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 return sysctl_sched_rt_runtime >= 0;
212 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
216 if (rt_bandwidth_enabled() && rt_b->rt_runtime == RUNTIME_INF)
219 if (hrtimer_active(&rt_b->rt_period_timer))
222 spin_lock(&rt_b->rt_runtime_lock);
224 if (hrtimer_active(&rt_b->rt_period_timer))
227 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
228 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
229 hrtimer_start(&rt_b->rt_period_timer,
230 rt_b->rt_period_timer.expires,
233 spin_unlock(&rt_b->rt_runtime_lock);
236 #ifdef CONFIG_RT_GROUP_SCHED
237 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
239 hrtimer_cancel(&rt_b->rt_period_timer);
244 * sched_domains_mutex serializes calls to arch_init_sched_domains,
245 * detach_destroy_domains and partition_sched_domains.
247 static DEFINE_MUTEX(sched_domains_mutex);
249 #ifdef CONFIG_GROUP_SCHED
251 #include <linux/cgroup.h>
255 static LIST_HEAD(task_groups);
257 /* task group related information */
259 #ifdef CONFIG_CGROUP_SCHED
260 struct cgroup_subsys_state css;
263 #ifdef CONFIG_FAIR_GROUP_SCHED
264 /* schedulable entities of this group on each cpu */
265 struct sched_entity **se;
266 /* runqueue "owned" by this group on each cpu */
267 struct cfs_rq **cfs_rq;
268 unsigned long shares;
271 #ifdef CONFIG_RT_GROUP_SCHED
272 struct sched_rt_entity **rt_se;
273 struct rt_rq **rt_rq;
275 struct rt_bandwidth rt_bandwidth;
279 struct list_head list;
281 struct task_group *parent;
282 struct list_head siblings;
283 struct list_head children;
286 #ifdef CONFIG_USER_SCHED
290 * Every UID task group (including init_task_group aka UID-0) will
291 * be a child to this group.
293 struct task_group root_task_group;
295 #ifdef CONFIG_FAIR_GROUP_SCHED
296 /* Default task group's sched entity on each cpu */
297 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
298 /* Default task group's cfs_rq on each cpu */
299 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
300 #endif /* CONFIG_FAIR_GROUP_SCHED */
302 #ifdef CONFIG_RT_GROUP_SCHED
303 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
304 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
305 #endif /* CONFIG_RT_GROUP_SCHED */
306 #else /* !CONFIG_USER_SCHED */
307 #define root_task_group init_task_group
308 #endif /* CONFIG_USER_SCHED */
310 /* task_group_lock serializes add/remove of task groups and also changes to
311 * a task group's cpu shares.
313 static DEFINE_SPINLOCK(task_group_lock);
315 #ifdef CONFIG_FAIR_GROUP_SCHED
316 #ifdef CONFIG_USER_SCHED
317 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
318 #else /* !CONFIG_USER_SCHED */
319 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
320 #endif /* CONFIG_USER_SCHED */
323 * A weight of 0 or 1 can cause arithmetics problems.
324 * A weight of a cfs_rq is the sum of weights of which entities
325 * are queued on this cfs_rq, so a weight of a entity should not be
326 * too large, so as the shares value of a task group.
327 * (The default weight is 1024 - so there's no practical
328 * limitation from this.)
331 #define MAX_SHARES (1UL << 18)
333 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
336 /* Default task group.
337 * Every task in system belong to this group at bootup.
339 struct task_group init_task_group;
341 /* return group to which a task belongs */
342 static inline struct task_group *task_group(struct task_struct *p)
344 struct task_group *tg;
346 #ifdef CONFIG_USER_SCHED
348 #elif defined(CONFIG_CGROUP_SCHED)
349 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
350 struct task_group, css);
352 tg = &init_task_group;
357 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
358 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
360 #ifdef CONFIG_FAIR_GROUP_SCHED
361 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
362 p->se.parent = task_group(p)->se[cpu];
365 #ifdef CONFIG_RT_GROUP_SCHED
366 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
367 p->rt.parent = task_group(p)->rt_se[cpu];
373 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
374 static inline struct task_group *task_group(struct task_struct *p)
379 #endif /* CONFIG_GROUP_SCHED */
381 /* CFS-related fields in a runqueue */
383 struct load_weight load;
384 unsigned long nr_running;
390 struct rb_root tasks_timeline;
391 struct rb_node *rb_leftmost;
393 struct list_head tasks;
394 struct list_head *balance_iterator;
397 * 'curr' points to currently running entity on this cfs_rq.
398 * It is set to NULL otherwise (i.e when none are currently running).
400 struct sched_entity *curr, *next;
402 unsigned long nr_spread_over;
404 #ifdef CONFIG_FAIR_GROUP_SCHED
405 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
408 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
409 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
410 * (like users, containers etc.)
412 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
413 * list is used during load balance.
415 struct list_head leaf_cfs_rq_list;
416 struct task_group *tg; /* group that "owns" this runqueue */
420 * the part of load.weight contributed by tasks
422 unsigned long task_weight;
425 * h_load = weight * f(tg)
427 * Where f(tg) is the recursive weight fraction assigned to
430 unsigned long h_load;
433 * this cpu's part of tg->shares
435 unsigned long shares;
438 * load.weight at the time we set shares
440 unsigned long rq_weight;
445 /* Real-Time classes' related field in a runqueue: */
447 struct rt_prio_array active;
448 unsigned long rt_nr_running;
449 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
450 int highest_prio; /* highest queued rt task prio */
453 unsigned long rt_nr_migratory;
459 /* Nests inside the rq lock: */
460 spinlock_t rt_runtime_lock;
462 #ifdef CONFIG_RT_GROUP_SCHED
463 unsigned long rt_nr_boosted;
466 struct list_head leaf_rt_rq_list;
467 struct task_group *tg;
468 struct sched_rt_entity *rt_se;
475 * We add the notion of a root-domain which will be used to define per-domain
476 * variables. Each exclusive cpuset essentially defines an island domain by
477 * fully partitioning the member cpus from any other cpuset. Whenever a new
478 * exclusive cpuset is created, we also create and attach a new root-domain
488 * The "RT overload" flag: it gets set if a CPU has more than
489 * one runnable RT task.
494 struct cpupri cpupri;
499 * By default the system creates a single root-domain with all cpus as
500 * members (mimicking the global state we have today).
502 static struct root_domain def_root_domain;
507 * This is the main, per-CPU runqueue data structure.
509 * Locking rule: those places that want to lock multiple runqueues
510 * (such as the load balancing or the thread migration code), lock
511 * acquire operations must be ordered by ascending &runqueue.
518 * nr_running and cpu_load should be in the same cacheline because
519 * remote CPUs use both these fields when doing load calculation.
521 unsigned long nr_running;
522 #define CPU_LOAD_IDX_MAX 5
523 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
524 unsigned char idle_at_tick;
526 unsigned long last_tick_seen;
527 unsigned char in_nohz_recently;
529 /* capture load from *all* tasks on this cpu: */
530 struct load_weight load;
531 unsigned long nr_load_updates;
537 #ifdef CONFIG_FAIR_GROUP_SCHED
538 /* list of leaf cfs_rq on this cpu: */
539 struct list_head leaf_cfs_rq_list;
541 #ifdef CONFIG_RT_GROUP_SCHED
542 struct list_head leaf_rt_rq_list;
546 * This is part of a global counter where only the total sum
547 * over all CPUs matters. A task can increase this counter on
548 * one CPU and if it got migrated afterwards it may decrease
549 * it on another CPU. Always updated under the runqueue lock:
551 unsigned long nr_uninterruptible;
553 struct task_struct *curr, *idle;
554 unsigned long next_balance;
555 struct mm_struct *prev_mm;
562 struct root_domain *rd;
563 struct sched_domain *sd;
565 /* For active balancing */
568 /* cpu of this runqueue: */
572 unsigned long avg_load_per_task;
574 struct task_struct *migration_thread;
575 struct list_head migration_queue;
578 #ifdef CONFIG_SCHED_HRTICK
580 int hrtick_csd_pending;
581 struct call_single_data hrtick_csd;
583 struct hrtimer hrtick_timer;
586 #ifdef CONFIG_SCHEDSTATS
588 struct sched_info rq_sched_info;
590 /* sys_sched_yield() stats */
591 unsigned int yld_exp_empty;
592 unsigned int yld_act_empty;
593 unsigned int yld_both_empty;
594 unsigned int yld_count;
596 /* schedule() stats */
597 unsigned int sched_switch;
598 unsigned int sched_count;
599 unsigned int sched_goidle;
601 /* try_to_wake_up() stats */
602 unsigned int ttwu_count;
603 unsigned int ttwu_local;
606 unsigned int bkl_count;
610 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
612 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
614 rq->curr->sched_class->check_preempt_curr(rq, p);
617 static inline int cpu_of(struct rq *rq)
627 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
628 * See detach_destroy_domains: synchronize_sched for details.
630 * The domain tree of any CPU may only be accessed from within
631 * preempt-disabled sections.
633 #define for_each_domain(cpu, __sd) \
634 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
636 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
637 #define this_rq() (&__get_cpu_var(runqueues))
638 #define task_rq(p) cpu_rq(task_cpu(p))
639 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
641 static inline void update_rq_clock(struct rq *rq)
643 rq->clock = sched_clock_cpu(cpu_of(rq));
647 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
649 #ifdef CONFIG_SCHED_DEBUG
650 # define const_debug __read_mostly
652 # define const_debug static const
658 * Returns true if the current cpu runqueue is locked.
659 * This interface allows printk to be called with the runqueue lock
660 * held and know whether or not it is OK to wake up the klogd.
662 int runqueue_is_locked(void)
665 struct rq *rq = cpu_rq(cpu);
668 ret = spin_is_locked(&rq->lock);
674 * Debugging: various feature bits
677 #define SCHED_FEAT(name, enabled) \
678 __SCHED_FEAT_##name ,
681 #include "sched_features.h"
686 #define SCHED_FEAT(name, enabled) \
687 (1UL << __SCHED_FEAT_##name) * enabled |
689 const_debug unsigned int sysctl_sched_features =
690 #include "sched_features.h"
695 #ifdef CONFIG_SCHED_DEBUG
696 #define SCHED_FEAT(name, enabled) \
699 static __read_mostly char *sched_feat_names[] = {
700 #include "sched_features.h"
706 static int sched_feat_open(struct inode *inode, struct file *filp)
708 filp->private_data = inode->i_private;
713 sched_feat_read(struct file *filp, char __user *ubuf,
714 size_t cnt, loff_t *ppos)
721 for (i = 0; sched_feat_names[i]; i++) {
722 len += strlen(sched_feat_names[i]);
726 buf = kmalloc(len + 2, GFP_KERNEL);
730 for (i = 0; sched_feat_names[i]; i++) {
731 if (sysctl_sched_features & (1UL << i))
732 r += sprintf(buf + r, "%s ", sched_feat_names[i]);
734 r += sprintf(buf + r, "NO_%s ", sched_feat_names[i]);
737 r += sprintf(buf + r, "\n");
738 WARN_ON(r >= len + 2);
740 r = simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
748 sched_feat_write(struct file *filp, const char __user *ubuf,
749 size_t cnt, loff_t *ppos)
759 if (copy_from_user(&buf, ubuf, cnt))
764 if (strncmp(buf, "NO_", 3) == 0) {
769 for (i = 0; sched_feat_names[i]; i++) {
770 int len = strlen(sched_feat_names[i]);
772 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
774 sysctl_sched_features &= ~(1UL << i);
776 sysctl_sched_features |= (1UL << i);
781 if (!sched_feat_names[i])
789 static struct file_operations sched_feat_fops = {
790 .open = sched_feat_open,
791 .read = sched_feat_read,
792 .write = sched_feat_write,
795 static __init int sched_init_debug(void)
797 debugfs_create_file("sched_features", 0644, NULL, NULL,
802 late_initcall(sched_init_debug);
806 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
809 * Number of tasks to iterate in a single balance run.
810 * Limited because this is done with IRQs disabled.
812 const_debug unsigned int sysctl_sched_nr_migrate = 32;
815 * ratelimit for updating the group shares.
818 unsigned int sysctl_sched_shares_ratelimit = 250000;
821 * period over which we measure -rt task cpu usage in us.
824 unsigned int sysctl_sched_rt_period = 1000000;
826 static __read_mostly int scheduler_running;
829 * part of the period that we allow rt tasks to run in us.
832 int sysctl_sched_rt_runtime = 950000;
834 static inline u64 global_rt_period(void)
836 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
839 static inline u64 global_rt_runtime(void)
841 if (sysctl_sched_rt_runtime < 0)
844 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
847 #ifndef prepare_arch_switch
848 # define prepare_arch_switch(next) do { } while (0)
850 #ifndef finish_arch_switch
851 # define finish_arch_switch(prev) do { } while (0)
854 static inline int task_current(struct rq *rq, struct task_struct *p)
856 return rq->curr == p;
859 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
860 static inline int task_running(struct rq *rq, struct task_struct *p)
862 return task_current(rq, p);
865 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
869 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
871 #ifdef CONFIG_DEBUG_SPINLOCK
872 /* this is a valid case when another task releases the spinlock */
873 rq->lock.owner = current;
876 * If we are tracking spinlock dependencies then we have to
877 * fix up the runqueue lock - which gets 'carried over' from
880 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
882 spin_unlock_irq(&rq->lock);
885 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
886 static inline int task_running(struct rq *rq, struct task_struct *p)
891 return task_current(rq, p);
895 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
899 * We can optimise this out completely for !SMP, because the
900 * SMP rebalancing from interrupt is the only thing that cares
905 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
906 spin_unlock_irq(&rq->lock);
908 spin_unlock(&rq->lock);
912 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
916 * After ->oncpu is cleared, the task can be moved to a different CPU.
917 * We must ensure this doesn't happen until the switch is completely
923 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
927 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
930 * __task_rq_lock - lock the runqueue a given task resides on.
931 * Must be called interrupts disabled.
933 static inline struct rq *__task_rq_lock(struct task_struct *p)
937 struct rq *rq = task_rq(p);
938 spin_lock(&rq->lock);
939 if (likely(rq == task_rq(p)))
941 spin_unlock(&rq->lock);
946 * task_rq_lock - lock the runqueue a given task resides on and disable
947 * interrupts. Note the ordering: we can safely lookup the task_rq without
948 * explicitly disabling preemption.
950 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
956 local_irq_save(*flags);
958 spin_lock(&rq->lock);
959 if (likely(rq == task_rq(p)))
961 spin_unlock_irqrestore(&rq->lock, *flags);
965 static void __task_rq_unlock(struct rq *rq)
968 spin_unlock(&rq->lock);
971 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
974 spin_unlock_irqrestore(&rq->lock, *flags);
978 * this_rq_lock - lock this runqueue and disable interrupts.
980 static struct rq *this_rq_lock(void)
987 spin_lock(&rq->lock);
992 #ifdef CONFIG_SCHED_HRTICK
994 * Use HR-timers to deliver accurate preemption points.
996 * Its all a bit involved since we cannot program an hrt while holding the
997 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1000 * When we get rescheduled we reprogram the hrtick_timer outside of the
1006 * - enabled by features
1007 * - hrtimer is actually high res
1009 static inline int hrtick_enabled(struct rq *rq)
1011 if (!sched_feat(HRTICK))
1013 if (!cpu_active(cpu_of(rq)))
1015 return hrtimer_is_hres_active(&rq->hrtick_timer);
1018 static void hrtick_clear(struct rq *rq)
1020 if (hrtimer_active(&rq->hrtick_timer))
1021 hrtimer_cancel(&rq->hrtick_timer);
1025 * High-resolution timer tick.
1026 * Runs from hardirq context with interrupts disabled.
1028 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1030 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1032 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1034 spin_lock(&rq->lock);
1035 update_rq_clock(rq);
1036 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1037 spin_unlock(&rq->lock);
1039 return HRTIMER_NORESTART;
1044 * called from hardirq (IPI) context
1046 static void __hrtick_start(void *arg)
1048 struct rq *rq = arg;
1050 spin_lock(&rq->lock);
1051 hrtimer_restart(&rq->hrtick_timer);
1052 rq->hrtick_csd_pending = 0;
1053 spin_unlock(&rq->lock);
1057 * Called to set the hrtick timer state.
1059 * called with rq->lock held and irqs disabled
1061 static void hrtick_start(struct rq *rq, u64 delay)
1063 struct hrtimer *timer = &rq->hrtick_timer;
1064 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1066 timer->expires = time;
1068 if (rq == this_rq()) {
1069 hrtimer_restart(timer);
1070 } else if (!rq->hrtick_csd_pending) {
1071 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd);
1072 rq->hrtick_csd_pending = 1;
1077 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1079 int cpu = (int)(long)hcpu;
1082 case CPU_UP_CANCELED:
1083 case CPU_UP_CANCELED_FROZEN:
1084 case CPU_DOWN_PREPARE:
1085 case CPU_DOWN_PREPARE_FROZEN:
1087 case CPU_DEAD_FROZEN:
1088 hrtick_clear(cpu_rq(cpu));
1095 static void init_hrtick(void)
1097 hotcpu_notifier(hotplug_hrtick, 0);
1101 * Called to set the hrtick timer state.
1103 * called with rq->lock held and irqs disabled
1105 static void hrtick_start(struct rq *rq, u64 delay)
1107 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay), HRTIMER_MODE_REL);
1110 static void init_hrtick(void)
1113 #endif /* CONFIG_SMP */
1115 static void init_rq_hrtick(struct rq *rq)
1118 rq->hrtick_csd_pending = 0;
1120 rq->hrtick_csd.flags = 0;
1121 rq->hrtick_csd.func = __hrtick_start;
1122 rq->hrtick_csd.info = rq;
1125 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1126 rq->hrtick_timer.function = hrtick;
1127 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1130 static inline void hrtick_clear(struct rq *rq)
1134 static inline void init_rq_hrtick(struct rq *rq)
1138 static inline void init_hrtick(void)
1144 * resched_task - mark a task 'to be rescheduled now'.
1146 * On UP this means the setting of the need_resched flag, on SMP it
1147 * might also involve a cross-CPU call to trigger the scheduler on
1152 #ifndef tsk_is_polling
1153 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1156 static void resched_task(struct task_struct *p)
1160 assert_spin_locked(&task_rq(p)->lock);
1162 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1165 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1168 if (cpu == smp_processor_id())
1171 /* NEED_RESCHED must be visible before we test polling */
1173 if (!tsk_is_polling(p))
1174 smp_send_reschedule(cpu);
1177 static void resched_cpu(int cpu)
1179 struct rq *rq = cpu_rq(cpu);
1180 unsigned long flags;
1182 if (!spin_trylock_irqsave(&rq->lock, flags))
1184 resched_task(cpu_curr(cpu));
1185 spin_unlock_irqrestore(&rq->lock, flags);
1190 * When add_timer_on() enqueues a timer into the timer wheel of an
1191 * idle CPU then this timer might expire before the next timer event
1192 * which is scheduled to wake up that CPU. In case of a completely
1193 * idle system the next event might even be infinite time into the
1194 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1195 * leaves the inner idle loop so the newly added timer is taken into
1196 * account when the CPU goes back to idle and evaluates the timer
1197 * wheel for the next timer event.
1199 void wake_up_idle_cpu(int cpu)
1201 struct rq *rq = cpu_rq(cpu);
1203 if (cpu == smp_processor_id())
1207 * This is safe, as this function is called with the timer
1208 * wheel base lock of (cpu) held. When the CPU is on the way
1209 * to idle and has not yet set rq->curr to idle then it will
1210 * be serialized on the timer wheel base lock and take the new
1211 * timer into account automatically.
1213 if (rq->curr != rq->idle)
1217 * We can set TIF_RESCHED on the idle task of the other CPU
1218 * lockless. The worst case is that the other CPU runs the
1219 * idle task through an additional NOOP schedule()
1221 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1223 /* NEED_RESCHED must be visible before we test polling */
1225 if (!tsk_is_polling(rq->idle))
1226 smp_send_reschedule(cpu);
1228 #endif /* CONFIG_NO_HZ */
1230 #else /* !CONFIG_SMP */
1231 static void resched_task(struct task_struct *p)
1233 assert_spin_locked(&task_rq(p)->lock);
1234 set_tsk_need_resched(p);
1236 #endif /* CONFIG_SMP */
1238 #if BITS_PER_LONG == 32
1239 # define WMULT_CONST (~0UL)
1241 # define WMULT_CONST (1UL << 32)
1244 #define WMULT_SHIFT 32
1247 * Shift right and round:
1249 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1252 * delta *= weight / lw
1254 static unsigned long
1255 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1256 struct load_weight *lw)
1260 if (!lw->inv_weight) {
1261 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1264 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1268 tmp = (u64)delta_exec * weight;
1270 * Check whether we'd overflow the 64-bit multiplication:
1272 if (unlikely(tmp > WMULT_CONST))
1273 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1276 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1278 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1281 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1287 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1294 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1295 * of tasks with abnormal "nice" values across CPUs the contribution that
1296 * each task makes to its run queue's load is weighted according to its
1297 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1298 * scaled version of the new time slice allocation that they receive on time
1302 #define WEIGHT_IDLEPRIO 2
1303 #define WMULT_IDLEPRIO (1 << 31)
1306 * Nice levels are multiplicative, with a gentle 10% change for every
1307 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1308 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1309 * that remained on nice 0.
1311 * The "10% effect" is relative and cumulative: from _any_ nice level,
1312 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1313 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1314 * If a task goes up by ~10% and another task goes down by ~10% then
1315 * the relative distance between them is ~25%.)
1317 static const int prio_to_weight[40] = {
1318 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1319 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1320 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1321 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1322 /* 0 */ 1024, 820, 655, 526, 423,
1323 /* 5 */ 335, 272, 215, 172, 137,
1324 /* 10 */ 110, 87, 70, 56, 45,
1325 /* 15 */ 36, 29, 23, 18, 15,
1329 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1331 * In cases where the weight does not change often, we can use the
1332 * precalculated inverse to speed up arithmetics by turning divisions
1333 * into multiplications:
1335 static const u32 prio_to_wmult[40] = {
1336 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1337 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1338 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1339 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1340 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1341 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1342 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1343 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1346 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1349 * runqueue iterator, to support SMP load-balancing between different
1350 * scheduling classes, without having to expose their internal data
1351 * structures to the load-balancing proper:
1353 struct rq_iterator {
1355 struct task_struct *(*start)(void *);
1356 struct task_struct *(*next)(void *);
1360 static unsigned long
1361 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1362 unsigned long max_load_move, struct sched_domain *sd,
1363 enum cpu_idle_type idle, int *all_pinned,
1364 int *this_best_prio, struct rq_iterator *iterator);
1367 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1368 struct sched_domain *sd, enum cpu_idle_type idle,
1369 struct rq_iterator *iterator);
1372 #ifdef CONFIG_CGROUP_CPUACCT
1373 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1375 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1378 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1380 update_load_add(&rq->load, load);
1383 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1385 update_load_sub(&rq->load, load);
1388 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1389 typedef int (*tg_visitor)(struct task_group *, void *);
1392 * Iterate the full tree, calling @down when first entering a node and @up when
1393 * leaving it for the final time.
1395 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1397 struct task_group *parent, *child;
1401 parent = &root_task_group;
1403 ret = (*down)(parent, data);
1406 list_for_each_entry_rcu(child, &parent->children, siblings) {
1413 ret = (*up)(parent, data);
1418 parent = parent->parent;
1427 static int tg_nop(struct task_group *tg, void *data)
1434 static unsigned long source_load(int cpu, int type);
1435 static unsigned long target_load(int cpu, int type);
1436 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1438 static unsigned long cpu_avg_load_per_task(int cpu)
1440 struct rq *rq = cpu_rq(cpu);
1443 rq->avg_load_per_task = rq->load.weight / rq->nr_running;
1445 return rq->avg_load_per_task;
1448 #ifdef CONFIG_FAIR_GROUP_SCHED
1450 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1453 * Calculate and set the cpu's group shares.
1456 __update_group_shares_cpu(struct task_group *tg, int cpu,
1457 unsigned long sd_shares, unsigned long sd_rq_weight)
1460 unsigned long shares;
1461 unsigned long rq_weight;
1466 rq_weight = tg->cfs_rq[cpu]->load.weight;
1469 * If there are currently no tasks on the cpu pretend there is one of
1470 * average load so that when a new task gets to run here it will not
1471 * get delayed by group starvation.
1475 rq_weight = NICE_0_LOAD;
1478 if (unlikely(rq_weight > sd_rq_weight))
1479 rq_weight = sd_rq_weight;
1482 * \Sum shares * rq_weight
1483 * shares = -----------------------
1487 shares = (sd_shares * rq_weight) / (sd_rq_weight + 1);
1490 * record the actual number of shares, not the boosted amount.
1492 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1493 tg->cfs_rq[cpu]->rq_weight = rq_weight;
1495 if (shares < MIN_SHARES)
1496 shares = MIN_SHARES;
1497 else if (shares > MAX_SHARES)
1498 shares = MAX_SHARES;
1500 __set_se_shares(tg->se[cpu], shares);
1504 * Re-compute the task group their per cpu shares over the given domain.
1505 * This needs to be done in a bottom-up fashion because the rq weight of a
1506 * parent group depends on the shares of its child groups.
1508 static int tg_shares_up(struct task_group *tg, void *data)
1510 unsigned long rq_weight = 0;
1511 unsigned long shares = 0;
1512 struct sched_domain *sd = data;
1515 for_each_cpu_mask(i, sd->span) {
1516 rq_weight += tg->cfs_rq[i]->load.weight;
1517 shares += tg->cfs_rq[i]->shares;
1520 if ((!shares && rq_weight) || shares > tg->shares)
1521 shares = tg->shares;
1523 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1524 shares = tg->shares;
1527 rq_weight = cpus_weight(sd->span) * NICE_0_LOAD;
1529 for_each_cpu_mask(i, sd->span) {
1530 struct rq *rq = cpu_rq(i);
1531 unsigned long flags;
1533 spin_lock_irqsave(&rq->lock, flags);
1534 __update_group_shares_cpu(tg, i, shares, rq_weight);
1535 spin_unlock_irqrestore(&rq->lock, flags);
1542 * Compute the cpu's hierarchical load factor for each task group.
1543 * This needs to be done in a top-down fashion because the load of a child
1544 * group is a fraction of its parents load.
1546 static int tg_load_down(struct task_group *tg, void *data)
1549 long cpu = (long)data;
1552 load = cpu_rq(cpu)->load.weight;
1554 load = tg->parent->cfs_rq[cpu]->h_load;
1555 load *= tg->cfs_rq[cpu]->shares;
1556 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1559 tg->cfs_rq[cpu]->h_load = load;
1564 static void update_shares(struct sched_domain *sd)
1566 u64 now = cpu_clock(raw_smp_processor_id());
1567 s64 elapsed = now - sd->last_update;
1569 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1570 sd->last_update = now;
1571 walk_tg_tree(tg_nop, tg_shares_up, sd);
1575 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1577 spin_unlock(&rq->lock);
1579 spin_lock(&rq->lock);
1582 static void update_h_load(long cpu)
1584 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1589 static inline void update_shares(struct sched_domain *sd)
1593 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1601 #ifdef CONFIG_FAIR_GROUP_SCHED
1602 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1605 cfs_rq->shares = shares;
1610 #include "sched_stats.h"
1611 #include "sched_idletask.c"
1612 #include "sched_fair.c"
1613 #include "sched_rt.c"
1614 #ifdef CONFIG_SCHED_DEBUG
1615 # include "sched_debug.c"
1618 #define sched_class_highest (&rt_sched_class)
1619 #define for_each_class(class) \
1620 for (class = sched_class_highest; class; class = class->next)
1622 static void inc_nr_running(struct rq *rq)
1627 static void dec_nr_running(struct rq *rq)
1632 static void set_load_weight(struct task_struct *p)
1634 if (task_has_rt_policy(p)) {
1635 p->se.load.weight = prio_to_weight[0] * 2;
1636 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1641 * SCHED_IDLE tasks get minimal weight:
1643 if (p->policy == SCHED_IDLE) {
1644 p->se.load.weight = WEIGHT_IDLEPRIO;
1645 p->se.load.inv_weight = WMULT_IDLEPRIO;
1649 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1650 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1653 static void update_avg(u64 *avg, u64 sample)
1655 s64 diff = sample - *avg;
1659 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1661 sched_info_queued(p);
1662 p->sched_class->enqueue_task(rq, p, wakeup);
1666 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1668 if (sleep && p->se.last_wakeup) {
1669 update_avg(&p->se.avg_overlap,
1670 p->se.sum_exec_runtime - p->se.last_wakeup);
1671 p->se.last_wakeup = 0;
1674 sched_info_dequeued(p);
1675 p->sched_class->dequeue_task(rq, p, sleep);
1680 * __normal_prio - return the priority that is based on the static prio
1682 static inline int __normal_prio(struct task_struct *p)
1684 return p->static_prio;
1688 * Calculate the expected normal priority: i.e. priority
1689 * without taking RT-inheritance into account. Might be
1690 * boosted by interactivity modifiers. Changes upon fork,
1691 * setprio syscalls, and whenever the interactivity
1692 * estimator recalculates.
1694 static inline int normal_prio(struct task_struct *p)
1698 if (task_has_rt_policy(p))
1699 prio = MAX_RT_PRIO-1 - p->rt_priority;
1701 prio = __normal_prio(p);
1706 * Calculate the current priority, i.e. the priority
1707 * taken into account by the scheduler. This value might
1708 * be boosted by RT tasks, or might be boosted by
1709 * interactivity modifiers. Will be RT if the task got
1710 * RT-boosted. If not then it returns p->normal_prio.
1712 static int effective_prio(struct task_struct *p)
1714 p->normal_prio = normal_prio(p);
1716 * If we are RT tasks or we were boosted to RT priority,
1717 * keep the priority unchanged. Otherwise, update priority
1718 * to the normal priority:
1720 if (!rt_prio(p->prio))
1721 return p->normal_prio;
1726 * activate_task - move a task to the runqueue.
1728 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1730 if (task_contributes_to_load(p))
1731 rq->nr_uninterruptible--;
1733 enqueue_task(rq, p, wakeup);
1738 * deactivate_task - remove a task from the runqueue.
1740 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1742 if (task_contributes_to_load(p))
1743 rq->nr_uninterruptible++;
1745 dequeue_task(rq, p, sleep);
1750 * task_curr - is this task currently executing on a CPU?
1751 * @p: the task in question.
1753 inline int task_curr(const struct task_struct *p)
1755 return cpu_curr(task_cpu(p)) == p;
1758 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1760 set_task_rq(p, cpu);
1763 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1764 * successfuly executed on another CPU. We must ensure that updates of
1765 * per-task data have been completed by this moment.
1768 task_thread_info(p)->cpu = cpu;
1772 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1773 const struct sched_class *prev_class,
1774 int oldprio, int running)
1776 if (prev_class != p->sched_class) {
1777 if (prev_class->switched_from)
1778 prev_class->switched_from(rq, p, running);
1779 p->sched_class->switched_to(rq, p, running);
1781 p->sched_class->prio_changed(rq, p, oldprio, running);
1786 /* Used instead of source_load when we know the type == 0 */
1787 static unsigned long weighted_cpuload(const int cpu)
1789 return cpu_rq(cpu)->load.weight;
1793 * Is this task likely cache-hot:
1796 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1801 * Buddy candidates are cache hot:
1803 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
1806 if (p->sched_class != &fair_sched_class)
1809 if (sysctl_sched_migration_cost == -1)
1811 if (sysctl_sched_migration_cost == 0)
1814 delta = now - p->se.exec_start;
1816 return delta < (s64)sysctl_sched_migration_cost;
1820 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1822 int old_cpu = task_cpu(p);
1823 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1824 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1825 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1828 clock_offset = old_rq->clock - new_rq->clock;
1830 #ifdef CONFIG_SCHEDSTATS
1831 if (p->se.wait_start)
1832 p->se.wait_start -= clock_offset;
1833 if (p->se.sleep_start)
1834 p->se.sleep_start -= clock_offset;
1835 if (p->se.block_start)
1836 p->se.block_start -= clock_offset;
1837 if (old_cpu != new_cpu) {
1838 schedstat_inc(p, se.nr_migrations);
1839 if (task_hot(p, old_rq->clock, NULL))
1840 schedstat_inc(p, se.nr_forced2_migrations);
1843 p->se.vruntime -= old_cfsrq->min_vruntime -
1844 new_cfsrq->min_vruntime;
1846 __set_task_cpu(p, new_cpu);
1849 struct migration_req {
1850 struct list_head list;
1852 struct task_struct *task;
1855 struct completion done;
1859 * The task's runqueue lock must be held.
1860 * Returns true if you have to wait for migration thread.
1863 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1865 struct rq *rq = task_rq(p);
1868 * If the task is not on a runqueue (and not running), then
1869 * it is sufficient to simply update the task's cpu field.
1871 if (!p->se.on_rq && !task_running(rq, p)) {
1872 set_task_cpu(p, dest_cpu);
1876 init_completion(&req->done);
1878 req->dest_cpu = dest_cpu;
1879 list_add(&req->list, &rq->migration_queue);
1885 * wait_task_inactive - wait for a thread to unschedule.
1887 * If @match_state is nonzero, it's the @p->state value just checked and
1888 * not expected to change. If it changes, i.e. @p might have woken up,
1889 * then return zero. When we succeed in waiting for @p to be off its CPU,
1890 * we return a positive number (its total switch count). If a second call
1891 * a short while later returns the same number, the caller can be sure that
1892 * @p has remained unscheduled the whole time.
1894 * The caller must ensure that the task *will* unschedule sometime soon,
1895 * else this function might spin for a *long* time. This function can't
1896 * be called with interrupts off, or it may introduce deadlock with
1897 * smp_call_function() if an IPI is sent by the same process we are
1898 * waiting to become inactive.
1900 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1902 unsigned long flags;
1909 * We do the initial early heuristics without holding
1910 * any task-queue locks at all. We'll only try to get
1911 * the runqueue lock when things look like they will
1917 * If the task is actively running on another CPU
1918 * still, just relax and busy-wait without holding
1921 * NOTE! Since we don't hold any locks, it's not
1922 * even sure that "rq" stays as the right runqueue!
1923 * But we don't care, since "task_running()" will
1924 * return false if the runqueue has changed and p
1925 * is actually now running somewhere else!
1927 while (task_running(rq, p)) {
1928 if (match_state && unlikely(p->state != match_state))
1934 * Ok, time to look more closely! We need the rq
1935 * lock now, to be *sure*. If we're wrong, we'll
1936 * just go back and repeat.
1938 rq = task_rq_lock(p, &flags);
1939 running = task_running(rq, p);
1940 on_rq = p->se.on_rq;
1942 if (!match_state || p->state == match_state) {
1943 ncsw = p->nivcsw + p->nvcsw;
1944 if (unlikely(!ncsw))
1947 task_rq_unlock(rq, &flags);
1950 * If it changed from the expected state, bail out now.
1952 if (unlikely(!ncsw))
1956 * Was it really running after all now that we
1957 * checked with the proper locks actually held?
1959 * Oops. Go back and try again..
1961 if (unlikely(running)) {
1967 * It's not enough that it's not actively running,
1968 * it must be off the runqueue _entirely_, and not
1971 * So if it wa still runnable (but just not actively
1972 * running right now), it's preempted, and we should
1973 * yield - it could be a while.
1975 if (unlikely(on_rq)) {
1976 schedule_timeout_uninterruptible(1);
1981 * Ahh, all good. It wasn't running, and it wasn't
1982 * runnable, which means that it will never become
1983 * running in the future either. We're all done!
1992 * kick_process - kick a running thread to enter/exit the kernel
1993 * @p: the to-be-kicked thread
1995 * Cause a process which is running on another CPU to enter
1996 * kernel-mode, without any delay. (to get signals handled.)
1998 * NOTE: this function doesnt have to take the runqueue lock,
1999 * because all it wants to ensure is that the remote task enters
2000 * the kernel. If the IPI races and the task has been migrated
2001 * to another CPU then no harm is done and the purpose has been
2004 void kick_process(struct task_struct *p)
2010 if ((cpu != smp_processor_id()) && task_curr(p))
2011 smp_send_reschedule(cpu);
2016 * Return a low guess at the load of a migration-source cpu weighted
2017 * according to the scheduling class and "nice" value.
2019 * We want to under-estimate the load of migration sources, to
2020 * balance conservatively.
2022 static unsigned long source_load(int cpu, int type)
2024 struct rq *rq = cpu_rq(cpu);
2025 unsigned long total = weighted_cpuload(cpu);
2027 if (type == 0 || !sched_feat(LB_BIAS))
2030 return min(rq->cpu_load[type-1], total);
2034 * Return a high guess at the load of a migration-target cpu weighted
2035 * according to the scheduling class and "nice" value.
2037 static unsigned long target_load(int cpu, int type)
2039 struct rq *rq = cpu_rq(cpu);
2040 unsigned long total = weighted_cpuload(cpu);
2042 if (type == 0 || !sched_feat(LB_BIAS))
2045 return max(rq->cpu_load[type-1], total);
2049 * find_idlest_group finds and returns the least busy CPU group within the
2052 static struct sched_group *
2053 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2055 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2056 unsigned long min_load = ULONG_MAX, this_load = 0;
2057 int load_idx = sd->forkexec_idx;
2058 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2061 unsigned long load, avg_load;
2065 /* Skip over this group if it has no CPUs allowed */
2066 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
2069 local_group = cpu_isset(this_cpu, group->cpumask);
2071 /* Tally up the load of all CPUs in the group */
2074 for_each_cpu_mask_nr(i, group->cpumask) {
2075 /* Bias balancing toward cpus of our domain */
2077 load = source_load(i, load_idx);
2079 load = target_load(i, load_idx);
2084 /* Adjust by relative CPU power of the group */
2085 avg_load = sg_div_cpu_power(group,
2086 avg_load * SCHED_LOAD_SCALE);
2089 this_load = avg_load;
2091 } else if (avg_load < min_load) {
2092 min_load = avg_load;
2095 } while (group = group->next, group != sd->groups);
2097 if (!idlest || 100*this_load < imbalance*min_load)
2103 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2106 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2109 unsigned long load, min_load = ULONG_MAX;
2113 /* Traverse only the allowed CPUs */
2114 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
2116 for_each_cpu_mask_nr(i, *tmp) {
2117 load = weighted_cpuload(i);
2119 if (load < min_load || (load == min_load && i == this_cpu)) {
2129 * sched_balance_self: balance the current task (running on cpu) in domains
2130 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2133 * Balance, ie. select the least loaded group.
2135 * Returns the target CPU number, or the same CPU if no balancing is needed.
2137 * preempt must be disabled.
2139 static int sched_balance_self(int cpu, int flag)
2141 struct task_struct *t = current;
2142 struct sched_domain *tmp, *sd = NULL;
2144 for_each_domain(cpu, tmp) {
2146 * If power savings logic is enabled for a domain, stop there.
2148 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2150 if (tmp->flags & flag)
2158 cpumask_t span, tmpmask;
2159 struct sched_group *group;
2160 int new_cpu, weight;
2162 if (!(sd->flags & flag)) {
2168 group = find_idlest_group(sd, t, cpu);
2174 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2175 if (new_cpu == -1 || new_cpu == cpu) {
2176 /* Now try balancing at a lower domain level of cpu */
2181 /* Now try balancing at a lower domain level of new_cpu */
2184 weight = cpus_weight(span);
2185 for_each_domain(cpu, tmp) {
2186 if (weight <= cpus_weight(tmp->span))
2188 if (tmp->flags & flag)
2191 /* while loop will break here if sd == NULL */
2197 #endif /* CONFIG_SMP */
2200 * try_to_wake_up - wake up a thread
2201 * @p: the to-be-woken-up thread
2202 * @state: the mask of task states that can be woken
2203 * @sync: do a synchronous wakeup?
2205 * Put it on the run-queue if it's not already there. The "current"
2206 * thread is always on the run-queue (except when the actual
2207 * re-schedule is in progress), and as such you're allowed to do
2208 * the simpler "current->state = TASK_RUNNING" to mark yourself
2209 * runnable without the overhead of this.
2211 * returns failure only if the task is already active.
2213 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2215 int cpu, orig_cpu, this_cpu, success = 0;
2216 unsigned long flags;
2220 if (!sched_feat(SYNC_WAKEUPS))
2224 if (sched_feat(LB_WAKEUP_UPDATE)) {
2225 struct sched_domain *sd;
2227 this_cpu = raw_smp_processor_id();
2230 for_each_domain(this_cpu, sd) {
2231 if (cpu_isset(cpu, sd->span)) {
2240 rq = task_rq_lock(p, &flags);
2241 old_state = p->state;
2242 if (!(old_state & state))
2250 this_cpu = smp_processor_id();
2253 if (unlikely(task_running(rq, p)))
2256 cpu = p->sched_class->select_task_rq(p, sync);
2257 if (cpu != orig_cpu) {
2258 set_task_cpu(p, cpu);
2259 task_rq_unlock(rq, &flags);
2260 /* might preempt at this point */
2261 rq = task_rq_lock(p, &flags);
2262 old_state = p->state;
2263 if (!(old_state & state))
2268 this_cpu = smp_processor_id();
2272 #ifdef CONFIG_SCHEDSTATS
2273 schedstat_inc(rq, ttwu_count);
2274 if (cpu == this_cpu)
2275 schedstat_inc(rq, ttwu_local);
2277 struct sched_domain *sd;
2278 for_each_domain(this_cpu, sd) {
2279 if (cpu_isset(cpu, sd->span)) {
2280 schedstat_inc(sd, ttwu_wake_remote);
2285 #endif /* CONFIG_SCHEDSTATS */
2288 #endif /* CONFIG_SMP */
2289 schedstat_inc(p, se.nr_wakeups);
2291 schedstat_inc(p, se.nr_wakeups_sync);
2292 if (orig_cpu != cpu)
2293 schedstat_inc(p, se.nr_wakeups_migrate);
2294 if (cpu == this_cpu)
2295 schedstat_inc(p, se.nr_wakeups_local);
2297 schedstat_inc(p, se.nr_wakeups_remote);
2298 update_rq_clock(rq);
2299 activate_task(rq, p, 1);
2303 trace_mark(kernel_sched_wakeup,
2304 "pid %d state %ld ## rq %p task %p rq->curr %p",
2305 p->pid, p->state, rq, p, rq->curr);
2306 check_preempt_curr(rq, p);
2308 p->state = TASK_RUNNING;
2310 if (p->sched_class->task_wake_up)
2311 p->sched_class->task_wake_up(rq, p);
2314 current->se.last_wakeup = current->se.sum_exec_runtime;
2316 task_rq_unlock(rq, &flags);
2321 int wake_up_process(struct task_struct *p)
2323 return try_to_wake_up(p, TASK_ALL, 0);
2325 EXPORT_SYMBOL(wake_up_process);
2327 int wake_up_state(struct task_struct *p, unsigned int state)
2329 return try_to_wake_up(p, state, 0);
2333 * Perform scheduler related setup for a newly forked process p.
2334 * p is forked by current.
2336 * __sched_fork() is basic setup used by init_idle() too:
2338 static void __sched_fork(struct task_struct *p)
2340 p->se.exec_start = 0;
2341 p->se.sum_exec_runtime = 0;
2342 p->se.prev_sum_exec_runtime = 0;
2343 p->se.last_wakeup = 0;
2344 p->se.avg_overlap = 0;
2346 #ifdef CONFIG_SCHEDSTATS
2347 p->se.wait_start = 0;
2348 p->se.sum_sleep_runtime = 0;
2349 p->se.sleep_start = 0;
2350 p->se.block_start = 0;
2351 p->se.sleep_max = 0;
2352 p->se.block_max = 0;
2354 p->se.slice_max = 0;
2358 INIT_LIST_HEAD(&p->rt.run_list);
2360 INIT_LIST_HEAD(&p->se.group_node);
2362 #ifdef CONFIG_PREEMPT_NOTIFIERS
2363 INIT_HLIST_HEAD(&p->preempt_notifiers);
2367 * We mark the process as running here, but have not actually
2368 * inserted it onto the runqueue yet. This guarantees that
2369 * nobody will actually run it, and a signal or other external
2370 * event cannot wake it up and insert it on the runqueue either.
2372 p->state = TASK_RUNNING;
2376 * fork()/clone()-time setup:
2378 void sched_fork(struct task_struct *p, int clone_flags)
2380 int cpu = get_cpu();
2385 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2387 set_task_cpu(p, cpu);
2390 * Make sure we do not leak PI boosting priority to the child:
2392 p->prio = current->normal_prio;
2393 if (!rt_prio(p->prio))
2394 p->sched_class = &fair_sched_class;
2396 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2397 if (likely(sched_info_on()))
2398 memset(&p->sched_info, 0, sizeof(p->sched_info));
2400 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2403 #ifdef CONFIG_PREEMPT
2404 /* Want to start with kernel preemption disabled. */
2405 task_thread_info(p)->preempt_count = 1;
2411 * wake_up_new_task - wake up a newly created task for the first time.
2413 * This function will do some initial scheduler statistics housekeeping
2414 * that must be done for every newly created context, then puts the task
2415 * on the runqueue and wakes it.
2417 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2419 unsigned long flags;
2422 rq = task_rq_lock(p, &flags);
2423 BUG_ON(p->state != TASK_RUNNING);
2424 update_rq_clock(rq);
2426 p->prio = effective_prio(p);
2428 if (!p->sched_class->task_new || !current->se.on_rq) {
2429 activate_task(rq, p, 0);
2432 * Let the scheduling class do new task startup
2433 * management (if any):
2435 p->sched_class->task_new(rq, p);
2438 trace_mark(kernel_sched_wakeup_new,
2439 "pid %d state %ld ## rq %p task %p rq->curr %p",
2440 p->pid, p->state, rq, p, rq->curr);
2441 check_preempt_curr(rq, p);
2443 if (p->sched_class->task_wake_up)
2444 p->sched_class->task_wake_up(rq, p);
2446 task_rq_unlock(rq, &flags);
2449 #ifdef CONFIG_PREEMPT_NOTIFIERS
2452 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2453 * @notifier: notifier struct to register
2455 void preempt_notifier_register(struct preempt_notifier *notifier)
2457 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2459 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2462 * preempt_notifier_unregister - no longer interested in preemption notifications
2463 * @notifier: notifier struct to unregister
2465 * This is safe to call from within a preemption notifier.
2467 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2469 hlist_del(¬ifier->link);
2471 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2473 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
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_in(notifier, raw_smp_processor_id());
2483 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2484 struct task_struct *next)
2486 struct preempt_notifier *notifier;
2487 struct hlist_node *node;
2489 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2490 notifier->ops->sched_out(notifier, next);
2493 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2495 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2500 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2501 struct task_struct *next)
2505 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2508 * prepare_task_switch - prepare to switch tasks
2509 * @rq: the runqueue preparing to switch
2510 * @prev: the current task that is being switched out
2511 * @next: the task we are going to switch to.
2513 * This is called with the rq lock held and interrupts off. It must
2514 * be paired with a subsequent finish_task_switch after the context
2517 * prepare_task_switch sets up locking and calls architecture specific
2521 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2522 struct task_struct *next)
2524 fire_sched_out_preempt_notifiers(prev, next);
2525 prepare_lock_switch(rq, next);
2526 prepare_arch_switch(next);
2530 * finish_task_switch - clean up after a task-switch
2531 * @rq: runqueue associated with task-switch
2532 * @prev: the thread we just switched away from.
2534 * finish_task_switch must be called after the context switch, paired
2535 * with a prepare_task_switch call before the context switch.
2536 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2537 * and do any other architecture-specific cleanup actions.
2539 * Note that we may have delayed dropping an mm in context_switch(). If
2540 * so, we finish that here outside of the runqueue lock. (Doing it
2541 * with the lock held can cause deadlocks; see schedule() for
2544 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2545 __releases(rq->lock)
2547 struct mm_struct *mm = rq->prev_mm;
2553 * A task struct has one reference for the use as "current".
2554 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2555 * schedule one last time. The schedule call will never return, and
2556 * the scheduled task must drop that reference.
2557 * The test for TASK_DEAD must occur while the runqueue locks are
2558 * still held, otherwise prev could be scheduled on another cpu, die
2559 * there before we look at prev->state, and then the reference would
2561 * Manfred Spraul <manfred@colorfullife.com>
2563 prev_state = prev->state;
2564 finish_arch_switch(prev);
2565 finish_lock_switch(rq, prev);
2567 if (current->sched_class->post_schedule)
2568 current->sched_class->post_schedule(rq);
2571 fire_sched_in_preempt_notifiers(current);
2574 if (unlikely(prev_state == TASK_DEAD)) {
2576 * Remove function-return probe instances associated with this
2577 * task and put them back on the free list.
2579 kprobe_flush_task(prev);
2580 put_task_struct(prev);
2585 * schedule_tail - first thing a freshly forked thread must call.
2586 * @prev: the thread we just switched away from.
2588 asmlinkage void schedule_tail(struct task_struct *prev)
2589 __releases(rq->lock)
2591 struct rq *rq = this_rq();
2593 finish_task_switch(rq, prev);
2594 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2595 /* In this case, finish_task_switch does not reenable preemption */
2598 if (current->set_child_tid)
2599 put_user(task_pid_vnr(current), current->set_child_tid);
2603 * context_switch - switch to the new MM and the new
2604 * thread's register state.
2607 context_switch(struct rq *rq, struct task_struct *prev,
2608 struct task_struct *next)
2610 struct mm_struct *mm, *oldmm;
2612 prepare_task_switch(rq, prev, next);
2613 trace_mark(kernel_sched_schedule,
2614 "prev_pid %d next_pid %d prev_state %ld "
2615 "## rq %p prev %p next %p",
2616 prev->pid, next->pid, prev->state,
2619 oldmm = prev->active_mm;
2621 * For paravirt, this is coupled with an exit in switch_to to
2622 * combine the page table reload and the switch backend into
2625 arch_enter_lazy_cpu_mode();
2627 if (unlikely(!mm)) {
2628 next->active_mm = oldmm;
2629 atomic_inc(&oldmm->mm_count);
2630 enter_lazy_tlb(oldmm, next);
2632 switch_mm(oldmm, mm, next);
2634 if (unlikely(!prev->mm)) {
2635 prev->active_mm = NULL;
2636 rq->prev_mm = oldmm;
2639 * Since the runqueue lock will be released by the next
2640 * task (which is an invalid locking op but in the case
2641 * of the scheduler it's an obvious special-case), so we
2642 * do an early lockdep release here:
2644 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2645 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2648 /* Here we just switch the register state and the stack. */
2649 switch_to(prev, next, prev);
2653 * this_rq must be evaluated again because prev may have moved
2654 * CPUs since it called schedule(), thus the 'rq' on its stack
2655 * frame will be invalid.
2657 finish_task_switch(this_rq(), prev);
2661 * nr_running, nr_uninterruptible and nr_context_switches:
2663 * externally visible scheduler statistics: current number of runnable
2664 * threads, current number of uninterruptible-sleeping threads, total
2665 * number of context switches performed since bootup.
2667 unsigned long nr_running(void)
2669 unsigned long i, sum = 0;
2671 for_each_online_cpu(i)
2672 sum += cpu_rq(i)->nr_running;
2677 unsigned long nr_uninterruptible(void)
2679 unsigned long i, sum = 0;
2681 for_each_possible_cpu(i)
2682 sum += cpu_rq(i)->nr_uninterruptible;
2685 * Since we read the counters lockless, it might be slightly
2686 * inaccurate. Do not allow it to go below zero though:
2688 if (unlikely((long)sum < 0))
2694 unsigned long long nr_context_switches(void)
2697 unsigned long long sum = 0;
2699 for_each_possible_cpu(i)
2700 sum += cpu_rq(i)->nr_switches;
2705 unsigned long nr_iowait(void)
2707 unsigned long i, sum = 0;
2709 for_each_possible_cpu(i)
2710 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2715 unsigned long nr_active(void)
2717 unsigned long i, running = 0, uninterruptible = 0;
2719 for_each_online_cpu(i) {
2720 running += cpu_rq(i)->nr_running;
2721 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2724 if (unlikely((long)uninterruptible < 0))
2725 uninterruptible = 0;
2727 return running + uninterruptible;
2731 * Update rq->cpu_load[] statistics. This function is usually called every
2732 * scheduler tick (TICK_NSEC).
2734 static void update_cpu_load(struct rq *this_rq)
2736 unsigned long this_load = this_rq->load.weight;
2739 this_rq->nr_load_updates++;
2741 /* Update our load: */
2742 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2743 unsigned long old_load, new_load;
2745 /* scale is effectively 1 << i now, and >> i divides by scale */
2747 old_load = this_rq->cpu_load[i];
2748 new_load = this_load;
2750 * Round up the averaging division if load is increasing. This
2751 * prevents us from getting stuck on 9 if the load is 10, for
2754 if (new_load > old_load)
2755 new_load += scale-1;
2756 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2763 * double_rq_lock - safely lock two runqueues
2765 * Note this does not disable interrupts like task_rq_lock,
2766 * you need to do so manually before calling.
2768 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2769 __acquires(rq1->lock)
2770 __acquires(rq2->lock)
2772 BUG_ON(!irqs_disabled());
2774 spin_lock(&rq1->lock);
2775 __acquire(rq2->lock); /* Fake it out ;) */
2778 spin_lock(&rq1->lock);
2779 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2781 spin_lock(&rq2->lock);
2782 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2785 update_rq_clock(rq1);
2786 update_rq_clock(rq2);
2790 * double_rq_unlock - safely unlock two runqueues
2792 * Note this does not restore interrupts like task_rq_unlock,
2793 * you need to do so manually after calling.
2795 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2796 __releases(rq1->lock)
2797 __releases(rq2->lock)
2799 spin_unlock(&rq1->lock);
2801 spin_unlock(&rq2->lock);
2803 __release(rq2->lock);
2807 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2809 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2810 __releases(this_rq->lock)
2811 __acquires(busiest->lock)
2812 __acquires(this_rq->lock)
2816 if (unlikely(!irqs_disabled())) {
2817 /* printk() doesn't work good under rq->lock */
2818 spin_unlock(&this_rq->lock);
2821 if (unlikely(!spin_trylock(&busiest->lock))) {
2822 if (busiest < this_rq) {
2823 spin_unlock(&this_rq->lock);
2824 spin_lock(&busiest->lock);
2825 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
2828 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
2833 static void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
2834 __releases(busiest->lock)
2836 spin_unlock(&busiest->lock);
2837 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
2841 * If dest_cpu is allowed for this process, migrate the task to it.
2842 * This is accomplished by forcing the cpu_allowed mask to only
2843 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2844 * the cpu_allowed mask is restored.
2846 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2848 struct migration_req req;
2849 unsigned long flags;
2852 rq = task_rq_lock(p, &flags);
2853 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2854 || unlikely(!cpu_active(dest_cpu)))
2857 /* force the process onto the specified CPU */
2858 if (migrate_task(p, dest_cpu, &req)) {
2859 /* Need to wait for migration thread (might exit: take ref). */
2860 struct task_struct *mt = rq->migration_thread;
2862 get_task_struct(mt);
2863 task_rq_unlock(rq, &flags);
2864 wake_up_process(mt);
2865 put_task_struct(mt);
2866 wait_for_completion(&req.done);
2871 task_rq_unlock(rq, &flags);
2875 * sched_exec - execve() is a valuable balancing opportunity, because at
2876 * this point the task has the smallest effective memory and cache footprint.
2878 void sched_exec(void)
2880 int new_cpu, this_cpu = get_cpu();
2881 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2883 if (new_cpu != this_cpu)
2884 sched_migrate_task(current, new_cpu);
2888 * pull_task - move a task from a remote runqueue to the local runqueue.
2889 * Both runqueues must be locked.
2891 static void pull_task(struct rq *src_rq, struct task_struct *p,
2892 struct rq *this_rq, int this_cpu)
2894 deactivate_task(src_rq, p, 0);
2895 set_task_cpu(p, this_cpu);
2896 activate_task(this_rq, p, 0);
2898 * Note that idle threads have a prio of MAX_PRIO, for this test
2899 * to be always true for them.
2901 check_preempt_curr(this_rq, p);
2905 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2908 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2909 struct sched_domain *sd, enum cpu_idle_type idle,
2913 * We do not migrate tasks that are:
2914 * 1) running (obviously), or
2915 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2916 * 3) are cache-hot on their current CPU.
2918 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2919 schedstat_inc(p, se.nr_failed_migrations_affine);
2924 if (task_running(rq, p)) {
2925 schedstat_inc(p, se.nr_failed_migrations_running);
2930 * Aggressive migration if:
2931 * 1) task is cache cold, or
2932 * 2) too many balance attempts have failed.
2935 if (!task_hot(p, rq->clock, sd) ||
2936 sd->nr_balance_failed > sd->cache_nice_tries) {
2937 #ifdef CONFIG_SCHEDSTATS
2938 if (task_hot(p, rq->clock, sd)) {
2939 schedstat_inc(sd, lb_hot_gained[idle]);
2940 schedstat_inc(p, se.nr_forced_migrations);
2946 if (task_hot(p, rq->clock, sd)) {
2947 schedstat_inc(p, se.nr_failed_migrations_hot);
2953 static unsigned long
2954 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2955 unsigned long max_load_move, struct sched_domain *sd,
2956 enum cpu_idle_type idle, int *all_pinned,
2957 int *this_best_prio, struct rq_iterator *iterator)
2959 int loops = 0, pulled = 0, pinned = 0;
2960 struct task_struct *p;
2961 long rem_load_move = max_load_move;
2963 if (max_load_move == 0)
2969 * Start the load-balancing iterator:
2971 p = iterator->start(iterator->arg);
2973 if (!p || loops++ > sysctl_sched_nr_migrate)
2976 if ((p->se.load.weight >> 1) > rem_load_move ||
2977 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2978 p = iterator->next(iterator->arg);
2982 pull_task(busiest, p, this_rq, this_cpu);
2984 rem_load_move -= p->se.load.weight;
2987 * We only want to steal up to the prescribed amount of weighted load.
2989 if (rem_load_move > 0) {
2990 if (p->prio < *this_best_prio)
2991 *this_best_prio = p->prio;
2992 p = iterator->next(iterator->arg);
2997 * Right now, this is one of only two places pull_task() is called,
2998 * so we can safely collect pull_task() stats here rather than
2999 * inside pull_task().
3001 schedstat_add(sd, lb_gained[idle], pulled);
3004 *all_pinned = pinned;
3006 return max_load_move - rem_load_move;
3010 * move_tasks tries to move up to max_load_move weighted load from busiest to
3011 * this_rq, as part of a balancing operation within domain "sd".
3012 * Returns 1 if successful and 0 otherwise.
3014 * Called with both runqueues locked.
3016 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3017 unsigned long max_load_move,
3018 struct sched_domain *sd, enum cpu_idle_type idle,
3021 const struct sched_class *class = sched_class_highest;
3022 unsigned long total_load_moved = 0;
3023 int this_best_prio = this_rq->curr->prio;
3027 class->load_balance(this_rq, this_cpu, busiest,
3028 max_load_move - total_load_moved,
3029 sd, idle, all_pinned, &this_best_prio);
3030 class = class->next;
3032 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3035 } while (class && max_load_move > total_load_moved);
3037 return total_load_moved > 0;
3041 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3042 struct sched_domain *sd, enum cpu_idle_type idle,
3043 struct rq_iterator *iterator)
3045 struct task_struct *p = iterator->start(iterator->arg);
3049 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3050 pull_task(busiest, p, this_rq, this_cpu);
3052 * Right now, this is only the second place pull_task()
3053 * is called, so we can safely collect pull_task()
3054 * stats here rather than inside pull_task().
3056 schedstat_inc(sd, lb_gained[idle]);
3060 p = iterator->next(iterator->arg);
3067 * move_one_task tries to move exactly one task from busiest to this_rq, as
3068 * part of active balancing operations within "domain".
3069 * Returns 1 if successful and 0 otherwise.
3071 * Called with both runqueues locked.
3073 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3074 struct sched_domain *sd, enum cpu_idle_type idle)
3076 const struct sched_class *class;
3078 for (class = sched_class_highest; class; class = class->next)
3079 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3086 * find_busiest_group finds and returns the busiest CPU group within the
3087 * domain. It calculates and returns the amount of weighted load which
3088 * should be moved to restore balance via the imbalance parameter.
3090 static struct sched_group *
3091 find_busiest_group(struct sched_domain *sd, int this_cpu,
3092 unsigned long *imbalance, enum cpu_idle_type idle,
3093 int *sd_idle, const cpumask_t *cpus, int *balance)
3095 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3096 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3097 unsigned long max_pull;
3098 unsigned long busiest_load_per_task, busiest_nr_running;
3099 unsigned long this_load_per_task, this_nr_running;
3100 int load_idx, group_imb = 0;
3101 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3102 int power_savings_balance = 1;
3103 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3104 unsigned long min_nr_running = ULONG_MAX;
3105 struct sched_group *group_min = NULL, *group_leader = NULL;
3108 max_load = this_load = total_load = total_pwr = 0;
3109 busiest_load_per_task = busiest_nr_running = 0;
3110 this_load_per_task = this_nr_running = 0;
3112 if (idle == CPU_NOT_IDLE)
3113 load_idx = sd->busy_idx;
3114 else if (idle == CPU_NEWLY_IDLE)
3115 load_idx = sd->newidle_idx;
3117 load_idx = sd->idle_idx;
3120 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3123 int __group_imb = 0;
3124 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3125 unsigned long sum_nr_running, sum_weighted_load;
3126 unsigned long sum_avg_load_per_task;
3127 unsigned long avg_load_per_task;
3129 local_group = cpu_isset(this_cpu, group->cpumask);
3132 balance_cpu = first_cpu(group->cpumask);
3134 /* Tally up the load of all CPUs in the group */
3135 sum_weighted_load = sum_nr_running = avg_load = 0;
3136 sum_avg_load_per_task = avg_load_per_task = 0;
3139 min_cpu_load = ~0UL;
3141 for_each_cpu_mask_nr(i, group->cpumask) {
3144 if (!cpu_isset(i, *cpus))
3149 if (*sd_idle && rq->nr_running)
3152 /* Bias balancing toward cpus of our domain */
3154 if (idle_cpu(i) && !first_idle_cpu) {
3159 load = target_load(i, load_idx);
3161 load = source_load(i, load_idx);
3162 if (load > max_cpu_load)
3163 max_cpu_load = load;
3164 if (min_cpu_load > load)
3165 min_cpu_load = load;
3169 sum_nr_running += rq->nr_running;
3170 sum_weighted_load += weighted_cpuload(i);
3172 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3176 * First idle cpu or the first cpu(busiest) in this sched group
3177 * is eligible for doing load balancing at this and above
3178 * domains. In the newly idle case, we will allow all the cpu's
3179 * to do the newly idle load balance.
3181 if (idle != CPU_NEWLY_IDLE && local_group &&
3182 balance_cpu != this_cpu && balance) {
3187 total_load += avg_load;
3188 total_pwr += group->__cpu_power;
3190 /* Adjust by relative CPU power of the group */
3191 avg_load = sg_div_cpu_power(group,
3192 avg_load * SCHED_LOAD_SCALE);
3196 * Consider the group unbalanced when the imbalance is larger
3197 * than the average weight of two tasks.
3199 * APZ: with cgroup the avg task weight can vary wildly and
3200 * might not be a suitable number - should we keep a
3201 * normalized nr_running number somewhere that negates
3204 avg_load_per_task = sg_div_cpu_power(group,
3205 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3207 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3210 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3213 this_load = avg_load;
3215 this_nr_running = sum_nr_running;
3216 this_load_per_task = sum_weighted_load;
3217 } else if (avg_load > max_load &&
3218 (sum_nr_running > group_capacity || __group_imb)) {
3219 max_load = avg_load;
3221 busiest_nr_running = sum_nr_running;
3222 busiest_load_per_task = sum_weighted_load;
3223 group_imb = __group_imb;
3226 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3228 * Busy processors will not participate in power savings
3231 if (idle == CPU_NOT_IDLE ||
3232 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3236 * If the local group is idle or completely loaded
3237 * no need to do power savings balance at this domain
3239 if (local_group && (this_nr_running >= group_capacity ||
3241 power_savings_balance = 0;
3244 * If a group is already running at full capacity or idle,
3245 * don't include that group in power savings calculations
3247 if (!power_savings_balance || sum_nr_running >= group_capacity
3252 * Calculate the group which has the least non-idle load.
3253 * This is the group from where we need to pick up the load
3256 if ((sum_nr_running < min_nr_running) ||
3257 (sum_nr_running == min_nr_running &&
3258 first_cpu(group->cpumask) <
3259 first_cpu(group_min->cpumask))) {
3261 min_nr_running = sum_nr_running;
3262 min_load_per_task = sum_weighted_load /
3267 * Calculate the group which is almost near its
3268 * capacity but still has some space to pick up some load
3269 * from other group and save more power
3271 if (sum_nr_running <= group_capacity - 1) {
3272 if (sum_nr_running > leader_nr_running ||
3273 (sum_nr_running == leader_nr_running &&
3274 first_cpu(group->cpumask) >
3275 first_cpu(group_leader->cpumask))) {
3276 group_leader = group;
3277 leader_nr_running = sum_nr_running;
3282 group = group->next;
3283 } while (group != sd->groups);
3285 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3288 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3290 if (this_load >= avg_load ||
3291 100*max_load <= sd->imbalance_pct*this_load)
3294 busiest_load_per_task /= busiest_nr_running;
3296 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3299 * We're trying to get all the cpus to the average_load, so we don't
3300 * want to push ourselves above the average load, nor do we wish to
3301 * reduce the max loaded cpu below the average load, as either of these
3302 * actions would just result in more rebalancing later, and ping-pong
3303 * tasks around. Thus we look for the minimum possible imbalance.
3304 * Negative imbalances (*we* are more loaded than anyone else) will
3305 * be counted as no imbalance for these purposes -- we can't fix that
3306 * by pulling tasks to us. Be careful of negative numbers as they'll
3307 * appear as very large values with unsigned longs.
3309 if (max_load <= busiest_load_per_task)
3313 * In the presence of smp nice balancing, certain scenarios can have
3314 * max load less than avg load(as we skip the groups at or below
3315 * its cpu_power, while calculating max_load..)
3317 if (max_load < avg_load) {
3319 goto small_imbalance;
3322 /* Don't want to pull so many tasks that a group would go idle */
3323 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3325 /* How much load to actually move to equalise the imbalance */
3326 *imbalance = min(max_pull * busiest->__cpu_power,
3327 (avg_load - this_load) * this->__cpu_power)
3331 * if *imbalance is less than the average load per runnable task
3332 * there is no gaurantee that any tasks will be moved so we'll have
3333 * a think about bumping its value to force at least one task to be
3336 if (*imbalance < busiest_load_per_task) {
3337 unsigned long tmp, pwr_now, pwr_move;
3341 pwr_move = pwr_now = 0;
3343 if (this_nr_running) {
3344 this_load_per_task /= this_nr_running;
3345 if (busiest_load_per_task > this_load_per_task)
3348 this_load_per_task = cpu_avg_load_per_task(this_cpu);
3350 if (max_load - this_load + 2*busiest_load_per_task >=
3351 busiest_load_per_task * imbn) {
3352 *imbalance = busiest_load_per_task;
3357 * OK, we don't have enough imbalance to justify moving tasks,
3358 * however we may be able to increase total CPU power used by
3362 pwr_now += busiest->__cpu_power *
3363 min(busiest_load_per_task, max_load);
3364 pwr_now += this->__cpu_power *
3365 min(this_load_per_task, this_load);
3366 pwr_now /= SCHED_LOAD_SCALE;
3368 /* Amount of load we'd subtract */
3369 tmp = sg_div_cpu_power(busiest,
3370 busiest_load_per_task * SCHED_LOAD_SCALE);
3372 pwr_move += busiest->__cpu_power *
3373 min(busiest_load_per_task, max_load - tmp);
3375 /* Amount of load we'd add */
3376 if (max_load * busiest->__cpu_power <
3377 busiest_load_per_task * SCHED_LOAD_SCALE)
3378 tmp = sg_div_cpu_power(this,
3379 max_load * busiest->__cpu_power);
3381 tmp = sg_div_cpu_power(this,
3382 busiest_load_per_task * SCHED_LOAD_SCALE);
3383 pwr_move += this->__cpu_power *
3384 min(this_load_per_task, this_load + tmp);
3385 pwr_move /= SCHED_LOAD_SCALE;
3387 /* Move if we gain throughput */
3388 if (pwr_move > pwr_now)
3389 *imbalance = busiest_load_per_task;
3395 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3396 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3399 if (this == group_leader && group_leader != group_min) {
3400 *imbalance = min_load_per_task;
3410 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3413 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3414 unsigned long imbalance, const cpumask_t *cpus)
3416 struct rq *busiest = NULL, *rq;
3417 unsigned long max_load = 0;
3420 for_each_cpu_mask_nr(i, group->cpumask) {
3423 if (!cpu_isset(i, *cpus))
3427 wl = weighted_cpuload(i);
3429 if (rq->nr_running == 1 && wl > imbalance)
3432 if (wl > max_load) {
3442 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3443 * so long as it is large enough.
3445 #define MAX_PINNED_INTERVAL 512
3448 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3449 * tasks if there is an imbalance.
3451 static int load_balance(int this_cpu, struct rq *this_rq,
3452 struct sched_domain *sd, enum cpu_idle_type idle,
3453 int *balance, cpumask_t *cpus)
3455 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3456 struct sched_group *group;
3457 unsigned long imbalance;
3459 unsigned long flags;
3464 * When power savings policy is enabled for the parent domain, idle
3465 * sibling can pick up load irrespective of busy siblings. In this case,
3466 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3467 * portraying it as CPU_NOT_IDLE.
3469 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3470 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3473 schedstat_inc(sd, lb_count[idle]);
3477 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3484 schedstat_inc(sd, lb_nobusyg[idle]);
3488 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3490 schedstat_inc(sd, lb_nobusyq[idle]);
3494 BUG_ON(busiest == this_rq);
3496 schedstat_add(sd, lb_imbalance[idle], imbalance);
3499 if (busiest->nr_running > 1) {
3501 * Attempt to move tasks. If find_busiest_group has found
3502 * an imbalance but busiest->nr_running <= 1, the group is
3503 * still unbalanced. ld_moved simply stays zero, so it is
3504 * correctly treated as an imbalance.
3506 local_irq_save(flags);
3507 double_rq_lock(this_rq, busiest);
3508 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3509 imbalance, sd, idle, &all_pinned);
3510 double_rq_unlock(this_rq, busiest);
3511 local_irq_restore(flags);
3514 * some other cpu did the load balance for us.
3516 if (ld_moved && this_cpu != smp_processor_id())
3517 resched_cpu(this_cpu);
3519 /* All tasks on this runqueue were pinned by CPU affinity */
3520 if (unlikely(all_pinned)) {
3521 cpu_clear(cpu_of(busiest), *cpus);
3522 if (!cpus_empty(*cpus))
3529 schedstat_inc(sd, lb_failed[idle]);
3530 sd->nr_balance_failed++;
3532 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3534 spin_lock_irqsave(&busiest->lock, flags);
3536 /* don't kick the migration_thread, if the curr
3537 * task on busiest cpu can't be moved to this_cpu
3539 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3540 spin_unlock_irqrestore(&busiest->lock, flags);
3542 goto out_one_pinned;
3545 if (!busiest->active_balance) {
3546 busiest->active_balance = 1;
3547 busiest->push_cpu = this_cpu;
3550 spin_unlock_irqrestore(&busiest->lock, flags);
3552 wake_up_process(busiest->migration_thread);
3555 * We've kicked active balancing, reset the failure
3558 sd->nr_balance_failed = sd->cache_nice_tries+1;
3561 sd->nr_balance_failed = 0;
3563 if (likely(!active_balance)) {
3564 /* We were unbalanced, so reset the balancing interval */
3565 sd->balance_interval = sd->min_interval;
3568 * If we've begun active balancing, start to back off. This
3569 * case may not be covered by the all_pinned logic if there
3570 * is only 1 task on the busy runqueue (because we don't call
3573 if (sd->balance_interval < sd->max_interval)
3574 sd->balance_interval *= 2;
3577 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3578 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3584 schedstat_inc(sd, lb_balanced[idle]);
3586 sd->nr_balance_failed = 0;
3589 /* tune up the balancing interval */
3590 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3591 (sd->balance_interval < sd->max_interval))
3592 sd->balance_interval *= 2;
3594 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3595 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3606 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3607 * tasks if there is an imbalance.
3609 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3610 * this_rq is locked.
3613 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3616 struct sched_group *group;
3617 struct rq *busiest = NULL;
3618 unsigned long imbalance;
3626 * When power savings policy is enabled for the parent domain, idle
3627 * sibling can pick up load irrespective of busy siblings. In this case,
3628 * let the state of idle sibling percolate up as IDLE, instead of
3629 * portraying it as CPU_NOT_IDLE.
3631 if (sd->flags & SD_SHARE_CPUPOWER &&
3632 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3635 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3637 update_shares_locked(this_rq, sd);
3638 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3639 &sd_idle, cpus, NULL);
3641 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3645 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3647 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3651 BUG_ON(busiest == this_rq);
3653 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3656 if (busiest->nr_running > 1) {
3657 /* Attempt to move tasks */
3658 double_lock_balance(this_rq, busiest);
3659 /* this_rq->clock is already updated */
3660 update_rq_clock(busiest);
3661 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3662 imbalance, sd, CPU_NEWLY_IDLE,
3664 double_unlock_balance(this_rq, busiest);
3666 if (unlikely(all_pinned)) {
3667 cpu_clear(cpu_of(busiest), *cpus);
3668 if (!cpus_empty(*cpus))
3674 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3675 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3676 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3679 sd->nr_balance_failed = 0;
3681 update_shares_locked(this_rq, sd);
3685 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3686 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3687 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3689 sd->nr_balance_failed = 0;
3695 * idle_balance is called by schedule() if this_cpu is about to become
3696 * idle. Attempts to pull tasks from other CPUs.
3698 static void idle_balance(int this_cpu, struct rq *this_rq)
3700 struct sched_domain *sd;
3701 int pulled_task = -1;
3702 unsigned long next_balance = jiffies + HZ;
3705 for_each_domain(this_cpu, sd) {
3706 unsigned long interval;
3708 if (!(sd->flags & SD_LOAD_BALANCE))
3711 if (sd->flags & SD_BALANCE_NEWIDLE)
3712 /* If we've pulled tasks over stop searching: */
3713 pulled_task = load_balance_newidle(this_cpu, this_rq,
3716 interval = msecs_to_jiffies(sd->balance_interval);
3717 if (time_after(next_balance, sd->last_balance + interval))
3718 next_balance = sd->last_balance + interval;
3722 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3724 * We are going idle. next_balance may be set based on
3725 * a busy processor. So reset next_balance.
3727 this_rq->next_balance = next_balance;
3732 * active_load_balance is run by migration threads. It pushes running tasks
3733 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3734 * running on each physical CPU where possible, and avoids physical /
3735 * logical imbalances.
3737 * Called with busiest_rq locked.
3739 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3741 int target_cpu = busiest_rq->push_cpu;
3742 struct sched_domain *sd;
3743 struct rq *target_rq;
3745 /* Is there any task to move? */
3746 if (busiest_rq->nr_running <= 1)
3749 target_rq = cpu_rq(target_cpu);
3752 * This condition is "impossible", if it occurs
3753 * we need to fix it. Originally reported by
3754 * Bjorn Helgaas on a 128-cpu setup.
3756 BUG_ON(busiest_rq == target_rq);
3758 /* move a task from busiest_rq to target_rq */
3759 double_lock_balance(busiest_rq, target_rq);
3760 update_rq_clock(busiest_rq);
3761 update_rq_clock(target_rq);
3763 /* Search for an sd spanning us and the target CPU. */
3764 for_each_domain(target_cpu, sd) {
3765 if ((sd->flags & SD_LOAD_BALANCE) &&
3766 cpu_isset(busiest_cpu, sd->span))
3771 schedstat_inc(sd, alb_count);
3773 if (move_one_task(target_rq, target_cpu, busiest_rq,
3775 schedstat_inc(sd, alb_pushed);
3777 schedstat_inc(sd, alb_failed);
3779 double_unlock_balance(busiest_rq, target_rq);
3784 atomic_t load_balancer;
3786 } nohz ____cacheline_aligned = {
3787 .load_balancer = ATOMIC_INIT(-1),
3788 .cpu_mask = CPU_MASK_NONE,
3792 * This routine will try to nominate the ilb (idle load balancing)
3793 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3794 * load balancing on behalf of all those cpus. If all the cpus in the system
3795 * go into this tickless mode, then there will be no ilb owner (as there is
3796 * no need for one) and all the cpus will sleep till the next wakeup event
3799 * For the ilb owner, tick is not stopped. And this tick will be used
3800 * for idle load balancing. ilb owner will still be part of
3803 * While stopping the tick, this cpu will become the ilb owner if there
3804 * is no other owner. And will be the owner till that cpu becomes busy
3805 * or if all cpus in the system stop their ticks at which point
3806 * there is no need for ilb owner.
3808 * When the ilb owner becomes busy, it nominates another owner, during the
3809 * next busy scheduler_tick()
3811 int select_nohz_load_balancer(int stop_tick)
3813 int cpu = smp_processor_id();
3816 cpu_set(cpu, nohz.cpu_mask);
3817 cpu_rq(cpu)->in_nohz_recently = 1;
3820 * If we are going offline and still the leader, give up!
3822 if (!cpu_active(cpu) &&
3823 atomic_read(&nohz.load_balancer) == cpu) {
3824 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3829 /* time for ilb owner also to sleep */
3830 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3831 if (atomic_read(&nohz.load_balancer) == cpu)
3832 atomic_set(&nohz.load_balancer, -1);
3836 if (atomic_read(&nohz.load_balancer) == -1) {
3837 /* make me the ilb owner */
3838 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3840 } else if (atomic_read(&nohz.load_balancer) == cpu)
3843 if (!cpu_isset(cpu, nohz.cpu_mask))
3846 cpu_clear(cpu, nohz.cpu_mask);
3848 if (atomic_read(&nohz.load_balancer) == cpu)
3849 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3856 static DEFINE_SPINLOCK(balancing);
3859 * It checks each scheduling domain to see if it is due to be balanced,
3860 * and initiates a balancing operation if so.
3862 * Balancing parameters are set up in arch_init_sched_domains.
3864 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3867 struct rq *rq = cpu_rq(cpu);
3868 unsigned long interval;
3869 struct sched_domain *sd;
3870 /* Earliest time when we have to do rebalance again */
3871 unsigned long next_balance = jiffies + 60*HZ;
3872 int update_next_balance = 0;
3876 for_each_domain(cpu, sd) {
3877 if (!(sd->flags & SD_LOAD_BALANCE))
3880 interval = sd->balance_interval;
3881 if (idle != CPU_IDLE)
3882 interval *= sd->busy_factor;
3884 /* scale ms to jiffies */
3885 interval = msecs_to_jiffies(interval);
3886 if (unlikely(!interval))
3888 if (interval > HZ*NR_CPUS/10)
3889 interval = HZ*NR_CPUS/10;
3891 need_serialize = sd->flags & SD_SERIALIZE;
3893 if (need_serialize) {
3894 if (!spin_trylock(&balancing))
3898 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3899 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3901 * We've pulled tasks over so either we're no
3902 * longer idle, or one of our SMT siblings is
3905 idle = CPU_NOT_IDLE;
3907 sd->last_balance = jiffies;
3910 spin_unlock(&balancing);
3912 if (time_after(next_balance, sd->last_balance + interval)) {
3913 next_balance = sd->last_balance + interval;
3914 update_next_balance = 1;
3918 * Stop the load balance at this level. There is another
3919 * CPU in our sched group which is doing load balancing more
3927 * next_balance will be updated only when there is a need.
3928 * When the cpu is attached to null domain for ex, it will not be
3931 if (likely(update_next_balance))
3932 rq->next_balance = next_balance;
3936 * run_rebalance_domains is triggered when needed from the scheduler tick.
3937 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3938 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3940 static void run_rebalance_domains(struct softirq_action *h)
3942 int this_cpu = smp_processor_id();
3943 struct rq *this_rq = cpu_rq(this_cpu);
3944 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3945 CPU_IDLE : CPU_NOT_IDLE;
3947 rebalance_domains(this_cpu, idle);
3951 * If this cpu is the owner for idle load balancing, then do the
3952 * balancing on behalf of the other idle cpus whose ticks are
3955 if (this_rq->idle_at_tick &&
3956 atomic_read(&nohz.load_balancer) == this_cpu) {
3957 cpumask_t cpus = nohz.cpu_mask;
3961 cpu_clear(this_cpu, cpus);
3962 for_each_cpu_mask_nr(balance_cpu, cpus) {
3964 * If this cpu gets work to do, stop the load balancing
3965 * work being done for other cpus. Next load
3966 * balancing owner will pick it up.
3971 rebalance_domains(balance_cpu, CPU_IDLE);
3973 rq = cpu_rq(balance_cpu);
3974 if (time_after(this_rq->next_balance, rq->next_balance))
3975 this_rq->next_balance = rq->next_balance;
3982 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3984 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3985 * idle load balancing owner or decide to stop the periodic load balancing,
3986 * if the whole system is idle.
3988 static inline void trigger_load_balance(struct rq *rq, int cpu)
3992 * If we were in the nohz mode recently and busy at the current
3993 * scheduler tick, then check if we need to nominate new idle
3996 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3997 rq->in_nohz_recently = 0;
3999 if (atomic_read(&nohz.load_balancer) == cpu) {
4000 cpu_clear(cpu, nohz.cpu_mask);
4001 atomic_set(&nohz.load_balancer, -1);
4004 if (atomic_read(&nohz.load_balancer) == -1) {
4006 * simple selection for now: Nominate the
4007 * first cpu in the nohz list to be the next
4010 * TBD: Traverse the sched domains and nominate
4011 * the nearest cpu in the nohz.cpu_mask.
4013 int ilb = first_cpu(nohz.cpu_mask);
4015 if (ilb < nr_cpu_ids)
4021 * If this cpu is idle and doing idle load balancing for all the
4022 * cpus with ticks stopped, is it time for that to stop?
4024 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4025 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4031 * If this cpu is idle and the idle load balancing is done by
4032 * someone else, then no need raise the SCHED_SOFTIRQ
4034 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4035 cpu_isset(cpu, nohz.cpu_mask))
4038 if (time_after_eq(jiffies, rq->next_balance))
4039 raise_softirq(SCHED_SOFTIRQ);
4042 #else /* CONFIG_SMP */
4045 * on UP we do not need to balance between CPUs:
4047 static inline void idle_balance(int cpu, struct rq *rq)
4053 DEFINE_PER_CPU(struct kernel_stat, kstat);
4055 EXPORT_PER_CPU_SYMBOL(kstat);
4058 * Return p->sum_exec_runtime plus any more ns on the sched_clock
4059 * that have not yet been banked in case the task is currently running.
4061 unsigned long long task_sched_runtime(struct task_struct *p)
4063 unsigned long flags;
4067 rq = task_rq_lock(p, &flags);
4068 ns = p->se.sum_exec_runtime;
4069 if (task_current(rq, p)) {
4070 update_rq_clock(rq);
4071 delta_exec = rq->clock - p->se.exec_start;
4072 if ((s64)delta_exec > 0)
4075 task_rq_unlock(rq, &flags);
4081 * Account user cpu time to a process.
4082 * @p: the process that the cpu time gets accounted to
4083 * @cputime: the cpu time spent in user space since the last update
4085 void account_user_time(struct task_struct *p, cputime_t cputime)
4087 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4090 p->utime = cputime_add(p->utime, cputime);
4092 /* Add user time to cpustat. */
4093 tmp = cputime_to_cputime64(cputime);
4094 if (TASK_NICE(p) > 0)
4095 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4097 cpustat->user = cputime64_add(cpustat->user, tmp);
4098 /* Account for user time used */
4099 acct_update_integrals(p);
4103 * Account guest cpu time to a process.
4104 * @p: the process that the cpu time gets accounted to
4105 * @cputime: the cpu time spent in virtual machine since the last update
4107 static void account_guest_time(struct task_struct *p, cputime_t cputime)
4110 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4112 tmp = cputime_to_cputime64(cputime);
4114 p->utime = cputime_add(p->utime, cputime);
4115 p->gtime = cputime_add(p->gtime, cputime);
4117 cpustat->user = cputime64_add(cpustat->user, tmp);
4118 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4122 * Account scaled user cpu time to a process.
4123 * @p: the process that the cpu time gets accounted to
4124 * @cputime: the cpu time spent in user space since the last update
4126 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4128 p->utimescaled = cputime_add(p->utimescaled, cputime);
4132 * Account system cpu time to a process.
4133 * @p: the process that the cpu time gets accounted to
4134 * @hardirq_offset: the offset to subtract from hardirq_count()
4135 * @cputime: the cpu time spent in kernel space since the last update
4137 void account_system_time(struct task_struct *p, int hardirq_offset,
4140 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4141 struct rq *rq = this_rq();
4144 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4145 account_guest_time(p, cputime);
4149 p->stime = cputime_add(p->stime, cputime);
4151 /* Add system time to cpustat. */
4152 tmp = cputime_to_cputime64(cputime);
4153 if (hardirq_count() - hardirq_offset)
4154 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4155 else if (softirq_count())
4156 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4157 else if (p != rq->idle)
4158 cpustat->system = cputime64_add(cpustat->system, tmp);
4159 else if (atomic_read(&rq->nr_iowait) > 0)
4160 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4162 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4163 /* Account for system time used */
4164 acct_update_integrals(p);
4168 * Account scaled system cpu time to a process.
4169 * @p: the process that the cpu time gets accounted to
4170 * @hardirq_offset: the offset to subtract from hardirq_count()
4171 * @cputime: the cpu time spent in kernel space since the last update
4173 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4175 p->stimescaled = cputime_add(p->stimescaled, cputime);
4179 * Account for involuntary wait time.
4180 * @p: the process from which the cpu time has been stolen
4181 * @steal: the cpu time spent in involuntary wait
4183 void account_steal_time(struct task_struct *p, cputime_t steal)
4185 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4186 cputime64_t tmp = cputime_to_cputime64(steal);
4187 struct rq *rq = this_rq();
4189 if (p == rq->idle) {
4190 p->stime = cputime_add(p->stime, steal);
4191 if (atomic_read(&rq->nr_iowait) > 0)
4192 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4194 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4196 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4200 * This function gets called by the timer code, with HZ frequency.
4201 * We call it with interrupts disabled.
4203 * It also gets called by the fork code, when changing the parent's
4206 void scheduler_tick(void)
4208 int cpu = smp_processor_id();
4209 struct rq *rq = cpu_rq(cpu);
4210 struct task_struct *curr = rq->curr;
4214 spin_lock(&rq->lock);
4215 update_rq_clock(rq);
4216 update_cpu_load(rq);
4217 curr->sched_class->task_tick(rq, curr, 0);
4218 spin_unlock(&rq->lock);
4221 rq->idle_at_tick = idle_cpu(cpu);
4222 trigger_load_balance(rq, cpu);
4226 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4227 defined(CONFIG_PREEMPT_TRACER))
4229 static inline unsigned long get_parent_ip(unsigned long addr)
4231 if (in_lock_functions(addr)) {
4232 addr = CALLER_ADDR2;
4233 if (in_lock_functions(addr))
4234 addr = CALLER_ADDR3;
4239 void __kprobes add_preempt_count(int val)
4241 #ifdef CONFIG_DEBUG_PREEMPT
4245 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4248 preempt_count() += val;
4249 #ifdef CONFIG_DEBUG_PREEMPT
4251 * Spinlock count overflowing soon?
4253 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4256 if (preempt_count() == val)
4257 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4259 EXPORT_SYMBOL(add_preempt_count);
4261 void __kprobes sub_preempt_count(int val)
4263 #ifdef CONFIG_DEBUG_PREEMPT
4267 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4270 * Is the spinlock portion underflowing?
4272 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4273 !(preempt_count() & PREEMPT_MASK)))
4277 if (preempt_count() == val)
4278 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4279 preempt_count() -= val;
4281 EXPORT_SYMBOL(sub_preempt_count);
4286 * Print scheduling while atomic bug:
4288 static noinline void __schedule_bug(struct task_struct *prev)
4290 struct pt_regs *regs = get_irq_regs();
4292 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4293 prev->comm, prev->pid, preempt_count());
4295 debug_show_held_locks(prev);
4297 if (irqs_disabled())
4298 print_irqtrace_events(prev);
4307 * Various schedule()-time debugging checks and statistics:
4309 static inline void schedule_debug(struct task_struct *prev)
4312 * Test if we are atomic. Since do_exit() needs to call into
4313 * schedule() atomically, we ignore that path for now.
4314 * Otherwise, whine if we are scheduling when we should not be.
4316 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4317 __schedule_bug(prev);
4319 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4321 schedstat_inc(this_rq(), sched_count);
4322 #ifdef CONFIG_SCHEDSTATS
4323 if (unlikely(prev->lock_depth >= 0)) {
4324 schedstat_inc(this_rq(), bkl_count);
4325 schedstat_inc(prev, sched_info.bkl_count);
4331 * Pick up the highest-prio task:
4333 static inline struct task_struct *
4334 pick_next_task(struct rq *rq, struct task_struct *prev)
4336 const struct sched_class *class;
4337 struct task_struct *p;
4340 * Optimization: we know that if all tasks are in
4341 * the fair class we can call that function directly:
4343 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4344 p = fair_sched_class.pick_next_task(rq);
4349 class = sched_class_highest;
4351 p = class->pick_next_task(rq);
4355 * Will never be NULL as the idle class always
4356 * returns a non-NULL p:
4358 class = class->next;
4363 * schedule() is the main scheduler function.
4365 asmlinkage void __sched schedule(void)
4367 struct task_struct *prev, *next;
4368 unsigned long *switch_count;
4374 cpu = smp_processor_id();
4378 switch_count = &prev->nivcsw;
4380 release_kernel_lock(prev);
4381 need_resched_nonpreemptible:
4383 schedule_debug(prev);
4385 if (sched_feat(HRTICK))
4389 * Do the rq-clock update outside the rq lock:
4391 local_irq_disable();
4392 update_rq_clock(rq);
4393 spin_lock(&rq->lock);
4394 clear_tsk_need_resched(prev);
4396 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4397 if (unlikely(signal_pending_state(prev->state, prev)))
4398 prev->state = TASK_RUNNING;
4400 deactivate_task(rq, prev, 1);
4401 switch_count = &prev->nvcsw;
4405 if (prev->sched_class->pre_schedule)
4406 prev->sched_class->pre_schedule(rq, prev);
4409 if (unlikely(!rq->nr_running))
4410 idle_balance(cpu, rq);
4412 prev->sched_class->put_prev_task(rq, prev);
4413 next = pick_next_task(rq, prev);
4415 if (likely(prev != next)) {
4416 sched_info_switch(prev, next);
4422 context_switch(rq, prev, next); /* unlocks the rq */
4424 * the context switch might have flipped the stack from under
4425 * us, hence refresh the local variables.
4427 cpu = smp_processor_id();
4430 spin_unlock_irq(&rq->lock);
4432 if (unlikely(reacquire_kernel_lock(current) < 0))
4433 goto need_resched_nonpreemptible;
4435 preempt_enable_no_resched();
4436 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4439 EXPORT_SYMBOL(schedule);
4441 #ifdef CONFIG_PREEMPT
4443 * this is the entry point to schedule() from in-kernel preemption
4444 * off of preempt_enable. Kernel preemptions off return from interrupt
4445 * occur there and call schedule directly.
4447 asmlinkage void __sched preempt_schedule(void)
4449 struct thread_info *ti = current_thread_info();
4452 * If there is a non-zero preempt_count or interrupts are disabled,
4453 * we do not want to preempt the current task. Just return..
4455 if (likely(ti->preempt_count || irqs_disabled()))
4459 add_preempt_count(PREEMPT_ACTIVE);
4461 sub_preempt_count(PREEMPT_ACTIVE);
4464 * Check again in case we missed a preemption opportunity
4465 * between schedule and now.
4468 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4470 EXPORT_SYMBOL(preempt_schedule);
4473 * this is the entry point to schedule() from kernel preemption
4474 * off of irq context.
4475 * Note, that this is called and return with irqs disabled. This will
4476 * protect us against recursive calling from irq.
4478 asmlinkage void __sched preempt_schedule_irq(void)
4480 struct thread_info *ti = current_thread_info();
4482 /* Catch callers which need to be fixed */
4483 BUG_ON(ti->preempt_count || !irqs_disabled());
4486 add_preempt_count(PREEMPT_ACTIVE);
4489 local_irq_disable();
4490 sub_preempt_count(PREEMPT_ACTIVE);
4493 * Check again in case we missed a preemption opportunity
4494 * between schedule and now.
4497 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4500 #endif /* CONFIG_PREEMPT */
4502 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4505 return try_to_wake_up(curr->private, mode, sync);
4507 EXPORT_SYMBOL(default_wake_function);
4510 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4511 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4512 * number) then we wake all the non-exclusive tasks and one exclusive task.
4514 * There are circumstances in which we can try to wake a task which has already
4515 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4516 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4518 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4519 int nr_exclusive, int sync, void *key)
4521 wait_queue_t *curr, *next;
4523 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4524 unsigned flags = curr->flags;
4526 if (curr->func(curr, mode, sync, key) &&
4527 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4533 * __wake_up - wake up threads blocked on a waitqueue.
4535 * @mode: which threads
4536 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4537 * @key: is directly passed to the wakeup function
4539 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4540 int nr_exclusive, void *key)
4542 unsigned long flags;
4544 spin_lock_irqsave(&q->lock, flags);
4545 __wake_up_common(q, mode, nr_exclusive, 0, key);
4546 spin_unlock_irqrestore(&q->lock, flags);
4548 EXPORT_SYMBOL(__wake_up);
4551 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4553 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4555 __wake_up_common(q, mode, 1, 0, NULL);
4559 * __wake_up_sync - wake up threads blocked on a waitqueue.
4561 * @mode: which threads
4562 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4564 * The sync wakeup differs that the waker knows that it will schedule
4565 * away soon, so while the target thread will be woken up, it will not
4566 * be migrated to another CPU - ie. the two threads are 'synchronized'
4567 * with each other. This can prevent needless bouncing between CPUs.
4569 * On UP it can prevent extra preemption.
4572 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4574 unsigned long flags;
4580 if (unlikely(!nr_exclusive))
4583 spin_lock_irqsave(&q->lock, flags);
4584 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4585 spin_unlock_irqrestore(&q->lock, flags);
4587 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4589 void complete(struct completion *x)
4591 unsigned long flags;
4593 spin_lock_irqsave(&x->wait.lock, flags);
4595 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4596 spin_unlock_irqrestore(&x->wait.lock, flags);
4598 EXPORT_SYMBOL(complete);
4600 void complete_all(struct completion *x)
4602 unsigned long flags;
4604 spin_lock_irqsave(&x->wait.lock, flags);
4605 x->done += UINT_MAX/2;
4606 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4607 spin_unlock_irqrestore(&x->wait.lock, flags);
4609 EXPORT_SYMBOL(complete_all);
4611 static inline long __sched
4612 do_wait_for_common(struct completion *x, long timeout, int state)
4615 DECLARE_WAITQUEUE(wait, current);
4617 wait.flags |= WQ_FLAG_EXCLUSIVE;
4618 __add_wait_queue_tail(&x->wait, &wait);
4620 if ((state == TASK_INTERRUPTIBLE &&
4621 signal_pending(current)) ||
4622 (state == TASK_KILLABLE &&
4623 fatal_signal_pending(current))) {
4624 timeout = -ERESTARTSYS;
4627 __set_current_state(state);
4628 spin_unlock_irq(&x->wait.lock);
4629 timeout = schedule_timeout(timeout);
4630 spin_lock_irq(&x->wait.lock);
4631 } while (!x->done && timeout);
4632 __remove_wait_queue(&x->wait, &wait);
4637 return timeout ?: 1;
4641 wait_for_common(struct completion *x, long timeout, int state)
4645 spin_lock_irq(&x->wait.lock);
4646 timeout = do_wait_for_common(x, timeout, state);
4647 spin_unlock_irq(&x->wait.lock);
4651 void __sched wait_for_completion(struct completion *x)
4653 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4655 EXPORT_SYMBOL(wait_for_completion);
4657 unsigned long __sched
4658 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4660 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4662 EXPORT_SYMBOL(wait_for_completion_timeout);
4664 int __sched wait_for_completion_interruptible(struct completion *x)
4666 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4667 if (t == -ERESTARTSYS)
4671 EXPORT_SYMBOL(wait_for_completion_interruptible);
4673 unsigned long __sched
4674 wait_for_completion_interruptible_timeout(struct completion *x,
4675 unsigned long timeout)
4677 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4679 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4681 int __sched wait_for_completion_killable(struct completion *x)
4683 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4684 if (t == -ERESTARTSYS)
4688 EXPORT_SYMBOL(wait_for_completion_killable);
4691 * try_wait_for_completion - try to decrement a completion without blocking
4692 * @x: completion structure
4694 * Returns: 0 if a decrement cannot be done without blocking
4695 * 1 if a decrement succeeded.
4697 * If a completion is being used as a counting completion,
4698 * attempt to decrement the counter without blocking. This
4699 * enables us to avoid waiting if the resource the completion
4700 * is protecting is not available.
4702 bool try_wait_for_completion(struct completion *x)
4706 spin_lock_irq(&x->wait.lock);
4711 spin_unlock_irq(&x->wait.lock);
4714 EXPORT_SYMBOL(try_wait_for_completion);
4717 * completion_done - Test to see if a completion has any waiters
4718 * @x: completion structure
4720 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4721 * 1 if there are no waiters.
4724 bool completion_done(struct completion *x)
4728 spin_lock_irq(&x->wait.lock);
4731 spin_unlock_irq(&x->wait.lock);
4734 EXPORT_SYMBOL(completion_done);
4737 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4739 unsigned long flags;
4742 init_waitqueue_entry(&wait, current);
4744 __set_current_state(state);
4746 spin_lock_irqsave(&q->lock, flags);
4747 __add_wait_queue(q, &wait);
4748 spin_unlock(&q->lock);
4749 timeout = schedule_timeout(timeout);
4750 spin_lock_irq(&q->lock);
4751 __remove_wait_queue(q, &wait);
4752 spin_unlock_irqrestore(&q->lock, flags);
4757 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4759 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4761 EXPORT_SYMBOL(interruptible_sleep_on);
4764 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4766 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4768 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4770 void __sched sleep_on(wait_queue_head_t *q)
4772 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4774 EXPORT_SYMBOL(sleep_on);
4776 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4778 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4780 EXPORT_SYMBOL(sleep_on_timeout);
4782 #ifdef CONFIG_RT_MUTEXES
4785 * rt_mutex_setprio - set the current priority of a task
4787 * @prio: prio value (kernel-internal form)
4789 * This function changes the 'effective' priority of a task. It does
4790 * not touch ->normal_prio like __setscheduler().
4792 * Used by the rt_mutex code to implement priority inheritance logic.
4794 void rt_mutex_setprio(struct task_struct *p, int prio)
4796 unsigned long flags;
4797 int oldprio, on_rq, running;
4799 const struct sched_class *prev_class = p->sched_class;
4801 BUG_ON(prio < 0 || prio > MAX_PRIO);
4803 rq = task_rq_lock(p, &flags);
4804 update_rq_clock(rq);
4807 on_rq = p->se.on_rq;
4808 running = task_current(rq, p);
4810 dequeue_task(rq, p, 0);
4812 p->sched_class->put_prev_task(rq, p);
4815 p->sched_class = &rt_sched_class;
4817 p->sched_class = &fair_sched_class;
4822 p->sched_class->set_curr_task(rq);
4824 enqueue_task(rq, p, 0);
4826 check_class_changed(rq, p, prev_class, oldprio, running);
4828 task_rq_unlock(rq, &flags);
4833 void set_user_nice(struct task_struct *p, long nice)
4835 int old_prio, delta, on_rq;
4836 unsigned long flags;
4839 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4842 * We have to be careful, if called from sys_setpriority(),
4843 * the task might be in the middle of scheduling on another CPU.
4845 rq = task_rq_lock(p, &flags);
4846 update_rq_clock(rq);
4848 * The RT priorities are set via sched_setscheduler(), but we still
4849 * allow the 'normal' nice value to be set - but as expected
4850 * it wont have any effect on scheduling until the task is
4851 * SCHED_FIFO/SCHED_RR:
4853 if (task_has_rt_policy(p)) {
4854 p->static_prio = NICE_TO_PRIO(nice);
4857 on_rq = p->se.on_rq;
4859 dequeue_task(rq, p, 0);
4861 p->static_prio = NICE_TO_PRIO(nice);
4864 p->prio = effective_prio(p);
4865 delta = p->prio - old_prio;
4868 enqueue_task(rq, p, 0);
4870 * If the task increased its priority or is running and
4871 * lowered its priority, then reschedule its CPU:
4873 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4874 resched_task(rq->curr);
4877 task_rq_unlock(rq, &flags);
4879 EXPORT_SYMBOL(set_user_nice);
4882 * can_nice - check if a task can reduce its nice value
4886 int can_nice(const struct task_struct *p, const int nice)
4888 /* convert nice value [19,-20] to rlimit style value [1,40] */
4889 int nice_rlim = 20 - nice;
4891 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4892 capable(CAP_SYS_NICE));
4895 #ifdef __ARCH_WANT_SYS_NICE
4898 * sys_nice - change the priority of the current process.
4899 * @increment: priority increment
4901 * sys_setpriority is a more generic, but much slower function that
4902 * does similar things.
4904 asmlinkage long sys_nice(int increment)
4909 * Setpriority might change our priority at the same moment.
4910 * We don't have to worry. Conceptually one call occurs first
4911 * and we have a single winner.
4913 if (increment < -40)
4918 nice = PRIO_TO_NICE(current->static_prio) + increment;
4924 if (increment < 0 && !can_nice(current, nice))
4927 retval = security_task_setnice(current, nice);
4931 set_user_nice(current, nice);
4938 * task_prio - return the priority value of a given task.
4939 * @p: the task in question.
4941 * This is the priority value as seen by users in /proc.
4942 * RT tasks are offset by -200. Normal tasks are centered
4943 * around 0, value goes from -16 to +15.
4945 int task_prio(const struct task_struct *p)
4947 return p->prio - MAX_RT_PRIO;
4951 * task_nice - return the nice value of a given task.
4952 * @p: the task in question.
4954 int task_nice(const struct task_struct *p)
4956 return TASK_NICE(p);
4958 EXPORT_SYMBOL(task_nice);
4961 * idle_cpu - is a given cpu idle currently?
4962 * @cpu: the processor in question.
4964 int idle_cpu(int cpu)
4966 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4970 * idle_task - return the idle task for a given cpu.
4971 * @cpu: the processor in question.
4973 struct task_struct *idle_task(int cpu)
4975 return cpu_rq(cpu)->idle;
4979 * find_process_by_pid - find a process with a matching PID value.
4980 * @pid: the pid in question.
4982 static struct task_struct *find_process_by_pid(pid_t pid)
4984 return pid ? find_task_by_vpid(pid) : current;
4987 /* Actually do priority change: must hold rq lock. */
4989 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4991 BUG_ON(p->se.on_rq);
4994 switch (p->policy) {
4998 p->sched_class = &fair_sched_class;
5002 p->sched_class = &rt_sched_class;
5006 p->rt_priority = prio;
5007 p->normal_prio = normal_prio(p);
5008 /* we are holding p->pi_lock already */
5009 p->prio = rt_mutex_getprio(p);
5013 static int __sched_setscheduler(struct task_struct *p, int policy,
5014 struct sched_param *param, bool user)
5016 int retval, oldprio, oldpolicy = -1, on_rq, running;
5017 unsigned long flags;
5018 const struct sched_class *prev_class = p->sched_class;
5021 /* may grab non-irq protected spin_locks */
5022 BUG_ON(in_interrupt());
5024 /* double check policy once rq lock held */
5026 policy = oldpolicy = p->policy;
5027 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5028 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5029 policy != SCHED_IDLE)
5032 * Valid priorities for SCHED_FIFO and SCHED_RR are
5033 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5034 * SCHED_BATCH and SCHED_IDLE is 0.
5036 if (param->sched_priority < 0 ||
5037 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5038 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5040 if (rt_policy(policy) != (param->sched_priority != 0))
5044 * Allow unprivileged RT tasks to decrease priority:
5046 if (user && !capable(CAP_SYS_NICE)) {
5047 if (rt_policy(policy)) {
5048 unsigned long rlim_rtprio;
5050 if (!lock_task_sighand(p, &flags))
5052 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5053 unlock_task_sighand(p, &flags);
5055 /* can't set/change the rt policy */
5056 if (policy != p->policy && !rlim_rtprio)
5059 /* can't increase priority */
5060 if (param->sched_priority > p->rt_priority &&
5061 param->sched_priority > rlim_rtprio)
5065 * Like positive nice levels, dont allow tasks to
5066 * move out of SCHED_IDLE either:
5068 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5071 /* can't change other user's priorities */
5072 if ((current->euid != p->euid) &&
5073 (current->euid != p->uid))
5078 #ifdef CONFIG_RT_GROUP_SCHED
5080 * Do not allow realtime tasks into groups that have no runtime
5083 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5084 task_group(p)->rt_bandwidth.rt_runtime == 0)
5088 retval = security_task_setscheduler(p, policy, param);
5094 * make sure no PI-waiters arrive (or leave) while we are
5095 * changing the priority of the task:
5097 spin_lock_irqsave(&p->pi_lock, flags);
5099 * To be able to change p->policy safely, the apropriate
5100 * runqueue lock must be held.
5102 rq = __task_rq_lock(p);
5103 /* recheck policy now with rq lock held */
5104 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5105 policy = oldpolicy = -1;
5106 __task_rq_unlock(rq);
5107 spin_unlock_irqrestore(&p->pi_lock, flags);
5110 update_rq_clock(rq);
5111 on_rq = p->se.on_rq;
5112 running = task_current(rq, p);
5114 deactivate_task(rq, p, 0);
5116 p->sched_class->put_prev_task(rq, p);
5119 __setscheduler(rq, p, policy, param->sched_priority);
5122 p->sched_class->set_curr_task(rq);
5124 activate_task(rq, p, 0);
5126 check_class_changed(rq, p, prev_class, oldprio, running);
5128 __task_rq_unlock(rq);
5129 spin_unlock_irqrestore(&p->pi_lock, flags);
5131 rt_mutex_adjust_pi(p);
5137 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5138 * @p: the task in question.
5139 * @policy: new policy.
5140 * @param: structure containing the new RT priority.
5142 * NOTE that the task may be already dead.
5144 int sched_setscheduler(struct task_struct *p, int policy,
5145 struct sched_param *param)
5147 return __sched_setscheduler(p, policy, param, true);
5149 EXPORT_SYMBOL_GPL(sched_setscheduler);
5152 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5153 * @p: the task in question.
5154 * @policy: new policy.
5155 * @param: structure containing the new RT priority.
5157 * Just like sched_setscheduler, only don't bother checking if the
5158 * current context has permission. For example, this is needed in
5159 * stop_machine(): we create temporary high priority worker threads,
5160 * but our caller might not have that capability.
5162 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5163 struct sched_param *param)
5165 return __sched_setscheduler(p, policy, param, false);
5169 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5171 struct sched_param lparam;
5172 struct task_struct *p;
5175 if (!param || pid < 0)
5177 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5182 p = find_process_by_pid(pid);
5184 retval = sched_setscheduler(p, policy, &lparam);
5191 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5192 * @pid: the pid in question.
5193 * @policy: new policy.
5194 * @param: structure containing the new RT priority.
5197 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5199 /* negative values for policy are not valid */
5203 return do_sched_setscheduler(pid, policy, param);
5207 * sys_sched_setparam - set/change the RT priority of a thread
5208 * @pid: the pid in question.
5209 * @param: structure containing the new RT priority.
5211 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5213 return do_sched_setscheduler(pid, -1, param);
5217 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5218 * @pid: the pid in question.
5220 asmlinkage long sys_sched_getscheduler(pid_t pid)
5222 struct task_struct *p;
5229 read_lock(&tasklist_lock);
5230 p = find_process_by_pid(pid);
5232 retval = security_task_getscheduler(p);
5236 read_unlock(&tasklist_lock);
5241 * sys_sched_getscheduler - get the RT priority of a thread
5242 * @pid: the pid in question.
5243 * @param: structure containing the RT priority.
5245 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5247 struct sched_param lp;
5248 struct task_struct *p;
5251 if (!param || pid < 0)
5254 read_lock(&tasklist_lock);
5255 p = find_process_by_pid(pid);
5260 retval = security_task_getscheduler(p);
5264 lp.sched_priority = p->rt_priority;
5265 read_unlock(&tasklist_lock);
5268 * This one might sleep, we cannot do it with a spinlock held ...
5270 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5275 read_unlock(&tasklist_lock);
5279 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5281 cpumask_t cpus_allowed;
5282 cpumask_t new_mask = *in_mask;
5283 struct task_struct *p;
5287 read_lock(&tasklist_lock);
5289 p = find_process_by_pid(pid);
5291 read_unlock(&tasklist_lock);
5297 * It is not safe to call set_cpus_allowed with the
5298 * tasklist_lock held. We will bump the task_struct's
5299 * usage count and then drop tasklist_lock.
5302 read_unlock(&tasklist_lock);
5305 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5306 !capable(CAP_SYS_NICE))
5309 retval = security_task_setscheduler(p, 0, NULL);
5313 cpuset_cpus_allowed(p, &cpus_allowed);
5314 cpus_and(new_mask, new_mask, cpus_allowed);
5316 retval = set_cpus_allowed_ptr(p, &new_mask);
5319 cpuset_cpus_allowed(p, &cpus_allowed);
5320 if (!cpus_subset(new_mask, cpus_allowed)) {
5322 * We must have raced with a concurrent cpuset
5323 * update. Just reset the cpus_allowed to the
5324 * cpuset's cpus_allowed
5326 new_mask = cpus_allowed;
5336 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5337 cpumask_t *new_mask)
5339 if (len < sizeof(cpumask_t)) {
5340 memset(new_mask, 0, sizeof(cpumask_t));
5341 } else if (len > sizeof(cpumask_t)) {
5342 len = sizeof(cpumask_t);
5344 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5348 * sys_sched_setaffinity - set the cpu affinity of a process
5349 * @pid: pid of the process
5350 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5351 * @user_mask_ptr: user-space pointer to the new cpu mask
5353 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5354 unsigned long __user *user_mask_ptr)
5359 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5363 return sched_setaffinity(pid, &new_mask);
5366 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5368 struct task_struct *p;
5372 read_lock(&tasklist_lock);
5375 p = find_process_by_pid(pid);
5379 retval = security_task_getscheduler(p);
5383 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5386 read_unlock(&tasklist_lock);
5393 * sys_sched_getaffinity - get the cpu affinity of a process
5394 * @pid: pid of the process
5395 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5396 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5398 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5399 unsigned long __user *user_mask_ptr)
5404 if (len < sizeof(cpumask_t))
5407 ret = sched_getaffinity(pid, &mask);
5411 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5414 return sizeof(cpumask_t);
5418 * sys_sched_yield - yield the current processor to other threads.
5420 * This function yields the current CPU to other tasks. If there are no
5421 * other threads running on this CPU then this function will return.
5423 asmlinkage long sys_sched_yield(void)
5425 struct rq *rq = this_rq_lock();
5427 schedstat_inc(rq, yld_count);
5428 current->sched_class->yield_task(rq);
5431 * Since we are going to call schedule() anyway, there's
5432 * no need to preempt or enable interrupts:
5434 __release(rq->lock);
5435 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5436 _raw_spin_unlock(&rq->lock);
5437 preempt_enable_no_resched();
5444 static void __cond_resched(void)
5446 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5447 __might_sleep(__FILE__, __LINE__);
5450 * The BKS might be reacquired before we have dropped
5451 * PREEMPT_ACTIVE, which could trigger a second
5452 * cond_resched() call.
5455 add_preempt_count(PREEMPT_ACTIVE);
5457 sub_preempt_count(PREEMPT_ACTIVE);
5458 } while (need_resched());
5461 int __sched _cond_resched(void)
5463 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5464 system_state == SYSTEM_RUNNING) {
5470 EXPORT_SYMBOL(_cond_resched);
5473 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5474 * call schedule, and on return reacquire the lock.
5476 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5477 * operations here to prevent schedule() from being called twice (once via
5478 * spin_unlock(), once by hand).
5480 int cond_resched_lock(spinlock_t *lock)
5482 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5485 if (spin_needbreak(lock) || resched) {
5487 if (resched && need_resched())
5496 EXPORT_SYMBOL(cond_resched_lock);
5498 int __sched cond_resched_softirq(void)
5500 BUG_ON(!in_softirq());
5502 if (need_resched() && system_state == SYSTEM_RUNNING) {
5510 EXPORT_SYMBOL(cond_resched_softirq);
5513 * yield - yield the current processor to other threads.
5515 * This is a shortcut for kernel-space yielding - it marks the
5516 * thread runnable and calls sys_sched_yield().
5518 void __sched yield(void)
5520 set_current_state(TASK_RUNNING);
5523 EXPORT_SYMBOL(yield);
5526 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5527 * that process accounting knows that this is a task in IO wait state.
5529 * But don't do that if it is a deliberate, throttling IO wait (this task
5530 * has set its backing_dev_info: the queue against which it should throttle)
5532 void __sched io_schedule(void)
5534 struct rq *rq = &__raw_get_cpu_var(runqueues);
5536 delayacct_blkio_start();
5537 atomic_inc(&rq->nr_iowait);
5539 atomic_dec(&rq->nr_iowait);
5540 delayacct_blkio_end();
5542 EXPORT_SYMBOL(io_schedule);
5544 long __sched io_schedule_timeout(long timeout)
5546 struct rq *rq = &__raw_get_cpu_var(runqueues);
5549 delayacct_blkio_start();
5550 atomic_inc(&rq->nr_iowait);
5551 ret = schedule_timeout(timeout);
5552 atomic_dec(&rq->nr_iowait);
5553 delayacct_blkio_end();
5558 * sys_sched_get_priority_max - return maximum RT priority.
5559 * @policy: scheduling class.
5561 * this syscall returns the maximum rt_priority that can be used
5562 * by a given scheduling class.
5564 asmlinkage long sys_sched_get_priority_max(int policy)
5571 ret = MAX_USER_RT_PRIO-1;
5583 * sys_sched_get_priority_min - return minimum RT priority.
5584 * @policy: scheduling class.
5586 * this syscall returns the minimum rt_priority that can be used
5587 * by a given scheduling class.
5589 asmlinkage long sys_sched_get_priority_min(int policy)
5607 * sys_sched_rr_get_interval - return the default timeslice of a process.
5608 * @pid: pid of the process.
5609 * @interval: userspace pointer to the timeslice value.
5611 * this syscall writes the default timeslice value of a given process
5612 * into the user-space timespec buffer. A value of '0' means infinity.
5615 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5617 struct task_struct *p;
5618 unsigned int time_slice;
5626 read_lock(&tasklist_lock);
5627 p = find_process_by_pid(pid);
5631 retval = security_task_getscheduler(p);
5636 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5637 * tasks that are on an otherwise idle runqueue:
5640 if (p->policy == SCHED_RR) {
5641 time_slice = DEF_TIMESLICE;
5642 } else if (p->policy != SCHED_FIFO) {
5643 struct sched_entity *se = &p->se;
5644 unsigned long flags;
5647 rq = task_rq_lock(p, &flags);
5648 if (rq->cfs.load.weight)
5649 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5650 task_rq_unlock(rq, &flags);
5652 read_unlock(&tasklist_lock);
5653 jiffies_to_timespec(time_slice, &t);
5654 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5658 read_unlock(&tasklist_lock);
5662 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5664 void sched_show_task(struct task_struct *p)
5666 unsigned long free = 0;
5669 state = p->state ? __ffs(p->state) + 1 : 0;
5670 printk(KERN_INFO "%-13.13s %c", p->comm,
5671 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5672 #if BITS_PER_LONG == 32
5673 if (state == TASK_RUNNING)
5674 printk(KERN_CONT " running ");
5676 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5678 if (state == TASK_RUNNING)
5679 printk(KERN_CONT " running task ");
5681 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5683 #ifdef CONFIG_DEBUG_STACK_USAGE
5685 unsigned long *n = end_of_stack(p);
5688 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5691 printk(KERN_CONT "%5lu %5d %6d\n", free,
5692 task_pid_nr(p), task_pid_nr(p->real_parent));
5694 show_stack(p, NULL);
5697 void show_state_filter(unsigned long state_filter)
5699 struct task_struct *g, *p;
5701 #if BITS_PER_LONG == 32
5703 " task PC stack pid father\n");
5706 " task PC stack pid father\n");
5708 read_lock(&tasklist_lock);
5709 do_each_thread(g, p) {
5711 * reset the NMI-timeout, listing all files on a slow
5712 * console might take alot of time:
5714 touch_nmi_watchdog();
5715 if (!state_filter || (p->state & state_filter))
5717 } while_each_thread(g, p);
5719 touch_all_softlockup_watchdogs();
5721 #ifdef CONFIG_SCHED_DEBUG
5722 sysrq_sched_debug_show();
5724 read_unlock(&tasklist_lock);
5726 * Only show locks if all tasks are dumped:
5728 if (state_filter == -1)
5729 debug_show_all_locks();
5732 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5734 idle->sched_class = &idle_sched_class;
5738 * init_idle - set up an idle thread for a given CPU
5739 * @idle: task in question
5740 * @cpu: cpu the idle task belongs to
5742 * NOTE: this function does not set the idle thread's NEED_RESCHED
5743 * flag, to make booting more robust.
5745 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5747 struct rq *rq = cpu_rq(cpu);
5748 unsigned long flags;
5751 idle->se.exec_start = sched_clock();
5753 idle->prio = idle->normal_prio = MAX_PRIO;
5754 idle->cpus_allowed = cpumask_of_cpu(cpu);
5755 __set_task_cpu(idle, cpu);
5757 spin_lock_irqsave(&rq->lock, flags);
5758 rq->curr = rq->idle = idle;
5759 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5762 spin_unlock_irqrestore(&rq->lock, flags);
5764 /* Set the preempt count _outside_ the spinlocks! */
5765 #if defined(CONFIG_PREEMPT)
5766 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5768 task_thread_info(idle)->preempt_count = 0;
5771 * The idle tasks have their own, simple scheduling class:
5773 idle->sched_class = &idle_sched_class;
5777 * In a system that switches off the HZ timer nohz_cpu_mask
5778 * indicates which cpus entered this state. This is used
5779 * in the rcu update to wait only for active cpus. For system
5780 * which do not switch off the HZ timer nohz_cpu_mask should
5781 * always be CPU_MASK_NONE.
5783 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5786 * Increase the granularity value when there are more CPUs,
5787 * because with more CPUs the 'effective latency' as visible
5788 * to users decreases. But the relationship is not linear,
5789 * so pick a second-best guess by going with the log2 of the
5792 * This idea comes from the SD scheduler of Con Kolivas:
5794 static inline void sched_init_granularity(void)
5796 unsigned int factor = 1 + ilog2(num_online_cpus());
5797 const unsigned long limit = 200000000;
5799 sysctl_sched_min_granularity *= factor;
5800 if (sysctl_sched_min_granularity > limit)
5801 sysctl_sched_min_granularity = limit;
5803 sysctl_sched_latency *= factor;
5804 if (sysctl_sched_latency > limit)
5805 sysctl_sched_latency = limit;
5807 sysctl_sched_wakeup_granularity *= factor;
5809 sysctl_sched_shares_ratelimit *= factor;
5814 * This is how migration works:
5816 * 1) we queue a struct migration_req structure in the source CPU's
5817 * runqueue and wake up that CPU's migration thread.
5818 * 2) we down() the locked semaphore => thread blocks.
5819 * 3) migration thread wakes up (implicitly it forces the migrated
5820 * thread off the CPU)
5821 * 4) it gets the migration request and checks whether the migrated
5822 * task is still in the wrong runqueue.
5823 * 5) if it's in the wrong runqueue then the migration thread removes
5824 * it and puts it into the right queue.
5825 * 6) migration thread up()s the semaphore.
5826 * 7) we wake up and the migration is done.
5830 * Change a given task's CPU affinity. Migrate the thread to a
5831 * proper CPU and schedule it away if the CPU it's executing on
5832 * is removed from the allowed bitmask.
5834 * NOTE: the caller must have a valid reference to the task, the
5835 * task must not exit() & deallocate itself prematurely. The
5836 * call is not atomic; no spinlocks may be held.
5838 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5840 struct migration_req req;
5841 unsigned long flags;
5845 rq = task_rq_lock(p, &flags);
5846 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5851 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5852 !cpus_equal(p->cpus_allowed, *new_mask))) {
5857 if (p->sched_class->set_cpus_allowed)
5858 p->sched_class->set_cpus_allowed(p, new_mask);
5860 p->cpus_allowed = *new_mask;
5861 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5864 /* Can the task run on the task's current CPU? If so, we're done */
5865 if (cpu_isset(task_cpu(p), *new_mask))
5868 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5869 /* Need help from migration thread: drop lock and wait. */
5870 task_rq_unlock(rq, &flags);
5871 wake_up_process(rq->migration_thread);
5872 wait_for_completion(&req.done);
5873 tlb_migrate_finish(p->mm);
5877 task_rq_unlock(rq, &flags);
5881 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5884 * Move (not current) task off this cpu, onto dest cpu. We're doing
5885 * this because either it can't run here any more (set_cpus_allowed()
5886 * away from this CPU, or CPU going down), or because we're
5887 * attempting to rebalance this task on exec (sched_exec).
5889 * So we race with normal scheduler movements, but that's OK, as long
5890 * as the task is no longer on this CPU.
5892 * Returns non-zero if task was successfully migrated.
5894 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5896 struct rq *rq_dest, *rq_src;
5899 if (unlikely(!cpu_active(dest_cpu)))
5902 rq_src = cpu_rq(src_cpu);
5903 rq_dest = cpu_rq(dest_cpu);
5905 double_rq_lock(rq_src, rq_dest);
5906 /* Already moved. */
5907 if (task_cpu(p) != src_cpu)
5909 /* Affinity changed (again). */
5910 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5913 on_rq = p->se.on_rq;
5915 deactivate_task(rq_src, p, 0);
5917 set_task_cpu(p, dest_cpu);
5919 activate_task(rq_dest, p, 0);
5920 check_preempt_curr(rq_dest, p);
5925 double_rq_unlock(rq_src, rq_dest);
5930 * migration_thread - this is a highprio system thread that performs
5931 * thread migration by bumping thread off CPU then 'pushing' onto
5934 static int migration_thread(void *data)
5936 int cpu = (long)data;
5940 BUG_ON(rq->migration_thread != current);
5942 set_current_state(TASK_INTERRUPTIBLE);
5943 while (!kthread_should_stop()) {
5944 struct migration_req *req;
5945 struct list_head *head;
5947 spin_lock_irq(&rq->lock);
5949 if (cpu_is_offline(cpu)) {
5950 spin_unlock_irq(&rq->lock);
5954 if (rq->active_balance) {
5955 active_load_balance(rq, cpu);
5956 rq->active_balance = 0;
5959 head = &rq->migration_queue;
5961 if (list_empty(head)) {
5962 spin_unlock_irq(&rq->lock);
5964 set_current_state(TASK_INTERRUPTIBLE);
5967 req = list_entry(head->next, struct migration_req, list);
5968 list_del_init(head->next);
5970 spin_unlock(&rq->lock);
5971 __migrate_task(req->task, cpu, req->dest_cpu);
5974 complete(&req->done);
5976 __set_current_state(TASK_RUNNING);
5980 /* Wait for kthread_stop */
5981 set_current_state(TASK_INTERRUPTIBLE);
5982 while (!kthread_should_stop()) {
5984 set_current_state(TASK_INTERRUPTIBLE);
5986 __set_current_state(TASK_RUNNING);
5990 #ifdef CONFIG_HOTPLUG_CPU
5992 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5996 local_irq_disable();
5997 ret = __migrate_task(p, src_cpu, dest_cpu);
6003 * Figure out where task on dead CPU should go, use force if necessary.
6004 * NOTE: interrupts should be disabled by the caller
6006 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6008 unsigned long flags;
6015 mask = node_to_cpumask(cpu_to_node(dead_cpu));
6016 cpus_and(mask, mask, p->cpus_allowed);
6017 dest_cpu = any_online_cpu(mask);
6019 /* On any allowed CPU? */
6020 if (dest_cpu >= nr_cpu_ids)
6021 dest_cpu = any_online_cpu(p->cpus_allowed);
6023 /* No more Mr. Nice Guy. */
6024 if (dest_cpu >= nr_cpu_ids) {
6025 cpumask_t cpus_allowed;
6027 cpuset_cpus_allowed_locked(p, &cpus_allowed);
6029 * Try to stay on the same cpuset, where the
6030 * current cpuset may be a subset of all cpus.
6031 * The cpuset_cpus_allowed_locked() variant of
6032 * cpuset_cpus_allowed() will not block. It must be
6033 * called within calls to cpuset_lock/cpuset_unlock.
6035 rq = task_rq_lock(p, &flags);
6036 p->cpus_allowed = cpus_allowed;
6037 dest_cpu = any_online_cpu(p->cpus_allowed);
6038 task_rq_unlock(rq, &flags);
6041 * Don't tell them about moving exiting tasks or
6042 * kernel threads (both mm NULL), since they never
6045 if (p->mm && printk_ratelimit()) {
6046 printk(KERN_INFO "process %d (%s) no "
6047 "longer affine to cpu%d\n",
6048 task_pid_nr(p), p->comm, dead_cpu);
6051 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
6055 * While a dead CPU has no uninterruptible tasks queued at this point,
6056 * it might still have a nonzero ->nr_uninterruptible counter, because
6057 * for performance reasons the counter is not stricly tracking tasks to
6058 * their home CPUs. So we just add the counter to another CPU's counter,
6059 * to keep the global sum constant after CPU-down:
6061 static void migrate_nr_uninterruptible(struct rq *rq_src)
6063 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
6064 unsigned long flags;
6066 local_irq_save(flags);
6067 double_rq_lock(rq_src, rq_dest);
6068 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6069 rq_src->nr_uninterruptible = 0;
6070 double_rq_unlock(rq_src, rq_dest);
6071 local_irq_restore(flags);
6074 /* Run through task list and migrate tasks from the dead cpu. */
6075 static void migrate_live_tasks(int src_cpu)
6077 struct task_struct *p, *t;
6079 read_lock(&tasklist_lock);
6081 do_each_thread(t, p) {
6085 if (task_cpu(p) == src_cpu)
6086 move_task_off_dead_cpu(src_cpu, p);
6087 } while_each_thread(t, p);
6089 read_unlock(&tasklist_lock);
6093 * Schedules idle task to be the next runnable task on current CPU.
6094 * It does so by boosting its priority to highest possible.
6095 * Used by CPU offline code.
6097 void sched_idle_next(void)
6099 int this_cpu = smp_processor_id();
6100 struct rq *rq = cpu_rq(this_cpu);
6101 struct task_struct *p = rq->idle;
6102 unsigned long flags;
6104 /* cpu has to be offline */
6105 BUG_ON(cpu_online(this_cpu));
6108 * Strictly not necessary since rest of the CPUs are stopped by now
6109 * and interrupts disabled on the current cpu.
6111 spin_lock_irqsave(&rq->lock, flags);
6113 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6115 update_rq_clock(rq);
6116 activate_task(rq, p, 0);
6118 spin_unlock_irqrestore(&rq->lock, flags);
6122 * Ensures that the idle task is using init_mm right before its cpu goes
6125 void idle_task_exit(void)
6127 struct mm_struct *mm = current->active_mm;
6129 BUG_ON(cpu_online(smp_processor_id()));
6132 switch_mm(mm, &init_mm, current);
6136 /* called under rq->lock with disabled interrupts */
6137 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6139 struct rq *rq = cpu_rq(dead_cpu);
6141 /* Must be exiting, otherwise would be on tasklist. */
6142 BUG_ON(!p->exit_state);
6144 /* Cannot have done final schedule yet: would have vanished. */
6145 BUG_ON(p->state == TASK_DEAD);
6150 * Drop lock around migration; if someone else moves it,
6151 * that's OK. No task can be added to this CPU, so iteration is
6154 spin_unlock_irq(&rq->lock);
6155 move_task_off_dead_cpu(dead_cpu, p);
6156 spin_lock_irq(&rq->lock);
6161 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6162 static void migrate_dead_tasks(unsigned int dead_cpu)
6164 struct rq *rq = cpu_rq(dead_cpu);
6165 struct task_struct *next;
6168 if (!rq->nr_running)
6170 update_rq_clock(rq);
6171 next = pick_next_task(rq, rq->curr);
6174 next->sched_class->put_prev_task(rq, next);
6175 migrate_dead(dead_cpu, next);
6179 #endif /* CONFIG_HOTPLUG_CPU */
6181 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6183 static struct ctl_table sd_ctl_dir[] = {
6185 .procname = "sched_domain",
6191 static struct ctl_table sd_ctl_root[] = {
6193 .ctl_name = CTL_KERN,
6194 .procname = "kernel",
6196 .child = sd_ctl_dir,
6201 static struct ctl_table *sd_alloc_ctl_entry(int n)
6203 struct ctl_table *entry =
6204 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6209 static void sd_free_ctl_entry(struct ctl_table **tablep)
6211 struct ctl_table *entry;
6214 * In the intermediate directories, both the child directory and
6215 * procname are dynamically allocated and could fail but the mode
6216 * will always be set. In the lowest directory the names are
6217 * static strings and all have proc handlers.
6219 for (entry = *tablep; entry->mode; entry++) {
6221 sd_free_ctl_entry(&entry->child);
6222 if (entry->proc_handler == NULL)
6223 kfree(entry->procname);
6231 set_table_entry(struct ctl_table *entry,
6232 const char *procname, void *data, int maxlen,
6233 mode_t mode, proc_handler *proc_handler)
6235 entry->procname = procname;
6237 entry->maxlen = maxlen;
6239 entry->proc_handler = proc_handler;
6242 static struct ctl_table *
6243 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6245 struct ctl_table *table = sd_alloc_ctl_entry(12);
6250 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6251 sizeof(long), 0644, proc_doulongvec_minmax);
6252 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6253 sizeof(long), 0644, proc_doulongvec_minmax);
6254 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6255 sizeof(int), 0644, proc_dointvec_minmax);
6256 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6257 sizeof(int), 0644, proc_dointvec_minmax);
6258 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6259 sizeof(int), 0644, proc_dointvec_minmax);
6260 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6261 sizeof(int), 0644, proc_dointvec_minmax);
6262 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6263 sizeof(int), 0644, proc_dointvec_minmax);
6264 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6265 sizeof(int), 0644, proc_dointvec_minmax);
6266 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6267 sizeof(int), 0644, proc_dointvec_minmax);
6268 set_table_entry(&table[9], "cache_nice_tries",
6269 &sd->cache_nice_tries,
6270 sizeof(int), 0644, proc_dointvec_minmax);
6271 set_table_entry(&table[10], "flags", &sd->flags,
6272 sizeof(int), 0644, proc_dointvec_minmax);
6273 /* &table[11] is terminator */
6278 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6280 struct ctl_table *entry, *table;
6281 struct sched_domain *sd;
6282 int domain_num = 0, i;
6285 for_each_domain(cpu, sd)
6287 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6292 for_each_domain(cpu, sd) {
6293 snprintf(buf, 32, "domain%d", i);
6294 entry->procname = kstrdup(buf, GFP_KERNEL);
6296 entry->child = sd_alloc_ctl_domain_table(sd);
6303 static struct ctl_table_header *sd_sysctl_header;
6304 static void register_sched_domain_sysctl(void)
6306 int i, cpu_num = num_online_cpus();
6307 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6310 WARN_ON(sd_ctl_dir[0].child);
6311 sd_ctl_dir[0].child = entry;
6316 for_each_online_cpu(i) {
6317 snprintf(buf, 32, "cpu%d", i);
6318 entry->procname = kstrdup(buf, GFP_KERNEL);
6320 entry->child = sd_alloc_ctl_cpu_table(i);
6324 WARN_ON(sd_sysctl_header);
6325 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6328 /* may be called multiple times per register */
6329 static void unregister_sched_domain_sysctl(void)
6331 if (sd_sysctl_header)
6332 unregister_sysctl_table(sd_sysctl_header);
6333 sd_sysctl_header = NULL;
6334 if (sd_ctl_dir[0].child)
6335 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6338 static void register_sched_domain_sysctl(void)
6341 static void unregister_sched_domain_sysctl(void)
6346 static void set_rq_online(struct rq *rq)
6349 const struct sched_class *class;
6351 cpu_set(rq->cpu, rq->rd->online);
6354 for_each_class(class) {
6355 if (class->rq_online)
6356 class->rq_online(rq);
6361 static void set_rq_offline(struct rq *rq)
6364 const struct sched_class *class;
6366 for_each_class(class) {
6367 if (class->rq_offline)
6368 class->rq_offline(rq);
6371 cpu_clear(rq->cpu, rq->rd->online);
6377 * migration_call - callback that gets triggered when a CPU is added.
6378 * Here we can start up the necessary migration thread for the new CPU.
6380 static int __cpuinit
6381 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6383 struct task_struct *p;
6384 int cpu = (long)hcpu;
6385 unsigned long flags;
6390 case CPU_UP_PREPARE:
6391 case CPU_UP_PREPARE_FROZEN:
6392 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6395 kthread_bind(p, cpu);
6396 /* Must be high prio: stop_machine expects to yield to it. */
6397 rq = task_rq_lock(p, &flags);
6398 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6399 task_rq_unlock(rq, &flags);
6400 cpu_rq(cpu)->migration_thread = p;
6404 case CPU_ONLINE_FROZEN:
6405 /* Strictly unnecessary, as first user will wake it. */
6406 wake_up_process(cpu_rq(cpu)->migration_thread);
6408 /* Update our root-domain */
6410 spin_lock_irqsave(&rq->lock, flags);
6412 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6416 spin_unlock_irqrestore(&rq->lock, flags);
6419 #ifdef CONFIG_HOTPLUG_CPU
6420 case CPU_UP_CANCELED:
6421 case CPU_UP_CANCELED_FROZEN:
6422 if (!cpu_rq(cpu)->migration_thread)
6424 /* Unbind it from offline cpu so it can run. Fall thru. */
6425 kthread_bind(cpu_rq(cpu)->migration_thread,
6426 any_online_cpu(cpu_online_map));
6427 kthread_stop(cpu_rq(cpu)->migration_thread);
6428 cpu_rq(cpu)->migration_thread = NULL;
6432 case CPU_DEAD_FROZEN:
6433 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6434 migrate_live_tasks(cpu);
6436 kthread_stop(rq->migration_thread);
6437 rq->migration_thread = NULL;
6438 /* Idle task back to normal (off runqueue, low prio) */
6439 spin_lock_irq(&rq->lock);
6440 update_rq_clock(rq);
6441 deactivate_task(rq, rq->idle, 0);
6442 rq->idle->static_prio = MAX_PRIO;
6443 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6444 rq->idle->sched_class = &idle_sched_class;
6445 migrate_dead_tasks(cpu);
6446 spin_unlock_irq(&rq->lock);
6448 migrate_nr_uninterruptible(rq);
6449 BUG_ON(rq->nr_running != 0);
6452 * No need to migrate the tasks: it was best-effort if
6453 * they didn't take sched_hotcpu_mutex. Just wake up
6456 spin_lock_irq(&rq->lock);
6457 while (!list_empty(&rq->migration_queue)) {
6458 struct migration_req *req;
6460 req = list_entry(rq->migration_queue.next,
6461 struct migration_req, list);
6462 list_del_init(&req->list);
6463 complete(&req->done);
6465 spin_unlock_irq(&rq->lock);
6469 case CPU_DYING_FROZEN:
6470 /* Update our root-domain */
6472 spin_lock_irqsave(&rq->lock, flags);
6474 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6477 spin_unlock_irqrestore(&rq->lock, flags);
6484 /* Register at highest priority so that task migration (migrate_all_tasks)
6485 * happens before everything else.
6487 static struct notifier_block __cpuinitdata migration_notifier = {
6488 .notifier_call = migration_call,
6492 static int __init migration_init(void)
6494 void *cpu = (void *)(long)smp_processor_id();
6497 /* Start one for the boot CPU: */
6498 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6499 BUG_ON(err == NOTIFY_BAD);
6500 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6501 register_cpu_notifier(&migration_notifier);
6505 early_initcall(migration_init);
6510 #ifdef CONFIG_SCHED_DEBUG
6512 static inline const char *sd_level_to_string(enum sched_domain_level lvl)
6525 case SD_LV_ALLNODES:
6534 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6535 cpumask_t *groupmask)
6537 struct sched_group *group = sd->groups;
6540 cpulist_scnprintf(str, sizeof(str), sd->span);
6541 cpus_clear(*groupmask);
6543 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6545 if (!(sd->flags & SD_LOAD_BALANCE)) {
6546 printk("does not load-balance\n");
6548 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6553 printk(KERN_CONT "span %s level %s\n",
6554 str, sd_level_to_string(sd->level));
6556 if (!cpu_isset(cpu, sd->span)) {
6557 printk(KERN_ERR "ERROR: domain->span does not contain "
6560 if (!cpu_isset(cpu, group->cpumask)) {
6561 printk(KERN_ERR "ERROR: domain->groups does not contain"
6565 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6569 printk(KERN_ERR "ERROR: group is NULL\n");
6573 if (!group->__cpu_power) {
6574 printk(KERN_CONT "\n");
6575 printk(KERN_ERR "ERROR: domain->cpu_power not "
6580 if (!cpus_weight(group->cpumask)) {
6581 printk(KERN_CONT "\n");
6582 printk(KERN_ERR "ERROR: empty group\n");
6586 if (cpus_intersects(*groupmask, group->cpumask)) {
6587 printk(KERN_CONT "\n");
6588 printk(KERN_ERR "ERROR: repeated CPUs\n");
6592 cpus_or(*groupmask, *groupmask, group->cpumask);
6594 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6595 printk(KERN_CONT " %s", str);
6597 group = group->next;
6598 } while (group != sd->groups);
6599 printk(KERN_CONT "\n");
6601 if (!cpus_equal(sd->span, *groupmask))
6602 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6604 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6605 printk(KERN_ERR "ERROR: parent span is not a superset "
6606 "of domain->span\n");
6610 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6612 cpumask_t *groupmask;
6616 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6620 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6622 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6624 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6629 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6638 #else /* !CONFIG_SCHED_DEBUG */
6639 # define sched_domain_debug(sd, cpu) do { } while (0)
6640 #endif /* CONFIG_SCHED_DEBUG */
6642 static int sd_degenerate(struct sched_domain *sd)
6644 if (cpus_weight(sd->span) == 1)
6647 /* Following flags need at least 2 groups */
6648 if (sd->flags & (SD_LOAD_BALANCE |
6649 SD_BALANCE_NEWIDLE |
6653 SD_SHARE_PKG_RESOURCES)) {
6654 if (sd->groups != sd->groups->next)
6658 /* Following flags don't use groups */
6659 if (sd->flags & (SD_WAKE_IDLE |
6668 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6670 unsigned long cflags = sd->flags, pflags = parent->flags;
6672 if (sd_degenerate(parent))
6675 if (!cpus_equal(sd->span, parent->span))
6678 /* Does parent contain flags not in child? */
6679 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6680 if (cflags & SD_WAKE_AFFINE)
6681 pflags &= ~SD_WAKE_BALANCE;
6682 /* Flags needing groups don't count if only 1 group in parent */
6683 if (parent->groups == parent->groups->next) {
6684 pflags &= ~(SD_LOAD_BALANCE |
6685 SD_BALANCE_NEWIDLE |
6689 SD_SHARE_PKG_RESOURCES);
6691 if (~cflags & pflags)
6697 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6699 unsigned long flags;
6701 spin_lock_irqsave(&rq->lock, flags);
6704 struct root_domain *old_rd = rq->rd;
6706 if (cpu_isset(rq->cpu, old_rd->online))
6709 cpu_clear(rq->cpu, old_rd->span);
6711 if (atomic_dec_and_test(&old_rd->refcount))
6715 atomic_inc(&rd->refcount);
6718 cpu_set(rq->cpu, rd->span);
6719 if (cpu_isset(rq->cpu, cpu_online_map))
6722 spin_unlock_irqrestore(&rq->lock, flags);
6725 static void init_rootdomain(struct root_domain *rd)
6727 memset(rd, 0, sizeof(*rd));
6729 cpus_clear(rd->span);
6730 cpus_clear(rd->online);
6732 cpupri_init(&rd->cpupri);
6735 static void init_defrootdomain(void)
6737 init_rootdomain(&def_root_domain);
6738 atomic_set(&def_root_domain.refcount, 1);
6741 static struct root_domain *alloc_rootdomain(void)
6743 struct root_domain *rd;
6745 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6749 init_rootdomain(rd);
6755 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6756 * hold the hotplug lock.
6759 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6761 struct rq *rq = cpu_rq(cpu);
6762 struct sched_domain *tmp;
6764 /* Remove the sched domains which do not contribute to scheduling. */
6765 for (tmp = sd; tmp; tmp = tmp->parent) {
6766 struct sched_domain *parent = tmp->parent;
6769 if (sd_parent_degenerate(tmp, parent)) {
6770 tmp->parent = parent->parent;
6772 parent->parent->child = tmp;
6776 if (sd && sd_degenerate(sd)) {
6782 sched_domain_debug(sd, cpu);
6784 rq_attach_root(rq, rd);
6785 rcu_assign_pointer(rq->sd, sd);
6788 /* cpus with isolated domains */
6789 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6791 /* Setup the mask of cpus configured for isolated domains */
6792 static int __init isolated_cpu_setup(char *str)
6794 static int __initdata ints[NR_CPUS];
6797 str = get_options(str, ARRAY_SIZE(ints), ints);
6798 cpus_clear(cpu_isolated_map);
6799 for (i = 1; i <= ints[0]; i++)
6800 if (ints[i] < NR_CPUS)
6801 cpu_set(ints[i], cpu_isolated_map);
6805 __setup("isolcpus=", isolated_cpu_setup);
6808 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6809 * to a function which identifies what group(along with sched group) a CPU
6810 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6811 * (due to the fact that we keep track of groups covered with a cpumask_t).
6813 * init_sched_build_groups will build a circular linked list of the groups
6814 * covered by the given span, and will set each group's ->cpumask correctly,
6815 * and ->cpu_power to 0.
6818 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6819 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6820 struct sched_group **sg,
6821 cpumask_t *tmpmask),
6822 cpumask_t *covered, cpumask_t *tmpmask)
6824 struct sched_group *first = NULL, *last = NULL;
6827 cpus_clear(*covered);
6829 for_each_cpu_mask_nr(i, *span) {
6830 struct sched_group *sg;
6831 int group = group_fn(i, cpu_map, &sg, tmpmask);
6834 if (cpu_isset(i, *covered))
6837 cpus_clear(sg->cpumask);
6838 sg->__cpu_power = 0;
6840 for_each_cpu_mask_nr(j, *span) {
6841 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6844 cpu_set(j, *covered);
6845 cpu_set(j, sg->cpumask);
6856 #define SD_NODES_PER_DOMAIN 16
6861 * find_next_best_node - find the next node to include in a sched_domain
6862 * @node: node whose sched_domain we're building
6863 * @used_nodes: nodes already in the sched_domain
6865 * Find the next node to include in a given scheduling domain. Simply
6866 * finds the closest node not already in the @used_nodes map.
6868 * Should use nodemask_t.
6870 static int find_next_best_node(int node, nodemask_t *used_nodes)
6872 int i, n, val, min_val, best_node = 0;
6876 for (i = 0; i < nr_node_ids; i++) {
6877 /* Start at @node */
6878 n = (node + i) % nr_node_ids;
6880 if (!nr_cpus_node(n))
6883 /* Skip already used nodes */
6884 if (node_isset(n, *used_nodes))
6887 /* Simple min distance search */
6888 val = node_distance(node, n);
6890 if (val < min_val) {
6896 node_set(best_node, *used_nodes);
6901 * sched_domain_node_span - get a cpumask for a node's sched_domain
6902 * @node: node whose cpumask we're constructing
6903 * @span: resulting cpumask
6905 * Given a node, construct a good cpumask for its sched_domain to span. It
6906 * should be one that prevents unnecessary balancing, but also spreads tasks
6909 static void sched_domain_node_span(int node, cpumask_t *span)
6911 nodemask_t used_nodes;
6912 node_to_cpumask_ptr(nodemask, node);
6916 nodes_clear(used_nodes);
6918 cpus_or(*span, *span, *nodemask);
6919 node_set(node, used_nodes);
6921 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6922 int next_node = find_next_best_node(node, &used_nodes);
6924 node_to_cpumask_ptr_next(nodemask, next_node);
6925 cpus_or(*span, *span, *nodemask);
6928 #endif /* CONFIG_NUMA */
6930 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6933 * SMT sched-domains:
6935 #ifdef CONFIG_SCHED_SMT
6936 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6937 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6940 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6944 *sg = &per_cpu(sched_group_cpus, cpu);
6947 #endif /* CONFIG_SCHED_SMT */
6950 * multi-core sched-domains:
6952 #ifdef CONFIG_SCHED_MC
6953 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6954 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6955 #endif /* CONFIG_SCHED_MC */
6957 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6959 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6964 *mask = per_cpu(cpu_sibling_map, cpu);
6965 cpus_and(*mask, *mask, *cpu_map);
6966 group = first_cpu(*mask);
6968 *sg = &per_cpu(sched_group_core, group);
6971 #elif defined(CONFIG_SCHED_MC)
6973 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6977 *sg = &per_cpu(sched_group_core, cpu);
6982 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6983 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6986 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6990 #ifdef CONFIG_SCHED_MC
6991 *mask = cpu_coregroup_map(cpu);
6992 cpus_and(*mask, *mask, *cpu_map);
6993 group = first_cpu(*mask);
6994 #elif defined(CONFIG_SCHED_SMT)
6995 *mask = per_cpu(cpu_sibling_map, cpu);
6996 cpus_and(*mask, *mask, *cpu_map);
6997 group = first_cpu(*mask);
7002 *sg = &per_cpu(sched_group_phys, group);
7008 * The init_sched_build_groups can't handle what we want to do with node
7009 * groups, so roll our own. Now each node has its own list of groups which
7010 * gets dynamically allocated.
7012 static DEFINE_PER_CPU(struct sched_domain, node_domains);
7013 static struct sched_group ***sched_group_nodes_bycpu;
7015 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
7016 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
7018 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
7019 struct sched_group **sg, cpumask_t *nodemask)
7023 *nodemask = node_to_cpumask(cpu_to_node(cpu));
7024 cpus_and(*nodemask, *nodemask, *cpu_map);
7025 group = first_cpu(*nodemask);
7028 *sg = &per_cpu(sched_group_allnodes, group);
7032 static void init_numa_sched_groups_power(struct sched_group *group_head)
7034 struct sched_group *sg = group_head;
7040 for_each_cpu_mask_nr(j, sg->cpumask) {
7041 struct sched_domain *sd;
7043 sd = &per_cpu(phys_domains, j);
7044 if (j != first_cpu(sd->groups->cpumask)) {
7046 * Only add "power" once for each
7052 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7055 } while (sg != group_head);
7057 #endif /* CONFIG_NUMA */
7060 /* Free memory allocated for various sched_group structures */
7061 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7065 for_each_cpu_mask_nr(cpu, *cpu_map) {
7066 struct sched_group **sched_group_nodes
7067 = sched_group_nodes_bycpu[cpu];
7069 if (!sched_group_nodes)
7072 for (i = 0; i < nr_node_ids; i++) {
7073 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7075 *nodemask = node_to_cpumask(i);
7076 cpus_and(*nodemask, *nodemask, *cpu_map);
7077 if (cpus_empty(*nodemask))
7087 if (oldsg != sched_group_nodes[i])
7090 kfree(sched_group_nodes);
7091 sched_group_nodes_bycpu[cpu] = NULL;
7094 #else /* !CONFIG_NUMA */
7095 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7098 #endif /* CONFIG_NUMA */
7101 * Initialize sched groups cpu_power.
7103 * cpu_power indicates the capacity of sched group, which is used while
7104 * distributing the load between different sched groups in a sched domain.
7105 * Typically cpu_power for all the groups in a sched domain will be same unless
7106 * there are asymmetries in the topology. If there are asymmetries, group
7107 * having more cpu_power will pickup more load compared to the group having
7110 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7111 * the maximum number of tasks a group can handle in the presence of other idle
7112 * or lightly loaded groups in the same sched domain.
7114 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7116 struct sched_domain *child;
7117 struct sched_group *group;
7119 WARN_ON(!sd || !sd->groups);
7121 if (cpu != first_cpu(sd->groups->cpumask))
7126 sd->groups->__cpu_power = 0;
7129 * For perf policy, if the groups in child domain share resources
7130 * (for example cores sharing some portions of the cache hierarchy
7131 * or SMT), then set this domain groups cpu_power such that each group
7132 * can handle only one task, when there are other idle groups in the
7133 * same sched domain.
7135 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7137 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7138 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7143 * add cpu_power of each child group to this groups cpu_power
7145 group = child->groups;
7147 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7148 group = group->next;
7149 } while (group != child->groups);
7153 * Initializers for schedule domains
7154 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7157 #define SD_INIT(sd, type) sd_init_##type(sd)
7158 #define SD_INIT_FUNC(type) \
7159 static noinline void sd_init_##type(struct sched_domain *sd) \
7161 memset(sd, 0, sizeof(*sd)); \
7162 *sd = SD_##type##_INIT; \
7163 sd->level = SD_LV_##type; \
7168 SD_INIT_FUNC(ALLNODES)
7171 #ifdef CONFIG_SCHED_SMT
7172 SD_INIT_FUNC(SIBLING)
7174 #ifdef CONFIG_SCHED_MC
7179 * To minimize stack usage kmalloc room for cpumasks and share the
7180 * space as the usage in build_sched_domains() dictates. Used only
7181 * if the amount of space is significant.
7184 cpumask_t tmpmask; /* make this one first */
7187 cpumask_t this_sibling_map;
7188 cpumask_t this_core_map;
7190 cpumask_t send_covered;
7193 cpumask_t domainspan;
7195 cpumask_t notcovered;
7200 #define SCHED_CPUMASK_ALLOC 1
7201 #define SCHED_CPUMASK_FREE(v) kfree(v)
7202 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7204 #define SCHED_CPUMASK_ALLOC 0
7205 #define SCHED_CPUMASK_FREE(v)
7206 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7209 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7210 ((unsigned long)(a) + offsetof(struct allmasks, v))
7212 static int default_relax_domain_level = -1;
7214 static int __init setup_relax_domain_level(char *str)
7218 val = simple_strtoul(str, NULL, 0);
7219 if (val < SD_LV_MAX)
7220 default_relax_domain_level = val;
7224 __setup("relax_domain_level=", setup_relax_domain_level);
7226 static void set_domain_attribute(struct sched_domain *sd,
7227 struct sched_domain_attr *attr)
7231 if (!attr || attr->relax_domain_level < 0) {
7232 if (default_relax_domain_level < 0)
7235 request = default_relax_domain_level;
7237 request = attr->relax_domain_level;
7238 if (request < sd->level) {
7239 /* turn off idle balance on this domain */
7240 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7242 /* turn on idle balance on this domain */
7243 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7248 * Build sched domains for a given set of cpus and attach the sched domains
7249 * to the individual cpus
7251 static int __build_sched_domains(const cpumask_t *cpu_map,
7252 struct sched_domain_attr *attr)
7255 struct root_domain *rd;
7256 SCHED_CPUMASK_DECLARE(allmasks);
7259 struct sched_group **sched_group_nodes = NULL;
7260 int sd_allnodes = 0;
7263 * Allocate the per-node list of sched groups
7265 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
7267 if (!sched_group_nodes) {
7268 printk(KERN_WARNING "Can not alloc sched group node list\n");
7273 rd = alloc_rootdomain();
7275 printk(KERN_WARNING "Cannot alloc root domain\n");
7277 kfree(sched_group_nodes);
7282 #if SCHED_CPUMASK_ALLOC
7283 /* get space for all scratch cpumask variables */
7284 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
7286 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7289 kfree(sched_group_nodes);
7294 tmpmask = (cpumask_t *)allmasks;
7298 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7302 * Set up domains for cpus specified by the cpu_map.
7304 for_each_cpu_mask_nr(i, *cpu_map) {
7305 struct sched_domain *sd = NULL, *p;
7306 SCHED_CPUMASK_VAR(nodemask, allmasks);
7308 *nodemask = node_to_cpumask(cpu_to_node(i));
7309 cpus_and(*nodemask, *nodemask, *cpu_map);
7312 if (cpus_weight(*cpu_map) >
7313 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7314 sd = &per_cpu(allnodes_domains, i);
7315 SD_INIT(sd, ALLNODES);
7316 set_domain_attribute(sd, attr);
7317 sd->span = *cpu_map;
7318 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7324 sd = &per_cpu(node_domains, i);
7326 set_domain_attribute(sd, attr);
7327 sched_domain_node_span(cpu_to_node(i), &sd->span);
7331 cpus_and(sd->span, sd->span, *cpu_map);
7335 sd = &per_cpu(phys_domains, i);
7337 set_domain_attribute(sd, attr);
7338 sd->span = *nodemask;
7342 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7344 #ifdef CONFIG_SCHED_MC
7346 sd = &per_cpu(core_domains, i);
7348 set_domain_attribute(sd, attr);
7349 sd->span = cpu_coregroup_map(i);
7350 cpus_and(sd->span, sd->span, *cpu_map);
7353 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7356 #ifdef CONFIG_SCHED_SMT
7358 sd = &per_cpu(cpu_domains, i);
7359 SD_INIT(sd, SIBLING);
7360 set_domain_attribute(sd, attr);
7361 sd->span = per_cpu(cpu_sibling_map, i);
7362 cpus_and(sd->span, sd->span, *cpu_map);
7365 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7369 #ifdef CONFIG_SCHED_SMT
7370 /* Set up CPU (sibling) groups */
7371 for_each_cpu_mask_nr(i, *cpu_map) {
7372 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7373 SCHED_CPUMASK_VAR(send_covered, allmasks);
7375 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7376 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7377 if (i != first_cpu(*this_sibling_map))
7380 init_sched_build_groups(this_sibling_map, cpu_map,
7382 send_covered, tmpmask);
7386 #ifdef CONFIG_SCHED_MC
7387 /* Set up multi-core groups */
7388 for_each_cpu_mask_nr(i, *cpu_map) {
7389 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7390 SCHED_CPUMASK_VAR(send_covered, allmasks);
7392 *this_core_map = cpu_coregroup_map(i);
7393 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7394 if (i != first_cpu(*this_core_map))
7397 init_sched_build_groups(this_core_map, cpu_map,
7399 send_covered, tmpmask);
7403 /* Set up physical groups */
7404 for (i = 0; i < nr_node_ids; i++) {
7405 SCHED_CPUMASK_VAR(nodemask, allmasks);
7406 SCHED_CPUMASK_VAR(send_covered, allmasks);
7408 *nodemask = node_to_cpumask(i);
7409 cpus_and(*nodemask, *nodemask, *cpu_map);
7410 if (cpus_empty(*nodemask))
7413 init_sched_build_groups(nodemask, cpu_map,
7415 send_covered, tmpmask);
7419 /* Set up node groups */
7421 SCHED_CPUMASK_VAR(send_covered, allmasks);
7423 init_sched_build_groups(cpu_map, cpu_map,
7424 &cpu_to_allnodes_group,
7425 send_covered, tmpmask);
7428 for (i = 0; i < nr_node_ids; i++) {
7429 /* Set up node groups */
7430 struct sched_group *sg, *prev;
7431 SCHED_CPUMASK_VAR(nodemask, allmasks);
7432 SCHED_CPUMASK_VAR(domainspan, allmasks);
7433 SCHED_CPUMASK_VAR(covered, allmasks);
7436 *nodemask = node_to_cpumask(i);
7437 cpus_clear(*covered);
7439 cpus_and(*nodemask, *nodemask, *cpu_map);
7440 if (cpus_empty(*nodemask)) {
7441 sched_group_nodes[i] = NULL;
7445 sched_domain_node_span(i, domainspan);
7446 cpus_and(*domainspan, *domainspan, *cpu_map);
7448 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7450 printk(KERN_WARNING "Can not alloc domain group for "
7454 sched_group_nodes[i] = sg;
7455 for_each_cpu_mask_nr(j, *nodemask) {
7456 struct sched_domain *sd;
7458 sd = &per_cpu(node_domains, j);
7461 sg->__cpu_power = 0;
7462 sg->cpumask = *nodemask;
7464 cpus_or(*covered, *covered, *nodemask);
7467 for (j = 0; j < nr_node_ids; j++) {
7468 SCHED_CPUMASK_VAR(notcovered, allmasks);
7469 int n = (i + j) % nr_node_ids;
7470 node_to_cpumask_ptr(pnodemask, n);
7472 cpus_complement(*notcovered, *covered);
7473 cpus_and(*tmpmask, *notcovered, *cpu_map);
7474 cpus_and(*tmpmask, *tmpmask, *domainspan);
7475 if (cpus_empty(*tmpmask))
7478 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7479 if (cpus_empty(*tmpmask))
7482 sg = kmalloc_node(sizeof(struct sched_group),
7486 "Can not alloc domain group for node %d\n", j);
7489 sg->__cpu_power = 0;
7490 sg->cpumask = *tmpmask;
7491 sg->next = prev->next;
7492 cpus_or(*covered, *covered, *tmpmask);
7499 /* Calculate CPU power for physical packages and nodes */
7500 #ifdef CONFIG_SCHED_SMT
7501 for_each_cpu_mask_nr(i, *cpu_map) {
7502 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7504 init_sched_groups_power(i, sd);
7507 #ifdef CONFIG_SCHED_MC
7508 for_each_cpu_mask_nr(i, *cpu_map) {
7509 struct sched_domain *sd = &per_cpu(core_domains, i);
7511 init_sched_groups_power(i, sd);
7515 for_each_cpu_mask_nr(i, *cpu_map) {
7516 struct sched_domain *sd = &per_cpu(phys_domains, i);
7518 init_sched_groups_power(i, sd);
7522 for (i = 0; i < nr_node_ids; i++)
7523 init_numa_sched_groups_power(sched_group_nodes[i]);
7526 struct sched_group *sg;
7528 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7530 init_numa_sched_groups_power(sg);
7534 /* Attach the domains */
7535 for_each_cpu_mask_nr(i, *cpu_map) {
7536 struct sched_domain *sd;
7537 #ifdef CONFIG_SCHED_SMT
7538 sd = &per_cpu(cpu_domains, i);
7539 #elif defined(CONFIG_SCHED_MC)
7540 sd = &per_cpu(core_domains, i);
7542 sd = &per_cpu(phys_domains, i);
7544 cpu_attach_domain(sd, rd, i);
7547 SCHED_CPUMASK_FREE((void *)allmasks);
7552 free_sched_groups(cpu_map, tmpmask);
7553 SCHED_CPUMASK_FREE((void *)allmasks);
7558 static int build_sched_domains(const cpumask_t *cpu_map)
7560 return __build_sched_domains(cpu_map, NULL);
7563 static cpumask_t *doms_cur; /* current sched domains */
7564 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7565 static struct sched_domain_attr *dattr_cur;
7566 /* attribues of custom domains in 'doms_cur' */
7569 * Special case: If a kmalloc of a doms_cur partition (array of
7570 * cpumask_t) fails, then fallback to a single sched domain,
7571 * as determined by the single cpumask_t fallback_doms.
7573 static cpumask_t fallback_doms;
7575 void __attribute__((weak)) arch_update_cpu_topology(void)
7580 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7581 * For now this just excludes isolated cpus, but could be used to
7582 * exclude other special cases in the future.
7584 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7588 arch_update_cpu_topology();
7590 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7592 doms_cur = &fallback_doms;
7593 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7595 err = build_sched_domains(doms_cur);
7596 register_sched_domain_sysctl();
7601 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7604 free_sched_groups(cpu_map, tmpmask);
7608 * Detach sched domains from a group of cpus specified in cpu_map
7609 * These cpus will now be attached to the NULL domain
7611 static void detach_destroy_domains(const cpumask_t *cpu_map)
7616 unregister_sched_domain_sysctl();
7618 for_each_cpu_mask_nr(i, *cpu_map)
7619 cpu_attach_domain(NULL, &def_root_domain, i);
7620 synchronize_sched();
7621 arch_destroy_sched_domains(cpu_map, &tmpmask);
7624 /* handle null as "default" */
7625 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7626 struct sched_domain_attr *new, int idx_new)
7628 struct sched_domain_attr tmp;
7635 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7636 new ? (new + idx_new) : &tmp,
7637 sizeof(struct sched_domain_attr));
7641 * Partition sched domains as specified by the 'ndoms_new'
7642 * cpumasks in the array doms_new[] of cpumasks. This compares
7643 * doms_new[] to the current sched domain partitioning, doms_cur[].
7644 * It destroys each deleted domain and builds each new domain.
7646 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7647 * The masks don't intersect (don't overlap.) We should setup one
7648 * sched domain for each mask. CPUs not in any of the cpumasks will
7649 * not be load balanced. If the same cpumask appears both in the
7650 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7653 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7654 * ownership of it and will kfree it when done with it. If the caller
7655 * failed the kmalloc call, then it can pass in doms_new == NULL,
7656 * and partition_sched_domains() will fallback to the single partition
7657 * 'fallback_doms', it also forces the domains to be rebuilt.
7659 * Call with hotplug lock held
7661 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7662 struct sched_domain_attr *dattr_new)
7666 mutex_lock(&sched_domains_mutex);
7668 /* always unregister in case we don't destroy any domains */
7669 unregister_sched_domain_sysctl();
7671 if (doms_new == NULL)
7674 /* Destroy deleted domains */
7675 for (i = 0; i < ndoms_cur; i++) {
7676 for (j = 0; j < ndoms_new; j++) {
7677 if (cpus_equal(doms_cur[i], doms_new[j])
7678 && dattrs_equal(dattr_cur, i, dattr_new, j))
7681 /* no match - a current sched domain not in new doms_new[] */
7682 detach_destroy_domains(doms_cur + i);
7687 if (doms_new == NULL) {
7690 doms_new = &fallback_doms;
7691 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7695 /* Build new domains */
7696 for (i = 0; i < ndoms_new; i++) {
7697 for (j = 0; j < ndoms_cur; j++) {
7698 if (cpus_equal(doms_new[i], doms_cur[j])
7699 && dattrs_equal(dattr_new, i, dattr_cur, j))
7702 /* no match - add a new doms_new */
7703 __build_sched_domains(doms_new + i,
7704 dattr_new ? dattr_new + i : NULL);
7709 /* Remember the new sched domains */
7710 if (doms_cur != &fallback_doms)
7712 kfree(dattr_cur); /* kfree(NULL) is safe */
7713 doms_cur = doms_new;
7714 dattr_cur = dattr_new;
7715 ndoms_cur = ndoms_new;
7717 register_sched_domain_sysctl();
7719 mutex_unlock(&sched_domains_mutex);
7722 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7723 int arch_reinit_sched_domains(void)
7726 rebuild_sched_domains();
7731 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7735 if (buf[0] != '0' && buf[0] != '1')
7739 sched_smt_power_savings = (buf[0] == '1');
7741 sched_mc_power_savings = (buf[0] == '1');
7743 ret = arch_reinit_sched_domains();
7745 return ret ? ret : count;
7748 #ifdef CONFIG_SCHED_MC
7749 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7752 return sprintf(page, "%u\n", sched_mc_power_savings);
7754 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7755 const char *buf, size_t count)
7757 return sched_power_savings_store(buf, count, 0);
7759 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7760 sched_mc_power_savings_show,
7761 sched_mc_power_savings_store);
7764 #ifdef CONFIG_SCHED_SMT
7765 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7768 return sprintf(page, "%u\n", sched_smt_power_savings);
7770 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7771 const char *buf, size_t count)
7773 return sched_power_savings_store(buf, count, 1);
7775 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7776 sched_smt_power_savings_show,
7777 sched_smt_power_savings_store);
7780 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7784 #ifdef CONFIG_SCHED_SMT
7786 err = sysfs_create_file(&cls->kset.kobj,
7787 &attr_sched_smt_power_savings.attr);
7789 #ifdef CONFIG_SCHED_MC
7790 if (!err && mc_capable())
7791 err = sysfs_create_file(&cls->kset.kobj,
7792 &attr_sched_mc_power_savings.attr);
7796 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7798 #ifndef CONFIG_CPUSETS
7800 * Add online and remove offline CPUs from the scheduler domains.
7801 * When cpusets are enabled they take over this function.
7803 static int update_sched_domains(struct notifier_block *nfb,
7804 unsigned long action, void *hcpu)
7808 case CPU_ONLINE_FROZEN:
7810 case CPU_DEAD_FROZEN:
7811 partition_sched_domains(0, NULL, NULL);
7820 static int update_runtime(struct notifier_block *nfb,
7821 unsigned long action, void *hcpu)
7823 int cpu = (int)(long)hcpu;
7826 case CPU_DOWN_PREPARE:
7827 case CPU_DOWN_PREPARE_FROZEN:
7828 disable_runtime(cpu_rq(cpu));
7831 case CPU_DOWN_FAILED:
7832 case CPU_DOWN_FAILED_FROZEN:
7834 case CPU_ONLINE_FROZEN:
7835 enable_runtime(cpu_rq(cpu));
7843 void __init sched_init_smp(void)
7845 cpumask_t non_isolated_cpus;
7847 #if defined(CONFIG_NUMA)
7848 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7850 BUG_ON(sched_group_nodes_bycpu == NULL);
7853 mutex_lock(&sched_domains_mutex);
7854 arch_init_sched_domains(&cpu_online_map);
7855 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7856 if (cpus_empty(non_isolated_cpus))
7857 cpu_set(smp_processor_id(), non_isolated_cpus);
7858 mutex_unlock(&sched_domains_mutex);
7861 #ifndef CONFIG_CPUSETS
7862 /* XXX: Theoretical race here - CPU may be hotplugged now */
7863 hotcpu_notifier(update_sched_domains, 0);
7866 /* RT runtime code needs to handle some hotplug events */
7867 hotcpu_notifier(update_runtime, 0);
7871 /* Move init over to a non-isolated CPU */
7872 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7874 sched_init_granularity();
7877 void __init sched_init_smp(void)
7879 sched_init_granularity();
7881 #endif /* CONFIG_SMP */
7883 int in_sched_functions(unsigned long addr)
7885 return in_lock_functions(addr) ||
7886 (addr >= (unsigned long)__sched_text_start
7887 && addr < (unsigned long)__sched_text_end);
7890 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7892 cfs_rq->tasks_timeline = RB_ROOT;
7893 INIT_LIST_HEAD(&cfs_rq->tasks);
7894 #ifdef CONFIG_FAIR_GROUP_SCHED
7897 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7900 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7902 struct rt_prio_array *array;
7905 array = &rt_rq->active;
7906 for (i = 0; i < MAX_RT_PRIO; i++) {
7907 INIT_LIST_HEAD(array->queue + i);
7908 __clear_bit(i, array->bitmap);
7910 /* delimiter for bitsearch: */
7911 __set_bit(MAX_RT_PRIO, array->bitmap);
7913 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7914 rt_rq->highest_prio = MAX_RT_PRIO;
7917 rt_rq->rt_nr_migratory = 0;
7918 rt_rq->overloaded = 0;
7922 rt_rq->rt_throttled = 0;
7923 rt_rq->rt_runtime = 0;
7924 spin_lock_init(&rt_rq->rt_runtime_lock);
7926 #ifdef CONFIG_RT_GROUP_SCHED
7927 rt_rq->rt_nr_boosted = 0;
7932 #ifdef CONFIG_FAIR_GROUP_SCHED
7933 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7934 struct sched_entity *se, int cpu, int add,
7935 struct sched_entity *parent)
7937 struct rq *rq = cpu_rq(cpu);
7938 tg->cfs_rq[cpu] = cfs_rq;
7939 init_cfs_rq(cfs_rq, rq);
7942 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7945 /* se could be NULL for init_task_group */
7950 se->cfs_rq = &rq->cfs;
7952 se->cfs_rq = parent->my_q;
7955 se->load.weight = tg->shares;
7956 se->load.inv_weight = 0;
7957 se->parent = parent;
7961 #ifdef CONFIG_RT_GROUP_SCHED
7962 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7963 struct sched_rt_entity *rt_se, int cpu, int add,
7964 struct sched_rt_entity *parent)
7966 struct rq *rq = cpu_rq(cpu);
7968 tg->rt_rq[cpu] = rt_rq;
7969 init_rt_rq(rt_rq, rq);
7971 rt_rq->rt_se = rt_se;
7972 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7974 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7976 tg->rt_se[cpu] = rt_se;
7981 rt_se->rt_rq = &rq->rt;
7983 rt_se->rt_rq = parent->my_q;
7985 rt_se->my_q = rt_rq;
7986 rt_se->parent = parent;
7987 INIT_LIST_HEAD(&rt_se->run_list);
7991 void __init sched_init(void)
7994 unsigned long alloc_size = 0, ptr;
7996 #ifdef CONFIG_FAIR_GROUP_SCHED
7997 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7999 #ifdef CONFIG_RT_GROUP_SCHED
8000 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8002 #ifdef CONFIG_USER_SCHED
8006 * As sched_init() is called before page_alloc is setup,
8007 * we use alloc_bootmem().
8010 ptr = (unsigned long)alloc_bootmem(alloc_size);
8012 #ifdef CONFIG_FAIR_GROUP_SCHED
8013 init_task_group.se = (struct sched_entity **)ptr;
8014 ptr += nr_cpu_ids * sizeof(void **);
8016 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8017 ptr += nr_cpu_ids * sizeof(void **);
8019 #ifdef CONFIG_USER_SCHED
8020 root_task_group.se = (struct sched_entity **)ptr;
8021 ptr += nr_cpu_ids * sizeof(void **);
8023 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8024 ptr += nr_cpu_ids * sizeof(void **);
8025 #endif /* CONFIG_USER_SCHED */
8026 #endif /* CONFIG_FAIR_GROUP_SCHED */
8027 #ifdef CONFIG_RT_GROUP_SCHED
8028 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8029 ptr += nr_cpu_ids * sizeof(void **);
8031 init_task_group.rt_rq = (struct rt_rq **)ptr;
8032 ptr += nr_cpu_ids * sizeof(void **);
8034 #ifdef CONFIG_USER_SCHED
8035 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8036 ptr += nr_cpu_ids * sizeof(void **);
8038 root_task_group.rt_rq = (struct rt_rq **)ptr;
8039 ptr += nr_cpu_ids * sizeof(void **);
8040 #endif /* CONFIG_USER_SCHED */
8041 #endif /* CONFIG_RT_GROUP_SCHED */
8045 init_defrootdomain();
8048 init_rt_bandwidth(&def_rt_bandwidth,
8049 global_rt_period(), global_rt_runtime());
8051 #ifdef CONFIG_RT_GROUP_SCHED
8052 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8053 global_rt_period(), global_rt_runtime());
8054 #ifdef CONFIG_USER_SCHED
8055 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8056 global_rt_period(), RUNTIME_INF);
8057 #endif /* CONFIG_USER_SCHED */
8058 #endif /* CONFIG_RT_GROUP_SCHED */
8060 #ifdef CONFIG_GROUP_SCHED
8061 list_add(&init_task_group.list, &task_groups);
8062 INIT_LIST_HEAD(&init_task_group.children);
8064 #ifdef CONFIG_USER_SCHED
8065 INIT_LIST_HEAD(&root_task_group.children);
8066 init_task_group.parent = &root_task_group;
8067 list_add(&init_task_group.siblings, &root_task_group.children);
8068 #endif /* CONFIG_USER_SCHED */
8069 #endif /* CONFIG_GROUP_SCHED */
8071 for_each_possible_cpu(i) {
8075 spin_lock_init(&rq->lock);
8077 init_cfs_rq(&rq->cfs, rq);
8078 init_rt_rq(&rq->rt, rq);
8079 #ifdef CONFIG_FAIR_GROUP_SCHED
8080 init_task_group.shares = init_task_group_load;
8081 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8082 #ifdef CONFIG_CGROUP_SCHED
8084 * How much cpu bandwidth does init_task_group get?
8086 * In case of task-groups formed thr' the cgroup filesystem, it
8087 * gets 100% of the cpu resources in the system. This overall
8088 * system cpu resource is divided among the tasks of
8089 * init_task_group and its child task-groups in a fair manner,
8090 * based on each entity's (task or task-group's) weight
8091 * (se->load.weight).
8093 * In other words, if init_task_group has 10 tasks of weight
8094 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8095 * then A0's share of the cpu resource is:
8097 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8099 * We achieve this by letting init_task_group's tasks sit
8100 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8102 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8103 #elif defined CONFIG_USER_SCHED
8104 root_task_group.shares = NICE_0_LOAD;
8105 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8107 * In case of task-groups formed thr' the user id of tasks,
8108 * init_task_group represents tasks belonging to root user.
8109 * Hence it forms a sibling of all subsequent groups formed.
8110 * In this case, init_task_group gets only a fraction of overall
8111 * system cpu resource, based on the weight assigned to root
8112 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8113 * by letting tasks of init_task_group sit in a separate cfs_rq
8114 * (init_cfs_rq) and having one entity represent this group of
8115 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8117 init_tg_cfs_entry(&init_task_group,
8118 &per_cpu(init_cfs_rq, i),
8119 &per_cpu(init_sched_entity, i), i, 1,
8120 root_task_group.se[i]);
8123 #endif /* CONFIG_FAIR_GROUP_SCHED */
8125 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8126 #ifdef CONFIG_RT_GROUP_SCHED
8127 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8128 #ifdef CONFIG_CGROUP_SCHED
8129 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8130 #elif defined CONFIG_USER_SCHED
8131 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8132 init_tg_rt_entry(&init_task_group,
8133 &per_cpu(init_rt_rq, i),
8134 &per_cpu(init_sched_rt_entity, i), i, 1,
8135 root_task_group.rt_se[i]);
8139 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8140 rq->cpu_load[j] = 0;
8144 rq->active_balance = 0;
8145 rq->next_balance = jiffies;
8149 rq->migration_thread = NULL;
8150 INIT_LIST_HEAD(&rq->migration_queue);
8151 rq_attach_root(rq, &def_root_domain);
8154 atomic_set(&rq->nr_iowait, 0);
8157 set_load_weight(&init_task);
8159 #ifdef CONFIG_PREEMPT_NOTIFIERS
8160 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8164 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8167 #ifdef CONFIG_RT_MUTEXES
8168 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8172 * The boot idle thread does lazy MMU switching as well:
8174 atomic_inc(&init_mm.mm_count);
8175 enter_lazy_tlb(&init_mm, current);
8178 * Make us the idle thread. Technically, schedule() should not be
8179 * called from this thread, however somewhere below it might be,
8180 * but because we are the idle thread, we just pick up running again
8181 * when this runqueue becomes "idle".
8183 init_idle(current, smp_processor_id());
8185 * During early bootup we pretend to be a normal task:
8187 current->sched_class = &fair_sched_class;
8189 scheduler_running = 1;
8192 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8193 void __might_sleep(char *file, int line)
8196 static unsigned long prev_jiffy; /* ratelimiting */
8198 if ((in_atomic() || irqs_disabled()) &&
8199 system_state == SYSTEM_RUNNING && !oops_in_progress) {
8200 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8202 prev_jiffy = jiffies;
8203 printk(KERN_ERR "BUG: sleeping function called from invalid"
8204 " context at %s:%d\n", file, line);
8205 printk("in_atomic():%d, irqs_disabled():%d\n",
8206 in_atomic(), irqs_disabled());
8207 debug_show_held_locks(current);
8208 if (irqs_disabled())
8209 print_irqtrace_events(current);
8214 EXPORT_SYMBOL(__might_sleep);
8217 #ifdef CONFIG_MAGIC_SYSRQ
8218 static void normalize_task(struct rq *rq, struct task_struct *p)
8222 update_rq_clock(rq);
8223 on_rq = p->se.on_rq;
8225 deactivate_task(rq, p, 0);
8226 __setscheduler(rq, p, SCHED_NORMAL, 0);
8228 activate_task(rq, p, 0);
8229 resched_task(rq->curr);
8233 void normalize_rt_tasks(void)
8235 struct task_struct *g, *p;
8236 unsigned long flags;
8239 read_lock_irqsave(&tasklist_lock, flags);
8240 do_each_thread(g, p) {
8242 * Only normalize user tasks:
8247 p->se.exec_start = 0;
8248 #ifdef CONFIG_SCHEDSTATS
8249 p->se.wait_start = 0;
8250 p->se.sleep_start = 0;
8251 p->se.block_start = 0;
8256 * Renice negative nice level userspace
8259 if (TASK_NICE(p) < 0 && p->mm)
8260 set_user_nice(p, 0);
8264 spin_lock(&p->pi_lock);
8265 rq = __task_rq_lock(p);
8267 normalize_task(rq, p);
8269 __task_rq_unlock(rq);
8270 spin_unlock(&p->pi_lock);
8271 } while_each_thread(g, p);
8273 read_unlock_irqrestore(&tasklist_lock, flags);
8276 #endif /* CONFIG_MAGIC_SYSRQ */
8280 * These functions are only useful for the IA64 MCA handling.
8282 * They can only be called when the whole system has been
8283 * stopped - every CPU needs to be quiescent, and no scheduling
8284 * activity can take place. Using them for anything else would
8285 * be a serious bug, and as a result, they aren't even visible
8286 * under any other configuration.
8290 * curr_task - return the current task for a given cpu.
8291 * @cpu: the processor in question.
8293 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8295 struct task_struct *curr_task(int cpu)
8297 return cpu_curr(cpu);
8301 * set_curr_task - set the current task for a given cpu.
8302 * @cpu: the processor in question.
8303 * @p: the task pointer to set.
8305 * Description: This function must only be used when non-maskable interrupts
8306 * are serviced on a separate stack. It allows the architecture to switch the
8307 * notion of the current task on a cpu in a non-blocking manner. This function
8308 * must be called with all CPU's synchronized, and interrupts disabled, the
8309 * and caller must save the original value of the current task (see
8310 * curr_task() above) and restore that value before reenabling interrupts and
8311 * re-starting the system.
8313 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8315 void set_curr_task(int cpu, struct task_struct *p)
8322 #ifdef CONFIG_FAIR_GROUP_SCHED
8323 static void free_fair_sched_group(struct task_group *tg)
8327 for_each_possible_cpu(i) {
8329 kfree(tg->cfs_rq[i]);
8339 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8341 struct cfs_rq *cfs_rq;
8342 struct sched_entity *se, *parent_se;
8346 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8349 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8353 tg->shares = NICE_0_LOAD;
8355 for_each_possible_cpu(i) {
8358 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8359 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8363 se = kmalloc_node(sizeof(struct sched_entity),
8364 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8368 parent_se = parent ? parent->se[i] : NULL;
8369 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
8378 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8380 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8381 &cpu_rq(cpu)->leaf_cfs_rq_list);
8384 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8386 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8388 #else /* !CONFG_FAIR_GROUP_SCHED */
8389 static inline void free_fair_sched_group(struct task_group *tg)
8394 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8399 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8403 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8406 #endif /* CONFIG_FAIR_GROUP_SCHED */
8408 #ifdef CONFIG_RT_GROUP_SCHED
8409 static void free_rt_sched_group(struct task_group *tg)
8413 destroy_rt_bandwidth(&tg->rt_bandwidth);
8415 for_each_possible_cpu(i) {
8417 kfree(tg->rt_rq[i]);
8419 kfree(tg->rt_se[i]);
8427 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8429 struct rt_rq *rt_rq;
8430 struct sched_rt_entity *rt_se, *parent_se;
8434 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8437 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8441 init_rt_bandwidth(&tg->rt_bandwidth,
8442 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8444 for_each_possible_cpu(i) {
8447 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8448 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8452 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8453 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8457 parent_se = parent ? parent->rt_se[i] : NULL;
8458 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
8467 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8469 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8470 &cpu_rq(cpu)->leaf_rt_rq_list);
8473 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8475 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8477 #else /* !CONFIG_RT_GROUP_SCHED */
8478 static inline void free_rt_sched_group(struct task_group *tg)
8483 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8488 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8492 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8495 #endif /* CONFIG_RT_GROUP_SCHED */
8497 #ifdef CONFIG_GROUP_SCHED
8498 static void free_sched_group(struct task_group *tg)
8500 free_fair_sched_group(tg);
8501 free_rt_sched_group(tg);
8505 /* allocate runqueue etc for a new task group */
8506 struct task_group *sched_create_group(struct task_group *parent)
8508 struct task_group *tg;
8509 unsigned long flags;
8512 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8514 return ERR_PTR(-ENOMEM);
8516 if (!alloc_fair_sched_group(tg, parent))
8519 if (!alloc_rt_sched_group(tg, parent))
8522 spin_lock_irqsave(&task_group_lock, flags);
8523 for_each_possible_cpu(i) {
8524 register_fair_sched_group(tg, i);
8525 register_rt_sched_group(tg, i);
8527 list_add_rcu(&tg->list, &task_groups);
8529 WARN_ON(!parent); /* root should already exist */
8531 tg->parent = parent;
8532 INIT_LIST_HEAD(&tg->children);
8533 list_add_rcu(&tg->siblings, &parent->children);
8534 spin_unlock_irqrestore(&task_group_lock, flags);
8539 free_sched_group(tg);
8540 return ERR_PTR(-ENOMEM);
8543 /* rcu callback to free various structures associated with a task group */
8544 static void free_sched_group_rcu(struct rcu_head *rhp)
8546 /* now it should be safe to free those cfs_rqs */
8547 free_sched_group(container_of(rhp, struct task_group, rcu));
8550 /* Destroy runqueue etc associated with a task group */
8551 void sched_destroy_group(struct task_group *tg)
8553 unsigned long flags;
8556 spin_lock_irqsave(&task_group_lock, flags);
8557 for_each_possible_cpu(i) {
8558 unregister_fair_sched_group(tg, i);
8559 unregister_rt_sched_group(tg, i);
8561 list_del_rcu(&tg->list);
8562 list_del_rcu(&tg->siblings);
8563 spin_unlock_irqrestore(&task_group_lock, flags);
8565 /* wait for possible concurrent references to cfs_rqs complete */
8566 call_rcu(&tg->rcu, free_sched_group_rcu);
8569 /* change task's runqueue when it moves between groups.
8570 * The caller of this function should have put the task in its new group
8571 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8572 * reflect its new group.
8574 void sched_move_task(struct task_struct *tsk)
8577 unsigned long flags;
8580 rq = task_rq_lock(tsk, &flags);
8582 update_rq_clock(rq);
8584 running = task_current(rq, tsk);
8585 on_rq = tsk->se.on_rq;
8588 dequeue_task(rq, tsk, 0);
8589 if (unlikely(running))
8590 tsk->sched_class->put_prev_task(rq, tsk);
8592 set_task_rq(tsk, task_cpu(tsk));
8594 #ifdef CONFIG_FAIR_GROUP_SCHED
8595 if (tsk->sched_class->moved_group)
8596 tsk->sched_class->moved_group(tsk);
8599 if (unlikely(running))
8600 tsk->sched_class->set_curr_task(rq);
8602 enqueue_task(rq, tsk, 0);
8604 task_rq_unlock(rq, &flags);
8606 #endif /* CONFIG_GROUP_SCHED */
8608 #ifdef CONFIG_FAIR_GROUP_SCHED
8609 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8611 struct cfs_rq *cfs_rq = se->cfs_rq;
8616 dequeue_entity(cfs_rq, se, 0);
8618 se->load.weight = shares;
8619 se->load.inv_weight = 0;
8622 enqueue_entity(cfs_rq, se, 0);
8625 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8627 struct cfs_rq *cfs_rq = se->cfs_rq;
8628 struct rq *rq = cfs_rq->rq;
8629 unsigned long flags;
8631 spin_lock_irqsave(&rq->lock, flags);
8632 __set_se_shares(se, shares);
8633 spin_unlock_irqrestore(&rq->lock, flags);
8636 static DEFINE_MUTEX(shares_mutex);
8638 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8641 unsigned long flags;
8644 * We can't change the weight of the root cgroup.
8649 if (shares < MIN_SHARES)
8650 shares = MIN_SHARES;
8651 else if (shares > MAX_SHARES)
8652 shares = MAX_SHARES;
8654 mutex_lock(&shares_mutex);
8655 if (tg->shares == shares)
8658 spin_lock_irqsave(&task_group_lock, flags);
8659 for_each_possible_cpu(i)
8660 unregister_fair_sched_group(tg, i);
8661 list_del_rcu(&tg->siblings);
8662 spin_unlock_irqrestore(&task_group_lock, flags);
8664 /* wait for any ongoing reference to this group to finish */
8665 synchronize_sched();
8668 * Now we are free to modify the group's share on each cpu
8669 * w/o tripping rebalance_share or load_balance_fair.
8671 tg->shares = shares;
8672 for_each_possible_cpu(i) {
8676 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8677 set_se_shares(tg->se[i], shares);
8681 * Enable load balance activity on this group, by inserting it back on
8682 * each cpu's rq->leaf_cfs_rq_list.
8684 spin_lock_irqsave(&task_group_lock, flags);
8685 for_each_possible_cpu(i)
8686 register_fair_sched_group(tg, i);
8687 list_add_rcu(&tg->siblings, &tg->parent->children);
8688 spin_unlock_irqrestore(&task_group_lock, flags);
8690 mutex_unlock(&shares_mutex);
8694 unsigned long sched_group_shares(struct task_group *tg)
8700 #ifdef CONFIG_RT_GROUP_SCHED
8702 * Ensure that the real time constraints are schedulable.
8704 static DEFINE_MUTEX(rt_constraints_mutex);
8706 static unsigned long to_ratio(u64 period, u64 runtime)
8708 if (runtime == RUNTIME_INF)
8711 return div64_u64(runtime << 20, period);
8714 /* Must be called with tasklist_lock held */
8715 static inline int tg_has_rt_tasks(struct task_group *tg)
8717 struct task_struct *g, *p;
8719 do_each_thread(g, p) {
8720 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8722 } while_each_thread(g, p);
8727 struct rt_schedulable_data {
8728 struct task_group *tg;
8733 static int tg_schedulable(struct task_group *tg, void *data)
8735 struct rt_schedulable_data *d = data;
8736 struct task_group *child;
8737 unsigned long total, sum = 0;
8738 u64 period, runtime;
8740 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8741 runtime = tg->rt_bandwidth.rt_runtime;
8744 period = d->rt_period;
8745 runtime = d->rt_runtime;
8748 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8751 total = to_ratio(period, runtime);
8753 list_for_each_entry_rcu(child, &tg->children, siblings) {
8754 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8755 runtime = child->rt_bandwidth.rt_runtime;
8757 if (child == d->tg) {
8758 period = d->rt_period;
8759 runtime = d->rt_runtime;
8762 sum += to_ratio(period, runtime);
8771 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8773 struct rt_schedulable_data data = {
8775 .rt_period = period,
8776 .rt_runtime = runtime,
8779 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8782 static int tg_set_bandwidth(struct task_group *tg,
8783 u64 rt_period, u64 rt_runtime)
8787 mutex_lock(&rt_constraints_mutex);
8788 read_lock(&tasklist_lock);
8789 err = __rt_schedulable(tg, rt_period, rt_runtime);
8793 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8794 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8795 tg->rt_bandwidth.rt_runtime = rt_runtime;
8797 for_each_possible_cpu(i) {
8798 struct rt_rq *rt_rq = tg->rt_rq[i];
8800 spin_lock(&rt_rq->rt_runtime_lock);
8801 rt_rq->rt_runtime = rt_runtime;
8802 spin_unlock(&rt_rq->rt_runtime_lock);
8804 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8806 read_unlock(&tasklist_lock);
8807 mutex_unlock(&rt_constraints_mutex);
8812 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8814 u64 rt_runtime, rt_period;
8816 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8817 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8818 if (rt_runtime_us < 0)
8819 rt_runtime = RUNTIME_INF;
8821 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8824 long sched_group_rt_runtime(struct task_group *tg)
8828 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8831 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8832 do_div(rt_runtime_us, NSEC_PER_USEC);
8833 return rt_runtime_us;
8836 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8838 u64 rt_runtime, rt_period;
8840 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8841 rt_runtime = tg->rt_bandwidth.rt_runtime;
8846 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8849 long sched_group_rt_period(struct task_group *tg)
8853 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8854 do_div(rt_period_us, NSEC_PER_USEC);
8855 return rt_period_us;
8858 static int sched_rt_global_constraints(void)
8860 struct task_group *tg = &root_task_group;
8861 u64 rt_runtime, rt_period;
8864 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8865 rt_runtime = tg->rt_bandwidth.rt_runtime;
8867 mutex_lock(&rt_constraints_mutex);
8868 read_lock(&tasklist_lock);
8869 ret = __rt_schedulable(tg, rt_period, rt_runtime);
8870 read_unlock(&tasklist_lock);
8871 mutex_unlock(&rt_constraints_mutex);
8875 #else /* !CONFIG_RT_GROUP_SCHED */
8876 static int sched_rt_global_constraints(void)
8878 unsigned long flags;
8881 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8882 for_each_possible_cpu(i) {
8883 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8885 spin_lock(&rt_rq->rt_runtime_lock);
8886 rt_rq->rt_runtime = global_rt_runtime();
8887 spin_unlock(&rt_rq->rt_runtime_lock);
8889 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8893 #endif /* CONFIG_RT_GROUP_SCHED */
8895 int sched_rt_handler(struct ctl_table *table, int write,
8896 struct file *filp, void __user *buffer, size_t *lenp,
8900 int old_period, old_runtime;
8901 static DEFINE_MUTEX(mutex);
8904 old_period = sysctl_sched_rt_period;
8905 old_runtime = sysctl_sched_rt_runtime;
8907 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8909 if (!ret && write) {
8910 ret = sched_rt_global_constraints();
8912 sysctl_sched_rt_period = old_period;
8913 sysctl_sched_rt_runtime = old_runtime;
8915 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8916 def_rt_bandwidth.rt_period =
8917 ns_to_ktime(global_rt_period());
8920 mutex_unlock(&mutex);
8925 #ifdef CONFIG_CGROUP_SCHED
8927 /* return corresponding task_group object of a cgroup */
8928 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8930 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8931 struct task_group, css);
8934 static struct cgroup_subsys_state *
8935 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8937 struct task_group *tg, *parent;
8939 if (!cgrp->parent) {
8940 /* This is early initialization for the top cgroup */
8941 init_task_group.css.cgroup = cgrp;
8942 return &init_task_group.css;
8945 parent = cgroup_tg(cgrp->parent);
8946 tg = sched_create_group(parent);
8948 return ERR_PTR(-ENOMEM);
8950 /* Bind the cgroup to task_group object we just created */
8951 tg->css.cgroup = cgrp;
8957 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8959 struct task_group *tg = cgroup_tg(cgrp);
8961 sched_destroy_group(tg);
8965 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8966 struct task_struct *tsk)
8968 #ifdef CONFIG_RT_GROUP_SCHED
8969 /* Don't accept realtime tasks when there is no way for them to run */
8970 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
8973 /* We don't support RT-tasks being in separate groups */
8974 if (tsk->sched_class != &fair_sched_class)
8982 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8983 struct cgroup *old_cont, struct task_struct *tsk)
8985 sched_move_task(tsk);
8988 #ifdef CONFIG_FAIR_GROUP_SCHED
8989 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8992 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8995 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8997 struct task_group *tg = cgroup_tg(cgrp);
8999 return (u64) tg->shares;
9001 #endif /* CONFIG_FAIR_GROUP_SCHED */
9003 #ifdef CONFIG_RT_GROUP_SCHED
9004 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9007 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9010 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9012 return sched_group_rt_runtime(cgroup_tg(cgrp));
9015 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9018 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9021 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9023 return sched_group_rt_period(cgroup_tg(cgrp));
9025 #endif /* CONFIG_RT_GROUP_SCHED */
9027 static struct cftype cpu_files[] = {
9028 #ifdef CONFIG_FAIR_GROUP_SCHED
9031 .read_u64 = cpu_shares_read_u64,
9032 .write_u64 = cpu_shares_write_u64,
9035 #ifdef CONFIG_RT_GROUP_SCHED
9037 .name = "rt_runtime_us",
9038 .read_s64 = cpu_rt_runtime_read,
9039 .write_s64 = cpu_rt_runtime_write,
9042 .name = "rt_period_us",
9043 .read_u64 = cpu_rt_period_read_uint,
9044 .write_u64 = cpu_rt_period_write_uint,
9049 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9051 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9054 struct cgroup_subsys cpu_cgroup_subsys = {
9056 .create = cpu_cgroup_create,
9057 .destroy = cpu_cgroup_destroy,
9058 .can_attach = cpu_cgroup_can_attach,
9059 .attach = cpu_cgroup_attach,
9060 .populate = cpu_cgroup_populate,
9061 .subsys_id = cpu_cgroup_subsys_id,
9065 #endif /* CONFIG_CGROUP_SCHED */
9067 #ifdef CONFIG_CGROUP_CPUACCT
9070 * CPU accounting code for task groups.
9072 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9073 * (balbir@in.ibm.com).
9076 /* track cpu usage of a group of tasks */
9078 struct cgroup_subsys_state css;
9079 /* cpuusage holds pointer to a u64-type object on every cpu */
9083 struct cgroup_subsys cpuacct_subsys;
9085 /* return cpu accounting group corresponding to this container */
9086 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9088 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9089 struct cpuacct, css);
9092 /* return cpu accounting group to which this task belongs */
9093 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9095 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9096 struct cpuacct, css);
9099 /* create a new cpu accounting group */
9100 static struct cgroup_subsys_state *cpuacct_create(
9101 struct cgroup_subsys *ss, struct cgroup *cgrp)
9103 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9106 return ERR_PTR(-ENOMEM);
9108 ca->cpuusage = alloc_percpu(u64);
9109 if (!ca->cpuusage) {
9111 return ERR_PTR(-ENOMEM);
9117 /* destroy an existing cpu accounting group */
9119 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9121 struct cpuacct *ca = cgroup_ca(cgrp);
9123 free_percpu(ca->cpuusage);
9127 /* return total cpu usage (in nanoseconds) of a group */
9128 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9130 struct cpuacct *ca = cgroup_ca(cgrp);
9131 u64 totalcpuusage = 0;
9134 for_each_possible_cpu(i) {
9135 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9138 * Take rq->lock to make 64-bit addition safe on 32-bit
9141 spin_lock_irq(&cpu_rq(i)->lock);
9142 totalcpuusage += *cpuusage;
9143 spin_unlock_irq(&cpu_rq(i)->lock);
9146 return totalcpuusage;
9149 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9152 struct cpuacct *ca = cgroup_ca(cgrp);
9161 for_each_possible_cpu(i) {
9162 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9164 spin_lock_irq(&cpu_rq(i)->lock);
9166 spin_unlock_irq(&cpu_rq(i)->lock);
9172 static struct cftype files[] = {
9175 .read_u64 = cpuusage_read,
9176 .write_u64 = cpuusage_write,
9180 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9182 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9186 * charge this task's execution time to its accounting group.
9188 * called with rq->lock held.
9190 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9194 if (!cpuacct_subsys.active)
9199 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
9201 *cpuusage += cputime;
9205 struct cgroup_subsys cpuacct_subsys = {
9207 .create = cpuacct_create,
9208 .destroy = cpuacct_destroy,
9209 .populate = cpuacct_populate,
9210 .subsys_id = cpuacct_subsys_id,
9212 #endif /* CONFIG_CGROUP_CPUACCT */