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>
73 #include <linux/ftrace.h>
76 #include <asm/irq_regs.h>
78 #include "sched_cpupri.h"
81 * Convert user-nice values [ -20 ... 0 ... 19 ]
82 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
85 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
86 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
87 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
90 * 'User priority' is the nice value converted to something we
91 * can work with better when scaling various scheduler parameters,
92 * it's a [ 0 ... 39 ] range.
94 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
95 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
96 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
99 * Helpers for converting nanosecond timing to jiffy resolution
101 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
103 #define NICE_0_LOAD SCHED_LOAD_SCALE
104 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
107 * These are the 'tuning knobs' of the scheduler:
109 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
110 * Timeslices get refilled after they expire.
112 #define DEF_TIMESLICE (100 * HZ / 1000)
115 * single value that denotes runtime == period, ie unlimited time.
117 #define RUNTIME_INF ((u64)~0ULL)
121 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
122 * Since cpu_power is a 'constant', we can use a reciprocal divide.
124 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
126 return reciprocal_divide(load, sg->reciprocal_cpu_power);
130 * Each time a sched group cpu_power is changed,
131 * we must compute its reciprocal value
133 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
135 sg->__cpu_power += val;
136 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
140 static inline int rt_policy(int policy)
142 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
147 static inline int task_has_rt_policy(struct task_struct *p)
149 return rt_policy(p->policy);
153 * This is the priority-queue data structure of the RT scheduling class:
155 struct rt_prio_array {
156 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
157 struct list_head queue[MAX_RT_PRIO];
160 struct rt_bandwidth {
161 /* nests inside the rq lock: */
162 spinlock_t rt_runtime_lock;
165 struct hrtimer rt_period_timer;
168 static struct rt_bandwidth def_rt_bandwidth;
170 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
172 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
174 struct rt_bandwidth *rt_b =
175 container_of(timer, struct rt_bandwidth, rt_period_timer);
181 now = hrtimer_cb_get_time(timer);
182 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
187 idle = do_sched_rt_period_timer(rt_b, overrun);
190 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
194 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
196 rt_b->rt_period = ns_to_ktime(period);
197 rt_b->rt_runtime = runtime;
199 spin_lock_init(&rt_b->rt_runtime_lock);
201 hrtimer_init(&rt_b->rt_period_timer,
202 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
203 rt_b->rt_period_timer.function = sched_rt_period_timer;
204 rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
207 static inline int rt_bandwidth_enabled(void);
209 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
213 if (rt_bandwidth_enabled() && rt_b->rt_runtime == RUNTIME_INF)
216 if (hrtimer_active(&rt_b->rt_period_timer))
219 spin_lock(&rt_b->rt_runtime_lock);
221 if (hrtimer_active(&rt_b->rt_period_timer))
224 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
225 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
226 hrtimer_start(&rt_b->rt_period_timer,
227 rt_b->rt_period_timer.expires,
230 spin_unlock(&rt_b->rt_runtime_lock);
233 #ifdef CONFIG_RT_GROUP_SCHED
234 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
236 hrtimer_cancel(&rt_b->rt_period_timer);
241 * sched_domains_mutex serializes calls to arch_init_sched_domains,
242 * detach_destroy_domains and partition_sched_domains.
244 static DEFINE_MUTEX(sched_domains_mutex);
246 #ifdef CONFIG_GROUP_SCHED
248 #include <linux/cgroup.h>
252 static LIST_HEAD(task_groups);
254 /* task group related information */
256 #ifdef CONFIG_CGROUP_SCHED
257 struct cgroup_subsys_state css;
260 #ifdef CONFIG_FAIR_GROUP_SCHED
261 /* schedulable entities of this group on each cpu */
262 struct sched_entity **se;
263 /* runqueue "owned" by this group on each cpu */
264 struct cfs_rq **cfs_rq;
265 unsigned long shares;
268 #ifdef CONFIG_RT_GROUP_SCHED
269 struct sched_rt_entity **rt_se;
270 struct rt_rq **rt_rq;
272 struct rt_bandwidth rt_bandwidth;
276 struct list_head list;
278 struct task_group *parent;
279 struct list_head siblings;
280 struct list_head children;
283 #ifdef CONFIG_USER_SCHED
287 * Every UID task group (including init_task_group aka UID-0) will
288 * be a child to this group.
290 struct task_group root_task_group;
292 #ifdef CONFIG_FAIR_GROUP_SCHED
293 /* Default task group's sched entity on each cpu */
294 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
295 /* Default task group's cfs_rq on each cpu */
296 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
297 #endif /* CONFIG_FAIR_GROUP_SCHED */
299 #ifdef CONFIG_RT_GROUP_SCHED
300 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
301 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
302 #endif /* CONFIG_RT_GROUP_SCHED */
303 #else /* !CONFIG_FAIR_GROUP_SCHED */
304 #define root_task_group init_task_group
305 #endif /* CONFIG_FAIR_GROUP_SCHED */
307 /* task_group_lock serializes add/remove of task groups and also changes to
308 * a task group's cpu shares.
310 static DEFINE_SPINLOCK(task_group_lock);
312 #ifdef CONFIG_FAIR_GROUP_SCHED
313 #ifdef CONFIG_USER_SCHED
314 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
315 #else /* !CONFIG_USER_SCHED */
316 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
317 #endif /* CONFIG_USER_SCHED */
320 * A weight of 0 or 1 can cause arithmetics problems.
321 * A weight of a cfs_rq is the sum of weights of which entities
322 * are queued on this cfs_rq, so a weight of a entity should not be
323 * too large, so as the shares value of a task group.
324 * (The default weight is 1024 - so there's no practical
325 * limitation from this.)
328 #define MAX_SHARES (1UL << 18)
330 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
333 /* Default task group.
334 * Every task in system belong to this group at bootup.
336 struct task_group init_task_group;
338 /* return group to which a task belongs */
339 static inline struct task_group *task_group(struct task_struct *p)
341 struct task_group *tg;
343 #ifdef CONFIG_USER_SCHED
345 #elif defined(CONFIG_CGROUP_SCHED)
346 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
347 struct task_group, css);
349 tg = &init_task_group;
354 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
355 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
357 #ifdef CONFIG_FAIR_GROUP_SCHED
358 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
359 p->se.parent = task_group(p)->se[cpu];
362 #ifdef CONFIG_RT_GROUP_SCHED
363 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
364 p->rt.parent = task_group(p)->rt_se[cpu];
370 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
371 static inline struct task_group *task_group(struct task_struct *p)
376 #endif /* CONFIG_GROUP_SCHED */
378 /* CFS-related fields in a runqueue */
380 struct load_weight load;
381 unsigned long nr_running;
387 struct rb_root tasks_timeline;
388 struct rb_node *rb_leftmost;
390 struct list_head tasks;
391 struct list_head *balance_iterator;
394 * 'curr' points to currently running entity on this cfs_rq.
395 * It is set to NULL otherwise (i.e when none are currently running).
397 struct sched_entity *curr, *next;
399 unsigned long nr_spread_over;
401 #ifdef CONFIG_FAIR_GROUP_SCHED
402 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
405 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
406 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
407 * (like users, containers etc.)
409 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
410 * list is used during load balance.
412 struct list_head leaf_cfs_rq_list;
413 struct task_group *tg; /* group that "owns" this runqueue */
417 * the part of load.weight contributed by tasks
419 unsigned long task_weight;
422 * h_load = weight * f(tg)
424 * Where f(tg) is the recursive weight fraction assigned to
427 unsigned long h_load;
430 * this cpu's part of tg->shares
432 unsigned long shares;
435 * load.weight at the time we set shares
437 unsigned long rq_weight;
442 /* Real-Time classes' related field in a runqueue: */
444 struct rt_prio_array active;
445 unsigned long rt_nr_running;
446 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
447 int highest_prio; /* highest queued rt task prio */
450 unsigned long rt_nr_migratory;
456 /* Nests inside the rq lock: */
457 spinlock_t rt_runtime_lock;
459 #ifdef CONFIG_RT_GROUP_SCHED
460 unsigned long rt_nr_boosted;
463 struct list_head leaf_rt_rq_list;
464 struct task_group *tg;
465 struct sched_rt_entity *rt_se;
472 * We add the notion of a root-domain which will be used to define per-domain
473 * variables. Each exclusive cpuset essentially defines an island domain by
474 * fully partitioning the member cpus from any other cpuset. Whenever a new
475 * exclusive cpuset is created, we also create and attach a new root-domain
485 * The "RT overload" flag: it gets set if a CPU has more than
486 * one runnable RT task.
491 struct cpupri cpupri;
496 * By default the system creates a single root-domain with all cpus as
497 * members (mimicking the global state we have today).
499 static struct root_domain def_root_domain;
504 * This is the main, per-CPU runqueue data structure.
506 * Locking rule: those places that want to lock multiple runqueues
507 * (such as the load balancing or the thread migration code), lock
508 * acquire operations must be ordered by ascending &runqueue.
515 * nr_running and cpu_load should be in the same cacheline because
516 * remote CPUs use both these fields when doing load calculation.
518 unsigned long nr_running;
519 #define CPU_LOAD_IDX_MAX 5
520 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
521 unsigned char idle_at_tick;
523 unsigned long last_tick_seen;
524 unsigned char in_nohz_recently;
526 /* capture load from *all* tasks on this cpu: */
527 struct load_weight load;
528 unsigned long nr_load_updates;
534 #ifdef CONFIG_FAIR_GROUP_SCHED
535 /* list of leaf cfs_rq on this cpu: */
536 struct list_head leaf_cfs_rq_list;
538 #ifdef CONFIG_RT_GROUP_SCHED
539 struct list_head leaf_rt_rq_list;
543 * This is part of a global counter where only the total sum
544 * over all CPUs matters. A task can increase this counter on
545 * one CPU and if it got migrated afterwards it may decrease
546 * it on another CPU. Always updated under the runqueue lock:
548 unsigned long nr_uninterruptible;
550 struct task_struct *curr, *idle;
551 unsigned long next_balance;
552 struct mm_struct *prev_mm;
559 struct root_domain *rd;
560 struct sched_domain *sd;
562 /* For active balancing */
565 /* cpu of this runqueue: */
569 unsigned long avg_load_per_task;
571 struct task_struct *migration_thread;
572 struct list_head migration_queue;
575 #ifdef CONFIG_SCHED_HRTICK
577 int hrtick_csd_pending;
578 struct call_single_data hrtick_csd;
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;
607 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
609 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
611 rq->curr->sched_class->check_preempt_curr(rq, p);
614 static inline int cpu_of(struct rq *rq)
624 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
625 * See detach_destroy_domains: synchronize_sched for details.
627 * The domain tree of any CPU may only be accessed from within
628 * preempt-disabled sections.
630 #define for_each_domain(cpu, __sd) \
631 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
633 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
634 #define this_rq() (&__get_cpu_var(runqueues))
635 #define task_rq(p) cpu_rq(task_cpu(p))
636 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
638 static inline void update_rq_clock(struct rq *rq)
640 rq->clock = sched_clock_cpu(cpu_of(rq));
644 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
646 #ifdef CONFIG_SCHED_DEBUG
647 # define const_debug __read_mostly
649 # define const_debug static const
655 * Returns true if the current cpu runqueue is locked.
656 * This interface allows printk to be called with the runqueue lock
657 * held and know whether or not it is OK to wake up the klogd.
659 int runqueue_is_locked(void)
662 struct rq *rq = cpu_rq(cpu);
665 ret = spin_is_locked(&rq->lock);
671 * Debugging: various feature bits
674 #define SCHED_FEAT(name, enabled) \
675 __SCHED_FEAT_##name ,
678 #include "sched_features.h"
683 #define SCHED_FEAT(name, enabled) \
684 (1UL << __SCHED_FEAT_##name) * enabled |
686 const_debug unsigned int sysctl_sched_features =
687 #include "sched_features.h"
692 #ifdef CONFIG_SCHED_DEBUG
693 #define SCHED_FEAT(name, enabled) \
696 static __read_mostly char *sched_feat_names[] = {
697 #include "sched_features.h"
703 static int sched_feat_open(struct inode *inode, struct file *filp)
705 filp->private_data = inode->i_private;
710 sched_feat_read(struct file *filp, char __user *ubuf,
711 size_t cnt, loff_t *ppos)
718 for (i = 0; sched_feat_names[i]; i++) {
719 len += strlen(sched_feat_names[i]);
723 buf = kmalloc(len + 2, GFP_KERNEL);
727 for (i = 0; sched_feat_names[i]; i++) {
728 if (sysctl_sched_features & (1UL << i))
729 r += sprintf(buf + r, "%s ", sched_feat_names[i]);
731 r += sprintf(buf + r, "NO_%s ", sched_feat_names[i]);
734 r += sprintf(buf + r, "\n");
735 WARN_ON(r >= len + 2);
737 r = simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
745 sched_feat_write(struct file *filp, const char __user *ubuf,
746 size_t cnt, loff_t *ppos)
756 if (copy_from_user(&buf, ubuf, cnt))
761 if (strncmp(buf, "NO_", 3) == 0) {
766 for (i = 0; sched_feat_names[i]; i++) {
767 int len = strlen(sched_feat_names[i]);
769 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
771 sysctl_sched_features &= ~(1UL << i);
773 sysctl_sched_features |= (1UL << i);
778 if (!sched_feat_names[i])
786 static struct file_operations sched_feat_fops = {
787 .open = sched_feat_open,
788 .read = sched_feat_read,
789 .write = sched_feat_write,
792 static __init int sched_init_debug(void)
794 debugfs_create_file("sched_features", 0644, NULL, NULL,
799 late_initcall(sched_init_debug);
803 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
806 * Number of tasks to iterate in a single balance run.
807 * Limited because this is done with IRQs disabled.
809 const_debug unsigned int sysctl_sched_nr_migrate = 32;
812 * ratelimit for updating the group shares.
815 unsigned int sysctl_sched_shares_ratelimit = 250000;
818 * period over which we measure -rt task cpu usage in us.
821 unsigned int sysctl_sched_rt_period = 1000000;
823 static __read_mostly int scheduler_running;
826 * part of the period that we allow rt tasks to run in us.
829 int sysctl_sched_rt_runtime = 950000;
831 static inline u64 global_rt_period(void)
833 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
836 static inline u64 global_rt_runtime(void)
838 if (sysctl_sched_rt_runtime < 0)
841 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
844 static inline int rt_bandwidth_enabled(void)
846 return sysctl_sched_rt_runtime >= 0;
849 #ifndef prepare_arch_switch
850 # define prepare_arch_switch(next) do { } while (0)
852 #ifndef finish_arch_switch
853 # define finish_arch_switch(prev) do { } while (0)
856 static inline int task_current(struct rq *rq, struct task_struct *p)
858 return rq->curr == p;
861 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
862 static inline int task_running(struct rq *rq, struct task_struct *p)
864 return task_current(rq, p);
867 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
871 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
873 #ifdef CONFIG_DEBUG_SPINLOCK
874 /* this is a valid case when another task releases the spinlock */
875 rq->lock.owner = current;
878 * If we are tracking spinlock dependencies then we have to
879 * fix up the runqueue lock - which gets 'carried over' from
882 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
884 spin_unlock_irq(&rq->lock);
887 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
888 static inline int task_running(struct rq *rq, struct task_struct *p)
893 return task_current(rq, p);
897 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
901 * We can optimise this out completely for !SMP, because the
902 * SMP rebalancing from interrupt is the only thing that cares
907 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
908 spin_unlock_irq(&rq->lock);
910 spin_unlock(&rq->lock);
914 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
918 * After ->oncpu is cleared, the task can be moved to a different CPU.
919 * We must ensure this doesn't happen until the switch is completely
925 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
929 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
932 * __task_rq_lock - lock the runqueue a given task resides on.
933 * Must be called interrupts disabled.
935 static inline struct rq *__task_rq_lock(struct task_struct *p)
939 struct rq *rq = task_rq(p);
940 spin_lock(&rq->lock);
941 if (likely(rq == task_rq(p)))
943 spin_unlock(&rq->lock);
948 * task_rq_lock - lock the runqueue a given task resides on and disable
949 * interrupts. Note the ordering: we can safely lookup the task_rq without
950 * explicitly disabling preemption.
952 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
958 local_irq_save(*flags);
960 spin_lock(&rq->lock);
961 if (likely(rq == task_rq(p)))
963 spin_unlock_irqrestore(&rq->lock, *flags);
967 static void __task_rq_unlock(struct rq *rq)
970 spin_unlock(&rq->lock);
973 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
976 spin_unlock_irqrestore(&rq->lock, *flags);
980 * this_rq_lock - lock this runqueue and disable interrupts.
982 static struct rq *this_rq_lock(void)
989 spin_lock(&rq->lock);
994 #ifdef CONFIG_SCHED_HRTICK
996 * Use HR-timers to deliver accurate preemption points.
998 * Its all a bit involved since we cannot program an hrt while holding the
999 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1002 * When we get rescheduled we reprogram the hrtick_timer outside of the
1008 * - enabled by features
1009 * - hrtimer is actually high res
1011 static inline int hrtick_enabled(struct rq *rq)
1013 if (!sched_feat(HRTICK))
1015 if (!cpu_active(cpu_of(rq)))
1017 return hrtimer_is_hres_active(&rq->hrtick_timer);
1020 static void hrtick_clear(struct rq *rq)
1022 if (hrtimer_active(&rq->hrtick_timer))
1023 hrtimer_cancel(&rq->hrtick_timer);
1027 * High-resolution timer tick.
1028 * Runs from hardirq context with interrupts disabled.
1030 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1032 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1034 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1036 spin_lock(&rq->lock);
1037 update_rq_clock(rq);
1038 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1039 spin_unlock(&rq->lock);
1041 return HRTIMER_NORESTART;
1046 * called from hardirq (IPI) context
1048 static void __hrtick_start(void *arg)
1050 struct rq *rq = arg;
1052 spin_lock(&rq->lock);
1053 hrtimer_restart(&rq->hrtick_timer);
1054 rq->hrtick_csd_pending = 0;
1055 spin_unlock(&rq->lock);
1059 * Called to set the hrtick timer state.
1061 * called with rq->lock held and irqs disabled
1063 static void hrtick_start(struct rq *rq, u64 delay)
1065 struct hrtimer *timer = &rq->hrtick_timer;
1066 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1068 timer->expires = time;
1070 if (rq == this_rq()) {
1071 hrtimer_restart(timer);
1072 } else if (!rq->hrtick_csd_pending) {
1073 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd);
1074 rq->hrtick_csd_pending = 1;
1079 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1081 int cpu = (int)(long)hcpu;
1084 case CPU_UP_CANCELED:
1085 case CPU_UP_CANCELED_FROZEN:
1086 case CPU_DOWN_PREPARE:
1087 case CPU_DOWN_PREPARE_FROZEN:
1089 case CPU_DEAD_FROZEN:
1090 hrtick_clear(cpu_rq(cpu));
1097 static void init_hrtick(void)
1099 hotcpu_notifier(hotplug_hrtick, 0);
1103 * Called to set the hrtick timer state.
1105 * called with rq->lock held and irqs disabled
1107 static void hrtick_start(struct rq *rq, u64 delay)
1109 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay), HRTIMER_MODE_REL);
1112 static void init_hrtick(void)
1115 #endif /* CONFIG_SMP */
1117 static void init_rq_hrtick(struct rq *rq)
1120 rq->hrtick_csd_pending = 0;
1122 rq->hrtick_csd.flags = 0;
1123 rq->hrtick_csd.func = __hrtick_start;
1124 rq->hrtick_csd.info = rq;
1127 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1128 rq->hrtick_timer.function = hrtick;
1129 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1132 static inline void hrtick_clear(struct rq *rq)
1136 static inline void init_rq_hrtick(struct rq *rq)
1140 static inline void init_hrtick(void)
1146 * resched_task - mark a task 'to be rescheduled now'.
1148 * On UP this means the setting of the need_resched flag, on SMP it
1149 * might also involve a cross-CPU call to trigger the scheduler on
1154 #ifndef tsk_is_polling
1155 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1158 static void resched_task(struct task_struct *p)
1162 assert_spin_locked(&task_rq(p)->lock);
1164 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1167 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1170 if (cpu == smp_processor_id())
1173 /* NEED_RESCHED must be visible before we test polling */
1175 if (!tsk_is_polling(p))
1176 smp_send_reschedule(cpu);
1179 static void resched_cpu(int cpu)
1181 struct rq *rq = cpu_rq(cpu);
1182 unsigned long flags;
1184 if (!spin_trylock_irqsave(&rq->lock, flags))
1186 resched_task(cpu_curr(cpu));
1187 spin_unlock_irqrestore(&rq->lock, flags);
1192 * When add_timer_on() enqueues a timer into the timer wheel of an
1193 * idle CPU then this timer might expire before the next timer event
1194 * which is scheduled to wake up that CPU. In case of a completely
1195 * idle system the next event might even be infinite time into the
1196 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1197 * leaves the inner idle loop so the newly added timer is taken into
1198 * account when the CPU goes back to idle and evaluates the timer
1199 * wheel for the next timer event.
1201 void wake_up_idle_cpu(int cpu)
1203 struct rq *rq = cpu_rq(cpu);
1205 if (cpu == smp_processor_id())
1209 * This is safe, as this function is called with the timer
1210 * wheel base lock of (cpu) held. When the CPU is on the way
1211 * to idle and has not yet set rq->curr to idle then it will
1212 * be serialized on the timer wheel base lock and take the new
1213 * timer into account automatically.
1215 if (rq->curr != rq->idle)
1219 * We can set TIF_RESCHED on the idle task of the other CPU
1220 * lockless. The worst case is that the other CPU runs the
1221 * idle task through an additional NOOP schedule()
1223 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1225 /* NEED_RESCHED must be visible before we test polling */
1227 if (!tsk_is_polling(rq->idle))
1228 smp_send_reschedule(cpu);
1230 #endif /* CONFIG_NO_HZ */
1232 #else /* !CONFIG_SMP */
1233 static void resched_task(struct task_struct *p)
1235 assert_spin_locked(&task_rq(p)->lock);
1236 set_tsk_need_resched(p);
1238 #endif /* CONFIG_SMP */
1240 #if BITS_PER_LONG == 32
1241 # define WMULT_CONST (~0UL)
1243 # define WMULT_CONST (1UL << 32)
1246 #define WMULT_SHIFT 32
1249 * Shift right and round:
1251 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1254 * delta *= weight / lw
1256 static unsigned long
1257 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1258 struct load_weight *lw)
1262 if (!lw->inv_weight) {
1263 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1266 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1270 tmp = (u64)delta_exec * weight;
1272 * Check whether we'd overflow the 64-bit multiplication:
1274 if (unlikely(tmp > WMULT_CONST))
1275 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1278 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1280 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1283 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1289 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1296 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1297 * of tasks with abnormal "nice" values across CPUs the contribution that
1298 * each task makes to its run queue's load is weighted according to its
1299 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1300 * scaled version of the new time slice allocation that they receive on time
1304 #define WEIGHT_IDLEPRIO 2
1305 #define WMULT_IDLEPRIO (1 << 31)
1308 * Nice levels are multiplicative, with a gentle 10% change for every
1309 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1310 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1311 * that remained on nice 0.
1313 * The "10% effect" is relative and cumulative: from _any_ nice level,
1314 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1315 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1316 * If a task goes up by ~10% and another task goes down by ~10% then
1317 * the relative distance between them is ~25%.)
1319 static const int prio_to_weight[40] = {
1320 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1321 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1322 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1323 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1324 /* 0 */ 1024, 820, 655, 526, 423,
1325 /* 5 */ 335, 272, 215, 172, 137,
1326 /* 10 */ 110, 87, 70, 56, 45,
1327 /* 15 */ 36, 29, 23, 18, 15,
1331 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1333 * In cases where the weight does not change often, we can use the
1334 * precalculated inverse to speed up arithmetics by turning divisions
1335 * into multiplications:
1337 static const u32 prio_to_wmult[40] = {
1338 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1339 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1340 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1341 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1342 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1343 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1344 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1345 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1348 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1351 * runqueue iterator, to support SMP load-balancing between different
1352 * scheduling classes, without having to expose their internal data
1353 * structures to the load-balancing proper:
1355 struct rq_iterator {
1357 struct task_struct *(*start)(void *);
1358 struct task_struct *(*next)(void *);
1362 static unsigned long
1363 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1364 unsigned long max_load_move, struct sched_domain *sd,
1365 enum cpu_idle_type idle, int *all_pinned,
1366 int *this_best_prio, struct rq_iterator *iterator);
1369 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1370 struct sched_domain *sd, enum cpu_idle_type idle,
1371 struct rq_iterator *iterator);
1374 #ifdef CONFIG_CGROUP_CPUACCT
1375 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1377 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1380 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1382 update_load_add(&rq->load, load);
1385 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1387 update_load_sub(&rq->load, load);
1391 static unsigned long source_load(int cpu, int type);
1392 static unsigned long target_load(int cpu, int type);
1393 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1395 static unsigned long cpu_avg_load_per_task(int cpu)
1397 struct rq *rq = cpu_rq(cpu);
1400 rq->avg_load_per_task = rq->load.weight / rq->nr_running;
1402 return rq->avg_load_per_task;
1405 #ifdef CONFIG_FAIR_GROUP_SCHED
1407 typedef void (*tg_visitor)(struct task_group *, int, struct sched_domain *);
1410 * Iterate the full tree, calling @down when first entering a node and @up when
1411 * leaving it for the final time.
1414 walk_tg_tree(tg_visitor down, tg_visitor up, int cpu, struct sched_domain *sd)
1416 struct task_group *parent, *child;
1419 parent = &root_task_group;
1421 (*down)(parent, cpu, sd);
1422 list_for_each_entry_rcu(child, &parent->children, siblings) {
1429 (*up)(parent, cpu, sd);
1432 parent = parent->parent;
1438 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1441 * Calculate and set the cpu's group shares.
1444 __update_group_shares_cpu(struct task_group *tg, int cpu,
1445 unsigned long sd_shares, unsigned long sd_rq_weight)
1448 unsigned long shares;
1449 unsigned long rq_weight;
1454 rq_weight = tg->cfs_rq[cpu]->load.weight;
1457 * If there are currently no tasks on the cpu pretend there is one of
1458 * average load so that when a new task gets to run here it will not
1459 * get delayed by group starvation.
1463 rq_weight = NICE_0_LOAD;
1466 if (unlikely(rq_weight > sd_rq_weight))
1467 rq_weight = sd_rq_weight;
1470 * \Sum shares * rq_weight
1471 * shares = -----------------------
1475 shares = (sd_shares * rq_weight) / (sd_rq_weight + 1);
1478 * record the actual number of shares, not the boosted amount.
1480 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1481 tg->cfs_rq[cpu]->rq_weight = rq_weight;
1483 if (shares < MIN_SHARES)
1484 shares = MIN_SHARES;
1485 else if (shares > MAX_SHARES)
1486 shares = MAX_SHARES;
1488 __set_se_shares(tg->se[cpu], shares);
1492 * Re-compute the task group their per cpu shares over the given domain.
1493 * This needs to be done in a bottom-up fashion because the rq weight of a
1494 * parent group depends on the shares of its child groups.
1497 tg_shares_up(struct task_group *tg, int cpu, struct sched_domain *sd)
1499 unsigned long rq_weight = 0;
1500 unsigned long shares = 0;
1503 for_each_cpu_mask(i, sd->span) {
1504 rq_weight += tg->cfs_rq[i]->load.weight;
1505 shares += tg->cfs_rq[i]->shares;
1508 if ((!shares && rq_weight) || shares > tg->shares)
1509 shares = tg->shares;
1511 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1512 shares = tg->shares;
1515 rq_weight = cpus_weight(sd->span) * NICE_0_LOAD;
1517 for_each_cpu_mask(i, sd->span) {
1518 struct rq *rq = cpu_rq(i);
1519 unsigned long flags;
1521 spin_lock_irqsave(&rq->lock, flags);
1522 __update_group_shares_cpu(tg, i, shares, rq_weight);
1523 spin_unlock_irqrestore(&rq->lock, flags);
1528 * Compute the cpu's hierarchical load factor for each task group.
1529 * This needs to be done in a top-down fashion because the load of a child
1530 * group is a fraction of its parents load.
1533 tg_load_down(struct task_group *tg, int cpu, struct sched_domain *sd)
1538 load = cpu_rq(cpu)->load.weight;
1540 load = tg->parent->cfs_rq[cpu]->h_load;
1541 load *= tg->cfs_rq[cpu]->shares;
1542 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1545 tg->cfs_rq[cpu]->h_load = load;
1549 tg_nop(struct task_group *tg, int cpu, struct sched_domain *sd)
1553 static void update_shares(struct sched_domain *sd)
1555 u64 now = cpu_clock(raw_smp_processor_id());
1556 s64 elapsed = now - sd->last_update;
1558 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1559 sd->last_update = now;
1560 walk_tg_tree(tg_nop, tg_shares_up, 0, sd);
1564 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1566 spin_unlock(&rq->lock);
1568 spin_lock(&rq->lock);
1571 static void update_h_load(int cpu)
1573 walk_tg_tree(tg_load_down, tg_nop, cpu, NULL);
1578 static inline void update_shares(struct sched_domain *sd)
1582 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1590 #ifdef CONFIG_FAIR_GROUP_SCHED
1591 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1594 cfs_rq->shares = shares;
1599 #include "sched_stats.h"
1600 #include "sched_idletask.c"
1601 #include "sched_fair.c"
1602 #include "sched_rt.c"
1603 #ifdef CONFIG_SCHED_DEBUG
1604 # include "sched_debug.c"
1607 #define sched_class_highest (&rt_sched_class)
1608 #define for_each_class(class) \
1609 for (class = sched_class_highest; class; class = class->next)
1611 static void inc_nr_running(struct rq *rq)
1616 static void dec_nr_running(struct rq *rq)
1621 static void set_load_weight(struct task_struct *p)
1623 if (task_has_rt_policy(p)) {
1624 p->se.load.weight = prio_to_weight[0] * 2;
1625 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1630 * SCHED_IDLE tasks get minimal weight:
1632 if (p->policy == SCHED_IDLE) {
1633 p->se.load.weight = WEIGHT_IDLEPRIO;
1634 p->se.load.inv_weight = WMULT_IDLEPRIO;
1638 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1639 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1642 static void update_avg(u64 *avg, u64 sample)
1644 s64 diff = sample - *avg;
1648 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1650 sched_info_queued(p);
1651 p->sched_class->enqueue_task(rq, p, wakeup);
1655 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1657 if (sleep && p->se.last_wakeup) {
1658 update_avg(&p->se.avg_overlap,
1659 p->se.sum_exec_runtime - p->se.last_wakeup);
1660 p->se.last_wakeup = 0;
1663 sched_info_dequeued(p);
1664 p->sched_class->dequeue_task(rq, p, sleep);
1669 * __normal_prio - return the priority that is based on the static prio
1671 static inline int __normal_prio(struct task_struct *p)
1673 return p->static_prio;
1677 * Calculate the expected normal priority: i.e. priority
1678 * without taking RT-inheritance into account. Might be
1679 * boosted by interactivity modifiers. Changes upon fork,
1680 * setprio syscalls, and whenever the interactivity
1681 * estimator recalculates.
1683 static inline int normal_prio(struct task_struct *p)
1687 if (task_has_rt_policy(p))
1688 prio = MAX_RT_PRIO-1 - p->rt_priority;
1690 prio = __normal_prio(p);
1695 * Calculate the current priority, i.e. the priority
1696 * taken into account by the scheduler. This value might
1697 * be boosted by RT tasks, or might be boosted by
1698 * interactivity modifiers. Will be RT if the task got
1699 * RT-boosted. If not then it returns p->normal_prio.
1701 static int effective_prio(struct task_struct *p)
1703 p->normal_prio = normal_prio(p);
1705 * If we are RT tasks or we were boosted to RT priority,
1706 * keep the priority unchanged. Otherwise, update priority
1707 * to the normal priority:
1709 if (!rt_prio(p->prio))
1710 return p->normal_prio;
1715 * activate_task - move a task to the runqueue.
1717 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1719 if (task_contributes_to_load(p))
1720 rq->nr_uninterruptible--;
1722 enqueue_task(rq, p, wakeup);
1727 * deactivate_task - remove a task from the runqueue.
1729 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1731 if (task_contributes_to_load(p))
1732 rq->nr_uninterruptible++;
1734 dequeue_task(rq, p, sleep);
1739 * task_curr - is this task currently executing on a CPU?
1740 * @p: the task in question.
1742 inline int task_curr(const struct task_struct *p)
1744 return cpu_curr(task_cpu(p)) == p;
1747 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1749 set_task_rq(p, cpu);
1752 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1753 * successfuly executed on another CPU. We must ensure that updates of
1754 * per-task data have been completed by this moment.
1757 task_thread_info(p)->cpu = cpu;
1761 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1762 const struct sched_class *prev_class,
1763 int oldprio, int running)
1765 if (prev_class != p->sched_class) {
1766 if (prev_class->switched_from)
1767 prev_class->switched_from(rq, p, running);
1768 p->sched_class->switched_to(rq, p, running);
1770 p->sched_class->prio_changed(rq, p, oldprio, running);
1775 /* Used instead of source_load when we know the type == 0 */
1776 static unsigned long weighted_cpuload(const int cpu)
1778 return cpu_rq(cpu)->load.weight;
1782 * Is this task likely cache-hot:
1785 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1790 * Buddy candidates are cache hot:
1792 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
1795 if (p->sched_class != &fair_sched_class)
1798 if (sysctl_sched_migration_cost == -1)
1800 if (sysctl_sched_migration_cost == 0)
1803 delta = now - p->se.exec_start;
1805 return delta < (s64)sysctl_sched_migration_cost;
1809 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1811 int old_cpu = task_cpu(p);
1812 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1813 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1814 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1817 clock_offset = old_rq->clock - new_rq->clock;
1819 #ifdef CONFIG_SCHEDSTATS
1820 if (p->se.wait_start)
1821 p->se.wait_start -= clock_offset;
1822 if (p->se.sleep_start)
1823 p->se.sleep_start -= clock_offset;
1824 if (p->se.block_start)
1825 p->se.block_start -= clock_offset;
1826 if (old_cpu != new_cpu) {
1827 schedstat_inc(p, se.nr_migrations);
1828 if (task_hot(p, old_rq->clock, NULL))
1829 schedstat_inc(p, se.nr_forced2_migrations);
1832 p->se.vruntime -= old_cfsrq->min_vruntime -
1833 new_cfsrq->min_vruntime;
1835 __set_task_cpu(p, new_cpu);
1838 struct migration_req {
1839 struct list_head list;
1841 struct task_struct *task;
1844 struct completion done;
1848 * The task's runqueue lock must be held.
1849 * Returns true if you have to wait for migration thread.
1852 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1854 struct rq *rq = task_rq(p);
1857 * If the task is not on a runqueue (and not running), then
1858 * it is sufficient to simply update the task's cpu field.
1860 if (!p->se.on_rq && !task_running(rq, p)) {
1861 set_task_cpu(p, dest_cpu);
1865 init_completion(&req->done);
1867 req->dest_cpu = dest_cpu;
1868 list_add(&req->list, &rq->migration_queue);
1874 * wait_task_inactive - wait for a thread to unschedule.
1876 * If @match_state is nonzero, it's the @p->state value just checked and
1877 * not expected to change. If it changes, i.e. @p might have woken up,
1878 * then return zero. When we succeed in waiting for @p to be off its CPU,
1879 * we return a positive number (its total switch count). If a second call
1880 * a short while later returns the same number, the caller can be sure that
1881 * @p has remained unscheduled the whole time.
1883 * The caller must ensure that the task *will* unschedule sometime soon,
1884 * else this function might spin for a *long* time. This function can't
1885 * be called with interrupts off, or it may introduce deadlock with
1886 * smp_call_function() if an IPI is sent by the same process we are
1887 * waiting to become inactive.
1889 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1891 unsigned long flags;
1898 * We do the initial early heuristics without holding
1899 * any task-queue locks at all. We'll only try to get
1900 * the runqueue lock when things look like they will
1906 * If the task is actively running on another CPU
1907 * still, just relax and busy-wait without holding
1910 * NOTE! Since we don't hold any locks, it's not
1911 * even sure that "rq" stays as the right runqueue!
1912 * But we don't care, since "task_running()" will
1913 * return false if the runqueue has changed and p
1914 * is actually now running somewhere else!
1916 while (task_running(rq, p)) {
1917 if (match_state && unlikely(p->state != match_state))
1923 * Ok, time to look more closely! We need the rq
1924 * lock now, to be *sure*. If we're wrong, we'll
1925 * just go back and repeat.
1927 rq = task_rq_lock(p, &flags);
1928 running = task_running(rq, p);
1929 on_rq = p->se.on_rq;
1931 if (!match_state || p->state == match_state) {
1932 ncsw = p->nivcsw + p->nvcsw;
1933 if (unlikely(!ncsw))
1936 task_rq_unlock(rq, &flags);
1939 * If it changed from the expected state, bail out now.
1941 if (unlikely(!ncsw))
1945 * Was it really running after all now that we
1946 * checked with the proper locks actually held?
1948 * Oops. Go back and try again..
1950 if (unlikely(running)) {
1956 * It's not enough that it's not actively running,
1957 * it must be off the runqueue _entirely_, and not
1960 * So if it wa still runnable (but just not actively
1961 * running right now), it's preempted, and we should
1962 * yield - it could be a while.
1964 if (unlikely(on_rq)) {
1965 schedule_timeout_uninterruptible(1);
1970 * Ahh, all good. It wasn't running, and it wasn't
1971 * runnable, which means that it will never become
1972 * running in the future either. We're all done!
1981 * kick_process - kick a running thread to enter/exit the kernel
1982 * @p: the to-be-kicked thread
1984 * Cause a process which is running on another CPU to enter
1985 * kernel-mode, without any delay. (to get signals handled.)
1987 * NOTE: this function doesnt have to take the runqueue lock,
1988 * because all it wants to ensure is that the remote task enters
1989 * the kernel. If the IPI races and the task has been migrated
1990 * to another CPU then no harm is done and the purpose has been
1993 void kick_process(struct task_struct *p)
1999 if ((cpu != smp_processor_id()) && task_curr(p))
2000 smp_send_reschedule(cpu);
2005 * Return a low guess at the load of a migration-source cpu weighted
2006 * according to the scheduling class and "nice" value.
2008 * We want to under-estimate the load of migration sources, to
2009 * balance conservatively.
2011 static unsigned long source_load(int cpu, int type)
2013 struct rq *rq = cpu_rq(cpu);
2014 unsigned long total = weighted_cpuload(cpu);
2016 if (type == 0 || !sched_feat(LB_BIAS))
2019 return min(rq->cpu_load[type-1], total);
2023 * Return a high guess at the load of a migration-target cpu weighted
2024 * according to the scheduling class and "nice" value.
2026 static unsigned long target_load(int cpu, int type)
2028 struct rq *rq = cpu_rq(cpu);
2029 unsigned long total = weighted_cpuload(cpu);
2031 if (type == 0 || !sched_feat(LB_BIAS))
2034 return max(rq->cpu_load[type-1], total);
2038 * find_idlest_group finds and returns the least busy CPU group within the
2041 static struct sched_group *
2042 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2044 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2045 unsigned long min_load = ULONG_MAX, this_load = 0;
2046 int load_idx = sd->forkexec_idx;
2047 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2050 unsigned long load, avg_load;
2054 /* Skip over this group if it has no CPUs allowed */
2055 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
2058 local_group = cpu_isset(this_cpu, group->cpumask);
2060 /* Tally up the load of all CPUs in the group */
2063 for_each_cpu_mask_nr(i, group->cpumask) {
2064 /* Bias balancing toward cpus of our domain */
2066 load = source_load(i, load_idx);
2068 load = target_load(i, load_idx);
2073 /* Adjust by relative CPU power of the group */
2074 avg_load = sg_div_cpu_power(group,
2075 avg_load * SCHED_LOAD_SCALE);
2078 this_load = avg_load;
2080 } else if (avg_load < min_load) {
2081 min_load = avg_load;
2084 } while (group = group->next, group != sd->groups);
2086 if (!idlest || 100*this_load < imbalance*min_load)
2092 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2095 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2098 unsigned long load, min_load = ULONG_MAX;
2102 /* Traverse only the allowed CPUs */
2103 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
2105 for_each_cpu_mask_nr(i, *tmp) {
2106 load = weighted_cpuload(i);
2108 if (load < min_load || (load == min_load && i == this_cpu)) {
2118 * sched_balance_self: balance the current task (running on cpu) in domains
2119 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2122 * Balance, ie. select the least loaded group.
2124 * Returns the target CPU number, or the same CPU if no balancing is needed.
2126 * preempt must be disabled.
2128 static int sched_balance_self(int cpu, int flag)
2130 struct task_struct *t = current;
2131 struct sched_domain *tmp, *sd = NULL;
2133 for_each_domain(cpu, tmp) {
2135 * If power savings logic is enabled for a domain, stop there.
2137 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2139 if (tmp->flags & flag)
2147 cpumask_t span, tmpmask;
2148 struct sched_group *group;
2149 int new_cpu, weight;
2151 if (!(sd->flags & flag)) {
2157 group = find_idlest_group(sd, t, cpu);
2163 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2164 if (new_cpu == -1 || new_cpu == cpu) {
2165 /* Now try balancing at a lower domain level of cpu */
2170 /* Now try balancing at a lower domain level of new_cpu */
2173 weight = cpus_weight(span);
2174 for_each_domain(cpu, tmp) {
2175 if (weight <= cpus_weight(tmp->span))
2177 if (tmp->flags & flag)
2180 /* while loop will break here if sd == NULL */
2186 #endif /* CONFIG_SMP */
2189 * try_to_wake_up - wake up a thread
2190 * @p: the to-be-woken-up thread
2191 * @state: the mask of task states that can be woken
2192 * @sync: do a synchronous wakeup?
2194 * Put it on the run-queue if it's not already there. The "current"
2195 * thread is always on the run-queue (except when the actual
2196 * re-schedule is in progress), and as such you're allowed to do
2197 * the simpler "current->state = TASK_RUNNING" to mark yourself
2198 * runnable without the overhead of this.
2200 * returns failure only if the task is already active.
2202 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2204 int cpu, orig_cpu, this_cpu, success = 0;
2205 unsigned long flags;
2209 if (!sched_feat(SYNC_WAKEUPS))
2213 if (sched_feat(LB_WAKEUP_UPDATE)) {
2214 struct sched_domain *sd;
2216 this_cpu = raw_smp_processor_id();
2219 for_each_domain(this_cpu, sd) {
2220 if (cpu_isset(cpu, sd->span)) {
2229 rq = task_rq_lock(p, &flags);
2230 old_state = p->state;
2231 if (!(old_state & state))
2239 this_cpu = smp_processor_id();
2242 if (unlikely(task_running(rq, p)))
2245 cpu = p->sched_class->select_task_rq(p, sync);
2246 if (cpu != orig_cpu) {
2247 set_task_cpu(p, cpu);
2248 task_rq_unlock(rq, &flags);
2249 /* might preempt at this point */
2250 rq = task_rq_lock(p, &flags);
2251 old_state = p->state;
2252 if (!(old_state & state))
2257 this_cpu = smp_processor_id();
2261 #ifdef CONFIG_SCHEDSTATS
2262 schedstat_inc(rq, ttwu_count);
2263 if (cpu == this_cpu)
2264 schedstat_inc(rq, ttwu_local);
2266 struct sched_domain *sd;
2267 for_each_domain(this_cpu, sd) {
2268 if (cpu_isset(cpu, sd->span)) {
2269 schedstat_inc(sd, ttwu_wake_remote);
2274 #endif /* CONFIG_SCHEDSTATS */
2277 #endif /* CONFIG_SMP */
2278 schedstat_inc(p, se.nr_wakeups);
2280 schedstat_inc(p, se.nr_wakeups_sync);
2281 if (orig_cpu != cpu)
2282 schedstat_inc(p, se.nr_wakeups_migrate);
2283 if (cpu == this_cpu)
2284 schedstat_inc(p, se.nr_wakeups_local);
2286 schedstat_inc(p, se.nr_wakeups_remote);
2287 update_rq_clock(rq);
2288 activate_task(rq, p, 1);
2292 trace_mark(kernel_sched_wakeup,
2293 "pid %d state %ld ## rq %p task %p rq->curr %p",
2294 p->pid, p->state, rq, p, rq->curr);
2295 check_preempt_curr(rq, p);
2297 p->state = TASK_RUNNING;
2299 if (p->sched_class->task_wake_up)
2300 p->sched_class->task_wake_up(rq, p);
2303 current->se.last_wakeup = current->se.sum_exec_runtime;
2305 task_rq_unlock(rq, &flags);
2310 int wake_up_process(struct task_struct *p)
2312 return try_to_wake_up(p, TASK_ALL, 0);
2314 EXPORT_SYMBOL(wake_up_process);
2316 int wake_up_state(struct task_struct *p, unsigned int state)
2318 return try_to_wake_up(p, state, 0);
2322 * Perform scheduler related setup for a newly forked process p.
2323 * p is forked by current.
2325 * __sched_fork() is basic setup used by init_idle() too:
2327 static void __sched_fork(struct task_struct *p)
2329 p->se.exec_start = 0;
2330 p->se.sum_exec_runtime = 0;
2331 p->se.prev_sum_exec_runtime = 0;
2332 p->se.last_wakeup = 0;
2333 p->se.avg_overlap = 0;
2335 #ifdef CONFIG_SCHEDSTATS
2336 p->se.wait_start = 0;
2337 p->se.sum_sleep_runtime = 0;
2338 p->se.sleep_start = 0;
2339 p->se.block_start = 0;
2340 p->se.sleep_max = 0;
2341 p->se.block_max = 0;
2343 p->se.slice_max = 0;
2347 INIT_LIST_HEAD(&p->rt.run_list);
2349 INIT_LIST_HEAD(&p->se.group_node);
2351 #ifdef CONFIG_PREEMPT_NOTIFIERS
2352 INIT_HLIST_HEAD(&p->preempt_notifiers);
2356 * We mark the process as running here, but have not actually
2357 * inserted it onto the runqueue yet. This guarantees that
2358 * nobody will actually run it, and a signal or other external
2359 * event cannot wake it up and insert it on the runqueue either.
2361 p->state = TASK_RUNNING;
2365 * fork()/clone()-time setup:
2367 void sched_fork(struct task_struct *p, int clone_flags)
2369 int cpu = get_cpu();
2374 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2376 set_task_cpu(p, cpu);
2379 * Make sure we do not leak PI boosting priority to the child:
2381 p->prio = current->normal_prio;
2382 if (!rt_prio(p->prio))
2383 p->sched_class = &fair_sched_class;
2385 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2386 if (likely(sched_info_on()))
2387 memset(&p->sched_info, 0, sizeof(p->sched_info));
2389 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2392 #ifdef CONFIG_PREEMPT
2393 /* Want to start with kernel preemption disabled. */
2394 task_thread_info(p)->preempt_count = 1;
2400 * wake_up_new_task - wake up a newly created task for the first time.
2402 * This function will do some initial scheduler statistics housekeeping
2403 * that must be done for every newly created context, then puts the task
2404 * on the runqueue and wakes it.
2406 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2408 unsigned long flags;
2411 rq = task_rq_lock(p, &flags);
2412 BUG_ON(p->state != TASK_RUNNING);
2413 update_rq_clock(rq);
2415 p->prio = effective_prio(p);
2417 if (!p->sched_class->task_new || !current->se.on_rq) {
2418 activate_task(rq, p, 0);
2421 * Let the scheduling class do new task startup
2422 * management (if any):
2424 p->sched_class->task_new(rq, p);
2427 trace_mark(kernel_sched_wakeup_new,
2428 "pid %d state %ld ## rq %p task %p rq->curr %p",
2429 p->pid, p->state, rq, p, rq->curr);
2430 check_preempt_curr(rq, p);
2432 if (p->sched_class->task_wake_up)
2433 p->sched_class->task_wake_up(rq, p);
2435 task_rq_unlock(rq, &flags);
2438 #ifdef CONFIG_PREEMPT_NOTIFIERS
2441 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2442 * @notifier: notifier struct to register
2444 void preempt_notifier_register(struct preempt_notifier *notifier)
2446 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2448 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2451 * preempt_notifier_unregister - no longer interested in preemption notifications
2452 * @notifier: notifier struct to unregister
2454 * This is safe to call from within a preemption notifier.
2456 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2458 hlist_del(¬ifier->link);
2460 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2462 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2464 struct preempt_notifier *notifier;
2465 struct hlist_node *node;
2467 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2468 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2472 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2473 struct task_struct *next)
2475 struct preempt_notifier *notifier;
2476 struct hlist_node *node;
2478 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2479 notifier->ops->sched_out(notifier, next);
2482 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2484 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2489 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2490 struct task_struct *next)
2494 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2497 * prepare_task_switch - prepare to switch tasks
2498 * @rq: the runqueue preparing to switch
2499 * @prev: the current task that is being switched out
2500 * @next: the task we are going to switch to.
2502 * This is called with the rq lock held and interrupts off. It must
2503 * be paired with a subsequent finish_task_switch after the context
2506 * prepare_task_switch sets up locking and calls architecture specific
2510 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2511 struct task_struct *next)
2513 fire_sched_out_preempt_notifiers(prev, next);
2514 prepare_lock_switch(rq, next);
2515 prepare_arch_switch(next);
2519 * finish_task_switch - clean up after a task-switch
2520 * @rq: runqueue associated with task-switch
2521 * @prev: the thread we just switched away from.
2523 * finish_task_switch must be called after the context switch, paired
2524 * with a prepare_task_switch call before the context switch.
2525 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2526 * and do any other architecture-specific cleanup actions.
2528 * Note that we may have delayed dropping an mm in context_switch(). If
2529 * so, we finish that here outside of the runqueue lock. (Doing it
2530 * with the lock held can cause deadlocks; see schedule() for
2533 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2534 __releases(rq->lock)
2536 struct mm_struct *mm = rq->prev_mm;
2542 * A task struct has one reference for the use as "current".
2543 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2544 * schedule one last time. The schedule call will never return, and
2545 * the scheduled task must drop that reference.
2546 * The test for TASK_DEAD must occur while the runqueue locks are
2547 * still held, otherwise prev could be scheduled on another cpu, die
2548 * there before we look at prev->state, and then the reference would
2550 * Manfred Spraul <manfred@colorfullife.com>
2552 prev_state = prev->state;
2553 finish_arch_switch(prev);
2554 finish_lock_switch(rq, prev);
2556 if (current->sched_class->post_schedule)
2557 current->sched_class->post_schedule(rq);
2560 fire_sched_in_preempt_notifiers(current);
2563 if (unlikely(prev_state == TASK_DEAD)) {
2565 * Remove function-return probe instances associated with this
2566 * task and put them back on the free list.
2568 kprobe_flush_task(prev);
2569 put_task_struct(prev);
2574 * schedule_tail - first thing a freshly forked thread must call.
2575 * @prev: the thread we just switched away from.
2577 asmlinkage void schedule_tail(struct task_struct *prev)
2578 __releases(rq->lock)
2580 struct rq *rq = this_rq();
2582 finish_task_switch(rq, prev);
2583 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2584 /* In this case, finish_task_switch does not reenable preemption */
2587 if (current->set_child_tid)
2588 put_user(task_pid_vnr(current), current->set_child_tid);
2592 * context_switch - switch to the new MM and the new
2593 * thread's register state.
2596 context_switch(struct rq *rq, struct task_struct *prev,
2597 struct task_struct *next)
2599 struct mm_struct *mm, *oldmm;
2601 prepare_task_switch(rq, prev, next);
2602 trace_mark(kernel_sched_schedule,
2603 "prev_pid %d next_pid %d prev_state %ld "
2604 "## rq %p prev %p next %p",
2605 prev->pid, next->pid, prev->state,
2608 oldmm = prev->active_mm;
2610 * For paravirt, this is coupled with an exit in switch_to to
2611 * combine the page table reload and the switch backend into
2614 arch_enter_lazy_cpu_mode();
2616 if (unlikely(!mm)) {
2617 next->active_mm = oldmm;
2618 atomic_inc(&oldmm->mm_count);
2619 enter_lazy_tlb(oldmm, next);
2621 switch_mm(oldmm, mm, next);
2623 if (unlikely(!prev->mm)) {
2624 prev->active_mm = NULL;
2625 rq->prev_mm = oldmm;
2628 * Since the runqueue lock will be released by the next
2629 * task (which is an invalid locking op but in the case
2630 * of the scheduler it's an obvious special-case), so we
2631 * do an early lockdep release here:
2633 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2634 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2637 /* Here we just switch the register state and the stack. */
2638 switch_to(prev, next, prev);
2642 * this_rq must be evaluated again because prev may have moved
2643 * CPUs since it called schedule(), thus the 'rq' on its stack
2644 * frame will be invalid.
2646 finish_task_switch(this_rq(), prev);
2650 * nr_running, nr_uninterruptible and nr_context_switches:
2652 * externally visible scheduler statistics: current number of runnable
2653 * threads, current number of uninterruptible-sleeping threads, total
2654 * number of context switches performed since bootup.
2656 unsigned long nr_running(void)
2658 unsigned long i, sum = 0;
2660 for_each_online_cpu(i)
2661 sum += cpu_rq(i)->nr_running;
2666 unsigned long nr_uninterruptible(void)
2668 unsigned long i, sum = 0;
2670 for_each_possible_cpu(i)
2671 sum += cpu_rq(i)->nr_uninterruptible;
2674 * Since we read the counters lockless, it might be slightly
2675 * inaccurate. Do not allow it to go below zero though:
2677 if (unlikely((long)sum < 0))
2683 unsigned long long nr_context_switches(void)
2686 unsigned long long sum = 0;
2688 for_each_possible_cpu(i)
2689 sum += cpu_rq(i)->nr_switches;
2694 unsigned long nr_iowait(void)
2696 unsigned long i, sum = 0;
2698 for_each_possible_cpu(i)
2699 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2704 unsigned long nr_active(void)
2706 unsigned long i, running = 0, uninterruptible = 0;
2708 for_each_online_cpu(i) {
2709 running += cpu_rq(i)->nr_running;
2710 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2713 if (unlikely((long)uninterruptible < 0))
2714 uninterruptible = 0;
2716 return running + uninterruptible;
2720 * Update rq->cpu_load[] statistics. This function is usually called every
2721 * scheduler tick (TICK_NSEC).
2723 static void update_cpu_load(struct rq *this_rq)
2725 unsigned long this_load = this_rq->load.weight;
2728 this_rq->nr_load_updates++;
2730 /* Update our load: */
2731 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2732 unsigned long old_load, new_load;
2734 /* scale is effectively 1 << i now, and >> i divides by scale */
2736 old_load = this_rq->cpu_load[i];
2737 new_load = this_load;
2739 * Round up the averaging division if load is increasing. This
2740 * prevents us from getting stuck on 9 if the load is 10, for
2743 if (new_load > old_load)
2744 new_load += scale-1;
2745 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2752 * double_rq_lock - safely lock two runqueues
2754 * Note this does not disable interrupts like task_rq_lock,
2755 * you need to do so manually before calling.
2757 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2758 __acquires(rq1->lock)
2759 __acquires(rq2->lock)
2761 BUG_ON(!irqs_disabled());
2763 spin_lock(&rq1->lock);
2764 __acquire(rq2->lock); /* Fake it out ;) */
2767 spin_lock(&rq1->lock);
2768 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2770 spin_lock(&rq2->lock);
2771 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2774 update_rq_clock(rq1);
2775 update_rq_clock(rq2);
2779 * double_rq_unlock - safely unlock two runqueues
2781 * Note this does not restore interrupts like task_rq_unlock,
2782 * you need to do so manually after calling.
2784 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2785 __releases(rq1->lock)
2786 __releases(rq2->lock)
2788 spin_unlock(&rq1->lock);
2790 spin_unlock(&rq2->lock);
2792 __release(rq2->lock);
2796 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2798 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2799 __releases(this_rq->lock)
2800 __acquires(busiest->lock)
2801 __acquires(this_rq->lock)
2805 if (unlikely(!irqs_disabled())) {
2806 /* printk() doesn't work good under rq->lock */
2807 spin_unlock(&this_rq->lock);
2810 if (unlikely(!spin_trylock(&busiest->lock))) {
2811 if (busiest < this_rq) {
2812 spin_unlock(&this_rq->lock);
2813 spin_lock(&busiest->lock);
2814 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
2817 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
2822 static void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
2823 __releases(busiest->lock)
2825 spin_unlock(&busiest->lock);
2826 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
2830 * If dest_cpu is allowed for this process, migrate the task to it.
2831 * This is accomplished by forcing the cpu_allowed mask to only
2832 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2833 * the cpu_allowed mask is restored.
2835 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2837 struct migration_req req;
2838 unsigned long flags;
2841 rq = task_rq_lock(p, &flags);
2842 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2843 || unlikely(!cpu_active(dest_cpu)))
2846 /* force the process onto the specified CPU */
2847 if (migrate_task(p, dest_cpu, &req)) {
2848 /* Need to wait for migration thread (might exit: take ref). */
2849 struct task_struct *mt = rq->migration_thread;
2851 get_task_struct(mt);
2852 task_rq_unlock(rq, &flags);
2853 wake_up_process(mt);
2854 put_task_struct(mt);
2855 wait_for_completion(&req.done);
2860 task_rq_unlock(rq, &flags);
2864 * sched_exec - execve() is a valuable balancing opportunity, because at
2865 * this point the task has the smallest effective memory and cache footprint.
2867 void sched_exec(void)
2869 int new_cpu, this_cpu = get_cpu();
2870 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2872 if (new_cpu != this_cpu)
2873 sched_migrate_task(current, new_cpu);
2877 * pull_task - move a task from a remote runqueue to the local runqueue.
2878 * Both runqueues must be locked.
2880 static void pull_task(struct rq *src_rq, struct task_struct *p,
2881 struct rq *this_rq, int this_cpu)
2883 deactivate_task(src_rq, p, 0);
2884 set_task_cpu(p, this_cpu);
2885 activate_task(this_rq, p, 0);
2887 * Note that idle threads have a prio of MAX_PRIO, for this test
2888 * to be always true for them.
2890 check_preempt_curr(this_rq, p);
2894 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2897 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2898 struct sched_domain *sd, enum cpu_idle_type idle,
2902 * We do not migrate tasks that are:
2903 * 1) running (obviously), or
2904 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2905 * 3) are cache-hot on their current CPU.
2907 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2908 schedstat_inc(p, se.nr_failed_migrations_affine);
2913 if (task_running(rq, p)) {
2914 schedstat_inc(p, se.nr_failed_migrations_running);
2919 * Aggressive migration if:
2920 * 1) task is cache cold, or
2921 * 2) too many balance attempts have failed.
2924 if (!task_hot(p, rq->clock, sd) ||
2925 sd->nr_balance_failed > sd->cache_nice_tries) {
2926 #ifdef CONFIG_SCHEDSTATS
2927 if (task_hot(p, rq->clock, sd)) {
2928 schedstat_inc(sd, lb_hot_gained[idle]);
2929 schedstat_inc(p, se.nr_forced_migrations);
2935 if (task_hot(p, rq->clock, sd)) {
2936 schedstat_inc(p, se.nr_failed_migrations_hot);
2942 static unsigned long
2943 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2944 unsigned long max_load_move, struct sched_domain *sd,
2945 enum cpu_idle_type idle, int *all_pinned,
2946 int *this_best_prio, struct rq_iterator *iterator)
2948 int loops = 0, pulled = 0, pinned = 0;
2949 struct task_struct *p;
2950 long rem_load_move = max_load_move;
2952 if (max_load_move == 0)
2958 * Start the load-balancing iterator:
2960 p = iterator->start(iterator->arg);
2962 if (!p || loops++ > sysctl_sched_nr_migrate)
2965 if ((p->se.load.weight >> 1) > rem_load_move ||
2966 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2967 p = iterator->next(iterator->arg);
2971 pull_task(busiest, p, this_rq, this_cpu);
2973 rem_load_move -= p->se.load.weight;
2976 * We only want to steal up to the prescribed amount of weighted load.
2978 if (rem_load_move > 0) {
2979 if (p->prio < *this_best_prio)
2980 *this_best_prio = p->prio;
2981 p = iterator->next(iterator->arg);
2986 * Right now, this is one of only two places pull_task() is called,
2987 * so we can safely collect pull_task() stats here rather than
2988 * inside pull_task().
2990 schedstat_add(sd, lb_gained[idle], pulled);
2993 *all_pinned = pinned;
2995 return max_load_move - rem_load_move;
2999 * move_tasks tries to move up to max_load_move weighted load from busiest to
3000 * this_rq, as part of a balancing operation within domain "sd".
3001 * Returns 1 if successful and 0 otherwise.
3003 * Called with both runqueues locked.
3005 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3006 unsigned long max_load_move,
3007 struct sched_domain *sd, enum cpu_idle_type idle,
3010 const struct sched_class *class = sched_class_highest;
3011 unsigned long total_load_moved = 0;
3012 int this_best_prio = this_rq->curr->prio;
3016 class->load_balance(this_rq, this_cpu, busiest,
3017 max_load_move - total_load_moved,
3018 sd, idle, all_pinned, &this_best_prio);
3019 class = class->next;
3021 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3024 } while (class && max_load_move > total_load_moved);
3026 return total_load_moved > 0;
3030 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3031 struct sched_domain *sd, enum cpu_idle_type idle,
3032 struct rq_iterator *iterator)
3034 struct task_struct *p = iterator->start(iterator->arg);
3038 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3039 pull_task(busiest, p, this_rq, this_cpu);
3041 * Right now, this is only the second place pull_task()
3042 * is called, so we can safely collect pull_task()
3043 * stats here rather than inside pull_task().
3045 schedstat_inc(sd, lb_gained[idle]);
3049 p = iterator->next(iterator->arg);
3056 * move_one_task tries to move exactly one task from busiest to this_rq, as
3057 * part of active balancing operations within "domain".
3058 * Returns 1 if successful and 0 otherwise.
3060 * Called with both runqueues locked.
3062 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3063 struct sched_domain *sd, enum cpu_idle_type idle)
3065 const struct sched_class *class;
3067 for (class = sched_class_highest; class; class = class->next)
3068 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3075 * find_busiest_group finds and returns the busiest CPU group within the
3076 * domain. It calculates and returns the amount of weighted load which
3077 * should be moved to restore balance via the imbalance parameter.
3079 static struct sched_group *
3080 find_busiest_group(struct sched_domain *sd, int this_cpu,
3081 unsigned long *imbalance, enum cpu_idle_type idle,
3082 int *sd_idle, const cpumask_t *cpus, int *balance)
3084 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3085 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3086 unsigned long max_pull;
3087 unsigned long busiest_load_per_task, busiest_nr_running;
3088 unsigned long this_load_per_task, this_nr_running;
3089 int load_idx, group_imb = 0;
3090 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3091 int power_savings_balance = 1;
3092 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3093 unsigned long min_nr_running = ULONG_MAX;
3094 struct sched_group *group_min = NULL, *group_leader = NULL;
3097 max_load = this_load = total_load = total_pwr = 0;
3098 busiest_load_per_task = busiest_nr_running = 0;
3099 this_load_per_task = this_nr_running = 0;
3101 if (idle == CPU_NOT_IDLE)
3102 load_idx = sd->busy_idx;
3103 else if (idle == CPU_NEWLY_IDLE)
3104 load_idx = sd->newidle_idx;
3106 load_idx = sd->idle_idx;
3109 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3112 int __group_imb = 0;
3113 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3114 unsigned long sum_nr_running, sum_weighted_load;
3115 unsigned long sum_avg_load_per_task;
3116 unsigned long avg_load_per_task;
3118 local_group = cpu_isset(this_cpu, group->cpumask);
3121 balance_cpu = first_cpu(group->cpumask);
3123 /* Tally up the load of all CPUs in the group */
3124 sum_weighted_load = sum_nr_running = avg_load = 0;
3125 sum_avg_load_per_task = avg_load_per_task = 0;
3128 min_cpu_load = ~0UL;
3130 for_each_cpu_mask_nr(i, group->cpumask) {
3133 if (!cpu_isset(i, *cpus))
3138 if (*sd_idle && rq->nr_running)
3141 /* Bias balancing toward cpus of our domain */
3143 if (idle_cpu(i) && !first_idle_cpu) {
3148 load = target_load(i, load_idx);
3150 load = source_load(i, load_idx);
3151 if (load > max_cpu_load)
3152 max_cpu_load = load;
3153 if (min_cpu_load > load)
3154 min_cpu_load = load;
3158 sum_nr_running += rq->nr_running;
3159 sum_weighted_load += weighted_cpuload(i);
3161 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3165 * First idle cpu or the first cpu(busiest) in this sched group
3166 * is eligible for doing load balancing at this and above
3167 * domains. In the newly idle case, we will allow all the cpu's
3168 * to do the newly idle load balance.
3170 if (idle != CPU_NEWLY_IDLE && local_group &&
3171 balance_cpu != this_cpu && balance) {
3176 total_load += avg_load;
3177 total_pwr += group->__cpu_power;
3179 /* Adjust by relative CPU power of the group */
3180 avg_load = sg_div_cpu_power(group,
3181 avg_load * SCHED_LOAD_SCALE);
3185 * Consider the group unbalanced when the imbalance is larger
3186 * than the average weight of two tasks.
3188 * APZ: with cgroup the avg task weight can vary wildly and
3189 * might not be a suitable number - should we keep a
3190 * normalized nr_running number somewhere that negates
3193 avg_load_per_task = sg_div_cpu_power(group,
3194 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3196 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3199 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3202 this_load = avg_load;
3204 this_nr_running = sum_nr_running;
3205 this_load_per_task = sum_weighted_load;
3206 } else if (avg_load > max_load &&
3207 (sum_nr_running > group_capacity || __group_imb)) {
3208 max_load = avg_load;
3210 busiest_nr_running = sum_nr_running;
3211 busiest_load_per_task = sum_weighted_load;
3212 group_imb = __group_imb;
3215 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3217 * Busy processors will not participate in power savings
3220 if (idle == CPU_NOT_IDLE ||
3221 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3225 * If the local group is idle or completely loaded
3226 * no need to do power savings balance at this domain
3228 if (local_group && (this_nr_running >= group_capacity ||
3230 power_savings_balance = 0;
3233 * If a group is already running at full capacity or idle,
3234 * don't include that group in power savings calculations
3236 if (!power_savings_balance || sum_nr_running >= group_capacity
3241 * Calculate the group which has the least non-idle load.
3242 * This is the group from where we need to pick up the load
3245 if ((sum_nr_running < min_nr_running) ||
3246 (sum_nr_running == min_nr_running &&
3247 first_cpu(group->cpumask) <
3248 first_cpu(group_min->cpumask))) {
3250 min_nr_running = sum_nr_running;
3251 min_load_per_task = sum_weighted_load /
3256 * Calculate the group which is almost near its
3257 * capacity but still has some space to pick up some load
3258 * from other group and save more power
3260 if (sum_nr_running <= group_capacity - 1) {
3261 if (sum_nr_running > leader_nr_running ||
3262 (sum_nr_running == leader_nr_running &&
3263 first_cpu(group->cpumask) >
3264 first_cpu(group_leader->cpumask))) {
3265 group_leader = group;
3266 leader_nr_running = sum_nr_running;
3271 group = group->next;
3272 } while (group != sd->groups);
3274 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3277 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3279 if (this_load >= avg_load ||
3280 100*max_load <= sd->imbalance_pct*this_load)
3283 busiest_load_per_task /= busiest_nr_running;
3285 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3288 * We're trying to get all the cpus to the average_load, so we don't
3289 * want to push ourselves above the average load, nor do we wish to
3290 * reduce the max loaded cpu below the average load, as either of these
3291 * actions would just result in more rebalancing later, and ping-pong
3292 * tasks around. Thus we look for the minimum possible imbalance.
3293 * Negative imbalances (*we* are more loaded than anyone else) will
3294 * be counted as no imbalance for these purposes -- we can't fix that
3295 * by pulling tasks to us. Be careful of negative numbers as they'll
3296 * appear as very large values with unsigned longs.
3298 if (max_load <= busiest_load_per_task)
3302 * In the presence of smp nice balancing, certain scenarios can have
3303 * max load less than avg load(as we skip the groups at or below
3304 * its cpu_power, while calculating max_load..)
3306 if (max_load < avg_load) {
3308 goto small_imbalance;
3311 /* Don't want to pull so many tasks that a group would go idle */
3312 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3314 /* How much load to actually move to equalise the imbalance */
3315 *imbalance = min(max_pull * busiest->__cpu_power,
3316 (avg_load - this_load) * this->__cpu_power)
3320 * if *imbalance is less than the average load per runnable task
3321 * there is no gaurantee that any tasks will be moved so we'll have
3322 * a think about bumping its value to force at least one task to be
3325 if (*imbalance < busiest_load_per_task) {
3326 unsigned long tmp, pwr_now, pwr_move;
3330 pwr_move = pwr_now = 0;
3332 if (this_nr_running) {
3333 this_load_per_task /= this_nr_running;
3334 if (busiest_load_per_task > this_load_per_task)
3337 this_load_per_task = cpu_avg_load_per_task(this_cpu);
3339 if (max_load - this_load + 2*busiest_load_per_task >=
3340 busiest_load_per_task * imbn) {
3341 *imbalance = busiest_load_per_task;
3346 * OK, we don't have enough imbalance to justify moving tasks,
3347 * however we may be able to increase total CPU power used by
3351 pwr_now += busiest->__cpu_power *
3352 min(busiest_load_per_task, max_load);
3353 pwr_now += this->__cpu_power *
3354 min(this_load_per_task, this_load);
3355 pwr_now /= SCHED_LOAD_SCALE;
3357 /* Amount of load we'd subtract */
3358 tmp = sg_div_cpu_power(busiest,
3359 busiest_load_per_task * SCHED_LOAD_SCALE);
3361 pwr_move += busiest->__cpu_power *
3362 min(busiest_load_per_task, max_load - tmp);
3364 /* Amount of load we'd add */
3365 if (max_load * busiest->__cpu_power <
3366 busiest_load_per_task * SCHED_LOAD_SCALE)
3367 tmp = sg_div_cpu_power(this,
3368 max_load * busiest->__cpu_power);
3370 tmp = sg_div_cpu_power(this,
3371 busiest_load_per_task * SCHED_LOAD_SCALE);
3372 pwr_move += this->__cpu_power *
3373 min(this_load_per_task, this_load + tmp);
3374 pwr_move /= SCHED_LOAD_SCALE;
3376 /* Move if we gain throughput */
3377 if (pwr_move > pwr_now)
3378 *imbalance = busiest_load_per_task;
3384 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3385 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3388 if (this == group_leader && group_leader != group_min) {
3389 *imbalance = min_load_per_task;
3399 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3402 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3403 unsigned long imbalance, const cpumask_t *cpus)
3405 struct rq *busiest = NULL, *rq;
3406 unsigned long max_load = 0;
3409 for_each_cpu_mask_nr(i, group->cpumask) {
3412 if (!cpu_isset(i, *cpus))
3416 wl = weighted_cpuload(i);
3418 if (rq->nr_running == 1 && wl > imbalance)
3421 if (wl > max_load) {
3431 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3432 * so long as it is large enough.
3434 #define MAX_PINNED_INTERVAL 512
3437 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3438 * tasks if there is an imbalance.
3440 static int load_balance(int this_cpu, struct rq *this_rq,
3441 struct sched_domain *sd, enum cpu_idle_type idle,
3442 int *balance, cpumask_t *cpus)
3444 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3445 struct sched_group *group;
3446 unsigned long imbalance;
3448 unsigned long flags;
3453 * When power savings policy is enabled for the parent domain, idle
3454 * sibling can pick up load irrespective of busy siblings. In this case,
3455 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3456 * portraying it as CPU_NOT_IDLE.
3458 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3459 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3462 schedstat_inc(sd, lb_count[idle]);
3466 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3473 schedstat_inc(sd, lb_nobusyg[idle]);
3477 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3479 schedstat_inc(sd, lb_nobusyq[idle]);
3483 BUG_ON(busiest == this_rq);
3485 schedstat_add(sd, lb_imbalance[idle], imbalance);
3488 if (busiest->nr_running > 1) {
3490 * Attempt to move tasks. If find_busiest_group has found
3491 * an imbalance but busiest->nr_running <= 1, the group is
3492 * still unbalanced. ld_moved simply stays zero, so it is
3493 * correctly treated as an imbalance.
3495 local_irq_save(flags);
3496 double_rq_lock(this_rq, busiest);
3497 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3498 imbalance, sd, idle, &all_pinned);
3499 double_rq_unlock(this_rq, busiest);
3500 local_irq_restore(flags);
3503 * some other cpu did the load balance for us.
3505 if (ld_moved && this_cpu != smp_processor_id())
3506 resched_cpu(this_cpu);
3508 /* All tasks on this runqueue were pinned by CPU affinity */
3509 if (unlikely(all_pinned)) {
3510 cpu_clear(cpu_of(busiest), *cpus);
3511 if (!cpus_empty(*cpus))
3518 schedstat_inc(sd, lb_failed[idle]);
3519 sd->nr_balance_failed++;
3521 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3523 spin_lock_irqsave(&busiest->lock, flags);
3525 /* don't kick the migration_thread, if the curr
3526 * task on busiest cpu can't be moved to this_cpu
3528 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3529 spin_unlock_irqrestore(&busiest->lock, flags);
3531 goto out_one_pinned;
3534 if (!busiest->active_balance) {
3535 busiest->active_balance = 1;
3536 busiest->push_cpu = this_cpu;
3539 spin_unlock_irqrestore(&busiest->lock, flags);
3541 wake_up_process(busiest->migration_thread);
3544 * We've kicked active balancing, reset the failure
3547 sd->nr_balance_failed = sd->cache_nice_tries+1;
3550 sd->nr_balance_failed = 0;
3552 if (likely(!active_balance)) {
3553 /* We were unbalanced, so reset the balancing interval */
3554 sd->balance_interval = sd->min_interval;
3557 * If we've begun active balancing, start to back off. This
3558 * case may not be covered by the all_pinned logic if there
3559 * is only 1 task on the busy runqueue (because we don't call
3562 if (sd->balance_interval < sd->max_interval)
3563 sd->balance_interval *= 2;
3566 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3567 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3573 schedstat_inc(sd, lb_balanced[idle]);
3575 sd->nr_balance_failed = 0;
3578 /* tune up the balancing interval */
3579 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3580 (sd->balance_interval < sd->max_interval))
3581 sd->balance_interval *= 2;
3583 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3584 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3595 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3596 * tasks if there is an imbalance.
3598 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3599 * this_rq is locked.
3602 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3605 struct sched_group *group;
3606 struct rq *busiest = NULL;
3607 unsigned long imbalance;
3615 * When power savings policy is enabled for the parent domain, idle
3616 * sibling can pick up load irrespective of busy siblings. In this case,
3617 * let the state of idle sibling percolate up as IDLE, instead of
3618 * portraying it as CPU_NOT_IDLE.
3620 if (sd->flags & SD_SHARE_CPUPOWER &&
3621 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3624 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3626 update_shares_locked(this_rq, sd);
3627 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3628 &sd_idle, cpus, NULL);
3630 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3634 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3636 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3640 BUG_ON(busiest == this_rq);
3642 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3645 if (busiest->nr_running > 1) {
3646 /* Attempt to move tasks */
3647 double_lock_balance(this_rq, busiest);
3648 /* this_rq->clock is already updated */
3649 update_rq_clock(busiest);
3650 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3651 imbalance, sd, CPU_NEWLY_IDLE,
3653 double_unlock_balance(this_rq, busiest);
3655 if (unlikely(all_pinned)) {
3656 cpu_clear(cpu_of(busiest), *cpus);
3657 if (!cpus_empty(*cpus))
3663 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3664 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3665 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3668 sd->nr_balance_failed = 0;
3670 update_shares_locked(this_rq, sd);
3674 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3675 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3676 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3678 sd->nr_balance_failed = 0;
3684 * idle_balance is called by schedule() if this_cpu is about to become
3685 * idle. Attempts to pull tasks from other CPUs.
3687 static void idle_balance(int this_cpu, struct rq *this_rq)
3689 struct sched_domain *sd;
3690 int pulled_task = -1;
3691 unsigned long next_balance = jiffies + HZ;
3694 for_each_domain(this_cpu, sd) {
3695 unsigned long interval;
3697 if (!(sd->flags & SD_LOAD_BALANCE))
3700 if (sd->flags & SD_BALANCE_NEWIDLE)
3701 /* If we've pulled tasks over stop searching: */
3702 pulled_task = load_balance_newidle(this_cpu, this_rq,
3705 interval = msecs_to_jiffies(sd->balance_interval);
3706 if (time_after(next_balance, sd->last_balance + interval))
3707 next_balance = sd->last_balance + interval;
3711 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3713 * We are going idle. next_balance may be set based on
3714 * a busy processor. So reset next_balance.
3716 this_rq->next_balance = next_balance;
3721 * active_load_balance is run by migration threads. It pushes running tasks
3722 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3723 * running on each physical CPU where possible, and avoids physical /
3724 * logical imbalances.
3726 * Called with busiest_rq locked.
3728 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3730 int target_cpu = busiest_rq->push_cpu;
3731 struct sched_domain *sd;
3732 struct rq *target_rq;
3734 /* Is there any task to move? */
3735 if (busiest_rq->nr_running <= 1)
3738 target_rq = cpu_rq(target_cpu);
3741 * This condition is "impossible", if it occurs
3742 * we need to fix it. Originally reported by
3743 * Bjorn Helgaas on a 128-cpu setup.
3745 BUG_ON(busiest_rq == target_rq);
3747 /* move a task from busiest_rq to target_rq */
3748 double_lock_balance(busiest_rq, target_rq);
3749 update_rq_clock(busiest_rq);
3750 update_rq_clock(target_rq);
3752 /* Search for an sd spanning us and the target CPU. */
3753 for_each_domain(target_cpu, sd) {
3754 if ((sd->flags & SD_LOAD_BALANCE) &&
3755 cpu_isset(busiest_cpu, sd->span))
3760 schedstat_inc(sd, alb_count);
3762 if (move_one_task(target_rq, target_cpu, busiest_rq,
3764 schedstat_inc(sd, alb_pushed);
3766 schedstat_inc(sd, alb_failed);
3768 double_unlock_balance(busiest_rq, target_rq);
3773 atomic_t load_balancer;
3775 } nohz ____cacheline_aligned = {
3776 .load_balancer = ATOMIC_INIT(-1),
3777 .cpu_mask = CPU_MASK_NONE,
3781 * This routine will try to nominate the ilb (idle load balancing)
3782 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3783 * load balancing on behalf of all those cpus. If all the cpus in the system
3784 * go into this tickless mode, then there will be no ilb owner (as there is
3785 * no need for one) and all the cpus will sleep till the next wakeup event
3788 * For the ilb owner, tick is not stopped. And this tick will be used
3789 * for idle load balancing. ilb owner will still be part of
3792 * While stopping the tick, this cpu will become the ilb owner if there
3793 * is no other owner. And will be the owner till that cpu becomes busy
3794 * or if all cpus in the system stop their ticks at which point
3795 * there is no need for ilb owner.
3797 * When the ilb owner becomes busy, it nominates another owner, during the
3798 * next busy scheduler_tick()
3800 int select_nohz_load_balancer(int stop_tick)
3802 int cpu = smp_processor_id();
3805 cpu_set(cpu, nohz.cpu_mask);
3806 cpu_rq(cpu)->in_nohz_recently = 1;
3809 * If we are going offline and still the leader, give up!
3811 if (!cpu_active(cpu) &&
3812 atomic_read(&nohz.load_balancer) == cpu) {
3813 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3818 /* time for ilb owner also to sleep */
3819 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3820 if (atomic_read(&nohz.load_balancer) == cpu)
3821 atomic_set(&nohz.load_balancer, -1);
3825 if (atomic_read(&nohz.load_balancer) == -1) {
3826 /* make me the ilb owner */
3827 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3829 } else if (atomic_read(&nohz.load_balancer) == cpu)
3832 if (!cpu_isset(cpu, nohz.cpu_mask))
3835 cpu_clear(cpu, nohz.cpu_mask);
3837 if (atomic_read(&nohz.load_balancer) == cpu)
3838 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3845 static DEFINE_SPINLOCK(balancing);
3848 * It checks each scheduling domain to see if it is due to be balanced,
3849 * and initiates a balancing operation if so.
3851 * Balancing parameters are set up in arch_init_sched_domains.
3853 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3856 struct rq *rq = cpu_rq(cpu);
3857 unsigned long interval;
3858 struct sched_domain *sd;
3859 /* Earliest time when we have to do rebalance again */
3860 unsigned long next_balance = jiffies + 60*HZ;
3861 int update_next_balance = 0;
3865 for_each_domain(cpu, sd) {
3866 if (!(sd->flags & SD_LOAD_BALANCE))
3869 interval = sd->balance_interval;
3870 if (idle != CPU_IDLE)
3871 interval *= sd->busy_factor;
3873 /* scale ms to jiffies */
3874 interval = msecs_to_jiffies(interval);
3875 if (unlikely(!interval))
3877 if (interval > HZ*NR_CPUS/10)
3878 interval = HZ*NR_CPUS/10;
3880 need_serialize = sd->flags & SD_SERIALIZE;
3882 if (need_serialize) {
3883 if (!spin_trylock(&balancing))
3887 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3888 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3890 * We've pulled tasks over so either we're no
3891 * longer idle, or one of our SMT siblings is
3894 idle = CPU_NOT_IDLE;
3896 sd->last_balance = jiffies;
3899 spin_unlock(&balancing);
3901 if (time_after(next_balance, sd->last_balance + interval)) {
3902 next_balance = sd->last_balance + interval;
3903 update_next_balance = 1;
3907 * Stop the load balance at this level. There is another
3908 * CPU in our sched group which is doing load balancing more
3916 * next_balance will be updated only when there is a need.
3917 * When the cpu is attached to null domain for ex, it will not be
3920 if (likely(update_next_balance))
3921 rq->next_balance = next_balance;
3925 * run_rebalance_domains is triggered when needed from the scheduler tick.
3926 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3927 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3929 static void run_rebalance_domains(struct softirq_action *h)
3931 int this_cpu = smp_processor_id();
3932 struct rq *this_rq = cpu_rq(this_cpu);
3933 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3934 CPU_IDLE : CPU_NOT_IDLE;
3936 rebalance_domains(this_cpu, idle);
3940 * If this cpu is the owner for idle load balancing, then do the
3941 * balancing on behalf of the other idle cpus whose ticks are
3944 if (this_rq->idle_at_tick &&
3945 atomic_read(&nohz.load_balancer) == this_cpu) {
3946 cpumask_t cpus = nohz.cpu_mask;
3950 cpu_clear(this_cpu, cpus);
3951 for_each_cpu_mask_nr(balance_cpu, cpus) {
3953 * If this cpu gets work to do, stop the load balancing
3954 * work being done for other cpus. Next load
3955 * balancing owner will pick it up.
3960 rebalance_domains(balance_cpu, CPU_IDLE);
3962 rq = cpu_rq(balance_cpu);
3963 if (time_after(this_rq->next_balance, rq->next_balance))
3964 this_rq->next_balance = rq->next_balance;
3971 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3973 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3974 * idle load balancing owner or decide to stop the periodic load balancing,
3975 * if the whole system is idle.
3977 static inline void trigger_load_balance(struct rq *rq, int cpu)
3981 * If we were in the nohz mode recently and busy at the current
3982 * scheduler tick, then check if we need to nominate new idle
3985 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3986 rq->in_nohz_recently = 0;
3988 if (atomic_read(&nohz.load_balancer) == cpu) {
3989 cpu_clear(cpu, nohz.cpu_mask);
3990 atomic_set(&nohz.load_balancer, -1);
3993 if (atomic_read(&nohz.load_balancer) == -1) {
3995 * simple selection for now: Nominate the
3996 * first cpu in the nohz list to be the next
3999 * TBD: Traverse the sched domains and nominate
4000 * the nearest cpu in the nohz.cpu_mask.
4002 int ilb = first_cpu(nohz.cpu_mask);
4004 if (ilb < nr_cpu_ids)
4010 * If this cpu is idle and doing idle load balancing for all the
4011 * cpus with ticks stopped, is it time for that to stop?
4013 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4014 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4020 * If this cpu is idle and the idle load balancing is done by
4021 * someone else, then no need raise the SCHED_SOFTIRQ
4023 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4024 cpu_isset(cpu, nohz.cpu_mask))
4027 if (time_after_eq(jiffies, rq->next_balance))
4028 raise_softirq(SCHED_SOFTIRQ);
4031 #else /* CONFIG_SMP */
4034 * on UP we do not need to balance between CPUs:
4036 static inline void idle_balance(int cpu, struct rq *rq)
4042 DEFINE_PER_CPU(struct kernel_stat, kstat);
4044 EXPORT_PER_CPU_SYMBOL(kstat);
4047 * Return p->sum_exec_runtime plus any more ns on the sched_clock
4048 * that have not yet been banked in case the task is currently running.
4050 unsigned long long task_sched_runtime(struct task_struct *p)
4052 unsigned long flags;
4056 rq = task_rq_lock(p, &flags);
4057 ns = p->se.sum_exec_runtime;
4058 if (task_current(rq, p)) {
4059 update_rq_clock(rq);
4060 delta_exec = rq->clock - p->se.exec_start;
4061 if ((s64)delta_exec > 0)
4064 task_rq_unlock(rq, &flags);
4070 * Account user cpu time to a process.
4071 * @p: the process that the cpu time gets accounted to
4072 * @cputime: the cpu time spent in user space since the last update
4074 void account_user_time(struct task_struct *p, cputime_t cputime)
4076 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4079 p->utime = cputime_add(p->utime, cputime);
4081 /* Add user time to cpustat. */
4082 tmp = cputime_to_cputime64(cputime);
4083 if (TASK_NICE(p) > 0)
4084 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4086 cpustat->user = cputime64_add(cpustat->user, tmp);
4087 /* Account for user time used */
4088 acct_update_integrals(p);
4092 * Account guest cpu time to a process.
4093 * @p: the process that the cpu time gets accounted to
4094 * @cputime: the cpu time spent in virtual machine since the last update
4096 static void account_guest_time(struct task_struct *p, cputime_t cputime)
4099 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4101 tmp = cputime_to_cputime64(cputime);
4103 p->utime = cputime_add(p->utime, cputime);
4104 p->gtime = cputime_add(p->gtime, cputime);
4106 cpustat->user = cputime64_add(cpustat->user, tmp);
4107 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4111 * Account scaled user cpu time to a process.
4112 * @p: the process that the cpu time gets accounted to
4113 * @cputime: the cpu time spent in user space since the last update
4115 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4117 p->utimescaled = cputime_add(p->utimescaled, cputime);
4121 * Account system cpu time to a process.
4122 * @p: the process that the cpu time gets accounted to
4123 * @hardirq_offset: the offset to subtract from hardirq_count()
4124 * @cputime: the cpu time spent in kernel space since the last update
4126 void account_system_time(struct task_struct *p, int hardirq_offset,
4129 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4130 struct rq *rq = this_rq();
4133 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4134 account_guest_time(p, cputime);
4138 p->stime = cputime_add(p->stime, cputime);
4140 /* Add system time to cpustat. */
4141 tmp = cputime_to_cputime64(cputime);
4142 if (hardirq_count() - hardirq_offset)
4143 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4144 else if (softirq_count())
4145 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4146 else if (p != rq->idle)
4147 cpustat->system = cputime64_add(cpustat->system, tmp);
4148 else if (atomic_read(&rq->nr_iowait) > 0)
4149 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4151 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4152 /* Account for system time used */
4153 acct_update_integrals(p);
4157 * Account scaled system cpu time to a process.
4158 * @p: the process that the cpu time gets accounted to
4159 * @hardirq_offset: the offset to subtract from hardirq_count()
4160 * @cputime: the cpu time spent in kernel space since the last update
4162 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4164 p->stimescaled = cputime_add(p->stimescaled, cputime);
4168 * Account for involuntary wait time.
4169 * @p: the process from which the cpu time has been stolen
4170 * @steal: the cpu time spent in involuntary wait
4172 void account_steal_time(struct task_struct *p, cputime_t steal)
4174 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4175 cputime64_t tmp = cputime_to_cputime64(steal);
4176 struct rq *rq = this_rq();
4178 if (p == rq->idle) {
4179 p->stime = cputime_add(p->stime, steal);
4180 if (atomic_read(&rq->nr_iowait) > 0)
4181 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4183 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4185 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4189 * This function gets called by the timer code, with HZ frequency.
4190 * We call it with interrupts disabled.
4192 * It also gets called by the fork code, when changing the parent's
4195 void scheduler_tick(void)
4197 int cpu = smp_processor_id();
4198 struct rq *rq = cpu_rq(cpu);
4199 struct task_struct *curr = rq->curr;
4203 spin_lock(&rq->lock);
4204 update_rq_clock(rq);
4205 update_cpu_load(rq);
4206 curr->sched_class->task_tick(rq, curr, 0);
4207 spin_unlock(&rq->lock);
4210 rq->idle_at_tick = idle_cpu(cpu);
4211 trigger_load_balance(rq, cpu);
4215 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4216 defined(CONFIG_PREEMPT_TRACER))
4218 static inline unsigned long get_parent_ip(unsigned long addr)
4220 if (in_lock_functions(addr)) {
4221 addr = CALLER_ADDR2;
4222 if (in_lock_functions(addr))
4223 addr = CALLER_ADDR3;
4228 void __kprobes add_preempt_count(int val)
4230 #ifdef CONFIG_DEBUG_PREEMPT
4234 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4237 preempt_count() += val;
4238 #ifdef CONFIG_DEBUG_PREEMPT
4240 * Spinlock count overflowing soon?
4242 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4245 if (preempt_count() == val)
4246 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4248 EXPORT_SYMBOL(add_preempt_count);
4250 void __kprobes sub_preempt_count(int val)
4252 #ifdef CONFIG_DEBUG_PREEMPT
4256 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4259 * Is the spinlock portion underflowing?
4261 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4262 !(preempt_count() & PREEMPT_MASK)))
4266 if (preempt_count() == val)
4267 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4268 preempt_count() -= val;
4270 EXPORT_SYMBOL(sub_preempt_count);
4275 * Print scheduling while atomic bug:
4277 static noinline void __schedule_bug(struct task_struct *prev)
4279 struct pt_regs *regs = get_irq_regs();
4281 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4282 prev->comm, prev->pid, preempt_count());
4284 debug_show_held_locks(prev);
4286 if (irqs_disabled())
4287 print_irqtrace_events(prev);
4296 * Various schedule()-time debugging checks and statistics:
4298 static inline void schedule_debug(struct task_struct *prev)
4301 * Test if we are atomic. Since do_exit() needs to call into
4302 * schedule() atomically, we ignore that path for now.
4303 * Otherwise, whine if we are scheduling when we should not be.
4305 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4306 __schedule_bug(prev);
4308 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4310 schedstat_inc(this_rq(), sched_count);
4311 #ifdef CONFIG_SCHEDSTATS
4312 if (unlikely(prev->lock_depth >= 0)) {
4313 schedstat_inc(this_rq(), bkl_count);
4314 schedstat_inc(prev, sched_info.bkl_count);
4320 * Pick up the highest-prio task:
4322 static inline struct task_struct *
4323 pick_next_task(struct rq *rq, struct task_struct *prev)
4325 const struct sched_class *class;
4326 struct task_struct *p;
4329 * Optimization: we know that if all tasks are in
4330 * the fair class we can call that function directly:
4332 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4333 p = fair_sched_class.pick_next_task(rq);
4338 class = sched_class_highest;
4340 p = class->pick_next_task(rq);
4344 * Will never be NULL as the idle class always
4345 * returns a non-NULL p:
4347 class = class->next;
4352 * schedule() is the main scheduler function.
4354 asmlinkage void __sched schedule(void)
4356 struct task_struct *prev, *next;
4357 unsigned long *switch_count;
4363 cpu = smp_processor_id();
4367 switch_count = &prev->nivcsw;
4369 release_kernel_lock(prev);
4370 need_resched_nonpreemptible:
4372 schedule_debug(prev);
4374 if (sched_feat(HRTICK))
4378 * Do the rq-clock update outside the rq lock:
4380 local_irq_disable();
4381 update_rq_clock(rq);
4382 spin_lock(&rq->lock);
4383 clear_tsk_need_resched(prev);
4385 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4386 if (unlikely(signal_pending_state(prev->state, prev)))
4387 prev->state = TASK_RUNNING;
4389 deactivate_task(rq, prev, 1);
4390 switch_count = &prev->nvcsw;
4394 if (prev->sched_class->pre_schedule)
4395 prev->sched_class->pre_schedule(rq, prev);
4398 if (unlikely(!rq->nr_running))
4399 idle_balance(cpu, rq);
4401 prev->sched_class->put_prev_task(rq, prev);
4402 next = pick_next_task(rq, prev);
4404 if (likely(prev != next)) {
4405 sched_info_switch(prev, next);
4411 context_switch(rq, prev, next); /* unlocks the rq */
4413 * the context switch might have flipped the stack from under
4414 * us, hence refresh the local variables.
4416 cpu = smp_processor_id();
4419 spin_unlock_irq(&rq->lock);
4421 if (unlikely(reacquire_kernel_lock(current) < 0))
4422 goto need_resched_nonpreemptible;
4424 preempt_enable_no_resched();
4425 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4428 EXPORT_SYMBOL(schedule);
4430 #ifdef CONFIG_PREEMPT
4432 * this is the entry point to schedule() from in-kernel preemption
4433 * off of preempt_enable. Kernel preemptions off return from interrupt
4434 * occur there and call schedule directly.
4436 asmlinkage void __sched preempt_schedule(void)
4438 struct thread_info *ti = current_thread_info();
4441 * If there is a non-zero preempt_count or interrupts are disabled,
4442 * we do not want to preempt the current task. Just return..
4444 if (likely(ti->preempt_count || irqs_disabled()))
4448 add_preempt_count(PREEMPT_ACTIVE);
4450 sub_preempt_count(PREEMPT_ACTIVE);
4453 * Check again in case we missed a preemption opportunity
4454 * between schedule and now.
4457 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4459 EXPORT_SYMBOL(preempt_schedule);
4462 * this is the entry point to schedule() from kernel preemption
4463 * off of irq context.
4464 * Note, that this is called and return with irqs disabled. This will
4465 * protect us against recursive calling from irq.
4467 asmlinkage void __sched preempt_schedule_irq(void)
4469 struct thread_info *ti = current_thread_info();
4471 /* Catch callers which need to be fixed */
4472 BUG_ON(ti->preempt_count || !irqs_disabled());
4475 add_preempt_count(PREEMPT_ACTIVE);
4478 local_irq_disable();
4479 sub_preempt_count(PREEMPT_ACTIVE);
4482 * Check again in case we missed a preemption opportunity
4483 * between schedule and now.
4486 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4489 #endif /* CONFIG_PREEMPT */
4491 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4494 return try_to_wake_up(curr->private, mode, sync);
4496 EXPORT_SYMBOL(default_wake_function);
4499 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4500 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4501 * number) then we wake all the non-exclusive tasks and one exclusive task.
4503 * There are circumstances in which we can try to wake a task which has already
4504 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4505 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4507 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4508 int nr_exclusive, int sync, void *key)
4510 wait_queue_t *curr, *next;
4512 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4513 unsigned flags = curr->flags;
4515 if (curr->func(curr, mode, sync, key) &&
4516 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4522 * __wake_up - wake up threads blocked on a waitqueue.
4524 * @mode: which threads
4525 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4526 * @key: is directly passed to the wakeup function
4528 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4529 int nr_exclusive, void *key)
4531 unsigned long flags;
4533 spin_lock_irqsave(&q->lock, flags);
4534 __wake_up_common(q, mode, nr_exclusive, 0, key);
4535 spin_unlock_irqrestore(&q->lock, flags);
4537 EXPORT_SYMBOL(__wake_up);
4540 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4542 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4544 __wake_up_common(q, mode, 1, 0, NULL);
4548 * __wake_up_sync - wake up threads blocked on a waitqueue.
4550 * @mode: which threads
4551 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4553 * The sync wakeup differs that the waker knows that it will schedule
4554 * away soon, so while the target thread will be woken up, it will not
4555 * be migrated to another CPU - ie. the two threads are 'synchronized'
4556 * with each other. This can prevent needless bouncing between CPUs.
4558 * On UP it can prevent extra preemption.
4561 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4563 unsigned long flags;
4569 if (unlikely(!nr_exclusive))
4572 spin_lock_irqsave(&q->lock, flags);
4573 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4574 spin_unlock_irqrestore(&q->lock, flags);
4576 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4578 void complete(struct completion *x)
4580 unsigned long flags;
4582 spin_lock_irqsave(&x->wait.lock, flags);
4584 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4585 spin_unlock_irqrestore(&x->wait.lock, flags);
4587 EXPORT_SYMBOL(complete);
4589 void complete_all(struct completion *x)
4591 unsigned long flags;
4593 spin_lock_irqsave(&x->wait.lock, flags);
4594 x->done += UINT_MAX/2;
4595 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4596 spin_unlock_irqrestore(&x->wait.lock, flags);
4598 EXPORT_SYMBOL(complete_all);
4600 static inline long __sched
4601 do_wait_for_common(struct completion *x, long timeout, int state)
4604 DECLARE_WAITQUEUE(wait, current);
4606 wait.flags |= WQ_FLAG_EXCLUSIVE;
4607 __add_wait_queue_tail(&x->wait, &wait);
4609 if ((state == TASK_INTERRUPTIBLE &&
4610 signal_pending(current)) ||
4611 (state == TASK_KILLABLE &&
4612 fatal_signal_pending(current))) {
4613 timeout = -ERESTARTSYS;
4616 __set_current_state(state);
4617 spin_unlock_irq(&x->wait.lock);
4618 timeout = schedule_timeout(timeout);
4619 spin_lock_irq(&x->wait.lock);
4620 } while (!x->done && timeout);
4621 __remove_wait_queue(&x->wait, &wait);
4626 return timeout ?: 1;
4630 wait_for_common(struct completion *x, long timeout, int state)
4634 spin_lock_irq(&x->wait.lock);
4635 timeout = do_wait_for_common(x, timeout, state);
4636 spin_unlock_irq(&x->wait.lock);
4640 void __sched wait_for_completion(struct completion *x)
4642 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4644 EXPORT_SYMBOL(wait_for_completion);
4646 unsigned long __sched
4647 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4649 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4651 EXPORT_SYMBOL(wait_for_completion_timeout);
4653 int __sched wait_for_completion_interruptible(struct completion *x)
4655 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4656 if (t == -ERESTARTSYS)
4660 EXPORT_SYMBOL(wait_for_completion_interruptible);
4662 unsigned long __sched
4663 wait_for_completion_interruptible_timeout(struct completion *x,
4664 unsigned long timeout)
4666 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4668 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4670 int __sched wait_for_completion_killable(struct completion *x)
4672 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4673 if (t == -ERESTARTSYS)
4677 EXPORT_SYMBOL(wait_for_completion_killable);
4680 * try_wait_for_completion - try to decrement a completion without blocking
4681 * @x: completion structure
4683 * Returns: 0 if a decrement cannot be done without blocking
4684 * 1 if a decrement succeeded.
4686 * If a completion is being used as a counting completion,
4687 * attempt to decrement the counter without blocking. This
4688 * enables us to avoid waiting if the resource the completion
4689 * is protecting is not available.
4691 bool try_wait_for_completion(struct completion *x)
4695 spin_lock_irq(&x->wait.lock);
4700 spin_unlock_irq(&x->wait.lock);
4703 EXPORT_SYMBOL(try_wait_for_completion);
4706 * completion_done - Test to see if a completion has any waiters
4707 * @x: completion structure
4709 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4710 * 1 if there are no waiters.
4713 bool completion_done(struct completion *x)
4717 spin_lock_irq(&x->wait.lock);
4720 spin_unlock_irq(&x->wait.lock);
4723 EXPORT_SYMBOL(completion_done);
4726 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4728 unsigned long flags;
4731 init_waitqueue_entry(&wait, current);
4733 __set_current_state(state);
4735 spin_lock_irqsave(&q->lock, flags);
4736 __add_wait_queue(q, &wait);
4737 spin_unlock(&q->lock);
4738 timeout = schedule_timeout(timeout);
4739 spin_lock_irq(&q->lock);
4740 __remove_wait_queue(q, &wait);
4741 spin_unlock_irqrestore(&q->lock, flags);
4746 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4748 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4750 EXPORT_SYMBOL(interruptible_sleep_on);
4753 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4755 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4757 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4759 void __sched sleep_on(wait_queue_head_t *q)
4761 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4763 EXPORT_SYMBOL(sleep_on);
4765 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4767 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4769 EXPORT_SYMBOL(sleep_on_timeout);
4771 #ifdef CONFIG_RT_MUTEXES
4774 * rt_mutex_setprio - set the current priority of a task
4776 * @prio: prio value (kernel-internal form)
4778 * This function changes the 'effective' priority of a task. It does
4779 * not touch ->normal_prio like __setscheduler().
4781 * Used by the rt_mutex code to implement priority inheritance logic.
4783 void rt_mutex_setprio(struct task_struct *p, int prio)
4785 unsigned long flags;
4786 int oldprio, on_rq, running;
4788 const struct sched_class *prev_class = p->sched_class;
4790 BUG_ON(prio < 0 || prio > MAX_PRIO);
4792 rq = task_rq_lock(p, &flags);
4793 update_rq_clock(rq);
4796 on_rq = p->se.on_rq;
4797 running = task_current(rq, p);
4799 dequeue_task(rq, p, 0);
4801 p->sched_class->put_prev_task(rq, p);
4804 p->sched_class = &rt_sched_class;
4806 p->sched_class = &fair_sched_class;
4811 p->sched_class->set_curr_task(rq);
4813 enqueue_task(rq, p, 0);
4815 check_class_changed(rq, p, prev_class, oldprio, running);
4817 task_rq_unlock(rq, &flags);
4822 void set_user_nice(struct task_struct *p, long nice)
4824 int old_prio, delta, on_rq;
4825 unsigned long flags;
4828 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4831 * We have to be careful, if called from sys_setpriority(),
4832 * the task might be in the middle of scheduling on another CPU.
4834 rq = task_rq_lock(p, &flags);
4835 update_rq_clock(rq);
4837 * The RT priorities are set via sched_setscheduler(), but we still
4838 * allow the 'normal' nice value to be set - but as expected
4839 * it wont have any effect on scheduling until the task is
4840 * SCHED_FIFO/SCHED_RR:
4842 if (task_has_rt_policy(p)) {
4843 p->static_prio = NICE_TO_PRIO(nice);
4846 on_rq = p->se.on_rq;
4848 dequeue_task(rq, p, 0);
4850 p->static_prio = NICE_TO_PRIO(nice);
4853 p->prio = effective_prio(p);
4854 delta = p->prio - old_prio;
4857 enqueue_task(rq, p, 0);
4859 * If the task increased its priority or is running and
4860 * lowered its priority, then reschedule its CPU:
4862 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4863 resched_task(rq->curr);
4866 task_rq_unlock(rq, &flags);
4868 EXPORT_SYMBOL(set_user_nice);
4871 * can_nice - check if a task can reduce its nice value
4875 int can_nice(const struct task_struct *p, const int nice)
4877 /* convert nice value [19,-20] to rlimit style value [1,40] */
4878 int nice_rlim = 20 - nice;
4880 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4881 capable(CAP_SYS_NICE));
4884 #ifdef __ARCH_WANT_SYS_NICE
4887 * sys_nice - change the priority of the current process.
4888 * @increment: priority increment
4890 * sys_setpriority is a more generic, but much slower function that
4891 * does similar things.
4893 asmlinkage long sys_nice(int increment)
4898 * Setpriority might change our priority at the same moment.
4899 * We don't have to worry. Conceptually one call occurs first
4900 * and we have a single winner.
4902 if (increment < -40)
4907 nice = PRIO_TO_NICE(current->static_prio) + increment;
4913 if (increment < 0 && !can_nice(current, nice))
4916 retval = security_task_setnice(current, nice);
4920 set_user_nice(current, nice);
4927 * task_prio - return the priority value of a given task.
4928 * @p: the task in question.
4930 * This is the priority value as seen by users in /proc.
4931 * RT tasks are offset by -200. Normal tasks are centered
4932 * around 0, value goes from -16 to +15.
4934 int task_prio(const struct task_struct *p)
4936 return p->prio - MAX_RT_PRIO;
4940 * task_nice - return the nice value of a given task.
4941 * @p: the task in question.
4943 int task_nice(const struct task_struct *p)
4945 return TASK_NICE(p);
4947 EXPORT_SYMBOL(task_nice);
4950 * idle_cpu - is a given cpu idle currently?
4951 * @cpu: the processor in question.
4953 int idle_cpu(int cpu)
4955 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4959 * idle_task - return the idle task for a given cpu.
4960 * @cpu: the processor in question.
4962 struct task_struct *idle_task(int cpu)
4964 return cpu_rq(cpu)->idle;
4968 * find_process_by_pid - find a process with a matching PID value.
4969 * @pid: the pid in question.
4971 static struct task_struct *find_process_by_pid(pid_t pid)
4973 return pid ? find_task_by_vpid(pid) : current;
4976 /* Actually do priority change: must hold rq lock. */
4978 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4980 BUG_ON(p->se.on_rq);
4983 switch (p->policy) {
4987 p->sched_class = &fair_sched_class;
4991 p->sched_class = &rt_sched_class;
4995 p->rt_priority = prio;
4996 p->normal_prio = normal_prio(p);
4997 /* we are holding p->pi_lock already */
4998 p->prio = rt_mutex_getprio(p);
5002 static int __sched_setscheduler(struct task_struct *p, int policy,
5003 struct sched_param *param, bool user)
5005 int retval, oldprio, oldpolicy = -1, on_rq, running;
5006 unsigned long flags;
5007 const struct sched_class *prev_class = p->sched_class;
5010 /* may grab non-irq protected spin_locks */
5011 BUG_ON(in_interrupt());
5013 /* double check policy once rq lock held */
5015 policy = oldpolicy = p->policy;
5016 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5017 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5018 policy != SCHED_IDLE)
5021 * Valid priorities for SCHED_FIFO and SCHED_RR are
5022 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5023 * SCHED_BATCH and SCHED_IDLE is 0.
5025 if (param->sched_priority < 0 ||
5026 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5027 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5029 if (rt_policy(policy) != (param->sched_priority != 0))
5033 * Allow unprivileged RT tasks to decrease priority:
5035 if (user && !capable(CAP_SYS_NICE)) {
5036 if (rt_policy(policy)) {
5037 unsigned long rlim_rtprio;
5039 if (!lock_task_sighand(p, &flags))
5041 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5042 unlock_task_sighand(p, &flags);
5044 /* can't set/change the rt policy */
5045 if (policy != p->policy && !rlim_rtprio)
5048 /* can't increase priority */
5049 if (param->sched_priority > p->rt_priority &&
5050 param->sched_priority > rlim_rtprio)
5054 * Like positive nice levels, dont allow tasks to
5055 * move out of SCHED_IDLE either:
5057 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5060 /* can't change other user's priorities */
5061 if ((current->euid != p->euid) &&
5062 (current->euid != p->uid))
5067 #ifdef CONFIG_RT_GROUP_SCHED
5069 * Do not allow realtime tasks into groups that have no runtime
5072 if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
5076 retval = security_task_setscheduler(p, policy, param);
5082 * make sure no PI-waiters arrive (or leave) while we are
5083 * changing the priority of the task:
5085 spin_lock_irqsave(&p->pi_lock, flags);
5087 * To be able to change p->policy safely, the apropriate
5088 * runqueue lock must be held.
5090 rq = __task_rq_lock(p);
5091 /* recheck policy now with rq lock held */
5092 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5093 policy = oldpolicy = -1;
5094 __task_rq_unlock(rq);
5095 spin_unlock_irqrestore(&p->pi_lock, flags);
5098 update_rq_clock(rq);
5099 on_rq = p->se.on_rq;
5100 running = task_current(rq, p);
5102 deactivate_task(rq, p, 0);
5104 p->sched_class->put_prev_task(rq, p);
5107 __setscheduler(rq, p, policy, param->sched_priority);
5110 p->sched_class->set_curr_task(rq);
5112 activate_task(rq, p, 0);
5114 check_class_changed(rq, p, prev_class, oldprio, running);
5116 __task_rq_unlock(rq);
5117 spin_unlock_irqrestore(&p->pi_lock, flags);
5119 rt_mutex_adjust_pi(p);
5125 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5126 * @p: the task in question.
5127 * @policy: new policy.
5128 * @param: structure containing the new RT priority.
5130 * NOTE that the task may be already dead.
5132 int sched_setscheduler(struct task_struct *p, int policy,
5133 struct sched_param *param)
5135 return __sched_setscheduler(p, policy, param, true);
5137 EXPORT_SYMBOL_GPL(sched_setscheduler);
5140 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5141 * @p: the task in question.
5142 * @policy: new policy.
5143 * @param: structure containing the new RT priority.
5145 * Just like sched_setscheduler, only don't bother checking if the
5146 * current context has permission. For example, this is needed in
5147 * stop_machine(): we create temporary high priority worker threads,
5148 * but our caller might not have that capability.
5150 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5151 struct sched_param *param)
5153 return __sched_setscheduler(p, policy, param, false);
5157 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5159 struct sched_param lparam;
5160 struct task_struct *p;
5163 if (!param || pid < 0)
5165 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5170 p = find_process_by_pid(pid);
5172 retval = sched_setscheduler(p, policy, &lparam);
5179 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5180 * @pid: the pid in question.
5181 * @policy: new policy.
5182 * @param: structure containing the new RT priority.
5185 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5187 /* negative values for policy are not valid */
5191 return do_sched_setscheduler(pid, policy, param);
5195 * sys_sched_setparam - set/change the RT priority of a thread
5196 * @pid: the pid in question.
5197 * @param: structure containing the new RT priority.
5199 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5201 return do_sched_setscheduler(pid, -1, param);
5205 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5206 * @pid: the pid in question.
5208 asmlinkage long sys_sched_getscheduler(pid_t pid)
5210 struct task_struct *p;
5217 read_lock(&tasklist_lock);
5218 p = find_process_by_pid(pid);
5220 retval = security_task_getscheduler(p);
5224 read_unlock(&tasklist_lock);
5229 * sys_sched_getscheduler - get the RT priority of a thread
5230 * @pid: the pid in question.
5231 * @param: structure containing the RT priority.
5233 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5235 struct sched_param lp;
5236 struct task_struct *p;
5239 if (!param || pid < 0)
5242 read_lock(&tasklist_lock);
5243 p = find_process_by_pid(pid);
5248 retval = security_task_getscheduler(p);
5252 lp.sched_priority = p->rt_priority;
5253 read_unlock(&tasklist_lock);
5256 * This one might sleep, we cannot do it with a spinlock held ...
5258 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5263 read_unlock(&tasklist_lock);
5267 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5269 cpumask_t cpus_allowed;
5270 cpumask_t new_mask = *in_mask;
5271 struct task_struct *p;
5275 read_lock(&tasklist_lock);
5277 p = find_process_by_pid(pid);
5279 read_unlock(&tasklist_lock);
5285 * It is not safe to call set_cpus_allowed with the
5286 * tasklist_lock held. We will bump the task_struct's
5287 * usage count and then drop tasklist_lock.
5290 read_unlock(&tasklist_lock);
5293 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5294 !capable(CAP_SYS_NICE))
5297 retval = security_task_setscheduler(p, 0, NULL);
5301 cpuset_cpus_allowed(p, &cpus_allowed);
5302 cpus_and(new_mask, new_mask, cpus_allowed);
5304 retval = set_cpus_allowed_ptr(p, &new_mask);
5307 cpuset_cpus_allowed(p, &cpus_allowed);
5308 if (!cpus_subset(new_mask, cpus_allowed)) {
5310 * We must have raced with a concurrent cpuset
5311 * update. Just reset the cpus_allowed to the
5312 * cpuset's cpus_allowed
5314 new_mask = cpus_allowed;
5324 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5325 cpumask_t *new_mask)
5327 if (len < sizeof(cpumask_t)) {
5328 memset(new_mask, 0, sizeof(cpumask_t));
5329 } else if (len > sizeof(cpumask_t)) {
5330 len = sizeof(cpumask_t);
5332 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5336 * sys_sched_setaffinity - set the cpu affinity of a process
5337 * @pid: pid of the process
5338 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5339 * @user_mask_ptr: user-space pointer to the new cpu mask
5341 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5342 unsigned long __user *user_mask_ptr)
5347 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5351 return sched_setaffinity(pid, &new_mask);
5354 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5356 struct task_struct *p;
5360 read_lock(&tasklist_lock);
5363 p = find_process_by_pid(pid);
5367 retval = security_task_getscheduler(p);
5371 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5374 read_unlock(&tasklist_lock);
5381 * sys_sched_getaffinity - get the cpu affinity of a process
5382 * @pid: pid of the process
5383 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5384 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5386 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5387 unsigned long __user *user_mask_ptr)
5392 if (len < sizeof(cpumask_t))
5395 ret = sched_getaffinity(pid, &mask);
5399 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5402 return sizeof(cpumask_t);
5406 * sys_sched_yield - yield the current processor to other threads.
5408 * This function yields the current CPU to other tasks. If there are no
5409 * other threads running on this CPU then this function will return.
5411 asmlinkage long sys_sched_yield(void)
5413 struct rq *rq = this_rq_lock();
5415 schedstat_inc(rq, yld_count);
5416 current->sched_class->yield_task(rq);
5419 * Since we are going to call schedule() anyway, there's
5420 * no need to preempt or enable interrupts:
5422 __release(rq->lock);
5423 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5424 _raw_spin_unlock(&rq->lock);
5425 preempt_enable_no_resched();
5432 static void __cond_resched(void)
5434 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5435 __might_sleep(__FILE__, __LINE__);
5438 * The BKS might be reacquired before we have dropped
5439 * PREEMPT_ACTIVE, which could trigger a second
5440 * cond_resched() call.
5443 add_preempt_count(PREEMPT_ACTIVE);
5445 sub_preempt_count(PREEMPT_ACTIVE);
5446 } while (need_resched());
5449 int __sched _cond_resched(void)
5451 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5452 system_state == SYSTEM_RUNNING) {
5458 EXPORT_SYMBOL(_cond_resched);
5461 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5462 * call schedule, and on return reacquire the lock.
5464 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5465 * operations here to prevent schedule() from being called twice (once via
5466 * spin_unlock(), once by hand).
5468 int cond_resched_lock(spinlock_t *lock)
5470 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5473 if (spin_needbreak(lock) || resched) {
5475 if (resched && need_resched())
5484 EXPORT_SYMBOL(cond_resched_lock);
5486 int __sched cond_resched_softirq(void)
5488 BUG_ON(!in_softirq());
5490 if (need_resched() && system_state == SYSTEM_RUNNING) {
5498 EXPORT_SYMBOL(cond_resched_softirq);
5501 * yield - yield the current processor to other threads.
5503 * This is a shortcut for kernel-space yielding - it marks the
5504 * thread runnable and calls sys_sched_yield().
5506 void __sched yield(void)
5508 set_current_state(TASK_RUNNING);
5511 EXPORT_SYMBOL(yield);
5514 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5515 * that process accounting knows that this is a task in IO wait state.
5517 * But don't do that if it is a deliberate, throttling IO wait (this task
5518 * has set its backing_dev_info: the queue against which it should throttle)
5520 void __sched io_schedule(void)
5522 struct rq *rq = &__raw_get_cpu_var(runqueues);
5524 delayacct_blkio_start();
5525 atomic_inc(&rq->nr_iowait);
5527 atomic_dec(&rq->nr_iowait);
5528 delayacct_blkio_end();
5530 EXPORT_SYMBOL(io_schedule);
5532 long __sched io_schedule_timeout(long timeout)
5534 struct rq *rq = &__raw_get_cpu_var(runqueues);
5537 delayacct_blkio_start();
5538 atomic_inc(&rq->nr_iowait);
5539 ret = schedule_timeout(timeout);
5540 atomic_dec(&rq->nr_iowait);
5541 delayacct_blkio_end();
5546 * sys_sched_get_priority_max - return maximum RT priority.
5547 * @policy: scheduling class.
5549 * this syscall returns the maximum rt_priority that can be used
5550 * by a given scheduling class.
5552 asmlinkage long sys_sched_get_priority_max(int policy)
5559 ret = MAX_USER_RT_PRIO-1;
5571 * sys_sched_get_priority_min - return minimum RT priority.
5572 * @policy: scheduling class.
5574 * this syscall returns the minimum rt_priority that can be used
5575 * by a given scheduling class.
5577 asmlinkage long sys_sched_get_priority_min(int policy)
5595 * sys_sched_rr_get_interval - return the default timeslice of a process.
5596 * @pid: pid of the process.
5597 * @interval: userspace pointer to the timeslice value.
5599 * this syscall writes the default timeslice value of a given process
5600 * into the user-space timespec buffer. A value of '0' means infinity.
5603 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5605 struct task_struct *p;
5606 unsigned int time_slice;
5614 read_lock(&tasklist_lock);
5615 p = find_process_by_pid(pid);
5619 retval = security_task_getscheduler(p);
5624 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5625 * tasks that are on an otherwise idle runqueue:
5628 if (p->policy == SCHED_RR) {
5629 time_slice = DEF_TIMESLICE;
5630 } else if (p->policy != SCHED_FIFO) {
5631 struct sched_entity *se = &p->se;
5632 unsigned long flags;
5635 rq = task_rq_lock(p, &flags);
5636 if (rq->cfs.load.weight)
5637 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5638 task_rq_unlock(rq, &flags);
5640 read_unlock(&tasklist_lock);
5641 jiffies_to_timespec(time_slice, &t);
5642 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5646 read_unlock(&tasklist_lock);
5650 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5652 void sched_show_task(struct task_struct *p)
5654 unsigned long free = 0;
5657 state = p->state ? __ffs(p->state) + 1 : 0;
5658 printk(KERN_INFO "%-13.13s %c", p->comm,
5659 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5660 #if BITS_PER_LONG == 32
5661 if (state == TASK_RUNNING)
5662 printk(KERN_CONT " running ");
5664 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5666 if (state == TASK_RUNNING)
5667 printk(KERN_CONT " running task ");
5669 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5671 #ifdef CONFIG_DEBUG_STACK_USAGE
5673 unsigned long *n = end_of_stack(p);
5676 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5679 printk(KERN_CONT "%5lu %5d %6d\n", free,
5680 task_pid_nr(p), task_pid_nr(p->real_parent));
5682 show_stack(p, NULL);
5685 void show_state_filter(unsigned long state_filter)
5687 struct task_struct *g, *p;
5689 #if BITS_PER_LONG == 32
5691 " task PC stack pid father\n");
5694 " task PC stack pid father\n");
5696 read_lock(&tasklist_lock);
5697 do_each_thread(g, p) {
5699 * reset the NMI-timeout, listing all files on a slow
5700 * console might take alot of time:
5702 touch_nmi_watchdog();
5703 if (!state_filter || (p->state & state_filter))
5705 } while_each_thread(g, p);
5707 touch_all_softlockup_watchdogs();
5709 #ifdef CONFIG_SCHED_DEBUG
5710 sysrq_sched_debug_show();
5712 read_unlock(&tasklist_lock);
5714 * Only show locks if all tasks are dumped:
5716 if (state_filter == -1)
5717 debug_show_all_locks();
5720 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5722 idle->sched_class = &idle_sched_class;
5726 * init_idle - set up an idle thread for a given CPU
5727 * @idle: task in question
5728 * @cpu: cpu the idle task belongs to
5730 * NOTE: this function does not set the idle thread's NEED_RESCHED
5731 * flag, to make booting more robust.
5733 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5735 struct rq *rq = cpu_rq(cpu);
5736 unsigned long flags;
5739 idle->se.exec_start = sched_clock();
5741 idle->prio = idle->normal_prio = MAX_PRIO;
5742 idle->cpus_allowed = cpumask_of_cpu(cpu);
5743 __set_task_cpu(idle, cpu);
5745 spin_lock_irqsave(&rq->lock, flags);
5746 rq->curr = rq->idle = idle;
5747 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5750 spin_unlock_irqrestore(&rq->lock, flags);
5752 /* Set the preempt count _outside_ the spinlocks! */
5753 #if defined(CONFIG_PREEMPT)
5754 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5756 task_thread_info(idle)->preempt_count = 0;
5759 * The idle tasks have their own, simple scheduling class:
5761 idle->sched_class = &idle_sched_class;
5765 * In a system that switches off the HZ timer nohz_cpu_mask
5766 * indicates which cpus entered this state. This is used
5767 * in the rcu update to wait only for active cpus. For system
5768 * which do not switch off the HZ timer nohz_cpu_mask should
5769 * always be CPU_MASK_NONE.
5771 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5774 * Increase the granularity value when there are more CPUs,
5775 * because with more CPUs the 'effective latency' as visible
5776 * to users decreases. But the relationship is not linear,
5777 * so pick a second-best guess by going with the log2 of the
5780 * This idea comes from the SD scheduler of Con Kolivas:
5782 static inline void sched_init_granularity(void)
5784 unsigned int factor = 1 + ilog2(num_online_cpus());
5785 const unsigned long limit = 200000000;
5787 sysctl_sched_min_granularity *= factor;
5788 if (sysctl_sched_min_granularity > limit)
5789 sysctl_sched_min_granularity = limit;
5791 sysctl_sched_latency *= factor;
5792 if (sysctl_sched_latency > limit)
5793 sysctl_sched_latency = limit;
5795 sysctl_sched_wakeup_granularity *= factor;
5797 sysctl_sched_shares_ratelimit *= factor;
5802 * This is how migration works:
5804 * 1) we queue a struct migration_req structure in the source CPU's
5805 * runqueue and wake up that CPU's migration thread.
5806 * 2) we down() the locked semaphore => thread blocks.
5807 * 3) migration thread wakes up (implicitly it forces the migrated
5808 * thread off the CPU)
5809 * 4) it gets the migration request and checks whether the migrated
5810 * task is still in the wrong runqueue.
5811 * 5) if it's in the wrong runqueue then the migration thread removes
5812 * it and puts it into the right queue.
5813 * 6) migration thread up()s the semaphore.
5814 * 7) we wake up and the migration is done.
5818 * Change a given task's CPU affinity. Migrate the thread to a
5819 * proper CPU and schedule it away if the CPU it's executing on
5820 * is removed from the allowed bitmask.
5822 * NOTE: the caller must have a valid reference to the task, the
5823 * task must not exit() & deallocate itself prematurely. The
5824 * call is not atomic; no spinlocks may be held.
5826 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5828 struct migration_req req;
5829 unsigned long flags;
5833 rq = task_rq_lock(p, &flags);
5834 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5839 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5840 !cpus_equal(p->cpus_allowed, *new_mask))) {
5845 if (p->sched_class->set_cpus_allowed)
5846 p->sched_class->set_cpus_allowed(p, new_mask);
5848 p->cpus_allowed = *new_mask;
5849 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5852 /* Can the task run on the task's current CPU? If so, we're done */
5853 if (cpu_isset(task_cpu(p), *new_mask))
5856 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5857 /* Need help from migration thread: drop lock and wait. */
5858 task_rq_unlock(rq, &flags);
5859 wake_up_process(rq->migration_thread);
5860 wait_for_completion(&req.done);
5861 tlb_migrate_finish(p->mm);
5865 task_rq_unlock(rq, &flags);
5869 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5872 * Move (not current) task off this cpu, onto dest cpu. We're doing
5873 * this because either it can't run here any more (set_cpus_allowed()
5874 * away from this CPU, or CPU going down), or because we're
5875 * attempting to rebalance this task on exec (sched_exec).
5877 * So we race with normal scheduler movements, but that's OK, as long
5878 * as the task is no longer on this CPU.
5880 * Returns non-zero if task was successfully migrated.
5882 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5884 struct rq *rq_dest, *rq_src;
5887 if (unlikely(!cpu_active(dest_cpu)))
5890 rq_src = cpu_rq(src_cpu);
5891 rq_dest = cpu_rq(dest_cpu);
5893 double_rq_lock(rq_src, rq_dest);
5894 /* Already moved. */
5895 if (task_cpu(p) != src_cpu)
5897 /* Affinity changed (again). */
5898 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5901 on_rq = p->se.on_rq;
5903 deactivate_task(rq_src, p, 0);
5905 set_task_cpu(p, dest_cpu);
5907 activate_task(rq_dest, p, 0);
5908 check_preempt_curr(rq_dest, p);
5913 double_rq_unlock(rq_src, rq_dest);
5918 * migration_thread - this is a highprio system thread that performs
5919 * thread migration by bumping thread off CPU then 'pushing' onto
5922 static int migration_thread(void *data)
5924 int cpu = (long)data;
5928 BUG_ON(rq->migration_thread != current);
5930 set_current_state(TASK_INTERRUPTIBLE);
5931 while (!kthread_should_stop()) {
5932 struct migration_req *req;
5933 struct list_head *head;
5935 spin_lock_irq(&rq->lock);
5937 if (cpu_is_offline(cpu)) {
5938 spin_unlock_irq(&rq->lock);
5942 if (rq->active_balance) {
5943 active_load_balance(rq, cpu);
5944 rq->active_balance = 0;
5947 head = &rq->migration_queue;
5949 if (list_empty(head)) {
5950 spin_unlock_irq(&rq->lock);
5952 set_current_state(TASK_INTERRUPTIBLE);
5955 req = list_entry(head->next, struct migration_req, list);
5956 list_del_init(head->next);
5958 spin_unlock(&rq->lock);
5959 __migrate_task(req->task, cpu, req->dest_cpu);
5962 complete(&req->done);
5964 __set_current_state(TASK_RUNNING);
5968 /* Wait for kthread_stop */
5969 set_current_state(TASK_INTERRUPTIBLE);
5970 while (!kthread_should_stop()) {
5972 set_current_state(TASK_INTERRUPTIBLE);
5974 __set_current_state(TASK_RUNNING);
5978 #ifdef CONFIG_HOTPLUG_CPU
5980 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5984 local_irq_disable();
5985 ret = __migrate_task(p, src_cpu, dest_cpu);
5991 * Figure out where task on dead CPU should go, use force if necessary.
5992 * NOTE: interrupts should be disabled by the caller
5994 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5996 unsigned long flags;
6003 mask = node_to_cpumask(cpu_to_node(dead_cpu));
6004 cpus_and(mask, mask, p->cpus_allowed);
6005 dest_cpu = any_online_cpu(mask);
6007 /* On any allowed CPU? */
6008 if (dest_cpu >= nr_cpu_ids)
6009 dest_cpu = any_online_cpu(p->cpus_allowed);
6011 /* No more Mr. Nice Guy. */
6012 if (dest_cpu >= nr_cpu_ids) {
6013 cpumask_t cpus_allowed;
6015 cpuset_cpus_allowed_locked(p, &cpus_allowed);
6017 * Try to stay on the same cpuset, where the
6018 * current cpuset may be a subset of all cpus.
6019 * The cpuset_cpus_allowed_locked() variant of
6020 * cpuset_cpus_allowed() will not block. It must be
6021 * called within calls to cpuset_lock/cpuset_unlock.
6023 rq = task_rq_lock(p, &flags);
6024 p->cpus_allowed = cpus_allowed;
6025 dest_cpu = any_online_cpu(p->cpus_allowed);
6026 task_rq_unlock(rq, &flags);
6029 * Don't tell them about moving exiting tasks or
6030 * kernel threads (both mm NULL), since they never
6033 if (p->mm && printk_ratelimit()) {
6034 printk(KERN_INFO "process %d (%s) no "
6035 "longer affine to cpu%d\n",
6036 task_pid_nr(p), p->comm, dead_cpu);
6039 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
6043 * While a dead CPU has no uninterruptible tasks queued at this point,
6044 * it might still have a nonzero ->nr_uninterruptible counter, because
6045 * for performance reasons the counter is not stricly tracking tasks to
6046 * their home CPUs. So we just add the counter to another CPU's counter,
6047 * to keep the global sum constant after CPU-down:
6049 static void migrate_nr_uninterruptible(struct rq *rq_src)
6051 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
6052 unsigned long flags;
6054 local_irq_save(flags);
6055 double_rq_lock(rq_src, rq_dest);
6056 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6057 rq_src->nr_uninterruptible = 0;
6058 double_rq_unlock(rq_src, rq_dest);
6059 local_irq_restore(flags);
6062 /* Run through task list and migrate tasks from the dead cpu. */
6063 static void migrate_live_tasks(int src_cpu)
6065 struct task_struct *p, *t;
6067 read_lock(&tasklist_lock);
6069 do_each_thread(t, p) {
6073 if (task_cpu(p) == src_cpu)
6074 move_task_off_dead_cpu(src_cpu, p);
6075 } while_each_thread(t, p);
6077 read_unlock(&tasklist_lock);
6081 * Schedules idle task to be the next runnable task on current CPU.
6082 * It does so by boosting its priority to highest possible.
6083 * Used by CPU offline code.
6085 void sched_idle_next(void)
6087 int this_cpu = smp_processor_id();
6088 struct rq *rq = cpu_rq(this_cpu);
6089 struct task_struct *p = rq->idle;
6090 unsigned long flags;
6092 /* cpu has to be offline */
6093 BUG_ON(cpu_online(this_cpu));
6096 * Strictly not necessary since rest of the CPUs are stopped by now
6097 * and interrupts disabled on the current cpu.
6099 spin_lock_irqsave(&rq->lock, flags);
6101 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6103 update_rq_clock(rq);
6104 activate_task(rq, p, 0);
6106 spin_unlock_irqrestore(&rq->lock, flags);
6110 * Ensures that the idle task is using init_mm right before its cpu goes
6113 void idle_task_exit(void)
6115 struct mm_struct *mm = current->active_mm;
6117 BUG_ON(cpu_online(smp_processor_id()));
6120 switch_mm(mm, &init_mm, current);
6124 /* called under rq->lock with disabled interrupts */
6125 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6127 struct rq *rq = cpu_rq(dead_cpu);
6129 /* Must be exiting, otherwise would be on tasklist. */
6130 BUG_ON(!p->exit_state);
6132 /* Cannot have done final schedule yet: would have vanished. */
6133 BUG_ON(p->state == TASK_DEAD);
6138 * Drop lock around migration; if someone else moves it,
6139 * that's OK. No task can be added to this CPU, so iteration is
6142 spin_unlock_irq(&rq->lock);
6143 move_task_off_dead_cpu(dead_cpu, p);
6144 spin_lock_irq(&rq->lock);
6149 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6150 static void migrate_dead_tasks(unsigned int dead_cpu)
6152 struct rq *rq = cpu_rq(dead_cpu);
6153 struct task_struct *next;
6156 if (!rq->nr_running)
6158 update_rq_clock(rq);
6159 next = pick_next_task(rq, rq->curr);
6162 next->sched_class->put_prev_task(rq, next);
6163 migrate_dead(dead_cpu, next);
6167 #endif /* CONFIG_HOTPLUG_CPU */
6169 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6171 static struct ctl_table sd_ctl_dir[] = {
6173 .procname = "sched_domain",
6179 static struct ctl_table sd_ctl_root[] = {
6181 .ctl_name = CTL_KERN,
6182 .procname = "kernel",
6184 .child = sd_ctl_dir,
6189 static struct ctl_table *sd_alloc_ctl_entry(int n)
6191 struct ctl_table *entry =
6192 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6197 static void sd_free_ctl_entry(struct ctl_table **tablep)
6199 struct ctl_table *entry;
6202 * In the intermediate directories, both the child directory and
6203 * procname are dynamically allocated and could fail but the mode
6204 * will always be set. In the lowest directory the names are
6205 * static strings and all have proc handlers.
6207 for (entry = *tablep; entry->mode; entry++) {
6209 sd_free_ctl_entry(&entry->child);
6210 if (entry->proc_handler == NULL)
6211 kfree(entry->procname);
6219 set_table_entry(struct ctl_table *entry,
6220 const char *procname, void *data, int maxlen,
6221 mode_t mode, proc_handler *proc_handler)
6223 entry->procname = procname;
6225 entry->maxlen = maxlen;
6227 entry->proc_handler = proc_handler;
6230 static struct ctl_table *
6231 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6233 struct ctl_table *table = sd_alloc_ctl_entry(12);
6238 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6239 sizeof(long), 0644, proc_doulongvec_minmax);
6240 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6241 sizeof(long), 0644, proc_doulongvec_minmax);
6242 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6243 sizeof(int), 0644, proc_dointvec_minmax);
6244 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6245 sizeof(int), 0644, proc_dointvec_minmax);
6246 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6247 sizeof(int), 0644, proc_dointvec_minmax);
6248 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6249 sizeof(int), 0644, proc_dointvec_minmax);
6250 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6251 sizeof(int), 0644, proc_dointvec_minmax);
6252 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6253 sizeof(int), 0644, proc_dointvec_minmax);
6254 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6255 sizeof(int), 0644, proc_dointvec_minmax);
6256 set_table_entry(&table[9], "cache_nice_tries",
6257 &sd->cache_nice_tries,
6258 sizeof(int), 0644, proc_dointvec_minmax);
6259 set_table_entry(&table[10], "flags", &sd->flags,
6260 sizeof(int), 0644, proc_dointvec_minmax);
6261 /* &table[11] is terminator */
6266 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6268 struct ctl_table *entry, *table;
6269 struct sched_domain *sd;
6270 int domain_num = 0, i;
6273 for_each_domain(cpu, sd)
6275 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6280 for_each_domain(cpu, sd) {
6281 snprintf(buf, 32, "domain%d", i);
6282 entry->procname = kstrdup(buf, GFP_KERNEL);
6284 entry->child = sd_alloc_ctl_domain_table(sd);
6291 static struct ctl_table_header *sd_sysctl_header;
6292 static void register_sched_domain_sysctl(void)
6294 int i, cpu_num = num_online_cpus();
6295 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6298 WARN_ON(sd_ctl_dir[0].child);
6299 sd_ctl_dir[0].child = entry;
6304 for_each_online_cpu(i) {
6305 snprintf(buf, 32, "cpu%d", i);
6306 entry->procname = kstrdup(buf, GFP_KERNEL);
6308 entry->child = sd_alloc_ctl_cpu_table(i);
6312 WARN_ON(sd_sysctl_header);
6313 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6316 /* may be called multiple times per register */
6317 static void unregister_sched_domain_sysctl(void)
6319 if (sd_sysctl_header)
6320 unregister_sysctl_table(sd_sysctl_header);
6321 sd_sysctl_header = NULL;
6322 if (sd_ctl_dir[0].child)
6323 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6326 static void register_sched_domain_sysctl(void)
6329 static void unregister_sched_domain_sysctl(void)
6334 static void set_rq_online(struct rq *rq)
6337 const struct sched_class *class;
6339 cpu_set(rq->cpu, rq->rd->online);
6342 for_each_class(class) {
6343 if (class->rq_online)
6344 class->rq_online(rq);
6349 static void set_rq_offline(struct rq *rq)
6352 const struct sched_class *class;
6354 for_each_class(class) {
6355 if (class->rq_offline)
6356 class->rq_offline(rq);
6359 cpu_clear(rq->cpu, rq->rd->online);
6365 * migration_call - callback that gets triggered when a CPU is added.
6366 * Here we can start up the necessary migration thread for the new CPU.
6368 static int __cpuinit
6369 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6371 struct task_struct *p;
6372 int cpu = (long)hcpu;
6373 unsigned long flags;
6378 case CPU_UP_PREPARE:
6379 case CPU_UP_PREPARE_FROZEN:
6380 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6383 kthread_bind(p, cpu);
6384 /* Must be high prio: stop_machine expects to yield to it. */
6385 rq = task_rq_lock(p, &flags);
6386 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6387 task_rq_unlock(rq, &flags);
6388 cpu_rq(cpu)->migration_thread = p;
6392 case CPU_ONLINE_FROZEN:
6393 /* Strictly unnecessary, as first user will wake it. */
6394 wake_up_process(cpu_rq(cpu)->migration_thread);
6396 /* Update our root-domain */
6398 spin_lock_irqsave(&rq->lock, flags);
6400 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6404 spin_unlock_irqrestore(&rq->lock, flags);
6407 #ifdef CONFIG_HOTPLUG_CPU
6408 case CPU_UP_CANCELED:
6409 case CPU_UP_CANCELED_FROZEN:
6410 if (!cpu_rq(cpu)->migration_thread)
6412 /* Unbind it from offline cpu so it can run. Fall thru. */
6413 kthread_bind(cpu_rq(cpu)->migration_thread,
6414 any_online_cpu(cpu_online_map));
6415 kthread_stop(cpu_rq(cpu)->migration_thread);
6416 cpu_rq(cpu)->migration_thread = NULL;
6420 case CPU_DEAD_FROZEN:
6421 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6422 migrate_live_tasks(cpu);
6424 kthread_stop(rq->migration_thread);
6425 rq->migration_thread = NULL;
6426 /* Idle task back to normal (off runqueue, low prio) */
6427 spin_lock_irq(&rq->lock);
6428 update_rq_clock(rq);
6429 deactivate_task(rq, rq->idle, 0);
6430 rq->idle->static_prio = MAX_PRIO;
6431 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6432 rq->idle->sched_class = &idle_sched_class;
6433 migrate_dead_tasks(cpu);
6434 spin_unlock_irq(&rq->lock);
6436 migrate_nr_uninterruptible(rq);
6437 BUG_ON(rq->nr_running != 0);
6440 * No need to migrate the tasks: it was best-effort if
6441 * they didn't take sched_hotcpu_mutex. Just wake up
6444 spin_lock_irq(&rq->lock);
6445 while (!list_empty(&rq->migration_queue)) {
6446 struct migration_req *req;
6448 req = list_entry(rq->migration_queue.next,
6449 struct migration_req, list);
6450 list_del_init(&req->list);
6451 complete(&req->done);
6453 spin_unlock_irq(&rq->lock);
6457 case CPU_DYING_FROZEN:
6458 /* Update our root-domain */
6460 spin_lock_irqsave(&rq->lock, flags);
6462 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6465 spin_unlock_irqrestore(&rq->lock, flags);
6472 /* Register at highest priority so that task migration (migrate_all_tasks)
6473 * happens before everything else.
6475 static struct notifier_block __cpuinitdata migration_notifier = {
6476 .notifier_call = migration_call,
6480 static int __init migration_init(void)
6482 void *cpu = (void *)(long)smp_processor_id();
6485 /* Start one for the boot CPU: */
6486 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6487 BUG_ON(err == NOTIFY_BAD);
6488 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6489 register_cpu_notifier(&migration_notifier);
6493 early_initcall(migration_init);
6498 #ifdef CONFIG_SCHED_DEBUG
6500 static inline const char *sd_level_to_string(enum sched_domain_level lvl)
6513 case SD_LV_ALLNODES:
6522 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6523 cpumask_t *groupmask)
6525 struct sched_group *group = sd->groups;
6528 cpulist_scnprintf(str, sizeof(str), sd->span);
6529 cpus_clear(*groupmask);
6531 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6533 if (!(sd->flags & SD_LOAD_BALANCE)) {
6534 printk("does not load-balance\n");
6536 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6541 printk(KERN_CONT "span %s level %s\n",
6542 str, sd_level_to_string(sd->level));
6544 if (!cpu_isset(cpu, sd->span)) {
6545 printk(KERN_ERR "ERROR: domain->span does not contain "
6548 if (!cpu_isset(cpu, group->cpumask)) {
6549 printk(KERN_ERR "ERROR: domain->groups does not contain"
6553 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6557 printk(KERN_ERR "ERROR: group is NULL\n");
6561 if (!group->__cpu_power) {
6562 printk(KERN_CONT "\n");
6563 printk(KERN_ERR "ERROR: domain->cpu_power not "
6568 if (!cpus_weight(group->cpumask)) {
6569 printk(KERN_CONT "\n");
6570 printk(KERN_ERR "ERROR: empty group\n");
6574 if (cpus_intersects(*groupmask, group->cpumask)) {
6575 printk(KERN_CONT "\n");
6576 printk(KERN_ERR "ERROR: repeated CPUs\n");
6580 cpus_or(*groupmask, *groupmask, group->cpumask);
6582 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6583 printk(KERN_CONT " %s", str);
6585 group = group->next;
6586 } while (group != sd->groups);
6587 printk(KERN_CONT "\n");
6589 if (!cpus_equal(sd->span, *groupmask))
6590 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6592 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6593 printk(KERN_ERR "ERROR: parent span is not a superset "
6594 "of domain->span\n");
6598 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6600 cpumask_t *groupmask;
6604 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6608 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6610 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6612 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6617 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6626 #else /* !CONFIG_SCHED_DEBUG */
6627 # define sched_domain_debug(sd, cpu) do { } while (0)
6628 #endif /* CONFIG_SCHED_DEBUG */
6630 static int sd_degenerate(struct sched_domain *sd)
6632 if (cpus_weight(sd->span) == 1)
6635 /* Following flags need at least 2 groups */
6636 if (sd->flags & (SD_LOAD_BALANCE |
6637 SD_BALANCE_NEWIDLE |
6641 SD_SHARE_PKG_RESOURCES)) {
6642 if (sd->groups != sd->groups->next)
6646 /* Following flags don't use groups */
6647 if (sd->flags & (SD_WAKE_IDLE |
6656 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6658 unsigned long cflags = sd->flags, pflags = parent->flags;
6660 if (sd_degenerate(parent))
6663 if (!cpus_equal(sd->span, parent->span))
6666 /* Does parent contain flags not in child? */
6667 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6668 if (cflags & SD_WAKE_AFFINE)
6669 pflags &= ~SD_WAKE_BALANCE;
6670 /* Flags needing groups don't count if only 1 group in parent */
6671 if (parent->groups == parent->groups->next) {
6672 pflags &= ~(SD_LOAD_BALANCE |
6673 SD_BALANCE_NEWIDLE |
6677 SD_SHARE_PKG_RESOURCES);
6679 if (~cflags & pflags)
6685 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6687 unsigned long flags;
6689 spin_lock_irqsave(&rq->lock, flags);
6692 struct root_domain *old_rd = rq->rd;
6694 if (cpu_isset(rq->cpu, old_rd->online))
6697 cpu_clear(rq->cpu, old_rd->span);
6699 if (atomic_dec_and_test(&old_rd->refcount))
6703 atomic_inc(&rd->refcount);
6706 cpu_set(rq->cpu, rd->span);
6707 if (cpu_isset(rq->cpu, cpu_online_map))
6710 spin_unlock_irqrestore(&rq->lock, flags);
6713 static void init_rootdomain(struct root_domain *rd)
6715 memset(rd, 0, sizeof(*rd));
6717 cpus_clear(rd->span);
6718 cpus_clear(rd->online);
6720 cpupri_init(&rd->cpupri);
6723 static void init_defrootdomain(void)
6725 init_rootdomain(&def_root_domain);
6726 atomic_set(&def_root_domain.refcount, 1);
6729 static struct root_domain *alloc_rootdomain(void)
6731 struct root_domain *rd;
6733 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6737 init_rootdomain(rd);
6743 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6744 * hold the hotplug lock.
6747 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6749 struct rq *rq = cpu_rq(cpu);
6750 struct sched_domain *tmp;
6752 /* Remove the sched domains which do not contribute to scheduling. */
6753 for (tmp = sd; tmp; tmp = tmp->parent) {
6754 struct sched_domain *parent = tmp->parent;
6757 if (sd_parent_degenerate(tmp, parent)) {
6758 tmp->parent = parent->parent;
6760 parent->parent->child = tmp;
6764 if (sd && sd_degenerate(sd)) {
6770 sched_domain_debug(sd, cpu);
6772 rq_attach_root(rq, rd);
6773 rcu_assign_pointer(rq->sd, sd);
6776 /* cpus with isolated domains */
6777 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6779 /* Setup the mask of cpus configured for isolated domains */
6780 static int __init isolated_cpu_setup(char *str)
6782 static int __initdata ints[NR_CPUS];
6785 str = get_options(str, ARRAY_SIZE(ints), ints);
6786 cpus_clear(cpu_isolated_map);
6787 for (i = 1; i <= ints[0]; i++)
6788 if (ints[i] < NR_CPUS)
6789 cpu_set(ints[i], cpu_isolated_map);
6793 __setup("isolcpus=", isolated_cpu_setup);
6796 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6797 * to a function which identifies what group(along with sched group) a CPU
6798 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6799 * (due to the fact that we keep track of groups covered with a cpumask_t).
6801 * init_sched_build_groups will build a circular linked list of the groups
6802 * covered by the given span, and will set each group's ->cpumask correctly,
6803 * and ->cpu_power to 0.
6806 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6807 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6808 struct sched_group **sg,
6809 cpumask_t *tmpmask),
6810 cpumask_t *covered, cpumask_t *tmpmask)
6812 struct sched_group *first = NULL, *last = NULL;
6815 cpus_clear(*covered);
6817 for_each_cpu_mask_nr(i, *span) {
6818 struct sched_group *sg;
6819 int group = group_fn(i, cpu_map, &sg, tmpmask);
6822 if (cpu_isset(i, *covered))
6825 cpus_clear(sg->cpumask);
6826 sg->__cpu_power = 0;
6828 for_each_cpu_mask_nr(j, *span) {
6829 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6832 cpu_set(j, *covered);
6833 cpu_set(j, sg->cpumask);
6844 #define SD_NODES_PER_DOMAIN 16
6849 * find_next_best_node - find the next node to include in a sched_domain
6850 * @node: node whose sched_domain we're building
6851 * @used_nodes: nodes already in the sched_domain
6853 * Find the next node to include in a given scheduling domain. Simply
6854 * finds the closest node not already in the @used_nodes map.
6856 * Should use nodemask_t.
6858 static int find_next_best_node(int node, nodemask_t *used_nodes)
6860 int i, n, val, min_val, best_node = 0;
6864 for (i = 0; i < nr_node_ids; i++) {
6865 /* Start at @node */
6866 n = (node + i) % nr_node_ids;
6868 if (!nr_cpus_node(n))
6871 /* Skip already used nodes */
6872 if (node_isset(n, *used_nodes))
6875 /* Simple min distance search */
6876 val = node_distance(node, n);
6878 if (val < min_val) {
6884 node_set(best_node, *used_nodes);
6889 * sched_domain_node_span - get a cpumask for a node's sched_domain
6890 * @node: node whose cpumask we're constructing
6891 * @span: resulting cpumask
6893 * Given a node, construct a good cpumask for its sched_domain to span. It
6894 * should be one that prevents unnecessary balancing, but also spreads tasks
6897 static void sched_domain_node_span(int node, cpumask_t *span)
6899 nodemask_t used_nodes;
6900 node_to_cpumask_ptr(nodemask, node);
6904 nodes_clear(used_nodes);
6906 cpus_or(*span, *span, *nodemask);
6907 node_set(node, used_nodes);
6909 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6910 int next_node = find_next_best_node(node, &used_nodes);
6912 node_to_cpumask_ptr_next(nodemask, next_node);
6913 cpus_or(*span, *span, *nodemask);
6916 #endif /* CONFIG_NUMA */
6918 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6921 * SMT sched-domains:
6923 #ifdef CONFIG_SCHED_SMT
6924 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6925 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6928 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6932 *sg = &per_cpu(sched_group_cpus, cpu);
6935 #endif /* CONFIG_SCHED_SMT */
6938 * multi-core sched-domains:
6940 #ifdef CONFIG_SCHED_MC
6941 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6942 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6943 #endif /* CONFIG_SCHED_MC */
6945 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6947 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6952 *mask = per_cpu(cpu_sibling_map, cpu);
6953 cpus_and(*mask, *mask, *cpu_map);
6954 group = first_cpu(*mask);
6956 *sg = &per_cpu(sched_group_core, group);
6959 #elif defined(CONFIG_SCHED_MC)
6961 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6965 *sg = &per_cpu(sched_group_core, cpu);
6970 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6971 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6974 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6978 #ifdef CONFIG_SCHED_MC
6979 *mask = cpu_coregroup_map(cpu);
6980 cpus_and(*mask, *mask, *cpu_map);
6981 group = first_cpu(*mask);
6982 #elif defined(CONFIG_SCHED_SMT)
6983 *mask = per_cpu(cpu_sibling_map, cpu);
6984 cpus_and(*mask, *mask, *cpu_map);
6985 group = first_cpu(*mask);
6990 *sg = &per_cpu(sched_group_phys, group);
6996 * The init_sched_build_groups can't handle what we want to do with node
6997 * groups, so roll our own. Now each node has its own list of groups which
6998 * gets dynamically allocated.
7000 static DEFINE_PER_CPU(struct sched_domain, node_domains);
7001 static struct sched_group ***sched_group_nodes_bycpu;
7003 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
7004 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
7006 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
7007 struct sched_group **sg, cpumask_t *nodemask)
7011 *nodemask = node_to_cpumask(cpu_to_node(cpu));
7012 cpus_and(*nodemask, *nodemask, *cpu_map);
7013 group = first_cpu(*nodemask);
7016 *sg = &per_cpu(sched_group_allnodes, group);
7020 static void init_numa_sched_groups_power(struct sched_group *group_head)
7022 struct sched_group *sg = group_head;
7028 for_each_cpu_mask_nr(j, sg->cpumask) {
7029 struct sched_domain *sd;
7031 sd = &per_cpu(phys_domains, j);
7032 if (j != first_cpu(sd->groups->cpumask)) {
7034 * Only add "power" once for each
7040 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7043 } while (sg != group_head);
7045 #endif /* CONFIG_NUMA */
7048 /* Free memory allocated for various sched_group structures */
7049 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7053 for_each_cpu_mask_nr(cpu, *cpu_map) {
7054 struct sched_group **sched_group_nodes
7055 = sched_group_nodes_bycpu[cpu];
7057 if (!sched_group_nodes)
7060 for (i = 0; i < nr_node_ids; i++) {
7061 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7063 *nodemask = node_to_cpumask(i);
7064 cpus_and(*nodemask, *nodemask, *cpu_map);
7065 if (cpus_empty(*nodemask))
7075 if (oldsg != sched_group_nodes[i])
7078 kfree(sched_group_nodes);
7079 sched_group_nodes_bycpu[cpu] = NULL;
7082 #else /* !CONFIG_NUMA */
7083 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7086 #endif /* CONFIG_NUMA */
7089 * Initialize sched groups cpu_power.
7091 * cpu_power indicates the capacity of sched group, which is used while
7092 * distributing the load between different sched groups in a sched domain.
7093 * Typically cpu_power for all the groups in a sched domain will be same unless
7094 * there are asymmetries in the topology. If there are asymmetries, group
7095 * having more cpu_power will pickup more load compared to the group having
7098 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7099 * the maximum number of tasks a group can handle in the presence of other idle
7100 * or lightly loaded groups in the same sched domain.
7102 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7104 struct sched_domain *child;
7105 struct sched_group *group;
7107 WARN_ON(!sd || !sd->groups);
7109 if (cpu != first_cpu(sd->groups->cpumask))
7114 sd->groups->__cpu_power = 0;
7117 * For perf policy, if the groups in child domain share resources
7118 * (for example cores sharing some portions of the cache hierarchy
7119 * or SMT), then set this domain groups cpu_power such that each group
7120 * can handle only one task, when there are other idle groups in the
7121 * same sched domain.
7123 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7125 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7126 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7131 * add cpu_power of each child group to this groups cpu_power
7133 group = child->groups;
7135 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7136 group = group->next;
7137 } while (group != child->groups);
7141 * Initializers for schedule domains
7142 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7145 #define SD_INIT(sd, type) sd_init_##type(sd)
7146 #define SD_INIT_FUNC(type) \
7147 static noinline void sd_init_##type(struct sched_domain *sd) \
7149 memset(sd, 0, sizeof(*sd)); \
7150 *sd = SD_##type##_INIT; \
7151 sd->level = SD_LV_##type; \
7156 SD_INIT_FUNC(ALLNODES)
7159 #ifdef CONFIG_SCHED_SMT
7160 SD_INIT_FUNC(SIBLING)
7162 #ifdef CONFIG_SCHED_MC
7167 * To minimize stack usage kmalloc room for cpumasks and share the
7168 * space as the usage in build_sched_domains() dictates. Used only
7169 * if the amount of space is significant.
7172 cpumask_t tmpmask; /* make this one first */
7175 cpumask_t this_sibling_map;
7176 cpumask_t this_core_map;
7178 cpumask_t send_covered;
7181 cpumask_t domainspan;
7183 cpumask_t notcovered;
7188 #define SCHED_CPUMASK_ALLOC 1
7189 #define SCHED_CPUMASK_FREE(v) kfree(v)
7190 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7192 #define SCHED_CPUMASK_ALLOC 0
7193 #define SCHED_CPUMASK_FREE(v)
7194 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7197 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7198 ((unsigned long)(a) + offsetof(struct allmasks, v))
7200 static int default_relax_domain_level = -1;
7202 static int __init setup_relax_domain_level(char *str)
7206 val = simple_strtoul(str, NULL, 0);
7207 if (val < SD_LV_MAX)
7208 default_relax_domain_level = val;
7212 __setup("relax_domain_level=", setup_relax_domain_level);
7214 static void set_domain_attribute(struct sched_domain *sd,
7215 struct sched_domain_attr *attr)
7219 if (!attr || attr->relax_domain_level < 0) {
7220 if (default_relax_domain_level < 0)
7223 request = default_relax_domain_level;
7225 request = attr->relax_domain_level;
7226 if (request < sd->level) {
7227 /* turn off idle balance on this domain */
7228 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7230 /* turn on idle balance on this domain */
7231 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7236 * Build sched domains for a given set of cpus and attach the sched domains
7237 * to the individual cpus
7239 static int __build_sched_domains(const cpumask_t *cpu_map,
7240 struct sched_domain_attr *attr)
7243 struct root_domain *rd;
7244 SCHED_CPUMASK_DECLARE(allmasks);
7247 struct sched_group **sched_group_nodes = NULL;
7248 int sd_allnodes = 0;
7251 * Allocate the per-node list of sched groups
7253 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
7255 if (!sched_group_nodes) {
7256 printk(KERN_WARNING "Can not alloc sched group node list\n");
7261 rd = alloc_rootdomain();
7263 printk(KERN_WARNING "Cannot alloc root domain\n");
7265 kfree(sched_group_nodes);
7270 #if SCHED_CPUMASK_ALLOC
7271 /* get space for all scratch cpumask variables */
7272 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
7274 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7277 kfree(sched_group_nodes);
7282 tmpmask = (cpumask_t *)allmasks;
7286 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7290 * Set up domains for cpus specified by the cpu_map.
7292 for_each_cpu_mask_nr(i, *cpu_map) {
7293 struct sched_domain *sd = NULL, *p;
7294 SCHED_CPUMASK_VAR(nodemask, allmasks);
7296 *nodemask = node_to_cpumask(cpu_to_node(i));
7297 cpus_and(*nodemask, *nodemask, *cpu_map);
7300 if (cpus_weight(*cpu_map) >
7301 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7302 sd = &per_cpu(allnodes_domains, i);
7303 SD_INIT(sd, ALLNODES);
7304 set_domain_attribute(sd, attr);
7305 sd->span = *cpu_map;
7306 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7312 sd = &per_cpu(node_domains, i);
7314 set_domain_attribute(sd, attr);
7315 sched_domain_node_span(cpu_to_node(i), &sd->span);
7319 cpus_and(sd->span, sd->span, *cpu_map);
7323 sd = &per_cpu(phys_domains, i);
7325 set_domain_attribute(sd, attr);
7326 sd->span = *nodemask;
7330 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7332 #ifdef CONFIG_SCHED_MC
7334 sd = &per_cpu(core_domains, i);
7336 set_domain_attribute(sd, attr);
7337 sd->span = cpu_coregroup_map(i);
7338 cpus_and(sd->span, sd->span, *cpu_map);
7341 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7344 #ifdef CONFIG_SCHED_SMT
7346 sd = &per_cpu(cpu_domains, i);
7347 SD_INIT(sd, SIBLING);
7348 set_domain_attribute(sd, attr);
7349 sd->span = per_cpu(cpu_sibling_map, i);
7350 cpus_and(sd->span, sd->span, *cpu_map);
7353 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7357 #ifdef CONFIG_SCHED_SMT
7358 /* Set up CPU (sibling) groups */
7359 for_each_cpu_mask_nr(i, *cpu_map) {
7360 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7361 SCHED_CPUMASK_VAR(send_covered, allmasks);
7363 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7364 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7365 if (i != first_cpu(*this_sibling_map))
7368 init_sched_build_groups(this_sibling_map, cpu_map,
7370 send_covered, tmpmask);
7374 #ifdef CONFIG_SCHED_MC
7375 /* Set up multi-core groups */
7376 for_each_cpu_mask_nr(i, *cpu_map) {
7377 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7378 SCHED_CPUMASK_VAR(send_covered, allmasks);
7380 *this_core_map = cpu_coregroup_map(i);
7381 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7382 if (i != first_cpu(*this_core_map))
7385 init_sched_build_groups(this_core_map, cpu_map,
7387 send_covered, tmpmask);
7391 /* Set up physical groups */
7392 for (i = 0; i < nr_node_ids; i++) {
7393 SCHED_CPUMASK_VAR(nodemask, allmasks);
7394 SCHED_CPUMASK_VAR(send_covered, allmasks);
7396 *nodemask = node_to_cpumask(i);
7397 cpus_and(*nodemask, *nodemask, *cpu_map);
7398 if (cpus_empty(*nodemask))
7401 init_sched_build_groups(nodemask, cpu_map,
7403 send_covered, tmpmask);
7407 /* Set up node groups */
7409 SCHED_CPUMASK_VAR(send_covered, allmasks);
7411 init_sched_build_groups(cpu_map, cpu_map,
7412 &cpu_to_allnodes_group,
7413 send_covered, tmpmask);
7416 for (i = 0; i < nr_node_ids; i++) {
7417 /* Set up node groups */
7418 struct sched_group *sg, *prev;
7419 SCHED_CPUMASK_VAR(nodemask, allmasks);
7420 SCHED_CPUMASK_VAR(domainspan, allmasks);
7421 SCHED_CPUMASK_VAR(covered, allmasks);
7424 *nodemask = node_to_cpumask(i);
7425 cpus_clear(*covered);
7427 cpus_and(*nodemask, *nodemask, *cpu_map);
7428 if (cpus_empty(*nodemask)) {
7429 sched_group_nodes[i] = NULL;
7433 sched_domain_node_span(i, domainspan);
7434 cpus_and(*domainspan, *domainspan, *cpu_map);
7436 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7438 printk(KERN_WARNING "Can not alloc domain group for "
7442 sched_group_nodes[i] = sg;
7443 for_each_cpu_mask_nr(j, *nodemask) {
7444 struct sched_domain *sd;
7446 sd = &per_cpu(node_domains, j);
7449 sg->__cpu_power = 0;
7450 sg->cpumask = *nodemask;
7452 cpus_or(*covered, *covered, *nodemask);
7455 for (j = 0; j < nr_node_ids; j++) {
7456 SCHED_CPUMASK_VAR(notcovered, allmasks);
7457 int n = (i + j) % nr_node_ids;
7458 node_to_cpumask_ptr(pnodemask, n);
7460 cpus_complement(*notcovered, *covered);
7461 cpus_and(*tmpmask, *notcovered, *cpu_map);
7462 cpus_and(*tmpmask, *tmpmask, *domainspan);
7463 if (cpus_empty(*tmpmask))
7466 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7467 if (cpus_empty(*tmpmask))
7470 sg = kmalloc_node(sizeof(struct sched_group),
7474 "Can not alloc domain group for node %d\n", j);
7477 sg->__cpu_power = 0;
7478 sg->cpumask = *tmpmask;
7479 sg->next = prev->next;
7480 cpus_or(*covered, *covered, *tmpmask);
7487 /* Calculate CPU power for physical packages and nodes */
7488 #ifdef CONFIG_SCHED_SMT
7489 for_each_cpu_mask_nr(i, *cpu_map) {
7490 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7492 init_sched_groups_power(i, sd);
7495 #ifdef CONFIG_SCHED_MC
7496 for_each_cpu_mask_nr(i, *cpu_map) {
7497 struct sched_domain *sd = &per_cpu(core_domains, i);
7499 init_sched_groups_power(i, sd);
7503 for_each_cpu_mask_nr(i, *cpu_map) {
7504 struct sched_domain *sd = &per_cpu(phys_domains, i);
7506 init_sched_groups_power(i, sd);
7510 for (i = 0; i < nr_node_ids; i++)
7511 init_numa_sched_groups_power(sched_group_nodes[i]);
7514 struct sched_group *sg;
7516 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7518 init_numa_sched_groups_power(sg);
7522 /* Attach the domains */
7523 for_each_cpu_mask_nr(i, *cpu_map) {
7524 struct sched_domain *sd;
7525 #ifdef CONFIG_SCHED_SMT
7526 sd = &per_cpu(cpu_domains, i);
7527 #elif defined(CONFIG_SCHED_MC)
7528 sd = &per_cpu(core_domains, i);
7530 sd = &per_cpu(phys_domains, i);
7532 cpu_attach_domain(sd, rd, i);
7535 SCHED_CPUMASK_FREE((void *)allmasks);
7540 free_sched_groups(cpu_map, tmpmask);
7541 SCHED_CPUMASK_FREE((void *)allmasks);
7546 static int build_sched_domains(const cpumask_t *cpu_map)
7548 return __build_sched_domains(cpu_map, NULL);
7551 static cpumask_t *doms_cur; /* current sched domains */
7552 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7553 static struct sched_domain_attr *dattr_cur;
7554 /* attribues of custom domains in 'doms_cur' */
7557 * Special case: If a kmalloc of a doms_cur partition (array of
7558 * cpumask_t) fails, then fallback to a single sched domain,
7559 * as determined by the single cpumask_t fallback_doms.
7561 static cpumask_t fallback_doms;
7563 void __attribute__((weak)) arch_update_cpu_topology(void)
7568 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7569 * For now this just excludes isolated cpus, but could be used to
7570 * exclude other special cases in the future.
7572 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7576 arch_update_cpu_topology();
7578 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7580 doms_cur = &fallback_doms;
7581 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7583 err = build_sched_domains(doms_cur);
7584 register_sched_domain_sysctl();
7589 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7592 free_sched_groups(cpu_map, tmpmask);
7596 * Detach sched domains from a group of cpus specified in cpu_map
7597 * These cpus will now be attached to the NULL domain
7599 static void detach_destroy_domains(const cpumask_t *cpu_map)
7604 unregister_sched_domain_sysctl();
7606 for_each_cpu_mask_nr(i, *cpu_map)
7607 cpu_attach_domain(NULL, &def_root_domain, i);
7608 synchronize_sched();
7609 arch_destroy_sched_domains(cpu_map, &tmpmask);
7612 /* handle null as "default" */
7613 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7614 struct sched_domain_attr *new, int idx_new)
7616 struct sched_domain_attr tmp;
7623 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7624 new ? (new + idx_new) : &tmp,
7625 sizeof(struct sched_domain_attr));
7629 * Partition sched domains as specified by the 'ndoms_new'
7630 * cpumasks in the array doms_new[] of cpumasks. This compares
7631 * doms_new[] to the current sched domain partitioning, doms_cur[].
7632 * It destroys each deleted domain and builds each new domain.
7634 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7635 * The masks don't intersect (don't overlap.) We should setup one
7636 * sched domain for each mask. CPUs not in any of the cpumasks will
7637 * not be load balanced. If the same cpumask appears both in the
7638 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7641 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7642 * ownership of it and will kfree it when done with it. If the caller
7643 * failed the kmalloc call, then it can pass in doms_new == NULL,
7644 * and partition_sched_domains() will fallback to the single partition
7645 * 'fallback_doms', it also forces the domains to be rebuilt.
7647 * Call with hotplug lock held
7649 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7650 struct sched_domain_attr *dattr_new)
7654 mutex_lock(&sched_domains_mutex);
7656 /* always unregister in case we don't destroy any domains */
7657 unregister_sched_domain_sysctl();
7659 if (doms_new == NULL)
7662 /* Destroy deleted domains */
7663 for (i = 0; i < ndoms_cur; i++) {
7664 for (j = 0; j < ndoms_new; j++) {
7665 if (cpus_equal(doms_cur[i], doms_new[j])
7666 && dattrs_equal(dattr_cur, i, dattr_new, j))
7669 /* no match - a current sched domain not in new doms_new[] */
7670 detach_destroy_domains(doms_cur + i);
7675 if (doms_new == NULL) {
7678 doms_new = &fallback_doms;
7679 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7683 /* Build new domains */
7684 for (i = 0; i < ndoms_new; i++) {
7685 for (j = 0; j < ndoms_cur; j++) {
7686 if (cpus_equal(doms_new[i], doms_cur[j])
7687 && dattrs_equal(dattr_new, i, dattr_cur, j))
7690 /* no match - add a new doms_new */
7691 __build_sched_domains(doms_new + i,
7692 dattr_new ? dattr_new + i : NULL);
7697 /* Remember the new sched domains */
7698 if (doms_cur != &fallback_doms)
7700 kfree(dattr_cur); /* kfree(NULL) is safe */
7701 doms_cur = doms_new;
7702 dattr_cur = dattr_new;
7703 ndoms_cur = ndoms_new;
7705 register_sched_domain_sysctl();
7707 mutex_unlock(&sched_domains_mutex);
7710 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7711 int arch_reinit_sched_domains(void)
7714 rebuild_sched_domains();
7719 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7723 if (buf[0] != '0' && buf[0] != '1')
7727 sched_smt_power_savings = (buf[0] == '1');
7729 sched_mc_power_savings = (buf[0] == '1');
7731 ret = arch_reinit_sched_domains();
7733 return ret ? ret : count;
7736 #ifdef CONFIG_SCHED_MC
7737 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7740 return sprintf(page, "%u\n", sched_mc_power_savings);
7742 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7743 const char *buf, size_t count)
7745 return sched_power_savings_store(buf, count, 0);
7747 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7748 sched_mc_power_savings_show,
7749 sched_mc_power_savings_store);
7752 #ifdef CONFIG_SCHED_SMT
7753 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7756 return sprintf(page, "%u\n", sched_smt_power_savings);
7758 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7759 const char *buf, size_t count)
7761 return sched_power_savings_store(buf, count, 1);
7763 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7764 sched_smt_power_savings_show,
7765 sched_smt_power_savings_store);
7768 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7772 #ifdef CONFIG_SCHED_SMT
7774 err = sysfs_create_file(&cls->kset.kobj,
7775 &attr_sched_smt_power_savings.attr);
7777 #ifdef CONFIG_SCHED_MC
7778 if (!err && mc_capable())
7779 err = sysfs_create_file(&cls->kset.kobj,
7780 &attr_sched_mc_power_savings.attr);
7784 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7786 #ifndef CONFIG_CPUSETS
7788 * Add online and remove offline CPUs from the scheduler domains.
7789 * When cpusets are enabled they take over this function.
7791 static int update_sched_domains(struct notifier_block *nfb,
7792 unsigned long action, void *hcpu)
7796 case CPU_ONLINE_FROZEN:
7798 case CPU_DEAD_FROZEN:
7799 partition_sched_domains(0, NULL, NULL);
7808 static int update_runtime(struct notifier_block *nfb,
7809 unsigned long action, void *hcpu)
7811 int cpu = (int)(long)hcpu;
7814 case CPU_DOWN_PREPARE:
7815 case CPU_DOWN_PREPARE_FROZEN:
7816 disable_runtime(cpu_rq(cpu));
7819 case CPU_DOWN_FAILED:
7820 case CPU_DOWN_FAILED_FROZEN:
7822 case CPU_ONLINE_FROZEN:
7823 enable_runtime(cpu_rq(cpu));
7831 void __init sched_init_smp(void)
7833 cpumask_t non_isolated_cpus;
7835 #if defined(CONFIG_NUMA)
7836 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7838 BUG_ON(sched_group_nodes_bycpu == NULL);
7841 mutex_lock(&sched_domains_mutex);
7842 arch_init_sched_domains(&cpu_online_map);
7843 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7844 if (cpus_empty(non_isolated_cpus))
7845 cpu_set(smp_processor_id(), non_isolated_cpus);
7846 mutex_unlock(&sched_domains_mutex);
7849 #ifndef CONFIG_CPUSETS
7850 /* XXX: Theoretical race here - CPU may be hotplugged now */
7851 hotcpu_notifier(update_sched_domains, 0);
7854 /* RT runtime code needs to handle some hotplug events */
7855 hotcpu_notifier(update_runtime, 0);
7859 /* Move init over to a non-isolated CPU */
7860 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7862 sched_init_granularity();
7865 void __init sched_init_smp(void)
7867 sched_init_granularity();
7869 #endif /* CONFIG_SMP */
7871 int in_sched_functions(unsigned long addr)
7873 return in_lock_functions(addr) ||
7874 (addr >= (unsigned long)__sched_text_start
7875 && addr < (unsigned long)__sched_text_end);
7878 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7880 cfs_rq->tasks_timeline = RB_ROOT;
7881 INIT_LIST_HEAD(&cfs_rq->tasks);
7882 #ifdef CONFIG_FAIR_GROUP_SCHED
7885 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7888 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7890 struct rt_prio_array *array;
7893 array = &rt_rq->active;
7894 for (i = 0; i < MAX_RT_PRIO; i++) {
7895 INIT_LIST_HEAD(array->queue + i);
7896 __clear_bit(i, array->bitmap);
7898 /* delimiter for bitsearch: */
7899 __set_bit(MAX_RT_PRIO, array->bitmap);
7901 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7902 rt_rq->highest_prio = MAX_RT_PRIO;
7905 rt_rq->rt_nr_migratory = 0;
7906 rt_rq->overloaded = 0;
7910 rt_rq->rt_throttled = 0;
7911 rt_rq->rt_runtime = 0;
7912 spin_lock_init(&rt_rq->rt_runtime_lock);
7914 #ifdef CONFIG_RT_GROUP_SCHED
7915 rt_rq->rt_nr_boosted = 0;
7920 #ifdef CONFIG_FAIR_GROUP_SCHED
7921 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7922 struct sched_entity *se, int cpu, int add,
7923 struct sched_entity *parent)
7925 struct rq *rq = cpu_rq(cpu);
7926 tg->cfs_rq[cpu] = cfs_rq;
7927 init_cfs_rq(cfs_rq, rq);
7930 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7933 /* se could be NULL for init_task_group */
7938 se->cfs_rq = &rq->cfs;
7940 se->cfs_rq = parent->my_q;
7943 se->load.weight = tg->shares;
7944 se->load.inv_weight = 0;
7945 se->parent = parent;
7949 #ifdef CONFIG_RT_GROUP_SCHED
7950 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7951 struct sched_rt_entity *rt_se, int cpu, int add,
7952 struct sched_rt_entity *parent)
7954 struct rq *rq = cpu_rq(cpu);
7956 tg->rt_rq[cpu] = rt_rq;
7957 init_rt_rq(rt_rq, rq);
7959 rt_rq->rt_se = rt_se;
7960 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7962 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7964 tg->rt_se[cpu] = rt_se;
7969 rt_se->rt_rq = &rq->rt;
7971 rt_se->rt_rq = parent->my_q;
7973 rt_se->my_q = rt_rq;
7974 rt_se->parent = parent;
7975 INIT_LIST_HEAD(&rt_se->run_list);
7979 void __init sched_init(void)
7982 unsigned long alloc_size = 0, ptr;
7984 #ifdef CONFIG_FAIR_GROUP_SCHED
7985 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7987 #ifdef CONFIG_RT_GROUP_SCHED
7988 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7990 #ifdef CONFIG_USER_SCHED
7994 * As sched_init() is called before page_alloc is setup,
7995 * we use alloc_bootmem().
7998 ptr = (unsigned long)alloc_bootmem(alloc_size);
8000 #ifdef CONFIG_FAIR_GROUP_SCHED
8001 init_task_group.se = (struct sched_entity **)ptr;
8002 ptr += nr_cpu_ids * sizeof(void **);
8004 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8005 ptr += nr_cpu_ids * sizeof(void **);
8007 #ifdef CONFIG_USER_SCHED
8008 root_task_group.se = (struct sched_entity **)ptr;
8009 ptr += nr_cpu_ids * sizeof(void **);
8011 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8012 ptr += nr_cpu_ids * sizeof(void **);
8013 #endif /* CONFIG_USER_SCHED */
8014 #endif /* CONFIG_FAIR_GROUP_SCHED */
8015 #ifdef CONFIG_RT_GROUP_SCHED
8016 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8017 ptr += nr_cpu_ids * sizeof(void **);
8019 init_task_group.rt_rq = (struct rt_rq **)ptr;
8020 ptr += nr_cpu_ids * sizeof(void **);
8022 #ifdef CONFIG_USER_SCHED
8023 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8024 ptr += nr_cpu_ids * sizeof(void **);
8026 root_task_group.rt_rq = (struct rt_rq **)ptr;
8027 ptr += nr_cpu_ids * sizeof(void **);
8028 #endif /* CONFIG_USER_SCHED */
8029 #endif /* CONFIG_RT_GROUP_SCHED */
8033 init_defrootdomain();
8036 init_rt_bandwidth(&def_rt_bandwidth,
8037 global_rt_period(), global_rt_runtime());
8039 #ifdef CONFIG_RT_GROUP_SCHED
8040 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8041 global_rt_period(), global_rt_runtime());
8042 #ifdef CONFIG_USER_SCHED
8043 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8044 global_rt_period(), RUNTIME_INF);
8045 #endif /* CONFIG_USER_SCHED */
8046 #endif /* CONFIG_RT_GROUP_SCHED */
8048 #ifdef CONFIG_GROUP_SCHED
8049 list_add(&init_task_group.list, &task_groups);
8050 INIT_LIST_HEAD(&init_task_group.children);
8052 #ifdef CONFIG_USER_SCHED
8053 INIT_LIST_HEAD(&root_task_group.children);
8054 init_task_group.parent = &root_task_group;
8055 list_add(&init_task_group.siblings, &root_task_group.children);
8056 #endif /* CONFIG_USER_SCHED */
8057 #endif /* CONFIG_GROUP_SCHED */
8059 for_each_possible_cpu(i) {
8063 spin_lock_init(&rq->lock);
8065 init_cfs_rq(&rq->cfs, rq);
8066 init_rt_rq(&rq->rt, rq);
8067 #ifdef CONFIG_FAIR_GROUP_SCHED
8068 init_task_group.shares = init_task_group_load;
8069 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8070 #ifdef CONFIG_CGROUP_SCHED
8072 * How much cpu bandwidth does init_task_group get?
8074 * In case of task-groups formed thr' the cgroup filesystem, it
8075 * gets 100% of the cpu resources in the system. This overall
8076 * system cpu resource is divided among the tasks of
8077 * init_task_group and its child task-groups in a fair manner,
8078 * based on each entity's (task or task-group's) weight
8079 * (se->load.weight).
8081 * In other words, if init_task_group has 10 tasks of weight
8082 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8083 * then A0's share of the cpu resource is:
8085 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8087 * We achieve this by letting init_task_group's tasks sit
8088 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8090 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8091 #elif defined CONFIG_USER_SCHED
8092 root_task_group.shares = NICE_0_LOAD;
8093 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8095 * In case of task-groups formed thr' the user id of tasks,
8096 * init_task_group represents tasks belonging to root user.
8097 * Hence it forms a sibling of all subsequent groups formed.
8098 * In this case, init_task_group gets only a fraction of overall
8099 * system cpu resource, based on the weight assigned to root
8100 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8101 * by letting tasks of init_task_group sit in a separate cfs_rq
8102 * (init_cfs_rq) and having one entity represent this group of
8103 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8105 init_tg_cfs_entry(&init_task_group,
8106 &per_cpu(init_cfs_rq, i),
8107 &per_cpu(init_sched_entity, i), i, 1,
8108 root_task_group.se[i]);
8111 #endif /* CONFIG_FAIR_GROUP_SCHED */
8113 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8114 #ifdef CONFIG_RT_GROUP_SCHED
8115 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8116 #ifdef CONFIG_CGROUP_SCHED
8117 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8118 #elif defined CONFIG_USER_SCHED
8119 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8120 init_tg_rt_entry(&init_task_group,
8121 &per_cpu(init_rt_rq, i),
8122 &per_cpu(init_sched_rt_entity, i), i, 1,
8123 root_task_group.rt_se[i]);
8127 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8128 rq->cpu_load[j] = 0;
8132 rq->active_balance = 0;
8133 rq->next_balance = jiffies;
8137 rq->migration_thread = NULL;
8138 INIT_LIST_HEAD(&rq->migration_queue);
8139 rq_attach_root(rq, &def_root_domain);
8142 atomic_set(&rq->nr_iowait, 0);
8145 set_load_weight(&init_task);
8147 #ifdef CONFIG_PREEMPT_NOTIFIERS
8148 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8152 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8155 #ifdef CONFIG_RT_MUTEXES
8156 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8160 * The boot idle thread does lazy MMU switching as well:
8162 atomic_inc(&init_mm.mm_count);
8163 enter_lazy_tlb(&init_mm, current);
8166 * Make us the idle thread. Technically, schedule() should not be
8167 * called from this thread, however somewhere below it might be,
8168 * but because we are the idle thread, we just pick up running again
8169 * when this runqueue becomes "idle".
8171 init_idle(current, smp_processor_id());
8173 * During early bootup we pretend to be a normal task:
8175 current->sched_class = &fair_sched_class;
8177 scheduler_running = 1;
8180 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8181 void __might_sleep(char *file, int line)
8184 static unsigned long prev_jiffy; /* ratelimiting */
8186 if ((in_atomic() || irqs_disabled()) &&
8187 system_state == SYSTEM_RUNNING && !oops_in_progress) {
8188 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8190 prev_jiffy = jiffies;
8191 printk(KERN_ERR "BUG: sleeping function called from invalid"
8192 " context at %s:%d\n", file, line);
8193 printk("in_atomic():%d, irqs_disabled():%d\n",
8194 in_atomic(), irqs_disabled());
8195 debug_show_held_locks(current);
8196 if (irqs_disabled())
8197 print_irqtrace_events(current);
8202 EXPORT_SYMBOL(__might_sleep);
8205 #ifdef CONFIG_MAGIC_SYSRQ
8206 static void normalize_task(struct rq *rq, struct task_struct *p)
8210 update_rq_clock(rq);
8211 on_rq = p->se.on_rq;
8213 deactivate_task(rq, p, 0);
8214 __setscheduler(rq, p, SCHED_NORMAL, 0);
8216 activate_task(rq, p, 0);
8217 resched_task(rq->curr);
8221 void normalize_rt_tasks(void)
8223 struct task_struct *g, *p;
8224 unsigned long flags;
8227 read_lock_irqsave(&tasklist_lock, flags);
8228 do_each_thread(g, p) {
8230 * Only normalize user tasks:
8235 p->se.exec_start = 0;
8236 #ifdef CONFIG_SCHEDSTATS
8237 p->se.wait_start = 0;
8238 p->se.sleep_start = 0;
8239 p->se.block_start = 0;
8244 * Renice negative nice level userspace
8247 if (TASK_NICE(p) < 0 && p->mm)
8248 set_user_nice(p, 0);
8252 spin_lock(&p->pi_lock);
8253 rq = __task_rq_lock(p);
8255 normalize_task(rq, p);
8257 __task_rq_unlock(rq);
8258 spin_unlock(&p->pi_lock);
8259 } while_each_thread(g, p);
8261 read_unlock_irqrestore(&tasklist_lock, flags);
8264 #endif /* CONFIG_MAGIC_SYSRQ */
8268 * These functions are only useful for the IA64 MCA handling.
8270 * They can only be called when the whole system has been
8271 * stopped - every CPU needs to be quiescent, and no scheduling
8272 * activity can take place. Using them for anything else would
8273 * be a serious bug, and as a result, they aren't even visible
8274 * under any other configuration.
8278 * curr_task - return the current task for a given cpu.
8279 * @cpu: the processor in question.
8281 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8283 struct task_struct *curr_task(int cpu)
8285 return cpu_curr(cpu);
8289 * set_curr_task - set the current task for a given cpu.
8290 * @cpu: the processor in question.
8291 * @p: the task pointer to set.
8293 * Description: This function must only be used when non-maskable interrupts
8294 * are serviced on a separate stack. It allows the architecture to switch the
8295 * notion of the current task on a cpu in a non-blocking manner. This function
8296 * must be called with all CPU's synchronized, and interrupts disabled, the
8297 * and caller must save the original value of the current task (see
8298 * curr_task() above) and restore that value before reenabling interrupts and
8299 * re-starting the system.
8301 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8303 void set_curr_task(int cpu, struct task_struct *p)
8310 #ifdef CONFIG_FAIR_GROUP_SCHED
8311 static void free_fair_sched_group(struct task_group *tg)
8315 for_each_possible_cpu(i) {
8317 kfree(tg->cfs_rq[i]);
8327 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8329 struct cfs_rq *cfs_rq;
8330 struct sched_entity *se, *parent_se;
8334 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8337 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8341 tg->shares = NICE_0_LOAD;
8343 for_each_possible_cpu(i) {
8346 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8347 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8351 se = kmalloc_node(sizeof(struct sched_entity),
8352 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8356 parent_se = parent ? parent->se[i] : NULL;
8357 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
8366 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8368 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8369 &cpu_rq(cpu)->leaf_cfs_rq_list);
8372 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8374 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8376 #else /* !CONFG_FAIR_GROUP_SCHED */
8377 static inline void free_fair_sched_group(struct task_group *tg)
8382 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8387 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8391 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8394 #endif /* CONFIG_FAIR_GROUP_SCHED */
8396 #ifdef CONFIG_RT_GROUP_SCHED
8397 static void free_rt_sched_group(struct task_group *tg)
8401 destroy_rt_bandwidth(&tg->rt_bandwidth);
8403 for_each_possible_cpu(i) {
8405 kfree(tg->rt_rq[i]);
8407 kfree(tg->rt_se[i]);
8415 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8417 struct rt_rq *rt_rq;
8418 struct sched_rt_entity *rt_se, *parent_se;
8422 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8425 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8429 init_rt_bandwidth(&tg->rt_bandwidth,
8430 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8432 for_each_possible_cpu(i) {
8435 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8436 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8440 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8441 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8445 parent_se = parent ? parent->rt_se[i] : NULL;
8446 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
8455 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8457 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8458 &cpu_rq(cpu)->leaf_rt_rq_list);
8461 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8463 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8465 #else /* !CONFIG_RT_GROUP_SCHED */
8466 static inline void free_rt_sched_group(struct task_group *tg)
8471 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8476 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8480 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8483 #endif /* CONFIG_RT_GROUP_SCHED */
8485 #ifdef CONFIG_GROUP_SCHED
8486 static void free_sched_group(struct task_group *tg)
8488 free_fair_sched_group(tg);
8489 free_rt_sched_group(tg);
8493 /* allocate runqueue etc for a new task group */
8494 struct task_group *sched_create_group(struct task_group *parent)
8496 struct task_group *tg;
8497 unsigned long flags;
8500 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8502 return ERR_PTR(-ENOMEM);
8504 if (!alloc_fair_sched_group(tg, parent))
8507 if (!alloc_rt_sched_group(tg, parent))
8510 spin_lock_irqsave(&task_group_lock, flags);
8511 for_each_possible_cpu(i) {
8512 register_fair_sched_group(tg, i);
8513 register_rt_sched_group(tg, i);
8515 list_add_rcu(&tg->list, &task_groups);
8517 WARN_ON(!parent); /* root should already exist */
8519 tg->parent = parent;
8520 INIT_LIST_HEAD(&tg->children);
8521 list_add_rcu(&tg->siblings, &parent->children);
8522 spin_unlock_irqrestore(&task_group_lock, flags);
8527 free_sched_group(tg);
8528 return ERR_PTR(-ENOMEM);
8531 /* rcu callback to free various structures associated with a task group */
8532 static void free_sched_group_rcu(struct rcu_head *rhp)
8534 /* now it should be safe to free those cfs_rqs */
8535 free_sched_group(container_of(rhp, struct task_group, rcu));
8538 /* Destroy runqueue etc associated with a task group */
8539 void sched_destroy_group(struct task_group *tg)
8541 unsigned long flags;
8544 spin_lock_irqsave(&task_group_lock, flags);
8545 for_each_possible_cpu(i) {
8546 unregister_fair_sched_group(tg, i);
8547 unregister_rt_sched_group(tg, i);
8549 list_del_rcu(&tg->list);
8550 list_del_rcu(&tg->siblings);
8551 spin_unlock_irqrestore(&task_group_lock, flags);
8553 /* wait for possible concurrent references to cfs_rqs complete */
8554 call_rcu(&tg->rcu, free_sched_group_rcu);
8557 /* change task's runqueue when it moves between groups.
8558 * The caller of this function should have put the task in its new group
8559 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8560 * reflect its new group.
8562 void sched_move_task(struct task_struct *tsk)
8565 unsigned long flags;
8568 rq = task_rq_lock(tsk, &flags);
8570 update_rq_clock(rq);
8572 running = task_current(rq, tsk);
8573 on_rq = tsk->se.on_rq;
8576 dequeue_task(rq, tsk, 0);
8577 if (unlikely(running))
8578 tsk->sched_class->put_prev_task(rq, tsk);
8580 set_task_rq(tsk, task_cpu(tsk));
8582 #ifdef CONFIG_FAIR_GROUP_SCHED
8583 if (tsk->sched_class->moved_group)
8584 tsk->sched_class->moved_group(tsk);
8587 if (unlikely(running))
8588 tsk->sched_class->set_curr_task(rq);
8590 enqueue_task(rq, tsk, 0);
8592 task_rq_unlock(rq, &flags);
8594 #endif /* CONFIG_GROUP_SCHED */
8596 #ifdef CONFIG_FAIR_GROUP_SCHED
8597 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8599 struct cfs_rq *cfs_rq = se->cfs_rq;
8604 dequeue_entity(cfs_rq, se, 0);
8606 se->load.weight = shares;
8607 se->load.inv_weight = 0;
8610 enqueue_entity(cfs_rq, se, 0);
8613 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8615 struct cfs_rq *cfs_rq = se->cfs_rq;
8616 struct rq *rq = cfs_rq->rq;
8617 unsigned long flags;
8619 spin_lock_irqsave(&rq->lock, flags);
8620 __set_se_shares(se, shares);
8621 spin_unlock_irqrestore(&rq->lock, flags);
8624 static DEFINE_MUTEX(shares_mutex);
8626 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8629 unsigned long flags;
8632 * We can't change the weight of the root cgroup.
8637 if (shares < MIN_SHARES)
8638 shares = MIN_SHARES;
8639 else if (shares > MAX_SHARES)
8640 shares = MAX_SHARES;
8642 mutex_lock(&shares_mutex);
8643 if (tg->shares == shares)
8646 spin_lock_irqsave(&task_group_lock, flags);
8647 for_each_possible_cpu(i)
8648 unregister_fair_sched_group(tg, i);
8649 list_del_rcu(&tg->siblings);
8650 spin_unlock_irqrestore(&task_group_lock, flags);
8652 /* wait for any ongoing reference to this group to finish */
8653 synchronize_sched();
8656 * Now we are free to modify the group's share on each cpu
8657 * w/o tripping rebalance_share or load_balance_fair.
8659 tg->shares = shares;
8660 for_each_possible_cpu(i) {
8664 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8665 set_se_shares(tg->se[i], shares);
8669 * Enable load balance activity on this group, by inserting it back on
8670 * each cpu's rq->leaf_cfs_rq_list.
8672 spin_lock_irqsave(&task_group_lock, flags);
8673 for_each_possible_cpu(i)
8674 register_fair_sched_group(tg, i);
8675 list_add_rcu(&tg->siblings, &tg->parent->children);
8676 spin_unlock_irqrestore(&task_group_lock, flags);
8678 mutex_unlock(&shares_mutex);
8682 unsigned long sched_group_shares(struct task_group *tg)
8688 #ifdef CONFIG_RT_GROUP_SCHED
8690 * Ensure that the real time constraints are schedulable.
8692 static DEFINE_MUTEX(rt_constraints_mutex);
8694 static unsigned long to_ratio(u64 period, u64 runtime)
8696 if (runtime == RUNTIME_INF)
8699 return div64_u64(runtime << 16, period);
8702 #ifdef CONFIG_CGROUP_SCHED
8703 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8705 struct task_group *tgi, *parent = tg->parent;
8706 unsigned long total = 0;
8709 if (global_rt_period() < period)
8712 return to_ratio(period, runtime) <
8713 to_ratio(global_rt_period(), global_rt_runtime());
8716 if (ktime_to_ns(parent->rt_bandwidth.rt_period) < period)
8720 list_for_each_entry_rcu(tgi, &parent->children, siblings) {
8724 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8725 tgi->rt_bandwidth.rt_runtime);
8729 return total + to_ratio(period, runtime) <=
8730 to_ratio(ktime_to_ns(parent->rt_bandwidth.rt_period),
8731 parent->rt_bandwidth.rt_runtime);
8733 #elif defined CONFIG_USER_SCHED
8734 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8736 struct task_group *tgi;
8737 unsigned long total = 0;
8738 unsigned long global_ratio =
8739 to_ratio(global_rt_period(), global_rt_runtime());
8742 list_for_each_entry_rcu(tgi, &task_groups, list) {
8746 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8747 tgi->rt_bandwidth.rt_runtime);
8751 return total + to_ratio(period, runtime) < global_ratio;
8755 /* Must be called with tasklist_lock held */
8756 static inline int tg_has_rt_tasks(struct task_group *tg)
8758 struct task_struct *g, *p;
8759 do_each_thread(g, p) {
8760 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8762 } while_each_thread(g, p);
8766 static int tg_set_bandwidth(struct task_group *tg,
8767 u64 rt_period, u64 rt_runtime)
8771 mutex_lock(&rt_constraints_mutex);
8772 read_lock(&tasklist_lock);
8773 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
8777 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8782 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8783 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8784 tg->rt_bandwidth.rt_runtime = rt_runtime;
8786 for_each_possible_cpu(i) {
8787 struct rt_rq *rt_rq = tg->rt_rq[i];
8789 spin_lock(&rt_rq->rt_runtime_lock);
8790 rt_rq->rt_runtime = rt_runtime;
8791 spin_unlock(&rt_rq->rt_runtime_lock);
8793 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8795 read_unlock(&tasklist_lock);
8796 mutex_unlock(&rt_constraints_mutex);
8801 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8803 u64 rt_runtime, rt_period;
8805 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8806 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8807 if (rt_runtime_us < 0)
8808 rt_runtime = RUNTIME_INF;
8810 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8813 long sched_group_rt_runtime(struct task_group *tg)
8817 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8820 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8821 do_div(rt_runtime_us, NSEC_PER_USEC);
8822 return rt_runtime_us;
8825 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8827 u64 rt_runtime, rt_period;
8829 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8830 rt_runtime = tg->rt_bandwidth.rt_runtime;
8835 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8838 long sched_group_rt_period(struct task_group *tg)
8842 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8843 do_div(rt_period_us, NSEC_PER_USEC);
8844 return rt_period_us;
8847 static int sched_rt_global_constraints(void)
8849 struct task_group *tg = &root_task_group;
8850 u64 rt_runtime, rt_period;
8853 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8854 rt_runtime = tg->rt_bandwidth.rt_runtime;
8856 mutex_lock(&rt_constraints_mutex);
8857 if (!__rt_schedulable(tg, rt_period, rt_runtime))
8859 mutex_unlock(&rt_constraints_mutex);
8863 #else /* !CONFIG_RT_GROUP_SCHED */
8864 static int sched_rt_global_constraints(void)
8866 unsigned long flags;
8869 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8870 for_each_possible_cpu(i) {
8871 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8873 spin_lock(&rt_rq->rt_runtime_lock);
8874 rt_rq->rt_runtime = global_rt_runtime();
8875 spin_unlock(&rt_rq->rt_runtime_lock);
8877 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8881 #endif /* CONFIG_RT_GROUP_SCHED */
8883 int sched_rt_handler(struct ctl_table *table, int write,
8884 struct file *filp, void __user *buffer, size_t *lenp,
8888 int old_period, old_runtime;
8889 static DEFINE_MUTEX(mutex);
8892 old_period = sysctl_sched_rt_period;
8893 old_runtime = sysctl_sched_rt_runtime;
8895 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8897 if (!ret && write) {
8898 ret = sched_rt_global_constraints();
8900 sysctl_sched_rt_period = old_period;
8901 sysctl_sched_rt_runtime = old_runtime;
8903 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8904 def_rt_bandwidth.rt_period =
8905 ns_to_ktime(global_rt_period());
8908 mutex_unlock(&mutex);
8913 #ifdef CONFIG_CGROUP_SCHED
8915 /* return corresponding task_group object of a cgroup */
8916 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8918 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8919 struct task_group, css);
8922 static struct cgroup_subsys_state *
8923 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8925 struct task_group *tg, *parent;
8927 if (!cgrp->parent) {
8928 /* This is early initialization for the top cgroup */
8929 init_task_group.css.cgroup = cgrp;
8930 return &init_task_group.css;
8933 parent = cgroup_tg(cgrp->parent);
8934 tg = sched_create_group(parent);
8936 return ERR_PTR(-ENOMEM);
8938 /* Bind the cgroup to task_group object we just created */
8939 tg->css.cgroup = cgrp;
8945 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8947 struct task_group *tg = cgroup_tg(cgrp);
8949 sched_destroy_group(tg);
8953 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8954 struct task_struct *tsk)
8956 #ifdef CONFIG_RT_GROUP_SCHED
8957 /* Don't accept realtime tasks when there is no way for them to run */
8958 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
8961 /* We don't support RT-tasks being in separate groups */
8962 if (tsk->sched_class != &fair_sched_class)
8970 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8971 struct cgroup *old_cont, struct task_struct *tsk)
8973 sched_move_task(tsk);
8976 #ifdef CONFIG_FAIR_GROUP_SCHED
8977 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8980 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8983 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8985 struct task_group *tg = cgroup_tg(cgrp);
8987 return (u64) tg->shares;
8989 #endif /* CONFIG_FAIR_GROUP_SCHED */
8991 #ifdef CONFIG_RT_GROUP_SCHED
8992 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8995 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8998 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9000 return sched_group_rt_runtime(cgroup_tg(cgrp));
9003 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9006 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9009 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9011 return sched_group_rt_period(cgroup_tg(cgrp));
9013 #endif /* CONFIG_RT_GROUP_SCHED */
9015 static struct cftype cpu_files[] = {
9016 #ifdef CONFIG_FAIR_GROUP_SCHED
9019 .read_u64 = cpu_shares_read_u64,
9020 .write_u64 = cpu_shares_write_u64,
9023 #ifdef CONFIG_RT_GROUP_SCHED
9025 .name = "rt_runtime_us",
9026 .read_s64 = cpu_rt_runtime_read,
9027 .write_s64 = cpu_rt_runtime_write,
9030 .name = "rt_period_us",
9031 .read_u64 = cpu_rt_period_read_uint,
9032 .write_u64 = cpu_rt_period_write_uint,
9037 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9039 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9042 struct cgroup_subsys cpu_cgroup_subsys = {
9044 .create = cpu_cgroup_create,
9045 .destroy = cpu_cgroup_destroy,
9046 .can_attach = cpu_cgroup_can_attach,
9047 .attach = cpu_cgroup_attach,
9048 .populate = cpu_cgroup_populate,
9049 .subsys_id = cpu_cgroup_subsys_id,
9053 #endif /* CONFIG_CGROUP_SCHED */
9055 #ifdef CONFIG_CGROUP_CPUACCT
9058 * CPU accounting code for task groups.
9060 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9061 * (balbir@in.ibm.com).
9064 /* track cpu usage of a group of tasks */
9066 struct cgroup_subsys_state css;
9067 /* cpuusage holds pointer to a u64-type object on every cpu */
9071 struct cgroup_subsys cpuacct_subsys;
9073 /* return cpu accounting group corresponding to this container */
9074 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9076 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9077 struct cpuacct, css);
9080 /* return cpu accounting group to which this task belongs */
9081 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9083 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9084 struct cpuacct, css);
9087 /* create a new cpu accounting group */
9088 static struct cgroup_subsys_state *cpuacct_create(
9089 struct cgroup_subsys *ss, struct cgroup *cgrp)
9091 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9094 return ERR_PTR(-ENOMEM);
9096 ca->cpuusage = alloc_percpu(u64);
9097 if (!ca->cpuusage) {
9099 return ERR_PTR(-ENOMEM);
9105 /* destroy an existing cpu accounting group */
9107 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9109 struct cpuacct *ca = cgroup_ca(cgrp);
9111 free_percpu(ca->cpuusage);
9115 /* return total cpu usage (in nanoseconds) of a group */
9116 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9118 struct cpuacct *ca = cgroup_ca(cgrp);
9119 u64 totalcpuusage = 0;
9122 for_each_possible_cpu(i) {
9123 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9126 * Take rq->lock to make 64-bit addition safe on 32-bit
9129 spin_lock_irq(&cpu_rq(i)->lock);
9130 totalcpuusage += *cpuusage;
9131 spin_unlock_irq(&cpu_rq(i)->lock);
9134 return totalcpuusage;
9137 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9140 struct cpuacct *ca = cgroup_ca(cgrp);
9149 for_each_possible_cpu(i) {
9150 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9152 spin_lock_irq(&cpu_rq(i)->lock);
9154 spin_unlock_irq(&cpu_rq(i)->lock);
9160 static struct cftype files[] = {
9163 .read_u64 = cpuusage_read,
9164 .write_u64 = cpuusage_write,
9168 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9170 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9174 * charge this task's execution time to its accounting group.
9176 * called with rq->lock held.
9178 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9182 if (!cpuacct_subsys.active)
9187 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
9189 *cpuusage += cputime;
9193 struct cgroup_subsys cpuacct_subsys = {
9195 .create = cpuacct_create,
9196 .destroy = cpuacct_destroy,
9197 .populate = cpuacct_populate,
9198 .subsys_id = cpuacct_subsys_id,
9200 #endif /* CONFIG_CGROUP_CPUACCT */