init_granularity();
7876 #endif /* CONFIG_SMP */
7878 int in_sched_functions(unsigned long addr)
7880 return in_lock_functions(addr) ||
7881 (addr >= (unsigned long)__sched_text_start
7882 && addr < (unsigned long)__sched_text_end);
7885 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7887 cfs_rq->tasks_timeline = RB_ROOT;
7888 INIT_LIST_HEAD(&cfs_rq->tasks);
7889 #ifdef CONFIG_FAIR_GROUP_SCHED
7892 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7895 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7897 struct rt_prio_array *array;
7900 array = &rt_rq->active;
7901 for (i = 0; i < MAX_RT_PRIO; i++) {
7902 INIT_LIST_HEAD(array->queue + i);
7903 __clear_bit(i, array->bitmap);
7905 /* delimiter for bitsearch: */
7906 __set_bit(MAX_RT_PRIO, array->bitmap);
7908 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7909 rt_rq->highest_prio = MAX_RT_PRIO;
7912 rt_rq->rt_nr_migratory = 0;
7913 rt_rq->overloaded = 0;
7917 rt_rq->rt_throttled = 0;
7918 rt_rq->rt_runtime = 0;
7919 spin_lock_init(&rt_rq->rt_runtime_lock);
7921 #ifdef CONFIG_RT_GROUP_SCHED
7922 rt_rq->rt_nr_boosted = 0;
7927 #ifdef CONFIG_FAIR_GROUP_SCHED
7928 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7929 struct sched_entity *se, int cpu, int add,
7930 struct sched_entity *parent)
7932 struct rq *rq = cpu_rq(cpu);
7933 tg->cfs_rq[cpu] = cfs_rq;
7934 init_cfs_rq(cfs_rq, rq);
7937 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7940 /* se could be NULL for init_task_group */
7945 se->cfs_rq = &rq->cfs;
7947 se->cfs_rq = parent->my_q;
7950 se->load.weight = tg->shares;
7951 se->load.inv_weight = 0;
7952 se->parent = parent;
7956 #ifdef CONFIG_RT_GROUP_SCHED
7957 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7958 struct sched_rt_entity *rt_se, int cpu, int add,
7959 struct sched_rt_entity *parent)
7961 struct rq *rq = cpu_rq(cpu);
7963 tg->rt_rq[cpu] = rt_rq;
7964 init_rt_rq(rt_rq, rq);
7966 rt_rq->rt_se = rt_se;
7967 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7969 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7971 tg->rt_se[cpu] = rt_se;
7976 rt_se->rt_rq = &rq->rt;
7978 rt_se->rt_rq = parent->my_q;
7980 rt_se->my_q = rt_rq;
7981 rt_se->parent = parent;
7982 INIT_LIST_HEAD(&rt_se->run_list);
7986 void __init sched_init(void)
7989 unsigned long alloc_size = 0, ptr;
7991 #ifdef CONFIG_FAIR_GROUP_SCHED
7992 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7994 #ifdef CONFIG_RT_GROUP_SCHED
7995 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7997 #ifdef CONFIG_USER_SCHED
8001 * As sched_init() is called before page_alloc is setup,
8002 * we use alloc_bootmem().
8005 ptr = (unsigned long)alloc_bootmem(alloc_size);
8007 #ifdef CONFIG_FAIR_GROUP_SCHED
8008 init_task_group.se = (struct sched_entity **)ptr;
8009 ptr += nr_cpu_ids * sizeof(void **);
8011 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8012 ptr += nr_cpu_ids * sizeof(void **);
8014 #ifdef CONFIG_USER_SCHED
8015 root_task_group.se = (struct sched_entity **)ptr;
8016 ptr += nr_cpu_ids * sizeof(void **);
8018 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8019 ptr += nr_cpu_ids * sizeof(void **);
8020 #endif /* CONFIG_USER_SCHED */
8021 #endif /* CONFIG_FAIR_GROUP_SCHED */
8022 #ifdef CONFIG_RT_GROUP_SCHED
8023 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8024 ptr += nr_cpu_ids * sizeof(void **);
8026 init_task_group.rt_rq = (struct rt_rq **)ptr;
8027 ptr += nr_cpu_ids * sizeof(void **);
8029 #ifdef CONFIG_USER_SCHED
8030 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8031 ptr += nr_cpu_ids * sizeof(void **);
8033 root_task_group.rt_rq = (struct rt_rq **)ptr;
8034 ptr += nr_cpu_ids * sizeof(void **);
8035 #endif /* CONFIG_USER_SCHED */
8036 #endif /* CONFIG_RT_GROUP_SCHED */
8041 init_defrootdomain();
8044 init_rt_bandwidth(&def_rt_bandwidth,
8045 global_rt_period(), global_rt_runtime());
8047 #ifdef CONFIG_RT_GROUP_SCHED
8048 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8049 global_rt_period(), global_rt_runtime());
8050 #ifdef CONFIG_USER_SCHED
8051 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8052 global_rt_period(), RUNTIME_INF);
8053 #endif /* CONFIG_USER_SCHED */
8054 #endif /* CONFIG_RT_GROUP_SCHED */
8056 #ifdef CONFIG_GROUP_SCHED
8057 list_add(&init_task_group.list, &task_groups);
8058 INIT_LIST_HEAD(&init_task_group.children);
8060 #ifdef CONFIG_USER_SCHED
8061 INIT_LIST_HEAD(&root_task_group.children);
8062 init_task_group.parent = &root_task_group;
8063 list_add(&init_task_group.siblings, &root_task_group.children);
8064 #endif /* CONFIG_USER_SCHED */
8065 #endif /* CONFIG_GROUP_SCHED */
8067 for_each_possible_cpu(i) {
8071 spin_lock_init(&rq->lock);
8072 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
8074 init_cfs_rq(&rq->cfs, rq);
8075 init_rt_rq(&rq->rt, rq);
8076 #ifdef CONFIG_FAIR_GROUP_SCHED
8077 init_task_group.shares = init_task_group_load;
8078 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8079 #ifdef CONFIG_CGROUP_SCHED
8081 * How much cpu bandwidth does init_task_group get?
8083 * In case of task-groups formed thr' the cgroup filesystem, it
8084 * gets 100% of the cpu resources in the system. This overall
8085 * system cpu resource is divided among the tasks of
8086 * init_task_group and its child task-groups in a fair manner,
8087 * based on each entity's (task or task-group's) weight
8088 * (se->load.weight).
8090 * In other words, if init_task_group has 10 tasks of weight
8091 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8092 * then A0's share of the cpu resource is:
8094 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8096 * We achieve this by letting init_task_group's tasks sit
8097 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8099 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8100 #elif defined CONFIG_USER_SCHED
8101 root_task_group.shares = NICE_0_LOAD;
8102 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8104 * In case of task-groups formed thr' the user id of tasks,
8105 * init_task_group represents tasks belonging to root user.
8106 * Hence it forms a sibling of all subsequent groups formed.
8107 * In this case, init_task_group gets only a fraction of overall
8108 * system cpu resource, based on the weight assigned to root
8109 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8110 * by letting tasks of init_task_group sit in a separate cfs_rq
8111 * (init_cfs_rq) and having one entity represent this group of
8112 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8114 init_tg_cfs_entry(&init_task_group,
8115 &per_cpu(init_cfs_rq, i),
8116 &per_cpu(init_sched_entity, i), i, 1,
8117 root_task_group.se[i]);
8120 #endif /* CONFIG_FAIR_GROUP_SCHED */
8122 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8123 #ifdef CONFIG_RT_GROUP_SCHED
8124 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8125 #ifdef CONFIG_CGROUP_SCHED
8126 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8127 #elif defined CONFIG_USER_SCHED
8128 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8129 init_tg_rt_entry(&init_task_group,
8130 &per_cpu(init_rt_rq, i),
8131 &per_cpu(init_sched_rt_entity, i), i, 1,
8132 root_task_group.rt_se[i]);
8136 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8137 rq->cpu_load[j] = 0;
8141 rq->active_balance = 0;
8142 rq->next_balance = jiffies;
8146 rq->migration_thread = NULL;
8147 INIT_LIST_HEAD(&rq->migration_queue);
8148 rq_attach_root(rq, &def_root_domain);
8151 atomic_set(&rq->nr_iowait, 0);
8154 set_load_weight(&init_task);
8156 #ifdef CONFIG_PREEMPT_NOTIFIERS
8157 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8161 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
8164 #ifdef CONFIG_RT_MUTEXES
8165 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8169 * The boot idle thread does lazy MMU switching as well:
8171 atomic_inc(&init_mm.mm_count);
8172 enter_lazy_tlb(&init_mm, current);
8175 * Make us the idle thread. Technically, schedule() should not be
8176 * called from this thread, however somewhere below it might be,
8177 * but because we are the idle thread, we just pick up running again
8178 * when this runqueue becomes "idle".
8180 init_idle(current, smp_processor_id());
8182 * During early bootup we pretend to be a normal task:
8184 current->sched_class = &fair_sched_class;
8186 scheduler_running = 1;
8189 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8190 void __might_sleep(char *file, int line)
8193 static unsigned long prev_jiffy; /* ratelimiting */
8195 if ((in_atomic() || irqs_disabled()) &&
8196 system_state == SYSTEM_RUNNING && !oops_in_progress) {
8197 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8199 prev_jiffy = jiffies;
8200 printk(KERN_ERR "BUG: sleeping function called from invalid"
8201 " context at %s:%d\n", file, line);
8202 printk("in_atomic():%d, irqs_disabled():%d\n",
8203 in_atomic(), irqs_disabled());
8204 debug_show_held_locks(current);
8205 if (irqs_disabled())
8206 print_irqtrace_events(current);
8211 EXPORT_SYMBOL(__might_sleep);
8214 #ifdef CONFIG_MAGIC_SYSRQ
8215 static void normalize_task(struct rq *rq, struct task_struct *p)
8219 update_rq_clock(rq);
8220 on_rq = p->se.on_rq;
8222 deactivate_task(rq, p, 0);
8223 __setscheduler(rq, p, SCHED_NORMAL, 0);
8225 activate_task(rq, p, 0);
8226 resched_task(rq->curr);
8230 void normalize_rt_tasks(void)
8232 struct task_struct *g, *p;
8233 unsigned long flags;
8236 read_lock_irqsave(&tasklist_lock, flags);
8237 do_each_thread(g, p) {
8239 * Only normalize user tasks:
8244 p->se.exec_start = 0;
8245 #ifdef CONFIG_SCHEDSTATS
8246 p->se.wait_start = 0;
8247 p->se.sleep_start = 0;
8248 p->se.block_start = 0;
8253 * Renice negative nice level userspace
8256 if (TASK_NICE(p) < 0 && p->mm)
8257 set_user_nice(p, 0);
8261 spin_lock(&p->pi_lock);
8262 rq = __task_rq_lock(p);
8264 normalize_task(rq, p);
8266 __task_rq_unlock(rq);
8267 spin_unlock(&p->pi_lock);
8268 } while_each_thread(g, p);
8270 read_unlock_irqrestore(&tasklist_lock, flags);
8273 #endif /* CONFIG_MAGIC_SYSRQ */
8277 * These functions are only useful for the IA64 MCA handling.
8279 * They can only be called when the whole system has been
8280 * stopped - every CPU needs to be quiescent, and no scheduling
8281 * activity can take place. Using them for anything else would
8282 * be a serious bug, and as a result, they aren't even visible
8283 * under any other configuration.
8287 * curr_task - return the current task for a given cpu.
8288 * @cpu: the processor in question.
8290 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8292 struct task_struct *curr_task(int cpu)
8294 return cpu_curr(cpu);
8298 * set_curr_task - set the current task for a given cpu.
8299 * @cpu: the processor in question.
8300 * @p: the task pointer to set.
8302 * Description: This function must only be used when non-maskable interrupts
8303 * are serviced on a separate stack. It allows the architecture to switch the
8304 * notion of the current task on a cpu in a non-blocking manner. This function
8305 * must be called with all CPU's synchronized, and interrupts disabled, the
8306 * and caller must save the original value of the current task (see
8307 * curr_task() above) and restore that value before reenabling interrupts and
8308 * re-starting the system.
8310 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8312 void set_curr_task(int cpu, struct task_struct *p)
8319 #ifdef CONFIG_FAIR_GROUP_SCHED
8320 static void free_fair_sched_group(struct task_group *tg)
8324 for_each_possible_cpu(i) {
8326 kfree(tg->cfs_rq[i]);
8336 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8338 struct cfs_rq *cfs_rq;
8339 struct sched_entity *se, *parent_se;
8343 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8346 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8350 tg->shares = NICE_0_LOAD;
8352 for_each_possible_cpu(i) {
8355 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8356 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8360 se = kmalloc_node(sizeof(struct sched_entity),
8361 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8365 parent_se = parent ? parent->se[i] : NULL;
8366 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
8375 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8377 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8378 &cpu_rq(cpu)->leaf_cfs_rq_list);
8381 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8383 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8385 #else /* !CONFG_FAIR_GROUP_SCHED */
8386 static inline void free_fair_sched_group(struct task_group *tg)
8391 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8396 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8400 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8403 #endif /* CONFIG_FAIR_GROUP_SCHED */
8405 #ifdef CONFIG_RT_GROUP_SCHED
8406 static void free_rt_sched_group(struct task_group *tg)
8410 destroy_rt_bandwidth(&tg->rt_bandwidth);
8412 for_each_possible_cpu(i) {
8414 kfree(tg->rt_rq[i]);
8416 kfree(tg->rt_se[i]);
8424 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8426 struct rt_rq *rt_rq;
8427 struct sched_rt_entity *rt_se, *parent_se;
8431 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8434 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8438 init_rt_bandwidth(&tg->rt_bandwidth,
8439 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8441 for_each_possible_cpu(i) {
8444 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8445 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8449 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8450 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8454 parent_se = parent ? parent->rt_se[i] : NULL;
8455 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
8464 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8466 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8467 &cpu_rq(cpu)->leaf_rt_rq_list);
8470 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8472 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8474 #else /* !CONFIG_RT_GROUP_SCHED */
8475 static inline void free_rt_sched_group(struct task_group *tg)
8480 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8485 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8489 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8492 #endif /* CONFIG_RT_GROUP_SCHED */
8494 #ifdef CONFIG_GROUP_SCHED
8495 static void free_sched_group(struct task_group *tg)
8497 free_fair_sched_group(tg);
8498 free_rt_sched_group(tg);
8502 /* allocate runqueue etc for a new task group */
8503 struct task_group *sched_create_group(struct task_group *parent)
8505 struct task_group *tg;
8506 unsigned long flags;
8509 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8511 return ERR_PTR(-ENOMEM);
8513 if (!alloc_fair_sched_group(tg, parent))
8516 if (!alloc_rt_sched_group(tg, parent))
8519 spin_lock_irqsave(&task_group_lock, flags);
8520 for_each_possible_cpu(i) {
8521 register_fair_sched_group(tg, i);
8522 register_rt_sched_group(tg, i);
8524 list_add_rcu(&tg->list, &task_groups);
8526 WARN_ON(!parent); /* root should already exist */
8528 tg->parent = parent;
8529 list_add_rcu(&tg->siblings, &parent->children);
8530 INIT_LIST_HEAD(&tg->children);
8531 spin_unlock_irqrestore(&task_group_lock, flags);
8536 free_sched_group(tg);
8537 return ERR_PTR(-ENOMEM);
8540 /* rcu callback to free various structures associated with a task group */
8541 static void free_sched_group_rcu(struct rcu_head *rhp)
8543 /* now it should be safe to free those cfs_rqs */
8544 free_sched_group(container_of(rhp, struct task_group, rcu));
8547 /* Destroy runqueue etc associated with a task group */
8548 void sched_destroy_group(struct task_group *tg)
8550 unsigned long flags;
8553 spin_lock_irqsave(&task_group_lock, flags);
8554 for_each_possible_cpu(i) {
8555 unregister_fair_sched_group(tg, i);
8556 unregister_rt_sched_group(tg, i);
8558 list_del_rcu(&tg->list);
8559 list_del_rcu(&tg->siblings);
8560 spin_unlock_irqrestore(&task_group_lock, flags);
8562 /* wait for possible concurrent references to cfs_rqs complete */
8563 call_rcu(&tg->rcu, free_sched_group_rcu);
8566 /* change task's runqueue when it moves between groups.
8567 * The caller of this function should have put the task in its new group
8568 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8569 * reflect its new group.
8571 void sched_move_task(struct task_struct *tsk)
8574 unsigned long flags;
8577 rq = task_rq_lock(tsk, &flags);
8579 update_rq_clock(rq);
8581 running = task_current(rq, tsk);
8582 on_rq = tsk->se.on_rq;
8585 dequeue_task(rq, tsk, 0);
8586 if (unlikely(running))
8587 tsk->sched_class->put_prev_task(rq, tsk);
8589 set_task_rq(tsk, task_cpu(tsk));
8591 #ifdef CONFIG_FAIR_GROUP_SCHED
8592 if (tsk->sched_class->moved_group)
8593 tsk->sched_class->moved_group(tsk);
8596 if (unlikely(running))
8597 tsk->sched_class->set_curr_task(rq);
8599 enqueue_task(rq, tsk, 0);
8601 task_rq_unlock(rq, &flags);
8603 #endif /* CONFIG_GROUP_SCHED */
8605 #ifdef CONFIG_FAIR_GROUP_SCHED
8606 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8608 struct cfs_rq *cfs_rq = se->cfs_rq;
8613 dequeue_entity(cfs_rq, se, 0);
8615 se->load.weight = shares;
8616 se->load.inv_weight = 0;
8619 enqueue_entity(cfs_rq, se, 0);
8622 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8624 struct cfs_rq *cfs_rq = se->cfs_rq;
8625 struct rq *rq = cfs_rq->rq;
8626 unsigned long flags;
8628 spin_lock_irqsave(&rq->lock, flags);
8629 __set_se_shares(se, shares);
8630 spin_unlock_irqrestore(&rq->lock, flags);
8633 static DEFINE_MUTEX(shares_mutex);
8635 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8638 unsigned long flags;
8641 * We can't change the weight of the root cgroup.
8646 if (shares < MIN_SHARES)
8647 shares = MIN_SHARES;
8648 else if (shares > MAX_SHARES)
8649 shares = MAX_SHARES;
8651 mutex_lock(&shares_mutex);
8652 if (tg->shares == shares)
8655 spin_lock_irqsave(&task_group_lock, flags);
8656 for_each_possible_cpu(i)
8657 unregister_fair_sched_group(tg, i);
8658 list_del_rcu(&tg->siblings);
8659 spin_unlock_irqrestore(&task_group_lock, flags);
8661 /* wait for any ongoing reference to this group to finish */
8662 synchronize_sched();
8665 * Now we are free to modify the group's share on each cpu
8666 * w/o tripping rebalance_share or load_balance_fair.
8668 tg->shares = shares;
8669 for_each_possible_cpu(i) {
8673 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8674 set_se_shares(tg->se[i], shares);
8678 * Enable load balance activity on this group, by inserting it back on
8679 * each cpu's rq->leaf_cfs_rq_list.
8681 spin_lock_irqsave(&task_group_lock, flags);
8682 for_each_possible_cpu(i)
8683 register_fair_sched_group(tg, i);
8684 list_add_rcu(&tg->siblings, &tg->parent->children);
8685 spin_unlock_irqrestore(&task_group_lock, flags);
8687 mutex_unlock(&shares_mutex);
8691 unsigned long sched_group_shares(struct task_group *tg)
8697 #ifdef CONFIG_RT_GROUP_SCHED
8699 * Ensure that the real time constraints are schedulable.
8701 static DEFINE_MUTEX(rt_constraints_mutex);
8703 static unsigned long to_ratio(u64 period, u64 runtime)
8705 if (runtime == RUNTIME_INF)
8708 return div64_u64(runtime << 16, period);
8711 #ifdef CONFIG_CGROUP_SCHED
8712 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8714 struct task_group *tgi, *parent = tg->parent;
8715 unsigned long total = 0;
8718 if (global_rt_period() < period)
8721 return to_ratio(period, runtime) <
8722 to_ratio(global_rt_period(), global_rt_runtime());
8725 if (ktime_to_ns(parent->rt_bandwidth.rt_period) < period)
8729 list_for_each_entry_rcu(tgi, &parent->children, siblings) {
8733 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8734 tgi->rt_bandwidth.rt_runtime);
8738 return total + to_ratio(period, runtime) <=
8739 to_ratio(ktime_to_ns(parent->rt_bandwidth.rt_period),
8740 parent->rt_bandwidth.rt_runtime);
8742 #elif defined CONFIG_USER_SCHED
8743 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8745 struct task_group *tgi;
8746 unsigned long total = 0;
8747 unsigned long global_ratio =
8748 to_ratio(global_rt_period(), global_rt_runtime());
8751 list_for_each_entry_rcu(tgi, &task_groups, list) {
8755 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8756 tgi->rt_bandwidth.rt_runtime);
8760 return total + to_ratio(period, runtime) < global_ratio;
8764 /* Must be called with tasklist_lock held */
8765 static inline int tg_has_rt_tasks(struct task_group *tg)
8767 struct task_struct *g, *p;
8768 do_each_thread(g, p) {
8769 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8771 } while_each_thread(g, p);
8775 static int tg_set_bandwidth(struct task_group *tg,
8776 u64 rt_period, u64 rt_runtime)
8780 mutex_lock(&rt_constraints_mutex);
8781 read_lock(&tasklist_lock);
8782 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
8786 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8791 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8792 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8793 tg->rt_bandwidth.rt_runtime = rt_runtime;
8795 for_each_possible_cpu(i) {
8796 struct rt_rq *rt_rq = tg->rt_rq[i];
8798 spin_lock(&rt_rq->rt_runtime_lock);
8799 rt_rq->rt_runtime = rt_runtime;
8800 spin_unlock(&rt_rq->rt_runtime_lock);
8802 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8804 read_unlock(&tasklist_lock);
8805 mutex_unlock(&rt_constraints_mutex);
8810 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8812 u64 rt_runtime, rt_period;
8814 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8815 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8816 if (rt_runtime_us < 0)
8817 rt_runtime = RUNTIME_INF;
8819 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8822 long sched_group_rt_runtime(struct task_group *tg)
8826 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8829 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8830 do_div(rt_runtime_us, NSEC_PER_USEC);
8831 return rt_runtime_us;
8834 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8836 u64 rt_runtime, rt_period;
8838 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8839 rt_runtime = tg->rt_bandwidth.rt_runtime;
8841 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8844 long sched_group_rt_period(struct task_group *tg)
8848 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8849 do_div(rt_period_us, NSEC_PER_USEC);
8850 return rt_period_us;
8853 static int sched_rt_global_constraints(void)
8855 struct task_group *tg = &root_task_group;
8856 u64 rt_runtime, rt_period;
8859 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8860 rt_runtime = tg->rt_bandwidth.rt_runtime;
8862 mutex_lock(&rt_constraints_mutex);
8863 if (!__rt_schedulable(tg, rt_period, rt_runtime))
8865 mutex_unlock(&rt_constraints_mutex);
8869 #else /* !CONFIG_RT_GROUP_SCHED */
8870 static int sched_rt_global_constraints(void)
8872 unsigned long flags;
8875 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8876 for_each_possible_cpu(i) {
8877 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8879 spin_lock(&rt_rq->rt_runtime_lock);
8880 rt_rq->rt_runtime = global_rt_runtime();
8881 spin_unlock(&rt_rq->rt_runtime_lock);
8883 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8887 #endif /* CONFIG_RT_GROUP_SCHED */
8889 int sched_rt_handler(struct ctl_table *table, int write,
8890 struct file *filp, void __user *buffer, size_t *lenp,
8894 int old_period, old_runtime;
8895 static DEFINE_MUTEX(mutex);
8898 old_period = sysctl_sched_rt_period;
8899 old_runtime = sysctl_sched_rt_runtime;
8901 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8903 if (!ret && write) {
8904 ret = sched_rt_global_constraints();
8906 sysctl_sched_rt_period = old_period;
8907 sysctl_sched_rt_runtime = old_runtime;
8909 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8910 def_rt_bandwidth.rt_period =
8911 ns_to_ktime(global_rt_period());
8914 mutex_unlock(&mutex);
8919 #ifdef CONFIG_CGROUP_SCHED
8921 /* return corresponding task_group object of a cgroup */
8922 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8924 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8925 struct task_group, css);
8928 static struct cgroup_subsys_state *
8929 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8931 struct task_group *tg, *parent;
8933 if (!cgrp->parent) {
8934 /* This is early initialization for the top cgroup */
8935 init_task_group.css.cgroup = cgrp;
8936 return &init_task_group.css;
8939 parent = cgroup_tg(cgrp->parent);
8940 tg = sched_create_group(parent);
8942 return ERR_PTR(-ENOMEM);
8944 /* Bind the cgroup to task_group object we just created */
8945 tg->css.cgroup = cgrp;
8951 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8953 struct task_group *tg = cgroup_tg(cgrp);
8955 sched_destroy_group(tg);
8959 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8960 struct task_struct *tsk)
8962 #ifdef CONFIG_RT_GROUP_SCHED
8963 /* Don't accept realtime tasks when there is no way for them to run */
8964 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
8967 /* We don't support RT-tasks being in separate groups */
8968 if (tsk->sched_class != &fair_sched_class)
8976 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8977 struct cgroup *old_cont, struct task_struct *tsk)
8979 sched_move_task(tsk);
8982 #ifdef CONFIG_FAIR_GROUP_SCHED
8983 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8986 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8989 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8991 struct task_group *tg = cgroup_tg(cgrp);
8993 return (u64) tg->shares;
8995 #endif /* CONFIG_FAIR_GROUP_SCHED */
8997 #ifdef CONFIG_RT_GROUP_SCHED
8998 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9001 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9004 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9006 return sched_group_rt_runtime(cgroup_tg(cgrp));
9009 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9012 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9015 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9017 return sched_group_rt_period(cgroup_tg(cgrp));
9019 #endif /* CONFIG_RT_GROUP_SCHED */
9021 static struct cftype cpu_files[] = {
9022 #ifdef CONFIG_FAIR_GROUP_SCHED
9025 .read_u64 = cpu_shares_read_u64,
9026 .write_u64 = cpu_shares_write_u64,
9029 #ifdef CONFIG_RT_GROUP_SCHED
9031 .name = "rt_runtime_us",
9032 .read_s64 = cpu_rt_runtime_read,
9033 .write_s64 = cpu_rt_runtime_write,
9036 .name = "rt_period_us",
9037 .read_u64 = cpu_rt_period_read_uint,
9038 .write_u64 = cpu_rt_period_write_uint,
9043 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9045 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9048 struct cgroup_subsys cpu_cgroup_subsys = {
9050 .create = cpu_cgroup_create,
9051 .destroy = cpu_cgroup_destroy,
9052 .can_attach = cpu_cgroup_can_attach,
9053 .attach = cpu_cgroup_attach,
9054 .populate = cpu_cgroup_populate,
9055 .subsys_id = cpu_cgroup_subsys_id,
9059 #endif /* CONFIG_CGROUP_SCHED */
9061 #ifdef CONFIG_CGROUP_CPUACCT
9064 * CPU accounting code for task groups.
9066 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9067 * (balbir@in.ibm.com).
9070 /* track cpu usage of a group of tasks */
9072 struct cgroup_subsys_state css;
9073 /* cpuusage holds pointer to a u64-type object on every cpu */
9077 struct cgroup_subsys cpuacct_subsys;
9079 /* return cpu accounting group corresponding to this container */
9080 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9082 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9083 struct cpuacct, css);
9086 /* return cpu accounting group to which this task belongs */
9087 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9089 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9090 struct cpuacct, css);
9093 /* create a new cpu accounting group */
9094 static struct cgroup_subsys_state *cpuacct_create(
9095 struct cgroup_subsys *ss, struct cgroup *cgrp)
9097 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9100 return ERR_PTR(-ENOMEM);
9102 ca->cpuusage = alloc_percpu(u64);
9103 if (!ca->cpuusage) {
9105 return ERR_PTR(-ENOMEM);
9111 /* destroy an existing cpu accounting group */
9113 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9115 struct cpuacct *ca = cgroup_ca(cgrp);
9117 free_percpu(ca->cpuusage);
9121 /* return total cpu usage (in nanoseconds) of a group */
9122 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9124 struct cpuacct *ca = cgroup_ca(cgrp);
9125 u64 totalcpuusage = 0;
9128 for_each_possible_cpu(i) {
9129 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9132 * Take rq->lock to make 64-bit addition safe on 32-bit
9135 spin_lock_irq(&cpu_rq(i)->lock);
9136 totalcpuusage += *cpuusage;
9137 spin_unlock_irq(&cpu_rq(i)->lock);
9140 return totalcpuusage;
9143 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9146 struct cpuacct *ca = cgroup_ca(cgrp);
9155 for_each_possible_cpu(i) {
9156 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9158 spin_lock_irq(&cpu_rq(i)->lock);
9160 spin_unlock_irq(&cpu_rq(i)->lock);
9166 static struct cftype files[] = {
9169 .read_u64 = cpuusage_read,
9170 .write_u64 = cpuusage_write,
9174 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9176 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9180 * charge this task's execution time to its accounting group.
9182 * called with rq->lock held.
9184 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9188 if (!cpuacct_subsys.active)
9193 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
9195 *cpuusage += cputime;
9199 struct cgroup_subsys cpuacct_subsys = {
9201 .create = cpuacct_create,
9202 .destroy = cpuacct_destroy,
9203 .populate = cpuacct_populate,
9204 .subsys_id = cpuacct_subsys_id,
9206 #endif /* CONFIG_CGROUP_CPUACCT */